1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2020 Mellanox Technologies, Ltd
13 #include <linux/rtnetlink.h>
14 #include <linux/sockios.h>
15 #include <linux/ethtool.h>
18 #include <rte_malloc.h>
19 #include <ethdev_driver.h>
20 #include <ethdev_pci.h>
22 #include <rte_bus_pci.h>
23 #include <rte_common.h>
24 #include <rte_kvargs.h>
25 #include <rte_rwlock.h>
26 #include <rte_spinlock.h>
27 #include <rte_string_fns.h>
28 #include <rte_alarm.h>
29 #include <rte_eal_paging.h>
31 #include <mlx5_glue.h>
32 #include <mlx5_devx_cmds.h>
33 #include <mlx5_common.h>
34 #include <mlx5_common_mp.h>
35 #include <mlx5_common_mr.h>
36 #include <mlx5_malloc.h>
38 #include "mlx5_defs.h"
40 #include "mlx5_common_os.h"
41 #include "mlx5_utils.h"
42 #include "mlx5_rxtx.h"
43 #include "mlx5_autoconf.h"
45 #include "mlx5_flow.h"
46 #include "rte_pmd_mlx5.h"
47 #include "mlx5_verbs.h"
49 #include "mlx5_devx.h"
51 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
53 #ifndef HAVE_IBV_MLX5_MOD_MPW
54 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
55 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
58 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
59 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
62 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
64 /* Spinlock for mlx5_shared_data allocation. */
65 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
67 /* Process local data for secondary processes. */
68 static struct mlx5_local_data mlx5_local_data;
71 * Set the completion channel file descriptor interrupt as non-blocking.
74 * Pointer to RQ channel object, which includes the channel fd
77 * The file descriptor (representing the intetrrupt) used in this channel.
80 * 0 on successfully setting the fd to non-blocking, non-zero otherwise.
83 mlx5_os_set_nonblock_channel_fd(int fd)
87 flags = fcntl(fd, F_GETFL);
88 return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
92 * Get mlx5 device attributes. The glue function query_device_ex() is called
93 * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
94 * device attributes from the glue out parameter.
97 * Pointer to ibv context.
100 * Pointer to mlx5 device attributes.
103 * 0 on success, non zero error number otherwise
106 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
109 struct ibv_device_attr_ex attr_ex;
110 memset(device_attr, 0, sizeof(*device_attr));
111 err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
115 device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
116 device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
117 device_attr->max_sge = attr_ex.orig_attr.max_sge;
118 device_attr->max_cq = attr_ex.orig_attr.max_cq;
119 device_attr->max_cqe = attr_ex.orig_attr.max_cqe;
120 device_attr->max_mr = attr_ex.orig_attr.max_mr;
121 device_attr->max_pd = attr_ex.orig_attr.max_pd;
122 device_attr->max_qp = attr_ex.orig_attr.max_qp;
123 device_attr->max_srq = attr_ex.orig_attr.max_srq;
124 device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
125 device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
126 device_attr->max_rwq_indirection_table_size =
127 attr_ex.rss_caps.max_rwq_indirection_table_size;
128 device_attr->max_tso = attr_ex.tso_caps.max_tso;
129 device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
131 struct mlx5dv_context dv_attr = { .comp_mask = 0 };
132 err = mlx5_glue->dv_query_device(ctx, &dv_attr);
136 device_attr->flags = dv_attr.flags;
137 device_attr->comp_mask = dv_attr.comp_mask;
138 #ifdef HAVE_IBV_MLX5_MOD_SWP
139 device_attr->sw_parsing_offloads =
140 dv_attr.sw_parsing_caps.sw_parsing_offloads;
142 device_attr->min_single_stride_log_num_of_bytes =
143 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
144 device_attr->max_single_stride_log_num_of_bytes =
145 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
146 device_attr->min_single_wqe_log_num_of_strides =
147 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
148 device_attr->max_single_wqe_log_num_of_strides =
149 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
150 device_attr->stride_supported_qpts =
151 dv_attr.striding_rq_caps.supported_qpts;
152 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
153 device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
160 * Verbs callback to allocate a memory. This function should allocate the space
161 * according to the size provided residing inside a huge page.
162 * Please note that all allocation must respect the alignment from libmlx5
163 * (i.e. currently rte_mem_page_size()).
166 * The size in bytes of the memory to allocate.
168 * A pointer to the callback data.
171 * Allocated buffer, NULL otherwise and rte_errno is set.
174 mlx5_alloc_verbs_buf(size_t size, void *data)
176 struct mlx5_dev_ctx_shared *sh = data;
178 size_t alignment = rte_mem_page_size();
179 if (alignment == (size_t)-1) {
180 DRV_LOG(ERR, "Failed to get mem page size");
185 MLX5_ASSERT(data != NULL);
186 ret = mlx5_malloc(0, size, alignment, sh->numa_node);
193 * Verbs callback to free a memory.
196 * A pointer to the memory to free.
198 * A pointer to the callback data.
201 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
203 MLX5_ASSERT(data != NULL);
208 * Initialize DR related data within private structure.
209 * Routine checks the reference counter and does actual
210 * resources creation/initialization only if counter is zero.
213 * Pointer to the private device data structure.
216 * Zero on success, positive error code otherwise.
219 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
221 struct mlx5_dev_ctx_shared *sh = priv->sh;
222 char s[MLX5_HLIST_NAMESIZE] __rte_unused;
225 MLX5_ASSERT(sh && sh->refcnt);
228 err = mlx5_alloc_table_hash_list(priv);
231 /* The resources below are only valid with DV support. */
232 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
233 /* Init port id action cache list. */
234 snprintf(s, sizeof(s), "%s_port_id_action_cache", sh->ibdev_name);
235 mlx5_cache_list_init(&sh->port_id_action_list, s, 0, sh,
236 flow_dv_port_id_create_cb,
237 flow_dv_port_id_match_cb,
238 flow_dv_port_id_remove_cb);
239 /* Init push vlan action cache list. */
240 snprintf(s, sizeof(s), "%s_push_vlan_action_cache", sh->ibdev_name);
241 mlx5_cache_list_init(&sh->push_vlan_action_list, s, 0, sh,
242 flow_dv_push_vlan_create_cb,
243 flow_dv_push_vlan_match_cb,
244 flow_dv_push_vlan_remove_cb);
245 /* Init sample action cache list. */
246 snprintf(s, sizeof(s), "%s_sample_action_cache", sh->ibdev_name);
247 mlx5_cache_list_init(&sh->sample_action_list, s, 0, sh,
248 flow_dv_sample_create_cb,
249 flow_dv_sample_match_cb,
250 flow_dv_sample_remove_cb);
251 /* Init dest array action cache list. */
252 snprintf(s, sizeof(s), "%s_dest_array_cache", sh->ibdev_name);
253 mlx5_cache_list_init(&sh->dest_array_list, s, 0, sh,
254 flow_dv_dest_array_create_cb,
255 flow_dv_dest_array_match_cb,
256 flow_dv_dest_array_remove_cb);
257 /* Create tags hash list table. */
258 snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
259 sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
260 MLX5_HLIST_WRITE_MOST,
261 flow_dv_tag_create_cb,
262 flow_dv_tag_match_cb,
263 flow_dv_tag_remove_cb);
264 if (!sh->tag_table) {
265 DRV_LOG(ERR, "tags with hash creation failed.");
269 sh->tag_table->ctx = sh;
270 snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
271 sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
272 0, MLX5_HLIST_WRITE_MOST |
273 MLX5_HLIST_DIRECT_KEY,
274 flow_dv_modify_create_cb,
275 flow_dv_modify_match_cb,
276 flow_dv_modify_remove_cb);
277 if (!sh->modify_cmds) {
278 DRV_LOG(ERR, "hdr modify hash creation failed");
282 sh->modify_cmds->ctx = sh;
283 snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
284 sh->encaps_decaps = mlx5_hlist_create(s,
285 MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
286 0, MLX5_HLIST_DIRECT_KEY |
287 MLX5_HLIST_WRITE_MOST,
288 flow_dv_encap_decap_create_cb,
289 flow_dv_encap_decap_match_cb,
290 flow_dv_encap_decap_remove_cb);
291 if (!sh->encaps_decaps) {
292 DRV_LOG(ERR, "encap decap hash creation failed");
296 sh->encaps_decaps->ctx = sh;
298 #ifdef HAVE_MLX5DV_DR
301 /* Reference counter is zero, we should initialize structures. */
302 domain = mlx5_glue->dr_create_domain(sh->ctx,
303 MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
305 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
309 sh->rx_domain = domain;
310 domain = mlx5_glue->dr_create_domain(sh->ctx,
311 MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
313 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
317 sh->tx_domain = domain;
318 #ifdef HAVE_MLX5DV_DR_ESWITCH
319 if (priv->config.dv_esw_en) {
320 domain = mlx5_glue->dr_create_domain
321 (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
323 DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
327 sh->fdb_domain = domain;
330 * The drop action is just some dummy placeholder in rdma-core. It
331 * does not belong to domains and has no any attributes, and, can be
332 * shared by the entire device.
334 sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
335 if (!sh->dr_drop_action) {
336 DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
342 err = mlx5_alloc_tunnel_hub(sh);
344 DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
347 if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
348 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
349 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
351 mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
353 sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
354 #endif /* HAVE_MLX5DV_DR */
355 sh->default_miss_action =
356 mlx5_glue->dr_create_flow_action_default_miss();
357 if (!sh->default_miss_action)
358 DRV_LOG(WARNING, "Default miss action is not supported.");
361 /* Rollback the created objects. */
363 mlx5_glue->dr_destroy_domain(sh->rx_domain);
364 sh->rx_domain = NULL;
367 mlx5_glue->dr_destroy_domain(sh->tx_domain);
368 sh->tx_domain = NULL;
370 if (sh->fdb_domain) {
371 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
372 sh->fdb_domain = NULL;
374 if (sh->dr_drop_action) {
375 mlx5_glue->destroy_flow_action(sh->dr_drop_action);
376 sh->dr_drop_action = NULL;
378 if (sh->pop_vlan_action) {
379 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
380 sh->pop_vlan_action = NULL;
382 if (sh->encaps_decaps) {
383 mlx5_hlist_destroy(sh->encaps_decaps);
384 sh->encaps_decaps = NULL;
386 if (sh->modify_cmds) {
387 mlx5_hlist_destroy(sh->modify_cmds);
388 sh->modify_cmds = NULL;
391 /* tags should be destroyed with flow before. */
392 mlx5_hlist_destroy(sh->tag_table);
393 sh->tag_table = NULL;
395 if (sh->tunnel_hub) {
396 mlx5_release_tunnel_hub(sh, priv->dev_port);
397 sh->tunnel_hub = NULL;
399 mlx5_free_table_hash_list(priv);
404 * Destroy DR related data within private structure.
407 * Pointer to the private device data structure.
410 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
412 struct mlx5_dev_ctx_shared *sh = priv->sh;
414 MLX5_ASSERT(sh && sh->refcnt);
417 #ifdef HAVE_MLX5DV_DR
419 mlx5_glue->dr_destroy_domain(sh->rx_domain);
420 sh->rx_domain = NULL;
423 mlx5_glue->dr_destroy_domain(sh->tx_domain);
424 sh->tx_domain = NULL;
426 #ifdef HAVE_MLX5DV_DR_ESWITCH
427 if (sh->fdb_domain) {
428 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
429 sh->fdb_domain = NULL;
431 if (sh->dr_drop_action) {
432 mlx5_glue->destroy_flow_action(sh->dr_drop_action);
433 sh->dr_drop_action = NULL;
436 if (sh->pop_vlan_action) {
437 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
438 sh->pop_vlan_action = NULL;
440 #endif /* HAVE_MLX5DV_DR */
441 if (sh->default_miss_action)
442 mlx5_glue->destroy_flow_action
443 (sh->default_miss_action);
444 if (sh->encaps_decaps) {
445 mlx5_hlist_destroy(sh->encaps_decaps);
446 sh->encaps_decaps = NULL;
448 if (sh->modify_cmds) {
449 mlx5_hlist_destroy(sh->modify_cmds);
450 sh->modify_cmds = NULL;
453 /* tags should be destroyed with flow before. */
454 mlx5_hlist_destroy(sh->tag_table);
455 sh->tag_table = NULL;
457 if (sh->tunnel_hub) {
458 mlx5_release_tunnel_hub(sh, priv->dev_port);
459 sh->tunnel_hub = NULL;
461 mlx5_cache_list_destroy(&sh->port_id_action_list);
462 mlx5_cache_list_destroy(&sh->push_vlan_action_list);
463 mlx5_free_table_hash_list(priv);
467 * Initialize shared data between primary and secondary process.
469 * A memzone is reserved by primary process and secondary processes attach to
473 * 0 on success, a negative errno value otherwise and rte_errno is set.
476 mlx5_init_shared_data(void)
478 const struct rte_memzone *mz;
481 rte_spinlock_lock(&mlx5_shared_data_lock);
482 if (mlx5_shared_data == NULL) {
483 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
484 /* Allocate shared memory. */
485 mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
486 sizeof(*mlx5_shared_data),
490 "Cannot allocate mlx5 shared data");
494 mlx5_shared_data = mz->addr;
495 memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
496 rte_spinlock_init(&mlx5_shared_data->lock);
498 /* Lookup allocated shared memory. */
499 mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
502 "Cannot attach mlx5 shared data");
506 mlx5_shared_data = mz->addr;
507 memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
511 rte_spinlock_unlock(&mlx5_shared_data_lock);
516 * PMD global initialization.
518 * Independent from individual device, this function initializes global
519 * per-PMD data structures distinguishing primary and secondary processes.
520 * Hence, each initialization is called once per a process.
523 * 0 on success, a negative errno value otherwise and rte_errno is set.
528 struct mlx5_shared_data *sd;
529 struct mlx5_local_data *ld = &mlx5_local_data;
532 if (mlx5_init_shared_data())
534 sd = mlx5_shared_data;
536 rte_spinlock_lock(&sd->lock);
537 switch (rte_eal_process_type()) {
538 case RTE_PROC_PRIMARY:
541 LIST_INIT(&sd->mem_event_cb_list);
542 rte_rwlock_init(&sd->mem_event_rwlock);
543 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
544 mlx5_mr_mem_event_cb, NULL);
545 ret = mlx5_mp_init_primary(MLX5_MP_NAME,
546 mlx5_mp_os_primary_handle);
549 sd->init_done = true;
551 case RTE_PROC_SECONDARY:
554 ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
555 mlx5_mp_os_secondary_handle);
559 ld->init_done = true;
565 rte_spinlock_unlock(&sd->lock);
570 * Create the Tx queue DevX/Verbs object.
573 * Pointer to Ethernet device.
575 * Queue index in DPDK Tx queue array.
578 * 0 on success, a negative errno value otherwise and rte_errno is set.
581 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
583 struct mlx5_priv *priv = dev->data->dev_private;
584 struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
585 struct mlx5_txq_ctrl *txq_ctrl =
586 container_of(txq_data, struct mlx5_txq_ctrl, txq);
588 if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
589 return mlx5_txq_devx_obj_new(dev, idx);
590 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
591 if (!priv->config.dv_esw_en)
592 return mlx5_txq_devx_obj_new(dev, idx);
594 return mlx5_txq_ibv_obj_new(dev, idx);
598 * Release an Tx DevX/verbs queue object.
601 * DevX/Verbs Tx queue object.
604 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
606 if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
607 mlx5_txq_devx_obj_release(txq_obj);
610 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
611 if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
612 mlx5_txq_devx_obj_release(txq_obj);
616 mlx5_txq_ibv_obj_release(txq_obj);
620 * DV flow counter mode detect and config.
623 * Pointer to rte_eth_dev structure.
627 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
629 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
630 struct mlx5_priv *priv = dev->data->dev_private;
631 struct mlx5_dev_ctx_shared *sh = priv->sh;
634 #ifndef HAVE_IBV_DEVX_ASYNC
638 if (!priv->config.devx || !priv->config.dv_flow_en ||
639 !priv->config.hca_attr.flow_counters_dump ||
640 !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
641 (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
645 DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
646 "counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
647 priv->config.hca_attr.flow_counters_dump,
648 priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
649 /* Initialize fallback mode only on the port initializes sh. */
651 sh->cmng.counter_fallback = fallback;
652 else if (fallback != sh->cmng.counter_fallback)
653 DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
654 "with others:%d.", PORT_ID(priv), fallback);
659 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
661 struct mlx5_priv *priv = dev->data->dev_private;
662 void *ctx = priv->sh->ctx;
664 priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
665 if (!priv->q_counters) {
666 struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
669 DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
670 "by DevX - fall-back to use the kernel driver global "
671 "queue counter.", dev->data->port_id);
672 /* Create WQ by kernel and query its queue counter ID. */
674 wq = mlx5_glue->create_wq(ctx,
675 &(struct ibv_wq_init_attr){
676 .wq_type = IBV_WQT_RQ,
683 /* Counter is assigned only on RDY state. */
684 int ret = mlx5_glue->modify_wq(wq,
685 &(struct ibv_wq_attr){
686 .attr_mask = IBV_WQ_ATTR_STATE,
687 .wq_state = IBV_WQS_RDY,
691 mlx5_devx_cmd_wq_query(wq,
692 &priv->counter_set_id);
693 claim_zero(mlx5_glue->destroy_wq(wq));
695 claim_zero(mlx5_glue->destroy_cq(cq));
698 priv->counter_set_id = priv->q_counters->id;
700 if (priv->counter_set_id == 0)
701 DRV_LOG(INFO, "Part of the port %d statistics will not be "
702 "available.", dev->data->port_id);
706 * Check if representor spawn info match devargs.
709 * Verbs device parameters (name, port, switch_info) to spawn.
711 * Device devargs to probe.
717 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
718 struct rte_eth_devargs *eth_da)
720 struct mlx5_switch_info *switch_info = &spawn->info;
723 uint16_t repr_id = mlx5_representor_id_encode(switch_info,
726 switch (eth_da->type) {
727 case RTE_ETH_REPRESENTOR_SF:
728 if (!(spawn->info.port_name == -1 &&
729 switch_info->name_type ==
730 MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
731 switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
736 case RTE_ETH_REPRESENTOR_VF:
737 /* Allows HPF representor index -1 as exception. */
738 if (!(spawn->info.port_name == -1 &&
739 switch_info->name_type ==
740 MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
741 switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
746 case RTE_ETH_REPRESENTOR_NONE:
751 DRV_LOG(ERR, "unsupported representor type");
754 /* Check representor ID: */
755 for (p = 0; p < eth_da->nb_ports; ++p) {
756 if (spawn->pf_bond < 0) {
757 /* For non-LAG mode, allow and ignore pf. */
758 switch_info->pf_num = eth_da->ports[p];
759 repr_id = mlx5_representor_id_encode(switch_info,
762 for (f = 0; f < eth_da->nb_representor_ports; ++f) {
763 id = MLX5_REPRESENTOR_ID
764 (eth_da->ports[p], eth_da->type,
765 eth_da->representor_ports[f]);
776 * Spawn an Ethernet device from Verbs information.
779 * Backing DPDK device.
781 * Verbs device parameters (name, port, switch_info) to spawn.
783 * Device configuration parameters.
788 * A valid Ethernet device object on success, NULL otherwise and rte_errno
789 * is set. The following errors are defined:
791 * EBUSY: device is not supposed to be spawned.
792 * EEXIST: device is already spawned
794 static struct rte_eth_dev *
795 mlx5_dev_spawn(struct rte_device *dpdk_dev,
796 struct mlx5_dev_spawn_data *spawn,
797 struct mlx5_dev_config *config,
798 struct rte_eth_devargs *eth_da)
800 const struct mlx5_switch_info *switch_info = &spawn->info;
801 struct mlx5_dev_ctx_shared *sh = NULL;
802 struct ibv_port_attr port_attr;
803 struct mlx5dv_context dv_attr = { .comp_mask = 0 };
804 struct rte_eth_dev *eth_dev = NULL;
805 struct mlx5_priv *priv = NULL;
807 unsigned int hw_padding = 0;
809 unsigned int tunnel_en = 0;
810 unsigned int mpls_en = 0;
811 unsigned int swp = 0;
812 unsigned int mprq = 0;
813 unsigned int mprq_min_stride_size_n = 0;
814 unsigned int mprq_max_stride_size_n = 0;
815 unsigned int mprq_min_stride_num_n = 0;
816 unsigned int mprq_max_stride_num_n = 0;
817 struct rte_ether_addr mac;
818 char name[RTE_ETH_NAME_MAX_LEN];
819 int own_domain_id = 0;
821 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
822 struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
825 /* Determine if this port representor is supposed to be spawned. */
826 if (switch_info->representor && dpdk_dev->devargs &&
827 !mlx5_representor_match(spawn, eth_da))
829 /* Build device name. */
830 if (spawn->pf_bond < 0) {
832 if (!switch_info->representor)
833 strlcpy(name, dpdk_dev->name, sizeof(name));
835 err = snprintf(name, sizeof(name), "%s_representor_%s%u",
837 switch_info->name_type ==
838 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
839 switch_info->port_name);
841 /* Bonding device. */
842 if (!switch_info->representor) {
843 err = snprintf(name, sizeof(name), "%s_%s",
845 mlx5_os_get_dev_device_name(spawn->phys_dev));
847 err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
849 mlx5_os_get_dev_device_name(spawn->phys_dev),
850 switch_info->ctrl_num,
852 switch_info->name_type ==
853 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
854 switch_info->port_name);
857 if (err >= (int)sizeof(name))
858 DRV_LOG(WARNING, "device name overflow %s", name);
859 /* check if the device is already spawned */
860 if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
864 DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
865 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
866 struct mlx5_mp_id mp_id;
868 eth_dev = rte_eth_dev_attach_secondary(name);
869 if (eth_dev == NULL) {
870 DRV_LOG(ERR, "can not attach rte ethdev");
874 eth_dev->device = dpdk_dev;
875 eth_dev->dev_ops = &mlx5_dev_sec_ops;
876 eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
877 eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
878 err = mlx5_proc_priv_init(eth_dev);
881 mp_id.port_id = eth_dev->data->port_id;
882 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
883 /* Receive command fd from primary process */
884 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
887 /* Remap UAR for Tx queues. */
888 err = mlx5_tx_uar_init_secondary(eth_dev, err);
892 * Ethdev pointer is still required as input since
893 * the primary device is not accessible from the
896 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
897 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
900 mlx5_dev_close(eth_dev);
904 * Some parameters ("tx_db_nc" in particularly) are needed in
905 * advance to create dv/verbs device context. We proceed the
906 * devargs here to get ones, and later proceed devargs again
907 * to override some hardware settings.
909 err = mlx5_args(config, dpdk_dev->devargs);
912 DRV_LOG(ERR, "failed to process device arguments: %s",
913 strerror(rte_errno));
916 if (config->dv_miss_info) {
917 if (switch_info->master || switch_info->representor)
918 config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
920 mlx5_malloc_mem_select(config->sys_mem_en);
921 sh = mlx5_alloc_shared_dev_ctx(spawn, config);
924 config->devx = sh->devx;
925 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
926 config->dest_tir = 1;
928 #ifdef HAVE_IBV_MLX5_MOD_SWP
929 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
932 * Multi-packet send is supported by ConnectX-4 Lx PF as well
933 * as all ConnectX-5 devices.
935 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
936 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
938 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
939 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
941 mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
942 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
943 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
944 DRV_LOG(DEBUG, "enhanced MPW is supported");
945 mps = MLX5_MPW_ENHANCED;
947 DRV_LOG(DEBUG, "MPW is supported");
951 DRV_LOG(DEBUG, "MPW isn't supported");
952 mps = MLX5_MPW_DISABLED;
954 #ifdef HAVE_IBV_MLX5_MOD_SWP
955 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
956 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
957 DRV_LOG(DEBUG, "SWP support: %u", swp);
960 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
961 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
962 struct mlx5dv_striding_rq_caps mprq_caps =
963 dv_attr.striding_rq_caps;
965 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
966 mprq_caps.min_single_stride_log_num_of_bytes);
967 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
968 mprq_caps.max_single_stride_log_num_of_bytes);
969 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
970 mprq_caps.min_single_wqe_log_num_of_strides);
971 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
972 mprq_caps.max_single_wqe_log_num_of_strides);
973 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
974 mprq_caps.supported_qpts);
975 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
977 mprq_min_stride_size_n =
978 mprq_caps.min_single_stride_log_num_of_bytes;
979 mprq_max_stride_size_n =
980 mprq_caps.max_single_stride_log_num_of_bytes;
981 mprq_min_stride_num_n =
982 mprq_caps.min_single_wqe_log_num_of_strides;
983 mprq_max_stride_num_n =
984 mprq_caps.max_single_wqe_log_num_of_strides;
987 /* Rx CQE compression is enabled by default. */
988 config->cqe_comp = 1;
989 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
990 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
991 tunnel_en = ((dv_attr.tunnel_offloads_caps &
992 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
993 (dv_attr.tunnel_offloads_caps &
994 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
995 (dv_attr.tunnel_offloads_caps &
996 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
998 DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
999 tunnel_en ? "" : "not ");
1002 "tunnel offloading disabled due to old OFED/rdma-core version");
1004 config->tunnel_en = tunnel_en;
1005 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1006 mpls_en = ((dv_attr.tunnel_offloads_caps &
1007 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1008 (dv_attr.tunnel_offloads_caps &
1009 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1010 DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1011 mpls_en ? "" : "not ");
1013 DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1014 " old OFED/rdma-core version or firmware configuration");
1016 config->mpls_en = mpls_en;
1017 /* Check port status. */
1018 err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1020 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1023 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1024 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1028 if (port_attr.state != IBV_PORT_ACTIVE)
1029 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1030 mlx5_glue->port_state_str(port_attr.state),
1032 /* Allocate private eth device data. */
1033 priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1035 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1037 DRV_LOG(ERR, "priv allocation failure");
1042 priv->dev_port = spawn->phys_port;
1043 priv->pci_dev = spawn->pci_dev;
1044 priv->mtu = RTE_ETHER_MTU;
1045 /* Some internal functions rely on Netlink sockets, open them now. */
1046 priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1047 priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1048 priv->representor = !!switch_info->representor;
1049 priv->master = !!switch_info->master;
1050 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1051 priv->vport_meta_tag = 0;
1052 priv->vport_meta_mask = 0;
1053 priv->pf_bond = spawn->pf_bond;
1054 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1056 * The DevX port query API is implemented. E-Switch may use
1057 * either vport or reg_c[0] metadata register to match on
1058 * vport index. The engaged part of metadata register is
1061 if (switch_info->representor || switch_info->master) {
1062 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
1063 MLX5DV_DEVX_PORT_MATCH_REG_C_0;
1064 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
1068 "can't query devx port %d on device %s",
1070 mlx5_os_get_dev_device_name(spawn->phys_dev));
1071 devx_port.comp_mask = 0;
1074 if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
1075 priv->vport_meta_tag = devx_port.reg_c_0.value;
1076 priv->vport_meta_mask = devx_port.reg_c_0.mask;
1077 if (!priv->vport_meta_mask) {
1078 DRV_LOG(ERR, "vport zero mask for port %d"
1079 " on bonding device %s",
1081 mlx5_os_get_dev_device_name
1086 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1087 DRV_LOG(ERR, "invalid vport tag for port %d"
1088 " on bonding device %s",
1090 mlx5_os_get_dev_device_name
1096 if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
1097 priv->vport_id = devx_port.vport_num;
1098 } else if (spawn->pf_bond >= 0) {
1099 DRV_LOG(ERR, "can't deduce vport index for port %d"
1100 " on bonding device %s",
1102 mlx5_os_get_dev_device_name(spawn->phys_dev));
1106 /* Suppose vport index in compatible way. */
1107 priv->vport_id = switch_info->representor ?
