1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
15 #include <linux/rtnetlink.h>
18 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
20 #pragma GCC diagnostic ignored "-Wpedantic"
22 #include <infiniband/verbs.h>
24 #pragma GCC diagnostic error "-Wpedantic"
27 #include <rte_malloc.h>
28 #include <rte_ethdev_driver.h>
29 #include <rte_ethdev_pci.h>
31 #include <rte_bus_pci.h>
32 #include <rte_common.h>
33 #include <rte_config.h>
34 #include <rte_kvargs.h>
35 #include <rte_rwlock.h>
36 #include <rte_spinlock.h>
37 #include <rte_string_fns.h>
38 #include <rte_alarm.h>
40 #include <mlx5_glue.h>
41 #include <mlx5_devx_cmds.h>
42 #include <mlx5_common.h>
44 #include "mlx5_defs.h"
46 #include "mlx5_utils.h"
47 #include "mlx5_rxtx.h"
48 #include "mlx5_autoconf.h"
50 #include "mlx5_flow.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 limit the size of memcpy'd packet for MPRQ. */
68 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
70 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
71 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
73 /* Device parameter to configure inline send. Deprecated, ignored.*/
74 #define MLX5_TXQ_INLINE "txq_inline"
76 /* Device parameter to limit packet size to inline with ordinary SEND. */
77 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
79 /* Device parameter to configure minimal data size to inline. */
80 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
82 /* Device parameter to limit packet size to inline with Enhanced MPW. */
83 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
86 * Device parameter to configure the number of TX queues threshold for
87 * enabling inline send.
89 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
92 * Device parameter to configure the number of TX queues threshold for
93 * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
95 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
97 /* Device parameter to enable multi-packet send WQEs. */
98 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
101 * Device parameter to force doorbell register mapping
102 * to non-cahed region eliminating the extra write memory barrier.
104 #define MLX5_TX_DB_NC "tx_db_nc"
107 * Device parameter to include 2 dsegs in the title WQEBB.
108 * Deprecated, ignored.
110 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
113 * Device parameter to limit the size of inlining packet.
114 * Deprecated, ignored.
116 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
119 * Device parameter to enable hardware Tx vector.
120 * Deprecated, ignored (no vectorized Tx routines anymore).
122 #define MLX5_TX_VEC_EN "tx_vec_en"
124 /* Device parameter to enable hardware Rx vector. */
125 #define MLX5_RX_VEC_EN "rx_vec_en"
127 /* Allow L3 VXLAN flow creation. */
128 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
130 /* Activate DV E-Switch flow steering. */
131 #define MLX5_DV_ESW_EN "dv_esw_en"
133 /* Activate DV flow steering. */
134 #define MLX5_DV_FLOW_EN "dv_flow_en"
136 /* Enable extensive flow metadata support. */
137 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
139 /* Activate Netlink support in VF mode. */
140 #define MLX5_VF_NL_EN "vf_nl_en"
142 /* Enable extending memsegs when creating a MR. */
143 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
145 /* Select port representors to instantiate. */
146 #define MLX5_REPRESENTOR "representor"
148 /* Device parameter to configure the maximum number of dump files per queue. */
149 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
151 /* Configure timeout of LRO session (in microseconds). */
152 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
154 #ifndef HAVE_IBV_MLX5_MOD_MPW
155 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
156 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
159 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
160 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
163 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
165 /* Shared memory between primary and secondary processes. */
166 struct mlx5_shared_data *mlx5_shared_data;
168 /* Spinlock for mlx5_shared_data allocation. */
169 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
171 /* Process local data for secondary processes. */
172 static struct mlx5_local_data mlx5_local_data;
174 /** Driver-specific log messages type. */
177 /** Data associated with devices to spawn. */
178 struct mlx5_dev_spawn_data {
179 uint32_t ifindex; /**< Network interface index. */
180 uint32_t max_port; /**< IB device maximal port index. */
181 uint32_t ibv_port; /**< IB device physical port index. */
182 int pf_bond; /**< bonding device PF index. < 0 - no bonding */
183 struct mlx5_switch_info info; /**< Switch information. */
184 struct ibv_device *ibv_dev; /**< Associated IB device. */
185 struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
186 struct rte_pci_device *pci_dev; /**< Backend PCI device. */
189 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
190 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
192 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
193 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
195 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
196 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
199 * Allocate ID pool structure.
202 * The maximum id can be allocated from the pool.
205 * Pointer to pool object, NULL value otherwise.
207 struct mlx5_flow_id_pool *
208 mlx5_flow_id_pool_alloc(uint32_t max_id)
210 struct mlx5_flow_id_pool *pool;
213 pool = rte_zmalloc("id pool allocation", sizeof(*pool),
214 RTE_CACHE_LINE_SIZE);
216 DRV_LOG(ERR, "can't allocate id pool");
220 mem = rte_zmalloc("", MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
221 RTE_CACHE_LINE_SIZE);
223 DRV_LOG(ERR, "can't allocate mem for id pool");
227 pool->free_arr = mem;
228 pool->curr = pool->free_arr;
229 pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
230 pool->base_index = 0;
231 pool->max_id = max_id;
239 * Release ID pool structure.
242 * Pointer to flow id pool object to free.
245 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
247 rte_free(pool->free_arr);
255 * Pointer to flow id pool.
260 * 0 on success, error value otherwise.
263 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
265 if (pool->curr == pool->free_arr) {
266 if (pool->base_index == pool->max_id) {
268 DRV_LOG(ERR, "no free id");
271 *id = ++pool->base_index;
274 *id = *(--pool->curr);
282 * Pointer to flow id pool.
287 * 0 on success, error value otherwise.
290 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
296 if (pool->curr == pool->last) {
297 size = pool->curr - pool->free_arr;
298 size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
299 assert(size2 > size);
300 mem = rte_malloc("", size2 * sizeof(uint32_t), 0);
302 DRV_LOG(ERR, "can't allocate mem for id pool");
306 memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
307 rte_free(pool->free_arr);
308 pool->free_arr = mem;
309 pool->curr = pool->free_arr + size;
310 pool->last = pool->free_arr + size2;
318 * Initialize the counters management structure.
321 * Pointer to mlx5_ibv_shared object to free
324 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
328 TAILQ_INIT(&sh->cmng.flow_counters);
329 for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
330 TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
334 * Destroy all the resources allocated for a counter memory management.
337 * Pointer to the memory management structure.
340 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
342 uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
344 LIST_REMOVE(mng, next);
345 claim_zero(mlx5_devx_cmd_destroy(mng->dm));
346 claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
351 * Close and release all the resources of the counters management.
354 * Pointer to mlx5_ibv_shared object to free.
357 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
359 struct mlx5_counter_stats_mem_mng *mng;
366 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
367 if (rte_errno != EINPROGRESS)
371 for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
372 struct mlx5_flow_counter_pool *pool;
373 uint32_t batch = !!(i % 2);
375 if (!sh->cmng.ccont[i].pools)
377 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
382 (mlx5_devx_cmd_destroy(pool->min_dcs));
384 for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
385 if (pool->counters_raw[j].action)
387 (mlx5_glue->destroy_flow_action
388 (pool->counters_raw[j].action));
389 if (!batch && pool->counters_raw[j].dcs)
390 claim_zero(mlx5_devx_cmd_destroy
391 (pool->counters_raw[j].dcs));
393 TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
396 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
398 rte_free(sh->cmng.ccont[i].pools);
400 mng = LIST_FIRST(&sh->cmng.mem_mngs);
402 mlx5_flow_destroy_counter_stat_mem_mng(mng);
403 mng = LIST_FIRST(&sh->cmng.mem_mngs);
405 memset(&sh->cmng, 0, sizeof(sh->cmng));
409 * Extract pdn of PD object using DV API.
412 * Pointer to the verbs PD object.
414 * Pointer to the PD object number variable.
417 * 0 on success, error value otherwise.
419 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
421 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
423 struct mlx5dv_obj obj;
424 struct mlx5dv_pd pd_info;
428 obj.pd.out = &pd_info;
429 ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
431 DRV_LOG(DEBUG, "Fail to get PD object info");
437 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
440 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
445 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
446 /* Get environment variable to store. */
447 env = getenv(MLX5_SHUT_UP_BF);
448 value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
449 if (config->dbnc == MLX5_ARG_UNSET)
450 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
452 setenv(MLX5_SHUT_UP_BF,
453 config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
458 mlx5_restore_doorbell_mapping_env(int value)
460 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
461 /* Restore the original environment variable state. */
462 if (value == MLX5_ARG_UNSET)
463 unsetenv(MLX5_SHUT_UP_BF);
465 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
469 * Allocate shared IB device context. If there is multiport device the
470 * master and representors will share this context, if there is single
471 * port dedicated IB device, the context will be used by only given
472 * port due to unification.
474 * Routine first searches the context for the specified IB device name,
475 * if found the shared context assumed and reference counter is incremented.
476 * If no context found the new one is created and initialized with specified
477 * IB device context and parameters.
480 * Pointer to the IB device attributes (name, port, etc).
482 * Pointer to device configuration structure.
485 * Pointer to mlx5_ibv_shared object on success,
486 * otherwise NULL and rte_errno is set.
488 static struct mlx5_ibv_shared *
489 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
490 const struct mlx5_dev_config *config)
492 struct mlx5_ibv_shared *sh;
496 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
497 struct mlx5_devx_tis_attr tis_attr = { 0 };
501 /* Secondary process should not create the shared context. */
502 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
503 pthread_mutex_lock(&mlx5_ibv_list_mutex);
504 /* Search for IB context by device name. */
505 LIST_FOREACH(sh, &mlx5_ibv_list, next) {
506 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
511 /* No device found, we have to create new shared context. */
512 assert(spawn->max_port);
513 sh = rte_zmalloc("ethdev shared ib context",
514 sizeof(struct mlx5_ibv_shared) +
516 sizeof(struct mlx5_ibv_shared_port),
517 RTE_CACHE_LINE_SIZE);
519 DRV_LOG(ERR, "shared context allocation failure");
524 * Configure environment variable "MLX5_BF_SHUT_UP"
525 * before the device creation. The rdma_core library
526 * checks the variable at device creation and
527 * stores the result internally.
529 dbmap_env = mlx5_config_doorbell_mapping_env(config);
530 /* Try to open IB device with DV first, then usual Verbs. */
532 sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
535 DRV_LOG(DEBUG, "DevX is supported");
536 /* The device is created, no need for environment. */
537 mlx5_restore_doorbell_mapping_env(dbmap_env);
539 /* The environment variable is still configured. */
540 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
541 err = errno ? errno : ENODEV;
543 * The environment variable is not needed anymore,
544 * all device creation attempts are completed.
546 mlx5_restore_doorbell_mapping_env(dbmap_env);
549 DRV_LOG(DEBUG, "DevX is NOT supported");
551 err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
553 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
557 sh->max_port = spawn->max_port;
558 strncpy(sh->ibdev_name, sh->ctx->device->name,
559 sizeof(sh->ibdev_name));
560 strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
561 sizeof(sh->ibdev_path));
562 pthread_mutex_init(&sh->intr_mutex, NULL);
564 * Setting port_id to max unallowed value means
565 * there is no interrupt subhandler installed for
566 * the given port index i.
568 for (i = 0; i < sh->max_port; i++) {
569 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
570 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
572 sh->pd = mlx5_glue->alloc_pd(sh->ctx);
573 if (sh->pd == NULL) {
574 DRV_LOG(ERR, "PD allocation failure");
578 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
580 err = mlx5_get_pdn(sh->pd, &sh->pdn);
582 DRV_LOG(ERR, "Fail to extract pdn from PD");
585 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
587 DRV_LOG(ERR, "TD allocation failure");
591 tis_attr.transport_domain = sh->td->id;
592 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
594 DRV_LOG(ERR, "TIS allocation failure");
599 sh->flow_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX);
600 if (!sh->flow_id_pool) {
601 DRV_LOG(ERR, "can't create flow id pool");
605 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
607 * Once the device is added to the list of memory event
608 * callback, its global MR cache table cannot be expanded
609 * on the fly because of deadlock. If it overflows, lookup
610 * should be done by searching MR list linearly, which is slow.
612 * At this point the device is not added to the memory
613 * event list yet, context is just being created.
615 err = mlx5_mr_btree_init(&sh->mr.cache,
616 MLX5_MR_BTREE_CACHE_N * 2,
617 spawn->pci_dev->device.numa_node);
622 mlx5_flow_counters_mng_init(sh);
623 /* Add device to memory callback list. */
624 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
625 LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
627 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
628 /* Add context to the global device list. */
629 LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
631 pthread_mutex_unlock(&mlx5_ibv_list_mutex);
634 pthread_mutex_unlock(&mlx5_ibv_list_mutex);
637 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
639 claim_zero(mlx5_devx_cmd_destroy(sh->td));
641 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
643 claim_zero(mlx5_glue->close_device(sh->ctx));
644 if (sh->flow_id_pool)
645 mlx5_flow_id_pool_release(sh->flow_id_pool);
653 * Free shared IB device context. Decrement counter and if zero free
654 * all allocated resources and close handles.
657 * Pointer to mlx5_ibv_shared object to free
660 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
662 pthread_mutex_lock(&mlx5_ibv_list_mutex);
664 /* Check the object presence in the list. */
665 struct mlx5_ibv_shared *lctx;
667 LIST_FOREACH(lctx, &mlx5_ibv_list, next)
672 DRV_LOG(ERR, "Freeing non-existing shared IB context");
678 /* Secondary process should not free the shared context. */
679 assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
682 /* Release created Memory Regions. */
684 /* Remove from memory callback device list. */
685 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
686 LIST_REMOVE(sh, mem_event_cb);
687 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
688 /* Remove context from the global device list. */
689 LIST_REMOVE(sh, next);
691 * Ensure there is no async event handler installed.
692 * Only primary process handles async device events.
694 mlx5_flow_counters_mng_close(sh);
695 assert(!sh->intr_cnt);
697 mlx5_intr_callback_unregister
698 (&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
699 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
700 if (sh->devx_intr_cnt) {
701 if (sh->intr_handle_devx.fd)
702 rte_intr_callback_unregister(&sh->intr_handle_devx,
703 mlx5_dev_interrupt_handler_devx, sh);
705 mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
708 pthread_mutex_destroy(&sh->intr_mutex);
710 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
712 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
714 claim_zero(mlx5_devx_cmd_destroy(sh->td));
716 claim_zero(mlx5_glue->close_device(sh->ctx));
717 if (sh->flow_id_pool)
718 mlx5_flow_id_pool_release(sh->flow_id_pool);
721 pthread_mutex_unlock(&mlx5_ibv_list_mutex);
725 * Destroy table hash list and all the root entries per domain.
728 * Pointer to the private device data structure.
731 mlx5_free_table_hash_list(struct mlx5_priv *priv)
733 struct mlx5_ibv_shared *sh = priv->sh;
734 struct mlx5_flow_tbl_data_entry *tbl_data;
735 union mlx5_flow_tbl_key table_key = {
743 struct mlx5_hlist_entry *pos;
747 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
749 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
752 mlx5_hlist_remove(sh->flow_tbls, pos);
755 table_key.direction = 1;
756 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
758 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
761 mlx5_hlist_remove(sh->flow_tbls, pos);
764 table_key.direction = 0;
765 table_key.domain = 1;
766 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
768 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
771 mlx5_hlist_remove(sh->flow_tbls, pos);
774 mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
778 * Initialize flow table hash list and create the root tables entry
782 * Pointer to the private device data structure.
785 * Zero on success, positive error code otherwise.
788 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
790 struct mlx5_ibv_shared *sh = priv->sh;
791 char s[MLX5_HLIST_NAMESIZE];
795 snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
796 sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
797 if (!sh->flow_tbls) {
798 DRV_LOG(ERR, "flow tables with hash creation failed.\n");
802 #ifndef HAVE_MLX5DV_DR
804 * In case we have not DR support, the zero tables should be created
805 * because DV expect to see them even if they cannot be created by
808 union mlx5_flow_tbl_key table_key = {
816 struct mlx5_flow_tbl_data_entry *tbl_data = rte_zmalloc(NULL,
817 sizeof(*tbl_data), 0);
823 tbl_data->entry.key = table_key.v64;
824 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
827 rte_atomic32_init(&tbl_data->tbl.refcnt);
828 rte_atomic32_inc(&tbl_data->tbl.refcnt);
829 table_key.direction = 1;
830 tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
835 tbl_data->entry.key = table_key.v64;
836 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
839 rte_atomic32_init(&tbl_data->tbl.refcnt);
840 rte_atomic32_inc(&tbl_data->tbl.refcnt);
841 table_key.direction = 0;
842 table_key.domain = 1;
843 tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
848 tbl_data->entry.key = table_key.v64;
849 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
852 rte_atomic32_init(&tbl_data->tbl.refcnt);
853 rte_atomic32_inc(&tbl_data->tbl.refcnt);
856 mlx5_free_table_hash_list(priv);
857 #endif /* HAVE_MLX5DV_DR */
862 * Initialize DR related data within private structure.
863 * Routine checks the reference counter and does actual
864 * resources creation/initialization only if counter is zero.
867 * Pointer to the private device data structure.
870 * Zero on success, positive error code otherwise.
873 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
875 struct mlx5_ibv_shared *sh = priv->sh;
876 char s[MLX5_HLIST_NAMESIZE];
880 err = mlx5_alloc_table_hash_list(priv);
882 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
883 (void *)sh->flow_tbls);
886 /* Create tags hash list table. */
887 snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
888 sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
889 if (!sh->tag_table) {
890 DRV_LOG(ERR, "tags with hash creation failed.\n");
894 #ifdef HAVE_MLX5DV_DR
898 /* Shared DV/DR structures is already initialized. */
903 /* Reference counter is zero, we should initialize structures. */
904 domain = mlx5_glue->dr_create_domain(sh->ctx,
905 MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
907 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
911 sh->rx_domain = domain;
912 domain = mlx5_glue->dr_create_domain(sh->ctx,
913 MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
915 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
919 pthread_mutex_init(&sh->dv_mutex, NULL);
920 sh->tx_domain = domain;
921 #ifdef HAVE_MLX5DV_DR_ESWITCH
922 if (priv->config.dv_esw_en) {
923 domain = mlx5_glue->dr_create_domain
924 (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
926 DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
930 sh->fdb_domain = domain;
931 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
934 sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
935 #endif /* HAVE_MLX5DV_DR */
940 /* Rollback the created objects. */
942 mlx5_glue->dr_destroy_domain(sh->rx_domain);
943 sh->rx_domain = NULL;
946 mlx5_glue->dr_destroy_domain(sh->tx_domain);
947 sh->tx_domain = NULL;
949 if (sh->fdb_domain) {
950 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
951 sh->fdb_domain = NULL;
953 if (sh->esw_drop_action) {
954 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
955 sh->esw_drop_action = NULL;
957 if (sh->pop_vlan_action) {
958 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
959 sh->pop_vlan_action = NULL;
962 /* tags should be destroyed with flow before. */
963 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
964 sh->tag_table = NULL;
966 mlx5_free_table_hash_list(priv);
971 * Destroy DR related data within private structure.
974 * Pointer to the private device data structure.
977 mlx5_free_shared_dr(struct mlx5_priv *priv)
979 struct mlx5_ibv_shared *sh;
981 if (!priv->dr_shared)
986 #ifdef HAVE_MLX5DV_DR
987 assert(sh->dv_refcnt);
988 if (sh->dv_refcnt && --sh->dv_refcnt)
991 mlx5_glue->dr_destroy_domain(sh->rx_domain);
992 sh->rx_domain = NULL;
995 mlx5_glue->dr_destroy_domain(sh->tx_domain);
996 sh->tx_domain = NULL;
998 #ifdef HAVE_MLX5DV_DR_ESWITCH
999 if (sh->fdb_domain) {
1000 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1001 sh->fdb_domain = NULL;
1003 if (sh->esw_drop_action) {
1004 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1005 sh->esw_drop_action = NULL;
1008 if (sh->pop_vlan_action) {
1009 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1010 sh->pop_vlan_action = NULL;
1012 pthread_mutex_destroy(&sh->dv_mutex);
1013 #endif /* HAVE_MLX5DV_DR */
1014 if (sh->tag_table) {
1015 /* tags should be destroyed with flow before. */
1016 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1017 sh->tag_table = NULL;
1019 mlx5_free_table_hash_list(priv);
1023 * Initialize shared data between primary and secondary process.
1025 * A memzone is reserved by primary process and secondary processes attach to
1029 * 0 on success, a negative errno value otherwise and rte_errno is set.
1032 mlx5_init_shared_data(void)
1034 const struct rte_memzone *mz;
1037 rte_spinlock_lock(&mlx5_shared_data_lock);
1038 if (mlx5_shared_data == NULL) {
1039 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1040 /* Allocate shared memory. */
1041 mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1042 sizeof(*mlx5_shared_data),
1046 "Cannot allocate mlx5 shared data");
1050 mlx5_shared_data = mz->addr;
1051 memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1052 rte_spinlock_init(&mlx5_shared_data->lock);
1054 /* Lookup allocated shared memory. */
1055 mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1058 "Cannot attach mlx5 shared data");
1062 mlx5_shared_data = mz->addr;
1063 memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1067 rte_spinlock_unlock(&mlx5_shared_data_lock);
1072 * Retrieve integer value from environment variable.
1075 * Environment variable name.
1078 * Integer value, 0 if the variable is not set.
1081 mlx5_getenv_int(const char *name)
1083 const char *val = getenv(name);
1091 * Verbs callback to allocate a memory. This function should allocate the space
1092 * according to the size provided residing inside a huge page.
1093 * Please note that all allocation must respect the alignment from libmlx5
1094 * (i.e. currently sysconf(_SC_PAGESIZE)).
1097 * The size in bytes of the memory to allocate.
1099 * A pointer to the callback data.
1102 * Allocated buffer, NULL otherwise and rte_errno is set.
1105 mlx5_alloc_verbs_buf(size_t size, void *data)
1107 struct mlx5_priv *priv = data;
1109 size_t alignment = sysconf(_SC_PAGESIZE);
1110 unsigned int socket = SOCKET_ID_ANY;
1112 if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
1113 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1115 socket = ctrl->socket;
1116 } else if (priv->verbs_alloc_ctx.type ==
1117 MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
1118 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1120 socket = ctrl->socket;
1122 assert(data != NULL);
1123 ret = rte_malloc_socket(__func__, size, alignment, socket);
1130 * Verbs callback to free a memory.
1133 * A pointer to the memory to free.
1135 * A pointer to the callback data.
1138 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1140 assert(data != NULL);
1145 * DPDK callback to add udp tunnel port
1148 * A pointer to eth_dev
1149 * @param[in] udp_tunnel
1150 * A pointer to udp tunnel
1153 * 0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1156 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1157 struct rte_eth_udp_tunnel *udp_tunnel)
1159 assert(udp_tunnel != NULL);
1160 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1161 udp_tunnel->udp_port == 4789)
1163 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1164 udp_tunnel->udp_port == 4790)
1170 * Initialize process private data structure.
1173 * Pointer to Ethernet device structure.
1176 * 0 on success, a negative errno value otherwise and rte_errno is set.
1179 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1181 struct mlx5_priv *priv = dev->data->dev_private;
1182 struct mlx5_proc_priv *ppriv;
1186 * UAR register table follows the process private structure. BlueFlame
1187 * registers for Tx queues are stored in the table.
1190 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1191 ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1192 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1197 ppriv->uar_table_sz = ppriv_size;
1198 dev->process_private = ppriv;
1203 * Un-initialize process private data structure.
1206 * Pointer to Ethernet device structure.
1209 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1211 if (!dev->process_private)
1213 rte_free(dev->process_private);
1214 dev->process_private = NULL;
1218 * DPDK callback to close the device.
1220 * Destroy all queues and objects, free memory.
1223 * Pointer to Ethernet device structure.
1226 mlx5_dev_close(struct rte_eth_dev *dev)
1228 struct mlx5_priv *priv = dev->data->dev_private;
1232 DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1234 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1235 /* In case mlx5_dev_stop() has not been called. */
1236 mlx5_dev_interrupt_handler_uninstall(dev);
1237 mlx5_dev_interrupt_handler_devx_uninstall(dev);
1238 mlx5_traffic_disable(dev);
1239 mlx5_flow_flush(dev, NULL);
1240 mlx5_flow_meter_flush(dev, NULL);
1241 /* Prevent crashes when queues are still in use. */
1242 dev->rx_pkt_burst = removed_rx_burst;
1243 dev->tx_pkt_burst = removed_tx_burst;
1245 /* Disable datapath on secondary process. */
1246 mlx5_mp_req_stop_rxtx(dev);
1247 if (priv->rxqs != NULL) {
1248 /* XXX race condition if mlx5_rx_burst() is still running. */
1250 for (i = 0; (i != priv->rxqs_n); ++i)
1251 mlx5_rxq_release(dev, i);
1255 if (priv->txqs != NULL) {
1256 /* XXX race condition if mlx5_tx_burst() is still running. */
1258 for (i = 0; (i != priv->txqs_n); ++i)
1259 mlx5_txq_release(dev, i);
1263 mlx5_proc_priv_uninit(dev);
1264 if (priv->mreg_cp_tbl)
1265 mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1266 mlx5_mprq_free_mp(dev);
1267 mlx5_free_shared_dr(priv);
1268 if (priv->rss_conf.rss_key != NULL)
1269 rte_free(priv->rss_conf.rss_key);
1270 if (priv->reta_idx != NULL)
1271 rte_free(priv->reta_idx);
1272 if (priv->config.vf)
1273 mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
1274 dev->data->mac_addrs,
1275 MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
1276 if (priv->nl_socket_route >= 0)
1277 close(priv->nl_socket_route);
1278 if (priv->nl_socket_rdma >= 0)
1279 close(priv->nl_socket_rdma);
1280 if (priv->vmwa_context)
1281 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1284 * Free the shared context in last turn, because the cleanup
1285 * routines above may use some shared fields, like
1286 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1287 * ifindex if Netlink fails.
1289 mlx5_free_shared_ibctx(priv->sh);
1292 ret = mlx5_hrxq_verify(dev);
1294 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1295 dev->data->port_id);
1296 ret = mlx5_ind_table_obj_verify(dev);
1298 DRV_LOG(WARNING, "port %u some indirection table still remain",
1299 dev->data->port_id);
1300 ret = mlx5_rxq_obj_verify(dev);
1302 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1303 dev->data->port_id);
1304 ret = mlx5_rxq_verify(dev);
1306 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1307 dev->data->port_id);
1308 ret = mlx5_txq_obj_verify(dev);
1310 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1311 dev->data->port_id);
1312 ret = mlx5_txq_verify(dev);
1314 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1315 dev->data->port_id);
1316 ret = mlx5_flow_verify(dev);
1318 DRV_LOG(WARNING, "port %u some flows still remain",
1319 dev->data->port_id);
1320 if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1324 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1325 struct mlx5_priv *opriv =
1326 rte_eth_devices[port_id].data->dev_private;
1329 opriv->domain_id != priv->domain_id ||
1330 &rte_eth_devices[port_id] == dev)
1336 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1338 memset(priv, 0, sizeof(*priv));
1339 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1341 * Reset mac_addrs to NULL such that it is not freed as part of
1342 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1343 * it is freed when dev_private is freed.
1345 dev->data->mac_addrs = NULL;
1348 const struct eth_dev_ops mlx5_dev_ops = {
1349 .dev_configure = mlx5_dev_configure,
1350 .dev_start = mlx5_dev_start,
1351 .dev_stop = mlx5_dev_stop,
1352 .dev_set_link_down = mlx5_set_link_down,
1353 .dev_set_link_up = mlx5_set_link_up,
1354 .dev_close = mlx5_dev_close,
1355 .promiscuous_enable = mlx5_promiscuous_enable,
1356 .promiscuous_disable = mlx5_promiscuous_disable,
1357 .allmulticast_enable = mlx5_allmulticast_enable,
1358 .allmulticast_disable = mlx5_allmulticast_disable,
1359 .link_update = mlx5_link_update,
1360 .stats_get = mlx5_stats_get,
1361 .stats_reset = mlx5_stats_reset,
1362 .xstats_get = mlx5_xstats_get,
1363 .xstats_reset = mlx5_xstats_reset,
1364 .xstats_get_names = mlx5_xstats_get_names,
1365 .fw_version_get = mlx5_fw_version_get,
1366 .dev_infos_get = mlx5_dev_infos_get,
1367 .read_clock = mlx5_read_clock,
1368 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1369 .vlan_filter_set = mlx5_vlan_filter_set,
1370 .rx_queue_setup = mlx5_rx_queue_setup,
1371 .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1372 .tx_queue_setup = mlx5_tx_queue_setup,
1373 .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1374 .rx_queue_release = mlx5_rx_queue_release,
1375 .tx_queue_release = mlx5_tx_queue_release,
1376 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1377 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1378 .mac_addr_remove = mlx5_mac_addr_remove,
1379 .mac_addr_add = mlx5_mac_addr_add,
1380 .mac_addr_set = mlx5_mac_addr_set,
1381 .set_mc_addr_list = mlx5_set_mc_addr_list,
1382 .mtu_set = mlx5_dev_set_mtu,
1383 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1384 .vlan_offload_set = mlx5_vlan_offload_set,
1385 .reta_update = mlx5_dev_rss_reta_update,
1386 .reta_query = mlx5_dev_rss_reta_query,
1387 .rss_hash_update = mlx5_rss_hash_update,
1388 .rss_hash_conf_get = mlx5_rss_hash_conf_get,
1389 .filter_ctrl = mlx5_dev_filter_ctrl,
1390 .rx_descriptor_status = mlx5_rx_descriptor_status,
1391 .tx_descriptor_status = mlx5_tx_descriptor_status,
1392 .rx_queue_count = mlx5_rx_queue_count,
1393 .rx_queue_intr_enable = mlx5_rx_intr_enable,
1394 .rx_queue_intr_disable = mlx5_rx_intr_disable,
1395 .is_removed = mlx5_is_removed,
1396 .udp_tunnel_port_add = mlx5_udp_tunnel_port_add,
1397 .get_module_info = mlx5_get_module_info,
1398 .get_module_eeprom = mlx5_get_module_eeprom,
1399 .hairpin_cap_get = mlx5_hairpin_cap_get,
1400 .mtr_ops_get = mlx5_flow_meter_ops_get,
1403 /* Available operations from secondary process. */
1404 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1405 .stats_get = mlx5_stats_get,
1406 .stats_reset = mlx5_stats_reset,
1407 .xstats_get = mlx5_xstats_get,
1408 .xstats_reset = mlx5_xstats_reset,
1409 .xstats_get_names = mlx5_xstats_get_names,
1410 .fw_version_get = mlx5_fw_version_get,
1411 .dev_infos_get = mlx5_dev_infos_get,
1412 .rx_descriptor_status = mlx5_rx_descriptor_status,
1413 .tx_descriptor_status = mlx5_tx_descriptor_status,
1414 .get_module_info = mlx5_get_module_info,
1415 .get_module_eeprom = mlx5_get_module_eeprom,
1418 /* Available operations in flow isolated mode. */
1419 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1420 .dev_configure = mlx5_dev_configure,
1421 .dev_start = mlx5_dev_start,
1422 .dev_stop = mlx5_dev_stop,
1423 .dev_set_link_down = mlx5_set_link_down,
1424 .dev_set_link_up = mlx5_set_link_up,
1425 .dev_close = mlx5_dev_close,
1426 .promiscuous_enable = mlx5_promiscuous_enable,
1427 .promiscuous_disable = mlx5_promiscuous_disable,
1428 .allmulticast_enable = mlx5_allmulticast_enable,
1429 .allmulticast_disable = mlx5_allmulticast_disable,
1430 .link_update = mlx5_link_update,
1431 .stats_get = mlx5_stats_get,
1432 .stats_reset = mlx5_stats_reset,
1433 .xstats_get = mlx5_xstats_get,
1434 .xstats_reset = mlx5_xstats_reset,
1435 .xstats_get_names = mlx5_xstats_get_names,
1436 .fw_version_get = mlx5_fw_version_get,
1437 .dev_infos_get = mlx5_dev_infos_get,
1438 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1439 .vlan_filter_set = mlx5_vlan_filter_set,
1440 .rx_queue_setup = mlx5_rx_queue_setup,
1441 .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1442 .tx_queue_setup = mlx5_tx_queue_setup,
1443 .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1444 .rx_queue_release = mlx5_rx_queue_release,
1445 .tx_queue_release = mlx5_tx_queue_release,
1446 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1447 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1448 .mac_addr_remove = mlx5_mac_addr_remove,
1449 .mac_addr_add = mlx5_mac_addr_add,
1450 .mac_addr_set = mlx5_mac_addr_set,
1451 .set_mc_addr_list = mlx5_set_mc_addr_list,
1452 .mtu_set = mlx5_dev_set_mtu,
1453 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1454 .vlan_offload_set = mlx5_vlan_offload_set,
1455 .filter_ctrl = mlx5_dev_filter_ctrl,
1456 .rx_descriptor_status = mlx5_rx_descriptor_status,
1457 .tx_descriptor_status = mlx5_tx_descriptor_status,
1458 .rx_queue_intr_enable = mlx5_rx_intr_enable,
1459 .rx_queue_intr_disable = mlx5_rx_intr_disable,
1460 .is_removed = mlx5_is_removed,
1461 .get_module_info = mlx5_get_module_info,
1462 .get_module_eeprom = mlx5_get_module_eeprom,
1463 .hairpin_cap_get = mlx5_hairpin_cap_get,
1464 .mtr_ops_get = mlx5_flow_meter_ops_get,
1468 * Verify and store value for device argument.
1471 * Key argument to verify.
1473 * Value associated with key.
1478 * 0 on success, a negative errno value otherwise and rte_errno is set.
1481 mlx5_args_check(const char *key, const char *val, void *opaque)
1483 struct mlx5_dev_config *config = opaque;
1486 /* No-op, port representors are processed in mlx5_dev_spawn(). */
1487 if (!strcmp(MLX5_REPRESENTOR, key))
1490 tmp = strtoul(val, NULL, 0);
1493 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1496 if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1497 config->cqe_comp = !!tmp;
1498 } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1499 config->cqe_pad = !!tmp;
1500 } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1501 config->hw_padding = !!tmp;
1502 } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1503 config->mprq.enabled = !!tmp;
1504 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1505 config->mprq.stride_num_n = tmp;
1506 } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1507 config->mprq.max_memcpy_len = tmp;
1508 } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1509 config->mprq.min_rxqs_num = tmp;
1510 } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1511 DRV_LOG(WARNING, "%s: deprecated parameter,"
1512 " converted to txq_inline_max", key);
1513 config->txq_inline_max = tmp;
1514 } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1515 config->txq_inline_max = tmp;
1516 } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1517 config->txq_inline_min = tmp;
1518 } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1519 config->txq_inline_mpw = tmp;
1520 } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1521 config->txqs_inline = tmp;
1522 } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1523 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1524 } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1525 config->mps = !!tmp;
1526 } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1527 if (tmp != MLX5_TXDB_CACHED &&
1528 tmp != MLX5_TXDB_NCACHED &&
1529 tmp != MLX5_TXDB_HEURISTIC) {
1530 DRV_LOG(ERR, "invalid Tx doorbell "
1531 "mapping parameter");
1536 } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1537 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1538 } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1539 DRV_LOG(WARNING, "%s: deprecated parameter,"
1540 " converted to txq_inline_mpw", key);
1541 config->txq_inline_mpw = tmp;
1542 } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1543 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1544 } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1545 config->rx_vec_en = !!tmp;
1546 } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1547 config->l3_vxlan_en = !!tmp;
1548 } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1549 config->vf_nl_en = !!tmp;
1550 } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1551 config->dv_esw_en = !!tmp;
1552 } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1553 config->dv_flow_en = !!tmp;
1554 } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1555 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1556 tmp != MLX5_XMETA_MODE_META16 &&
1557 tmp != MLX5_XMETA_MODE_META32) {
1558 DRV_LOG(ERR, "invalid extensive "
1559 "metadata parameter");
1563 config->dv_xmeta_en = tmp;
1564 } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1565 config->mr_ext_memseg_en = !!tmp;
1566 } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1567 config->max_dump_files_num = tmp;
1568 } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1569 config->lro.timeout = tmp;
1570 } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1571 DRV_LOG(DEBUG, "class argument is %s.", val);
1573 DRV_LOG(WARNING, "%s: unknown parameter", key);
1581 * Parse device parameters.
1584 * Pointer to device configuration structure.
1586 * Device arguments structure.
1589 * 0 on success, a negative errno value otherwise and rte_errno is set.
1592 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1594 const char **params = (const char *[]){
1595 MLX5_RXQ_CQE_COMP_EN,
1596 MLX5_RXQ_CQE_PAD_EN,
1597 MLX5_RXQ_PKT_PAD_EN,
1599 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1600 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1603 MLX5_TXQ_INLINE_MIN,
1604 MLX5_TXQ_INLINE_MAX,
1605 MLX5_TXQ_INLINE_MPW,
1606 MLX5_TXQS_MIN_INLINE,
1609 MLX5_TXQ_MPW_HDR_DSEG_EN,
1610 MLX5_TXQ_MAX_INLINE_LEN,
1619 MLX5_MR_EXT_MEMSEG_EN,
1621 MLX5_MAX_DUMP_FILES_NUM,
1622 MLX5_LRO_TIMEOUT_USEC,
1623 MLX5_CLASS_ARG_NAME,
1626 struct rte_kvargs *kvlist;
1630 if (devargs == NULL)
1632 /* Following UGLY cast is done to pass checkpatch. */
1633 kvlist = rte_kvargs_parse(devargs->args, params);
1634 if (kvlist == NULL) {
1638 /* Process parameters. */
1639 for (i = 0; (params[i] != NULL); ++i) {
1640 if (rte_kvargs_count(kvlist, params[i])) {
1641 ret = rte_kvargs_process(kvlist, params[i],
1642 mlx5_args_check, config);
1645 rte_kvargs_free(kvlist);
1650 rte_kvargs_free(kvlist);
1654 static struct rte_pci_driver mlx5_driver;
1657 * PMD global initialization.
1659 * Independent from individual device, this function initializes global
1660 * per-PMD data structures distinguishing primary and secondary processes.
1661 * Hence, each initialization is called once per a process.
1664 * 0 on success, a negative errno value otherwise and rte_errno is set.
1667 mlx5_init_once(void)
1669 struct mlx5_shared_data *sd;
1670 struct mlx5_local_data *ld = &mlx5_local_data;
1673 if (mlx5_init_shared_data())
1675 sd = mlx5_shared_data;
1677 rte_spinlock_lock(&sd->lock);
1678 switch (rte_eal_process_type()) {
1679 case RTE_PROC_PRIMARY:
1682 LIST_INIT(&sd->mem_event_cb_list);
1683 rte_rwlock_init(&sd->mem_event_rwlock);
1684 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1685 mlx5_mr_mem_event_cb, NULL);
1686 ret = mlx5_mp_init_primary();
1689 sd->init_done = true;
1691 case RTE_PROC_SECONDARY:
1694 ret = mlx5_mp_init_secondary();
1697 ++sd->secondary_cnt;
1698 ld->init_done = true;
1704 rte_spinlock_unlock(&sd->lock);
1709 * Configures the minimal amount of data to inline into WQE
1710 * while sending packets.
1712 * - the txq_inline_min has the maximal priority, if this
1713 * key is specified in devargs
1714 * - if DevX is enabled the inline mode is queried from the
1715 * device (HCA attributes and NIC vport context if needed).
1716 * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1717 * and none (0 bytes) for other NICs
1720 * Verbs device parameters (name, port, switch_info) to spawn.
1722 * Device configuration parameters.
1725 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1726 struct mlx5_dev_config *config)
1728 if (config->txq_inline_min != MLX5_ARG_UNSET) {
1729 /* Application defines size of inlined data explicitly. */
1730 switch (spawn->pci_dev->id.device_id) {
1731 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1732 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1733 if (config->txq_inline_min <
1734 (int)MLX5_INLINE_HSIZE_L2) {
1736 "txq_inline_mix aligned to minimal"
1737 " ConnectX-4 required value %d",
1738 (int)MLX5_INLINE_HSIZE_L2);
1739 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1745 if (config->hca_attr.eth_net_offloads) {
1746 /* We have DevX enabled, inline mode queried successfully. */
1747 switch (config->hca_attr.wqe_inline_mode) {
1748 case MLX5_CAP_INLINE_MODE_L2:
1749 /* outer L2 header must be inlined. */
1750 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1752 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1753 /* No inline data are required by NIC. */
1754 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1755 config->hw_vlan_insert =
1756 config->hca_attr.wqe_vlan_insert;
1757 DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1759 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1760 /* inline mode is defined by NIC vport context. */
1761 if (!config->hca_attr.eth_virt)
1763 switch (config->hca_attr.vport_inline_mode) {
1764 case MLX5_INLINE_MODE_NONE:
1765 config->txq_inline_min =
1766 MLX5_INLINE_HSIZE_NONE;
1768 case MLX5_INLINE_MODE_L2:
1769 config->txq_inline_min =
1770 MLX5_INLINE_HSIZE_L2;
1772 case MLX5_INLINE_MODE_IP:
1773 config->txq_inline_min =
1774 MLX5_INLINE_HSIZE_L3;
1776 case MLX5_INLINE_MODE_TCP_UDP:
1777 config->txq_inline_min =
1778 MLX5_INLINE_HSIZE_L4;
1780 case MLX5_INLINE_MODE_INNER_L2:
1781 config->txq_inline_min =
1782 MLX5_INLINE_HSIZE_INNER_L2;
1784 case MLX5_INLINE_MODE_INNER_IP:
1785 config->txq_inline_min =
1786 MLX5_INLINE_HSIZE_INNER_L3;
1788 case MLX5_INLINE_MODE_INNER_TCP_UDP:
1789 config->txq_inline_min =
1790 MLX5_INLINE_HSIZE_INNER_L4;
1796 * We get here if we are unable to deduce
1797 * inline data size with DevX. Try PCI ID
1798 * to determine old NICs.
1800 switch (spawn->pci_dev->id.device_id) {
1801 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1802 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1803 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1804 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1805 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1806 config->hw_vlan_insert = 0;
1808 case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1809 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1810 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1811 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1813 * These NICs support VLAN insertion from WQE and
1814 * report the wqe_vlan_insert flag. But there is the bug
1815 * and PFC control may be broken, so disable feature.
1817 config->hw_vlan_insert = 0;
1818 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1821 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1825 DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1829 * Configures the metadata mask fields in the shared context.
1832 * Pointer to Ethernet device.
1835 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1837 struct mlx5_priv *priv = dev->data->dev_private;
1838 struct mlx5_ibv_shared *sh = priv->sh;
1839 uint32_t meta, mark, reg_c0;
1841 reg_c0 = ~priv->vport_meta_mask;
1842 switch (priv->config.dv_xmeta_en) {
1843 case MLX5_XMETA_MODE_LEGACY:
1845 mark = MLX5_FLOW_MARK_MASK;
1847 case MLX5_XMETA_MODE_META16:
1848 meta = reg_c0 >> rte_bsf32(reg_c0);
1849 mark = MLX5_FLOW_MARK_MASK;
1851 case MLX5_XMETA_MODE_META32:
1853 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1861 if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1862 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1863 sh->dv_mark_mask, mark);
1865 sh->dv_mark_mask = mark;
1866 if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1867 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1868 sh->dv_meta_mask, meta);
1870 sh->dv_meta_mask = meta;
1871 if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1872 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1873 sh->dv_meta_mask, reg_c0);
1875 sh->dv_regc0_mask = reg_c0;
1876 DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1877 DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1878 DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1879 DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1883 * Allocate page of door-bells and register it using DevX API.
1886 * Pointer to Ethernet device.
1889 * Pointer to new page on success, NULL otherwise.
1891 static struct mlx5_devx_dbr_page *
1892 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1894 struct mlx5_priv *priv = dev->data->dev_private;
1895 struct mlx5_devx_dbr_page *page;
1897 /* Allocate space for door-bell page and management data. */
1898 page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1899 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1901 DRV_LOG(ERR, "port %u cannot allocate dbr page",
1902 dev->data->port_id);
1905 /* Register allocated memory. */
1906 page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1907 MLX5_DBR_PAGE_SIZE, 0);
1909 DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1910 dev->data->port_id);
1918 * Find the next available door-bell, allocate new page if needed.
1921 * Pointer to Ethernet device.
1922 * @param [out] dbr_page
1923 * Door-bell page containing the page data.
1926 * Door-bell address offset on success, a negative error value otherwise.
1929 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1931 struct mlx5_priv *priv = dev->data->dev_private;
1932 struct mlx5_devx_dbr_page *page = NULL;
1935 LIST_FOREACH(page, &priv->dbrpgs, next)
1936 if (page->dbr_count < MLX5_DBR_PER_PAGE)
1938 if (!page) { /* No page with free door-bell exists. */
1939 page = mlx5_alloc_dbr_page(dev);
1940 if (!page) /* Failed to allocate new page. */
1942 LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1944 /* Loop to find bitmap part with clear bit. */
1946 i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1949 /* Find the first clear bit. */
1950 j = rte_bsf64(~page->dbr_bitmap[i]);
1951 assert(i < (MLX5_DBR_PER_PAGE / 64));
1952 page->dbr_bitmap[i] |= (1 << j);
1955 return (((i * 64) + j) * sizeof(uint64_t));
1959 * Release a door-bell record.
1962 * Pointer to Ethernet device.
1963 * @param [in] umem_id
1964 * UMEM ID of page containing the door-bell record to release.
1965 * @param [in] offset
1966 * Offset of door-bell record in page.
1969 * 0 on success, a negative error value otherwise.
1972 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1974 struct mlx5_priv *priv = dev->data->dev_private;
1975 struct mlx5_devx_dbr_page *page = NULL;
1978 LIST_FOREACH(page, &priv->dbrpgs, next)
1979 /* Find the page this address belongs to. */
1980 if (page->umem->umem_id == umem_id)
1985 if (!page->dbr_count) {
1986 /* Page not used, free it and remove from list. */
1987 LIST_REMOVE(page, next);
1989 ret = -mlx5_glue->devx_umem_dereg(page->umem);
1992 /* Mark in bitmap that this door-bell is not in use. */
1993 offset /= MLX5_DBR_SIZE;
1994 int i = offset / 64;
1995 int j = offset % 64;
1997 page->dbr_bitmap[i] &= ~(1 << j);
2003 * Check sibling device configurations.
2005 * Sibling devices sharing the Infiniband device context
2006 * should have compatible configurations. This regards
2007 * representors and bonding slaves.
2010 * Private device descriptor.
2012 * Configuration of the device is going to be created.
2015 * 0 on success, EINVAL otherwise
2018 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2019 struct mlx5_dev_config *config)
2021 struct mlx5_ibv_shared *sh = priv->sh;
2022 struct mlx5_dev_config *sh_conf = NULL;
2026 /* Nothing to compare for the single/first device. */
2027 if (sh->refcnt == 1)
2029 /* Find the device with shared context. */
2030 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2031 struct mlx5_priv *opriv =
2032 rte_eth_devices[port_id].data->dev_private;
2034 if (opriv && opriv != priv && opriv->sh == sh) {
2035 sh_conf = &opriv->config;
2041 if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2042 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2043 " for shared %s context", sh->ibdev_name);
2047 if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2048 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2049 " for shared %s context", sh->ibdev_name);
2056 * Spawn an Ethernet device from Verbs information.
2059 * Backing DPDK device.
2061 * Verbs device parameters (name, port, switch_info) to spawn.
2063 * Device configuration parameters.
2066 * A valid Ethernet device object on success, NULL otherwise and rte_errno
2067 * is set. The following errors are defined:
2069 * EBUSY: device is not supposed to be spawned.
2070 * EEXIST: device is already spawned
2072 static struct rte_eth_dev *
2073 mlx5_dev_spawn(struct rte_device *dpdk_dev,
2074 struct mlx5_dev_spawn_data *spawn,
2075 struct mlx5_dev_config config)
2077 const struct mlx5_switch_info *switch_info = &spawn->info;
2078 struct mlx5_ibv_shared *sh = NULL;
2079 struct ibv_port_attr port_attr;
2080 struct mlx5dv_context dv_attr = { .comp_mask = 0 };
2081 struct rte_eth_dev *eth_dev = NULL;
2082 struct mlx5_priv *priv = NULL;
2084 unsigned int hw_padding = 0;
2086 unsigned int cqe_comp;
2087 unsigned int cqe_pad = 0;
2088 unsigned int tunnel_en = 0;
2089 unsigned int mpls_en = 0;
2090 unsigned int swp = 0;
2091 unsigned int mprq = 0;
2092 unsigned int mprq_min_stride_size_n = 0;
2093 unsigned int mprq_max_stride_size_n = 0;
2094 unsigned int mprq_min_stride_num_n = 0;
2095 unsigned int mprq_max_stride_num_n = 0;
2096 struct rte_ether_addr mac;
2097 char name[RTE_ETH_NAME_MAX_LEN];
2098 int own_domain_id = 0;
2101 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2102 struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
2105 /* Determine if this port representor is supposed to be spawned. */
2106 if (switch_info->representor && dpdk_dev->devargs) {
2107 struct rte_eth_devargs eth_da;
2109 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, ð_da);
2112 DRV_LOG(ERR, "failed to process device arguments: %s",
2113 strerror(rte_errno));
2116 for (i = 0; i < eth_da.nb_representor_ports; ++i)
2117 if (eth_da.representor_ports[i] ==
2118 (uint16_t)switch_info->port_name)
2120 if (i == eth_da.nb_representor_ports) {
2125 /* Build device name. */
2126 if (spawn->pf_bond < 0) {
2127 /* Single device. */
2128 if (!switch_info->representor)
2129 strlcpy(name, dpdk_dev->name, sizeof(name));
2131 snprintf(name, sizeof(name), "%s_representor_%u",
2132 dpdk_dev->name, switch_info->port_name);
2134 /* Bonding device. */
2135 if (!switch_info->representor)
2136 snprintf(name, sizeof(name), "%s_%s",
2137 dpdk_dev->name, spawn->ibv_dev->name);
2139 snprintf(name, sizeof(name), "%s_%s_representor_%u",
2140 dpdk_dev->name, spawn->ibv_dev->name,
2141 switch_info->port_name);
2143 /* check if the device is already spawned */
2144 if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2148 DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2149 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2150 eth_dev = rte_eth_dev_attach_secondary(name);
2151 if (eth_dev == NULL) {
2152 DRV_LOG(ERR, "can not attach rte ethdev");
2156 eth_dev->device = dpdk_dev;
2157 eth_dev->dev_ops = &mlx5_dev_sec_ops;
2158 err = mlx5_proc_priv_init(eth_dev);
2161 /* Receive command fd from primary process */
2162 err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
2165 /* Remap UAR for Tx queues. */
2166 err = mlx5_tx_uar_init_secondary(eth_dev, err);
2170 * Ethdev pointer is still required as input since
2171 * the primary device is not accessible from the
2172 * secondary process.
2174 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2175 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2179 * Some parameters ("tx_db_nc" in particularly) are needed in
2180 * advance to create dv/verbs device context. We proceed the
2181 * devargs here to get ones, and later proceed devargs again
2182 * to override some hardware settings.
2184 err = mlx5_args(&config, dpdk_dev->devargs);
2187 DRV_LOG(ERR, "failed to process device arguments: %s",
2188 strerror(rte_errno));
2191 sh = mlx5_alloc_shared_ibctx(spawn, &config);
2194 config.devx = sh->devx;
2195 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2196 config.dest_tir = 1;
2198 #ifdef HAVE_IBV_MLX5_MOD_SWP
2199 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2202 * Multi-packet send is supported by ConnectX-4 Lx PF as well
2203 * as all ConnectX-5 devices.
2205 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2206 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2208 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2209 dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2211 mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2212 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2213 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2214 DRV_LOG(DEBUG, "enhanced MPW is supported");
2215 mps = MLX5_MPW_ENHANCED;
2217 DRV_LOG(DEBUG, "MPW is supported");
2221 DRV_LOG(DEBUG, "MPW isn't supported");
2222 mps = MLX5_MPW_DISABLED;
2224 #ifdef HAVE_IBV_MLX5_MOD_SWP
2225 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2226 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2227 DRV_LOG(DEBUG, "SWP support: %u", swp);
2230 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2231 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2232 struct mlx5dv_striding_rq_caps mprq_caps =
2233 dv_attr.striding_rq_caps;
2235 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2236 mprq_caps.min_single_stride_log_num_of_bytes);
2237 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2238 mprq_caps.max_single_stride_log_num_of_bytes);
2239 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2240 mprq_caps.min_single_wqe_log_num_of_strides);
2241 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2242 mprq_caps.max_single_wqe_log_num_of_strides);
2243 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2244 mprq_caps.supported_qpts);
2245 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2247 mprq_min_stride_size_n =
2248 mprq_caps.min_single_stride_log_num_of_bytes;
2249 mprq_max_stride_size_n =
2250 mprq_caps.max_single_stride_log_num_of_bytes;
2251 mprq_min_stride_num_n =
2252 mprq_caps.min_single_wqe_log_num_of_strides;
2253 mprq_max_stride_num_n =
2254 mprq_caps.max_single_wqe_log_num_of_strides;
2255 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2256 mprq_min_stride_num_n);
2259 if (RTE_CACHE_LINE_SIZE == 128 &&
2260 !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2264 config.cqe_comp = cqe_comp;
2265 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2266 /* Whether device supports 128B Rx CQE padding. */
2267 cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2268 (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2270 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2271 if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2272 tunnel_en = ((dv_attr.tunnel_offloads_caps &
2273 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2274 (dv_attr.tunnel_offloads_caps &
2275 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
2276 (dv_attr.tunnel_offloads_caps &
2277 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
2279 DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2280 tunnel_en ? "" : "not ");
2283 "tunnel offloading disabled due to old OFED/rdma-core version");
2285 config.tunnel_en = tunnel_en;
2286 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2287 mpls_en = ((dv_attr.tunnel_offloads_caps &
2288 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2289 (dv_attr.tunnel_offloads_caps &
2290 MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2291 DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2292 mpls_en ? "" : "not ");
2294 DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2295 " old OFED/rdma-core version or firmware configuration");
2297 config.mpls_en = mpls_en;
2298 /* Check port status. */
2299 err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2301 DRV_LOG(ERR, "port query failed: %s", strerror(err));
2304 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2305 DRV_LOG(ERR, "port is not configured in Ethernet mode");
2309 if (port_attr.state != IBV_PORT_ACTIVE)
2310 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2311 mlx5_glue->port_state_str(port_attr.state),
2313 /* Allocate private eth device data. */
2314 priv = rte_zmalloc("ethdev private structure",
2316 RTE_CACHE_LINE_SIZE);
2318 DRV_LOG(ERR, "priv allocation failure");
2323 priv->ibv_port = spawn->ibv_port;
2324 priv->pci_dev = spawn->pci_dev;
2325 priv->mtu = RTE_ETHER_MTU;
2327 /* Initialize UAR access locks for 32bit implementations. */
2328 rte_spinlock_init(&priv->uar_lock_cq);
2329 for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2330 rte_spinlock_init(&priv->uar_lock[i]);
2332 /* Some internal functions rely on Netlink sockets, open them now. */
2333 priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2334 priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
2335 priv->representor = !!switch_info->representor;
2336 priv->master = !!switch_info->master;
2337 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2338 priv->vport_meta_tag = 0;
2339 priv->vport_meta_mask = 0;
2340 priv->pf_bond = spawn->pf_bond;
2341 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2343 * The DevX port query API is implemented. E-Switch may use
2344 * either vport or reg_c[0] metadata register to match on
2345 * vport index. The engaged part of metadata register is
2348 if (switch_info->representor || switch_info->master) {
2349 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2350 MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2351 err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2355 "can't query devx port %d on device %s",
2356 spawn->ibv_port, spawn->ibv_dev->name);
2357 devx_port.comp_mask = 0;
2360 if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2361 priv->vport_meta_tag = devx_port.reg_c_0.value;
2362 priv->vport_meta_mask = devx_port.reg_c_0.mask;
2363 if (!priv->vport_meta_mask) {
2364 DRV_LOG(ERR, "vport zero mask for port %d"
2365 " on bonding device %s",
2366 spawn->ibv_port, spawn->ibv_dev->name);
2370 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2371 DRV_LOG(ERR, "invalid vport tag for port %d"
2372 " on bonding device %s",
2373 spawn->ibv_port, spawn->ibv_dev->name);
2378 if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2379 priv->vport_id = devx_port.vport_num;
2380 } else if (spawn->pf_bond >= 0) {
2381 DRV_LOG(ERR, "can't deduce vport index for port %d"
2382 " on bonding device %s",
2383 spawn->ibv_port, spawn->ibv_dev->name);
2387 /* Suppose vport index in compatible way. */
2388 priv->vport_id = switch_info->representor ?
2389 switch_info->port_name + 1 : -1;
2393 * Kernel/rdma_core support single E-Switch per PF configurations
2394 * only and vport_id field contains the vport index for
2395 * associated VF, which is deduced from representor port name.
2396 * For example, let's have the IB device port 10, it has
2397 * attached network device eth0, which has port name attribute
2398 * pf0vf2, we can deduce the VF number as 2, and set vport index
2399 * as 3 (2+1). This assigning schema should be changed if the
2400 * multiple E-Switch instances per PF configurations or/and PCI
2401 * subfunctions are added.
2403 priv->vport_id = switch_info->representor ?
2404 switch_info->port_name + 1 : -1;
2406 /* representor_id field keeps the unmodified VF index. */
2407 priv->representor_id = switch_info->representor ?
2408 switch_info->port_name : -1;
2410 * Look for sibling devices in order to reuse their switch domain
2411 * if any, otherwise allocate one.
2413 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2414 const struct mlx5_priv *opriv =
2415 rte_eth_devices[port_id].data->dev_private;
2418 opriv->sh != priv->sh ||
2420 RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2422 priv->domain_id = opriv->domain_id;
2425 if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2426 err = rte_eth_switch_domain_alloc(&priv->domain_id);
2429 DRV_LOG(ERR, "unable to allocate switch domain: %s",
2430 strerror(rte_errno));
2435 /* Override some values set by hardware configuration. */
2436 mlx5_args(&config, dpdk_dev->devargs);
2437 err = mlx5_dev_check_sibling_config(priv, &config);
2440 config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2441 IBV_DEVICE_RAW_IP_CSUM);
2442 DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2443 (config.hw_csum ? "" : "not "));
2444 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2445 !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2446 DRV_LOG(DEBUG, "counters are not supported");
2448 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
2449 if (config.dv_flow_en) {
2450 DRV_LOG(WARNING, "DV flow is not supported");
2451 config.dv_flow_en = 0;
2454 config.ind_table_max_size =
2455 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2457 * Remove this check once DPDK supports larger/variable
2458 * indirection tables.
2460 if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2461 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2462 DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2463 config.ind_table_max_size);
2464 config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2465 IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2466 DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2467 (config.hw_vlan_strip ? "" : "not "));
2468 config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2469 IBV_RAW_PACKET_CAP_SCATTER_FCS);
2470 DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2471 (config.hw_fcs_strip ? "" : "not "));
2472 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2473 hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2474 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2475 hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2476 IBV_DEVICE_PCI_WRITE_END_PADDING);
2478 if (config.hw_padding && !hw_padding) {
2479 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2480 config.hw_padding = 0;
2481 } else if (config.hw_padding) {
2482 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2484 config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2485 (sh->device_attr.tso_caps.supported_qpts &
2486 (1 << IBV_QPT_RAW_PACKET)));
2488 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2490 * MPW is disabled by default, while the Enhanced MPW is enabled
2493 if (config.mps == MLX5_ARG_UNSET)
2494 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2497 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2498 DRV_LOG(INFO, "%sMPS is %s",
2499 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
2500 config.mps == MLX5_MPW ? "legacy " : "",
2501 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2502 if (config.cqe_comp && !cqe_comp) {
2503 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2504 config.cqe_comp = 0;
2506 if (config.cqe_pad && !cqe_pad) {
2507 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2509 } else if (config.cqe_pad) {
2510 DRV_LOG(INFO, "Rx CQE padding is enabled");
2513 priv->counter_fallback = 0;
2514 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2519 if (!config.hca_attr.flow_counters_dump)
2520 priv->counter_fallback = 1;
2521 #ifndef HAVE_IBV_DEVX_ASYNC
2522 priv->counter_fallback = 1;
2524 if (priv->counter_fallback)
2525 DRV_LOG(INFO, "Use fall-back DV counter management");
2526 /* Check for LRO support. */
2527 if (config.dest_tir && config.hca_attr.lro_cap &&
2528 config.dv_flow_en) {
2529 /* TBD check tunnel lro caps. */
2530 config.lro.supported = config.hca_attr.lro_cap;
2531 DRV_LOG(DEBUG, "Device supports LRO");
2533 * If LRO timeout is not configured by application,
2534 * use the minimal supported value.
2536 if (!config.lro.timeout)
2537 config.lro.timeout =
2538 config.hca_attr.lro_timer_supported_periods[0];
2539 DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2540 config.lro.timeout);
2542 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2543 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2544 config.dv_flow_en) {
2545 uint8_t reg_c_mask =
2546 config.hca_attr.qos.flow_meter_reg_c_ids;
2548 * Meter needs two REG_C's for color match and pre-sfx
2549 * flow match. Here get the REG_C for color match.
2550 * REG_C_0 and REG_C_1 is reserved for metadata feature.
2553 if (__builtin_popcount(reg_c_mask) < 1) {
2555 DRV_LOG(WARNING, "No available register for"
2558 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2561 priv->mtr_reg_share =
2562 config.hca_attr.qos.flow_meter_reg_share;
2563 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2564 priv->mtr_color_reg);
2569 if (config.mprq.enabled && mprq) {
2570 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2571 config.mprq.stride_num_n < mprq_min_stride_num_n) {
2572 config.mprq.stride_num_n =
2573 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2574 mprq_min_stride_num_n);
2576 "the number of strides"
2577 " for Multi-Packet RQ is out of range,"
2578 " setting default value (%u)",
2579 1 << config.mprq.stride_num_n);
2581 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2582 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2583 } else if (config.mprq.enabled && !mprq) {
2584 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2585 config.mprq.enabled = 0;
2587 if (config.max_dump_files_num == 0)
2588 config.max_dump_files_num = 128;
2589 eth_dev = rte_eth_dev_allocate(name);
2590 if (eth_dev == NULL) {
2591 DRV_LOG(ERR, "can not allocate rte ethdev");
2595 /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2596 eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2597 if (priv->representor) {
2598 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2599 eth_dev->data->representor_id = priv->representor_id;
2602 * Store associated network device interface index. This index
2603 * is permanent throughout the lifetime of device. So, we may store
2604 * the ifindex here and use the cached value further.
2606 assert(spawn->ifindex);
2607 priv->if_index = spawn->ifindex;
2608 eth_dev->data->dev_private = priv;
2609 priv->dev_data = eth_dev->data;
2610 eth_dev->data->mac_addrs = priv->mac;
2611 eth_dev->device = dpdk_dev;
2612 /* Configure the first MAC address by default. */
2613 if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2615 "port %u cannot get MAC address, is mlx5_en"
2616 " loaded? (errno: %s)",
2617 eth_dev->data->port_id, strerror(rte_errno));
2622 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2623 eth_dev->data->port_id,
2624 mac.addr_bytes[0], mac.addr_bytes[1],
2625 mac.addr_bytes[2], mac.addr_bytes[3],
2626 mac.addr_bytes[4], mac.addr_bytes[5]);
2629 char ifname[IF_NAMESIZE];
2631 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2632 DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2633 eth_dev->data->port_id, ifname);
2635 DRV_LOG(DEBUG, "port %u ifname is unknown",
2636 eth_dev->data->port_id);
2639 /* Get actual MTU if possible. */
2640 err = mlx5_get_mtu(eth_dev, &priv->mtu);
2645 DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2647 /* Initialize burst functions to prevent crashes before link-up. */
2648 eth_dev->rx_pkt_burst = removed_rx_burst;
2649 eth_dev->tx_pkt_burst = removed_tx_burst;
2650 eth_dev->dev_ops = &mlx5_dev_ops;
2651 /* Register MAC address. */
2652 claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2653 if (config.vf && config.vf_nl_en)
2654 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
2655 mlx5_ifindex(eth_dev),
2656 eth_dev->data->mac_addrs,
2657 MLX5_MAX_MAC_ADDRESSES);
2658 TAILQ_INIT(&priv->flows);
2659 TAILQ_INIT(&priv->ctrl_flows);
2660 TAILQ_INIT(&priv->flow_meters);
2661 TAILQ_INIT(&priv->flow_meter_profiles);
2662 /* Hint libmlx5 to use PMD allocator for data plane resources */
2663 struct mlx5dv_ctx_allocators alctr = {
2664 .alloc = &mlx5_alloc_verbs_buf,
2665 .free = &mlx5_free_verbs_buf,
2668 mlx5_glue->dv_set_context_attr(sh->ctx,
2669 MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2670 (void *)((uintptr_t)&alctr));
2671 /* Bring Ethernet device up. */
2672 DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2673 eth_dev->data->port_id);
2674 mlx5_set_link_up(eth_dev);
2676 * Even though the interrupt handler is not installed yet,
2677 * interrupts will still trigger on the async_fd from
2678 * Verbs context returned by ibv_open_device().
2680 mlx5_link_update(eth_dev, 0);
2681 #ifdef HAVE_MLX5DV_DR_ESWITCH
2682 if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2683 (switch_info->representor || switch_info->master)))
2684 config.dv_esw_en = 0;
2686 config.dv_esw_en = 0;
2688 /* Detect minimal data bytes to inline. */
2689 mlx5_set_min_inline(spawn, &config);
2690 /* Store device configuration on private structure. */
2691 priv->config = config;
2692 /* Create context for virtual machine VLAN workaround. */
2693 priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2694 if (config.dv_flow_en) {
2695 err = mlx5_alloc_shared_dr(priv);
2699 * RSS id is shared with meter flow id. Meter flow id can only
2700 * use the 24 MSB of the register.
2702 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
2703 MLX5_MTR_COLOR_BITS);
2704 if (!priv->qrss_id_pool) {
2705 DRV_LOG(ERR, "can't create flow id pool");
2710 /* Supported Verbs flow priority number detection. */
2711 err = mlx5_flow_discover_priorities(eth_dev);
2716 priv->config.flow_prio = err;
2717 if (!priv->config.dv_esw_en &&
2718 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2719 DRV_LOG(WARNING, "metadata mode %u is not supported "
2720 "(no E-Switch)", priv->config.dv_xmeta_en);
2721 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2723 mlx5_set_metadata_mask(eth_dev);
2724 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2725 !priv->sh->dv_regc0_mask) {
2726 DRV_LOG(ERR, "metadata mode %u is not supported "
2727 "(no metadata reg_c[0] is available)",
2728 priv->config.dv_xmeta_en);
2732 /* Query availibility of metadata reg_c's. */
2733 err = mlx5_flow_discover_mreg_c(eth_dev);
2738 if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
2740 "port %u extensive metadata register is not supported",
2741 eth_dev->data->port_id);
2742 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2743 DRV_LOG(ERR, "metadata mode %u is not supported "
2744 "(no metadata registers available)",
2745 priv->config.dv_xmeta_en);
2750 if (priv->config.dv_flow_en &&
2751 priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2752 mlx5_flow_ext_mreg_supported(eth_dev) &&
2753 priv->sh->dv_regc0_mask) {
2754 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
2755 MLX5_FLOW_MREG_HTABLE_SZ);
2756 if (!priv->mreg_cp_tbl) {
2764 if (priv->mreg_cp_tbl)
2765 mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
2767 mlx5_free_shared_dr(priv);
2768 if (priv->nl_socket_route >= 0)
2769 close(priv->nl_socket_route);
2770 if (priv->nl_socket_rdma >= 0)
2771 close(priv->nl_socket_rdma);
2772 if (priv->vmwa_context)
2773 mlx5_vlan_vmwa_exit(priv->vmwa_context);
2774 if (priv->qrss_id_pool)
2775 mlx5_flow_id_pool_release(priv->qrss_id_pool);
2777 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2779 if (eth_dev != NULL)
2780 eth_dev->data->dev_private = NULL;
2782 if (eth_dev != NULL) {
2783 /* mac_addrs must not be freed alone because part of dev_private */
2784 eth_dev->data->mac_addrs = NULL;
2785 rte_eth_dev_release_port(eth_dev);
2788 mlx5_free_shared_ibctx(sh);
2795 * Comparison callback to sort device data.
2797 * This is meant to be used with qsort().
2800 * Pointer to pointer to first data object.
2802 * Pointer to pointer to second data object.
2805 * 0 if both objects are equal, less than 0 if the first argument is less
2806 * than the second, greater than 0 otherwise.
2809 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2811 const struct mlx5_switch_info *si_a =
2812 &((const struct mlx5_dev_spawn_data *)a)->info;
2813 const struct mlx5_switch_info *si_b =
2814 &((const struct mlx5_dev_spawn_data *)b)->info;
2817 /* Master device first. */
2818 ret = si_b->master - si_a->master;
2821 /* Then representor devices. */
2822 ret = si_b->representor - si_a->representor;
2825 /* Unidentified devices come last in no specific order. */
2826 if (!si_a->representor)
2828 /* Order representors by name. */
2829 return si_a->port_name - si_b->port_name;
2833 * Match PCI information for possible slaves of bonding device.
2835 * @param[in] ibv_dev
2836 * Pointer to Infiniband device structure.
2837 * @param[in] pci_dev
2838 * Pointer to PCI device structure to match PCI address.
2839 * @param[in] nl_rdma
2840 * Netlink RDMA group socket handle.
2843 * negative value if no bonding device found, otherwise
2844 * positive index of slave PF in bonding.
2847 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2848 const struct rte_pci_device *pci_dev,
2851 char ifname[IF_NAMESIZE + 1];
2852 unsigned int ifindex;
2858 * Try to get master device name. If something goes
2859 * wrong suppose the lack of kernel support and no
2864 if (!strstr(ibv_dev->name, "bond"))
2866 np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2870 * The Master device might not be on the predefined
2871 * port (not on port index 1, it is not garanted),
2872 * we have to scan all Infiniband device port and
2875 for (i = 1; i <= np; ++i) {
2876 /* Check whether Infiniband port is populated. */
2877 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2880 if (!if_indextoname(ifindex, ifname))
2882 /* Try to read bonding slave names from sysfs. */
2884 "/sys/class/net/%s/master/bonding/slaves", ifname);
2885 file = fopen(slaves, "r");
2891 /* Use safe format to check maximal buffer length. */
2892 assert(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2893 while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2894 char tmp_str[IF_NAMESIZE + 32];
2895 struct rte_pci_addr pci_addr;
2896 struct mlx5_switch_info info;
2898 /* Process slave interface names in the loop. */
2899 snprintf(tmp_str, sizeof(tmp_str),
2900 "/sys/class/net/%s", ifname);
2901 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2902 DRV_LOG(WARNING, "can not get PCI address"
2903 " for netdev \"%s\"", ifname);
2906 if (pci_dev->addr.domain != pci_addr.domain ||
2907 pci_dev->addr.bus != pci_addr.bus ||
2908 pci_dev->addr.devid != pci_addr.devid ||
2909 pci_dev->addr.function != pci_addr.function)
2911 /* Slave interface PCI address match found. */
2913 snprintf(tmp_str, sizeof(tmp_str),
2914 "/sys/class/net/%s/phys_port_name", ifname);
2915 file = fopen(tmp_str, "rb");
2918 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2919 if (fscanf(file, "%32s", tmp_str) == 1)
2920 mlx5_translate_port_name(tmp_str, &info);
2921 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2922 info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2923 pf = info.port_name;
2932 * DPDK callback to register a PCI device.
2934 * This function spawns Ethernet devices out of a given PCI device.
2936 * @param[in] pci_drv
2937 * PCI driver structure (mlx5_driver).
2938 * @param[in] pci_dev
2939 * PCI device information.
2942 * 0 on success, a negative errno value otherwise and rte_errno is set.
2945 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2946 struct rte_pci_device *pci_dev)
2948 struct ibv_device **ibv_list;
2950 * Number of found IB Devices matching with requested PCI BDF.
2951 * nd != 1 means there are multiple IB devices over the same
2952 * PCI device and we have representors and master.
2954 unsigned int nd = 0;
2956 * Number of found IB device Ports. nd = 1 and np = 1..n means
2957 * we have the single multiport IB device, and there may be
2958 * representors attached to some of found ports.
2960 unsigned int np = 0;
2962 * Number of DPDK ethernet devices to Spawn - either over
2963 * multiple IB devices or multiple ports of single IB device.
2964 * Actually this is the number of iterations to spawn.
2966 unsigned int ns = 0;
2969 * < 0 - no bonding device (single one)
2970 * >= 0 - bonding device (value is slave PF index)
2973 struct mlx5_dev_spawn_data *list = NULL;
2974 struct mlx5_dev_config dev_config;
2977 if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
2978 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
2982 if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2983 mlx5_pmd_socket_init();
2984 ret = mlx5_init_once();
2986 DRV_LOG(ERR, "unable to init PMD global data: %s",
2987 strerror(rte_errno));
2990 assert(pci_drv == &mlx5_driver);
2992 ibv_list = mlx5_glue->get_device_list(&ret);
2994 rte_errno = errno ? errno : ENOSYS;
2995 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2999 * First scan the list of all Infiniband devices to find
3000 * matching ones, gathering into the list.
3002 struct ibv_device *ibv_match[ret + 1];
3003 int nl_route = mlx5_nl_init(NETLINK_ROUTE);
3004 int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
3008 struct rte_pci_addr pci_addr;
3010 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
3011 bd = mlx5_device_bond_pci_match
3012 (ibv_list[ret], pci_dev, nl_rdma);
3015 * Bonding device detected. Only one match is allowed,
3016 * the bonding is supported over multi-port IB device,
3017 * there should be no matches on representor PCI
3018 * functions or non VF LAG bonding devices with
3019 * specified address.
3023 "multiple PCI match on bonding device"
3024 "\"%s\" found", ibv_list[ret]->name);
3029 DRV_LOG(INFO, "PCI information matches for"
3030 " slave %d bonding device \"%s\"",
3031 bd, ibv_list[ret]->name);
3032 ibv_match[nd++] = ibv_list[ret];
3035 if (mlx5_dev_to_pci_addr
3036 (ibv_list[ret]->ibdev_path, &pci_addr))
3038 if (pci_dev->addr.domain != pci_addr.domain ||
3039 pci_dev->addr.bus != pci_addr.bus ||
3040 pci_dev->addr.devid != pci_addr.devid ||
3041 pci_dev->addr.function != pci_addr.function)
3043 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
3044 ibv_list[ret]->name);
3045 ibv_match[nd++] = ibv_list[ret];
3047 ibv_match[nd] = NULL;
3049 /* No device matches, just complain and bail out. */
3051 "no Verbs device matches PCI device " PCI_PRI_FMT ","
3052 " are kernel drivers loaded?",
3053 pci_dev->addr.domain, pci_dev->addr.bus,
3054 pci_dev->addr.devid, pci_dev->addr.function);
3061 * Found single matching device may have multiple ports.
3062 * Each port may be representor, we have to check the port
3063 * number and check the representors existence.
3066 np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
3068 DRV_LOG(WARNING, "can not get IB device \"%s\""
3069 " ports number", ibv_match[0]->name);
3070 if (bd >= 0 && !np) {
3071 DRV_LOG(ERR, "can not get ports"
3072 " for bonding device");
3078 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
3081 * This may happen if there is VF LAG kernel support and
3082 * application is compiled with older rdma_core library.
3085 "No kernel/verbs support for VF LAG bonding found.");
3086 rte_errno = ENOTSUP;
3092 * Now we can determine the maximal
3093 * amount of devices to be spawned.
3095 list = rte_zmalloc("device spawn data",
3096 sizeof(struct mlx5_dev_spawn_data) *
3098 RTE_CACHE_LINE_SIZE);
3100 DRV_LOG(ERR, "spawn data array allocation failure");
3105 if (bd >= 0 || np > 1) {
3107 * Single IB device with multiple ports found,
3108 * it may be E-Switch master device and representors.
3109 * We have to perform identification trough the ports.
3111 assert(nl_rdma >= 0);
3115 for (i = 1; i <= np; ++i) {
3116 list[ns].max_port = np;
3117 list[ns].ibv_port = i;
3118 list[ns].ibv_dev = ibv_match[0];
3119 list[ns].eth_dev = NULL;
3120 list[ns].pci_dev = pci_dev;
3121 list[ns].pf_bond = bd;
3122 list[ns].ifindex = mlx5_nl_ifindex
3123 (nl_rdma, list[ns].ibv_dev->name, i);
3124 if (!list[ns].ifindex) {
3126 * No network interface index found for the
3127 * specified port, it means there is no
3128 * representor on this port. It's OK,
3129 * there can be disabled ports, for example
3130 * if sriov_numvfs < sriov_totalvfs.
3136 ret = mlx5_nl_switch_info
3140 if (ret || (!list[ns].info.representor &&
3141 !list[ns].info.master)) {
3143 * We failed to recognize representors with
3144 * Netlink, let's try to perform the task
3147 ret = mlx5_sysfs_switch_info
3151 if (!ret && bd >= 0) {
3152 switch (list[ns].info.name_type) {
3153 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
3154 if (list[ns].info.port_name == bd)
3157 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
3158 if (list[ns].info.pf_num == bd)
3166 if (!ret && (list[ns].info.representor ^
3167 list[ns].info.master))
3172 "unable to recognize master/representors"
3173 " on the IB device with multiple ports");
3180 * The existence of several matching entries (nd > 1) means
3181 * port representors have been instantiated. No existing Verbs
3182 * call nor sysfs entries can tell them apart, this can only
3183 * be done through Netlink calls assuming kernel drivers are
3184 * recent enough to support them.
3186 * In the event of identification failure through Netlink,
3187 * try again through sysfs, then:
3189 * 1. A single IB device matches (nd == 1) with single
3190 * port (np=0/1) and is not a representor, assume
3191 * no switch support.
3193 * 2. Otherwise no safe assumptions can be made;
3194 * complain louder and bail out.
3197 for (i = 0; i != nd; ++i) {
3198 memset(&list[ns].info, 0, sizeof(list[ns].info));
3199 list[ns].max_port = 1;
3200 list[ns].ibv_port = 1;
3201 list[ns].ibv_dev = ibv_match[i];
3202 list[ns].eth_dev = NULL;
3203 list[ns].pci_dev = pci_dev;
3204 list[ns].pf_bond = -1;
3205 list[ns].ifindex = 0;
3207 list[ns].ifindex = mlx5_nl_ifindex
3208 (nl_rdma, list[ns].ibv_dev->name, 1);
3209 if (!list[ns].ifindex) {
3210 char ifname[IF_NAMESIZE];
3213 * Netlink failed, it may happen with old
3214 * ib_core kernel driver (before 4.16).
3215 * We can assume there is old driver because
3216 * here we are processing single ports IB
3217 * devices. Let's try sysfs to retrieve
3218 * the ifindex. The method works for
3219 * master device only.
3223 * Multiple devices found, assume
3224 * representors, can not distinguish
3225 * master/representor and retrieve
3226 * ifindex via sysfs.
3230 ret = mlx5_get_master_ifname
3231 (ibv_match[i]->ibdev_path, &ifname);
3234 if_nametoindex(ifname);
3235 if (!list[ns].ifindex) {
3237 * No network interface index found
3238 * for the specified device, it means
3239 * there it is neither representor
3247 ret = mlx5_nl_switch_info
3251 if (ret || (!list[ns].info.representor &&
3252 !list[ns].info.master)) {
3254 * We failed to recognize representors with
3255 * Netlink, let's try to perform the task
3258 ret = mlx5_sysfs_switch_info
3262 if (!ret && (list[ns].info.representor ^
3263 list[ns].info.master)) {
3265 } else if ((nd == 1) &&
3266 !list[ns].info.representor &&
3267 !list[ns].info.master) {
3269 * Single IB device with
3270 * one physical port and
3271 * attached network device.
3272 * May be SRIOV is not enabled
3273 * or there is no representors.
3275 DRV_LOG(INFO, "no E-Switch support detected");
3282 "unable to recognize master/representors"
3283 " on the multiple IB devices");
3291 * Sort list to probe devices in natural order for users convenience
3292 * (i.e. master first, then representors from lowest to highest ID).
3294 qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3295 /* Default configuration. */
3296 dev_config = (struct mlx5_dev_config){
3298 .mps = MLX5_ARG_UNSET,
3299 .dbnc = MLX5_ARG_UNSET,
3301 .txq_inline_max = MLX5_ARG_UNSET,
3302 .txq_inline_min = MLX5_ARG_UNSET,
3303 .txq_inline_mpw = MLX5_ARG_UNSET,
3304 .txqs_inline = MLX5_ARG_UNSET,
3306 .mr_ext_memseg_en = 1,
3308 .enabled = 0, /* Disabled by default. */
3309 .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
3310 .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3311 .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3316 /* Device specific configuration. */
3317 switch (pci_dev->id.device_id) {
3318 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3319 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3320 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3321 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3322 case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3323 case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3324 case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3330 for (i = 0; i != ns; ++i) {
3333 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3336 if (!list[i].eth_dev) {
3337 if (rte_errno != EBUSY && rte_errno != EEXIST)
3339 /* Device is disabled or already spawned. Ignore it. */
3342 restore = list[i].eth_dev->data->dev_flags;
3343 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3344 /* Restore non-PCI flags cleared by the above call. */
3345 list[i].eth_dev->data->dev_flags |= restore;
3346 mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3347 rte_eth_dev_probing_finish(list[i].eth_dev);
3351 "probe of PCI device " PCI_PRI_FMT " aborted after"
3352 " encountering an error: %s",
3353 pci_dev->addr.domain, pci_dev->addr.bus,
3354 pci_dev->addr.devid, pci_dev->addr.function,
3355 strerror(rte_errno));
3359 if (!list[i].eth_dev)
3361 mlx5_dev_close(list[i].eth_dev);
3362 /* mac_addrs must not be freed because in dev_private */
3363 list[i].eth_dev->data->mac_addrs = NULL;
3364 claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3366 /* Restore original error. */
3373 * Do the routine cleanup:
3374 * - close opened Netlink sockets
3375 * - free allocated spawn data array
3376 * - free the Infiniband device list
3385 mlx5_glue->free_device_list(ibv_list);
3390 * Look for the ethernet device belonging to mlx5 driver.
3392 * @param[in] port_id
3393 * port_id to start looking for device.
3394 * @param[in] pci_dev
3395 * Pointer to the hint PCI device. When device is being probed
3396 * the its siblings (master and preceding representors might
3397 * not have assigned driver yet (because the mlx5_pci_probe()
3398 * is not completed yet, for this case match on hint PCI
3399 * device may be used to detect sibling device.
3402 * port_id of found device, RTE_MAX_ETHPORT if not found.
3405 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3407 while (port_id < RTE_MAX_ETHPORTS) {
3408 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3410 if (dev->state != RTE_ETH_DEV_UNUSED &&
3412 (dev->device == &pci_dev->device ||
3413 (dev->device->driver &&
3414 dev->device->driver->name &&
3415 !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3419 if (port_id >= RTE_MAX_ETHPORTS)
3420 return RTE_MAX_ETHPORTS;
3425 * DPDK callback to remove a PCI device.
3427 * This function removes all Ethernet devices belong to a given PCI device.
3429 * @param[in] pci_dev
3430 * Pointer to the PCI device.
3433 * 0 on success, the function cannot fail.
3436 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3440 RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3441 rte_eth_dev_close(port_id);
3445 static const struct rte_pci_id mlx5_pci_id_map[] = {
3447 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3448 PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3451 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3452 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3455 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3456 PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3459 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3460 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3463 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3464 PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3467 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3468 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3471 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3472 PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3475 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3476 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3479 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3480 PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3483 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3484 PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3487 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3488 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3491 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3492 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3495 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3496 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3499 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3500 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3507 static struct rte_pci_driver mlx5_driver = {
3509 .name = MLX5_DRIVER_NAME
3511 .id_table = mlx5_pci_id_map,
3512 .probe = mlx5_pci_probe,
3513 .remove = mlx5_pci_remove,
3514 .dma_map = mlx5_dma_map,
3515 .dma_unmap = mlx5_dma_unmap,
3516 .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3517 RTE_PCI_DRV_PROBE_AGAIN,
3521 * Driver initialization routine.
3523 RTE_INIT(rte_mlx5_pmd_init)
3525 /* Initialize driver log type. */
3526 mlx5_logtype = rte_log_register("pmd.net.mlx5");
3527 if (mlx5_logtype >= 0)
3528 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3530 /* Build the static tables for Verbs conversion. */
3531 mlx5_set_ptype_table();
3532 mlx5_set_cksum_table();
3533 mlx5_set_swp_types_table();
3535 rte_pci_register(&mlx5_driver);
3538 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3539 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3540 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");