drivers/net/mlx5/mlx5.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2015 6WIND S.A.
   3  * Copyright 2015 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <stddef.h>
   7 #include <unistd.h>
   8 #include <string.h>
   9 #include <stdint.h>
  10 #include <stdlib.h>
  11 #include <errno.h>
  12
  13 #include <rte_malloc.h>
  14 #include <ethdev_driver.h>
  15 #include <rte_pci.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>
  24
  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>
  31
  32 #include "mlx5_defs.h"
  33 #include "mlx5.h"
  34 #include "mlx5_utils.h"
  35 #include "mlx5_rxtx.h"
  36 #include "mlx5_rx.h"
  37 #include "mlx5_tx.h"
  38 #include "mlx5_autoconf.h"
  39 #include "mlx5_mr.h"
  40 #include "mlx5_flow.h"
  41 #include "mlx5_flow_os.h"
  42 #include "rte_pmd_mlx5.h"
  43
  44 #define MLX5_ETH_DRIVER_NAME mlx5_eth
  45
  46 /* Driver type key for new device global syntax. */
  47 #define MLX5_DRIVER_KEY "driver"
  48
  49 /* Device parameter to enable RX completion queue compression. */
  50 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
  51
  52 /* Device parameter to enable padding Rx packet to cacheline size. */
  53 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
  54
  55 /* Device parameter to enable Multi-Packet Rx queue. */
  56 #define MLX5_RX_MPRQ_EN "mprq_en"
  57
  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"
  60
  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"
  63
  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"
  66
  67 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
  68 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
  69
  70 /* Device parameter to configure inline send. Deprecated, ignored.*/
  71 #define MLX5_TXQ_INLINE "txq_inline"
  72
  73 /* Device parameter to limit packet size to inline with ordinary SEND. */
  74 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
  75
  76 /* Device parameter to configure minimal data size to inline. */
  77 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
  78
  79 /* Device parameter to limit packet size to inline with Enhanced MPW. */
  80 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
  81
  82 /*
  83  * Device parameter to configure the number of TX queues threshold for
  84  * enabling inline send.
  85  */
  86 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
  87
  88 /*
  89  * Device parameter to configure the number of TX queues threshold for
  90  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
  91  */
  92 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
  93
  94 /* Device parameter to enable multi-packet send WQEs. */
  95 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
  96
  97 /*
  98  * Device parameter to force doorbell register mapping
  99  * to non-cahed region eliminating the extra write memory barrier.
 100  */
 101 #define MLX5_TX_DB_NC "tx_db_nc"
 102
 103 /*
 104  * Device parameter to include 2 dsegs in the title WQEBB.
 105  * Deprecated, ignored.
 106  */
 107 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
 108
 109 /*
 110  * Device parameter to limit the size of inlining packet.
 111  * Deprecated, ignored.
 112  */
 113 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
 114
 115 /*
 116  * Device parameter to enable Tx scheduling on timestamps
 117  * and specify the packet pacing granularity in nanoseconds.
 118  */
 119 #define MLX5_TX_PP "tx_pp"
 120
 121 /*
 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.
 125  */
 126 #define MLX5_TX_SKEW "tx_skew"
 127
 128 /*
 129  * Device parameter to enable hardware Tx vector.
 130  * Deprecated, ignored (no vectorized Tx routines anymore).
 131  */
 132 #define MLX5_TX_VEC_EN "tx_vec_en"
 133
 134 /* Device parameter to enable hardware Rx vector. */
 135 #define MLX5_RX_VEC_EN "rx_vec_en"
 136
 137 /* Allow L3 VXLAN flow creation. */
 138 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
 139
 140 /* Activate DV E-Switch flow steering. */
 141 #define MLX5_DV_ESW_EN "dv_esw_en"
 142
 143 /* Activate DV flow steering. */
 144 #define MLX5_DV_FLOW_EN "dv_flow_en"
 145
 146 /* Enable extensive flow metadata support. */
 147 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
 148
 149 /* Device parameter to let the user manage the lacp traffic of bonded device */
 150 #define MLX5_LACP_BY_USER "lacp_by_user"
 151
 152 /* Activate Netlink support in VF mode. */
 153 #define MLX5_VF_NL_EN "vf_nl_en"
 154
 155 /* Enable extending memsegs when creating a MR. */
 156 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
 157
 158 /* Select port representors to instantiate. */
 159 #define MLX5_REPRESENTOR "representor"
 160
 161 /* Device parameter to configure the maximum number of dump files per queue. */
 162 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
 163
 164 /* Configure timeout of LRO session (in microseconds). */
 165 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
 166
 167 /*
 168  * Device parameter to configure the total data buffer size for a single
 169  * hairpin queue (logarithm value).
 170  */
 171 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
 172
 173 /* Flow memory reclaim mode. */
 174 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
 175
 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"
 180
 181 /* Device parameter to configure allow or prevent duplicate rules pattern. */
 182 #define MLX5_ALLOW_DUPLICATE_PATTERN "allow_duplicate_pattern"
 183
 184 /* Device parameter to configure implicit registration of mempool memory. */
 185 #define MLX5_MR_MEMPOOL_REG_EN "mr_mempool_reg_en"
 186
 187 /* Shared memory between primary and secondary processes. */
 188 struct mlx5_shared_data *mlx5_shared_data;
 189
 190 /** Driver-specific log messages type. */
 191 int mlx5_logtype;
 192
 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),
 200                 .trunk_size = 64,
 201                 .grow_trunk = 3,
 202                 .grow_shift = 2,
 203                 .need_lock = 1,
 204                 .release_mem_en = 1,
 205                 .malloc = mlx5_malloc,
 206                 .free = mlx5_free,
 207                 .type = "mlx5_encap_decap_ipool",
 208         },
 209         [MLX5_IPOOL_PUSH_VLAN] = {
 210                 .size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
 211                 .trunk_size = 64,
 212                 .grow_trunk = 3,
 213                 .grow_shift = 2,
 214                 .need_lock = 1,
 215                 .release_mem_en = 1,
 216                 .malloc = mlx5_malloc,
 217                 .free = mlx5_free,
 218                 .type = "mlx5_push_vlan_ipool",
 219         },
 220         [MLX5_IPOOL_TAG] = {
 221                 .size = sizeof(struct mlx5_flow_dv_tag_resource),
 222                 .trunk_size = 64,
 223                 .grow_trunk = 3,
 224                 .grow_shift = 2,
 225                 .need_lock = 1,
 226                 .release_mem_en = 0,
 227                 .per_core_cache = (1 << 16),
 228                 .malloc = mlx5_malloc,
 229                 .free = mlx5_free,
 230                 .type = "mlx5_tag_ipool",
 231         },
 232         [MLX5_IPOOL_PORT_ID] = {
 233                 .size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
 234                 .trunk_size = 64,
 235                 .grow_trunk = 3,
 236                 .grow_shift = 2,
 237                 .need_lock = 1,
 238                 .release_mem_en = 1,
 239                 .malloc = mlx5_malloc,
 240                 .free = mlx5_free,
 241                 .type = "mlx5_port_id_ipool",
 242         },
 243         [MLX5_IPOOL_JUMP] = {
 244                 .size = sizeof(struct mlx5_flow_tbl_data_entry),
 245                 .trunk_size = 64,
 246                 .grow_trunk = 3,
 247                 .grow_shift = 2,
 248                 .need_lock = 1,
 249                 .release_mem_en = 1,
 250                 .malloc = mlx5_malloc,
 251                 .free = mlx5_free,
 252                 .type = "mlx5_jump_ipool",
 253         },
 254         [MLX5_IPOOL_SAMPLE] = {
 255                 .size = sizeof(struct mlx5_flow_dv_sample_resource),
 256                 .trunk_size = 64,
 257                 .grow_trunk = 3,
 258                 .grow_shift = 2,
 259                 .need_lock = 1,
 260                 .release_mem_en = 1,
 261                 .malloc = mlx5_malloc,
 262                 .free = mlx5_free,
 263                 .type = "mlx5_sample_ipool",
 264         },
 265         [MLX5_IPOOL_DEST_ARRAY] = {
 266                 .size = sizeof(struct mlx5_flow_dv_dest_array_resource),
 267                 .trunk_size = 64,
 268                 .grow_trunk = 3,
 269                 .grow_shift = 2,
 270                 .need_lock = 1,
 271                 .release_mem_en = 1,
 272                 .malloc = mlx5_malloc,
 273                 .free = mlx5_free,
 274                 .type = "mlx5_dest_array_ipool",
 275         },
 276         [MLX5_IPOOL_TUNNEL_ID] = {
 277                 .size = sizeof(struct mlx5_flow_tunnel),
 278                 .trunk_size = MLX5_MAX_TUNNELS,
 279                 .need_lock = 1,
 280                 .release_mem_en = 1,
 281                 .type = "mlx5_tunnel_offload",
 282         },
 283         [MLX5_IPOOL_TNL_TBL_ID] = {
 284                 .size = 0,
 285                 .need_lock = 1,
 286                 .type = "mlx5_flow_tnl_tbl_ipool",
 287         },
 288 #endif
 289         [MLX5_IPOOL_MTR] = {
 290                 /**
 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.
 294                  */
 295                 .size = sizeof(struct mlx5_legacy_flow_meter),
 296                 .trunk_size = 64,
 297                 .need_lock = 1,
 298                 .release_mem_en = 1,
 299                 .malloc = mlx5_malloc,
 300                 .free = mlx5_free,
 301                 .type = "mlx5_meter_ipool",
 302         },
 303         [MLX5_IPOOL_MCP] = {
 304                 .size = sizeof(struct mlx5_flow_mreg_copy_resource),
 305                 .trunk_size = 64,
 306                 .grow_trunk = 3,
 307                 .grow_shift = 2,
 308                 .need_lock = 1,
 309                 .release_mem_en = 1,
 310                 .malloc = mlx5_malloc,
 311                 .free = mlx5_free,
 312                 .type = "mlx5_mcp_ipool",
 313         },
 314         [MLX5_IPOOL_HRXQ] = {
 315                 .size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
 316                 .trunk_size = 64,
 317                 .grow_trunk = 3,
 318                 .grow_shift = 2,
 319                 .need_lock = 1,
 320                 .release_mem_en = 1,
 321                 .malloc = mlx5_malloc,
 322                 .free = mlx5_free,
 323                 .type = "mlx5_hrxq_ipool",
 324         },
 325         [MLX5_IPOOL_MLX5_FLOW] = {
 326                 /*
 327                  * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
 328                  * It set in run time according to PCI function configuration.
 329                  */
 330                 .size = 0,
 331                 .trunk_size = 64,
 332                 .grow_trunk = 3,
 333                 .grow_shift = 2,
 334                 .need_lock = 1,
 335                 .release_mem_en = 0,
 336                 .per_core_cache = 1 << 19,
 337                 .malloc = mlx5_malloc,
 338                 .free = mlx5_free,
 339                 .type = "mlx5_flow_handle_ipool",
 340         },
 341         [MLX5_IPOOL_RTE_FLOW] = {
 342                 .size = sizeof(struct rte_flow),
 343                 .trunk_size = 4096,
 344                 .need_lock = 1,
 345                 .release_mem_en = 1,
 346                 .malloc = mlx5_malloc,
 347                 .free = mlx5_free,
 348                 .type = "rte_flow_ipool",
 349         },
 350         [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
 351                 .size = 0,
 352                 .need_lock = 1,
 353                 .type = "mlx5_flow_rss_id_ipool",
 354         },
 355         [MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
 356                 .size = sizeof(struct mlx5_shared_action_rss),
 357                 .trunk_size = 64,
 358                 .grow_trunk = 3,
 359                 .grow_shift = 2,
 360                 .need_lock = 1,
 361                 .release_mem_en = 1,
 362                 .malloc = mlx5_malloc,
 363                 .free = mlx5_free,
 364                 .type = "mlx5_shared_action_rss",
 365         },
 366         [MLX5_IPOOL_MTR_POLICY] = {
 367                 /**
 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.
 371                  */
 372                 .size = sizeof(struct mlx5_flow_meter_sub_policy),
 373                 .trunk_size = 64,
 374                 .need_lock = 1,
 375                 .release_mem_en = 1,
 376                 .malloc = mlx5_malloc,
 377                 .free = mlx5_free,
 378                 .type = "mlx5_meter_policy_ipool",
 379         },
 380 };
 381
 382
 383 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
 384 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
 385
 386 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 1024
 387
 388 /**
 389  * Decide whether representor ID is a HPF(host PF) port on BF2.
 390  *
 391  * @param dev
 392  *   Pointer to Ethernet device structure.
 393  *
 394  * @return
 395  *   Non-zero if HPF, otherwise 0.
 396  */
 397 bool
 398 mlx5_is_hpf(struct rte_eth_dev *dev)
 399 {
 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);
 403
 404         return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_VF &&
 405                MLX5_REPRESENTOR_REPR(-1) == repr;
 406 }
 407
 408 /**
 409  * Decide whether representor ID is a SF port representor.
 410  *
 411  * @param dev
 412  *   Pointer to Ethernet device structure.
 413  *
 414  * @return
 415  *   Non-zero if HPF, otherwise 0.
 416  */
 417 bool
 418 mlx5_is_sf_repr(struct rte_eth_dev *dev)
 419 {
 420         struct mlx5_priv *priv = dev->data->dev_private;
 421         int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
 422
 423         return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_SF;
 424 }
 425
 426 /**
 427  * Initialize the ASO aging management structure.
 428  *
 429  * @param[in] sh
 430  *   Pointer to mlx5_dev_ctx_shared object to free
 431  *
 432  * @return
 433  *   0 on success, a negative errno value otherwise and rte_errno is set.
 434  */
 435 int
 436 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
 437 {
 438         int err;
 439
 440         if (sh->aso_age_mng)
 441                 return 0;
 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.");
 446                 rte_errno = ENOMEM;
 447                 return -ENOMEM;
 448         }
 449         err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_FLOW_HIT);
 450         if (err) {
 451                 mlx5_free(sh->aso_age_mng);
 452                 return -1;
 453         }
 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);
 457         return 0;
 458 }
 459
 460 /**
 461  * Close and release all the resources of the ASO aging management structure.
 462  *
 463  * @param[in] sh
 464  *   Pointer to mlx5_dev_ctx_shared object to free.
 465  */
 466 static void
 467 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
 468 {
 469         int i, j;
 470
 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;
 475
 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)
 482                                         claim_zero
 483                                             (mlx5_flow_os_destroy_flow_action
 484                                               (pool->actions[j].dr_action));
 485                         mlx5_free(pool);
 486                 }
 487                 mlx5_free(sh->aso_age_mng->pools);
 488         }
 489         mlx5_free(sh->aso_age_mng);
 490 }
 491
 492 /**
 493  * Initialize the shared aging list information per port.
 494  *
 495  * @param[in] sh
 496  *   Pointer to mlx5_dev_ctx_shared object.
 497  */
 498 static void
 499 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
 500 {
 501         uint32_t i;
 502         struct mlx5_age_info *age_info;
 503
 504         for (i = 0; i < sh->max_port; i++) {
 505                 age_info = &sh->port[i].age_info;
 506                 age_info->flags = 0;
 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);
 511         }
 512 }
 513
 514 /**
 515  * Initialize the counters management structure.
 516  *
 517  * @param[in] sh
 518  *   Pointer to mlx5_dev_ctx_shared object to free
 519  */
 520 static void
 521 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
 522 {
 523         int i;
 524
 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]);
 534         }
 535 }
 536
 537 /**
 538  * Destroy all the resources allocated for a counter memory management.
 539  *
 540  * @param[in] mng
 541  *   Pointer to the memory management structure.
 542  */
 543 static void
 544 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
 545 {
 546         uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
 547
 548         LIST_REMOVE(mng, next);
 549         claim_zero(mlx5_devx_cmd_destroy(mng->dm));
 550         claim_zero(mlx5_os_umem_dereg(mng->umem));
 551         mlx5_free(mem);
 552 }
 553
 554 /**
 555  * Close and release all the resources of the counters management.
 556  *
 557  * @param[in] sh
 558  *   Pointer to mlx5_dev_ctx_shared object to free.
 559  */
 560 static void
 561 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
 562 {
 563         struct mlx5_counter_stats_mem_mng *mng;
 564         int i, j;
 565         int retries = 1024;
 566
 567         rte_errno = 0;
 568         while (--retries) {
 569                 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
 570                 if (rte_errno != EINPROGRESS)
 571                         break;
 572                 rte_pause();
 573         }
 574
 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;
 579
 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
 584                                                                (pool->min_dcs));
 585                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
 586                                 struct mlx5_flow_counter *cnt =
 587                                                 MLX5_POOL_GET_CNT(pool, j);
 588
 589                                 if (cnt->action)
 590                                         claim_zero
 591                                          (mlx5_flow_os_destroy_flow_action
 592                                           (cnt->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));
 597                         }
 598                         mlx5_free(pool);
 599                 }
 600                 mlx5_free(sh->cmng.pools);
 601         }
 602         mng = LIST_FIRST(&sh->cmng.mem_mngs);
 603         while (mng) {
 604                 mlx5_flow_destroy_counter_stat_mem_mng(mng);
 605                 mng = LIST_FIRST(&sh->cmng.mem_mngs);
 606         }
 607         memset(&sh->cmng, 0, sizeof(sh->cmng));
 608 }
 609
 610 /**
 611  * Initialize the aso flow meters management structure.
 612  *
 613  * @param[in] sh
 614  *   Pointer to mlx5_dev_ctx_shared object to free
 615  */
 616 int
 617 mlx5_aso_flow_mtrs_mng_init(struct mlx5_dev_ctx_shared *sh)
 618 {
 619         if (!sh->mtrmng) {
 620                 sh->mtrmng = mlx5_malloc(MLX5_MEM_ZERO,
 621                         sizeof(*sh->mtrmng),
 622                         RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
 623                 if (!sh->mtrmng) {
 624                         DRV_LOG(ERR,
 625                         "meter management allocation was failed.");
 626                         rte_errno = ENOMEM;
 627                         return -ENOMEM;
 628                 }
 629                 if (sh->meter_aso_en) {
 630                         rte_spinlock_init(&sh->mtrmng->pools_mng.mtrsl);
 631                         LIST_INIT(&sh->mtrmng->pools_mng.meters);
 632                 }
 633                 sh->mtrmng->def_policy_id = MLX5_INVALID_POLICY_ID;
 634         }
 635         return 0;
 636 }
 637
 638 /**
 639  * Close and release all the resources of
 640  * the ASO flow meter management structure.
 641  *
 642  * @param[in] sh
 643  *   Pointer to mlx5_dev_ctx_shared object to free.
 644  */
 645 static void
 646 mlx5_aso_flow_mtrs_mng_close(struct mlx5_dev_ctx_shared *sh)
 647 {
 648         struct mlx5_aso_mtr_pool *mtr_pool;
 649         struct mlx5_flow_mtr_mng *mtrmng = sh->mtrmng;
 650         uint32_t idx;
 651 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
 652         struct mlx5_aso_mtr *aso_mtr;
 653         int i;
 654 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
 655
 656         if (sh->meter_aso_en) {
 657                 mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_POLICER);
 658                 idx = mtrmng->pools_mng.n_valid;
 659                 while (idx--) {
 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)
 665                                         claim_zero
 666                                         (mlx5_glue->destroy_flow_action
 667                                         (aso_mtr->fm.meter_action));
 668                         }
 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--;
 673                         mlx5_free(mtr_pool);
 674                 }
 675                 mlx5_free(sh->mtrmng->pools_mng.pools);
 676         }
 677         mlx5_free(sh->mtrmng);
 678         sh->mtrmng = NULL;
 679 }
 680
 681 /* Send FLOW_AGED event if needed. */
 682 void
 683 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
 684 {
 685         struct mlx5_age_info *age_info;
 686         uint32_t i;
 687
 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))
 691                         continue;
 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);
 698                 }
 699         }
 700 }
 701
 702 /*
 703  * Initialize the ASO connection tracking structure.
 704  *
 705  * @param[in] sh
 706  *   Pointer to mlx5_dev_ctx_shared object.
 707  *
 708  * @return
 709  *   0 on success, a negative errno value otherwise and rte_errno is set.
 710  */
 711 int
 712 mlx5_flow_aso_ct_mng_init(struct mlx5_dev_ctx_shared *sh)
 713 {
 714         int err;
 715
 716         if (sh->ct_mng)
 717                 return 0;
 718         sh->ct_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->ct_mng),
 719                                  RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
 720         if (!sh->ct_mng) {
 721                 DRV_LOG(ERR, "ASO CT management allocation failed.");
 722                 rte_errno = ENOMEM;
 723                 return -rte_errno;
 724         }
 725         err = mlx5_aso_queue_init(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
 726         if (err) {
 727                 mlx5_free(sh->ct_mng);
 728                 /* rte_errno should be extracted from the failure. */
 729                 rte_errno = EINVAL;
 730                 return -rte_errno;
 731         }
 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);
 735         return 0;
 736 }
 737
 738 /*
 739  * Close and release all the resources of the
 740  * ASO connection tracking management structure.
 741  *
 742  * @param[in] sh
 743  *   Pointer to mlx5_dev_ctx_shared object to free.
 744  */
 745 static void
 746 mlx5_flow_aso_ct_mng_close(struct mlx5_dev_ctx_shared *sh)
 747 {
 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;
 751         uint32_t idx;
 752         uint32_t val;
 753         uint32_t cnt;
 754         int i;
 755
 756         mlx5_aso_queue_uninit(sh, ASO_OPC_MOD_CONNECTION_TRACKING);
 757         idx = mng->next;
 758         while (idx--) {
 759                 cnt = 0;
 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,
 764                                                  __ATOMIC_RELAXED);
 765                         MLX5_ASSERT(val == 1);
 766                         if (val > 1)
 767                                 cnt++;
 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));
 775 #endif
 776                 }
 777                 claim_zero(mlx5_devx_cmd_destroy(ct_pool->devx_obj));
 778                 if (cnt) {
 779                         DRV_LOG(DEBUG, "%u ASO CT objects are being used in the pool %u",
 780                                 cnt, i);
 781                 }
 782                 mlx5_free(ct_pool);
 783                 /* in case of failure. */
 784                 mng->next--;
 785         }
 786         mlx5_free(mng->pools);
 787         mlx5_free(mng);
 788         /* Management structure must be cleared to 0s during allocation. */
 789         sh->ct_mng = NULL;
 790 }
 791
 792 /**
 793  * Initialize the flow resources' indexed mempool.
 794  *
 795  * @param[in] sh
 796  *   Pointer to mlx5_dev_ctx_shared object.
 797  * @param[in] config
 798  *   Pointer to user dev config.
 799  */
 800 static void
 801 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
 802                        const struct mlx5_dev_config *config)
 803 {
 804         uint8_t i;
 805         struct mlx5_indexed_pool_config cfg;
 806
 807         for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
 808                 cfg = mlx5_ipool_cfg[i];
 809                 switch (i) {
 810                 default:
 811                         break;
 812                 /*
 813                  * Set MLX5_IPOOL_MLX5_FLOW ipool size
 814                  * according to PCI function flow configuration.
 815                  */
 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;
 820                         break;
 821                 }
 822                 if (config->reclaim_mode) {
 823                         cfg.release_mem_en = 1;
 824                         cfg.per_core_cache = 0;
 825                 } else {
 826                         cfg.release_mem_en = 0;
 827                 }
 828                 sh->ipool[i] = mlx5_ipool_create(&cfg);
 829         }
 830 }
 831
 832
 833 /**
 834  * Release the flow resources' indexed mempool.
 835  *
 836  * @param[in] sh
 837  *   Pointer to mlx5_dev_ctx_shared object.
 838  */
 839 static void
 840 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
 841 {
 842         uint8_t i;
 843
 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]);
 849 }
 850
 851 /*
 852  * Check if dynamic flex parser for eCPRI already exists.
 853  *
 854  * @param dev
 855  *   Pointer to Ethernet device structure.
 856  *
 857  * @return
 858  *   true on exists, false on not.
 859  */
 860 bool
 861 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
 862 {
 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];
 866
 867         return !!prf->obj;
 868 }
 869
 870 /*
 871  * Allocation of a flex parser for eCPRI. Once created, this parser related
 872  * resources will be held until the device is closed.
 873  *
 874  * @param dev
 875  *   Pointer to Ethernet device structure.
 876  *
 877  * @return
 878  *   0 on success, a negative errno value otherwise and rte_errno is set.
 879  */
 880 int
 881 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
 882 {
 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,
 888         };
 889         uint32_t ids[8];
 890         int ret;
 891
 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);
 895                 return -ENOTSUP;
 896         }
 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;
 906         /* Fixed offset. */
 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;
 912         /* Fixed offset. */
 913         node.sample[1].flow_match_sample_offset_mode = 0x0;
 914         /*
 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).
 917          */
 918         node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
 919         prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
 920         if (!prf->obj) {
 921                 DRV_LOG(ERR, "Failed to create flex parser node object.");
 922                 return (rte_errno == 0) ? -ENODEV : -rte_errno;
 923         }
 924         prf->num = 2;
 925         ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
 926         if (ret) {
 927                 DRV_LOG(ERR, "Failed to query sample IDs.");
 928                 return (rte_errno == 0) ? -ENODEV : -rte_errno;
 929         }
 930         prf->offset[0] = 0x0;
 931         prf->offset[1] = sizeof(uint32_t);
 932         prf->ids[0] = ids[0];
 933         prf->ids[1] = ids[1];
 934         return 0;
 935 }
 936
 937 /*
 938  * Destroy the flex parser node, including the parser itself, input / output
 939  * arcs and DW samples. Resources could be reused then.
 940  *
 941  * @param dev
 942  *   Pointer to Ethernet device structure.
 943  */
 944 static void
 945 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
 946 {
 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];
 950
 951         if (prf->obj)
 952                 mlx5_devx_cmd_destroy(prf->obj);
 953         prf->obj = NULL;
 954 }
 955
 956 uint32_t
 957 mlx5_get_supported_sw_parsing_offloads(const struct mlx5_hca_attr *attr)
 958 {
 959         uint32_t sw_parsing_offloads = 0;
 960
 961         if (attr->swp) {
 962                 sw_parsing_offloads |= MLX5_SW_PARSING_CAP;
 963                 if (attr->swp_csum)
 964                         sw_parsing_offloads |= MLX5_SW_PARSING_CSUM_CAP;
 965
 966                 if (attr->swp_lso)
 967                         sw_parsing_offloads |= MLX5_SW_PARSING_TSO_CAP;
 968         }
 969         return sw_parsing_offloads;
 970 }
 971
 972 uint32_t
 973 mlx5_get_supported_tunneling_offloads(const struct mlx5_hca_attr *attr)
 974 {
 975         uint32_t tn_offloads = 0;
 976
 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;
 983         return tn_offloads;
 984 }
 985
 986 /*
 987  * Allocate Rx and Tx UARs in robust fashion.
 988  * This routine handles the following UAR allocation issues:
 989  *
 990  *  - tries to allocate the UAR with the most appropriate memory
 991  *    mapping type from the ones supported by the host
 992  *
 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.
 998  */
 999 static int
1000 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
1001                      const struct mlx5_dev_config *config)
1002 {
1003         uint32_t uar_mapping, retry;
1004         int err = 0;
1005         void *base_addr;
1006
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;
1013 #else
1014                 RTE_SET_USED(config);
1015                 /*
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.
1021                  */
1022                 uar_mapping = 0;
1023 #endif
1024                 sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, uar_mapping);
1025 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1026                 if (!sh->tx_uar &&
1027                     uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
1028                         if (config->dbnc == MLX5_TXDB_CACHED ||
1029                             config->dbnc == MLX5_TXDB_HEURISTIC)
1030                                 DRV_LOG(WARNING, "Devarg tx_db_nc setting "
1031                                                  "is not supported by DevX");
1032                         /*
1033                          * In some environments like virtual machine
1034                          * the Write Combining mapped might be not supported
1035                          * and UAR allocation fails. We try "Non-Cached"
1036                          * mapping for the case. The tx_burst routines take
1037                          * the UAR mapping type into account on UAR setup
1038                          * on queue creation.
1039                          */
1040                         DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (BF)");
1041                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
1042                         sh->tx_uar = mlx5_glue->devx_alloc_uar
1043                                                         (sh->ctx, uar_mapping);
1044                 } else if (!sh->tx_uar &&
1045                            uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
1046                         if (config->dbnc == MLX5_TXDB_NCACHED)
1047                                 DRV_LOG(WARNING, "Devarg tx_db_nc settings "
1048                                                  "is not supported by DevX");
1049                         /*
1050                          * If Verbs/kernel does not support "Non-Cached"
1051                          * try the "Write-Combining".
1052                          */
1053                         DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (NC)");
1054                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
1055                         sh->tx_uar = mlx5_glue->devx_alloc_uar
1056                                                         (sh->ctx, uar_mapping);
1057                 }
1058 #endif
1059                 if (!sh->tx_uar) {
1060                         DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
1061                         err = ENOMEM;
1062                         goto exit;
1063                 }
1064                 base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
1065                 if (base_addr)
1066                         break;
1067                 /*
1068                  * The UARs are allocated by rdma_core within the
1069                  * IB device context, on context closure all UARs
1070                  * will be freed, should be no memory/object leakage.
1071                  */
1072                 DRV_LOG(DEBUG, "Retrying to allocate Tx DevX UAR");
1073                 sh->tx_uar = NULL;
1074         }
1075         /* Check whether we finally succeeded with valid UAR allocation. */
1076         if (!sh->tx_uar) {
1077                 DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
1078                 err = ENOMEM;
1079                 goto exit;
1080         }
1081         for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
1082                 uar_mapping = 0;
1083                 sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
1084                                                         (sh->ctx, uar_mapping);
1085 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
1086                 if (!sh->devx_rx_uar &&
1087                     uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
1088                         /*
1089                          * Rx UAR is used to control interrupts only,
1090                          * should be no datapath noticeable impact,
1091                          * can try "Non-Cached" mapping safely.
1092                          */
1093                         DRV_LOG(DEBUG, "Failed to allocate Rx DevX UAR (BF)");
1094                         uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
1095                         sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
1096                                                         (sh->ctx, uar_mapping);
1097                 }
1098 #endif
1099                 if (!sh->devx_rx_uar) {
1100                         DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
1101                         err = ENOMEM;
1102                         goto exit;
1103                 }
1104                 base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
1105                 if (base_addr)
1106                         break;
1107                 /*
1108                  * The UARs are allocated by rdma_core within the
1109                  * IB device context, on context closure all UARs
1110                  * will be freed, should be no memory/object leakage.
1111                  */
1112                 DRV_LOG(DEBUG, "Retrying to allocate Rx DevX UAR");
1113                 sh->devx_rx_uar = NULL;
1114         }
1115         /* Check whether we finally succeeded with valid UAR allocation. */
1116         if (!sh->devx_rx_uar) {
1117                 DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
1118                 err = ENOMEM;
1119         }
1120 exit:
1121         return err;
1122 }
1123
1124 /**
1125  * Unregister the mempool from the protection domain.
1126  *
1127  * @param sh
1128  *   Pointer to the device shared context.
1129  * @param mp
1130  *   Mempool being unregistered.
1131  */
1132 static void
1133 mlx5_dev_ctx_shared_mempool_unregister(struct mlx5_dev_ctx_shared *sh,
1134                                        struct rte_mempool *mp)
1135 {
1136         struct mlx5_mp_id mp_id;
1137
1138         mlx5_mp_id_init(&mp_id, 0);
1139         if (mlx5_mr_mempool_unregister(&sh->share_cache, mp, &mp_id) < 0)
1140                 DRV_LOG(WARNING, "Failed to unregister mempool %s for PD %p: %s",
1141                         mp->name, sh->pd, rte_strerror(rte_errno));
1142 }
1143
1144 /**
1145  * rte_mempool_walk() callback to register mempools
1146  * for the protection domain.
1147  *
1148  * @param mp
1149  *   The mempool being walked.
1150  * @param arg
1151  *   Pointer to the device shared context.
1152  */
1153 static void
1154 mlx5_dev_ctx_shared_mempool_register_cb(struct rte_mempool *mp, void *arg)
1155 {
1156         struct mlx5_dev_ctx_shared *sh = arg;
1157         struct mlx5_mp_id mp_id;
1158         int ret;
1159
1160         mlx5_mp_id_init(&mp_id, 0);
1161         ret = mlx5_mr_mempool_register(&sh->share_cache, sh->pd, mp, &mp_id);
1162         if (ret < 0 && rte_errno != EEXIST)
1163                 DRV_LOG(ERR, "Failed to register existing mempool %s for PD %p: %s",
1164                         mp->name, sh->pd, rte_strerror(rte_errno));
1165 }
1166
1167 /**
1168  * rte_mempool_walk() callback to unregister mempools
1169  * from the protection domain.
1170  *
1171  * @param mp
1172  *   The mempool being walked.
1173  * @param arg
1174  *   Pointer to the device shared context.
1175  */
1176 static void
1177 mlx5_dev_ctx_shared_mempool_unregister_cb(struct rte_mempool *mp, void *arg)
1178 {
1179         mlx5_dev_ctx_shared_mempool_unregister
1180                                 ((struct mlx5_dev_ctx_shared *)arg, mp);
1181 }
1182
1183 /**
1184  * Mempool life cycle callback for Ethernet devices.
1185  *
1186  * @param event
1187  *   Mempool life cycle event.
1188  * @param mp
1189  *   Associated mempool.
1190  * @param arg
1191  *   Pointer to a device shared context.
1192  */
1193 static void
1194 mlx5_dev_ctx_shared_mempool_event_cb(enum rte_mempool_event event,
1195                                      struct rte_mempool *mp, void *arg)
1196 {
1197         struct mlx5_dev_ctx_shared *sh = arg;
1198         struct mlx5_mp_id mp_id;
1199
1200         switch (event) {
1201         case RTE_MEMPOOL_EVENT_READY:
1202                 mlx5_mp_id_init(&mp_id, 0);
1203                 if (mlx5_mr_mempool_register(&sh->share_cache, sh->pd, mp,
1204                                              &mp_id) < 0)
1205                         DRV_LOG(ERR, "Failed to register new mempool %s for PD %p: %s",
1206                                 mp->name, sh->pd, rte_strerror(rte_errno));
1207                 break;
1208         case RTE_MEMPOOL_EVENT_DESTROY:
1209                 mlx5_dev_ctx_shared_mempool_unregister(sh, mp);
1210                 break;
1211         }
1212 }
1213
1214 /**
1215  * Callback used when implicit mempool registration is disabled
1216  * in order to track Rx mempool destruction.
1217  *
1218  * @param event
1219  *   Mempool life cycle event.
1220  * @param mp
1221  *   An Rx mempool registered explicitly when the port is started.
1222  * @param arg
1223  *   Pointer to a device shared context.
1224  */
1225 static void
1226 mlx5_dev_ctx_shared_rx_mempool_event_cb(enum rte_mempool_event event,
1227                                         struct rte_mempool *mp, void *arg)
1228 {
1229         struct mlx5_dev_ctx_shared *sh = arg;
1230
1231         if (event == RTE_MEMPOOL_EVENT_DESTROY)
1232                 mlx5_dev_ctx_shared_mempool_unregister(sh, mp);
1233 }
1234
1235 int
1236 mlx5_dev_ctx_shared_mempool_subscribe(struct rte_eth_dev *dev)
1237 {
1238         struct mlx5_priv *priv = dev->data->dev_private;
1239         struct mlx5_dev_ctx_shared *sh = priv->sh;
1240         int ret;
1241
1242         /* Check if we only need to track Rx mempool destruction. */
1243         if (!priv->config.mr_mempool_reg_en) {
1244                 ret = rte_mempool_event_callback_register
1245                                 (mlx5_dev_ctx_shared_rx_mempool_event_cb, sh);
1246                 return ret == 0 || rte_errno == EEXIST ? 0 : ret;
1247         }
1248         /* Callback for this shared context may be already registered. */
1249         ret = rte_mempool_event_callback_register
1250                                 (mlx5_dev_ctx_shared_mempool_event_cb, sh);
1251         if (ret != 0 && rte_errno != EEXIST)
1252                 return ret;
1253         /* Register mempools only once for this shared context. */
1254         if (ret == 0)
1255                 rte_mempool_walk(mlx5_dev_ctx_shared_mempool_register_cb, sh);
1256         return 0;
1257 }
1258
1259 /**
1260  * Allocate shared device context. If there is multiport device the
1261  * master and representors will share this context, if there is single
1262  * port dedicated device, the context will be used by only given
1263  * port due to unification.
1264  *
1265  * Routine first searches the context for the specified device name,
1266  * if found the shared context assumed and reference counter is incremented.
1267  * If no context found the new one is created and initialized with specified
1268  * device context and parameters.
1269  *
1270  * @param[in] spawn
1271  *   Pointer to the device attributes (name, port, etc).
1272  * @param[in] config
1273  *   Pointer to device configuration structure.
1274  *
1275  * @return
1276  *   Pointer to mlx5_dev_ctx_shared object on success,
1277  *   otherwise NULL and rte_errno is set.
1278  */
1279 struct mlx5_dev_ctx_shared *
1280 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
1281                            const struct mlx5_dev_config *config)
1282 {
1283         struct mlx5_dev_ctx_shared *sh;
1284         int err = 0;
1285         uint32_t i;
1286         struct mlx5_devx_tis_attr tis_attr = { 0 };
1287
1288         MLX5_ASSERT(spawn);
1289         /* Secondary process should not create the shared context. */
1290         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1291         pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1292         /* Search for IB context by device name. */
1293         LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
1294                 if (!strcmp(sh->ibdev_name,
1295                         mlx5_os_get_dev_device_name(spawn->phys_dev))) {
1296                         sh->refcnt++;
1297                         goto exit;
1298                 }
1299         }
1300         /* No device found, we have to create new shared context. */
1301         MLX5_ASSERT(spawn->max_port);
1302         sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1303                          sizeof(struct mlx5_dev_ctx_shared) +
1304                          spawn->max_port *
1305                          sizeof(struct mlx5_dev_shared_port),
1306                          RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1307         if (!sh) {
1308                 DRV_LOG(ERR, "shared context allocation failure");
1309                 rte_errno  = ENOMEM;
1310                 goto exit;
1311         }
1312         sh->numa_node = spawn->cdev->dev->numa_node;
1313         sh->cdev = spawn->cdev;
1314         if (spawn->bond_info)
1315                 sh->bond = *spawn->bond_info;
1316         err = mlx5_os_open_device(spawn, config, sh);
1317         if (!sh->ctx)
1318                 goto error;
1319         err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
1320         if (err) {
1321                 DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
1322                 goto error;
1323         }
1324         sh->refcnt = 1;
1325         sh->max_port = spawn->max_port;
1326         sh->reclaim_mode = config->reclaim_mode;
1327         strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
1328                 sizeof(sh->ibdev_name) - 1);
1329         strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
1330                 sizeof(sh->ibdev_path) - 1);
1331         /*
1332          * Setting port_id to max unallowed value means
1333          * there is no interrupt subhandler installed for
1334          * the given port index i.
1335          */
1336         for (i = 0; i < sh->max_port; i++) {
1337                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
1338                 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
1339         }
1340         sh->pd = mlx5_os_alloc_pd(sh->ctx);
1341         if (sh->pd == NULL) {
1342                 DRV_LOG(ERR, "PD allocation failure");
1343                 err = ENOMEM;
1344                 goto error;
1345         }
1346         if (sh->devx) {
1347                 err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
1348                 if (err) {
1349                         DRV_LOG(ERR, "Fail to extract pdn from PD");
1350                         goto error;
1351                 }
1352                 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
1353                 if (!sh->td) {
1354                         DRV_LOG(ERR, "TD allocation failure");
1355                         err = ENOMEM;
1356                         goto error;
1357                 }
1358                 tis_attr.transport_domain = sh->td->id;
1359                 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
1360                 if (!sh->tis) {
1361                         DRV_LOG(ERR, "TIS allocation failure");
1362                         err = ENOMEM;
1363                         goto error;
1364                 }
1365                 err = mlx5_alloc_rxtx_uars(sh, config);
1366                 if (err)
1367                         goto error;
1368                 MLX5_ASSERT(sh->tx_uar);
1369                 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
1370
1371                 MLX5_ASSERT(sh->devx_rx_uar);
1372                 MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
1373         }
1374 #ifndef RTE_ARCH_64
1375         /* Initialize UAR access locks for 32bit implementations. */
1376         rte_spinlock_init(&sh->uar_lock_cq);
1377         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1378                 rte_spinlock_init(&sh->uar_lock[i]);
1379 #endif
1380         /*
1381          * Once the device is added to the list of memory event
1382          * callback, its global MR cache table cannot be expanded
1383          * on the fly because of deadlock. If it overflows, lookup
1384          * should be done by searching MR list linearly, which is slow.
1385          *
1386          * At this point the device is not added to the memory
1387          * event list yet, context is just being created.
1388          */
1389         err = mlx5_mr_btree_init(&sh->share_cache.cache,
1390                                  MLX5_MR_BTREE_CACHE_N * 2,
1391                                  sh->numa_node);
1392         if (err) {
1393                 err = rte_errno;
1394                 goto error;
1395         }
1396         mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
1397                               &sh->share_cache.dereg_mr_cb);
1398         mlx5_os_dev_shared_handler_install(sh);
1399         if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1400                 err = mlx5_flow_os_init_workspace_once();
1401                 if (err)
1402                         goto error;
1403         }
1404         mlx5_flow_aging_init(sh);
1405         mlx5_flow_counters_mng_init(sh);
1406         mlx5_flow_ipool_create(sh, config);
1407         /* Add device to memory callback list. */
1408         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1409         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1410                          sh, mem_event_cb);
1411         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1412         /* Add context to the global device list. */
1413         LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1414         rte_spinlock_init(&sh->geneve_tlv_opt_sl);
1415 exit:
1416         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1417         return sh;
1418 error:
1419         pthread_mutex_destroy(&sh->txpp.mutex);
1420         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1421         MLX5_ASSERT(sh);
1422         if (sh->share_cache.cache.table)
1423                 mlx5_mr_btree_free(&sh->share_cache.cache);
1424         if (sh->tis)
1425                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1426         if (sh->td)
1427                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1428         if (sh->devx_rx_uar)
1429                 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1430         if (sh->tx_uar)
1431                 mlx5_glue->devx_free_uar(sh->tx_uar);
1432         if (sh->pd)
1433                 claim_zero(mlx5_os_dealloc_pd(sh->pd));
1434         if (sh->ctx)
1435                 claim_zero(mlx5_glue->close_device(sh->ctx));
1436         mlx5_free(sh);
1437         MLX5_ASSERT(err > 0);
1438         rte_errno = err;
1439         return NULL;
1440 }
1441
1442 /**
1443  * Free shared IB device context. Decrement counter and if zero free
1444  * all allocated resources and close handles.
1445  *
1446  * @param[in] sh
1447  *   Pointer to mlx5_dev_ctx_shared object to free
1448  */
1449 void
1450 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1451 {
1452         int ret;
1453
1454         pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1455 #ifdef RTE_LIBRTE_MLX5_DEBUG
1456         /* Check the object presence in the list. */
1457         struct mlx5_dev_ctx_shared *lctx;
1458
1459         LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1460                 if (lctx == sh)
1461                         break;
1462         MLX5_ASSERT(lctx);
1463         if (lctx != sh) {
1464                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
1465                 goto exit;
1466         }
1467 #endif
1468         MLX5_ASSERT(sh);
1469         MLX5_ASSERT(sh->refcnt);
1470         /* Secondary process should not free the shared context. */
1471         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1472         if (--sh->refcnt)
1473                 goto exit;
1474         /* Stop watching for mempool events and unregister all mempools. */
1475         ret = rte_mempool_event_callback_unregister
1476                                 (mlx5_dev_ctx_shared_mempool_event_cb, sh);
1477         if (ret < 0 && rte_errno == ENOENT)
1478                 ret = rte_mempool_event_callback_unregister
1479                                 (mlx5_dev_ctx_shared_rx_mempool_event_cb, sh);
1480         if (ret == 0)
1481                 rte_mempool_walk(mlx5_dev_ctx_shared_mempool_unregister_cb,
1482                                  sh);
1483         /* Remove from memory callback device list. */
1484         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1485         LIST_REMOVE(sh, mem_event_cb);
1486         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1487         /* Release created Memory Regions. */
1488         mlx5_mr_release_cache(&sh->share_cache);
1489         /* Remove context from the global device list. */
1490         LIST_REMOVE(sh, next);
1491         /* Release flow workspaces objects on the last device. */
1492         if (LIST_EMPTY(&mlx5_dev_ctx_list))
1493                 mlx5_flow_os_release_workspace();
1494         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1495         /*
1496          *  Ensure there is no async event handler installed.
1497          *  Only primary process handles async device events.
1498          **/
1499         mlx5_flow_counters_mng_close(sh);
1500         if (sh->aso_age_mng) {
1501                 mlx5_flow_aso_age_mng_close(sh);
1502                 sh->aso_age_mng = NULL;
1503         }
1504         if (sh->mtrmng)
1505                 mlx5_aso_flow_mtrs_mng_close(sh);
1506         mlx5_flow_ipool_destroy(sh);
1507         mlx5_os_dev_shared_handler_uninstall(sh);
1508         if (sh->tx_uar) {
1509                 mlx5_glue->devx_free_uar(sh->tx_uar);
1510                 sh->tx_uar = NULL;
1511         }
1512         if (sh->pd)
1513                 claim_zero(mlx5_os_dealloc_pd(sh->pd));
1514         if (sh->tis)
1515                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1516         if (sh->td)
1517                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
1518         if (sh->devx_rx_uar)
1519                 mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1520         if (sh->ctx)
1521                 claim_zero(mlx5_glue->close_device(sh->ctx));
1522         MLX5_ASSERT(sh->geneve_tlv_option_resource == NULL);
1523         pthread_mutex_destroy(&sh->txpp.mutex);
1524         mlx5_free(sh);
1525         return;
1526 exit:
1527         pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1528 }
1529
1530 /**
1531  * Destroy table hash list.
1532  *
1533  * @param[in] priv
1534  *   Pointer to the private device data structure.
1535  */
1536 void
1537 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1538 {
1539         struct mlx5_dev_ctx_shared *sh = priv->sh;
1540
1541         if (!sh->flow_tbls)
1542                 return;
1543         mlx5_hlist_destroy(sh->flow_tbls);
1544         sh->flow_tbls = NULL;
1545 }
1546
1547 /**
1548  * Initialize flow table hash list and create the root tables entry
1549  * for each domain.
1550  *
1551  * @param[in] priv
1552  *   Pointer to the private device data structure.
1553  *
1554  * @return
1555  *   Zero on success, positive error code otherwise.
1556  */
1557 int
1558 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1559 {
1560         int err = 0;
1561         /* Tables are only used in DV and DR modes. */
1562 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1563         struct mlx5_dev_ctx_shared *sh = priv->sh;
1564         char s[MLX5_NAME_SIZE];
1565
1566         MLX5_ASSERT(sh);
1567         snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1568         sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1569                                           false, true, sh,
1570                                           flow_dv_tbl_create_cb,
1571                                           flow_dv_tbl_match_cb,
1572                                           flow_dv_tbl_remove_cb,
1573                                           flow_dv_tbl_clone_cb,
1574                                           flow_dv_tbl_clone_free_cb);
1575         if (!sh->flow_tbls) {
1576                 DRV_LOG(ERR, "flow tables with hash creation failed.");
1577                 err = ENOMEM;
1578                 return err;
1579         }
1580 #ifndef HAVE_MLX5DV_DR
1581         struct rte_flow_error error;
1582         struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1583
1584         /*
1585          * In case we have not DR support, the zero tables should be created
1586          * because DV expect to see them even if they cannot be created by
1587          * RDMA-CORE.
1588          */
1589         if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0,
1590                 NULL, 0, 1, 0, &error) ||
1591             !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0,
1592                 NULL, 0, 1, 0, &error) ||
1593             !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0,
1594                 NULL, 0, 1, 0, &error)) {
1595                 err = ENOMEM;
1596                 goto error;
1597         }
1598         return err;
1599 error:
1600         mlx5_free_table_hash_list(priv);
1601 #endif /* HAVE_MLX5DV_DR */
1602 #endif
1603         return err;
1604 }
1605
1606 /**
1607  * Retrieve integer value from environment variable.
1608  *
1609  * @param[in] name
1610  *   Environment variable name.
1611  *
1612  * @return
1613  *   Integer value, 0 if the variable is not set.
1614  */
1615 int
1616 mlx5_getenv_int(const char *name)
1617 {
1618         const char *val = getenv(name);
1619
1620         if (val == NULL)
1621                 return 0;
1622         return atoi(val);
1623 }
1624
1625 /**
1626  * DPDK callback to add udp tunnel port
1627  *
1628  * @param[in] dev
1629  *   A pointer to eth_dev
1630  * @param[in] udp_tunnel
1631  *   A pointer to udp tunnel
1632  *
1633  * @return
1634  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1635  */
1636 int
1637 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1638                          struct rte_eth_udp_tunnel *udp_tunnel)
1639 {
1640         MLX5_ASSERT(udp_tunnel != NULL);
1641         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1642             udp_tunnel->udp_port == 4789)
1643                 return 0;
1644         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1645             udp_tunnel->udp_port == 4790)
1646                 return 0;
1647         return -ENOTSUP;
1648 }
1649
1650 /**
1651  * Initialize process private data structure.
1652  *
1653  * @param dev
1654  *   Pointer to Ethernet device structure.
1655  *
1656  * @return
1657  *   0 on success, a negative errno value otherwise and rte_errno is set.
1658  */
1659 int
1660 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1661 {
1662         struct mlx5_priv *priv = dev->data->dev_private;
1663         struct mlx5_proc_priv *ppriv;
1664         size_t ppriv_size;
1665
1666         mlx5_proc_priv_uninit(dev);
1667         /*
1668          * UAR register table follows the process private structure. BlueFlame
1669          * registers for Tx queues are stored in the table.
1670          */
1671         ppriv_size =
1672                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1673         ppriv = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, ppriv_size,
1674                             RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1675         if (!ppriv) {
1676                 rte_errno = ENOMEM;
1677                 return -rte_errno;
1678         }
1679         ppriv->uar_table_sz = priv->txqs_n;
1680         dev->process_private = ppriv;
1681         return 0;
1682 }
1683
1684 /**
1685  * Un-initialize process private data structure.
1686  *
1687  * @param dev
1688  *   Pointer to Ethernet device structure.
1689  */
1690 void
1691 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1692 {
1693         if (!dev->process_private)
1694                 return;
1695         mlx5_free(dev->process_private);
1696         dev->process_private = NULL;
1697 }
1698
1699 /**
1700  * DPDK callback to close the device.
1701  *
1702  * Destroy all queues and objects, free memory.
1703  *
1704  * @param dev
1705  *   Pointer to Ethernet device structure.
1706  */
1707 int
1708 mlx5_dev_close(struct rte_eth_dev *dev)
1709 {
1710         struct mlx5_priv *priv = dev->data->dev_private;
1711         unsigned int i;
1712         int ret;
1713
1714         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1715                 /* Check if process_private released. */
1716                 if (!dev->process_private)
1717                         return 0;
1718                 mlx5_tx_uar_uninit_secondary(dev);
1719                 mlx5_proc_priv_uninit(dev);
1720                 rte_eth_dev_release_port(dev);
1721                 return 0;
1722         }
1723         if (!priv->sh)
1724                 return 0;
1725         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1726                 dev->data->port_id,
1727                 ((priv->sh->ctx != NULL) ?
1728                 mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1729         /*
1730          * If default mreg copy action is removed at the stop stage,
1731          * the search will return none and nothing will be done anymore.
1732          */
1733         mlx5_flow_stop_default(dev);
1734         mlx5_traffic_disable(dev);
1735         /*
1736          * If all the flows are already flushed in the device stop stage,
1737          * then this will return directly without any action.
1738          */
1739         mlx5_flow_list_flush(dev, MLX5_FLOW_TYPE_GEN, true);
1740         mlx5_action_handle_flush(dev);
1741         mlx5_flow_meter_flush(dev, NULL);
1742         /* Prevent crashes when queues are still in use. */
1743         dev->rx_pkt_burst = removed_rx_burst;
1744         dev->tx_pkt_burst = removed_tx_burst;
1745         rte_wmb();
1746         /* Disable datapath on secondary process. */
1747         mlx5_mp_os_req_stop_rxtx(dev);
1748         /* Free the eCPRI flex parser resource. */
1749         mlx5_flex_parser_ecpri_release(dev);
1750         if (priv->rxqs != NULL) {
1751                 /* XXX race condition if mlx5_rx_burst() is still running. */
1752                 rte_delay_us_sleep(1000);
1753                 for (i = 0; (i != priv->rxqs_n); ++i)
1754                         mlx5_rxq_release(dev, i);
1755                 priv->rxqs_n = 0;
1756                 priv->rxqs = NULL;
1757         }
1758         if (priv->representor) {
1759                 /* Each representor has a dedicated interrupts handler */
1760                 mlx5_free(dev->intr_handle);
1761                 dev->intr_handle = NULL;
1762         }
1763         if (priv->txqs != NULL) {
1764                 /* XXX race condition if mlx5_tx_burst() is still running. */
1765                 rte_delay_us_sleep(1000);
1766                 for (i = 0; (i != priv->txqs_n); ++i)
1767                         mlx5_txq_release(dev, i);
1768                 priv->txqs_n = 0;
1769                 priv->txqs = NULL;
1770         }
1771         mlx5_proc_priv_uninit(dev);
1772         if (priv->q_counters) {
1773                 mlx5_devx_cmd_destroy(priv->q_counters);
1774                 priv->q_counters = NULL;
1775         }
1776         if (priv->drop_queue.hrxq)
1777                 mlx5_drop_action_destroy(dev);
1778         if (priv->mreg_cp_tbl)
1779                 mlx5_hlist_destroy(priv->mreg_cp_tbl);
1780         mlx5_mprq_free_mp(dev);
1781         if (priv->sh->ct_mng)
1782                 mlx5_flow_aso_ct_mng_close(priv->sh);
1783         mlx5_os_free_shared_dr(priv);
1784         if (priv->rss_conf.rss_key != NULL)
1785                 mlx5_free(priv->rss_conf.rss_key);
1786         if (priv->reta_idx != NULL)
1787                 mlx5_free(priv->reta_idx);
1788         if (priv->config.vf)
1789                 mlx5_os_mac_addr_flush(dev);
1790         if (priv->nl_socket_route >= 0)
1791                 close(priv->nl_socket_route);
1792         if (priv->nl_socket_rdma >= 0)
1793                 close(priv->nl_socket_rdma);
1794         if (priv->vmwa_context)
1795                 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1796         ret = mlx5_hrxq_verify(dev);
1797         if (ret)
1798                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1799                         dev->data->port_id);
1800         ret = mlx5_ind_table_obj_verify(dev);
1801         if (ret)
1802                 DRV_LOG(WARNING, "port %u some indirection table still remain",
1803                         dev->data->port_id);
1804         ret = mlx5_rxq_obj_verify(dev);
1805         if (ret)
1806                 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1807                         dev->data->port_id);
1808         ret = mlx5_rxq_verify(dev);
1809         if (ret)
1810                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1811                         dev->data->port_id);
1812         ret = mlx5_txq_obj_verify(dev);
1813         if (ret)
1814                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1815                         dev->data->port_id);
1816         ret = mlx5_txq_verify(dev);
1817         if (ret)
1818                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1819                         dev->data->port_id);
1820         ret = mlx5_flow_verify(dev);
1821         if (ret)
1822                 DRV_LOG(WARNING, "port %u some flows still remain",
1823                         dev->data->port_id);
1824         if (priv->hrxqs)
1825                 mlx5_list_destroy(priv->hrxqs);
1826         /*
1827          * Free the shared context in last turn, because the cleanup
1828          * routines above may use some shared fields, like
1829          * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1830          * ifindex if Netlink fails.
1831          */
1832         mlx5_free_shared_dev_ctx(priv->sh);
1833         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1834                 unsigned int c = 0;
1835                 uint16_t port_id;
1836
1837                 MLX5_ETH_FOREACH_DEV(port_id, dev->device) {
1838                         struct mlx5_priv *opriv =
1839                                 rte_eth_devices[port_id].data->dev_private;
1840
1841                         if (!opriv ||
1842                             opriv->domain_id != priv->domain_id ||
1843                             &rte_eth_devices[port_id] == dev)
1844                                 continue;
1845                         ++c;
1846                         break;
1847                 }
1848                 if (!c)
1849                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1850         }
1851         memset(priv, 0, sizeof(*priv));
1852         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1853         /*
1854          * Reset mac_addrs to NULL such that it is not freed as part of
1855          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1856          * it is freed when dev_private is freed.
1857          */
1858         dev->data->mac_addrs = NULL;
1859         return 0;
1860 }
1861
1862 const struct eth_dev_ops mlx5_dev_ops = {
1863         .dev_configure = mlx5_dev_configure,
1864         .dev_start = mlx5_dev_start,
1865         .dev_stop = mlx5_dev_stop,
1866         .dev_set_link_down = mlx5_set_link_down,
1867         .dev_set_link_up = mlx5_set_link_up,
1868         .dev_close = mlx5_dev_close,
1869         .promiscuous_enable = mlx5_promiscuous_enable,
1870         .promiscuous_disable = mlx5_promiscuous_disable,
1871         .allmulticast_enable = mlx5_allmulticast_enable,
1872         .allmulticast_disable = mlx5_allmulticast_disable,
1873         .link_update = mlx5_link_update,
1874         .stats_get = mlx5_stats_get,
1875         .stats_reset = mlx5_stats_reset,
1876         .xstats_get = mlx5_xstats_get,
1877         .xstats_reset = mlx5_xstats_reset,
1878         .xstats_get_names = mlx5_xstats_get_names,
1879         .fw_version_get = mlx5_fw_version_get,
1880         .dev_infos_get = mlx5_dev_infos_get,
1881         .representor_info_get = mlx5_representor_info_get,
1882         .read_clock = mlx5_txpp_read_clock,
1883         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1884         .vlan_filter_set = mlx5_vlan_filter_set,
1885         .rx_queue_setup = mlx5_rx_queue_setup,
1886         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1887         .tx_queue_setup = mlx5_tx_queue_setup,
1888         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1889         .rx_queue_release = mlx5_rx_queue_release,
1890         .tx_queue_release = mlx5_tx_queue_release,
1891         .rx_queue_start = mlx5_rx_queue_start,
1892         .rx_queue_stop = mlx5_rx_queue_stop,
1893         .tx_queue_start = mlx5_tx_queue_start,
1894         .tx_queue_stop = mlx5_tx_queue_stop,
1895         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1896         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1897         .mac_addr_remove = mlx5_mac_addr_remove,
1898         .mac_addr_add = mlx5_mac_addr_add,
1899         .mac_addr_set = mlx5_mac_addr_set,
1900         .set_mc_addr_list = mlx5_set_mc_addr_list,
1901         .mtu_set = mlx5_dev_set_mtu,
1902         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1903         .vlan_offload_set = mlx5_vlan_offload_set,
1904         .reta_update = mlx5_dev_rss_reta_update,
1905         .reta_query = mlx5_dev_rss_reta_query,
1906         .rss_hash_update = mlx5_rss_hash_update,
1907         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
1908         .flow_ops_get = mlx5_flow_ops_get,
1909         .rxq_info_get = mlx5_rxq_info_get,
1910         .txq_info_get = mlx5_txq_info_get,
1911         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1912         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1913         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1914         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1915         .is_removed = mlx5_is_removed,
1916         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1917         .get_module_info = mlx5_get_module_info,
1918         .get_module_eeprom = mlx5_get_module_eeprom,
1919         .hairpin_cap_get = mlx5_hairpin_cap_get,
1920         .mtr_ops_get = mlx5_flow_meter_ops_get,
1921         .hairpin_bind = mlx5_hairpin_bind,
1922         .hairpin_unbind = mlx5_hairpin_unbind,
1923         .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1924         .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1925         .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1926         .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1927         .get_monitor_addr = mlx5_get_monitor_addr,
1928 };
1929
1930 /* Available operations from secondary process. */
1931 const struct eth_dev_ops mlx5_dev_sec_ops = {
1932         .stats_get = mlx5_stats_get,
1933         .stats_reset = mlx5_stats_reset,
1934         .xstats_get = mlx5_xstats_get,
1935         .xstats_reset = mlx5_xstats_reset,
1936         .xstats_get_names = mlx5_xstats_get_names,
1937         .fw_version_get = mlx5_fw_version_get,
1938         .dev_infos_get = mlx5_dev_infos_get,
1939         .representor_info_get = mlx5_representor_info_get,
1940         .read_clock = mlx5_txpp_read_clock,
1941         .rx_queue_start = mlx5_rx_queue_start,
1942         .rx_queue_stop = mlx5_rx_queue_stop,
1943         .tx_queue_start = mlx5_tx_queue_start,
1944         .tx_queue_stop = mlx5_tx_queue_stop,
1945         .rxq_info_get = mlx5_rxq_info_get,
1946         .txq_info_get = mlx5_txq_info_get,
1947         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1948         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1949         .get_module_info = mlx5_get_module_info,
1950         .get_module_eeprom = mlx5_get_module_eeprom,
1951 };
1952
1953 /* Available operations in flow isolated mode. */
1954 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1955         .dev_configure = mlx5_dev_configure,
1956         .dev_start = mlx5_dev_start,
1957         .dev_stop = mlx5_dev_stop,
1958         .dev_set_link_down = mlx5_set_link_down,
1959         .dev_set_link_up = mlx5_set_link_up,
1960         .dev_close = mlx5_dev_close,
1961         .promiscuous_enable = mlx5_promiscuous_enable,
1962         .promiscuous_disable = mlx5_promiscuous_disable,
1963         .allmulticast_enable = mlx5_allmulticast_enable,
1964         .allmulticast_disable = mlx5_allmulticast_disable,
1965         .link_update = mlx5_link_update,
1966         .stats_get = mlx5_stats_get,
1967         .stats_reset = mlx5_stats_reset,
1968         .xstats_get = mlx5_xstats_get,
1969         .xstats_reset = mlx5_xstats_reset,
1970         .xstats_get_names = mlx5_xstats_get_names,
1971         .fw_version_get = mlx5_fw_version_get,
1972         .dev_infos_get = mlx5_dev_infos_get,
1973         .representor_info_get = mlx5_representor_info_get,
1974         .read_clock = mlx5_txpp_read_clock,
1975         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1976         .vlan_filter_set = mlx5_vlan_filter_set,
1977         .rx_queue_setup = mlx5_rx_queue_setup,
1978         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1979         .tx_queue_setup = mlx5_tx_queue_setup,
1980         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1981         .rx_queue_release = mlx5_rx_queue_release,
1982         .tx_queue_release = mlx5_tx_queue_release,
1983         .rx_queue_start = mlx5_rx_queue_start,
1984         .rx_queue_stop = mlx5_rx_queue_stop,
1985         .tx_queue_start = mlx5_tx_queue_start,
1986         .tx_queue_stop = mlx5_tx_queue_stop,
1987         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1988         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1989         .mac_addr_remove = mlx5_mac_addr_remove,
1990         .mac_addr_add = mlx5_mac_addr_add,
1991         .mac_addr_set = mlx5_mac_addr_set,
1992         .set_mc_addr_list = mlx5_set_mc_addr_list,
1993         .mtu_set = mlx5_dev_set_mtu,
1994         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1995         .vlan_offload_set = mlx5_vlan_offload_set,
1996         .flow_ops_get = mlx5_flow_ops_get,
1997         .rxq_info_get = mlx5_rxq_info_get,
1998         .txq_info_get = mlx5_txq_info_get,
1999         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2000         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2001         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2002         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2003         .is_removed = mlx5_is_removed,
2004         .get_module_info = mlx5_get_module_info,
2005         .get_module_eeprom = mlx5_get_module_eeprom,
2006         .hairpin_cap_get = mlx5_hairpin_cap_get,
2007         .mtr_ops_get = mlx5_flow_meter_ops_get,
2008         .hairpin_bind = mlx5_hairpin_bind,
2009         .hairpin_unbind = mlx5_hairpin_unbind,
2010         .hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
2011         .hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2012         .hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2013         .hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
2014         .get_monitor_addr = mlx5_get_monitor_addr,
2015 };
2016
2017 /**
2018  * Verify and store value for device argument.
2019  *
2020  * @param[in] key
2021  *   Key argument to verify.
2022  * @param[in] val
2023  *   Value associated with key.
2024  * @param opaque
2025  *   User data.
2026  *
2027  * @return
2028  *   0 on success, a negative errno value otherwise and rte_errno is set.
2029  */
2030 static int
2031 mlx5_args_check(const char *key, const char *val, void *opaque)
2032 {
2033         struct mlx5_dev_config *config = opaque;
2034         unsigned long mod;
2035         signed long tmp;
2036
2037         /* No-op, port representors are processed in mlx5_dev_spawn(). */
2038         if (!strcmp(MLX5_DRIVER_KEY, key) || !strcmp(MLX5_REPRESENTOR, key))
2039                 return 0;
2040         errno = 0;
2041         tmp = strtol(val, NULL, 0);
2042         if (errno) {
2043                 rte_errno = errno;
2044                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
2045                 return -rte_errno;
2046         }
2047         if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
2048                 /* Negative values are acceptable for some keys only. */
2049                 rte_errno = EINVAL;
2050                 DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
2051                 return -rte_errno;
2052         }
2053         mod = tmp >= 0 ? tmp : -tmp;
2054         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
2055                 if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
2056                         DRV_LOG(ERR, "invalid CQE compression "
2057                                      "format parameter");
2058                         rte_errno = EINVAL;
2059                         return -rte_errno;
2060                 }
2061                 config->cqe_comp = !!tmp;
2062                 config->cqe_comp_fmt = tmp;
2063         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
2064                 config->hw_padding = !!tmp;
2065         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
2066                 config->mprq.enabled = !!tmp;
2067         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
2068                 config->mprq.stride_num_n = tmp;
2069         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
2070                 config->mprq.stride_size_n = tmp;
2071         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
2072                 config->mprq.max_memcpy_len = tmp;
2073         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
2074                 config->mprq.min_rxqs_num = tmp;
2075         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
2076                 DRV_LOG(WARNING, "%s: deprecated parameter,"
2077                                  " converted to txq_inline_max", key);
2078                 config->txq_inline_max = tmp;
2079         } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
2080                 config->txq_inline_max = tmp;
2081         } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
2082                 config->txq_inline_min = tmp;
2083         } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
2084                 config->txq_inline_mpw = tmp;
2085         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
2086                 config->txqs_inline = tmp;
2087         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
2088                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2089         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
2090                 config->mps = !!tmp;
2091         } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
2092                 if (tmp != MLX5_TXDB_CACHED &&
2093                     tmp != MLX5_TXDB_NCACHED &&
2094                     tmp != MLX5_TXDB_HEURISTIC) {
2095                         DRV_LOG(ERR, "invalid Tx doorbell "
2096                                      "mapping parameter");
2097                         rte_errno = EINVAL;
2098                         return -rte_errno;
2099                 }
2100                 config->dbnc = tmp;
2101         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
2102                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2103         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
2104                 DRV_LOG(WARNING, "%s: deprecated parameter,"
2105                                  " converted to txq_inline_mpw", key);
2106                 config->txq_inline_mpw = tmp;
2107         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
2108                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
2109         } else if (strcmp(MLX5_TX_PP, key) == 0) {
2110                 if (!mod) {
2111                         DRV_LOG(ERR, "Zero Tx packet pacing parameter");
2112                         rte_errno = EINVAL;
2113                         return -rte_errno;
2114                 }
2115                 config->tx_pp = tmp;
2116         } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
2117                 config->tx_skew = tmp;
2118         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
2119                 config->rx_vec_en = !!tmp;
2120         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
2121                 config->l3_vxlan_en = !!tmp;
2122         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
2123                 config->vf_nl_en = !!tmp;
2124         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
2125                 config->dv_esw_en = !!tmp;
2126         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
2127                 config->dv_flow_en = !!tmp;
2128         } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
2129                 if (tmp != MLX5_XMETA_MODE_LEGACY &&
2130                     tmp != MLX5_XMETA_MODE_META16 &&
2131                     tmp != MLX5_XMETA_MODE_META32 &&
2132                     tmp != MLX5_XMETA_MODE_MISS_INFO) {
2133                         DRV_LOG(ERR, "invalid extensive "
2134                                      "metadata parameter");
2135                         rte_errno = EINVAL;
2136                         return -rte_errno;
2137                 }
2138                 if (tmp != MLX5_XMETA_MODE_MISS_INFO)
2139                         config->dv_xmeta_en = tmp;
2140                 else
2141                         config->dv_miss_info = 1;
2142         } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
2143                 config->lacp_by_user = !!tmp;
2144         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
2145                 config->mr_ext_memseg_en = !!tmp;
2146         } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
2147                 config->max_dump_files_num = tmp;
2148         } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
2149                 config->lro.timeout = tmp;
2150         } else if (strcmp(RTE_DEVARGS_KEY_CLASS, key) == 0) {
2151                 DRV_LOG(DEBUG, "class argument is %s.", val);
2152         } else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
2153                 config->log_hp_size = tmp;
2154         } else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
2155                 if (tmp != MLX5_RCM_NONE &&
2156                     tmp != MLX5_RCM_LIGHT &&
2157                     tmp != MLX5_RCM_AGGR) {
2158                         DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
2159                         rte_errno = EINVAL;
2160                         return -rte_errno;
2161                 }
2162                 config->reclaim_mode = tmp;
2163         } else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
2164                 config->sys_mem_en = !!tmp;
2165         } else if (strcmp(MLX5_DECAP_EN, key) == 0) {
2166                 config->decap_en = !!tmp;
2167         } else if (strcmp(MLX5_ALLOW_DUPLICATE_PATTERN, key) == 0) {
2168                 config->allow_duplicate_pattern = !!tmp;
2169         } else if (strcmp(MLX5_MR_MEMPOOL_REG_EN, key) == 0) {
2170                 config->mr_mempool_reg_en = !!tmp;
2171         } else {
2172                 DRV_LOG(WARNING, "%s: unknown parameter", key);
2173                 rte_errno = EINVAL;
2174                 return -rte_errno;
2175         }
2176         return 0;
2177 }
2178
2179 /**
2180  * Parse device parameters.
2181  *
2182  * @param config
2183  *   Pointer to device configuration structure.
2184  * @param devargs
2185  *   Device arguments structure.
2186  *
2187  * @return
2188  *   0 on success, a negative errno value otherwise and rte_errno is set.
2189  */
2190 int
2191 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
2192 {
2193         const char **params = (const char *[]){
2194                 MLX5_DRIVER_KEY,
2195                 MLX5_RXQ_CQE_COMP_EN,
2196                 MLX5_RXQ_PKT_PAD_EN,
2197                 MLX5_RX_MPRQ_EN,
2198                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
2199                 MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
2200                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
2201                 MLX5_RXQS_MIN_MPRQ,
2202                 MLX5_TXQ_INLINE,
2203                 MLX5_TXQ_INLINE_MIN,
2204                 MLX5_TXQ_INLINE_MAX,
2205                 MLX5_TXQ_INLINE_MPW,
2206                 MLX5_TXQS_MIN_INLINE,
2207                 MLX5_TXQS_MAX_VEC,
2208                 MLX5_TXQ_MPW_EN,
2209                 MLX5_TXQ_MPW_HDR_DSEG_EN,
2210                 MLX5_TXQ_MAX_INLINE_LEN,
2211                 MLX5_TX_DB_NC,
2212                 MLX5_TX_PP,
2213                 MLX5_TX_SKEW,
2214                 MLX5_TX_VEC_EN,
2215                 MLX5_RX_VEC_EN,
2216                 MLX5_L3_VXLAN_EN,
2217                 MLX5_VF_NL_EN,
2218                 MLX5_DV_ESW_EN,
2219                 MLX5_DV_FLOW_EN,
2220                 MLX5_DV_XMETA_EN,
2221                 MLX5_LACP_BY_USER,
2222                 MLX5_MR_EXT_MEMSEG_EN,
2223                 MLX5_REPRESENTOR,
2224                 MLX5_MAX_DUMP_FILES_NUM,
2225                 MLX5_LRO_TIMEOUT_USEC,
2226                 RTE_DEVARGS_KEY_CLASS,
2227                 MLX5_HP_BUF_SIZE,
2228                 MLX5_RECLAIM_MEM,
2229                 MLX5_SYS_MEM_EN,
2230                 MLX5_DECAP_EN,
2231                 MLX5_ALLOW_DUPLICATE_PATTERN,
2232                 MLX5_MR_MEMPOOL_REG_EN,
2233                 NULL,
2234         };
2235         struct rte_kvargs *kvlist;
2236         int ret = 0;
2237         int i;
2238
2239         if (devargs == NULL)
2240                 return 0;
2241         /* Following UGLY cast is done to pass checkpatch. */
2242         kvlist = rte_kvargs_parse(devargs->args, params);
2243         if (kvlist == NULL) {
2244                 rte_errno = EINVAL;
2245                 return -rte_errno;
2246         }
2247         /* Process parameters. */
2248         for (i = 0; (params[i] != NULL); ++i) {
2249                 if (rte_kvargs_count(kvlist, params[i])) {
2250                         ret = rte_kvargs_process(kvlist, params[i],
2251                                                  mlx5_args_check, config);
2252                         if (ret) {
2253                                 rte_errno = EINVAL;
2254                                 rte_kvargs_free(kvlist);
2255                                 return -rte_errno;
2256                         }
2257                 }
2258         }
2259         rte_kvargs_free(kvlist);
2260         return 0;
2261 }
2262
2263 /**
2264  * Configures the minimal amount of data to inline into WQE
2265  * while sending packets.
2266  *
2267  * - the txq_inline_min has the maximal priority, if this
2268  *   key is specified in devargs
2269  * - if DevX is enabled the inline mode is queried from the
2270  *   device (HCA attributes and NIC vport context if needed).
2271  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
2272  *   and none (0 bytes) for other NICs
2273  *
2274  * @param spawn
2275  *   Verbs device parameters (name, port, switch_info) to spawn.
2276  * @param config
2277  *   Device configuration parameters.
2278  */
2279 void
2280 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
2281                     struct mlx5_dev_config *config)
2282 {
2283         if (config->txq_inline_min != MLX5_ARG_UNSET) {
2284                 /* Application defines size of inlined data explicitly. */
2285                 if (spawn->pci_dev != NULL) {
2286                         switch (spawn->pci_dev->id.device_id) {
2287                         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2288                         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2289                                 if (config->txq_inline_min <
2290                                                (int)MLX5_INLINE_HSIZE_L2) {
2291                                         DRV_LOG(DEBUG,
2292                                                 "txq_inline_mix aligned to minimal ConnectX-4 required value %d",
2293                                                 (int)MLX5_INLINE_HSIZE_L2);
2294                                         config->txq_inline_min =
2295                                                         MLX5_INLINE_HSIZE_L2;
2296                                 }
2297                                 break;
2298                         }
2299                 }
2300                 goto exit;
2301         }
2302         if (config->hca_attr.eth_net_offloads) {
2303                 /* We have DevX enabled, inline mode queried successfully. */
2304                 switch (config->hca_attr.wqe_inline_mode) {
2305                 case MLX5_CAP_INLINE_MODE_L2:
2306                         /* outer L2 header must be inlined. */
2307                         config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2308                         goto exit;
2309                 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
2310                         /* No inline data are required by NIC. */
2311                         config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2312                         config->hw_vlan_insert =
2313                                 config->hca_attr.wqe_vlan_insert;
2314                         DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
2315                         goto exit;
2316                 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
2317                         /* inline mode is defined by NIC vport context. */
2318                         if (!config->hca_attr.eth_virt)
2319                                 break;
2320                         switch (config->hca_attr.vport_inline_mode) {
2321                         case MLX5_INLINE_MODE_NONE:
2322                                 config->txq_inline_min =
2323                                         MLX5_INLINE_HSIZE_NONE;
2324                                 goto exit;
2325                         case MLX5_INLINE_MODE_L2:
2326                                 config->txq_inline_min =
2327                                         MLX5_INLINE_HSIZE_L2;
2328                                 goto exit;
2329                         case MLX5_INLINE_MODE_IP:
2330                                 config->txq_inline_min =
2331                                         MLX5_INLINE_HSIZE_L3;
2332                                 goto exit;
2333                         case MLX5_INLINE_MODE_TCP_UDP:
2334                                 config->txq_inline_min =
2335                                         MLX5_INLINE_HSIZE_L4;
2336                                 goto exit;
2337                         case MLX5_INLINE_MODE_INNER_L2:
2338                                 config->txq_inline_min =
2339                                         MLX5_INLINE_HSIZE_INNER_L2;
2340                                 goto exit;
2341                         case MLX5_INLINE_MODE_INNER_IP:
2342                                 config->txq_inline_min =
2343                                         MLX5_INLINE_HSIZE_INNER_L3;
2344                                 goto exit;
2345                         case MLX5_INLINE_MODE_INNER_TCP_UDP:
2346                                 config->txq_inline_min =
2347                                         MLX5_INLINE_HSIZE_INNER_L4;
2348                                 goto exit;
2349                         }
2350                 }
2351         }
2352         if (spawn->pci_dev == NULL) {
2353                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2354                 goto exit;
2355         }
2356         /*
2357          * We get here if we are unable to deduce
2358          * inline data size with DevX. Try PCI ID
2359          * to determine old NICs.
2360          */
2361         switch (spawn->pci_dev->id.device_id) {
2362         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
2363         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2364         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
2365         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2366                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
2367                 config->hw_vlan_insert = 0;
2368                 break;
2369         case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
2370         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2371         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
2372         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2373                 /*
2374                  * These NICs support VLAN insertion from WQE and
2375                  * report the wqe_vlan_insert flag. But there is the bug
2376                  * and PFC control may be broken, so disable feature.
2377                  */
2378                 config->hw_vlan_insert = 0;
2379                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2380                 break;
2381         default:
2382                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
2383                 break;
2384         }
2385 exit:
2386         DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
2387 }
2388
2389 /**
2390  * Configures the metadata mask fields in the shared context.
2391  *
2392  * @param [in] dev
2393  *   Pointer to Ethernet device.
2394  */
2395 void
2396 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
2397 {
2398         struct mlx5_priv *priv = dev->data->dev_private;
2399         struct mlx5_dev_ctx_shared *sh = priv->sh;
2400         uint32_t meta, mark, reg_c0;
2401
2402         reg_c0 = ~priv->vport_meta_mask;
2403         switch (priv->config.dv_xmeta_en) {
2404         case MLX5_XMETA_MODE_LEGACY:
2405                 meta = UINT32_MAX;
2406                 mark = MLX5_FLOW_MARK_MASK;
2407                 break;
2408         case MLX5_XMETA_MODE_META16:
2409                 meta = reg_c0 >> rte_bsf32(reg_c0);
2410                 mark = MLX5_FLOW_MARK_MASK;
2411                 break;
2412         case MLX5_XMETA_MODE_META32:
2413                 meta = UINT32_MAX;
2414                 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
2415                 break;
2416         default:
2417                 meta = 0;
2418                 mark = 0;
2419                 MLX5_ASSERT(false);
2420                 break;
2421         }
2422         if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2423                 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2424                                  sh->dv_mark_mask, mark);
2425         else
2426                 sh->dv_mark_mask = mark;
2427         if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2428                 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2429                                  sh->dv_meta_mask, meta);
2430         else
2431                 sh->dv_meta_mask = meta;
2432         if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2433                 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2434                                  sh->dv_meta_mask, reg_c0);
2435         else
2436                 sh->dv_regc0_mask = reg_c0;
2437         DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2438         DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2439         DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2440         DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2441 }
2442
2443 int
2444 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2445 {
2446         static const char *const dynf_names[] = {
2447                 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2448                 RTE_MBUF_DYNFLAG_METADATA_NAME,
2449                 RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
2450         };
2451         unsigned int i;
2452
2453         if (n < RTE_DIM(dynf_names))
2454                 return -ENOMEM;
2455         for (i = 0; i < RTE_DIM(dynf_names); i++) {
2456                 if (names[i] == NULL)
2457                         return -EINVAL;
2458                 strcpy(names[i], dynf_names[i]);
2459         }
2460         return RTE_DIM(dynf_names);
2461 }
2462
2463 /**
2464  * Comparison callback to sort device data.
2465  *
2466  * This is meant to be used with qsort().
2467  *
2468  * @param a[in]
2469  *   Pointer to pointer to first data object.
2470  * @param b[in]
2471  *   Pointer to pointer to second data object.
2472  *
2473  * @return
2474  *   0 if both objects are equal, less than 0 if the first argument is less
2475  *   than the second, greater than 0 otherwise.
2476  */
2477 int
2478 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2479                               struct mlx5_dev_config *config,
2480                               struct rte_device *dpdk_dev)
2481 {
2482         struct mlx5_dev_ctx_shared *sh = priv->sh;
2483         struct mlx5_dev_config *sh_conf = NULL;
2484         uint16_t port_id;
2485
2486         MLX5_ASSERT(sh);
2487         /* Nothing to compare for the single/first device. */
2488         if (sh->refcnt == 1)
2489                 return 0;
2490         /* Find the device with shared context. */
2491         MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
2492                 struct mlx5_priv *opriv =
2493                         rte_eth_devices[port_id].data->dev_private;
2494
2495                 if (opriv && opriv != priv && opriv->sh == sh) {
2496                         sh_conf = &opriv->config;
2497                         break;
2498                 }
2499         }
2500         if (!sh_conf)
2501                 return 0;
2502         if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2503                 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2504                              " for shared %s context", sh->ibdev_name);
2505                 rte_errno = EINVAL;
2506                 return rte_errno;
2507         }
2508         if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2509                 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2510                              " for shared %s context", sh->ibdev_name);
2511                 rte_errno = EINVAL;
2512                 return rte_errno;
2513         }
2514         return 0;
2515 }
2516
2517 /**
2518  * Look for the ethernet device belonging to mlx5 driver.
2519  *
2520  * @param[in] port_id
2521  *   port_id to start looking for device.
2522  * @param[in] odev
2523  *   Pointer to the hint device. When device is being probed
2524  *   the its siblings (master and preceding representors might
2525  *   not have assigned driver yet (because the mlx5_os_pci_probe()
2526  *   is not completed yet, for this case match on hint
2527  *   device may be used to detect sibling device.
2528  *
2529  * @return
2530  *   port_id of found device, RTE_MAX_ETHPORT if not found.
2531  */
2532 uint16_t
2533 mlx5_eth_find_next(uint16_t port_id, struct rte_device *odev)
2534 {
2535         while (port_id < RTE_MAX_ETHPORTS) {
2536                 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2537
2538                 if (dev->state != RTE_ETH_DEV_UNUSED &&
2539                     dev->device &&
2540                     (dev->device == odev ||
2541                      (dev->device->driver &&
2542                      dev->device->driver->name &&
2543                      ((strcmp(dev->device->driver->name,
2544                               MLX5_PCI_DRIVER_NAME) == 0) ||
2545                       (strcmp(dev->device->driver->name,
2546                               MLX5_AUXILIARY_DRIVER_NAME) == 0)))))
2547                         break;
2548                 port_id++;
2549         }
2550         if (port_id >= RTE_MAX_ETHPORTS)
2551                 return RTE_MAX_ETHPORTS;
2552         return port_id;
2553 }
2554
2555 /**
2556  * Callback to remove a device.
2557  *
2558  * This function removes all Ethernet devices belong to a given device.
2559  *
2560  * @param[in] cdev
2561  *   Pointer to the generic device.
2562  *
2563  * @return
2564  *   0 on success, the function cannot fail.
2565  */
2566 int
2567 mlx5_net_remove(struct mlx5_common_device *cdev)
2568 {
2569         uint16_t port_id;
2570         int ret = 0;
2571
2572         RTE_ETH_FOREACH_DEV_OF(port_id, cdev->dev) {
2573                 /*
2574                  * mlx5_dev_close() is not registered to secondary process,
2575                  * call the close function explicitly for secondary process.
2576                  */
2577                 if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2578                         ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2579                 else
2580                         ret |= rte_eth_dev_close(port_id);
2581         }
2582         return ret == 0 ? 0 : -EIO;
2583 }
2584
2585 static const struct rte_pci_id mlx5_pci_id_map[] = {
2586         {
2587                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2588                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2589         },
2590         {
2591                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2592                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2593         },
2594         {
2595                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2596                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2597         },
2598         {
2599                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2600                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2601         },
2602         {
2603                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2604                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2605         },
2606         {
2607                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2608                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2609         },
2610         {
2611                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2612                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2613         },
2614         {
2615                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2616                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2617         },
2618         {
2619                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2620                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2621         },
2622         {
2623                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2624                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2625         },
2626         {
2627                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2628                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2629         },
2630         {
2631                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2632                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2633         },
2634         {
2635                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2636                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2637         },
2638         {
2639                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2640                                 PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2641         },
2642         {
2643                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2644                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2645         },
2646         {
2647                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2648                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2649         },
2650         {
2651                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2652                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2653         },
2654         {
2655                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2656                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2657         },
2658         {
2659                 .vendor_id = 0
2660         }
2661 };
2662
2663 static struct mlx5_class_driver mlx5_net_driver = {
2664         .drv_class = MLX5_CLASS_ETH,
2665         .name = RTE_STR(MLX5_ETH_DRIVER_NAME),
2666         .id_table = mlx5_pci_id_map,
2667         .probe = mlx5_os_net_probe,
2668         .remove = mlx5_net_remove,
2669         .dma_map = mlx5_net_dma_map,
2670         .dma_unmap = mlx5_net_dma_unmap,
2671         .probe_again = 1,
2672         .intr_lsc = 1,
2673         .intr_rmv = 1,
2674 };
2675
2676 /* Initialize driver log type. */
2677 RTE_LOG_REGISTER_DEFAULT(mlx5_logtype, NOTICE)
2678
2679 /**
2680  * Driver initialization routine.
2681  */
2682 RTE_INIT(rte_mlx5_pmd_init)
2683 {
2684         pthread_mutex_init(&mlx5_dev_ctx_list_mutex, NULL);
2685         mlx5_common_init();
2686         /* Build the static tables for Verbs conversion. */
2687         mlx5_set_ptype_table();
2688         mlx5_set_cksum_table();
2689         mlx5_set_swp_types_table();
2690         if (mlx5_glue)
2691                 mlx5_class_driver_register(&mlx5_net_driver);
2692 }
2693
2694 RTE_PMD_EXPORT_NAME(MLX5_ETH_DRIVER_NAME, __COUNTER__);
2695 RTE_PMD_REGISTER_PCI_TABLE(MLX5_ETH_DRIVER_NAME, mlx5_pci_id_map);
2696 RTE_PMD_REGISTER_KMOD_DEP(MLX5_ETH_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib");