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 <rte_ethdev_driver.h>
20 #include <rte_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_qp = attr_ex.orig_attr.max_qp;
120 device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
121 device_attr->max_rwq_indirection_table_size =
122 attr_ex.rss_caps.max_rwq_indirection_table_size;
123 device_attr->max_tso = attr_ex.tso_caps.max_tso;
124 device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
126 struct mlx5dv_context dv_attr = { .comp_mask = 0 };
127 err = mlx5_glue->dv_query_device(ctx, &dv_attr);
131 device_attr->flags = dv_attr.flags;
132 device_attr->comp_mask = dv_attr.comp_mask;
133 #ifdef HAVE_IBV_MLX5_MOD_SWP
134 device_attr->sw_parsing_offloads =
135 dv_attr.sw_parsing_caps.sw_parsing_offloads;
137 device_attr->min_single_stride_log_num_of_bytes =
138 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
139 device_attr->max_single_stride_log_num_of_bytes =
140 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
141 device_attr->min_single_wqe_log_num_of_strides =
142 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
143 device_attr->max_single_wqe_log_num_of_strides =
144 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
145 device_attr->stride_supported_qpts =
146 dv_attr.striding_rq_caps.supported_qpts;
147 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
148 device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
155 * Verbs callback to allocate a memory. This function should allocate the space
156 * according to the size provided residing inside a huge page.
157 * Please note that all allocation must respect the alignment from libmlx5
158 * (i.e. currently rte_mem_page_size()).
161 * The size in bytes of the memory to allocate.
163 * A pointer to the callback data.
166 * Allocated buffer, NULL otherwise and rte_errno is set.
169 mlx5_alloc_verbs_buf(size_t size, void *data)
171 struct mlx5_priv *priv = data;
173 unsigned int socket = SOCKET_ID_ANY;
174 size_t alignment = rte_mem_page_size();
175 if (alignment == (size_t)-1) {
176 DRV_LOG(ERR, "Failed to get mem page size");
181 if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
182 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
184 socket = ctrl->socket;
185 } else if (priv->verbs_alloc_ctx.type ==
186 MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
187 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
189 socket = ctrl->socket;
191 MLX5_ASSERT(data != NULL);
192 ret = mlx5_malloc(0, size, alignment, socket);
199 * Verbs callback to free a memory.
202 * A pointer to the memory to free.
204 * A pointer to the callback data.
207 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
209 MLX5_ASSERT(data != NULL);
214 * Initialize DR related data within private structure.
215 * Routine checks the reference counter and does actual
216 * resources creation/initialization only if counter is zero.
219 * Pointer to the private device data structure.
222 * Zero on success, positive error code otherwise.
225 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
227 struct mlx5_dev_ctx_shared *sh = priv->sh;
228 char s[MLX5_HLIST_NAMESIZE];
231 MLX5_ASSERT(sh && sh->refcnt);
234 err = mlx5_alloc_table_hash_list(priv);
237 /* Create tags hash list table. */
238 snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
239 sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
240 0, NULL, NULL, NULL);
241 if (!sh->tag_table) {
242 DRV_LOG(ERR, "tags with hash creation failed.");
246 snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
247 sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
248 0, 0, NULL, NULL, NULL);
249 if (!sh->modify_cmds) {
250 DRV_LOG(ERR, "hdr modify hash creation failed");
254 snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
255 sh->encaps_decaps = mlx5_hlist_create(s,
256 MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
257 0, 0, NULL, NULL, NULL);
258 if (!sh->encaps_decaps) {
259 DRV_LOG(ERR, "encap decap hash creation failed");
263 #ifdef HAVE_MLX5DV_DR
266 /* Reference counter is zero, we should initialize structures. */
267 domain = mlx5_glue->dr_create_domain(sh->ctx,
268 MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
270 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
274 sh->rx_domain = domain;
275 domain = mlx5_glue->dr_create_domain(sh->ctx,
276 MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
278 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
282 pthread_mutex_init(&sh->dv_mutex, NULL);
283 sh->tx_domain = domain;
284 #ifdef HAVE_MLX5DV_DR_ESWITCH
285 if (priv->config.dv_esw_en) {
286 domain = mlx5_glue->dr_create_domain
287 (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
289 DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
293 sh->fdb_domain = domain;
294 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
298 err = mlx5_alloc_tunnel_hub(sh);
300 DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
303 if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
304 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
305 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
307 mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
309 sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
310 #endif /* HAVE_MLX5DV_DR */
311 sh->default_miss_action =
312 mlx5_glue->dr_create_flow_action_default_miss();
313 if (!sh->default_miss_action)
314 DRV_LOG(WARNING, "Default miss action is not supported.");
317 /* Rollback the created objects. */
319 mlx5_glue->dr_destroy_domain(sh->rx_domain);
320 sh->rx_domain = NULL;
323 mlx5_glue->dr_destroy_domain(sh->tx_domain);
324 sh->tx_domain = NULL;
326 if (sh->fdb_domain) {
327 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
328 sh->fdb_domain = NULL;
330 if (sh->esw_drop_action) {
331 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
332 sh->esw_drop_action = NULL;
334 if (sh->pop_vlan_action) {
335 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
336 sh->pop_vlan_action = NULL;
338 if (sh->encaps_decaps) {
339 mlx5_hlist_destroy(sh->encaps_decaps);
340 sh->encaps_decaps = NULL;
342 if (sh->modify_cmds) {
343 mlx5_hlist_destroy(sh->modify_cmds);
344 sh->modify_cmds = NULL;
347 /* tags should be destroyed with flow before. */
348 mlx5_hlist_destroy(sh->tag_table);
349 sh->tag_table = NULL;
351 if (sh->tunnel_hub) {
352 mlx5_release_tunnel_hub(sh, priv->dev_port);
353 sh->tunnel_hub = NULL;
355 mlx5_free_table_hash_list(priv);
360 * Destroy DR related data within private structure.
363 * Pointer to the private device data structure.
366 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
368 struct mlx5_dev_ctx_shared *sh = priv->sh;
370 MLX5_ASSERT(sh && sh->refcnt);
373 #ifdef HAVE_MLX5DV_DR
375 mlx5_glue->dr_destroy_domain(sh->rx_domain);
376 sh->rx_domain = NULL;
379 mlx5_glue->dr_destroy_domain(sh->tx_domain);
380 sh->tx_domain = NULL;
382 #ifdef HAVE_MLX5DV_DR_ESWITCH
383 if (sh->fdb_domain) {
384 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
385 sh->fdb_domain = NULL;
387 if (sh->esw_drop_action) {
388 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
389 sh->esw_drop_action = NULL;
392 if (sh->pop_vlan_action) {
393 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
394 sh->pop_vlan_action = NULL;
396 pthread_mutex_destroy(&sh->dv_mutex);
397 #endif /* HAVE_MLX5DV_DR */
398 if (sh->default_miss_action)
399 mlx5_glue->destroy_flow_action
400 (sh->default_miss_action);
401 if (sh->encaps_decaps) {
402 mlx5_hlist_destroy(sh->encaps_decaps);
403 sh->encaps_decaps = NULL;
405 if (sh->modify_cmds) {
406 mlx5_hlist_destroy(sh->modify_cmds);
407 sh->modify_cmds = NULL;
410 /* tags should be destroyed with flow before. */
411 mlx5_hlist_destroy(sh->tag_table);
412 sh->tag_table = NULL;
414 if (sh->tunnel_hub) {
415 mlx5_release_tunnel_hub(sh, priv->dev_port);
416 sh->tunnel_hub = NULL;
418 mlx5_free_table_hash_list(priv);
422 * Initialize shared data between primary and secondary process.
424 * A memzone is reserved by primary process and secondary processes attach to
428 * 0 on success, a negative errno value otherwise and rte_errno is set.
431 mlx5_init_shared_data(void)
433 const struct rte_memzone *mz;
436 rte_spinlock_lock(&mlx5_shared_data_lock);
437 if (mlx5_shared_data == NULL) {
438 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
439 /* Allocate shared memory. */
440 mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
441 sizeof(*mlx5_shared_data),
445 "Cannot allocate mlx5 shared data");
449 mlx5_shared_data = mz->addr;
450 memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
451 rte_spinlock_init(&mlx5_shared_data->lock);
453 /* Lookup allocated shared memory. */
454 mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
457 "Cannot attach mlx5 shared data");
461 mlx5_shared_data = mz->addr;
462 memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
466 rte_spinlock_unlock(&mlx5_shared_data_lock);
471 * PMD global initialization.