1108 switch_info->port_name + 1 : -1;
1112 * Kernel/rdma_core support single E-Switch per PF configurations
1113 * only and vport_id field contains the vport index for
1114 * associated VF, which is deduced from representor port name.
1115 * For example, let's have the IB device port 10, it has
1116 * attached network device eth0, which has port name attribute
1117 * pf0vf2, we can deduce the VF number as 2, and set vport index
1118 * as 3 (2+1). This assigning schema should be changed if the
1119 * multiple E-Switch instances per PF configurations or/and PCI
1120 * subfunctions are added.
1122 priv->vport_id = switch_info->representor ?
1123 switch_info->port_name + 1 : -1;
1125 priv->representor_id = mlx5_representor_id_encode(switch_info,
1128 * Look for sibling devices in order to reuse their switch domain
1129 * if any, otherwise allocate one.
1131 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1132 const struct mlx5_priv *opriv =
1133 rte_eth_devices[port_id].data->dev_private;
1136 opriv->sh != priv->sh ||
1138 RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1140 priv->domain_id = opriv->domain_id;
1143 if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1144 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1147 DRV_LOG(ERR, "unable to allocate switch domain: %s",
1148 strerror(rte_errno));
1153 /* Override some values set by hardware configuration. */
1154 mlx5_args(config, dpdk_dev->devargs);
1155 err = mlx5_dev_check_sibling_config(priv, config);
1158 config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1159 IBV_DEVICE_RAW_IP_CSUM);
1160 DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1161 (config->hw_csum ? "" : "not "));
1162 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1163 !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1164 DRV_LOG(DEBUG, "counters are not supported");
1166 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1167 if (config->dv_flow_en) {
1168 DRV_LOG(WARNING, "DV flow is not supported");
1169 config->dv_flow_en = 0;
1172 config->ind_table_max_size =
1173 sh->device_attr.max_rwq_indirection_table_size;
1175 * Remove this check once DPDK supports larger/variable
1176 * indirection tables.
1178 if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1179 config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1180 DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1181 config->ind_table_max_size);
1182 config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1183 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1184 DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1185 (config->hw_vlan_strip ? "" : "not "));
1186 config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1187 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1188 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1189 hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1190 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1191 hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1192 IBV_DEVICE_PCI_WRITE_END_PADDING);
1194 if (config->hw_padding && !hw_padding) {
1195 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1196 config->hw_padding = 0;
1197 } else if (config->hw_padding) {
1198 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1200 config->tso = (sh->device_attr.max_tso > 0 &&
1201 (sh->device_attr.tso_supported_qpts &
1202 (1 << IBV_QPT_RAW_PACKET)));
1204 config->tso_max_payload_sz = sh->device_attr.max_tso;
1206 * MPW is disabled by default, while the Enhanced MPW is enabled
1209 if (config->mps == MLX5_ARG_UNSET)
1210 config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1213 config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1214 DRV_LOG(INFO, "%sMPS is %s",
1215 config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1216 config->mps == MLX5_MPW ? "legacy " : "",
1217 config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1219 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1224 /* Check relax ordering support. */
1225 if (!haswell_broadwell_cpu) {
1226 sh->cmng.relaxed_ordering_write =
1227 config->hca_attr.relaxed_ordering_write;
1228 sh->cmng.relaxed_ordering_read =
1229 config->hca_attr.relaxed_ordering_read;
1231 sh->cmng.relaxed_ordering_read = 0;
1232 sh->cmng.relaxed_ordering_write = 0;
1234 sh->rq_ts_format = config->hca_attr.rq_ts_format;
1235 sh->sq_ts_format = config->hca_attr.sq_ts_format;
1236 sh->qp_ts_format = config->hca_attr.qp_ts_format;
1237 /* Check for LRO support. */
1238 if (config->dest_tir && config->hca_attr.lro_cap &&
1239 config->dv_flow_en) {
1240 /* TBD check tunnel lro caps. */
1241 config->lro.supported = config->hca_attr.lro_cap;
1242 DRV_LOG(DEBUG, "Device supports LRO");
1244 * If LRO timeout is not configured by application,
1245 * use the minimal supported value.
1247 if (!config->lro.timeout)
1248 config->lro.timeout =
1249 config->hca_attr.lro_timer_supported_periods[0];
1250 DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1251 config->lro.timeout);
1252 DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1253 "required for coalescing is %d bytes",
1254 config->hca_attr.lro_min_mss_size);
1256 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1257 if (config->hca_attr.qos.sup &&
1258 config->hca_attr.qos.flow_meter_old &&
1259 config->dv_flow_en) {
1260 uint8_t reg_c_mask =
1261 config->hca_attr.qos.flow_meter_reg_c_ids;
1263 * Meter needs two REG_C's for color match and pre-sfx
1264 * flow match. Here get the REG_C for color match.
1265 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1268 if (__builtin_popcount(reg_c_mask) < 1) {
1270 DRV_LOG(WARNING, "No available register for"
1274 * The meter color register is used by the
1275 * flow-hit feature as well.
1276 * The flow-hit feature must use REG_C_3
1277 * Prefer REG_C_3 if it is available.
1279 if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1280 priv->mtr_color_reg = REG_C_3;
1282 priv->mtr_color_reg = ffs(reg_c_mask)
1285 priv->mtr_reg_share =
1286 config->hca_attr.qos.flow_meter;
1287 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1288 priv->mtr_color_reg);
1292 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1293 if (config->hca_attr.flow_hit_aso &&
1294 priv->mtr_color_reg == REG_C_3) {
1295 sh->flow_hit_aso_en = 1;
1296 err = mlx5_flow_aso_age_mng_init(sh);
1301 DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1303 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1304 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1305 if (config->hca_attr.log_max_ft_sampler_num > 0 &&
1306 config->dv_flow_en) {
1307 priv->sampler_en = 1;
1308 DRV_LOG(DEBUG, "Sampler enabled!");
1310 priv->sampler_en = 0;
1311 if (!config->hca_attr.log_max_ft_sampler_num)
1313 "No available register for sampler.");
1315 DRV_LOG(DEBUG, "DV flow is not supported!");
1319 if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1320 !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1321 DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1322 config->cqe_comp = 0;
1324 if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1325 (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1326 DRV_LOG(WARNING, "Flow Tag CQE compression"
1327 " format isn't supported.");
1328 config->cqe_comp = 0;
1330 if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1331 (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1332 DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1333 " format isn't supported.");
1334 config->cqe_comp = 0;
1336 DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1337 config->cqe_comp ? "" : "not ");
1338 if (config->tx_pp) {
1339 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1340 config->hca_attr.dev_freq_khz);
1341 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1342 config->hca_attr.qos.packet_pacing ? "" : "not ");
1343 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1344 config->hca_attr.cross_channel ? "" : "not ");
1345 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1346 config->hca_attr.wqe_index_ignore ? "" : "not ");
1347 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1348 config->hca_attr.non_wire_sq ? "" : "not ");
1349 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1350 config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1351 config->hca_attr.log_max_static_sq_wq);
1352 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1353 config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1354 if (!config->devx) {
1355 DRV_LOG(ERR, "DevX is required for packet pacing");
1359 if (!config->hca_attr.qos.packet_pacing) {
1360 DRV_LOG(ERR, "Packet pacing is not supported");
1364 if (!config->hca_attr.cross_channel) {
1365 DRV_LOG(ERR, "Cross channel operations are"
1366 " required for packet pacing");
1370 if (!config->hca_attr.wqe_index_ignore) {
1371 DRV_LOG(ERR, "WQE index ignore feature is"
1372 " required for packet pacing");
1376 if (!config->hca_attr.non_wire_sq) {
1377 DRV_LOG(ERR, "Non-wire SQ feature is"
1378 " required for packet pacing");
1382 if (!config->hca_attr.log_max_static_sq_wq) {
1383 DRV_LOG(ERR, "Static WQE SQ feature is"
1384 " required for packet pacing");
1388 if (!config->hca_attr.qos.wqe_rate_pp) {
1389 DRV_LOG(ERR, "WQE rate mode is required"
1390 " for packet pacing");
1394 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1395 DRV_LOG(ERR, "DevX does not provide UAR offset,"
1396 " can't create queues for packet pacing");
1402 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1404 err = config->hca_attr.access_register_user ?
