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 <ethdev_driver.h>
15 #include <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"
40 #include "mlx5_autoconf.h"
42 #include "mlx5_flow.h"
43 #include "mlx5_flow_os.h"
44 #include "rte_pmd_mlx5.h"
46 /* Device parameter to enable RX completion queue compression. */
47 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_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;
187 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
188 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
189 [MLX5_IPOOL_DECAP_ENCAP] = {
190 .size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
196 .malloc = mlx5_malloc,
198 .type = "mlx5_encap_decap_ipool",
200 [MLX5_IPOOL_PUSH_VLAN] = {
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",
222 [MLX5_IPOOL_PORT_ID] = {
223 .size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
229 .malloc = mlx5_malloc,
231 .type = "mlx5_port_id_ipool",
233 [MLX5_IPOOL_JUMP] = {
234 .size = sizeof(struct mlx5_flow_tbl_data_entry),
240 .malloc = mlx5_malloc,
242 .type = "mlx5_jump_ipool",
244 [MLX5_IPOOL_SAMPLE] = {
245 .size = sizeof(struct mlx5_flow_dv_sample_resource),
251 .malloc = mlx5_malloc,
253 .type = "mlx5_sample_ipool",
255 [MLX5_IPOOL_DEST_ARRAY] = {
256 .size = sizeof(struct mlx5_flow_dv_dest_array_resource),
262 .malloc = mlx5_malloc,
264 .type = "mlx5_dest_array_ipool",
266 [MLX5_IPOOL_TUNNEL_ID] = {
267 .size = sizeof(struct mlx5_flow_tunnel),
268 .trunk_size = MLX5_MAX_TUNNELS,
271 .type = "mlx5_tunnel_offload",
273 [MLX5_IPOOL_TNL_TBL_ID] = {
276 .type = "mlx5_flow_tnl_tbl_ipool",
281 * The ipool index should grow continually from small to big,
282 * for meter idx, so not set grow_trunk to avoid meter index
283 * not jump continually.
285 .size = sizeof(struct mlx5_legacy_flow_meter),
289 .malloc = mlx5_malloc,
291 .type = "mlx5_meter_ipool",
294 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
300 .malloc = mlx5_malloc,
302 .type = "mlx5_mcp_ipool",
304 [MLX5_IPOOL_HRXQ] = {
305 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
311 .malloc = mlx5_malloc,
313 .type = "mlx5_hrxq_ipool",
315 [MLX5_IPOOL_MLX5_FLOW] = {
317 * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
318 * It set in run time according to PCI function configuration.
326 .malloc = mlx5_malloc,
328 .type = "mlx5_flow_handle_ipool",
330 [MLX5_IPOOL_RTE_FLOW] = {
331 .size = sizeof(struct rte_flow),
335 .malloc = mlx5_malloc,
337 .type = "rte_flow_ipool",
339 [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
342 .type = "mlx5_flow_rss_id_ipool",
344 [MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
345 .size = sizeof(struct mlx5_shared_action_rss),
351 .malloc = mlx5_malloc,
353 .type = "mlx5_shared_action_rss",
358 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
359 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
361 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
364 * Decide whether representor ID is a HPF(host PF) port on BF2.
367 * Pointer to Ethernet device structure.
370 * Non-zero if HPF, otherwise 0.
373 mlx5_is_hpf(struct rte_eth_dev *dev)
375 struct mlx5_priv *priv = dev->data->dev_private;
376 uint16_t repr = MLX5_REPRESENTOR_REPR(priv->representor_id);
377 int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
379 return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_VF &&
380 MLX5_REPRESENTOR_REPR(-1) == repr;
384 * Initialize the ASO aging management structure.
387 * Pointer to mlx5_dev_ctx_shared object to free
390 * 0 on success, a negative errno value otherwise and rte_errno is set.
393 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
399 sh->aso_age_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->aso_age_mng),
400 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
401 if (!sh->aso_age_mng) {
402 DRV_LOG(ERR, "aso_age_mng allocation was failed.");
406 err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_FLOW_HIT);
408 mlx5_free(sh->aso_age_mng);
411 rte_spinlock_init(&sh->aso_age_mng->resize_sl);
412 rte_spinlock_init(&sh->aso_age_mng->free_sl);
413 LIST_INIT(&sh->aso_age_mng->free);
418 * Close and release all the resources of the ASO aging management structure.
421 * Pointer to mlx5_dev_ctx_shared object to free.
424 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
428 mlx5_aso_flow_hit_queue_poll_stop(sh);
429 mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_FLOW_HIT);
430 if (sh->aso_age_mng->pools) {
431 struct mlx5_aso_age_pool *pool;
433 for (i = 0; i < sh->aso_age_mng->next; ++i) {
434 pool = sh->aso_age_mng->pools[i];
435 claim_zero(mlx5_devx_cmd_destroy
436 (pool->flow_hit_aso_obj));
437 for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j)
438 if (pool->actions[j].dr_action)
440 (mlx5_flow_os_destroy_flow_action
441 (pool->actions[j].dr_action));
444 mlx5_free(sh->aso_age_mng->pools);
446 mlx5_free(sh->aso_age_mng);
450 * Initialize the shared aging list information per port.
453 * Pointer to mlx5_dev_ctx_shared object.
456 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
459 struct mlx5_age_info *age_info;
461 for (i = 0; i < sh->max_port; i++) {
462 age_info = &sh->port[i].age_info;
464 TAILQ_INIT(&age_info->aged_counters);
465 LIST_INIT(&age_info->aged_aso);
466 rte_spinlock_init(&age_info->aged_sl);
467 MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
472 * Initialize the counters management structure.
475 * Pointer to mlx5_dev_ctx_shared object to free
478 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
482 memset(&sh->cmng, 0, sizeof(sh->cmng));
483 TAILQ_INIT(&sh->cmng.flow_counters);
484 sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
485 sh->cmng.max_id = -1;
486 sh->cmng.last_pool_idx = POOL_IDX_INVALID;
487 rte_spinlock_init(&sh->cmng.pool_update_sl);
488 for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
489 TAILQ_INIT(&sh->cmng.counters[i]);
490 rte_spinlock_init(&sh->cmng.csl[i]);
495 * Destroy all the resources allocated for a counter memory management.
498 * Pointer to the memory management structure.
501 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
503 uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
505 LIST_REMOVE(mng, next);
506 claim_zero(mlx5_devx_cmd_destroy(mng->dm));
507 claim_zero(mlx5_os_umem_dereg(mng->umem));
512 * Close and release all the resources of the counters management.
515 * Pointer to mlx5_dev_ctx_shared object to free.
518 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
520 struct mlx5_counter_stats_mem_mng *mng;
526 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
527 if (rte_errno != EINPROGRESS)
532 if (sh->cmng.pools) {
533 struct mlx5_flow_counter_pool *pool;
534 uint16_t n_valid = sh->cmng.n_valid;
535 bool fallback = sh->cmng.counter_fallback;
537 for (i = 0; i < n_valid; ++i) {
538 pool = sh->cmng.pools[i];
539 if (!fallback && pool->min_dcs)
540 claim_zero(mlx5_devx_cmd_destroy
542 for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
543 struct mlx5_flow_counter *cnt =
544 MLX5_POOL_GET_CNT(pool, j);
548 (mlx5_flow_os_destroy_flow_action
550 if (fallback && MLX5_POOL_GET_CNT
551 (pool, j)->dcs_when_free)
552 claim_zero(mlx5_devx_cmd_destroy
553 (cnt->dcs_when_free));
557 mlx5_free(sh->cmng.pools);
559 mng = LIST_FIRST(&sh->cmng.mem_mngs);
561 mlx5_flow_destroy_counter_stat_mem_mng(mng);
562 mng = LIST_FIRST(&sh->cmng.mem_mngs);
564 memset(&sh->cmng, 0, sizeof(sh->cmng));
568 * Initialize the aso flow meters management structure.
571 * Pointer to mlx5_dev_ctx_shared object to free
574 mlx5_aso_flow_mtrs_mng_init(struct mlx5_priv *priv)
576 if (!priv->mtr_idx_tbl) {
577 priv->mtr_idx_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
578 if (!priv->mtr_idx_tbl) {
579 DRV_LOG(ERR, "fail to create meter lookup table.");
584 if (!priv->sh->mtrmng) {
585 priv->sh->mtrmng = mlx5_malloc(MLX5_MEM_ZERO,
586 sizeof(*priv->sh->mtrmng),
587 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
588 if (!priv->sh->mtrmng) {
589 DRV_LOG(ERR, "mlx5_aso_mtr_pools_mng allocation was failed.");
593 rte_spinlock_init(&priv->sh->mtrmng->mtrsl);
594 LIST_INIT(&priv->sh->mtrmng->meters);
600 * Close and release all the resources of
601 * the ASO flow meter management structure.
604 * Pointer to mlx5_dev_ctx_shared object to free.
607 mlx5_aso_flow_mtrs_mng_close(struct mlx5_dev_ctx_shared *sh)
609 struct mlx5_aso_mtr_pool *mtr_pool;
610 struct mlx5_aso_mtr_pools_mng *mtrmng = sh->mtrmng;
613 mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_POLICER);
614 idx = mtrmng->n_valid;
616 mtr_pool = mtrmng->pools[idx];
617 claim_zero(mlx5_devx_cmd_destroy
618 (mtr_pool->devx_obj));
622 mlx5_free(sh->mtrmng->pools);
623 mlx5_free(sh->mtrmng);
627 /* Send FLOW_AGED event if needed. */
629 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
631 struct mlx5_age_info *age_info;
634 for (i = 0; i < sh->max_port; i++) {
635 age_info = &sh->port[i].age_info;
636 if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
638 if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER))
639 rte_eth_dev_callback_process
640 (&rte_eth_devices[sh->port[i].devx_ih_port_id],
641 RTE_ETH_EVENT_FLOW_AGED, NULL);
647 * Initialize the flow resources' indexed mempool.
650 * Pointer to mlx5_dev_ctx_shared object.
652 * Pointer to user dev config.
655 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
656 const struct mlx5_dev_config *config)
659 struct mlx5_indexed_pool_config cfg;
661 for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
662 cfg = mlx5_ipool_cfg[i];
667 * Set MLX5_IPOOL_MLX5_FLOW ipool size
668 * according to PCI function flow configuration.
670 case MLX5_IPOOL_MLX5_FLOW:
671 cfg.size = config->dv_flow_en ?
672 sizeof(struct mlx5_flow_handle) :
673 MLX5_FLOW_HANDLE_VERBS_SIZE;
676 if (config->reclaim_mode)
677 cfg.release_mem_en = 1;
678 sh->ipool[i] = mlx5_ipool_create(&cfg);
683 * Release the flow resources' indexed mempool.
686 * Pointer to mlx5_dev_ctx_shared object.
689 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
693 for (i = 0; i < MLX5_IPOOL_MAX; ++i)
694 mlx5_ipool_destroy(sh->ipool[i]);
698 * Check if dynamic flex parser for eCPRI already exists.
701 * Pointer to Ethernet device structure.
704 * true on exists, false on not.
707 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
709 struct mlx5_priv *priv = dev->data->dev_private;
710 struct mlx5_flex_parser_profiles *prf =
711 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
717 * Allocation of a flex parser for eCPRI. Once created, this parser related
718 * resources will be held until the device is closed.
721 * Pointer to Ethernet device structure.
724 * 0 on success, a negative errno value otherwise and rte_errno is set.
727 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
729 struct mlx5_priv *priv = dev->data->dev_private;
730 struct mlx5_flex_parser_profiles *prf =
731 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
732 struct mlx5_devx_graph_node_attr node = {
733 .modify_field_select = 0,
738 if (!priv->config.hca_attr.parse_graph_flex_node) {
739 DRV_LOG(ERR, "Dynamic flex parser is not supported "
740 "for device %s.", priv->dev_data->name);
743 node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
744 /* 8 bytes now: 4B common header + 4B message body header. */
745 node.header_length_base_value = 0x8;
746 /* After MAC layer: Ether / VLAN. */
747 node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
748 /* Type of compared condition should be 0xAEFE in the L2 layer. */
749 node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
750 /* Sample #0: type in common header. */
751 node.sample[0].flow_match_sample_en = 1;
753 node.sample[0].flow_match_sample_offset_mode = 0x0;
754 /* Only the 2nd byte will be used. */
755 node.sample[0].flow_match_sample_field_base_offset = 0x0;
756 /* Sample #1: message payload. */
757 node.sample[1].flow_match_sample_en = 1;
759 node.sample[1].flow_match_sample_offset_mode = 0x0;
761 * Only the first two bytes will be used right now, and its offset will
762 * start after the common header that with the length of a DW(u32).
764 node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
765 prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
767 DRV_LOG(ERR, "Failed to create flex parser node object.");
768 return (rte_errno == 0) ? -ENODEV : -rte_errno;
771 ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
773 DRV_LOG(ERR, "Failed to query sample IDs.");
774 return (rte_errno == 0) ? -ENODEV : -rte_errno;
776 prf->offset[0] = 0x0;
777 prf->offset[1] = sizeof(uint32_t);
778 prf->ids[0] = ids[0];
779 prf->ids[1] = ids[1];
784 * Destroy the flex parser node, including the parser itself, input / output
785 * arcs and DW samples. Resources could be reused then.
788 * Pointer to Ethernet device structure.
791 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
793 struct mlx5_priv *priv = dev->data->dev_private;
794 struct mlx5_flex_parser_profiles *prf =
795 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
798 mlx5_devx_cmd_destroy(prf->obj);
803 * Allocate Rx and Tx UARs in robust fashion.
804 * This routine handles the following UAR allocation issues:
806 * - tries to allocate the UAR with the most appropriate memory
807 * mapping type from the ones supported by the host
809 * - tries to allocate the UAR with non-NULL base address
810 * OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
811 * UAR base address if UAR was not the first object in the UAR page.
812 * It caused the PMD failure and we should try to get another UAR
813 * till we get the first one with non-NULL base address returned.
816 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
817 const struct mlx5_dev_config *config)
819 uint32_t uar_mapping, retry;
823 for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
824 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
825 /* Control the mapping type according to the settings. */
826 uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
827 MLX5DV_UAR_ALLOC_TYPE_NC :
828 MLX5DV_UAR_ALLOC_TYPE_BF;
830 RTE_SET_USED(config);
832 * It seems we have no way to control the memory mapping type
833 * for the UAR, the default "Write-Combining" type is supposed.
834 * The UAR initialization on queue creation queries the
835 * actual mapping type done by Verbs/kernel and setups the
836 * PMD datapath accordingly.
840 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, uar_mapping);
841 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
843 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
844 if (config->dbnc == MLX5_TXDB_CACHED ||
845 config->dbnc == MLX5_TXDB_HEURISTIC)
846 DRV_LOG(WARNING, "Devarg tx_db_nc setting "
847 "is not supported by DevX");
849 * In some environments like virtual machine
850 * the Write Combining mapped might be not supported
851 * and UAR allocation fails. We try "Non-Cached"
852 * mapping for the case. The tx_burst routines take
853 * the UAR mapping type into account on UAR setup
856 DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (BF)");
857 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
858 sh->tx_uar = mlx5_glue->devx_alloc_uar
859 (sh->ctx, uar_mapping);
860 } else if (!sh->tx_uar &&
861 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
862 if (config->dbnc == MLX5_TXDB_NCACHED)
863 DRV_LOG(WARNING, "Devarg tx_db_nc settings "
864 "is not supported by DevX");
866 * If Verbs/kernel does not support "Non-Cached"
867 * try the "Write-Combining".
869 DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (NC)");
870 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
871 sh->tx_uar = mlx5_glue->devx_alloc_uar
872 (sh->ctx, uar_mapping);
876 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
880 base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
884 * The UARs are allocated by rdma_core within the
885 * IB device context, on context closure all UARs
886 * will be freed, should be no memory/object leakage.
888 DRV_LOG(DEBUG, "Retrying to allocate Tx DevX UAR");
891 /* Check whether we finally succeeded with valid UAR allocation. */
893 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
897 for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
899 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
900 (sh->ctx, uar_mapping);
901 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
902 if (!sh->devx_rx_uar &&
903 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
905 * Rx UAR is used to control interrupts only,
906 * should be no datapath noticeable impact,
907 * can try "Non-Cached" mapping safely.
909 DRV_LOG(DEBUG, "Failed to allocate Rx DevX UAR (BF)");
910 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
911 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
912 (sh->ctx, uar_mapping);
915 if (!sh->devx_rx_uar) {
916 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
920 base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
924 * The UARs are allocated by rdma_core within the
925 * IB device context, on context closure all UARs
926 * will be freed, should be no memory/object leakage.
928 DRV_LOG(DEBUG, "Retrying to allocate Rx DevX UAR");
929 sh->devx_rx_uar = NULL;
931 /* Check whether we finally succeeded with valid UAR allocation. */
932 if (!sh->devx_rx_uar) {
933 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
941 * Allocate shared device context. If there is multiport device the
942 * master and representors will share this context, if there is single
943 * port dedicated device, the context will be used by only given
944 * port due to unification.
946 * Routine first searches the context for the specified device name,
947 * if found the shared context assumed and reference counter is incremented.
948 * If no context found the new one is created and initialized with specified
949 * device context and parameters.
952 * Pointer to the device attributes (name, port, etc).
954 * Pointer to device configuration structure.
957 * Pointer to mlx5_dev_ctx_shared object on success,
958 * otherwise NULL and rte_errno is set.
960 struct mlx5_dev_ctx_shared *
961 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
962 const struct mlx5_dev_config *config)
964 struct mlx5_dev_ctx_shared *sh;
967 struct mlx5_devx_tis_attr tis_attr = { 0 };
970 /* Secondary process should not create the shared context. */
971 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
972 pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
973 /* Search for IB context by device name. */
974 LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
975 if (!strcmp(sh->ibdev_name,
976 mlx5_os_get_dev_device_name(spawn->phys_dev))) {
981 /* No device found, we have to create new shared context. */
982 MLX5_ASSERT(spawn->max_port);
983 sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
984 sizeof(struct mlx5_dev_ctx_shared) +
986 sizeof(struct mlx5_dev_shared_port),
987 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
989 DRV_LOG(ERR, "shared context allocation failure");
993 if (spawn->bond_info)
994 sh->bond = *spawn->bond_info;
995 err = mlx5_os_open_device(spawn, config, sh);
998 err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
1000 DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
1004 sh->max_port = spawn->max_port;
1005 strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
1006 sizeof(sh->ibdev_name) - 1);
1007 strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
1008 sizeof(sh->ibdev_path) - 1);
1010 * Setting port_id to max unallowed value means
1011 * there is no interrupt subhandler installed for
1012 * the given port index i.
1014 for (i = 0; i < sh->max_port; i++) {
1015 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
1016 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
1018 sh->pd = mlx5_os_alloc_pd(sh->ctx);
1019 if (sh->pd == NULL) {
1020 DRV_LOG(ERR, "PD allocation failure");
1025 err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
1027 DRV_LOG(ERR, "Fail to extract pdn from PD");
1030 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
1032 DRV_LOG(ERR, "TD allocation failure");
1036 tis_attr.transport_domain = sh->td->id;
1037 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
1039 DRV_LOG(ERR, "TIS allocation failure");
1043 err = mlx5_alloc_rxtx_uars(sh, config);
1046 MLX5_ASSERT(sh->tx_uar);
1047 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
1049 MLX5_ASSERT(sh->devx_rx_uar);
1050 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
1053 /* Initialize UAR access locks for 32bit implementations. */
1054 rte_spinlock_init(&sh->uar_lock_cq);
1055 for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1056 rte_spinlock_init(&sh->uar_lock[i]);
1059 * Once the device is added to the list of memory event
1060 * callback, its global MR cache table cannot be expanded
1061 * on the fly because of deadlock. If it overflows, lookup
1062 * should be done by searching MR list linearly, which is slow.
1064 * At this point the device is not added to the memory
1065 * event list yet, context is just being created.
1067 err = mlx5_mr_btree_init(&sh->share_cache.cache,
1068 MLX5_MR_BTREE_CACHE_N * 2,
1069 spawn->pci_dev->device.numa_node);
1074 mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
1075 &sh->share_cache.dereg_mr_cb);
1076 mlx5_os_dev_shared_handler_install(sh);
1077 sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
1078 if (!sh->cnt_id_tbl) {
1082 if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1083 err = mlx5_flow_os_init_workspace_once();
1087 mlx5_flow_aging_init(sh);
1088 mlx5_flow_counters_mng_init(sh);
1089 mlx5_flow_ipool_create(sh, config);
1090 /* Add device to memory callback list. */
1091 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1092 LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1094 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1095 /* Add context to the global device list. */
1096 LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1097 rte_spinlock_init(&sh->geneve_tlv_opt_sl);
1099 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1102 pthread_mutex_destroy(&sh->txpp.mutex);
1103 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1106 mlx5_l3t_destroy(sh->cnt_id_tbl);
1108 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1110 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1111 if (sh->devx_rx_uar)
1112 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1114 mlx5_glue->devx_free_uar(sh->tx_uar);
1116 claim_zero(mlx5_os_dealloc_pd(sh->pd));
1118 claim_zero(mlx5_glue->close_device(sh->ctx));
1120 MLX5_ASSERT(err > 0);
1126 * Free shared IB device context. Decrement counter and if zero free
1127 * all allocated resources and close handles.
1130 * Pointer to mlx5_dev_ctx_shared object to free
1133 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1135 pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1136 #ifdef RTE_LIBRTE_MLX5_DEBUG
1137 /* Check the object presence in the list. */
1138 struct mlx5_dev_ctx_shared *lctx;
1140 LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1145 DRV_LOG(ERR, "Freeing non-existing shared IB context");
1150 MLX5_ASSERT(sh->refcnt);
1151 /* Secondary process should not free the shared context. */
1152 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1155 /* Remove from memory callback device list. */
1156 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1157 LIST_REMOVE(sh, mem_event_cb);
1158 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1159 /* Release created Memory Regions. */
1160 mlx5_mr_release_cache(&sh->share_cache);
1161 /* Remove context from the global device list. */
1162 LIST_REMOVE(sh, next);
1163 /* Release flow workspaces objects on the last device. */
1164 if (LIST_EMPTY(&mlx5_dev_ctx_list))
1165 mlx5_flow_os_release_workspace();
1166 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1168 * Ensure there is no async event handler installed.
1169 * Only primary process handles async device events.
1171 mlx5_flow_counters_mng_close(sh);
1172 if (sh->aso_age_mng) {
1173 mlx5_flow_aso_age_mng_close(sh);
1174 sh->aso_age_mng = NULL;
1177 mlx5_aso_flow_mtrs_mng_close(sh);
1178 mlx5_flow_ipool_destroy(sh);
1179 mlx5_os_dev_shared_handler_uninstall(sh);
1180 if (sh->cnt_id_tbl) {
1181 mlx5_l3t_destroy(sh->cnt_id_tbl);
1182 sh->cnt_id_tbl = NULL;
1185 mlx5_glue->devx_free_uar(sh->tx_uar);
1189 claim_zero(mlx5_os_dealloc_pd(sh->pd));
1191 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1193 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1194 if (sh->devx_rx_uar)
1195 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1197 claim_zero(mlx5_glue->close_device(sh->ctx));
1198 MLX5_ASSERT(sh->geneve_tlv_option_resource == NULL);
1199 pthread_mutex_destroy(&sh->txpp.mutex);
1203 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1207 * Destroy table hash list.
1210 * Pointer to the private device data structure.
1213 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1215 struct mlx5_dev_ctx_shared *sh = priv->sh;
1219 mlx5_hlist_destroy(sh->flow_tbls);
1223 * Initialize flow table hash list and create the root tables entry
1227 * Pointer to the private device data structure.
1230 * Zero on success, positive error code otherwise.
1233 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1236 /* Tables are only used in DV and DR modes. */
1237 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1238 struct mlx5_dev_ctx_shared *sh = priv->sh;
1239 char s[MLX5_HLIST_NAMESIZE];
1242 snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1243 sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1244 0, 0, flow_dv_tbl_create_cb,
1245 flow_dv_tbl_match_cb,
1246 flow_dv_tbl_remove_cb);
1247 if (!sh->flow_tbls) {
1248 DRV_LOG(ERR, "flow tables with hash creation failed.");
1252 sh->flow_tbls->ctx = sh;
1253 #ifndef HAVE_MLX5DV_DR
1254 struct rte_flow_error error;
1255 struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1258 * In case we have not DR support, the zero tables should be created
1259 * because DV expect to see them even if they cannot be created by
1262 if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0, NULL, 0, 1, &error) ||
1263 !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0, NULL, 0, 1, &error) ||
1264 !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0, NULL, 0, 1, &error)) {
1270 mlx5_free_table_hash_list(priv);
1271 #endif /* HAVE_MLX5DV_DR */
1277 * Retrieve integer value from environment variable.
1280 * Environment variable name.
1283 * Integer value, 0 if the variable is not set.
1286 mlx5_getenv_int(const char *name)
1288 const char *val = getenv(name);
1296 * DPDK callback to add udp tunnel port
1299 * A pointer to eth_dev
1300 * @param[in] udp_tunnel
1301 * A pointer to udp tunnel
1304 * 0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1307 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1308 struct rte_eth_udp_tunnel *udp_tunnel)
1310 MLX5_ASSERT(udp_tunnel != NULL);
1311 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1312 udp_tunnel->udp_port == 4789)
1314 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1315 udp_tunnel->udp_port == 4790)
1321 * Initialize process private data structure.
1324 * Pointer to Ethernet device structure.
1327 * 0 on success, a negative errno value otherwise and rte_errno is set.
1330 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1332 struct mlx5_priv *priv = dev->data->dev_private;
1333 struct mlx5_proc_priv *ppriv;
1337 * UAR register table follows the process private structure. BlueFlame
1338 * registers for Tx queues are stored in the table.
1341 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1342 ppriv = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, ppriv_size,
1343 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1348 ppriv->uar_table_sz = priv->txqs_n;
1349 dev->process_private = ppriv;
1354 * Un-initialize process private data structure.
1357 * Pointer to Ethernet device structure.
1360 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1362 if (!dev->process_private)
1364 mlx5_free(dev->process_private);
1365 dev->process_private = NULL;
1369 * DPDK callback to close the device.
1371 * Destroy all queues and objects, free memory.
1374 * Pointer to Ethernet device structure.
1377 mlx5_dev_close(struct rte_eth_dev *dev)
1379 struct mlx5_priv *priv = dev->data->dev_private;
1383 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1384 /* Check if process_private released. */
1385 if (!dev->process_private)
1387 mlx5_tx_uar_uninit_secondary(dev);
1388 mlx5_proc_priv_uninit(dev);
1389 rte_eth_dev_release_port(dev);
1394 DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1396 ((priv->sh->ctx != NULL) ?
1397 mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1399 * If default mreg copy action is removed at the stop stage,
1400 * the search will return none and nothing will be done anymore.
1402 mlx5_flow_stop_default(dev);
1403 mlx5_traffic_disable(dev);
1405 * If all the flows are already flushed in the device stop stage,
1406 * then this will return directly without any action.
1408 mlx5_flow_list_flush(dev, &priv->flows, true);
1409 mlx5_action_handle_flush(dev);
1410 mlx5_flow_meter_flush(dev, NULL);
1411 /* Prevent crashes when queues are still in use. */
1412 dev->rx_pkt_burst = removed_rx_burst;
1413 dev->tx_pkt_burst = removed_tx_burst;
1415 /* Disable datapath on secondary process. */
1416 mlx5_mp_os_req_stop_rxtx(dev);
1417 /* Free the eCPRI flex parser resource. */
1418 mlx5_flex_parser_ecpri_release(dev);
1419 if (priv->rxqs != NULL) {
1420 /* XXX race condition if mlx5_rx_burst() is still running. */
1421 rte_delay_us_sleep(1000);
1422 for (i = 0; (i != priv->rxqs_n); ++i)
1423 mlx5_rxq_release(dev, i);
1427 if (priv->txqs != NULL) {
1428 /* XXX race condition if mlx5_tx_burst() is still running. */
1429 rte_delay_us_sleep(1000);
1430 for (i = 0; (i != priv->txqs_n); ++i)
1431 mlx5_txq_release(dev, i);
1435 mlx5_proc_priv_uninit(dev);
1436 if (priv->q_counters) {
1437 mlx5_devx_cmd_destroy(priv->q_counters);
1438 priv->q_counters = NULL;
1440 if (priv->drop_queue.hrxq)
1441 mlx5_drop_action_destroy(dev);
1442 if (priv->mreg_cp_tbl)
1443 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1444 mlx5_mprq_free_mp(dev);
1445 mlx5_os_free_shared_dr(priv);
1446 if (priv->rss_conf.rss_key != NULL)
1447 mlx5_free(priv->rss_conf.rss_key);
1448 if (priv->reta_idx != NULL)
1449 mlx5_free(priv->reta_idx);
1450 if (priv->config.vf)
1451 mlx5_os_mac_addr_flush(dev);
1452 if (priv->nl_socket_route >= 0)
1453 close(priv->nl_socket_route);
1454 if (priv->nl_socket_rdma >= 0)
1455 close(priv->nl_socket_rdma);
1456 if (priv->vmwa_context)
1457 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1458 ret = mlx5_hrxq_verify(dev);
1460 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1461 dev->data->port_id);
1462 ret = mlx5_ind_table_obj_verify(dev);
1464 DRV_LOG(WARNING, "port %u some indirection table still remain",
1465 dev->data->port_id);
1466 ret = mlx5_rxq_obj_verify(dev);
1468 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1469 dev->data->port_id);
1470 ret = mlx5_rxq_verify(dev);
1472 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1473 dev->data->port_id);
1474 ret = mlx5_txq_obj_verify(dev);
1476 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1477 dev->data->port_id);
1478 ret = mlx5_txq_verify(dev);
1480 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1481 dev->data->port_id);
1482 ret = mlx5_flow_verify(dev);
1484 DRV_LOG(WARNING, "port %u some flows still remain",
1485 dev->data->port_id);
1486 mlx5_cache_list_destroy(&priv->hrxqs);
1488 * Free the shared context in last turn, because the cleanup
1489 * routines above may use some shared fields, like
1490 * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1491 * ifindex if Netlink fails.
1493 mlx5_free_shared_dev_ctx(priv->sh);
1494 if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1498 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1499 struct mlx5_priv *opriv =
1500 rte_eth_devices[port_id].data->dev_private;
1503 opriv->domain_id != priv->domain_id ||
1504 &rte_eth_devices[port_id] == dev)
1510 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1512 memset(priv, 0, sizeof(*priv));
1513 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1515 * Reset mac_addrs to NULL such that it is not freed as part of
1516 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1517 * it is freed when dev_private is freed.
1519 dev->data->mac_addrs = NULL;
1523 const struct eth_dev_ops mlx5_dev_ops = {
1524 .dev_configure = mlx5_dev_configure,
1525 .dev_start = mlx5_dev_start,
1526 .dev_stop = mlx5_dev_stop,
1527 .dev_set_link_down = mlx5_set_link_down,
1528 .dev_set_link_up = mlx5_set_link_up,
1529 .dev_close = mlx5_dev_close,
1530 .promiscuous_enable = mlx5_promiscuous_enable,
1531 .promiscuous_disable = mlx5_promiscuous_disable,
1532 .allmulticast_enable = mlx5_allmulticast_enable,
1533 .allmulticast_disable = mlx5_allmulticast_disable,
1534 .link_update = mlx5_link_update,
1535 .stats_get = mlx5_stats_get,
1536 .stats_reset = mlx5_stats_reset,
1537 .xstats_get = mlx5_xstats_get,
1538 .xstats_reset = mlx5_xstats_reset,
1539 .xstats_get_names = mlx5_xstats_get_names,
1540 .fw_version_get = mlx5_fw_version_get,
1541 .dev_infos_get = mlx5_dev_infos_get,
1542 .representor_info_get = mlx5_representor_info_get,
1543 .read_clock = mlx5_txpp_read_clock,
1544 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1545 .vlan_filter_set = mlx5_vlan_filter_set,
1546 .rx_queue_setup = mlx5_rx_queue_setup,
1547 .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1548 .tx_queue_setup = mlx5_tx_queue_setup,
1549 .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1550 .rx_queue_release = mlx5_rx_queue_release,
1551 .tx_queue_release = mlx5_tx_queue_release,
1552 .rx_queue_start = mlx5_rx_queue_start,
1553 .rx_queue_stop = mlx5_rx_queue_stop,
1554 .tx_queue_start = mlx5_tx_queue_start,
1555 .tx_queue_stop = mlx5_tx_queue_stop,
1556 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1557 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1558 .mac_addr_remove = mlx5_mac_addr_remove,
1559 .mac_addr_add = mlx5_mac_addr_add,
1560 .mac_addr_set = mlx5_mac_addr_set,
1561 .set_mc_addr_list = mlx5_set_mc_addr_list,
1562 .mtu_set = mlx5_dev_set_mtu,
1563 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1564 .vlan_offload_set = mlx5_vlan_offload_set,
1565 .reta_update = mlx5_dev_rss_reta_update,
1566 .reta_query = mlx5_dev_rss_reta_query,
1567 .rss_hash_update = mlx5_rss_hash_update,
1568 .rss_hash_conf_get = mlx5_rss_hash_conf_get,
1569 .flow_ops_get = mlx5_flow_ops_get,
1570 .rxq_info_get = mlx5_rxq_info_get,
1571 .txq_info_get = mlx5_txq_info_get,
1572 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1573 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1574 .rx_queue_intr_enable = mlx5_rx_intr_enable,
1575 .rx_queue_intr_disable = mlx5_rx_intr_disable,
1576 .is_removed = mlx5_is_removed,
1577 .udp_tunnel_port_add = mlx5_udp_tunnel_port_add,
1578 .get_module_info = mlx5_get_module_info,
1579 .get_module_eeprom = mlx5_get_module_eeprom,
1580 .hairpin_cap_get = mlx5_hairpin_cap_get,
1581 .mtr_ops_get = mlx5_flow_meter_ops_get,
1582 .hairpin_bind = mlx5_hairpin_bind,
1583 .hairpin_unbind = mlx5_hairpin_unbind,
1584 .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1585 .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1586 .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1587 .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1590 /* Available operations from secondary process. */
1591 const struct eth_dev_ops mlx5_dev_sec_ops = {
1592 .stats_get = mlx5_stats_get,
1593 .stats_reset = mlx5_stats_reset,
1594 .xstats_get = mlx5_xstats_get,
1595 .xstats_reset = mlx5_xstats_reset,
1596 .xstats_get_names = mlx5_xstats_get_names,
1597 .fw_version_get = mlx5_fw_version_get,
1598 .dev_infos_get = mlx5_dev_infos_get,
1599 .read_clock = mlx5_txpp_read_clock,
1600 .rx_queue_start = mlx5_rx_queue_start,
1601 .rx_queue_stop = mlx5_rx_queue_stop,
1602 .tx_queue_start = mlx5_tx_queue_start,
1603 .tx_queue_stop = mlx5_tx_queue_stop,
1604 .rxq_info_get = mlx5_rxq_info_get,
1605 .txq_info_get = mlx5_txq_info_get,
1606 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1607 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1608 .get_module_info = mlx5_get_module_info,
1609 .get_module_eeprom = mlx5_get_module_eeprom,
1612 /* Available operations in flow isolated mode. */
1613 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1614 .dev_configure = mlx5_dev_configure,
1615 .dev_start = mlx5_dev_start,
1616 .dev_stop = mlx5_dev_stop,
1617 .dev_set_link_down = mlx5_set_link_down,
1618 .dev_set_link_up = mlx5_set_link_up,
1619 .dev_close = mlx5_dev_close,
1620 .promiscuous_enable = mlx5_promiscuous_enable,
1621 .promiscuous_disable = mlx5_promiscuous_disable,
1622 .allmulticast_enable = mlx5_allmulticast_enable,
1623 .allmulticast_disable = mlx5_allmulticast_disable,
1624 .link_update = mlx5_link_update,
1625 .stats_get = mlx5_stats_get,
1626 .stats_reset = mlx5_stats_reset,
1627 .xstats_get = mlx5_xstats_get,
1628 .xstats_reset = mlx5_xstats_reset,
1629 .xstats_get_names = mlx5_xstats_get_names,
1630 .fw_version_get = mlx5_fw_version_get,
1631 .dev_infos_get = mlx5_dev_infos_get,
1632 .read_clock = mlx5_txpp_read_clock,
1633 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1634 .vlan_filter_set = mlx5_vlan_filter_set,
1635 .rx_queue_setup = mlx5_rx_queue_setup,
1636 .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1637 .tx_queue_setup = mlx5_tx_queue_setup,
1638 .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1639 .rx_queue_release = mlx5_rx_queue_release,
1640 .tx_queue_release = mlx5_tx_queue_release,
1641 .rx_queue_start = mlx5_rx_queue_start,
1642 .rx_queue_stop = mlx5_rx_queue_stop,
1643 .tx_queue_start = mlx5_tx_queue_start,
1644 .tx_queue_stop = mlx5_tx_queue_stop,
1645 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1646 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1647 .mac_addr_remove = mlx5_mac_addr_remove,
1648 .mac_addr_add = mlx5_mac_addr_add,
1649 .mac_addr_set = mlx5_mac_addr_set,
1650 .set_mc_addr_list = mlx5_set_mc_addr_list,
1651 .mtu_set = mlx5_dev_set_mtu,
1652 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1653 .vlan_offload_set = mlx5_vlan_offload_set,
1654 .flow_ops_get = mlx5_flow_ops_get,
1655 .rxq_info_get = mlx5_rxq_info_get,
1656 .txq_info_get = mlx5_txq_info_get,
1657 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1658 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1659 .rx_queue_intr_enable = mlx5_rx_intr_enable,
1660 .rx_queue_intr_disable = mlx5_rx_intr_disable,
1661 .is_removed = mlx5_is_removed,
1662 .get_module_info = mlx5_get_module_info,
1663 .get_module_eeprom = mlx5_get_module_eeprom,
1664 .hairpin_cap_get = mlx5_hairpin_cap_get,
1665 .mtr_ops_get = mlx5_flow_meter_ops_get,
1666 .hairpin_bind = mlx5_hairpin_bind,
1667 .hairpin_unbind = mlx5_hairpin_unbind,
1668 .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1669 .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1670 .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1671 .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1675 * Verify and store value for device argument.
1678 * Key argument to verify.
1680 * Value associated with key.
1685 * 0 on success, a negative errno value otherwise and rte_errno is set.
1688 mlx5_args_check(const char *key, const char *val, void *opaque)
1690 struct mlx5_dev_config *config = opaque;
1694 /* No-op, port representors are processed in mlx5_dev_spawn(). */
1695 if (!strcmp(MLX5_REPRESENTOR, key))
1698 tmp = strtol(val, NULL, 0);
1701 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1704 if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1705 /* Negative values are acceptable for some keys only. */
1707 DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1710 mod = tmp >= 0 ? tmp : -tmp;
1711 if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1712 if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1713 DRV_LOG(ERR, "invalid CQE compression "
1714 "format parameter");
1718 config->cqe_comp = !!tmp;
1719 config->cqe_comp_fmt = tmp;
1720 } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1721 config->hw_padding = !!tmp;
1722 } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1723 config->mprq.enabled = !!tmp;
1724 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1725 config->mprq.stride_num_n = tmp;
1726 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1727 config->mprq.stride_size_n = tmp;
1728 } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1729 config->mprq.max_memcpy_len = tmp;
1730 } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1731 config->mprq.min_rxqs_num = tmp;
1732 } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1733 DRV_LOG(WARNING, "%s: deprecated parameter,"
1734 " converted to txq_inline_max", key);
1735 config->txq_inline_max = tmp;
1736 } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1737 config->txq_inline_max = tmp;
1738 } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1739 config->txq_inline_min = tmp;
1740 } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1741 config->txq_inline_mpw = tmp;
1742 } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1743 config->txqs_inline = tmp;
1744 } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1745 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1746 } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1747 config->mps = !!tmp;
1748 } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1749 if (tmp != MLX5_TXDB_CACHED &&
1750 tmp != MLX5_TXDB_NCACHED &&
1751 tmp != MLX5_TXDB_HEURISTIC) {
1752 DRV_LOG(ERR, "invalid Tx doorbell "
1753 "mapping parameter");
1758 } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1759 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1760 } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1761 DRV_LOG(WARNING, "%s: deprecated parameter,"
1762 " converted to txq_inline_mpw", key);
1763 config->txq_inline_mpw = tmp;
1764 } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1765 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1766 } else if (strcmp(MLX5_TX_PP, key) == 0) {
1768 DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1772 config->tx_pp = tmp;
1773 } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1774 config->tx_skew = tmp;
1775 } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1776 config->rx_vec_en = !!tmp;
1777 } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1778 config->l3_vxlan_en = !!tmp;
1779 } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1780 config->vf_nl_en = !!tmp;
1781 } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1782 config->dv_esw_en = !!tmp;
1783 } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1784 config->dv_flow_en = !!tmp;
1785 } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1786 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1787 tmp != MLX5_XMETA_MODE_META16 &&
1788 tmp != MLX5_XMETA_MODE_META32 &&
1789 tmp != MLX5_XMETA_MODE_MISS_INFO) {
1790 DRV_LOG(ERR, "invalid extensive "
1791 "metadata parameter");
1795 if (tmp != MLX5_XMETA_MODE_MISS_INFO)
1796 config->dv_xmeta_en = tmp;
1798 config->dv_miss_info = 1;
1799 } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1800 config->lacp_by_user = !!tmp;
1801 } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1802 config->mr_ext_memseg_en = !!tmp;
1803 } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1804 config->max_dump_files_num = tmp;
1805 } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1806 config->lro.timeout = tmp;
1807 } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1808 DRV_LOG(DEBUG, "class argument is %s.", val);
1809 } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1810 config->log_hp_size = tmp;
1811 } else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1812 if (tmp != MLX5_RCM_NONE &&
1813 tmp != MLX5_RCM_LIGHT &&
1814 tmp != MLX5_RCM_AGGR) {
1815 DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1819 config->reclaim_mode = tmp;
1820 } else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1821 config->sys_mem_en = !!tmp;
1822 } else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1823 config->decap_en = !!tmp;
1825 DRV_LOG(WARNING, "%s: unknown parameter", key);
1833 * Parse device parameters.
1836 * Pointer to device configuration structure.
1838 * Device arguments structure.
1841 * 0 on success, a negative errno value otherwise and rte_errno is set.
1844 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1846 const char **params = (const char *[]){
1847 MLX5_RXQ_CQE_COMP_EN,
1848 MLX5_RXQ_PKT_PAD_EN,
1850 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1851 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1852 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1855 MLX5_TXQ_INLINE_MIN,
1856 MLX5_TXQ_INLINE_MAX,
1857 MLX5_TXQ_INLINE_MPW,
1858 MLX5_TXQS_MIN_INLINE,
1861 MLX5_TXQ_MPW_HDR_DSEG_EN,
1862 MLX5_TXQ_MAX_INLINE_LEN,
1874 MLX5_MR_EXT_MEMSEG_EN,
1876 MLX5_MAX_DUMP_FILES_NUM,
1877 MLX5_LRO_TIMEOUT_USEC,
1878 MLX5_CLASS_ARG_NAME,
1885 struct rte_kvargs *kvlist;
1889 if (devargs == NULL)
1891 /* Following UGLY cast is done to pass checkpatch. */
1892 kvlist = rte_kvargs_parse(devargs->args, params);
1893 if (kvlist == NULL) {
1897 /* Process parameters. */
1898 for (i = 0; (params[i] != NULL); ++i) {
1899 if (rte_kvargs_count(kvlist, params[i])) {
1900 ret = rte_kvargs_process(kvlist, params[i],
1901 mlx5_args_check, config);
1904 rte_kvargs_free(kvlist);
1909 rte_kvargs_free(kvlist);
1914 * Configures the minimal amount of data to inline into WQE
1915 * while sending packets.
1917 * - the txq_inline_min has the maximal priority, if this
1918 * key is specified in devargs
1919 * - if DevX is enabled the inline mode is queried from the
1920 * device (HCA attributes and NIC vport context if needed).
1921 * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1922 * and none (0 bytes) for other NICs
1925 * Verbs device parameters (name, port, switch_info) to spawn.
1927 * Device configuration parameters.
1930 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1931 struct mlx5_dev_config *config)
1933 if (config->txq_inline_min != MLX5_ARG_UNSET) {
1934 /* Application defines size of inlined data explicitly. */
1935 switch (spawn->pci_dev->id.device_id) {
1936 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1937 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1938 if (config->txq_inline_min <
1939 (int)MLX5_INLINE_HSIZE_L2) {
1941 "txq_inline_mix aligned to minimal"
1942 " ConnectX-4 required value %d",
1943 (int)MLX5_INLINE_HSIZE_L2);
1944 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1950 if (config->hca_attr.eth_net_offloads) {
1951 /* We have DevX enabled, inline mode queried successfully. */
1952 switch (config->hca_attr.wqe_inline_mode) {
1953 case MLX5_CAP_INLINE_MODE_L2:
1954 /* outer L2 header must be inlined. */
1955 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1957 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1958 /* No inline data are required by NIC. */
1959 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1960 config->hw_vlan_insert =
1961 config->hca_attr.wqe_vlan_insert;
1962 DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1964 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1965 /* inline mode is defined by NIC vport context. */
1966 if (!config->hca_attr.eth_virt)
1968 switch (config->hca_attr.vport_inline_mode) {
1969 case MLX5_INLINE_MODE_NONE:
1970 config->txq_inline_min =
1971 MLX5_INLINE_HSIZE_NONE;
1973 case MLX5_INLINE_MODE_L2:
1974 config->txq_inline_min =
1975 MLX5_INLINE_HSIZE_L2;
1977 case MLX5_INLINE_MODE_IP:
1978 config->txq_inline_min =
1979 MLX5_INLINE_HSIZE_L3;
1981 case MLX5_INLINE_MODE_TCP_UDP:
1982 config->txq_inline_min =
1983 MLX5_INLINE_HSIZE_L4;
1985 case MLX5_INLINE_MODE_INNER_L2:
1986 config->txq_inline_min =
1987 MLX5_INLINE_HSIZE_INNER_L2;
1989 case MLX5_INLINE_MODE_INNER_IP:
1990 config->txq_inline_min =
1991 MLX5_INLINE_HSIZE_INNER_L3;
1993 case MLX5_INLINE_MODE_INNER_TCP_UDP:
1994 config->txq_inline_min =
1995 MLX5_INLINE_HSIZE_INNER_L4;
2001 * We get here if we are unable to deduce
2002 * inline data size with DevX. Try PCI ID
2003 * to determine old NICs.
2005 switch (spawn->pci_dev->id.device_id) {
2006 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2007 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2008 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2009 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2010 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2011 config->hw_vlan_insert = 0;
2013 case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2014 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2015 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2016 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2018 * These NICs support VLAN insertion from WQE and
2019 * report the wqe_vlan_insert flag. But there is the bug
2020 * and PFC control may be broken, so disable feature.
2022 config->hw_vlan_insert = 0;
2023 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2026 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2030 DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2034 * Configures the metadata mask fields in the shared context.
2037 * Pointer to Ethernet device.
2040 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2042 struct mlx5_priv *priv = dev->data->dev_private;
2043 struct mlx5_dev_ctx_shared *sh = priv->sh;
2044 uint32_t meta, mark, reg_c0;
2046 reg_c0 = ~priv->vport_meta_mask;
2047 switch (priv->config.dv_xmeta_en) {
2048 case MLX5_XMETA_MODE_LEGACY:
2050 mark = MLX5_FLOW_MARK_MASK;
2052 case MLX5_XMETA_MODE_META16:
2053 meta = reg_c0 >> rte_bsf32(reg_c0);
2054 mark = MLX5_FLOW_MARK_MASK;
2056 case MLX5_XMETA_MODE_META32:
2058 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2066 if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2067 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2068 sh->dv_mark_mask, mark);
2070 sh->dv_mark_mask = mark;
2071 if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2072 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2073 sh->dv_meta_mask, meta);
2075 sh->dv_meta_mask = meta;
2076 if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2077 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2078 sh->dv_meta_mask, reg_c0);
2080 sh->dv_regc0_mask = reg_c0;
2081 DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2082 DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2083 DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2084 DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2088 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2090 static const char *const dynf_names[] = {
2091 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2092 RTE_MBUF_DYNFLAG_METADATA_NAME,
2093 RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
2097 if (n < RTE_DIM(dynf_names))
2099 for (i = 0; i < RTE_DIM(dynf_names); i++) {
2100 if (names[i] == NULL)
2102 strcpy(names[i], dynf_names[i]);
2104 return RTE_DIM(dynf_names);
2108 * Comparison callback to sort device data.
2110 * This is meant to be used with qsort().
2113 * Pointer to pointer to first data object.
2115 * Pointer to pointer to second data object.
2118 * 0 if both objects are equal, less than 0 if the first argument is less
2119 * than the second, greater than 0 otherwise.
2122 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2123 struct mlx5_dev_config *config)
2125 struct mlx5_dev_ctx_shared *sh = priv->sh;
2126 struct mlx5_dev_config *sh_conf = NULL;
2130 /* Nothing to compare for the single/first device. */
2131 if (sh->refcnt == 1)
2133 /* Find the device with shared context. */
2134 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2135 struct mlx5_priv *opriv =
2136 rte_eth_devices[port_id].data->dev_private;
2138 if (opriv && opriv != priv && opriv->sh == sh) {
2139 sh_conf = &opriv->config;
2145 if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2146 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2147 " for shared %s context", sh->ibdev_name);
2151 if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2152 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2153 " for shared %s context", sh->ibdev_name);
2161 * Look for the ethernet device belonging to mlx5 driver.
2163 * @param[in] port_id
2164 * port_id to start looking for device.
2165 * @param[in] pci_dev
2166 * Pointer to the hint PCI device. When device is being probed
2167 * the its siblings (master and preceding representors might
2168 * not have assigned driver yet (because the mlx5_os_pci_probe()
2169 * is not completed yet, for this case match on hint PCI
2170 * device may be used to detect sibling device.
2173 * port_id of found device, RTE_MAX_ETHPORT if not found.
2176 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
2178 while (port_id < RTE_MAX_ETHPORTS) {
2179 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2181 if (dev->state != RTE_ETH_DEV_UNUSED &&
2183 (dev->device == &pci_dev->device ||
2184 (dev->device->driver &&
2185 dev->device->driver->name &&
2186 !strcmp(dev->device->driver->name, MLX5_PCI_DRIVER_NAME))))
2190 if (port_id >= RTE_MAX_ETHPORTS)
2191 return RTE_MAX_ETHPORTS;
2196 * DPDK callback to remove a PCI device.
2198 * This function removes all Ethernet devices belong to a given PCI device.
2200 * @param[in] pci_dev
2201 * Pointer to the PCI device.
2204 * 0 on success, the function cannot fail.
2207 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2212 RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
2214 * mlx5_dev_close() is not registered to secondary process,
2215 * call the close function explicitly for secondary process.
2217 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2218 ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2220 ret |= rte_eth_dev_close(port_id);
2222 return ret == 0 ? 0 : -EIO;
2225 static const struct rte_pci_id mlx5_pci_id_map[] = {
2227 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2228 PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2231 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2232 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2235 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2236 PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2239 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2240 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2243 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2244 PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2247 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2248 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2251 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2252 PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2255 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2256 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2259 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2260 PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2263 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2264 PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2267 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2268 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2271 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2272 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2275 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2276 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2279 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2280 PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2283 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2284 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2287 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2288 PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2291 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2292 PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2295 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2296 PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2303 static struct mlx5_pci_driver mlx5_driver = {
2304 .driver_class = MLX5_CLASS_NET,
2307 .name = MLX5_PCI_DRIVER_NAME,
2309 .id_table = mlx5_pci_id_map,
2310 .probe = mlx5_os_pci_probe,
2311 .remove = mlx5_pci_remove,
2312 .dma_map = mlx5_dma_map,
2313 .dma_unmap = mlx5_dma_unmap,
2314 .drv_flags = PCI_DRV_FLAGS,
2318 /* Initialize driver log type. */
2319 RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
2322 * Driver initialization routine.
2324 RTE_INIT(rte_mlx5_pmd_init)
2326 pthread_mutex_init(&mlx5_dev_ctx_list_mutex, NULL);
2328 /* Build the static tables for Verbs conversion. */
2329 mlx5_set_ptype_table();
2330 mlx5_set_cksum_table();
2331 mlx5_set_swp_types_table();
2333 mlx5_pci_driver_register(&mlx5_driver);
2336 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2337 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2338 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");