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_eal_memconfig.h>
  36 #include <rte_kvargs.h>
  37 #include <rte_rwlock.h>
  38 #include <rte_spinlock.h>
  39 #include <rte_string_fns.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. */
  72 #define MLX5_TXQ_INLINE "txq_inline"
  73
  74 /*
  75  * Device parameter to configure the number of TX queues threshold for
  76  * enabling inline send.
  77  */
  78 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
  79
  80 /*
  81  * Device parameter to configure the number of TX queues threshold for
  82  * enabling vectorized Tx.
  83  */
  84 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
  85
  86 /* Device parameter to enable multi-packet send WQEs. */
  87 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
  88
  89 /* Device parameter to include 2 dsegs in the title WQEBB. */
  90 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
  91
  92 /* Device parameter to limit the size of inlining packet. */
  93 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
  94
  95 /* Device parameter to enable hardware Tx vector. */
  96 #define MLX5_TX_VEC_EN "tx_vec_en"
  97
  98 /* Device parameter to enable hardware Rx vector. */
  99 #define MLX5_RX_VEC_EN "rx_vec_en"
 100
 101 /* Allow L3 VXLAN flow creation. */
 102 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
 103
 104 /* Activate DV E-Switch flow steering. */
 105 #define MLX5_DV_ESW_EN "dv_esw_en"
 106
 107 /* Activate DV flow steering. */
 108 #define MLX5_DV_FLOW_EN "dv_flow_en"
 109
 110 /* Activate Netlink support in VF mode. */
 111 #define MLX5_VF_NL_EN "vf_nl_en"
 112
 113 /* Enable extending memsegs when creating a MR. */
 114 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
 115
 116 /* Select port representors to instantiate. */
 117 #define MLX5_REPRESENTOR "representor"
 118
 119 #ifndef HAVE_IBV_MLX5_MOD_MPW
 120 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
 121 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
 122 #endif
 123
 124 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
 125 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
 126 #endif
 127
 128 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
 129
 130 /* Shared memory between primary and secondary processes. */
 131 struct mlx5_shared_data *mlx5_shared_data;
 132
 133 /* Spinlock for mlx5_shared_data allocation. */
 134 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
 135
 136 /* Process local data for secondary processes. */
 137 static struct mlx5_local_data mlx5_local_data;
 138
 139 /** Driver-specific log messages type. */
 140 int mlx5_logtype;
 141
 142 /** Data associated with devices to spawn. */
 143 struct mlx5_dev_spawn_data {
 144         uint32_t ifindex; /**< Network interface index. */
 145         uint32_t max_port; /**< IB device maximal port index. */
 146         uint32_t ibv_port; /**< IB device physical port index. */
 147         struct mlx5_switch_info info; /**< Switch information. */
 148         struct ibv_device *ibv_dev; /**< Associated IB device. */
 149         struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
 150         struct rte_pci_device *pci_dev; /**< Backend PCI device. */
 151 };
 152
 153 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
 154 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
 155
 156 /**
 157  * Allocate shared IB device context. If there is multiport device the
 158  * master and representors will share this context, if there is single
 159  * port dedicated IB device, the context will be used by only given
 160  * port due to unification.
 161  *
 162  * Routine first searches the context for the specified IB device name,
 163  * if found the shared context assumed and reference counter is incremented.
 164  * If no context found the new one is created and initialized with specified
 165  * IB device context and parameters.
 166  *
 167  * @param[in] spawn
 168  *   Pointer to the IB device attributes (name, port, etc).
 169  *
 170  * @return
 171  *   Pointer to mlx5_ibv_shared object on success,
 172  *   otherwise NULL and rte_errno is set.
 173  */
 174 static struct mlx5_ibv_shared *
 175 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn)
 176 {
 177         struct mlx5_ibv_shared *sh;
 178         int err = 0;
 179         uint32_t i;
 180
 181         assert(spawn);
 182         /* Secondary process should not create the shared context. */
 183         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 184         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 185         /* Search for IB context by device name. */
 186         LIST_FOREACH(sh, &mlx5_ibv_list, next) {
 187                 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
 188                         sh->refcnt++;
 189                         goto exit;
 190                 }
 191         }
 192         /* No device found, we have to create new shared context. */
 193         assert(spawn->max_port);
 194         sh = rte_zmalloc("ethdev shared ib context",
 195                          sizeof(struct mlx5_ibv_shared) +
 196                          spawn->max_port *
 197                          sizeof(struct mlx5_ibv_shared_port),
 198                          RTE_CACHE_LINE_SIZE);
 199         if (!sh) {
 200                 DRV_LOG(ERR, "shared context allocation failure");
 201                 rte_errno  = ENOMEM;
 202                 goto exit;
 203         }
 204         /* Try to open IB device with DV first, then usual Verbs. */
 205         errno = 0;
 206         sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
 207         if (sh->ctx) {
 208                 sh->devx = 1;
 209                 DRV_LOG(DEBUG, "DevX is supported");
 210         } else {
 211                 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
 212                 if (!sh->ctx) {
 213                         err = errno ? errno : ENODEV;
 214                         goto error;
 215                 }
 216                 DRV_LOG(DEBUG, "DevX is NOT supported");
 217         }
 218         err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
 219         if (err) {
 220                 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
 221                 goto error;
 222         }
 223         sh->refcnt = 1;
 224         sh->max_port = spawn->max_port;
 225         strncpy(sh->ibdev_name, sh->ctx->device->name,
 226                 sizeof(sh->ibdev_name));
 227         strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
 228                 sizeof(sh->ibdev_path));
 229         sh->pci_dev = spawn->pci_dev;
 230         pthread_mutex_init(&sh->intr_mutex, NULL);
 231         /*
 232          * Setting port_id to max unallowed value means
 233          * there is no interrupt subhandler installed for
 234          * the given port index i.
 235          */
 236         for (i = 0; i < sh->max_port; i++)
 237                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
 238         sh->pd = mlx5_glue->alloc_pd(sh->ctx);
 239         if (sh->pd == NULL) {
 240                 DRV_LOG(ERR, "PD allocation failure");
 241                 err = ENOMEM;
 242                 goto error;
 243         }
 244         /*
 245          * Once the device is added to the list of memory event
 246          * callback, its global MR cache table cannot be expanded
 247          * on the fly because of deadlock. If it overflows, lookup
 248          * should be done by searching MR list linearly, which is slow.
 249          *
 250          * At this point the device is not added to the memory
 251          * event list yet, context is just being created.
 252          */
 253         err = mlx5_mr_btree_init(&sh->mr.cache,
 254                                  MLX5_MR_BTREE_CACHE_N * 2,
 255                                  sh->pci_dev->device.numa_node);
 256         if (err) {
 257                 err = rte_errno;
 258                 goto error;
 259         }
 260         LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
 261 exit:
 262         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 263         return sh;
 264 error:
 265         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 266         assert(sh);
 267         if (sh->pd)
 268                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
 269         if (sh->ctx)
 270                 claim_zero(mlx5_glue->close_device(sh->ctx));
 271         rte_free(sh);
 272         assert(err > 0);
 273         rte_errno = err;
 274         return NULL;
 275 }
 276
 277 /**
 278  * Free shared IB device context. Decrement counter and if zero free
 279  * all allocated resources and close handles.
 280  *
 281  * @param[in] sh
 282  *   Pointer to mlx5_ibv_shared object to free
 283  */
 284 static void
 285 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
 286 {
 287         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 288 #ifndef NDEBUG
 289         /* Check the object presence in the list. */
 290         struct mlx5_ibv_shared *lctx;
 291
 292         LIST_FOREACH(lctx, &mlx5_ibv_list, next)
 293                 if (lctx == sh)
 294                         break;
 295         assert(lctx);
 296         if (lctx != sh) {
 297                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
 298                 goto exit;
 299         }
 300 #endif
 301         assert(sh);
 302         assert(sh->refcnt);
 303         /* Secondary process should not free the shared context. */
 304         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 305         if (--sh->refcnt)
 306                 goto exit;
 307         /* Release created Memory Regions. */
 308         mlx5_mr_release(sh);
 309         LIST_REMOVE(sh, next);
 310         /*
 311          *  Ensure there is no async event handler installed.
 312          *  Only primary process handles async device events.
 313          **/
 314         assert(!sh->intr_cnt);
 315         if (sh->intr_cnt)
 316                 rte_intr_callback_unregister
 317                         (&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
 318         pthread_mutex_destroy(&sh->intr_mutex);
 319         if (sh->pd)
 320                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
 321         if (sh->ctx)
 322                 claim_zero(mlx5_glue->close_device(sh->ctx));
 323         rte_free(sh);
 324 exit:
 325         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 326 }
 327
 328 /**
 329  * Initialize DR related data within private structure.
 330  * Routine checks the reference counter and does actual
 331  * resources creation/initialization only if counter is zero.
 332  *
 333  * @param[in] priv
 334  *   Pointer to the private device data structure.
 335  *
 336  * @return
 337  *   Zero on success, positive error code otherwise.