1405 mlx5_devx_cmd_register_read
1406 (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1407 reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1411 /* MTUTC register is read successfully. */
1412 ts_mode = MLX5_GET(register_mtutc, reg,
1414 if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1415 config->rt_timestamp = 1;
1417 /* Kernel does not support register reading. */
1418 if (config->hca_attr.dev_freq_khz ==
1419 (NS_PER_S / MS_PER_S))
1420 config->rt_timestamp = 1;
1424 * If HW has bug working with tunnel packet decapsulation and
1425 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1426 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1428 if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1429 config->hw_fcs_strip = 0;
1430 DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1431 (config->hw_fcs_strip ? "" : "not "));
1432 if (config->mprq.enabled && mprq) {
1433 if (config->mprq.stride_num_n &&
1434 (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1435 config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1436 config->mprq.stride_num_n =
1437 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1438 mprq_min_stride_num_n),
1439 mprq_max_stride_num_n);
1441 "the number of strides"
1442 " for Multi-Packet RQ is out of range,"
1443 " setting default value (%u)",
1444 1 << config->mprq.stride_num_n);
1446 if (config->mprq.stride_size_n &&
1447 (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1448 config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1449 config->mprq.stride_size_n =
1450 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1451 mprq_min_stride_size_n),
1452 mprq_max_stride_size_n);
1454 "the size of a stride"
1455 " for Multi-Packet RQ is out of range,"
1456 " setting default value (%u)",
1457 1 << config->mprq.stride_size_n);
1459 config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1460 config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1461 } else if (config->mprq.enabled && !mprq) {
1462 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1463 config->mprq.enabled = 0;
1465 if (config->max_dump_files_num == 0)
1466 config->max_dump_files_num = 128;
1467 eth_dev = rte_eth_dev_allocate(name);
1468 if (eth_dev == NULL) {
1469 DRV_LOG(ERR, "can not allocate rte ethdev");
1473 if (priv->representor) {
1474 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1475 eth_dev->data->representor_id = priv->representor_id;
1477 priv->mp_id.port_id = eth_dev->data->port_id;
1478 strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1480 * Store associated network device interface index. This index
1481 * is permanent throughout the lifetime of device. So, we may store
1482 * the ifindex here and use the cached value further.
1484 MLX5_ASSERT(spawn->ifindex);
1485 priv->if_index = spawn->ifindex;
1486 eth_dev->data->dev_private = priv;
1487 priv->dev_data = eth_dev->data;
1488 eth_dev->data->mac_addrs = priv->mac;
1489 eth_dev->device = dpdk_dev;
1490 eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1491 /* Configure the first MAC address by default. */
1492 if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1494 "port %u cannot get MAC address, is mlx5_en"
1495 " loaded? (errno: %s)",
1496 eth_dev->data->port_id, strerror(rte_errno));
1501 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1502 eth_dev->data->port_id,
1503 mac.addr_bytes[0], mac.addr_bytes[1],
1504 mac.addr_bytes[2], mac.addr_bytes[3],
1505 mac.addr_bytes[4], mac.addr_bytes[5]);
1506 #ifdef RTE_LIBRTE_MLX5_DEBUG
1508 char ifname[MLX5_NAMESIZE];
1510 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1511 DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1512 eth_dev->data->port_id, ifname);
1514 DRV_LOG(DEBUG, "port %u ifname is unknown",
1515 eth_dev->data->port_id);
1518 /* Get actual MTU if possible. */
1519 err = mlx5_get_mtu(eth_dev, &priv->mtu);
1524 DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1526 /* Initialize burst functions to prevent crashes before link-up. */
1527 eth_dev->rx_pkt_burst = removed_rx_burst;
1528 eth_dev->tx_pkt_burst = removed_tx_burst;
1529 eth_dev->dev_ops = &mlx5_dev_ops;
1530 eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1531 eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1532 eth_dev->rx_queue_count = mlx5_rx_queue_count;
1533 /* Register MAC address. */
1534 claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1535 if (config->vf && config->vf_nl_en)
1536 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1537 mlx5_ifindex(eth_dev),
1538 eth_dev->data->mac_addrs,
1539 MLX5_MAX_MAC_ADDRESSES);
1541 priv->ctrl_flows = 0;
1542 rte_spinlock_init(&priv->flow_list_lock);
1543 TAILQ_INIT(&priv->flow_meters);
1544 TAILQ_INIT(&priv->flow_meter_profiles);
1545 /* Hint libmlx5 to use PMD allocator for data plane resources */
1546 mlx5_glue->dv_set_context_attr(sh->ctx,
1547 MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1548 (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1549 .alloc = &mlx5_alloc_verbs_buf,
1550 .free = &mlx5_free_verbs_buf,
1553 /* Bring Ethernet device up. */
1554 DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1555 eth_dev->data->port_id);
1556 mlx5_set_link_up(eth_dev);
1558 * Even though the interrupt handler is not installed yet,
1559 * interrupts will still trigger on the async_fd from
1560 * Verbs context returned by ibv_open_device().
1562 mlx5_link_update(eth_dev, 0);
1563 #ifdef HAVE_MLX5DV_DR_ESWITCH
1564 if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1565 (switch_info->representor || switch_info->master)))
1566 config->dv_esw_en = 0;
1568 config->dv_esw_en = 0;
1570 /* Detect minimal data bytes to inline. */
1571 mlx5_set_min_inline(spawn, config);
1572 /* Store device configuration on private structure. */
1573 priv->config = *config;
1574 /* Create context for virtual machine VLAN workaround. */
1575 priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1576 if (config->dv_flow_en) {
1577 err = mlx5_alloc_shared_dr(priv);
1581 if (config->devx && config->dv_flow_en && config->dest_tir) {
1582 priv->obj_ops = devx_obj_ops;
1583 priv->obj_ops.drop_action_create =
1584 ibv_obj_ops.drop_action_create;
1585 priv->obj_ops.drop_action_destroy =
1586 ibv_obj_ops.drop_action_destroy;
1587 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1588 priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1590 if (config->dv_esw_en)
1591 priv->obj_ops.txq_obj_modify =
1592 ibv_obj_ops.txq_obj_modify;
1594 /* Use specific wrappers for Tx object. */
1595 priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1596 priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1597 mlx5_queue_counter_id_prepare(eth_dev);
1600 priv->obj_ops = ibv_obj_ops;
1602 priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1603 if (!priv->drop_queue.hrxq)
1605 /* Supported Verbs flow priority number detection. */
1606 err = mlx5_flow_discover_priorities(eth_dev);
1611 priv->config.flow_prio = err;
1612 if (!priv->config.dv_esw_en &&
1613 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1614 DRV_LOG(WARNING, "metadata mode %u is not supported "
1615 "(no E-Switch)", priv->config.dv_xmeta_en);
1616 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1618 mlx5_set_metadata_mask(eth_dev);
1619 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1620 !priv->sh->dv_regc0_mask) {
1621 DRV_LOG(ERR, "metadata mode %u is not supported "
1622 "(no metadata reg_c[0] is available)",
1623 priv->config.dv_xmeta_en);
1627 mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1628 mlx5_hrxq_create_cb,
1630 mlx5_hrxq_remove_cb);
1631 /* Query availability of metadata reg_c's. */
1632 err = mlx5_flow_discover_mreg_c(eth_dev);
1637 if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1639 "port %u extensive metadata register is not supported",
1640 eth_dev->data->port_id);
1641 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1642 DRV_LOG(ERR, "metadata mode %u is not supported "
1643 "(no metadata registers available)",
1644 priv->config.dv_xmeta_en);
1649 if (priv->config.dv_flow_en &&
1650 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1651 mlx5_flow_ext_mreg_supported(eth_dev) &&
1652 priv->sh->dv_regc0_mask) {
1653 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1654 MLX5_FLOW_MREG_HTABLE_SZ,
1656 flow_dv_mreg_create_cb,
1657 flow_dv_mreg_match_cb,
1658 flow_dv_mreg_remove_cb);
1659 if (!priv->mreg_cp_tbl) {
1663 priv->mreg_cp_tbl->ctx = eth_dev;
1665 rte_spinlock_init(&priv->shared_act_sl);
1666 mlx5_flow_counter_mode_config(eth_dev);
1667 if (priv->config.dv_flow_en)
1668 eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1672 if (priv->mreg_cp_tbl)
1673 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1675 mlx5_os_free_shared_dr(priv);
1676 if (priv->nl_socket_route >= 0)
1677 close(priv->nl_socket_route);
1678 if (priv->nl_socket_rdma >= 0)
1679 close(priv->nl_socket_rdma);
1680 if (priv->vmwa_context)
1681 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1682 if (eth_dev && priv->drop_queue.hrxq)
1683 mlx5_drop_action_destroy(eth_dev);
1685 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1686 mlx5_cache_list_destroy(&priv->hrxqs);
1688 if (eth_dev != NULL)
1689 eth_dev->data->dev_private = NULL;
1691 if (eth_dev != NULL) {
1692 /* mac_addrs must not be freed alone because part of
1695 eth_dev->data->mac_addrs = NULL;
1696 rte_eth_dev_release_port(eth_dev);
1699 mlx5_free_shared_dev_ctx(sh);
1700 MLX5_ASSERT(err > 0);
1706 * Comparison callback to sort device data.
1708 * This is meant to be used with qsort().
1711 * Pointer to pointer to first data object.
1713 * Pointer to pointer to second data object.
1716 * 0 if both objects are equal, less than 0 if the first argument is less
1717 * than the second, greater than 0 otherwise.
1720 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1722 const struct mlx5_switch_info *si_a =
1723 &((const struct mlx5_dev_spawn_data *)a)->info;
1724 const struct mlx5_switch_info *si_b =
1725 &((const struct mlx5_dev_spawn_data *)b)->info;
1728 /* Master device first. */
1729 ret = si_b->master - si_a->master;
1732 /* Then representor devices. */
1733 ret = si_b->representor - si_a->representor;
1736 /* Unidentified devices come last in no specific order. */
1737 if (!si_a->representor)
1739 /* Order representors by name. */
1740 return si_a->port_name - si_b->port_name;
1744 * Match PCI information for possible slaves of bonding device.
1746 * @param[in] ibv_dev
1747 * Pointer to Infiniband device structure.
1748 * @param[in] pci_dev
1749 * Pointer to primary PCI address structure to match.
1750 * @param[in] nl_rdma
1751 * Netlink RDMA group socket handle.
1753 * Rerepsentor owner PF index.
1754 * @param[out] bond_info
1755 * Pointer to bonding information.
1758 * negative value if no bonding device found, otherwise
1759 * positive index of slave PF in bonding.
1762 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1763 const struct rte_pci_addr *pci_dev,
1764 int nl_rdma, uint16_t owner,
1765 struct mlx5_bond_info *bond_info)
1767 char ifname[IF_NAMESIZE + 1];
1768 unsigned int ifindex;
1770 FILE *bond_file = NULL, *file;
1775 * Try to get master device name. If something goes
1776 * wrong suppose the lack of kernel support and no
1779 memset(bond_info, 0, sizeof(*bond_info));
1782 if (!strstr(ibv_dev->name, "bond"))
1784 np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1788 * The Master device might not be on the predefined
1789 * port (not on port index 1, it is not garanted),
1790 * we have to scan all Infiniband device port and
1793 for (i = 1; i <= np; ++i) {
1794 /* Check whether Infiniband port is populated. */
1795 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1798 if (!if_indextoname(ifindex, ifname))
1800 /* Try to read bonding slave names from sysfs. */
1802 "/sys/class/net/%s/master/bonding/slaves", ifname);
1803 bond_file = fopen(slaves, "r");
1809 /* Use safe format to check maximal buffer length. */
1810 MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1811 while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1812 char tmp_str[IF_NAMESIZE + 32];
1813 struct rte_pci_addr pci_addr;
1814 struct mlx5_switch_info info;
1816 /* Process slave interface names in the loop. */
1817 snprintf(tmp_str, sizeof(tmp_str),
1818 "/sys/class/net/%s", ifname);
1819 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1820 DRV_LOG(WARNING, "can not get PCI address"
1821 " for netdev \"%s\"", ifname);
1824 /* Slave interface PCI address match found. */
1825 snprintf(tmp_str, sizeof(tmp_str),
1826 "/sys/class/net/%s/phys_port_name", ifname);
1827 file = fopen(tmp_str, "rb");
1830 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1831 if (fscanf(file, "%32s", tmp_str) == 1)
1832 mlx5_translate_port_name(tmp_str, &info);
1834 /* Only process PF ports. */
1835 if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1836 info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1838 /* Check max bonding member. */
1839 if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1840 DRV_LOG(WARNING, "bonding index out of range, "
1841 "please increase MLX5_BOND_MAX_PORTS: %s",
1845 /* Match PCI address. */
1846 if (pci_dev->domain == pci_addr.domain &&
1847 pci_dev->bus == pci_addr.bus &&
1848 pci_dev->devid == pci_addr.devid &&
1849 pci_dev->function + owner == pci_addr.function)
1850 pf = info.port_name;
1852 snprintf(tmp_str, sizeof(tmp_str),
1853 "/sys/class/net/%s/ifindex", ifname);
1854 file = fopen(tmp_str, "rb");
1857 ret = fscanf(file, "%u", &ifindex);
1861 /* Save bonding info. */
1862 strncpy(bond_info->ports[info.port_name].ifname, ifname,
1863 sizeof(bond_info->ports[0].ifname));
1864 bond_info->ports[info.port_name].pci_addr = pci_addr;
1865 bond_info->ports[info.port_name].ifindex = ifindex;
1866 bond_info->n_port++;
1869 /* Get bond interface info */
1870 ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1873 DRV_LOG(ERR, "unable to get bond info: %s",
1874 strerror(rte_errno));
1876 DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1877 ifindex, bond_info->ifindex, bond_info->ifname);
1883 * Register a PCI device within bonding.
1885 * This function spawns Ethernet devices out of a given PCI device and
1886 * bonding owner PF index.
1888 * @param[in] pci_dev
1889 * PCI device information.
1890 * @param[in] req_eth_da
1891 * Requested ethdev device argument.
1892 * @param[in] owner_id
1893 * Requested owner PF port ID within bonding device, default to 0.
1896 * 0 on success, a negative errno value otherwise and rte_errno is set.
1899 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
1900 struct rte_eth_devargs *req_eth_da,
1903 struct ibv_device **ibv_list;
1905 * Number of found IB Devices matching with requested PCI BDF.
1906 * nd != 1 means there are multiple IB devices over the same
1907 * PCI device and we have representors and master.
1909 unsigned int nd = 0;
1911 * Number of found IB device Ports. nd = 1 and np = 1..n means
1912 * we have the single multiport IB device, and there may be
1913 * representors attached to some of found ports.
1915 unsigned int np = 0;
1917 * Number of DPDK ethernet devices to Spawn - either over
1918 * multiple IB devices or multiple ports of single IB device.
1919 * Actually this is the number of iterations to spawn.
1921 unsigned int ns = 0;
1924 * < 0 - no bonding device (single one)
1925 * >= 0 - bonding device (value is slave PF index)
1928 struct mlx5_dev_spawn_data *list = NULL;
1929 struct mlx5_dev_config dev_config;
1930 unsigned int dev_config_vf;
1931 struct rte_eth_devargs eth_da = *req_eth_da;
1932 struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1933 struct mlx5_bond_info bond_info;
1936 if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1937 mlx5_pmd_socket_init();
1938 ret = mlx5_init_once();
1940 DRV_LOG(ERR, "unable to init PMD global data: %s",
1941 strerror(rte_errno));
1945 ibv_list = mlx5_glue->get_device_list(&ret);
1947 rte_errno = errno ? errno : ENOSYS;
1948 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1952 * First scan the list of all Infiniband devices to find
1953 * matching ones, gathering into the list.
1955 struct ibv_device *ibv_match[ret + 1];
1956 int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1957 int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1961 struct rte_pci_addr pci_addr;
1963 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1964 bd = mlx5_device_bond_pci_match
1965 (ibv_list[ret], &owner_pci, nl_rdma, owner_id,
1969 * Bonding device detected. Only one match is allowed,
1970 * the bonding is supported over multi-port IB device,
1971 * there should be no matches on representor PCI
1972 * functions or non VF LAG bonding devices with
1973 * specified address.
1977 "multiple PCI match on bonding device"
1978 "\"%s\" found", ibv_list[ret]->name);
1983 /* Amend owner pci address if owner PF ID specified. */
1984 if (eth_da.nb_representor_ports)
1985 owner_pci.function += owner_id;
1986 DRV_LOG(INFO, "PCI information matches for"
1987 " slave %d bonding device \"%s\"",
1988 bd, ibv_list[ret]->name);
1989 ibv_match[nd++] = ibv_list[ret];
1992 /* Bonding device not found. */
1993 if (mlx5_dev_to_pci_addr
1994 (ibv_list[ret]->ibdev_path, &pci_addr))
1996 if (owner_pci.domain != pci_addr.domain ||
1997 owner_pci.bus != pci_addr.bus ||
1998 owner_pci.devid != pci_addr.devid ||
1999 owner_pci.function != pci_addr.function)
2001 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2002 ibv_list[ret]->name);
2003 ibv_match[nd++] = ibv_list[ret];
2006 ibv_match[nd] = NULL;
2008 /* No device matches, just complain and bail out. */
2010 "no Verbs device matches PCI device " PCI_PRI_FMT ","
2011 " are kernel drivers loaded?",
2012 owner_pci.domain, owner_pci.bus,
2013 owner_pci.devid, owner_pci.function);
2020 * Found single matching device may have multiple ports.
2021 * Each port may be representor, we have to check the port
2022 * number and check the representors existence.
2025 np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2027 DRV_LOG(WARNING, "can not get IB device \"%s\""
2028 " ports number", ibv_match[0]->name);
2029 if (bd >= 0 && !np) {
2030 DRV_LOG(ERR, "can not get ports"
2031 " for bonding device");
2037 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2040 * This may happen if there is VF LAG kernel support and
2041 * application is compiled with older rdma_core library.
