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;
328 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
332 err = mlx5_alloc_tunnel_hub(sh);
334 DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
337 if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
338 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
339 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
341 mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
343 sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
344 #endif /* HAVE_MLX5DV_DR */
345 sh->default_miss_action =
346 mlx5_glue->dr_create_flow_action_default_miss();
347 if (!sh->default_miss_action)
348 DRV_LOG(WARNING, "Default miss action is not supported.");
351 /* Rollback the created objects. */
353 mlx5_glue->dr_destroy_domain(sh->rx_domain);
354 sh->rx_domain = NULL;
357 mlx5_glue->dr_destroy_domain(sh->tx_domain);
358 sh->tx_domain = NULL;
360 if (sh->fdb_domain) {
361 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
362 sh->fdb_domain = NULL;
364 if (sh->esw_drop_action) {
365 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
366 sh->esw_drop_action = NULL;
368 if (sh->pop_vlan_action) {
369 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
370 sh->pop_vlan_action = NULL;
372 if (sh->encaps_decaps) {
373 mlx5_hlist_destroy(sh->encaps_decaps);
374 sh->encaps_decaps = NULL;
376 if (sh->modify_cmds) {
377 mlx5_hlist_destroy(sh->modify_cmds);
378 sh->modify_cmds = NULL;
381 /* tags should be destroyed with flow before. */
382 mlx5_hlist_destroy(sh->tag_table);
383 sh->tag_table = NULL;
385 if (sh->tunnel_hub) {
386 mlx5_release_tunnel_hub(sh, priv->dev_port);
387 sh->tunnel_hub = NULL;
389 mlx5_free_table_hash_list(priv);
394 * Destroy DR related data within private structure.
397 * Pointer to the private device data structure.
400 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
402 struct mlx5_dev_ctx_shared *sh = priv->sh;
404 MLX5_ASSERT(sh && sh->refcnt);
407 #ifdef HAVE_MLX5DV_DR
409 mlx5_glue->dr_destroy_domain(sh->rx_domain);
410 sh->rx_domain = NULL;
413 mlx5_glue->dr_destroy_domain(sh->tx_domain);
414 sh->tx_domain = NULL;
416 #ifdef HAVE_MLX5DV_DR_ESWITCH
417 if (sh->fdb_domain) {
418 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
419 sh->fdb_domain = NULL;
421 if (sh->esw_drop_action) {
422 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
423 sh->esw_drop_action = NULL;
426 if (sh->pop_vlan_action) {
427 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
428 sh->pop_vlan_action = NULL;
430 #endif /* HAVE_MLX5DV_DR */
431 if (sh->default_miss_action)
432 mlx5_glue->destroy_flow_action
433 (sh->default_miss_action);
434 if (sh->encaps_decaps) {
435 mlx5_hlist_destroy(sh->encaps_decaps);
436 sh->encaps_decaps = NULL;
438 if (sh->modify_cmds) {
439 mlx5_hlist_destroy(sh->modify_cmds);
440 sh->modify_cmds = NULL;
443 /* tags should be destroyed with flow before. */
444 mlx5_hlist_destroy(sh->tag_table);
445 sh->tag_table = NULL;
447 if (sh->tunnel_hub) {
448 mlx5_release_tunnel_hub(sh, priv->dev_port);
449 sh->tunnel_hub = NULL;
451 mlx5_cache_list_destroy(&sh->port_id_action_list);
452 mlx5_cache_list_destroy(&sh->push_vlan_action_list);
453 mlx5_free_table_hash_list(priv);
457 * Initialize shared data between primary and secondary process.
459 * A memzone is reserved by primary process and secondary processes attach to
463 * 0 on success, a negative errno value otherwise and rte_errno is set.
466 mlx5_init_shared_data(void)
468 const struct rte_memzone *mz;
471 rte_spinlock_lock(&mlx5_shared_data_lock);
472 if (mlx5_shared_data == NULL) {
473 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
474 /* Allocate shared memory. */
475 mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
476 sizeof(*mlx5_shared_data),
480 "Cannot allocate mlx5 shared data");
484 mlx5_shared_data = mz->addr;
485 memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
486 rte_spinlock_init(&mlx5_shared_data->lock);
488 /* Lookup allocated shared memory. */
489 mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
492 "Cannot attach mlx5 shared data");
496 mlx5_shared_data = mz->addr;
497 memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
501 rte_spinlock_unlock(&mlx5_shared_data_lock);
506 * PMD global initialization.
508 * Independent from individual device, this function initializes global
509 * per-PMD data structures distinguishing primary and secondary processes.
510 * Hence, each initialization is called once per a process.
513 * 0 on success, a negative errno value otherwise and rte_errno is set.
518 struct mlx5_shared_data *sd;
519 struct mlx5_local_data *ld = &mlx5_local_data;
522 if (mlx5_init_shared_data())
524 sd = mlx5_shared_data;
526 rte_spinlock_lock(&sd->lock);
527 switch (rte_eal_process_type()) {
528 case RTE_PROC_PRIMARY:
531 LIST_INIT(&sd->mem_event_cb_list);
532 rte_rwlock_init(&sd->mem_event_rwlock);
533 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
534 mlx5_mr_mem_event_cb, NULL);
535 ret = mlx5_mp_init_primary(MLX5_MP_NAME,
536 mlx5_mp_os_primary_handle);
539 sd->init_done = true;
541 case RTE_PROC_SECONDARY:
544 ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
545 mlx5_mp_os_secondary_handle);
549 ld->init_done = true;
555 rte_spinlock_unlock(&sd->lock);
560 * Create the Tx queue DevX/Verbs object.
563 * Pointer to Ethernet device.
565 * Queue index in DPDK Tx queue array.
568 * 0 on success, a negative errno value otherwise and rte_errno is set.
571 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
573 struct mlx5_priv *priv = dev->data->dev_private;
574 struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
575 struct mlx5_txq_ctrl *txq_ctrl =
576 container_of(txq_data, struct mlx5_txq_ctrl, txq);
578 if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
579 return mlx5_txq_devx_obj_new(dev, idx);
580 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
581 if (!priv->config.dv_esw_en)
582 return mlx5_txq_devx_obj_new(dev, idx);
584 return mlx5_txq_ibv_obj_new(dev, idx);
588 * Release an Tx DevX/verbs queue object.
591 * DevX/Verbs Tx queue object.
594 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
596 if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
597 mlx5_txq_devx_obj_release(txq_obj);
600 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
601 if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
602 mlx5_txq_devx_obj_release(txq_obj);
606 mlx5_txq_ibv_obj_release(txq_obj);
610 * DV flow counter mode detect and config.
613 * Pointer to rte_eth_dev structure.
617 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
619 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
620 struct mlx5_priv *priv = dev->data->dev_private;
621 struct mlx5_dev_ctx_shared *sh = priv->sh;
624 #ifndef HAVE_IBV_DEVX_ASYNC
628 if (!priv->config.devx || !priv->config.dv_flow_en ||
629 !priv->config.hca_attr.flow_counters_dump ||
630 !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
631 (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
635 DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
636 "counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
637 priv->config.hca_attr.flow_counters_dump,
638 priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
639 /* Initialize fallback mode only on the port initializes sh. */
641 sh->cmng.counter_fallback = fallback;
642 else if (fallback != sh->cmng.counter_fallback)
643 DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
644 "with others:%d.", PORT_ID(priv), fallback);
649 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
651 struct mlx5_priv *priv = dev->data->dev_private;
652 void *ctx = priv->sh->ctx;
654 priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
655 if (!priv->q_counters) {
656 struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
659 DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
660 "by DevX - fall-back to use the kernel driver global "
661 "queue counter.", dev->data->port_id);
662 /* Create WQ by kernel and query its queue counter ID. */
664 wq = mlx5_glue->create_wq(ctx,
665 &(struct ibv_wq_init_attr){
666 .wq_type = IBV_WQT_RQ,
673 /* Counter is assigned only on RDY state. */
674 int ret = mlx5_glue->modify_wq(wq,
675 &(struct ibv_wq_attr){
676 .attr_mask = IBV_WQ_ATTR_STATE,
677 .wq_state = IBV_WQS_RDY,
681 mlx5_devx_cmd_wq_query(wq,
682 &priv->counter_set_id);
683 claim_zero(mlx5_glue->destroy_wq(wq));
685 claim_zero(mlx5_glue->destroy_cq(cq));
688 priv->counter_set_id = priv->q_counters->id;
690 if (priv->counter_set_id == 0)
691 DRV_LOG(INFO, "Part of the port %d statistics will not be "
692 "available.", dev->data->port_id);
696 * Check if representor spawn info match devargs.
