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 };
 151
 152 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
 153 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
 154
 155 /**
 156  * Allocate shared IB device context. If there is multiport device the
 157  * master and representors will share this context, if there is single
 158  * port dedicated IB device, the context will be used by only given
 159  * port due to unification.
 160  *
 161  * Routine first searches the context for the specified IB device name,
 162  * if found the shared context assumed and reference counter is incremented.
 163  * If no context found the new one is created and initialized with specified
 164  * IB device context and parameters.
 165  *
 166  * @param[in] spawn
 167  *   Pointer to the IB device attributes (name, port, etc).
 168  *
 169  * @return
 170  *   Pointer to mlx5_ibv_shared object on success,
 171  *   otherwise NULL and rte_errno is set.
 172  */
 173 static struct mlx5_ibv_shared *
 174 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn)
 175 {
 176         struct mlx5_ibv_shared *sh;
 177         int err = 0;
 178         uint32_t i;
 179
 180         assert(spawn);
 181         /* Secondary process should not create the shared context. */
 182         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 183         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 184         /* Search for IB context by device name. */
 185         LIST_FOREACH(sh, &mlx5_ibv_list, next) {
 186                 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
 187                         sh->refcnt++;
 188                         goto exit;
 189                 }
 190         }
 191         /* No device found, we have to create new shared context. */
 192         assert(spawn->max_port);
 193         sh = rte_zmalloc("ethdev shared ib context",
 194                          sizeof(struct mlx5_ibv_shared) +
 195                          spawn->max_port *
 196                          sizeof(struct mlx5_ibv_shared_port),
 197                          RTE_CACHE_LINE_SIZE);
 198         if (!sh) {
 199                 DRV_LOG(ERR, "shared context allocation failure");
 200                 rte_errno  = ENOMEM;
 201                 goto exit;
 202         }
 203         /* Try to open IB device with DV first, then usual Verbs. */
 204         errno = 0;
 205         sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
 206         if (sh->ctx) {
 207                 sh->devx = 1;
 208                 DRV_LOG(DEBUG, "DevX is supported");
 209         } else {
 210                 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
 211                 if (!sh->ctx) {
 212                         err = errno ? errno : ENODEV;
 213                         goto error;
 214                 }
 215                 DRV_LOG(DEBUG, "DevX is NOT supported");
 216         }
 217         err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
 218         if (err) {
 219                 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
 220                 goto error;
 221         }
 222         sh->refcnt = 1;
 223         sh->max_port = spawn->max_port;
 224         strncpy(sh->ibdev_name, sh->ctx->device->name,
 225                 sizeof(sh->ibdev_name));
 226         strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
 227                 sizeof(sh->ibdev_path));
 228         pthread_mutex_init(&sh->intr_mutex, NULL);
 229         /*
 230          * Setting port_id to max unallowed value means
 231          * there is no interrupt subhandler installed for
 232          * the given port index i.
 233          */
 234         for (i = 0; i < sh->max_port; i++)
 235                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
 236         sh->pd = mlx5_glue->alloc_pd(sh->ctx);
 237         if (sh->pd == NULL) {
 238                 DRV_LOG(ERR, "PD allocation failure");
 239                 err = ENOMEM;
 240                 goto error;
 241         }
 242         LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
 243 exit:
 244         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 245         return sh;
 246 error:
 247         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 248         assert(sh);
 249         if (sh->pd)
 250                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
 251         if (sh->ctx)
 252                 claim_zero(mlx5_glue->close_device(sh->ctx));
 253         rte_free(sh);
 254         assert(err > 0);
 255         rte_errno = err;
 256         return NULL;
 257 }
 258
 259 /**
 260  * Free shared IB device context. Decrement counter and if zero free
 261  * all allocated resources and close handles.
 262  *
 263  * @param[in] sh
 264  *   Pointer to mlx5_ibv_shared object to free
 265  */
 266 static void
 267 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
 268 {
 269         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 270 #ifndef NDEBUG
 271         /* Check the object presence in the list. */
 272         struct mlx5_ibv_shared *lctx;
 273
 274         LIST_FOREACH(lctx, &mlx5_ibv_list, next)
 275                 if (lctx == sh)
 276                         break;
 277         assert(lctx);
 278         if (lctx != sh) {
 279                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
 280                 goto exit;
 281         }
 282 #endif
 283         assert(sh);
 284         assert(sh->refcnt);
 285         /* Secondary process should not free the shared context. */
 286         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 287         if (--sh->refcnt)
 288                 goto exit;
 289         LIST_REMOVE(sh, next);
 290         /*
 291          *  Ensure there is no async event handler installed.
 292          *  Only primary process handles async device events.
 293          **/
 294         assert(!sh->intr_cnt);
 295         if (sh->intr_cnt)
 296                 rte_intr_callback_unregister
 297                         (&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
 298         pthread_mutex_destroy(&sh->intr_mutex);
 299         if (sh->pd)
 300                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
 301         if (sh->ctx)
 302                 claim_zero(mlx5_glue->close_device(sh->ctx));
 303         rte_free(sh);
 304 exit:
 305         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 306 }
 307
 308 /**
 309  * Initialize DR related data within private structure.
 310  * Routine checks the reference counter and does actual
 311  * resources creation/initialization only if counter is zero.
 312  *
 313  * @param[in] priv
 314  *   Pointer to the private device data structure.
 315  *
 316  * @return
 317  *   Zero on success, positive error code otherwise.
 318  */
 319 static int
 320 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 321 {
 322 #ifdef HAVE_MLX5DV_DR
 323         struct mlx5_ibv_shared *sh = priv->sh;
 324         int err = 0;
 325         void *ns;
 326
 327         assert(sh);
 328         if (sh->dv_refcnt) {
 329                 /* Shared DV/DR structures is already initialized. */
 330                 sh->dv_refcnt++;
 331                 priv->dr_shared = 1;
 332                 return 0;
 333         }
 334         /* Reference counter is zero, we should initialize structures. */
 335         ns = mlx5_glue->dr_create_ns(sh->ctx,
 336                                      MLX5DV_DR_NS_DOMAIN_INGRESS_BYPASS);
 337         if (!ns) {
 338                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_ns failed");
 339                 err = errno;
 340                 goto error;
 341         }
 342         sh->rx_ns = ns;
 343         ns = mlx5_glue->dr_create_ns(sh->ctx,
 344                                      MLX5DV_DR_NS_DOMAIN_EGRESS_BYPASS);
 345         if (!ns) {
 346                 DRV_LOG(ERR, "egress mlx5dv_dr_create_ns failed");
 347                 err = errno;
 348                 goto error;
 349         }
 350         pthread_mutex_init(&sh->dv_mutex, NULL);
 351         sh->tx_ns = ns;
 352 #ifdef HAVE_MLX5DV_DR_ESWITCH
 353         if (priv->config.dv_esw_en) {
 354                 ns  = mlx5_glue->dr_create_ns(sh->ctx,
 355                                               MLX5DV_DR_NS_DOMAIN_FDB_BYPASS);
 356                 if (!ns) {
 357                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_ns failed");
 358                         err = errno;
 359                         goto error;
 360                 }
 361                 sh->fdb_ns = ns;
 362                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 363         }
 364 #endif
 365         sh->dv_refcnt++;
 366         priv->dr_shared = 1;
 367         return 0;
 368
 369 error:
 370        /* Rollback the created objects. */
 371         if (sh->rx_ns) {
 372                 mlx5_glue->dr_destroy_ns(sh->rx_ns);
 373                 sh->rx_ns = NULL;
 374         }
 375         if (sh->tx_ns) {
 376                 mlx5_glue->dr_destroy_ns(sh->tx_ns);
 377                 sh->tx_ns = NULL;
 378         }
 379         if (sh->fdb_ns) {
 380                 mlx5_glue->dr_destroy_ns(sh->fdb_ns);
 381                 sh->fdb_ns = NULL;
 382         }
 383         if (sh->esw_drop_action) {
 384                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 385                 sh->esw_drop_action = NULL;
 386         }
 387         return err;
 388 #else
 389         (void)priv;
 390         return 0;
 391 #endif
 392 }
 393
 394 /**
 395  * Destroy DR related data within private structure.
 396  *
 397  * @param[in] priv
 398  *   Pointer to the private device data structure.
 399  */
 400 static void
 401 mlx5_free_shared_dr(struct mlx5_priv *priv)
 402 {
 403 #ifdef HAVE_MLX5DV_DR
 404         struct mlx5_ibv_shared *sh;
 405
 406         if (!priv->dr_shared)
 407                 return;
 408         priv->dr_shared = 0;
 409         sh = priv->sh;
 410         assert(sh);
 411         assert(sh->dv_refcnt);
 412         if (sh->dv_refcnt && --sh->dv_refcnt)
 413                 return;
 414         if (sh->rx_ns) {
 415                 mlx5_glue->dr_destroy_ns(sh->rx_ns);
 416                 sh->rx_ns = NULL;
 417         }
 418         if (sh->tx_ns) {
 419                 mlx5_glue->dr_destroy_ns(sh->tx_ns);
 420                 sh->tx_ns = NULL;
 421         }
 422 #ifdef HAVE_MLX5DV_DR_ESWITCH
 423         if (sh->fdb_ns) {
 424                 mlx5_glue->dr_destroy_ns(sh->fdb_ns);
 425                 sh->fdb_ns = NULL;
 426         }
 427         if (sh->esw_drop_action) {
 428                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 429                 sh->esw_drop_action = NULL;
 430         }
 431 #endif
 432         pthread_mutex_destroy(&sh->dv_mutex);
 433 #else
 434         (void)priv;
 435 #endif
 436 }
 437
 438 /**
 439  * Initialize shared data between primary and secondary process.
