From 7eba44d1ec54982636f830a4859027218ca56832 Mon Sep 17 00:00:00 2001 From: Steven Luong Date: Wed, 19 Oct 2022 12:46:29 -0700 Subject: vhost: convert vhost device driver to a plugin convert vhost device driver to a plugin as described in https://jira.fd.io/browse/VPP-2065 Type: improvement Signed-off-by: Steven Luong Change-Id: Ibfe2f351bcaed36a04b136d082ae414145dd37b5 --- src/plugins/vhost/CMakeLists.txt | 32 + src/plugins/vhost/FEATURE.yaml | 13 + src/plugins/vhost/plugin.c | 22 + src/plugins/vhost/vhost_std.h | 69 + src/plugins/vhost/vhost_user.api | 201 +++ src/plugins/vhost/vhost_user.c | 2613 +++++++++++++++++++++++++++++++++ src/plugins/vhost/vhost_user.h | 388 +++++ src/plugins/vhost/vhost_user_api.c | 358 +++++ src/plugins/vhost/vhost_user_inline.h | 496 +++++++ src/plugins/vhost/vhost_user_input.c | 1474 +++++++++++++++++++ src/plugins/vhost/vhost_user_output.c | 1145 +++++++++++++++ src/plugins/vhost/virtio_std.h | 188 +++ 12 files changed, 6999 insertions(+) create mode 100644 src/plugins/vhost/CMakeLists.txt create mode 100644 src/plugins/vhost/FEATURE.yaml create mode 100644 src/plugins/vhost/plugin.c create mode 100644 src/plugins/vhost/vhost_std.h create mode 100644 src/plugins/vhost/vhost_user.api create mode 100644 src/plugins/vhost/vhost_user.c create mode 100644 src/plugins/vhost/vhost_user.h create mode 100644 src/plugins/vhost/vhost_user_api.c create mode 100644 src/plugins/vhost/vhost_user_inline.h create mode 100644 src/plugins/vhost/vhost_user_input.c create mode 100644 src/plugins/vhost/vhost_user_output.c create mode 100644 src/plugins/vhost/virtio_std.h (limited to 'src/plugins/vhost') diff --git a/src/plugins/vhost/CMakeLists.txt b/src/plugins/vhost/CMakeLists.txt new file mode 100644 index 00000000000..f72d9f20346 --- /dev/null +++ b/src/plugins/vhost/CMakeLists.txt @@ -0,0 +1,32 @@ +# Copyright (c) 2020 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. + +add_vpp_plugin(vhost + SOURCES + plugin.c + vhost_user.c + vhost_user_api.c + vhost_user_input.c + vhost_user_output.c + vhost_std.h + vhost_user.h + vhost_user_inline.h + virtio_std.h + + MULTIARCH_SOURCES + vhost_user_input.c + vhost_user_output.c + + API_FILES + vhost_user.api +) diff --git a/src/plugins/vhost/FEATURE.yaml b/src/plugins/vhost/FEATURE.yaml new file mode 100644 index 00000000000..7104dda1dc5 --- /dev/null +++ b/src/plugins/vhost/FEATURE.yaml @@ -0,0 +1,13 @@ +--- +name: Vhost-user Device Driver +maintainer: sluong@cisco.com +features: + - Device mode to emulate vhost-user interface presented to VPP from the + guest VM. + - Support virtio 1.0 in virtio + - Support virtio 1.1 packed ring in virtio [experimental] + - Support multi-queue, GSO, checksum offload, indirect descriptor, + jumbo frame, and packed ring. +description: "Vhost-user implementation" +state: production +properties: [API, CLI, STATS, MULTITHREAD] diff --git a/src/plugins/vhost/plugin.c b/src/plugins/vhost/plugin.c new file mode 100644 index 00000000000..dc5de1c9ed8 --- /dev/null +++ b/src/plugins/vhost/plugin.c @@ -0,0 +1,22 @@ +/* SPDX-License-Identifier: Apache-2.0 + * Copyright (c) 2022 Cisco Systems, Inc. + * License: Cisco Proprietary Closed Source License - Cisco Internal. + * The software, documentation and any fonts accompanying this License whether + * on disk, in read only memory, on any other media or in any other form (col- + * lectively the “Software”) are licensed, not sold, to you by Cisco, Inc. + * (“Cisco”) for use only under the terms of this License, and Cisco reserves + * all rights not expressly granted to you. The rights granted herein are + * limited to Cisco’s intel- lectual property rights in the Cisco Software and + * do not include any other patents or intellectual property rights. You own + * the media on which the Cisco Software is recorded but Cisco and/or Cisco’s + * licensor(s) retain ownership of the Software itself. + */ + +#include +#include +#include + +VLIB_PLUGIN_REGISTER () = { + .version = VPP_BUILD_VER, + .description = "Vhost-User", +}; diff --git a/src/plugins/vhost/vhost_std.h b/src/plugins/vhost/vhost_std.h new file mode 100644 index 00000000000..7799093bac3 --- /dev/null +++ b/src/plugins/vhost/vhost_std.h @@ -0,0 +1,69 @@ +/* + * 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 __VHOST_STD_H__ +#define __VHOST_STD_H__ + +typedef struct +{ + u64 guest_phys_addr; + u64 memory_size; + u64 userspace_addr; + u64 mmap_offset; +} vhost_memory_region_t; + +typedef struct +{ + u32 nregions; + u32 padding; + vhost_memory_region_t regions[0]; +} vhost_memory_t; + +typedef struct +{ + u32 index; + u32 num; +} vhost_vring_state_t; + +typedef struct +{ + u32 index; + int fd; +} vhost_vring_file_t; + +typedef struct +{ + u32 index; + u32 flags; + u64 desc_user_addr; + u64 used_user_addr; + u64 avail_user_addr; + u64 log_guest_addr; +} vhost_vring_addr_t; + +typedef struct +{ + u64 size; + u64 offset; +} vhost_user_log_t; + +#endif + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/plugins/vhost/vhost_user.api b/src/plugins/vhost/vhost_user.api new file mode 100644 index 00000000000..b026ba768a9 --- /dev/null +++ b/src/plugins/vhost/vhost_user.api @@ -0,0 +1,201 @@ +/* + * 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. + */ + +option version = "4.1.1"; + +import "vnet/interface_types.api"; +import "vnet/ethernet/ethernet_types.api"; +import "vnet/devices/virtio/virtio_types.api"; + +/** \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 disable_mrg_rxbuf - disable the use of merge receive buffers + @param disable_indirect_desc - disable the use of indirect descriptors which driver can use + @param enable_gso - enable gso support (default 0) + @param enable_packed - enable packed ring support (default 0) + @param mac_address - hardware address to use if 'use_custom_mac' is set +*/ +define create_vhost_user_if +{ + option deprecated; + u32 client_index; + u32 context; + bool is_server; + string sock_filename[256]; + bool renumber; + bool disable_mrg_rxbuf; + bool disable_indirect_desc; + bool enable_gso; + bool enable_packed; + u32 custom_dev_instance; + bool use_custom_mac; + vl_api_mac_address_t mac_address; + string 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 +{ + option deprecated; + u32 context; + i32 retval; + vl_api_interface_index_t 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 + @param enable_gso - enable gso support (default 0) + @param enable_packed - enable packed ring support (default 0) +*/ +autoreply define modify_vhost_user_if +{ + option deprecated; + u32 client_index; + u32 context; + vl_api_interface_index_t sw_if_index; + bool is_server; + string sock_filename[256]; + bool renumber; + bool enable_gso; + bool enable_packed; + u32 custom_dev_instance; +}; + +/** \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 disable_mrg_rxbuf - disable the use of merge receive buffers + @param disable_indirect_desc - disable the use of indirect descriptors which driver can use + @param enable_gso - enable gso support (default 0) + @param enable_packed - enable packed ring support (default 0) + @param enable_event_idx - enable event_idx support (default 0) + @param mac_address - hardware address to use if 'use_custom_mac' is set + @param renumber - if true, use custom_dev_instance is valid + @param custom_dev_instance - custom device instance number +*/ +define create_vhost_user_if_v2 +{ + u32 client_index; + u32 context; + bool is_server; + string sock_filename[256]; + bool renumber; + bool disable_mrg_rxbuf; + bool disable_indirect_desc; + bool enable_gso; + bool enable_packed; + bool enable_event_idx; + u32 custom_dev_instance; + bool use_custom_mac; + vl_api_mac_address_t mac_address; + string 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_v2_reply +{ + u32 context; + i32 retval; + vl_api_interface_index_t 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 + @param enable_gso - enable gso support (default 0) + @param enable_packed - enable packed ring support (default 0) + @param enable_event_idx - enable event idx support (default 0) + @param renumber - if true, use custom_dev_instance is valid + @param custom_dev_instance - custom device instance number +*/ +autoreply define modify_vhost_user_if_v2 +{ + u32 client_index; + u32 context; + vl_api_interface_index_t sw_if_index; + bool is_server; + string sock_filename[256]; + bool renumber; + bool enable_gso; + bool enable_packed; + bool enable_event_idx; + u32 custom_dev_instance; +}; + +/** \brief vhost-user interface delete request + @param client_index - opaque cookie to identify the sender +*/ +autoreply define delete_vhost_user_if +{ + u32 client_index; + u32 context; + vl_api_interface_index_t sw_if_index; +}; + +/** \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_first_32 - interface features, first 32 bits + @param features_last_32 - interface features, last 32 bits + @param is_server - vhost-user server socket + @param sock_filename - socket filename + @param num_regions - number of used memory regions + @param sock_errno - socket errno +*/ +define sw_interface_vhost_user_details +{ + u32 context; + vl_api_interface_index_t sw_if_index; + string interface_name[64]; + u32 virtio_net_hdr_sz; + vl_api_virtio_net_features_first_32_t features_first_32; + vl_api_virtio_net_features_last_32_t features_last_32; + bool is_server; + string sock_filename[256]; + u32 num_regions; + i32 sock_errno; +}; + +/** \brief Vhost-user interface dump request + @param sw_if_index - filter by sw_if_index +*/ +define sw_interface_vhost_user_dump +{ + u32 client_index; + u32 context; + vl_api_interface_index_t sw_if_index [default=0xffffffff]; +}; +/* + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/plugins/vhost/vhost_user.c b/src/plugins/vhost/vhost_user.c new file mode 100644 index 00000000000..fac20800848 --- /dev/null +++ b/src/plugins/vhost/vhost_user.c @@ -0,0 +1,2613 @@ +/* + *------------------------------------------------------------------ + * vhost.c - vhost-user + * + * Copyright (c) 2014-2018 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 + +#include +#include + +/** + * @file + * @brief vHost User Device Driver. + * + * This file contains the source code for vHost User interface. + */ + + +vlib_node_registration_t vhost_user_send_interrupt_node; + +/* *INDENT-OFF* */ +vhost_user_main_t vhost_user_main = { + .mtu_bytes = 1518, +}; + +VNET_HW_INTERFACE_CLASS (vhost_interface_class, static) = { + .name = "vhost-user", +}; +/* *INDENT-ON* */ + +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, q; + vhost_user_vring_t *vq; + + for (i = 0; i < vui->nregions; i++) + { + if (vui->region_mmap_addr[i] != MAP_FAILED) + { + + 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 - 1) & ~(page_sz - 1); + + r = + munmap (vui->region_mmap_addr[i] - vui->regions[i].mmap_offset, + map_sz); + + vu_log_debug (vui, "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] = MAP_FAILED; + + if (r == -1) + { + vu_log_err (vui, "failed to unmap memory region (errno %d)", + errno); + } + close (vui->region_mmap_fd[i]); + } + } + vui->nregions = 0; + + FOR_ALL_VHOST_RX_TXQ (q, vui) + { + vq = &vui->vrings[q]; + vq->avail = 0; + vq->used = 0; + vq->desc = 0; + } +} + +static_always_inline void +vhost_user_tx_thread_placement (vhost_user_intf_t *vui, u32 qid) +{ + vnet_main_t *vnm = vnet_get_main (); + vhost_user_vring_t *rxvq = &vui->vrings[qid]; + u32 q = qid >> 1, rxvq_count; + + ASSERT ((qid & 1) == 0); + if (!rxvq->started || !rxvq->enabled) + return; + + rxvq_count = (qid >> 1) + 1; + if (rxvq->queue_index == ~0) + { + rxvq->queue_index = + vnet_hw_if_register_tx_queue (vnm, vui->hw_if_index, q); + rxvq->qid = q; + } + + FOR_ALL_VHOST_RXQ (q, vui) + { + vhost_user_vring_t *rxvq = &vui->vrings[q]; + u32 qi = rxvq->queue_index; + + if (rxvq->queue_index == ~0) + break; + for (u32 i = 0; i < vlib_get_n_threads (); i++) + vnet_hw_if_tx_queue_unassign_thread (vnm, qi, i); + } + + for (u32 i = 0; i < vlib_get_n_threads (); i++) + { + vhost_user_vring_t *rxvq = + &vui->vrings[VHOST_VRING_IDX_RX (i % rxvq_count)]; + u32 qi = rxvq->queue_index; + + vnet_hw_if_tx_queue_assign_thread (vnm, qi, i); + } + + vnet_hw_if_update_runtime_data (vnm, vui->hw_if_index); +} + +/** + * @brief Unassign existing interface/queue to thread mappings and re-assign + * new interface/queue to thread mappings + */ +static_always_inline void +vhost_user_rx_thread_placement (vhost_user_intf_t * vui, u32 qid) +{ + vhost_user_vring_t *txvq = &vui->vrings[qid]; + vnet_main_t *vnm = vnet_get_main (); + int rv; + u32 q = qid >> 1; + vhost_user_main_t *vum = &vhost_user_main; + + ASSERT ((qid & 1) == 1); // should be odd + // Assign new queue mappings for the interface + if (txvq->queue_index != ~0) + return; + vnet_hw_if_set_input_node (vnm, vui->hw_if_index, + vhost_user_input_node.index); + txvq->queue_index = vnet_hw_if_register_rx_queue (vnm, vui->hw_if_index, q, + VNET_HW_IF_RXQ_THREAD_ANY); + txvq->thread_index = + vnet_hw_if_get_rx_queue_thread_index (vnm, txvq->queue_index); + + if (txvq->mode == VNET_HW_IF_RX_MODE_UNKNOWN) + /* Set polling as the default */ + txvq->mode = VNET_HW_IF_RX_MODE_POLLING; + if (txvq->mode == VNET_HW_IF_RX_MODE_POLLING) + { + vhost_cpu_t *cpu = vec_elt_at_index (vum->cpus, txvq->thread_index); + /* Keep a polling queue count for each thread */ + cpu->polling_q_count++; + } + txvq->qid = q; + rv = vnet_hw_if_set_rx_queue_mode (vnm, txvq->queue_index, txvq->mode); + if (rv) + vu_log_warn (vui, "unable to set rx mode for interface %d, " + "queue %d: rc=%d", vui->hw_if_index, q, rv); + vnet_hw_if_update_runtime_data (vnm, vui->hw_if_index); +} + +/** @brief Returns whether at least one TX and one RX vring are enabled */ +static_always_inline int +vhost_user_intf_ready (vhost_user_intf_t * vui) +{ + int i, found[2] = { }; //RX + TX + + for (i = 0; i < vui->num_qid; i++) + if (vui->vrings[i].started && vui->vrings[i].enabled) + found[i & 1] = 1; + + return found[0] && found[1]; +} + +static_always_inline void +vhost_user_update_iface_state (vhost_user_intf_t * vui) +{ + /* if we have pointers to descriptor table, go up */ + int is_ready = vhost_user_intf_ready (vui); + if (is_ready != vui->is_ready) + { + vu_log_debug (vui, "interface %d %s", vui->sw_if_index, + is_ready ? "ready" : "down"); + if (vui->admin_up) + vnet_hw_interface_set_flags (vnet_get_main (), vui->hw_if_index, + is_ready ? VNET_HW_INTERFACE_FLAG_LINK_UP + : 0); + vui->is_ready = is_ready; + } +} + +static clib_error_t * +vhost_user_callfd_read_ready (clib_file_t * uf) +{ + __attribute__ ((unused)) int n; + u8 buff[8]; + + n = read (uf->file_descriptor, ((char *) &buff), 8); + + return 0; +} + +static_always_inline void +vhost_user_thread_placement (vhost_user_intf_t * vui, u32 qid) +{ + if (qid & 1) // RX is odd, TX is even + { + if (vui->vrings[qid].queue_index == ~0) + vhost_user_rx_thread_placement (vui, qid); + } + else + vhost_user_tx_thread_placement (vui, qid); +} + +static clib_error_t * +vhost_user_kickfd_read_ready (clib_file_t * uf) +{ + __attribute__ ((unused)) ssize_t n; + u8 buff[8]; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui = + pool_elt_at_index (vum->vhost_user_interfaces, uf->private_data >> 8); + u32 qid = uf->private_data & 0xff; + u32 is_txq = qid & 1; + vhost_user_vring_t *vq = &vui->vrings[qid]; + vnet_main_t *vnm = vnet_get_main (); + + n = read (uf->file_descriptor, buff, 8); + if (vq->started == 0) + { + vq->started = 1; + vhost_user_thread_placement (vui, qid); + vhost_user_update_iface_state (vui); + if (is_txq) + vnet_hw_if_set_rx_queue_file_index (vnm, vq->queue_index, + vq->kickfd_idx); + } + + if (is_txq && (vq->mode != VNET_HW_IF_RX_MODE_POLLING) && + vhost_user_intf_ready (vui)) + { + vhost_cpu_t *cpu = vec_elt_at_index (vum->cpus, vq->thread_index); + /* + * If the thread has more than 1 queue and the other queue is in polling + * mode, there is no need to trigger an interrupt + */ + if (cpu->polling_q_count == 0) + vnet_hw_if_rx_queue_set_int_pending (vnm, vq->queue_index); + } + + return 0; +} + +static_always_inline void +vhost_user_vring_init (vhost_user_intf_t * vui, u32 qid) +{ + vhost_user_vring_t *vring = &vui->vrings[qid]; + + clib_memset (vring, 0, sizeof (*vring)); + vring->kickfd_idx = ~0; + vring->callfd_idx = ~0; + vring->errfd = -1; + vring->qid = -1; + vring->queue_index = ~0; + vring->thread_index = ~0; + vring->mode = VNET_HW_IF_RX_MODE_POLLING; + + clib_spinlock_init (&vring->vring_lock); + + /* + * 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_always_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) + { + clib_file_t *uf = pool_elt_at_index (file_main.file_pool, + vring->kickfd_idx); + clib_file_del (&file_main, uf); + vring->kickfd_idx = ~0; + } + if (vring->callfd_idx != ~0) + { + clib_file_t *uf = pool_elt_at_index (file_main.file_pool, + vring->callfd_idx); + clib_file_del (&file_main, uf); + vring->callfd_idx = ~0; + } + if (vring->errfd != -1) + { + close (vring->errfd); + vring->errfd = -1; + } + + clib_spinlock_free (&vring->vring_lock); + + // save the needed information in vrings prior to being wiped out + u16 q = vui->vrings[qid].qid; + u32 queue_index = vui->vrings[qid].queue_index; + u32 mode = vui->vrings[qid].mode; + u32 thread_index = vui->vrings[qid].thread_index; + vhost_user_vring_init (vui, qid); + vui->vrings[qid].qid = q; + vui->vrings[qid].queue_index = queue_index; + vui->vrings[qid].mode = mode; + vui->vrings[qid].thread_index = thread_index; +} + +static_always_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->clib_file_index != ~0) + { + clib_file_del (&file_main, file_main.file_pool + vui->clib_file_index); + vui->clib_file_index = ~0; + } + + vui->is_ready = 0; + + FOR_ALL_VHOST_RX_TXQ (q, vui) { vhost_user_vring_close (vui, q); } + + unmap_all_mem_regions (vui); + vu_log_debug (vui, "interface ifindex %d disconnected", vui->sw_if_index); +} + +void +vhost_user_set_operation_mode (vhost_user_intf_t *vui, + vhost_user_vring_t *txvq) +{ + if (vhost_user_is_packed_ring_supported (vui)) + { + if (txvq->used_event) + { + if (txvq->mode == VNET_HW_IF_RX_MODE_POLLING) + txvq->used_event->flags = VRING_EVENT_F_DISABLE; + else + txvq->used_event->flags = 0; + } + } + else + { + if (txvq->used) + { + if (txvq->mode == VNET_HW_IF_RX_MODE_POLLING) + txvq->used->flags = VRING_USED_F_NO_NOTIFY; + else + txvq->used->flags = 0; + } + } +} + +static clib_error_t * +vhost_user_socket_read (clib_file_t * uf) +{ + int n, i, j; + 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; + clib_file_t template = { 0 }; + vnet_main_t *vnm = vnet_get_main (); + vlib_main_t *vm = vlib_get_main (); + + vui = pool_elt_at_index (vum->vhost_user_interfaces, uf->private_data); + + char control[CMSG_SPACE (VHOST_MEMORY_MAX_NREGIONS * sizeof (int))]; + + clib_memset (&mh, 0, sizeof (mh)); + clib_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); + + if (n != VHOST_USER_MSG_HDR_SZ) + { + if (n == -1) + { + vu_log_debug (vui, "recvmsg returned error %d %s", errno, + strerror (errno)); + } + else + { + vu_log_debug (vui, "n (%d) != VHOST_USER_MSG_HDR_SZ (%d)", + n, VHOST_USER_MSG_HDR_SZ); + } + goto close_socket; + } + + if (mh.msg_flags & MSG_CTRUNC) + { + vu_log_debug (vui, "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_fast (fds, CMSG_DATA (cmsg), number_of_fds * sizeof (int)); + } + + /* version 1, no reply bit set */ + if ((msg.flags & 7) != 1) + { + vu_log_debug (vui, "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) + { + vu_log_debug (vui, "read failed %s", strerror (errno)); + goto close_socket; + } + else if (rv != msg.size) + { + vu_log_debug (vui, "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 = VIRTIO_FEATURE (VIRTIO_NET_F_MRG_RXBUF) | + VIRTIO_FEATURE (VIRTIO_NET_F_CTRL_VQ) | + VIRTIO_FEATURE (VIRTIO_F_ANY_LAYOUT) | + VIRTIO_FEATURE (VIRTIO_RING_F_INDIRECT_DESC) | + VIRTIO_FEATURE (VHOST_F_LOG_ALL) | + VIRTIO_FEATURE (VIRTIO_NET_F_GUEST_ANNOUNCE) | + VIRTIO_FEATURE (VIRTIO_NET_F_MQ) | + VIRTIO_FEATURE (VHOST_USER_F_PROTOCOL_FEATURES) | + VIRTIO_FEATURE (VIRTIO_F_VERSION_1); + msg.u64 &= vui->feature_mask; + + if (vui->enable_event_idx) + msg.u64 |= VIRTIO_FEATURE (VIRTIO_RING_F_EVENT_IDX); + if (vui->enable_gso) + msg.u64 |= FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS; + if (vui->enable_packed) + msg.u64 |= VIRTIO_FEATURE (VIRTIO_F_RING_PACKED); + + msg.size = sizeof (msg.u64); + vu_log_debug (vui, "if %d msg VHOST_USER_GET_FEATURES - reply " + "0x%016llx", vui->hw_if_index, msg.u64); + n = + send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0); + if (n != (msg.size + VHOST_USER_MSG_HDR_SZ)) + { + vu_log_debug (vui, "could not send message response"); + goto close_socket; + } + break; + + case VHOST_USER_SET_FEATURES: + vu_log_debug (vui, "if %d msg VHOST_USER_SET_FEATURES features " + "0x%016llx", vui->hw_if_index, msg.u64); + + vui->features = msg.u64; + + if (vui->features & + (VIRTIO_FEATURE (VIRTIO_NET_F_MRG_RXBUF) | + VIRTIO_FEATURE (VIRTIO_F_VERSION_1))) + vui->virtio_net_hdr_sz = 12; + else + vui->virtio_net_hdr_sz = 10; + + vui->is_any_layout = + (vui->features & VIRTIO_FEATURE (VIRTIO_F_ANY_LAYOUT)) ? 1 : 0; + + ASSERT (vui->virtio_net_hdr_sz < VLIB_BUFFER_PRE_DATA_SIZE); + if (vui->enable_gso && + ((vui->features & FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS) + == FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS)) + { + vnet_hw_if_set_caps (vnm, vui->hw_if_index, + VNET_HW_IF_CAP_TCP_GSO | + VNET_HW_IF_CAP_TX_TCP_CKSUM | + VNET_HW_IF_CAP_TX_UDP_CKSUM); + } + else + { + vnet_hw_if_unset_caps (vnm, vui->hw_if_index, + VNET_HW_IF_CAP_TCP_GSO | + VNET_HW_IF_CAP_L4_TX_CKSUM); + } + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); + vui->is_ready = 0; + vhost_user_update_iface_state (vui); + break; + + case VHOST_USER_SET_MEM_TABLE: + vu_log_debug (vui, "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)) + { + vu_log_debug (vui, "number of mem regions must be between 1 and %i", + VHOST_MEMORY_MAX_NREGIONS); + goto close_socket; + } + + if (msg.memory.nregions != number_of_fds) + { + vu_log_debug (vui, "each memory region must have FD"); + goto close_socket; + } + + /* Do the mmap without barrier sync */ + void *region_mmap_addr[VHOST_MEMORY_MAX_NREGIONS]; + for (i = 0; i < msg.memory.nregions; i++) + { + long page_sz = get_huge_page_size (fds[i]); + + /* align size to page */ + ssize_t map_sz = (msg.memory.regions[i].memory_size + + msg.memory.regions[i].mmap_offset + + page_sz - 1) & ~(page_sz - 1); + + region_mmap_addr[i] = mmap (0, map_sz, PROT_READ | PROT_WRITE, + MAP_SHARED, fds[i], 0); + if (region_mmap_addr[i] == MAP_FAILED) + { + vu_log_err (vui, "failed to map memory. errno is %d", errno); + for (j = 0; j < i; j++) + munmap (region_mmap_addr[j], map_sz); + goto close_socket; + } + vu_log_debug (vui, "map memory region %d addr 0 len 0x%lx fd %d " + "mapped 0x%lx page_sz 0x%x", i, map_sz, fds[i], + region_mmap_addr[i], page_sz); + } + + vlib_worker_thread_barrier_sync (vm); + unmap_all_mem_regions (vui); + for (i = 0; i < msg.memory.nregions; i++) + { + clib_memcpy_fast (&(vui->regions[i]), &msg.memory.regions[i], + sizeof (vhost_user_memory_region_t)); + + vui->region_mmap_addr[i] = region_mmap_addr[i]; + 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; + + vui->region_mmap_addr[i] += vui->regions[i].mmap_offset; + vui->region_mmap_fd[i] = fds[i]; + + vui->nregions++; + } + + /* + * Re-compute desc, used, and avail descriptor table if vring address + * is set. + */ + FOR_ALL_VHOST_RX_TXQ (q, vui) + { + if (vui->vrings[q].desc_user_addr && vui->vrings[q].used_user_addr && + vui->vrings[q].avail_user_addr) + { + vui->vrings[q].desc = + map_user_mem (vui, vui->vrings[q].desc_user_addr); + vui->vrings[q].used = + map_user_mem (vui, vui->vrings[q].used_user_addr); + vui->vrings[q].avail = + map_user_mem (vui, vui->vrings[q].avail_user_addr); + } + } + vlib_worker_thread_barrier_release (vm); + break; + + case VHOST_USER_SET_VRING_NUM: + vu_log_debug (vui, "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 */ + (msg.state.index >= vui->num_qid)) + { + vu_log_debug (vui, "invalid VHOST_USER_SET_VRING_NUM: msg.state.num" + " %d, msg.state.index %d, curruent max q %d", + msg.state.num, msg.state.index, vui->num_qid); + goto close_socket; + } + vui->vrings[msg.state.index].qsz_mask = msg.state.num - 1; + break; + + case VHOST_USER_SET_VRING_ADDR: + vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_ADDR idx %d", + vui->hw_if_index, msg.state.index); + + if (msg.state.index >= vui->num_qid) + { + vu_log_debug (vui, "invalid vring index VHOST_USER_SET_VRING_ADDR:" + " %u >= %u", msg.state.index, vui->num_qid); + goto close_socket; + } + + if (msg.size < sizeof (msg.addr)) + { + vu_log_debug (vui, "vhost message is too short (%d < %d)", + msg.size, sizeof (msg.addr)); + goto close_socket; + } + + vnet_virtio_vring_desc_t *desc = + map_user_mem (vui, msg.addr.desc_user_addr); + vnet_virtio_vring_used_t *used = + map_user_mem (vui, msg.addr.used_user_addr); + vnet_virtio_vring_avail_t *avail = + map_user_mem (vui, msg.addr.avail_user_addr); + + if ((desc == NULL) || (used == NULL) || (avail == NULL)) + { + vu_log_debug (vui, "failed to map user memory for hw_if_index %d", + vui->hw_if_index); + goto close_socket; + } + + vui->vrings[msg.state.index].desc_user_addr = msg.addr.desc_user_addr; + vui->vrings[msg.state.index].used_user_addr = msg.addr.used_user_addr; + vui->vrings[msg.state.index].avail_user_addr = msg.addr.avail_user_addr; + + vlib_worker_thread_barrier_sync (vm); + vui->vrings[msg.state.index].desc = desc; + vui->vrings[msg.state.index].used = used; + vui->vrings[msg.state.index].avail = avail; + + 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 & VIRTIO_FEATURE (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; + vui->vrings[msg.state.index].last_kick = + vui->vrings[msg.state.index].last_used_idx; + + /* tell driver that we want interrupts or not */ + vhost_user_set_operation_mode (vui, &vui->vrings[msg.state.index]); + vlib_worker_thread_barrier_release (vm); + vhost_user_update_iface_state (vui); + break; + + case VHOST_USER_SET_OWNER: + vu_log_debug (vui, "if %d msg VHOST_USER_SET_OWNER", vui->hw_if_index); + break; + + case VHOST_USER_RESET_OWNER: + vu_log_debug (vui, "if %d msg VHOST_USER_RESET_OWNER", + vui->hw_if_index); + break; + + case VHOST_USER_SET_VRING_CALL: + vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_CALL %d", + vui->hw_if_index, msg.u64); + + q = (u8) (msg.u64 & 0xFF); + if (vui->num_qid > q) + { + /* if there is old fd, delete and close it */ + if (vui->vrings[q].callfd_idx != ~0) + { + clib_file_t *uf = pool_elt_at_index (file_main.file_pool, + vui->vrings[q].callfd_idx); + clib_file_del (&file_main, uf); + vui->vrings[q].callfd_idx = ~0; + } + } + else if (vec_len (vui->vrings) > q) + { + /* grow vrings by pair (RX + TX) */ + vui->num_qid = (q & 1) ? (q + 1) : (q + 2); + } + else + { + u32 i, new_max_q, old_max_q = vec_len (vui->vrings); + + /* + * Double the array size if it is less than 64 entries. + * Slow down thereafter. + */ + if (vec_len (vui->vrings) < (VHOST_VRING_INIT_MQ_PAIR_SZ << 3)) + new_max_q = vec_len (vui->vrings) << 1; + else + new_max_q = vec_len (vui->vrings) + + (VHOST_VRING_INIT_MQ_PAIR_SZ << 2); + if (new_max_q > (VHOST_VRING_MAX_MQ_PAIR_SZ << 1)) + new_max_q = (VHOST_VRING_MAX_MQ_PAIR_SZ << 1); + + /* sync with the worker threads, vrings may move due to realloc */ + vlib_worker_thread_barrier_sync (vm); + vec_validate_aligned (vui->vrings, new_max_q - 1, + CLIB_CACHE_LINE_BYTES); + vlib_worker_thread_barrier_release (vm); + + for (i = old_max_q; i < vec_len (vui->vrings); i++) + vhost_user_vring_init (vui, i); + + /* grow vrings by pair (RX + TX) */ + vui->num_qid = (q & 1) ? (q + 1) : (q + 2); + } + + if (!(msg.u64 & VHOST_USER_VRING_NOFD_MASK)) + { + if (number_of_fds != 1) + { + vu_log_debug (vui, "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; + template.description = format (0, "vhost user"); + vui->vrings[q].callfd_idx = clib_file_add (&file_main, &template); + } + else + vui->vrings[q].callfd_idx = ~0; + break; + + case VHOST_USER_SET_VRING_KICK: + vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_KICK %d", + vui->hw_if_index, msg.u64); + + q = (u8) (msg.u64 & 0xFF); + if (q >= vui->num_qid) + { + vu_log_debug (vui, "invalid vring index VHOST_USER_SET_VRING_KICK:" + " %u >= %u", q, vui->num_qid); + goto close_socket; + } + + if (vui->vrings[q].kickfd_idx != ~0) + { + clib_file_t *uf = pool_elt_at_index (file_main.file_pool, + vui->vrings[q].kickfd_idx); + clib_file_del (&file_main, uf); + vui->vrings[q].kickfd_idx = ~0; + } + + if (!(msg.u64 & VHOST_USER_VRING_NOFD_MASK)) + { + if (number_of_fds != 1) + { + vu_log_debug (vui, "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 = clib_file_add (&file_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; + vhost_user_thread_placement (vui, q); + } + vhost_user_update_iface_state (vui); + break; + + case VHOST_USER_SET_VRING_ERR: + vu_log_debug (vui, "if %d msg VHOST_USER_SET_VRING_ERR %d", + vui->hw_if_index, msg.u64); + + q = (u8) (msg.u64 & 0xFF); + if (q >= vui->num_qid) + { + vu_log_debug (vui, "invalid vring index VHOST_USER_SET_VRING_ERR:" + " %u >= %u", q, vui->num_qid); + goto close_socket; + } + + if (vui->vrings[q].errfd != -1) + close (vui->vrings[q].errfd); + + if (!(msg.u64 & VHOST_USER_VRING_NOFD_MASK)) + { + 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: + vu_log_debug (vui, + "if %d msg VHOST_USER_SET_VRING_BASE idx %d num 0x%x", + vui->hw_if_index, msg.state.index, msg.state.num); + if (msg.state.index >= vui->num_qid) + { + vu_log_debug (vui, "invalid vring index VHOST_USER_SET_VRING_ADDR:" + " %u >= %u", msg.state.index, vui->num_qid); + goto close_socket; + } + vlib_worker_thread_barrier_sync (vm); + vui->vrings[msg.state.index].last_avail_idx = msg.state.num; + if (vhost_user_is_packed_ring_supported (vui)) + { + /* + * 0 1 2 3 + * 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 + * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ + * | last avail idx | | last used idx | | + * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ + * ^ ^ + * | | + * avail wrap counter used wrap counter + */ + /* last avail idx at bit 0-14. */ + vui->vrings[msg.state.index].last_avail_idx = + msg.state.num & 0x7fff; + /* avail wrap counter at bit 15 */ + vui->vrings[msg.state.index].avail_wrap_counter = + ! !(msg.state.num & (1 << 15)); + + /* + * Although last_used_idx is passed in the upper 16 bits in qemu + * implementation, in practice, last_avail_idx and last_used_idx are + * usually the same. As a result, DPDK does not bother to pass us + * last_used_idx. The spec is not clear on thex coding. I figured it + * out by reading the qemu code. So let's just read last_avail_idx + * and set last_used_idx equals to last_avail_idx. + */ + vui->vrings[msg.state.index].last_used_idx = + vui->vrings[msg.state.index].last_avail_idx; + vui->vrings[msg.state.index].last_kick = + vui->vrings[msg.state.index].last_used_idx; + vui->vrings[msg.state.index].used_wrap_counter = + vui->vrings[msg.state.index].avail_wrap_counter; + + if (vui->vrings[msg.state.index].avail_wrap_counter == 1) + vui->vrings[msg.state.index].avail_wrap_counter = + VRING_DESC_F_AVAIL; + } + vlib_worker_thread_barrier_release (vm); + break; + + case VHOST_USER_GET_VRING_BASE: + if (msg.state.index >= vui->num_qid) + { + vu_log_debug (vui, "invalid vring index VHOST_USER_GET_VRING_BASE:" + " %u >= %u", msg.state.index, vui->num_qid); + goto close_socket; + } + + /* protection is needed to prevent rx/tx from changing last_avail_idx */ + vlib_worker_thread_barrier_sync (vm); + /* + * Copy last_avail_idx from the vring before closing it because + * closing the vring also initializes the vring last_avail_idx + */ + msg.state.num = vui->vrings[msg.state.index].last_avail_idx; + if (vhost_user_is_packed_ring_supported (vui)) + { + msg.state.num = + (vui->vrings[msg.state.index].last_avail_idx & 0x7fff) | + (! !vui->vrings[msg.state.index].avail_wrap_counter << 15); + msg.state.num |= + ((vui->vrings[msg.state.index].last_used_idx & 0x7fff) | + (! !vui->vrings[msg.state.index].used_wrap_counter << 15)) << 16; + } + msg.flags |= 4; + msg.size = sizeof (msg.state); + + /* + * Spec says: Client must [...] stop ring upon receiving + * VHOST_USER_GET_VRING_BASE + */ + vhost_user_vring_close (vui, msg.state.index); + vlib_worker_thread_barrier_release (vm); + vu_log_debug (vui, + "if %d msg VHOST_USER_GET_VRING_BASE idx %d num 0x%x", + vui->hw_if_index, msg.state.index, msg.state.num); + n = + send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0); + if (n != (msg.size + VHOST_USER_MSG_HDR_SZ)) + { + vu_log_debug (vui, "could not send message response"); + goto close_socket; + } + vhost_user_update_iface_state (vui); + break; + + case VHOST_USER_NONE: + vu_log_debug (vui, "if %d msg VHOST_USER_NONE", vui->hw_if_index); + break; + + case VHOST_USER_SET_LOG_BASE: + vu_log_debug (vui, "if %d msg VHOST_USER_SET_LOG_BASE", + vui->hw_if_index); + + if (msg.size != sizeof (msg.log)) + { + vu_log_debug (vui, "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))) + { + vu_log_debug (vui, "VHOST_USER_PROTOCOL_F_LOG_SHMFD not set but " + "VHOST_USER_SET_LOG_BASE received"); + goto close_socket; + } + + fd = fds[0]; + /* align size to page */ + long page_sz = get_huge_page_size (fd); + ssize_t map_sz = + (msg.log.size + msg.log.offset + page_sz - 1) & ~(page_sz - 1); + + void *log_base_addr = mmap (0, map_sz, PROT_READ | PROT_WRITE, + MAP_SHARED, fd, 0); + + vu_log_debug (vui, "map log region addr 0 len 0x%lx off 0x%lx fd %d " + "mapped 0x%lx", map_sz, msg.log.offset, fd, + log_base_addr); + + if (log_base_addr == MAP_FAILED) + { + vu_log_err (vui, "failed to map memory. errno is %d", errno); + goto close_socket; + } + + vlib_worker_thread_barrier_sync (vm); + vui->log_base_addr = log_base_addr; + vui->log_base_addr += msg.log.offset; + vui->log_size = msg.log.size; + vlib_worker_thread_barrier_release (vm); + + msg.flags |= 4; + msg.size = sizeof (msg.u64); + n = + send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0); + if (n != (msg.size + VHOST_USER_MSG_HDR_SZ)) + { + vu_log_debug (vui, "could not send message response"); + goto close_socket; + } + break; + + case VHOST_USER_SET_LOG_FD: + vu_log_debug (vui, "if %d msg VHOST_USER_SET_LOG_FD", vui->hw_if_index); + break; + + case VHOST_USER_GET_PROTOCOL_FEATURES: + msg.flags |= 4; + msg.u64 = (1 << VHOST_USER_PROTOCOL_F_LOG_SHMFD) | + (1 << VHOST_USER_PROTOCOL_F_MQ); + msg.size = sizeof (msg.u64); + vu_log_debug (vui, "if %d msg VHOST_USER_GET_PROTOCOL_FEATURES - " + "reply 0x%016llx", vui->hw_if_index, msg.u64); + n = + send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0); + if (n != (msg.size + VHOST_USER_MSG_HDR_SZ)) + { + vu_log_debug (vui, "could not send message response"); + goto close_socket; + } + break; + + case VHOST_USER_SET_PROTOCOL_FEATURES: + vu_log_debug (vui, "if %d msg VHOST_USER_SET_PROTOCOL_FEATURES " + "features 0x%016llx", vui->hw_if_index, msg.u64); + vui->protocol_features = msg.u64; + break; + + case VHOST_USER_GET_QUEUE_NUM: + msg.flags |= 4; + msg.u64 = VHOST_VRING_MAX_MQ_PAIR_SZ; + msg.size = sizeof (msg.u64); + vu_log_debug (vui, "if %d msg VHOST_USER_GET_QUEUE_NUM - reply %d", + vui->hw_if_index, msg.u64); + n = + send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0); + if (n != (msg.size + VHOST_USER_MSG_HDR_SZ)) + { + vu_log_debug (vui, "could not send message response"); + goto close_socket; + } + break; + + case VHOST_USER_SET_VRING_ENABLE: + vu_log_debug (vui, "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 >= vui->num_qid) + { + vu_log_debug (vui, "invalid vring idx VHOST_USER_SET_VRING_ENABLE:" + " %u >= %u", msg.state.index, vui->num_qid); + goto close_socket; + } + + vui->vrings[msg.state.index].enabled = msg.state.num; + vhost_user_thread_placement (vui, msg.state.index); + vhost_user_update_iface_state (vui); + break; + + default: + vu_log_debug (vui, "unknown vhost-user message %d received. " + "closing socket", msg.request); + goto close_socket; + } + + return 0; + +close_socket: + vlib_worker_thread_barrier_sync (vm); + vhost_user_if_disconnect (vui); + vlib_worker_thread_barrier_release (vm); + vhost_user_update_iface_state (vui); + return 0; +} + +static clib_error_t * +vhost_user_socket_error (clib_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); + + vu_log_debug (vui, "socket error on if %d", vui->sw_if_index); + vlib_worker_thread_barrier_sync (vm); + vhost_user_if_disconnect (vui); + vlib_worker_thread_barrier_release (vm); + return 0; +} + +static clib_error_t * +vhost_user_socksvr_accept_ready (clib_file_t * uf) +{ + int client_fd, client_len; + struct sockaddr_un client; + clib_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"); + + if (vui->clib_file_index != ~0) + { + vu_log_debug (vui, "Close client socket for vhost interface %d, fd %d", + vui->sw_if_index, UNIX_GET_FD (vui->clib_file_index)); + clib_file_del (&file_main, file_main.file_pool + vui->clib_file_index); + } + + vu_log_debug (vui, "New client socket for vhost interface %d, fd %d", + vui->sw_if_index, client_fd); + 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; + template.description = format (0, "vhost interface %d", vui->sw_if_index); + vui->clib_file_index = clib_file_add (&file_main, &template); + vui->num_qid = 2; + return 0; +} + +static clib_error_t * +vhost_user_init (vlib_main_t * vm) +{ + vhost_user_main_t *vum = &vhost_user_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); + + vum->log_default = vlib_log_register_class ("vhost-user", 0); + + 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; + } + + vum->random = random_default_seed (); + + mhash_init_c_string (&vum->if_index_by_sock_name, sizeof (uword)); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_INIT_FUNCTION (vhost_user_init) = +{ + .runs_after = VLIB_INITS("ip4_init"), +}; +/* *INDENT-ON* */ + +static uword +vhost_user_send_interrupt_process (vlib_main_t * vm, + vlib_node_runtime_t * rt, vlib_frame_t * f) +{ + vhost_user_intf_t *vui; + f64 timeout = 3153600000.0 /* 100 years */ ; + uword event_type, *event_data = 0; + vhost_user_main_t *vum = &vhost_user_main; + u16 qid; + f64 now, poll_time_remaining; + f64 next_timeout; + u8 stop_timer = 0; + + while (1) + { + poll_time_remaining = + vlib_process_wait_for_event_or_clock (vm, timeout); + event_type = vlib_process_get_events (vm, &event_data); + vec_reset_length (event_data); + + /* + * Use the remaining timeout if it is less than coalesce time to avoid + * resetting the existing timer in the middle of expiration + */ + timeout = poll_time_remaining; + if (vlib_process_suspend_time_is_zero (timeout) || + (timeout > vum->coalesce_time)) + timeout = vum->coalesce_time; + + now = vlib_time_now (vm); + switch (event_type) + { + case VHOST_USER_EVENT_STOP_TIMER: + stop_timer = 1; + break; + + case VHOST_USER_EVENT_START_TIMER: + stop_timer = 0; + timeout = 1e-3; + if (!vlib_process_suspend_time_is_zero (poll_time_remaining)) + break; + /* fall through */ + + case ~0: + /* *INDENT-OFF* */ + pool_foreach (vui, vum->vhost_user_interfaces) { + next_timeout = timeout; + FOR_ALL_VHOST_RX_TXQ (qid, vui) + { + vhost_user_vring_t *vq = &vui->vrings[qid]; + + if (vq->started == 0) + continue; + if (vq->n_since_last_int) + { + if (now >= vq->int_deadline) + vhost_user_send_call (vm, vui, vq); + else + next_timeout = vq->int_deadline - now; + } + + if ((next_timeout < timeout) && (next_timeout > 0.0)) + timeout = next_timeout; + } + } + /* *INDENT-ON* */ + break; + + default: + clib_warning ("BUG: unhandled event type %d", event_type); + break; + } + /* No less than 1 millisecond */ + if (timeout < 1e-3) + timeout = 1e-3; + if (stop_timer) + timeout = 3153600000.0; + } + return 0; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (vhost_user_send_interrupt_node) = { + .function = vhost_user_send_interrupt_process, + .type = VLIB_NODE_TYPE_PROCESS, + .name = "vhost-user-send-interrupt-process", +}; +/* *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; + clib_file_t template = { 0 }; + f64 timeout = 3153600000.0 /* 100 years */ ; + uword *event_data = 0; + + sockfd = -1; + sun.sun_family = AF_UNIX; + template.read_function = vhost_user_socket_read; + template.error_function = vhost_user_socket_error; + + 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->clib_file_index == ~0) + { + if ((sockfd < 0) && + ((sockfd = socket (AF_UNIX, SOCK_STREAM, 0)) < 0)) + { + /* + * 1st time error or new error for this interface, + * spit out the message and record the error + */ + if (!vui->sock_errno || (vui->sock_errno != errno)) + { + clib_unix_warning + ("Error: Could not open unix socket for %s", + vui->sock_filename); + vui->sock_errno = errno; + } + continue; + } + + /* try to connect */ + strncpy (sun.sun_path, (char *) vui->sock_filename, + sizeof (sun.sun_path) - 1); + sun.sun_path[sizeof (sun.sun_path) - 1] = 0; + + /* Avoid hanging VPP if the other end does not accept */ + if (fcntl(sockfd, F_SETFL, O_NONBLOCK) < 0) + clib_unix_warning ("fcntl"); + + if (connect (sockfd, (struct sockaddr *) &sun, + sizeof (struct sockaddr_un)) == 0) + { + /* Set the socket to blocking as it was before */ + if (fcntl(sockfd, F_SETFL, 0) < 0) + clib_unix_warning ("fcntl2"); + + vui->sock_errno = 0; + template.file_descriptor = sockfd; + template.private_data = + vui - vhost_user_main.vhost_user_interfaces; + template.description = format (0, "vhost user process"); + vui->clib_file_index = clib_file_add (&file_main, &template); + vui->num_qid = 2; + + /* This sockfd is considered consumed */ + sockfd = -1; + } + 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->clib_file_index); + int retval = + getsockopt (fd, SOL_SOCKET, SO_ERROR, &error, &len); + + if (retval) + { + vu_log_debug (vui, "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) +{ + int q; + vhost_user_main_t *vum = &vhost_user_main; + + // disconnect interface sockets + vhost_user_if_disconnect (vui); + vhost_user_update_gso_interface_count (vui, 0 /* delete */ ); + vhost_user_update_iface_state (vui); + + for (q = 0; q < vec_len (vui->vrings); q++) + clib_spinlock_free (&vui->vrings[q].vring_lock); + + if (vui->unix_server_index != ~0) + { + //Close server socket + clib_file_t *uf = pool_elt_at_index (file_main.file_pool, + vui->unix_server_index); + clib_file_del (&file_main, uf); + vui->unix_server_index = ~0; + unlink (vui->sock_filename); + } + + mhash_unset (&vum->if_index_by_sock_name, vui->sock_filename, + &vui->if_index); +} + +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; + u16 qid; + + if (! + (hwif = + vnet_get_sup_hw_interface_api_visible_or_null (vnm, sw_if_index)) + || hwif->dev_class_index != vhost_user_device_class.index) + return VNET_API_ERROR_INVALID_SW_IF_INDEX; + + vui = pool_elt_at_index (vum->vhost_user_interfaces, hwif->dev_instance); + + vu_log_debug (vui, "Deleting vhost-user interface %s (instance %d)", + hwif->name, hwif->dev_instance); + + FOR_ALL_VHOST_TXQ (qid, vui) + { + vhost_user_vring_t *txvq = &vui->vrings[qid]; + + if ((txvq->mode == VNET_HW_IF_RX_MODE_POLLING) && + (txvq->thread_index != ~0)) + { + vhost_cpu_t *cpu = vec_elt_at_index (vum->cpus, txvq->thread_index); + ASSERT (cpu->polling_q_count != 0); + cpu->polling_q_count--; + } + + if ((vum->ifq_count > 0) && + ((txvq->mode == VNET_HW_IF_RX_MODE_INTERRUPT) || + (txvq->mode == VNET_HW_IF_RX_MODE_ADAPTIVE))) + { + vum->ifq_count--; + // Stop the timer if there is no more interrupt interface/queue + if (vum->ifq_count == 0) + { + vlib_process_signal_event (vm, + vhost_user_send_interrupt_node.index, + VHOST_USER_EVENT_STOP_TIMER, 0); + break; + } + } + } + + // Disable and reset interface + vhost_user_term_if (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); + + // free vrings + vec_free (vui->vrings); + + // Back to pool + pool_put (vum->vhost_user_interfaces, vui); + + return rv; +} + +static clib_error_t * +vhost_user_exit (vlib_main_t * vm) +{ + vnet_main_t *vnm = vnet_get_main (); + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + + vlib_worker_thread_barrier_sync (vlib_get_main ()); + /* *INDENT-OFF* */ + pool_foreach (vui, vum->vhost_user_interfaces) { + vhost_user_delete_if (vnm, vm, vui->sw_if_index); + } + /* *INDENT-ON* */ + vlib_worker_thread_barrier_release (vlib_get_main ()); + return 0; +} + +VLIB_MAIN_LOOP_EXIT_FUNCTION (vhost_user_exit); + +/** + * 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, + vhost_user_create_if_args_t *args) +{ + vhost_user_main_t *vum = &vhost_user_main; + vnet_eth_interface_registration_t eir = {}; + u8 hwaddr[6]; + + /* create hw and sw interface */ + if (args->use_custom_mac) + { + clib_memcpy (hwaddr, args->hwaddr, 6); + } + else + { + random_u32 (&vum->random); + clib_memcpy (hwaddr + 2, &vum->random, sizeof (vum->random)); + hwaddr[0] = 2; + hwaddr[1] = 0xfe; + } + + eir.dev_class_index = vhost_user_device_class.index; + eir.dev_instance = vui - vum->vhost_user_interfaces /* device instance */, + eir.address = hwaddr; + vui->hw_if_index = vnet_eth_register_interface (vnm, &eir); +} + +/* + * Initialize vui with specified attributes + */ +static void +vhost_user_vui_init (vnet_main_t * vnm, vhost_user_intf_t * vui, + int server_sock_fd, vhost_user_create_if_args_t * args, + u32 * sw_if_index) +{ + vnet_sw_interface_t *sw; + int q; + vhost_user_main_t *vum = &vhost_user_main; + + sw = vnet_get_hw_sw_interface (vnm, vui->hw_if_index); + if (server_sock_fd != -1) + { + clib_file_t template = { 0 }; + template.read_function = vhost_user_socksvr_accept_ready; + template.file_descriptor = server_sock_fd; + template.private_data = vui - vum->vhost_user_interfaces; //hw index + template.description = format (0, "vhost user %d", sw); + vui->unix_server_index = clib_file_add (&file_main, &template); + } + else + { + vui->unix_server_index = ~0; + } + + vui->sw_if_index = sw->sw_if_index; + strncpy (vui->sock_filename, args->sock_filename, + ARRAY_LEN (vui->sock_filename) - 1); + vui->sock_errno = 0; + vui->is_ready = 0; + vui->feature_mask = args->feature_mask; + vui->clib_file_index = ~0; + vui->log_base_addr = 0; + vui->if_index = vui - vum->vhost_user_interfaces; + vui->enable_gso = args->enable_gso; + vui->enable_event_idx = args->enable_event_idx; + vui->enable_packed = args->enable_packed; + /* + * enable_gso takes precedence over configurable feature mask if there + * is a clash. + * if feature mask disables gso, but enable_gso is configured, + * then gso is enable + * if feature mask enables gso, but enable_gso is not configured, + * then gso is enable + * + * if gso is enable via feature mask, it must enable both host and guest + * gso feature mask, we don't support one sided GSO or partial GSO. + */ + if ((vui->enable_gso == 0) && + ((args->feature_mask & FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS) + == (FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS))) + vui->enable_gso = 1; + vhost_user_update_gso_interface_count (vui, 1 /* add */ ); + mhash_set_mem (&vum->if_index_by_sock_name, vui->sock_filename, + &vui->if_index, 0); + + vec_validate_aligned (vui->vrings, (VHOST_VRING_INIT_MQ_PAIR_SZ << 1) - 1, + CLIB_CACHE_LINE_BYTES); + vui->num_qid = 2; + for (q = 0; q < vec_len (vui->vrings); q++) + vhost_user_vring_init (vui, q); + + vnet_hw_if_set_caps (vnm, vui->hw_if_index, VNET_HW_IF_CAP_INT_MODE); + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); + + if (sw_if_index) + *sw_if_index = vui->sw_if_index; +} + +int +vhost_user_create_if (vnet_main_t * vnm, vlib_main_t * vm, + vhost_user_create_if_args_t * args) +{ + vhost_user_intf_t *vui = NULL; + u32 sw_if_idx = ~0; + int rv = 0; + int server_sock_fd = -1; + vhost_user_main_t *vum = &vhost_user_main; + uword *if_index; + + if (args->sock_filename == NULL || !(strlen (args->sock_filename) > 0)) + { + return VNET_API_ERROR_INVALID_ARGUMENT; + } + + if_index = mhash_get (&vum->if_index_by_sock_name, + (void *) args->sock_filename); + if (if_index) + { + vui = &vum->vhost_user_interfaces[*if_index]; + args->sw_if_index = vui->sw_if_index; + return VNET_API_ERROR_IF_ALREADY_EXISTS; + } + + if (args->is_server) + { + if ((rv = + vhost_user_init_server_sock (args->sock_filename, + &server_sock_fd)) != 0) + { + return rv; + } + } + + /* Protect the uninitialized vui from being dispatched by rx/tx */ + vlib_worker_thread_barrier_sync (vm); + pool_get (vhost_user_main.vhost_user_interfaces, vui); + vhost_user_create_ethernet (vnm, vm, vui, args); + vlib_worker_thread_barrier_release (vm); + + vhost_user_vui_init (vnm, vui, server_sock_fd, args, &sw_if_idx); + vnet_sw_interface_set_mtu (vnm, vui->sw_if_index, 9000); + vhost_user_rx_thread_placement (vui, 1); + + if (args->renumber) + vnet_interface_name_renumber (sw_if_idx, args->custom_dev_instance); + + args->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, + vhost_user_create_if_args_t * args) +{ + 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; + uword *if_index; + + if (!(hwif = vnet_get_sup_hw_interface_api_visible_or_null (vnm, + args->sw_if_index)) + || hwif->dev_class_index != vhost_user_device_class.index) + return VNET_API_ERROR_INVALID_SW_IF_INDEX; + + if (args->sock_filename == NULL || !(strlen (args->sock_filename) > 0)) + return VNET_API_ERROR_INVALID_ARGUMENT; + + vui = vec_elt_at_index (vum->vhost_user_interfaces, hwif->dev_instance); + + /* + * Disallow changing the interface to have the same path name + * as other interface + */ + if_index = mhash_get (&vum->if_index_by_sock_name, + (void *) args->sock_filename); + if (if_index && (*if_index != vui->if_index)) + return VNET_API_ERROR_IF_ALREADY_EXISTS; + + // First try to open server socket + if (args->is_server) + if ((rv = vhost_user_init_server_sock (args->sock_filename, + &server_sock_fd)) != 0) + return rv; + + vhost_user_term_if (vui); + vhost_user_vui_init (vnm, vui, server_sock_fd, args, &sw_if_idx); + + if (args->renumber) + vnet_interface_name_renumber (sw_if_idx, args->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) +{ + vnet_main_t *vnm = vnet_get_main (); + unformat_input_t _line_input, *line_input = &_line_input; + clib_error_t *error = NULL; + vhost_user_create_if_args_t args = { 0 }; + int rv; + + /* Get a line of input. */ + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + args.feature_mask = (u64) ~ (0ULL); + args.custom_dev_instance = ~0; + /* GSO feature is disable by default */ + args.feature_mask &= ~FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS; + /* packed-ring feature is disable by default */ + args.feature_mask &= ~VIRTIO_FEATURE (VIRTIO_F_RING_PACKED); + /* event_idx feature is disable by default */ + args.feature_mask &= ~VIRTIO_FEATURE (VIRTIO_RING_F_EVENT_IDX); + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "socket %s", &args.sock_filename)) + ; + else if (unformat (line_input, "server")) + args.is_server = 1; + else if (unformat (line_input, "gso")) + args.enable_gso = 1; + else if (unformat (line_input, "packed")) + args.enable_packed = 1; + else if (unformat (line_input, "event-idx")) + args.enable_event_idx = 1; + else if (unformat (line_input, "feature-mask 0x%llx", + &args.feature_mask)) + ; + else if (unformat (line_input, "hwaddr %U", unformat_ethernet_address, + args.hwaddr)) + args.use_custom_mac = 1; + else if (unformat (line_input, "renumber %d", + &args.custom_dev_instance)) + args.renumber = 1; + else + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, line_input); + goto done; + } + } + + if ((rv = vhost_user_create_if (vnm, vm, &args))) + { + error = clib_error_return (0, "vhost_user_create_if returned %d", rv); + goto done; + } + + vlib_cli_output (vm, "%U\n", format_vnet_sw_if_index_name, vnm, + args.sw_if_index); + +done: + vec_free (args.sock_filename); + unformat_free (line_input); + + return error; +} + +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 (); + clib_error_t *error = 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, "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_api_visible_or_null (vnm, sw_if_index); + if (hwif == NULL || + vhost_user_device_class.index != hwif->dev_class_index) + { + error = clib_error_return (0, "Not a vhost interface"); + goto done; + } + } + else + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, line_input); + goto done; + } + } + + vhost_user_delete_if (vnm, vm, sw_if_index); + +done: + unformat_free (line_input); + + return error; +} + +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; + 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->is_server = vui->unix_server_index != ~0; + vuid->sock_errno = vui->sock_errno; + snprintf ((char *) vuid->sock_filename, sizeof (vuid->sock_filename), + "%s", vui->sock_filename); + memcpy_s (vuid->if_name, sizeof (vuid->if_name), hi->name, + clib_min (vec_len (hi->name), sizeof (vuid->if_name) - 1)); + vuid->if_name[sizeof (vuid->if_name) - 1] = 0; + } + + vec_free (hw_if_indices); + + *out_vuids = r_vuids; + + return rv; +} + +static u8 * +format_vhost_user_desc (u8 * s, va_list * args) +{ + char *fmt = va_arg (*args, char *); + vhost_user_intf_t *vui = va_arg (*args, vhost_user_intf_t *); + vnet_virtio_vring_desc_t *desc_table = + va_arg (*args, vnet_virtio_vring_desc_t *); + int idx = va_arg (*args, int); + u32 *mem_hint = va_arg (*args, u32 *); + + s = format (s, fmt, idx, desc_table[idx].addr, desc_table[idx].len, + desc_table[idx].flags, desc_table[idx].next, + pointer_to_uword (map_guest_mem (vui, desc_table[idx].addr, + mem_hint))); + return s; +} + +static void +vhost_user_show_fds (vlib_main_t * vm, vhost_user_vring_t * vq) +{ + int kickfd = UNIX_GET_FD (vq->kickfd_idx); + int callfd = UNIX_GET_FD (vq->callfd_idx); + + vlib_cli_output (vm, " kickfd %d callfd %d errfd %d\n", kickfd, callfd, + vq->errfd); +} + +static void +vhost_user_show_desc (vlib_main_t * vm, vhost_user_intf_t * vui, int q, + int show_descr, int show_verbose) +{ + int j; + u32 mem_hint = 0; + u32 idx; + u32 n_entries; + vnet_virtio_vring_desc_t *desc_table; + vhost_user_vring_t *vq = &vui->vrings[q]; + + if (vq->avail && vq->used) + vlib_cli_output (vm, + " avail.flags %x avail event idx %u avail.idx %d " + "used.flags %x used event idx %u used.idx %d\n", + vq->avail->flags, vhost_user_avail_event_idx (vq), + vq->avail->idx, vq->used->flags, + vhost_user_used_event_idx (vq), vq->used->idx); + + vhost_user_show_fds (vm, vq); + + if (show_descr) + { + vlib_cli_output (vm, "\n descriptor table:\n"); + vlib_cli_output (vm, + " slot addr len flags next " + "user_addr\n"); + vlib_cli_output (vm, + " ===== ================== ===== ====== ===== " + "==================\n"); + for (j = 0; j < vq->qsz_mask + 1; j++) + { + desc_table = vq->desc; + vlib_cli_output (vm, "%U", format_vhost_user_desc, + " %-5d 0x%016lx %-5d 0x%04x %-5d 0x%016lx\n", vui, + desc_table, j, &mem_hint); + if (show_verbose && (desc_table[j].flags & VRING_DESC_F_INDIRECT)) + { + n_entries = + desc_table[j].len / sizeof (vnet_virtio_vring_desc_t); + desc_table = map_guest_mem (vui, desc_table[j].addr, &mem_hint); + if (desc_table) + { + for (idx = 0; idx < clib_min (20, n_entries); idx++) + { + vlib_cli_output + (vm, "%U", format_vhost_user_desc, + "> %-4u 0x%016lx %-5u 0x%04x %-5u 0x%016lx\n", vui, + desc_table, idx, &mem_hint); + } + if (n_entries >= 20) + vlib_cli_output (vm, "Skip displaying entries 20...