From 7cd468a3d7dee7d6c92f69a0bb7061ae208ec727 Mon Sep 17 00:00:00 2001 From: Damjan Marion Date: Mon, 19 Dec 2016 23:05:39 +0100 Subject: Reorganize source tree to use single autotools instance Change-Id: I7b51f88292e057c6443b12224486f2d0c9f8ae23 Signed-off-by: Damjan Marion --- src/vnet/devices/af_packet/af_packet.api | 71 + src/vnet/devices/af_packet/af_packet.c | 366 +++ src/vnet/devices/af_packet/af_packet.h | 69 + src/vnet/devices/af_packet/af_packet_api.c | 143 + src/vnet/devices/af_packet/cli.c | 144 + src/vnet/devices/af_packet/device.c | 250 ++ src/vnet/devices/af_packet/node.c | 288 ++ src/vnet/devices/devices.c | 91 + src/vnet/devices/devices.h | 53 + src/vnet/devices/dpdk/cli.c | 1296 ++++++++ src/vnet/devices/dpdk/device.c | 840 +++++ src/vnet/devices/dpdk/dpdk.h | 534 ++++ src/vnet/devices/dpdk/dpdk_priv.h | 132 + src/vnet/devices/dpdk/format.c | 763 +++++ src/vnet/devices/dpdk/hqos.c | 775 +++++ src/vnet/devices/dpdk/init.c | 1803 +++++++++++ src/vnet/devices/dpdk/ipsec/cli.c | 141 + src/vnet/devices/dpdk/ipsec/crypto_node.c | 210 ++ src/vnet/devices/dpdk/ipsec/dir.dox | 18 + .../devices/dpdk/ipsec/dpdk_crypto_ipsec_doc.md | 73 + src/vnet/devices/dpdk/ipsec/esp.h | 295 ++ src/vnet/devices/dpdk/ipsec/esp_decrypt.c | 583 ++++ src/vnet/devices/dpdk/ipsec/esp_encrypt.c | 598 ++++ src/vnet/devices/dpdk/ipsec/ipsec.c | 313 ++ src/vnet/devices/dpdk/ipsec/ipsec.h | 227 ++ src/vnet/devices/dpdk/node.c | 687 ++++ src/vnet/devices/dpdk/qos_doc.md | 404 +++ src/vnet/devices/netmap/cli.c | 146 + src/vnet/devices/netmap/device.c | 261 ++ src/vnet/devices/netmap/net_netmap.h | 650 ++++ src/vnet/devices/netmap/netmap.api | 74 + src/vnet/devices/netmap/netmap.c | 316 ++ src/vnet/devices/netmap/netmap.h | 164 + src/vnet/devices/netmap/netmap_api.c | 137 + src/vnet/devices/netmap/node.c | 300 ++ src/vnet/devices/nic/ixge.c | 2938 +++++++++++++++++ src/vnet/devices/nic/ixge.h | 1293 ++++++++ src/vnet/devices/nic/sfp.c | 117 + src/vnet/devices/nic/sfp.h | 117 + src/vnet/devices/ssvm/node.c | 343 ++ src/vnet/devices/ssvm/ssvm_eth.c | 491 +++ src/vnet/devices/ssvm/ssvm_eth.h | 141 + src/vnet/devices/virtio/dir.dox | 27 + src/vnet/devices/virtio/vhost-user.c | 3314 ++++++++++++++++++++ src/vnet/devices/virtio/vhost-user.h | 350 +++ src/vnet/devices/virtio/vhost_user.api | 125 + src/vnet/devices/virtio/vhost_user_api.c | 262 ++ 47 files changed, 22733 insertions(+) create mode 100644 src/vnet/devices/af_packet/af_packet.api create mode 100644 src/vnet/devices/af_packet/af_packet.c create mode 100644 src/vnet/devices/af_packet/af_packet.h create mode 100644 src/vnet/devices/af_packet/af_packet_api.c create mode 100644 src/vnet/devices/af_packet/cli.c create mode 100644 src/vnet/devices/af_packet/device.c create mode 100644 src/vnet/devices/af_packet/node.c create mode 100644 src/vnet/devices/devices.c create mode 100644 src/vnet/devices/devices.h create mode 100644 src/vnet/devices/dpdk/cli.c create mode 100644 src/vnet/devices/dpdk/device.c create mode 100644 src/vnet/devices/dpdk/dpdk.h create mode 100644 src/vnet/devices/dpdk/dpdk_priv.h create mode 100644 src/vnet/devices/dpdk/format.c create mode 100644 src/vnet/devices/dpdk/hqos.c create mode 100755 src/vnet/devices/dpdk/init.c create mode 100644 src/vnet/devices/dpdk/ipsec/cli.c create mode 100644 src/vnet/devices/dpdk/ipsec/crypto_node.c create mode 100644 src/vnet/devices/dpdk/ipsec/dir.dox create mode 100644 src/vnet/devices/dpdk/ipsec/dpdk_crypto_ipsec_doc.md create mode 100644 src/vnet/devices/dpdk/ipsec/esp.h create mode 100644 src/vnet/devices/dpdk/ipsec/esp_decrypt.c create mode 100644 src/vnet/devices/dpdk/ipsec/esp_encrypt.c create mode 100644 src/vnet/devices/dpdk/ipsec/ipsec.c create mode 100644 src/vnet/devices/dpdk/ipsec/ipsec.h create mode 100644 src/vnet/devices/dpdk/node.c create mode 100644 src/vnet/devices/dpdk/qos_doc.md create mode 100644 src/vnet/devices/netmap/cli.c create mode 100644 src/vnet/devices/netmap/device.c create mode 100644 src/vnet/devices/netmap/net_netmap.h create mode 100644 src/vnet/devices/netmap/netmap.api create mode 100644 src/vnet/devices/netmap/netmap.c create mode 100644 src/vnet/devices/netmap/netmap.h create mode 100644 src/vnet/devices/netmap/netmap_api.c create mode 100644 src/vnet/devices/netmap/node.c create mode 100644 src/vnet/devices/nic/ixge.c create mode 100644 src/vnet/devices/nic/ixge.h create mode 100644 src/vnet/devices/nic/sfp.c create mode 100644 src/vnet/devices/nic/sfp.h create mode 100644 src/vnet/devices/ssvm/node.c create mode 100644 src/vnet/devices/ssvm/ssvm_eth.c create mode 100644 src/vnet/devices/ssvm/ssvm_eth.h create mode 100644 src/vnet/devices/virtio/dir.dox create mode 100644 src/vnet/devices/virtio/vhost-user.c create mode 100644 src/vnet/devices/virtio/vhost-user.h create mode 100644 src/vnet/devices/virtio/vhost_user.api create mode 100644 src/vnet/devices/virtio/vhost_user_api.c (limited to 'src/vnet/devices') diff --git a/src/vnet/devices/af_packet/af_packet.api b/src/vnet/devices/af_packet/af_packet.api new file mode 100644 index 00000000000..9fb2a2070f2 --- /dev/null +++ b/src/vnet/devices/af_packet/af_packet.api @@ -0,0 +1,71 @@ +/* + * Copyright (c) 2015-2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/** \brief Create host-interface + @param client_index - opaque cookie to identify the sender + @param context - sender context, to match reply w/ request + @param host_if_name - interface name + @param hw_addr - interface MAC + @param use_random_hw_addr - use random generated MAC +*/ +define af_packet_create +{ + u32 client_index; + u32 context; + + u8 host_if_name[64]; + u8 hw_addr[6]; + u8 use_random_hw_addr; +}; + +/** \brief Create host-interface response + @param context - sender context, to match reply w/ request + @param retval - return value for request +*/ +define af_packet_create_reply +{ + u32 context; + i32 retval; + u32 sw_if_index; +}; + +/** \brief Delete host-interface + @param client_index - opaque cookie to identify the sender + @param context - sender context, to match reply w/ request + @param host_if_name - interface name +*/ +define af_packet_delete +{ + u32 client_index; + u32 context; + + u8 host_if_name[64]; +}; + +/** \brief Delete host-interface response + @param context - sender context, to match reply w/ request + @param retval - return value for request +*/ +define af_packet_delete_reply +{ + u32 context; + i32 retval; +}; + +/* + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/af_packet/af_packet.c b/src/vnet/devices/af_packet/af_packet.c new file mode 100644 index 00000000000..91c3988b439 --- /dev/null +++ b/src/vnet/devices/af_packet/af_packet.c @@ -0,0 +1,366 @@ +/* + *------------------------------------------------------------------ + * af_packet.c - linux kernel packet interface + * + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include +#include + +#include +#include +#include +#include + +#include + +#define AF_PACKET_DEBUG_SOCKET 0 + +#define AF_PACKET_TX_FRAMES_PER_BLOCK 1024 +#define AF_PACKET_TX_FRAME_SIZE (2048 * 5) +#define AF_PACKET_TX_BLOCK_NR 1 +#define AF_PACKET_TX_FRAME_NR (AF_PACKET_TX_BLOCK_NR * \ + AF_PACKET_TX_FRAMES_PER_BLOCK) +#define AF_PACKET_TX_BLOCK_SIZE (AF_PACKET_TX_FRAME_SIZE * \ + AF_PACKET_TX_FRAMES_PER_BLOCK) + +#define AF_PACKET_RX_FRAMES_PER_BLOCK 1024 +#define AF_PACKET_RX_FRAME_SIZE (2048 * 5) +#define AF_PACKET_RX_BLOCK_NR 1 +#define AF_PACKET_RX_FRAME_NR (AF_PACKET_RX_BLOCK_NR * \ + AF_PACKET_RX_FRAMES_PER_BLOCK) +#define AF_PACKET_RX_BLOCK_SIZE (AF_PACKET_RX_FRAME_SIZE * \ + AF_PACKET_RX_FRAMES_PER_BLOCK) + +#if AF_PACKET_DEBUG_SOCKET == 1 +#define DBG_SOCK(args...) clib_warning(args); +#else +#define DBG_SOCK(args...) +#endif + +/*defined in net/if.h but clashes with dpdk headers */ +unsigned int if_nametoindex (const char *ifname); + +typedef struct tpacket_req tpacket_req_t; + +static u32 +af_packet_eth_flag_change (vnet_main_t * vnm, vnet_hw_interface_t * hi, + u32 flags) +{ + /* nothing for now */ + return 0; +} + +static clib_error_t * +af_packet_fd_read_ready (unix_file_t * uf) +{ + vlib_main_t *vm = vlib_get_main (); + af_packet_main_t *apm = &af_packet_main; + u32 idx = uf->private_data; + + apm->pending_input_bitmap = + clib_bitmap_set (apm->pending_input_bitmap, idx, 1); + + /* Schedule the rx node */ + vlib_node_set_interrupt_pending (vm, af_packet_input_node.index); + + return 0; +} + +static int +create_packet_v2_sock (u8 * name, tpacket_req_t * rx_req, + tpacket_req_t * tx_req, int *fd, u8 ** ring) +{ + int ret, err; + struct sockaddr_ll sll; + uint host_if_index; + int ver = TPACKET_V2; + socklen_t req_sz = sizeof (struct tpacket_req); + u32 ring_sz = rx_req->tp_block_size * rx_req->tp_block_nr + + tx_req->tp_block_size * tx_req->tp_block_nr; + + host_if_index = if_nametoindex ((const char *) name); + + if (!host_if_index) + { + DBG_SOCK ("Wrong host interface name"); + ret = VNET_API_ERROR_INVALID_INTERFACE; + goto error; + } + + if ((*fd = socket (AF_PACKET, SOCK_RAW, htons (ETH_P_ALL))) < 0) + { + DBG_SOCK ("Failed to create socket"); + ret = VNET_API_ERROR_SYSCALL_ERROR_1; + goto error; + } + + if ((err = + setsockopt (*fd, SOL_PACKET, PACKET_VERSION, &ver, sizeof (ver))) < 0) + { + DBG_SOCK ("Failed to set rx packet interface version"); + ret = VNET_API_ERROR_SYSCALL_ERROR_1; + goto error; + } + + int opt = 1; + if ((err = + setsockopt (*fd, SOL_PACKET, PACKET_LOSS, &opt, sizeof (opt))) < 0) + { + DBG_SOCK ("Failed to set packet tx ring error handling option"); + ret = VNET_API_ERROR_SYSCALL_ERROR_1; + goto error; + } + + if ((err = + setsockopt (*fd, SOL_PACKET, PACKET_RX_RING, rx_req, req_sz)) < 0) + { + DBG_SOCK ("Failed to set packet rx ring options"); + ret = VNET_API_ERROR_SYSCALL_ERROR_1; + goto error; + } + + if ((err = + setsockopt (*fd, SOL_PACKET, PACKET_TX_RING, tx_req, req_sz)) < 0) + { + DBG_SOCK ("Failed to set packet rx ring options"); + ret = VNET_API_ERROR_SYSCALL_ERROR_1; + goto error; + } + + *ring = + mmap (NULL, ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_LOCKED, *fd, + 0); + if (*ring == MAP_FAILED) + { + DBG_SOCK ("mmap failure"); + ret = VNET_API_ERROR_SYSCALL_ERROR_1; + goto error; + } + + memset (&sll, 0, sizeof (sll)); + sll.sll_family = PF_PACKET; + sll.sll_protocol = htons (ETH_P_ALL); + sll.sll_ifindex = host_if_index; + + if ((err = bind (*fd, (struct sockaddr *) &sll, sizeof (sll))) < 0) + { + DBG_SOCK ("Failed to bind rx packet socket (error %d)", err); + ret = VNET_API_ERROR_SYSCALL_ERROR_1; + goto error; + } + + return 0; +error: + if (*fd >= 0) + close (*fd); + *fd = -1; + return ret; +} + +int +af_packet_create_if (vlib_main_t * vm, u8 * host_if_name, u8 * hw_addr_set, + u32 * sw_if_index) +{ + af_packet_main_t *apm = &af_packet_main; + int ret, fd = -1; + struct tpacket_req *rx_req = 0; + struct tpacket_req *tx_req = 0; + u8 *ring = 0; + af_packet_if_t *apif = 0; + u8 hw_addr[6]; + clib_error_t *error; + vnet_sw_interface_t *sw; + vnet_main_t *vnm = vnet_get_main (); + uword *p; + uword if_index; + u8 *host_if_name_dup = vec_dup (host_if_name); + + p = mhash_get (&apm->if_index_by_host_if_name, host_if_name); + if (p) + { + return VNET_API_ERROR_SUBIF_ALREADY_EXISTS; + } + + vec_validate (rx_req, 0); + rx_req->tp_block_size = AF_PACKET_RX_BLOCK_SIZE; + rx_req->tp_frame_size = AF_PACKET_RX_FRAME_SIZE; + rx_req->tp_block_nr = AF_PACKET_RX_BLOCK_NR; + rx_req->tp_frame_nr = AF_PACKET_RX_FRAME_NR; + + vec_validate (tx_req, 0); + tx_req->tp_block_size = AF_PACKET_TX_BLOCK_SIZE; + tx_req->tp_frame_size = AF_PACKET_TX_FRAME_SIZE; + tx_req->tp_block_nr = AF_PACKET_TX_BLOCK_NR; + tx_req->tp_frame_nr = AF_PACKET_TX_FRAME_NR; + + ret = create_packet_v2_sock (host_if_name, rx_req, tx_req, &fd, &ring); + + if (ret != 0) + goto error; + + /* So far everything looks good, let's create interface */ + pool_get (apm->interfaces, apif); + if_index = apif - apm->interfaces; + + apif->fd = fd; + apif->rx_ring = ring; + apif->tx_ring = ring + rx_req->tp_block_size * rx_req->tp_block_nr; + apif->rx_req = rx_req; + apif->tx_req = tx_req; + apif->host_if_name = host_if_name_dup; + apif->per_interface_next_index = ~0; + apif->next_tx_frame = 0; + apif->next_rx_frame = 0; + + { + unix_file_t template = { 0 }; + template.read_function = af_packet_fd_read_ready; + template.file_descriptor = fd; + template.private_data = if_index; + template.flags = UNIX_FILE_EVENT_EDGE_TRIGGERED; + apif->unix_file_index = unix_file_add (&unix_main, &template); + } + + /*use configured or generate random MAC address */ + if (hw_addr_set) + clib_memcpy (hw_addr, hw_addr_set, 6); + else + { + f64 now = vlib_time_now (vm); + u32 rnd; + rnd = (u32) (now * 1e6); + rnd = random_u32 (&rnd); + + clib_memcpy (hw_addr + 2, &rnd, sizeof (rnd)); + hw_addr[0] = 2; + hw_addr[1] = 0xfe; + } + + error = ethernet_register_interface (vnm, af_packet_device_class.index, + if_index, hw_addr, &apif->hw_if_index, + af_packet_eth_flag_change); + + if (error) + { + memset (apif, 0, sizeof (*apif)); + pool_put (apm->interfaces, apif); + clib_error_report (error); + ret = VNET_API_ERROR_SYSCALL_ERROR_1; + goto error; + } + + sw = vnet_get_hw_sw_interface (vnm, apif->hw_if_index); + apif->sw_if_index = sw->sw_if_index; + + vnet_hw_interface_set_flags (vnm, apif->hw_if_index, + VNET_HW_INTERFACE_FLAG_LINK_UP); + + mhash_set_mem (&apm->if_index_by_host_if_name, host_if_name_dup, &if_index, + 0); + if (sw_if_index) + *sw_if_index = apif->sw_if_index; + return 0; + +error: + vec_free (host_if_name_dup); + vec_free (rx_req); + vec_free (tx_req); + return ret; +} + +int +af_packet_delete_if (vlib_main_t * vm, u8 * host_if_name) +{ + vnet_main_t *vnm = vnet_get_main (); + af_packet_main_t *apm = &af_packet_main; + af_packet_if_t *apif; + uword *p; + uword if_index; + u32 ring_sz; + + p = mhash_get (&apm->if_index_by_host_if_name, host_if_name); + if (p == NULL) + { + clib_warning ("Host interface %s does not exist", host_if_name); + return VNET_API_ERROR_SYSCALL_ERROR_1; + } + apif = pool_elt_at_index (apm->interfaces, p[0]); + if_index = apif - apm->interfaces; + + /* bring down the interface */ + vnet_hw_interface_set_flags (vnm, apif->hw_if_index, 0); + + /* clean up */ + if (apif->unix_file_index != ~0) + { + unix_file_del (&unix_main, unix_main.file_pool + apif->unix_file_index); + apif->unix_file_index = ~0; + } + else + close (apif->fd); + + ring_sz = apif->rx_req->tp_block_size * apif->rx_req->tp_block_nr + + apif->tx_req->tp_block_size * apif->tx_req->tp_block_nr; + if (munmap (apif->rx_ring, ring_sz)) + clib_warning ("Host interface %s could not free rx/tx ring", + host_if_name); + apif->rx_ring = NULL; + apif->tx_ring = NULL; + apif->fd = -1; + + vec_free (apif->rx_req); + apif->rx_req = NULL; + vec_free (apif->tx_req); + apif->tx_req = NULL; + + vec_free (apif->host_if_name); + apif->host_if_name = NULL; + + mhash_unset (&apm->if_index_by_host_if_name, host_if_name, &if_index); + + ethernet_delete_interface (vnm, apif->hw_if_index); + + pool_put (apm->interfaces, apif); + + return 0; +} + +static clib_error_t * +af_packet_init (vlib_main_t * vm) +{ + af_packet_main_t *apm = &af_packet_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); + + memset (apm, 0, sizeof (af_packet_main_t)); + + mhash_init_vec_string (&apm->if_index_by_host_if_name, sizeof (uword)); + + vec_validate_aligned (apm->rx_buffers, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + + return 0; +} + +VLIB_INIT_FUNCTION (af_packet_init); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/af_packet/af_packet.h b/src/vnet/devices/af_packet/af_packet.h new file mode 100644 index 00000000000..19e2523d6c9 --- /dev/null +++ b/src/vnet/devices/af_packet/af_packet.h @@ -0,0 +1,69 @@ +/* + *------------------------------------------------------------------ + * af_packet.h - linux kernel packet interface header file + * + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + u8 *host_if_name; + int fd; + struct tpacket_req *rx_req; + struct tpacket_req *tx_req; + u8 *rx_ring; + u8 *tx_ring; + u32 hw_if_index; + u32 sw_if_index; + u32 unix_file_index; + + u32 next_rx_frame; + u32 next_tx_frame; + + u32 per_interface_next_index; + u8 is_admin_up; +} af_packet_if_t; + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + af_packet_if_t *interfaces; + + /* bitmap of pending rx interfaces */ + uword *pending_input_bitmap; + + /* rx buffer cache */ + u32 **rx_buffers; + + /* hash of host interface names */ + mhash_t if_index_by_host_if_name; +} af_packet_main_t; + +af_packet_main_t af_packet_main; +extern vnet_device_class_t af_packet_device_class; +extern vlib_node_registration_t af_packet_input_node; + +int af_packet_create_if (vlib_main_t * vm, u8 * host_if_name, + u8 * hw_addr_set, u32 * sw_if_index); +int af_packet_delete_if (vlib_main_t * vm, u8 * host_if_name); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/af_packet/af_packet_api.c b/src/vnet/devices/af_packet/af_packet_api.c new file mode 100644 index 00000000000..414c838cdf7 --- /dev/null +++ b/src/vnet/devices/af_packet/af_packet_api.c @@ -0,0 +1,143 @@ +/* + *------------------------------------------------------------------ + * af_packet_api.c - af-packet api + * + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include +#include + +#include +#include +#include + +#include + +#define vl_typedefs /* define message structures */ +#include +#undef vl_typedefs + +#define vl_endianfun /* define message structures */ +#include +#undef vl_endianfun + +/* instantiate all the print functions we know about */ +#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__) +#define vl_printfun +#include +#undef vl_printfun + +#include + +#define foreach_vpe_api_msg \ +_(AF_PACKET_CREATE, af_packet_create) \ +_(AF_PACKET_DELETE, af_packet_delete) + +static void +vl_api_af_packet_create_t_handler (vl_api_af_packet_create_t * mp) +{ + vlib_main_t *vm = vlib_get_main (); + vl_api_af_packet_create_reply_t *rmp; + int rv = 0; + u8 *host_if_name = NULL; + u32 sw_if_index; + + host_if_name = format (0, "%s", mp->host_if_name); + vec_add1 (host_if_name, 0); + + rv = af_packet_create_if (vm, host_if_name, + mp->use_random_hw_addr ? 0 : mp->hw_addr, + &sw_if_index); + + vec_free (host_if_name); + + /* *INDENT-OFF* */ + REPLY_MACRO2(VL_API_AF_PACKET_CREATE_REPLY, + ({ + rmp->sw_if_index = clib_host_to_net_u32(sw_if_index); + })); + /* *INDENT-ON* */ +} + +static void +vl_api_af_packet_delete_t_handler (vl_api_af_packet_delete_t * mp) +{ + vlib_main_t *vm = vlib_get_main (); + vl_api_af_packet_delete_reply_t *rmp; + int rv = 0; + u8 *host_if_name = NULL; + + host_if_name = format (0, "%s", mp->host_if_name); + vec_add1 (host_if_name, 0); + + rv = af_packet_delete_if (vm, host_if_name); + + vec_free (host_if_name); + + REPLY_MACRO (VL_API_AF_PACKET_DELETE_REPLY); +} + +/* + * af_packet_api_hookup + * Add vpe's API message handlers to the table. + * vlib has alread mapped shared memory and + * added the client registration handlers. + * See .../vlib-api/vlibmemory/memclnt_vlib.c:memclnt_process() + */ +#define vl_msg_name_crc_list +#include +#undef vl_msg_name_crc_list + +static void +setup_message_id_table (api_main_t * am) +{ +#define _(id,n,crc) vl_msg_api_add_msg_name_crc (am, #n "_" #crc, id); + foreach_vl_msg_name_crc_af_packet; +#undef _ +} + +static clib_error_t * +af_packet_api_hookup (vlib_main_t * vm) +{ + api_main_t *am = &api_main; + +#define _(N,n) \ + vl_msg_api_set_handlers(VL_API_##N, #n, \ + vl_api_##n##_t_handler, \ + vl_noop_handler, \ + vl_api_##n##_t_endian, \ + vl_api_##n##_t_print, \ + sizeof(vl_api_##n##_t), 1); + foreach_vpe_api_msg; +#undef _ + + /* + * Set up the (msg_name, crc, message-id) table + */ + setup_message_id_table (am); + + return 0; +} + +VLIB_API_INIT_FUNCTION (af_packet_api_hookup); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/af_packet/cli.c b/src/vnet/devices/af_packet/cli.c new file mode 100644 index 00000000000..2cbd415289e --- /dev/null +++ b/src/vnet/devices/af_packet/cli.c @@ -0,0 +1,144 @@ +/* + *------------------------------------------------------------------ + * af_packet.c - linux kernel packet interface + * + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include /* for open */ +#include +#include +#include +#include +#include /* for iovec */ +#include + +#include +#include +#include +#include + +#include + +static clib_error_t * +af_packet_create_command_fn (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + u8 *host_if_name = NULL; + u8 hwaddr[6]; + u8 *hw_addr_ptr = 0; + u32 sw_if_index; + int r; + + /* Get a line of input. */ + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "name %s", &host_if_name)) + ; + else + if (unformat + (line_input, "hw-addr %U", unformat_ethernet_address, hwaddr)) + hw_addr_ptr = hwaddr; + else + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + unformat_free (line_input); + + if (host_if_name == NULL) + return clib_error_return (0, "missing host interface name"); + + r = af_packet_create_if (vm, host_if_name, hw_addr_ptr, &sw_if_index); + vec_free (host_if_name); + + if (r == VNET_API_ERROR_SYSCALL_ERROR_1) + return clib_error_return (0, "%s (errno %d)", strerror (errno), errno); + + if (r == VNET_API_ERROR_INVALID_INTERFACE) + return clib_error_return (0, "Invalid interface name"); + + if (r == VNET_API_ERROR_SUBIF_ALREADY_EXISTS) + return clib_error_return (0, "Interface elready exists"); + + vlib_cli_output (vm, "%U\n", format_vnet_sw_if_index_name, vnet_get_main (), + sw_if_index); + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (af_packet_create_command, static) = { + .path = "create host-interface", + .short_help = "create host-interface name [hw-addr ]", + .function = af_packet_create_command_fn, +}; +/* *INDENT-ON* */ + +static clib_error_t * +af_packet_delete_command_fn (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + u8 *host_if_name = NULL; + + /* Get a line of input. */ + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "name %s", &host_if_name)) + ; + else + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + unformat_free (line_input); + + if (host_if_name == NULL) + return clib_error_return (0, "missing host interface name"); + + af_packet_delete_if (vm, host_if_name); + vec_free (host_if_name); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (af_packet_delete_command, static) = { + .path = "delete host-interface", + .short_help = "delete host-interface name ", + .function = af_packet_delete_command_fn, +}; +/* *INDENT-ON* */ + +clib_error_t * +af_packet_cli_init (vlib_main_t * vm) +{ + return 0; +} + +VLIB_INIT_FUNCTION (af_packet_cli_init); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/af_packet/device.c b/src/vnet/devices/af_packet/device.c new file mode 100644 index 00000000000..1fb4000f6e6 --- /dev/null +++ b/src/vnet/devices/af_packet/device.c @@ -0,0 +1,250 @@ +/* + *------------------------------------------------------------------ + * af_packet.c - linux kernel packet interface + * + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include + +#include +#include +#include +#include + +#include + +#define foreach_af_packet_tx_func_error \ +_(FRAME_NOT_READY, "tx frame not ready") \ +_(TXRING_EAGAIN, "tx sendto temporary failure") \ +_(TXRING_FATAL, "tx sendto fatal failure") \ +_(TXRING_OVERRUN, "tx ring overrun") + +typedef enum +{ +#define _(f,s) AF_PACKET_TX_ERROR_##f, + foreach_af_packet_tx_func_error +#undef _ + AF_PACKET_TX_N_ERROR, +} af_packet_tx_func_error_t; + +static char *af_packet_tx_func_error_strings[] = { +#define _(n,s) s, + foreach_af_packet_tx_func_error +#undef _ +}; + + +static u8 * +format_af_packet_device_name (u8 * s, va_list * args) +{ + u32 i = va_arg (*args, u32); + af_packet_main_t *apm = &af_packet_main; + af_packet_if_t *apif = pool_elt_at_index (apm->interfaces, i); + + s = format (s, "host-%s", apif->host_if_name); + return s; +} + +static u8 * +format_af_packet_device (u8 * s, va_list * args) +{ + s = format (s, "Linux PACKET socket interface"); + return s; +} + +static u8 * +format_af_packet_tx_trace (u8 * s, va_list * args) +{ + s = format (s, "Unimplemented..."); + return s; +} + +static uword +af_packet_interface_tx (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * frame) +{ + af_packet_main_t *apm = &af_packet_main; + u32 *buffers = vlib_frame_args (frame); + u32 n_left = frame->n_vectors; + u32 n_sent = 0; + vnet_interface_output_runtime_t *rd = (void *) node->runtime_data; + af_packet_if_t *apif = + pool_elt_at_index (apm->interfaces, rd->dev_instance); + int block = 0; + u32 block_size = apif->tx_req->tp_block_size; + u32 frame_size = apif->tx_req->tp_frame_size; + u32 frame_num = apif->tx_req->tp_frame_nr; + u8 *block_start = apif->tx_ring + block * block_size; + u32 tx_frame = apif->next_tx_frame; + struct tpacket2_hdr *tph; + u32 frame_not_ready = 0; + + while (n_left > 0) + { + u32 len; + u32 offset = 0; + vlib_buffer_t *b0; + n_left--; + u32 bi = buffers[0]; + buffers++; + + tph = (struct tpacket2_hdr *) (block_start + tx_frame * frame_size); + + if (PREDICT_FALSE + (tph->tp_status & (TP_STATUS_SEND_REQUEST | TP_STATUS_SENDING))) + { + frame_not_ready++; + goto next; + } + + do + { + b0 = vlib_get_buffer (vm, bi); + len = b0->current_length; + clib_memcpy ((u8 *) tph + + TPACKET_ALIGN (sizeof (struct tpacket2_hdr)) + offset, + vlib_buffer_get_current (b0), len); + offset += len; + } + while ((bi = b0->next_buffer)); + + tph->tp_len = tph->tp_snaplen = offset; + tph->tp_status = TP_STATUS_SEND_REQUEST; + n_sent++; + next: + /* check if we've exhausted the ring */ + if (PREDICT_FALSE (frame_not_ready + n_sent == frame_num)) + break; + + tx_frame = (tx_frame + 1) % frame_num; + } + + CLIB_MEMORY_BARRIER (); + + if (PREDICT_TRUE (n_sent)) + { + apif->next_tx_frame = tx_frame; + + if (PREDICT_FALSE (sendto (apif->fd, NULL, 0, + MSG_DONTWAIT, NULL, 0) == -1)) + { + /* Uh-oh, drop & move on, but count whether it was fatal or not. + * Note that we have no reliable way to properly determine the + * disposition of the packets we just enqueued for delivery. + */ + vlib_error_count (vm, node->node_index, + unix_error_is_fatal (errno) ? + AF_PACKET_TX_ERROR_TXRING_FATAL : + AF_PACKET_TX_ERROR_TXRING_EAGAIN, n_sent); + } + } + + if (PREDICT_FALSE (frame_not_ready)) + vlib_error_count (vm, node->node_index, + AF_PACKET_TX_ERROR_FRAME_NOT_READY, frame_not_ready); + + if (PREDICT_FALSE (frame_not_ready + n_sent == frame_num)) + vlib_error_count (vm, node->node_index, AF_PACKET_TX_ERROR_TXRING_OVERRUN, + n_left); + + vlib_buffer_free (vm, vlib_frame_args (frame), frame->n_vectors); + return frame->n_vectors; +} + +static void +af_packet_set_interface_next_node (vnet_main_t * vnm, u32 hw_if_index, + u32 node_index) +{ + af_packet_main_t *apm = &af_packet_main; + vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index); + af_packet_if_t *apif = + pool_elt_at_index (apm->interfaces, hw->dev_instance); + + /* Shut off redirection */ + if (node_index == ~0) + { + apif->per_interface_next_index = node_index; + return; + } + + apif->per_interface_next_index = + vlib_node_add_next (vlib_get_main (), af_packet_input_node.index, + node_index); +} + +static void +af_packet_clear_hw_interface_counters (u32 instance) +{ + /* Nothing for now */ +} + +static clib_error_t * +af_packet_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, + u32 flags) +{ + af_packet_main_t *apm = &af_packet_main; + vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index); + af_packet_if_t *apif = + pool_elt_at_index (apm->interfaces, hw->dev_instance); + u32 hw_flags; + + apif->is_admin_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0; + + if (apif->is_admin_up) + hw_flags = VNET_HW_INTERFACE_FLAG_LINK_UP; + else + hw_flags = 0; + + vnet_hw_interface_set_flags (vnm, hw_if_index, hw_flags); + + return 0; +} + +static clib_error_t * +af_packet_subif_add_del_function (vnet_main_t * vnm, + u32 hw_if_index, + struct vnet_sw_interface_t *st, int is_add) +{ + /* Nothing for now */ + return 0; +} + +/* *INDENT-OFF* */ +VNET_DEVICE_CLASS (af_packet_device_class) = { + .name = "af-packet", + .tx_function = af_packet_interface_tx, + .format_device_name = format_af_packet_device_name, + .format_device = format_af_packet_device, + .format_tx_trace = format_af_packet_tx_trace, + .tx_function_n_errors = AF_PACKET_TX_N_ERROR, + .tx_function_error_strings = af_packet_tx_func_error_strings, + .rx_redirect_to_node = af_packet_set_interface_next_node, + .clear_counters = af_packet_clear_hw_interface_counters, + .admin_up_down_function = af_packet_interface_admin_up_down, + .subif_add_del_function = af_packet_subif_add_del_function, +}; + +VLIB_DEVICE_TX_FUNCTION_MULTIARCH (af_packet_device_class, + af_packet_interface_tx) +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/af_packet/node.c b/src/vnet/devices/af_packet/node.c new file mode 100644 index 00000000000..72004320c67 --- /dev/null +++ b/src/vnet/devices/af_packet/node.c @@ -0,0 +1,288 @@ +/* + *------------------------------------------------------------------ + * af_packet.c - linux kernel packet interface + * + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include + +#include +#include +#include +#include +#include +#include + +#include + +#define foreach_af_packet_input_error + +typedef enum +{ +#define _(f,s) AF_PACKET_INPUT_ERROR_##f, + foreach_af_packet_input_error +#undef _ + AF_PACKET_INPUT_N_ERROR, +} af_packet_input_error_t; + +static char *af_packet_input_error_strings[] = { +#define _(n,s) s, + foreach_af_packet_input_error +#undef _ +}; + +typedef struct +{ + u32 next_index; + u32 hw_if_index; + int block; + struct tpacket2_hdr tph; +} af_packet_input_trace_t; + +static u8 * +format_af_packet_input_trace (u8 * s, va_list * args) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *); + af_packet_input_trace_t *t = va_arg (*args, af_packet_input_trace_t *); + uword indent = format_get_indent (s); + + s = format (s, "af_packet: hw_if_index %d next-index %d", + t->hw_if_index, t->next_index); + + s = + format (s, + "\n%Utpacket2_hdr:\n%Ustatus 0x%x len %u snaplen %u mac %u net %u" + "\n%Usec 0x%x nsec 0x%x vlan %U" +#ifdef TP_STATUS_VLAN_TPID_VALID + " vlan_tpid %u" +#endif + , + format_white_space, indent + 2, + format_white_space, indent + 4, + t->tph.tp_status, + t->tph.tp_len, + t->tph.tp_snaplen, + t->tph.tp_mac, + t->tph.tp_net, + format_white_space, indent + 4, + t->tph.tp_sec, + t->tph.tp_nsec, format_ethernet_vlan_tci, t->tph.tp_vlan_tci +#ifdef TP_STATUS_VLAN_TPID_VALID + , t->tph.tp_vlan_tpid +#endif + ); + return s; +} + +always_inline void +buffer_add_to_chain (vlib_main_t * vm, u32 bi, u32 first_bi, u32 prev_bi) +{ + vlib_buffer_t *b = vlib_get_buffer (vm, bi); + vlib_buffer_t *first_b = vlib_get_buffer (vm, first_bi); + vlib_buffer_t *prev_b = vlib_get_buffer (vm, prev_bi); + + /* update first buffer */ + first_b->total_length_not_including_first_buffer += b->current_length; + + /* update previous buffer */ + prev_b->next_buffer = bi; + prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT; + + /* update current buffer */ + b->next_buffer = 0; +} + +always_inline uword +af_packet_device_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node, + vlib_frame_t * frame, u32 device_idx) +{ + af_packet_main_t *apm = &af_packet_main; + af_packet_if_t *apif = pool_elt_at_index (apm->interfaces, device_idx); + struct tpacket2_hdr *tph; + u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; + u32 block = 0; + u32 rx_frame; + u32 n_free_bufs; + u32 n_rx_packets = 0; + u32 n_rx_bytes = 0; + u32 *to_next = 0; + u32 block_size = apif->rx_req->tp_block_size; + u32 frame_size = apif->rx_req->tp_frame_size; + u32 frame_num = apif->rx_req->tp_frame_nr; + u8 *block_start = apif->rx_ring + block * block_size; + uword n_trace = vlib_get_trace_count (vm, node); + u32 n_buffer_bytes = vlib_buffer_free_list_buffer_size (vm, + VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); + u32 min_bufs = apif->rx_req->tp_frame_size / n_buffer_bytes; + int cpu_index = node->cpu_index; + + if (apif->per_interface_next_index != ~0) + next_index = apif->per_interface_next_index; + + n_free_bufs = vec_len (apm->rx_buffers[cpu_index]); + if (PREDICT_FALSE (n_free_bufs < VLIB_FRAME_SIZE)) + { + vec_validate (apm->rx_buffers[cpu_index], + VLIB_FRAME_SIZE + n_free_bufs - 1); + n_free_bufs += + vlib_buffer_alloc (vm, &apm->rx_buffers[cpu_index][n_free_bufs], + VLIB_FRAME_SIZE); + _vec_len (apm->rx_buffers[cpu_index]) = n_free_bufs; + } + + rx_frame = apif->next_rx_frame; + tph = (struct tpacket2_hdr *) (block_start + rx_frame * frame_size); + while ((tph->tp_status & TP_STATUS_USER) && (n_free_bufs > min_bufs)) + { + vlib_buffer_t *b0 = 0, *first_b0 = 0; + u32 next0 = next_index; + + u32 n_left_to_next; + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + while ((tph->tp_status & TP_STATUS_USER) && (n_free_bufs > min_bufs) && + n_left_to_next) + { + u32 data_len = tph->tp_snaplen; + u32 offset = 0; + u32 bi0 = 0, first_bi0 = 0, prev_bi0; + + while (data_len) + { + /* grab free buffer */ + u32 last_empty_buffer = + vec_len (apm->rx_buffers[cpu_index]) - 1; + prev_bi0 = bi0; + bi0 = apm->rx_buffers[cpu_index][last_empty_buffer]; + b0 = vlib_get_buffer (vm, bi0); + _vec_len (apm->rx_buffers[cpu_index]) = last_empty_buffer; + n_free_bufs--; + + /* copy data */ + u32 bytes_to_copy = + data_len > n_buffer_bytes ? n_buffer_bytes : data_len; + b0->current_data = 0; + clib_memcpy (vlib_buffer_get_current (b0), + (u8 *) tph + tph->tp_mac + offset, bytes_to_copy); + + /* fill buffer header */ + b0->current_length = bytes_to_copy; + + if (offset == 0) + { + b0->total_length_not_including_first_buffer = 0; + b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID; + vnet_buffer (b0)->sw_if_index[VLIB_RX] = apif->sw_if_index; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + first_bi0 = bi0; + first_b0 = vlib_get_buffer (vm, first_bi0); + } + else + buffer_add_to_chain (vm, bi0, first_bi0, prev_bi0); + + offset += bytes_to_copy; + data_len -= bytes_to_copy; + } + n_rx_packets++; + n_rx_bytes += tph->tp_snaplen; + to_next[0] = first_bi0; + to_next += 1; + n_left_to_next--; + + /* trace */ + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (first_b0); + if (PREDICT_FALSE (n_trace > 0)) + { + af_packet_input_trace_t *tr; + vlib_trace_buffer (vm, node, next0, first_b0, /* follow_chain */ + 0); + vlib_set_trace_count (vm, node, --n_trace); + tr = vlib_add_trace (vm, node, first_b0, sizeof (*tr)); + tr->next_index = next0; + tr->hw_if_index = apif->hw_if_index; + clib_memcpy (&tr->tph, tph, sizeof (struct tpacket2_hdr)); + } + + /* redirect if feature path enabled */ + vnet_feature_start_device_input_x1 (apif->sw_if_index, &next0, b0, + 0); + + /* enque and take next packet */ + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next, + n_left_to_next, first_bi0, next0); + + /* next packet */ + tph->tp_status = TP_STATUS_KERNEL; + rx_frame = (rx_frame + 1) % frame_num; + tph = (struct tpacket2_hdr *) (block_start + rx_frame * frame_size); + } + + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + + apif->next_rx_frame = rx_frame; + + vlib_increment_combined_counter + (vnet_get_main ()->interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, + os_get_cpu_number (), apif->hw_if_index, n_rx_packets, n_rx_bytes); + + return n_rx_packets; +} + +static uword +af_packet_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node, + vlib_frame_t * frame) +{ + int i; + u32 n_rx_packets = 0; + + af_packet_main_t *apm = &af_packet_main; + + /* *INDENT-OFF* */ + clib_bitmap_foreach (i, apm->pending_input_bitmap, + ({ + clib_bitmap_set (apm->pending_input_bitmap, i, 0); + n_rx_packets += af_packet_device_input_fn(vm, node, frame, i); + })); + /* *INDENT-ON* */ + + return n_rx_packets; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (af_packet_input_node) = { + .function = af_packet_input_fn, + .name = "af-packet-input", + .sibling_of = "device-input", + .format_trace = format_af_packet_input_trace, + .type = VLIB_NODE_TYPE_INPUT, + .state = VLIB_NODE_STATE_INTERRUPT, + .n_errors = AF_PACKET_INPUT_N_ERROR, + .error_strings = af_packet_input_error_strings, +}; + +VLIB_NODE_FUNCTION_MULTIARCH (af_packet_input_node, af_packet_input_fn) +/* *INDENT-ON* */ + + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/devices.c b/src/vnet/devices/devices.c new file mode 100644 index 00000000000..cd4386ebdca --- /dev/null +++ b/src/vnet/devices/devices.c @@ -0,0 +1,91 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include +#include +#include +#include +#include + +static uword +device_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node, + vlib_frame_t * frame) +{ + return 0; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (device_input_node) = { + .function = device_input_fn, + .name = "device-input", + .type = VLIB_NODE_TYPE_INPUT, + .state = VLIB_NODE_STATE_DISABLED, + .n_next_nodes = VNET_DEVICE_INPUT_N_NEXT_NODES, + .next_nodes = VNET_DEVICE_INPUT_NEXT_NODES, +}; + +/* Table defines how much we need to advance current data pointer + in the buffer if we shortcut to l3 nodes */ + +const u32 __attribute__((aligned (CLIB_CACHE_LINE_BYTES))) +device_input_next_node_advance[((VNET_DEVICE_INPUT_N_NEXT_NODES / + CLIB_CACHE_LINE_BYTES) +1) * CLIB_CACHE_LINE_BYTES] = +{ + [VNET_DEVICE_INPUT_NEXT_IP4_INPUT] = sizeof (ethernet_header_t), + [VNET_DEVICE_INPUT_NEXT_IP4_NCS_INPUT] = sizeof (ethernet_header_t), + [VNET_DEVICE_INPUT_NEXT_IP6_INPUT] = sizeof (ethernet_header_t), + [VNET_DEVICE_INPUT_NEXT_MPLS_INPUT] = sizeof (ethernet_header_t), +}; + +VNET_FEATURE_ARC_INIT (device_input, static) = +{ + .arc_name = "device-input", + .start_nodes = VNET_FEATURES ("device-input"), + .end_node = "ethernet-input", + .arc_index_ptr = &feature_main.device_input_feature_arc_index, +}; + +VNET_FEATURE_INIT (l2_patch, static) = { + .arc_name = "device-input", + .node_name = "l2-patch", + .runs_before = VNET_FEATURES ("ethernet-input"), +}; + +VNET_FEATURE_INIT (worker_handoff, static) = { + .arc_name = "device-input", + .node_name = "worker-handoff", + .runs_before = VNET_FEATURES ("ethernet-input"), +}; + +VNET_FEATURE_INIT (span_input, static) = { + .arc_name = "device-input", + .node_name = "span-input", + .runs_before = VNET_FEATURES ("ethernet-input"), +}; + +VNET_FEATURE_INIT (ethernet_input, static) = { + .arc_name = "device-input", + .node_name = "ethernet-input", + .runs_before = 0, /* not before any other features */ +}; +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/devices.h b/src/vnet/devices/devices.h new file mode 100644 index 00000000000..c46dab904c3 --- /dev/null +++ b/src/vnet/devices/devices.h @@ -0,0 +1,53 @@ +/* + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef included_vnet_vnet_device_h +#define included_vnet_vnet_device_h + +#include +#include + +typedef enum +{ + VNET_DEVICE_INPUT_NEXT_IP4_NCS_INPUT, + VNET_DEVICE_INPUT_NEXT_IP4_INPUT, + VNET_DEVICE_INPUT_NEXT_IP6_INPUT, + VNET_DEVICE_INPUT_NEXT_MPLS_INPUT, + VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT, + VNET_DEVICE_INPUT_NEXT_DROP, + VNET_DEVICE_INPUT_N_NEXT_NODES, +} vnet_device_input_next_t; + +#define VNET_DEVICE_INPUT_NEXT_NODES { \ + [VNET_DEVICE_INPUT_NEXT_DROP] = "error-drop", \ + [VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT] = "ethernet-input", \ + [VNET_DEVICE_INPUT_NEXT_IP4_NCS_INPUT] = "ip4-input-no-checksum", \ + [VNET_DEVICE_INPUT_NEXT_IP4_INPUT] = "ip4-input", \ + [VNET_DEVICE_INPUT_NEXT_IP6_INPUT] = "ip6-input", \ + [VNET_DEVICE_INPUT_NEXT_MPLS_INPUT] = "mpls-input", \ +} + +extern vlib_node_registration_t device_input_node; +extern const u32 device_input_next_node_advance[]; + +#endif /* included_vnet_vnet_device_h */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/cli.c b/src/vnet/devices/dpdk/cli.c new file mode 100644 index 00000000000..538a00fd975 --- /dev/null +++ b/src/vnet/devices/dpdk/cli.c @@ -0,0 +1,1296 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include +#include +#include +#include +#include + +#include +#include +#include +#include + +#include "dpdk_priv.h" + +static clib_error_t * +pcap_trace_command_fn (vlib_main_t * vm, + unformat_input_t * input, vlib_cli_command_t * cmd) +{ + dpdk_main_t *dm = &dpdk_main; + u8 *filename; + u32 max; + int matched = 0; + clib_error_t *error = 0; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "on")) + { + if (dm->tx_pcap_enable == 0) + { + if (dm->pcap_filename == 0) + dm->pcap_filename = format (0, "/tmp/vpe.pcap%c", 0); + + memset (&dm->pcap_main, 0, sizeof (dm->pcap_main)); + dm->pcap_main.file_name = (char *) dm->pcap_filename; + dm->pcap_main.n_packets_to_capture = 100; + if (dm->pcap_pkts_to_capture) + dm->pcap_main.n_packets_to_capture = dm->pcap_pkts_to_capture; + + dm->pcap_main.packet_type = PCAP_PACKET_TYPE_ethernet; + dm->tx_pcap_enable = 1; + matched = 1; + vlib_cli_output (vm, "pcap tx capture on..."); + } + else + { + vlib_cli_output (vm, "pcap tx capture already on..."); + } + matched = 1; + } + else if (unformat (input, "off")) + { + if (dm->tx_pcap_enable) + { + vlib_cli_output (vm, "captured %d pkts...", + dm->pcap_main.n_packets_captured + 1); + if (dm->pcap_main.n_packets_captured) + { + dm->pcap_main.n_packets_to_capture = + dm->pcap_main.n_packets_captured; + error = pcap_write (&dm->pcap_main); + if (error) + clib_error_report (error); + else + vlib_cli_output (vm, "saved to %s...", dm->pcap_filename); + } + } + else + { + vlib_cli_output (vm, "pcap tx capture already off..."); + } + + dm->tx_pcap_enable = 0; + matched = 1; + } + else if (unformat (input, "max %d", &max)) + { + dm->pcap_pkts_to_capture = max; + matched = 1; + } + + else if (unformat (input, "intfc %U", + unformat_vnet_sw_interface, dm->vnet_main, + &dm->pcap_sw_if_index)) + matched = 1; + else if (unformat (input, "intfc any")) + { + dm->pcap_sw_if_index = 0; + matched = 1; + } + else if (unformat (input, "file %s", &filename)) + { + u8 *chroot_filename; + /* Brain-police user path input */ + if (strstr ((char *) filename, "..") + || index ((char *) filename, '/')) + { + vlib_cli_output (vm, "illegal characters in filename '%s'", + filename); + continue; + } + + chroot_filename = format (0, "/tmp/%s%c", filename, 0); + vec_free (filename); + + if (dm->pcap_filename) + vec_free (dm->pcap_filename); + vec_add1 (filename, 0); + dm->pcap_filename = chroot_filename; + matched = 1; + } + else if (unformat (input, "status")) + { + if (dm->tx_pcap_enable == 0) + { + vlib_cli_output (vm, "pcap tx capture is off..."); + continue; + } + + vlib_cli_output (vm, "pcap tx capture: %d of %d pkts...", + dm->pcap_main.n_packets_captured, + dm->pcap_main.n_packets_to_capture); + matched = 1; + } + + else + break; + } + + if (matched == 0) + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (pcap_trace_command, static) = { + .path = "pcap tx trace", + .short_help = + "pcap tx trace on off max intfc file status", + .function = pcap_trace_command_fn, +}; +/* *INDENT-ON* */ + + +static clib_error_t * +show_dpdk_buffer (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + struct rte_mempool *rmp; + int i; + + for (i = 0; i < vec_len (vm->buffer_main->pktmbuf_pools); i++) + { + rmp = vm->buffer_main->pktmbuf_pools[i]; + if (rmp) + { + unsigned count = rte_mempool_avail_count (rmp); + unsigned free_count = rte_mempool_in_use_count (rmp); + + vlib_cli_output (vm, + "name=\"%s\" available = %7d allocated = %7d total = %7d\n", + rmp->name, (u32) count, (u32) free_count, + (u32) (count + free_count)); + } + else + { + vlib_cli_output (vm, "rte_mempool is NULL (!)\n"); + } + } + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_show_dpdk_bufferr,static) = { + .path = "show dpdk buffer", + .short_help = "show dpdk buffer state", + .function = show_dpdk_buffer, + .is_mp_safe = 1, +}; +/* *INDENT-ON* */ + +static clib_error_t * +test_dpdk_buffer (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + static u32 *allocated_buffers; + u32 n_alloc = 0; + u32 n_free = 0; + u32 first, actual_alloc; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "allocate %d", &n_alloc)) + ; + else if (unformat (input, "free %d", &n_free)) + ; + else + break; + } + + if (n_free) + { + if (vec_len (allocated_buffers) < n_free) + return clib_error_return (0, "Can't free %d, only %d allocated", + n_free, vec_len (allocated_buffers)); + + first = vec_len (allocated_buffers) - n_free; + vlib_buffer_free (vm, allocated_buffers + first, n_free); + _vec_len (allocated_buffers) = first; + } + if (n_alloc) + { + first = vec_len (allocated_buffers); + vec_validate (allocated_buffers, + vec_len (allocated_buffers) + n_alloc - 1); + + actual_alloc = vlib_buffer_alloc (vm, allocated_buffers + first, + n_alloc); + _vec_len (allocated_buffers) = first + actual_alloc; + + if (actual_alloc < n_alloc) + vlib_cli_output (vm, "WARNING: only allocated %d buffers", + actual_alloc); + } + + vlib_cli_output (vm, "Currently %d buffers allocated", + vec_len (allocated_buffers)); + + if (allocated_buffers && vec_len (allocated_buffers) == 0) + vec_free (allocated_buffers); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_test_dpdk_buffer,static) = { + .path = "test dpdk buffer", + .short_help = "test dpdk buffer [allocate ][free ]", + .function = test_dpdk_buffer, + .is_mp_safe = 1, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_desc (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 nb_rx_desc = (u32) ~ 0; + u32 nb_tx_desc = (u32) ~ 0; + clib_error_t *rv; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "tx %d", &nb_tx_desc)) + ; + else if (unformat (line_input, "rx %d", &nb_rx_desc)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + if ((xd->flags & DPDK_DEVICE_FLAG_PMD) == 0) + return clib_error_return (0, "number of descriptors can be set only for " + "physical devices"); + + if ((nb_rx_desc == (u32) ~ 0 || nb_rx_desc == xd->nb_rx_desc) && + (nb_tx_desc == (u32) ~ 0 || nb_tx_desc == xd->nb_tx_desc)) + return clib_error_return (0, "nothing changed"); + + if (nb_rx_desc != (u32) ~ 0) + xd->nb_rx_desc = nb_rx_desc; + + if (nb_tx_desc != (u32) ~ 0) + xd->nb_tx_desc = nb_tx_desc; + + rv = dpdk_port_setup (dm, xd); + + return rv; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_desc,static) = { + .path = "set dpdk interface descriptors", + .short_help = "set dpdk interface descriptors [rx ] [tx ]", + .function = set_dpdk_if_desc, +}; +/* *INDENT-ON* */ + +static clib_error_t * +show_dpdk_if_placement (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + dpdk_device_and_queue_t *dq; + int cpu; + + if (tm->n_vlib_mains == 1) + vlib_cli_output (vm, "All interfaces are handled by main thread"); + + for (cpu = 0; cpu < vec_len (dm->devices_by_cpu); cpu++) + { + if (vec_len (dm->devices_by_cpu[cpu])) + vlib_cli_output (vm, "Thread %u (%s at lcore %u):", cpu, + vlib_worker_threads[cpu].name, + vlib_worker_threads[cpu].lcore_id); + + /* *INDENT-OFF* */ + vec_foreach(dq, dm->devices_by_cpu[cpu]) + { + u32 hw_if_index = dm->devices[dq->device].vlib_hw_if_index; + vnet_hw_interface_t * hi = vnet_get_hw_interface(dm->vnet_main, hw_if_index); + vlib_cli_output(vm, " %v queue %u", hi->name, dq->queue_id); + } + /* *INDENT-ON* */ + } + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_show_dpdk_if_placement,static) = { + .path = "show dpdk interface placement", + .short_help = "show dpdk interface placement", + .function = show_dpdk_if_placement, +}; +/* *INDENT-ON* */ + +static int +dpdk_device_queue_sort (void *a1, void *a2) +{ + dpdk_device_and_queue_t *dq1 = a1; + dpdk_device_and_queue_t *dq2 = a2; + + if (dq1->device > dq2->device) + return 1; + else if (dq1->device < dq2->device) + return -1; + else if (dq1->queue_id > dq2->queue_id) + return 1; + else if (dq1->queue_id < dq2->queue_id) + return -1; + else + return 0; +} + +static clib_error_t * +set_dpdk_if_placement (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + dpdk_device_and_queue_t *dq; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 queue = (u32) 0; + u32 cpu = (u32) ~ 0; + int i; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "queue %d", &queue)) + ; + else if (unformat (line_input, "thread %d", &cpu)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + if (cpu < dm->input_cpu_first_index || + cpu >= (dm->input_cpu_first_index + dm->input_cpu_count)) + return clib_error_return (0, "please specify valid thread id"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + for (i = 0; i < vec_len (dm->devices_by_cpu); i++) + { + /* *INDENT-OFF* */ + vec_foreach(dq, dm->devices_by_cpu[i]) + { + if (hw_if_index == dm->devices[dq->device].vlib_hw_if_index && + queue == dq->queue_id) + { + if (cpu == i) /* nothing to do */ + return 0; + + vec_del1(dm->devices_by_cpu[i], dq - dm->devices_by_cpu[i]); + vec_add2(dm->devices_by_cpu[cpu], dq, 1); + dq->queue_id = queue; + dq->device = xd->device_index; + xd->cpu_socket_id_by_queue[queue] = + rte_lcore_to_socket_id(vlib_worker_threads[cpu].lcore_id); + + vec_sort_with_function(dm->devices_by_cpu[i], + dpdk_device_queue_sort); + + vec_sort_with_function(dm->devices_by_cpu[cpu], + dpdk_device_queue_sort); + + if (vec_len(dm->devices_by_cpu[i]) == 0) + vlib_node_set_state (vlib_mains[i], dpdk_input_node.index, + VLIB_NODE_STATE_DISABLED); + + if (vec_len(dm->devices_by_cpu[cpu]) == 1) + vlib_node_set_state (vlib_mains[cpu], dpdk_input_node.index, + VLIB_NODE_STATE_POLLING); + + return 0; + } + } + /* *INDENT-ON* */ + } + + return clib_error_return (0, "not found"); +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_placement,static) = { + .path = "set dpdk interface placement", + .short_help = "set dpdk interface placement [queue ] thread ", + .function = set_dpdk_if_placement, +}; +/* *INDENT-ON* */ + +static clib_error_t * +show_dpdk_if_hqos_placement (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + dpdk_device_and_queue_t *dq; + int cpu; + + if (tm->n_vlib_mains == 1) + vlib_cli_output (vm, "All interfaces are handled by main thread"); + + for (cpu = 0; cpu < vec_len (dm->devices_by_hqos_cpu); cpu++) + { + if (vec_len (dm->devices_by_hqos_cpu[cpu])) + vlib_cli_output (vm, "Thread %u (%s at lcore %u):", cpu, + vlib_worker_threads[cpu].name, + vlib_worker_threads[cpu].lcore_id); + + vec_foreach (dq, dm->devices_by_hqos_cpu[cpu]) + { + u32 hw_if_index = dm->devices[dq->device].vlib_hw_if_index; + vnet_hw_interface_t *hi = + vnet_get_hw_interface (dm->vnet_main, hw_if_index); + vlib_cli_output (vm, " %v queue %u", hi->name, dq->queue_id); + } + } + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_show_dpdk_if_hqos_placement, static) = { + .path = "show dpdk interface hqos placement", + .short_help = "show dpdk interface hqos placement", + .function = show_dpdk_if_hqos_placement, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_hqos_placement (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + dpdk_device_and_queue_t *dq; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 cpu = (u32) ~ 0; + int i; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "thread %d", &cpu)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + if (cpu < dm->hqos_cpu_first_index || + cpu >= (dm->hqos_cpu_first_index + dm->hqos_cpu_count)) + return clib_error_return (0, "please specify valid thread id"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + for (i = 0; i < vec_len (dm->devices_by_hqos_cpu); i++) + { + vec_foreach (dq, dm->devices_by_hqos_cpu[i]) + { + if (hw_if_index == dm->devices[dq->device].vlib_hw_if_index) + { + if (cpu == i) /* nothing to do */ + return 0; + + vec_del1 (dm->devices_by_hqos_cpu[i], + dq - dm->devices_by_hqos_cpu[i]); + vec_add2 (dm->devices_by_hqos_cpu[cpu], dq, 1); + dq->queue_id = 0; + dq->device = xd->device_index; + + vec_sort_with_function (dm->devices_by_hqos_cpu[i], + dpdk_device_queue_sort); + + vec_sort_with_function (dm->devices_by_hqos_cpu[cpu], + dpdk_device_queue_sort); + + return 0; + } + } + } + + return clib_error_return (0, "not found"); +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_hqos_placement, static) = { + .path = "set dpdk interface hqos placement", + .short_help = "set dpdk interface hqos placement thread ", + .function = set_dpdk_if_hqos_placement, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_hqos_pipe (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 subport_id = (u32) ~ 0; + u32 pipe_id = (u32) ~ 0; + u32 profile_id = (u32) ~ 0; + int rv; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "subport %d", &subport_id)) + ; + else if (unformat (line_input, "pipe %d", &pipe_id)) + ; + else if (unformat (line_input, "profile %d", &profile_id)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + rv = + rte_sched_pipe_config (xd->hqos_ht->hqos, subport_id, pipe_id, + profile_id); + if (rv) + return clib_error_return (0, "pipe configuration failed"); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_hqos_pipe, static) = +{ + .path = "set dpdk interface hqos pipe", + .short_help = "set dpdk interface hqos pipe subport pipe " + "profile ", + .function = set_dpdk_if_hqos_pipe, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_hqos_subport (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 subport_id = (u32) ~ 0; + struct rte_sched_subport_params p = { + .tb_rate = 1250000000, /* 10GbE */ + .tb_size = 1000000, + .tc_rate = {1250000000, 1250000000, 1250000000, 1250000000}, + .tc_period = 10, + }; + int rv; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "subport %d", &subport_id)) + ; + else if (unformat (line_input, "rate %d", &p.tb_rate)) + { + p.tc_rate[0] = p.tb_rate; + p.tc_rate[1] = p.tb_rate; + p.tc_rate[2] = p.tb_rate; + p.tc_rate[3] = p.tb_rate; + } + else if (unformat (line_input, "bktsize %d", &p.tb_size)) + ; + else if (unformat (line_input, "tc0 %d", &p.tc_rate[0])) + ; + else if (unformat (line_input, "tc1 %d", &p.tc_rate[1])) + ; + else if (unformat (line_input, "tc2 %d", &p.tc_rate[2])) + ; + else if (unformat (line_input, "tc3 %d", &p.tc_rate[3])) + ; + else if (unformat (line_input, "period %d", &p.tc_period)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + rv = rte_sched_subport_config (xd->hqos_ht->hqos, subport_id, &p); + if (rv) + return clib_error_return (0, "subport configuration failed"); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_hqos_subport, static) = { + .path = "set dpdk interface hqos subport", + .short_help = "set dpdk interface hqos subport subport " + "[rate ] [bktsize ] [tc0 ] [tc1 ] [tc2 ] [tc3 ] " + "[period ]", + .function = set_dpdk_if_hqos_subport, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_hqos_tctbl (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 tc = (u32) ~ 0; + u32 queue = (u32) ~ 0; + u32 entry = (u32) ~ 0; + u32 val, i; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "entry %d", &entry)) + ; + else if (unformat (line_input, "tc %d", &tc)) + ; + else if (unformat (line_input, "queue %d", &queue)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + if (entry >= 64) + return clib_error_return (0, "invalid entry"); + if (tc >= RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE) + return clib_error_return (0, "invalid traffic class"); + if (queue >= RTE_SCHED_QUEUES_PER_TRAFFIC_CLASS) + return clib_error_return (0, "invalid traffic class"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + /* Detect the set of worker threads */ + uword *p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + /* Should never happen, shut up Coverity warning */ + if (p == 0) + return clib_error_return (0, "no worker registrations?"); + + vlib_thread_registration_t *tr = (vlib_thread_registration_t *) p[0]; + int worker_thread_first = tr->first_index; + int worker_thread_count = tr->count; + + val = tc * RTE_SCHED_QUEUES_PER_TRAFFIC_CLASS + queue; + for (i = 0; i < worker_thread_count; i++) + xd->hqos_wt[worker_thread_first + i].hqos_tc_table[entry] = val; + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_hqos_tctbl, static) = { + .path = "set dpdk interface hqos tctbl", + .short_help = "set dpdk interface hqos tctbl entry tc queue ", + .function = set_dpdk_if_hqos_tctbl, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_hqos_pktfield (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + + /* Device specific data */ + struct rte_eth_dev_info dev_info; + dpdk_device_config_t *devconf = 0; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + + /* Detect the set of worker threads */ + uword *p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + /* Should never happen, shut up Coverity warning */ + if (p == 0) + return clib_error_return (0, "no worker registrations?"); + + vlib_thread_registration_t *tr = (vlib_thread_registration_t *) p[0]; + int worker_thread_first = tr->first_index; + int worker_thread_count = tr->count; + + /* Packet field configuration */ + u64 mask = (u64) ~ 0; + u32 id = (u32) ~ 0; + u32 offset = (u32) ~ 0; + + /* HQoS params */ + u32 n_subports_per_port, n_pipes_per_subport, tctbl_size; + + u32 i; + + /* Parse input arguments */ + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "id %d", &id)) + ; + else if (unformat (line_input, "offset %d", &offset)) + ; + else if (unformat (line_input, "mask %llx", &mask)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + /* Get interface */ + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + rte_eth_dev_info_get (xd->device_index, &dev_info); + if (dev_info.pci_dev) + { /* bonded interface has no pci info */ + vlib_pci_addr_t pci_addr; + + pci_addr.domain = dev_info.pci_dev->addr.domain; + pci_addr.bus = dev_info.pci_dev->addr.bus; + pci_addr.slot = dev_info.pci_dev->addr.devid; + pci_addr.function = dev_info.pci_dev->addr.function; + + p = + hash_get (dm->conf->device_config_index_by_pci_addr, pci_addr.as_u32); + } + + if (p) + devconf = pool_elt_at_index (dm->conf->dev_confs, p[0]); + else + devconf = &dm->conf->default_devconf; + + if (devconf->hqos_enabled == 0) + { + vlib_cli_output (vm, "HQoS disabled for this interface"); + return 0; + } + + n_subports_per_port = devconf->hqos.port.n_subports_per_port; + n_pipes_per_subport = devconf->hqos.port.n_pipes_per_subport; + tctbl_size = RTE_DIM (devconf->hqos.tc_table); + + /* Validate packet field configuration: id, offset and mask */ + if (id >= 3) + return clib_error_return (0, "invalid packet field id"); + + switch (id) + { + case 0: + if (dpdk_hqos_validate_mask (mask, n_subports_per_port) != 0) + return clib_error_return (0, "invalid subport ID mask " + "(n_subports_per_port = %u)", + n_subports_per_port); + break; + case 1: + if (dpdk_hqos_validate_mask (mask, n_pipes_per_subport) != 0) + return clib_error_return (0, "invalid pipe ID mask " + "(n_pipes_per_subport = %u)", + n_pipes_per_subport); + break; + case 2: + default: + if (dpdk_hqos_validate_mask (mask, tctbl_size) != 0) + return clib_error_return (0, "invalid TC table index mask " + "(TC table size = %u)", tctbl_size); + } + + /* Propagate packet field configuration to all workers */ + for (i = 0; i < worker_thread_count; i++) + switch (id) + { + case 0: + xd->hqos_wt[worker_thread_first + i].hqos_field0_slabpos = offset; + xd->hqos_wt[worker_thread_first + i].hqos_field0_slabmask = mask; + xd->hqos_wt[worker_thread_first + i].hqos_field0_slabshr = + __builtin_ctzll (mask); + break; + case 1: + xd->hqos_wt[worker_thread_first + i].hqos_field1_slabpos = offset; + xd->hqos_wt[worker_thread_first + i].hqos_field1_slabmask = mask; + xd->hqos_wt[worker_thread_first + i].hqos_field1_slabshr = + __builtin_ctzll (mask); + break; + case 2: + default: + xd->hqos_wt[worker_thread_first + i].hqos_field2_slabpos = offset; + xd->hqos_wt[worker_thread_first + i].hqos_field2_slabmask = mask; + xd->hqos_wt[worker_thread_first + i].hqos_field2_slabshr = + __builtin_ctzll (mask); + } + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_hqos_pktfield, static) = { + .path = "set dpdk interface hqos pktfield", + .short_help = "set dpdk interface hqos pktfield id offset " + "mask ", + .function = set_dpdk_if_hqos_pktfield, +}; +/* *INDENT-ON* */ + +static clib_error_t * +show_dpdk_if_hqos (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + dpdk_device_config_hqos_t *cfg; + dpdk_device_hqos_per_hqos_thread_t *ht; + dpdk_device_hqos_per_worker_thread_t *wk; + u32 *tctbl; + u32 hw_if_index = (u32) ~ 0; + u32 profile_id, i; + struct rte_eth_dev_info dev_info; + dpdk_device_config_t *devconf = 0; + vlib_thread_registration_t *tr; + uword *p = 0; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify interface name!!"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + rte_eth_dev_info_get (xd->device_index, &dev_info); + if (dev_info.pci_dev) + { /* bonded interface has no pci info */ + vlib_pci_addr_t pci_addr; + + pci_addr.domain = dev_info.pci_dev->addr.domain; + pci_addr.bus = dev_info.pci_dev->addr.bus; + pci_addr.slot = dev_info.pci_dev->addr.devid; + pci_addr.function = dev_info.pci_dev->addr.function; + + p = + hash_get (dm->conf->device_config_index_by_pci_addr, pci_addr.as_u32); + } + + if (p) + devconf = pool_elt_at_index (dm->conf->dev_confs, p[0]); + else + devconf = &dm->conf->default_devconf; + + if (devconf->hqos_enabled == 0) + { + vlib_cli_output (vm, "HQoS disabled for this interface"); + return 0; + } + + /* Detect the set of worker threads */ + p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + + /* Should never happen, shut up Coverity warning */ + if (p == 0) + return clib_error_return (0, "no worker registrations?"); + + tr = (vlib_thread_registration_t *) p[0]; + + cfg = &devconf->hqos; + ht = xd->hqos_ht; + wk = &xd->hqos_wt[tr->first_index]; + tctbl = wk->hqos_tc_table; + + vlib_cli_output (vm, " Thread:"); + vlib_cli_output (vm, " Input SWQ size = %u packets", cfg->swq_size); + vlib_cli_output (vm, " Enqueue burst size = %u packets", + ht->hqos_burst_enq); + vlib_cli_output (vm, " Dequeue burst size = %u packets", + ht->hqos_burst_deq); + + vlib_cli_output (vm, + " Packet field 0: slab position = %4u, slab bitmask = 0x%016llx", + wk->hqos_field0_slabpos, wk->hqos_field0_slabmask); + vlib_cli_output (vm, + " Packet field 1: slab position = %4u, slab bitmask = 0x%016llx", + wk->hqos_field1_slabpos, wk->hqos_field1_slabmask); + vlib_cli_output (vm, + " Packet field 2: slab position = %4u, slab bitmask = 0x%016llx", + wk->hqos_field2_slabpos, wk->hqos_field2_slabmask); + vlib_cli_output (vm, " Packet field 2 translation table:"); + vlib_cli_output (vm, " [ 0 .. 15]: " + "%2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u", + tctbl[0], tctbl[1], tctbl[2], tctbl[3], + tctbl[4], tctbl[5], tctbl[6], tctbl[7], + tctbl[8], tctbl[9], tctbl[10], tctbl[11], + tctbl[12], tctbl[13], tctbl[14], tctbl[15]); + vlib_cli_output (vm, " [16 .. 31]: " + "%2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u", + tctbl[16], tctbl[17], tctbl[18], tctbl[19], + tctbl[20], tctbl[21], tctbl[22], tctbl[23], + tctbl[24], tctbl[25], tctbl[26], tctbl[27], + tctbl[28], tctbl[29], tctbl[30], tctbl[31]); + vlib_cli_output (vm, " [32 .. 47]: " + "%2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u", + tctbl[32], tctbl[33], tctbl[34], tctbl[35], + tctbl[36], tctbl[37], tctbl[38], tctbl[39], + tctbl[40], tctbl[41], tctbl[42], tctbl[43], + tctbl[44], tctbl[45], tctbl[46], tctbl[47]); + vlib_cli_output (vm, " [48 .. 63]: " + "%2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u", + tctbl[48], tctbl[49], tctbl[50], tctbl[51], + tctbl[52], tctbl[53], tctbl[54], tctbl[55], + tctbl[56], tctbl[57], tctbl[58], tctbl[59], + tctbl[60], tctbl[61], tctbl[62], tctbl[63]); + + vlib_cli_output (vm, " Port:"); + vlib_cli_output (vm, " Rate = %u bytes/second", cfg->port.rate); + vlib_cli_output (vm, " MTU = %u bytes", cfg->port.mtu); + vlib_cli_output (vm, " Frame overhead = %u bytes", + cfg->port.frame_overhead); + vlib_cli_output (vm, " Number of subports = %u", + cfg->port.n_subports_per_port); + vlib_cli_output (vm, " Number of pipes per subport = %u", + cfg->port.n_pipes_per_subport); + vlib_cli_output (vm, + " Packet queue size: TC0 = %u, TC1 = %u, TC2 = %u, TC3 = %u packets", + cfg->port.qsize[0], cfg->port.qsize[1], cfg->port.qsize[2], + cfg->port.qsize[3]); + vlib_cli_output (vm, " Number of pipe profiles = %u", + cfg->port.n_pipe_profiles); + + for (profile_id = 0; profile_id < vec_len (cfg->pipe); profile_id++) + { + vlib_cli_output (vm, " Pipe profile %u:", profile_id); + vlib_cli_output (vm, " Rate = %u bytes/second", + cfg->pipe[profile_id].tb_rate); + vlib_cli_output (vm, " Token bucket size = %u bytes", + cfg->pipe[profile_id].tb_size); + vlib_cli_output (vm, + " Traffic class rate: TC0 = %u, TC1 = %u, TC2 = %u, TC3 = %u bytes/second", + cfg->pipe[profile_id].tc_rate[0], + cfg->pipe[profile_id].tc_rate[1], + cfg->pipe[profile_id].tc_rate[2], + cfg->pipe[profile_id].tc_rate[3]); + vlib_cli_output (vm, " TC period = %u milliseconds", + cfg->pipe[profile_id].tc_period); +#ifdef RTE_SCHED_SUBPORT_TC_OV + vlib_cli_output (vm, " TC3 oversubscription_weight = %u", + cfg->pipe[profile_id].tc_ov_weight); +#endif + + for (i = 0; i < RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE; i++) + { + vlib_cli_output (vm, + " TC%u WRR weights: Q0 = %u, Q1 = %u, Q2 = %u, Q3 = %u", + i, cfg->pipe[profile_id].wrr_weights[i * 4], + cfg->pipe[profile_id].wrr_weights[i * 4 + 1], + cfg->pipe[profile_id].wrr_weights[i * 4 + 2], + cfg->pipe[profile_id].wrr_weights[i * 4 + 3]); + } + } + +#ifdef RTE_SCHED_RED + vlib_cli_output (vm, " Weighted Random Early Detection (WRED):"); + for (i = 0; i < RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE; i++) + { + vlib_cli_output (vm, " TC%u min: G = %u, Y = %u, R = %u", i, + cfg->port.red_params[i][e_RTE_METER_GREEN].min_th, + cfg->port.red_params[i][e_RTE_METER_YELLOW].min_th, + cfg->port.red_params[i][e_RTE_METER_RED].min_th); + + vlib_cli_output (vm, " TC%u max: G = %u, Y = %u, R = %u", i, + cfg->port.red_params[i][e_RTE_METER_GREEN].max_th, + cfg->port.red_params[i][e_RTE_METER_YELLOW].max_th, + cfg->port.red_params[i][e_RTE_METER_RED].max_th); + + vlib_cli_output (vm, + " TC%u inverted probability: G = %u, Y = %u, R = %u", + i, cfg->port.red_params[i][e_RTE_METER_GREEN].maxp_inv, + cfg->port.red_params[i][e_RTE_METER_YELLOW].maxp_inv, + cfg->port.red_params[i][e_RTE_METER_RED].maxp_inv); + + vlib_cli_output (vm, " TC%u weight: R = %u, Y = %u, R = %u", i, + cfg->port.red_params[i][e_RTE_METER_GREEN].wq_log2, + cfg->port.red_params[i][e_RTE_METER_YELLOW].wq_log2, + cfg->port.red_params[i][e_RTE_METER_RED].wq_log2); + } +#endif + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_show_dpdk_if_hqos, static) = { + .path = "show dpdk interface hqos", + .short_help = "show dpdk interface hqos ", + .function = show_dpdk_if_hqos, +}; + +/* *INDENT-ON* */ + +static clib_error_t * +show_dpdk_hqos_queue_stats (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + u32 hw_if_index = (u32) ~ 0; + u32 subport = (u32) ~ 0; + u32 pipe = (u32) ~ 0; + u32 tc = (u32) ~ 0; + u32 tc_q = (u32) ~ 0; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + uword *p = 0; + struct rte_eth_dev_info dev_info; + dpdk_device_config_t *devconf = 0; + u32 qindex; + struct rte_sched_queue_stats stats; + u16 qlen; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + + else if (unformat (line_input, "subport %d", &subport)) + ; + + else if (unformat (line_input, "pipe %d", &pipe)) + ; + + else if (unformat (line_input, "tc %d", &tc)) + ; + + else if (unformat (line_input, "tc_q %d", &tc_q)) + ; + + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify interface name!!"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + rte_eth_dev_info_get (xd->device_index, &dev_info); + if (dev_info.pci_dev) + { /* bonded interface has no pci info */ + vlib_pci_addr_t pci_addr; + + pci_addr.domain = dev_info.pci_dev->addr.domain; + pci_addr.bus = dev_info.pci_dev->addr.bus; + pci_addr.slot = dev_info.pci_dev->addr.devid; + pci_addr.function = dev_info.pci_dev->addr.function; + + p = + hash_get (dm->conf->device_config_index_by_pci_addr, pci_addr.as_u32); + } + + if (p) + devconf = pool_elt_at_index (dm->conf->dev_confs, p[0]); + else + devconf = &dm->conf->default_devconf; + + if (devconf->hqos_enabled == 0) + { + vlib_cli_output (vm, "HQoS disabled for this interface"); + return 0; + } + + /* + * Figure out which queue to query. cf rte_sched_port_qindex. (Not sure why + * that method isn't made public by DPDK - how _should_ we get the queue ID?) + */ + qindex = subport * devconf->hqos.port.n_pipes_per_subport + pipe; + qindex = qindex * RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE + tc; + qindex = qindex * RTE_SCHED_QUEUES_PER_TRAFFIC_CLASS + tc_q; + + if (rte_sched_queue_read_stats (xd->hqos_ht->hqos, qindex, &stats, &qlen) != + 0) + return clib_error_return (0, "failed to read stats"); + + vlib_cli_output (vm, "%=24s%=16s", "Stats Parameter", "Value"); + vlib_cli_output (vm, "%=24s%=16d", "Packets", stats.n_pkts); + vlib_cli_output (vm, "%=24s%=16d", "Packets dropped", stats.n_pkts_dropped); +#ifdef RTE_SCHED_RED + vlib_cli_output (vm, "%=24s%=16d", "Packets dropped (RED)", + stats.n_pkts_red_dropped); +#endif + vlib_cli_output (vm, "%=24s%=16d", "Bytes", stats.n_bytes); + vlib_cli_output (vm, "%=24s%=16d", "Bytes dropped", stats.n_bytes_dropped); + + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_show_dpdk_hqos_queue_stats, static) = { + .path = "show dpdk hqos queue", + .short_help = "show dpdk hqos queue subport pipe tc tc_q ", + .function = show_dpdk_hqos_queue_stats, +}; +/* *INDENT-ON* */ + +clib_error_t * +dpdk_cli_init (vlib_main_t * vm) +{ + return 0; +} + +VLIB_INIT_FUNCTION (dpdk_cli_init); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/device.c b/src/vnet/devices/dpdk/device.c new file mode 100644 index 00000000000..b22fbf2e69e --- /dev/null +++ b/src/vnet/devices/dpdk/device.c @@ -0,0 +1,840 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include +#include +#include +#include +#include + +#include +#include + +#include "dpdk_priv.h" +#include + +#define foreach_dpdk_tx_func_error \ + _(BAD_RETVAL, "DPDK tx function returned an error") \ + _(RING_FULL, "Tx packet drops (ring full)") \ + _(PKT_DROP, "Tx packet drops (dpdk tx failure)") \ + _(REPL_FAIL, "Tx packet drops (replication failure)") + +typedef enum +{ +#define _(f,s) DPDK_TX_FUNC_ERROR_##f, + foreach_dpdk_tx_func_error +#undef _ + DPDK_TX_FUNC_N_ERROR, +} dpdk_tx_func_error_t; + +static char *dpdk_tx_func_error_strings[] = { +#define _(n,s) s, + foreach_dpdk_tx_func_error +#undef _ +}; + +clib_error_t * +dpdk_set_mac_address (vnet_hw_interface_t * hi, char *address) +{ + int error; + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, hi->dev_instance); + + error = rte_eth_dev_default_mac_addr_set (xd->device_index, + (struct ether_addr *) address); + + if (error) + { + return clib_error_return (0, "mac address set failed: %d", error); + } + else + { + return NULL; + } +} + +clib_error_t * +dpdk_set_mc_filter (vnet_hw_interface_t * hi, + struct ether_addr mc_addr_vec[], int naddr) +{ + int error; + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, hi->dev_instance); + + error = rte_eth_dev_set_mc_addr_list (xd->device_index, mc_addr_vec, naddr); + + if (error) + { + return clib_error_return (0, "mc addr list failed: %d", error); + } + else + { + return NULL; + } +} + +struct rte_mbuf * +dpdk_replicate_packet_mb (vlib_buffer_t * b) +{ + vlib_main_t *vm = vlib_get_main (); + vlib_buffer_main_t *bm = vm->buffer_main; + struct rte_mbuf **mbufs = 0, *s, *d; + u8 nb_segs; + unsigned socket_id = rte_socket_id (); + int i; + + ASSERT (bm->pktmbuf_pools[socket_id]); + s = rte_mbuf_from_vlib_buffer (b); + nb_segs = s->nb_segs; + vec_validate (mbufs, nb_segs - 1); + + if (rte_pktmbuf_alloc_bulk (bm->pktmbuf_pools[socket_id], mbufs, nb_segs)) + { + vec_free (mbufs); + return 0; + } + + d = mbufs[0]; + d->nb_segs = s->nb_segs; + d->data_len = s->data_len; + d->pkt_len = s->pkt_len; + d->data_off = s->data_off; + clib_memcpy (d->buf_addr, s->buf_addr, RTE_PKTMBUF_HEADROOM + s->data_len); + + for (i = 1; i < nb_segs; i++) + { + d->next = mbufs[i]; + d = mbufs[i]; + s = s->next; + d->data_len = s->data_len; + clib_memcpy (d->buf_addr, s->buf_addr, + RTE_PKTMBUF_HEADROOM + s->data_len); + } + + d = mbufs[0]; + vec_free (mbufs); + return d; +} + +static void +dpdk_tx_trace_buffer (dpdk_main_t * dm, + vlib_node_runtime_t * node, + dpdk_device_t * xd, + u16 queue_id, u32 buffer_index, vlib_buffer_t * buffer) +{ + vlib_main_t *vm = vlib_get_main (); + dpdk_tx_dma_trace_t *t0; + struct rte_mbuf *mb; + + mb = rte_mbuf_from_vlib_buffer (buffer); + + t0 = vlib_add_trace (vm, node, buffer, sizeof (t0[0])); + t0->queue_index = queue_id; + t0->device_index = xd->device_index; + t0->buffer_index = buffer_index; + clib_memcpy (&t0->mb, mb, sizeof (t0->mb)); + clib_memcpy (&t0->buffer, buffer, + sizeof (buffer[0]) - sizeof (buffer->pre_data)); + clib_memcpy (t0->buffer.pre_data, buffer->data + buffer->current_data, + sizeof (t0->buffer.pre_data)); +} + +static_always_inline void +dpdk_validate_rte_mbuf (vlib_main_t * vm, vlib_buffer_t * b, + int maybe_multiseg) +{ + struct rte_mbuf *mb, *first_mb, *last_mb; + + /* buffer is coming from non-dpdk source so we need to init + rte_mbuf header */ + if (PREDICT_FALSE ((b->flags & VNET_BUFFER_RTE_MBUF_VALID) == 0)) + { + vlib_buffer_t *b2 = b; + last_mb = mb = rte_mbuf_from_vlib_buffer (b2); + rte_pktmbuf_reset (mb); + while (maybe_multiseg && (b2->flags & VLIB_BUFFER_NEXT_PRESENT)) + { + b2 = vlib_get_buffer (vm, b2->next_buffer); + mb = rte_mbuf_from_vlib_buffer (b2); + last_mb->next = mb; + last_mb = mb; + rte_pktmbuf_reset (mb); + } + } + + first_mb = mb = rte_mbuf_from_vlib_buffer (b); + first_mb->nb_segs = 1; + mb->data_len = b->current_length; + mb->pkt_len = maybe_multiseg ? vlib_buffer_length_in_chain (vm, b) : + b->current_length; + mb->data_off = VLIB_BUFFER_PRE_DATA_SIZE + b->current_data; + + while (maybe_multiseg && (b->flags & VLIB_BUFFER_NEXT_PRESENT)) + { + b = vlib_get_buffer (vm, b->next_buffer); + mb = rte_mbuf_from_vlib_buffer (b); + mb->data_len = b->current_length; + mb->pkt_len = b->current_length; + mb->data_off = VLIB_BUFFER_PRE_DATA_SIZE + b->current_data; + first_mb->nb_segs++; + } +} + +/* + * This function calls the dpdk's tx_burst function to transmit the packets + * on the tx_vector. It manages a lock per-device if the device does not + * support multiple queues. It returns the number of packets untransmitted + * on the tx_vector. If all packets are transmitted (the normal case), the + * function returns 0. + * + * The function assumes there is at least one packet on the tx_vector. + */ +static_always_inline + u32 tx_burst_vector_internal (vlib_main_t * vm, + dpdk_device_t * xd, + struct rte_mbuf **tx_vector) +{ + dpdk_main_t *dm = &dpdk_main; + u32 n_packets; + u32 tx_head; + u32 tx_tail; + u32 n_retry; + int rv; + int queue_id; + tx_ring_hdr_t *ring; + + ring = vec_header (tx_vector, sizeof (*ring)); + + n_packets = ring->tx_head - ring->tx_tail; + + tx_head = ring->tx_head % xd->nb_tx_desc; + + /* + * Ensure rte_eth_tx_burst is not called with 0 packets, which can lead to + * unpredictable results. + */ + ASSERT (n_packets > 0); + + /* + * Check for tx_vector overflow. If this fails it is a system configuration + * error. The ring should be sized big enough to handle the largest un-flowed + * off burst from a traffic manager. A larger size also helps performance + * a bit because it decreases the probability of having to issue two tx_burst + * calls due to a ring wrap. + */ + ASSERT (n_packets < xd->nb_tx_desc); + ASSERT (ring->tx_tail == 0); + + n_retry = 16; + queue_id = vm->cpu_index; + + do + { + /* start the burst at the tail */ + tx_tail = ring->tx_tail % xd->nb_tx_desc; + + /* + * This device only supports one TX queue, + * and we're running multi-threaded... + */ + if (PREDICT_FALSE (xd->lockp != 0)) + { + queue_id = queue_id % xd->tx_q_used; + while (__sync_lock_test_and_set (xd->lockp[queue_id], 1)) + /* zzzz */ + queue_id = (queue_id + 1) % xd->tx_q_used; + } + + if (PREDICT_FALSE (xd->flags & DPDK_DEVICE_FLAG_HQOS)) /* HQoS ON */ + { + /* no wrap, transmit in one burst */ + dpdk_device_hqos_per_worker_thread_t *hqos = + &xd->hqos_wt[vm->cpu_index]; + + ASSERT (hqos->swq != NULL); + + dpdk_hqos_metadata_set (hqos, + &tx_vector[tx_tail], tx_head - tx_tail); + rv = rte_ring_sp_enqueue_burst (hqos->swq, + (void **) &tx_vector[tx_tail], + (uint16_t) (tx_head - tx_tail)); + } + else if (PREDICT_TRUE (xd->flags & DPDK_DEVICE_FLAG_PMD)) + { + /* no wrap, transmit in one burst */ + rv = rte_eth_tx_burst (xd->device_index, + (uint16_t) queue_id, + &tx_vector[tx_tail], + (uint16_t) (tx_head - tx_tail)); + } + else + { + ASSERT (0); + rv = 0; + } + + if (PREDICT_FALSE (xd->lockp != 0)) + *xd->lockp[queue_id] = 0; + + if (PREDICT_FALSE (rv < 0)) + { + // emit non-fatal message, bump counter + vnet_main_t *vnm = dm->vnet_main; + vnet_interface_main_t *im = &vnm->interface_main; + u32 node_index; + + node_index = vec_elt_at_index (im->hw_interfaces, + xd->vlib_hw_if_index)->tx_node_index; + + vlib_error_count (vm, node_index, DPDK_TX_FUNC_ERROR_BAD_RETVAL, 1); + clib_warning ("rte_eth_tx_burst[%d]: error %d", xd->device_index, + rv); + return n_packets; // untransmitted packets + } + ring->tx_tail += (u16) rv; + n_packets -= (uint16_t) rv; + } + while (rv && n_packets && (n_retry > 0)); + + return n_packets; +} + +static_always_inline void +dpdk_prefetch_buffer_by_index (vlib_main_t * vm, u32 bi) +{ + vlib_buffer_t *b; + struct rte_mbuf *mb; + b = vlib_get_buffer (vm, bi); + mb = rte_mbuf_from_vlib_buffer (b); + CLIB_PREFETCH (mb, CLIB_CACHE_LINE_BYTES, LOAD); + CLIB_PREFETCH (b, CLIB_CACHE_LINE_BYTES, LOAD); +} + +static_always_inline void +dpdk_buffer_recycle (vlib_main_t * vm, vlib_node_runtime_t * node, + vlib_buffer_t * b, u32 bi, struct rte_mbuf **mbp) +{ + dpdk_main_t *dm = &dpdk_main; + u32 my_cpu = vm->cpu_index; + struct rte_mbuf *mb_new; + + if (PREDICT_FALSE (b->flags & VLIB_BUFFER_RECYCLE) == 0) + return; + + mb_new = dpdk_replicate_packet_mb (b); + if (PREDICT_FALSE (mb_new == 0)) + { + vlib_error_count (vm, node->node_index, + DPDK_TX_FUNC_ERROR_REPL_FAIL, 1); + b->flags |= VLIB_BUFFER_REPL_FAIL; + } + else + *mbp = mb_new; + + vec_add1 (dm->recycle[my_cpu], bi); +} + +/* + * Transmits the packets on the frame to the interface associated with the + * node. It first copies packets on the frame to a tx_vector containing the + * rte_mbuf pointers. It then passes this vector to tx_burst_vector_internal + * which calls the dpdk tx_burst function. + */ +static uword +dpdk_interface_tx (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * f) +{ + dpdk_main_t *dm = &dpdk_main; + vnet_interface_output_runtime_t *rd = (void *) node->runtime_data; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, rd->dev_instance); + u32 n_packets = f->n_vectors; + u32 n_left; + u32 *from; + struct rte_mbuf **tx_vector; + u16 i; + u16 nb_tx_desc = xd->nb_tx_desc; + int queue_id; + u32 my_cpu; + u32 tx_pkts = 0; + tx_ring_hdr_t *ring; + u32 n_on_ring; + + my_cpu = vm->cpu_index; + + queue_id = my_cpu; + + tx_vector = xd->tx_vectors[queue_id]; + ring = vec_header (tx_vector, sizeof (*ring)); + + n_on_ring = ring->tx_head - ring->tx_tail; + from = vlib_frame_vector_args (f); + + ASSERT (n_packets <= VLIB_FRAME_SIZE); + + if (PREDICT_FALSE (n_on_ring + n_packets > nb_tx_desc)) + { + /* + * Overflowing the ring should never happen. + * If it does then drop the whole frame. + */ + vlib_error_count (vm, node->node_index, DPDK_TX_FUNC_ERROR_RING_FULL, + n_packets); + + while (n_packets--) + { + u32 bi0 = from[n_packets]; + vlib_buffer_t *b0 = vlib_get_buffer (vm, bi0); + struct rte_mbuf *mb0 = rte_mbuf_from_vlib_buffer (b0); + rte_pktmbuf_free (mb0); + } + return n_on_ring; + } + + if (PREDICT_FALSE (dm->tx_pcap_enable)) + { + n_left = n_packets; + while (n_left > 0) + { + u32 bi0 = from[0]; + vlib_buffer_t *b0 = vlib_get_buffer (vm, bi0); + if (dm->pcap_sw_if_index == 0 || + dm->pcap_sw_if_index == vnet_buffer (b0)->sw_if_index[VLIB_TX]) + pcap_add_buffer (&dm->pcap_main, vm, bi0, 512); + from++; + n_left--; + } + } + + from = vlib_frame_vector_args (f); + n_left = n_packets; + i = ring->tx_head % nb_tx_desc; + + while (n_left >= 8) + { + u32 bi0, bi1, bi2, bi3; + struct rte_mbuf *mb0, *mb1, *mb2, *mb3; + vlib_buffer_t *b0, *b1, *b2, *b3; + u32 or_flags; + + dpdk_prefetch_buffer_by_index (vm, from[4]); + dpdk_prefetch_buffer_by_index (vm, from[5]); + dpdk_prefetch_buffer_by_index (vm, from[6]); + dpdk_prefetch_buffer_by_index (vm, from[7]); + + bi0 = from[0]; + bi1 = from[1]; + bi2 = from[2]; + bi3 = from[3]; + from += 4; + + b0 = vlib_get_buffer (vm, bi0); + b1 = vlib_get_buffer (vm, bi1); + b2 = vlib_get_buffer (vm, bi2); + b3 = vlib_get_buffer (vm, bi3); + + or_flags = b0->flags | b1->flags | b2->flags | b3->flags; + + if (or_flags & VLIB_BUFFER_NEXT_PRESENT) + { + dpdk_validate_rte_mbuf (vm, b0, 1); + dpdk_validate_rte_mbuf (vm, b1, 1); + dpdk_validate_rte_mbuf (vm, b2, 1); + dpdk_validate_rte_mbuf (vm, b3, 1); + } + else + { + dpdk_validate_rte_mbuf (vm, b0, 0); + dpdk_validate_rte_mbuf (vm, b1, 0); + dpdk_validate_rte_mbuf (vm, b2, 0); + dpdk_validate_rte_mbuf (vm, b3, 0); + } + + mb0 = rte_mbuf_from_vlib_buffer (b0); + mb1 = rte_mbuf_from_vlib_buffer (b1); + mb2 = rte_mbuf_from_vlib_buffer (b2); + mb3 = rte_mbuf_from_vlib_buffer (b3); + + if (PREDICT_FALSE (or_flags & VLIB_BUFFER_RECYCLE)) + { + dpdk_buffer_recycle (vm, node, b0, bi0, &mb0); + dpdk_buffer_recycle (vm, node, b1, bi1, &mb1); + dpdk_buffer_recycle (vm, node, b2, bi2, &mb2); + dpdk_buffer_recycle (vm, node, b3, bi3, &mb3); + + /* dont enqueue packets if replication failed as they must + be sent back to recycle */ + if (PREDICT_TRUE ((b0->flags & VLIB_BUFFER_REPL_FAIL) == 0)) + tx_vector[i++ % nb_tx_desc] = mb0; + if (PREDICT_TRUE ((b1->flags & VLIB_BUFFER_REPL_FAIL) == 0)) + tx_vector[i++ % nb_tx_desc] = mb1; + if (PREDICT_TRUE ((b2->flags & VLIB_BUFFER_REPL_FAIL) == 0)) + tx_vector[i++ % nb_tx_desc] = mb2; + if (PREDICT_TRUE ((b3->flags & VLIB_BUFFER_REPL_FAIL) == 0)) + tx_vector[i++ % nb_tx_desc] = mb3; + } + else + { + if (PREDICT_FALSE (i + 3 >= nb_tx_desc)) + { + tx_vector[i++ % nb_tx_desc] = mb0; + tx_vector[i++ % nb_tx_desc] = mb1; + tx_vector[i++ % nb_tx_desc] = mb2; + tx_vector[i++ % nb_tx_desc] = mb3; + i %= nb_tx_desc; + } + else + { + tx_vector[i++] = mb0; + tx_vector[i++] = mb1; + tx_vector[i++] = mb2; + tx_vector[i++] = mb3; + } + } + + + if (PREDICT_FALSE (node->flags & VLIB_NODE_FLAG_TRACE)) + { + if (b0->flags & VLIB_BUFFER_IS_TRACED) + dpdk_tx_trace_buffer (dm, node, xd, queue_id, bi0, b0); + if (b1->flags & VLIB_BUFFER_IS_TRACED) + dpdk_tx_trace_buffer (dm, node, xd, queue_id, bi1, b1); + if (b2->flags & VLIB_BUFFER_IS_TRACED) + dpdk_tx_trace_buffer (dm, node, xd, queue_id, bi2, b2); + if (b3->flags & VLIB_BUFFER_IS_TRACED) + dpdk_tx_trace_buffer (dm, node, xd, queue_id, bi3, b3); + } + + n_left -= 4; + } + while (n_left > 0) + { + u32 bi0; + struct rte_mbuf *mb0; + vlib_buffer_t *b0; + + bi0 = from[0]; + from++; + + b0 = vlib_get_buffer (vm, bi0); + + dpdk_validate_rte_mbuf (vm, b0, 1); + + mb0 = rte_mbuf_from_vlib_buffer (b0); + dpdk_buffer_recycle (vm, node, b0, bi0, &mb0); + + if (PREDICT_FALSE (node->flags & VLIB_NODE_FLAG_TRACE)) + if (b0->flags & VLIB_BUFFER_IS_TRACED) + dpdk_tx_trace_buffer (dm, node, xd, queue_id, bi0, b0); + + if (PREDICT_TRUE ((b0->flags & VLIB_BUFFER_REPL_FAIL) == 0)) + { + tx_vector[i % nb_tx_desc] = mb0; + i++; + } + n_left--; + } + + /* account for additional packets in the ring */ + ring->tx_head += n_packets; + n_on_ring = ring->tx_head - ring->tx_tail; + + /* transmit as many packets as possible */ + n_packets = tx_burst_vector_internal (vm, xd, tx_vector); + + /* + * tx_pkts is the number of packets successfully transmitted + * This is the number originally on ring minus the number remaining on ring + */ + tx_pkts = n_on_ring - n_packets; + + { + /* If there is no callback then drop any non-transmitted packets */ + if (PREDICT_FALSE (n_packets)) + { + vlib_simple_counter_main_t *cm; + vnet_main_t *vnm = vnet_get_main (); + + cm = vec_elt_at_index (vnm->interface_main.sw_if_counters, + VNET_INTERFACE_COUNTER_TX_ERROR); + + vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index, + n_packets); + + vlib_error_count (vm, node->node_index, DPDK_TX_FUNC_ERROR_PKT_DROP, + n_packets); + + while (n_packets--) + rte_pktmbuf_free (tx_vector[ring->tx_tail + n_packets]); + } + + /* Reset head/tail to avoid unnecessary wrap */ + ring->tx_head = 0; + ring->tx_tail = 0; + } + + /* Recycle replicated buffers */ + if (PREDICT_FALSE (vec_len (dm->recycle[my_cpu]))) + { + vlib_buffer_free (vm, dm->recycle[my_cpu], + vec_len (dm->recycle[my_cpu])); + _vec_len (dm->recycle[my_cpu]) = 0; + } + + ASSERT (ring->tx_head >= ring->tx_tail); + + return tx_pkts; +} + +static void +dpdk_clear_hw_interface_counters (u32 instance) +{ + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, instance); + + /* + * Set the "last_cleared_stats" to the current stats, so that + * things appear to clear from a display perspective. + */ + dpdk_update_counters (xd, vlib_time_now (dm->vlib_main)); + + clib_memcpy (&xd->last_cleared_stats, &xd->stats, sizeof (xd->stats)); + clib_memcpy (xd->last_cleared_xstats, xd->xstats, + vec_len (xd->last_cleared_xstats) * + sizeof (xd->last_cleared_xstats[0])); + +} + +static clib_error_t * +dpdk_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, u32 flags) +{ + vnet_hw_interface_t *hif = vnet_get_hw_interface (vnm, hw_if_index); + uword is_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0; + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, hif->dev_instance); + int rv = 0; + + if (is_up) + { + f64 now = vlib_time_now (dm->vlib_main); + + if ((xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) == 0) + rv = rte_eth_dev_start (xd->device_index); + + if (xd->flags & DPDK_DEVICE_FLAG_PROMISC) + rte_eth_promiscuous_enable (xd->device_index); + else + rte_eth_promiscuous_disable (xd->device_index); + + rte_eth_allmulticast_enable (xd->device_index); + xd->flags |= DPDK_DEVICE_FLAG_ADMIN_UP; + dpdk_update_counters (xd, now); + dpdk_update_link_state (xd, now); + } + else + { + xd->flags &= ~DPDK_DEVICE_FLAG_ADMIN_UP; + + rte_eth_allmulticast_disable (xd->device_index); + vnet_hw_interface_set_flags (vnm, xd->vlib_hw_if_index, 0); + rte_eth_dev_stop (xd->device_index); + + /* For bonded interface, stop slave links */ + if (xd->pmd == VNET_DPDK_PMD_BOND) + { + u8 slink[16]; + int nlink = rte_eth_bond_slaves_get (xd->device_index, slink, 16); + while (nlink >= 1) + { + u8 dpdk_port = slink[--nlink]; + rte_eth_dev_stop (dpdk_port); + } + } + } + + if (rv < 0) + clib_warning ("rte_eth_dev_%s error: %d", is_up ? "start" : "stop", rv); + + return /* no error */ 0; +} + +/* + * Dynamically redirect all pkts from a specific interface + * to the specified node + */ +static void +dpdk_set_interface_next_node (vnet_main_t * vnm, u32 hw_if_index, + u32 node_index) +{ + dpdk_main_t *xm = &dpdk_main; + vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index); + dpdk_device_t *xd = vec_elt_at_index (xm->devices, hw->dev_instance); + + /* Shut off redirection */ + if (node_index == ~0) + { + xd->per_interface_next_index = node_index; + return; + } + + xd->per_interface_next_index = + vlib_node_add_next (xm->vlib_main, dpdk_input_node.index, node_index); +} + + +static clib_error_t * +dpdk_subif_add_del_function (vnet_main_t * vnm, + u32 hw_if_index, + struct vnet_sw_interface_t *st, int is_add) +{ + dpdk_main_t *xm = &dpdk_main; + vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index); + dpdk_device_t *xd = vec_elt_at_index (xm->devices, hw->dev_instance); + vnet_sw_interface_t *t = (vnet_sw_interface_t *) st; + int r, vlan_offload; + u32 prev_subifs = xd->num_subifs; + clib_error_t *err = 0; + + if (is_add) + xd->num_subifs++; + else if (xd->num_subifs) + xd->num_subifs--; + + if ((xd->flags & DPDK_DEVICE_FLAG_PMD) == 0) + goto done; + + /* currently we program VLANS only for IXGBE VF and I40E VF */ + if ((xd->pmd != VNET_DPDK_PMD_IXGBEVF) && (xd->pmd != VNET_DPDK_PMD_I40EVF)) + goto done; + + if (t->sub.eth.flags.no_tags == 1) + goto done; + + if ((t->sub.eth.flags.one_tag != 1) || (t->sub.eth.flags.exact_match != 1)) + { + xd->num_subifs = prev_subifs; + err = clib_error_return (0, "unsupported VLAN setup"); + goto done; + } + + vlan_offload = rte_eth_dev_get_vlan_offload (xd->device_index); + vlan_offload |= ETH_VLAN_FILTER_OFFLOAD; + + if ((r = rte_eth_dev_set_vlan_offload (xd->device_index, vlan_offload))) + { + xd->num_subifs = prev_subifs; + err = clib_error_return (0, "rte_eth_dev_set_vlan_offload[%d]: err %d", + xd->device_index, r); + goto done; + } + + + if ((r = + rte_eth_dev_vlan_filter (xd->device_index, t->sub.eth.outer_vlan_id, + is_add))) + { + xd->num_subifs = prev_subifs; + err = clib_error_return (0, "rte_eth_dev_vlan_filter[%d]: err %d", + xd->device_index, r); + goto done; + } + +done: + if (xd->num_subifs) + xd->flags |= DPDK_DEVICE_FLAG_HAVE_SUBIF; + else + xd->flags &= ~DPDK_DEVICE_FLAG_HAVE_SUBIF; + + return err; +} + +/* *INDENT-OFF* */ +VNET_DEVICE_CLASS (dpdk_device_class) = { + .name = "dpdk", + .tx_function = dpdk_interface_tx, + .tx_function_n_errors = DPDK_TX_FUNC_N_ERROR, + .tx_function_error_strings = dpdk_tx_func_error_strings, + .format_device_name = format_dpdk_device_name, + .format_device = format_dpdk_device, + .format_tx_trace = format_dpdk_tx_dma_trace, + .clear_counters = dpdk_clear_hw_interface_counters, + .admin_up_down_function = dpdk_interface_admin_up_down, + .subif_add_del_function = dpdk_subif_add_del_function, + .rx_redirect_to_node = dpdk_set_interface_next_node, + .mac_addr_change_function = dpdk_set_mac_address, +}; + +VLIB_DEVICE_TX_FUNCTION_MULTIARCH (dpdk_device_class, dpdk_interface_tx) +/* *INDENT-ON* */ + +#define UP_DOWN_FLAG_EVENT 1 + +uword +admin_up_down_process (vlib_main_t * vm, + vlib_node_runtime_t * rt, vlib_frame_t * f) +{ + clib_error_t *error = 0; + uword event_type; + uword *event_data = 0; + u32 sw_if_index; + u32 flags; + + while (1) + { + vlib_process_wait_for_event (vm); + + event_type = vlib_process_get_events (vm, &event_data); + + dpdk_main.admin_up_down_in_progress = 1; + + switch (event_type) + { + case UP_DOWN_FLAG_EVENT: + { + if (vec_len (event_data) == 2) + { + sw_if_index = event_data[0]; + flags = event_data[1]; + error = + vnet_sw_interface_set_flags (vnet_get_main (), sw_if_index, + flags); + clib_error_report (error); + } + } + break; + } + + vec_reset_length (event_data); + + dpdk_main.admin_up_down_in_progress = 0; + + } + return 0; /* or not */ +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (admin_up_down_process_node,static) = { + .function = admin_up_down_process, + .type = VLIB_NODE_TYPE_PROCESS, + .name = "admin-up-down-process", + .process_log2_n_stack_bytes = 17, // 256KB +}; +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/dpdk.h b/src/vnet/devices/dpdk/dpdk.h new file mode 100644 index 00000000000..d8f378d2b54 --- /dev/null +++ b/src/vnet/devices/dpdk/dpdk.h @@ -0,0 +1,534 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#ifndef __included_dpdk_h__ +#define __included_dpdk_h__ + +/* $$$$ We should rename always_inline -> clib_always_inline */ +#undef always_inline + +#include + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include + +#if CLIB_DEBUG > 0 +#define always_inline static inline +#else +#define always_inline static inline __attribute__ ((__always_inline__)) +#endif + +#include + +#define NB_MBUF (16<<10) + +extern vnet_device_class_t dpdk_device_class; +extern vlib_node_registration_t dpdk_input_node; +extern vlib_node_registration_t handoff_dispatch_node; + +#if RTE_VERSION >= RTE_VERSION_NUM(16, 11, 0, 0) +#define foreach_dpdk_pmd \ + _ ("net_thunderx", THUNDERX) \ + _ ("net_e1000_em", E1000EM) \ + _ ("net_e1000_igb", IGB) \ + _ ("net_e1000_igb_vf", IGBVF) \ + _ ("net_ixgbe", IXGBE) \ + _ ("net_ixgbe_vf", IXGBEVF) \ + _ ("net_i40e", I40E) \ + _ ("net_i40e_vf", I40EVF) \ + _ ("net_virtio", VIRTIO) \ + _ ("net_enic", ENIC) \ + _ ("net_vmxnet3", VMXNET3) \ + _ ("net_af_packet", AF_PACKET) \ + _ ("rte_bond_pmd", BOND) \ + _ ("net_fm10k", FM10K) \ + _ ("net_cxgbe", CXGBE) \ + _ ("net_mlx5", MLX5) \ + _ ("net_dpaa2", DPAA2) +#else +#define foreach_dpdk_pmd \ + _ ("rte_nicvf_pmd", THUNDERX) \ + _ ("rte_em_pmd", E1000EM) \ + _ ("rte_igb_pmd", IGB) \ + _ ("rte_igbvf_pmd", IGBVF) \ + _ ("rte_ixgbe_pmd", IXGBE) \ + _ ("rte_ixgbevf_pmd", IXGBEVF) \ + _ ("rte_i40e_pmd", I40E) \ + _ ("rte_i40evf_pmd", I40EVF) \ + _ ("rte_virtio_pmd", VIRTIO) \ + _ ("rte_enic_pmd", ENIC) \ + _ ("rte_vmxnet3_pmd", VMXNET3) \ + _ ("AF_PACKET PMD", AF_PACKET) \ + _ ("rte_bond_pmd", BOND) \ + _ ("rte_pmd_fm10k", FM10K) \ + _ ("rte_cxgbe_pmd", CXGBE) \ + _ ("rte_dpaa2_dpni", DPAA2) +#endif + +typedef enum +{ + VNET_DPDK_PMD_NONE, +#define _(s,f) VNET_DPDK_PMD_##f, + foreach_dpdk_pmd +#undef _ + VNET_DPDK_PMD_UNKNOWN, /* must be last */ +} dpdk_pmd_t; + +typedef enum +{ + VNET_DPDK_PORT_TYPE_ETH_1G, + VNET_DPDK_PORT_TYPE_ETH_10G, + VNET_DPDK_PORT_TYPE_ETH_40G, + VNET_DPDK_PORT_TYPE_ETH_100G, + VNET_DPDK_PORT_TYPE_ETH_BOND, + VNET_DPDK_PORT_TYPE_ETH_SWITCH, + VNET_DPDK_PORT_TYPE_AF_PACKET, + VNET_DPDK_PORT_TYPE_UNKNOWN, +} dpdk_port_type_t; + +/* + * The header for the tx_vector in dpdk_device_t. + * Head and tail are indexes into the tx_vector and are of type + * u64 so they never overflow. + */ +typedef struct +{ + u64 tx_head; + u64 tx_tail; +} tx_ring_hdr_t; + +typedef struct +{ + struct rte_ring *swq; + + u64 hqos_field0_slabmask; + u32 hqos_field0_slabpos; + u32 hqos_field0_slabshr; + u64 hqos_field1_slabmask; + u32 hqos_field1_slabpos; + u32 hqos_field1_slabshr; + u64 hqos_field2_slabmask; + u32 hqos_field2_slabpos; + u32 hqos_field2_slabshr; + u32 hqos_tc_table[64]; +} dpdk_device_hqos_per_worker_thread_t; + +typedef struct +{ + struct rte_ring **swq; + struct rte_mbuf **pkts_enq; + struct rte_mbuf **pkts_deq; + struct rte_sched_port *hqos; + u32 hqos_burst_enq; + u32 hqos_burst_deq; + u32 pkts_enq_len; + u32 swq_pos; + u32 flush_count; +} dpdk_device_hqos_per_hqos_thread_t; + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + volatile u32 **lockp; + + /* Instance ID */ + u32 device_index; + + u32 vlib_hw_if_index; + u32 vlib_sw_if_index; + + /* next node index if we decide to steal the rx graph arc */ + u32 per_interface_next_index; + + /* dpdk rte_mbuf rx and tx vectors, VLIB_FRAME_SIZE */ + struct rte_mbuf ***tx_vectors; /* one per worker thread */ + struct rte_mbuf ***rx_vectors; + + /* vector of traced contexts, per device */ + u32 **d_trace_buffers; + + dpdk_pmd_t pmd:8; + i8 cpu_socket; + + u16 flags; +#define DPDK_DEVICE_FLAG_ADMIN_UP (1 << 0) +#define DPDK_DEVICE_FLAG_PROMISC (1 << 1) +#define DPDK_DEVICE_FLAG_PMD (1 << 2) +#define DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE (1 << 3) +#define DPDK_DEVICE_FLAG_MAYBE_MULTISEG (1 << 4) +#define DPDK_DEVICE_FLAG_HAVE_SUBIF (1 << 5) +#define DPDK_DEVICE_FLAG_HQOS (1 << 6) + + u16 nb_tx_desc; + CLIB_CACHE_LINE_ALIGN_MARK (cacheline1); + + u8 *interface_name_suffix; + + /* number of sub-interfaces */ + u16 num_subifs; + + /* PMD related */ + u16 tx_q_used; + u16 rx_q_used; + u16 nb_rx_desc; + u16 *cpu_socket_id_by_queue; + struct rte_eth_conf port_conf; + struct rte_eth_txconf tx_conf; + + /* HQoS related */ + dpdk_device_hqos_per_worker_thread_t *hqos_wt; + dpdk_device_hqos_per_hqos_thread_t *hqos_ht; + + /* af_packet */ + u8 af_packet_port_id; + + struct rte_eth_link link; + f64 time_last_link_update; + + struct rte_eth_stats stats; + struct rte_eth_stats last_stats; + struct rte_eth_stats last_cleared_stats; + struct rte_eth_xstat *xstats; + struct rte_eth_xstat *last_cleared_xstats; + f64 time_last_stats_update; + dpdk_port_type_t port_type; +} dpdk_device_t; + +#define DPDK_STATS_POLL_INTERVAL (10.0) +#define DPDK_MIN_STATS_POLL_INTERVAL (0.001) /* 1msec */ + +#define DPDK_LINK_POLL_INTERVAL (3.0) +#define DPDK_MIN_LINK_POLL_INTERVAL (0.001) /* 1msec */ + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + + /* total input packet counter */ + u64 aggregate_rx_packets; +} dpdk_worker_t; + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + + /* total input packet counter */ + u64 aggregate_rx_packets; +} dpdk_hqos_thread_t; + +typedef struct +{ + u32 device; + u16 queue_id; +} dpdk_device_and_queue_t; + +#ifndef DPDK_HQOS_DBG_BYPASS +#define DPDK_HQOS_DBG_BYPASS 0 +#endif + +#ifndef HQOS_FLUSH_COUNT_THRESHOLD +#define HQOS_FLUSH_COUNT_THRESHOLD 100000 +#endif + +typedef struct dpdk_device_config_hqos_t +{ + u32 hqos_thread; + u32 hqos_thread_valid; + + u32 swq_size; + u32 burst_enq; + u32 burst_deq; + + u32 pktfield0_slabpos; + u32 pktfield1_slabpos; + u32 pktfield2_slabpos; + u64 pktfield0_slabmask; + u64 pktfield1_slabmask; + u64 pktfield2_slabmask; + u32 tc_table[64]; + + struct rte_sched_port_params port; + struct rte_sched_subport_params *subport; + struct rte_sched_pipe_params *pipe; + uint32_t *pipe_map; +} dpdk_device_config_hqos_t; + +int dpdk_hqos_validate_mask (u64 mask, u32 n); +void dpdk_device_config_hqos_pipe_profile_default (dpdk_device_config_hqos_t * + hqos, u32 pipe_profile_id); +void dpdk_device_config_hqos_default (dpdk_device_config_hqos_t * hqos); +clib_error_t *dpdk_port_setup_hqos (dpdk_device_t * xd, + dpdk_device_config_hqos_t * hqos); +void dpdk_hqos_metadata_set (dpdk_device_hqos_per_worker_thread_t * hqos, + struct rte_mbuf **pkts, u32 n_pkts); + +#define foreach_dpdk_device_config_item \ + _ (num_rx_queues) \ + _ (num_tx_queues) \ + _ (num_rx_desc) \ + _ (num_tx_desc) \ + _ (rss_fn) + +typedef struct +{ + vlib_pci_addr_t pci_addr; + u8 is_blacklisted; + u8 vlan_strip_offload; +#define DPDK_DEVICE_VLAN_STRIP_DEFAULT 0 +#define DPDK_DEVICE_VLAN_STRIP_OFF 1 +#define DPDK_DEVICE_VLAN_STRIP_ON 2 + +#define _(x) uword x; + foreach_dpdk_device_config_item +#undef _ + clib_bitmap_t * workers; + u32 hqos_enabled; + dpdk_device_config_hqos_t hqos; +} dpdk_device_config_t; + +typedef struct +{ + + /* Config stuff */ + u8 **eal_init_args; + u8 *eal_init_args_str; + u8 *uio_driver_name; + u8 no_multi_seg; + u8 enable_tcp_udp_checksum; + + /* Required config parameters */ + u8 coremask_set_manually; + u8 nchannels_set_manually; + u32 coremask; + u32 nchannels; + u32 num_mbufs; + u8 num_kni; /* while kni_init allows u32, port_id in callback fn is only u8 */ + + /* + * format interface names ala xxxEthernet%d/%d/%d instead of + * xxxEthernet%x/%x/%x. + */ + u8 interface_name_format_decimal; + + /* per-device config */ + dpdk_device_config_t default_devconf; + dpdk_device_config_t *dev_confs; + uword *device_config_index_by_pci_addr; + +} dpdk_config_main_t; + +dpdk_config_main_t dpdk_config_main; + +typedef struct +{ + + /* Devices */ + dpdk_device_t *devices; + dpdk_device_and_queue_t **devices_by_cpu; + dpdk_device_and_queue_t **devices_by_hqos_cpu; + + /* per-thread recycle lists */ + u32 **recycle; + + /* buffer flags template, configurable to enable/disable tcp / udp cksum */ + u32 buffer_flags_template; + + /* vlib buffer free list, must be same size as an rte_mbuf */ + u32 vlib_buffer_free_list_index; + + /* dpdk worker "threads" */ + dpdk_worker_t *workers; + + /* dpdk HQoS "threads" */ + dpdk_hqos_thread_t *hqos_threads; + + /* Ethernet input node index */ + u32 ethernet_input_node_index; + + /* pcap tracing [only works if (CLIB_DEBUG > 0)] */ + int tx_pcap_enable; + pcap_main_t pcap_main; + u8 *pcap_filename; + u32 pcap_sw_if_index; + u32 pcap_pkts_to_capture; + + /* hashes */ + uword *dpdk_device_by_kni_port_id; + uword *vu_sw_if_index_by_listener_fd; + uword *vu_sw_if_index_by_sock_fd; + u32 *vu_inactive_interfaces_device_index; + + /* + * flag indicating that a posted admin up/down + * (via post_sw_interface_set_flags) is in progress + */ + u8 admin_up_down_in_progress; + + u8 use_rss; + + /* which cpus are running dpdk-input */ + int input_cpu_first_index; + int input_cpu_count; + + /* which cpus are running I/O TX */ + int hqos_cpu_first_index; + int hqos_cpu_count; + + /* control interval of dpdk link state and stat polling */ + f64 link_state_poll_interval; + f64 stat_poll_interval; + + /* Sleep for this many MS after each device poll */ + u32 poll_sleep; + + /* convenience */ + vlib_main_t *vlib_main; + vnet_main_t *vnet_main; + dpdk_config_main_t *conf; +} dpdk_main_t; + +dpdk_main_t dpdk_main; + +typedef struct +{ + u32 buffer_index; + u16 device_index; + u8 queue_index; + struct rte_mbuf mb; + /* Copy of VLIB buffer; packet data stored in pre_data. */ + vlib_buffer_t buffer; +} dpdk_tx_dma_trace_t; + +typedef struct +{ + u32 buffer_index; + u16 device_index; + u16 queue_index; + struct rte_mbuf mb; + vlib_buffer_t buffer; /* Copy of VLIB buffer; pkt data stored in pre_data. */ + u8 data[256]; /* First 256 data bytes, used for hexdump */ +} dpdk_rx_dma_trace_t; + +void vnet_buffer_needs_dpdk_mb (vlib_buffer_t * b); + +clib_error_t *dpdk_set_mac_address (vnet_hw_interface_t * hi, char *address); + +clib_error_t *dpdk_set_mc_filter (vnet_hw_interface_t * hi, + struct ether_addr mc_addr_vec[], int naddr); + +void dpdk_thread_input (dpdk_main_t * dm, dpdk_device_t * xd); + +clib_error_t *dpdk_port_setup (dpdk_main_t * dm, dpdk_device_t * xd); + +u32 dpdk_interface_tx_vector (vlib_main_t * vm, u32 dev_instance); + +struct rte_mbuf *dpdk_replicate_packet_mb (vlib_buffer_t * b); +struct rte_mbuf *dpdk_zerocopy_replicate_packet_mb (vlib_buffer_t * b); + +#define foreach_dpdk_error \ + _(NONE, "no error") \ + _(RX_PACKET_ERROR, "Rx packet errors") \ + _(RX_BAD_FCS, "Rx bad fcs") \ + _(IP_CHECKSUM_ERROR, "Rx ip checksum errors") \ + _(RX_ALLOC_FAIL, "rx buf alloc from free list failed") \ + _(RX_ALLOC_NO_PHYSMEM, "rx buf alloc failed no physmem") \ + _(RX_ALLOC_DROP_PKTS, "rx packets dropped due to alloc error") + +typedef enum +{ +#define _(f,s) DPDK_ERROR_##f, + foreach_dpdk_error +#undef _ + DPDK_N_ERROR, +} dpdk_error_t; + +int dpdk_set_stat_poll_interval (f64 interval); +int dpdk_set_link_state_poll_interval (f64 interval); +void dpdk_update_link_state (dpdk_device_t * xd, f64 now); +void dpdk_device_lock_init (dpdk_device_t * xd); +void dpdk_device_lock_free (dpdk_device_t * xd); + +static inline u64 +vnet_get_aggregate_rx_packets (void) +{ + dpdk_main_t *dm = &dpdk_main; + u64 sum = 0; + dpdk_worker_t *dw; + + vec_foreach (dw, dm->workers) sum += dw->aggregate_rx_packets; + + return sum; +} + +void dpdk_rx_trace (dpdk_main_t * dm, + vlib_node_runtime_t * node, + dpdk_device_t * xd, + u16 queue_id, u32 * buffers, uword n_buffers); + +#define EFD_OPERATION_LESS_THAN 0 +#define EFD_OPERATION_GREATER_OR_EQUAL 1 + +format_function_t format_dpdk_device_name; +format_function_t format_dpdk_device; +format_function_t format_dpdk_tx_dma_trace; +format_function_t format_dpdk_rx_dma_trace; +format_function_t format_dpdk_rte_mbuf; +format_function_t format_dpdk_rx_rte_mbuf; +unformat_function_t unformat_socket_mem; +clib_error_t *unformat_rss_fn (unformat_input_t * input, uword * rss_fn); +clib_error_t *unformat_hqos (unformat_input_t * input, + dpdk_device_config_hqos_t * hqos); + +uword +admin_up_down_process (vlib_main_t * vm, + vlib_node_runtime_t * rt, vlib_frame_t * f); + +#endif /* __included_dpdk_h__ */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/dpdk_priv.h b/src/vnet/devices/dpdk/dpdk_priv.h new file mode 100644 index 00000000000..0c81dbc3beb --- /dev/null +++ b/src/vnet/devices/dpdk/dpdk_priv.h @@ -0,0 +1,132 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#define DPDK_NB_RX_DESC_DEFAULT 1024 +#define DPDK_NB_TX_DESC_DEFAULT 1024 +#define DPDK_NB_RX_DESC_VIRTIO 256 +#define DPDK_NB_TX_DESC_VIRTIO 256 + +#define I40E_DEV_ID_SFP_XL710 0x1572 +#define I40E_DEV_ID_QSFP_A 0x1583 +#define I40E_DEV_ID_QSFP_B 0x1584 +#define I40E_DEV_ID_QSFP_C 0x1585 +#define I40E_DEV_ID_10G_BASE_T 0x1586 +#define I40E_DEV_ID_VF 0x154C + +/* These args appear by themselves */ +#define foreach_eal_double_hyphen_predicate_arg \ +_(no-shconf) \ +_(no-hpet) \ +_(no-huge) \ +_(vmware-tsc-map) + +#define foreach_eal_single_hyphen_mandatory_arg \ +_(coremask, c) \ +_(nchannels, n) \ + +#define foreach_eal_single_hyphen_arg \ +_(blacklist, b) \ +_(mem-alloc-request, m) \ +_(force-ranks, r) + +/* These args are preceeded by "--" and followed by a single string */ +#define foreach_eal_double_hyphen_arg \ +_(huge-dir) \ +_(proc-type) \ +_(file-prefix) \ +_(vdev) + +static inline void +dpdk_get_xstats (dpdk_device_t * xd) +{ + int len; + if ((len = rte_eth_xstats_get (xd->device_index, NULL, 0)) > 0) + { + vec_validate (xd->xstats, len - 1); + vec_validate (xd->last_cleared_xstats, len - 1); + + len = + rte_eth_xstats_get (xd->device_index, xd->xstats, + vec_len (xd->xstats)); + + ASSERT (vec_len (xd->xstats) == len); + ASSERT (vec_len (xd->last_cleared_xstats) == len); + + _vec_len (xd->xstats) = len; + _vec_len (xd->last_cleared_xstats) = len; + + } +} + + +static inline void +dpdk_update_counters (dpdk_device_t * xd, f64 now) +{ + vlib_simple_counter_main_t *cm; + vnet_main_t *vnm = vnet_get_main (); + u32 my_cpu = os_get_cpu_number (); + u64 rxerrors, last_rxerrors; + + /* only update counters for PMD interfaces */ + if ((xd->flags & DPDK_DEVICE_FLAG_PMD) == 0) + return; + + xd->time_last_stats_update = now ? now : xd->time_last_stats_update; + clib_memcpy (&xd->last_stats, &xd->stats, sizeof (xd->last_stats)); + rte_eth_stats_get (xd->device_index, &xd->stats); + + /* maybe bump interface rx no buffer counter */ + if (PREDICT_FALSE (xd->stats.rx_nombuf != xd->last_stats.rx_nombuf)) + { + cm = vec_elt_at_index (vnm->interface_main.sw_if_counters, + VNET_INTERFACE_COUNTER_RX_NO_BUF); + + vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index, + xd->stats.rx_nombuf - + xd->last_stats.rx_nombuf); + } + + /* missed pkt counter */ + if (PREDICT_FALSE (xd->stats.imissed != xd->last_stats.imissed)) + { + cm = vec_elt_at_index (vnm->interface_main.sw_if_counters, + VNET_INTERFACE_COUNTER_RX_MISS); + + vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index, + xd->stats.imissed - + xd->last_stats.imissed); + } + rxerrors = xd->stats.ierrors; + last_rxerrors = xd->last_stats.ierrors; + + if (PREDICT_FALSE (rxerrors != last_rxerrors)) + { + cm = vec_elt_at_index (vnm->interface_main.sw_if_counters, + VNET_INTERFACE_COUNTER_RX_ERROR); + + vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index, + rxerrors - last_rxerrors); + } + + dpdk_get_xstats (xd); +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/format.c b/src/vnet/devices/dpdk/format.c new file mode 100644 index 00000000000..ff7c7a5a41c --- /dev/null +++ b/src/vnet/devices/dpdk/format.c @@ -0,0 +1,763 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include +#include +#include +#include +#include + +#include +#include + +#include "dpdk_priv.h" +#include + +#define foreach_dpdk_counter \ + _ (tx_frames_ok, opackets) \ + _ (tx_bytes_ok, obytes) \ + _ (tx_errors, oerrors) \ + _ (rx_frames_ok, ipackets) \ + _ (rx_bytes_ok, ibytes) \ + _ (rx_errors, ierrors) \ + _ (rx_missed, imissed) \ + _ (rx_no_bufs, rx_nombuf) + +#define foreach_dpdk_q_counter \ + _ (rx_frames_ok, q_ipackets) \ + _ (tx_frames_ok, q_opackets) \ + _ (rx_bytes_ok, q_ibytes) \ + _ (tx_bytes_ok, q_obytes) \ + _ (rx_errors, q_errors) + +#define foreach_dpdk_rss_hf \ + _(ETH_RSS_FRAG_IPV4, "ipv4-frag") \ + _(ETH_RSS_NONFRAG_IPV4_TCP, "ipv4-tcp") \ + _(ETH_RSS_NONFRAG_IPV4_UDP, "ipv4-udp") \ + _(ETH_RSS_NONFRAG_IPV4_SCTP, "ipv4-sctp") \ + _(ETH_RSS_NONFRAG_IPV4_OTHER, "ipv4-other") \ + _(ETH_RSS_IPV4, "ipv4") \ + _(ETH_RSS_IPV6_TCP_EX, "ipv6-tcp-ex") \ + _(ETH_RSS_IPV6_UDP_EX, "ipv6-udp-ex") \ + _(ETH_RSS_FRAG_IPV6, "ipv6-frag") \ + _(ETH_RSS_NONFRAG_IPV6_TCP, "ipv6-tcp") \ + _(ETH_RSS_NONFRAG_IPV6_UDP, "ipv6-udp") \ + _(ETH_RSS_NONFRAG_IPV6_SCTP, "ipv6-sctp") \ + _(ETH_RSS_NONFRAG_IPV6_OTHER, "ipv6-other") \ + _(ETH_RSS_L2_PAYLOAD, "l2-payload") \ + _(ETH_RSS_IPV6_EX, "ipv6-ex") \ + _(ETH_RSS_IPV6, "ipv6") + + +#define foreach_dpdk_rx_offload_caps \ + _(DEV_RX_OFFLOAD_VLAN_STRIP, "vlan-strip") \ + _(DEV_RX_OFFLOAD_IPV4_CKSUM, "ipv4-cksum") \ + _(DEV_RX_OFFLOAD_UDP_CKSUM , "udp-cksum") \ + _(DEV_RX_OFFLOAD_TCP_CKSUM , "tcp-cksum") \ + _(DEV_RX_OFFLOAD_TCP_LRO , "rcp-lro") \ + _(DEV_RX_OFFLOAD_QINQ_STRIP, "qinq-strip") + +#define foreach_dpdk_tx_offload_caps \ + _(DEV_TX_OFFLOAD_VLAN_INSERT, "vlan-insert") \ + _(DEV_TX_OFFLOAD_IPV4_CKSUM, "ipv4-cksum") \ + _(DEV_TX_OFFLOAD_UDP_CKSUM , "udp-cksum") \ + _(DEV_TX_OFFLOAD_TCP_CKSUM , "tcp-cksum") \ + _(DEV_TX_OFFLOAD_SCTP_CKSUM , "sctp-cksum") \ + _(DEV_TX_OFFLOAD_TCP_TSO , "tcp-tso") \ + _(DEV_TX_OFFLOAD_UDP_TSO , "udp-tso") \ + _(DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM, "outer-ipv4-cksum") \ + _(DEV_TX_OFFLOAD_QINQ_INSERT, "qinq-insert") + +#if RTE_VERSION < RTE_VERSION_NUM(16, 11, 0, 0) +/* New ol_flags bits added in DPDK-16.11 */ +#define PKT_RX_IP_CKSUM_GOOD (1ULL << 7) +#define PKT_RX_L4_CKSUM_GOOD (1ULL << 8) +#endif + +#define foreach_dpdk_pkt_rx_offload_flag \ + _ (PKT_RX_VLAN_PKT, "RX packet is a 802.1q VLAN packet") \ + _ (PKT_RX_RSS_HASH, "RX packet with RSS hash result") \ + _ (PKT_RX_FDIR, "RX packet with FDIR infos") \ + _ (PKT_RX_L4_CKSUM_BAD, "L4 cksum of RX pkt. is not OK") \ + _ (PKT_RX_IP_CKSUM_BAD, "IP cksum of RX pkt. is not OK") \ + _ (PKT_RX_VLAN_STRIPPED, "RX packet VLAN tag stripped") \ + _ (PKT_RX_IP_CKSUM_GOOD, "IP cksum of RX pkt. is valid") \ + _ (PKT_RX_L4_CKSUM_GOOD, "L4 cksum of RX pkt. is valid") \ + _ (PKT_RX_IEEE1588_PTP, "RX IEEE1588 L2 Ethernet PT Packet") \ + _ (PKT_RX_IEEE1588_TMST, "RX IEEE1588 L2/L4 timestamped packet") \ + _ (PKT_RX_QINQ_STRIPPED, "RX packet QinQ tags stripped") + +#if RTE_VERSION < RTE_VERSION_NUM(16, 11, 0, 0) +/* PTYPE added in DPDK-16.11 */ +#define RTE_PTYPE_L2_ETHER_VLAN 0x00000006 +#define RTE_PTYPE_L2_ETHER_QINQ 0x00000007 +#endif + +#define foreach_dpdk_pkt_type \ + _ (L2, ETHER, "Ethernet packet") \ + _ (L2, ETHER_TIMESYNC, "Ethernet packet for time sync") \ + _ (L2, ETHER_ARP, "ARP packet") \ + _ (L2, ETHER_LLDP, "LLDP (Link Layer Discovery Protocol) packet") \ + _ (L2, ETHER_NSH, "NSH (Network Service Header) packet") \ + _ (L2, ETHER_VLAN, "VLAN packet") \ + _ (L2, ETHER_QINQ, "QinQ packet") \ + _ (L3, IPV4, "IPv4 packet without extension headers") \ + _ (L3, IPV4_EXT, "IPv4 packet with extension headers") \ + _ (L3, IPV4_EXT_UNKNOWN, "IPv4 packet with or without extension headers") \ + _ (L3, IPV6, "IPv6 packet without extension headers") \ + _ (L3, IPV6_EXT, "IPv6 packet with extension headers") \ + _ (L3, IPV6_EXT_UNKNOWN, "IPv6 packet with or without extension headers") \ + _ (L4, TCP, "TCP packet") \ + _ (L4, UDP, "UDP packet") \ + _ (L4, FRAG, "Fragmented IP packet") \ + _ (L4, SCTP, "SCTP (Stream Control Transmission Protocol) packet") \ + _ (L4, ICMP, "ICMP packet") \ + _ (L4, NONFRAG, "Non-fragmented IP packet") \ + _ (TUNNEL, GRE, "GRE tunneling packet") \ + _ (TUNNEL, VXLAN, "VXLAN tunneling packet") \ + _ (TUNNEL, NVGRE, "NVGRE Tunneling packet") \ + _ (TUNNEL, GENEVE, "GENEVE Tunneling packet") \ + _ (TUNNEL, GRENAT, "Teredo, VXLAN or GRE Tunneling packet") \ + _ (INNER_L2, ETHER, "Inner Ethernet packet") \ + _ (INNER_L2, ETHER_VLAN, "Inner Ethernet packet with VLAN") \ + _ (INNER_L3, IPV4, "Inner IPv4 packet without extension headers") \ + _ (INNER_L3, IPV4_EXT, "Inner IPv4 packet with extension headers") \ + _ (INNER_L3, IPV4_EXT_UNKNOWN, "Inner IPv4 packet with or without extension headers") \ + _ (INNER_L3, IPV6, "Inner IPv6 packet without extension headers") \ + _ (INNER_L3, IPV6_EXT, "Inner IPv6 packet with extension headers") \ + _ (INNER_L3, IPV6_EXT_UNKNOWN, "Inner IPv6 packet with or without extension headers") \ + _ (INNER_L4, TCP, "Inner TCP packet") \ + _ (INNER_L4, UDP, "Inner UDP packet") \ + _ (INNER_L4, FRAG, "Inner fagmented IP packet") \ + _ (INNER_L4, SCTP, "Inner SCTP (Stream Control Transmission Protocol) packet") \ + _ (INNER_L4, ICMP, "Inner ICMP packet") \ + _ (INNER_L4, NONFRAG, "Inner non-fragmented IP packet") + +#define foreach_dpdk_pkt_tx_offload_flag \ + _ (PKT_TX_VLAN_PKT, "TX packet is a 802.1q VLAN packet") \ + _ (PKT_TX_IP_CKSUM, "IP cksum of TX pkt. computed by NIC") \ + _ (PKT_TX_TCP_CKSUM, "TCP cksum of TX pkt. computed by NIC") \ + _ (PKT_TX_SCTP_CKSUM, "SCTP cksum of TX pkt. computed by NIC") \ + _ (PKT_TX_IEEE1588_TMST, "TX IEEE1588 packet to timestamp") + +#define foreach_dpdk_pkt_offload_flag \ + foreach_dpdk_pkt_rx_offload_flag \ + foreach_dpdk_pkt_tx_offload_flag + +u8 * +format_dpdk_device_name (u8 * s, va_list * args) +{ + dpdk_main_t *dm = &dpdk_main; + char *devname_format; + char *device_name; + u32 i = va_arg (*args, u32); + struct rte_eth_dev_info dev_info; + u8 *ret; + + if (dm->conf->interface_name_format_decimal) + devname_format = "%s%d/%d/%d"; + else + devname_format = "%s%x/%x/%x"; + + switch (dm->devices[i].port_type) + { + case VNET_DPDK_PORT_TYPE_ETH_1G: + device_name = "GigabitEthernet"; + break; + + case VNET_DPDK_PORT_TYPE_ETH_10G: + device_name = "TenGigabitEthernet"; + break; + + case VNET_DPDK_PORT_TYPE_ETH_40G: + device_name = "FortyGigabitEthernet"; + break; + + case VNET_DPDK_PORT_TYPE_ETH_100G: + device_name = "HundredGigabitEthernet"; + break; + + case VNET_DPDK_PORT_TYPE_ETH_BOND: + return format (s, "BondEthernet%d", dm->devices[i].device_index); + + case VNET_DPDK_PORT_TYPE_ETH_SWITCH: + device_name = "EthernetSwitch"; + break; + + case VNET_DPDK_PORT_TYPE_AF_PACKET: + rte_eth_dev_info_get (i, &dev_info); + return format (s, "af_packet%d", dm->devices[i].af_packet_port_id); + + default: + case VNET_DPDK_PORT_TYPE_UNKNOWN: + device_name = "UnknownEthernet"; + break; + } + + rte_eth_dev_info_get (i, &dev_info); + + if (dev_info.pci_dev) + ret = format (s, devname_format, device_name, dev_info.pci_dev->addr.bus, + dev_info.pci_dev->addr.devid, + dev_info.pci_dev->addr.function); + else + ret = format (s, "%s%d", device_name, dm->devices[i].device_index); + + if (dm->devices[i].interface_name_suffix) + return format (ret, "/%s", dm->devices[i].interface_name_suffix); + return ret; +} + +static u8 * +format_dpdk_device_type (u8 * s, va_list * args) +{ + dpdk_main_t *dm = &dpdk_main; + char *dev_type; + u32 i = va_arg (*args, u32); + + switch (dm->devices[i].pmd) + { + case VNET_DPDK_PMD_E1000EM: + dev_type = "Intel 82540EM (e1000)"; + break; + + case VNET_DPDK_PMD_IGB: + dev_type = "Intel e1000"; + break; + + case VNET_DPDK_PMD_I40E: + dev_type = "Intel X710/XL710 Family"; + break; + + case VNET_DPDK_PMD_I40EVF: + dev_type = "Intel X710/XL710 Family VF"; + break; + + case VNET_DPDK_PMD_FM10K: + dev_type = "Intel FM10000 Family Ethernet Switch"; + break; + + case VNET_DPDK_PMD_IGBVF: + dev_type = "Intel e1000 VF"; + break; + + case VNET_DPDK_PMD_VIRTIO: + dev_type = "Red Hat Virtio"; + break; + + case VNET_DPDK_PMD_IXGBEVF: + dev_type = "Intel 82599 VF"; + break; + + case VNET_DPDK_PMD_IXGBE: + dev_type = "Intel 82599"; + break; + + case VNET_DPDK_PMD_ENIC: + dev_type = "Cisco VIC"; + break; + + case VNET_DPDK_PMD_CXGBE: + dev_type = "Chelsio T4/T5"; + break; + + case VNET_DPDK_PMD_MLX5: + dev_type = "Mellanox ConnectX-4 Family"; + break; + + case VNET_DPDK_PMD_VMXNET3: + dev_type = "VMware VMXNET3"; + break; + + case VNET_DPDK_PMD_AF_PACKET: + dev_type = "af_packet"; + break; + + case VNET_DPDK_PMD_BOND: + dev_type = "Ethernet Bonding"; + break; + + case VNET_DPDK_PMD_DPAA2: + dev_type = "NXP DPAA2 Mac"; + break; + + default: + case VNET_DPDK_PMD_UNKNOWN: + dev_type = "### UNKNOWN ###"; + break; + } + + return format (s, dev_type); +} + +static u8 * +format_dpdk_link_status (u8 * s, va_list * args) +{ + dpdk_device_t *xd = va_arg (*args, dpdk_device_t *); + struct rte_eth_link *l = &xd->link; + vnet_main_t *vnm = vnet_get_main (); + vnet_hw_interface_t *hi = vnet_get_hw_interface (vnm, xd->vlib_hw_if_index); + + s = format (s, "%s ", l->link_status ? "up" : "down"); + if (l->link_status) + { + u32 promisc = rte_eth_promiscuous_get (xd->device_index); + + s = format (s, "%s duplex ", (l->link_duplex == ETH_LINK_FULL_DUPLEX) ? + "full" : "half"); + s = format (s, "speed %u mtu %d %s\n", l->link_speed, + hi->max_packet_bytes, promisc ? " promisc" : ""); + } + else + s = format (s, "\n"); + + return s; +} + +#define _line_len 72 +#define _(v, str) \ +if (bitmap & v) { \ + if (format_get_indent (s) > next_split ) { \ + next_split += _line_len; \ + s = format(s,"\n%U", format_white_space, indent); \ + } \ + s = format(s, "%s ", str); \ +} + +static u8 * +format_dpdk_rss_hf_name (u8 * s, va_list * args) +{ + u64 bitmap = va_arg (*args, u64); + int next_split = _line_len; + int indent = format_get_indent (s); + + if (!bitmap) + return format (s, "none"); + + foreach_dpdk_rss_hf return s; +} + +static u8 * +format_dpdk_rx_offload_caps (u8 * s, va_list * args) +{ + u32 bitmap = va_arg (*args, u32); + int next_split = _line_len; + int indent = format_get_indent (s); + + if (!bitmap) + return format (s, "none"); + + foreach_dpdk_rx_offload_caps return s; +} + +static u8 * +format_dpdk_tx_offload_caps (u8 * s, va_list * args) +{ + u32 bitmap = va_arg (*args, u32); + int next_split = _line_len; + int indent = format_get_indent (s); + if (!bitmap) + return format (s, "none"); + + foreach_dpdk_tx_offload_caps return s; +} + +#undef _line_len +#undef _ + +u8 * +format_dpdk_device (u8 * s, va_list * args) +{ + u32 dev_instance = va_arg (*args, u32); + int verbose = va_arg (*args, int); + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, dev_instance); + uword indent = format_get_indent (s); + f64 now = vlib_time_now (dm->vlib_main); + struct rte_eth_dev_info di; + + dpdk_update_counters (xd, now); + dpdk_update_link_state (xd, now); + + s = format (s, "%U\n%Ucarrier %U", + format_dpdk_device_type, xd->device_index, + format_white_space, indent + 2, format_dpdk_link_status, xd); + + rte_eth_dev_info_get (xd->device_index, &di); + + if (verbose > 1 && xd->flags & DPDK_DEVICE_FLAG_PMD) + { + struct rte_pci_device *pci; + struct rte_eth_rss_conf rss_conf; + int vlan_off; + int retval; + + rss_conf.rss_key = 0; + retval = rte_eth_dev_rss_hash_conf_get (xd->device_index, &rss_conf); + if (retval < 0) + clib_warning ("rte_eth_dev_rss_hash_conf_get returned %d", retval); + pci = di.pci_dev; + + if (pci) + s = + format (s, + "%Upci id: device %04x:%04x subsystem %04x:%04x\n" + "%Upci address: %04x:%02x:%02x.%02x\n", + format_white_space, indent + 2, pci->id.vendor_id, + pci->id.device_id, pci->id.subsystem_vendor_id, + pci->id.subsystem_device_id, format_white_space, indent + 2, + pci->addr.domain, pci->addr.bus, pci->addr.devid, + pci->addr.function); + s = + format (s, "%Umax rx packet len: %d\n", format_white_space, + indent + 2, di.max_rx_pktlen); + s = + format (s, "%Umax num of queues: rx %d tx %d\n", format_white_space, + indent + 2, di.max_rx_queues, di.max_tx_queues); + s = + format (s, "%Upromiscuous: unicast %s all-multicast %s\n", + format_white_space, indent + 2, + rte_eth_promiscuous_get (xd->device_index) ? "on" : "off", + rte_eth_promiscuous_get (xd->device_index) ? "on" : "off"); + vlan_off = rte_eth_dev_get_vlan_offload (xd->device_index); + s = format (s, "%Uvlan offload: strip %s filter %s qinq %s\n", + format_white_space, indent + 2, + vlan_off & ETH_VLAN_STRIP_OFFLOAD ? "on" : "off", + vlan_off & ETH_VLAN_FILTER_OFFLOAD ? "on" : "off", + vlan_off & ETH_VLAN_EXTEND_OFFLOAD ? "on" : "off"); + s = format (s, "%Urx offload caps: %U\n", + format_white_space, indent + 2, + format_dpdk_rx_offload_caps, di.rx_offload_capa); + s = format (s, "%Utx offload caps: %U\n", + format_white_space, indent + 2, + format_dpdk_tx_offload_caps, di.tx_offload_capa); + s = format (s, "%Urss active: %U\n" + "%Urss supported: %U\n", + format_white_space, indent + 2, + format_dpdk_rss_hf_name, rss_conf.rss_hf, + format_white_space, indent + 2, + format_dpdk_rss_hf_name, di.flow_type_rss_offloads); + } + + s = format (s, "%Urx queues %d, rx desc %d, tx queues %d, tx desc %d\n", + format_white_space, indent + 2, + xd->rx_q_used, xd->nb_rx_desc, xd->tx_q_used, xd->nb_tx_desc); + + if (xd->cpu_socket > -1) + s = format (s, "%Ucpu socket %d\n", + format_white_space, indent + 2, xd->cpu_socket); + + /* $$$ MIB counters */ + { +#define _(N, V) \ + if ((xd->stats.V - xd->last_cleared_stats.V) != 0) { \ + s = format (s, "\n%U%-40U%16Ld", \ + format_white_space, indent + 2, \ + format_c_identifier, #N, \ + xd->stats.V - xd->last_cleared_stats.V); \ + } \ + + foreach_dpdk_counter +#undef _ + } + + u8 *xs = 0; + u32 i = 0; + struct rte_eth_xstat *xstat, *last_xstat; + struct rte_eth_xstat_name *xstat_names = 0; + int len = rte_eth_xstats_get_names (xd->device_index, NULL, 0); + vec_validate (xstat_names, len - 1); + rte_eth_xstats_get_names (xd->device_index, xstat_names, len); + + ASSERT (vec_len (xd->xstats) == vec_len (xd->last_cleared_xstats)); + + /* *INDENT-OFF* */ + vec_foreach_index(i, xd->xstats) + { + u64 delta = 0; + xstat = vec_elt_at_index(xd->xstats, i); + last_xstat = vec_elt_at_index(xd->last_cleared_xstats, i); + + delta = xstat->value - last_xstat->value; + if (verbose == 2 || (verbose && delta)) + { + /* format_c_identifier doesn't like c strings inside vector */ + u8 * name = format(0,"%s", xstat_names[i].name); + xs = format(xs, "\n%U%-38U%16Ld", + format_white_space, indent + 4, + format_c_identifier, name, delta); + vec_free(name); + } + } + /* *INDENT-ON* */ + + vec_free (xstat_names); + + if (xs) + { + s = format (s, "\n%Uextended stats:%v", + format_white_space, indent + 2, xs); + vec_free (xs); + } + + return s; +} + +u8 * +format_dpdk_tx_dma_trace (u8 * s, va_list * va) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*va, vlib_node_t *); + CLIB_UNUSED (vnet_main_t * vnm) = vnet_get_main (); + dpdk_tx_dma_trace_t *t = va_arg (*va, dpdk_tx_dma_trace_t *); + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, t->device_index); + uword indent = format_get_indent (s); + vnet_sw_interface_t *sw = vnet_get_sw_interface (vnm, xd->vlib_sw_if_index); + + s = format (s, "%U tx queue %d", + format_vnet_sw_interface_name, vnm, sw, t->queue_index); + + s = format (s, "\n%Ubuffer 0x%x: %U", + format_white_space, indent, + t->buffer_index, format_vlib_buffer, &t->buffer); + + s = format (s, "\n%U%U", format_white_space, indent, + format_ethernet_header_with_length, t->buffer.pre_data, + sizeof (t->buffer.pre_data)); + + return s; +} + +u8 * +format_dpdk_rx_dma_trace (u8 * s, va_list * va) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*va, vlib_node_t *); + CLIB_UNUSED (vnet_main_t * vnm) = vnet_get_main (); + dpdk_rx_dma_trace_t *t = va_arg (*va, dpdk_rx_dma_trace_t *); + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, t->device_index); + format_function_t *f; + uword indent = format_get_indent (s); + vnet_sw_interface_t *sw = vnet_get_sw_interface (vnm, xd->vlib_sw_if_index); + + s = format (s, "%U rx queue %d", + format_vnet_sw_interface_name, vnm, sw, t->queue_index); + + s = format (s, "\n%Ubuffer 0x%x: %U", + format_white_space, indent, + t->buffer_index, format_vlib_buffer, &t->buffer); + + s = format (s, "\n%U%U", + format_white_space, indent, + format_dpdk_rte_mbuf, &t->mb, &t->data); + + if (vm->trace_main.verbose) + { + s = format (s, "\n%UPacket Dump%s", format_white_space, indent + 2, + t->mb.data_len > sizeof (t->data) ? " (truncated)" : ""); + s = format (s, "\n%U%U", format_white_space, indent + 4, + format_hexdump, &t->data, + t->mb.data_len > + sizeof (t->data) ? sizeof (t->data) : t->mb.data_len); + } + f = node->format_buffer; + if (!f) + f = format_hex_bytes; + s = format (s, "\n%U%U", format_white_space, indent, + f, t->buffer.pre_data, sizeof (t->buffer.pre_data)); + + return s; +} + + +static inline u8 * +format_dpdk_pkt_types (u8 * s, va_list * va) +{ + u32 *pkt_types = va_arg (*va, u32 *); + uword indent __attribute__ ((unused)) = format_get_indent (s) + 2; + + if (!*pkt_types) + return s; + + s = format (s, "Packet Types"); + +#define _(L, F, S) \ + if ((*pkt_types & RTE_PTYPE_##L##_MASK) == RTE_PTYPE_##L##_##F) \ + { \ + s = format (s, "\n%U%s (0x%04x) %s", format_white_space, indent, \ + "RTE_PTYPE_" #L "_" #F, RTE_PTYPE_##L##_##F, S); \ + } + + foreach_dpdk_pkt_type +#undef _ + return s; +} + +static inline u8 * +format_dpdk_pkt_offload_flags (u8 * s, va_list * va) +{ + u64 *ol_flags = va_arg (*va, u64 *); + uword indent = format_get_indent (s) + 2; + + if (!*ol_flags) + return s; + + s = format (s, "Packet Offload Flags"); + +#define _(F, S) \ + if (*ol_flags & F) \ + { \ + s = format (s, "\n%U%s (0x%04x) %s", \ + format_white_space, indent, #F, F, S); \ + } + + foreach_dpdk_pkt_offload_flag +#undef _ + return s; +} + +u8 * +format_dpdk_rte_mbuf_vlan (u8 * s, va_list * va) +{ + ethernet_vlan_header_tv_t *vlan_hdr = + va_arg (*va, ethernet_vlan_header_tv_t *); + + if (clib_net_to_host_u16 (vlan_hdr->type) == ETHERNET_TYPE_DOT1AD) + { + s = format (s, "%U 802.1q vlan ", + format_ethernet_vlan_tci, + clib_net_to_host_u16 (vlan_hdr->priority_cfi_and_id)); + vlan_hdr++; + } + + s = format (s, "%U", + format_ethernet_vlan_tci, + clib_net_to_host_u16 (vlan_hdr->priority_cfi_and_id)); + + return s; +} + +u8 * +format_dpdk_rte_mbuf (u8 * s, va_list * va) +{ + struct rte_mbuf *mb = va_arg (*va, struct rte_mbuf *); + ethernet_header_t *eth_hdr = va_arg (*va, ethernet_header_t *); + uword indent = format_get_indent (s) + 2; + + s = format (s, "PKT MBUF: port %d, nb_segs %d, pkt_len %d" + "\n%Ubuf_len %d, data_len %d, ol_flags 0x%x, data_off %d, phys_addr 0x%x" + "\n%Upacket_type 0x%x", + mb->port, mb->nb_segs, mb->pkt_len, + format_white_space, indent, + mb->buf_len, mb->data_len, mb->ol_flags, mb->data_off, + mb->buf_physaddr, format_white_space, indent, mb->packet_type); + + if (mb->ol_flags) + s = format (s, "\n%U%U", format_white_space, indent, + format_dpdk_pkt_offload_flags, &mb->ol_flags); + + if ((mb->ol_flags & PKT_RX_VLAN_PKT) && + ((mb->ol_flags & (PKT_RX_VLAN_STRIPPED | PKT_RX_QINQ_STRIPPED)) == 0)) + { + ethernet_vlan_header_tv_t *vlan_hdr = + ((ethernet_vlan_header_tv_t *) & (eth_hdr->type)); + s = format (s, " %U", format_dpdk_rte_mbuf_vlan, vlan_hdr); + } + + if (mb->packet_type) + s = format (s, "\n%U%U", format_white_space, indent, + format_dpdk_pkt_types, &mb->packet_type); + + return s; +} + +uword +unformat_socket_mem (unformat_input_t * input, va_list * va) +{ + uword **r = va_arg (*va, uword **); + int i = 0; + u32 mem; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, ",")) + hash_set (*r, i, 1024); + else if (unformat (input, "%u,", &mem)) + hash_set (*r, i, mem); + else if (unformat (input, "%u", &mem)) + hash_set (*r, i, mem); + else + { + unformat_put_input (input); + goto done; + } + i++; + } + +done: + return 1; +} + +clib_error_t * +unformat_rss_fn (unformat_input_t * input, uword * rss_fn) +{ + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (0) + ; +#undef _ +#define _(f, s) \ + else if (unformat (input, s)) \ + *rss_fn |= f; + + foreach_dpdk_rss_hf +#undef _ + else + { + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + } + return 0; +} + +clib_error_t * +unformat_hqos (unformat_input_t * input, dpdk_device_config_hqos_t * hqos) +{ + clib_error_t *error = 0; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "hqos-thread %u", &hqos->hqos_thread)) + hqos->hqos_thread_valid = 1; + else + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + break; + } + } + + return error; +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/hqos.c b/src/vnet/devices/dpdk/hqos.c new file mode 100644 index 00000000000..d68bc48f80b --- /dev/null +++ b/src/vnet/devices/dpdk/hqos.c @@ -0,0 +1,775 @@ +/* + * Copyright(c) 2016 Intel Corporation. All rights reserved. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include + +#include +#include +#include + +#include +#include +#include +#include /* enumerate all vlib messages */ + +#define vl_typedefs /* define message structures */ +#include +#undef vl_typedefs + +/* instantiate all the print functions we know about */ +#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__) +#define vl_printfun +#include +#undef vl_printfun + +#include "dpdk_priv.h" + +dpdk_main_t dpdk_main; + +/*** + * + * HQoS default configuration values + * + ***/ + +static dpdk_device_config_hqos_t hqos_params_default = { + .hqos_thread_valid = 0, + + .swq_size = 4096, + .burst_enq = 256, + .burst_deq = 220, + + /* + * Packet field to identify the subport. + * + * Default value: Since only one subport is defined by default (see below: + * n_subports_per_port = 1), the subport ID is hardcoded to 0. + */ + .pktfield0_slabpos = 0, + .pktfield0_slabmask = 0, + + /* + * Packet field to identify the pipe. + * + * Default value: Assuming Ethernet/IPv4/UDP packets, UDP payload bits 12 .. 23 + */ + .pktfield1_slabpos = 40, + .pktfield1_slabmask = 0x0000000FFF000000LLU, + + /* Packet field used as index into TC translation table to identify the traffic + * class and queue. + * + * Default value: Assuming Ethernet/IPv4 packets, IPv4 DSCP field + */ + .pktfield2_slabpos = 8, + .pktfield2_slabmask = 0x00000000000000FCLLU, + .tc_table = { + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + }, + + /* port */ + .port = { + .name = NULL, /* Set at init */ + .socket = 0, /* Set at init */ + .rate = 1250000000, /* Assuming 10GbE port */ + .mtu = 14 + 1500, /* Assuming Ethernet/IPv4 pkt (Ethernet FCS not included) */ + .frame_overhead = RTE_SCHED_FRAME_OVERHEAD_DEFAULT, + .n_subports_per_port = 1, + .n_pipes_per_subport = 4096, + .qsize = {64, 64, 64, 64}, + .pipe_profiles = NULL, /* Set at config */ + .n_pipe_profiles = 1, + +#ifdef RTE_SCHED_RED + .red_params = { + /* Traffic Class 0 Colors Green / Yellow / Red */ + [0][0] = {.min_th = 48,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [0][1] = {.min_th = 40,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [0][2] = {.min_th = 32,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + + /* Traffic Class 1 - Colors Green / Yellow / Red */ + [1][0] = {.min_th = 48,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [1][1] = {.min_th = 40,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [1][2] = {.min_th = 32,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + + /* Traffic Class 2 - Colors Green / Yellow / Red */ + [2][0] = {.min_th = 48,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [2][1] = {.min_th = 40,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [2][2] = {.min_th = 32,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + + /* Traffic Class 3 - Colors Green / Yellow / Red */ + [3][0] = {.min_th = 48,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [3][1] = {.min_th = 40,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [3][2] = {.min_th = 32,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9} + }, +#endif /* RTE_SCHED_RED */ + }, +}; + +static struct rte_sched_subport_params hqos_subport_params_default = { + .tb_rate = 1250000000, /* 10GbE line rate (measured in bytes/second) */ + .tb_size = 1000000, + .tc_rate = {1250000000, 1250000000, 1250000000, 1250000000}, + .tc_period = 10, +}; + +static struct rte_sched_pipe_params hqos_pipe_params_default = { + .tb_rate = 305175, /* 10GbE line rate divided by 4K pipes */ + .tb_size = 1000000, + .tc_rate = {305175, 305175, 305175, 305175}, + .tc_period = 40, +#ifdef RTE_SCHED_SUBPORT_TC_OV + .tc_ov_weight = 1, +#endif + .wrr_weights = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, +}; + +/*** + * + * HQoS configuration + * + ***/ + +int +dpdk_hqos_validate_mask (u64 mask, u32 n) +{ + int count = __builtin_popcountll (mask); + int pos_lead = sizeof (u64) * 8 - __builtin_clzll (mask); + int pos_trail = __builtin_ctzll (mask); + int count_expected = __builtin_popcount (n - 1); + + /* Handle the exceptions */ + if (n == 0) + return -1; /* Error */ + + if ((mask == 0) && (n == 1)) + return 0; /* OK */ + + if (((mask == 0) && (n != 1)) || ((mask != 0) && (n == 1))) + return -2; /* Error */ + + /* Check that mask is contiguous */ + if ((pos_lead - pos_trail) != count) + return -3; /* Error */ + + /* Check that mask contains the expected number of bits set */ + if (count != count_expected) + return -4; /* Error */ + + return 0; /* OK */ +} + +void +dpdk_device_config_hqos_pipe_profile_default (dpdk_device_config_hqos_t * + hqos, u32 pipe_profile_id) +{ + memcpy (&hqos->pipe[pipe_profile_id], &hqos_pipe_params_default, + sizeof (hqos_pipe_params_default)); +} + +void +dpdk_device_config_hqos_default (dpdk_device_config_hqos_t * hqos) +{ + struct rte_sched_subport_params *subport_params; + struct rte_sched_pipe_params *pipe_params; + u32 *pipe_map; + u32 i; + + memcpy (hqos, &hqos_params_default, sizeof (hqos_params_default)); + + /* pipe */ + vec_add2 (hqos->pipe, pipe_params, hqos->port.n_pipe_profiles); + + for (i = 0; i < vec_len (hqos->pipe); i++) + memcpy (&pipe_params[i], + &hqos_pipe_params_default, sizeof (hqos_pipe_params_default)); + + hqos->port.pipe_profiles = hqos->pipe; + + /* subport */ + vec_add2 (hqos->subport, subport_params, hqos->port.n_subports_per_port); + + for (i = 0; i < vec_len (hqos->subport); i++) + memcpy (&subport_params[i], + &hqos_subport_params_default, + sizeof (hqos_subport_params_default)); + + /* pipe profile */ + vec_add2 (hqos->pipe_map, + pipe_map, + hqos->port.n_subports_per_port * hqos->port.n_pipes_per_subport); + + for (i = 0; i < vec_len (hqos->pipe_map); i++) + pipe_map[i] = 0; +} + +/*** + * + * HQoS init + * + ***/ + +clib_error_t * +dpdk_port_setup_hqos (dpdk_device_t * xd, dpdk_device_config_hqos_t * hqos) +{ + vlib_thread_main_t *tm = vlib_get_thread_main (); + char name[32]; + u32 subport_id, i; + int rv; + + /* Detect the set of worker threads */ + int worker_thread_first = 0; + int worker_thread_count = 0; + + uword *p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + vlib_thread_registration_t *tr = + p ? (vlib_thread_registration_t *) p[0] : 0; + + if (tr && tr->count > 0) + { + worker_thread_first = tr->first_index; + worker_thread_count = tr->count; + } + + /* Allocate the per-thread device data array */ + vec_validate_aligned (xd->hqos_wt, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + memset (xd->hqos_wt, 0, tm->n_vlib_mains * sizeof (xd->hqos_wt[0])); + + vec_validate_aligned (xd->hqos_ht, 0, CLIB_CACHE_LINE_BYTES); + memset (xd->hqos_ht, 0, sizeof (xd->hqos_ht[0])); + + /* Allocate space for one SWQ per worker thread in the I/O TX thread data structure */ + vec_validate (xd->hqos_ht->swq, worker_thread_count); + + /* SWQ */ + for (i = 0; i < worker_thread_count + 1; i++) + { + u32 swq_flags = RING_F_SP_ENQ | RING_F_SC_DEQ; + + snprintf (name, sizeof (name), "SWQ-worker%u-to-device%u", i, + xd->device_index); + xd->hqos_ht->swq[i] = + rte_ring_create (name, hqos->swq_size, xd->cpu_socket, swq_flags); + if (xd->hqos_ht->swq[i] == NULL) + return clib_error_return (0, + "SWQ-worker%u-to-device%u: rte_ring_create err", + i, xd->device_index); + } + + /* + * HQoS + */ + + /* HQoS port */ + snprintf (name, sizeof (name), "HQoS%u", xd->device_index); + hqos->port.name = strdup (name); + if (hqos->port.name == NULL) + return clib_error_return (0, "HQoS%u: strdup err", xd->device_index); + + hqos->port.socket = rte_eth_dev_socket_id (xd->device_index); + if (hqos->port.socket == SOCKET_ID_ANY) + hqos->port.socket = 0; + + xd->hqos_ht->hqos = rte_sched_port_config (&hqos->port); + if (xd->hqos_ht->hqos == NULL) + return clib_error_return (0, "HQoS%u: rte_sched_port_config err", + xd->device_index); + + /* HQoS subport */ + for (subport_id = 0; subport_id < hqos->port.n_subports_per_port; + subport_id++) + { + u32 pipe_id; + + rv = + rte_sched_subport_config (xd->hqos_ht->hqos, subport_id, + &hqos->subport[subport_id]); + if (rv) + return clib_error_return (0, + "HQoS%u subport %u: rte_sched_subport_config err (%d)", + xd->device_index, subport_id, rv); + + /* HQoS pipe */ + for (pipe_id = 0; pipe_id < hqos->port.n_pipes_per_subport; pipe_id++) + { + u32 pos = subport_id * hqos->port.n_pipes_per_subport + pipe_id; + u32 profile_id = hqos->pipe_map[pos]; + + rv = + rte_sched_pipe_config (xd->hqos_ht->hqos, subport_id, pipe_id, + profile_id); + if (rv) + return clib_error_return (0, + "HQoS%u subport %u pipe %u: rte_sched_pipe_config err (%d)", + xd->device_index, subport_id, pipe_id, + rv); + } + } + + /* Set up per-thread device data for the I/O TX thread */ + xd->hqos_ht->hqos_burst_enq = hqos->burst_enq; + xd->hqos_ht->hqos_burst_deq = hqos->burst_deq; + vec_validate (xd->hqos_ht->pkts_enq, 2 * hqos->burst_enq - 1); + vec_validate (xd->hqos_ht->pkts_deq, hqos->burst_deq - 1); + xd->hqos_ht->pkts_enq_len = 0; + xd->hqos_ht->swq_pos = 0; + xd->hqos_ht->flush_count = 0; + + /* Set up per-thread device data for each worker thread */ + for (i = 0; i < worker_thread_count + 1; i++) + { + u32 tid; + if (i) + tid = worker_thread_first + (i - 1); + else + tid = i; + + xd->hqos_wt[tid].swq = xd->hqos_ht->swq[i]; + xd->hqos_wt[tid].hqos_field0_slabpos = hqos->pktfield0_slabpos; + xd->hqos_wt[tid].hqos_field0_slabmask = hqos->pktfield0_slabmask; + xd->hqos_wt[tid].hqos_field0_slabshr = + __builtin_ctzll (hqos->pktfield0_slabmask); + xd->hqos_wt[tid].hqos_field1_slabpos = hqos->pktfield1_slabpos; + xd->hqos_wt[tid].hqos_field1_slabmask = hqos->pktfield1_slabmask; + xd->hqos_wt[tid].hqos_field1_slabshr = + __builtin_ctzll (hqos->pktfield1_slabmask); + xd->hqos_wt[tid].hqos_field2_slabpos = hqos->pktfield2_slabpos; + xd->hqos_wt[tid].hqos_field2_slabmask = hqos->pktfield2_slabmask; + xd->hqos_wt[tid].hqos_field2_slabshr = + __builtin_ctzll (hqos->pktfield2_slabmask); + memcpy (xd->hqos_wt[tid].hqos_tc_table, hqos->tc_table, + sizeof (hqos->tc_table)); + } + + return 0; +} + +/*** + * + * HQoS run-time + * + ***/ +/* + * dpdk_hqos_thread - Contains the main loop of an HQoS thread. + * + * w + * Information for the current thread + */ +static_always_inline void +dpdk_hqos_thread_internal_hqos_dbg_bypass (vlib_main_t * vm) +{ + dpdk_main_t *dm = &dpdk_main; + u32 cpu_index = vm->cpu_index; + u32 dev_pos; + + dev_pos = 0; + while (1) + { + vlib_worker_thread_barrier_check (); + + u32 n_devs = vec_len (dm->devices_by_hqos_cpu[cpu_index]); + if (dev_pos >= n_devs) + dev_pos = 0; + + dpdk_device_and_queue_t *dq = + vec_elt_at_index (dm->devices_by_hqos_cpu[cpu_index], dev_pos); + dpdk_device_t *xd = vec_elt_at_index (dm->devices, dq->device); + + dpdk_device_hqos_per_hqos_thread_t *hqos = xd->hqos_ht; + u32 device_index = xd->device_index; + u16 queue_id = dq->queue_id; + + struct rte_mbuf **pkts_enq = hqos->pkts_enq; + u32 pkts_enq_len = hqos->pkts_enq_len; + u32 swq_pos = hqos->swq_pos; + u32 n_swq = vec_len (hqos->swq), i; + u32 flush_count = hqos->flush_count; + + for (i = 0; i < n_swq; i++) + { + /* Get current SWQ for this device */ + struct rte_ring *swq = hqos->swq[swq_pos]; + + /* Read SWQ burst to packet buffer of this device */ + pkts_enq_len += rte_ring_sc_dequeue_burst (swq, + (void **) + &pkts_enq[pkts_enq_len], + hqos->hqos_burst_enq); + + /* Get next SWQ for this device */ + swq_pos++; + if (swq_pos >= n_swq) + swq_pos = 0; + hqos->swq_pos = swq_pos; + + /* HWQ TX enqueue when burst available */ + if (pkts_enq_len >= hqos->hqos_burst_enq) + { + u32 n_pkts = rte_eth_tx_burst (device_index, + (uint16_t) queue_id, + pkts_enq, + (uint16_t) pkts_enq_len); + + for (; n_pkts < pkts_enq_len; n_pkts++) + rte_pktmbuf_free (pkts_enq[n_pkts]); + + pkts_enq_len = 0; + flush_count = 0; + break; + } + } + if (pkts_enq_len) + { + flush_count++; + if (PREDICT_FALSE (flush_count == HQOS_FLUSH_COUNT_THRESHOLD)) + { + rte_sched_port_enqueue (hqos->hqos, pkts_enq, pkts_enq_len); + + pkts_enq_len = 0; + flush_count = 0; + } + } + hqos->pkts_enq_len = pkts_enq_len; + hqos->flush_count = flush_count; + + /* Advance to next device */ + dev_pos++; + } +} + +static_always_inline void +dpdk_hqos_thread_internal (vlib_main_t * vm) +{ + dpdk_main_t *dm = &dpdk_main; + u32 cpu_index = vm->cpu_index; + u32 dev_pos; + + dev_pos = 0; + while (1) + { + vlib_worker_thread_barrier_check (); + + u32 n_devs = vec_len (dm->devices_by_hqos_cpu[cpu_index]); + if (PREDICT_FALSE (n_devs == 0)) + { + dev_pos = 0; + continue; + } + if (dev_pos >= n_devs) + dev_pos = 0; + + dpdk_device_and_queue_t *dq = + vec_elt_at_index (dm->devices_by_hqos_cpu[cpu_index], dev_pos); + dpdk_device_t *xd = vec_elt_at_index (dm->devices, dq->device); + + dpdk_device_hqos_per_hqos_thread_t *hqos = xd->hqos_ht; + u32 device_index = xd->device_index; + u16 queue_id = dq->queue_id; + + struct rte_mbuf **pkts_enq = hqos->pkts_enq; + struct rte_mbuf **pkts_deq = hqos->pkts_deq; + u32 pkts_enq_len = hqos->pkts_enq_len; + u32 swq_pos = hqos->swq_pos; + u32 n_swq = vec_len (hqos->swq), i; + u32 flush_count = hqos->flush_count; + + /* + * SWQ dequeue and HQoS enqueue for current device + */ + for (i = 0; i < n_swq; i++) + { + /* Get current SWQ for this device */ + struct rte_ring *swq = hqos->swq[swq_pos]; + + /* Read SWQ burst to packet buffer of this device */ + pkts_enq_len += rte_ring_sc_dequeue_burst (swq, + (void **) + &pkts_enq[pkts_enq_len], + hqos->hqos_burst_enq); + + /* Get next SWQ for this device */ + swq_pos++; + if (swq_pos >= n_swq) + swq_pos = 0; + hqos->swq_pos = swq_pos; + + /* HQoS enqueue when burst available */ + if (pkts_enq_len >= hqos->hqos_burst_enq) + { + rte_sched_port_enqueue (hqos->hqos, pkts_enq, pkts_enq_len); + + pkts_enq_len = 0; + flush_count = 0; + break; + } + } + if (pkts_enq_len) + { + flush_count++; + if (PREDICT_FALSE (flush_count == HQOS_FLUSH_COUNT_THRESHOLD)) + { + rte_sched_port_enqueue (hqos->hqos, pkts_enq, pkts_enq_len); + + pkts_enq_len = 0; + flush_count = 0; + } + } + hqos->pkts_enq_len = pkts_enq_len; + hqos->flush_count = flush_count; + + /* + * HQoS dequeue and HWQ TX enqueue for current device + */ + { + u32 pkts_deq_len, n_pkts; + + pkts_deq_len = rte_sched_port_dequeue (hqos->hqos, + pkts_deq, + hqos->hqos_burst_deq); + + for (n_pkts = 0; n_pkts < pkts_deq_len;) + n_pkts += rte_eth_tx_burst (device_index, + (uint16_t) queue_id, + &pkts_deq[n_pkts], + (uint16_t) (pkts_deq_len - n_pkts)); + } + + /* Advance to next device */ + dev_pos++; + } +} + +void +dpdk_hqos_thread (vlib_worker_thread_t * w) +{ + vlib_main_t *vm; + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + + vm = vlib_get_main (); + + ASSERT (vm->cpu_index == os_get_cpu_number ()); + + clib_time_init (&vm->clib_time); + clib_mem_set_heap (w->thread_mheap); + + /* Wait until the dpdk init sequence is complete */ + while (tm->worker_thread_release == 0) + vlib_worker_thread_barrier_check (); + + if (vec_len (dm->devices_by_hqos_cpu[vm->cpu_index]) == 0) + return + clib_error + ("current I/O TX thread does not have any devices assigned to it"); + + if (DPDK_HQOS_DBG_BYPASS) + dpdk_hqos_thread_internal_hqos_dbg_bypass (vm); + else + dpdk_hqos_thread_internal (vm); +} + +void +dpdk_hqos_thread_fn (void *arg) +{ + vlib_worker_thread_t *w = (vlib_worker_thread_t *) arg; + vlib_worker_thread_init (w); + dpdk_hqos_thread (w); +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_THREAD (hqos_thread_reg, static) = +{ + .name = "hqos-threads", + .short_name = "hqos-threads", + .function = dpdk_hqos_thread_fn, +}; +/* *INDENT-ON* */ + +/* + * HQoS run-time code to be called by the worker threads + */ +#define BITFIELD(byte_array, slab_pos, slab_mask, slab_shr) \ +({ \ + u64 slab = *((u64 *) &byte_array[slab_pos]); \ + u64 val = (rte_be_to_cpu_64(slab) & slab_mask) >> slab_shr; \ + val; \ +}) + +#define RTE_SCHED_PORT_HIERARCHY(subport, pipe, traffic_class, queue, color) \ + ((((u64) (queue)) & 0x3) | \ + ((((u64) (traffic_class)) & 0x3) << 2) | \ + ((((u64) (color)) & 0x3) << 4) | \ + ((((u64) (subport)) & 0xFFFF) << 16) | \ + ((((u64) (pipe)) & 0xFFFFFFFF) << 32)) + +void +dpdk_hqos_metadata_set (dpdk_device_hqos_per_worker_thread_t * hqos, + struct rte_mbuf **pkts, u32 n_pkts) +{ + u32 i; + + for (i = 0; i < (n_pkts & (~0x3)); i += 4) + { + struct rte_mbuf *pkt0 = pkts[i]; + struct rte_mbuf *pkt1 = pkts[i + 1]; + struct rte_mbuf *pkt2 = pkts[i + 2]; + struct rte_mbuf *pkt3 = pkts[i + 3]; + + u8 *pkt0_data = rte_pktmbuf_mtod (pkt0, u8 *); + u8 *pkt1_data = rte_pktmbuf_mtod (pkt1, u8 *); + u8 *pkt2_data = rte_pktmbuf_mtod (pkt2, u8 *); + u8 *pkt3_data = rte_pktmbuf_mtod (pkt3, u8 *); + + u64 pkt0_subport = BITFIELD (pkt0_data, hqos->hqos_field0_slabpos, + hqos->hqos_field0_slabmask, + hqos->hqos_field0_slabshr); + u64 pkt0_pipe = BITFIELD (pkt0_data, hqos->hqos_field1_slabpos, + hqos->hqos_field1_slabmask, + hqos->hqos_field1_slabshr); + u64 pkt0_dscp = BITFIELD (pkt0_data, hqos->hqos_field2_slabpos, + hqos->hqos_field2_slabmask, + hqos->hqos_field2_slabshr); + u32 pkt0_tc = hqos->hqos_tc_table[pkt0_dscp & 0x3F] >> 2; + u32 pkt0_tc_q = hqos->hqos_tc_table[pkt0_dscp & 0x3F] & 0x3; + + u64 pkt1_subport = BITFIELD (pkt1_data, hqos->hqos_field0_slabpos, + hqos->hqos_field0_slabmask, + hqos->hqos_field0_slabshr); + u64 pkt1_pipe = BITFIELD (pkt1_data, hqos->hqos_field1_slabpos, + hqos->hqos_field1_slabmask, + hqos->hqos_field1_slabshr); + u64 pkt1_dscp = BITFIELD (pkt1_data, hqos->hqos_field2_slabpos, + hqos->hqos_field2_slabmask, + hqos->hqos_field2_slabshr); + u32 pkt1_tc = hqos->hqos_tc_table[pkt1_dscp & 0x3F] >> 2; + u32 pkt1_tc_q = hqos->hqos_tc_table[pkt1_dscp & 0x3F] & 0x3; + + u64 pkt2_subport = BITFIELD (pkt2_data, hqos->hqos_field0_slabpos, + hqos->hqos_field0_slabmask, + hqos->hqos_field0_slabshr); + u64 pkt2_pipe = BITFIELD (pkt2_data, hqos->hqos_field1_slabpos, + hqos->hqos_field1_slabmask, + hqos->hqos_field1_slabshr); + u64 pkt2_dscp = BITFIELD (pkt2_data, hqos->hqos_field2_slabpos, + hqos->hqos_field2_slabmask, + hqos->hqos_field2_slabshr); + u32 pkt2_tc = hqos->hqos_tc_table[pkt2_dscp & 0x3F] >> 2; + u32 pkt2_tc_q = hqos->hqos_tc_table[pkt2_dscp & 0x3F] & 0x3; + + u64 pkt3_subport = BITFIELD (pkt3_data, hqos->hqos_field0_slabpos, + hqos->hqos_field0_slabmask, + hqos->hqos_field0_slabshr); + u64 pkt3_pipe = BITFIELD (pkt3_data, hqos->hqos_field1_slabpos, + hqos->hqos_field1_slabmask, + hqos->hqos_field1_slabshr); + u64 pkt3_dscp = BITFIELD (pkt3_data, hqos->hqos_field2_slabpos, + hqos->hqos_field2_slabmask, + hqos->hqos_field2_slabshr); + u32 pkt3_tc = hqos->hqos_tc_table[pkt3_dscp & 0x3F] >> 2; + u32 pkt3_tc_q = hqos->hqos_tc_table[pkt3_dscp & 0x3F] & 0x3; + + u64 pkt0_sched = RTE_SCHED_PORT_HIERARCHY (pkt0_subport, + pkt0_pipe, + pkt0_tc, + pkt0_tc_q, + 0); + u64 pkt1_sched = RTE_SCHED_PORT_HIERARCHY (pkt1_subport, + pkt1_pipe, + pkt1_tc, + pkt1_tc_q, + 0); + u64 pkt2_sched = RTE_SCHED_PORT_HIERARCHY (pkt2_subport, + pkt2_pipe, + pkt2_tc, + pkt2_tc_q, + 0); + u64 pkt3_sched = RTE_SCHED_PORT_HIERARCHY (pkt3_subport, + pkt3_pipe, + pkt3_tc, + pkt3_tc_q, + 0); + + pkt0->hash.sched.lo = pkt0_sched & 0xFFFFFFFF; + pkt0->hash.sched.hi = pkt0_sched >> 32; + pkt1->hash.sched.lo = pkt1_sched & 0xFFFFFFFF; + pkt1->hash.sched.hi = pkt1_sched >> 32; + pkt2->hash.sched.lo = pkt2_sched & 0xFFFFFFFF; + pkt2->hash.sched.hi = pkt2_sched >> 32; + pkt3->hash.sched.lo = pkt3_sched & 0xFFFFFFFF; + pkt3->hash.sched.hi = pkt3_sched >> 32; + } + + for (; i < n_pkts; i++) + { + struct rte_mbuf *pkt = pkts[i]; + + u8 *pkt_data = rte_pktmbuf_mtod (pkt, u8 *); + + u64 pkt_subport = BITFIELD (pkt_data, hqos->hqos_field0_slabpos, + hqos->hqos_field0_slabmask, + hqos->hqos_field0_slabshr); + u64 pkt_pipe = BITFIELD (pkt_data, hqos->hqos_field1_slabpos, + hqos->hqos_field1_slabmask, + hqos->hqos_field1_slabshr); + u64 pkt_dscp = BITFIELD (pkt_data, hqos->hqos_field2_slabpos, + hqos->hqos_field2_slabmask, + hqos->hqos_field2_slabshr); + u32 pkt_tc = hqos->hqos_tc_table[pkt_dscp & 0x3F] >> 2; + u32 pkt_tc_q = hqos->hqos_tc_table[pkt_dscp & 0x3F] & 0x3; + + u64 pkt_sched = RTE_SCHED_PORT_HIERARCHY (pkt_subport, + pkt_pipe, + pkt_tc, + pkt_tc_q, + 0); + + pkt->hash.sched.lo = pkt_sched & 0xFFFFFFFF; + pkt->hash.sched.hi = pkt_sched >> 32; + } +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/init.c b/src/vnet/devices/dpdk/init.c new file mode 100755 index 00000000000..693ca985130 --- /dev/null +++ b/src/vnet/devices/dpdk/init.c @@ -0,0 +1,1803 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include +#include +#include +#include +#include + +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include + +#include "dpdk_priv.h" + +dpdk_main_t dpdk_main; + +/* force linker to link functions used by vlib and declared weak */ +void *vlib_weakly_linked_functions[] = { + &rte_pktmbuf_init, + &rte_pktmbuf_pool_init, +}; + +#define LINK_STATE_ELOGS 0 + +#define DEFAULT_HUGE_DIR "/run/vpp/hugepages" +#define VPP_RUN_DIR "/run/vpp" + +/* Port configuration, mildly modified Intel app values */ + +static struct rte_eth_conf port_conf_template = { + .rxmode = { + .split_hdr_size = 0, + .header_split = 0, /**< Header Split disabled */ + .hw_ip_checksum = 0, /**< IP checksum offload disabled */ + .hw_vlan_filter = 0, /**< VLAN filtering disabled */ + .hw_strip_crc = 0, /**< CRC stripped by hardware */ + }, + .txmode = { + .mq_mode = ETH_MQ_TX_NONE, + }, +}; + +clib_error_t * +dpdk_port_setup (dpdk_main_t * dm, dpdk_device_t * xd) +{ + vlib_main_t *vm = vlib_get_main (); + vlib_buffer_main_t *bm = vm->buffer_main; + int rv; + int j; + + ASSERT (os_get_cpu_number () == 0); + + if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) + { + vnet_hw_interface_set_flags (dm->vnet_main, xd->vlib_hw_if_index, 0); + rte_eth_dev_stop (xd->device_index); + } + + rv = rte_eth_dev_configure (xd->device_index, xd->rx_q_used, + xd->tx_q_used, &xd->port_conf); + + if (rv < 0) + return clib_error_return (0, "rte_eth_dev_configure[%d]: err %d", + xd->device_index, rv); + + /* Set up one TX-queue per worker thread */ + for (j = 0; j < xd->tx_q_used; j++) + { + rv = rte_eth_tx_queue_setup (xd->device_index, j, xd->nb_tx_desc, + xd->cpu_socket, &xd->tx_conf); + + /* retry with any other CPU socket */ + if (rv < 0) + rv = rte_eth_tx_queue_setup (xd->device_index, j, xd->nb_tx_desc, + SOCKET_ID_ANY, &xd->tx_conf); + if (rv < 0) + break; + } + + if (rv < 0) + return clib_error_return (0, "rte_eth_tx_queue_setup[%d]: err %d", + xd->device_index, rv); + + for (j = 0; j < xd->rx_q_used; j++) + { + + rv = rte_eth_rx_queue_setup (xd->device_index, j, xd->nb_rx_desc, + xd->cpu_socket, 0, + bm-> + pktmbuf_pools[xd->cpu_socket_id_by_queue + [j]]); + + /* retry with any other CPU socket */ + if (rv < 0) + rv = rte_eth_rx_queue_setup (xd->device_index, j, xd->nb_rx_desc, + SOCKET_ID_ANY, 0, + bm-> + pktmbuf_pools[xd->cpu_socket_id_by_queue + [j]]); + if (rv < 0) + return clib_error_return (0, "rte_eth_rx_queue_setup[%d]: err %d", + xd->device_index, rv); + } + + if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) + { + int rv; + rv = rte_eth_dev_start (xd->device_index); + if (rv < 0) + clib_warning ("rte_eth_dev_start %d returned %d", + xd->device_index, rv); + } + return 0; +} + +static u32 +dpdk_flag_change (vnet_main_t * vnm, vnet_hw_interface_t * hi, u32 flags) +{ + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, hi->dev_instance); + u32 old = 0; + + if (ETHERNET_INTERFACE_FLAG_CONFIG_PROMISC (flags)) + { + old = (xd->flags & DPDK_DEVICE_FLAG_PROMISC) != 0; + + if (flags & ETHERNET_INTERFACE_FLAG_ACCEPT_ALL) + xd->flags |= DPDK_DEVICE_FLAG_PROMISC; + else + xd->flags &= ~DPDK_DEVICE_FLAG_PROMISC; + + if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) + { + if (xd->flags & DPDK_DEVICE_FLAG_PROMISC) + rte_eth_promiscuous_enable (xd->device_index); + else + rte_eth_promiscuous_disable (xd->device_index); + } + } + else if (ETHERNET_INTERFACE_FLAG_CONFIG_MTU (flags)) + { + /* + * DAW-FIXME: The Cisco VIC firmware does not provide an api for a + * driver to dynamically change the mtu. If/when the + * VIC firmware gets fixed, then this should be removed. + */ + if (xd->pmd == VNET_DPDK_PMD_ENIC) + { + struct rte_eth_dev_info dev_info; + + /* + * Restore mtu to what has been set by CIMC in the firmware cfg. + */ + rte_eth_dev_info_get (xd->device_index, &dev_info); + hi->max_packet_bytes = dev_info.max_rx_pktlen; + + vlib_cli_output (vlib_get_main (), + "Cisco VIC mtu can only be changed " + "using CIMC then rebooting the server!"); + } + else + { + int rv; + + xd->port_conf.rxmode.max_rx_pkt_len = hi->max_packet_bytes; + + if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) + rte_eth_dev_stop (xd->device_index); + + rv = rte_eth_dev_configure + (xd->device_index, xd->rx_q_used, xd->tx_q_used, &xd->port_conf); + + if (rv < 0) + vlib_cli_output (vlib_get_main (), + "rte_eth_dev_configure[%d]: err %d", + xd->device_index, rv); + + rte_eth_dev_set_mtu (xd->device_index, hi->max_packet_bytes); + + if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) + { + int rv = rte_eth_dev_start (xd->device_index); + if (rv < 0) + clib_warning ("rte_eth_dev_start %d returned %d", + xd->device_index, rv); + } + } + } + return old; +} + +void +dpdk_device_lock_init (dpdk_device_t * xd) +{ + int q; + vec_validate (xd->lockp, xd->tx_q_used - 1); + for (q = 0; q < xd->tx_q_used; q++) + { + xd->lockp[q] = clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES, + CLIB_CACHE_LINE_BYTES); + memset ((void *) xd->lockp[q], 0, CLIB_CACHE_LINE_BYTES); + } +} + +void +dpdk_device_lock_free (dpdk_device_t * xd) +{ + int q; + + for (q = 0; q < vec_len (xd->lockp); q++) + clib_mem_free ((void *) xd->lockp[q]); + vec_free (xd->lockp); + xd->lockp = 0; +} + +static clib_error_t * +dpdk_lib_init (dpdk_main_t * dm) +{ + u32 nports; + u32 nb_desc = 0; + int i; + clib_error_t *error; + vlib_main_t *vm = vlib_get_main (); + vlib_thread_main_t *tm = vlib_get_thread_main (); + vnet_sw_interface_t *sw; + vnet_hw_interface_t *hi; + dpdk_device_t *xd; + vlib_pci_addr_t last_pci_addr; + u32 last_pci_addr_port = 0; + vlib_thread_registration_t *tr, *tr_hqos; + uword *p, *p_hqos; + + u32 next_cpu = 0, next_hqos_cpu = 0; + u8 af_packet_port_id = 0; + last_pci_addr.as_u32 = ~0; + + dm->input_cpu_first_index = 0; + dm->input_cpu_count = 1; + + /* find out which cpus will be used for input */ + p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + tr = p ? (vlib_thread_registration_t *) p[0] : 0; + + if (tr && tr->count > 0) + { + dm->input_cpu_first_index = tr->first_index; + dm->input_cpu_count = tr->count; + } + + vec_validate_aligned (dm->devices_by_cpu, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + + vec_validate_aligned (dm->workers, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + + dm->hqos_cpu_first_index = 0; + dm->hqos_cpu_count = 0; + + /* find out which cpus will be used for I/O TX */ + p_hqos = hash_get_mem (tm->thread_registrations_by_name, "hqos-threads"); + tr_hqos = p_hqos ? (vlib_thread_registration_t *) p_hqos[0] : 0; + + if (tr_hqos && tr_hqos->count > 0) + { + dm->hqos_cpu_first_index = tr_hqos->first_index; + dm->hqos_cpu_count = tr_hqos->count; + } + + vec_validate_aligned (dm->devices_by_hqos_cpu, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + + vec_validate_aligned (dm->hqos_threads, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + + nports = rte_eth_dev_count (); + if (nports < 1) + { + clib_warning ("DPDK drivers found no ports..."); + } + + if (CLIB_DEBUG > 0) + clib_warning ("DPDK drivers found %d ports...", nports); + + /* + * All buffers are all allocated from the same rte_mempool. + * Thus they all have the same number of data bytes. + */ + dm->vlib_buffer_free_list_index = + vlib_buffer_get_or_create_free_list (vm, + VLIB_BUFFER_DEFAULT_FREE_LIST_BYTES, + "dpdk rx"); + + if (dm->conf->enable_tcp_udp_checksum) + dm->buffer_flags_template &= ~(IP_BUFFER_L4_CHECKSUM_CORRECT + | IP_BUFFER_L4_CHECKSUM_COMPUTED); + + for (i = 0; i < nports; i++) + { + u8 addr[6]; + u8 vlan_strip = 0; + int j; + struct rte_eth_dev_info dev_info; + clib_error_t *rv; + struct rte_eth_link l; + dpdk_device_config_t *devconf = 0; + vlib_pci_addr_t pci_addr; + uword *p = 0; + + rte_eth_dev_info_get (i, &dev_info); + if (dev_info.pci_dev) /* bonded interface has no pci info */ + { + pci_addr.domain = dev_info.pci_dev->addr.domain; + pci_addr.bus = dev_info.pci_dev->addr.bus; + pci_addr.slot = dev_info.pci_dev->addr.devid; + pci_addr.function = dev_info.pci_dev->addr.function; + p = + hash_get (dm->conf->device_config_index_by_pci_addr, + pci_addr.as_u32); + } + + if (p) + devconf = pool_elt_at_index (dm->conf->dev_confs, p[0]); + else + devconf = &dm->conf->default_devconf; + + /* Create vnet interface */ + vec_add2_aligned (dm->devices, xd, 1, CLIB_CACHE_LINE_BYTES); + xd->nb_rx_desc = DPDK_NB_RX_DESC_DEFAULT; + xd->nb_tx_desc = DPDK_NB_TX_DESC_DEFAULT; + xd->cpu_socket = (i8) rte_eth_dev_socket_id (i); + + /* Handle interface naming for devices with multiple ports sharing same PCI ID */ + if (dev_info.pci_dev) + { + struct rte_eth_dev_info di = { 0 }; + rte_eth_dev_info_get (i + 1, &di); + if (di.pci_dev && pci_addr.as_u32 != last_pci_addr.as_u32 && + memcmp (&dev_info.pci_dev->addr, &di.pci_dev->addr, + sizeof (struct rte_pci_addr)) == 0) + { + xd->interface_name_suffix = format (0, "0"); + last_pci_addr.as_u32 = pci_addr.as_u32; + last_pci_addr_port = i; + } + else if (pci_addr.as_u32 == last_pci_addr.as_u32) + { + xd->interface_name_suffix = + format (0, "%u", i - last_pci_addr_port); + } + else + { + last_pci_addr.as_u32 = ~0; + } + } + else + last_pci_addr.as_u32 = ~0; + + clib_memcpy (&xd->tx_conf, &dev_info.default_txconf, + sizeof (struct rte_eth_txconf)); + if (dm->conf->no_multi_seg) + { + xd->tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS; + port_conf_template.rxmode.jumbo_frame = 0; + } + else + { + xd->tx_conf.txq_flags &= ~ETH_TXQ_FLAGS_NOMULTSEGS; + port_conf_template.rxmode.jumbo_frame = 1; + xd->flags |= DPDK_DEVICE_FLAG_MAYBE_MULTISEG; + } + + clib_memcpy (&xd->port_conf, &port_conf_template, + sizeof (struct rte_eth_conf)); + + xd->tx_q_used = clib_min (dev_info.max_tx_queues, tm->n_vlib_mains); + + if (devconf->num_tx_queues > 0 + && devconf->num_tx_queues < xd->tx_q_used) + xd->tx_q_used = clib_min (xd->tx_q_used, devconf->num_tx_queues); + + if (devconf->num_rx_queues > 1 && dm->use_rss == 0) + { + dm->use_rss = 1; + } + + if (devconf->num_rx_queues > 1 + && dev_info.max_rx_queues >= devconf->num_rx_queues) + { + xd->rx_q_used = devconf->num_rx_queues; + xd->port_conf.rxmode.mq_mode = ETH_MQ_RX_RSS; + if (devconf->rss_fn == 0) + xd->port_conf.rx_adv_conf.rss_conf.rss_hf = + ETH_RSS_IP | ETH_RSS_UDP | ETH_RSS_TCP; + else + xd->port_conf.rx_adv_conf.rss_conf.rss_hf = devconf->rss_fn; + } + else + xd->rx_q_used = 1; + + xd->flags |= DPDK_DEVICE_FLAG_PMD; + + /* workaround for drivers not setting driver_name */ + if ((!dev_info.driver_name) && (dev_info.pci_dev)) +#if RTE_VERSION < RTE_VERSION_NUM(16, 11, 0, 0) + dev_info.driver_name = dev_info.pci_dev->driver->name; +#else + dev_info.driver_name = dev_info.pci_dev->driver->driver.name; +#endif + ASSERT (dev_info.driver_name); + + if (!xd->pmd) + { + + +#define _(s,f) else if (dev_info.driver_name && \ + !strcmp(dev_info.driver_name, s)) \ + xd->pmd = VNET_DPDK_PMD_##f; + if (0) + ; + foreach_dpdk_pmd +#undef _ + else + xd->pmd = VNET_DPDK_PMD_UNKNOWN; + + xd->port_type = VNET_DPDK_PORT_TYPE_UNKNOWN; + xd->nb_rx_desc = DPDK_NB_RX_DESC_DEFAULT; + xd->nb_tx_desc = DPDK_NB_TX_DESC_DEFAULT; + + switch (xd->pmd) + { + /* 1G adapters */ + case VNET_DPDK_PMD_E1000EM: + case VNET_DPDK_PMD_IGB: + case VNET_DPDK_PMD_IGBVF: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_1G; + break; + + /* 10G adapters */ + case VNET_DPDK_PMD_IXGBE: + case VNET_DPDK_PMD_IXGBEVF: + case VNET_DPDK_PMD_THUNDERX: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + case VNET_DPDK_PMD_DPAA2: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + + /* Cisco VIC */ + case VNET_DPDK_PMD_ENIC: + rte_eth_link_get_nowait (i, &l); + xd->flags |= DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE; + if (l.link_speed == 40000) + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G; + else + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + + /* Intel Fortville */ + case VNET_DPDK_PMD_I40E: + case VNET_DPDK_PMD_I40EVF: + xd->flags |= DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE; + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G; + + switch (dev_info.pci_dev->id.device_id) + { + case I40E_DEV_ID_10G_BASE_T: + case I40E_DEV_ID_SFP_XL710: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + case I40E_DEV_ID_QSFP_A: + case I40E_DEV_ID_QSFP_B: + case I40E_DEV_ID_QSFP_C: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G; + break; + case I40E_DEV_ID_VF: + rte_eth_link_get_nowait (i, &l); + xd->port_type = l.link_speed == 10000 ? + VNET_DPDK_PORT_TYPE_ETH_10G : VNET_DPDK_PORT_TYPE_ETH_40G; + break; + default: + xd->port_type = VNET_DPDK_PORT_TYPE_UNKNOWN; + } + break; + + case VNET_DPDK_PMD_CXGBE: + switch (dev_info.pci_dev->id.device_id) + { + case 0x540d: /* T580-CR */ + case 0x5410: /* T580-LP-cr */ + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G; + break; + case 0x5403: /* T540-CR */ + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + default: + xd->port_type = VNET_DPDK_PORT_TYPE_UNKNOWN; + } + break; + + case VNET_DPDK_PMD_MLX5: + { + char *pn_100g[] = { "MCX415A-CCAT", "MCX416A-CCAT", 0 }; + char *pn_40g[] = { "MCX413A-BCAT", "MCX414A-BCAT", + "MCX415A-BCAT", "MCX416A-BCAT", "MCX4131A-BCAT", 0 + }; + char *pn_10g[] = { "MCX4111A-XCAT", "MCX4121A-XCAT", 0 }; + + vlib_pci_device_t *pd = vlib_get_pci_device (&pci_addr); + u8 *pn = 0; + char **c; + int found = 0; + pn = format (0, "%U%c", + format_vlib_pci_vpd, pd->vpd_r, "PN", 0); + + if (!pn) + break; + + c = pn_100g; + while (!found && c[0]) + { + if (strncmp ((char *) pn, c[0], strlen (c[0])) == 0) + { + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_100G; + break; + } + c++; + } + + c = pn_40g; + while (!found && c[0]) + { + if (strncmp ((char *) pn, c[0], strlen (c[0])) == 0) + { + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G; + break; + } + c++; + } + + c = pn_10g; + while (!found && c[0]) + { + if (strncmp ((char *) pn, c[0], strlen (c[0])) == 0) + { + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + } + c++; + } + + vec_free (pn); + } + + break; + /* Intel Red Rock Canyon */ + case VNET_DPDK_PMD_FM10K: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_SWITCH; + break; + + /* virtio */ + case VNET_DPDK_PMD_VIRTIO: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_1G; + xd->nb_rx_desc = DPDK_NB_RX_DESC_VIRTIO; + xd->nb_tx_desc = DPDK_NB_TX_DESC_VIRTIO; + break; + + /* vmxnet3 */ + case VNET_DPDK_PMD_VMXNET3: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_1G; + xd->tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS; + break; + + case VNET_DPDK_PMD_AF_PACKET: + xd->port_type = VNET_DPDK_PORT_TYPE_AF_PACKET; + xd->af_packet_port_id = af_packet_port_id++; + break; + + case VNET_DPDK_PMD_BOND: + xd->flags |= DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE; + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_BOND; + break; + + default: + xd->port_type = VNET_DPDK_PORT_TYPE_UNKNOWN; + } + + if (devconf->num_rx_desc) + xd->nb_rx_desc = devconf->num_rx_desc; + + if (devconf->num_tx_desc) + xd->nb_tx_desc = devconf->num_tx_desc; + } + + /* + * Ensure default mtu is not > the mtu read from the hardware. + * Otherwise rte_eth_dev_configure() will fail and the port will + * not be available. + */ + if (ETHERNET_MAX_PACKET_BYTES > dev_info.max_rx_pktlen) + { + /* + * This device does not support the platforms's max frame + * size. Use it's advertised mru instead. + */ + xd->port_conf.rxmode.max_rx_pkt_len = dev_info.max_rx_pktlen; + } + else + { + xd->port_conf.rxmode.max_rx_pkt_len = ETHERNET_MAX_PACKET_BYTES; + + /* + * Some platforms do not account for Ethernet FCS (4 bytes) in + * MTU calculations. To interop with them increase mru but only + * if the device's settings can support it. + */ + if ((dev_info.max_rx_pktlen >= (ETHERNET_MAX_PACKET_BYTES + 4)) && + xd->port_conf.rxmode.hw_strip_crc) + { + /* + * Allow additional 4 bytes (for Ethernet FCS). These bytes are + * stripped by h/w and so will not consume any buffer memory. + */ + xd->port_conf.rxmode.max_rx_pkt_len += 4; + } + } + + if (xd->pmd == VNET_DPDK_PMD_AF_PACKET) + { + f64 now = vlib_time_now (vm); + u32 rnd; + rnd = (u32) (now * 1e6); + rnd = random_u32 (&rnd); + clib_memcpy (addr + 2, &rnd, sizeof (rnd)); + addr[0] = 2; + addr[1] = 0xfe; + } + else + rte_eth_macaddr_get (i, (struct ether_addr *) addr); + + if (xd->tx_q_used < tm->n_vlib_mains) + dpdk_device_lock_init (xd); + + xd->device_index = xd - dm->devices; + ASSERT (i == xd->device_index); + xd->per_interface_next_index = ~0; + + /* assign interface to input thread */ + dpdk_device_and_queue_t *dq; + int q; + + if (devconf->workers) + { + int i; + q = 0; + /* *INDENT-OFF* */ + clib_bitmap_foreach (i, devconf->workers, ({ + int cpu = dm->input_cpu_first_index + i; + unsigned lcore = vlib_worker_threads[cpu].lcore_id; + vec_validate(xd->cpu_socket_id_by_queue, q); + xd->cpu_socket_id_by_queue[q] = rte_lcore_to_socket_id(lcore); + vec_add2(dm->devices_by_cpu[cpu], dq, 1); + dq->device = xd->device_index; + dq->queue_id = q++; + })); + /* *INDENT-ON* */ + } + else + for (q = 0; q < xd->rx_q_used; q++) + { + int cpu = dm->input_cpu_first_index + next_cpu; + unsigned lcore = vlib_worker_threads[cpu].lcore_id; + + /* + * numa node for worker thread handling this queue + * needed for taking buffers from the right mempool + */ + vec_validate (xd->cpu_socket_id_by_queue, q); + xd->cpu_socket_id_by_queue[q] = rte_lcore_to_socket_id (lcore); + + /* + * construct vector of (device,queue) pairs for each worker thread + */ + vec_add2 (dm->devices_by_cpu[cpu], dq, 1); + dq->device = xd->device_index; + dq->queue_id = q; + + next_cpu++; + if (next_cpu == dm->input_cpu_count) + next_cpu = 0; + } + + + if (devconf->hqos_enabled) + { + xd->flags |= DPDK_DEVICE_FLAG_HQOS; + + if (devconf->hqos.hqos_thread_valid) + { + int cpu = dm->hqos_cpu_first_index + devconf->hqos.hqos_thread; + + if (devconf->hqos.hqos_thread >= dm->hqos_cpu_count) + return clib_error_return (0, "invalid HQoS thread index"); + + vec_add2 (dm->devices_by_hqos_cpu[cpu], dq, 1); + dq->device = xd->device_index; + dq->queue_id = 0; + } + else + { + int cpu = dm->hqos_cpu_first_index + next_hqos_cpu; + + if (dm->hqos_cpu_count == 0) + return clib_error_return (0, "no HQoS threads available"); + + vec_add2 (dm->devices_by_hqos_cpu[cpu], dq, 1); + dq->device = xd->device_index; + dq->queue_id = 0; + + next_hqos_cpu++; + if (next_hqos_cpu == dm->hqos_cpu_count) + next_hqos_cpu = 0; + + devconf->hqos.hqos_thread_valid = 1; + devconf->hqos.hqos_thread = cpu; + } + } + + vec_validate_aligned (xd->tx_vectors, tm->n_vlib_mains, + CLIB_CACHE_LINE_BYTES); + for (j = 0; j < tm->n_vlib_mains; j++) + { + vec_validate_ha (xd->tx_vectors[j], xd->nb_tx_desc, + sizeof (tx_ring_hdr_t), CLIB_CACHE_LINE_BYTES); + vec_reset_length (xd->tx_vectors[j]); + } + + vec_validate_aligned (xd->rx_vectors, xd->rx_q_used, + CLIB_CACHE_LINE_BYTES); + for (j = 0; j < xd->rx_q_used; j++) + { + vec_validate_aligned (xd->rx_vectors[j], VLIB_FRAME_SIZE - 1, + CLIB_CACHE_LINE_BYTES); + vec_reset_length (xd->rx_vectors[j]); + } + + vec_validate_aligned (xd->d_trace_buffers, tm->n_vlib_mains, + CLIB_CACHE_LINE_BYTES); + + rv = dpdk_port_setup (dm, xd); + + if (rv) + return rv; + + if (devconf->hqos_enabled) + { + rv = dpdk_port_setup_hqos (xd, &devconf->hqos); + if (rv) + return rv; + } + + /* count the number of descriptors used for this device */ + nb_desc += xd->nb_rx_desc + xd->nb_tx_desc * xd->tx_q_used; + + error = ethernet_register_interface + (dm->vnet_main, dpdk_device_class.index, xd->device_index, + /* ethernet address */ addr, + &xd->vlib_hw_if_index, dpdk_flag_change); + if (error) + return error; + + sw = vnet_get_hw_sw_interface (dm->vnet_main, xd->vlib_hw_if_index); + xd->vlib_sw_if_index = sw->sw_if_index; + hi = vnet_get_hw_interface (dm->vnet_main, xd->vlib_hw_if_index); + + /* + * DAW-FIXME: The Cisco VIC firmware does not provide an api for a + * driver to dynamically change the mtu. If/when the + * VIC firmware gets fixed, then this should be removed. + */ + if (xd->pmd == VNET_DPDK_PMD_ENIC) + { + /* + * Initialize mtu to what has been set by CIMC in the firmware cfg. + */ + hi->max_packet_bytes = dev_info.max_rx_pktlen; + if (devconf->vlan_strip_offload != DPDK_DEVICE_VLAN_STRIP_OFF) + vlan_strip = 1; /* remove vlan tag from VIC port by default */ + else + clib_warning ("VLAN strip disabled for interface\n"); + } + else if (devconf->vlan_strip_offload == DPDK_DEVICE_VLAN_STRIP_ON) + vlan_strip = 1; + + if (vlan_strip) + { + int vlan_off; + vlan_off = rte_eth_dev_get_vlan_offload (xd->device_index); + vlan_off |= ETH_VLAN_STRIP_OFFLOAD; + xd->port_conf.rxmode.hw_vlan_strip = vlan_off; + if (rte_eth_dev_set_vlan_offload (xd->device_index, vlan_off) == 0) + clib_warning ("VLAN strip enabled for interface\n"); + else + clib_warning ("VLAN strip cannot be supported by interface\n"); + } + + hi->max_l3_packet_bytes[VLIB_RX] = hi->max_l3_packet_bytes[VLIB_TX] = + xd->port_conf.rxmode.max_rx_pkt_len - sizeof (ethernet_header_t); + + rte_eth_dev_set_mtu (xd->device_index, hi->max_packet_bytes); + } + + if (nb_desc > dm->conf->num_mbufs) + clib_warning ("%d mbufs allocated but total rx/tx ring size is %d\n", + dm->conf->num_mbufs, nb_desc); + + return 0; +} + +static void +dpdk_bind_devices_to_uio (dpdk_config_main_t * conf) +{ + vlib_pci_main_t *pm = &pci_main; + clib_error_t *error; + vlib_pci_device_t *d; + u8 *pci_addr = 0; + int num_whitelisted = vec_len (conf->dev_confs); + + /* *INDENT-OFF* */ + pool_foreach (d, pm->pci_devs, ({ + dpdk_device_config_t * devconf = 0; + vec_reset_length (pci_addr); + pci_addr = format (pci_addr, "%U%c", format_vlib_pci_addr, &d->bus_address, 0); + + if (d->device_class != PCI_CLASS_NETWORK_ETHERNET) + continue; + + if (num_whitelisted) + { + uword * p = hash_get (conf->device_config_index_by_pci_addr, d->bus_address.as_u32); + + if (!p) + continue; + + devconf = pool_elt_at_index (conf->dev_confs, p[0]); + } + + /* virtio */ + if (d->vendor_id == 0x1af4 && d->device_id == 0x1000) + ; + /* vmxnet3 */ + else if (d->vendor_id == 0x15ad && d->device_id == 0x07b0) + ; + /* all Intel devices */ + else if (d->vendor_id == 0x8086) + ; + /* Cisco VIC */ + else if (d->vendor_id == 0x1137 && d->device_id == 0x0043) + ; + /* Chelsio T4/T5 */ + else if (d->vendor_id == 0x1425 && (d->device_id & 0xe000) == 0x4000) + ; + else + { + clib_warning ("Unsupported Ethernet PCI device 0x%04x:0x%04x found " + "at PCI address %s\n", (u16) d->vendor_id, (u16) d->device_id, + pci_addr); + continue; + } + + error = vlib_pci_bind_to_uio (d, (char *) conf->uio_driver_name); + + if (error) + { + if (devconf == 0) + { + pool_get (conf->dev_confs, devconf); + hash_set (conf->device_config_index_by_pci_addr, d->bus_address.as_u32, + devconf - conf->dev_confs); + devconf->pci_addr.as_u32 = d->bus_address.as_u32; + } + devconf->is_blacklisted = 1; + clib_error_report (error); + } + })); + /* *INDENT-ON* */ + vec_free (pci_addr); +} + +static clib_error_t * +dpdk_device_config (dpdk_config_main_t * conf, vlib_pci_addr_t pci_addr, + unformat_input_t * input, u8 is_default) +{ + clib_error_t *error = 0; + uword *p; + dpdk_device_config_t *devconf; + unformat_input_t sub_input; + + if (is_default) + { + devconf = &conf->default_devconf; + } + else + { + p = hash_get (conf->device_config_index_by_pci_addr, pci_addr.as_u32); + + if (!p) + { + pool_get (conf->dev_confs, devconf); + hash_set (conf->device_config_index_by_pci_addr, pci_addr.as_u32, + devconf - conf->dev_confs); + } + else + return clib_error_return (0, + "duplicate configuration for PCI address %U", + format_vlib_pci_addr, &pci_addr); + } + + devconf->pci_addr.as_u32 = pci_addr.as_u32; + devconf->hqos_enabled = 0; + dpdk_device_config_hqos_default (&devconf->hqos); + + if (!input) + return 0; + + unformat_skip_white_space (input); + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "num-rx-queues %u", &devconf->num_rx_queues)) + ; + else if (unformat (input, "num-tx-queues %u", &devconf->num_tx_queues)) + ; + else if (unformat (input, "num-rx-desc %u", &devconf->num_rx_desc)) + ; + else if (unformat (input, "num-tx-desc %u", &devconf->num_tx_desc)) + ; + else if (unformat (input, "workers %U", unformat_bitmap_list, + &devconf->workers)) + ; + else + if (unformat + (input, "rss %U", unformat_vlib_cli_sub_input, &sub_input)) + { + error = unformat_rss_fn (&sub_input, &devconf->rss_fn); + if (error) + break; + } + else if (unformat (input, "vlan-strip-offload off")) + devconf->vlan_strip_offload = DPDK_DEVICE_VLAN_STRIP_OFF; + else if (unformat (input, "vlan-strip-offload on")) + devconf->vlan_strip_offload = DPDK_DEVICE_VLAN_STRIP_ON; + else + if (unformat + (input, "hqos %U", unformat_vlib_cli_sub_input, &sub_input)) + { + devconf->hqos_enabled = 1; + error = unformat_hqos (&sub_input, &devconf->hqos); + if (error) + break; + } + else if (unformat (input, "hqos")) + { + devconf->hqos_enabled = 1; + } + else + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + break; + } + } + + if (error) + return error; + + if (devconf->workers && devconf->num_rx_queues == 0) + devconf->num_rx_queues = clib_bitmap_count_set_bits (devconf->workers); + else if (devconf->workers && + clib_bitmap_count_set_bits (devconf->workers) != + devconf->num_rx_queues) + error = + clib_error_return (0, + "%U: number of worker threadds must be " + "equal to number of rx queues", format_vlib_pci_addr, + &pci_addr); + + return error; +} + +static clib_error_t * +dpdk_config (vlib_main_t * vm, unformat_input_t * input) +{ + clib_error_t *error = 0; + dpdk_main_t *dm = &dpdk_main; + dpdk_config_main_t *conf = &dpdk_config_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_device_config_t *devconf; + vlib_pci_addr_t pci_addr; + unformat_input_t sub_input; + u8 *s, *tmp = 0; + u8 *rte_cmd = 0, *ethname = 0; + u32 log_level; + int ret, i; + int num_whitelisted = 0; + u8 no_pci = 0; + u8 no_huge = 0; + u8 huge_dir = 0; + u8 file_prefix = 0; + u8 *socket_mem = 0; + + conf->device_config_index_by_pci_addr = hash_create (0, sizeof (uword)); + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + /* Prime the pump */ + if (unformat (input, "no-hugetlb")) + { + vec_add1 (conf->eal_init_args, (u8 *) "no-huge"); + no_huge = 1; + } + + else if (unformat (input, "enable-tcp-udp-checksum")) + conf->enable_tcp_udp_checksum = 1; + + else if (unformat (input, "decimal-interface-names")) + conf->interface_name_format_decimal = 1; + + else if (unformat (input, "no-multi-seg")) + conf->no_multi_seg = 1; + + else if (unformat (input, "dev default %U", unformat_vlib_cli_sub_input, + &sub_input)) + { + error = + dpdk_device_config (conf, (vlib_pci_addr_t) (u32) ~ 1, &sub_input, + 1); + + if (error) + return error; + } + else + if (unformat + (input, "dev %U %U", unformat_vlib_pci_addr, &pci_addr, + unformat_vlib_cli_sub_input, &sub_input)) + { + error = dpdk_device_config (conf, pci_addr, &sub_input, 0); + + if (error) + return error; + + num_whitelisted++; + } + else if (unformat (input, "dev %U", unformat_vlib_pci_addr, &pci_addr)) + { + error = dpdk_device_config (conf, pci_addr, 0, 0); + + if (error) + return error; + + num_whitelisted++; + } + else if (unformat (input, "num-mbufs %d", &conf->num_mbufs)) + ; + else if (unformat (input, "kni %d", &conf->num_kni)) + ; + else if (unformat (input, "uio-driver %s", &conf->uio_driver_name)) + ; + else if (unformat (input, "socket-mem %s", &socket_mem)) + ; + else if (unformat (input, "no-pci")) + { + no_pci = 1; + tmp = format (0, "--no-pci%c", 0); + vec_add1 (conf->eal_init_args, tmp); + } + else if (unformat (input, "poll-sleep %d", &dm->poll_sleep)) + ; + +#define _(a) \ + else if (unformat(input, #a)) \ + { \ + tmp = format (0, "--%s%c", #a, 0); \ + vec_add1 (conf->eal_init_args, tmp); \ + } + foreach_eal_double_hyphen_predicate_arg +#undef _ +#define _(a) \ + else if (unformat(input, #a " %s", &s)) \ + { \ + if (!strncmp(#a, "huge-dir", 8)) \ + huge_dir = 1; \ + else if (!strncmp(#a, "file-prefix", 11)) \ + file_prefix = 1; \ + tmp = format (0, "--%s%c", #a, 0); \ + vec_add1 (conf->eal_init_args, tmp); \ + vec_add1 (s, 0); \ + vec_add1 (conf->eal_init_args, s); \ + } + foreach_eal_double_hyphen_arg +#undef _ +#define _(a,b) \ + else if (unformat(input, #a " %s", &s)) \ + { \ + tmp = format (0, "-%s%c", #b, 0); \ + vec_add1 (conf->eal_init_args, tmp); \ + vec_add1 (s, 0); \ + vec_add1 (conf->eal_init_args, s); \ + } + foreach_eal_single_hyphen_arg +#undef _ +#define _(a,b) \ + else if (unformat(input, #a " %s", &s)) \ + { \ + tmp = format (0, "-%s%c", #b, 0); \ + vec_add1 (conf->eal_init_args, tmp); \ + vec_add1 (s, 0); \ + vec_add1 (conf->eal_init_args, s); \ + conf->a##_set_manually = 1; \ + } + foreach_eal_single_hyphen_mandatory_arg +#undef _ + else if (unformat (input, "default")) + ; + + else if (unformat_skip_white_space (input)) + ; + else + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + goto done; + } + } + + if (!conf->uio_driver_name) + conf->uio_driver_name = format (0, "igb_uio%c", 0); + + /* + * Use 1G huge pages if available. + */ + if (!no_huge && !huge_dir) + { + u32 x, *mem_by_socket = 0; + uword c = 0; + u8 use_1g = 1; + u8 use_2m = 1; + u8 less_than_1g = 1; + int rv; + + umount (DEFAULT_HUGE_DIR); + + /* Process "socket-mem" parameter value */ + if (vec_len (socket_mem)) + { + unformat_input_t in; + unformat_init_vector (&in, socket_mem); + while (unformat_check_input (&in) != UNFORMAT_END_OF_INPUT) + { + if (unformat (&in, "%u,", &x)) + ; + else if (unformat (&in, "%u", &x)) + ; + else if (unformat (&in, ",")) + x = 0; + else + break; + + vec_add1 (mem_by_socket, x); + + if (x > 1023) + less_than_1g = 0; + } + /* Note: unformat_free vec_frees(in.buffer), aka socket_mem... */ + unformat_free (&in); + socket_mem = 0; + } + else + { + /* *INDENT-OFF* */ + clib_bitmap_foreach (c, tm->cpu_socket_bitmap, ( + { + vec_validate(mem_by_socket, c); + mem_by_socket[c] = 256; /* default per-socket mem */ + } + )); + /* *INDENT-ON* */ + } + + /* check if available enough 1GB pages for each socket */ + /* *INDENT-OFF* */ + clib_bitmap_foreach (c, tm->cpu_socket_bitmap, ( + { + int pages_avail, page_size, mem; + + vec_validate(mem_by_socket, c); + mem = mem_by_socket[c]; + + page_size = 1024; + pages_avail = vlib_sysfs_get_free_hugepages(c, page_size * 1024); + + if (pages_avail < 0 || page_size * pages_avail < mem) + use_1g = 0; + + page_size = 2; + pages_avail = vlib_sysfs_get_free_hugepages(c, page_size * 1024); + + if (pages_avail < 0 || page_size * pages_avail < mem) + use_2m = 0; + })); + /* *INDENT-ON* */ + + if (mem_by_socket == 0) + { + error = clib_error_return (0, "mem_by_socket NULL"); + goto done; + } + _vec_len (mem_by_socket) = c + 1; + + /* regenerate socket_mem string */ + vec_foreach_index (x, mem_by_socket) + socket_mem = format (socket_mem, "%s%u", + socket_mem ? "," : "", mem_by_socket[x]); + socket_mem = format (socket_mem, "%c", 0); + + vec_free (mem_by_socket); + + rv = mkdir (VPP_RUN_DIR, 0755); + if (rv && errno != EEXIST) + { + error = clib_error_return (0, "mkdir '%s' failed errno %d", + VPP_RUN_DIR, errno); + goto done; + } + + rv = mkdir (DEFAULT_HUGE_DIR, 0755); + if (rv && errno != EEXIST) + { + error = clib_error_return (0, "mkdir '%s' failed errno %d", + DEFAULT_HUGE_DIR, errno); + goto done; + } + + if (use_1g && !(less_than_1g && use_2m)) + { + rv = + mount ("none", DEFAULT_HUGE_DIR, "hugetlbfs", 0, "pagesize=1G"); + } + else if (use_2m) + { + rv = mount ("none", DEFAULT_HUGE_DIR, "hugetlbfs", 0, NULL); + } + else + { + return clib_error_return (0, "not enough free huge pages"); + } + + if (rv) + { + error = clib_error_return (0, "mount failed %d", errno); + goto done; + } + + tmp = format (0, "--huge-dir%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "%s%c", DEFAULT_HUGE_DIR, 0); + vec_add1 (conf->eal_init_args, tmp); + if (!file_prefix) + { + tmp = format (0, "--file-prefix%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "vpp%c", 0); + vec_add1 (conf->eal_init_args, tmp); + } + } + + vec_free (rte_cmd); + vec_free (ethname); + + if (error) + return error; + + /* I'll bet that -c and -n must be the first and second args... */ + if (!conf->coremask_set_manually) + { + vlib_thread_registration_t *tr; + uword *coremask = 0; + int i; + + /* main thread core */ + coremask = clib_bitmap_set (coremask, tm->main_lcore, 1); + + for (i = 0; i < vec_len (tm->registrations); i++) + { + tr = tm->registrations[i]; + coremask = clib_bitmap_or (coremask, tr->coremask); + } + + vec_insert (conf->eal_init_args, 2, 1); + conf->eal_init_args[1] = (u8 *) "-c"; + tmp = format (0, "%U%c", format_bitmap_hex, coremask, 0); + conf->eal_init_args[2] = tmp; + clib_bitmap_free (coremask); + } + + if (!conf->nchannels_set_manually) + { + vec_insert (conf->eal_init_args, 2, 3); + conf->eal_init_args[3] = (u8 *) "-n"; + tmp = format (0, "%d", conf->nchannels); + conf->eal_init_args[4] = tmp; + } + + if (no_pci == 0 && geteuid () == 0) + dpdk_bind_devices_to_uio (conf); + +#define _(x) \ + if (devconf->x == 0 && conf->default_devconf.x > 0) \ + devconf->x = conf->default_devconf.x ; + + /* *INDENT-OFF* */ + pool_foreach (devconf, conf->dev_confs, ({ + + /* default per-device config items */ + foreach_dpdk_device_config_item + + /* add DPDK EAL whitelist/blacklist entry */ + if (num_whitelisted > 0 && devconf->is_blacklisted == 0) + { + tmp = format (0, "-w%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "%U%c", format_vlib_pci_addr, &devconf->pci_addr, 0); + vec_add1 (conf->eal_init_args, tmp); + } + else if (num_whitelisted == 0 && devconf->is_blacklisted != 0) + { + tmp = format (0, "-b%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "%U%c", format_vlib_pci_addr, &devconf->pci_addr, 0); + vec_add1 (conf->eal_init_args, tmp); + } + })); + /* *INDENT-ON* */ + +#undef _ + + /* set master-lcore */ + tmp = format (0, "--master-lcore%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "%u%c", tm->main_lcore, 0); + vec_add1 (conf->eal_init_args, tmp); + + /* set socket-mem */ + tmp = format (0, "--socket-mem%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "%s%c", socket_mem, 0); + vec_add1 (conf->eal_init_args, tmp); + + /* NULL terminate the "argv" vector, in case of stupidity */ + vec_add1 (conf->eal_init_args, 0); + _vec_len (conf->eal_init_args) -= 1; + + /* Set up DPDK eal and packet mbuf pool early. */ + + log_level = (CLIB_DEBUG > 0) ? RTE_LOG_DEBUG : RTE_LOG_NOTICE; + + rte_set_log_level (log_level); + + vm = vlib_get_main (); + + /* make copy of args as rte_eal_init tends to mess up with arg array */ + for (i = 1; i < vec_len (conf->eal_init_args); i++) + conf->eal_init_args_str = format (conf->eal_init_args_str, "%s ", + conf->eal_init_args[i]); + + ret = + rte_eal_init (vec_len (conf->eal_init_args), + (char **) conf->eal_init_args); + + /* lazy umount hugepages */ + umount2 (DEFAULT_HUGE_DIR, MNT_DETACH); + + if (ret < 0) + return clib_error_return (0, "rte_eal_init returned %d", ret); + + /* Dump the physical memory layout prior to creating the mbuf_pool */ + fprintf (stdout, "DPDK physical memory layout:\n"); + rte_dump_physmem_layout (stdout); + + /* main thread 1st */ + error = vlib_buffer_pool_create (vm, conf->num_mbufs, rte_socket_id ()); + if (error) + return error; + + for (i = 0; i < RTE_MAX_LCORE; i++) + { + error = vlib_buffer_pool_create (vm, conf->num_mbufs, + rte_lcore_to_socket_id (i)); + if (error) + return error; + } + +done: + return error; +} + +VLIB_CONFIG_FUNCTION (dpdk_config, "dpdk"); + +void +dpdk_update_link_state (dpdk_device_t * xd, f64 now) +{ + vnet_main_t *vnm = vnet_get_main (); + struct rte_eth_link prev_link = xd->link; + u32 hw_flags = 0; + u8 hw_flags_chg = 0; + + /* only update link state for PMD interfaces */ + if ((xd->flags & DPDK_DEVICE_FLAG_PMD) == 0) + return; + + xd->time_last_link_update = now ? now : xd->time_last_link_update; + memset (&xd->link, 0, sizeof (xd->link)); + rte_eth_link_get_nowait (xd->device_index, &xd->link); + + if (LINK_STATE_ELOGS) + { + vlib_main_t *vm = vlib_get_main (); + ELOG_TYPE_DECLARE (e) = + { + .format = + "update-link-state: sw_if_index %d, admin_up %d," + "old link_state %d new link_state %d",.format_args = "i4i1i1i1",}; + + struct + { + u32 sw_if_index; + u8 admin_up; + u8 old_link_state; + u8 new_link_state; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->sw_if_index = xd->vlib_sw_if_index; + ed->admin_up = (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) != 0; + ed->old_link_state = (u8) + vnet_hw_interface_is_link_up (vnm, xd->vlib_hw_if_index); + ed->new_link_state = (u8) xd->link.link_status; + } + + if ((xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) && + ((xd->link.link_status != 0) ^ + vnet_hw_interface_is_link_up (vnm, xd->vlib_hw_if_index))) + { + hw_flags_chg = 1; + hw_flags |= (xd->link.link_status ? VNET_HW_INTERFACE_FLAG_LINK_UP : 0); + } + + if (hw_flags_chg || (xd->link.link_duplex != prev_link.link_duplex)) + { + hw_flags_chg = 1; + switch (xd->link.link_duplex) + { + case ETH_LINK_HALF_DUPLEX: + hw_flags |= VNET_HW_INTERFACE_FLAG_HALF_DUPLEX; + break; + case ETH_LINK_FULL_DUPLEX: + hw_flags |= VNET_HW_INTERFACE_FLAG_FULL_DUPLEX; + break; + default: + break; + } + } + if (hw_flags_chg || (xd->link.link_speed != prev_link.link_speed)) + { + hw_flags_chg = 1; + switch (xd->link.link_speed) + { + case ETH_SPEED_NUM_10M: + hw_flags |= VNET_HW_INTERFACE_FLAG_SPEED_10M; + break; + case ETH_SPEED_NUM_100M: + hw_flags |= VNET_HW_INTERFACE_FLAG_SPEED_100M; + break; + case ETH_SPEED_NUM_1G: + hw_flags |= VNET_HW_INTERFACE_FLAG_SPEED_1G; + break; + case ETH_SPEED_NUM_10G: + hw_flags |= VNET_HW_INTERFACE_FLAG_SPEED_10G; + break; + case ETH_SPEED_NUM_40G: + hw_flags |= VNET_HW_INTERFACE_FLAG_SPEED_40G; + break; + case 0: + break; + default: + clib_warning ("unknown link speed %d", xd->link.link_speed); + break; + } + } + if (hw_flags_chg) + { + if (LINK_STATE_ELOGS) + { + vlib_main_t *vm = vlib_get_main (); + + ELOG_TYPE_DECLARE (e) = + { + .format = + "update-link-state: sw_if_index %d, new flags %d",.format_args + = "i4i4",}; + + struct + { + u32 sw_if_index; + u32 flags; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->sw_if_index = xd->vlib_sw_if_index; + ed->flags = hw_flags; + } + vnet_hw_interface_set_flags (vnm, xd->vlib_hw_if_index, hw_flags); + } +} + +static uword +dpdk_process (vlib_main_t * vm, vlib_node_runtime_t * rt, vlib_frame_t * f) +{ + clib_error_t *error; + vnet_main_t *vnm = vnet_get_main (); + dpdk_main_t *dm = &dpdk_main; + ethernet_main_t *em = ðernet_main; + dpdk_device_t *xd; + vlib_thread_main_t *tm = vlib_get_thread_main (); + int i; + + error = dpdk_lib_init (dm); + + /* + * Turn on the input node if we found some devices to drive + * and we're not running worker threads or i/o threads + */ + + if (error == 0 && vec_len (dm->devices) > 0) + { + if (tm->n_vlib_mains == 1) + vlib_node_set_state (vm, dpdk_input_node.index, + VLIB_NODE_STATE_POLLING); + else + for (i = 0; i < tm->n_vlib_mains; i++) + if (vec_len (dm->devices_by_cpu[i]) > 0) + vlib_node_set_state (vlib_mains[i], dpdk_input_node.index, + VLIB_NODE_STATE_POLLING); + } + + if (error) + clib_error_report (error); + + tm->worker_thread_release = 1; + + f64 now = vlib_time_now (vm); + vec_foreach (xd, dm->devices) + { + dpdk_update_link_state (xd, now); + } + + { + /* + * Extra set up for bond interfaces: + * 1. Setup MACs for bond interfaces and their slave links which was set + * in dpdk_port_setup() but needs to be done again here to take effect. + * 2. Set up info for bond interface related CLI support. + */ + int nports = rte_eth_dev_count (); + if (nports > 0) + { + for (i = 0; i < nports; i++) + { + struct rte_eth_dev_info dev_info; + rte_eth_dev_info_get (i, &dev_info); + if (!dev_info.driver_name) +#if RTE_VERSION < RTE_VERSION_NUM(16, 11, 0, 0) + dev_info.driver_name = dev_info.pci_dev->driver->name; +#else + dev_info.driver_name = dev_info.pci_dev->driver->driver.name; +#endif + ASSERT (dev_info.driver_name); + if (strncmp (dev_info.driver_name, "rte_bond_pmd", 12) == 0) + { + u8 addr[6]; + u8 slink[16]; + int nlink = rte_eth_bond_slaves_get (i, slink, 16); + if (nlink > 0) + { + vnet_hw_interface_t *bhi; + ethernet_interface_t *bei; + int rv; + + /* Get MAC of 1st slave link */ + rte_eth_macaddr_get (slink[0], + (struct ether_addr *) addr); + /* Set MAC of bounded interface to that of 1st slave link */ + rv = + rte_eth_bond_mac_address_set (i, + (struct ether_addr *) + addr); + if (rv < 0) + clib_warning ("Failed to set MAC address"); + + /* Populate MAC of bonded interface in VPP hw tables */ + bhi = + vnet_get_hw_interface (vnm, + dm->devices[i].vlib_hw_if_index); + bei = + pool_elt_at_index (em->interfaces, bhi->hw_instance); + clib_memcpy (bhi->hw_address, addr, 6); + clib_memcpy (bei->address, addr, 6); + /* Init l3 packet size allowed on bonded interface */ + bhi->max_packet_bytes = ETHERNET_MAX_PACKET_BYTES; + bhi->max_l3_packet_bytes[VLIB_RX] = + bhi->max_l3_packet_bytes[VLIB_TX] = + ETHERNET_MAX_PACKET_BYTES - sizeof (ethernet_header_t); + while (nlink >= 1) + { /* for all slave links */ + int slave = slink[--nlink]; + dpdk_device_t *sdev = &dm->devices[slave]; + vnet_hw_interface_t *shi; + vnet_sw_interface_t *ssi; + /* Add MAC to all slave links except the first one */ + if (nlink) + rte_eth_dev_mac_addr_add (slave, + (struct ether_addr *) + addr, 0); + /* Set slaves bitmap for bonded interface */ + bhi->bond_info = + clib_bitmap_set (bhi->bond_info, + sdev->vlib_hw_if_index, 1); + /* Set slave link flags on slave interface */ + shi = + vnet_get_hw_interface (vnm, sdev->vlib_hw_if_index); + ssi = + vnet_get_sw_interface (vnm, sdev->vlib_sw_if_index); + shi->bond_info = VNET_HW_INTERFACE_BOND_INFO_SLAVE; + ssi->flags |= VNET_SW_INTERFACE_FLAG_BOND_SLAVE; + + /* Set l3 packet size allowed as the lowest of slave */ + if (bhi->max_l3_packet_bytes[VLIB_RX] > + shi->max_l3_packet_bytes[VLIB_RX]) + bhi->max_l3_packet_bytes[VLIB_RX] = + bhi->max_l3_packet_bytes[VLIB_TX] = + shi->max_l3_packet_bytes[VLIB_RX]; + + /* Set max packet size allowed as the lowest of slave */ + if (bhi->max_packet_bytes > shi->max_packet_bytes) + bhi->max_packet_bytes = shi->max_packet_bytes; + } + } + } + } + } + } + + while (1) + { + /* + * check each time through the loop in case intervals are changed + */ + f64 min_wait = dm->link_state_poll_interval < dm->stat_poll_interval ? + dm->link_state_poll_interval : dm->stat_poll_interval; + + vlib_process_wait_for_event_or_clock (vm, min_wait); + + if (dm->admin_up_down_in_progress) + /* skip the poll if an admin up down is in progress (on any interface) */ + continue; + + vec_foreach (xd, dm->devices) + { + f64 now = vlib_time_now (vm); + if ((now - xd->time_last_stats_update) >= dm->stat_poll_interval) + dpdk_update_counters (xd, now); + if ((now - xd->time_last_link_update) >= dm->link_state_poll_interval) + dpdk_update_link_state (xd, now); + + } + } + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (dpdk_process_node,static) = { + .function = dpdk_process, + .type = VLIB_NODE_TYPE_PROCESS, + .name = "dpdk-process", + .process_log2_n_stack_bytes = 17, +}; +/* *INDENT-ON* */ + +int +dpdk_set_stat_poll_interval (f64 interval) +{ + if (interval < DPDK_MIN_STATS_POLL_INTERVAL) + return (VNET_API_ERROR_INVALID_VALUE); + + dpdk_main.stat_poll_interval = interval; + + return 0; +} + +int +dpdk_set_link_state_poll_interval (f64 interval) +{ + if (interval < DPDK_MIN_LINK_POLL_INTERVAL) + return (VNET_API_ERROR_INVALID_VALUE); + + dpdk_main.link_state_poll_interval = interval; + + return 0; +} + +clib_error_t * +dpdk_init (vlib_main_t * vm) +{ + dpdk_main_t *dm = &dpdk_main; + vlib_node_t *ei; + clib_error_t *error = 0; + vlib_thread_main_t *tm = vlib_get_thread_main (); + + /* verify that structs are cacheline aligned */ + STATIC_ASSERT (offsetof (dpdk_device_t, cacheline0) == 0, + "Cache line marker must be 1st element in dpdk_device_t"); + STATIC_ASSERT (offsetof (dpdk_device_t, cacheline1) == + CLIB_CACHE_LINE_BYTES, + "Data in cache line 0 is bigger than cache line size"); + STATIC_ASSERT (offsetof (dpdk_worker_t, cacheline0) == 0, + "Cache line marker must be 1st element in dpdk_worker_t"); + STATIC_ASSERT (offsetof (frame_queue_trace_t, cacheline0) == 0, + "Cache line marker must be 1st element in frame_queue_trace_t"); + + dm->vlib_main = vm; + dm->vnet_main = vnet_get_main (); + dm->conf = &dpdk_config_main; + + ei = vlib_get_node_by_name (vm, (u8 *) "ethernet-input"); + if (ei == 0) + return clib_error_return (0, "ethernet-input node AWOL"); + + dm->ethernet_input_node_index = ei->index; + + dm->conf->nchannels = 4; + dm->conf->num_mbufs = dm->conf->num_mbufs ? dm->conf->num_mbufs : NB_MBUF; + vec_add1 (dm->conf->eal_init_args, (u8 *) "vnet"); + + dm->dpdk_device_by_kni_port_id = hash_create (0, sizeof (uword)); + dm->vu_sw_if_index_by_listener_fd = hash_create (0, sizeof (uword)); + dm->vu_sw_if_index_by_sock_fd = hash_create (0, sizeof (uword)); + + /* $$$ use n_thread_stacks since it's known-good at this point */ + vec_validate (dm->recycle, tm->n_thread_stacks - 1); + + /* Default vlib_buffer_t flags, DISABLES tcp/udp checksumming... */ + dm->buffer_flags_template = + (VLIB_BUFFER_TOTAL_LENGTH_VALID | VNET_BUFFER_RTE_MBUF_VALID + | IP_BUFFER_L4_CHECKSUM_COMPUTED | IP_BUFFER_L4_CHECKSUM_CORRECT); + + dm->stat_poll_interval = DPDK_STATS_POLL_INTERVAL; + dm->link_state_poll_interval = DPDK_LINK_POLL_INTERVAL; + + /* init CLI */ + if ((error = vlib_call_init_function (vm, dpdk_cli_init))) + return error; + + return error; +} + +VLIB_INIT_FUNCTION (dpdk_init); + + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/cli.c b/src/vnet/devices/dpdk/ipsec/cli.c new file mode 100644 index 00000000000..3b634e036da --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/cli.c @@ -0,0 +1,141 @@ +/* + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include +#include + +static void +dpdk_ipsec_show_mapping (vlib_main_t * vm, u16 detail_display) +{ + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); + u32 i, skip_master; + + if (detail_display) + vlib_cli_output (vm, "worker\t%10s\t%15s\tdir\tdev\tqp\n", + "cipher", "auth"); + else + vlib_cli_output (vm, "worker\tcrypto device id(type)\n"); + + skip_master = vlib_num_workers () > 0; + + for (i = 0; i < tm->n_vlib_mains; i++) + { + uword key, data; + u32 cpu_index = vlib_mains[i]->cpu_index; + crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index]; + u8 *s = 0; + + if (skip_master) + { + skip_master = 0; + continue; + } + + if (!detail_display) + { + i32 last_cdev = -1; + crypto_qp_data_t *qpd; + + s = format (s, "%u\t", cpu_index); + + /* *INDENT-OFF* */ + vec_foreach (qpd, cwm->qp_data) + { + u32 dev_id = qpd->dev_id; + + if ((u16) last_cdev != dev_id) + { + struct rte_cryptodev_info cdev_info; + + rte_cryptodev_info_get (dev_id, &cdev_info); + + s = format(s, "%u(%s)\t", dev_id, cdev_info.feature_flags & + RTE_CRYPTODEV_FF_HW_ACCELERATED ? "HW" : "SW"); + } + last_cdev = dev_id; + } + /* *INDENT-ON* */ + vlib_cli_output (vm, "%s", s); + } + else + { + char cipher_str[15], auth_str[15]; + struct rte_cryptodev_capabilities cap; + crypto_worker_qp_key_t *p_key = (crypto_worker_qp_key_t *) & key; + /* *INDENT-OFF* */ + hash_foreach (key, data, cwm->algo_qp_map, + ({ + cap.op = RTE_CRYPTO_OP_TYPE_SYMMETRIC; + cap.sym.xform_type = RTE_CRYPTO_SYM_XFORM_CIPHER; + cap.sym.cipher.algo = p_key->cipher_algo; + check_algo_is_supported (&cap, cipher_str); + cap.op = RTE_CRYPTO_OP_TYPE_SYMMETRIC; + cap.sym.xform_type = RTE_CRYPTO_SYM_XFORM_AUTH; + cap.sym.auth.algo = p_key->auth_algo; + check_algo_is_supported (&cap, auth_str); + vlib_cli_output (vm, "%u\t%10s\t%15s\t%3s\t%u\t%u\n", + vlib_mains[i]->cpu_index, cipher_str, auth_str, + p_key->is_outbound ? "out" : "in", + cwm->qp_data[data].dev_id, + cwm->qp_data[data].qp_id); + })); + /* *INDENT-ON* */ + } + } +} + +static clib_error_t * +lcore_cryptodev_map_fn (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + u16 detail = 0; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "verbose")) + detail = 1; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + dpdk_ipsec_show_mapping (vm, detail); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (lcore_cryptodev_map, static) = { + .path = "show crypto device mapping", + .short_help = + "show cryptodev device mapping ", + .function = lcore_cryptodev_map_fn, +}; +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/crypto_node.c b/src/vnet/devices/dpdk/ipsec/crypto_node.c new file mode 100644 index 00000000000..7b32704ec05 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/crypto_node.c @@ -0,0 +1,210 @@ +/* + *------------------------------------------------------------------ + * crypto_node.c - DPDK Cryptodev input node + * + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include +#include +#include +#include + +#include + +#define foreach_dpdk_crypto_input_next \ + _(DROP, "error-drop") \ + _(ENCRYPT_POST, "dpdk-esp-encrypt-post") \ + _(DECRYPT_POST, "dpdk-esp-decrypt-post") + +typedef enum +{ +#define _(f,s) DPDK_CRYPTO_INPUT_NEXT_##f, + foreach_dpdk_crypto_input_next +#undef _ + DPDK_CRYPTO_INPUT_N_NEXT, +} dpdk_crypto_input_next_t; + +#define foreach_dpdk_crypto_input_error \ + _(DQ_COPS, "Crypto ops dequeued") \ + _(COP_FAILED, "Crypto op failed") + +typedef enum +{ +#define _(f,s) DPDK_CRYPTO_INPUT_ERROR_##f, + foreach_dpdk_crypto_input_error +#undef _ + DPDK_CRYPTO_INPUT_N_ERROR, +} dpdk_crypto_input_error_t; + +static char *dpdk_crypto_input_error_strings[] = { +#define _(n, s) s, + foreach_dpdk_crypto_input_error +#undef _ +}; + +vlib_node_registration_t dpdk_crypto_input_node; + +typedef struct +{ + u32 cdev; + u32 qp; + u32 status; + u32 sa_idx; + u32 next_index; +} dpdk_crypto_input_trace_t; + +static u8 * +format_dpdk_crypto_input_trace (u8 * s, va_list * args) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *); + dpdk_crypto_input_trace_t *t = va_arg (*args, dpdk_crypto_input_trace_t *); + + s = format (s, "dpdk_crypto: cryptodev-id %u queue-pair %u next-index %d", + t->cdev, t->qp, t->next_index); + + s = format (s, " status %u sa-idx %u\n", t->status, t->sa_idx); + + return s; +} + +static_always_inline u32 +dpdk_crypto_dequeue (vlib_main_t * vm, vlib_node_runtime_t * node, + crypto_qp_data_t * qpd) +{ + u32 n_deq, *to_next = 0, next_index, n_cops, def_next_index; + struct rte_crypto_op **cops = qpd->cops; + + if (qpd->inflights == 0) + return 0; + + if (qpd->is_outbound) + def_next_index = DPDK_CRYPTO_INPUT_NEXT_ENCRYPT_POST; + else + def_next_index = DPDK_CRYPTO_INPUT_NEXT_DECRYPT_POST; + + n_cops = rte_cryptodev_dequeue_burst (qpd->dev_id, qpd->qp_id, + cops, VLIB_FRAME_SIZE); + n_deq = n_cops; + next_index = def_next_index; + + qpd->inflights -= n_cops; + ASSERT (qpd->inflights >= 0); + + while (n_cops > 0) + { + u32 n_left_to_next; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_cops > 0 && n_left_to_next > 0) + { + u32 bi0, next0; + vlib_buffer_t *b0 = 0; + struct rte_crypto_op *cop; + struct rte_crypto_sym_op *sym_cop; + + cop = cops[0]; + cops += 1; + n_cops -= 1; + n_left_to_next -= 1; + + next0 = def_next_index; + + if (PREDICT_FALSE (cop->status != RTE_CRYPTO_OP_STATUS_SUCCESS)) + { + next0 = DPDK_CRYPTO_INPUT_NEXT_DROP; + vlib_node_increment_counter (vm, dpdk_crypto_input_node.index, + DPDK_CRYPTO_INPUT_ERROR_COP_FAILED, + 1); + } + cop->status = RTE_CRYPTO_OP_STATUS_NOT_PROCESSED; + + sym_cop = (struct rte_crypto_sym_op *) (cop + 1); + b0 = vlib_buffer_from_rte_mbuf (sym_cop->m_src); + bi0 = vlib_get_buffer_index (vm, b0); + + to_next[0] = bi0; + to_next += 1; + + if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED)) + { + vlib_trace_next_frame (vm, node, next0); + dpdk_crypto_input_trace_t *tr = + vlib_add_trace (vm, node, b0, sizeof (*tr)); + tr->cdev = qpd->dev_id; + tr->qp = qpd->qp_id; + tr->status = cop->status; + tr->next_index = next0; + tr->sa_idx = vnet_buffer (b0)->ipsec.sad_index; + } + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next, + n_left_to_next, bi0, next0); + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + + crypto_free_cop (qpd, qpd->cops, n_deq); + + vlib_node_increment_counter (vm, dpdk_crypto_input_node.index, + DPDK_CRYPTO_INPUT_ERROR_DQ_COPS, n_deq); + return n_deq; +} + +static uword +dpdk_crypto_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node, + vlib_frame_t * frame) +{ + u32 cpu_index = os_get_cpu_number (); + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index]; + crypto_qp_data_t *qpd; + u32 n_deq = 0; + + /* *INDENT-OFF* */ + vec_foreach (qpd, cwm->qp_data) + n_deq += dpdk_crypto_dequeue(vm, node, qpd); + /* *INDENT-ON* */ + + return n_deq; +} + +VLIB_REGISTER_NODE (dpdk_crypto_input_node) = +{ + .function = dpdk_crypto_input_fn,.name = "dpdk-crypto-input",.format_trace = + format_dpdk_crypto_input_trace,.type = VLIB_NODE_TYPE_INPUT,.state = + VLIB_NODE_STATE_DISABLED,.n_errors = + DPDK_CRYPTO_INPUT_N_ERROR,.error_strings = + dpdk_crypto_input_error_strings,.n_next_nodes = + DPDK_CRYPTO_INPUT_N_NEXT,.next_nodes = + { +#define _(s,n) [DPDK_CRYPTO_INPUT_NEXT_##s] = n, + foreach_dpdk_crypto_input_next +#undef _ + } +,}; + +#if DPDK_CRYPTO==1 +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_crypto_input_node, dpdk_crypto_input_fn) +#endif +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/dir.dox b/src/vnet/devices/dpdk/ipsec/dir.dox new file mode 100644 index 00000000000..ffebfc4d62e --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/dir.dox @@ -0,0 +1,18 @@ +/* + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +/** + @dir vnet/vnet/devices/dpdk/ipsec + @brief IPSec ESP encrypt/decrypt using DPDK Cryptodev API +*/ diff --git a/src/vnet/devices/dpdk/ipsec/dpdk_crypto_ipsec_doc.md b/src/vnet/devices/dpdk/ipsec/dpdk_crypto_ipsec_doc.md new file mode 100644 index 00000000000..8089696f4a0 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/dpdk_crypto_ipsec_doc.md @@ -0,0 +1,73 @@ +# VPP IPSec implementation using DPDK Cryptodev API {#dpdk_crypto_ipsec_doc} + +This document is meant to contain all related information about implementation and usability. + + +## VPP IPsec with DPDK Cryptodev + +DPDK Cryptodev is an asynchronous crypto API that supports both Hardware and Software implementations (for more details refer to [DPDK Cryptography Device Library documentation](http://dpdk.org/doc/guides/prog_guide/cryptodev_lib.html)). + +When DPDK Cryptodev support is enabled, the node graph is modified by adding and replacing some of the nodes. + +The following nodes are replaced: +* esp-encrypt -> dpdk-esp-encrypt +* esp-decrypt -> dpdk-esp-decrypt + +The following nodes are added: +* dpdk-crypto-input : polling input node, basically dequeuing from crypto devices. +* dpdk-esp-encrypt-post : internal node. +* dpdk-esp-decrypt-post : internal node. + + +### How to enable VPP IPSec with DPDK Cryptodev support + +To enable DPDK Cryptodev support (disabled by default), we need the following env option: + + vpp_uses_dpdk_cryptodev=yes + +A couple of ways to achive this: +* uncomment/add it in the platforms config (ie. build-data/platforms/vpp.mk) +* set the option when building vpp (ie. make vpp_uses_dpdk_cryptodev=yes build-release) + + +### Crypto Resources allocation + +VPP allocates crypto resources based on a best effort approach: +* first allocate Hardware crypto resources, then Software. +* if there are not enough crypto resources for all workers, all packets will be dropped if they reach ESP encrypt/decrypt nodes, displaying the warning: + + 0: dpdk_ipsec_init: not enough cryptodevs for ipsec + + +### Configuration example + +No especial IPsec configuration is required. + +Once DPDK Cryptodev is enabled, the user just needs to provide cryptodevs in the startup.conf. + +Example startup.conf: + +``` +dpdk { + socket-mem 1024,1024 + num-mbufs 131072 + dev 0000:81:00.0 + dev 0000:81:00.1 + dev 0000:85:01.0 + dev 0000:85:01.1 + vdev cryptodev_aesni_mb_pmd,socket_id=1 + vdev cryptodev_aesni_mb_pmd,socket_id=1 +} +``` + +In the above configuration: +* 0000:85:01.0 and 0000:85:01.1 are crypto BDFs and they require the same driver binding as DPDK Ethernet devices but they do not support any extra configuration options. +* Two AESNI-MB Software Cryptodev PMDs are created in NUMA node 1. + +For further details refer to [DPDK Crypto Device Driver documentation](http://dpdk.org/doc/guides/cryptodevs/index.html) + +### Operational data + +The following CLI command displays the Cryptodev/Worker mapping: + + show crypto device mapping [verbose] diff --git a/src/vnet/devices/dpdk/ipsec/esp.h b/src/vnet/devices/dpdk/ipsec/esp.h new file mode 100644 index 00000000000..7ef90c49816 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/esp.h @@ -0,0 +1,295 @@ +/* + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#ifndef __DPDK_ESP_H__ +#define __DPDK_ESP_H__ + +#include +#include +#include + +typedef struct +{ + enum rte_crypto_cipher_algorithm algo; + u8 key_len; + u8 iv_len; +} dpdk_esp_crypto_alg_t; + +typedef struct +{ + enum rte_crypto_auth_algorithm algo; + u8 trunc_size; +} dpdk_esp_integ_alg_t; + +typedef struct +{ + dpdk_esp_crypto_alg_t *esp_crypto_algs; + dpdk_esp_integ_alg_t *esp_integ_algs; +} dpdk_esp_main_t; + +dpdk_esp_main_t dpdk_esp_main; + +static_always_inline void +dpdk_esp_init () +{ + dpdk_esp_main_t *em = &dpdk_esp_main; + dpdk_esp_integ_alg_t *i; + dpdk_esp_crypto_alg_t *c; + + vec_validate (em->esp_crypto_algs, IPSEC_CRYPTO_N_ALG - 1); + + c = &em->esp_crypto_algs[IPSEC_CRYPTO_ALG_AES_CBC_128]; + c->algo = RTE_CRYPTO_CIPHER_AES_CBC; + c->key_len = 16; + c->iv_len = 16; + + c = &em->esp_crypto_algs[IPSEC_CRYPTO_ALG_AES_CBC_192]; + c->algo = RTE_CRYPTO_CIPHER_AES_CBC; + c->key_len = 24; + c->iv_len = 16; + + c = &em->esp_crypto_algs[IPSEC_CRYPTO_ALG_AES_CBC_256]; + c->algo = RTE_CRYPTO_CIPHER_AES_CBC; + c->key_len = 32; + c->iv_len = 16; + + c = &em->esp_crypto_algs[IPSEC_CRYPTO_ALG_AES_GCM_128]; + c->algo = RTE_CRYPTO_CIPHER_AES_GCM; + c->key_len = 16; + c->iv_len = 8; + + vec_validate (em->esp_integ_algs, IPSEC_INTEG_N_ALG - 1); + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_SHA1_96]; + i->algo = RTE_CRYPTO_AUTH_SHA1_HMAC; + i->trunc_size = 12; + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_SHA_256_96]; + i->algo = RTE_CRYPTO_AUTH_SHA256_HMAC; + i->trunc_size = 12; + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_SHA_256_128]; + i->algo = RTE_CRYPTO_AUTH_SHA256_HMAC; + i->trunc_size = 16; + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_SHA_384_192]; + i->algo = RTE_CRYPTO_AUTH_SHA384_HMAC; + i->trunc_size = 24; + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_SHA_512_256]; + i->algo = RTE_CRYPTO_AUTH_SHA512_HMAC; + i->trunc_size = 32; + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_AES_GCM_128]; + i->algo = RTE_CRYPTO_AUTH_AES_GCM; + i->trunc_size = 16; +} + +static_always_inline int +add_del_sa_sess (u32 sa_index, u8 is_add) +{ + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + crypto_worker_main_t *cwm; + u8 skip_master = vlib_num_workers () > 0; + + /* *INDENT-OFF* */ + vec_foreach (cwm, dcm->workers_main) + { + crypto_sa_session_t *sa_sess; + u8 is_outbound; + + if (skip_master) + { + skip_master = 0; + continue; + } + + for (is_outbound = 0; is_outbound < 2; is_outbound++) + { + if (is_add) + { + pool_get (cwm->sa_sess_d[is_outbound], sa_sess); + } + else + { + u8 dev_id; + + sa_sess = pool_elt_at_index (cwm->sa_sess_d[is_outbound], sa_index); + dev_id = cwm->qp_data[sa_sess->qp_index].dev_id; + + if (!sa_sess->sess) + continue; + + if (rte_cryptodev_sym_session_free(dev_id, sa_sess->sess)) + { + clib_warning("failed to free session"); + return -1; + } + memset(sa_sess, 0, sizeof(sa_sess[0])); + } + } + } + /* *INDENT-OFF* */ + + return 0; +} + +static_always_inline int +translate_crypto_algo(ipsec_crypto_alg_t crypto_algo, + struct rte_crypto_sym_xform *cipher_xform) +{ + switch (crypto_algo) + { + case IPSEC_CRYPTO_ALG_NONE: + cipher_xform->cipher.algo = RTE_CRYPTO_CIPHER_NULL; + break; + case IPSEC_CRYPTO_ALG_AES_CBC_128: + case IPSEC_CRYPTO_ALG_AES_CBC_192: + case IPSEC_CRYPTO_ALG_AES_CBC_256: + cipher_xform->cipher.algo = RTE_CRYPTO_CIPHER_AES_CBC; + break; + case IPSEC_CRYPTO_ALG_AES_GCM_128: + cipher_xform->cipher.algo = RTE_CRYPTO_CIPHER_AES_GCM; + break; + default: + return -1; + } + + cipher_xform->type = RTE_CRYPTO_SYM_XFORM_CIPHER; + + return 0; +} + +static_always_inline int +translate_integ_algo(ipsec_integ_alg_t integ_alg, + struct rte_crypto_sym_xform *auth_xform, int use_esn) +{ + switch (integ_alg) { + case IPSEC_INTEG_ALG_NONE: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_NULL; + auth_xform->auth.digest_length = 0; + break; + case IPSEC_INTEG_ALG_SHA1_96: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_SHA1_HMAC; + auth_xform->auth.digest_length = 12; + break; + case IPSEC_INTEG_ALG_SHA_256_96: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_SHA256_HMAC; + auth_xform->auth.digest_length = 12; + break; + case IPSEC_INTEG_ALG_SHA_256_128: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_SHA256_HMAC; + auth_xform->auth.digest_length = 16; + break; + case IPSEC_INTEG_ALG_SHA_384_192: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_SHA384_HMAC; + auth_xform->auth.digest_length = 24; + break; + case IPSEC_INTEG_ALG_SHA_512_256: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_SHA512_HMAC; + auth_xform->auth.digest_length = 32; + break; + case IPSEC_INTEG_ALG_AES_GCM_128: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_AES_GCM; + auth_xform->auth.digest_length = 16; + auth_xform->auth.add_auth_data_length = use_esn? 12 : 8; + break; + default: + return -1; + } + + auth_xform->type = RTE_CRYPTO_SYM_XFORM_AUTH; + + return 0; +} + +static_always_inline int +create_sym_sess(ipsec_sa_t *sa, crypto_sa_session_t *sa_sess, u8 is_outbound) +{ + u32 cpu_index = os_get_cpu_number(); + dpdk_crypto_main_t * dcm = &dpdk_crypto_main; + crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index]; + struct rte_crypto_sym_xform cipher_xform = {0}; + struct rte_crypto_sym_xform auth_xform = {0}; + struct rte_crypto_sym_xform *xfs; + uword key = 0, *data; + crypto_worker_qp_key_t *p_key = (crypto_worker_qp_key_t *)&key; + + if (sa->crypto_alg == IPSEC_CRYPTO_ALG_AES_GCM_128) + { + sa->crypto_key_len -= 4; + clib_memcpy(&sa->salt, &sa->crypto_key[sa->crypto_key_len], 4); + } + else + { + sa->salt = (u32) rand(); + } + + cipher_xform.type = RTE_CRYPTO_SYM_XFORM_CIPHER; + cipher_xform.cipher.key.data = sa->crypto_key; + cipher_xform.cipher.key.length = sa->crypto_key_len; + + auth_xform.type = RTE_CRYPTO_SYM_XFORM_AUTH; + auth_xform.auth.key.data = sa->integ_key; + auth_xform.auth.key.length = sa->integ_key_len; + + if (translate_crypto_algo(sa->crypto_alg, &cipher_xform) < 0) + return -1; + p_key->cipher_algo = cipher_xform.cipher.algo; + + if (translate_integ_algo(sa->integ_alg, &auth_xform, sa->use_esn) < 0) + return -1; + p_key->auth_algo = auth_xform.auth.algo; + + if (is_outbound) + { + cipher_xform.cipher.op = RTE_CRYPTO_CIPHER_OP_ENCRYPT; + auth_xform.auth.op = RTE_CRYPTO_AUTH_OP_GENERATE; + cipher_xform.next = &auth_xform; + xfs = &cipher_xform; + } + else + { + cipher_xform.cipher.op = RTE_CRYPTO_CIPHER_OP_DECRYPT; + auth_xform.auth.op = RTE_CRYPTO_AUTH_OP_VERIFY; + auth_xform.next = &cipher_xform; + xfs = &auth_xform; + } + + p_key->is_outbound = is_outbound; + + data = hash_get(cwm->algo_qp_map, key); + if (!data) + return -1; + + sa_sess->sess = + rte_cryptodev_sym_session_create(cwm->qp_data[*data].dev_id, xfs); + + if (!sa_sess->sess) + return -1; + + sa_sess->qp_index = (u8)*data; + + return 0; +} + +#endif /* __DPDK_ESP_H__ */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/esp_decrypt.c b/src/vnet/devices/dpdk/ipsec/esp_decrypt.c new file mode 100644 index 00000000000..89ab9f9bc43 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/esp_decrypt.c @@ -0,0 +1,583 @@ +/* + * esp_decrypt.c : IPSec ESP Decrypt node using DPDK Cryptodev + * + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include +#include +#include + +#include +#include +#include + +#define foreach_esp_decrypt_next \ +_(DROP, "error-drop") \ +_(IP4_INPUT, "ip4-input") \ +_(IP6_INPUT, "ip6-input") + +#define _(v, s) ESP_DECRYPT_NEXT_##v, +typedef enum { + foreach_esp_decrypt_next +#undef _ + ESP_DECRYPT_N_NEXT, +} esp_decrypt_next_t; + +#define foreach_esp_decrypt_error \ + _(RX_PKTS, "ESP pkts received") \ + _(DECRYPTION_FAILED, "ESP decryption failed") \ + _(REPLAY, "SA replayed packet") \ + _(NOT_IP, "Not IP packet (dropped)") \ + _(ENQ_FAIL, "Enqueue failed (buffer full)") \ + _(NO_CRYPTODEV, "Cryptodev not configured") \ + _(BAD_LEN, "Invalid ciphertext length") \ + _(UNSUPPORTED, "Cipher/Auth not supported") + + +typedef enum { +#define _(sym,str) ESP_DECRYPT_ERROR_##sym, + foreach_esp_decrypt_error +#undef _ + ESP_DECRYPT_N_ERROR, +} esp_decrypt_error_t; + +static char * esp_decrypt_error_strings[] = { +#define _(sym,string) string, + foreach_esp_decrypt_error +#undef _ +}; + +vlib_node_registration_t dpdk_esp_decrypt_node; + +typedef struct { + ipsec_crypto_alg_t crypto_alg; + ipsec_integ_alg_t integ_alg; +} esp_decrypt_trace_t; + +/* packet trace format function */ +static u8 * format_esp_decrypt_trace (u8 * s, va_list * args) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *); + esp_decrypt_trace_t * t = va_arg (*args, esp_decrypt_trace_t *); + + s = format (s, "esp: crypto %U integrity %U", + format_ipsec_crypto_alg, t->crypto_alg, + format_ipsec_integ_alg, t->integ_alg); + return s; +} + +static uword +dpdk_esp_decrypt_node_fn (vlib_main_t * vm, + vlib_node_runtime_t * node, + vlib_frame_t * from_frame) +{ + u32 n_left_from, *from, *to_next, next_index; + ipsec_main_t *im = &ipsec_main; + u32 cpu_index = os_get_cpu_number(); + dpdk_crypto_main_t * dcm = &dpdk_crypto_main; + dpdk_esp_main_t * em = &dpdk_esp_main; + u32 i; + + from = vlib_frame_vector_args (from_frame); + n_left_from = from_frame->n_vectors; + + if (PREDICT_FALSE(!dcm->workers_main)) + { + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_NO_CRYPTODEV, n_left_from); + vlib_buffer_free(vm, from, n_left_from); + return n_left_from; + } + + crypto_worker_main_t *cwm = vec_elt_at_index(dcm->workers_main, cpu_index); + u32 n_qps = vec_len(cwm->qp_data); + struct rte_crypto_op ** cops_to_enq[n_qps]; + u32 n_cop_qp[n_qps], * bi_to_enq[n_qps]; + + for (i = 0; i < n_qps; i++) + { + bi_to_enq[i] = cwm->qp_data[i].bi; + cops_to_enq[i] = cwm->qp_data[i].cops; + } + + memset(n_cop_qp, 0, n_qps * sizeof(u32)); + + crypto_alloc_cops(); + + next_index = ESP_DECRYPT_NEXT_DROP; + + while (n_left_from > 0) + { + u32 n_left_to_next; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_left_from > 0 && n_left_to_next > 0) + { + u32 bi0, sa_index0 = ~0, seq, icv_size, iv_size; + vlib_buffer_t * b0; + esp_header_t * esp0; + ipsec_sa_t * sa0; + struct rte_mbuf * mb0 = 0; + const int BLOCK_SIZE = 16; + crypto_sa_session_t * sa_sess; + void * sess; + u16 qp_index; + struct rte_crypto_op * cop = 0; + + bi0 = from[0]; + from += 1; + n_left_from -= 1; + + b0 = vlib_get_buffer (vm, bi0); + esp0 = vlib_buffer_get_current (b0); + + sa_index0 = vnet_buffer(b0)->ipsec.sad_index; + sa0 = pool_elt_at_index (im->sad, sa_index0); + + seq = clib_host_to_net_u32(esp0->seq); + + /* anti-replay check */ + if (sa0->use_anti_replay) + { + int rv = 0; + + if (PREDICT_TRUE(sa0->use_esn)) + rv = esp_replay_check_esn(sa0, seq); + else + rv = esp_replay_check(sa0, seq); + + if (PREDICT_FALSE(rv)) + { + clib_warning ("anti-replay SPI %u seq %u", sa0->spi, seq); + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_REPLAY, 1); + to_next[0] = bi0; + to_next += 1; + n_left_to_next -= 1; + goto trace; + } + } + + if (PREDICT_FALSE(sa0->integ_alg == IPSEC_INTEG_ALG_NONE) || + PREDICT_FALSE(sa0->crypto_alg == IPSEC_CRYPTO_ALG_NONE)) + { + clib_warning ("SPI %u : only cipher + auth supported", sa0->spi); + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_UNSUPPORTED, 1); + to_next[0] = bi0; + to_next += 1; + n_left_to_next -= 1; + goto trace; + } + + sa_sess = pool_elt_at_index(cwm->sa_sess_d[0], sa_index0); + + if (PREDICT_FALSE(!sa_sess->sess)) + { + int ret = create_sym_sess(sa0, sa_sess, 0); + ASSERT(ret == 0); + } + + sess = sa_sess->sess; + qp_index = sa_sess->qp_index; + + ASSERT (vec_len (vec_elt (cwm->qp_data, qp_index).free_cops) > 0); + cop = vec_pop (vec_elt (cwm->qp_data, qp_index).free_cops); + ASSERT (cop->status == RTE_CRYPTO_OP_STATUS_NOT_PROCESSED); + + cops_to_enq[qp_index][0] = cop; + cops_to_enq[qp_index] += 1; + n_cop_qp[qp_index] += 1; + bi_to_enq[qp_index][0] = bi0; + bi_to_enq[qp_index] += 1; + + rte_crypto_op_attach_sym_session(cop, sess); + + icv_size = em->esp_integ_algs[sa0->integ_alg].trunc_size; + iv_size = em->esp_crypto_algs[sa0->crypto_alg].iv_len; + + /* Convert vlib buffer to mbuf */ + mb0 = rte_mbuf_from_vlib_buffer(b0); + mb0->data_len = b0->current_length; + mb0->pkt_len = b0->current_length; + mb0->data_off = RTE_PKTMBUF_HEADROOM + b0->current_data; + + /* Outer IP header has already been stripped */ + u16 payload_len = rte_pktmbuf_pkt_len(mb0) - sizeof (esp_header_t) - + iv_size - icv_size; + + if ((payload_len & (BLOCK_SIZE - 1)) || (payload_len <= 0)) + { + clib_warning ("payload %u not multiple of %d\n", + payload_len, BLOCK_SIZE); + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_BAD_LEN, 1); + vec_add (vec_elt (cwm->qp_data, qp_index).free_cops, &cop, 1); + bi_to_enq[qp_index] -= 1; + cops_to_enq[qp_index] -= 1; + n_cop_qp[qp_index] -= 1; + to_next[0] = bi0; + to_next += 1; + n_left_to_next -= 1; + goto trace; + } + + struct rte_crypto_sym_op *sym_cop = (struct rte_crypto_sym_op *)(cop + 1); + + sym_cop->m_src = mb0; + sym_cop->cipher.data.offset = sizeof (esp_header_t) + iv_size; + sym_cop->cipher.data.length = payload_len; + + u8 *iv = rte_pktmbuf_mtod_offset(mb0, void*, sizeof (esp_header_t)); + dpdk_cop_priv_t * priv = (dpdk_cop_priv_t *)(sym_cop + 1); + + if (sa0->crypto_alg == IPSEC_CRYPTO_ALG_AES_GCM_128) + { + dpdk_gcm_cnt_blk *icb = &priv->cb; + icb->salt = sa0->salt; + clib_memcpy(icb->iv, iv, 8); + icb->cnt = clib_host_to_net_u32(1); + sym_cop->cipher.iv.data = (u8 *)icb; + sym_cop->cipher.iv.phys_addr = cop->phys_addr + + (uintptr_t)icb - (uintptr_t)cop; + sym_cop->cipher.iv.length = 16; + + u8 *aad = priv->aad; + clib_memcpy(aad, iv - sizeof(esp_header_t), 8); + sym_cop->auth.aad.data = aad; + sym_cop->auth.aad.phys_addr = cop->phys_addr + + (uintptr_t)aad - (uintptr_t)cop; + if (sa0->use_esn) + { + *((u32*)&aad[8]) = sa0->seq_hi; + sym_cop->auth.aad.length = 12; + } + else + { + sym_cop->auth.aad.length = 8; + } + + sym_cop->auth.digest.data = rte_pktmbuf_mtod_offset(mb0, void*, + rte_pktmbuf_pkt_len(mb0) - icv_size); + sym_cop->auth.digest.phys_addr = rte_pktmbuf_mtophys_offset(mb0, + rte_pktmbuf_pkt_len(mb0) - icv_size); + sym_cop->auth.digest.length = icv_size; + + } + else + { + sym_cop->cipher.iv.data = rte_pktmbuf_mtod_offset(mb0, void*, + sizeof (esp_header_t)); + sym_cop->cipher.iv.phys_addr = rte_pktmbuf_mtophys_offset(mb0, + sizeof (esp_header_t)); + sym_cop->cipher.iv.length = iv_size; + + if (sa0->use_esn) + { + dpdk_cop_priv_t* priv = (dpdk_cop_priv_t*) (sym_cop + 1); + u8* payload_end = rte_pktmbuf_mtod_offset( + mb0, u8*, sizeof(esp_header_t) + iv_size + payload_len); + + clib_memcpy (priv->icv, payload_end, icv_size); + *((u32*) payload_end) = sa0->seq_hi; + sym_cop->auth.data.offset = 0; + sym_cop->auth.data.length = sizeof(esp_header_t) + iv_size + + payload_len + sizeof(sa0->seq_hi); + sym_cop->auth.digest.data = priv->icv; + sym_cop->auth.digest.phys_addr = cop->phys_addr + + (uintptr_t) priv->icv - (uintptr_t) cop; + sym_cop->auth.digest.length = icv_size; + } + else + { + sym_cop->auth.data.offset = 