1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
13 #include <rte_malloc.h>
14 #include <rte_ethdev_driver.h>
15 #include <rte_ethdev_pci.h>
17 #include <rte_bus_pci.h>
18 #include <rte_common.h>
19 #include <rte_kvargs.h>
20 #include <rte_rwlock.h>
21 #include <rte_spinlock.h>
22 #include <rte_string_fns.h>
23 #include <rte_alarm.h>
25 #include <mlx5_glue.h>
26 #include <mlx5_devx_cmds.h>
27 #include <mlx5_common.h>
28 #include <mlx5_common_os.h>
29 #include <mlx5_common_mp.h>
30 #include <mlx5_common_pci.h>
31 #include <mlx5_malloc.h>
33 #include "mlx5_defs.h"
35 #include "mlx5_utils.h"
36 #include "mlx5_rxtx.h"
37 #include "mlx5_autoconf.h"
39 #include "mlx5_flow.h"
40 #include "rte_pmd_mlx5.h"
42 /* Device parameter to enable RX completion queue compression. */
43 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
45 /* Device parameter to enable RX completion entry padding to 128B. */
46 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
48 /* Device parameter to enable padding Rx packet to cacheline size. */
49 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
51 /* Device parameter to enable Multi-Packet Rx queue. */
52 #define MLX5_RX_MPRQ_EN "mprq_en"
54 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
55 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
57 /* Device parameter to configure log 2 of the stride size for MPRQ. */
58 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
60 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
61 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
63 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
64 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
66 /* Device parameter to configure inline send. Deprecated, ignored.*/
67 #define MLX5_TXQ_INLINE "txq_inline"
69 /* Device parameter to limit packet size to inline with ordinary SEND. */
70 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
72 /* Device parameter to configure minimal data size to inline. */
73 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
75 /* Device parameter to limit packet size to inline with Enhanced MPW. */
76 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
79 * Device parameter to configure the number of TX queues threshold for
80 * enabling inline send.
82 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
85 * Device parameter to configure the number of TX queues threshold for
86 * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
88 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
90 /* Device parameter to enable multi-packet send WQEs. */
91 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
94 * Device parameter to force doorbell register mapping
95 * to non-cahed region eliminating the extra write memory barrier.
97 #define MLX5_TX_DB_NC "tx_db_nc"
100 * Device parameter to include 2 dsegs in the title WQEBB.
101 * Deprecated, ignored.
103 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
106 * Device parameter to limit the size of inlining packet.
107 * Deprecated, ignored.
109 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
112 * Device parameter to enable Tx scheduling on timestamps
113 * and specify the packet pacing granularity in nanoseconds.
115 #define MLX5_TX_PP "tx_pp"
118 * Device parameter to specify skew in nanoseconds on Tx datapath,
119 * it represents the time between SQ start WQE processing and
120 * appearing actual packet data on the wire.
122 #define MLX5_TX_SKEW "tx_skew"
125 * Device parameter to enable hardware Tx vector.
126 * Deprecated, ignored (no vectorized Tx routines anymore).
128 #define MLX5_TX_VEC_EN "tx_vec_en"
130 /* Device parameter to enable hardware Rx vector. */
131 #define MLX5_RX_VEC_EN "rx_vec_en"
133 /* Allow L3 VXLAN flow creation. */
134 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
136 /* Activate DV E-Switch flow steering. */
137 #define MLX5_DV_ESW_EN "dv_esw_en"
139 /* Activate DV flow steering. */
140 #define MLX5_DV_FLOW_EN "dv_flow_en"
142 /* Enable extensive flow metadata support. */
143 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
145 /* Device parameter to let the user manage the lacp traffic of bonded device */
146 #define MLX5_LACP_BY_USER "lacp_by_user"
148 /* Activate Netlink support in VF mode. */
149 #define MLX5_VF_NL_EN "vf_nl_en"
151 /* Enable extending memsegs when creating a MR. */
152 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
154 /* Select port representors to instantiate. */
155 #define MLX5_REPRESENTOR "representor"
157 /* Device parameter to configure the maximum number of dump files per queue. */
158 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
160 /* Configure timeout of LRO session (in microseconds). */
161 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
164 * Device parameter to configure the total data buffer size for a single
165 * hairpin queue (logarithm value).
167 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
169 /* Flow memory reclaim mode. */
170 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
172 /* The default memory allocator used in PMD. */
173 #define MLX5_SYS_MEM_EN "sys_mem_en"
174 /* Decap will be used or not. */
175 #define MLX5_DECAP_EN "decap_en"
177 /* Shared memory between primary and secondary processes. */
178 struct mlx5_shared_data *mlx5_shared_data;
180 /** Driver-specific log messages type. */
183 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
184 LIST_HEAD_INITIALIZER();
185 static pthread_mutex_t mlx5_dev_ctx_list_mutex = PTHREAD_MUTEX_INITIALIZER;
187 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
188 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
190 .size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
196 .malloc = mlx5_malloc,
198 .type = "mlx5_encap_decap_ipool",
201 .size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
207 .malloc = mlx5_malloc,
209 .type = "mlx5_push_vlan_ipool",
212 .size = sizeof(struct mlx5_flow_dv_tag_resource),
218 .malloc = mlx5_malloc,
220 .type = "mlx5_tag_ipool",
223 .size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
229 .malloc = mlx5_malloc,
231 .type = "mlx5_port_id_ipool",
234 .size = sizeof(struct mlx5_flow_tbl_data_entry),
240 .malloc = mlx5_malloc,
242 .type = "mlx5_jump_ipool",
245 .size = sizeof(struct mlx5_flow_dv_sample_resource),
251 .malloc = mlx5_malloc,
253 .type = "mlx5_sample_ipool",
256 .size = sizeof(struct mlx5_flow_dv_dest_array_resource),
262 .malloc = mlx5_malloc,
264 .type = "mlx5_dest_array_ipool",
268 .size = sizeof(struct mlx5_flow_meter),
274 .malloc = mlx5_malloc,
276 .type = "mlx5_meter_ipool",
279 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
285 .malloc = mlx5_malloc,
287 .type = "mlx5_mcp_ipool",
290 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
296 .malloc = mlx5_malloc,
298 .type = "mlx5_hrxq_ipool",
302 * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
303 * It set in run time according to PCI function configuration.