 440  *
 441  * A memzone is reserved by primary process and secondary processes attach to
 442  * the memzone.
 443  *
 444  * @return
 445  *   0 on success, a negative errno value otherwise and rte_errno is set.
 446  */
 447 static int
 448 mlx5_init_shared_data(void)
 449 {
 450         const struct rte_memzone *mz;
 451         int ret = 0;
 452
 453         rte_spinlock_lock(&mlx5_shared_data_lock);
 454         if (mlx5_shared_data == NULL) {
 455                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 456                         /* Allocate shared memory. */
 457                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
 458                                                  sizeof(*mlx5_shared_data),
 459                                                  SOCKET_ID_ANY, 0);
 460                         if (mz == NULL) {
 461                                 DRV_LOG(ERR,
 462                                         "Cannot allocate mlx5 shared data\n");
 463                                 ret = -rte_errno;
 464                                 goto error;
 465                         }
 466                         mlx5_shared_data = mz->addr;
 467                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
 468                         rte_spinlock_init(&mlx5_shared_data->lock);
 469                 } else {
 470                         /* Lookup allocated shared memory. */
 471                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
 472                         if (mz == NULL) {
 473                                 DRV_LOG(ERR,
 474                                         "Cannot attach mlx5 shared data\n");
 475                                 ret = -rte_errno;
 476                                 goto error;
 477                         }
 478                         mlx5_shared_data = mz->addr;
 479                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
 480                 }
 481         }
 482 error:
 483         rte_spinlock_unlock(&mlx5_shared_data_lock);
 484         return ret;
 485 }
 486
 487 /**
 488  * Retrieve integer value from environment variable.
 489  *
 490  * @param[in] name
 491  *   Environment variable name.
 492  *
 493  * @return
 494  *   Integer value, 0 if the variable is not set.
 495  */
 496 int
 497 mlx5_getenv_int(const char *name)
 498 {
 499         const char *val = getenv(name);
 500
 501         if (val == NULL)
 502                 return 0;
 503         return atoi(val);
 504 }
 505
 506 /**
 507  * Verbs callback to allocate a memory. This function should allocate the space
 508  * according to the size provided residing inside a huge page.
 509  * Please note that all allocation must respect the alignment from libmlx5
 510  * (i.e. currently sysconf(_SC_PAGESIZE)).
 511  *
 512  * @param[in] size
 513  *   The size in bytes of the memory to allocate.
 514  * @param[in] data
 515  *   A pointer to the callback data.
 516  *
 517  * @return
 518  *   Allocated buffer, NULL otherwise and rte_errno is set.
 519  */
 520 static void *
 521 mlx5_alloc_verbs_buf(size_t size, void *data)
 522 {
 523         struct mlx5_priv *priv = data;
 524         void *ret;
 525         size_t alignment = sysconf(_SC_PAGESIZE);
 526         unsigned int socket = SOCKET_ID_ANY;
 527
 528         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 529                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 530
 531                 socket = ctrl->socket;
 532         } else if (priv->verbs_alloc_ctx.type ==
 533                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 534                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 535
 536                 socket = ctrl->socket;
 537         }
 538         assert(data != NULL);
 539         ret = rte_malloc_socket(__func__, size, alignment, socket);
 540         if (!ret && size)
 541                 rte_errno = ENOMEM;
 542         return ret;
 543 }
 544
 545 /**
 546  * Verbs callback to free a memory.
 547  *
 548  * @param[in] ptr
 549  *   A pointer to the memory to free.
 550  * @param[in] data
 551  *   A pointer to the callback data.
 552  */
 553 static void
 554 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 555 {
 556         assert(data != NULL);
 557         rte_free(ptr);
 558 }
 559
 560 /**
 561  * Initialize process private data structure.
 562  *
 563  * @param dev
 564  *   Pointer to Ethernet device structure.
 565  *
 566  * @return
 567  *   0 on success, a negative errno value otherwise and rte_errno is set.
 568  */
 569 int
 570 mlx5_proc_priv_init(struct rte_eth_dev *dev)
 571 {
 572         struct mlx5_priv *priv = dev->data->dev_private;
 573         struct mlx5_proc_priv *ppriv;
 574         size_t ppriv_size;
 575
 576         /*
 577          * UAR register table follows the process private structure. BlueFlame
 578          * registers for Tx queues are stored in the table.
 579          */
 580         ppriv_size =
 581                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
 582         ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
 583                                   RTE_CACHE_LINE_SIZE, dev->device->numa_node);
 584         if (!ppriv) {
 585                 rte_errno = ENOMEM;
 586                 return -rte_errno;
 587         }
 588         ppriv->uar_table_sz = ppriv_size;
 589         dev->process_private = ppriv;
 590         return 0;
 591 }
 592
 593 /**
 594  * Un-initialize process private data structure.
 595  *
 596  * @param dev
 597  *   Pointer to Ethernet device structure.
 598  */
 599 static void
 600 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
 601 {
 602         if (!dev->process_private)
 603                 return;
 604         rte_free(dev->process_private);
 605         dev->process_private = NULL;
 606 }
 607
 608 /**
 609  * DPDK callback to close the device.
 610  *
 611  * Destroy all queues and objects, free memory.
 612  *
 613  * @param dev
 614  *   Pointer to Ethernet device structure.
 615  */
 616 static void
 617 mlx5_dev_close(struct rte_eth_dev *dev)
 618 {
 619         struct mlx5_priv *priv = dev->data->dev_private;
 620         unsigned int i;
 621         int ret;
 622
 623         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
 624                 dev->data->port_id,
 625                 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
 626         /* In case mlx5_dev_stop() has not been called. */
 627         mlx5_dev_interrupt_handler_uninstall(dev);
 628         mlx5_traffic_disable(dev);
 629         mlx5_flow_flush(dev, NULL);
 630         /* Prevent crashes when queues are still in use. */
 631         dev->rx_pkt_burst = removed_rx_burst;
 632         dev->tx_pkt_burst = removed_tx_burst;
 633         rte_wmb();
 634         /* Disable datapath on secondary process. */
 635         mlx5_mp_req_stop_rxtx(dev);
 636         if (priv->rxqs != NULL) {
 637                 /* XXX race condition if mlx5_rx_burst() is still running. */
 638                 usleep(1000);
 639                 for (i = 0; (i != priv->rxqs_n); ++i)
 640                         mlx5_rxq_release(dev, i);
 641                 priv->rxqs_n = 0;
 642                 priv->rxqs = NULL;
 643         }
 644         if (priv->txqs != NULL) {
 645                 /* XXX race condition if mlx5_tx_burst() is still running. */
 646                 usleep(1000);
 647                 for (i = 0; (i != priv->txqs_n); ++i)
 648                         mlx5_txq_release(dev, i);
 649                 priv->txqs_n = 0;
 650                 priv->txqs = NULL;
 651         }
 652         mlx5_proc_priv_uninit(dev);
 653         mlx5_mprq_free_mp(dev);
 654         mlx5_mr_release(dev);
 655         assert(priv->sh);
 656         mlx5_free_shared_dr(priv);
 657         if (priv->rss_conf.rss_key != NULL)
 658                 rte_free(priv->rss_conf.rss_key);
 659         if (priv->reta_idx != NULL)
 660                 rte_free(priv->reta_idx);
 661         if (priv->config.vf)
 662                 mlx5_nl_mac_addr_flush(dev);
 663         if (priv->nl_socket_route >= 0)
 664                 close(priv->nl_socket_route);
 665         if (priv->nl_socket_rdma >= 0)
 666                 close(priv->nl_socket_rdma);
 667         if (priv->tcf_context)
 668                 mlx5_flow_tcf_context_destroy(priv->tcf_context);
 669         if (priv->sh) {
 670                 /*
 671                  * Free the shared context in last turn, because the cleanup
 672                  * routines above may use some shared fields, like
 673                  * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
 674                  * ifindex if Netlink fails.
