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>
16 #include <rte_bus_pci.h>
17 #include <rte_common.h>
18 #include <rte_kvargs.h>
19 #include <rte_rwlock.h>
20 #include <rte_spinlock.h>
21 #include <rte_string_fns.h>
22 #include <rte_alarm.h>
23 #include <rte_cycles.h>
25 #include <mlx5_glue.h>
26 #include <mlx5_devx_cmds.h>
27 #include <mlx5_common.h>
28 #include <mlx5_common_os.h>
29 #include <mlx5_common_mp.h>
30 #include <mlx5_malloc.h>
32 #include "mlx5_defs.h"
34 #include "mlx5_utils.h"
35 #include "mlx5_rxtx.h"
38 #include "mlx5_autoconf.h"
40 #include "mlx5_flow.h"
41 #include "mlx5_flow_os.h"
42 #include "rte_pmd_mlx5.h"
44 #define MLX5_ETH_DRIVER_NAME mlx5_eth
46 /* Driver type key for new device global syntax. */
47 #define MLX5_DRIVER_KEY "driver"
49 /* Device parameter to enable RX completion queue compression. */
50 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
52 /* Device parameter to enable padding Rx packet to cacheline size. */
53 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
55 /* Device parameter to enable Multi-Packet Rx queue. */
56 #define MLX5_RX_MPRQ_EN "mprq_en"
58 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
59 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
61 /* Device parameter to configure log 2 of the stride size for MPRQ. */
62 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
64 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
65 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
67 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
68 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
70 /* Device parameter to configure inline send. Deprecated, ignored.*/
71 #define MLX5_TXQ_INLINE "txq_inline"
73 /* Device parameter to limit packet size to inline with ordinary SEND. */
74 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
76 /* Device parameter to configure minimal data size to inline. */
77 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
79 /* Device parameter to limit packet size to inline with Enhanced MPW. */
80 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
83 * Device parameter to configure the number of TX queues threshold for
84 * enabling inline send.
86 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
89 * Device parameter to configure the number of TX queues threshold for
90 * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
92 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
94 /* Device parameter to enable multi-packet send WQEs. */
95 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
98 * Device parameter to force doorbell register mapping
99 * to non-cahed region eliminating the extra write memory barrier.
101 #define MLX5_TX_DB_NC "tx_db_nc"
104 * Device parameter to include 2 dsegs in the title WQEBB.
105 * Deprecated, ignored.
107 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
110 * Device parameter to limit the size of inlining packet.
111 * Deprecated, ignored.
113 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
116 * Device parameter to enable Tx scheduling on timestamps
117 * and specify the packet pacing granularity in nanoseconds.
119 #define MLX5_TX_PP "tx_pp"
122 * Device parameter to specify skew in nanoseconds on Tx datapath,
123 * it represents the time between SQ start WQE processing and
124 * appearing actual packet data on the wire.
126 #define MLX5_TX_SKEW "tx_skew"
129 * Device parameter to enable hardware Tx vector.
130 * Deprecated, ignored (no vectorized Tx routines anymore).
132 #define MLX5_TX_VEC_EN "tx_vec_en"
134 /* Device parameter to enable hardware Rx vector. */
135 #define MLX5_RX_VEC_EN "rx_vec_en"
137 /* Allow L3 VXLAN flow creation. */
138 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
140 /* Activate DV E-Switch flow steering. */
141 #define MLX5_DV_ESW_EN "dv_esw_en"
143 /* Activate DV flow steering. */
144 #define MLX5_DV_FLOW_EN "dv_flow_en"
146 /* Enable extensive flow metadata support. */
147 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
149 /* Device parameter to let the user manage the lacp traffic of bonded device */
150 #define MLX5_LACP_BY_USER "lacp_by_user"
152 /* Activate Netlink support in VF mode. */
153 #define MLX5_VF_NL_EN "vf_nl_en"
155 /* Enable extending memsegs when creating a MR. */
156 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
158 /* Select port representors to instantiate. */
159 #define MLX5_REPRESENTOR "representor"
161 /* Device parameter to configure the maximum number of dump files per queue. */
162 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
164 /* Configure timeout of LRO session (in microseconds). */
165 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
168 * Device parameter to configure the total data buffer size for a single
169 * hairpin queue (logarithm value).
171 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
173 /* Flow memory reclaim mode. */
174 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
176 /* The default memory allocator used in PMD. */
177 #define MLX5_SYS_MEM_EN "sys_mem_en"
178 /* Decap will be used or not. */
179 #define MLX5_DECAP_EN "decap_en"
181 /* Device parameter to configure allow or prevent duplicate rules pattern. */
182 #define MLX5_ALLOW_DUPLICATE_PATTERN "allow_duplicate_pattern"
184 /* Device parameter to configure implicit registration of mempool memory. */
185 #define MLX5_MR_MEMPOOL_REG_EN "mr_mempool_reg_en"
187 /* Shared memory between primary and secondary processes. */
188 struct mlx5_shared_data *mlx5_shared_data;
190 /** Driver-specific log messages type. */
193 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
194 LIST_HEAD_INITIALIZER();
195 static pthread_mutex_t mlx5_dev_ctx_list_mutex;
196 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
197 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
198 [MLX5_IPOOL_DECAP_ENCAP] = {
199 .size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
205 .malloc = mlx5_malloc,
207 .type = "mlx5_encap_decap_ipool",
209 [MLX5_IPOOL_PUSH_VLAN] = {
210 .size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
216 .malloc = mlx5_malloc,
218 .type = "mlx5_push_vlan_ipool",
221 .size = sizeof(struct mlx5_flow_dv_tag_resource),
227 .per_core_cache = (1 << 16),
228 .malloc = mlx5_malloc,
230 .type = "mlx5_tag_ipool",
232 [MLX5_IPOOL_PORT_ID] = {
233 .size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
239 .malloc = mlx5_malloc,
241 .type = "mlx5_port_id_ipool",
243 [MLX5_IPOOL_JUMP] = {
244 .size = sizeof(struct mlx5_flow_tbl_data_entry),
250 .malloc = mlx5_malloc,
252 .type = "mlx5_jump_ipool",
254 [MLX5_IPOOL_SAMPLE] = {
255 .size = sizeof(struct mlx5_flow_dv_sample_resource),
261 .malloc = mlx5_malloc,
263 .type = "mlx5_sample_ipool",
265 [MLX5_IPOOL_DEST_ARRAY] = {
266 .size = sizeof(struct mlx5_flow_dv_dest_array_resource),
272 .malloc = mlx5_malloc,
274 .type = "mlx5_dest_array_ipool",
276 [MLX5_IPOOL_TUNNEL_ID] = {
277 .size = sizeof(struct mlx5_flow_tunnel),
278 .trunk_size = MLX5_MAX_TUNNELS,
281 .type = "mlx5_tunnel_offload",
283 [MLX5_IPOOL_TNL_TBL_ID] = {
286 .type = "mlx5_flow_tnl_tbl_ipool",
291 * The ipool index should grow continually from small to big,
292 * for meter idx, so not set grow_trunk to avoid meter index
293 * not jump continually.
295 .size = sizeof(struct mlx5_legacy_flow_meter),
299 .malloc = mlx5_malloc,
301 .type = "mlx5_meter_ipool",
304 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
310 .malloc = mlx5_malloc,
312 .type = "mlx5_mcp_ipool",
314 [MLX5_IPOOL_HRXQ] = {
315 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
321 .malloc = mlx5_malloc,
323 .type = "mlx5_hrxq_ipool",
325 [MLX5_IPOOL_MLX5_FLOW] = {
327 * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
328 * It set in run time according to PCI function configuration.
336 .per_core_cache = 1 << 19,
337 .malloc = mlx5_malloc,
339 .type = "mlx5_flow_handle_ipool",
341 [MLX5_IPOOL_RTE_FLOW] = {
342 .size = sizeof(struct rte_flow),
346 .malloc = mlx5_malloc,
348 .type = "rte_flow_ipool",
350 [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
353 .type = "mlx5_flow_rss_id_ipool",
355 [MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
356 .size = sizeof(struct mlx5_shared_action_rss),
362 .malloc = mlx5_malloc,
364 .type = "mlx5_shared_action_rss",
366 [MLX5_IPOOL_MTR_POLICY] = {
368 * The ipool index should grow continually from small to big,
369 * for policy idx, so not set grow_trunk to avoid policy index
370 * not jump continually.
372 .size = sizeof(struct mlx5_flow_meter_sub_policy),
376 .malloc = mlx5_malloc,
378 .type = "mlx5_meter_policy_ipool",
383 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
384 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
386 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 1024
389 * Decide whether representor ID is a HPF(host PF) port on BF2.
392 * Pointer to Ethernet device structure.
395 * Non-zero if HPF, otherwise 0.