2044 "No kernel/verbs support for VF LAG bonding found.");
2045 rte_errno = ENOTSUP;
2051 * Now we can determine the maximal
2052 * amount of devices to be spawned.
2054 list = mlx5_malloc(MLX5_MEM_ZERO,
2055 sizeof(struct mlx5_dev_spawn_data) *
2057 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2059 DRV_LOG(ERR, "spawn data array allocation failure");
2064 if (bd >= 0 || np > 1) {
2066 * Single IB device with multiple ports found,
2067 * it may be E-Switch master device and representors.
2068 * We have to perform identification through the ports.
2070 MLX5_ASSERT(nl_rdma >= 0);
2071 MLX5_ASSERT(ns == 0);
2072 MLX5_ASSERT(nd == 1);
2074 for (i = 1; i <= np; ++i) {
2075 list[ns].bond_info = &bond_info;
2076 list[ns].max_port = np;
2077 list[ns].phys_port = i;
2078 list[ns].phys_dev = ibv_match[0];
2079 list[ns].eth_dev = NULL;
2080 list[ns].pci_dev = pci_dev;
2081 list[ns].pf_bond = bd;
2082 list[ns].ifindex = mlx5_nl_ifindex
2084 mlx5_os_get_dev_device_name
2085 (list[ns].phys_dev), i);
2086 if (!list[ns].ifindex) {
2088 * No network interface index found for the
2089 * specified port, it means there is no
2090 * representor on this port. It's OK,
2091 * there can be disabled ports, for example
2092 * if sriov_numvfs < sriov_totalvfs.
2098 ret = mlx5_nl_switch_info
2102 if (ret || (!list[ns].info.representor &&
2103 !list[ns].info.master)) {
2105 * We failed to recognize representors with
2106 * Netlink, let's try to perform the task
2109 ret = mlx5_sysfs_switch_info
2113 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2114 if (!ret && bd >= 0) {
2115 switch (list[ns].info.name_type) {
2116 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2117 if (list[ns].info.port_name == bd)
2120 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2122 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2124 case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2125 if (list[ns].info.pf_num == bd)
2134 if (!ret && (list[ns].info.representor ^
2135 list[ns].info.master))
2140 "unable to recognize master/representors"
2141 " on the IB device with multiple ports");
2148 * The existence of several matching entries (nd > 1) means
2149 * port representors have been instantiated. No existing Verbs
2150 * call nor sysfs entries can tell them apart, this can only
2151 * be done through Netlink calls assuming kernel drivers are
2152 * recent enough to support them.
2154 * In the event of identification failure through Netlink,
2155 * try again through sysfs, then:
2157 * 1. A single IB device matches (nd == 1) with single
2158 * port (np=0/1) and is not a representor, assume
2159 * no switch support.
2161 * 2. Otherwise no safe assumptions can be made;
2162 * complain louder and bail out.
2164 for (i = 0; i != nd; ++i) {
2165 memset(&list[ns].info, 0, sizeof(list[ns].info));
2166 list[ns].bond_info = NULL;
2167 list[ns].max_port = 1;
2168 list[ns].phys_port = 1;
2169 list[ns].phys_dev = ibv_match[i];
2170 list[ns].eth_dev = NULL;
2171 list[ns].pci_dev = pci_dev;
2172 list[ns].pf_bond = -1;
2173 list[ns].ifindex = 0;
2175 list[ns].ifindex = mlx5_nl_ifindex
2177 mlx5_os_get_dev_device_name
2178 (list[ns].phys_dev), 1);
2179 if (!list[ns].ifindex) {
2180 char ifname[IF_NAMESIZE];
2183 * Netlink failed, it may happen with old
2184 * ib_core kernel driver (before 4.16).
2185 * We can assume there is old driver because
2186 * here we are processing single ports IB
2187 * devices. Let's try sysfs to retrieve
2188 * the ifindex. The method works for
2189 * master device only.
2193 * Multiple devices found, assume
2194 * representors, can not distinguish
2195 * master/representor and retrieve
2196 * ifindex via sysfs.
2200 ret = mlx5_get_ifname_sysfs
2201 (ibv_match[i]->ibdev_path, ifname);
2204 if_nametoindex(ifname);
2205 if (!list[ns].ifindex) {
2207 * No network interface index found
2208 * for the specified device, it means
2209 * there it is neither representor
2217 ret = mlx5_nl_switch_info
2221 if (ret || (!list[ns].info.representor &&
2222 !list[ns].info.master)) {
2224 * We failed to recognize representors with
2225 * Netlink, let's try to perform the task
2228 ret = mlx5_sysfs_switch_info
2232 if (!ret && (list[ns].info.representor ^
2233 list[ns].info.master)) {
2235 } else if ((nd == 1) &&
2236 !list[ns].info.representor &&
2237 !list[ns].info.master) {
2239 * Single IB device with
2240 * one physical port and
2241 * attached network device.
2242 * May be SRIOV is not enabled
2243 * or there is no representors.
2245 DRV_LOG(INFO, "no E-Switch support detected");
2252 "unable to recognize master/representors"
2253 " on the multiple IB devices");
2261 * Sort list to probe devices in natural order for users convenience
2262 * (i.e. master first, then representors from lowest to highest ID).
2264 qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2265 /* Device specific configuration. */
2266 switch (pci_dev->id.device_id) {
2267 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2268 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2269 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2270 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2271 case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2272 case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2273 case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2280 if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2281 /* Set devargs default values. */
2282 if (eth_da.nb_mh_controllers == 0) {
2283 eth_da.nb_mh_controllers = 1;
2284 eth_da.mh_controllers[0] = 0;
2286 if (eth_da.nb_ports == 0 && ns > 0) {
2287 if (list[0].pf_bond >= 0 && list[0].info.representor)
2288 DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2289 pci_dev->device.devargs->args);
2290 eth_da.nb_ports = 1;
2291 eth_da.ports[0] = list[0].info.pf_num;
2293 if (eth_da.nb_representor_ports == 0) {
2294 eth_da.nb_representor_ports = 1;
2295 eth_da.representor_ports[0] = 0;
2298 for (i = 0; i != ns; ++i) {
2301 /* Default configuration. */
2302 memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2303 dev_config.vf = dev_config_vf;
2304 dev_config.mps = MLX5_ARG_UNSET;
2305 dev_config.dbnc = MLX5_ARG_UNSET;
2306 dev_config.rx_vec_en = 1;
2307 dev_config.txq_inline_max = MLX5_ARG_UNSET;
2308 dev_config.txq_inline_min = MLX5_ARG_UNSET;
2309 dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2310 dev_config.txqs_inline = MLX5_ARG_UNSET;
2311 dev_config.vf_nl_en = 1;
2312 dev_config.mr_ext_memseg_en = 1;
2313 dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2314 dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2315 dev_config.dv_esw_en = 1;
2316 dev_config.dv_flow_en = 1;
2317 dev_config.decap_en = 1;
2318 dev_config.log_hp_size = MLX5_ARG_UNSET;
2319 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2323 if (!list[i].eth_dev) {
2324 if (rte_errno != EBUSY && rte_errno != EEXIST)
2326 /* Device is disabled or already spawned. Ignore it. */
2329 restore = list[i].eth_dev->data->dev_flags;
2330 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2331 /* Restore non-PCI flags cleared by the above call. */
2332 list[i].eth_dev->data->dev_flags |= restore;
2333 rte_eth_dev_probing_finish(list[i].eth_dev);
2337 "probe of PCI device " PCI_PRI_FMT " aborted after"
2338 " encountering an error: %s",
2339 owner_pci.domain, owner_pci.bus,
2340 owner_pci.devid, owner_pci.function,
2341 strerror(rte_errno));
2345 if (!list[i].eth_dev)
2347 mlx5_dev_close(list[i].eth_dev);
2348 /* mac_addrs must not be freed because in dev_private */
2349 list[i].eth_dev->data->mac_addrs = NULL;
2350 claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2352 /* Restore original error. */
2359 * Do the routine cleanup:
2360 * - close opened Netlink sockets
2361 * - free allocated spawn data array
2362 * - free the Infiniband device list
2370 MLX5_ASSERT(ibv_list);
2371 mlx5_glue->free_device_list(ibv_list);
2376 * DPDK callback to register a PCI device.
2378 * This function spawns Ethernet devices out of a given PCI device.
2380 * @param[in] pci_drv
2381 * PCI driver structure (mlx5_driver).
2382 * @param[in] pci_dev
2383 * PCI device information.
2386 * 0 on success, a negative errno value otherwise and rte_errno is set.