473 * Independent from individual device, this function initializes global
474 * per-PMD data structures distinguishing primary and secondary processes.
475 * Hence, each initialization is called once per a process.
478 * 0 on success, a negative errno value otherwise and rte_errno is set.
483 struct mlx5_shared_data *sd;
484 struct mlx5_local_data *ld = &mlx5_local_data;
487 if (mlx5_init_shared_data())
489 sd = mlx5_shared_data;
491 rte_spinlock_lock(&sd->lock);
492 switch (rte_eal_process_type()) {
493 case RTE_PROC_PRIMARY:
496 LIST_INIT(&sd->mem_event_cb_list);
497 rte_rwlock_init(&sd->mem_event_rwlock);
498 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
499 mlx5_mr_mem_event_cb, NULL);
500 ret = mlx5_mp_init_primary(MLX5_MP_NAME,
501 mlx5_mp_os_primary_handle);
504 sd->init_done = true;
506 case RTE_PROC_SECONDARY:
509 ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
510 mlx5_mp_os_secondary_handle);
514 ld->init_done = true;
520 rte_spinlock_unlock(&sd->lock);
525 * Create the Tx queue DevX/Verbs object.
528 * Pointer to Ethernet device.
530 * Queue index in DPDK Tx queue array.
533 * 0 on success, a negative errno value otherwise and rte_errno is set.
536 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
538 struct mlx5_priv *priv = dev->data->dev_private;
539 struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
540 struct mlx5_txq_ctrl *txq_ctrl =
541 container_of(txq_data, struct mlx5_txq_ctrl, txq);
543 if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
544 return mlx5_txq_devx_obj_new(dev, idx);
545 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
546 if (!priv->config.dv_esw_en)
547 return mlx5_txq_devx_obj_new(dev, idx);
549 return mlx5_txq_ibv_obj_new(dev, idx);
553 * Release an Tx DevX/verbs queue object.
556 * DevX/Verbs Tx queue object.
559 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
561 if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
562 mlx5_txq_devx_obj_release(txq_obj);
565 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
566 if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
567 mlx5_txq_devx_obj_release(txq_obj);
571 mlx5_txq_ibv_obj_release(txq_obj);
575 * DV flow counter mode detect and config.
578 * Pointer to rte_eth_dev structure.
582 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
584 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
585 struct mlx5_priv *priv = dev->data->dev_private;
586 struct mlx5_dev_ctx_shared *sh = priv->sh;
589 #ifndef HAVE_IBV_DEVX_ASYNC
593 if (!priv->config.devx || !priv->config.dv_flow_en ||
594 !priv->config.hca_attr.flow_counters_dump ||
595 !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
596 (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
600 DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
601 "counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
602 priv->config.hca_attr.flow_counters_dump,
603 priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
604 /* Initialize fallback mode only on the port initializes sh. */
606 sh->cmng.counter_fallback = fallback;
607 else if (fallback != sh->cmng.counter_fallback)
608 DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
609 "with others:%d.", PORT_ID(priv), fallback);
614 * Spawn an Ethernet device from Verbs information.
617 * Backing DPDK device.
619 * Verbs device parameters (name, port, switch_info) to spawn.
621 * Device configuration parameters.
624 * A valid Ethernet device object on success, NULL otherwise and rte_errno
625 * is set. The following errors are defined:
627 * EBUSY: device is not supposed to be spawned.
628 * EEXIST: device is already spawned
630 static struct rte_eth_dev *
631 mlx5_dev_spawn(struct rte_device *dpdk_dev,
632 struct mlx5_dev_spawn_data *spawn,
633 struct mlx5_dev_config *config)
635 const struct mlx5_switch_info *switch_info = &spawn->info;
636 struct mlx5_dev_ctx_shared *sh = NULL;
637 struct ibv_port_attr port_attr;
638 struct mlx5dv_context dv_attr = { .comp_mask = 0 };
639 struct rte_eth_dev *eth_dev = NULL;
640 struct mlx5_priv *priv = NULL;
642 unsigned int hw_padding = 0;
644 unsigned int cqe_comp;
645 unsigned int cqe_pad = 0;
646 unsigned int tunnel_en = 0;
647 unsigned int mpls_en = 0;
648 unsigned int swp = 0;
649 unsigned int mprq = 0;
650 unsigned int mprq_min_stride_size_n = 0;
651 unsigned int mprq_max_stride_size_n = 0;
652 unsigned int mprq_min_stride_num_n = 0;
653 unsigned int mprq_max_stride_num_n = 0;
654 struct rte_ether_addr mac;
655 char name[RTE_ETH_NAME_MAX_LEN];
656 int own_domain_id = 0;
659 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
660 struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
663 /* Determine if this port representor is supposed to be spawned. */
664 if (switch_info->representor && dpdk_dev->devargs) {
665 struct rte_eth_devargs eth_da;
667 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, ð_da);
670 DRV_LOG(ERR, "failed to process device arguments: %s",
671 strerror(rte_errno));
674 for (i = 0; i < eth_da.nb_representor_ports; ++i)
675 if (eth_da.representor_ports[i] ==
676 (uint16_t)switch_info->port_name)
678 if (i == eth_da.nb_representor_ports) {
683 /* Build device name. */
684 if (spawn->pf_bond < 0) {
686 if (!switch_info->representor)
687 strlcpy(name, dpdk_dev->name, sizeof(name));
689 snprintf(name, sizeof(name), "%s_representor_%u",
690 dpdk_dev->name, switch_info->port_name);
692 /* Bonding device. */
693 if (!switch_info->representor)
694 snprintf(name, sizeof(name), "%s_%s",
696 mlx5_os_get_dev_device_name(spawn->phys_dev));
698 snprintf(name, sizeof(name), "%s_%s_representor_%u",
700 mlx5_os_get_dev_device_name(spawn->phys_dev),
701 switch_info->port_name);
703 /* check if the device is already spawned */
704 if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
708 DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
709 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
710 struct mlx5_mp_id mp_id;
712 eth_dev = rte_eth_dev_attach_secondary(name);
713 if (eth_dev == NULL) {
714 DRV_LOG(ERR, "can not attach rte ethdev");
718 eth_dev->device = dpdk_dev;
719 eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
720 eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
721 eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
722 err = mlx5_proc_priv_init(eth_dev);
725 mp_id.port_id = eth_dev->data->port_id;
726 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
727 /* Receive command fd from primary process */
728 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
731 /* Remap UAR for Tx queues. */
732 err = mlx5_tx_uar_init_secondary(eth_dev, err);
736 * Ethdev pointer is still required as input since
737 * the primary device is not accessible from the
740 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
741 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
744 mlx5_dev_close(eth_dev);
748 * Some parameters ("tx_db_nc" in particularly) are needed in
749 * advance to create dv/verbs device context. We proceed the
750 * devargs here to get ones, and later proceed devargs again
751 * to override some hardware settings.
753 err = mlx5_args(config, dpdk_dev->devargs);
756 DRV_LOG(ERR, "failed to process device arguments: %s",
757 strerror(rte_errno));
760 if (config->dv_miss_info) {
761 if (switch_info->master || switch_info->representor)
762 config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
764 mlx5_malloc_mem_select(config->sys_mem_en);
765 sh = mlx5_alloc_shared_dev_ctx(spawn, config);
768 config->devx = sh->devx;
769 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
770 config->dest_tir = 1;
772 #ifdef HAVE_IBV_MLX5_MOD_SWP
773 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
776 * Multi-packet send is supported by ConnectX-4 Lx PF as well
777 * as all ConnectX-5 devices.