311 .malloc = mlx5_malloc,
313 .type = "mlx5_flow_handle_ipool",
316 .size = sizeof(struct rte_flow),
320 .malloc = mlx5_malloc,
322 .type = "rte_flow_ipool",
327 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
328 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
330 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
333 * Allocate ID pool structure.
336 * The maximum id can be allocated from the pool.
339 * Pointer to pool object, NULL value otherwise.
341 struct mlx5_flow_id_pool *
342 mlx5_flow_id_pool_alloc(uint32_t max_id)
344 struct mlx5_flow_id_pool *pool;
347 pool = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*pool),
348 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
350 DRV_LOG(ERR, "can't allocate id pool");
354 mem = mlx5_malloc(MLX5_MEM_ZERO,
355 MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
356 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
358 DRV_LOG(ERR, "can't allocate mem for id pool");
362 pool->free_arr = mem;
363 pool->curr = pool->free_arr;
364 pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
365 pool->base_index = 0;
366 pool->max_id = max_id;
374 * Release ID pool structure.
377 * Pointer to flow id pool object to free.
380 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
382 mlx5_free(pool->free_arr);
390 * Pointer to flow id pool.
395 * 0 on success, error value otherwise.
398 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
400 if (pool->curr == pool->free_arr) {
401 if (pool->base_index == pool->max_id) {
403 DRV_LOG(ERR, "no free id");
406 *id = ++pool->base_index;
409 *id = *(--pool->curr);
417 * Pointer to flow id pool.
422 * 0 on success, error value otherwise.
425 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
431 if (pool->curr == pool->last) {
432 size = pool->curr - pool->free_arr;
433 size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
434 MLX5_ASSERT(size2 > size);
435 mem = mlx5_malloc(0, size2 * sizeof(uint32_t), 0,
438 DRV_LOG(ERR, "can't allocate mem for id pool");
442 memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
443 mlx5_free(pool->free_arr);
444 pool->free_arr = mem;
445 pool->curr = pool->free_arr + size;
446 pool->last = pool->free_arr + size2;
454 * Initialize the shared aging list information per port.
457 * Pointer to mlx5_dev_ctx_shared object.
460 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
463 struct mlx5_age_info *age_info;
465 for (i = 0; i < sh->max_port; i++) {
466 age_info = &sh->port[i].age_info;
468 TAILQ_INIT(&age_info->aged_counters);
469 rte_spinlock_init(&age_info->aged_sl);
470 MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
475 * Initialize the counters management structure.
478 * Pointer to mlx5_dev_ctx_shared object to free
481 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
485 memset(&sh->cmng, 0, sizeof(sh->cmng));
486 TAILQ_INIT(&sh->cmng.flow_counters);
487 sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
488 sh->cmng.max_id = -1;
489 sh->cmng.last_pool_idx = POOL_IDX_INVALID;
490 rte_spinlock_init(&sh->cmng.pool_update_sl);
491 for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
492 TAILQ_INIT(&sh->cmng.counters[i]);
493 rte_spinlock_init(&sh->cmng.csl[i]);
498 * Destroy all the resources allocated for a counter memory management.
501 * Pointer to the memory management structure.
504 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
506 uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
508 LIST_REMOVE(mng, next);
509 claim_zero(mlx5_devx_cmd_destroy(mng->dm));
510 claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
515 * Close and release all the resources of the counters management.
518 * Pointer to mlx5_dev_ctx_shared object to free.
521 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
523 struct mlx5_counter_stats_mem_mng *mng;
529 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
530 if (rte_errno != EINPROGRESS)
535 if (sh->cmng.pools) {
536 struct mlx5_flow_counter_pool *pool;
537 uint16_t n_valid = sh->cmng.n_valid;
538 bool fallback = sh->cmng.counter_fallback;
540 for (i = 0; i < n_valid; ++i) {
541 pool = sh->cmng.pools[i];
542 if (!fallback && pool->min_dcs)
543 claim_zero(mlx5_devx_cmd_destroy
545 for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
546 struct mlx5_flow_counter *cnt =
547 MLX5_POOL_GET_CNT(pool, j);
551 (mlx5_glue->destroy_flow_action
553 if (fallback && MLX5_POOL_GET_CNT
554 (pool, j)->dcs_when_free)
555 claim_zero(mlx5_devx_cmd_destroy
556 (cnt->dcs_when_free));
560 mlx5_free(sh->cmng.pools);
562 mng = LIST_FIRST(&sh->cmng.mem_mngs);
564 mlx5_flow_destroy_counter_stat_mem_mng(mng);
565 mng = LIST_FIRST(&sh->cmng.mem_mngs);
567 memset(&sh->cmng, 0, sizeof(sh->cmng));
571 * Initialize the flow resources' indexed mempool.
574 * Pointer to mlx5_dev_ctx_shared object.
576 * Pointer to user dev config.
579 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
580 const struct mlx5_dev_config *config)
583 struct mlx5_indexed_pool_config cfg;
585 for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
586 cfg = mlx5_ipool_cfg[i];
591 * Set MLX5_IPOOL_MLX5_FLOW ipool size
592 * according to PCI function flow configuration.
594 case MLX5_IPOOL_MLX5_FLOW:
595 cfg.size = config->dv_flow_en ?
596 sizeof(struct mlx5_flow_handle) :
597 MLX5_FLOW_HANDLE_VERBS_SIZE;
600 if (config->reclaim_mode)
601 cfg.release_mem_en = 1;
602 sh->ipool[i] = mlx5_ipool_create(&cfg);
607 * Release the flow resources' indexed mempool.
610 * Pointer to mlx5_dev_ctx_shared object.
