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>
24 #include <rte_cycles.h>
26 #include <mlx5_glue.h>
27 #include <mlx5_devx_cmds.h>
28 #include <mlx5_common.h>
29 #include <mlx5_common_os.h>
30 #include <mlx5_common_mp.h>
31 #include <mlx5_common_pci.h>
32 #include <mlx5_malloc.h>
34 #include "mlx5_defs.h"
36 #include "mlx5_utils.h"
37 #include "mlx5_rxtx.h"
38 #include "mlx5_autoconf.h"
40 #include "mlx5_flow.h"
41 #include "rte_pmd_mlx5.h"
43 /* Device parameter to enable RX completion queue compression. */
44 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
46 /* Device parameter to enable RX completion entry padding to 128B. */
47 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
49 /* Device parameter to enable padding Rx packet to cacheline size. */
50 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
52 /* Device parameter to enable Multi-Packet Rx queue. */
53 #define MLX5_RX_MPRQ_EN "mprq_en"
55 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
56 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
58 /* Device parameter to configure log 2 of the stride size for MPRQ. */
59 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
61 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
62 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
64 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
65 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
67 /* Device parameter to configure inline send. Deprecated, ignored.*/
68 #define MLX5_TXQ_INLINE "txq_inline"
70 /* Device parameter to limit packet size to inline with ordinary SEND. */
71 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
73 /* Device parameter to configure minimal data size to inline. */
74 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
76 /* Device parameter to limit packet size to inline with Enhanced MPW. */
77 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
80 * Device parameter to configure the number of TX queues threshold for
81 * enabling inline send.
83 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
86 * Device parameter to configure the number of TX queues threshold for
87 * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
89 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
91 /* Device parameter to enable multi-packet send WQEs. */
92 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
95 * Device parameter to force doorbell register mapping
96 * to non-cahed region eliminating the extra write memory barrier.
98 #define MLX5_TX_DB_NC "tx_db_nc"
101 * Device parameter to include 2 dsegs in the title WQEBB.
102 * Deprecated, ignored.
104 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
107 * Device parameter to limit the size of inlining packet.
108 * Deprecated, ignored.
110 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
113 * Device parameter to enable Tx scheduling on timestamps
114 * and specify the packet pacing granularity in nanoseconds.
116 #define MLX5_TX_PP "tx_pp"
119 * Device parameter to specify skew in nanoseconds on Tx datapath,
120 * it represents the time between SQ start WQE processing and
121 * appearing actual packet data on the wire.
123 #define MLX5_TX_SKEW "tx_skew"
126 * Device parameter to enable hardware Tx vector.
127 * Deprecated, ignored (no vectorized Tx routines anymore).
129 #define MLX5_TX_VEC_EN "tx_vec_en"
131 /* Device parameter to enable hardware Rx vector. */
132 #define MLX5_RX_VEC_EN "rx_vec_en"
134 /* Allow L3 VXLAN flow creation. */
135 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
137 /* Activate DV E-Switch flow steering. */
138 #define MLX5_DV_ESW_EN "dv_esw_en"
140 /* Activate DV flow steering. */
141 #define MLX5_DV_FLOW_EN "dv_flow_en"
143 /* Enable extensive flow metadata support. */
144 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
146 /* Device parameter to let the user manage the lacp traffic of bonded device */
147 #define MLX5_LACP_BY_USER "lacp_by_user"
149 /* Activate Netlink support in VF mode. */
150 #define MLX5_VF_NL_EN "vf_nl_en"
152 /* Enable extending memsegs when creating a MR. */
153 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
155 /* Select port representors to instantiate. */
156 #define MLX5_REPRESENTOR "representor"
158 /* Device parameter to configure the maximum number of dump files per queue. */
159 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
161 /* Configure timeout of LRO session (in microseconds). */
162 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
165 * Device parameter to configure the total data buffer size for a single
166 * hairpin queue (logarithm value).
168 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
170 /* Flow memory reclaim mode. */
171 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
173 /* The default memory allocator used in PMD. */
174 #define MLX5_SYS_MEM_EN "sys_mem_en"
175 /* Decap will be used or not. */
176 #define MLX5_DECAP_EN "decap_en"
178 /* Shared memory between primary and secondary processes. */
179 struct mlx5_shared_data *mlx5_shared_data;
181 /** Driver-specific log messages type. */
184 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
185 LIST_HEAD_INITIALIZER();
186 static pthread_mutex_t mlx5_dev_ctx_list_mutex = PTHREAD_MUTEX_INITIALIZER;
188 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
189 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
190 [MLX5_IPOOL_DECAP_ENCAP] = {
191 .size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
197 .malloc = mlx5_malloc,
199 .type = "mlx5_encap_decap_ipool",
201 [MLX5_IPOOL_PUSH_VLAN] = {
202 .size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
208 .malloc = mlx5_malloc,
210 .type = "mlx5_push_vlan_ipool",
213 .size = sizeof(struct mlx5_flow_dv_tag_resource),
219 .malloc = mlx5_malloc,
221 .type = "mlx5_tag_ipool",
223 [MLX5_IPOOL_PORT_ID] = {
224 .size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
230 .malloc = mlx5_malloc,
232 .type = "mlx5_port_id_ipool",
234 [MLX5_IPOOL_JUMP] = {
235 .size = sizeof(struct mlx5_flow_tbl_data_entry),
241 .malloc = mlx5_malloc,
243 .type = "mlx5_jump_ipool",
245 [MLX5_IPOOL_SAMPLE] = {
246 .size = sizeof(struct mlx5_flow_dv_sample_resource),
252 .malloc = mlx5_malloc,
254 .type = "mlx5_sample_ipool",
256 [MLX5_IPOOL_DEST_ARRAY] = {
257 .size = sizeof(struct mlx5_flow_dv_dest_array_resource),
263 .malloc = mlx5_malloc,
265 .type = "mlx5_dest_array_ipool",
267 [MLX5_IPOOL_TUNNEL_ID] = {
268 .size = sizeof(struct mlx5_flow_tunnel),
269 .trunk_size = MLX5_MAX_TUNNELS,
272 .type = "mlx5_tunnel_offload",
274 [MLX5_IPOOL_TNL_TBL_ID] = {
277 .type = "mlx5_flow_tnl_tbl_ipool",
281 .size = sizeof(struct mlx5_flow_meter),
287 .malloc = mlx5_malloc,
289 .type = "mlx5_meter_ipool",
292 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
298 .malloc = mlx5_malloc,
300 .type = "mlx5_mcp_ipool",
302 [MLX5_IPOOL_HRXQ] = {
303 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
309 .malloc = mlx5_malloc,
311 .type = "mlx5_hrxq_ipool",
313 [MLX5_IPOOL_MLX5_FLOW] = {
315 * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
316 * It set in run time according to PCI function configuration.
324 .malloc = mlx5_malloc,
326 .type = "mlx5_flow_handle_ipool",
328 [MLX5_IPOOL_RTE_FLOW] = {
329 .size = sizeof(struct rte_flow),
333 .malloc = mlx5_malloc,
335 .type = "rte_flow_ipool",
337 [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
340 .type = "mlx5_flow_rss_id_ipool",
342 [MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
343 .size = sizeof(struct mlx5_shared_action_rss),
349 .malloc = mlx5_malloc,
351 .type = "mlx5_shared_action_rss",
356 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
357 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
359 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
362 * Initialize the ASO aging management structure.
365 * Pointer to mlx5_dev_ctx_shared object to free
368 * 0 on success, a negative errno value otherwise and rte_errno is set.
371 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
377 sh->aso_age_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->aso_age_mng),
378 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
379 if (!sh->aso_age_mng) {
380 DRV_LOG(ERR, "aso_age_mng allocation was failed.");
384 err = mlx5_aso_queue_init(sh);
386 mlx5_free(sh->aso_age_mng);
389 rte_spinlock_init(&sh->aso_age_mng->resize_sl);
390 rte_spinlock_init(&sh->aso_age_mng->free_sl);
391 LIST_INIT(&sh->aso_age_mng->free);
396 * Close and release all the resources of the ASO aging management structure.
399 * Pointer to mlx5_dev_ctx_shared object to free.
402 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
406 mlx5_aso_queue_stop(sh);
407 mlx5_aso_queue_uninit(sh);
408 if (sh->aso_age_mng->pools) {
409 struct mlx5_aso_age_pool *pool;
411 for (i = 0; i < sh->aso_age_mng->next; ++i) {
412 pool = sh->aso_age_mng->pools[i];
413 claim_zero(mlx5_devx_cmd_destroy
414 (pool->flow_hit_aso_obj));
415 for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j)
416 if (pool->actions[j].dr_action)
418 (mlx5_glue->destroy_flow_action
419 (pool->actions[j].dr_action));
422 mlx5_free(sh->aso_age_mng->pools);
424 mlx5_free(sh->aso_age_mng);
428 * Initialize the shared aging list information per port.
431 * Pointer to mlx5_dev_ctx_shared object.
434 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
437 struct mlx5_age_info *age_info;
439 for (i = 0; i < sh->max_port; i++) {
440 age_info = &sh->port[i].age_info;
442 TAILQ_INIT(&age_info->aged_counters);
443 LIST_INIT(&age_info->aged_aso);
444 rte_spinlock_init(&age_info->aged_sl);
445 MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
450 * Initialize the counters management structure.
453 * Pointer to mlx5_dev_ctx_shared object to free
456 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
460 memset(&sh->cmng, 0, sizeof(sh->cmng));
461 TAILQ_INIT(&sh->cmng.flow_counters);
462 sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
463 sh->cmng.max_id = -1;
464 sh->cmng.last_pool_idx = POOL_IDX_INVALID;
465 rte_spinlock_init(&sh->cmng.pool_update_sl);
466 for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
467 TAILQ_INIT(&sh->cmng.counters[i]);
468 rte_spinlock_init(&sh->cmng.csl[i]);
473 * Destroy all the resources allocated for a counter memory management.
476 * Pointer to the memory management structure.
479 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
481 uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
483 LIST_REMOVE(mng, next);
484 claim_zero(mlx5_devx_cmd_destroy(mng->dm));
485 claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
490 * Close and release all the resources of the counters management.
493 * Pointer to mlx5_dev_ctx_shared object to free.
496 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
498 struct mlx5_counter_stats_mem_mng *mng;
504 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
505 if (rte_errno != EINPROGRESS)
510 if (sh->cmng.pools) {
511 struct mlx5_flow_counter_pool *pool;
512 uint16_t n_valid = sh->cmng.n_valid;
513 bool fallback = sh->cmng.counter_fallback;
515 for (i = 0; i < n_valid; ++i) {
516 pool = sh->cmng.pools[i];
517 if (!fallback && pool->min_dcs)
518 claim_zero(mlx5_devx_cmd_destroy
520 for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
521 struct mlx5_flow_counter *cnt =
522 MLX5_POOL_GET_CNT(pool, j);
526 (mlx5_glue->destroy_flow_action
528 if (fallback && MLX5_POOL_GET_CNT
529 (pool, j)->dcs_when_free)
530 claim_zero(mlx5_devx_cmd_destroy
531 (cnt->dcs_when_free));
535 mlx5_free(sh->cmng.pools);
537 mng = LIST_FIRST(&sh->cmng.mem_mngs);
539 mlx5_flow_destroy_counter_stat_mem_mng(mng);
540 mng = LIST_FIRST(&sh->cmng.mem_mngs);
542 memset(&sh->cmng, 0, sizeof(sh->cmng));
545 /* Send FLOW_AGED event if needed. */
547 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
549 struct mlx5_age_info *age_info;
552 for (i = 0; i < sh->max_port; i++) {
553 age_info = &sh->port[i].age_info;
554 if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
556 if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER))
557 rte_eth_dev_callback_process
558 (&rte_eth_devices[sh->port[i].devx_ih_port_id],
559 RTE_ETH_EVENT_FLOW_AGED, NULL);
565 * Initialize the flow resources' indexed mempool.
568 * Pointer to mlx5_dev_ctx_shared object.
570 * Pointer to user dev config.
573 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
574 const struct mlx5_dev_config *config)
577 struct mlx5_indexed_pool_config cfg;
579 for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
580 cfg = mlx5_ipool_cfg[i];
585 * Set MLX5_IPOOL_MLX5_FLOW ipool size
586 * according to PCI function flow configuration.
588 case MLX5_IPOOL_MLX5_FLOW:
589 cfg.size = config->dv_flow_en ?
590 sizeof(struct mlx5_flow_handle) :
591 MLX5_FLOW_HANDLE_VERBS_SIZE;
594 if (config->reclaim_mode)
595 cfg.release_mem_en = 1;
596 sh->ipool[i] = mlx5_ipool_create(&cfg);
601 * Release the flow resources' indexed mempool.
604 * Pointer to mlx5_dev_ctx_shared object.
607 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
611 for (i = 0; i < MLX5_IPOOL_MAX; ++i)
612 mlx5_ipool_destroy(sh->ipool[i]);
616 * Check if dynamic flex parser for eCPRI already exists.
619 * Pointer to Ethernet device structure.
622 * true on exists, false on not.
625 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
627 struct mlx5_priv *priv = dev->data->dev_private;
628 struct mlx5_flex_parser_profiles *prf =
629 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
635 * Allocation of a flex parser for eCPRI. Once created, this parser related
636 * resources will be held until the device is closed.
639 * Pointer to Ethernet device structure.
642 * 0 on success, a negative errno value otherwise and rte_errno is set.
645 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
647 struct mlx5_priv *priv = dev->data->dev_private;
648 struct mlx5_flex_parser_profiles *prf =
649 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
650 struct mlx5_devx_graph_node_attr node = {
651 .modify_field_select = 0,
656 if (!priv->config.hca_attr.parse_graph_flex_node) {
657 DRV_LOG(ERR, "Dynamic flex parser is not supported "
658 "for device %s.", priv->dev_data->name);
661 node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
662 /* 8 bytes now: 4B common header + 4B message body header. */
663 node.header_length_base_value = 0x8;
664 /* After MAC layer: Ether / VLAN. */
665 node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
666 /* Type of compared condition should be 0xAEFE in the L2 layer. */
667 node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
668 /* Sample #0: type in common header. */
669 node.sample[0].flow_match_sample_en = 1;
671 node.sample[0].flow_match_sample_offset_mode = 0x0;
672 /* Only the 2nd byte will be used. */
673 node.sample[0].flow_match_sample_field_base_offset = 0x0;
674 /* Sample #1: message payload. */
675 node.sample[1].flow_match_sample_en = 1;
677 node.sample[1].flow_match_sample_offset_mode = 0x0;
679 * Only the first two bytes will be used right now, and its offset will
680 * start after the common header that with the length of a DW(u32).
682 node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
683 prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
685 DRV_LOG(ERR, "Failed to create flex parser node object.");
686 return (rte_errno == 0) ? -ENODEV : -rte_errno;
689 ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
691 DRV_LOG(ERR, "Failed to query sample IDs.");
692 return (rte_errno == 0) ? -ENODEV : -rte_errno;
694 prf->offset[0] = 0x0;
695 prf->offset[1] = sizeof(uint32_t);
696 prf->ids[0] = ids[0];
697 prf->ids[1] = ids[1];
702 * Destroy the flex parser node, including the parser itself, input / output
703 * arcs and DW samples. Resources could be reused then.
706 * Pointer to Ethernet device structure.
709 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
711 struct mlx5_priv *priv = dev->data->dev_private;
712 struct mlx5_flex_parser_profiles *prf =
713 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
716 mlx5_devx_cmd_destroy(prf->obj);
721 * Allocate Rx and Tx UARs in robust fashion.
722 * This routine handles the following UAR allocation issues:
724 * - tries to allocate the UAR with the most appropriate memory
725 * mapping type from the ones supported by the host
727 * - tries to allocate the UAR with non-NULL base address
728 * OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
729 * UAR base address if UAR was not the first object in the UAR page.
730 * It caused the PMD failure and we should try to get another UAR
731 * till we get the first one with non-NULL base address returned.
734 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
735 const struct mlx5_dev_config *config)
737 uint32_t uar_mapping, retry;
741 for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
742 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
743 /* Control the mapping type according to the settings. */
744 uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
745 MLX5DV_UAR_ALLOC_TYPE_NC :
746 MLX5DV_UAR_ALLOC_TYPE_BF;
748 RTE_SET_USED(config);
750 * It seems we have no way to control the memory mapping type
751 * for the UAR, the default "Write-Combining" type is supposed.
752 * The UAR initialization on queue creation queries the
753 * actual mapping type done by Verbs/kernel and setups the
754 * PMD datapath accordingly.
758 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, uar_mapping);
759 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
761 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
762 if (config->dbnc == MLX5_TXDB_CACHED ||
763 config->dbnc == MLX5_TXDB_HEURISTIC)
764 DRV_LOG(WARNING, "Devarg tx_db_nc setting "
765 "is not supported by DevX");
767 * In some environments like virtual machine
768 * the Write Combining mapped might be not supported
769 * and UAR allocation fails. We try "Non-Cached"
770 * mapping for the case. The tx_burst routines take
771 * the UAR mapping type into account on UAR setup
774 DRV_LOG(WARNING, "Failed to allocate Tx DevX UAR (BF)");
775 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
776 sh->tx_uar = mlx5_glue->devx_alloc_uar
777 (sh->ctx, uar_mapping);
778 } else if (!sh->tx_uar &&
779 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
780 if (config->dbnc == MLX5_TXDB_NCACHED)
781 DRV_LOG(WARNING, "Devarg tx_db_nc settings "
782 "is not supported by DevX");
784 * If Verbs/kernel does not support "Non-Cached"
785 * try the "Write-Combining".
787 DRV_LOG(WARNING, "Failed to allocate Tx DevX UAR (NC)");
788 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
789 sh->tx_uar = mlx5_glue->devx_alloc_uar
790 (sh->ctx, uar_mapping);
794 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
798 base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
802 * The UARs are allocated by rdma_core within the
803 * IB device context, on context closure all UARs
804 * will be freed, should be no memory/object leakage.
806 DRV_LOG(WARNING, "Retrying to allocate Tx DevX UAR");
809 /* Check whether we finally succeeded with valid UAR allocation. */
811 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
815 for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
817 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
818 (sh->ctx, uar_mapping);
819 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
820 if (!sh->devx_rx_uar &&
821 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
823 * Rx UAR is used to control interrupts only,
824 * should be no datapath noticeable impact,
825 * can try "Non-Cached" mapping safely.
827 DRV_LOG(WARNING, "Failed to allocate Rx DevX UAR (BF)");
828 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
829 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
830 (sh->ctx, uar_mapping);
833 if (!sh->devx_rx_uar) {
834 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
838 base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
842 * The UARs are allocated by rdma_core within the
843 * IB device context, on context closure all UARs
844 * will be freed, should be no memory/object leakage.
846 DRV_LOG(WARNING, "Retrying to allocate Rx DevX UAR");
847 sh->devx_rx_uar = NULL;
849 /* Check whether we finally succeeded with valid UAR allocation. */
850 if (!sh->devx_rx_uar) {
851 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
859 * Allocate shared device context. If there is multiport device the
860 * master and representors will share this context, if there is single
861 * port dedicated device, the context will be used by only given
862 * port due to unification.
864 * Routine first searches the context for the specified device name,
865 * if found the shared context assumed and reference counter is incremented.
866 * If no context found the new one is created and initialized with specified
867 * device context and parameters.
870 * Pointer to the device attributes (name, port, etc).
872 * Pointer to device configuration structure.
875 * Pointer to mlx5_dev_ctx_shared object on success,
876 * otherwise NULL and rte_errno is set.
878 struct mlx5_dev_ctx_shared *
879 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
880 const struct mlx5_dev_config *config)
882 struct mlx5_dev_ctx_shared *sh;
885 struct mlx5_devx_tis_attr tis_attr = { 0 };
888 /* Secondary process should not create the shared context. */
889 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
890 pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
891 /* Search for IB context by device name. */
892 LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
893 if (!strcmp(sh->ibdev_name,
894 mlx5_os_get_dev_device_name(spawn->phys_dev))) {
899 /* No device found, we have to create new shared context. */
900 MLX5_ASSERT(spawn->max_port);
901 sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
902 sizeof(struct mlx5_dev_ctx_shared) +
904 sizeof(struct mlx5_dev_shared_port),
905 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
907 DRV_LOG(ERR, "shared context allocation failure");
911 err = mlx5_os_open_device(spawn, config, sh);
914 err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
916 DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
920 sh->bond_dev = UINT16_MAX;
921 sh->max_port = spawn->max_port;
922 strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
923 sizeof(sh->ibdev_name) - 1);
924 strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
925 sizeof(sh->ibdev_path) - 1);
927 * Setting port_id to max unallowed value means
928 * there is no interrupt subhandler installed for
929 * the given port index i.
931 for (i = 0; i < sh->max_port; i++) {
932 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
933 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
935 sh->pd = mlx5_os_alloc_pd(sh->ctx);
936 if (sh->pd == NULL) {
937 DRV_LOG(ERR, "PD allocation failure");
942 /* Query the EQN for this core. */
943 err = mlx5_glue->devx_query_eqn(sh->ctx, 0, &sh->eqn);
946 DRV_LOG(ERR, "Failed to query event queue number %d.",
950 err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
952 DRV_LOG(ERR, "Fail to extract pdn from PD");
955 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
957 DRV_LOG(ERR, "TD allocation failure");
961 tis_attr.transport_domain = sh->td->id;
962 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
964 DRV_LOG(ERR, "TIS allocation failure");
968 err = mlx5_alloc_rxtx_uars(sh, config);
971 MLX5_ASSERT(sh->tx_uar);
972 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
974 MLX5_ASSERT(sh->devx_rx_uar);
975 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
978 /* Initialize UAR access locks for 32bit implementations. */
979 rte_spinlock_init(&sh->uar_lock_cq);
980 for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
981 rte_spinlock_init(&sh->uar_lock[i]);
984 * Once the device is added to the list of memory event
985 * callback, its global MR cache table cannot be expanded
986 * on the fly because of deadlock. If it overflows, lookup
987 * should be done by searching MR list linearly, which is slow.
989 * At this point the device is not added to the memory
990 * event list yet, context is just being created.
992 err = mlx5_mr_btree_init(&sh->share_cache.cache,
993 MLX5_MR_BTREE_CACHE_N * 2,
994 spawn->pci_dev->device.numa_node);
999 mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
1000 &sh->share_cache.dereg_mr_cb);
1001 mlx5_os_dev_shared_handler_install(sh);
1002 sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
1003 if (!sh->cnt_id_tbl) {
1007 mlx5_flow_aging_init(sh);
1008 mlx5_flow_counters_mng_init(sh);
1009 mlx5_flow_ipool_create(sh, config);
1010 /* Add device to memory callback list. */
1011 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1012 LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1014 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1015 /* Add context to the global device list. */
1016 LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1018 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1021 pthread_mutex_destroy(&sh->txpp.mutex);
1022 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1025 mlx5_l3t_destroy(sh->cnt_id_tbl);
1027 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1029 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1030 if (sh->devx_rx_uar)
1031 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1033 mlx5_glue->devx_free_uar(sh->tx_uar);
1035 claim_zero(mlx5_os_dealloc_pd(sh->pd));
1037 claim_zero(mlx5_glue->close_device(sh->ctx));
1039 MLX5_ASSERT(err > 0);
1045 * Free shared IB device context. Decrement counter and if zero free
1046 * all allocated resources and close handles.
1049 * Pointer to mlx5_dev_ctx_shared object to free
1052 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1054 pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1055 #ifdef RTE_LIBRTE_MLX5_DEBUG
1056 /* Check the object presence in the list. */
1057 struct mlx5_dev_ctx_shared *lctx;
1059 LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1064 DRV_LOG(ERR, "Freeing non-existing shared IB context");
1069 MLX5_ASSERT(sh->refcnt);
1070 /* Secondary process should not free the shared context. */
1071 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1074 /* Remove from memory callback device list. */
1075 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1076 LIST_REMOVE(sh, mem_event_cb);
1077 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1078 /* Release created Memory Regions. */
1079 mlx5_mr_release_cache(&sh->share_cache);
1080 /* Remove context from the global device list. */
1081 LIST_REMOVE(sh, next);
1082 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1084 * Ensure there is no async event handler installed.
1085 * Only primary process handles async device events.
1087 mlx5_flow_counters_mng_close(sh);
1088 if (sh->aso_age_mng) {
1089 mlx5_flow_aso_age_mng_close(sh);
1090 sh->aso_age_mng = NULL;
1092 mlx5_flow_ipool_destroy(sh);
1093 mlx5_os_dev_shared_handler_uninstall(sh);
1094 if (sh->cnt_id_tbl) {
1095 mlx5_l3t_destroy(sh->cnt_id_tbl);
1096 sh->cnt_id_tbl = NULL;
1099 mlx5_glue->devx_free_uar(sh->tx_uar);
1103 claim_zero(mlx5_os_dealloc_pd(sh->pd));
1105 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1107 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1108 if (sh->devx_rx_uar)
1109 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1111 claim_zero(mlx5_glue->close_device(sh->ctx));
1112 pthread_mutex_destroy(&sh->txpp.mutex);
1116 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1120 * Destroy table hash list.
1123 * Pointer to the private device data structure.
1126 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1128 struct mlx5_dev_ctx_shared *sh = priv->sh;
1132 mlx5_hlist_destroy(sh->flow_tbls);
1136 * Initialize flow table hash list and create the root tables entry
1140 * Pointer to the private device data structure.
1143 * Zero on success, positive error code otherwise.
1146 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1149 /* Tables are only used in DV and DR modes. */
1150 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1151 struct mlx5_dev_ctx_shared *sh = priv->sh;
1152 char s[MLX5_HLIST_NAMESIZE];
1155 snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1156 sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1157 0, 0, flow_dv_tbl_create_cb,
1158 flow_dv_tbl_match_cb,
1159 flow_dv_tbl_remove_cb);
1160 if (!sh->flow_tbls) {
1161 DRV_LOG(ERR, "flow tables with hash creation failed.");
1165 sh->flow_tbls->ctx = sh;
1166 #ifndef HAVE_MLX5DV_DR
1167 struct rte_flow_error error;
1168 struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1171 * In case we have not DR support, the zero tables should be created
1172 * because DV expect to see them even if they cannot be created by
1175 if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0, NULL, 0, 1, &error) ||
1176 !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0, NULL, 0, 1, &error) ||
1177 !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0, NULL, 0, 1, &error)) {
1183 mlx5_free_table_hash_list(priv);
1184 #endif /* HAVE_MLX5DV_DR */
1190 * Retrieve integer value from environment variable.
1193 * Environment variable name.
1196 * Integer value, 0 if the variable is not set.
1199 mlx5_getenv_int(const char *name)
1201 const char *val = getenv(name);
1209 * DPDK callback to add udp tunnel port
1212 * A pointer to eth_dev
1213 * @param[in] udp_tunnel
1214 * A pointer to udp tunnel
1217 * 0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1220 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1221 struct rte_eth_udp_tunnel *udp_tunnel)
1223 MLX5_ASSERT(udp_tunnel != NULL);
1224 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1225 udp_tunnel->udp_port == 4789)
1227 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1228 udp_tunnel->udp_port == 4790)
1234 * Initialize process private data structure.
1237 * Pointer to Ethernet device structure.
1240 * 0 on success, a negative errno value otherwise and rte_errno is set.
1243 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1245 struct mlx5_priv *priv = dev->data->dev_private;
1246 struct mlx5_proc_priv *ppriv;
1250 * UAR register table follows the process private structure. BlueFlame
1251 * registers for Tx queues are stored in the table.
1254 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1255 ppriv = mlx5_malloc(MLX5_MEM_RTE, ppriv_size, RTE_CACHE_LINE_SIZE,
1256 dev->device->numa_node);
1261 ppriv->uar_table_sz = ppriv_size;
1262 dev->process_private = ppriv;
1267 * Un-initialize process private data structure.
1270 * Pointer to Ethernet device structure.
1273 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1275 if (!dev->process_private)
1277 mlx5_free(dev->process_private);
1278 dev->process_private = NULL;
1282 * DPDK callback to close the device.
1284 * Destroy all queues and objects, free memory.
1287 * Pointer to Ethernet device structure.
1290 mlx5_dev_close(struct rte_eth_dev *dev)
1292 struct mlx5_priv *priv = dev->data->dev_private;
1296 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1297 /* Check if process_private released. */
1298 if (!dev->process_private)
1300 mlx5_tx_uar_uninit_secondary(dev);
1301 mlx5_proc_priv_uninit(dev);
1302 rte_eth_dev_release_port(dev);
1307 DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1309 ((priv->sh->ctx != NULL) ?
1310 mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1312 * If default mreg copy action is removed at the stop stage,
1313 * the search will return none and nothing will be done anymore.
1315 mlx5_flow_stop_default(dev);
1316 mlx5_traffic_disable(dev);
1318 * If all the flows are already flushed in the device stop stage,
1319 * then this will return directly without any action.
1321 mlx5_flow_list_flush(dev, &priv->flows, true);
1322 mlx5_shared_action_flush(dev);
1323 mlx5_flow_meter_flush(dev, NULL);
1324 /* Prevent crashes when queues are still in use. */
1325 dev->rx_pkt_burst = removed_rx_burst;
1326 dev->tx_pkt_burst = removed_tx_burst;
1328 /* Disable datapath on secondary process. */
1329 mlx5_mp_os_req_stop_rxtx(dev);
1330 /* Free the eCPRI flex parser resource. */
1331 mlx5_flex_parser_ecpri_release(dev);
1332 if (priv->rxqs != NULL) {
1333 /* XXX race condition if mlx5_rx_burst() is still running. */
1334 rte_delay_us_sleep(1000);
1335 for (i = 0; (i != priv->rxqs_n); ++i)
1336 mlx5_rxq_release(dev, i);
1340 if (priv->txqs != NULL) {
1341 /* XXX race condition if mlx5_tx_burst() is still running. */
1342 rte_delay_us_sleep(1000);
1343 for (i = 0; (i != priv->txqs_n); ++i)
1344 mlx5_txq_release(dev, i);
1348 mlx5_proc_priv_uninit(dev);
1349 if (priv->drop_queue.hrxq)
1350 mlx5_drop_action_destroy(dev);
1351 if (priv->mreg_cp_tbl)
1352 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1353 mlx5_mprq_free_mp(dev);
1354 mlx5_os_free_shared_dr(priv);
1355 if (priv->rss_conf.rss_key != NULL)
1356 mlx5_free(priv->rss_conf.rss_key);
1357 if (priv->reta_idx != NULL)
1358 mlx5_free(priv->reta_idx);
1359 if (priv->config.vf)
1360 mlx5_os_mac_addr_flush(dev);
1361 if (priv->nl_socket_route >= 0)
1362 close(priv->nl_socket_route);
1363 if (priv->nl_socket_rdma >= 0)
1364 close(priv->nl_socket_rdma);
1365 if (priv->vmwa_context)
1366 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1367 ret = mlx5_hrxq_verify(dev);
1369 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1370 dev->data->port_id);
1371 ret = mlx5_ind_table_obj_verify(dev);
1373 DRV_LOG(WARNING, "port %u some indirection table still remain",
1374 dev->data->port_id);
1375 ret = mlx5_rxq_obj_verify(dev);
1377 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1378 dev->data->port_id);
1379 ret = mlx5_rxq_verify(dev);
1381 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1382 dev->data->port_id);
1383 ret = mlx5_txq_obj_verify(dev);
1385 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1386 dev->data->port_id);
1387 ret = mlx5_txq_verify(dev);
1389 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1390 dev->data->port_id);
1391 ret = mlx5_flow_verify(dev);
1393 DRV_LOG(WARNING, "port %u some flows still remain",
1394 dev->data->port_id);
1395 mlx5_cache_list_destroy(&priv->hrxqs);
1397 * Free the shared context in last turn, because the cleanup
1398 * routines above may use some shared fields, like
1399 * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1400 * ifindex if Netlink fails.
1402 mlx5_free_shared_dev_ctx(priv->sh);
1403 if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1407 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1408 struct mlx5_priv *opriv =
1409 rte_eth_devices[port_id].data->dev_private;
1412 opriv->domain_id != priv->domain_id ||
1413 &rte_eth_devices[port_id] == dev)
1419 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1421 memset(priv, 0, sizeof(*priv));
1422 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1424 * Reset mac_addrs to NULL such that it is not freed as part of
1425 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1426 * it is freed when dev_private is freed.
1428 dev->data->mac_addrs = NULL;
1433 * Verify and store value for device argument.
1436 * Key argument to verify.
1438 * Value associated with key.
1443 * 0 on success, a negative errno value otherwise and rte_errno is set.
1446 mlx5_args_check(const char *key, const char *val, void *opaque)
1448 struct mlx5_dev_config *config = opaque;
1452 /* No-op, port representors are processed in mlx5_dev_spawn(). */
1453 if (!strcmp(MLX5_REPRESENTOR, key))
1456 tmp = strtol(val, NULL, 0);
1459 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1462 if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1463 /* Negative values are acceptable for some keys only. */
1465 DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1468 mod = tmp >= 0 ? tmp : -tmp;
1469 if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1470 if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1471 DRV_LOG(ERR, "invalid CQE compression "
1472 "format parameter");
1476 config->cqe_comp = !!tmp;
1477 config->cqe_comp_fmt = tmp;
1478 } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1479 config->cqe_pad = !!tmp;
1480 } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1481 config->hw_padding = !!tmp;
1482 } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1483 config->mprq.enabled = !!tmp;
1484 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1485 config->mprq.stride_num_n = tmp;
1486 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1487 config->mprq.stride_size_n = tmp;
1488 } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1489 config->mprq.max_memcpy_len = tmp;
1490 } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1491 config->mprq.min_rxqs_num = tmp;
1492 } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1493 DRV_LOG(WARNING, "%s: deprecated parameter,"
1494 " converted to txq_inline_max", key);
1495 config->txq_inline_max = tmp;
1496 } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1497 config->txq_inline_max = tmp;
1498 } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1499 config->txq_inline_min = tmp;
1500 } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1501 config->txq_inline_mpw = tmp;
1502 } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1503 config->txqs_inline = tmp;
1504 } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1505 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1506 } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1507 config->mps = !!tmp;
1508 } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1509 if (tmp != MLX5_TXDB_CACHED &&
1510 tmp != MLX5_TXDB_NCACHED &&
1511 tmp != MLX5_TXDB_HEURISTIC) {
1512 DRV_LOG(ERR, "invalid Tx doorbell "
1513 "mapping parameter");
1518 } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1519 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1520 } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1521 DRV_LOG(WARNING, "%s: deprecated parameter,"
1522 " converted to txq_inline_mpw", key);
1523 config->txq_inline_mpw = tmp;
1524 } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1525 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1526 } else if (strcmp(MLX5_TX_PP, key) == 0) {
1528 DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1532 config->tx_pp = tmp;
1533 } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1534 config->tx_skew = tmp;
1535 } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1536 config->rx_vec_en = !!tmp;
1537 } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1538 config->l3_vxlan_en = !!tmp;
1539 } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1540 config->vf_nl_en = !!tmp;
1541 } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1542 config->dv_esw_en = !!tmp;
1543 } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1544 config->dv_flow_en = !!tmp;
1545 } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1546 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1547 tmp != MLX5_XMETA_MODE_META16 &&
1548 tmp != MLX5_XMETA_MODE_META32 &&
1549 tmp != MLX5_XMETA_MODE_MISS_INFO) {
1550 DRV_LOG(ERR, "invalid extensive "
1551 "metadata parameter");
1555 if (tmp != MLX5_XMETA_MODE_MISS_INFO)
1556 config->dv_xmeta_en = tmp;
1558 config->dv_miss_info = 1;
1559 } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1560 config->lacp_by_user = !!tmp;
1561 } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1562 config->mr_ext_memseg_en = !!tmp;
1563 } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1564 config->max_dump_files_num = tmp;
1565 } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1566 config->lro.timeout = tmp;
1567 } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1568 DRV_LOG(DEBUG, "class argument is %s.", val);
1569 } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1570 config->log_hp_size = tmp;
1571 } else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1572 if (tmp != MLX5_RCM_NONE &&
1573 tmp != MLX5_RCM_LIGHT &&
1574 tmp != MLX5_RCM_AGGR) {
1575 DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1579 config->reclaim_mode = tmp;
1580 } else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1581 config->sys_mem_en = !!tmp;
1582 } else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1583 config->decap_en = !!tmp;
1585 DRV_LOG(WARNING, "%s: unknown parameter", key);
1593 * Parse device parameters.
1596 * Pointer to device configuration structure.
1598 * Device arguments structure.
1601 * 0 on success, a negative errno value otherwise and rte_errno is set.
1604 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1606 const char **params = (const char *[]){
1607 MLX5_RXQ_CQE_COMP_EN,
1608 MLX5_RXQ_CQE_PAD_EN,
1609 MLX5_RXQ_PKT_PAD_EN,
1611 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1612 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1613 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1616 MLX5_TXQ_INLINE_MIN,
1617 MLX5_TXQ_INLINE_MAX,
1618 MLX5_TXQ_INLINE_MPW,
1619 MLX5_TXQS_MIN_INLINE,
1622 MLX5_TXQ_MPW_HDR_DSEG_EN,
1623 MLX5_TXQ_MAX_INLINE_LEN,
1635 MLX5_MR_EXT_MEMSEG_EN,
1637 MLX5_MAX_DUMP_FILES_NUM,
1638 MLX5_LRO_TIMEOUT_USEC,
1639 MLX5_CLASS_ARG_NAME,
1646 struct rte_kvargs *kvlist;
1650 if (devargs == NULL)
1652 /* Following UGLY cast is done to pass checkpatch. */
1653 kvlist = rte_kvargs_parse(devargs->args, params);
1654 if (kvlist == NULL) {
1658 /* Process parameters. */
1659 for (i = 0; (params[i] != NULL); ++i) {
1660 if (rte_kvargs_count(kvlist, params[i])) {
1661 ret = rte_kvargs_process(kvlist, params[i],
1662 mlx5_args_check, config);
1665 rte_kvargs_free(kvlist);
1670 rte_kvargs_free(kvlist);
1675 * Configures the minimal amount of data to inline into WQE
1676 * while sending packets.
1678 * - the txq_inline_min has the maximal priority, if this
1679 * key is specified in devargs
1680 * - if DevX is enabled the inline mode is queried from the
1681 * device (HCA attributes and NIC vport context if needed).
1682 * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1683 * and none (0 bytes) for other NICs
1686 * Verbs device parameters (name, port, switch_info) to spawn.
1688 * Device configuration parameters.
1691 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1692 struct mlx5_dev_config *config)
1694 if (config->txq_inline_min != MLX5_ARG_UNSET) {
1695 /* Application defines size of inlined data explicitly. */
1696 switch (spawn->pci_dev->id.device_id) {
1697 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1698 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1699 if (config->txq_inline_min <
1700 (int)MLX5_INLINE_HSIZE_L2) {
1702 "txq_inline_mix aligned to minimal"
1703 " ConnectX-4 required value %d",
1704 (int)MLX5_INLINE_HSIZE_L2);
1705 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1711 if (config->hca_attr.eth_net_offloads) {
1712 /* We have DevX enabled, inline mode queried successfully. */
1713 switch (config->hca_attr.wqe_inline_mode) {
1714 case MLX5_CAP_INLINE_MODE_L2:
1715 /* outer L2 header must be inlined. */
1716 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1718 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1719 /* No inline data are required by NIC. */
1720 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1721 config->hw_vlan_insert =
1722 config->hca_attr.wqe_vlan_insert;
1723 DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1725 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1726 /* inline mode is defined by NIC vport context. */
1727 if (!config->hca_attr.eth_virt)
1729 switch (config->hca_attr.vport_inline_mode) {
1730 case MLX5_INLINE_MODE_NONE:
1731 config->txq_inline_min =
1732 MLX5_INLINE_HSIZE_NONE;
1734 case MLX5_INLINE_MODE_L2:
1735 config->txq_inline_min =
1736 MLX5_INLINE_HSIZE_L2;
1738 case MLX5_INLINE_MODE_IP:
1739 config->txq_inline_min =
1740 MLX5_INLINE_HSIZE_L3;
1742 case MLX5_INLINE_MODE_TCP_UDP:
1743 config->txq_inline_min =
1744 MLX5_INLINE_HSIZE_L4;
1746 case MLX5_INLINE_MODE_INNER_L2:
1747 config->txq_inline_min =
1748 MLX5_INLINE_HSIZE_INNER_L2;
1750 case MLX5_INLINE_MODE_INNER_IP:
1751 config->txq_inline_min =
1752 MLX5_INLINE_HSIZE_INNER_L3;
1754 case MLX5_INLINE_MODE_INNER_TCP_UDP:
1755 config->txq_inline_min =
1756 MLX5_INLINE_HSIZE_INNER_L4;
1762 * We get here if we are unable to deduce
1763 * inline data size with DevX. Try PCI ID
1764 * to determine old NICs.
1766 switch (spawn->pci_dev->id.device_id) {
1767 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1768 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1769 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1770 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1771 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1772 config->hw_vlan_insert = 0;
1774 case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1775 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1776 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1777 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1779 * These NICs support VLAN insertion from WQE and
1780 * report the wqe_vlan_insert flag. But there is the bug
1781 * and PFC control may be broken, so disable feature.
1783 config->hw_vlan_insert = 0;
1784 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1787 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1791 DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1795 * Configures the metadata mask fields in the shared context.
1798 * Pointer to Ethernet device.
1801 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1803 struct mlx5_priv *priv = dev->data->dev_private;
1804 struct mlx5_dev_ctx_shared *sh = priv->sh;
1805 uint32_t meta, mark, reg_c0;
1807 reg_c0 = ~priv->vport_meta_mask;
1808 switch (priv->config.dv_xmeta_en) {
1809 case MLX5_XMETA_MODE_LEGACY:
1811 mark = MLX5_FLOW_MARK_MASK;
1813 case MLX5_XMETA_MODE_META16:
1814 meta = reg_c0 >> rte_bsf32(reg_c0);
1815 mark = MLX5_FLOW_MARK_MASK;
1817 case MLX5_XMETA_MODE_META32:
1819 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1827 if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1828 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1829 sh->dv_mark_mask, mark);
1831 sh->dv_mark_mask = mark;
1832 if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1833 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1834 sh->dv_meta_mask, meta);
1836 sh->dv_meta_mask = meta;
1837 if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1838 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1839 sh->dv_meta_mask, reg_c0);
1841 sh->dv_regc0_mask = reg_c0;
1842 DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1843 DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1844 DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1845 DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1849 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
1851 static const char *const dynf_names[] = {
1852 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
1853 RTE_MBUF_DYNFLAG_METADATA_NAME,
1854 RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
1858 if (n < RTE_DIM(dynf_names))
1860 for (i = 0; i < RTE_DIM(dynf_names); i++) {
1861 if (names[i] == NULL)
1863 strcpy(names[i], dynf_names[i]);
1865 return RTE_DIM(dynf_names);
1869 * Comparison callback to sort device data.
1871 * This is meant to be used with qsort().
1874 * Pointer to pointer to first data object.
1876 * Pointer to pointer to second data object.
1879 * 0 if both objects are equal, less than 0 if the first argument is less
1880 * than the second, greater than 0 otherwise.
1883 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1884 struct mlx5_dev_config *config)
1886 struct mlx5_dev_ctx_shared *sh = priv->sh;
1887 struct mlx5_dev_config *sh_conf = NULL;
1891 /* Nothing to compare for the single/first device. */
1892 if (sh->refcnt == 1)
1894 /* Find the device with shared context. */
1895 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1896 struct mlx5_priv *opriv =
1897 rte_eth_devices[port_id].data->dev_private;
1899 if (opriv && opriv != priv && opriv->sh == sh) {
1900 sh_conf = &opriv->config;
1906 if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
1907 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
1908 " for shared %s context", sh->ibdev_name);
1912 if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
1913 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
1914 " for shared %s context", sh->ibdev_name);
1922 * Look for the ethernet device belonging to mlx5 driver.
1924 * @param[in] port_id
1925 * port_id to start looking for device.
1926 * @param[in] pci_dev
1927 * Pointer to the hint PCI device. When device is being probed
1928 * the its siblings (master and preceding representors might
1929 * not have assigned driver yet (because the mlx5_os_pci_probe()
1930 * is not completed yet, for this case match on hint PCI
1931 * device may be used to detect sibling device.
1934 * port_id of found device, RTE_MAX_ETHPORT if not found.
1937 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
1939 while (port_id < RTE_MAX_ETHPORTS) {
1940 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
1942 if (dev->state != RTE_ETH_DEV_UNUSED &&
1944 (dev->device == &pci_dev->device ||
1945 (dev->device->driver &&
1946 dev->device->driver->name &&
1947 !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
1951 if (port_id >= RTE_MAX_ETHPORTS)
1952 return RTE_MAX_ETHPORTS;
1957 * DPDK callback to remove a PCI device.
1959 * This function removes all Ethernet devices belong to a given PCI device.
1961 * @param[in] pci_dev
1962 * Pointer to the PCI device.
1965 * 0 on success, the function cannot fail.
1968 mlx5_pci_remove(struct rte_pci_device *pci_dev)
1973 RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
1975 * mlx5_dev_close() is not registered to secondary process,
1976 * call the close function explicitly for secondary process.
1978 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
1979 ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
1981 ret |= rte_eth_dev_close(port_id);
1983 return ret == 0 ? 0 : -EIO;
1986 static const struct rte_pci_id mlx5_pci_id_map[] = {
1988 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1989 PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1992 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1993 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1996 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1997 PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2000 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2001 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2004 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2005 PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2008 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2009 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2012 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2013 PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2016 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2017 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2020 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2021 PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2024 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2025 PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2028 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2029 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2032 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2033 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2036 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2037 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2040 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2041 PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2044 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2045 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2048 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2049 PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2052 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2053 PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2056 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2057 PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2064 static struct mlx5_pci_driver mlx5_driver = {
2065 .driver_class = MLX5_CLASS_NET,
2068 .name = MLX5_DRIVER_NAME,
2070 .id_table = mlx5_pci_id_map,
2071 .probe = mlx5_os_pci_probe,
2072 .remove = mlx5_pci_remove,
2073 .dma_map = mlx5_dma_map,
2074 .dma_unmap = mlx5_dma_unmap,
2075 .drv_flags = PCI_DRV_FLAGS,
2079 /* Initialize driver log type. */
2080 RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
2083 * Driver initialization routine.
2085 RTE_INIT(rte_mlx5_pmd_init)
2088 /* Build the static tables for Verbs conversion. */
2089 mlx5_set_ptype_table();
2090 mlx5_set_cksum_table();
2091 mlx5_set_swp_types_table();
2093 mlx5_pci_driver_register(&mlx5_driver);
2096 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2097 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2098 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");