%u\n", + n_entries); + } + } + } + } +} + +static u8 * +format_vhost_user_packed_desc (u8 * s, va_list * args) +{ + char *fmt = va_arg (*args, char *); + vhost_user_intf_t *vui = va_arg (*args, vhost_user_intf_t *); + vnet_virtio_vring_packed_desc_t *desc_table = + va_arg (*args, vnet_virtio_vring_packed_desc_t *); + int idx = va_arg (*args, int); + u32 *mem_hint = va_arg (*args, u32 *); + + s = format (s, fmt, idx, desc_table[idx].addr, desc_table[idx].len, + desc_table[idx].flags, desc_table[idx].id, + pointer_to_uword (map_guest_mem (vui, desc_table[idx].addr, + mem_hint))); + return s; +} + +static u8 * +format_vhost_user_event_idx_flags (u8 * s, va_list * args) +{ + u32 flags = va_arg (*args, u32); + typedef struct + { + u8 value; + char *str; + } event_idx_flags; + static event_idx_flags event_idx_array[] = { +#define _(s,v) { .str = #s, .value = v, }, + foreach_virtio_event_idx_flags +#undef _ + }; + u32 num_entries = sizeof (event_idx_array) / sizeof (event_idx_flags); + + if (flags < num_entries) + s = format (s, "%s", event_idx_array[flags].str); + else + s = format (s, "%u", flags); + return s; +} + +static void +vhost_user_show_desc_packed (vlib_main_t * vm, vhost_user_intf_t * vui, int q, + int show_descr, int show_verbose) +{ + int j; + u32 mem_hint = 0; + u32 idx; + u32 n_entries; + vnet_virtio_vring_packed_desc_t *desc_table; + vhost_user_vring_t *vq = &vui->vrings[q]; + u16 off_wrap, event_idx; + + off_wrap = vq->avail_event->off_wrap; + event_idx = off_wrap & 0x7fff; + vlib_cli_output (vm, " avail_event.flags %U avail_event.off_wrap %u " + "avail event idx %u\n", format_vhost_user_event_idx_flags, + (u32) vq->avail_event->flags, off_wrap, event_idx); + + off_wrap = vq->used_event->off_wrap; + event_idx = off_wrap & 0x7fff; + vlib_cli_output (vm, " used_event.flags %U used_event.off_wrap %u " + "used event idx %u\n", format_vhost_user_event_idx_flags, + (u32) vq->used_event->flags, off_wrap, event_idx); + + vlib_cli_output (vm, " avail wrap counter %u, used wrap counter %u\n", + vq->avail_wrap_counter, vq->used_wrap_counter); + + vhost_user_show_fds (vm, vq); + + if (show_descr) + { + vlib_cli_output (vm, "\n descriptor table:\n"); + vlib_cli_output (vm, + " slot addr len flags id " + "user_addr\n"); + vlib_cli_output (vm, + " ===== ================== ===== ====== ===== " + "==================\n"); + for (j = 0; j < vq->qsz_mask + 1; j++) + { + desc_table = vq->packed_desc; + vlib_cli_output (vm, "%U", format_vhost_user_packed_desc, + " %-5u 0x%016lx %-5u 0x%04x %-5u 0x%016lx\n", vui, + desc_table, j, &mem_hint); + if (show_verbose && (desc_table[j].flags & VRING_DESC_F_INDIRECT)) + { + n_entries = desc_table[j].len >> 4; + desc_table = map_guest_mem (vui, desc_table[j].addr, &mem_hint); + if (desc_table) + { + for (idx = 0; idx < clib_min (20, n_entries); idx++) + { + vlib_cli_output + (vm, "%U", format_vhost_user_packed_desc, + "> %-4u 0x%016lx %-5u 0x%04x %-5u 0x%016lx\n", vui, + desc_table, idx, &mem_hint); + } + if (n_entries >= 20) + vlib_cli_output (vm, "Skip displaying entries 20...%u\n", + n_entries); + } + } + } + } +} + +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; + u16 qid; + int i, j, q; + int show_descr = 0; + int show_verbose = 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_features +#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)) + { + hi = vnet_get_hw_interface (vnm, hw_if_index); + if (vhost_user_device_class.index != hi->dev_class_index) + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + goto done; + } + vec_add1 (hw_if_indices, hw_if_index); + } + else if (unformat (input, "descriptors") || unformat (input, "desc")) + show_descr = 1; + else if (unformat (input, "verbose")) + show_verbose = 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); + vlib_cli_output (vm, " Number of rx virtqueues in interrupt mode: %d", + vum->ifq_count); + vlib_cli_output (vm, " Number of GSO interfaces: %d", vum->gso_count); + for (u32 tid = 0; tid <= vlib_num_workers (); tid++) + { + vhost_cpu_t *cpu = vec_elt_at_index (vum->cpus, tid); + vlib_cli_output (vm, " Thread %u: Polling queue count %u", tid, + cpu->polling_q_count); + } + + 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: %U (ifindex %d)", + format_vnet_hw_if_index_name, vnm, hw_if_indices[i], + hw_if_indices[i]); + vlib_cli_output (vm, " Number of qids %u", vui->num_qid); + if (vui->enable_gso) + vlib_cli_output (vm, " GSO enable"); + if (vui->enable_packed) + vlib_cli_output (vm, " Packed ring enable"); + if (vui->enable_event_idx) + vlib_cli_output (vm, " Event index enable"); + + 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: "); + + FOR_ALL_VHOST_TXQ (qid, vui) + { + vhost_user_vring_t *txvq = &vui->vrings[qid]; + + if (txvq->qid == -1) + continue; + vlib_cli_output (vm, " thread %d on vring %d, %U\n", + txvq->thread_index, qid, format_vnet_hw_if_rx_mode, + txvq->mode); + } + + vlib_cli_output (vm, " tx placement\n"); + + FOR_ALL_VHOST_RXQ (qid, vui) + { + vhost_user_vring_t *rxvq = &vui->vrings[qid]; + vnet_hw_if_tx_queue_t *txq; + + if (rxvq->queue_index == ~0) + continue; + txq = vnet_hw_if_get_tx_queue (vnm, rxvq->queue_index); + if (txq->threads) + vlib_cli_output (vm, " threads %U on vring %u: %s\n", + format_bitmap_list, txq->threads, qid, + txq->shared_queue ? "spin-lock" : "lock-free"); + } + + 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_ALL_VHOST_RX_TXQ (q, vui) + { + 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, " global %s queue index %u\n", + (q & 1) ? "RX" : "TX", vui->vrings[q].queue_index); + + vlib_cli_output ( + vm, + " qsz %d last_avail_idx %d last_used_idx %d" + " last_kick %u\n", + vui->vrings[q].qsz_mask + 1, vui->vrings[q].last_avail_idx, + vui->vrings[q].last_used_idx, vui->vrings[q].last_kick); + + if (vhost_user_is_packed_ring_supported (vui)) + vhost_user_show_desc_packed (vm, vui, q, show_descr, show_verbose); + else + vhost_user_show_desc (vm, vui, q, show_descr, show_verbose); + } + 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 + * hypervisor and VPP to manage the vHost interface. If in server + * mode, VPP will create the socket if it does not already exist. If in + * client mode, hypervisor will create the socket if it does not + * already exist. The VPP code is indifferent to the file location. However, + * if SELinux is enabled, then the socket needs to be created in + * /var/run/vpp/. + * + * - server - Optional flag to indicate that VPP should be the server + * for the linux socket. If not provided, VPP will be the client. In + * server mode, the VM can be reset without tearing down the vHost + * Interface. In client mode, VPP can be reset without bringing down + * the VM and tearing down the vHost Interface. + * + * - feature-mask - Optional virtio/vhost feature set negotiated + * at startup. This is intended for degugging only. It is recommended + * that this parameter not be used except by experienced users. 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 /var/run/vpp/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 /var/run/vpp/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 ] [gso] " + "[packed] [event-idx]", + .function = vhost_user_connect_command_fn, + .is_mp_safe = 1, +}; +/* *INDENT-ON* */ + +/*? + * Delete a vHost User interface using the interface name or the + * software interface index. Use the 'show interface' + * 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 + * separated 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 /var/run/vpp/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 0x00000000 0x000a0000 0xaac00000 0x00000000 0x2b400000 + * 1 61 0x000c0000 0x3ff40000 0xaacc0000 0x000c0000 0xabcc0000 + * + * 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] [verbose]]", + .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); + } +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/plugins/vhost/vhost_user.h b/src/plugins/vhost/vhost_user.h new file mode 100644 index 00000000000..3479d65cbc6 --- /dev/null +++ b/src/plugins/vhost/vhost_user.h @@ -0,0 +1,388 @@ +/* + * 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__ + +#include +#include + +/* vhost-user data structures */ + +#define VHOST_MEMORY_MAX_NREGIONS 8 +#define VHOST_USER_MSG_HDR_SZ 12 +#define VHOST_VRING_INIT_MQ_PAIR_SZ 8 //8TX + 8RX + +/* + * qid is one byte in size in the spec. Please see VHOST_USER_SET_VRING_CALL, + * VHOST_USER_SET_VRING_KICK, and VHOST_USER_SET_VRING_ERR. + * The max number for q pair is naturally 128. + */ +#define VHOST_VRING_MAX_MQ_PAIR_SZ 128 +#define VHOST_VRING_IDX_RX(qid) (2 * (qid)) +#define VHOST_VRING_IDX_TX(qid) (2 * (qid) + 1) + +#define VHOST_USER_VRING_NOFD_MASK 0x100 + +#define VHOST_USER_PROTOCOL_F_MQ 0 +#define VHOST_USER_PROTOCOL_F_LOG_SHMFD 1 +#define VHOST_VRING_F_LOG 0 + +#define VHOST_USER_PROTOCOL_FEATURES ((1ULL << VHOST_USER_PROTOCOL_F_MQ) | \ + (1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD)) + +#define vu_log_debug(dev, f, ...) \ +{ \ + vlib_log(VLIB_LOG_LEVEL_DEBUG, vhost_user_main.log_default, "%U: " f, \ + format_vnet_hw_if_index_name, vnet_get_main(), \ + dev->hw_if_index, ##__VA_ARGS__); \ +}; + +#define vu_log_warn(dev, f, ...) \ +{ \ + vlib_log(VLIB_LOG_LEVEL_WARNING, vhost_user_main.log_default, "%U: " f, \ + format_vnet_hw_if_index_name, vnet_get_main(), \ + dev->hw_if_index, ##__VA_ARGS__); \ +}; +#define vu_log_err(dev, f, ...) \ +{ \ + vlib_log(VLIB_LOG_LEVEL_ERR, vhost_user_main.log_default, "%U: " f, \ + format_vnet_hw_if_index_name, vnet_get_main(), \ + dev->hw_if_index, ##__VA_ARGS__); \ +}; + +#define UNIX_GET_FD(unixfd_idx) ({ \ + typeof(unixfd_idx) __unixfd_idx = (unixfd_idx); \ + (__unixfd_idx != ~0) ? \ + pool_elt_at_index (file_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; + +#define FEATURE_VIRTIO_NET_F_HOST_TSO_FEATURE_BITS \ + (VIRTIO_FEATURE (VIRTIO_NET_F_CSUM) | \ + VIRTIO_FEATURE (VIRTIO_NET_F_HOST_UFO) | \ + VIRTIO_FEATURE (VIRTIO_NET_F_HOST_TSO4) | \ + VIRTIO_FEATURE (VIRTIO_NET_F_HOST_TSO6)) + +#define FEATURE_VIRTIO_NET_F_GUEST_TSO_FEATURE_BITS \ + (VIRTIO_FEATURE (VIRTIO_NET_F_GUEST_CSUM) | \ + VIRTIO_FEATURE (VIRTIO_NET_F_GUEST_UFO) | \ + VIRTIO_FEATURE (VIRTIO_NET_F_GUEST_TSO4) | \ + VIRTIO_FEATURE (VIRTIO_NET_F_GUEST_TSO6)) + +#define FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS \ + (FEATURE_VIRTIO_NET_F_HOST_TSO_FEATURE_BITS | \ + FEATURE_VIRTIO_NET_F_GUEST_TSO_FEATURE_BITS) + + +typedef struct +{ + char *sock_filename; + u64 feature_mask; + u32 custom_dev_instance; + u8 hwaddr[6]; + u8 renumber; + u8 is_server; + u8 enable_gso; + u8 enable_packed; + u8 enable_event_idx; + u8 use_custom_mac; + + /* return */ + u32 sw_if_index; +} vhost_user_create_if_args_t; + +int vhost_user_create_if (vnet_main_t * vnm, vlib_main_t * vm, + vhost_user_create_if_args_t * args); +int vhost_user_modify_if (vnet_main_t * vnm, vlib_main_t * vm, + vhost_user_create_if_args_t * args); +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 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; + +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_mask; + u16 last_avail_idx; + u16 last_used_idx; + u16 n_since_last_int; + union + { + vnet_virtio_vring_desc_t *desc; + vnet_virtio_vring_packed_desc_t *packed_desc; + }; + union + { + vnet_virtio_vring_avail_t *avail; + vnet_virtio_vring_desc_event_t *avail_event; + }; + union + { + vnet_virtio_vring_used_t *used; + vnet_virtio_vring_desc_event_t *used_event; + }; + uword desc_user_addr; + uword used_user_addr; + uword avail_user_addr; + f64 int_deadline; + u8 started; + u8 enabled; + u8 log_used; + clib_spinlock_t vring_lock; + + //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; + + /* The rx queue policy (interrupt/adaptive/polling) for this queue */ + u32 mode; + + /* + * It contains the device queue number. -1 if it does not. The idea is + * to not invoke vnet_hw_interface_assign_rx_thread and + * vnet_hw_interface_unassign_rx_thread more than once for the duration of + * the interface even if it is disconnected and reconnected. + */ + i16 qid; + + u16 used_wrap_counter; + u16 avail_wrap_counter; + u16 last_kick; + u8 first_kick; + u32 queue_index; + u32 thread_index; +} vhost_user_vring_t; + +#define VHOST_USER_EVENT_START_TIMER 1 +#define VHOST_USER_EVENT_STOP_TIMER 2 + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + u32 is_ready; + u32 admin_up; + u32 unix_server_index; + u32 clib_file_index; + char sock_filename[256]; + int sock_errno; + uword if_index; + 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; + + /* + * vrings is a dynamic array. It may have more elements than it is + * currently used. num_qid indicates the current total qid's in the + * vrings. For example, vec_len(vrings) = 64, num_qid = 60, so the + * current valid/used qid is (0, 59) in the vrings array. + */ + u32 num_qid; + + int virtio_net_hdr_sz; + int is_any_layout; + + void *log_base_addr; + u64 log_size; + + u8 enable_gso; + + /* Packed ring configured */ + u8 enable_packed; + + u8 enable_event_idx; +} vhost_user_intf_t; + +#define FOR_ALL_VHOST_TXQ(qid, vui) for (qid = 1; qid < vui->num_qid; qid += 2) + +#define FOR_ALL_VHOST_RXQ(qid, vui) for (qid = 0; qid < vui->num_qid; qid += 2) + +#define FOR_ALL_VHOST_RX_TXQ(qid, vui) for (qid = 0; qid < vui->num_qid; qid++) + +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 **/ + vnet_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 +{ + u32 rx_buffers_len; + u32 rx_buffers[VHOST_USER_RX_BUFFERS_N]; + + vnet_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; + + u32 *to_next_list; + vlib_buffer_t **rx_buffers_pdesc; + u32 polling_q_count; +} vhost_cpu_t; + +typedef struct +{ + mhash_t if_index_by_sock_name; + 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; + + /** Per-CPU data for vhost-user */ + vhost_cpu_t *cpus; + + /** Pseudo random iterator */ + u32 random; + + /* The number of rx interface/queue pairs in interrupt mode */ + u32 ifq_count; + + /* logging */ + vlib_log_class_t log_default; + + /* gso interface count */ + u32 gso_count; +} 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); +void vhost_user_set_operation_mode (vhost_user_intf_t *vui, + vhost_user_vring_t *txvq); + +extern vlib_node_registration_t vhost_user_send_interrupt_node; +extern vnet_device_class_t vhost_user_device_class; +extern vlib_node_registration_t vhost_user_input_node; +extern vhost_user_main_t vhost_user_main; + +#endif + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/plugins/vhost/vhost_user_api.c b/src/plugins/vhost/vhost_user_api.c new file mode 100644 index 00000000000..d522803a011 --- /dev/null +++ b/src/plugins/vhost/vhost_user_api.c @@ -0,0 +1,358 @@ +/* + *------------------------------------------------------------------ + * 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 +#include + +#include +#include +#include + +#define REPLY_MSG_ID_BASE msg_id_base +#include + +static u16 msg_id_base; + +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; + vnet_main_t *vnm = vnet_get_main (); + vlib_main_t *vm = vlib_get_main (); + u64 disabled_features = (u64) (0ULL); + vhost_user_create_if_args_t args = { 0 }; + + args.sw_if_index = (u32) ~ 0; + args.feature_mask = (u64) ~ (0ULL); + if (mp->disable_mrg_rxbuf) + disabled_features = VIRTIO_FEATURE (VIRTIO_NET_F_MRG_RXBUF); + + if (mp->disable_indirect_desc) + disabled_features |= VIRTIO_FEATURE (VIRTIO_RING_F_INDIRECT_DESC); + + /* + * GSO and PACKED are not supported by feature mask via binary API. We + * disable GSO and PACKED feature in the feature mask. They may be enabled + * explicitly via enable_gso and enable_packed argument + */ + disabled_features |= FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS | + VIRTIO_FEATURE (VIRTIO_F_RING_PACKED); + + /* EVENT_IDX is disabled by default */ + disabled_features |= VIRTIO_FEATURE (VIRTIO_RING_F_EVENT_IDX); + args.feature_mask &= ~disabled_features; + + if (mp->use_custom_mac) + mac_address_decode (mp->mac_address, (mac_address_t *) args.hwaddr); + + args.use_custom_mac = mp->use_custom_mac; + args.is_server = mp->is_server; + args.sock_filename = (char *) mp->sock_filename; + args.renumber = mp->renumber; + args.custom_dev_instance = ntohl (mp->custom_dev_instance); + args.enable_gso = mp->enable_gso; + args.enable_packed = mp->enable_packed; + rv = vhost_user_create_if (vnm, vm, &args); + + /* 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, args.sw_if_index); + } + } + + /* *INDENT-OFF* */ + REPLY_MACRO2(VL_API_CREATE_VHOST_USER_IF_REPLY, + ({ + rmp->sw_if_index = ntohl (args.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; + u64 disabled_features = (u64) (0ULL); + vhost_user_create_if_args_t args = { 0 }; + vnet_main_t *vnm = vnet_get_main (); + vlib_main_t *vm = vlib_get_main (); + + args.feature_mask = (u64) ~ (0ULL); + /* + * GSO and PACKED are not supported by feature mask via binary API. We + * disable GSO and PACKED feature in the feature mask. They may be enabled + * explicitly via enable_gso and enable_packed argument + */ + disabled_features |= FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS | + VIRTIO_FEATURE (VIRTIO_F_RING_PACKED); + + /* EVENT_IDX is disabled by default */ + disabled_features |= VIRTIO_FEATURE (VIRTIO_RING_F_EVENT_IDX); + args.feature_mask &= ~disabled_features; + + args.sw_if_index = ntohl (mp->sw_if_index); + args.sock_filename = (char *) mp->sock_filename; + args.is_server = mp->is_server; + args.renumber = mp->renumber; + args.custom_dev_instance = ntohl (mp->custom_dev_instance); + args.enable_gso = mp->enable_gso; + args.enable_packed = mp->enable_packed; + rv = vhost_user_modify_if (vnm, vm, &args); + + REPLY_MACRO (VL_API_MODIFY_VHOST_USER_IF_REPLY); +} + +static void +vl_api_create_vhost_user_if_v2_t_handler (vl_api_create_vhost_user_if_v2_t * + mp) +{ + int rv = 0; + vl_api_create_vhost_user_if_v2_reply_t *rmp; + vnet_main_t *vnm = vnet_get_main (); + vlib_main_t *vm = vlib_get_main (); + u64 disabled_features = (u64) (0ULL); + vhost_user_create_if_args_t args = { 0 }; + + args.sw_if_index = (u32) ~ 0; + args.feature_mask = (u64) ~ (0ULL); + if (mp->disable_mrg_rxbuf) + disabled_features = VIRTIO_FEATURE (VIRTIO_NET_F_MRG_RXBUF); + + if (mp->disable_indirect_desc) + disabled_features |= VIRTIO_FEATURE (VIRTIO_RING_F_INDIRECT_DESC); + + /* + * GSO and PACKED are not supported by feature mask via binary API. We + * disable GSO and PACKED feature in the feature mask. They may be enabled + * explicitly via enable_gso and enable_packed argument + */ + disabled_features |= FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS | + VIRTIO_FEATURE (VIRTIO_F_RING_PACKED); + + /* EVENT_IDX is disabled by default */ + disabled_features |= VIRTIO_FEATURE (VIRTIO_RING_F_EVENT_IDX); + args.feature_mask &= ~disabled_features; + + if (mp->use_custom_mac) + mac_address_decode (mp->mac_address, (mac_address_t *) args.hwaddr); + + args.use_custom_mac = mp->use_custom_mac; + args.is_server = mp->is_server; + args.sock_filename = (char *) mp->sock_filename; + args.renumber = mp->renumber; + args.custom_dev_instance = ntohl (mp->custom_dev_instance); + args.enable_gso = mp->enable_gso; + args.enable_packed = mp->enable_packed; + args.enable_event_idx = mp->enable_event_idx; + rv = vhost_user_create_if (vnm, vm, &args); + + /* 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, args.sw_if_index); + } + } + + /* *INDENT-OFF* */ + REPLY_MACRO2(VL_API_CREATE_VHOST_USER_IF_V2_REPLY, + ({ + rmp->sw_if_index = ntohl (args.sw_if_index); + })); + /* *INDENT-ON* */ +} + +static void +vl_api_modify_vhost_user_if_v2_t_handler (vl_api_modify_vhost_user_if_v2_t * + mp) +{ + int rv = 0; + vl_api_modify_vhost_user_if_v2_reply_t *rmp; + u64 disabled_features = (u64) (0ULL); + vhost_user_create_if_args_t args = { 0 }; + vnet_main_t *vnm = vnet_get_main (); + vlib_main_t *vm = vlib_get_main (); + + args.feature_mask = (u64) ~ (0ULL); + /* + * GSO and PACKED are not supported by feature mask via binary API. We + * disable GSO and PACKED feature in the feature mask. They may be enabled + * explicitly via enable_gso and enable_packed argument + */ + disabled_features |= FEATURE_VIRTIO_NET_F_HOST_GUEST_TSO_FEATURE_BITS | + VIRTIO_FEATURE (VIRTIO_F_RING_PACKED); + + /* EVENT_IDX is disabled by default */ + disabled_features |= VIRTIO_FEATURE (VIRTIO_RING_F_EVENT_IDX); + args.feature_mask &= ~disabled_features; + + args.sw_if_index = ntohl (mp->sw_if_index); + args.sock_filename = (char *) mp->sock_filename; + args.is_server = mp->is_server; + args.renumber = mp->renumber; + args.custom_dev_instance = ntohl (mp->custom_dev_instance); + args.enable_gso = mp->enable_gso; + args.enable_packed = mp->enable_packed; + args.enable_event_idx = mp->enable_event_idx; + rv = vhost_user_modify_if (vnm, vm, &args); + + REPLY_MACRO (VL_API_MODIFY_VHOST_USER_IF_V2_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; + u32 sw_if_index = ntohl (mp->sw_if_index); + vl_api_registration_t *reg; + + 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) + { + reg = vl_api_client_index_to_registration (mp->client_index); + if (!reg) + return; + + vnet_clear_sw_interface_tag (vnm, sw_if_index); + } +} + +static void +vhost_user_features_encode (u64 features, u32 *first, u32 *last) +{ + *first = clib_net_to_host_u32 (features); + *last = clib_net_to_host_u32 (features >> 32); +} + +static void +send_sw_interface_vhost_user_details (vpe_api_main_t * am, + vl_api_registration_t * reg, + vhost_user_intf_details_t * vui, + u32 context) +{ + vl_api_sw_interface_vhost_user_details_t *mp; + + mp = vl_msg_api_alloc (sizeof (*mp)); + clib_memset (mp, 0, sizeof (*mp)); + mp->_vl_msg_id = + ntohs (REPLY_MSG_ID_BASE + 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); + vhost_user_features_encode (vui->features, (u32 *) &mp->features_first_32, + (u32 *) &mp->features_last_32); + 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_api_send_msg (reg, (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; + vl_api_registration_t *reg; + u32 filter_sw_if_index; + + reg = vl_api_client_index_to_registration (mp->client_index); + if (!reg) + return; + + filter_sw_if_index = htonl (mp->sw_if_index); + if (filter_sw_if_index != ~0) + VALIDATE_SW_IF_INDEX (mp); + + rv = vhost_user_dump_ifs (vnm, vm, &ifaces); + if (rv) + return; + + vec_foreach (vuid, ifaces) + { + if ((filter_sw_if_index == ~0) || + (vuid->sw_if_index == filter_sw_if_index)) + send_sw_interface_vhost_user_details (am, reg, vuid, mp->context); + } + BAD_SW_IF_INDEX_LABEL; + vec_free (ifaces); +} + +#include +static clib_error_t * +vhost_user_api_hookup (vlib_main_t * vm) +{ + api_main_t *am = vlibapi_get_main (); + /* Mark CREATE_VHOST_USER_IF as mp safe */ + vl_api_set_msg_thread_safe (am, VL_API_CREATE_VHOST_USER_IF, 1); + vl_api_set_msg_thread_safe (am, VL_API_CREATE_VHOST_USER_IF_V2, 1); + + /* + * Set up the (msg_name, crc, message-id) table + */ + REPLY_MSG_ID_BASE = setup_message_id_table (); + + 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: + */ diff --git a/src/plugins/vhost/vhost_user_inline.h b/src/plugins/vhost/vhost_user_inline.h new file mode 100644 index 00000000000..8bdff3733a7 --- /dev/null +++ b/src/plugins/vhost/vhost_user_inline.h @@ -0,0 +1,496 @@ +/* + * Copyright (c) 2018 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_INLINE_H__ +#define __VIRTIO_VHOST_USER_INLINE_H__ +/* vhost-user inline functions */ +#include + +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 = count_trailing_zeros (_mm_movemask_epi8 (r) | + (1 << VHOST_MEMORY_MAX_NREGIONS)); + + if (i < vui->nregions) + { + *hint = i; + return (void *) (vui->region_mmap_addr[i] + addr - + vui->regions[i].guest_phys_addr); + } +#elif __aarch64__ && __ARM_NEON + uint64x2_t al, ah, rl, rh, r; + uint32_t u32 = 0; + + al = vdupq_n_u64 (addr + 1); + ah = vdupq_n_u64 (addr); + + /*First Iteration */ + rl = vld1q_u64 (&vui->region_guest_addr_lo[0]); + rl = vcgtq_u64 (al, rl); + rh = vld1q_u64 (&vui->region_guest_addr_hi[0]); + rh = vcgtq_u64 (rh, ah); + r = vandq_u64 (rl, rh); + u32 |= (vgetq_lane_u8 (vreinterpretq_u8_u64 (r), 0) & 0x1); + u32 |= ((vgetq_lane_u8 (vreinterpretq_u8_u64 (r), 8) & 0x1) << 1); + + if (u32) + { + i = count_trailing_zeros (u32); + goto vhost_map_guest_mem_done; + } + + /*Second Iteration */ + rl = vld1q_u64 (&vui->region_guest_addr_lo[2]); + rl = vcgtq_u64 (al, rl); + rh = vld1q_u64 (&vui->region_guest_addr_hi[2]); + rh = vcgtq_u64 (rh, ah); + r = vandq_u64 (rl, rh); + u32 |= ((vgetq_lane_u8 (vreinterpretq_u8_u64 (r), 0) & 0x1) << 2); + u32 |= ((vgetq_lane_u8 (vreinterpretq_u8_u64 (r), 8) & 0x1) << 3); + + if (u32) + { + i = count_trailing_zeros (u32); + goto vhost_map_guest_mem_done; + } + + /*Third Iteration */ + rl = vld1q_u64 (&vui->region_guest_addr_lo[4]); + rl = vcgtq_u64 (al, rl); + rh = vld1q_u64 (&vui->region_guest_addr_hi[4]); + rh = vcgtq_u64 (rh, ah); + r = vandq_u64 (rl, rh); + u32 |= ((vgetq_lane_u8 (vreinterpretq_u8_u64 (r), 0) & 0x1) << 6); + u32 |= ((vgetq_lane_u8 (vreinterpretq_u8_u64 (r), 8) & 0x1) << 7); + + i = count_trailing_zeros (u32 | (1 << VHOST_MEMORY_MAX_NREGIONS)); + +vhost_map_guest_mem_done: + 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 + /* *INDENT-OFF* */ + ELOG_TYPE_DECLARE (el) = + { + .format = "failed to map guest mem addr %lx", + .format_args = "i8", + }; + /* *INDENT-ON* */ + struct + { + uword addr; + } *ed; + ed = ELOG_DATA (&vlib_global_main.elog_main, el); + ed->addr = addr; + *hint = 0; + return 0; +} + +static_always_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; +} + +#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 & VIRTIO_FEATURE (VHOST_F_LOG_ALL)))) + { + return; + } + if (is_host_address) + { + addr = pointer_to_uword (map_user_mem (vui, (uword) addr)); + } + if (PREDICT_FALSE ((addr + len - 1) / VHOST_LOG_PAGE / 8 >= vui->log_size)) + { + vu_log_debug (vui, "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++; + } +} + +#define vhost_user_log_dirty_ring(vui, vq, member) \ + if (PREDICT_FALSE (vq->log_used)) \ + { \ + vhost_user_log_dirty_pages_2 ( \ + vui, \ + vq->log_guest_addr + \ + STRUCT_OFFSET_OF (vnet_virtio_vring_used_t, member), \ + sizeof (vq->used->member), 0); \ + } + +static_always_inline 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 = vum->vhost_user_interfaces + t->device_index; + vnet_sw_interface_t *sw; + u32 indent; + + if (pool_is_free (vum->vhost_user_interfaces, vui)) + { + s = format (s, "vhost-user interface is deleted"); + return s; + } + sw = vnet_get_sw_interface (vnm, vui->sw_if_index); + 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; +} + +static_always_inline u64 +vhost_user_is_packed_ring_supported (vhost_user_intf_t * vui) +{ + return (vui->features & VIRTIO_FEATURE (VIRTIO_F_RING_PACKED)); +} + +static_always_inline u64 +vhost_user_is_event_idx_supported (vhost_user_intf_t * vui) +{ + return (vui->features & VIRTIO_FEATURE (VIRTIO_RING_F_EVENT_IDX)); +} + +static_always_inline void +vhost_user_kick (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; + + rv = write (fd, &x, sizeof (x)); + if (PREDICT_FALSE (rv <= 0)) + { + clib_unix_warning + ("Error: Could not write to unix socket for callfd %d", fd); + return; + } + + vq->n_since_last_int = 0; + vq->int_deadline = vlib_time_now (vm) + vum->coalesce_time; +} + +static_always_inline u16 +vhost_user_avail_event_idx (vhost_user_vring_t * vq) +{ + volatile u16 *event_idx = (u16 *) & (vq->used->ring[vq->qsz_mask + 1]); + + return *event_idx; +} + +static_always_inline u16 +vhost_user_used_event_idx (vhost_user_vring_t * vq) +{ + volatile u16 *event_idx = (u16 *) & (vq->avail->ring[vq->qsz_mask + 1]); + + return *event_idx; +} + +static_always_inline u16 +vhost_user_need_event (u16 event_idx, u16 new_idx, u16 old_idx) +{ + return ((u16) (new_idx - event_idx - 1) < (u16) (new_idx - old_idx)); +} + +static_always_inline void +vhost_user_send_call_event_idx (vlib_main_t * vm, vhost_user_vring_t * vq) +{ + vhost_user_main_t *vum = &vhost_user_main; + u8 first_kick = vq->first_kick; + u16 event_idx = vhost_user_used_event_idx (vq); + + vq->first_kick = 1; + if (vhost_user_need_event (event_idx, vq->last_used_idx, vq->last_kick) || + PREDICT_FALSE (!first_kick)) + { + vhost_user_kick (vm, vq); + vq->last_kick = event_idx; + } + else + { + vq->n_since_last_int = 0; + vq->int_deadline = vlib_time_now (vm) + vum->coalesce_time; + } +} + +static_always_inline void +vhost_user_send_call_event_idx_packed (vlib_main_t * vm, + vhost_user_vring_t * vq) +{ + vhost_user_main_t *vum = &vhost_user_main; + u8 first_kick = vq->first_kick; + u16 off_wrap; + u16 event_idx; + u16 new_idx = vq->last_used_idx; + u16 old_idx = vq->last_kick; + + if (PREDICT_TRUE (vq->avail_event->flags == VRING_EVENT_F_DESC)) + { + CLIB_COMPILER_BARRIER (); + off_wrap = vq->avail_event->off_wrap; + event_idx = off_wrap & 0x7fff; + if (vq->used_wrap_counter != (off_wrap >> 15)) + event_idx -= (vq->qsz_mask + 1); + + if (new_idx <= old_idx) + old_idx -= (vq->qsz_mask + 1); + + vq->first_kick = 1; + vq->last_kick = event_idx; + if (vhost_user_need_event (event_idx, new_idx, old_idx) || + PREDICT_FALSE (!first_kick)) + vhost_user_kick (vm, vq); + else + { + vq->n_since_last_int = 0; + vq->int_deadline = vlib_time_now (vm) + vum->coalesce_time; + } + } + else + vhost_user_kick (vm, vq); +} + +static_always_inline void +vhost_user_send_call (vlib_main_t * vm, vhost_user_intf_t * vui, + vhost_user_vring_t * vq) +{ + if (vhost_user_is_event_idx_supported (vui)) + { + if (vhost_user_is_packed_ring_supported (vui)) + vhost_user_send_call_event_idx_packed (vm, vq); + else + vhost_user_send_call_event_idx (vm, vq); + } + else + vhost_user_kick (vm, vq); +} + +static_always_inline u8 +vui_is_link_up (vhost_user_intf_t * vui) +{ + return vui->admin_up && vui->is_ready; +} + +static_always_inline void +vhost_user_update_gso_interface_count (vhost_user_intf_t * vui, u8 add) +{ + vhost_user_main_t *vum = &vhost_user_main; + + if (vui->enable_gso) + { + if (add) + { + vum->gso_count++; + } + else + { + ASSERT (vum->gso_count > 0); + vum->gso_count--; + } + } +} + +static_always_inline u8 +vhost_user_packed_desc_available (vhost_user_vring_t * vring, u16 idx) +{ + return (((vring->packed_desc[idx].flags & VRING_DESC_F_AVAIL) == + vring->avail_wrap_counter)); +} + +static_always_inline void +vhost_user_advance_last_avail_idx (vhost_user_vring_t * vring) +{ + vring->last_avail_idx++; + if (PREDICT_FALSE ((vring->last_avail_idx & vring->qsz_mask) == 0)) + { + vring->avail_wrap_counter ^= VRING_DESC_F_AVAIL; + vring->last_avail_idx = 0; + } +} + +static_always_inline void +vhost_user_advance_last_avail_table_idx (vhost_user_intf_t * vui, + vhost_user_vring_t * vring, + u8 chained) +{ + if (chained) + { + vnet_virtio_vring_packed_desc_t *desc_table = vring->packed_desc; + + /* pick up the slot of the next avail idx */ + while (desc_table[vring->last_avail_idx & vring->qsz_mask].flags & + VRING_DESC_F_NEXT) + vhost_user_advance_last_avail_idx (vring); + } + + vhost_user_advance_last_avail_idx (vring); +} + +static_always_inline void +vhost_user_undo_advanced_last_avail_idx (vhost_user_vring_t * vring) +{ + if (PREDICT_FALSE ((vring->last_avail_idx & vring->qsz_mask) == 0)) + vring->avail_wrap_counter ^= VRING_DESC_F_AVAIL; + + if (PREDICT_FALSE (vring->last_avail_idx == 0)) + vring->last_avail_idx = vring->qsz_mask; + else + vring->last_avail_idx--; +} + +static_always_inline void +vhost_user_dequeue_descs (vhost_user_vring_t *rxvq, + vnet_virtio_net_hdr_mrg_rxbuf_t *hdr, + u16 *n_descs_processed) +{ + u16 i; + + *n_descs_processed -= (hdr->num_buffers - 1); + for (i = 0; i < hdr->num_buffers - 1; i++) + vhost_user_undo_advanced_last_avail_idx (rxvq); +} + +static_always_inline void +vhost_user_dequeue_chained_descs (vhost_user_vring_t * rxvq, + u16 * n_descs_processed) +{ + while (*n_descs_processed) + { + vhost_user_undo_advanced_last_avail_idx (rxvq); + (*n_descs_processed)--; + } +} + +static_always_inline void +vhost_user_advance_last_used_idx (vhost_user_vring_t * vring) +{ + vring->last_used_idx++; + if (PREDICT_FALSE ((vring->last_used_idx & vring->qsz_mask) == 0)) + { + vring->used_wrap_counter ^= 1; + vring->last_used_idx = 0; + } +} + +#endif + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/plugins/vhost/vhost_user_input.c b/src/plugins/vhost/vhost_user_input.c new file mode 100644 index 00000000000..c083f43ba7c --- /dev/null +++ b/src/plugins/vhost/vhost_user_input.c @@ -0,0 +1,1474 @@ +/* + *------------------------------------------------------------------ + * vhost-user-input + * + * Copyright (c) 2014-2018 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 +#include + +#include +#include + +#include +#include + +/* + * 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 + +extern vlib_node_registration_t vhost_user_input_node; + +#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)") \ + _(NOT_READY, "vhost interface not ready or down") \ + _(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 __clib_unused char *vhost_user_input_func_error_strings[] = { +#define _(n,s) s, + foreach_vhost_user_input_func_error +#undef _ +}; + +static_always_inline 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, + u16 last_avail_idx) +{ + vhost_user_main_t *vum = &vhost_user_main; + u32 desc_current = txvq->avail->ring[last_avail_idx & txvq->qsz_mask]; + vnet_virtio_vring_desc_t *hdr_desc = 0; + vnet_virtio_net_hdr_mrg_rxbuf_t *hdr; + u32 hint = 0; + + clib_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 & VRING_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 & VRING_DESC_F_NEXT) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SIMPLE_CHAINED; + } + if (!(txvq->desc[desc_current].flags & VRING_DESC_F_NEXT) && + !(txvq->desc[desc_current].flags & VRING_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_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_load (src2); + clib_prefetch_load (src3); + + clib_memcpy_fast ((void *) cpy[0].dst, src0, cpy[0].len); + clib_memcpy_fast ((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_fast ((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. + */ +static_always_inline 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 mask = txvq->qsz_mask; + u16 last_avail_idx = txvq->last_avail_idx; + u16 last_used_idx = txvq->last_used_idx; + while (discarded_packets != discard_max) + { + if (avail_idx == last_avail_idx) + goto out; + + u16 desc_chain_head = txvq->avail->ring[last_avail_idx & mask]; + last_avail_idx++; + txvq->used->ring[last_used_idx & mask].id = desc_chain_head; + txvq->used->ring[last_used_idx & mask].len = 0; + vhost_user_log_dirty_ring (vui, txvq, ring[last_used_idx & mask]); + last_used_idx++; + discarded_packets++; + } + +out: + txvq->last_avail_idx = last_avail_idx; + txvq->last_used_idx = last_used_idx; + CLIB_MEMORY_STORE_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_always_inline 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; + } + cpu->rx_buffers_len++; +} + +static_always_inline void +vhost_user_handle_rx_offload (vlib_buffer_t *b0, u8 *b0_data, + vnet_virtio_net_hdr_t *hdr) +{ + u8 l4_hdr_sz = 0; + u8 l4_proto = 0; + ethernet_header_t *eh = (ethernet_header_t *) b0_data; + u16 ethertype = clib_net_to_host_u16 (eh->type); + u16 l2hdr_sz = sizeof (ethernet_header_t); + vnet_buffer_oflags_t oflags = 0; + + if (ethernet_frame_is_tagged (ethertype)) + { + ethernet_vlan_header_t *vlan = (ethernet_vlan_header_t *) (eh + 1); + + ethertype = clib_net_to_host_u16 (vlan->type); + l2hdr_sz += sizeof (*vlan); + if (ethertype == ETHERNET_TYPE_VLAN) + { + vlan++; + ethertype = clib_net_to_host_u16 (vlan->type); + l2hdr_sz += sizeof (*vlan); + } + } + vnet_buffer (b0)->l2_hdr_offset = 0; + vnet_buffer (b0)->l3_hdr_offset = l2hdr_sz; + vnet_buffer (b0)->l4_hdr_offset = hdr->csum_start; + b0->flags |= (VNET_BUFFER_F_L2_HDR_OFFSET_VALID | + VNET_BUFFER_F_L3_HDR_OFFSET_VALID | + VNET_BUFFER_F_L4_HDR_OFFSET_VALID); + + if (PREDICT_TRUE (ethertype == ETHERNET_TYPE_IP4)) + { + ip4_header_t *ip4 = (ip4_header_t *) (b0_data + l2hdr_sz); + l4_proto = ip4->protocol; + b0->flags |= VNET_BUFFER_F_IS_IP4; + oflags |= VNET_BUFFER_OFFLOAD_F_IP_CKSUM; + } + else if (PREDICT_TRUE (ethertype == ETHERNET_TYPE_IP6)) + { + ip6_header_t *ip6 = (ip6_header_t *) (b0_data + l2hdr_sz); + l4_proto = ip6->protocol; + b0->flags |= VNET_BUFFER_F_IS_IP6; + } + + if (l4_proto == IP_PROTOCOL_TCP) + { + tcp_header_t *tcp = (tcp_header_t *) + (b0_data + vnet_buffer (b0)->l4_hdr_offset); + l4_hdr_sz = tcp_header_bytes (tcp); + oflags |= VNET_BUFFER_OFFLOAD_F_TCP_CKSUM; + } + else if (l4_proto == IP_PROTOCOL_UDP) + { + l4_hdr_sz = sizeof (udp_header_t); + oflags |= VNET_BUFFER_OFFLOAD_F_UDP_CKSUM; + } + + if (hdr->gso_type == VIRTIO_NET_HDR_GSO_UDP) + { + vnet_buffer2 (b0)->gso_size = hdr->gso_size; + vnet_buffer2 (b0)->gso_l4_hdr_sz = l4_hdr_sz; + b0->flags |= VNET_BUFFER_F_GSO; + } + else if (hdr->gso_type == VIRTIO_NET_HDR_GSO_TCPV4) + { + vnet_buffer2 (b0)->gso_size = hdr->gso_size; + vnet_buffer2 (b0)->gso_l4_hdr_sz = l4_hdr_sz; + b0->flags |= (VNET_BUFFER_F_GSO | VNET_BUFFER_F_IS_IP4); + } + else if (hdr->gso_type == VIRTIO_NET_HDR_GSO_TCPV6) + { + vnet_buffer2 (b0)->gso_size = hdr->gso_size; + vnet_buffer2 (b0)->gso_l4_hdr_sz = l4_hdr_sz; + b0->flags |= (VNET_BUFFER_F_GSO | VNET_BUFFER_F_IS_IP6); + } + + if (oflags) + vnet_buffer_offload_flags_set (b0, oflags); +} + +static_always_inline void +vhost_user_input_do_interrupt (vlib_main_t * vm, vhost_user_intf_t * vui, + vhost_user_vring_t * txvq, + vhost_user_vring_t * rxvq) +{ + f64 now = vlib_time_now (vm); + + if ((txvq->n_since_last_int) && (txvq->int_deadline < now)) + vhost_user_send_call (vm, vui, txvq); + + if ((rxvq->n_since_last_int) && (rxvq->int_deadline < now)) + vhost_user_send_call (vm, vui, rxvq); +} + +static_always_inline void +vhost_user_input_setup_frame (vlib_main_t * vm, vlib_node_runtime_t * node, + vhost_user_intf_t * vui, + u32 * current_config_index, u32 * next_index, + u32 ** to_next, u32 * n_left_to_next) +{ + vnet_feature_main_t *fm = &feature_main; + u8 feature_arc_idx = fm->device_input_feature_arc_index; + + if (PREDICT_FALSE (vnet_have_features (feature_arc_idx, vui->sw_if_index))) + { + vnet_feature_config_main_t *cm; + cm = &fm->feature_config_mains[feature_arc_idx]; + *current_config_index = vec_elt (cm->config_index_by_sw_if_index, + vui->sw_if_index); + vnet_get_config_data (&cm->config_main, current_config_index, + next_index, 0); + } + + vlib_get_new_next_frame (vm, node, *next_index, *to_next, *n_left_to_next); + + if (*next_index == VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT) + { + /* give some hints to ethernet-input */ + vlib_next_frame_t *nf; + vlib_frame_t *f; + ethernet_input_frame_t *ef; + nf = vlib_node_runtime_get_next_frame (vm, node, *next_index); + f = vlib_get_frame (vm, nf->frame); + f->flags = ETH_INPUT_FRAME_F_SINGLE_SW_IF_IDX; + + ef = vlib_frame_scalar_args (f); + ef->sw_if_index = vui->sw_if_index; + ef->hw_if_index = vui->hw_if_index; + vlib_frame_no_append (f); + } +} + +static_always_inline 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, u8 enable_csum) +{ + vhost_user_vring_t *txvq = &vui->vrings[VHOST_VRING_IDX_TX (qid)]; + vnet_feature_main_t *fm = &feature_main; + 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); + u32 buffer_data_size = vlib_buffer_get_default_data_size (vm); + u32 map_hint = 0; + vhost_cpu_t *cpu = &vum->cpus[vm->thread_index]; + u16 copy_len = 0; + u8 feature_arc_idx = fm->device_input_feature_arc_index; + u32 current_config_index = ~(u32) 0; + u16 mask = txvq->qsz_mask; + + /* The descriptor table is not ready yet */ + if (PREDICT_FALSE (txvq->avail == 0)) + goto done; + + { + /* do we have pending interrupts ? */ + vhost_user_vring_t *rxvq = &vui->vrings[VHOST_VRING_IDX_RX (qid)]; + vhost_user_input_do_interrupt (vm, vui, txvq, rxvq); + } + + /* + * For adaptive mode, it is optimized to reduce interrupts. + * If the scheduler switches the input node to polling due + * to burst of traffic, we tell the driver no interrupt. + * When the traffic subsides, the scheduler switches the node back to + * interrupt mode. We must tell the driver we want interrupt. + */ + if (PREDICT_FALSE (txvq->mode == VNET_HW_IF_RX_MODE_ADAPTIVE)) + { + if ((node->flags & + VLIB_NODE_FLAG_SWITCH_FROM_POLLING_TO_INTERRUPT_MODE) || + !(node->flags & + VLIB_NODE_FLAG_SWITCH_FROM_INTERRUPT_TO_POLLING_MODE)) + /* Tell driver we want notification */ + txvq->used->flags = 0; + else + /* Tell driver we don't want notification */ + txvq->used->flags = VRING_USED_F_NO_NOTIFY; + } + + if (PREDICT_FALSE (txvq->avail->flags & 0xFFFE)) + goto done; + + n_left = (u16) (txvq->avail->idx - txvq->last_avail_idx); + + /* nothing to do */ + if (PREDICT_FALSE (n_left == 0)) + goto done; + + 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); + goto done; + } + + if (PREDICT_FALSE (n_left == (mask + 1))) + { + /* + * 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); + } + + 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 yield 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. + * The assumption is that big packets will fit in 40 buffers. + */ + if (PREDICT_FALSE (cpu->rx_buffers_len < n_left + 1 || + cpu->rx_buffers_len < 40)) + { + u32 curr_len = cpu->rx_buffers_len; + cpu->rx_buffers_len += + vlib_buffer_alloc (vm, cpu->rx_buffers + curr_len, + VHOST_USER_RX_BUFFERS_N - curr_len); + + if (PREDICT_FALSE + (cpu->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 > cpu->rx_buffers_len) ? + n_left + 1 - cpu->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, + vm->thread_index, vui->sw_if_index, + flush); + + vlib_error_count (vm, vhost_user_input_node.index, + VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER, flush); + } + } + + vhost_user_input_setup_frame (vm, node, vui, ¤t_config_index, + &next_index, &to_next, &n_left_to_next); + + u16 last_avail_idx = txvq->last_avail_idx; + u16 last_used_idx = txvq->last_used_idx; + + while (n_left > 0) + { + vlib_buffer_t *b_head, *b_current; + u32 bi_current; + u16 desc_current; + u32 desc_data_offset; + vnet_virtio_vring_desc_t *desc_table = txvq->desc; + + if (PREDICT_FALSE (cpu->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[last_avail_idx & mask]; + cpu->rx_buffers_len--; + bi_current = cpu->rx_buffers[cpu->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, cpu->rx_buffers[cpu->rx_buffers_len - 1], LOAD); + + /* Just preset the used descriptor id and length for later */ + txvq->used->ring[last_used_idx & mask].id = desc_current; + txvq->used->ring[last_used_idx & mask].len = 0; + vhost_user_log_dirty_ring (vui, txvq, ring[last_used_idx & 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 > 0 && 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, last_avail_idx); + 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 & VRING_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)) + { + vlib_error_count (vm, node->node_index, + VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1); + goto out; + } + } + + desc_data_offset = vui->virtio_net_hdr_sz; + + if (enable_csum) + { + vnet_virtio_net_hdr_mrg_rxbuf_t *hdr; + u8 *b_data; + u16 current; + + hdr = map_guest_mem (vui, desc_table[desc_current].addr, &map_hint); + if (PREDICT_FALSE (hdr == 0)) + { + vlib_error_count (vm, node->node_index, + VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1); + goto out; + } + if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) + { + if ((desc_data_offset == desc_table[desc_current].len) && + (desc_table[desc_current].flags & VRING_DESC_F_NEXT)) + { + current = desc_table[desc_current].next; + b_data = map_guest_mem (vui, desc_table[current].addr, + &map_hint); + if (PREDICT_FALSE (b_data == 0)) + { + vlib_error_count (vm, node->node_index, + VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, + 1); + goto out; + } + } + else + b_data = (u8 *) hdr + desc_data_offset; + + vhost_user_handle_rx_offload (b_head, b_data, &hdr->hdr); + } + } + + 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 & + VRING_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 == buffer_data_size)) + { + if (PREDICT_FALSE (cpu->rx_buffers_len == 0)) + { + /* Cancel speculation */ + to_next--; + n_left_to_next++; + + /* + * 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, cpu, b_head); + n_left = 0; + goto stop; + } + + /* Get next output */ + cpu->rx_buffers_len--; + u32 bi_next = cpu->rx_buffers[cpu->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 */ + ASSERT (copy_len < VHOST_USER_COPY_ARRAY_N); + vhost_copy_t *cpy = &cpu->copy[copy_len]; + copy_len++; + u32 desc_data_l = desc_table[desc_current].len - desc_data_offset; + cpy->len = 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) + + b_current->current_length); + 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: + + 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 */ + last_avail_idx++; + last_used_idx++; + + 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; + + if (current_config_index != ~(u32) 0) + { + b_head->current_config_index = current_config_index; + vnet_buffer (b_head)->feature_arc_index = feature_arc_idx; + } + + 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, cpu->copy, + copy_len, &map_hint))) + { + vlib_error_count (vm, node->node_index, + VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1); + } + copy_len = 0; + + /* give buffers back to driver */ + CLIB_MEMORY_STORE_BARRIER (); + txvq->used->idx = last_used_idx; + vhost_user_log_dirty_ring (vui, txvq, idx); + } + } +stop: + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + + txvq->last_used_idx = last_used_idx; + txvq->last_avail_idx = last_avail_idx; + + /* Do the memory copies */ + if (PREDICT_FALSE (vhost_user_input_copy (vui, cpu->copy, copy_len, + &map_hint))) + { + vlib_error_count (vm, node->node_index, + VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1); + } + + /* give buffers back to driver */ + CLIB_MEMORY_STORE_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 & VRING_AVAIL_F_NO_INTERRUPT)) + { + txvq->n_since_last_int += n_rx_packets; + + if (txvq->n_since_last_int > vum->coalesce_frames) + vhost_user_send_call (vm, vui, txvq); + } + + /* increase rx counters */ + vlib_increment_combined_counter + (vnet_main.interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, vm->thread_index, vui->sw_if_index, + n_rx_packets, n_rx_bytes); + + vnet_device_increment_rx_packets (vm->thread_index, n_rx_packets); + +done: + return n_rx_packets; +} + +static_always_inline void +vhost_user_mark_desc_consumed (vhost_user_intf_t * vui, + vhost_user_vring_t * txvq, u16 desc_head, + u16 n_descs_processed) +{ + vnet_virtio_vring_packed_desc_t *desc_table = txvq->packed_desc; + u16 desc_idx; + u16 mask = txvq->qsz_mask; + + for (desc_idx = 0; desc_idx < n_descs_processed; desc_idx++) + { + if (txvq->used_wrap_counter) + desc_table[(desc_head + desc_idx) & mask].flags |= + (VRING_DESC_F_AVAIL | VRING_DESC_F_USED); + else + desc_table[(desc_head + desc_idx) & mask].flags &= + ~(VRING_DESC_F_AVAIL | VRING_DESC_F_USED); + vhost_user_advance_last_used_idx (txvq); + } +} + +static_always_inline void +vhost_user_rx_trace_packed (vhost_trace_t * t, vhost_user_intf_t * vui, + u16 qid, vhost_user_vring_t * txvq, + u16 desc_current) +{ + vhost_user_main_t *vum = &vhost_user_main; + vnet_virtio_vring_packed_desc_t *hdr_desc; + vnet_virtio_net_hdr_mrg_rxbuf_t *hdr; + u32 hint = 0; + + clib_memset (t, 0, sizeof (*t)); + t->device_index = vui - vum->vhost_user_interfaces; + t->qid = qid; + + hdr_desc = &txvq->packed_desc[desc_current]; + if (txvq->packed_desc[desc_current].flags & VRING_DESC_F_INDIRECT) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_INDIRECT; + /* Header is the first here */ + hdr_desc = map_guest_mem (vui, txvq->packed_desc[desc_current].addr, + &hint); + } + if (txvq->packed_desc[desc_current].flags & VRING_DESC_F_NEXT) + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SIMPLE_CHAINED; + + if (!(txvq->packed_desc[desc_current].flags & VRING_DESC_F_NEXT) && + !(txvq->packed_desc[desc_current].flags & VRING_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; + clib_memcpy_fast (&t->hdr, hdr, + len > hdr_desc->len ? hdr_desc->len : len); + } +} + +static_always_inline u32 +vhost_user_rx_discard_packet_packed (vlib_main_t * vm, + vhost_user_intf_t * vui, + vhost_user_vring_t * txvq, + u32 discard_max) +{ + u32 discarded_packets = 0; + u16 mask = txvq->qsz_mask; + u16 desc_current, desc_head; + + desc_head = desc_current = txvq->last_used_idx & mask; + + /* + * 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. + */ + while ((discarded_packets != discard_max) && + vhost_user_packed_desc_available (txvq, desc_current)) + { + vhost_user_advance_last_avail_idx (txvq); + discarded_packets++; + desc_current = (desc_current + 1) & mask; + } + + if (PREDICT_TRUE (discarded_packets)) + vhost_user_mark_desc_consumed (vui, txvq, desc_head, discarded_packets); + return (discarded_packets); +} + +static_always_inline u32 +vhost_user_input_copy_packed (vhost_user_intf_t * vui, vhost_copy_t * cpy, + u16 copy_len, u32 * map_hint) +{ + void *src0, *src1, *src2, *src3, *src4, *src5, *src6, *src7; + u8 bad; + u32 rc = VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR; + + if (PREDICT_TRUE (copy_len >= 8)) + { + src4 = map_guest_mem (vui, cpy[0].src, map_hint); + src5 = map_guest_mem (vui, cpy[1].src, map_hint); + src6 = map_guest_mem (vui, cpy[2].src, map_hint); + src7 = map_guest_mem (vui, cpy[3].src, map_hint); + bad = (src4 == 0) + (src5 == 0) + (src6 == 0) + (src7 == 0); + if (PREDICT_FALSE (bad)) + goto one_by_one; + clib_prefetch_load (src4); + clib_prefetch_load (src5); + clib_prefetch_load (src6); + clib_prefetch_load (src7); + + while (PREDICT_TRUE (copy_len >= 8)) + { + src0 = src4; + src1 = src5; + src2 = src6; + src3 = src7; + + src4 = map_guest_mem (vui, cpy[4].src, map_hint); + src5 = map_guest_mem (vui, cpy[5].src, map_hint); + src6 = map_guest_mem (vui, cpy[6].src, map_hint); + src7 = map_guest_mem (vui, cpy[7].src, map_hint); + bad = (src4 == 0) + (src5 == 0) + (src6 == 0) + (src7 == 0); + if (PREDICT_FALSE (bad)) + break; + + clib_prefetch_load (src4); + clib_prefetch_load (src5); + clib_prefetch_load (src6); + clib_prefetch_load (src7); + + clib_memcpy_fast ((void *) cpy[0].dst, src0, cpy[0].len); + clib_memcpy_fast ((void *) cpy[1].dst, src1, cpy[1].len); + clib_memcpy_fast ((void *) cpy[2].dst, src2, cpy[2].len); + clib_memcpy_fast ((void *) cpy[3].dst, src3, cpy[3].len); + copy_len -= 4; + cpy += 4; + } + } + +one_by_one: + while (copy_len) + { + if (PREDICT_FALSE (!(src0 = map_guest_mem (vui, cpy->src, map_hint)))) + { + rc = VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL; + break; + } + clib_memcpy_fast ((void *) cpy->dst, src0, cpy->len); + copy_len -= 1; + cpy += 1; + } + return rc; +} + +static_always_inline u32 +vhost_user_do_offload (vhost_user_intf_t *vui, + vnet_virtio_vring_packed_desc_t *desc_table, + u16 desc_current, u16 mask, vlib_buffer_t *b_head, + u32 *map_hint) +{ + u32 rc = VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR; + vnet_virtio_net_hdr_mrg_rxbuf_t *hdr; + u8 *b_data; + u32 desc_data_offset = vui->virtio_net_hdr_sz; + + hdr = map_guest_mem (vui, desc_table[desc_current].addr, map_hint); + if (PREDICT_FALSE (hdr == 0)) + rc = VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL; + else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) + { + if (desc_data_offset == desc_table[desc_current].len) + { + desc_current = (desc_current + 1) & mask; + b_data = + map_guest_mem (vui, desc_table[desc_current].addr, map_hint); + if (PREDICT_FALSE (b_data == 0)) + rc = VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL; + else + vhost_user_handle_rx_offload (b_head, b_data, &hdr->hdr); + } + else + { + b_data = (u8 *) hdr + desc_data_offset; + vhost_user_handle_rx_offload (b_head, b_data, &hdr->hdr); + } + } + + return rc; +} + +static_always_inline u32 +vhost_user_compute_buffers_required (u32 desc_len, u32 buffer_data_size) +{ + div_t result; + u32 buffers_required; + + if (PREDICT_TRUE (buffer_data_size == 2048)) + { + buffers_required = desc_len >> 11; + if ((desc_len & 2047) != 0) + buffers_required++; + return (buffers_required); + } + + if (desc_len < buffer_data_size) + return 1; + + result = div (desc_len, buffer_data_size); + if (result.rem) + buffers_required = result.quot + 1; + else + buffers_required = result.quot; + + return (buffers_required); +} + +static_always_inline u32 +vhost_user_compute_indirect_desc_len (vhost_user_intf_t * vui, + vhost_user_vring_t * txvq, + u32 buffer_data_size, u16 desc_current, + u32 * map_hint) +{ + vnet_virtio_vring_packed_desc_t *desc_table = txvq->packed_desc; + u32 desc_len = 0; + u16 desc_data_offset = vui->virtio_net_hdr_sz; + u16 desc_idx = desc_current; + u32 n_descs; + + n_descs = desc_table[desc_idx].len >> 4; + desc_table = map_guest_mem (vui, desc_table[desc_idx].addr, map_hint); + if (PREDICT_FALSE (desc_table == 0)) + return 0; + + for (desc_idx = 0; desc_idx < n_descs; desc_idx++) + desc_len += desc_table[desc_idx].len; + + if (PREDICT_TRUE (desc_len > desc_data_offset)) + desc_len -= desc_data_offset; + + return vhost_user_compute_buffers_required (desc_len, buffer_data_size); +} + +static_always_inline u32 +vhost_user_compute_chained_desc_len (vhost_user_intf_t * vui, + vhost_user_vring_t * txvq, + u32 buffer_data_size, u16 * current, + u16 * n_left) +{ + vnet_virtio_vring_packed_desc_t *desc_table = txvq->packed_desc; + u32 desc_len = 0; + u16 mask = txvq->qsz_mask; + + while (desc_table[*current].flags & VRING_DESC_F_NEXT) + { + desc_len += desc_table[*current].len; + (*n_left)++; + *current = (*current + 1) & mask; + vhost_user_advance_last_avail_idx (txvq); + } + desc_len += desc_table[*current].len; + (*n_left)++; + *current = (*current + 1) & mask; + vhost_user_advance_last_avail_idx (txvq); + + if (PREDICT_TRUE (desc_len > vui->virtio_net_hdr_sz)) + desc_len -= vui->virtio_net_hdr_sz; + + return vhost_user_compute_buffers_required (desc_len, buffer_data_size); +} + +static_always_inline void +vhost_user_assemble_packet (vnet_virtio_vring_packed_desc_t *desc_table, + u16 *desc_idx, vlib_buffer_t *b_head, + vlib_buffer_t **b_current, u32 **next, + vlib_buffer_t ***b, u32 *bi_current, + vhost_cpu_t *cpu, u16 *copy_len, u32 *buffers_used, + u32 buffers_required, u32 *desc_data_offset, + u32 buffer_data_size, u16 mask) +{ + u32 desc_data_l; + + while (*desc_data_offset < desc_table[*desc_idx].len) + { + /* Get more output if necessary. Or end of packet. */ + if (PREDICT_FALSE ((*b_current)->current_length == buffer_data_size)) + { + /* Get next output */ + u32 bi_next = **next; + (*next)++; + (*b_current)->next_buffer = bi_next; + (*b_current)->flags |= VLIB_BUFFER_NEXT_PRESENT; + *bi_current = bi_next; + *b_current = **b; + (*b)++; + (*buffers_used)++; + ASSERT (*buffers_used <= buffers_required); + } + + /* Prepare a copy order executed later for the data */ + ASSERT (*copy_len < VHOST_USER_COPY_ARRAY_N); + vhost_copy_t *cpy = &cpu->copy[*copy_len]; + (*copy_len)++; + desc_data_l = desc_table[*desc_idx].len - *desc_data_offset; + cpy->len = 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) + + (*b_current)->current_length); + cpy->src = desc_table[*desc_idx].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; + } + *desc_idx = (*desc_idx + 1) & mask;; + *desc_data_offset = 0; +} + +static_always_inline u32 +vhost_user_if_input_packed (vlib_main_t *vm, vhost_user_main_t *vum, + vhost_user_intf_t *vui, u16 qid, + vlib_node_runtime_t *node, u8 enable_csum) +{ + vhost_user_vring_t *txvq = &vui->vrings[VHOST_VRING_IDX_TX (qid)]; + vnet_feature_main_t *fm = &feature_main; + u8 feature_arc_idx = fm->device_input_feature_arc_index; + u16 n_rx_packets = 0; + u32 n_rx_bytes = 0; + u16 n_left = 0; + u32 buffers_required = 0; + 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); + u32 buffer_data_size = vlib_buffer_get_default_data_size (vm); + u32 map_hint = 0; + vhost_cpu_t *cpu = &vum->cpus[vm->thread_index]; + u16 copy_len = 0; + u32 current_config_index = ~0; + u16 mask = txvq->qsz_mask; + u16 desc_current, desc_head, last_used_idx; + vnet_virtio_vring_packed_desc_t *desc_table = 0; + u32 n_descs_processed = 0; + u32 rv; + vlib_buffer_t **b; + u32 *next; + u32 buffers_used = 0; + u16 current, n_descs_to_process; + + /* The descriptor table is not ready yet */ + if (PREDICT_FALSE (txvq->packed_desc == 0)) + goto done; + + /* do we have pending interrupts ? */ + vhost_user_vring_t *rxvq = &vui->vrings[VHOST_VRING_IDX_RX (qid)]; + vhost_user_input_do_interrupt (vm, vui, txvq, rxvq); + + /* + * For adaptive mode, it is optimized to reduce interrupts. + * If the scheduler switches the input node to polling due + * to burst of traffic, we tell the driver no interrupt. + * When the traffic subsides, the scheduler switches the node back to + * interrupt mode. We must tell the driver we want interrupt. + */ + if (PREDICT_FALSE (txvq->mode == VNET_HW_IF_RX_MODE_ADAPTIVE)) + { + if ((node->flags & + VLIB_NODE_FLAG_SWITCH_FROM_POLLING_TO_INTERRUPT_MODE) || + !(node->flags & + VLIB_NODE_FLAG_SWITCH_FROM_INTERRUPT_TO_POLLING_MODE)) + /* Tell driver we want notification */ + txvq->used_event->flags = 0; + else + /* Tell driver we don't want notification */ + txvq->used_event->flags = VRING_EVENT_F_DISABLE; + } + + last_used_idx = txvq->last_used_idx & mask; + desc_head = desc_current = last_used_idx; + + if (vhost_user_packed_desc_available (txvq, desc_current) == 0) + goto done; + + if (PREDICT_FALSE (!vui->admin_up || !vui->is_ready || !(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." + */ + rv = vhost_user_rx_discard_packet_packed (vm, vui, txvq, + VHOST_USER_DOWN_DISCARD_COUNT); + vlib_error_count (vm, vhost_user_input_node.index, + VHOST_USER_INPUT_FUNC_ERROR_NOT_READY, rv); + goto done; + } + + vhost_user_input_setup_frame (vm, node, vui, ¤t_config_index, + &next_index, &to_next, &n_left_to_next); + + /* + * Compute n_left and total buffers needed + */ + desc_table = txvq->packed_desc; + current = desc_current; + while (vhost_user_packed_desc_available (txvq, current) && + (n_left < VLIB_FRAME_SIZE)) + { + if (desc_table[current].flags & VRING_DESC_F_INDIRECT) + { + buffers_required += + vhost_user_compute_indirect_desc_len (vui, txvq, buffer_data_size, + current, &map_hint); + n_left++; + current = (current + 1) & mask; + vhost_user_advance_last_avail_idx (txvq); + } + else + { + buffers_required += + vhost_user_compute_chained_desc_len (vui, txvq, buffer_data_size, + ¤t, &n_left); + } + } + + /* Something is broken if we need more than 10000 buffers */ + if (PREDICT_FALSE ((buffers_required == 0) || (buffers_required > 10000))) + { + rv = vhost_user_rx_discard_packet_packed (vm, vui, txvq, n_left); + vlib_error_count (vm, vhost_user_input_node.index, + VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER, rv); + goto done; + } + + vec_validate (cpu->to_next_list, buffers_required); + rv = vlib_buffer_alloc (vm, cpu->to_next_list, buffers_required); + if (PREDICT_FALSE (rv != buffers_required)) + { + vlib_buffer_free (vm, cpu->to_next_list, rv); + rv = vhost_user_rx_discard_packet_packed (vm, vui, txvq, n_left); + vlib_error_count (vm, vhost_user_input_node.index, + VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER, rv); + goto done; + } + + next = cpu->to_next_list; + vec_validate (cpu->rx_buffers_pdesc, buffers_required); + vlib_get_buffers (vm, next, cpu->rx_buffers_pdesc, buffers_required); + b = cpu->rx_buffers_pdesc; + n_descs_processed = n_left; + + while (n_left) + { + vlib_buffer_t *b_head, *b_current; + u32 bi_current; + u32 desc_data_offset; + u16 desc_idx = desc_current; + u32 n_descs; + + desc_table = txvq->packed_desc; + to_next[0] = bi_current = next[0]; + b_head = b_current = b[0]; + b++; + buffers_used++; + ASSERT (buffers_used <= buffers_required); + to_next++; + next++; + n_left_to_next--; + + /* The buffer should already be initialized */ + b_head->total_length_not_including_first_buffer = 0; + b_head->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID; + desc_data_offset = vui->virtio_net_hdr_sz; + n_descs_to_process = 1; + + if (desc_table[desc_idx].flags & VRING_DESC_F_INDIRECT) + { + n_descs = desc_table[desc_idx].len >> 4; + desc_table = map_guest_mem (vui, desc_table[desc_idx].addr, + &map_hint); + desc_idx = 0; + if (PREDICT_FALSE (desc_table == 0) || + (enable_csum && + (PREDICT_FALSE + (vhost_user_do_offload + (vui, desc_table, desc_idx, mask, b_head, + &map_hint) != VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR)))) + { + vlib_error_count (vm, node->node_index, + VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1); + to_next--; + next--; + n_left_to_next++; + buffers_used--; + b--; + goto out; + } + while (n_descs) + { + vhost_user_assemble_packet (desc_table, &desc_idx, b_head, + &b_current, &next, &b, &bi_current, + cpu, ©_len, &buffers_used, + buffers_required, &desc_data_offset, + buffer_data_size, mask); + n_descs--; + } + } + else + { + if (enable_csum) + { + rv = vhost_user_do_offload (vui, desc_table, desc_idx, mask, + b_head, &map_hint); + if (PREDICT_FALSE (rv != VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR)) + { + vlib_error_count (vm, node->node_index, rv, 1); + to_next--; + next--; + n_left_to_next++; + buffers_used--; + b--; + goto out; + } + } + /* + * For chained descriptor, we process all chains in a single while + * loop. So count how many descriptors in the chain. + */ + n_descs_to_process = 1; + while (desc_table[desc_idx].flags & VRING_DESC_F_NEXT) + { + vhost_user_assemble_packet (desc_table, &desc_idx, b_head, + &b_current, &next, &b, &bi_current, + cpu, ©_len, &buffers_used, + buffers_required, &desc_data_offset, + buffer_data_size, mask); + n_descs_to_process++; + } + vhost_user_assemble_packet (desc_table, &desc_idx, b_head, + &b_current, &next, &b, &bi_current, + cpu, ©_len, &buffers_used, + buffers_required, &desc_data_offset, + buffer_data_size, mask); + } + + 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; + + vnet_buffer (b_head)->sw_if_index[VLIB_RX] = vui->sw_if_index; + vnet_buffer (b_head)->sw_if_index[VLIB_TX] = ~0; + b_head->error = 0; + + if (current_config_index != ~0) + { + b_head->current_config_index = current_config_index; + vnet_buffer (b_head)->feature_arc_index = feature_arc_idx; + } + + out: + ASSERT (n_left >= n_descs_to_process); + n_left -= n_descs_to_process; + + /* advance to next descrptor */ + desc_current = (desc_current + n_descs_to_process) & mask; + + /* + * 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)) + { + rv = vhost_user_input_copy_packed (vui, cpu->copy, copy_len, + &map_hint); + if (PREDICT_FALSE (rv != VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR)) + vlib_error_count (vm, node->node_index, rv, 1); + copy_len = 0; + } + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + + /* Do the memory copies */ + rv = vhost_user_input_copy_packed (vui, cpu->copy, copy_len, &map_hint); + if (PREDICT_FALSE (rv != VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR)) + vlib_error_count (vm, node->node_index, rv, 1); + + /* Must do the tracing before giving buffers back to driver */ + if (PREDICT_FALSE (n_trace)) + { + u32 left = n_rx_packets; + + b = cpu->rx_buffers_pdesc; + while (n_trace && left) + { + if (PREDICT_TRUE + (vlib_trace_buffer + (vm, node, next_index, b[0], /* follow_chain */ 0))) + { + vhost_trace_t *t0; + t0 = vlib_add_trace (vm, node, b[0], sizeof (t0[0])); + vhost_user_rx_trace_packed (t0, vui, qid, txvq, last_used_idx); + last_used_idx = (last_used_idx + 1) & mask; + n_trace--; + vlib_set_trace_count (vm, node, n_trace); + } + left--; + b++; + } + } + + /* + * Give buffers back to driver. + */ + vhost_user_mark_desc_consumed (vui, txvq, desc_head, n_descs_processed); + + /* interrupt (call) handling */ + if ((txvq->callfd_idx != ~0) && + (txvq->avail_event->flags != VRING_EVENT_F_DISABLE)) + { + txvq->n_since_last_int += n_rx_packets; + if (txvq->n_since_last_int > vum->coalesce_frames) + vhost_user_send_call (vm, vui, txvq); + } + + /* increase rx counters */ + vlib_increment_combined_counter + (vnet_main.interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, vm->thread_index, vui->sw_if_index, + n_rx_packets, n_rx_bytes); + + vnet_device_increment_rx_packets (vm->thread_index, n_rx_packets); + + if (PREDICT_FALSE (buffers_used < buffers_required)) + vlib_buffer_free (vm, next, buffers_required - buffers_used); + +done: + return n_rx_packets; +} + +VLIB_NODE_FN (vhost_user_input_node) (vlib_main_t * vm, + vlib_node_runtime_t * node, + vlib_frame_t * frame) +{ + vhost_user_main_t *vum = &vhost_user_main; + uword n_rx_packets = 0; + vhost_user_intf_t *vui; + vnet_hw_if_rxq_poll_vector_t *pv = vnet_hw_if_get_rxq_poll_vector (vm, node); + vnet_hw_if_rxq_poll_vector_t *pve; + + vec_foreach (pve, pv) + { + vui = pool_elt_at_index (vum->vhost_user_interfaces, pve->dev_instance); + if (vhost_user_is_packed_ring_supported (vui)) + { + if (vui->features & VIRTIO_FEATURE (VIRTIO_NET_F_CSUM)) + n_rx_packets += vhost_user_if_input_packed ( + vm, vum, vui, pve->queue_id, node, 1); + else + n_rx_packets += vhost_user_if_input_packed ( + vm, vum, vui, pve->queue_id, node, 0); + } + else + { + if (vui->features & VIRTIO_FEATURE (VIRTIO_NET_F_CSUM)) + n_rx_packets += + vhost_user_if_input (vm, vum, vui, pve->queue_id, node, 1); + else + n_rx_packets += + vhost_user_if_input (vm, vum, vui, pve->queue_id, node, 0); + } + } + + return n_rx_packets; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (vhost_user_input_node) = { + .type = VLIB_NODE_TYPE_INPUT, + .name = "vhost-user-input", + .sibling_of = "device-input", + .flags = VLIB_NODE_FLAG_TRACE_SUPPORTED, + + /* 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, +}; +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/plugins/vhost/vhost_user_output.c b/src/plugins/vhost/vhost_user_output.c new file mode 100644 index 00000000000..9b52fc6265c --- /dev/null +++ b/src/plugins/vhost/vhost_user_output.c @@ -0,0 +1,1145 @@ +/* + *------------------------------------------------------------------ + * vhost-user-output + * + * Copyright (c) 2014-2018 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 /* for open */ +#include +#include +#include +#include +#include +#include /* for iovec */ +#include +#include + +#include +#include + +#include +#include + +#include +#include +#include +#include + +#include +#include + +#include +/* + * On the transmit side, we keep processing the buffers from vlib in the while + * loop and prepare the copy order to be executed later. However, the static + * array which we keep the copy order is limited to VHOST_USER_COPY_ARRAY_N + * entries. In order to not corrupt memory, we have to do the copy when the + * static array reaches the copy threshold. We subtract 40 in case the code + * goes into the inner loop for a maximum of 64k frames which may require + * more array entries. We subtract 200 because our default buffer size is + * 2048 and the default desc len is likely 1536. While it takes less than 40 + * vlib buffers for the jumbo frame, it may take twice as much descriptors + * for the same jumbo frame. Use 200 for the extra head room. + */ +#define VHOST_USER_TX_COPY_THRESHOLD (VHOST_USER_COPY_ARRAY_N - 200) + +extern vnet_device_class_t vhost_user_device_class; + +#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 __clib_unused char *vhost_user_tx_func_error_strings[] = { +#define _(n,s) s, + foreach_vhost_user_tx_func_error +#undef _ +}; + +static __clib_unused 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 __clib_unused 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; + vhost_user_intf_t *vui = pool_elt_at_index (vum->vhost_user_interfaces, + hi->dev_instance); + + 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; + + vu_log_debug (vui, "renumbered vhost-user interface dev_instance %d to %d", + hi->dev_instance, new_dev_instance); + + return 0; +} + +static_always_inline 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 last_avail_idx = rxvq->last_avail_idx; + u32 desc_current = rxvq->avail->ring[last_avail_idx & rxvq->qsz_mask]; + vnet_virtio_vring_desc_t *hdr_desc = 0; + u32 hint = 0; + + clib_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 & VRING_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 & VRING_DESC_F_NEXT) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SIMPLE_CHAINED; + } + if (!(rxvq->desc[desc_current].flags & VRING_DESC_F_NEXT) && + !(rxvq->desc[desc_current].flags & VRING_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_load ((void *) cpy[2].src); + clib_prefetch_load ((void *) cpy[3].src); + + clib_memcpy_fast (dst0, (void *) cpy[0].src, cpy[0].len); + clib_memcpy_fast (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_fast (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_always_inline void +vhost_user_handle_tx_offload (vhost_user_intf_t *vui, vlib_buffer_t *b, + vnet_virtio_net_hdr_t *hdr) +{ + generic_header_offset_t gho = { 0 }; + int is_ip4 = b->flags & VNET_BUFFER_F_IS_IP4; + int is_ip6 = b->flags & VNET_BUFFER_F_IS_IP6; + vnet_buffer_oflags_t oflags = vnet_buffer (b)->oflags; + u16 psh_cksum = 0; + ip4_header_t *ip4 = 0; + ip6_header_t *ip6 = 0; + + ASSERT (!(is_ip4 && is_ip6)); + vnet_generic_header_offset_parser (b, &gho, 1 /* l2 */ , is_ip4, is_ip6); + if (oflags & VNET_BUFFER_OFFLOAD_F_IP_CKSUM) + { + ip4 = + (ip4_header_t *) (vlib_buffer_get_current (b) + gho.l3_hdr_offset); + ip4->checksum = ip4_header_checksum (ip4); + psh_cksum = ip4_pseudo_header_cksum (ip4); + } + else + { + ip6 = (ip6_header_t *) (vlib_buffer_get_current (b) + gho.l3_hdr_offset); + psh_cksum = ip6_pseudo_header_cksum (ip6); + } + + /* checksum offload */ + if (oflags & VNET_BUFFER_OFFLOAD_F_UDP_CKSUM) + { + udp_header_t *udp = + (udp_header_t *) (vlib_buffer_get_current (b) + gho.l4_hdr_offset); + udp->checksum = psh_cksum; + hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM; + hdr->csum_start = gho.l4_hdr_offset; + hdr->csum_offset = offsetof (udp_header_t, checksum); + } + else if (oflags & VNET_BUFFER_OFFLOAD_F_TCP_CKSUM) + { + tcp_header_t *tcp = + (tcp_header_t *) (vlib_buffer_get_current (b) + gho.l4_hdr_offset); + tcp->checksum = psh_cksum; + hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM; + hdr->csum_start = gho.l4_hdr_offset; + hdr->csum_offset = offsetof (tcp_header_t, checksum); + } + + /* GSO offload */ + if (b->flags & VNET_BUFFER_F_GSO) + { + if (oflags & VNET_BUFFER_OFFLOAD_F_TCP_CKSUM) + { + if (is_ip4 && + (vui->features & VIRTIO_FEATURE (VIRTIO_NET_F_GUEST_TSO4))) + { + hdr->gso_size = vnet_buffer2 (b)->gso_size; + hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4; + } + else if (is_ip6 && + (vui->features & VIRTIO_FEATURE (VIRTIO_NET_F_GUEST_TSO6))) + { + hdr->gso_size = vnet_buffer2 (b)->gso_size; + hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6; + } + } + else if ((vui->features & VIRTIO_FEATURE (VIRTIO_NET_F_GUEST_UFO)) && + (oflags & VNET_BUFFER_OFFLOAD_F_UDP_CKSUM)) + { + hdr->gso_size = vnet_buffer2 (b)->gso_size; + hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP; + } + } +} + +static_always_inline void +vhost_user_mark_desc_available (vlib_main_t * vm, vhost_user_intf_t * vui, + vhost_user_vring_t * rxvq, + u16 * n_descs_processed, u8 chained, + vlib_frame_t * frame, u32 n_left) +{ + u16 desc_idx, flags; + vnet_virtio_vring_packed_desc_t *desc_table = rxvq->packed_desc; + u16 last_used_idx = rxvq->last_used_idx; + + if (PREDICT_FALSE (*n_descs_processed == 0)) + return; + + if (rxvq->used_wrap_counter) + flags = desc_table[last_used_idx & rxvq->qsz_mask].flags | + (VRING_DESC_F_AVAIL | VRING_DESC_F_USED); + else + flags = desc_table[last_used_idx & rxvq->qsz_mask].flags & + ~(VRING_DESC_F_AVAIL | VRING_DESC_F_USED); + + vhost_user_advance_last_used_idx (rxvq); + + for (desc_idx = 1; desc_idx < *n_descs_processed; desc_idx++) + { + if (rxvq->used_wrap_counter) + desc_table[rxvq->last_used_idx & rxvq->qsz_mask].flags |= + (VRING_DESC_F_AVAIL | VRING_DESC_F_USED); + else + desc_table[rxvq->last_used_idx & rxvq->qsz_mask].flags &= + ~(VRING_DESC_F_AVAIL | VRING_DESC_F_USED); + vhost_user_advance_last_used_idx (rxvq); + } + + desc_table[last_used_idx & rxvq->qsz_mask].flags = flags; + + *n_descs_processed = 0; + + if (chained) + { + vnet_virtio_vring_packed_desc_t *desc_table = rxvq->packed_desc; + + while (desc_table[rxvq->last_used_idx & rxvq->qsz_mask].flags & + VRING_DESC_F_NEXT) + vhost_user_advance_last_used_idx (rxvq); + + /* Advance past the current chained table entries */ + vhost_user_advance_last_used_idx (rxvq); + } + + /* interrupt (call) handling */ + if ((rxvq->callfd_idx != ~0) && + (rxvq->avail_event->flags != VRING_EVENT_F_DISABLE)) + { + vhost_user_main_t *vum = &vhost_user_main; + + rxvq->n_since_last_int += frame->n_vectors - n_left; + if (rxvq->n_since_last_int > vum->coalesce_frames) + vhost_user_send_call (vm, vui, rxvq); + } +} + +static_always_inline void +vhost_user_tx_trace_packed (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 last_avail_idx = rxvq->last_avail_idx; + u32 desc_current = last_avail_idx & rxvq->qsz_mask; + vnet_virtio_vring_packed_desc_t *hdr_desc = 0; + u32 hint = 0; + + clib_memset (t, 0, sizeof (*t)); + t->device_index = vui - vum->vhost_user_interfaces; + t->qid = qid; + + hdr_desc = &rxvq->packed_desc[desc_current]; + if (rxvq->packed_desc[desc_current].flags & VRING_DESC_F_INDIRECT) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_INDIRECT; + /* Header is the first here */ + hdr_desc = map_guest_mem (vui, rxvq->packed_desc[desc_current].addr, + &hint); + } + if (rxvq->packed_desc[desc_current].flags & VRING_DESC_F_NEXT) + { + t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SIMPLE_CHAINED; + } + if (!(rxvq->packed_desc[desc_current].flags & VRING_DESC_F_NEXT) && + !(rxvq->packed_desc[desc_current].flags & VRING_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 uword +vhost_user_device_class_packed (vlib_main_t *vm, vlib_node_runtime_t *node, + vlib_frame_t *frame, vhost_user_intf_t *vui, + vhost_user_vring_t *rxvq) +{ + u32 *buffers = vlib_frame_vector_args (frame); + u32 n_left = frame->n_vectors; + vhost_user_main_t *vum = &vhost_user_main; + u32 qid = rxvq->qid; + u8 error; + u32 thread_index = vm->thread_index; + vhost_cpu_t *cpu = &vum->cpus[thread_index]; + u32 map_hint = 0; + u8 retry = 8; + u16 copy_len; + u16 tx_headers_len; + vnet_virtio_vring_packed_desc_t *desc_table; + u32 or_flags; + u16 desc_head, desc_index, desc_len; + u16 n_descs_processed; + u8 indirect, chained; + +retry: + error = VHOST_USER_TX_FUNC_ERROR_NONE; + tx_headers_len = 0; + copy_len = 0; + n_descs_processed = 0; + + while (n_left > 0) + { + vlib_buffer_t *b0, *current_b0; + uword buffer_map_addr; + u32 buffer_len; + u16 bytes_left; + u32 total_desc_len = 0; + u16 n_entries = 0; + + indirect = 0; + chained = 0; + 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)) + { + cpu->current_trace = vlib_add_trace (vm, node, b0, + sizeof (*cpu->current_trace)); + vhost_user_tx_trace_packed (cpu->current_trace, vui, qid / 2, b0, + rxvq); + } + + desc_table = rxvq->packed_desc; + desc_head = desc_index = rxvq->last_avail_idx & rxvq->qsz_mask; + if (PREDICT_FALSE (!vhost_user_packed_desc_available (rxvq, desc_head))) + { + error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; + goto done; + } + /* + * Go deeper in case of indirect descriptor. + * To test it, turn off mrg_rxbuf. + */ + if (desc_table[desc_head].flags & VRING_DESC_F_INDIRECT) + { + indirect = 1; + if (PREDICT_FALSE (desc_table[desc_head].len < + sizeof (vnet_virtio_vring_packed_desc_t))) + { + error = VHOST_USER_TX_FUNC_ERROR_INDIRECT_OVERFLOW; + goto done; + } + n_entries = desc_table[desc_head].len >> 4; + desc_table = map_guest_mem (vui, desc_table[desc_index].addr, + &map_hint); + if (PREDICT_FALSE (desc_table == 0)) + { + error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; + goto done; + } + desc_index = 0; + } + else if (rxvq->packed_desc[desc_head].flags & VRING_DESC_F_NEXT) + chained = 1; + + 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 */ + vnet_virtio_net_hdr_mrg_rxbuf_t *hdr = &cpu->tx_headers[tx_headers_len]; + tx_headers_len++; + hdr->hdr.flags = 0; + hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE; + hdr->num_buffers = 1; + + or_flags = (b0->flags & VNET_BUFFER_F_OFFLOAD); + + /* Guest supports csum offload and buffer requires checksum offload? */ + if (or_flags && + (vui->features & VIRTIO_FEATURE (VIRTIO_NET_F_GUEST_CSUM))) + vhost_user_handle_tx_offload (vui, b0, &hdr->hdr); + + /* Prepare a copy order executed later for the header */ + ASSERT (copy_len < VHOST_USER_COPY_ARRAY_N); + vhost_copy_t *cpy = &cpu->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 (chained) + { + /* + * Next one is chained + * Test it with both indirect and mrg_rxbuf off + */ + if (PREDICT_FALSE (!(desc_table[desc_index].flags & + VRING_DESC_F_NEXT))) + { + /* + * Last descriptor in chain. + * Dequeue queued descriptors for this packet + */ + vhost_user_dequeue_chained_descs (rxvq, + &n_descs_processed); + error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; + goto done; + } + vhost_user_advance_last_avail_idx (rxvq); + desc_index = rxvq->last_avail_idx & rxvq->qsz_mask; + n_descs_processed++; + buffer_map_addr = desc_table[desc_index].addr; + buffer_len = desc_table[desc_index].len; + total_desc_len += desc_len; + desc_len = 0; + } + else if (indirect) + { + /* + * Indirect table + * Test it with mrg_rxnuf off + */ + if (PREDICT_TRUE (n_entries > 0)) + n_entries--; + else + { + /* Dequeue queued descriptors for this packet */ + vhost_user_dequeue_chained_descs (rxvq, + &n_descs_processed); + error = VHOST_USER_TX_FUNC_ERROR_INDIRECT_OVERFLOW; + goto done; + } + total_desc_len += desc_len; + desc_index = (desc_index + 1) & rxvq->qsz_mask; + buffer_map_addr = desc_table[desc_index].addr; + buffer_len = desc_table[desc_index].len; + desc_len = 0; + } + else if (vui->virtio_net_hdr_sz == 12) + { + /* + * MRG is available + * This is the default setting for the guest VM + */ + vnet_virtio_net_hdr_mrg_rxbuf_t *hdr = + &cpu->tx_headers[tx_headers_len - 1]; + + desc_table[desc_index].len = desc_len; + vhost_user_advance_last_avail_idx (rxvq); + desc_head = desc_index = + rxvq->last_avail_idx & rxvq->qsz_mask; + hdr->num_buffers++; + n_descs_processed++; + desc_len = 0; + + if (PREDICT_FALSE (!vhost_user_packed_desc_available + (rxvq, desc_index))) + { + /* Dequeue queued descriptors for this packet */ + vhost_user_dequeue_descs (rxvq, hdr, + &n_descs_processed); + error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; + goto done; + } + + 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; + } + } + + ASSERT (copy_len < VHOST_USER_COPY_ARRAY_N); + vhost_copy_t *cpy = &cpu->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_load (&rxvq->packed_desc); + + /* 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 */ + total_desc_len += desc_len; + rxvq->packed_desc[desc_head].len = total_desc_len; + + vhost_user_advance_last_avail_table_idx (vui, rxvq, chained); + n_descs_processed++; + + if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED)) + cpu->current_trace->hdr = cpu->tx_headers[tx_headers_len - 1]; + + n_left--; + + /* + * Do the copy periodically to prevent + * cpu->copy array overflow and corrupt memory + */ + if (PREDICT_FALSE (copy_len >= VHOST_USER_TX_COPY_THRESHOLD) || chained) + { + if (PREDICT_FALSE (vhost_user_tx_copy (vui, cpu->copy, copy_len, + &map_hint))) + vlib_error_count (vm, node->node_index, + VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL, 1); + copy_len = 0; + + /* give buffers back to driver */ + vhost_user_mark_desc_available (vm, vui, rxvq, &n_descs_processed, + chained, frame, n_left); + } + + buffers++; + } + +done: + if (PREDICT_TRUE (copy_len)) + { + if (PREDICT_FALSE (vhost_user_tx_copy (vui, cpu->copy, copy_len, + &map_hint))) + vlib_error_count (vm, node->node_index, + VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL, 1); + + vhost_user_mark_desc_available (vm, vui, rxvq, &n_descs_processed, + chained, frame, n_left); + } + + /* + * 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; + } + + clib_spinlock_unlock (&rxvq->vring_lock); + + 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, thread_index, vui->sw_if_index, n_left); + } + + vlib_buffer_free (vm, vlib_frame_vector_args (frame), frame->n_vectors); + return frame->n_vectors; +} + +VNET_DEVICE_CLASS_TX_FN (vhost_user_device_class) (vlib_main_t * vm, + vlib_node_runtime_t * + node, vlib_frame_t * frame) +{ + u32 *buffers = vlib_frame_vector_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; + vhost_user_vring_t *rxvq; + u8 error; + u32 thread_index = vm->thread_index; + vhost_cpu_t *cpu = &vum->cpus[thread_index]; + u32 map_hint = 0; + u8 retry = 8; + u16 copy_len; + u16 tx_headers_len; + u32 or_flags; + vnet_hw_if_tx_frame_t *tf = vlib_frame_scalar_args (frame); + + if (PREDICT_FALSE (!vui->admin_up)) + { + error = VHOST_USER_TX_FUNC_ERROR_DOWN; + goto done3; + } + + if (PREDICT_FALSE (!vui->is_ready)) + { + error = VHOST_USER_TX_FUNC_ERROR_NOT_READY; + goto done3; + } + + qid = VHOST_VRING_IDX_RX (tf->queue_id); + rxvq = &vui->vrings[qid]; + ASSERT (tf->queue_id == rxvq->qid); + + if (PREDICT_FALSE (rxvq->avail == 0)) + { + error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; + goto done3; + } + if (tf->shared_queue) + clib_spinlock_lock (&rxvq->vring_lock); + + if (vhost_user_is_packed_ring_supported (vui)) + return (vhost_user_device_class_packed (vm, node, frame, vui, rxvq)); + +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; + vnet_virtio_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)) + { + cpu->current_trace = vlib_add_trace (vm, node, b0, + sizeof (*cpu->current_trace)); + vhost_user_tx_trace (cpu->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 & rxvq->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 & VRING_DESC_F_INDIRECT)) + { + if (PREDICT_FALSE (rxvq->desc[desc_head].len < + sizeof (vnet_virtio_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 + vnet_virtio_net_hdr_mrg_rxbuf_t *hdr = + &cpu->tx_headers[tx_headers_len]; + tx_headers_len++; + hdr->hdr.flags = 0; + hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE; + hdr->num_buffers = 1; //This is local, no need to check + + or_flags = (b0->flags & VNET_BUFFER_F_OFFLOAD); + + /* Guest supports csum offload and buffer requires checksum offload? */ + if (or_flags + && (vui->features & VIRTIO_FEATURE (VIRTIO_NET_F_GUEST_CSUM))) + vhost_user_handle_tx_offload (vui, b0, &hdr->hdr); + + // Prepare a copy order executed later for the header + ASSERT (copy_len < VHOST_USER_COPY_ARRAY_N); + vhost_copy_t *cpy = &cpu->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 & VRING_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 + { + vnet_virtio_net_hdr_mrg_rxbuf_t *hdr = + &cpu->tx_headers[tx_headers_len - 1]; + + //Move from available to used buffer + rxvq->used->ring[rxvq->last_used_idx & rxvq->qsz_mask].id = + desc_head; + rxvq->used->ring[rxvq->last_used_idx & rxvq->qsz_mask].len = + desc_len; + vhost_user_log_dirty_ring (vui, rxvq, + ring[rxvq->last_used_idx & + rxvq->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 & rxvq->qsz_mask]; + if (PREDICT_FALSE + (rxvq->desc[desc_head].flags & VRING_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 (vnet_virtio_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; + } + } + + { + ASSERT (copy_len < VHOST_USER_COPY_ARRAY_N); + vhost_copy_t *cpy = &cpu->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_load (&rxvq->desc); + } + + // 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 & rxvq->qsz_mask].id = desc_head; + rxvq->used->ring[rxvq->last_used_idx & rxvq->qsz_mask].len = desc_len; + vhost_user_log_dirty_ring (vui, rxvq, + ring[rxvq->last_used_idx & rxvq->qsz_mask]); + rxvq->last_avail_idx++; + rxvq->last_used_idx++; + + if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED)) + { + cpu->current_trace->hdr = cpu->tx_headers[tx_headers_len - 1]; + } + + n_left--; //At the end for error counting when 'goto done' is invoked + + /* + * Do the copy periodically to prevent + * cpu->copy array overflow and corrupt memory + */ + if (PREDICT_FALSE (copy_len >= VHOST_USER_TX_COPY_THRESHOLD)) + { + if (PREDICT_FALSE (vhost_user_tx_copy (vui, cpu->copy, copy_len, + &map_hint))) + { + vlib_error_count (vm, node->node_index, + VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL, 1); + } + copy_len = 0; + + /* give buffers back to driver */ + CLIB_MEMORY_BARRIER (); + rxvq->used->idx = rxvq->last_used_idx; + vhost_user_log_dirty_ring (vui, rxvq, idx); + } + buffers++; + } + +done: + //Do the memory copies + if (PREDICT_FALSE (vhost_user_tx_copy (vui, cpu->copy, copy_len, + &map_hint))) + { + vlib_error_count (vm, node->node_index, + VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL, 1); + } + + 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 likelihood 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 & VRING_AVAIL_F_NO_INTERRUPT)) + { + rxvq->n_since_last_int += frame->n_vectors - n_left; + + if (rxvq->n_since_last_int > vum->coalesce_frames) + vhost_user_send_call (vm, vui, rxvq); + } + + clib_spinlock_unlock (&rxvq->vring_lock); + +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, + thread_index, vui->sw_if_index, n_left); + } + + vlib_buffer_free (vm, vlib_frame_vector_args (frame), frame->n_vectors); + return frame->n_vectors; +} + +static __clib_unused clib_error_t * +vhost_user_interface_rx_mode_change (vnet_main_t * vnm, u32 hw_if_index, + u32 qid, vnet_hw_if_rx_mode mode) +{ + vlib_main_t *vm = vnm->vlib_main; + vnet_hw_interface_t *hif = vnet_get_hw_interface (vnm, hw_if_index); + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui = + pool_elt_at_index (vum->vhost_user_interfaces, hif->dev_instance); + vhost_user_vring_t *txvq = &vui->vrings[VHOST_VRING_IDX_TX (qid)]; + vhost_cpu_t *cpu; + + if (mode == txvq->mode) + return 0; + + if ((mode != VNET_HW_IF_RX_MODE_POLLING) && + (mode != VNET_HW_IF_RX_MODE_ADAPTIVE) && + (mode != VNET_HW_IF_RX_MODE_INTERRUPT)) + { + vu_log_err (vui, "unhandled mode %d changed for if %d queue %d", mode, + hw_if_index, qid); + return clib_error_return (0, "unsupported"); + } + + if (txvq->thread_index == ~0) + return clib_error_return (0, "Queue initialization is not finished yet"); + + cpu = vec_elt_at_index (vum->cpus, txvq->thread_index); + if ((mode == VNET_HW_IF_RX_MODE_INTERRUPT) || + (mode == VNET_HW_IF_RX_MODE_ADAPTIVE)) + { + if (txvq->kickfd_idx == ~0) + { + // We cannot support interrupt mode if the driver opts out + return clib_error_return (0, "Driver does not support interrupt"); + } + if (txvq->mode == VNET_HW_IF_RX_MODE_POLLING) + { + ASSERT (cpu->polling_q_count != 0); + if (cpu->polling_q_count) + cpu->polling_q_count--; + vum->ifq_count++; + // Start the timer if this is the first encounter on interrupt + // interface/queue + if ((vum->ifq_count == 1) && + ((vum->coalesce_time > 0.0) || (vum->coalesce_frames > 0))) + vlib_process_signal_event (vm, + vhost_user_send_interrupt_node.index, + VHOST_USER_EVENT_START_TIMER, 0); + } + } + else if (mode == VNET_HW_IF_RX_MODE_POLLING) + { + if (((txvq->mode == VNET_HW_IF_RX_MODE_INTERRUPT) || + (txvq->mode == VNET_HW_IF_RX_MODE_ADAPTIVE)) && vum->ifq_count) + { + cpu->polling_q_count++; + vum->ifq_count--; + // Stop the timer if there is no more interrupt interface/queue + if (vum->ifq_count == 0) + vlib_process_signal_event (vm, + vhost_user_send_interrupt_node.index, + VHOST_USER_EVENT_STOP_TIMER, 0); + } + } + + txvq->mode = mode; + vhost_user_set_operation_mode (vui, txvq); + + return 0; +} + +static __clib_unused 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); + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui = + pool_elt_at_index (vum->vhost_user_interfaces, hif->dev_instance); + u8 link_old, link_new; + + link_old = vui_is_link_up (vui); + + vui->admin_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0; + + link_new = vui_is_link_up (vui); + + if (link_old != link_new) + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, link_new ? + VNET_HW_INTERFACE_FLAG_LINK_UP : 0); + + return /* no error */ 0; +} + +/* *INDENT-OFF* */ +VNET_DEVICE_CLASS (vhost_user_device_class) = { + .name = "vhost-user", + .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, + .rx_mode_change_function = vhost_user_interface_rx_mode_change, + .format_tx_trace = format_vhost_trace, +}; + +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/plugins/vhost/virtio_std.h b/src/plugins/vhost/virtio_std.h new file mode 100644 index 00000000000..fa826933a9c --- /dev/null +++ b/src/plugins/vhost/virtio_std.h @@ -0,0 +1,188 @@ +/* + * 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_STD_H__ +#define __VIRTIO_STD_H__ + +#define foreach_virtio_net_features \ + _ (VIRTIO_NET_F_CSUM, 0) /* Host handles pkts w/ partial csum */ \ + _ (VIRTIO_NET_F_GUEST_CSUM, 1) /* Guest handles pkts w/ partial csum */ \ + _ (VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \ + 2) /* Dynamic offload configuration. */ \ + _ (VIRTIO_NET_F_MTU, 3) /* Initial MTU advice. */ \ + _ (VIRTIO_NET_F_MAC, 5) /* Host has given MAC address. */ \ + _ (VIRTIO_NET_F_GSO, 6) /* Host handles pkts w/ any GSO. */ \ + _ (VIRTIO_NET_F_GUEST_TSO4, 7) /* Guest can handle TSOv4 in. */ \ + _ (VIRTIO_NET_F_GUEST_TSO6, 8) /* Guest can handle TSOv6 in. */ \ + _ (VIRTIO_NET_F_GUEST_ECN, 9) /* Guest can handle TSO[6] w/ ECN in. */ \ + _ (VIRTIO_NET_F_GUEST_UFO, 10) /* Guest can handle UFO in. */ \ + _ (VIRTIO_NET_F_HOST_TSO4, 11) /* Host can handle TSOv4 in. */ \ + _ (VIRTIO_NET_F_HOST_TSO6, 12) /* Host can handle TSOv6 in. */ \ + _ (VIRTIO_NET_F_HOST_ECN, 13) /* Host can handle TSO[6] w/ ECN in. */ \ + _ (VIRTIO_NET_F_HOST_UFO, 14) /* Host can handle UFO in. */ \ + _ (VIRTIO_NET_F_MRG_RXBUF, 15) /* Host can merge receive buffers. */ \ + _ (VIRTIO_NET_F_STATUS, 16) /* virtio_net_config.status available */ \ + _ (VIRTIO_NET_F_CTRL_VQ, 17) /* Control channel available */ \ + _ (VIRTIO_NET_F_CTRL_RX, 18) /* Control channel RX mode support */ \ + _ (VIRTIO_NET_F_CTRL_VLAN, 19) /* Control channel VLAN filtering */ \ + _ (VIRTIO_NET_F_CTRL_RX_EXTRA, 20) /* Extra RX mode control support */ \ + _ (VIRTIO_NET_F_GUEST_ANNOUNCE, \ + 21) /* Guest can announce device on the network */ \ + _ (VIRTIO_NET_F_MQ, 22) /* Device supports Receive Flow Steering */ \ + _ (VIRTIO_NET_F_CTRL_MAC_ADDR, 23) /* Set MAC address */ \ + _ (VIRTIO_F_NOTIFY_ON_EMPTY, 24) \ + _ (VHOST_F_LOG_ALL, 26) /* Log all write descriptors */ \ + _ (VIRTIO_F_ANY_LAYOUT, \ + 27) /* Can the device handle any descriptor layout */ \ + _ (VIRTIO_RING_F_INDIRECT_DESC, \ + 28) /* Support indirect buffer descriptors */ \ + _ (VIRTIO_RING_F_EVENT_IDX, \ + 29) /* The Guest publishes the used index for which it expects an \ + * interrupt at the end of the avail ring. Host should ignore the \ + * avail->flags field. */ \ + /* The Host publishes the avail index for which it expects a kick \ + * at the end of the used ring. Guest should ignore the used->flags field. \ + */ \ + _ (VHOST_USER_F_PROTOCOL_FEATURES, 30) \ + _ (VIRTIO_F_VERSION_1, 32) /* v1.0 compliant. */ \ + _ (VIRTIO_F_IOMMU_PLATFORM, 33) \ + _ (VIRTIO_F_RING_PACKED, 34) \ + _ (VIRTIO_F_IN_ORDER, 35) /* all buffers are used by the device in the */ \ + /* same order in which they have been made available */ \ + _ (VIRTIO_F_ORDER_PLATFORM, 36) /* memory accesses by the driver and the */ \ + /* device are ordered in a way described by the platfor */ \ + _ (VIRTIO_F_NOTIFICATION_DATA, \ + 38) /* the driver passes extra data (besides */ \ + /* identifying the virtqueue) in its device notifications. */ \ + _ (VIRTIO_NET_F_SPEED_DUPLEX, 63) /* Device set linkspeed and duplex */ + +typedef enum +{ +#define _(f, n) f = n, + foreach_virtio_net_features +#undef _ +} vnet_virtio_net_feature_t; + +#define VIRTIO_FEATURE(X) (1ULL << X) + +#define VRING_MAX_SIZE 32768 + +#define VRING_DESC_F_NEXT 1 +#define VRING_DESC_F_WRITE 2 +#define VRING_DESC_F_INDIRECT 4 + +#define VRING_DESC_F_AVAIL (1 << 7) +#define VRING_DESC_F_USED (1 << 15) + +#define foreach_virtio_event_idx_flags \ + _ (VRING_EVENT_F_ENABLE, 0) \ + _ (VRING_EVENT_F_DISABLE, 1) \ + _ (VRING_EVENT_F_DESC, 2) + +typedef enum +{ +#define _(f, n) f = n, + foreach_virtio_event_idx_flags +#undef _ +} vnet_virtio_event_idx_flags_t; + +#define VRING_USED_F_NO_NOTIFY 1 +#define VRING_AVAIL_F_NO_INTERRUPT 1 + +typedef struct +{ + u64 addr; + u32 len; + u16 flags; + u16 next; +} vnet_virtio_vring_desc_t; + +typedef struct +{ + u16 flags; + u16 idx; + u16 ring[0]; + /* u16 used_event; */ +} vnet_virtio_vring_avail_t; + +typedef struct +{ + u32 id; + u32 len; +} vnet_virtio_vring_used_elem_t; + +typedef struct +{ + u16 flags; + u16 idx; + vnet_virtio_vring_used_elem_t ring[0]; + /* u16 avail_event; */ +} vnet_virtio_vring_used_t; + +typedef CLIB_PACKED (struct { + u64 addr; // packet data buffer address + u32 len; // packet data buffer size + u16 id; // buffer id + u16 flags; // flags +}) vnet_virtio_vring_packed_desc_t; + +STATIC_ASSERT_SIZEOF (vnet_virtio_vring_packed_desc_t, 16); + +typedef CLIB_PACKED (struct { + u16 off_wrap; + u16 flags; +}) vnet_virtio_vring_desc_event_t; + +#define VIRTIO_NET_HDR_F_NEEDS_CSUM 1 /* Use csum_start, csum_offset */ +#define VIRTIO_NET_HDR_F_DATA_VALID 2 /* Csum is valid */ + +#define VIRTIO_NET_HDR_GSO_NONE 0 /* Not a GSO frame */ +#define VIRTIO_NET_HDR_GSO_TCPV4 1 /* GSO frame, IPv4 TCP (TSO) */ +#define VIRTIO_NET_HDR_GSO_UDP 3 /* GSO frame, IPv4 UDP (UFO) */ +#define VIRTIO_NET_HDR_GSO_TCPV6 4 /* GSO frame, IPv6 TCP */ +#define VIRTIO_NET_HDR_GSO_ECN 0x80 /* TCP has ECN set */ + +typedef CLIB_PACKED (struct { + u8 flags; + u8 gso_type; + u16 hdr_len; /* Ethernet + IP + tcp/udp hdrs */ + u16 gso_size; /* Bytes to append to hdr_len per frame */ + u16 csum_start; /* Position to start checksumming from */ + u16 csum_offset; /* Offset after that to place checksum */ + u16 num_buffers; /* Number of merged rx buffers */ +}) vnet_virtio_net_hdr_v1_t; + +typedef CLIB_PACKED (struct { + u8 flags; + u8 gso_type; + u16 hdr_len; + u16 gso_size; + u16 csum_start; + u16 csum_offset; +}) vnet_virtio_net_hdr_t; + +typedef CLIB_PACKED (struct { + vnet_virtio_net_hdr_t hdr; + u16 num_buffers; +}) vnet_virtio_net_hdr_mrg_rxbuf_t; + +#endif + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ -- cgit 1.2.3-korg