613 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
617 for (i = 0; i < MLX5_IPOOL_MAX; ++i)
618 mlx5_ipool_destroy(sh->ipool[i]);
622 * Check if dynamic flex parser for eCPRI already exists.
625 * Pointer to Ethernet device structure.
628 * true on exists, false on not.
631 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
633 struct mlx5_priv *priv = dev->data->dev_private;
634 struct mlx5_flex_parser_profiles *prf =
635 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
641 * Allocation of a flex parser for eCPRI. Once created, this parser related
642 * resources will be held until the device is closed.
645 * Pointer to Ethernet device structure.
648 * 0 on success, a negative errno value otherwise and rte_errno is set.
651 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
653 struct mlx5_priv *priv = dev->data->dev_private;
654 struct mlx5_flex_parser_profiles *prf =
655 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
656 struct mlx5_devx_graph_node_attr node = {
657 .modify_field_select = 0,
662 if (!priv->config.hca_attr.parse_graph_flex_node) {
663 DRV_LOG(ERR, "Dynamic flex parser is not supported "
664 "for device %s.", priv->dev_data->name);
667 node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
668 /* 8 bytes now: 4B common header + 4B message body header. */
669 node.header_length_base_value = 0x8;
670 /* After MAC layer: Ether / VLAN. */
671 node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
672 /* Type of compared condition should be 0xAEFE in the L2 layer. */
673 node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
674 /* Sample #0: type in common header. */
675 node.sample[0].flow_match_sample_en = 1;
677 node.sample[0].flow_match_sample_offset_mode = 0x0;
678 /* Only the 2nd byte will be used. */
679 node.sample[0].flow_match_sample_field_base_offset = 0x0;
680 /* Sample #1: message payload. */
681 node.sample[1].flow_match_sample_en = 1;
683 node.sample[1].flow_match_sample_offset_mode = 0x0;
685 * Only the first two bytes will be used right now, and its offset will
686 * start after the common header that with the length of a DW(u32).
688 node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
689 prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
691 DRV_LOG(ERR, "Failed to create flex parser node object.");
692 return (rte_errno == 0) ? -ENODEV : -rte_errno;
695 ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
697 DRV_LOG(ERR, "Failed to query sample IDs.");
698 return (rte_errno == 0) ? -ENODEV : -rte_errno;
700 prf->offset[0] = 0x0;
701 prf->offset[1] = sizeof(uint32_t);
702 prf->ids[0] = ids[0];
703 prf->ids[1] = ids[1];
708 * Destroy the flex parser node, including the parser itself, input / output
709 * arcs and DW samples. Resources could be reused then.
712 * Pointer to Ethernet device structure.
715 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
717 struct mlx5_priv *priv = dev->data->dev_private;
718 struct mlx5_flex_parser_profiles *prf =
719 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
722 mlx5_devx_cmd_destroy(prf->obj);
727 * Allocate Rx and Tx UARs in robust fashion.
728 * This routine handles the following UAR allocation issues:
730 * - tries to allocate the UAR with the most appropriate memory
731 * mapping type from the ones supported by the host
733 * - tries to allocate the UAR with non-NULL base address
734 * OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
735 * UAR base address if UAR was not the first object in the UAR page.
736 * It caused the PMD failure and we should try to get another UAR
737 * till we get the first one with non-NULL base address returned.
740 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
741 const struct mlx5_dev_config *config)
743 uint32_t uar_mapping, retry;
747 for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
748 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
749 /* Control the mapping type according to the settings. */
750 uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
751 MLX5DV_UAR_ALLOC_TYPE_NC :
752 MLX5DV_UAR_ALLOC_TYPE_BF;
754 RTE_SET_USED(config);
756 * It seems we have no way to control the memory mapping type
757 * for the UAR, the default "Write-Combining" type is supposed.
758 * The UAR initialization on queue creation queries the
759 * actual mapping type done by Verbs/kernel and setups the
760 * PMD datapath accordingly.
764 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, uar_mapping);
765 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
767 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
768 if (config->dbnc == MLX5_TXDB_CACHED ||
769 config->dbnc == MLX5_TXDB_HEURISTIC)
770 DRV_LOG(WARNING, "Devarg tx_db_nc setting "
771 "is not supported by DevX");
773 * In some environments like virtual machine
774 * the Write Combining mapped might be not supported
775 * and UAR allocation fails. We try "Non-Cached"
776 * mapping for the case. The tx_burst routines take
777 * the UAR mapping type into account on UAR setup
780 DRV_LOG(WARNING, "Failed to allocate Tx DevX UAR (BF)");
781 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
782 sh->tx_uar = mlx5_glue->devx_alloc_uar
783 (sh->ctx, uar_mapping);
784 } else if (!sh->tx_uar &&
785 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
786 if (config->dbnc == MLX5_TXDB_NCACHED)
787 DRV_LOG(WARNING, "Devarg tx_db_nc settings "
788 "is not supported by DevX");
790 * If Verbs/kernel does not support "Non-Cached"
791 * try the "Write-Combining".
793 DRV_LOG(WARNING, "Failed to allocate Tx DevX UAR (NC)");
794 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
795 sh->tx_uar = mlx5_glue->devx_alloc_uar
796 (sh->ctx, uar_mapping);
800 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
804 base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
808 * The UARs are allocated by rdma_core within the
809 * IB device context, on context closure all UARs
810 * will be freed, should be no memory/object leakage.
812 DRV_LOG(WARNING, "Retrying to allocate Tx DevX UAR");
815 /* Check whether we finally succeeded with valid UAR allocation. */
817 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
821 for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
823 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
824 (sh->ctx, uar_mapping);
825 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
826 if (!sh->devx_rx_uar &&
827 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
829 * Rx UAR is used to control interrupts only,
830 * should be no datapath noticeable impact,
831 * can try "Non-Cached" mapping safely.
833 DRV_LOG(WARNING, "Failed to allocate Rx DevX UAR (BF)");
834 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
835 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
836 (sh->ctx, uar_mapping);
839 if (!sh->devx_rx_uar) {
840 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
844 base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
848 * The UARs are allocated by rdma_core within the
849 * IB device context, on context closure all UARs
850 * will be freed, should be no memory/object leakage.
852 DRV_LOG(WARNING, "Retrying to allocate Rx DevX UAR");
853 sh->devx_rx_uar = NULL;
855 /* Check whether we finally succeeded with valid UAR allocation. */
856 if (!sh->devx_rx_uar) {
857 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
865 * Allocate shared device context. If there is multiport device the
866 * master and representors will share this context, if there is single
867 * port dedicated device, the context will be used by only given
868 * port due to unification.
870 * Routine first searches the context for the specified device name,
871 * if found the shared context assumed and reference counter is incremented.
872 * If no context found the new one is created and initialized with specified
873 * device context and parameters.
876 * Pointer to the device attributes (name, port, etc).
878 * Pointer to device configuration structure.
881 * Pointer to mlx5_dev_ctx_shared object on success,
882 * otherwise NULL and rte_errno is set.
884 struct mlx5_dev_ctx_shared *
885 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
886 const struct mlx5_dev_config *config)
888 struct mlx5_dev_ctx_shared *sh;
891 struct mlx5_devx_tis_attr tis_attr = { 0 };
894 /* Secondary process should not create the shared context. */
895 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
896 pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
897 /* Search for IB context by device name. */
898 LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
899 if (!strcmp(sh->ibdev_name,
900 mlx5_os_get_dev_device_name(spawn->phys_dev))) {
905 /* No device found, we have to create new shared context. */
906 MLX5_ASSERT(spawn->max_port);
907 sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
908 sizeof(struct mlx5_dev_ctx_shared) +
910 sizeof(struct mlx5_dev_shared_port),
911 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
913 DRV_LOG(ERR, "shared context allocation failure");
917 err = mlx5_os_open_device(spawn, config, sh);
920 err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
922 DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
926 sh->max_port = spawn->max_port;
927 strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
928 sizeof(sh->ibdev_name) - 1);
929 strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
930 sizeof(sh->ibdev_path) - 1);
932 * Setting port_id to max unallowed value means
933 * there is no interrupt subhandler installed for
934 * the given port index i.
936 for (i = 0; i < sh->max_port; i++) {
937 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
938 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
940 sh->pd = mlx5_glue->alloc_pd(sh->ctx);
941 if (sh->pd == NULL) {
942 DRV_LOG(ERR, "PD allocation failure");
947 /* Query the EQN for this core. */
948 err = mlx5_glue->devx_query_eqn(sh->ctx, 0, &sh->eqn);
951 DRV_LOG(ERR, "Failed to query event queue number %d.",
955 err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
957 DRV_LOG(ERR, "Fail to extract pdn from PD");
960 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
962 DRV_LOG(ERR, "TD allocation failure");
966 tis_attr.transport_domain = sh->td->id;
967 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
969 DRV_LOG(ERR, "TIS allocation failure");
973 err = mlx5_alloc_rxtx_uars(sh, config);
976 MLX5_ASSERT(sh->tx_uar);
977 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
979 MLX5_ASSERT(sh->devx_rx_uar);
980 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
982 sh->flow_id_pool = mlx5_flow_id_pool_alloc
983 ((1 << HAIRPIN_FLOW_ID_BITS) - 1);
984 if (!sh->flow_id_pool) {
985 DRV_LOG(ERR, "can't create flow id pool");
990 /* Initialize UAR access locks for 32bit implementations. */
991 rte_spinlock_init(&sh->uar_lock_cq);
992 for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
993 rte_spinlock_init(&sh->uar_lock[i]);
996 * Once the device is added to the list of memory event
997 * callback, its global MR cache table cannot be expanded
998 * on the fly because of deadlock. If it overflows, lookup
999 * should be done by searching MR list linearly, which is slow.
1001 * At this point the device is not added to the memory
1002 * event list yet, context is just being created.
1004 err = mlx5_mr_btree_init(&sh->share_cache.cache,
1005 MLX5_MR_BTREE_CACHE_N * 2,
1006 spawn->pci_dev->device.numa_node);
1011 mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
1012 &sh->share_cache.dereg_mr_cb);
1013 mlx5_os_dev_shared_handler_install(sh);
1014 sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
1015 if (!sh->cnt_id_tbl) {
1019 mlx5_flow_aging_init(sh);
1020 mlx5_flow_counters_mng_init(sh);
1021 mlx5_flow_ipool_create(sh, config);
1022 /* Add device to memory callback list. */
1023 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1024 LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1026 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1027 /* Add context to the global device list. */
1028 LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1030 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1033 pthread_mutex_destroy(&sh->txpp.mutex);
1034 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1037 mlx5_l3t_destroy(sh->cnt_id_tbl);
1039 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1041 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1042 if (sh->devx_rx_uar)
1043 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1045 mlx5_glue->devx_free_uar(sh->tx_uar);
1047 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
1049 claim_zero(mlx5_glue->close_device(sh->ctx));
1050 if (sh->flow_id_pool)
1051 mlx5_flow_id_pool_release(sh->flow_id_pool);
1053 MLX5_ASSERT(err > 0);
1059 * Free shared IB device context. Decrement counter and if zero free
1060 * all allocated resources and close handles.
1063 * Pointer to mlx5_dev_ctx_shared object to free
1066 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1068 pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1069 #ifdef RTE_LIBRTE_MLX5_DEBUG
1070 /* Check the object presence in the list. */
1071 struct mlx5_dev_ctx_shared *lctx;
1073 LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1078 DRV_LOG(ERR, "Freeing non-existing shared IB context");
1083 MLX5_ASSERT(sh->refcnt);
1084 /* Secondary process should not free the shared context. */
1085 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1088 /* Remove from memory callback device list. */
1089 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1090 LIST_REMOVE(sh, mem_event_cb);
1091 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1092 /* Release created Memory Regions. */
1093 mlx5_mr_release_cache(&sh->share_cache);
1094 /* Remove context from the global device list. */
1095 LIST_REMOVE(sh, next);
1096 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1098 * Ensure there is no async event handler installed.
1099 * Only primary process handles async device events.
1101 mlx5_flow_counters_mng_close(sh);
1102 mlx5_flow_ipool_destroy(sh);
1103 mlx5_os_dev_shared_handler_uninstall(sh);
1104 if (sh->cnt_id_tbl) {
1105 mlx5_l3t_destroy(sh->cnt_id_tbl);
1106 sh->cnt_id_tbl = NULL;
1109 mlx5_glue->devx_free_uar(sh->tx_uar);
1113 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
1115 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1117 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1118 if (sh->devx_rx_uar)
1119 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1121 claim_zero(mlx5_glue->close_device(sh->ctx));
1122 if (sh->flow_id_pool)
1123 mlx5_flow_id_pool_release(sh->flow_id_pool);
1124 pthread_mutex_destroy(&sh->txpp.mutex);
1128 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1132 * Destroy table hash list and all the root entries per domain.
1135 * Pointer to the private device data structure.
1138 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1140 struct mlx5_dev_ctx_shared *sh = priv->sh;
1141 struct mlx5_flow_tbl_data_entry *tbl_data;
1142 union mlx5_flow_tbl_key table_key = {
1150 struct mlx5_hlist_entry *pos;
1154 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
1156 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
1158 MLX5_ASSERT(tbl_data);
1159 mlx5_hlist_remove(sh->flow_tbls, pos);
1160 mlx5_free(tbl_data);
1162 table_key.direction = 1;
1163 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
1165 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
1167 MLX5_ASSERT(tbl_data);
1168 mlx5_hlist_remove(sh->flow_tbls, pos);
1169 mlx5_free(tbl_data);
1171 table_key.direction = 0;
1172 table_key.domain = 1;
1173 pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
1175 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
1177 MLX5_ASSERT(tbl_data);
1178 mlx5_hlist_remove(sh->flow_tbls, pos);
1179 mlx5_free(tbl_data);
1181 mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
1185 * Initialize flow table hash list and create the root tables entry
1189 * Pointer to the private device data structure.
1192 * Zero on success, positive error code otherwise.
1195 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
1197 struct mlx5_dev_ctx_shared *sh = priv->sh;
1198 char s[MLX5_HLIST_NAMESIZE];
1202 snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1203 sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
1204 if (!sh->flow_tbls) {
1205 DRV_LOG(ERR, "flow tables with hash creation failed.");
1209 #ifndef HAVE_MLX5DV_DR
1211 * In case we have not DR support, the zero tables should be created
1212 * because DV expect to see them even if they cannot be created by
1215 union mlx5_flow_tbl_key table_key = {
1223 struct mlx5_flow_tbl_data_entry *tbl_data = mlx5_malloc(MLX5_MEM_ZERO,
1224 sizeof(*tbl_data), 0,
1231 tbl_data->entry.key = table_key.v64;
1232 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1235 rte_atomic32_init(&tbl_data->tbl.refcnt);
1236 rte_atomic32_inc(&tbl_data->tbl.refcnt);
1237 table_key.direction = 1;
1238 tbl_data = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*tbl_data), 0,
1244 tbl_data->entry.key = table_key.v64;
1245 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1248 rte_atomic32_init(&tbl_data->tbl.refcnt);
1249 rte_atomic32_inc(&tbl_data->tbl.refcnt);
1250 table_key.direction = 0;
1251 table_key.domain = 1;
1252 tbl_data = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*tbl_data), 0,
1258 tbl_data->entry.key = table_key.v64;
1259 err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
1262 rte_atomic32_init(&tbl_data->tbl.refcnt);
1263 rte_atomic32_inc(&tbl_data->tbl.refcnt);
1266 mlx5_free_table_hash_list(priv);
1267 #endif /* HAVE_MLX5DV_DR */
1272 * Retrieve integer value from environment variable.
1275 * Environment variable name.
1278 * Integer value, 0 if the variable is not set.
1281 mlx5_getenv_int(const char *name)
1283 const char *val = getenv(name);
1291 * DPDK callback to add udp tunnel port
1294 * A pointer to eth_dev
1295 * @param[in] udp_tunnel
1296 * A pointer to udp tunnel
1299 * 0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1302 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1303 struct rte_eth_udp_tunnel *udp_tunnel)
1305 MLX5_ASSERT(udp_tunnel != NULL);
1306 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1307 udp_tunnel->udp_port == 4789)
1309 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1310 udp_tunnel->udp_port == 4790)
1316 * Initialize process private data structure.
1319 * Pointer to Ethernet device structure.
1322 * 0 on success, a negative errno value otherwise and rte_errno is set.
1325 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1327 struct mlx5_priv *priv = dev->data->dev_private;
1328 struct mlx5_proc_priv *ppriv;
1332 * UAR register table follows the process private structure. BlueFlame
1333 * registers for Tx queues are stored in the table.
1336 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1337 ppriv = mlx5_malloc(MLX5_MEM_RTE, ppriv_size, RTE_CACHE_LINE_SIZE,
1338 dev->device->numa_node);
1343 ppriv->uar_table_sz = ppriv_size;
1344 dev->process_private = ppriv;
1349 * Un-initialize process private data structure.
1352 * Pointer to Ethernet device structure.
1355 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1357 if (!dev->process_private)
1359 mlx5_free(dev->process_private);
1360 dev->process_private = NULL;
1364 * DPDK callback to close the device.
1366 * Destroy all queues and objects, free memory.
1369 * Pointer to Ethernet device structure.
1372 mlx5_dev_close(struct rte_eth_dev *dev)
1374 struct mlx5_priv *priv = dev->data->dev_private;
1378 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1379 /* Check if process_private released. */
1380 if (!dev->process_private)
1382 mlx5_tx_uar_uninit_secondary(dev);
1383 mlx5_proc_priv_uninit(dev);
1384 rte_eth_dev_release_port(dev);
1389 DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1391 ((priv->sh->ctx != NULL) ?
1392 mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1394 * If default mreg copy action is removed at the stop stage,
1395 * the search will return none and nothing will be done anymore.
1397 mlx5_flow_stop_default(dev);
1398 mlx5_traffic_disable(dev);
1400 * If all the flows are already flushed in the device stop stage,
1401 * then this will return directly without any action.
1403 mlx5_flow_list_flush(dev, &priv->flows, true);
1404 mlx5_flow_meter_flush(dev, NULL);
1405 /* Free the intermediate buffers for flow creation. */
1406 mlx5_flow_free_intermediate(dev);
1407 /* Prevent crashes when queues are still in use. */
1408 dev->rx_pkt_burst = removed_rx_burst;
1409 dev->tx_pkt_burst = removed_tx_burst;
1411 /* Disable datapath on secondary process. */
1412 mlx5_mp_os_req_stop_rxtx(dev);
1413 /* Free the eCPRI flex parser resource. */
1414 mlx5_flex_parser_ecpri_release(dev);
1415 if (priv->rxqs != NULL) {
1416 /* XXX race condition if mlx5_rx_burst() is still running. */
1418 for (i = 0; (i != priv->rxqs_n); ++i)
1419 mlx5_rxq_release(dev, i);
1423 if (priv->txqs != NULL) {
1424 /* XXX race condition if mlx5_tx_burst() is still running. */
1426 for (i = 0; (i != priv->txqs_n); ++i)
1427 mlx5_txq_release(dev, i);
1431 mlx5_proc_priv_uninit(dev);
1432 if (priv->mreg_cp_tbl)
1433 mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1434 mlx5_mprq_free_mp(dev);
1435 mlx5_os_free_shared_dr(priv);
1436 if (priv->rss_conf.rss_key != NULL)
1437 mlx5_free(priv->rss_conf.rss_key);
1438 if (priv->reta_idx != NULL)
1439 mlx5_free(priv->reta_idx);
1440 if (priv->config.vf)
1441 mlx5_os_mac_addr_flush(dev);
1442 if (priv->nl_socket_route >= 0)
1443 close(priv->nl_socket_route);
1444 if (priv->nl_socket_rdma >= 0)
1445 close(priv->nl_socket_rdma);
1446 if (priv->vmwa_context)
1447 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1448 ret = mlx5_hrxq_verify(dev);
1450 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1451 dev->data->port_id);
1452 ret = mlx5_ind_table_obj_verify(dev);
1454 DRV_LOG(WARNING, "port %u some indirection table still remain",
1455 dev->data->port_id);
1456 ret = mlx5_rxq_obj_verify(dev);
1458 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1459 dev->data->port_id);
1460 ret = mlx5_rxq_verify(dev);
1462 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1463 dev->data->port_id);
1464 ret = mlx5_txq_obj_verify(dev);
1466 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1467 dev->data->port_id);
1468 ret = mlx5_txq_verify(dev);
1470 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1471 dev->data->port_id);
1472 ret = mlx5_flow_verify(dev);
1474 DRV_LOG(WARNING, "port %u some flows still remain",
1475 dev->data->port_id);
1477 * Free the shared context in last turn, because the cleanup
1478 * routines above may use some shared fields, like
1479 * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1480 * ifindex if Netlink fails.
1482 mlx5_free_shared_dev_ctx(priv->sh);
1483 if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1487 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1488 struct mlx5_priv *opriv =
1489 rte_eth_devices[port_id].data->dev_private;
1492 opriv->domain_id != priv->domain_id ||
1493 &rte_eth_devices[port_id] == dev)
1499 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1501 memset(priv, 0, sizeof(*priv));
1502 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1504 * Reset mac_addrs to NULL such that it is not freed as part of
1505 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1506 * it is freed when dev_private is freed.
1508 dev->data->mac_addrs = NULL;
1513 * Verify and store value for device argument.
1516 * Key argument to verify.
1518 * Value associated with key.
1523 * 0 on success, a negative errno value otherwise and rte_errno is set.
1526 mlx5_args_check(const char *key, const char *val, void *opaque)
1528 struct mlx5_dev_config *config = opaque;
1532 /* No-op, port representors are processed in mlx5_dev_spawn(). */
1533 if (!strcmp(MLX5_REPRESENTOR, key))
1536 tmp = strtol(val, NULL, 0);
1539 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1542 if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1543 /* Negative values are acceptable for some keys only. */
1545 DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1548 mod = tmp >= 0 ? tmp : -tmp;
1549 if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1550 config->cqe_comp = !!tmp;
1551 } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1552 config->cqe_pad = !!tmp;
1553 } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1554 config->hw_padding = !!tmp;
1555 } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1556 config->mprq.enabled = !!tmp;
1557 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1558 config->mprq.stride_num_n = tmp;
1559 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1560 config->mprq.stride_size_n = tmp;
1561 } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1562 config->mprq.max_memcpy_len = tmp;
1563 } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1564 config->mprq.min_rxqs_num = tmp;
1565 } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1566 DRV_LOG(WARNING, "%s: deprecated parameter,"
1567 " converted to txq_inline_max", key);
1568 config->txq_inline_max = tmp;
1569 } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1570 config->txq_inline_max = tmp;
1571 } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1572 config->txq_inline_min = tmp;
1573 } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1574 config->txq_inline_mpw = tmp;
1575 } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1576 config->txqs_inline = tmp;
1577 } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1578 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1579 } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1580 config->mps = !!tmp;
1581 } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1582 if (tmp != MLX5_TXDB_CACHED &&
1583 tmp != MLX5_TXDB_NCACHED &&
1584 tmp != MLX5_TXDB_HEURISTIC) {
1585 DRV_LOG(ERR, "invalid Tx doorbell "
1586 "mapping parameter");
1591 } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1592 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1593 } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1594 DRV_LOG(WARNING, "%s: deprecated parameter,"
1595 " converted to txq_inline_mpw", key);
1596 config->txq_inline_mpw = tmp;
1597 } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1598 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1599 } else if (strcmp(MLX5_TX_PP, key) == 0) {
1601 DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1605 config->tx_pp = tmp;
1606 } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1607 config->tx_skew = tmp;
1608 } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1609 config->rx_vec_en = !!tmp;
1610 } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1611 config->l3_vxlan_en = !!tmp;
1612 } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1613 config->vf_nl_en = !!tmp;
1614 } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1615 config->dv_esw_en = !!tmp;
1616 } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1617 config->dv_flow_en = !!tmp;
1618 } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1619 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1620 tmp != MLX5_XMETA_MODE_META16 &&
1621 tmp != MLX5_XMETA_MODE_META32) {
1622 DRV_LOG(ERR, "invalid extensive "
1623 "metadata parameter");
1627 config->dv_xmeta_en = tmp;
1628 } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1629 config->lacp_by_user = !!tmp;
1630 } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1631 config->mr_ext_memseg_en = !!tmp;
1632 } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1633 config->max_dump_files_num = tmp;
1634 } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1635 config->lro.timeout = tmp;
1636 } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1637 DRV_LOG(DEBUG, "class argument is %s.", val);
1638 } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1639 config->log_hp_size = tmp;
1640 } else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1641 if (tmp != MLX5_RCM_NONE &&
1642 tmp != MLX5_RCM_LIGHT &&
1643 tmp != MLX5_RCM_AGGR) {
1644 DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1648 config->reclaim_mode = tmp;
1649 } else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1650 config->sys_mem_en = !!tmp;
1651 } else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1652 config->decap_en = !!tmp;
1654 DRV_LOG(WARNING, "%s: unknown parameter", key);
1662 * Parse device parameters.
1665 * Pointer to device configuration structure.
1667 * Device arguments structure.
1670 * 0 on success, a negative errno value otherwise and rte_errno is set.
1673 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1675 const char **params = (const char *[]){
1676 MLX5_RXQ_CQE_COMP_EN,
1677 MLX5_RXQ_CQE_PAD_EN,
1678 MLX5_RXQ_PKT_PAD_EN,
1680 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1681 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1682 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1685 MLX5_TXQ_INLINE_MIN,
1686 MLX5_TXQ_INLINE_MAX,
1687 MLX5_TXQ_INLINE_MPW,
1688 MLX5_TXQS_MIN_INLINE,
1691 MLX5_TXQ_MPW_HDR_DSEG_EN,
1692 MLX5_TXQ_MAX_INLINE_LEN,
1704 MLX5_MR_EXT_MEMSEG_EN,
1706 MLX5_MAX_DUMP_FILES_NUM,
1707 MLX5_LRO_TIMEOUT_USEC,
1708 MLX5_CLASS_ARG_NAME,
1715 struct rte_kvargs *kvlist;
1719 if (devargs == NULL)
1721 /* Following UGLY cast is done to pass checkpatch. */
1722 kvlist = rte_kvargs_parse(devargs->args, params);
1723 if (kvlist == NULL) {
1727 /* Process parameters. */
1728 for (i = 0; (params[i] != NULL); ++i) {
1729 if (rte_kvargs_count(kvlist, params[i])) {
1730 ret = rte_kvargs_process(kvlist, params[i],
1731 mlx5_args_check, config);
1734 rte_kvargs_free(kvlist);
1739 rte_kvargs_free(kvlist);
1744 * Configures the minimal amount of data to inline into WQE
1745 * while sending packets.
1747 * - the txq_inline_min has the maximal priority, if this
1748 * key is specified in devargs
1749 * - if DevX is enabled the inline mode is queried from the
1750 * device (HCA attributes and NIC vport context if needed).
1751 * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1752 * and none (0 bytes) for other NICs
1755 * Verbs device parameters (name, port, switch_info) to spawn.
1757 * Device configuration parameters.
1760 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1761 struct mlx5_dev_config *config)
1763 if (config->txq_inline_min != MLX5_ARG_UNSET) {
1764 /* Application defines size of inlined data explicitly. */
1765 switch (spawn->pci_dev->id.device_id) {
1766 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1767 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1768 if (config->txq_inline_min <
1769 (int)MLX5_INLINE_HSIZE_L2) {
1771 "txq_inline_mix aligned to minimal"
1772 " ConnectX-4 required value %d",
1773 (int)MLX5_INLINE_HSIZE_L2);
1774 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1780 if (config->hca_attr.eth_net_offloads) {
1781 /* We have DevX enabled, inline mode queried successfully. */
1782 switch (config->hca_attr.wqe_inline_mode) {
1783 case MLX5_CAP_INLINE_MODE_L2:
1784 /* outer L2 header must be inlined. */
1785 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1787 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1788 /* No inline data are required by NIC. */
1789 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1790 config->hw_vlan_insert =
1791 config->hca_attr.wqe_vlan_insert;
1792 DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1794 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1795 /* inline mode is defined by NIC vport context. */
1796 if (!config->hca_attr.eth_virt)
1798 switch (config->hca_attr.vport_inline_mode) {
1799 case MLX5_INLINE_MODE_NONE:
1800 config->txq_inline_min =
1801 MLX5_INLINE_HSIZE_NONE;
1803 case MLX5_INLINE_MODE_L2:
1804 config->txq_inline_min =
1805 MLX5_INLINE_HSIZE_L2;
1807 case MLX5_INLINE_MODE_IP:
1808 config->txq_inline_min =
1809 MLX5_INLINE_HSIZE_L3;
1811 case MLX5_INLINE_MODE_TCP_UDP:
1812 config->txq_inline_min =
1813 MLX5_INLINE_HSIZE_L4;
1815 case MLX5_INLINE_MODE_INNER_L2:
1816 config->txq_inline_min =
1817 MLX5_INLINE_HSIZE_INNER_L2;
1819 case MLX5_INLINE_MODE_INNER_IP:
1820 config->txq_inline_min =
1821 MLX5_INLINE_HSIZE_INNER_L3;
1823 case MLX5_INLINE_MODE_INNER_TCP_UDP:
1824 config->txq_inline_min =
1825 MLX5_INLINE_HSIZE_INNER_L4;
1831 * We get here if we are unable to deduce
1832 * inline data size with DevX. Try PCI ID
1833 * to determine old NICs.
1835 switch (spawn->pci_dev->id.device_id) {
1836 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1837 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1838 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1839 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1840 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1841 config->hw_vlan_insert = 0;
1843 case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1844 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1845 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1846 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1848 * These NICs support VLAN insertion from WQE and
1849 * report the wqe_vlan_insert flag. But there is the bug
1850 * and PFC control may be broken, so disable feature.
1852 config->hw_vlan_insert = 0;
1853 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1856 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1860 DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1864 * Configures the metadata mask fields in the shared context.
1867 * Pointer to Ethernet device.
1870 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1872 struct mlx5_priv *priv = dev->data->dev_private;
1873 struct mlx5_dev_ctx_shared *sh = priv->sh;
1874 uint32_t meta, mark, reg_c0;
1876 reg_c0 = ~priv->vport_meta_mask;
1877 switch (priv->config.dv_xmeta_en) {
1878 case MLX5_XMETA_MODE_LEGACY:
1880 mark = MLX5_FLOW_MARK_MASK;
1882 case MLX5_XMETA_MODE_META16:
1883 meta = reg_c0 >> rte_bsf32(reg_c0);
1884 mark = MLX5_FLOW_MARK_MASK;
1886 case MLX5_XMETA_MODE_META32:
1888 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1896 if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1897 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1898 sh->dv_mark_mask, mark);
1900 sh->dv_mark_mask = mark;
1901 if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1902 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1903 sh->dv_meta_mask, meta);
1905 sh->dv_meta_mask = meta;
1906 if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1907 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1908 sh->dv_meta_mask, reg_c0);
1910 sh->dv_regc0_mask = reg_c0;
1911 DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1912 DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1913 DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1914 DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1918 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
1920 static const char *const dynf_names[] = {
1921 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
1922 RTE_MBUF_DYNFLAG_METADATA_NAME,
1923 RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
1927 if (n < RTE_DIM(dynf_names))
1929 for (i = 0; i < RTE_DIM(dynf_names); i++) {
1930 if (names[i] == NULL)
1932 strcpy(names[i], dynf_names[i]);
1934 return RTE_DIM(dynf_names);
1938 * Comparison callback to sort device data.
1940 * This is meant to be used with qsort().
1943 * Pointer to pointer to first data object.
1945 * Pointer to pointer to second data object.
1948 * 0 if both objects are equal, less than 0 if the first argument is less
1949 * than the second, greater than 0 otherwise.
1952 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1953 struct mlx5_dev_config *config)
1955 struct mlx5_dev_ctx_shared *sh = priv->sh;
1956 struct mlx5_dev_config *sh_conf = NULL;
1960 /* Nothing to compare for the single/first device. */
1961 if (sh->refcnt == 1)
1963 /* Find the device with shared context. */
1964 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1965 struct mlx5_priv *opriv =
1966 rte_eth_devices[port_id].data->dev_private;
1968 if (opriv && opriv != priv && opriv->sh == sh) {
1969 sh_conf = &opriv->config;
1975 if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
1976 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
1977 " for shared %s context", sh->ibdev_name);
1981 if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
1982 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
1983 " for shared %s context", sh->ibdev_name);
1991 * Look for the ethernet device belonging to mlx5 driver.
1993 * @param[in] port_id
1994 * port_id to start looking for device.
1995 * @param[in] pci_dev
1996 * Pointer to the hint PCI device. When device is being probed
1997 * the its siblings (master and preceding representors might
1998 * not have assigned driver yet (because the mlx5_os_pci_probe()
1999 * is not completed yet, for this case match on hint PCI
2000 * device may be used to detect sibling device.
2003 * port_id of found device, RTE_MAX_ETHPORT if not found.
2006 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
2008 while (port_id < RTE_MAX_ETHPORTS) {
2009 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2011 if (dev->state != RTE_ETH_DEV_UNUSED &&
2013 (dev->device == &pci_dev->device ||
2014 (dev->device->driver &&
2015 dev->device->driver->name &&
2016 !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
2020 if (port_id >= RTE_MAX_ETHPORTS)
2021 return RTE_MAX_ETHPORTS;
2026 * DPDK callback to remove a PCI device.
2028 * This function removes all Ethernet devices belong to a given PCI device.
2030 * @param[in] pci_dev
2031 * Pointer to the PCI device.
2034 * 0 on success, the function cannot fail.
2037 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2042 RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
2044 * mlx5_dev_close() is not registered to secondary process,
2045 * call the close function explicitly for secondary process.
2047 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2048 ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2050 ret |= rte_eth_dev_close(port_id);
2052 return ret == 0 ? 0 : -EIO;
2055 static const struct rte_pci_id mlx5_pci_id_map[] = {
2057 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2058 PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2061 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2062 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2065 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2066 PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2069 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2070 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2073 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2074 PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2077 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2078 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2081 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2082 PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2085 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2086 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2089 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2090 PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2093 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2094 PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2097 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2098 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2101 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2102 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2105 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2106 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2109 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2110 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
2113 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2114 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2117 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2118 PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2125 static struct mlx5_pci_driver mlx5_driver = {
2126 .driver_class = MLX5_CLASS_NET,
2129 .name = MLX5_DRIVER_NAME,
2131 .id_table = mlx5_pci_id_map,
2132 .probe = mlx5_os_pci_probe,
2133 .remove = mlx5_pci_remove,
2134 .dma_map = mlx5_dma_map,
2135 .dma_unmap = mlx5_dma_unmap,
2136 .drv_flags = PCI_DRV_FLAGS,
2140 /* Initialize driver log type. */
2141 RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
2144 * Driver initialization routine.
2146 RTE_INIT(rte_mlx5_pmd_init)
2149 /* Build the static tables for Verbs conversion. */
2150 mlx5_set_ptype_table();
2151 mlx5_set_cksum_table();
2152 mlx5_set_swp_types_table();
2154 mlx5_pci_driver_register(&mlx5_driver);
2157 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2158 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2159 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");