1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
16 #include <linux/rtnetlink.h>
19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
21 #pragma GCC diagnostic ignored "-Wpedantic"
23 #include <infiniband/verbs.h>
25 #pragma GCC diagnostic error "-Wpedantic"
28 #include <rte_malloc.h>
29 #include <rte_ethdev_driver.h>
30 #include <rte_ethdev_pci.h>
32 #include <rte_bus_pci.h>
33 #include <rte_common.h>
34 #include <rte_config.h>
35 #include <rte_eal_memconfig.h>
36 #include <rte_kvargs.h>
37 #include <rte_rwlock.h>
38 #include <rte_spinlock.h>
39 #include <rte_string_fns.h>
42 #include "mlx5_utils.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_autoconf.h"
45 #include "mlx5_defs.h"
46 #include "mlx5_glue.h"
48 #include "mlx5_flow.h"
50 /* Device parameter to enable RX completion queue compression. */
51 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
53 /* Device parameter to enable RX completion entry padding to 128B. */
54 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
56 /* Device parameter to enable padding Rx packet to cacheline size. */
57 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
59 /* Device parameter to enable Multi-Packet Rx queue. */
60 #define MLX5_RX_MPRQ_EN "mprq_en"
62 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
63 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
65 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
66 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
68 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
69 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
71 /* Device parameter to configure inline send. */
72 #define MLX5_TXQ_INLINE "txq_inline"
75 * Device parameter to configure the number of TX queues threshold for
76 * enabling inline send.
78 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
81 * Device parameter to configure the number of TX queues threshold for
82 * enabling vectorized Tx.
84 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
86 /* Device parameter to enable multi-packet send WQEs. */
87 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
89 /* Device parameter to include 2 dsegs in the title WQEBB. */
90 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
92 /* Device parameter to limit the size of inlining packet. */
93 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
95 /* Device parameter to enable hardware Tx vector. */
96 #define MLX5_TX_VEC_EN "tx_vec_en"
98 /* Device parameter to enable hardware Rx vector. */
99 #define MLX5_RX_VEC_EN "rx_vec_en"
101 /* Allow L3 VXLAN flow creation. */
102 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
104 /* Activate DV flow steering. */
105 #define MLX5_DV_FLOW_EN "dv_flow_en"
107 /* Activate Netlink support in VF mode. */
108 #define MLX5_VF_NL_EN "vf_nl_en"
110 /* Enable extending memsegs when creating a MR. */
111 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
113 /* Select port representors to instantiate. */
114 #define MLX5_REPRESENTOR "representor"
116 #ifndef HAVE_IBV_MLX5_MOD_MPW
117 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
118 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
121 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
122 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
125 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
127 /* Shared memory between primary and secondary processes. */
128 struct mlx5_shared_data *mlx5_shared_data;
130 /* Spinlock for mlx5_shared_data allocation. */
131 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
133 /* Process local data for secondary processes. */
134 static struct mlx5_local_data mlx5_local_data;
136 /** Driver-specific log messages type. */
139 /** Data associated with devices to spawn. */
140 struct mlx5_dev_spawn_data {
141 uint32_t ifindex; /**< Network interface index. */
142 uint32_t max_port; /**< IB device maximal port index. */
143 uint32_t ibv_port; /**< IB device physical port index. */
144 struct mlx5_switch_info info; /**< Switch information. */
145 struct ibv_device *ibv_dev; /**< Associated IB device. */
146 struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
149 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
150 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
153 * Allocate shared IB device context. If there is multiport device the
154 * master and representors will share this context, if there is single
155 * port dedicated IB device, the context will be used by only given
156 * port due to unification.
158 * Routine first searches the context for the specified IB device name,
159 * if found the shared context assumed and reference counter is incremented.
160 * If no context found the new one is created and initialized with specified
161 * IB device context and parameters.
164 * Pointer to the IB device attributes (name, port, etc).
167 * Pointer to mlx5_ibv_shared object on success,
168 * otherwise NULL and rte_errno is set.
170 static struct mlx5_ibv_shared *
171 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn)
173 struct mlx5_ibv_shared *sh;
178 /* Secondary process should not create the shared context. */
179 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
180 pthread_mutex_lock(&mlx5_ibv_list_mutex);
181 /* Search for IB context by device name. */
182 LIST_FOREACH(sh, &mlx5_ibv_list, next) {
183 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
188 /* No device found, we have to create new shared context. */
189 assert(spawn->max_port);
190 sh = rte_zmalloc("ethdev shared ib context",
191 sizeof(struct mlx5_ibv_shared) +
193 sizeof(struct mlx5_ibv_shared_port),
194 RTE_CACHE_LINE_SIZE);
196 DRV_LOG(ERR, "shared context allocation failure");
200 /* Try to open IB device with DV first, then usual Verbs. */
202 sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
205 DRV_LOG(DEBUG, "DevX is supported");
207 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
209 err = errno ? errno : ENODEV;
212 DRV_LOG(DEBUG, "DevX is NOT supported");
214 err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
216 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
220 sh->max_port = spawn->max_port;
221 strncpy(sh->ibdev_name, sh->ctx->device->name,
222 sizeof(sh->ibdev_name));
223 strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
224 sizeof(sh->ibdev_path));
225 pthread_mutex_init(&sh->intr_mutex, NULL);
227 * Setting port_id to max unallowed value means
228 * there is no interrupt subhandler installed for
229 * the given port index i.
231 for (i = 0; i < sh->max_port; i++)
232 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
233 sh->pd = mlx5_glue->alloc_pd(sh->ctx);
234 if (sh->pd == NULL) {
235 DRV_LOG(ERR, "PD allocation failure");
239 LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
241 pthread_mutex_unlock(&mlx5_ibv_list_mutex);
244 pthread_mutex_unlock(&mlx5_ibv_list_mutex);
247 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
249 claim_zero(mlx5_glue->close_device(sh->ctx));
257 * Free shared IB device context. Decrement counter and if zero free
258 * all allocated resources and close handles.
261 * Pointer to mlx5_ibv_shared object to free
264 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
266 pthread_mutex_lock(&mlx5_ibv_list_mutex);
268 /* Check the object presence in the list. */
269 struct mlx5_ibv_shared *lctx;
271 LIST_FOREACH(lctx, &mlx5_ibv_list, next)
276 DRV_LOG(ERR, "Freeing non-existing shared IB context");
282 /* Secondary process should not free the shared context. */
283 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
286 LIST_REMOVE(sh, next);
288 * Ensure there is no async event handler installed.
289 * Only primary process handles async device events.
291 assert(!sh->intr_cnt);
293 rte_intr_callback_unregister
294 (&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
295 pthread_mutex_destroy(&sh->intr_mutex);
297 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
299 claim_zero(mlx5_glue->close_device(sh->ctx));
302 pthread_mutex_unlock(&mlx5_ibv_list_mutex);
306 * Initialize DR related data within private structure.
307 * Routine checks the reference counter and does actual
308 * resources creation/initialization only if counter is zero.
311 * Pointer to the private device data structure.
314 * Zero on success, positive error code otherwise.
317 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
319 #ifdef HAVE_MLX5DV_DR
320 struct mlx5_ibv_shared *sh = priv->sh;
326 /* Shared DV/DR structures is already initialized. */
331 /* Reference counter is zero, we should initialize structures. */
332 ns = mlx5dv_dr_create_ns(sh->ctx, MLX5DV_DR_NS_DOMAIN_INGRESS_BYPASS);
334 DRV_LOG(ERR, "ingress mlx5dv_dr_create_ns failed");
339 ns = mlx5dv_dr_create_ns(sh->ctx, MLX5DV_DR_NS_DOMAIN_EGRESS_BYPASS);
341 DRV_LOG(ERR, "egress mlx5dv_dr_create_ns failed");
345 pthread_mutex_init(&sh->dv_mutex, NULL);
352 /* Rollback the created objects. */
354 mlx5dv_dr_destroy_ns(sh->rx_ns);
358 mlx5dv_dr_destroy_ns(sh->tx_ns);
369 * Destroy DR related data within private structure.
372 * Pointer to the private device data structure.
375 mlx5_free_shared_dr(struct mlx5_priv *priv)
377 #ifdef HAVE_MLX5DV_DR
378 struct mlx5_ibv_shared *sh;
380 if (!priv->dr_shared)
385 assert(sh->dv_refcnt);
386 if (sh->dv_refcnt && --sh->dv_refcnt)
389 mlx5dv_dr_destroy_ns(sh->rx_ns);
393 mlx5dv_dr_destroy_ns(sh->tx_ns);
396 pthread_mutex_destroy(&sh->dv_mutex);
403 * Initialize shared data between primary and secondary process.
405 * A memzone is reserved by primary process and secondary processes attach to
409 * 0 on success, a negative errno value otherwise and rte_errno is set.
412 mlx5_init_shared_data(void)
414 const struct rte_memzone *mz;
417 rte_spinlock_lock(&mlx5_shared_data_lock);
418 if (mlx5_shared_data == NULL) {
419 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
420 /* Allocate shared memory. */
421 mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
422 sizeof(*mlx5_shared_data),
426 "Cannot allocate mlx5 shared data\n");
430 mlx5_shared_data = mz->addr;
431 memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
432 rte_spinlock_init(&mlx5_shared_data->lock);
434 /* Lookup allocated shared memory. */
435 mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
438 "Cannot attach mlx5 shared data\n");
442 mlx5_shared_data = mz->addr;
443 memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
447 rte_spinlock_unlock(&mlx5_shared_data_lock);
452 * Retrieve integer value from environment variable.
455 * Environment variable name.
458 * Integer value, 0 if the variable is not set.
461 mlx5_getenv_int(const char *name)
463 const char *val = getenv(name);
471 * Verbs callback to allocate a memory. This function should allocate the space
472 * according to the size provided residing inside a huge page.
473 * Please note that all allocation must respect the alignment from libmlx5
474 * (i.e. currently sysconf(_SC_PAGESIZE)).
477 * The size in bytes of the memory to allocate.
479 * A pointer to the callback data.
482 * Allocated buffer, NULL otherwise and rte_errno is set.
485 mlx5_alloc_verbs_buf(size_t size, void *data)
487 struct mlx5_priv *priv = data;
489 size_t alignment = sysconf(_SC_PAGESIZE);
490 unsigned int socket = SOCKET_ID_ANY;
492 if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
493 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
495 socket = ctrl->socket;
496 } else if (priv->verbs_alloc_ctx.type ==
497 MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
498 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
500 socket = ctrl->socket;
502 assert(data != NULL);
503 ret = rte_malloc_socket(__func__, size, alignment, socket);
510 * Verbs callback to free a memory.
513 * A pointer to the memory to free.
515 * A pointer to the callback data.
518 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
520 assert(data != NULL);
525 * Initialize process private data structure.
528 * Pointer to Ethernet device structure.
531 * 0 on success, a negative errno value otherwise and rte_errno is set.
534 mlx5_proc_priv_init(struct rte_eth_dev *dev)
536 struct mlx5_priv *priv = dev->data->dev_private;
537 struct mlx5_proc_priv *ppriv;
541 * UAR register table follows the process private structure. BlueFlame
542 * registers for Tx queues are stored in the table.
545 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
546 ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
547 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
552 ppriv->uar_table_sz = ppriv_size;
553 dev->process_private = ppriv;
558 * Un-initialize process private data structure.
561 * Pointer to Ethernet device structure.
564 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
566 if (!dev->process_private)
568 rte_free(dev->process_private);
569 dev->process_private = NULL;
573 * DPDK callback to close the device.
575 * Destroy all queues and objects, free memory.
578 * Pointer to Ethernet device structure.
581 mlx5_dev_close(struct rte_eth_dev *dev)
583 struct mlx5_priv *priv = dev->data->dev_private;
587 DRV_LOG(DEBUG, "port %u closing device \"%s\"",
589 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
590 /* In case mlx5_dev_stop() has not been called. */
591 mlx5_dev_interrupt_handler_uninstall(dev);
592 mlx5_traffic_disable(dev);
593 mlx5_flow_flush(dev, NULL);
594 /* Prevent crashes when queues are still in use. */
595 dev->rx_pkt_burst = removed_rx_burst;
596 dev->tx_pkt_burst = removed_tx_burst;
598 /* Disable datapath on secondary process. */
599 mlx5_mp_req_stop_rxtx(dev);
600 if (priv->rxqs != NULL) {
601 /* XXX race condition if mlx5_rx_burst() is still running. */
603 for (i = 0; (i != priv->rxqs_n); ++i)
604 mlx5_rxq_release(dev, i);
608 if (priv->txqs != NULL) {
609 /* XXX race condition if mlx5_tx_burst() is still running. */
611 for (i = 0; (i != priv->txqs_n); ++i)
612 mlx5_txq_release(dev, i);
616 mlx5_proc_priv_uninit(dev);
617 mlx5_mprq_free_mp(dev);
618 mlx5_mr_release(dev);
620 mlx5_free_shared_dr(priv);
621 if (priv->rss_conf.rss_key != NULL)
622 rte_free(priv->rss_conf.rss_key);
623 if (priv->reta_idx != NULL)
624 rte_free(priv->reta_idx);
626 mlx5_nl_mac_addr_flush(dev);
627 if (priv->nl_socket_route >= 0)
628 close(priv->nl_socket_route);
629 if (priv->nl_socket_rdma >= 0)
630 close(priv->nl_socket_rdma);
631 if (priv->tcf_context)
632 mlx5_flow_tcf_context_destroy(priv->tcf_context);
635 * Free the shared context in last turn, because the cleanup
636 * routines above may use some shared fields, like
637 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
638 * ifindex if Netlink fails.
