diff options
Diffstat (limited to 'src/vnet/devices/dpdk')
| -rw-r--r-- | src/vnet/devices/dpdk/cli.c | 1296 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/device.c | 840 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/dpdk.h | 534 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/dpdk_priv.h | 132 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/format.c | 763 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/hqos.c | 775 | ||||
| -rwxr-xr-x | src/vnet/devices/dpdk/init.c | 1803 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/ipsec/cli.c | 141 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/ipsec/crypto_node.c | 210 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/ipsec/dir.dox | 18 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/ipsec/dpdk_crypto_ipsec_doc.md | 73 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/ipsec/esp.h | 295 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/ipsec/esp_decrypt.c | 583 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/ipsec/esp_encrypt.c | 598 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/ipsec/ipsec.c | 313 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/ipsec/ipsec.h | 227 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/node.c | 687 | ||||
| -rw-r--r-- | src/vnet/devices/dpdk/qos_doc.md | 404 |
18 files changed, 9692 insertions, 0 deletions
diff --git a/src/vnet/devices/dpdk/cli.c b/src/vnet/devices/dpdk/cli.c new file mode 100644 index 00000000000..538a00fd975 --- /dev/null +++ b/src/vnet/devices/dpdk/cli.c @@ -0,0 +1,1296 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include <vnet/vnet.h> +#include <vppinfra/vec.h> +#include <vppinfra/error.h> +#include <vppinfra/format.h> +#include <vppinfra/xxhash.h> + +#include <vnet/ethernet/ethernet.h> +#include <vnet/devices/dpdk/dpdk.h> +#include <vnet/classify/vnet_classify.h> +#include <vnet/mpls/packet.h> + +#include "dpdk_priv.h" + +static clib_error_t * +pcap_trace_command_fn (vlib_main_t * vm, + unformat_input_t * input, vlib_cli_command_t * cmd) +{ + dpdk_main_t *dm = &dpdk_main; + u8 *filename; + u32 max; + int matched = 0; + clib_error_t *error = 0; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "on")) + { + if (dm->tx_pcap_enable == 0) + { + if (dm->pcap_filename == 0) + dm->pcap_filename = format (0, "/tmp/vpe.pcap%c", 0); + + memset (&dm->pcap_main, 0, sizeof (dm->pcap_main)); + dm->pcap_main.file_name = (char *) dm->pcap_filename; + dm->pcap_main.n_packets_to_capture = 100; + if (dm->pcap_pkts_to_capture) + dm->pcap_main.n_packets_to_capture = dm->pcap_pkts_to_capture; + + dm->pcap_main.packet_type = PCAP_PACKET_TYPE_ethernet; + dm->tx_pcap_enable = 1; + matched = 1; + vlib_cli_output (vm, "pcap tx capture on..."); + } + else + { + vlib_cli_output (vm, "pcap tx capture already on..."); + } + matched = 1; + } + else if (unformat (input, "off")) + { + if (dm->tx_pcap_enable) + { + vlib_cli_output (vm, "captured %d pkts...", + dm->pcap_main.n_packets_captured + 1); + if (dm->pcap_main.n_packets_captured) + { + dm->pcap_main.n_packets_to_capture = + dm->pcap_main.n_packets_captured; + error = pcap_write (&dm->pcap_main); + if (error) + clib_error_report (error); + else + vlib_cli_output (vm, "saved to %s...", dm->pcap_filename); + } + } + else + { + vlib_cli_output (vm, "pcap tx capture already off..."); + } + + dm->tx_pcap_enable = 0; + matched = 1; + } + else if (unformat (input, "max %d", &max)) + { + dm->pcap_pkts_to_capture = max; + matched = 1; + } + + else if (unformat (input, "intfc %U", + unformat_vnet_sw_interface, dm->vnet_main, + &dm->pcap_sw_if_index)) + matched = 1; + else if (unformat (input, "intfc any")) + { + dm->pcap_sw_if_index = 0; + matched = 1; + } + else if (unformat (input, "file %s", &filename)) + { + u8 *chroot_filename; + /* Brain-police user path input */ + if (strstr ((char *) filename, "..") + || index ((char *) filename, '/')) + { + vlib_cli_output (vm, "illegal characters in filename '%s'", + filename); + continue; + } + + chroot_filename = format (0, "/tmp/%s%c", filename, 0); + vec_free (filename); + + if (dm->pcap_filename) + vec_free (dm->pcap_filename); + vec_add1 (filename, 0); + dm->pcap_filename = chroot_filename; + matched = 1; + } + else if (unformat (input, "status")) + { + if (dm->tx_pcap_enable == 0) + { + vlib_cli_output (vm, "pcap tx capture is off..."); + continue; + } + + vlib_cli_output (vm, "pcap tx capture: %d of %d pkts...", + dm->pcap_main.n_packets_captured, + dm->pcap_main.n_packets_to_capture); + matched = 1; + } + + else + break; + } + + if (matched == 0) + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (pcap_trace_command, static) = { + .path = "pcap tx trace", + .short_help = + "pcap tx trace on off max <nn> intfc <intfc> file <name> status", + .function = pcap_trace_command_fn, +}; +/* *INDENT-ON* */ + + +static clib_error_t * +show_dpdk_buffer (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + struct rte_mempool *rmp; + int i; + + for (i = 0; i < vec_len (vm->buffer_main->pktmbuf_pools); i++) + { + rmp = vm->buffer_main->pktmbuf_pools[i]; + if (rmp) + { + unsigned count = rte_mempool_avail_count (rmp); + unsigned free_count = rte_mempool_in_use_count (rmp); + + vlib_cli_output (vm, + "name=\"%s\" available = %7d allocated = %7d total = %7d\n", + rmp->name, (u32) count, (u32) free_count, + (u32) (count + free_count)); + } + else + { + vlib_cli_output (vm, "rte_mempool is NULL (!)\n"); + } + } + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_show_dpdk_bufferr,static) = { + .path = "show dpdk buffer", + .short_help = "show dpdk buffer state", + .function = show_dpdk_buffer, + .is_mp_safe = 1, +}; +/* *INDENT-ON* */ + +static clib_error_t * +test_dpdk_buffer (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + static u32 *allocated_buffers; + u32 n_alloc = 0; + u32 n_free = 0; + u32 first, actual_alloc; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "allocate %d", &n_alloc)) + ; + else if (unformat (input, "free %d", &n_free)) + ; + else + break; + } + + if (n_free) + { + if (vec_len (allocated_buffers) < n_free) + return clib_error_return (0, "Can't free %d, only %d allocated", + n_free, vec_len (allocated_buffers)); + + first = vec_len (allocated_buffers) - n_free; + vlib_buffer_free (vm, allocated_buffers + first, n_free); + _vec_len (allocated_buffers) = first; + } + if (n_alloc) + { + first = vec_len (allocated_buffers); + vec_validate (allocated_buffers, + vec_len (allocated_buffers) + n_alloc - 1); + + actual_alloc = vlib_buffer_alloc (vm, allocated_buffers + first, + n_alloc); + _vec_len (allocated_buffers) = first + actual_alloc; + + if (actual_alloc < n_alloc) + vlib_cli_output (vm, "WARNING: only allocated %d buffers", + actual_alloc); + } + + vlib_cli_output (vm, "Currently %d buffers allocated", + vec_len (allocated_buffers)); + + if (allocated_buffers && vec_len (allocated_buffers) == 0) + vec_free (allocated_buffers); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_test_dpdk_buffer,static) = { + .path = "test dpdk buffer", + .short_help = "test dpdk buffer [allocate <nn>][free <nn>]", + .function = test_dpdk_buffer, + .is_mp_safe = 1, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_desc (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 nb_rx_desc = (u32) ~ 0; + u32 nb_tx_desc = (u32) ~ 0; + clib_error_t *rv; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "tx %d", &nb_tx_desc)) + ; + else if (unformat (line_input, "rx %d", &nb_rx_desc)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + if ((xd->flags & DPDK_DEVICE_FLAG_PMD) == 0) + return clib_error_return (0, "number of descriptors can be set only for " + "physical devices"); + + if ((nb_rx_desc == (u32) ~ 0 || nb_rx_desc == xd->nb_rx_desc) && + (nb_tx_desc == (u32) ~ 0 || nb_tx_desc == xd->nb_tx_desc)) + return clib_error_return (0, "nothing changed"); + + if (nb_rx_desc != (u32) ~ 0) + xd->nb_rx_desc = nb_rx_desc; + + if (nb_tx_desc != (u32) ~ 0) + xd->nb_tx_desc = nb_tx_desc; + + rv = dpdk_port_setup (dm, xd); + + return rv; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_desc,static) = { + .path = "set dpdk interface descriptors", + .short_help = "set dpdk interface descriptors <if-name> [rx <n>] [tx <n>]", + .function = set_dpdk_if_desc, +}; +/* *INDENT-ON* */ + +static clib_error_t * +show_dpdk_if_placement (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + dpdk_device_and_queue_t *dq; + int cpu; + + if (tm->n_vlib_mains == 1) + vlib_cli_output (vm, "All interfaces are handled by main thread"); + + for (cpu = 0; cpu < vec_len (dm->devices_by_cpu); cpu++) + { + if (vec_len (dm->devices_by_cpu[cpu])) + vlib_cli_output (vm, "Thread %u (%s at lcore %u):", cpu, + vlib_worker_threads[cpu].name, + vlib_worker_threads[cpu].lcore_id); + + /* *INDENT-OFF* */ + vec_foreach(dq, dm->devices_by_cpu[cpu]) + { + u32 hw_if_index = dm->devices[dq->device].vlib_hw_if_index; + vnet_hw_interface_t * hi = vnet_get_hw_interface(dm->vnet_main, hw_if_index); + vlib_cli_output(vm, " %v queue %u", hi->name, dq->queue_id); + } + /* *INDENT-ON* */ + } + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_show_dpdk_if_placement,static) = { + .path = "show dpdk interface placement", + .short_help = "show dpdk interface placement", + .function = show_dpdk_if_placement, +}; +/* *INDENT-ON* */ + +static int +dpdk_device_queue_sort (void *a1, void *a2) +{ + dpdk_device_and_queue_t *dq1 = a1; + dpdk_device_and_queue_t *dq2 = a2; + + if (dq1->device > dq2->device) + return 1; + else if (dq1->device < dq2->device) + return -1; + else if (dq1->queue_id > dq2->queue_id) + return 1; + else if (dq1->queue_id < dq2->queue_id) + return -1; + else + return 0; +} + +static clib_error_t * +set_dpdk_if_placement (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + dpdk_device_and_queue_t *dq; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 queue = (u32) 0; + u32 cpu = (u32) ~ 0; + int i; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "queue %d", &queue)) + ; + else if (unformat (line_input, "thread %d", &cpu)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + if (cpu < dm->input_cpu_first_index || + cpu >= (dm->input_cpu_first_index + dm->input_cpu_count)) + return clib_error_return (0, "please specify valid thread id"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + for (i = 0; i < vec_len (dm->devices_by_cpu); i++) + { + /* *INDENT-OFF* */ + vec_foreach(dq, dm->devices_by_cpu[i]) + { + if (hw_if_index == dm->devices[dq->device].vlib_hw_if_index && + queue == dq->queue_id) + { + if (cpu == i) /* nothing to do */ + return 0; + + vec_del1(dm->devices_by_cpu[i], dq - dm->devices_by_cpu[i]); + vec_add2(dm->devices_by_cpu[cpu], dq, 1); + dq->queue_id = queue; + dq->device = xd->device_index; + xd->cpu_socket_id_by_queue[queue] = + rte_lcore_to_socket_id(vlib_worker_threads[cpu].lcore_id); + + vec_sort_with_function(dm->devices_by_cpu[i], + dpdk_device_queue_sort); + + vec_sort_with_function(dm->devices_by_cpu[cpu], + dpdk_device_queue_sort); + + if (vec_len(dm->devices_by_cpu[i]) == 0) + vlib_node_set_state (vlib_mains[i], dpdk_input_node.index, + VLIB_NODE_STATE_DISABLED); + + if (vec_len(dm->devices_by_cpu[cpu]) == 1) + vlib_node_set_state (vlib_mains[cpu], dpdk_input_node.index, + VLIB_NODE_STATE_POLLING); + + return 0; + } + } + /* *INDENT-ON* */ + } + + return clib_error_return (0, "not found"); +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_placement,static) = { + .path = "set dpdk interface placement", + .short_help = "set dpdk interface placement <if-name> [queue <n>] thread <n>", + .function = set_dpdk_if_placement, +}; +/* *INDENT-ON* */ + +static clib_error_t * +show_dpdk_if_hqos_placement (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + dpdk_device_and_queue_t *dq; + int cpu; + + if (tm->n_vlib_mains == 1) + vlib_cli_output (vm, "All interfaces are handled by main thread"); + + for (cpu = 0; cpu < vec_len (dm->devices_by_hqos_cpu); cpu++) + { + if (vec_len (dm->devices_by_hqos_cpu[cpu])) + vlib_cli_output (vm, "Thread %u (%s at lcore %u):", cpu, + vlib_worker_threads[cpu].name, + vlib_worker_threads[cpu].lcore_id); + + vec_foreach (dq, dm->devices_by_hqos_cpu[cpu]) + { + u32 hw_if_index = dm->devices[dq->device].vlib_hw_if_index; + vnet_hw_interface_t *hi = + vnet_get_hw_interface (dm->vnet_main, hw_if_index); + vlib_cli_output (vm, " %v queue %u", hi->name, dq->queue_id); + } + } + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_show_dpdk_if_hqos_placement, static) = { + .path = "show dpdk interface hqos placement", + .short_help = "show dpdk interface hqos placement", + .function = show_dpdk_if_hqos_placement, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_hqos_placement (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + dpdk_device_and_queue_t *dq; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 cpu = (u32) ~ 0; + int i; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "thread %d", &cpu)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + if (cpu < dm->hqos_cpu_first_index || + cpu >= (dm->hqos_cpu_first_index + dm->hqos_cpu_count)) + return clib_error_return (0, "please specify valid thread id"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + for (i = 0; i < vec_len (dm->devices_by_hqos_cpu); i++) + { + vec_foreach (dq, dm->devices_by_hqos_cpu[i]) + { + if (hw_if_index == dm->devices[dq->device].vlib_hw_if_index) + { + if (cpu == i) /* nothing to do */ + return 0; + + vec_del1 (dm->devices_by_hqos_cpu[i], + dq - dm->devices_by_hqos_cpu[i]); + vec_add2 (dm->devices_by_hqos_cpu[cpu], dq, 1); + dq->queue_id = 0; + dq->device = xd->device_index; + + vec_sort_with_function (dm->devices_by_hqos_cpu[i], + dpdk_device_queue_sort); + + vec_sort_with_function (dm->devices_by_hqos_cpu[cpu], + dpdk_device_queue_sort); + + return 0; + } + } + } + + return clib_error_return (0, "not found"); +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_hqos_placement, static) = { + .path = "set dpdk interface hqos placement", + .short_help = "set dpdk interface hqos placement <if-name> thread <n>", + .function = set_dpdk_if_hqos_placement, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_hqos_pipe (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 subport_id = (u32) ~ 0; + u32 pipe_id = (u32) ~ 0; + u32 profile_id = (u32) ~ 0; + int rv; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "subport %d", &subport_id)) + ; + else if (unformat (line_input, "pipe %d", &pipe_id)) + ; + else if (unformat (line_input, "profile %d", &profile_id)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + rv = + rte_sched_pipe_config (xd->hqos_ht->hqos, subport_id, pipe_id, + profile_id); + if (rv) + return clib_error_return (0, "pipe configuration failed"); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_hqos_pipe, static) = +{ + .path = "set dpdk interface hqos pipe", + .short_help = "set dpdk interface hqos pipe <if-name> subport <n> pipe <n> " + "profile <n>", + .function = set_dpdk_if_hqos_pipe, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_hqos_subport (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 subport_id = (u32) ~ 0; + struct rte_sched_subport_params p = { + .tb_rate = 1250000000, /* 10GbE */ + .tb_size = 1000000, + .tc_rate = {1250000000, 1250000000, 1250000000, 1250000000}, + .tc_period = 10, + }; + int rv; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "subport %d", &subport_id)) + ; + else if (unformat (line_input, "rate %d", &p.tb_rate)) + { + p.tc_rate[0] = p.tb_rate; + p.tc_rate[1] = p.tb_rate; + p.tc_rate[2] = p.tb_rate; + p.tc_rate[3] = p.tb_rate; + } + else if (unformat (line_input, "bktsize %d", &p.tb_size)) + ; + else if (unformat (line_input, "tc0 %d", &p.tc_rate[0])) + ; + else if (unformat (line_input, "tc1 %d", &p.tc_rate[1])) + ; + else if (unformat (line_input, "tc2 %d", &p.tc_rate[2])) + ; + else if (unformat (line_input, "tc3 %d", &p.tc_rate[3])) + ; + else if (unformat (line_input, "period %d", &p.tc_period)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + rv = rte_sched_subport_config (xd->hqos_ht->hqos, subport_id, &p); + if (rv) + return clib_error_return (0, "subport configuration failed"); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_hqos_subport, static) = { + .path = "set dpdk interface hqos subport", + .short_help = "set dpdk interface hqos subport <if-name> subport <n> " + "[rate <n>] [bktsize <n>] [tc0 <n>] [tc1 <n>] [tc2 <n>] [tc3 <n>] " + "[period <n>]", + .function = set_dpdk_if_hqos_subport, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_hqos_tctbl (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + u32 tc = (u32) ~ 0; + u32 queue = (u32) ~ 0; + u32 entry = (u32) ~ 0; + u32 val, i; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "entry %d", &entry)) + ; + else if (unformat (line_input, "tc %d", &tc)) + ; + else if (unformat (line_input, "queue %d", &queue)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + if (entry >= 64) + return clib_error_return (0, "invalid entry"); + if (tc >= RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE) + return clib_error_return (0, "invalid traffic class"); + if (queue >= RTE_SCHED_QUEUES_PER_TRAFFIC_CLASS) + return clib_error_return (0, "invalid traffic class"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + /* Detect the set of worker threads */ + uword *p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + /* Should never happen, shut up Coverity warning */ + if (p == 0) + return clib_error_return (0, "no worker registrations?"); + + vlib_thread_registration_t *tr = (vlib_thread_registration_t *) p[0]; + int worker_thread_first = tr->first_index; + int worker_thread_count = tr->count; + + val = tc * RTE_SCHED_QUEUES_PER_TRAFFIC_CLASS + queue; + for (i = 0; i < worker_thread_count; i++) + xd->hqos_wt[worker_thread_first + i].hqos_tc_table[entry] = val; + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_hqos_tctbl, static) = { + .path = "set dpdk interface hqos tctbl", + .short_help = "set dpdk interface hqos tctbl <if-name> entry <n> tc <n> queue <n>", + .function = set_dpdk_if_hqos_tctbl, +}; +/* *INDENT-ON* */ + +static clib_error_t * +set_dpdk_if_hqos_pktfield (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + + /* Device specific data */ + struct rte_eth_dev_info dev_info; + dpdk_device_config_t *devconf = 0; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + u32 hw_if_index = (u32) ~ 0; + + /* Detect the set of worker threads */ + uword *p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + /* Should never happen, shut up Coverity warning */ + if (p == 0) + return clib_error_return (0, "no worker registrations?"); + + vlib_thread_registration_t *tr = (vlib_thread_registration_t *) p[0]; + int worker_thread_first = tr->first_index; + int worker_thread_count = tr->count; + + /* Packet field configuration */ + u64 mask = (u64) ~ 0; + u32 id = (u32) ~ 0; + u32 offset = (u32) ~ 0; + + /* HQoS params */ + u32 n_subports_per_port, n_pipes_per_subport, tctbl_size; + + u32 i; + + /* Parse input arguments */ + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else if (unformat (line_input, "id %d", &id)) + ; + else if (unformat (line_input, "offset %d", &offset)) + ; + else if (unformat (line_input, "mask %llx", &mask)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + /* Get interface */ + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify valid interface name"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + rte_eth_dev_info_get (xd->device_index, &dev_info); + if (dev_info.pci_dev) + { /* bonded interface has no pci info */ + vlib_pci_addr_t pci_addr; + + pci_addr.domain = dev_info.pci_dev->addr.domain; + pci_addr.bus = dev_info.pci_dev->addr.bus; + pci_addr.slot = dev_info.pci_dev->addr.devid; + pci_addr.function = dev_info.pci_dev->addr.function; + + p = + hash_get (dm->conf->device_config_index_by_pci_addr, pci_addr.as_u32); + } + + if (p) + devconf = pool_elt_at_index (dm->conf->dev_confs, p[0]); + else + devconf = &dm->conf->default_devconf; + + if (devconf->hqos_enabled == 0) + { + vlib_cli_output (vm, "HQoS disabled for this interface"); + return 0; + } + + n_subports_per_port = devconf->hqos.port.n_subports_per_port; + n_pipes_per_subport = devconf->hqos.port.n_pipes_per_subport; + tctbl_size = RTE_DIM (devconf->hqos.tc_table); + + /* Validate packet field configuration: id, offset and mask */ + if (id >= 3) + return clib_error_return (0, "invalid packet field id"); + + switch (id) + { + case 0: + if (dpdk_hqos_validate_mask (mask, n_subports_per_port) != 0) + return clib_error_return (0, "invalid subport ID mask " + "(n_subports_per_port = %u)", + n_subports_per_port); + break; + case 1: + if (dpdk_hqos_validate_mask (mask, n_pipes_per_subport) != 0) + return clib_error_return (0, "invalid pipe ID mask " + "(n_pipes_per_subport = %u)", + n_pipes_per_subport); + break; + case 2: + default: + if (dpdk_hqos_validate_mask (mask, tctbl_size) != 0) + return clib_error_return (0, "invalid TC table index mask " + "(TC table size = %u)", tctbl_size); + } + + /* Propagate packet field configuration to all workers */ + for (i = 0; i < worker_thread_count; i++) + switch (id) + { + case 0: + xd->hqos_wt[worker_thread_first + i].hqos_field0_slabpos = offset; + xd->hqos_wt[worker_thread_first + i].hqos_field0_slabmask = mask; + xd->hqos_wt[worker_thread_first + i].hqos_field0_slabshr = + __builtin_ctzll (mask); + break; + case 1: + xd->hqos_wt[worker_thread_first + i].hqos_field1_slabpos = offset; + xd->hqos_wt[worker_thread_first + i].hqos_field1_slabmask = mask; + xd->hqos_wt[worker_thread_first + i].hqos_field1_slabshr = + __builtin_ctzll (mask); + break; + case 2: + default: + xd->hqos_wt[worker_thread_first + i].hqos_field2_slabpos = offset; + xd->hqos_wt[worker_thread_first + i].hqos_field2_slabmask = mask; + xd->hqos_wt[worker_thread_first + i].hqos_field2_slabshr = + __builtin_ctzll (mask); + } + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_set_dpdk_if_hqos_pktfield, static) = { + .path = "set dpdk interface hqos pktfield", + .short_help = "set dpdk interface hqos pktfield <if-name> id <n> offset <n> " + "mask <n>", + .function = set_dpdk_if_hqos_pktfield, +}; +/* *INDENT-ON* */ + +static clib_error_t * +show_dpdk_if_hqos (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + dpdk_device_config_hqos_t *cfg; + dpdk_device_hqos_per_hqos_thread_t *ht; + dpdk_device_hqos_per_worker_thread_t *wk; + u32 *tctbl; + u32 hw_if_index = (u32) ~ 0; + u32 profile_id, i; + struct rte_eth_dev_info dev_info; + dpdk_device_config_t *devconf = 0; + vlib_thread_registration_t *tr; + uword *p = 0; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify interface name!!"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + rte_eth_dev_info_get (xd->device_index, &dev_info); + if (dev_info.pci_dev) + { /* bonded interface has no pci info */ + vlib_pci_addr_t pci_addr; + + pci_addr.domain = dev_info.pci_dev->addr.domain; + pci_addr.bus = dev_info.pci_dev->addr.bus; + pci_addr.slot = dev_info.pci_dev->addr.devid; + pci_addr.function = dev_info.pci_dev->addr.function; + + p = + hash_get (dm->conf->device_config_index_by_pci_addr, pci_addr.as_u32); + } + + if (p) + devconf = pool_elt_at_index (dm->conf->dev_confs, p[0]); + else + devconf = &dm->conf->default_devconf; + + if (devconf->hqos_enabled == 0) + { + vlib_cli_output (vm, "HQoS disabled for this interface"); + return 0; + } + + /* Detect the set of worker threads */ + p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + + /* Should never happen, shut up Coverity warning */ + if (p == 0) + return clib_error_return (0, "no worker registrations?"); + + tr = (vlib_thread_registration_t *) p[0]; + + cfg = &devconf->hqos; + ht = xd->hqos_ht; + wk = &xd->hqos_wt[tr->first_index]; + tctbl = wk->hqos_tc_table; + + vlib_cli_output (vm, " Thread:"); + vlib_cli_output (vm, " Input SWQ size = %u packets", cfg->swq_size); + vlib_cli_output (vm, " Enqueue burst size = %u packets", + ht->hqos_burst_enq); + vlib_cli_output (vm, " Dequeue burst size = %u packets", + ht->hqos_burst_deq); + + vlib_cli_output (vm, + " Packet field 0: slab position = %4u, slab bitmask = 0x%016llx", + wk->hqos_field0_slabpos, wk->hqos_field0_slabmask); + vlib_cli_output (vm, + " Packet field 1: slab position = %4u, slab bitmask = 0x%016llx", + wk->hqos_field1_slabpos, wk->hqos_field1_slabmask); + vlib_cli_output (vm, + " Packet field 2: slab position = %4u, slab bitmask = 0x%016llx", + wk->hqos_field2_slabpos, wk->hqos_field2_slabmask); + vlib_cli_output (vm, " Packet field 2 translation table:"); + vlib_cli_output (vm, " [ 0 .. 15]: " + "%2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u", + tctbl[0], tctbl[1], tctbl[2], tctbl[3], + tctbl[4], tctbl[5], tctbl[6], tctbl[7], + tctbl[8], tctbl[9], tctbl[10], tctbl[11], + tctbl[12], tctbl[13], tctbl[14], tctbl[15]); + vlib_cli_output (vm, " [16 .. 31]: " + "%2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u", + tctbl[16], tctbl[17], tctbl[18], tctbl[19], + tctbl[20], tctbl[21], tctbl[22], tctbl[23], + tctbl[24], tctbl[25], tctbl[26], tctbl[27], + tctbl[28], tctbl[29], tctbl[30], tctbl[31]); + vlib_cli_output (vm, " [32 .. 47]: " + "%2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u", + tctbl[32], tctbl[33], tctbl[34], tctbl[35], + tctbl[36], tctbl[37], tctbl[38], tctbl[39], + tctbl[40], tctbl[41], tctbl[42], tctbl[43], + tctbl[44], tctbl[45], tctbl[46], tctbl[47]); + vlib_cli_output (vm, " [48 .. 63]: " + "%2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u %2u", + tctbl[48], tctbl[49], tctbl[50], tctbl[51], + tctbl[52], tctbl[53], tctbl[54], tctbl[55], + tctbl[56], tctbl[57], tctbl[58], tctbl[59], + tctbl[60], tctbl[61], tctbl[62], tctbl[63]); + + vlib_cli_output (vm, " Port:"); + vlib_cli_output (vm, " Rate = %u bytes/second", cfg->port.rate); + vlib_cli_output (vm, " MTU = %u bytes", cfg->port.mtu); + vlib_cli_output (vm, " Frame overhead = %u bytes", + cfg->port.frame_overhead); + vlib_cli_output (vm, " Number of subports = %u", + cfg->port.n_subports_per_port); + vlib_cli_output (vm, " Number of pipes per subport = %u", + cfg->port.n_pipes_per_subport); + vlib_cli_output (vm, + " Packet queue size: TC0 = %u, TC1 = %u, TC2 = %u, TC3 = %u packets", + cfg->port.qsize[0], cfg->port.qsize[1], cfg->port.qsize[2], + cfg->port.qsize[3]); + vlib_cli_output (vm, " Number of pipe profiles = %u", + cfg->port.n_pipe_profiles); + + for (profile_id = 0; profile_id < vec_len (cfg->pipe); profile_id++) + { + vlib_cli_output (vm, " Pipe profile %u:", profile_id); + vlib_cli_output (vm, " Rate = %u bytes/second", + cfg->pipe[profile_id].tb_rate); + vlib_cli_output (vm, " Token bucket size = %u bytes", + cfg->pipe[profile_id].tb_size); + vlib_cli_output (vm, + " Traffic class rate: TC0 = %u, TC1 = %u, TC2 = %u, TC3 = %u bytes/second", + cfg->pipe[profile_id].tc_rate[0], + cfg->pipe[profile_id].tc_rate[1], + cfg->pipe[profile_id].tc_rate[2], + cfg->pipe[profile_id].tc_rate[3]); + vlib_cli_output (vm, " TC period = %u milliseconds", + cfg->pipe[profile_id].tc_period); +#ifdef RTE_SCHED_SUBPORT_TC_OV + vlib_cli_output (vm, " TC3 oversubscription_weight = %u", + cfg->pipe[profile_id].tc_ov_weight); +#endif + + for (i = 0; i < RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE; i++) + { + vlib_cli_output (vm, + " TC%u WRR weights: Q0 = %u, Q1 = %u, Q2 = %u, Q3 = %u", + i, cfg->pipe[profile_id].wrr_weights[i * 4], + cfg->pipe[profile_id].wrr_weights[i * 4 + 1], + cfg->pipe[profile_id].wrr_weights[i * 4 + 2], + cfg->pipe[profile_id].wrr_weights[i * 4 + 3]); + } + } + +#ifdef RTE_SCHED_RED + vlib_cli_output (vm, " Weighted Random Early Detection (WRED):"); + for (i = 0; i < RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE; i++) + { + vlib_cli_output (vm, " TC%u min: G = %u, Y = %u, R = %u", i, + cfg->port.red_params[i][e_RTE_METER_GREEN].min_th, + cfg->port.red_params[i][e_RTE_METER_YELLOW].min_th, + cfg->port.red_params[i][e_RTE_METER_RED].min_th); + + vlib_cli_output (vm, " TC%u max: G = %u, Y = %u, R = %u", i, + cfg->port.red_params[i][e_RTE_METER_GREEN].max_th, + cfg->port.red_params[i][e_RTE_METER_YELLOW].max_th, + cfg->port.red_params[i][e_RTE_METER_RED].max_th); + + vlib_cli_output (vm, + " TC%u inverted probability: G = %u, Y = %u, R = %u", + i, cfg->port.red_params[i][e_RTE_METER_GREEN].maxp_inv, + cfg->port.red_params[i][e_RTE_METER_YELLOW].maxp_inv, + cfg->port.red_params[i][e_RTE_METER_RED].maxp_inv); + + vlib_cli_output (vm, " TC%u weight: R = %u, Y = %u, R = %u", i, + cfg->port.red_params[i][e_RTE_METER_GREEN].wq_log2, + cfg->port.red_params[i][e_RTE_METER_YELLOW].wq_log2, + cfg->port.red_params[i][e_RTE_METER_RED].wq_log2); + } +#endif + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_show_dpdk_if_hqos, static) = { + .path = "show dpdk interface hqos", + .short_help = "show dpdk interface hqos <if-name>", + .function = show_dpdk_if_hqos, +}; + +/* *INDENT-ON* */ + +static clib_error_t * +show_dpdk_hqos_queue_stats (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + dpdk_main_t *dm = &dpdk_main; + u32 hw_if_index = (u32) ~ 0; + u32 subport = (u32) ~ 0; + u32 pipe = (u32) ~ 0; + u32 tc = (u32) ~ 0; + u32 tc_q = (u32) ~ 0; + vnet_hw_interface_t *hw; + dpdk_device_t *xd; + uword *p = 0; + struct rte_eth_dev_info dev_info; + dpdk_device_config_t *devconf = 0; + u32 qindex; + struct rte_sched_queue_stats stats; + u16 qlen; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (line_input, "%U", unformat_vnet_hw_interface, dm->vnet_main, + &hw_if_index)) + ; + + else if (unformat (line_input, "subport %d", &subport)) + ; + + else if (unformat (line_input, "pipe %d", &pipe)) + ; + + else if (unformat (line_input, "tc %d", &tc)) + ; + + else if (unformat (line_input, "tc_q %d", &tc_q)) + ; + + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + if (hw_if_index == (u32) ~ 0) + return clib_error_return (0, "please specify interface name!!"); + + hw = vnet_get_hw_interface (dm->vnet_main, hw_if_index); + xd = vec_elt_at_index (dm->devices, hw->dev_instance); + + rte_eth_dev_info_get (xd->device_index, &dev_info); + if (dev_info.pci_dev) + { /* bonded interface has no pci info */ + vlib_pci_addr_t pci_addr; + + pci_addr.domain = dev_info.pci_dev->addr.domain; + pci_addr.bus = dev_info.pci_dev->addr.bus; + pci_addr.slot = dev_info.pci_dev->addr.devid; + pci_addr.function = dev_info.pci_dev->addr.function; + + p = + hash_get (dm->conf->device_config_index_by_pci_addr, pci_addr.as_u32); + } + + if (p) + devconf = pool_elt_at_index (dm->conf->dev_confs, p[0]); + else + devconf = &dm->conf->default_devconf; + + if (devconf->hqos_enabled == 0) + { + vlib_cli_output (vm, "HQoS disabled for this interface"); + return 0; + } + + /* + * Figure out which queue to query. cf rte_sched_port_qindex. (Not sure why + * that method isn't made public by DPDK - how _should_ we get the queue ID?) + */ + qindex = subport * devconf->hqos.port.n_pipes_per_subport + pipe; + qindex = qindex * RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE + tc; + qindex = qindex * RTE_SCHED_QUEUES_PER_TRAFFIC_CLASS + tc_q; + + if (rte_sched_queue_read_stats (xd->hqos_ht->hqos, qindex, &stats, &qlen) != + 0) + return clib_error_return (0, "failed to read stats"); + + vlib_cli_output (vm, "%=24s%=16s", "Stats Parameter", "Value"); + vlib_cli_output (vm, "%=24s%=16d", "Packets", stats.n_pkts); + vlib_cli_output (vm, "%=24s%=16d", "Packets dropped", stats.n_pkts_dropped); +#ifdef RTE_SCHED_RED + vlib_cli_output (vm, "%=24s%=16d", "Packets dropped (RED)", + stats.n_pkts_red_dropped); +#endif + vlib_cli_output (vm, "%=24s%=16d", "Bytes", stats.n_bytes); + vlib_cli_output (vm, "%=24s%=16d", "Bytes dropped", stats.n_bytes_dropped); + + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cmd_show_dpdk_hqos_queue_stats, static) = { + .path = "show dpdk hqos queue", + .short_help = "show dpdk hqos queue <if-name> subport <subport> pipe <pipe> tc <tc> tc_q <tc_q>", + .function = show_dpdk_hqos_queue_stats, +}; +/* *INDENT-ON* */ + +clib_error_t * +dpdk_cli_init (vlib_main_t * vm) +{ + return 0; +} + +VLIB_INIT_FUNCTION (dpdk_cli_init); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/device.c b/src/vnet/devices/dpdk/device.c new file mode 100644 index 00000000000..b22fbf2e69e --- /dev/null +++ b/src/vnet/devices/dpdk/device.c @@ -0,0 +1,840 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include <vnet/vnet.h> +#include <vppinfra/vec.h> +#include <vppinfra/format.h> +#include <vlib/unix/cj.h> +#include <assert.h> + +#include <vnet/ethernet/ethernet.h> +#include <vnet/devices/dpdk/dpdk.h> + +#include "dpdk_priv.h" +#include <vppinfra/error.h> + +#define foreach_dpdk_tx_func_error \ + _(BAD_RETVAL, "DPDK tx function returned an error") \ + _(RING_FULL, "Tx packet drops (ring full)") \ + _(PKT_DROP, "Tx packet drops (dpdk tx failure)") \ + _(REPL_FAIL, "Tx packet drops (replication failure)") + +typedef enum +{ +#define _(f,s) DPDK_TX_FUNC_ERROR_##f, + foreach_dpdk_tx_func_error +#undef _ + DPDK_TX_FUNC_N_ERROR, +} dpdk_tx_func_error_t; + +static char *dpdk_tx_func_error_strings[] = { +#define _(n,s) s, + foreach_dpdk_tx_func_error +#undef _ +}; + +clib_error_t * +dpdk_set_mac_address (vnet_hw_interface_t * hi, char *address) +{ + int error; + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, hi->dev_instance); + + error = rte_eth_dev_default_mac_addr_set (xd->device_index, + (struct ether_addr *) address); + + if (error) + { + return clib_error_return (0, "mac address set failed: %d", error); + } + else + { + return NULL; + } +} + +clib_error_t * +dpdk_set_mc_filter (vnet_hw_interface_t * hi, + struct ether_addr mc_addr_vec[], int naddr) +{ + int error; + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, hi->dev_instance); + + error = rte_eth_dev_set_mc_addr_list (xd->device_index, mc_addr_vec, naddr); + + if (error) + { + return clib_error_return (0, "mc addr list failed: %d", error); + } + else + { + return NULL; + } +} + +struct rte_mbuf * +dpdk_replicate_packet_mb (vlib_buffer_t * b) +{ + vlib_main_t *vm = vlib_get_main (); + vlib_buffer_main_t *bm = vm->buffer_main; + struct rte_mbuf **mbufs = 0, *s, *d; + u8 nb_segs; + unsigned socket_id = rte_socket_id (); + int i; + + ASSERT (bm->pktmbuf_pools[socket_id]); + s = rte_mbuf_from_vlib_buffer (b); + nb_segs = s->nb_segs; + vec_validate (mbufs, nb_segs - 1); + + if (rte_pktmbuf_alloc_bulk (bm->pktmbuf_pools[socket_id], mbufs, nb_segs)) + { + vec_free (mbufs); + return 0; + } + + d = mbufs[0]; + d->nb_segs = s->nb_segs; + d->data_len = s->data_len; + d->pkt_len = s->pkt_len; + d->data_off = s->data_off; + clib_memcpy (d->buf_addr, s->buf_addr, RTE_PKTMBUF_HEADROOM + s->data_len); + + for (i = 1; i < nb_segs; i++) + { + d->next = mbufs[i]; + d = mbufs[i]; + s = s->next; + d->data_len = s->data_len; + clib_memcpy (d->buf_addr, s->buf_addr, + RTE_PKTMBUF_HEADROOM + s->data_len); + } + + d = mbufs[0]; + vec_free (mbufs); + return d; +} + +static void +dpdk_tx_trace_buffer (dpdk_main_t * dm, + vlib_node_runtime_t * node, + dpdk_device_t * xd, + u16 queue_id, u32 buffer_index, vlib_buffer_t * buffer) +{ + vlib_main_t *vm = vlib_get_main (); + dpdk_tx_dma_trace_t *t0; + struct rte_mbuf *mb; + + mb = rte_mbuf_from_vlib_buffer (buffer); + + t0 = vlib_add_trace (vm, node, buffer, sizeof (t0[0])); + t0->queue_index = queue_id; + t0->device_index = xd->device_index; + t0->buffer_index = buffer_index; + clib_memcpy (&t0->mb, mb, sizeof (t0->mb)); + clib_memcpy (&t0->buffer, buffer, + sizeof (buffer[0]) - sizeof (buffer->pre_data)); + clib_memcpy (t0->buffer.pre_data, buffer->data + buffer->current_data, + sizeof (t0->buffer.pre_data)); +} + +static_always_inline void +dpdk_validate_rte_mbuf (vlib_main_t * vm, vlib_buffer_t * b, + int maybe_multiseg) +{ + struct rte_mbuf *mb, *first_mb, *last_mb; + + /* buffer is coming from non-dpdk source so we need to init + rte_mbuf header */ + if (PREDICT_FALSE ((b->flags & VNET_BUFFER_RTE_MBUF_VALID) == 0)) + { + vlib_buffer_t *b2 = b; + last_mb = mb = rte_mbuf_from_vlib_buffer (b2); + rte_pktmbuf_reset (mb); + while (maybe_multiseg && (b2->flags & VLIB_BUFFER_NEXT_PRESENT)) + { + b2 = vlib_get_buffer (vm, b2->next_buffer); + mb = rte_mbuf_from_vlib_buffer (b2); + last_mb->next = mb; + last_mb = mb; + rte_pktmbuf_reset (mb); + } + } + + first_mb = mb = rte_mbuf_from_vlib_buffer (b); + first_mb->nb_segs = 1; + mb->data_len = b->current_length; + mb->pkt_len = maybe_multiseg ? vlib_buffer_length_in_chain (vm, b) : + b->current_length; + mb->data_off = VLIB_BUFFER_PRE_DATA_SIZE + b->current_data; + + while (maybe_multiseg && (b->flags & VLIB_BUFFER_NEXT_PRESENT)) + { + b = vlib_get_buffer (vm, b->next_buffer); + mb = rte_mbuf_from_vlib_buffer (b); + mb->data_len = b->current_length; + mb->pkt_len = b->current_length; + mb->data_off = VLIB_BUFFER_PRE_DATA_SIZE + b->current_data; + first_mb->nb_segs++; + } +} + +/* + * This function calls the dpdk's tx_burst function to transmit the packets + * on the tx_vector. It manages a lock per-device if the device does not + * support multiple queues. It returns the number of packets untransmitted + * on the tx_vector. If all packets are transmitted (the normal case), the + * function returns 0. + * + * The function assumes there is at least one packet on the tx_vector. + */ +static_always_inline + u32 tx_burst_vector_internal (vlib_main_t * vm, + dpdk_device_t * xd, + struct rte_mbuf **tx_vector) +{ + dpdk_main_t *dm = &dpdk_main; + u32 n_packets; + u32 tx_head; + u32 tx_tail; + u32 n_retry; + int rv; + int queue_id; + tx_ring_hdr_t *ring; + + ring = vec_header (tx_vector, sizeof (*ring)); + + n_packets = ring->tx_head - ring->tx_tail; + + tx_head = ring->tx_head % xd->nb_tx_desc; + + /* + * Ensure rte_eth_tx_burst is not called with 0 packets, which can lead to + * unpredictable results. + */ + ASSERT (n_packets > 0); + + /* + * Check for tx_vector overflow. If this fails it is a system configuration + * error. The ring should be sized big enough to handle the largest un-flowed + * off burst from a traffic manager. A larger size also helps performance + * a bit because it decreases the probability of having to issue two tx_burst + * calls due to a ring wrap. + */ + ASSERT (n_packets < xd->nb_tx_desc); + ASSERT (ring->tx_tail == 0); + + n_retry = 16; + queue_id = vm->cpu_index; + + do + { + /* start the burst at the tail */ + tx_tail = ring->tx_tail % xd->nb_tx_desc; + + /* + * This device only supports one TX queue, + * and we're running multi-threaded... + */ + if (PREDICT_FALSE (xd->lockp != 0)) + { + queue_id = queue_id % xd->tx_q_used; + while (__sync_lock_test_and_set (xd->lockp[queue_id], 1)) + /* zzzz */ + queue_id = (queue_id + 1) % xd->tx_q_used; + } + + if (PREDICT_FALSE (xd->flags & DPDK_DEVICE_FLAG_HQOS)) /* HQoS ON */ + { + /* no wrap, transmit in one burst */ + dpdk_device_hqos_per_worker_thread_t *hqos = + &xd->hqos_wt[vm->cpu_index]; + + ASSERT (hqos->swq != NULL); + + dpdk_hqos_metadata_set (hqos, + &tx_vector[tx_tail], tx_head - tx_tail); + rv = rte_ring_sp_enqueue_burst (hqos->swq, + (void **) &tx_vector[tx_tail], + (uint16_t) (tx_head - tx_tail)); + } + else if (PREDICT_TRUE (xd->flags & DPDK_DEVICE_FLAG_PMD)) + { + /* no wrap, transmit in one burst */ + rv = rte_eth_tx_burst (xd->device_index, + (uint16_t) queue_id, + &tx_vector[tx_tail], + (uint16_t) (tx_head - tx_tail)); + } + else + { + ASSERT (0); + rv = 0; + } + + if (PREDICT_FALSE (xd->lockp != 0)) + *xd->lockp[queue_id] = 0; + + if (PREDICT_FALSE (rv < 0)) + { + // emit non-fatal message, bump counter + vnet_main_t *vnm = dm->vnet_main; + vnet_interface_main_t *im = &vnm->interface_main; + u32 node_index; + + node_index = vec_elt_at_index (im->hw_interfaces, + xd->vlib_hw_if_index)->tx_node_index; + + vlib_error_count (vm, node_index, DPDK_TX_FUNC_ERROR_BAD_RETVAL, 1); + clib_warning ("rte_eth_tx_burst[%d]: error %d", xd->device_index, + rv); + return n_packets; // untransmitted packets + } + ring->tx_tail += (u16) rv; + n_packets -= (uint16_t) rv; + } + while (rv && n_packets && (n_retry > 0)); + + return n_packets; +} + +static_always_inline void +dpdk_prefetch_buffer_by_index (vlib_main_t * vm, u32 bi) +{ + vlib_buffer_t *b; + struct rte_mbuf *mb; + b = vlib_get_buffer (vm, bi); + mb = rte_mbuf_from_vlib_buffer (b); + CLIB_PREFETCH (mb, CLIB_CACHE_LINE_BYTES, LOAD); + CLIB_PREFETCH (b, CLIB_CACHE_LINE_BYTES, LOAD); +} + +static_always_inline void +dpdk_buffer_recycle (vlib_main_t * vm, vlib_node_runtime_t * node, + vlib_buffer_t * b, u32 bi, struct rte_mbuf **mbp) +{ + dpdk_main_t *dm = &dpdk_main; + u32 my_cpu = vm->cpu_index; + struct rte_mbuf *mb_new; + + if (PREDICT_FALSE (b->flags & VLIB_BUFFER_RECYCLE) == 0) + return; + + mb_new = dpdk_replicate_packet_mb (b); + if (PREDICT_FALSE (mb_new == 0)) + { + vlib_error_count (vm, node->node_index, + DPDK_TX_FUNC_ERROR_REPL_FAIL, 1); + b->flags |= VLIB_BUFFER_REPL_FAIL; + } + else + *mbp = mb_new; + + vec_add1 (dm->recycle[my_cpu], bi); +} + +/* + * Transmits the packets on the frame to the interface associated with the + * node. It first copies packets on the frame to a tx_vector containing the + * rte_mbuf pointers. It then passes this vector to tx_burst_vector_internal + * which calls the dpdk tx_burst function. + */ +static uword +dpdk_interface_tx (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * f) +{ + dpdk_main_t *dm = &dpdk_main; + vnet_interface_output_runtime_t *rd = (void *) node->runtime_data; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, rd->dev_instance); + u32 n_packets = f->n_vectors; + u32 n_left; + u32 *from; + struct rte_mbuf **tx_vector; + u16 i; + u16 nb_tx_desc = xd->nb_tx_desc; + int queue_id; + u32 my_cpu; + u32 tx_pkts = 0; + tx_ring_hdr_t *ring; + u32 n_on_ring; + + my_cpu = vm->cpu_index; + + queue_id = my_cpu; + + tx_vector = xd->tx_vectors[queue_id]; + ring = vec_header (tx_vector, sizeof (*ring)); + + n_on_ring = ring->tx_head - ring->tx_tail; + from = vlib_frame_vector_args (f); + + ASSERT (n_packets <= VLIB_FRAME_SIZE); + + if (PREDICT_FALSE (n_on_ring + n_packets > nb_tx_desc)) + { + /* + * Overflowing the ring should never happen. + * If it does then drop the whole frame. + */ + vlib_error_count (vm, node->node_index, DPDK_TX_FUNC_ERROR_RING_FULL, + n_packets); + + while (n_packets--) + { + u32 bi0 = from[n_packets]; + vlib_buffer_t *b0 = vlib_get_buffer (vm, bi0); + struct rte_mbuf *mb0 = rte_mbuf_from_vlib_buffer (b0); + rte_pktmbuf_free (mb0); + } + return n_on_ring; + } + + if (PREDICT_FALSE (dm->tx_pcap_enable)) + { + n_left = n_packets; + while (n_left > 0) + { + u32 bi0 = from[0]; + vlib_buffer_t *b0 = vlib_get_buffer (vm, bi0); + if (dm->pcap_sw_if_index == 0 || + dm->pcap_sw_if_index == vnet_buffer (b0)->sw_if_index[VLIB_TX]) + pcap_add_buffer (&dm->pcap_main, vm, bi0, 512); + from++; + n_left--; + } + } + + from = vlib_frame_vector_args (f); + n_left = n_packets; + i = ring->tx_head % nb_tx_desc; + + while (n_left >= 8) + { + u32 bi0, bi1, bi2, bi3; + struct rte_mbuf *mb0, *mb1, *mb2, *mb3; + vlib_buffer_t *b0, *b1, *b2, *b3; + u32 or_flags; + + dpdk_prefetch_buffer_by_index (vm, from[4]); + dpdk_prefetch_buffer_by_index (vm, from[5]); + dpdk_prefetch_buffer_by_index (vm, from[6]); + dpdk_prefetch_buffer_by_index (vm, from[7]); + + bi0 = from[0]; + bi1 = from[1]; + bi2 = from[2]; + bi3 = from[3]; + from += 4; + + b0 = vlib_get_buffer (vm, bi0); + b1 = vlib_get_buffer (vm, bi1); + b2 = vlib_get_buffer (vm, bi2); + b3 = vlib_get_buffer (vm, bi3); + + or_flags = b0->flags | b1->flags | b2->flags | b3->flags; + + if (or_flags & VLIB_BUFFER_NEXT_PRESENT) + { + dpdk_validate_rte_mbuf (vm, b0, 1); + dpdk_validate_rte_mbuf (vm, b1, 1); + dpdk_validate_rte_mbuf (vm, b2, 1); + dpdk_validate_rte_mbuf (vm, b3, 1); + } + else + { + dpdk_validate_rte_mbuf (vm, b0, 0); + dpdk_validate_rte_mbuf (vm, b1, 0); + dpdk_validate_rte_mbuf (vm, b2, 0); + dpdk_validate_rte_mbuf (vm, b3, 0); + } + + mb0 = rte_mbuf_from_vlib_buffer (b0); + mb1 = rte_mbuf_from_vlib_buffer (b1); + mb2 = rte_mbuf_from_vlib_buffer (b2); + mb3 = rte_mbuf_from_vlib_buffer (b3); + + if (PREDICT_FALSE (or_flags & VLIB_BUFFER_RECYCLE)) + { + dpdk_buffer_recycle (vm, node, b0, bi0, &mb0); + dpdk_buffer_recycle (vm, node, b1, bi1, &mb1); + dpdk_buffer_recycle (vm, node, b2, bi2, &mb2); + dpdk_buffer_recycle (vm, node, b3, bi3, &mb3); + + /* dont enqueue packets if replication failed as they must + be sent back to recycle */ + if (PREDICT_TRUE ((b0->flags & VLIB_BUFFER_REPL_FAIL) == 0)) + tx_vector[i++ % nb_tx_desc] = mb0; + if (PREDICT_TRUE ((b1->flags & VLIB_BUFFER_REPL_FAIL) == 0)) + tx_vector[i++ % nb_tx_desc] = mb1; + if (PREDICT_TRUE ((b2->flags & VLIB_BUFFER_REPL_FAIL) == 0)) + tx_vector[i++ % nb_tx_desc] = mb2; + if (PREDICT_TRUE ((b3->flags & VLIB_BUFFER_REPL_FAIL) == 0)) + tx_vector[i++ % nb_tx_desc] = mb3; + } + else + { + if (PREDICT_FALSE (i + 3 >= nb_tx_desc)) + { + tx_vector[i++ % nb_tx_desc] = mb0; + tx_vector[i++ % nb_tx_desc] = mb1; + tx_vector[i++ % nb_tx_desc] = mb2; + tx_vector[i++ % nb_tx_desc] = mb3; + i %= nb_tx_desc; + } + else + { + tx_vector[i++] = mb0; + tx_vector[i++] = mb1; + tx_vector[i++] = mb2; + tx_vector[i++] = mb3; + } + } + + + if (PREDICT_FALSE (node->flags & VLIB_NODE_FLAG_TRACE)) + { + if (b0->flags & VLIB_BUFFER_IS_TRACED) + dpdk_tx_trace_buffer (dm, node, xd, queue_id, bi0, b0); + if (b1->flags & VLIB_BUFFER_IS_TRACED) + dpdk_tx_trace_buffer (dm, node, xd, queue_id, bi1, b1); + if (b2->flags & VLIB_BUFFER_IS_TRACED) + dpdk_tx_trace_buffer (dm, node, xd, queue_id, bi2, b2); + if (b3->flags & VLIB_BUFFER_IS_TRACED) + dpdk_tx_trace_buffer (dm, node, xd, queue_id, bi3, b3); + } + + n_left -= 4; + } + while (n_left > 0) + { + u32 bi0; + struct rte_mbuf *mb0; + vlib_buffer_t *b0; + + bi0 = from[0]; + from++; + + b0 = vlib_get_buffer (vm, bi0); + + dpdk_validate_rte_mbuf (vm, b0, 1); + + mb0 = rte_mbuf_from_vlib_buffer (b0); + dpdk_buffer_recycle (vm, node, b0, bi0, &mb0); + + if (PREDICT_FALSE (node->flags & VLIB_NODE_FLAG_TRACE)) + if (b0->flags & VLIB_BUFFER_IS_TRACED) + dpdk_tx_trace_buffer (dm, node, xd, queue_id, bi0, b0); + + if (PREDICT_TRUE ((b0->flags & VLIB_BUFFER_REPL_FAIL) == 0)) + { + tx_vector[i % nb_tx_desc] = mb0; + i++; + } + n_left--; + } + + /* account for additional packets in the ring */ + ring->tx_head += n_packets; + n_on_ring = ring->tx_head - ring->tx_tail; + + /* transmit as many packets as possible */ + n_packets = tx_burst_vector_internal (vm, xd, tx_vector); + + /* + * tx_pkts is the number of packets successfully transmitted + * This is the number originally on ring minus the number remaining on ring + */ + tx_pkts = n_on_ring - n_packets; + + { + /* If there is no callback then drop any non-transmitted packets */ + if (PREDICT_FALSE (n_packets)) + { + vlib_simple_counter_main_t *cm; + vnet_main_t *vnm = vnet_get_main (); + + cm = vec_elt_at_index (vnm->interface_main.sw_if_counters, + VNET_INTERFACE_COUNTER_TX_ERROR); + + vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index, + n_packets); + + vlib_error_count (vm, node->node_index, DPDK_TX_FUNC_ERROR_PKT_DROP, + n_packets); + + while (n_packets--) + rte_pktmbuf_free (tx_vector[ring->tx_tail + n_packets]); + } + + /* Reset head/tail to avoid unnecessary wrap */ + ring->tx_head = 0; + ring->tx_tail = 0; + } + + /* Recycle replicated buffers */ + if (PREDICT_FALSE (vec_len (dm->recycle[my_cpu]))) + { + vlib_buffer_free (vm, dm->recycle[my_cpu], + vec_len (dm->recycle[my_cpu])); + _vec_len (dm->recycle[my_cpu]) = 0; + } + + ASSERT (ring->tx_head >= ring->tx_tail); + + return tx_pkts; +} + +static void +dpdk_clear_hw_interface_counters (u32 instance) +{ + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, instance); + + /* + * Set the "last_cleared_stats" to the current stats, so that + * things appear to clear from a display perspective. + */ + dpdk_update_counters (xd, vlib_time_now (dm->vlib_main)); + + clib_memcpy (&xd->last_cleared_stats, &xd->stats, sizeof (xd->stats)); + clib_memcpy (xd->last_cleared_xstats, xd->xstats, + vec_len (xd->last_cleared_xstats) * + sizeof (xd->last_cleared_xstats[0])); + +} + +static clib_error_t * +dpdk_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, u32 flags) +{ + vnet_hw_interface_t *hif = vnet_get_hw_interface (vnm, hw_if_index); + uword is_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0; + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, hif->dev_instance); + int rv = 0; + + if (is_up) + { + f64 now = vlib_time_now (dm->vlib_main); + + if ((xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) == 0) + rv = rte_eth_dev_start (xd->device_index); + + if (xd->flags & DPDK_DEVICE_FLAG_PROMISC) + rte_eth_promiscuous_enable (xd->device_index); + else + rte_eth_promiscuous_disable (xd->device_index); + + rte_eth_allmulticast_enable (xd->device_index); + xd->flags |= DPDK_DEVICE_FLAG_ADMIN_UP; + dpdk_update_counters (xd, now); + dpdk_update_link_state (xd, now); + } + else + { + xd->flags &= ~DPDK_DEVICE_FLAG_ADMIN_UP; + + rte_eth_allmulticast_disable (xd->device_index); + vnet_hw_interface_set_flags (vnm, xd->vlib_hw_if_index, 0); + rte_eth_dev_stop (xd->device_index); + + /* For bonded interface, stop slave links */ + if (xd->pmd == VNET_DPDK_PMD_BOND) + { + u8 slink[16]; + int nlink = rte_eth_bond_slaves_get (xd->device_index, slink, 16); + while (nlink >= 1) + { + u8 dpdk_port = slink[--nlink]; + rte_eth_dev_stop (dpdk_port); + } + } + } + + if (rv < 0) + clib_warning ("rte_eth_dev_%s error: %d", is_up ? "start" : "stop", rv); + + return /* no error */ 0; +} + +/* + * Dynamically redirect all pkts from a specific interface + * to the specified node + */ +static void +dpdk_set_interface_next_node (vnet_main_t * vnm, u32 hw_if_index, + u32 node_index) +{ + dpdk_main_t *xm = &dpdk_main; + vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index); + dpdk_device_t *xd = vec_elt_at_index (xm->devices, hw->dev_instance); + + /* Shut off redirection */ + if (node_index == ~0) + { + xd->per_interface_next_index = node_index; + return; + } + + xd->per_interface_next_index = + vlib_node_add_next (xm->vlib_main, dpdk_input_node.index, node_index); +} + + +static clib_error_t * +dpdk_subif_add_del_function (vnet_main_t * vnm, + u32 hw_if_index, + struct vnet_sw_interface_t *st, int is_add) +{ + dpdk_main_t *xm = &dpdk_main; + vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index); + dpdk_device_t *xd = vec_elt_at_index (xm->devices, hw->dev_instance); + vnet_sw_interface_t *t = (vnet_sw_interface_t *) st; + int r, vlan_offload; + u32 prev_subifs = xd->num_subifs; + clib_error_t *err = 0; + + if (is_add) + xd->num_subifs++; + else if (xd->num_subifs) + xd->num_subifs--; + + if ((xd->flags & DPDK_DEVICE_FLAG_PMD) == 0) + goto done; + + /* currently we program VLANS only for IXGBE VF and I40E VF */ + if ((xd->pmd != VNET_DPDK_PMD_IXGBEVF) && (xd->pmd != VNET_DPDK_PMD_I40EVF)) + goto done; + + if (t->sub.eth.flags.no_tags == 1) + goto done; + + if ((t->sub.eth.flags.one_tag != 1) || (t->sub.eth.flags.exact_match != 1)) + { + xd->num_subifs = prev_subifs; + err = clib_error_return (0, "unsupported VLAN setup"); + goto done; + } + + vlan_offload = rte_eth_dev_get_vlan_offload (xd->device_index); + vlan_offload |= ETH_VLAN_FILTER_OFFLOAD; + + if ((r = rte_eth_dev_set_vlan_offload (xd->device_index, vlan_offload))) + { + xd->num_subifs = prev_subifs; + err = clib_error_return (0, "rte_eth_dev_set_vlan_offload[%d]: err %d", + xd->device_index, r); + goto done; + } + + + if ((r = + rte_eth_dev_vlan_filter (xd->device_index, t->sub.eth.outer_vlan_id, + is_add))) + { + xd->num_subifs = prev_subifs; + err = clib_error_return (0, "rte_eth_dev_vlan_filter[%d]: err %d", + xd->device_index, r); + goto done; + } + +done: + if (xd->num_subifs) + xd->flags |= DPDK_DEVICE_FLAG_HAVE_SUBIF; + else + xd->flags &= ~DPDK_DEVICE_FLAG_HAVE_SUBIF; + + return err; +} + +/* *INDENT-OFF* */ +VNET_DEVICE_CLASS (dpdk_device_class) = { + .name = "dpdk", + .tx_function = dpdk_interface_tx, + .tx_function_n_errors = DPDK_TX_FUNC_N_ERROR, + .tx_function_error_strings = dpdk_tx_func_error_strings, + .format_device_name = format_dpdk_device_name, + .format_device = format_dpdk_device, + .format_tx_trace = format_dpdk_tx_dma_trace, + .clear_counters = dpdk_clear_hw_interface_counters, + .admin_up_down_function = dpdk_interface_admin_up_down, + .subif_add_del_function = dpdk_subif_add_del_function, + .rx_redirect_to_node = dpdk_set_interface_next_node, + .mac_addr_change_function = dpdk_set_mac_address, +}; + +VLIB_DEVICE_TX_FUNCTION_MULTIARCH (dpdk_device_class, dpdk_interface_tx) +/* *INDENT-ON* */ + +#define UP_DOWN_FLAG_EVENT 1 + +uword +admin_up_down_process (vlib_main_t * vm, + vlib_node_runtime_t * rt, vlib_frame_t * f) +{ + clib_error_t *error = 0; + uword event_type; + uword *event_data = 0; + u32 sw_if_index; + u32 flags; + + while (1) + { + vlib_process_wait_for_event (vm); + + event_type = vlib_process_get_events (vm, &event_data); + + dpdk_main.admin_up_down_in_progress = 1; + + switch (event_type) + { + case UP_DOWN_FLAG_EVENT: + { + if (vec_len (event_data) == 2) + { + sw_if_index = event_data[0]; + flags = event_data[1]; + error = + vnet_sw_interface_set_flags (vnet_get_main (), sw_if_index, + flags); + clib_error_report (error); + } + } + break; + } + + vec_reset_length (event_data); + + dpdk_main.admin_up_down_in_progress = 0; + + } + return 0; /* or not */ +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (admin_up_down_process_node,static) = { + .function = admin_up_down_process, + .type = VLIB_NODE_TYPE_PROCESS, + .name = "admin-up-down-process", + .process_log2_n_stack_bytes = 17, // 256KB +}; +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/dpdk.h b/src/vnet/devices/dpdk/dpdk.h new file mode 100644 index 00000000000..d8f378d2b54 --- /dev/null +++ b/src/vnet/devices/dpdk/dpdk.h @@ -0,0 +1,534 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#ifndef __included_dpdk_h__ +#define __included_dpdk_h__ + +/* $$$$ We should rename always_inline -> clib_always_inline */ +#undef always_inline + +#include <rte_config.h> + +#include <rte_common.h> +#include <rte_dev.h> +#include <rte_log.h> +#include <rte_memory.h> +#include <rte_memzone.h> +#include <rte_tailq.h> +#include <rte_eal.h> +#include <rte_per_lcore.h> +#include <rte_launch.h> +#include <rte_atomic.h> +#include <rte_cycles.h> +#include <rte_prefetch.h> +#include <rte_lcore.h> +#include <rte_per_lcore.h> +#include <rte_branch_prediction.h> +#include <rte_interrupts.h> +#include <rte_pci.h> +#include <rte_random.h> +#include <rte_debug.h> +#include <rte_ether.h> +#include <rte_ethdev.h> +#include <rte_ring.h> +#include <rte_mempool.h> +#include <rte_mbuf.h> +#include <rte_virtio_net.h> +#include <rte_version.h> +#include <rte_eth_bond.h> +#include <rte_sched.h> + +#include <vnet/unix/pcap.h> +#include <vnet/devices/devices.h> + +#if CLIB_DEBUG > 0 +#define always_inline static inline +#else +#define always_inline static inline __attribute__ ((__always_inline__)) +#endif + +#include <vlib/pci/pci.h> + +#define NB_MBUF (16<<10) + +extern vnet_device_class_t dpdk_device_class; +extern vlib_node_registration_t dpdk_input_node; +extern vlib_node_registration_t handoff_dispatch_node; + +#if RTE_VERSION >= RTE_VERSION_NUM(16, 11, 0, 0) +#define foreach_dpdk_pmd \ + _ ("net_thunderx", THUNDERX) \ + _ ("net_e1000_em", E1000EM) \ + _ ("net_e1000_igb", IGB) \ + _ ("net_e1000_igb_vf", IGBVF) \ + _ ("net_ixgbe", IXGBE) \ + _ ("net_ixgbe_vf", IXGBEVF) \ + _ ("net_i40e", I40E) \ + _ ("net_i40e_vf", I40EVF) \ + _ ("net_virtio", VIRTIO) \ + _ ("net_enic", ENIC) \ + _ ("net_vmxnet3", VMXNET3) \ + _ ("net_af_packet", AF_PACKET) \ + _ ("rte_bond_pmd", BOND) \ + _ ("net_fm10k", FM10K) \ + _ ("net_cxgbe", CXGBE) \ + _ ("net_mlx5", MLX5) \ + _ ("net_dpaa2", DPAA2) +#else +#define foreach_dpdk_pmd \ + _ ("rte_nicvf_pmd", THUNDERX) \ + _ ("rte_em_pmd", E1000EM) \ + _ ("rte_igb_pmd", IGB) \ + _ ("rte_igbvf_pmd", IGBVF) \ + _ ("rte_ixgbe_pmd", IXGBE) \ + _ ("rte_ixgbevf_pmd", IXGBEVF) \ + _ ("rte_i40e_pmd", I40E) \ + _ ("rte_i40evf_pmd", I40EVF) \ + _ ("rte_virtio_pmd", VIRTIO) \ + _ ("rte_enic_pmd", ENIC) \ + _ ("rte_vmxnet3_pmd", VMXNET3) \ + _ ("AF_PACKET PMD", AF_PACKET) \ + _ ("rte_bond_pmd", BOND) \ + _ ("rte_pmd_fm10k", FM10K) \ + _ ("rte_cxgbe_pmd", CXGBE) \ + _ ("rte_dpaa2_dpni", DPAA2) +#endif + +typedef enum +{ + VNET_DPDK_PMD_NONE, +#define _(s,f) VNET_DPDK_PMD_##f, + foreach_dpdk_pmd +#undef _ + VNET_DPDK_PMD_UNKNOWN, /* must be last */ +} dpdk_pmd_t; + +typedef enum +{ + VNET_DPDK_PORT_TYPE_ETH_1G, + VNET_DPDK_PORT_TYPE_ETH_10G, + VNET_DPDK_PORT_TYPE_ETH_40G, + VNET_DPDK_PORT_TYPE_ETH_100G, + VNET_DPDK_PORT_TYPE_ETH_BOND, + VNET_DPDK_PORT_TYPE_ETH_SWITCH, + VNET_DPDK_PORT_TYPE_AF_PACKET, + VNET_DPDK_PORT_TYPE_UNKNOWN, +} dpdk_port_type_t; + +/* + * The header for the tx_vector in dpdk_device_t. + * Head and tail are indexes into the tx_vector and are of type + * u64 so they never overflow. + */ +typedef struct +{ + u64 tx_head; + u64 tx_tail; +} tx_ring_hdr_t; + +typedef struct +{ + struct rte_ring *swq; + + u64 hqos_field0_slabmask; + u32 hqos_field0_slabpos; + u32 hqos_field0_slabshr; + u64 hqos_field1_slabmask; + u32 hqos_field1_slabpos; + u32 hqos_field1_slabshr; + u64 hqos_field2_slabmask; + u32 hqos_field2_slabpos; + u32 hqos_field2_slabshr; + u32 hqos_tc_table[64]; +} dpdk_device_hqos_per_worker_thread_t; + +typedef struct +{ + struct rte_ring **swq; + struct rte_mbuf **pkts_enq; + struct rte_mbuf **pkts_deq; + struct rte_sched_port *hqos; + u32 hqos_burst_enq; + u32 hqos_burst_deq; + u32 pkts_enq_len; + u32 swq_pos; + u32 flush_count; +} dpdk_device_hqos_per_hqos_thread_t; + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + volatile u32 **lockp; + + /* Instance ID */ + u32 device_index; + + u32 vlib_hw_if_index; + u32 vlib_sw_if_index; + + /* next node index if we decide to steal the rx graph arc */ + u32 per_interface_next_index; + + /* dpdk rte_mbuf rx and tx vectors, VLIB_FRAME_SIZE */ + struct rte_mbuf ***tx_vectors; /* one per worker thread */ + struct rte_mbuf ***rx_vectors; + + /* vector of traced contexts, per device */ + u32 **d_trace_buffers; + + dpdk_pmd_t pmd:8; + i8 cpu_socket; + + u16 flags; +#define DPDK_DEVICE_FLAG_ADMIN_UP (1 << 0) +#define DPDK_DEVICE_FLAG_PROMISC (1 << 1) +#define DPDK_DEVICE_FLAG_PMD (1 << 2) +#define DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE (1 << 3) +#define DPDK_DEVICE_FLAG_MAYBE_MULTISEG (1 << 4) +#define DPDK_DEVICE_FLAG_HAVE_SUBIF (1 << 5) +#define DPDK_DEVICE_FLAG_HQOS (1 << 6) + + u16 nb_tx_desc; + CLIB_CACHE_LINE_ALIGN_MARK (cacheline1); + + u8 *interface_name_suffix; + + /* number of sub-interfaces */ + u16 num_subifs; + + /* PMD related */ + u16 tx_q_used; + u16 rx_q_used; + u16 nb_rx_desc; + u16 *cpu_socket_id_by_queue; + struct rte_eth_conf port_conf; + struct rte_eth_txconf tx_conf; + + /* HQoS related */ + dpdk_device_hqos_per_worker_thread_t *hqos_wt; + dpdk_device_hqos_per_hqos_thread_t *hqos_ht; + + /* af_packet */ + u8 af_packet_port_id; + + struct rte_eth_link link; + f64 time_last_link_update; + + struct rte_eth_stats stats; + struct rte_eth_stats last_stats; + struct rte_eth_stats last_cleared_stats; + struct rte_eth_xstat *xstats; + struct rte_eth_xstat *last_cleared_xstats; + f64 time_last_stats_update; + dpdk_port_type_t port_type; +} dpdk_device_t; + +#define DPDK_STATS_POLL_INTERVAL (10.0) +#define DPDK_MIN_STATS_POLL_INTERVAL (0.001) /* 1msec */ + +#define DPDK_LINK_POLL_INTERVAL (3.0) +#define DPDK_MIN_LINK_POLL_INTERVAL (0.001) /* 1msec */ + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + + /* total input packet counter */ + u64 aggregate_rx_packets; +} dpdk_worker_t; + +typedef struct +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + + /* total input packet counter */ + u64 aggregate_rx_packets; +} dpdk_hqos_thread_t; + +typedef struct +{ + u32 device; + u16 queue_id; +} dpdk_device_and_queue_t; + +#ifndef DPDK_HQOS_DBG_BYPASS +#define DPDK_HQOS_DBG_BYPASS 0 +#endif + +#ifndef HQOS_FLUSH_COUNT_THRESHOLD +#define HQOS_FLUSH_COUNT_THRESHOLD 100000 +#endif + +typedef struct dpdk_device_config_hqos_t +{ + u32 hqos_thread; + u32 hqos_thread_valid; + + u32 swq_size; + u32 burst_enq; + u32 burst_deq; + + u32 pktfield0_slabpos; + u32 pktfield1_slabpos; + u32 pktfield2_slabpos; + u64 pktfield0_slabmask; + u64 pktfield1_slabmask; + u64 pktfield2_slabmask; + u32 tc_table[64]; + + struct rte_sched_port_params port; + struct rte_sched_subport_params *subport; + struct rte_sched_pipe_params *pipe; + uint32_t *pipe_map; +} dpdk_device_config_hqos_t; + +int dpdk_hqos_validate_mask (u64 mask, u32 n); +void dpdk_device_config_hqos_pipe_profile_default (dpdk_device_config_hqos_t * + hqos, u32 pipe_profile_id); +void dpdk_device_config_hqos_default (dpdk_device_config_hqos_t * hqos); +clib_error_t *dpdk_port_setup_hqos (dpdk_device_t * xd, + dpdk_device_config_hqos_t * hqos); +void dpdk_hqos_metadata_set (dpdk_device_hqos_per_worker_thread_t * hqos, + struct rte_mbuf **pkts, u32 n_pkts); + +#define foreach_dpdk_device_config_item \ + _ (num_rx_queues) \ + _ (num_tx_queues) \ + _ (num_rx_desc) \ + _ (num_tx_desc) \ + _ (rss_fn) + +typedef struct +{ + vlib_pci_addr_t pci_addr; + u8 is_blacklisted; + u8 vlan_strip_offload; +#define DPDK_DEVICE_VLAN_STRIP_DEFAULT 0 +#define DPDK_DEVICE_VLAN_STRIP_OFF 1 +#define DPDK_DEVICE_VLAN_STRIP_ON 2 + +#define _(x) uword x; + foreach_dpdk_device_config_item +#undef _ + clib_bitmap_t * workers; + u32 hqos_enabled; + dpdk_device_config_hqos_t hqos; +} dpdk_device_config_t; + +typedef struct +{ + + /* Config stuff */ + u8 **eal_init_args; + u8 *eal_init_args_str; + u8 *uio_driver_name; + u8 no_multi_seg; + u8 enable_tcp_udp_checksum; + + /* Required config parameters */ + u8 coremask_set_manually; + u8 nchannels_set_manually; + u32 coremask; + u32 nchannels; + u32 num_mbufs; + u8 num_kni; /* while kni_init allows u32, port_id in callback fn is only u8 */ + + /* + * format interface names ala xxxEthernet%d/%d/%d instead of + * xxxEthernet%x/%x/%x. + */ + u8 interface_name_format_decimal; + + /* per-device config */ + dpdk_device_config_t default_devconf; + dpdk_device_config_t *dev_confs; + uword *device_config_index_by_pci_addr; + +} dpdk_config_main_t; + +dpdk_config_main_t dpdk_config_main; + +typedef struct +{ + + /* Devices */ + dpdk_device_t *devices; + dpdk_device_and_queue_t **devices_by_cpu; + dpdk_device_and_queue_t **devices_by_hqos_cpu; + + /* per-thread recycle lists */ + u32 **recycle; + + /* buffer flags template, configurable to enable/disable tcp / udp cksum */ + u32 buffer_flags_template; + + /* vlib buffer free list, must be same size as an rte_mbuf */ + u32 vlib_buffer_free_list_index; + + /* dpdk worker "threads" */ + dpdk_worker_t *workers; + + /* dpdk HQoS "threads" */ + dpdk_hqos_thread_t *hqos_threads; + + /* Ethernet input node index */ + u32 ethernet_input_node_index; + + /* pcap tracing [only works if (CLIB_DEBUG > 0)] */ + int tx_pcap_enable; + pcap_main_t pcap_main; + u8 *pcap_filename; + u32 pcap_sw_if_index; + u32 pcap_pkts_to_capture; + + /* hashes */ + uword *dpdk_device_by_kni_port_id; + uword *vu_sw_if_index_by_listener_fd; + uword *vu_sw_if_index_by_sock_fd; + u32 *vu_inactive_interfaces_device_index; + + /* + * flag indicating that a posted admin up/down + * (via post_sw_interface_set_flags) is in progress + */ + u8 admin_up_down_in_progress; + + u8 use_rss; + + /* which cpus are running dpdk-input */ + int input_cpu_first_index; + int input_cpu_count; + + /* which cpus are running I/O TX */ + int hqos_cpu_first_index; + int hqos_cpu_count; + + /* control interval of dpdk link state and stat polling */ + f64 link_state_poll_interval; + f64 stat_poll_interval; + + /* Sleep for this many MS after each device poll */ + u32 poll_sleep; + + /* convenience */ + vlib_main_t *vlib_main; + vnet_main_t *vnet_main; + dpdk_config_main_t *conf; +} dpdk_main_t; + +dpdk_main_t dpdk_main; + +typedef struct +{ + u32 buffer_index; + u16 device_index; + u8 queue_index; + struct rte_mbuf mb; + /* Copy of VLIB buffer; packet data stored in pre_data. */ + vlib_buffer_t buffer; +} dpdk_tx_dma_trace_t; + +typedef struct +{ + u32 buffer_index; + u16 device_index; + u16 queue_index; + struct rte_mbuf mb; + vlib_buffer_t buffer; /* Copy of VLIB buffer; pkt data stored in pre_data. */ + u8 data[256]; /* First 256 data bytes, used for hexdump */ +} dpdk_rx_dma_trace_t; + +void vnet_buffer_needs_dpdk_mb (vlib_buffer_t * b); + +clib_error_t *dpdk_set_mac_address (vnet_hw_interface_t * hi, char *address); + +clib_error_t *dpdk_set_mc_filter (vnet_hw_interface_t * hi, + struct ether_addr mc_addr_vec[], int naddr); + +void dpdk_thread_input (dpdk_main_t * dm, dpdk_device_t * xd); + +clib_error_t *dpdk_port_setup (dpdk_main_t * dm, dpdk_device_t * xd); + +u32 dpdk_interface_tx_vector (vlib_main_t * vm, u32 dev_instance); + +struct rte_mbuf *dpdk_replicate_packet_mb (vlib_buffer_t * b); +struct rte_mbuf *dpdk_zerocopy_replicate_packet_mb (vlib_buffer_t * b); + +#define foreach_dpdk_error \ + _(NONE, "no error") \ + _(RX_PACKET_ERROR, "Rx packet errors") \ + _(RX_BAD_FCS, "Rx bad fcs") \ + _(IP_CHECKSUM_ERROR, "Rx ip checksum errors") \ + _(RX_ALLOC_FAIL, "rx buf alloc from free list failed") \ + _(RX_ALLOC_NO_PHYSMEM, "rx buf alloc failed no physmem") \ + _(RX_ALLOC_DROP_PKTS, "rx packets dropped due to alloc error") + +typedef enum +{ +#define _(f,s) DPDK_ERROR_##f, + foreach_dpdk_error +#undef _ + DPDK_N_ERROR, +} dpdk_error_t; + +int dpdk_set_stat_poll_interval (f64 interval); +int dpdk_set_link_state_poll_interval (f64 interval); +void dpdk_update_link_state (dpdk_device_t * xd, f64 now); +void dpdk_device_lock_init (dpdk_device_t * xd); +void dpdk_device_lock_free (dpdk_device_t * xd); + +static inline u64 +vnet_get_aggregate_rx_packets (void) +{ + dpdk_main_t *dm = &dpdk_main; + u64 sum = 0; + dpdk_worker_t *dw; + + vec_foreach (dw, dm->workers) sum += dw->aggregate_rx_packets; + + return sum; +} + +void dpdk_rx_trace (dpdk_main_t * dm, + vlib_node_runtime_t * node, + dpdk_device_t * xd, + u16 queue_id, u32 * buffers, uword n_buffers); + +#define EFD_OPERATION_LESS_THAN 0 +#define EFD_OPERATION_GREATER_OR_EQUAL 1 + +format_function_t format_dpdk_device_name; +format_function_t format_dpdk_device; +format_function_t format_dpdk_tx_dma_trace; +format_function_t format_dpdk_rx_dma_trace; +format_function_t format_dpdk_rte_mbuf; +format_function_t format_dpdk_rx_rte_mbuf; +unformat_function_t unformat_socket_mem; +clib_error_t *unformat_rss_fn (unformat_input_t * input, uword * rss_fn); +clib_error_t *unformat_hqos (unformat_input_t * input, + dpdk_device_config_hqos_t * hqos); + +uword +admin_up_down_process (vlib_main_t * vm, + vlib_node_runtime_t * rt, vlib_frame_t * f); + +#endif /* __included_dpdk_h__ */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/dpdk_priv.h b/src/vnet/devices/dpdk/dpdk_priv.h new file mode 100644 index 00000000000..0c81dbc3beb --- /dev/null +++ b/src/vnet/devices/dpdk/dpdk_priv.h @@ -0,0 +1,132 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#define DPDK_NB_RX_DESC_DEFAULT 1024 +#define DPDK_NB_TX_DESC_DEFAULT 1024 +#define DPDK_NB_RX_DESC_VIRTIO 256 +#define DPDK_NB_TX_DESC_VIRTIO 256 + +#define I40E_DEV_ID_SFP_XL710 0x1572 +#define I40E_DEV_ID_QSFP_A 0x1583 +#define I40E_DEV_ID_QSFP_B 0x1584 +#define I40E_DEV_ID_QSFP_C 0x1585 +#define I40E_DEV_ID_10G_BASE_T 0x1586 +#define I40E_DEV_ID_VF 0x154C + +/* These args appear by themselves */ +#define foreach_eal_double_hyphen_predicate_arg \ +_(no-shconf) \ +_(no-hpet) \ +_(no-huge) \ +_(vmware-tsc-map) + +#define foreach_eal_single_hyphen_mandatory_arg \ +_(coremask, c) \ +_(nchannels, n) \ + +#define foreach_eal_single_hyphen_arg \ +_(blacklist, b) \ +_(mem-alloc-request, m) \ +_(force-ranks, r) + +/* These args are preceeded by "--" and followed by a single string */ +#define foreach_eal_double_hyphen_arg \ +_(huge-dir) \ +_(proc-type) \ +_(file-prefix) \ +_(vdev) + +static inline void +dpdk_get_xstats (dpdk_device_t * xd) +{ + int len; + if ((len = rte_eth_xstats_get (xd->device_index, NULL, 0)) > 0) + { + vec_validate (xd->xstats, len - 1); + vec_validate (xd->last_cleared_xstats, len - 1); + + len = + rte_eth_xstats_get (xd->device_index, xd->xstats, + vec_len (xd->xstats)); + + ASSERT (vec_len (xd->xstats) == len); + ASSERT (vec_len (xd->last_cleared_xstats) == len); + + _vec_len (xd->xstats) = len; + _vec_len (xd->last_cleared_xstats) = len; + + } +} + + +static inline void +dpdk_update_counters (dpdk_device_t * xd, f64 now) +{ + vlib_simple_counter_main_t *cm; + vnet_main_t *vnm = vnet_get_main (); + u32 my_cpu = os_get_cpu_number (); + u64 rxerrors, last_rxerrors; + + /* only update counters for PMD interfaces */ + if ((xd->flags & DPDK_DEVICE_FLAG_PMD) == 0) + return; + + xd->time_last_stats_update = now ? now : xd->time_last_stats_update; + clib_memcpy (&xd->last_stats, &xd->stats, sizeof (xd->last_stats)); + rte_eth_stats_get (xd->device_index, &xd->stats); + + /* maybe bump interface rx no buffer counter */ + if (PREDICT_FALSE (xd->stats.rx_nombuf != xd->last_stats.rx_nombuf)) + { + cm = vec_elt_at_index (vnm->interface_main.sw_if_counters, + VNET_INTERFACE_COUNTER_RX_NO_BUF); + + vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index, + xd->stats.rx_nombuf - + xd->last_stats.rx_nombuf); + } + + /* missed pkt counter */ + if (PREDICT_FALSE (xd->stats.imissed != xd->last_stats.imissed)) + { + cm = vec_elt_at_index (vnm->interface_main.sw_if_counters, + VNET_INTERFACE_COUNTER_RX_MISS); + + vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index, + xd->stats.imissed - + xd->last_stats.imissed); + } + rxerrors = xd->stats.ierrors; + last_rxerrors = xd->last_stats.ierrors; + + if (PREDICT_FALSE (rxerrors != last_rxerrors)) + { + cm = vec_elt_at_index (vnm->interface_main.sw_if_counters, + VNET_INTERFACE_COUNTER_RX_ERROR); + + vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index, + rxerrors - last_rxerrors); + } + + dpdk_get_xstats (xd); +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/format.c b/src/vnet/devices/dpdk/format.c new file mode 100644 index 00000000000..ff7c7a5a41c --- /dev/null +++ b/src/vnet/devices/dpdk/format.c @@ -0,0 +1,763 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include <vnet/vnet.h> +#include <vppinfra/vec.h> +#include <vppinfra/format.h> +#include <vlib/unix/cj.h> +#include <assert.h> + +#include <vnet/ethernet/ethernet.h> +#include <vnet/devices/dpdk/dpdk.h> + +#include "dpdk_priv.h" +#include <vppinfra/error.h> + +#define foreach_dpdk_counter \ + _ (tx_frames_ok, opackets) \ + _ (tx_bytes_ok, obytes) \ + _ (tx_errors, oerrors) \ + _ (rx_frames_ok, ipackets) \ + _ (rx_bytes_ok, ibytes) \ + _ (rx_errors, ierrors) \ + _ (rx_missed, imissed) \ + _ (rx_no_bufs, rx_nombuf) + +#define foreach_dpdk_q_counter \ + _ (rx_frames_ok, q_ipackets) \ + _ (tx_frames_ok, q_opackets) \ + _ (rx_bytes_ok, q_ibytes) \ + _ (tx_bytes_ok, q_obytes) \ + _ (rx_errors, q_errors) + +#define foreach_dpdk_rss_hf \ + _(ETH_RSS_FRAG_IPV4, "ipv4-frag") \ + _(ETH_RSS_NONFRAG_IPV4_TCP, "ipv4-tcp") \ + _(ETH_RSS_NONFRAG_IPV4_UDP, "ipv4-udp") \ + _(ETH_RSS_NONFRAG_IPV4_SCTP, "ipv4-sctp") \ + _(ETH_RSS_NONFRAG_IPV4_OTHER, "ipv4-other") \ + _(ETH_RSS_IPV4, "ipv4") \ + _(ETH_RSS_IPV6_TCP_EX, "ipv6-tcp-ex") \ + _(ETH_RSS_IPV6_UDP_EX, "ipv6-udp-ex") \ + _(ETH_RSS_FRAG_IPV6, "ipv6-frag") \ + _(ETH_RSS_NONFRAG_IPV6_TCP, "ipv6-tcp") \ + _(ETH_RSS_NONFRAG_IPV6_UDP, "ipv6-udp") \ + _(ETH_RSS_NONFRAG_IPV6_SCTP, "ipv6-sctp") \ + _(ETH_RSS_NONFRAG_IPV6_OTHER, "ipv6-other") \ + _(ETH_RSS_L2_PAYLOAD, "l2-payload") \ + _(ETH_RSS_IPV6_EX, "ipv6-ex") \ + _(ETH_RSS_IPV6, "ipv6") + + +#define foreach_dpdk_rx_offload_caps \ + _(DEV_RX_OFFLOAD_VLAN_STRIP, "vlan-strip") \ + _(DEV_RX_OFFLOAD_IPV4_CKSUM, "ipv4-cksum") \ + _(DEV_RX_OFFLOAD_UDP_CKSUM , "udp-cksum") \ + _(DEV_RX_OFFLOAD_TCP_CKSUM , "tcp-cksum") \ + _(DEV_RX_OFFLOAD_TCP_LRO , "rcp-lro") \ + _(DEV_RX_OFFLOAD_QINQ_STRIP, "qinq-strip") + +#define foreach_dpdk_tx_offload_caps \ + _(DEV_TX_OFFLOAD_VLAN_INSERT, "vlan-insert") \ + _(DEV_TX_OFFLOAD_IPV4_CKSUM, "ipv4-cksum") \ + _(DEV_TX_OFFLOAD_UDP_CKSUM , "udp-cksum") \ + _(DEV_TX_OFFLOAD_TCP_CKSUM , "tcp-cksum") \ + _(DEV_TX_OFFLOAD_SCTP_CKSUM , "sctp-cksum") \ + _(DEV_TX_OFFLOAD_TCP_TSO , "tcp-tso") \ + _(DEV_TX_OFFLOAD_UDP_TSO , "udp-tso") \ + _(DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM, "outer-ipv4-cksum") \ + _(DEV_TX_OFFLOAD_QINQ_INSERT, "qinq-insert") + +#if RTE_VERSION < RTE_VERSION_NUM(16, 11, 0, 0) +/* New ol_flags bits added in DPDK-16.11 */ +#define PKT_RX_IP_CKSUM_GOOD (1ULL << 7) +#define PKT_RX_L4_CKSUM_GOOD (1ULL << 8) +#endif + +#define foreach_dpdk_pkt_rx_offload_flag \ + _ (PKT_RX_VLAN_PKT, "RX packet is a 802.1q VLAN packet") \ + _ (PKT_RX_RSS_HASH, "RX packet with RSS hash result") \ + _ (PKT_RX_FDIR, "RX packet with FDIR infos") \ + _ (PKT_RX_L4_CKSUM_BAD, "L4 cksum of RX pkt. is not OK") \ + _ (PKT_RX_IP_CKSUM_BAD, "IP cksum of RX pkt. is not OK") \ + _ (PKT_RX_VLAN_STRIPPED, "RX packet VLAN tag stripped") \ + _ (PKT_RX_IP_CKSUM_GOOD, "IP cksum of RX pkt. is valid") \ + _ (PKT_RX_L4_CKSUM_GOOD, "L4 cksum of RX pkt. is valid") \ + _ (PKT_RX_IEEE1588_PTP, "RX IEEE1588 L2 Ethernet PT Packet") \ + _ (PKT_RX_IEEE1588_TMST, "RX IEEE1588 L2/L4 timestamped packet") \ + _ (PKT_RX_QINQ_STRIPPED, "RX packet QinQ tags stripped") + +#if RTE_VERSION < RTE_VERSION_NUM(16, 11, 0, 0) +/* PTYPE added in DPDK-16.11 */ +#define RTE_PTYPE_L2_ETHER_VLAN 0x00000006 +#define RTE_PTYPE_L2_ETHER_QINQ 0x00000007 +#endif + +#define foreach_dpdk_pkt_type \ + _ (L2, ETHER, "Ethernet packet") \ + _ (L2, ETHER_TIMESYNC, "Ethernet packet for time sync") \ + _ (L2, ETHER_ARP, "ARP packet") \ + _ (L2, ETHER_LLDP, "LLDP (Link Layer Discovery Protocol) packet") \ + _ (L2, ETHER_NSH, "NSH (Network Service Header) packet") \ + _ (L2, ETHER_VLAN, "VLAN packet") \ + _ (L2, ETHER_QINQ, "QinQ packet") \ + _ (L3, IPV4, "IPv4 packet without extension headers") \ + _ (L3, IPV4_EXT, "IPv4 packet with extension headers") \ + _ (L3, IPV4_EXT_UNKNOWN, "IPv4 packet with or without extension headers") \ + _ (L3, IPV6, "IPv6 packet without extension headers") \ + _ (L3, IPV6_EXT, "IPv6 packet with extension headers") \ + _ (L3, IPV6_EXT_UNKNOWN, "IPv6 packet with or without extension headers") \ + _ (L4, TCP, "TCP packet") \ + _ (L4, UDP, "UDP packet") \ + _ (L4, FRAG, "Fragmented IP packet") \ + _ (L4, SCTP, "SCTP (Stream Control Transmission Protocol) packet") \ + _ (L4, ICMP, "ICMP packet") \ + _ (L4, NONFRAG, "Non-fragmented IP packet") \ + _ (TUNNEL, GRE, "GRE tunneling packet") \ + _ (TUNNEL, VXLAN, "VXLAN tunneling packet") \ + _ (TUNNEL, NVGRE, "NVGRE Tunneling packet") \ + _ (TUNNEL, GENEVE, "GENEVE Tunneling packet") \ + _ (TUNNEL, GRENAT, "Teredo, VXLAN or GRE Tunneling packet") \ + _ (INNER_L2, ETHER, "Inner Ethernet packet") \ + _ (INNER_L2, ETHER_VLAN, "Inner Ethernet packet with VLAN") \ + _ (INNER_L3, IPV4, "Inner IPv4 packet without extension headers") \ + _ (INNER_L3, IPV4_EXT, "Inner IPv4 packet with extension headers") \ + _ (INNER_L3, IPV4_EXT_UNKNOWN, "Inner IPv4 packet with or without extension headers") \ + _ (INNER_L3, IPV6, "Inner IPv6 packet without extension headers") \ + _ (INNER_L3, IPV6_EXT, "Inner IPv6 packet with extension headers") \ + _ (INNER_L3, IPV6_EXT_UNKNOWN, "Inner IPv6 packet with or without extension headers") \ + _ (INNER_L4, TCP, "Inner TCP packet") \ + _ (INNER_L4, UDP, "Inner UDP packet") \ + _ (INNER_L4, FRAG, "Inner fagmented IP packet") \ + _ (INNER_L4, SCTP, "Inner SCTP (Stream Control Transmission Protocol) packet") \ + _ (INNER_L4, ICMP, "Inner ICMP packet") \ + _ (INNER_L4, NONFRAG, "Inner non-fragmented IP packet") + +#define foreach_dpdk_pkt_tx_offload_flag \ + _ (PKT_TX_VLAN_PKT, "TX packet is a 802.1q VLAN packet") \ + _ (PKT_TX_IP_CKSUM, "IP cksum of TX pkt. computed by NIC") \ + _ (PKT_TX_TCP_CKSUM, "TCP cksum of TX pkt. computed by NIC") \ + _ (PKT_TX_SCTP_CKSUM, "SCTP cksum of TX pkt. computed by NIC") \ + _ (PKT_TX_IEEE1588_TMST, "TX IEEE1588 packet to timestamp") + +#define foreach_dpdk_pkt_offload_flag \ + foreach_dpdk_pkt_rx_offload_flag \ + foreach_dpdk_pkt_tx_offload_flag + +u8 * +format_dpdk_device_name (u8 * s, va_list * args) +{ + dpdk_main_t *dm = &dpdk_main; + char *devname_format; + char *device_name; + u32 i = va_arg (*args, u32); + struct rte_eth_dev_info dev_info; + u8 *ret; + + if (dm->conf->interface_name_format_decimal) + devname_format = "%s%d/%d/%d"; + else + devname_format = "%s%x/%x/%x"; + + switch (dm->devices[i].port_type) + { + case VNET_DPDK_PORT_TYPE_ETH_1G: + device_name = "GigabitEthernet"; + break; + + case VNET_DPDK_PORT_TYPE_ETH_10G: + device_name = "TenGigabitEthernet"; + break; + + case VNET_DPDK_PORT_TYPE_ETH_40G: + device_name = "FortyGigabitEthernet"; + break; + + case VNET_DPDK_PORT_TYPE_ETH_100G: + device_name = "HundredGigabitEthernet"; + break; + + case VNET_DPDK_PORT_TYPE_ETH_BOND: + return format (s, "BondEthernet%d", dm->devices[i].device_index); + + case VNET_DPDK_PORT_TYPE_ETH_SWITCH: + device_name = "EthernetSwitch"; + break; + + case VNET_DPDK_PORT_TYPE_AF_PACKET: + rte_eth_dev_info_get (i, &dev_info); + return format (s, "af_packet%d", dm->devices[i].af_packet_port_id); + + default: + case VNET_DPDK_PORT_TYPE_UNKNOWN: + device_name = "UnknownEthernet"; + break; + } + + rte_eth_dev_info_get (i, &dev_info); + + if (dev_info.pci_dev) + ret = format (s, devname_format, device_name, dev_info.pci_dev->addr.bus, + dev_info.pci_dev->addr.devid, + dev_info.pci_dev->addr.function); + else + ret = format (s, "%s%d", device_name, dm->devices[i].device_index); + + if (dm->devices[i].interface_name_suffix) + return format (ret, "/%s", dm->devices[i].interface_name_suffix); + return ret; +} + +static u8 * +format_dpdk_device_type (u8 * s, va_list * args) +{ + dpdk_main_t *dm = &dpdk_main; + char *dev_type; + u32 i = va_arg (*args, u32); + + switch (dm->devices[i].pmd) + { + case VNET_DPDK_PMD_E1000EM: + dev_type = "Intel 82540EM (e1000)"; + break; + + case VNET_DPDK_PMD_IGB: + dev_type = "Intel e1000"; + break; + + case VNET_DPDK_PMD_I40E: + dev_type = "Intel X710/XL710 Family"; + break; + + case VNET_DPDK_PMD_I40EVF: + dev_type = "Intel X710/XL710 Family VF"; + break; + + case VNET_DPDK_PMD_FM10K: + dev_type = "Intel FM10000 Family Ethernet Switch"; + break; + + case VNET_DPDK_PMD_IGBVF: + dev_type = "Intel e1000 VF"; + break; + + case VNET_DPDK_PMD_VIRTIO: + dev_type = "Red Hat Virtio"; + break; + + case VNET_DPDK_PMD_IXGBEVF: + dev_type = "Intel 82599 VF"; + break; + + case VNET_DPDK_PMD_IXGBE: + dev_type = "Intel 82599"; + break; + + case VNET_DPDK_PMD_ENIC: + dev_type = "Cisco VIC"; + break; + + case VNET_DPDK_PMD_CXGBE: + dev_type = "Chelsio T4/T5"; + break; + + case VNET_DPDK_PMD_MLX5: + dev_type = "Mellanox ConnectX-4 Family"; + break; + + case VNET_DPDK_PMD_VMXNET3: + dev_type = "VMware VMXNET3"; + break; + + case VNET_DPDK_PMD_AF_PACKET: + dev_type = "af_packet"; + break; + + case VNET_DPDK_PMD_BOND: + dev_type = "Ethernet Bonding"; + break; + + case VNET_DPDK_PMD_DPAA2: + dev_type = "NXP DPAA2 Mac"; + break; + + default: + case VNET_DPDK_PMD_UNKNOWN: + dev_type = "### UNKNOWN ###"; + break; + } + + return format (s, dev_type); +} + +static u8 * +format_dpdk_link_status (u8 * s, va_list * args) +{ + dpdk_device_t *xd = va_arg (*args, dpdk_device_t *); + struct rte_eth_link *l = &xd->link; + vnet_main_t *vnm = vnet_get_main (); + vnet_hw_interface_t *hi = vnet_get_hw_interface (vnm, xd->vlib_hw_if_index); + + s = format (s, "%s ", l->link_status ? "up" : "down"); + if (l->link_status) + { + u32 promisc = rte_eth_promiscuous_get (xd->device_index); + + s = format (s, "%s duplex ", (l->link_duplex == ETH_LINK_FULL_DUPLEX) ? + "full" : "half"); + s = format (s, "speed %u mtu %d %s\n", l->link_speed, + hi->max_packet_bytes, promisc ? " promisc" : ""); + } + else + s = format (s, "\n"); + + return s; +} + +#define _line_len 72 +#define _(v, str) \ +if (bitmap & v) { \ + if (format_get_indent (s) > next_split ) { \ + next_split += _line_len; \ + s = format(s,"\n%U", format_white_space, indent); \ + } \ + s = format(s, "%s ", str); \ +} + +static u8 * +format_dpdk_rss_hf_name (u8 * s, va_list * args) +{ + u64 bitmap = va_arg (*args, u64); + int next_split = _line_len; + int indent = format_get_indent (s); + + if (!bitmap) + return format (s, "none"); + + foreach_dpdk_rss_hf return s; +} + +static u8 * +format_dpdk_rx_offload_caps (u8 * s, va_list * args) +{ + u32 bitmap = va_arg (*args, u32); + int next_split = _line_len; + int indent = format_get_indent (s); + + if (!bitmap) + return format (s, "none"); + + foreach_dpdk_rx_offload_caps return s; +} + +static u8 * +format_dpdk_tx_offload_caps (u8 * s, va_list * args) +{ + u32 bitmap = va_arg (*args, u32); + int next_split = _line_len; + int indent = format_get_indent (s); + if (!bitmap) + return format (s, "none"); + + foreach_dpdk_tx_offload_caps return s; +} + +#undef _line_len +#undef _ + +u8 * +format_dpdk_device (u8 * s, va_list * args) +{ + u32 dev_instance = va_arg (*args, u32); + int verbose = va_arg (*args, int); + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, dev_instance); + uword indent = format_get_indent (s); + f64 now = vlib_time_now (dm->vlib_main); + struct rte_eth_dev_info di; + + dpdk_update_counters (xd, now); + dpdk_update_link_state (xd, now); + + s = format (s, "%U\n%Ucarrier %U", + format_dpdk_device_type, xd->device_index, + format_white_space, indent + 2, format_dpdk_link_status, xd); + + rte_eth_dev_info_get (xd->device_index, &di); + + if (verbose > 1 && xd->flags & DPDK_DEVICE_FLAG_PMD) + { + struct rte_pci_device *pci; + struct rte_eth_rss_conf rss_conf; + int vlan_off; + int retval; + + rss_conf.rss_key = 0; + retval = rte_eth_dev_rss_hash_conf_get (xd->device_index, &rss_conf); + if (retval < 0) + clib_warning ("rte_eth_dev_rss_hash_conf_get returned %d", retval); + pci = di.pci_dev; + + if (pci) + s = + format (s, + "%Upci id: device %04x:%04x subsystem %04x:%04x\n" + "%Upci address: %04x:%02x:%02x.%02x\n", + format_white_space, indent + 2, pci->id.vendor_id, + pci->id.device_id, pci->id.subsystem_vendor_id, + pci->id.subsystem_device_id, format_white_space, indent + 2, + pci->addr.domain, pci->addr.bus, pci->addr.devid, + pci->addr.function); + s = + format (s, "%Umax rx packet len: %d\n", format_white_space, + indent + 2, di.max_rx_pktlen); + s = + format (s, "%Umax num of queues: rx %d tx %d\n", format_white_space, + indent + 2, di.max_rx_queues, di.max_tx_queues); + s = + format (s, "%Upromiscuous: unicast %s all-multicast %s\n", + format_white_space, indent + 2, + rte_eth_promiscuous_get (xd->device_index) ? "on" : "off", + rte_eth_promiscuous_get (xd->device_index) ? "on" : "off"); + vlan_off = rte_eth_dev_get_vlan_offload (xd->device_index); + s = format (s, "%Uvlan offload: strip %s filter %s qinq %s\n", + format_white_space, indent + 2, + vlan_off & ETH_VLAN_STRIP_OFFLOAD ? "on" : "off", + vlan_off & ETH_VLAN_FILTER_OFFLOAD ? "on" : "off", + vlan_off & ETH_VLAN_EXTEND_OFFLOAD ? "on" : "off"); + s = format (s, "%Urx offload caps: %U\n", + format_white_space, indent + 2, + format_dpdk_rx_offload_caps, di.rx_offload_capa); + s = format (s, "%Utx offload caps: %U\n", + format_white_space, indent + 2, + format_dpdk_tx_offload_caps, di.tx_offload_capa); + s = format (s, "%Urss active: %U\n" + "%Urss supported: %U\n", + format_white_space, indent + 2, + format_dpdk_rss_hf_name, rss_conf.rss_hf, + format_white_space, indent + 2, + format_dpdk_rss_hf_name, di.flow_type_rss_offloads); + } + + s = format (s, "%Urx queues %d, rx desc %d, tx queues %d, tx desc %d\n", + format_white_space, indent + 2, + xd->rx_q_used, xd->nb_rx_desc, xd->tx_q_used, xd->nb_tx_desc); + + if (xd->cpu_socket > -1) + s = format (s, "%Ucpu socket %d\n", + format_white_space, indent + 2, xd->cpu_socket); + + /* $$$ MIB counters */ + { +#define _(N, V) \ + if ((xd->stats.V - xd->last_cleared_stats.V) != 0) { \ + s = format (s, "\n%U%-40U%16Ld", \ + format_white_space, indent + 2, \ + format_c_identifier, #N, \ + xd->stats.V - xd->last_cleared_stats.V); \ + } \ + + foreach_dpdk_counter +#undef _ + } + + u8 *xs = 0; + u32 i = 0; + struct rte_eth_xstat *xstat, *last_xstat; + struct rte_eth_xstat_name *xstat_names = 0; + int len = rte_eth_xstats_get_names (xd->device_index, NULL, 0); + vec_validate (xstat_names, len - 1); + rte_eth_xstats_get_names (xd->device_index, xstat_names, len); + + ASSERT (vec_len (xd->xstats) == vec_len (xd->last_cleared_xstats)); + + /* *INDENT-OFF* */ + vec_foreach_index(i, xd->xstats) + { + u64 delta = 0; + xstat = vec_elt_at_index(xd->xstats, i); + last_xstat = vec_elt_at_index(xd->last_cleared_xstats, i); + + delta = xstat->value - last_xstat->value; + if (verbose == 2 || (verbose && delta)) + { + /* format_c_identifier doesn't like c strings inside vector */ + u8 * name = format(0,"%s", xstat_names[i].name); + xs = format(xs, "\n%U%-38U%16Ld", + format_white_space, indent + 4, + format_c_identifier, name, delta); + vec_free(name); + } + } + /* *INDENT-ON* */ + + vec_free (xstat_names); + + if (xs) + { + s = format (s, "\n%Uextended stats:%v", + format_white_space, indent + 2, xs); + vec_free (xs); + } + + return s; +} + +u8 * +format_dpdk_tx_dma_trace (u8 * s, va_list * va) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*va, vlib_node_t *); + CLIB_UNUSED (vnet_main_t * vnm) = vnet_get_main (); + dpdk_tx_dma_trace_t *t = va_arg (*va, dpdk_tx_dma_trace_t *); + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, t->device_index); + uword indent = format_get_indent (s); + vnet_sw_interface_t *sw = vnet_get_sw_interface (vnm, xd->vlib_sw_if_index); + + s = format (s, "%U tx queue %d", + format_vnet_sw_interface_name, vnm, sw, t->queue_index); + + s = format (s, "\n%Ubuffer 0x%x: %U", + format_white_space, indent, + t->buffer_index, format_vlib_buffer, &t->buffer); + + s = format (s, "\n%U%U", format_white_space, indent, + format_ethernet_header_with_length, t->buffer.pre_data, + sizeof (t->buffer.pre_data)); + + return s; +} + +u8 * +format_dpdk_rx_dma_trace (u8 * s, va_list * va) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*va, vlib_node_t *); + CLIB_UNUSED (vnet_main_t * vnm) = vnet_get_main (); + dpdk_rx_dma_trace_t *t = va_arg (*va, dpdk_rx_dma_trace_t *); + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, t->device_index); + format_function_t *f; + uword indent = format_get_indent (s); + vnet_sw_interface_t *sw = vnet_get_sw_interface (vnm, xd->vlib_sw_if_index); + + s = format (s, "%U rx queue %d", + format_vnet_sw_interface_name, vnm, sw, t->queue_index); + + s = format (s, "\n%Ubuffer 0x%x: %U", + format_white_space, indent, + t->buffer_index, format_vlib_buffer, &t->buffer); + + s = format (s, "\n%U%U", + format_white_space, indent, + format_dpdk_rte_mbuf, &t->mb, &t->data); + + if (vm->trace_main.verbose) + { + s = format (s, "\n%UPacket Dump%s", format_white_space, indent + 2, + t->mb.data_len > sizeof (t->data) ? " (truncated)" : ""); + s = format (s, "\n%U%U", format_white_space, indent + 4, + format_hexdump, &t->data, + t->mb.data_len > + sizeof (t->data) ? sizeof (t->data) : t->mb.data_len); + } + f = node->format_buffer; + if (!f) + f = format_hex_bytes; + s = format (s, "\n%U%U", format_white_space, indent, + f, t->buffer.pre_data, sizeof (t->buffer.pre_data)); + + return s; +} + + +static inline u8 * +format_dpdk_pkt_types (u8 * s, va_list * va) +{ + u32 *pkt_types = va_arg (*va, u32 *); + uword indent __attribute__ ((unused)) = format_get_indent (s) + 2; + + if (!*pkt_types) + return s; + + s = format (s, "Packet Types"); + +#define _(L, F, S) \ + if ((*pkt_types & RTE_PTYPE_##L##_MASK) == RTE_PTYPE_##L##_##F) \ + { \ + s = format (s, "\n%U%s (0x%04x) %s", format_white_space, indent, \ + "RTE_PTYPE_" #L "_" #F, RTE_PTYPE_##L##_##F, S); \ + } + + foreach_dpdk_pkt_type +#undef _ + return s; +} + +static inline u8 * +format_dpdk_pkt_offload_flags (u8 * s, va_list * va) +{ + u64 *ol_flags = va_arg (*va, u64 *); + uword indent = format_get_indent (s) + 2; + + if (!*ol_flags) + return s; + + s = format (s, "Packet Offload Flags"); + +#define _(F, S) \ + if (*ol_flags & F) \ + { \ + s = format (s, "\n%U%s (0x%04x) %s", \ + format_white_space, indent, #F, F, S); \ + } + + foreach_dpdk_pkt_offload_flag +#undef _ + return s; +} + +u8 * +format_dpdk_rte_mbuf_vlan (u8 * s, va_list * va) +{ + ethernet_vlan_header_tv_t *vlan_hdr = + va_arg (*va, ethernet_vlan_header_tv_t *); + + if (clib_net_to_host_u16 (vlan_hdr->type) == ETHERNET_TYPE_DOT1AD) + { + s = format (s, "%U 802.1q vlan ", + format_ethernet_vlan_tci, + clib_net_to_host_u16 (vlan_hdr->priority_cfi_and_id)); + vlan_hdr++; + } + + s = format (s, "%U", + format_ethernet_vlan_tci, + clib_net_to_host_u16 (vlan_hdr->priority_cfi_and_id)); + + return s; +} + +u8 * +format_dpdk_rte_mbuf (u8 * s, va_list * va) +{ + struct rte_mbuf *mb = va_arg (*va, struct rte_mbuf *); + ethernet_header_t *eth_hdr = va_arg (*va, ethernet_header_t *); + uword indent = format_get_indent (s) + 2; + + s = format (s, "PKT MBUF: port %d, nb_segs %d, pkt_len %d" + "\n%Ubuf_len %d, data_len %d, ol_flags 0x%x, data_off %d, phys_addr 0x%x" + "\n%Upacket_type 0x%x", + mb->port, mb->nb_segs, mb->pkt_len, + format_white_space, indent, + mb->buf_len, mb->data_len, mb->ol_flags, mb->data_off, + mb->buf_physaddr, format_white_space, indent, mb->packet_type); + + if (mb->ol_flags) + s = format (s, "\n%U%U", format_white_space, indent, + format_dpdk_pkt_offload_flags, &mb->ol_flags); + + if ((mb->ol_flags & PKT_RX_VLAN_PKT) && + ((mb->ol_flags & (PKT_RX_VLAN_STRIPPED | PKT_RX_QINQ_STRIPPED)) == 0)) + { + ethernet_vlan_header_tv_t *vlan_hdr = + ((ethernet_vlan_header_tv_t *) & (eth_hdr->type)); + s = format (s, " %U", format_dpdk_rte_mbuf_vlan, vlan_hdr); + } + + if (mb->packet_type) + s = format (s, "\n%U%U", format_white_space, indent, + format_dpdk_pkt_types, &mb->packet_type); + + return s; +} + +uword +unformat_socket_mem (unformat_input_t * input, va_list * va) +{ + uword **r = va_arg (*va, uword **); + int i = 0; + u32 mem; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, ",")) + hash_set (*r, i, 1024); + else if (unformat (input, "%u,", &mem)) + hash_set (*r, i, mem); + else if (unformat (input, "%u", &mem)) + hash_set (*r, i, mem); + else + { + unformat_put_input (input); + goto done; + } + i++; + } + +done: + return 1; +} + +clib_error_t * +unformat_rss_fn (unformat_input_t * input, uword * rss_fn) +{ + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (0) + ; +#undef _ +#define _(f, s) \ + else if (unformat (input, s)) \ + *rss_fn |= f; + + foreach_dpdk_rss_hf +#undef _ + else + { + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + } + return 0; +} + +clib_error_t * +unformat_hqos (unformat_input_t * input, dpdk_device_config_hqos_t * hqos) +{ + clib_error_t *error = 0; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "hqos-thread %u", &hqos->hqos_thread)) + hqos->hqos_thread_valid = 1; + else + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + break; + } + } + + return error; +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/hqos.c b/src/vnet/devices/dpdk/hqos.c new file mode 100644 index 00000000000..d68bc48f80b --- /dev/null +++ b/src/vnet/devices/dpdk/hqos.c @@ -0,0 +1,775 @@ +/* + * Copyright(c) 2016 Intel Corporation. All rights reserved. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include <stdio.h> +#include <stdlib.h> +#include <unistd.h> +#include <sys/stat.h> +#include <sys/mount.h> +#include <string.h> +#include <fcntl.h> + +#include <vppinfra/vec.h> +#include <vppinfra/error.h> +#include <vppinfra/format.h> +#include <vppinfra/bitmap.h> + +#include <vnet/vnet.h> +#include <vnet/ethernet/ethernet.h> +#include <vnet/devices/dpdk/dpdk.h> + +#include <vlib/unix/physmem.h> +#include <vlib/pci/pci.h> +#include <vlibmemory/api.h> +#include <vlibmemory/vl_memory_msg_enum.h> /* enumerate all vlib messages */ + +#define vl_typedefs /* define message structures */ +#include <vlibmemory/vl_memory_api_h.h> +#undef vl_typedefs + +/* instantiate all the print functions we know about */ +#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__) +#define vl_printfun +#include <vlibmemory/vl_memory_api_h.h> +#undef vl_printfun + +#include "dpdk_priv.h" + +dpdk_main_t dpdk_main; + +/*** + * + * HQoS default configuration values + * + ***/ + +static dpdk_device_config_hqos_t hqos_params_default = { + .hqos_thread_valid = 0, + + .swq_size = 4096, + .burst_enq = 256, + .burst_deq = 220, + + /* + * Packet field to identify the subport. + * + * Default value: Since only one subport is defined by default (see below: + * n_subports_per_port = 1), the subport ID is hardcoded to 0. + */ + .pktfield0_slabpos = 0, + .pktfield0_slabmask = 0, + + /* + * Packet field to identify the pipe. + * + * Default value: Assuming Ethernet/IPv4/UDP packets, UDP payload bits 12 .. 23 + */ + .pktfield1_slabpos = 40, + .pktfield1_slabmask = 0x0000000FFF000000LLU, + + /* Packet field used as index into TC translation table to identify the traffic + * class and queue. + * + * Default value: Assuming Ethernet/IPv4 packets, IPv4 DSCP field + */ + .pktfield2_slabpos = 8, + .pktfield2_slabmask = 0x00000000000000FCLLU, + .tc_table = { + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + }, + + /* port */ + .port = { + .name = NULL, /* Set at init */ + .socket = 0, /* Set at init */ + .rate = 1250000000, /* Assuming 10GbE port */ + .mtu = 14 + 1500, /* Assuming Ethernet/IPv4 pkt (Ethernet FCS not included) */ + .frame_overhead = RTE_SCHED_FRAME_OVERHEAD_DEFAULT, + .n_subports_per_port = 1, + .n_pipes_per_subport = 4096, + .qsize = {64, 64, 64, 64}, + .pipe_profiles = NULL, /* Set at config */ + .n_pipe_profiles = 1, + +#ifdef RTE_SCHED_RED + .red_params = { + /* Traffic Class 0 Colors Green / Yellow / Red */ + [0][0] = {.min_th = 48,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [0][1] = {.min_th = 40,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [0][2] = {.min_th = 32,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + + /* Traffic Class 1 - Colors Green / Yellow / Red */ + [1][0] = {.min_th = 48,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [1][1] = {.min_th = 40,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [1][2] = {.min_th = 32,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + + /* Traffic Class 2 - Colors Green / Yellow / Red */ + [2][0] = {.min_th = 48,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [2][1] = {.min_th = 40,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [2][2] = {.min_th = 32,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + + /* Traffic Class 3 - Colors Green / Yellow / Red */ + [3][0] = {.min_th = 48,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [3][1] = {.min_th = 40,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9}, + [3][2] = {.min_th = 32,.max_th = 64,.maxp_inv = + 10,.wq_log2 = 9} + }, +#endif /* RTE_SCHED_RED */ + }, +}; + +static struct rte_sched_subport_params hqos_subport_params_default = { + .tb_rate = 1250000000, /* 10GbE line rate (measured in bytes/second) */ + .tb_size = 1000000, + .tc_rate = {1250000000, 1250000000, 1250000000, 1250000000}, + .tc_period = 10, +}; + +static struct rte_sched_pipe_params hqos_pipe_params_default = { + .tb_rate = 305175, /* 10GbE line rate divided by 4K pipes */ + .tb_size = 1000000, + .tc_rate = {305175, 305175, 305175, 305175}, + .tc_period = 40, +#ifdef RTE_SCHED_SUBPORT_TC_OV + .tc_ov_weight = 1, +#endif + .wrr_weights = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, +}; + +/*** + * + * HQoS configuration + * + ***/ + +int +dpdk_hqos_validate_mask (u64 mask, u32 n) +{ + int count = __builtin_popcountll (mask); + int pos_lead = sizeof (u64) * 8 - __builtin_clzll (mask); + int pos_trail = __builtin_ctzll (mask); + int count_expected = __builtin_popcount (n - 1); + + /* Handle the exceptions */ + if (n == 0) + return -1; /* Error */ + + if ((mask == 0) && (n == 1)) + return 0; /* OK */ + + if (((mask == 0) && (n != 1)) || ((mask != 0) && (n == 1))) + return -2; /* Error */ + + /* Check that mask is contiguous */ + if ((pos_lead - pos_trail) != count) + return -3; /* Error */ + + /* Check that mask contains the expected number of bits set */ + if (count != count_expected) + return -4; /* Error */ + + return 0; /* OK */ +} + +void +dpdk_device_config_hqos_pipe_profile_default (dpdk_device_config_hqos_t * + hqos, u32 pipe_profile_id) +{ + memcpy (&hqos->pipe[pipe_profile_id], &hqos_pipe_params_default, + sizeof (hqos_pipe_params_default)); +} + +void +dpdk_device_config_hqos_default (dpdk_device_config_hqos_t * hqos) +{ + struct rte_sched_subport_params *subport_params; + struct rte_sched_pipe_params *pipe_params; + u32 *pipe_map; + u32 i; + + memcpy (hqos, &hqos_params_default, sizeof (hqos_params_default)); + + /* pipe */ + vec_add2 (hqos->pipe, pipe_params, hqos->port.n_pipe_profiles); + + for (i = 0; i < vec_len (hqos->pipe); i++) + memcpy (&pipe_params[i], + &hqos_pipe_params_default, sizeof (hqos_pipe_params_default)); + + hqos->port.pipe_profiles = hqos->pipe; + + /* subport */ + vec_add2 (hqos->subport, subport_params, hqos->port.n_subports_per_port); + + for (i = 0; i < vec_len (hqos->subport); i++) + memcpy (&subport_params[i], + &hqos_subport_params_default, + sizeof (hqos_subport_params_default)); + + /* pipe profile */ + vec_add2 (hqos->pipe_map, + pipe_map, + hqos->port.n_subports_per_port * hqos->port.n_pipes_per_subport); + + for (i = 0; i < vec_len (hqos->pipe_map); i++) + pipe_map[i] = 0; +} + +/*** + * + * HQoS init + * + ***/ + +clib_error_t * +dpdk_port_setup_hqos (dpdk_device_t * xd, dpdk_device_config_hqos_t * hqos) +{ + vlib_thread_main_t *tm = vlib_get_thread_main (); + char name[32]; + u32 subport_id, i; + int rv; + + /* Detect the set of worker threads */ + int worker_thread_first = 0; + int worker_thread_count = 0; + + uword *p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + vlib_thread_registration_t *tr = + p ? (vlib_thread_registration_t *) p[0] : 0; + + if (tr && tr->count > 0) + { + worker_thread_first = tr->first_index; + worker_thread_count = tr->count; + } + + /* Allocate the per-thread device data array */ + vec_validate_aligned (xd->hqos_wt, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + memset (xd->hqos_wt, 0, tm->n_vlib_mains * sizeof (xd->hqos_wt[0])); + + vec_validate_aligned (xd->hqos_ht, 0, CLIB_CACHE_LINE_BYTES); + memset (xd->hqos_ht, 0, sizeof (xd->hqos_ht[0])); + + /* Allocate space for one SWQ per worker thread in the I/O TX thread data structure */ + vec_validate (xd->hqos_ht->swq, worker_thread_count); + + /* SWQ */ + for (i = 0; i < worker_thread_count + 1; i++) + { + u32 swq_flags = RING_F_SP_ENQ | RING_F_SC_DEQ; + + snprintf (name, sizeof (name), "SWQ-worker%u-to-device%u", i, + xd->device_index); + xd->hqos_ht->swq[i] = + rte_ring_create (name, hqos->swq_size, xd->cpu_socket, swq_flags); + if (xd->hqos_ht->swq[i] == NULL) + return clib_error_return (0, + "SWQ-worker%u-to-device%u: rte_ring_create err", + i, xd->device_index); + } + + /* + * HQoS + */ + + /* HQoS port */ + snprintf (name, sizeof (name), "HQoS%u", xd->device_index); + hqos->port.name = strdup (name); + if (hqos->port.name == NULL) + return clib_error_return (0, "HQoS%u: strdup err", xd->device_index); + + hqos->port.socket = rte_eth_dev_socket_id (xd->device_index); + if (hqos->port.socket == SOCKET_ID_ANY) + hqos->port.socket = 0; + + xd->hqos_ht->hqos = rte_sched_port_config (&hqos->port); + if (xd->hqos_ht->hqos == NULL) + return clib_error_return (0, "HQoS%u: rte_sched_port_config err", + xd->device_index); + + /* HQoS subport */ + for (subport_id = 0; subport_id < hqos->port.n_subports_per_port; + subport_id++) + { + u32 pipe_id; + + rv = + rte_sched_subport_config (xd->hqos_ht->hqos, subport_id, + &hqos->subport[subport_id]); + if (rv) + return clib_error_return (0, + "HQoS%u subport %u: rte_sched_subport_config err (%d)", + xd->device_index, subport_id, rv); + + /* HQoS pipe */ + for (pipe_id = 0; pipe_id < hqos->port.n_pipes_per_subport; pipe_id++) + { + u32 pos = subport_id * hqos->port.n_pipes_per_subport + pipe_id; + u32 profile_id = hqos->pipe_map[pos]; + + rv = + rte_sched_pipe_config (xd->hqos_ht->hqos, subport_id, pipe_id, + profile_id); + if (rv) + return clib_error_return (0, + "HQoS%u subport %u pipe %u: rte_sched_pipe_config err (%d)", + xd->device_index, subport_id, pipe_id, + rv); + } + } + + /* Set up per-thread device data for the I/O TX thread */ + xd->hqos_ht->hqos_burst_enq = hqos->burst_enq; + xd->hqos_ht->hqos_burst_deq = hqos->burst_deq; + vec_validate (xd->hqos_ht->pkts_enq, 2 * hqos->burst_enq - 1); + vec_validate (xd->hqos_ht->pkts_deq, hqos->burst_deq - 1); + xd->hqos_ht->pkts_enq_len = 0; + xd->hqos_ht->swq_pos = 0; + xd->hqos_ht->flush_count = 0; + + /* Set up per-thread device data for each worker thread */ + for (i = 0; i < worker_thread_count + 1; i++) + { + u32 tid; + if (i) + tid = worker_thread_first + (i - 1); + else + tid = i; + + xd->hqos_wt[tid].swq = xd->hqos_ht->swq[i]; + xd->hqos_wt[tid].hqos_field0_slabpos = hqos->pktfield0_slabpos; + xd->hqos_wt[tid].hqos_field0_slabmask = hqos->pktfield0_slabmask; + xd->hqos_wt[tid].hqos_field0_slabshr = + __builtin_ctzll (hqos->pktfield0_slabmask); + xd->hqos_wt[tid].hqos_field1_slabpos = hqos->pktfield1_slabpos; + xd->hqos_wt[tid].hqos_field1_slabmask = hqos->pktfield1_slabmask; + xd->hqos_wt[tid].hqos_field1_slabshr = + __builtin_ctzll (hqos->pktfield1_slabmask); + xd->hqos_wt[tid].hqos_field2_slabpos = hqos->pktfield2_slabpos; + xd->hqos_wt[tid].hqos_field2_slabmask = hqos->pktfield2_slabmask; + xd->hqos_wt[tid].hqos_field2_slabshr = + __builtin_ctzll (hqos->pktfield2_slabmask); + memcpy (xd->hqos_wt[tid].hqos_tc_table, hqos->tc_table, + sizeof (hqos->tc_table)); + } + + return 0; +} + +/*** + * + * HQoS run-time + * + ***/ +/* + * dpdk_hqos_thread - Contains the main loop of an HQoS thread. + * + * w + * Information for the current thread + */ +static_always_inline void +dpdk_hqos_thread_internal_hqos_dbg_bypass (vlib_main_t * vm) +{ + dpdk_main_t *dm = &dpdk_main; + u32 cpu_index = vm->cpu_index; + u32 dev_pos; + + dev_pos = 0; + while (1) + { + vlib_worker_thread_barrier_check (); + + u32 n_devs = vec_len (dm->devices_by_hqos_cpu[cpu_index]); + if (dev_pos >= n_devs) + dev_pos = 0; + + dpdk_device_and_queue_t *dq = + vec_elt_at_index (dm->devices_by_hqos_cpu[cpu_index], dev_pos); + dpdk_device_t *xd = vec_elt_at_index (dm->devices, dq->device); + + dpdk_device_hqos_per_hqos_thread_t *hqos = xd->hqos_ht; + u32 device_index = xd->device_index; + u16 queue_id = dq->queue_id; + + struct rte_mbuf **pkts_enq = hqos->pkts_enq; + u32 pkts_enq_len = hqos->pkts_enq_len; + u32 swq_pos = hqos->swq_pos; + u32 n_swq = vec_len (hqos->swq), i; + u32 flush_count = hqos->flush_count; + + for (i = 0; i < n_swq; i++) + { + /* Get current SWQ for this device */ + struct rte_ring *swq = hqos->swq[swq_pos]; + + /* Read SWQ burst to packet buffer of this device */ + pkts_enq_len += rte_ring_sc_dequeue_burst (swq, + (void **) + &pkts_enq[pkts_enq_len], + hqos->hqos_burst_enq); + + /* Get next SWQ for this device */ + swq_pos++; + if (swq_pos >= n_swq) + swq_pos = 0; + hqos->swq_pos = swq_pos; + + /* HWQ TX enqueue when burst available */ + if (pkts_enq_len >= hqos->hqos_burst_enq) + { + u32 n_pkts = rte_eth_tx_burst (device_index, + (uint16_t) queue_id, + pkts_enq, + (uint16_t) pkts_enq_len); + + for (; n_pkts < pkts_enq_len; n_pkts++) + rte_pktmbuf_free (pkts_enq[n_pkts]); + + pkts_enq_len = 0; + flush_count = 0; + break; + } + } + if (pkts_enq_len) + { + flush_count++; + if (PREDICT_FALSE (flush_count == HQOS_FLUSH_COUNT_THRESHOLD)) + { + rte_sched_port_enqueue (hqos->hqos, pkts_enq, pkts_enq_len); + + pkts_enq_len = 0; + flush_count = 0; + } + } + hqos->pkts_enq_len = pkts_enq_len; + hqos->flush_count = flush_count; + + /* Advance to next device */ + dev_pos++; + } +} + +static_always_inline void +dpdk_hqos_thread_internal (vlib_main_t * vm) +{ + dpdk_main_t *dm = &dpdk_main; + u32 cpu_index = vm->cpu_index; + u32 dev_pos; + + dev_pos = 0; + while (1) + { + vlib_worker_thread_barrier_check (); + + u32 n_devs = vec_len (dm->devices_by_hqos_cpu[cpu_index]); + if (PREDICT_FALSE (n_devs == 0)) + { + dev_pos = 0; + continue; + } + if (dev_pos >= n_devs) + dev_pos = 0; + + dpdk_device_and_queue_t *dq = + vec_elt_at_index (dm->devices_by_hqos_cpu[cpu_index], dev_pos); + dpdk_device_t *xd = vec_elt_at_index (dm->devices, dq->device); + + dpdk_device_hqos_per_hqos_thread_t *hqos = xd->hqos_ht; + u32 device_index = xd->device_index; + u16 queue_id = dq->queue_id; + + struct rte_mbuf **pkts_enq = hqos->pkts_enq; + struct rte_mbuf **pkts_deq = hqos->pkts_deq; + u32 pkts_enq_len = hqos->pkts_enq_len; + u32 swq_pos = hqos->swq_pos; + u32 n_swq = vec_len (hqos->swq), i; + u32 flush_count = hqos->flush_count; + + /* + * SWQ dequeue and HQoS enqueue for current device + */ + for (i = 0; i < n_swq; i++) + { + /* Get current SWQ for this device */ + struct rte_ring *swq = hqos->swq[swq_pos]; + + /* Read SWQ burst to packet buffer of this device */ + pkts_enq_len += rte_ring_sc_dequeue_burst (swq, + (void **) + &pkts_enq[pkts_enq_len], + hqos->hqos_burst_enq); + + /* Get next SWQ for this device */ + swq_pos++; + if (swq_pos >= n_swq) + swq_pos = 0; + hqos->swq_pos = swq_pos; + + /* HQoS enqueue when burst available */ + if (pkts_enq_len >= hqos->hqos_burst_enq) + { + rte_sched_port_enqueue (hqos->hqos, pkts_enq, pkts_enq_len); + + pkts_enq_len = 0; + flush_count = 0; + break; + } + } + if (pkts_enq_len) + { + flush_count++; + if (PREDICT_FALSE (flush_count == HQOS_FLUSH_COUNT_THRESHOLD)) + { + rte_sched_port_enqueue (hqos->hqos, pkts_enq, pkts_enq_len); + + pkts_enq_len = 0; + flush_count = 0; + } + } + hqos->pkts_enq_len = pkts_enq_len; + hqos->flush_count = flush_count; + + /* + * HQoS dequeue and HWQ TX enqueue for current device + */ + { + u32 pkts_deq_len, n_pkts; + + pkts_deq_len = rte_sched_port_dequeue (hqos->hqos, + pkts_deq, + hqos->hqos_burst_deq); + + for (n_pkts = 0; n_pkts < pkts_deq_len;) + n_pkts += rte_eth_tx_burst (device_index, + (uint16_t) queue_id, + &pkts_deq[n_pkts], + (uint16_t) (pkts_deq_len - n_pkts)); + } + + /* Advance to next device */ + dev_pos++; + } +} + +void +dpdk_hqos_thread (vlib_worker_thread_t * w) +{ + vlib_main_t *vm; + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_main_t *dm = &dpdk_main; + + vm = vlib_get_main (); + + ASSERT (vm->cpu_index == os_get_cpu_number ()); + + clib_time_init (&vm->clib_time); + clib_mem_set_heap (w->thread_mheap); + + /* Wait until the dpdk init sequence is complete */ + while (tm->worker_thread_release == 0) + vlib_worker_thread_barrier_check (); + + if (vec_len (dm->devices_by_hqos_cpu[vm->cpu_index]) == 0) + return + clib_error + ("current I/O TX thread does not have any devices assigned to it"); + + if (DPDK_HQOS_DBG_BYPASS) + dpdk_hqos_thread_internal_hqos_dbg_bypass (vm); + else + dpdk_hqos_thread_internal (vm); +} + +void +dpdk_hqos_thread_fn (void *arg) +{ + vlib_worker_thread_t *w = (vlib_worker_thread_t *) arg; + vlib_worker_thread_init (w); + dpdk_hqos_thread (w); +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_THREAD (hqos_thread_reg, static) = +{ + .name = "hqos-threads", + .short_name = "hqos-threads", + .function = dpdk_hqos_thread_fn, +}; +/* *INDENT-ON* */ + +/* + * HQoS run-time code to be called by the worker threads + */ +#define BITFIELD(byte_array, slab_pos, slab_mask, slab_shr) \ +({ \ + u64 slab = *((u64 *) &byte_array[slab_pos]); \ + u64 val = (rte_be_to_cpu_64(slab) & slab_mask) >> slab_shr; \ + val; \ +}) + +#define RTE_SCHED_PORT_HIERARCHY(subport, pipe, traffic_class, queue, color) \ + ((((u64) (queue)) & 0x3) | \ + ((((u64) (traffic_class)) & 0x3) << 2) | \ + ((((u64) (color)) & 0x3) << 4) | \ + ((((u64) (subport)) & 0xFFFF) << 16) | \ + ((((u64) (pipe)) & 0xFFFFFFFF) << 32)) + +void +dpdk_hqos_metadata_set (dpdk_device_hqos_per_worker_thread_t * hqos, + struct rte_mbuf **pkts, u32 n_pkts) +{ + u32 i; + + for (i = 0; i < (n_pkts & (~0x3)); i += 4) + { + struct rte_mbuf *pkt0 = pkts[i]; + struct rte_mbuf *pkt1 = pkts[i + 1]; + struct rte_mbuf *pkt2 = pkts[i + 2]; + struct rte_mbuf *pkt3 = pkts[i + 3]; + + u8 *pkt0_data = rte_pktmbuf_mtod (pkt0, u8 *); + u8 *pkt1_data = rte_pktmbuf_mtod (pkt1, u8 *); + u8 *pkt2_data = rte_pktmbuf_mtod (pkt2, u8 *); + u8 *pkt3_data = rte_pktmbuf_mtod (pkt3, u8 *); + + u64 pkt0_subport = BITFIELD (pkt0_data, hqos->hqos_field0_slabpos, + hqos->hqos_field0_slabmask, + hqos->hqos_field0_slabshr); + u64 pkt0_pipe = BITFIELD (pkt0_data, hqos->hqos_field1_slabpos, + hqos->hqos_field1_slabmask, + hqos->hqos_field1_slabshr); + u64 pkt0_dscp = BITFIELD (pkt0_data, hqos->hqos_field2_slabpos, + hqos->hqos_field2_slabmask, + hqos->hqos_field2_slabshr); + u32 pkt0_tc = hqos->hqos_tc_table[pkt0_dscp & 0x3F] >> 2; + u32 pkt0_tc_q = hqos->hqos_tc_table[pkt0_dscp & 0x3F] & 0x3; + + u64 pkt1_subport = BITFIELD (pkt1_data, hqos->hqos_field0_slabpos, + hqos->hqos_field0_slabmask, + hqos->hqos_field0_slabshr); + u64 pkt1_pipe = BITFIELD (pkt1_data, hqos->hqos_field1_slabpos, + hqos->hqos_field1_slabmask, + hqos->hqos_field1_slabshr); + u64 pkt1_dscp = BITFIELD (pkt1_data, hqos->hqos_field2_slabpos, + hqos->hqos_field2_slabmask, + hqos->hqos_field2_slabshr); + u32 pkt1_tc = hqos->hqos_tc_table[pkt1_dscp & 0x3F] >> 2; + u32 pkt1_tc_q = hqos->hqos_tc_table[pkt1_dscp & 0x3F] & 0x3; + + u64 pkt2_subport = BITFIELD (pkt2_data, hqos->hqos_field0_slabpos, + hqos->hqos_field0_slabmask, + hqos->hqos_field0_slabshr); + u64 pkt2_pipe = BITFIELD (pkt2_data, hqos->hqos_field1_slabpos, + hqos->hqos_field1_slabmask, + hqos->hqos_field1_slabshr); + u64 pkt2_dscp = BITFIELD (pkt2_data, hqos->hqos_field2_slabpos, + hqos->hqos_field2_slabmask, + hqos->hqos_field2_slabshr); + u32 pkt2_tc = hqos->hqos_tc_table[pkt2_dscp & 0x3F] >> 2; + u32 pkt2_tc_q = hqos->hqos_tc_table[pkt2_dscp & 0x3F] & 0x3; + + u64 pkt3_subport = BITFIELD (pkt3_data, hqos->hqos_field0_slabpos, + hqos->hqos_field0_slabmask, + hqos->hqos_field0_slabshr); + u64 pkt3_pipe = BITFIELD (pkt3_data, hqos->hqos_field1_slabpos, + hqos->hqos_field1_slabmask, + hqos->hqos_field1_slabshr); + u64 pkt3_dscp = BITFIELD (pkt3_data, hqos->hqos_field2_slabpos, + hqos->hqos_field2_slabmask, + hqos->hqos_field2_slabshr); + u32 pkt3_tc = hqos->hqos_tc_table[pkt3_dscp & 0x3F] >> 2; + u32 pkt3_tc_q = hqos->hqos_tc_table[pkt3_dscp & 0x3F] & 0x3; + + u64 pkt0_sched = RTE_SCHED_PORT_HIERARCHY (pkt0_subport, + pkt0_pipe, + pkt0_tc, + pkt0_tc_q, + 0); + u64 pkt1_sched = RTE_SCHED_PORT_HIERARCHY (pkt1_subport, + pkt1_pipe, + pkt1_tc, + pkt1_tc_q, + 0); + u64 pkt2_sched = RTE_SCHED_PORT_HIERARCHY (pkt2_subport, + pkt2_pipe, + pkt2_tc, + pkt2_tc_q, + 0); + u64 pkt3_sched = RTE_SCHED_PORT_HIERARCHY (pkt3_subport, + pkt3_pipe, + pkt3_tc, + pkt3_tc_q, + 0); + + pkt0->hash.sched.lo = pkt0_sched & 0xFFFFFFFF; + pkt0->hash.sched.hi = pkt0_sched >> 32; + pkt1->hash.sched.lo = pkt1_sched & 0xFFFFFFFF; + pkt1->hash.sched.hi = pkt1_sched >> 32; + pkt2->hash.sched.lo = pkt2_sched & 0xFFFFFFFF; + pkt2->hash.sched.hi = pkt2_sched >> 32; + pkt3->hash.sched.lo = pkt3_sched & 0xFFFFFFFF; + pkt3->hash.sched.hi = pkt3_sched >> 32; + } + + for (; i < n_pkts; i++) + { + struct rte_mbuf *pkt = pkts[i]; + + u8 *pkt_data = rte_pktmbuf_mtod (pkt, u8 *); + + u64 pkt_subport = BITFIELD (pkt_data, hqos->hqos_field0_slabpos, + hqos->hqos_field0_slabmask, + hqos->hqos_field0_slabshr); + u64 pkt_pipe = BITFIELD (pkt_data, hqos->hqos_field1_slabpos, + hqos->hqos_field1_slabmask, + hqos->hqos_field1_slabshr); + u64 pkt_dscp = BITFIELD (pkt_data, hqos->hqos_field2_slabpos, + hqos->hqos_field2_slabmask, + hqos->hqos_field2_slabshr); + u32 pkt_tc = hqos->hqos_tc_table[pkt_dscp & 0x3F] >> 2; + u32 pkt_tc_q = hqos->hqos_tc_table[pkt_dscp & 0x3F] & 0x3; + + u64 pkt_sched = RTE_SCHED_PORT_HIERARCHY (pkt_subport, + pkt_pipe, + pkt_tc, + pkt_tc_q, + 0); + + pkt->hash.sched.lo = pkt_sched & 0xFFFFFFFF; + pkt->hash.sched.hi = pkt_sched >> 32; + } +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/init.c b/src/vnet/devices/dpdk/init.c new file mode 100755 index 00000000000..693ca985130 --- /dev/null +++ b/src/vnet/devices/dpdk/init.c @@ -0,0 +1,1803 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include <vnet/vnet.h> +#include <vppinfra/vec.h> +#include <vppinfra/error.h> +#include <vppinfra/format.h> +#include <vppinfra/bitmap.h> + +#include <vnet/ethernet/ethernet.h> +#include <vnet/devices/dpdk/dpdk.h> +#include <vlib/unix/physmem.h> +#include <vlib/pci/pci.h> + +#include <stdio.h> +#include <stdlib.h> +#include <unistd.h> +#include <sys/stat.h> +#include <sys/mount.h> +#include <string.h> +#include <fcntl.h> + +#include "dpdk_priv.h" + +dpdk_main_t dpdk_main; + +/* force linker to link functions used by vlib and declared weak */ +void *vlib_weakly_linked_functions[] = { + &rte_pktmbuf_init, + &rte_pktmbuf_pool_init, +}; + +#define LINK_STATE_ELOGS 0 + +#define DEFAULT_HUGE_DIR "/run/vpp/hugepages" +#define VPP_RUN_DIR "/run/vpp" + +/* Port configuration, mildly modified Intel app values */ + +static struct rte_eth_conf port_conf_template = { + .rxmode = { + .split_hdr_size = 0, + .header_split = 0, /**< Header Split disabled */ + .hw_ip_checksum = 0, /**< IP checksum offload disabled */ + .hw_vlan_filter = 0, /**< VLAN filtering disabled */ + .hw_strip_crc = 0, /**< CRC stripped by hardware */ + }, + .txmode = { + .mq_mode = ETH_MQ_TX_NONE, + }, +}; + +clib_error_t * +dpdk_port_setup (dpdk_main_t * dm, dpdk_device_t * xd) +{ + vlib_main_t *vm = vlib_get_main (); + vlib_buffer_main_t *bm = vm->buffer_main; + int rv; + int j; + + ASSERT (os_get_cpu_number () == 0); + + if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) + { + vnet_hw_interface_set_flags (dm->vnet_main, xd->vlib_hw_if_index, 0); + rte_eth_dev_stop (xd->device_index); + } + + rv = rte_eth_dev_configure (xd->device_index, xd->rx_q_used, + xd->tx_q_used, &xd->port_conf); + + if (rv < 0) + return clib_error_return (0, "rte_eth_dev_configure[%d]: err %d", + xd->device_index, rv); + + /* Set up one TX-queue per worker thread */ + for (j = 0; j < xd->tx_q_used; j++) + { + rv = rte_eth_tx_queue_setup (xd->device_index, j, xd->nb_tx_desc, + xd->cpu_socket, &xd->tx_conf); + + /* retry with any other CPU socket */ + if (rv < 0) + rv = rte_eth_tx_queue_setup (xd->device_index, j, xd->nb_tx_desc, + SOCKET_ID_ANY, &xd->tx_conf); + if (rv < 0) + break; + } + + if (rv < 0) + return clib_error_return (0, "rte_eth_tx_queue_setup[%d]: err %d", + xd->device_index, rv); + + for (j = 0; j < xd->rx_q_used; j++) + { + + rv = rte_eth_rx_queue_setup (xd->device_index, j, xd->nb_rx_desc, + xd->cpu_socket, 0, + bm-> + pktmbuf_pools[xd->cpu_socket_id_by_queue + [j]]); + + /* retry with any other CPU socket */ + if (rv < 0) + rv = rte_eth_rx_queue_setup (xd->device_index, j, xd->nb_rx_desc, + SOCKET_ID_ANY, 0, + bm-> + pktmbuf_pools[xd->cpu_socket_id_by_queue + [j]]); + if (rv < 0) + return clib_error_return (0, "rte_eth_rx_queue_setup[%d]: err %d", + xd->device_index, rv); + } + + if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) + { + int rv; + rv = rte_eth_dev_start (xd->device_index); + if (rv < 0) + clib_warning ("rte_eth_dev_start %d returned %d", + xd->device_index, rv); + } + return 0; +} + +static u32 +dpdk_flag_change (vnet_main_t * vnm, vnet_hw_interface_t * hi, u32 flags) +{ + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd = vec_elt_at_index (dm->devices, hi->dev_instance); + u32 old = 0; + + if (ETHERNET_INTERFACE_FLAG_CONFIG_PROMISC (flags)) + { + old = (xd->flags & DPDK_DEVICE_FLAG_PROMISC) != 0; + + if (flags & ETHERNET_INTERFACE_FLAG_ACCEPT_ALL) + xd->flags |= DPDK_DEVICE_FLAG_PROMISC; + else + xd->flags &= ~DPDK_DEVICE_FLAG_PROMISC; + + if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) + { + if (xd->flags & DPDK_DEVICE_FLAG_PROMISC) + rte_eth_promiscuous_enable (xd->device_index); + else + rte_eth_promiscuous_disable (xd->device_index); + } + } + else if (ETHERNET_INTERFACE_FLAG_CONFIG_MTU (flags)) + { + /* + * DAW-FIXME: The Cisco VIC firmware does not provide an api for a + * driver to dynamically change the mtu. If/when the + * VIC firmware gets fixed, then this should be removed. + */ + if (xd->pmd == VNET_DPDK_PMD_ENIC) + { + struct rte_eth_dev_info dev_info; + + /* + * Restore mtu to what has been set by CIMC in the firmware cfg. + */ + rte_eth_dev_info_get (xd->device_index, &dev_info); + hi->max_packet_bytes = dev_info.max_rx_pktlen; + + vlib_cli_output (vlib_get_main (), + "Cisco VIC mtu can only be changed " + "using CIMC then rebooting the server!"); + } + else + { + int rv; + + xd->port_conf.rxmode.max_rx_pkt_len = hi->max_packet_bytes; + + if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) + rte_eth_dev_stop (xd->device_index); + + rv = rte_eth_dev_configure + (xd->device_index, xd->rx_q_used, xd->tx_q_used, &xd->port_conf); + + if (rv < 0) + vlib_cli_output (vlib_get_main (), + "rte_eth_dev_configure[%d]: err %d", + xd->device_index, rv); + + rte_eth_dev_set_mtu (xd->device_index, hi->max_packet_bytes); + + if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) + { + int rv = rte_eth_dev_start (xd->device_index); + if (rv < 0) + clib_warning ("rte_eth_dev_start %d returned %d", + xd->device_index, rv); + } + } + } + return old; +} + +void +dpdk_device_lock_init (dpdk_device_t * xd) +{ + int q; + vec_validate (xd->lockp, xd->tx_q_used - 1); + for (q = 0; q < xd->tx_q_used; q++) + { + xd->lockp[q] = clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES, + CLIB_CACHE_LINE_BYTES); + memset ((void *) xd->lockp[q], 0, CLIB_CACHE_LINE_BYTES); + } +} + +void +dpdk_device_lock_free (dpdk_device_t * xd) +{ + int q; + + for (q = 0; q < vec_len (xd->lockp); q++) + clib_mem_free ((void *) xd->lockp[q]); + vec_free (xd->lockp); + xd->lockp = 0; +} + +static clib_error_t * +dpdk_lib_init (dpdk_main_t * dm) +{ + u32 nports; + u32 nb_desc = 0; + int i; + clib_error_t *error; + vlib_main_t *vm = vlib_get_main (); + vlib_thread_main_t *tm = vlib_get_thread_main (); + vnet_sw_interface_t *sw; + vnet_hw_interface_t *hi; + dpdk_device_t *xd; + vlib_pci_addr_t last_pci_addr; + u32 last_pci_addr_port = 0; + vlib_thread_registration_t *tr, *tr_hqos; + uword *p, *p_hqos; + + u32 next_cpu = 0, next_hqos_cpu = 0; + u8 af_packet_port_id = 0; + last_pci_addr.as_u32 = ~0; + + dm->input_cpu_first_index = 0; + dm->input_cpu_count = 1; + + /* find out which cpus will be used for input */ + p = hash_get_mem (tm->thread_registrations_by_name, "workers"); + tr = p ? (vlib_thread_registration_t *) p[0] : 0; + + if (tr && tr->count > 0) + { + dm->input_cpu_first_index = tr->first_index; + dm->input_cpu_count = tr->count; + } + + vec_validate_aligned (dm->devices_by_cpu, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + + vec_validate_aligned (dm->workers, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + + dm->hqos_cpu_first_index = 0; + dm->hqos_cpu_count = 0; + + /* find out which cpus will be used for I/O TX */ + p_hqos = hash_get_mem (tm->thread_registrations_by_name, "hqos-threads"); + tr_hqos = p_hqos ? (vlib_thread_registration_t *) p_hqos[0] : 0; + + if (tr_hqos && tr_hqos->count > 0) + { + dm->hqos_cpu_first_index = tr_hqos->first_index; + dm->hqos_cpu_count = tr_hqos->count; + } + + vec_validate_aligned (dm->devices_by_hqos_cpu, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + + vec_validate_aligned (dm->hqos_threads, tm->n_vlib_mains - 1, + CLIB_CACHE_LINE_BYTES); + + nports = rte_eth_dev_count (); + if (nports < 1) + { + clib_warning ("DPDK drivers found no ports..."); + } + + if (CLIB_DEBUG > 0) + clib_warning ("DPDK drivers found %d ports...", nports); + + /* + * All buffers are all allocated from the same rte_mempool. + * Thus they all have the same number of data bytes. + */ + dm->vlib_buffer_free_list_index = + vlib_buffer_get_or_create_free_list (vm, + VLIB_BUFFER_DEFAULT_FREE_LIST_BYTES, + "dpdk rx"); + + if (dm->conf->enable_tcp_udp_checksum) + dm->buffer_flags_template &= ~(IP_BUFFER_L4_CHECKSUM_CORRECT + | IP_BUFFER_L4_CHECKSUM_COMPUTED); + + for (i = 0; i < nports; i++) + { + u8 addr[6]; + u8 vlan_strip = 0; + int j; + struct rte_eth_dev_info dev_info; + clib_error_t *rv; + struct rte_eth_link l; + dpdk_device_config_t *devconf = 0; + vlib_pci_addr_t pci_addr; + uword *p = 0; + + rte_eth_dev_info_get (i, &dev_info); + if (dev_info.pci_dev) /* bonded interface has no pci info */ + { + pci_addr.domain = dev_info.pci_dev->addr.domain; + pci_addr.bus = dev_info.pci_dev->addr.bus; + pci_addr.slot = dev_info.pci_dev->addr.devid; + pci_addr.function = dev_info.pci_dev->addr.function; + p = + hash_get (dm->conf->device_config_index_by_pci_addr, + pci_addr.as_u32); + } + + if (p) + devconf = pool_elt_at_index (dm->conf->dev_confs, p[0]); + else + devconf = &dm->conf->default_devconf; + + /* Create vnet interface */ + vec_add2_aligned (dm->devices, xd, 1, CLIB_CACHE_LINE_BYTES); + xd->nb_rx_desc = DPDK_NB_RX_DESC_DEFAULT; + xd->nb_tx_desc = DPDK_NB_TX_DESC_DEFAULT; + xd->cpu_socket = (i8) rte_eth_dev_socket_id (i); + + /* Handle interface naming for devices with multiple ports sharing same PCI ID */ + if (dev_info.pci_dev) + { + struct rte_eth_dev_info di = { 0 }; + rte_eth_dev_info_get (i + 1, &di); + if (di.pci_dev && pci_addr.as_u32 != last_pci_addr.as_u32 && + memcmp (&dev_info.pci_dev->addr, &di.pci_dev->addr, + sizeof (struct rte_pci_addr)) == 0) + { + xd->interface_name_suffix = format (0, "0"); + last_pci_addr.as_u32 = pci_addr.as_u32; + last_pci_addr_port = i; + } + else if (pci_addr.as_u32 == last_pci_addr.as_u32) + { + xd->interface_name_suffix = + format (0, "%u", i - last_pci_addr_port); + } + else + { + last_pci_addr.as_u32 = ~0; + } + } + else + last_pci_addr.as_u32 = ~0; + + clib_memcpy (&xd->tx_conf, &dev_info.default_txconf, + sizeof (struct rte_eth_txconf)); + if (dm->conf->no_multi_seg) + { + xd->tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS; + port_conf_template.rxmode.jumbo_frame = 0; + } + else + { + xd->tx_conf.txq_flags &= ~ETH_TXQ_FLAGS_NOMULTSEGS; + port_conf_template.rxmode.jumbo_frame = 1; + xd->flags |= DPDK_DEVICE_FLAG_MAYBE_MULTISEG; + } + + clib_memcpy (&xd->port_conf, &port_conf_template, + sizeof (struct rte_eth_conf)); + + xd->tx_q_used = clib_min (dev_info.max_tx_queues, tm->n_vlib_mains); + + if (devconf->num_tx_queues > 0 + && devconf->num_tx_queues < xd->tx_q_used) + xd->tx_q_used = clib_min (xd->tx_q_used, devconf->num_tx_queues); + + if (devconf->num_rx_queues > 1 && dm->use_rss == 0) + { + dm->use_rss = 1; + } + + if (devconf->num_rx_queues > 1 + && dev_info.max_rx_queues >= devconf->num_rx_queues) + { + xd->rx_q_used = devconf->num_rx_queues; + xd->port_conf.rxmode.mq_mode = ETH_MQ_RX_RSS; + if (devconf->rss_fn == 0) + xd->port_conf.rx_adv_conf.rss_conf.rss_hf = + ETH_RSS_IP | ETH_RSS_UDP | ETH_RSS_TCP; + else + xd->port_conf.rx_adv_conf.rss_conf.rss_hf = devconf->rss_fn; + } + else + xd->rx_q_used = 1; + + xd->flags |= DPDK_DEVICE_FLAG_PMD; + + /* workaround for drivers not setting driver_name */ + if ((!dev_info.driver_name) && (dev_info.pci_dev)) +#if RTE_VERSION < RTE_VERSION_NUM(16, 11, 0, 0) + dev_info.driver_name = dev_info.pci_dev->driver->name; +#else + dev_info.driver_name = dev_info.pci_dev->driver->driver.name; +#endif + ASSERT (dev_info.driver_name); + + if (!xd->pmd) + { + + +#define _(s,f) else if (dev_info.driver_name && \ + !strcmp(dev_info.driver_name, s)) \ + xd->pmd = VNET_DPDK_PMD_##f; + if (0) + ; + foreach_dpdk_pmd +#undef _ + else + xd->pmd = VNET_DPDK_PMD_UNKNOWN; + + xd->port_type = VNET_DPDK_PORT_TYPE_UNKNOWN; + xd->nb_rx_desc = DPDK_NB_RX_DESC_DEFAULT; + xd->nb_tx_desc = DPDK_NB_TX_DESC_DEFAULT; + + switch (xd->pmd) + { + /* 1G adapters */ + case VNET_DPDK_PMD_E1000EM: + case VNET_DPDK_PMD_IGB: + case VNET_DPDK_PMD_IGBVF: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_1G; + break; + + /* 10G adapters */ + case VNET_DPDK_PMD_IXGBE: + case VNET_DPDK_PMD_IXGBEVF: + case VNET_DPDK_PMD_THUNDERX: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + case VNET_DPDK_PMD_DPAA2: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + + /* Cisco VIC */ + case VNET_DPDK_PMD_ENIC: + rte_eth_link_get_nowait (i, &l); + xd->flags |= DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE; + if (l.link_speed == 40000) + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G; + else + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + + /* Intel Fortville */ + case VNET_DPDK_PMD_I40E: + case VNET_DPDK_PMD_I40EVF: + xd->flags |= DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE; + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G; + + switch (dev_info.pci_dev->id.device_id) + { + case I40E_DEV_ID_10G_BASE_T: + case I40E_DEV_ID_SFP_XL710: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + case I40E_DEV_ID_QSFP_A: + case I40E_DEV_ID_QSFP_B: + case I40E_DEV_ID_QSFP_C: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G; + break; + case I40E_DEV_ID_VF: + rte_eth_link_get_nowait (i, &l); + xd->port_type = l.link_speed == 10000 ? + VNET_DPDK_PORT_TYPE_ETH_10G : VNET_DPDK_PORT_TYPE_ETH_40G; + break; + default: + xd->port_type = VNET_DPDK_PORT_TYPE_UNKNOWN; + } + break; + + case VNET_DPDK_PMD_CXGBE: + switch (dev_info.pci_dev->id.device_id) + { + case 0x540d: /* T580-CR */ + case 0x5410: /* T580-LP-cr */ + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G; + break; + case 0x5403: /* T540-CR */ + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + default: + xd->port_type = VNET_DPDK_PORT_TYPE_UNKNOWN; + } + break; + + case VNET_DPDK_PMD_MLX5: + { + char *pn_100g[] = { "MCX415A-CCAT", "MCX416A-CCAT", 0 }; + char *pn_40g[] = { "MCX413A-BCAT", "MCX414A-BCAT", + "MCX415A-BCAT", "MCX416A-BCAT", "MCX4131A-BCAT", 0 + }; + char *pn_10g[] = { "MCX4111A-XCAT", "MCX4121A-XCAT", 0 }; + + vlib_pci_device_t *pd = vlib_get_pci_device (&pci_addr); + u8 *pn = 0; + char **c; + int found = 0; + pn = format (0, "%U%c", + format_vlib_pci_vpd, pd->vpd_r, "PN", 0); + + if (!pn) + break; + + c = pn_100g; + while (!found && c[0]) + { + if (strncmp ((char *) pn, c[0], strlen (c[0])) == 0) + { + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_100G; + break; + } + c++; + } + + c = pn_40g; + while (!found && c[0]) + { + if (strncmp ((char *) pn, c[0], strlen (c[0])) == 0) + { + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G; + break; + } + c++; + } + + c = pn_10g; + while (!found && c[0]) + { + if (strncmp ((char *) pn, c[0], strlen (c[0])) == 0) + { + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G; + break; + } + c++; + } + + vec_free (pn); + } + + break; + /* Intel Red Rock Canyon */ + case VNET_DPDK_PMD_FM10K: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_SWITCH; + break; + + /* virtio */ + case VNET_DPDK_PMD_VIRTIO: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_1G; + xd->nb_rx_desc = DPDK_NB_RX_DESC_VIRTIO; + xd->nb_tx_desc = DPDK_NB_TX_DESC_VIRTIO; + break; + + /* vmxnet3 */ + case VNET_DPDK_PMD_VMXNET3: + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_1G; + xd->tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS; + break; + + case VNET_DPDK_PMD_AF_PACKET: + xd->port_type = VNET_DPDK_PORT_TYPE_AF_PACKET; + xd->af_packet_port_id = af_packet_port_id++; + break; + + case VNET_DPDK_PMD_BOND: + xd->flags |= DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE; + xd->port_type = VNET_DPDK_PORT_TYPE_ETH_BOND; + break; + + default: + xd->port_type = VNET_DPDK_PORT_TYPE_UNKNOWN; + } + + if (devconf->num_rx_desc) + xd->nb_rx_desc = devconf->num_rx_desc; + + if (devconf->num_tx_desc) + xd->nb_tx_desc = devconf->num_tx_desc; + } + + /* + * Ensure default mtu is not > the mtu read from the hardware. + * Otherwise rte_eth_dev_configure() will fail and the port will + * not be available. + */ + if (ETHERNET_MAX_PACKET_BYTES > dev_info.max_rx_pktlen) + { + /* + * This device does not support the platforms's max frame + * size. Use it's advertised mru instead. + */ + xd->port_conf.rxmode.max_rx_pkt_len = dev_info.max_rx_pktlen; + } + else + { + xd->port_conf.rxmode.max_rx_pkt_len = ETHERNET_MAX_PACKET_BYTES; + + /* + * Some platforms do not account for Ethernet FCS (4 bytes) in + * MTU calculations. To interop with them increase mru but only + * if the device's settings can support it. + */ + if ((dev_info.max_rx_pktlen >= (ETHERNET_MAX_PACKET_BYTES + 4)) && + xd->port_conf.rxmode.hw_strip_crc) + { + /* + * Allow additional 4 bytes (for Ethernet FCS). These bytes are + * stripped by h/w and so will not consume any buffer memory. + */ + xd->port_conf.rxmode.max_rx_pkt_len += 4; + } + } + + if (xd->pmd == VNET_DPDK_PMD_AF_PACKET) + { + f64 now = vlib_time_now (vm); + u32 rnd; + rnd = (u32) (now * 1e6); + rnd = random_u32 (&rnd); + clib_memcpy (addr + 2, &rnd, sizeof (rnd)); + addr[0] = 2; + addr[1] = 0xfe; + } + else + rte_eth_macaddr_get (i, (struct ether_addr *) addr); + + if (xd->tx_q_used < tm->n_vlib_mains) + dpdk_device_lock_init (xd); + + xd->device_index = xd - dm->devices; + ASSERT (i == xd->device_index); + xd->per_interface_next_index = ~0; + + /* assign interface to input thread */ + dpdk_device_and_queue_t *dq; + int q; + + if (devconf->workers) + { + int i; + q = 0; + /* *INDENT-OFF* */ + clib_bitmap_foreach (i, devconf->workers, ({ + int cpu = dm->input_cpu_first_index + i; + unsigned lcore = vlib_worker_threads[cpu].lcore_id; + vec_validate(xd->cpu_socket_id_by_queue, q); + xd->cpu_socket_id_by_queue[q] = rte_lcore_to_socket_id(lcore); + vec_add2(dm->devices_by_cpu[cpu], dq, 1); + dq->device = xd->device_index; + dq->queue_id = q++; + })); + /* *INDENT-ON* */ + } + else + for (q = 0; q < xd->rx_q_used; q++) + { + int cpu = dm->input_cpu_first_index + next_cpu; + unsigned lcore = vlib_worker_threads[cpu].lcore_id; + + /* + * numa node for worker thread handling this queue + * needed for taking buffers from the right mempool + */ + vec_validate (xd->cpu_socket_id_by_queue, q); + xd->cpu_socket_id_by_queue[q] = rte_lcore_to_socket_id (lcore); + + /* + * construct vector of (device,queue) pairs for each worker thread + */ + vec_add2 (dm->devices_by_cpu[cpu], dq, 1); + dq->device = xd->device_index; + dq->queue_id = q; + + next_cpu++; + if (next_cpu == dm->input_cpu_count) + next_cpu = 0; + } + + + if (devconf->hqos_enabled) + { + xd->flags |= DPDK_DEVICE_FLAG_HQOS; + + if (devconf->hqos.hqos_thread_valid) + { + int cpu = dm->hqos_cpu_first_index + devconf->hqos.hqos_thread; + + if (devconf->hqos.hqos_thread >= dm->hqos_cpu_count) + return clib_error_return (0, "invalid HQoS thread index"); + + vec_add2 (dm->devices_by_hqos_cpu[cpu], dq, 1); + dq->device = xd->device_index; + dq->queue_id = 0; + } + else + { + int cpu = dm->hqos_cpu_first_index + next_hqos_cpu; + + if (dm->hqos_cpu_count == 0) + return clib_error_return (0, "no HQoS threads available"); + + vec_add2 (dm->devices_by_hqos_cpu[cpu], dq, 1); + dq->device = xd->device_index; + dq->queue_id = 0; + + next_hqos_cpu++; + if (next_hqos_cpu == dm->hqos_cpu_count) + next_hqos_cpu = 0; + + devconf->hqos.hqos_thread_valid = 1; + devconf->hqos.hqos_thread = cpu; + } + } + + vec_validate_aligned (xd->tx_vectors, tm->n_vlib_mains, + CLIB_CACHE_LINE_BYTES); + for (j = 0; j < tm->n_vlib_mains; j++) + { + vec_validate_ha (xd->tx_vectors[j], xd->nb_tx_desc, + sizeof (tx_ring_hdr_t), CLIB_CACHE_LINE_BYTES); + vec_reset_length (xd->tx_vectors[j]); + } + + vec_validate_aligned (xd->rx_vectors, xd->rx_q_used, + CLIB_CACHE_LINE_BYTES); + for (j = 0; j < xd->rx_q_used; j++) + { + vec_validate_aligned (xd->rx_vectors[j], VLIB_FRAME_SIZE - 1, + CLIB_CACHE_LINE_BYTES); + vec_reset_length (xd->rx_vectors[j]); + } + + vec_validate_aligned (xd->d_trace_buffers, tm->n_vlib_mains, + CLIB_CACHE_LINE_BYTES); + + rv = dpdk_port_setup (dm, xd); + + if (rv) + return rv; + + if (devconf->hqos_enabled) + { + rv = dpdk_port_setup_hqos (xd, &devconf->hqos); + if (rv) + return rv; + } + + /* count the number of descriptors used for this device */ + nb_desc += xd->nb_rx_desc + xd->nb_tx_desc * xd->tx_q_used; + + error = ethernet_register_interface + (dm->vnet_main, dpdk_device_class.index, xd->device_index, + /* ethernet address */ addr, + &xd->vlib_hw_if_index, dpdk_flag_change); + if (error) + return error; + + sw = vnet_get_hw_sw_interface (dm->vnet_main, xd->vlib_hw_if_index); + xd->vlib_sw_if_index = sw->sw_if_index; + hi = vnet_get_hw_interface (dm->vnet_main, xd->vlib_hw_if_index); + + /* + * DAW-FIXME: The Cisco VIC firmware does not provide an api for a + * driver to dynamically change the mtu. If/when the + * VIC firmware gets fixed, then this should be removed. + */ + if (xd->pmd == VNET_DPDK_PMD_ENIC) + { + /* + * Initialize mtu to what has been set by CIMC in the firmware cfg. + */ + hi->max_packet_bytes = dev_info.max_rx_pktlen; + if (devconf->vlan_strip_offload != DPDK_DEVICE_VLAN_STRIP_OFF) + vlan_strip = 1; /* remove vlan tag from VIC port by default */ + else + clib_warning ("VLAN strip disabled for interface\n"); + } + else if (devconf->vlan_strip_offload == DPDK_DEVICE_VLAN_STRIP_ON) + vlan_strip = 1; + + if (vlan_strip) + { + int vlan_off; + vlan_off = rte_eth_dev_get_vlan_offload (xd->device_index); + vlan_off |= ETH_VLAN_STRIP_OFFLOAD; + xd->port_conf.rxmode.hw_vlan_strip = vlan_off; + if (rte_eth_dev_set_vlan_offload (xd->device_index, vlan_off) == 0) + clib_warning ("VLAN strip enabled for interface\n"); + else + clib_warning ("VLAN strip cannot be supported by interface\n"); + } + + hi->max_l3_packet_bytes[VLIB_RX] = hi->max_l3_packet_bytes[VLIB_TX] = + xd->port_conf.rxmode.max_rx_pkt_len - sizeof (ethernet_header_t); + + rte_eth_dev_set_mtu (xd->device_index, hi->max_packet_bytes); + } + + if (nb_desc > dm->conf->num_mbufs) + clib_warning ("%d mbufs allocated but total rx/tx ring size is %d\n", + dm->conf->num_mbufs, nb_desc); + + return 0; +} + +static void +dpdk_bind_devices_to_uio (dpdk_config_main_t * conf) +{ + vlib_pci_main_t *pm = &pci_main; + clib_error_t *error; + vlib_pci_device_t *d; + u8 *pci_addr = 0; + int num_whitelisted = vec_len (conf->dev_confs); + + /* *INDENT-OFF* */ + pool_foreach (d, pm->pci_devs, ({ + dpdk_device_config_t * devconf = 0; + vec_reset_length (pci_addr); + pci_addr = format (pci_addr, "%U%c", format_vlib_pci_addr, &d->bus_address, 0); + + if (d->device_class != PCI_CLASS_NETWORK_ETHERNET) + continue; + + if (num_whitelisted) + { + uword * p = hash_get (conf->device_config_index_by_pci_addr, d->bus_address.as_u32); + + if (!p) + continue; + + devconf = pool_elt_at_index (conf->dev_confs, p[0]); + } + + /* virtio */ + if (d->vendor_id == 0x1af4 && d->device_id == 0x1000) + ; + /* vmxnet3 */ + else if (d->vendor_id == 0x15ad && d->device_id == 0x07b0) + ; + /* all Intel devices */ + else if (d->vendor_id == 0x8086) + ; + /* Cisco VIC */ + else if (d->vendor_id == 0x1137 && d->device_id == 0x0043) + ; + /* Chelsio T4/T5 */ + else if (d->vendor_id == 0x1425 && (d->device_id & 0xe000) == 0x4000) + ; + else + { + clib_warning ("Unsupported Ethernet PCI device 0x%04x:0x%04x found " + "at PCI address %s\n", (u16) d->vendor_id, (u16) d->device_id, + pci_addr); + continue; + } + + error = vlib_pci_bind_to_uio (d, (char *) conf->uio_driver_name); + + if (error) + { + if (devconf == 0) + { + pool_get (conf->dev_confs, devconf); + hash_set (conf->device_config_index_by_pci_addr, d->bus_address.as_u32, + devconf - conf->dev_confs); + devconf->pci_addr.as_u32 = d->bus_address.as_u32; + } + devconf->is_blacklisted = 1; + clib_error_report (error); + } + })); + /* *INDENT-ON* */ + vec_free (pci_addr); +} + +static clib_error_t * +dpdk_device_config (dpdk_config_main_t * conf, vlib_pci_addr_t pci_addr, + unformat_input_t * input, u8 is_default) +{ + clib_error_t *error = 0; + uword *p; + dpdk_device_config_t *devconf; + unformat_input_t sub_input; + + if (is_default) + { + devconf = &conf->default_devconf; + } + else + { + p = hash_get (conf->device_config_index_by_pci_addr, pci_addr.as_u32); + + if (!p) + { + pool_get (conf->dev_confs, devconf); + hash_set (conf->device_config_index_by_pci_addr, pci_addr.as_u32, + devconf - conf->dev_confs); + } + else + return clib_error_return (0, + "duplicate configuration for PCI address %U", + format_vlib_pci_addr, &pci_addr); + } + + devconf->pci_addr.as_u32 = pci_addr.as_u32; + devconf->hqos_enabled = 0; + dpdk_device_config_hqos_default (&devconf->hqos); + + if (!input) + return 0; + + unformat_skip_white_space (input); + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "num-rx-queues %u", &devconf->num_rx_queues)) + ; + else if (unformat (input, "num-tx-queues %u", &devconf->num_tx_queues)) + ; + else if (unformat (input, "num-rx-desc %u", &devconf->num_rx_desc)) + ; + else if (unformat (input, "num-tx-desc %u", &devconf->num_tx_desc)) + ; + else if (unformat (input, "workers %U", unformat_bitmap_list, + &devconf->workers)) + ; + else + if (unformat + (input, "rss %U", unformat_vlib_cli_sub_input, &sub_input)) + { + error = unformat_rss_fn (&sub_input, &devconf->rss_fn); + if (error) + break; + } + else if (unformat (input, "vlan-strip-offload off")) + devconf->vlan_strip_offload = DPDK_DEVICE_VLAN_STRIP_OFF; + else if (unformat (input, "vlan-strip-offload on")) + devconf->vlan_strip_offload = DPDK_DEVICE_VLAN_STRIP_ON; + else + if (unformat + (input, "hqos %U", unformat_vlib_cli_sub_input, &sub_input)) + { + devconf->hqos_enabled = 1; + error = unformat_hqos (&sub_input, &devconf->hqos); + if (error) + break; + } + else if (unformat (input, "hqos")) + { + devconf->hqos_enabled = 1; + } + else + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + break; + } + } + + if (error) + return error; + + if (devconf->workers && devconf->num_rx_queues == 0) + devconf->num_rx_queues = clib_bitmap_count_set_bits (devconf->workers); + else if (devconf->workers && + clib_bitmap_count_set_bits (devconf->workers) != + devconf->num_rx_queues) + error = + clib_error_return (0, + "%U: number of worker threadds must be " + "equal to number of rx queues", format_vlib_pci_addr, + &pci_addr); + + return error; +} + +static clib_error_t * +dpdk_config (vlib_main_t * vm, unformat_input_t * input) +{ + clib_error_t *error = 0; + dpdk_main_t *dm = &dpdk_main; + dpdk_config_main_t *conf = &dpdk_config_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); + dpdk_device_config_t *devconf; + vlib_pci_addr_t pci_addr; + unformat_input_t sub_input; + u8 *s, *tmp = 0; + u8 *rte_cmd = 0, *ethname = 0; + u32 log_level; + int ret, i; + int num_whitelisted = 0; + u8 no_pci = 0; + u8 no_huge = 0; + u8 huge_dir = 0; + u8 file_prefix = 0; + u8 *socket_mem = 0; + + conf->device_config_index_by_pci_addr = hash_create (0, sizeof (uword)); + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + /* Prime the pump */ + if (unformat (input, "no-hugetlb")) + { + vec_add1 (conf->eal_init_args, (u8 *) "no-huge"); + no_huge = 1; + } + + else if (unformat (input, "enable-tcp-udp-checksum")) + conf->enable_tcp_udp_checksum = 1; + + else if (unformat (input, "decimal-interface-names")) + conf->interface_name_format_decimal = 1; + + else if (unformat (input, "no-multi-seg")) + conf->no_multi_seg = 1; + + else if (unformat (input, "dev default %U", unformat_vlib_cli_sub_input, + &sub_input)) + { + error = + dpdk_device_config (conf, (vlib_pci_addr_t) (u32) ~ 1, &sub_input, + 1); + + if (error) + return error; + } + else + if (unformat + (input, "dev %U %U", unformat_vlib_pci_addr, &pci_addr, + unformat_vlib_cli_sub_input, &sub_input)) + { + error = dpdk_device_config (conf, pci_addr, &sub_input, 0); + + if (error) + return error; + + num_whitelisted++; + } + else if (unformat (input, "dev %U", unformat_vlib_pci_addr, &pci_addr)) + { + error = dpdk_device_config (conf, pci_addr, 0, 0); + + if (error) + return error; + + num_whitelisted++; + } + else if (unformat (input, "num-mbufs %d", &conf->num_mbufs)) + ; + else if (unformat (input, "kni %d", &conf->num_kni)) + ; + else if (unformat (input, "uio-driver %s", &conf->uio_driver_name)) + ; + else if (unformat (input, "socket-mem %s", &socket_mem)) + ; + else if (unformat (input, "no-pci")) + { + no_pci = 1; + tmp = format (0, "--no-pci%c", 0); + vec_add1 (conf->eal_init_args, tmp); + } + else if (unformat (input, "poll-sleep %d", &dm->poll_sleep)) + ; + +#define _(a) \ + else if (unformat(input, #a)) \ + { \ + tmp = format (0, "--%s%c", #a, 0); \ + vec_add1 (conf->eal_init_args, tmp); \ + } + foreach_eal_double_hyphen_predicate_arg +#undef _ +#define _(a) \ + else if (unformat(input, #a " %s", &s)) \ + { \ + if (!strncmp(#a, "huge-dir", 8)) \ + huge_dir = 1; \ + else if (!strncmp(#a, "file-prefix", 11)) \ + file_prefix = 1; \ + tmp = format (0, "--%s%c", #a, 0); \ + vec_add1 (conf->eal_init_args, tmp); \ + vec_add1 (s, 0); \ + vec_add1 (conf->eal_init_args, s); \ + } + foreach_eal_double_hyphen_arg +#undef _ +#define _(a,b) \ + else if (unformat(input, #a " %s", &s)) \ + { \ + tmp = format (0, "-%s%c", #b, 0); \ + vec_add1 (conf->eal_init_args, tmp); \ + vec_add1 (s, 0); \ + vec_add1 (conf->eal_init_args, s); \ + } + foreach_eal_single_hyphen_arg +#undef _ +#define _(a,b) \ + else if (unformat(input, #a " %s", &s)) \ + { \ + tmp = format (0, "-%s%c", #b, 0); \ + vec_add1 (conf->eal_init_args, tmp); \ + vec_add1 (s, 0); \ + vec_add1 (conf->eal_init_args, s); \ + conf->a##_set_manually = 1; \ + } + foreach_eal_single_hyphen_mandatory_arg +#undef _ + else if (unformat (input, "default")) + ; + + else if (unformat_skip_white_space (input)) + ; + else + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + goto done; + } + } + + if (!conf->uio_driver_name) + conf->uio_driver_name = format (0, "igb_uio%c", 0); + + /* + * Use 1G huge pages if available. + */ + if (!no_huge && !huge_dir) + { + u32 x, *mem_by_socket = 0; + uword c = 0; + u8 use_1g = 1; + u8 use_2m = 1; + u8 less_than_1g = 1; + int rv; + + umount (DEFAULT_HUGE_DIR); + + /* Process "socket-mem" parameter value */ + if (vec_len (socket_mem)) + { + unformat_input_t in; + unformat_init_vector (&in, socket_mem); + while (unformat_check_input (&in) != UNFORMAT_END_OF_INPUT) + { + if (unformat (&in, "%u,", &x)) + ; + else if (unformat (&in, "%u", &x)) + ; + else if (unformat (&in, ",")) + x = 0; + else + break; + + vec_add1 (mem_by_socket, x); + + if (x > 1023) + less_than_1g = 0; + } + /* Note: unformat_free vec_frees(in.buffer), aka socket_mem... */ + unformat_free (&in); + socket_mem = 0; + } + else + { + /* *INDENT-OFF* */ + clib_bitmap_foreach (c, tm->cpu_socket_bitmap, ( + { + vec_validate(mem_by_socket, c); + mem_by_socket[c] = 256; /* default per-socket mem */ + } + )); + /* *INDENT-ON* */ + } + + /* check if available enough 1GB pages for each socket */ + /* *INDENT-OFF* */ + clib_bitmap_foreach (c, tm->cpu_socket_bitmap, ( + { + int pages_avail, page_size, mem; + + vec_validate(mem_by_socket, c); + mem = mem_by_socket[c]; + + page_size = 1024; + pages_avail = vlib_sysfs_get_free_hugepages(c, page_size * 1024); + + if (pages_avail < 0 || page_size * pages_avail < mem) + use_1g = 0; + + page_size = 2; + pages_avail = vlib_sysfs_get_free_hugepages(c, page_size * 1024); + + if (pages_avail < 0 || page_size * pages_avail < mem) + use_2m = 0; + })); + /* *INDENT-ON* */ + + if (mem_by_socket == 0) + { + error = clib_error_return (0, "mem_by_socket NULL"); + goto done; + } + _vec_len (mem_by_socket) = c + 1; + + /* regenerate socket_mem string */ + vec_foreach_index (x, mem_by_socket) + socket_mem = format (socket_mem, "%s%u", + socket_mem ? "," : "", mem_by_socket[x]); + socket_mem = format (socket_mem, "%c", 0); + + vec_free (mem_by_socket); + + rv = mkdir (VPP_RUN_DIR, 0755); + if (rv && errno != EEXIST) + { + error = clib_error_return (0, "mkdir '%s' failed errno %d", + VPP_RUN_DIR, errno); + goto done; + } + + rv = mkdir (DEFAULT_HUGE_DIR, 0755); + if (rv && errno != EEXIST) + { + error = clib_error_return (0, "mkdir '%s' failed errno %d", + DEFAULT_HUGE_DIR, errno); + goto done; + } + + if (use_1g && !(less_than_1g && use_2m)) + { + rv = + mount ("none", DEFAULT_HUGE_DIR, "hugetlbfs", 0, "pagesize=1G"); + } + else if (use_2m) + { + rv = mount ("none", DEFAULT_HUGE_DIR, "hugetlbfs", 0, NULL); + } + else + { + return clib_error_return (0, "not enough free huge pages"); + } + + if (rv) + { + error = clib_error_return (0, "mount failed %d", errno); + goto done; + } + + tmp = format (0, "--huge-dir%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "%s%c", DEFAULT_HUGE_DIR, 0); + vec_add1 (conf->eal_init_args, tmp); + if (!file_prefix) + { + tmp = format (0, "--file-prefix%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "vpp%c", 0); + vec_add1 (conf->eal_init_args, tmp); + } + } + + vec_free (rte_cmd); + vec_free (ethname); + + if (error) + return error; + + /* I'll bet that -c and -n must be the first and second args... */ + if (!conf->coremask_set_manually) + { + vlib_thread_registration_t *tr; + uword *coremask = 0; + int i; + + /* main thread core */ + coremask = clib_bitmap_set (coremask, tm->main_lcore, 1); + + for (i = 0; i < vec_len (tm->registrations); i++) + { + tr = tm->registrations[i]; + coremask = clib_bitmap_or (coremask, tr->coremask); + } + + vec_insert (conf->eal_init_args, 2, 1); + conf->eal_init_args[1] = (u8 *) "-c"; + tmp = format (0, "%U%c", format_bitmap_hex, coremask, 0); + conf->eal_init_args[2] = tmp; + clib_bitmap_free (coremask); + } + + if (!conf->nchannels_set_manually) + { + vec_insert (conf->eal_init_args, 2, 3); + conf->eal_init_args[3] = (u8 *) "-n"; + tmp = format (0, "%d", conf->nchannels); + conf->eal_init_args[4] = tmp; + } + + if (no_pci == 0 && geteuid () == 0) + dpdk_bind_devices_to_uio (conf); + +#define _(x) \ + if (devconf->x == 0 && conf->default_devconf.x > 0) \ + devconf->x = conf->default_devconf.x ; + + /* *INDENT-OFF* */ + pool_foreach (devconf, conf->dev_confs, ({ + + /* default per-device config items */ + foreach_dpdk_device_config_item + + /* add DPDK EAL whitelist/blacklist entry */ + if (num_whitelisted > 0 && devconf->is_blacklisted == 0) + { + tmp = format (0, "-w%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "%U%c", format_vlib_pci_addr, &devconf->pci_addr, 0); + vec_add1 (conf->eal_init_args, tmp); + } + else if (num_whitelisted == 0 && devconf->is_blacklisted != 0) + { + tmp = format (0, "-b%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "%U%c", format_vlib_pci_addr, &devconf->pci_addr, 0); + vec_add1 (conf->eal_init_args, tmp); + } + })); + /* *INDENT-ON* */ + +#undef _ + + /* set master-lcore */ + tmp = format (0, "--master-lcore%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "%u%c", tm->main_lcore, 0); + vec_add1 (conf->eal_init_args, tmp); + + /* set socket-mem */ + tmp = format (0, "--socket-mem%c", 0); + vec_add1 (conf->eal_init_args, tmp); + tmp = format (0, "%s%c", socket_mem, 0); + vec_add1 (conf->eal_init_args, tmp); + + /* NULL terminate the "argv" vector, in case of stupidity */ + vec_add1 (conf->eal_init_args, 0); + _vec_len (conf->eal_init_args) -= 1; + + /* Set up DPDK eal and packet mbuf pool early. */ + + log_level = (CLIB_DEBUG > 0) ? RTE_LOG_DEBUG : RTE_LOG_NOTICE; + + rte_set_log_level (log_level); + + vm = vlib_get_main (); + + /* make copy of args as rte_eal_init tends to mess up with arg array */ + for (i = 1; i < vec_len (conf->eal_init_args); i++) + conf->eal_init_args_str = format (conf->eal_init_args_str, "%s ", + conf->eal_init_args[i]); + + ret = + rte_eal_init (vec_len (conf->eal_init_args), + (char **) conf->eal_init_args); + + /* lazy umount hugepages */ + umount2 (DEFAULT_HUGE_DIR, MNT_DETACH); + + if (ret < 0) + return clib_error_return (0, "rte_eal_init returned %d", ret); + + /* Dump the physical memory layout prior to creating the mbuf_pool */ + fprintf (stdout, "DPDK physical memory layout:\n"); + rte_dump_physmem_layout (stdout); + + /* main thread 1st */ + error = vlib_buffer_pool_create (vm, conf->num_mbufs, rte_socket_id ()); + if (error) + return error; + + for (i = 0; i < RTE_MAX_LCORE; i++) + { + error = vlib_buffer_pool_create (vm, conf->num_mbufs, + rte_lcore_to_socket_id (i)); + if (error) + return error; + } + +done: + return error; +} + +VLIB_CONFIG_FUNCTION (dpdk_config, "dpdk"); + +void +dpdk_update_link_state (dpdk_device_t * xd, f64 now) +{ + vnet_main_t *vnm = vnet_get_main (); + struct rte_eth_link prev_link = xd->link; + u32 hw_flags = 0; + u8 hw_flags_chg = 0; + + /* only update link state for PMD interfaces */ + if ((xd->flags & DPDK_DEVICE_FLAG_PMD) == 0) + return; + + xd->time_last_link_update = now ? now : xd->time_last_link_update; + memset (&xd->link, 0, sizeof (xd->link)); + rte_eth_link_get_nowait (xd->device_index, &xd->link); + + if (LINK_STATE_ELOGS) + { + vlib_main_t *vm = vlib_get_main (); + ELOG_TYPE_DECLARE (e) = + { + .format = + "update-link-state: sw_if_index %d, admin_up %d," + "old link_state %d new link_state %d",.format_args = "i4i1i1i1",}; + + struct + { + u32 sw_if_index; + u8 admin_up; + u8 old_link_state; + u8 new_link_state; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->sw_if_index = xd->vlib_sw_if_index; + ed->admin_up = (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) != 0; + ed->old_link_state = (u8) + vnet_hw_interface_is_link_up (vnm, xd->vlib_hw_if_index); + ed->new_link_state = (u8) xd->link.link_status; + } + + if ((xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) && + ((xd->link.link_status != 0) ^ + vnet_hw_interface_is_link_up (vnm, xd->vlib_hw_if_index))) + { + hw_flags_chg = 1; + hw_flags |= (xd->link.link_status ? VNET_HW_INTERFACE_FLAG_LINK_UP : 0); + } + + if (hw_flags_chg || (xd->link.link_duplex != prev_link.link_duplex)) + { + hw_flags_chg = 1; + switch (xd->link.link_duplex) + { + case ETH_LINK_HALF_DUPLEX: + hw_flags |= VNET_HW_INTERFACE_FLAG_HALF_DUPLEX; + break; + case ETH_LINK_FULL_DUPLEX: + hw_flags |= VNET_HW_INTERFACE_FLAG_FULL_DUPLEX; + break; + default: + break; + } + } + if (hw_flags_chg || (xd->link.link_speed != prev_link.link_speed)) + { + hw_flags_chg = 1; + switch (xd->link.link_speed) + { + case ETH_SPEED_NUM_10M: + hw_flags |= VNET_HW_INTERFACE_FLAG_SPEED_10M; + break; + case ETH_SPEED_NUM_100M: + hw_flags |= VNET_HW_INTERFACE_FLAG_SPEED_100M; + break; + case ETH_SPEED_NUM_1G: + hw_flags |= VNET_HW_INTERFACE_FLAG_SPEED_1G; + break; + case ETH_SPEED_NUM_10G: + hw_flags |= VNET_HW_INTERFACE_FLAG_SPEED_10G; + break; + case ETH_SPEED_NUM_40G: + hw_flags |= VNET_HW_INTERFACE_FLAG_SPEED_40G; + break; + case 0: + break; + default: + clib_warning ("unknown link speed %d", xd->link.link_speed); + break; + } + } + if (hw_flags_chg) + { + if (LINK_STATE_ELOGS) + { + vlib_main_t *vm = vlib_get_main (); + + ELOG_TYPE_DECLARE (e) = + { + .format = + "update-link-state: sw_if_index %d, new flags %d",.format_args + = "i4i4",}; + + struct + { + u32 sw_if_index; + u32 flags; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->sw_if_index = xd->vlib_sw_if_index; + ed->flags = hw_flags; + } + vnet_hw_interface_set_flags (vnm, xd->vlib_hw_if_index, hw_flags); + } +} + +static uword +dpdk_process (vlib_main_t * vm, vlib_node_runtime_t * rt, vlib_frame_t * f) +{ + clib_error_t *error; + vnet_main_t *vnm = vnet_get_main (); + dpdk_main_t *dm = &dpdk_main; + ethernet_main_t *em = ðernet_main; + dpdk_device_t *xd; + vlib_thread_main_t *tm = vlib_get_thread_main (); + int i; + + error = dpdk_lib_init (dm); + + /* + * Turn on the input node if we found some devices to drive + * and we're not running worker threads or i/o threads + */ + + if (error == 0 && vec_len (dm->devices) > 0) + { + if (tm->n_vlib_mains == 1) + vlib_node_set_state (vm, dpdk_input_node.index, + VLIB_NODE_STATE_POLLING); + else + for (i = 0; i < tm->n_vlib_mains; i++) + if (vec_len (dm->devices_by_cpu[i]) > 0) + vlib_node_set_state (vlib_mains[i], dpdk_input_node.index, + VLIB_NODE_STATE_POLLING); + } + + if (error) + clib_error_report (error); + + tm->worker_thread_release = 1; + + f64 now = vlib_time_now (vm); + vec_foreach (xd, dm->devices) + { + dpdk_update_link_state (xd, now); + } + + { + /* + * Extra set up for bond interfaces: + * 1. Setup MACs for bond interfaces and their slave links which was set + * in dpdk_port_setup() but needs to be done again here to take effect. + * 2. Set up info for bond interface related CLI support. + */ + int nports = rte_eth_dev_count (); + if (nports > 0) + { + for (i = 0; i < nports; i++) + { + struct rte_eth_dev_info dev_info; + rte_eth_dev_info_get (i, &dev_info); + if (!dev_info.driver_name) +#if RTE_VERSION < RTE_VERSION_NUM(16, 11, 0, 0) + dev_info.driver_name = dev_info.pci_dev->driver->name; +#else + dev_info.driver_name = dev_info.pci_dev->driver->driver.name; +#endif + ASSERT (dev_info.driver_name); + if (strncmp (dev_info.driver_name, "rte_bond_pmd", 12) == 0) + { + u8 addr[6]; + u8 slink[16]; + int nlink = rte_eth_bond_slaves_get (i, slink, 16); + if (nlink > 0) + { + vnet_hw_interface_t *bhi; + ethernet_interface_t *bei; + int rv; + + /* Get MAC of 1st slave link */ + rte_eth_macaddr_get (slink[0], + (struct ether_addr *) addr); + /* Set MAC of bounded interface to that of 1st slave link */ + rv = + rte_eth_bond_mac_address_set (i, + (struct ether_addr *) + addr); + if (rv < 0) + clib_warning ("Failed to set MAC address"); + + /* Populate MAC of bonded interface in VPP hw tables */ + bhi = + vnet_get_hw_interface (vnm, + dm->devices[i].vlib_hw_if_index); + bei = + pool_elt_at_index (em->interfaces, bhi->hw_instance); + clib_memcpy (bhi->hw_address, addr, 6); + clib_memcpy (bei->address, addr, 6); + /* Init l3 packet size allowed on bonded interface */ + bhi->max_packet_bytes = ETHERNET_MAX_PACKET_BYTES; + bhi->max_l3_packet_bytes[VLIB_RX] = + bhi->max_l3_packet_bytes[VLIB_TX] = + ETHERNET_MAX_PACKET_BYTES - sizeof (ethernet_header_t); + while (nlink >= 1) + { /* for all slave links */ + int slave = slink[--nlink]; + dpdk_device_t *sdev = &dm->devices[slave]; + vnet_hw_interface_t *shi; + vnet_sw_interface_t *ssi; + /* Add MAC to all slave links except the first one */ + if (nlink) + rte_eth_dev_mac_addr_add (slave, + (struct ether_addr *) + addr, 0); + /* Set slaves bitmap for bonded interface */ + bhi->bond_info = + clib_bitmap_set (bhi->bond_info, + sdev->vlib_hw_if_index, 1); + /* Set slave link flags on slave interface */ + shi = + vnet_get_hw_interface (vnm, sdev->vlib_hw_if_index); + ssi = + vnet_get_sw_interface (vnm, sdev->vlib_sw_if_index); + shi->bond_info = VNET_HW_INTERFACE_BOND_INFO_SLAVE; + ssi->flags |= VNET_SW_INTERFACE_FLAG_BOND_SLAVE; + + /* Set l3 packet size allowed as the lowest of slave */ + if (bhi->max_l3_packet_bytes[VLIB_RX] > + shi->max_l3_packet_bytes[VLIB_RX]) + bhi->max_l3_packet_bytes[VLIB_RX] = + bhi->max_l3_packet_bytes[VLIB_TX] = + shi->max_l3_packet_bytes[VLIB_RX]; + + /* Set max packet size allowed as the lowest of slave */ + if (bhi->max_packet_bytes > shi->max_packet_bytes) + bhi->max_packet_bytes = shi->max_packet_bytes; + } + } + } + } + } + } + + while (1) + { + /* + * check each time through the loop in case intervals are changed + */ + f64 min_wait = dm->link_state_poll_interval < dm->stat_poll_interval ? + dm->link_state_poll_interval : dm->stat_poll_interval; + + vlib_process_wait_for_event_or_clock (vm, min_wait); + + if (dm->admin_up_down_in_progress) + /* skip the poll if an admin up down is in progress (on any interface) */ + continue; + + vec_foreach (xd, dm->devices) + { + f64 now = vlib_time_now (vm); + if ((now - xd->time_last_stats_update) >= dm->stat_poll_interval) + dpdk_update_counters (xd, now); + if ((now - xd->time_last_link_update) >= dm->link_state_poll_interval) + dpdk_update_link_state (xd, now); + + } + } + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (dpdk_process_node,static) = { + .function = dpdk_process, + .type = VLIB_NODE_TYPE_PROCESS, + .name = "dpdk-process", + .process_log2_n_stack_bytes = 17, +}; +/* *INDENT-ON* */ + +int +dpdk_set_stat_poll_interval (f64 interval) +{ + if (interval < DPDK_MIN_STATS_POLL_INTERVAL) + return (VNET_API_ERROR_INVALID_VALUE); + + dpdk_main.stat_poll_interval = interval; + + return 0; +} + +int +dpdk_set_link_state_poll_interval (f64 interval) +{ + if (interval < DPDK_MIN_LINK_POLL_INTERVAL) + return (VNET_API_ERROR_INVALID_VALUE); + + dpdk_main.link_state_poll_interval = interval; + + return 0; +} + +clib_error_t * +dpdk_init (vlib_main_t * vm) +{ + dpdk_main_t *dm = &dpdk_main; + vlib_node_t *ei; + clib_error_t *error = 0; + vlib_thread_main_t *tm = vlib_get_thread_main (); + + /* verify that structs are cacheline aligned */ + STATIC_ASSERT (offsetof (dpdk_device_t, cacheline0) == 0, + "Cache line marker must be 1st element in dpdk_device_t"); + STATIC_ASSERT (offsetof (dpdk_device_t, cacheline1) == + CLIB_CACHE_LINE_BYTES, + "Data in cache line 0 is bigger than cache line size"); + STATIC_ASSERT (offsetof (dpdk_worker_t, cacheline0) == 0, + "Cache line marker must be 1st element in dpdk_worker_t"); + STATIC_ASSERT (offsetof (frame_queue_trace_t, cacheline0) == 0, + "Cache line marker must be 1st element in frame_queue_trace_t"); + + dm->vlib_main = vm; + dm->vnet_main = vnet_get_main (); + dm->conf = &dpdk_config_main; + + ei = vlib_get_node_by_name (vm, (u8 *) "ethernet-input"); + if (ei == 0) + return clib_error_return (0, "ethernet-input node AWOL"); + + dm->ethernet_input_node_index = ei->index; + + dm->conf->nchannels = 4; + dm->conf->num_mbufs = dm->conf->num_mbufs ? dm->conf->num_mbufs : NB_MBUF; + vec_add1 (dm->conf->eal_init_args, (u8 *) "vnet"); + + dm->dpdk_device_by_kni_port_id = hash_create (0, sizeof (uword)); + dm->vu_sw_if_index_by_listener_fd = hash_create (0, sizeof (uword)); + dm->vu_sw_if_index_by_sock_fd = hash_create (0, sizeof (uword)); + + /* $$$ use n_thread_stacks since it's known-good at this point */ + vec_validate (dm->recycle, tm->n_thread_stacks - 1); + + /* Default vlib_buffer_t flags, DISABLES tcp/udp checksumming... */ + dm->buffer_flags_template = + (VLIB_BUFFER_TOTAL_LENGTH_VALID | VNET_BUFFER_RTE_MBUF_VALID + | IP_BUFFER_L4_CHECKSUM_COMPUTED | IP_BUFFER_L4_CHECKSUM_CORRECT); + + dm->stat_poll_interval = DPDK_STATS_POLL_INTERVAL; + dm->link_state_poll_interval = DPDK_LINK_POLL_INTERVAL; + + /* init CLI */ + if ((error = vlib_call_init_function (vm, dpdk_cli_init))) + return error; + + return error; +} + +VLIB_INIT_FUNCTION (dpdk_init); + + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/cli.c b/src/vnet/devices/dpdk/ipsec/cli.c new file mode 100644 index 00000000000..3b634e036da --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/cli.c @@ -0,0 +1,141 @@ +/* + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include <vnet/vnet.h> +#include <vnet/devices/dpdk/ipsec/ipsec.h> + +static void +dpdk_ipsec_show_mapping (vlib_main_t * vm, u16 detail_display) +{ + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); + u32 i, skip_master; + + if (detail_display) + vlib_cli_output (vm, "worker\t%10s\t%15s\tdir\tdev\tqp\n", + "cipher", "auth"); + else + vlib_cli_output (vm, "worker\tcrypto device id(type)\n"); + + skip_master = vlib_num_workers () > 0; + + for (i = 0; i < tm->n_vlib_mains; i++) + { + uword key, data; + u32 cpu_index = vlib_mains[i]->cpu_index; + crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index]; + u8 *s = 0; + + if (skip_master) + { + skip_master = 0; + continue; + } + + if (!detail_display) + { + i32 last_cdev = -1; + crypto_qp_data_t *qpd; + + s = format (s, "%u\t", cpu_index); + + /* *INDENT-OFF* */ + vec_foreach (qpd, cwm->qp_data) + { + u32 dev_id = qpd->dev_id; + + if ((u16) last_cdev != dev_id) + { + struct rte_cryptodev_info cdev_info; + + rte_cryptodev_info_get (dev_id, &cdev_info); + + s = format(s, "%u(%s)\t", dev_id, cdev_info.feature_flags & + RTE_CRYPTODEV_FF_HW_ACCELERATED ? "HW" : "SW"); + } + last_cdev = dev_id; + } + /* *INDENT-ON* */ + vlib_cli_output (vm, "%s", s); + } + else + { + char cipher_str[15], auth_str[15]; + struct rte_cryptodev_capabilities cap; + crypto_worker_qp_key_t *p_key = (crypto_worker_qp_key_t *) & key; + /* *INDENT-OFF* */ + hash_foreach (key, data, cwm->algo_qp_map, + ({ + cap.op = RTE_CRYPTO_OP_TYPE_SYMMETRIC; + cap.sym.xform_type = RTE_CRYPTO_SYM_XFORM_CIPHER; + cap.sym.cipher.algo = p_key->cipher_algo; + check_algo_is_supported (&cap, cipher_str); + cap.op = RTE_CRYPTO_OP_TYPE_SYMMETRIC; + cap.sym.xform_type = RTE_CRYPTO_SYM_XFORM_AUTH; + cap.sym.auth.algo = p_key->auth_algo; + check_algo_is_supported (&cap, auth_str); + vlib_cli_output (vm, "%u\t%10s\t%15s\t%3s\t%u\t%u\n", + vlib_mains[i]->cpu_index, cipher_str, auth_str, + p_key->is_outbound ? "out" : "in", + cwm->qp_data[data].dev_id, + cwm->qp_data[data].qp_id); + })); + /* *INDENT-ON* */ + } + } +} + +static clib_error_t * +lcore_cryptodev_map_fn (vlib_main_t * vm, unformat_input_t * input, + vlib_cli_command_t * cmd) +{ + unformat_input_t _line_input, *line_input = &_line_input; + u16 detail = 0; + + if (!unformat_user (input, unformat_line_input, line_input)) + return 0; + + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "verbose")) + detail = 1; + else + return clib_error_return (0, "parse error: '%U'", + format_unformat_error, line_input); + } + + unformat_free (line_input); + + dpdk_ipsec_show_mapping (vm, detail); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (lcore_cryptodev_map, static) = { + .path = "show crypto device mapping", + .short_help = + "show cryptodev device mapping <verbose>", + .function = lcore_cryptodev_map_fn, +}; +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/crypto_node.c b/src/vnet/devices/dpdk/ipsec/crypto_node.c new file mode 100644 index 00000000000..7b32704ec05 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/crypto_node.c @@ -0,0 +1,210 @@ +/* + *------------------------------------------------------------------ + * crypto_node.c - DPDK Cryptodev input node + * + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include <vlib/vlib.h> +#include <vnet/ip/ip.h> +#include <vnet/ethernet/ethernet.h> +#include <vnet/ipsec/ipsec.h> + +#include <vnet/devices/dpdk/ipsec/ipsec.h> + +#define foreach_dpdk_crypto_input_next \ + _(DROP, "error-drop") \ + _(ENCRYPT_POST, "dpdk-esp-encrypt-post") \ + _(DECRYPT_POST, "dpdk-esp-decrypt-post") + +typedef enum +{ +#define _(f,s) DPDK_CRYPTO_INPUT_NEXT_##f, + foreach_dpdk_crypto_input_next +#undef _ + DPDK_CRYPTO_INPUT_N_NEXT, +} dpdk_crypto_input_next_t; + +#define foreach_dpdk_crypto_input_error \ + _(DQ_COPS, "Crypto ops dequeued") \ + _(COP_FAILED, "Crypto op failed") + +typedef enum +{ +#define _(f,s) DPDK_CRYPTO_INPUT_ERROR_##f, + foreach_dpdk_crypto_input_error +#undef _ + DPDK_CRYPTO_INPUT_N_ERROR, +} dpdk_crypto_input_error_t; + +static char *dpdk_crypto_input_error_strings[] = { +#define _(n, s) s, + foreach_dpdk_crypto_input_error +#undef _ +}; + +vlib_node_registration_t dpdk_crypto_input_node; + +typedef struct +{ + u32 cdev; + u32 qp; + u32 status; + u32 sa_idx; + u32 next_index; +} dpdk_crypto_input_trace_t; + +static u8 * +format_dpdk_crypto_input_trace (u8 * s, va_list * args) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *); + dpdk_crypto_input_trace_t *t = va_arg (*args, dpdk_crypto_input_trace_t *); + + s = format (s, "dpdk_crypto: cryptodev-id %u queue-pair %u next-index %d", + t->cdev, t->qp, t->next_index); + + s = format (s, " status %u sa-idx %u\n", t->status, t->sa_idx); + + return s; +} + +static_always_inline u32 +dpdk_crypto_dequeue (vlib_main_t * vm, vlib_node_runtime_t * node, + crypto_qp_data_t * qpd) +{ + u32 n_deq, *to_next = 0, next_index, n_cops, def_next_index; + struct rte_crypto_op **cops = qpd->cops; + + if (qpd->inflights == 0) + return 0; + + if (qpd->is_outbound) + def_next_index = DPDK_CRYPTO_INPUT_NEXT_ENCRYPT_POST; + else + def_next_index = DPDK_CRYPTO_INPUT_NEXT_DECRYPT_POST; + + n_cops = rte_cryptodev_dequeue_burst (qpd->dev_id, qpd->qp_id, + cops, VLIB_FRAME_SIZE); + n_deq = n_cops; + next_index = def_next_index; + + qpd->inflights -= n_cops; + ASSERT (qpd->inflights >= 0); + + while (n_cops > 0) + { + u32 n_left_to_next; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_cops > 0 && n_left_to_next > 0) + { + u32 bi0, next0; + vlib_buffer_t *b0 = 0; + struct rte_crypto_op *cop; + struct rte_crypto_sym_op *sym_cop; + + cop = cops[0]; + cops += 1; + n_cops -= 1; + n_left_to_next -= 1; + + next0 = def_next_index; + + if (PREDICT_FALSE (cop->status != RTE_CRYPTO_OP_STATUS_SUCCESS)) + { + next0 = DPDK_CRYPTO_INPUT_NEXT_DROP; + vlib_node_increment_counter (vm, dpdk_crypto_input_node.index, + DPDK_CRYPTO_INPUT_ERROR_COP_FAILED, + 1); + } + cop->status = RTE_CRYPTO_OP_STATUS_NOT_PROCESSED; + + sym_cop = (struct rte_crypto_sym_op *) (cop + 1); + b0 = vlib_buffer_from_rte_mbuf (sym_cop->m_src); + bi0 = vlib_get_buffer_index (vm, b0); + + to_next[0] = bi0; + to_next += 1; + + if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED)) + { + vlib_trace_next_frame (vm, node, next0); + dpdk_crypto_input_trace_t *tr = + vlib_add_trace (vm, node, b0, sizeof (*tr)); + tr->cdev = qpd->dev_id; + tr->qp = qpd->qp_id; + tr->status = cop->status; + tr->next_index = next0; + tr->sa_idx = vnet_buffer (b0)->ipsec.sad_index; + } + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next, + n_left_to_next, bi0, next0); + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + + crypto_free_cop (qpd, qpd->cops, n_deq); + + vlib_node_increment_counter (vm, dpdk_crypto_input_node.index, + DPDK_CRYPTO_INPUT_ERROR_DQ_COPS, n_deq); + return n_deq; +} + +static uword +dpdk_crypto_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node, + vlib_frame_t * frame) +{ + u32 cpu_index = os_get_cpu_number (); + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index]; + crypto_qp_data_t *qpd; + u32 n_deq = 0; + + /* *INDENT-OFF* */ + vec_foreach (qpd, cwm->qp_data) + n_deq += dpdk_crypto_dequeue(vm, node, qpd); + /* *INDENT-ON* */ + + return n_deq; +} + +VLIB_REGISTER_NODE (dpdk_crypto_input_node) = +{ + .function = dpdk_crypto_input_fn,.name = "dpdk-crypto-input",.format_trace = + format_dpdk_crypto_input_trace,.type = VLIB_NODE_TYPE_INPUT,.state = + VLIB_NODE_STATE_DISABLED,.n_errors = + DPDK_CRYPTO_INPUT_N_ERROR,.error_strings = + dpdk_crypto_input_error_strings,.n_next_nodes = + DPDK_CRYPTO_INPUT_N_NEXT,.next_nodes = + { +#define _(s,n) [DPDK_CRYPTO_INPUT_NEXT_##s] = n, + foreach_dpdk_crypto_input_next +#undef _ + } +,}; + +#if DPDK_CRYPTO==1 +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_crypto_input_node, dpdk_crypto_input_fn) +#endif +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/dir.dox b/src/vnet/devices/dpdk/ipsec/dir.dox new file mode 100644 index 00000000000..ffebfc4d62e --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/dir.dox @@ -0,0 +1,18 @@ +/* + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +/** + @dir vnet/vnet/devices/dpdk/ipsec + @brief IPSec ESP encrypt/decrypt using DPDK Cryptodev API +*/ diff --git a/src/vnet/devices/dpdk/ipsec/dpdk_crypto_ipsec_doc.md b/src/vnet/devices/dpdk/ipsec/dpdk_crypto_ipsec_doc.md new file mode 100644 index 00000000000..8089696f4a0 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/dpdk_crypto_ipsec_doc.md @@ -0,0 +1,73 @@ +# VPP IPSec implementation using DPDK Cryptodev API {#dpdk_crypto_ipsec_doc} + +This document is meant to contain all related information about implementation and usability. + + +## VPP IPsec with DPDK Cryptodev + +DPDK Cryptodev is an asynchronous crypto API that supports both Hardware and Software implementations (for more details refer to [DPDK Cryptography Device Library documentation](http://dpdk.org/doc/guides/prog_guide/cryptodev_lib.html)). + +When DPDK Cryptodev support is enabled, the node graph is modified by adding and replacing some of the nodes. + +The following nodes are replaced: +* esp-encrypt -> dpdk-esp-encrypt +* esp-decrypt -> dpdk-esp-decrypt + +The following nodes are added: +* dpdk-crypto-input : polling input node, basically dequeuing from crypto devices. +* dpdk-esp-encrypt-post : internal node. +* dpdk-esp-decrypt-post : internal node. + + +### How to enable VPP IPSec with DPDK Cryptodev support + +To enable DPDK Cryptodev support (disabled by default), we need the following env option: + + vpp_uses_dpdk_cryptodev=yes + +A couple of ways to achive this: +* uncomment/add it in the platforms config (ie. build-data/platforms/vpp.mk) +* set the option when building vpp (ie. make vpp_uses_dpdk_cryptodev=yes build-release) + + +### Crypto Resources allocation + +VPP allocates crypto resources based on a best effort approach: +* first allocate Hardware crypto resources, then Software. +* if there are not enough crypto resources for all workers, all packets will be dropped if they reach ESP encrypt/decrypt nodes, displaying the warning: + + 0: dpdk_ipsec_init: not enough cryptodevs for ipsec + + +### Configuration example + +No especial IPsec configuration is required. + +Once DPDK Cryptodev is enabled, the user just needs to provide cryptodevs in the startup.conf. + +Example startup.conf: + +``` +dpdk { + socket-mem 1024,1024 + num-mbufs 131072 + dev 0000:81:00.0 + dev 0000:81:00.1 + dev 0000:85:01.0 + dev 0000:85:01.1 + vdev cryptodev_aesni_mb_pmd,socket_id=1 + vdev cryptodev_aesni_mb_pmd,socket_id=1 +} +``` + +In the above configuration: +* 0000:85:01.0 and 0000:85:01.1 are crypto BDFs and they require the same driver binding as DPDK Ethernet devices but they do not support any extra configuration options. +* Two AESNI-MB Software Cryptodev PMDs are created in NUMA node 1. + +For further details refer to [DPDK Crypto Device Driver documentation](http://dpdk.org/doc/guides/cryptodevs/index.html) + +### Operational data + +The following CLI command displays the Cryptodev/Worker mapping: + + show crypto device mapping [verbose] diff --git a/src/vnet/devices/dpdk/ipsec/esp.h b/src/vnet/devices/dpdk/ipsec/esp.h new file mode 100644 index 00000000000..7ef90c49816 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/esp.h @@ -0,0 +1,295 @@ +/* + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#ifndef __DPDK_ESP_H__ +#define __DPDK_ESP_H__ + +#include <vnet/devices/dpdk/ipsec/ipsec.h> +#include <vnet/ipsec/ipsec.h> +#include <vnet/ipsec/esp.h> + +typedef struct +{ + enum rte_crypto_cipher_algorithm algo; + u8 key_len; + u8 iv_len; +} dpdk_esp_crypto_alg_t; + +typedef struct +{ + enum rte_crypto_auth_algorithm algo; + u8 trunc_size; +} dpdk_esp_integ_alg_t; + +typedef struct +{ + dpdk_esp_crypto_alg_t *esp_crypto_algs; + dpdk_esp_integ_alg_t *esp_integ_algs; +} dpdk_esp_main_t; + +dpdk_esp_main_t dpdk_esp_main; + +static_always_inline void +dpdk_esp_init () +{ + dpdk_esp_main_t *em = &dpdk_esp_main; + dpdk_esp_integ_alg_t *i; + dpdk_esp_crypto_alg_t *c; + + vec_validate (em->esp_crypto_algs, IPSEC_CRYPTO_N_ALG - 1); + + c = &em->esp_crypto_algs[IPSEC_CRYPTO_ALG_AES_CBC_128]; + c->algo = RTE_CRYPTO_CIPHER_AES_CBC; + c->key_len = 16; + c->iv_len = 16; + + c = &em->esp_crypto_algs[IPSEC_CRYPTO_ALG_AES_CBC_192]; + c->algo = RTE_CRYPTO_CIPHER_AES_CBC; + c->key_len = 24; + c->iv_len = 16; + + c = &em->esp_crypto_algs[IPSEC_CRYPTO_ALG_AES_CBC_256]; + c->algo = RTE_CRYPTO_CIPHER_AES_CBC; + c->key_len = 32; + c->iv_len = 16; + + c = &em->esp_crypto_algs[IPSEC_CRYPTO_ALG_AES_GCM_128]; + c->algo = RTE_CRYPTO_CIPHER_AES_GCM; + c->key_len = 16; + c->iv_len = 8; + + vec_validate (em->esp_integ_algs, IPSEC_INTEG_N_ALG - 1); + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_SHA1_96]; + i->algo = RTE_CRYPTO_AUTH_SHA1_HMAC; + i->trunc_size = 12; + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_SHA_256_96]; + i->algo = RTE_CRYPTO_AUTH_SHA256_HMAC; + i->trunc_size = 12; + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_SHA_256_128]; + i->algo = RTE_CRYPTO_AUTH_SHA256_HMAC; + i->trunc_size = 16; + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_SHA_384_192]; + i->algo = RTE_CRYPTO_AUTH_SHA384_HMAC; + i->trunc_size = 24; + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_SHA_512_256]; + i->algo = RTE_CRYPTO_AUTH_SHA512_HMAC; + i->trunc_size = 32; + + i = &em->esp_integ_algs[IPSEC_INTEG_ALG_AES_GCM_128]; + i->algo = RTE_CRYPTO_AUTH_AES_GCM; + i->trunc_size = 16; +} + +static_always_inline int +add_del_sa_sess (u32 sa_index, u8 is_add) +{ + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + crypto_worker_main_t *cwm; + u8 skip_master = vlib_num_workers () > 0; + + /* *INDENT-OFF* */ + vec_foreach (cwm, dcm->workers_main) + { + crypto_sa_session_t *sa_sess; + u8 is_outbound; + + if (skip_master) + { + skip_master = 0; + continue; + } + + for (is_outbound = 0; is_outbound < 2; is_outbound++) + { + if (is_add) + { + pool_get (cwm->sa_sess_d[is_outbound], sa_sess); + } + else + { + u8 dev_id; + + sa_sess = pool_elt_at_index (cwm->sa_sess_d[is_outbound], sa_index); + dev_id = cwm->qp_data[sa_sess->qp_index].dev_id; + + if (!sa_sess->sess) + continue; + + if (rte_cryptodev_sym_session_free(dev_id, sa_sess->sess)) + { + clib_warning("failed to free session"); + return -1; + } + memset(sa_sess, 0, sizeof(sa_sess[0])); + } + } + } + /* *INDENT-OFF* */ + + return 0; +} + +static_always_inline int +translate_crypto_algo(ipsec_crypto_alg_t crypto_algo, + struct rte_crypto_sym_xform *cipher_xform) +{ + switch (crypto_algo) + { + case IPSEC_CRYPTO_ALG_NONE: + cipher_xform->cipher.algo = RTE_CRYPTO_CIPHER_NULL; + break; + case IPSEC_CRYPTO_ALG_AES_CBC_128: + case IPSEC_CRYPTO_ALG_AES_CBC_192: + case IPSEC_CRYPTO_ALG_AES_CBC_256: + cipher_xform->cipher.algo = RTE_CRYPTO_CIPHER_AES_CBC; + break; + case IPSEC_CRYPTO_ALG_AES_GCM_128: + cipher_xform->cipher.algo = RTE_CRYPTO_CIPHER_AES_GCM; + break; + default: + return -1; + } + + cipher_xform->type = RTE_CRYPTO_SYM_XFORM_CIPHER; + + return 0; +} + +static_always_inline int +translate_integ_algo(ipsec_integ_alg_t integ_alg, + struct rte_crypto_sym_xform *auth_xform, int use_esn) +{ + switch (integ_alg) { + case IPSEC_INTEG_ALG_NONE: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_NULL; + auth_xform->auth.digest_length = 0; + break; + case IPSEC_INTEG_ALG_SHA1_96: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_SHA1_HMAC; + auth_xform->auth.digest_length = 12; + break; + case IPSEC_INTEG_ALG_SHA_256_96: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_SHA256_HMAC; + auth_xform->auth.digest_length = 12; + break; + case IPSEC_INTEG_ALG_SHA_256_128: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_SHA256_HMAC; + auth_xform->auth.digest_length = 16; + break; + case IPSEC_INTEG_ALG_SHA_384_192: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_SHA384_HMAC; + auth_xform->auth.digest_length = 24; + break; + case IPSEC_INTEG_ALG_SHA_512_256: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_SHA512_HMAC; + auth_xform->auth.digest_length = 32; + break; + case IPSEC_INTEG_ALG_AES_GCM_128: + auth_xform->auth.algo = RTE_CRYPTO_AUTH_AES_GCM; + auth_xform->auth.digest_length = 16; + auth_xform->auth.add_auth_data_length = use_esn? 12 : 8; + break; + default: + return -1; + } + + auth_xform->type = RTE_CRYPTO_SYM_XFORM_AUTH; + + return 0; +} + +static_always_inline int +create_sym_sess(ipsec_sa_t *sa, crypto_sa_session_t *sa_sess, u8 is_outbound) +{ + u32 cpu_index = os_get_cpu_number(); + dpdk_crypto_main_t * dcm = &dpdk_crypto_main; + crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index]; + struct rte_crypto_sym_xform cipher_xform = {0}; + struct rte_crypto_sym_xform auth_xform = {0}; + struct rte_crypto_sym_xform *xfs; + uword key = 0, *data; + crypto_worker_qp_key_t *p_key = (crypto_worker_qp_key_t *)&key; + + if (sa->crypto_alg == IPSEC_CRYPTO_ALG_AES_GCM_128) + { + sa->crypto_key_len -= 4; + clib_memcpy(&sa->salt, &sa->crypto_key[sa->crypto_key_len], 4); + } + else + { + sa->salt = (u32) rand(); + } + + cipher_xform.type = RTE_CRYPTO_SYM_XFORM_CIPHER; + cipher_xform.cipher.key.data = sa->crypto_key; + cipher_xform.cipher.key.length = sa->crypto_key_len; + + auth_xform.type = RTE_CRYPTO_SYM_XFORM_AUTH; + auth_xform.auth.key.data = sa->integ_key; + auth_xform.auth.key.length = sa->integ_key_len; + + if (translate_crypto_algo(sa->crypto_alg, &cipher_xform) < 0) + return -1; + p_key->cipher_algo = cipher_xform.cipher.algo; + + if (translate_integ_algo(sa->integ_alg, &auth_xform, sa->use_esn) < 0) + return -1; + p_key->auth_algo = auth_xform.auth.algo; + + if (is_outbound) + { + cipher_xform.cipher.op = RTE_CRYPTO_CIPHER_OP_ENCRYPT; + auth_xform.auth.op = RTE_CRYPTO_AUTH_OP_GENERATE; + cipher_xform.next = &auth_xform; + xfs = &cipher_xform; + } + else + { + cipher_xform.cipher.op = RTE_CRYPTO_CIPHER_OP_DECRYPT; + auth_xform.auth.op = RTE_CRYPTO_AUTH_OP_VERIFY; + auth_xform.next = &cipher_xform; + xfs = &auth_xform; + } + + p_key->is_outbound = is_outbound; + + data = hash_get(cwm->algo_qp_map, key); + if (!data) + return -1; + + sa_sess->sess = + rte_cryptodev_sym_session_create(cwm->qp_data[*data].dev_id, xfs); + + if (!sa_sess->sess) + return -1; + + sa_sess->qp_index = (u8)*data; + + return 0; +} + +#endif /* __DPDK_ESP_H__ */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/esp_decrypt.c b/src/vnet/devices/dpdk/ipsec/esp_decrypt.c new file mode 100644 index 00000000000..89ab9f9bc43 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/esp_decrypt.c @@ -0,0 +1,583 @@ +/* + * esp_decrypt.c : IPSec ESP Decrypt node using DPDK Cryptodev + * + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include <vnet/vnet.h> +#include <vnet/api_errno.h> +#include <vnet/ip/ip.h> + +#include <vnet/ipsec/ipsec.h> +#include <vnet/devices/dpdk/ipsec/ipsec.h> +#include <vnet/devices/dpdk/ipsec/esp.h> + +#define foreach_esp_decrypt_next \ +_(DROP, "error-drop") \ +_(IP4_INPUT, "ip4-input") \ +_(IP6_INPUT, "ip6-input") + +#define _(v, s) ESP_DECRYPT_NEXT_##v, +typedef enum { + foreach_esp_decrypt_next +#undef _ + ESP_DECRYPT_N_NEXT, +} esp_decrypt_next_t; + +#define foreach_esp_decrypt_error \ + _(RX_PKTS, "ESP pkts received") \ + _(DECRYPTION_FAILED, "ESP decryption failed") \ + _(REPLAY, "SA replayed packet") \ + _(NOT_IP, "Not IP packet (dropped)") \ + _(ENQ_FAIL, "Enqueue failed (buffer full)") \ + _(NO_CRYPTODEV, "Cryptodev not configured") \ + _(BAD_LEN, "Invalid ciphertext length") \ + _(UNSUPPORTED, "Cipher/Auth not supported") + + +typedef enum { +#define _(sym,str) ESP_DECRYPT_ERROR_##sym, + foreach_esp_decrypt_error +#undef _ + ESP_DECRYPT_N_ERROR, +} esp_decrypt_error_t; + +static char * esp_decrypt_error_strings[] = { +#define _(sym,string) string, + foreach_esp_decrypt_error +#undef _ +}; + +vlib_node_registration_t dpdk_esp_decrypt_node; + +typedef struct { + ipsec_crypto_alg_t crypto_alg; + ipsec_integ_alg_t integ_alg; +} esp_decrypt_trace_t; + +/* packet trace format function */ +static u8 * format_esp_decrypt_trace (u8 * s, va_list * args) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *); + esp_decrypt_trace_t * t = va_arg (*args, esp_decrypt_trace_t *); + + s = format (s, "esp: crypto %U integrity %U", + format_ipsec_crypto_alg, t->crypto_alg, + format_ipsec_integ_alg, t->integ_alg); + return s; +} + +static uword +dpdk_esp_decrypt_node_fn (vlib_main_t * vm, + vlib_node_runtime_t * node, + vlib_frame_t * from_frame) +{ + u32 n_left_from, *from, *to_next, next_index; + ipsec_main_t *im = &ipsec_main; + u32 cpu_index = os_get_cpu_number(); + dpdk_crypto_main_t * dcm = &dpdk_crypto_main; + dpdk_esp_main_t * em = &dpdk_esp_main; + u32 i; + + from = vlib_frame_vector_args (from_frame); + n_left_from = from_frame->n_vectors; + + if (PREDICT_FALSE(!dcm->workers_main)) + { + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_NO_CRYPTODEV, n_left_from); + vlib_buffer_free(vm, from, n_left_from); + return n_left_from; + } + + crypto_worker_main_t *cwm = vec_elt_at_index(dcm->workers_main, cpu_index); + u32 n_qps = vec_len(cwm->qp_data); + struct rte_crypto_op ** cops_to_enq[n_qps]; + u32 n_cop_qp[n_qps], * bi_to_enq[n_qps]; + + for (i = 0; i < n_qps; i++) + { + bi_to_enq[i] = cwm->qp_data[i].bi; + cops_to_enq[i] = cwm->qp_data[i].cops; + } + + memset(n_cop_qp, 0, n_qps * sizeof(u32)); + + crypto_alloc_cops(); + + next_index = ESP_DECRYPT_NEXT_DROP; + + while (n_left_from > 0) + { + u32 n_left_to_next; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_left_from > 0 && n_left_to_next > 0) + { + u32 bi0, sa_index0 = ~0, seq, icv_size, iv_size; + vlib_buffer_t * b0; + esp_header_t * esp0; + ipsec_sa_t * sa0; + struct rte_mbuf * mb0 = 0; + const int BLOCK_SIZE = 16; + crypto_sa_session_t * sa_sess; + void * sess; + u16 qp_index; + struct rte_crypto_op * cop = 0; + + bi0 = from[0]; + from += 1; + n_left_from -= 1; + + b0 = vlib_get_buffer (vm, bi0); + esp0 = vlib_buffer_get_current (b0); + + sa_index0 = vnet_buffer(b0)->ipsec.sad_index; + sa0 = pool_elt_at_index (im->sad, sa_index0); + + seq = clib_host_to_net_u32(esp0->seq); + + /* anti-replay check */ + if (sa0->use_anti_replay) + { + int rv = 0; + + if (PREDICT_TRUE(sa0->use_esn)) + rv = esp_replay_check_esn(sa0, seq); + else + rv = esp_replay_check(sa0, seq); + + if (PREDICT_FALSE(rv)) + { + clib_warning ("anti-replay SPI %u seq %u", sa0->spi, seq); + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_REPLAY, 1); + to_next[0] = bi0; + to_next += 1; + n_left_to_next -= 1; + goto trace; + } + } + + if (PREDICT_FALSE(sa0->integ_alg == IPSEC_INTEG_ALG_NONE) || + PREDICT_FALSE(sa0->crypto_alg == IPSEC_CRYPTO_ALG_NONE)) + { + clib_warning ("SPI %u : only cipher + auth supported", sa0->spi); + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_UNSUPPORTED, 1); + to_next[0] = bi0; + to_next += 1; + n_left_to_next -= 1; + goto trace; + } + + sa_sess = pool_elt_at_index(cwm->sa_sess_d[0], sa_index0); + + if (PREDICT_FALSE(!sa_sess->sess)) + { + int ret = create_sym_sess(sa0, sa_sess, 0); + ASSERT(ret == 0); + } + + sess = sa_sess->sess; + qp_index = sa_sess->qp_index; + + ASSERT (vec_len (vec_elt (cwm->qp_data, qp_index).free_cops) > 0); + cop = vec_pop (vec_elt (cwm->qp_data, qp_index).free_cops); + ASSERT (cop->status == RTE_CRYPTO_OP_STATUS_NOT_PROCESSED); + + cops_to_enq[qp_index][0] = cop; + cops_to_enq[qp_index] += 1; + n_cop_qp[qp_index] += 1; + bi_to_enq[qp_index][0] = bi0; + bi_to_enq[qp_index] += 1; + + rte_crypto_op_attach_sym_session(cop, sess); + + icv_size = em->esp_integ_algs[sa0->integ_alg].trunc_size; + iv_size = em->esp_crypto_algs[sa0->crypto_alg].iv_len; + + /* Convert vlib buffer to mbuf */ + mb0 = rte_mbuf_from_vlib_buffer(b0); + mb0->data_len = b0->current_length; + mb0->pkt_len = b0->current_length; + mb0->data_off = RTE_PKTMBUF_HEADROOM + b0->current_data; + + /* Outer IP header has already been stripped */ + u16 payload_len = rte_pktmbuf_pkt_len(mb0) - sizeof (esp_header_t) - + iv_size - icv_size; + + if ((payload_len & (BLOCK_SIZE - 1)) || (payload_len <= 0)) + { + clib_warning ("payload %u not multiple of %d\n", + payload_len, BLOCK_SIZE); + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_BAD_LEN, 1); + vec_add (vec_elt (cwm->qp_data, qp_index).free_cops, &cop, 1); + bi_to_enq[qp_index] -= 1; + cops_to_enq[qp_index] -= 1; + n_cop_qp[qp_index] -= 1; + to_next[0] = bi0; + to_next += 1; + n_left_to_next -= 1; + goto trace; + } + + struct rte_crypto_sym_op *sym_cop = (struct rte_crypto_sym_op *)(cop + 1); + + sym_cop->m_src = mb0; + sym_cop->cipher.data.offset = sizeof (esp_header_t) + iv_size; + sym_cop->cipher.data.length = payload_len; + + u8 *iv = rte_pktmbuf_mtod_offset(mb0, void*, sizeof (esp_header_t)); + dpdk_cop_priv_t * priv = (dpdk_cop_priv_t *)(sym_cop + 1); + + if (sa0->crypto_alg == IPSEC_CRYPTO_ALG_AES_GCM_128) + { + dpdk_gcm_cnt_blk *icb = &priv->cb; + icb->salt = sa0->salt; + clib_memcpy(icb->iv, iv, 8); + icb->cnt = clib_host_to_net_u32(1); + sym_cop->cipher.iv.data = (u8 *)icb; + sym_cop->cipher.iv.phys_addr = cop->phys_addr + + (uintptr_t)icb - (uintptr_t)cop; + sym_cop->cipher.iv.length = 16; + + u8 *aad = priv->aad; + clib_memcpy(aad, iv - sizeof(esp_header_t), 8); + sym_cop->auth.aad.data = aad; + sym_cop->auth.aad.phys_addr = cop->phys_addr + + (uintptr_t)aad - (uintptr_t)cop; + if (sa0->use_esn) + { + *((u32*)&aad[8]) = sa0->seq_hi; + sym_cop->auth.aad.length = 12; + } + else + { + sym_cop->auth.aad.length = 8; + } + + sym_cop->auth.digest.data = rte_pktmbuf_mtod_offset(mb0, void*, + rte_pktmbuf_pkt_len(mb0) - icv_size); + sym_cop->auth.digest.phys_addr = rte_pktmbuf_mtophys_offset(mb0, + rte_pktmbuf_pkt_len(mb0) - icv_size); + sym_cop->auth.digest.length = icv_size; + + } + else + { + sym_cop->cipher.iv.data = rte_pktmbuf_mtod_offset(mb0, void*, + sizeof (esp_header_t)); + sym_cop->cipher.iv.phys_addr = rte_pktmbuf_mtophys_offset(mb0, + sizeof (esp_header_t)); + sym_cop->cipher.iv.length = iv_size; + + if (sa0->use_esn) + { + dpdk_cop_priv_t* priv = (dpdk_cop_priv_t*) (sym_cop + 1); + u8* payload_end = rte_pktmbuf_mtod_offset( + mb0, u8*, sizeof(esp_header_t) + iv_size + payload_len); + + clib_memcpy (priv->icv, payload_end, icv_size); + *((u32*) payload_end) = sa0->seq_hi; + sym_cop->auth.data.offset = 0; + sym_cop->auth.data.length = sizeof(esp_header_t) + iv_size + + payload_len + sizeof(sa0->seq_hi); + sym_cop->auth.digest.data = priv->icv; + sym_cop->auth.digest.phys_addr = cop->phys_addr + + (uintptr_t) priv->icv - (uintptr_t) cop; + sym_cop->auth.digest.length = icv_size; + } + else + { + sym_cop->auth.data.offset = 0; + sym_cop->auth.data.length = sizeof(esp_header_t) + + iv_size + payload_len; + + sym_cop->auth.digest.data = rte_pktmbuf_mtod_offset(mb0, void*, + rte_pktmbuf_pkt_len(mb0) - icv_size); + sym_cop->auth.digest.phys_addr = rte_pktmbuf_mtophys_offset(mb0, + rte_pktmbuf_pkt_len(mb0) - icv_size); + sym_cop->auth.digest.length = icv_size; + } + } + +trace: + if (PREDICT_FALSE(b0->flags & VLIB_BUFFER_IS_TRACED)) + { + esp_decrypt_trace_t *tr = vlib_add_trace (vm, node, b0, sizeof (*tr)); + tr->crypto_alg = sa0->crypto_alg; + tr->integ_alg = sa0->integ_alg; + } + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_RX_PKTS, + from_frame->n_vectors); + crypto_qp_data_t *qpd; + /* *INDENT-OFF* */ + vec_foreach_index (i, cwm->qp_data) + { + u32 enq; + + qpd = vec_elt_at_index(cwm->qp_data, i); + enq = rte_cryptodev_enqueue_burst(qpd->dev_id, qpd->qp_id, + qpd->cops, n_cop_qp[i]); + qpd->inflights += enq; + + if (PREDICT_FALSE(enq < n_cop_qp[i])) + { + crypto_free_cop (qpd, &qpd->cops[enq], n_cop_qp[i] - enq); + vlib_buffer_free (vm, &qpd->bi[enq], n_cop_qp[i] - enq); + + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_ENQ_FAIL, + n_cop_qp[i] - enq); + } + } + /* *INDENT-ON* */ + + return from_frame->n_vectors; +} + +VLIB_REGISTER_NODE (dpdk_esp_decrypt_node) = { + .function = dpdk_esp_decrypt_node_fn, + .name = "dpdk-esp-decrypt", + .vector_size = sizeof (u32), + .format_trace = format_esp_decrypt_trace, + .type = VLIB_NODE_TYPE_INTERNAL, + + .n_errors = ARRAY_LEN(esp_decrypt_error_strings), + .error_strings = esp_decrypt_error_strings, + + .n_next_nodes = ESP_DECRYPT_N_NEXT, + .next_nodes = { +#define _(s,n) [ESP_DECRYPT_NEXT_##s] = n, + foreach_esp_decrypt_next +#undef _ + }, +}; + +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_esp_decrypt_node, dpdk_esp_decrypt_node_fn) + +/* + * Decrypt Post Node + */ + +#define foreach_esp_decrypt_post_error \ + _(PKTS, "ESP post pkts") + +typedef enum { +#define _(sym,str) ESP_DECRYPT_POST_ERROR_##sym, + foreach_esp_decrypt_post_error +#undef _ + ESP_DECRYPT_POST_N_ERROR, +} esp_decrypt_post_error_t; + +static char * esp_decrypt_post_error_strings[] = { +#define _(sym,string) string, + foreach_esp_decrypt_post_error +#undef _ +}; + +vlib_node_registration_t dpdk_esp_decrypt_post_node; + +static u8 * format_esp_decrypt_post_trace (u8 * s, va_list * args) +{ + return s; +} + +static uword +dpdk_esp_decrypt_post_node_fn (vlib_main_t * vm, + vlib_node_runtime_t * node, + vlib_frame_t * from_frame) +{ + u32 n_left_from, *from, *to_next = 0, next_index; + ipsec_sa_t * sa0; + u32 sa_index0 = ~0; + ipsec_main_t *im = &ipsec_main; + dpdk_esp_main_t *em = &dpdk_esp_main; + + from = vlib_frame_vector_args (from_frame); + n_left_from = from_frame->n_vectors; + + next_index = node->cached_next_index; + + while (n_left_from > 0) + { + u32 n_left_to_next; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_left_from > 0 && n_left_to_next > 0) + { + esp_footer_t * f0; + u32 bi0, next0, icv_size, iv_size; + vlib_buffer_t * b0 = 0; + ip4_header_t *ih4 = 0, *oh4 = 0; + ip6_header_t *ih6 = 0, *oh6 = 0; + u8 tunnel_mode = 1; + u8 transport_ip6 = 0; + + next0 = ESP_DECRYPT_NEXT_DROP; + + bi0 = from[0]; + from += 1; + n_left_from -= 1; + n_left_to_next -= 1; + + b0 = vlib_get_buffer (vm, bi0); + + sa_index0 = vnet_buffer(b0)->ipsec.sad_index; + sa0 = pool_elt_at_index (im->sad, sa_index0); + + to_next[0] = bi0; + to_next += 1; + + icv_size = em->esp_integ_algs[sa0->integ_alg].trunc_size; + iv_size = em->esp_crypto_algs[sa0->crypto_alg].iv_len; + + if (sa0->use_anti_replay) + { + esp_header_t * esp0 = vlib_buffer_get_current (b0); + u32 seq; + seq = clib_host_to_net_u32(esp0->seq); + if (PREDICT_TRUE(sa0->use_esn)) + esp_replay_advance_esn(sa0, seq); + else + esp_replay_advance(sa0, seq); + } + + ih4 = (ip4_header_t *) (b0->data + sizeof(ethernet_header_t)); + vlib_buffer_advance (b0, sizeof (esp_header_t) + iv_size); + + b0->current_length -= (icv_size + 2); + b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID; + f0 = (esp_footer_t *) ((u8 *) vlib_buffer_get_current (b0) + + b0->current_length); + b0->current_length -= f0->pad_length; + + /* transport mode */ + if (PREDICT_FALSE(!sa0->is_tunnel && !sa0->is_tunnel_ip6)) + { + tunnel_mode = 0; + + if (PREDICT_TRUE((ih4->ip_version_and_header_length & 0xF0) != 0x40)) + { + if (PREDICT_TRUE((ih4->ip_version_and_header_length & 0xF0) == 0x60)) + transport_ip6 = 1; + else + { + clib_warning("next header: 0x%x", f0->next_header); + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_NOT_IP, 1); + goto trace; + } + } + } + + if (PREDICT_TRUE (tunnel_mode)) + { + if (PREDICT_TRUE(f0->next_header == IP_PROTOCOL_IP_IN_IP)) + next0 = ESP_DECRYPT_NEXT_IP4_INPUT; + else if (f0->next_header == IP_PROTOCOL_IPV6) + next0 = ESP_DECRYPT_NEXT_IP6_INPUT; + else + { + clib_warning("next header: 0x%x", f0->next_header); + vlib_node_increment_counter (vm, dpdk_esp_decrypt_node.index, + ESP_DECRYPT_ERROR_DECRYPTION_FAILED, + 1); + goto trace; + } + } + /* transport mode */ + else + { + if (PREDICT_FALSE(transport_ip6)) + { + ih6 = (ip6_header_t *) (b0->data + sizeof(ethernet_header_t)); + vlib_buffer_advance (b0, -sizeof(ip6_header_t)); + oh6 = vlib_buffer_get_current (b0); + memmove(oh6, ih6, sizeof(ip6_header_t)); + + next0 = ESP_DECRYPT_NEXT_IP6_INPUT; + oh6->protocol = f0->next_header; + oh6->payload_length = + clib_host_to_net_u16 ( + vlib_buffer_length_in_chain(vm, b0) - + sizeof (ip6_header_t)); + } + else + { + vlib_buffer_advance (b0, -sizeof(ip4_header_t)); + oh4 = vlib_buffer_get_current (b0); + memmove(oh4, ih4, sizeof(ip4_header_t)); + + next0 = ESP_DECRYPT_NEXT_IP4_INPUT; + oh4->ip_version_and_header_length = 0x45; + oh4->fragment_id = 0; + oh4->flags_and_fragment_offset = 0; + oh4->protocol = f0->next_header; + oh4->length = clib_host_to_net_u16 ( + vlib_buffer_length_in_chain (vm, b0)); + oh4->checksum = ip4_header_checksum (oh4); + } + } + + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32)~0; + +trace: + if (PREDICT_FALSE(b0->flags & VLIB_BUFFER_IS_TRACED)) + { + esp_decrypt_trace_t *tr = vlib_add_trace (vm, node, b0, sizeof (*tr)); + tr->crypto_alg = sa0->crypto_alg; + tr->integ_alg = sa0->integ_alg; + } + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, bi0, next0); + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + vlib_node_increment_counter (vm, dpdk_esp_decrypt_post_node.index, + ESP_DECRYPT_POST_ERROR_PKTS, + from_frame->n_vectors); + + return from_frame->n_vectors; +} + +VLIB_REGISTER_NODE (dpdk_esp_decrypt_post_node) = { + .function = dpdk_esp_decrypt_post_node_fn, + .name = "dpdk-esp-decrypt-post", + .vector_size = sizeof (u32), + .format_trace = format_esp_decrypt_post_trace, + .type = VLIB_NODE_TYPE_INTERNAL, + + .n_errors = ARRAY_LEN(esp_decrypt_post_error_strings), + .error_strings = esp_decrypt_post_error_strings, + + .n_next_nodes = ESP_DECRYPT_N_NEXT, + .next_nodes = { +#define _(s,n) [ESP_DECRYPT_NEXT_##s] = n, + foreach_esp_decrypt_next +#undef _ + }, +}; + +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_esp_decrypt_post_node, dpdk_esp_decrypt_post_node_fn) diff --git a/src/vnet/devices/dpdk/ipsec/esp_encrypt.c b/src/vnet/devices/dpdk/ipsec/esp_encrypt.c new file mode 100644 index 00000000000..10bb4616eef --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/esp_encrypt.c @@ -0,0 +1,598 @@ +/* + * esp_encrypt.c : IPSec ESP encrypt node using DPDK Cryptodev + * + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include <vnet/vnet.h> +#include <vnet/api_errno.h> +#include <vnet/ip/ip.h> + +#include <vnet/ipsec/ipsec.h> +#include <vnet/devices/dpdk/ipsec/ipsec.h> +#include <vnet/devices/dpdk/ipsec/esp.h> + +#define foreach_esp_encrypt_next \ +_(DROP, "error-drop") \ +_(IP4_LOOKUP, "ip4-lookup") \ +_(IP6_LOOKUP, "ip6-lookup") \ +_(INTERFACE_OUTPUT, "interface-output") + +#define _(v, s) ESP_ENCRYPT_NEXT_##v, +typedef enum +{ + foreach_esp_encrypt_next +#undef _ + ESP_ENCRYPT_N_NEXT, +} esp_encrypt_next_t; + +#define foreach_esp_encrypt_error \ + _(RX_PKTS, "ESP pkts received") \ + _(SEQ_CYCLED, "sequence number cycled") \ + _(ENQ_FAIL, "Enqueue failed (buffer full)") \ + _(NO_CRYPTODEV, "Cryptodev not configured") \ + _(UNSUPPORTED, "Cipher/Auth not supported") + + +typedef enum +{ +#define _(sym,str) ESP_ENCRYPT_ERROR_##sym, + foreach_esp_encrypt_error +#undef _ + ESP_ENCRYPT_N_ERROR, +} esp_encrypt_error_t; + +static char *esp_encrypt_error_strings[] = { +#define _(sym,string) string, + foreach_esp_encrypt_error +#undef _ +}; + +vlib_node_registration_t dpdk_esp_encrypt_node; + +typedef struct +{ + u32 spi; + u32 seq; + ipsec_crypto_alg_t crypto_alg; + ipsec_integ_alg_t integ_alg; +} esp_encrypt_trace_t; + +/* packet trace format function */ +static u8 * +format_esp_encrypt_trace (u8 * s, va_list * args) +{ + CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *); + CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *); + esp_encrypt_trace_t *t = va_arg (*args, esp_encrypt_trace_t *); + + s = format (s, "esp: spi %u seq %u crypto %U integrity %U", + t->spi, t->seq, + format_ipsec_crypto_alg, t->crypto_alg, + format_ipsec_integ_alg, t->integ_alg); + return s; +} + +static uword +dpdk_esp_encrypt_node_fn (vlib_main_t * vm, + vlib_node_runtime_t * node, + vlib_frame_t * from_frame) +{ + u32 n_left_from, *from, *to_next, next_index; + ipsec_main_t *im = &ipsec_main; + u32 cpu_index = os_get_cpu_number (); + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + dpdk_esp_main_t *em = &dpdk_esp_main; + u32 i; + + from = vlib_frame_vector_args (from_frame); + n_left_from = from_frame->n_vectors; + + if (PREDICT_FALSE (!dcm->workers_main)) + { + /* Likely there are not enough cryptodevs, so drop frame */ + vlib_node_increment_counter (vm, dpdk_esp_encrypt_node.index, + ESP_ENCRYPT_ERROR_NO_CRYPTODEV, + n_left_from); + vlib_buffer_free (vm, from, n_left_from); + return n_left_from; + } + + crypto_worker_main_t *cwm = vec_elt_at_index (dcm->workers_main, cpu_index); + u32 n_qps = vec_len (cwm->qp_data); + struct rte_crypto_op **cops_to_enq[n_qps]; + u32 n_cop_qp[n_qps], *bi_to_enq[n_qps]; + + for (i = 0; i < n_qps; i++) + { + bi_to_enq[i] = cwm->qp_data[i].bi; + cops_to_enq[i] = cwm->qp_data[i].cops; + } + + memset (n_cop_qp, 0, n_qps * sizeof (u32)); + + crypto_alloc_cops (); + + next_index = ESP_ENCRYPT_NEXT_DROP; + + while (n_left_from > 0) + { + u32 n_left_to_next; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_left_from > 0 && n_left_to_next > 0) + { + u32 bi0, next0; + vlib_buffer_t *b0 = 0; + u32 sa_index0; + ipsec_sa_t *sa0; + ip4_and_esp_header_t *ih0, *oh0 = 0; + ip6_and_esp_header_t *ih6_0, *oh6_0 = 0; + struct rte_mbuf *mb0 = 0; + esp_footer_t *f0; + u8 is_ipv6; + u8 ip_hdr_size; + u8 next_hdr_type; + u8 transport_mode = 0; + const int BLOCK_SIZE = 16; + u32 iv_size; + u16 orig_sz; + crypto_sa_session_t *sa_sess; + void *sess; + struct rte_crypto_op *cop = 0; + u16 qp_index; + + bi0 = from[0]; + from += 1; + n_left_from -= 1; + + b0 = vlib_get_buffer (vm, bi0); + sa_index0 = vnet_buffer (b0)->ipsec.sad_index; + sa0 = pool_elt_at_index (im->sad, sa_index0); + + if (PREDICT_FALSE (esp_seq_advance (sa0))) + { + clib_warning ("sequence number counter has cycled SPI %u", + sa0->spi); + vlib_node_increment_counter (vm, dpdk_esp_encrypt_node.index, + ESP_ENCRYPT_ERROR_SEQ_CYCLED, 1); + //TODO: rekey SA + to_next[0] = bi0; + to_next += 1; + n_left_to_next -= 1; + goto trace; + } + + sa_sess = pool_elt_at_index (cwm->sa_sess_d[1], sa_index0); + if (PREDICT_FALSE (!sa_sess->sess)) + { + int ret = create_sym_sess (sa0, sa_sess, 1); + ASSERT (ret == 0); + } + + qp_index = sa_sess->qp_index; + sess = sa_sess->sess; + + ASSERT (vec_len (vec_elt (cwm->qp_data, qp_index).free_cops) > 0); + cop = vec_pop (vec_elt (cwm->qp_data, qp_index).free_cops); + ASSERT (cop->status == RTE_CRYPTO_OP_STATUS_NOT_PROCESSED); + + cops_to_enq[qp_index][0] = cop; + cops_to_enq[qp_index] += 1; + n_cop_qp[qp_index] += 1; + bi_to_enq[qp_index][0] = bi0; + bi_to_enq[qp_index] += 1; + + ssize_t adv; + iv_size = em->esp_crypto_algs[sa0->crypto_alg].iv_len; + ih0 = vlib_buffer_get_current (b0); + orig_sz = b0->current_length; + is_ipv6 = (ih0->ip4.ip_version_and_header_length & 0xF0) == 0x60; + /* is ipv6 */ + if (PREDICT_TRUE (sa0->is_tunnel)) + { + if (PREDICT_TRUE (!is_ipv6)) + adv = -sizeof (ip4_and_esp_header_t); + else + adv = -sizeof (ip6_and_esp_header_t); + } + else + { + adv = -sizeof (esp_header_t); + if (PREDICT_TRUE (!is_ipv6)) + orig_sz -= sizeof (ip4_header_t); + else + orig_sz -= sizeof (ip6_header_t); + } + + /*transport mode save the eth header before it is overwritten */ + if (PREDICT_FALSE (!sa0->is_tunnel)) + { + ethernet_header_t *ieh0 = (ethernet_header_t *) + ((u8 *) vlib_buffer_get_current (b0) - + sizeof (ethernet_header_t)); + ethernet_header_t *oeh0 = + (ethernet_header_t *) ((u8 *) ieh0 + (adv - iv_size)); + clib_memcpy (oeh0, ieh0, sizeof (ethernet_header_t)); + } + + vlib_buffer_advance (b0, adv - iv_size); + + /* XXX IP6/ip4 and IP4/IP6 not supported, only IP4/IP4 and IP6/IP6 */ + + /* is ipv6 */ + if (PREDICT_FALSE (is_ipv6)) + { + ih6_0 = (ip6_and_esp_header_t *) ih0; + ip_hdr_size = sizeof (ip6_header_t); + oh6_0 = vlib_buffer_get_current (b0); + + if (PREDICT_TRUE (sa0->is_tunnel)) + { + next_hdr_type = IP_PROTOCOL_IPV6; + oh6_0->ip6.ip_version_traffic_class_and_flow_label = + ih6_0->ip6.ip_version_traffic_class_and_flow_label; + } + else + { + next_hdr_type = ih6_0->ip6.protocol; + memmove (oh6_0, ih6_0, sizeof (ip6_header_t)); + } + + oh6_0->ip6.protocol = IP_PROTOCOL_IPSEC_ESP; + oh6_0->ip6.hop_limit = 254; + oh6_0->esp.spi = clib_net_to_host_u32 (sa0->spi); + oh6_0->esp.seq = clib_net_to_host_u32 (sa0->seq); + } + else + { + ip_hdr_size = sizeof (ip4_header_t); + oh0 = vlib_buffer_get_current (b0); + + if (PREDICT_TRUE (sa0->is_tunnel)) + { + next_hdr_type = IP_PROTOCOL_IP_IN_IP; + oh0->ip4.tos = ih0->ip4.tos; + } + else + { + next_hdr_type = ih0->ip4.protocol; + memmove (oh0, ih0, sizeof (ip4_header_t)); + } + + oh0->ip4.ip_version_and_header_length = 0x45; + oh0->ip4.fragment_id = 0; + oh0->ip4.flags_and_fragment_offset = 0; + oh0->ip4.ttl = 254; + oh0->ip4.protocol = IP_PROTOCOL_IPSEC_ESP; + oh0->esp.spi = clib_net_to_host_u32 (sa0->spi); + oh0->esp.seq = clib_net_to_host_u32 (sa0->seq); + } + + if (PREDICT_TRUE (sa0->is_tunnel && !sa0->is_tunnel_ip6)) + { + oh0->ip4.src_address.as_u32 = sa0->tunnel_src_addr.ip4.as_u32; + oh0->ip4.dst_address.as_u32 = sa0->tunnel_dst_addr.ip4.as_u32; + + /* in tunnel mode send it back to FIB */ + next0 = ESP_ENCRYPT_NEXT_IP4_LOOKUP; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + } + else if (sa0->is_tunnel && sa0->is_tunnel_ip6) + { + oh6_0->ip6.src_address.as_u64[0] = + sa0->tunnel_src_addr.ip6.as_u64[0]; + oh6_0->ip6.src_address.as_u64[1] = + sa0->tunnel_src_addr.ip6.as_u64[1]; + oh6_0->ip6.dst_address.as_u64[0] = + sa0->tunnel_dst_addr.ip6.as_u64[0]; + oh6_0->ip6.dst_address.as_u64[1] = + sa0->tunnel_dst_addr.ip6.as_u64[1]; + + /* in tunnel mode send it back to FIB */ + next0 = ESP_ENCRYPT_NEXT_IP6_LOOKUP; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + } + else + { + next0 = ESP_ENCRYPT_NEXT_INTERFACE_OUTPUT; + transport_mode = 1; + } + + ASSERT (sa0->crypto_alg < IPSEC_CRYPTO_N_ALG); + ASSERT (sa0->crypto_alg != IPSEC_CRYPTO_ALG_NONE); + + int blocks = 1 + (orig_sz + 1) / BLOCK_SIZE; + + /* pad packet in input buffer */ + u8 pad_bytes = BLOCK_SIZE * blocks - 2 - orig_sz; + u8 i; + u8 *padding = vlib_buffer_get_current (b0) + b0->current_length; + + for (i = 0; i < pad_bytes; ++i) + padding[i] = i + 1; + + f0 = vlib_buffer_get_current (b0) + b0->current_length + pad_bytes; + f0->pad_length = pad_bytes; + f0->next_header = next_hdr_type; + b0->current_length += pad_bytes + 2 + + em->esp_integ_algs[sa0->integ_alg].trunc_size; + + vnet_buffer (b0)->sw_if_index[VLIB_RX] = + vnet_buffer (b0)->sw_if_index[VLIB_RX]; + b0->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID; + + struct rte_crypto_sym_op *sym_cop; + sym_cop = (struct rte_crypto_sym_op *) (cop + 1); + + dpdk_cop_priv_t *priv = (dpdk_cop_priv_t *) (sym_cop + 1); + + vnet_buffer (b0)->unused[0] = next0; + + mb0 = rte_mbuf_from_vlib_buffer (b0); + mb0->data_len = b0->current_length; + mb0->pkt_len = b0->current_length; + mb0->data_off = RTE_PKTMBUF_HEADROOM + b0->current_data; + + rte_crypto_op_attach_sym_session (cop, sess); + + sym_cop->m_src = mb0; + + dpdk_gcm_cnt_blk *icb = &priv->cb; + icb->salt = sa0->salt; + icb->iv[0] = sa0->seq; + icb->iv[1] = sa0->seq_hi; + + if (sa0->crypto_alg == IPSEC_CRYPTO_ALG_AES_GCM_128) + { + icb->cnt = clib_host_to_net_u32 (1); + clib_memcpy (vlib_buffer_get_current (b0) + ip_hdr_size + + sizeof (esp_header_t), icb->iv, 8); + sym_cop->cipher.data.offset = + ip_hdr_size + sizeof (esp_header_t) + iv_size; + sym_cop->cipher.data.length = BLOCK_SIZE * blocks; + sym_cop->cipher.iv.length = 16; + } + else + { + sym_cop->cipher.data.offset = + ip_hdr_size + sizeof (esp_header_t); + sym_cop->cipher.data.length = BLOCK_SIZE * blocks + iv_size; + sym_cop->cipher.iv.length = iv_size; + } + + sym_cop->cipher.iv.data = (u8 *) icb; + sym_cop->cipher.iv.phys_addr = cop->phys_addr + (uintptr_t) icb + - (uintptr_t) cop; + + + ASSERT (sa0->integ_alg < IPSEC_INTEG_N_ALG); + ASSERT (sa0->integ_alg != IPSEC_INTEG_ALG_NONE); + + if (PREDICT_FALSE (sa0->integ_alg == IPSEC_INTEG_ALG_AES_GCM_128)) + { + u8 *aad = priv->aad; + clib_memcpy (aad, vlib_buffer_get_current (b0) + ip_hdr_size, + 8); + sym_cop->auth.aad.data = aad; + sym_cop->auth.aad.phys_addr = cop->phys_addr + + (uintptr_t) aad - (uintptr_t) cop; + + if (PREDICT_FALSE (sa0->use_esn)) + { + *((u32 *) & aad[8]) = sa0->seq_hi; + sym_cop->auth.aad.length = 12; + } + else + { + sym_cop->auth.aad.length = 8; + } + } + else + { + sym_cop->auth.data.offset = ip_hdr_size; + sym_cop->auth.data.length = b0->current_length - ip_hdr_size + - em->esp_integ_algs[sa0->integ_alg].trunc_size; + + if (PREDICT_FALSE (sa0->use_esn)) + { + u8 *payload_end = + vlib_buffer_get_current (b0) + b0->current_length; + *((u32 *) payload_end) = sa0->seq_hi; + sym_cop->auth.data.length += sizeof (sa0->seq_hi); + } + } + sym_cop->auth.digest.data = vlib_buffer_get_current (b0) + + b0->current_length - + em->esp_integ_algs[sa0->integ_alg].trunc_size; + sym_cop->auth.digest.phys_addr = rte_pktmbuf_mtophys_offset (mb0, + b0->current_length + - + em->esp_integ_algs + [sa0->integ_alg].trunc_size); + sym_cop->auth.digest.length = + em->esp_integ_algs[sa0->integ_alg].trunc_size; + + + if (PREDICT_FALSE (is_ipv6)) + { + oh6_0->ip6.payload_length = + clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, b0) - + sizeof (ip6_header_t)); + } + else + { + oh0->ip4.length = + clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, b0)); + oh0->ip4.checksum = ip4_header_checksum (&oh0->ip4); + } + + if (transport_mode) + vlib_buffer_advance (b0, -sizeof (ethernet_header_t)); + + trace: + if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED)) + { + esp_encrypt_trace_t *tr = + vlib_add_trace (vm, node, b0, sizeof (*tr)); + tr->spi = sa0->spi; + tr->seq = sa0->seq - 1; + tr->crypto_alg = sa0->crypto_alg; + tr->integ_alg = sa0->integ_alg; + } + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + vlib_node_increment_counter (vm, dpdk_esp_encrypt_node.index, + ESP_ENCRYPT_ERROR_RX_PKTS, + from_frame->n_vectors); + crypto_qp_data_t *qpd; + /* *INDENT-OFF* */ + vec_foreach_index (i, cwm->qp_data) + { + u32 enq; + + qpd = vec_elt_at_index(cwm->qp_data, i); + enq = rte_cryptodev_enqueue_burst(qpd->dev_id, qpd->qp_id, + qpd->cops, n_cop_qp[i]); + qpd->inflights += enq; + + if (PREDICT_FALSE(enq < n_cop_qp[i])) + { + crypto_free_cop (qpd, &qpd->cops[enq], n_cop_qp[i] - enq); + vlib_buffer_free (vm, &qpd->bi[enq], n_cop_qp[i] - enq); + + vlib_node_increment_counter (vm, dpdk_esp_encrypt_node.index, + ESP_ENCRYPT_ERROR_ENQ_FAIL, + n_cop_qp[i] - enq); + } + } + /* *INDENT-ON* */ + + return from_frame->n_vectors; +} + +VLIB_REGISTER_NODE (dpdk_esp_encrypt_node) = +{ + .function = dpdk_esp_encrypt_node_fn,.name = "dpdk-esp-encrypt",.flags = + VLIB_NODE_FLAG_IS_OUTPUT,.vector_size = sizeof (u32),.format_trace = + format_esp_encrypt_trace,.n_errors = + ARRAY_LEN (esp_encrypt_error_strings),.error_strings = + esp_encrypt_error_strings,.n_next_nodes = 1,.next_nodes = + { + [ESP_ENCRYPT_NEXT_DROP] = "error-drop",} +}; + +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_esp_encrypt_node, dpdk_esp_encrypt_node_fn) +/* + * ESP Encrypt Post Node + */ +#define foreach_esp_encrypt_post_error \ + _(PKTS, "ESP post pkts") + typedef enum + { +#define _(sym,str) ESP_ENCRYPT_POST_ERROR_##sym, + foreach_esp_encrypt_post_error +#undef _ + ESP_ENCRYPT_POST_N_ERROR, + } esp_encrypt_post_error_t; + + static char *esp_encrypt_post_error_strings[] = { +#define _(sym,string) string, + foreach_esp_encrypt_post_error +#undef _ + }; + +vlib_node_registration_t dpdk_esp_encrypt_post_node; + +static u8 * +format_esp_encrypt_post_trace (u8 * s, va_list * args) +{ + return s; +} + +static uword +dpdk_esp_encrypt_post_node_fn (vlib_main_t * vm, + vlib_node_runtime_t * node, + vlib_frame_t * from_frame) +{ + u32 n_left_from, *from, *to_next = 0, next_index; + + from = vlib_frame_vector_args (from_frame); + n_left_from = from_frame->n_vectors; + + next_index = node->cached_next_index; + + while (n_left_from > 0) + { + u32 n_left_to_next; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_left_from > 0 && n_left_to_next > 0) + { + u32 bi0, next0; + vlib_buffer_t *b0 = 0; + + bi0 = from[0]; + from += 1; + n_left_from -= 1; + n_left_to_next -= 1; + + b0 = vlib_get_buffer (vm, bi0); + + to_next[0] = bi0; + to_next += 1; + + next0 = vnet_buffer (b0)->unused[0]; + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, bi0, + next0); + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + + vlib_node_increment_counter (vm, dpdk_esp_encrypt_post_node.index, + ESP_ENCRYPT_POST_ERROR_PKTS, + from_frame->n_vectors); + + return from_frame->n_vectors; +} + +VLIB_REGISTER_NODE (dpdk_esp_encrypt_post_node) = +{ + .function = dpdk_esp_encrypt_post_node_fn,.name = + "dpdk-esp-encrypt-post",.vector_size = sizeof (u32),.format_trace = + format_esp_encrypt_post_trace,.type = VLIB_NODE_TYPE_INTERNAL,.n_errors = + ARRAY_LEN (esp_encrypt_post_error_strings),.error_strings = + esp_encrypt_post_error_strings,.n_next_nodes = + ESP_ENCRYPT_N_NEXT,.next_nodes = + { +#define _(s,n) [ESP_ENCRYPT_NEXT_##s] = n, + foreach_esp_encrypt_next +#undef _ + } +}; + +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_esp_encrypt_post_node, + dpdk_esp_encrypt_post_node_fn) +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/ipsec.c b/src/vnet/devices/dpdk/ipsec/ipsec.c new file mode 100644 index 00000000000..de253f02636 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/ipsec.c @@ -0,0 +1,313 @@ +/* + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include <vnet/vnet.h> +#include <vnet/ip/ip.h> +#include <vnet/api_errno.h> +#include <vnet/devices/dpdk/dpdk.h> +#include <vnet/devices/dpdk/ipsec/ipsec.h> +#include <vnet/devices/dpdk/ipsec/esp.h> +#include <vnet/ipsec/ipsec.h> + +#define DPDK_CRYPTO_NB_OBJS 2048 +#define DPDK_CRYPTO_CACHE_SIZE 512 +#define DPDK_CRYPTO_PRIV_SIZE 128 +#define DPDK_CRYPTO_N_QUEUE_DESC 512 +#define DPDK_CRYPTO_NB_COPS (1024 * 4) + +/* + * return: + * -1: update failed + * 0: already exist + * 1: mapped + */ +static int +update_qp_data (crypto_worker_main_t * cwm, + u8 cdev_id, u16 qp_id, u8 is_outbound, u16 * idx) +{ + crypto_qp_data_t *qpd; + + /* *INDENT-OFF* */ + vec_foreach_index (*idx, cwm->qp_data) + { + qpd = vec_elt_at_index(cwm->qp_data, *idx); + + if (qpd->dev_id == cdev_id && qpd->qp_id == qp_id && + qpd->is_outbound == is_outbound) + return 0; + } + /* *INDENT-ON* */ + + vec_add2 (cwm->qp_data, qpd, 1); + + qpd->dev_id = cdev_id; + qpd->qp_id = qp_id; + qpd->is_outbound = is_outbound; + + return 1; +} + +/* + * return: + * -1: error + * 0: already exist + * 1: mapped + */ +static int +add_mapping (crypto_worker_main_t * cwm, + u8 cdev_id, u16 qp, u8 is_outbound, + const struct rte_cryptodev_capabilities *cipher_cap, + const struct rte_cryptodev_capabilities *auth_cap) +{ + int mapped; + u16 qp_index; + uword key = 0, data, *ret; + crypto_worker_qp_key_t *p_key = (crypto_worker_qp_key_t *) & key; + + p_key->cipher_algo = (u8) cipher_cap->sym.cipher.algo; + p_key->auth_algo = (u8) auth_cap->sym.auth.algo; + p_key->is_outbound = is_outbound; + + ret = hash_get (cwm->algo_qp_map, key); + if (ret) + return 0; + + mapped = update_qp_data (cwm, cdev_id, qp, is_outbound, &qp_index); + if (mapped < 0) + return -1; + + data = (uword) qp_index; + + ret = hash_set (cwm->algo_qp_map, key, data); + if (!ret) + rte_panic ("Failed to insert hash table\n"); + + return mapped; +} + +/* + * return: + * 0: already exist + * 1: mapped + */ +static int +add_cdev_mapping (crypto_worker_main_t * cwm, + struct rte_cryptodev_info *dev_info, u8 cdev_id, + u16 qp, u8 is_outbound) +{ + const struct rte_cryptodev_capabilities *i, *j; + u32 mapped = 0; + + for (i = dev_info->capabilities; i->op != RTE_CRYPTO_OP_TYPE_UNDEFINED; i++) + { + if (i->sym.xform_type != RTE_CRYPTO_SYM_XFORM_CIPHER) + continue; + + if (check_algo_is_supported (i, NULL) != 0) + continue; + + for (j = dev_info->capabilities; j->op != RTE_CRYPTO_OP_TYPE_UNDEFINED; + j++) + { + int status = 0; + + if (j->sym.xform_type != RTE_CRYPTO_SYM_XFORM_AUTH) + continue; + + if (check_algo_is_supported (j, NULL) != 0) + continue; + + status = add_mapping (cwm, cdev_id, qp, is_outbound, i, j); + if (status == 1) + mapped += 1; + if (status < 0) + return status; + } + } + + return mapped; +} + +static int +check_cryptodev_queues () +{ + u32 n_qs = 0; + u8 cdev_id; + u32 n_req_qs = 2; + + if (vlib_num_workers () > 0) + n_req_qs = vlib_num_workers () * 2; + + for (cdev_id = 0; cdev_id < rte_cryptodev_count (); cdev_id++) + { + struct rte_cryptodev_info cdev_info; + + rte_cryptodev_info_get (cdev_id, &cdev_info); + + if (! + (cdev_info.feature_flags & RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING)) + continue; + + n_qs += cdev_info.max_nb_queue_pairs; + } + + if (n_qs >= n_req_qs) + return 0; + else + return -1; +} + +static clib_error_t * +dpdk_ipsec_init (vlib_main_t * vm) +{ + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); + struct rte_cryptodev_config dev_conf; + struct rte_cryptodev_qp_conf qp_conf; + struct rte_cryptodev_info cdev_info; + struct rte_mempool *rmp; + i32 dev_id, ret; + u32 i, skip_master; + + if (check_cryptodev_queues () < 0) + return clib_error_return (0, "not enough cryptodevs for ipsec"); + + vec_alloc (dcm->workers_main, tm->n_vlib_mains); + _vec_len (dcm->workers_main) = tm->n_vlib_mains; + + fprintf (stdout, "DPDK Cryptodevs info:\n"); + fprintf (stdout, "dev_id\tn_qp\tnb_obj\tcache_size\n"); + /* HW cryptodevs have higher dev_id, use HW first */ + for (dev_id = rte_cryptodev_count () - 1; dev_id >= 0; dev_id--) + { + u16 max_nb_qp, qp = 0; + skip_master = vlib_num_workers () > 0; + + rte_cryptodev_info_get (dev_id, &cdev_info); + + if (! + (cdev_info.feature_flags & RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING)) + continue; + + max_nb_qp = cdev_info.max_nb_queue_pairs; + + for (i = 0; i < tm->n_vlib_mains; i++) + { + u8 is_outbound; + crypto_worker_main_t *cwm; + uword *map; + + if (skip_master) + { + skip_master = 0; + continue; + } + + cwm = vec_elt_at_index (dcm->workers_main, i); + map = cwm->algo_qp_map; + + if (!map) + { + map = hash_create (0, sizeof (crypto_worker_qp_key_t)); + if (!map) + return clib_error_return (0, "unable to create hash table " + "for worker %u", + vlib_mains[i]->cpu_index); + cwm->algo_qp_map = map; + } + + for (is_outbound = 0; is_outbound < 2 && qp < max_nb_qp; + is_outbound++) + { + int mapped = add_cdev_mapping (cwm, &cdev_info, + dev_id, qp, is_outbound); + if (mapped > 0) + qp++; + + if (mapped < 0) + return clib_error_return (0, + "too many queues for one worker"); + } + } + + if (qp == 0) + continue; + + dev_conf.socket_id = rte_cryptodev_socket_id (dev_id); + dev_conf.nb_queue_pairs = cdev_info.max_nb_queue_pairs; + dev_conf.session_mp.nb_objs = DPDK_CRYPTO_NB_OBJS; + dev_conf.session_mp.cache_size = DPDK_CRYPTO_CACHE_SIZE; + + ret = rte_cryptodev_configure (dev_id, &dev_conf); + if (ret < 0) + return clib_error_return (0, "cryptodev %u config error", dev_id); + + qp_conf.nb_descriptors = DPDK_CRYPTO_N_QUEUE_DESC; + for (qp = 0; qp < dev_conf.nb_queue_pairs; qp++) + { + ret = rte_cryptodev_queue_pair_setup (dev_id, qp, &qp_conf, + dev_conf.socket_id); + if (ret < 0) + return clib_error_return (0, "cryptodev %u qp %u setup error", + dev_id, qp); + } + fprintf (stdout, "%u\t%u\t%u\t%u\n", dev_id, dev_conf.nb_queue_pairs, + DPDK_CRYPTO_NB_OBJS, DPDK_CRYPTO_CACHE_SIZE); + } + + u32 socket_id = rte_socket_id (); + + vec_validate_aligned (dcm->cop_pools, socket_id, CLIB_CACHE_LINE_BYTES); + + /* pool already exists, nothing to do */ + if (dcm->cop_pools[socket_id]) + return 0; + + u8 *pool_name = format (0, "crypto_op_pool_socket%u%c", socket_id, 0); + + rmp = rte_crypto_op_pool_create ((char *) pool_name, + RTE_CRYPTO_OP_TYPE_SYMMETRIC, + DPDK_CRYPTO_NB_COPS * + (1 + vlib_num_workers ()), + DPDK_CRYPTO_CACHE_SIZE, + DPDK_CRYPTO_PRIV_SIZE, socket_id); + vec_free (pool_name); + + if (!rmp) + return clib_error_return (0, "failed to allocate mempool on socket %u", + socket_id); + dcm->cop_pools[socket_id] = rmp; + + dpdk_esp_init (); + + if (vec_len (vlib_mains) == 0) + vlib_node_set_state (&vlib_global_main, dpdk_crypto_input_node.index, + VLIB_NODE_STATE_POLLING); + else + for (i = 1; i < tm->n_vlib_mains; i++) + vlib_node_set_state (vlib_mains[i], dpdk_crypto_input_node.index, + VLIB_NODE_STATE_POLLING); + + return 0; +} + +VLIB_MAIN_LOOP_ENTER_FUNCTION (dpdk_ipsec_init); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/ipsec/ipsec.h b/src/vnet/devices/dpdk/ipsec/ipsec.h new file mode 100644 index 00000000000..e6c7498c0d3 --- /dev/null +++ b/src/vnet/devices/dpdk/ipsec/ipsec.h @@ -0,0 +1,227 @@ +/* + * Copyright (c) 2016 Intel and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#ifndef __DPDK_IPSEC_H__ +#define __DPDK_IPSEC_H__ + +#include <vnet/vnet.h> + +#undef always_inline +#include <rte_crypto.h> +#include <rte_cryptodev.h> + +#if CLIB_DEBUG > 0 +#define always_inline static inline +#else +#define always_inline static inline __attribute__ ((__always_inline__)) +#endif + + +#define MAX_QP_PER_LCORE 16 + +typedef struct +{ + u32 salt; + u32 iv[2]; + u32 cnt; +} dpdk_gcm_cnt_blk; + +typedef struct +{ + dpdk_gcm_cnt_blk cb; + union + { + u8 aad[12]; + u8 icv[64]; + }; +} dpdk_cop_priv_t; + +typedef struct +{ + u8 cipher_algo; + u8 auth_algo; + u8 is_outbound; +} crypto_worker_qp_key_t; + +typedef struct +{ + u16 dev_id; + u16 qp_id; + u16 is_outbound; + i16 inflights; + u32 bi[VLIB_FRAME_SIZE]; + struct rte_crypto_op *cops[VLIB_FRAME_SIZE]; + struct rte_crypto_op **free_cops; +} crypto_qp_data_t; + +typedef struct +{ + u8 qp_index; + void *sess; +} crypto_sa_session_t; + +typedef struct +{ + crypto_sa_session_t *sa_sess_d[2]; + crypto_qp_data_t *qp_data; + uword *algo_qp_map; +} crypto_worker_main_t; + +typedef struct +{ + struct rte_mempool **cop_pools; + crypto_worker_main_t *workers_main; +} dpdk_crypto_main_t; + +dpdk_crypto_main_t dpdk_crypto_main; + +extern vlib_node_registration_t dpdk_crypto_input_node; + +#define CRYPTO_N_FREE_COPS (VLIB_FRAME_SIZE * 3) + +static_always_inline void +crypto_alloc_cops () +{ + dpdk_crypto_main_t *dcm = &dpdk_crypto_main; + u32 cpu_index = os_get_cpu_number (); + crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index]; + unsigned socket_id = rte_socket_id (); + crypto_qp_data_t *qpd; + + /* *INDENT-OFF* */ + vec_foreach (qpd, cwm->qp_data) + { + u32 l = vec_len (qpd->free_cops); + + if (PREDICT_FALSE (l < VLIB_FRAME_SIZE)) + { + u32 n_alloc; + + if (PREDICT_FALSE (!qpd->free_cops)) + vec_alloc (qpd->free_cops, CRYPTO_N_FREE_COPS); + + n_alloc = rte_crypto_op_bulk_alloc (dcm->cop_pools[socket_id], + RTE_CRYPTO_OP_TYPE_SYMMETRIC, + &qpd->free_cops[l], + CRYPTO_N_FREE_COPS - l - 1); + + _vec_len (qpd->free_cops) = l + n_alloc; + } + } + /* *INDENT-ON* */ +} + +static_always_inline void +crypto_free_cop (crypto_qp_data_t * qpd, struct rte_crypto_op **cops, u32 n) +{ + u32 l = vec_len (qpd->free_cops); + + if (l + n >= CRYPTO_N_FREE_COPS) + { + l -= VLIB_FRAME_SIZE; + rte_mempool_put_bulk (cops[0]->mempool, + (void **) &qpd->free_cops[l], VLIB_FRAME_SIZE); + } + clib_memcpy (&qpd->free_cops[l], cops, sizeof (*cops) * n); + + _vec_len (qpd->free_cops) = l + n; +} + +static_always_inline int +check_algo_is_supported (const struct rte_cryptodev_capabilities *cap, + char *name) +{ + struct + { + uint8_t cipher_algo; + enum rte_crypto_sym_xform_type type; + union + { + enum rte_crypto_auth_algorithm auth; + enum rte_crypto_cipher_algorithm cipher; + }; + char *name; + } supported_algo[] = + { + { + .type = RTE_CRYPTO_SYM_XFORM_CIPHER,.cipher = + RTE_CRYPTO_CIPHER_NULL,.name = "NULL"}, + { + .type = RTE_CRYPTO_SYM_XFORM_CIPHER,.cipher = + RTE_CRYPTO_CIPHER_AES_CBC,.name = "AES_CBC"}, + { + .type = RTE_CRYPTO_SYM_XFORM_CIPHER,.cipher = + RTE_CRYPTO_CIPHER_AES_CTR,.name = "AES_CTR"}, + { + .type = RTE_CRYPTO_SYM_XFORM_CIPHER,.cipher = + RTE_CRYPTO_CIPHER_3DES_CBC,.name = "3DES-CBC"}, + { + .type = RTE_CRYPTO_SYM_XFORM_CIPHER,.auth = + RTE_CRYPTO_CIPHER_AES_GCM,.name = "AES-GCM"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_SHA1_HMAC,.name = "HMAC-SHA1"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_SHA256_HMAC,.name = "HMAC-SHA256"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_SHA384_HMAC,.name = "HMAC-SHA384"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_SHA512_HMAC,.name = "HMAC-SHA512"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_AES_XCBC_MAC,.name = "AES-XCBC-MAC"}, + { + .type = RTE_CRYPTO_SYM_XFORM_AUTH,.auth = + RTE_CRYPTO_AUTH_AES_GCM,.name = "AES-GCM"}, + { + /* tail */ + .type = RTE_CRYPTO_SYM_XFORM_NOT_SPECIFIED},}; + uint32_t i = 0; + + if (cap->op != RTE_CRYPTO_OP_TYPE_SYMMETRIC) + return -1; + + while (supported_algo[i].type != RTE_CRYPTO_SYM_XFORM_NOT_SPECIFIED) + { + if (cap->sym.xform_type == supported_algo[i].type) + { + if ((cap->sym.xform_type == RTE_CRYPTO_SYM_XFORM_CIPHER && + cap->sym.cipher.algo == supported_algo[i].cipher) || + (cap->sym.xform_type == RTE_CRYPTO_SYM_XFORM_AUTH && + cap->sym.auth.algo == supported_algo[i].auth)) + { + if (name) + strcpy (name, supported_algo[i].name); + return 0; + } + } + + i++; + } + + return -1; +} + +#endif /* __DPDK_IPSEC_H__ */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/node.c b/src/vnet/devices/dpdk/node.c new file mode 100644 index 00000000000..e541cdbcbd2 --- /dev/null +++ b/src/vnet/devices/dpdk/node.c @@ -0,0 +1,687 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include <vnet/vnet.h> +#include <vppinfra/vec.h> +#include <vppinfra/error.h> +#include <vppinfra/format.h> +#include <vppinfra/xxhash.h> + +#include <vnet/ethernet/ethernet.h> +#include <vnet/devices/dpdk/dpdk.h> +#include <vnet/classify/vnet_classify.h> +#include <vnet/mpls/packet.h> +#include <vnet/handoff.h> +#include <vnet/devices/devices.h> +#include <vnet/feature/feature.h> + +#include "dpdk_priv.h" + +static char *dpdk_error_strings[] = { +#define _(n,s) s, + foreach_dpdk_error +#undef _ +}; + +always_inline int +vlib_buffer_is_ip4 (vlib_buffer_t * b) +{ + ethernet_header_t *h = (ethernet_header_t *) b->data; + return (h->type == clib_host_to_net_u16 (ETHERNET_TYPE_IP4)); +} + +always_inline int +vlib_buffer_is_ip6 (vlib_buffer_t * b) +{ + ethernet_header_t *h = (ethernet_header_t *) b->data; + return (h->type == clib_host_to_net_u16 (ETHERNET_TYPE_IP6)); +} + +always_inline int +vlib_buffer_is_mpls (vlib_buffer_t * b) +{ + ethernet_header_t *h = (ethernet_header_t *) b->data; + return (h->type == clib_host_to_net_u16 (ETHERNET_TYPE_MPLS_UNICAST)); +} + +#if RTE_VERSION < RTE_VERSION_NUM(16, 11, 0, 0) +/* New ol_flags bits added in DPDK-16.11 */ +#define PKT_RX_IP_CKSUM_GOOD (1ULL << 7) +#endif + +always_inline u32 +dpdk_rx_next_from_etype (struct rte_mbuf * mb, vlib_buffer_t * b0) +{ + if (PREDICT_TRUE (vlib_buffer_is_ip4 (b0))) + if (PREDICT_TRUE ((mb->ol_flags & PKT_RX_IP_CKSUM_GOOD) != 0)) + return VNET_DEVICE_INPUT_NEXT_IP4_NCS_INPUT; + else + return VNET_DEVICE_INPUT_NEXT_IP4_INPUT; + else if (PREDICT_TRUE (vlib_buffer_is_ip6 (b0))) + return VNET_DEVICE_INPUT_NEXT_IP6_INPUT; + else if (PREDICT_TRUE (vlib_buffer_is_mpls (b0))) + return VNET_DEVICE_INPUT_NEXT_MPLS_INPUT; + else + return VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; +} + +always_inline int +dpdk_mbuf_is_vlan (struct rte_mbuf *mb) +{ +#if RTE_VERSION >= RTE_VERSION_NUM(16, 11, 0, 0) + return (mb->packet_type & RTE_PTYPE_L2_ETHER_VLAN) == + RTE_PTYPE_L2_ETHER_VLAN; +#else + return + (mb->ol_flags & + (PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED | PKT_RX_QINQ_STRIPPED)) == + PKT_RX_VLAN_PKT; +#endif +} + +always_inline int +dpdk_mbuf_is_ip4 (struct rte_mbuf *mb) +{ + return RTE_ETH_IS_IPV4_HDR (mb->packet_type) != 0; +} + +always_inline int +dpdk_mbuf_is_ip6 (struct rte_mbuf *mb) +{ + return RTE_ETH_IS_IPV6_HDR (mb->packet_type) != 0; +} + +always_inline u32 +dpdk_rx_next_from_mb (struct rte_mbuf * mb, vlib_buffer_t * b0) +{ + if (PREDICT_FALSE (dpdk_mbuf_is_vlan (mb))) + return VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; + else if (PREDICT_TRUE (dpdk_mbuf_is_ip4 (mb))) + return VNET_DEVICE_INPUT_NEXT_IP4_NCS_INPUT; + else if (PREDICT_TRUE (dpdk_mbuf_is_ip6 (mb))) + return VNET_DEVICE_INPUT_NEXT_IP6_INPUT; + else if (PREDICT_TRUE (vlib_buffer_is_mpls (b0))) + return VNET_DEVICE_INPUT_NEXT_MPLS_INPUT; + else + return dpdk_rx_next_from_etype (mb, b0); +} + +always_inline void +dpdk_rx_error_from_mb (struct rte_mbuf *mb, u32 * next, u8 * error) +{ + if (mb->ol_flags & PKT_RX_IP_CKSUM_BAD) + { + *error = DPDK_ERROR_IP_CHECKSUM_ERROR; + *next = VNET_DEVICE_INPUT_NEXT_DROP; + } + else + *error = DPDK_ERROR_NONE; +} + +void +dpdk_rx_trace (dpdk_main_t * dm, + vlib_node_runtime_t * node, + dpdk_device_t * xd, + u16 queue_id, u32 * buffers, uword n_buffers) +{ + vlib_main_t *vm = vlib_get_main (); + u32 *b, n_left; + u32 next0; + + n_left = n_buffers; + b = buffers; + + while (n_left >= 1) + { + u32 bi0; + vlib_buffer_t *b0; + dpdk_rx_dma_trace_t *t0; + struct rte_mbuf *mb; + u8 error0; + + bi0 = b[0]; + n_left -= 1; + + b0 = vlib_get_buffer (vm, bi0); + mb = rte_mbuf_from_vlib_buffer (b0); + + if (PREDICT_FALSE (xd->per_interface_next_index != ~0)) + next0 = xd->per_interface_next_index; + else if (PREDICT_TRUE + ((xd->flags & DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE) != 0)) + next0 = dpdk_rx_next_from_mb (mb, b0); + else + next0 = dpdk_rx_next_from_etype (mb, b0); + + dpdk_rx_error_from_mb (mb, &next0, &error0); + + vlib_trace_buffer (vm, node, next0, b0, /* follow_chain */ 0); + t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0])); + t0->queue_index = queue_id; + t0->device_index = xd->device_index; + t0->buffer_index = bi0; + + clib_memcpy (&t0->mb, mb, sizeof (t0->mb)); + clib_memcpy (&t0->buffer, b0, sizeof (b0[0]) - sizeof (b0->pre_data)); + clib_memcpy (t0->buffer.pre_data, b0->data, + sizeof (t0->buffer.pre_data)); + clib_memcpy (&t0->data, mb->buf_addr + mb->data_off, sizeof (t0->data)); + + b += 1; + } +} + +static inline u32 +dpdk_rx_burst (dpdk_main_t * dm, dpdk_device_t * xd, u16 queue_id) +{ + u32 n_buffers; + u32 n_left; + u32 n_this_chunk; + + n_left = VLIB_FRAME_SIZE; + n_buffers = 0; + + if (PREDICT_TRUE (xd->flags & DPDK_DEVICE_FLAG_PMD)) + { + while (n_left) + { + n_this_chunk = rte_eth_rx_burst (xd->device_index, queue_id, + xd->rx_vectors[queue_id] + + n_buffers, n_left); + n_buffers += n_this_chunk; + n_left -= n_this_chunk; + + /* Empirically, DPDK r1.8 produces vectors w/ 32 or fewer elts */ + if (n_this_chunk < 32) + break; + } + } + else + { + ASSERT (0); + } + + return n_buffers; +} + + +static_always_inline void +dpdk_process_subseq_segs (vlib_main_t * vm, vlib_buffer_t * b, + struct rte_mbuf *mb, vlib_buffer_free_list_t * fl) +{ + u8 nb_seg = 1; + struct rte_mbuf *mb_seg = 0; + vlib_buffer_t *b_seg, *b_chain = 0; + mb_seg = mb->next; + b_chain = b; + + while ((mb->nb_segs > 1) && (nb_seg < mb->nb_segs)) + { + ASSERT (mb_seg != 0); + + b_seg = vlib_buffer_from_rte_mbuf (mb_seg); + vlib_buffer_init_for_free_list (b_seg, fl); + + ASSERT ((b_seg->flags & VLIB_BUFFER_NEXT_PRESENT) == 0); + ASSERT (b_seg->current_data == 0); + + /* + * The driver (e.g. virtio) may not put the packet data at the start + * of the segment, so don't assume b_seg->current_data == 0 is correct. + */ + b_seg->current_data = + (mb_seg->buf_addr + mb_seg->data_off) - (void *) b_seg->data; + + b_seg->current_length = mb_seg->data_len; + b->total_length_not_including_first_buffer += mb_seg->data_len; + + b_chain->flags |= VLIB_BUFFER_NEXT_PRESENT; + b_chain->next_buffer = vlib_get_buffer_index (vm, b_seg); + + b_chain = b_seg; + mb_seg = mb_seg->next; + nb_seg++; + } +} + +static_always_inline void +dpdk_prefetch_buffer (struct rte_mbuf *mb) +{ + vlib_buffer_t *b = vlib_buffer_from_rte_mbuf (mb); + CLIB_PREFETCH (mb, CLIB_CACHE_LINE_BYTES, LOAD); + CLIB_PREFETCH (b, CLIB_CACHE_LINE_BYTES, STORE); +} + +/* + * This function is used when there are no worker threads. + * The main thread performs IO and forwards the packets. + */ +static_always_inline u32 +dpdk_device_input (dpdk_main_t * dm, dpdk_device_t * xd, + vlib_node_runtime_t * node, u32 cpu_index, u16 queue_id) +{ + u32 n_buffers; + u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT; + u32 n_left_to_next, *to_next; + u32 mb_index; + vlib_main_t *vm = vlib_get_main (); + uword n_rx_bytes = 0; + u32 n_trace, trace_cnt __attribute__ ((unused)); + vlib_buffer_free_list_t *fl; + u32 buffer_flags_template; + + if ((xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) == 0) + return 0; + + n_buffers = dpdk_rx_burst (dm, xd, queue_id); + + if (n_buffers == 0) + { + return 0; + } + + buffer_flags_template = dm->buffer_flags_template; + + vec_reset_length (xd->d_trace_buffers[cpu_index]); + trace_cnt = n_trace = vlib_get_trace_count (vm, node); + + if (n_trace > 0) + { + u32 n = clib_min (n_trace, n_buffers); + mb_index = 0; + + while (n--) + { + struct rte_mbuf *mb = xd->rx_vectors[queue_id][mb_index++]; + vlib_buffer_t *b = vlib_buffer_from_rte_mbuf (mb); + vec_add1 (xd->d_trace_buffers[cpu_index], + vlib_get_buffer_index (vm, b)); + } + } + + fl = vlib_buffer_get_free_list (vm, VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); + + mb_index = 0; + + while (n_buffers > 0) + { + vlib_buffer_t *b0, *b1, *b2, *b3; + u32 bi0, next0, l3_offset0; + u32 bi1, next1, l3_offset1; + u32 bi2, next2, l3_offset2; + u32 bi3, next3, l3_offset3; + u8 error0, error1, error2, error3; + u64 or_ol_flags; + + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_buffers > 8 && n_left_to_next > 4) + { + struct rte_mbuf *mb0 = xd->rx_vectors[queue_id][mb_index]; + struct rte_mbuf *mb1 = xd->rx_vectors[queue_id][mb_index + 1]; + struct rte_mbuf *mb2 = xd->rx_vectors[queue_id][mb_index + 2]; + struct rte_mbuf *mb3 = xd->rx_vectors[queue_id][mb_index + 3]; + + dpdk_prefetch_buffer (xd->rx_vectors[queue_id][mb_index + 4]); + dpdk_prefetch_buffer (xd->rx_vectors[queue_id][mb_index + 5]); + dpdk_prefetch_buffer (xd->rx_vectors[queue_id][mb_index + 6]); + dpdk_prefetch_buffer (xd->rx_vectors[queue_id][mb_index + 7]); + + if (xd->flags & DPDK_DEVICE_FLAG_MAYBE_MULTISEG) + { + if (PREDICT_FALSE (mb0->nb_segs > 1)) + dpdk_prefetch_buffer (mb0->next); + if (PREDICT_FALSE (mb1->nb_segs > 1)) + dpdk_prefetch_buffer (mb1->next); + if (PREDICT_FALSE (mb2->nb_segs > 1)) + dpdk_prefetch_buffer (mb2->next); + if (PREDICT_FALSE (mb3->nb_segs > 1)) + dpdk_prefetch_buffer (mb3->next); + } + + ASSERT (mb0); + ASSERT (mb1); + ASSERT (mb2); + ASSERT (mb3); + + or_ol_flags = (mb0->ol_flags | mb1->ol_flags | + mb2->ol_flags | mb3->ol_flags); + b0 = vlib_buffer_from_rte_mbuf (mb0); + b1 = vlib_buffer_from_rte_mbuf (mb1); + b2 = vlib_buffer_from_rte_mbuf (mb2); + b3 = vlib_buffer_from_rte_mbuf (mb3); + + vlib_buffer_init_for_free_list (b0, fl); + vlib_buffer_init_for_free_list (b1, fl); + vlib_buffer_init_for_free_list (b2, fl); + vlib_buffer_init_for_free_list (b3, fl); + + bi0 = vlib_get_buffer_index (vm, b0); + bi1 = vlib_get_buffer_index (vm, b1); + bi2 = vlib_get_buffer_index (vm, b2); + bi3 = vlib_get_buffer_index (vm, b3); + + to_next[0] = bi0; + to_next[1] = bi1; + to_next[2] = bi2; + to_next[3] = bi3; + to_next += 4; + n_left_to_next -= 4; + + if (PREDICT_FALSE (xd->per_interface_next_index != ~0)) + { + next0 = next1 = next2 = next3 = xd->per_interface_next_index; + } + else if (PREDICT_TRUE + ((xd->flags & DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE) != 0)) + { + next0 = dpdk_rx_next_from_mb (mb0, b0); + next1 = dpdk_rx_next_from_mb (mb1, b1); + next2 = dpdk_rx_next_from_mb (mb2, b2); + next3 = dpdk_rx_next_from_mb (mb3, b3); + } + else + { + next0 = dpdk_rx_next_from_etype (mb0, b0); + next1 = dpdk_rx_next_from_etype (mb1, b1); + next2 = dpdk_rx_next_from_etype (mb2, b2); + next3 = dpdk_rx_next_from_etype (mb3, b3); + } + + if (PREDICT_FALSE (or_ol_flags & PKT_RX_IP_CKSUM_BAD)) + { + dpdk_rx_error_from_mb (mb0, &next0, &error0); + dpdk_rx_error_from_mb (mb1, &next1, &error1); + dpdk_rx_error_from_mb (mb2, &next2, &error2); + dpdk_rx_error_from_mb (mb3, &next3, &error3); + b0->error = node->errors[error0]; + b1->error = node->errors[error1]; + b2->error = node->errors[error2]; + b3->error = node->errors[error3]; + } + else + { + b0->error = b1->error = node->errors[DPDK_ERROR_NONE]; + b2->error = b3->error = node->errors[DPDK_ERROR_NONE]; + } + + l3_offset0 = device_input_next_node_advance[next0]; + l3_offset1 = device_input_next_node_advance[next1]; + l3_offset2 = device_input_next_node_advance[next2]; + l3_offset3 = device_input_next_node_advance[next3]; + + b0->current_data = l3_offset0 + mb0->data_off; + b1->current_data = l3_offset1 + mb1->data_off; + b2->current_data = l3_offset2 + mb2->data_off; + b3->current_data = l3_offset3 + mb3->data_off; + + b0->current_data -= RTE_PKTMBUF_HEADROOM; + b1->current_data -= RTE_PKTMBUF_HEADROOM; + b2->current_data -= RTE_PKTMBUF_HEADROOM; + b3->current_data -= RTE_PKTMBUF_HEADROOM; + + b0->current_length = mb0->data_len - l3_offset0; + b1->current_length = mb1->data_len - l3_offset1; + b2->current_length = mb2->data_len - l3_offset2; + b3->current_length = mb3->data_len - l3_offset3; + + b0->flags = buffer_flags_template; + b1->flags = buffer_flags_template; + b2->flags = buffer_flags_template; + b3->flags = buffer_flags_template; + + vnet_buffer (b0)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b1)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b2)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b3)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + vnet_buffer (b1)->sw_if_index[VLIB_TX] = (u32) ~ 0; + vnet_buffer (b2)->sw_if_index[VLIB_TX] = (u32) ~ 0; + vnet_buffer (b3)->sw_if_index[VLIB_TX] = (u32) ~ 0; + + n_rx_bytes += mb0->pkt_len; + n_rx_bytes += mb1->pkt_len; + n_rx_bytes += mb2->pkt_len; + n_rx_bytes += mb3->pkt_len; + + /* Process subsequent segments of multi-segment packets */ + if (xd->flags & DPDK_DEVICE_FLAG_MAYBE_MULTISEG) + { + dpdk_process_subseq_segs (vm, b0, mb0, fl); + dpdk_process_subseq_segs (vm, b1, mb1, fl); + dpdk_process_subseq_segs (vm, b2, mb2, fl); + dpdk_process_subseq_segs (vm, b3, mb3, fl); + } + + /* + * Turn this on if you run into + * "bad monkey" contexts, and you want to know exactly + * which nodes they've visited... See main.c... + */ + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b1); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b2); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b3); + + /* Do we have any driver RX features configured on the interface? */ + vnet_feature_start_device_input_x4 (xd->vlib_sw_if_index, + &next0, &next1, &next2, &next3, + b0, b1, b2, b3, + l3_offset0, l3_offset1, + l3_offset2, l3_offset3); + + vlib_validate_buffer_enqueue_x4 (vm, node, next_index, + to_next, n_left_to_next, + bi0, bi1, bi2, bi3, + next0, next1, next2, next3); + n_buffers -= 4; + mb_index += 4; + } + while (n_buffers > 0 && n_left_to_next > 0) + { + struct rte_mbuf *mb0 = xd->rx_vectors[queue_id][mb_index]; + + ASSERT (mb0); + + b0 = vlib_buffer_from_rte_mbuf (mb0); + + /* Prefetch one next segment if it exists. */ + if (PREDICT_FALSE (mb0->nb_segs > 1)) + dpdk_prefetch_buffer (mb0->next); + + vlib_buffer_init_for_free_list (b0, fl); + + bi0 = vlib_get_buffer_index (vm, b0); + + to_next[0] = bi0; + to_next++; + n_left_to_next--; + + if (PREDICT_FALSE (xd->per_interface_next_index != ~0)) + next0 = xd->per_interface_next_index; + else if (PREDICT_TRUE + ((xd->flags & DPDK_DEVICE_FLAG_PMD_SUPPORTS_PTYPE) != 0)) + next0 = dpdk_rx_next_from_mb (mb0, b0); + else + next0 = dpdk_rx_next_from_etype (mb0, b0); + + dpdk_rx_error_from_mb (mb0, &next0, &error0); + b0->error = node->errors[error0]; + + l3_offset0 = device_input_next_node_advance[next0]; + + b0->current_data = l3_offset0; + b0->current_data += mb0->data_off - RTE_PKTMBUF_HEADROOM; + b0->current_length = mb0->data_len - l3_offset0; + + b0->flags = buffer_flags_template; + + vnet_buffer (b0)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index; + vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0; + n_rx_bytes += mb0->pkt_len; + + /* Process subsequent segments of multi-segment packets */ + dpdk_process_subseq_segs (vm, b0, mb0, fl); + + /* + * Turn this on if you run into + * "bad monkey" contexts, and you want to know exactly + * which nodes they've visited... See main.c... + */ + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); + + /* Do we have any driver RX features configured on the interface? */ + vnet_feature_start_device_input_x1 (xd->vlib_sw_if_index, &next0, + b0, l3_offset0); + + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, + bi0, next0); + n_buffers--; + mb_index++; + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next); + } + + if (PREDICT_FALSE (vec_len (xd->d_trace_buffers[cpu_index]) > 0)) + { + dpdk_rx_trace (dm, node, xd, queue_id, xd->d_trace_buffers[cpu_index], + vec_len (xd->d_trace_buffers[cpu_index])); + vlib_set_trace_count (vm, node, n_trace - + vec_len (xd->d_trace_buffers[cpu_index])); + } + + vlib_increment_combined_counter + (vnet_get_main ()->interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, + cpu_index, xd->vlib_sw_if_index, mb_index, n_rx_bytes); + + dpdk_worker_t *dw = vec_elt_at_index (dm->workers, cpu_index); + dw->aggregate_rx_packets += mb_index; + + return mb_index; +} + +static inline void +poll_rate_limit (dpdk_main_t * dm) +{ + /* Limit the poll rate by sleeping for N msec between polls */ + if (PREDICT_FALSE (dm->poll_sleep != 0)) + { + struct timespec ts, tsrem; + + ts.tv_sec = 0; + ts.tv_nsec = 1000 * 1000 * dm->poll_sleep; /* 1ms */ + + while (nanosleep (&ts, &tsrem) < 0) + { + ts = tsrem; + } + } +} + +/** \brief Main DPDK input node + @node dpdk-input + + This is the main DPDK input node: across each assigned interface, + call rte_eth_rx_burst(...) or similar to obtain a vector of + packets to process. Handle early packet discard. Derive @c + vlib_buffer_t metadata from <code>struct rte_mbuf</code> metadata, + Depending on the resulting metadata: adjust <code>b->current_data, + b->current_length </code> and dispatch directly to + ip4-input-no-checksum, or ip6-input. Trace the packet if required. + + @param vm vlib_main_t corresponding to the current thread + @param node vlib_node_runtime_t + @param f vlib_frame_t input-node, not used. + + @par Graph mechanics: buffer metadata, next index usage + + @em Uses: + - <code>struct rte_mbuf mb->ol_flags</code> + - PKT_RX_IP_CKSUM_BAD + - <code> RTE_ETH_IS_xxx_HDR(mb->packet_type) </code> + - packet classification result + + @em Sets: + - <code>b->error</code> if the packet is to be dropped immediately + - <code>b->current_data, b->current_length</code> + - adjusted as needed to skip the L2 header in direct-dispatch cases + - <code>vnet_buffer(b)->sw_if_index[VLIB_RX]</code> + - rx interface sw_if_index + - <code>vnet_buffer(b)->sw_if_index[VLIB_TX] = ~0</code> + - required by ipX-lookup + - <code>b->flags</code> + - to indicate multi-segment pkts (VLIB_BUFFER_NEXT_PRESENT), etc. + + <em>Next Nodes:</em> + - Static arcs to: error-drop, ethernet-input, + ip4-input-no-checksum, ip6-input, mpls-input + - per-interface redirection, controlled by + <code>xd->per_interface_next_index</code> +*/ + +static uword +dpdk_input (vlib_main_t * vm, vlib_node_runtime_t * node, vlib_frame_t * f) +{ + dpdk_main_t *dm = &dpdk_main; + dpdk_device_t *xd; + uword n_rx_packets = 0; + dpdk_device_and_queue_t *dq; + u32 cpu_index = os_get_cpu_number (); + + /* + * Poll all devices on this cpu for input/interrupts. + */ + /* *INDENT-OFF* */ + vec_foreach (dq, dm->devices_by_cpu[cpu_index]) + { + xd = vec_elt_at_index(dm->devices, dq->device); + n_rx_packets += dpdk_device_input (dm, xd, node, cpu_index, dq->queue_id); + } + /* *INDENT-ON* */ + + poll_rate_limit (dm); + + return n_rx_packets; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (dpdk_input_node) = { + .function = dpdk_input, + .type = VLIB_NODE_TYPE_INPUT, + .name = "dpdk-input", + .sibling_of = "device-input", + + /* Will be enabled if/when hardware is detected. */ + .state = VLIB_NODE_STATE_DISABLED, + + .format_buffer = format_ethernet_header_with_length, + .format_trace = format_dpdk_rx_dma_trace, + + .n_errors = DPDK_N_ERROR, + .error_strings = dpdk_error_strings, +}; + +VLIB_NODE_FUNCTION_MULTIARCH (dpdk_input_node, dpdk_input); +/* *INDENT-ON* */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vnet/devices/dpdk/qos_doc.md b/src/vnet/devices/dpdk/qos_doc.md new file mode 100644 index 00000000000..9bd0659d616 --- /dev/null +++ b/src/vnet/devices/dpdk/qos_doc.md @@ -0,0 +1,404 @@ +# QoS Hierarchical Scheduler {#qos_doc} + +The Quality-of-Service (QoS) scheduler performs egress-traffic management by +prioritizing the transmission of the packets of different type services and +subcribers based on the Service Level Agreements (SLAs). The QoS scheduler can +be enabled on one or more NIC output interfaces depending upon the +requirement. + + +## Overview + +The QoS schdeuler supports a number of scheduling and shaping levels which +construct hierarchical-tree. The first level in the hierarchy is port (i.e. +the physical interface) that constitutes the root node of the tree. The +subsequent level is subport which represents the group of the +users/subscribers. The individual user/subscriber is represented by the pipe +at the next level. Each user can have different traffic type based on the +criteria of specific loss rate, jitter, and latency. These traffic types are +represented at the traffic-class level in the form of different traffic- +classes. The last level contains number of queues which are grouped together +to host the packets of the specific class type traffic. + +The QoS scheduler implementation requires flow classification, enqueue and +dequeue operations. The flow classification is mandatory stage for HQoS where +incoming packets are classified by mapping the packet fields information to +5-tuple (HQoS subport, pipe, traffic class, queue within traffic class, and +color) and storing that information in mbuf sched field. The enqueue operation +uses this information to determine the queue for storing the packet, and at +this stage, if the specific queue is full, QoS drops the packet. The dequeue +operation consists of scheduling the packet based on its length and available +credits, and handing over the scheduled packet to the output interface. + +For more information on QoS Scheduler, please refer DPDK Programmer's Guide- +http://dpdk.org/doc/guides/prog_guide/qos_framework.html + + +### QoS Schdeuler Parameters + +Following illustrates the default HQoS configuration for each 10GbE output +port: + +Single subport (subport 0): + - Subport rate set to 100% of port rate + - Each of the 4 traffic classes has rate set to 100% of port rate + +4K pipes per subport 0 (pipes 0 .. 4095) with identical configuration: + - Pipe rate set to 1/4K of port rate + - Each of the 4 traffic classes has rate set to 100% of pipe rate + - Within each traffic class, the byte-level WRR weights for the 4 queues are set to 1:1:1:1 + + +#### Port configuration + +``` +port { + rate 1250000000 /* Assuming 10GbE port */ + frame_overhead 24 /* Overhead fields per Ethernet frame: + * 7B (Preamble) + + * 1B (Start of Frame Delimiter (SFD)) + + * 4B (Frame Check Sequence (FCS)) + + * 12B (Inter Frame Gap (IFG)) + */ + mtu 1522 /* Assuming Ethernet/IPv4 pkt (FCS not included) */ + n_subports_per_port 1 /* Number of subports per output interface */ + n_pipes_per_subport 4096 /* Number of pipes (users/subscribers) */ + queue_sizes 64 64 64 64 /* Packet queue size for each traffic class. + * All queues within the same pipe traffic class + * have the same size. Queues from different + * pipes serving the same traffic class have + * the same size. */ +} +``` + + +#### Subport configuration + +``` +subport 0 { + tb_rate 1250000000 /* Subport level token bucket rate (bytes per second) */ + tb_size 1000000 /* Subport level token bucket size (bytes) */ + tc0_rate 1250000000 /* Subport level token bucket rate for traffic class 0 (bytes per second) */ + tc1_rate 1250000000 /* Subport level token bucket rate for traffic class 1 (bytes per second) */ + tc2_rate 1250000000 /* Subport level token bucket rate for traffic class 2 (bytes per second) */ + tc3_rate 1250000000 /* Subport level token bucket rate for traffic class 3 (bytes per second) */ + tc_period 10 /* Time interval for refilling the token bucket associated with traffic class (Milliseconds) */ + pipe 0 4095 profile 0 /* pipes (users/subscribers) configured with pipe profile 0 */ +} +``` + + +#### Pipe configuration + +``` +pipe_profile 0 { + tb_rate 305175 /* Pipe level token bucket rate (bytes per second) */ + tb_size 1000000 /* Pipe level token bucket size (bytes) */ + tc0_rate 305175 /* Pipe level token bucket rate for traffic class 0 (bytes per second) */ + tc1_rate 305175 /* Pipe level token bucket rate for traffic class 1 (bytes per second) */ + tc2_rate 305175 /* Pipe level token bucket rate for traffic class 2 (bytes per second) */ + tc3_rate 305175 /* Pipe level token bucket rate for traffic class 3 (bytes per second) */ + tc_period 40 /* Time interval for refilling the token bucket associated with traffic class at pipe level (Milliseconds) */ + tc3_oversubscription_weight 1 /* Weight traffic class 3 oversubscription */ + tc0_wrr_weights 1 1 1 1 /* Pipe queues WRR weights for traffic class 0 */ + tc1_wrr_weights 1 1 1 1 /* Pipe queues WRR weights for traffic class 1 */ + tc2_wrr_weights 1 1 1 1 /* Pipe queues WRR weights for traffic class 2 */ + tc3_wrr_weights 1 1 1 1 /* Pipe queues WRR weights for traffic class 3 */ +} +``` + + +#### Random Early Detection (RED) parameters per traffic class and color (Green / Yellow / Red) + +``` +red { + tc0_wred_min 48 40 32 /* Minimum threshold for traffic class 0 queue (min_th) in number of packets */ + tc0_wred_max 64 64 64 /* Maximum threshold for traffic class 0 queue (max_th) in number of packets */ + tc0_wred_inv_prob 10 10 10 /* Inverse of packet marking probability for traffic class 0 queue (maxp = 1 / maxp_inv) */ + tc0_wred_weight 9 9 9 /* Traffic Class 0 queue weight */ + tc1_wred_min 48 40 32 /* Minimum threshold for traffic class 1 queue (min_th) in number of packets */ + tc1_wred_max 64 64 64 /* Maximum threshold for traffic class 1 queue (max_th) in number of packets */ + tc1_wred_inv_prob 10 10 10 /* Inverse of packet marking probability for traffic class 1 queue (maxp = 1 / maxp_inv) */ + tc1_wred_weight 9 9 9 /* Traffic Class 1 queue weight */ + tc2_wred_min 48 40 32 /* Minimum threshold for traffic class 2 queue (min_th) in number of packets */ + tc2_wred_max 64 64 64 /* Maximum threshold for traffic class 2 queue (max_th) in number of packets */ + tc2_wred_inv_prob 10 10 10 /* Inverse of packet marking probability for traffic class 2 queue (maxp = 1 / maxp_inv) */ + tc2_wred_weight 9 9 9 /* Traffic Class 2 queue weight */ + tc3_wred_min 48 40 32 /* Minimum threshold for traffic class 3 queue (min_th) in number of packets */ + tc3_wred_max 64 64 64 /* Maximum threshold for traffic class 3 queue (max_th) in number of packets */ + tc3_wred_inv_prob 10 10 10 /* Inverse of packet marking probability for traffic class 3 queue (maxp = 1 / maxp_inv) */ + tc3_wred_weight 9 9 9 /* Traffic Class 3 queue weight */ +} +``` + + +### DPDK QoS Scheduler Integration in VPP + +The Hierarchical Quaity-of-Service (HQoS) scheduler object could be seen as +part of the logical NIC output interface. To enable HQoS on specific output +interface, vpp startup.conf file has to be configured accordingly. The output +interface that requires HQoS, should have "hqos" parameter specified in dpdk +section. Another optional parameter "hqos-thread" has been defined which can +be used to associate the output interface with specific hqos thread. In cpu +section of the config file, "corelist-hqos-threads" is introduced to assign +logical cpu cores to run the HQoS threads. A HQoS thread can run multiple HQoS +objects each associated with different output interfaces. All worker threads +instead of writing packets to NIC TX queue directly, write the packets to a +software queues. The hqos_threads read the software queues, and enqueue the +packets to HQoS objects, as well as dequeue packets from HQOS objects and +write them to NIC output interfaces. The worker threads need to be able to +send the packets to any output interface, therefore, each HQoS object +associated with NIC output interface should have software queues equal to +worker threads count. + +Following illustrates the sample startup configuration file with 4x worker +threads feeding 2x hqos threads that handle each QoS scheduler for 1x output +interface. + +``` +dpdk { + socket-mem 16384,16384 + + dev 0000:02:00.0 { + num-rx-queues 2 + hqos + } + dev 0000:06:00.0 { + num-rx-queues 2 + hqos + } + + num-mbufs 1000000 +} + +cpu { + main-core 0 + corelist-workers 1, 2, 3, 4 + corelist-hqos-threads 5, 6 +} +``` + + +### QoS scheduler CLI Commands + +Each QoS scheduler instance is initialised with default parameters required to +configure hqos port, subport, pipe and queues. Some of the parameters can be +re-configured in run-time through CLI commands. + + +#### Configuration + +Following commands can be used to configure QoS scheduler parameters. + +The command below can be used to set the subport level parameters such as +token bucket rate (bytes per seconds), token bucket size (bytes), traffic +class rates (bytes per seconds) and token update period (Milliseconds). + +``` +set dpdk interface hqos subport <if-name> subport <n> [rate <n>] + [bktsize <n>] [tc0 <n>] [tc1 <n>] [tc2 <n>] [tc3 <n>] [period <n>] +``` + +For setting the pipe profile, following command can be used. + +``` +set dpdk interface hqos pipe <if-name> subport <n> pipe <n> profile <n> +``` + +To assign QoS scheduler instance to the specific thread, following command can +be used. + +``` +set dpdk interface hqos placement <if-name> thread <n> +``` + +The command below is used to set the packet fields required for classifiying +the incoming packet. As a result of classification process, packet field +information will be mapped to 5 tuples (subport, pipe, traffic class, pipe, +color) and stored in packet mbuf. + +``` +set dpdk interface hqos pktfield <if-name> id <n> offset <n> mask <n> +``` + +The DSCP table entries used for idenfiying the traffic class and queue can be set using the command below; + +``` +set dpdk interface hqos tctbl <if-name> entry <n> tc <n> queue <n> +``` + + +#### Show Command + +The QoS Scheduler configuration can displayed using the command below. + +``` + vpp# show dpdk interface hqos TenGigabitEthernet2/0/0 + Thread: + Input SWQ size = 4096 packets + Enqueue burst size = 256 packets + Dequeue burst size = 220 packets + Packet field 0: slab position = 0, slab bitmask = 0x0000000000000000 + Packet field 1: slab position = 40, slab bitmask = 0x0000000fff000000 + Packet field 2: slab position = 8, slab bitmask = 0x00000000000000fc + Packet field 2 translation table: + [ 0 .. 15]: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 + [16 .. 31]: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 + [32 .. 47]: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 + [48 .. 63]: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 + Port: + Rate = 1250000000 bytes/second + MTU = 1514 bytes + Frame overhead = 24 bytes + Number of subports = 1 + Number of pipes per subport = 4096 + Packet queue size: TC0 = 64, TC1 = 64, TC2 = 64, TC3 = 64 packets + Number of pipe profiles = 1 + Pipe profile 0: + Rate = 305175 bytes/second + Token bucket size = 1000000 bytes + Traffic class rate: TC0 = 305175, TC1 = 305175, TC2 = 305175, TC3 = 305175 bytes/second + TC period = 40 milliseconds + TC0 WRR weights: Q0 = 1, Q1 = 1, Q2 = 1, Q3 = 1 + TC1 WRR weights: Q0 = 1, Q1 = 1, Q2 = 1, Q3 = 1 + TC2 WRR weights: Q0 = 1, Q1 = 1, Q2 = 1, Q3 = 1 + TC3 WRR weights: Q0 = 1, Q1 = 1, Q2 = 1, Q3 = 1 +``` + +The QoS Scheduler placement over the logical cpu cores can be displayed using +below command. + +``` + vpp# show dpdk interface hqos placement + Thread 5 (vpp_hqos-threads_0 at lcore 5): + TenGigabitEthernet2/0/0 queue 0 + Thread 6 (vpp_hqos-threads_1 at lcore 6): + TenGigabitEthernet4/0/1 queue 0 +``` + + +### QoS Scheduler Binary APIs + +This section explans the available binary APIs for configuring QoS scheduler +parameters in run-time. + +The following API can be used to set the pipe profile of a pipe that belongs +to a given subport: + +``` +sw_interface_set_dpdk_hqos_pipe rx <intfc> | sw_if_index <id> + subport <subport-id> pipe <pipe-id> profile <profile-id> +``` + +The data structures used for set the pipe profile parameter are as follows; + +``` + /** \\brief DPDK interface HQoS pipe profile set request + @param client_index - opaque cookie to identify the sender + @param context - sender context, to match reply w/ request + @param sw_if_index - the interface + @param subport - subport ID + @param pipe - pipe ID within its subport + @param profile - pipe profile ID + */ + define sw_interface_set_dpdk_hqos_pipe { + u32 client_index; + u32 context; + u32 sw_if_index; + u32 subport; + u32 pipe; + u32 profile; + }; + + /** \\brief DPDK interface HQoS pipe profile set reply + @param context - sender context, to match reply w/ request + @param retval - request return code + */ + define sw_interface_set_dpdk_hqos_pipe_reply { + u32 context; + i32 retval; + }; +``` + +The following API can be used to set the subport level parameters, for +example- token bucket rate (bytes per seconds), token bucket size (bytes), +traffic class rate (bytes per seconds) and tokens update period. + +``` +sw_interface_set_dpdk_hqos_subport rx <intfc> | sw_if_index <id> + subport <subport-id> [rate <n>] [bktsize <n>] + [tc0 <n>] [tc1 <n>] [tc2 <n>] [tc3 <n>] [period <n>] +``` + +The data structures used for set the subport level parameter are as follows; + +``` + /** \\brief DPDK interface HQoS subport parameters set request + @param client_index - opaque cookie to identify the sender + @param context - sender context, to match reply w/ request + @param sw_if_index - the interface + @param subport - subport ID + @param tb_rate - subport token bucket rate (measured in bytes/second) + @param tb_size - subport token bucket size (measured in credits) + @param tc_rate - subport traffic class 0 .. 3 rates (measured in bytes/second) + @param tc_period - enforcement period for rates (measured in milliseconds) + */ + define sw_interface_set_dpdk_hqos_subport { + u32 client_index; + u32 context; + u32 sw_if_index; + u32 subport; + u32 tb_rate; + u32 tb_size; + u32 tc_rate[4]; + u32 tc_period; + }; + + /** \\brief DPDK interface HQoS subport parameters set reply + @param context - sender context, to match reply w/ request + @param retval - request return code + */ + define sw_interface_set_dpdk_hqos_subport_reply { + u32 context; + i32 retval; + }; +``` + +The following API can be used set the DSCP table entry. The DSCP table have +64 entries to map the packet DSCP field onto traffic class and hqos input +queue. + +``` +sw_interface_set_dpdk_hqos_tctbl rx <intfc> | sw_if_index <id> + entry <n> tc <n> queue <n> +``` + +The data structures used for setting DSCP table entries are given below. + +``` + /** \\brief DPDK interface HQoS tctbl entry set request + @param client_index - opaque cookie to identify the sender + @param context - sender context, to match reply w/ request + @param sw_if_index - the interface + @param entry - entry index ID + @param tc - traffic class (0 .. 3) + @param queue - traffic class queue (0 .. 3) + */ + define sw_interface_set_dpdk_hqos_tctbl { + u32 client_index; + u32 context; + u32 sw_if_index; + u32 entry; + u32 tc; + u32 queue; + }; + + /** \\brief DPDK interface HQoS tctbl entry set reply + @param context - sender context, to match reply w/ request + @param retval - request return code + */ + define sw_interface_set_dpdk_hqos_tctbl_reply { + u32 context; + i32 retval; + }; +``` |