779 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
780 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
782 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
783 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
785 mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
786 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
787 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
788 DRV_LOG(DEBUG, "enhanced MPW is supported");
789 mps = MLX5_MPW_ENHANCED;
791 DRV_LOG(DEBUG, "MPW is supported");
795 DRV_LOG(DEBUG, "MPW isn't supported");
796 mps = MLX5_MPW_DISABLED;
798 #ifdef HAVE_IBV_MLX5_MOD_SWP
799 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
800 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
801 DRV_LOG(DEBUG, "SWP support: %u", swp);
804 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
805 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
806 struct mlx5dv_striding_rq_caps mprq_caps =
807 dv_attr.striding_rq_caps;
809 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
810 mprq_caps.min_single_stride_log_num_of_bytes);
811 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
812 mprq_caps.max_single_stride_log_num_of_bytes);
813 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
814 mprq_caps.min_single_wqe_log_num_of_strides);
815 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
816 mprq_caps.max_single_wqe_log_num_of_strides);
817 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
818 mprq_caps.supported_qpts);
819 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
821 mprq_min_stride_size_n =
822 mprq_caps.min_single_stride_log_num_of_bytes;
823 mprq_max_stride_size_n =
824 mprq_caps.max_single_stride_log_num_of_bytes;
825 mprq_min_stride_num_n =
826 mprq_caps.min_single_wqe_log_num_of_strides;
827 mprq_max_stride_num_n =
828 mprq_caps.max_single_wqe_log_num_of_strides;
831 if (RTE_CACHE_LINE_SIZE == 128 &&
832 !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
836 config->cqe_comp = cqe_comp;
837 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
838 /* Whether device supports 128B Rx CQE padding. */
839 cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
840 (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
842 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
843 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
844 tunnel_en = ((dv_attr.tunnel_offloads_caps &
845 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
846 (dv_attr.tunnel_offloads_caps &
847 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
848 (dv_attr.tunnel_offloads_caps &
849 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
851 DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
852 tunnel_en ? "" : "not ");
855 "tunnel offloading disabled due to old OFED/rdma-core version");
857 config->tunnel_en = tunnel_en;
858 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
859 mpls_en = ((dv_attr.tunnel_offloads_caps &
860 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
861 (dv_attr.tunnel_offloads_caps &
862 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
863 DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
864 mpls_en ? "" : "not ");
866 DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
867 " old OFED/rdma-core version or firmware configuration");
869 config->mpls_en = mpls_en;
870 /* Check port status. */
871 err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
873 DRV_LOG(ERR, "port query failed: %s", strerror(err));
876 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
877 DRV_LOG(ERR, "port is not configured in Ethernet mode");
881 if (port_attr.state != IBV_PORT_ACTIVE)
882 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
883 mlx5_glue->port_state_str(port_attr.state),
885 /* Allocate private eth device data. */
886 priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
888 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
890 DRV_LOG(ERR, "priv allocation failure");
895 priv->dev_port = spawn->phys_port;
896 priv->pci_dev = spawn->pci_dev;
897 priv->mtu = RTE_ETHER_MTU;
898 priv->mp_id.port_id = port_id;
899 strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
900 /* Some internal functions rely on Netlink sockets, open them now. */
901 priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
902 priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
903 priv->representor = !!switch_info->representor;
904 priv->master = !!switch_info->master;
905 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
906 priv->vport_meta_tag = 0;
907 priv->vport_meta_mask = 0;
908 priv->pf_bond = spawn->pf_bond;
909 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
911 * The DevX port query API is implemented. E-Switch may use
912 * either vport or reg_c[0] metadata register to match on
913 * vport index. The engaged part of metadata register is
916 if (switch_info->representor || switch_info->master) {
917 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
918 MLX5DV_DEVX_PORT_MATCH_REG_C_0;
919 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
923 "can't query devx port %d on device %s",
925 mlx5_os_get_dev_device_name(spawn->phys_dev));
926 devx_port.comp_mask = 0;
929 if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
930 priv->vport_meta_tag = devx_port.reg_c_0.value;
931 priv->vport_meta_mask = devx_port.reg_c_0.mask;
932 if (!priv->vport_meta_mask) {
933 DRV_LOG(ERR, "vport zero mask for port %d"
934 " on bonding device %s",
936 mlx5_os_get_dev_device_name
941 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
942 DRV_LOG(ERR, "invalid vport tag for port %d"
943 " on bonding device %s",
945 mlx5_os_get_dev_device_name
951 if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
952 priv->vport_id = devx_port.vport_num;
953 } else if (spawn->pf_bond >= 0) {
954 DRV_LOG(ERR, "can't deduce vport index for port %d"
955 " on bonding device %s",
957 mlx5_os_get_dev_device_name(spawn->phys_dev));
961 /* Suppose vport index in compatible way. */
962 priv->vport_id = switch_info->representor ?
963 switch_info->port_name + 1 : -1;
967 * Kernel/rdma_core support single E-Switch per PF configurations
968 * only and vport_id field contains the vport index for
969 * associated VF, which is deduced from representor port name.
970 * For example, let's have the IB device port 10, it has
971 * attached network device eth0, which has port name attribute
972 * pf0vf2, we can deduce the VF number as 2, and set vport index
973 * as 3 (2+1). This assigning schema should be changed if the
974 * multiple E-Switch instances per PF configurations or/and PCI
975 * subfunctions are added.
977 priv->vport_id = switch_info->representor ?
978 switch_info->port_name + 1 : -1;
980 /* representor_id field keeps the unmodified VF index. */
981 priv->representor_id = switch_info->representor ?
982 switch_info->port_name : -1;
984 * Look for sibling devices in order to reuse their switch domain
985 * if any, otherwise allocate one.
987 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
988 const struct mlx5_priv *opriv =
989 rte_eth_devices[port_id].data->dev_private;
992 opriv->sh != priv->sh ||
994 RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
996 priv->domain_id = opriv->domain_id;
999 if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1000 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1003 DRV_LOG(ERR, "unable to allocate switch domain: %s",
1004 strerror(rte_errno));
1009 /* Override some values set by hardware configuration. */
1010 mlx5_args(config, dpdk_dev->devargs);
1011 err = mlx5_dev_check_sibling_config(priv, config);
1014 config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1015 IBV_DEVICE_RAW_IP_CSUM);
1016 DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1017 (config->hw_csum ? "" : "not "));
1018 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1019 !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1020 DRV_LOG(DEBUG, "counters are not supported");
1022 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1023 if (config->dv_flow_en) {
1024 DRV_LOG(WARNING, "DV flow is not supported");
1025 config->dv_flow_en = 0;
1028 config->ind_table_max_size =
1029 sh->device_attr.max_rwq_indirection_table_size;
1031 * Remove this check once DPDK supports larger/variable
1032 * indirection tables.
1034 if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1035 config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1036 DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1037 config->ind_table_max_size);
1038 config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1039 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1040 DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1041 (config->hw_vlan_strip ? "" : "not "));
1042 config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1043 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1044 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1045 hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1046 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1047 hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1048 IBV_DEVICE_PCI_WRITE_END_PADDING);
1050 if (config->hw_padding && !hw_padding) {
1051 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1052 config->hw_padding = 0;
1053 } else if (config->hw_padding) {
1054 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1056 config->tso = (sh->device_attr.max_tso > 0 &&
1057 (sh->device_attr.tso_supported_qpts &
1058 (1 << IBV_QPT_RAW_PACKET)));
1060 config->tso_max_payload_sz = sh->device_attr.max_tso;
1062 * MPW is disabled by default, while the Enhanced MPW is enabled
1065 if (config->mps == MLX5_ARG_UNSET)
1066 config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1069 config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1070 DRV_LOG(INFO, "%sMPS is %s",
1071 config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1072 config->mps == MLX5_MPW ? "legacy " : "",
1073 config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1074 if (config->cqe_comp && !cqe_comp) {
1075 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1076 config->cqe_comp = 0;
1078 if (config->cqe_pad && !cqe_pad) {
1079 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1080 config->cqe_pad = 0;
1081 } else if (config->cqe_pad) {
1082 DRV_LOG(INFO, "Rx CQE padding is enabled");
1085 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1090 /* Check relax ordering support. */
1091 if (config->hca_attr.relaxed_ordering_write &&
1092 config->hca_attr.relaxed_ordering_read &&
1093 !haswell_broadwell_cpu)
1094 sh->cmng.relaxed_ordering = 1;
1095 /* Check for LRO support. */
1096 if (config->dest_tir && config->hca_attr.lro_cap &&
1097 config->dv_flow_en) {
1098 /* TBD check tunnel lro caps. */
1099 config->lro.supported = config->hca_attr.lro_cap;
1100 DRV_LOG(DEBUG, "Device supports LRO");
1102 * If LRO timeout is not configured by application,
1103 * use the minimal supported value.
1105 if (!config->lro.timeout)
1106 config->lro.timeout =
1107 config->hca_attr.lro_timer_supported_periods[0];
1108 DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1109 config->lro.timeout);
1110 DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1111 "required for coalescing is %d bytes",
1112 config->hca_attr.lro_min_mss_size);
1114 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1115 if (config->hca_attr.qos.sup &&
1116 config->hca_attr.qos.srtcm_sup &&
1117 config->dv_flow_en) {
1118 uint8_t reg_c_mask =
1119 config->hca_attr.qos.flow_meter_reg_c_ids;
1121 * Meter needs two REG_C's for color match and pre-sfx
1122 * flow match. Here get the REG_C for color match.