0; + sym_cop->auth.data.length = sizeof(esp_header_t) + + iv_size + payload_len; + + sym_cop->auth.digest.data = rte_pktmbuf_mtod_offset(mb0, void*, + rte_pktmbuf_pkt_len(mb0) - icv_size); + sym_cop->auth.digest.phys_addr = rte_pktmbuf_mtophys_offset(mb0, + rte_pktmbuf_pkt_len(mb0) - icv_size); + sym_cop->auth.digest.length = icv_size; + } + } + +trace: + if (PREDICT_FALSE(b0->flags & VLIB_BUFFER_IS_TRACED)) + { + esp_decrypt_trace_t *tr = vlib_add_trace (vm, node, b0, sizeof (*tr)); + tr->crypto_alg = sa0->crypto_alg; + tr->integ_alg = sa0->integ_alg; + } + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_RX_PKTS, + from_frame->n_vectors); + crypto_qp_data_t *qpd; + /* *INDENT-OFF* */ + vec_foreach_index (i, cwm->qp_data) + { + u32 enq; + + qpd = vec_elt_at_index(cwm->qp_data, i); + enq = rte_cryptodev_enqueue_burst(qpd->dev_id, qpd->qp_id, + qpd->cops, n_cop_qp[i]); + qpd->inflights += enq; + + if (PREDICT_FALSE(enq < n_cop_qp[i])) + { + crypto_free_cop (qpd, &qpd->cops[enq], n_cop_qp[i] - enq); + vlib_buffer_free (vm, &qpd->bi[enq], n_cop_qp[i] - enq); + + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_ENQ_FAIL, + n_cop_qp[i] - enq); + } + } + /* *INDENT-ON* */ + + return from_frame->n_vectors; +} + +VLIB_REGISTER_NODE (dpdk_esp_decrypt_node) = { + .function = dpdk_esp_decrypt_node_fn, + .name = "dpdk-esp-decrypt", + .vector_size = sizeof (u32), + .format_trace = format_esp_decrypt_trace, + .type = VLIB_NODE_TYPE_INTERNAL, + + .n_errors = ARRAY_LEN(esp_decrypt_error_strings), + .error_strings = esp_decrypt_error_strings, + + .n_next_nodes = ESP_DECRYPT_N_NEXT, + .next_nodes = { +#define _(s,n) [ESP_DECRYPT_NEXT_##s] = n, + foreach_esp_decrypt_next +#undef _ + }, +}; + +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_esp_decrypt_node, dpdk_esp_decrypt_node_fn) + +/* + * Decrypt Post Node + */ + +#define foreach_esp_decrypt_post_error \ + _(PKTS, "ESP post pkts") + +typedef enum { +#define _(sym,str) ESP_DECRYPT_POST_ERROR_##sym, + foreach_esp_decrypt_post_error +#undef _ + ESP_DECRYPT_POST_N_ERROR, +} esp_decrypt_post_error_t; + +static char * esp_decrypt_post_error_strings[] = { +#define _(sym,string) string, + foreach_esp_decrypt_post_error +#undef _ +}; + +vlib_node_registration_t dpdk_esp_decrypt_post_node; + +static u8 * format_esp_decrypt_post_trace (u8 * s, va_list * args) +{ + return s; +} + +static uword +dpdk_esp_decrypt_post_node_fn (vlib_main_t * vm, + vlib_node_runtime_t * node, + vlib_frame_t * from_frame) +{ + u32 n_left_from, *from, *to_next = 0, next_index; + ipsec_sa_t * sa0; + u32 sa_index0 = ~0; + ipsec_main_t *im = &ipsec_main; + dpdk_esp_main_t *em = &dpdk_esp_main; + + from = vlib_frame_vector_args (from_frame); + n_left_from = from_frame->n_vectors; + + next_index = node->cached_next_index; + + while (n_left_from > 0) + { + u32 n_left_to_next; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_left_from > 0 && n_left_to_next > 0) + { + esp_footer_t * f0; + u32 bi0, next0, icv_size, iv_size; + vlib_buffer_t * b0 = 0; + ip4_header_t *ih4 = 0, *oh4 = 0; + ip6_header_t *ih6 = 0, *oh6 = 0; + u8 tunnel_mode = 1; + u8 transport_ip6 = 0; + + next0 = ESP_DECRYPT_NEXT_DROP; + + bi0 = from[0]; + from += 1; + n_left_from -= 1; + n_left_to_next -= 1; + + b0 = vlib_get_buffer (vm, bi0); + + sa_index0 = vnet_buffer(b0)->ipsec.sad_index; + sa0 = pool_elt_at_index (im->sad, sa_index0); + + to_next[0] = bi0; + to_next += 1; + + icv_size = em->esp_integ_algs[sa0->integ_alg].trunc_size; + iv_size = em->esp_crypto_algs[sa0->crypto_alg].iv_len; + + if (sa0->use_anti_replay) + { + esp_header_t * esp0 = vlib_buffer_get_current (b0); + u32 seq; + seq = clib_host_to_net_u32(esp0->seq); + if (PREDICT_TRUE(sa0->use_esn)) + esp_replay_advance_esn(sa0, seq); + else + esp_replay_advance(sa0, seq); + } + + ih4 = (ip4_header_t *) (b0->data + sizeof(ethernet_header_t)); + vlib_buffer_advance (b0, sizeof (esp_header_t) + iv_size); + + b0->current_length -= (icv_size + 2); + b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID; + f0 = (esp_footer_t *) ((u8 *) vlib_buffer_get_current (b0) + + b0->current_length); + b0->current_length -= f0->pad_length; + + /* transport mode */ + if (PREDICT_FALSE(!sa0->is_tunnel && !sa0->is_tunnel_ip6)) + { + tunnel_mode = 0; + + if (PREDICT_TRUE((ih4->ip_version_and_header_length & 0xF0) != 0x40)) + { + if (PREDICT_TRUE((ih4->ip_version_and_header_length & 0xF0) == 0x60)) + transport_ip6 = 1; + else + { + clib_warning("next header: 0x%x", f0->next_header); + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_NOT_IP, 1); + goto trace; + } + } + } + + if (PREDICT_TRUE (tunnel_mode)) + { + if (PREDICT_TRUE(f0->next_header == IP_PROTOCOL_IP_IN_IP)) + next0 = ESP_DECRYPT_NEXT_IP4_INPUT; + else if (f0->next_header == IP_PROTOCOL_IPV6) + next0 = ESP_DECRYPT_NEXT_IP6_INPUT; + else + { + clib_warning("next header: 0x%x", f0->next_header); + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_DECRYPTION_FAILED, + 1); + goto trace; + } + } + /* transport mode */ + else + { + if (PREDICT_FALSE(transport_ip6)) + { + ih6 = (ip6_header_t *) (b0->data + sizeof(ethernet_header_t)); + vlib_buffer_advance (b0, -sizeof(ip6_header_t)); + oh6 = vlib_buffer_get_current (b0); + memmove(oh6, ih6, sizeof(ip6_header_t)); + + next0 = ESP_DECRYPT_NEXT_IP6_INPUT; + oh6->protocol = f0->next_header; + oh6->payload_length = + clib_host_to_net_u16 ( + vlib_buffer_length_in_chain(vm, b0) - + sizeof (ip6_header_t)); + } + else + { + vlib_buffer_advance (b0, -sizeof(ip4_header_t)); + oh4 = vlib_buffer_get_current (b0); + memmove(oh4, ih4, sizeof(ip4_header_t)); + + next0 = ESP_DECRYPT_NEXT_IP4_INPUT; + oh4->ip_version_and_header_length = 0x45; + oh4->fragment_id = 0; + oh4->flags_and_fragment_offset = 0; + oh4->protocol = f0->next_header; + oh4->length = clib_host_to_net_u16 ( + vlib_buffer_length_in_chain (vm, b0)); + oh4->checksum = ip4_header_checksum (oh4); + } + } + + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32)~0; + +trace: + if (PREDICT_FALSE(b0->flags & VLIB_BUFFER_IS_TRACED)) + { + esp_decrypt_trace_t *tr = vlib_add_trace (vm, node, b0, sizeof (*tr)); + tr->crypto_alg = sa0->crypto_alg; + tr->integ_alg = sa0->integ_alg; + } + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, bi0, next0); + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + vlib_node_increment_counter (vm, dpdk_esp_decrypt_post_node.index, + ESP_DECRYPT_POST_ERROR_PKTS, + from_frame->n_vectors); + + return from_frame->n_vectors; +} + +VLIB_REGISTER_NODE (dpdk_esp_decrypt_post_node) = { + .function = dpdk_esp_decrypt_post_node_fn, + .name = "dpdk-esp-decrypt-post", + .vector_size = sizeof (u32), + .format_trace = format_esp_decrypt_post_trace, + .type = VLIB_NODE_TYPE_INTERNAL, + + .n_errors = ARRAY_LEN(esp_decrypt_post_error_strings), + .error_strings = esp_decrypt_post_error_strings, + + .n_next_nodes = ESP_DECRYPT_N_NEXT, + .next_nodes = { +#define _(s,n) [ESP_DECRYPT_NEXT_##s] = n, + foreach_esp_decrypt_next +#undef _ + }, +}; + +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_esp_decrypt_post_node, dpdk_esp_decrypt_post_node_fn) diff --git a/src/vnet/devices/dpdk/ipsec/esp_encrypt.c b/src/vnet/devices/dpdk/ipsec/esp_encrypt.c new file mode 100644 index 00000000000..10bb4616eef --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/esp_encrypt.c @@ -0,0 +1,598 @@ +/* + * esp_encrypt.c : IPSec ESP encrypt node using DPDK Cryptodev + * + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include +#include +#include + +#include +#include +#include + +#define foreach_esp_encrypt_next \ +_(DROP, "error-drop") \ +_(IP4_LOOKUP, "ip4-lookup") \ +_(IP6_LOOKUP, "ip6-lookup") \ +_(INTERFACE_OUTPUT, "interface-output") + +#define _(v, s) ESP_ENCRYPT_NEXT_##v, +typedef enum +{ + foreach_esp_encrypt_next +#undef _ + ESP_ENCRYPT_N_NEXT, +} esp_encrypt_next_t; + +#define foreach_esp_encrypt_error \ + _(RX_PKTS, "ESP pkts received") \ + _(SEQ_CYCLED, "sequence number cycled") \ + _(ENQ_FAIL, "Enqueue failed (buffer full)") \ + _(NO_CRYPTODEV, "Cryptodev not configured") \ + _(UNSUPPORTED, "Cipher/Auth not supported") + + +typedef enum +{ +#define _(sym,str) ESP_ENCRYPT_ERROR_##sym, + foreach_esp_encrypt_error +#undef _ + ESP_ENCRYPT_N_ERROR, +} esp_encrypt_error_t; + +static char *esp_encrypt_error_strings[] = { +#define _(sym,string) string, + foreach_esp_encrypt_error +#undef _ +}; + +vlib_node_registration_t dpdk_esp_encrypt_node; + +typedef struct +{ + u32 spi; + u32 seq; + ipsec_crypto_alg_t crypto_alg; + ipsec_integ_alg_t integ_alg; +} esp_encrypt_trace_t; + +/* packet trace format function */ +static u8 * +format_esp_encrypt_trace (u8 * s, va_list * args) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *); + esp_encrypt_trace_t *t = va_arg (*args, esp_encrypt_trace_t *); + + s = format (s, "esp: spi %u seq %u crypto %U integrity %U", + t->spi, t->seq, + format_ipsec_crypto_alg, t->crypto_alg, + format_ipsec_integ_alg, t->integ_alg); + return s; +} + +static uword +dpdk_esp_encrypt_node_fn (vlib_main_t * vm, + vlib_node_runtime_t * node, + vlib_frame_t * from_frame) +{ + u32 n_left_from, *from, *to_next, next_index; + ipsec_main_t *im = &ipsec_main; + u32 cpu_index = os_get_cpu_number (); + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + dpdk_esp_main_t *em = &dpdk_esp_main; + u32 i; + + from = vlib_frame_vector_args (from_frame); + n_left_from = from_frame->n_vectors; + + if (PREDICT_FALSE (!dcm->workers_main)) + { + /* Likely there are not enough cryptodevs, so drop frame */ + vlib_node_increment_counter (vm, dpdk_esp_encrypt_node.index, + ESP_ENCRYPT_ERROR_NO_CRYPTODEV, + n_left_from); + vlib_buffer_free (vm, from, n_left_from); + return n_left_from; + } + + crypto_worker_main_t *cwm = vec_elt_at_index (dcm->workers_main, cpu_index); + u32 n_qps = vec_len (cwm->qp_data); + struct rte_crypto_op **cops_to_enq[n_qps]; + u32 n_cop_qp[n_qps], *bi_to_enq[n_qps]; + + for (i = 0; i < n_qps; i++) + { + bi_to_enq[i] = cwm->qp_data[i].bi; + cops_to_enq[i] = cwm->qp_data[i].cops; + } + + memset (n_cop_qp, 0, n_qps * sizeof (u32)); + + crypto_alloc_cops (); + + next_index = ESP_ENCRYPT_NEXT_DROP; + + while (n_left_from > 0) + { + u32 n_left_to_next; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_left_from > 0 && n_left_to_next > 0) + { + u32 bi0, next0; + vlib_buffer_t *b0 = 0; + u32 sa_index0; + ipsec_sa_t *sa0; + ip4_and_esp_header_t *ih0, *oh0 = 0; + ip6_and_esp_header_t *ih6_0, *oh6_0 = 0; + struct rte_mbuf *mb0 = 0; + esp_footer_t *f0; + u8 is_ipv6; + u8 ip_hdr_size; + u8 next_hdr_type; + u8 transport_mode = 0; + const int BLOCK_SIZE = 16; + u32 iv_size; + u16 orig_sz; + crypto_sa_session_t *sa_sess; + void *sess; + struct rte_crypto_op *cop = 0; + u16 qp_index; + + bi0 = from[0]; + from += 1; + n_left_from -= 1; + + b0 = vlib_get_buffer (vm, bi0); + sa_index0 = vnet_buffer (b0)->ipsec.sad_index; + sa0 = pool_elt_at_index (im->sad, sa_index0); + + if (PREDICT_FALSE (esp_seq_advance (sa0))) + { + clib_warning ("sequence number counter has cycled SPI %u", + sa0->spi); + vlib_node_increment_counter (vm, dpdk_esp_encrypt_node.index, + ESP_ENCRYPT_ERROR_SEQ_CYCLED, 1); + //TODO: rekey SA + to_next[0] = bi0; + to_next += 1; + n_left_to_next -= 1; + goto trace; + } + + sa_sess = pool_elt_at_index (cwm->sa_sess_d[1], sa_index0); + if (PREDICT_FALSE (!sa_sess->sess)) + { + int ret = create_sym_sess (sa0, sa_sess, 1); + ASSERT (ret == 0); + } + + qp_index = sa_sess->qp_index; + sess = sa_sess->sess; + + ASSERT (vec_len (vec_elt (cwm->qp_data, qp_index).free_cops) > 0); + cop = vec_pop (vec_elt (cwm->qp_data, qp_index).free_cops); + ASSERT (cop->status == RTE_CRYPTO_OP_STATUS_NOT_PROCESSED); + + cops_to_enq[qp_index][0] = cop; + cops_to_enq[qp_index] += 1; + n_cop_qp[qp_index] += 1; + bi_to_enq[qp_index][0] = bi0; + bi_to_enq[qp_index] += 1; + + ssize_t adv; + iv_size = em->esp_crypto_algs[sa0->crypto_alg].iv_len; + ih0 = vlib_buffer_get_current (b0); + orig_sz = b0->current_length; + is_ipv6 = (ih0->ip4.ip_version_and_header_length & 0xF0) == 0x60; + /* is ipv6 */ + if (PREDICT_TRUE (sa0->is_tunnel)) + { + if (PREDICT_TRUE (!is_ipv6)) + adv = -sizeof (ip4_and_esp_header_t); + else + adv = -sizeof (ip6_and_esp_header_t); + } + else + { + adv = -sizeof (esp_header_t); + if (PREDICT_TRUE (!is_ipv6)) + orig_sz -= sizeof (ip4_header_t); + else + orig_sz -= sizeof (ip6_header_t); + } + + /*transport mode save the eth header before it is overwritten */ + if (PREDICT_FALSE (!sa0->is_tunnel)) + { + ethernet_header_t *ieh0 = (ethernet_header_t *) + ((u8 *) vlib_buffer_get_current (b0) - + sizeof (ethernet_header_t)); + ethernet_header_t *oeh0 = + (ethernet_header_t *) ((u8 *) ieh0 + (adv - iv_size)); + clib_memcpy (oeh0, ieh0, sizeof (ethernet_header_t)); + } + + vlib_buffer_advance (b0, adv - iv_size); + + /* XXX IP6/ip4 and IP4/IP6 not supported, only IP4/IP4 and IP6/IP6 */ + + /* is ipv6 */ + if (PREDICT_FALSE (is_ipv6)) + { + ih6_0 = (ip6_and_esp_header_t *) ih0; + ip_hdr_size = sizeof (ip6_header_t); + oh6_0 = vlib_buffer_get_current (b0); + + if (PREDICT_TRUE (sa0->is_tunnel)) + { + next_hdr_type = IP_PROTOCOL_IPV6; + oh6_0->ip6.ip_version_traffic_class_and_flow_label = + ih6_0->ip6.ip_version_traffic_class_and_flow_label; + } + else + { + next_hdr_type = ih6_0->ip6.protocol; + memmove (oh6_0, ih6_0, sizeof (ip6_header_t)); + } + + oh6_0->ip6.protocol = IP_PROTOCOL_IPSEC_ESP; + oh6_0->ip6.hop_limit = 254; + oh6_0->esp.spi = clib_net_to_host_u32 (sa0->spi); + oh6_0->esp.seq = clib_net_to_host_u32 (sa0->seq); + } + else + { + ip_hdr_size = sizeof (ip4_header_t); + oh0 = vlib_buffer_get_current (b0); + + if (PREDICT_TRUE (sa0->is_tunnel)) + { + next_hdr_type = IP_PROTOCOL_IP_IN_IP; + oh0->ip4.tos = ih0->ip4.tos; + } + else + { + next_hdr_type = ih0->ip4.protocol; + memmove (oh0, ih0, sizeof (ip4_header_t)); + } + + oh0->ip4.ip_version_and_header_length = 0x45; + oh0->ip4.fragment_id = 0; + oh0->ip4.flags_and_fragment_offset = 0; + oh0->ip4.ttl = 254; + oh0->ip4.protocol = IP_PROTOCOL_IPSEC_ESP; + oh0->esp.spi = clib_net_to_host_u32 (sa0->spi); + oh0->esp.seq = clib_net_to_host_u32 (sa0->seq); + } + + if (PREDICT_TRUE (sa0->is_tunnel && !sa0->is_tunnel_ip6)) + { + oh0->ip4.src_address.as_u32 = sa0->tunnel_src_addr.ip4.as_u32; + oh0->ip4.dst_address.as_u32 = sa0->tunnel_dst_addr.ip4.as_u32; + + /* in tunnel mode send it back to FIB */ + next0 = ESP_ENCRYPT_NEXT_IP4_LOOKUP; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + } + else if (sa0->is_tunnel && sa0->is_tunnel_ip6) + { + oh6_0->ip6.src_address.as_u64[0] = + sa0->tunnel_src_addr.ip6.as_u64[0]; + oh6_0->ip6.src_address.as_u64[1] = + sa0->tunnel_src_addr.ip6.as_u64[1]; + oh6_0->ip6.dst_address.as_u64[0] = + sa0->tunnel_dst_addr.ip6.as_u64[0]; + oh6_0->ip6.dst_address.as_u64[1] = + sa0->tunnel_dst_addr.ip6.as_u64[1]; + + /* in tunnel mode send it back to FIB */ + next0 = ESP_ENCRYPT_NEXT_IP6_LOOKUP; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + } + else + { + next0 = ESP_ENCRYPT_NEXT_INTERFACE_OUTPUT; + transport_mode = 1; + } + + ASSERT (sa0->crypto_alg < IPSEC_CRYPTO_N_ALG); + ASSERT (sa0->crypto_alg != IPSEC_CRYPTO_ALG_NONE); + + int blocks = 1 + (orig_sz + 1) / BLOCK_SIZE; + + /* pad packet in input buffer */ + u8 pad_bytes = BLOCK_SIZE * blocks - 2 - orig_sz; + u8 i; + u8 *padding = vlib_buffer_get_current (b0) + b0->current_length; + + for (i = 0; i < pad_bytes; ++i) + padding[i] = i + 1; + + f0 = vlib_buffer_get_current (b0) + b0->current_length + pad_bytes; + f0->pad_length = pad_bytes; + f0->next_header = next_hdr_type; + b0->current_length += pad_bytes + 2 + + em->esp_integ_algs[sa0->integ_alg].trunc_size; + + vnet_buffer (b0)->sw_if_index[VLIB_RX] = + vnet_buffer (b0)->sw_if_index[VLIB_RX]; + b0->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID; + + struct rte_crypto_sym_op *sym_cop; + sym_cop = (struct rte_crypto_sym_op *) (cop + 1); + + dpdk_cop_priv_t *priv = (dpdk_cop_priv_t *) (sym_cop + 1); + + vnet_buffer (b0)->unused[0] = next0; + + mb0 = rte_mbuf_from_vlib_buffer (b0); + mb0->data_len = b0->current_length; + mb0->pkt_len = b0->current_length; + mb0->data_off = RTE_PKTMBUF_HEADROOM + b0->current_data; + + rte_crypto_op_attach_sym_session (cop, sess); + + sym_cop->m_src = mb0; + + dpdk_gcm_cnt_blk *icb = &priv->cb; + icb->salt = sa0->salt; + icb->iv[0] = sa0->seq; + icb->iv[1] = sa0->seq_hi; + + if (sa0->crypto_alg == IPSEC_CRYPTO_ALG_AES_GCM_128) + { + icb->cnt = clib_host_to_net_u32 (1); + clib_memcpy (vlib_buffer_get_current (b0) + ip_hdr_size + + sizeof (esp_header_t), icb->iv, 8); + sym_cop->cipher.data.offset = + ip_hdr_size + sizeof (esp_header_t) + iv_size; + sym_cop->cipher.data.length = BLOCK_SIZE * blocks; + sym_cop->cipher.iv.length = 16; + } + else + { + sym_cop->cipher.data.offset = + ip_hdr_size + sizeof (esp_header_t); + sym_cop->cipher.data.length = BLOCK_SIZE * blocks + iv_size; + sym_cop->cipher.iv.length = iv_size; + } + + sym_cop->cipher.iv.data = (u8 *) icb; + sym_cop->cipher.iv.phys_addr = cop->phys_addr + (uintptr_t) icb + - (uintptr_t) cop; + + + ASSERT (sa0->integ_alg < IPSEC_INTEG_N_ALG); + ASSERT (sa0->integ_alg != IPSEC_INTEG_ALG_NONE); + + if (PREDICT_FALSE (sa0->integ_alg == IPSEC_INTEG_ALG_AES_GCM_128)) + { + u8 *aad = priv->aad; + clib_memcpy (aad, vlib_buffer_get_current (b0) + ip_hdr_size, + 8); + sym_cop->auth.aad.data = aad; + sym_cop->auth.aad.phys_addr = cop->phys_addr + + (uintptr_t) aad - (uintptr_t) cop; + + if (PREDICT_FALSE (sa0->use_esn)) + { + *((u32 *) & aad[8]) = sa0->seq_hi; + sym_cop->auth.aad.length = 12; + } + else + { + sym_cop->auth.aad.length = 8; + } + } + else + { + sym_cop->auth.data.offset = ip_hdr_size; + sym_cop->auth.data.length = b0->current_length - ip_hdr_size + - em->esp_integ_algs[sa0->integ_alg].trunc_size; + + if (PREDICT_FALSE (sa0->use_esn)) + { + u8 *payload_end = + vlib_buffer_get_current (b0) + b0->current_length; + *((u32 *) payload_end) = sa0->seq_hi; + sym_cop->auth.data.length += sizeof (sa0->seq_hi); + } + } + sym_cop->auth.digest.data = vlib_buffer_get_current (b0) + + b0->current_length - + em->esp_integ_algs[sa0->integ_alg].trunc_size; + sym_cop->auth.digest.phys_addr = rte_pktmbuf_mtophys_offset (mb0, + b0->current_length + - + em->esp_integ_algs + [sa0->integ_alg].trunc_size); + sym_cop->auth.digest.length = + em->esp_integ_algs[sa0->integ_alg].trunc_size; + + + if (PREDICT_FALSE (is_ipv6)) + { + oh6_0->ip6.payload_length = + clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, b0) - + sizeof (ip6_header_t)); + } + else + { + oh0->ip4.length = + clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, b0)); + oh0->ip4.checksum = ip4_header_checksum (&oh0->ip4); + } + + if (transport_mode) + vlib_buffer_advance (b0, -sizeof (ethernet_header_t)); + + trace: + if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED)) + { + esp_encrypt_trace_t *tr = + vlib_add_trace (vm, node, b0, sizeof (*tr)); + tr->spi = sa0->spi; + tr->seq = sa0->seq - 1; + tr->crypto_alg = sa0->crypto_alg; + tr->integ_alg = sa0->integ_alg; + } + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + vlib_node_increment_counter (vm, dpdk_esp_encrypt_node.index, + ESP_ENCRYPT_ERROR_RX_PKTS, + from_frame->n_vectors); + crypto_qp_data_t *qpd; + /* *INDENT-OFF* */ + vec_foreach_index (i, cwm->qp_data) + { + u32 enq; + + qpd = vec_elt_at_index(cwm->qp_data, i); + enq = rte_cryptodev_enqueue_burst(qpd->dev_id, qpd->qp_id, + qpd->cops, n_cop_qp[i]); + qpd->inflights += enq; + + if (PREDICT_FALSE(enq < n_cop_qp[i])) + { + crypto_free_cop (qpd, &qpd->cops[enq], n_cop_qp[i] - enq); + vlib_buffer_free (vm, &qpd->bi[enq], n_cop_qp[i] - enq); + + vlib_node_increment_counter (vm, dpdk_esp_encrypt_node.index, + ESP_ENCRYPT_ERROR_ENQ_FAIL, + n_cop_qp[i] - enq); + } + } + /* *INDENT-ON* */ + + return from_frame->n_vectors; +} + +VLIB_REGISTER_NODE (dpdk_esp_encrypt_node) = +{ + .function = dpdk_esp_encrypt_node_fn,.name = "dpdk-esp-encrypt",.flags = + VLIB_NODE_FLAG_IS_OUTPUT,.vector_size = sizeof (u32),.format_trace = + format_esp_encrypt_trace,.n_errors = + ARRAY_LEN (esp_encrypt_error_strings),.error_strings = + esp_encrypt_error_strings,.n_next_nodes = 1,.next_nodes = + { + [ESP_ENCRYPT_NEXT_DROP] = "error-drop",} +}; + +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_esp_encrypt_node, dpdk_esp_encrypt_node_fn) +/* + * ESP Encrypt Post Node + */ +#define foreach_esp_encrypt_post_error \ + _(PKTS, "ESP post pkts") + typedef enum + { +#define _(sym,str) ESP_ENCRYPT_POST_ERROR_##sym, + foreach_esp_encrypt_post_error +#undef _ + ESP_ENCRYPT_POST_N_ERROR, + } esp_encrypt_post_error_t; + + static char *esp_encrypt_post_error_strings[] = { +#define _(sym,string) string, + foreach_esp_encrypt_post_error +#undef _ + }; + +vlib_node_registration_t dpdk_esp_encrypt_post_node; + +static u8 * +format_esp_encrypt_post_trace (u8 * s, va_list * args) +{ + return s; +} + +static uword +dpdk_esp_encrypt_post_node_fn (vlib_main_t * vm, + vlib_node_runtime_t * node, + vlib_frame_t * from_frame) +{ + u32 n_left_from, *from, *to_next = 0, next_index; + + from = vlib_frame_vector_args (from_frame); + n_left_from = from_frame->n_vectors; + + next_index = node->cached_next_index; + + while (n_left_from > 0) + { + u32 n_left_to_next; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_left_from > 0 && n_left_to_next > 0) + { + u32 bi0, next0; + vlib_buffer_t *b0 = 0; + + bi0 = from[0]; + from += 1; + n_left_from -= 1; + n_left_to_next -= 1; + + b0 = vlib_get_buffer (vm, bi0); + + to_next[0] = bi0; + to_next += 1; + + next0 = vnet_buffer (b0)->unused[0]; + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, bi0, + next0); + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + + vlib_node_increment_counter (vm, dpdk_esp_encrypt_post_node.index, + ESP_ENCRYPT_POST_ERROR_PKTS, + from_frame->n_vectors); + + return from_frame->n_vectors; +} + +VLIB_REGISTER_NODE (dpdk_esp_encrypt_post_node) = +{ + .function = dpdk_esp_encrypt_post_node_fn,.name = + "dpdk-esp-encrypt-post",.vector_size = sizeof (u32),.format_trace = + format_esp_encrypt_post_trace,.type = VLIB_NODE_TYPE_INTERNAL,.n_errors = + ARRAY_LEN (esp_encrypt_post_error_strings),.error_strings = + esp_encrypt_post_error_strings,.n_next_nodes = + ESP_ENCRYPT_N_NEXT,.next_nodes = + { +#define _(s,n) [ESP_ENCRYPT_NEXT_##s] = n, + foreach_esp_encrypt_next +#undef _ + } +}; + +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_esp_encrypt_post_node, + dpdk_esp_encrypt_post_node_fn) +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/ipsec.c b/src/vnet/devices/dpdk/ipsec/ipsec.c new file mode 100644 index 00000000000..de253f02636 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/ipsec.c @@ -0,0 +1,313 @@ +/* + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include +#include +#include +#include +#include +#include +#include + +#define DPDK_CRYPTO_NB_OBJS 2048 +#define DPDK_CRYPTO_CACHE_SIZE 512 +#define DPDK_CRYPTO_PRIV_SIZE 128 +#define DPDK_CRYPTO_N_QUEUE_DESC 512 +#define DPDK_CRYPTO_NB_COPS (1024 * 4) + +/* + * return: + * -1: update failed + * 0: already exist + * 1: mapped + */ +static int +update_qp_data (crypto_worker_main_t * cwm, + u8 cdev_id, u16 qp_id, u8 is_outbound, u16 * idx) +{ + crypto_qp_data_t *qpd; + + /* *INDENT-OFF* */ + vec_foreach_index (*idx, cwm->qp_data) + { + qpd = vec_elt_at_index(cwm->qp_data, *idx); + + if (qpd->dev_id == cdev_id && qpd->qp_id == qp_id && + qpd->is_outbound == is_outbound) + return 0; + } + /* *INDENT-ON* */ + + vec_add2 (cwm->qp_data, qpd, 1); + + qpd->dev_id = cdev_id; + qpd->qp_id = qp_id; + qpd->is_outbound = is_outbound; + + return 1; +} + +/* + * return: + * -1: error + * 0: already exist + * 1: mapped + */ +static int +add_mapping (crypto_worker_main_t * cwm, + u8 cdev_id, u16 qp, u8 is_outbound, + const struct rte_cryptodev_capabilities *cipher_cap, + const struct rte_cryptodev_capabilities *auth_cap) +{ + int mapped; + u16 qp_index; + uword key = 0, data, *ret; + crypto_worker_qp_key_t *p_key = (crypto_worker_qp_key_t *) & key; + + p_key->cipher_algo = (u8) cipher_cap->sym.cipher.algo; + p_key->auth_algo = (u8) auth_cap->sym.auth.algo; + p_key->is_outbound = is_outbound; + + ret = hash_get (cwm->algo_qp_map, key); + if (ret) + return 0; + + mapped = update_qp_data (cwm, cdev_id, qp, is_outbound, &qp_index); + if (mapped < 0) + return -1; + + data = (uword) qp_index; + + ret = hash_set (cwm->algo_qp_map, key, data); + if (!ret) + rte_panic ("Failed to insert hash table\n"); + + return mapped; +} + +/* + * return: + * 0: already exist + * 1: mapped + */ +static int +add_cdev_mapping (crypto_worker_main_t * cwm, + struct rte_cryptodev_info *dev_info, u8 cdev_id, + u16 qp, u8 is_outbound) +{ + const struct rte_cryptodev_capabilities *i, *j; + u32 mapped = 0; + + for (i = dev_info->capabilities; i->op != RTE_CRYPTO_OP_TYPE_UNDEFINED; i++) + { + if (i->sym.xform_type != RTE_CRYPTO_SYM_XFORM_CIPHER) + continue; + + if (check_algo_is_supported (i, NULL) != 0) + continue; + + for (j = dev_info->capabilities; j->op != RTE_CRYPTO_OP_TYPE_UNDEFINED; + j++) + { + int status = 0; + + if (j->sym.xform_type != RTE_CRYPTO_SYM_XFORM_AUTH) + continue; + + if (check_algo_is_supported (j, NULL) != 0) + continue; + + status = add_mapping (cwm, cdev_id, qp, is_outbound, i, j); + if (status == 1) + mapped += 1; + if (status < 0) + return status; + } + } + + return mapped; +} + +static int +check_cryptodev_queues () +{ + u32 n_qs = 0; + u8 cdev_id; + u32 n_req_qs = 2; + + if (vlib_num_workers () > 0) + n_req_qs = vlib_num_workers () * 2; + + for (cdev_id = 0; cdev_id < rte_cryptodev_count (); cdev_id++) + { + struct rte_cryptodev_info cdev_info; + + rte_cryptodev_info_get (cdev_id, &cdev_info); + + if (! + (cdev_info.feature_flags & RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING)) + continue; + + n_qs += cdev_info.max_nb_queue_pairs; + } + + if (n_qs >= n_req_qs) + return 0; + else + return -1; +} + +static clib_error_t * +dpdk_ipsec_init (vlib_main_t * vm) +{ + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); + struct rte_cryptodev_config dev_conf; + struct rte_cryptodev_qp_conf qp_conf; + struct rte_cryptodev_info cdev_info; + struct rte_mempool *rmp; + i32 dev_id, ret; + u32 i, skip_master; + + if (check_cryptodev_queues () < 0) + return clib_error_return (0, "not enough cryptodevs for ipsec"); + + vec_alloc (dcm->workers_main, tm->n_vlib_mains); + _vec_len (dcm->workers_main) = tm->n_vlib_mains; + + fprintf (stdout, "DPDK Cryptodevs info:\n"); + fprintf (stdout, "dev_id\tn_qp\tnb_obj\tcache_size\n"); + /* HW cryptodevs have higher dev_id, use HW first */ + for (dev_id = rte_cryptodev_count () - 1; dev_id >= 0; dev_id--) + { + u16 max_nb_qp, qp = 0; + skip_master = vlib_num_workers () > 0; + + rte_cryptodev_info_get (dev_id, &cdev_info); + + if (! + (cdev_info.feature_flags & RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING)) + continue; + + max_nb_qp = cdev_info.max_nb_queue_pairs; + + for (i = 0; i < tm->n_vlib_mains; i++) + { + u8 is_outbound; + crypto_worker_main_t *cwm; + uword *map; + + if (skip_master) + { + skip_master = 0; + continue; + } + + cwm = vec_elt_at_index (dcm->workers_main, i); + map = cwm->algo_qp_map; + + if (!map) + { + map = hash_create (0, sizeof (crypto_worker_qp_key_t)); + if (!map) + return clib_error_return (0, "unable to create hash table " + "for worker %u", + vlib_mains[i]->cpu_index); + cwm->algo_qp_map = map; + } + + for (is_outbound = 0; is_outbound < 2 && qp < max_nb_qp; + is_outbound++) + { + int mapped = add_cdev_mapping (cwm, &cdev_info, + dev_id, qp, is_outbound); + if (mapped > 0) + qp++; + + if (mapped < 0) + return clib_error_return (0, + "too many queues for one worker"); + } + } + + if (qp == 0) + continue; + + dev_conf.socket_id = rte_cryptodev_socket_id (dev_id); + dev_conf.nb_queue_pairs = cdev_info.max_nb_queue_pairs; + dev_conf.session_mp.nb_objs = DPDK_CRYPTO_NB_OBJS; + dev_conf.session_mp.cache_size = DPDK_CRYPTO_CACHE_SIZE; + + ret = rte_cryptodev_configure (dev_id, &dev_conf); + if (ret < 0) + return clib_error_return (0, "cryptodev %u config error", dev_id); + + qp_conf.nb_descriptors = DPDK_CRYPTO_N_QUEUE_DESC; + for (qp = 0; qp < dev_conf.nb_queue_pairs; qp++) + { + ret = rte_cryptodev_queue_pair_setup (dev_id, qp, &qp_conf, + dev_conf.socket_id); + if (ret < 0) + return clib_error_return (0, "cryptodev %u qp %u setup error", + dev_id, qp); + } + fprintf (stdout, "%u\t%u\t%u\t%u\n", dev_id, dev_conf.nb_queue_pairs, + DPDK_CRYPTO_NB_OBJS, DPDK_CRYPTO_CACHE_SIZE); + } + + u32 socket_id = rte_socket_id (); + + vec_validate_aligned (dcm->cop_pools, socket_id, CLIB_CACHE_LINE_BYTES); + + /* pool already exists, nothing to do */ + if (dcm->cop_pools[socket_id]) + return 0; + + u8 *pool_name = format (0, "crypto_op_pool_socket%u%c", socket_id, 0); + + rmp = rte_crypto_op_pool_create ((char *) pool_name, + RTE_CRYPTO_OP_TYPE_SYMMETRIC, + DPDK_CRYPTO_NB_COPS * + (1 + vlib_num_workers ()), + DPDK_CRYPTO_CACHE_SIZE, + DPDK_CRYPTO_PRIV_SIZE, socket_id); + vec_free (pool_name); + + if (!rmp) + return clib_error_return (0, "failed to allocate mempool on socket %u", + socket_id); + dcm->cop_pools[socket_id] = rmp; + + dpdk_esp_init (); + + if (vec_len (vlib_mains) == 0) + vlib_node_set_state (&vlib_global_main, dpdk_crypto_input_node.index, + VLIB_NODE_STATE_POLLING); + else + for (i = 1; i < tm->n_vlib_mains; i++) + vlib_node_set_state (vlib_mains[i], dpdk_crypto_input_node.index, + VLIB_NODE_STATE_POLLING); + + return 0; +} + +VLIB_MAIN_LOOP_ENTER_FUNCTION (dpdk_ipsec_init); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/ipsec.h b/src/vnet/devices/dpdk/ipsec/ipsec.h new file mode 100644 index 00000000000..e6c7498c0d3 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/ipsec.h @@ -0,0 +1,227 @@ +/* + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#ifndef __DPDK_IPSEC_H__ +#define __DPDK_IPSEC_H__ + +#include + +#undef always_inline +#include +#include + +#if CLIB_DEBUG > 0 +#define always_inline static inline +#else +#define always_inline static inline __attribute__ ((__always_inline__)) +#endif + + +#define MAX_QP_PER_LCORE 16 + +typedef struct +{ + u32 salt; + u32 iv[2]; + u32 cnt; +} dpdk_gcm_cnt_blk; + +typedef struct +{ + dpdk_gcm_cnt_blk cb; + union + { + u8 aad[12]; + u8 icv[64]; + }; +} dpdk_cop_priv_t; + +typedef struct +{ + u8 cipher_algo; + u8 auth_algo; + u8 is_outbound; +} crypto_worker_qp_key_t; + +typedef struct +{ + u16 dev_id; + u16 qp_id; + u16 is_outbound; + i16 inflights; + u32 bi[VLIB_FRAME_SIZE]; + struct rte_crypto_op *cops[VLIB_FRAME_SIZE]; + struct rte_crypto_op **free_cops; +} crypto_qp_data_t; + +typedef struct +{ + u8 qp_index; + void *sess; +} crypto_sa_session_t; + +typedef struct +{ + crypto_sa_session_t *sa_sess_d[2]; + crypto_qp_data_t *qp_data; + uword *algo_qp_map; +} crypto_worker_main_t; + +typedef struct +{ + struct rte_mempool **cop_pools; + crypto_worker_main_t *workers_main; +} dpdk_crypto_main_t; + +dpdk_crypto_main_t dpdk_crypto_main; + +extern vlib_node_registration_t dpdk_crypto_input_node; + +#define CRYPTO_N_FREE_COPS (VLIB_FRAME_SIZE * 3) + +static_always_inline void +crypto_alloc_cops () +{ + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + u32 cpu_index = os_get_cpu_number (); + crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index]; + unsigned socket_id = rte_socket_id (); + crypto_qp_data_t *qpd; + + /* *INDENT-OFF* */ + vec_foreach (qpd, cwm->qp_data) + { + u32 l = vec_len (qpd->free_cops); + + if (PREDICT_FALSE (l < VLIB_FRAME_SIZE)) + { + u32 n_alloc; + + if (PREDICT_FALSE (!qpd->free_cops)) + vec_alloc (qpd->free_cops, CRYPTO_N_FREE_COPS); + + n_alloc = rte_crypto_op_bulk_alloc (dcm->cop_pools[socket_id], + RTE_CRYPTO_OP_TYPE_SYMMETRIC, + &qpd->free_cops[l], + CRYPTO_N_FREE_COPS - l - 1); + + _vec_len (qpd->free_cops) = l + n_alloc; + } + } + /* *INDENT-ON* */ +} + +static_always_inline void +crypto_free_cop (crypto_qp_data_t * qpd, struct rte_crypto_op **cops, u32 n) +{ + u32 l = vec_len (qpd->free_cops); + + if (l + n >= CRYPTO_N_FREE_COPS) + { + l -= VLIB_FRAME_SIZE; + rte_mempool_put_bulk (cops[0]->mempool, + (void **) &qpd->free_cops[l], VLIB_FRAME_SIZE); + } + clib_memcpy (&qpd->free_cops[l], cops, sizeof (*cops) * n); + + _vec_len (qpd->free_cops) = l + n; +} + +static_always_inline int +check_algo_is_supported (const struct rte_cryptodev_capabilities *cap, + char *name) +{ + struct + { + uint8_t cipher_algo; + enum rte_crypto_sym_xform_type type; + union + { + enum rte_crypto_auth_algorithm auth; + enum rte_crypto_cipher_algorithm cipher; + }; + char *name; + } supported_algo[] = + { + { + .type = RTE_CRYPTO_SYM_XFORM_CIPHER,.cipher = + RTE_CRYPTO_CIPHER_NULL,.name = "NULL"}, + { + .type = RTE_CRYPTO_SYM_XFORM_CIPHER,.cipher = + RTE_CRYPTO_CIPHER_AES_CBC,.name = "AES_CBC"}, + { + .type = RTE_CRYPTO_SYM_XFORM_CIPHER,.cipher = + RTE_CRYPTO_CIPHER_AES_CTR,.name = "AES_CTR"}, + { + .type = RTE_CRYPTO_SYM_XFORM_CIPHER,.cipher = + RTE_CRYPTO_CIPHER_3DES_CBC,.name = "3DES-CBC"}, + { + .type = RTE_CRYPTO_SYM_XFORM_CIPHER,.auth = + RTE_CRYPTO_CIPHER_AES_GCM,.name = "AES-GCM"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_SHA1_HMAC,.name = "HMAC-SHA1"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_SHA256_HMAC,.name = "HMAC-SHA256"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_SHA384_HMAC,.name = "HMAC-SHA384"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_SHA512_HMAC,.name = "HMAC-SHA512"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_AES_XCBC_MAC,.name = "AES-XCBC-MAC"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_AES_GCM,.name = "AES-GCM"}, + { + /* tail */ + .type = RTE_CRYPTO_SYM_XFORM_NOT_SPECIFIED},}; + uint32_t i = 0; + + if (cap->op != RTE_CRYPTO_OP_TYPE_SYMMETRIC) + return -1; + + while (supported_algo[i].type != RTE_CRYPTO_SYM_XFORM_NOT_SPECIFIED) + { + if (cap->sym.xform_type == supported_algo[i].type) + { + if ((cap->sym.xform_type == RTE_CRYPTO_SYM_XFORM_CIPHER && + cap->sym.cipher.algo == supported_algo[i].cipher) || + (cap->sym.xform_type == RTE_CRYPTO_SYM_XFORM_AUTH && + cap->sym.auth.algo == supported_algo[i].auth)) + { + if (name) + strcpy (name, supported_algo[i].name); + return 0; + } + } + + i++; + } + + return -1; +} + +#endif /* __DPDK_IPSEC_H__ */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/node.c b/src/vnet/devices/dpdk/node.c new file mode 100644 index 00000000000..e541cdbcbd2 --- /dev/null +++ b/src/vnet/devices/dpdk/node.c @@ -0,0 +1,687 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include + +#include "dpdk_priv.h" + +static char *dpdk_error_strings[] = { +#define _(n,s) s, + foreach_dpdk_error +#undef _ +}; + +always_inline int +vlib_buffer_is_ip4 (vlib_buffer_t * b) +{ + ethernet_header_t *h = (ethernet_header_t *) b->data; + return (h->type == clib_host_to_net_u16 (ETHERNET_TYPE_IP4)); +} + +always_inline int +vlib_buffer_is_ip6 (vlib_buffer_t * b) +{ + ethernet_header_t *h = (ethernet_header_t *) b->data; + return (h->type == clib_host_to_net_u16 (ETHERNET_TYPE_IP6)); +} + +always_inline int +vlib_buffer_is_mpls (vlib_buffer_t * b) +{ + ethernet_header_t *h = (ethernet_header_t *) b->data; + return (h->type == clib_host_to_net_u16 (ETHERNET_TYPE_MPLS_UNICAST)); +} + +#if RTE_VERSION < RTE_VERSION_NUM(16, 11, 0, 0) +/* New ol_flags bits added in DPDK-16.11 */ +#define PKT_RX_IP_CKSUM_GOOD (1ULL << 7) +#endif + +always_inline u32 +dpdk_rx_next_from_etype (struct rte_mbuf * mb, vlib_buffer_t * b0) +{ + if (PREDICT_TRUE (vlib_buffer_is_ip4 (b0))) + if (PREDICT_TRUE ((mb->ol_flags & PKT_RX_IP_CKSUM_GOOD) != 0)) + return VNET_DEVICE_INPUT_NEXT_IP4_NCS_INPUT; + else + return VNET_DEVICE_INPUT_NEXT_IP4_INPUT; + else if (PREDICT_TRUE (vlib_buffer_is_ip6 (b0))) + return VNET_DEVICE_INPUT_NEXT_IP6_INPUT; + else if (PREDICT_TRUE (vlib_buffer_is_mpls (b0))) + return VNET_DEVICE_INPUT_NEXT_MPLS_INPUT; + else + return VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; +} + +always_inline int +dpdk_mbuf_is_vlan (struct rte_mbuf *mb) +{ +#if RTE_VERSION >= RTE_VERSION_NUM(16, 11, 0, 0) + return (mb->packet_type & RTE_PTYPE_L2_ETHER_VLAN) == + RTE_PTYPE_L2_ETHER_VLAN; +#else + return + (mb->ol_flags & + (PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED | PKT_RX_QINQ_STRIPPED)) == + PKT_RX_VLAN_PKT; +#endif +} + +always_inline int +dpdk_mbuf_is_ip4 (struct rte_mbuf *mb) +{ + return RTE_ETH_IS_IPV4_HDR (mb->packet_type) != 0; +} + +always_inline int +dpdk_mbuf_is_ip6 (struct rte_mbuf *mb) +{ + return RTE_ETH_IS_IPV6_HDR (mb->packet_type) != 0; +} + +always_inline u32 +dpdk_rx_next_from_mb (struct rte_mbuf * mb, vlib_buffer_t * b0) +{ + if (PREDICT_FALSE (dpdk_mbuf_is_vlan (mb))) + return VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; + else if (PREDICT_TRUE (dpdk_mbuf_is_ip4 (mb))) + return VNET_DEVICE_INPUT_NEXT_IP4_NCS_INPUT; + else if (PREDICT_TRUE (dpdk_mbuf_is_ip6 (mb))) + return VNET_DEVICE_INPUT_NEXT_IP6_INPUT; + else if (PREDICT_TRUE (vlib_buffer_is_mpls (b0))) + return VNET_DEVICE_INPUT_NEXT_MPLS_INPUT; + else + return dpdk_rx_next_from_etype (mb, b0); +} + +always_inline void +dpdk_rx_error_from_mb (struct rte_mbuf *mb, u32 * next, u8 * error) +{ + if (mb->ol_flags & PKT_RX_IP_CKSUM_BAD) + { + *error = DPDK_ERROR_IP_CHECKSUM_ERROR; + *next = VNET_DEVICE_INPUT_NEXT_DROP; + } + else + *error = DPDK_ERROR_NONE; +} + +void +dpdk_rx_trace (dpdk_main_t * dm, + vlib_node_runtime_t * node, + dpdk_device_t * xd, + u16 queue_id, u32 * buffers, uword n_buffers) +{ + vlib_main_t *vm = vlib_get_main (); + u32 *b, n_left; + u32 next0; + + n_left = n_buffers; + b = buffers; + + while (n_left >= 1) + { + u32 bi0; + vlib_buffer_t *b0; + dpdk_rx_dma_trace_t *t0; + struct rte_mbuf *mb; + u8 error0; + + bi0 = b[0]; + n_left -= 1; + + b0 = vlib_get_buffer (vm, bi0); + mb = rte_mbuf_from_vlib_buffer (b0); + + if (PREDICT_FALSE (xd->per_interface_next_index != ~0)) + next0 = xd->per_interface_next_index; + else if (PREDICT_TRUE + ((xd->flags & DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE) != 0)) + next0 = dpdk_rx_next_from_mb (mb, b0); + else + next0 = dpdk_rx_next_from_etype (mb, b0); + + dpdk_rx_error_from_mb (mb, &next0, &error0); + + vlib_trace_buffer (vm, node, next0, b0, /* follow_chain */ 0); + t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0])); + t0->queue_index = queue_id; + t0->device_index = xd->device_index; + t0->buffer_index = bi0; + + clib_memcpy (&t0->mb, mb, sizeof (t0->mb)); + clib_memcpy (&t0->buffer, b0, sizeof (b0[0]) - sizeof (b0->pre_data)); + clib_memcpy (t0->buffer.pre_data, b0->data, + sizeof (t0->buffer.pre_data)); + clib_memcpy (&t0->data, mb->buf_addr + mb->data_off, sizeof (t0->data)); + + b += 1; + } +} + +static inline u32 +dpdk_rx_burst (dpdk_main_t * dm, dpdk_device_t * xd, u16 queue_id) +{ + u32 n_buffers; + u32 n_left; + u32 n_this_chunk; + + n_left = VLIB_FRAME_SIZE; + n_buffers = 0; + + if (PREDICT_TRUE (xd->flags & DPDK_DEVICE_FLAG_PMD)) + { + while (n_left) + { + n_this_chunk = rte_eth_rx_burst (xd->device_index, queue_id, + xd->rx_vectors[queue_id] + + n_buffers, n_left); + n_buffers += n_this_chunk; + n_left -= n_this_chunk; + + /* Empirically, DPDK r1.8 produces vectors w/ 32 or fewer elts */ + if (n_this_chunk < 32) + break; + } + } + else + { + ASSERT (0); + } + + return n_buffers; +} + + +static_always_inline void +dpdk_process_subseq_segs (vlib_main_t * vm, vlib_buffer_t * b, + struct rte_mbuf *mb, vlib_buffer_free_list_t * fl) +{ + u8 nb_seg = 1; + struct rte_mbuf *mb_seg = 0; + vlib_buffer_t *b_seg, *b_chain = 0; + mb_seg = mb->next; + b_chain = b; + + while ((mb->nb_segs > 1) && (nb_seg < mb->nb_segs)) + { + ASSERT (mb_seg != 0); + + b_seg = vlib_buffer_from_rte_mbuf (mb_seg); + vlib_buffer_init_for_free_list (b_seg, fl); + + ASSERT ((b_seg->flags & VLIB_BUFFER_NEXT_PRESENT) == 0); + ASSERT (b_seg->current_data == 0); + + /* + * The driver (e.g. virtio) may not put the packet data at the start + * of the segment, so don't assume b_seg->current_data == 0 is correct. + */ + b_seg->current_data = + (mb_seg->buf_addr + mb_seg->data_off) - (void *) b_seg->data; + + b_seg->current_length = mb_seg->data_len; + b->total_length_not_including_first_buffer += mb_seg->data_len; + + b_chain->flags |= VLIB_BUFFER_NEXT_PRESENT; + b_chain->next_buffer = vlib_get_buffer_index (vm, b_seg); + + b_chain = b_seg; + mb_seg = mb_seg->next; + nb_seg++; + } +} + +static_always_inline void +dpdk_prefetch_buffer (struct rte_mbuf *mb) +{ + vlib_buffer_t *b = vlib_buffer_from_rte_mbuf (mb); + CLIB_PREFETCH (mb, CLIB_CACHE_LINE_BYTES, LOAD); + CLIB_PREFETCH (b, CLIB_CACHE_LINE_BYTES, STORE); +} + +/* + * This function is used when there are no worker threads. + * The main thread performs IO and forwards the packets. + */ +static_always_inline u32 +dpdk_device_input (dpdk_main_t * dm, dpdk_device_t * xd, + vlib_node_runtime_t * node, u32 cpu_index, u16 queue_id) +{ + u32 n_buffers; + u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; + u32 n_left_to_next, *to_next; + u32 mb_index; + vlib_main_t *vm = vlib_get_main (); + uword n_rx_bytes = 0; + u32 n_trace, trace_cnt __attribute__ ((unused)); + vlib_buffer_free_list_t *fl; + u32 buffer_flags_template; + + if ((xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) == 0) + return 0; + + n_buffers = dpdk_rx_burst (dm, xd, queue_id); + + if (n_buffers == 0) + { + return 0; + } + + buffer_flags_template = dm->buffer_flags_template; + + vec_reset_length (xd->d_trace_buffers[cpu_index]); + trace_cnt = n_trace = vlib_get_trace_count (vm, node); + + if (n_trace > 0) + { + u32 n = clib_min (n_trace, n_buffers); + mb_index = 0; + + while (n--) + { + struct rte_mbuf *mb = xd->rx_vectors[queue_id][mb_index++]; + vlib_buffer_t *b = vlib_buffer_from_rte_mbuf (mb); + vec_add1 (xd->d_trace_buffers[cpu_index], + vlib_get_buffer_index (vm, b)); + } + } + + fl = vlib_buffer_get_free_list (vm, VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); + + mb_index = 0; + + while (n_buffers > 0) + { + vlib_buffer_t *b0, *b1, *b2, *b3; + u32 bi0, next0, l3_offset0; + u32 bi1, next1, l3_offset1; + u32 bi2, next2, l3_offset2; + u32 bi3, next3, l3_offset3; + u8 error0, error1, error2, error3; + u64 or_ol_flags; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_buffers > 8 && n_left_to_next > 4) + { + struct rte_mbuf *mb0 = xd->rx_vectors[queue_id][mb_index]; + struct rte_mbuf *mb1 = xd->rx_vectors[queue_id][mb_index + 1]; + struct rte_mbuf *mb2 = xd->rx_vectors[queue_id][mb_index + 2]; + struct rte_mbuf *mb3 = xd->rx_vectors[queue_id][mb_index + 3]; + + dpdk_prefetch_buffer (xd->rx_vectors[queue_id][mb_index + 4]); + dpdk_prefetch_buffer (xd->rx_vectors[queue_id][mb_index + 5]); + dpdk_prefetch_buffer (xd->rx_vectors[queue_id][mb_index + 6]); + dpdk_prefetch_buffer (xd->rx_vectors[queue_id][mb_index + 7]); + + if (xd->flags & DPDK_DEVICE_FLAG_MAYBE_MULTISEG) + { + if (PREDICT_FALSE (mb0->nb_segs > 1)) + dpdk_prefetch_buffer (mb0->next); + if (PREDICT_FALSE (mb1->nb_segs > 1)) + dpdk_prefetch_buffer (mb1->next); + if (PREDICT_FALSE (mb2->nb_segs > 1)) + dpdk_prefetch_buffer (mb2->next); + if (PREDICT_FALSE (mb3->nb_segs > 1)) + dpdk_prefetch_buffer (mb3->next); + } + + ASSERT (mb0); + ASSERT (mb1); + ASSERT (mb2); + ASSERT (mb3); + + or_ol_flags = (mb0->ol_flags | mb1->ol_flags | + mb2->ol_flags | mb3->ol_flags); + b0 = vlib_buffer_from_rte_mbuf (mb0); + b1 = vlib_buffer_from_rte_mbuf (mb1); + b2 = vlib_buffer_from_rte_mbuf (mb2); + b3 = vlib_buffer_from_rte_mbuf (mb3); + + vlib_buffer_init_for_free_list (b0, fl); + vlib_buffer_init_for_free_list (b1, fl); + vlib_buffer_init_for_free_list (b2, fl); + vlib_buffer_init_for_free_list (b3, fl); + + bi0 = vlib_get_buffer_index (vm, b0); + bi1 = vlib_get_buffer_index (vm, b1); + bi2 = vlib_get_buffer_index (vm, b2); + bi3 = vlib_get_buffer_index (vm, b3); + + to_next[0] = bi0; + to_next[1] = bi1; + to_next[2] = bi2; + to_next[3] = bi3; + to_next += 4; + n_left_to_next -= 4; + + if (PREDICT_FALSE (xd->per_interface_next_index != ~0)) + { + next0 = next1 = next2 = next3 = xd->per_interface_next_index; + } + else if (PREDICT_TRUE + ((xd->flags & DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE) != 0)) + { + next0 = dpdk_rx_next_from_mb (mb0, b0); + next1 = dpdk_rx_next_from_mb (mb1, b1); + next2 = dpdk_rx_next_from_mb (mb2, b2); + next3 = dpdk_rx_next_from_mb (mb3, b3); + } + else + { + next0 = dpdk_rx_next_from_etype (mb0, b0); + next1 = dpdk_rx_next_from_etype (mb1, b1); + next2 = dpdk_rx_next_from_etype (mb2, b2); + next3 = dpdk_rx_next_from_etype (mb3, b3); + } + + if (PREDICT_FALSE (or_ol_flags & PKT_RX_IP_CKSUM_BAD)) + { + dpdk_rx_error_from_mb (mb0, &next0, &error0); + dpdk_rx_error_from_mb (mb1, &next1, &error1); + dpdk_rx_error_from_mb (mb2, &next2, &error2); + dpdk_rx_error_from_mb (mb3, &next3, &error3); + b0->error = node->errors[error0]; + b1->error = node->errors[error1]; + b2->error = node->errors[error2]; + b3->error = node->errors[error3]; + } + else + { + b0->error = b1->error = node->errors[DPDK_ERROR_NONE]; + b2->error = b3->error = node->errors[DPDK_ERROR_NONE]; + } + + l3_offset0 = device_input_next_node_advance[next0]; + l3_offset1 = device_input_next_node_advance[next1]; + l3_offset2 = device_input_next_node_advance[next2]; + l3_offset3 = device_input_next_node_advance[next3]; + + b0->current_data = l3_offset0 + mb0->data_off; + b1->current_data = l3_offset1 + mb1->data_off; + b2->current_data = l3_offset2 + mb2->data_off; + b3->current_data = l3_offset3 + mb3->data_off; + + b0->current_data -= RTE_PKTMBUF_HEADROOM; + b1->current_data -= RTE_PKTMBUF_HEADROOM; + b2->current_data -= RTE_PKTMBUF_HEADROOM; + b3->current_data -= RTE_PKTMBUF_HEADROOM; + + b0->current_length = mb0->data_len - l3_offset0; + b1->current_length = mb1->data_len - l3_offset1; + b2->current_length = mb2->data_len - l3_offset2; + b3->current_length = mb3->data_len - l3_offset3; + + b0->flags = buffer_flags_template; + b1->flags = buffer_flags_template; + b2->flags = buffer_flags_template; + b3->flags = buffer_flags_template; + + vnet_buffer (b0)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b1)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b2)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b3)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + vnet_buffer (b1)->sw_if_index[VLIB_TX] = (u32) ~ 0; + vnet_buffer (b2)->sw_if_index[VLIB_TX] = (u32) ~ 0; + vnet_buffer (b3)->sw_if_index[VLIB_TX] = (u32) ~ 0; + + n_rx_bytes += mb0->pkt_len; + n_rx_bytes += mb1->pkt_len; + n_rx_bytes += mb2->pkt_len; + n_rx_bytes += mb3->pkt_len; + + /* Process subsequent segments of multi-segment packets */ + if (xd->flags & DPDK_DEVICE_FLAG_MAYBE_MULTISEG) + { + dpdk_process_subseq_segs (vm, b0, mb0, fl); + dpdk_process_subseq_segs (vm, b1, mb1, fl); + dpdk_process_subseq_segs (vm, b2, mb2, fl); + dpdk_process_subseq_segs (vm, b3, mb3, fl); + } + + /* + * Turn this on if you run into + * "bad monkey" contexts, and you want to know exactly + * which nodes they've visited... See main.c... + */ + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b1); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b2); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b3); + + /* Do we have any driver RX features configured on the interface? */ + vnet_feature_start_device_input_x4 (xd->vlib_sw_if_index, + &next0, &next1, &next2, &next3, + b0, b1, b2, b3, + l3_offset0, l3_offset1, + l3_offset2, l3_offset3); + + vlib_validate_buffer_enqueue_x4 (vm, node, next_index, + to_next, n_left_to_next, + bi0, bi1, bi2, bi3, + next0, next1, next2, next3); + n_buffers -= 4; + mb_index += 4; + } + while (n_buffers > 0 && n_left_to_next > 0) + { + struct rte_mbuf *mb0 = xd->rx_vectors[queue_id][mb_index]; + + ASSERT (mb0); + + b0 = vlib_buffer_from_rte_mbuf (mb0); + + /* Prefetch one next segment if it exists. */ + if (PREDICT_FALSE (mb0->nb_segs > 1)) + dpdk_prefetch_buffer (mb0->next); + + vlib_buffer_init_for_free_list (b0, fl); + + bi0 = vlib_get_buffer_index (vm, b0); + + to_next[0] = bi0; + to_next++; + n_left_to_next--; + + if (PREDICT_FALSE (xd->per_interface_next_index != ~0)) + next0 = xd->per_interface_next_index; + else if (PREDICT_TRUE + ((xd->flags & DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE) != 0)) + next0 = dpdk_rx_next_from_mb (mb0, b0); + else + next0 = dpdk_rx_next_from_etype (mb0, b0); + + dpdk_rx_error_from_mb (mb0, &next0, &error0); + b0->error = node->errors[error0]; + + l3_offset0 = device_input_next_node_advance[next0]; + + b0->current_data = l3_offset0; + b0->current_data += mb0->data_off - RTE_PKTMBUF_HEADROOM; + b0->current_length = mb0->data_len - l3_offset0; + + b0->flags = buffer_flags_template; + + vnet_buffer (b0)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + n_rx_bytes += mb0->pkt_len; + + /* Process subsequent segments of multi-segment packets */ + dpdk_process_subseq_segs (vm, b0, mb0, fl); + + /* + * Turn this on if you run into + * "bad monkey" contexts, and you want to know exactly + * which nodes they've visited... See main.c... + */ + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); + + /* Do we have any driver RX features configured on the interface? */ + vnet_feature_start_device_input_x1 (xd->vlib_sw_if_index, &next0, + b0, l3_offset0); + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, + bi0, next0); + n_buffers--; + mb_index++; + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + + if (PREDICT_FALSE (vec_len (xd->d_trace_buffers[cpu_index]) > 0)) + { + dpdk_rx_trace (dm, node, xd, queue_id, xd->d_trace_buffers[cpu_index], + vec_len (xd->d_trace_buffers[cpu_index])); + vlib_set_trace_count (vm, node, n_trace - + vec_len (xd->d_trace_buffers[cpu_index])); + } + + vlib_increment_combined_counter + (vnet_get_main ()->interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, + cpu_index, xd->vlib_sw_if_index, mb_index, n_rx_bytes); + + dpdk_worker_t *dw = vec_elt_at_index (dm->workers, cpu_index); + dw->aggregate_rx_packets += mb_index; + + return mb_index; +} + +static inline void +poll_rate_limit (dpdk_main_t * dm) +{ + /* Limit the poll rate by sleeping for N msec between polls */ + if (PREDICT_FALSE (dm->poll_sleep != 0)) + { + struct timespec ts, tsrem; + + ts.tv_sec = 0; + ts.tv_nsec = 1000 * 1000 * dm->poll_sleep; /* 1ms */ + + while (nanosleep (&ts, &tsrem) < 0) + { + ts = tsrem; + } + } +} + +/** \brief Main DPDK input node + @node dpdk-input + + This is the main DPDK input node: across each assigned interface, + call rte_eth_rx_burst(...) or similar to obtain a vector of + packets to process. Handle early packet discard. Derive @c + vlib_buffer_t metadata from struct rte_mbuf metadata, + Depending on the resulting metadata: adjust b->current_data, + b->current_length and dispatch directly to + ip4-input-no-checksum, or ip6-input. Trace the packet if required. + + @param vm vlib_main_t corresponding to the current thread + @param node vlib_node_runtime_t + @param f vlib_frame_t input-node, not used. + + @par Graph mechanics: buffer metadata, next index usage + + @em Uses: + - struct rte_mbuf mb->ol_flags + - PKT_RX_IP_CKSUM_BAD + - RTE_ETH_IS_xxx_HDR(mb->packet_type) + - packet classification result + + @em Sets: + - b->error if the packet is to be dropped immediately + - b->current_data, b->current_length + - adjusted as needed to skip the L2 header in direct-dispatch cases + - vnet_buffer(b)->sw_if_index[VLIB_RX] + - rx interface sw_if_index + - vnet_buffer(b)->sw_if_index[VLIB_TX] = ~0 + - required by ipX-lookup + - b->flags + - to indicate multi-segment pkts (VLIB_BUFFER_NEXT_PRESENT), etc. + + Next Nodes: + - Static arcs to: error-drop, ethernet-input, + ip4-input-no-checksum, ip6-input, mpls-input + - per-interface redirection, controlled by + xd->per_interface_next_index +*/ + +static uword +dpdk_input (vlib_main_t * vm, vlib_node_runtime_t * node, vlib_frame_t * f) +{ + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd; + uword n_rx_packets = 0; + dpdk_device_and_queue_t *dq; + u32 cpu_index = os_get_cpu_number (); + + /* + * Poll all devices on this cpu for input/interrupts. + */ + /* *INDENT-OFF* */ + vec_foreach (dq, dm->devices_by_cpu[cpu_index]) + { + xd = vec_elt_at_index(dm->devices, dq->device); + n_rx_packets += dpdk_device_input (dm, xd, node, cpu_index, dq->queue_id); + } + /* *INDENT-ON* */ + + poll_rate_limit (dm); + + return n_rx_packets; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (dpdk_input_node) = { + .function = dpdk_input, + .type = VLIB_NODE_TYPE_INPUT, + .name = "dpdk-input", + .sibling_of = "device-input", + + /* Will be enabled if/when hardware is detected. */ + .state = VLIB_NODE_STATE_DISABLED, + + .format_buffer = format_ethernet_header_with_length, + .format_trace = format_dpdk_rx_dma_trace, + + .n_errors = DPDK_N_ERROR, + .error_strings = dpdk_error_strings, +}; + +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_input_node, dpdk_input); +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/qos_doc.md b/src/vnet/devices/dpdk/qos_doc.md new file mode 100644 index 00000000000..9bd0659d616 --- /dev/null +++ b/src/vnet/devices/dpdk/qos_doc.md @@ -0,0 +1,404 @@ +# QoS Hierarchical Scheduler {#qos_doc} + +The Quality-of-Service (QoS) scheduler performs egress-traffic management by +prioritizing the transmission of the packets of different type services and +subcribers based on the Service Level Agreements (SLAs). The QoS scheduler can +be enabled on one or more NIC output interfaces depending upon the +requirement. + + +## Overview + +The QoS schdeuler supports a number of scheduling and shaping levels which +construct hierarchical-tree. The first level in the hierarchy is port (i.e. +the physical interface) that constitutes the root node of the tree. The +subsequent level is subport which represents the group of the +users/subscribers. The individual user/subscriber is represented by the pipe +at the next level. Each user can have different traffic type based on the +criteria of specific loss rate, jitter, and latency. These traffic types are +represented at the traffic-class level in the form of different traffic- +classes. The last level contains number of queues which are grouped together +to host the packets of the specific class type traffic. + +The QoS scheduler implementation requires flow classification, enqueue and +dequeue operations. The flow classification is mandatory stage for HQoS where +incoming packets are classified by mapping the packet fields information to +5-tuple (HQoS subport, pipe, traffic class, queue within traffic class, and +color) and storing that information in mbuf sched field. The enqueue operation +uses this information to determine the queue for storing the packet, and at +this stage, if the specific queue is full, QoS drops the packet. The dequeue +operation consists of scheduling the packet based on its length and available +credits, and handing over the scheduled packet to the output interface. + +For more information on QoS Scheduler, please refer DPDK Programmer's Guide- +http://dpdk.org/doc/guides/prog_guide/qos_framework.html + + +### QoS Schdeuler Parameters + +Following illustrates the default HQoS configuration for each 10GbE output +port: + +Single subport (subport 0): + - Subport rate set to 100% of port rate + - Each of the 4 traffic classes has rate set to 100% of port rate + +4K pipes per subport 0 (pipes 0 .. 4095) with identical configuration: + - Pipe rate set to 1/4K of port rate + - Each of the 4 traffic classes has rate set to 100% of pipe rate + - Within each traffic class, the byte-level WRR weights for the 4 queues are set to 1:1:1:1 + + +#### Port configuration + +``` +port { + rate 1250000000 /* Assuming 10GbE port */ + frame_overhead 24 /* Overhead fields per Ethernet frame: + * 7B (Preamble) + + * 1B (Start of Frame Delimiter (SFD)) + + * 4B (Frame Check Sequence (FCS)) + + * 12B (Inter Frame Gap (IFG)) + */ + mtu 1522 /* Assuming Ethernet/IPv4 pkt (FCS not included) */ + n_subports_per_port 1 /* Number of subports per output interface */ + n_pipes_per_subport 4096 /* Number of pipes (users/subscribers) */ + queue_sizes 64 64 64 64 /* Packet queue size for each traffic class. + * All queues within the same pipe traffic class + * have the same size. Queues from different + * pipes serving the same traffic class have + * the same size. */ +} +``` + + +#### Subport configuration + +``` +subport 0 { + tb_rate 1250000000 /* Subport level token bucket rate (bytes per second) */ + tb_size 1000000 /* Subport level token bucket size (bytes) */ + tc0_rate 1250000000 /* Subport level token bucket rate for traffic class 0 (bytes per second) */ + tc1_rate 1250000000 /* Subport level token bucket rate for traffic class 1 (bytes per second) */ + tc2_rate 1250000000 /* Subport level token bucket rate for traffic class 2 (bytes per second) */ + tc3_rate 1250000000 /* Subport level token bucket rate for traffic class 3 (bytes per second) */ + tc_period 10 /* Time interval for refilling the token bucket associated with traffic class (Milliseconds) */ + pipe 0 4095 profile 0 /* pipes (users/subscribers) configured with pipe profile 0 */ +} +``` + + +#### Pipe configuration + +``` +pipe_profile 0 { + tb_rate 305175 /* Pipe level token bucket rate (bytes per second) */ + tb_size 1000000 /* Pipe level token bucket size (bytes) */ + tc0_rate 305175 /* Pipe level token bucket rate for traffic class 0 (bytes per second) */ + tc1_rate 305175 /* Pipe level token bucket rate for traffic class 1 (bytes per second) */ + tc2_rate 305175 /* Pipe level token bucket rate for traffic class 2 (bytes per second) */ + tc3_rate 305175 /* Pipe level token bucket rate for traffic class 3 (bytes per second) */ + tc_period 40 /* Time interval for refilling the token bucket associated with traffic class at pipe level (Milliseconds) */ + tc3_oversubscription_weight 1 /* Weight traffic class 3 oversubscription */ + tc0_wrr_weights 1 1 1 1 /* Pipe queues WRR weights for traffic class 0 */ + tc1_wrr_weights 1 1 1 1 /* Pipe queues WRR weights for traffic class 1 */ + tc2_wrr_weights 1 1 1 1 /* Pipe queues WRR weights for traffic class 2 */ + tc3_wrr_weights 1 1 1 1 /* Pipe queues WRR weights for traffic class 3 */ +} +``` + + +#### Random Early Detection (RED) parameters per traffic class and color (Green / Yellow / Red) + +``` +red { + tc0_wred_min 48 40 32 /* Minimum threshold for traffic class 0 queue (min_th) in number of packets */ + tc0_wred_max 64 64 64 /* Maximum threshold for traffic class 0 queue (max_th) in number of packets */ + tc0_wred_inv_prob 10 10 10 /* Inverse of packet marking probability for traffic class 0 queue (maxp = 1 / maxp_inv) */ + tc0_wred_weight 9 9 9 /* Traffic Class 0 queue weight */ + tc1_wred_min 48 40 32 /* Minimum threshold for traffic class 1 queue (min_th) in number of packets */ + tc1_wred_max 64 64 64 /* Maximum threshold for traffic class 1 queue (max_th) in number of packets */ + tc1_wred_inv_prob 10 10 10 /* Inverse of packet marking probability for traffic class 1 queue (maxp = 1 / maxp_inv) */ + tc1_wred_weight 9 9 9 /* Traffic Class 1 queue weight */ + tc2_wred_min 48 40 32 /* Minimum threshold for traffic class 2 queue (min_th) in number of packets */ + tc2_wred_max 64 64 64 /* Maximum threshold for traffic class 2 queue (max_th) in number of packets */ + tc2_wred_inv_prob 10 10 10 /* Inverse of packet marking probability for traffic class 2 queue (maxp = 1 / maxp_inv) */ + tc2_wred_weight 9 9 9 /* Traffic Class 2 queue weight */ + tc3_wred_min 48 40 32 /* Minimum threshold for traffic class 3 queue (min_th) in number of packets */ + tc3_wred_max 64 64 64 /* Maximum threshold for traffic class 3 queue (max_th) in number of packets */ + tc3_wred_inv_prob 10 10 10 /* Inverse of packet marking probability for traffic class 3 queue (maxp = 1 / maxp_inv) */ + tc3_wred_weight 9 9 9 /* Traffic Class 3 queue weight */ +} +``` + + +### DPDK QoS Scheduler Integration in VPP + +The Hierarchical Quaity-of-Service (HQoS) scheduler object could be seen as +part of the logical NIC output interface. To enable HQoS on specific output +interface, vpp startup.conf file has to be configured accordingly. The output +interface that requires HQoS, should have "hqos" parameter specified in dpdk +section. Another optional parameter "hqos-thread" has been defined which can +be used to associate the output interface with specific hqos thread. In cpu +section of the config file, "corelist-hqos-threads" is introduced to assign +logical cpu cores to run the HQoS threads. A HQoS thread can run multiple HQoS +objects each associated with different output interfaces. All worker threads +instead of writing packets to NIC TX queue directly, write the packets to a +software queues. The hqos_threads read the software queues, and enqueue the +packets to HQoS objects, as well as dequeue packets from HQOS objects and +write them to NIC output interfaces. The worker threads need to be able to +send the packets to any output interface, therefore, each HQoS object +associated with NIC output interface should have software queues equal to +worker threads count. + +Following illustrates the sample startup configuration file with 4x worker +threads feeding 2x hqos threads that handle each QoS scheduler for 1x output +interface. + +``` +dpdk { + socket-mem 16384,16384 + + dev 0000:02:00.0 { + num-rx-queues 2 + hqos + } + dev 0000:06:00.0 { + num-rx-queues 2 + hqos + } + + num-mbufs 1000000 +} + +cpu { + main-core 0 + corelist-workers 1, 2, 3, 4 + corelist-hqos-threads 5, 6 +} +``` + + +### QoS scheduler CLI Commands + +Each QoS scheduler instance is initialised with default parameters required to +configure hqos port, subport, pipe and queues. Some of the parameters can be +re-configured in run-time through CLI commands. + + +#### Configuration + +Following commands can be used to configure QoS scheduler parameters. + +The command below can be used to set the subport level parameters such as +token bucket rate (bytes per seconds), token bucket size (bytes), traffic +class rates (bytes per seconds) and token update period (Milliseconds). + +``` +set dpdk interface hqos subport subport [rate ] + [bktsize ] [tc0 ] [tc1 ] [tc2 ] [tc3 ] [period ] +``` + +For setting the pipe profile, following command can be used. + +``` +set dpdk interface hqos pipe subport pipe profile +``` + +To assign QoS scheduler instance to the specific thread, following command can +be used. + +``` +set dpdk interface hqos placement thread +``` + +The command below is used to set the packet fields required for classifiying +the incoming packet. As a result of classification process, packet field +information will be mapped to 5 tuples (subport, pipe, traffic class, pipe, +color) and stored in packet mbuf. + +``` +set dpdk interface hqos pktfield id offset mask +``` + +The DSCP table entries used for idenfiying the traffic class and queue can be set using the command below; + +``` +set dpdk interface hqos tctbl entry tc queue +``` + + +#### Show Command + +The QoS Scheduler configuration can displayed using the command below. + +``` + vpp# show dpdk interface hqos TenGigabitEthernet2/0/0 + Thread: + Input SWQ size = 4096 packets + Enqueue burst size = 256 packets + Dequeue burst size = 220 packets + Packet field 0: slab position = 0, slab bitmask = 0x0000000000000000 + Packet field 1: slab position = 40, slab bitmask = 0x0000000fff000000 + Packet field 2: slab position = 8, slab bitmask = 0x00000000000000fc + Packet field 2 translation table: + [ 0 .. 15]: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 + [16 .. 31]: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 + [32 .. 47]: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 + [48 .. 63]: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 + Port: + Rate = 1250000000 bytes/second + MTU = 1514 bytes + Frame overhead = 24 bytes + Number of subports = 1 + Number of pipes per subport = 4096 + Packet queue size: TC0 = 64, TC1 = 64, TC2 = 64, TC3 = 64 packets + Number of pipe profiles = 1 + Pipe profile 0: + Rate = 305175 bytes/second + Token bucket size = 1000000 bytes + Traffic class rate: TC0 = 305175, TC1 = 305175, TC2 = 305175, TC3 = 305175 bytes/second + TC period = 40 milliseconds + TC0 WRR weights: Q0 = 1, Q1 = 1, Q2 = 1, Q3 = 1 + TC1 WRR weights: Q0 = 1, Q1 = 1, Q2 = 1, Q3 = 1 + TC2 WRR weights: Q0 = 1, Q1 = 1, Q2 = 1, Q3 = 1 + TC3 WRR weights: Q0 = 1, Q1 = 1, Q2 = 1, Q3 = 1 +``` + +The QoS Scheduler placement over the logical cpu cores can be displayed using +below command. + +``` + vpp# show dpdk interface hqos placement + Thread 5 (vpp_hqos-threads_0 at lcore 5): + TenGigabitEthernet2/0/0 queue 0 + Thread 6 (vpp_hqos-threads_1 at lcore 6): + TenGigabitEthernet4/0/1 queue 0 +``` + + +### QoS Scheduler Binary APIs + +This section explans the available binary APIs for configuring QoS scheduler +parameters in run-time. + +The following API can be used to set the pipe profile of a pipe that belongs +to a given subport: + +``` +sw_interface_set_dpdk_hqos_pipe rx | sw_if_index + subport pipe profile +``` + +The data structures used for set the pipe profile parameter are as follows; + +``` + /** \\brief DPDK interface HQoS pipe profile set request + @param client_index - opaque cookie to identify the sender + @param context - sender context, to match reply w/ request + @param sw_if_index - the interface + @param subport - subport ID + @param pipe - pipe ID within its subport + @param profile - pipe profile ID + */ + define sw_interface_set_dpdk_hqos_pipe { + u32 client_index; + u32 context; + u32 sw_if_index; + u32 subport; + u32 pipe; + u32 profile; + }; + + /** \\brief DPDK interface HQoS pipe profile set reply + @param context - sender context, to match reply w/ request + @param retval - request return code + */ + define sw_interface_set_dpdk_hqos_pipe_reply { + u32 context; + i32 retval; + }; +``` + +The following API can be used to set the subport level parameters, for +example- token bucket rate (bytes per seconds), token bucket size (bytes), +traffic class rate (bytes per seconds) and tokens update period. + +``` +sw_interface_set_dpdk_hqos_subport rx | sw_if_index + subport [rate ] [bktsize ] + [tc0 ] [tc1 ] [tc2 ] [tc3 ] [period ] +``` + +The data structures used for set the subport level parameter are as follows; + +``` + /** \\brief DPDK interface HQoS subport parameters set request + @param client_index - opaque cookie to identify the sender + @param context - sender context, to match reply w/ request + @param sw_if_index - the interface + @param subport - subport ID + @param tb_rate - subport token bucket rate (measured in bytes/second) + @param tb_size - subport token bucket size (measured in credits) + @param tc_rate - subport traffic class 0 .. 3 rates (measured in bytes/second) + @param tc_period - enforcement period for rates (measured in milliseconds) + */ + define sw_interface_set_dpdk_hqos_subport { + u32 client_index; + u32 context; + u32 sw_if_index; + u32 subport; + u32 tb_rate; + u32 tb_size; + u32 tc_rate[4]; + u32 tc_period; + }; + + /** \\brief DPDK interface HQoS subport parameters set reply + @param context - sender context, to match reply w/ request + @param retval - request return code + */ + define sw_interface_set_dpdk_hqos_subport_reply { + u32 context; + i32 retval; + }; +``` + +The following API can be used set the DSCP table entry. The DSCP table have +64 entries to map the packet DSCP field onto traffic class and hqos input +queue. + +``` +sw_interface_set_dpdk_hqos_tctbl rx | sw_if_index + entry tc queue +``` + +The data structures used for setting DSCP table entries are given below. + +``` + /** \\brief DPDK interface HQoS tctbl entry set request + @param client_index - opaque cookie to identify the sender + @param context - sender context, to match reply w/ request + @param sw_if_index - the interface + @param entry - entry index ID + @param tc - traffic class (0 .. 3) + @param queue - traffic class queue (0 .. 3) + */ + define sw_interface_set_dpdk_hqos_tctbl { + u32 client_index; + u32 context; + u32 sw_if_index; + u32 entry; + u32 tc; + u32 queue; + }; + + /** \\brief DPDK interface HQoS tctbl entry set reply + @param context - sender context, to match reply w/ request + @param retval - request return code + */ + define sw_interface_set_dpdk_hqos_tctbl_reply { + u32 context; + i32 retval; + }; +``` diff --git a/src/vnet/devices/netmap/cli.c b/src/vnet/devices/netmap/cli.c new file mode 100644 index 00000000000..68695250506 --- /dev/null +++ b/src/vnet/devices/netmap/cli.c @@ -0,0 +1,146 @@ +/* + *------------------------------------------------------------------ + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ +#include +#include +#include + +#include +#include +#include + +#include +#include + +static clib_error_t * +netmap_create_command_fn (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + u8 *host_if_name = NULL; + u8 hwaddr[6]; + u8 *hw_addr_ptr = 0; + int r; + u8 is_pipe = 0; + u8 is_master = 0; + u32 sw_if_index = ~0; + + /* Get a line of input. */ + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "name %s", &host_if_name)) + ; + else + if (unformat + (line_input, "hw-addr %U", unformat_ethernet_address, hwaddr)) + hw_addr_ptr = hwaddr; + else if (unformat (line_input, "pipe")) + is_pipe = 1; + else if (unformat (line_input, "master")) + is_master = 1; + else if (unformat (line_input, "slave")) + is_master = 0; + else + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + unformat_free (line_input); + + if (host_if_name == NULL) + return clib_error_return (0, "missing host interface name"); + + r = + netmap_create_if (vm, host_if_name, hw_addr_ptr, is_pipe, is_master, + &sw_if_index); + + if (r == VNET_API_ERROR_SYSCALL_ERROR_1) + return clib_error_return (0, "%s (errno %d)", strerror (errno), errno); + + if (r == VNET_API_ERROR_INVALID_INTERFACE) + return clib_error_return (0, "Invalid interface name"); + + if (r == VNET_API_ERROR_SUBIF_ALREADY_EXISTS) + return clib_error_return (0, "Interface already exists"); + + vlib_cli_output (vm, "%U\n", format_vnet_sw_if_index_name, vnet_get_main (), + sw_if_index); + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (netmap_create_command, static) = { + .path = "create netmap", + .short_help = "create netmap name [|valeXXX:YYY] " + "[hw-addr ] [pipe] [master|slave]", + .function = netmap_create_command_fn, +}; +/* *INDENT-ON* */ + +static clib_error_t * +netmap_delete_command_fn (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + u8 *host_if_name = NULL; + + /* Get a line of input. */ + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "name %s", &host_if_name)) + ; + else + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + unformat_free (line_input); + + if (host_if_name == NULL) + return clib_error_return (0, "missing host interface name"); + + netmap_delete_if (vm, host_if_name); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (netmap_delete_command, static) = { + .path = "delete netmap", + .short_help = "delete netmap name ", + .function = netmap_delete_command_fn, +}; +/* *INDENT-ON* */ + +clib_error_t * +netmap_cli_init (vlib_main_t * vm) +{ + return 0; +} + +VLIB_INIT_FUNCTION (netmap_cli_init); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/netmap/device.c b/src/vnet/devices/netmap/device.c new file mode 100644 index 00000000000..2152824f733 --- /dev/null +++ b/src/vnet/devices/netmap/device.c @@ -0,0 +1,261 @@ +/* + *------------------------------------------------------------------ + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include +#include +#include + +#include +#include +#include + +#include +#include + +#define foreach_netmap_tx_func_error \ +_(NO_FREE_SLOTS, "no free tx slots") \ +_(PENDING_MSGS, "pending msgs in tx ring") + +typedef enum +{ +#define _(f,s) NETMAP_TX_ERROR_##f, + foreach_netmap_tx_func_error +#undef _ + NETMAP_TX_N_ERROR, +} netmap_tx_func_error_t; + +static char *netmap_tx_func_error_strings[] = { +#define _(n,s) s, + foreach_netmap_tx_func_error +#undef _ +}; + + +static u8 * +format_netmap_device_name (u8 * s, va_list * args) +{ + u32 i = va_arg (*args, u32); + netmap_main_t *apm = &netmap_main; + netmap_if_t *nif = pool_elt_at_index (apm->interfaces, i); + + s = format (s, "netmap-%s", nif->host_if_name); + return s; +} + +static u8 * +format_netmap_device (u8 * s, va_list * args) +{ + u32 dev_instance = va_arg (*args, u32); + int verbose = va_arg (*args, int); + netmap_main_t *nm = &netmap_main; + netmap_if_t *nif = vec_elt_at_index (nm->interfaces, dev_instance); + uword indent = format_get_indent (s); + + s = format (s, "NETMAP interface"); + if (verbose) + { + s = format (s, "\n%U version %d flags 0x%x" + "\n%U region %u memsize 0x%x offset 0x%x" + "\n%U tx_slots %u rx_slots %u tx_rings %u rx_rings %u", + format_white_space, indent + 2, + nif->req->nr_version, + nif->req->nr_flags, + format_white_space, indent + 2, + nif->mem_region, + nif->req->nr_memsize, + nif->req->nr_offset, + format_white_space, indent + 2, + nif->req->nr_tx_slots, + nif->req->nr_rx_slots, + nif->req->nr_tx_rings, nif->req->nr_rx_rings); + } + return s; +} + +static u8 * +format_netmap_tx_trace (u8 * s, va_list * args) +{ + s = format (s, "Unimplemented..."); + return s; +} + +static uword +netmap_interface_tx (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * frame) +{ + netmap_main_t *nm = &netmap_main; + u32 *buffers = vlib_frame_args (frame); + u32 n_left = frame->n_vectors; + f64 const time_constant = 1e3; + vnet_interface_output_runtime_t *rd = (void *) node->runtime_data; + netmap_if_t *nif = pool_elt_at_index (nm->interfaces, rd->dev_instance); + int cur_ring; + + if (PREDICT_FALSE (nif->lockp != 0)) + { + while (__sync_lock_test_and_set (nif->lockp, 1)) + ; + } + + cur_ring = nif->first_tx_ring; + + while (n_left && cur_ring <= nif->last_tx_ring) + { + struct netmap_ring *ring = NETMAP_TXRING (nif->nifp, cur_ring); + int n_free_slots = nm_ring_space (ring); + uint cur = ring->cur; + + if (nm_tx_pending (ring)) + { + if (ioctl (nif->fd, NIOCTXSYNC, NULL) < 0) + clib_unix_warning ("NIOCTXSYNC"); + clib_cpu_time_wait (time_constant); + + if (nm_tx_pending (ring) && !n_free_slots) + { + cur_ring++; + continue; + } + } + + while (n_left && n_free_slots) + { + vlib_buffer_t *b0 = 0; + u32 bi = buffers[0]; + u32 len; + u32 offset = 0; + buffers++; + + struct netmap_slot *slot = &ring->slot[cur]; + + do + { + b0 = vlib_get_buffer (vm, bi); + len = b0->current_length; + /* memcpy */ + clib_memcpy ((u8 *) NETMAP_BUF (ring, slot->buf_idx) + offset, + vlib_buffer_get_current (b0), len); + offset += len; + } + while ((bi = b0->next_buffer)); + + slot->len = offset; + cur = (cur + 1) % ring->num_slots; + n_free_slots--; + n_left--; + } + CLIB_MEMORY_BARRIER (); + ring->head = ring->cur = cur; + } + + if (n_left < frame->n_vectors) + ioctl (nif->fd, NIOCTXSYNC, NULL); + + if (PREDICT_FALSE (nif->lockp != 0)) + *nif->lockp = 0; + + if (n_left) + vlib_error_count (vm, node->node_index, + (n_left == + frame->n_vectors ? NETMAP_TX_ERROR_PENDING_MSGS : + NETMAP_TX_ERROR_NO_FREE_SLOTS), n_left); + + vlib_buffer_free (vm, vlib_frame_args (frame), frame->n_vectors); + return frame->n_vectors; +} + +static void +netmap_set_interface_next_node (vnet_main_t * vnm, u32 hw_if_index, + u32 node_index) +{ + netmap_main_t *apm = &netmap_main; + vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index); + netmap_if_t *nif = pool_elt_at_index (apm->interfaces, hw->dev_instance); + + /* Shut off redirection */ + if (node_index == ~0) + { + nif->per_interface_next_index = node_index; + return; + } + + nif->per_interface_next_index = + vlib_node_add_next (vlib_get_main (), netmap_input_node.index, + node_index); +} + +static void +netmap_clear_hw_interface_counters (u32 instance) +{ + /* Nothing for now */ +} + +static clib_error_t * +netmap_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, u32 flags) +{ + netmap_main_t *apm = &netmap_main; + vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index); + netmap_if_t *nif = pool_elt_at_index (apm->interfaces, hw->dev_instance); + u32 hw_flags; + + nif->is_admin_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0; + + if (nif->is_admin_up) + hw_flags = VNET_HW_INTERFACE_FLAG_LINK_UP; + else + hw_flags = 0; + + vnet_hw_interface_set_flags (vnm, hw_if_index, hw_flags); + + return 0; +} + +static clib_error_t * +netmap_subif_add_del_function (vnet_main_t * vnm, + u32 hw_if_index, + struct vnet_sw_interface_t *st, int is_add) +{ + /* Nothing for now */ + return 0; +} + +/* *INDENT-OFF* */ +VNET_DEVICE_CLASS (netmap_device_class) = { + .name = "netmap", + .tx_function = netmap_interface_tx, + .format_device_name = format_netmap_device_name, + .format_device = format_netmap_device, + .format_tx_trace = format_netmap_tx_trace, + .tx_function_n_errors = NETMAP_TX_N_ERROR, + .tx_function_error_strings = netmap_tx_func_error_strings, + .rx_redirect_to_node = netmap_set_interface_next_node, + .clear_counters = netmap_clear_hw_interface_counters, + .admin_up_down_function = netmap_interface_admin_up_down, + .subif_add_del_function = netmap_subif_add_del_function, +}; + +VLIB_DEVICE_TX_FUNCTION_MULTIARCH(netmap_device_class, + netmap_interface_tx) +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/netmap/net_netmap.h b/src/vnet/devices/netmap/net_netmap.h new file mode 100644 index 00000000000..fd4253b7c0c --- /dev/null +++ b/src/vnet/devices/netmap/net_netmap.h @@ -0,0 +1,650 @@ +/* + * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``S IS''AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + +/* + * $FreeBSD: head/sys/net/netmap.h 251139 2013-05-30 14:07:14Z luigi $ + * + * Definitions of constants and the structures used by the netmap + * framework, for the part visible to both kernel and userspace. + * Detailed info on netmap is available with "man netmap" or at + * + * http://info.iet.unipi.it/~luigi/netmap/ + * + * This API is also used to communicate with the VALE software switch + */ + +#ifndef _NET_NETMAP_H_ +#define _NET_NETMAP_H_ + +#define NETMAP_API 11 /* current API version */ + +#define NETMAP_MIN_API 11 /* min and max versions accepted */ +#define NETMAP_MAX_API 15 +/* + * Some fields should be cache-aligned to reduce contention. + * The alignment is architecture and OS dependent, but rather than + * digging into OS headers to find the exact value we use an estimate + * that should cover most architectures. + */ +#define NM_CACHE_ALIGN 128 + +/* + * --- Netmap data structures --- + * + * The userspace data structures used by netmap are shown below. + * They are allocated by the kernel and mmap()ed by userspace threads. + * Pointers are implemented as memory offsets or indexes, + * so that they can be easily dereferenced in kernel and userspace. + + KERNEL (opaque, obviously) + + ==================================================================== + | + USERSPACE | struct netmap_ring + +---->+---------------+ + / | head,cur,tail | + struct netmap_if (nifp, 1 per fd) / | buf_ofs | + +---------------+ / | other fields | + | ni_tx_rings | / +===============+ + | ni_rx_rings | / | buf_idx, len | slot[0] + | | / | flags, ptr | + | | / +---------------+ + +===============+ / | buf_idx, len | slot[1] + | txring_ofs[0] | (rel.to nifp)--' | flags, ptr | + | txring_ofs[1] | +---------------+ + (tx+1 entries) (num_slots entries) + | txring_ofs[t] | | buf_idx, len | slot[n-1] + +---------------+ | flags, ptr | + | rxring_ofs[0] | +---------------+ + | rxring_ofs[1] | + (rx+1 entries) + | rxring_ofs[r] | + +---------------+ + + * For each "interface" (NIC, host stack, PIPE, VALE switch port) bound to + * a file descriptor, the mmap()ed region contains a (logically readonly) + * struct netmap_if pointing to struct netmap_ring's. + * + * There is one netmap_ring per physical NIC ring, plus one tx/rx ring + * pair attached to the host stack (this pair is unused for non-NIC ports). + * + * All physical/host stack ports share the same memory region, + * so that zero-copy can be implemented between them. + * VALE switch ports instead have separate memory regions. + * + * The netmap_ring is the userspace-visible replica of the NIC ring. + * Each slot has the index of a buffer (MTU-sized and residing in the + * mmapped region), its length and some flags. An extra 64-bit pointer + * is provided for user-supplied buffers in the tx path. + * + * In user space, the buffer address is computed as + * (char *)ring + buf_ofs + index * NETMAP_BUF_SIZE + * + * Added in NETMAP_API 11: + * + * + NIOCREGIF can request the allocation of extra spare buffers from + * the same memory pool. The desired number of buffers must be in + * nr_arg3. The ioctl may return fewer buffers, depending on memory + * availability. nr_arg3 will return the actual value, and, once + * mapped, nifp->ni_bufs_head will be the index of the first buffer. + * + * The buffers are linked to each other using the first uint32_t + * as the index. On close, ni_bufs_head must point to the list of + * buffers to be released. + * + * + NIOCREGIF can request space for extra rings (and buffers) + * allocated in the same memory space. The number of extra rings + * is in nr_arg1, and is advisory. This is a no-op on NICs where + * the size of the memory space is fixed. + * + * + NIOCREGIF can attach to PIPE rings sharing the same memory + * space with a parent device. The ifname indicates the parent device, + * which must already exist. Flags in nr_flags indicate if we want to + * bind the master or slave side, the index (from nr_ringid) + * is just a cookie and does not need to be sequential. + * + * + NIOCREGIF can also attach to 'monitor' rings that replicate + * the content of specific rings, also from the same memory space. + * + * Extra flags in nr_flags support the above functions. + * Application libraries may use the following naming scheme: + * netmap:foo all NIC ring pairs + * netmap:foo^ only host ring pair + * netmap:foo+ all NIC ring + host ring pairs + * netmap:foo-k the k-th NIC ring pair + * netmap:foo{k PIPE ring pair k, master side + * netmap:foo}k PIPE ring pair k, slave side + */ + +/* + * struct netmap_slot is a buffer descriptor + */ +struct netmap_slot { + uint32_t buf_idx; /* buffer index */ + uint16_t len; /* length for this slot */ + uint16_t flags; /* buf changed, etc. */ + uint64_t ptr; /* pointer for indirect buffers */ +}; + +/* + * The following flags control how the slot is used + */ + +#define NS_BUF_CHANGED 0x0001 /* buf_idx changed */ + /* + * must be set whenever buf_idx is changed (as it might be + * necessary to recompute the physical address and mapping) + * + * It is also set by the kernel whenever the buf_idx is + * changed internally (e.g., by pipes). Applications may + * use this information to know when they can reuse the + * contents of previously prepared buffers. + */ + +#define NS_REPORT 0x0002 /* ask the hardware to report results */ + /* + * Request notification when slot is used by the hardware. + * Normally transmit completions are handled lazily and + * may be unreported. This flag lets us know when a slot + * has been sent (e.g. to terminate the sender). + */ + +#define NS_FORWARD 0x0004 /* pass packet 'forward' */ + /* + * (Only for physical ports, rx rings with NR_FORWARD set). + * Slot released to the kernel (i.e. before ring->head) with + * this flag set are passed to the peer ring (host/NIC), + * thus restoring the host-NIC connection for these slots. + * This supports efficient traffic monitoring or firewalling. + */ + +#define NS_NO_LEARN 0x0008 /* disable bridge learning */ + /* + * On a VALE switch, do not 'learn' the source port for + * this buffer. + */ + +#define NS_INDIRECT 0x0010 /* userspace buffer */ + /* + * (VALE tx rings only) data is in a userspace buffer, + * whose address is in the 'ptr' field in the slot. + */ + +#define NS_MOREFRAG 0x0020 /* packet has more fragments */ + /* + * (VALE ports only) + * Set on all but the last slot of a multi-segment packet. + * The 'len' field refers to the individual fragment. + */ + +#define NS_PORT_SHIFT 8 +#define NS_PORT_MASK (0xff << NS_PORT_SHIFT) + /* + * The high 8 bits of the flag, if not zero, indicate the + * destination port for the VALE switch, overriding + * the lookup table. + */ + +#define NS_RFRAGS(_slot) ( ((_slot)->flags >> 8) & 0xff) + /* + * (VALE rx rings only) the high 8 bits + * are the number of fragments. + */ + + +/* + * struct netmap_ring + * + * Netmap representation of a TX or RX ring (also known as "queue"). + * This is a queue implemented as a fixed-size circular array. + * At the software level the important fields are: head, cur, tail. + * + * In TX rings: + * + * head first slot available for transmission. + * cur wakeup point. select() and poll() will unblock + * when 'tail' moves past 'cur' + * tail (readonly) first slot reserved to the kernel + * + * [head .. tail-1] can be used for new packets to send; + * 'head' and 'cur' must be incremented as slots are filled + * with new packets to be sent; + * 'cur' can be moved further ahead if we need more space + * for new transmissions. XXX todo (2014-03-12) + * + * In RX rings: + * + * head first valid received packet + * cur wakeup point. select() and poll() will unblock + * when 'tail' moves past 'cur' + * tail (readonly) first slot reserved to the kernel + * + * [head .. tail-1] contain received packets; + * 'head' and 'cur' must be incremented as slots are consumed + * and can be returned to the kernel; + * 'cur' can be moved further ahead if we want to wait for + * new packets without returning the previous ones. + * + * DATA OWNERSHIP/LOCKING: + * The netmap_ring, and all slots and buffers in the range + * [head .. tail-1] are owned by the user program; + * the kernel only accesses them during a netmap system call + * and in the user thread context. + * + * Other slots and buffers are reserved for use by the kernel + */ +struct netmap_ring { + /* + * buf_ofs is meant to be used through macros. + * It contains the offset of the buffer region from this + * descriptor. + */ + const int64_t buf_ofs; + const uint32_t num_slots; /* number of slots in the ring. */ + const uint32_t nr_buf_size; + const uint16_t ringid; + const uint16_t dir; /* 0: tx, 1: rx */ + + uint32_t head; /* (u) first user slot */ + uint32_t cur; /* (u) wakeup point */ + uint32_t tail; /* (k) first kernel slot */ + + uint32_t flags; + + struct timeval ts; /* (k) time of last *sync() */ + + /* opaque room for a mutex or similar object */ +#if !defined(_WIN32) || defined(__CYGWIN__) + uint8_t __attribute__((__aligned__(NM_CACHE_ALIGN))) sem[128]; +#else + uint8_t __declspec(align(NM_CACHE_ALIGN)) sem[128]; +#endif + + /* the slots follow. This struct has variable size */ + struct netmap_slot slot[0]; /* array of slots. */ +}; + + +/* + * RING FLAGS + */ +#define NR_TIMESTAMP 0x0002 /* set timestamp on *sync() */ + /* + * updates the 'ts' field on each netmap syscall. This saves + * saves a separate gettimeofday(), and is not much worse than + * software timestamps generated in the interrupt handler. + */ + +#define NR_FORWARD 0x0004 /* enable NS_FORWARD for ring */ + /* + * Enables the NS_FORWARD slot flag for the ring. + */ + + +/* + * Netmap representation of an interface and its queue(s). + * This is initialized by the kernel when binding a file + * descriptor to a port, and should be considered as readonly + * by user programs. The kernel never uses it. + * + * There is one netmap_if for each file descriptor on which we want + * to select/poll. + * select/poll operates on one or all pairs depending on the value of + * nmr_queueid passed on the ioctl. + */ +struct netmap_if { + char ni_name[IFNAMSIZ]; /* name of the interface. */ + const uint32_t ni_version; /* API version, currently unused */ + const uint32_t ni_flags; /* properties */ +#define NI_PRIV_MEM 0x1 /* private memory region */ + + /* + * The number of packet rings available in netmap mode. + * Physical NICs can have different numbers of tx and rx rings. + * Physical NICs also have a 'host' ring pair. + * Additionally, clients can request additional ring pairs to + * be used for internal communication. + */ + const uint32_t ni_tx_rings; /* number of HW tx rings */ + const uint32_t ni_rx_rings; /* number of HW rx rings */ + + uint32_t ni_bufs_head; /* head index for extra bufs */ + uint32_t ni_spare1[5]; + /* + * The following array contains the offset of each netmap ring + * from this structure, in the following order: + * NIC tx rings (ni_tx_rings); host tx ring (1); extra tx rings; + * NIC rx rings (ni_rx_rings); host tx ring (1); extra rx rings. + * + * The area is filled up by the kernel on NIOCREGIF, + * and then only read by userspace code. + */ + const ssize_t ring_ofs[0]; +}; + + +#ifndef NIOCREGIF +/* + * ioctl names and related fields + * + * NIOCTXSYNC, NIOCRXSYNC synchronize tx or rx queues, + * whose identity is set in NIOCREGIF through nr_ringid. + * These are non blocking and take no argument. + * + * NIOCGINFO takes a struct ifreq, the interface name is the input, + * the outputs are number of queues and number of descriptor + * for each queue (useful to set number of threads etc.). + * The info returned is only advisory and may change before + * the interface is bound to a file descriptor. + * + * NIOCREGIF takes an interface name within a struct nmre, + * and activates netmap mode on the interface (if possible). + * + * The argument to NIOCGINFO/NIOCREGIF overlays struct ifreq so we + * can pass it down to other NIC-related ioctls. + * + * The actual argument (struct nmreq) has a number of options to request + * different functions. + * The following are used in NIOCREGIF when nr_cmd == 0: + * + * nr_name (in) + * The name of the port (em0, valeXXX:YYY, etc.) + * limited to IFNAMSIZ for backward compatibility. + * + * nr_version (in/out) + * Must match NETMAP_API as used in the kernel, error otherwise. + * Always returns the desired value on output. + * + * nr_tx_slots, nr_tx_slots, nr_tx_rings, nr_rx_rings (in/out) + * On input, non-zero values may be used to reconfigure the port + * according to the requested values, but this is not guaranteed. + * On output the actual values in use are reported. + * + * nr_ringid (in) + * Indicates how rings should be bound to the file descriptors. + * If nr_flags != 0, then the low bits (in NETMAP_RING_MASK) + * are used to indicate the ring number, and nr_flags specifies + * the actual rings to bind. NETMAP_NO_TX_POLL is unaffected. + * + * NOTE: THE FOLLOWING (nr_flags == 0) IS DEPRECATED: + * If nr_flags == 0, NETMAP_HW_RING and NETMAP_SW_RING control + * the binding as follows: + * 0 (default) binds all physical rings + * NETMAP_HW_RING | ring number binds a single ring pair + * NETMAP_SW_RING binds only the host tx/rx rings + * + * NETMAP_NO_TX_POLL can be OR-ed to make select()/poll() push + * packets on tx rings only if POLLOUT is set. + * The default is to push any pending packet. + * + * NETMAP_DO_RX_POLL can be OR-ed to make select()/poll() release + * packets on rx rings also when POLLIN is NOT set. + * The default is to touch the rx ring only with POLLIN. + * Note that this is the opposite of TX because it + * reflects the common usage. + * + * NOTE: NETMAP_PRIV_MEM IS DEPRECATED, use nr_arg2 instead. + * NETMAP_PRIV_MEM is set on return for ports that do not use + * the global memory allocator. + * This information is not significant and applications + * should look at the region id in nr_arg2 + * + * nr_flags is the recommended mode to indicate which rings should + * be bound to a file descriptor. Values are NR_REG_* + * + * nr_arg1 (in) The number of extra rings to be reserved. + * Especially when allocating a VALE port the system only + * allocates the amount of memory needed for the port. + * If more shared memory rings are desired (e.g. for pipes), + * the first invocation for the same basename/allocator + * should specify a suitable number. Memory cannot be + * extended after the first allocation without closing + * all ports on the same region. + * + * nr_arg2 (in/out) The identity of the memory region used. + * On input, 0 means the system decides autonomously, + * other values may try to select a specific region. + * On return the actual value is reported. + * Region '1' is the global allocator, normally shared + * by all interfaces. Other values are private regions. + * If two ports the same region zero-copy is possible. + * + * nr_arg3 (in/out) number of extra buffers to be allocated. + * + * + * + * nr_cmd (in) if non-zero indicates a special command: + * NETMAP_BDG_ATTACH and nr_name = vale*:ifname + * attaches the NIC to the switch; nr_ringid specifies + * which rings to use. Used by vale-ctl -a ... + * nr_arg1 = NETMAP_BDG_HOST also attaches the host port + * as in vale-ctl -h ... + * + * NETMAP_BDG_DETACH and nr_name = vale*:ifname + * disconnects a previously attached NIC. + * Used by vale-ctl -d ... + * + * NETMAP_BDG_LIST + * list the configuration of VALE switches. + * + * NETMAP_BDG_VNET_HDR + * Set the virtio-net header length used by the client + * of a VALE switch port. + * + * NETMAP_BDG_NEWIF + * create a persistent VALE port with name nr_name. + * Used by vale-ctl -n ... + * + * NETMAP_BDG_DELIF + * delete a persistent VALE port. Used by vale-ctl -d ... + * + * nr_arg1, nr_arg2, nr_arg3 (in/out) command specific + * + * + * + */ + + +/* + * struct nmreq overlays a struct ifreq (just the name) + */ +struct nmreq { + char nr_name[IFNAMSIZ]; + uint32_t nr_version; /* API version */ + uint32_t nr_offset; /* nifp offset in the shared region */ + uint32_t nr_memsize; /* size of the shared region */ + uint32_t nr_tx_slots; /* slots in tx rings */ + uint32_t nr_rx_slots; /* slots in rx rings */ + uint16_t nr_tx_rings; /* number of tx rings */ + uint16_t nr_rx_rings; /* number of rx rings */ + + uint16_t nr_ringid; /* ring(s) we care about */ +#define NETMAP_HW_RING 0x4000 /* single NIC ring pair */ +#define NETMAP_SW_RING 0x2000 /* only host ring pair */ + +#define NETMAP_RING_MASK 0x0fff /* the ring number */ + +#define NETMAP_NO_TX_POLL 0x1000 /* no automatic txsync on poll */ + +#define NETMAP_DO_RX_POLL 0x8000 /* DO automatic rxsync on poll */ + + uint16_t nr_cmd; +#define NETMAP_BDG_ATTACH 1 /* attach the NIC */ +#define NETMAP_BDG_DETACH 2 /* detach the NIC */ +#define NETMAP_BDG_REGOPS 3 /* register bridge callbacks */ +#define NETMAP_BDG_LIST 4 /* get bridge's info */ +#define NETMAP_BDG_VNET_HDR 5 /* set the port virtio-net-hdr length */ +#define NETMAP_BDG_OFFSET NETMAP_BDG_VNET_HDR /* deprecated alias */ +#define NETMAP_BDG_NEWIF 6 /* create a virtual port */ +#define NETMAP_BDG_DELIF 7 /* destroy a virtual port */ +#define NETMAP_PT_HOST_CREATE 8 /* create ptnetmap kthreads */ +#define NETMAP_PT_HOST_DELETE 9 /* delete ptnetmap kthreads */ +#define NETMAP_BDG_POLLING_ON 10 /* delete polling kthread */ +#define NETMAP_BDG_POLLING_OFF 11 /* delete polling kthread */ +#define NETMAP_VNET_HDR_GET 12 /* get the port virtio-net-hdr length */ + uint16_t nr_arg1; /* reserve extra rings in NIOCREGIF */ +#define NETMAP_BDG_HOST 1 /* attach the host stack on ATTACH */ + + uint16_t nr_arg2; + uint32_t nr_arg3; /* req. extra buffers in NIOCREGIF */ + uint32_t nr_flags; + /* various modes, extends nr_ringid */ + uint32_t spare2[1]; +}; + +#define NR_REG_MASK 0xf /* values for nr_flags */ +enum { NR_REG_DEFAULT = 0, /* backward compat, should not be used. */ + NR_REG_ALL_NIC = 1, + NR_REG_SW = 2, + NR_REG_NIC_SW = 3, + NR_REG_ONE_NIC = 4, + NR_REG_PIPE_MASTER = 5, + NR_REG_PIPE_SLAVE = 6, +}; +/* monitor uses the NR_REG to select the rings to monitor */ +#define NR_MONITOR_TX 0x100 +#define NR_MONITOR_RX 0x200 +#define NR_ZCOPY_MON 0x400 +/* request exclusive access to the selected rings */ +#define NR_EXCLUSIVE 0x800 +/* request ptnetmap host support */ +#define NR_PASSTHROUGH_HOST NR_PTNETMAP_HOST /* deprecated */ +#define NR_PTNETMAP_HOST 0x1000 +#define NR_RX_RINGS_ONLY 0x2000 +#define NR_TX_RINGS_ONLY 0x4000 +/* Applications set this flag if they are able to deal with virtio-net headers, + * that is send/receive frames that start with a virtio-net header. + * If not set, NIOCREGIF will fail with netmap ports that require applications + * to use those headers. If the flag is set, the application can use the + * NETMAP_VNET_HDR_GET command to figure out the header length. */ +#define NR_ACCEPT_VNET_HDR 0x8000 + + +/* + * Windows does not have _IOWR(). _IO(), _IOW() and _IOR() are defined + * in ws2def.h but not sure if they are in the form we need. + * XXX so we redefine them + * in a convenient way to use for DeviceIoControl signatures + */ +#ifdef _WIN32 +#undef _IO // ws2def.h +#define _WIN_NM_IOCTL_TYPE 40000 +#define _IO(_c, _n) CTL_CODE(_WIN_NM_IOCTL_TYPE, ((_n) + 0x800) , \ + METHOD_BUFFERED, FILE_ANY_ACCESS ) +#define _IO_direct(_c, _n) CTL_CODE(_WIN_NM_IOCTL_TYPE, ((_n) + 0x800) , \ + METHOD_OUT_DIRECT, FILE_ANY_ACCESS ) + +#define _IOWR(_c, _n, _s) _IO(_c, _n) + +/* We havesome internal sysctl in addition to the externally visible ones */ +#define NETMAP_MMAP _IO_direct('i', 160) // note METHOD_OUT_DIRECT +#define NETMAP_POLL _IO('i', 162) + +/* and also two setsockopt for sysctl emulation */ +#define NETMAP_SETSOCKOPT _IO('i', 140) +#define NETMAP_GETSOCKOPT _IO('i', 141) + + +//These linknames are for the Netmap Core Driver +#define NETMAP_NT_DEVICE_NAME L"\\Device\\NETMAP" +#define NETMAP_DOS_DEVICE_NAME L"\\DosDevices\\netmap" + +//Definition of a structure used to pass a virtual address within an IOCTL +typedef struct _MEMORY_ENTRY { + PVOID pUsermodeVirtualAddress; +} MEMORY_ENTRY, *PMEMORY_ENTRY; + +typedef struct _POLL_REQUEST_DATA { + int events; + int timeout; + int revents; +} POLL_REQUEST_DATA; + +#endif /* _WIN32 */ + +/* + * FreeBSD uses the size value embedded in the _IOWR to determine + * how much to copy in/out. So we need it to match the actual + * data structure we pass. We put some spares in the structure + * to ease compatibility with other versions + */ +#define NIOCGINFO _IOWR('i', 145, struct nmreq) /* return IF info */ +#define NIOCREGIF _IOWR('i', 146, struct nmreq) /* interface register */ +#define NIOCTXSYNC _IO('i', 148) /* sync tx queues */ +#define NIOCRXSYNC _IO('i', 149) /* sync rx queues */ +#define NIOCCONFIG _IOWR('i',150, struct nm_ifreq) /* for ext. modules */ +#endif /* !NIOCREGIF */ + + +/* + * Helper functions for kernel and userspace + */ + +/* + * check if space is available in the ring. + */ +static inline int +nm_ring_empty(struct netmap_ring *ring) +{ + return (ring->cur == ring->tail); +} + +/* + * Opaque structure that is passed to an external kernel + * module via ioctl(fd, NIOCCONFIG, req) for a user-owned + * bridge port (at this point ephemeral VALE interface). + */ +#define NM_IFRDATA_LEN 256 +struct nm_ifreq { + char nifr_name[IFNAMSIZ]; + char data[NM_IFRDATA_LEN]; +}; + +/* + * netmap kernel thread configuration + */ +/* bhyve/vmm.ko MSIX parameters for IOCTL */ +struct ptn_vmm_ioctl_msix { + uint64_t msg; + uint64_t addr; +}; + +/* IOCTL parameters */ +struct nm_kth_ioctl { + u_long com; + /* TODO: use union */ + union { + struct ptn_vmm_ioctl_msix msix; + } data; +}; + +/* Configuration of a ptnetmap ring */ +struct ptnet_ring_cfg { + uint64_t ioeventfd; /* eventfd in linux, tsleep() parameter in FreeBSD */ + uint64_t irqfd; /* eventfd in linux, ioctl fd in FreeBSD */ + struct nm_kth_ioctl ioctl; /* ioctl parameter to send irq (only used in bhyve/FreeBSD) */ +}; +#endif /* _NET_NETMAP_H_ */ diff --git a/src/vnet/devices/netmap/netmap.api b/src/vnet/devices/netmap/netmap.api new file mode 100644 index 00000000000..377ccffda4c --- /dev/null +++ b/src/vnet/devices/netmap/netmap.api @@ -0,0 +1,74 @@ +/* + * Copyright (c) 2015-2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/** \brief Create netmap + @param client_index - opaque cookie to identify the sender + @param context - sender context, to match reply w/ request + @param netmap_if_name - interface name + @param hw_addr - interface MAC + @param use_random_hw_addr - use random generated MAC + @param is_pipe - is pipe + @param is_master - 0=slave, 1=master +*/ +define netmap_create +{ + u32 client_index; + u32 context; + + u8 netmap_if_name[64]; + u8 hw_addr[6]; + u8 use_random_hw_addr; + u8 is_pipe; + u8 is_master; +}; + +/** \brief Create netmap response + @param context - sender context, to match reply w/ request + @param retval - return value for request +*/ +define netmap_create_reply +{ + u32 context; + i32 retval; +}; + +/** \brief Delete netmap + @param client_index - opaque cookie to identify the sender + @param context - sender context, to match reply w/ request + @param netmap_if_name - interface name +*/ +define netmap_delete +{ + u32 client_index; + u32 context; + + u8 netmap_if_name[64]; +}; + +/** \brief Delete netmap response + @param context - sender context, to match reply w/ request + @param retval - return value for request +*/ +define netmap_delete_reply +{ + u32 context; + i32 retval; +}; + +/* + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/netmap/netmap.c b/src/vnet/devices/netmap/netmap.c new file mode 100644 index 00000000000..3bdb442dda2 --- /dev/null +++ b/src/vnet/devices/netmap/netmap.c @@ -0,0 +1,316 @@ +/* + *------------------------------------------------------------------ + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include + +static u32 +netmap_eth_flag_change (vnet_main_t * vnm, vnet_hw_interface_t * hi, + u32 flags) +{ + /* nothing for now */ + return 0; +} + +static clib_error_t * +netmap_fd_read_ready (unix_file_t * uf) +{ + vlib_main_t *vm = vlib_get_main (); + netmap_main_t *nm = &netmap_main; + u32 idx = uf->private_data; + + nm->pending_input_bitmap = + clib_bitmap_set (nm->pending_input_bitmap, idx, 1); + + /* Schedule the rx node */ + vlib_node_set_interrupt_pending (vm, netmap_input_node.index); + + return 0; +} + +static void +close_netmap_if (netmap_main_t * nm, netmap_if_t * nif) +{ + if (nif->unix_file_index != ~0) + { + unix_file_del (&unix_main, unix_main.file_pool + nif->unix_file_index); + nif->unix_file_index = ~0; + } + else if (nif->fd > -1) + close (nif->fd); + + if (nif->mem_region) + { + netmap_mem_region_t *reg = &nm->mem_regions[nif->mem_region]; + if (--reg->refcnt == 0) + { + munmap (reg->mem, reg->region_size); + reg->region_size = 0; + } + } + + + mhash_unset (&nm->if_index_by_host_if_name, nif->host_if_name, + &nif->if_index); + vec_free (nif->host_if_name); + vec_free (nif->req); + + memset (nif, 0, sizeof (*nif)); + pool_put (nm->interfaces, nif); +} + +int +netmap_worker_thread_enable () +{ + /* if worker threads are enabled, switch to polling mode */ + foreach_vlib_main (( + { + vlib_node_set_state (this_vlib_main, + netmap_input_node.index, + VLIB_NODE_STATE_POLLING); + })); + + return 0; +} + +int +netmap_worker_thread_disable () +{ + foreach_vlib_main (( + { + vlib_node_set_state (this_vlib_main, + netmap_input_node.index, + VLIB_NODE_STATE_INTERRUPT); + })); + + return 0; +} + +int +netmap_create_if (vlib_main_t * vm, u8 * if_name, u8 * hw_addr_set, + u8 is_pipe, u8 is_master, u32 * sw_if_index) +{ + netmap_main_t *nm = &netmap_main; + int ret = 0; + netmap_if_t *nif = 0; + u8 hw_addr[6]; + clib_error_t *error = 0; + vnet_sw_interface_t *sw; + vnet_main_t *vnm = vnet_get_main (); + uword *p; + struct nmreq *req = 0; + netmap_mem_region_t *reg; + vlib_thread_main_t *tm = vlib_get_thread_main (); + int fd; + + p = mhash_get (&nm->if_index_by_host_if_name, if_name); + if (p) + return VNET_API_ERROR_SUBIF_ALREADY_EXISTS; + + fd = open ("/dev/netmap", O_RDWR); + if (fd < 0) + return VNET_API_ERROR_SUBIF_ALREADY_EXISTS; + + pool_get (nm->interfaces, nif); + nif->if_index = nif - nm->interfaces; + nif->fd = fd; + nif->unix_file_index = ~0; + + vec_validate (req, 0); + nif->req = req; + req->nr_version = NETMAP_API; + req->nr_flags = NR_REG_ALL_NIC; + + if (is_pipe) + req->nr_flags = is_master ? NR_REG_PIPE_MASTER : NR_REG_PIPE_SLAVE; + else + req->nr_flags = NR_REG_ALL_NIC; + + req->nr_flags |= NR_ACCEPT_VNET_HDR; + snprintf (req->nr_name, IFNAMSIZ, "%s", if_name); + req->nr_name[IFNAMSIZ - 1] = 0; + + if (ioctl (nif->fd, NIOCREGIF, req)) + { + ret = VNET_API_ERROR_NOT_CONNECTED; + goto error; + } + + nif->mem_region = req->nr_arg2; + vec_validate (nm->mem_regions, nif->mem_region); + reg = &nm->mem_regions[nif->mem_region]; + if (reg->region_size == 0) + { + reg->mem = mmap (NULL, req->nr_memsize, PROT_READ | PROT_WRITE, + MAP_SHARED, fd, 0); + clib_warning ("mem %p", reg->mem); + if (reg->mem == MAP_FAILED) + { + ret = VNET_API_ERROR_NOT_CONNECTED; + goto error; + } + reg->region_size = req->nr_memsize; + } + reg->refcnt++; + + nif->nifp = NETMAP_IF (reg->mem, req->nr_offset); + nif->first_rx_ring = 0; + nif->last_rx_ring = 0; + nif->first_tx_ring = 0; + nif->last_tx_ring = 0; + nif->host_if_name = if_name; + nif->per_interface_next_index = ~0; + + if (tm->n_vlib_mains > 1) + { + nif->lockp = clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES, + CLIB_CACHE_LINE_BYTES); + memset ((void *) nif->lockp, 0, CLIB_CACHE_LINE_BYTES); + } + + { + unix_file_t template = { 0 }; + template.read_function = netmap_fd_read_ready; + template.file_descriptor = nif->fd; + template.private_data = nif->if_index; + nif->unix_file_index = unix_file_add (&unix_main, &template); + } + + /*use configured or generate random MAC address */ + if (hw_addr_set) + memcpy (hw_addr, hw_addr_set, 6); + else + { + f64 now = vlib_time_now (vm); + u32 rnd; + rnd = (u32) (now * 1e6); + rnd = random_u32 (&rnd); + + memcpy (hw_addr + 2, &rnd, sizeof (rnd)); + hw_addr[0] = 2; + hw_addr[1] = 0xfe; + } + + error = ethernet_register_interface (vnm, netmap_device_class.index, + nif->if_index, hw_addr, + &nif->hw_if_index, + netmap_eth_flag_change); + + if (error) + { + clib_error_report (error); + ret = VNET_API_ERROR_SYSCALL_ERROR_1; + goto error; + } + + sw = vnet_get_hw_sw_interface (vnm, nif->hw_if_index); + nif->sw_if_index = sw->sw_if_index; + + mhash_set_mem (&nm->if_index_by_host_if_name, if_name, &nif->if_index, 0); + + if (sw_if_index) + *sw_if_index = nif->sw_if_index; + + if (tm->n_vlib_mains > 1 && pool_elts (nm->interfaces) == 1) + netmap_worker_thread_enable (); + + return 0; + +error: + close_netmap_if (nm, nif); + return ret; +} + +int +netmap_delete_if (vlib_main_t * vm, u8 * host_if_name) +{ + vnet_main_t *vnm = vnet_get_main (); + netmap_main_t *nm = &netmap_main; + netmap_if_t *nif; + uword *p; + vlib_thread_main_t *tm = vlib_get_thread_main (); + + p = mhash_get (&nm->if_index_by_host_if_name, host_if_name); + if (p == NULL) + { + clib_warning ("Host interface %s does not exist", host_if_name); + return VNET_API_ERROR_SYSCALL_ERROR_1; + } + nif = pool_elt_at_index (nm->interfaces, p[0]); + + /* bring down the interface */ + vnet_hw_interface_set_flags (vnm, nif->hw_if_index, 0); + + ethernet_delete_interface (vnm, nif->hw_if_index); + + close_netmap_if (nm, nif); + + if (tm->n_vlib_mains > 1 && pool_elts (nm->interfaces) == 0) + netmap_worker_thread_disable (); + + return 0; +} + +static clib_error_t * +netmap_init (vlib_main_t * vm) +{ + netmap_main_t *nm = &netmap_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); + vlib_thread_registration_t *tr; + uword *p; + + memset (nm, 0, sizeof (netmap_main_t)); + + nm->input_cpu_first_index = 0; + nm->input_cpu_count = 1; + + /* find out which cpus will be used for input */ + p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + tr = p ? (vlib_thread_registration_t *) p[0] : 0; + + if (tr && tr->count > 0) + { + nm->input_cpu_first_index = tr->first_index; + nm->input_cpu_count = tr->count; + } + + mhash_init_vec_string (&nm->if_index_by_host_if_name, sizeof (uword)); + + vec_validate_aligned (nm->rx_buffers, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + + return 0; +} + +VLIB_INIT_FUNCTION (netmap_init); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/netmap/netmap.h b/src/vnet/devices/netmap/netmap.h new file mode 100644 index 00000000000..39a94043c3c --- /dev/null +++ b/src/vnet/devices/netmap/netmap.h @@ -0,0 +1,164 @@ +/* + *------------------------------------------------------------------ + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ +/* + * Copyright (C) 2011-2014 Universita` di Pisa. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + volatile u32 *lockp; + u8 *host_if_name; + uword if_index; + u32 hw_if_index; + u32 sw_if_index; + u32 unix_file_index; + + u32 per_interface_next_index; + u8 is_admin_up; + + /* netmap */ + struct nmreq *req; + u16 mem_region; + int fd; + struct netmap_if *nifp; + u16 first_tx_ring; + u16 last_tx_ring; + u16 first_rx_ring; + u16 last_rx_ring; + +} netmap_if_t; + +typedef struct +{ + char *mem; + u32 region_size; + int refcnt; +} netmap_mem_region_t; + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + netmap_if_t *interfaces; + + /* bitmap of pending rx interfaces */ + uword *pending_input_bitmap; + + /* rx buffer cache */ + u32 **rx_buffers; + + /* hash of host interface names */ + mhash_t if_index_by_host_if_name; + + /* vector of memory regions */ + netmap_mem_region_t *mem_regions; + + /* first cpu index */ + u32 input_cpu_first_index; + + /* total cpu count */ + u32 input_cpu_count; +} netmap_main_t; + +netmap_main_t netmap_main; +extern vnet_device_class_t netmap_device_class; +extern vlib_node_registration_t netmap_input_node; + +int netmap_create_if (vlib_main_t * vm, u8 * host_if_name, u8 * hw_addr_set, + u8 is_pipe, u8 is_master, u32 * sw_if_index); +int netmap_delete_if (vlib_main_t * vm, u8 * host_if_name); + + +/* Macros and helper functions from sys/net/netmap_user.h */ + +#ifdef _NET_NETMAP_H_ + +#define _NETMAP_OFFSET(type, ptr, offset) \ + ((type)(void *)((char *)(ptr) + (offset))) + +#define NETMAP_IF(_base, _ofs) _NETMAP_OFFSET(struct netmap_if *, _base, _ofs) + +#define NETMAP_TXRING(nifp, index) _NETMAP_OFFSET(struct netmap_ring *, \ + nifp, (nifp)->ring_ofs[index] ) + +#define NETMAP_RXRING(nifp, index) _NETMAP_OFFSET(struct netmap_ring *, \ + nifp, (nifp)->ring_ofs[index + (nifp)->ni_tx_rings + 1] ) + +#define NETMAP_BUF(ring, index) \ + ((char *)(ring) + (ring)->buf_ofs + ((index)*(ring)->nr_buf_size)) + +#define NETMAP_BUF_IDX(ring, buf) \ + ( ((char *)(buf) - ((char *)(ring) + (ring)->buf_ofs) ) / \ + (ring)->nr_buf_size ) + +static inline uint32_t +nm_ring_next (struct netmap_ring *ring, uint32_t i) +{ + return (PREDICT_FALSE (i + 1 == ring->num_slots) ? 0 : i + 1); +} + + +/* + * Return 1 if we have pending transmissions in the tx ring. + * When everything is complete ring->head = ring->tail + 1 (modulo ring size) + */ +static inline int +nm_tx_pending (struct netmap_ring *ring) +{ + return nm_ring_next (ring, ring->tail) != ring->head; +} + +static inline uint32_t +nm_ring_space (struct netmap_ring *ring) +{ + int ret = ring->tail - ring->cur; + if (ret < 0) + ret += ring->num_slots; + return ret; +} +#endif + + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/netmap/netmap_api.c b/src/vnet/devices/netmap/netmap_api.c new file mode 100644 index 00000000000..9a393b1fda4 --- /dev/null +++ b/src/vnet/devices/netmap/netmap_api.c @@ -0,0 +1,137 @@ +/* + *------------------------------------------------------------------ + * netmap_api.c - netmap api + * + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include +#include + +#include +#include +#include + +#include + +#define vl_typedefs /* define message structures */ +#include +#undef vl_typedefs + +#define vl_endianfun /* define message structures */ +#include +#undef vl_endianfun + +/* instantiate all the print functions we know about */ +#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__) +#define vl_printfun +#include +#undef vl_printfun + +#include + +#define foreach_vpe_api_msg \ +_(NETMAP_CREATE, netmap_create) \ +_(NETMAP_DELETE, netmap_delete) \ + +static void +vl_api_netmap_create_t_handler (vl_api_netmap_create_t * mp) +{ + vlib_main_t *vm = vlib_get_main (); + vl_api_netmap_create_reply_t *rmp; + int rv = 0; + u8 *if_name = NULL; + + if_name = format (0, "%s", mp->netmap_if_name); + vec_add1 (if_name, 0); + + rv = + netmap_create_if (vm, if_name, mp->use_random_hw_addr ? 0 : mp->hw_addr, + mp->is_pipe, mp->is_master, 0); + + vec_free (if_name); + + REPLY_MACRO (VL_API_NETMAP_CREATE_REPLY); +} + +static void +vl_api_netmap_delete_t_handler (vl_api_netmap_delete_t * mp) +{ + vlib_main_t *vm = vlib_get_main (); + vl_api_netmap_delete_reply_t *rmp; + int rv = 0; + u8 *if_name = NULL; + + if_name = format (0, "%s", mp->netmap_if_name); + vec_add1 (if_name, 0); + + rv = netmap_delete_if (vm, if_name); + + vec_free (if_name); + + REPLY_MACRO (VL_API_NETMAP_DELETE_REPLY); +} + +/* + * netmap_api_hookup + * Add vpe's API message handlers to the table. + * vlib has alread mapped shared memory and + * added the client registration handlers. + * See .../vlib-api/vlibmemory/memclnt_vlib.c:memclnt_process() + */ +#define vl_msg_name_crc_list +#include +#undef vl_msg_name_crc_list + +static void +setup_message_id_table (api_main_t * am) +{ +#define _(id,n,crc) vl_msg_api_add_msg_name_crc (am, #n "_" #crc, id); + foreach_vl_msg_name_crc_netmap; +#undef _ +} + +static clib_error_t * +netmap_api_hookup (vlib_main_t * vm) +{ + api_main_t *am = &api_main; + +#define _(N,n) \ + vl_msg_api_set_handlers(VL_API_##N, #n, \ + vl_api_##n##_t_handler, \ + vl_noop_handler, \ + vl_api_##n##_t_endian, \ + vl_api_##n##_t_print, \ + sizeof(vl_api_##n##_t), 1); + foreach_vpe_api_msg; +#undef _ + + /* + * Set up the (msg_name, crc, message-id) table + */ + setup_message_id_table (am); + + return 0; +} + +VLIB_API_INIT_FUNCTION (netmap_api_hookup); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/netmap/node.c b/src/vnet/devices/netmap/node.c new file mode 100644 index 00000000000..19895e4754a --- /dev/null +++ b/src/vnet/devices/netmap/node.c @@ -0,0 +1,300 @@ +/* + *------------------------------------------------------------------ + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include +#include +#include + +#include +#include +#include +#include +#include + +#include +#include + +#define foreach_netmap_input_error + +typedef enum +{ +#define _(f,s) NETMAP_INPUT_ERROR_##f, + foreach_netmap_input_error +#undef _ + NETMAP_INPUT_N_ERROR, +} netmap_input_error_t; + +static char *netmap_input_error_strings[] = { +#define _(n,s) s, + foreach_netmap_input_error +#undef _ +}; + +typedef struct +{ + u32 next_index; + u32 hw_if_index; + struct netmap_slot slot; +} netmap_input_trace_t; + +static u8 * +format_netmap_input_trace (u8 * s, va_list * args) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *); + netmap_input_trace_t *t = va_arg (*args, netmap_input_trace_t *); + uword indent = format_get_indent (s); + + s = format (s, "netmap: hw_if_index %d next-index %d", + t->hw_if_index, t->next_index); + s = format (s, "\n%Uslot: flags 0x%x len %u buf_idx %u", + format_white_space, indent + 2, + t->slot.flags, t->slot.len, t->slot.buf_idx); + return s; +} + +always_inline void +buffer_add_to_chain (vlib_main_t * vm, u32 bi, u32 first_bi, u32 prev_bi) +{ + vlib_buffer_t *b = vlib_get_buffer (vm, bi); + vlib_buffer_t *first_b = vlib_get_buffer (vm, first_bi); + vlib_buffer_t *prev_b = vlib_get_buffer (vm, prev_bi); + + /* update first buffer */ + first_b->total_length_not_including_first_buffer += b->current_length; + + /* update previous buffer */ + prev_b->next_buffer = bi; + prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT; + + /* update current buffer */ + b->next_buffer = 0; +} + +always_inline uword +netmap_device_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node, + vlib_frame_t * frame, netmap_if_t * nif) +{ + u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; + uword n_trace = vlib_get_trace_count (vm, node); + netmap_main_t *nm = &netmap_main; + u32 n_rx_packets = 0; + u32 n_rx_bytes = 0; + u32 *to_next = 0; + u32 n_free_bufs; + struct netmap_ring *ring; + int cur_ring; + u32 cpu_index = os_get_cpu_number (); + u32 n_buffer_bytes = vlib_buffer_free_list_buffer_size (vm, + VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); + + if (nif->per_interface_next_index != ~0) + next_index = nif->per_interface_next_index; + + n_free_bufs = vec_len (nm->rx_buffers[cpu_index]); + if (PREDICT_FALSE (n_free_bufs < VLIB_FRAME_SIZE)) + { + vec_validate (nm->rx_buffers[cpu_index], + VLIB_FRAME_SIZE + n_free_bufs - 1); + n_free_bufs += + vlib_buffer_alloc (vm, &nm->rx_buffers[cpu_index][n_free_bufs], + VLIB_FRAME_SIZE); + _vec_len (nm->rx_buffers[cpu_index]) = n_free_bufs; + } + + cur_ring = nif->first_rx_ring; + while (cur_ring <= nif->last_rx_ring && n_free_bufs) + { + int r = 0; + u32 cur_slot_index; + ring = NETMAP_RXRING (nif->nifp, cur_ring); + r = nm_ring_space (ring); + + if (!r) + { + cur_ring++; + continue; + } + + if (r > n_free_bufs) + r = n_free_bufs; + + cur_slot_index = ring->cur; + while (r) + { + u32 n_left_to_next; + u32 next0 = next_index; + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (r && n_left_to_next) + { + vlib_buffer_t *first_b0 = 0; + u32 offset = 0; + u32 bi0 = 0, first_bi0 = 0, prev_bi0; + u32 next_slot_index = (cur_slot_index + 1) % ring->num_slots; + u32 next2_slot_index = (cur_slot_index + 2) % ring->num_slots; + struct netmap_slot *slot = &ring->slot[cur_slot_index]; + u32 data_len = slot->len; + + /* prefetch 2 slots in advance */ + CLIB_PREFETCH (&ring->slot[next2_slot_index], + CLIB_CACHE_LINE_BYTES, LOAD); + /* prefetch start of next packet */ + CLIB_PREFETCH (NETMAP_BUF + (ring, ring->slot[next_slot_index].buf_idx), + CLIB_CACHE_LINE_BYTES, LOAD); + + while (data_len && n_free_bufs) + { + vlib_buffer_t *b0; + /* grab free buffer */ + u32 last_empty_buffer = + vec_len (nm->rx_buffers[cpu_index]) - 1; + prev_bi0 = bi0; + bi0 = nm->rx_buffers[cpu_index][last_empty_buffer]; + b0 = vlib_get_buffer (vm, bi0); + _vec_len (nm->rx_buffers[cpu_index]) = last_empty_buffer; + n_free_bufs--; + + /* copy data */ + u32 bytes_to_copy = + data_len > n_buffer_bytes ? n_buffer_bytes : data_len; + b0->current_data = 0; + clib_memcpy (vlib_buffer_get_current (b0), + (u8 *) NETMAP_BUF (ring, + slot->buf_idx) + offset, + bytes_to_copy); + + /* fill buffer header */ + b0->current_length = bytes_to_copy; + + if (offset == 0) + { + b0->total_length_not_including_first_buffer = 0; + b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID; + vnet_buffer (b0)->sw_if_index[VLIB_RX] = + nif->sw_if_index; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + first_bi0 = bi0; + first_b0 = vlib_get_buffer (vm, first_bi0); + } + else + buffer_add_to_chain (vm, bi0, first_bi0, prev_bi0); + + offset += bytes_to_copy; + data_len -= bytes_to_copy; + } + + /* trace */ + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (first_b0); + if (PREDICT_FALSE (n_trace > 0)) + { + if (PREDICT_TRUE (first_b0 != 0)) + { + netmap_input_trace_t *tr; + vlib_trace_buffer (vm, node, next0, first_b0, + /* follow_chain */ 0); + vlib_set_trace_count (vm, node, --n_trace); + tr = vlib_add_trace (vm, node, first_b0, sizeof (*tr)); + tr->next_index = next0; + tr->hw_if_index = nif->hw_if_index; + memcpy (&tr->slot, slot, sizeof (struct netmap_slot)); + } + } + + /* redirect if feature path enabled */ + vnet_feature_start_device_input_x1 (nif->sw_if_index, &next0, + first_b0, 0); + + /* enque and take next packet */ + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next, + n_left_to_next, first_bi0, + next0); + + /* next packet */ + n_rx_packets++; + n_rx_bytes += slot->len; + to_next[0] = first_bi0; + to_next += 1; + n_left_to_next--; + cur_slot_index = next_slot_index; + + r--; + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + ring->head = ring->cur = cur_slot_index; + cur_ring++; + } + + if (n_rx_packets) + ioctl (nif->fd, NIOCRXSYNC, NULL); + + vlib_increment_combined_counter + (vnet_get_main ()->interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, + os_get_cpu_number (), nif->hw_if_index, n_rx_packets, n_rx_bytes); + + return n_rx_packets; +} + +static uword +netmap_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node, + vlib_frame_t * frame) +{ + int i; + u32 n_rx_packets = 0; + u32 cpu_index = os_get_cpu_number (); + netmap_main_t *nm = &netmap_main; + netmap_if_t *nmi; + + for (i = 0; i < vec_len (nm->interfaces); i++) + { + nmi = vec_elt_at_index (nm->interfaces, i); + if (nmi->is_admin_up && + (i % nm->input_cpu_count) == + (cpu_index - nm->input_cpu_first_index)) + n_rx_packets += netmap_device_input_fn (vm, node, frame, nmi); + } + + return n_rx_packets; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (netmap_input_node) = { + .function = netmap_input_fn, + .name = "netmap-input", + .sibling_of = "device-input", + .format_trace = format_netmap_input_trace, + .type = VLIB_NODE_TYPE_INPUT, + /* default state is INTERRUPT mode, switch to POLLING if worker threads are enabled */ + .state = VLIB_NODE_STATE_INTERRUPT, + .n_errors = NETMAP_INPUT_N_ERROR, + .error_strings = netmap_input_error_strings, +}; + +VLIB_NODE_FUNCTION_MULTIARCH (netmap_input_node, netmap_input_fn) +/* *INDENT-ON* */ + + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/nic/ixge.c b/src/vnet/devices/nic/ixge.c new file mode 100644 index 00000000000..d4c4c6b7414 --- /dev/null +++ b/src/vnet/devices/nic/ixge.c @@ -0,0 +1,2938 @@ +/* + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* + * WARNING! + * This driver is not intended for production use and it is unsupported. + * It is provided for educational use only. + * Please use supported DPDK driver instead. + */ + +#if __x86_64__ +#include + +#ifndef CLIB_HAVE_VEC128 +#warning HACK: ixge driver wont really work, missing u32x4 +typedef unsigned long long u32x4; +#endif + +#include +#include +#include +#include +#include +#include + +#define IXGE_ALWAYS_POLL 0 + +#define EVENT_SET_FLAGS 0 +#define IXGE_HWBP_RACE_ELOG 0 + +#define PCI_VENDOR_ID_INTEL 0x8086 + +/* 10 GIG E (XGE) PHY IEEE 802.3 clause 45 definitions. */ +#define XGE_PHY_DEV_TYPE_PMA_PMD 1 +#define XGE_PHY_DEV_TYPE_PHY_XS 4 +#define XGE_PHY_ID1 0x2 +#define XGE_PHY_ID2 0x3 +#define XGE_PHY_CONTROL 0x0 +#define XGE_PHY_CONTROL_RESET (1 << 15) + +ixge_main_t ixge_main; +static vlib_node_registration_t ixge_input_node; +static vlib_node_registration_t ixge_process_node; + +static void +ixge_semaphore_get (ixge_device_t * xd) +{ + ixge_main_t *xm = &ixge_main; + vlib_main_t *vm = xm->vlib_main; + ixge_regs_t *r = xd->regs; + u32 i; + + i = 0; + while (!(r->software_semaphore & (1 << 0))) + { + if (i > 0) + vlib_process_suspend (vm, 100e-6); + i++; + } + do + { + r->software_semaphore |= 1 << 1; + } + while (!(r->software_semaphore & (1 << 1))); +} + +static void +ixge_semaphore_release (ixge_device_t * xd) +{ + ixge_regs_t *r = xd->regs; + r->software_semaphore &= ~3; +} + +static void +ixge_software_firmware_sync (ixge_device_t * xd, u32 sw_mask) +{ + ixge_main_t *xm = &ixge_main; + vlib_main_t *vm = xm->vlib_main; + ixge_regs_t *r = xd->regs; + u32 fw_mask = sw_mask << 5; + u32 m, done = 0; + + while (!done) + { + ixge_semaphore_get (xd); + m = r->software_firmware_sync; + done = (m & fw_mask) == 0; + if (done) + r->software_firmware_sync = m | sw_mask; + ixge_semaphore_release (xd); + if (!done) + vlib_process_suspend (vm, 10e-3); + } +} + +static void +ixge_software_firmware_sync_release (ixge_device_t * xd, u32 sw_mask) +{ + ixge_regs_t *r = xd->regs; + ixge_semaphore_get (xd); + r->software_firmware_sync &= ~sw_mask; + ixge_semaphore_release (xd); +} + +u32 +ixge_read_write_phy_reg (ixge_device_t * xd, u32 dev_type, u32 reg_index, + u32 v, u32 is_read) +{ + ixge_regs_t *r = xd->regs; + const u32 busy_bit = 1 << 30; + u32 x; + + ASSERT (xd->phy_index < 2); + ixge_software_firmware_sync (xd, 1 << (1 + xd->phy_index)); + + ASSERT (reg_index < (1 << 16)); + ASSERT (dev_type < (1 << 5)); + if (!is_read) + r->xge_mac.phy_data = v; + + /* Address cycle. */ + x = + reg_index | (dev_type << 16) | (xd-> + phys[xd->phy_index].mdio_address << 21); + r->xge_mac.phy_command = x | busy_bit; + /* Busy wait timed to take 28e-6 secs. No suspend. */ + while (r->xge_mac.phy_command & busy_bit) + ; + + r->xge_mac.phy_command = x | ((is_read ? 2 : 1) << 26) | busy_bit; + while (r->xge_mac.phy_command & busy_bit) + ; + + if (is_read) + v = r->xge_mac.phy_data >> 16; + + ixge_software_firmware_sync_release (xd, 1 << (1 + xd->phy_index)); + + return v; +} + +static u32 +ixge_read_phy_reg (ixge_device_t * xd, u32 dev_type, u32 reg_index) +{ + return ixge_read_write_phy_reg (xd, dev_type, reg_index, 0, /* is_read */ + 1); +} + +static void +ixge_write_phy_reg (ixge_device_t * xd, u32 dev_type, u32 reg_index, u32 v) +{ + (void) ixge_read_write_phy_reg (xd, dev_type, reg_index, v, /* is_read */ + 0); +} + +static void +ixge_i2c_put_bits (i2c_bus_t * b, int scl, int sda) +{ + ixge_main_t *xm = &ixge_main; + ixge_device_t *xd = vec_elt_at_index (xm->devices, b->private_data); + u32 v; + + v = 0; + v |= (sda != 0) << 3; + v |= (scl != 0) << 1; + xd->regs->i2c_control = v; +} + +static void +ixge_i2c_get_bits (i2c_bus_t * b, int *scl, int *sda) +{ + ixge_main_t *xm = &ixge_main; + ixge_device_t *xd = vec_elt_at_index (xm->devices, b->private_data); + u32 v; + + v = xd->regs->i2c_control; + *sda = (v & (1 << 2)) != 0; + *scl = (v & (1 << 0)) != 0; +} + +static u16 +ixge_read_eeprom (ixge_device_t * xd, u32 address) +{ + ixge_regs_t *r = xd->regs; + u32 v; + r->eeprom_read = (( /* start bit */ (1 << 0)) | (address << 2)); + /* Wait for done bit. */ + while (!((v = r->eeprom_read) & (1 << 1))) + ; + return v >> 16; +} + +static void +ixge_sfp_enable_disable_laser (ixge_device_t * xd, uword enable) +{ + u32 tx_disable_bit = 1 << 3; + if (enable) + xd->regs->sdp_control &= ~tx_disable_bit; + else + xd->regs->sdp_control |= tx_disable_bit; +} + +static void +ixge_sfp_enable_disable_10g (ixge_device_t * xd, uword enable) +{ + u32 is_10g_bit = 1 << 5; + if (enable) + xd->regs->sdp_control |= is_10g_bit; + else + xd->regs->sdp_control &= ~is_10g_bit; +} + +static clib_error_t * +ixge_sfp_phy_init_from_eeprom (ixge_device_t * xd, u16 sfp_type) +{ + u16 a, id, reg_values_addr = 0; + + a = ixge_read_eeprom (xd, 0x2b); + if (a == 0 || a == 0xffff) + return clib_error_create ("no init sequence in eeprom"); + + while (1) + { + id = ixge_read_eeprom (xd, ++a); + if (id == 0xffff) + break; + reg_values_addr = ixge_read_eeprom (xd, ++a); + if (id == sfp_type) + break; + } + if (id != sfp_type) + return clib_error_create ("failed to find id 0x%x", sfp_type); + + ixge_software_firmware_sync (xd, 1 << 3); + while (1) + { + u16 v = ixge_read_eeprom (xd, ++reg_values_addr); + if (v == 0xffff) + break; + xd->regs->core_analog_config = v; + } + ixge_software_firmware_sync_release (xd, 1 << 3); + + /* Make sure laser is off. We'll turn on the laser when + the interface is brought up. */ + ixge_sfp_enable_disable_laser (xd, /* enable */ 0); + ixge_sfp_enable_disable_10g (xd, /* is_10g */ 1); + + return 0; +} + +static void +ixge_sfp_device_up_down (ixge_device_t * xd, uword is_up) +{ + u32 v; + + if (is_up) + { + /* pma/pmd 10g serial SFI. */ + xd->regs->xge_mac.auto_negotiation_control2 &= ~(3 << 16); + xd->regs->xge_mac.auto_negotiation_control2 |= 2 << 16; + + v = xd->regs->xge_mac.auto_negotiation_control; + v &= ~(7 << 13); + v |= (0 << 13); + /* Restart autoneg. */ + v |= (1 << 12); + xd->regs->xge_mac.auto_negotiation_control = v; + + while (!(xd->regs->xge_mac.link_partner_ability[0] & 0xf0000)) + ; + + v = xd->regs->xge_mac.auto_negotiation_control; + + /* link mode 10g sfi serdes */ + v &= ~(7 << 13); + v |= (3 << 13); + + /* Restart autoneg. */ + v |= (1 << 12); + xd->regs->xge_mac.auto_negotiation_control = v; + + xd->regs->xge_mac.link_status; + } + + ixge_sfp_enable_disable_laser (xd, /* enable */ is_up); + + /* Give time for link partner to notice that we're up. */ + if (is_up && vlib_in_process_context (vlib_get_main ())) + { + vlib_process_suspend (vlib_get_main (), 300e-3); + } +} + +always_inline ixge_dma_regs_t * +get_dma_regs (ixge_device_t * xd, vlib_rx_or_tx_t rt, u32 qi) +{ + ixge_regs_t *r = xd->regs; + ASSERT (qi < 128); + if (rt == VLIB_RX) + return qi < 64 ? &r->rx_dma0[qi] : &r->rx_dma1[qi - 64]; + else + return &r->tx_dma[qi]; +} + +static clib_error_t * +ixge_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, u32 flags) +{ + vnet_hw_interface_t *hif = vnet_get_hw_interface (vnm, hw_if_index); + uword is_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0; + ixge_main_t *xm = &ixge_main; + ixge_device_t *xd = vec_elt_at_index (xm->devices, hif->dev_instance); + ixge_dma_regs_t *dr = get_dma_regs (xd, VLIB_RX, 0); + + if (is_up) + { + xd->regs->rx_enable |= 1; + xd->regs->tx_dma_control |= 1; + dr->control |= 1 << 25; + while (!(dr->control & (1 << 25))) + ; + } + else + { + xd->regs->rx_enable &= ~1; + xd->regs->tx_dma_control &= ~1; + } + + ixge_sfp_device_up_down (xd, is_up); + + return /* no error */ 0; +} + +static void +ixge_sfp_phy_init (ixge_device_t * xd) +{ + ixge_phy_t *phy = xd->phys + xd->phy_index; + i2c_bus_t *ib = &xd->i2c_bus; + + ib->private_data = xd->device_index; + ib->put_bits = ixge_i2c_put_bits; + ib->get_bits = ixge_i2c_get_bits; + vlib_i2c_init (ib); + + vlib_i2c_read_eeprom (ib, 0x50, 0, 128, (u8 *) & xd->sfp_eeprom); + + if (vlib_i2c_bus_timed_out (ib) || !sfp_eeprom_is_valid (&xd->sfp_eeprom)) + xd->sfp_eeprom.id = SFP_ID_unknown; + else + { + /* FIXME 5 => SR/LR eeprom ID. */ + clib_error_t *e = + ixge_sfp_phy_init_from_eeprom (xd, 5 + xd->pci_function); + if (e) + clib_error_report (e); + } + + phy->mdio_address = ~0; +} + +static void +ixge_phy_init (ixge_device_t * xd) +{ + ixge_main_t *xm = &ixge_main; + vlib_main_t *vm = xm->vlib_main; + ixge_phy_t *phy = xd->phys + xd->phy_index; + + switch (xd->device_id) + { + case IXGE_82599_sfp: + case IXGE_82599_sfp_em: + case IXGE_82599_sfp_fcoe: + /* others? */ + return ixge_sfp_phy_init (xd); + + default: + break; + } + + /* Probe address of phy. */ + { + u32 i, v; + + phy->mdio_address = ~0; + for (i = 0; i < 32; i++) + { + phy->mdio_address = i; + v = ixge_read_phy_reg (xd, XGE_PHY_DEV_TYPE_PMA_PMD, XGE_PHY_ID1); + if (v != 0xffff && v != 0) + break; + } + + /* No PHY found? */ + if (i >= 32) + return; + } + + phy->id = + ((ixge_read_phy_reg (xd, XGE_PHY_DEV_TYPE_PMA_PMD, XGE_PHY_ID1) << 16) | + ixge_read_phy_reg (xd, XGE_PHY_DEV_TYPE_PMA_PMD, XGE_PHY_ID2)); + + { + ELOG_TYPE_DECLARE (e) = + { + .function = (char *) __FUNCTION__,.format = + "ixge %d, phy id 0x%d mdio address %d",.format_args = "i4i4i4",}; + struct + { + u32 instance, id, address; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->instance = xd->device_index; + ed->id = phy->id; + ed->address = phy->mdio_address; + } + + /* Reset phy. */ + ixge_write_phy_reg (xd, XGE_PHY_DEV_TYPE_PHY_XS, XGE_PHY_CONTROL, + XGE_PHY_CONTROL_RESET); + + /* Wait for self-clearning reset bit to clear. */ + do + { + vlib_process_suspend (vm, 1e-3); + } + while (ixge_read_phy_reg (xd, XGE_PHY_DEV_TYPE_PHY_XS, XGE_PHY_CONTROL) & + XGE_PHY_CONTROL_RESET); +} + +static u8 * +format_ixge_rx_from_hw_descriptor (u8 * s, va_list * va) +{ + ixge_rx_from_hw_descriptor_t *d = + va_arg (*va, ixge_rx_from_hw_descriptor_t *); + u32 s0 = d->status[0], s2 = d->status[2]; + u32 is_ip4, is_ip6, is_ip, is_tcp, is_udp; + uword indent = format_get_indent (s); + + s = format (s, "%s-owned", + (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_OWNED_BY_SOFTWARE) ? "sw" : + "hw"); + s = + format (s, ", length this descriptor %d, l3 offset %d", + d->n_packet_bytes_this_descriptor, + IXGE_RX_DESCRIPTOR_STATUS0_L3_OFFSET (s0)); + if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_END_OF_PACKET) + s = format (s, ", end-of-packet"); + + s = format (s, "\n%U", format_white_space, indent); + + if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_ETHERNET_ERROR) + s = format (s, "layer2 error"); + + if (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_LAYER2) + { + s = format (s, "layer 2 type %d", (s0 & 0x1f)); + return s; + } + + if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_VLAN) + s = format (s, "vlan header 0x%x\n%U", d->vlan_tag, + format_white_space, indent); + + if ((is_ip4 = (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP4))) + { + s = format (s, "ip4%s", + (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP4_EXT) ? " options" : + ""); + if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_IP4_CHECKSUMMED) + s = format (s, " checksum %s", + (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IP4_CHECKSUM_ERROR) ? + "bad" : "ok"); + } + if ((is_ip6 = (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6))) + s = format (s, "ip6%s", + (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6_EXT) ? " extended" : + ""); + is_tcp = is_udp = 0; + if ((is_ip = (is_ip4 | is_ip6))) + { + is_tcp = (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_TCP) != 0; + is_udp = (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_UDP) != 0; + if (is_tcp) + s = format (s, ", tcp"); + if (is_udp) + s = format (s, ", udp"); + } + + if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_TCP_CHECKSUMMED) + s = format (s, ", tcp checksum %s", + (s2 & IXGE_RX_DESCRIPTOR_STATUS2_TCP_CHECKSUM_ERROR) ? "bad" : + "ok"); + if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_UDP_CHECKSUMMED) + s = format (s, ", udp checksum %s", + (s2 & IXGE_RX_DESCRIPTOR_STATUS2_UDP_CHECKSUM_ERROR) ? "bad" : + "ok"); + + return s; +} + +static u8 * +format_ixge_tx_descriptor (u8 * s, va_list * va) +{ + ixge_tx_descriptor_t *d = va_arg (*va, ixge_tx_descriptor_t *); + u32 s0 = d->status0, s1 = d->status1; + uword indent = format_get_indent (s); + u32 v; + + s = format (s, "buffer 0x%Lx, %d packet bytes, %d bytes this buffer", + d->buffer_address, s1 >> 14, d->n_bytes_this_buffer); + + s = format (s, "\n%U", format_white_space, indent); + + if ((v = (s0 >> 0) & 3)) + s = format (s, "reserved 0x%x, ", v); + + if ((v = (s0 >> 2) & 3)) + s = format (s, "mac 0x%x, ", v); + + if ((v = (s0 >> 4) & 0xf) != 3) + s = format (s, "type 0x%x, ", v); + + s = format (s, "%s%s%s%s%s%s%s%s", + (s0 & (1 << 8)) ? "eop, " : "", + (s0 & (1 << 9)) ? "insert-fcs, " : "", + (s0 & (1 << 10)) ? "reserved26, " : "", + (s0 & (1 << 11)) ? "report-status, " : "", + (s0 & (1 << 12)) ? "reserved28, " : "", + (s0 & (1 << 13)) ? "is-advanced, " : "", + (s0 & (1 << 14)) ? "vlan-enable, " : "", + (s0 & (1 << 15)) ? "tx-segmentation, " : ""); + + if ((v = s1 & 0xf) != 0) + s = format (s, "status 0x%x, ", v); + + if ((v = (s1 >> 4) & 0xf)) + s = format (s, "context 0x%x, ", v); + + if ((v = (s1 >> 8) & 0x3f)) + s = format (s, "options 0x%x, ", v); + + return s; +} + +typedef struct +{ + ixge_descriptor_t before, after; + + u32 buffer_index; + + u16 device_index; + + u8 queue_index; + + u8 is_start_of_packet; + + /* Copy of VLIB buffer; packet data stored in pre_data. */ + vlib_buffer_t buffer; +} ixge_rx_dma_trace_t; + +static u8 * +format_ixge_rx_dma_trace (u8 * s, va_list * va) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *); + vlib_node_t *node = va_arg (*va, vlib_node_t *); + vnet_main_t *vnm = vnet_get_main (); + ixge_rx_dma_trace_t *t = va_arg (*va, ixge_rx_dma_trace_t *); + ixge_main_t *xm = &ixge_main; + ixge_device_t *xd = vec_elt_at_index (xm->devices, t->device_index); + format_function_t *f; + uword indent = format_get_indent (s); + + { + vnet_sw_interface_t *sw = + vnet_get_sw_interface (vnm, xd->vlib_sw_if_index); + s = + format (s, "%U rx queue %d", format_vnet_sw_interface_name, vnm, sw, + t->queue_index); + } + + s = format (s, "\n%Ubefore: %U", + format_white_space, indent, + format_ixge_rx_from_hw_descriptor, &t->before); + s = format (s, "\n%Uafter : head/tail address 0x%Lx/0x%Lx", + format_white_space, indent, + t->after.rx_to_hw.head_address, t->after.rx_to_hw.tail_address); + + s = format (s, "\n%Ubuffer 0x%x: %U", + format_white_space, indent, + t->buffer_index, format_vlib_buffer, &t->buffer); + + s = format (s, "\n%U", format_white_space, indent); + + f = node->format_buffer; + if (!f || !t->is_start_of_packet) + f = format_hex_bytes; + s = format (s, "%U", f, t->buffer.pre_data, sizeof (t->buffer.pre_data)); + + return s; +} + +#define foreach_ixge_error \ + _ (none, "no error") \ + _ (tx_full_drops, "tx ring full drops") \ + _ (ip4_checksum_error, "ip4 checksum errors") \ + _ (rx_alloc_fail, "rx buf alloc from free list failed") \ + _ (rx_alloc_no_physmem, "rx buf alloc failed no physmem") + +typedef enum +{ +#define _(f,s) IXGE_ERROR_##f, + foreach_ixge_error +#undef _ + IXGE_N_ERROR, +} ixge_error_t; + +always_inline void +ixge_rx_next_and_error_from_status_x1 (ixge_device_t * xd, + u32 s00, u32 s02, + u8 * next0, u8 * error0, u32 * flags0) +{ + u8 is0_ip4, is0_ip6, n0, e0; + u32 f0; + + e0 = IXGE_ERROR_none; + n0 = IXGE_RX_NEXT_ETHERNET_INPUT; + + is0_ip4 = s02 & IXGE_RX_DESCRIPTOR_STATUS2_IS_IP4_CHECKSUMMED; + n0 = is0_ip4 ? IXGE_RX_NEXT_IP4_INPUT : n0; + + e0 = (is0_ip4 && (s02 & IXGE_RX_DESCRIPTOR_STATUS2_IP4_CHECKSUM_ERROR) + ? IXGE_ERROR_ip4_checksum_error : e0); + + is0_ip6 = s00 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6; + n0 = is0_ip6 ? IXGE_RX_NEXT_IP6_INPUT : n0; + + n0 = (xd->per_interface_next_index != ~0) ? + xd->per_interface_next_index : n0; + + /* Check for error. */ + n0 = e0 != IXGE_ERROR_none ? IXGE_RX_NEXT_DROP : n0; + + f0 = ((s02 & (IXGE_RX_DESCRIPTOR_STATUS2_IS_TCP_CHECKSUMMED + | IXGE_RX_DESCRIPTOR_STATUS2_IS_UDP_CHECKSUMMED)) + ? IP_BUFFER_L4_CHECKSUM_COMPUTED : 0); + + f0 |= ((s02 & (IXGE_RX_DESCRIPTOR_STATUS2_TCP_CHECKSUM_ERROR + | IXGE_RX_DESCRIPTOR_STATUS2_UDP_CHECKSUM_ERROR)) + ? 0 : IP_BUFFER_L4_CHECKSUM_CORRECT); + + *error0 = e0; + *next0 = n0; + *flags0 = f0; +} + +always_inline void +ixge_rx_next_and_error_from_status_x2 (ixge_device_t * xd, + u32 s00, u32 s02, + u32 s10, u32 s12, + u8 * next0, u8 * error0, u32 * flags0, + u8 * next1, u8 * error1, u32 * flags1) +{ + u8 is0_ip4, is0_ip6, n0, e0; + u8 is1_ip4, is1_ip6, n1, e1; + u32 f0, f1; + + e0 = e1 = IXGE_ERROR_none; + n0 = n1 = IXGE_RX_NEXT_IP4_INPUT; + + is0_ip4 = s02 & IXGE_RX_DESCRIPTOR_STATUS2_IS_IP4_CHECKSUMMED; + is1_ip4 = s12 & IXGE_RX_DESCRIPTOR_STATUS2_IS_IP4_CHECKSUMMED; + + n0 = is0_ip4 ? IXGE_RX_NEXT_IP4_INPUT : n0; + n1 = is1_ip4 ? IXGE_RX_NEXT_IP4_INPUT : n1; + + e0 = (is0_ip4 && (s02 & IXGE_RX_DESCRIPTOR_STATUS2_IP4_CHECKSUM_ERROR) + ? IXGE_ERROR_ip4_checksum_error : e0); + e1 = (is1_ip4 && (s12 & IXGE_RX_DESCRIPTOR_STATUS2_IP4_CHECKSUM_ERROR) + ? IXGE_ERROR_ip4_checksum_error : e1); + + is0_ip6 = s00 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6; + is1_ip6 = s10 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6; + + n0 = is0_ip6 ? IXGE_RX_NEXT_IP6_INPUT : n0; + n1 = is1_ip6 ? IXGE_RX_NEXT_IP6_INPUT : n1; + + n0 = (xd->per_interface_next_index != ~0) ? + xd->per_interface_next_index : n0; + n1 = (xd->per_interface_next_index != ~0) ? + xd->per_interface_next_index : n1; + + /* Check for error. */ + n0 = e0 != IXGE_ERROR_none ? IXGE_RX_NEXT_DROP : n0; + n1 = e1 != IXGE_ERROR_none ? IXGE_RX_NEXT_DROP : n1; + + *error0 = e0; + *error1 = e1; + + *next0 = n0; + *next1 = n1; + + f0 = ((s02 & (IXGE_RX_DESCRIPTOR_STATUS2_IS_TCP_CHECKSUMMED + | IXGE_RX_DESCRIPTOR_STATUS2_IS_UDP_CHECKSUMMED)) + ? IP_BUFFER_L4_CHECKSUM_COMPUTED : 0); + f1 = ((s12 & (IXGE_RX_DESCRIPTOR_STATUS2_IS_TCP_CHECKSUMMED + | IXGE_RX_DESCRIPTOR_STATUS2_IS_UDP_CHECKSUMMED)) + ? IP_BUFFER_L4_CHECKSUM_COMPUTED : 0); + + f0 |= ((s02 & (IXGE_RX_DESCRIPTOR_STATUS2_TCP_CHECKSUM_ERROR + | IXGE_RX_DESCRIPTOR_STATUS2_UDP_CHECKSUM_ERROR)) + ? 0 : IP_BUFFER_L4_CHECKSUM_CORRECT); + f1 |= ((s12 & (IXGE_RX_DESCRIPTOR_STATUS2_TCP_CHECKSUM_ERROR + | IXGE_RX_DESCRIPTOR_STATUS2_UDP_CHECKSUM_ERROR)) + ? 0 : IP_BUFFER_L4_CHECKSUM_CORRECT); + + *flags0 = f0; + *flags1 = f1; +} + +static void +ixge_rx_trace (ixge_main_t * xm, + ixge_device_t * xd, + ixge_dma_queue_t * dq, + ixge_descriptor_t * before_descriptors, + u32 * before_buffers, + ixge_descriptor_t * after_descriptors, uword n_descriptors) +{ + vlib_main_t *vm = xm->vlib_main; + vlib_node_runtime_t *node = dq->rx.node; + ixge_rx_from_hw_descriptor_t *bd; + ixge_rx_to_hw_descriptor_t *ad; + u32 *b, n_left, is_sop, next_index_sop; + + n_left = n_descriptors; + b = before_buffers; + bd = &before_descriptors->rx_from_hw; + ad = &after_descriptors->rx_to_hw; + is_sop = dq->rx.is_start_of_packet; + next_index_sop = dq->rx.saved_start_of_packet_next_index; + + while (n_left >= 2) + { + u32 bi0, bi1, flags0, flags1; + vlib_buffer_t *b0, *b1; + ixge_rx_dma_trace_t *t0, *t1; + u8 next0, error0, next1, error1; + + bi0 = b[0]; + bi1 = b[1]; + n_left -= 2; + + b0 = vlib_get_buffer (vm, bi0); + b1 = vlib_get_buffer (vm, bi1); + + ixge_rx_next_and_error_from_status_x2 (xd, + bd[0].status[0], bd[0].status[2], + bd[1].status[0], bd[1].status[2], + &next0, &error0, &flags0, + &next1, &error1, &flags1); + + next_index_sop = is_sop ? next0 : next_index_sop; + vlib_trace_buffer (vm, node, next_index_sop, b0, /* follow_chain */ 0); + t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0])); + t0->is_start_of_packet = is_sop; + is_sop = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0; + + next_index_sop = is_sop ? next1 : next_index_sop; + vlib_trace_buffer (vm, node, next_index_sop, b1, /* follow_chain */ 0); + t1 = vlib_add_trace (vm, node, b1, sizeof (t1[0])); + t1->is_start_of_packet = is_sop; + is_sop = (b1->flags & VLIB_BUFFER_NEXT_PRESENT) == 0; + + t0->queue_index = dq->queue_index; + t1->queue_index = dq->queue_index; + t0->device_index = xd->device_index; + t1->device_index = xd->device_index; + t0->before.rx_from_hw = bd[0]; + t1->before.rx_from_hw = bd[1]; + t0->after.rx_to_hw = ad[0]; + t1->after.rx_to_hw = ad[1]; + t0->buffer_index = bi0; + t1->buffer_index = bi1; + memcpy (&t0->buffer, b0, sizeof (b0[0]) - sizeof (b0->pre_data)); + memcpy (&t1->buffer, b1, sizeof (b1[0]) - sizeof (b0->pre_data)); + memcpy (t0->buffer.pre_data, b0->data + b0->current_data, + sizeof (t0->buffer.pre_data)); + memcpy (t1->buffer.pre_data, b1->data + b1->current_data, + sizeof (t1->buffer.pre_data)); + + b += 2; + bd += 2; + ad += 2; + } + + while (n_left >= 1) + { + u32 bi0, flags0; + vlib_buffer_t *b0; + ixge_rx_dma_trace_t *t0; + u8 next0, error0; + + bi0 = b[0]; + n_left -= 1; + + b0 = vlib_get_buffer (vm, bi0); + + ixge_rx_next_and_error_from_status_x1 (xd, + bd[0].status[0], bd[0].status[2], + &next0, &error0, &flags0); + + next_index_sop = is_sop ? next0 : next_index_sop; + vlib_trace_buffer (vm, node, next_index_sop, b0, /* follow_chain */ 0); + t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0])); + t0->is_start_of_packet = is_sop; + is_sop = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0; + + t0->queue_index = dq->queue_index; + t0->device_index = xd->device_index; + t0->before.rx_from_hw = bd[0]; + t0->after.rx_to_hw = ad[0]; + t0->buffer_index = bi0; + memcpy (&t0->buffer, b0, sizeof (b0[0]) - sizeof (b0->pre_data)); + memcpy (t0->buffer.pre_data, b0->data + b0->current_data, + sizeof (t0->buffer.pre_data)); + + b += 1; + bd += 1; + ad += 1; + } +} + +typedef struct +{ + ixge_tx_descriptor_t descriptor; + + u32 buffer_index; + + u16 device_index; + + u8 queue_index; + + u8 is_start_of_packet; + + /* Copy of VLIB buffer; packet data stored in pre_data. */ + vlib_buffer_t buffer; +} ixge_tx_dma_trace_t; + +static u8 * +format_ixge_tx_dma_trace (u8 * s, va_list * va) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*va, vlib_node_t *); + ixge_tx_dma_trace_t *t = va_arg (*va, ixge_tx_dma_trace_t *); + vnet_main_t *vnm = vnet_get_main (); + ixge_main_t *xm = &ixge_main; + ixge_device_t *xd = vec_elt_at_index (xm->devices, t->device_index); + format_function_t *f; + uword indent = format_get_indent (s); + + { + vnet_sw_interface_t *sw = + vnet_get_sw_interface (vnm, xd->vlib_sw_if_index); + s = + format (s, "%U tx queue %d", format_vnet_sw_interface_name, vnm, sw, + t->queue_index); + } + + s = format (s, "\n%Udescriptor: %U", + format_white_space, indent, + format_ixge_tx_descriptor, &t->descriptor); + + s = format (s, "\n%Ubuffer 0x%x: %U", + format_white_space, indent, + t->buffer_index, format_vlib_buffer, &t->buffer); + + s = format (s, "\n%U", format_white_space, indent); + + f = format_ethernet_header_with_length; + if (!f || !t->is_start_of_packet) + f = format_hex_bytes; + s = format (s, "%U", f, t->buffer.pre_data, sizeof (t->buffer.pre_data)); + + return s; +} + +typedef struct +{ + vlib_node_runtime_t *node; + + u32 is_start_of_packet; + + u32 n_bytes_in_packet; + + ixge_tx_descriptor_t *start_of_packet_descriptor; +} ixge_tx_state_t; + +static void +ixge_tx_trace (ixge_main_t * xm, + ixge_device_t * xd, + ixge_dma_queue_t * dq, + ixge_tx_state_t * tx_state, + ixge_tx_descriptor_t * descriptors, + u32 * buffers, uword n_descriptors) +{ + vlib_main_t *vm = xm->vlib_main; + vlib_node_runtime_t *node = tx_state->node; + ixge_tx_descriptor_t *d; + u32 *b, n_left, is_sop; + + n_left = n_descriptors; + b = buffers; + d = descriptors; + is_sop = tx_state->is_start_of_packet; + + while (n_left >= 2) + { + u32 bi0, bi1; + vlib_buffer_t *b0, *b1; + ixge_tx_dma_trace_t *t0, *t1; + + bi0 = b[0]; + bi1 = b[1]; + n_left -= 2; + + b0 = vlib_get_buffer (vm, bi0); + b1 = vlib_get_buffer (vm, bi1); + + t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0])); + t0->is_start_of_packet = is_sop; + is_sop = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0; + + t1 = vlib_add_trace (vm, node, b1, sizeof (t1[0])); + t1->is_start_of_packet = is_sop; + is_sop = (b1->flags & VLIB_BUFFER_NEXT_PRESENT) == 0; + + t0->queue_index = dq->queue_index; + t1->queue_index = dq->queue_index; + t0->device_index = xd->device_index; + t1->device_index = xd->device_index; + t0->descriptor = d[0]; + t1->descriptor = d[1]; + t0->buffer_index = bi0; + t1->buffer_index = bi1; + memcpy (&t0->buffer, b0, sizeof (b0[0]) - sizeof (b0->pre_data)); + memcpy (&t1->buffer, b1, sizeof (b1[0]) - sizeof (b0->pre_data)); + memcpy (t0->buffer.pre_data, b0->data + b0->current_data, + sizeof (t0->buffer.pre_data)); + memcpy (t1->buffer.pre_data, b1->data + b1->current_data, + sizeof (t1->buffer.pre_data)); + + b += 2; + d += 2; + } + + while (n_left >= 1) + { + u32 bi0; + vlib_buffer_t *b0; + ixge_tx_dma_trace_t *t0; + + bi0 = b[0]; + n_left -= 1; + + b0 = vlib_get_buffer (vm, bi0); + + t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0])); + t0->is_start_of_packet = is_sop; + is_sop = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0; + + t0->queue_index = dq->queue_index; + t0->device_index = xd->device_index; + t0->descriptor = d[0]; + t0->buffer_index = bi0; + memcpy (&t0->buffer, b0, sizeof (b0[0]) - sizeof (b0->pre_data)); + memcpy (t0->buffer.pre_data, b0->data + b0->current_data, + sizeof (t0->buffer.pre_data)); + + b += 1; + d += 1; + } +} + +always_inline uword +ixge_ring_sub (ixge_dma_queue_t * q, u32 i0, u32 i1) +{ + i32 d = i1 - i0; + ASSERT (i0 < q->n_descriptors); + ASSERT (i1 < q->n_descriptors); + return d < 0 ? q->n_descriptors + d : d; +} + +always_inline uword +ixge_ring_add (ixge_dma_queue_t * q, u32 i0, u32 i1) +{ + u32 d = i0 + i1; + ASSERT (i0 < q->n_descriptors); + ASSERT (i1 < q->n_descriptors); + d -= d >= q->n_descriptors ? q->n_descriptors : 0; + return d; +} + +always_inline uword +ixge_tx_descriptor_matches_template (ixge_main_t * xm, + ixge_tx_descriptor_t * d) +{ + u32 cmp; + + cmp = ((d->status0 & xm->tx_descriptor_template_mask.status0) + ^ xm->tx_descriptor_template.status0); + if (cmp) + return 0; + cmp = ((d->status1 & xm->tx_descriptor_template_mask.status1) + ^ xm->tx_descriptor_template.status1); + if (cmp) + return 0; + + return 1; +} + +static uword +ixge_tx_no_wrap (ixge_main_t * xm, + ixge_device_t * xd, + ixge_dma_queue_t * dq, + u32 * buffers, + u32 start_descriptor_index, + u32 n_descriptors, ixge_tx_state_t * tx_state) +{ + vlib_main_t *vm = xm->vlib_main; + ixge_tx_descriptor_t *d, *d_sop; + u32 n_left = n_descriptors; + u32 *to_free = vec_end (xm->tx_buffers_pending_free); + u32 *to_tx = + vec_elt_at_index (dq->descriptor_buffer_indices, start_descriptor_index); + u32 is_sop = tx_state->is_start_of_packet; + u32 len_sop = tx_state->n_bytes_in_packet; + u16 template_status = xm->tx_descriptor_template.status0; + u32 descriptor_prefetch_rotor = 0; + + ASSERT (start_descriptor_index + n_descriptors <= dq->n_descriptors); + d = &dq->descriptors[start_descriptor_index].tx; + d_sop = is_sop ? d : tx_state->start_of_packet_descriptor; + + while (n_left >= 4) + { + vlib_buffer_t *b0, *b1; + u32 bi0, fi0, len0; + u32 bi1, fi1, len1; + u8 is_eop0, is_eop1; + + /* Prefetch next iteration. */ + vlib_prefetch_buffer_with_index (vm, buffers[2], LOAD); + vlib_prefetch_buffer_with_index (vm, buffers[3], LOAD); + + if ((descriptor_prefetch_rotor & 0x3) == 0) + CLIB_PREFETCH (d + 4, CLIB_CACHE_LINE_BYTES, STORE); + + descriptor_prefetch_rotor += 2; + + bi0 = buffers[0]; + bi1 = buffers[1]; + + to_free[0] = fi0 = to_tx[0]; + to_tx[0] = bi0; + to_free += fi0 != 0; + + to_free[0] = fi1 = to_tx[1]; + to_tx[1] = bi1; + to_free += fi1 != 0; + + buffers += 2; + n_left -= 2; + to_tx += 2; + + b0 = vlib_get_buffer (vm, bi0); + b1 = vlib_get_buffer (vm, bi1); + + is_eop0 = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0; + is_eop1 = (b1->flags & VLIB_BUFFER_NEXT_PRESENT) == 0; + + len0 = b0->current_length; + len1 = b1->current_length; + + ASSERT (ixge_tx_descriptor_matches_template (xm, d + 0)); + ASSERT (ixge_tx_descriptor_matches_template (xm, d + 1)); + + d[0].buffer_address = + vlib_get_buffer_data_physical_address (vm, bi0) + b0->current_data; + d[1].buffer_address = + vlib_get_buffer_data_physical_address (vm, bi1) + b1->current_data; + + d[0].n_bytes_this_buffer = len0; + d[1].n_bytes_this_buffer = len1; + + d[0].status0 = + template_status | (is_eop0 << + IXGE_TX_DESCRIPTOR_STATUS0_LOG2_IS_END_OF_PACKET); + d[1].status0 = + template_status | (is_eop1 << + IXGE_TX_DESCRIPTOR_STATUS0_LOG2_IS_END_OF_PACKET); + + len_sop = (is_sop ? 0 : len_sop) + len0; + d_sop[0].status1 = + IXGE_TX_DESCRIPTOR_STATUS1_N_BYTES_IN_PACKET (len_sop); + d += 1; + d_sop = is_eop0 ? d : d_sop; + + is_sop = is_eop0; + + len_sop = (is_sop ? 0 : len_sop) + len1; + d_sop[0].status1 = + IXGE_TX_DESCRIPTOR_STATUS1_N_BYTES_IN_PACKET (len_sop); + d += 1; + d_sop = is_eop1 ? d : d_sop; + + is_sop = is_eop1; + } + + while (n_left > 0) + { + vlib_buffer_t *b0; + u32 bi0, fi0, len0; + u8 is_eop0; + + bi0 = buffers[0]; + + to_free[0] = fi0 = to_tx[0]; + to_tx[0] = bi0; + to_free += fi0 != 0; + + buffers += 1; + n_left -= 1; + to_tx += 1; + + b0 = vlib_get_buffer (vm, bi0); + + is_eop0 = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0; + + len0 = b0->current_length; + + ASSERT (ixge_tx_descriptor_matches_template (xm, d + 0)); + + d[0].buffer_address = + vlib_get_buffer_data_physical_address (vm, bi0) + b0->current_data; + + d[0].n_bytes_this_buffer = len0; + + d[0].status0 = + template_status | (is_eop0 << + IXGE_TX_DESCRIPTOR_STATUS0_LOG2_IS_END_OF_PACKET); + + len_sop = (is_sop ? 0 : len_sop) + len0; + d_sop[0].status1 = + IXGE_TX_DESCRIPTOR_STATUS1_N_BYTES_IN_PACKET (len_sop); + d += 1; + d_sop = is_eop0 ? d : d_sop; + + is_sop = is_eop0; + } + + if (tx_state->node->flags & VLIB_NODE_FLAG_TRACE) + { + to_tx = + vec_elt_at_index (dq->descriptor_buffer_indices, + start_descriptor_index); + ixge_tx_trace (xm, xd, dq, tx_state, + &dq->descriptors[start_descriptor_index].tx, to_tx, + n_descriptors); + } + + _vec_len (xm->tx_buffers_pending_free) = + to_free - xm->tx_buffers_pending_free; + + /* When we are done d_sop can point to end of ring. Wrap it if so. */ + { + ixge_tx_descriptor_t *d_start = &dq->descriptors[0].tx; + + ASSERT (d_sop - d_start <= dq->n_descriptors); + d_sop = d_sop - d_start == dq->n_descriptors ? d_start : d_sop; + } + + tx_state->is_start_of_packet = is_sop; + tx_state->start_of_packet_descriptor = d_sop; + tx_state->n_bytes_in_packet = len_sop; + + return n_descriptors; +} + +static uword +ixge_interface_tx (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * f) +{ + ixge_main_t *xm = &ixge_main; + vnet_interface_output_runtime_t *rd = (void *) node->runtime_data; + ixge_device_t *xd = vec_elt_at_index (xm->devices, rd->dev_instance); + ixge_dma_queue_t *dq; + u32 *from, n_left_tx, n_descriptors_to_tx, n_tail_drop; + u32 queue_index = 0; /* fixme parameter */ + ixge_tx_state_t tx_state; + + tx_state.node = node; + tx_state.is_start_of_packet = 1; + tx_state.start_of_packet_descriptor = 0; + tx_state.n_bytes_in_packet = 0; + + from = vlib_frame_vector_args (f); + + dq = vec_elt_at_index (xd->dma_queues[VLIB_TX], queue_index); + + dq->head_index = dq->tx.head_index_write_back[0]; + + /* Since head == tail means ring is empty we can send up to dq->n_descriptors - 1. */ + n_left_tx = dq->n_descriptors - 1; + n_left_tx -= ixge_ring_sub (dq, dq->head_index, dq->tail_index); + + _vec_len (xm->tx_buffers_pending_free) = 0; + + n_descriptors_to_tx = f->n_vectors; + n_tail_drop = 0; + if (PREDICT_FALSE (n_descriptors_to_tx > n_left_tx)) + { + i32 i, n_ok, i_eop, i_sop; + + i_sop = i_eop = ~0; + for (i = n_left_tx - 1; i >= 0; i--) + { + vlib_buffer_t *b = vlib_get_buffer (vm, from[i]); + if (!(b->flags & VLIB_BUFFER_NEXT_PRESENT)) + { + if (i_sop != ~0 && i_eop != ~0) + break; + i_eop = i; + i_sop = i + 1; + } + } + if (i == 0) + n_ok = 0; + else + n_ok = i_eop + 1; + + { + ELOG_TYPE_DECLARE (e) = + { + .function = (char *) __FUNCTION__,.format = + "ixge %d, ring full to tx %d head %d tail %d",.format_args = + "i2i2i2i2",}; + struct + { + u16 instance, to_tx, head, tail; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->instance = xd->device_index; + ed->to_tx = n_descriptors_to_tx; + ed->head = dq->head_index; + ed->tail = dq->tail_index; + } + + if (n_ok < n_descriptors_to_tx) + { + n_tail_drop = n_descriptors_to_tx - n_ok; + vec_add (xm->tx_buffers_pending_free, from + n_ok, n_tail_drop); + vlib_error_count (vm, ixge_input_node.index, + IXGE_ERROR_tx_full_drops, n_tail_drop); + } + + n_descriptors_to_tx = n_ok; + } + + dq->tx.n_buffers_on_ring += n_descriptors_to_tx; + + /* Process from tail to end of descriptor ring. */ + if (n_descriptors_to_tx > 0 && dq->tail_index < dq->n_descriptors) + { + u32 n = + clib_min (dq->n_descriptors - dq->tail_index, n_descriptors_to_tx); + n = ixge_tx_no_wrap (xm, xd, dq, from, dq->tail_index, n, &tx_state); + from += n; + n_descriptors_to_tx -= n; + dq->tail_index += n; + ASSERT (dq->tail_index <= dq->n_descriptors); + if (dq->tail_index == dq->n_descriptors) + dq->tail_index = 0; + } + + if (n_descriptors_to_tx > 0) + { + u32 n = + ixge_tx_no_wrap (xm, xd, dq, from, 0, n_descriptors_to_tx, &tx_state); + from += n; + ASSERT (n == n_descriptors_to_tx); + dq->tail_index += n; + ASSERT (dq->tail_index <= dq->n_descriptors); + if (dq->tail_index == dq->n_descriptors) + dq->tail_index = 0; + } + + /* We should only get full packets. */ + ASSERT (tx_state.is_start_of_packet); + + /* Report status when last descriptor is done. */ + { + u32 i = dq->tail_index == 0 ? dq->n_descriptors - 1 : dq->tail_index - 1; + ixge_tx_descriptor_t *d = &dq->descriptors[i].tx; + d->status0 |= IXGE_TX_DESCRIPTOR_STATUS0_REPORT_STATUS; + } + + /* Give new descriptors to hardware. */ + { + ixge_dma_regs_t *dr = get_dma_regs (xd, VLIB_TX, queue_index); + + CLIB_MEMORY_BARRIER (); + + dr->tail_index = dq->tail_index; + } + + /* Free any buffers that are done. */ + { + u32 n = _vec_len (xm->tx_buffers_pending_free); + if (n > 0) + { + vlib_buffer_free_no_next (vm, xm->tx_buffers_pending_free, n); + _vec_len (xm->tx_buffers_pending_free) = 0; + ASSERT (dq->tx.n_buffers_on_ring >= n); + dq->tx.n_buffers_on_ring -= (n - n_tail_drop); + } + } + + return f->n_vectors; +} + +static uword +ixge_rx_queue_no_wrap (ixge_main_t * xm, + ixge_device_t * xd, + ixge_dma_queue_t * dq, + u32 start_descriptor_index, u32 n_descriptors) +{ + vlib_main_t *vm = xm->vlib_main; + vlib_node_runtime_t *node = dq->rx.node; + ixge_descriptor_t *d; + static ixge_descriptor_t *d_trace_save; + static u32 *d_trace_buffers; + u32 n_descriptors_left = n_descriptors; + u32 *to_rx = + vec_elt_at_index (dq->descriptor_buffer_indices, start_descriptor_index); + u32 *to_add; + u32 bi_sop = dq->rx.saved_start_of_packet_buffer_index; + u32 bi_last = dq->rx.saved_last_buffer_index; + u32 next_index_sop = dq->rx.saved_start_of_packet_next_index; + u32 is_sop = dq->rx.is_start_of_packet; + u32 next_index, n_left_to_next, *to_next; + u32 n_packets = 0; + u32 n_bytes = 0; + u32 n_trace = vlib_get_trace_count (vm, node); + vlib_buffer_t *b_last, b_dummy; + + ASSERT (start_descriptor_index + n_descriptors <= dq->n_descriptors); + d = &dq->descriptors[start_descriptor_index]; + + b_last = bi_last != ~0 ? vlib_get_buffer (vm, bi_last) : &b_dummy; + next_index = dq->rx.next_index; + + if (n_trace > 0) + { + u32 n = clib_min (n_trace, n_descriptors); + if (d_trace_save) + { + _vec_len (d_trace_save) = 0; + _vec_len (d_trace_buffers) = 0; + } + vec_add (d_trace_save, (ixge_descriptor_t *) d, n); + vec_add (d_trace_buffers, to_rx, n); + } + + { + uword l = vec_len (xm->rx_buffers_to_add); + + if (l < n_descriptors_left) + { + u32 n_to_alloc = 2 * dq->n_descriptors - l; + u32 n_allocated; + + vec_resize (xm->rx_buffers_to_add, n_to_alloc); + + _vec_len (xm->rx_buffers_to_add) = l; + n_allocated = vlib_buffer_alloc_from_free_list + (vm, xm->rx_buffers_to_add + l, n_to_alloc, + xm->vlib_buffer_free_list_index); + _vec_len (xm->rx_buffers_to_add) += n_allocated; + + /* Handle transient allocation failure */ + if (PREDICT_FALSE (l + n_allocated <= n_descriptors_left)) + { + if (n_allocated == 0) + vlib_error_count (vm, ixge_input_node.index, + IXGE_ERROR_rx_alloc_no_physmem, 1); + else + vlib_error_count (vm, ixge_input_node.index, + IXGE_ERROR_rx_alloc_fail, 1); + + n_descriptors_left = l + n_allocated; + } + n_descriptors = n_descriptors_left; + } + + /* Add buffers from end of vector going backwards. */ + to_add = vec_end (xm->rx_buffers_to_add) - 1; + } + + while (n_descriptors_left > 0) + { + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_descriptors_left >= 4 && n_left_to_next >= 2) + { + vlib_buffer_t *b0, *b1; + u32 bi0, fi0, len0, l3_offset0, s20, s00, flags0; + u32 bi1, fi1, len1, l3_offset1, s21, s01, flags1; + u8 is_eop0, error0, next0; + u8 is_eop1, error1, next1; + ixge_descriptor_t d0, d1; + + vlib_prefetch_buffer_with_index (vm, to_rx[2], STORE); + vlib_prefetch_buffer_with_index (vm, to_rx[3], STORE); + + CLIB_PREFETCH (d + 2, 32, STORE); + + d0.as_u32x4 = d[0].as_u32x4; + d1.as_u32x4 = d[1].as_u32x4; + + s20 = d0.rx_from_hw.status[2]; + s21 = d1.rx_from_hw.status[2]; + + s00 = d0.rx_from_hw.status[0]; + s01 = d1.rx_from_hw.status[0]; + + if (! + ((s20 & s21) & IXGE_RX_DESCRIPTOR_STATUS2_IS_OWNED_BY_SOFTWARE)) + goto found_hw_owned_descriptor_x2; + + bi0 = to_rx[0]; + bi1 = to_rx[1]; + + ASSERT (to_add - 1 >= xm->rx_buffers_to_add); + fi0 = to_add[0]; + fi1 = to_add[-1]; + + to_rx[0] = fi0; + to_rx[1] = fi1; + to_rx += 2; + to_add -= 2; + + ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == + vlib_buffer_is_known (vm, bi0)); + ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == + vlib_buffer_is_known (vm, bi1)); + ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == + vlib_buffer_is_known (vm, fi0)); + ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == + vlib_buffer_is_known (vm, fi1)); + + b0 = vlib_get_buffer (vm, bi0); + b1 = vlib_get_buffer (vm, bi1); + + /* + * Turn this on if you run into + * "bad monkey" contexts, and you want to know exactly + * which nodes they've visited... See main.c... + */ + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b1); + + CLIB_PREFETCH (b0->data, CLIB_CACHE_LINE_BYTES, LOAD); + CLIB_PREFETCH (b1->data, CLIB_CACHE_LINE_BYTES, LOAD); + + is_eop0 = (s20 & IXGE_RX_DESCRIPTOR_STATUS2_IS_END_OF_PACKET) != 0; + is_eop1 = (s21 & IXGE_RX_DESCRIPTOR_STATUS2_IS_END_OF_PACKET) != 0; + + ixge_rx_next_and_error_from_status_x2 (xd, s00, s20, s01, s21, + &next0, &error0, &flags0, + &next1, &error1, &flags1); + + next0 = is_sop ? next0 : next_index_sop; + next1 = is_eop0 ? next1 : next0; + next_index_sop = next1; + + b0->flags |= flags0 | (!is_eop0 << VLIB_BUFFER_LOG2_NEXT_PRESENT); + b1->flags |= flags1 | (!is_eop1 << VLIB_BUFFER_LOG2_NEXT_PRESENT); + + vnet_buffer (b0)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b1)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + vnet_buffer (b1)->sw_if_index[VLIB_TX] = (u32) ~ 0; + + b0->error = node->errors[error0]; + b1->error = node->errors[error1]; + + len0 = d0.rx_from_hw.n_packet_bytes_this_descriptor; + len1 = d1.rx_from_hw.n_packet_bytes_this_descriptor; + n_bytes += len0 + len1; + n_packets += is_eop0 + is_eop1; + + /* Give new buffers to hardware. */ + d0.rx_to_hw.tail_address = + vlib_get_buffer_data_physical_address (vm, fi0); + d1.rx_to_hw.tail_address = + vlib_get_buffer_data_physical_address (vm, fi1); + d0.rx_to_hw.head_address = d[0].rx_to_hw.tail_address; + d1.rx_to_hw.head_address = d[1].rx_to_hw.tail_address; + d[0].as_u32x4 = d0.as_u32x4; + d[1].as_u32x4 = d1.as_u32x4; + + d += 2; + n_descriptors_left -= 2; + + /* Point to either l2 or l3 header depending on next. */ + l3_offset0 = (is_sop && (next0 != IXGE_RX_NEXT_ETHERNET_INPUT)) + ? IXGE_RX_DESCRIPTOR_STATUS0_L3_OFFSET (s00) : 0; + l3_offset1 = (is_eop0 && (next1 != IXGE_RX_NEXT_ETHERNET_INPUT)) + ? IXGE_RX_DESCRIPTOR_STATUS0_L3_OFFSET (s01) : 0; + + b0->current_length = len0 - l3_offset0; + b1->current_length = len1 - l3_offset1; + b0->current_data = l3_offset0; + b1->current_data = l3_offset1; + + b_last->next_buffer = is_sop ? ~0 : bi0; + b0->next_buffer = is_eop0 ? ~0 : bi1; + bi_last = bi1; + b_last = b1; + + if (CLIB_DEBUG > 0) + { + u32 bi_sop0 = is_sop ? bi0 : bi_sop; + u32 bi_sop1 = is_eop0 ? bi1 : bi_sop0; + + if (is_eop0) + { + u8 *msg = vlib_validate_buffer (vm, bi_sop0, + /* follow_buffer_next */ 1); + ASSERT (!msg); + } + if (is_eop1) + { + u8 *msg = vlib_validate_buffer (vm, bi_sop1, + /* follow_buffer_next */ 1); + ASSERT (!msg); + } + } + if (0) /* "Dave" version */ + { + u32 bi_sop0 = is_sop ? bi0 : bi_sop; + u32 bi_sop1 = is_eop0 ? bi1 : bi_sop0; + + if (is_eop0) + { + to_next[0] = bi_sop0; + to_next++; + n_left_to_next--; + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, + bi_sop0, next0); + } + if (is_eop1) + { + to_next[0] = bi_sop1; + to_next++; + n_left_to_next--; + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, + bi_sop1, next1); + } + is_sop = is_eop1; + bi_sop = bi_sop1; + } + if (1) /* "Eliot" version */ + { + /* Speculatively enqueue to cached next. */ + u8 saved_is_sop = is_sop; + u32 bi_sop_save = bi_sop; + + bi_sop = saved_is_sop ? bi0 : bi_sop; + to_next[0] = bi_sop; + to_next += is_eop0; + n_left_to_next -= is_eop0; + + bi_sop = is_eop0 ? bi1 : bi_sop; + to_next[0] = bi_sop; + to_next += is_eop1; + n_left_to_next -= is_eop1; + + is_sop = is_eop1; + + if (PREDICT_FALSE + (!(next0 == next_index && next1 == next_index))) + { + /* Undo speculation. */ + to_next -= is_eop0 + is_eop1; + n_left_to_next += is_eop0 + is_eop1; + + /* Re-do both descriptors being careful about where we enqueue. */ + bi_sop = saved_is_sop ? bi0 : bi_sop_save; + if (is_eop0) + { + if (next0 != next_index) + vlib_set_next_frame_buffer (vm, node, next0, bi_sop); + else + { + to_next[0] = bi_sop; + to_next += 1; + n_left_to_next -= 1; + } + } + + bi_sop = is_eop0 ? bi1 : bi_sop; + if (is_eop1) + { + if (next1 != next_index) + vlib_set_next_frame_buffer (vm, node, next1, bi_sop); + else + { + to_next[0] = bi_sop; + to_next += 1; + n_left_to_next -= 1; + } + } + + /* Switch cached next index when next for both packets is the same. */ + if (is_eop0 && is_eop1 && next0 == next1) + { + vlib_put_next_frame (vm, node, next_index, + n_left_to_next); + next_index = next0; + vlib_get_next_frame (vm, node, next_index, + to_next, n_left_to_next); + } + } + } + } + + /* Bail out of dual loop and proceed with single loop. */ + found_hw_owned_descriptor_x2: + + while (n_descriptors_left > 0 && n_left_to_next > 0) + { + vlib_buffer_t *b0; + u32 bi0, fi0, len0, l3_offset0, s20, s00, flags0; + u8 is_eop0, error0, next0; + ixge_descriptor_t d0; + + d0.as_u32x4 = d[0].as_u32x4; + + s20 = d0.rx_from_hw.status[2]; + s00 = d0.rx_from_hw.status[0]; + + if (!(s20 & IXGE_RX_DESCRIPTOR_STATUS2_IS_OWNED_BY_SOFTWARE)) + goto found_hw_owned_descriptor_x1; + + bi0 = to_rx[0]; + ASSERT (to_add >= xm->rx_buffers_to_add); + fi0 = to_add[0]; + + to_rx[0] = fi0; + to_rx += 1; + to_add -= 1; + + ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == + vlib_buffer_is_known (vm, bi0)); + ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == + vlib_buffer_is_known (vm, fi0)); + + b0 = vlib_get_buffer (vm, bi0); + + /* + * Turn this on if you run into + * "bad monkey" contexts, and you want to know exactly + * which nodes they've visited... + */ + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); + + is_eop0 = (s20 & IXGE_RX_DESCRIPTOR_STATUS2_IS_END_OF_PACKET) != 0; + ixge_rx_next_and_error_from_status_x1 + (xd, s00, s20, &next0, &error0, &flags0); + + next0 = is_sop ? next0 : next_index_sop; + next_index_sop = next0; + + b0->flags |= flags0 | (!is_eop0 << VLIB_BUFFER_LOG2_NEXT_PRESENT); + + vnet_buffer (b0)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + + b0->error = node->errors[error0]; + + len0 = d0.rx_from_hw.n_packet_bytes_this_descriptor; + n_bytes += len0; + n_packets += is_eop0; + + /* Give new buffer to hardware. */ + d0.rx_to_hw.tail_address = + vlib_get_buffer_data_physical_address (vm, fi0); + d0.rx_to_hw.head_address = d0.rx_to_hw.tail_address; + d[0].as_u32x4 = d0.as_u32x4; + + d += 1; + n_descriptors_left -= 1; + + /* Point to either l2 or l3 header depending on next. */ + l3_offset0 = (is_sop && (next0 != IXGE_RX_NEXT_ETHERNET_INPUT)) + ? IXGE_RX_DESCRIPTOR_STATUS0_L3_OFFSET (s00) : 0; + b0->current_length = len0 - l3_offset0; + b0->current_data = l3_offset0; + + b_last->next_buffer = is_sop ? ~0 : bi0; + bi_last = bi0; + b_last = b0; + + bi_sop = is_sop ? bi0 : bi_sop; + + if (CLIB_DEBUG > 0 && is_eop0) + { + u8 *msg = + vlib_validate_buffer (vm, bi_sop, /* follow_buffer_next */ 1); + ASSERT (!msg); + } + + if (0) /* "Dave" version */ + { + if (is_eop0) + { + to_next[0] = bi_sop; + to_next++; + n_left_to_next--; + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, + bi_sop, next0); + } + } + if (1) /* "Eliot" version */ + { + if (PREDICT_TRUE (next0 == next_index)) + { + to_next[0] = bi_sop; + to_next += is_eop0; + n_left_to_next -= is_eop0; + } + else + { + if (next0 != next_index && is_eop0) + vlib_set_next_frame_buffer (vm, node, next0, bi_sop); + + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + next_index = next0; + vlib_get_next_frame (vm, node, next_index, + to_next, n_left_to_next); + } + } + is_sop = is_eop0; + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + +found_hw_owned_descriptor_x1: + if (n_descriptors_left > 0) + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + + _vec_len (xm->rx_buffers_to_add) = (to_add + 1) - xm->rx_buffers_to_add; + + { + u32 n_done = n_descriptors - n_descriptors_left; + + if (n_trace > 0 && n_done > 0) + { + u32 n = clib_min (n_trace, n_done); + ixge_rx_trace (xm, xd, dq, + d_trace_save, + d_trace_buffers, + &dq->descriptors[start_descriptor_index], n); + vlib_set_trace_count (vm, node, n_trace - n); + } + if (d_trace_save) + { + _vec_len (d_trace_save) = 0; + _vec_len (d_trace_buffers) = 0; + } + + /* Don't keep a reference to b_last if we don't have to. + Otherwise we can over-write a next_buffer pointer after already haven + enqueued a packet. */ + if (is_sop) + { + b_last->next_buffer = ~0; + bi_last = ~0; + } + + dq->rx.n_descriptors_done_this_call = n_done; + dq->rx.n_descriptors_done_total += n_done; + dq->rx.is_start_of_packet = is_sop; + dq->rx.saved_start_of_packet_buffer_index = bi_sop; + dq->rx.saved_last_buffer_index = bi_last; + dq->rx.saved_start_of_packet_next_index = next_index_sop; + dq->rx.next_index = next_index; + dq->rx.n_bytes += n_bytes; + + return n_packets; + } +} + +static uword +ixge_rx_queue (ixge_main_t * xm, + ixge_device_t * xd, + vlib_node_runtime_t * node, u32 queue_index) +{ + ixge_dma_queue_t *dq = + vec_elt_at_index (xd->dma_queues[VLIB_RX], queue_index); + ixge_dma_regs_t *dr = get_dma_regs (xd, VLIB_RX, dq->queue_index); + uword n_packets = 0; + u32 hw_head_index, sw_head_index; + + /* One time initialization. */ + if (!dq->rx.node) + { + dq->rx.node = node; + dq->rx.is_start_of_packet = 1; + dq->rx.saved_start_of_packet_buffer_index = ~0; + dq->rx.saved_last_buffer_index = ~0; + } + + dq->rx.next_index = node->cached_next_index; + + dq->rx.n_descriptors_done_total = 0; + dq->rx.n_descriptors_done_this_call = 0; + dq->rx.n_bytes = 0; + + /* Fetch head from hardware and compare to where we think we are. */ + hw_head_index = dr->head_index; + sw_head_index = dq->head_index; + + if (hw_head_index == sw_head_index) + goto done; + + if (hw_head_index < sw_head_index) + { + u32 n_tried = dq->n_descriptors - sw_head_index; + n_packets += ixge_rx_queue_no_wrap (xm, xd, dq, sw_head_index, n_tried); + sw_head_index = + ixge_ring_add (dq, sw_head_index, + dq->rx.n_descriptors_done_this_call); + + if (dq->rx.n_descriptors_done_this_call != n_tried) + goto done; + } + if (hw_head_index >= sw_head_index) + { + u32 n_tried = hw_head_index - sw_head_index; + n_packets += ixge_rx_queue_no_wrap (xm, xd, dq, sw_head_index, n_tried); + sw_head_index = + ixge_ring_add (dq, sw_head_index, + dq->rx.n_descriptors_done_this_call); + } + +done: + dq->head_index = sw_head_index; + dq->tail_index = + ixge_ring_add (dq, dq->tail_index, dq->rx.n_descriptors_done_total); + + /* Give tail back to hardware. */ + CLIB_MEMORY_BARRIER (); + + dr->tail_index = dq->tail_index; + + vlib_increment_combined_counter (vnet_main. + interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, + 0 /* cpu_index */ , + xd->vlib_sw_if_index, n_packets, + dq->rx.n_bytes); + + return n_packets; +} + +static void +ixge_interrupt (ixge_main_t * xm, ixge_device_t * xd, u32 i) +{ + vlib_main_t *vm = xm->vlib_main; + ixge_regs_t *r = xd->regs; + + if (i != 20) + { + ELOG_TYPE_DECLARE (e) = + { + .function = (char *) __FUNCTION__,.format = + "ixge %d, %s",.format_args = "i1t1",.n_enum_strings = + 16,.enum_strings = + { + "flow director", + "rx miss", + "pci exception", + "mailbox", + "link status change", + "linksec key exchange", + "manageability event", + "reserved23", + "sdp0", + "sdp1", + "sdp2", + "sdp3", + "ecc", "descriptor handler error", "tcp timer", "other",},}; + struct + { + u8 instance; + u8 index; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->instance = xd->device_index; + ed->index = i - 16; + } + else + { + u32 v = r->xge_mac.link_status; + uword is_up = (v & (1 << 30)) != 0; + + ELOG_TYPE_DECLARE (e) = + { + .function = (char *) __FUNCTION__,.format = + "ixge %d, link status change 0x%x",.format_args = "i4i4",}; + struct + { + u32 instance, link_status; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->instance = xd->device_index; + ed->link_status = v; + xd->link_status_at_last_link_change = v; + + vlib_process_signal_event (vm, ixge_process_node.index, + EVENT_SET_FLAGS, + ((is_up << 31) | xd->vlib_hw_if_index)); + } +} + +always_inline u32 +clean_block (u32 * b, u32 * t, u32 n_left) +{ + u32 *t0 = t; + + while (n_left >= 4) + { + u32 bi0, bi1, bi2, bi3; + + t[0] = bi0 = b[0]; + b[0] = 0; + t += bi0 != 0; + + t[0] = bi1 = b[1]; + b[1] = 0; + t += bi1 != 0; + + t[0] = bi2 = b[2]; + b[2] = 0; + t += bi2 != 0; + + t[0] = bi3 = b[3]; + b[3] = 0; + t += bi3 != 0; + + b += 4; + n_left -= 4; + } + + while (n_left > 0) + { + u32 bi0; + + t[0] = bi0 = b[0]; + b[0] = 0; + t += bi0 != 0; + b += 1; + n_left -= 1; + } + + return t - t0; +} + +static void +ixge_tx_queue (ixge_main_t * xm, ixge_device_t * xd, u32 queue_index) +{ + vlib_main_t *vm = xm->vlib_main; + ixge_dma_queue_t *dq = + vec_elt_at_index (xd->dma_queues[VLIB_TX], queue_index); + u32 n_clean, *b, *t, *t0; + i32 n_hw_owned_descriptors; + i32 first_to_clean, last_to_clean; + u64 hwbp_race = 0; + + /* Handle case where head write back pointer update + * arrives after the interrupt during high PCI bus loads. + */ + while ((dq->head_index == dq->tx.head_index_write_back[0]) && + dq->tx.n_buffers_on_ring && (dq->head_index != dq->tail_index)) + { + hwbp_race++; + if (IXGE_HWBP_RACE_ELOG && (hwbp_race == 1)) + { + ELOG_TYPE_DECLARE (e) = + { + .function = (char *) __FUNCTION__,.format = + "ixge %d tx head index race: head %4d, tail %4d, buffs %4d",.format_args + = "i4i4i4i4",}; + struct + { + u32 instance, head_index, tail_index, n_buffers_on_ring; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->instance = xd->device_index; + ed->head_index = dq->head_index; + ed->tail_index = dq->tail_index; + ed->n_buffers_on_ring = dq->tx.n_buffers_on_ring; + } + } + + dq->head_index = dq->tx.head_index_write_back[0]; + n_hw_owned_descriptors = ixge_ring_sub (dq, dq->head_index, dq->tail_index); + ASSERT (dq->tx.n_buffers_on_ring >= n_hw_owned_descriptors); + n_clean = dq->tx.n_buffers_on_ring - n_hw_owned_descriptors; + + if (IXGE_HWBP_RACE_ELOG && hwbp_race) + { + ELOG_TYPE_DECLARE (e) = + { + .function = (char *) __FUNCTION__,.format = + "ixge %d tx head index race: head %4d, hw_owned %4d, n_clean %4d, retries %d",.format_args + = "i4i4i4i4i4",}; + struct + { + u32 instance, head_index, n_hw_owned_descriptors, n_clean, retries; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->instance = xd->device_index; + ed->head_index = dq->head_index; + ed->n_hw_owned_descriptors = n_hw_owned_descriptors; + ed->n_clean = n_clean; + ed->retries = hwbp_race; + } + + /* + * This function used to wait until hardware owned zero descriptors. + * At high PPS rates, that doesn't happen until the TX ring is + * completely full of descriptors which need to be cleaned up. + * That, in turn, causes TX ring-full drops and/or long RX service + * interruptions. + */ + if (n_clean == 0) + return; + + /* Clean the n_clean descriptors prior to the reported hardware head */ + last_to_clean = dq->head_index - 1; + last_to_clean = (last_to_clean < 0) ? last_to_clean + dq->n_descriptors : + last_to_clean; + + first_to_clean = (last_to_clean) - (n_clean - 1); + first_to_clean = (first_to_clean < 0) ? first_to_clean + dq->n_descriptors : + first_to_clean; + + vec_resize (xm->tx_buffers_pending_free, dq->n_descriptors - 1); + t0 = t = xm->tx_buffers_pending_free; + b = dq->descriptor_buffer_indices + first_to_clean; + + /* Wrap case: clean from first to end, then start to last */ + if (first_to_clean > last_to_clean) + { + t += clean_block (b, t, (dq->n_descriptors - 1) - first_to_clean); + first_to_clean = 0; + b = dq->descriptor_buffer_indices; + } + + /* Typical case: clean from first to last */ + if (first_to_clean <= last_to_clean) + t += clean_block (b, t, (last_to_clean - first_to_clean) + 1); + + if (t > t0) + { + u32 n = t - t0; + vlib_buffer_free_no_next (vm, t0, n); + ASSERT (dq->tx.n_buffers_on_ring >= n); + dq->tx.n_buffers_on_ring -= n; + _vec_len (xm->tx_buffers_pending_free) = 0; + } +} + +/* RX queue interrupts 0 thru 7; TX 8 thru 15. */ +always_inline uword +ixge_interrupt_is_rx_queue (uword i) +{ + return i < 8; +} + +always_inline uword +ixge_interrupt_is_tx_queue (uword i) +{ + return i >= 8 && i < 16; +} + +always_inline uword +ixge_tx_queue_to_interrupt (uword i) +{ + return 8 + i; +} + +always_inline uword +ixge_rx_queue_to_interrupt (uword i) +{ + return 0 + i; +} + +always_inline uword +ixge_interrupt_rx_queue (uword i) +{ + ASSERT (ixge_interrupt_is_rx_queue (i)); + return i - 0; +} + +always_inline uword +ixge_interrupt_tx_queue (uword i) +{ + ASSERT (ixge_interrupt_is_tx_queue (i)); + return i - 8; +} + +static uword +ixge_device_input (ixge_main_t * xm, + ixge_device_t * xd, vlib_node_runtime_t * node) +{ + ixge_regs_t *r = xd->regs; + u32 i, s; + uword n_rx_packets = 0; + + s = r->interrupt.status_write_1_to_set; + if (s) + r->interrupt.status_write_1_to_clear = s; + + /* *INDENT-OFF* */ + foreach_set_bit (i, s, ({ + if (ixge_interrupt_is_rx_queue (i)) + n_rx_packets += ixge_rx_queue (xm, xd, node, ixge_interrupt_rx_queue (i)); + + else if (ixge_interrupt_is_tx_queue (i)) + ixge_tx_queue (xm, xd, ixge_interrupt_tx_queue (i)); + + else + ixge_interrupt (xm, xd, i); + })); + /* *INDENT-ON* */ + + return n_rx_packets; +} + +static uword +ixge_input (vlib_main_t * vm, vlib_node_runtime_t * node, vlib_frame_t * f) +{ + ixge_main_t *xm = &ixge_main; + ixge_device_t *xd; + uword n_rx_packets = 0; + + if (node->state == VLIB_NODE_STATE_INTERRUPT) + { + uword i; + + /* Loop over devices with interrupts. */ + /* *INDENT-OFF* */ + foreach_set_bit (i, node->runtime_data[0], ({ + xd = vec_elt_at_index (xm->devices, i); + n_rx_packets += ixge_device_input (xm, xd, node); + + /* Re-enable interrupts since we're going to stay in interrupt mode. */ + if (! (node->flags & VLIB_NODE_FLAG_SWITCH_FROM_INTERRUPT_TO_POLLING_MODE)) + xd->regs->interrupt.enable_write_1_to_set = ~0; + })); + /* *INDENT-ON* */ + + /* Clear mask of devices with pending interrupts. */ + node->runtime_data[0] = 0; + } + else + { + /* Poll all devices for input/interrupts. */ + vec_foreach (xd, xm->devices) + { + n_rx_packets += ixge_device_input (xm, xd, node); + + /* Re-enable interrupts when switching out of polling mode. */ + if (node->flags & + VLIB_NODE_FLAG_SWITCH_FROM_POLLING_TO_INTERRUPT_MODE) + xd->regs->interrupt.enable_write_1_to_set = ~0; + } + } + + return n_rx_packets; +} + +static char *ixge_error_strings[] = { +#define _(n,s) s, + foreach_ixge_error +#undef _ +}; + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (ixge_input_node, static) = { + .function = ixge_input, + .type = VLIB_NODE_TYPE_INPUT, + .name = "ixge-input", + + /* Will be enabled if/when hardware is detected. */ + .state = VLIB_NODE_STATE_DISABLED, + + .format_buffer = format_ethernet_header_with_length, + .format_trace = format_ixge_rx_dma_trace, + + .n_errors = IXGE_N_ERROR, + .error_strings = ixge_error_strings, + + .n_next_nodes = IXGE_RX_N_NEXT, + .next_nodes = { + [IXGE_RX_NEXT_DROP] = "error-drop", + [IXGE_RX_NEXT_ETHERNET_INPUT] = "ethernet-input", + [IXGE_RX_NEXT_IP4_INPUT] = "ip4-input", + [IXGE_RX_NEXT_IP6_INPUT] = "ip6-input", + }, +}; + +VLIB_NODE_FUNCTION_MULTIARCH_CLONE (ixge_input) +CLIB_MULTIARCH_SELECT_FN (ixge_input) +/* *INDENT-ON* */ + +static u8 * +format_ixge_device_name (u8 * s, va_list * args) +{ + u32 i = va_arg (*args, u32); + ixge_main_t *xm = &ixge_main; + ixge_device_t *xd = vec_elt_at_index (xm->devices, i); + return format (s, "TenGigabitEthernet%U", + format_vlib_pci_handle, &xd->pci_device.bus_address); +} + +#define IXGE_COUNTER_IS_64_BIT (1 << 0) +#define IXGE_COUNTER_NOT_CLEAR_ON_READ (1 << 1) + +static u8 ixge_counter_flags[] = { +#define _(a,f) 0, +#define _64(a,f) IXGE_COUNTER_IS_64_BIT, + foreach_ixge_counter +#undef _ +#undef _64 +}; + +static void +ixge_update_counters (ixge_device_t * xd) +{ + /* Byte offset for counter registers. */ + static u32 reg_offsets[] = { +#define _(a,f) (a) / sizeof (u32), +#define _64(a,f) _(a,f) + foreach_ixge_counter +#undef _ +#undef _64 + }; + volatile u32 *r = (volatile u32 *) xd->regs; + int i; + + for (i = 0; i < ARRAY_LEN (xd->counters); i++) + { + u32 o = reg_offsets[i]; + xd->counters[i] += r[o]; + if (ixge_counter_flags[i] & IXGE_COUNTER_NOT_CLEAR_ON_READ) + r[o] = 0; + if (ixge_counter_flags[i] & IXGE_COUNTER_IS_64_BIT) + xd->counters[i] += (u64) r[o + 1] << (u64) 32; + } +} + +static u8 * +format_ixge_device_id (u8 * s, va_list * args) +{ + u32 device_id = va_arg (*args, u32); + char *t = 0; + switch (device_id) + { +#define _(f,n) case n: t = #f; break; + foreach_ixge_pci_device_id; +#undef _ + default: + t = 0; + break; + } + if (t == 0) + s = format (s, "unknown 0x%x", device_id); + else + s = format (s, "%s", t); + return s; +} + +static u8 * +format_ixge_link_status (u8 * s, va_list * args) +{ + ixge_device_t *xd = va_arg (*args, ixge_device_t *); + u32 v = xd->link_status_at_last_link_change; + + s = format (s, "%s", (v & (1 << 30)) ? "up" : "down"); + + { + char *modes[] = { + "1g", "10g parallel", "10g serial", "autoneg", + }; + char *speeds[] = { + "unknown", "100m", "1g", "10g", + }; + s = format (s, ", mode %s, speed %s", + modes[(v >> 26) & 3], speeds[(v >> 28) & 3]); + } + + return s; +} + +static u8 * +format_ixge_device (u8 * s, va_list * args) +{ + u32 dev_instance = va_arg (*args, u32); + CLIB_UNUSED (int verbose) = va_arg (*args, int); + ixge_main_t *xm = &ixge_main; + ixge_device_t *xd = vec_elt_at_index (xm->devices, dev_instance); + ixge_phy_t *phy = xd->phys + xd->phy_index; + uword indent = format_get_indent (s); + + ixge_update_counters (xd); + xd->link_status_at_last_link_change = xd->regs->xge_mac.link_status; + + s = format (s, "Intel 8259X: id %U\n%Ulink %U", + format_ixge_device_id, xd->device_id, + format_white_space, indent + 2, format_ixge_link_status, xd); + + { + + s = format (s, "\n%UPCIe %U", format_white_space, indent + 2, + format_vlib_pci_link_speed, &xd->pci_device); + } + + s = format (s, "\n%U", format_white_space, indent + 2); + if (phy->mdio_address != ~0) + s = format (s, "PHY address %d, id 0x%x", phy->mdio_address, phy->id); + else if (xd->sfp_eeprom.id == SFP_ID_sfp) + s = format (s, "SFP %U", format_sfp_eeprom, &xd->sfp_eeprom); + else + s = format (s, "PHY not found"); + + /* FIXME */ + { + ixge_dma_queue_t *dq = vec_elt_at_index (xd->dma_queues[VLIB_RX], 0); + ixge_dma_regs_t *dr = get_dma_regs (xd, VLIB_RX, 0); + u32 hw_head_index = dr->head_index; + u32 sw_head_index = dq->head_index; + u32 nitems; + + nitems = ixge_ring_sub (dq, hw_head_index, sw_head_index); + s = format (s, "\n%U%d unprocessed, %d total buffers on rx queue 0 ring", + format_white_space, indent + 2, nitems, dq->n_descriptors); + + s = format (s, "\n%U%d buffers in driver rx cache", + format_white_space, indent + 2, + vec_len (xm->rx_buffers_to_add)); + + s = format (s, "\n%U%d buffers on tx queue 0 ring", + format_white_space, indent + 2, + xd->dma_queues[VLIB_TX][0].tx.n_buffers_on_ring); + } + { + u32 i; + u64 v; + static char *names[] = { +#define _(a,f) #f, +#define _64(a,f) _(a,f) + foreach_ixge_counter +#undef _ +#undef _64 + }; + + for (i = 0; i < ARRAY_LEN (names); i++) + { + v = xd->counters[i] - xd->counters_last_clear[i]; + if (v != 0) + s = format (s, "\n%U%-40U%16Ld", + format_white_space, indent + 2, + format_c_identifier, names[i], v); + } + } + + return s; +} + +static void +ixge_clear_hw_interface_counters (u32 instance) +{ + ixge_main_t *xm = &ixge_main; + ixge_device_t *xd = vec_elt_at_index (xm->devices, instance); + ixge_update_counters (xd); + memcpy (xd->counters_last_clear, xd->counters, sizeof (xd->counters)); +} + +/* + * Dynamically redirect all pkts from a specific interface + * to the specified node + */ +static void +ixge_set_interface_next_node (vnet_main_t * vnm, u32 hw_if_index, + u32 node_index) +{ + ixge_main_t *xm = &ixge_main; + vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index); + ixge_device_t *xd = vec_elt_at_index (xm->devices, hw->dev_instance); + + /* Shut off redirection */ + if (node_index == ~0) + { + xd->per_interface_next_index = node_index; + return; + } + + xd->per_interface_next_index = + vlib_node_add_next (xm->vlib_main, ixge_input_node.index, node_index); +} + + +/* *INDENT-OFF* */ +VNET_DEVICE_CLASS (ixge_device_class) = { + .name = "ixge", + .tx_function = ixge_interface_tx, + .format_device_name = format_ixge_device_name, + .format_device = format_ixge_device, + .format_tx_trace = format_ixge_tx_dma_trace, + .clear_counters = ixge_clear_hw_interface_counters, + .admin_up_down_function = ixge_interface_admin_up_down, + .rx_redirect_to_node = ixge_set_interface_next_node, + .flatten_output_chains = 1, +}; +/* *INDENT-ON* */ + +#define IXGE_N_BYTES_IN_RX_BUFFER (2048) // DAW-HACK: Set Rx buffer size so all packets < ETH_MTU_SIZE fit in the buffer (i.e. sop & eop for all descriptors). + +static clib_error_t * +ixge_dma_init (ixge_device_t * xd, vlib_rx_or_tx_t rt, u32 queue_index) +{ + ixge_main_t *xm = &ixge_main; + vlib_main_t *vm = xm->vlib_main; + ixge_dma_queue_t *dq; + clib_error_t *error = 0; + + vec_validate (xd->dma_queues[rt], queue_index); + dq = vec_elt_at_index (xd->dma_queues[rt], queue_index); + + if (!xm->n_descriptors_per_cache_line) + xm->n_descriptors_per_cache_line = + CLIB_CACHE_LINE_BYTES / sizeof (dq->descriptors[0]); + + if (!xm->n_bytes_in_rx_buffer) + xm->n_bytes_in_rx_buffer = IXGE_N_BYTES_IN_RX_BUFFER; + xm->n_bytes_in_rx_buffer = round_pow2 (xm->n_bytes_in_rx_buffer, 1024); + if (!xm->vlib_buffer_free_list_index) + { + xm->vlib_buffer_free_list_index = + vlib_buffer_get_or_create_free_list (vm, xm->n_bytes_in_rx_buffer, + "ixge rx"); + ASSERT (xm->vlib_buffer_free_list_index != 0); + } + + if (!xm->n_descriptors[rt]) + xm->n_descriptors[rt] = 4 * VLIB_FRAME_SIZE; + + dq->queue_index = queue_index; + dq->n_descriptors = + round_pow2 (xm->n_descriptors[rt], xm->n_descriptors_per_cache_line); + dq->head_index = dq->tail_index = 0; + + dq->descriptors = vlib_physmem_alloc_aligned (vm, &error, + dq->n_descriptors * + sizeof (dq->descriptors[0]), + 128 /* per chip spec */ ); + if (error) + return error; + + memset (dq->descriptors, 0, + dq->n_descriptors * sizeof (dq->descriptors[0])); + vec_resize (dq->descriptor_buffer_indices, dq->n_descriptors); + + if (rt == VLIB_RX) + { + u32 n_alloc, i; + + n_alloc = vlib_buffer_alloc_from_free_list + (vm, dq->descriptor_buffer_indices, + vec_len (dq->descriptor_buffer_indices), + xm->vlib_buffer_free_list_index); + ASSERT (n_alloc == vec_len (dq->descriptor_buffer_indices)); + for (i = 0; i < n_alloc; i++) + { + vlib_buffer_t *b = + vlib_get_buffer (vm, dq->descriptor_buffer_indices[i]); + dq->descriptors[i].rx_to_hw.tail_address = + vlib_physmem_virtual_to_physical (vm, b->data); + } + } + else + { + u32 i; + + dq->tx.head_index_write_back = + vlib_physmem_alloc (vm, &error, CLIB_CACHE_LINE_BYTES); + + for (i = 0; i < dq->n_descriptors; i++) + dq->descriptors[i].tx = xm->tx_descriptor_template; + + vec_validate (xm->tx_buffers_pending_free, dq->n_descriptors - 1); + } + + { + ixge_dma_regs_t *dr = get_dma_regs (xd, rt, queue_index); + u64 a; + + a = vlib_physmem_virtual_to_physical (vm, dq->descriptors); + dr->descriptor_address[0] = a & 0xFFFFFFFF; + dr->descriptor_address[1] = a >> (u64) 32; + dr->n_descriptor_bytes = dq->n_descriptors * sizeof (dq->descriptors[0]); + dq->head_index = dq->tail_index = 0; + + if (rt == VLIB_RX) + { + ASSERT ((xm->n_bytes_in_rx_buffer / 1024) < 32); + dr->rx_split_control = + ( /* buffer size */ ((xm->n_bytes_in_rx_buffer / 1024) << 0) + | ( /* lo free descriptor threshold (units of 64 descriptors) */ + (1 << 22)) | ( /* descriptor type: advanced one buffer */ + (1 << 25)) | ( /* drop if no descriptors available */ + (1 << 28))); + + /* Give hardware all but last 16 cache lines' worth of descriptors. */ + dq->tail_index = dq->n_descriptors - + 16 * xm->n_descriptors_per_cache_line; + } + else + { + /* Make sure its initialized before hardware can get to it. */ + dq->tx.head_index_write_back[0] = dq->head_index; + + a = + vlib_physmem_virtual_to_physical (vm, dq->tx.head_index_write_back); + dr->tx.head_index_write_back_address[0] = /* enable bit */ 1 | a; + dr->tx.head_index_write_back_address[1] = (u64) a >> (u64) 32; + } + + /* DMA on 82599 does not work with [13] rx data write relaxed ordering + and [12] undocumented set. */ + if (rt == VLIB_RX) + dr->dca_control &= ~((1 << 13) | (1 << 12)); + + CLIB_MEMORY_BARRIER (); + + if (rt == VLIB_TX) + { + xd->regs->tx_dma_control |= (1 << 0); + dr->control |= ((32 << 0) /* prefetch threshold */ + | (64 << 8) /* host threshold */ + | (0 << 16) /* writeback threshold */ ); + } + + /* Enable this queue and wait for hardware to initialize + before adding to tail. */ + if (rt == VLIB_TX) + { + dr->control |= 1 << 25; + while (!(dr->control & (1 << 25))) + ; + } + + /* Set head/tail indices and enable DMA. */ + dr->head_index = dq->head_index; + dr->tail_index = dq->tail_index; + } + + return error; +} + +static u32 +ixge_flag_change (vnet_main_t * vnm, vnet_hw_interface_t * hw, u32 flags) +{ + ixge_device_t *xd; + ixge_regs_t *r; + u32 old; + ixge_main_t *xm = &ixge_main; + + xd = vec_elt_at_index (xm->devices, hw->dev_instance); + r = xd->regs; + + old = r->filter_control; + + if (flags & ETHERNET_INTERFACE_FLAG_ACCEPT_ALL) + r->filter_control = old | (1 << 9) /* unicast promiscuous */ ; + else + r->filter_control = old & ~(1 << 9); + + return old; +} + +static void +ixge_device_init (ixge_main_t * xm) +{ + vnet_main_t *vnm = vnet_get_main (); + ixge_device_t *xd; + + /* Reset chip(s). */ + vec_foreach (xd, xm->devices) + { + ixge_regs_t *r = xd->regs; + const u32 reset_bit = (1 << 26) | (1 << 3); + + r->control |= reset_bit; + + /* No need to suspend. Timed to take ~1e-6 secs */ + while (r->control & reset_bit) + ; + + /* Software loaded. */ + r->extended_control |= (1 << 28); + + ixge_phy_init (xd); + + /* Register ethernet interface. */ + { + u8 addr8[6]; + u32 i, addr32[2]; + clib_error_t *error; + + addr32[0] = r->rx_ethernet_address0[0][0]; + addr32[1] = r->rx_ethernet_address0[0][1]; + for (i = 0; i < 6; i++) + addr8[i] = addr32[i / 4] >> ((i % 4) * 8); + + error = ethernet_register_interface + (vnm, ixge_device_class.index, xd->device_index, + /* ethernet address */ addr8, + &xd->vlib_hw_if_index, ixge_flag_change); + if (error) + clib_error_report (error); + } + + { + vnet_sw_interface_t *sw = + vnet_get_hw_sw_interface (vnm, xd->vlib_hw_if_index); + xd->vlib_sw_if_index = sw->sw_if_index; + } + + ixge_dma_init (xd, VLIB_RX, /* queue_index */ 0); + + xm->n_descriptors[VLIB_TX] = 20 * VLIB_FRAME_SIZE; + + ixge_dma_init (xd, VLIB_TX, /* queue_index */ 0); + + /* RX/TX queue 0 gets mapped to interrupt bits 0 & 8. */ + r->interrupt.queue_mapping[0] = (( /* valid bit */ (1 << 7) | + ixge_rx_queue_to_interrupt (0)) << 0); + + r->interrupt.queue_mapping[0] |= (( /* valid bit */ (1 << 7) | + ixge_tx_queue_to_interrupt (0)) << 8); + + /* No use in getting too many interrupts. + Limit them to one every 3/4 ring size at line rate + min sized packets. + No need for this since kernel/vlib main loop provides adequate interrupt + limiting scheme. */ + if (0) + { + f64 line_rate_max_pps = + 10e9 / (8 * (64 + /* interframe padding */ 20)); + ixge_throttle_queue_interrupt (r, 0, + .75 * xm->n_descriptors[VLIB_RX] / + line_rate_max_pps); + } + + /* Accept all multicast and broadcast packets. Should really add them + to the dst_ethernet_address register array. */ + r->filter_control |= (1 << 10) | (1 << 8); + + /* Enable frames up to size in mac frame size register. */ + r->xge_mac.control |= 1 << 2; + r->xge_mac.rx_max_frame_size = (9216 + 14) << 16; + + /* Enable all interrupts. */ + if (!IXGE_ALWAYS_POLL) + r->interrupt.enable_write_1_to_set = ~0; + } +} + +static uword +ixge_process (vlib_main_t * vm, vlib_node_runtime_t * rt, vlib_frame_t * f) +{ + vnet_main_t *vnm = vnet_get_main (); + ixge_main_t *xm = &ixge_main; + ixge_device_t *xd; + uword event_type, *event_data = 0; + f64 timeout, link_debounce_deadline; + + ixge_device_init (xm); + + /* Clear all counters. */ + vec_foreach (xd, xm->devices) + { + ixge_update_counters (xd); + memset (xd->counters, 0, sizeof (xd->counters)); + } + + timeout = 30.0; + link_debounce_deadline = 1e70; + + while (1) + { + /* 36 bit stat counters could overflow in ~50 secs. + We poll every 30 secs to be conservative. */ + vlib_process_wait_for_event_or_clock (vm, timeout); + + event_type = vlib_process_get_events (vm, &event_data); + + switch (event_type) + { + case EVENT_SET_FLAGS: + /* 1 ms */ + link_debounce_deadline = vlib_time_now (vm) + 1e-3; + timeout = 1e-3; + break; + + case ~0: + /* No events found: timer expired. */ + if (vlib_time_now (vm) > link_debounce_deadline) + { + vec_foreach (xd, xm->devices) + { + ixge_regs_t *r = xd->regs; + u32 v = r->xge_mac.link_status; + uword is_up = (v & (1 << 30)) != 0; + + vnet_hw_interface_set_flags + (vnm, xd->vlib_hw_if_index, + is_up ? VNET_HW_INTERFACE_FLAG_LINK_UP : 0); + } + link_debounce_deadline = 1e70; + timeout = 30.0; + } + break; + + default: + ASSERT (0); + } + + if (event_data) + _vec_len (event_data) = 0; + + /* Query stats every 30 secs. */ + { + f64 now = vlib_time_now (vm); + if (now - xm->time_last_stats_update > 30) + { + xm->time_last_stats_update = now; + vec_foreach (xd, xm->devices) ixge_update_counters (xd); + } + } + } + + return 0; +} + +static vlib_node_registration_t ixge_process_node = { + .function = ixge_process, + .type = VLIB_NODE_TYPE_PROCESS, + .name = "ixge-process", +}; + +clib_error_t * +ixge_init (vlib_main_t * vm) +{ + ixge_main_t *xm = &ixge_main; + clib_error_t *error; + + xm->vlib_main = vm; + memset (&xm->tx_descriptor_template, 0, + sizeof (xm->tx_descriptor_template)); + memset (&xm->tx_descriptor_template_mask, 0, + sizeof (xm->tx_descriptor_template_mask)); + xm->tx_descriptor_template.status0 = + (IXGE_TX_DESCRIPTOR_STATUS0_ADVANCED | + IXGE_TX_DESCRIPTOR_STATUS0_IS_ADVANCED | + IXGE_TX_DESCRIPTOR_STATUS0_INSERT_FCS); + xm->tx_descriptor_template_mask.status0 = 0xffff; + xm->tx_descriptor_template_mask.status1 = 0x00003fff; + + xm->tx_descriptor_template_mask.status0 &= + ~(IXGE_TX_DESCRIPTOR_STATUS0_IS_END_OF_PACKET + | IXGE_TX_DESCRIPTOR_STATUS0_REPORT_STATUS); + xm->tx_descriptor_template_mask.status1 &= + ~(IXGE_TX_DESCRIPTOR_STATUS1_DONE); + + error = vlib_call_init_function (vm, pci_bus_init); + + return error; +} + +VLIB_INIT_FUNCTION (ixge_init); + + +static void +ixge_pci_intr_handler (vlib_pci_device_t * dev) +{ + ixge_main_t *xm = &ixge_main; + vlib_main_t *vm = xm->vlib_main; + + vlib_node_set_interrupt_pending (vm, ixge_input_node.index); + + /* Let node know which device is interrupting. */ + { + vlib_node_runtime_t *rt = + vlib_node_get_runtime (vm, ixge_input_node.index); + rt->runtime_data[0] |= 1 << dev->private_data; + } +} + +static clib_error_t * +ixge_pci_init (vlib_main_t * vm, vlib_pci_device_t * dev) +{ + ixge_main_t *xm = &ixge_main; + clib_error_t *error; + void *r; + ixge_device_t *xd; + + /* Device found: make sure we have dma memory. */ + if (unix_physmem_is_fake (vm)) + return clib_error_return (0, "no physical memory available"); + + error = vlib_pci_map_resource (dev, 0, &r); + if (error) + return error; + + vec_add2 (xm->devices, xd, 1); + + if (vec_len (xm->devices) == 1) + { + ixge_input_node.function = ixge_input_multiarch_select (); + } + + xd->pci_device = dev[0]; + xd->device_id = xd->pci_device.config0.header.device_id; + xd->regs = r; + xd->device_index = xd - xm->devices; + xd->pci_function = dev->bus_address.function; + xd->per_interface_next_index = ~0; + + + /* Chip found so enable node. */ + { + vlib_node_set_state (vm, ixge_input_node.index, + (IXGE_ALWAYS_POLL + ? VLIB_NODE_STATE_POLLING + : VLIB_NODE_STATE_INTERRUPT)); + + dev->private_data = xd->device_index; + } + + if (vec_len (xm->devices) == 1) + { + vlib_register_node (vm, &ixge_process_node); + xm->process_node_index = ixge_process_node.index; + } + + error = vlib_pci_bus_master_enable (dev); + + if (error) + return error; + + return vlib_pci_intr_enable (dev); +} + +/* *INDENT-OFF* */ +PCI_REGISTER_DEVICE (ixge_pci_device_registration,static) = { + .init_function = ixge_pci_init, + .interrupt_handler = ixge_pci_intr_handler, + .supported_devices = { +#define _(t,i) { .vendor_id = PCI_VENDOR_ID_INTEL, .device_id = i, }, + foreach_ixge_pci_device_id +#undef _ + { 0 }, + }, +}; +/* *INDENT-ON* */ + +void +ixge_set_next_node (ixge_rx_next_t next, char *name) +{ + vlib_node_registration_t *r = &ixge_input_node; + + switch (next) + { + case IXGE_RX_NEXT_IP4_INPUT: + case IXGE_RX_NEXT_IP6_INPUT: + case IXGE_RX_NEXT_ETHERNET_INPUT: + r->next_nodes[next] = name; + break; + + default: + clib_warning ("%s: illegal next %d\n", __FUNCTION__, next); + break; + } +} +#endif + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/nic/ixge.h b/src/vnet/devices/nic/ixge.h new file mode 100644 index 00000000000..a8e652dcdab --- /dev/null +++ b/src/vnet/devices/nic/ixge.h @@ -0,0 +1,1293 @@ +/* + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef included_ixge_h +#define included_ixge_h + +#include +#include +#include +#include +#include +#include + +typedef volatile struct +{ + /* [31:7] 128 byte aligned. */ + u32 descriptor_address[2]; + u32 n_descriptor_bytes; + + /* [5] rx/tx descriptor dca enable + [6] rx packet head dca enable + [7] rx packet tail dca enable + [9] rx/tx descriptor relaxed order + [11] rx/tx descriptor write back relaxed order + [13] rx/tx data write/read relaxed order + [15] rx head data write relaxed order + [31:24] apic id for cpu's cache. */ + u32 dca_control; + + u32 head_index; + + /* [4:0] tail buffer size (in 1k byte units) + [13:8] head buffer size (in 64 byte units) + [24:22] lo free descriptors threshold (units of 64 descriptors) + [27:25] descriptor type 0 = legacy, 1 = advanced one buffer (e.g. tail), + 2 = advanced header splitting (head + tail), 5 = advanced header + splitting (head only). + [28] drop if no descriptors available. */ + u32 rx_split_control; + + u32 tail_index; + CLIB_PAD_FROM_TO (0x1c, 0x28); + + /* [7:0] rx/tx prefetch threshold + [15:8] rx/tx host threshold + [24:16] rx/tx write back threshold + [25] rx/tx enable + [26] tx descriptor writeback flush + [30] rx strip vlan enable */ + u32 control; + + u32 rx_coallesce_control; + + union + { + struct + { + /* packets bytes lo hi */ + u32 stats[3]; + + u32 unused; + } rx; + + struct + { + u32 unused[2]; + + /* [0] enables head write back. */ + u32 head_index_write_back_address[2]; + } tx; + }; +} ixge_dma_regs_t; + +/* Only advanced descriptors are supported. */ +typedef struct +{ + u64 tail_address; + u64 head_address; +} ixge_rx_to_hw_descriptor_t; + +typedef struct +{ + u32 status[3]; + u16 n_packet_bytes_this_descriptor; + u16 vlan_tag; +} ixge_rx_from_hw_descriptor_t; + +#define IXGE_RX_DESCRIPTOR_STATUS0_IS_LAYER2 (1 << (4 + 11)) +/* Valid if not layer2. */ +#define IXGE_RX_DESCRIPTOR_STATUS0_IS_IP4 (1 << (4 + 0)) +#define IXGE_RX_DESCRIPTOR_STATUS0_IS_IP4_EXT (1 << (4 + 1)) +#define IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6 (1 << (4 + 2)) +#define IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6_EXT (1 << (4 + 3)) +#define IXGE_RX_DESCRIPTOR_STATUS0_IS_TCP (1 << (4 + 4)) +#define IXGE_RX_DESCRIPTOR_STATUS0_IS_UDP (1 << (4 + 5)) +#define IXGE_RX_DESCRIPTOR_STATUS0_L3_OFFSET(s) (((s) >> 21) & 0x3ff) + +#define IXGE_RX_DESCRIPTOR_STATUS2_IS_OWNED_BY_SOFTWARE (1 << (0 + 0)) +#define IXGE_RX_DESCRIPTOR_STATUS2_IS_END_OF_PACKET (1 << (0 + 1)) +#define IXGE_RX_DESCRIPTOR_STATUS2_IS_VLAN (1 << (0 + 3)) +#define IXGE_RX_DESCRIPTOR_STATUS2_IS_UDP_CHECKSUMMED (1 << (0 + 4)) +#define IXGE_RX_DESCRIPTOR_STATUS2_IS_TCP_CHECKSUMMED (1 << (0 + 5)) +#define IXGE_RX_DESCRIPTOR_STATUS2_IS_IP4_CHECKSUMMED (1 << (0 + 6)) +#define IXGE_RX_DESCRIPTOR_STATUS2_NOT_UNICAST (1 << (0 + 7)) +#define IXGE_RX_DESCRIPTOR_STATUS2_IS_DOUBLE_VLAN (1 << (0 + 9)) +#define IXGE_RX_DESCRIPTOR_STATUS2_UDP_CHECKSUM_ERROR (1 << (0 + 10)) +#define IXGE_RX_DESCRIPTOR_STATUS2_ETHERNET_ERROR (1 << (20 + 9)) +#define IXGE_RX_DESCRIPTOR_STATUS2_TCP_CHECKSUM_ERROR (1 << (20 + 10)) +#define IXGE_RX_DESCRIPTOR_STATUS2_IP4_CHECKSUM_ERROR (1 << (20 + 11)) + +/* For layer2 packets stats0 bottom 3 bits give ether type index from filter. */ +#define IXGE_RX_DESCRIPTOR_STATUS0_LAYER2_ETHERNET_TYPE(s) ((s) & 7) + +typedef struct +{ + u64 buffer_address; + u16 n_bytes_this_buffer; + u16 status0; + u32 status1; +#define IXGE_TX_DESCRIPTOR_STATUS0_ADVANCED (3 << 4) +#define IXGE_TX_DESCRIPTOR_STATUS0_IS_ADVANCED (1 << (8 + 5)) +#define IXGE_TX_DESCRIPTOR_STATUS0_LOG2_REPORT_STATUS (8 + 3) +#define IXGE_TX_DESCRIPTOR_STATUS0_REPORT_STATUS (1 << IXGE_TX_DESCRIPTOR_STATUS0_LOG2_REPORT_STATUS) +#define IXGE_TX_DESCRIPTOR_STATUS0_INSERT_FCS (1 << (8 + 1)) +#define IXGE_TX_DESCRIPTOR_STATUS0_LOG2_IS_END_OF_PACKET (8 + 0) +#define IXGE_TX_DESCRIPTOR_STATUS0_IS_END_OF_PACKET (1 << IXGE_TX_DESCRIPTOR_STATUS0_LOG2_IS_END_OF_PACKET) +#define IXGE_TX_DESCRIPTOR_STATUS1_DONE (1 << 0) +#define IXGE_TX_DESCRIPTOR_STATUS1_CONTEXT(i) (/* valid */ (1 << 7) | ((i) << 4)) +#define IXGE_TX_DESCRIPTOR_STATUS1_IPSEC_OFFLOAD (1 << (8 + 2)) +#define IXGE_TX_DESCRIPTOR_STATUS1_INSERT_TCP_UDP_CHECKSUM (1 << (8 + 1)) +#define IXGE_TX_DESCRIPTOR_STATUS1_INSERT_IP4_CHECKSUM (1 << (8 + 0)) +#define IXGE_TX_DESCRIPTOR_STATUS0_N_BYTES_THIS_BUFFER(l) ((l) << 0) +#define IXGE_TX_DESCRIPTOR_STATUS1_N_BYTES_IN_PACKET(l) ((l) << 14) +} ixge_tx_descriptor_t; + +typedef struct +{ + struct + { + u8 checksum_start_offset; + u8 checksum_insert_offset; + u16 checksum_end_offset; + } ip, tcp; + u32 status0; + + u8 status1; + + /* Byte offset after UDP/TCP header. */ + u8 payload_offset; + + u16 max_tcp_segment_size; +} __attribute__ ((packed)) ixge_tx_context_descriptor_t; + +typedef union +{ + ixge_rx_to_hw_descriptor_t rx_to_hw; + ixge_rx_from_hw_descriptor_t rx_from_hw; + ixge_tx_descriptor_t tx; + u32x4 as_u32x4; +} ixge_descriptor_t; + +typedef volatile struct +{ + /* [2] pcie master disable + [3] mac reset + [26] global device reset */ + u32 control; + u32 control_alias; + /* [3:2] device id (0 or 1 for dual port chips) + [7] link is up + [17:10] num vfs + [18] io active + [19] pcie master enable status */ + u32 status_read_only; + CLIB_PAD_FROM_TO (0xc, 0x18); + /* [14] pf reset done + [17] relaxed ordering disable + [26] extended vlan enable + [28] driver loaded */ + u32 extended_control; + CLIB_PAD_FROM_TO (0x1c, 0x20); + + /* software definable pins. + sdp_data [7:0] + sdp_is_output [15:8] + sdp_is_native [23:16] + sdp_function [31:24]. + */ + u32 sdp_control; + CLIB_PAD_FROM_TO (0x24, 0x28); + + /* [0] i2c clock in + [1] i2c clock out + [2] i2c data in + [3] i2c data out */ + u32 i2c_control; + CLIB_PAD_FROM_TO (0x2c, 0x4c); + u32 tcp_timer; + + CLIB_PAD_FROM_TO (0x50, 0x200); + + u32 led_control; + + CLIB_PAD_FROM_TO (0x204, 0x600); + u32 core_spare; + CLIB_PAD_FROM_TO (0x604, 0x700); + + struct + { + u32 vflr_events_clear[4]; + u32 mailbox_interrupt_status[4]; + u32 mailbox_interrupt_enable[4]; + CLIB_PAD_FROM_TO (0x730, 0x800); + } pf_foo; + + struct + { + u32 status_write_1_to_clear; + CLIB_PAD_FROM_TO (0x804, 0x808); + u32 status_write_1_to_set; + CLIB_PAD_FROM_TO (0x80c, 0x810); + u32 status_auto_clear_enable; + CLIB_PAD_FROM_TO (0x814, 0x820); + + /* [11:3] minimum inter-interrupt interval + (2e-6 units; 20e-6 units for fast ethernet). + [15] low-latency interrupt moderation enable + [20:16] low-latency interrupt credit + [27:21] interval counter + [31] write disable for credit and counter (write only). */ + u32 throttle0[24]; + + u32 enable_write_1_to_set; + CLIB_PAD_FROM_TO (0x884, 0x888); + u32 enable_write_1_to_clear; + CLIB_PAD_FROM_TO (0x88c, 0x890); + u32 enable_auto_clear; + u32 msi_to_eitr_select; + /* [3:0] spd 0-3 interrupt detection enable + [4] msi-x enable + [5] other clear disable (makes other bits in status not clear on read) + etc. */ + u32 control; + CLIB_PAD_FROM_TO (0x89c, 0x900); + + /* Defines interrupt mapping for 128 rx + 128 tx queues. + 64 x 4 8 bit entries. + For register [i]: + [5:0] bit in interrupt status for rx queue 2*i + 0 + [7] valid bit + [13:8] bit for tx queue 2*i + 0 + [15] valid bit + similar for rx 2*i + 1 and tx 2*i + 1. */ + u32 queue_mapping[64]; + + /* tcp timer [7:0] and other interrupts [15:8] */ + u32 misc_mapping; + CLIB_PAD_FROM_TO (0xa04, 0xa90); + + /* 64 interrupts determined by mappings. */ + u32 status1_write_1_to_clear[4]; + u32 enable1_write_1_to_set[4]; + u32 enable1_write_1_to_clear[4]; + CLIB_PAD_FROM_TO (0xac0, 0xad0); + u32 status1_enable_auto_clear[4]; + CLIB_PAD_FROM_TO (0xae0, 0x1000); + } interrupt; + + ixge_dma_regs_t rx_dma0[64]; + + CLIB_PAD_FROM_TO (0x2000, 0x2140); + u32 dcb_rx_packet_plane_t4_config[8]; + u32 dcb_rx_packet_plane_t4_status[8]; + CLIB_PAD_FROM_TO (0x2180, 0x2300); + + /* reg i defines mapping for 4 rx queues starting at 4*i + 0. */ + u32 rx_queue_stats_mapping[32]; + u32 rx_queue_stats_control; + + CLIB_PAD_FROM_TO (0x2384, 0x2410); + u32 fc_user_descriptor_ptr[2]; + u32 fc_buffer_control; + CLIB_PAD_FROM_TO (0x241c, 0x2420); + u32 fc_rx_dma; + CLIB_PAD_FROM_TO (0x2424, 0x2430); + u32 dcb_packet_plane_control; + CLIB_PAD_FROM_TO (0x2434, 0x2f00); + + u32 rx_dma_control; + u32 pf_queue_drop_enable; + CLIB_PAD_FROM_TO (0x2f08, 0x2f20); + u32 rx_dma_descriptor_cache_config; + CLIB_PAD_FROM_TO (0x2f24, 0x3000); + + /* 1 bit. */ + u32 rx_enable; + CLIB_PAD_FROM_TO (0x3004, 0x3008); + /* [15:0] ether type (little endian) + [31:16] opcode (big endian) */ + u32 flow_control_control; + CLIB_PAD_FROM_TO (0x300c, 0x3020); + /* 3 bit traffic class for each of 8 priorities. */ + u32 rx_priority_to_traffic_class; + CLIB_PAD_FROM_TO (0x3024, 0x3028); + u32 rx_coallesce_data_buffer_control; + CLIB_PAD_FROM_TO (0x302c, 0x3190); + u32 rx_packet_buffer_flush_detect; + CLIB_PAD_FROM_TO (0x3194, 0x3200); + u32 flow_control_tx_timers[4]; /* 2 timer values */ + CLIB_PAD_FROM_TO (0x3210, 0x3220); + u32 flow_control_rx_threshold_lo[8]; + CLIB_PAD_FROM_TO (0x3240, 0x3260); + u32 flow_control_rx_threshold_hi[8]; + CLIB_PAD_FROM_TO (0x3280, 0x32a0); + u32 flow_control_refresh_threshold; + CLIB_PAD_FROM_TO (0x32a4, 0x3c00); + /* For each of 8 traffic classes (units of bytes). */ + u32 rx_packet_buffer_size[8]; + CLIB_PAD_FROM_TO (0x3c20, 0x3d00); + u32 flow_control_config; + CLIB_PAD_FROM_TO (0x3d04, 0x4200); + + struct + { + u32 pcs_config; + CLIB_PAD_FROM_TO (0x4204, 0x4208); + u32 link_control; + u32 link_status; + u32 pcs_debug[2]; + u32 auto_negotiation; + u32 link_partner_ability; + u32 auto_negotiation_tx_next_page; + u32 auto_negotiation_link_partner_next_page; + CLIB_PAD_FROM_TO (0x4228, 0x4240); + } gige_mac; + + struct + { + /* [0] tx crc enable + [2] enable frames up to max frame size register [31:16] + [10] pad frames < 64 bytes if specified by user + [15] loopback enable + [16] mdc hi speed + [17] turn off mdc between mdio packets */ + u32 control; + + /* [5] rx symbol error (all bits clear on read) + [6] rx illegal symbol + [7] rx idle error + [8] rx local fault + [9] rx remote fault */ + u32 status; + + u32 pause_and_pace_control; + CLIB_PAD_FROM_TO (0x424c, 0x425c); + u32 phy_command; + u32 phy_data; + CLIB_PAD_FROM_TO (0x4264, 0x4268); + + /* [31:16] max frame size in bytes. */ + u32 rx_max_frame_size; + CLIB_PAD_FROM_TO (0x426c, 0x4288); + + /* [0] + [2] pcs receive link up? (latch lo) + [7] local fault + [1] + [0] pcs 10g base r capable + [1] pcs 10g base x capable + [2] pcs 10g base w capable + [10] rx local fault + [11] tx local fault + [15:14] 2 => device present at this address (else not present) */ + u32 xgxs_status[2]; + + u32 base_x_pcs_status; + + /* [0] pass unrecognized flow control frames + [1] discard pause frames + [2] rx priority flow control enable (only in dcb mode) + [3] rx flow control enable. */ + u32 flow_control; + + /* [3:0] tx lanes change polarity + [7:4] rx lanes change polarity + [11:8] swizzle tx lanes + [15:12] swizzle rx lanes + 4 x 2 bit tx lane swap + 4 x 2 bit rx lane swap. */ + u32 serdes_control; + + u32 fifo_control; + + /* [0] force link up + [1] autoneg ack2 bit to transmit + [6:2] autoneg selector field to transmit + [8:7] 10g pma/pmd type 0 => xaui, 1 kx4, 2 cx4 + [9] 1g pma/pmd type 0 => sfi, 1 => kx/bx + [10] disable 10g on without main power + [11] restart autoneg on transition to dx power state + [12] restart autoneg + [15:13] link mode: + 0 => 1g no autoneg + 1 => 10g kx4 parallel link no autoneg + 2 => 1g bx autoneg + 3 => 10g sfi serdes + 4 => kx4/kx/kr + 5 => xgmii 1g/100m + 6 => kx4/kx/kr 1g an + 7 kx4/kx/kr sgmii. + [16] kr support + [17] fec requested + [18] fec ability + etc. */ + u32 auto_negotiation_control; + + /* [0] signal detect 1g/100m + [1] fec signal detect + [2] 10g serial pcs fec block lock + [3] 10g serial high error rate + [4] 10g serial pcs block lock + [5] kx/kx4/kr autoneg next page received + [6] kx/kx4/kr backplane autoneg next page received + [7] link status clear to read + [11:8] 10g signal detect (4 lanes) (for serial just lane 0) + [12] 10g serial signal detect + [16:13] 10g parallel lane sync status + [17] 10g parallel align status + [18] 1g sync status + [19] kx/kx4/kr backplane autoneg is idle + [20] 1g autoneg enabled + [21] 1g pcs enabled for sgmii + [22] 10g xgxs enabled + [23] 10g serial fec enabled (forward error detection) + [24] 10g kr pcs enabled + [25] sgmii enabled + [27:26] mac link mode + 0 => 1g + 1 => 10g parallel + 2 => 10g serial + 3 => autoneg + [29:28] link speed + 1 => 100m + 2 => 1g + 3 => 10g + [30] link is up + [31] kx/kx4/kr backplane autoneg completed successfully. */ + u32 link_status; + + /* [17:16] pma/pmd for 10g serial + 0 => kr, 2 => sfi + [18] disable dme pages */ + u32 auto_negotiation_control2; + + CLIB_PAD_FROM_TO (0x42ac, 0x42b0); + u32 link_partner_ability[2]; + CLIB_PAD_FROM_TO (0x42b8, 0x42d0); + u32 manageability_control; + u32 link_partner_next_page[2]; + CLIB_PAD_FROM_TO (0x42dc, 0x42e0); + u32 kr_pcs_control; + u32 kr_pcs_status; + u32 fec_status[2]; + CLIB_PAD_FROM_TO (0x42f0, 0x4314); + u32 sgmii_control; + CLIB_PAD_FROM_TO (0x4318, 0x4324); + u32 link_status2; + CLIB_PAD_FROM_TO (0x4328, 0x4900); + } xge_mac; + + u32 tx_dcb_control; + u32 tx_dcb_descriptor_plane_queue_select; + u32 tx_dcb_descriptor_plane_t1_config; + u32 tx_dcb_descriptor_plane_t1_status; + CLIB_PAD_FROM_TO (0x4910, 0x4950); + + /* For each TC in units of 1k bytes. */ + u32 tx_packet_buffer_thresholds[8]; + CLIB_PAD_FROM_TO (0x4970, 0x4980); + struct + { + u32 mmw; + u32 config; + u32 status; + u32 rate_drift; + } dcb_tx_rate_scheduler; + CLIB_PAD_FROM_TO (0x4990, 0x4a80); + u32 tx_dma_control; + CLIB_PAD_FROM_TO (0x4a84, 0x4a88); + u32 tx_dma_tcp_flags_control[2]; + CLIB_PAD_FROM_TO (0x4a90, 0x4b00); + u32 pf_mailbox[64]; + CLIB_PAD_FROM_TO (0x4c00, 0x5000); + + /* RX */ + u32 checksum_control; + CLIB_PAD_FROM_TO (0x5004, 0x5008); + u32 rx_filter_control; + CLIB_PAD_FROM_TO (0x500c, 0x5010); + u32 management_vlan_tag[8]; + u32 management_udp_tcp_ports[8]; + CLIB_PAD_FROM_TO (0x5050, 0x5078); + /* little endian. */ + u32 extended_vlan_ether_type; + CLIB_PAD_FROM_TO (0x507c, 0x5080); + /* [1] store/dma bad packets + [8] accept all multicast + [9] accept all unicast + [10] accept all broadcast. */ + u32 filter_control; + CLIB_PAD_FROM_TO (0x5084, 0x5088); + /* [15:0] vlan ethernet type (0x8100) little endian + [28] cfi bit expected + [29] drop packets with unexpected cfi bit + [30] vlan filter enable. */ + u32 vlan_control; + CLIB_PAD_FROM_TO (0x508c, 0x5090); + /* [1:0] hi bit of ethernet address for 12 bit index into multicast table + 0 => 47, 1 => 46, 2 => 45, 3 => 43. + [2] enable multicast filter + */ + u32 multicast_control; + CLIB_PAD_FROM_TO (0x5094, 0x5100); + u32 fcoe_rx_control; + CLIB_PAD_FROM_TO (0x5104, 0x5108); + u32 fc_flt_context; + CLIB_PAD_FROM_TO (0x510c, 0x5110); + u32 fc_filter_control; + CLIB_PAD_FROM_TO (0x5114, 0x5120); + u32 rx_message_type_lo; + CLIB_PAD_FROM_TO (0x5124, 0x5128); + /* [15:0] ethernet type (little endian) + [18:16] matche pri in vlan tag + [19] priority match enable + [25:20] virtualization pool + [26] pool enable + [27] is fcoe + [30] ieee 1588 timestamp enable + [31] filter enable. + (See ethernet_type_queue_select.) */ + u32 ethernet_type_queue_filter[8]; + CLIB_PAD_FROM_TO (0x5148, 0x5160); + /* [7:0] l2 ethernet type and + [15:8] l2 ethernet type or */ + u32 management_decision_filters1[8]; + u32 vf_vm_tx_switch_loopback_enable[2]; + u32 rx_time_sync_control; + CLIB_PAD_FROM_TO (0x518c, 0x5190); + u32 management_ethernet_type_filters[4]; + u32 rx_timestamp_attributes_lo; + u32 rx_timestamp_hi; + u32 rx_timestamp_attributes_hi; + CLIB_PAD_FROM_TO (0x51ac, 0x51b0); + u32 pf_virtual_control; + CLIB_PAD_FROM_TO (0x51b4, 0x51d8); + u32 fc_offset_parameter; + CLIB_PAD_FROM_TO (0x51dc, 0x51e0); + u32 vf_rx_enable[2]; + u32 rx_timestamp_lo; + CLIB_PAD_FROM_TO (0x51ec, 0x5200); + /* 12 bits determined by multicast_control + lookup bits in this vector. */ + u32 multicast_enable[128]; + + /* [0] ethernet address [31:0] + [1] [15:0] ethernet address [47:32] + [31] valid bit. + Index 0 is read from eeprom after reset. */ + u32 rx_ethernet_address0[16][2]; + + CLIB_PAD_FROM_TO (0x5480, 0x5800); + u32 wake_up_control; + CLIB_PAD_FROM_TO (0x5804, 0x5808); + u32 wake_up_filter_control; + CLIB_PAD_FROM_TO (0x580c, 0x5818); + u32 multiple_rx_queue_command_82598; + CLIB_PAD_FROM_TO (0x581c, 0x5820); + u32 management_control; + u32 management_filter_control; + CLIB_PAD_FROM_TO (0x5828, 0x5838); + u32 wake_up_ip4_address_valid; + CLIB_PAD_FROM_TO (0x583c, 0x5840); + u32 wake_up_ip4_address_table[4]; + u32 management_control_to_host; + CLIB_PAD_FROM_TO (0x5854, 0x5880); + u32 wake_up_ip6_address_table[4]; + + /* unicast_and broadcast_and vlan_and ip_address_and + etc. */ + u32 management_decision_filters[8]; + + u32 management_ip4_or_ip6_address_filters[4][4]; + CLIB_PAD_FROM_TO (0x58f0, 0x5900); + u32 wake_up_packet_length; + CLIB_PAD_FROM_TO (0x5904, 0x5910); + u32 management_ethernet_address_filters[4][2]; + CLIB_PAD_FROM_TO (0x5930, 0x5a00); + u32 wake_up_packet_memory[32]; + CLIB_PAD_FROM_TO (0x5a80, 0x5c00); + u32 redirection_table_82598[32]; + u32 rss_random_keys_82598[10]; + CLIB_PAD_FROM_TO (0x5ca8, 0x6000); + + ixge_dma_regs_t tx_dma[128]; + + u32 pf_vm_vlan_insert[64]; + u32 tx_dma_tcp_max_alloc_size_requests; + CLIB_PAD_FROM_TO (0x8104, 0x8110); + u32 vf_tx_enable[2]; + CLIB_PAD_FROM_TO (0x8118, 0x8120); + /* [0] dcb mode enable + [1] virtualization mode enable + [3:2] number of tcs/qs per pool. */ + u32 multiple_tx_queues_command; + CLIB_PAD_FROM_TO (0x8124, 0x8200); + u32 pf_vf_anti_spoof[8]; + u32 pf_dma_tx_switch_control; + CLIB_PAD_FROM_TO (0x8224, 0x82e0); + u32 tx_strict_low_latency_queues[4]; + CLIB_PAD_FROM_TO (0x82f0, 0x8600); + u32 tx_queue_stats_mapping_82599[32]; + u32 tx_queue_packet_counts[32]; + u32 tx_queue_byte_counts[32][2]; + + struct + { + u32 control; + u32 status; + u32 buffer_almost_full; + CLIB_PAD_FROM_TO (0x880c, 0x8810); + u32 buffer_min_ifg; + CLIB_PAD_FROM_TO (0x8814, 0x8900); + } tx_security; + + struct + { + u32 index; + u32 salt; + u32 key[4]; + CLIB_PAD_FROM_TO (0x8918, 0x8a00); + } tx_ipsec; + + struct + { + u32 capabilities; + u32 control; + u32 tx_sci[2]; + u32 sa; + u32 sa_pn[2]; + u32 key[2][4]; + /* untagged packets, encrypted packets, protected packets, + encrypted bytes, protected bytes */ + u32 stats[5]; + CLIB_PAD_FROM_TO (0x8a50, 0x8c00); + } tx_link_security; + + struct + { + u32 control; + u32 timestamp_value[2]; + u32 system_time[2]; + u32 increment_attributes; + u32 time_adjustment_offset[2]; + u32 aux_control; + u32 target_time[2][2]; + CLIB_PAD_FROM_TO (0x8c34, 0x8c3c); + u32 aux_time_stamp[2][2]; + CLIB_PAD_FROM_TO (0x8c4c, 0x8d00); + } tx_timesync; + + struct + { + u32 control; + u32 status; + CLIB_PAD_FROM_TO (0x8d08, 0x8e00); + } rx_security; + + struct + { + u32 index; + u32 ip_address[4]; + u32 spi; + u32 ip_index; + u32 key[4]; + u32 salt; + u32 mode; + CLIB_PAD_FROM_TO (0x8e34, 0x8f00); + } rx_ipsec; + + struct + { + u32 capabilities; + u32 control; + u32 sci[2]; + u32 sa[2]; + u32 sa_pn[2]; + u32 key[2][4]; + /* see datasheet */ + u32 stats[17]; + CLIB_PAD_FROM_TO (0x8f84, 0x9000); + } rx_link_security; + + /* 4 wake up, 2 management, 2 wake up. */ + u32 flexible_filters[8][16][4]; + CLIB_PAD_FROM_TO (0x9800, 0xa000); + + /* 4096 bits. */ + u32 vlan_filter[128]; + + /* [0] ethernet address [31:0] + [1] [15:0] ethernet address [47:32] + [31] valid bit. + Index 0 is read from eeprom after reset. */ + u32 rx_ethernet_address1[128][2]; + + /* select one of 64 pools for each rx address. */ + u32 rx_ethernet_address_pool_select[128][2]; + CLIB_PAD_FROM_TO (0xaa00, 0xc800); + u32 tx_priority_to_traffic_class; + CLIB_PAD_FROM_TO (0xc804, 0xcc00); + + /* In bytes units of 1k. Total packet buffer is 160k. */ + u32 tx_packet_buffer_size[8]; + + CLIB_PAD_FROM_TO (0xcc20, 0xcd10); + u32 tx_manageability_tc_mapping; + CLIB_PAD_FROM_TO (0xcd14, 0xcd20); + u32 dcb_tx_packet_plane_t2_config[8]; + u32 dcb_tx_packet_plane_t2_status[8]; + CLIB_PAD_FROM_TO (0xcd60, 0xce00); + + u32 tx_flow_control_status; + CLIB_PAD_FROM_TO (0xce04, 0xd000); + + ixge_dma_regs_t rx_dma1[64]; + + struct + { + /* Bigendian ip4 src/dst address. */ + u32 src_address[128]; + u32 dst_address[128]; + + /* TCP/UDP ports [15:0] src [31:16] dst; bigendian. */ + u32 tcp_udp_port[128]; + + /* [1:0] protocol tcp, udp, sctp, other + [4:2] match priority (highest wins) + [13:8] pool + [25] src address match disable + [26] dst address match disable + [27] src port match disable + [28] dst port match disable + [29] protocol match disable + [30] pool match disable + [31] enable. */ + u32 control[128]; + + /* [12] size bypass + [19:13] must be 0x80 + [20] low-latency interrupt + [27:21] rx queue. */ + u32 interrupt[128]; + } ip4_filters; + + CLIB_PAD_FROM_TO (0xea00, 0xeb00); + /* 4 bit rss output index indexed by 7 bit hash. + 128 8 bit fields = 32 registers. */ + u32 redirection_table_82599[32]; + + u32 rss_random_key_82599[10]; + CLIB_PAD_FROM_TO (0xeba8, 0xec00); + /* [15:0] reserved + [22:16] rx queue index + [29] low-latency interrupt on match + [31] enable */ + u32 ethernet_type_queue_select[8]; + CLIB_PAD_FROM_TO (0xec20, 0xec30); + u32 syn_packet_queue_filter; + CLIB_PAD_FROM_TO (0xec34, 0xec60); + u32 immediate_interrupt_rx_vlan_priority; + CLIB_PAD_FROM_TO (0xec64, 0xec70); + u32 rss_queues_per_traffic_class; + CLIB_PAD_FROM_TO (0xec74, 0xec90); + u32 lli_size_threshold; + CLIB_PAD_FROM_TO (0xec94, 0xed00); + + struct + { + u32 control; + CLIB_PAD_FROM_TO (0xed04, 0xed10); + u32 table[8]; + CLIB_PAD_FROM_TO (0xed30, 0xee00); + } fcoe_redirection; + + struct + { + /* [1:0] packet buffer allocation 0 => disabled, else 64k*2^(f-1) + [3] packet buffer initialization done + [4] perfetch match mode + [5] report status in rss field of rx descriptors + [7] report status always + [14:8] drop queue + [20:16] flex 2 byte packet offset (units of 2 bytes) + [27:24] max linked list length + [31:28] full threshold. */ + u32 control; + CLIB_PAD_FROM_TO (0xee04, 0xee0c); + + u32 data[8]; + + /* [1:0] 0 => no action, 1 => add, 2 => remove, 3 => query. + [2] valid filter found by query command + [3] filter update override + [4] ip6 adress table + [6:5] l4 protocol reserved, udp, tcp, sctp + [7] is ip6 + [8] clear head/tail + [9] packet drop action + [10] matched packet generates low-latency interrupt + [11] last in linked list + [12] collision + [15] rx queue enable + [22:16] rx queue + [29:24] pool. */ + u32 command; + + CLIB_PAD_FROM_TO (0xee30, 0xee3c); + /* ip4 dst/src address, tcp ports, udp ports. + set bits mean bit is ignored. */ + u32 ip4_masks[4]; + u32 filter_length; + u32 usage_stats; + u32 failed_usage_stats; + u32 filters_match_stats; + u32 filters_miss_stats; + CLIB_PAD_FROM_TO (0xee60, 0xee68); + /* Lookup, signature. */ + u32 hash_keys[2]; + /* [15:0] ip6 src address 1 bit per byte + [31:16] ip6 dst address. */ + u32 ip6_mask; + /* [0] vlan id + [1] vlan priority + [2] pool + [3] ip protocol + [4] flex + [5] dst ip6. */ + u32 other_mask; + CLIB_PAD_FROM_TO (0xee78, 0xf000); + } flow_director; + + struct + { + u32 l2_control[64]; + u32 vlan_pool_filter[64]; + u32 vlan_pool_filter_bitmap[128]; + u32 dst_ethernet_address[128]; + u32 mirror_rule[4]; + u32 mirror_rule_vlan[8]; + u32 mirror_rule_pool[8]; + CLIB_PAD_FROM_TO (0xf650, 0x10010); + } pf_bar; + + u32 eeprom_flash_control; + /* [0] start + [1] done + [15:2] address + [31:16] read data. */ + u32 eeprom_read; + CLIB_PAD_FROM_TO (0x10018, 0x1001c); + u32 flash_access; + CLIB_PAD_FROM_TO (0x10020, 0x10114); + u32 flash_data; + u32 flash_control; + u32 flash_read_data; + CLIB_PAD_FROM_TO (0x10120, 0x1013c); + u32 flash_opcode; + u32 software_semaphore; + CLIB_PAD_FROM_TO (0x10144, 0x10148); + u32 firmware_semaphore; + CLIB_PAD_FROM_TO (0x1014c, 0x10160); + u32 software_firmware_sync; + CLIB_PAD_FROM_TO (0x10164, 0x10200); + u32 general_rx_control; + CLIB_PAD_FROM_TO (0x10204, 0x11000); + + struct + { + u32 control; + CLIB_PAD_FROM_TO (0x11004, 0x11010); + /* [3:0] enable counters + [7:4] leaky bucket counter mode + [29] reset + [30] stop + [31] start. */ + u32 counter_control; + /* [7:0],[15:8],[23:16],[31:24] event for counters 0-3. + event codes: + 0x0 bad tlp + 0x10 reqs that reached timeout + etc. */ + u32 counter_event; + CLIB_PAD_FROM_TO (0x11018, 0x11020); + u32 counters_clear_on_read[4]; + u32 counter_config[4]; + struct + { + u32 address; + u32 data; + } indirect_access; + CLIB_PAD_FROM_TO (0x11048, 0x11050); + u32 extended_control; + CLIB_PAD_FROM_TO (0x11054, 0x11064); + u32 mirrored_revision_id; + CLIB_PAD_FROM_TO (0x11068, 0x11070); + u32 dca_requester_id_information; + + /* [0] global disable + [4:1] mode: 0 => legacy, 1 => dca 1.0. */ + u32 dca_control; + CLIB_PAD_FROM_TO (0x11078, 0x110b0); + /* [0] pci completion abort + [1] unsupported i/o address + [2] wrong byte enable + [3] pci timeout */ + u32 pcie_interrupt_status; + CLIB_PAD_FROM_TO (0x110b4, 0x110b8); + u32 pcie_interrupt_enable; + CLIB_PAD_FROM_TO (0x110bc, 0x110c0); + u32 msi_x_pba_clear[8]; + CLIB_PAD_FROM_TO (0x110e0, 0x12300); + } pcie; + + u32 interrupt_throttle1[128 - 24]; + CLIB_PAD_FROM_TO (0x124a0, 0x14f00); + + u32 core_analog_config; + CLIB_PAD_FROM_TO (0x14f04, 0x14f10); + u32 core_common_config; + CLIB_PAD_FROM_TO (0x14f14, 0x15f14); + + u32 link_sec_software_firmware_interface; +} ixge_regs_t; + +typedef union +{ + struct + { + /* Addresses bigendian. */ + union + { + struct + { + ip6_address_t src_address; + u32 unused[1]; + } ip6; + struct + { + u32 unused[3]; + ip4_address_t src_address, dst_address; + } ip4; + }; + + /* [15:0] src port (little endian). + [31:16] dst port. */ + u32 tcp_udp_ports; + + /* [15:0] vlan (cfi bit set to 0). + [31:16] flex bytes. bigendian. */ + u32 vlan_and_flex_word; + + /* [14:0] hash + [15] bucket valid + [31:16] signature (signature filers)/sw-index (perfect match). */ + u32 hash; + }; + + u32 as_u32[8]; +} ixge_flow_director_key_t; + +always_inline void +ixge_throttle_queue_interrupt (ixge_regs_t * r, + u32 queue_interrupt_index, + f64 inter_interrupt_interval_in_secs) +{ + volatile u32 *tr = + (queue_interrupt_index < ARRAY_LEN (r->interrupt.throttle0) + ? &r->interrupt.throttle0[queue_interrupt_index] + : &r->interrupt_throttle1[queue_interrupt_index]); + ASSERT (queue_interrupt_index < 128); + u32 v; + i32 i, mask = (1 << 9) - 1; + + i = flt_round_nearest (inter_interrupt_interval_in_secs / 2e-6); + i = i < 1 ? 1 : i; + i = i >= mask ? mask : i; + + v = tr[0]; + v &= ~(mask << 3); + v |= i << 3; + tr[0] = v; +} + +#define foreach_ixge_counter \ + _ (0x40d0, rx_total_packets) \ + _64 (0x40c0, rx_total_bytes) \ + _ (0x41b0, rx_good_packets_before_filtering) \ + _64 (0x41b4, rx_good_bytes_before_filtering) \ + _ (0x2f50, rx_dma_good_packets) \ + _64 (0x2f54, rx_dma_good_bytes) \ + _ (0x2f5c, rx_dma_duplicated_good_packets) \ + _64 (0x2f60, rx_dma_duplicated_good_bytes) \ + _ (0x2f68, rx_dma_good_loopback_packets) \ + _64 (0x2f6c, rx_dma_good_loopback_bytes) \ + _ (0x2f74, rx_dma_good_duplicated_loopback_packets) \ + _64 (0x2f78, rx_dma_good_duplicated_loopback_bytes) \ + _ (0x4074, rx_good_packets) \ + _64 (0x4088, rx_good_bytes) \ + _ (0x407c, rx_multicast_packets) \ + _ (0x4078, rx_broadcast_packets) \ + _ (0x405c, rx_64_byte_packets) \ + _ (0x4060, rx_65_127_byte_packets) \ + _ (0x4064, rx_128_255_byte_packets) \ + _ (0x4068, rx_256_511_byte_packets) \ + _ (0x406c, rx_512_1023_byte_packets) \ + _ (0x4070, rx_gt_1023_byte_packets) \ + _ (0x4000, rx_crc_errors) \ + _ (0x4120, rx_ip_checksum_errors) \ + _ (0x4004, rx_illegal_symbol_errors) \ + _ (0x4008, rx_error_symbol_errors) \ + _ (0x4034, rx_mac_local_faults) \ + _ (0x4038, rx_mac_remote_faults) \ + _ (0x4040, rx_length_errors) \ + _ (0x41a4, rx_xons) \ + _ (0x41a8, rx_xoffs) \ + _ (0x40a4, rx_undersize_packets) \ + _ (0x40a8, rx_fragments) \ + _ (0x40ac, rx_oversize_packets) \ + _ (0x40b0, rx_jabbers) \ + _ (0x40b4, rx_management_packets) \ + _ (0x40b8, rx_management_drops) \ + _ (0x3fa0, rx_missed_packets_pool_0) \ + _ (0x40d4, tx_total_packets) \ + _ (0x4080, tx_good_packets) \ + _64 (0x4090, tx_good_bytes) \ + _ (0x40f0, tx_multicast_packets) \ + _ (0x40f4, tx_broadcast_packets) \ + _ (0x87a0, tx_dma_good_packets) \ + _64 (0x87a4, tx_dma_good_bytes) \ + _ (0x40d8, tx_64_byte_packets) \ + _ (0x40dc, tx_65_127_byte_packets) \ + _ (0x40e0, tx_128_255_byte_packets) \ + _ (0x40e4, tx_256_511_byte_packets) \ + _ (0x40e8, tx_512_1023_byte_packets) \ + _ (0x40ec, tx_gt_1023_byte_packets) \ + _ (0x4010, tx_undersize_drops) \ + _ (0x8780, switch_security_violation_packets) \ + _ (0x5118, fc_crc_errors) \ + _ (0x241c, fc_rx_drops) \ + _ (0x2424, fc_last_error_count) \ + _ (0x2428, fcoe_rx_packets) \ + _ (0x242c, fcoe_rx_dwords) \ + _ (0x8784, fcoe_tx_packets) \ + _ (0x8788, fcoe_tx_dwords) \ + _ (0x1030, queue_0_rx_count) \ + _ (0x1430, queue_0_drop_count) \ + _ (0x1070, queue_1_rx_count) \ + _ (0x1470, queue_1_drop_count) \ + _ (0x10b0, queue_2_rx_count) \ + _ (0x14b0, queue_2_drop_count) \ + _ (0x10f0, queue_3_rx_count) \ + _ (0x14f0, queue_3_drop_count) \ + _ (0x1130, queue_4_rx_count) \ + _ (0x1530, queue_4_drop_count) \ + _ (0x1170, queue_5_rx_count) \ + _ (0x1570, queue_5_drop_count) \ + _ (0x11b0, queue_6_rx_count) \ + _ (0x15b0, queue_6_drop_count) \ + _ (0x11f0, queue_7_rx_count) \ + _ (0x15f0, queue_7_drop_count) \ + _ (0x1230, queue_8_rx_count) \ + _ (0x1630, queue_8_drop_count) \ + _ (0x1270, queue_9_rx_count) \ + _ (0x1270, queue_9_drop_count) + + + + +typedef enum +{ +#define _(a,f) IXGE_COUNTER_##f, +#define _64(a,f) _(a,f) + foreach_ixge_counter +#undef _ +#undef _64 + IXGE_N_COUNTER, +} ixge_counter_type_t; + +typedef struct +{ + u32 mdio_address; + + /* 32 bit ID read from ID registers. */ + u32 id; +} ixge_phy_t; + +typedef struct +{ + /* Cache aligned descriptors. */ + ixge_descriptor_t *descriptors; + + /* Number of descriptors in table. */ + u32 n_descriptors; + + /* Software head and tail pointers into descriptor ring. */ + u32 head_index, tail_index; + + /* Index into dma_queues vector. */ + u32 queue_index; + + /* Buffer indices corresponding to each active descriptor. */ + u32 *descriptor_buffer_indices; + + union + { + struct + { + u32 *volatile head_index_write_back; + + u32 n_buffers_on_ring; + } tx; + + struct + { + /* Buffer indices to use to replenish each descriptor. */ + u32 *replenish_buffer_indices; + + vlib_node_runtime_t *node; + u32 next_index; + + u32 saved_start_of_packet_buffer_index; + + u32 saved_start_of_packet_next_index; + u32 saved_last_buffer_index; + + u32 is_start_of_packet; + + u32 n_descriptors_done_total; + + u32 n_descriptors_done_this_call; + + u32 n_bytes; + } rx; + }; +} ixge_dma_queue_t; + +#define foreach_ixge_pci_device_id \ + _ (82598, 0x10b6) \ + _ (82598_bx, 0x1508) \ + _ (82598af_dual_port, 0x10c6) \ + _ (82598af_single_port, 0x10c7) \ + _ (82598at, 0x10c8) \ + _ (82598at2, 0x150b) \ + _ (82598eb_sfp_lom, 0x10db) \ + _ (82598eb_cx4, 0x10dd) \ + _ (82598_cx4_dual_port, 0x10ec) \ + _ (82598_da_dual_port, 0x10f1) \ + _ (82598_sr_dual_port_em, 0x10e1) \ + _ (82598eb_xf_lr, 0x10f4) \ + _ (82599_kx4, 0x10f7) \ + _ (82599_kx4_mezz, 0x1514) \ + _ (82599_kr, 0x1517) \ + _ (82599_combo_backplane, 0x10f8) \ + _ (82599_cx4, 0x10f9) \ + _ (82599_sfp, 0x10fb) \ + _ (82599_backplane_fcoe, 0x152a) \ + _ (82599_sfp_fcoe, 0x1529) \ + _ (82599_sfp_em, 0x1507) \ + _ (82599_xaui_lom, 0x10fc) \ + _ (82599_t3_lom, 0x151c) \ + _ (x540t, 0x1528) + +typedef enum +{ +#define _(f,n) IXGE_##f = n, + foreach_ixge_pci_device_id +#undef _ +} ixge_pci_device_id_t; + +typedef struct +{ + /* registers */ + ixge_regs_t *regs; + + /* Specific next index when using dynamic redirection */ + u32 per_interface_next_index; + + /* PCI bus info. */ + vlib_pci_device_t pci_device; + + /* From PCI config space header. */ + ixge_pci_device_id_t device_id; + + u16 device_index; + + /* 0 or 1. */ + u16 pci_function; + + /* VLIB interface for this instance. */ + u32 vlib_hw_if_index, vlib_sw_if_index; + + ixge_dma_queue_t *dma_queues[VLIB_N_RX_TX]; + + /* Phy index (0 or 1) and address on MDI bus. */ + u32 phy_index; + ixge_phy_t phys[2]; + + /* Value of link_status register at last link change. */ + u32 link_status_at_last_link_change; + + i2c_bus_t i2c_bus; + sfp_eeprom_t sfp_eeprom; + + /* Counters. */ + u64 counters[IXGE_N_COUNTER], counters_last_clear[IXGE_N_COUNTER]; +} ixge_device_t; + +typedef struct +{ + vlib_main_t *vlib_main; + + /* Vector of devices. */ + ixge_device_t *devices; + + /* Descriptor ring sizes. */ + u32 n_descriptors[VLIB_N_RX_TX]; + + /* RX buffer size. Must be at least 1k; will be rounded to + next largest 1k size. */ + u32 n_bytes_in_rx_buffer; + + u32 n_descriptors_per_cache_line; + + u32 vlib_buffer_free_list_index; + + u32 process_node_index; + + /* Template and mask for initializing/validating TX descriptors. */ + ixge_tx_descriptor_t tx_descriptor_template, tx_descriptor_template_mask; + + /* Vector of buffers for which TX is done and can be freed. */ + u32 *tx_buffers_pending_free; + + u32 *rx_buffers_to_add; + + f64 time_last_stats_update; +} ixge_main_t; + +ixge_main_t ixge_main; +vnet_device_class_t ixge_device_class; + +typedef enum +{ + IXGE_RX_NEXT_IP4_INPUT, + IXGE_RX_NEXT_IP6_INPUT, + IXGE_RX_NEXT_ETHERNET_INPUT, + IXGE_RX_NEXT_DROP, + IXGE_RX_N_NEXT, +} ixge_rx_next_t; + +void ixge_set_next_node (ixge_rx_next_t, char *); + +#endif /* included_ixge_h */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/nic/sfp.c b/src/vnet/devices/nic/sfp.c new file mode 100644 index 00000000000..9e9c008dc15 --- /dev/null +++ b/src/vnet/devices/nic/sfp.c @@ -0,0 +1,117 @@ +/* + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include + +static u8 * +format_space_terminated (u8 * s, va_list * args) +{ + u32 l = va_arg (*args, u32); + u8 *v = va_arg (*args, u8 *); + u8 *p; + + for (p = v + l - 1; p >= v && p[0] == ' '; p--) + ; + vec_add (s, v, clib_min (p - v + 1, l)); + return s; +} + +static u8 * +format_sfp_id (u8 * s, va_list * args) +{ + u32 id = va_arg (*args, u32); + char *t = 0; + switch (id) + { +#define _(f) case SFP_ID_##f: t = #f; break; + foreach_sfp_id +#undef _ + default: + return format (s, "unknown 0x%x", id); + } + return format (s, "%s", t); +} + +static u8 * +format_sfp_compatibility (u8 * s, va_list * args) +{ + u32 c = va_arg (*args, u32); + char *t = 0; + switch (c) + { +#define _(a,b,f) case SFP_COMPATIBILITY_##f: t = #f; break; + foreach_sfp_compatibility +#undef _ + default: + return format (s, "unknown 0x%x", c); + } + return format (s, "%s", t); +} + +u32 +sfp_is_comatible (sfp_eeprom_t * e, sfp_compatibility_t c) +{ + static struct + { + u8 byte, bit; + } t[] = + { +#define _(a,b,f) { .byte = a, .bit = b, }, + foreach_sfp_compatibility +#undef _ + }; + + ASSERT (c < ARRAY_LEN (t)); + return (e->compatibility[t[c].byte] & (1 << t[c].bit)) != 0; +} + +u8 * +format_sfp_eeprom (u8 * s, va_list * args) +{ + sfp_eeprom_t *e = va_arg (*args, sfp_eeprom_t *); + uword indent = format_get_indent (s); + int i; + + if (e->id != SFP_ID_sfp) + s = format (s, "id %U, ", format_sfp_id, e->id); + + s = format (s, "compatibility:"); + for (i = 0; i < SFP_N_COMPATIBILITY; i++) + if (sfp_is_comatible (e, i)) + s = format (s, " %U", format_sfp_compatibility, i); + + s = format (s, "\n%Uvendor: %U, part %U", + format_white_space, indent, + format_space_terminated, sizeof (e->vendor_name), + e->vendor_name, format_space_terminated, + sizeof (e->vendor_part_number), e->vendor_part_number); + s = + format (s, "\n%Urevision: %U, serial: %U, date code: %U", + format_white_space, indent, format_space_terminated, + sizeof (e->vendor_revision), e->vendor_revision, + format_space_terminated, sizeof (e->vendor_serial_number), + e->vendor_serial_number, format_space_terminated, + sizeof (e->vendor_date_code), e->vendor_date_code); + + return s; +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/nic/sfp.h b/src/vnet/devices/nic/sfp.h new file mode 100644 index 00000000000..a1ac7997a44 --- /dev/null +++ b/src/vnet/devices/nic/sfp.h @@ -0,0 +1,117 @@ +/* + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef included_vnet_optics_sfp_h +#define included_vnet_optics_sfp_h + +#include + +#define foreach_sfp_id \ + _ (unknown) \ + _ (gbic) \ + _ (on_motherboard) \ + _ (sfp) + +typedef enum +{ +#define _(f) SFP_ID_##f, + foreach_sfp_id +#undef _ +} sfp_id_t; + +typedef struct +{ + u8 id; + u8 extended_id; + u8 connector_type; + u8 compatibility[8]; + u8 encoding; + u8 nominal_bit_rate_100mbits_per_sec; + u8 reserved13; + u8 link_length[5]; + u8 reserved19; + u8 vendor_name[16]; + u8 reserved36; + u8 vendor_oui[3]; + u8 vendor_part_number[16]; + u8 vendor_revision[4]; + /* 16 bit value network byte order. */ + u8 laser_wavelength_in_nm[2]; + u8 reserved62; + u8 checksum_0_to_62; + + u8 options[2]; + u8 max_bit_rate_margin_percent; + u8 min_bit_rate_margin_percent; + u8 vendor_serial_number[16]; + u8 vendor_date_code[8]; + u8 reserved92[3]; + u8 checksum_63_to_94; + u8 vendor_specific[32]; + u8 reserved128[384]; + + /* Vendor specific data follows. */ + u8 vendor_specific1[0]; +} sfp_eeprom_t; + +always_inline uword +sfp_eeprom_is_valid (sfp_eeprom_t * e) +{ + int i; + u8 sum = 0; + for (i = 0; i < 63; i++) + sum += ((u8 *) e)[i]; + return sum == e->checksum_0_to_62; +} + +/* _ (byte_index, bit_index, name) */ +#define foreach_sfp_compatibility \ + _ (0, 4, 10g_base_sr) \ + _ (0, 5, 10g_base_lr) \ + _ (1, 2, oc48_long_reach) \ + _ (1, 1, oc48_intermediate_reach) \ + _ (1, 0, oc48_short_reach) \ + _ (2, 6, oc12_long_reach) \ + _ (2, 5, oc12_intermediate_reach) \ + _ (2, 4, oc12_short_reach) \ + _ (2, 2, oc3_long_reach) \ + _ (2, 1, oc3_intermediate_reach) \ + _ (2, 0, oc3_short_reach) \ + _ (3, 3, 1g_base_t) \ + _ (3, 2, 1g_base_cx) \ + _ (3, 1, 1g_base_lx) \ + _ (3, 0, 1g_base_sx) + +typedef enum +{ +#define _(a,b,f) SFP_COMPATIBILITY_##f, + foreach_sfp_compatibility +#undef _ + SFP_N_COMPATIBILITY, +} sfp_compatibility_t; + +u32 sfp_is_comatible (sfp_eeprom_t * e, sfp_compatibility_t c); + +format_function_t format_sfp_eeprom; + +#endif /* included_vnet_optics_sfp_h */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/ssvm/node.c b/src/vnet/devices/ssvm/node.c new file mode 100644 index 00000000000..3a695b1d8c0 --- /dev/null +++ b/src/vnet/devices/ssvm/node.c @@ -0,0 +1,343 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include "ssvm_eth.h" + +vlib_node_registration_t ssvm_eth_input_node; + +typedef struct +{ + u32 next_index; + u32 sw_if_index; +} ssvm_eth_input_trace_t; + +/* packet trace format function */ +static u8 * +format_ssvm_eth_input_trace (u8 * s, va_list * args) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *); + ssvm_eth_input_trace_t *t = va_arg (*args, ssvm_eth_input_trace_t *); + + s = format (s, "SSVM_ETH_INPUT: sw_if_index %d, next index %d", + t->sw_if_index, t->next_index); + return s; +} + +vlib_node_registration_t ssvm_eth_input_node; + +#define foreach_ssvm_eth_input_error \ +_(NO_BUFFERS, "Rx packet drops (no buffers)") + +typedef enum +{ +#define _(sym,str) SSVM_ETH_INPUT_ERROR_##sym, + foreach_ssvm_eth_input_error +#undef _ + SSVM_ETH_INPUT_N_ERROR, +} ssvm_eth_input_error_t; + +static char *ssvm_eth_input_error_strings[] = { +#define _(sym,string) string, + foreach_ssvm_eth_input_error +#undef _ +}; + +typedef enum +{ + SSVM_ETH_INPUT_NEXT_DROP, + SSVM_ETH_INPUT_NEXT_ETHERNET_INPUT, + SSVM_ETH_INPUT_NEXT_IP4_INPUT, + SSVM_ETH_INPUT_NEXT_IP6_INPUT, + SSVM_ETH_INPUT_NEXT_MPLS_INPUT, + SSVM_ETH_INPUT_N_NEXT, +} ssvm_eth_input_next_t; + +static inline uword +ssvm_eth_device_input (ssvm_eth_main_t * em, + ssvm_private_t * intfc, vlib_node_runtime_t * node) +{ + ssvm_shared_header_t *sh = intfc->sh; + vlib_main_t *vm = em->vlib_main; + unix_shared_memory_queue_t *q; + ssvm_eth_queue_elt_t *elt, *elts; + u32 elt_index; + u32 my_pid = intfc->my_pid; + int rx_queue_index; + u32 n_to_alloc = VLIB_FRAME_SIZE * 2; + u32 n_allocated, n_present_in_cache; + u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; + vlib_buffer_free_list_t *fl; + u32 n_left_to_next, *to_next; + u32 next0; + u32 n_buffers; + u32 n_available; + u32 bi0, saved_bi0; + vlib_buffer_t *b0, *prev; + u32 saved_cache_size = 0; + ethernet_header_t *eh0; + u16 type0; + u32 n_rx_bytes = 0, l3_offset0; + u32 cpu_index = os_get_cpu_number (); + u32 trace_cnt __attribute__ ((unused)) = vlib_get_trace_count (vm, node); + volatile u32 *lock; + u32 *elt_indices; + uword n_trace = vlib_get_trace_count (vm, node); + + /* Either side down? buh-bye... */ + if (pointer_to_uword (sh->opaque[MASTER_ADMIN_STATE_INDEX]) == 0 || + pointer_to_uword (sh->opaque[SLAVE_ADMIN_STATE_INDEX]) == 0) + return 0; + + if (intfc->i_am_master) + q = (unix_shared_memory_queue_t *) (sh->opaque[TO_MASTER_Q_INDEX]); + else + q = (unix_shared_memory_queue_t *) (sh->opaque[TO_SLAVE_Q_INDEX]); + + /* Nothing to do? */ + if (q->cursize == 0) + return 0; + + fl = vlib_buffer_get_free_list (vm, VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); + + vec_reset_length (intfc->rx_queue); + + lock = (u32 *) q; + while (__sync_lock_test_and_set (lock, 1)) + ; + while (q->cursize > 0) + { + unix_shared_memory_queue_sub_raw (q, (u8 *) & elt_index); + ASSERT (elt_index < 2048); + vec_add1 (intfc->rx_queue, elt_index); + } + CLIB_MEMORY_BARRIER (); + *lock = 0; + + n_present_in_cache = vec_len (em->buffer_cache); + + if (vec_len (em->buffer_cache) < vec_len (intfc->rx_queue) * 2) + { + vec_validate (em->buffer_cache, + n_to_alloc + vec_len (em->buffer_cache) - 1); + n_allocated = + vlib_buffer_alloc (vm, &em->buffer_cache[n_present_in_cache], + n_to_alloc); + + n_present_in_cache += n_allocated; + _vec_len (em->buffer_cache) = n_present_in_cache; + } + + elts = (ssvm_eth_queue_elt_t *) (sh->opaque[CHUNK_POOL_INDEX]); + + n_buffers = vec_len (intfc->rx_queue); + rx_queue_index = 0; + + while (n_buffers > 0) + { + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_buffers > 0 && n_left_to_next > 0) + { + elt = elts + intfc->rx_queue[rx_queue_index]; + + saved_cache_size = n_present_in_cache; + if (PREDICT_FALSE (saved_cache_size == 0)) + { + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + goto out; + } + saved_bi0 = bi0 = em->buffer_cache[--n_present_in_cache]; + b0 = vlib_get_buffer (vm, bi0); + prev = 0; + + while (1) + { + vlib_buffer_init_for_free_list (b0, fl); + + b0->current_data = elt->current_data_hint; + b0->current_length = elt->length_this_buffer; + b0->total_length_not_including_first_buffer = + elt->total_length_not_including_first_buffer; + + clib_memcpy (b0->data + b0->current_data, elt->data, + b0->current_length); + + if (PREDICT_FALSE (prev != 0)) + prev->next_buffer = bi0; + + if (PREDICT_FALSE (elt->flags & SSVM_BUFFER_NEXT_PRESENT)) + { + prev = b0; + if (PREDICT_FALSE (n_present_in_cache == 0)) + { + vlib_put_next_frame (vm, node, next_index, + n_left_to_next); + goto out; + } + bi0 = em->buffer_cache[--n_present_in_cache]; + b0 = vlib_get_buffer (vm, bi0); + } + else + break; + } + + saved_cache_size = n_present_in_cache; + + to_next[0] = saved_bi0; + to_next++; + n_left_to_next--; + + b0 = vlib_get_buffer (vm, saved_bi0); + eh0 = vlib_buffer_get_current (b0); + + type0 = clib_net_to_host_u16 (eh0->type); + + next0 = SSVM_ETH_INPUT_NEXT_ETHERNET_INPUT; + + if (type0 == ETHERNET_TYPE_IP4) + next0 = SSVM_ETH_INPUT_NEXT_IP4_INPUT; + else if (type0 == ETHERNET_TYPE_IP6) + next0 = SSVM_ETH_INPUT_NEXT_IP6_INPUT; + else if (type0 == ETHERNET_TYPE_MPLS_UNICAST) + next0 = SSVM_ETH_INPUT_NEXT_MPLS_INPUT; + + l3_offset0 = ((next0 == SSVM_ETH_INPUT_NEXT_IP4_INPUT || + next0 == SSVM_ETH_INPUT_NEXT_IP6_INPUT || + next0 == SSVM_ETH_INPUT_NEXT_MPLS_INPUT) ? + sizeof (ethernet_header_t) : 0); + + n_rx_bytes += b0->current_length + + b0->total_length_not_including_first_buffer; + + b0->current_data += l3_offset0; + b0->current_length -= l3_offset0; + b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID; + + vnet_buffer (b0)->sw_if_index[VLIB_RX] = intfc->vlib_hw_if_index; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + + /* + * Turn this on if you run into + * "bad monkey" contexts, and you want to know exactly + * which nodes they've visited... See main.c... + */ + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); + + if (PREDICT_FALSE (n_trace > 0)) + { + ssvm_eth_input_trace_t *tr; + + vlib_trace_buffer (vm, node, next0, b0, /* follow_chain */ 1); + vlib_set_trace_count (vm, node, --n_trace); + + tr = vlib_add_trace (vm, node, b0, sizeof (*tr)); + + tr->next_index = next0; + tr->sw_if_index = intfc->vlib_hw_if_index; + } + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, + bi0, next0); + n_buffers--; + rx_queue_index++; + } + + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + +out: + if (em->buffer_cache) + _vec_len (em->buffer_cache) = saved_cache_size; + else + ASSERT (saved_cache_size == 0); + + ssvm_lock (sh, my_pid, 2); + + ASSERT (vec_len (intfc->rx_queue) > 0); + + n_available = (u32) pointer_to_uword (sh->opaque[CHUNK_POOL_NFREE]); + elt_indices = (u32 *) (sh->opaque[CHUNK_POOL_FREELIST_INDEX]); + + clib_memcpy (&elt_indices[n_available], intfc->rx_queue, + vec_len (intfc->rx_queue) * sizeof (u32)); + + n_available += vec_len (intfc->rx_queue); + sh->opaque[CHUNK_POOL_NFREE] = uword_to_pointer (n_available, void *); + + ssvm_unlock (sh); + + vlib_error_count (vm, node->node_index, SSVM_ETH_INPUT_ERROR_NO_BUFFERS, + n_buffers); + + vlib_increment_combined_counter + (vnet_get_main ()->interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, cpu_index, + intfc->vlib_hw_if_index, rx_queue_index, n_rx_bytes); + + return rx_queue_index; +} + +static uword +ssvm_eth_input_node_fn (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * frame) +{ + ssvm_eth_main_t *em = &ssvm_eth_main; + ssvm_private_t *intfc; + uword n_rx_packets = 0; + + vec_foreach (intfc, em->intfcs) + { + n_rx_packets += ssvm_eth_device_input (em, intfc, node); + } + + return n_rx_packets; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (ssvm_eth_input_node) = { + .function = ssvm_eth_input_node_fn, + .name = "ssvm_eth_input", + .vector_size = sizeof (u32), + .format_trace = format_ssvm_eth_input_trace, + .type = VLIB_NODE_TYPE_INPUT, + .state = VLIB_NODE_STATE_DISABLED, + + .n_errors = ARRAY_LEN(ssvm_eth_input_error_strings), + .error_strings = ssvm_eth_input_error_strings, + + .n_next_nodes = SSVM_ETH_INPUT_N_NEXT, + + /* edit / add dispositions here */ + .next_nodes = { + [SSVM_ETH_INPUT_NEXT_DROP] = "error-drop", + [SSVM_ETH_INPUT_NEXT_ETHERNET_INPUT] = "ethernet-input", + [SSVM_ETH_INPUT_NEXT_IP4_INPUT] = "ip4-input", + [SSVM_ETH_INPUT_NEXT_IP6_INPUT] = "ip6-input", + [SSVM_ETH_INPUT_NEXT_MPLS_INPUT] = "mpls-input", + }, +}; + +VLIB_NODE_FUNCTION_MULTIARCH (ssvm_eth_input_node, ssvm_eth_input_node_fn) +/* *INDENT-ON* */ + + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/ssvm/ssvm_eth.c b/src/vnet/devices/ssvm/ssvm_eth.c new file mode 100644 index 00000000000..db4fafa9a14 --- /dev/null +++ b/src/vnet/devices/ssvm/ssvm_eth.c @@ -0,0 +1,491 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include "ssvm_eth.h" + +ssvm_eth_main_t ssvm_eth_main; + +#define foreach_ssvm_eth_tx_func_error \ +_(RING_FULL, "Tx packet drops (ring full)") \ +_(NO_BUFFERS, "Tx packet drops (no buffers)") \ +_(ADMIN_DOWN, "Tx packet drops (admin down)") + +typedef enum +{ +#define _(f,s) SSVM_ETH_TX_ERROR_##f, + foreach_ssvm_eth_tx_func_error +#undef _ + SSVM_ETH_TX_N_ERROR, +} ssvm_eth_tx_func_error_t; + +static u32 ssvm_eth_flag_change (vnet_main_t * vnm, + vnet_hw_interface_t * hi, u32 flags); + +int +ssvm_eth_create (ssvm_eth_main_t * em, u8 * name, int is_master) +{ + ssvm_private_t *intfc; + void *oldheap; + clib_error_t *e; + unix_shared_memory_queue_t *q; + ssvm_shared_header_t *sh; + ssvm_eth_queue_elt_t *elts; + u32 *elt_indices; + u8 enet_addr[6]; + int i, rv; + + vec_add2 (em->intfcs, intfc, 1); + + intfc->ssvm_size = em->segment_size; + intfc->i_am_master = 1; + intfc->name = name; + intfc->my_pid = getpid (); + if (is_master == 0) + { + rv = ssvm_slave_init (intfc, 20 /* timeout in seconds */ ); + if (rv < 0) + return rv; + goto create_vnet_interface; + } + + intfc->requested_va = em->next_base_va; + em->next_base_va += em->segment_size; + rv = ssvm_master_init (intfc, intfc - em->intfcs /* master index */ ); + + if (rv < 0) + return rv; + + /* OK, segment created, set up queues and so forth. */ + + sh = intfc->sh; + oldheap = ssvm_push_heap (sh); + + q = unix_shared_memory_queue_init (em->queue_elts, sizeof (u32), + 0 /* consumer pid not interesting */ , + 0 /* signal not sent */ ); + sh->opaque[TO_MASTER_Q_INDEX] = (void *) q; + q = unix_shared_memory_queue_init (em->queue_elts, sizeof (u32), + 0 /* consumer pid not interesting */ , + 0 /* signal not sent */ ); + sh->opaque[TO_SLAVE_Q_INDEX] = (void *) q; + + /* + * Preallocate the requested number of buffer chunks + * There must be a better way to do this, etc. + * Add some slop to avoid pool reallocation, which will not go well + */ + elts = 0; + elt_indices = 0; + + vec_validate_aligned (elts, em->nbuffers - 1, CLIB_CACHE_LINE_BYTES); + vec_validate_aligned (elt_indices, em->nbuffers - 1, CLIB_CACHE_LINE_BYTES); + + for (i = 0; i < em->nbuffers; i++) + elt_indices[i] = i; + + sh->opaque[CHUNK_POOL_INDEX] = (void *) elts; + sh->opaque[CHUNK_POOL_FREELIST_INDEX] = (void *) elt_indices; + sh->opaque[CHUNK_POOL_NFREE] = (void *) (uword) em->nbuffers; + + ssvm_pop_heap (oldheap); + +create_vnet_interface: + + sh = intfc->sh; + + memset (enet_addr, 0, sizeof (enet_addr)); + enet_addr[0] = 2; + enet_addr[1] = 0xFE; + enet_addr[2] = is_master; + enet_addr[5] = sh->master_index; + + e = ethernet_register_interface + (em->vnet_main, ssvm_eth_device_class.index, intfc - em->intfcs, + /* ethernet address */ enet_addr, + &intfc->vlib_hw_if_index, ssvm_eth_flag_change); + + if (e) + { + clib_error_report (e); + /* $$$$ unmap offending region? */ + return VNET_API_ERROR_INVALID_INTERFACE; + } + + /* Declare link up */ + vnet_hw_interface_set_flags (em->vnet_main, intfc->vlib_hw_if_index, + VNET_HW_INTERFACE_FLAG_LINK_UP); + + /* Let the games begin... */ + if (is_master) + sh->ready = 1; + return 0; +} + +static clib_error_t * +ssvm_config (vlib_main_t * vm, unformat_input_t * input) +{ + u8 *name; + int is_master = 1; + int i, rv; + ssvm_eth_main_t *em = &ssvm_eth_main; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "base-va %llx", &em->next_base_va)) + ; + else if (unformat (input, "segment-size %lld", &em->segment_size)) + em->segment_size = 1ULL << (max_log2 (em->segment_size)); + else if (unformat (input, "nbuffers %lld", &em->nbuffers)) + ; + else if (unformat (input, "queue-elts %lld", &em->queue_elts)) + ; + else if (unformat (input, "slave")) + is_master = 0; + else if (unformat (input, "%s", &name)) + vec_add1 (em->names, name); + else + break; + } + + /* No configured instances, we're done... */ + if (vec_len (em->names) == 0) + return 0; + + for (i = 0; i < vec_len (em->names); i++) + { + rv = ssvm_eth_create (em, em->names[i], is_master); + if (rv < 0) + return clib_error_return (0, "ssvm_eth_create '%s' failed, error %d", + em->names[i], rv); + } + + vlib_node_set_state (vm, ssvm_eth_input_node.index, + VLIB_NODE_STATE_POLLING); + + return 0; +} + +VLIB_CONFIG_FUNCTION (ssvm_config, "ssvm_eth"); + + +static clib_error_t * +ssvm_eth_init (vlib_main_t * vm) +{ + ssvm_eth_main_t *em = &ssvm_eth_main; + + if (((sizeof (ssvm_eth_queue_elt_t) / CLIB_CACHE_LINE_BYTES) + * CLIB_CACHE_LINE_BYTES) != sizeof (ssvm_eth_queue_elt_t)) + clib_warning ("ssvm_eth_queue_elt_t size %d not a multiple of %d", + sizeof (ssvm_eth_queue_elt_t), CLIB_CACHE_LINE_BYTES); + + em->vlib_main = vm; + em->vnet_main = vnet_get_main (); + em->elog_main = &vm->elog_main; + + /* default config param values... */ + + em->next_base_va = 0x600000000ULL; + /* + * Allocate 2 full superframes in each dir (256 x 2 x 2 x 2048 bytes), + * 2mb; double that so we have plenty of space... 4mb + */ + em->segment_size = 8 << 20; + em->nbuffers = 1024; + em->queue_elts = 512; + return 0; +} + +VLIB_INIT_FUNCTION (ssvm_eth_init); + +static char *ssvm_eth_tx_func_error_strings[] = { +#define _(n,s) s, + foreach_ssvm_eth_tx_func_error +#undef _ +}; + +static u8 * +format_ssvm_eth_device_name (u8 * s, va_list * args) +{ + u32 i = va_arg (*args, u32); + + s = format (s, "ssvmEthernet%d", i); + return s; +} + +static u8 * +format_ssvm_eth_device (u8 * s, va_list * args) +{ + s = format (s, "SSVM Ethernet"); + return s; +} + +static u8 * +format_ssvm_eth_tx_trace (u8 * s, va_list * args) +{ + s = format (s, "Unimplemented..."); + return s; +} + + +static uword +ssvm_eth_interface_tx (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * f) +{ + ssvm_eth_main_t *em = &ssvm_eth_main; + vnet_interface_output_runtime_t *rd = (void *) node->runtime_data; + ssvm_private_t *intfc = vec_elt_at_index (em->intfcs, rd->dev_instance); + ssvm_shared_header_t *sh = intfc->sh; + unix_shared_memory_queue_t *q; + u32 *from; + u32 n_left; + ssvm_eth_queue_elt_t *elts, *elt, *prev_elt; + u32 my_pid = intfc->my_pid; + vlib_buffer_t *b0; + u32 bi0; + u32 size_this_buffer; + u32 chunks_this_buffer; + u8 i_am_master = intfc->i_am_master; + u32 elt_index; + int is_ring_full, interface_down; + int i; + volatile u32 *queue_lock; + u32 n_to_alloc = VLIB_FRAME_SIZE; + u32 n_allocated, n_present_in_cache, n_available; + u32 *elt_indices; + + if (i_am_master) + q = (unix_shared_memory_queue_t *) sh->opaque[TO_SLAVE_Q_INDEX]; + else + q = (unix_shared_memory_queue_t *) sh->opaque[TO_MASTER_Q_INDEX]; + + queue_lock = (u32 *) q; + + from = vlib_frame_vector_args (f); + n_left = f->n_vectors; + is_ring_full = 0; + interface_down = 0; + + n_present_in_cache = vec_len (em->chunk_cache); + + /* admin / link up/down check */ + if (sh->opaque[MASTER_ADMIN_STATE_INDEX] == 0 || + sh->opaque[SLAVE_ADMIN_STATE_INDEX] == 0) + { + interface_down = 1; + goto out; + } + + ssvm_lock (sh, my_pid, 1); + + elts = (ssvm_eth_queue_elt_t *) (sh->opaque[CHUNK_POOL_INDEX]); + elt_indices = (u32 *) (sh->opaque[CHUNK_POOL_FREELIST_INDEX]); + n_available = (u32) pointer_to_uword (sh->opaque[CHUNK_POOL_NFREE]); + + if (n_present_in_cache < n_left * 2) + { + vec_validate (em->chunk_cache, n_to_alloc + n_present_in_cache - 1); + + n_allocated = n_to_alloc < n_available ? n_to_alloc : n_available; + + if (PREDICT_TRUE (n_allocated > 0)) + { + clib_memcpy (&em->chunk_cache[n_present_in_cache], + &elt_indices[n_available - n_allocated], + sizeof (u32) * n_allocated); + } + + n_present_in_cache += n_allocated; + n_available -= n_allocated; + sh->opaque[CHUNK_POOL_NFREE] = uword_to_pointer (n_available, void *); + _vec_len (em->chunk_cache) = n_present_in_cache; + } + + ssvm_unlock (sh); + + while (n_left) + { + bi0 = from[0]; + b0 = vlib_get_buffer (vm, bi0); + + size_this_buffer = vlib_buffer_length_in_chain (vm, b0); + chunks_this_buffer = (size_this_buffer + (SSVM_BUFFER_SIZE - 1)) + / SSVM_BUFFER_SIZE; + + /* If we're not going to be able to enqueue the buffer, tail drop. */ + if (q->cursize >= q->maxsize) + { + is_ring_full = 1; + break; + } + + prev_elt = 0; + elt_index = ~0; + for (i = 0; i < chunks_this_buffer; i++) + { + if (PREDICT_FALSE (n_present_in_cache == 0)) + goto out; + + elt_index = em->chunk_cache[--n_present_in_cache]; + elt = elts + elt_index; + + elt->type = SSVM_PACKET_TYPE; + elt->flags = 0; + elt->total_length_not_including_first_buffer = + b0->total_length_not_including_first_buffer; + elt->length_this_buffer = b0->current_length; + elt->current_data_hint = b0->current_data; + elt->owner = !i_am_master; + elt->tag = 1; + + clib_memcpy (elt->data, b0->data + b0->current_data, + b0->current_length); + + if (PREDICT_FALSE (prev_elt != 0)) + prev_elt->next_index = elt - elts; + + if (PREDICT_FALSE (i < (chunks_this_buffer - 1))) + { + elt->flags = SSVM_BUFFER_NEXT_PRESENT; + ASSERT (b0->flags & VLIB_BUFFER_NEXT_PRESENT); + b0 = vlib_get_buffer (vm, b0->next_buffer); + } + prev_elt = elt; + } + + while (__sync_lock_test_and_set (queue_lock, 1)) + ; + + unix_shared_memory_queue_add_raw (q, (u8 *) & elt_index); + CLIB_MEMORY_BARRIER (); + *queue_lock = 0; + + from++; + n_left--; + } + +out: + if (PREDICT_FALSE (n_left)) + { + if (is_ring_full) + vlib_error_count (vm, node->node_index, SSVM_ETH_TX_ERROR_RING_FULL, + n_left); + else if (interface_down) + vlib_error_count (vm, node->node_index, SSVM_ETH_TX_ERROR_ADMIN_DOWN, + n_left); + else + vlib_error_count (vm, node->node_index, SSVM_ETH_TX_ERROR_NO_BUFFERS, + n_left); + + vlib_buffer_free (vm, from, n_left); + } + else + vlib_buffer_free (vm, vlib_frame_vector_args (f), f->n_vectors); + + if (PREDICT_TRUE (vec_len (em->chunk_cache))) + _vec_len (em->chunk_cache) = n_present_in_cache; + + return f->n_vectors; +} + +static void +ssvm_eth_clear_hw_interface_counters (u32 instance) +{ + /* Nothing for now */ +} + +static clib_error_t * +ssvm_eth_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, + u32 flags) +{ + vnet_hw_interface_t *hif = vnet_get_hw_interface (vnm, hw_if_index); + uword is_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0; + ssvm_eth_main_t *em = &ssvm_eth_main; + ssvm_private_t *intfc = vec_elt_at_index (em->intfcs, hif->dev_instance); + ssvm_shared_header_t *sh; + + /* publish link-state in shared-memory, to discourage buffer-wasting */ + sh = intfc->sh; + if (intfc->i_am_master) + sh->opaque[MASTER_ADMIN_STATE_INDEX] = (void *) is_up; + else + sh->opaque[SLAVE_ADMIN_STATE_INDEX] = (void *) is_up; + + return 0; +} + +static clib_error_t * +ssvm_eth_subif_add_del_function (vnet_main_t * vnm, + u32 hw_if_index, + struct vnet_sw_interface_t *st, int is_add) +{ + /* Nothing for now */ + return 0; +} + +/* + * Dynamically redirect all pkts from a specific interface + * to the specified node + */ +static void +ssvm_eth_set_interface_next_node (vnet_main_t * vnm, u32 hw_if_index, + u32 node_index) +{ + ssvm_eth_main_t *em = &ssvm_eth_main; + vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index); + ssvm_private_t *intfc = pool_elt_at_index (em->intfcs, hw->dev_instance); + + /* Shut off redirection */ + if (node_index == ~0) + { + intfc->per_interface_next_index = node_index; + return; + } + + intfc->per_interface_next_index = + vlib_node_add_next (em->vlib_main, ssvm_eth_input_node.index, node_index); +} + +static u32 +ssvm_eth_flag_change (vnet_main_t * vnm, vnet_hw_interface_t * hi, u32 flags) +{ + /* nothing for now */ + return 0; +} + +/* *INDENT-OFF* */ +VNET_DEVICE_CLASS (ssvm_eth_device_class) = { + .name = "ssvm-eth", + .tx_function = ssvm_eth_interface_tx, + .tx_function_n_errors = SSVM_ETH_TX_N_ERROR, + .tx_function_error_strings = ssvm_eth_tx_func_error_strings, + .format_device_name = format_ssvm_eth_device_name, + .format_device = format_ssvm_eth_device, + .format_tx_trace = format_ssvm_eth_tx_trace, + .clear_counters = ssvm_eth_clear_hw_interface_counters, + .admin_up_down_function = ssvm_eth_interface_admin_up_down, + .subif_add_del_function = ssvm_eth_subif_add_del_function, + .rx_redirect_to_node = ssvm_eth_set_interface_next_node, +}; + +VLIB_DEVICE_TX_FUNCTION_MULTIARCH (ssvm_eth_device_class, + ssvm_eth_interface_tx) +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/ssvm/ssvm_eth.h b/src/vnet/devices/ssvm/ssvm_eth.h new file mode 100644 index 00000000000..f877df3cd33 --- /dev/null +++ b/src/vnet/devices/ssvm/ssvm_eth.h @@ -0,0 +1,141 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#ifndef __included_ssvm_eth_h__ +#define __included_ssvm_eth_h__ + +#include + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include + +extern vnet_device_class_t ssvm_eth_device_class; +extern vlib_node_registration_t ssvm_eth_input_node; + +#define SSVM_BUFFER_SIZE \ + (VLIB_BUFFER_DATA_SIZE + VLIB_BUFFER_PRE_DATA_SIZE) +#define SSVM_PACKET_TYPE 1 + +typedef struct +{ + /* Type of queue element */ + u8 type; + u8 flags; +#define SSVM_BUFFER_NEXT_PRESENT (1<<0) + u8 owner; + u8 tag; + i16 current_data_hint; + u16 length_this_buffer; + u16 total_length_not_including_first_buffer; + u16 pad; + u32 next_index; + /* offset 16 */ + u8 data[SSVM_BUFFER_SIZE]; + /* pad to an even multiple of 64 octets */ + u8 pad2[CLIB_CACHE_LINE_BYTES - 16]; +} ssvm_eth_queue_elt_t; + +typedef struct +{ + /* vector of point-to-point connections */ + ssvm_private_t *intfcs; + + u32 *buffer_cache; + u32 *chunk_cache; + + /* Configurable parameters */ + /* base address for next placement */ + u64 next_base_va; + u64 segment_size; + u64 nbuffers; + u64 queue_elts; + + /* Segment names */ + u8 **names; + + /* convenience */ + vlib_main_t *vlib_main; + vnet_main_t *vnet_main; + elog_main_t *elog_main; +} ssvm_eth_main_t; + +ssvm_eth_main_t ssvm_eth_main; + +typedef enum +{ + CHUNK_POOL_FREELIST_INDEX = 0, + CHUNK_POOL_INDEX, + CHUNK_POOL_NFREE, + TO_MASTER_Q_INDEX, + TO_SLAVE_Q_INDEX, + MASTER_ADMIN_STATE_INDEX, + SLAVE_ADMIN_STATE_INDEX, +} ssvm_eth_opaque_index_t; + +/* + * debug scaffolding. + */ +static inline void +ssvm_eth_validate_freelists (int need_lock) +{ +#if CLIB_DEBUG > 0 + ssvm_eth_main_t *em = &ssvm_eth_main; + ssvm_private_t *intfc; + ssvm_shared_header_t *sh; + u32 *elt_indices; + u32 n_available; + int i; + + for (i = 0; i < vec_len (em->intfcs); i++) + { + intfc = em->intfcs + i; + sh = intfc->sh; + u32 my_pid = intfc->my_pid; + + if (need_lock) + ssvm_lock (sh, my_pid, 15); + + elt_indices = (u32 *) (sh->opaque[CHUNK_POOL_FREELIST_INDEX]); + n_available = (u32) (uword) (sh->opaque[CHUNK_POOL_NFREE]); + + for (i = 0; i < n_available; i++) + ASSERT (elt_indices[i] < 2048); + + if (need_lock) + ssvm_unlock (sh); + } +#endif +} + +#endif /* __included_ssvm_eth_h__ */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/virtio/dir.dox b/src/vnet/devices/virtio/dir.dox new file mode 100644 index 00000000000..50150799e62 --- /dev/null +++ b/src/vnet/devices/virtio/dir.dox @@ -0,0 +1,27 @@ +/* + * Copyright (c) 2016 Cisco and/or its affiliates. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* Doxygen directory documentation */ + +/** +@dir +@brief vHost User Interface Implementation. + +This directory contains the source code for vHost User driver. + +*/ +/*? %%clicmd:group_label vHost User %% ?*/ +/*? %%syscfg:group_label vHost User %% ?*/ diff --git a/src/vnet/devices/virtio/vhost-user.c b/src/vnet/devices/virtio/vhost-user.c new file mode 100644 index 00000000000..bde8106c501 --- /dev/null +++ b/src/vnet/devices/virtio/vhost-user.c @@ -0,0 +1,3314 @@ +/* + *------------------------------------------------------------------ + * vhost.c - vhost-user + * + * Copyright (c) 2014 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include /* for open */ +#include +#include +#include +#include +#include +#include /* for iovec */ +#include +#include + +#include +#include + +#include +#include + +#include + +#include +#include +#include + +#include + +/** + * @file + * @brief vHost User Device Driver. + * + * This file contains the source code for vHost User interface. + */ + + +#define VHOST_USER_DEBUG_SOCKET 0 +#define VHOST_DEBUG_VQ 0 + +#if VHOST_USER_DEBUG_SOCKET == 1 +#define DBG_SOCK(args...) clib_warning(args); +#else +#define DBG_SOCK(args...) +#endif + +#if VHOST_DEBUG_VQ == 1 +#define DBG_VQ(args...) clib_warning(args); +#else +#define DBG_VQ(args...) +#endif + +/* + * When an RX queue is down but active, received packets + * must be discarded. This value controls up to how many + * packets will be discarded during each round. + */ +#define VHOST_USER_DOWN_DISCARD_COUNT 256 + +/* + * When the number of available buffers gets under this threshold, + * RX node will start discarding packets. + */ +#define VHOST_USER_RX_BUFFER_STARVATION 32 + +/* + * On the receive side, the host should free descriptors as soon + * as possible in order to avoid TX drop in the VM. + * This value controls the number of copy operations that are stacked + * before copy is done for all and descriptors are given back to + * the guest. + * The value 64 was obtained by testing (48 and 128 were not as good). + */ +#define VHOST_USER_RX_COPY_THRESHOLD 64 + +#define UNIX_GET_FD(unixfd_idx) \ + (unixfd_idx != ~0) ? \ + pool_elt_at_index (unix_main.file_pool, \ + unixfd_idx)->file_descriptor : -1; + +#define foreach_virtio_trace_flags \ + _ (SIMPLE_CHAINED, 0, "Simple descriptor chaining") \ + _ (SINGLE_DESC, 1, "Single descriptor packet") \ + _ (INDIRECT, 2, "Indirect descriptor") \ + _ (MAP_ERROR, 4, "Memory mapping error") + +typedef enum +{ +#define _(n,i,s) VIRTIO_TRACE_F_##n, + foreach_virtio_trace_flags +#undef _ +} virtio_trace_flag_t; + +vlib_node_registration_t vhost_user_input_node; + +#define foreach_vhost_user_tx_func_error \ + _(NONE, "no error") \ + _(NOT_READY, "vhost vring not ready") \ + _(DOWN, "vhost interface is down") \ + _(PKT_DROP_NOBUF, "tx packet drops (no available descriptors)") \ + _(PKT_DROP_NOMRG, "tx packet drops (cannot merge descriptors)") \ + _(MMAP_FAIL, "mmap failure") \ + _(INDIRECT_OVERFLOW, "indirect descriptor table overflow") + +typedef enum +{ +#define _(f,s) VHOST_USER_TX_FUNC_ERROR_##f, + foreach_vhost_user_tx_func_error +#undef _ + VHOST_USER_TX_FUNC_N_ERROR, +} vhost_user_tx_func_error_t; + +static char *vhost_user_tx_func_error_strings[] = { +#define _(n,s) s, + foreach_vhost_user_tx_func_error +#undef _ +}; + +#define foreach_vhost_user_input_func_error \ + _(NO_ERROR, "no error") \ + _(NO_BUFFER, "no available buffer") \ + _(MMAP_FAIL, "mmap failure") \ + _(INDIRECT_OVERFLOW, "indirect descriptor overflows table") \ + _(UNDERSIZED_FRAME, "undersized ethernet frame received (< 14 bytes)") \ + _(FULL_RX_QUEUE, "full rx queue (possible driver tx drop)") + +typedef enum +{ +#define _(f,s) VHOST_USER_INPUT_FUNC_ERROR_##f, + foreach_vhost_user_input_func_error +#undef _ + VHOST_USER_INPUT_FUNC_N_ERROR, +} vhost_user_input_func_error_t; + +static char *vhost_user_input_func_error_strings[] = { +#define _(n,s) s, + foreach_vhost_user_input_func_error +#undef _ +}; + +/* *INDENT-OFF* */ +static vhost_user_main_t vhost_user_main = { + .mtu_bytes = 1518, +}; + +VNET_HW_INTERFACE_CLASS (vhost_interface_class, static) = { + .name = "vhost-user", +}; +/* *INDENT-ON* */ + +static u8 * +format_vhost_user_interface_name (u8 * s, va_list * args) +{ + u32 i = va_arg (*args, u32); + u32 show_dev_instance = ~0; + vhost_user_main_t *vum = &vhost_user_main; + + if (i < vec_len (vum->show_dev_instance_by_real_dev_instance)) + show_dev_instance = vum->show_dev_instance_by_real_dev_instance[i]; + + if (show_dev_instance != ~0) + i = show_dev_instance; + + s = format (s, "VirtualEthernet0/0/%d", i); + return s; +} + +static int +vhost_user_name_renumber (vnet_hw_interface_t * hi, u32 new_dev_instance) +{ + // FIXME: check if the new dev instance is already used + vhost_user_main_t *vum = &vhost_user_main; + vec_validate_init_empty (vum->show_dev_instance_by_real_dev_instance, + hi->dev_instance, ~0); + + vum->show_dev_instance_by_real_dev_instance[hi->dev_instance] = + new_dev_instance; + + DBG_SOCK ("renumbered vhost-user interface dev_instance %d to %d", + hi->dev_instance, new_dev_instance); + + return 0; +} + +static_always_inline void * +map_guest_mem (vhost_user_intf_t * vui, uword addr, u32 * hint) +{ + int i = *hint; + if (PREDICT_TRUE ((vui->regions[i].guest_phys_addr <= addr) && + ((vui->regions[i].guest_phys_addr + + vui->regions[i].memory_size) > addr))) + { + return (void *) (vui->region_mmap_addr[i] + addr - + vui->regions[i].guest_phys_addr); + } +#if __SSE4_2__ + __m128i rl, rh, al, ah, r; + al = _mm_set1_epi64x (addr + 1); + ah = _mm_set1_epi64x (addr); + + rl = _mm_loadu_si128 ((__m128i *) & vui->region_guest_addr_lo[0]); + rl = _mm_cmpgt_epi64 (al, rl); + rh = _mm_loadu_si128 ((__m128i *) & vui->region_guest_addr_hi[0]); + rh = _mm_cmpgt_epi64 (rh, ah); + r = _mm_and_si128 (rl, rh); + + rl = _mm_loadu_si128 ((__m128i *) & vui->region_guest_addr_lo[2]); + rl = _mm_cmpgt_epi64 (al, rl); + rh = _mm_loadu_si128 ((__m128i *) & vui->region_guest_addr_hi[2]); + rh = _mm_cmpgt_epi64 (rh, ah); + r = _mm_blend_epi16 (r, _mm_and_si128 (rl, rh), 0x22); + + rl = _mm_loadu_si128 ((__m128i *) & vui->region_guest_addr_lo[4]); + rl = _mm_cmpgt_epi64 (al, rl); + rh = _mm_loadu_si128 ((__m128i *) & vui->region_guest_addr_hi[4]); + rh = _mm_cmpgt_epi64 (rh, ah); + r = _mm_blend_epi16 (r, _mm_and_si128 (rl, rh), 0x44); + + rl = _mm_loadu_si128 ((__m128i *) & vui->region_guest_addr_lo[6]); + rl = _mm_cmpgt_epi64 (al, rl); + rh = _mm_loadu_si128 ((__m128i *) & vui->region_guest_addr_hi[6]); + rh = _mm_cmpgt_epi64 (rh, ah); + r = _mm_blend_epi16 (r, _mm_and_si128 (rl, rh), 0x88); + + r = _mm_shuffle_epi8 (r, _mm_set_epi64x (0, 0x0e060c040a020800)); + i = __builtin_ctzll (_mm_movemask_epi8 (r)); + + if (i < vui->nregions) + { + *hint = i; + return (void *) (vui->region_mmap_addr[i] + addr - + vui->regions[i].guest_phys_addr); + } + +#else + for (i = 0; i < vui->nregions; i++) + { + if ((vui->regions[i].guest_phys_addr <= addr) && + ((vui->regions[i].guest_phys_addr + vui->regions[i].memory_size) > + addr)) + { + *hint = i; + return (void *) (vui->region_mmap_addr[i] + addr - + vui->regions[i].guest_phys_addr); + } + } +#endif + DBG_VQ ("failed to map guest mem addr %llx", addr); + *hint = 0; + return 0; +} + +static inline void * +map_user_mem (vhost_user_intf_t * vui, uword addr) +{ + int i; + for (i = 0; i < vui->nregions; i++) + { + if ((vui->regions[i].userspace_addr <= addr) && + ((vui->regions[i].userspace_addr + vui->regions[i].memory_size) > + addr)) + { + return (void *) (vui->region_mmap_addr[i] + addr - + vui->regions[i].userspace_addr); + } + } + return 0; +} + +static long +get_huge_page_size (int fd) +{ + struct statfs s; + fstatfs (fd, &s); + return s.f_bsize; +} + +static void +unmap_all_mem_regions (vhost_user_intf_t * vui) +{ + int i, r; + for (i = 0; i < vui->nregions; i++) + { + if (vui->region_mmap_addr[i] != (void *) -1) + { + + long page_sz = get_huge_page_size (vui->region_mmap_fd[i]); + + ssize_t map_sz = (vui->regions[i].memory_size + + vui->regions[i].mmap_offset + + page_sz) & ~(page_sz - 1); + + r = + munmap (vui->region_mmap_addr[i] - vui->regions[i].mmap_offset, + map_sz); + + DBG_SOCK + ("unmap memory region %d addr 0x%lx len 0x%lx page_sz 0x%x", i, + vui->region_mmap_addr[i], map_sz, page_sz); + + vui->region_mmap_addr[i] = (void *) -1; + + if (r == -1) + { + clib_warning ("failed to unmap memory region (errno %d)", + errno); + } + close (vui->region_mmap_fd[i]); + } + } + vui->nregions = 0; +} + +static void +vhost_user_tx_thread_placement (vhost_user_intf_t * vui) +{ + //Let's try to assign one queue to each thread + u32 qid = 0; + u32 cpu_index = 0; + vui->use_tx_spinlock = 0; + while (1) + { + for (qid = 0; qid < VHOST_VRING_MAX_N / 2; qid++) + { + vhost_user_vring_t *rxvq = &vui->vrings[VHOST_VRING_IDX_RX (qid)]; + if (!rxvq->started || !rxvq->enabled) + continue; + + vui->per_cpu_tx_qid[cpu_index] = qid; + cpu_index++; + if (cpu_index == vlib_get_thread_main ()->n_vlib_mains) + return; + } + //We need to loop, meaning the spinlock has to be used + vui->use_tx_spinlock = 1; + if (cpu_index == 0) + { + //Could not find a single valid one + for (cpu_index = 0; + cpu_index < vlib_get_thread_main ()->n_vlib_mains; cpu_index++) + { + vui->per_cpu_tx_qid[cpu_index] = 0; + } + return; + } + } +} + +static void +vhost_user_rx_thread_placement () +{ + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + vhost_cpu_t *vhc; + u32 *workers = 0; + + //Let's list all workers cpu indexes + u32 i; + for (i = vum->input_cpu_first_index; + i < vum->input_cpu_first_index + vum->input_cpu_count; i++) + { + vlib_node_set_state (vlib_mains ? vlib_mains[i] : &vlib_global_main, + vhost_user_input_node.index, + VLIB_NODE_STATE_DISABLED); + vec_add1 (workers, i); + } + + vec_foreach (vhc, vum->cpus) + { + vec_reset_length (vhc->rx_queues); + } + + i = 0; + vhost_iface_and_queue_t iaq; + /* *INDENT-OFF* */ + pool_foreach (vui, vum->vhost_user_interfaces, { + u32 *vui_workers = vec_len (vui->workers) ? vui->workers : workers; + u32 qid; + for (qid = 0; qid < VHOST_VRING_MAX_N / 2; qid++) + { + vhost_user_vring_t *txvq = + &vui->vrings[VHOST_VRING_IDX_TX (qid)]; + if (!txvq->started) + continue; + + i %= vec_len (vui_workers); + u32 cpu_index = vui_workers[i]; + i++; + vhc = &vum->cpus[cpu_index]; + + iaq.qid = qid; + iaq.vhost_iface_index = vui - vum->vhost_user_interfaces; + vec_add1 (vhc->rx_queues, iaq); + vlib_node_set_state (vlib_mains ? vlib_mains[cpu_index] : + &vlib_global_main, vhost_user_input_node.index, + VLIB_NODE_STATE_POLLING); + } + }); + /* *INDENT-ON* */ +} + +static int +vhost_user_thread_placement (u32 sw_if_index, u32 worker_thread_index, u8 del) +{ + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + vnet_hw_interface_t *hw; + + if (worker_thread_index < vum->input_cpu_first_index || + worker_thread_index >= + vum->input_cpu_first_index + vum->input_cpu_count) + return -1; + + if (!(hw = vnet_get_sup_hw_interface (vnet_get_main (), sw_if_index))) + return -2; + + vui = pool_elt_at_index (vum->vhost_user_interfaces, hw->dev_instance); + u32 found = ~0, *w; + vec_foreach (w, vui->workers) + { + if (*w == worker_thread_index) + { + found = w - vui->workers; + break; + } + } + + if (del) + { + if (found == ~0) + return -3; + vec_del1 (vui->workers, found); + } + else if (found == ~0) + { + vec_add1 (vui->workers, worker_thread_index); + } + + vhost_user_rx_thread_placement (); + return 0; +} + +/** @brief Returns whether at least one TX and one RX vring are enabled */ +int +vhost_user_intf_ready (vhost_user_intf_t * vui) +{ + int i, found[2] = { }; //RX + TX + + for (i = 0; i < VHOST_VRING_MAX_N; i++) + if (vui->vrings[i].started && vui->vrings[i].enabled) + found[i & 1] = 1; + + return found[0] && found[1]; +} + +static void +vhost_user_update_iface_state (vhost_user_intf_t * vui) +{ + /* if we have pointers to descriptor table, go up */ + int is_up = vhost_user_intf_ready (vui); + if (is_up != vui->is_up) + { + DBG_SOCK ("interface %d %s", vui->sw_if_index, + is_up ? "ready" : "down"); + vnet_hw_interface_set_flags (vnet_get_main (), vui->hw_if_index, + is_up ? VNET_HW_INTERFACE_FLAG_LINK_UP : + 0); + vui->is_up = is_up; + } + vhost_user_rx_thread_placement (); + vhost_user_tx_thread_placement (vui); +} + +static clib_error_t * +vhost_user_callfd_read_ready (unix_file_t * uf) +{ + __attribute__ ((unused)) int n; + u8 buff[8]; + n = read (uf->file_descriptor, ((char *) &buff), 8); + return 0; +} + +static clib_error_t * +vhost_user_kickfd_read_ready (unix_file_t * uf) +{ + __attribute__ ((unused)) int n; + u8 buff[8]; + vhost_user_intf_t *vui = + pool_elt_at_index (vhost_user_main.vhost_user_interfaces, + uf->private_data >> 8); + u32 qid = uf->private_data & 0xff; + n = read (uf->file_descriptor, ((char *) &buff), 8); + DBG_SOCK ("if %d KICK queue %d", uf->private_data >> 8, qid); + + vlib_worker_thread_barrier_sync (vlib_get_main ()); + vui->vrings[qid].started = 1; + vhost_user_update_iface_state (vui); + vlib_worker_thread_barrier_release (vlib_get_main ()); + return 0; +} + +/** + * @brief Try once to lock the vring + * @return 0 on success, non-zero on failure. + */ +static inline int +vhost_user_vring_try_lock (vhost_user_intf_t * vui, u32 qid) +{ + return __sync_lock_test_and_set (vui->vring_locks[qid], 1); +} + +/** + * @brief Spin until the vring is successfully locked + */ +static inline void +vhost_user_vring_lock (vhost_user_intf_t * vui, u32 qid) +{ + while (vhost_user_vring_try_lock (vui, qid)) + ; +} + +/** + * @brief Unlock the vring lock + */ +static inline void +vhost_user_vring_unlock (vhost_user_intf_t * vui, u32 qid) +{ + *vui->vring_locks[qid] = 0; +} + +static inline void +vhost_user_vring_init (vhost_user_intf_t * vui, u32 qid) +{ + vhost_user_vring_t *vring = &vui->vrings[qid]; + memset (vring, 0, sizeof (*vring)); + vring->kickfd_idx = ~0; + vring->callfd_idx = ~0; + vring->errfd = -1; + + /* + * We have a bug with some qemu 2.5, and this may be a fix. + * Feel like interpretation holy text, but this is from vhost-user.txt. + * " + * One queue pair is enabled initially. More queues are enabled + * dynamically, by sending message VHOST_USER_SET_VRING_ENABLE. + * " + * Don't know who's right, but this is what DPDK does. + */ + if (qid == 0 || qid == 1) + vring->enabled = 1; +} + +static inline void +vhost_user_vring_close (vhost_user_intf_t * vui, u32 qid) +{ + vhost_user_vring_t *vring = &vui->vrings[qid]; + if (vring->kickfd_idx != ~0) + { + unix_file_t *uf = pool_elt_at_index (unix_main.file_pool, + vring->kickfd_idx); + unix_file_del (&unix_main, uf); + vring->kickfd_idx = ~0; + } + if (vring->callfd_idx != ~0) + { + unix_file_t *uf = pool_elt_at_index (unix_main.file_pool, + vring->callfd_idx); + unix_file_del (&unix_main, uf); + vring->callfd_idx = ~0; + } + if (vring->errfd != -1) + close (vring->errfd); + vhost_user_vring_init (vui, qid); +} + +static inline void +vhost_user_if_disconnect (vhost_user_intf_t * vui) +{ + vnet_main_t *vnm = vnet_get_main (); + int q; + + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); + + if (vui->unix_file_index != ~0) + { + unix_file_del (&unix_main, unix_main.file_pool + vui->unix_file_index); + vui->unix_file_index = ~0; + } + + vui->is_up = 0; + + for (q = 0; q < VHOST_VRING_MAX_N; q++) + vhost_user_vring_close (vui, q); + + unmap_all_mem_regions (vui); + DBG_SOCK ("interface ifindex %d disconnected", vui->sw_if_index); +} + +#define VHOST_LOG_PAGE 0x1000 +static_always_inline void +vhost_user_log_dirty_pages_2 (vhost_user_intf_t * vui, + u64 addr, u64 len, u8 is_host_address) +{ + if (PREDICT_TRUE (vui->log_base_addr == 0 + || !(vui->features & (1 << FEAT_VHOST_F_LOG_ALL)))) + { + return; + } + if (is_host_address) + { + addr = (u64) map_user_mem (vui, (uword) addr); + } + if (PREDICT_FALSE ((addr + len - 1) / VHOST_LOG_PAGE / 8 >= vui->log_size)) + { + DBG_SOCK ("vhost_user_log_dirty_pages(): out of range\n"); + return; + } + + CLIB_MEMORY_BARRIER (); + u64 page = addr / VHOST_LOG_PAGE; + while (page * VHOST_LOG_PAGE < addr + len) + { + ((u8 *) vui->log_base_addr)[page / 8] |= 1 << page % 8; + page++; + } +} + +static_always_inline void +vhost_user_log_dirty_pages (vhost_user_intf_t * vui, u64 addr, u64 len) +{ + vhost_user_log_dirty_pages_2 (vui, addr, len, 0); +} + +#define vhost_user_log_dirty_ring(vui, vq, member) \ + if (PREDICT_FALSE(vq->log_used)) { \ + vhost_user_log_dirty_pages(vui, vq->log_guest_addr + STRUCT_OFFSET_OF(vring_used_t, member), \ + sizeof(vq->used->member)); \ + } + +static clib_error_t * +vhost_user_socket_read (unix_file_t * uf) +{ + int n, i; + int fd, number_of_fds = 0; + int fds[VHOST_MEMORY_MAX_NREGIONS]; + vhost_user_msg_t msg; + struct msghdr mh; + struct iovec iov[1]; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + struct cmsghdr *cmsg; + u8 q; + unix_file_t template = { 0 }; + vnet_main_t *vnm = vnet_get_main (); + + vui = pool_elt_at_index (vum->vhost_user_interfaces, uf->private_data); + + char control[CMSG_SPACE (VHOST_MEMORY_MAX_NREGIONS * sizeof (int))]; + + memset (&mh, 0, sizeof (mh)); + memset (control, 0, sizeof (control)); + + for (i = 0; i < VHOST_MEMORY_MAX_NREGIONS; i++) + fds[i] = -1; + + /* set the payload */ + iov[0].iov_base = (void *) &msg; + iov[0].iov_len = VHOST_USER_MSG_HDR_SZ; + + mh.msg_iov = iov; + mh.msg_iovlen = 1; + mh.msg_control = control; + mh.msg_controllen = sizeof (control); + + n = recvmsg (uf->file_descriptor, &mh, 0); + + /* Stop workers to avoid end of the world */ + vlib_worker_thread_barrier_sync (vlib_get_main ()); + + if (n != VHOST_USER_MSG_HDR_SZ) + { + if (n == -1) + { + DBG_SOCK ("recvmsg returned error %d %s", errno, strerror (errno)); + } + else + { + DBG_SOCK ("n (%d) != VHOST_USER_MSG_HDR_SZ (%d)", + n, VHOST_USER_MSG_HDR_SZ); + } + goto close_socket; + } + + if (mh.msg_flags & MSG_CTRUNC) + { + DBG_SOCK ("MSG_CTRUNC is set"); + goto close_socket; + } + + cmsg = CMSG_FIRSTHDR (&mh); + + if (cmsg && (cmsg->cmsg_len > 0) && (cmsg->cmsg_level == SOL_SOCKET) && + (cmsg->cmsg_type == SCM_RIGHTS) && + (cmsg->cmsg_len - CMSG_LEN (0) <= + VHOST_MEMORY_MAX_NREGIONS * sizeof (int))) + { + number_of_fds = (cmsg->cmsg_len - CMSG_LEN (0)) / sizeof (int); + clib_memcpy (fds, CMSG_DATA (cmsg), number_of_fds * sizeof (int)); + } + + /* version 1, no reply bit set */ + if ((msg.flags & 7) != 1) + { + DBG_SOCK ("malformed message received. closing socket"); + goto close_socket; + } + + { + int rv; + rv = + read (uf->file_descriptor, ((char *) &msg) + VHOST_USER_MSG_HDR_SZ, + msg.size); + if (rv < 0) + { + DBG_SOCK ("read failed %s", strerror (errno)); + goto close_socket; + } + else if (rv != msg.size) + { + DBG_SOCK ("message too short (read %dB should be %dB)", rv, msg.size); + goto close_socket; + } + } + + switch (msg.request) + { + case VHOST_USER_GET_FEATURES: + msg.flags |= 4; + msg.u64 = (1ULL << FEAT_VIRTIO_NET_F_MRG_RXBUF) | + (1ULL << FEAT_VIRTIO_NET_F_CTRL_VQ) | + (1ULL << FEAT_VIRTIO_F_ANY_LAYOUT) | + (1ULL << FEAT_VIRTIO_F_INDIRECT_DESC) | + (1ULL << FEAT_VHOST_F_LOG_ALL) | + (1ULL << FEAT_VIRTIO_NET_F_GUEST_ANNOUNCE) | + (1ULL << FEAT_VIRTIO_NET_F_MQ) | + (1ULL << FEAT_VHOST_USER_F_PROTOCOL_FEATURES) | + (1ULL << FEAT_VIRTIO_F_VERSION_1); + msg.u64 &= vui->feature_mask; + msg.size = sizeof (msg.u64); + DBG_SOCK ("if %d msg VHOST_USER_GET_FEATURES - reply 0x%016llx", + vui->hw_if_index, msg.u64); + break; + + case VHOST_USER_SET_FEATURES: + DBG_SOCK ("if %d msg VHOST_USER_SET_FEATURES features 0x%016llx", + vui->hw_if_index, msg.u64); + + vui->features = msg.u64; + + if (vui->features & + ((1 << FEAT_VIRTIO_NET_F_MRG_RXBUF) | + (1ULL << FEAT_VIRTIO_F_VERSION_1))) + vui->virtio_net_hdr_sz = 12; + else + vui->virtio_net_hdr_sz = 10; + + vui->is_any_layout = + (vui->features & (1 << FEAT_VIRTIO_F_ANY_LAYOUT)) ? 1 : 0; + + ASSERT (vui->virtio_net_hdr_sz < VLIB_BUFFER_PRE_DATA_SIZE); + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); + vui->is_up = 0; + + /*for (q = 0; q < VHOST_VRING_MAX_N; q++) + vhost_user_vring_close(&vui->vrings[q]); */ + + break; + + case VHOST_USER_SET_MEM_TABLE: + DBG_SOCK ("if %d msg VHOST_USER_SET_MEM_TABLE nregions %d", + vui->hw_if_index, msg.memory.nregions); + + if ((msg.memory.nregions < 1) || + (msg.memory.nregions > VHOST_MEMORY_MAX_NREGIONS)) + { + + DBG_SOCK ("number of mem regions must be between 1 and %i", + VHOST_MEMORY_MAX_NREGIONS); + + goto close_socket; + } + + if (msg.memory.nregions != number_of_fds) + { + DBG_SOCK ("each memory region must have FD"); + goto close_socket; + } + unmap_all_mem_regions (vui); + for (i = 0; i < msg.memory.nregions; i++) + { + clib_memcpy (&(vui->regions[i]), &msg.memory.regions[i], + sizeof (vhost_user_memory_region_t)); + + long page_sz = get_huge_page_size (fds[i]); + + /* align size to 2M page */ + ssize_t map_sz = (vui->regions[i].memory_size + + vui->regions[i].mmap_offset + + page_sz) & ~(page_sz - 1); + + vui->region_mmap_addr[i] = mmap (0, map_sz, PROT_READ | PROT_WRITE, + MAP_SHARED, fds[i], 0); + vui->region_guest_addr_lo[i] = vui->regions[i].guest_phys_addr; + vui->region_guest_addr_hi[i] = vui->regions[i].guest_phys_addr + + vui->regions[i].memory_size; + + DBG_SOCK + ("map memory region %d addr 0 len 0x%lx fd %d mapped 0x%lx " + "page_sz 0x%x", i, map_sz, fds[i], vui->region_mmap_addr[i], + page_sz); + + if (vui->region_mmap_addr[i] == MAP_FAILED) + { + clib_warning ("failed to map memory. errno is %d", errno); + goto close_socket; + } + vui->region_mmap_addr[i] += vui->regions[i].mmap_offset; + vui->region_mmap_fd[i] = fds[i]; + } + vui->nregions = msg.memory.nregions; + break; + + case VHOST_USER_SET_VRING_NUM: + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_NUM idx %d num %d", + vui->hw_if_index, msg.state.index, msg.state.num); + + if ((msg.state.num > 32768) || /* maximum ring size is 32768 */ + (msg.state.num == 0) || /* it cannot be zero */ + ((msg.state.num - 1) & msg.state.num)) /* must be power of 2 */ + goto close_socket; + vui->vrings[msg.state.index].qsz = msg.state.num; + break; + + case VHOST_USER_SET_VRING_ADDR: + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_ADDR idx %d", + vui->hw_if_index, msg.state.index); + + if (msg.state.index >= VHOST_VRING_MAX_N) + { + DBG_SOCK ("invalid vring index VHOST_USER_SET_VRING_ADDR:" + " %d >= %d", msg.state.index, VHOST_VRING_MAX_N); + goto close_socket; + } + + if (msg.size < sizeof (msg.addr)) + { + DBG_SOCK ("vhost message is too short (%d < %d)", + msg.size, sizeof (msg.addr)); + goto close_socket; + } + + vui->vrings[msg.state.index].desc = (vring_desc_t *) + map_user_mem (vui, msg.addr.desc_user_addr); + vui->vrings[msg.state.index].used = (vring_used_t *) + map_user_mem (vui, msg.addr.used_user_addr); + vui->vrings[msg.state.index].avail = (vring_avail_t *) + map_user_mem (vui, msg.addr.avail_user_addr); + + if ((vui->vrings[msg.state.index].desc == NULL) || + (vui->vrings[msg.state.index].used == NULL) || + (vui->vrings[msg.state.index].avail == NULL)) + { + DBG_SOCK ("failed to map user memory for hw_if_index %d", + vui->hw_if_index); + goto close_socket; + } + + vui->vrings[msg.state.index].log_guest_addr = msg.addr.log_guest_addr; + vui->vrings[msg.state.index].log_used = + (msg.addr.flags & (1 << VHOST_VRING_F_LOG)) ? 1 : 0; + + /* Spec says: If VHOST_USER_F_PROTOCOL_FEATURES has not been negotiated, + the ring is initialized in an enabled state. */ + if (!(vui->features & (1 << FEAT_VHOST_USER_F_PROTOCOL_FEATURES))) + { + vui->vrings[msg.state.index].enabled = 1; + } + + vui->vrings[msg.state.index].last_used_idx = + vui->vrings[msg.state.index].last_avail_idx = + vui->vrings[msg.state.index].used->idx; + + /* tell driver that we don't want interrupts */ + vui->vrings[msg.state.index].used->flags = VRING_USED_F_NO_NOTIFY; + break; + + case VHOST_USER_SET_OWNER: + DBG_SOCK ("if %d msg VHOST_USER_SET_OWNER", vui->hw_if_index); + break; + + case VHOST_USER_RESET_OWNER: + DBG_SOCK ("if %d msg VHOST_USER_RESET_OWNER", vui->hw_if_index); + break; + + case VHOST_USER_SET_VRING_CALL: + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_CALL u64 %d", + vui->hw_if_index, msg.u64); + + q = (u8) (msg.u64 & 0xFF); + + /* if there is old fd, delete and close it */ + if (vui->vrings[q].callfd_idx != ~0) + { + unix_file_t *uf = pool_elt_at_index (unix_main.file_pool, + vui->vrings[q].callfd_idx); + unix_file_del (&unix_main, uf); + vui->vrings[q].callfd_idx = ~0; + } + + if (!(msg.u64 & 0x100)) + { + if (number_of_fds != 1) + { + DBG_SOCK ("More than one fd received !"); + goto close_socket; + } + + template.read_function = vhost_user_callfd_read_ready; + template.file_descriptor = fds[0]; + template.private_data = + ((vui - vhost_user_main.vhost_user_interfaces) << 8) + q; + vui->vrings[q].callfd_idx = unix_file_add (&unix_main, &template); + } + else + vui->vrings[q].callfd_idx = ~0; + break; + + case VHOST_USER_SET_VRING_KICK: + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_KICK u64 %d", + vui->hw_if_index, msg.u64); + + q = (u8) (msg.u64 & 0xFF); + + if (vui->vrings[q].kickfd_idx != ~0) + { + unix_file_t *uf = pool_elt_at_index (unix_main.file_pool, + vui->vrings[q].kickfd_idx); + unix_file_del (&unix_main, uf); + vui->vrings[q].kickfd_idx = ~0; + } + + if (!(msg.u64 & 0x100)) + { + if (number_of_fds != 1) + { + DBG_SOCK ("More than one fd received !"); + goto close_socket; + } + + template.read_function = vhost_user_kickfd_read_ready; + template.file_descriptor = fds[0]; + template.private_data = + (((uword) (vui - vhost_user_main.vhost_user_interfaces)) << 8) + + q; + vui->vrings[q].kickfd_idx = unix_file_add (&unix_main, &template); + } + else + { + //When no kickfd is set, the queue is initialized as started + vui->vrings[q].kickfd_idx = ~0; + vui->vrings[q].started = 1; + } + + break; + + case VHOST_USER_SET_VRING_ERR: + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_ERR u64 %d", + vui->hw_if_index, msg.u64); + + q = (u8) (msg.u64 & 0xFF); + + if (vui->vrings[q].errfd != -1) + close (vui->vrings[q].errfd); + + if (!(msg.u64 & 0x100)) + { + if (number_of_fds != 1) + goto close_socket; + + vui->vrings[q].errfd = fds[0]; + } + else + vui->vrings[q].errfd = -1; + + break; + + case VHOST_USER_SET_VRING_BASE: + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_BASE idx %d num %d", + vui->hw_if_index, msg.state.index, msg.state.num); + + vui->vrings[msg.state.index].last_avail_idx = msg.state.num; + break; + + case VHOST_USER_GET_VRING_BASE: + DBG_SOCK ("if %d msg VHOST_USER_GET_VRING_BASE idx %d num %d", + vui->hw_if_index, msg.state.index, msg.state.num); + + if (msg.state.index >= VHOST_VRING_MAX_N) + { + DBG_SOCK ("invalid vring index VHOST_USER_GET_VRING_BASE:" + " %d >= %d", msg.state.index, VHOST_VRING_MAX_N); + goto close_socket; + } + + /* Spec says: Client must [...] stop ring upon receiving VHOST_USER_GET_VRING_BASE. */ + vhost_user_vring_close (vui, msg.state.index); + + msg.state.num = vui->vrings[msg.state.index].last_avail_idx; + msg.flags |= 4; + msg.size = sizeof (msg.state); + break; + + case VHOST_USER_NONE: + DBG_SOCK ("if %d msg VHOST_USER_NONE", vui->hw_if_index); + + break; + + case VHOST_USER_SET_LOG_BASE: + { + DBG_SOCK ("if %d msg VHOST_USER_SET_LOG_BASE", vui->hw_if_index); + + if (msg.size != sizeof (msg.log)) + { + DBG_SOCK + ("invalid msg size for VHOST_USER_SET_LOG_BASE: %d instead of %d", + msg.size, sizeof (msg.log)); + goto close_socket; + } + + if (! + (vui->protocol_features & (1 << VHOST_USER_PROTOCOL_F_LOG_SHMFD))) + { + DBG_SOCK + ("VHOST_USER_PROTOCOL_F_LOG_SHMFD not set but VHOST_USER_SET_LOG_BASE received"); + goto close_socket; + } + + fd = fds[0]; + /* align size to 2M page */ + long page_sz = get_huge_page_size (fd); + ssize_t map_sz = + (msg.log.size + msg.log.offset + page_sz) & ~(page_sz - 1); + + vui->log_base_addr = mmap (0, map_sz, PROT_READ | PROT_WRITE, + MAP_SHARED, fd, 0); + + DBG_SOCK + ("map log region addr 0 len 0x%lx off 0x%lx fd %d mapped 0x%lx", + map_sz, msg.log.offset, fd, vui->log_base_addr); + + if (vui->log_base_addr == MAP_FAILED) + { + clib_warning ("failed to map memory. errno is %d", errno); + goto close_socket; + } + + vui->log_base_addr += msg.log.offset; + vui->log_size = msg.log.size; + + msg.flags |= 4; + msg.size = sizeof (msg.u64); + + break; + } + + case VHOST_USER_SET_LOG_FD: + DBG_SOCK ("if %d msg VHOST_USER_SET_LOG_FD", vui->hw_if_index); + + break; + + case VHOST_USER_GET_PROTOCOL_FEATURES: + DBG_SOCK ("if %d msg VHOST_USER_GET_PROTOCOL_FEATURES", + vui->hw_if_index); + + msg.flags |= 4; + msg.u64 = (1 << VHOST_USER_PROTOCOL_F_LOG_SHMFD) | + (1 << VHOST_USER_PROTOCOL_F_MQ); + msg.size = sizeof (msg.u64); + break; + + case VHOST_USER_SET_PROTOCOL_FEATURES: + DBG_SOCK ("if %d msg VHOST_USER_SET_PROTOCOL_FEATURES features 0x%lx", + vui->hw_if_index, msg.u64); + + vui->protocol_features = msg.u64; + + break; + + case VHOST_USER_GET_QUEUE_NUM: + DBG_SOCK ("if %d msg VHOST_USER_GET_QUEUE_NUM", vui->hw_if_index); + msg.flags |= 4; + msg.u64 = VHOST_VRING_MAX_N; + msg.size = sizeof (msg.u64); + break; + + case VHOST_USER_SET_VRING_ENABLE: + DBG_SOCK ("if %d VHOST_USER_SET_VRING_ENABLE: %s queue %d", + vui->hw_if_index, msg.state.num ? "enable" : "disable", + msg.state.index); + if (msg.state.index >= VHOST_VRING_MAX_N) + { + DBG_SOCK ("invalid vring index VHOST_USER_SET_VRING_ENABLE:" + " %d >= %d", msg.state.index, VHOST_VRING_MAX_N); + goto close_socket; + } + + vui->vrings[msg.state.index].enabled = msg.state.num; + break; + + default: + DBG_SOCK ("unknown vhost-user message %d received. closing socket", + msg.request); + goto close_socket; + } + + /* if we need to reply */ + if (msg.flags & 4) + { + n = + send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0); + if (n != (msg.size + VHOST_USER_MSG_HDR_SZ)) + { + DBG_SOCK ("could not send message response"); + goto close_socket; + } + } + + vhost_user_update_iface_state (vui); + vlib_worker_thread_barrier_release (vlib_get_main ()); + return 0; + +close_socket: + vhost_user_if_disconnect (vui); + vhost_user_update_iface_state (vui); + vlib_worker_thread_barrier_release (vlib_get_main ()); + return 0; +} + +static clib_error_t * +vhost_user_socket_error (unix_file_t * uf) +{ + vlib_main_t *vm = vlib_get_main (); + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui = + pool_elt_at_index (vum->vhost_user_interfaces, uf->private_data); + + DBG_SOCK ("socket error on if %d", vui->sw_if_index); + vlib_worker_thread_barrier_sync (vm); + vhost_user_if_disconnect (vui); + vhost_user_rx_thread_placement (); + vlib_worker_thread_barrier_release (vm); + return 0; +} + +static clib_error_t * +vhost_user_socksvr_accept_ready (unix_file_t * uf) +{ + int client_fd, client_len; + struct sockaddr_un client; + unix_file_t template = { 0 }; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + + vui = pool_elt_at_index (vum->vhost_user_interfaces, uf->private_data); + + client_len = sizeof (client); + client_fd = accept (uf->file_descriptor, + (struct sockaddr *) &client, + (socklen_t *) & client_len); + + if (client_fd < 0) + return clib_error_return_unix (0, "accept"); + + DBG_SOCK ("New client socket for vhost interface %d", vui->sw_if_index); + template.read_function = vhost_user_socket_read; + template.error_function = vhost_user_socket_error; + template.file_descriptor = client_fd; + template.private_data = vui - vhost_user_main.vhost_user_interfaces; + vui->unix_file_index = unix_file_add (&unix_main, &template); + return 0; +} + +static clib_error_t * +vhost_user_init (vlib_main_t * vm) +{ + clib_error_t *error; + vhost_user_main_t *vum = &vhost_user_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); + vlib_thread_registration_t *tr; + uword *p; + + error = vlib_call_init_function (vm, ip4_init); + if (error) + return error; + + vum->coalesce_frames = 32; + vum->coalesce_time = 1e-3; + + vec_validate (vum->cpus, tm->n_vlib_mains - 1); + + vhost_cpu_t *cpu; + vec_foreach (cpu, vum->cpus) + { + /* This is actually not necessary as validate already zeroes it + * Just keeping the loop here for later because I am lazy. */ + cpu->rx_buffers_len = 0; + } + + /* find out which cpus will be used for input */ + vum->input_cpu_first_index = 0; + vum->input_cpu_count = 1; + p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + tr = p ? (vlib_thread_registration_t *) p[0] : 0; + + if (tr && tr->count > 0) + { + vum->input_cpu_first_index = tr->first_index; + vum->input_cpu_count = tr->count; + } + + vum->random = random_default_seed (); + + return 0; +} + +VLIB_INIT_FUNCTION (vhost_user_init); + +static clib_error_t * +vhost_user_exit (vlib_main_t * vm) +{ + /* TODO cleanup */ + return 0; +} + +VLIB_MAIN_LOOP_EXIT_FUNCTION (vhost_user_exit); + +static u8 * +format_vhost_trace (u8 * s, va_list * va) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*va, vlib_node_t *); + CLIB_UNUSED (vnet_main_t * vnm) = vnet_get_main (); + vhost_user_main_t *vum = &vhost_user_main; + vhost_trace_t *t = va_arg (*va, vhost_trace_t *); + vhost_user_intf_t *vui = pool_elt_at_index (vum->vhost_user_interfaces, + t->device_index); + + vnet_sw_interface_t *sw = vnet_get_sw_interface (vnm, vui->sw_if_index); + + uword indent = format_get_indent (s); + + s = format (s, "%U %U queue %d\n", format_white_space, indent, + format_vnet_sw_interface_name, vnm, sw, t->qid); + + s = format (s, "%U virtio flags:\n", format_white_space, indent); +#define _(n,i,st) \ + if (t->virtio_ring_flags & (1 << VIRTIO_TRACE_F_##n)) \ + s = format (s, "%U %s %s\n", format_white_space, indent, #n, st); + foreach_virtio_trace_flags +#undef _ + s = format (s, "%U virtio_net_hdr first_desc_len %u\n", + format_white_space, indent, t->first_desc_len); + + s = format (s, "%U flags 0x%02x gso_type %u\n", + format_white_space, indent, + t->hdr.hdr.flags, t->hdr.hdr.gso_type); + + if (vui->virtio_net_hdr_sz == 12) + s = format (s, "%U num_buff %u", + format_white_space, indent, t->hdr.num_buffers); + + return s; +} + +void +vhost_user_rx_trace (vhost_trace_t * t, + vhost_user_intf_t * vui, u16 qid, + vlib_buffer_t * b, vhost_user_vring_t * txvq) +{ + vhost_user_main_t *vum = &vhost_user_main; + u32 qsz_mask = txvq->qsz - 1; + u32 last_avail_idx = txvq->last_avail_idx; + u32 desc_current = txvq->avail->ring[last_avail_idx & qsz_mask]; + vring_desc_t *hdr_desc = 0; + virtio_net_hdr_mrg_rxbuf_t *hdr; + u32 hint = 0; + + memset (t, 0, sizeof (*t)); + t->device_index = vui - vum->vhost_user_interfaces; + t->qid = qid; + + hdr_desc = &txvq->desc[desc_current]; + if (txvq->desc[desc_current].flags & VIRTQ_DESC_F_INDIRECT) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_INDIRECT; + /* Header is the first here */ + hdr_desc = map_guest_mem (vui, txvq->desc[desc_current].addr, &hint); + } + if (txvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SIMPLE_CHAINED; + } + if (!(txvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT) && + !(txvq->desc[desc_current].flags & VIRTQ_DESC_F_INDIRECT)) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SINGLE_DESC; + } + + t->first_desc_len = hdr_desc ? hdr_desc->len : 0; + + if (!hdr_desc || !(hdr = map_guest_mem (vui, hdr_desc->addr, &hint))) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_MAP_ERROR; + } + else + { + u32 len = vui->virtio_net_hdr_sz; + memcpy (&t->hdr, hdr, len > hdr_desc->len ? hdr_desc->len : len); + } +} + +static inline void +vhost_user_send_call (vlib_main_t * vm, vhost_user_vring_t * vq) +{ + vhost_user_main_t *vum = &vhost_user_main; + u64 x = 1; + int fd = UNIX_GET_FD (vq->callfd_idx); + int rv __attribute__ ((unused)); + /* TODO: pay attention to rv */ + rv = write (fd, &x, sizeof (x)); + vq->n_since_last_int = 0; + vq->int_deadline = vlib_time_now (vm) + vum->coalesce_time; +} + +static_always_inline u32 +vhost_user_input_copy (vhost_user_intf_t * vui, vhost_copy_t * cpy, + u16 copy_len, u32 * map_hint) +{ + void *src0, *src1, *src2, *src3; + if (PREDICT_TRUE (copy_len >= 4)) + { + if (PREDICT_FALSE (!(src2 = map_guest_mem (vui, cpy[0].src, map_hint)))) + return 1; + if (PREDICT_FALSE (!(src3 = map_guest_mem (vui, cpy[1].src, map_hint)))) + return 1; + + while (PREDICT_TRUE (copy_len >= 4)) + { + src0 = src2; + src1 = src3; + + if (PREDICT_FALSE + (!(src2 = map_guest_mem (vui, cpy[2].src, map_hint)))) + return 1; + if (PREDICT_FALSE + (!(src3 = map_guest_mem (vui, cpy[3].src, map_hint)))) + return 1; + + CLIB_PREFETCH (src2, 64, LOAD); + CLIB_PREFETCH (src3, 64, LOAD); + + clib_memcpy ((void *) cpy[0].dst, src0, cpy[0].len); + clib_memcpy ((void *) cpy[1].dst, src1, cpy[1].len); + copy_len -= 2; + cpy += 2; + } + } + while (copy_len) + { + if (PREDICT_FALSE (!(src0 = map_guest_mem (vui, cpy->src, map_hint)))) + return 1; + clib_memcpy ((void *) cpy->dst, src0, cpy->len); + copy_len -= 1; + cpy += 1; + } + return 0; +} + +/** + * Try to discard packets from the tx ring (VPP RX path). + * Returns the number of discarded packets. + */ +u32 +vhost_user_rx_discard_packet (vlib_main_t * vm, + vhost_user_intf_t * vui, + vhost_user_vring_t * txvq, u32 discard_max) +{ + /* + * On the RX side, each packet corresponds to one descriptor + * (it is the same whether it is a shallow descriptor, chained, or indirect). + * Therefore, discarding a packet is like discarding a descriptor. + */ + u32 discarded_packets = 0; + u32 avail_idx = txvq->avail->idx; + u16 qsz_mask = txvq->qsz - 1; + while (discarded_packets != discard_max) + { + if (avail_idx == txvq->last_avail_idx) + goto out; + + u16 desc_chain_head = + txvq->avail->ring[txvq->last_avail_idx & qsz_mask]; + txvq->last_avail_idx++; + txvq->used->ring[txvq->last_used_idx & qsz_mask].id = desc_chain_head; + txvq->used->ring[txvq->last_used_idx & qsz_mask].len = 0; + vhost_user_log_dirty_ring (vui, txvq, + ring[txvq->last_used_idx & qsz_mask]); + txvq->last_used_idx++; + discarded_packets++; + } + +out: + CLIB_MEMORY_BARRIER (); + txvq->used->idx = txvq->last_used_idx; + vhost_user_log_dirty_ring (vui, txvq, idx); + return discarded_packets; +} + +/* + * In case of overflow, we need to rewind the array of allocated buffers. + */ +static void +vhost_user_input_rewind_buffers (vlib_main_t * vm, + vhost_cpu_t * cpu, vlib_buffer_t * b_head) +{ + u32 bi_current = cpu->rx_buffers[cpu->rx_buffers_len]; + vlib_buffer_t *b_current = vlib_get_buffer (vm, bi_current); + b_current->current_length = 0; + b_current->flags = 0; + while (b_current != b_head) + { + cpu->rx_buffers_len++; + bi_current = cpu->rx_buffers[cpu->rx_buffers_len]; + b_current = vlib_get_buffer (vm, bi_current); + b_current->current_length = 0; + b_current->flags = 0; + } +} + +static u32 +vhost_user_if_input (vlib_main_t * vm, + vhost_user_main_t * vum, + vhost_user_intf_t * vui, + u16 qid, vlib_node_runtime_t * node) +{ + vhost_user_vring_t *txvq = &vui->vrings[VHOST_VRING_IDX_TX (qid)]; + u16 n_rx_packets = 0; + u32 n_rx_bytes = 0; + u16 n_left; + u32 n_left_to_next, *to_next; + u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; + u32 n_trace = vlib_get_trace_count (vm, node); + u16 qsz_mask; + u32 map_hint = 0; + u16 cpu_index = os_get_cpu_number (); + u16 copy_len = 0; + + { + /* do we have pending interrupts ? */ + vhost_user_vring_t *rxvq = &vui->vrings[VHOST_VRING_IDX_RX (qid)]; + f64 now = vlib_time_now (vm); + + if ((txvq->n_since_last_int) && (txvq->int_deadline < now)) + vhost_user_send_call (vm, txvq); + + if ((rxvq->n_since_last_int) && (rxvq->int_deadline < now)) + vhost_user_send_call (vm, rxvq); + } + + if (PREDICT_FALSE (txvq->avail->flags & 0xFFFE)) + return 0; + + n_left = (u16) (txvq->avail->idx - txvq->last_avail_idx); + + /* nothing to do */ + if (PREDICT_FALSE (n_left == 0)) + return 0; + + if (PREDICT_FALSE (!vui->admin_up || !(txvq->enabled))) + { + /* + * Discard input packet if interface is admin down or vring is not + * enabled. + * "For example, for a networking device, in the disabled state + * client must not supply any new RX packets, but must process + * and discard any TX packets." + */ + vhost_user_rx_discard_packet (vm, vui, txvq, + VHOST_USER_DOWN_DISCARD_COUNT); + return 0; + } + + if (PREDICT_FALSE (n_left == txvq->qsz)) + { + /* + * Informational error logging when VPP is not + * receiving packets fast enough. + */ + vlib_error_count (vm, node->node_index, + VHOST_USER_INPUT_FUNC_ERROR_FULL_RX_QUEUE, 1); + } + + qsz_mask = txvq->qsz - 1; + + if (n_left > VLIB_FRAME_SIZE) + n_left = VLIB_FRAME_SIZE; + + /* + * For small packets (<2kB), we will not need more than one vlib buffer + * per packet. In case packets are bigger, we will just yeld at some point + * in the loop and come back later. This is not an issue as for big packet, + * processing cost really comes from the memory copy. + */ + if (PREDICT_FALSE (vum->cpus[cpu_index].rx_buffers_len < n_left + 1)) + { + u32 curr_len = vum->cpus[cpu_index].rx_buffers_len; + vum->cpus[cpu_index].rx_buffers_len += + vlib_buffer_alloc_from_free_list (vm, + vum->cpus[cpu_index].rx_buffers + + curr_len, + VHOST_USER_RX_BUFFERS_N - curr_len, + VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); + + if (PREDICT_FALSE + (vum->cpus[cpu_index].rx_buffers_len < + VHOST_USER_RX_BUFFER_STARVATION)) + { + /* In case of buffer starvation, discard some packets from the queue + * and log the event. + * We keep doing best effort for the remaining packets. */ + u32 flush = (n_left + 1 > vum->cpus[cpu_index].rx_buffers_len) ? + n_left + 1 - vum->cpus[cpu_index].rx_buffers_len : 1; + flush = vhost_user_rx_discard_packet (vm, vui, txvq, flush); + + n_left -= flush; + vlib_increment_simple_counter (vnet_main. + interface_main.sw_if_counters + + VNET_INTERFACE_COUNTER_DROP, + os_get_cpu_number (), + vui->sw_if_index, flush); + + vlib_error_count (vm, vhost_user_input_node.index, + VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER, flush); + } + } + + while (n_left > 0) + { + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_left > 0 && n_left_to_next > 0) + { + vlib_buffer_t *b_head, *b_current; + u32 bi_current; + u16 desc_current; + u32 desc_data_offset; + vring_desc_t *desc_table = txvq->desc; + + if (PREDICT_FALSE (vum->cpus[cpu_index].rx_buffers_len <= 1)) + { + /* Not enough rx_buffers + * Note: We yeld on 1 so we don't need to do an additional + * check for the next buffer prefetch. + */ + n_left = 0; + break; + } + + desc_current = txvq->avail->ring[txvq->last_avail_idx & qsz_mask]; + vum->cpus[cpu_index].rx_buffers_len--; + bi_current = (vum->cpus[cpu_index].rx_buffers) + [vum->cpus[cpu_index].rx_buffers_len]; + b_head = b_current = vlib_get_buffer (vm, bi_current); + to_next[0] = bi_current; //We do that now so we can forget about bi_current + to_next++; + n_left_to_next--; + + vlib_prefetch_buffer_with_index (vm, + (vum->cpus[cpu_index].rx_buffers) + [vum->cpus[cpu_index]. + rx_buffers_len - 1], LOAD); + + /* Just preset the used descriptor id and length for later */ + txvq->used->ring[txvq->last_used_idx & qsz_mask].id = desc_current; + txvq->used->ring[txvq->last_used_idx & qsz_mask].len = 0; + vhost_user_log_dirty_ring (vui, txvq, + ring[txvq->last_used_idx & qsz_mask]); + + /* The buffer should already be initialized */ + b_head->total_length_not_including_first_buffer = 0; + b_head->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID; + + if (PREDICT_FALSE (n_trace)) + { + //TODO: next_index is not exactly known at that point + vlib_trace_buffer (vm, node, next_index, b_head, + /* follow_chain */ 0); + vhost_trace_t *t0 = + vlib_add_trace (vm, node, b_head, sizeof (t0[0])); + vhost_user_rx_trace (t0, vui, qid, b_head, txvq); + n_trace--; + vlib_set_trace_count (vm, node, n_trace); + } + + /* This depends on the setup but is very consistent + * So I think the CPU branch predictor will make a pretty good job + * at optimizing the decision. */ + if (txvq->desc[desc_current].flags & VIRTQ_DESC_F_INDIRECT) + { + desc_table = map_guest_mem (vui, txvq->desc[desc_current].addr, + &map_hint); + desc_current = 0; + if (PREDICT_FALSE (desc_table == 0)) + { + //FIXME: Handle error by shutdown the queue + goto out; + } + } + + if (PREDICT_TRUE (vui->is_any_layout) || + (!(desc_table[desc_current].flags & VIRTQ_DESC_F_NEXT))) + { + /* ANYLAYOUT or single buffer */ + desc_data_offset = vui->virtio_net_hdr_sz; + } + else + { + /* CSR case without ANYLAYOUT, skip 1st buffer */ + desc_data_offset = desc_table[desc_current].len; + } + + while (1) + { + /* Get more input if necessary. Or end of packet. */ + if (desc_data_offset == desc_table[desc_current].len) + { + if (PREDICT_FALSE (desc_table[desc_current].flags & + VIRTQ_DESC_F_NEXT)) + { + desc_current = desc_table[desc_current].next; + desc_data_offset = 0; + } + else + { + goto out; + } + } + + /* Get more output if necessary. Or end of packet. */ + if (PREDICT_FALSE + (b_current->current_length == VLIB_BUFFER_DATA_SIZE)) + { + if (PREDICT_FALSE + (vum->cpus[cpu_index].rx_buffers_len == 0)) + { + /* + * Checking if there are some left buffers. + * If not, just rewind the used buffers and stop. + * Note: Scheduled copies are not cancelled. This is + * not an issue as they would still be valid. Useless, + * but valid. + */ + vhost_user_input_rewind_buffers (vm, + &vum->cpus[cpu_index], + b_head); + n_left = 0; + goto stop; + } + + /* Get next output */ + vum->cpus[cpu_index].rx_buffers_len--; + u32 bi_next = + (vum->cpus[cpu_index].rx_buffers)[vum->cpus + [cpu_index].rx_buffers_len]; + b_current->next_buffer = bi_next; + b_current->flags |= VLIB_BUFFER_NEXT_PRESENT; + bi_current = bi_next; + b_current = vlib_get_buffer (vm, bi_current); + } + + /* Prepare a copy order executed later for the data */ + vhost_copy_t *cpy = &vum->cpus[cpu_index].copy[copy_len]; + copy_len++; + u32 desc_data_l = + desc_table[desc_current].len - desc_data_offset; + cpy->len = VLIB_BUFFER_DATA_SIZE - b_current->current_length; + cpy->len = (cpy->len > desc_data_l) ? desc_data_l : cpy->len; + cpy->dst = (uword) vlib_buffer_get_current (b_current); + cpy->src = desc_table[desc_current].addr + desc_data_offset; + + desc_data_offset += cpy->len; + + b_current->current_length += cpy->len; + b_head->total_length_not_including_first_buffer += cpy->len; + } + + out: + CLIB_PREFETCH (&n_left, sizeof (n_left), LOAD); + + n_rx_bytes += b_head->total_length_not_including_first_buffer; + n_rx_packets++; + + b_head->total_length_not_including_first_buffer -= + b_head->current_length; + + /* consume the descriptor and return it as used */ + txvq->last_avail_idx++; + txvq->last_used_idx++; + + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b_head); + + vnet_buffer (b_head)->sw_if_index[VLIB_RX] = vui->sw_if_index; + vnet_buffer (b_head)->sw_if_index[VLIB_TX] = (u32) ~ 0; + b_head->error = 0; + + { + u32 next0 = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; + + /* redirect if feature path enabled */ + vnet_feature_start_device_input_x1 (vui->sw_if_index, &next0, + b_head, 0); + + u32 bi = to_next[-1]; //Cannot use to_next[-1] in the macro + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, + bi, next0); + } + + n_left--; + + /* + * Although separating memory copies from virtio ring parsing + * is beneficial, we can offer to perform the copies from time + * to time in order to free some space in the ring. + */ + if (PREDICT_FALSE (copy_len >= VHOST_USER_RX_COPY_THRESHOLD)) + { + if (PREDICT_FALSE + (vhost_user_input_copy (vui, vum->cpus[cpu_index].copy, + copy_len, &map_hint))) + { + clib_warning + ("Memory mapping error on interface hw_if_index=%d " + "(Shutting down - Switch interface down and up to restart)", + vui->hw_if_index); + vui->admin_up = 0; + copy_len = 0; + break; + } + copy_len = 0; + + /* give buffers back to driver */ + CLIB_MEMORY_BARRIER (); + txvq->used->idx = txvq->last_used_idx; + vhost_user_log_dirty_ring (vui, txvq, idx); + } + } + stop: + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + + /* Do the memory copies */ + if (PREDICT_FALSE + (vhost_user_input_copy (vui, vum->cpus[cpu_index].copy, + copy_len, &map_hint))) + { + clib_warning ("Memory mapping error on interface hw_if_index=%d " + "(Shutting down - Switch interface down and up to restart)", + vui->hw_if_index); + vui->admin_up = 0; + } + + /* give buffers back to driver */ + CLIB_MEMORY_BARRIER (); + txvq->used->idx = txvq->last_used_idx; + vhost_user_log_dirty_ring (vui, txvq, idx); + + /* interrupt (call) handling */ + if ((txvq->callfd_idx != ~0) && !(txvq->avail->flags & 1)) + { + txvq->n_since_last_int += n_rx_packets; + + if (txvq->n_since_last_int > vum->coalesce_frames) + vhost_user_send_call (vm, txvq); + } + + /* increase rx counters */ + vlib_increment_combined_counter + (vnet_main.interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, + os_get_cpu_number (), vui->sw_if_index, n_rx_packets, n_rx_bytes); + + return n_rx_packets; +} + +static uword +vhost_user_input (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * f) +{ + vhost_user_main_t *vum = &vhost_user_main; + uword n_rx_packets = 0; + u32 cpu_index = os_get_cpu_number (); + + + vhost_iface_and_queue_t *vhiq; + vec_foreach (vhiq, vum->cpus[cpu_index].rx_queues) + { + vhost_user_intf_t *vui = + &vum->vhost_user_interfaces[vhiq->vhost_iface_index]; + n_rx_packets += vhost_user_if_input (vm, vum, vui, vhiq->qid, node); + } + + return n_rx_packets; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (vhost_user_input_node) = { + .function = vhost_user_input, + .type = VLIB_NODE_TYPE_INPUT, + .name = "vhost-user-input", + .sibling_of = "device-input", + + /* Will be enabled if/when hardware is detected. */ + .state = VLIB_NODE_STATE_DISABLED, + + .format_buffer = format_ethernet_header_with_length, + .format_trace = format_vhost_trace, + + .n_errors = VHOST_USER_INPUT_FUNC_N_ERROR, + .error_strings = vhost_user_input_func_error_strings, +}; + +VLIB_NODE_FUNCTION_MULTIARCH (vhost_user_input_node, vhost_user_input) +/* *INDENT-ON* */ + + +void +vhost_user_tx_trace (vhost_trace_t * t, + vhost_user_intf_t * vui, u16 qid, + vlib_buffer_t * b, vhost_user_vring_t * rxvq) +{ + vhost_user_main_t *vum = &vhost_user_main; + u32 qsz_mask = rxvq->qsz - 1; + u32 last_avail_idx = rxvq->last_avail_idx; + u32 desc_current = rxvq->avail->ring[last_avail_idx & qsz_mask]; + vring_desc_t *hdr_desc = 0; + u32 hint = 0; + + memset (t, 0, sizeof (*t)); + t->device_index = vui - vum->vhost_user_interfaces; + t->qid = qid; + + hdr_desc = &rxvq->desc[desc_current]; + if (rxvq->desc[desc_current].flags & VIRTQ_DESC_F_INDIRECT) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_INDIRECT; + /* Header is the first here */ + hdr_desc = map_guest_mem (vui, rxvq->desc[desc_current].addr, &hint); + } + if (rxvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SIMPLE_CHAINED; + } + if (!(rxvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT) && + !(rxvq->desc[desc_current].flags & VIRTQ_DESC_F_INDIRECT)) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SINGLE_DESC; + } + + t->first_desc_len = hdr_desc ? hdr_desc->len : 0; +} + +static_always_inline u32 +vhost_user_tx_copy (vhost_user_intf_t * vui, vhost_copy_t * cpy, + u16 copy_len, u32 * map_hint) +{ + void *dst0, *dst1, *dst2, *dst3; + if (PREDICT_TRUE (copy_len >= 4)) + { + if (PREDICT_FALSE (!(dst2 = map_guest_mem (vui, cpy[0].dst, map_hint)))) + return 1; + if (PREDICT_FALSE (!(dst3 = map_guest_mem (vui, cpy[1].dst, map_hint)))) + return 1; + while (PREDICT_TRUE (copy_len >= 4)) + { + dst0 = dst2; + dst1 = dst3; + + if (PREDICT_FALSE + (!(dst2 = map_guest_mem (vui, cpy[2].dst, map_hint)))) + return 1; + if (PREDICT_FALSE + (!(dst3 = map_guest_mem (vui, cpy[3].dst, map_hint)))) + return 1; + + CLIB_PREFETCH ((void *) cpy[2].src, 64, LOAD); + CLIB_PREFETCH ((void *) cpy[3].src, 64, LOAD); + + clib_memcpy (dst0, (void *) cpy[0].src, cpy[0].len); + clib_memcpy (dst1, (void *) cpy[1].src, cpy[1].len); + + vhost_user_log_dirty_pages_2 (vui, cpy[0].dst, cpy[0].len, 1); + vhost_user_log_dirty_pages_2 (vui, cpy[1].dst, cpy[1].len, 1); + copy_len -= 2; + cpy += 2; + } + } + while (copy_len) + { + if (PREDICT_FALSE (!(dst0 = map_guest_mem (vui, cpy->dst, map_hint)))) + return 1; + clib_memcpy (dst0, (void *) cpy->src, cpy->len); + vhost_user_log_dirty_pages_2 (vui, cpy->dst, cpy->len, 1); + copy_len -= 1; + cpy += 1; + } + return 0; +} + + +static uword +vhost_user_tx (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * frame) +{ + u32 *buffers = vlib_frame_args (frame); + u32 n_left = frame->n_vectors; + vhost_user_main_t *vum = &vhost_user_main; + vnet_interface_output_runtime_t *rd = (void *) node->runtime_data; + vhost_user_intf_t *vui = + pool_elt_at_index (vum->vhost_user_interfaces, rd->dev_instance); + u32 qid = ~0; + vhost_user_vring_t *rxvq; + u16 qsz_mask; + u8 error; + u32 cpu_index = os_get_cpu_number (); + u32 map_hint = 0; + u8 retry = 8; + u16 copy_len; + u16 tx_headers_len; + + if (PREDICT_FALSE (!vui->admin_up)) + { + error = VHOST_USER_TX_FUNC_ERROR_DOWN; + goto done3; + } + + if (PREDICT_FALSE (!vui->is_up)) + { + error = VHOST_USER_TX_FUNC_ERROR_NOT_READY; + goto done3; + } + + qid = + VHOST_VRING_IDX_RX (*vec_elt_at_index + (vui->per_cpu_tx_qid, os_get_cpu_number ())); + rxvq = &vui->vrings[qid]; + if (PREDICT_FALSE (vui->use_tx_spinlock)) + vhost_user_vring_lock (vui, qid); + + qsz_mask = rxvq->qsz - 1; /* qsz is always power of 2 */ + +retry: + error = VHOST_USER_TX_FUNC_ERROR_NONE; + tx_headers_len = 0; + copy_len = 0; + while (n_left > 0) + { + vlib_buffer_t *b0, *current_b0; + u16 desc_head, desc_index, desc_len; + vring_desc_t *desc_table; + uword buffer_map_addr; + u32 buffer_len; + u16 bytes_left; + + if (PREDICT_TRUE (n_left > 1)) + vlib_prefetch_buffer_with_index (vm, buffers[1], LOAD); + + b0 = vlib_get_buffer (vm, buffers[0]); + + if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED)) + { + vum->cpus[cpu_index].current_trace = + vlib_add_trace (vm, node, b0, + sizeof (*vum->cpus[cpu_index].current_trace)); + vhost_user_tx_trace (vum->cpus[cpu_index].current_trace, + vui, qid / 2, b0, rxvq); + } + + if (PREDICT_FALSE (rxvq->last_avail_idx == rxvq->avail->idx)) + { + error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; + goto done; + } + + desc_table = rxvq->desc; + desc_head = desc_index = + rxvq->avail->ring[rxvq->last_avail_idx & qsz_mask]; + + /* Go deeper in case of indirect descriptor + * I don't know of any driver providing indirect for RX. */ + if (PREDICT_FALSE (rxvq->desc[desc_head].flags & VIRTQ_DESC_F_INDIRECT)) + { + if (PREDICT_FALSE + (rxvq->desc[desc_head].len < sizeof (vring_desc_t))) + { + error = VHOST_USER_TX_FUNC_ERROR_INDIRECT_OVERFLOW; + goto done; + } + if (PREDICT_FALSE + (!(desc_table = + map_guest_mem (vui, rxvq->desc[desc_index].addr, + &map_hint)))) + { + error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; + goto done; + } + desc_index = 0; + } + + desc_len = vui->virtio_net_hdr_sz; + buffer_map_addr = desc_table[desc_index].addr; + buffer_len = desc_table[desc_index].len; + + { + // Get a header from the header array + virtio_net_hdr_mrg_rxbuf_t *hdr = + &vum->cpus[cpu_index].tx_headers[tx_headers_len]; + tx_headers_len++; + hdr->hdr.flags = 0; + hdr->hdr.gso_type = 0; + hdr->num_buffers = 1; //This is local, no need to check + + // Prepare a copy order executed later for the header + vhost_copy_t *cpy = &vum->cpus[cpu_index].copy[copy_len]; + copy_len++; + cpy->len = vui->virtio_net_hdr_sz; + cpy->dst = buffer_map_addr; + cpy->src = (uword) hdr; + } + + buffer_map_addr += vui->virtio_net_hdr_sz; + buffer_len -= vui->virtio_net_hdr_sz; + bytes_left = b0->current_length; + current_b0 = b0; + while (1) + { + if (buffer_len == 0) + { //Get new output + if (desc_table[desc_index].flags & VIRTQ_DESC_F_NEXT) + { + //Next one is chained + desc_index = desc_table[desc_index].next; + buffer_map_addr = desc_table[desc_index].addr; + buffer_len = desc_table[desc_index].len; + } + else if (vui->virtio_net_hdr_sz == 12) //MRG is available + { + virtio_net_hdr_mrg_rxbuf_t *hdr = + &vum->cpus[cpu_index].tx_headers[tx_headers_len - 1]; + + //Move from available to used buffer + rxvq->used->ring[rxvq->last_used_idx & qsz_mask].id = + desc_head; + rxvq->used->ring[rxvq->last_used_idx & qsz_mask].len = + desc_len; + vhost_user_log_dirty_ring (vui, rxvq, + ring[rxvq->last_used_idx & + qsz_mask]); + + rxvq->last_avail_idx++; + rxvq->last_used_idx++; + hdr->num_buffers++; + desc_len = 0; + + if (PREDICT_FALSE + (rxvq->last_avail_idx == rxvq->avail->idx)) + { + //Dequeue queued descriptors for this packet + rxvq->last_used_idx -= hdr->num_buffers - 1; + rxvq->last_avail_idx -= hdr->num_buffers - 1; + error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; + goto done; + } + + desc_table = rxvq->desc; + desc_head = desc_index = + rxvq->avail->ring[rxvq->last_avail_idx & qsz_mask]; + if (PREDICT_FALSE + (rxvq->desc[desc_head].flags & VIRTQ_DESC_F_INDIRECT)) + { + //It is seriously unlikely that a driver will put indirect descriptor + //after non-indirect descriptor. + if (PREDICT_FALSE + (rxvq->desc[desc_head].len < sizeof (vring_desc_t))) + { + error = VHOST_USER_TX_FUNC_ERROR_INDIRECT_OVERFLOW; + goto done; + } + if (PREDICT_FALSE + (!(desc_table = + map_guest_mem (vui, + rxvq->desc[desc_index].addr, + &map_hint)))) + { + error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; + goto done; + } + desc_index = 0; + } + buffer_map_addr = desc_table[desc_index].addr; + buffer_len = desc_table[desc_index].len; + } + else + { + error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOMRG; + goto done; + } + } + + { + vhost_copy_t *cpy = &vum->cpus[cpu_index].copy[copy_len]; + copy_len++; + cpy->len = bytes_left; + cpy->len = (cpy->len > buffer_len) ? buffer_len : cpy->len; + cpy->dst = buffer_map_addr; + cpy->src = (uword) vlib_buffer_get_current (current_b0) + + current_b0->current_length - bytes_left; + + bytes_left -= cpy->len; + buffer_len -= cpy->len; + buffer_map_addr += cpy->len; + desc_len += cpy->len; + + CLIB_PREFETCH (&rxvq->desc, CLIB_CACHE_LINE_BYTES, LOAD); + } + + // Check if vlib buffer has more data. If not, get more or break. + if (PREDICT_TRUE (!bytes_left)) + { + if (PREDICT_FALSE + (current_b0->flags & VLIB_BUFFER_NEXT_PRESENT)) + { + current_b0 = vlib_get_buffer (vm, current_b0->next_buffer); + bytes_left = current_b0->current_length; + } + else + { + //End of packet + break; + } + } + } + + //Move from available to used ring + rxvq->used->ring[rxvq->last_used_idx & qsz_mask].id = desc_head; + rxvq->used->ring[rxvq->last_used_idx & qsz_mask].len = desc_len; + vhost_user_log_dirty_ring (vui, rxvq, + ring[rxvq->last_used_idx & qsz_mask]); + rxvq->last_avail_idx++; + rxvq->last_used_idx++; + + if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED)) + { + vum->cpus[cpu_index].current_trace->hdr = + vum->cpus[cpu_index].tx_headers[tx_headers_len - 1]; + } + + n_left--; //At the end for error counting when 'goto done' is invoked + buffers++; + } + +done: + //Do the memory copies + if (PREDICT_FALSE + (vhost_user_tx_copy (vui, vum->cpus[cpu_index].copy, + copy_len, &map_hint))) + { + clib_warning ("Memory mapping error on interface hw_if_index=%d " + "(Shutting down - Switch interface down and up to restart)", + vui->hw_if_index); + vui->admin_up = 0; + } + + CLIB_MEMORY_BARRIER (); + rxvq->used->idx = rxvq->last_used_idx; + vhost_user_log_dirty_ring (vui, rxvq, idx); + + /* + * When n_left is set, error is always set to something too. + * In case error is due to lack of remaining buffers, we go back up and + * retry. + * The idea is that it is better to waste some time on packets + * that have been processed already than dropping them and get + * more fresh packets with a good likelyhood that they will be dropped too. + * This technique also gives more time to VM driver to pick-up packets. + * In case the traffic flows from physical to virtual interfaces, this + * technique will end-up leveraging the physical NIC buffer in order to + * absorb the VM's CPU jitter. + */ + if (n_left && (error == VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF) && retry) + { + retry--; + goto retry; + } + + /* interrupt (call) handling */ + if ((rxvq->callfd_idx != ~0) && !(rxvq->avail->flags & 1)) + { + rxvq->n_since_last_int += frame->n_vectors - n_left; + + if (rxvq->n_since_last_int > vum->coalesce_frames) + vhost_user_send_call (vm, rxvq); + } + + vhost_user_vring_unlock (vui, qid); + +done3: + if (PREDICT_FALSE (n_left && error != VHOST_USER_TX_FUNC_ERROR_NONE)) + { + vlib_error_count (vm, node->node_index, error, n_left); + vlib_increment_simple_counter + (vnet_main.interface_main.sw_if_counters + + VNET_INTERFACE_COUNTER_DROP, + os_get_cpu_number (), vui->sw_if_index, n_left); + } + + vlib_buffer_free (vm, vlib_frame_args (frame), frame->n_vectors); + return frame->n_vectors; +} + +static clib_error_t * +vhost_user_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, + u32 flags) +{ + vnet_hw_interface_t *hif = vnet_get_hw_interface (vnm, hw_if_index); + uword is_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui = + pool_elt_at_index (vum->vhost_user_interfaces, hif->dev_instance); + + vui->admin_up = is_up; + + if (is_up) + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, + VNET_HW_INTERFACE_FLAG_LINK_UP); + + return /* no error */ 0; +} + +/* *INDENT-OFF* */ +VNET_DEVICE_CLASS (vhost_user_dev_class,static) = { + .name = "vhost-user", + .tx_function = vhost_user_tx, + .tx_function_n_errors = VHOST_USER_TX_FUNC_N_ERROR, + .tx_function_error_strings = vhost_user_tx_func_error_strings, + .format_device_name = format_vhost_user_interface_name, + .name_renumber = vhost_user_name_renumber, + .admin_up_down_function = vhost_user_interface_admin_up_down, + .format_tx_trace = format_vhost_trace, +}; + +VLIB_DEVICE_TX_FUNCTION_MULTIARCH (vhost_user_dev_class, + vhost_user_tx) +/* *INDENT-ON* */ + +static uword +vhost_user_process (vlib_main_t * vm, + vlib_node_runtime_t * rt, vlib_frame_t * f) +{ + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + struct sockaddr_un sun; + int sockfd; + unix_file_t template = { 0 }; + f64 timeout = 3153600000.0 /* 100 years */ ; + uword *event_data = 0; + + sockfd = socket (AF_UNIX, SOCK_STREAM, 0); + sun.sun_family = AF_UNIX; + template.read_function = vhost_user_socket_read; + template.error_function = vhost_user_socket_error; + + if (sockfd < 0) + return 0; + + while (1) + { + vlib_process_wait_for_event_or_clock (vm, timeout); + vlib_process_get_events (vm, &event_data); + vec_reset_length (event_data); + + timeout = 3.0; + + /* *INDENT-OFF* */ + pool_foreach (vui, vum->vhost_user_interfaces, { + + if (vui->unix_server_index == ~0) { //Nothing to do for server sockets + if (vui->unix_file_index == ~0) + { + /* try to connect */ + strncpy (sun.sun_path, (char *) vui->sock_filename, + sizeof (sun.sun_path) - 1); + + if (connect (sockfd, (struct sockaddr *) &sun, + sizeof (struct sockaddr_un)) == 0) + { + vui->sock_errno = 0; + template.file_descriptor = sockfd; + template.private_data = + vui - vhost_user_main.vhost_user_interfaces; + vui->unix_file_index = unix_file_add (&unix_main, &template); + + //Re-open for next connect + if ((sockfd = socket (AF_UNIX, SOCK_STREAM, 0)) < 0) { + clib_warning("Critical: Could not open unix socket"); + return 0; + } + } + else + { + vui->sock_errno = errno; + } + } + else + { + /* check if socket is alive */ + int error = 0; + socklen_t len = sizeof (error); + int fd = UNIX_GET_FD(vui->unix_file_index); + int retval = + getsockopt (fd, SOL_SOCKET, SO_ERROR, &error, &len); + + if (retval) + { + DBG_SOCK ("getsockopt returned %d", retval); + vhost_user_if_disconnect (vui); + } + } + } + }); + /* *INDENT-ON* */ + } + return 0; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (vhost_user_process_node,static) = { + .function = vhost_user_process, + .type = VLIB_NODE_TYPE_PROCESS, + .name = "vhost-user-process", +}; +/* *INDENT-ON* */ + +/** + * Disables and reset interface structure. + * It can then be either init again, or removed from used interfaces. + */ +static void +vhost_user_term_if (vhost_user_intf_t * vui) +{ + // Delete configured thread pinning + vec_reset_length (vui->workers); + // disconnect interface sockets + vhost_user_if_disconnect (vui); + vhost_user_update_iface_state (vui); + + if (vui->unix_server_index != ~0) + { + //Close server socket + unix_file_t *uf = pool_elt_at_index (unix_main.file_pool, + vui->unix_server_index); + unix_file_del (&unix_main, uf); + vui->unix_server_index = ~0; + } +} + +int +vhost_user_delete_if (vnet_main_t * vnm, vlib_main_t * vm, u32 sw_if_index) +{ + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + int rv = 0; + vnet_hw_interface_t *hwif; + + if (!(hwif = vnet_get_sup_hw_interface (vnm, sw_if_index)) || + hwif->dev_class_index != vhost_user_dev_class.index) + return VNET_API_ERROR_INVALID_SW_IF_INDEX; + + DBG_SOCK ("Deleting vhost-user interface %s (instance %d)", + hwif->name, hwif->dev_instance); + + vui = pool_elt_at_index (vum->vhost_user_interfaces, hwif->dev_instance); + + // Disable and reset interface + vhost_user_term_if (vui); + + // Back to pool + pool_put (vum->vhost_user_interfaces, vui); + + // Reset renumbered iface + if (hwif->dev_instance < + vec_len (vum->show_dev_instance_by_real_dev_instance)) + vum->show_dev_instance_by_real_dev_instance[hwif->dev_instance] = ~0; + + // Delete ethernet interface + ethernet_delete_interface (vnm, vui->hw_if_index); + return rv; +} + +/** + * Open server unix socket on specified sock_filename. + */ +static int +vhost_user_init_server_sock (const char *sock_filename, int *sock_fd) +{ + int rv = 0; + struct sockaddr_un un = { }; + int fd; + /* create listening socket */ + if ((fd = socket (AF_UNIX, SOCK_STREAM, 0)) < 0) + return VNET_API_ERROR_SYSCALL_ERROR_1; + + un.sun_family = AF_UNIX; + strncpy ((char *) un.sun_path, (char *) sock_filename, + sizeof (un.sun_path) - 1); + + /* remove if exists */ + unlink ((char *) sock_filename); + + if (bind (fd, (struct sockaddr *) &un, sizeof (un)) == -1) + { + rv = VNET_API_ERROR_SYSCALL_ERROR_2; + goto error; + } + + if (listen (fd, 1) == -1) + { + rv = VNET_API_ERROR_SYSCALL_ERROR_3; + goto error; + } + + *sock_fd = fd; + return 0; + +error: + close (fd); + return rv; +} + +/** + * Create ethernet interface for vhost user interface. + */ +static void +vhost_user_create_ethernet (vnet_main_t * vnm, vlib_main_t * vm, + vhost_user_intf_t * vui, u8 * hwaddress) +{ + vhost_user_main_t *vum = &vhost_user_main; + u8 hwaddr[6]; + clib_error_t *error; + + /* create hw and sw interface */ + if (hwaddress) + { + clib_memcpy (hwaddr, hwaddress, 6); + } + else + { + random_u32 (&vum->random); + clib_memcpy (hwaddr + 2, &vum->random, sizeof (vum->random)); + hwaddr[0] = 2; + hwaddr[1] = 0xfe; + } + + error = ethernet_register_interface + (vnm, + vhost_user_dev_class.index, + vui - vum->vhost_user_interfaces /* device instance */ , + hwaddr /* ethernet address */ , + &vui->hw_if_index, 0 /* flag change */ ); + + if (error) + clib_error_report (error); + + vnet_hw_interface_t *hi = vnet_get_hw_interface (vnm, vui->hw_if_index); + hi->max_l3_packet_bytes[VLIB_RX] = hi->max_l3_packet_bytes[VLIB_TX] = 9000; +} + +/* + * Initialize vui with specified attributes + */ +static void +vhost_user_vui_init (vnet_main_t * vnm, + vhost_user_intf_t * vui, + int server_sock_fd, + const char *sock_filename, + u64 feature_mask, u32 * sw_if_index) +{ + vnet_sw_interface_t *sw; + sw = vnet_get_hw_sw_interface (vnm, vui->hw_if_index); + int q; + + if (server_sock_fd != -1) + { + unix_file_t template = { 0 }; + template.read_function = vhost_user_socksvr_accept_ready; + template.file_descriptor = server_sock_fd; + template.private_data = vui - vhost_user_main.vhost_user_interfaces; //hw index + vui->unix_server_index = unix_file_add (&unix_main, &template); + } + else + { + vui->unix_server_index = ~0; + } + + vui->sw_if_index = sw->sw_if_index; + strncpy (vui->sock_filename, sock_filename, + ARRAY_LEN (vui->sock_filename) - 1); + vui->sock_errno = 0; + vui->is_up = 0; + vui->feature_mask = feature_mask; + vui->unix_file_index = ~0; + vui->log_base_addr = 0; + + for (q = 0; q < VHOST_VRING_MAX_N; q++) + vhost_user_vring_init (vui, q); + + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); + + if (sw_if_index) + *sw_if_index = vui->sw_if_index; + + for (q = 0; q < VHOST_VRING_MAX_N; q++) + { + vui->vring_locks[q] = clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES, + CLIB_CACHE_LINE_BYTES); + memset ((void *) vui->vring_locks[q], 0, CLIB_CACHE_LINE_BYTES); + } + + vec_validate (vui->per_cpu_tx_qid, + vlib_get_thread_main ()->n_vlib_mains - 1); + vhost_user_tx_thread_placement (vui); +} + +int +vhost_user_create_if (vnet_main_t * vnm, vlib_main_t * vm, + const char *sock_filename, + u8 is_server, + u32 * sw_if_index, + u64 feature_mask, + u8 renumber, u32 custom_dev_instance, u8 * hwaddr) +{ + vhost_user_intf_t *vui = NULL; + u32 sw_if_idx = ~0; + int rv = 0; + int server_sock_fd = -1; + + if (is_server) + { + if ((rv = + vhost_user_init_server_sock (sock_filename, &server_sock_fd)) != 0) + { + return rv; + } + } + + pool_get (vhost_user_main.vhost_user_interfaces, vui); + + vhost_user_create_ethernet (vnm, vm, vui, hwaddr); + vhost_user_vui_init (vnm, vui, server_sock_fd, sock_filename, + feature_mask, &sw_if_idx); + + if (renumber) + vnet_interface_name_renumber (sw_if_idx, custom_dev_instance); + + if (sw_if_index) + *sw_if_index = sw_if_idx; + + // Process node must connect + vlib_process_signal_event (vm, vhost_user_process_node.index, 0, 0); + return rv; +} + +int +vhost_user_modify_if (vnet_main_t * vnm, vlib_main_t * vm, + const char *sock_filename, + u8 is_server, + u32 sw_if_index, + u64 feature_mask, u8 renumber, u32 custom_dev_instance) +{ + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui = NULL; + u32 sw_if_idx = ~0; + int server_sock_fd = -1; + int rv = 0; + vnet_hw_interface_t *hwif; + + if (!(hwif = vnet_get_sup_hw_interface (vnm, sw_if_index)) || + hwif->dev_class_index != vhost_user_dev_class.index) + return VNET_API_ERROR_INVALID_SW_IF_INDEX; + + vui = vec_elt_at_index (vum->vhost_user_interfaces, hwif->dev_instance); + + // First try to open server socket + if (is_server) + if ((rv = vhost_user_init_server_sock (sock_filename, + &server_sock_fd)) != 0) + return rv; + + vhost_user_term_if (vui); + vhost_user_vui_init (vnm, vui, server_sock_fd, + sock_filename, feature_mask, &sw_if_idx); + + if (renumber) + vnet_interface_name_renumber (sw_if_idx, custom_dev_instance); + + // Process node must connect + vlib_process_signal_event (vm, vhost_user_process_node.index, 0, 0); + return rv; +} + +clib_error_t * +vhost_user_connect_command_fn (vlib_main_t * vm, + unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + u8 *sock_filename = NULL; + u32 sw_if_index; + u8 is_server = 0; + u64 feature_mask = (u64) ~ (0ULL); + u8 renumber = 0; + u32 custom_dev_instance = ~0; + u8 hwaddr[6]; + u8 *hw = NULL; + + /* Get a line of input. */ + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "socket %s", &sock_filename)) + ; + else if (unformat (line_input, "server")) + is_server = 1; + else if (unformat (line_input, "feature-mask 0x%llx", &feature_mask)) + ; + else + if (unformat + (line_input, "hwaddr %U", unformat_ethernet_address, hwaddr)) + hw = hwaddr; + else if (unformat (line_input, "renumber %d", &custom_dev_instance)) + { + renumber = 1; + } + else + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + unformat_free (line_input); + + vnet_main_t *vnm = vnet_get_main (); + + int rv; + if ((rv = vhost_user_create_if (vnm, vm, (char *) sock_filename, + is_server, &sw_if_index, feature_mask, + renumber, custom_dev_instance, hw))) + { + vec_free (sock_filename); + return clib_error_return (0, "vhost_user_create_if returned %d", rv); + } + + vec_free (sock_filename); + vlib_cli_output (vm, "%U\n", format_vnet_sw_if_index_name, vnet_get_main (), + sw_if_index); + return 0; +} + +clib_error_t * +vhost_user_delete_command_fn (vlib_main_t * vm, + unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + u32 sw_if_index = ~0; + vnet_main_t *vnm = vnet_get_main (); + + /* Get a line of input. */ + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "sw_if_index %d", &sw_if_index)) + ; + else if (unformat + (line_input, "%U", unformat_vnet_sw_interface, vnm, + &sw_if_index)) + { + vnet_hw_interface_t *hwif = + vnet_get_sup_hw_interface (vnm, sw_if_index); + if (hwif == NULL || + vhost_user_dev_class.index != hwif->dev_class_index) + return clib_error_return (0, "Not a vhost interface"); + } + else + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + unformat_free (line_input); + vhost_user_delete_if (vnm, vm, sw_if_index); + return 0; +} + +int +vhost_user_dump_ifs (vnet_main_t * vnm, vlib_main_t * vm, + vhost_user_intf_details_t ** out_vuids) +{ + int rv = 0; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + vhost_user_intf_details_t *r_vuids = NULL; + vhost_user_intf_details_t *vuid = NULL; + u32 *hw_if_indices = 0; + vnet_hw_interface_t *hi; + u8 *s = NULL; + int i; + + if (!out_vuids) + return -1; + + pool_foreach (vui, vum->vhost_user_interfaces, + vec_add1 (hw_if_indices, vui->hw_if_index); + ); + + for (i = 0; i < vec_len (hw_if_indices); i++) + { + hi = vnet_get_hw_interface (vnm, hw_if_indices[i]); + vui = pool_elt_at_index (vum->vhost_user_interfaces, hi->dev_instance); + + vec_add2 (r_vuids, vuid, 1); + vuid->sw_if_index = vui->sw_if_index; + vuid->virtio_net_hdr_sz = vui->virtio_net_hdr_sz; + vuid->features = vui->features; + vuid->num_regions = vui->nregions; + vuid->sock_errno = vui->sock_errno; + strncpy ((char *) vuid->sock_filename, (char *) vui->sock_filename, + ARRAY_LEN (vuid->sock_filename) - 1); + + s = format (s, "%v%c", hi->name, 0); + + strncpy ((char *) vuid->if_name, (char *) s, + ARRAY_LEN (vuid->if_name) - 1); + _vec_len (s) = 0; + } + + vec_free (s); + vec_free (hw_if_indices); + + *out_vuids = r_vuids; + + return rv; +} + +clib_error_t * +show_vhost_user_command_fn (vlib_main_t * vm, + unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + clib_error_t *error = 0; + vnet_main_t *vnm = vnet_get_main (); + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + u32 hw_if_index, *hw_if_indices = 0; + vnet_hw_interface_t *hi; + vhost_cpu_t *vhc; + vhost_iface_and_queue_t *vhiq; + u32 ci; + + int i, j, q; + int show_descr = 0; + struct feat_struct + { + u8 bit; + char *str; + }; + struct feat_struct *feat_entry; + + static struct feat_struct feat_array[] = { +#define _(s,b) { .str = #s, .bit = b, }, + foreach_virtio_net_feature +#undef _ + {.str = NULL} + }; + +#define foreach_protocol_feature \ + _(VHOST_USER_PROTOCOL_F_MQ) \ + _(VHOST_USER_PROTOCOL_F_LOG_SHMFD) + + static struct feat_struct proto_feat_array[] = { +#define _(s) { .str = #s, .bit = s}, + foreach_protocol_feature +#undef _ + {.str = NULL} + }; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (input, "%U", unformat_vnet_hw_interface, vnm, &hw_if_index)) + { + vec_add1 (hw_if_indices, hw_if_index); + } + else if (unformat (input, "descriptors") || unformat (input, "desc")) + show_descr = 1; + else + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + goto done; + } + } + if (vec_len (hw_if_indices) == 0) + { + pool_foreach (vui, vum->vhost_user_interfaces, + vec_add1 (hw_if_indices, vui->hw_if_index); + ); + } + vlib_cli_output (vm, "Virtio vhost-user interfaces"); + vlib_cli_output (vm, "Global:\n coalesce frames %d time %e", + vum->coalesce_frames, vum->coalesce_time); + + for (i = 0; i < vec_len (hw_if_indices); i++) + { + hi = vnet_get_hw_interface (vnm, hw_if_indices[i]); + vui = pool_elt_at_index (vum->vhost_user_interfaces, hi->dev_instance); + vlib_cli_output (vm, "Interface: %s (ifindex %d)", + hi->name, hw_if_indices[i]); + + vlib_cli_output (vm, "virtio_net_hdr_sz %d\n" + " features mask (0x%llx): \n" + " features (0x%llx): \n", + vui->virtio_net_hdr_sz, vui->feature_mask, + vui->features); + + feat_entry = (struct feat_struct *) &feat_array; + while (feat_entry->str) + { + if (vui->features & (1ULL << feat_entry->bit)) + vlib_cli_output (vm, " %s (%d)", feat_entry->str, + feat_entry->bit); + feat_entry++; + } + + vlib_cli_output (vm, " protocol features (0x%llx)", + vui->protocol_features); + feat_entry = (struct feat_struct *) &proto_feat_array; + while (feat_entry->str) + { + if (vui->protocol_features & (1ULL << feat_entry->bit)) + vlib_cli_output (vm, " %s (%d)", feat_entry->str, + feat_entry->bit); + feat_entry++; + } + + vlib_cli_output (vm, "\n"); + + vlib_cli_output (vm, " socket filename %s type %s errno \"%s\"\n\n", + vui->sock_filename, + (vui->unix_server_index != ~0) ? "server" : "client", + strerror (vui->sock_errno)); + + vlib_cli_output (vm, " rx placement: "); + vec_foreach (vhc, vum->cpus) + { + vec_foreach (vhiq, vhc->rx_queues) + { + if (vhiq->vhost_iface_index == vui - vum->vhost_user_interfaces) + vlib_cli_output (vm, " thread %d on vring %d\n", + vhc - vum->cpus, VHOST_VRING_IDX_TX (vhiq->qid)); + } + } + + vlib_cli_output (vm, " tx placement: %s\n", + vui->use_tx_spinlock ? "spin-lock" : "lock-free"); + + vec_foreach_index (ci, vui->per_cpu_tx_qid) + { + vlib_cli_output (vm, " thread %d on vring %d\n", ci, + VHOST_VRING_IDX_RX (vui->per_cpu_tx_qid[ci])); + } + + vlib_cli_output (vm, "\n"); + + vlib_cli_output (vm, " Memory regions (total %d)\n", vui->nregions); + + if (vui->nregions) + { + vlib_cli_output (vm, + " region fd guest_phys_addr memory_size userspace_addr mmap_offset mmap_addr\n"); + vlib_cli_output (vm, + " ====== ===== ================== ================== ================== ================== ==================\n"); + } + for (j = 0; j < vui->nregions; j++) + { + vlib_cli_output (vm, + " %d %-5d 0x%016lx 0x%016lx 0x%016lx 0x%016lx 0x%016lx\n", + j, vui->region_mmap_fd[j], + vui->regions[j].guest_phys_addr, + vui->regions[j].memory_size, + vui->regions[j].userspace_addr, + vui->regions[j].mmap_offset, + pointer_to_uword (vui->region_mmap_addr[j])); + } + for (q = 0; q < VHOST_VRING_MAX_N; q++) + { + if (!vui->vrings[q].started) + continue; + + vlib_cli_output (vm, "\n Virtqueue %d (%s%s)\n", q, + (q & 1) ? "RX" : "TX", + vui->vrings[q].enabled ? "" : " disabled"); + + vlib_cli_output (vm, + " qsz %d last_avail_idx %d last_used_idx %d\n", + vui->vrings[q].qsz, vui->vrings[q].last_avail_idx, + vui->vrings[q].last_used_idx); + + if (vui->vrings[q].avail && vui->vrings[q].used) + vlib_cli_output (vm, + " avail.flags %x avail.idx %d used.flags %x used.idx %d\n", + vui->vrings[q].avail->flags, + vui->vrings[q].avail->idx, + vui->vrings[q].used->flags, + vui->vrings[q].used->idx); + + int kickfd = UNIX_GET_FD (vui->vrings[q].kickfd_idx); + int callfd = UNIX_GET_FD (vui->vrings[q].callfd_idx); + vlib_cli_output (vm, " kickfd %d callfd %d errfd %d\n", + kickfd, callfd, vui->vrings[q].errfd); + + if (show_descr) + { + vlib_cli_output (vm, "\n descriptor table:\n"); + vlib_cli_output (vm, + " id addr len flags next user_addr\n"); + vlib_cli_output (vm, + " ===== ================== ===== ====== ===== ==================\n"); + for (j = 0; j < vui->vrings[q].qsz; j++) + { + u32 mem_hint = 0; + vlib_cli_output (vm, + " %-5d 0x%016lx %-5d 0x%04x %-5d 0x%016lx\n", + j, vui->vrings[q].desc[j].addr, + vui->vrings[q].desc[j].len, + vui->vrings[q].desc[j].flags, + vui->vrings[q].desc[j].next, + pointer_to_uword (map_guest_mem + (vui, + vui->vrings[q].desc[j]. + addr, &mem_hint))); + } + } + } + vlib_cli_output (vm, "\n"); + } +done: + vec_free (hw_if_indices); + return error; +} + +/* + * CLI functions + */ + +/*? + * Create a vHost User interface. Once created, a new virtual interface + * will exist with the name 'VirtualEthernet0/0/x', where 'x' + * is the next free index. + * + * There are several parameters associated with a vHost interface: + * + * - socket - Name of the linux socket used by QEMU/VM and + * VPP to manage the vHost interface. If socket does not already exist, VPP will + * create the socket. + * + * - server - Optional flag to indicate that VPP should be the server for the + * linux socket. If not provided, VPP will be the client. + * + * - feature-mask - Optional virtio/vhost feature set negotiated at + * startup. By default, all supported features will be advertised. Otherwise, + * provide the set of features desired. + * - 0x000008000 (15) - VIRTIO_NET_F_MRG_RXBUF + * - 0x000020000 (17) - VIRTIO_NET_F_CTRL_VQ + * - 0x000200000 (21) - VIRTIO_NET_F_GUEST_ANNOUNCE + * - 0x000400000 (22) - VIRTIO_NET_F_MQ + * - 0x004000000 (26) - VHOST_F_LOG_ALL + * - 0x008000000 (27) - VIRTIO_F_ANY_LAYOUT + * - 0x010000000 (28) - VIRTIO_F_INDIRECT_DESC + * - 0x040000000 (30) - VHOST_USER_F_PROTOCOL_FEATURES + * - 0x100000000 (32) - VIRTIO_F_VERSION_1 + * + * - hwaddr - Optional ethernet address, can be in either + * X:X:X:X:X:X unix or X.X.X cisco format. + * + * - renumber - Optional parameter which allows the instance + * in the name to be specified. If instance already exists, name will be used + * anyway and multiple instances will have the same name. Use with caution. + * + * @cliexpar + * Example of how to create a vhost interface with VPP as the client and all features enabled: + * @cliexstart{create vhost-user socket /tmp/vhost1.sock} + * VirtualEthernet0/0/0 + * @cliexend + * Example of how to create a vhost interface with VPP as the server and with just + * multiple queues enabled: + * @cliexstart{create vhost-user socket /tmp/vhost2.sock server feature-mask 0x40400000} + * VirtualEthernet0/0/1 + * @cliexend + * Once the vHost interface is created, enable the interface using: + * @cliexcmd{set interface state VirtualEthernet0/0/0 up} +?*/ +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (vhost_user_connect_command, static) = { + .path = "create vhost-user", + .short_help = "create vhost-user socket [server] [feature-mask ] [hwaddr ] [renumber ]", + .function = vhost_user_connect_command_fn, +}; +/* *INDENT-ON* */ + +/*? + * Delete a vHost User interface using the interface name or the + * software interface index. Use the 'show interfaces' + * command to determine the software interface index. On deletion, + * the linux socket will not be deleted. + * + * @cliexpar + * Example of how to delete a vhost interface by name: + * @cliexcmd{delete vhost-user VirtualEthernet0/0/1} + * Example of how to delete a vhost interface by software interface index: + * @cliexcmd{delete vhost-user sw_if_index 1} +?*/ +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (vhost_user_delete_command, static) = { + .path = "delete vhost-user", + .short_help = "delete vhost-user { | sw_if_index }", + .function = vhost_user_delete_command_fn, +}; + +/*? + * Display the attributes of a single vHost User interface (provide interface + * name), multiple vHost User interfaces (provide a list of interface names seperated + * by spaces) or all Vhost User interfaces (omit an interface name to display all + * vHost interfaces). + * + * @cliexpar + * @parblock + * Example of how to display a vhost interface: + * @cliexstart{show vhost-user VirtualEthernet0/0/0} + * Virtio vhost-user interfaces + * Global: + * coalesce frames 32 time 1e-3 + * Interface: VirtualEthernet0/0/0 (ifindex 1) + * virtio_net_hdr_sz 12 + * features mask (0xffffffffffffffff): + * features (0x50408000): + * VIRTIO_NET_F_MRG_RXBUF (15) + * VIRTIO_NET_F_MQ (22) + * VIRTIO_F_INDIRECT_DESC (28) + * VHOST_USER_F_PROTOCOL_FEATURES (30) + * protocol features (0x3) + * VHOST_USER_PROTOCOL_F_MQ (0) + * VHOST_USER_PROTOCOL_F_LOG_SHMFD (1) + * + * socket filename /tmp/vhost1.sock type client errno "Success" + * + * rx placement: + * thread 1 on vring 1 + * thread 1 on vring 5 + * thread 2 on vring 3 + * thread 2 on vring 7 + * tx placement: spin-lock + * thread 0 on vring 0 + * thread 1 on vring 2 + * thread 2 on vring 0 + * + * Memory regions (total 2) + * region fd guest_phys_addr memory_size userspace_addr mmap_offset mmap_addr + * ====== ===== ================== ================== ================== ================== ================== + * 0 60 0x0000000000000000 0x00000000000a0000 0x00002aaaaac00000 0x0000000000000000 0x00002aab2b400000 + * 1 61 0x00000000000c0000 0x000000003ff40000 0x00002aaaaacc0000 0x00000000000c0000 0x00002aababcc0000 + * + * Virtqueue 0 (TX) + * qsz 256 last_avail_idx 0 last_used_idx 0 + * avail.flags 1 avail.idx 128 used.flags 1 used.idx 0 + * kickfd 62 callfd 64 errfd -1 + * + * Virtqueue 1 (RX) + * qsz 256 last_avail_idx 0 last_used_idx 0 + * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0 + * kickfd 65 callfd 66 errfd -1 + * + * Virtqueue 2 (TX) + * qsz 256 last_avail_idx 0 last_used_idx 0 + * avail.flags 1 avail.idx 128 used.flags 1 used.idx 0 + * kickfd 63 callfd 70 errfd -1 + * + * Virtqueue 3 (RX) + * qsz 256 last_avail_idx 0 last_used_idx 0 + * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0 + * kickfd 72 callfd 74 errfd -1 + * + * Virtqueue 4 (TX disabled) + * qsz 256 last_avail_idx 0 last_used_idx 0 + * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0 + * kickfd 76 callfd 78 errfd -1 + * + * Virtqueue 5 (RX disabled) + * qsz 256 last_avail_idx 0 last_used_idx 0 + * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0 + * kickfd 80 callfd 82 errfd -1 + * + * Virtqueue 6 (TX disabled) + * qsz 256 last_avail_idx 0 last_used_idx 0 + * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0 + * kickfd 84 callfd 86 errfd -1 + * + * Virtqueue 7 (RX disabled) + * qsz 256 last_avail_idx 0 last_used_idx 0 + * avail.flags 1 avail.idx 0 used.flags 1 used.idx 0 + * kickfd 88 callfd 90 errfd -1 + * + * @cliexend + * + * The optional 'descriptors' parameter will display the same output as + * the previous example but will include the descriptor table for each queue. + * The output is truncated below: + * @cliexstart{show vhost-user VirtualEthernet0/0/0 descriptors} + * Virtio vhost-user interfaces + * Global: + * coalesce frames 32 time 1e-3 + * Interface: VirtualEthernet0/0/0 (ifindex 1) + * virtio_net_hdr_sz 12 + * features mask (0xffffffffffffffff): + * features (0x50408000): + * VIRTIO_NET_F_MRG_RXBUF (15) + * VIRTIO_NET_F_MQ (22) + * : + * Virtqueue 0 (TX) + * qsz 256 last_avail_idx 0 last_used_idx 0 + * avail.flags 1 avail.idx 128 used.flags 1 used.idx 0 + * kickfd 62 callfd 64 errfd -1 + * + * descriptor table: + * id addr len flags next user_addr + * ===== ================== ===== ====== ===== ================== + * 0 0x0000000010b6e974 2060 0x0002 1 0x00002aabbc76e974 + * 1 0x0000000010b6e034 2060 0x0002 2 0x00002aabbc76e034 + * 2 0x0000000010b6d6f4 2060 0x0002 3 0x00002aabbc76d6f4 + * 3 0x0000000010b6cdb4 2060 0x0002 4 0x00002aabbc76cdb4 + * 4 0x0000000010b6c474 2060 0x0002 5 0x00002aabbc76c474 + * 5 0x0000000010b6bb34 2060 0x0002 6 0x00002aabbc76bb34 + * 6 0x0000000010b6b1f4 2060 0x0002 7 0x00002aabbc76b1f4 + * 7 0x0000000010b6a8b4 2060 0x0002 8 0x00002aabbc76a8b4 + * 8 0x0000000010b69f74 2060 0x0002 9 0x00002aabbc769f74 + * 9 0x0000000010b69634 2060 0x0002 10 0x00002aabbc769634 + * 10 0x0000000010b68cf4 2060 0x0002 11 0x00002aabbc768cf4 + * : + * 249 0x0000000000000000 0 0x0000 250 0x00002aab2b400000 + * 250 0x0000000000000000 0 0x0000 251 0x00002aab2b400000 + * 251 0x0000000000000000 0 0x0000 252 0x00002aab2b400000 + * 252 0x0000000000000000 0 0x0000 253 0x00002aab2b400000 + * 253 0x0000000000000000 0 0x0000 254 0x00002aab2b400000 + * 254 0x0000000000000000 0 0x0000 255 0x00002aab2b400000 + * 255 0x0000000000000000 0 0x0000 32768 0x00002aab2b400000 + * + * Virtqueue 1 (RX) + * qsz 256 last_avail_idx 0 last_used_idx 0 + * : + * @cliexend + * @endparblock +?*/ +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (show_vhost_user_command, static) = { + .path = "show vhost-user", + .short_help = "show vhost-user [ [ [..]]] [descriptors]", + .function = show_vhost_user_command_fn, +}; +/* *INDENT-ON* */ + +static clib_error_t * +vhost_user_config (vlib_main_t * vm, unformat_input_t * input) +{ + vhost_user_main_t *vum = &vhost_user_main; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "coalesce-frames %d", &vum->coalesce_frames)) + ; + else if (unformat (input, "coalesce-time %f", &vum->coalesce_time)) + ; + else if (unformat (input, "dont-dump-memory")) + vum->dont_dump_vhost_user_memory = 1; + else + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + + return 0; +} + +/* vhost-user { ... } configuration. */ +VLIB_CONFIG_FUNCTION (vhost_user_config, "vhost-user"); + +void +vhost_user_unmap_all (void) +{ + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + + if (vum->dont_dump_vhost_user_memory) + { + pool_foreach (vui, vum->vhost_user_interfaces, + unmap_all_mem_regions (vui); + ); + } +} + +static clib_error_t * +vhost_thread_command_fn (vlib_main_t * vm, + unformat_input_t * input, vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + u32 worker_thread_index; + u32 sw_if_index; + u8 del = 0; + int rv; + + /* Get a line of input. */ + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + if (!unformat + (line_input, "%U %d", unformat_vnet_sw_interface, vnet_get_main (), + &sw_if_index, &worker_thread_index)) + { + unformat_free (line_input); + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + + if (unformat (line_input, "del")) + del = 1; + + if ((rv = + vhost_user_thread_placement (sw_if_index, worker_thread_index, del))) + return clib_error_return (0, "vhost_user_thread_placement returned %d", + rv); + return 0; +} + + +/*? + * This command is used to move the RX processing for the given + * interfaces to the provided thread. If the 'del' option is used, + * the forced thread assignment is removed and the thread assigment is + * reassigned automatically. Use 'show vhost-user ' + * to see the thread assignment. + * + * @cliexpar + * Example of how to move the RX processing for a given interface to a given thread: + * @cliexcmd{vhost thread VirtualEthernet0/0/0 1} + * Example of how to remove the forced thread assignment for a given interface: + * @cliexcmd{vhost thread VirtualEthernet0/0/0 1 del} +?*/ +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (vhost_user_thread_command, static) = { + .path = "vhost thread", + .short_help = "vhost thread [del]", + .function = vhost_thread_command_fn, +}; +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/virtio/vhost-user.h b/src/vnet/devices/virtio/vhost-user.h new file mode 100644 index 00000000000..3083b614016 --- /dev/null +++ b/src/vnet/devices/virtio/vhost-user.h @@ -0,0 +1,350 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#ifndef __VIRTIO_VHOST_USER_H__ +#define __VIRTIO_VHOST_USER_H__ +/* vhost-user data structures */ + +#define VHOST_MEMORY_MAX_NREGIONS 8 +#define VHOST_USER_MSG_HDR_SZ 12 +#define VHOST_VRING_MAX_SIZE 32768 +#define VHOST_VRING_MAX_N 16 //8TX + 8RX +#define VHOST_VRING_IDX_RX(qid) (2*qid) +#define VHOST_VRING_IDX_TX(qid) (2*qid + 1) + +#define VIRTQ_DESC_F_NEXT 1 +#define VIRTQ_DESC_F_INDIRECT 4 +#define VHOST_USER_REPLY_MASK (0x1 << 2) + +#define VHOST_USER_PROTOCOL_F_MQ 0 +#define VHOST_USER_PROTOCOL_F_LOG_SHMFD 1 +#define VHOST_VRING_F_LOG 0 + +#define VHOST_USER_F_PROTOCOL_FEATURES 30 +#define VHOST_USER_PROTOCOL_FEATURES ((1ULL << VHOST_USER_PROTOCOL_F_MQ) | \ + (1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD)) + +/* If multiqueue is provided by host, then we suppport it. */ +#define VIRTIO_NET_CTRL_MQ 4 +#define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET 0 +#define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN 1 +#define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX 0x8000 + +#define VRING_USED_F_NO_NOTIFY 1 + +#define foreach_virtio_net_feature \ + _ (VIRTIO_NET_F_MRG_RXBUF, 15) \ + _ (VIRTIO_NET_F_CTRL_VQ, 17) \ + _ (VIRTIO_NET_F_GUEST_ANNOUNCE, 21) \ + _ (VIRTIO_NET_F_MQ, 22) \ + _ (VHOST_F_LOG_ALL, 26) \ + _ (VIRTIO_F_ANY_LAYOUT, 27) \ + _ (VIRTIO_F_INDIRECT_DESC, 28) \ + _ (VHOST_USER_F_PROTOCOL_FEATURES, 30) \ + _ (VIRTIO_F_VERSION_1, 32) + + +typedef enum +{ +#define _(f,n) FEAT_##f = (n), + foreach_virtio_net_feature +#undef _ +} virtio_net_feature_t; + +int vhost_user_create_if (vnet_main_t * vnm, vlib_main_t * vm, + const char *sock_filename, u8 is_server, + u32 * sw_if_index, u64 feature_mask, + u8 renumber, u32 custom_dev_instance, u8 * hwaddr); +int vhost_user_modify_if (vnet_main_t * vnm, vlib_main_t * vm, + const char *sock_filename, u8 is_server, + u32 sw_if_index, u64 feature_mask, + u8 renumber, u32 custom_dev_instance); +int vhost_user_delete_if (vnet_main_t * vnm, vlib_main_t * vm, + u32 sw_if_index); + +/* *INDENT-OFF* */ +typedef struct vhost_user_memory_region +{ + u64 guest_phys_addr; + u64 memory_size; + u64 userspace_addr; + u64 mmap_offset; +} __attribute ((packed)) vhost_user_memory_region_t; + +typedef struct vhost_user_memory +{ + u32 nregions; + u32 padding; + vhost_user_memory_region_t regions[VHOST_MEMORY_MAX_NREGIONS]; +} __attribute ((packed)) vhost_user_memory_t; + +typedef struct +{ + u32 index, num; +} __attribute ((packed)) vhost_vring_state_t; + +typedef struct +{ + u32 index, flags; + u64 desc_user_addr, used_user_addr, avail_user_addr, log_guest_addr; +} __attribute ((packed)) vhost_vring_addr_t; + +typedef struct vhost_user_log +{ + u64 size; + u64 offset; +} __attribute ((packed)) vhost_user_log_t; + +typedef enum vhost_user_req +{ + VHOST_USER_NONE = 0, + VHOST_USER_GET_FEATURES = 1, + VHOST_USER_SET_FEATURES = 2, + VHOST_USER_SET_OWNER = 3, + VHOST_USER_RESET_OWNER = 4, + VHOST_USER_SET_MEM_TABLE = 5, + VHOST_USER_SET_LOG_BASE = 6, + VHOST_USER_SET_LOG_FD = 7, + VHOST_USER_SET_VRING_NUM = 8, + VHOST_USER_SET_VRING_ADDR = 9, + VHOST_USER_SET_VRING_BASE = 10, + VHOST_USER_GET_VRING_BASE = 11, + VHOST_USER_SET_VRING_KICK = 12, + VHOST_USER_SET_VRING_CALL = 13, + VHOST_USER_SET_VRING_ERR = 14, + VHOST_USER_GET_PROTOCOL_FEATURES = 15, + VHOST_USER_SET_PROTOCOL_FEATURES = 16, + VHOST_USER_GET_QUEUE_NUM = 17, + VHOST_USER_SET_VRING_ENABLE = 18, + VHOST_USER_MAX +} vhost_user_req_t; + +// vring_desc I/O buffer descriptor +typedef struct +{ + uint64_t addr; // packet data buffer address + uint32_t len; // packet data buffer size + uint16_t flags; // (see below) + uint16_t next; // optional index next descriptor in chain +} __attribute ((packed)) vring_desc_t; + +typedef struct +{ + uint16_t flags; + volatile uint16_t idx; + uint16_t ring[VHOST_VRING_MAX_SIZE]; +} __attribute ((packed)) vring_avail_t; + +typedef struct +{ + uint16_t flags; + uint16_t idx; + struct /* vring_used_elem */ + { + uint32_t id; + uint32_t len; + } ring[VHOST_VRING_MAX_SIZE]; +} __attribute ((packed)) vring_used_t; + +typedef struct +{ + u8 flags; + u8 gso_type; + u16 hdr_len; + u16 gso_size; + u16 csum_start; + u16 csum_offset; +} __attribute ((packed)) virtio_net_hdr_t; + +typedef struct { + virtio_net_hdr_t hdr; + u16 num_buffers; +} __attribute ((packed)) virtio_net_hdr_mrg_rxbuf_t; + +typedef struct vhost_user_msg { + vhost_user_req_t request; + u32 flags; + u32 size; + union + { + u64 u64; + vhost_vring_state_t state; + vhost_vring_addr_t addr; + vhost_user_memory_t memory; + vhost_user_log_t log; + }; +} __attribute ((packed)) vhost_user_msg_t; +/* *INDENT-ON* */ + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + u16 qsz; + u16 last_avail_idx; + u16 last_used_idx; + u16 n_since_last_int; + vring_desc_t *desc; + vring_avail_t *avail; + vring_used_t *used; + f64 int_deadline; + u8 started; + u8 enabled; + u8 log_used; + //Put non-runtime in a different cache line + CLIB_CACHE_LINE_ALIGN_MARK (cacheline1); + int errfd; + u32 callfd_idx; + u32 kickfd_idx; + u64 log_guest_addr; +} vhost_user_vring_t; + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + u32 is_up; + u32 admin_up; + u32 unix_server_index; + u32 unix_file_index; + char sock_filename[256]; + int sock_errno; + u32 hw_if_index, sw_if_index; + + //Feature negotiation + u64 features; + u64 feature_mask; + u64 protocol_features; + + //Memory region information + u32 nregions; + vhost_user_memory_region_t regions[VHOST_MEMORY_MAX_NREGIONS]; + void *region_mmap_addr[VHOST_MEMORY_MAX_NREGIONS]; + u64 region_guest_addr_lo[VHOST_MEMORY_MAX_NREGIONS]; + u64 region_guest_addr_hi[VHOST_MEMORY_MAX_NREGIONS]; + u32 region_mmap_fd[VHOST_MEMORY_MAX_NREGIONS]; + + //Virtual rings + vhost_user_vring_t vrings[VHOST_VRING_MAX_N]; + volatile u32 *vring_locks[VHOST_VRING_MAX_N]; + + int virtio_net_hdr_sz; + int is_any_layout; + + void *log_base_addr; + u64 log_size; + + /* Whether to use spinlock or per_cpu_tx_qid assignment */ + u8 use_tx_spinlock; + u16 *per_cpu_tx_qid; + + /* Vector of workers for this interface */ + u32 *workers; +} vhost_user_intf_t; + +typedef struct +{ + u16 vhost_iface_index; + u16 qid; +} vhost_iface_and_queue_t; + +typedef struct +{ + uword dst; + uword src; + u32 len; +} vhost_copy_t; + +typedef struct +{ + u16 qid; /** The interface queue index (Not the virtio vring idx) */ + u16 device_index; /** The device index */ + u32 virtio_ring_flags; /** Runtime queue flags **/ + u16 first_desc_len; /** Length of the first data descriptor **/ + virtio_net_hdr_mrg_rxbuf_t hdr; /** Virtio header **/ +} vhost_trace_t; + + +#define VHOST_USER_RX_BUFFERS_N (2 * VLIB_FRAME_SIZE + 2) +#define VHOST_USER_COPY_ARRAY_N (4 * VLIB_FRAME_SIZE) + +typedef struct +{ + vhost_iface_and_queue_t *rx_queues; + u32 rx_buffers_len; + u32 rx_buffers[VHOST_USER_RX_BUFFERS_N]; + + virtio_net_hdr_mrg_rxbuf_t tx_headers[VLIB_FRAME_SIZE]; + vhost_copy_t copy[VHOST_USER_COPY_ARRAY_N]; + + /* This is here so it doesn't end-up + * using stack or registers. */ + vhost_trace_t *current_trace; +} vhost_cpu_t; + +typedef struct +{ + u32 mtu_bytes; + vhost_user_intf_t *vhost_user_interfaces; + u32 *show_dev_instance_by_real_dev_instance; + u32 coalesce_frames; + f64 coalesce_time; + int dont_dump_vhost_user_memory; + + /** first cpu index */ + u32 input_cpu_first_index; + + /** total cpu count */ + u32 input_cpu_count; + + /** Per-CPU data for vhost-user */ + vhost_cpu_t *cpus; + + /** Pseudo random iterator */ + u32 random; +} vhost_user_main_t; + +typedef struct +{ + u8 if_name[64]; + u32 sw_if_index; + u32 virtio_net_hdr_sz; + u64 features; + u8 is_server; + u8 sock_filename[256]; + u32 num_regions; + int sock_errno; +} vhost_user_intf_details_t; + +int vhost_user_dump_ifs (vnet_main_t * vnm, vlib_main_t * vm, + vhost_user_intf_details_t ** out_vuids); + +// CLI commands to be used from dpdk +clib_error_t *vhost_user_connect_command_fn (vlib_main_t * vm, + unformat_input_t * input, + vlib_cli_command_t * cmd); +clib_error_t *vhost_user_delete_command_fn (vlib_main_t * vm, + unformat_input_t * input, + vlib_cli_command_t * cmd); +clib_error_t *show_vhost_user_command_fn (vlib_main_t * vm, + unformat_input_t * input, + vlib_cli_command_t * cmd); + +#endif + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/virtio/vhost_user.api b/src/vnet/devices/virtio/vhost_user.api new file mode 100644 index 00000000000..21e42298361 --- /dev/null +++ b/src/vnet/devices/virtio/vhost_user.api @@ -0,0 +1,125 @@ +/* + * Copyright (c) 2015-2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/** \brief vhost-user interface create request + @param client_index - opaque cookie to identify the sender + @param is_server - our side is socket server + @param sock_filename - unix socket filename, used to speak with frontend + @param use_custom_mac - enable or disable the use of the provided hardware address + @param mac_address - hardware address to use if 'use_custom_mac' is set +*/ +define create_vhost_user_if +{ + u32 client_index; + u32 context; + u8 is_server; + u8 sock_filename[256]; + u8 renumber; + u32 custom_dev_instance; + u8 use_custom_mac; + u8 mac_address[6]; + u8 tag[64]; +}; + +/** \brief vhost-user interface create response + @param context - sender context, to match reply w/ request + @param retval - return code for the request + @param sw_if_index - interface the operation is applied to +*/ +define create_vhost_user_if_reply +{ + u32 context; + i32 retval; + u32 sw_if_index; +}; + +/** \brief vhost-user interface modify request + @param client_index - opaque cookie to identify the sender + @param is_server - our side is socket server + @param sock_filename - unix socket filename, used to speak with frontend +*/ +define modify_vhost_user_if +{ + u32 client_index; + u32 context; + u32 sw_if_index; + u8 is_server; + u8 sock_filename[256]; + u8 renumber; + u32 custom_dev_instance; +}; + +/** \brief vhost-user interface modify response + @param context - sender context, to match reply w/ request + @param retval - return code for the request +*/ +define modify_vhost_user_if_reply +{ + u32 context; + i32 retval; +}; + +/** \brief vhost-user interface delete request + @param client_index - opaque cookie to identify the sender +*/ +define delete_vhost_user_if +{ + u32 client_index; + u32 context; + u32 sw_if_index; +}; + +/** \brief vhost-user interface delete response + @param context - sender context, to match reply w/ request + @param retval - return code for the request +*/ +define delete_vhost_user_if_reply +{ + u32 context; + i32 retval; +}; + +/** \brief Vhost-user interface details structure (fix this) + @param sw_if_index - index of the interface + @param interface_name - name of interface + @param virtio_net_hdr_sz - net header size + @param features - interface features + @param is_server - vhost-user server socket + @param sock_filename - socket filename + @param num_regions - number of used memory regions +*/ +define sw_interface_vhost_user_details +{ + u32 context; + u32 sw_if_index; + u8 interface_name[64]; + u32 virtio_net_hdr_sz; + u64 features; + u8 is_server; + u8 sock_filename[256]; + u32 num_regions; + i32 sock_errno; +}; + +define sw_interface_vhost_user_dump +{ + u32 client_index; + u32 context; +}; +/* + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/virtio/vhost_user_api.c b/src/vnet/devices/virtio/vhost_user_api.c new file mode 100644 index 00000000000..dd517c26f55 --- /dev/null +++ b/src/vnet/devices/virtio/vhost_user_api.c @@ -0,0 +1,262 @@ +/* + *------------------------------------------------------------------ + * vhost-user_api.c - vhost-user api + * + * Copyright (c) 2016 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include +#include + +#include +#include +#include + +#include + +#define vl_typedefs /* define message structures */ +#include +#undef vl_typedefs + +#define vl_endianfun /* define message structures */ +#include +#undef vl_endianfun + +/* instantiate all the print functions we know about */ +#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__) +#define vl_printfun +#include +#undef vl_printfun + +#include + +#define foreach_vpe_api_msg \ +_(CREATE_VHOST_USER_IF, create_vhost_user_if) \ +_(MODIFY_VHOST_USER_IF, modify_vhost_user_if) \ +_(DELETE_VHOST_USER_IF, delete_vhost_user_if) \ +_(SW_INTERFACE_VHOST_USER_DUMP, sw_interface_vhost_user_dump) \ +_(SW_INTERFACE_VHOST_USER_DETAILS, sw_interface_vhost_user_details) + +/* + * WARNING: replicated pending api refactor completion + */ +static void +send_sw_interface_flags_deleted (vpe_api_main_t * am, + unix_shared_memory_queue_t * q, + u32 sw_if_index) +{ + vl_api_sw_interface_set_flags_t *mp; + + mp = vl_msg_api_alloc (sizeof (*mp)); + memset (mp, 0, sizeof (*mp)); + mp->_vl_msg_id = ntohs (VL_API_SW_INTERFACE_SET_FLAGS); + mp->sw_if_index = ntohl (sw_if_index); + + mp->admin_up_down = 0; + mp->link_up_down = 0; + mp->deleted = 1; + vl_msg_api_send_shmem (q, (u8 *) & mp); +} + +static void +vl_api_create_vhost_user_if_t_handler (vl_api_create_vhost_user_if_t * mp) +{ + int rv = 0; + vl_api_create_vhost_user_if_reply_t *rmp; + u32 sw_if_index = (u32) ~ 0; + vnet_main_t *vnm = vnet_get_main (); + vlib_main_t *vm = vlib_get_main (); + + rv = vhost_user_create_if (vnm, vm, (char *) mp->sock_filename, + mp->is_server, &sw_if_index, (u64) ~ 0, + mp->renumber, ntohl (mp->custom_dev_instance), + (mp->use_custom_mac) ? mp->mac_address : NULL); + + /* Remember an interface tag for the new interface */ + if (rv == 0) + { + /* If a tag was supplied... */ + if (mp->tag[0]) + { + /* Make sure it's a proper C-string */ + mp->tag[ARRAY_LEN (mp->tag) - 1] = 0; + u8 *tag = format (0, "%s%c", mp->tag, 0); + vnet_set_sw_interface_tag (vnm, tag, sw_if_index); + } + } + + /* *INDENT-OFF* */ + REPLY_MACRO2(VL_API_CREATE_VHOST_USER_IF_REPLY, + ({ + rmp->sw_if_index = ntohl (sw_if_index); + })); + /* *INDENT-ON* */ +} + +static void +vl_api_modify_vhost_user_if_t_handler (vl_api_modify_vhost_user_if_t * mp) +{ + int rv = 0; + vl_api_modify_vhost_user_if_reply_t *rmp; + u32 sw_if_index = ntohl (mp->sw_if_index); + + vnet_main_t *vnm = vnet_get_main (); + vlib_main_t *vm = vlib_get_main (); + + rv = vhost_user_modify_if (vnm, vm, (char *) mp->sock_filename, + mp->is_server, sw_if_index, (u64) ~ 0, + mp->renumber, ntohl (mp->custom_dev_instance)); + + REPLY_MACRO (VL_API_MODIFY_VHOST_USER_IF_REPLY); +} + +static void +vl_api_delete_vhost_user_if_t_handler (vl_api_delete_vhost_user_if_t * mp) +{ + int rv = 0; + vl_api_delete_vhost_user_if_reply_t *rmp; + vpe_api_main_t *vam = &vpe_api_main; + u32 sw_if_index = ntohl (mp->sw_if_index); + + vnet_main_t *vnm = vnet_get_main (); + vlib_main_t *vm = vlib_get_main (); + + rv = vhost_user_delete_if (vnm, vm, sw_if_index); + + REPLY_MACRO (VL_API_DELETE_VHOST_USER_IF_REPLY); + if (!rv) + { + unix_shared_memory_queue_t *q = + vl_api_client_index_to_input_queue (mp->client_index); + if (!q) + return; + + vnet_clear_sw_interface_tag (vnm, sw_if_index); + send_sw_interface_flags_deleted (vam, q, sw_if_index); + } +} + +static void + vl_api_sw_interface_vhost_user_details_t_handler + (vl_api_sw_interface_vhost_user_details_t * mp) +{ + clib_warning ("BUG"); +} + +static void +send_sw_interface_vhost_user_details (vpe_api_main_t * am, + unix_shared_memory_queue_t * q, + vhost_user_intf_details_t * vui, + u32 context) +{ + vl_api_sw_interface_vhost_user_details_t *mp; + + mp = vl_msg_api_alloc (sizeof (*mp)); + memset (mp, 0, sizeof (*mp)); + mp->_vl_msg_id = ntohs (VL_API_SW_INTERFACE_VHOST_USER_DETAILS); + mp->sw_if_index = ntohl (vui->sw_if_index); + mp->virtio_net_hdr_sz = ntohl (vui->virtio_net_hdr_sz); + mp->features = clib_net_to_host_u64 (vui->features); + mp->is_server = vui->is_server; + mp->num_regions = ntohl (vui->num_regions); + mp->sock_errno = ntohl (vui->sock_errno); + mp->context = context; + + strncpy ((char *) mp->sock_filename, + (char *) vui->sock_filename, ARRAY_LEN (mp->sock_filename) - 1); + strncpy ((char *) mp->interface_name, + (char *) vui->if_name, ARRAY_LEN (mp->interface_name) - 1); + + vl_msg_api_send_shmem (q, (u8 *) & mp); +} + +static void + vl_api_sw_interface_vhost_user_dump_t_handler + (vl_api_sw_interface_vhost_user_dump_t * mp) +{ + int rv = 0; + vpe_api_main_t *am = &vpe_api_main; + vnet_main_t *vnm = vnet_get_main (); + vlib_main_t *vm = vlib_get_main (); + vhost_user_intf_details_t *ifaces = NULL; + vhost_user_intf_details_t *vuid = NULL; + unix_shared_memory_queue_t *q; + + q = vl_api_client_index_to_input_queue (mp->client_index); + if (q == 0) + return; + + rv = vhost_user_dump_ifs (vnm, vm, &ifaces); + if (rv) + return; + + vec_foreach (vuid, ifaces) + { + send_sw_interface_vhost_user_details (am, q, vuid, mp->context); + } + vec_free (ifaces); +} + +/* + * vhost-user_api_hookup + * Add vpe's API message handlers to the table. + * vlib has alread mapped shared memory and + * added the client registration handlers. + * See .../vlib-api/vlibmemory/memclnt_vlib.c:memclnt_process() + */ +#define vl_msg_name_crc_list +#include +#undef vl_msg_name_crc_list + +static void +setup_message_id_table (api_main_t * am) +{ +#define _(id,n,crc) vl_msg_api_add_msg_name_crc (am, #n "_" #crc, id); + foreach_vl_msg_name_crc_vhost_user; +#undef _ +} + +static clib_error_t * +vhost_user_api_hookup (vlib_main_t * vm) +{ + api_main_t *am = &api_main; + +#define _(N,n) \ + vl_msg_api_set_handlers(VL_API_##N, #n, \ + vl_api_##n##_t_handler, \ + vl_noop_handler, \ + vl_api_##n##_t_endian, \ + vl_api_##n##_t_print, \ + sizeof(vl_api_##n##_t), 1); + foreach_vpe_api_msg; +#undef _ + + /* + * Set up the (msg_name, crc, message-id) table + */ + setup_message_id_table (am); + + return 0; +} + +VLIB_API_INIT_FUNCTION (vhost_user_api_hookup); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ -- cgit 1.2.3-korg