 675                  */
 676                 mlx5_free_shared_ibctx(priv->sh);
 677                 priv->sh = NULL;
 678         }
 679         ret = mlx5_hrxq_ibv_verify(dev);
 680         if (ret)
 681                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
 682                         dev->data->port_id);
 683         ret = mlx5_ind_table_ibv_verify(dev);
 684         if (ret)
 685                 DRV_LOG(WARNING, "port %u some indirection table still remain",
 686                         dev->data->port_id);
 687         ret = mlx5_rxq_ibv_verify(dev);
 688         if (ret)
 689                 DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
 690                         dev->data->port_id);
 691         ret = mlx5_rxq_verify(dev);
 692         if (ret)
 693                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
 694                         dev->data->port_id);
 695         ret = mlx5_txq_ibv_verify(dev);
 696         if (ret)
 697                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
 698                         dev->data->port_id);
 699         ret = mlx5_txq_verify(dev);
 700         if (ret)
 701                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
 702                         dev->data->port_id);
 703         ret = mlx5_flow_verify(dev);
 704         if (ret)
 705                 DRV_LOG(WARNING, "port %u some flows still remain",
 706                         dev->data->port_id);
 707         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 708                 unsigned int c = 0;
 709                 uint16_t port_id;
 710
 711                 RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
 712                         struct mlx5_priv *opriv =
 713                                 rte_eth_devices[port_id].data->dev_private;
 714
 715                         if (!opriv ||
 716                             opriv->domain_id != priv->domain_id ||
 717                             &rte_eth_devices[port_id] == dev)
 718                                 continue;
 719                         ++c;
 720                 }
 721                 if (!c)
 722                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
 723         }
 724         memset(priv, 0, sizeof(*priv));
 725         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 726         /*
 727          * Reset mac_addrs to NULL such that it is not freed as part of
 728          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
 729          * it is freed when dev_private is freed.
 730          */
 731         dev->data->mac_addrs = NULL;
 732 }
 733
 734 const struct eth_dev_ops mlx5_dev_ops = {
 735         .dev_configure = mlx5_dev_configure,
 736         .dev_start = mlx5_dev_start,
 737         .dev_stop = mlx5_dev_stop,
 738         .dev_set_link_down = mlx5_set_link_down,
 739         .dev_set_link_up = mlx5_set_link_up,
 740         .dev_close = mlx5_dev_close,
 741         .promiscuous_enable = mlx5_promiscuous_enable,
 742         .promiscuous_disable = mlx5_promiscuous_disable,
 743         .allmulticast_enable = mlx5_allmulticast_enable,
 744         .allmulticast_disable = mlx5_allmulticast_disable,
 745         .link_update = mlx5_link_update,
 746         .stats_get = mlx5_stats_get,
 747         .stats_reset = mlx5_stats_reset,
 748         .xstats_get = mlx5_xstats_get,
 749         .xstats_reset = mlx5_xstats_reset,
 750         .xstats_get_names = mlx5_xstats_get_names,
 751         .fw_version_get = mlx5_fw_version_get,
 752         .dev_infos_get = mlx5_dev_infos_get,
 753         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 754         .vlan_filter_set = mlx5_vlan_filter_set,
 755         .rx_queue_setup = mlx5_rx_queue_setup,
 756         .tx_queue_setup = mlx5_tx_queue_setup,
 757         .rx_queue_release = mlx5_rx_queue_release,
 758         .tx_queue_release = mlx5_tx_queue_release,
 759         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 760         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 761         .mac_addr_remove = mlx5_mac_addr_remove,
 762         .mac_addr_add = mlx5_mac_addr_add,
 763         .mac_addr_set = mlx5_mac_addr_set,
 764         .set_mc_addr_list = mlx5_set_mc_addr_list,
 765         .mtu_set = mlx5_dev_set_mtu,
 766         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 767         .vlan_offload_set = mlx5_vlan_offload_set,
 768         .reta_update = mlx5_dev_rss_reta_update,
 769         .reta_query = mlx5_dev_rss_reta_query,
 770         .rss_hash_update = mlx5_rss_hash_update,
 771         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
 772         .filter_ctrl = mlx5_dev_filter_ctrl,
 773         .rx_descriptor_status = mlx5_rx_descriptor_status,
 774         .tx_descriptor_status = mlx5_tx_descriptor_status,
 775         .rx_queue_count = mlx5_rx_queue_count,
 776         .rx_queue_intr_enable = mlx5_rx_intr_enable,
 777         .rx_queue_intr_disable = mlx5_rx_intr_disable,
 778         .is_removed = mlx5_is_removed,
 779 };
 780
 781 /* Available operations from secondary process. */
 782 static const struct eth_dev_ops mlx5_dev_sec_ops = {
 783         .stats_get = mlx5_stats_get,
 784         .stats_reset = mlx5_stats_reset,
 785         .xstats_get = mlx5_xstats_get,
 786         .xstats_reset = mlx5_xstats_reset,
 787         .xstats_get_names = mlx5_xstats_get_names,
 788         .fw_version_get = mlx5_fw_version_get,
 789         .dev_infos_get = mlx5_dev_infos_get,
 790         .rx_descriptor_status = mlx5_rx_descriptor_status,
 791         .tx_descriptor_status = mlx5_tx_descriptor_status,
 792 };
 793
 794 /* Available operations in flow isolated mode. */
 795 const struct eth_dev_ops mlx5_dev_ops_isolate = {
 796         .dev_configure = mlx5_dev_configure,
 797         .dev_start = mlx5_dev_start,
 798         .dev_stop = mlx5_dev_stop,
 799         .dev_set_link_down = mlx5_set_link_down,
 800         .dev_set_link_up = mlx5_set_link_up,
 801         .dev_close = mlx5_dev_close,
 802         .promiscuous_enable = mlx5_promiscuous_enable,
 803         .promiscuous_disable = mlx5_promiscuous_disable,
 804         .allmulticast_enable = mlx5_allmulticast_enable,
 805         .allmulticast_disable = mlx5_allmulticast_disable,
 806         .link_update = mlx5_link_update,
 807         .stats_get = mlx5_stats_get,
 808         .stats_reset = mlx5_stats_reset,
 809         .xstats_get = mlx5_xstats_get,
 810         .xstats_reset = mlx5_xstats_reset,
 811         .xstats_get_names = mlx5_xstats_get_names,
 812         .fw_version_get = mlx5_fw_version_get,
 813         .dev_infos_get = mlx5_dev_infos_get,
 814         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 815         .vlan_filter_set = mlx5_vlan_filter_set,
 816         .rx_queue_setup = mlx5_rx_queue_setup,
 817         .tx_queue_setup = mlx5_tx_queue_setup,
 818         .rx_queue_release = mlx5_rx_queue_release,
 819         .tx_queue_release = mlx5_tx_queue_release,
 820         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 821         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 822         .mac_addr_remove = mlx5_mac_addr_remove,
 823         .mac_addr_add = mlx5_mac_addr_add,
 824         .mac_addr_set = mlx5_mac_addr_set,
 825         .set_mc_addr_list = mlx5_set_mc_addr_list,
 826         .mtu_set = mlx5_dev_set_mtu,
 827         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 828         .vlan_offload_set = mlx5_vlan_offload_set,
 829         .filter_ctrl = mlx5_dev_filter_ctrl,
 830         .rx_descriptor_status = mlx5_rx_descriptor_status,
 831         .tx_descriptor_status = mlx5_tx_descriptor_status,
 832         .rx_queue_intr_enable = mlx5_rx_intr_enable,
 833         .rx_queue_intr_disable = mlx5_rx_intr_disable,
 834         .is_removed = mlx5_is_removed,
 835 };
 836
 837 /**
 838  * Verify and store value for device argument.
 839  *
 840  * @param[in] key
 841  *   Key argument to verify.
 842  * @param[in] val
 843  *   Value associated with key.
 844  * @param opaque
 845  *   User data.
 846  *
 847  * @return
 848  *   0 on success, a negative errno value otherwise and rte_errno is set.
 849  */
 850 static int
 851 mlx5_args_check(const char *key, const char *val, void *opaque)
 852 {
 853         struct mlx5_dev_config *config = opaque;
 854         unsigned long tmp;
 855
 856         /* No-op, port representors are processed in mlx5_dev_spawn(). */
 857         if (!strcmp(MLX5_REPRESENTOR, key))
 858                 return 0;
 859         errno = 0;
 860         tmp = strtoul(val, NULL, 0);
 861         if (errno) {
 862                 rte_errno = errno;
 863                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
 864                 return -rte_errno;
 865         }
 866         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
 867                 config->cqe_comp = !!tmp;
 868         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
 869                 config->cqe_pad = !!tmp;
 870         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
 871                 config->hw_padding = !!tmp;
 872         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
 873                 config->mprq.enabled = !!tmp;
 874         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
 875                 config->mprq.stride_num_n = tmp;
 876         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
 877                 config->mprq.max_memcpy_len = tmp;
 878         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
 879                 config->mprq.min_rxqs_num = tmp;
 880         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
 881                 config->txq_inline = tmp;
 882         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
 883                 config->txqs_inline = tmp;
 884         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
 885                 config->txqs_vec = tmp;
 886         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
 887                 config->mps = !!tmp;
 888         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
 889                 config->mpw_hdr_dseg = !!tmp;
 890         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
 891                 config->inline_max_packet_sz = tmp;
 892         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
 893                 config->tx_vec_en = !!tmp;
 894         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
 895                 config->rx_vec_en = !!tmp;
 896         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
 897                 config->l3_vxlan_en = !!tmp;
 898         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
 899                 config->vf_nl_en = !!tmp;
 900         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
 901                 config->dv_esw_en = !!tmp;
 902         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
 903                 config->dv_flow_en = !!tmp;
 904         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
 905                 config->mr_ext_memseg_en = !!tmp;
 906         } else {
 907                 DRV_LOG(WARNING, "%s: unknown parameter", key);
 908                 rte_errno = EINVAL;
 909                 return -rte_errno;
 910         }
 911         return 0;
 912 }
 913
 914 /**
 915  * Parse device parameters.
 916  *
 917  * @param config
 918  *   Pointer to device configuration structure.
 919  * @param devargs
 920  *   Device arguments structure.
 921  *
 922  * @return
 923  *   0 on success, a negative errno value otherwise and rte_errno is set.
 924  */
 925 static int
 926 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
 927 {
 928         const char **params = (const char *[]){
 929                 MLX5_RXQ_CQE_COMP_EN,
 930                 MLX5_RXQ_CQE_PAD_EN,
 931                 MLX5_RXQ_PKT_PAD_EN,
 932                 MLX5_RX_MPRQ_EN,
 933                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
 934                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
 935                 MLX5_RXQS_MIN_MPRQ,
 936                 MLX5_TXQ_INLINE,
 937                 MLX5_TXQS_MIN_INLINE,
 938                 MLX5_TXQS_MAX_VEC,
 939                 MLX5_TXQ_MPW_EN,
 940                 MLX5_TXQ_MPW_HDR_DSEG_EN,
 941                 MLX5_TXQ_MAX_INLINE_LEN,
 942                 MLX5_TX_VEC_EN,
 943                 MLX5_RX_VEC_EN,
 944                 MLX5_L3_VXLAN_EN,
 945                 MLX5_VF_NL_EN,
 946                 MLX5_DV_ESW_EN,
 947                 MLX5_DV_FLOW_EN,
 948                 MLX5_MR_EXT_MEMSEG_EN,
 949                 MLX5_REPRESENTOR,
 950                 NULL,
 951         };
 952         struct rte_kvargs *kvlist;
 953         int ret = 0;
 954         int i;
 955
 956         if (devargs == NULL)
 957                 return 0;
 958         /* Following UGLY cast is done to pass checkpatch. */
 959         kvlist = rte_kvargs_parse(devargs->args, params);
 960         if (kvlist == NULL)
 961                 return 0;
 962         /* Process parameters. */
 963         for (i = 0; (params[i] != NULL); ++i) {
 964                 if (rte_kvargs_count(kvlist, params[i])) {
 965                         ret = rte_kvargs_process(kvlist, params[i],
 966                                                  mlx5_args_check, config);
 967                         if (ret) {
 968                                 rte_errno = EINVAL;
 969                                 rte_kvargs_free(kvlist);
 970                                 return -rte_errno;
 971                         }
 972                 }
 973         }
 974         rte_kvargs_free(kvlist);
 975         return 0;
 976 }
 977
 978 static struct rte_pci_driver mlx5_driver;
 979
 980 /**
 981  * PMD global initialization.
 982  *
 983  * Independent from individual device, this function initializes global
 984  * per-PMD data structures distinguishing primary and secondary processes.
 985  * Hence, each initialization is called once per a process.
 986  *
 987  * @return
 988  *   0 on success, a negative errno value otherwise and rte_errno is set.
 989  */
 990 static int
 991 mlx5_init_once(void)
 992 {
 993         struct mlx5_shared_data *sd;
 994         struct mlx5_local_data *ld = &mlx5_local_data;
 995
 996         if (mlx5_init_shared_data())
 997                 return -rte_errno;
 998         sd = mlx5_shared_data;
 999         assert(sd);
1000         rte_spinlock_lock(&sd->lock);
1001         switch (rte_eal_process_type()) {
1002         case RTE_PROC_PRIMARY:
1003                 if (sd->init_done)
1004                         break;
1005                 LIST_INIT(&sd->mem_event_cb_list);
1006                 rte_rwlock_init(&sd->mem_event_rwlock);
1007                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1008                                                 mlx5_mr_mem_event_cb, NULL);
1009                 mlx5_mp_init_primary();
1010                 sd->init_done = true;
1011                 break;
1012         case RTE_PROC_SECONDARY:
1013                 if (ld->init_done)
1014                         break;
1015                 mlx5_mp_init_secondary();
1016                 ++sd->secondary_cnt;
1017                 ld->init_done = true;
1018                 break;
1019         default:
1020                 break;
1021         }
1022         rte_spinlock_unlock(&sd->lock);
1023         return 0;
1024 }
1025
1026 /**
1027  * Spawn an Ethernet device from Verbs information.
1028  *
1029  * @param dpdk_dev
1030  *   Backing DPDK device.
1031  * @param spawn
1032  *   Verbs device parameters (name, port, switch_info) to spawn.
1033  * @param config
1034  *   Device configuration parameters.
1035  *
1036  * @return
1037  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1038  *   is set. The following errors are defined:
1039  *
1040  *   EBUSY: device is not supposed to be spawned.
1041  *   EEXIST: device is already spawned
1042  */
1043 static struct rte_eth_dev *
1044 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1045                struct mlx5_dev_spawn_data *spawn,
1046                struct mlx5_dev_config config)
1047 {
1048         const struct mlx5_switch_info *switch_info = &spawn->info;
1049         struct mlx5_ibv_shared *sh = NULL;
1050         struct ibv_port_attr port_attr;
1051         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1052         struct rte_eth_dev *eth_dev = NULL;
1053         struct mlx5_priv *priv = NULL;
1054         int err = 0;
1055         unsigned int hw_padding = 0;
1056         unsigned int mps;
1057         unsigned int cqe_comp;
1058         unsigned int cqe_pad = 0;
1059         unsigned int tunnel_en = 0;
1060         unsigned int mpls_en = 0;
1061         unsigned int swp = 0;
1062         unsigned int mprq = 0;
1063         unsigned int mprq_min_stride_size_n = 0;
1064         unsigned int mprq_max_stride_size_n = 0;
1065         unsigned int mprq_min_stride_num_n = 0;
1066         unsigned int mprq_max_stride_num_n = 0;
1067         struct ether_addr mac;
1068         char name[RTE_ETH_NAME_MAX_LEN];
1069         int own_domain_id = 0;
1070         uint16_t port_id;
1071         unsigned int i;
1072
1073         /* Determine if this port representor is supposed to be spawned. */
1074         if (switch_info->representor && dpdk_dev->devargs) {
1075                 struct rte_eth_devargs eth_da;
1076
1077                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
1078                 if (err) {
1079                         rte_errno = -err;
1080                         DRV_LOG(ERR, "failed to process device arguments: %s",
1081                                 strerror(rte_errno));
1082                         return NULL;
1083                 }
1084                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
1085                         if (eth_da.representor_ports[i] ==
1086                             (uint16_t)switch_info->port_name)
1087                                 break;
1088                 if (i == eth_da.nb_representor_ports) {
1089                         rte_errno = EBUSY;
1090                         return NULL;
1091                 }
1092         }
1093         /* Build device name. */
1094         if (!switch_info->representor)
1095                 strlcpy(name, dpdk_dev->name, sizeof(name));
1096         else
1097                 snprintf(name, sizeof(name), "%s_representor_%u",
1098                          dpdk_dev->name, switch_info->port_name);
1099         /* check if the device is already spawned */
1100         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1101                 rte_errno = EEXIST;
1102                 return NULL;
1103         }
1104         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1105         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1106                 eth_dev = rte_eth_dev_attach_secondary(name);
1107                 if (eth_dev == NULL) {
1108                         DRV_LOG(ERR, "can not attach rte ethdev");
1109                         rte_errno = ENOMEM;
1110                         return NULL;
1111                 }
1112                 eth_dev->device = dpdk_dev;
1113                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
1114                 err = mlx5_proc_priv_init(eth_dev);
1115                 if (err)
1116                         return NULL;
1117                 /* Receive command fd from primary process */
1118                 err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
1119                 if (err < 0)
1120                         return NULL;
1121                 /* Remap UAR for Tx queues. */
1122                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
1123                 if (err)
1124                         return NULL;
1125                 /*
1126                  * Ethdev pointer is still required as input since
1127                  * the primary device is not accessible from the
1128                  * secondary process.
1129                  */
1130                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1131                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1132                 return eth_dev;
1133         }
1134         sh = mlx5_alloc_shared_ibctx(spawn);
1135         if (!sh)
1136                 return NULL;
1137         config.devx = sh->devx;
1138 #ifdef HAVE_IBV_MLX5_MOD_SWP
1139         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1140 #endif
1141         /*
1142          * Multi-packet send is supported by ConnectX-4 Lx PF as well
1143          * as all ConnectX-5 devices.
1144          */
1145 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1146         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1147 #endif
1148 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1149         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1150 #endif
1151         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1152         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1153                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1154                         DRV_LOG(DEBUG, "enhanced MPW is supported");
1155                         mps = MLX5_MPW_ENHANCED;
1156                 } else {
1157                         DRV_LOG(DEBUG, "MPW is supported");
1158                         mps = MLX5_MPW;
1159                 }
1160         } else {
1161                 DRV_LOG(DEBUG, "MPW isn't supported");
1162                 mps = MLX5_MPW_DISABLED;
1163         }
1164 #ifdef HAVE_IBV_MLX5_MOD_SWP
1165         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1166                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1167         DRV_LOG(DEBUG, "SWP support: %u", swp);
1168 #endif
1169         config.swp = !!swp;
1170 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1171         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1172                 struct mlx5dv_striding_rq_caps mprq_caps =
1173                         dv_attr.striding_rq_caps;
1174
1175                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1176                         mprq_caps.min_single_stride_log_num_of_bytes);
1177                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1178                         mprq_caps.max_single_stride_log_num_of_bytes);
1179                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1180                         mprq_caps.min_single_wqe_log_num_of_strides);
1181                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1182                         mprq_caps.max_single_wqe_log_num_of_strides);
1183                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1184                         mprq_caps.supported_qpts);
1185                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1186                 mprq = 1;
1187                 mprq_min_stride_size_n =
1188                         mprq_caps.min_single_stride_log_num_of_bytes;
1189                 mprq_max_stride_size_n =
1190                         mprq_caps.max_single_stride_log_num_of_bytes;
1191                 mprq_min_stride_num_n =
1192                         mprq_caps.min_single_wqe_log_num_of_strides;
1193                 mprq_max_stride_num_n =
1194                         mprq_caps.max_single_wqe_log_num_of_strides;
1195                 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1196                                                    mprq_min_stride_num_n);
1197         }
1198 #endif
1199         if (RTE_CACHE_LINE_SIZE == 128 &&
1200             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1201                 cqe_comp = 0;
1202         else
1203                 cqe_comp = 1;
1204         config.cqe_comp = cqe_comp;
1205 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1206         /* Whether device supports 128B Rx CQE padding. */
1207         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1208                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1209 #endif
1210 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1211         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1212                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
1213                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1214                              (dv_attr.tunnel_offloads_caps &
1215                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1216         }
1217         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1218                 tunnel_en ? "" : "not ");
1219 #else
1220         DRV_LOG(WARNING,
1221                 "tunnel offloading disabled due to old OFED/rdma-core version");
1222 #endif
1223         config.tunnel_en = tunnel_en;
1224 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1225         mpls_en = ((dv_attr.tunnel_offloads_caps &
1226                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1227                    (dv_attr.tunnel_offloads_caps &
1228                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1229         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1230                 mpls_en ? "" : "not ");
1231 #else
1232         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1233                 " old OFED/rdma-core version or firmware configuration");
1234 #endif
1235         config.mpls_en = mpls_en;
1236         /* Check port status. */
1237         err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1238         if (err) {
1239                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1240                 goto error;
1241         }
1242         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1243                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1244                 err = EINVAL;
1245                 goto error;
1246         }
1247         if (port_attr.state != IBV_PORT_ACTIVE)
1248                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1249                         mlx5_glue->port_state_str(port_attr.state),
1250                         port_attr.state);
1251         /* Allocate private eth device data. */
1252         priv = rte_zmalloc("ethdev private structure",
1253                            sizeof(*priv),
1254                            RTE_CACHE_LINE_SIZE);
1255         if (priv == NULL) {
1256                 DRV_LOG(ERR, "priv allocation failure");
1257                 err = ENOMEM;
1258                 goto error;
1259         }
1260         priv->sh = sh;
1261         priv->ibv_port = spawn->ibv_port;
1262         priv->mtu = ETHER_MTU;
1263 #ifndef RTE_ARCH_64
1264         /* Initialize UAR access locks for 32bit implementations. */
1265         rte_spinlock_init(&priv->uar_lock_cq);
1266         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1267                 rte_spinlock_init(&priv->uar_lock[i]);
1268 #endif
1269         /* Some internal functions rely on Netlink sockets, open them now. */
1270         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1271         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1272         priv->nl_sn = 0;
1273         priv->representor = !!switch_info->representor;
1274         priv->master = !!switch_info->master;
1275         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1276         /*
1277          * Currently we support single E-Switch per PF configurations
1278          * only and vport_id field contains the vport index for
1279          * associated VF, which is deduced from representor port name.
1280          * For example, let's have the IB device port 10, it has
1281          * attached network device eth0, which has port name attribute
1282          * pf0vf2, we can deduce the VF number as 2, and set vport index
1283          * as 3 (2+1). This assigning schema should be changed if the
1284          * multiple E-Switch instances per PF configurations or/and PCI
1285          * subfunctions are added.
1286          */
1287         priv->vport_id = switch_info->representor ?
1288                          switch_info->port_name + 1 : -1;
1289         /* representor_id field keeps the unmodified port/VF index. */
1290         priv->representor_id = switch_info->representor ?
1291                                switch_info->port_name : -1;
1292         /*
1293          * Look for sibling devices in order to reuse their switch domain
1294          * if any, otherwise allocate one.
1295          */
1296         RTE_ETH_FOREACH_DEV_OF(port_id, dpdk_dev) {
1297                 const struct mlx5_priv *opriv =
1298                         rte_eth_devices[port_id].data->dev_private;
1299
1300                 if (!opriv ||
1301                         opriv->domain_id ==
1302                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1303                         continue;
1304                 priv->domain_id = opriv->domain_id;
1305                 break;
1306         }
1307         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1308                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1309                 if (err) {
1310                         err = rte_errno;
1311                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
1312                                 strerror(rte_errno));
1313                         goto error;
1314                 }
1315                 own_domain_id = 1;
1316         }
1317         err = mlx5_args(&config, dpdk_dev->devargs);
1318         if (err) {
1319                 err = rte_errno;
1320                 DRV_LOG(ERR, "failed to process device arguments: %s",
1321                         strerror(rte_errno));
1322                 goto error;
1323         }
1324         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1325                             IBV_DEVICE_RAW_IP_CSUM);
1326         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1327                 (config.hw_csum ? "" : "not "));
1328 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1329         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1330         DRV_LOG(DEBUG, "counters are not supported");
1331 #endif
1332 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1333         if (config.dv_flow_en) {
1334                 DRV_LOG(WARNING, "DV flow is not supported");
1335                 config.dv_flow_en = 0;
1336         }
1337 #endif
1338         config.ind_table_max_size =
1339                 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1340         /*
1341          * Remove this check once DPDK supports larger/variable
1342          * indirection tables.
1343          */
1344         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1345                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1346         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1347                 config.ind_table_max_size);
1348         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1349                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1350         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1351                 (config.hw_vlan_strip ? "" : "not "));
1352         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1353                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
1354         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1355                 (config.hw_fcs_strip ? "" : "not "));
1356 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1357         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1358 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1359         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1360                         IBV_DEVICE_PCI_WRITE_END_PADDING);
1361 #endif
1362         if (config.hw_padding && !hw_padding) {
1363                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1364                 config.hw_padding = 0;
1365         } else if (config.hw_padding) {
1366                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1367         }
1368         config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1369                       (sh->device_attr.tso_caps.supported_qpts &
1370                        (1 << IBV_QPT_RAW_PACKET)));
1371         if (config.tso)
1372                 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1373         /*
1374          * MPW is disabled by default, while the Enhanced MPW is enabled
1375          * by default.
1376          */
1377         if (config.mps == MLX5_ARG_UNSET)
1378                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1379                                                           MLX5_MPW_DISABLED;
1380         else
1381                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1382         DRV_LOG(INFO, "%sMPS is %s",
1383                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1384                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1385         if (config.cqe_comp && !cqe_comp) {
1386                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1387                 config.cqe_comp = 0;
1388         }
1389         if (config.cqe_pad && !cqe_pad) {
1390                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1391                 config.cqe_pad = 0;
1392         } else if (config.cqe_pad) {
1393                 DRV_LOG(INFO, "Rx CQE padding is enabled");
1394         }
1395         if (config.mprq.enabled && mprq) {
1396                 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1397                     config.mprq.stride_num_n < mprq_min_stride_num_n) {
1398                         config.mprq.stride_num_n =
1399                                 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1400                                         mprq_min_stride_num_n);
1401                         DRV_LOG(WARNING,
1402                                 "the number of strides"
1403                                 " for Multi-Packet RQ is out of range,"
1404                                 " setting default value (%u)",
1405                                 1 << config.mprq.stride_num_n);
1406                 }
1407                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1408                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1409         } else if (config.mprq.enabled && !mprq) {
1410                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1411                 config.mprq.enabled = 0;
1412         }
1413         eth_dev = rte_eth_dev_allocate(name);
1414         if (eth_dev == NULL) {
1415                 DRV_LOG(ERR, "can not allocate rte ethdev");
1416                 err = ENOMEM;
1417                 goto error;
1418         }
1419         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1420         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1421         if (priv->representor) {
1422                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1423                 eth_dev->data->representor_id = priv->representor_id;
1424         }
1425         eth_dev->data->dev_private = priv;
1426         priv->dev_data = eth_dev->data;
1427         eth_dev->data->mac_addrs = priv->mac;
1428         eth_dev->device = dpdk_dev;
1429         /* Configure the first MAC address by default. */
1430         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1431                 DRV_LOG(ERR,
1432                         "port %u cannot get MAC address, is mlx5_en"
1433                         " loaded? (errno: %s)",
1434                         eth_dev->data->port_id, strerror(rte_errno));
1435                 err = ENODEV;
1436                 goto error;
1437         }
1438         DRV_LOG(INFO,
1439                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1440                 eth_dev->data->port_id,
1441                 mac.addr_bytes[0], mac.addr_bytes[1],
1442                 mac.addr_bytes[2], mac.addr_bytes[3],
1443                 mac.addr_bytes[4], mac.addr_bytes[5]);
1444 #ifndef NDEBUG
1445         {
1446                 char ifname[IF_NAMESIZE];
1447
1448                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1449                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1450                                 eth_dev->data->port_id, ifname);
1451                 else
1452                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1453                                 eth_dev->data->port_id);
1454         }
1455 #endif
1456         /* Get actual MTU if possible. */
1457         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1458         if (err) {
1459                 err = rte_errno;
1460                 goto error;
1461         }
1462         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1463                 priv->mtu);
1464         /* Initialize burst functions to prevent crashes before link-up. */
1465         eth_dev->rx_pkt_burst = removed_rx_burst;
1466         eth_dev->tx_pkt_burst = removed_tx_burst;
1467         eth_dev->dev_ops = &mlx5_dev_ops;
1468         /* Register MAC address. */
1469         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1470         if (config.vf && config.vf_nl_en)
1471                 mlx5_nl_mac_addr_sync(eth_dev);
1472         priv->tcf_context = mlx5_flow_tcf_context_create();
1473         if (!priv->tcf_context) {
1474                 err = -rte_errno;
1475                 DRV_LOG(WARNING,
1476                         "flow rules relying on switch offloads will not be"
1477                         " supported: cannot open libmnl socket: %s",
1478                         strerror(rte_errno));
1479         } else {
1480                 struct rte_flow_error error;
1481                 unsigned int ifindex = mlx5_ifindex(eth_dev);
1482
1483                 if (!ifindex) {
1484                         err = -rte_errno;
1485                         error.message =
1486                                 "cannot retrieve network interface index";
1487                 } else {
1488                         err = mlx5_flow_tcf_init(priv->tcf_context,
1489                                                  ifindex, &error);
1490                 }
1491                 if (err) {
1492                         DRV_LOG(WARNING,
1493                                 "flow rules relying on switch offloads will"
1494                                 " not be supported: %s: %s",
1495                                 error.message, strerror(rte_errno));
1496                         mlx5_flow_tcf_context_destroy(priv->tcf_context);
1497                         priv->tcf_context = NULL;
1498                 }
1499         }
1500         TAILQ_INIT(&priv->flows);
1501         TAILQ_INIT(&priv->ctrl_flows);
1502         /* Hint libmlx5 to use PMD allocator for data plane resources */
1503         struct mlx5dv_ctx_allocators alctr = {
1504                 .alloc = &mlx5_alloc_verbs_buf,
1505                 .free = &mlx5_free_verbs_buf,
1506                 .data = priv,
1507         };
1508         mlx5_glue->dv_set_context_attr(sh->ctx,
1509                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1510                                        (void *)((uintptr_t)&alctr));
1511         /* Bring Ethernet device up. */
1512         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1513                 eth_dev->data->port_id);
1514         mlx5_set_link_up(eth_dev);
1515         /*
1516          * Even though the interrupt handler is not installed yet,
1517          * interrupts will still trigger on the async_fd from
1518          * Verbs context returned by ibv_open_device().
1519          */
1520         mlx5_link_update(eth_dev, 0);
1521 #ifdef HAVE_IBV_DEVX_OBJ
1522         if (config.devx) {
1523                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
1524                 if (err) {
1525                         err = -err;
1526                         goto error;
1527                 }
1528         }
1529 #endif
1530 #ifdef HAVE_MLX5DV_DR_ESWITCH
1531         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1532               (switch_info->representor || switch_info->master)))
1533                 config.dv_esw_en = 0;
1534 #else
1535         config.dv_esw_en = 0;
1536 #endif
1537         /* Store device configuration on private structure. */
1538         priv->config = config;
1539         if (config.dv_flow_en) {
1540                 err = mlx5_alloc_shared_dr(priv);
1541                 if (err)
1542                         goto error;
1543         }
1544         /* Supported Verbs flow priority number detection. */
1545         err = mlx5_flow_discover_priorities(eth_dev);
1546         if (err < 0) {
1547                 err = -err;
1548                 goto error;
1549         }
1550         priv->config.flow_prio = err;
1551         /*
1552          * Once the device is added to the list of memory event
1553          * callback, its global MR cache table cannot be expanded
1554          * on the fly because of deadlock. If it overflows, lookup
1555          * should be done by searching MR list linearly, which is slow.
1556          */
1557         err = mlx5_mr_btree_init(&priv->mr.cache,
1558                                  MLX5_MR_BTREE_CACHE_N * 2,
1559                                  eth_dev->device->numa_node);
1560         if (err) {
1561                 err = rte_errno;
1562                 goto error;
1563         }
1564         /* Add device to memory callback list. */
1565         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1566         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1567                          priv, mem_event_cb);
1568         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1569         return eth_dev;
1570 error:
1571         if (priv) {
1572                 if (priv->sh)
1573                         mlx5_free_shared_dr(priv);
1574                 if (priv->nl_socket_route >= 0)
1575                         close(priv->nl_socket_route);
1576                 if (priv->nl_socket_rdma >= 0)
1577                         close(priv->nl_socket_rdma);
1578                 if (priv->tcf_context)
1579                         mlx5_flow_tcf_context_destroy(priv->tcf_context);
1580                 if (own_domain_id)
1581                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1582                 rte_free(priv);
1583                 if (eth_dev != NULL)
1584                         eth_dev->data->dev_private = NULL;
1585         }
1586         if (eth_dev != NULL) {
1587                 /* mac_addrs must not be freed alone because part of dev_private */
1588                 eth_dev->data->mac_addrs = NULL;
1589                 rte_eth_dev_release_port(eth_dev);
1590         }
1591         if (sh)
1592                 mlx5_free_shared_ibctx(sh);
1593         assert(err > 0);
1594         rte_errno = err;
1595         return NULL;
1596 }
1597
1598 /**
1599  * Comparison callback to sort device data.
1600  *
1601  * This is meant to be used with qsort().
1602  *
1603  * @param a[in]
1604  *   Pointer to pointer to first data object.
1605  * @param b[in]
1606  *   Pointer to pointer to second data object.
1607  *
1608  * @return
1609  *   0 if both objects are equal, less than 0 if the first argument is less
1610  *   than the second, greater than 0 otherwise.
1611  */
1612 static int
1613 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1614 {
1615         const struct mlx5_switch_info *si_a =
1616                 &((const struct mlx5_dev_spawn_data *)a)->info;
1617         const struct mlx5_switch_info *si_b =
1618                 &((const struct mlx5_dev_spawn_data *)b)->info;
1619         int ret;
1620
1621         /* Master device first. */
1622         ret = si_b->master - si_a->master;
1623         if (ret)
1624                 return ret;
1625         /* Then representor devices. */
1626         ret = si_b->representor - si_a->representor;
1627         if (ret)
1628                 return ret;
1629         /* Unidentified devices come last in no specific order. */
1630         if (!si_a->representor)
1631                 return 0;
1632         /* Order representors by name. */
1633         return si_a->port_name - si_b->port_name;
1634 }
1635
1636 /**
1637  * DPDK callback to register a PCI device.
1638  *
1639  * This function spawns Ethernet devices out of a given PCI device.
1640  *
1641  * @param[in] pci_drv
1642  *   PCI driver structure (mlx5_driver).
1643  * @param[in] pci_dev
1644  *   PCI device information.
1645  *
1646  * @return
1647  *   0 on success, a negative errno value otherwise and rte_errno is set.
1648  */
1649 static int
1650 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1651                struct rte_pci_device *pci_dev)
1652 {
1653         struct ibv_device **ibv_list;
1654         /*
1655          * Number of found IB Devices matching with requested PCI BDF.
1656          * nd != 1 means there are multiple IB devices over the same
1657          * PCI device and we have representors and master.
1658          */
1659         unsigned int nd = 0;
1660         /*
1661          * Number of found IB device Ports. nd = 1 and np = 1..n means
1662          * we have the single multiport IB device, and there may be
1663          * representors attached to some of found ports.
1664          */
1665         unsigned int np = 0;
1666         /*
1667          * Number of DPDK ethernet devices to Spawn - either over
1668          * multiple IB devices or multiple ports of single IB device.
1669          * Actually this is the number of iterations to spawn.
1670          */
1671         unsigned int ns = 0;
1672         struct mlx5_dev_config dev_config;
1673         int ret;
1674
1675         ret = mlx5_init_once();
1676         if (ret) {
1677                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1678                         strerror(rte_errno));
1679                 return -rte_errno;
1680         }
1681         assert(pci_drv == &mlx5_driver);
1682         errno = 0;
1683         ibv_list = mlx5_glue->get_device_list(&ret);
1684         if (!ibv_list) {
1685                 rte_errno = errno ? errno : ENOSYS;
1686                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1687                 return -rte_errno;
1688         }
1689         /*
1690          * First scan the list of all Infiniband devices to find
1691          * matching ones, gathering into the list.
1692          */
1693         struct ibv_device *ibv_match[ret + 1];
1694         int nl_route = -1;
1695         int nl_rdma = -1;
1696         unsigned int i;
1697
1698         while (ret-- > 0) {
1699                 struct rte_pci_addr pci_addr;
1700
1701                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1702                 if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1703                         continue;
1704                 if (pci_dev->addr.domain != pci_addr.domain ||
1705                     pci_dev->addr.bus != pci_addr.bus ||
1706                     pci_dev->addr.devid != pci_addr.devid ||
1707                     pci_dev->addr.function != pci_addr.function)
1708                         continue;
1709                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1710                         ibv_list[ret]->name);
1711                 ibv_match[nd++] = ibv_list[ret];
1712         }
1713         ibv_match[nd] = NULL;
1714         if (!nd) {
1715                 /* No device matches, just complain and bail out. */
1716                 mlx5_glue->free_device_list(ibv_list);
1717                 DRV_LOG(WARNING,
1718                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1719                         " are kernel drivers loaded?",
1720                         pci_dev->addr.domain, pci_dev->addr.bus,
1721                         pci_dev->addr.devid, pci_dev->addr.function);
1722                 rte_errno = ENOENT;
1723                 ret = -rte_errno;
1724                 return ret;
1725         }
1726         nl_route = mlx5_nl_init(NETLINK_ROUTE);
1727         nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1728         if (nd == 1) {
1729                 /*
1730                  * Found single matching device may have multiple ports.
1731                  * Each port may be representor, we have to check the port
1732                  * number and check the representors existence.
1733                  */
1734                 if (nl_rdma >= 0)
1735                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1736                 if (!np)
1737                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1738                                          " ports number", ibv_match[0]->name);
1739         }
1740         /*
1741          * Now we can determine the maximal
1742          * amount of devices to be spawned.
1743          */
1744         struct mlx5_dev_spawn_data list[np ? np : nd];
1745
1746         if (np > 1) {
1747                 /*
1748                  * Single IB device with multiple ports found,
1749                  * it may be E-Switch master device and representors.
1750                  * We have to perform identification trough the ports.
1751                  */
1752                 assert(nl_rdma >= 0);
1753                 assert(ns == 0);
1754                 assert(nd == 1);
1755                 for (i = 1; i <= np; ++i) {
1756                         list[ns].max_port = np;
1757                         list[ns].ibv_port = i;
1758                         list[ns].ibv_dev = ibv_match[0];
1759                         list[ns].eth_dev = NULL;
1760                         list[ns].ifindex = mlx5_nl_ifindex
1761                                         (nl_rdma, list[ns].ibv_dev->name, i);
1762                         if (!list[ns].ifindex) {
1763                                 /*
1764                                  * No network interface index found for the
1765                                  * specified port, it means there is no
1766                                  * representor on this port. It's OK,
1767                                  * there can be disabled ports, for example
1768                                  * if sriov_numvfs < sriov_totalvfs.
1769                                  */
1770                                 continue;
1771                         }
1772                         ret = -1;
1773                         if (nl_route >= 0)
1774                                 ret = mlx5_nl_switch_info
1775                                                (nl_route,
1776                                                 list[ns].ifindex,
1777                                                 &list[ns].info);
1778                         if (ret || (!list[ns].info.representor &&
1779                                     !list[ns].info.master)) {
1780                                 /*
1781                                  * We failed to recognize representors with
1782                                  * Netlink, let's try to perform the task
1783                                  * with sysfs.
1784                                  */
1785                                 ret =  mlx5_sysfs_switch_info
1786                                                 (list[ns].ifindex,
1787                                                  &list[ns].info);
1788                         }
1789                         if (!ret && (list[ns].info.representor ^
1790                                      list[ns].info.master))
1791                                 ns++;
1792                 }
1793                 if (!ns) {
1794                         DRV_LOG(ERR,
1795                                 "unable to recognize master/representors"
1796                                 " on the IB device with multiple ports");
1797                         rte_errno = ENOENT;
1798                         ret = -rte_errno;
1799                         goto exit;
1800                 }
1801         } else {
1802                 /*
1803                  * The existence of several matching entries (nd > 1) means
1804                  * port representors have been instantiated. No existing Verbs
1805                  * call nor sysfs entries can tell them apart, this can only
1806                  * be done through Netlink calls assuming kernel drivers are
1807                  * recent enough to support them.
1808                  *
1809                  * In the event of identification failure through Netlink,
1810                  * try again through sysfs, then:
1811                  *
1812                  * 1. A single IB device matches (nd == 1) with single
1813                  *    port (np=0/1) and is not a representor, assume
1814                  *    no switch support.
1815                  *
1816                  * 2. Otherwise no safe assumptions can be made;
1817                  *    complain louder and bail out.
1818                  */
1819                 np = 1;
1820                 for (i = 0; i != nd; ++i) {
1821                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1822                         list[ns].max_port = 1;
1823                         list[ns].ibv_port = 1;
1824                         list[ns].ibv_dev = ibv_match[i];
1825                         list[ns].eth_dev = NULL;
1826                         list[ns].ifindex = 0;
1827                         if (nl_rdma >= 0)
1828                                 list[ns].ifindex = mlx5_nl_ifindex
1829                                         (nl_rdma, list[ns].ibv_dev->name, 1);
1830                         if (!list[ns].ifindex) {
1831                                 char ifname[IF_NAMESIZE];
1832
1833                                 /*
1834                                  * Netlink failed, it may happen with old
1835                                  * ib_core kernel driver (before 4.16).
1836                                  * We can assume there is old driver because
1837                                  * here we are processing single ports IB
1838                                  * devices. Let's try sysfs to retrieve
1839                                  * the ifindex. The method works for
1840                                  * master device only.
1841                                  */
1842                                 if (nd > 1) {
1843                                         /*
1844                                          * Multiple devices found, assume
1845                                          * representors, can not distinguish
1846                                          * master/representor and retrieve
1847                                          * ifindex via sysfs.
1848                                          */
1849                                         continue;
1850                                 }
1851                                 ret = mlx5_get_master_ifname
1852                                         (ibv_match[i]->ibdev_path, &ifname);
1853                                 if (!ret)
1854                                         list[ns].ifindex =
1855                                                 if_nametoindex(ifname);
1856                                 if (!list[ns].ifindex) {
1857                                         /*
1858                                          * No network interface index found
1859                                          * for the specified device, it means
1860                                          * there it is neither representor
1861                                          * nor master.
1862                                          */
1863                                         continue;
1864                                 }
1865                         }
1866                         ret = -1;
1867                         if (nl_route >= 0)
1868                                 ret = mlx5_nl_switch_info
1869                                                (nl_route,
1870                                                 list[ns].ifindex,
1871                                                 &list[ns].info);
1872                         if (ret || (!list[ns].info.representor &&
1873                                     !list[ns].info.master)) {
1874                                 /*
1875                                  * We failed to recognize representors with
1876                                  * Netlink, let's try to perform the task
1877                                  * with sysfs.
1878                                  */
1879                                 ret =  mlx5_sysfs_switch_info
1880                                                 (list[ns].ifindex,
1881                                                  &list[ns].info);
1882                         }
1883                         if (!ret && (list[ns].info.representor ^
1884                                      list[ns].info.master)) {
1885                                 ns++;
1886                         } else if ((nd == 1) &&
1887                                    !list[ns].info.representor &&
1888                                    !list[ns].info.master) {
1889                                 /*
1890                                  * Single IB device with
1891                                  * one physical port and
1892                                  * attached network device.
1893                                  * May be SRIOV is not enabled
1894                                  * or there is no representors.
1895                                  */
1896                                 DRV_LOG(INFO, "no E-Switch support detected");
1897                                 ns++;
1898                                 break;
1899                         }
1900                 }
1901                 if (!ns) {
1902                         DRV_LOG(ERR,
1903                                 "unable to recognize master/representors"
1904                                 " on the multiple IB devices");
1905                         rte_errno = ENOENT;
1906                         ret = -rte_errno;
1907                         goto exit;
1908                 }
1909         }
1910         assert(ns);
1911         /*
1912          * Sort list to probe devices in natural order for users convenience
1913          * (i.e. master first, then representors from lowest to highest ID).
1914          */
1915         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1916         /* Default configuration. */
1917         dev_config = (struct mlx5_dev_config){
1918                 .hw_padding = 0,
1919                 .mps = MLX5_ARG_UNSET,
1920                 .tx_vec_en = 1,
1921                 .rx_vec_en = 1,
1922                 .txq_inline = MLX5_ARG_UNSET,
1923                 .txqs_inline = MLX5_ARG_UNSET,
1924                 .txqs_vec = MLX5_ARG_UNSET,
1925                 .inline_max_packet_sz = MLX5_ARG_UNSET,
1926                 .vf_nl_en = 1,
1927                 .mr_ext_memseg_en = 1,
1928                 .mprq = {
1929                         .enabled = 0, /* Disabled by default. */
1930                         .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
1931                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1932                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1933                 },
1934                 .dv_esw_en = 1,
1935         };
1936         /* Device specific configuration. */
1937         switch (pci_dev->id.device_id) {
1938         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BF:
1939                 dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS_BLUEFIELD;
1940                 break;
1941         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1942         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1943         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1944         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1945                 dev_config.vf = 1;
1946                 break;
1947         default:
1948                 break;
1949         }
1950         /* Set architecture-dependent default value if unset. */
1951         if (dev_config.txqs_vec == MLX5_ARG_UNSET)
1952                 dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS;
1953         for (i = 0; i != ns; ++i) {
1954                 uint32_t restore;
1955
1956                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1957                                                  &list[i],
1958                                                  dev_config);
1959                 if (!list[i].eth_dev) {
1960                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1961                                 break;
1962                         /* Device is disabled or already spawned. Ignore it. */
1963                         continue;
1964                 }
1965                 restore = list[i].eth_dev->data->dev_flags;
1966                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1967                 /* Restore non-PCI flags cleared by the above call. */
1968                 list[i].eth_dev->data->dev_flags |= restore;
1969                 rte_eth_dev_probing_finish(list[i].eth_dev);
1970         }
1971         if (i != ns) {
1972                 DRV_LOG(ERR,
1973                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1974                         " encountering an error: %s",
1975                         pci_dev->addr.domain, pci_dev->addr.bus,
1976                         pci_dev->addr.devid, pci_dev->addr.function,
1977                         strerror(rte_errno));
1978                 ret = -rte_errno;
1979                 /* Roll back. */
1980                 while (i--) {
1981                         if (!list[i].eth_dev)
1982                                 continue;
1983                         mlx5_dev_close(list[i].eth_dev);
1984                         /* mac_addrs must not be freed because in dev_private */
1985                         list[i].eth_dev->data->mac_addrs = NULL;
1986                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1987                 }
1988                 /* Restore original error. */
1989                 rte_errno = -ret;
1990         } else {
1991                 ret = 0;
1992         }
1993 exit:
1994         /*
1995          * Do the routine cleanup:
1996          * - close opened Netlink sockets
1997          * - free the Infiniband device list
1998          */
1999         if (nl_rdma >= 0)
2000                 close(nl_rdma);
2001         if (nl_route >= 0)
2002                 close(nl_route);
2003         assert(ibv_list);
2004         mlx5_glue->free_device_list(ibv_list);
2005         return ret;
2006 }
2007
2008 /**
2009  * DPDK callback to remove a PCI device.
2010  *
2011  * This function removes all Ethernet devices belong to a given PCI device.
2012  *
2013  * @param[in] pci_dev
2014  *   Pointer to the PCI device.
2015  *
2016  * @return
2017  *   0 on success, the function cannot fail.
2018  */
2019 static int
2020 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2021 {
2022         uint16_t port_id;
2023
2024         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
2025                 rte_eth_dev_close(port_id);
2026         return 0;
2027 }
2028
2029 static const struct rte_pci_id mlx5_pci_id_map[] = {
2030         {
2031                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2032                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2033         },
2034         {
2035                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2036                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2037         },
2038         {
2039                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2040                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2041         },
2042         {
2043                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2044                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2045         },
2046         {
2047                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2048                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2049         },
2050         {
2051                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2052                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2053         },
2054         {
2055                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2056                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2057         },
2058         {
2059                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2060                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2061         },
2062         {
2063                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2064                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2065         },
2066         {
2067                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2068                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2069         },
2070         {
2071                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2072                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2073         },
2074         {
2075                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2076                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2077         },
2078         {
2079                 .vendor_id = 0
2080         }
2081 };
2082
2083 static struct rte_pci_driver mlx5_driver = {
2084         .driver = {
2085                 .name = MLX5_DRIVER_NAME
2086         },
2087         .id_table = mlx5_pci_id_map,
2088         .probe = mlx5_pci_probe,
2089         .remove = mlx5_pci_remove,
2090         .dma_map = mlx5_dma_map,
2091         .dma_unmap = mlx5_dma_unmap,
2092         .drv_flags = (RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
2093                       RTE_PCI_DRV_PROBE_AGAIN),
2094 };
2095
2096 #ifdef RTE_IBVERBS_LINK_DLOPEN
2097
2098 /**
2099  * Suffix RTE_EAL_PMD_PATH with "-glue".
2100  *
2101  * This function performs a sanity check on RTE_EAL_PMD_PATH before
2102  * suffixing its last component.
2103  *
2104  * @param buf[out]
2105  *   Output buffer, should be large enough otherwise NULL is returned.
2106  * @param size
2107  *   Size of @p out.
2108  *
2109  * @return
2110  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
2111  */
2112 static char *
2113 mlx5_glue_path(char *buf, size_t size)
2114 {
2115         static const char *const bad[] = { "/", ".", "..", NULL };
2116         const char *path = RTE_EAL_PMD_PATH;
2117         size_t len = strlen(path);
2118         size_t off;
2119         int i;
2120
2121         while (len && path[len - 1] == '/')
2122                 --len;
2123         for (off = len; off && path[off - 1] != '/'; --off)
2124                 ;
2125         for (i = 0; bad[i]; ++i)
2126                 if (!strncmp(path + off, bad[i], (int)(len - off)))
2127                         goto error;
2128         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
2129         if (i == -1 || (size_t)i >= size)
2130                 goto error;
2131         return buf;
2132 error:
2133         DRV_LOG(ERR,
2134                 "unable to append \"-glue\" to last component of"
2135                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
2136                 " please re-configure DPDK");
2137         return NULL;
2138 }
2139
2140 /**
2141  * Initialization routine for run-time dependency on rdma-core.
2142  */
2143 static int
2144 mlx5_glue_init(void)
2145 {
2146         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
2147         const char *path[] = {
2148                 /*
2149                  * A basic security check is necessary before trusting
2150                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
2151                  */
2152                 (geteuid() == getuid() && getegid() == getgid() ?
2153                  getenv("MLX5_GLUE_PATH") : NULL),
2154                 /*
2155                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
2156                  * variant, otherwise let dlopen() look up libraries on its
2157                  * own.
2158                  */
2159                 (*RTE_EAL_PMD_PATH ?
2160                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
2161         };
2162         unsigned int i = 0;
2163         void *handle = NULL;
2164         void **sym;
2165         const char *dlmsg;
2166
2167         while (!handle && i != RTE_DIM(path)) {
2168                 const char *end;
2169                 size_t len;
2170                 int ret;
2171
2172                 if (!path[i]) {
2173                         ++i;
2174                         continue;
2175                 }
2176                 end = strpbrk(path[i], ":;");
2177                 if (!end)
2178                         end = path[i] + strlen(path[i]);
2179                 len = end - path[i];
2180                 ret = 0;
2181                 do {
2182                         char name[ret + 1];
2183
2184                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
2185                                        (int)len, path[i],
2186                                        (!len || *(end - 1) == '/') ? "" : "/");
2187                         if (ret == -1)
2188                                 break;
2189                         if (sizeof(name) != (size_t)ret + 1)
2190                                 continue;
2191                         DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
2192                                 name);
2193                         handle = dlopen(name, RTLD_LAZY);
2194                         break;
2195                 } while (1);
2196                 path[i] = end + 1;
2197                 if (!*end)
2198                         ++i;
2199         }
2200         if (!handle) {
2201                 rte_errno = EINVAL;
2202                 dlmsg = dlerror();
2203                 if (dlmsg)
2204                         DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
2205                 goto glue_error;
2206         }
2207         sym = dlsym(handle, "mlx5_glue");
2208         if (!sym || !*sym) {
2209                 rte_errno = EINVAL;
2210                 dlmsg = dlerror();
2211                 if (dlmsg)
2212                         DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
2213                 goto glue_error;
2214         }
2215         mlx5_glue = *sym;
2216         return 0;
2217 glue_error:
2218         if (handle)
2219                 dlclose(handle);
2220         DRV_LOG(WARNING,
2221                 "cannot initialize PMD due to missing run-time dependency on"
2222                 " rdma-core libraries (libibverbs, libmlx5)");
2223         return -rte_errno;
2224 }
2225
2226 #endif
2227
2228 /**
2229  * Driver initialization routine.
2230  */
2231 RTE_INIT(rte_mlx5_pmd_init)
2232 {
2233         /* Initialize driver log type. */
2234         mlx5_logtype = rte_log_register("pmd.net.mlx5");
2235         if (mlx5_logtype >= 0)
2236                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
2237
2238         /* Build the static tables for Verbs conversion. */
2239         mlx5_set_ptype_table();
2240         mlx5_set_cksum_table();
2241         mlx5_set_swp_types_table();
2242         /*
2243          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
2244          * huge pages. Calling ibv_fork_init() during init allows
2245          * applications to use fork() safely for purposes other than
2246          * using this PMD, which is not supported in forked processes.
2247          */
2248         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
2249         /* Match the size of Rx completion entry to the size of a cacheline. */
2250         if (RTE_CACHE_LINE_SIZE == 128)
2251                 setenv("MLX5_CQE_SIZE", "128", 0);
2252         /*
2253          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
2254          * cleanup all the Verbs resources even when the device was removed.
2255          */
2256         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
2257 #ifdef RTE_IBVERBS_LINK_DLOPEN
2258         if (mlx5_glue_init())
2259                 return;
2260         assert(mlx5_glue);
2261 #endif
2262 #ifndef NDEBUG
2263         /* Glue structure must not contain any NULL pointers. */
2264         {
2265                 unsigned int i;
2266
2267                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
2268                         assert(((const void *const *)mlx5_glue)[i]);
2269         }
2270 #endif
2271         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
2272                 DRV_LOG(ERR,
2273                         "rdma-core glue \"%s\" mismatch: \"%s\" is required",
2274                         mlx5_glue->version, MLX5_GLUE_VERSION);
2275                 return;
2276         }
2277         mlx5_glue->fork_init();
2278         rte_pci_register(&mlx5_driver);
2279 }
2280
2281 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2282 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2283 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");