1123 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1126 if (__builtin_popcount(reg_c_mask) < 1) {
1128 DRV_LOG(WARNING, "No available register for"
1131 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
1134 priv->mtr_reg_share =
1135 config->hca_attr.qos.flow_meter_reg_share;
1136 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1137 priv->mtr_color_reg);
1141 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1142 if (config->hca_attr.log_max_ft_sampler_num > 0 &&
1143 config->dv_flow_en) {
1144 priv->sampler_en = 1;
1145 DRV_LOG(DEBUG, "The Sampler enabled!\n");
1147 priv->sampler_en = 0;
1148 if (!config->hca_attr.log_max_ft_sampler_num)
1149 DRV_LOG(WARNING, "No available register for"
1152 DRV_LOG(DEBUG, "DV flow is not supported!\n");
1156 if (config->tx_pp) {
1157 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1158 config->hca_attr.dev_freq_khz);
1159 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1160 config->hca_attr.qos.packet_pacing ? "" : "not ");
1161 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1162 config->hca_attr.cross_channel ? "" : "not ");
1163 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1164 config->hca_attr.wqe_index_ignore ? "" : "not ");
1165 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1166 config->hca_attr.non_wire_sq ? "" : "not ");
1167 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1168 config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1169 config->hca_attr.log_max_static_sq_wq);
1170 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1171 config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1172 if (!config->devx) {
1173 DRV_LOG(ERR, "DevX is required for packet pacing");
1177 if (!config->hca_attr.qos.packet_pacing) {
1178 DRV_LOG(ERR, "Packet pacing is not supported");
1182 if (!config->hca_attr.cross_channel) {
1183 DRV_LOG(ERR, "Cross channel operations are"
1184 " required for packet pacing");
1188 if (!config->hca_attr.wqe_index_ignore) {
1189 DRV_LOG(ERR, "WQE index ignore feature is"
1190 " required for packet pacing");
1194 if (!config->hca_attr.non_wire_sq) {
1195 DRV_LOG(ERR, "Non-wire SQ feature is"
1196 " required for packet pacing");
1200 if (!config->hca_attr.log_max_static_sq_wq) {
1201 DRV_LOG(ERR, "Static WQE SQ feature is"
1202 " required for packet pacing");
1206 if (!config->hca_attr.qos.wqe_rate_pp) {
1207 DRV_LOG(ERR, "WQE rate mode is required"
1208 " for packet pacing");
1212 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1213 DRV_LOG(ERR, "DevX does not provide UAR offset,"
1214 " can't create queues for packet pacing");
1220 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1222 err = config->hca_attr.access_register_user ?
1223 mlx5_devx_cmd_register_read
1224 (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1225 reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1229 /* MTUTC register is read successfully. */
1230 ts_mode = MLX5_GET(register_mtutc, reg,
1232 if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1233 config->rt_timestamp = 1;
1235 /* Kernel does not support register reading. */
1236 if (config->hca_attr.dev_freq_khz ==
1237 (NS_PER_S / MS_PER_S))
1238 config->rt_timestamp = 1;
1242 * If HW has bug working with tunnel packet decapsulation and
1243 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1244 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1246 if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1247 config->hw_fcs_strip = 0;
1248 DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1249 (config->hw_fcs_strip ? "" : "not "));
1250 if (config->mprq.enabled && mprq) {
1251 if (config->mprq.stride_num_n &&
1252 (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1253 config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1254 config->mprq.stride_num_n =
1255 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1256 mprq_min_stride_num_n),
1257 mprq_max_stride_num_n);
1259 "the number of strides"
1260 " for Multi-Packet RQ is out of range,"
1261 " setting default value (%u)",
1262 1 << config->mprq.stride_num_n);
1264 if (config->mprq.stride_size_n &&
1265 (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1266 config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1267 config->mprq.stride_size_n =
1268 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1269 mprq_min_stride_size_n),
1270 mprq_max_stride_size_n);
1272 "the size of a stride"
1273 " for Multi-Packet RQ is out of range,"
1274 " setting default value (%u)",
1275 1 << config->mprq.stride_size_n);
1277 config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1278 config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1279 } else if (config->mprq.enabled && !mprq) {
1280 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1281 config->mprq.enabled = 0;
1283 if (config->max_dump_files_num == 0)
1284 config->max_dump_files_num = 128;
1285 eth_dev = rte_eth_dev_allocate(name);
1286 if (eth_dev == NULL) {
1287 DRV_LOG(ERR, "can not allocate rte ethdev");
1291 if (priv->representor) {
1292 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1293 eth_dev->data->representor_id = priv->representor_id;
1296 * Store associated network device interface index. This index
1297 * is permanent throughout the lifetime of device. So, we may store
1298 * the ifindex here and use the cached value further.
1300 MLX5_ASSERT(spawn->ifindex);
1301 priv->if_index = spawn->ifindex;
1302 if (priv->pf_bond >= 0 && priv->master) {
1303 /* Get bond interface info */
1304 err = mlx5_sysfs_bond_info(priv->if_index,
1305 &priv->bond_ifindex,
1308 DRV_LOG(ERR, "unable to get bond info: %s",
1309 strerror(rte_errno));
1311 DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1312 priv->if_index, priv->bond_ifindex,
1315 eth_dev->data->dev_private = priv;
1316 priv->dev_data = eth_dev->data;
1317 eth_dev->data->mac_addrs = priv->mac;
1318 eth_dev->device = dpdk_dev;
1319 eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1320 /* Configure the first MAC address by default. */
1321 if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1323 "port %u cannot get MAC address, is mlx5_en"
1324 " loaded? (errno: %s)",
1325 eth_dev->data->port_id, strerror(rte_errno));
1330 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1331 eth_dev->data->port_id,
1332 mac.addr_bytes[0], mac.addr_bytes[1],
1333 mac.addr_bytes[2], mac.addr_bytes[3],
1334 mac.addr_bytes[4], mac.addr_bytes[5]);
1335 #ifdef RTE_LIBRTE_MLX5_DEBUG
1337 char ifname[IF_NAMESIZE];
1339 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1340 DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1341 eth_dev->data->port_id, ifname);
1343 DRV_LOG(DEBUG, "port %u ifname is unknown",
1344 eth_dev->data->port_id);
1347 /* Get actual MTU if possible. */
1348 err = mlx5_get_mtu(eth_dev, &priv->mtu);
1353 DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1355 /* Initialize burst functions to prevent crashes before link-up. */
1356 eth_dev->rx_pkt_burst = removed_rx_burst;
1357 eth_dev->tx_pkt_burst = removed_tx_burst;
1358 eth_dev->dev_ops = &mlx5_os_dev_ops;
1359 eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1360 eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1361 eth_dev->rx_queue_count = mlx5_rx_queue_count;
1362 /* Register MAC address. */
1363 claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1364 if (config->vf && config->vf_nl_en)
1365 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1366 mlx5_ifindex(eth_dev),
1367 eth_dev->data->mac_addrs,
1368 MLX5_MAX_MAC_ADDRESSES);
1370 priv->ctrl_flows = 0;
1371 rte_spinlock_init(&priv->flow_list_lock);
1372 TAILQ_INIT(&priv->flow_meters);
1373 TAILQ_INIT(&priv->flow_meter_profiles);
1374 /* Hint libmlx5 to use PMD allocator for data plane resources */
1375 mlx5_glue->dv_set_context_attr(sh->ctx,
1376 MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1377 (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1378 .alloc = &mlx5_alloc_verbs_buf,
1379 .free = &mlx5_free_verbs_buf,
1382 /* Bring Ethernet device up. */
1383 DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1384 eth_dev->data->port_id);
1385 mlx5_set_link_up(eth_dev);
1387 * Even though the interrupt handler is not installed yet,
1388 * interrupts will still trigger on the async_fd from
1389 * Verbs context returned by ibv_open_device().
