1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
14 #include <linux/rtnetlink.h>
17 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
19 #pragma GCC diagnostic ignored "-Wpedantic"
21 #include <infiniband/verbs.h>
23 #pragma GCC diagnostic error "-Wpedantic"
26 #include <rte_malloc.h>
27 #include <rte_ethdev_driver.h>
28 #include <rte_ethdev_pci.h>
30 #include <rte_bus_pci.h>
31 #include <rte_common.h>
32 #include <rte_kvargs.h>
33 #include <rte_rwlock.h>
34 #include <rte_spinlock.h>
35 #include <rte_string_fns.h>
36 #include <rte_alarm.h>
38 #include <mlx5_glue.h>
39 #include <mlx5_devx_cmds.h>
40 #include <mlx5_common.h>
41 #include <mlx5_common_os.h>
42 #include <mlx5_common_mp.h>
43 #include <mlx5_malloc.h>
45 #include "mlx5_defs.h"
47 #include "mlx5_utils.h"
48 #include "mlx5_rxtx.h"
49 #include "mlx5_autoconf.h"
51 #include "mlx5_flow.h"
52 #include "rte_pmd_mlx5.h"
54 /* Device parameter to enable RX completion queue compression. */
55 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
57 /* Device parameter to enable RX completion entry padding to 128B. */
58 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
60 /* Device parameter to enable padding Rx packet to cacheline size. */
61 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
63 /* Device parameter to enable Multi-Packet Rx queue. */
64 #define MLX5_RX_MPRQ_EN "mprq_en"
66 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
67 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
69 /* Device parameter to configure log 2 of the stride size for MPRQ. */
70 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
72 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
73 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
75 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
76 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
78 /* Device parameter to configure inline send. Deprecated, ignored.*/
79 #define MLX5_TXQ_INLINE "txq_inline"
81 /* Device parameter to limit packet size to inline with ordinary SEND. */
82 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
84 /* Device parameter to configure minimal data size to inline. */
85 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
87 /* Device parameter to limit packet size to inline with Enhanced MPW. */
88 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
91 * Device parameter to configure the number of TX queues threshold for
92 * enabling inline send.
94 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
97 * Device parameter to configure the number of TX queues threshold for
98 * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
100 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
102 /* Device parameter to enable multi-packet send WQEs. */
103 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
106 * Device parameter to force doorbell register mapping
107 * to non-cahed region eliminating the extra write memory barrier.
109 #define MLX5_TX_DB_NC "tx_db_nc"
112 * Device parameter to include 2 dsegs in the title WQEBB.
113 * Deprecated, ignored.
115 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
118 * Device parameter to limit the size of inlining packet.
119 * Deprecated, ignored.
121 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
124 * Device parameter to enable Tx scheduling on timestamps
125 * and specify the packet pacing granularity in nanoseconds.
127 #define MLX5_TX_PP "tx_pp"
130 * Device parameter to specify skew in nanoseconds on Tx datapath,
131 * it represents the time between SQ start WQE processing and
132 * appearing actual packet data on the wire.
134 #define MLX5_TX_SKEW "tx_skew"
137 * Device parameter to enable hardware Tx vector.
138 * Deprecated, ignored (no vectorized Tx routines anymore).
140 #define MLX5_TX_VEC_EN "tx_vec_en"
142 /* Device parameter to enable hardware Rx vector. */
143 #define MLX5_RX_VEC_EN "rx_vec_en"
145 /* Allow L3 VXLAN flow creation. */
146 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
148 /* Activate DV E-Switch flow steering. */
149 #define MLX5_DV_ESW_EN "dv_esw_en"
151 /* Activate DV flow steering. */
152 #define MLX5_DV_FLOW_EN "dv_flow_en"
154 /* Enable extensive flow metadata support. */
155 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
157 /* Device parameter to let the user manage the lacp traffic of bonded device */
158 #define MLX5_LACP_BY_USER "lacp_by_user"
160 /* Activate Netlink support in VF mode. */
161 #define MLX5_VF_NL_EN "vf_nl_en"
163 /* Enable extending memsegs when creating a MR. */
164 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
166 /* Select port representors to instantiate. */
167 #define MLX5_REPRESENTOR "representor"
169 /* Device parameter to configure the maximum number of dump files per queue. */
170 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
172 /* Configure timeout of LRO session (in microseconds). */
173 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
176 * Device parameter to configure the total data buffer size for a single
177 * hairpin queue (logarithm value).
179 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
181 /* Flow memory reclaim mode. */
182 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
184 /* The default memory allocator used in PMD. */
185 #define MLX5_SYS_MEM_EN "sys_mem_en"
186 /* Decap will be used or not. */
187 #define MLX5_DECAP_EN "decap_en"
189 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
191 /* Shared memory between primary and secondary processes. */
192 struct mlx5_shared_data *mlx5_shared_data;
194 /* Spinlock for mlx5_shared_data allocation. */
195 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
197 /* Process local data for secondary processes. */
198 static struct mlx5_local_data mlx5_local_data;
200 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
201 LIST_HEAD_INITIALIZER();
202 static pthread_mutex_t mlx5_dev_ctx_list_mutex = PTHREAD_MUTEX_INITIALIZER;
204 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
205 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
207 .size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
213 .malloc = mlx5_malloc,
215 .type = "mlx5_encap_decap_ipool",
218 .size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
224 .malloc = mlx5_malloc,
226 .type = "mlx5_push_vlan_ipool",
229 .size = sizeof(struct mlx5_flow_dv_tag_resource),
235 .malloc = mlx5_malloc,
237 .type = "mlx5_tag_ipool",
240 .size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
246 .malloc = mlx5_malloc,
248 .type = "mlx5_port_id_ipool",
251 .size = sizeof(struct mlx5_flow_tbl_data_entry),
257 .malloc = mlx5_malloc,
259 .type = "mlx5_jump_ipool",
263 .size = sizeof(struct mlx5_flow_meter),
269 .malloc = mlx5_malloc,
271 .type = "mlx5_meter_ipool",
274 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
280 .malloc = mlx5_malloc,
282 .type = "mlx5_mcp_ipool",
285 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
291 .malloc = mlx5_malloc,
293 .type = "mlx5_hrxq_ipool",
297 * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
298 * It set in run time according to PCI function configuration.
306 .malloc = mlx5_malloc,
308 .type = "mlx5_flow_handle_ipool",
311 .size = sizeof(struct rte_flow),
315 .malloc = mlx5_malloc,
317 .type = "rte_flow_ipool",
322 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
323 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
325 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
328 * Allocate ID pool structure.
331 * The maximum id can be allocated from the pool.
334 * Pointer to pool object, NULL value otherwise.
336 struct mlx5_flow_id_pool *
337 mlx5_flow_id_pool_alloc(uint32_t max_id)
339 struct mlx5_flow_id_pool *pool;
342 pool = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*pool),
343 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
345 DRV_LOG(ERR, "can't allocate id pool");
349 mem = mlx5_malloc(MLX5_MEM_ZERO,
350 MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
351 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
353 DRV_LOG(ERR, "can't allocate mem for id pool");
357 pool->free_arr = mem;
358 pool->curr = pool->free_arr;
359 pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
360 pool->base_index = 0;
361 pool->max_id = max_id;
369 * Release ID pool structure.
372 * Pointer to flow id pool object to free.
375 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
377 mlx5_free(pool->free_arr);
385 * Pointer to flow id pool.
390 * 0 on success, error value otherwise.
393 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
395 if (pool->curr == pool->free_arr) {
396 if (pool->base_index == pool->max_id) {
398 DRV_LOG(ERR, "no free id");
401 *id = ++pool->base_index;
404 *id = *(--pool->curr);
412 * Pointer to flow id pool.
417 * 0 on success, error value otherwise.
420 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
426 if (pool->curr == pool->last) {
427 size = pool->curr - pool->free_arr;
428 size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
429 MLX5_ASSERT(size2 > size);
430 mem = mlx5_malloc(0, size2 * sizeof(uint32_t), 0,
433 DRV_LOG(ERR, "can't allocate mem for id pool");
437 memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
438 mlx5_free(pool->free_arr);
439 pool->free_arr = mem;
440 pool->curr = pool->free_arr + size;
441 pool->last = pool->free_arr + size2;
449 * Initialize the shared aging list information per port.
452 * Pointer to mlx5_dev_ctx_shared object.
455 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
458 struct mlx5_age_info *age_info;
460 for (i = 0; i < sh->max_port; i++) {
461 age_info = &sh->port[i].age_info;
463 TAILQ_INIT(&age_info->aged_counters);
464 rte_spinlock_init(&age_info->aged_sl);
465 MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
470 * Initialize the counters management structure.
473 * Pointer to mlx5_dev_ctx_shared object to free
476 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
480 memset(&sh->cmng, 0, sizeof(sh->cmng));
481 TAILQ_INIT(&sh->cmng.flow_counters);
482 for (i = 0; i < MLX5_CCONT_TYPE_MAX; ++i) {
483 sh->cmng.ccont[i].min_id = MLX5_CNT_BATCH_OFFSET;
484 sh->cmng.ccont[i].max_id = -1;
485 sh->cmng.ccont[i].last_pool_idx = POOL_IDX_INVALID;
486 TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
487 rte_spinlock_init(&sh->cmng.ccont[i].resize_sl);
488 TAILQ_INIT(&sh->cmng.ccont[i].counters);
489 rte_spinlock_init(&sh->cmng.ccont[i].csl);
494 * Destroy all the resources allocated for a counter memory management.
497 * Pointer to the memory management structure.
500 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
502 uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
504 LIST_REMOVE(mng, next);
505 claim_zero(mlx5_devx_cmd_destroy(mng->dm));
506 claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
511 * Close and release all the resources of the counters management.
514 * Pointer to mlx5_dev_ctx_shared object to free.
517 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
519 struct mlx5_counter_stats_mem_mng *mng;
526 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
527 if (rte_errno != EINPROGRESS)
531 for (i = 0; i < MLX5_CCONT_TYPE_MAX; ++i) {
532 struct mlx5_flow_counter_pool *pool;
533 uint32_t batch = !!(i > 1);
535 if (!sh->cmng.ccont[i].pools)
537 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
539 if (batch && pool->min_dcs)
540 claim_zero(mlx5_devx_cmd_destroy
542 for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
543 if (MLX5_POOL_GET_CNT(pool, j)->action)
545 (mlx5_glue->destroy_flow_action
548 if (!batch && MLX5_GET_POOL_CNT_EXT
550 claim_zero(mlx5_devx_cmd_destroy
551 (MLX5_GET_POOL_CNT_EXT
554 TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool, next);
556 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
558 mlx5_free(sh->cmng.ccont[i].pools);
560 mng = LIST_FIRST(&sh->cmng.mem_mngs);
562 mlx5_flow_destroy_counter_stat_mem_mng(mng);
563 mng = LIST_FIRST(&sh->cmng.mem_mngs);
565 memset(&sh->cmng, 0, sizeof(sh->cmng));
569 * Initialize the flow resources' indexed mempool.
572 * Pointer to mlx5_dev_ctx_shared object.
574 * Pointer to user dev config.
577 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
578 const struct mlx5_dev_config *config)
581 struct mlx5_indexed_pool_config cfg;
583 for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
584 cfg = mlx5_ipool_cfg[i];
589 * Set MLX5_IPOOL_MLX5_FLOW ipool size
590 * according to PCI function flow configuration.
592 case MLX5_IPOOL_MLX5_FLOW:
593 cfg.size = config->dv_flow_en ?
594 sizeof(struct mlx5_flow_handle) :
595 MLX5_FLOW_HANDLE_VERBS_SIZE;
598 if (config->reclaim_mode)
599 cfg.release_mem_en = 1;
600 sh->ipool[i] = mlx5_ipool_create(&cfg);
605 * Release the flow resources' indexed mempool.
608 * Pointer to mlx5_dev_ctx_shared object.
611 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
615 for (i = 0; i < MLX5_IPOOL_MAX; ++i)
616 mlx5_ipool_destroy(sh->ipool[i]);
620 * Check if dynamic flex parser for eCPRI already exists.
623 * Pointer to Ethernet device structure.
626 * true on exists, false on not.
629 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
631 struct mlx5_priv *priv = dev->data->dev_private;
632 struct mlx5_flex_parser_profiles *prf =
633 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
639 * Allocation of a flex parser for eCPRI. Once created, this parser related
640 * resources will be held until the device is closed.
643 * Pointer to Ethernet device structure.
646 * 0 on success, a negative errno value otherwise and rte_errno is set.
649 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
651 struct mlx5_priv *priv = dev->data->dev_private;
652 struct mlx5_flex_parser_profiles *prf =
653 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
654 struct mlx5_devx_graph_node_attr node = {
655 .modify_field_select = 0,
660 if (!priv->config.hca_attr.parse_graph_flex_node) {
661 DRV_LOG(ERR, "Dynamic flex parser is not supported "
662 "for device %s.", priv->dev_data->name);
665 node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
666 /* 8 bytes now: 4B common header + 4B message body header. */
667 node.header_length_base_value = 0x8;
668 /* After MAC layer: Ether / VLAN. */
669 node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
670 /* Type of compared condition should be 0xAEFE in the L2 layer. */
671 node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
672 /* Sample #0: type in common header. */
673 node.sample[0].flow_match_sample_en = 1;
675 node.sample[0].flow_match_sample_offset_mode = 0x0;
676 /* Only the 2nd byte will be used. */
677 node.sample[0].flow_match_sample_field_base_offset = 0x0;
678 /* Sample #1: message payload. */
679 node.sample[1].flow_match_sample_en = 1;
681 node.sample[1].flow_match_sample_offset_mode = 0x0;
683 * Only the first two bytes will be used right now, and its offset will
684 * start after the common header that with the length of a DW(u32).
686 node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
687 prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
689 DRV_LOG(ERR, "Failed to create flex parser node object.");
690 return (rte_errno == 0) ? -ENODEV : -rte_errno;
693 ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
695 DRV_LOG(ERR, "Failed to query sample IDs.");
696 return (rte_errno == 0) ? -ENODEV : -rte_errno;
698 prf->offset[0] = 0x0;
699 prf->offset[1] = sizeof(uint32_t);
700 prf->ids[0] = ids[0];
701 prf->ids[1] = ids[1];
706 * Destroy the flex parser node, including the parser itself, input / output
707 * arcs and DW samples. Resources could be reused then.
710 * Pointer to Ethernet device structure.
713 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
715 struct mlx5_priv *priv = dev->data->dev_private;
716 struct mlx5_flex_parser_profiles *prf =
717 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
720 mlx5_devx_cmd_destroy(prf->obj);
725 * Allocate shared device context. If there is multiport device the
726 * master and representors will share this context, if there is single
727 * port dedicated device, the context will be used by only given
728 * port due to unification.
730 * Routine first searches the context for the specified device name,
731 * if found the shared context assumed and reference counter is incremented.
732 * If no context found the new one is created and initialized with specified
733 * device context and parameters.
736 * Pointer to the device attributes (name, port, etc).
738 * Pointer to device configuration structure.
741 * Pointer to mlx5_dev_ctx_shared object on success,
742 * otherwise NULL and rte_errno is set.
744 struct mlx5_dev_ctx_shared *
745 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
746 const struct mlx5_dev_config *config)
748 struct mlx5_dev_ctx_shared *sh;
751 struct mlx5_devx_tis_attr tis_attr = { 0 };
754 /* Secondary process should not create the shared context. */
755 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
756 pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
757 /* Search for IB context by device name. */
758 LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
759 if (!strcmp(sh->ibdev_name,
760 mlx5_os_get_dev_device_name(spawn->phys_dev))) {
765 /* No device found, we have to create new shared context. */
766 MLX5_ASSERT(spawn->max_port);
767 sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
768 sizeof(struct mlx5_dev_ctx_shared) +
770 sizeof(struct mlx5_dev_shared_port),
771 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
773 DRV_LOG(ERR, "shared context allocation failure");
777 err = mlx5_os_open_device(spawn, config, sh);
780 err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
782 DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
786 sh->max_port = spawn->max_port;
787 strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
788 sizeof(sh->ibdev_name) - 1);
789 strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
790 sizeof(sh->ibdev_path) - 1);
792 * Setting port_id to max unallowed value means
793 * there is no interrupt subhandler installed for
794 * the given port index i.
796 for (i = 0; i < sh->max_port; i++) {
797 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
798 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
800 sh->pd = mlx5_glue->alloc_pd(sh->ctx);
801 if (sh->pd == NULL) {
802 DRV_LOG(ERR, "PD allocation failure");
807 err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
809 DRV_LOG(ERR, "Fail to extract pdn from PD");
812 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
814 DRV_LOG(ERR, "TD allocation failure");
818 tis_attr.transport_domain = sh->td->id;
819 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
821 DRV_LOG(ERR, "TIS allocation failure");
825 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, 0);
827 DRV_LOG(ERR, "Failed to allocate DevX UAR.");
832 sh->flow_id_pool = mlx5_flow_id_pool_alloc
833 ((1 << HAIRPIN_FLOW_ID_BITS) - 1);
834 if (!sh->flow_id_pool) {
835 DRV_LOG(ERR, "can't create flow id pool");
840 /* Initialize UAR access locks for 32bit implementations. */
841 rte_spinlock_init(&sh->uar_lock_cq);
842 for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
843 rte_spinlock_init(&sh->uar_lock[i]);
846 * Once the device is added to the list of memory event
847 * callback, its global MR cache table cannot be expanded
848 * on the fly because of deadlock. If it overflows, lookup
849 * should be done by searching MR list linearly, which is slow.
851 * At this point the device is not added to the memory
852 * event list yet, context is just being created.
854 err = mlx5_mr_btree_init(&sh->share_cache.cache,
855 MLX5_MR_BTREE_CACHE_N * 2,
856 spawn->pci_dev->device.numa_node);
861 mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
862 &sh->share_cache.dereg_mr_cb);
863 mlx5_os_dev_shared_handler_install(sh);
864 sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
865 if (!sh->cnt_id_tbl) {
869 mlx5_flow_aging_init(sh);
870 mlx5_flow_counters_mng_init(sh);
871 mlx5_flow_ipool_create(sh, config);
872 /* Add device to memory callback list. */
873 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
874 LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
876 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
877 /* Add context to the global device list. */
878 LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
880 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
883 pthread_mutex_destroy(&sh->txpp.mutex);
884 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
886 if (sh->cnt_id_tbl) {
887 mlx5_l3t_destroy(sh->cnt_id_tbl);
888 sh->cnt_id_tbl = NULL;
891 mlx5_glue->devx_free_uar(sh->tx_uar);
895 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
897 claim_zero(mlx5_devx_cmd_destroy(sh->td));
899 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
901 claim_zero(mlx5_glue->close_device(sh->ctx));
902 if (sh->flow_id_pool)
903 mlx5_flow_id_pool_release(sh->flow_id_pool);
905 MLX5_ASSERT(err > 0);
911 * Free shared IB device context. Decrement counter and if zero free
912 * all allocated resources and close handles.
915 * Pointer to mlx5_dev_ctx_shared object to free
918 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
920 pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
921 #ifdef RTE_LIBRTE_MLX5_DEBUG
922 /* Check the object presence in the list. */
923 struct mlx5_dev_ctx_shared *lctx;
925 LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
930 DRV_LOG(ERR, "Freeing non-existing shared IB context");
935 MLX5_ASSERT(sh->refcnt);
936 /* Secondary process should not free the shared context. */
937 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
940 /* Remove from memory callback device list. */
941 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
942 LIST_REMOVE(sh, mem_event_cb);
943 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
944 /* Release created Memory Regions. */
945 mlx5_mr_release_cache(&sh->share_cache);
946 /* Remove context from the global device list. */
947 LIST_REMOVE(sh, next);
949 * Ensure there is no async event handler installed.
950 * Only primary process handles async device events.
952 mlx5_flow_counters_mng_close(sh);
953 mlx5_flow_ipool_destroy(sh);
954 mlx5_os_dev_shared_handler_uninstall(sh);
955 if (sh->cnt_id_tbl) {
956 mlx5_l3t_destroy(sh->cnt_id_tbl);
957 sh->cnt_id_tbl = NULL;
960 mlx5_glue->devx_free_uar(sh->tx_uar);
964 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
966 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
968 claim_zero(mlx5_devx_cmd_destroy(sh->td));
970 claim_zero(mlx5_glue->close_device(sh->ctx));
971 if (sh->flow_id_pool)
972 mlx5_flow_id_pool_release(sh->flow_id_pool);
973 pthread_mutex_destroy(&sh->txpp.mutex);
976 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
980 * Destroy table hash list and all the root entries per domain.
983 * Pointer to the private device data structure.
986 mlx5_free_table_hash_list(struct mlx5_priv *priv)
988 struct mlx5_dev_ctx_shared *sh = priv->sh;
989 struct mlx5_flow_tbl_data_entry *tbl_data;
990 union mlx5_flow_tbl_key table_key = {
998 struct mlx5_hlist_entry *pos;
1002 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
1004 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
1006 MLX5_ASSERT(tbl_data);
1007 mlx5_hlist_remove(sh->flow_tbls, pos);
1008 mlx5_free(tbl_data);
1010 table_key.direction = 1;
1011 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
1013 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
1015 MLX5_ASSERT(tbl_data);
1016 mlx5_hlist_remove(sh->flow_tbls, pos);
1017 mlx5_free(tbl_data);
1019 table_key.direction = 0;
1020 table_key.domain = 1;
1021 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
1023 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
1025 MLX5_ASSERT(tbl_data);
1026 mlx5_hlist_remove(sh->flow_tbls, pos);
1027 mlx5_free(tbl_data);
1029 mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
1033 * Initialize flow table hash list and create the root tables entry
1037 * Pointer to the private device data structure.
1040 * Zero on success, positive error code otherwise.
1043 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
1045 struct mlx5_dev_ctx_shared *sh = priv->sh;
1046 char s[MLX5_HLIST_NAMESIZE];
1050 snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1051 sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
1052 if (!sh->flow_tbls) {
1053 DRV_LOG(ERR, "flow tables with hash creation failed.");
1057 #ifndef HAVE_MLX5DV_DR
1059 * In case we have not DR support, the zero tables should be created
1060 * because DV expect to see them even if they cannot be created by
1063 union mlx5_flow_tbl_key table_key = {
1071 struct mlx5_flow_tbl_data_entry *tbl_data = mlx5_malloc(MLX5_MEM_ZERO,
1072 sizeof(*tbl_data), 0,
1079 tbl_data->entry.key = table_key.v64;
1080 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1083 rte_atomic32_init(&tbl_data->tbl.refcnt);
1084 rte_atomic32_inc(&tbl_data->tbl.refcnt);
1085 table_key.direction = 1;
1086 tbl_data = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*tbl_data), 0,
1092 tbl_data->entry.key = table_key.v64;
1093 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1096 rte_atomic32_init(&tbl_data->tbl.refcnt);
1097 rte_atomic32_inc(&tbl_data->tbl.refcnt);
1098 table_key.direction = 0;
1099 table_key.domain = 1;
1100 tbl_data = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*tbl_data), 0,
1106 tbl_data->entry.key = table_key.v64;
1107 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1110 rte_atomic32_init(&tbl_data->tbl.refcnt);
1111 rte_atomic32_inc(&tbl_data->tbl.refcnt);
1114 mlx5_free_table_hash_list(priv);
1115 #endif /* HAVE_MLX5DV_DR */
1120 * Initialize shared data between primary and secondary process.
1122 * A memzone is reserved by primary process and secondary processes attach to
1126 * 0 on success, a negative errno value otherwise and rte_errno is set.
1129 mlx5_init_shared_data(void)
1131 const struct rte_memzone *mz;
1134 rte_spinlock_lock(&mlx5_shared_data_lock);
1135 if (mlx5_shared_data == NULL) {
1136 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1137 /* Allocate shared memory. */
1138 mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1139 sizeof(*mlx5_shared_data),
1143 "Cannot allocate mlx5 shared data");
1147 mlx5_shared_data = mz->addr;
1148 memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1149 rte_spinlock_init(&mlx5_shared_data->lock);
1151 /* Lookup allocated shared memory. */
1152 mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1155 "Cannot attach mlx5 shared data");
1159 mlx5_shared_data = mz->addr;
1160 memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1164 rte_spinlock_unlock(&mlx5_shared_data_lock);
1169 * Retrieve integer value from environment variable.
1172 * Environment variable name.
1175 * Integer value, 0 if the variable is not set.
1178 mlx5_getenv_int(const char *name)
1180 const char *val = getenv(name);
1188 * DPDK callback to add udp tunnel port
1191 * A pointer to eth_dev
1192 * @param[in] udp_tunnel
1193 * A pointer to udp tunnel
1196 * 0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1199 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1200 struct rte_eth_udp_tunnel *udp_tunnel)
1202 MLX5_ASSERT(udp_tunnel != NULL);
1203 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1204 udp_tunnel->udp_port == 4789)
1206 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1207 udp_tunnel->udp_port == 4790)
1213 * Initialize process private data structure.
1216 * Pointer to Ethernet device structure.
1219 * 0 on success, a negative errno value otherwise and rte_errno is set.
1222 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1224 struct mlx5_priv *priv = dev->data->dev_private;
1225 struct mlx5_proc_priv *ppriv;
1229 * UAR register table follows the process private structure. BlueFlame
1230 * registers for Tx queues are stored in the table.
1233 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1234 ppriv = mlx5_malloc(MLX5_MEM_RTE, ppriv_size, RTE_CACHE_LINE_SIZE,
1235 dev->device->numa_node);
1240 ppriv->uar_table_sz = ppriv_size;
1241 dev->process_private = ppriv;
1246 * Un-initialize process private data structure.
1249 * Pointer to Ethernet device structure.
1252 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1254 if (!dev->process_private)
1256 mlx5_free(dev->process_private);
1257 dev->process_private = NULL;
1261 * DPDK callback to close the device.
1263 * Destroy all queues and objects, free memory.
1266 * Pointer to Ethernet device structure.
1269 mlx5_dev_close(struct rte_eth_dev *dev)
1271 struct mlx5_priv *priv = dev->data->dev_private;
1275 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1276 /* Check if process_private released. */
1277 if (!dev->process_private)
1279 mlx5_tx_uar_uninit_secondary(dev);
1280 mlx5_proc_priv_uninit(dev);
1281 rte_eth_dev_release_port(dev);
1286 DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1288 ((priv->sh->ctx != NULL) ?
1289 mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1291 * If default mreg copy action is removed at the stop stage,
1292 * the search will return none and nothing will be done anymore.
1294 mlx5_flow_stop_default(dev);
1295 mlx5_traffic_disable(dev);
1297 * If all the flows are already flushed in the device stop stage,
1298 * then this will return directly without any action.
1300 mlx5_flow_list_flush(dev, &priv->flows, true);
1301 mlx5_flow_meter_flush(dev, NULL);
1302 /* Free the intermediate buffers for flow creation. */
1303 mlx5_flow_free_intermediate(dev);
1304 /* Prevent crashes when queues are still in use. */
1305 dev->rx_pkt_burst = removed_rx_burst;
1306 dev->tx_pkt_burst = removed_tx_burst;
1308 /* Disable datapath on secondary process. */
1309 mlx5_mp_req_stop_rxtx(dev);
1310 /* Free the eCPRI flex parser resource. */
1311 mlx5_flex_parser_ecpri_release(dev);
1312 if (priv->rxqs != NULL) {
1313 /* XXX race condition if mlx5_rx_burst() is still running. */
1315 for (i = 0; (i != priv->rxqs_n); ++i)
1316 mlx5_rxq_release(dev, i);
1320 if (priv->txqs != NULL) {
1321 /* XXX race condition if mlx5_tx_burst() is still running. */
1323 for (i = 0; (i != priv->txqs_n); ++i)
1324 mlx5_txq_release(dev, i);
1328 mlx5_proc_priv_uninit(dev);
1329 if (priv->mreg_cp_tbl)
1330 mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1331 mlx5_mprq_free_mp(dev);
1332 mlx5_os_free_shared_dr(priv);
1333 if (priv->rss_conf.rss_key != NULL)
1334 mlx5_free(priv->rss_conf.rss_key);
1335 if (priv->reta_idx != NULL)
1336 mlx5_free(priv->reta_idx);
1337 if (priv->config.vf)
1338 mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
1339 dev->data->mac_addrs,
1340 MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
1341 if (priv->nl_socket_route >= 0)
1342 close(priv->nl_socket_route);
1343 if (priv->nl_socket_rdma >= 0)
1344 close(priv->nl_socket_rdma);
1345 if (priv->vmwa_context)
1346 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1347 ret = mlx5_hrxq_verify(dev);
1349 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1350 dev->data->port_id);
1351 ret = mlx5_ind_table_obj_verify(dev);
1353 DRV_LOG(WARNING, "port %u some indirection table still remain",
1354 dev->data->port_id);
1355 ret = mlx5_rxq_obj_verify(dev);
1357 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1358 dev->data->port_id);
1359 ret = mlx5_rxq_verify(dev);
1361 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1362 dev->data->port_id);
1363 ret = mlx5_txq_obj_verify(dev);
1365 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1366 dev->data->port_id);
1367 ret = mlx5_txq_verify(dev);
1369 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1370 dev->data->port_id);
1371 ret = mlx5_flow_verify(dev);
1373 DRV_LOG(WARNING, "port %u some flows still remain",
1374 dev->data->port_id);
1376 * Free the shared context in last turn, because the cleanup
1377 * routines above may use some shared fields, like
1378 * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1379 * ifindex if Netlink fails.
1381 mlx5_free_shared_dev_ctx(priv->sh);
1382 if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1386 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1387 struct mlx5_priv *opriv =
1388 rte_eth_devices[port_id].data->dev_private;
1391 opriv->domain_id != priv->domain_id ||
1392 &rte_eth_devices[port_id] == dev)
1398 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1400 memset(priv, 0, sizeof(*priv));
1401 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1403 * Reset mac_addrs to NULL such that it is not freed as part of
1404 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1405 * it is freed when dev_private is freed.
1407 dev->data->mac_addrs = NULL;
1411 * Verify and store value for device argument.
1414 * Key argument to verify.
1416 * Value associated with key.
1421 * 0 on success, a negative errno value otherwise and rte_errno is set.
1424 mlx5_args_check(const char *key, const char *val, void *opaque)
1426 struct mlx5_dev_config *config = opaque;
1430 /* No-op, port representors are processed in mlx5_dev_spawn(). */
1431 if (!strcmp(MLX5_REPRESENTOR, key))
1434 tmp = strtol(val, NULL, 0);
1437 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1440 if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1441 /* Negative values are acceptable for some keys only. */
1443 DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1446 mod = tmp >= 0 ? tmp : -tmp;
1447 if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1448 config->cqe_comp = !!tmp;
1449 } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1450 config->cqe_pad = !!tmp;
1451 } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1452 config->hw_padding = !!tmp;
1453 } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1454 config->mprq.enabled = !!tmp;
1455 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1456 config->mprq.stride_num_n = tmp;
1457 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1458 config->mprq.stride_size_n = tmp;
1459 } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1460 config->mprq.max_memcpy_len = tmp;
1461 } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1462 config->mprq.min_rxqs_num = tmp;
1463 } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1464 DRV_LOG(WARNING, "%s: deprecated parameter,"
1465 " converted to txq_inline_max", key);
1466 config->txq_inline_max = tmp;
1467 } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1468 config->txq_inline_max = tmp;
1469 } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1470 config->txq_inline_min = tmp;
1471 } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1472 config->txq_inline_mpw = tmp;
1473 } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1474 config->txqs_inline = tmp;
1475 } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1476 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1477 } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1478 config->mps = !!tmp;
1479 } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1480 if (tmp != MLX5_TXDB_CACHED &&
1481 tmp != MLX5_TXDB_NCACHED &&
1482 tmp != MLX5_TXDB_HEURISTIC) {
1483 DRV_LOG(ERR, "invalid Tx doorbell "
1484 "mapping parameter");
1489 } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1490 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1491 } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1492 DRV_LOG(WARNING, "%s: deprecated parameter,"
1493 " converted to txq_inline_mpw", key);
1494 config->txq_inline_mpw = tmp;
1495 } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1496 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1497 } else if (strcmp(MLX5_TX_PP, key) == 0) {
1499 DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1503 config->tx_pp = tmp;
1504 } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1505 config->tx_skew = tmp;
1506 } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1507 config->rx_vec_en = !!tmp;
1508 } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1509 config->l3_vxlan_en = !!tmp;
1510 } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1511 config->vf_nl_en = !!tmp;
1512 } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1513 config->dv_esw_en = !!tmp;
1514 } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1515 config->dv_flow_en = !!tmp;
1516 } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1517 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1518 tmp != MLX5_XMETA_MODE_META16 &&
1519 tmp != MLX5_XMETA_MODE_META32) {
1520 DRV_LOG(ERR, "invalid extensive "
1521 "metadata parameter");
1525 config->dv_xmeta_en = tmp;
1526 } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1527 config->lacp_by_user = !!tmp;
1528 } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1529 config->mr_ext_memseg_en = !!tmp;
1530 } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1531 config->max_dump_files_num = tmp;
1532 } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1533 config->lro.timeout = tmp;
1534 } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1535 DRV_LOG(DEBUG, "class argument is %s.", val);
1536 } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1537 config->log_hp_size = tmp;
1538 } else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1539 if (tmp != MLX5_RCM_NONE &&
1540 tmp != MLX5_RCM_LIGHT &&
1541 tmp != MLX5_RCM_AGGR) {
1542 DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1546 config->reclaim_mode = tmp;
1547 } else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1548 config->sys_mem_en = !!tmp;
1549 } else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1550 config->decap_en = !!tmp;
1552 DRV_LOG(WARNING, "%s: unknown parameter", key);
1560 * Parse device parameters.
1563 * Pointer to device configuration structure.
1565 * Device arguments structure.
1568 * 0 on success, a negative errno value otherwise and rte_errno is set.
1571 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1573 const char **params = (const char *[]){
1574 MLX5_RXQ_CQE_COMP_EN,
1575 MLX5_RXQ_CQE_PAD_EN,
1576 MLX5_RXQ_PKT_PAD_EN,
1578 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1579 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1580 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1583 MLX5_TXQ_INLINE_MIN,
1584 MLX5_TXQ_INLINE_MAX,
1585 MLX5_TXQ_INLINE_MPW,
1586 MLX5_TXQS_MIN_INLINE,
1589 MLX5_TXQ_MPW_HDR_DSEG_EN,
1590 MLX5_TXQ_MAX_INLINE_LEN,
1602 MLX5_MR_EXT_MEMSEG_EN,
1604 MLX5_MAX_DUMP_FILES_NUM,
1605 MLX5_LRO_TIMEOUT_USEC,
1606 MLX5_CLASS_ARG_NAME,
1613 struct rte_kvargs *kvlist;
1617 if (devargs == NULL)
1619 /* Following UGLY cast is done to pass checkpatch. */
1620 kvlist = rte_kvargs_parse(devargs->args, params);
1621 if (kvlist == NULL) {
1625 /* Process parameters. */
1626 for (i = 0; (params[i] != NULL); ++i) {
1627 if (rte_kvargs_count(kvlist, params[i])) {
1628 ret = rte_kvargs_process(kvlist, params[i],
1629 mlx5_args_check, config);
1632 rte_kvargs_free(kvlist);
1637 rte_kvargs_free(kvlist);
1642 * PMD global initialization.
1644 * Independent from individual device, this function initializes global
1645 * per-PMD data structures distinguishing primary and secondary processes.
1646 * Hence, each initialization is called once per a process.
1649 * 0 on success, a negative errno value otherwise and rte_errno is set.
1652 mlx5_init_once(void)
1654 struct mlx5_shared_data *sd;
1655 struct mlx5_local_data *ld = &mlx5_local_data;
1658 if (mlx5_init_shared_data())
1660 sd = mlx5_shared_data;
1662 rte_spinlock_lock(&sd->lock);
1663 switch (rte_eal_process_type()) {
1664 case RTE_PROC_PRIMARY:
1667 LIST_INIT(&sd->mem_event_cb_list);
1668 rte_rwlock_init(&sd->mem_event_rwlock);
1669 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1670 mlx5_mr_mem_event_cb, NULL);
1671 ret = mlx5_mp_init_primary(MLX5_MP_NAME,
1672 mlx5_mp_primary_handle);
1675 sd->init_done = true;
1677 case RTE_PROC_SECONDARY:
1680 ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
1681 mlx5_mp_secondary_handle);
1684 ++sd->secondary_cnt;
1685 ld->init_done = true;
1691 rte_spinlock_unlock(&sd->lock);
1696 * Configures the minimal amount of data to inline into WQE
1697 * while sending packets.
1699 * - the txq_inline_min has the maximal priority, if this
1700 * key is specified in devargs
1701 * - if DevX is enabled the inline mode is queried from the
1702 * device (HCA attributes and NIC vport context if needed).
1703 * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1704 * and none (0 bytes) for other NICs
1707 * Verbs device parameters (name, port, switch_info) to spawn.
1709 * Device configuration parameters.
1712 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1713 struct mlx5_dev_config *config)
1715 if (config->txq_inline_min != MLX5_ARG_UNSET) {
1716 /* Application defines size of inlined data explicitly. */
1717 switch (spawn->pci_dev->id.device_id) {
1718 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1719 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1720 if (config->txq_inline_min <
1721 (int)MLX5_INLINE_HSIZE_L2) {
1723 "txq_inline_mix aligned to minimal"
1724 " ConnectX-4 required value %d",
1725 (int)MLX5_INLINE_HSIZE_L2);
1726 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1732 if (config->hca_attr.eth_net_offloads) {
1733 /* We have DevX enabled, inline mode queried successfully. */
1734 switch (config->hca_attr.wqe_inline_mode) {
1735 case MLX5_CAP_INLINE_MODE_L2:
1736 /* outer L2 header must be inlined. */
1737 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1739 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1740 /* No inline data are required by NIC. */
1741 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1742 config->hw_vlan_insert =
1743 config->hca_attr.wqe_vlan_insert;
1744 DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1746 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1747 /* inline mode is defined by NIC vport context. */
1748 if (!config->hca_attr.eth_virt)
1750 switch (config->hca_attr.vport_inline_mode) {
1751 case MLX5_INLINE_MODE_NONE:
1752 config->txq_inline_min =
1753 MLX5_INLINE_HSIZE_NONE;
1755 case MLX5_INLINE_MODE_L2:
1756 config->txq_inline_min =
1757 MLX5_INLINE_HSIZE_L2;
1759 case MLX5_INLINE_MODE_IP:
1760 config->txq_inline_min =
1761 MLX5_INLINE_HSIZE_L3;
1763 case MLX5_INLINE_MODE_TCP_UDP:
1764 config->txq_inline_min =
1765 MLX5_INLINE_HSIZE_L4;
1767 case MLX5_INLINE_MODE_INNER_L2:
1768 config->txq_inline_min =
1769 MLX5_INLINE_HSIZE_INNER_L2;
1771 case MLX5_INLINE_MODE_INNER_IP:
1772 config->txq_inline_min =
1773 MLX5_INLINE_HSIZE_INNER_L3;
1775 case MLX5_INLINE_MODE_INNER_TCP_UDP:
1776 config->txq_inline_min =
1777 MLX5_INLINE_HSIZE_INNER_L4;
1783 * We get here if we are unable to deduce
1784 * inline data size with DevX. Try PCI ID
1785 * to determine old NICs.
1787 switch (spawn->pci_dev->id.device_id) {
1788 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1789 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1790 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1791 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1792 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1793 config->hw_vlan_insert = 0;
1795 case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1796 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1797 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1798 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1800 * These NICs support VLAN insertion from WQE and
1801 * report the wqe_vlan_insert flag. But there is the bug
1802 * and PFC control may be broken, so disable feature.
1804 config->hw_vlan_insert = 0;
1805 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1808 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1812 DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1816 * Configures the metadata mask fields in the shared context.
1819 * Pointer to Ethernet device.
1822 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1824 struct mlx5_priv *priv = dev->data->dev_private;
1825 struct mlx5_dev_ctx_shared *sh = priv->sh;
1826 uint32_t meta, mark, reg_c0;
1828 reg_c0 = ~priv->vport_meta_mask;
1829 switch (priv->config.dv_xmeta_en) {
1830 case MLX5_XMETA_MODE_LEGACY:
1832 mark = MLX5_FLOW_MARK_MASK;
1834 case MLX5_XMETA_MODE_META16:
1835 meta = reg_c0 >> rte_bsf32(reg_c0);
1836 mark = MLX5_FLOW_MARK_MASK;
1838 case MLX5_XMETA_MODE_META32:
1840 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1848 if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1849 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1850 sh->dv_mark_mask, mark);
1852 sh->dv_mark_mask = mark;
1853 if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1854 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1855 sh->dv_meta_mask, meta);
1857 sh->dv_meta_mask = meta;
1858 if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1859 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1860 sh->dv_meta_mask, reg_c0);
1862 sh->dv_regc0_mask = reg_c0;
1863 DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1864 DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1865 DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1866 DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1870 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
1872 static const char *const dynf_names[] = {
1873 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
1874 RTE_MBUF_DYNFLAG_METADATA_NAME,
1875 RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
1879 if (n < RTE_DIM(dynf_names))
1881 for (i = 0; i < RTE_DIM(dynf_names); i++) {
1882 if (names[i] == NULL)
1884 strcpy(names[i], dynf_names[i]);
1886 return RTE_DIM(dynf_names);
1890 * Comparison callback to sort device data.
1892 * This is meant to be used with qsort().
1895 * Pointer to pointer to first data object.
1897 * Pointer to pointer to second data object.
1900 * 0 if both objects are equal, less than 0 if the first argument is less
1901 * than the second, greater than 0 otherwise.
1904 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1905 struct mlx5_dev_config *config)
1907 struct mlx5_dev_ctx_shared *sh = priv->sh;
1908 struct mlx5_dev_config *sh_conf = NULL;
1912 /* Nothing to compare for the single/first device. */
1913 if (sh->refcnt == 1)
1915 /* Find the device with shared context. */
1916 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1917 struct mlx5_priv *opriv =
1918 rte_eth_devices[port_id].data->dev_private;
1920 if (opriv && opriv != priv && opriv->sh == sh) {
1921 sh_conf = &opriv->config;
1927 if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
1928 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
1929 " for shared %s context", sh->ibdev_name);
1933 if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
1934 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
1935 " for shared %s context", sh->ibdev_name);
1943 * Look for the ethernet device belonging to mlx5 driver.
1945 * @param[in] port_id
1946 * port_id to start looking for device.
1947 * @param[in] pci_dev
1948 * Pointer to the hint PCI device. When device is being probed
1949 * the its siblings (master and preceding representors might
1950 * not have assigned driver yet (because the mlx5_os_pci_probe()
1951 * is not completed yet, for this case match on hint PCI
1952 * device may be used to detect sibling device.
1955 * port_id of found device, RTE_MAX_ETHPORT if not found.
1958 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
1960 while (port_id < RTE_MAX_ETHPORTS) {
1961 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
1963 if (dev->state != RTE_ETH_DEV_UNUSED &&
1965 (dev->device == &pci_dev->device ||
1966 (dev->device->driver &&
1967 dev->device->driver->name &&
1968 !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
1972 if (port_id >= RTE_MAX_ETHPORTS)
1973 return RTE_MAX_ETHPORTS;
1978 * DPDK callback to remove a PCI device.
1980 * This function removes all Ethernet devices belong to a given PCI device.
1982 * @param[in] pci_dev
1983 * Pointer to the PCI device.
1986 * 0 on success, the function cannot fail.
1989 mlx5_pci_remove(struct rte_pci_device *pci_dev)
1993 RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
1995 * mlx5_dev_close() is not registered to secondary process,
1996 * call the close function explicitly for secondary process.
1998 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1999 mlx5_dev_close(&rte_eth_devices[port_id]);
2001 rte_eth_dev_close(port_id);
2006 static const struct rte_pci_id mlx5_pci_id_map[] = {
2008 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2009 PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2012 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2013 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2016 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2017 PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2020 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2021 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2024 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2025 PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2028 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2029 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2032 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2033 PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2036 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2037 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2040 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2041 PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2044 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2045 PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2048 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2049 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2052 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2053 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2056 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2057 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2060 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2061 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
2064 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2065 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2068 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2069 PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2076 struct rte_pci_driver mlx5_driver = {
2078 .name = MLX5_DRIVER_NAME
2080 .id_table = mlx5_pci_id_map,
2081 .probe = mlx5_os_pci_probe,
2082 .remove = mlx5_pci_remove,
2083 .dma_map = mlx5_dma_map,
2084 .dma_unmap = mlx5_dma_unmap,
2085 .drv_flags = PCI_DRV_FLAGS,
2088 /* Initialize driver log type. */
2089 RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
2092 * Driver initialization routine.
2094 RTE_INIT(rte_mlx5_pmd_init)
2096 /* Build the static tables for Verbs conversion. */
2097 mlx5_set_ptype_table();
2098 mlx5_set_cksum_table();
2099 mlx5_set_swp_types_table();
2101 rte_pci_register(&mlx5_driver);
2104 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2105 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2106 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");