398 mlx5_is_hpf(struct rte_eth_dev *dev)
400 struct mlx5_priv *priv = dev->data->dev_private;
401 uint16_t repr = MLX5_REPRESENTOR_REPR(priv->representor_id);
402 int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
404 return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_VF &&
405 MLX5_REPRESENTOR_REPR(-1) == repr;
409 * Decide whether representor ID is a SF port representor.
412 * Pointer to Ethernet device structure.
415 * Non-zero if HPF, otherwise 0.
418 mlx5_is_sf_repr(struct rte_eth_dev *dev)
420 struct mlx5_priv *priv = dev->data->dev_private;
421 int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
423 return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_SF;
427 * Initialize the ASO aging management structure.
430 * Pointer to mlx5_dev_ctx_shared object to free
433 * 0 on success, a negative errno value otherwise and rte_errno is set.
436 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
442 sh->aso_age_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->aso_age_mng),
443 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
444 if (!sh->aso_age_mng) {
445 DRV_LOG(ERR, "aso_age_mng allocation was failed.");
449 err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_FLOW_HIT);
451 mlx5_free(sh->aso_age_mng);
454 rte_spinlock_init(&sh->aso_age_mng->resize_sl);
455 rte_spinlock_init(&sh->aso_age_mng->free_sl);
456 LIST_INIT(&sh->aso_age_mng->free);
461 * Close and release all the resources of the ASO aging management structure.
464 * Pointer to mlx5_dev_ctx_shared object to free.
467 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
471 mlx5_aso_flow_hit_queue_poll_stop(sh);
472 mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_FLOW_HIT);
473 if (sh->aso_age_mng->pools) {
474 struct mlx5_aso_age_pool *pool;
476 for (i = 0; i < sh->aso_age_mng->next; ++i) {
477 pool = sh->aso_age_mng->pools[i];
478 claim_zero(mlx5_devx_cmd_destroy
479 (pool->flow_hit_aso_obj));
480 for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j)
481 if (pool->actions[j].dr_action)
483 (mlx5_flow_os_destroy_flow_action
484 (pool->actions[j].dr_action));
487 mlx5_free(sh->aso_age_mng->pools);
489 mlx5_free(sh->aso_age_mng);
493 * Initialize the shared aging list information per port.
496 * Pointer to mlx5_dev_ctx_shared object.
499 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
502 struct mlx5_age_info *age_info;
504 for (i = 0; i < sh->max_port; i++) {
505 age_info = &sh->port[i].age_info;
507 TAILQ_INIT(&age_info->aged_counters);
508 LIST_INIT(&age_info->aged_aso);
509 rte_spinlock_init(&age_info->aged_sl);
510 MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
515 * Initialize the counters management structure.
518 * Pointer to mlx5_dev_ctx_shared object to free
521 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
525 memset(&sh->cmng, 0, sizeof(sh->cmng));
526 TAILQ_INIT(&sh->cmng.flow_counters);
527 sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
528 sh->cmng.max_id = -1;
529 sh->cmng.last_pool_idx = POOL_IDX_INVALID;
530 rte_spinlock_init(&sh->cmng.pool_update_sl);
531 for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
532 TAILQ_INIT(&sh->cmng.counters[i]);
533 rte_spinlock_init(&sh->cmng.csl[i]);
538 * Destroy all the resources allocated for a counter memory management.
541 * Pointer to the memory management structure.
544 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
546 uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
548 LIST_REMOVE(mng, next);
549 claim_zero(mlx5_devx_cmd_destroy(mng->dm));
550 claim_zero(mlx5_os_umem_dereg(mng->umem));
555 * Close and release all the resources of the counters management.
558 * Pointer to mlx5_dev_ctx_shared object to free.
561 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
563 struct mlx5_counter_stats_mem_mng *mng;
569 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
570 if (rte_errno != EINPROGRESS)
575 if (sh->cmng.pools) {
576 struct mlx5_flow_counter_pool *pool;
577 uint16_t n_valid = sh->cmng.n_valid;
578 bool fallback = sh->cmng.counter_fallback;
580 for (i = 0; i < n_valid; ++i) {
581 pool = sh->cmng.pools[i];
582 if (!fallback && pool->min_dcs)
583 claim_zero(mlx5_devx_cmd_destroy
585 for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
586 struct mlx5_flow_counter *cnt =
587 MLX5_POOL_GET_CNT(pool, j);
591 (mlx5_flow_os_destroy_flow_action
593 if (fallback && MLX5_POOL_GET_CNT
594 (pool, j)->dcs_when_free)
595 claim_zero(mlx5_devx_cmd_destroy
596 (cnt->dcs_when_free));
600 mlx5_free(sh->cmng.pools);
602 mng = LIST_FIRST(&sh->cmng.mem_mngs);
604 mlx5_flow_destroy_counter_stat_mem_mng(mng);
605 mng = LIST_FIRST(&sh->cmng.mem_mngs);
607 memset(&sh->cmng, 0, sizeof(sh->cmng));
611 * Initialize the aso flow meters management structure.
614 * Pointer to mlx5_dev_ctx_shared object to free
617 mlx5_aso_flow_mtrs_mng_init(struct mlx5_dev_ctx_shared *sh)
620 sh->mtrmng = mlx5_malloc(MLX5_MEM_ZERO,
622 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
625 "meter management allocation was failed.");
629 if (sh->meter_aso_en) {
630 rte_spinlock_init(&sh->mtrmng->pools_mng.mtrsl);
631 LIST_INIT(&sh->mtrmng->pools_mng.meters);
633 sh->mtrmng->def_policy_id = MLX5_INVALID_POLICY_ID;
639 * Close and release all the resources of
640 * the ASO flow meter management structure.
643 * Pointer to mlx5_dev_ctx_shared object to free.
646 mlx5_aso_flow_mtrs_mng_close(struct mlx5_dev_ctx_shared *sh)
648 struct mlx5_aso_mtr_pool *mtr_pool;
649 struct mlx5_flow_mtr_mng *mtrmng = sh->mtrmng;
651 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
652 struct mlx5_aso_mtr *aso_mtr;
654 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
656 if (sh->meter_aso_en) {
657 mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_POLICER);
658 idx = mtrmng->pools_mng.n_valid;
660 mtr_pool = mtrmng->pools_mng.pools[idx];
661 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
662 for (i = 0; i < MLX5_ASO_MTRS_PER_POOL; i++) {
663 aso_mtr = &mtr_pool->mtrs[i];
664 if (aso_mtr->fm.meter_action)
666 (mlx5_glue->destroy_flow_action
667 (aso_mtr->fm.meter_action));
669 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
670 claim_zero(mlx5_devx_cmd_destroy
671 (mtr_pool->devx_obj));
672 mtrmng->pools_mng.n_valid--;
675 mlx5_free(sh->mtrmng->pools_mng.pools);
677 mlx5_free(sh->mtrmng);
681 /* Send FLOW_AGED event if needed. */
683 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
685 struct mlx5_age_info *age_info;
688 for (i = 0; i < sh->max_port; i++) {
689 age_info = &sh->port[i].age_info;
690 if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
692 MLX5_AGE_UNSET(age_info, MLX5_AGE_EVENT_NEW);
693 if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER)) {
694 MLX5_AGE_UNSET(age_info, MLX5_AGE_TRIGGER);
695 rte_eth_dev_callback_process
696 (&rte_eth_devices[sh->port[i].devx_ih_port_id],
697 RTE_ETH_EVENT_FLOW_AGED, NULL);
703 * Initialize the ASO connection tracking structure.
706 * Pointer to mlx5_dev_ctx_shared object.
709 * 0 on success, a negative errno value otherwise and rte_errno is set.
712 mlx5_flow_aso_ct_mng_init(struct mlx5_dev_ctx_shared *sh)
718 sh->ct_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->ct_mng),
719 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
721 DRV_LOG(ERR, "ASO CT management allocation failed.");
725 err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
727 mlx5_free(sh->ct_mng);
728 /* rte_errno should be extracted from the failure. */
732 rte_spinlock_init(&sh->ct_mng->ct_sl);
733 rte_rwlock_init(&sh->ct_mng->resize_rwl);
734 LIST_INIT(&sh->ct_mng->free_cts);
739 * Close and release all the resources of the
740 * ASO connection tracking management structure.
743 * Pointer to mlx5_dev_ctx_shared object to free.