1391 mlx5_link_update(eth_dev, 0);
1392 #ifdef HAVE_MLX5DV_DR_ESWITCH
1393 if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1394 (switch_info->representor || switch_info->master)))
1395 config->dv_esw_en = 0;
1397 config->dv_esw_en = 0;
1399 /* Detect minimal data bytes to inline. */
1400 mlx5_set_min_inline(spawn, config);
1401 /* Store device configuration on private structure. */
1402 priv->config = *config;
1403 /* Create context for virtual machine VLAN workaround. */
1404 priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1405 if (config->dv_flow_en) {
1406 err = mlx5_alloc_shared_dr(priv);
1410 if (config->devx && config->dv_flow_en && config->dest_tir) {
1411 priv->obj_ops = devx_obj_ops;
1412 priv->obj_ops.drop_action_create =
1413 ibv_obj_ops.drop_action_create;
1414 priv->obj_ops.drop_action_destroy =
1415 ibv_obj_ops.drop_action_destroy;
1416 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1417 priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1419 if (config->dv_esw_en)
1420 priv->obj_ops.txq_obj_modify =
1421 ibv_obj_ops.txq_obj_modify;
1423 /* Use specific wrappers for Tx object. */
1424 priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1425 priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1428 priv->obj_ops = ibv_obj_ops;
1430 priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1431 if (!priv->drop_queue.hrxq)
1433 /* Supported Verbs flow priority number detection. */
1434 err = mlx5_flow_discover_priorities(eth_dev);
1439 priv->config.flow_prio = err;
1440 if (!priv->config.dv_esw_en &&
1441 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1442 DRV_LOG(WARNING, "metadata mode %u is not supported "
1443 "(no E-Switch)", priv->config.dv_xmeta_en);
1444 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1446 mlx5_set_metadata_mask(eth_dev);
1447 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1448 !priv->sh->dv_regc0_mask) {
1449 DRV_LOG(ERR, "metadata mode %u is not supported "
1450 "(no metadata reg_c[0] is available)",
1451 priv->config.dv_xmeta_en);
1455 /* Query availability of metadata reg_c's. */
1456 err = mlx5_flow_discover_mreg_c(eth_dev);
1461 if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1463 "port %u extensive metadata register is not supported",
1464 eth_dev->data->port_id);
1465 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1466 DRV_LOG(ERR, "metadata mode %u is not supported "
1467 "(no metadata registers available)",
1468 priv->config.dv_xmeta_en);
1473 if (priv->config.dv_flow_en &&
1474 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1475 mlx5_flow_ext_mreg_supported(eth_dev) &&
1476 priv->sh->dv_regc0_mask) {
1477 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1478 MLX5_FLOW_MREG_HTABLE_SZ,
1481 if (!priv->mreg_cp_tbl) {
1486 mlx5_flow_counter_mode_config(eth_dev);
1490 if (priv->mreg_cp_tbl)
1491 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1493 mlx5_os_free_shared_dr(priv);
1494 if (priv->nl_socket_route >= 0)
1495 close(priv->nl_socket_route);
1496 if (priv->nl_socket_rdma >= 0)
1497 close(priv->nl_socket_rdma);
1498 if (priv->vmwa_context)
1499 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1500 if (eth_dev && priv->drop_queue.hrxq)
1501 mlx5_drop_action_destroy(eth_dev);
1503 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1505 if (eth_dev != NULL)
1506 eth_dev->data->dev_private = NULL;
1508 if (eth_dev != NULL) {
1509 /* mac_addrs must not be freed alone because part of
1512 eth_dev->data->mac_addrs = NULL;
1513 rte_eth_dev_release_port(eth_dev);
1516 mlx5_free_shared_dev_ctx(sh);
1517 MLX5_ASSERT(err > 0);
1523 * Comparison callback to sort device data.
1525 * This is meant to be used with qsort().
1528 * Pointer to pointer to first data object.
1530 * Pointer to pointer to second data object.
1533 * 0 if both objects are equal, less than 0 if the first argument is less
1534 * than the second, greater than 0 otherwise.
1537 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1539 const struct mlx5_switch_info *si_a =
1540 &((const struct mlx5_dev_spawn_data *)a)->info;
1541 const struct mlx5_switch_info *si_b =
1542 &((const struct mlx5_dev_spawn_data *)b)->info;
1545 /* Master device first. */
1546 ret = si_b->master - si_a->master;
1549 /* Then representor devices. */
1550 ret = si_b->representor - si_a->representor;
1553 /* Unidentified devices come last in no specific order. */
1554 if (!si_a->representor)
1556 /* Order representors by name. */
1557 return si_a->port_name - si_b->port_name;
1561 * Match PCI information for possible slaves of bonding device.
1563 * @param[in] ibv_dev
1564 * Pointer to Infiniband device structure.
1565 * @param[in] pci_dev
1566 * Pointer to PCI device structure to match PCI address.
1567 * @param[in] nl_rdma
1568 * Netlink RDMA group socket handle.
1571 * negative value if no bonding device found, otherwise
1572 * positive index of slave PF in bonding.
1575 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1576 const struct rte_pci_device *pci_dev,
1579 char ifname[IF_NAMESIZE + 1];
1580 unsigned int ifindex;
1586 * Try to get master device name. If something goes
1587 * wrong suppose the lack of kernel support and no
1592 if (!strstr(ibv_dev->name, "bond"))
1594 np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1598 * The Master device might not be on the predefined
1599 * port (not on port index 1, it is not garanted),
1600 * we have to scan all Infiniband device port and
1603 for (i = 1; i <= np; ++i) {
1604 /* Check whether Infiniband port is populated. */
1605 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1608 if (!if_indextoname(ifindex, ifname))
1610 /* Try to read bonding slave names from sysfs. */
1612 "/sys/class/net/%s/master/bonding/slaves", ifname);
1613 file = fopen(slaves, "r");
1619 /* Use safe format to check maximal buffer length. */
1620 MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1621 while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1622 char tmp_str[IF_NAMESIZE + 32];
1623 struct rte_pci_addr pci_addr;
1624 struct mlx5_switch_info info;
1626 /* Process slave interface names in the loop. */
1627 snprintf(tmp_str, sizeof(tmp_str),
1628 "/sys/class/net/%s", ifname);
1629 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1630 DRV_LOG(WARNING, "can not get PCI address"
1631 " for netdev \"%s\"", ifname);
1634 if (pci_dev->addr.domain != pci_addr.domain ||
1635 pci_dev->addr.bus != pci_addr.bus ||
1636 pci_dev->addr.devid != pci_addr.devid ||
1637 pci_dev->addr.function != pci_addr.function)
1639 /* Slave interface PCI address match found. */
1641 snprintf(tmp_str, sizeof(tmp_str),
1642 "/sys/class/net/%s/phys_port_name", ifname);
1643 file = fopen(tmp_str, "rb");
1646 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1647 if (fscanf(file, "%32s", tmp_str) == 1)
1648 mlx5_translate_port_name(tmp_str, &info);
1649 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1650 info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1651 pf = info.port_name;
1660 * DPDK callback to register a PCI device.
1662 * This function spawns Ethernet devices out of a given PCI device.
1664 * @param[in] pci_drv
1665 * PCI driver structure (mlx5_driver).
1666 * @param[in] pci_dev
1667 * PCI device information.
1670 * 0 on success, a negative errno value otherwise and rte_errno is set.
1673 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1674 struct rte_pci_device *pci_dev)
1676 struct ibv_device **ibv_list;
1678 * Number of found IB Devices matching with requested PCI BDF.
1679 * nd != 1 means there are multiple IB devices over the same
1680 * PCI device and we have representors and master.
1682 unsigned int nd = 0;
1684 * Number of found IB device Ports. nd = 1 and np = 1..n means
1685 * we have the single multiport IB device, and there may be
1686 * representors attached to some of found ports.
1688 unsigned int np = 0;
1690 * Number of DPDK ethernet devices to Spawn - either over
1691 * multiple IB devices or multiple ports of single IB device.
1692 * Actually this is the number of iterations to spawn.
1694 unsigned int ns = 0;
1697 * < 0 - no bonding device (single one)
1698 * >= 0 - bonding device (value is slave PF index)
1701 struct mlx5_dev_spawn_data *list = NULL;
1702 struct mlx5_dev_config dev_config;
1703 unsigned int dev_config_vf;
1706 if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1707 mlx5_pmd_socket_init();
1708 ret = mlx5_init_once();
1710 DRV_LOG(ERR, "unable to init PMD global data: %s",
1711 strerror(rte_errno));
1715 ibv_list = mlx5_glue->get_device_list(&ret);
1717 rte_errno = errno ? errno : ENOSYS;
1718 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1722 * First scan the list of all Infiniband devices to find
1723 * matching ones, gathering into the list.
1725 struct ibv_device *ibv_match[ret + 1];
1726 int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1727 int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1731 struct rte_pci_addr pci_addr;
1733 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1734 bd = mlx5_device_bond_pci_match
1735 (ibv_list[ret], pci_dev, nl_rdma);
1738 * Bonding device detected. Only one match is allowed,
1739 * the bonding is supported over multi-port IB device,
1740 * there should be no matches on representor PCI
1741 * functions or non VF LAG bonding devices with
1742 * specified address.
1746 "multiple PCI match on bonding device"
1747 "\"%s\" found", ibv_list[ret]->name);
1752 DRV_LOG(INFO, "PCI information matches for"
1753 " slave %d bonding device \"%s\"",
1754 bd, ibv_list[ret]->name);
1755 ibv_match[nd++] = ibv_list[ret];
1758 if (mlx5_dev_to_pci_addr
1759 (ibv_list[ret]->ibdev_path, &pci_addr))
1761 if (pci_dev->addr.domain != pci_addr.domain ||
1762 pci_dev->addr.bus != pci_addr.bus ||
1763 pci_dev->addr.devid != pci_addr.devid ||
1764 pci_dev->addr.function != pci_addr.function)
1766 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1767 ibv_list[ret]->name);
1768 ibv_match[nd++] = ibv_list[ret];
1770 ibv_match[nd] = NULL;
1772 /* No device matches, just complain and bail out. */
1774 "no Verbs device matches PCI device " PCI_PRI_FMT ","
1775 " are kernel drivers loaded?",
1776 pci_dev->addr.domain, pci_dev->addr.bus,
1777 pci_dev->addr.devid, pci_dev->addr.function);
1784 * Found single matching device may have multiple ports.
1785 * Each port may be representor, we have to check the port
1786 * number and check the representors existence.
1789 np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1791 DRV_LOG(WARNING, "can not get IB device \"%s\""
1792 " ports number", ibv_match[0]->name);
1793 if (bd >= 0 && !np) {
1794 DRV_LOG(ERR, "can not get ports"
1795 " for bonding device");
1801 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1804 * This may happen if there is VF LAG kernel support and
1805 * application is compiled with older rdma_core library.
1808 "No kernel/verbs support for VF LAG bonding found.");
1809 rte_errno = ENOTSUP;
1815 * Now we can determine the maximal
1816 * amount of devices to be spawned.
1818 list = mlx5_malloc(MLX5_MEM_ZERO,
1819 sizeof(struct mlx5_dev_spawn_data) *
1821 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1823 DRV_LOG(ERR, "spawn data array allocation failure");
1828 if (bd >= 0 || np > 1) {
1830 * Single IB device with multiple ports found,
1831 * it may be E-Switch master device and representors.
1832 * We have to perform identification through the ports.
1834 MLX5_ASSERT(nl_rdma >= 0);
1835 MLX5_ASSERT(ns == 0);
1836 MLX5_ASSERT(nd == 1);
1838 for (i = 1; i <= np; ++i) {
1839 list[ns].max_port = np;
1840 list[ns].phys_port = i;
1841 list[ns].phys_dev = ibv_match[0];
1842 list[ns].eth_dev = NULL;
1843 list[ns].pci_dev = pci_dev;
1844 list[ns].pf_bond = bd;
1845 list[ns].ifindex = mlx5_nl_ifindex
1847 mlx5_os_get_dev_device_name
1848 (list[ns].phys_dev), i);
1849 if (!list[ns].ifindex) {
1851 * No network interface index found for the
1852 * specified port, it means there is no
1853 * representor on this port. It's OK,
1854 * there can be disabled ports, for example
1855 * if sriov_numvfs < sriov_totalvfs.
1861 ret = mlx5_nl_switch_info
1865 if (ret || (!list[ns].info.representor &&
1866 !list[ns].info.master)) {
1868 * We failed to recognize representors with
1869 * Netlink, let's try to perform the task
1872 ret = mlx5_sysfs_switch_info
1876 if (!ret && bd >= 0) {
1877 switch (list[ns].info.name_type) {
1878 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1879 if (list[ns].info.port_name == bd)
1882 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1884 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1885 if (list[ns].info.pf_num == bd)
1893 if (!ret && (list[ns].info.representor ^
1894 list[ns].info.master))
1899 "unable to recognize master/representors"
1900 " on the IB device with multiple ports");
1907 * The existence of several matching entries (nd > 1) means
1908 * port representors have been instantiated. No existing Verbs
1909 * call nor sysfs entries can tell them apart, this can only
1910 * be done through Netlink calls assuming kernel drivers are
1911 * recent enough to support them.
1913 * In the event of identification failure through Netlink,
1914 * try again through sysfs, then:
1916 * 1. A single IB device matches (nd == 1) with single
1917 * port (np=0/1) and is not a representor, assume
1918 * no switch support.
1920 * 2. Otherwise no safe assumptions can be made;
1921 * complain louder and bail out.
1923 for (i = 0; i != nd; ++i) {
1924 memset(&list[ns].info, 0, sizeof(list[ns].info));
1925 list[ns].max_port = 1;
1926 list[ns].phys_port = 1;
1927 list[ns].phys_dev = ibv_match[i];
1928 list[ns].eth_dev = NULL;
1929 list[ns].pci_dev = pci_dev;
1930 list[ns].pf_bond = -1;
1931 list[ns].ifindex = 0;
1933 list[ns].ifindex = mlx5_nl_ifindex
1935 mlx5_os_get_dev_device_name
1936 (list[ns].phys_dev), 1);
1937 if (!list[ns].ifindex) {
1938 char ifname[IF_NAMESIZE];
1941 * Netlink failed, it may happen with old
1942 * ib_core kernel driver (before 4.16).
1943 * We can assume there is old driver because
1944 * here we are processing single ports IB
1945 * devices. Let's try sysfs to retrieve
1946 * the ifindex. The method works for
1947 * master device only.
1951 * Multiple devices found, assume
1952 * representors, can not distinguish
1953 * master/representor and retrieve
1954 * ifindex via sysfs.
1958 ret = mlx5_get_ifname_sysfs
1959 (ibv_match[i]->ibdev_path, ifname);
1962 if_nametoindex(ifname);
1963 if (!list[ns].ifindex) {
1965 * No network interface index found
1966 * for the specified device, it means
1967 * there it is neither representor
1975 ret = mlx5_nl_switch_info
1979 if (ret || (!list[ns].info.representor &&
1980 !list[ns].info.master)) {
1982 * We failed to recognize representors with
1983 * Netlink, let's try to perform the task
1986 ret = mlx5_sysfs_switch_info
1990 if (!ret && (list[ns].info.representor ^
1991 list[ns].info.master)) {
1993 } else if ((nd == 1) &&
1994 !list[ns].info.representor &&
1995 !list[ns].info.master) {
1997 * Single IB device with
1998 * one physical port and
1999 * attached network device.
2000 * May be SRIOV is not enabled
2001 * or there is no representors.
2003 DRV_LOG(INFO, "no E-Switch support detected");
2010 "unable to recognize master/representors"
2011 " on the multiple IB devices");
2019 * Sort list to probe devices in natural order for users convenience
2020 * (i.e. master first, then representors from lowest to highest ID).
2022 qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2023 /* Device specific configuration. */
2024 switch (pci_dev->id.device_id) {
2025 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2026 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2027 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2028 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2029 case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2030 case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2031 case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
2038 for (i = 0; i != ns; ++i) {
2041 /* Default configuration. */
2042 memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2043 dev_config.vf = dev_config_vf;
2044 dev_config.mps = MLX5_ARG_UNSET;
2045 dev_config.dbnc = MLX5_ARG_UNSET;
2046 dev_config.rx_vec_en = 1;
2047 dev_config.txq_inline_max = MLX5_ARG_UNSET;
2048 dev_config.txq_inline_min = MLX5_ARG_UNSET;
2049 dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2050 dev_config.txqs_inline = MLX5_ARG_UNSET;
2051 dev_config.vf_nl_en = 1;
2052 dev_config.mr_ext_memseg_en = 1;
2053 dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2054 dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2055 dev_config.dv_esw_en = 1;
2056 dev_config.dv_flow_en = 1;
2057 dev_config.decap_en = 1;
2058 dev_config.log_hp_size = MLX5_ARG_UNSET;
2059 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2062 if (!list[i].eth_dev) {
2063 if (rte_errno != EBUSY && rte_errno != EEXIST)
2065 /* Device is disabled or already spawned. Ignore it. */
2068 restore = list[i].eth_dev->data->dev_flags;
2069 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2070 /* Restore non-PCI flags cleared by the above call. */
2071 list[i].eth_dev->data->dev_flags |= restore;
2072 rte_eth_dev_probing_finish(list[i].eth_dev);
2076 "probe of PCI device " PCI_PRI_FMT " aborted after"
2077 " encountering an error: %s",
2078 pci_dev->addr.domain, pci_dev->addr.bus,
2079 pci_dev->addr.devid, pci_dev->addr.function,
2080 strerror(rte_errno));
2084 if (!list[i].eth_dev)
2086 mlx5_dev_close(list[i].eth_dev);
2087 /* mac_addrs must not be freed because in dev_private */
2088 list[i].eth_dev->data->mac_addrs = NULL;
2089 claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2091 /* Restore original error. */
2098 * Do the routine cleanup:
2099 * - close opened Netlink sockets
2100 * - free allocated spawn data array
2101 * - free the Infiniband device list
2109 MLX5_ASSERT(ibv_list);
2110 mlx5_glue->free_device_list(ibv_list);
2115 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2120 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2121 /* Get environment variable to store. */
2122 env = getenv(MLX5_SHUT_UP_BF);
2123 value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2124 if (config->dbnc == MLX5_ARG_UNSET)
2125 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2127 setenv(MLX5_SHUT_UP_BF,
2128 config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2133 mlx5_restore_doorbell_mapping_env(int value)
2135 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2136 /* Restore the original environment variable state. */
2137 if (value == MLX5_ARG_UNSET)
2138 unsetenv(MLX5_SHUT_UP_BF);
2140 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2144 * Extract pdn of PD object using DV API.
2147 * Pointer to the verbs PD object.
2149 * Pointer to the PD object number variable.
2152 * 0 on success, error value otherwise.
2155 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2157 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2158 struct mlx5dv_obj obj;
2159 struct mlx5dv_pd pd_info;
2163 obj.pd.out = &pd_info;
2164 ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2166 DRV_LOG(DEBUG, "Fail to get PD object info");
2175 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2179 * Function API to open IB device.
2181 * This function calls the Linux glue APIs to open a device.
2184 * Pointer to the IB device attributes (name, port, etc).
2185 * @param[out] config
2186 * Pointer to device configuration structure.
2188 * Pointer to shared context structure.
2191 * 0 on success, a positive error value otherwise.
2194 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2195 const struct mlx5_dev_config *config,
2196 struct mlx5_dev_ctx_shared *sh)
2201 sh->numa_node = spawn->pci_dev->device.numa_node;
2202 pthread_mutex_init(&sh->txpp.mutex, NULL);
2204 * Configure environment variable "MLX5_BF_SHUT_UP"
2205 * before the device creation. The rdma_core library
2206 * checks the variable at device creation and
2207 * stores the result internally.
2209 dbmap_env = mlx5_config_doorbell_mapping_env(config);
2210 /* Try to open IB device with DV first, then usual Verbs. */
2212 sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2215 DRV_LOG(DEBUG, "DevX is supported");
2216 /* The device is created, no need for environment. */
2217 mlx5_restore_doorbell_mapping_env(dbmap_env);
2219 /* The environment variable is still configured. */
2220 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2221 err = errno ? errno : ENODEV;
2223 * The environment variable is not needed anymore,
2224 * all device creation attempts are completed.
2226 mlx5_restore_doorbell_mapping_env(dbmap_env);
2229 DRV_LOG(DEBUG, "DevX is NOT supported");
2236 * Install shared asynchronous device events handler.
2237 * This function is implemented to support event sharing
2238 * between multiple ports of single IB device.
2241 * Pointer to mlx5_dev_ctx_shared object.
2244 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2249 sh->intr_handle.fd = -1;
2250 flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2251 ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2252 F_SETFL, flags | O_NONBLOCK);
2254 DRV_LOG(INFO, "failed to change file descriptor async event"
2257 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2258 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2259 if (rte_intr_callback_register(&sh->intr_handle,
2260 mlx5_dev_interrupt_handler, sh)) {
2261 DRV_LOG(INFO, "Fail to install the shared interrupt.");
2262 sh->intr_handle.fd = -1;
2266 #ifdef HAVE_IBV_DEVX_ASYNC
2267 sh->intr_handle_devx.fd = -1;
2269 (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2270 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2272 DRV_LOG(INFO, "failed to allocate devx_comp.");
2275 flags = fcntl(devx_comp->fd, F_GETFL);
2276 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2278 DRV_LOG(INFO, "failed to change file descriptor"
2282 sh->intr_handle_devx.fd = devx_comp->fd;
2283 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2284 if (rte_intr_callback_register(&sh->intr_handle_devx,
2285 mlx5_dev_interrupt_handler_devx, sh)) {
2286 DRV_LOG(INFO, "Fail to install the devx shared"
2288 sh->intr_handle_devx.fd = -1;
2290 #endif /* HAVE_IBV_DEVX_ASYNC */
2295 * Uninstall shared asynchronous device events handler.
2296 * This function is implemented to support event sharing
2297 * between multiple ports of single IB device.
2300 * Pointer to mlx5_dev_ctx_shared object.
2303 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2305 if (sh->intr_handle.fd >= 0)
2306 mlx5_intr_callback_unregister(&sh->intr_handle,
2307 mlx5_dev_interrupt_handler, sh);
2308 #ifdef HAVE_IBV_DEVX_ASYNC
2309 if (sh->intr_handle_devx.fd >= 0)
2310 rte_intr_callback_unregister(&sh->intr_handle_devx,
2311 mlx5_dev_interrupt_handler_devx, sh);
2313 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2318 * Read statistics by a named counter.
2321 * Pointer to the private device data structure.
2322 * @param[in] ctr_name
2323 * Pointer to the name of the statistic counter to read
2325 * Pointer to read statistic value.
2327 * 0 on success and stat is valud, 1 if failed to read the value
2332 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2338 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2339 priv->sh->ibdev_path,
2342 fd = open(path, O_RDONLY);
2344 * in switchdev the file location is not per port
2345 * but rather in <ibdev_path>/hw_counters/<file_name>.
2348 MKSTR(path1, "%s/hw_counters/%s",
2349 priv->sh->ibdev_path,
2351 fd = open(path1, O_RDONLY);
2354 char buf[21] = {'\0'};
2355 ssize_t n = read(fd, buf, sizeof(buf));
2359 *stat = strtoull(buf, NULL, 10);
2369 * Set the reg_mr and dereg_mr call backs
2371 * @param reg_mr_cb[out]
2372 * Pointer to reg_mr func
2373 * @param dereg_mr_cb[out]
2374 * Pointer to dereg_mr func
2378 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2379 mlx5_dereg_mr_t *dereg_mr_cb)
2381 *reg_mr_cb = mlx5_verbs_ops.reg_mr;
2382 *dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2386 * Remove a MAC address from device
2389 * Pointer to Ethernet device structure.
2391 * MAC address index.
2394 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2396 struct mlx5_priv *priv = dev->data->dev_private;
2397 const int vf = priv->config.vf;
2400 mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2401 mlx5_ifindex(dev), priv->mac_own,
2402 &dev->data->mac_addrs[index], index);
2406 * Adds a MAC address to the device
2409 * Pointer to Ethernet device structure.
2411 * MAC address to register.
2413 * MAC address index.
2416 * 0 on success, a negative errno value otherwise
2419 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2422 struct mlx5_priv *priv = dev->data->dev_private;
2423 const int vf = priv->config.vf;
2427 ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2428 mlx5_ifindex(dev), priv->mac_own,
2434 * Modify a VF MAC address
2437 * Pointer to device private data.
2439 * MAC address to modify into.
2441 * Net device interface index
2446 * 0 on success, a negative errno value otherwise
2449 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2450 unsigned int iface_idx,
2451 struct rte_ether_addr *mac_addr,
2454 return mlx5_nl_vf_mac_addr_modify
2455 (priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2459 * Set device promiscuous mode
2462 * Pointer to Ethernet device structure.
2464 * 0 - promiscuous is disabled, otherwise - enabled
2467 * 0 on success, a negative error value otherwise
2470 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2472 struct mlx5_priv *priv = dev->data->dev_private;
2474 return mlx5_nl_promisc(priv->nl_socket_route,
2475 mlx5_ifindex(dev), !!enable);
2479 * Set device promiscuous mode
2482 * Pointer to Ethernet device structure.
2484 * 0 - all multicase is disabled, otherwise - enabled
2487 * 0 on success, a negative error value otherwise
2490 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2492 struct mlx5_priv *priv = dev->data->dev_private;
2494 return mlx5_nl_allmulti(priv->nl_socket_route,
2495 mlx5_ifindex(dev), !!enable);
2499 * Flush device MAC addresses
2502 * Pointer to Ethernet device structure.
2506 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2508 struct mlx5_priv *priv = dev->data->dev_private;
2510 mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2511 dev->data->mac_addrs,
2512 MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2515 const struct eth_dev_ops mlx5_os_dev_ops = {
2516 .dev_configure = mlx5_dev_configure,
2517 .dev_start = mlx5_dev_start,
2518 .dev_stop = mlx5_dev_stop,
2519 .dev_set_link_down = mlx5_set_link_down,
2520 .dev_set_link_up = mlx5_set_link_up,
2521 .dev_close = mlx5_dev_close,
2522 .promiscuous_enable = mlx5_promiscuous_enable,
2523 .promiscuous_disable = mlx5_promiscuous_disable,
2524 .allmulticast_enable = mlx5_allmulticast_enable,
2525 .allmulticast_disable = mlx5_allmulticast_disable,
2526 .link_update = mlx5_link_update,
2527 .stats_get = mlx5_stats_get,
2528 .stats_reset = mlx5_stats_reset,
2529 .xstats_get = mlx5_xstats_get,
2530 .xstats_reset = mlx5_xstats_reset,
2531 .xstats_get_names = mlx5_xstats_get_names,
2532 .fw_version_get = mlx5_fw_version_get,
2533 .dev_infos_get = mlx5_dev_infos_get,
2534 .read_clock = mlx5_txpp_read_clock,
2535 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2536 .vlan_filter_set = mlx5_vlan_filter_set,
2537 .rx_queue_setup = mlx5_rx_queue_setup,
2538 .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2539 .tx_queue_setup = mlx5_tx_queue_setup,
2540 .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2541 .rx_queue_release = mlx5_rx_queue_release,
2542 .tx_queue_release = mlx5_tx_queue_release,
2543 .rx_queue_start = mlx5_rx_queue_start,
2544 .rx_queue_stop = mlx5_rx_queue_stop,
2545 .tx_queue_start = mlx5_tx_queue_start,
2546 .tx_queue_stop = mlx5_tx_queue_stop,
2547 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2548 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2549 .mac_addr_remove = mlx5_mac_addr_remove,
2550 .mac_addr_add = mlx5_mac_addr_add,
2551 .mac_addr_set = mlx5_mac_addr_set,
2552 .set_mc_addr_list = mlx5_set_mc_addr_list,
2553 .mtu_set = mlx5_dev_set_mtu,
2554 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2555 .vlan_offload_set = mlx5_vlan_offload_set,
2556 .reta_update = mlx5_dev_rss_reta_update,
2557 .reta_query = mlx5_dev_rss_reta_query,
2558 .rss_hash_update = mlx5_rss_hash_update,
2559 .rss_hash_conf_get = mlx5_rss_hash_conf_get,
2560 .filter_ctrl = mlx5_dev_filter_ctrl,
2561 .rxq_info_get = mlx5_rxq_info_get,
2562 .txq_info_get = mlx5_txq_info_get,
2563 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2564 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2565 .rx_queue_intr_enable = mlx5_rx_intr_enable,
2566 .rx_queue_intr_disable = mlx5_rx_intr_disable,
2567 .is_removed = mlx5_is_removed,
2568 .udp_tunnel_port_add = mlx5_udp_tunnel_port_add,
2569 .get_module_info = mlx5_get_module_info,
2570 .get_module_eeprom = mlx5_get_module_eeprom,
2571 .hairpin_cap_get = mlx5_hairpin_cap_get,
2572 .mtr_ops_get = mlx5_flow_meter_ops_get,
2573 .hairpin_bind = mlx5_hairpin_bind,
2574 .hairpin_unbind = mlx5_hairpin_unbind,
2575 .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
2576 .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2577 .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2578 .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
2581 /* Available operations from secondary process. */
2582 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2583 .stats_get = mlx5_stats_get,
2584 .stats_reset = mlx5_stats_reset,
2585 .xstats_get = mlx5_xstats_get,
2586 .xstats_reset = mlx5_xstats_reset,
2587 .xstats_get_names = mlx5_xstats_get_names,
2588 .fw_version_get = mlx5_fw_version_get,
2589 .dev_infos_get = mlx5_dev_infos_get,
2590 .read_clock = mlx5_txpp_read_clock,
2591 .rx_queue_start = mlx5_rx_queue_start,
2592 .rx_queue_stop = mlx5_rx_queue_stop,
2593 .tx_queue_start = mlx5_tx_queue_start,
2594 .tx_queue_stop = mlx5_tx_queue_stop,
2595 .rxq_info_get = mlx5_rxq_info_get,
2596 .txq_info_get = mlx5_txq_info_get,
2597 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2598 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2599 .get_module_info = mlx5_get_module_info,
2600 .get_module_eeprom = mlx5_get_module_eeprom,
2603 /* Available operations in flow isolated mode. */
2604 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2605 .dev_configure = mlx5_dev_configure,
2606 .dev_start = mlx5_dev_start,
2607 .dev_stop = mlx5_dev_stop,
2608 .dev_set_link_down = mlx5_set_link_down,
2609 .dev_set_link_up = mlx5_set_link_up,
2610 .dev_close = mlx5_dev_close,
2611 .promiscuous_enable = mlx5_promiscuous_enable,
2612 .promiscuous_disable = mlx5_promiscuous_disable,
2613 .allmulticast_enable = mlx5_allmulticast_enable,
2614 .allmulticast_disable = mlx5_allmulticast_disable,
2615 .link_update = mlx5_link_update,
2616 .stats_get = mlx5_stats_get,
2617 .stats_reset = mlx5_stats_reset,
2618 .xstats_get = mlx5_xstats_get,
2619 .xstats_reset = mlx5_xstats_reset,
2620 .xstats_get_names = mlx5_xstats_get_names,
2621 .fw_version_get = mlx5_fw_version_get,
2622 .dev_infos_get = mlx5_dev_infos_get,
2623 .read_clock = mlx5_txpp_read_clock,
2624 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2625 .vlan_filter_set = mlx5_vlan_filter_set,
2626 .rx_queue_setup = mlx5_rx_queue_setup,
2627 .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2628 .tx_queue_setup = mlx5_tx_queue_setup,
2629 .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2630 .rx_queue_release = mlx5_rx_queue_release,
2631 .tx_queue_release = mlx5_tx_queue_release,
2632 .rx_queue_start = mlx5_rx_queue_start,
2633 .rx_queue_stop = mlx5_rx_queue_stop,
2634 .tx_queue_start = mlx5_tx_queue_start,
2635 .tx_queue_stop = mlx5_tx_queue_stop,
2636 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2637 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2638 .mac_addr_remove = mlx5_mac_addr_remove,
2639 .mac_addr_add = mlx5_mac_addr_add,
2640 .mac_addr_set = mlx5_mac_addr_set,
2641 .set_mc_addr_list = mlx5_set_mc_addr_list,
2642 .mtu_set = mlx5_dev_set_mtu,
2643 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2644 .vlan_offload_set = mlx5_vlan_offload_set,
2645 .filter_ctrl = mlx5_dev_filter_ctrl,
2646 .rxq_info_get = mlx5_rxq_info_get,
2647 .txq_info_get = mlx5_txq_info_get,
2648 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2649 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2650 .rx_queue_intr_enable = mlx5_rx_intr_enable,
2651 .rx_queue_intr_disable = mlx5_rx_intr_disable,
2652 .is_removed = mlx5_is_removed,
2653 .get_module_info = mlx5_get_module_info,
2654 .get_module_eeprom = mlx5_get_module_eeprom,
2655 .hairpin_cap_get = mlx5_hairpin_cap_get,
2656 .mtr_ops_get = mlx5_flow_meter_ops_get,
2657 .hairpin_bind = mlx5_hairpin_bind,
2658 .hairpin_unbind = mlx5_hairpin_unbind,
2659 .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
2660 .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2661 .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2662 .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,