640 mlx5_free_shared_ibctx(priv->sh);
643 ret = mlx5_hrxq_ibv_verify(dev);
645 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
647 ret = mlx5_ind_table_ibv_verify(dev);
649 DRV_LOG(WARNING, "port %u some indirection table still remain",
651 ret = mlx5_rxq_ibv_verify(dev);
653 DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
655 ret = mlx5_rxq_verify(dev);
657 DRV_LOG(WARNING, "port %u some Rx queues still remain",
659 ret = mlx5_txq_ibv_verify(dev);
661 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
663 ret = mlx5_txq_verify(dev);
665 DRV_LOG(WARNING, "port %u some Tx queues still remain",
667 ret = mlx5_flow_verify(dev);
669 DRV_LOG(WARNING, "port %u some flows still remain",
671 if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
675 RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
676 struct mlx5_priv *opriv =
677 rte_eth_devices[port_id].data->dev_private;
680 opriv->domain_id != priv->domain_id ||
681 &rte_eth_devices[port_id] == dev)
686 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
688 memset(priv, 0, sizeof(*priv));
689 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
691 * Reset mac_addrs to NULL such that it is not freed as part of
692 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
693 * it is freed when dev_private is freed.
695 dev->data->mac_addrs = NULL;
698 const struct eth_dev_ops mlx5_dev_ops = {
699 .dev_configure = mlx5_dev_configure,
700 .dev_start = mlx5_dev_start,
701 .dev_stop = mlx5_dev_stop,
702 .dev_set_link_down = mlx5_set_link_down,
703 .dev_set_link_up = mlx5_set_link_up,
704 .dev_close = mlx5_dev_close,
705 .promiscuous_enable = mlx5_promiscuous_enable,
706 .promiscuous_disable = mlx5_promiscuous_disable,
707 .allmulticast_enable = mlx5_allmulticast_enable,
708 .allmulticast_disable = mlx5_allmulticast_disable,
709 .link_update = mlx5_link_update,
710 .stats_get = mlx5_stats_get,
711 .stats_reset = mlx5_stats_reset,
712 .xstats_get = mlx5_xstats_get,
713 .xstats_reset = mlx5_xstats_reset,
714 .xstats_get_names = mlx5_xstats_get_names,
715 .fw_version_get = mlx5_fw_version_get,
716 .dev_infos_get = mlx5_dev_infos_get,
717 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
718 .vlan_filter_set = mlx5_vlan_filter_set,
719 .rx_queue_setup = mlx5_rx_queue_setup,
720 .tx_queue_setup = mlx5_tx_queue_setup,
721 .rx_queue_release = mlx5_rx_queue_release,
722 .tx_queue_release = mlx5_tx_queue_release,
723 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
724 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
725 .mac_addr_remove = mlx5_mac_addr_remove,
726 .mac_addr_add = mlx5_mac_addr_add,
727 .mac_addr_set = mlx5_mac_addr_set,
728 .set_mc_addr_list = mlx5_set_mc_addr_list,
729 .mtu_set = mlx5_dev_set_mtu,
730 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
731 .vlan_offload_set = mlx5_vlan_offload_set,
732 .reta_update = mlx5_dev_rss_reta_update,
733 .reta_query = mlx5_dev_rss_reta_query,
734 .rss_hash_update = mlx5_rss_hash_update,
735 .rss_hash_conf_get = mlx5_rss_hash_conf_get,
736 .filter_ctrl = mlx5_dev_filter_ctrl,
737 .rx_descriptor_status = mlx5_rx_descriptor_status,
738 .tx_descriptor_status = mlx5_tx_descriptor_status,
739 .rx_queue_count = mlx5_rx_queue_count,
740 .rx_queue_intr_enable = mlx5_rx_intr_enable,
741 .rx_queue_intr_disable = mlx5_rx_intr_disable,
742 .is_removed = mlx5_is_removed,
745 /* Available operations from secondary process. */
746 static const struct eth_dev_ops mlx5_dev_sec_ops = {
747 .stats_get = mlx5_stats_get,
748 .stats_reset = mlx5_stats_reset,
749 .xstats_get = mlx5_xstats_get,
750 .xstats_reset = mlx5_xstats_reset,
751 .xstats_get_names = mlx5_xstats_get_names,
752 .fw_version_get = mlx5_fw_version_get,
753 .dev_infos_get = mlx5_dev_infos_get,
754 .rx_descriptor_status = mlx5_rx_descriptor_status,
755 .tx_descriptor_status = mlx5_tx_descriptor_status,
758 /* Available operations in flow isolated mode. */
759 const struct eth_dev_ops mlx5_dev_ops_isolate = {
760 .dev_configure = mlx5_dev_configure,
761 .dev_start = mlx5_dev_start,
762 .dev_stop = mlx5_dev_stop,
763 .dev_set_link_down = mlx5_set_link_down,
764 .dev_set_link_up = mlx5_set_link_up,
765 .dev_close = mlx5_dev_close,
766 .promiscuous_enable = mlx5_promiscuous_enable,
767 .promiscuous_disable = mlx5_promiscuous_disable,
768 .allmulticast_enable = mlx5_allmulticast_enable,
769 .allmulticast_disable = mlx5_allmulticast_disable,
770 .link_update = mlx5_link_update,
771 .stats_get = mlx5_stats_get,
772 .stats_reset = mlx5_stats_reset,
773 .xstats_get = mlx5_xstats_get,
774 .xstats_reset = mlx5_xstats_reset,
775 .xstats_get_names = mlx5_xstats_get_names,
776 .fw_version_get = mlx5_fw_version_get,
777 .dev_infos_get = mlx5_dev_infos_get,
778 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
779 .vlan_filter_set = mlx5_vlan_filter_set,
780 .rx_queue_setup = mlx5_rx_queue_setup,
781 .tx_queue_setup = mlx5_tx_queue_setup,
782 .rx_queue_release = mlx5_rx_queue_release,
783 .tx_queue_release = mlx5_tx_queue_release,
784 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
785 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
786 .mac_addr_remove = mlx5_mac_addr_remove,
787 .mac_addr_add = mlx5_mac_addr_add,
788 .mac_addr_set = mlx5_mac_addr_set,
789 .set_mc_addr_list = mlx5_set_mc_addr_list,
790 .mtu_set = mlx5_dev_set_mtu,
791 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
792 .vlan_offload_set = mlx5_vlan_offload_set,
793 .filter_ctrl = mlx5_dev_filter_ctrl,
794 .rx_descriptor_status = mlx5_rx_descriptor_status,
795 .tx_descriptor_status = mlx5_tx_descriptor_status,
796 .rx_queue_intr_enable = mlx5_rx_intr_enable,
797 .rx_queue_intr_disable = mlx5_rx_intr_disable,
798 .is_removed = mlx5_is_removed,
802 * Verify and store value for device argument.
805 * Key argument to verify.
807 * Value associated with key.
812 * 0 on success, a negative errno value otherwise and rte_errno is set.
815 mlx5_args_check(const char *key, const char *val, void *opaque)
817 struct mlx5_dev_config *config = opaque;
820 /* No-op, port representors are processed in mlx5_dev_spawn(). */
821 if (!strcmp(MLX5_REPRESENTOR, key))
824 tmp = strtoul(val, NULL, 0);
827 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
830 if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
831 config->cqe_comp = !!tmp;
832 } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
833 config->cqe_pad = !!tmp;
834 } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
835 config->hw_padding = !!tmp;
836 } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
837 config->mprq.enabled = !!tmp;
838 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
839 config->mprq.stride_num_n = tmp;
840 } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
841 config->mprq.max_memcpy_len = tmp;
842 } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
843 config->mprq.min_rxqs_num = tmp;
844 } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
845 config->txq_inline = tmp;
846 } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
847 config->txqs_inline = tmp;
848 } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
849 config->txqs_vec = tmp;
850 } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
852 } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
853 config->mpw_hdr_dseg = !!tmp;
854 } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
855 config->inline_max_packet_sz = tmp;
856 } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
857 config->tx_vec_en = !!tmp;
858 } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
859 config->rx_vec_en = !!tmp;
860 } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
861 config->l3_vxlan_en = !!tmp;
862 } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
863 config->vf_nl_en = !!tmp;
864 } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
865 config->dv_flow_en = !!tmp;
866 } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
867 config->mr_ext_memseg_en = !!tmp;
869 DRV_LOG(WARNING, "%s: unknown parameter", key);
877 * Parse device parameters.
880 * Pointer to device configuration structure.
882 * Device arguments structure.
885 * 0 on success, a negative errno value otherwise and rte_errno is set.
888 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
890 const char **params = (const char *[]){
891 MLX5_RXQ_CQE_COMP_EN,
895 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
896 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
899 MLX5_TXQS_MIN_INLINE,
902 MLX5_TXQ_MPW_HDR_DSEG_EN,
903 MLX5_TXQ_MAX_INLINE_LEN,
909 MLX5_MR_EXT_MEMSEG_EN,
913 struct rte_kvargs *kvlist;
919 /* Following UGLY cast is done to pass checkpatch. */
920 kvlist = rte_kvargs_parse(devargs->args, params);
923 /* Process parameters. */
924 for (i = 0; (params[i] != NULL); ++i) {
925 if (rte_kvargs_count(kvlist, params[i])) {
926 ret = rte_kvargs_process(kvlist, params[i],
927 mlx5_args_check, config);
930 rte_kvargs_free(kvlist);
935 rte_kvargs_free(kvlist);
939 static struct rte_pci_driver mlx5_driver;
942 * PMD global initialization.
944 * Independent from individual device, this function initializes global
945 * per-PMD data structures distinguishing primary and secondary processes.
946 * Hence, each initialization is called once per a process.
949 * 0 on success, a negative errno value otherwise and rte_errno is set.
954 struct mlx5_shared_data *sd;
955 struct mlx5_local_data *ld = &mlx5_local_data;
957 if (mlx5_init_shared_data())
959 sd = mlx5_shared_data;
961 rte_spinlock_lock(&sd->lock);
962 switch (rte_eal_process_type()) {
963 case RTE_PROC_PRIMARY:
966 LIST_INIT(&sd->mem_event_cb_list);
967 rte_rwlock_init(&sd->mem_event_rwlock);
968 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
969 mlx5_mr_mem_event_cb, NULL);
970 mlx5_mp_init_primary();
971 sd->init_done = true;
973 case RTE_PROC_SECONDARY:
976 mlx5_mp_init_secondary();
978 ld->init_done = true;
983 rte_spinlock_unlock(&sd->lock);
988 * Spawn an Ethernet device from Verbs information.
991 * Backing DPDK device.
993 * Verbs device parameters (name, port, switch_info) to spawn.
995 * Device configuration parameters.
998 * A valid Ethernet device object on success, NULL otherwise and rte_errno
999 * is set. The following errors are defined:
1001 * EBUSY: device is not supposed to be spawned.
1002 * EEXIST: device is already spawned
1004 static struct rte_eth_dev *
1005 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1006 struct mlx5_dev_spawn_data *spawn,
1007 struct mlx5_dev_config config)
1009 const struct mlx5_switch_info *switch_info = &spawn->info;
1010 struct mlx5_ibv_shared *sh = NULL;
1011 struct ibv_port_attr port_attr;
1012 struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1013 struct rte_eth_dev *eth_dev = NULL;
1014 struct mlx5_priv *priv = NULL;
1016 unsigned int hw_padding = 0;
1018 unsigned int cqe_comp;
1019 unsigned int cqe_pad = 0;
1020 unsigned int tunnel_en = 0;
1021 unsigned int mpls_en = 0;
1022 unsigned int swp = 0;
1023 unsigned int mprq = 0;
1024 unsigned int mprq_min_stride_size_n = 0;
1025 unsigned int mprq_max_stride_size_n = 0;
1026 unsigned int mprq_min_stride_num_n = 0;
1027 unsigned int mprq_max_stride_num_n = 0;
1028 struct ether_addr mac;
1029 char name[RTE_ETH_NAME_MAX_LEN];
1030 int own_domain_id = 0;
1034 /* Determine if this port representor is supposed to be spawned. */
1035 if (switch_info->representor && dpdk_dev->devargs) {
1036 struct rte_eth_devargs eth_da;
1038 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, ð_da);
1041 DRV_LOG(ERR, "failed to process device arguments: %s",
1042 strerror(rte_errno));
1045 for (i = 0; i < eth_da.nb_representor_ports; ++i)
1046 if (eth_da.representor_ports[i] ==
1047 (uint16_t)switch_info->port_name)
1049 if (i == eth_da.nb_representor_ports) {
1054 /* Build device name. */
1055 if (!switch_info->representor)
1056 strlcpy(name, dpdk_dev->name, sizeof(name));
1058 snprintf(name, sizeof(name), "%s_representor_%u",
1059 dpdk_dev->name, switch_info->port_name);
1060 /* check if the device is already spawned */
1061 if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1065 DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1066 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1067 eth_dev = rte_eth_dev_attach_secondary(name);
1068 if (eth_dev == NULL) {
1069 DRV_LOG(ERR, "can not attach rte ethdev");
1073 eth_dev->device = dpdk_dev;
1074 eth_dev->dev_ops = &mlx5_dev_sec_ops;
1075 err = mlx5_proc_priv_init(eth_dev);
1078 /* Receive command fd from primary process */
1079 err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
1082 /* Remap UAR for Tx queues. */
1083 err = mlx5_tx_uar_init_secondary(eth_dev, err);
1087 * Ethdev pointer is still required as input since
1088 * the primary device is not accessible from the
1089 * secondary process.
1091 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1092 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1095 sh = mlx5_alloc_shared_ibctx(spawn);
1098 config.devx = sh->devx;
1099 #ifdef HAVE_IBV_MLX5_MOD_SWP
1100 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1103 * Multi-packet send is supported by ConnectX-4 Lx PF as well
1104 * as all ConnectX-5 devices.
1106 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1107 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1109 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1110 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1112 mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1113 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1114 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1115 DRV_LOG(DEBUG, "enhanced MPW is supported");
1116 mps = MLX5_MPW_ENHANCED;
1118 DRV_LOG(DEBUG, "MPW is supported");
1122 DRV_LOG(DEBUG, "MPW isn't supported");
1123 mps = MLX5_MPW_DISABLED;
1125 #ifdef HAVE_IBV_MLX5_MOD_SWP
1126 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1127 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1128 DRV_LOG(DEBUG, "SWP support: %u", swp);
1131 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1132 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1133 struct mlx5dv_striding_rq_caps mprq_caps =
1134 dv_attr.striding_rq_caps;
1136 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1137 mprq_caps.min_single_stride_log_num_of_bytes);
1138 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1139 mprq_caps.max_single_stride_log_num_of_bytes);
1140 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1141 mprq_caps.min_single_wqe_log_num_of_strides);
1142 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1143 mprq_caps.max_single_wqe_log_num_of_strides);
1144 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1145 mprq_caps.supported_qpts);
1146 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1148 mprq_min_stride_size_n =
1149 mprq_caps.min_single_stride_log_num_of_bytes;
1150 mprq_max_stride_size_n =
1151 mprq_caps.max_single_stride_log_num_of_bytes;
1152 mprq_min_stride_num_n =
1153 mprq_caps.min_single_wqe_log_num_of_strides;
1154 mprq_max_stride_num_n =
1155 mprq_caps.max_single_wqe_log_num_of_strides;
1156 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1157 mprq_min_stride_num_n);
1160 if (RTE_CACHE_LINE_SIZE == 128 &&
1161 !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1165 config.cqe_comp = cqe_comp;
1166 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1167 /* Whether device supports 128B Rx CQE padding. */
1168 cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1169 (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1171 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1172 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1173 tunnel_en = ((dv_attr.tunnel_offloads_caps &
1174 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1175 (dv_attr.tunnel_offloads_caps &
1176 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1178 DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1179 tunnel_en ? "" : "not ");
1182 "tunnel offloading disabled due to old OFED/rdma-core version");
1184 config.tunnel_en = tunnel_en;
1185 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1186 mpls_en = ((dv_attr.tunnel_offloads_caps &
1187 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1188 (dv_attr.tunnel_offloads_caps &
1189 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1190 DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1191 mpls_en ? "" : "not ");
1193 DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1194 " old OFED/rdma-core version or firmware configuration");
1196 config.mpls_en = mpls_en;
1197 /* Check port status. */
1198 err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1200 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1203 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1204 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1208 if (port_attr.state != IBV_PORT_ACTIVE)
1209 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1210 mlx5_glue->port_state_str(port_attr.state),
1212 /* Allocate private eth device data. */
1213 priv = rte_zmalloc("ethdev private structure",
1215 RTE_CACHE_LINE_SIZE);
1217 DRV_LOG(ERR, "priv allocation failure");
1222 priv->ibv_port = spawn->ibv_port;
1223 priv->mtu = ETHER_MTU;
1225 /* Initialize UAR access locks for 32bit implementations. */
1226 rte_spinlock_init(&priv->uar_lock_cq);
1227 for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1228 rte_spinlock_init(&priv->uar_lock[i]);
1230 /* Some internal functions rely on Netlink sockets, open them now. */
1231 priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1232 priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1234 priv->representor = !!switch_info->representor;
1235 priv->master = !!switch_info->master;
1236 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1238 * Currently we support single E-Switch per PF configurations
1239 * only and vport_id field contains the vport index for
1240 * associated VF, which is deduced from representor port name.
1241 * For example, let's have the IB device port 10, it has
1242 * attached network device eth0, which has port name attribute
1243 * pf0vf2, we can deduce the VF number as 2, and set vport index
1244 * as 3 (2+1). This assigning schema should be changed if the
1245 * multiple E-Switch instances per PF configurations or/and PCI
1246 * subfunctions are added.
1248 priv->vport_id = switch_info->representor ?
1249 switch_info->port_name + 1 : -1;
1250 /* representor_id field keeps the unmodified port/VF index. */
1251 priv->representor_id = switch_info->representor ?
1252 switch_info->port_name : -1;
1254 * Look for sibling devices in order to reuse their switch domain
1255 * if any, otherwise allocate one.
1257 RTE_ETH_FOREACH_DEV_OF(port_id, dpdk_dev) {
1258 const struct mlx5_priv *opriv =
1259 rte_eth_devices[port_id].data->dev_private;
1263 RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1265 priv->domain_id = opriv->domain_id;
1268 if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1269 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1272 DRV_LOG(ERR, "unable to allocate switch domain: %s",
1273 strerror(rte_errno));
1278 err = mlx5_args(&config, dpdk_dev->devargs);
1281 DRV_LOG(ERR, "failed to process device arguments: %s",
1282 strerror(rte_errno));
1285 config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1286 IBV_DEVICE_RAW_IP_CSUM);
1287 DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1288 (config.hw_csum ? "" : "not "));
1289 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1290 !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1291 DRV_LOG(DEBUG, "counters are not supported");
1293 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1294 if (config.dv_flow_en) {
1295 DRV_LOG(WARNING, "DV flow is not supported");
1296 config.dv_flow_en = 0;
1299 config.ind_table_max_size =
1300 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1302 * Remove this check once DPDK supports larger/variable
1303 * indirection tables.
1305 if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1306 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1307 DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1308 config.ind_table_max_size);
1309 config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1310 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1311 DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1312 (config.hw_vlan_strip ? "" : "not "));
1313 config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1314 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1315 DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1316 (config.hw_fcs_strip ? "" : "not "));
1317 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1318 hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1319 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1320 hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1321 IBV_DEVICE_PCI_WRITE_END_PADDING);
1323 if (config.hw_padding && !hw_padding) {
1324 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1325 config.hw_padding = 0;
1326 } else if (config.hw_padding) {
1327 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1329 config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1330 (sh->device_attr.tso_caps.supported_qpts &
1331 (1 << IBV_QPT_RAW_PACKET)));
1333 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1335 * MPW is disabled by default, while the Enhanced MPW is enabled
1338 if (config.mps == MLX5_ARG_UNSET)
1339 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1342 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1343 DRV_LOG(INFO, "%sMPS is %s",
1344 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1345 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1346 if (config.cqe_comp && !cqe_comp) {
1347 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1348 config.cqe_comp = 0;
1350 if (config.cqe_pad && !cqe_pad) {
1351 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1353 } else if (config.cqe_pad) {
1354 DRV_LOG(INFO, "Rx CQE padding is enabled");
1356 if (config.mprq.enabled && mprq) {
1357 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1358 config.mprq.stride_num_n < mprq_min_stride_num_n) {
1359 config.mprq.stride_num_n =
1360 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1361 mprq_min_stride_num_n);
1363 "the number of strides"
1364 " for Multi-Packet RQ is out of range,"
1365 " setting default value (%u)",
1366 1 << config.mprq.stride_num_n);
1368 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1369 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1370 } else if (config.mprq.enabled && !mprq) {
1371 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1372 config.mprq.enabled = 0;
1374 eth_dev = rte_eth_dev_allocate(name);
1375 if (eth_dev == NULL) {
1376 DRV_LOG(ERR, "can not allocate rte ethdev");
1380 /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1381 eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1382 if (priv->representor) {
1383 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1384 eth_dev->data->representor_id = priv->representor_id;
1386 eth_dev->data->dev_private = priv;
1387 priv->dev_data = eth_dev->data;
1388 eth_dev->data->mac_addrs = priv->mac;
1389 eth_dev->device = dpdk_dev;
1390 /* Configure the first MAC address by default. */
1391 if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1393 "port %u cannot get MAC address, is mlx5_en"
1394 " loaded? (errno: %s)",
1395 eth_dev->data->port_id, strerror(rte_errno));
1400 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1401 eth_dev->data->port_id,
1402 mac.addr_bytes[0], mac.addr_bytes[1],
1403 mac.addr_bytes[2], mac.addr_bytes[3],
1404 mac.addr_bytes[4], mac.addr_bytes[5]);
1407 char ifname[IF_NAMESIZE];
1409 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1410 DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1411 eth_dev->data->port_id, ifname);
1413 DRV_LOG(DEBUG, "port %u ifname is unknown",
1414 eth_dev->data->port_id);
1417 /* Get actual MTU if possible. */
1418 err = mlx5_get_mtu(eth_dev, &priv->mtu);
1423 DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1425 /* Initialize burst functions to prevent crashes before link-up. */
1426 eth_dev->rx_pkt_burst = removed_rx_burst;
1427 eth_dev->tx_pkt_burst = removed_tx_burst;
1428 eth_dev->dev_ops = &mlx5_dev_ops;
1429 /* Register MAC address. */
1430 claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1431 if (config.vf && config.vf_nl_en)
1432 mlx5_nl_mac_addr_sync(eth_dev);
1433 priv->tcf_context = mlx5_flow_tcf_context_create();
1434 if (!priv->tcf_context) {
1437 "flow rules relying on switch offloads will not be"
1438 " supported: cannot open libmnl socket: %s",
1439 strerror(rte_errno));
1441 struct rte_flow_error error;
1442 unsigned int ifindex = mlx5_ifindex(eth_dev);
1447 "cannot retrieve network interface index";
1449 err = mlx5_flow_tcf_init(priv->tcf_context,
1454 "flow rules relying on switch offloads will"
1455 " not be supported: %s: %s",
1456 error.message, strerror(rte_errno));
1457 mlx5_flow_tcf_context_destroy(priv->tcf_context);
1458 priv->tcf_context = NULL;
1461 if (config.dv_flow_en) {
1462 err = mlx5_alloc_shared_dr(priv);
1466 TAILQ_INIT(&priv->flows);
1467 TAILQ_INIT(&priv->ctrl_flows);
1468 /* Hint libmlx5 to use PMD allocator for data plane resources */
1469 struct mlx5dv_ctx_allocators alctr = {
1470 .alloc = &mlx5_alloc_verbs_buf,
1471 .free = &mlx5_free_verbs_buf,
1474 mlx5_glue->dv_set_context_attr(sh->ctx,
1475 MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1476 (void *)((uintptr_t)&alctr));
1477 /* Bring Ethernet device up. */
1478 DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1479 eth_dev->data->port_id);
1480 mlx5_set_link_up(eth_dev);
1482 * Even though the interrupt handler is not installed yet,
1483 * interrupts will still trigger on the async_fd from
1484 * Verbs context returned by ibv_open_device().
1486 mlx5_link_update(eth_dev, 0);
1487 /* Store device configuration on private structure. */
1488 priv->config = config;
1489 /* Supported Verbs flow priority number detection. */
1490 err = mlx5_flow_discover_priorities(eth_dev);
1495 priv->config.flow_prio = err;
1497 * Once the device is added to the list of memory event
1498 * callback, its global MR cache table cannot be expanded
1499 * on the fly because of deadlock. If it overflows, lookup
1500 * should be done by searching MR list linearly, which is slow.
1502 err = mlx5_mr_btree_init(&priv->mr.cache,
1503 MLX5_MR_BTREE_CACHE_N * 2,
1504 eth_dev->device->numa_node);
1509 /* Add device to memory callback list. */
1510 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1511 LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1512 priv, mem_event_cb);
1513 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1518 mlx5_free_shared_dr(priv);
1519 if (priv->nl_socket_route >= 0)
1520 close(priv->nl_socket_route);
1521 if (priv->nl_socket_rdma >= 0)
1522 close(priv->nl_socket_rdma);
1523 if (priv->tcf_context)
1524 mlx5_flow_tcf_context_destroy(priv->tcf_context);
1526 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1528 if (eth_dev != NULL)
1529 eth_dev->data->dev_private = NULL;
1531 if (eth_dev != NULL) {
1532 /* mac_addrs must not be freed alone because part of dev_private */
1533 eth_dev->data->mac_addrs = NULL;
1534 rte_eth_dev_release_port(eth_dev);
1537 mlx5_free_shared_ibctx(sh);
1544 * Comparison callback to sort device data.
1546 * This is meant to be used with qsort().
1549 * Pointer to pointer to first data object.
1551 * Pointer to pointer to second data object.
1554 * 0 if both objects are equal, less than 0 if the first argument is less
1555 * than the second, greater than 0 otherwise.
1558 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1560 const struct mlx5_switch_info *si_a =
1561 &((const struct mlx5_dev_spawn_data *)a)->info;
1562 const struct mlx5_switch_info *si_b =
1563 &((const struct mlx5_dev_spawn_data *)b)->info;
1566 /* Master device first. */
1567 ret = si_b->master - si_a->master;
1570 /* Then representor devices. */
1571 ret = si_b->representor - si_a->representor;
1574 /* Unidentified devices come last in no specific order. */
1575 if (!si_a->representor)
1577 /* Order representors by name. */
1578 return si_a->port_name - si_b->port_name;
1582 * DPDK callback to register a PCI device.
1584 * This function spawns Ethernet devices out of a given PCI device.
1586 * @param[in] pci_drv
1587 * PCI driver structure (mlx5_driver).
1588 * @param[in] pci_dev
1589 * PCI device information.
1592 * 0 on success, a negative errno value otherwise and rte_errno is set.
1595 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1596 struct rte_pci_device *pci_dev)
1598 struct ibv_device **ibv_list;
1600 * Number of found IB Devices matching with requested PCI BDF.
1601 * nd != 1 means there are multiple IB devices over the same
1602 * PCI device and we have representors and master.
1604 unsigned int nd = 0;
1606 * Number of found IB device Ports. nd = 1 and np = 1..n means
1607 * we have the single multiport IB device, and there may be
1608 * representors attached to some of found ports.
1610 unsigned int np = 0;
1612 * Number of DPDK ethernet devices to Spawn - either over
1613 * multiple IB devices or multiple ports of single IB device.
1614 * Actually this is the number of iterations to spawn.
1616 unsigned int ns = 0;
1617 struct mlx5_dev_config dev_config;
1620 ret = mlx5_init_once();
1622 DRV_LOG(ERR, "unable to init PMD global data: %s",
1623 strerror(rte_errno));
1626 assert(pci_drv == &mlx5_driver);
1628 ibv_list = mlx5_glue->get_device_list(&ret);
1630 rte_errno = errno ? errno : ENOSYS;
1631 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1635 * First scan the list of all Infiniband devices to find
1636 * matching ones, gathering into the list.
1638 struct ibv_device *ibv_match[ret + 1];
1644 struct rte_pci_addr pci_addr;
1646 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1647 if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1649 if (pci_dev->addr.domain != pci_addr.domain ||
1650 pci_dev->addr.bus != pci_addr.bus ||
1651 pci_dev->addr.devid != pci_addr.devid ||
1652 pci_dev->addr.function != pci_addr.function)
1654 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1655 ibv_list[ret]->name);
1656 ibv_match[nd++] = ibv_list[ret];
1658 ibv_match[nd] = NULL;
1660 /* No device matches, just complain and bail out. */
1661 mlx5_glue->free_device_list(ibv_list);
1663 "no Verbs device matches PCI device " PCI_PRI_FMT ","
1664 " are kernel drivers loaded?",
1665 pci_dev->addr.domain, pci_dev->addr.bus,
1666 pci_dev->addr.devid, pci_dev->addr.function);
1671 nl_route = mlx5_nl_init(NETLINK_ROUTE);
1672 nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1675 * Found single matching device may have multiple ports.
1676 * Each port may be representor, we have to check the port
1677 * number and check the representors existence.
1680 np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1682 DRV_LOG(WARNING, "can not get IB device \"%s\""
1683 " ports number", ibv_match[0]->name);
1686 * Now we can determine the maximal
1687 * amount of devices to be spawned.
1689 struct mlx5_dev_spawn_data list[np ? np : nd];
1693 * Single IB device with multiple ports found,
1694 * it may be E-Switch master device and representors.
1695 * We have to perform identification trough the ports.
1697 assert(nl_rdma >= 0);
1700 for (i = 1; i <= np; ++i) {
1701 list[ns].max_port = np;
1702 list[ns].ibv_port = i;
1703 list[ns].ibv_dev = ibv_match[0];
1704 list[ns].eth_dev = NULL;
1705 list[ns].ifindex = mlx5_nl_ifindex
1706 (nl_rdma, list[ns].ibv_dev->name, i);
1707 if (!list[ns].ifindex) {
1709 * No network interface index found for the
1710 * specified port, it means there is no
1711 * representor on this port. It's OK,
1712 * there can be disabled ports, for example
1713 * if sriov_numvfs < sriov_totalvfs.
1719 ret = mlx5_nl_switch_info
1723 if (ret || (!list[ns].info.representor &&
1724 !list[ns].info.master)) {
1726 * We failed to recognize representors with
1727 * Netlink, let's try to perform the task
1730 ret = mlx5_sysfs_switch_info
1734 if (!ret && (list[ns].info.representor ^
1735 list[ns].info.master))
1740 "unable to recognize master/representors"
1741 " on the IB device with multiple ports");
1748 * The existence of several matching entries (nd > 1) means
1749 * port representors have been instantiated. No existing Verbs
1750 * call nor sysfs entries can tell them apart, this can only
1751 * be done through Netlink calls assuming kernel drivers are
1752 * recent enough to support them.
1754 * In the event of identification failure through Netlink,
1755 * try again through sysfs, then:
1757 * 1. A single IB device matches (nd == 1) with single
1758 * port (np=0/1) and is not a representor, assume
1759 * no switch support.
1761 * 2. Otherwise no safe assumptions can be made;
1762 * complain louder and bail out.
1765 for (i = 0; i != nd; ++i) {
1766 memset(&list[ns].info, 0, sizeof(list[ns].info));
1767 list[ns].max_port = 1;
1768 list[ns].ibv_port = 1;
1769 list[ns].ibv_dev = ibv_match[i];
1770 list[ns].eth_dev = NULL;
1771 list[ns].ifindex = 0;
1773 list[ns].ifindex = mlx5_nl_ifindex
1774 (nl_rdma, list[ns].ibv_dev->name, 1);
1775 if (!list[ns].ifindex) {
1776 char ifname[IF_NAMESIZE];
1779 * Netlink failed, it may happen with old
1780 * ib_core kernel driver (before 4.16).
1781 * We can assume there is old driver because
1782 * here we are processing single ports IB
1783 * devices. Let's try sysfs to retrieve
1784 * the ifindex. The method works for
1785 * master device only.
1789 * Multiple devices found, assume
1790 * representors, can not distinguish
1791 * master/representor and retrieve
1792 * ifindex via sysfs.
1796 ret = mlx5_get_master_ifname
1797 (ibv_match[i]->ibdev_path, &ifname);
1800 if_nametoindex(ifname);
1801 if (!list[ns].ifindex) {
1803 * No network interface index found
1804 * for the specified device, it means
1805 * there it is neither representor
1813 ret = mlx5_nl_switch_info
1817 if (ret || (!list[ns].info.representor &&
1818 !list[ns].info.master)) {
1820 * We failed to recognize representors with
1821 * Netlink, let's try to perform the task
1824 ret = mlx5_sysfs_switch_info
1828 if (!ret && (list[ns].info.representor ^
1829 list[ns].info.master)) {
1831 } else if ((nd == 1) &&
1832 !list[ns].info.representor &&
1833 !list[ns].info.master) {
1835 * Single IB device with
1836 * one physical port and
1837 * attached network device.
1838 * May be SRIOV is not enabled
1839 * or there is no representors.
1841 DRV_LOG(INFO, "no E-Switch support detected");
1848 "unable to recognize master/representors"
1849 " on the multiple IB devices");
1857 * Sort list to probe devices in natural order for users convenience
1858 * (i.e. master first, then representors from lowest to highest ID).
1860 qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1861 /* Default configuration. */
1862 dev_config = (struct mlx5_dev_config){
1864 .mps = MLX5_ARG_UNSET,
1867 .txq_inline = MLX5_ARG_UNSET,
1868 .txqs_inline = MLX5_ARG_UNSET,
1869 .txqs_vec = MLX5_ARG_UNSET,
1870 .inline_max_packet_sz = MLX5_ARG_UNSET,
1872 .mr_ext_memseg_en = 1,
1874 .enabled = 0, /* Disabled by default. */
1875 .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
1876 .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1877 .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1880 /* Device specific configuration. */
1881 switch (pci_dev->id.device_id) {
1882 case PCI_DEVICE_ID_MELLANOX_CONNECTX5BF:
1883 dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS_BLUEFIELD;
1885 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1886 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1887 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1888 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1894 /* Set architecture-dependent default value if unset. */
1895 if (dev_config.txqs_vec == MLX5_ARG_UNSET)
1896 dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS;
1897 for (i = 0; i != ns; ++i) {
1900 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1903 if (!list[i].eth_dev) {
1904 if (rte_errno != EBUSY && rte_errno != EEXIST)
1906 /* Device is disabled or already spawned. Ignore it. */
1909 restore = list[i].eth_dev->data->dev_flags;
1910 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1911 /* Restore non-PCI flags cleared by the above call. */
1912 list[i].eth_dev->data->dev_flags |= restore;
1913 rte_eth_dev_probing_finish(list[i].eth_dev);
1917 "probe of PCI device " PCI_PRI_FMT " aborted after"
1918 " encountering an error: %s",
1919 pci_dev->addr.domain, pci_dev->addr.bus,
1920 pci_dev->addr.devid, pci_dev->addr.function,
1921 strerror(rte_errno));
1925 if (!list[i].eth_dev)
1927 mlx5_dev_close(list[i].eth_dev);
1928 /* mac_addrs must not be freed because in dev_private */
1929 list[i].eth_dev->data->mac_addrs = NULL;
1930 claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1932 /* Restore original error. */
1939 * Do the routine cleanup:
1940 * - close opened Netlink sockets
1941 * - free the Infiniband device list
1948 mlx5_glue->free_device_list(ibv_list);
1953 * DPDK callback to remove a PCI device.
1955 * This function removes all Ethernet devices belong to a given PCI device.
1957 * @param[in] pci_dev
1958 * Pointer to the PCI device.
1961 * 0 on success, the function cannot fail.
1964 mlx5_pci_remove(struct rte_pci_device *pci_dev)
1968 RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
1969 rte_eth_dev_close(port_id);
1973 static const struct rte_pci_id mlx5_pci_id_map[] = {
1975 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1976 PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1979 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1980 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1983 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1984 PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1987 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1988 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1991 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1992 PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1995 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1996 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1999 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2000 PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2003 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2004 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2007 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2008 PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2011 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2012 PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2015 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2016 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2019 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2020 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2027 static struct rte_pci_driver mlx5_driver = {
2029 .name = MLX5_DRIVER_NAME
2031 .id_table = mlx5_pci_id_map,
2032 .probe = mlx5_pci_probe,
2033 .remove = mlx5_pci_remove,
2034 .dma_map = mlx5_dma_map,
2035 .dma_unmap = mlx5_dma_unmap,
2036 .drv_flags = (RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
2037 RTE_PCI_DRV_PROBE_AGAIN),
2040 #ifdef RTE_IBVERBS_LINK_DLOPEN
2043 * Suffix RTE_EAL_PMD_PATH with "-glue".
2045 * This function performs a sanity check on RTE_EAL_PMD_PATH before
2046 * suffixing its last component.
2049 * Output buffer, should be large enough otherwise NULL is returned.
2054 * Pointer to @p buf or @p NULL in case suffix cannot be appended.
2057 mlx5_glue_path(char *buf, size_t size)
2059 static const char *const bad[] = { "/", ".", "..", NULL };
2060 const char *path = RTE_EAL_PMD_PATH;
2061 size_t len = strlen(path);
2065 while (len && path[len - 1] == '/')
2067 for (off = len; off && path[off - 1] != '/'; --off)
2069 for (i = 0; bad[i]; ++i)
2070 if (!strncmp(path + off, bad[i], (int)(len - off)))
2072 i = snprintf(buf, size, "%.*s-glue", (int)len, path);
2073 if (i == -1 || (size_t)i >= size)
2078 "unable to append \"-glue\" to last component of"
2079 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
2080 " please re-configure DPDK");
2085 * Initialization routine for run-time dependency on rdma-core.
2088 mlx5_glue_init(void)
2090 char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
2091 const char *path[] = {
2093 * A basic security check is necessary before trusting
2094 * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
2096 (geteuid() == getuid() && getegid() == getgid() ?
2097 getenv("MLX5_GLUE_PATH") : NULL),
2099 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
2100 * variant, otherwise let dlopen() look up libraries on its
2103 (*RTE_EAL_PMD_PATH ?
2104 mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
2107 void *handle = NULL;
2111 while (!handle && i != RTE_DIM(path)) {
2120 end = strpbrk(path[i], ":;");
2122 end = path[i] + strlen(path[i]);
2123 len = end - path[i];
2128 ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
2130 (!len || *(end - 1) == '/') ? "" : "/");
2133 if (sizeof(name) != (size_t)ret + 1)
2135 DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
2137 handle = dlopen(name, RTLD_LAZY);
2148 DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
2151 sym = dlsym(handle, "mlx5_glue");
2152 if (!sym || !*sym) {
2156 DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
2165 "cannot initialize PMD due to missing run-time dependency on"
2166 " rdma-core libraries (libibverbs, libmlx5)");
2173 * Driver initialization routine.
2175 RTE_INIT(rte_mlx5_pmd_init)
2177 /* Initialize driver log type. */
2178 mlx5_logtype = rte_log_register("pmd.net.mlx5");
2179 if (mlx5_logtype >= 0)
2180 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
2182 /* Build the static tables for Verbs conversion. */
2183 mlx5_set_ptype_table();
2184 mlx5_set_cksum_table();
2185 mlx5_set_swp_types_table();
2187 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
2188 * huge pages. Calling ibv_fork_init() during init allows
2189 * applications to use fork() safely for purposes other than
2190 * using this PMD, which is not supported in forked processes.
2192 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
2193 /* Match the size of Rx completion entry to the size of a cacheline. */
2194 if (RTE_CACHE_LINE_SIZE == 128)
2195 setenv("MLX5_CQE_SIZE", "128", 0);
2197 * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
2198 * cleanup all the Verbs resources even when the device was removed.
2200 setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
2201 #ifdef RTE_IBVERBS_LINK_DLOPEN
2202 if (mlx5_glue_init())
2207 /* Glue structure must not contain any NULL pointers. */
2211 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
2212 assert(((const void *const *)mlx5_glue)[i]);
2215 if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
2217 "rdma-core glue \"%s\" mismatch: \"%s\" is required",
2218 mlx5_glue->version, MLX5_GLUE_VERSION);
2221 mlx5_glue->fork_init();
2222 rte_pci_register(&mlx5_driver);
2225 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2226 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2227 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");