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 <assert.h>
  10 #include <stdint.h>
  11 #include <stdlib.h>
  12 #include <errno.h>
  13 #include <net/if.h>
  14 #include <sys/mman.h>
  15 #include <linux/rtnetlink.h>
  16
  17 /* Verbs header. */
  18 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  19 #ifdef PEDANTIC
  20 #pragma GCC diagnostic ignored "-Wpedantic"
  21 #endif
  22 #include <infiniband/verbs.h>
  23 #ifdef PEDANTIC
  24 #pragma GCC diagnostic error "-Wpedantic"
  25 #endif
  26
  27 #include <rte_malloc.h>
  28 #include <rte_ethdev_driver.h>
  29 #include <rte_ethdev_pci.h>
  30 #include <rte_pci.h>
  31 #include <rte_bus_pci.h>
  32 #include <rte_common.h>
  33 #include <rte_config.h>
  34 #include <rte_kvargs.h>
  35 #include <rte_rwlock.h>
  36 #include <rte_spinlock.h>
  37 #include <rte_string_fns.h>
  38 #include <rte_alarm.h>
  39
  40 #include <mlx5_glue.h>
  41 #include <mlx5_devx_cmds.h>
  42 #include <mlx5_common.h>
  43
  44 #include "mlx5_defs.h"
  45 #include "mlx5.h"
  46 #include "mlx5_utils.h"
  47 #include "mlx5_rxtx.h"
  48 #include "mlx5_autoconf.h"
  49 #include "mlx5_mr.h"
  50 #include "mlx5_flow.h"
  51
  52 /* Device parameter to enable RX completion queue compression. */
  53 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
  54
  55 /* Device parameter to enable RX completion entry padding to 128B. */
  56 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
  57
  58 /* Device parameter to enable padding Rx packet to cacheline size. */
  59 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
  60
  61 /* Device parameter to enable Multi-Packet Rx queue. */
  62 #define MLX5_RX_MPRQ_EN "mprq_en"
  63
  64 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
  65 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
  66
  67 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
  68 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
  69
  70 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
  71 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
  72
  73 /* Device parameter to configure inline send. Deprecated, ignored.*/
  74 #define MLX5_TXQ_INLINE "txq_inline"
  75
  76 /* Device parameter to limit packet size to inline with ordinary SEND. */
  77 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
  78
  79 /* Device parameter to configure minimal data size to inline. */
  80 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
  81
  82 /* Device parameter to limit packet size to inline with Enhanced MPW. */
  83 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
  84
  85 /*
  86  * Device parameter to configure the number of TX queues threshold for
  87  * enabling inline send.
  88  */
  89 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
  90
  91 /*
  92  * Device parameter to configure the number of TX queues threshold for
  93  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
  94  */
  95 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
  96
  97 /* Device parameter to enable multi-packet send WQEs. */
  98 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
  99
 100 /*
 101  * Device parameter to force doorbell register mapping
 102  * to non-cahed region eliminating the extra write memory barrier.
 103  */
 104 #define MLX5_TX_DB_NC "tx_db_nc"
 105
 106 /*
 107  * Device parameter to include 2 dsegs in the title WQEBB.
 108  * Deprecated, ignored.
 109  */
 110 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
 111
 112 /*
 113  * Device parameter to limit the size of inlining packet.
 114  * Deprecated, ignored.
 115  */
 116 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
 117
 118 /*
 119  * Device parameter to enable hardware Tx vector.
 120  * Deprecated, ignored (no vectorized Tx routines anymore).
 121  */
 122 #define MLX5_TX_VEC_EN "tx_vec_en"
 123
 124 /* Device parameter to enable hardware Rx vector. */
 125 #define MLX5_RX_VEC_EN "rx_vec_en"
 126
 127 /* Allow L3 VXLAN flow creation. */
 128 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
 129
 130 /* Activate DV E-Switch flow steering. */
 131 #define MLX5_DV_ESW_EN "dv_esw_en"
 132
 133 /* Activate DV flow steering. */
 134 #define MLX5_DV_FLOW_EN "dv_flow_en"
 135
 136 /* Enable extensive flow metadata support. */
 137 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
 138
 139 /* Activate Netlink support in VF mode. */
 140 #define MLX5_VF_NL_EN "vf_nl_en"
 141
 142 /* Enable extending memsegs when creating a MR. */
 143 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
 144
 145 /* Select port representors to instantiate. */
 146 #define MLX5_REPRESENTOR "representor"
 147
 148 /* Device parameter to configure the maximum number of dump files per queue. */
 149 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
 150
 151 /* Configure timeout of LRO session (in microseconds). */
 152 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
 153
 154 #ifndef HAVE_IBV_MLX5_MOD_MPW
 155 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
 156 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
 157 #endif
 158
 159 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
 160 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
 161 #endif
 162
 163 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
 164
 165 /* Shared memory between primary and secondary processes. */
 166 struct mlx5_shared_data *mlx5_shared_data;
 167
 168 /* Spinlock for mlx5_shared_data allocation. */
 169 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
 170
 171 /* Process local data for secondary processes. */
 172 static struct mlx5_local_data mlx5_local_data;
 173
 174 /** Driver-specific log messages type. */
 175 int mlx5_logtype;
 176
 177 /** Data associated with devices to spawn. */
 178 struct mlx5_dev_spawn_data {
 179         uint32_t ifindex; /**< Network interface index. */
 180         uint32_t max_port; /**< IB device maximal port index. */
 181         uint32_t ibv_port; /**< IB device physical port index. */
 182         int pf_bond; /**< bonding device PF index. < 0 - no bonding */
 183         struct mlx5_switch_info info; /**< Switch information. */
 184         struct ibv_device *ibv_dev; /**< Associated IB device. */
 185         struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
 186         struct rte_pci_device *pci_dev; /**< Backend PCI device. */
 187 };
 188
 189 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
 190 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
 191
 192 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
 193 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
 194
 195 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
 196 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
 197
 198 /**
 199  * Allocate ID pool structure.
 200  *
 201  * @param[in] max_id
 202  *   The maximum id can be allocated from the pool.
 203  *
 204  * @return
 205  *   Pointer to pool object, NULL value otherwise.
 206  */
 207 struct mlx5_flow_id_pool *
 208 mlx5_flow_id_pool_alloc(uint32_t max_id)
 209 {
 210         struct mlx5_flow_id_pool *pool;
 211         void *mem;
 212
 213         pool = rte_zmalloc("id pool allocation", sizeof(*pool),
 214                            RTE_CACHE_LINE_SIZE);
 215         if (!pool) {
 216                 DRV_LOG(ERR, "can't allocate id pool");
 217                 rte_errno  = ENOMEM;
 218                 return NULL;
 219         }
 220         mem = rte_zmalloc("", MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
 221                           RTE_CACHE_LINE_SIZE);
 222         if (!mem) {
 223                 DRV_LOG(ERR, "can't allocate mem for id pool");
 224                 rte_errno  = ENOMEM;
 225                 goto error;
 226         }
 227         pool->free_arr = mem;
 228         pool->curr = pool->free_arr;
 229         pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
 230         pool->base_index = 0;
 231         pool->max_id = max_id;
 232         return pool;
 233 error:
 234         rte_free(pool);
 235         return NULL;
 236 }
 237
 238 /**
 239  * Release ID pool structure.
 240  *
 241  * @param[in] pool
 242  *   Pointer to flow id pool object to free.
 243  */
 244 void
 245 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
 246 {
 247         rte_free(pool->free_arr);
 248         rte_free(pool);
 249 }
 250
 251 /**
 252  * Generate ID.
 253  *
 254  * @param[in] pool
 255  *   Pointer to flow id pool.
 256  * @param[out] id
 257  *   The generated ID.
 258  *
 259  * @return
 260  *   0 on success, error value otherwise.
 261  */
 262 uint32_t
 263 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
 264 {
 265         if (pool->curr == pool->free_arr) {
 266                 if (pool->base_index == pool->max_id) {
 267                         rte_errno  = ENOMEM;
 268                         DRV_LOG(ERR, "no free id");
 269                         return -rte_errno;
 270                 }
 271                 *id = ++pool->base_index;
 272                 return 0;
 273         }
 274         *id = *(--pool->curr);
 275         return 0;
 276 }
 277
 278 /**
 279  * Release ID.
 280  *
 281  * @param[in] pool
 282  *   Pointer to flow id pool.
 283  * @param[out] id
 284  *   The generated ID.
 285  *
 286  * @return
 287  *   0 on success, error value otherwise.
 288  */
 289 uint32_t
 290 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
 291 {
 292         uint32_t size;
 293         uint32_t size2;
 294         void *mem;
 295
 296         if (pool->curr == pool->last) {
 297                 size = pool->curr - pool->free_arr;
 298                 size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
 299                 assert(size2 > size);
 300                 mem = rte_malloc("", size2 * sizeof(uint32_t), 0);
 301                 if (!mem) {
 302                         DRV_LOG(ERR, "can't allocate mem for id pool");
 303                         rte_errno  = ENOMEM;
 304                         return -rte_errno;
 305                 }
 306                 memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
 307                 rte_free(pool->free_arr);
 308                 pool->free_arr = mem;
 309                 pool->curr = pool->free_arr + size;
 310                 pool->last = pool->free_arr + size2;
 311         }
 312         *pool->curr = id;
 313         pool->curr++;
 314         return 0;
 315 }
 316
 317 /**
 318  * Initialize the counters management structure.
 319  *
 320  * @param[in] sh
 321  *   Pointer to mlx5_ibv_shared object to free
 322  */
 323 static void
 324 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
 325 {
 326         uint8_t i;
 327
 328         TAILQ_INIT(&sh->cmng.flow_counters);
 329         for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
 330                 TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
 331 }
 332
 333 /**
 334  * Destroy all the resources allocated for a counter memory management.
 335  *
 336  * @param[in] mng
 337  *   Pointer to the memory management structure.
 338  */
 339 static void
 340 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
 341 {
 342         uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
 343
 344         LIST_REMOVE(mng, next);
 345         claim_zero(mlx5_devx_cmd_destroy(mng->dm));
 346         claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
 347         rte_free(mem);
 348 }
 349
 350 /**
 351  * Close and release all the resources of the counters management.
 352  *
 353  * @param[in] sh
 354  *   Pointer to mlx5_ibv_shared object to free.
 355  */
 356 static void
 357 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
 358 {
 359         struct mlx5_counter_stats_mem_mng *mng;
 360         uint8_t i;
 361         int j;
 362         int retries = 1024;
 363
 364         rte_errno = 0;
 365         while (--retries) {
 366                 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
 367                 if (rte_errno != EINPROGRESS)
 368                         break;
 369                 rte_pause();
 370         }
 371         for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
 372                 struct mlx5_flow_counter_pool *pool;
 373                 uint32_t batch = !!(i % 2);
 374
 375                 if (!sh->cmng.ccont[i].pools)
 376                         continue;
 377                 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
 378                 while (pool) {
 379                         if (batch) {
 380                                 if (pool->min_dcs)
 381                                         claim_zero
 382                                         (mlx5_devx_cmd_destroy(pool->min_dcs));
 383                         }
 384                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
 385                                 if (pool->counters_raw[j].action)
 386                                         claim_zero
 387                                         (mlx5_glue->destroy_flow_action
 388                                                (pool->counters_raw[j].action));
 389                                 if (!batch && pool->counters_raw[j].dcs)
 390                                         claim_zero(mlx5_devx_cmd_destroy
 391                                                   (pool->counters_raw[j].dcs));
 392                         }
 393                         TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
 394                                      next);
 395                         rte_free(pool);
 396                         pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
 397                 }
 398                 rte_free(sh->cmng.ccont[i].pools);
 399         }
 400         mng = LIST_FIRST(&sh->cmng.mem_mngs);
 401         while (mng) {
 402                 mlx5_flow_destroy_counter_stat_mem_mng(mng);
 403                 mng = LIST_FIRST(&sh->cmng.mem_mngs);
 404         }
 405         memset(&sh->cmng, 0, sizeof(sh->cmng));
 406 }
 407
 408 /**
 409  * Extract pdn of PD object using DV API.
 410  *
 411  * @param[in] pd
 412  *   Pointer to the verbs PD object.
 413  * @param[out] pdn
 414  *   Pointer to the PD object number variable.
 415  *
 416  * @return
 417  *   0 on success, error value otherwise.
 418  */
 419 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 420 static int
 421 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
 422 {
 423         struct mlx5dv_obj obj;
 424         struct mlx5dv_pd pd_info;
 425         int ret = 0;
 426
 427         obj.pd.in = pd;
 428         obj.pd.out = &pd_info;
 429         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
 430         if (ret) {
 431                 DRV_LOG(DEBUG, "Fail to get PD object info");
 432                 return ret;
 433         }
 434         *pdn = pd_info.pdn;
 435         return 0;
 436 }
 437 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 438
 439 static int
 440 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
 441 {
 442         char *env;
 443         int value;
 444
 445         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 446         /* Get environment variable to store. */
 447         env = getenv(MLX5_SHUT_UP_BF);
 448         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
 449         if (config->dbnc == MLX5_ARG_UNSET)
 450                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
 451         else
 452                 setenv(MLX5_SHUT_UP_BF,
 453                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
 454         return value;
 455 }
 456
 457 static void
 458 mlx5_restore_doorbell_mapping_env(int value)
 459 {
 460         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 461         /* Restore the original environment variable state. */
 462         if (value == MLX5_ARG_UNSET)
 463                 unsetenv(MLX5_SHUT_UP_BF);
 464         else
 465                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
 466 }
 467
 468 /**
 469  * Allocate shared IB device context. If there is multiport device the
 470  * master and representors will share this context, if there is single
 471  * port dedicated IB device, the context will be used by only given
 472  * port due to unification.
 473  *
 474  * Routine first searches the context for the specified IB device name,
 475  * if found the shared context assumed and reference counter is incremented.
 476  * If no context found the new one is created and initialized with specified
 477  * IB device context and parameters.
 478  *
 479  * @param[in] spawn
 480  *   Pointer to the IB device attributes (name, port, etc).
 481  * @param[in] config
 482  *   Pointer to device configuration structure.
 483  *
 484  * @return
 485  *   Pointer to mlx5_ibv_shared object on success,
 486  *   otherwise NULL and rte_errno is set.