746 mlx5_flow_aso_ct_mng_close(struct mlx5_dev_ctx_shared *sh)
748 struct mlx5_aso_ct_pools_mng *mng = sh->ct_mng;
749 struct mlx5_aso_ct_pool *ct_pool;
750 struct mlx5_aso_ct_action *ct;
756 mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
760 ct_pool = mng->pools[idx];
761 for (i = 0; i < MLX5_ASO_CT_ACTIONS_PER_POOL; i++) {
762 ct = &ct_pool->actions[i];
763 val = __atomic_fetch_sub(&ct->refcnt, 1,
765 MLX5_ASSERT(val == 1);
768 #ifdef HAVE_MLX5_DR_ACTION_ASO_CT
769 if (ct->dr_action_orig)
770 claim_zero(mlx5_glue->destroy_flow_action
771 (ct->dr_action_orig));
772 if (ct->dr_action_rply)
773 claim_zero(mlx5_glue->destroy_flow_action
774 (ct->dr_action_rply));
777 claim_zero(mlx5_devx_cmd_destroy(ct_pool->devx_obj));
779 DRV_LOG(DEBUG, "%u ASO CT objects are being used in the pool %u",
783 /* in case of failure. */
786 mlx5_free(mng->pools);
788 /* Management structure must be cleared to 0s during allocation. */
793 * Initialize the flow resources' indexed mempool.
796 * Pointer to mlx5_dev_ctx_shared object.
798 * Pointer to user dev config.
801 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
802 const struct mlx5_dev_config *config)
805 struct mlx5_indexed_pool_config cfg;
807 for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
808 cfg = mlx5_ipool_cfg[i];
813 * Set MLX5_IPOOL_MLX5_FLOW ipool size
814 * according to PCI function flow configuration.
816 case MLX5_IPOOL_MLX5_FLOW:
817 cfg.size = config->dv_flow_en ?
818 sizeof(struct mlx5_flow_handle) :
819 MLX5_FLOW_HANDLE_VERBS_SIZE;
822 if (config->reclaim_mode) {
823 cfg.release_mem_en = 1;
824 cfg.per_core_cache = 0;
826 cfg.release_mem_en = 0;
828 sh->ipool[i] = mlx5_ipool_create(&cfg);
834 * Release the flow resources' indexed mempool.
837 * Pointer to mlx5_dev_ctx_shared object.
840 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
844 for (i = 0; i < MLX5_IPOOL_MAX; ++i)
845 mlx5_ipool_destroy(sh->ipool[i]);
846 for (i = 0; i < MLX5_MAX_MODIFY_NUM; ++i)
847 if (sh->mdh_ipools[i])
848 mlx5_ipool_destroy(sh->mdh_ipools[i]);
852 * Check if dynamic flex parser for eCPRI already exists.
855 * Pointer to Ethernet device structure.
858 * true on exists, false on not.
861 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
863 struct mlx5_priv *priv = dev->data->dev_private;
864 struct mlx5_flex_parser_profiles *prf =
865 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
871 * Allocation of a flex parser for eCPRI. Once created, this parser related
872 * resources will be held until the device is closed.
875 * Pointer to Ethernet device structure.
878 * 0 on success, a negative errno value otherwise and rte_errno is set.
881 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
883 struct mlx5_priv *priv = dev->data->dev_private;
884 struct mlx5_flex_parser_profiles *prf =
885 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
886 struct mlx5_devx_graph_node_attr node = {
887 .modify_field_select = 0,
892 if (!priv->config.hca_attr.parse_graph_flex_node) {
893 DRV_LOG(ERR, "Dynamic flex parser is not supported "
894 "for device %s.", priv->dev_data->name);
897 node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
898 /* 8 bytes now: 4B common header + 4B message body header. */
899 node.header_length_base_value = 0x8;
900 /* After MAC layer: Ether / VLAN. */
901 node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
902 /* Type of compared condition should be 0xAEFE in the L2 layer. */
903 node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
904 /* Sample #0: type in common header. */
905 node.sample[0].flow_match_sample_en = 1;
907 node.sample[0].flow_match_sample_offset_mode = 0x0;
908 /* Only the 2nd byte will be used. */
909 node.sample[0].flow_match_sample_field_base_offset = 0x0;
910 /* Sample #1: message payload. */
911 node.sample[1].flow_match_sample_en = 1;
913 node.sample[1].flow_match_sample_offset_mode = 0x0;
915 * Only the first two bytes will be used right now, and its offset will
916 * start after the common header that with the length of a DW(u32).
918 node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
919 prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->cdev->ctx, &node);
921 DRV_LOG(ERR, "Failed to create flex parser node object.");
922 return (rte_errno == 0) ? -ENODEV : -rte_errno;
925 ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
927 DRV_LOG(ERR, "Failed to query sample IDs.");
928 return (rte_errno == 0) ? -ENODEV : -rte_errno;
930 prf->offset[0] = 0x0;
931 prf->offset[1] = sizeof(uint32_t);
932 prf->ids[0] = ids[0];
933 prf->ids[1] = ids[1];
938 * Destroy the flex parser node, including the parser itself, input / output
939 * arcs and DW samples. Resources could be reused then.
942 * Pointer to Ethernet device structure.
945 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
947 struct mlx5_priv *priv = dev->data->dev_private;
948 struct mlx5_flex_parser_profiles *prf =
949 &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
952 mlx5_devx_cmd_destroy(prf->obj);
957 mlx5_get_supported_sw_parsing_offloads(const struct mlx5_hca_attr *attr)
959 uint32_t sw_parsing_offloads = 0;
962 sw_parsing_offloads |= MLX5_SW_PARSING_CAP;
964 sw_parsing_offloads |= MLX5_SW_PARSING_CSUM_CAP;
967 sw_parsing_offloads |= MLX5_SW_PARSING_TSO_CAP;
969 return sw_parsing_offloads;
973 mlx5_get_supported_tunneling_offloads(const struct mlx5_hca_attr *attr)
975 uint32_t tn_offloads = 0;
977 if (attr->tunnel_stateless_vxlan)
978 tn_offloads |= MLX5_TUNNELED_OFFLOADS_VXLAN_CAP;
979 if (attr->tunnel_stateless_gre)
980 tn_offloads |= MLX5_TUNNELED_OFFLOADS_GRE_CAP;
981 if (attr->tunnel_stateless_geneve_rx)
982 tn_offloads |= MLX5_TUNNELED_OFFLOADS_GENEVE_CAP;
987 * Allocate Rx and Tx UARs in robust fashion.
988 * This routine handles the following UAR allocation issues:
990 * - tries to allocate the UAR with the most appropriate memory
991 * mapping type from the ones supported by the host
993 * - tries to allocate the UAR with non-NULL base address
994 * OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
995 * UAR base address if UAR was not the first object in the UAR page.
996 * It caused the PMD failure and we should try to get another UAR
997 * till we get the first one with non-NULL base address returned.
1000 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
1001 const struct mlx5_common_dev_config *config)
1003 uint32_t uar_mapping, retry;
1007 for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
1008 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1009 /* Control the mapping type according to the settings. */
1010 uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
1011 MLX5DV_UAR_ALLOC_TYPE_NC :
1012 MLX5DV_UAR_ALLOC_TYPE_BF;
1014 RTE_SET_USED(config);
1016 * It seems we have no way to control the memory mapping type
1017 * for the UAR, the default "Write-Combining" type is supposed.
1018 * The UAR initialization on queue creation queries the
1019 * actual mapping type done by Verbs/kernel and setups the
1020 * PMD datapath accordingly.
1024 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1026 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1028 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
1029 if (config->dbnc == MLX5_TXDB_CACHED ||
1030 config->dbnc == MLX5_TXDB_HEURISTIC)
1031 DRV_LOG(WARNING, "Devarg tx_db_nc setting "
1032 "is not supported by DevX");
1034 * In some environments like virtual machine
1035 * the Write Combining mapped might be not supported
1036 * and UAR allocation fails. We try "Non-Cached"
1037 * mapping for the case. The tx_burst routines take
1038 * the UAR mapping type into account on UAR setup
1039 * on queue creation.
1041 DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (BF)");
1042 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
1043 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1045 } else if (!sh->tx_uar &&
1046 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
1047 if (config->dbnc == MLX5_TXDB_NCACHED)
1048 DRV_LOG(WARNING, "Devarg tx_db_nc settings "
1049 "is not supported by DevX");
1051 * If Verbs/kernel does not support "Non-Cached"
1052 * try the "Write-Combining".
1054 DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (NC)");
1055 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
1056 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1061 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
1065 base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
1069 * The UARs are allocated by rdma_core within the
1070 * IB device context, on context closure all UARs
1071 * will be freed, should be no memory/object leakage.
1073 DRV_LOG(DEBUG, "Retrying to allocate Tx DevX UAR");
1076 /* Check whether we finally succeeded with valid UAR allocation. */
1078 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
1082 for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
1084 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar(sh->cdev->ctx,
1086 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1087 if (!sh->devx_rx_uar &&
1088 uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
1090 * Rx UAR is used to control interrupts only,
1091 * should be no datapath noticeable impact,
1092 * can try "Non-Cached" mapping safely.
1094 DRV_LOG(DEBUG, "Failed to allocate Rx DevX UAR (BF)");
1095 uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
1096 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
1097 (sh->cdev->ctx, uar_mapping);
1100 if (!sh->devx_rx_uar) {
1101 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
1105 base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
1109 * The UARs are allocated by rdma_core within the
1110 * IB device context, on context closure all UARs
1111 * will be freed, should be no memory/object leakage.
1113 DRV_LOG(DEBUG, "Retrying to allocate Rx DevX UAR");
1114 sh->devx_rx_uar = NULL;
1116 /* Check whether we finally succeeded with valid UAR allocation. */
1117 if (!sh->devx_rx_uar) {
1118 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
1126 * Unregister the mempool from the protection domain.
1129 * Pointer to the device shared context.
1131 * Mempool being unregistered.
1134 mlx5_dev_ctx_shared_mempool_unregister(struct mlx5_dev_ctx_shared *sh,
1135 struct rte_mempool *mp)
1137 struct mlx5_mp_id mp_id;
1139 mlx5_mp_id_init(&mp_id, 0);
1140 if (mlx5_mr_mempool_unregister(&sh->share_cache, mp, &mp_id) < 0)
1141 DRV_LOG(WARNING, "Failed to unregister mempool %s for PD %p: %s",
1142 mp->name, sh->cdev->pd, rte_strerror(rte_errno));
1146 * rte_mempool_walk() callback to register mempools
1147 * for the protection domain.
1150 * The mempool being walked.
1152 * Pointer to the device shared context.
1155 mlx5_dev_ctx_shared_mempool_register_cb(struct rte_mempool *mp, void *arg)
1157 struct mlx5_dev_ctx_shared *sh = arg;
1158 struct mlx5_mp_id mp_id;
1161 mlx5_mp_id_init(&mp_id, 0);
1162 ret = mlx5_mr_mempool_register(&sh->share_cache, sh->cdev->pd, mp,
1164 if (ret < 0 && rte_errno != EEXIST)
1165 DRV_LOG(ERR, "Failed to register existing mempool %s for PD %p: %s",
1166 mp->name, sh->cdev->pd, rte_strerror(rte_errno));
1170 * rte_mempool_walk() callback to unregister mempools
1171 * from the protection domain.
1174 * The mempool being walked.
1176 * Pointer to the device shared context.
1179 mlx5_dev_ctx_shared_mempool_unregister_cb(struct rte_mempool *mp, void *arg)
1181 mlx5_dev_ctx_shared_mempool_unregister
1182 ((struct mlx5_dev_ctx_shared *)arg, mp);
1186 * Mempool life cycle callback for Ethernet devices.
1189 * Mempool life cycle event.
1191 * Associated mempool.
1193 * Pointer to a device shared context.
1196 mlx5_dev_ctx_shared_mempool_event_cb(enum rte_mempool_event event,
1197 struct rte_mempool *mp, void *arg)
1199 struct mlx5_dev_ctx_shared *sh = arg;
1200 struct mlx5_mp_id mp_id;
1203 case RTE_MEMPOOL_EVENT_READY:
1204 mlx5_mp_id_init(&mp_id, 0);
1205 if (mlx5_mr_mempool_register(&sh->share_cache, sh->cdev->pd, mp,
1207 DRV_LOG(ERR, "Failed to register new mempool %s for PD %p: %s",
1208 mp->name, sh->cdev->pd,
1209 rte_strerror(rte_errno));
1211 case RTE_MEMPOOL_EVENT_DESTROY:
1212 mlx5_dev_ctx_shared_mempool_unregister(sh, mp);
1218 * Callback used when implicit mempool registration is disabled
1219 * in order to track Rx mempool destruction.
1222 * Mempool life cycle event.
1224 * An Rx mempool registered explicitly when the port is started.
1226 * Pointer to a device shared context.
1229 mlx5_dev_ctx_shared_rx_mempool_event_cb(enum rte_mempool_event event,
1230 struct rte_mempool *mp, void *arg)
1232 struct mlx5_dev_ctx_shared *sh = arg;
1234 if (event == RTE_MEMPOOL_EVENT_DESTROY)
1235 mlx5_dev_ctx_shared_mempool_unregister(sh, mp);
1239 mlx5_dev_ctx_shared_mempool_subscribe(struct rte_eth_dev *dev)
1241 struct mlx5_priv *priv = dev->data->dev_private;
1242 struct mlx5_dev_ctx_shared *sh = priv->sh;
1245 /* Check if we only need to track Rx mempool destruction. */
1246 if (!sh->cdev->config.mr_mempool_reg_en) {
1247 ret = rte_mempool_event_callback_register
1248 (mlx5_dev_ctx_shared_rx_mempool_event_cb, sh);
1249 return ret == 0 || rte_errno == EEXIST ? 0 : ret;
1251 /* Callback for this shared context may be already registered. */
1252 ret = rte_mempool_event_callback_register
1253 (mlx5_dev_ctx_shared_mempool_event_cb, sh);
1254 if (ret != 0 && rte_errno != EEXIST)
1256 /* Register mempools only once for this shared context. */
1258 rte_mempool_walk(mlx5_dev_ctx_shared_mempool_register_cb, sh);
1263 * Allocate shared device context. If there is multiport device the
1264 * master and representors will share this context, if there is single
1265 * port dedicated device, the context will be used by only given
1266 * port due to unification.
1268 * Routine first searches the context for the specified device name,
1269 * if found the shared context assumed and reference counter is incremented.
1270 * If no context found the new one is created and initialized with specified
1271 * device context and parameters.
1274 * Pointer to the device attributes (name, port, etc).
1276 * Pointer to device configuration structure.
1279 * Pointer to mlx5_dev_ctx_shared object on success,
1280 * otherwise NULL and rte_errno is set.
1282 struct mlx5_dev_ctx_shared *
1283 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
1284 const struct mlx5_dev_config *config)
1286 struct mlx5_dev_ctx_shared *sh;
1289 struct mlx5_devx_tis_attr tis_attr = { 0 };
1292 /* Secondary process should not create the shared context. */
1293 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1294 pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1295 /* Search for IB context by device name. */
1296 LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
1297 if (!strcmp(sh->ibdev_name, spawn->phys_dev_name)) {
1302 /* No device found, we have to create new shared context. */
1303 MLX5_ASSERT(spawn->max_port);
1304 sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1305 sizeof(struct mlx5_dev_ctx_shared) +
1307 sizeof(struct mlx5_dev_shared_port),
1308 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1310 DRV_LOG(ERR, "shared context allocation failure");
1314 pthread_mutex_init(&sh->txpp.mutex, NULL);
1315 sh->numa_node = spawn->cdev->dev->numa_node;
1316 sh->cdev = spawn->cdev;
1317 sh->devx = sh->cdev->config.devx;
1318 if (spawn->bond_info)
1319 sh->bond = *spawn->bond_info;
1320 err = mlx5_os_get_dev_attr(sh->cdev->ctx, &sh->device_attr);
1322 DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
1326 sh->max_port = spawn->max_port;
1327 sh->reclaim_mode = config->reclaim_mode;
1328 strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->cdev->ctx),
1329 sizeof(sh->ibdev_name) - 1);
1330 strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->cdev->ctx),
1331 sizeof(sh->ibdev_path) - 1);
1333 * Setting port_id to max unallowed value means
1334 * there is no interrupt subhandler installed for
1335 * the given port index i.
1337 for (i = 0; i < sh->max_port; i++) {
1338 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
1339 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
1342 sh->td = mlx5_devx_cmd_create_td(sh->cdev->ctx);
1344 DRV_LOG(ERR, "TD allocation failure");
1348 tis_attr.transport_domain = sh->td->id;
1349 sh->tis = mlx5_devx_cmd_create_tis(sh->cdev->ctx, &tis_attr);
1351 DRV_LOG(ERR, "TIS allocation failure");
1355 err = mlx5_alloc_rxtx_uars(sh, &sh->cdev->config);
1358 MLX5_ASSERT(sh->tx_uar);
1359 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
1361 MLX5_ASSERT(sh->devx_rx_uar);
1362 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
1365 /* Initialize UAR access locks for 32bit implementations. */
1366 rte_spinlock_init(&sh->uar_lock_cq);
1367 for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1368 rte_spinlock_init(&sh->uar_lock[i]);
1371 * Once the device is added to the list of memory event
1372 * callback, its global MR cache table cannot be expanded
1373 * on the fly because of deadlock. If it overflows, lookup
1374 * should be done by searching MR list linearly, which is slow.
1376 * At this point the device is not added to the memory
1377 * event list yet, context is just being created.
1379 err = mlx5_mr_btree_init(&sh->share_cache.cache,
1380 MLX5_MR_BTREE_CACHE_N * 2,
1386 mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
1387 &sh->share_cache.dereg_mr_cb);
1388 mlx5_os_dev_shared_handler_install(sh);
1389 if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1390 err = mlx5_flow_os_init_workspace_once();
1394 mlx5_flow_aging_init(sh);
1395 mlx5_flow_counters_mng_init(sh);
1396 mlx5_flow_ipool_create(sh, config);
1397 /* Add device to memory callback list. */
1398 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1399 LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1401 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1402 /* Add context to the global device list. */
1403 LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1404 rte_spinlock_init(&sh->geneve_tlv_opt_sl);
1406 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1409 pthread_mutex_destroy(&sh->txpp.mutex);
1410 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1412 if (sh->share_cache.cache.table)
1413 mlx5_mr_btree_free(&sh->share_cache.cache);
1415 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1417 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1418 if (sh->devx_rx_uar)
1419 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1421 mlx5_glue->devx_free_uar(sh->tx_uar);
1423 MLX5_ASSERT(err > 0);
1429 * Free shared IB device context. Decrement counter and if zero free
1430 * all allocated resources and close handles.
1433 * Pointer to mlx5_dev_ctx_shared object to free
1436 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1440 pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1441 #ifdef RTE_LIBRTE_MLX5_DEBUG
1442 /* Check the object presence in the list. */
1443 struct mlx5_dev_ctx_shared *lctx;
1445 LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1450 DRV_LOG(ERR, "Freeing non-existing shared IB context");
1455 MLX5_ASSERT(sh->refcnt);
1456 /* Secondary process should not free the shared context. */
1457 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1460 /* Stop watching for mempool events and unregister all mempools. */
1461 ret = rte_mempool_event_callback_unregister
1462 (mlx5_dev_ctx_shared_mempool_event_cb, sh);
1463 if (ret < 0 && rte_errno == ENOENT)
1464 ret = rte_mempool_event_callback_unregister
1465 (mlx5_dev_ctx_shared_rx_mempool_event_cb, sh);
1467 rte_mempool_walk(mlx5_dev_ctx_shared_mempool_unregister_cb,
1469 /* Remove from memory callback device list. */
1470 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1471 LIST_REMOVE(sh, mem_event_cb);
1472 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1473 /* Release created Memory Regions. */
1474 mlx5_mr_release_cache(&sh->share_cache);
1475 /* Remove context from the global device list. */
1476 LIST_REMOVE(sh, next);
1477 /* Release flow workspaces objects on the last device. */
1478 if (LIST_EMPTY(&mlx5_dev_ctx_list))
1479 mlx5_flow_os_release_workspace();
1480 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1482 * Ensure there is no async event handler installed.
1483 * Only primary process handles async device events.
1485 mlx5_flow_counters_mng_close(sh);
1486 if (sh->aso_age_mng) {
1487 mlx5_flow_aso_age_mng_close(sh);
1488 sh->aso_age_mng = NULL;
1491 mlx5_aso_flow_mtrs_mng_close(sh);
1492 mlx5_flow_ipool_destroy(sh);
1493 mlx5_os_dev_shared_handler_uninstall(sh);
1495 mlx5_glue->devx_free_uar(sh->tx_uar);
1499 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1501 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1502 if (sh->devx_rx_uar)
1503 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1504 MLX5_ASSERT(sh->geneve_tlv_option_resource == NULL);
1505 pthread_mutex_destroy(&sh->txpp.mutex);
1509 pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1513 * Destroy table hash list.
1516 * Pointer to the private device data structure.
1519 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1521 struct mlx5_dev_ctx_shared *sh = priv->sh;
1525 mlx5_hlist_destroy(sh->flow_tbls);
1526 sh->flow_tbls = NULL;
1530 * Initialize flow table hash list and create the root tables entry
1534 * Pointer to the private device data structure.
1537 * Zero on success, positive error code otherwise.
1540 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1543 /* Tables are only used in DV and DR modes. */
1544 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1545 struct mlx5_dev_ctx_shared *sh = priv->sh;
1546 char s[MLX5_NAME_SIZE];
1549 snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1550 sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1552 flow_dv_tbl_create_cb,
1553 flow_dv_tbl_match_cb,
1554 flow_dv_tbl_remove_cb,
1555 flow_dv_tbl_clone_cb,
1556 flow_dv_tbl_clone_free_cb);
1557 if (!sh->flow_tbls) {
1558 DRV_LOG(ERR, "flow tables with hash creation failed.");
1562 #ifndef HAVE_MLX5DV_DR
1563 struct rte_flow_error error;
1564 struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1567 * In case we have not DR support, the zero tables should be created
1568 * because DV expect to see them even if they cannot be created by
1571 if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0,
1572 NULL, 0, 1, 0, &error) ||
1573 !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0,
1574 NULL, 0, 1, 0, &error) ||
1575 !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0,
1576 NULL, 0, 1, 0, &error)) {
1582 mlx5_free_table_hash_list(priv);
1583 #endif /* HAVE_MLX5DV_DR */
1589 * Retrieve integer value from environment variable.
1592 * Environment variable name.
1595 * Integer value, 0 if the variable is not set.
1598 mlx5_getenv_int(const char *name)
1600 const char *val = getenv(name);
1608 * DPDK callback to add udp tunnel port
1611 * A pointer to eth_dev
1612 * @param[in] udp_tunnel
1613 * A pointer to udp tunnel
1616 * 0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1619 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1620 struct rte_eth_udp_tunnel *udp_tunnel)
1622 MLX5_ASSERT(udp_tunnel != NULL);
1623 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1624 udp_tunnel->udp_port == 4789)
1626 if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1627 udp_tunnel->udp_port == 4790)
1633 * Initialize process private data structure.
1636 * Pointer to Ethernet device structure.
1639 * 0 on success, a negative errno value otherwise and rte_errno is set.
1642 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1644 struct mlx5_priv *priv = dev->data->dev_private;
1645 struct mlx5_proc_priv *ppriv;
1648 mlx5_proc_priv_uninit(dev);
1650 * UAR register table follows the process private structure. BlueFlame
1651 * registers for Tx queues are stored in the table.
1654 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1655 ppriv = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, ppriv_size,
1656 RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1661 ppriv->uar_table_sz = priv->txqs_n;
1662 dev->process_private = ppriv;
1667 * Un-initialize process private data structure.
1670 * Pointer to Ethernet device structure.
1673 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1675 if (!dev->process_private)
1677 mlx5_free(dev->process_private);
1678 dev->process_private = NULL;
1682 * DPDK callback to close the device.
1684 * Destroy all queues and objects, free memory.
1687 * Pointer to Ethernet device structure.
1690 mlx5_dev_close(struct rte_eth_dev *dev)
1692 struct mlx5_priv *priv = dev->data->dev_private;
1696 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1697 /* Check if process_private released. */
1698 if (!dev->process_private)
1700 mlx5_tx_uar_uninit_secondary(dev);
1701 mlx5_proc_priv_uninit(dev);
1702 rte_eth_dev_release_port(dev);
1707 DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1709 ((priv->sh->cdev->ctx != NULL) ?
1710 mlx5_os_get_ctx_device_name(priv->sh->cdev->ctx) : ""));
1712 * If default mreg copy action is removed at the stop stage,
1713 * the search will return none and nothing will be done anymore.
1715 mlx5_flow_stop_default(dev);
1716 mlx5_traffic_disable(dev);
1718 * If all the flows are already flushed in the device stop stage,
1719 * then this will return directly without any action.
1721 mlx5_flow_list_flush(dev, MLX5_FLOW_TYPE_GEN, true);
1722 mlx5_action_handle_flush(dev);
1723 mlx5_flow_meter_flush(dev, NULL);
1724 /* Prevent crashes when queues are still in use. */
1725 dev->rx_pkt_burst = removed_rx_burst;
1726 dev->tx_pkt_burst = removed_tx_burst;
1728 /* Disable datapath on secondary process. */
1729 mlx5_mp_os_req_stop_rxtx(dev);
1730 /* Free the eCPRI flex parser resource. */
1731 mlx5_flex_parser_ecpri_release(dev);
1732 if (priv->rxqs != NULL) {
1733 /* XXX race condition if mlx5_rx_burst() is still running. */
1734 rte_delay_us_sleep(1000);
1735 for (i = 0; (i != priv->rxqs_n); ++i)
1736 mlx5_rxq_release(dev, i);
1740 if (priv->representor) {
1741 /* Each representor has a dedicated interrupts handler */
1742 mlx5_free(dev->intr_handle);
1743 dev->intr_handle = NULL;
1745 if (priv->txqs != NULL) {
1746 /* XXX race condition if mlx5_tx_burst() is still running. */
1747 rte_delay_us_sleep(1000);
1748 for (i = 0; (i != priv->txqs_n); ++i)
1749 mlx5_txq_release(dev, i);
1753 mlx5_proc_priv_uninit(dev);
1754 if (priv->q_counters) {
1755 mlx5_devx_cmd_destroy(priv->q_counters);
1756 priv->q_counters = NULL;
1758 if (priv->drop_queue.hrxq)
1759 mlx5_drop_action_destroy(dev);
1760 if (priv->mreg_cp_tbl)
1761 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1762 mlx5_mprq_free_mp(dev);
1763 if (priv->sh->ct_mng)
1764 mlx5_flow_aso_ct_mng_close(priv->sh);
1765 mlx5_os_free_shared_dr(priv);
1766 if (priv->rss_conf.rss_key != NULL)
1767 mlx5_free(priv->rss_conf.rss_key);
1768 if (priv->reta_idx != NULL)
1769 mlx5_free(priv->reta_idx);
1770 if (priv->config.vf)
1771 mlx5_os_mac_addr_flush(dev);
1772 if (priv->nl_socket_route >= 0)
1773 close(priv->nl_socket_route);
1774 if (priv->nl_socket_rdma >= 0)
1775 close(priv->nl_socket_rdma);
1776 if (priv->vmwa_context)
1777 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1778 ret = mlx5_hrxq_verify(dev);
1780 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1781 dev->data->port_id);
1782 ret = mlx5_ind_table_obj_verify(dev);
1784 DRV_LOG(WARNING, "port %u some indirection table still remain",
1785 dev->data->port_id);
1786 ret = mlx5_rxq_obj_verify(dev);
1788 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1789 dev->data->port_id);
1790 ret = mlx5_rxq_verify(dev);
1792 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1793 dev->data->port_id);
1794 ret = mlx5_txq_obj_verify(dev);
1796 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1797 dev->data->port_id);
1798 ret = mlx5_txq_verify(dev);
1800 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1801 dev->data->port_id);
1802 ret = mlx5_flow_verify(dev);
1804 DRV_LOG(WARNING, "port %u some flows still remain",
1805 dev->data->port_id);
1807 mlx5_list_destroy(priv->hrxqs);
1809 * Free the shared context in last turn, because the cleanup
1810 * routines above may use some shared fields, like
1811 * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1812 * ifindex if Netlink fails.
1814 mlx5_free_shared_dev_ctx(priv->sh);
1815 if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1819 MLX5_ETH_FOREACH_DEV(port_id, dev->device) {
1820 struct mlx5_priv *opriv =
1821 rte_eth_devices[port_id].data->dev_private;
1824 opriv->domain_id != priv->domain_id ||
1825 &rte_eth_devices[port_id] == dev)
1831 claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1833 memset(priv, 0, sizeof(*priv));
1834 priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1836 * Reset mac_addrs to NULL such that it is not freed as part of
1837 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1838 * it is freed when dev_private is freed.
1840 dev->data->mac_addrs = NULL;
1844 const struct eth_dev_ops mlx5_dev_ops = {
1845 .dev_configure = mlx5_dev_configure,
1846 .dev_start = mlx5_dev_start,
1847 .dev_stop = mlx5_dev_stop,
1848 .dev_set_link_down = mlx5_set_link_down,
1849 .dev_set_link_up = mlx5_set_link_up,
1850 .dev_close = mlx5_dev_close,
1851 .promiscuous_enable = mlx5_promiscuous_enable,
1852 .promiscuous_disable = mlx5_promiscuous_disable,
1853 .allmulticast_enable = mlx5_allmulticast_enable,
1854 .allmulticast_disable = mlx5_allmulticast_disable,
1855 .link_update = mlx5_link_update,
1856 .stats_get = mlx5_stats_get,
1857 .stats_reset = mlx5_stats_reset,
1858 .xstats_get = mlx5_xstats_get,
1859 .xstats_reset = mlx5_xstats_reset,
1860 .xstats_get_names = mlx5_xstats_get_names,
1861 .fw_version_get = mlx5_fw_version_get,
1862 .dev_infos_get = mlx5_dev_infos_get,
1863 .representor_info_get = mlx5_representor_info_get,
1864 .read_clock = mlx5_txpp_read_clock,
1865 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1866 .vlan_filter_set = mlx5_vlan_filter_set,
1867 .rx_queue_setup = mlx5_rx_queue_setup,
1868 .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1869 .tx_queue_setup = mlx5_tx_queue_setup,
1870 .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1871 .rx_queue_release = mlx5_rx_queue_release,
1872 .tx_queue_release = mlx5_tx_queue_release,
1873 .rx_queue_start = mlx5_rx_queue_start,
1874 .rx_queue_stop = mlx5_rx_queue_stop,
1875 .tx_queue_start = mlx5_tx_queue_start,
1876 .tx_queue_stop = mlx5_tx_queue_stop,
1877 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1878 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1879 .mac_addr_remove = mlx5_mac_addr_remove,
1880 .mac_addr_add = mlx5_mac_addr_add,
1881 .mac_addr_set = mlx5_mac_addr_set,
1882 .set_mc_addr_list = mlx5_set_mc_addr_list,
1883 .mtu_set = mlx5_dev_set_mtu,
1884 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1885 .vlan_offload_set = mlx5_vlan_offload_set,
1886 .reta_update = mlx5_dev_rss_reta_update,
1887 .reta_query = mlx5_dev_rss_reta_query,
1888 .rss_hash_update = mlx5_rss_hash_update,
1889 .rss_hash_conf_get = mlx5_rss_hash_conf_get,
1890 .flow_ops_get = mlx5_flow_ops_get,
1891 .rxq_info_get = mlx5_rxq_info_get,
1892 .txq_info_get = mlx5_txq_info_get,
1893 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1894 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1895 .rx_queue_intr_enable = mlx5_rx_intr_enable,
1896 .rx_queue_intr_disable = mlx5_rx_intr_disable,
1897 .is_removed = mlx5_is_removed,
1898 .udp_tunnel_port_add = mlx5_udp_tunnel_port_add,
1899 .get_module_info = mlx5_get_module_info,
1900 .get_module_eeprom = mlx5_get_module_eeprom,
1901 .hairpin_cap_get = mlx5_hairpin_cap_get,
1902 .mtr_ops_get = mlx5_flow_meter_ops_get,
1903 .hairpin_bind = mlx5_hairpin_bind,
1904 .hairpin_unbind = mlx5_hairpin_unbind,
1905 .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1906 .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1907 .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1908 .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1909 .get_monitor_addr = mlx5_get_monitor_addr,
1912 /* Available operations from secondary process. */
1913 const struct eth_dev_ops mlx5_dev_sec_ops = {
1914 .stats_get = mlx5_stats_get,
1915 .stats_reset = mlx5_stats_reset,
1916 .xstats_get = mlx5_xstats_get,
1917 .xstats_reset = mlx5_xstats_reset,
1918 .xstats_get_names = mlx5_xstats_get_names,
1919 .fw_version_get = mlx5_fw_version_get,
1920 .dev_infos_get = mlx5_dev_infos_get,
1921 .representor_info_get = mlx5_representor_info_get,
1922 .read_clock = mlx5_txpp_read_clock,
1923 .rx_queue_start = mlx5_rx_queue_start,
1924 .rx_queue_stop = mlx5_rx_queue_stop,
1925 .tx_queue_start = mlx5_tx_queue_start,
1926 .tx_queue_stop = mlx5_tx_queue_stop,
1927 .rxq_info_get = mlx5_rxq_info_get,
1928 .txq_info_get = mlx5_txq_info_get,
1929 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1930 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1931 .get_module_info = mlx5_get_module_info,
1932 .get_module_eeprom = mlx5_get_module_eeprom,
1935 /* Available operations in flow isolated mode. */
1936 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1937 .dev_configure = mlx5_dev_configure,
1938 .dev_start = mlx5_dev_start,
1939 .dev_stop = mlx5_dev_stop,
1940 .dev_set_link_down = mlx5_set_link_down,
1941 .dev_set_link_up = mlx5_set_link_up,
1942 .dev_close = mlx5_dev_close,
1943 .promiscuous_enable = mlx5_promiscuous_enable,
1944 .promiscuous_disable = mlx5_promiscuous_disable,
1945 .allmulticast_enable = mlx5_allmulticast_enable,
1946 .allmulticast_disable = mlx5_allmulticast_disable,
1947 .link_update = mlx5_link_update,
1948 .stats_get = mlx5_stats_get,
1949 .stats_reset = mlx5_stats_reset,
1950 .xstats_get = mlx5_xstats_get,
1951 .xstats_reset = mlx5_xstats_reset,
1952 .xstats_get_names = mlx5_xstats_get_names,
1953 .fw_version_get = mlx5_fw_version_get,
1954 .dev_infos_get = mlx5_dev_infos_get,
1955 .representor_info_get = mlx5_representor_info_get,
1956 .read_clock = mlx5_txpp_read_clock,
1957 .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1958 .vlan_filter_set = mlx5_vlan_filter_set,
1959 .rx_queue_setup = mlx5_rx_queue_setup,
1960 .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1961 .tx_queue_setup = mlx5_tx_queue_setup,
1962 .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1963 .rx_queue_release = mlx5_rx_queue_release,
1964 .tx_queue_release = mlx5_tx_queue_release,
1965 .rx_queue_start = mlx5_rx_queue_start,
1966 .rx_queue_stop = mlx5_rx_queue_stop,
1967 .tx_queue_start = mlx5_tx_queue_start,
1968 .tx_queue_stop = mlx5_tx_queue_stop,
1969 .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1970 .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1971 .mac_addr_remove = mlx5_mac_addr_remove,
1972 .mac_addr_add = mlx5_mac_addr_add,
1973 .mac_addr_set = mlx5_mac_addr_set,
1974 .set_mc_addr_list = mlx5_set_mc_addr_list,
1975 .mtu_set = mlx5_dev_set_mtu,
1976 .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1977 .vlan_offload_set = mlx5_vlan_offload_set,
1978 .flow_ops_get = mlx5_flow_ops_get,
1979 .rxq_info_get = mlx5_rxq_info_get,
1980 .txq_info_get = mlx5_txq_info_get,
1981 .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1982 .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1983 .rx_queue_intr_enable = mlx5_rx_intr_enable,
1984 .rx_queue_intr_disable = mlx5_rx_intr_disable,
1985 .is_removed = mlx5_is_removed,
1986 .get_module_info = mlx5_get_module_info,
1987 .get_module_eeprom = mlx5_get_module_eeprom,
1988 .hairpin_cap_get = mlx5_hairpin_cap_get,
1989 .mtr_ops_get = mlx5_flow_meter_ops_get,
1990 .hairpin_bind = mlx5_hairpin_bind,
1991 .hairpin_unbind = mlx5_hairpin_unbind,
1992 .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1993 .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1994 .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1995 .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1996 .get_monitor_addr = mlx5_get_monitor_addr,
2000 * Verify and store value for device argument.
2003 * Key argument to verify.
2005 * Value associated with key.
2010 * 0 on success, a negative errno value otherwise and rte_errno is set.
2013 mlx5_args_check(const char *key, const char *val, void *opaque)
2015 struct mlx5_dev_config *config = opaque;
2019 /* No-op, port representors are processed in mlx5_dev_spawn(). */
2020 if (!strcmp(MLX5_DRIVER_KEY, key) || !strcmp(MLX5_REPRESENTOR, key) ||
2021 !strcmp(MLX5_SYS_MEM_EN, key) || !strcmp(MLX5_TX_DB_NC, key) ||
2022 !strcmp(MLX5_MR_MEMPOOL_REG_EN, key) ||
2023 !strcmp(MLX5_MR_EXT_MEMSEG_EN, key))
2026 tmp = strtol(val, NULL, 0);
2029 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
2032 if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
2033 /* Negative values are acceptable for some keys only. */
2035 DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
2038 mod = tmp >= 0 ? tmp : -tmp;
2039 if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
2040 if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
2041 DRV_LOG(ERR, "invalid CQE compression "
2042 "format parameter");
2046 config->cqe_comp = !!tmp;
2047 config->cqe_comp_fmt = tmp;
2048 } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
2049 config->hw_padding = !!tmp;
2050 } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
2051 config->mprq.enabled = !!tmp;
2052 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
2053 config->mprq.stride_num_n = tmp;
2054 } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
2055 config->mprq.stride_size_n = tmp;
2056 } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
2057 config->mprq.max_memcpy_len = tmp;
2058 } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
2059 config->mprq.min_rxqs_num = tmp;
2060 } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
2061 DRV_LOG(WARNING, "%s: deprecated parameter,"
2062 " converted to txq_inline_max", key);
2063 config->txq_inline_max = tmp;
2064 } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
2065 config->txq_inline_max = tmp;
2066 } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
2067 config->txq_inline_min = tmp;
2068 } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
2069 config->txq_inline_mpw = tmp;
2070 } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
2071 config->txqs_inline = tmp;
2072 } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
2073 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2074 } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
2075 config->mps = !!tmp;
2076 } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
2077 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2078 } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
2079 DRV_LOG(WARNING, "%s: deprecated parameter,"
2080 " converted to txq_inline_mpw", key);
2081 config->txq_inline_mpw = tmp;
2082 } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
2083 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2084 } else if (strcmp(MLX5_TX_PP, key) == 0) {
2086 DRV_LOG(ERR, "Zero Tx packet pacing parameter");
2090 config->tx_pp = tmp;
2091 } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
2092 config->tx_skew = tmp;
2093 } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
2094 config->rx_vec_en = !!tmp;
2095 } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
2096 config->l3_vxlan_en = !!tmp;
2097 } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
2098 config->vf_nl_en = !!tmp;
2099 } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
2100 config->dv_esw_en = !!tmp;
2101 } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
2102 config->dv_flow_en = !!tmp;
2103 } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
2104 if (tmp != MLX5_XMETA_MODE_LEGACY &&
2105 tmp != MLX5_XMETA_MODE_META16 &&
2106 tmp != MLX5_XMETA_MODE_META32 &&
2107 tmp != MLX5_XMETA_MODE_MISS_INFO) {
2108 DRV_LOG(ERR, "invalid extensive "
2109 "metadata parameter");
2113 if (tmp != MLX5_XMETA_MODE_MISS_INFO)
2114 config->dv_xmeta_en = tmp;
2116 config->dv_miss_info = 1;
2117 } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
2118 config->lacp_by_user = !!tmp;
2119 } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
2120 config->max_dump_files_num = tmp;
2121 } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
2122 config->lro.timeout = tmp;
2123 } else if (strcmp(RTE_DEVARGS_KEY_CLASS, key) == 0) {
2124 DRV_LOG(DEBUG, "class argument is %s.", val);
2125 } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
2126 config->log_hp_size = tmp;
2127 } else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
2128 if (tmp != MLX5_RCM_NONE &&
2129 tmp != MLX5_RCM_LIGHT &&
2130 tmp != MLX5_RCM_AGGR) {
2131 DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
2135 config->reclaim_mode = tmp;
2136 } else if (strcmp(MLX5_DECAP_EN, key) == 0) {
2137 config->decap_en = !!tmp;
2138 } else if (strcmp(MLX5_ALLOW_DUPLICATE_PATTERN, key) == 0) {
2139 config->allow_duplicate_pattern = !!tmp;
2141 DRV_LOG(WARNING, "%s: unknown parameter", key);
2149 * Parse device parameters.
2152 * Pointer to device configuration structure.
2154 * Device arguments structure.
2157 * 0 on success, a negative errno value otherwise and rte_errno is set.
2160 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
2162 const char **params = (const char *[]){
2164 MLX5_RXQ_CQE_COMP_EN,
2165 MLX5_RXQ_PKT_PAD_EN,
2167 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
2168 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
2169 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
2172 MLX5_TXQ_INLINE_MIN,
2173 MLX5_TXQ_INLINE_MAX,
2174 MLX5_TXQ_INLINE_MPW,
2175 MLX5_TXQS_MIN_INLINE,
2178 MLX5_TXQ_MPW_HDR_DSEG_EN,
2179 MLX5_TXQ_MAX_INLINE_LEN,
2191 MLX5_MR_EXT_MEMSEG_EN,
2193 MLX5_MAX_DUMP_FILES_NUM,
2194 MLX5_LRO_TIMEOUT_USEC,
2195 RTE_DEVARGS_KEY_CLASS,
2200 MLX5_ALLOW_DUPLICATE_PATTERN,
2201 MLX5_MR_MEMPOOL_REG_EN,
2204 struct rte_kvargs *kvlist;
2208 if (devargs == NULL)
2210 /* Following UGLY cast is done to pass checkpatch. */
2211 kvlist = rte_kvargs_parse(devargs->args, params);
2212 if (kvlist == NULL) {
2216 /* Process parameters. */
2217 for (i = 0; (params[i] != NULL); ++i) {
2218 if (rte_kvargs_count(kvlist, params[i])) {
2219 ret = rte_kvargs_process(kvlist, params[i],
2220 mlx5_args_check, config);
2223 rte_kvargs_free(kvlist);
2228 rte_kvargs_free(kvlist);
2233 * Configures the minimal amount of data to inline into WQE
2234 * while sending packets.
2236 * - the txq_inline_min has the maximal priority, if this
2237 * key is specified in devargs
2238 * - if DevX is enabled the inline mode is queried from the
2239 * device (HCA attributes and NIC vport context if needed).
2240 * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
2241 * and none (0 bytes) for other NICs
2244 * Verbs device parameters (name, port, switch_info) to spawn.
2246 * Device configuration parameters.
2249 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
2250 struct mlx5_dev_config *config)
2252 if (config->txq_inline_min != MLX5_ARG_UNSET) {
2253 /* Application defines size of inlined data explicitly. */
2254 if (spawn->pci_dev != NULL) {
2255 switch (spawn->pci_dev->id.device_id) {
2256 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2257 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2258 if (config->txq_inline_min <
2259 (int)MLX5_INLINE_HSIZE_L2) {
2261 "txq_inline_mix aligned to minimal ConnectX-4 required value %d",
2262 (int)MLX5_INLINE_HSIZE_L2);
2263 config->txq_inline_min =
2264 MLX5_INLINE_HSIZE_L2;
2271 if (config->hca_attr.eth_net_offloads) {
2272 /* We have DevX enabled, inline mode queried successfully. */
2273 switch (config->hca_attr.wqe_inline_mode) {
2274 case MLX5_CAP_INLINE_MODE_L2:
2275 /* outer L2 header must be inlined. */
2276 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2278 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
2279 /* No inline data are required by NIC. */
2280 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2281 config->hw_vlan_insert =
2282 config->hca_attr.wqe_vlan_insert;
2283 DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
2285 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
2286 /* inline mode is defined by NIC vport context. */
2287 if (!config->hca_attr.eth_virt)
2289 switch (config->hca_attr.vport_inline_mode) {
2290 case MLX5_INLINE_MODE_NONE:
2291 config->txq_inline_min =
2292 MLX5_INLINE_HSIZE_NONE;
2294 case MLX5_INLINE_MODE_L2:
2295 config->txq_inline_min =
2296 MLX5_INLINE_HSIZE_L2;
2298 case MLX5_INLINE_MODE_IP:
2299 config->txq_inline_min =
2300 MLX5_INLINE_HSIZE_L3;
2302 case MLX5_INLINE_MODE_TCP_UDP:
2303 config->txq_inline_min =
2304 MLX5_INLINE_HSIZE_L4;
2306 case MLX5_INLINE_MODE_INNER_L2:
2307 config->txq_inline_min =
2308 MLX5_INLINE_HSIZE_INNER_L2;
2310 case MLX5_INLINE_MODE_INNER_IP:
2311 config->txq_inline_min =
2312 MLX5_INLINE_HSIZE_INNER_L3;
2314 case MLX5_INLINE_MODE_INNER_TCP_UDP:
2315 config->txq_inline_min =
2316 MLX5_INLINE_HSIZE_INNER_L4;
2321 if (spawn->pci_dev == NULL) {
2322 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2326 * We get here if we are unable to deduce
2327 * inline data size with DevX. Try PCI ID
2328 * to determine old NICs.
2330 switch (spawn->pci_dev->id.device_id) {
2331 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2332 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2333 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2334 case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2335 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2336 config->hw_vlan_insert = 0;
2338 case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2339 case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2340 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2341 case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2343 * These NICs support VLAN insertion from WQE and
2344 * report the wqe_vlan_insert flag. But there is the bug
2345 * and PFC control may be broken, so disable feature.
2347 config->hw_vlan_insert = 0;
2348 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2351 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2355 DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2359 * Configures the metadata mask fields in the shared context.
2362 * Pointer to Ethernet device.
2365 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2367 struct mlx5_priv *priv = dev->data->dev_private;
2368 struct mlx5_dev_ctx_shared *sh = priv->sh;
2369 uint32_t meta, mark, reg_c0;
2371 reg_c0 = ~priv->vport_meta_mask;
2372 switch (priv->config.dv_xmeta_en) {
2373 case MLX5_XMETA_MODE_LEGACY:
2375 mark = MLX5_FLOW_MARK_MASK;
2377 case MLX5_XMETA_MODE_META16:
2378 meta = reg_c0 >> rte_bsf32(reg_c0);
2379 mark = MLX5_FLOW_MARK_MASK;
2381 case MLX5_XMETA_MODE_META32:
2383 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2391 if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2392 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2393 sh->dv_mark_mask, mark);
2395 sh->dv_mark_mask = mark;
2396 if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2397 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2398 sh->dv_meta_mask, meta);
2400 sh->dv_meta_mask = meta;
2401 if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2402 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2403 sh->dv_meta_mask, reg_c0);
2405 sh->dv_regc0_mask = reg_c0;
2406 DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2407 DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2408 DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2409 DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2413 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2415 static const char *const dynf_names[] = {
2416 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2417 RTE_MBUF_DYNFLAG_METADATA_NAME,
2418 RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
2422 if (n < RTE_DIM(dynf_names))
2424 for (i = 0; i < RTE_DIM(dynf_names); i++) {
2425 if (names[i] == NULL)
2427 strcpy(names[i], dynf_names[i]);
2429 return RTE_DIM(dynf_names);
2433 * Comparison callback to sort device data.
2435 * This is meant to be used with qsort().
2438 * Pointer to pointer to first data object.
2440 * Pointer to pointer to second data object.
2443 * 0 if both objects are equal, less than 0 if the first argument is less
2444 * than the second, greater than 0 otherwise.
2447 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2448 struct mlx5_dev_config *config,
2449 struct rte_device *dpdk_dev)
2451 struct mlx5_dev_ctx_shared *sh = priv->sh;
2452 struct mlx5_dev_config *sh_conf = NULL;
2456 /* Nothing to compare for the single/first device. */
2457 if (sh->refcnt == 1)
2459 /* Find the device with shared context. */
2460 MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
2461 struct mlx5_priv *opriv =
2462 rte_eth_devices[port_id].data->dev_private;
2464 if (opriv && opriv != priv && opriv->sh == sh) {
2465 sh_conf = &opriv->config;
2471 if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2472 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2473 " for shared %s context", sh->ibdev_name);
2477 if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2478 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2479 " for shared %s context", sh->ibdev_name);
2487 * Look for the ethernet device belonging to mlx5 driver.
2489 * @param[in] port_id
2490 * port_id to start looking for device.
2492 * Pointer to the hint device. When device is being probed
2493 * the its siblings (master and preceding representors might
2494 * not have assigned driver yet (because the mlx5_os_pci_probe()
2495 * is not completed yet, for this case match on hint
2496 * device may be used to detect sibling device.
2499 * port_id of found device, RTE_MAX_ETHPORT if not found.
2502 mlx5_eth_find_next(uint16_t port_id, struct rte_device *odev)
2504 while (port_id < RTE_MAX_ETHPORTS) {
2505 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2507 if (dev->state != RTE_ETH_DEV_UNUSED &&
2509 (dev->device == odev ||
2510 (dev->device->driver &&
2511 dev->device->driver->name &&
2512 ((strcmp(dev->device->driver->name,
2513 MLX5_PCI_DRIVER_NAME) == 0) ||
2514 (strcmp(dev->device->driver->name,
2515 MLX5_AUXILIARY_DRIVER_NAME) == 0)))))
2519 if (port_id >= RTE_MAX_ETHPORTS)
2520 return RTE_MAX_ETHPORTS;
2525 * Callback to remove a device.
2527 * This function removes all Ethernet devices belong to a given device.
2530 * Pointer to the generic device.
2533 * 0 on success, the function cannot fail.
2536 mlx5_net_remove(struct mlx5_common_device *cdev)
2541 RTE_ETH_FOREACH_DEV_OF(port_id, cdev->dev) {
2543 * mlx5_dev_close() is not registered to secondary process,
2544 * call the close function explicitly for secondary process.
2546 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2547 ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2549 ret |= rte_eth_dev_close(port_id);
2551 return ret == 0 ? 0 : -EIO;
2554 static const struct rte_pci_id mlx5_pci_id_map[] = {
2556 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2557 PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2560 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2561 PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2564 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2565 PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2568 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2569 PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2572 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2573 PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2576 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2577 PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2580 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2581 PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2584 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2585 PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2588 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2589 PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2592 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2593 PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2596 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2597 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2600 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2601 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2604 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2605 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2608 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2609 PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2612 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2613 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2616 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2617 PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2620 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2621 PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2624 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2625 PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2632 static struct mlx5_class_driver mlx5_net_driver = {
2633 .drv_class = MLX5_CLASS_ETH,
2634 .name = RTE_STR(MLX5_ETH_DRIVER_NAME),
2635 .id_table = mlx5_pci_id_map,
2636 .probe = mlx5_os_net_probe,
2637 .remove = mlx5_net_remove,
2638 .dma_map = mlx5_net_dma_map,
2639 .dma_unmap = mlx5_net_dma_unmap,
2645 /* Initialize driver log type. */
2646 RTE_LOG_REGISTER_DEFAULT(mlx5_logtype, NOTICE)
2649 * Driver initialization routine.
2651 RTE_INIT(rte_mlx5_pmd_init)
2653 pthread_mutex_init(&mlx5_dev_ctx_list_mutex, NULL);
2655 /* Build the static tables for Verbs conversion. */
2656 mlx5_set_ptype_table();
2657 mlx5_set_cksum_table();
2658 mlx5_set_swp_types_table();
2660 mlx5_class_driver_register(&mlx5_net_driver);
2663 RTE_PMD_EXPORT_NAME(MLX5_ETH_DRIVER_NAME, __COUNTER__);
2664 RTE_PMD_REGISTER_PCI_TABLE(MLX5_ETH_DRIVER_NAME, mlx5_pci_id_map);
2665 RTE_PMD_REGISTER_KMOD_DEP(MLX5_ETH_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib");