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 <dlfcn.h>
  11 #include <stdint.h>
  12 #include <stdlib.h>
  13 #include <errno.h>
  14 #include <net/if.h>
  15 #include <sys/mman.h>
  16 #include <linux/rtnetlink.h>
  17
  18 /* Verbs header. */
  19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  20 #ifdef PEDANTIC
  21 #pragma GCC diagnostic ignored "-Wpedantic"
  22 #endif
  23 #include <infiniband/verbs.h>
  24 #ifdef PEDANTIC
  25 #pragma GCC diagnostic error "-Wpedantic"
  26 #endif
  27
  28 #include <rte_malloc.h>
  29 #include <rte_ethdev_driver.h>
  30 #include <rte_ethdev_pci.h>
  31 #include <rte_pci.h>
  32 #include <rte_bus_pci.h>
  33 #include <rte_common.h>
  34 #include <rte_config.h>
  35 #include <rte_kvargs.h>
  36 #include <rte_rwlock.h>
  37 #include <rte_spinlock.h>
  38 #include <rte_string_fns.h>
  39 #include <rte_alarm.h>
  40
  41 #include "mlx5.h"
  42 #include "mlx5_utils.h"
  43 #include "mlx5_rxtx.h"
  44 #include "mlx5_autoconf.h"
  45 #include "mlx5_defs.h"
  46 #include "mlx5_glue.h"
  47 #include "mlx5_mr.h"
  48 #include "mlx5_flow.h"
  49
  50 /* Device parameter to enable RX completion queue compression. */
  51 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
  52
  53 /* Device parameter to enable RX completion entry padding to 128B. */
  54 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
  55
  56 /* Device parameter to enable padding Rx packet to cacheline size. */
  57 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
  58
  59 /* Device parameter to enable Multi-Packet Rx queue. */
  60 #define MLX5_RX_MPRQ_EN "mprq_en"
  61
  62 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
  63 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
  64
  65 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
  66 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
  67
  68 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
  69 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
  70
  71 /* Device parameter to configure inline send. Deprecated, ignored.*/
  72 #define MLX5_TXQ_INLINE "txq_inline"
  73
  74 /* Device parameter to limit packet size to inline with ordinary SEND. */
  75 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
  76
  77 /* Device parameter to configure minimal data size to inline. */
  78 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
  79
  80 /* Device parameter to limit packet size to inline with Enhanced MPW. */
  81 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
  82
  83 /*
  84  * Device parameter to configure the number of TX queues threshold for
  85  * enabling inline send.
  86  */
  87 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
  88
  89 /*
  90  * Device parameter to configure the number of TX queues threshold for
  91  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
  92  */
  93 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
  94
  95 /* Device parameter to enable multi-packet send WQEs. */
  96 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
  97
  98 /*
  99  * Device parameter to include 2 dsegs in the title WQEBB.
 100  * Deprecated, ignored.
 101  */
 102 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
 103
 104 /*
 105  * Device parameter to limit the size of inlining packet.
 106  * Deprecated, ignored.
 107  */
 108 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
 109
 110 /*
 111  * Device parameter to enable hardware Tx vector.
 112  * Deprecated, ignored (no vectorized Tx routines anymore).
 113  */
 114 #define MLX5_TX_VEC_EN "tx_vec_en"
 115
 116 /* Device parameter to enable hardware Rx vector. */
 117 #define MLX5_RX_VEC_EN "rx_vec_en"
 118
 119 /* Allow L3 VXLAN flow creation. */
 120 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
 121
 122 /* Activate DV E-Switch flow steering. */
 123 #define MLX5_DV_ESW_EN "dv_esw_en"
 124
 125 /* Activate DV flow steering. */
 126 #define MLX5_DV_FLOW_EN "dv_flow_en"
 127
 128 /* Activate Netlink support in VF mode. */
 129 #define MLX5_VF_NL_EN "vf_nl_en"
 130
 131 /* Enable extending memsegs when creating a MR. */
 132 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
 133
 134 /* Select port representors to instantiate. */
 135 #define MLX5_REPRESENTOR "representor"
 136
 137 /* Device parameter to configure the maximum number of dump files per queue. */
 138 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
 139
 140 /* Configure timeout of LRO session (in microseconds). */
 141 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
 142
 143 #ifndef HAVE_IBV_MLX5_MOD_MPW
 144 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
 145 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
 146 #endif
 147
 148 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
 149 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
 150 #endif
 151
 152 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
 153
 154 /* Shared memory between primary and secondary processes. */
 155 struct mlx5_shared_data *mlx5_shared_data;
 156
 157 /* Spinlock for mlx5_shared_data allocation. */
 158 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
 159
 160 /* Process local data for secondary processes. */
 161 static struct mlx5_local_data mlx5_local_data;
 162
 163 /** Driver-specific log messages type. */
 164 int mlx5_logtype;
 165
 166 /** Data associated with devices to spawn. */
 167 struct mlx5_dev_spawn_data {
 168         uint32_t ifindex; /**< Network interface index. */
 169         uint32_t max_port; /**< IB device maximal port index. */
 170         uint32_t ibv_port; /**< IB device physical port index. */
 171         int pf_bond; /**< bonding device PF index. < 0 - no bonding */
 172         struct mlx5_switch_info info; /**< Switch information. */
 173         struct ibv_device *ibv_dev; /**< Associated IB device. */
 174         struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
 175         struct rte_pci_device *pci_dev; /**< Backend PCI device. */
 176 };
 177
 178 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
 179 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
 180
 181 /**
 182  * Initialize the counters management structure.
 183  *
 184  * @param[in] sh
 185  *   Pointer to mlx5_ibv_shared object to free
 186  */
 187 static void
 188 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
 189 {
 190         uint8_t i;
 191
 192         TAILQ_INIT(&sh->cmng.flow_counters);
 193         for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
 194                 TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
 195 }
 196
 197 /**
 198  * Destroy all the resources allocated for a counter memory management.
 199  *
 200  * @param[in] mng
 201  *   Pointer to the memory management structure.
 202  */
 203 static void
 204 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
 205 {
 206         uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
 207
 208         LIST_REMOVE(mng, next);
 209         claim_zero(mlx5_devx_cmd_destroy(mng->dm));
 210         claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
 211         rte_free(mem);
 212 }
 213
 214 /**
 215  * Close and release all the resources of the counters management.
 216  *
 217  * @param[in] sh
 218  *   Pointer to mlx5_ibv_shared object to free.
 219  */
 220 static void
 221 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
 222 {
 223         struct mlx5_counter_stats_mem_mng *mng;
 224         uint8_t i;
 225         int j;
 226         int retries = 1024;
 227
 228         rte_errno = 0;
 229         while (--retries) {
 230                 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
 231                 if (rte_errno != EINPROGRESS)
 232                         break;
 233                 rte_pause();
 234         }
 235         for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
 236                 struct mlx5_flow_counter_pool *pool;
 237                 uint32_t batch = !!(i % 2);
 238
 239                 if (!sh->cmng.ccont[i].pools)
 240                         continue;
 241                 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
 242                 while (pool) {
 243                         if (batch) {
 244                                 if (pool->min_dcs)
 245                                         claim_zero
 246                                         (mlx5_devx_cmd_destroy(pool->min_dcs));
 247                         }
 248                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
 249                                 if (pool->counters_raw[j].action)
 250                                         claim_zero
 251                                         (mlx5_glue->destroy_flow_action
 252                                                (pool->counters_raw[j].action));
 253                                 if (!batch && pool->counters_raw[j].dcs)
 254                                         claim_zero(mlx5_devx_cmd_destroy
 255                                                   (pool->counters_raw[j].dcs));
 256                         }
 257                         TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
 258                                      next);
 259                         rte_free(pool);
 260                         pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
 261                 }
 262                 rte_free(sh->cmng.ccont[i].pools);
 263         }
 264         mng = LIST_FIRST(&sh->cmng.mem_mngs);
 265         while (mng) {
 266                 mlx5_flow_destroy_counter_stat_mem_mng(mng);
 267                 mng = LIST_FIRST(&sh->cmng.mem_mngs);
 268         }
 269         memset(&sh->cmng, 0, sizeof(sh->cmng));
 270 }
 271
 272 /**
 273  * Extract pdn of PD object using DV API.
 274  *
 275  * @param[in] pd
 276  *   Pointer to the verbs PD object.
 277  * @param[out] pdn
 278  *   Pointer to the PD object number variable.
 279  *
 280  * @return
 281  *   0 on success, error value otherwise.
 282  */
 283 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 284 static int
 285 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
 286 {
 287         struct mlx5dv_obj obj;
 288         struct mlx5dv_pd pd_info;
 289         int ret = 0;
 290
 291         obj.pd.in = pd;
 292         obj.pd.out = &pd_info;
 293         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
 294         if (ret) {
 295                 DRV_LOG(DEBUG, "Fail to get PD object info");
 296                 return ret;
 297         }
 298         *pdn = pd_info.pdn;
 299         return 0;
 300 }
 301 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 302
 303 /**
 304  * Allocate shared IB device context. If there is multiport device the
 305  * master and representors will share this context, if there is single
 306  * port dedicated IB device, the context will be used by only given
 307  * port due to unification.
 308  *
 309  * Routine first searches the context for the specified IB device name,
 310  * if found the shared context assumed and reference counter is incremented.
 311  * If no context found the new one is created and initialized with specified
 312  * IB device context and parameters.
 313  *
 314  * @param[in] spawn
 315  *   Pointer to the IB device attributes (name, port, etc).
 316  *
 317  * @return
 318  *   Pointer to mlx5_ibv_shared object on success,
 319  *   otherwise NULL and rte_errno is set.
 320  */
 321 static struct mlx5_ibv_shared *
 322 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn)
 323 {
 324         struct mlx5_ibv_shared *sh;
 325         int err = 0;
 326         uint32_t i;
 327
 328         assert(spawn);
 329         /* Secondary process should not create the shared context. */
 330         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 331         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 332         /* Search for IB context by device name. */
 333         LIST_FOREACH(sh, &mlx5_ibv_list, next) {
 334                 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
 335                         sh->refcnt++;
 336                         goto exit;
 337                 }
 338         }
 339         /* No device found, we have to create new shared context. */
 340         assert(spawn->max_port);
 341         sh = rte_zmalloc("ethdev shared ib context",
 342                          sizeof(struct mlx5_ibv_shared) +
 343                          spawn->max_port *
 344                          sizeof(struct mlx5_ibv_shared_port),
 345                          RTE_CACHE_LINE_SIZE);
 346         if (!sh) {
 347                 DRV_LOG(ERR, "shared context allocation failure");
 348                 rte_errno  = ENOMEM;
 349                 goto exit;
 350         }
 351         /* Try to open IB device with DV first, then usual Verbs. */
 352         errno = 0;
 353         sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
 354         if (sh->ctx) {
 355                 sh->devx = 1;
 356                 DRV_LOG(DEBUG, "DevX is supported");
 357         } else {
 358                 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
 359                 if (!sh->ctx) {
 360                         err = errno ? errno : ENODEV;
 361                         goto error;
 362                 }
 363                 DRV_LOG(DEBUG, "DevX is NOT supported");
 364         }
 365         err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
 366         if (err) {
 367                 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
 368                 goto error;
 369         }
 370         sh->refcnt = 1;
 371         sh->max_port = spawn->max_port;
 372         strncpy(sh->ibdev_name, sh->ctx->device->name,
 373                 sizeof(sh->ibdev_name));
 374         strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
 375                 sizeof(sh->ibdev_path));
 376         pthread_mutex_init(&sh->intr_mutex, NULL);
 377         /*
 378          * Setting port_id to max unallowed value means
 379          * there is no interrupt subhandler installed for
 380          * the given port index i.
 381          */
 382         for (i = 0; i < sh->max_port; i++) {
 383                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
 384                 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
 385         }
 386         sh->pd = mlx5_glue->alloc_pd(sh->ctx);
 387         if (sh->pd == NULL) {
 388                 DRV_LOG(ERR, "PD allocation failure");
 389                 err = ENOMEM;
 390                 goto error;
 391         }
 392 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 393         err = mlx5_get_pdn(sh->pd, &sh->pdn);
 394         if (err) {
 395                 DRV_LOG(ERR, "Fail to extract pdn from PD");
 396                 goto error;
 397         }
 398 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 399         /*
 400          * Once the device is added to the list of memory event
 401          * callback, its global MR cache table cannot be expanded
 402          * on the fly because of deadlock. If it overflows, lookup
 403          * should be done by searching MR list linearly, which is slow.
 404          *
 405          * At this point the device is not added to the memory
 406          * event list yet, context is just being created.
 407          */
 408         err = mlx5_mr_btree_init(&sh->mr.cache,
 409                                  MLX5_MR_BTREE_CACHE_N * 2,
 410                                  spawn->pci_dev->device.numa_node);
 411         if (err) {
 412                 err = rte_errno;
 413                 goto error;
 414         }
 415         mlx5_flow_counters_mng_init(sh);
 416         /* Add device to memory callback list. */
 417         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
 418         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
 419                          sh, mem_event_cb);
 420         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
 421         /* Add context to the global device list. */
 422         LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
 423 exit:
 424         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 425         return sh;
 426 error:
 427         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 428         assert(sh);
 429         if (sh->pd)
 430                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
 431         if (sh->ctx)
 432                 claim_zero(mlx5_glue->close_device(sh->ctx));
 433         rte_free(sh);
 434         assert(err > 0);
 435         rte_errno = err;
 436         return NULL;
 437 }
 438
 439 /**
 440  * Free shared IB device context. Decrement counter and if zero free
 441  * all allocated resources and close handles.
 442  *
 443  * @param[in] sh
 444  *   Pointer to mlx5_ibv_shared object to free
 445  */
 446 static void
 447 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
 448 {
 449         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 450 #ifndef NDEBUG
 451         /* Check the object presence in the list. */
 452         struct mlx5_ibv_shared *lctx;
 453
 454         LIST_FOREACH(lctx, &mlx5_ibv_list, next)
 455                 if (lctx == sh)
 456                         break;
 457         assert(lctx);
 458         if (lctx != sh) {
 459                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
 460                 goto exit;
 461         }
 462 #endif
 463         assert(sh);
 464         assert(sh->refcnt);
 465         /* Secondary process should not free the shared context. */
 466         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 467         if (--sh->refcnt)
 468                 goto exit;
 469         /* Release created Memory Regions. */
 470         mlx5_mr_release(sh);
 471         /* Remove from memory callback device list. */
 472         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
 473         LIST_REMOVE(sh, mem_event_cb);
 474         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
 475         /* Remove context from the global device list. */
 476         LIST_REMOVE(sh, next);
 477         /*
 478          *  Ensure there is no async event handler installed.
 479          *  Only primary process handles async device events.
 480          **/
 481         mlx5_flow_counters_mng_close(sh);
 482         assert(!sh->intr_cnt);
 483         if (sh->intr_cnt)
 484                 mlx5_intr_callback_unregister
 485                         (&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
 486 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
 487         if (sh->devx_intr_cnt) {
 488                 if (sh->intr_handle_devx.fd)
 489                         rte_intr_callback_unregister(&sh->intr_handle_devx,
 490                                           mlx5_dev_interrupt_handler_devx, sh);
 491                 if (sh->devx_comp)
 492                         mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
 493         }
 494 #endif
 495         pthread_mutex_destroy(&sh->intr_mutex);
 496         if (sh->pd)
 497                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
 498         if (sh->ctx)
 499                 claim_zero(mlx5_glue->close_device(sh->ctx));
 500         rte_free(sh);
 501 exit:
 502         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 503 }
 504
 505 /**
 506  * Initialize DR related data within private structure.
 507  * Routine checks the reference counter and does actual
 508  * resources creation/initialization only if counter is zero.
 509  *
 510  * @param[in] priv
 511  *   Pointer to the private device data structure.
 512  *
 513  * @return
 514  *   Zero on success, positive error code otherwise.
 515  */
 516 static int
 517 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 518 {
 519 #ifdef HAVE_MLX5DV_DR
 520         struct mlx5_ibv_shared *sh = priv->sh;
 521         int err = 0;
 522         void *domain;
 523
 524         assert(sh);
 525         if (sh->dv_refcnt) {
 526                 /* Shared DV/DR structures is already initialized. */
 527                 sh->dv_refcnt++;
 528                 priv->dr_shared = 1;
 529                 return 0;
 530         }
 531         /* Reference counter is zero, we should initialize structures. */
 532         domain = mlx5_glue->dr_create_domain(sh->ctx,
 533                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 534         if (!domain) {
 535                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 536                 err = errno;
 537                 goto error;
 538         }
 539         sh->rx_domain = domain;
 540         domain = mlx5_glue->dr_create_domain(sh->ctx,
 541                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 542         if (!domain) {
 543                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 544                 err = errno;
 545                 goto error;
 546         }
 547         pthread_mutex_init(&sh->dv_mutex, NULL);
 548         sh->tx_domain = domain;
 549 #ifdef HAVE_MLX5DV_DR_ESWITCH
 550         if (priv->config.dv_esw_en) {
 551                 domain  = mlx5_glue->dr_create_domain
 552                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 553                 if (!domain) {
 554                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 555                         err = errno;
 556                         goto error;
 557                 }
 558                 sh->fdb_domain = domain;
 559                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 560         }
 561 #endif
 562         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 563         sh->dv_refcnt++;
 564         priv->dr_shared = 1;
 565         return 0;
 566
 567 error:
 568        /* Rollback the created objects. */
 569         if (sh->rx_domain) {
 570                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 571                 sh->rx_domain = NULL;
 572         }
 573         if (sh->tx_domain) {
 574                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 575                 sh->tx_domain = NULL;
 576         }
 577         if (sh->fdb_domain) {
 578                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 579                 sh->fdb_domain = NULL;
 580         }
 581         if (sh->esw_drop_action) {
 582                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 583                 sh->esw_drop_action = NULL;
 584         }
 585         if (sh->pop_vlan_action) {
 586                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 587                 sh->pop_vlan_action = NULL;
 588         }
 589         return err;
 590 #else
 591         (void)priv;
 592         return 0;
 593 #endif
 594 }
 595
 596 /**
 597  * Destroy DR related data within private structure.
 598  *
 599  * @param[in] priv
 600  *   Pointer to the private device data structure.
 601  */
 602 static void
 603 mlx5_free_shared_dr(struct mlx5_priv *priv)
 604 {
 605 #ifdef HAVE_MLX5DV_DR
 606         struct mlx5_ibv_shared *sh;
 607
 608         if (!priv->dr_shared)
 609                 return;
 610         priv->dr_shared = 0;
 611         sh = priv->sh;
 612         assert(sh);
 613         assert(sh->dv_refcnt);
 614         if (sh->dv_refcnt && --sh->dv_refcnt)
 615                 return;
 616         if (sh->rx_domain) {
 617                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 618                 sh->rx_domain = NULL;
 619         }
 620         if (sh->tx_domain) {
 621                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 622                 sh->tx_domain = NULL;
 623         }
 624 #ifdef HAVE_MLX5DV_DR_ESWITCH
 625         if (sh->fdb_domain) {
 626                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 627                 sh->fdb_domain = NULL;
 628         }
 629         if (sh->esw_drop_action) {
 630                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 631                 sh->esw_drop_action = NULL;
 632         }
 633 #endif
 634         if (sh->pop_vlan_action) {
 635                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 636                 sh->pop_vlan_action = NULL;
 637         }
 638         pthread_mutex_destroy(&sh->dv_mutex);
 639 #else
 640         (void)priv;
 641 #endif
 642 }
 643
 644 /**
 645  * Initialize shared data between primary and secondary process.
 646  *
 647  * A memzone is reserved by primary process and secondary processes attach to
 648  * the memzone.
 649  *
 650  * @return
 651  *   0 on success, a negative errno value otherwise and rte_errno is set.
 652  */
 653 static int
 654 mlx5_init_shared_data(void)
 655 {
 656         const struct rte_memzone *mz;
 657         int ret = 0;
 658
 659         rte_spinlock_lock(&mlx5_shared_data_lock);
 660         if (mlx5_shared_data == NULL) {
 661                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 662                         /* Allocate shared memory. */
 663                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
 664                                                  sizeof(*mlx5_shared_data),
 665                                                  SOCKET_ID_ANY, 0);
 666                         if (mz == NULL) {
 667                                 DRV_LOG(ERR,
 668                                         "Cannot allocate mlx5 shared data\n");
 669                                 ret = -rte_errno;
 670                                 goto error;
 671                         }
 672                         mlx5_shared_data = mz->addr;
 673                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
 674                         rte_spinlock_init(&mlx5_shared_data->lock);
 675                 } else {
 676                         /* Lookup allocated shared memory. */
 677                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
 678                         if (mz == NULL) {
 679                                 DRV_LOG(ERR,
 680                                         "Cannot attach mlx5 shared data\n");
 681                                 ret = -rte_errno;
 682                                 goto error;
 683                         }
 684                         mlx5_shared_data = mz->addr;
 685                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
 686                 }
 687         }
 688 error:
 689         rte_spinlock_unlock(&mlx5_shared_data_lock);
 690         return ret;
 691 }
 692
 693 /**
 694  * Retrieve integer value from environment variable.
 695  *
 696  * @param[in] name
 697  *   Environment variable name.
 698  *
 699  * @return
 700  *   Integer value, 0 if the variable is not set.
 701  */
 702 int
 703 mlx5_getenv_int(const char *name)
 704 {
 705         const char *val = getenv(name);
 706
 707         if (val == NULL)
 708                 return 0;
 709         return atoi(val);
 710 }
 711
 712 /**
 713  * Verbs callback to allocate a memory. This function should allocate the space
 714  * according to the size provided residing inside a huge page.
 715  * Please note that all allocation must respect the alignment from libmlx5
 716  * (i.e. currently sysconf(_SC_PAGESIZE)).
 717  *
 718  * @param[in] size
 719  *   The size in bytes of the memory to allocate.
 720  * @param[in] data
 721  *   A pointer to the callback data.
 722  *
 723  * @return
 724  *   Allocated buffer, NULL otherwise and rte_errno is set.
 725  */
 726 static void *
 727 mlx5_alloc_verbs_buf(size_t size, void *data)
 728 {
 729         struct mlx5_priv *priv = data;
 730         void *ret;
 731         size_t alignment = sysconf(_SC_PAGESIZE);
 732         unsigned int socket = SOCKET_ID_ANY;
 733
 734         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 735                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 736
 737                 socket = ctrl->socket;
 738         } else if (priv->verbs_alloc_ctx.type ==
 739                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 740                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 741
 742                 socket = ctrl->socket;
 743         }
 744         assert(data != NULL);
 745         ret = rte_malloc_socket(__func__, size, alignment, socket);
 746         if (!ret && size)
 747                 rte_errno = ENOMEM;
 748         return ret;
 749 }
 750
 751 /**
 752  * Verbs callback to free a memory.
 753  *
 754  * @param[in] ptr
 755  *   A pointer to the memory to free.
 756  * @param[in] data
 757  *   A pointer to the callback data.
 758  */
 759 static void
 760 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 761 {
 762         assert(data != NULL);
 763         rte_free(ptr);
 764 }
 765
 766 /**
 767  * DPDK callback to add udp tunnel port
 768  *
 769  * @param[in] dev
 770  *   A pointer to eth_dev
 771  * @param[in] udp_tunnel
 772  *   A pointer to udp tunnel
 773  *
 774  * @return
 775  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
 776  */
 777 int
 778 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
 779                          struct rte_eth_udp_tunnel *udp_tunnel)
 780 {
 781         assert(udp_tunnel != NULL);
 782         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
 783             udp_tunnel->udp_port == 4789)
 784                 return 0;
 785         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
 786             udp_tunnel->udp_port == 4790)
 787                 return 0;
 788         return -ENOTSUP;
 789 }
 790
 791 /**
 792  * Initialize process private data structure.
 793  *
 794  * @param dev
 795  *   Pointer to Ethernet device structure.
 796  *
 797  * @return
 798  *   0 on success, a negative errno value otherwise and rte_errno is set.
 799  */
 800 int
 801 mlx5_proc_priv_init(struct rte_eth_dev *dev)
 802 {
 803         struct mlx5_priv *priv = dev->data->dev_private;
 804         struct mlx5_proc_priv *ppriv;
 805         size_t ppriv_size;
 806
 807         /*
 808          * UAR register table follows the process private structure. BlueFlame
 809          * registers for Tx queues are stored in the table.
 810          */
 811         ppriv_size =
 812                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
 813         ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
 814                                   RTE_CACHE_LINE_SIZE, dev->device->numa_node);
 815         if (!ppriv) {
 816                 rte_errno = ENOMEM;
 817                 return -rte_errno;
 818         }
 819         ppriv->uar_table_sz = ppriv_size;
 820         dev->process_private = ppriv;
 821         return 0;
 822 }
 823
 824 /**
 825  * Un-initialize process private data structure.
 826  *
 827  * @param dev
 828  *   Pointer to Ethernet device structure.
 829  */
 830 static void
 831 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
 832 {
 833         if (!dev->process_private)
 834                 return;
 835         rte_free(dev->process_private);
 836         dev->process_private = NULL;
 837 }
 838
 839 /**
 840  * DPDK callback to close the device.
 841  *
 842  * Destroy all queues and objects, free memory.
 843  *
 844  * @param dev
 845  *   Pointer to Ethernet device structure.
 846  */
 847 static void
 848 mlx5_dev_close(struct rte_eth_dev *dev)
 849 {
 850         struct mlx5_priv *priv = dev->data->dev_private;
 851         unsigned int i;
 852         int ret;
 853
 854         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
 855                 dev->data->port_id,
 856                 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
 857         /* In case mlx5_dev_stop() has not been called. */
 858         mlx5_dev_interrupt_handler_uninstall(dev);
 859         mlx5_dev_interrupt_handler_devx_uninstall(dev);
 860         mlx5_traffic_disable(dev);
 861         mlx5_flow_flush(dev, NULL);
 862         /* Prevent crashes when queues are still in use. */
 863         dev->rx_pkt_burst = removed_rx_burst;
 864         dev->tx_pkt_burst = removed_tx_burst;
 865         rte_wmb();
 866         /* Disable datapath on secondary process. */
 867         mlx5_mp_req_stop_rxtx(dev);
 868         if (priv->rxqs != NULL) {
 869                 /* XXX race condition if mlx5_rx_burst() is still running. */
 870                 usleep(1000);
 871                 for (i = 0; (i != priv->rxqs_n); ++i)
 872                         mlx5_rxq_release(dev, i);
 873                 priv->rxqs_n = 0;
 874                 priv->rxqs = NULL;
 875         }
 876         if (priv->txqs != NULL) {
 877                 /* XXX race condition if mlx5_tx_burst() is still running. */
 878                 usleep(1000);
 879                 for (i = 0; (i != priv->txqs_n); ++i)
 880                         mlx5_txq_release(dev, i);
 881                 priv->txqs_n = 0;
 882                 priv->txqs = NULL;
 883         }
 884         mlx5_proc_priv_uninit(dev);
 885         mlx5_mprq_free_mp(dev);
 886         mlx5_free_shared_dr(priv);
 887         if (priv->rss_conf.rss_key != NULL)
 888                 rte_free(priv->rss_conf.rss_key);
 889         if (priv->reta_idx != NULL)
 890                 rte_free(priv->reta_idx);
 891         if (priv->config.vf)
 892                 mlx5_nl_mac_addr_flush(dev);
 893         if (priv->nl_socket_route >= 0)
 894                 close(priv->nl_socket_route);
 895         if (priv->nl_socket_rdma >= 0)
 896                 close(priv->nl_socket_rdma);
 897         if (priv->vmwa_context)
 898                 mlx5_vlan_vmwa_exit(priv->vmwa_context);
 899         if (priv->sh) {
 900                 /*
 901                  * Free the shared context in last turn, because the cleanup
 902                  * routines above may use some shared fields, like
 903                  * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
 904                  * ifindex if Netlink fails.
 905                  */
 906                 mlx5_free_shared_ibctx(priv->sh);
 907                 priv->sh = NULL;
 908         }
 909         ret = mlx5_hrxq_verify(dev);
 910         if (ret)
 911                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
 912                         dev->data->port_id);
 913         ret = mlx5_ind_table_obj_verify(dev);
 914         if (ret)
 915                 DRV_LOG(WARNING, "port %u some indirection table still remain",
 916                         dev->data->port_id);
 917         ret = mlx5_rxq_obj_verify(dev);
 918         if (ret)
 919                 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
 920                         dev->data->port_id);
 921         ret = mlx5_rxq_verify(dev);
 922         if (ret)
 923                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
 924                         dev->data->port_id);
 925         ret = mlx5_txq_ibv_verify(dev);
 926         if (ret)
 927                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
 928                         dev->data->port_id);
 929         ret = mlx5_txq_verify(dev);
 930         if (ret)
 931                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
 932                         dev->data->port_id);
 933         ret = mlx5_flow_verify(dev);
 934         if (ret)
 935                 DRV_LOG(WARNING, "port %u some flows still remain",
 936                         dev->data->port_id);
 937         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 938                 unsigned int c = 0;
 939                 uint16_t port_id;
 940
 941                 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 942                         struct mlx5_priv *opriv =
 943                                 rte_eth_devices[port_id].data->dev_private;
 944
 945                         if (!opriv ||
 946                             opriv->domain_id != priv->domain_id ||
 947                             &rte_eth_devices[port_id] == dev)
 948                                 continue;
 949                         ++c;
 950                         break;
 951                 }
 952                 if (!c)
 953                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
 954         }
 955         memset(priv, 0, sizeof(*priv));
 956         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 957         /*
 958          * Reset mac_addrs to NULL such that it is not freed as part of
 959          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
 960          * it is freed when dev_private is freed.
 961          */
 962         dev->data->mac_addrs = NULL;
 963 }
 964
 965 const struct eth_dev_ops mlx5_dev_ops = {
 966         .dev_configure = mlx5_dev_configure,
 967         .dev_start = mlx5_dev_start,
 968         .dev_stop = mlx5_dev_stop,
 969         .dev_set_link_down = mlx5_set_link_down,
 970         .dev_set_link_up = mlx5_set_link_up,
 971         .dev_close = mlx5_dev_close,
 972         .promiscuous_enable = mlx5_promiscuous_enable,
 973         .promiscuous_disable = mlx5_promiscuous_disable,
 974         .allmulticast_enable = mlx5_allmulticast_enable,
 975         .allmulticast_disable = mlx5_allmulticast_disable,
 976         .link_update = mlx5_link_update,
 977         .stats_get = mlx5_stats_get,
 978         .stats_reset = mlx5_stats_reset,
 979         .xstats_get = mlx5_xstats_get,
 980         .xstats_reset = mlx5_xstats_reset,
 981         .xstats_get_names = mlx5_xstats_get_names,
 982         .fw_version_get = mlx5_fw_version_get,
 983         .dev_infos_get = mlx5_dev_infos_get,
 984         .read_clock = mlx5_read_clock,
 985         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 986         .vlan_filter_set = mlx5_vlan_filter_set,
 987         .rx_queue_setup = mlx5_rx_queue_setup,
 988         .tx_queue_setup = mlx5_tx_queue_setup,
 989         .rx_queue_release = mlx5_rx_queue_release,
 990         .tx_queue_release = mlx5_tx_queue_release,
 991         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 992         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 993         .mac_addr_remove = mlx5_mac_addr_remove,
 994         .mac_addr_add = mlx5_mac_addr_add,
 995         .mac_addr_set = mlx5_mac_addr_set,
 996         .set_mc_addr_list = mlx5_set_mc_addr_list,
 997         .mtu_set = mlx5_dev_set_mtu,
 998         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 999         .vlan_offload_set = mlx5_vlan_offload_set,
1000         .reta_update = mlx5_dev_rss_reta_update,
1001         .reta_query = mlx5_dev_rss_reta_query,
1002         .rss_hash_update = mlx5_rss_hash_update,
1003         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
1004         .filter_ctrl = mlx5_dev_filter_ctrl,
1005         .rx_descriptor_status = mlx5_rx_descriptor_status,
1006         .tx_descriptor_status = mlx5_tx_descriptor_status,
1007         .rx_queue_count = mlx5_rx_queue_count,
1008         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1009         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1010         .is_removed = mlx5_is_removed,
1011         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1012         .get_module_info = mlx5_get_module_info,
1013         .get_module_eeprom = mlx5_get_module_eeprom,
1014 };
1015
1016 /* Available operations from secondary process. */
1017 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1018         .stats_get = mlx5_stats_get,
1019         .stats_reset = mlx5_stats_reset,
1020         .xstats_get = mlx5_xstats_get,
1021         .xstats_reset = mlx5_xstats_reset,
1022         .xstats_get_names = mlx5_xstats_get_names,
1023         .fw_version_get = mlx5_fw_version_get,
1024         .dev_infos_get = mlx5_dev_infos_get,
1025         .rx_descriptor_status = mlx5_rx_descriptor_status,
1026         .tx_descriptor_status = mlx5_tx_descriptor_status,
1027         .get_module_info = mlx5_get_module_info,
1028         .get_module_eeprom = mlx5_get_module_eeprom,
1029 };
1030
1031 /* Available operations in flow isolated mode. */
1032 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1033         .dev_configure = mlx5_dev_configure,
1034         .dev_start = mlx5_dev_start,
1035         .dev_stop = mlx5_dev_stop,
1036         .dev_set_link_down = mlx5_set_link_down,
1037         .dev_set_link_up = mlx5_set_link_up,
1038         .dev_close = mlx5_dev_close,
1039         .promiscuous_enable = mlx5_promiscuous_enable,
1040         .promiscuous_disable = mlx5_promiscuous_disable,
1041         .allmulticast_enable = mlx5_allmulticast_enable,
1042         .allmulticast_disable = mlx5_allmulticast_disable,
1043         .link_update = mlx5_link_update,
1044         .stats_get = mlx5_stats_get,
1045         .stats_reset = mlx5_stats_reset,
1046         .xstats_get = mlx5_xstats_get,
1047         .xstats_reset = mlx5_xstats_reset,
1048         .xstats_get_names = mlx5_xstats_get_names,
1049         .fw_version_get = mlx5_fw_version_get,
1050         .dev_infos_get = mlx5_dev_infos_get,
1051         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1052         .vlan_filter_set = mlx5_vlan_filter_set,
1053         .rx_queue_setup = mlx5_rx_queue_setup,
1054         .tx_queue_setup = mlx5_tx_queue_setup,
1055         .rx_queue_release = mlx5_rx_queue_release,
1056         .tx_queue_release = mlx5_tx_queue_release,
1057         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1058         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1059         .mac_addr_remove = mlx5_mac_addr_remove,
1060         .mac_addr_add = mlx5_mac_addr_add,
1061         .mac_addr_set = mlx5_mac_addr_set,
1062         .set_mc_addr_list = mlx5_set_mc_addr_list,
1063         .mtu_set = mlx5_dev_set_mtu,
1064         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1065         .vlan_offload_set = mlx5_vlan_offload_set,
1066         .filter_ctrl = mlx5_dev_filter_ctrl,
1067         .rx_descriptor_status = mlx5_rx_descriptor_status,
1068         .tx_descriptor_status = mlx5_tx_descriptor_status,
1069         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1070         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1071         .is_removed = mlx5_is_removed,
1072         .get_module_info = mlx5_get_module_info,
1073         .get_module_eeprom = mlx5_get_module_eeprom,
1074 };
1075
1076 /**
1077  * Verify and store value for device argument.
1078  *
1079  * @param[in] key
1080  *   Key argument to verify.
1081  * @param[in] val
1082  *   Value associated with key.
1083  * @param opaque
1084  *   User data.
1085  *
1086  * @return
1087  *   0 on success, a negative errno value otherwise and rte_errno is set.
1088  */
1089 static int
1090 mlx5_args_check(const char *key, const char *val, void *opaque)
1091 {
1092         struct mlx5_dev_config *config = opaque;
1093         unsigned long tmp;
1094
1095         /* No-op, port representors are processed in mlx5_dev_spawn(). */
1096         if (!strcmp(MLX5_REPRESENTOR, key))
1097                 return 0;
1098         errno = 0;
1099         tmp = strtoul(val, NULL, 0);
1100         if (errno) {
1101                 rte_errno = errno;
1102                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1103                 return -rte_errno;
1104         }
1105         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1106                 config->cqe_comp = !!tmp;
1107         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1108                 config->cqe_pad = !!tmp;
1109         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1110                 config->hw_padding = !!tmp;
1111         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1112                 config->mprq.enabled = !!tmp;
1113         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1114                 config->mprq.stride_num_n = tmp;
1115         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1116                 config->mprq.max_memcpy_len = tmp;
1117         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1118                 config->mprq.min_rxqs_num = tmp;
1119         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1120                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1121                                  " converted to txq_inline_max", key);
1122                 config->txq_inline_max = tmp;
1123         } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1124                 config->txq_inline_max = tmp;
1125         } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1126                 config->txq_inline_min = tmp;
1127         } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1128                 config->txq_inline_mpw = tmp;
1129         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1130                 config->txqs_inline = tmp;
1131         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1132                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1133         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1134                 config->mps = !!tmp;
1135         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1136                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1137         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1138                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1139                                  " converted to txq_inline_mpw", key);
1140                 config->txq_inline_mpw = tmp;
1141         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1142                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1143         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1144                 config->rx_vec_en = !!tmp;
1145         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1146                 config->l3_vxlan_en = !!tmp;
1147         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1148                 config->vf_nl_en = !!tmp;
1149         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1150                 config->dv_esw_en = !!tmp;
1151         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1152                 config->dv_flow_en = !!tmp;
1153         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1154                 config->mr_ext_memseg_en = !!tmp;
1155         } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1156                 config->max_dump_files_num = tmp;
1157         } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1158                 config->lro.timeout = tmp;
1159         } else {
1160                 DRV_LOG(WARNING, "%s: unknown parameter", key);
1161                 rte_errno = EINVAL;
1162                 return -rte_errno;
1163         }
1164         return 0;
1165 }
1166
1167 /**
1168  * Parse device parameters.
1169  *
1170  * @param config
1171  *   Pointer to device configuration structure.
1172  * @param devargs
1173  *   Device arguments structure.
1174  *
1175  * @return
1176  *   0 on success, a negative errno value otherwise and rte_errno is set.
1177  */
1178 static int
1179 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1180 {
1181         const char **params = (const char *[]){
1182                 MLX5_RXQ_CQE_COMP_EN,
1183                 MLX5_RXQ_CQE_PAD_EN,
1184                 MLX5_RXQ_PKT_PAD_EN,
1185                 MLX5_RX_MPRQ_EN,
1186                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1187                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1188                 MLX5_RXQS_MIN_MPRQ,
1189                 MLX5_TXQ_INLINE,
1190                 MLX5_TXQ_INLINE_MIN,
1191                 MLX5_TXQ_INLINE_MAX,
1192                 MLX5_TXQ_INLINE_MPW,
1193                 MLX5_TXQS_MIN_INLINE,
1194                 MLX5_TXQS_MAX_VEC,
1195                 MLX5_TXQ_MPW_EN,
1196                 MLX5_TXQ_MPW_HDR_DSEG_EN,
1197                 MLX5_TXQ_MAX_INLINE_LEN,
1198                 MLX5_TX_VEC_EN,
1199                 MLX5_RX_VEC_EN,
1200                 MLX5_L3_VXLAN_EN,
1201                 MLX5_VF_NL_EN,
1202                 MLX5_DV_ESW_EN,
1203                 MLX5_DV_FLOW_EN,
1204                 MLX5_MR_EXT_MEMSEG_EN,
1205                 MLX5_REPRESENTOR,
1206                 MLX5_MAX_DUMP_FILES_NUM,
1207                 MLX5_LRO_TIMEOUT_USEC,
1208                 NULL,
1209         };
1210         struct rte_kvargs *kvlist;
1211         int ret = 0;
1212         int i;
1213
1214         if (devargs == NULL)
1215                 return 0;
1216         /* Following UGLY cast is done to pass checkpatch. */
1217         kvlist = rte_kvargs_parse(devargs->args, params);
1218         if (kvlist == NULL) {
1219                 rte_errno = EINVAL;
1220                 return -rte_errno;
1221         }
1222         /* Process parameters. */
1223         for (i = 0; (params[i] != NULL); ++i) {
1224                 if (rte_kvargs_count(kvlist, params[i])) {
1225                         ret = rte_kvargs_process(kvlist, params[i],
1226                                                  mlx5_args_check, config);
1227                         if (ret) {
1228                                 rte_errno = EINVAL;
1229                                 rte_kvargs_free(kvlist);
1230                                 return -rte_errno;
1231                         }
1232                 }
1233         }
1234         rte_kvargs_free(kvlist);
1235         return 0;
1236 }
1237
1238 static struct rte_pci_driver mlx5_driver;
1239
1240 /**
1241  * PMD global initialization.
1242  *
1243  * Independent from individual device, this function initializes global
1244  * per-PMD data structures distinguishing primary and secondary processes.
1245  * Hence, each initialization is called once per a process.
1246  *
1247  * @return
1248  *   0 on success, a negative errno value otherwise and rte_errno is set.
1249  */
1250 static int
1251 mlx5_init_once(void)
1252 {
1253         struct mlx5_shared_data *sd;
1254         struct mlx5_local_data *ld = &mlx5_local_data;
1255         int ret = 0;
1256
1257         if (mlx5_init_shared_data())
1258                 return -rte_errno;
1259         sd = mlx5_shared_data;
1260         assert(sd);
1261         rte_spinlock_lock(&sd->lock);
1262         switch (rte_eal_process_type()) {
1263         case RTE_PROC_PRIMARY:
1264                 if (sd->init_done)
1265                         break;
1266                 LIST_INIT(&sd->mem_event_cb_list);
1267                 rte_rwlock_init(&sd->mem_event_rwlock);
1268                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1269                                                 mlx5_mr_mem_event_cb, NULL);
1270                 ret = mlx5_mp_init_primary();
1271                 if (ret)
1272                         goto out;
1273                 sd->init_done = true;
1274                 break;
1275         case RTE_PROC_SECONDARY:
1276                 if (ld->init_done)
1277                         break;
1278                 ret = mlx5_mp_init_secondary();
1279                 if (ret)
1280                         goto out;
1281                 ++sd->secondary_cnt;
1282                 ld->init_done = true;
1283                 break;
1284         default:
1285                 break;
1286         }
1287 out:
1288         rte_spinlock_unlock(&sd->lock);
1289         return ret;
1290 }
1291
1292 /**
1293  * Configures the minimal amount of data to inline into WQE
1294  * while sending packets.
1295  *
1296  * - the txq_inline_min has the maximal priority, if this
1297  *   key is specified in devargs
1298  * - if DevX is enabled the inline mode is queried from the
1299  *   device (HCA attributes and NIC vport context if needed).
1300  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1301  *   and none (0 bytes) for other NICs
1302  *
1303  * @param spawn
1304  *   Verbs device parameters (name, port, switch_info) to spawn.
1305  * @param config
1306  *   Device configuration parameters.
1307  */
1308 static void
1309 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1310                     struct mlx5_dev_config *config)
1311 {
1312         if (config->txq_inline_min != MLX5_ARG_UNSET) {
1313                 /* Application defines size of inlined data explicitly. */
1314                 switch (spawn->pci_dev->id.device_id) {
1315                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1316                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1317                         if (config->txq_inline_min <
1318                                        (int)MLX5_INLINE_HSIZE_L2) {
1319                                 DRV_LOG(DEBUG,
1320                                         "txq_inline_mix aligned to minimal"
1321                                         " ConnectX-4 required value %d",
1322                                         (int)MLX5_INLINE_HSIZE_L2);
1323                                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1324                         }
1325                         break;
1326                 }
1327                 goto exit;
1328         }
1329         if (config->hca_attr.eth_net_offloads) {
1330                 /* We have DevX enabled, inline mode queried successfully. */
1331                 switch (config->hca_attr.wqe_inline_mode) {
1332                 case MLX5_CAP_INLINE_MODE_L2:
1333                         /* outer L2 header must be inlined. */
1334                         config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1335                         goto exit;
1336                 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1337                         /* No inline data are required by NIC. */
1338                         config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1339                         config->hw_vlan_insert =
1340                                 config->hca_attr.wqe_vlan_insert;
1341                         DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1342                         goto exit;
1343                 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1344                         /* inline mode is defined by NIC vport context. */
1345                         if (!config->hca_attr.eth_virt)
1346                                 break;
1347                         switch (config->hca_attr.vport_inline_mode) {
1348                         case MLX5_INLINE_MODE_NONE:
1349                                 config->txq_inline_min =
1350                                         MLX5_INLINE_HSIZE_NONE;
1351                                 goto exit;
1352                         case MLX5_INLINE_MODE_L2:
1353                                 config->txq_inline_min =
1354                                         MLX5_INLINE_HSIZE_L2;
1355                                 goto exit;
1356                         case MLX5_INLINE_MODE_IP:
1357                                 config->txq_inline_min =
1358                                         MLX5_INLINE_HSIZE_L3;
1359                                 goto exit;
1360                         case MLX5_INLINE_MODE_TCP_UDP:
1361                                 config->txq_inline_min =
1362                                         MLX5_INLINE_HSIZE_L4;
1363                                 goto exit;
1364                         case MLX5_INLINE_MODE_INNER_L2:
1365                                 config->txq_inline_min =
1366                                         MLX5_INLINE_HSIZE_INNER_L2;
1367                                 goto exit;
1368                         case MLX5_INLINE_MODE_INNER_IP:
1369                                 config->txq_inline_min =
1370                                         MLX5_INLINE_HSIZE_INNER_L3;
1371                                 goto exit;
1372                         case MLX5_INLINE_MODE_INNER_TCP_UDP:
1373                                 config->txq_inline_min =
1374                                         MLX5_INLINE_HSIZE_INNER_L4;
1375                                 goto exit;
1376                         }
1377                 }
1378         }
1379         /*
1380          * We get here if we are unable to deduce
1381          * inline data size with DevX. Try PCI ID
1382          * to determine old NICs.
1383          */
1384         switch (spawn->pci_dev->id.device_id) {
1385         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1386         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1387         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1388         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1389                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1390                 config->hw_vlan_insert = 0;
1391                 break;
1392         case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1393         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1394         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1395         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1396                 /*
1397                  * These NICs support VLAN insertion from WQE and
1398                  * report the wqe_vlan_insert flag. But there is the bug
1399                  * and PFC control may be broken, so disable feature.
1400                  */
1401                 config->hw_vlan_insert = 0;
1402                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1403                 break;
1404         default:
1405                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1406                 break;
1407         }
1408 exit:
1409         DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1410 }
1411
1412 /**
1413  * Allocate page of door-bells and register it using DevX API.
1414  *
1415  * @param [in] dev
1416  *   Pointer to Ethernet device.
1417  *
1418  * @return
1419  *   Pointer to new page on success, NULL otherwise.
1420  */
1421 static struct mlx5_devx_dbr_page *
1422 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1423 {
1424         struct mlx5_priv *priv = dev->data->dev_private;
1425         struct mlx5_devx_dbr_page *page;
1426
1427         /* Allocate space for door-bell page and management data. */
1428         page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1429                                  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1430         if (!page) {
1431                 DRV_LOG(ERR, "port %u cannot allocate dbr page",
1432                         dev->data->port_id);
1433                 return NULL;
1434         }
1435         /* Register allocated memory. */
1436         page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1437                                               MLX5_DBR_PAGE_SIZE, 0);
1438         if (!page->umem) {
1439                 DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1440                         dev->data->port_id);
1441                 rte_free(page);
1442                 return NULL;
1443         }
1444         return page;
1445 }
1446
1447 /**
1448  * Find the next available door-bell, allocate new page if needed.
1449  *
1450  * @param [in] dev
1451  *   Pointer to Ethernet device.
1452  * @param [out] dbr_page
1453  *   Door-bell page containing the page data.
1454  *
1455  * @return
1456  *   Door-bell address offset on success, a negative error value otherwise.
1457  */
1458 int64_t
1459 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1460 {
1461         struct mlx5_priv *priv = dev->data->dev_private;
1462         struct mlx5_devx_dbr_page *page = NULL;
1463         uint32_t i, j;
1464
1465         LIST_FOREACH(page, &priv->dbrpgs, next)
1466                 if (page->dbr_count < MLX5_DBR_PER_PAGE)
1467                         break;
1468         if (!page) { /* No page with free door-bell exists. */
1469                 page = mlx5_alloc_dbr_page(dev);
1470                 if (!page) /* Failed to allocate new page. */
1471                         return (-1);
1472                 LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1473         }
1474         /* Loop to find bitmap part with clear bit. */
1475         for (i = 0;
1476              i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1477              i++)
1478                 ; /* Empty. */
1479         /* Find the first clear bit. */
1480         j = rte_bsf64(~page->dbr_bitmap[i]);
1481         assert(i < (MLX5_DBR_PER_PAGE / 64));
1482         page->dbr_bitmap[i] |= (1 << j);
1483         page->dbr_count++;
1484         *dbr_page = page;
1485         return (((i * 64) + j) * sizeof(uint64_t));
1486 }
1487
1488 /**
1489  * Release a door-bell record.
1490  *
1491  * @param [in] dev
1492  *   Pointer to Ethernet device.
1493  * @param [in] umem_id
1494  *   UMEM ID of page containing the door-bell record to release.
1495  * @param [in] offset
1496  *   Offset of door-bell record in page.
1497  *
1498  * @return
1499  *   0 on success, a negative error value otherwise.
1500  */
1501 int32_t
1502 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1503 {
1504         struct mlx5_priv *priv = dev->data->dev_private;
1505         struct mlx5_devx_dbr_page *page = NULL;
1506         int ret = 0;
1507
1508         LIST_FOREACH(page, &priv->dbrpgs, next)
1509                 /* Find the page this address belongs to. */
1510                 if (page->umem->umem_id == umem_id)
1511                         break;
1512         if (!page)
1513                 return -EINVAL;
1514         page->dbr_count--;
1515         if (!page->dbr_count) {
1516                 /* Page not used, free it and remove from list. */
1517                 LIST_REMOVE(page, next);
1518                 if (page->umem)
1519                         ret = -mlx5_glue->devx_umem_dereg(page->umem);
1520                 rte_free(page);
1521         } else {
1522                 /* Mark in bitmap that this door-bell is not in use. */
1523                 offset /= MLX5_DBR_SIZE;
1524                 int i = offset / 64;
1525                 int j = offset % 64;
1526
1527                 page->dbr_bitmap[i] &= ~(1 << j);
1528         }
1529         return ret;
1530 }
1531
1532 /**
1533  * Check sibling device configurations.
1534  *
1535  * Sibling devices sharing the Infiniband device context
1536  * should have compatible configurations. This regards
1537  * representors and bonding slaves.
1538  *
1539  * @param priv
1540  *   Private device descriptor.
1541  * @param config
1542  *   Configuration of the device is going to be created.
1543  *
1544  * @return
1545  *   0 on success, EINVAL otherwise
1546  */
1547 static int
1548 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
1549                               struct mlx5_dev_config *config)
1550 {
1551         struct mlx5_ibv_shared *sh = priv->sh;
1552         struct mlx5_dev_config *sh_conf = NULL;
1553         uint16_t port_id;
1554
1555         assert(sh);
1556         /* Nothing to compare for the single/first device. */
1557         if (sh->refcnt == 1)
1558                 return 0;
1559         /* Find the device with shared context. */
1560         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1561                 struct mlx5_priv *opriv =
1562                         rte_eth_devices[port_id].data->dev_private;
1563
1564                 if (opriv && opriv != priv && opriv->sh == sh) {
1565                         sh_conf = &opriv->config;
1566                         break;
1567                 }
1568         }
1569         if (!sh_conf)
1570                 return 0;
1571         if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
1572                 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
1573                              " for shared %s context", sh->ibdev_name);
1574                 rte_errno = EINVAL;
1575                 return rte_errno;
1576         }
1577         return 0;
1578 }
1579 /**
1580  * Spawn an Ethernet device from Verbs information.
1581  *
1582  * @param dpdk_dev
1583  *   Backing DPDK device.
1584  * @param spawn
1585  *   Verbs device parameters (name, port, switch_info) to spawn.
1586  * @param config
1587  *   Device configuration parameters.
1588  *
1589  * @return
1590  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1591  *   is set. The following errors are defined:
1592  *
1593  *   EBUSY: device is not supposed to be spawned.
1594  *   EEXIST: device is already spawned
1595  */
1596 static struct rte_eth_dev *
1597 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1598                struct mlx5_dev_spawn_data *spawn,
1599                struct mlx5_dev_config config)
1600 {
1601         const struct mlx5_switch_info *switch_info = &spawn->info;
1602         struct mlx5_ibv_shared *sh = NULL;
1603         struct ibv_port_attr port_attr;
1604         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1605         struct rte_eth_dev *eth_dev = NULL;
1606         struct mlx5_priv *priv = NULL;
1607         int err = 0;
1608         unsigned int hw_padding = 0;
1609         unsigned int mps;
1610         unsigned int cqe_comp;
1611         unsigned int cqe_pad = 0;
1612         unsigned int tunnel_en = 0;
1613         unsigned int mpls_en = 0;
1614         unsigned int swp = 0;
1615         unsigned int mprq = 0;
1616         unsigned int mprq_min_stride_size_n = 0;
1617         unsigned int mprq_max_stride_size_n = 0;
1618         unsigned int mprq_min_stride_num_n = 0;
1619         unsigned int mprq_max_stride_num_n = 0;
1620         struct rte_ether_addr mac;
1621         char name[RTE_ETH_NAME_MAX_LEN];
1622         int own_domain_id = 0;
1623         uint16_t port_id;
1624         unsigned int i;
1625 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1626         struct mlx5dv_devx_port devx_port;
1627 #endif
1628
1629         /* Determine if this port representor is supposed to be spawned. */
1630         if (switch_info->representor && dpdk_dev->devargs) {
1631                 struct rte_eth_devargs eth_da;
1632
1633                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
1634                 if (err) {
1635                         rte_errno = -err;
1636                         DRV_LOG(ERR, "failed to process device arguments: %s",
1637                                 strerror(rte_errno));
1638                         return NULL;
1639                 }
1640                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
1641                         if (eth_da.representor_ports[i] ==
1642                             (uint16_t)switch_info->port_name)
1643                                 break;
1644                 if (i == eth_da.nb_representor_ports) {
1645                         rte_errno = EBUSY;
1646                         return NULL;
1647                 }
1648         }
1649         /* Build device name. */
1650         if (spawn->pf_bond <  0) {
1651                 /* Single device. */
1652                 if (!switch_info->representor)
1653                         strlcpy(name, dpdk_dev->name, sizeof(name));
1654                 else
1655                         snprintf(name, sizeof(name), "%s_representor_%u",
1656                                  dpdk_dev->name, switch_info->port_name);
1657         } else {
1658                 /* Bonding device. */
1659                 if (!switch_info->representor)
1660                         snprintf(name, sizeof(name), "%s_%s",
1661                                  dpdk_dev->name, spawn->ibv_dev->name);
1662                 else
1663                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
1664                                  dpdk_dev->name, spawn->ibv_dev->name,
1665                                  switch_info->port_name);
1666         }
1667         /* check if the device is already spawned */
1668         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1669                 rte_errno = EEXIST;
1670                 return NULL;
1671         }
1672         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1673         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1674                 eth_dev = rte_eth_dev_attach_secondary(name);
1675                 if (eth_dev == NULL) {
1676                         DRV_LOG(ERR, "can not attach rte ethdev");
1677                         rte_errno = ENOMEM;
1678                         return NULL;
1679                 }
1680                 eth_dev->device = dpdk_dev;
1681                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
1682                 err = mlx5_proc_priv_init(eth_dev);
1683                 if (err)
1684                         return NULL;
1685                 /* Receive command fd from primary process */
1686                 err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
1687                 if (err < 0)
1688                         return NULL;
1689                 /* Remap UAR for Tx queues. */
1690                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
1691                 if (err)
1692                         return NULL;
1693                 /*
1694                  * Ethdev pointer is still required as input since
1695                  * the primary device is not accessible from the
1696                  * secondary process.
1697                  */
1698                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1699                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1700                 return eth_dev;
1701         }
1702         sh = mlx5_alloc_shared_ibctx(spawn);
1703         if (!sh)
1704                 return NULL;
1705         config.devx = sh->devx;
1706 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
1707         config.dest_tir = 1;
1708 #endif
1709 #ifdef HAVE_IBV_MLX5_MOD_SWP
1710         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1711 #endif
1712         /*
1713          * Multi-packet send is supported by ConnectX-4 Lx PF as well
1714          * as all ConnectX-5 devices.
1715          */
1716 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1717         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1718 #endif
1719 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1720         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1721 #endif
1722         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1723         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1724                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1725                         DRV_LOG(DEBUG, "enhanced MPW is supported");
1726                         mps = MLX5_MPW_ENHANCED;
1727                 } else {
1728                         DRV_LOG(DEBUG, "MPW is supported");
1729                         mps = MLX5_MPW;
1730                 }
1731         } else {
1732                 DRV_LOG(DEBUG, "MPW isn't supported");
1733                 mps = MLX5_MPW_DISABLED;
1734         }
1735 #ifdef HAVE_IBV_MLX5_MOD_SWP
1736         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1737                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1738         DRV_LOG(DEBUG, "SWP support: %u", swp);
1739 #endif
1740         config.swp = !!swp;
1741 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1742         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1743                 struct mlx5dv_striding_rq_caps mprq_caps =
1744                         dv_attr.striding_rq_caps;
1745
1746                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1747                         mprq_caps.min_single_stride_log_num_of_bytes);
1748                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1749                         mprq_caps.max_single_stride_log_num_of_bytes);
1750                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1751                         mprq_caps.min_single_wqe_log_num_of_strides);
1752                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1753                         mprq_caps.max_single_wqe_log_num_of_strides);
1754                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1755                         mprq_caps.supported_qpts);
1756                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1757                 mprq = 1;
1758                 mprq_min_stride_size_n =
1759                         mprq_caps.min_single_stride_log_num_of_bytes;
1760                 mprq_max_stride_size_n =
1761                         mprq_caps.max_single_stride_log_num_of_bytes;
1762                 mprq_min_stride_num_n =
1763                         mprq_caps.min_single_wqe_log_num_of_strides;
1764                 mprq_max_stride_num_n =
1765                         mprq_caps.max_single_wqe_log_num_of_strides;
1766                 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1767                                                    mprq_min_stride_num_n);
1768         }
1769 #endif
1770         if (RTE_CACHE_LINE_SIZE == 128 &&
1771             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1772                 cqe_comp = 0;
1773         else
1774                 cqe_comp = 1;
1775         config.cqe_comp = cqe_comp;
1776 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1777         /* Whether device supports 128B Rx CQE padding. */
1778         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1779                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1780 #endif
1781 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1782         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1783                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
1784                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1785                              (dv_attr.tunnel_offloads_caps &
1786                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1787         }
1788         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1789                 tunnel_en ? "" : "not ");
1790 #else
1791         DRV_LOG(WARNING,
1792                 "tunnel offloading disabled due to old OFED/rdma-core version");
1793 #endif
1794         config.tunnel_en = tunnel_en;
1795 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1796         mpls_en = ((dv_attr.tunnel_offloads_caps &
1797                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1798                    (dv_attr.tunnel_offloads_caps &
1799                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1800         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1801                 mpls_en ? "" : "not ");
1802 #else
1803         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1804                 " old OFED/rdma-core version or firmware configuration");
1805 #endif
1806         config.mpls_en = mpls_en;
1807         /* Check port status. */
1808         err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1809         if (err) {
1810                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1811                 goto error;
1812         }
1813         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1814                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1815                 err = EINVAL;
1816                 goto error;
1817         }
1818         if (port_attr.state != IBV_PORT_ACTIVE)
1819                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1820                         mlx5_glue->port_state_str(port_attr.state),
1821                         port_attr.state);
1822         /* Allocate private eth device data. */
1823         priv = rte_zmalloc("ethdev private structure",
1824                            sizeof(*priv),
1825                            RTE_CACHE_LINE_SIZE);
1826         if (priv == NULL) {
1827                 DRV_LOG(ERR, "priv allocation failure");
1828                 err = ENOMEM;
1829                 goto error;
1830         }
1831         priv->sh = sh;
1832         priv->ibv_port = spawn->ibv_port;
1833         priv->pci_dev = spawn->pci_dev;
1834         priv->mtu = RTE_ETHER_MTU;
1835 #ifndef RTE_ARCH_64
1836         /* Initialize UAR access locks for 32bit implementations. */
1837         rte_spinlock_init(&priv->uar_lock_cq);
1838         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1839                 rte_spinlock_init(&priv->uar_lock[i]);
1840 #endif
1841         /* Some internal functions rely on Netlink sockets, open them now. */
1842         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1843         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1844         priv->nl_sn = 0;
1845         priv->representor = !!switch_info->representor;
1846         priv->master = !!switch_info->master;
1847         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1848         priv->vport_meta_tag = 0;
1849         priv->vport_meta_mask = 0;
1850         priv->pf_bond = spawn->pf_bond;
1851 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1852         /*
1853          * The DevX port query API is implemented. E-Switch may use
1854          * either vport or reg_c[0] metadata register to match on
1855          * vport index. The engaged part of metadata register is
1856          * defined by mask.
1857          */
1858         devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
1859                               MLX5DV_DEVX_PORT_MATCH_REG_C_0;
1860         err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port, &devx_port);
1861         if (err) {
1862                 DRV_LOG(WARNING, "can't query devx port %d on device %s\n",
1863                         spawn->ibv_port, spawn->ibv_dev->name);
1864                 devx_port.comp_mask = 0;
1865         }
1866         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
1867                 priv->vport_meta_tag = devx_port.reg_c_0.value;
1868                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
1869                 if (!priv->vport_meta_mask) {
1870                         DRV_LOG(ERR, "vport zero mask for port %d"
1871                                      " on bonding device %s\n",
1872                                      spawn->ibv_port, spawn->ibv_dev->name);
1873                         err = ENOTSUP;
1874                         goto error;
1875                 }
1876                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1877                         DRV_LOG(ERR, "invalid vport tag for port %d"
1878                                      " on bonding device %s\n",
1879                                      spawn->ibv_port, spawn->ibv_dev->name);
1880                         err = ENOTSUP;
1881                         goto error;
1882                 }
1883         } else if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
1884                 priv->vport_id = devx_port.vport_num;
1885         } else if (spawn->pf_bond >= 0) {
1886                 DRV_LOG(ERR, "can't deduce vport index for port %d"
1887                              " on bonding device %s\n",
1888                              spawn->ibv_port, spawn->ibv_dev->name);
1889                 err = ENOTSUP;
1890                 goto error;
1891         } else {
1892                 /* Suppose vport index in compatible way. */
1893                 priv->vport_id = switch_info->representor ?
1894                                  switch_info->port_name + 1 : -1;
1895         }
1896 #else
1897         /*
1898          * Kernel/rdma_core support single E-Switch per PF configurations
1899          * only and vport_id field contains the vport index for
1900          * associated VF, which is deduced from representor port name.
1901          * For example, let's have the IB device port 10, it has
1902          * attached network device eth0, which has port name attribute
1903          * pf0vf2, we can deduce the VF number as 2, and set vport index
1904          * as 3 (2+1). This assigning schema should be changed if the
1905          * multiple E-Switch instances per PF configurations or/and PCI
1906          * subfunctions are added.
1907          */
1908         priv->vport_id = switch_info->representor ?
1909                          switch_info->port_name + 1 : -1;
1910 #endif
1911         /* representor_id field keeps the unmodified VF index. */
1912         priv->representor_id = switch_info->representor ?
1913                                switch_info->port_name : -1;
1914         /*
1915          * Look for sibling devices in order to reuse their switch domain
1916          * if any, otherwise allocate one.
1917          */
1918         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1919                 const struct mlx5_priv *opriv =
1920                         rte_eth_devices[port_id].data->dev_private;
1921
1922                 if (!opriv ||
1923                     opriv->sh != priv->sh ||
1924                         opriv->domain_id ==
1925                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1926                         continue;
1927                 priv->domain_id = opriv->domain_id;
1928                 break;
1929         }
1930         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1931                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1932                 if (err) {
1933                         err = rte_errno;
1934                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
1935                                 strerror(rte_errno));
1936                         goto error;
1937                 }
1938                 own_domain_id = 1;
1939         }
1940         err = mlx5_args(&config, dpdk_dev->devargs);
1941         if (err) {
1942                 err = rte_errno;
1943                 DRV_LOG(ERR, "failed to process device arguments: %s",
1944                         strerror(rte_errno));
1945                 goto error;
1946         }
1947         err = mlx5_dev_check_sibling_config(priv, &config);
1948         if (err)
1949                 goto error;
1950         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1951                             IBV_DEVICE_RAW_IP_CSUM);
1952         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1953                 (config.hw_csum ? "" : "not "));
1954 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1955         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1956         DRV_LOG(DEBUG, "counters are not supported");
1957 #endif
1958 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1959         if (config.dv_flow_en) {
1960                 DRV_LOG(WARNING, "DV flow is not supported");
1961                 config.dv_flow_en = 0;
1962         }
1963 #endif
1964         config.ind_table_max_size =
1965                 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1966         /*
1967          * Remove this check once DPDK supports larger/variable
1968          * indirection tables.
1969          */
1970         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1971                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1972         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1973                 config.ind_table_max_size);
1974         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1975                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1976         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1977                 (config.hw_vlan_strip ? "" : "not "));
1978         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1979                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
1980         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1981                 (config.hw_fcs_strip ? "" : "not "));
1982 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1983         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1984 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1985         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1986                         IBV_DEVICE_PCI_WRITE_END_PADDING);
1987 #endif
1988         if (config.hw_padding && !hw_padding) {
1989                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1990                 config.hw_padding = 0;
1991         } else if (config.hw_padding) {
1992                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1993         }
1994         config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1995                       (sh->device_attr.tso_caps.supported_qpts &
1996                        (1 << IBV_QPT_RAW_PACKET)));
1997         if (config.tso)
1998                 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1999         /*
2000          * MPW is disabled by default, while the Enhanced MPW is enabled
2001          * by default.
2002          */
2003         if (config.mps == MLX5_ARG_UNSET)
2004                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2005                                                           MLX5_MPW_DISABLED;
2006         else
2007                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2008         DRV_LOG(INFO, "%sMPS is %s",
2009                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
2010                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2011         if (config.cqe_comp && !cqe_comp) {
2012                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2013                 config.cqe_comp = 0;
2014         }
2015         if (config.cqe_pad && !cqe_pad) {
2016                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2017                 config.cqe_pad = 0;
2018         } else if (config.cqe_pad) {
2019                 DRV_LOG(INFO, "Rx CQE padding is enabled");
2020         }
2021         if (config.devx) {
2022                 priv->counter_fallback = 0;
2023                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2024                 if (err) {
2025                         err = -err;
2026                         goto error;
2027                 }
2028                 if (!config.hca_attr.flow_counters_dump)
2029                         priv->counter_fallback = 1;
2030 #ifndef HAVE_IBV_DEVX_ASYNC
2031                 priv->counter_fallback = 1;
2032 #endif
2033                 if (priv->counter_fallback)
2034                         DRV_LOG(INFO, "Use fall-back DV counter management\n");
2035                 /* Check for LRO support. */
2036                 if (config.dest_tir && config.hca_attr.lro_cap &&
2037                     config.dv_flow_en) {
2038                         /* TBD check tunnel lro caps. */
2039                         config.lro.supported = config.hca_attr.lro_cap;
2040                         DRV_LOG(DEBUG, "Device supports LRO");
2041                         /*
2042                          * If LRO timeout is not configured by application,
2043                          * use the minimal supported value.
2044                          */
2045                         if (!config.lro.timeout)
2046                                 config.lro.timeout =
2047                                 config.hca_attr.lro_timer_supported_periods[0];
2048                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2049                                 config.lro.timeout);
2050                 }
2051         }
2052         if (config.mprq.enabled && mprq) {
2053                 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2054                     config.mprq.stride_num_n < mprq_min_stride_num_n) {
2055                         config.mprq.stride_num_n =
2056                                 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2057                                         mprq_min_stride_num_n);
2058                         DRV_LOG(WARNING,
2059                                 "the number of strides"
2060                                 " for Multi-Packet RQ is out of range,"
2061                                 " setting default value (%u)",
2062                                 1 << config.mprq.stride_num_n);
2063                 }
2064                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2065                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2066         } else if (config.mprq.enabled && !mprq) {
2067                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2068                 config.mprq.enabled = 0;
2069         }
2070         if (config.max_dump_files_num == 0)
2071                 config.max_dump_files_num = 128;
2072         eth_dev = rte_eth_dev_allocate(name);
2073         if (eth_dev == NULL) {
2074                 DRV_LOG(ERR, "can not allocate rte ethdev");
2075                 err = ENOMEM;
2076                 goto error;
2077         }
2078         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2079         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2080         if (priv->representor) {
2081                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2082                 eth_dev->data->representor_id = priv->representor_id;
2083         }
2084         /*
2085          * Store associated network device interface index. This index
2086          * is permanent throughout the lifetime of device. So, we may store
2087          * the ifindex here and use the cached value further.
2088          */
2089         assert(spawn->ifindex);
2090         priv->if_index = spawn->ifindex;
2091         eth_dev->data->dev_private = priv;
2092         priv->dev_data = eth_dev->data;
2093         eth_dev->data->mac_addrs = priv->mac;
2094         eth_dev->device = dpdk_dev;
2095         /* Configure the first MAC address by default. */
2096         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2097                 DRV_LOG(ERR,
2098                         "port %u cannot get MAC address, is mlx5_en"
2099                         " loaded? (errno: %s)",
2100                         eth_dev->data->port_id, strerror(rte_errno));
2101                 err = ENODEV;
2102                 goto error;
2103         }
2104         DRV_LOG(INFO,
2105                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2106                 eth_dev->data->port_id,
2107                 mac.addr_bytes[0], mac.addr_bytes[1],
2108                 mac.addr_bytes[2], mac.addr_bytes[3],
2109                 mac.addr_bytes[4], mac.addr_bytes[5]);
2110 #ifndef NDEBUG
2111         {
2112                 char ifname[IF_NAMESIZE];
2113
2114                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2115                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2116                                 eth_dev->data->port_id, ifname);
2117                 else
2118                         DRV_LOG(DEBUG, "port %u ifname is unknown",
2119                                 eth_dev->data->port_id);
2120         }
2121 #endif
2122         /* Get actual MTU if possible. */
2123         err = mlx5_get_mtu(eth_dev, &priv->mtu);
2124         if (err) {
2125                 err = rte_errno;
2126                 goto error;
2127         }
2128         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2129                 priv->mtu);
2130         /* Initialize burst functions to prevent crashes before link-up. */
2131         eth_dev->rx_pkt_burst = removed_rx_burst;
2132         eth_dev->tx_pkt_burst = removed_tx_burst;
2133         eth_dev->dev_ops = &mlx5_dev_ops;
2134         /* Register MAC address. */
2135         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2136         if (config.vf && config.vf_nl_en)
2137                 mlx5_nl_mac_addr_sync(eth_dev);
2138         TAILQ_INIT(&priv->flows);
2139         TAILQ_INIT(&priv->ctrl_flows);
2140         /* Hint libmlx5 to use PMD allocator for data plane resources */
2141         struct mlx5dv_ctx_allocators alctr = {
2142                 .alloc = &mlx5_alloc_verbs_buf,
2143                 .free = &mlx5_free_verbs_buf,
2144                 .data = priv,
2145         };
2146         mlx5_glue->dv_set_context_attr(sh->ctx,
2147                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2148                                        (void *)((uintptr_t)&alctr));
2149         /* Bring Ethernet device up. */
2150         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2151                 eth_dev->data->port_id);
2152         mlx5_set_link_up(eth_dev);
2153         /*
2154          * Even though the interrupt handler is not installed yet,
2155          * interrupts will still trigger on the async_fd from
2156          * Verbs context returned by ibv_open_device().
2157          */
2158         mlx5_link_update(eth_dev, 0);
2159 #ifdef HAVE_MLX5DV_DR_ESWITCH
2160         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2161               (switch_info->representor || switch_info->master)))
2162                 config.dv_esw_en = 0;
2163 #else
2164         config.dv_esw_en = 0;
2165 #endif
2166         /* Detect minimal data bytes to inline. */
2167         mlx5_set_min_inline(spawn, &config);
2168         /* Store device configuration on private structure. */
2169         priv->config = config;
2170         /* Create context for virtual machine VLAN workaround. */
2171         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2172         if (config.dv_flow_en) {
2173                 err = mlx5_alloc_shared_dr(priv);
2174                 if (err)
2175                         goto error;
2176         }
2177         /* Supported Verbs flow priority number detection. */
2178         err = mlx5_flow_discover_priorities(eth_dev);
2179         if (err < 0) {
2180                 err = -err;
2181                 goto error;
2182         }
2183         priv->config.flow_prio = err;
2184         return eth_dev;
2185 error:
2186         if (priv) {
2187                 if (priv->sh)
2188                         mlx5_free_shared_dr(priv);
2189                 if (priv->nl_socket_route >= 0)
2190                         close(priv->nl_socket_route);
2191                 if (priv->nl_socket_rdma >= 0)
2192                         close(priv->nl_socket_rdma);
2193                 if (priv->vmwa_context)
2194                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
2195                 if (own_domain_id)
2196                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2197                 rte_free(priv);
2198                 if (eth_dev != NULL)
2199                         eth_dev->data->dev_private = NULL;
2200         }
2201         if (eth_dev != NULL) {
2202                 /* mac_addrs must not be freed alone because part of dev_private */
2203                 eth_dev->data->mac_addrs = NULL;
2204                 rte_eth_dev_release_port(eth_dev);
2205         }
2206         if (sh)
2207                 mlx5_free_shared_ibctx(sh);
2208         assert(err > 0);
2209         rte_errno = err;
2210         return NULL;
2211 }
2212
2213 /**
2214  * Comparison callback to sort device data.
2215  *
2216  * This is meant to be used with qsort().
2217  *
2218  * @param a[in]
2219  *   Pointer to pointer to first data object.
2220  * @param b[in]
2221  *   Pointer to pointer to second data object.
2222  *
2223  * @return
2224  *   0 if both objects are equal, less than 0 if the first argument is less
2225  *   than the second, greater than 0 otherwise.
2226  */
2227 static int
2228 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2229 {
2230         const struct mlx5_switch_info *si_a =
2231                 &((const struct mlx5_dev_spawn_data *)a)->info;
2232         const struct mlx5_switch_info *si_b =
2233                 &((const struct mlx5_dev_spawn_data *)b)->info;
2234         int ret;
2235
2236         /* Master device first. */
2237         ret = si_b->master - si_a->master;
2238         if (ret)
2239                 return ret;
2240         /* Then representor devices. */
2241         ret = si_b->representor - si_a->representor;
2242         if (ret)
2243                 return ret;
2244         /* Unidentified devices come last in no specific order. */
2245         if (!si_a->representor)
2246                 return 0;
2247         /* Order representors by name. */
2248         return si_a->port_name - si_b->port_name;
2249 }
2250
2251 /**
2252  * Match PCI information for possible slaves of bonding device.
2253  *
2254  * @param[in] ibv_dev
2255  *   Pointer to Infiniband device structure.
2256  * @param[in] pci_dev
2257  *   Pointer to PCI device structure to match PCI address.
2258  * @param[in] nl_rdma
2259  *   Netlink RDMA group socket handle.
2260  *
2261  * @return
2262  *   negative value if no bonding device found, otherwise
2263  *   positive index of slave PF in bonding.
2264  */
2265 static int
2266 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2267                            const struct rte_pci_device *pci_dev,
2268                            int nl_rdma)
2269 {
2270         char ifname[IF_NAMESIZE + 1];
2271         unsigned int ifindex;
2272         unsigned int np, i;
2273         FILE *file = NULL;
2274         int pf = -1;
2275
2276         /*
2277          * Try to get master device name. If something goes
2278          * wrong suppose the lack of kernel support and no
2279          * bonding devices.
2280          */
2281         if (nl_rdma < 0)
2282                 return -1;
2283         if (!strstr(ibv_dev->name, "bond"))
2284                 return -1;
2285         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2286         if (!np)
2287                 return -1;
2288         /*
2289          * The Master device might not be on the predefined
2290          * port (not on port index 1, it is not garanted),
2291          * we have to scan all Infiniband device port and
2292          * find master.
2293          */
2294         for (i = 1; i <= np; ++i) {
2295                 /* Check whether Infiniband port is populated. */
2296                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2297                 if (!ifindex)
2298                         continue;
2299                 if (!if_indextoname(ifindex, ifname))
2300                         continue;
2301                 /* Try to read bonding slave names from sysfs. */
2302                 MKSTR(slaves,
2303                       "/sys/class/net/%s/master/bonding/slaves", ifname);
2304                 file = fopen(slaves, "r");
2305                 if (file)
2306                         break;
2307         }
2308         if (!file)
2309                 return -1;
2310         /* Use safe format to check maximal buffer length. */
2311         assert(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2312         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2313                 char tmp_str[IF_NAMESIZE + 32];
2314                 struct rte_pci_addr pci_addr;
2315                 struct mlx5_switch_info info;
2316
2317                 /* Process slave interface names in the loop. */
2318                 snprintf(tmp_str, sizeof(tmp_str),
2319                          "/sys/class/net/%s", ifname);
2320                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2321                         DRV_LOG(WARNING, "can not get PCI address"
2322                                          " for netdev \"%s\"", ifname);
2323                         continue;
2324                 }
2325                 if (pci_dev->addr.domain != pci_addr.domain ||
2326                     pci_dev->addr.bus != pci_addr.bus ||
2327                     pci_dev->addr.devid != pci_addr.devid ||
2328                     pci_dev->addr.function != pci_addr.function)
2329                         continue;
2330                 /* Slave interface PCI address match found. */
2331                 fclose(file);
2332                 snprintf(tmp_str, sizeof(tmp_str),
2333                          "/sys/class/net/%s/phys_port_name", ifname);
2334                 file = fopen(tmp_str, "rb");
2335                 if (!file)
2336                         break;
2337                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2338                 if (fscanf(file, "%32s", tmp_str) == 1)
2339                         mlx5_translate_port_name(tmp_str, &info);
2340                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2341                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2342                         pf = info.port_name;
2343                 break;
2344         }
2345         if (file)
2346                 fclose(file);
2347         return pf;
2348 }
2349
2350 /**
2351  * DPDK callback to register a PCI device.
2352  *
2353  * This function spawns Ethernet devices out of a given PCI device.
2354  *
2355  * @param[in] pci_drv
2356  *   PCI driver structure (mlx5_driver).
2357  * @param[in] pci_dev
2358  *   PCI device information.
2359  *
2360  * @return
2361  *   0 on success, a negative errno value otherwise and rte_errno is set.
2362  */
2363 static int
2364 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2365                struct rte_pci_device *pci_dev)
2366 {
2367         struct ibv_device **ibv_list;
2368         /*
2369          * Number of found IB Devices matching with requested PCI BDF.
2370          * nd != 1 means there are multiple IB devices over the same
2371          * PCI device and we have representors and master.
2372          */
2373         unsigned int nd = 0;
2374         /*
2375          * Number of found IB device Ports. nd = 1 and np = 1..n means
2376          * we have the single multiport IB device, and there may be
2377          * representors attached to some of found ports.
2378          */
2379         unsigned int np = 0;
2380         /*
2381          * Number of DPDK ethernet devices to Spawn - either over
2382          * multiple IB devices or multiple ports of single IB device.
2383          * Actually this is the number of iterations to spawn.
2384          */
2385         unsigned int ns = 0;
2386         /*
2387          * Bonding device
2388          *   < 0 - no bonding device (single one)
2389          *  >= 0 - bonding device (value is slave PF index)
2390          */
2391         int bd = -1;
2392         struct mlx5_dev_spawn_data *list = NULL;
2393         struct mlx5_dev_config dev_config;
2394         int ret;
2395
2396         ret = mlx5_init_once();
2397         if (ret) {
2398                 DRV_LOG(ERR, "unable to init PMD global data: %s",
2399                         strerror(rte_errno));
2400                 return -rte_errno;
2401         }
2402         assert(pci_drv == &mlx5_driver);
2403         errno = 0;
2404         ibv_list = mlx5_glue->get_device_list(&ret);
2405         if (!ibv_list) {
2406                 rte_errno = errno ? errno : ENOSYS;
2407                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2408                 return -rte_errno;
2409         }
2410         /*
2411          * First scan the list of all Infiniband devices to find
2412          * matching ones, gathering into the list.
2413          */
2414         struct ibv_device *ibv_match[ret + 1];
2415         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2416         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2417         unsigned int i;
2418
2419         while (ret-- > 0) {
2420                 struct rte_pci_addr pci_addr;
2421
2422                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2423                 bd = mlx5_device_bond_pci_match
2424                                 (ibv_list[ret], pci_dev, nl_rdma);
2425                 if (bd >= 0) {
2426                         /*
2427                          * Bonding device detected. Only one match is allowed,
2428                          * the bonding is supported over multi-port IB device,
2429                          * there should be no matches on representor PCI
2430                          * functions or non VF LAG bonding devices with
2431                          * specified address.
2432                          */
2433                         if (nd) {
2434                                 DRV_LOG(ERR,
2435                                         "multiple PCI match on bonding device"
2436                                         "\"%s\" found", ibv_list[ret]->name);
2437                                 rte_errno = ENOENT;
2438                                 ret = -rte_errno;
2439                                 goto exit;
2440                         }
2441                         DRV_LOG(INFO, "PCI information matches for"
2442                                       " slave %d bonding device \"%s\"",
2443                                       bd, ibv_list[ret]->name);
2444                         ibv_match[nd++] = ibv_list[ret];
2445                         break;
2446                 }
2447                 if (mlx5_dev_to_pci_addr
2448                         (ibv_list[ret]->ibdev_path, &pci_addr))
2449                         continue;
2450                 if (pci_dev->addr.domain != pci_addr.domain ||
2451                     pci_dev->addr.bus != pci_addr.bus ||
2452                     pci_dev->addr.devid != pci_addr.devid ||
2453                     pci_dev->addr.function != pci_addr.function)
2454                         continue;
2455                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2456                         ibv_list[ret]->name);
2457                 ibv_match[nd++] = ibv_list[ret];
2458         }
2459         ibv_match[nd] = NULL;
2460         if (!nd) {
2461                 /* No device matches, just complain and bail out. */
2462                 DRV_LOG(WARNING,
2463                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
2464                         " are kernel drivers loaded?",
2465                         pci_dev->addr.domain, pci_dev->addr.bus,
2466                         pci_dev->addr.devid, pci_dev->addr.function);
2467                 rte_errno = ENOENT;
2468                 ret = -rte_errno;
2469                 goto exit;
2470         }
2471         if (nd == 1) {
2472                 /*
2473                  * Found single matching device may have multiple ports.
2474                  * Each port may be representor, we have to check the port
2475                  * number and check the representors existence.
2476                  */
2477                 if (nl_rdma >= 0)
2478                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2479                 if (!np)
2480                         DRV_LOG(WARNING, "can not get IB device \"%s\""
2481                                          " ports number", ibv_match[0]->name);
2482                 if (bd >= 0 && !np) {
2483                         DRV_LOG(ERR, "can not get ports"
2484                                      " for bonding device");
2485                         rte_errno = ENOENT;
2486                         ret = -rte_errno;
2487                         goto exit;
2488                 }
2489         }
2490 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2491         if (bd >= 0) {
2492                 /*
2493                  * This may happen if there is VF LAG kernel support and
2494                  * application is compiled with older rdma_core library.
2495                  */
2496                 DRV_LOG(ERR,
2497                         "No kernel/verbs support for VF LAG bonding found.");
2498                 rte_errno = ENOTSUP;
2499                 ret = -rte_errno;
2500                 goto exit;
2501         }
2502 #endif
2503         /*
2504          * Now we can determine the maximal
2505          * amount of devices to be spawned.
2506          */
2507         list = rte_zmalloc("device spawn data",
2508                          sizeof(struct mlx5_dev_spawn_data) *
2509                          (np ? np : nd),
2510                          RTE_CACHE_LINE_SIZE);
2511         if (!list) {
2512                 DRV_LOG(ERR, "spawn data array allocation failure");
2513                 rte_errno = ENOMEM;
2514                 ret = -rte_errno;
2515                 goto exit;
2516         }
2517         if (bd >= 0 || np > 1) {
2518                 /*
2519                  * Single IB device with multiple ports found,
2520                  * it may be E-Switch master device and representors.
2521                  * We have to perform identification trough the ports.
2522                  */
2523                 assert(nl_rdma >= 0);
2524                 assert(ns == 0);
2525                 assert(nd == 1);
2526                 assert(np);
2527                 for (i = 1; i <= np; ++i) {
2528                         list[ns].max_port = np;
2529                         list[ns].ibv_port = i;
2530                         list[ns].ibv_dev = ibv_match[0];
2531                         list[ns].eth_dev = NULL;
2532                         list[ns].pci_dev = pci_dev;
2533                         list[ns].pf_bond = bd;
2534                         list[ns].ifindex = mlx5_nl_ifindex
2535                                         (nl_rdma, list[ns].ibv_dev->name, i);
2536                         if (!list[ns].ifindex) {
2537                                 /*
2538                                  * No network interface index found for the
2539                                  * specified port, it means there is no
2540                                  * representor on this port. It's OK,
2541                                  * there can be disabled ports, for example
2542                                  * if sriov_numvfs < sriov_totalvfs.
2543                                  */
2544                                 continue;
2545                         }
2546                         ret = -1;
2547                         if (nl_route >= 0)
2548                                 ret = mlx5_nl_switch_info
2549                                                (nl_route,
2550                                                 list[ns].ifindex,
2551                                                 &list[ns].info);
2552                         if (ret || (!list[ns].info.representor &&
2553                                     !list[ns].info.master)) {
2554                                 /*
2555                                  * We failed to recognize representors with
2556                                  * Netlink, let's try to perform the task
2557                                  * with sysfs.
2558                                  */
2559                                 ret =  mlx5_sysfs_switch_info
2560                                                 (list[ns].ifindex,
2561                                                  &list[ns].info);
2562                         }
2563                         if (!ret && bd >= 0) {
2564                                 switch (list[ns].info.name_type) {
2565                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2566                                         if (list[ns].info.port_name == bd)
2567                                                 ns++;
2568                                         break;
2569                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2570                                         if (list[ns].info.pf_num == bd)
2571                                                 ns++;
2572                                         break;
2573                                 default:
2574                                         break;
2575                                 }
2576                                 continue;
2577                         }
2578                         if (!ret && (list[ns].info.representor ^
2579                                      list[ns].info.master))
2580                                 ns++;
2581                 }
2582                 if (!ns) {
2583                         DRV_LOG(ERR,
2584                                 "unable to recognize master/representors"
2585                                 " on the IB device with multiple ports");
2586                         rte_errno = ENOENT;
2587                         ret = -rte_errno;
2588                         goto exit;
2589                 }
2590         } else {
2591                 /*
2592                  * The existence of several matching entries (nd > 1) means
2593                  * port representors have been instantiated. No existing Verbs
2594                  * call nor sysfs entries can tell them apart, this can only
2595                  * be done through Netlink calls assuming kernel drivers are
2596                  * recent enough to support them.
2597                  *
2598                  * In the event of identification failure through Netlink,
2599                  * try again through sysfs, then:
2600                  *
2601                  * 1. A single IB device matches (nd == 1) with single
2602                  *    port (np=0/1) and is not a representor, assume
2603                  *    no switch support.
2604                  *
2605                  * 2. Otherwise no safe assumptions can be made;
2606                  *    complain louder and bail out.
2607                  */
2608                 np = 1;
2609                 for (i = 0; i != nd; ++i) {
2610                         memset(&list[ns].info, 0, sizeof(list[ns].info));
2611                         list[ns].max_port = 1;
2612                         list[ns].ibv_port = 1;
2613                         list[ns].ibv_dev = ibv_match[i];
2614                         list[ns].eth_dev = NULL;
2615                         list[ns].pci_dev = pci_dev;
2616                         list[ns].pf_bond = -1;
2617                         list[ns].ifindex = 0;
2618                         if (nl_rdma >= 0)
2619                                 list[ns].ifindex = mlx5_nl_ifindex
2620                                         (nl_rdma, list[ns].ibv_dev->name, 1);
2621                         if (!list[ns].ifindex) {
2622                                 char ifname[IF_NAMESIZE];
2623
2624                                 /*
2625                                  * Netlink failed, it may happen with old
2626                                  * ib_core kernel driver (before 4.16).
2627                                  * We can assume there is old driver because
2628                                  * here we are processing single ports IB
2629                                  * devices. Let's try sysfs to retrieve
2630                                  * the ifindex. The method works for
2631                                  * master device only.
2632                                  */
2633                                 if (nd > 1) {
2634                                         /*
2635                                          * Multiple devices found, assume
2636                                          * representors, can not distinguish
2637                                          * master/representor and retrieve
2638                                          * ifindex via sysfs.
2639                                          */
2640                                         continue;
2641                                 }
2642                                 ret = mlx5_get_master_ifname
2643                                         (ibv_match[i]->ibdev_path, &ifname);
2644                                 if (!ret)
2645                                         list[ns].ifindex =
2646                                                 if_nametoindex(ifname);
2647                                 if (!list[ns].ifindex) {
2648                                         /*
2649                                          * No network interface index found
2650                                          * for the specified device, it means
2651                                          * there it is neither representor
2652                                          * nor master.
2653                                          */
2654                                         continue;
2655                                 }
2656                         }
2657                         ret = -1;
2658                         if (nl_route >= 0)
2659                                 ret = mlx5_nl_switch_info
2660                                                (nl_route,
2661                                                 list[ns].ifindex,
2662                                                 &list[ns].info);
2663                         if (ret || (!list[ns].info.representor &&
2664                                     !list[ns].info.master)) {
2665                                 /*
2666                                  * We failed to recognize representors with
2667                                  * Netlink, let's try to perform the task
2668                                  * with sysfs.
2669                                  */
2670                                 ret =  mlx5_sysfs_switch_info
2671                                                 (list[ns].ifindex,
2672                                                  &list[ns].info);
2673                         }
2674                         if (!ret && (list[ns].info.representor ^
2675                                      list[ns].info.master)) {
2676                                 ns++;
2677                         } else if ((nd == 1) &&
2678                                    !list[ns].info.representor &&
2679                                    !list[ns].info.master) {
2680                                 /*
2681                                  * Single IB device with
2682                                  * one physical port and
2683                                  * attached network device.
2684                                  * May be SRIOV is not enabled
2685                                  * or there is no representors.
2686                                  */
2687                                 DRV_LOG(INFO, "no E-Switch support detected");
2688                                 ns++;
2689                                 break;
2690                         }
2691                 }
2692                 if (!ns) {
2693                         DRV_LOG(ERR,
2694                                 "unable to recognize master/representors"
2695                                 " on the multiple IB devices");
2696                         rte_errno = ENOENT;
2697                         ret = -rte_errno;
2698                         goto exit;
2699                 }
2700         }
2701         assert(ns);
2702         /*
2703          * Sort list to probe devices in natural order for users convenience
2704          * (i.e. master first, then representors from lowest to highest ID).
2705          */
2706         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2707         /* Default configuration. */
2708         dev_config = (struct mlx5_dev_config){
2709                 .hw_padding = 0,
2710                 .mps = MLX5_ARG_UNSET,
2711                 .rx_vec_en = 1,
2712                 .txq_inline_max = MLX5_ARG_UNSET,
2713                 .txq_inline_min = MLX5_ARG_UNSET,
2714                 .txq_inline_mpw = MLX5_ARG_UNSET,
2715                 .txqs_inline = MLX5_ARG_UNSET,
2716                 .vf_nl_en = 1,
2717                 .mr_ext_memseg_en = 1,
2718                 .mprq = {
2719                         .enabled = 0, /* Disabled by default. */
2720                         .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
2721                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
2722                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
2723                 },
2724                 .dv_esw_en = 1,
2725         };
2726         /* Device specific configuration. */
2727         switch (pci_dev->id.device_id) {
2728         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2729         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2730         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2731         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2732         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2733         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2734                 dev_config.vf = 1;
2735                 break;
2736         default:
2737                 break;
2738         }
2739         for (i = 0; i != ns; ++i) {
2740                 uint32_t restore;
2741
2742                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2743                                                  &list[i],
2744                                                  dev_config);
2745                 if (!list[i].eth_dev) {
2746                         if (rte_errno != EBUSY && rte_errno != EEXIST)
2747                                 break;
2748                         /* Device is disabled or already spawned. Ignore it. */
2749                         continue;
2750                 }
2751                 restore = list[i].eth_dev->data->dev_flags;
2752                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2753                 /* Restore non-PCI flags cleared by the above call. */
2754                 list[i].eth_dev->data->dev_flags |= restore;
2755                 mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
2756                 rte_eth_dev_probing_finish(list[i].eth_dev);
2757         }
2758         if (i != ns) {
2759                 DRV_LOG(ERR,
2760                         "probe of PCI device " PCI_PRI_FMT " aborted after"
2761                         " encountering an error: %s",
2762                         pci_dev->addr.domain, pci_dev->addr.bus,
2763                         pci_dev->addr.devid, pci_dev->addr.function,
2764                         strerror(rte_errno));
2765                 ret = -rte_errno;
2766                 /* Roll back. */
2767                 while (i--) {
2768                         if (!list[i].eth_dev)
2769                                 continue;
2770                         mlx5_dev_close(list[i].eth_dev);
2771                         /* mac_addrs must not be freed because in dev_private */
2772                         list[i].eth_dev->data->mac_addrs = NULL;
2773                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2774                 }
2775                 /* Restore original error. */
2776                 rte_errno = -ret;
2777         } else {
2778                 ret = 0;
2779         }
2780 exit:
2781         /*
2782          * Do the routine cleanup:
2783          * - close opened Netlink sockets
2784          * - free allocated spawn data array
2785          * - free the Infiniband device list
2786          */
2787         if (nl_rdma >= 0)
2788                 close(nl_rdma);
2789         if (nl_route >= 0)
2790                 close(nl_route);
2791         if (list)
2792                 rte_free(list);
2793         assert(ibv_list);
2794         mlx5_glue->free_device_list(ibv_list);
2795         return ret;
2796 }
2797
2798 /**
2799  * Look for the ethernet device belonging to mlx5 driver.
2800  *
2801  * @param[in] port_id
2802  *   port_id to start looking for device.
2803  * @param[in] pci_dev
2804  *   Pointer to the hint PCI device. When device is being probed
2805  *   the its siblings (master and preceding representors might
2806  *   not have assigned driver yet (because the mlx5_pci_probe()
2807  *   is not completed yet, for this case match on hint PCI
2808  *   device may be used to detect sibling device.
2809  *
2810  * @return
2811  *   port_id of found device, RTE_MAX_ETHPORT if not found.
2812  */
2813 uint16_t
2814 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
2815 {
2816         while (port_id < RTE_MAX_ETHPORTS) {
2817                 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2818
2819                 if (dev->state != RTE_ETH_DEV_UNUSED &&
2820                     dev->device &&
2821                     (dev->device == &pci_dev->device ||
2822                      (dev->device->driver &&
2823                      dev->device->driver->name &&
2824                      !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
2825                         break;
2826                 port_id++;
2827         }
2828         if (port_id >= RTE_MAX_ETHPORTS)
2829                 return RTE_MAX_ETHPORTS;
2830         return port_id;
2831 }
2832
2833 /**
2834  * DPDK callback to remove a PCI device.
2835  *
2836  * This function removes all Ethernet devices belong to a given PCI device.
2837  *
2838  * @param[in] pci_dev
2839  *   Pointer to the PCI device.
2840  *
2841  * @return
2842  *   0 on success, the function cannot fail.
2843  */
2844 static int
2845 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2846 {
2847         uint16_t port_id;
2848
2849         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
2850                 rte_eth_dev_close(port_id);
2851         return 0;
2852 }
2853
2854 static const struct rte_pci_id mlx5_pci_id_map[] = {
2855         {
2856                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2857                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2858         },
2859         {
2860                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2861                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2862         },
2863         {
2864                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2865                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2866         },
2867         {
2868                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2869                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2870         },
2871         {
2872                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2873                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2874         },
2875         {
2876                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2877                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2878         },
2879         {
2880                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2881                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2882         },
2883         {
2884                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2885                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2886         },
2887         {
2888                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2889                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2890         },
2891         {
2892                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2893                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2894         },
2895         {
2896                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2897                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2898         },
2899         {
2900                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2901                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2902         },
2903         {
2904                 .vendor_id = 0
2905         }
2906 };
2907
2908 static struct rte_pci_driver mlx5_driver = {
2909         .driver = {
2910                 .name = MLX5_DRIVER_NAME
2911         },
2912         .id_table = mlx5_pci_id_map,
2913         .probe = mlx5_pci_probe,
2914         .remove = mlx5_pci_remove,
2915         .dma_map = mlx5_dma_map,
2916         .dma_unmap = mlx5_dma_unmap,
2917         .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
2918                      RTE_PCI_DRV_PROBE_AGAIN,
2919 };
2920
2921 #ifdef RTE_IBVERBS_LINK_DLOPEN
2922
2923 /**
2924  * Suffix RTE_EAL_PMD_PATH with "-glue".
2925  *
2926  * This function performs a sanity check on RTE_EAL_PMD_PATH before
2927  * suffixing its last component.
2928  *
2929  * @param buf[out]
2930  *   Output buffer, should be large enough otherwise NULL is returned.
2931  * @param size
2932  *   Size of @p out.
2933  *
2934  * @return
2935  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
2936  */
2937 static char *
2938 mlx5_glue_path(char *buf, size_t size)
2939 {
2940         static const char *const bad[] = { "/", ".", "..", NULL };
2941         const char *path = RTE_EAL_PMD_PATH;
2942         size_t len = strlen(path);
2943         size_t off;
2944         int i;
2945
2946         while (len && path[len - 1] == '/')
2947                 --len;
2948         for (off = len; off && path[off - 1] != '/'; --off)
2949                 ;
2950         for (i = 0; bad[i]; ++i)
2951                 if (!strncmp(path + off, bad[i], (int)(len - off)))
2952                         goto error;
2953         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
2954         if (i == -1 || (size_t)i >= size)
2955                 goto error;
2956         return buf;
2957 error:
2958         DRV_LOG(ERR,
2959                 "unable to append \"-glue\" to last component of"
2960                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
2961                 " please re-configure DPDK");
2962         return NULL;
2963 }
2964
2965 /**
2966  * Initialization routine for run-time dependency on rdma-core.
2967  */
2968 static int
2969 mlx5_glue_init(void)
2970 {
2971         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
2972         const char *path[] = {
2973                 /*
2974                  * A basic security check is necessary before trusting
2975                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
2976                  */
2977                 (geteuid() == getuid() && getegid() == getgid() ?
2978                  getenv("MLX5_GLUE_PATH") : NULL),
2979                 /*
2980                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
2981                  * variant, otherwise let dlopen() look up libraries on its
2982                  * own.
2983                  */
2984                 (*RTE_EAL_PMD_PATH ?
2985                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
2986         };
2987         unsigned int i = 0;
2988         void *handle = NULL;
2989         void **sym;
2990         const char *dlmsg;
2991
2992         while (!handle && i != RTE_DIM(path)) {
2993                 const char *end;
2994                 size_t len;
2995                 int ret;
2996
2997                 if (!path[i]) {
2998                         ++i;
2999                         continue;
3000                 }
3001                 end = strpbrk(path[i], ":;");
3002                 if (!end)
3003                         end = path[i] + strlen(path[i]);
3004                 len = end - path[i];
3005                 ret = 0;
3006                 do {
3007                         char name[ret + 1];
3008
3009                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
3010                                        (int)len, path[i],
3011                                        (!len || *(end - 1) == '/') ? "" : "/");
3012                         if (ret == -1)
3013                                 break;
3014                         if (sizeof(name) != (size_t)ret + 1)
3015                                 continue;
3016                         DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
3017                                 name);
3018                         handle = dlopen(name, RTLD_LAZY);
3019                         break;
3020                 } while (1);
3021                 path[i] = end + 1;
3022                 if (!*end)
3023                         ++i;
3024         }
3025         if (!handle) {
3026                 rte_errno = EINVAL;
3027                 dlmsg = dlerror();
3028                 if (dlmsg)
3029                         DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
3030                 goto glue_error;
3031         }
3032         sym = dlsym(handle, "mlx5_glue");
3033         if (!sym || !*sym) {
3034                 rte_errno = EINVAL;
3035                 dlmsg = dlerror();
3036                 if (dlmsg)
3037                         DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
3038                 goto glue_error;
3039         }
3040         mlx5_glue = *sym;
3041         return 0;
3042 glue_error:
3043         if (handle)
3044                 dlclose(handle);
3045         DRV_LOG(WARNING,
3046                 "cannot initialize PMD due to missing run-time dependency on"
3047                 " rdma-core libraries (libibverbs, libmlx5)");
3048         return -rte_errno;
3049 }
3050
3051 #endif
3052
3053 /**
3054  * Driver initialization routine.
3055  */
3056 RTE_INIT(rte_mlx5_pmd_init)
3057 {
3058         /* Initialize driver log type. */
3059         mlx5_logtype = rte_log_register("pmd.net.mlx5");
3060         if (mlx5_logtype >= 0)
3061                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3062
3063         /* Build the static tables for Verbs conversion. */
3064         mlx5_set_ptype_table();
3065         mlx5_set_cksum_table();
3066         mlx5_set_swp_types_table();
3067         /*
3068          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
3069          * huge pages. Calling ibv_fork_init() during init allows
3070          * applications to use fork() safely for purposes other than
3071          * using this PMD, which is not supported in forked processes.
3072          */
3073         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
3074         /* Match the size of Rx completion entry to the size of a cacheline. */
3075         if (RTE_CACHE_LINE_SIZE == 128)
3076                 setenv("MLX5_CQE_SIZE", "128", 0);
3077         /*
3078          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
3079          * cleanup all the Verbs resources even when the device was removed.
3080          */
3081         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
3082 #ifdef RTE_IBVERBS_LINK_DLOPEN
3083         if (mlx5_glue_init())
3084                 return;
3085         assert(mlx5_glue);
3086 #endif
3087 #ifndef NDEBUG
3088         /* Glue structure must not contain any NULL pointers. */
3089         {
3090                 unsigned int i;
3091
3092                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
3093                         assert(((const void *const *)mlx5_glue)[i]);
3094         }
3095 #endif
3096         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
3097                 DRV_LOG(ERR,
3098                         "rdma-core glue \"%s\" mismatch: \"%s\" is required",
3099                         mlx5_glue->version, MLX5_GLUE_VERSION);
3100                 return;
3101         }
3102         mlx5_glue->fork_init();
3103         rte_pci_register(&mlx5_driver);
3104 }
3105
3106 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3107 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3108 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");