 487  */
 488 static struct mlx5_ibv_shared *
 489 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
 490                         const struct mlx5_dev_config *config)
 491 {
 492         struct mlx5_ibv_shared *sh;
 493         int dbmap_env;
 494         int err = 0;
 495         uint32_t i;
 496 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 497         struct mlx5_devx_tis_attr tis_attr = { 0 };
 498 #endif
 499
 500         assert(spawn);
 501         /* Secondary process should not create the shared context. */
 502         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 503         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 504         /* Search for IB context by device name. */
 505         LIST_FOREACH(sh, &mlx5_ibv_list, next) {
 506                 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
 507                         sh->refcnt++;
 508                         goto exit;
 509                 }
 510         }
 511         /* No device found, we have to create new shared context. */
 512         assert(spawn->max_port);
 513         sh = rte_zmalloc("ethdev shared ib context",
 514                          sizeof(struct mlx5_ibv_shared) +
 515                          spawn->max_port *
 516                          sizeof(struct mlx5_ibv_shared_port),
 517                          RTE_CACHE_LINE_SIZE);
 518         if (!sh) {
 519                 DRV_LOG(ERR, "shared context allocation failure");
 520                 rte_errno  = ENOMEM;
 521                 goto exit;
 522         }
 523         /*
 524          * Configure environment variable "MLX5_BF_SHUT_UP"
 525          * before the device creation. The rdma_core library
 526          * checks the variable at device creation and
 527          * stores the result internally.
 528          */
 529         dbmap_env = mlx5_config_doorbell_mapping_env(config);
 530         /* Try to open IB device with DV first, then usual Verbs. */
 531         errno = 0;
 532         sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
 533         if (sh->ctx) {
 534                 sh->devx = 1;
 535                 DRV_LOG(DEBUG, "DevX is supported");
 536                 /* The device is created, no need for environment. */
 537                 mlx5_restore_doorbell_mapping_env(dbmap_env);
 538         } else {
 539                 /* The environment variable is still configured. */
 540                 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
 541                 err = errno ? errno : ENODEV;
 542                 /*
 543                  * The environment variable is not needed anymore,
 544                  * all device creation attempts are completed.
 545                  */
 546                 mlx5_restore_doorbell_mapping_env(dbmap_env);
 547                 if (!sh->ctx)
 548                         goto error;
 549                 DRV_LOG(DEBUG, "DevX is NOT supported");
 550         }
 551         err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
 552         if (err) {
 553                 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
 554                 goto error;
 555         }
 556         sh->refcnt = 1;
 557         sh->max_port = spawn->max_port;
 558         strncpy(sh->ibdev_name, sh->ctx->device->name,
 559                 sizeof(sh->ibdev_name));
 560         strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
 561                 sizeof(sh->ibdev_path));
 562         pthread_mutex_init(&sh->intr_mutex, NULL);
 563         /*
 564          * Setting port_id to max unallowed value means
 565          * there is no interrupt subhandler installed for
 566          * the given port index i.
 567          */
 568         for (i = 0; i < sh->max_port; i++) {
 569                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
 570                 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
 571         }
 572         sh->pd = mlx5_glue->alloc_pd(sh->ctx);
 573         if (sh->pd == NULL) {
 574                 DRV_LOG(ERR, "PD allocation failure");
 575                 err = ENOMEM;
 576                 goto error;
 577         }
 578 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 579         if (sh->devx) {
 580                 err = mlx5_get_pdn(sh->pd, &sh->pdn);
 581                 if (err) {
 582                         DRV_LOG(ERR, "Fail to extract pdn from PD");
 583                         goto error;
 584                 }
 585                 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
 586                 if (!sh->td) {
 587                         DRV_LOG(ERR, "TD allocation failure");
 588                         err = ENOMEM;
 589                         goto error;
 590                 }
 591                 tis_attr.transport_domain = sh->td->id;
 592                 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
 593                 if (!sh->tis) {
 594                         DRV_LOG(ERR, "TIS allocation failure");
 595                         err = ENOMEM;
 596                         goto error;
 597                 }
 598         }
 599         sh->flow_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX);
 600         if (!sh->flow_id_pool) {
 601                 DRV_LOG(ERR, "can't create flow id pool");
 602                 err = ENOMEM;
 603                 goto error;
 604         }
 605 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 606         /*
 607          * Once the device is added to the list of memory event
 608          * callback, its global MR cache table cannot be expanded
 609          * on the fly because of deadlock. If it overflows, lookup
 610          * should be done by searching MR list linearly, which is slow.
 611          *
 612          * At this point the device is not added to the memory
 613          * event list yet, context is just being created.
 614          */
 615         err = mlx5_mr_btree_init(&sh->mr.cache,
 616                                  MLX5_MR_BTREE_CACHE_N * 2,
 617                                  spawn->pci_dev->device.numa_node);
 618         if (err) {
 619                 err = rte_errno;
 620                 goto error;
 621         }
 622         mlx5_flow_counters_mng_init(sh);
 623         /* Add device to memory callback list. */
 624         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
 625         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
 626                          sh, mem_event_cb);
 627         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
 628         /* Add context to the global device list. */
 629         LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
 630 exit:
 631         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 632         return sh;
 633 error:
 634         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 635         assert(sh);
 636         if (sh->tis)
 637                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
 638         if (sh->td)
 639                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
 640         if (sh->pd)
 641                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
 642         if (sh->ctx)
 643                 claim_zero(mlx5_glue->close_device(sh->ctx));
 644         if (sh->flow_id_pool)
 645                 mlx5_flow_id_pool_release(sh->flow_id_pool);
 646         rte_free(sh);
 647         assert(err > 0);
 648         rte_errno = err;
 649         return NULL;
 650 }
 651
 652 /**
 653  * Free shared IB device context. Decrement counter and if zero free
 654  * all allocated resources and close handles.
 655  *
 656  * @param[in] sh
 657  *   Pointer to mlx5_ibv_shared object to free
 658  */
 659 static void
 660 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
 661 {
 662         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 663 #ifndef NDEBUG
 664         /* Check the object presence in the list. */
 665         struct mlx5_ibv_shared *lctx;
 666
 667         LIST_FOREACH(lctx, &mlx5_ibv_list, next)
 668                 if (lctx == sh)
 669                         break;
 670         assert(lctx);
 671         if (lctx != sh) {
 672                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
 673                 goto exit;
 674         }
 675 #endif
 676         assert(sh);
 677         assert(sh->refcnt);
 678         /* Secondary process should not free the shared context. */
 679         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 680         if (--sh->refcnt)
 681                 goto exit;
 682         /* Release created Memory Regions. */
 683         mlx5_mr_release(sh);
 684         /* Remove from memory callback device list. */
 685         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
 686         LIST_REMOVE(sh, mem_event_cb);
 687         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
 688         /* Remove context from the global device list. */
 689         LIST_REMOVE(sh, next);
 690         /*
 691          *  Ensure there is no async event handler installed.
 692          *  Only primary process handles async device events.
 693          **/
 694         mlx5_flow_counters_mng_close(sh);
 695         assert(!sh->intr_cnt);
 696         if (sh->intr_cnt)
 697                 mlx5_intr_callback_unregister
 698                         (&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
 699 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
 700         if (sh->devx_intr_cnt) {
 701                 if (sh->intr_handle_devx.fd)
 702                         rte_intr_callback_unregister(&sh->intr_handle_devx,
 703                                           mlx5_dev_interrupt_handler_devx, sh);
 704                 if (sh->devx_comp)
 705                         mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
 706         }
 707 #endif
 708         pthread_mutex_destroy(&sh->intr_mutex);
 709         if (sh->pd)
 710                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
 711         if (sh->tis)
 712                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
 713         if (sh->td)
 714                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
 715         if (sh->ctx)
 716                 claim_zero(mlx5_glue->close_device(sh->ctx));
 717         if (sh->flow_id_pool)
 718                 mlx5_flow_id_pool_release(sh->flow_id_pool);
 719         rte_free(sh);
 720 exit:
 721         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 722 }
 723
 724 /**
 725  * Destroy table hash list and all the root entries per domain.
 726  *
 727  * @param[in] priv
 728  *   Pointer to the private device data structure.
 729  */
 730 static void
 731 mlx5_free_table_hash_list(struct mlx5_priv *priv)
 732 {
 733         struct mlx5_ibv_shared *sh = priv->sh;
 734         struct mlx5_flow_tbl_data_entry *tbl_data;
 735         union mlx5_flow_tbl_key table_key = {
 736                 {
 737                         .table_id = 0,
 738                         .reserved = 0,
 739                         .domain = 0,
 740                         .direction = 0,
 741                 }
 742         };
 743         struct mlx5_hlist_entry *pos;
 744
 745         if (!sh->flow_tbls)
 746                 return;
 747         pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
 748         if (pos) {
 749                 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
 750                                         entry);
 751                 assert(tbl_data);
 752                 mlx5_hlist_remove(sh->flow_tbls, pos);
 753                 rte_free(tbl_data);
 754         }
 755         table_key.direction = 1;
 756         pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
 757         if (pos) {
 758                 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
 759                                         entry);
 760                 assert(tbl_data);
 761                 mlx5_hlist_remove(sh->flow_tbls, pos);
 762                 rte_free(tbl_data);
 763         }
 764         table_key.direction = 0;
 765         table_key.domain = 1;
 766         pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
 767         if (pos) {
 768                 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
 769                                         entry);
 770                 assert(tbl_data);
 771                 mlx5_hlist_remove(sh->flow_tbls, pos);
 772                 rte_free(tbl_data);
 773         }
 774         mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
 775 }
 776
 777 /**
 778  * Initialize flow table hash list and create the root tables entry
 779  * for each domain.
 780  *
 781  * @param[in] priv
 782  *   Pointer to the private device data structure.
 783  *
 784  * @return
 785  *   Zero on success, positive error code otherwise.
 786  */
 787 static int
 788 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
 789 {
 790         struct mlx5_ibv_shared *sh = priv->sh;
 791         char s[MLX5_HLIST_NAMESIZE];
 792         int err = 0;
 793
 794         assert(sh);
 795         snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
 796         sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
 797         if (!sh->flow_tbls) {
 798                 DRV_LOG(ERR, "flow tables with hash creation failed.\n");
 799                 err = ENOMEM;
 800                 return err;
 801         }
 802 #ifndef HAVE_MLX5DV_DR
 803         /*
 804          * In case we have not DR support, the zero tables should be created
 805          * because DV expect to see them even if they cannot be created by
 806          * RDMA-CORE.
 807          */
 808         union mlx5_flow_tbl_key table_key = {
 809                 {
 810                         .table_id = 0,
 811                         .reserved = 0,
 812                         .domain = 0,
 813                         .direction = 0,
 814                 }
 815         };
 816         struct mlx5_flow_tbl_data_entry *tbl_data = rte_zmalloc(NULL,
 817                                                           sizeof(*tbl_data), 0);
 818
 819         if (!tbl_data) {
 820                 err = ENOMEM;
 821                 goto error;
 822         }
 823         tbl_data->entry.key = table_key.v64;
 824         err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
 825         if (err)
 826                 goto error;
 827         rte_atomic32_init(&tbl_data->tbl.refcnt);
 828         rte_atomic32_inc(&tbl_data->tbl.refcnt);
 829         table_key.direction = 1;
 830         tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
 831         if (!tbl_data) {
 832                 err = ENOMEM;
 833                 goto error;
 834         }
 835         tbl_data->entry.key = table_key.v64;
 836         err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
 837         if (err)
 838                 goto error;
 839         rte_atomic32_init(&tbl_data->tbl.refcnt);
 840         rte_atomic32_inc(&tbl_data->tbl.refcnt);
 841         table_key.direction = 0;
 842         table_key.domain = 1;
 843         tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
 844         if (!tbl_data) {
 845                 err = ENOMEM;
 846                 goto error;
 847         }
 848         tbl_data->entry.key = table_key.v64;
 849         err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
 850         if (err)
 851                 goto error;
 852         rte_atomic32_init(&tbl_data->tbl.refcnt);
 853         rte_atomic32_inc(&tbl_data->tbl.refcnt);
 854         return err;
 855 error:
 856         mlx5_free_table_hash_list(priv);
 857 #endif /* HAVE_MLX5DV_DR */
 858         return err;
 859 }
 860
 861 /**
 862  * Initialize DR related data within private structure.
 863  * Routine checks the reference counter and does actual
 864  * resources creation/initialization only if counter is zero.
 865  *
 866  * @param[in] priv
 867  *   Pointer to the private device data structure.
 868  *
 869  * @return
 870  *   Zero on success, positive error code otherwise.
 871  */
 872 static int
 873 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 874 {
 875         struct mlx5_ibv_shared *sh = priv->sh;
 876         char s[MLX5_HLIST_NAMESIZE];
 877         int err = 0;
 878
 879         if (!sh->flow_tbls)
 880                 err = mlx5_alloc_table_hash_list(priv);
 881         else
 882                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
 883                         (void *)sh->flow_tbls);
 884         if (err)
 885                 return err;
 886         /* Create tags hash list table. */
 887         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 888         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
 889         if (!sh->tag_table) {
 890                 DRV_LOG(ERR, "tags with hash creation failed.\n");
 891                 err = ENOMEM;
 892                 goto error;
 893         }
 894 #ifdef HAVE_MLX5DV_DR
 895         void *domain;
 896
 897         if (sh->dv_refcnt) {
 898                 /* Shared DV/DR structures is already initialized. */
 899                 sh->dv_refcnt++;
 900                 priv->dr_shared = 1;
 901                 return 0;
 902         }
 903         /* Reference counter is zero, we should initialize structures. */
 904         domain = mlx5_glue->dr_create_domain(sh->ctx,
 905                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 906         if (!domain) {
 907                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 908                 err = errno;
 909                 goto error;
 910         }
 911         sh->rx_domain = domain;
 912         domain = mlx5_glue->dr_create_domain(sh->ctx,
 913                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 914         if (!domain) {
 915                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 916                 err = errno;
 917                 goto error;
 918         }
 919         pthread_mutex_init(&sh->dv_mutex, NULL);
 920         sh->tx_domain = domain;
 921 #ifdef HAVE_MLX5DV_DR_ESWITCH
 922         if (priv->config.dv_esw_en) {
 923                 domain  = mlx5_glue->dr_create_domain
 924                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 925                 if (!domain) {
 926                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 927                         err = errno;
 928                         goto error;
 929                 }
 930                 sh->fdb_domain = domain;
 931                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 932         }
 933 #endif
 934         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 935 #endif /* HAVE_MLX5DV_DR */
 936         sh->dv_refcnt++;
 937         priv->dr_shared = 1;
 938         return 0;
 939 error:
 940         /* Rollback the created objects. */
 941         if (sh->rx_domain) {
 942                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 943                 sh->rx_domain = NULL;
 944         }
 945         if (sh->tx_domain) {
 946                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 947                 sh->tx_domain = NULL;
 948         }
 949         if (sh->fdb_domain) {
 950                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 951                 sh->fdb_domain = NULL;
 952         }
 953         if (sh->esw_drop_action) {
 954                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 955                 sh->esw_drop_action = NULL;
 956         }
 957         if (sh->pop_vlan_action) {
 958                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 959                 sh->pop_vlan_action = NULL;
 960         }
 961         if (sh->tag_table) {
 962                 /* tags should be destroyed with flow before. */
 963                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 964                 sh->tag_table = NULL;
 965         }
 966         mlx5_free_table_hash_list(priv);
 967         return err;
 968 }
 969
 970 /**
 971  * Destroy DR related data within private structure.
 972  *
 973  * @param[in] priv
 974  *   Pointer to the private device data structure.
 975  */
 976 static void
 977 mlx5_free_shared_dr(struct mlx5_priv *priv)
 978 {
 979         struct mlx5_ibv_shared *sh;
 980
 981         if (!priv->dr_shared)
 982                 return;
 983         priv->dr_shared = 0;
 984         sh = priv->sh;
 985         assert(sh);
 986 #ifdef HAVE_MLX5DV_DR
 987         assert(sh->dv_refcnt);
 988         if (sh->dv_refcnt && --sh->dv_refcnt)
 989                 return;
 990         if (sh->rx_domain) {
 991                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 992                 sh->rx_domain = NULL;
 993         }
 994         if (sh->tx_domain) {
 995                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 996                 sh->tx_domain = NULL;
 997         }
 998 #ifdef HAVE_MLX5DV_DR_ESWITCH
 999         if (sh->fdb_domain) {
1000                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1001                 sh->fdb_domain = NULL;
1002         }
1003         if (sh->esw_drop_action) {
1004                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1005                 sh->esw_drop_action = NULL;
1006         }
1007 #endif
1008         if (sh->pop_vlan_action) {
1009                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1010                 sh->pop_vlan_action = NULL;
1011         }
1012         pthread_mutex_destroy(&sh->dv_mutex);
1013 #endif /* HAVE_MLX5DV_DR */
1014         if (sh->tag_table) {
1015                 /* tags should be destroyed with flow before. */
1016                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1017                 sh->tag_table = NULL;
1018         }
1019         mlx5_free_table_hash_list(priv);
1020 }
1021
1022 /**
1023  * Initialize shared data between primary and secondary process.
1024  *
1025  * A memzone is reserved by primary process and secondary processes attach to
1026  * the memzone.
1027  *
1028  * @return
1029  *   0 on success, a negative errno value otherwise and rte_errno is set.
1030  */
1031 static int
1032 mlx5_init_shared_data(void)
1033 {
1034         const struct rte_memzone *mz;
1035         int ret = 0;
1036
1037         rte_spinlock_lock(&mlx5_shared_data_lock);
1038         if (mlx5_shared_data == NULL) {
1039                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1040                         /* Allocate shared memory. */
1041                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1042                                                  sizeof(*mlx5_shared_data),
1043                                                  SOCKET_ID_ANY, 0);
1044                         if (mz == NULL) {
1045                                 DRV_LOG(ERR,
1046                                         "Cannot allocate mlx5 shared data");
1047                                 ret = -rte_errno;
1048                                 goto error;
1049                         }
1050                         mlx5_shared_data = mz->addr;
1051                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1052                         rte_spinlock_init(&mlx5_shared_data->lock);
1053                 } else {
1054                         /* Lookup allocated shared memory. */
1055                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1056                         if (mz == NULL) {
1057                                 DRV_LOG(ERR,
1058                                         "Cannot attach mlx5 shared data");
1059                                 ret = -rte_errno;
1060                                 goto error;
1061                         }
1062                         mlx5_shared_data = mz->addr;
1063                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1064                 }
1065         }
1066 error:
1067         rte_spinlock_unlock(&mlx5_shared_data_lock);
1068         return ret;
1069 }
1070
1071 /**
1072  * Retrieve integer value from environment variable.
1073  *
1074  * @param[in] name
1075  *   Environment variable name.
1076  *
1077  * @return
1078  *   Integer value, 0 if the variable is not set.
1079  */
1080 int
1081 mlx5_getenv_int(const char *name)
1082 {
1083         const char *val = getenv(name);
1084
1085         if (val == NULL)
1086                 return 0;
1087         return atoi(val);
1088 }
1089
1090 /**
1091  * Verbs callback to allocate a memory. This function should allocate the space
1092  * according to the size provided residing inside a huge page.
1093  * Please note that all allocation must respect the alignment from libmlx5
1094  * (i.e. currently sysconf(_SC_PAGESIZE)).
1095  *
1096  * @param[in] size
1097  *   The size in bytes of the memory to allocate.
1098  * @param[in] data
1099  *   A pointer to the callback data.
1100  *
1101  * @return
1102  *   Allocated buffer, NULL otherwise and rte_errno is set.
1103  */
1104 static void *
1105 mlx5_alloc_verbs_buf(size_t size, void *data)
1106 {
1107         struct mlx5_priv *priv = data;
1108         void *ret;
1109         size_t alignment = sysconf(_SC_PAGESIZE);
1110         unsigned int socket = SOCKET_ID_ANY;
1111
1112         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
1113                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1114
1115                 socket = ctrl->socket;
1116         } else if (priv->verbs_alloc_ctx.type ==
1117                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
1118                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1119
1120                 socket = ctrl->socket;
1121         }
1122         assert(data != NULL);
1123         ret = rte_malloc_socket(__func__, size, alignment, socket);
1124         if (!ret && size)
1125                 rte_errno = ENOMEM;
1126         return ret;
1127 }
1128
1129 /**
1130  * Verbs callback to free a memory.
1131  *
1132  * @param[in] ptr
1133  *   A pointer to the memory to free.
1134  * @param[in] data
1135  *   A pointer to the callback data.
1136  */
1137 static void
1138 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1139 {
1140         assert(data != NULL);
1141         rte_free(ptr);
1142 }
1143
1144 /**
1145  * DPDK callback to add udp tunnel port
1146  *
1147  * @param[in] dev
1148  *   A pointer to eth_dev
1149  * @param[in] udp_tunnel
1150  *   A pointer to udp tunnel
1151  *
1152  * @return
1153  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1154  */
1155 int
1156 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1157                          struct rte_eth_udp_tunnel *udp_tunnel)
1158 {
1159         assert(udp_tunnel != NULL);
1160         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1161             udp_tunnel->udp_port == 4789)
1162                 return 0;
1163         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1164             udp_tunnel->udp_port == 4790)
1165                 return 0;
1166         return -ENOTSUP;
1167 }
1168
1169 /**
1170  * Initialize process private data structure.
1171  *
1172  * @param dev
1173  *   Pointer to Ethernet device structure.
1174  *
1175  * @return
1176  *   0 on success, a negative errno value otherwise and rte_errno is set.
1177  */
1178 int
1179 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1180 {
1181         struct mlx5_priv *priv = dev->data->dev_private;
1182         struct mlx5_proc_priv *ppriv;
1183         size_t ppriv_size;
1184
1185         /*
1186          * UAR register table follows the process private structure. BlueFlame
1187          * registers for Tx queues are stored in the table.
1188          */
1189         ppriv_size =
1190                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1191         ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1192                                   RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1193         if (!ppriv) {
1194                 rte_errno = ENOMEM;
1195                 return -rte_errno;
1196         }
1197         ppriv->uar_table_sz = ppriv_size;
1198         dev->process_private = ppriv;
1199         return 0;
1200 }
1201
1202 /**
1203  * Un-initialize process private data structure.
1204  *
1205  * @param dev
1206  *   Pointer to Ethernet device structure.
1207  */
1208 static void
1209 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1210 {
1211         if (!dev->process_private)
1212                 return;
1213         rte_free(dev->process_private);
1214         dev->process_private = NULL;
1215 }
1216
1217 /**
1218  * DPDK callback to close the device.
1219  *
1220  * Destroy all queues and objects, free memory.
1221  *
1222  * @param dev
1223  *   Pointer to Ethernet device structure.
1224  */
1225 static void
1226 mlx5_dev_close(struct rte_eth_dev *dev)
1227 {
1228         struct mlx5_priv *priv = dev->data->dev_private;
1229         unsigned int i;
1230         int ret;
1231
1232         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1233                 dev->data->port_id,
1234                 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1235         /* In case mlx5_dev_stop() has not been called. */
1236         mlx5_dev_interrupt_handler_uninstall(dev);
1237         mlx5_dev_interrupt_handler_devx_uninstall(dev);
1238         mlx5_traffic_disable(dev);
1239         mlx5_flow_flush(dev, NULL);
1240         mlx5_flow_meter_flush(dev, NULL);
1241         /* Prevent crashes when queues are still in use. */
1242         dev->rx_pkt_burst = removed_rx_burst;
1243         dev->tx_pkt_burst = removed_tx_burst;
1244         rte_wmb();
1245         /* Disable datapath on secondary process. */
1246         mlx5_mp_req_stop_rxtx(dev);
1247         if (priv->rxqs != NULL) {
1248                 /* XXX race condition if mlx5_rx_burst() is still running. */
1249                 usleep(1000);
1250                 for (i = 0; (i != priv->rxqs_n); ++i)
1251                         mlx5_rxq_release(dev, i);
1252                 priv->rxqs_n = 0;
1253                 priv->rxqs = NULL;
1254         }
1255         if (priv->txqs != NULL) {
1256                 /* XXX race condition if mlx5_tx_burst() is still running. */
1257                 usleep(1000);
1258                 for (i = 0; (i != priv->txqs_n); ++i)
1259                         mlx5_txq_release(dev, i);
1260                 priv->txqs_n = 0;
1261                 priv->txqs = NULL;
1262         }
1263         mlx5_proc_priv_uninit(dev);
1264         if (priv->mreg_cp_tbl)
1265                 mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1266         mlx5_mprq_free_mp(dev);
1267         mlx5_free_shared_dr(priv);
1268         if (priv->rss_conf.rss_key != NULL)
1269                 rte_free(priv->rss_conf.rss_key);
1270         if (priv->reta_idx != NULL)
1271                 rte_free(priv->reta_idx);
1272         if (priv->config.vf)
1273                 mlx5_nl_mac_addr_flush(dev);
1274         if (priv->nl_socket_route >= 0)
1275                 close(priv->nl_socket_route);
1276         if (priv->nl_socket_rdma >= 0)
1277                 close(priv->nl_socket_rdma);
1278         if (priv->vmwa_context)
1279                 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1280         if (priv->sh) {
1281                 /*
1282                  * Free the shared context in last turn, because the cleanup
1283                  * routines above may use some shared fields, like
1284                  * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1285                  * ifindex if Netlink fails.
1286                  */
1287                 mlx5_free_shared_ibctx(priv->sh);
1288                 priv->sh = NULL;
1289         }
1290         ret = mlx5_hrxq_verify(dev);
1291         if (ret)
1292                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1293                         dev->data->port_id);
1294         ret = mlx5_ind_table_obj_verify(dev);
1295         if (ret)
1296                 DRV_LOG(WARNING, "port %u some indirection table still remain",
1297                         dev->data->port_id);
1298         ret = mlx5_rxq_obj_verify(dev);
1299         if (ret)
1300                 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1301                         dev->data->port_id);
1302         ret = mlx5_rxq_verify(dev);
1303         if (ret)
1304                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1305                         dev->data->port_id);
1306         ret = mlx5_txq_obj_verify(dev);
1307         if (ret)
1308                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1309                         dev->data->port_id);
1310         ret = mlx5_txq_verify(dev);
1311         if (ret)
1312                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1313                         dev->data->port_id);
1314         ret = mlx5_flow_verify(dev);
1315         if (ret)
1316                 DRV_LOG(WARNING, "port %u some flows still remain",
1317                         dev->data->port_id);
1318         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1319                 unsigned int c = 0;
1320                 uint16_t port_id;
1321
1322                 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1323                         struct mlx5_priv *opriv =
1324                                 rte_eth_devices[port_id].data->dev_private;
1325
1326                         if (!opriv ||
1327                             opriv->domain_id != priv->domain_id ||
1328                             &rte_eth_devices[port_id] == dev)
1329                                 continue;
1330                         ++c;
1331                         break;
1332                 }
1333                 if (!c)
1334                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1335         }
1336         memset(priv, 0, sizeof(*priv));
1337         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1338         /*
1339          * Reset mac_addrs to NULL such that it is not freed as part of
1340          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1341          * it is freed when dev_private is freed.
1342          */
1343         dev->data->mac_addrs = NULL;
1344 }
1345
1346 const struct eth_dev_ops mlx5_dev_ops = {
1347         .dev_configure = mlx5_dev_configure,
1348         .dev_start = mlx5_dev_start,
1349         .dev_stop = mlx5_dev_stop,
1350         .dev_set_link_down = mlx5_set_link_down,
1351         .dev_set_link_up = mlx5_set_link_up,
1352         .dev_close = mlx5_dev_close,
1353         .promiscuous_enable = mlx5_promiscuous_enable,
1354         .promiscuous_disable = mlx5_promiscuous_disable,
1355         .allmulticast_enable = mlx5_allmulticast_enable,
1356         .allmulticast_disable = mlx5_allmulticast_disable,
1357         .link_update = mlx5_link_update,
1358         .stats_get = mlx5_stats_get,
1359         .stats_reset = mlx5_stats_reset,
1360         .xstats_get = mlx5_xstats_get,
1361         .xstats_reset = mlx5_xstats_reset,
1362         .xstats_get_names = mlx5_xstats_get_names,
1363         .fw_version_get = mlx5_fw_version_get,
1364         .dev_infos_get = mlx5_dev_infos_get,
1365         .read_clock = mlx5_read_clock,
1366         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1367         .vlan_filter_set = mlx5_vlan_filter_set,
1368         .rx_queue_setup = mlx5_rx_queue_setup,
1369         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1370         .tx_queue_setup = mlx5_tx_queue_setup,
1371         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1372         .rx_queue_release = mlx5_rx_queue_release,
1373         .tx_queue_release = mlx5_tx_queue_release,
1374         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1375         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1376         .mac_addr_remove = mlx5_mac_addr_remove,
1377         .mac_addr_add = mlx5_mac_addr_add,
1378         .mac_addr_set = mlx5_mac_addr_set,
1379         .set_mc_addr_list = mlx5_set_mc_addr_list,
1380         .mtu_set = mlx5_dev_set_mtu,
1381         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1382         .vlan_offload_set = mlx5_vlan_offload_set,
1383         .reta_update = mlx5_dev_rss_reta_update,
1384         .reta_query = mlx5_dev_rss_reta_query,
1385         .rss_hash_update = mlx5_rss_hash_update,
1386         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
1387         .filter_ctrl = mlx5_dev_filter_ctrl,
1388         .rx_descriptor_status = mlx5_rx_descriptor_status,
1389         .tx_descriptor_status = mlx5_tx_descriptor_status,
1390         .rx_queue_count = mlx5_rx_queue_count,
1391         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1392         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1393         .is_removed = mlx5_is_removed,
1394         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1395         .get_module_info = mlx5_get_module_info,
1396         .get_module_eeprom = mlx5_get_module_eeprom,
1397         .hairpin_cap_get = mlx5_hairpin_cap_get,
1398         .mtr_ops_get = mlx5_flow_meter_ops_get,
1399 };
1400
1401 /* Available operations from secondary process. */
1402 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1403         .stats_get = mlx5_stats_get,
1404         .stats_reset = mlx5_stats_reset,
1405         .xstats_get = mlx5_xstats_get,
1406         .xstats_reset = mlx5_xstats_reset,
1407         .xstats_get_names = mlx5_xstats_get_names,
1408         .fw_version_get = mlx5_fw_version_get,
1409         .dev_infos_get = mlx5_dev_infos_get,
1410         .rx_descriptor_status = mlx5_rx_descriptor_status,
1411         .tx_descriptor_status = mlx5_tx_descriptor_status,
1412         .get_module_info = mlx5_get_module_info,
1413         .get_module_eeprom = mlx5_get_module_eeprom,
1414 };
1415
1416 /* Available operations in flow isolated mode. */
1417 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1418         .dev_configure = mlx5_dev_configure,
1419         .dev_start = mlx5_dev_start,
1420         .dev_stop = mlx5_dev_stop,
1421         .dev_set_link_down = mlx5_set_link_down,
1422         .dev_set_link_up = mlx5_set_link_up,
1423         .dev_close = mlx5_dev_close,
1424         .promiscuous_enable = mlx5_promiscuous_enable,
1425         .promiscuous_disable = mlx5_promiscuous_disable,
1426         .allmulticast_enable = mlx5_allmulticast_enable,
1427         .allmulticast_disable = mlx5_allmulticast_disable,
1428         .link_update = mlx5_link_update,
1429         .stats_get = mlx5_stats_get,
1430         .stats_reset = mlx5_stats_reset,
1431         .xstats_get = mlx5_xstats_get,
1432         .xstats_reset = mlx5_xstats_reset,
1433         .xstats_get_names = mlx5_xstats_get_names,
1434         .fw_version_get = mlx5_fw_version_get,
1435         .dev_infos_get = mlx5_dev_infos_get,
1436         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1437         .vlan_filter_set = mlx5_vlan_filter_set,
1438         .rx_queue_setup = mlx5_rx_queue_setup,
1439         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1440         .tx_queue_setup = mlx5_tx_queue_setup,
1441         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1442         .rx_queue_release = mlx5_rx_queue_release,
1443         .tx_queue_release = mlx5_tx_queue_release,
1444         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1445         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1446         .mac_addr_remove = mlx5_mac_addr_remove,
1447         .mac_addr_add = mlx5_mac_addr_add,
1448         .mac_addr_set = mlx5_mac_addr_set,
1449         .set_mc_addr_list = mlx5_set_mc_addr_list,
1450         .mtu_set = mlx5_dev_set_mtu,
1451         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1452         .vlan_offload_set = mlx5_vlan_offload_set,
1453         .filter_ctrl = mlx5_dev_filter_ctrl,
1454         .rx_descriptor_status = mlx5_rx_descriptor_status,
1455         .tx_descriptor_status = mlx5_tx_descriptor_status,
1456         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1457         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1458         .is_removed = mlx5_is_removed,
1459         .get_module_info = mlx5_get_module_info,
1460         .get_module_eeprom = mlx5_get_module_eeprom,
1461         .hairpin_cap_get = mlx5_hairpin_cap_get,
1462         .mtr_ops_get = mlx5_flow_meter_ops_get,
1463 };
1464
1465 /**
1466  * Verify and store value for device argument.
1467  *
1468  * @param[in] key
1469  *   Key argument to verify.
1470  * @param[in] val
1471  *   Value associated with key.
1472  * @param opaque
1473  *   User data.
1474  *
1475  * @return
1476  *   0 on success, a negative errno value otherwise and rte_errno is set.
1477  */
1478 static int
1479 mlx5_args_check(const char *key, const char *val, void *opaque)
1480 {
1481         struct mlx5_dev_config *config = opaque;
1482         unsigned long tmp;
1483
1484         /* No-op, port representors are processed in mlx5_dev_spawn(). */
1485         if (!strcmp(MLX5_REPRESENTOR, key))
1486                 return 0;
1487         errno = 0;
1488         tmp = strtoul(val, NULL, 0);
1489         if (errno) {
1490                 rte_errno = errno;
1491                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1492                 return -rte_errno;
1493         }
1494         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1495                 config->cqe_comp = !!tmp;
1496         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1497                 config->cqe_pad = !!tmp;
1498         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1499                 config->hw_padding = !!tmp;
1500         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1501                 config->mprq.enabled = !!tmp;
1502         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1503                 config->mprq.stride_num_n = tmp;
1504         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1505                 config->mprq.max_memcpy_len = tmp;
1506         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1507                 config->mprq.min_rxqs_num = tmp;
1508         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1509                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1510                                  " converted to txq_inline_max", key);
1511                 config->txq_inline_max = tmp;
1512         } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1513                 config->txq_inline_max = tmp;
1514         } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1515                 config->txq_inline_min = tmp;
1516         } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1517                 config->txq_inline_mpw = tmp;
1518         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1519                 config->txqs_inline = tmp;
1520         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1521                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1522         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1523                 config->mps = !!tmp;
1524         } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1525                 if (tmp != MLX5_TXDB_CACHED &&
1526                     tmp != MLX5_TXDB_NCACHED &&
1527                     tmp != MLX5_TXDB_HEURISTIC) {
1528                         DRV_LOG(ERR, "invalid Tx doorbell "
1529                                      "mapping parameter");
1530                         rte_errno = EINVAL;
1531                         return -rte_errno;
1532                 }
1533                 config->dbnc = tmp;
1534         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1535                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1536         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1537                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1538                                  " converted to txq_inline_mpw", key);
1539                 config->txq_inline_mpw = tmp;
1540         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1541                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1542         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1543                 config->rx_vec_en = !!tmp;
1544         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1545                 config->l3_vxlan_en = !!tmp;
1546         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1547                 config->vf_nl_en = !!tmp;
1548         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1549                 config->dv_esw_en = !!tmp;
1550         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1551                 config->dv_flow_en = !!tmp;
1552         } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1553                 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1554                     tmp != MLX5_XMETA_MODE_META16 &&
1555                     tmp != MLX5_XMETA_MODE_META32) {
1556                         DRV_LOG(ERR, "invalid extensive "
1557                                      "metadata parameter");
1558                         rte_errno = EINVAL;
1559                         return -rte_errno;
1560                 }
1561                 config->dv_xmeta_en = tmp;
1562         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1563                 config->mr_ext_memseg_en = !!tmp;
1564         } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1565                 config->max_dump_files_num = tmp;
1566         } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1567                 config->lro.timeout = tmp;
1568         } else {
1569                 DRV_LOG(WARNING, "%s: unknown parameter", key);
1570                 rte_errno = EINVAL;
1571                 return -rte_errno;
1572         }
1573         return 0;
1574 }
1575
1576 /**
1577  * Parse device parameters.
1578  *
1579  * @param config
1580  *   Pointer to device configuration structure.
1581  * @param devargs
1582  *   Device arguments structure.
1583  *
1584  * @return
1585  *   0 on success, a negative errno value otherwise and rte_errno is set.
1586  */
1587 static int
1588 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1589 {
1590         const char **params = (const char *[]){
1591                 MLX5_RXQ_CQE_COMP_EN,
1592                 MLX5_RXQ_CQE_PAD_EN,
1593                 MLX5_RXQ_PKT_PAD_EN,
1594                 MLX5_RX_MPRQ_EN,
1595                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1596                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1597                 MLX5_RXQS_MIN_MPRQ,
1598                 MLX5_TXQ_INLINE,
1599                 MLX5_TXQ_INLINE_MIN,
1600                 MLX5_TXQ_INLINE_MAX,
1601                 MLX5_TXQ_INLINE_MPW,
1602                 MLX5_TXQS_MIN_INLINE,
1603                 MLX5_TXQS_MAX_VEC,
1604                 MLX5_TXQ_MPW_EN,
1605                 MLX5_TXQ_MPW_HDR_DSEG_EN,
1606                 MLX5_TXQ_MAX_INLINE_LEN,
1607                 MLX5_TX_DB_NC,
1608                 MLX5_TX_VEC_EN,
1609                 MLX5_RX_VEC_EN,
1610                 MLX5_L3_VXLAN_EN,
1611                 MLX5_VF_NL_EN,
1612                 MLX5_DV_ESW_EN,
1613                 MLX5_DV_FLOW_EN,
1614                 MLX5_DV_XMETA_EN,
1615                 MLX5_MR_EXT_MEMSEG_EN,
1616                 MLX5_REPRESENTOR,
1617                 MLX5_MAX_DUMP_FILES_NUM,
1618                 MLX5_LRO_TIMEOUT_USEC,
1619                 NULL,
1620         };
1621         struct rte_kvargs *kvlist;
1622         int ret = 0;
1623         int i;
1624
1625         if (devargs == NULL)
1626                 return 0;
1627         /* Following UGLY cast is done to pass checkpatch. */
1628         kvlist = rte_kvargs_parse(devargs->args, params);
1629         if (kvlist == NULL) {
1630                 rte_errno = EINVAL;
1631                 return -rte_errno;
1632         }
1633         /* Process parameters. */
1634         for (i = 0; (params[i] != NULL); ++i) {
1635                 if (rte_kvargs_count(kvlist, params[i])) {
1636                         ret = rte_kvargs_process(kvlist, params[i],
1637                                                  mlx5_args_check, config);
1638                         if (ret) {
1639                                 rte_errno = EINVAL;
1640                                 rte_kvargs_free(kvlist);
1641                                 return -rte_errno;
1642                         }
1643                 }
1644         }
1645         rte_kvargs_free(kvlist);
1646         return 0;
1647 }
1648
1649 static struct rte_pci_driver mlx5_driver;
1650
1651 /**
1652  * PMD global initialization.
1653  *
1654  * Independent from individual device, this function initializes global
1655  * per-PMD data structures distinguishing primary and secondary processes.
1656  * Hence, each initialization is called once per a process.
1657  *
1658  * @return
1659  *   0 on success, a negative errno value otherwise and rte_errno is set.
1660  */
1661 static int
1662 mlx5_init_once(void)
1663 {
1664         struct mlx5_shared_data *sd;
1665         struct mlx5_local_data *ld = &mlx5_local_data;
1666         int ret = 0;
1667
1668         if (mlx5_init_shared_data())
1669                 return -rte_errno;
1670         sd = mlx5_shared_data;
1671         assert(sd);
1672         rte_spinlock_lock(&sd->lock);
1673         switch (rte_eal_process_type()) {
1674         case RTE_PROC_PRIMARY:
1675                 if (sd->init_done)
1676                         break;
1677                 LIST_INIT(&sd->mem_event_cb_list);
1678                 rte_rwlock_init(&sd->mem_event_rwlock);
1679                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1680                                                 mlx5_mr_mem_event_cb, NULL);
1681                 ret = mlx5_mp_init_primary();
1682                 if (ret)
1683                         goto out;
1684                 sd->init_done = true;
1685                 break;
1686         case RTE_PROC_SECONDARY:
1687                 if (ld->init_done)
1688                         break;
1689                 ret = mlx5_mp_init_secondary();
1690                 if (ret)
1691                         goto out;
1692                 ++sd->secondary_cnt;
1693                 ld->init_done = true;
1694                 break;
1695         default:
1696                 break;
1697         }
1698 out:
1699         rte_spinlock_unlock(&sd->lock);
1700         return ret;
1701 }
1702
1703 /**
1704  * Configures the minimal amount of data to inline into WQE
1705  * while sending packets.
1706  *
1707  * - the txq_inline_min has the maximal priority, if this
1708  *   key is specified in devargs
1709  * - if DevX is enabled the inline mode is queried from the
1710  *   device (HCA attributes and NIC vport context if needed).
1711  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1712  *   and none (0 bytes) for other NICs
1713  *
1714  * @param spawn
1715  *   Verbs device parameters (name, port, switch_info) to spawn.
1716  * @param config
1717  *   Device configuration parameters.
1718  */
1719 static void
1720 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1721                     struct mlx5_dev_config *config)
1722 {
1723         if (config->txq_inline_min != MLX5_ARG_UNSET) {
1724                 /* Application defines size of inlined data explicitly. */
1725                 switch (spawn->pci_dev->id.device_id) {
1726                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1727                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1728                         if (config->txq_inline_min <
1729                                        (int)MLX5_INLINE_HSIZE_L2) {
1730                                 DRV_LOG(DEBUG,
1731                                         "txq_inline_mix aligned to minimal"
1732                                         " ConnectX-4 required value %d",
1733                                         (int)MLX5_INLINE_HSIZE_L2);
1734                                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1735                         }
1736                         break;
1737                 }
1738                 goto exit;
1739         }
1740         if (config->hca_attr.eth_net_offloads) {
1741                 /* We have DevX enabled, inline mode queried successfully. */
1742                 switch (config->hca_attr.wqe_inline_mode) {
1743                 case MLX5_CAP_INLINE_MODE_L2:
1744                         /* outer L2 header must be inlined. */
1745                         config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1746                         goto exit;
1747                 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1748                         /* No inline data are required by NIC. */
1749                         config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1750                         config->hw_vlan_insert =
1751                                 config->hca_attr.wqe_vlan_insert;
1752                         DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1753                         goto exit;
1754                 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1755                         /* inline mode is defined by NIC vport context. */
1756                         if (!config->hca_attr.eth_virt)
1757                                 break;
1758                         switch (config->hca_attr.vport_inline_mode) {
1759                         case MLX5_INLINE_MODE_NONE:
1760                                 config->txq_inline_min =
1761                                         MLX5_INLINE_HSIZE_NONE;
1762                                 goto exit;
1763                         case MLX5_INLINE_MODE_L2:
1764                                 config->txq_inline_min =
1765                                         MLX5_INLINE_HSIZE_L2;
1766                                 goto exit;
1767                         case MLX5_INLINE_MODE_IP:
1768                                 config->txq_inline_min =
1769                                         MLX5_INLINE_HSIZE_L3;
1770                                 goto exit;
1771                         case MLX5_INLINE_MODE_TCP_UDP:
1772                                 config->txq_inline_min =
1773                                         MLX5_INLINE_HSIZE_L4;
1774                                 goto exit;
1775                         case MLX5_INLINE_MODE_INNER_L2:
1776                                 config->txq_inline_min =
1777                                         MLX5_INLINE_HSIZE_INNER_L2;
1778                                 goto exit;
1779                         case MLX5_INLINE_MODE_INNER_IP:
1780                                 config->txq_inline_min =
1781                                         MLX5_INLINE_HSIZE_INNER_L3;
1782                                 goto exit;
1783                         case MLX5_INLINE_MODE_INNER_TCP_UDP:
1784                                 config->txq_inline_min =
1785                                         MLX5_INLINE_HSIZE_INNER_L4;
1786                                 goto exit;
1787                         }
1788                 }
1789         }
1790         /*
1791          * We get here if we are unable to deduce
1792          * inline data size with DevX. Try PCI ID
1793          * to determine old NICs.
1794          */
1795         switch (spawn->pci_dev->id.device_id) {
1796         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1797         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1798         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1799         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1800                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1801                 config->hw_vlan_insert = 0;
1802                 break;
1803         case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1804         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1805         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1806         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1807                 /*
1808                  * These NICs support VLAN insertion from WQE and
1809                  * report the wqe_vlan_insert flag. But there is the bug
1810                  * and PFC control may be broken, so disable feature.
1811                  */
1812                 config->hw_vlan_insert = 0;
1813                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1814                 break;
1815         default:
1816                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1817                 break;
1818         }
1819 exit:
1820         DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1821 }
1822
1823 /**
1824  * Configures the metadata mask fields in the shared context.
1825  *
1826  * @param [in] dev
1827  *   Pointer to Ethernet device.
1828  */
1829 static void
1830 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1831 {
1832         struct mlx5_priv *priv = dev->data->dev_private;
1833         struct mlx5_ibv_shared *sh = priv->sh;
1834         uint32_t meta, mark, reg_c0;
1835
1836         reg_c0 = ~priv->vport_meta_mask;
1837         switch (priv->config.dv_xmeta_en) {
1838         case MLX5_XMETA_MODE_LEGACY:
1839                 meta = UINT32_MAX;
1840                 mark = MLX5_FLOW_MARK_MASK;
1841                 break;
1842         case MLX5_XMETA_MODE_META16:
1843                 meta = reg_c0 >> rte_bsf32(reg_c0);
1844                 mark = MLX5_FLOW_MARK_MASK;
1845                 break;
1846         case MLX5_XMETA_MODE_META32:
1847                 meta = UINT32_MAX;
1848                 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1849                 break;
1850         default:
1851                 meta = 0;
1852                 mark = 0;
1853                 assert(false);
1854                 break;
1855         }
1856         if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1857                 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1858                                  sh->dv_mark_mask, mark);
1859         else
1860                 sh->dv_mark_mask = mark;
1861         if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1862                 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1863                                  sh->dv_meta_mask, meta);
1864         else
1865                 sh->dv_meta_mask = meta;
1866         if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1867                 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1868                                  sh->dv_meta_mask, reg_c0);
1869         else
1870                 sh->dv_regc0_mask = reg_c0;
1871         DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1872         DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1873         DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1874         DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1875 }
1876
1877 /**
1878  * Allocate page of door-bells and register it using DevX API.
1879  *
1880  * @param [in] dev
1881  *   Pointer to Ethernet device.
1882  *
1883  * @return
1884  *   Pointer to new page on success, NULL otherwise.
1885  */
1886 static struct mlx5_devx_dbr_page *
1887 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1888 {
1889         struct mlx5_priv *priv = dev->data->dev_private;
1890         struct mlx5_devx_dbr_page *page;
1891
1892         /* Allocate space for door-bell page and management data. */
1893         page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1894                                  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1895         if (!page) {
1896                 DRV_LOG(ERR, "port %u cannot allocate dbr page",
1897                         dev->data->port_id);
1898                 return NULL;
1899         }
1900         /* Register allocated memory. */
1901         page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1902                                               MLX5_DBR_PAGE_SIZE, 0);
1903         if (!page->umem) {
1904                 DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1905                         dev->data->port_id);
1906                 rte_free(page);
1907                 return NULL;
1908         }
1909         return page;
1910 }
1911
1912 /**
1913  * Find the next available door-bell, allocate new page if needed.
1914  *
1915  * @param [in] dev
1916  *   Pointer to Ethernet device.
1917  * @param [out] dbr_page
1918  *   Door-bell page containing the page data.
1919  *
1920  * @return
1921  *   Door-bell address offset on success, a negative error value otherwise.
1922  */
1923 int64_t
1924 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1925 {
1926         struct mlx5_priv *priv = dev->data->dev_private;
1927         struct mlx5_devx_dbr_page *page = NULL;
1928         uint32_t i, j;
1929
1930         LIST_FOREACH(page, &priv->dbrpgs, next)
1931                 if (page->dbr_count < MLX5_DBR_PER_PAGE)
1932                         break;
1933         if (!page) { /* No page with free door-bell exists. */
1934                 page = mlx5_alloc_dbr_page(dev);
1935                 if (!page) /* Failed to allocate new page. */
1936                         return (-1);
1937                 LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1938         }
1939         /* Loop to find bitmap part with clear bit. */
1940         for (i = 0;
1941              i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1942              i++)
1943                 ; /* Empty. */
1944         /* Find the first clear bit. */
1945         j = rte_bsf64(~page->dbr_bitmap[i]);
1946         assert(i < (MLX5_DBR_PER_PAGE / 64));
1947         page->dbr_bitmap[i] |= (1 << j);
1948         page->dbr_count++;
1949         *dbr_page = page;
1950         return (((i * 64) + j) * sizeof(uint64_t));
1951 }
1952
1953 /**
1954  * Release a door-bell record.
1955  *
1956  * @param [in] dev
1957  *   Pointer to Ethernet device.
1958  * @param [in] umem_id
1959  *   UMEM ID of page containing the door-bell record to release.
1960  * @param [in] offset
1961  *   Offset of door-bell record in page.
1962  *
1963  * @return
1964  *   0 on success, a negative error value otherwise.
1965  */
1966 int32_t
1967 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1968 {
1969         struct mlx5_priv *priv = dev->data->dev_private;
1970         struct mlx5_devx_dbr_page *page = NULL;
1971         int ret = 0;
1972
1973         LIST_FOREACH(page, &priv->dbrpgs, next)
1974                 /* Find the page this address belongs to. */
1975                 if (page->umem->umem_id == umem_id)
1976                         break;
1977         if (!page)
1978                 return -EINVAL;
1979         page->dbr_count--;
1980         if (!page->dbr_count) {
1981                 /* Page not used, free it and remove from list. */
1982                 LIST_REMOVE(page, next);
1983                 if (page->umem)
1984                         ret = -mlx5_glue->devx_umem_dereg(page->umem);
1985                 rte_free(page);
1986         } else {
1987                 /* Mark in bitmap that this door-bell is not in use. */
1988                 offset /= MLX5_DBR_SIZE;
1989                 int i = offset / 64;
1990                 int j = offset % 64;
1991
1992                 page->dbr_bitmap[i] &= ~(1 << j);
1993         }
1994         return ret;
1995 }
1996
1997 /**
1998  * Check sibling device configurations.
1999  *
2000  * Sibling devices sharing the Infiniband device context
2001  * should have compatible configurations. This regards
2002  * representors and bonding slaves.
2003  *
2004  * @param priv
2005  *   Private device descriptor.
2006  * @param config
2007  *   Configuration of the device is going to be created.
2008  *
2009  * @return
2010  *   0 on success, EINVAL otherwise
2011  */
2012 static int
2013 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2014                               struct mlx5_dev_config *config)
2015 {
2016         struct mlx5_ibv_shared *sh = priv->sh;
2017         struct mlx5_dev_config *sh_conf = NULL;
2018         uint16_t port_id;
2019
2020         assert(sh);
2021         /* Nothing to compare for the single/first device. */
2022         if (sh->refcnt == 1)
2023                 return 0;
2024         /* Find the device with shared context. */
2025         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2026                 struct mlx5_priv *opriv =
2027                         rte_eth_devices[port_id].data->dev_private;
2028
2029                 if (opriv && opriv != priv && opriv->sh == sh) {
2030                         sh_conf = &opriv->config;
2031                         break;
2032                 }
2033         }
2034         if (!sh_conf)
2035                 return 0;
2036         if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2037                 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2038                              " for shared %s context", sh->ibdev_name);
2039                 rte_errno = EINVAL;
2040                 return rte_errno;
2041         }
2042         if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2043                 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2044                              " for shared %s context", sh->ibdev_name);
2045                 rte_errno = EINVAL;
2046                 return rte_errno;
2047         }
2048         return 0;
2049 }
2050 /**
2051  * Spawn an Ethernet device from Verbs information.
2052  *
2053  * @param dpdk_dev
2054  *   Backing DPDK device.
2055  * @param spawn
2056  *   Verbs device parameters (name, port, switch_info) to spawn.
2057  * @param config
2058  *   Device configuration parameters.
2059  *
2060  * @return
2061  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
2062  *   is set. The following errors are defined:
2063  *
2064  *   EBUSY: device is not supposed to be spawned.
2065  *   EEXIST: device is already spawned
2066  */
2067 static struct rte_eth_dev *
2068 mlx5_dev_spawn(struct rte_device *dpdk_dev,
2069                struct mlx5_dev_spawn_data *spawn,
2070                struct mlx5_dev_config config)
2071 {
2072         const struct mlx5_switch_info *switch_info = &spawn->info;
2073         struct mlx5_ibv_shared *sh = NULL;
2074         struct ibv_port_attr port_attr;
2075         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
2076         struct rte_eth_dev *eth_dev = NULL;
2077         struct mlx5_priv *priv = NULL;
2078         int err = 0;
2079         unsigned int hw_padding = 0;
2080         unsigned int mps;
2081         unsigned int cqe_comp;
2082         unsigned int cqe_pad = 0;
2083         unsigned int tunnel_en = 0;
2084         unsigned int mpls_en = 0;
2085         unsigned int swp = 0;
2086         unsigned int mprq = 0;
2087         unsigned int mprq_min_stride_size_n = 0;
2088         unsigned int mprq_max_stride_size_n = 0;
2089         unsigned int mprq_min_stride_num_n = 0;
2090         unsigned int mprq_max_stride_num_n = 0;
2091         struct rte_ether_addr mac;
2092         char name[RTE_ETH_NAME_MAX_LEN];
2093         int own_domain_id = 0;
2094         uint16_t port_id;
2095         unsigned int i;
2096 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2097         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
2098 #endif
2099
2100         /* Determine if this port representor is supposed to be spawned. */
2101         if (switch_info->representor && dpdk_dev->devargs) {
2102                 struct rte_eth_devargs eth_da;
2103
2104                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
2105                 if (err) {
2106                         rte_errno = -err;
2107                         DRV_LOG(ERR, "failed to process device arguments: %s",
2108                                 strerror(rte_errno));
2109                         return NULL;
2110                 }
2111                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
2112                         if (eth_da.representor_ports[i] ==
2113                             (uint16_t)switch_info->port_name)
2114                                 break;
2115                 if (i == eth_da.nb_representor_ports) {
2116                         rte_errno = EBUSY;
2117                         return NULL;
2118                 }
2119         }
2120         /* Build device name. */
2121         if (spawn->pf_bond <  0) {
2122                 /* Single device. */
2123                 if (!switch_info->representor)
2124                         strlcpy(name, dpdk_dev->name, sizeof(name));
2125                 else
2126                         snprintf(name, sizeof(name), "%s_representor_%u",
2127                                  dpdk_dev->name, switch_info->port_name);
2128         } else {
2129                 /* Bonding device. */
2130                 if (!switch_info->representor)
2131                         snprintf(name, sizeof(name), "%s_%s",
2132                                  dpdk_dev->name, spawn->ibv_dev->name);
2133                 else
2134                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
2135                                  dpdk_dev->name, spawn->ibv_dev->name,
2136                                  switch_info->port_name);
2137         }
2138         /* check if the device is already spawned */
2139         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2140                 rte_errno = EEXIST;
2141                 return NULL;
2142         }
2143         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2144         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2145                 eth_dev = rte_eth_dev_attach_secondary(name);
2146                 if (eth_dev == NULL) {
2147                         DRV_LOG(ERR, "can not attach rte ethdev");
2148                         rte_errno = ENOMEM;
2149                         return NULL;
2150                 }
2151                 eth_dev->device = dpdk_dev;
2152                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
2153                 err = mlx5_proc_priv_init(eth_dev);
2154                 if (err)
2155                         return NULL;
2156                 /* Receive command fd from primary process */
2157                 err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
2158                 if (err < 0)
2159                         return NULL;
2160                 /* Remap UAR for Tx queues. */
2161                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
2162                 if (err)
2163                         return NULL;
2164                 /*
2165                  * Ethdev pointer is still required as input since
2166                  * the primary device is not accessible from the
2167                  * secondary process.
2168                  */
2169                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2170                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2171                 return eth_dev;
2172         }
2173         /*
2174          * Some parameters ("tx_db_nc" in particularly) are needed in
2175          * advance to create dv/verbs device context. We proceed the
2176          * devargs here to get ones, and later proceed devargs again
2177          * to override some hardware settings.
2178          */
2179         err = mlx5_args(&config, dpdk_dev->devargs);
2180         if (err) {
2181                 err = rte_errno;
2182                 DRV_LOG(ERR, "failed to process device arguments: %s",
2183                         strerror(rte_errno));
2184                 goto error;
2185         }
2186         sh = mlx5_alloc_shared_ibctx(spawn, &config);
2187         if (!sh)
2188                 return NULL;
2189         config.devx = sh->devx;
2190 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2191         config.dest_tir = 1;
2192 #endif
2193 #ifdef HAVE_IBV_MLX5_MOD_SWP
2194         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2195 #endif
2196         /*
2197          * Multi-packet send is supported by ConnectX-4 Lx PF as well
2198          * as all ConnectX-5 devices.
2199          */
2200 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2201         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2202 #endif
2203 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2204         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2205 #endif
2206         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2207         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2208                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2209                         DRV_LOG(DEBUG, "enhanced MPW is supported");
2210                         mps = MLX5_MPW_ENHANCED;
2211                 } else {
2212                         DRV_LOG(DEBUG, "MPW is supported");
2213                         mps = MLX5_MPW;
2214                 }
2215         } else {
2216                 DRV_LOG(DEBUG, "MPW isn't supported");
2217                 mps = MLX5_MPW_DISABLED;
2218         }
2219 #ifdef HAVE_IBV_MLX5_MOD_SWP
2220         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2221                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2222         DRV_LOG(DEBUG, "SWP support: %u", swp);
2223 #endif
2224         config.swp = !!swp;
2225 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2226         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2227                 struct mlx5dv_striding_rq_caps mprq_caps =
2228                         dv_attr.striding_rq_caps;
2229
2230                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2231                         mprq_caps.min_single_stride_log_num_of_bytes);
2232                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2233                         mprq_caps.max_single_stride_log_num_of_bytes);
2234                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2235                         mprq_caps.min_single_wqe_log_num_of_strides);
2236                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2237                         mprq_caps.max_single_wqe_log_num_of_strides);
2238                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2239                         mprq_caps.supported_qpts);
2240                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2241                 mprq = 1;
2242                 mprq_min_stride_size_n =
2243                         mprq_caps.min_single_stride_log_num_of_bytes;
2244                 mprq_max_stride_size_n =
2245                         mprq_caps.max_single_stride_log_num_of_bytes;
2246                 mprq_min_stride_num_n =
2247                         mprq_caps.min_single_wqe_log_num_of_strides;
2248                 mprq_max_stride_num_n =
2249                         mprq_caps.max_single_wqe_log_num_of_strides;
2250                 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2251                                                    mprq_min_stride_num_n);
2252         }
2253 #endif
2254         if (RTE_CACHE_LINE_SIZE == 128 &&
2255             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2256                 cqe_comp = 0;
2257         else
2258                 cqe_comp = 1;
2259         config.cqe_comp = cqe_comp;
2260 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2261         /* Whether device supports 128B Rx CQE padding. */
2262         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2263                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2264 #endif
2265 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2266         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2267                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
2268                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2269                              (dv_attr.tunnel_offloads_caps &
2270                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
2271                              (dv_attr.tunnel_offloads_caps &
2272                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
2273         }
2274         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2275                 tunnel_en ? "" : "not ");
2276 #else
2277         DRV_LOG(WARNING,
2278                 "tunnel offloading disabled due to old OFED/rdma-core version");
2279 #endif
2280         config.tunnel_en = tunnel_en;
2281 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2282         mpls_en = ((dv_attr.tunnel_offloads_caps &
2283                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2284                    (dv_attr.tunnel_offloads_caps &
2285                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2286         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2287                 mpls_en ? "" : "not ");
2288 #else
2289         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2290                 " old OFED/rdma-core version or firmware configuration");
2291 #endif
2292         config.mpls_en = mpls_en;
2293         /* Check port status. */
2294         err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2295         if (err) {
2296                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
2297                 goto error;
2298         }
2299         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2300                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
2301                 err = EINVAL;
2302                 goto error;
2303         }
2304         if (port_attr.state != IBV_PORT_ACTIVE)
2305                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2306                         mlx5_glue->port_state_str(port_attr.state),
2307                         port_attr.state);
2308         /* Allocate private eth device data. */
2309         priv = rte_zmalloc("ethdev private structure",
2310                            sizeof(*priv),
2311                            RTE_CACHE_LINE_SIZE);
2312         if (priv == NULL) {
2313                 DRV_LOG(ERR, "priv allocation failure");
2314                 err = ENOMEM;
2315                 goto error;
2316         }
2317         priv->sh = sh;
2318         priv->ibv_port = spawn->ibv_port;
2319         priv->pci_dev = spawn->pci_dev;
2320         priv->mtu = RTE_ETHER_MTU;
2321 #ifndef RTE_ARCH_64
2322         /* Initialize UAR access locks for 32bit implementations. */
2323         rte_spinlock_init(&priv->uar_lock_cq);
2324         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2325                 rte_spinlock_init(&priv->uar_lock[i]);
2326 #endif
2327         /* Some internal functions rely on Netlink sockets, open them now. */
2328         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2329         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
2330         priv->nl_sn = 0;
2331         priv->representor = !!switch_info->representor;
2332         priv->master = !!switch_info->master;
2333         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2334         priv->vport_meta_tag = 0;
2335         priv->vport_meta_mask = 0;
2336         priv->pf_bond = spawn->pf_bond;
2337 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2338         /*
2339          * The DevX port query API is implemented. E-Switch may use
2340          * either vport or reg_c[0] metadata register to match on
2341          * vport index. The engaged part of metadata register is
2342          * defined by mask.
2343          */
2344         if (switch_info->representor || switch_info->master) {
2345                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2346                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2347                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2348                                                  &devx_port);
2349                 if (err) {
2350                         DRV_LOG(WARNING,
2351                                 "can't query devx port %d on device %s",
2352                                 spawn->ibv_port, spawn->ibv_dev->name);
2353                         devx_port.comp_mask = 0;
2354                 }
2355         }
2356         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2357                 priv->vport_meta_tag = devx_port.reg_c_0.value;
2358                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
2359                 if (!priv->vport_meta_mask) {
2360                         DRV_LOG(ERR, "vport zero mask for port %d"
2361                                      " on bonding device %s",
2362                                      spawn->ibv_port, spawn->ibv_dev->name);
2363                         err = ENOTSUP;
2364                         goto error;
2365                 }
2366                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2367                         DRV_LOG(ERR, "invalid vport tag for port %d"
2368                                      " on bonding device %s",
2369                                      spawn->ibv_port, spawn->ibv_dev->name);
2370                         err = ENOTSUP;
2371                         goto error;
2372                 }
2373         }
2374         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2375                 priv->vport_id = devx_port.vport_num;
2376         } else if (spawn->pf_bond >= 0) {
2377                 DRV_LOG(ERR, "can't deduce vport index for port %d"
2378                              " on bonding device %s",
2379                              spawn->ibv_port, spawn->ibv_dev->name);
2380                 err = ENOTSUP;
2381                 goto error;
2382         } else {
2383                 /* Suppose vport index in compatible way. */
2384                 priv->vport_id = switch_info->representor ?
2385                                  switch_info->port_name + 1 : -1;
2386         }
2387 #else
2388         /*
2389          * Kernel/rdma_core support single E-Switch per PF configurations
2390          * only and vport_id field contains the vport index for
2391          * associated VF, which is deduced from representor port name.
2392          * For example, let's have the IB device port 10, it has
2393          * attached network device eth0, which has port name attribute
2394          * pf0vf2, we can deduce the VF number as 2, and set vport index
2395          * as 3 (2+1). This assigning schema should be changed if the
2396          * multiple E-Switch instances per PF configurations or/and PCI
2397          * subfunctions are added.
2398          */
2399         priv->vport_id = switch_info->representor ?
2400                          switch_info->port_name + 1 : -1;
2401 #endif
2402         /* representor_id field keeps the unmodified VF index. */
2403         priv->representor_id = switch_info->representor ?
2404                                switch_info->port_name : -1;
2405         /*
2406          * Look for sibling devices in order to reuse their switch domain
2407          * if any, otherwise allocate one.
2408          */
2409         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2410                 const struct mlx5_priv *opriv =
2411                         rte_eth_devices[port_id].data->dev_private;
2412
2413                 if (!opriv ||
2414                     opriv->sh != priv->sh ||
2415                         opriv->domain_id ==
2416                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2417                         continue;
2418                 priv->domain_id = opriv->domain_id;
2419                 break;
2420         }
2421         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2422                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
2423                 if (err) {
2424                         err = rte_errno;
2425                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
2426                                 strerror(rte_errno));
2427                         goto error;
2428                 }
2429                 own_domain_id = 1;
2430         }
2431         /* Override some values set by hardware configuration. */
2432         mlx5_args(&config, dpdk_dev->devargs);
2433         err = mlx5_dev_check_sibling_config(priv, &config);
2434         if (err)
2435                 goto error;
2436         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2437                             IBV_DEVICE_RAW_IP_CSUM);
2438         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2439                 (config.hw_csum ? "" : "not "));
2440 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2441         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2442         DRV_LOG(DEBUG, "counters are not supported");
2443 #endif
2444 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
2445         if (config.dv_flow_en) {
2446                 DRV_LOG(WARNING, "DV flow is not supported");
2447                 config.dv_flow_en = 0;
2448         }
2449 #endif
2450         config.ind_table_max_size =
2451                 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2452         /*
2453          * Remove this check once DPDK supports larger/variable
2454          * indirection tables.
2455          */
2456         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2457                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2458         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2459                 config.ind_table_max_size);
2460         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2461                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2462         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2463                 (config.hw_vlan_strip ? "" : "not "));
2464         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2465                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
2466         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2467                 (config.hw_fcs_strip ? "" : "not "));
2468 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2469         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2470 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2471         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2472                         IBV_DEVICE_PCI_WRITE_END_PADDING);
2473 #endif
2474         if (config.hw_padding && !hw_padding) {
2475                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2476                 config.hw_padding = 0;
2477         } else if (config.hw_padding) {
2478                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2479         }
2480         config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2481                       (sh->device_attr.tso_caps.supported_qpts &
2482                        (1 << IBV_QPT_RAW_PACKET)));
2483         if (config.tso)
2484                 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2485         /*
2486          * MPW is disabled by default, while the Enhanced MPW is enabled
2487          * by default.
2488          */
2489         if (config.mps == MLX5_ARG_UNSET)
2490                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2491                                                           MLX5_MPW_DISABLED;
2492         else
2493                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2494         DRV_LOG(INFO, "%sMPS is %s",
2495                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
2496                 config.mps == MLX5_MPW ? "legacy " : "",
2497                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2498         if (config.cqe_comp && !cqe_comp) {
2499                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2500                 config.cqe_comp = 0;
2501         }
2502         if (config.cqe_pad && !cqe_pad) {
2503                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2504                 config.cqe_pad = 0;
2505         } else if (config.cqe_pad) {
2506                 DRV_LOG(INFO, "Rx CQE padding is enabled");
2507         }
2508         if (config.devx) {
2509                 priv->counter_fallback = 0;
2510                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2511                 if (err) {
2512                         err = -err;
2513                         goto error;
2514                 }
2515                 if (!config.hca_attr.flow_counters_dump)
2516                         priv->counter_fallback = 1;
2517 #ifndef HAVE_IBV_DEVX_ASYNC
2518                 priv->counter_fallback = 1;
2519 #endif
2520                 if (priv->counter_fallback)
2521                         DRV_LOG(INFO, "Use fall-back DV counter management");
2522                 /* Check for LRO support. */
2523                 if (config.dest_tir && config.hca_attr.lro_cap &&
2524                     config.dv_flow_en) {
2525                         /* TBD check tunnel lro caps. */
2526                         config.lro.supported = config.hca_attr.lro_cap;
2527                         DRV_LOG(DEBUG, "Device supports LRO");
2528                         /*
2529                          * If LRO timeout is not configured by application,
2530                          * use the minimal supported value.
2531                          */
2532                         if (!config.lro.timeout)
2533                                 config.lro.timeout =
2534                                 config.hca_attr.lro_timer_supported_periods[0];
2535                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2536                                 config.lro.timeout);
2537                 }
2538 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2539                 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2540                     config.dv_flow_en) {
2541                         uint8_t reg_c_mask =
2542                                 config.hca_attr.qos.flow_meter_reg_c_ids;
2543                         /*
2544                          * Meter needs two REG_C's for color match and pre-sfx
2545                          * flow match. Here get the REG_C for color match.
2546                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
2547                          */
2548                         reg_c_mask &= 0xfc;
2549                         if (__builtin_popcount(reg_c_mask) < 1) {
2550                                 priv->mtr_en = 0;
2551                                 DRV_LOG(WARNING, "No available register for"
2552                                         " meter.");
2553                         } else {
2554                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2555                                                       REG_C_0;
2556                                 priv->mtr_en = 1;
2557                                 priv->mtr_reg_share =
2558                                       config.hca_attr.qos.flow_meter_reg_share;
2559                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2560                                         priv->mtr_color_reg);
2561                         }
2562                 }
2563 #endif
2564         }
2565         if (config.mprq.enabled && mprq) {
2566                 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2567                     config.mprq.stride_num_n < mprq_min_stride_num_n) {
2568                         config.mprq.stride_num_n =
2569                                 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2570                                         mprq_min_stride_num_n);
2571                         DRV_LOG(WARNING,
2572                                 "the number of strides"
2573                                 " for Multi-Packet RQ is out of range,"
2574                                 " setting default value (%u)",
2575                                 1 << config.mprq.stride_num_n);
2576                 }
2577                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2578                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2579         } else if (config.mprq.enabled && !mprq) {
2580                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2581                 config.mprq.enabled = 0;
2582         }
2583         if (config.max_dump_files_num == 0)
2584                 config.max_dump_files_num = 128;
2585         eth_dev = rte_eth_dev_allocate(name);
2586         if (eth_dev == NULL) {
2587                 DRV_LOG(ERR, "can not allocate rte ethdev");
2588                 err = ENOMEM;
2589                 goto error;
2590         }
2591         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2592         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2593         if (priv->representor) {
2594                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2595                 eth_dev->data->representor_id = priv->representor_id;
2596         }
2597         /*
2598          * Store associated network device interface index. This index
2599          * is permanent throughout the lifetime of device. So, we may store
2600          * the ifindex here and use the cached value further.
2601          */
2602         assert(spawn->ifindex);
2603         priv->if_index = spawn->ifindex;
2604         eth_dev->data->dev_private = priv;
2605         priv->dev_data = eth_dev->data;
2606         eth_dev->data->mac_addrs = priv->mac;
2607         eth_dev->device = dpdk_dev;
2608         /* Configure the first MAC address by default. */
2609         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2610                 DRV_LOG(ERR,
2611                         "port %u cannot get MAC address, is mlx5_en"
2612                         " loaded? (errno: %s)",
2613                         eth_dev->data->port_id, strerror(rte_errno));
2614                 err = ENODEV;
2615                 goto error;
2616         }
2617         DRV_LOG(INFO,
2618                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2619                 eth_dev->data->port_id,
2620                 mac.addr_bytes[0], mac.addr_bytes[1],
2621                 mac.addr_bytes[2], mac.addr_bytes[3],
2622                 mac.addr_bytes[4], mac.addr_bytes[5]);
2623 #ifndef NDEBUG
2624         {
2625                 char ifname[IF_NAMESIZE];
2626
2627                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2628                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2629                                 eth_dev->data->port_id, ifname);
2630                 else
2631                         DRV_LOG(DEBUG, "port %u ifname is unknown",
2632                                 eth_dev->data->port_id);
2633         }
2634 #endif
2635         /* Get actual MTU if possible. */
2636         err = mlx5_get_mtu(eth_dev, &priv->mtu);
2637         if (err) {
2638                 err = rte_errno;
2639                 goto error;
2640         }
2641         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2642                 priv->mtu);
2643         /* Initialize burst functions to prevent crashes before link-up. */
2644         eth_dev->rx_pkt_burst = removed_rx_burst;
2645         eth_dev->tx_pkt_burst = removed_tx_burst;
2646         eth_dev->dev_ops = &mlx5_dev_ops;
2647         /* Register MAC address. */
2648         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2649         if (config.vf && config.vf_nl_en)
2650                 mlx5_nl_mac_addr_sync(eth_dev);
2651         TAILQ_INIT(&priv->flows);
2652         TAILQ_INIT(&priv->ctrl_flows);
2653         TAILQ_INIT(&priv->flow_meters);
2654         TAILQ_INIT(&priv->flow_meter_profiles);
2655         /* Hint libmlx5 to use PMD allocator for data plane resources */
2656         struct mlx5dv_ctx_allocators alctr = {
2657                 .alloc = &mlx5_alloc_verbs_buf,
2658                 .free = &mlx5_free_verbs_buf,
2659                 .data = priv,
2660         };
2661         mlx5_glue->dv_set_context_attr(sh->ctx,
2662                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2663                                        (void *)((uintptr_t)&alctr));
2664         /* Bring Ethernet device up. */
2665         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2666                 eth_dev->data->port_id);
2667         mlx5_set_link_up(eth_dev);
2668         /*
2669          * Even though the interrupt handler is not installed yet,
2670          * interrupts will still trigger on the async_fd from
2671          * Verbs context returned by ibv_open_device().
2672          */
2673         mlx5_link_update(eth_dev, 0);
2674 #ifdef HAVE_MLX5DV_DR_ESWITCH
2675         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2676               (switch_info->representor || switch_info->master)))
2677                 config.dv_esw_en = 0;
2678 #else
2679         config.dv_esw_en = 0;
2680 #endif
2681         /* Detect minimal data bytes to inline. */
2682         mlx5_set_min_inline(spawn, &config);
2683         /* Store device configuration on private structure. */
2684         priv->config = config;
2685         /* Create context for virtual machine VLAN workaround. */
2686         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2687         if (config.dv_flow_en) {
2688                 err = mlx5_alloc_shared_dr(priv);
2689                 if (err)
2690                         goto error;
2691                 /*
2692                  * RSS id is shared with meter flow id. Meter flow id can only
2693                  * use the 24 MSB of the register.
2694                  */
2695                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
2696                                      MLX5_MTR_COLOR_BITS);
2697                 if (!priv->qrss_id_pool) {
2698                         DRV_LOG(ERR, "can't create flow id pool");
2699                         err = ENOMEM;
2700                         goto error;
2701                 }
2702         }
2703         /* Supported Verbs flow priority number detection. */
2704         err = mlx5_flow_discover_priorities(eth_dev);
2705         if (err < 0) {
2706                 err = -err;
2707                 goto error;
2708         }
2709         priv->config.flow_prio = err;
2710         if (!priv->config.dv_esw_en &&
2711             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2712                 DRV_LOG(WARNING, "metadata mode %u is not supported "
2713                                  "(no E-Switch)", priv->config.dv_xmeta_en);
2714                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2715         }
2716         mlx5_set_metadata_mask(eth_dev);
2717         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2718             !priv->sh->dv_regc0_mask) {
2719                 DRV_LOG(ERR, "metadata mode %u is not supported "
2720                              "(no metadata reg_c[0] is available)",
2721                              priv->config.dv_xmeta_en);
2722                         err = ENOTSUP;
2723                         goto error;
2724         }
2725         /* Query availibility of metadata reg_c's. */
2726         err = mlx5_flow_discover_mreg_c(eth_dev);
2727         if (err < 0) {
2728                 err = -err;
2729                 goto error;
2730         }
2731         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
2732                 DRV_LOG(DEBUG,
2733                         "port %u extensive metadata register is not supported",
2734                         eth_dev->data->port_id);
2735                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2736                         DRV_LOG(ERR, "metadata mode %u is not supported "
2737                                      "(no metadata registers available)",
2738                                      priv->config.dv_xmeta_en);
2739                         err = ENOTSUP;
2740                         goto error;
2741                 }
2742         }
2743         if (priv->config.dv_flow_en &&
2744             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2745             mlx5_flow_ext_mreg_supported(eth_dev) &&
2746             priv->sh->dv_regc0_mask) {
2747                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
2748                                                       MLX5_FLOW_MREG_HTABLE_SZ);
2749                 if (!priv->mreg_cp_tbl) {
2750                         err = ENOMEM;
2751                         goto error;
2752                 }
2753         }
2754         return eth_dev;
2755 error:
2756         if (priv) {
2757                 if (priv->mreg_cp_tbl)
2758                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
2759                 if (priv->sh)
2760                         mlx5_free_shared_dr(priv);
2761                 if (priv->nl_socket_route >= 0)
2762                         close(priv->nl_socket_route);
2763                 if (priv->nl_socket_rdma >= 0)
2764                         close(priv->nl_socket_rdma);
2765                 if (priv->vmwa_context)
2766                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
2767                 if (priv->qrss_id_pool)
2768                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
2769                 if (own_domain_id)
2770                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2771                 rte_free(priv);
2772                 if (eth_dev != NULL)
2773                         eth_dev->data->dev_private = NULL;
2774         }
2775         if (eth_dev != NULL) {
2776                 /* mac_addrs must not be freed alone because part of dev_private */
2777                 eth_dev->data->mac_addrs = NULL;
2778                 rte_eth_dev_release_port(eth_dev);
2779         }
2780         if (sh)
2781                 mlx5_free_shared_ibctx(sh);
2782         assert(err > 0);
2783         rte_errno = err;
2784         return NULL;
2785 }
2786
2787 /**
2788  * Comparison callback to sort device data.
2789  *
2790  * This is meant to be used with qsort().
2791  *
2792  * @param a[in]
2793  *   Pointer to pointer to first data object.
2794  * @param b[in]
2795  *   Pointer to pointer to second data object.
2796  *
2797  * @return
2798  *   0 if both objects are equal, less than 0 if the first argument is less
2799  *   than the second, greater than 0 otherwise.
2800  */
2801 static int
2802 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2803 {
2804         const struct mlx5_switch_info *si_a =
2805                 &((const struct mlx5_dev_spawn_data *)a)->info;
2806         const struct mlx5_switch_info *si_b =
2807                 &((const struct mlx5_dev_spawn_data *)b)->info;
2808         int ret;
2809
2810         /* Master device first. */
2811         ret = si_b->master - si_a->master;
2812         if (ret)
2813                 return ret;
2814         /* Then representor devices. */
2815         ret = si_b->representor - si_a->representor;
2816         if (ret)
2817                 return ret;
2818         /* Unidentified devices come last in no specific order. */
2819         if (!si_a->representor)
2820                 return 0;
2821         /* Order representors by name. */
2822         return si_a->port_name - si_b->port_name;
2823 }
2824
2825 /**
2826  * Match PCI information for possible slaves of bonding device.
2827  *
2828  * @param[in] ibv_dev
2829  *   Pointer to Infiniband device structure.
2830  * @param[in] pci_dev
2831  *   Pointer to PCI device structure to match PCI address.
2832  * @param[in] nl_rdma
2833  *   Netlink RDMA group socket handle.
2834  *
2835  * @return
2836  *   negative value if no bonding device found, otherwise
2837  *   positive index of slave PF in bonding.
2838  */
2839 static int
2840 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2841                            const struct rte_pci_device *pci_dev,
2842                            int nl_rdma)
2843 {
2844         char ifname[IF_NAMESIZE + 1];
2845         unsigned int ifindex;
2846         unsigned int np, i;
2847         FILE *file = NULL;
2848         int pf = -1;
2849
2850         /*
2851          * Try to get master device name. If something goes
2852          * wrong suppose the lack of kernel support and no
2853          * bonding devices.
2854          */
2855         if (nl_rdma < 0)
2856                 return -1;
2857         if (!strstr(ibv_dev->name, "bond"))
2858                 return -1;
2859         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2860         if (!np)
2861                 return -1;
2862         /*
2863          * The Master device might not be on the predefined
2864          * port (not on port index 1, it is not garanted),
2865          * we have to scan all Infiniband device port and
2866          * find master.
2867          */
2868         for (i = 1; i <= np; ++i) {
2869                 /* Check whether Infiniband port is populated. */
2870                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2871                 if (!ifindex)
2872                         continue;
2873                 if (!if_indextoname(ifindex, ifname))
2874                         continue;
2875                 /* Try to read bonding slave names from sysfs. */
2876                 MKSTR(slaves,
2877                       "/sys/class/net/%s/master/bonding/slaves", ifname);
2878                 file = fopen(slaves, "r");
2879                 if (file)
2880                         break;
2881         }
2882         if (!file)
2883                 return -1;
2884         /* Use safe format to check maximal buffer length. */
2885         assert(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2886         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2887                 char tmp_str[IF_NAMESIZE + 32];
2888                 struct rte_pci_addr pci_addr;
2889                 struct mlx5_switch_info info;
2890
2891                 /* Process slave interface names in the loop. */
2892                 snprintf(tmp_str, sizeof(tmp_str),
2893                          "/sys/class/net/%s", ifname);
2894                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2895                         DRV_LOG(WARNING, "can not get PCI address"
2896                                          " for netdev \"%s\"", ifname);
2897                         continue;
2898                 }
2899                 if (pci_dev->addr.domain != pci_addr.domain ||
2900                     pci_dev->addr.bus != pci_addr.bus ||
2901                     pci_dev->addr.devid != pci_addr.devid ||
2902                     pci_dev->addr.function != pci_addr.function)
2903                         continue;
2904                 /* Slave interface PCI address match found. */
2905                 fclose(file);
2906                 snprintf(tmp_str, sizeof(tmp_str),
2907                          "/sys/class/net/%s/phys_port_name", ifname);
2908                 file = fopen(tmp_str, "rb");
2909                 if (!file)
2910                         break;
2911                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2912                 if (fscanf(file, "%32s", tmp_str) == 1)
2913                         mlx5_translate_port_name(tmp_str, &info);
2914                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2915                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2916                         pf = info.port_name;
2917                 break;
2918         }
2919         if (file)
2920                 fclose(file);
2921         return pf;
2922 }
2923
2924 /**
2925  * DPDK callback to register a PCI device.
2926  *
2927  * This function spawns Ethernet devices out of a given PCI device.
2928  *
2929  * @param[in] pci_drv
2930  *   PCI driver structure (mlx5_driver).
2931  * @param[in] pci_dev
2932  *   PCI device information.
2933  *
2934  * @return
2935  *   0 on success, a negative errno value otherwise and rte_errno is set.
2936  */
2937 static int
2938 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2939                struct rte_pci_device *pci_dev)
2940 {
2941         struct ibv_device **ibv_list;
2942         /*
2943          * Number of found IB Devices matching with requested PCI BDF.
2944          * nd != 1 means there are multiple IB devices over the same
2945          * PCI device and we have representors and master.
2946          */
2947         unsigned int nd = 0;
2948         /*
2949          * Number of found IB device Ports. nd = 1 and np = 1..n means
2950          * we have the single multiport IB device, and there may be
2951          * representors attached to some of found ports.
2952          */
2953         unsigned int np = 0;
2954         /*
2955          * Number of DPDK ethernet devices to Spawn - either over
2956          * multiple IB devices or multiple ports of single IB device.
2957          * Actually this is the number of iterations to spawn.
2958          */
2959         unsigned int ns = 0;
2960         /*
2961          * Bonding device
2962          *   < 0 - no bonding device (single one)
2963          *  >= 0 - bonding device (value is slave PF index)
2964          */
2965         int bd = -1;
2966         struct mlx5_dev_spawn_data *list = NULL;
2967         struct mlx5_dev_config dev_config;
2968         int ret;
2969
2970         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2971                 mlx5_pmd_socket_init();
2972         ret = mlx5_init_once();
2973         if (ret) {
2974                 DRV_LOG(ERR, "unable to init PMD global data: %s",
2975                         strerror(rte_errno));
2976                 return -rte_errno;
2977         }
2978         assert(pci_drv == &mlx5_driver);
2979         errno = 0;
2980         ibv_list = mlx5_glue->get_device_list(&ret);
2981         if (!ibv_list) {
2982                 rte_errno = errno ? errno : ENOSYS;
2983                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2984                 return -rte_errno;
2985         }
2986         /*
2987          * First scan the list of all Infiniband devices to find
2988          * matching ones, gathering into the list.
2989          */
2990         struct ibv_device *ibv_match[ret + 1];
2991         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2992         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2993         unsigned int i;
2994
2995         while (ret-- > 0) {
2996                 struct rte_pci_addr pci_addr;
2997
2998                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2999                 bd = mlx5_device_bond_pci_match
3000                                 (ibv_list[ret], pci_dev, nl_rdma);
3001                 if (bd >= 0) {
3002                         /*
3003                          * Bonding device detected. Only one match is allowed,
3004                          * the bonding is supported over multi-port IB device,
3005                          * there should be no matches on representor PCI
3006                          * functions or non VF LAG bonding devices with
3007                          * specified address.
3008                          */
3009                         if (nd) {
3010                                 DRV_LOG(ERR,
3011                                         "multiple PCI match on bonding device"
3012                                         "\"%s\" found", ibv_list[ret]->name);
3013                                 rte_errno = ENOENT;
3014                                 ret = -rte_errno;
3015                                 goto exit;
3016                         }
3017                         DRV_LOG(INFO, "PCI information matches for"
3018                                       " slave %d bonding device \"%s\"",
3019                                       bd, ibv_list[ret]->name);
3020                         ibv_match[nd++] = ibv_list[ret];
3021                         break;
3022                 }
3023                 if (mlx5_dev_to_pci_addr
3024                         (ibv_list[ret]->ibdev_path, &pci_addr))
3025                         continue;
3026                 if (pci_dev->addr.domain != pci_addr.domain ||
3027                     pci_dev->addr.bus != pci_addr.bus ||
3028                     pci_dev->addr.devid != pci_addr.devid ||
3029                     pci_dev->addr.function != pci_addr.function)
3030                         continue;
3031                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
3032                         ibv_list[ret]->name);
3033                 ibv_match[nd++] = ibv_list[ret];
3034         }
3035         ibv_match[nd] = NULL;
3036         if (!nd) {
3037                 /* No device matches, just complain and bail out. */
3038                 DRV_LOG(WARNING,
3039                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
3040                         " are kernel drivers loaded?",
3041                         pci_dev->addr.domain, pci_dev->addr.bus,
3042                         pci_dev->addr.devid, pci_dev->addr.function);
3043                 rte_errno = ENOENT;
3044                 ret = -rte_errno;
3045                 goto exit;
3046         }
3047         if (nd == 1) {
3048                 /*
3049                  * Found single matching device may have multiple ports.
3050                  * Each port may be representor, we have to check the port
3051                  * number and check the representors existence.
3052                  */
3053                 if (nl_rdma >= 0)
3054                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
3055                 if (!np)
3056                         DRV_LOG(WARNING, "can not get IB device \"%s\""
3057                                          " ports number", ibv_match[0]->name);
3058                 if (bd >= 0 && !np) {
3059                         DRV_LOG(ERR, "can not get ports"
3060                                      " for bonding device");
3061                         rte_errno = ENOENT;
3062                         ret = -rte_errno;
3063                         goto exit;
3064                 }
3065         }
3066 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
3067         if (bd >= 0) {
3068                 /*
3069                  * This may happen if there is VF LAG kernel support and
3070                  * application is compiled with older rdma_core library.
3071                  */
3072                 DRV_LOG(ERR,
3073                         "No kernel/verbs support for VF LAG bonding found.");
3074                 rte_errno = ENOTSUP;
3075                 ret = -rte_errno;
3076                 goto exit;
3077         }
3078 #endif
3079         /*
3080          * Now we can determine the maximal
3081          * amount of devices to be spawned.
3082          */
3083         list = rte_zmalloc("device spawn data",
3084                          sizeof(struct mlx5_dev_spawn_data) *
3085                          (np ? np : nd),
3086                          RTE_CACHE_LINE_SIZE);
3087         if (!list) {
3088                 DRV_LOG(ERR, "spawn data array allocation failure");
3089                 rte_errno = ENOMEM;
3090                 ret = -rte_errno;
3091                 goto exit;
3092         }
3093         if (bd >= 0 || np > 1) {
3094                 /*
3095                  * Single IB device with multiple ports found,
3096                  * it may be E-Switch master device and representors.
3097                  * We have to perform identification trough the ports.
3098                  */
3099                 assert(nl_rdma >= 0);
3100                 assert(ns == 0);
3101                 assert(nd == 1);
3102                 assert(np);
3103                 for (i = 1; i <= np; ++i) {
3104                         list[ns].max_port = np;
3105                         list[ns].ibv_port = i;
3106                         list[ns].ibv_dev = ibv_match[0];
3107                         list[ns].eth_dev = NULL;
3108                         list[ns].pci_dev = pci_dev;
3109                         list[ns].pf_bond = bd;
3110                         list[ns].ifindex = mlx5_nl_ifindex
3111                                         (nl_rdma, list[ns].ibv_dev->name, i);
3112                         if (!list[ns].ifindex) {
3113                                 /*
3114                                  * No network interface index found for the
3115                                  * specified port, it means there is no
3116                                  * representor on this port. It's OK,
3117                                  * there can be disabled ports, for example
3118                                  * if sriov_numvfs < sriov_totalvfs.
3119                                  */
3120                                 continue;
3121                         }
3122                         ret = -1;
3123                         if (nl_route >= 0)
3124                                 ret = mlx5_nl_switch_info
3125                                                (nl_route,
3126                                                 list[ns].ifindex,
3127                                                 &list[ns].info);
3128                         if (ret || (!list[ns].info.representor &&
3129                                     !list[ns].info.master)) {
3130                                 /*
3131                                  * We failed to recognize representors with
3132                                  * Netlink, let's try to perform the task
3133                                  * with sysfs.
3134                                  */
3135                                 ret =  mlx5_sysfs_switch_info
3136                                                 (list[ns].ifindex,
3137                                                  &list[ns].info);
3138                         }
3139                         if (!ret && bd >= 0) {
3140                                 switch (list[ns].info.name_type) {
3141                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
3142                                         if (list[ns].info.port_name == bd)
3143                                                 ns++;
3144                                         break;
3145                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
3146                                         if (list[ns].info.pf_num == bd)
3147                                                 ns++;
3148                                         break;
3149                                 default:
3150                                         break;
3151                                 }
3152                                 continue;
3153                         }
3154                         if (!ret && (list[ns].info.representor ^
3155                                      list[ns].info.master))
3156                                 ns++;
3157                 }
3158                 if (!ns) {
3159                         DRV_LOG(ERR,
3160                                 "unable to recognize master/representors"
3161                                 " on the IB device with multiple ports");
3162                         rte_errno = ENOENT;
3163                         ret = -rte_errno;
3164                         goto exit;
3165                 }
3166         } else {
3167                 /*
3168                  * The existence of several matching entries (nd > 1) means
3169                  * port representors have been instantiated. No existing Verbs
3170                  * call nor sysfs entries can tell them apart, this can only
3171                  * be done through Netlink calls assuming kernel drivers are
3172                  * recent enough to support them.
3173                  *
3174                  * In the event of identification failure through Netlink,
3175                  * try again through sysfs, then:
3176                  *
3177                  * 1. A single IB device matches (nd == 1) with single
3178                  *    port (np=0/1) and is not a representor, assume
3179                  *    no switch support.
3180                  *
3181                  * 2. Otherwise no safe assumptions can be made;
3182                  *    complain louder and bail out.
3183                  */
3184                 np = 1;
3185                 for (i = 0; i != nd; ++i) {
3186                         memset(&list[ns].info, 0, sizeof(list[ns].info));
3187                         list[ns].max_port = 1;
3188                         list[ns].ibv_port = 1;
3189                         list[ns].ibv_dev = ibv_match[i];
3190                         list[ns].eth_dev = NULL;
3191                         list[ns].pci_dev = pci_dev;
3192                         list[ns].pf_bond = -1;
3193                         list[ns].ifindex = 0;
3194                         if (nl_rdma >= 0)
3195                                 list[ns].ifindex = mlx5_nl_ifindex
3196                                         (nl_rdma, list[ns].ibv_dev->name, 1);
3197                         if (!list[ns].ifindex) {
3198                                 char ifname[IF_NAMESIZE];
3199
3200                                 /*
3201                                  * Netlink failed, it may happen with old
3202                                  * ib_core kernel driver (before 4.16).
3203                                  * We can assume there is old driver because
3204                                  * here we are processing single ports IB
3205                                  * devices. Let's try sysfs to retrieve
3206                                  * the ifindex. The method works for
3207                                  * master device only.
3208                                  */
3209                                 if (nd > 1) {
3210                                         /*
3211                                          * Multiple devices found, assume
3212                                          * representors, can not distinguish
3213                                          * master/representor and retrieve
3214                                          * ifindex via sysfs.
3215                                          */
3216                                         continue;
3217                                 }
3218                                 ret = mlx5_get_master_ifname
3219                                         (ibv_match[i]->ibdev_path, &ifname);
3220                                 if (!ret)
3221                                         list[ns].ifindex =
3222                                                 if_nametoindex(ifname);
3223                                 if (!list[ns].ifindex) {
3224                                         /*
3225                                          * No network interface index found
3226                                          * for the specified device, it means
3227                                          * there it is neither representor
3228                                          * nor master.
3229                                          */
3230                                         continue;
3231                                 }
3232                         }
3233                         ret = -1;
3234                         if (nl_route >= 0)
3235                                 ret = mlx5_nl_switch_info
3236                                                (nl_route,
3237                                                 list[ns].ifindex,
3238                                                 &list[ns].info);
3239                         if (ret || (!list[ns].info.representor &&
3240                                     !list[ns].info.master)) {
3241                                 /*
3242                                  * We failed to recognize representors with
3243                                  * Netlink, let's try to perform the task
3244                                  * with sysfs.
3245                                  */
3246                                 ret =  mlx5_sysfs_switch_info
3247                                                 (list[ns].ifindex,
3248                                                  &list[ns].info);
3249                         }
3250                         if (!ret && (list[ns].info.representor ^
3251                                      list[ns].info.master)) {
3252                                 ns++;
3253                         } else if ((nd == 1) &&
3254                                    !list[ns].info.representor &&
3255                                    !list[ns].info.master) {
3256                                 /*
3257                                  * Single IB device with
3258                                  * one physical port and
3259                                  * attached network device.
3260                                  * May be SRIOV is not enabled
3261                                  * or there is no representors.
3262                                  */
3263                                 DRV_LOG(INFO, "no E-Switch support detected");
3264                                 ns++;
3265                                 break;
3266                         }
3267                 }
3268                 if (!ns) {
3269                         DRV_LOG(ERR,
3270                                 "unable to recognize master/representors"
3271                                 " on the multiple IB devices");
3272                         rte_errno = ENOENT;
3273                         ret = -rte_errno;
3274                         goto exit;
3275                 }
3276         }
3277         assert(ns);
3278         /*
3279          * Sort list to probe devices in natural order for users convenience
3280          * (i.e. master first, then representors from lowest to highest ID).
3281          */
3282         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3283         /* Default configuration. */
3284         dev_config = (struct mlx5_dev_config){
3285                 .hw_padding = 0,
3286                 .mps = MLX5_ARG_UNSET,
3287                 .dbnc = MLX5_ARG_UNSET,
3288                 .rx_vec_en = 1,
3289                 .txq_inline_max = MLX5_ARG_UNSET,
3290                 .txq_inline_min = MLX5_ARG_UNSET,
3291                 .txq_inline_mpw = MLX5_ARG_UNSET,
3292                 .txqs_inline = MLX5_ARG_UNSET,
3293                 .vf_nl_en = 1,
3294                 .mr_ext_memseg_en = 1,
3295                 .mprq = {
3296                         .enabled = 0, /* Disabled by default. */
3297                         .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
3298                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3299                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3300                 },
3301                 .dv_esw_en = 1,
3302                 .dv_flow_en = 1,
3303         };
3304         /* Device specific configuration. */
3305         switch (pci_dev->id.device_id) {
3306         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3307         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3308         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3309         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3310         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3311         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3312         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3313                 dev_config.vf = 1;
3314                 break;
3315         default:
3316                 break;
3317         }
3318         for (i = 0; i != ns; ++i) {
3319                 uint32_t restore;
3320
3321                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3322                                                  &list[i],
3323                                                  dev_config);
3324                 if (!list[i].eth_dev) {
3325                         if (rte_errno != EBUSY && rte_errno != EEXIST)
3326                                 break;
3327                         /* Device is disabled or already spawned. Ignore it. */
3328                         continue;
3329                 }
3330                 restore = list[i].eth_dev->data->dev_flags;
3331                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3332                 /* Restore non-PCI flags cleared by the above call. */
3333                 list[i].eth_dev->data->dev_flags |= restore;
3334                 mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3335                 rte_eth_dev_probing_finish(list[i].eth_dev);
3336         }
3337         if (i != ns) {
3338                 DRV_LOG(ERR,
3339                         "probe of PCI device " PCI_PRI_FMT " aborted after"
3340                         " encountering an error: %s",
3341                         pci_dev->addr.domain, pci_dev->addr.bus,
3342                         pci_dev->addr.devid, pci_dev->addr.function,
3343                         strerror(rte_errno));
3344                 ret = -rte_errno;
3345                 /* Roll back. */
3346                 while (i--) {
3347                         if (!list[i].eth_dev)
3348                                 continue;
3349                         mlx5_dev_close(list[i].eth_dev);
3350                         /* mac_addrs must not be freed because in dev_private */
3351                         list[i].eth_dev->data->mac_addrs = NULL;
3352                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3353                 }
3354                 /* Restore original error. */
3355                 rte_errno = -ret;
3356         } else {
3357                 ret = 0;
3358         }
3359 exit:
3360         /*
3361          * Do the routine cleanup:
3362          * - close opened Netlink sockets
3363          * - free allocated spawn data array
3364          * - free the Infiniband device list
3365          */
3366         if (nl_rdma >= 0)
3367                 close(nl_rdma);
3368         if (nl_route >= 0)
3369                 close(nl_route);
3370         if (list)
3371                 rte_free(list);
3372         assert(ibv_list);
3373         mlx5_glue->free_device_list(ibv_list);
3374         return ret;
3375 }
3376
3377 /**
3378  * Look for the ethernet device belonging to mlx5 driver.
3379  *
3380  * @param[in] port_id
3381  *   port_id to start looking for device.
3382  * @param[in] pci_dev
3383  *   Pointer to the hint PCI device. When device is being probed
3384  *   the its siblings (master and preceding representors might
3385  *   not have assigned driver yet (because the mlx5_pci_probe()
3386  *   is not completed yet, for this case match on hint PCI
3387  *   device may be used to detect sibling device.
3388  *
3389  * @return
3390  *   port_id of found device, RTE_MAX_ETHPORT if not found.
3391  */
3392 uint16_t
3393 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3394 {
3395         while (port_id < RTE_MAX_ETHPORTS) {
3396                 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3397
3398                 if (dev->state != RTE_ETH_DEV_UNUSED &&
3399                     dev->device &&
3400                     (dev->device == &pci_dev->device ||
3401                      (dev->device->driver &&
3402                      dev->device->driver->name &&
3403                      !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3404                         break;
3405                 port_id++;
3406         }
3407         if (port_id >= RTE_MAX_ETHPORTS)
3408                 return RTE_MAX_ETHPORTS;
3409         return port_id;
3410 }
3411
3412 /**
3413  * DPDK callback to remove a PCI device.
3414  *
3415  * This function removes all Ethernet devices belong to a given PCI device.
3416  *
3417  * @param[in] pci_dev
3418  *   Pointer to the PCI device.
3419  *
3420  * @return
3421  *   0 on success, the function cannot fail.
3422  */
3423 static int
3424 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3425 {
3426         uint16_t port_id;
3427
3428         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3429                 rte_eth_dev_close(port_id);
3430         return 0;
3431 }
3432
3433 static const struct rte_pci_id mlx5_pci_id_map[] = {
3434         {
3435                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3436                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3437         },
3438         {
3439                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3440                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3441         },
3442         {
3443                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3444                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3445         },
3446         {
3447                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3448                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3449         },
3450         {
3451                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3452                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3453         },
3454         {
3455                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3456                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3457         },
3458         {
3459                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3460                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3461         },
3462         {
3463                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3464                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3465         },
3466         {
3467                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3468                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3469         },
3470         {
3471                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3472                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3473         },
3474         {
3475                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3476                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3477         },
3478         {
3479                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3480                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3481         },
3482         {
3483                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3484                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3485         },
3486         {
3487                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3488                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3489         },
3490         {
3491                 .vendor_id = 0
3492         }
3493 };
3494
3495 static struct rte_pci_driver mlx5_driver = {
3496         .driver = {
3497                 .name = MLX5_DRIVER_NAME
3498         },
3499         .id_table = mlx5_pci_id_map,
3500         .probe = mlx5_pci_probe,
3501         .remove = mlx5_pci_remove,
3502         .dma_map = mlx5_dma_map,
3503         .dma_unmap = mlx5_dma_unmap,
3504         .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3505                      RTE_PCI_DRV_PROBE_AGAIN,
3506 };
3507
3508 /**
3509  * Driver initialization routine.
3510  */
3511 RTE_INIT(rte_mlx5_pmd_init)
3512 {
3513         /* Initialize driver log type. */
3514         mlx5_logtype = rte_log_register("pmd.net.mlx5");
3515         if (mlx5_logtype >= 0)
3516                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3517
3518         /* Build the static tables for Verbs conversion. */
3519         mlx5_set_ptype_table();
3520         mlx5_set_cksum_table();
3521         mlx5_set_swp_types_table();
3522         if (mlx5_glue)
3523                 rte_pci_register(&mlx5_driver);
3524 }
3525
3526 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3527 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3528 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");