1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
14 #include <linux/rtnetlink.h>
17 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
19 #pragma GCC diagnostic ignored "-Wpedantic"
21 #include <infiniband/verbs.h>
23 #pragma GCC diagnostic error "-Wpedantic"
26 #include <rte_malloc.h>
27 #include <rte_ethdev_driver.h>
28 #include <rte_ethdev_pci.h>
30 #include <rte_bus_pci.h>
31 #include <rte_common.h>
32 #include <rte_kvargs.h>
33 #include <rte_rwlock.h>
34 #include <rte_spinlock.h>
35 #include <rte_string_fns.h>
36 #include <rte_alarm.h>
38 #include <mlx5_glue.h>
39 #include <mlx5_devx_cmds.h>
40 #include <mlx5_common.h>
41 #include <mlx5_common_mp.h>
43 #include "mlx5_defs.h"
45 #include "mlx5_utils.h"
46 #include "mlx5_rxtx.h"
47 #include "mlx5_autoconf.h"
49 #include "mlx5_flow.h"
50 #include "rte_pmd_mlx5.h"
52 /* Device parameter to enable RX completion queue compression. */
53 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
55 /* Device parameter to enable RX completion entry padding to 128B. */
56 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
58 /* Device parameter to enable padding Rx packet to cacheline size. */
59 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
61 /* Device parameter to enable Multi-Packet Rx queue. */
62 #define MLX5_RX_MPRQ_EN "mprq_en"
64 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
65 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
67 /* Device parameter to configure log 2 of the stride size for MPRQ. */
68 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
70 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
71 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
73 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
74 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
76 /* Device parameter to configure inline send. Deprecated, ignored.*/
77 #define MLX5_TXQ_INLINE "txq_inline"
79 /* Device parameter to limit packet size to inline with ordinary SEND. */
80 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
82 /* Device parameter to configure minimal data size to inline. */
83 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
85 /* Device parameter to limit packet size to inline with Enhanced MPW. */
86 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
89 * Device parameter to configure the number of TX queues threshold for
90 * enabling inline send.
92 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
95 * Device parameter to configure the number of TX queues threshold for
96 * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
98 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
100 /* Device parameter to enable multi-packet send WQEs. */
101 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
104 * Device parameter to force doorbell register mapping
105 * to non-cahed region eliminating the extra write memory barrier.
107 #define MLX5_TX_DB_NC "tx_db_nc"
110 * Device parameter to include 2 dsegs in the title WQEBB.
111 * Deprecated, ignored.
113 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
116 * Device parameter to limit the size of inlining packet.
117 * Deprecated, ignored.
119 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
122 * Device parameter to enable hardware Tx vector.
123 * Deprecated, ignored (no vectorized Tx routines anymore).
125 #define MLX5_TX_VEC_EN "tx_vec_en"
127 /* Device parameter to enable hardware Rx vector. */
128 #define MLX5_RX_VEC_EN "rx_vec_en"
130 /* Allow L3 VXLAN flow creation. */
131 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
133 /* Activate DV E-Switch flow steering. */
134 #define MLX5_DV_ESW_EN "dv_esw_en"
136 /* Activate DV flow steering. */
137 #define MLX5_DV_FLOW_EN "dv_flow_en"
139 /* Enable extensive flow metadata support. */
140 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
142 /* Activate Netlink support in VF mode. */
143 #define MLX5_VF_NL_EN "vf_nl_en"
145 /* Enable extending memsegs when creating a MR. */
146 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
148 /* Select port representors to instantiate. */
149 #define MLX5_REPRESENTOR "representor"
151 /* Device parameter to configure the maximum number of dump files per queue. */
152 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
154 /* Configure timeout of LRO session (in microseconds). */
155 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
158 * Device parameter to configure the total data buffer size for a single
159 * hairpin queue (logarithm value).
161 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
163 #ifndef HAVE_IBV_MLX5_MOD_MPW
164 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
165 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
168 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
169 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
172 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
174 /* Shared memory between primary and secondary processes. */
175 struct mlx5_shared_data *mlx5_shared_data;
177 /* Spinlock for mlx5_shared_data allocation. */
178 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
180 /* Process local data for secondary processes. */
181 static struct mlx5_local_data mlx5_local_data;
183 /** Driver-specific log messages type. */
186 /** Data associated with devices to spawn. */
187 struct mlx5_dev_spawn_data {
188 uint32_t ifindex; /**< Network interface index. */
189 uint32_t max_port; /**< IB device maximal port index. */
190 uint32_t ibv_port; /**< IB device physical port index. */
191 int pf_bond; /**< bonding device PF index. < 0 - no bonding */
192 struct mlx5_switch_info info; /**< Switch information. */
193 struct ibv_device *ibv_dev; /**< Associated IB device. */
194 struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
195 struct rte_pci_device *pci_dev; /**< Backend PCI device. */
198 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
199 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
201 static struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
202 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
204 .size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
210 .malloc = rte_malloc_socket,
212 .type = "mlx5_encap_decap_ipool",
215 .size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
221 .malloc = rte_malloc_socket,
223 .type = "mlx5_push_vlan_ipool",
226 .size = sizeof(struct mlx5_flow_dv_tag_resource),
232 .malloc = rte_malloc_socket,
234 .type = "mlx5_tag_ipool",
237 .size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
243 .malloc = rte_malloc_socket,
245 .type = "mlx5_port_id_ipool",
248 .size = sizeof(struct mlx5_flow_tbl_data_entry),
254 .malloc = rte_malloc_socket,
256 .type = "mlx5_jump_ipool",
260 .size = sizeof(struct mlx5_flow_meter),
266 .malloc = rte_malloc_socket,
268 .type = "mlx5_meter_ipool",
271 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
277 .malloc = rte_malloc_socket,
279 .type = "mlx5_mcp_ipool",
282 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
288 .malloc = rte_malloc_socket,
290 .type = "mlx5_hrxq_ipool",
293 .size = sizeof(struct mlx5_flow_handle),
299 .malloc = rte_malloc_socket,
301 .type = "mlx5_flow_handle_ipool",
304 .size = sizeof(struct rte_flow),
308 .malloc = rte_malloc_socket,
310 .type = "rte_flow_ipool",
315 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
316 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
318 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
319 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
322 * Allocate ID pool structure.
325 * The maximum id can be allocated from the pool.
328 * Pointer to pool object, NULL value otherwise.
330 struct mlx5_flow_id_pool *
331 mlx5_flow_id_pool_alloc(uint32_t max_id)
333 struct mlx5_flow_id_pool *pool;
336 pool = rte_zmalloc("id pool allocation", sizeof(*pool),
337 RTE_CACHE_LINE_SIZE);
339 DRV_LOG(ERR, "can't allocate id pool");
343 mem = rte_zmalloc("", MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
344 RTE_CACHE_LINE_SIZE);
346 DRV_LOG(ERR, "can't allocate mem for id pool");
350 pool->free_arr = mem;
351 pool->curr = pool->free_arr;
352 pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
353 pool->base_index = 0;
354 pool->max_id = max_id;
362 * Release ID pool structure.
365 * Pointer to flow id pool object to free.
368 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
370 rte_free(pool->free_arr);
378 * Pointer to flow id pool.
383 * 0 on success, error value otherwise.
386 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
388 if (pool->curr == pool->free_arr) {
389 if (pool->base_index == pool->max_id) {
391 DRV_LOG(ERR, "no free id");
394 *id = ++pool->base_index;
397 *id = *(--pool->curr);
405 * Pointer to flow id pool.
410 * 0 on success, error value otherwise.
413 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
419 if (pool->curr == pool->last) {
420 size = pool->curr - pool->free_arr;
421 size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
422 MLX5_ASSERT(size2 > size);
423 mem = rte_malloc("", size2 * sizeof(uint32_t), 0);
425 DRV_LOG(ERR, "can't allocate mem for id pool");
429 memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
430 rte_free(pool->free_arr);
431 pool->free_arr = mem;
432 pool->curr = pool->free_arr + size;
433 pool->last = pool->free_arr + size2;
441 * Initialize the shared aging list information per port.
444 * Pointer to mlx5_ibv_shared object.
447 mlx5_flow_aging_init(struct mlx5_ibv_shared *sh)
450 struct mlx5_age_info *age_info;
452 for (i = 0; i < sh->max_port; i++) {
453 age_info = &sh->port[i].age_info;
455 TAILQ_INIT(&age_info->aged_counters);
456 rte_spinlock_init(&age_info->aged_sl);
457 MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
462 * Initialize the counters management structure.
465 * Pointer to mlx5_ibv_shared object to free
468 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
472 memset(&sh->cmng, 0, sizeof(sh->cmng));
473 TAILQ_INIT(&sh->cmng.flow_counters);
474 for (i = 0; i < MLX5_CCONT_TYPE_MAX; ++i) {
475 TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
476 rte_spinlock_init(&sh->cmng.ccont[i].resize_sl);
481 * Destroy all the resources allocated for a counter memory management.
484 * Pointer to the memory management structure.
487 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
489 uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
491 LIST_REMOVE(mng, next);
492 claim_zero(mlx5_devx_cmd_destroy(mng->dm));
493 claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
498 * Close and release all the resources of the counters management.
501 * Pointer to mlx5_ibv_shared object to free.
504 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
506 struct mlx5_counter_stats_mem_mng *mng;
513 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
514 if (rte_errno != EINPROGRESS)
518 for (i = 0; i < MLX5_CCONT_TYPE_MAX; ++i) {
519 struct mlx5_flow_counter_pool *pool;
520 uint32_t batch = !!(i > 1);
522 if (!sh->cmng.ccont[i].pools)
524 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
526 if (batch && pool->min_dcs)
527 claim_zero(mlx5_devx_cmd_destroy
529 for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
530 if (MLX5_POOL_GET_CNT(pool, j)->action)
532 (mlx5_glue->destroy_flow_action
535 if (!batch && MLX5_GET_POOL_CNT_EXT
537 claim_zero(mlx5_devx_cmd_destroy
538 (MLX5_GET_POOL_CNT_EXT
541 TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool, next);
543 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
545 rte_free(sh->cmng.ccont[i].pools);
547 mng = LIST_FIRST(&sh->cmng.mem_mngs);
549 mlx5_flow_destroy_counter_stat_mem_mng(mng);
550 mng = LIST_FIRST(&sh->cmng.mem_mngs);
552 memset(&sh->cmng, 0, sizeof(sh->cmng));
556 * Initialize the flow resources' indexed mempool.
559 * Pointer to mlx5_ibv_shared object.
561 * Pointer to user dev config.
564 mlx5_flow_ipool_create(struct mlx5_ibv_shared *sh,
565 const struct mlx5_dev_config *config __rte_unused)
569 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
571 * While DV is supported, user chooses the verbs mode,
572 * the mlx5 flow handle size is different with the
573 * MLX5_FLOW_HANDLE_VERBS_SIZE.
575 if (!config->dv_flow_en)
576 mlx5_ipool_cfg[MLX5_IPOOL_MLX5_FLOW].size =
577 MLX5_FLOW_HANDLE_VERBS_SIZE;
579 for (i = 0; i < MLX5_IPOOL_MAX; ++i)
580 sh->ipool[i] = mlx5_ipool_create(&mlx5_ipool_cfg[i]);
584 * Release the flow resources' indexed mempool.
587 * Pointer to mlx5_ibv_shared object.
590 mlx5_flow_ipool_destroy(struct mlx5_ibv_shared *sh)
594 for (i = 0; i < MLX5_IPOOL_MAX; ++i)
595 mlx5_ipool_destroy(sh->ipool[i]);
599 * Extract pdn of PD object using DV API.
602 * Pointer to the verbs PD object.
604 * Pointer to the PD object number variable.
607 * 0 on success, error value otherwise.
609 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
611 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
613 struct mlx5dv_obj obj;
614 struct mlx5dv_pd pd_info;
618 obj.pd.out = &pd_info;
619 ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
621 DRV_LOG(DEBUG, "Fail to get PD object info");
627 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
630 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
635 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
636 /* Get environment variable to store. */
637 env = getenv(MLX5_SHUT_UP_BF);
638 value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
639 if (config->dbnc == MLX5_ARG_UNSET)
640 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
642 setenv(MLX5_SHUT_UP_BF,
643 config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
648 mlx5_restore_doorbell_mapping_env(int value)
650 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
651 /* Restore the original environment variable state. */
652 if (value == MLX5_ARG_UNSET)
653 unsetenv(MLX5_SHUT_UP_BF);
655 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
659 * Allocate shared IB device context. If there is multiport device the
660 * master and representors will share this context, if there is single
661 * port dedicated IB device, the context will be used by only given
662 * port due to unification.
664 * Routine first searches the context for the specified IB device name,
665 * if found the shared context assumed and reference counter is incremented.
666 * If no context found the new one is created and initialized with specified
667 * IB device context and parameters.
670 * Pointer to the IB device attributes (name, port, etc).
672 * Pointer to device configuration structure.
675 * Pointer to mlx5_ibv_shared object on success,
676 * otherwise NULL and rte_errno is set.
678 static struct mlx5_ibv_shared *
679 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
680 const struct mlx5_dev_config *config)
682 struct mlx5_ibv_shared *sh;
686 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
687 struct mlx5_devx_tis_attr tis_attr = { 0 };
691 /* Secondary process should not create the shared context. */
692 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
693 pthread_mutex_lock(&mlx5_ibv_list_mutex);
694 /* Search for IB context by device name. */
695 LIST_FOREACH(sh, &mlx5_ibv_list, next) {
696 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
701 /* No device found, we have to create new shared context. */
702 MLX5_ASSERT(spawn->max_port);
703 sh = rte_zmalloc("ethdev shared ib context",
704 sizeof(struct mlx5_ibv_shared) +
706 sizeof(struct mlx5_ibv_shared_port),
707 RTE_CACHE_LINE_SIZE);
709 DRV_LOG(ERR, "shared context allocation failure");
714 * Configure environment variable "MLX5_BF_SHUT_UP"
715 * before the device creation. The rdma_core library
716 * checks the variable at device creation and
717 * stores the result internally.
719 dbmap_env = mlx5_config_doorbell_mapping_env(config);
720 /* Try to open IB device with DV first, then usual Verbs. */
722 sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
725 DRV_LOG(DEBUG, "DevX is supported");
726 /* The device is created, no need for environment. */
727 mlx5_restore_doorbell_mapping_env(dbmap_env);
729 /* The environment variable is still configured. */
730 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
731 err = errno ? errno : ENODEV;
733 * The environment variable is not needed anymore,
734 * all device creation attempts are completed.
736 mlx5_restore_doorbell_mapping_env(dbmap_env);
739 DRV_LOG(DEBUG, "DevX is NOT supported");
741 err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
743 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
747 sh->max_port = spawn->max_port;
748 strncpy(sh->ibdev_name, sh->ctx->device->name,
749 sizeof(sh->ibdev_name));
750 strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
751 sizeof(sh->ibdev_path));
752 pthread_mutex_init(&sh->intr_mutex, NULL);
754 * Setting port_id to max unallowed value means
755 * there is no interrupt subhandler installed for
756 * the given port index i.
758 for (i = 0; i < sh->max_port; i++) {
759 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
760 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
762 sh->pd = mlx5_glue->alloc_pd(sh->ctx);
763 if (sh->pd == NULL) {
764 DRV_LOG(ERR, "PD allocation failure");
768 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
770 err = mlx5_get_pdn(sh->pd, &sh->pdn);
772 DRV_LOG(ERR, "Fail to extract pdn from PD");
775 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
777 DRV_LOG(ERR, "TD allocation failure");
781 tis_attr.transport_domain = sh->td->id;
782 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
784 DRV_LOG(ERR, "TIS allocation failure");
789 sh->flow_id_pool = mlx5_flow_id_pool_alloc
790 ((1 << HAIRPIN_FLOW_ID_BITS) - 1);
791 if (!sh->flow_id_pool) {
792 DRV_LOG(ERR, "can't create flow id pool");
796 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
798 * Once the device is added to the list of memory event
799 * callback, its global MR cache table cannot be expanded
800 * on the fly because of deadlock. If it overflows, lookup
801 * should be done by searching MR list linearly, which is slow.
803 * At this point the device is not added to the memory
804 * event list yet, context is just being created.
806 err = mlx5_mr_btree_init(&sh->share_cache.cache,
807 MLX5_MR_BTREE_CACHE_N * 2,
808 spawn->pci_dev->device.numa_node);
813 mlx5_flow_aging_init(sh);
814 mlx5_flow_counters_mng_init(sh);
815 mlx5_flow_ipool_create(sh, config);
816 /* Add device to memory callback list. */
817 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
818 LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
820 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
821 /* Add context to the global device list. */
822 LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
824 pthread_mutex_unlock(&mlx5_ibv_list_mutex);
827 pthread_mutex_unlock(&mlx5_ibv_list_mutex);
830 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
832 claim_zero(mlx5_devx_cmd_destroy(sh->td));
834 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
836 claim_zero(mlx5_glue->close_device(sh->ctx));
837 if (sh->flow_id_pool)
838 mlx5_flow_id_pool_release(sh->flow_id_pool);
840 MLX5_ASSERT(err > 0);
846 * Free shared IB device context. Decrement counter and if zero free
847 * all allocated resources and close handles.
850 * Pointer to mlx5_ibv_shared object to free
853 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
855 pthread_mutex_lock(&mlx5_ibv_list_mutex);
856 #ifdef RTE_LIBRTE_MLX5_DEBUG
857 /* Check the object presence in the list. */
858 struct mlx5_ibv_shared *lctx;
860 LIST_FOREACH(lctx, &mlx5_ibv_list, next)
865 DRV_LOG(ERR, "Freeing non-existing shared IB context");
870 MLX5_ASSERT(sh->refcnt);
871 /* Secondary process should not free the shared context. */
872 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
875 /* Remove from memory callback device list. */
876 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
877 LIST_REMOVE(sh, mem_event_cb);
878 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
879 /* Release created Memory Regions. */
880 mlx5_mr_release_cache(&sh->share_cache);
881 /* Remove context from the global device list. */
882 LIST_REMOVE(sh, next);
884 * Ensure there is no async event handler installed.
885 * Only primary process handles async device events.
887 mlx5_flow_counters_mng_close(sh);
888 mlx5_flow_ipool_destroy(sh);
889 MLX5_ASSERT(!sh->intr_cnt);
891 mlx5_intr_callback_unregister
892 (&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
893 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
894 if (sh->devx_intr_cnt) {
895 if (sh->intr_handle_devx.fd)
896 rte_intr_callback_unregister(&sh->intr_handle_devx,
897 mlx5_dev_interrupt_handler_devx, sh);
899 mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
902 pthread_mutex_destroy(&sh->intr_mutex);
904 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
906 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
908 claim_zero(mlx5_devx_cmd_destroy(sh->td));
910 claim_zero(mlx5_glue->close_device(sh->ctx));
911 if (sh->flow_id_pool)
912 mlx5_flow_id_pool_release(sh->flow_id_pool);
915 pthread_mutex_unlock(&mlx5_ibv_list_mutex);
919 * Destroy table hash list and all the root entries per domain.
922 * Pointer to the private device data structure.
925 mlx5_free_table_hash_list(struct mlx5_priv *priv)
927 struct mlx5_ibv_shared *sh = priv->sh;
928 struct mlx5_flow_tbl_data_entry *tbl_data;
929 union mlx5_flow_tbl_key table_key = {
937 struct mlx5_hlist_entry *pos;
941 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
943 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
945 MLX5_ASSERT(tbl_data);
946 mlx5_hlist_remove(sh->flow_tbls, pos);
949 table_key.direction = 1;
950 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
952 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
954 MLX5_ASSERT(tbl_data);
955 mlx5_hlist_remove(sh->flow_tbls, pos);
958 table_key.direction = 0;
959 table_key.domain = 1;
960 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
962 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
964 MLX5_ASSERT(tbl_data);
965 mlx5_hlist_remove(sh->flow_tbls, pos);
968 mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
972 * Initialize flow table hash list and create the root tables entry
976 * Pointer to the private device data structure.
979 * Zero on success, positive error code otherwise.
982 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
984 struct mlx5_ibv_shared *sh = priv->sh;
985 char s[MLX5_HLIST_NAMESIZE];
989 snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
990 sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
991 if (!sh->flow_tbls) {
992 DRV_LOG(ERR, "flow tables with hash creation failed.\n");
996 #ifndef HAVE_MLX5DV_DR
998 * In case we have not DR support, the zero tables should be created
999 * because DV expect to see them even if they cannot be created by
1002 union mlx5_flow_tbl_key table_key = {
1010 struct mlx5_flow_tbl_data_entry *tbl_data = rte_zmalloc(NULL,
1011 sizeof(*tbl_data), 0);
1017 tbl_data->entry.key = table_key.v64;
1018 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1021 rte_atomic32_init(&tbl_data->tbl.refcnt);
1022 rte_atomic32_inc(&tbl_data->tbl.refcnt);
1023 table_key.direction = 1;
1024 tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
1029 tbl_data->entry.key = table_key.v64;
1030 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1033 rte_atomic32_init(&tbl_data->tbl.refcnt);
1034 rte_atomic32_inc(&tbl_data->tbl.refcnt);
1035 table_key.direction = 0;
1036 table_key.domain = 1;
1037 tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
1042 tbl_data->entry.key = table_key.v64;
1043 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1046 rte_atomic32_init(&tbl_data->tbl.refcnt);
1047 rte_atomic32_inc(&tbl_data->tbl.refcnt);
1050 mlx5_free_table_hash_list(priv);
1051 #endif /* HAVE_MLX5DV_DR */
1056 * Initialize DR related data within private structure.
1057 * Routine checks the reference counter and does actual
1058 * resources creation/initialization only if counter is zero.
1061 * Pointer to the private device data structure.
1064 * Zero on success, positive error code otherwise.
1067 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
1069 struct mlx5_ibv_shared *sh = priv->sh;
1070 char s[MLX5_HLIST_NAMESIZE];
1074 err = mlx5_alloc_table_hash_list(priv);
1076 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
1077 (void *)sh->flow_tbls);
1080 /* Create tags hash list table. */
1081 snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
1082 sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
1083 if (!sh->tag_table) {
1084 DRV_LOG(ERR, "tags with hash creation failed.\n");
1088 #ifdef HAVE_MLX5DV_DR
1091 if (sh->dv_refcnt) {
1092 /* Shared DV/DR structures is already initialized. */
1094 priv->dr_shared = 1;
1097 /* Reference counter is zero, we should initialize structures. */
1098 domain = mlx5_glue->dr_create_domain(sh->ctx,
1099 MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
1101 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
1105 sh->rx_domain = domain;
1106 domain = mlx5_glue->dr_create_domain(sh->ctx,
1107 MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
1109 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
1113 pthread_mutex_init(&sh->dv_mutex, NULL);
1114 sh->tx_domain = domain;
1115 #ifdef HAVE_MLX5DV_DR_ESWITCH
1116 if (priv->config.dv_esw_en) {
1117 domain = mlx5_glue->dr_create_domain
1118 (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
1120 DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
1124 sh->fdb_domain = domain;
1125 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
1128 sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
1129 #endif /* HAVE_MLX5DV_DR */
1131 priv->dr_shared = 1;
1134 /* Rollback the created objects. */
1135 if (sh->rx_domain) {
1136 mlx5_glue->dr_destroy_domain(sh->rx_domain);
1137 sh->rx_domain = NULL;
1139 if (sh->tx_domain) {
1140 mlx5_glue->dr_destroy_domain(sh->tx_domain);
1141 sh->tx_domain = NULL;
1143 if (sh->fdb_domain) {
1144 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1145 sh->fdb_domain = NULL;
1147 if (sh->esw_drop_action) {
1148 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1149 sh->esw_drop_action = NULL;
1151 if (sh->pop_vlan_action) {
1152 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1153 sh->pop_vlan_action = NULL;
1155 if (sh->tag_table) {
1156 /* tags should be destroyed with flow before. */
1157 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1158 sh->tag_table = NULL;
1160 mlx5_free_table_hash_list(priv);
1165 * Destroy DR related data within private structure.
1168 * Pointer to the private device data structure.
1171 mlx5_free_shared_dr(struct mlx5_priv *priv)
1173 struct mlx5_ibv_shared *sh;
1175 if (!priv->dr_shared)
1177 priv->dr_shared = 0;
1180 #ifdef HAVE_MLX5DV_DR
1181 MLX5_ASSERT(sh->dv_refcnt);
1182 if (sh->dv_refcnt && --sh->dv_refcnt)
1184 if (sh->rx_domain) {
1185 mlx5_glue->dr_destroy_domain(sh->rx_domain);
1186 sh->rx_domain = NULL;
1188 if (sh->tx_domain) {
1189 mlx5_glue->dr_destroy_domain(sh->tx_domain);
1190 sh->tx_domain = NULL;
1192 #ifdef HAVE_MLX5DV_DR_ESWITCH
1193 if (sh->fdb_domain) {
1194 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1195 sh->fdb_domain = NULL;
1197 if (sh->esw_drop_action) {
1198 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1199 sh->esw_drop_action = NULL;
1202 if (sh->pop_vlan_action) {
1203 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1204 sh->pop_vlan_action = NULL;
1206 pthread_mutex_destroy(&sh->dv_mutex);
1207 #endif /* HAVE_MLX5DV_DR */
1208 if (sh->tag_table) {
1209 /* tags should be destroyed with flow before. */
1210 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1211 sh->tag_table = NULL;
1213 mlx5_free_table_hash_list(priv);
1217 * Initialize shared data between primary and secondary process.
1219 * A memzone is reserved by primary process and secondary processes attach to
1223 * 0 on success, a negative errno value otherwise and rte_errno is set.
1226 mlx5_init_shared_data(void)
1228 const struct rte_memzone *mz;
1231 rte_spinlock_lock(&mlx5_shared_data_lock);
1232 if (mlx5_shared_data == NULL) {
1233 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1234 /* Allocate shared memory. */
1235 mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1236 sizeof(*mlx5_shared_data),
1240 "Cannot allocate mlx5 shared data");
1244 mlx5_shared_data = mz->addr;
1245 memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1246 rte_spinlock_init(&mlx5_shared_data->lock);
1248 /* Lookup allocated shared memory. */
1249 mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1252 "Cannot attach mlx5 shared data");
1256 mlx5_shared_data = mz->addr;
1257 memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1261 rte_spinlock_unlock(&mlx5_shared_data_lock);
1266 * Retrieve integer value from environment variable.
1269 * Environment variable name.
1272 * Integer value, 0 if the variable is not set.
1275 mlx5_getenv_int(const char *name)
1277 const char *val = getenv(name);
1285 * Verbs callback to allocate a memory. This function should allocate the space
1286 * according to the size provided residing inside a huge page.
1287 * Please note that all allocation must respect the alignment from libmlx5
1288 * (i.e. currently sysconf(_SC_PAGESIZE)).
1291 * The size in bytes of the memory to allocate.
1293 * A pointer to the callback data.
1296 * Allocated buffer, NULL otherwise and rte_errno is set.
1299 mlx5_alloc_verbs_buf(size_t size, void *data)
1301 struct mlx5_priv *priv = data;
1303 size_t alignment = sysconf(_SC_PAGESIZE);
1304 unsigned int socket = SOCKET_ID_ANY;
1306 if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
1307 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1309 socket = ctrl->socket;
1310 } else if (priv->verbs_alloc_ctx.type ==
1311 MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
1312 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1314 socket = ctrl->socket;
1316 MLX5_ASSERT(data != NULL);
1317 ret = rte_malloc_socket(__func__, size, alignment, socket);
1324 * Verbs callback to free a memory.
1327 * A pointer to the memory to free.
1329 * A pointer to the callback data.
1332 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1334 MLX5_ASSERT(data != NULL);
1339 * DPDK callback to add udp tunnel port
1342 * A pointer to eth_dev
1343 * @param[in] udp_tunnel
1344 * A pointer to udp tunnel
1347 * 0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1350 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1351 struct rte_eth_udp_tunnel *udp_tunnel)
1353 MLX5_ASSERT(udp_tunnel != NULL);
1354 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1355 udp_tunnel->udp_port == 4789)
1357 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1358 udp_tunnel->udp_port == 4790)
1364 * Initialize process private data structure.
1367 * Pointer to Ethernet device structure.
1370 * 0 on success, a negative errno value otherwise and rte_errno is set.
1373 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1375 struct mlx5_priv *priv = dev->data->dev_private;
1376 struct mlx5_proc_priv *ppriv;
1380 * UAR register table follows the process private structure. BlueFlame
1381 * registers for Tx queues are stored in the table.
1384 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1385 ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1386 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1391 ppriv->uar_table_sz = ppriv_size;
1392 dev->process_private = ppriv;
1397 * Un-initialize process private data structure.
1400 * Pointer to Ethernet device structure.
1403 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1405 if (!dev->process_private)
1407 rte_free(dev->process_private);
1408 dev->process_private = NULL;
1412 * DPDK callback to close the device.
1414 * Destroy all queues and objects, free memory.
1417 * Pointer to Ethernet device structure.
1420 mlx5_dev_close(struct rte_eth_dev *dev)
1422 struct mlx5_priv *priv = dev->data->dev_private;
1426 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1427 /* Check if process_private released. */
1428 if (!dev->process_private)
1430 mlx5_tx_uar_uninit_secondary(dev);
1431 mlx5_proc_priv_uninit(dev);
1432 rte_eth_dev_release_port(dev);
1437 DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1439 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1440 /* In case mlx5_dev_stop() has not been called. */
1441 mlx5_dev_interrupt_handler_uninstall(dev);
1442 mlx5_dev_interrupt_handler_devx_uninstall(dev);
1444 * If default mreg copy action is removed at the stop stage,
1445 * the search will return none and nothing will be done anymore.
1447 mlx5_flow_stop_default(dev);
1448 mlx5_traffic_disable(dev);
1450 * If all the flows are already flushed in the device stop stage,
1451 * then this will return directly without any action.
1453 mlx5_flow_list_flush(dev, &priv->flows, true);
1454 mlx5_flow_meter_flush(dev, NULL);
1455 /* Free the intermediate buffers for flow creation. */
1456 mlx5_flow_free_intermediate(dev);
1457 /* Prevent crashes when queues are still in use. */
1458 dev->rx_pkt_burst = removed_rx_burst;
1459 dev->tx_pkt_burst = removed_tx_burst;
1461 /* Disable datapath on secondary process. */
1462 mlx5_mp_req_stop_rxtx(dev);
1463 if (priv->rxqs != NULL) {
1464 /* XXX race condition if mlx5_rx_burst() is still running. */
1466 for (i = 0; (i != priv->rxqs_n); ++i)
1467 mlx5_rxq_release(dev, i);
1471 if (priv->txqs != NULL) {
1472 /* XXX race condition if mlx5_tx_burst() is still running. */
1474 for (i = 0; (i != priv->txqs_n); ++i)
1475 mlx5_txq_release(dev, i);
1479 mlx5_proc_priv_uninit(dev);
1480 if (priv->mreg_cp_tbl)
1481 mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1482 mlx5_mprq_free_mp(dev);
1483 mlx5_free_shared_dr(priv);
1484 if (priv->rss_conf.rss_key != NULL)
1485 rte_free(priv->rss_conf.rss_key);
1486 if (priv->reta_idx != NULL)
1487 rte_free(priv->reta_idx);
1488 if (priv->config.vf)
1489 mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
1490 dev->data->mac_addrs,
1491 MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
1492 if (priv->nl_socket_route >= 0)
1493 close(priv->nl_socket_route);
1494 if (priv->nl_socket_rdma >= 0)
1495 close(priv->nl_socket_rdma);
1496 if (priv->vmwa_context)
1497 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1498 ret = mlx5_hrxq_verify(dev);
1500 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1501 dev->data->port_id);
1502 ret = mlx5_ind_table_obj_verify(dev);
1504 DRV_LOG(WARNING, "port %u some indirection table still remain",
1505 dev->data->port_id);
1506 ret = mlx5_rxq_obj_verify(dev);
1508 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1509 dev->data->port_id);
1510 ret = mlx5_rxq_verify(dev);
1512 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1513 dev->data->port_id);
1514 ret = mlx5_txq_obj_verify(dev);
1516 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1517 dev->data->port_id);
1518 ret = mlx5_txq_verify(dev);
1520 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1521 dev->data->port_id);
1522 ret = mlx5_flow_verify(dev);
1524 DRV_LOG(WARNING, "port %u some flows still remain",
1525 dev->data->port_id);
1527 * Free the shared context in last turn, because the cleanup
1528 * routines above may use some shared fields, like
1529 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1530 * ifindex if Netlink fails.
1532 mlx5_free_shared_ibctx(priv->sh);
1533 if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1537 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1538 struct mlx5_priv *opriv =
1539 rte_eth_devices[port_id].data->dev_private;
1542 opriv->domain_id != priv->domain_id ||
1543 &rte_eth_devices[port_id] == dev)
1549 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1551 memset(priv, 0, sizeof(*priv));
1552 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1554 * Reset mac_addrs to NULL such that it is not freed as part of
1555 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1556 * it is freed when dev_private is freed.
1558 dev->data->mac_addrs = NULL;
1561 const struct eth_dev_ops mlx5_dev_ops = {
1562 .dev_configure = mlx5_dev_configure,
1563 .dev_start = mlx5_dev_start,
1564 .dev_stop = mlx5_dev_stop,
1565 .dev_set_link_down = mlx5_set_link_down,
1566 .dev_set_link_up = mlx5_set_link_up,
1567 .dev_close = mlx5_dev_close,
1568 .promiscuous_enable = mlx5_promiscuous_enable,
1569 .promiscuous_disable = mlx5_promiscuous_disable,
1570 .allmulticast_enable = mlx5_allmulticast_enable,
1571 .allmulticast_disable = mlx5_allmulticast_disable,
1572 .link_update = mlx5_link_update,
1573 .stats_get = mlx5_stats_get,
1574 .stats_reset = mlx5_stats_reset,
1575 .xstats_get = mlx5_xstats_get,
1576 .xstats_reset = mlx5_xstats_reset,
1577 .xstats_get_names = mlx5_xstats_get_names,
1578 .fw_version_get = mlx5_fw_version_get,
1579 .dev_infos_get = mlx5_dev_infos_get,
1580 .read_clock = mlx5_read_clock,
1581 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1582 .vlan_filter_set = mlx5_vlan_filter_set,
1583 .rx_queue_setup = mlx5_rx_queue_setup,
1584 .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1585 .tx_queue_setup = mlx5_tx_queue_setup,
1586 .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1587 .rx_queue_release = mlx5_rx_queue_release,
1588 .tx_queue_release = mlx5_tx_queue_release,
1589 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1590 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1591 .mac_addr_remove = mlx5_mac_addr_remove,
1592 .mac_addr_add = mlx5_mac_addr_add,
1593 .mac_addr_set = mlx5_mac_addr_set,
1594 .set_mc_addr_list = mlx5_set_mc_addr_list,
1595 .mtu_set = mlx5_dev_set_mtu,
1596 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1597 .vlan_offload_set = mlx5_vlan_offload_set,
1598 .reta_update = mlx5_dev_rss_reta_update,
1599 .reta_query = mlx5_dev_rss_reta_query,
1600 .rss_hash_update = mlx5_rss_hash_update,
1601 .rss_hash_conf_get = mlx5_rss_hash_conf_get,
1602 .filter_ctrl = mlx5_dev_filter_ctrl,
1603 .rx_descriptor_status = mlx5_rx_descriptor_status,
1604 .tx_descriptor_status = mlx5_tx_descriptor_status,
1605 .rxq_info_get = mlx5_rxq_info_get,
1606 .txq_info_get = mlx5_txq_info_get,
1607 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1608 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1609 .rx_queue_count = mlx5_rx_queue_count,
1610 .rx_queue_intr_enable = mlx5_rx_intr_enable,
1611 .rx_queue_intr_disable = mlx5_rx_intr_disable,
1612 .is_removed = mlx5_is_removed,
1613 .udp_tunnel_port_add = mlx5_udp_tunnel_port_add,
1614 .get_module_info = mlx5_get_module_info,
1615 .get_module_eeprom = mlx5_get_module_eeprom,
1616 .hairpin_cap_get = mlx5_hairpin_cap_get,
1617 .mtr_ops_get = mlx5_flow_meter_ops_get,
1620 /* Available operations from secondary process. */
1621 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1622 .stats_get = mlx5_stats_get,
1623 .stats_reset = mlx5_stats_reset,
1624 .xstats_get = mlx5_xstats_get,
1625 .xstats_reset = mlx5_xstats_reset,
1626 .xstats_get_names = mlx5_xstats_get_names,
1627 .fw_version_get = mlx5_fw_version_get,
1628 .dev_infos_get = mlx5_dev_infos_get,
1629 .rx_descriptor_status = mlx5_rx_descriptor_status,
1630 .tx_descriptor_status = mlx5_tx_descriptor_status,
1631 .rxq_info_get = mlx5_rxq_info_get,
1632 .txq_info_get = mlx5_txq_info_get,
1633 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1634 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1635 .get_module_info = mlx5_get_module_info,
1636 .get_module_eeprom = mlx5_get_module_eeprom,
1639 /* Available operations in flow isolated mode. */
1640 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1641 .dev_configure = mlx5_dev_configure,
1642 .dev_start = mlx5_dev_start,
1643 .dev_stop = mlx5_dev_stop,
1644 .dev_set_link_down = mlx5_set_link_down,
1645 .dev_set_link_up = mlx5_set_link_up,
1646 .dev_close = mlx5_dev_close,
1647 .promiscuous_enable = mlx5_promiscuous_enable,
1648 .promiscuous_disable = mlx5_promiscuous_disable,
1649 .allmulticast_enable = mlx5_allmulticast_enable,
1650 .allmulticast_disable = mlx5_allmulticast_disable,
1651 .link_update = mlx5_link_update,
1652 .stats_get = mlx5_stats_get,
1653 .stats_reset = mlx5_stats_reset,
1654 .xstats_get = mlx5_xstats_get,
1655 .xstats_reset = mlx5_xstats_reset,
1656 .xstats_get_names = mlx5_xstats_get_names,
1657 .fw_version_get = mlx5_fw_version_get,
1658 .dev_infos_get = mlx5_dev_infos_get,
1659 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1660 .vlan_filter_set = mlx5_vlan_filter_set,
1661 .rx_queue_setup = mlx5_rx_queue_setup,
1662 .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1663 .tx_queue_setup = mlx5_tx_queue_setup,
1664 .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1665 .rx_queue_release = mlx5_rx_queue_release,
1666 .tx_queue_release = mlx5_tx_queue_release,
1667 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1668 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1669 .mac_addr_remove = mlx5_mac_addr_remove,
1670 .mac_addr_add = mlx5_mac_addr_add,
1671 .mac_addr_set = mlx5_mac_addr_set,
1672 .set_mc_addr_list = mlx5_set_mc_addr_list,
1673 .mtu_set = mlx5_dev_set_mtu,
1674 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1675 .vlan_offload_set = mlx5_vlan_offload_set,
1676 .filter_ctrl = mlx5_dev_filter_ctrl,
1677 .rx_descriptor_status = mlx5_rx_descriptor_status,
1678 .tx_descriptor_status = mlx5_tx_descriptor_status,
1679 .rxq_info_get = mlx5_rxq_info_get,
1680 .txq_info_get = mlx5_txq_info_get,
1681 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1682 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1683 .rx_queue_intr_enable = mlx5_rx_intr_enable,
1684 .rx_queue_intr_disable = mlx5_rx_intr_disable,
1685 .is_removed = mlx5_is_removed,
1686 .get_module_info = mlx5_get_module_info,
1687 .get_module_eeprom = mlx5_get_module_eeprom,
1688 .hairpin_cap_get = mlx5_hairpin_cap_get,
1689 .mtr_ops_get = mlx5_flow_meter_ops_get,
1693 * Verify and store value for device argument.
1696 * Key argument to verify.
1698 * Value associated with key.
1703 * 0 on success, a negative errno value otherwise and rte_errno is set.
1706 mlx5_args_check(const char *key, const char *val, void *opaque)
1708 struct mlx5_dev_config *config = opaque;
1711 /* No-op, port representors are processed in mlx5_dev_spawn(). */
1712 if (!strcmp(MLX5_REPRESENTOR, key))
1715 tmp = strtoul(val, NULL, 0);
1718 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1721 if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1722 config->cqe_comp = !!tmp;
1723 } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1724 config->cqe_pad = !!tmp;
1725 } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1726 config->hw_padding = !!tmp;
1727 } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1728 config->mprq.enabled = !!tmp;
1729 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1730 config->mprq.stride_num_n = tmp;
1731 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1732 config->mprq.stride_size_n = tmp;
1733 } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1734 config->mprq.max_memcpy_len = tmp;
1735 } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1736 config->mprq.min_rxqs_num = tmp;
1737 } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1738 DRV_LOG(WARNING, "%s: deprecated parameter,"
1739 " converted to txq_inline_max", key);
1740 config->txq_inline_max = tmp;
1741 } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1742 config->txq_inline_max = tmp;
1743 } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1744 config->txq_inline_min = tmp;
1745 } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1746 config->txq_inline_mpw = tmp;
1747 } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1748 config->txqs_inline = tmp;
1749 } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1750 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1751 } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1752 config->mps = !!tmp;
1753 } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1754 if (tmp != MLX5_TXDB_CACHED &&
1755 tmp != MLX5_TXDB_NCACHED &&
1756 tmp != MLX5_TXDB_HEURISTIC) {
1757 DRV_LOG(ERR, "invalid Tx doorbell "
1758 "mapping parameter");
1763 } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1764 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1765 } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1766 DRV_LOG(WARNING, "%s: deprecated parameter,"
1767 " converted to txq_inline_mpw", key);
1768 config->txq_inline_mpw = tmp;
1769 } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1770 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1771 } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1772 config->rx_vec_en = !!tmp;
1773 } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1774 config->l3_vxlan_en = !!tmp;
1775 } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1776 config->vf_nl_en = !!tmp;
1777 } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1778 config->dv_esw_en = !!tmp;
1779 } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1780 config->dv_flow_en = !!tmp;
1781 } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1782 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1783 tmp != MLX5_XMETA_MODE_META16 &&
1784 tmp != MLX5_XMETA_MODE_META32) {
1785 DRV_LOG(ERR, "invalid extensive "
1786 "metadata parameter");
1790 config->dv_xmeta_en = tmp;
1791 } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1792 config->mr_ext_memseg_en = !!tmp;
1793 } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1794 config->max_dump_files_num = tmp;
1795 } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1796 config->lro.timeout = tmp;
1797 } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1798 DRV_LOG(DEBUG, "class argument is %s.", val);
1799 } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1800 config->log_hp_size = tmp;
1802 DRV_LOG(WARNING, "%s: unknown parameter", key);
1810 * Parse device parameters.
1813 * Pointer to device configuration structure.
1815 * Device arguments structure.
1818 * 0 on success, a negative errno value otherwise and rte_errno is set.
1821 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1823 const char **params = (const char *[]){
1824 MLX5_RXQ_CQE_COMP_EN,
1825 MLX5_RXQ_CQE_PAD_EN,
1826 MLX5_RXQ_PKT_PAD_EN,
1828 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1829 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1830 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1833 MLX5_TXQ_INLINE_MIN,
1834 MLX5_TXQ_INLINE_MAX,
1835 MLX5_TXQ_INLINE_MPW,
1836 MLX5_TXQS_MIN_INLINE,
1839 MLX5_TXQ_MPW_HDR_DSEG_EN,
1840 MLX5_TXQ_MAX_INLINE_LEN,
1849 MLX5_MR_EXT_MEMSEG_EN,
1851 MLX5_MAX_DUMP_FILES_NUM,
1852 MLX5_LRO_TIMEOUT_USEC,
1853 MLX5_CLASS_ARG_NAME,
1857 struct rte_kvargs *kvlist;
1861 if (devargs == NULL)
1863 /* Following UGLY cast is done to pass checkpatch. */
1864 kvlist = rte_kvargs_parse(devargs->args, params);
1865 if (kvlist == NULL) {
1869 /* Process parameters. */
1870 for (i = 0; (params[i] != NULL); ++i) {
1871 if (rte_kvargs_count(kvlist, params[i])) {
1872 ret = rte_kvargs_process(kvlist, params[i],
1873 mlx5_args_check, config);
1876 rte_kvargs_free(kvlist);
1881 rte_kvargs_free(kvlist);
1885 static struct rte_pci_driver mlx5_driver;
1888 * PMD global initialization.
1890 * Independent from individual device, this function initializes global
1891 * per-PMD data structures distinguishing primary and secondary processes.
1892 * Hence, each initialization is called once per a process.
1895 * 0 on success, a negative errno value otherwise and rte_errno is set.
1898 mlx5_init_once(void)
1900 struct mlx5_shared_data *sd;
1901 struct mlx5_local_data *ld = &mlx5_local_data;
1904 if (mlx5_init_shared_data())
1906 sd = mlx5_shared_data;
1908 rte_spinlock_lock(&sd->lock);
1909 switch (rte_eal_process_type()) {
1910 case RTE_PROC_PRIMARY:
1913 LIST_INIT(&sd->mem_event_cb_list);
1914 rte_rwlock_init(&sd->mem_event_rwlock);
1915 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1916 mlx5_mr_mem_event_cb, NULL);
1917 ret = mlx5_mp_init_primary(MLX5_MP_NAME,
1918 mlx5_mp_primary_handle);
1921 sd->init_done = true;
1923 case RTE_PROC_SECONDARY:
1926 ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
1927 mlx5_mp_secondary_handle);
1930 ++sd->secondary_cnt;
1931 ld->init_done = true;
1937 rte_spinlock_unlock(&sd->lock);
1942 * Configures the minimal amount of data to inline into WQE
1943 * while sending packets.
1945 * - the txq_inline_min has the maximal priority, if this
1946 * key is specified in devargs
1947 * - if DevX is enabled the inline mode is queried from the
1948 * device (HCA attributes and NIC vport context if needed).
1949 * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1950 * and none (0 bytes) for other NICs
1953 * Verbs device parameters (name, port, switch_info) to spawn.
1955 * Device configuration parameters.
1958 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1959 struct mlx5_dev_config *config)
1961 if (config->txq_inline_min != MLX5_ARG_UNSET) {
1962 /* Application defines size of inlined data explicitly. */
1963 switch (spawn->pci_dev->id.device_id) {
1964 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1965 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1966 if (config->txq_inline_min <
1967 (int)MLX5_INLINE_HSIZE_L2) {
1969 "txq_inline_mix aligned to minimal"
1970 " ConnectX-4 required value %d",
1971 (int)MLX5_INLINE_HSIZE_L2);
1972 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1978 if (config->hca_attr.eth_net_offloads) {
1979 /* We have DevX enabled, inline mode queried successfully. */
1980 switch (config->hca_attr.wqe_inline_mode) {
1981 case MLX5_CAP_INLINE_MODE_L2:
1982 /* outer L2 header must be inlined. */
1983 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1985 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1986 /* No inline data are required by NIC. */
1987 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1988 config->hw_vlan_insert =
1989 config->hca_attr.wqe_vlan_insert;
1990 DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1992 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1993 /* inline mode is defined by NIC vport context. */
1994 if (!config->hca_attr.eth_virt)
1996 switch (config->hca_attr.vport_inline_mode) {
1997 case MLX5_INLINE_MODE_NONE:
1998 config->txq_inline_min =
1999 MLX5_INLINE_HSIZE_NONE;
2001 case MLX5_INLINE_MODE_L2:
2002 config->txq_inline_min =
2003 MLX5_INLINE_HSIZE_L2;
2005 case MLX5_INLINE_MODE_IP:
2006 config->txq_inline_min =
2007 MLX5_INLINE_HSIZE_L3;
2009 case MLX5_INLINE_MODE_TCP_UDP:
2010 config->txq_inline_min =
2011 MLX5_INLINE_HSIZE_L4;
2013 case MLX5_INLINE_MODE_INNER_L2:
2014 config->txq_inline_min =
2015 MLX5_INLINE_HSIZE_INNER_L2;
2017 case MLX5_INLINE_MODE_INNER_IP:
2018 config->txq_inline_min =
2019 MLX5_INLINE_HSIZE_INNER_L3;
2021 case MLX5_INLINE_MODE_INNER_TCP_UDP:
2022 config->txq_inline_min =
2023 MLX5_INLINE_HSIZE_INNER_L4;
2029 * We get here if we are unable to deduce
2030 * inline data size with DevX. Try PCI ID
2031 * to determine old NICs.
2033 switch (spawn->pci_dev->id.device_id) {
2034 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2035 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2036 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2037 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2038 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2039 config->hw_vlan_insert = 0;
2041 case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2042 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2043 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2044 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2046 * These NICs support VLAN insertion from WQE and
2047 * report the wqe_vlan_insert flag. But there is the bug
2048 * and PFC control may be broken, so disable feature.
2050 config->hw_vlan_insert = 0;
2051 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2054 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2058 DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2062 * Configures the metadata mask fields in the shared context.
2065 * Pointer to Ethernet device.
2068 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2070 struct mlx5_priv *priv = dev->data->dev_private;
2071 struct mlx5_ibv_shared *sh = priv->sh;
2072 uint32_t meta, mark, reg_c0;
2074 reg_c0 = ~priv->vport_meta_mask;
2075 switch (priv->config.dv_xmeta_en) {
2076 case MLX5_XMETA_MODE_LEGACY:
2078 mark = MLX5_FLOW_MARK_MASK;
2080 case MLX5_XMETA_MODE_META16:
2081 meta = reg_c0 >> rte_bsf32(reg_c0);
2082 mark = MLX5_FLOW_MARK_MASK;
2084 case MLX5_XMETA_MODE_META32:
2086 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2094 if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2095 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2096 sh->dv_mark_mask, mark);
2098 sh->dv_mark_mask = mark;
2099 if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2100 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2101 sh->dv_meta_mask, meta);
2103 sh->dv_meta_mask = meta;
2104 if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2105 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2106 sh->dv_meta_mask, reg_c0);
2108 sh->dv_regc0_mask = reg_c0;
2109 DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2110 DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2111 DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2112 DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2116 * Allocate page of door-bells and register it using DevX API.
2119 * Pointer to Ethernet device.
2122 * Pointer to new page on success, NULL otherwise.
2124 static struct mlx5_devx_dbr_page *
2125 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
2127 struct mlx5_priv *priv = dev->data->dev_private;
2128 struct mlx5_devx_dbr_page *page;
2130 /* Allocate space for door-bell page and management data. */
2131 page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
2132 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
2134 DRV_LOG(ERR, "port %u cannot allocate dbr page",
2135 dev->data->port_id);
2138 /* Register allocated memory. */
2139 page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
2140 MLX5_DBR_PAGE_SIZE, 0);
2142 DRV_LOG(ERR, "port %u cannot umem reg dbr page",
2143 dev->data->port_id);
2151 * Find the next available door-bell, allocate new page if needed.
2154 * Pointer to Ethernet device.
2155 * @param [out] dbr_page
2156 * Door-bell page containing the page data.
2159 * Door-bell address offset on success, a negative error value otherwise.
2162 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
2164 struct mlx5_priv *priv = dev->data->dev_private;
2165 struct mlx5_devx_dbr_page *page = NULL;
2168 LIST_FOREACH(page, &priv->dbrpgs, next)
2169 if (page->dbr_count < MLX5_DBR_PER_PAGE)
2171 if (!page) { /* No page with free door-bell exists. */
2172 page = mlx5_alloc_dbr_page(dev);
2173 if (!page) /* Failed to allocate new page. */
2175 LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
2177 /* Loop to find bitmap part with clear bit. */
2179 i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
2182 /* Find the first clear bit. */
2183 MLX5_ASSERT(i < MLX5_DBR_BITMAP_SIZE);
2184 j = rte_bsf64(~page->dbr_bitmap[i]);
2185 page->dbr_bitmap[i] |= (UINT64_C(1) << j);
2188 return (((i * 64) + j) * sizeof(uint64_t));
2192 * Release a door-bell record.
2195 * Pointer to Ethernet device.
2196 * @param [in] umem_id
2197 * UMEM ID of page containing the door-bell record to release.
2198 * @param [in] offset
2199 * Offset of door-bell record in page.
2202 * 0 on success, a negative error value otherwise.
2205 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
2207 struct mlx5_priv *priv = dev->data->dev_private;
2208 struct mlx5_devx_dbr_page *page = NULL;
2211 LIST_FOREACH(page, &priv->dbrpgs, next)
2212 /* Find the page this address belongs to. */
2213 if (page->umem->umem_id == umem_id)
2218 if (!page->dbr_count) {
2219 /* Page not used, free it and remove from list. */
2220 LIST_REMOVE(page, next);
2222 ret = -mlx5_glue->devx_umem_dereg(page->umem);
2225 /* Mark in bitmap that this door-bell is not in use. */
2226 offset /= MLX5_DBR_SIZE;
2227 int i = offset / 64;
2228 int j = offset % 64;
2230 page->dbr_bitmap[i] &= ~(UINT64_C(1) << j);
2236 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2238 static const char *const dynf_names[] = {
2239 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2240 RTE_MBUF_DYNFLAG_METADATA_NAME
2244 if (n < RTE_DIM(dynf_names))
2246 for (i = 0; i < RTE_DIM(dynf_names); i++) {
2247 if (names[i] == NULL)
2249 strcpy(names[i], dynf_names[i]);
2251 return RTE_DIM(dynf_names);
2255 * Check sibling device configurations.
2257 * Sibling devices sharing the Infiniband device context
2258 * should have compatible configurations. This regards
2259 * representors and bonding slaves.
2262 * Private device descriptor.
2264 * Configuration of the device is going to be created.
2267 * 0 on success, EINVAL otherwise
2270 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2271 struct mlx5_dev_config *config)
2273 struct mlx5_ibv_shared *sh = priv->sh;
2274 struct mlx5_dev_config *sh_conf = NULL;
2278 /* Nothing to compare for the single/first device. */
2279 if (sh->refcnt == 1)
2281 /* Find the device with shared context. */
2282 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2283 struct mlx5_priv *opriv =
2284 rte_eth_devices[port_id].data->dev_private;
2286 if (opriv && opriv != priv && opriv->sh == sh) {
2287 sh_conf = &opriv->config;
2293 if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2294 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2295 " for shared %s context", sh->ibdev_name);
2299 if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2300 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2301 " for shared %s context", sh->ibdev_name);
2308 * Spawn an Ethernet device from Verbs information.
2311 * Backing DPDK device.
2313 * Verbs device parameters (name, port, switch_info) to spawn.
2315 * Device configuration parameters.
2318 * A valid Ethernet device object on success, NULL otherwise and rte_errno
2319 * is set. The following errors are defined:
2321 * EBUSY: device is not supposed to be spawned.
2322 * EEXIST: device is already spawned
2324 static struct rte_eth_dev *
2325 mlx5_dev_spawn(struct rte_device *dpdk_dev,
2326 struct mlx5_dev_spawn_data *spawn,
2327 struct mlx5_dev_config config)
2329 const struct mlx5_switch_info *switch_info = &spawn->info;
2330 struct mlx5_ibv_shared *sh = NULL;
2331 struct ibv_port_attr port_attr;
2332 struct mlx5dv_context dv_attr = { .comp_mask = 0 };
2333 struct rte_eth_dev *eth_dev = NULL;
2334 struct mlx5_priv *priv = NULL;
2336 unsigned int hw_padding = 0;
2338 unsigned int cqe_comp;
2339 unsigned int cqe_pad = 0;
2340 unsigned int tunnel_en = 0;
2341 unsigned int mpls_en = 0;
2342 unsigned int swp = 0;
2343 unsigned int mprq = 0;
2344 unsigned int mprq_min_stride_size_n = 0;
2345 unsigned int mprq_max_stride_size_n = 0;
2346 unsigned int mprq_min_stride_num_n = 0;
2347 unsigned int mprq_max_stride_num_n = 0;
2348 struct rte_ether_addr mac;
2349 char name[RTE_ETH_NAME_MAX_LEN];
2350 int own_domain_id = 0;
2353 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2354 struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
2357 /* Determine if this port representor is supposed to be spawned. */
2358 if (switch_info->representor && dpdk_dev->devargs) {
2359 struct rte_eth_devargs eth_da;
2361 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, ð_da);
2364 DRV_LOG(ERR, "failed to process device arguments: %s",
2365 strerror(rte_errno));
2368 for (i = 0; i < eth_da.nb_representor_ports; ++i)
2369 if (eth_da.representor_ports[i] ==
2370 (uint16_t)switch_info->port_name)
2372 if (i == eth_da.nb_representor_ports) {
2377 /* Build device name. */
2378 if (spawn->pf_bond < 0) {
2379 /* Single device. */
2380 if (!switch_info->representor)
2381 strlcpy(name, dpdk_dev->name, sizeof(name));
2383 snprintf(name, sizeof(name), "%s_representor_%u",
2384 dpdk_dev->name, switch_info->port_name);
2386 /* Bonding device. */
2387 if (!switch_info->representor)
2388 snprintf(name, sizeof(name), "%s_%s",
2389 dpdk_dev->name, spawn->ibv_dev->name);
2391 snprintf(name, sizeof(name), "%s_%s_representor_%u",
2392 dpdk_dev->name, spawn->ibv_dev->name,
2393 switch_info->port_name);
2395 /* check if the device is already spawned */
2396 if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2400 DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2401 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2402 struct mlx5_mp_id mp_id;
2404 eth_dev = rte_eth_dev_attach_secondary(name);
2405 if (eth_dev == NULL) {
2406 DRV_LOG(ERR, "can not attach rte ethdev");
2410 eth_dev->device = dpdk_dev;
2411 eth_dev->dev_ops = &mlx5_dev_sec_ops;
2412 err = mlx5_proc_priv_init(eth_dev);
2415 mp_id.port_id = eth_dev->data->port_id;
2416 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
2417 /* Receive command fd from primary process */
2418 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
2421 /* Remap UAR for Tx queues. */
2422 err = mlx5_tx_uar_init_secondary(eth_dev, err);
2426 * Ethdev pointer is still required as input since
2427 * the primary device is not accessible from the
2428 * secondary process.
2430 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2431 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2434 mlx5_dev_close(eth_dev);
2438 * Some parameters ("tx_db_nc" in particularly) are needed in
2439 * advance to create dv/verbs device context. We proceed the
2440 * devargs here to get ones, and later proceed devargs again
2441 * to override some hardware settings.
2443 err = mlx5_args(&config, dpdk_dev->devargs);
2446 DRV_LOG(ERR, "failed to process device arguments: %s",
2447 strerror(rte_errno));
2450 sh = mlx5_alloc_shared_ibctx(spawn, &config);
2453 config.devx = sh->devx;
2454 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2455 config.dest_tir = 1;
2457 #ifdef HAVE_IBV_MLX5_MOD_SWP
2458 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2461 * Multi-packet send is supported by ConnectX-4 Lx PF as well
2462 * as all ConnectX-5 devices.
2464 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2465 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2467 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2468 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2470 mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2471 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2472 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2473 DRV_LOG(DEBUG, "enhanced MPW is supported");
2474 mps = MLX5_MPW_ENHANCED;
2476 DRV_LOG(DEBUG, "MPW is supported");
2480 DRV_LOG(DEBUG, "MPW isn't supported");
2481 mps = MLX5_MPW_DISABLED;
2483 #ifdef HAVE_IBV_MLX5_MOD_SWP
2484 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2485 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2486 DRV_LOG(DEBUG, "SWP support: %u", swp);
2489 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2490 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2491 struct mlx5dv_striding_rq_caps mprq_caps =
2492 dv_attr.striding_rq_caps;
2494 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2495 mprq_caps.min_single_stride_log_num_of_bytes);
2496 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2497 mprq_caps.max_single_stride_log_num_of_bytes);
2498 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2499 mprq_caps.min_single_wqe_log_num_of_strides);
2500 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2501 mprq_caps.max_single_wqe_log_num_of_strides);
2502 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2503 mprq_caps.supported_qpts);
2504 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2506 mprq_min_stride_size_n =
2507 mprq_caps.min_single_stride_log_num_of_bytes;
2508 mprq_max_stride_size_n =
2509 mprq_caps.max_single_stride_log_num_of_bytes;
2510 mprq_min_stride_num_n =
2511 mprq_caps.min_single_wqe_log_num_of_strides;
2512 mprq_max_stride_num_n =
2513 mprq_caps.max_single_wqe_log_num_of_strides;
2516 if (RTE_CACHE_LINE_SIZE == 128 &&
2517 !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2521 config.cqe_comp = cqe_comp;
2522 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2523 /* Whether device supports 128B Rx CQE padding. */
2524 cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2525 (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2527 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2528 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2529 tunnel_en = ((dv_attr.tunnel_offloads_caps &
2530 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2531 (dv_attr.tunnel_offloads_caps &
2532 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
2533 (dv_attr.tunnel_offloads_caps &
2534 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
2536 DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2537 tunnel_en ? "" : "not ");
2540 "tunnel offloading disabled due to old OFED/rdma-core version");
2542 config.tunnel_en = tunnel_en;
2543 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2544 mpls_en = ((dv_attr.tunnel_offloads_caps &
2545 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2546 (dv_attr.tunnel_offloads_caps &
2547 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2548 DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2549 mpls_en ? "" : "not ");
2551 DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2552 " old OFED/rdma-core version or firmware configuration");
2554 config.mpls_en = mpls_en;
2555 /* Check port status. */
2556 err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2558 DRV_LOG(ERR, "port query failed: %s", strerror(err));
2561 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2562 DRV_LOG(ERR, "port is not configured in Ethernet mode");
2566 if (port_attr.state != IBV_PORT_ACTIVE)
2567 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2568 mlx5_glue->port_state_str(port_attr.state),
2570 /* Allocate private eth device data. */
2571 priv = rte_zmalloc("ethdev private structure",
2573 RTE_CACHE_LINE_SIZE);
2575 DRV_LOG(ERR, "priv allocation failure");
2580 priv->ibv_port = spawn->ibv_port;
2581 priv->pci_dev = spawn->pci_dev;
2582 priv->mtu = RTE_ETHER_MTU;
2583 priv->mp_id.port_id = port_id;
2584 strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
2586 /* Initialize UAR access locks for 32bit implementations. */
2587 rte_spinlock_init(&priv->uar_lock_cq);
2588 for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2589 rte_spinlock_init(&priv->uar_lock[i]);
2591 /* Some internal functions rely on Netlink sockets, open them now. */
2592 priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2593 priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
2594 priv->representor = !!switch_info->representor;
2595 priv->master = !!switch_info->master;
2596 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2597 priv->vport_meta_tag = 0;
2598 priv->vport_meta_mask = 0;
2599 priv->pf_bond = spawn->pf_bond;
2600 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2602 * The DevX port query API is implemented. E-Switch may use
2603 * either vport or reg_c[0] metadata register to match on
2604 * vport index. The engaged part of metadata register is
2607 if (switch_info->representor || switch_info->master) {
2608 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2609 MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2610 err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2614 "can't query devx port %d on device %s",
2615 spawn->ibv_port, spawn->ibv_dev->name);
2616 devx_port.comp_mask = 0;
2619 if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2620 priv->vport_meta_tag = devx_port.reg_c_0.value;
2621 priv->vport_meta_mask = devx_port.reg_c_0.mask;
2622 if (!priv->vport_meta_mask) {
2623 DRV_LOG(ERR, "vport zero mask for port %d"
2624 " on bonding device %s",
2625 spawn->ibv_port, spawn->ibv_dev->name);
2629 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2630 DRV_LOG(ERR, "invalid vport tag for port %d"
2631 " on bonding device %s",
2632 spawn->ibv_port, spawn->ibv_dev->name);
2637 if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2638 priv->vport_id = devx_port.vport_num;
2639 } else if (spawn->pf_bond >= 0) {
2640 DRV_LOG(ERR, "can't deduce vport index for port %d"
2641 " on bonding device %s",
2642 spawn->ibv_port, spawn->ibv_dev->name);
2646 /* Suppose vport index in compatible way. */
2647 priv->vport_id = switch_info->representor ?
2648 switch_info->port_name + 1 : -1;
2652 * Kernel/rdma_core support single E-Switch per PF configurations
2653 * only and vport_id field contains the vport index for
2654 * associated VF, which is deduced from representor port name.
2655 * For example, let's have the IB device port 10, it has
2656 * attached network device eth0, which has port name attribute
2657 * pf0vf2, we can deduce the VF number as 2, and set vport index
2658 * as 3 (2+1). This assigning schema should be changed if the
2659 * multiple E-Switch instances per PF configurations or/and PCI
2660 * subfunctions are added.
2662 priv->vport_id = switch_info->representor ?
2663 switch_info->port_name + 1 : -1;
2665 /* representor_id field keeps the unmodified VF index. */
2666 priv->representor_id = switch_info->representor ?
2667 switch_info->port_name : -1;
2669 * Look for sibling devices in order to reuse their switch domain
2670 * if any, otherwise allocate one.
2672 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2673 const struct mlx5_priv *opriv =
2674 rte_eth_devices[port_id].data->dev_private;
2677 opriv->sh != priv->sh ||
2679 RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2681 priv->domain_id = opriv->domain_id;
2684 if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2685 err = rte_eth_switch_domain_alloc(&priv->domain_id);
2688 DRV_LOG(ERR, "unable to allocate switch domain: %s",
2689 strerror(rte_errno));
2694 /* Override some values set by hardware configuration. */
2695 mlx5_args(&config, dpdk_dev->devargs);
2696 err = mlx5_dev_check_sibling_config(priv, &config);
2699 config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2700 IBV_DEVICE_RAW_IP_CSUM);
2701 DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2702 (config.hw_csum ? "" : "not "));
2703 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2704 !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2705 DRV_LOG(DEBUG, "counters are not supported");
2707 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
2708 if (config.dv_flow_en) {
2709 DRV_LOG(WARNING, "DV flow is not supported");
2710 config.dv_flow_en = 0;
2713 config.ind_table_max_size =
2714 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2716 * Remove this check once DPDK supports larger/variable
2717 * indirection tables.
2719 if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2720 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2721 DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2722 config.ind_table_max_size);
2723 config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2724 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2725 DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2726 (config.hw_vlan_strip ? "" : "not "));
2727 config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2728 IBV_RAW_PACKET_CAP_SCATTER_FCS);
2729 DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2730 (config.hw_fcs_strip ? "" : "not "));
2731 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2732 hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2733 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2734 hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2735 IBV_DEVICE_PCI_WRITE_END_PADDING);
2737 if (config.hw_padding && !hw_padding) {
2738 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2739 config.hw_padding = 0;
2740 } else if (config.hw_padding) {
2741 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2743 config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2744 (sh->device_attr.tso_caps.supported_qpts &
2745 (1 << IBV_QPT_RAW_PACKET)));
2747 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2749 * MPW is disabled by default, while the Enhanced MPW is enabled
2752 if (config.mps == MLX5_ARG_UNSET)
2753 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2756 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2757 DRV_LOG(INFO, "%sMPS is %s",
2758 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
2759 config.mps == MLX5_MPW ? "legacy " : "",
2760 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2761 if (config.cqe_comp && !cqe_comp) {
2762 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2763 config.cqe_comp = 0;
2765 if (config.cqe_pad && !cqe_pad) {
2766 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2768 } else if (config.cqe_pad) {
2769 DRV_LOG(INFO, "Rx CQE padding is enabled");
2772 priv->counter_fallback = 0;
2773 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2778 if (!config.hca_attr.flow_counters_dump)
2779 priv->counter_fallback = 1;
2780 #ifndef HAVE_IBV_DEVX_ASYNC
2781 priv->counter_fallback = 1;
2783 if (priv->counter_fallback)
2784 DRV_LOG(INFO, "Use fall-back DV counter management");
2785 /* Check for LRO support. */
2786 if (config.dest_tir && config.hca_attr.lro_cap &&
2787 config.dv_flow_en) {
2788 /* TBD check tunnel lro caps. */
2789 config.lro.supported = config.hca_attr.lro_cap;
2790 DRV_LOG(DEBUG, "Device supports LRO");
2792 * If LRO timeout is not configured by application,
2793 * use the minimal supported value.
2795 if (!config.lro.timeout)
2796 config.lro.timeout =
2797 config.hca_attr.lro_timer_supported_periods[0];
2798 DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2799 config.lro.timeout);
2801 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2802 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2803 config.dv_flow_en) {
2804 uint8_t reg_c_mask =
2805 config.hca_attr.qos.flow_meter_reg_c_ids;
2807 * Meter needs two REG_C's for color match and pre-sfx
2808 * flow match. Here get the REG_C for color match.
2809 * REG_C_0 and REG_C_1 is reserved for metadata feature.
2812 if (__builtin_popcount(reg_c_mask) < 1) {
2814 DRV_LOG(WARNING, "No available register for"
2817 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2820 priv->mtr_reg_share =
2821 config.hca_attr.qos.flow_meter_reg_share;
2822 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2823 priv->mtr_color_reg);
2828 if (config.mprq.enabled && mprq) {
2829 if (config.mprq.stride_num_n &&
2830 (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2831 config.mprq.stride_num_n < mprq_min_stride_num_n)) {
2832 config.mprq.stride_num_n =
2833 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2834 mprq_min_stride_num_n),
2835 mprq_max_stride_num_n);
2837 "the number of strides"
2838 " for Multi-Packet RQ is out of range,"
2839 " setting default value (%u)",
2840 1 << config.mprq.stride_num_n);
2842 if (config.mprq.stride_size_n &&
2843 (config.mprq.stride_size_n > mprq_max_stride_size_n ||
2844 config.mprq.stride_size_n < mprq_min_stride_size_n)) {
2845 config.mprq.stride_size_n =
2846 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
2847 mprq_min_stride_size_n),
2848 mprq_max_stride_size_n);
2850 "the size of a stride"
2851 " for Multi-Packet RQ is out of range,"
2852 " setting default value (%u)",
2853 1 << config.mprq.stride_size_n);
2855 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2856 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2857 } else if (config.mprq.enabled && !mprq) {
2858 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2859 config.mprq.enabled = 0;
2861 if (config.max_dump_files_num == 0)
2862 config.max_dump_files_num = 128;
2863 eth_dev = rte_eth_dev_allocate(name);
2864 if (eth_dev == NULL) {
2865 DRV_LOG(ERR, "can not allocate rte ethdev");
2869 /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2870 eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2871 if (priv->representor) {
2872 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2873 eth_dev->data->representor_id = priv->representor_id;
2876 * Store associated network device interface index. This index
2877 * is permanent throughout the lifetime of device. So, we may store
2878 * the ifindex here and use the cached value further.
2880 MLX5_ASSERT(spawn->ifindex);
2881 priv->if_index = spawn->ifindex;
2882 eth_dev->data->dev_private = priv;
2883 priv->dev_data = eth_dev->data;
2884 eth_dev->data->mac_addrs = priv->mac;
2885 eth_dev->device = dpdk_dev;
2886 /* Configure the first MAC address by default. */
2887 if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2889 "port %u cannot get MAC address, is mlx5_en"
2890 " loaded? (errno: %s)",
2891 eth_dev->data->port_id, strerror(rte_errno));
2896 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2897 eth_dev->data->port_id,
2898 mac.addr_bytes[0], mac.addr_bytes[1],
2899 mac.addr_bytes[2], mac.addr_bytes[3],
2900 mac.addr_bytes[4], mac.addr_bytes[5]);
2901 #ifdef RTE_LIBRTE_MLX5_DEBUG
2903 char ifname[IF_NAMESIZE];
2905 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2906 DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2907 eth_dev->data->port_id, ifname);
2909 DRV_LOG(DEBUG, "port %u ifname is unknown",
2910 eth_dev->data->port_id);
2913 /* Get actual MTU if possible. */
2914 err = mlx5_get_mtu(eth_dev, &priv->mtu);
2919 DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2921 /* Initialize burst functions to prevent crashes before link-up. */
2922 eth_dev->rx_pkt_burst = removed_rx_burst;
2923 eth_dev->tx_pkt_burst = removed_tx_burst;
2924 eth_dev->dev_ops = &mlx5_dev_ops;
2925 /* Register MAC address. */
2926 claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2927 if (config.vf && config.vf_nl_en)
2928 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
2929 mlx5_ifindex(eth_dev),
2930 eth_dev->data->mac_addrs,
2931 MLX5_MAX_MAC_ADDRESSES);
2933 priv->ctrl_flows = 0;
2934 TAILQ_INIT(&priv->flow_meters);
2935 TAILQ_INIT(&priv->flow_meter_profiles);
2936 /* Hint libmlx5 to use PMD allocator for data plane resources */
2937 struct mlx5dv_ctx_allocators alctr = {
2938 .alloc = &mlx5_alloc_verbs_buf,
2939 .free = &mlx5_free_verbs_buf,
2942 mlx5_glue->dv_set_context_attr(sh->ctx,
2943 MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2944 (void *)((uintptr_t)&alctr));
2945 /* Bring Ethernet device up. */
2946 DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2947 eth_dev->data->port_id);
2948 mlx5_set_link_up(eth_dev);
2950 * Even though the interrupt handler is not installed yet,
2951 * interrupts will still trigger on the async_fd from
2952 * Verbs context returned by ibv_open_device().
2954 mlx5_link_update(eth_dev, 0);
2955 #ifdef HAVE_MLX5DV_DR_ESWITCH
2956 if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2957 (switch_info->representor || switch_info->master)))
2958 config.dv_esw_en = 0;
2960 config.dv_esw_en = 0;
2962 /* Detect minimal data bytes to inline. */
2963 mlx5_set_min_inline(spawn, &config);
2964 /* Store device configuration on private structure. */
2965 priv->config = config;
2966 /* Create context for virtual machine VLAN workaround. */
2967 priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2968 if (config.dv_flow_en) {
2969 err = mlx5_alloc_shared_dr(priv);
2973 * RSS id is shared with meter flow id. Meter flow id can only
2974 * use the 24 MSB of the register.
2976 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
2977 MLX5_MTR_COLOR_BITS);
2978 if (!priv->qrss_id_pool) {
2979 DRV_LOG(ERR, "can't create flow id pool");
2984 /* Supported Verbs flow priority number detection. */
2985 err = mlx5_flow_discover_priorities(eth_dev);
2990 priv->config.flow_prio = err;
2991 if (!priv->config.dv_esw_en &&
2992 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2993 DRV_LOG(WARNING, "metadata mode %u is not supported "
2994 "(no E-Switch)", priv->config.dv_xmeta_en);
2995 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2997 mlx5_set_metadata_mask(eth_dev);
2998 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2999 !priv->sh->dv_regc0_mask) {
3000 DRV_LOG(ERR, "metadata mode %u is not supported "
3001 "(no metadata reg_c[0] is available)",
3002 priv->config.dv_xmeta_en);
3007 * Allocate the buffer for flow creating, just once.
3008 * The allocation must be done before any flow creating.
3010 mlx5_flow_alloc_intermediate(eth_dev);
3011 /* Query availibility of metadata reg_c's. */
3012 err = mlx5_flow_discover_mreg_c(eth_dev);
3017 if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
3019 "port %u extensive metadata register is not supported",
3020 eth_dev->data->port_id);
3021 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
3022 DRV_LOG(ERR, "metadata mode %u is not supported "
3023 "(no metadata registers available)",
3024 priv->config.dv_xmeta_en);
3029 if (priv->config.dv_flow_en &&
3030 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
3031 mlx5_flow_ext_mreg_supported(eth_dev) &&
3032 priv->sh->dv_regc0_mask) {
3033 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
3034 MLX5_FLOW_MREG_HTABLE_SZ);
3035 if (!priv->mreg_cp_tbl) {
3043 if (priv->mreg_cp_tbl)
3044 mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
3046 mlx5_free_shared_dr(priv);
3047 if (priv->nl_socket_route >= 0)
3048 close(priv->nl_socket_route);
3049 if (priv->nl_socket_rdma >= 0)
3050 close(priv->nl_socket_rdma);
3051 if (priv->vmwa_context)
3052 mlx5_vlan_vmwa_exit(priv->vmwa_context);
3053 if (priv->qrss_id_pool)
3054 mlx5_flow_id_pool_release(priv->qrss_id_pool);
3056 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
3058 if (eth_dev != NULL)
3059 eth_dev->data->dev_private = NULL;
3061 if (eth_dev != NULL) {
3062 /* mac_addrs must not be freed alone because part of dev_private */
3063 eth_dev->data->mac_addrs = NULL;
3064 rte_eth_dev_release_port(eth_dev);
3067 mlx5_free_shared_ibctx(sh);
3068 MLX5_ASSERT(err > 0);
3074 * Comparison callback to sort device data.
3076 * This is meant to be used with qsort().
3079 * Pointer to pointer to first data object.
3081 * Pointer to pointer to second data object.
3084 * 0 if both objects are equal, less than 0 if the first argument is less
3085 * than the second, greater than 0 otherwise.
3088 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
3090 const struct mlx5_switch_info *si_a =
3091 &((const struct mlx5_dev_spawn_data *)a)->info;
3092 const struct mlx5_switch_info *si_b =
3093 &((const struct mlx5_dev_spawn_data *)b)->info;
3096 /* Master device first. */
3097 ret = si_b->master - si_a->master;
3100 /* Then representor devices. */
3101 ret = si_b->representor - si_a->representor;
3104 /* Unidentified devices come last in no specific order. */
3105 if (!si_a->representor)
3107 /* Order representors by name. */
3108 return si_a->port_name - si_b->port_name;
3112 * Match PCI information for possible slaves of bonding device.
3114 * @param[in] ibv_dev
3115 * Pointer to Infiniband device structure.
3116 * @param[in] pci_dev
3117 * Pointer to PCI device structure to match PCI address.
3118 * @param[in] nl_rdma
3119 * Netlink RDMA group socket handle.
3122 * negative value if no bonding device found, otherwise
3123 * positive index of slave PF in bonding.
3126 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
3127 const struct rte_pci_device *pci_dev,
3130 char ifname[IF_NAMESIZE + 1];
3131 unsigned int ifindex;
3137 * Try to get master device name. If something goes
3138 * wrong suppose the lack of kernel support and no
3143 if (!strstr(ibv_dev->name, "bond"))
3145 np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
3149 * The Master device might not be on the predefined
3150 * port (not on port index 1, it is not garanted),
3151 * we have to scan all Infiniband device port and
3154 for (i = 1; i <= np; ++i) {
3155 /* Check whether Infiniband port is populated. */
3156 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
3159 if (!if_indextoname(ifindex, ifname))
3161 /* Try to read bonding slave names from sysfs. */
3163 "/sys/class/net/%s/master/bonding/slaves", ifname);
3164 file = fopen(slaves, "r");
3170 /* Use safe format to check maximal buffer length. */
3171 MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
3172 while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
3173 char tmp_str[IF_NAMESIZE + 32];
3174 struct rte_pci_addr pci_addr;
3175 struct mlx5_switch_info info;
3177 /* Process slave interface names in the loop. */
3178 snprintf(tmp_str, sizeof(tmp_str),
3179 "/sys/class/net/%s", ifname);
3180 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
3181 DRV_LOG(WARNING, "can not get PCI address"
3182 " for netdev \"%s\"", ifname);
3185 if (pci_dev->addr.domain != pci_addr.domain ||
3186 pci_dev->addr.bus != pci_addr.bus ||
3187 pci_dev->addr.devid != pci_addr.devid ||
3188 pci_dev->addr.function != pci_addr.function)
3190 /* Slave interface PCI address match found. */
3192 snprintf(tmp_str, sizeof(tmp_str),
3193 "/sys/class/net/%s/phys_port_name", ifname);
3194 file = fopen(tmp_str, "rb");
3197 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
3198 if (fscanf(file, "%32s", tmp_str) == 1)
3199 mlx5_translate_port_name(tmp_str, &info);
3200 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
3201 info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
3202 pf = info.port_name;
3211 * DPDK callback to register a PCI device.
3213 * This function spawns Ethernet devices out of a given PCI device.
3215 * @param[in] pci_drv
3216 * PCI driver structure (mlx5_driver).
3217 * @param[in] pci_dev
3218 * PCI device information.
3221 * 0 on success, a negative errno value otherwise and rte_errno is set.
3224 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
3225 struct rte_pci_device *pci_dev)
3227 struct ibv_device **ibv_list;
3229 * Number of found IB Devices matching with requested PCI BDF.
3230 * nd != 1 means there are multiple IB devices over the same
3231 * PCI device and we have representors and master.
3233 unsigned int nd = 0;
3235 * Number of found IB device Ports. nd = 1 and np = 1..n means
3236 * we have the single multiport IB device, and there may be
3237 * representors attached to some of found ports.
3239 unsigned int np = 0;
3241 * Number of DPDK ethernet devices to Spawn - either over
3242 * multiple IB devices or multiple ports of single IB device.
3243 * Actually this is the number of iterations to spawn.
3245 unsigned int ns = 0;
3248 * < 0 - no bonding device (single one)
3249 * >= 0 - bonding device (value is slave PF index)
3252 struct mlx5_dev_spawn_data *list = NULL;
3253 struct mlx5_dev_config dev_config;
3256 if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
3257 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
3261 if (rte_eal_process_type() == RTE_PROC_PRIMARY)
3262 mlx5_pmd_socket_init();
3263 ret = mlx5_init_once();
3265 DRV_LOG(ERR, "unable to init PMD global data: %s",
3266 strerror(rte_errno));
3269 MLX5_ASSERT(pci_drv == &mlx5_driver);
3271 ibv_list = mlx5_glue->get_device_list(&ret);
3273 rte_errno = errno ? errno : ENOSYS;
3274 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
3278 * First scan the list of all Infiniband devices to find
3279 * matching ones, gathering into the list.
3281 struct ibv_device *ibv_match[ret + 1];
3282 int nl_route = mlx5_nl_init(NETLINK_ROUTE);
3283 int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
3287 struct rte_pci_addr pci_addr;
3289 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
3290 bd = mlx5_device_bond_pci_match
3291 (ibv_list[ret], pci_dev, nl_rdma);
3294 * Bonding device detected. Only one match is allowed,
3295 * the bonding is supported over multi-port IB device,
3296 * there should be no matches on representor PCI
3297 * functions or non VF LAG bonding devices with
3298 * specified address.
3302 "multiple PCI match on bonding device"
3303 "\"%s\" found", ibv_list[ret]->name);
3308 DRV_LOG(INFO, "PCI information matches for"
3309 " slave %d bonding device \"%s\"",
3310 bd, ibv_list[ret]->name);
3311 ibv_match[nd++] = ibv_list[ret];
3314 if (mlx5_dev_to_pci_addr
3315 (ibv_list[ret]->ibdev_path, &pci_addr))
3317 if (pci_dev->addr.domain != pci_addr.domain ||
3318 pci_dev->addr.bus != pci_addr.bus ||
3319 pci_dev->addr.devid != pci_addr.devid ||
3320 pci_dev->addr.function != pci_addr.function)
3322 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
3323 ibv_list[ret]->name);
3324 ibv_match[nd++] = ibv_list[ret];
3326 ibv_match[nd] = NULL;
3328 /* No device matches, just complain and bail out. */
3330 "no Verbs device matches PCI device " PCI_PRI_FMT ","
3331 " are kernel drivers loaded?",
3332 pci_dev->addr.domain, pci_dev->addr.bus,
3333 pci_dev->addr.devid, pci_dev->addr.function);
3340 * Found single matching device may have multiple ports.
3341 * Each port may be representor, we have to check the port
3342 * number and check the representors existence.
3345 np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
3347 DRV_LOG(WARNING, "can not get IB device \"%s\""
3348 " ports number", ibv_match[0]->name);
3349 if (bd >= 0 && !np) {
3350 DRV_LOG(ERR, "can not get ports"
3351 " for bonding device");
3357 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
3360 * This may happen if there is VF LAG kernel support and
3361 * application is compiled with older rdma_core library.
3364 "No kernel/verbs support for VF LAG bonding found.");
3365 rte_errno = ENOTSUP;
3371 * Now we can determine the maximal
3372 * amount of devices to be spawned.
3374 list = rte_zmalloc("device spawn data",
3375 sizeof(struct mlx5_dev_spawn_data) *
3377 RTE_CACHE_LINE_SIZE);
3379 DRV_LOG(ERR, "spawn data array allocation failure");
3384 if (bd >= 0 || np > 1) {
3386 * Single IB device with multiple ports found,
3387 * it may be E-Switch master device and representors.
3388 * We have to perform identification through the ports.
3390 MLX5_ASSERT(nl_rdma >= 0);
3391 MLX5_ASSERT(ns == 0);
3392 MLX5_ASSERT(nd == 1);
3394 for (i = 1; i <= np; ++i) {
3395 list[ns].max_port = np;
3396 list[ns].ibv_port = i;
3397 list[ns].ibv_dev = ibv_match[0];
3398 list[ns].eth_dev = NULL;
3399 list[ns].pci_dev = pci_dev;
3400 list[ns].pf_bond = bd;
3401 list[ns].ifindex = mlx5_nl_ifindex
3402 (nl_rdma, list[ns].ibv_dev->name, i);
3403 if (!list[ns].ifindex) {
3405 * No network interface index found for the
3406 * specified port, it means there is no
3407 * representor on this port. It's OK,
3408 * there can be disabled ports, for example
3409 * if sriov_numvfs < sriov_totalvfs.
3415 ret = mlx5_nl_switch_info
3419 if (ret || (!list[ns].info.representor &&
3420 !list[ns].info.master)) {
3422 * We failed to recognize representors with
3423 * Netlink, let's try to perform the task
3426 ret = mlx5_sysfs_switch_info
3430 if (!ret && bd >= 0) {
3431 switch (list[ns].info.name_type) {
3432 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
3433 if (list[ns].info.port_name == bd)
3436 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
3437 if (list[ns].info.pf_num == bd)
3445 if (!ret && (list[ns].info.representor ^
3446 list[ns].info.master))
3451 "unable to recognize master/representors"
3452 " on the IB device with multiple ports");
3459 * The existence of several matching entries (nd > 1) means
3460 * port representors have been instantiated. No existing Verbs
3461 * call nor sysfs entries can tell them apart, this can only
3462 * be done through Netlink calls assuming kernel drivers are
3463 * recent enough to support them.
3465 * In the event of identification failure through Netlink,
3466 * try again through sysfs, then:
3468 * 1. A single IB device matches (nd == 1) with single
3469 * port (np=0/1) and is not a representor, assume
3470 * no switch support.
3472 * 2. Otherwise no safe assumptions can be made;
3473 * complain louder and bail out.
3476 for (i = 0; i != nd; ++i) {
3477 memset(&list[ns].info, 0, sizeof(list[ns].info));
3478 list[ns].max_port = 1;
3479 list[ns].ibv_port = 1;
3480 list[ns].ibv_dev = ibv_match[i];
3481 list[ns].eth_dev = NULL;
3482 list[ns].pci_dev = pci_dev;
3483 list[ns].pf_bond = -1;
3484 list[ns].ifindex = 0;
3486 list[ns].ifindex = mlx5_nl_ifindex
3487 (nl_rdma, list[ns].ibv_dev->name, 1);
3488 if (!list[ns].ifindex) {
3489 char ifname[IF_NAMESIZE];
3492 * Netlink failed, it may happen with old
3493 * ib_core kernel driver (before 4.16).
3494 * We can assume there is old driver because
3495 * here we are processing single ports IB
3496 * devices. Let's try sysfs to retrieve
3497 * the ifindex. The method works for
3498 * master device only.
3502 * Multiple devices found, assume
3503 * representors, can not distinguish
3504 * master/representor and retrieve
3505 * ifindex via sysfs.
3509 ret = mlx5_get_master_ifname
3510 (ibv_match[i]->ibdev_path, &ifname);
3513 if_nametoindex(ifname);
3514 if (!list[ns].ifindex) {
3516 * No network interface index found
3517 * for the specified device, it means
3518 * there it is neither representor
3526 ret = mlx5_nl_switch_info
3530 if (ret || (!list[ns].info.representor &&
3531 !list[ns].info.master)) {
3533 * We failed to recognize representors with
3534 * Netlink, let's try to perform the task
3537 ret = mlx5_sysfs_switch_info
3541 if (!ret && (list[ns].info.representor ^
3542 list[ns].info.master)) {
3544 } else if ((nd == 1) &&
3545 !list[ns].info.representor &&
3546 !list[ns].info.master) {
3548 * Single IB device with
3549 * one physical port and
3550 * attached network device.
3551 * May be SRIOV is not enabled
3552 * or there is no representors.
3554 DRV_LOG(INFO, "no E-Switch support detected");
3561 "unable to recognize master/representors"
3562 " on the multiple IB devices");
3570 * Sort list to probe devices in natural order for users convenience
3571 * (i.e. master first, then representors from lowest to highest ID).
3573 qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3574 /* Default configuration. */
3575 dev_config = (struct mlx5_dev_config){
3577 .mps = MLX5_ARG_UNSET,
3578 .dbnc = MLX5_ARG_UNSET,
3580 .txq_inline_max = MLX5_ARG_UNSET,
3581 .txq_inline_min = MLX5_ARG_UNSET,
3582 .txq_inline_mpw = MLX5_ARG_UNSET,
3583 .txqs_inline = MLX5_ARG_UNSET,
3585 .mr_ext_memseg_en = 1,
3587 .enabled = 0, /* Disabled by default. */
3590 .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3591 .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3595 .log_hp_size = MLX5_ARG_UNSET,
3597 /* Device specific configuration. */
3598 switch (pci_dev->id.device_id) {
3599 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3600 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3601 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3602 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3603 case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3604 case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3605 case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3611 for (i = 0; i != ns; ++i) {
3614 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3617 if (!list[i].eth_dev) {
3618 if (rte_errno != EBUSY && rte_errno != EEXIST)
3620 /* Device is disabled or already spawned. Ignore it. */
3623 restore = list[i].eth_dev->data->dev_flags;
3624 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3625 /* Restore non-PCI flags cleared by the above call. */
3626 list[i].eth_dev->data->dev_flags |= restore;
3627 mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3628 rte_eth_dev_probing_finish(list[i].eth_dev);
3632 "probe of PCI device " PCI_PRI_FMT " aborted after"
3633 " encountering an error: %s",
3634 pci_dev->addr.domain, pci_dev->addr.bus,
3635 pci_dev->addr.devid, pci_dev->addr.function,
3636 strerror(rte_errno));
3640 if (!list[i].eth_dev)
3642 mlx5_dev_close(list[i].eth_dev);
3643 /* mac_addrs must not be freed because in dev_private */
3644 list[i].eth_dev->data->mac_addrs = NULL;
3645 claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3647 /* Restore original error. */
3654 * Do the routine cleanup:
3655 * - close opened Netlink sockets
3656 * - free allocated spawn data array
3657 * - free the Infiniband device list
3665 MLX5_ASSERT(ibv_list);
3666 mlx5_glue->free_device_list(ibv_list);
3671 * Look for the ethernet device belonging to mlx5 driver.
3673 * @param[in] port_id
3674 * port_id to start looking for device.
3675 * @param[in] pci_dev
3676 * Pointer to the hint PCI device. When device is being probed
3677 * the its siblings (master and preceding representors might
3678 * not have assigned driver yet (because the mlx5_pci_probe()
3679 * is not completed yet, for this case match on hint PCI
3680 * device may be used to detect sibling device.
3683 * port_id of found device, RTE_MAX_ETHPORT if not found.
3686 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3688 while (port_id < RTE_MAX_ETHPORTS) {
3689 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3691 if (dev->state != RTE_ETH_DEV_UNUSED &&
3693 (dev->device == &pci_dev->device ||
3694 (dev->device->driver &&
3695 dev->device->driver->name &&
3696 !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3700 if (port_id >= RTE_MAX_ETHPORTS)
3701 return RTE_MAX_ETHPORTS;
3706 * DPDK callback to remove a PCI device.
3708 * This function removes all Ethernet devices belong to a given PCI device.
3710 * @param[in] pci_dev
3711 * Pointer to the PCI device.
3714 * 0 on success, the function cannot fail.
3717 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3721 RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
3723 * mlx5_dev_close() is not registered to secondary process,
3724 * call the close function explicitly for secondary process.
3726 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
3727 mlx5_dev_close(&rte_eth_devices[port_id]);
3729 rte_eth_dev_close(port_id);
3734 static const struct rte_pci_id mlx5_pci_id_map[] = {
3736 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3737 PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3740 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3741 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3744 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3745 PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3748 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3749 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3752 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3753 PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3756 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3757 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3760 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3761 PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3764 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3765 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3768 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3769 PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3772 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3773 PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3776 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3777 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3780 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3781 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3784 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3785 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3788 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3789 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3792 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3793 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
3800 static struct rte_pci_driver mlx5_driver = {
3802 .name = MLX5_DRIVER_NAME
3804 .id_table = mlx5_pci_id_map,
3805 .probe = mlx5_pci_probe,
3806 .remove = mlx5_pci_remove,
3807 .dma_map = mlx5_dma_map,
3808 .dma_unmap = mlx5_dma_unmap,
3809 .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3810 RTE_PCI_DRV_PROBE_AGAIN,
3814 * Driver initialization routine.
3816 RTE_INIT(rte_mlx5_pmd_init)
3818 /* Initialize driver log type. */
3819 mlx5_logtype = rte_log_register("pmd.net.mlx5");
3820 if (mlx5_logtype >= 0)
3821 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3823 /* Build the static tables for Verbs conversion. */
3824 mlx5_set_ptype_table();
3825 mlx5_set_cksum_table();
3826 mlx5_set_swp_types_table();
3828 rte_pci_register(&mlx5_driver);
3831 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3832 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3833 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");