 338  */
 339 static int
 340 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 341 {
 342 #ifdef HAVE_MLX5DV_DR
 343         struct mlx5_ibv_shared *sh = priv->sh;
 344         int err = 0;
 345         void *domain;
 346
 347         assert(sh);
 348         if (sh->dv_refcnt) {
 349                 /* Shared DV/DR structures is already initialized. */
 350                 sh->dv_refcnt++;
 351                 priv->dr_shared = 1;
 352                 return 0;
 353         }
 354         /* Reference counter is zero, we should initialize structures. */
 355         domain = mlx5_glue->dr_create_domain(sh->ctx,
 356                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 357         if (!domain) {
 358                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 359                 err = errno;
 360                 goto error;
 361         }
 362         sh->rx_domain = domain;
 363         domain = mlx5_glue->dr_create_domain(sh->ctx,
 364                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 365         if (!domain) {
 366                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 367                 err = errno;
 368                 goto error;
 369         }
 370         pthread_mutex_init(&sh->dv_mutex, NULL);
 371         sh->tx_domain = domain;
 372 #ifdef HAVE_MLX5DV_DR_ESWITCH
 373         if (priv->config.dv_esw_en) {
 374                 domain  = mlx5_glue->dr_create_domain
 375                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 376                 if (!domain) {
 377                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 378                         err = errno;
 379                         goto error;
 380                 }
 381                 sh->fdb_domain = domain;
 382                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 383         }
 384 #endif
 385         sh->dv_refcnt++;
 386         priv->dr_shared = 1;
 387         return 0;
 388
 389 error:
 390        /* Rollback the created objects. */
 391         if (sh->rx_domain) {
 392                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 393                 sh->rx_domain = NULL;
 394         }
 395         if (sh->tx_domain) {
 396                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 397                 sh->tx_domain = NULL;
 398         }
 399         if (sh->fdb_domain) {
 400                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 401                 sh->fdb_domain = NULL;
 402         }
 403         if (sh->esw_drop_action) {
 404                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 405                 sh->esw_drop_action = NULL;
 406         }
 407         return err;
 408 #else
 409         (void)priv;
 410         return 0;
 411 #endif
 412 }
 413
 414 /**
 415  * Destroy DR related data within private structure.
 416  *
 417  * @param[in] priv
 418  *   Pointer to the private device data structure.
 419  */
 420 static void
 421 mlx5_free_shared_dr(struct mlx5_priv *priv)
 422 {
 423 #ifdef HAVE_MLX5DV_DR
 424         struct mlx5_ibv_shared *sh;
 425
 426         if (!priv->dr_shared)
 427                 return;
 428         priv->dr_shared = 0;
 429         sh = priv->sh;
 430         assert(sh);
 431         assert(sh->dv_refcnt);
 432         if (sh->dv_refcnt && --sh->dv_refcnt)
 433                 return;
 434         if (sh->rx_domain) {
 435                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 436                 sh->rx_domain = NULL;
 437         }
 438         if (sh->tx_domain) {
 439                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 440                 sh->tx_domain = NULL;
 441         }
 442 #ifdef HAVE_MLX5DV_DR_ESWITCH
 443         if (sh->fdb_domain) {
 444                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 445                 sh->fdb_domain = NULL;
 446         }
 447         if (sh->esw_drop_action) {
 448                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 449                 sh->esw_drop_action = NULL;
 450         }
 451 #endif
 452         pthread_mutex_destroy(&sh->dv_mutex);
 453 #else
 454         (void)priv;
 455 #endif
 456 }
 457
 458 /**
 459  * Initialize shared data between primary and secondary process.
 460  *
 461  * A memzone is reserved by primary process and secondary processes attach to
 462  * the memzone.
 463  *
 464  * @return
 465  *   0 on success, a negative errno value otherwise and rte_errno is set.
 466  */
 467 static int
 468 mlx5_init_shared_data(void)
 469 {
 470         const struct rte_memzone *mz;
 471         int ret = 0;
 472
 473         rte_spinlock_lock(&mlx5_shared_data_lock);
 474         if (mlx5_shared_data == NULL) {
 475                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 476                         /* Allocate shared memory. */
 477                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
 478                                                  sizeof(*mlx5_shared_data),
 479                                                  SOCKET_ID_ANY, 0);
 480                         if (mz == NULL) {
 481                                 DRV_LOG(ERR,
 482                                         "Cannot allocate mlx5 shared data\n");
 483                                 ret = -rte_errno;
 484                                 goto error;
 485                         }
 486                         mlx5_shared_data = mz->addr;
 487                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
 488                         rte_spinlock_init(&mlx5_shared_data->lock);
 489                 } else {
 490                         /* Lookup allocated shared memory. */
 491                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
 492                         if (mz == NULL) {
 493                                 DRV_LOG(ERR,
 494                                         "Cannot attach mlx5 shared data\n");
 495                                 ret = -rte_errno;
 496                                 goto error;
 497                         }
 498                         mlx5_shared_data = mz->addr;
 499                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
 500                 }
 501         }
 502 error:
 503         rte_spinlock_unlock(&mlx5_shared_data_lock);
 504         return ret;
 505 }
 506
 507 /**
 508  * Retrieve integer value from environment variable.
 509  *
 510  * @param[in] name
 511  *   Environment variable name.
 512  *
 513  * @return
 514  *   Integer value, 0 if the variable is not set.
 515  */
 516 int
 517 mlx5_getenv_int(const char *name)
 518 {
 519         const char *val = getenv(name);
 520
 521         if (val == NULL)
 522                 return 0;
 523         return atoi(val);
 524 }
 525
 526 /**
 527  * Verbs callback to allocate a memory. This function should allocate the space
 528  * according to the size provided residing inside a huge page.
 529  * Please note that all allocation must respect the alignment from libmlx5
 530  * (i.e. currently sysconf(_SC_PAGESIZE)).
 531  *
 532  * @param[in] size
 533  *   The size in bytes of the memory to allocate.
 534  * @param[in] data
 535  *   A pointer to the callback data.
 536  *
 537  * @return
 538  *   Allocated buffer, NULL otherwise and rte_errno is set.
 539  */
 540 static void *
 541 mlx5_alloc_verbs_buf(size_t size, void *data)
 542 {
 543         struct mlx5_priv *priv = data;
 544         void *ret;
 545         size_t alignment = sysconf(_SC_PAGESIZE);
 546         unsigned int socket = SOCKET_ID_ANY;
 547
 548         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 549                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 550
 551                 socket = ctrl->socket;
 552         } else if (priv->verbs_alloc_ctx.type ==
 553                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 554                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 555
 556                 socket = ctrl->socket;
 557         }
 558         assert(data != NULL);
 559         ret = rte_malloc_socket(__func__, size, alignment, socket);
 560         if (!ret && size)
 561                 rte_errno = ENOMEM;
 562         return ret;
 563 }
 564
 565 /**
 566  * Verbs callback to free a memory.
 567  *
 568  * @param[in] ptr
 569  *   A pointer to the memory to free.
 570  * @param[in] data
 571  *   A pointer to the callback data.
 572  */
 573 static void
 574 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 575 {
 576         assert(data != NULL);
 577         rte_free(ptr);
 578 }
 579
 580 /**
 581  * Initialize process private data structure.
 582  *
 583  * @param dev
 584  *   Pointer to Ethernet device structure.
 585  *
 586  * @return
 587  *   0 on success, a negative errno value otherwise and rte_errno is set.
 588  */
 589 int
 590 mlx5_proc_priv_init(struct rte_eth_dev *dev)
 591 {
 592         struct mlx5_priv *priv = dev->data->dev_private;
 593         struct mlx5_proc_priv *ppriv;
 594         size_t ppriv_size;
 595
 596         /*
 597          * UAR register table follows the process private structure. BlueFlame
 598          * registers for Tx queues are stored in the table.
 599          */
 600         ppriv_size =
 601                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
 602         ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
 603                                   RTE_CACHE_LINE_SIZE, dev->device->numa_node);
 604         if (!ppriv) {
 605                 rte_errno = ENOMEM;
 606                 return -rte_errno;
 607         }
 608         ppriv->uar_table_sz = ppriv_size;
 609         dev->process_private = ppriv;
 610         return 0;
 611 }
 612
 613 /**
 614  * Un-initialize process private data structure.
 615  *
 616  * @param dev
 617  *   Pointer to Ethernet device structure.
 618  */
 619 static void
 620 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
 621 {
 622         if (!dev->process_private)
 623                 return;
 624         rte_free(dev->process_private);
 625         dev->process_private = NULL;
 626 }
 627
 628 /**
 629  * DPDK callback to close the device.
 630  *
 631  * Destroy all queues and objects, free memory.
 632  *
 633  * @param dev
 634  *   Pointer to Ethernet device structure.
 635  */
 636 static void
 637 mlx5_dev_close(struct rte_eth_dev *dev)
 638 {
 639         struct mlx5_priv *priv = dev->data->dev_private;
 640         unsigned int i;
 641         int ret;
 642
 643         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
 644                 dev->data->port_id,
 645                 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
 646         /* In case mlx5_dev_stop() has not been called. */
 647         mlx5_dev_interrupt_handler_uninstall(dev);
 648         mlx5_traffic_disable(dev);
 649         mlx5_flow_flush(dev, NULL);
 650         /* Prevent crashes when queues are still in use. */
 651         dev->rx_pkt_burst = removed_rx_burst;
 652         dev->tx_pkt_burst = removed_tx_burst;
 653         rte_wmb();
 654         /* Disable datapath on secondary process. */
 655         mlx5_mp_req_stop_rxtx(dev);
 656         if (priv->rxqs != NULL) {
 657                 /* XXX race condition if mlx5_rx_burst() is still running. */
 658                 usleep(1000);
 659                 for (i = 0; (i != priv->rxqs_n); ++i)
 660                         mlx5_rxq_release(dev, i);
 661                 priv->rxqs_n = 0;
 662                 priv->rxqs = NULL;
 663         }
 664         if (priv->txqs != NULL) {
 665                 /* XXX race condition if mlx5_tx_burst() is still running. */
 666                 usleep(1000);
 667                 for (i = 0; (i != priv->txqs_n); ++i)
 668                         mlx5_txq_release(dev, i);
 669                 priv->txqs_n = 0;
 670                 priv->txqs = NULL;
 671         }
 672         mlx5_proc_priv_uninit(dev);
 673         mlx5_mprq_free_mp(dev);
 674         /* Remove from memory callback device list. */
 675         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
 676         assert(priv->sh);
 677         LIST_REMOVE(priv->sh, mem_event_cb);
 678         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
 679         mlx5_free_shared_dr(priv);
 680         if (priv->rss_conf.rss_key != NULL)
 681                 rte_free(priv->rss_conf.rss_key);
 682         if (priv->reta_idx != NULL)
 683                 rte_free(priv->reta_idx);
 684         if (priv->config.vf)
 685                 mlx5_nl_mac_addr_flush(dev);
 686         if (priv->nl_socket_route >= 0)
 687                 close(priv->nl_socket_route);
 688         if (priv->nl_socket_rdma >= 0)
 689                 close(priv->nl_socket_rdma);
 690         if (priv->tcf_context)
 691                 mlx5_flow_tcf_context_destroy(priv->tcf_context);
 692         if (priv->sh) {
 693                 /*
 694                  * Free the shared context in last turn, because the cleanup
 695                  * routines above may use some shared fields, like
 696                  * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
 697                  * ifindex if Netlink fails.
 698                  */
 699                 mlx5_free_shared_ibctx(priv->sh);
 700                 priv->sh = NULL;
 701         }
 702         ret = mlx5_hrxq_ibv_verify(dev);
 703         if (ret)
 704                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
 705                         dev->data->port_id);
 706         ret = mlx5_ind_table_ibv_verify(dev);
 707         if (ret)
 708                 DRV_LOG(WARNING, "port %u some indirection table still remain",
 709                         dev->data->port_id);
 710         ret = mlx5_rxq_ibv_verify(dev);
 711         if (ret)
 712                 DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
 713                         dev->data->port_id);
 714         ret = mlx5_rxq_verify(dev);
 715         if (ret)
 716                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
 717                         dev->data->port_id);
 718         ret = mlx5_txq_ibv_verify(dev);
 719         if (ret)
 720                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
 721                         dev->data->port_id);
 722         ret = mlx5_txq_verify(dev);
 723         if (ret)
 724                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
 725                         dev->data->port_id);
 726         ret = mlx5_flow_verify(dev);
 727         if (ret)
 728                 DRV_LOG(WARNING, "port %u some flows still remain",
 729                         dev->data->port_id);
 730         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 731                 unsigned int c = 0;
 732                 uint16_t port_id;
 733
 734                 RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
 735                         struct mlx5_priv *opriv =
 736                                 rte_eth_devices[port_id].data->dev_private;
 737
 738                         if (!opriv ||
 739                             opriv->domain_id != priv->domain_id ||
 740                             &rte_eth_devices[port_id] == dev)
 741                                 continue;
 742                         ++c;
 743                 }
 744                 if (!c)
 745                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
 746         }
 747         memset(priv, 0, sizeof(*priv));
 748         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 749         /*
 750          * Reset mac_addrs to NULL such that it is not freed as part of
 751          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
 752          * it is freed when dev_private is freed.
 753          */
 754         dev->data->mac_addrs = NULL;
 755 }
 756
 757 const struct eth_dev_ops mlx5_dev_ops = {
 758         .dev_configure = mlx5_dev_configure,
 759         .dev_start = mlx5_dev_start,
 760         .dev_stop = mlx5_dev_stop,
 761         .dev_set_link_down = mlx5_set_link_down,
 762         .dev_set_link_up = mlx5_set_link_up,
 763         .dev_close = mlx5_dev_close,
 764         .promiscuous_enable = mlx5_promiscuous_enable,
 765         .promiscuous_disable = mlx5_promiscuous_disable,
 766         .allmulticast_enable = mlx5_allmulticast_enable,
 767         .allmulticast_disable = mlx5_allmulticast_disable,
 768         .link_update = mlx5_link_update,
 769         .stats_get = mlx5_stats_get,
 770         .stats_reset = mlx5_stats_reset,
 771         .xstats_get = mlx5_xstats_get,
 772         .xstats_reset = mlx5_xstats_reset,
 773         .xstats_get_names = mlx5_xstats_get_names,
 774         .fw_version_get = mlx5_fw_version_get,
 775         .dev_infos_get = mlx5_dev_infos_get,
 776         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 777         .vlan_filter_set = mlx5_vlan_filter_set,
 778         .rx_queue_setup = mlx5_rx_queue_setup,
 779         .tx_queue_setup = mlx5_tx_queue_setup,
 780         .rx_queue_release = mlx5_rx_queue_release,
 781         .tx_queue_release = mlx5_tx_queue_release,
 782         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 783         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 784         .mac_addr_remove = mlx5_mac_addr_remove,
 785         .mac_addr_add = mlx5_mac_addr_add,
 786         .mac_addr_set = mlx5_mac_addr_set,
 787         .set_mc_addr_list = mlx5_set_mc_addr_list,
 788         .mtu_set = mlx5_dev_set_mtu,
 789         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 790         .vlan_offload_set = mlx5_vlan_offload_set,
 791         .reta_update = mlx5_dev_rss_reta_update,
 792         .reta_query = mlx5_dev_rss_reta_query,
 793         .rss_hash_update = mlx5_rss_hash_update,
 794         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
 795         .filter_ctrl = mlx5_dev_filter_ctrl,
 796         .rx_descriptor_status = mlx5_rx_descriptor_status,
 797         .tx_descriptor_status = mlx5_tx_descriptor_status,
 798         .rx_queue_count = mlx5_rx_queue_count,
 799         .rx_queue_intr_enable = mlx5_rx_intr_enable,
 800         .rx_queue_intr_disable = mlx5_rx_intr_disable,
 801         .is_removed = mlx5_is_removed,
 802 };
 803
 804 /* Available operations from secondary process. */
 805 static const struct eth_dev_ops mlx5_dev_sec_ops = {
 806         .stats_get = mlx5_stats_get,
 807         .stats_reset = mlx5_stats_reset,
 808         .xstats_get = mlx5_xstats_get,
 809         .xstats_reset = mlx5_xstats_reset,
 810         .xstats_get_names = mlx5_xstats_get_names,
 811         .fw_version_get = mlx5_fw_version_get,
 812         .dev_infos_get = mlx5_dev_infos_get,
 813         .rx_descriptor_status = mlx5_rx_descriptor_status,
 814         .tx_descriptor_status = mlx5_tx_descriptor_status,
 815 };
 816
 817 /* Available operations in flow isolated mode. */
 818 const struct eth_dev_ops mlx5_dev_ops_isolate = {
 819         .dev_configure = mlx5_dev_configure,
 820         .dev_start = mlx5_dev_start,
 821         .dev_stop = mlx5_dev_stop,
 822         .dev_set_link_down = mlx5_set_link_down,
 823         .dev_set_link_up = mlx5_set_link_up,
 824         .dev_close = mlx5_dev_close,
 825         .promiscuous_enable = mlx5_promiscuous_enable,
 826         .promiscuous_disable = mlx5_promiscuous_disable,
 827         .allmulticast_enable = mlx5_allmulticast_enable,
 828         .allmulticast_disable = mlx5_allmulticast_disable,
 829         .link_update = mlx5_link_update,
 830         .stats_get = mlx5_stats_get,
 831         .stats_reset = mlx5_stats_reset,
 832         .xstats_get = mlx5_xstats_get,
 833         .xstats_reset = mlx5_xstats_reset,
 834         .xstats_get_names = mlx5_xstats_get_names,
 835         .fw_version_get = mlx5_fw_version_get,
 836         .dev_infos_get = mlx5_dev_infos_get,
 837         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 838         .vlan_filter_set = mlx5_vlan_filter_set,
 839         .rx_queue_setup = mlx5_rx_queue_setup,
 840         .tx_queue_setup = mlx5_tx_queue_setup,
 841         .rx_queue_release = mlx5_rx_queue_release,
 842         .tx_queue_release = mlx5_tx_queue_release,
 843         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 844         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 845         .mac_addr_remove = mlx5_mac_addr_remove,
 846         .mac_addr_add = mlx5_mac_addr_add,
 847         .mac_addr_set = mlx5_mac_addr_set,
 848         .set_mc_addr_list = mlx5_set_mc_addr_list,
 849         .mtu_set = mlx5_dev_set_mtu,
 850         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 851         .vlan_offload_set = mlx5_vlan_offload_set,
 852         .filter_ctrl = mlx5_dev_filter_ctrl,
 853         .rx_descriptor_status = mlx5_rx_descriptor_status,
 854         .tx_descriptor_status = mlx5_tx_descriptor_status,
 855         .rx_queue_intr_enable = mlx5_rx_intr_enable,
 856         .rx_queue_intr_disable = mlx5_rx_intr_disable,
 857         .is_removed = mlx5_is_removed,
 858 };
 859
 860 /**
 861  * Verify and store value for device argument.
 862  *
 863  * @param[in] key
 864  *   Key argument to verify.
 865  * @param[in] val
 866  *   Value associated with key.
 867  * @param opaque
 868  *   User data.
 869  *
 870  * @return
 871  *   0 on success, a negative errno value otherwise and rte_errno is set.
 872  */
 873 static int
 874 mlx5_args_check(const char *key, const char *val, void *opaque)
 875 {
 876         struct mlx5_dev_config *config = opaque;
 877         unsigned long tmp;
 878
 879         /* No-op, port representors are processed in mlx5_dev_spawn(). */
 880         if (!strcmp(MLX5_REPRESENTOR, key))
 881                 return 0;
 882         errno = 0;
 883         tmp = strtoul(val, NULL, 0);
 884         if (errno) {
 885                 rte_errno = errno;
 886                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
 887                 return -rte_errno;
 888         }
 889         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
 890                 config->cqe_comp = !!tmp;
 891         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
 892                 config->cqe_pad = !!tmp;
 893         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
 894                 config->hw_padding = !!tmp;
 895         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
 896                 config->mprq.enabled = !!tmp;
 897         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
 898                 config->mprq.stride_num_n = tmp;
 899         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
 900                 config->mprq.max_memcpy_len = tmp;
 901         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
 902                 config->mprq.min_rxqs_num = tmp;
 903         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
 904                 config->txq_inline = tmp;
 905         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
 906                 config->txqs_inline = tmp;
 907         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
 908                 config->txqs_vec = tmp;
 909         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
 910                 config->mps = !!tmp;
 911         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
 912                 config->mpw_hdr_dseg = !!tmp;
 913         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
 914                 config->inline_max_packet_sz = tmp;
 915         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
 916                 config->tx_vec_en = !!tmp;
 917         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
 918                 config->rx_vec_en = !!tmp;
 919         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
 920                 config->l3_vxlan_en = !!tmp;
 921         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
 922                 config->vf_nl_en = !!tmp;
 923         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
 924                 config->dv_esw_en = !!tmp;
 925         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
 926                 config->dv_flow_en = !!tmp;
 927         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
 928                 config->mr_ext_memseg_en = !!tmp;
 929         } else {
 930                 DRV_LOG(WARNING, "%s: unknown parameter", key);
 931                 rte_errno = EINVAL;
 932                 return -rte_errno;
 933         }
 934         return 0;
 935 }
 936
 937 /**
 938  * Parse device parameters.
 939  *
 940  * @param config
 941  *   Pointer to device configuration structure.
 942  * @param devargs
 943  *   Device arguments structure.
 944  *
 945  * @return
 946  *   0 on success, a negative errno value otherwise and rte_errno is set.
 947  */
 948 static int
 949 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
 950 {
 951         const char **params = (const char *[]){
 952                 MLX5_RXQ_CQE_COMP_EN,
 953                 MLX5_RXQ_CQE_PAD_EN,
 954                 MLX5_RXQ_PKT_PAD_EN,
 955                 MLX5_RX_MPRQ_EN,
 956                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
 957                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
 958                 MLX5_RXQS_MIN_MPRQ,
 959                 MLX5_TXQ_INLINE,
 960                 MLX5_TXQS_MIN_INLINE,
 961                 MLX5_TXQS_MAX_VEC,
 962                 MLX5_TXQ_MPW_EN,
 963                 MLX5_TXQ_MPW_HDR_DSEG_EN,
 964                 MLX5_TXQ_MAX_INLINE_LEN,
 965                 MLX5_TX_VEC_EN,
 966                 MLX5_RX_VEC_EN,
 967                 MLX5_L3_VXLAN_EN,
 968                 MLX5_VF_NL_EN,
 969                 MLX5_DV_ESW_EN,
 970                 MLX5_DV_FLOW_EN,
 971                 MLX5_MR_EXT_MEMSEG_EN,
 972                 MLX5_REPRESENTOR,
 973                 NULL,
 974         };
 975         struct rte_kvargs *kvlist;
 976         int ret = 0;
 977         int i;
 978
 979         if (devargs == NULL)
 980                 return 0;
 981         /* Following UGLY cast is done to pass checkpatch. */
 982         kvlist = rte_kvargs_parse(devargs->args, params);
 983         if (kvlist == NULL)
 984                 return 0;
 985         /* Process parameters. */
 986         for (i = 0; (params[i] != NULL); ++i) {
 987                 if (rte_kvargs_count(kvlist, params[i])) {
 988                         ret = rte_kvargs_process(kvlist, params[i],
 989                                                  mlx5_args_check, config);
 990                         if (ret) {
 991                                 rte_errno = EINVAL;
 992                                 rte_kvargs_free(kvlist);
 993                                 return -rte_errno;
 994                         }
 995                 }
 996         }
 997         rte_kvargs_free(kvlist);
 998         return 0;
 999 }
1000
1001 static struct rte_pci_driver mlx5_driver;
1002
1003 /**
1004  * PMD global initialization.
1005  *
1006  * Independent from individual device, this function initializes global
1007  * per-PMD data structures distinguishing primary and secondary processes.
1008  * Hence, each initialization is called once per a process.
1009  *
1010  * @return
1011  *   0 on success, a negative errno value otherwise and rte_errno is set.
1012  */
1013 static int
1014 mlx5_init_once(void)
1015 {
1016         struct mlx5_shared_data *sd;
1017         struct mlx5_local_data *ld = &mlx5_local_data;
1018
1019         if (mlx5_init_shared_data())
1020                 return -rte_errno;
1021         sd = mlx5_shared_data;
1022         assert(sd);
1023         rte_spinlock_lock(&sd->lock);
1024         switch (rte_eal_process_type()) {
1025         case RTE_PROC_PRIMARY:
1026                 if (sd->init_done)
1027                         break;
1028                 LIST_INIT(&sd->mem_event_cb_list);
1029                 rte_rwlock_init(&sd->mem_event_rwlock);
1030                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1031                                                 mlx5_mr_mem_event_cb, NULL);
1032                 mlx5_mp_init_primary();
1033                 sd->init_done = true;
1034                 break;
1035         case RTE_PROC_SECONDARY:
1036                 if (ld->init_done)
1037                         break;
1038                 mlx5_mp_init_secondary();
1039                 ++sd->secondary_cnt;
1040                 ld->init_done = true;
1041                 break;
1042         default:
1043                 break;
1044         }
1045         rte_spinlock_unlock(&sd->lock);
1046         return 0;
1047 }
1048
1049 /**
1050  * Spawn an Ethernet device from Verbs information.
1051  *
1052  * @param dpdk_dev
1053  *   Backing DPDK device.
1054  * @param spawn
1055  *   Verbs device parameters (name, port, switch_info) to spawn.
1056  * @param config
1057  *   Device configuration parameters.
1058  *
1059  * @return
1060  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1061  *   is set. The following errors are defined:
1062  *
1063  *   EBUSY: device is not supposed to be spawned.
1064  *   EEXIST: device is already spawned
1065  */
1066 static struct rte_eth_dev *
1067 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1068                struct mlx5_dev_spawn_data *spawn,
1069                struct mlx5_dev_config config)
1070 {
1071         const struct mlx5_switch_info *switch_info = &spawn->info;
1072         struct mlx5_ibv_shared *sh = NULL;
1073         struct ibv_port_attr port_attr;
1074         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1075         struct rte_eth_dev *eth_dev = NULL;
1076         struct mlx5_priv *priv = NULL;
1077         int err = 0;
1078         unsigned int hw_padding = 0;
1079         unsigned int mps;
1080         unsigned int cqe_comp;
1081         unsigned int cqe_pad = 0;
1082         unsigned int tunnel_en = 0;
1083         unsigned int mpls_en = 0;
1084         unsigned int swp = 0;
1085         unsigned int mprq = 0;
1086         unsigned int mprq_min_stride_size_n = 0;
1087         unsigned int mprq_max_stride_size_n = 0;
1088         unsigned int mprq_min_stride_num_n = 0;
1089         unsigned int mprq_max_stride_num_n = 0;
1090         struct rte_ether_addr mac;
1091         char name[RTE_ETH_NAME_MAX_LEN];
1092         int own_domain_id = 0;
1093         uint16_t port_id;
1094         unsigned int i;
1095
1096         /* Determine if this port representor is supposed to be spawned. */
1097         if (switch_info->representor && dpdk_dev->devargs) {
1098                 struct rte_eth_devargs eth_da;
1099
1100                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
1101                 if (err) {
1102                         rte_errno = -err;
1103                         DRV_LOG(ERR, "failed to process device arguments: %s",
1104                                 strerror(rte_errno));
1105                         return NULL;
1106                 }
1107                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
1108                         if (eth_da.representor_ports[i] ==
1109                             (uint16_t)switch_info->port_name)
1110                                 break;
1111                 if (i == eth_da.nb_representor_ports) {
1112                         rte_errno = EBUSY;
1113                         return NULL;
1114                 }
1115         }
1116         /* Build device name. */
1117         if (!switch_info->representor)
1118                 strlcpy(name, dpdk_dev->name, sizeof(name));
1119         else
1120                 snprintf(name, sizeof(name), "%s_representor_%u",
1121                          dpdk_dev->name, switch_info->port_name);
1122         /* check if the device is already spawned */
1123         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1124                 rte_errno = EEXIST;
1125                 return NULL;
1126         }
1127         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1128         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1129                 eth_dev = rte_eth_dev_attach_secondary(name);
1130                 if (eth_dev == NULL) {
1131                         DRV_LOG(ERR, "can not attach rte ethdev");
1132                         rte_errno = ENOMEM;
1133                         return NULL;
1134                 }
1135                 eth_dev->device = dpdk_dev;
1136                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
1137                 err = mlx5_proc_priv_init(eth_dev);
1138                 if (err)
1139                         return NULL;
1140                 /* Receive command fd from primary process */
1141                 err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
1142                 if (err < 0)
1143                         return NULL;
1144                 /* Remap UAR for Tx queues. */
1145                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
1146                 if (err)
1147                         return NULL;
1148                 /*
1149                  * Ethdev pointer is still required as input since
1150                  * the primary device is not accessible from the
1151                  * secondary process.
1152                  */
1153                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1154                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1155                 return eth_dev;
1156         }
1157         sh = mlx5_alloc_shared_ibctx(spawn);
1158         if (!sh)
1159                 return NULL;
1160         config.devx = sh->devx;
1161 #ifdef HAVE_IBV_MLX5_MOD_SWP
1162         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1163 #endif
1164         /*
1165          * Multi-packet send is supported by ConnectX-4 Lx PF as well
1166          * as all ConnectX-5 devices.
1167          */
1168 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1169         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1170 #endif
1171 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1172         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1173 #endif
1174         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1175         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1176                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1177                         DRV_LOG(DEBUG, "enhanced MPW is supported");
1178                         mps = MLX5_MPW_ENHANCED;
1179                 } else {
1180                         DRV_LOG(DEBUG, "MPW is supported");
1181                         mps = MLX5_MPW;
1182                 }
1183         } else {
1184                 DRV_LOG(DEBUG, "MPW isn't supported");
1185                 mps = MLX5_MPW_DISABLED;
1186         }
1187 #ifdef HAVE_IBV_MLX5_MOD_SWP
1188         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1189                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1190         DRV_LOG(DEBUG, "SWP support: %u", swp);
1191 #endif
1192         config.swp = !!swp;
1193 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1194         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1195                 struct mlx5dv_striding_rq_caps mprq_caps =
1196                         dv_attr.striding_rq_caps;
1197
1198                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1199                         mprq_caps.min_single_stride_log_num_of_bytes);
1200                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1201                         mprq_caps.max_single_stride_log_num_of_bytes);
1202                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1203                         mprq_caps.min_single_wqe_log_num_of_strides);
1204                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1205                         mprq_caps.max_single_wqe_log_num_of_strides);
1206                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1207                         mprq_caps.supported_qpts);
1208                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1209                 mprq = 1;
1210                 mprq_min_stride_size_n =
1211                         mprq_caps.min_single_stride_log_num_of_bytes;
1212                 mprq_max_stride_size_n =
1213                         mprq_caps.max_single_stride_log_num_of_bytes;
1214                 mprq_min_stride_num_n =
1215                         mprq_caps.min_single_wqe_log_num_of_strides;
1216                 mprq_max_stride_num_n =
1217                         mprq_caps.max_single_wqe_log_num_of_strides;
1218                 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1219                                                    mprq_min_stride_num_n);
1220         }
1221 #endif
1222         if (RTE_CACHE_LINE_SIZE == 128 &&
1223             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1224                 cqe_comp = 0;
1225         else
1226                 cqe_comp = 1;
1227         config.cqe_comp = cqe_comp;
1228 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1229         /* Whether device supports 128B Rx CQE padding. */
1230         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1231                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1232 #endif
1233 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1234         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1235                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
1236                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1237                              (dv_attr.tunnel_offloads_caps &
1238                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1239         }
1240         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1241                 tunnel_en ? "" : "not ");
1242 #else
1243         DRV_LOG(WARNING,
1244                 "tunnel offloading disabled due to old OFED/rdma-core version");
1245 #endif
1246         config.tunnel_en = tunnel_en;
1247 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1248         mpls_en = ((dv_attr.tunnel_offloads_caps &
1249                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1250                    (dv_attr.tunnel_offloads_caps &
1251                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1252         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1253                 mpls_en ? "" : "not ");
1254 #else
1255         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1256                 " old OFED/rdma-core version or firmware configuration");
1257 #endif
1258         config.mpls_en = mpls_en;
1259         /* Check port status. */
1260         err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1261         if (err) {
1262                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1263                 goto error;
1264         }
1265         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1266                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1267                 err = EINVAL;
1268                 goto error;
1269         }
1270         if (port_attr.state != IBV_PORT_ACTIVE)
1271                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1272                         mlx5_glue->port_state_str(port_attr.state),
1273                         port_attr.state);
1274         /* Allocate private eth device data. */
1275         priv = rte_zmalloc("ethdev private structure",
1276                            sizeof(*priv),
1277                            RTE_CACHE_LINE_SIZE);
1278         if (priv == NULL) {
1279                 DRV_LOG(ERR, "priv allocation failure");
1280                 err = ENOMEM;
1281                 goto error;
1282         }
1283         priv->sh = sh;
1284         priv->ibv_port = spawn->ibv_port;
1285         priv->mtu = RTE_ETHER_MTU;
1286 #ifndef RTE_ARCH_64
1287         /* Initialize UAR access locks for 32bit implementations. */
1288         rte_spinlock_init(&priv->uar_lock_cq);
1289         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1290                 rte_spinlock_init(&priv->uar_lock[i]);
1291 #endif
1292         /* Some internal functions rely on Netlink sockets, open them now. */
1293         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1294         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1295         priv->nl_sn = 0;
1296         priv->representor = !!switch_info->representor;
1297         priv->master = !!switch_info->master;
1298         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1299         /*
1300          * Currently we support single E-Switch per PF configurations
1301          * only and vport_id field contains the vport index for
1302          * associated VF, which is deduced from representor port name.
1303          * For example, let's have the IB device port 10, it has
1304          * attached network device eth0, which has port name attribute
1305          * pf0vf2, we can deduce the VF number as 2, and set vport index
1306          * as 3 (2+1). This assigning schema should be changed if the
1307          * multiple E-Switch instances per PF configurations or/and PCI
1308          * subfunctions are added.
1309          */
1310         priv->vport_id = switch_info->representor ?
1311                          switch_info->port_name + 1 : -1;
1312         /* representor_id field keeps the unmodified port/VF index. */
1313         priv->representor_id = switch_info->representor ?
1314                                switch_info->port_name : -1;
1315         /*
1316          * Look for sibling devices in order to reuse their switch domain
1317          * if any, otherwise allocate one.
1318          */
1319         RTE_ETH_FOREACH_DEV_OF(port_id, dpdk_dev) {
1320                 const struct mlx5_priv *opriv =
1321                         rte_eth_devices[port_id].data->dev_private;
1322
1323                 if (!opriv ||
1324                         opriv->domain_id ==
1325                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1326                         continue;
1327                 priv->domain_id = opriv->domain_id;
1328                 break;
1329         }
1330         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1331                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1332                 if (err) {
1333                         err = rte_errno;
1334                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
1335                                 strerror(rte_errno));
1336                         goto error;
1337                 }
1338                 own_domain_id = 1;
1339         }
1340         err = mlx5_args(&config, dpdk_dev->devargs);
1341         if (err) {
1342                 err = rte_errno;
1343                 DRV_LOG(ERR, "failed to process device arguments: %s",
1344                         strerror(rte_errno));
1345                 goto error;
1346         }
1347         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1348                             IBV_DEVICE_RAW_IP_CSUM);
1349         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1350                 (config.hw_csum ? "" : "not "));
1351 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1352         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1353         DRV_LOG(DEBUG, "counters are not supported");
1354 #endif
1355 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1356         if (config.dv_flow_en) {
1357                 DRV_LOG(WARNING, "DV flow is not supported");
1358                 config.dv_flow_en = 0;
1359         }
1360 #endif
1361         config.ind_table_max_size =
1362                 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1363         /*
1364          * Remove this check once DPDK supports larger/variable
1365          * indirection tables.
1366          */
1367         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1368                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1369         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1370                 config.ind_table_max_size);
1371         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1372                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1373         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1374                 (config.hw_vlan_strip ? "" : "not "));
1375         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1376                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
1377         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1378                 (config.hw_fcs_strip ? "" : "not "));
1379 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1380         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1381 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1382         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1383                         IBV_DEVICE_PCI_WRITE_END_PADDING);
1384 #endif
1385         if (config.hw_padding && !hw_padding) {
1386                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1387                 config.hw_padding = 0;
1388         } else if (config.hw_padding) {
1389                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1390         }
1391         config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1392                       (sh->device_attr.tso_caps.supported_qpts &
1393                        (1 << IBV_QPT_RAW_PACKET)));
1394         if (config.tso)
1395                 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1396         /*
1397          * MPW is disabled by default, while the Enhanced MPW is enabled
1398          * by default.
1399          */
1400         if (config.mps == MLX5_ARG_UNSET)
1401                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1402                                                           MLX5_MPW_DISABLED;
1403         else
1404                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1405         DRV_LOG(INFO, "%sMPS is %s",
1406                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1407                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1408         if (config.cqe_comp && !cqe_comp) {
1409                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1410                 config.cqe_comp = 0;
1411         }
1412         if (config.cqe_pad && !cqe_pad) {
1413                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1414                 config.cqe_pad = 0;
1415         } else if (config.cqe_pad) {
1416                 DRV_LOG(INFO, "Rx CQE padding is enabled");
1417         }
1418         if (config.mprq.enabled && mprq) {
1419                 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1420                     config.mprq.stride_num_n < mprq_min_stride_num_n) {
1421                         config.mprq.stride_num_n =
1422                                 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1423                                         mprq_min_stride_num_n);
1424                         DRV_LOG(WARNING,
1425                                 "the number of strides"
1426                                 " for Multi-Packet RQ is out of range,"
1427                                 " setting default value (%u)",
1428                                 1 << config.mprq.stride_num_n);
1429                 }
1430                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1431                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1432         } else if (config.mprq.enabled && !mprq) {
1433                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1434                 config.mprq.enabled = 0;
1435         }
1436         eth_dev = rte_eth_dev_allocate(name);
1437         if (eth_dev == NULL) {
1438                 DRV_LOG(ERR, "can not allocate rte ethdev");
1439                 err = ENOMEM;
1440                 goto error;
1441         }
1442         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1443         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1444         if (priv->representor) {
1445                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1446                 eth_dev->data->representor_id = priv->representor_id;
1447         }
1448         eth_dev->data->dev_private = priv;
1449         priv->dev_data = eth_dev->data;
1450         eth_dev->data->mac_addrs = priv->mac;
1451         eth_dev->device = dpdk_dev;
1452         /* Configure the first MAC address by default. */
1453         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1454                 DRV_LOG(ERR,
1455                         "port %u cannot get MAC address, is mlx5_en"
1456                         " loaded? (errno: %s)",
1457                         eth_dev->data->port_id, strerror(rte_errno));
1458                 err = ENODEV;
1459                 goto error;
1460         }
1461         DRV_LOG(INFO,
1462                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1463                 eth_dev->data->port_id,
1464                 mac.addr_bytes[0], mac.addr_bytes[1],
1465                 mac.addr_bytes[2], mac.addr_bytes[3],
1466                 mac.addr_bytes[4], mac.addr_bytes[5]);
1467 #ifndef NDEBUG
1468         {
1469                 char ifname[IF_NAMESIZE];
1470
1471                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1472                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1473                                 eth_dev->data->port_id, ifname);
1474                 else
1475                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1476                                 eth_dev->data->port_id);
1477         }
1478 #endif
1479         /* Get actual MTU if possible. */
1480         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1481         if (err) {
1482                 err = rte_errno;
1483                 goto error;
1484         }
1485         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1486                 priv->mtu);
1487         /* Initialize burst functions to prevent crashes before link-up. */
1488         eth_dev->rx_pkt_burst = removed_rx_burst;
1489         eth_dev->tx_pkt_burst = removed_tx_burst;
1490         eth_dev->dev_ops = &mlx5_dev_ops;
1491         /* Register MAC address. */
1492         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1493         if (config.vf && config.vf_nl_en)
1494                 mlx5_nl_mac_addr_sync(eth_dev);
1495         priv->tcf_context = mlx5_flow_tcf_context_create();
1496         if (!priv->tcf_context) {
1497                 err = -rte_errno;
1498                 DRV_LOG(WARNING,
1499                         "flow rules relying on switch offloads will not be"
1500                         " supported: cannot open libmnl socket: %s",
1501                         strerror(rte_errno));
1502         } else {
1503                 struct rte_flow_error error;
1504                 unsigned int ifindex = mlx5_ifindex(eth_dev);
1505
1506                 if (!ifindex) {
1507                         err = -rte_errno;
1508                         error.message =
1509                                 "cannot retrieve network interface index";
1510                 } else {
1511                         err = mlx5_flow_tcf_init(priv->tcf_context,
1512                                                  ifindex, &error);
1513                 }
1514                 if (err) {
1515                         DRV_LOG(WARNING,
1516                                 "flow rules relying on switch offloads will"
1517                                 " not be supported: %s: %s",
1518                                 error.message, strerror(rte_errno));
1519                         mlx5_flow_tcf_context_destroy(priv->tcf_context);
1520                         priv->tcf_context = NULL;
1521                 }
1522         }
1523         TAILQ_INIT(&priv->flows);
1524         TAILQ_INIT(&priv->ctrl_flows);
1525         /* Hint libmlx5 to use PMD allocator for data plane resources */
1526         struct mlx5dv_ctx_allocators alctr = {
1527                 .alloc = &mlx5_alloc_verbs_buf,
1528                 .free = &mlx5_free_verbs_buf,
1529                 .data = priv,
1530         };
1531         mlx5_glue->dv_set_context_attr(sh->ctx,
1532                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1533                                        (void *)((uintptr_t)&alctr));
1534         /* Bring Ethernet device up. */
1535         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1536                 eth_dev->data->port_id);
1537         mlx5_set_link_up(eth_dev);
1538         /*
1539          * Even though the interrupt handler is not installed yet,
1540          * interrupts will still trigger on the async_fd from
1541          * Verbs context returned by ibv_open_device().
1542          */
1543         mlx5_link_update(eth_dev, 0);
1544 #ifdef HAVE_IBV_DEVX_OBJ
1545         if (config.devx) {
1546                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
1547                 if (err) {
1548                         err = -err;
1549                         goto error;
1550                 }
1551         }
1552 #endif
1553 #ifdef HAVE_MLX5DV_DR_ESWITCH
1554         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1555               (switch_info->representor || switch_info->master)))
1556                 config.dv_esw_en = 0;
1557 #else
1558         config.dv_esw_en = 0;
1559 #endif
1560         /* Store device configuration on private structure. */
1561         priv->config = config;
1562         if (config.dv_flow_en) {
1563                 err = mlx5_alloc_shared_dr(priv);
1564                 if (err)
1565                         goto error;
1566         }
1567         /* Supported Verbs flow priority number detection. */
1568         err = mlx5_flow_discover_priorities(eth_dev);
1569         if (err < 0) {
1570                 err = -err;
1571                 goto error;
1572         }
1573         priv->config.flow_prio = err;
1574         /* Add device to memory callback list. */
1575         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1576         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1577                          sh, mem_event_cb);
1578         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1579         return eth_dev;
1580 error:
1581         if (priv) {
1582                 if (priv->sh)
1583                         mlx5_free_shared_dr(priv);
1584                 if (priv->nl_socket_route >= 0)
1585                         close(priv->nl_socket_route);
1586                 if (priv->nl_socket_rdma >= 0)
1587                         close(priv->nl_socket_rdma);
1588                 if (priv->tcf_context)
1589                         mlx5_flow_tcf_context_destroy(priv->tcf_context);
1590                 if (own_domain_id)
1591                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1592                 rte_free(priv);
1593                 if (eth_dev != NULL)
1594                         eth_dev->data->dev_private = NULL;
1595         }
1596         if (eth_dev != NULL) {
1597                 /* mac_addrs must not be freed alone because part of dev_private */
1598                 eth_dev->data->mac_addrs = NULL;
1599                 rte_eth_dev_release_port(eth_dev);
1600         }
1601         if (sh)
1602                 mlx5_free_shared_ibctx(sh);
1603         assert(err > 0);
1604         rte_errno = err;
1605         return NULL;
1606 }
1607
1608 /**
1609  * Comparison callback to sort device data.
1610  *
1611  * This is meant to be used with qsort().
1612  *
1613  * @param a[in]
1614  *   Pointer to pointer to first data object.
1615  * @param b[in]
1616  *   Pointer to pointer to second data object.
1617  *
1618  * @return
1619  *   0 if both objects are equal, less than 0 if the first argument is less
1620  *   than the second, greater than 0 otherwise.
1621  */
1622 static int
1623 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1624 {
1625         const struct mlx5_switch_info *si_a =
1626                 &((const struct mlx5_dev_spawn_data *)a)->info;
1627         const struct mlx5_switch_info *si_b =
1628                 &((const struct mlx5_dev_spawn_data *)b)->info;
1629         int ret;
1630
1631         /* Master device first. */
1632         ret = si_b->master - si_a->master;
1633         if (ret)
1634                 return ret;
1635         /* Then representor devices. */
1636         ret = si_b->representor - si_a->representor;
1637         if (ret)
1638                 return ret;
1639         /* Unidentified devices come last in no specific order. */
1640         if (!si_a->representor)
1641                 return 0;
1642         /* Order representors by name. */
1643         return si_a->port_name - si_b->port_name;
1644 }
1645
1646 /**
1647  * DPDK callback to register a PCI device.
1648  *
1649  * This function spawns Ethernet devices out of a given PCI device.
1650  *
1651  * @param[in] pci_drv
1652  *   PCI driver structure (mlx5_driver).
1653  * @param[in] pci_dev
1654  *   PCI device information.
1655  *
1656  * @return
1657  *   0 on success, a negative errno value otherwise and rte_errno is set.
1658  */
1659 static int
1660 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1661                struct rte_pci_device *pci_dev)
1662 {
1663         struct ibv_device **ibv_list;
1664         /*
1665          * Number of found IB Devices matching with requested PCI BDF.
1666          * nd != 1 means there are multiple IB devices over the same
1667          * PCI device and we have representors and master.
1668          */
1669         unsigned int nd = 0;
1670         /*
1671          * Number of found IB device Ports. nd = 1 and np = 1..n means
1672          * we have the single multiport IB device, and there may be
1673          * representors attached to some of found ports.
1674          */
1675         unsigned int np = 0;
1676         /*
1677          * Number of DPDK ethernet devices to Spawn - either over
1678          * multiple IB devices or multiple ports of single IB device.
1679          * Actually this is the number of iterations to spawn.
1680          */
1681         unsigned int ns = 0;
1682         struct mlx5_dev_config dev_config;
1683         int ret;
1684
1685         ret = mlx5_init_once();
1686         if (ret) {
1687                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1688                         strerror(rte_errno));
1689                 return -rte_errno;
1690         }
1691         assert(pci_drv == &mlx5_driver);
1692         errno = 0;
1693         ibv_list = mlx5_glue->get_device_list(&ret);
1694         if (!ibv_list) {
1695                 rte_errno = errno ? errno : ENOSYS;
1696                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1697                 return -rte_errno;
1698         }
1699         /*
1700          * First scan the list of all Infiniband devices to find
1701          * matching ones, gathering into the list.
1702          */
1703         struct ibv_device *ibv_match[ret + 1];
1704         int nl_route = -1;
1705         int nl_rdma = -1;
1706         unsigned int i;
1707
1708         while (ret-- > 0) {
1709                 struct rte_pci_addr pci_addr;
1710
1711                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1712                 if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1713                         continue;
1714                 if (pci_dev->addr.domain != pci_addr.domain ||
1715                     pci_dev->addr.bus != pci_addr.bus ||
1716                     pci_dev->addr.devid != pci_addr.devid ||
1717                     pci_dev->addr.function != pci_addr.function)
1718                         continue;
1719                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1720                         ibv_list[ret]->name);
1721                 ibv_match[nd++] = ibv_list[ret];
1722         }
1723         ibv_match[nd] = NULL;
1724         if (!nd) {
1725                 /* No device matches, just complain and bail out. */
1726                 mlx5_glue->free_device_list(ibv_list);
1727                 DRV_LOG(WARNING,
1728                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1729                         " are kernel drivers loaded?",
1730                         pci_dev->addr.domain, pci_dev->addr.bus,
1731                         pci_dev->addr.devid, pci_dev->addr.function);
1732                 rte_errno = ENOENT;
1733                 ret = -rte_errno;
1734                 return ret;
1735         }
1736         nl_route = mlx5_nl_init(NETLINK_ROUTE);
1737         nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1738         if (nd == 1) {
1739                 /*
1740                  * Found single matching device may have multiple ports.
1741                  * Each port may be representor, we have to check the port
1742                  * number and check the representors existence.
1743                  */
1744                 if (nl_rdma >= 0)
1745                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1746                 if (!np)
1747                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1748                                          " ports number", ibv_match[0]->name);
1749         }
1750         /*
1751          * Now we can determine the maximal
1752          * amount of devices to be spawned.
1753          */
1754         struct mlx5_dev_spawn_data list[np ? np : nd];
1755
1756         if (np > 1) {
1757                 /*
1758                  * Single IB device with multiple ports found,
1759                  * it may be E-Switch master device and representors.
1760                  * We have to perform identification trough the ports.
1761                  */
1762                 assert(nl_rdma >= 0);
1763                 assert(ns == 0);
1764                 assert(nd == 1);
1765                 for (i = 1; i <= np; ++i) {
1766                         list[ns].max_port = np;
1767                         list[ns].ibv_port = i;
1768                         list[ns].ibv_dev = ibv_match[0];
1769                         list[ns].eth_dev = NULL;
1770                         list[ns].pci_dev = pci_dev;
1771                         list[ns].ifindex = mlx5_nl_ifindex
1772                                         (nl_rdma, list[ns].ibv_dev->name, i);
1773                         if (!list[ns].ifindex) {
1774                                 /*
1775                                  * No network interface index found for the
1776                                  * specified port, it means there is no
1777                                  * representor on this port. It's OK,
1778                                  * there can be disabled ports, for example
1779                                  * if sriov_numvfs < sriov_totalvfs.
1780                                  */
1781                                 continue;
1782                         }
1783                         ret = -1;
1784                         if (nl_route >= 0)
1785                                 ret = mlx5_nl_switch_info
1786                                                (nl_route,
1787                                                 list[ns].ifindex,
1788                                                 &list[ns].info);
1789                         if (ret || (!list[ns].info.representor &&
1790                                     !list[ns].info.master)) {
1791                                 /*
1792                                  * We failed to recognize representors with
1793                                  * Netlink, let's try to perform the task
1794                                  * with sysfs.
1795                                  */
1796                                 ret =  mlx5_sysfs_switch_info
1797                                                 (list[ns].ifindex,
1798                                                  &list[ns].info);
1799                         }
1800                         if (!ret && (list[ns].info.representor ^
1801                                      list[ns].info.master))
1802                                 ns++;
1803                 }
1804                 if (!ns) {
1805                         DRV_LOG(ERR,
1806                                 "unable to recognize master/representors"
1807                                 " on the IB device with multiple ports");
1808                         rte_errno = ENOENT;
1809                         ret = -rte_errno;
1810                         goto exit;
1811                 }
1812         } else {
1813                 /*
1814                  * The existence of several matching entries (nd > 1) means
1815                  * port representors have been instantiated. No existing Verbs
1816                  * call nor sysfs entries can tell them apart, this can only
1817                  * be done through Netlink calls assuming kernel drivers are
1818                  * recent enough to support them.
1819                  *
1820                  * In the event of identification failure through Netlink,
1821                  * try again through sysfs, then:
1822                  *
1823                  * 1. A single IB device matches (nd == 1) with single
1824                  *    port (np=0/1) and is not a representor, assume
1825                  *    no switch support.
1826                  *
1827                  * 2. Otherwise no safe assumptions can be made;
1828                  *    complain louder and bail out.
1829                  */
1830                 np = 1;
1831                 for (i = 0; i != nd; ++i) {
1832                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1833                         list[ns].max_port = 1;
1834                         list[ns].ibv_port = 1;
1835                         list[ns].ibv_dev = ibv_match[i];
1836                         list[ns].eth_dev = NULL;
1837                         list[ns].pci_dev = pci_dev;
1838                         list[ns].ifindex = 0;
1839                         if (nl_rdma >= 0)
1840                                 list[ns].ifindex = mlx5_nl_ifindex
1841                                         (nl_rdma, list[ns].ibv_dev->name, 1);
1842                         if (!list[ns].ifindex) {
1843                                 char ifname[IF_NAMESIZE];
1844
1845                                 /*
1846                                  * Netlink failed, it may happen with old
1847                                  * ib_core kernel driver (before 4.16).
1848                                  * We can assume there is old driver because
1849                                  * here we are processing single ports IB
1850                                  * devices. Let's try sysfs to retrieve
1851                                  * the ifindex. The method works for
1852                                  * master device only.
1853                                  */
1854                                 if (nd > 1) {
1855                                         /*
1856                                          * Multiple devices found, assume
1857                                          * representors, can not distinguish
1858                                          * master/representor and retrieve
1859                                          * ifindex via sysfs.
1860                                          */
1861                                         continue;
1862                                 }
1863                                 ret = mlx5_get_master_ifname
1864                                         (ibv_match[i]->ibdev_path, &ifname);
1865                                 if (!ret)
1866                                         list[ns].ifindex =
1867                                                 if_nametoindex(ifname);
1868                                 if (!list[ns].ifindex) {
1869                                         /*
1870                                          * No network interface index found
1871                                          * for the specified device, it means
1872                                          * there it is neither representor
1873                                          * nor master.
1874                                          */
1875                                         continue;
1876                                 }
1877                         }
1878                         ret = -1;
1879                         if (nl_route >= 0)
1880                                 ret = mlx5_nl_switch_info
1881                                                (nl_route,
1882                                                 list[ns].ifindex,
1883                                                 &list[ns].info);
1884                         if (ret || (!list[ns].info.representor &&
1885                                     !list[ns].info.master)) {
1886                                 /*
1887                                  * We failed to recognize representors with
1888                                  * Netlink, let's try to perform the task
1889                                  * with sysfs.
1890                                  */
1891                                 ret =  mlx5_sysfs_switch_info
1892                                                 (list[ns].ifindex,
1893                                                  &list[ns].info);
1894                         }
1895                         if (!ret && (list[ns].info.representor ^
1896                                      list[ns].info.master)) {
1897                                 ns++;
1898                         } else if ((nd == 1) &&
1899                                    !list[ns].info.representor &&
1900                                    !list[ns].info.master) {
1901                                 /*
1902                                  * Single IB device with
1903                                  * one physical port and
1904                                  * attached network device.
1905                                  * May be SRIOV is not enabled
1906                                  * or there is no representors.
1907                                  */
1908                                 DRV_LOG(INFO, "no E-Switch support detected");
1909                                 ns++;
1910                                 break;
1911                         }
1912                 }
1913                 if (!ns) {
1914                         DRV_LOG(ERR,
1915                                 "unable to recognize master/representors"
1916                                 " on the multiple IB devices");
1917                         rte_errno = ENOENT;
1918                         ret = -rte_errno;
1919                         goto exit;
1920                 }
1921         }
1922         assert(ns);
1923         /*
1924          * Sort list to probe devices in natural order for users convenience
1925          * (i.e. master first, then representors from lowest to highest ID).
1926          */
1927         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1928         /* Default configuration. */
1929         dev_config = (struct mlx5_dev_config){
1930                 .hw_padding = 0,
1931                 .mps = MLX5_ARG_UNSET,
1932                 .tx_vec_en = 1,
1933                 .rx_vec_en = 1,
1934                 .txq_inline = MLX5_ARG_UNSET,
1935                 .txqs_inline = MLX5_ARG_UNSET,
1936                 .txqs_vec = MLX5_ARG_UNSET,
1937                 .inline_max_packet_sz = MLX5_ARG_UNSET,
1938                 .vf_nl_en = 1,
1939                 .mr_ext_memseg_en = 1,
1940                 .mprq = {
1941                         .enabled = 0, /* Disabled by default. */
1942                         .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
1943                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1944                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1945                 },
1946                 .dv_esw_en = 1,
1947         };
1948         /* Device specific configuration. */
1949         switch (pci_dev->id.device_id) {
1950         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BF:
1951                 dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS_BLUEFIELD;
1952                 break;
1953         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1954         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1955         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1956         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1957                 dev_config.vf = 1;
1958                 break;
1959         default:
1960                 break;
1961         }
1962         /* Set architecture-dependent default value if unset. */
1963         if (dev_config.txqs_vec == MLX5_ARG_UNSET)
1964                 dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS;
1965         for (i = 0; i != ns; ++i) {
1966                 uint32_t restore;
1967
1968                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1969                                                  &list[i],
1970                                                  dev_config);
1971                 if (!list[i].eth_dev) {
1972                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1973                                 break;
1974                         /* Device is disabled or already spawned. Ignore it. */
1975                         continue;
1976                 }
1977                 restore = list[i].eth_dev->data->dev_flags;
1978                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1979                 /* Restore non-PCI flags cleared by the above call. */
1980                 list[i].eth_dev->data->dev_flags |= restore;
1981                 rte_eth_dev_probing_finish(list[i].eth_dev);
1982         }
1983         if (i != ns) {
1984                 DRV_LOG(ERR,
1985                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1986                         " encountering an error: %s",
1987                         pci_dev->addr.domain, pci_dev->addr.bus,
1988                         pci_dev->addr.devid, pci_dev->addr.function,
1989                         strerror(rte_errno));
1990                 ret = -rte_errno;
1991                 /* Roll back. */
1992                 while (i--) {
1993                         if (!list[i].eth_dev)
1994                                 continue;
1995                         mlx5_dev_close(list[i].eth_dev);
1996                         /* mac_addrs must not be freed because in dev_private */
1997                         list[i].eth_dev->data->mac_addrs = NULL;
1998                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1999                 }
2000                 /* Restore original error. */
2001                 rte_errno = -ret;
2002         } else {
2003                 ret = 0;
2004         }
2005 exit:
2006         /*
2007          * Do the routine cleanup:
2008          * - close opened Netlink sockets
2009          * - free the Infiniband device list
2010          */
2011         if (nl_rdma >= 0)
2012                 close(nl_rdma);
2013         if (nl_route >= 0)
2014                 close(nl_route);
2015         assert(ibv_list);
2016         mlx5_glue->free_device_list(ibv_list);
2017         return ret;
2018 }
2019
2020 /**
2021  * DPDK callback to remove a PCI device.
2022  *
2023  * This function removes all Ethernet devices belong to a given PCI device.
2024  *
2025  * @param[in] pci_dev
2026  *   Pointer to the PCI device.
2027  *
2028  * @return
2029  *   0 on success, the function cannot fail.
2030  */
2031 static int
2032 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2033 {
2034         uint16_t port_id;
2035
2036         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
2037                 rte_eth_dev_close(port_id);
2038         return 0;
2039 }
2040
2041 static const struct rte_pci_id mlx5_pci_id_map[] = {
2042         {
2043                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2044                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2045         },
2046         {
2047                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2048                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2049         },
2050         {
2051                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2052                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2053         },
2054         {
2055                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2056                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2057         },
2058         {
2059                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2060                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2061         },
2062         {
2063                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2064                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2065         },
2066         {
2067                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2068                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2069         },
2070         {
2071                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2072                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2073         },
2074         {
2075                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2076                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2077         },
2078         {
2079                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2080                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2081         },
2082         {
2083                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2084                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2085         },
2086         {
2087                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2088                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2089         },
2090         {
2091                 .vendor_id = 0
2092         }
2093 };
2094
2095 static struct rte_pci_driver mlx5_driver = {
2096         .driver = {
2097                 .name = MLX5_DRIVER_NAME
2098         },
2099         .id_table = mlx5_pci_id_map,
2100         .probe = mlx5_pci_probe,
2101         .remove = mlx5_pci_remove,
2102         .dma_map = mlx5_dma_map,
2103         .dma_unmap = mlx5_dma_unmap,
2104         .drv_flags = (RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
2105                       RTE_PCI_DRV_PROBE_AGAIN),
2106 };
2107
2108 #ifdef RTE_IBVERBS_LINK_DLOPEN
2109
2110 /**
2111  * Suffix RTE_EAL_PMD_PATH with "-glue".
2112  *
2113  * This function performs a sanity check on RTE_EAL_PMD_PATH before
2114  * suffixing its last component.
2115  *
2116  * @param buf[out]
2117  *   Output buffer, should be large enough otherwise NULL is returned.
2118  * @param size
2119  *   Size of @p out.
2120  *
2121  * @return
2122  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
2123  */
2124 static char *
2125 mlx5_glue_path(char *buf, size_t size)
2126 {
2127         static const char *const bad[] = { "/", ".", "..", NULL };
2128         const char *path = RTE_EAL_PMD_PATH;
2129         size_t len = strlen(path);
2130         size_t off;
2131         int i;
2132
2133         while (len && path[len - 1] == '/')
2134                 --len;
2135         for (off = len; off && path[off - 1] != '/'; --off)
2136                 ;
2137         for (i = 0; bad[i]; ++i)
2138                 if (!strncmp(path + off, bad[i], (int)(len - off)))
2139                         goto error;
2140         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
2141         if (i == -1 || (size_t)i >= size)
2142                 goto error;
2143         return buf;
2144 error:
2145         DRV_LOG(ERR,
2146                 "unable to append \"-glue\" to last component of"
2147                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
2148                 " please re-configure DPDK");
2149         return NULL;
2150 }
2151
2152 /**
2153  * Initialization routine for run-time dependency on rdma-core.
2154  */
2155 static int
2156 mlx5_glue_init(void)
2157 {
2158         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
2159         const char *path[] = {
2160                 /*
2161                  * A basic security check is necessary before trusting
2162                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
2163                  */
2164                 (geteuid() == getuid() && getegid() == getgid() ?
2165                  getenv("MLX5_GLUE_PATH") : NULL),
2166                 /*
2167                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
2168                  * variant, otherwise let dlopen() look up libraries on its
2169                  * own.
2170                  */
2171                 (*RTE_EAL_PMD_PATH ?
2172                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
2173         };
2174         unsigned int i = 0;
2175         void *handle = NULL;
2176         void **sym;
2177         const char *dlmsg;
2178
2179         while (!handle && i != RTE_DIM(path)) {
2180                 const char *end;
2181                 size_t len;
2182                 int ret;
2183
2184                 if (!path[i]) {
2185                         ++i;
2186                         continue;
2187                 }
2188                 end = strpbrk(path[i], ":;");
2189                 if (!end)
2190                         end = path[i] + strlen(path[i]);
2191                 len = end - path[i];
2192                 ret = 0;
2193                 do {
2194                         char name[ret + 1];
2195
2196                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
2197                                        (int)len, path[i],
2198                                        (!len || *(end - 1) == '/') ? "" : "/");
2199                         if (ret == -1)
2200                                 break;
2201                         if (sizeof(name) != (size_t)ret + 1)
2202                                 continue;
2203                         DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
2204                                 name);
2205                         handle = dlopen(name, RTLD_LAZY);
2206                         break;
2207                 } while (1);
2208                 path[i] = end + 1;
2209                 if (!*end)
2210                         ++i;
2211         }
2212         if (!handle) {
2213                 rte_errno = EINVAL;
2214                 dlmsg = dlerror();
2215                 if (dlmsg)
2216                         DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
2217                 goto glue_error;
2218         }
2219         sym = dlsym(handle, "mlx5_glue");
2220         if (!sym || !*sym) {
2221                 rte_errno = EINVAL;
2222                 dlmsg = dlerror();
2223                 if (dlmsg)
2224                         DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
2225                 goto glue_error;
2226         }
2227         mlx5_glue = *sym;
2228         return 0;
2229 glue_error:
2230         if (handle)
2231                 dlclose(handle);
2232         DRV_LOG(WARNING,
2233                 "cannot initialize PMD due to missing run-time dependency on"
2234                 " rdma-core libraries (libibverbs, libmlx5)");
2235         return -rte_errno;
2236 }
2237
2238 #endif
2239
2240 /**
2241  * Driver initialization routine.
2242  */
2243 RTE_INIT(rte_mlx5_pmd_init)
2244 {
2245         /* Initialize driver log type. */
2246         mlx5_logtype = rte_log_register("pmd.net.mlx5");
2247         if (mlx5_logtype >= 0)
2248                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
2249
2250         /* Build the static tables for Verbs conversion. */
2251         mlx5_set_ptype_table();
2252         mlx5_set_cksum_table();
2253         mlx5_set_swp_types_table();
2254         /*
2255          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
2256          * huge pages. Calling ibv_fork_init() during init allows
2257          * applications to use fork() safely for purposes other than
2258          * using this PMD, which is not supported in forked processes.
2259          */
2260         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
2261         /* Match the size of Rx completion entry to the size of a cacheline. */
2262         if (RTE_CACHE_LINE_SIZE == 128)
2263                 setenv("MLX5_CQE_SIZE", "128", 0);
2264         /*
2265          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
2266          * cleanup all the Verbs resources even when the device was removed.
2267          */
2268         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
2269 #ifdef RTE_IBVERBS_LINK_DLOPEN
2270         if (mlx5_glue_init())
2271                 return;
2272         assert(mlx5_glue);
2273 #endif
2274 #ifndef NDEBUG
2275         /* Glue structure must not contain any NULL pointers. */
2276         {
2277                 unsigned int i;
2278
2279                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
2280                         assert(((const void *const *)mlx5_glue)[i]);
2281         }
2282 #endif
2283         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
2284                 DRV_LOG(ERR,
2285                         "rdma-core glue \"%s\" mismatch: \"%s\" is required",
2286                         mlx5_glue->version, MLX5_GLUE_VERSION);
2287                 return;
2288         }
2289         mlx5_glue->fork_init();
2290         rte_pci_register(&mlx5_driver);
2291 }
2292
2293 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2294 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2295 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");