699 * Verbs device parameters (name, port, switch_info) to spawn.
701 * Device devargs to probe.
707 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
708 struct rte_eth_devargs *eth_da)
710 struct mlx5_switch_info *switch_info = &spawn->info;
713 uint16_t repr_id = mlx5_representor_id_encode(switch_info,
716 switch (eth_da->type) {
717 case RTE_ETH_REPRESENTOR_SF:
718 if (!(spawn->info.port_name == -1 &&
719 switch_info->name_type ==
720 MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
721 switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
726 case RTE_ETH_REPRESENTOR_VF:
727 /* Allows HPF representor index -1 as exception. */
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_PFVF) {
736 case RTE_ETH_REPRESENTOR_NONE:
741 DRV_LOG(ERR, "unsupported representor type");
744 /* Check representor ID: */
745 for (p = 0; p < eth_da->nb_ports; ++p) {
746 if (spawn->pf_bond < 0) {
747 /* For non-LAG mode, allow and ignore pf. */
748 switch_info->pf_num = eth_da->ports[p];
749 repr_id = mlx5_representor_id_encode(switch_info,
752 for (f = 0; f < eth_da->nb_representor_ports; ++f) {
753 id = MLX5_REPRESENTOR_ID
754 (eth_da->ports[p], eth_da->type,
755 eth_da->representor_ports[f]);
766 * Spawn an Ethernet device from Verbs information.
769 * Backing DPDK device.
771 * Verbs device parameters (name, port, switch_info) to spawn.
773 * Device configuration parameters.
778 * A valid Ethernet device object on success, NULL otherwise and rte_errno
779 * is set. The following errors are defined:
781 * EBUSY: device is not supposed to be spawned.
782 * EEXIST: device is already spawned
784 static struct rte_eth_dev *
785 mlx5_dev_spawn(struct rte_device *dpdk_dev,
786 struct mlx5_dev_spawn_data *spawn,
787 struct mlx5_dev_config *config,
788 struct rte_eth_devargs *eth_da)
790 const struct mlx5_switch_info *switch_info = &spawn->info;
791 struct mlx5_dev_ctx_shared *sh = NULL;
792 struct ibv_port_attr port_attr;
793 struct mlx5dv_context dv_attr = { .comp_mask = 0 };
794 struct rte_eth_dev *eth_dev = NULL;
795 struct mlx5_priv *priv = NULL;
797 unsigned int hw_padding = 0;
799 unsigned int tunnel_en = 0;
800 unsigned int mpls_en = 0;
801 unsigned int swp = 0;
802 unsigned int mprq = 0;
803 unsigned int mprq_min_stride_size_n = 0;
804 unsigned int mprq_max_stride_size_n = 0;
805 unsigned int mprq_min_stride_num_n = 0;
806 unsigned int mprq_max_stride_num_n = 0;
807 struct rte_ether_addr mac;
808 char name[RTE_ETH_NAME_MAX_LEN];
809 int own_domain_id = 0;
811 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
812 struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
815 /* Determine if this port representor is supposed to be spawned. */
816 if (switch_info->representor && dpdk_dev->devargs &&
817 !mlx5_representor_match(spawn, eth_da))
819 /* Build device name. */
820 if (spawn->pf_bond < 0) {
822 if (!switch_info->representor)
823 strlcpy(name, dpdk_dev->name, sizeof(name));
825 err = snprintf(name, sizeof(name), "%s_representor_%s%u",
827 switch_info->name_type ==
828 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
829 switch_info->port_name);
831 /* Bonding device. */
832 if (!switch_info->representor) {
833 err = snprintf(name, sizeof(name), "%s_%s",
835 mlx5_os_get_dev_device_name(spawn->phys_dev));
837 err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
839 mlx5_os_get_dev_device_name(spawn->phys_dev),
840 switch_info->ctrl_num,
842 switch_info->name_type ==
843 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
844 switch_info->port_name);
847 if (err >= (int)sizeof(name))
848 DRV_LOG(WARNING, "device name overflow %s", name);
849 /* check if the device is already spawned */
850 if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
854 DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
855 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
856 struct mlx5_mp_id mp_id;
858 eth_dev = rte_eth_dev_attach_secondary(name);
859 if (eth_dev == NULL) {
860 DRV_LOG(ERR, "can not attach rte ethdev");
864 eth_dev->device = dpdk_dev;
865 eth_dev->dev_ops = &mlx5_dev_sec_ops;
866 eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
867 eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
868 err = mlx5_proc_priv_init(eth_dev);
871 mp_id.port_id = eth_dev->data->port_id;
872 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
873 /* Receive command fd from primary process */
874 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
877 /* Remap UAR for Tx queues. */
878 err = mlx5_tx_uar_init_secondary(eth_dev, err);
882 * Ethdev pointer is still required as input since
883 * the primary device is not accessible from the
886 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
887 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
890 mlx5_dev_close(eth_dev);
894 * Some parameters ("tx_db_nc" in particularly) are needed in
895 * advance to create dv/verbs device context. We proceed the
896 * devargs here to get ones, and later proceed devargs again
897 * to override some hardware settings.
899 err = mlx5_args(config, dpdk_dev->devargs);
902 DRV_LOG(ERR, "failed to process device arguments: %s",
903 strerror(rte_errno));
906 if (config->dv_miss_info) {
907 if (switch_info->master || switch_info->representor)
908 config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
910 mlx5_malloc_mem_select(config->sys_mem_en);
911 sh = mlx5_alloc_shared_dev_ctx(spawn, config);
914 config->devx = sh->devx;
915 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
916 config->dest_tir = 1;
918 #ifdef HAVE_IBV_MLX5_MOD_SWP
919 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
922 * Multi-packet send is supported by ConnectX-4 Lx PF as well
923 * as all ConnectX-5 devices.
925 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
926 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
928 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
929 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
931 mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
932 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
933 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
934 DRV_LOG(DEBUG, "enhanced MPW is supported");
935 mps = MLX5_MPW_ENHANCED;
937 DRV_LOG(DEBUG, "MPW is supported");
941 DRV_LOG(DEBUG, "MPW isn't supported");
942 mps = MLX5_MPW_DISABLED;
944 #ifdef HAVE_IBV_MLX5_MOD_SWP
945 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
946 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
947 DRV_LOG(DEBUG, "SWP support: %u", swp);
950 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
951 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
952 struct mlx5dv_striding_rq_caps mprq_caps =
953 dv_attr.striding_rq_caps;
955 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
956 mprq_caps.min_single_stride_log_num_of_bytes);
957 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
958 mprq_caps.max_single_stride_log_num_of_bytes);
959 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
960 mprq_caps.min_single_wqe_log_num_of_strides);
961 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
962 mprq_caps.max_single_wqe_log_num_of_strides);
963 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
964 mprq_caps.supported_qpts);
965 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
967 mprq_min_stride_size_n =
968 mprq_caps.min_single_stride_log_num_of_bytes;
969 mprq_max_stride_size_n =
970 mprq_caps.max_single_stride_log_num_of_bytes;
971 mprq_min_stride_num_n =
972 mprq_caps.min_single_wqe_log_num_of_strides;
973 mprq_max_stride_num_n =
974 mprq_caps.max_single_wqe_log_num_of_strides;
977 /* Rx CQE compression is enabled by default. */
978 config->cqe_comp = 1;
979 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
980 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
981 tunnel_en = ((dv_attr.tunnel_offloads_caps &
982 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
983 (dv_attr.tunnel_offloads_caps &
984 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
985 (dv_attr.tunnel_offloads_caps &
986 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
988 DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
989 tunnel_en ? "" : "not ");
992 "tunnel offloading disabled due to old OFED/rdma-core version");
994 config->tunnel_en = tunnel_en;
995 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
996 mpls_en = ((dv_attr.tunnel_offloads_caps &
997 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
998 (dv_attr.tunnel_offloads_caps &
999 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1000 DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1001 mpls_en ? "" : "not ");
1003 DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1004 " old OFED/rdma-core version or firmware configuration");
1006 config->mpls_en = mpls_en;
1007 /* Check port status. */
1008 err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1010 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1013 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1014 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1018 if (port_attr.state != IBV_PORT_ACTIVE)
1019 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1020 mlx5_glue->port_state_str(port_attr.state),
1022 /* Allocate private eth device data. */
1023 priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1025 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1027 DRV_LOG(ERR, "priv allocation failure");
1032 priv->dev_port = spawn->phys_port;
1033 priv->pci_dev = spawn->pci_dev;
1034 priv->mtu = RTE_ETHER_MTU;
1035 /* Some internal functions rely on Netlink sockets, open them now. */
1036 priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1037 priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1038 priv->representor = !!switch_info->representor;
1039 priv->master = !!switch_info->master;
1040 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1041 priv->vport_meta_tag = 0;
1042 priv->vport_meta_mask = 0;
1043 priv->pf_bond = spawn->pf_bond;
1044 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1046 * The DevX port query API is implemented. E-Switch may use
1047 * either vport or reg_c[0] metadata register to match on
1048 * vport index. The engaged part of metadata register is
1051 if (switch_info->representor || switch_info->master) {
1052 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
1053 MLX5DV_DEVX_PORT_MATCH_REG_C_0;
1054 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
1058 "can't query devx port %d on device %s",
1060 mlx5_os_get_dev_device_name(spawn->phys_dev));
1061 devx_port.comp_mask = 0;
1064 if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
1065 priv->vport_meta_tag = devx_port.reg_c_0.value;
1066 priv->vport_meta_mask = devx_port.reg_c_0.mask;
1067 if (!priv->vport_meta_mask) {
1068 DRV_LOG(ERR, "vport zero mask for port %d"
1069 " on bonding device %s",
1071 mlx5_os_get_dev_device_name
1076 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1077 DRV_LOG(ERR, "invalid vport tag for port %d"
1078 " on bonding device %s",
1080 mlx5_os_get_dev_device_name
1086 if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
1087 priv->vport_id = devx_port.vport_num;
1088 } else if (spawn->pf_bond >= 0) {
1089 DRV_LOG(ERR, "can't deduce vport index for port %d"
1090 " on bonding device %s",
1092 mlx5_os_get_dev_device_name(spawn->phys_dev));
1096 /* Suppose vport index in compatible way. */
1097 priv->vport_id = switch_info->representor ?
1098 switch_info->port_name + 1 : -1;
1102 * Kernel/rdma_core support single E-Switch per PF configurations
1103 * only and vport_id field contains the vport index for
1104 * associated VF, which is deduced from representor port name.
1105 * For example, let's have the IB device port 10, it has
1106 * attached network device eth0, which has port name attribute
1107 * pf0vf2, we can deduce the VF number as 2, and set vport index
1108 * as 3 (2+1). This assigning schema should be changed if the
1109 * multiple E-Switch instances per PF configurations or/and PCI
1110 * subfunctions are added.
1112 priv->vport_id = switch_info->representor ?
1113 switch_info->port_name + 1 : -1;
1115 priv->representor_id = mlx5_representor_id_encode(switch_info,
1118 * Look for sibling devices in order to reuse their switch domain
1119 * if any, otherwise allocate one.
1121 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1122 const struct mlx5_priv *opriv =
1123 rte_eth_devices[port_id].data->dev_private;
1126 opriv->sh != priv->sh ||
1128 RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1130 priv->domain_id = opriv->domain_id;
1133 if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1134 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1137 DRV_LOG(ERR, "unable to allocate switch domain: %s",
1138 strerror(rte_errno));
1143 /* Override some values set by hardware configuration. */
1144 mlx5_args(config, dpdk_dev->devargs);
1145 err = mlx5_dev_check_sibling_config(priv, config);
1148 config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1149 IBV_DEVICE_RAW_IP_CSUM);
1150 DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1151 (config->hw_csum ? "" : "not "));
1152 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1153 !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1154 DRV_LOG(DEBUG, "counters are not supported");
1156 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1157 if (config->dv_flow_en) {
1158 DRV_LOG(WARNING, "DV flow is not supported");
1159 config->dv_flow_en = 0;
1162 config->ind_table_max_size =
1163 sh->device_attr.max_rwq_indirection_table_size;
1165 * Remove this check once DPDK supports larger/variable
1166 * indirection tables.
1168 if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1169 config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1170 DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1171 config->ind_table_max_size);
1172 config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1173 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1174 DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1175 (config->hw_vlan_strip ? "" : "not "));
1176 config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1177 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1178 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1179 hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1180 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1181 hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1182 IBV_DEVICE_PCI_WRITE_END_PADDING);
1184 if (config->hw_padding && !hw_padding) {
1185 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1186 config->hw_padding = 0;
1187 } else if (config->hw_padding) {
1188 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1190 config->tso = (sh->device_attr.max_tso > 0 &&
1191 (sh->device_attr.tso_supported_qpts &
1192 (1 << IBV_QPT_RAW_PACKET)));
1194 config->tso_max_payload_sz = sh->device_attr.max_tso;
1196 * MPW is disabled by default, while the Enhanced MPW is enabled
1199 if (config->mps == MLX5_ARG_UNSET)
1200 config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1203 config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1204 DRV_LOG(INFO, "%sMPS is %s",
1205 config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1206 config->mps == MLX5_MPW ? "legacy " : "",
1207 config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1209 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1214 /* Check relax ordering support. */
1215 if (!haswell_broadwell_cpu) {
1216 sh->cmng.relaxed_ordering_write =
1217 config->hca_attr.relaxed_ordering_write;
1218 sh->cmng.relaxed_ordering_read =
1219 config->hca_attr.relaxed_ordering_read;
1221 sh->cmng.relaxed_ordering_read = 0;
1222 sh->cmng.relaxed_ordering_write = 0;
1224 sh->rq_ts_format = config->hca_attr.rq_ts_format;
1225 sh->sq_ts_format = config->hca_attr.sq_ts_format;
1226 sh->qp_ts_format = config->hca_attr.qp_ts_format;
1227 /* Check for LRO support. */
1228 if (config->dest_tir && config->hca_attr.lro_cap &&
1229 config->dv_flow_en) {
1230 /* TBD check tunnel lro caps. */
1231 config->lro.supported = config->hca_attr.lro_cap;
1232 DRV_LOG(DEBUG, "Device supports LRO");
1234 * If LRO timeout is not configured by application,
1235 * use the minimal supported value.
1237 if (!config->lro.timeout)
1238 config->lro.timeout =
1239 config->hca_attr.lro_timer_supported_periods[0];
1240 DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1241 config->lro.timeout);
1242 DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1243 "required for coalescing is %d bytes",
1244 config->hca_attr.lro_min_mss_size);
1246 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1247 if (config->hca_attr.qos.sup &&
1248 config->hca_attr.qos.flow_meter_old &&
1249 config->dv_flow_en) {
1250 uint8_t reg_c_mask =
1251 config->hca_attr.qos.flow_meter_reg_c_ids;
1253 * Meter needs two REG_C's for color match and pre-sfx
1254 * flow match. Here get the REG_C for color match.
1255 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1258 if (__builtin_popcount(reg_c_mask) < 1) {
1260 DRV_LOG(WARNING, "No available register for"
1264 * The meter color register is used by the
1265 * flow-hit feature as well.
1266 * The flow-hit feature must use REG_C_3
1267 * Prefer REG_C_3 if it is available.
1269 if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1270 priv->mtr_color_reg = REG_C_3;
1272 priv->mtr_color_reg = ffs(reg_c_mask)
1275 priv->mtr_reg_share =
1276 config->hca_attr.qos.flow_meter;
1277 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1278 priv->mtr_color_reg);
1282 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1283 if (config->hca_attr.flow_hit_aso &&
1284 priv->mtr_color_reg == REG_C_3) {
1285 sh->flow_hit_aso_en = 1;
1286 err = mlx5_flow_aso_age_mng_init(sh);
1291 DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1293 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1294 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1295 if (config->hca_attr.log_max_ft_sampler_num > 0 &&
1296 config->dv_flow_en) {
1297 priv->sampler_en = 1;
1298 DRV_LOG(DEBUG, "Sampler enabled!");
1300 priv->sampler_en = 0;
1301 if (!config->hca_attr.log_max_ft_sampler_num)
1303 "No available register for sampler.");
1305 DRV_LOG(DEBUG, "DV flow is not supported!");
1309 if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1310 !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1311 DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1312 config->cqe_comp = 0;
1314 if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1315 (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1316 DRV_LOG(WARNING, "Flow Tag CQE compression"
1317 " format isn't supported.");
1318 config->cqe_comp = 0;
1320 if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1321 (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1322 DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1323 " format isn't supported.");
1324 config->cqe_comp = 0;
1326 DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1327 config->cqe_comp ? "" : "not ");
1328 if (config->tx_pp) {
1329 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1330 config->hca_attr.dev_freq_khz);
1331 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1332 config->hca_attr.qos.packet_pacing ? "" : "not ");
1333 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1334 config->hca_attr.cross_channel ? "" : "not ");
1335 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1336 config->hca_attr.wqe_index_ignore ? "" : "not ");
1337 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1338 config->hca_attr.non_wire_sq ? "" : "not ");
1339 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1340 config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1341 config->hca_attr.log_max_static_sq_wq);
1342 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1343 config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1344 if (!config->devx) {
1345 DRV_LOG(ERR, "DevX is required for packet pacing");
1349 if (!config->hca_attr.qos.packet_pacing) {
1350 DRV_LOG(ERR, "Packet pacing is not supported");
1354 if (!config->hca_attr.cross_channel) {
1355 DRV_LOG(ERR, "Cross channel operations are"
1356 " required for packet pacing");
1360 if (!config->hca_attr.wqe_index_ignore) {
1361 DRV_LOG(ERR, "WQE index ignore feature is"
1362 " required for packet pacing");
1366 if (!config->hca_attr.non_wire_sq) {
1367 DRV_LOG(ERR, "Non-wire SQ feature is"
1368 " required for packet pacing");
1372 if (!config->hca_attr.log_max_static_sq_wq) {
1373 DRV_LOG(ERR, "Static WQE SQ feature is"
1374 " required for packet pacing");
1378 if (!config->hca_attr.qos.wqe_rate_pp) {
1379 DRV_LOG(ERR, "WQE rate mode is required"
1380 " for packet pacing");
1384 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1385 DRV_LOG(ERR, "DevX does not provide UAR offset,"
1386 " can't create queues for packet pacing");
1392 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1394 err = config->hca_attr.access_register_user ?
1395 mlx5_devx_cmd_register_read
1396 (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1397 reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1401 /* MTUTC register is read successfully. */
1402 ts_mode = MLX5_GET(register_mtutc, reg,
1404 if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1405 config->rt_timestamp = 1;
1407 /* Kernel does not support register reading. */
1408 if (config->hca_attr.dev_freq_khz ==
1409 (NS_PER_S / MS_PER_S))
1410 config->rt_timestamp = 1;
1414 * If HW has bug working with tunnel packet decapsulation and
1415 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1416 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1418 if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1419 config->hw_fcs_strip = 0;
1420 DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1421 (config->hw_fcs_strip ? "" : "not "));
1422 if (config->mprq.enabled && mprq) {
1423 if (config->mprq.stride_num_n &&
1424 (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1425 config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1426 config->mprq.stride_num_n =
1427 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1428 mprq_min_stride_num_n),
1429 mprq_max_stride_num_n);
1431 "the number of strides"
1432 " for Multi-Packet RQ is out of range,"
1433 " setting default value (%u)",
1434 1 << config->mprq.stride_num_n);
1436 if (config->mprq.stride_size_n &&
1437 (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1438 config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1439 config->mprq.stride_size_n =
1440 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1441 mprq_min_stride_size_n),
1442 mprq_max_stride_size_n);
1444 "the size of a stride"
1445 " for Multi-Packet RQ is out of range,"
1446 " setting default value (%u)",
1447 1 << config->mprq.stride_size_n);
1449 config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1450 config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1451 } else if (config->mprq.enabled && !mprq) {
1452 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1453 config->mprq.enabled = 0;
1455 if (config->max_dump_files_num == 0)
1456 config->max_dump_files_num = 128;
1457 eth_dev = rte_eth_dev_allocate(name);
1458 if (eth_dev == NULL) {
1459 DRV_LOG(ERR, "can not allocate rte ethdev");
1463 if (priv->representor) {
1464 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1465 eth_dev->data->representor_id = priv->representor_id;
1467 priv->mp_id.port_id = eth_dev->data->port_id;
1468 strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1470 * Store associated network device interface index. This index
1471 * is permanent throughout the lifetime of device. So, we may store
1472 * the ifindex here and use the cached value further.
1474 MLX5_ASSERT(spawn->ifindex);
1475 priv->if_index = spawn->ifindex;
1476 eth_dev->data->dev_private = priv;
1477 priv->dev_data = eth_dev->data;
1478 eth_dev->data->mac_addrs = priv->mac;
1479 eth_dev->device = dpdk_dev;
1480 eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1481 /* Configure the first MAC address by default. */
1482 if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1484 "port %u cannot get MAC address, is mlx5_en"
1485 " loaded? (errno: %s)",
1486 eth_dev->data->port_id, strerror(rte_errno));
1491 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1492 eth_dev->data->port_id,
1493 mac.addr_bytes[0], mac.addr_bytes[1],
1494 mac.addr_bytes[2], mac.addr_bytes[3],
1495 mac.addr_bytes[4], mac.addr_bytes[5]);
1496 #ifdef RTE_LIBRTE_MLX5_DEBUG
1498 char ifname[MLX5_NAMESIZE];
1500 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1501 DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1502 eth_dev->data->port_id, ifname);
1504 DRV_LOG(DEBUG, "port %u ifname is unknown",
1505 eth_dev->data->port_id);
1508 /* Get actual MTU if possible. */
1509 err = mlx5_get_mtu(eth_dev, &priv->mtu);
1514 DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1516 /* Initialize burst functions to prevent crashes before link-up. */
1517 eth_dev->rx_pkt_burst = removed_rx_burst;
1518 eth_dev->tx_pkt_burst = removed_tx_burst;
1519 eth_dev->dev_ops = &mlx5_dev_ops;
1520 eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1521 eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1522 eth_dev->rx_queue_count = mlx5_rx_queue_count;
1523 /* Register MAC address. */
1524 claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1525 if (config->vf && config->vf_nl_en)
1526 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1527 mlx5_ifindex(eth_dev),
1528 eth_dev->data->mac_addrs,
1529 MLX5_MAX_MAC_ADDRESSES);
1531 priv->ctrl_flows = 0;
1532 rte_spinlock_init(&priv->flow_list_lock);
1533 TAILQ_INIT(&priv->flow_meters);
1534 TAILQ_INIT(&priv->flow_meter_profiles);
1535 /* Hint libmlx5 to use PMD allocator for data plane resources */
1536 mlx5_glue->dv_set_context_attr(sh->ctx,
1537 MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1538 (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1539 .alloc = &mlx5_alloc_verbs_buf,
1540 .free = &mlx5_free_verbs_buf,
1543 /* Bring Ethernet device up. */
1544 DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1545 eth_dev->data->port_id);
1546 mlx5_set_link_up(eth_dev);
1548 * Even though the interrupt handler is not installed yet,
1549 * interrupts will still trigger on the async_fd from
1550 * Verbs context returned by ibv_open_device().
1552 mlx5_link_update(eth_dev, 0);
1553 #ifdef HAVE_MLX5DV_DR_ESWITCH
1554 if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1555 (switch_info->representor || switch_info->master)))
1556 config->dv_esw_en = 0;
1558 config->dv_esw_en = 0;
1560 /* Detect minimal data bytes to inline. */
1561 mlx5_set_min_inline(spawn, config);
1562 /* Store device configuration on private structure. */
1563 priv->config = *config;
1564 /* Create context for virtual machine VLAN workaround. */
1565 priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1566 if (config->dv_flow_en) {
1567 err = mlx5_alloc_shared_dr(priv);
1571 if (config->devx && config->dv_flow_en && config->dest_tir) {
1572 priv->obj_ops = devx_obj_ops;
1573 priv->obj_ops.drop_action_create =
1574 ibv_obj_ops.drop_action_create;
1575 priv->obj_ops.drop_action_destroy =
1576 ibv_obj_ops.drop_action_destroy;
1577 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1578 priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1580 if (config->dv_esw_en)
1581 priv->obj_ops.txq_obj_modify =
1582 ibv_obj_ops.txq_obj_modify;
1584 /* Use specific wrappers for Tx object. */
1585 priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1586 priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1587 mlx5_queue_counter_id_prepare(eth_dev);
1590 priv->obj_ops = ibv_obj_ops;
1592 priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1593 if (!priv->drop_queue.hrxq)
1595 /* Supported Verbs flow priority number detection. */
1596 err = mlx5_flow_discover_priorities(eth_dev);
1601 priv->config.flow_prio = err;
1602 if (!priv->config.dv_esw_en &&
1603 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1604 DRV_LOG(WARNING, "metadata mode %u is not supported "
1605 "(no E-Switch)", priv->config.dv_xmeta_en);
1606 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1608 mlx5_set_metadata_mask(eth_dev);
1609 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1610 !priv->sh->dv_regc0_mask) {
1611 DRV_LOG(ERR, "metadata mode %u is not supported "
1612 "(no metadata reg_c[0] is available)",
1613 priv->config.dv_xmeta_en);
1617 mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1618 mlx5_hrxq_create_cb,
1620 mlx5_hrxq_remove_cb);
1621 /* Query availability of metadata reg_c's. */
1622 err = mlx5_flow_discover_mreg_c(eth_dev);
1627 if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1629 "port %u extensive metadata register is not supported",
1630 eth_dev->data->port_id);
1631 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1632 DRV_LOG(ERR, "metadata mode %u is not supported "
1633 "(no metadata registers available)",
1634 priv->config.dv_xmeta_en);
1639 if (priv->config.dv_flow_en &&
1640 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1641 mlx5_flow_ext_mreg_supported(eth_dev) &&
1642 priv->sh->dv_regc0_mask) {
1643 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1644 MLX5_FLOW_MREG_HTABLE_SZ,
1646 flow_dv_mreg_create_cb,
1647 flow_dv_mreg_match_cb,
1648 flow_dv_mreg_remove_cb);
1649 if (!priv->mreg_cp_tbl) {
1653 priv->mreg_cp_tbl->ctx = eth_dev;
1655 rte_spinlock_init(&priv->shared_act_sl);
1656 mlx5_flow_counter_mode_config(eth_dev);
1657 if (priv->config.dv_flow_en)
1658 eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1662 if (priv->mreg_cp_tbl)
1663 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1665 mlx5_os_free_shared_dr(priv);
1666 if (priv->nl_socket_route >= 0)
1667 close(priv->nl_socket_route);
1668 if (priv->nl_socket_rdma >= 0)
1669 close(priv->nl_socket_rdma);
1670 if (priv->vmwa_context)
1671 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1672 if (eth_dev && priv->drop_queue.hrxq)
1673 mlx5_drop_action_destroy(eth_dev);
1675 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1676 mlx5_cache_list_destroy(&priv->hrxqs);
1678 if (eth_dev != NULL)
1679 eth_dev->data->dev_private = NULL;
1681 if (eth_dev != NULL) {
1682 /* mac_addrs must not be freed alone because part of
1685 eth_dev->data->mac_addrs = NULL;
1686 rte_eth_dev_release_port(eth_dev);
1689 mlx5_free_shared_dev_ctx(sh);
1690 MLX5_ASSERT(err > 0);
1696 * Comparison callback to sort device data.
1698 * This is meant to be used with qsort().
1701 * Pointer to pointer to first data object.
1703 * Pointer to pointer to second data object.
1706 * 0 if both objects are equal, less than 0 if the first argument is less
1707 * than the second, greater than 0 otherwise.
1710 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1712 const struct mlx5_switch_info *si_a =
1713 &((const struct mlx5_dev_spawn_data *)a)->info;
1714 const struct mlx5_switch_info *si_b =
1715 &((const struct mlx5_dev_spawn_data *)b)->info;
1718 /* Master device first. */
1719 ret = si_b->master - si_a->master;
1722 /* Then representor devices. */
1723 ret = si_b->representor - si_a->representor;
1726 /* Unidentified devices come last in no specific order. */
1727 if (!si_a->representor)
1729 /* Order representors by name. */
1730 return si_a->port_name - si_b->port_name;
1734 * Match PCI information for possible slaves of bonding device.
1736 * @param[in] ibv_dev
1737 * Pointer to Infiniband device structure.
1738 * @param[in] pci_dev
1739 * Pointer to primary PCI address structure to match.
1740 * @param[in] nl_rdma
1741 * Netlink RDMA group socket handle.
1743 * Rerepsentor owner PF index.
1744 * @param[out] bond_info
1745 * Pointer to bonding information.
1748 * negative value if no bonding device found, otherwise
1749 * positive index of slave PF in bonding.
1752 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1753 const struct rte_pci_addr *pci_dev,
1754 int nl_rdma, uint16_t owner,
1755 struct mlx5_bond_info *bond_info)
1757 char ifname[IF_NAMESIZE + 1];
1758 unsigned int ifindex;
1760 FILE *bond_file = NULL, *file;
1765 * Try to get master device name. If something goes
1766 * wrong suppose the lack of kernel support and no
1769 memset(bond_info, 0, sizeof(*bond_info));
1772 if (!strstr(ibv_dev->name, "bond"))
1774 np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1778 * The Master device might not be on the predefined
1779 * port (not on port index 1, it is not garanted),
1780 * we have to scan all Infiniband device port and
1783 for (i = 1; i <= np; ++i) {
1784 /* Check whether Infiniband port is populated. */
1785 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1788 if (!if_indextoname(ifindex, ifname))
1790 /* Try to read bonding slave names from sysfs. */
1792 "/sys/class/net/%s/master/bonding/slaves", ifname);
1793 bond_file = fopen(slaves, "r");
1799 /* Use safe format to check maximal buffer length. */
1800 MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1801 while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1802 char tmp_str[IF_NAMESIZE + 32];
1803 struct rte_pci_addr pci_addr;
1804 struct mlx5_switch_info info;
1806 /* Process slave interface names in the loop. */
1807 snprintf(tmp_str, sizeof(tmp_str),
1808 "/sys/class/net/%s", ifname);
1809 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1810 DRV_LOG(WARNING, "can not get PCI address"
1811 " for netdev \"%s\"", ifname);
1814 /* Slave interface PCI address match found. */
1815 snprintf(tmp_str, sizeof(tmp_str),
1816 "/sys/class/net/%s/phys_port_name", ifname);
1817 file = fopen(tmp_str, "rb");
1820 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1821 if (fscanf(file, "%32s", tmp_str) == 1)
1822 mlx5_translate_port_name(tmp_str, &info);
1824 /* Only process PF ports. */
1825 if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1826 info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1828 /* Check max bonding member. */
1829 if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1830 DRV_LOG(WARNING, "bonding index out of range, "
1831 "please increase MLX5_BOND_MAX_PORTS: %s",
1835 /* Match PCI address. */
1836 if (pci_dev->domain == pci_addr.domain &&
1837 pci_dev->bus == pci_addr.bus &&
1838 pci_dev->devid == pci_addr.devid &&
1839 pci_dev->function + owner == pci_addr.function)
1840 pf = info.port_name;
1842 snprintf(tmp_str, sizeof(tmp_str),
1843 "/sys/class/net/%s/ifindex", ifname);
1844 file = fopen(tmp_str, "rb");
1847 ret = fscanf(file, "%u", &ifindex);
1851 /* Save bonding info. */
1852 strncpy(bond_info->ports[info.port_name].ifname, ifname,
1853 sizeof(bond_info->ports[0].ifname));
1854 bond_info->ports[info.port_name].pci_addr = pci_addr;
1855 bond_info->ports[info.port_name].ifindex = ifindex;
1856 bond_info->n_port++;
1859 /* Get bond interface info */
1860 ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1863 DRV_LOG(ERR, "unable to get bond info: %s",
1864 strerror(rte_errno));
1866 DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1867 ifindex, bond_info->ifindex, bond_info->ifname);
1873 * Register a PCI device within bonding.
1875 * This function spawns Ethernet devices out of a given PCI device and
1876 * bonding owner PF index.
1878 * @param[in] pci_dev
1879 * PCI device information.
1880 * @param[in] req_eth_da
1881 * Requested ethdev device argument.
1882 * @param[in] owner_id
1883 * Requested owner PF port ID within bonding device, default to 0.
1886 * 0 on success, a negative errno value otherwise and rte_errno is set.
1889 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
1890 struct rte_eth_devargs *req_eth_da,
1893 struct ibv_device **ibv_list;
1895 * Number of found IB Devices matching with requested PCI BDF.
1896 * nd != 1 means there are multiple IB devices over the same
1897 * PCI device and we have representors and master.
1899 unsigned int nd = 0;
1901 * Number of found IB device Ports. nd = 1 and np = 1..n means
1902 * we have the single multiport IB device, and there may be
1903 * representors attached to some of found ports.
1905 unsigned int np = 0;
1907 * Number of DPDK ethernet devices to Spawn - either over
1908 * multiple IB devices or multiple ports of single IB device.
1909 * Actually this is the number of iterations to spawn.
1911 unsigned int ns = 0;
1914 * < 0 - no bonding device (single one)
1915 * >= 0 - bonding device (value is slave PF index)
1918 struct mlx5_dev_spawn_data *list = NULL;
1919 struct mlx5_dev_config dev_config;
1920 unsigned int dev_config_vf;
1921 struct rte_eth_devargs eth_da = *req_eth_da;
1922 struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1923 struct mlx5_bond_info bond_info;
1926 if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1927 mlx5_pmd_socket_init();
1928 ret = mlx5_init_once();
1930 DRV_LOG(ERR, "unable to init PMD global data: %s",
1931 strerror(rte_errno));
1935 ibv_list = mlx5_glue->get_device_list(&ret);
1937 rte_errno = errno ? errno : ENOSYS;
1938 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1942 * First scan the list of all Infiniband devices to find
1943 * matching ones, gathering into the list.
1945 struct ibv_device *ibv_match[ret + 1];
1946 int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1947 int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1951 struct rte_pci_addr pci_addr;
1953 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1954 bd = mlx5_device_bond_pci_match
1955 (ibv_list[ret], &owner_pci, nl_rdma, owner_id,
1959 * Bonding device detected. Only one match is allowed,
1960 * the bonding is supported over multi-port IB device,
1961 * there should be no matches on representor PCI
1962 * functions or non VF LAG bonding devices with
1963 * specified address.
1967 "multiple PCI match on bonding device"
1968 "\"%s\" found", ibv_list[ret]->name);
1973 /* Amend owner pci address if owner PF ID specified. */
1974 if (eth_da.nb_representor_ports)
1975 owner_pci.function += owner_id;
1976 DRV_LOG(INFO, "PCI information matches for"
1977 " slave %d bonding device \"%s\"",
1978 bd, ibv_list[ret]->name);
1979 ibv_match[nd++] = ibv_list[ret];
1982 /* Bonding device not found. */
1983 if (mlx5_dev_to_pci_addr
1984 (ibv_list[ret]->ibdev_path, &pci_addr))
1986 if (owner_pci.domain != pci_addr.domain ||
1987 owner_pci.bus != pci_addr.bus ||
1988 owner_pci.devid != pci_addr.devid ||
1989 owner_pci.function != pci_addr.function)
1991 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1992 ibv_list[ret]->name);
1993 ibv_match[nd++] = ibv_list[ret];
1996 ibv_match[nd] = NULL;
1998 /* No device matches, just complain and bail out. */
2000 "no Verbs device matches PCI device " PCI_PRI_FMT ","
2001 " are kernel drivers loaded?",
2002 owner_pci.domain, owner_pci.bus,
2003 owner_pci.devid, owner_pci.function);
2010 * Found single matching device may have multiple ports.
2011 * Each port may be representor, we have to check the port
2012 * number and check the representors existence.
2015 np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2017 DRV_LOG(WARNING, "can not get IB device \"%s\""
2018 " ports number", ibv_match[0]->name);
2019 if (bd >= 0 && !np) {
2020 DRV_LOG(ERR, "can not get ports"
2021 " for bonding device");
2027 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2030 * This may happen if there is VF LAG kernel support and
2031 * application is compiled with older rdma_core library.
2034 "No kernel/verbs support for VF LAG bonding found.");
2035 rte_errno = ENOTSUP;
2041 * Now we can determine the maximal
2042 * amount of devices to be spawned.
2044 list = mlx5_malloc(MLX5_MEM_ZERO,
2045 sizeof(struct mlx5_dev_spawn_data) *
2047 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2049 DRV_LOG(ERR, "spawn data array allocation failure");
2054 if (bd >= 0 || np > 1) {
2056 * Single IB device with multiple ports found,
2057 * it may be E-Switch master device and representors.
2058 * We have to perform identification through the ports.
2060 MLX5_ASSERT(nl_rdma >= 0);
2061 MLX5_ASSERT(ns == 0);
2062 MLX5_ASSERT(nd == 1);
2064 for (i = 1; i <= np; ++i) {
2065 list[ns].bond_info = &bond_info;
2066 list[ns].max_port = np;
2067 list[ns].phys_port = i;
2068 list[ns].phys_dev = ibv_match[0];
2069 list[ns].eth_dev = NULL;
2070 list[ns].pci_dev = pci_dev;
2071 list[ns].pf_bond = bd;
2072 list[ns].ifindex = mlx5_nl_ifindex
2074 mlx5_os_get_dev_device_name
2075 (list[ns].phys_dev), i);
2076 if (!list[ns].ifindex) {
2078 * No network interface index found for the
2079 * specified port, it means there is no
2080 * representor on this port. It's OK,
2081 * there can be disabled ports, for example
2082 * if sriov_numvfs < sriov_totalvfs.
2088 ret = mlx5_nl_switch_info
2092 if (ret || (!list[ns].info.representor &&
2093 !list[ns].info.master)) {
2095 * We failed to recognize representors with
2096 * Netlink, let's try to perform the task
2099 ret = mlx5_sysfs_switch_info
2103 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2104 if (!ret && bd >= 0) {
2105 switch (list[ns].info.name_type) {
2106 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2107 if (list[ns].info.port_name == bd)
2110 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2112 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2114 case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2115 if (list[ns].info.pf_num == bd)
2124 if (!ret && (list[ns].info.representor ^
2125 list[ns].info.master))
2130 "unable to recognize master/representors"
2131 " on the IB device with multiple ports");
2138 * The existence of several matching entries (nd > 1) means
2139 * port representors have been instantiated. No existing Verbs
2140 * call nor sysfs entries can tell them apart, this can only
2141 * be done through Netlink calls assuming kernel drivers are
2142 * recent enough to support them.
2144 * In the event of identification failure through Netlink,
2145 * try again through sysfs, then:
2147 * 1. A single IB device matches (nd == 1) with single
2148 * port (np=0/1) and is not a representor, assume
2149 * no switch support.
2151 * 2. Otherwise no safe assumptions can be made;
2152 * complain louder and bail out.
2154 for (i = 0; i != nd; ++i) {
2155 memset(&list[ns].info, 0, sizeof(list[ns].info));
2156 list[ns].bond_info = NULL;
2157 list[ns].max_port = 1;
2158 list[ns].phys_port = 1;
2159 list[ns].phys_dev = ibv_match[i];
2160 list[ns].eth_dev = NULL;
2161 list[ns].pci_dev = pci_dev;
2162 list[ns].pf_bond = -1;
2163 list[ns].ifindex = 0;
2165 list[ns].ifindex = mlx5_nl_ifindex
2167 mlx5_os_get_dev_device_name
2168 (list[ns].phys_dev), 1);
2169 if (!list[ns].ifindex) {
2170 char ifname[IF_NAMESIZE];
2173 * Netlink failed, it may happen with old
2174 * ib_core kernel driver (before 4.16).
2175 * We can assume there is old driver because
2176 * here we are processing single ports IB
2177 * devices. Let's try sysfs to retrieve
2178 * the ifindex. The method works for
2179 * master device only.
2183 * Multiple devices found, assume
2184 * representors, can not distinguish
2185 * master/representor and retrieve
2186 * ifindex via sysfs.
2190 ret = mlx5_get_ifname_sysfs
2191 (ibv_match[i]->ibdev_path, ifname);
2194 if_nametoindex(ifname);
2195 if (!list[ns].ifindex) {
2197 * No network interface index found
2198 * for the specified device, it means
2199 * there it is neither representor
2207 ret = mlx5_nl_switch_info
2211 if (ret || (!list[ns].info.representor &&
2212 !list[ns].info.master)) {
2214 * We failed to recognize representors with
2215 * Netlink, let's try to perform the task
2218 ret = mlx5_sysfs_switch_info
2222 if (!ret && (list[ns].info.representor ^
2223 list[ns].info.master)) {
2225 } else if ((nd == 1) &&
2226 !list[ns].info.representor &&
2227 !list[ns].info.master) {
2229 * Single IB device with
2230 * one physical port and
2231 * attached network device.
2232 * May be SRIOV is not enabled
2233 * or there is no representors.
2235 DRV_LOG(INFO, "no E-Switch support detected");
2242 "unable to recognize master/representors"
2243 " on the multiple IB devices");
2251 * Sort list to probe devices in natural order for users convenience
2252 * (i.e. master first, then representors from lowest to highest ID).
2254 qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2255 /* Device specific configuration. */
2256 switch (pci_dev->id.device_id) {
2257 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2258 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2259 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2260 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2261 case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2262 case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2263 case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2270 if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2271 /* Set devargs default values. */
2272 if (eth_da.nb_mh_controllers == 0) {
2273 eth_da.nb_mh_controllers = 1;
2274 eth_da.mh_controllers[0] = 0;
2276 if (eth_da.nb_ports == 0 && ns > 0) {
2277 if (list[0].pf_bond >= 0 && list[0].info.representor)
2278 DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2279 pci_dev->device.devargs->args);
2280 eth_da.nb_ports = 1;
2281 eth_da.ports[0] = list[0].info.pf_num;
2283 if (eth_da.nb_representor_ports == 0) {
2284 eth_da.nb_representor_ports = 1;
2285 eth_da.representor_ports[0] = 0;
2288 for (i = 0; i != ns; ++i) {
2291 /* Default configuration. */
2292 memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2293 dev_config.vf = dev_config_vf;
2294 dev_config.mps = MLX5_ARG_UNSET;
2295 dev_config.dbnc = MLX5_ARG_UNSET;
2296 dev_config.rx_vec_en = 1;
2297 dev_config.txq_inline_max = MLX5_ARG_UNSET;
2298 dev_config.txq_inline_min = MLX5_ARG_UNSET;
2299 dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2300 dev_config.txqs_inline = MLX5_ARG_UNSET;
2301 dev_config.vf_nl_en = 1;
2302 dev_config.mr_ext_memseg_en = 1;
2303 dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2304 dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2305 dev_config.dv_esw_en = 1;
2306 dev_config.dv_flow_en = 1;
2307 dev_config.decap_en = 1;
2308 dev_config.log_hp_size = MLX5_ARG_UNSET;
2309 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2313 if (!list[i].eth_dev) {
2314 if (rte_errno != EBUSY && rte_errno != EEXIST)
2316 /* Device is disabled or already spawned. Ignore it. */
2319 restore = list[i].eth_dev->data->dev_flags;
2320 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2321 /* Restore non-PCI flags cleared by the above call. */
2322 list[i].eth_dev->data->dev_flags |= restore;
2323 rte_eth_dev_probing_finish(list[i].eth_dev);
2327 "probe of PCI device " PCI_PRI_FMT " aborted after"
2328 " encountering an error: %s",
2329 owner_pci.domain, owner_pci.bus,
2330 owner_pci.devid, owner_pci.function,
2331 strerror(rte_errno));
2335 if (!list[i].eth_dev)
2337 mlx5_dev_close(list[i].eth_dev);
2338 /* mac_addrs must not be freed because in dev_private */
2339 list[i].eth_dev->data->mac_addrs = NULL;
2340 claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2342 /* Restore original error. */
2349 * Do the routine cleanup:
2350 * - close opened Netlink sockets
2351 * - free allocated spawn data array
2352 * - free the Infiniband device list
2360 MLX5_ASSERT(ibv_list);
2361 mlx5_glue->free_device_list(ibv_list);
2366 * DPDK callback to register a PCI device.
2368 * This function spawns Ethernet devices out of a given PCI device.
2370 * @param[in] pci_drv
2371 * PCI driver structure (mlx5_driver).
2372 * @param[in] pci_dev
2373 * PCI device information.
2376 * 0 on success, a negative errno value otherwise and rte_errno is set.
2379 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2380 struct rte_pci_device *pci_dev)
2382 struct rte_eth_devargs eth_da = { .type = RTE_ETH_REPRESENTOR_NONE };
2386 if (pci_dev->device.devargs) {
2387 /* Parse representor information from device argument. */
2388 if (pci_dev->device.devargs->cls_str)
2389 ret = rte_eth_devargs_parse
2390 (pci_dev->device.devargs->cls_str, ð_da);
2392 DRV_LOG(ERR, "failed to parse device arguments: %s",
2393 pci_dev->device.devargs->cls_str);
2396 if (eth_da.type == RTE_ETH_REPRESENTOR_NONE) {
2397 /* Support legacy device argument */
2398 ret = rte_eth_devargs_parse
2399 (pci_dev->device.devargs->args, ð_da);
2401 DRV_LOG(ERR, "failed to parse device arguments: %s",
2402 pci_dev->device.devargs->args);
2408 if (eth_da.nb_ports > 0) {
2409 /* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2410 for (p = 0; p < eth_da.nb_ports; p++)
2411 ret = mlx5_os_pci_probe_pf(pci_dev, ð_da,
2414 ret = mlx5_os_pci_probe_pf(pci_dev, ð_da, 0);
2420 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2425 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2426 /* Get environment variable to store. */
2427 env = getenv(MLX5_SHUT_UP_BF);
2428 value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2429 if (config->dbnc == MLX5_ARG_UNSET)
2430 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2432 setenv(MLX5_SHUT_UP_BF,
2433 config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2438 mlx5_restore_doorbell_mapping_env(int value)
2440 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2441 /* Restore the original environment variable state. */
2442 if (value == MLX5_ARG_UNSET)
2443 unsetenv(MLX5_SHUT_UP_BF);
2445 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2449 * Extract pdn of PD object using DV API.
2452 * Pointer to the verbs PD object.
2454 * Pointer to the PD object number variable.
2457 * 0 on success, error value otherwise.
2460 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2462 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2463 struct mlx5dv_obj obj;
2464 struct mlx5dv_pd pd_info;
2468 obj.pd.out = &pd_info;
2469 ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2471 DRV_LOG(DEBUG, "Fail to get PD object info");
2480 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2484 * Function API to open IB device.
2486 * This function calls the Linux glue APIs to open a device.
2489 * Pointer to the IB device attributes (name, port, etc).
2490 * @param[out] config
2491 * Pointer to device configuration structure.
2493 * Pointer to shared context structure.
2496 * 0 on success, a positive error value otherwise.
2499 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2500 const struct mlx5_dev_config *config,
2501 struct mlx5_dev_ctx_shared *sh)
2506 sh->numa_node = spawn->pci_dev->device.numa_node;
2507 pthread_mutex_init(&sh->txpp.mutex, NULL);
2509 * Configure environment variable "MLX5_BF_SHUT_UP"
2510 * before the device creation. The rdma_core library
2511 * checks the variable at device creation and
2512 * stores the result internally.
2514 dbmap_env = mlx5_config_doorbell_mapping_env(config);
2515 /* Try to open IB device with DV first, then usual Verbs. */
2517 sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2520 DRV_LOG(DEBUG, "DevX is supported");
2521 /* The device is created, no need for environment. */
2522 mlx5_restore_doorbell_mapping_env(dbmap_env);
2524 /* The environment variable is still configured. */
2525 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2526 err = errno ? errno : ENODEV;
2528 * The environment variable is not needed anymore,
2529 * all device creation attempts are completed.
2531 mlx5_restore_doorbell_mapping_env(dbmap_env);
2534 DRV_LOG(DEBUG, "DevX is NOT supported");
2537 if (!err && sh->ctx) {
2538 /* Hint libmlx5 to use PMD allocator for data plane resources */
2539 mlx5_glue->dv_set_context_attr(sh->ctx,
2540 MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2541 (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2542 .alloc = &mlx5_alloc_verbs_buf,
2543 .free = &mlx5_free_verbs_buf,
2551 * Install shared asynchronous device events handler.
2552 * This function is implemented to support event sharing
2553 * between multiple ports of single IB device.
2556 * Pointer to mlx5_dev_ctx_shared object.
2559 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2564 sh->intr_handle.fd = -1;
2565 flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2566 ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2567 F_SETFL, flags | O_NONBLOCK);
2569 DRV_LOG(INFO, "failed to change file descriptor async event"
2572 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2573 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2574 if (rte_intr_callback_register(&sh->intr_handle,
2575 mlx5_dev_interrupt_handler, sh)) {
2576 DRV_LOG(INFO, "Fail to install the shared interrupt.");
2577 sh->intr_handle.fd = -1;
2581 #ifdef HAVE_IBV_DEVX_ASYNC
2582 sh->intr_handle_devx.fd = -1;
2584 (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2585 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2587 DRV_LOG(INFO, "failed to allocate devx_comp.");
2590 flags = fcntl(devx_comp->fd, F_GETFL);
2591 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2593 DRV_LOG(INFO, "failed to change file descriptor"
2597 sh->intr_handle_devx.fd = devx_comp->fd;
2598 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2599 if (rte_intr_callback_register(&sh->intr_handle_devx,
2600 mlx5_dev_interrupt_handler_devx, sh)) {
2601 DRV_LOG(INFO, "Fail to install the devx shared"
2603 sh->intr_handle_devx.fd = -1;
2605 #endif /* HAVE_IBV_DEVX_ASYNC */
2610 * Uninstall shared asynchronous device events handler.
2611 * This function is implemented to support event sharing
2612 * between multiple ports of single IB device.
2615 * Pointer to mlx5_dev_ctx_shared object.
2618 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2620 if (sh->intr_handle.fd >= 0)
2621 mlx5_intr_callback_unregister(&sh->intr_handle,
2622 mlx5_dev_interrupt_handler, sh);
2623 #ifdef HAVE_IBV_DEVX_ASYNC
2624 if (sh->intr_handle_devx.fd >= 0)
2625 rte_intr_callback_unregister(&sh->intr_handle_devx,
2626 mlx5_dev_interrupt_handler_devx, sh);
2628 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2633 * Read statistics by a named counter.
2636 * Pointer to the private device data structure.
2637 * @param[in] ctr_name
2638 * Pointer to the name of the statistic counter to read
2640 * Pointer to read statistic value.
2642 * 0 on success and stat is valud, 1 if failed to read the value
2647 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2653 if (priv->q_counters != NULL &&
2654 strcmp(ctr_name, "out_of_buffer") == 0)
2655 return mlx5_devx_cmd_queue_counter_query(priv->sh->ctx,
2656 0, (uint32_t *)stat);
2657 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2658 priv->sh->ibdev_path,
2661 fd = open(path, O_RDONLY);
2663 * in switchdev the file location is not per port
2664 * but rather in <ibdev_path>/hw_counters/<file_name>.
2667 MKSTR(path1, "%s/hw_counters/%s",
2668 priv->sh->ibdev_path,
2670 fd = open(path1, O_RDONLY);
2673 char buf[21] = {'\0'};
2674 ssize_t n = read(fd, buf, sizeof(buf));
2678 *stat = strtoull(buf, NULL, 10);
2688 * Set the reg_mr and dereg_mr call backs
2690 * @param reg_mr_cb[out]
2691 * Pointer to reg_mr func
2692 * @param dereg_mr_cb[out]
2693 * Pointer to dereg_mr func
2697 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2698 mlx5_dereg_mr_t *dereg_mr_cb)
2700 *reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2701 *dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2705 * Remove a MAC address from device
2708 * Pointer to Ethernet device structure.
2710 * MAC address index.
2713 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2715 struct mlx5_priv *priv = dev->data->dev_private;
2716 const int vf = priv->config.vf;
2719 mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2720 mlx5_ifindex(dev), priv->mac_own,
2721 &dev->data->mac_addrs[index], index);
2725 * Adds a MAC address to the device
2728 * Pointer to Ethernet device structure.
2730 * MAC address to register.
2732 * MAC address index.
2735 * 0 on success, a negative errno value otherwise
2738 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2741 struct mlx5_priv *priv = dev->data->dev_private;
2742 const int vf = priv->config.vf;
2746 ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2747 mlx5_ifindex(dev), priv->mac_own,
2753 * Modify a VF MAC address
2756 * Pointer to device private data.
2758 * MAC address to modify into.
2760 * Net device interface index
2765 * 0 on success, a negative errno value otherwise
2768 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2769 unsigned int iface_idx,
2770 struct rte_ether_addr *mac_addr,
2773 return mlx5_nl_vf_mac_addr_modify
2774 (priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2778 * Set device promiscuous mode
2781 * Pointer to Ethernet device structure.
2783 * 0 - promiscuous is disabled, otherwise - enabled
2786 * 0 on success, a negative error value otherwise
2789 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2791 struct mlx5_priv *priv = dev->data->dev_private;
2793 return mlx5_nl_promisc(priv->nl_socket_route,
2794 mlx5_ifindex(dev), !!enable);
2798 * Set device promiscuous mode
2801 * Pointer to Ethernet device structure.
2803 * 0 - all multicase is disabled, otherwise - enabled
2806 * 0 on success, a negative error value otherwise
2809 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2811 struct mlx5_priv *priv = dev->data->dev_private;
2813 return mlx5_nl_allmulti(priv->nl_socket_route,
2814 mlx5_ifindex(dev), !!enable);
2818 * Flush device MAC addresses
2821 * Pointer to Ethernet device structure.
2825 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2827 struct mlx5_priv *priv = dev->data->dev_private;
2829 mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2830 dev->data->mac_addrs,
2831 MLX5_MAX_MAC_ADDRESSES, priv->mac_own);