2389 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2390 struct rte_pci_device *pci_dev)
2392 struct rte_eth_devargs eth_da = { .type = RTE_ETH_REPRESENTOR_NONE };
2396 if (pci_dev->device.devargs) {
2397 /* Parse representor information from device argument. */
2398 if (pci_dev->device.devargs->cls_str)
2399 ret = rte_eth_devargs_parse
2400 (pci_dev->device.devargs->cls_str, ð_da);
2402 DRV_LOG(ERR, "failed to parse device arguments: %s",
2403 pci_dev->device.devargs->cls_str);
2406 if (eth_da.type == RTE_ETH_REPRESENTOR_NONE) {
2407 /* Support legacy device argument */
2408 ret = rte_eth_devargs_parse
2409 (pci_dev->device.devargs->args, ð_da);
2411 DRV_LOG(ERR, "failed to parse device arguments: %s",
2412 pci_dev->device.devargs->args);
2418 if (eth_da.nb_ports > 0) {
2419 /* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2420 for (p = 0; p < eth_da.nb_ports; p++)
2421 ret = mlx5_os_pci_probe_pf(pci_dev, ð_da,
2424 ret = mlx5_os_pci_probe_pf(pci_dev, ð_da, 0);
2430 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2435 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2436 /* Get environment variable to store. */
2437 env = getenv(MLX5_SHUT_UP_BF);
2438 value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2439 if (config->dbnc == MLX5_ARG_UNSET)
2440 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2442 setenv(MLX5_SHUT_UP_BF,
2443 config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2448 mlx5_restore_doorbell_mapping_env(int value)
2450 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2451 /* Restore the original environment variable state. */
2452 if (value == MLX5_ARG_UNSET)
2453 unsetenv(MLX5_SHUT_UP_BF);
2455 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2459 * Extract pdn of PD object using DV API.
2462 * Pointer to the verbs PD object.
2464 * Pointer to the PD object number variable.
2467 * 0 on success, error value otherwise.
2470 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2472 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2473 struct mlx5dv_obj obj;
2474 struct mlx5dv_pd pd_info;
2478 obj.pd.out = &pd_info;
2479 ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2481 DRV_LOG(DEBUG, "Fail to get PD object info");
2490 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2494 * Function API to open IB device.
2496 * This function calls the Linux glue APIs to open a device.
2499 * Pointer to the IB device attributes (name, port, etc).
2500 * @param[out] config
2501 * Pointer to device configuration structure.
2503 * Pointer to shared context structure.
2506 * 0 on success, a positive error value otherwise.
2509 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2510 const struct mlx5_dev_config *config,
2511 struct mlx5_dev_ctx_shared *sh)
2516 sh->numa_node = spawn->pci_dev->device.numa_node;
2517 pthread_mutex_init(&sh->txpp.mutex, NULL);
2519 * Configure environment variable "MLX5_BF_SHUT_UP"
2520 * before the device creation. The rdma_core library
2521 * checks the variable at device creation and
2522 * stores the result internally.
2524 dbmap_env = mlx5_config_doorbell_mapping_env(config);
2525 /* Try to open IB device with DV first, then usual Verbs. */
2527 sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2530 DRV_LOG(DEBUG, "DevX is supported");
2531 /* The device is created, no need for environment. */
2532 mlx5_restore_doorbell_mapping_env(dbmap_env);
2534 /* The environment variable is still configured. */
2535 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2536 err = errno ? errno : ENODEV;
2538 * The environment variable is not needed anymore,
2539 * all device creation attempts are completed.
2541 mlx5_restore_doorbell_mapping_env(dbmap_env);
2544 DRV_LOG(DEBUG, "DevX is NOT supported");
2547 if (!err && sh->ctx) {
2548 /* Hint libmlx5 to use PMD allocator for data plane resources */
2549 mlx5_glue->dv_set_context_attr(sh->ctx,
2550 MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2551 (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2552 .alloc = &mlx5_alloc_verbs_buf,
2553 .free = &mlx5_free_verbs_buf,
2561 * Install shared asynchronous device events handler.
2562 * This function is implemented to support event sharing
2563 * between multiple ports of single IB device.
2566 * Pointer to mlx5_dev_ctx_shared object.
2569 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2574 sh->intr_handle.fd = -1;
2575 flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2576 ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2577 F_SETFL, flags | O_NONBLOCK);
2579 DRV_LOG(INFO, "failed to change file descriptor async event"
2582 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2583 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2584 if (rte_intr_callback_register(&sh->intr_handle,
2585 mlx5_dev_interrupt_handler, sh)) {
2586 DRV_LOG(INFO, "Fail to install the shared interrupt.");
2587 sh->intr_handle.fd = -1;
2591 #ifdef HAVE_IBV_DEVX_ASYNC
2592 sh->intr_handle_devx.fd = -1;
2594 (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2595 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2597 DRV_LOG(INFO, "failed to allocate devx_comp.");
2600 flags = fcntl(devx_comp->fd, F_GETFL);
2601 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2603 DRV_LOG(INFO, "failed to change file descriptor"
2607 sh->intr_handle_devx.fd = devx_comp->fd;
2608 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2609 if (rte_intr_callback_register(&sh->intr_handle_devx,
2610 mlx5_dev_interrupt_handler_devx, sh)) {
2611 DRV_LOG(INFO, "Fail to install the devx shared"
2613 sh->intr_handle_devx.fd = -1;
2615 #endif /* HAVE_IBV_DEVX_ASYNC */
2620 * Uninstall shared asynchronous device events handler.
2621 * This function is implemented to support event sharing
2622 * between multiple ports of single IB device.
2625 * Pointer to mlx5_dev_ctx_shared object.
2628 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2630 if (sh->intr_handle.fd >= 0)
2631 mlx5_intr_callback_unregister(&sh->intr_handle,
2632 mlx5_dev_interrupt_handler, sh);
2633 #ifdef HAVE_IBV_DEVX_ASYNC
2634 if (sh->intr_handle_devx.fd >= 0)
2635 rte_intr_callback_unregister(&sh->intr_handle_devx,
2636 mlx5_dev_interrupt_handler_devx, sh);
2638 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2643 * Read statistics by a named counter.
2646 * Pointer to the private device data structure.
2647 * @param[in] ctr_name
2648 * Pointer to the name of the statistic counter to read
2650 * Pointer to read statistic value.
2652 * 0 on success and stat is valud, 1 if failed to read the value
2657 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2663 if (priv->q_counters != NULL &&
2664 strcmp(ctr_name, "out_of_buffer") == 0)
2665 return mlx5_devx_cmd_queue_counter_query(priv->sh->ctx,
2666 0, (uint32_t *)stat);
2667 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2668 priv->sh->ibdev_path,
2671 fd = open(path, O_RDONLY);
2673 * in switchdev the file location is not per port
2674 * but rather in <ibdev_path>/hw_counters/<file_name>.
2677 MKSTR(path1, "%s/hw_counters/%s",
2678 priv->sh->ibdev_path,
2680 fd = open(path1, O_RDONLY);
2683 char buf[21] = {'\0'};
2684 ssize_t n = read(fd, buf, sizeof(buf));
2688 *stat = strtoull(buf, NULL, 10);
2698 * Set the reg_mr and dereg_mr call backs
2700 * @param reg_mr_cb[out]
2701 * Pointer to reg_mr func
2702 * @param dereg_mr_cb[out]
2703 * Pointer to dereg_mr func
2707 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2708 mlx5_dereg_mr_t *dereg_mr_cb)
2710 *reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2711 *dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2715 * Remove a MAC address from device
2718 * Pointer to Ethernet device structure.
2720 * MAC address index.
2723 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2725 struct mlx5_priv *priv = dev->data->dev_private;
2726 const int vf = priv->config.vf;
2729 mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2730 mlx5_ifindex(dev), priv->mac_own,
2731 &dev->data->mac_addrs[index], index);
2735 * Adds a MAC address to the device
2738 * Pointer to Ethernet device structure.
2740 * MAC address to register.
2742 * MAC address index.
2745 * 0 on success, a negative errno value otherwise
2748 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2751 struct mlx5_priv *priv = dev->data->dev_private;
2752 const int vf = priv->config.vf;
2756 ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2757 mlx5_ifindex(dev), priv->mac_own,
2763 * Modify a VF MAC address
2766 * Pointer to device private data.
2768 * MAC address to modify into.
2770 * Net device interface index
2775 * 0 on success, a negative errno value otherwise
2778 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2779 unsigned int iface_idx,
2780 struct rte_ether_addr *mac_addr,
2783 return mlx5_nl_vf_mac_addr_modify
2784 (priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2788 * Set device promiscuous mode
2791 * Pointer to Ethernet device structure.
2793 * 0 - promiscuous is disabled, otherwise - enabled
2796 * 0 on success, a negative error value otherwise
2799 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2801 struct mlx5_priv *priv = dev->data->dev_private;
2803 return mlx5_nl_promisc(priv->nl_socket_route,
2804 mlx5_ifindex(dev), !!enable);
2808 * Set device promiscuous mode
2811 * Pointer to Ethernet device structure.
2813 * 0 - all multicase is disabled, otherwise - enabled
2816 * 0 on success, a negative error value otherwise
2819 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2821 struct mlx5_priv *priv = dev->data->dev_private;
2823 return mlx5_nl_allmulti(priv->nl_socket_route,
2824 mlx5_ifindex(dev), !!enable);
2828 * Flush device MAC addresses
2831 * Pointer to Ethernet device structure.
2835 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2837 struct mlx5_priv *priv = dev->data->dev_private;
2839 mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2840 dev->data->mac_addrs,
2841 MLX5_MAX_MAC_ADDRESSES, priv->mac_own);