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 = mlx5dv_dr_create_ns(sh->ctx, MLX5DV_DR_NS_DOMAIN_INGRESS_BYPASS);
 336         if (!ns) {
 337                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_ns failed");
 338                 err = errno;
 339                 goto error;
 340         }
 341         sh->rx_ns = ns;
 342         ns = mlx5dv_dr_create_ns(sh->ctx, MLX5DV_DR_NS_DOMAIN_EGRESS_BYPASS);
 343         if (!ns) {
 344                 DRV_LOG(ERR, "egress mlx5dv_dr_create_ns failed");
 345                 err = errno;
 346                 goto error;
 347         }
 348         pthread_mutex_init(&sh->dv_mutex, NULL);
 349         sh->tx_ns = ns;
 350 #ifdef HAVE_MLX5DV_DR_ESWITCH
 351         if (priv->config.dv_esw_en) {
 352                 ns  = mlx5_glue->dr_create_ns(sh->ctx,
 353                                               MLX5DV_DR_NS_DOMAIN_FDB_BYPASS);
 354                 if (!ns) {
 355                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_ns failed");
 356                         err = errno;
 357                         goto error;
 358                 }
 359                 sh->fdb_ns = ns;
 360                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 361         }
 362 #endif
 363         sh->dv_refcnt++;
 364         priv->dr_shared = 1;
 365         return 0;
 366
 367 error:
 368        /* Rollback the created objects. */
 369         if (sh->rx_ns) {
 370                 mlx5dv_dr_destroy_ns(sh->rx_ns);
 371                 sh->rx_ns = NULL;
 372         }
 373         if (sh->tx_ns) {
 374                 mlx5dv_dr_destroy_ns(sh->tx_ns);
 375                 sh->tx_ns = NULL;
 376         }
 377         if (sh->fdb_ns) {
 378                 mlx5_glue->dr_destroy_ns(sh->fdb_ns);
 379                 sh->fdb_ns = NULL;
 380         }
 381         if (sh->esw_drop_action) {
 382                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 383                 sh->esw_drop_action = NULL;
 384         }
 385         return err;
 386 #else
 387         (void)priv;
 388         return 0;
 389 #endif
 390 }
 391
 392 /**
 393  * Destroy DR related data within private structure.
 394  *
 395  * @param[in] priv
 396  *   Pointer to the private device data structure.
 397  */
 398 static void
 399 mlx5_free_shared_dr(struct mlx5_priv *priv)
 400 {
 401 #ifdef HAVE_MLX5DV_DR
 402         struct mlx5_ibv_shared *sh;
 403
 404         if (!priv->dr_shared)
 405                 return;
 406         priv->dr_shared = 0;
 407         sh = priv->sh;
 408         assert(sh);
 409         assert(sh->dv_refcnt);
 410         if (sh->dv_refcnt && --sh->dv_refcnt)
 411                 return;
 412         if (sh->rx_ns) {
 413                 mlx5dv_dr_destroy_ns(sh->rx_ns);
 414                 sh->rx_ns = NULL;
 415         }
 416         if (sh->tx_ns) {
 417                 mlx5dv_dr_destroy_ns(sh->tx_ns);
 418                 sh->tx_ns = NULL;
 419         }
 420 #ifdef HAVE_MLX5DV_DR_ESWITCH
 421         if (sh->fdb_ns) {
 422                 mlx5_glue->dr_destroy_ns(sh->fdb_ns);
 423                 sh->fdb_ns = NULL;
 424         }
 425         if (sh->esw_drop_action) {
 426                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 427                 sh->esw_drop_action = NULL;
 428         }
 429 #endif
 430         pthread_mutex_destroy(&sh->dv_mutex);
 431 #else
 432         (void)priv;
 433 #endif
 434 }
 435
 436 /**
 437  * Initialize shared data between primary and secondary process.
 438  *
 439  * A memzone is reserved by primary process and secondary processes attach to
 440  * the memzone.
 441  *
 442  * @return
 443  *   0 on success, a negative errno value otherwise and rte_errno is set.
 444  */
 445 static int
 446 mlx5_init_shared_data(void)
 447 {
 448         const struct rte_memzone *mz;
 449         int ret = 0;
 450
 451         rte_spinlock_lock(&mlx5_shared_data_lock);
 452         if (mlx5_shared_data == NULL) {
 453                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 454                         /* Allocate shared memory. */
 455                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
 456                                                  sizeof(*mlx5_shared_data),
 457                                                  SOCKET_ID_ANY, 0);
 458                         if (mz == NULL) {
 459                                 DRV_LOG(ERR,
 460                                         "Cannot allocate mlx5 shared data\n");
 461                                 ret = -rte_errno;
 462                                 goto error;
 463                         }
 464                         mlx5_shared_data = mz->addr;
 465                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
 466                         rte_spinlock_init(&mlx5_shared_data->lock);
 467                 } else {
 468                         /* Lookup allocated shared memory. */
 469                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
 470                         if (mz == NULL) {
 471                                 DRV_LOG(ERR,
 472                                         "Cannot attach mlx5 shared data\n");
 473                                 ret = -rte_errno;
 474                                 goto error;
 475                         }
 476                         mlx5_shared_data = mz->addr;
 477                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
 478                 }
 479         }
 480 error:
 481         rte_spinlock_unlock(&mlx5_shared_data_lock);
 482         return ret;
 483 }
 484
 485 /**
 486  * Retrieve integer value from environment variable.
 487  *
 488  * @param[in] name
 489  *   Environment variable name.
 490  *
 491  * @return
 492  *   Integer value, 0 if the variable is not set.
 493  */
 494 int
 495 mlx5_getenv_int(const char *name)
 496 {
 497         const char *val = getenv(name);
 498
 499         if (val == NULL)
 500                 return 0;
 501         return atoi(val);
 502 }
 503
 504 /**
 505  * Verbs callback to allocate a memory. This function should allocate the space
 506  * according to the size provided residing inside a huge page.
 507  * Please note that all allocation must respect the alignment from libmlx5
 508  * (i.e. currently sysconf(_SC_PAGESIZE)).
 509  *
 510  * @param[in] size
 511  *   The size in bytes of the memory to allocate.
 512  * @param[in] data
 513  *   A pointer to the callback data.
 514  *
 515  * @return
 516  *   Allocated buffer, NULL otherwise and rte_errno is set.
 517  */
 518 static void *
 519 mlx5_alloc_verbs_buf(size_t size, void *data)
 520 {
 521         struct mlx5_priv *priv = data;
 522         void *ret;
 523         size_t alignment = sysconf(_SC_PAGESIZE);
 524         unsigned int socket = SOCKET_ID_ANY;
 525
 526         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 527                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 528
 529                 socket = ctrl->socket;
 530         } else if (priv->verbs_alloc_ctx.type ==
 531                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 532                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 533
 534                 socket = ctrl->socket;
 535         }
 536         assert(data != NULL);
 537         ret = rte_malloc_socket(__func__, size, alignment, socket);
 538         if (!ret && size)
 539                 rte_errno = ENOMEM;
 540         return ret;
 541 }
 542
 543 /**
 544  * Verbs callback to free a memory.
 545  *
 546  * @param[in] ptr
 547  *   A pointer to the memory to free.
 548  * @param[in] data
 549  *   A pointer to the callback data.
 550  */
 551 static void
 552 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 553 {
 554         assert(data != NULL);
 555         rte_free(ptr);
 556 }
 557
 558 /**
 559  * Initialize process private data structure.
 560  *
 561  * @param dev
 562  *   Pointer to Ethernet device structure.
 563  *
 564  * @return
 565  *   0 on success, a negative errno value otherwise and rte_errno is set.
 566  */
 567 int
 568 mlx5_proc_priv_init(struct rte_eth_dev *dev)
 569 {
 570         struct mlx5_priv *priv = dev->data->dev_private;
 571         struct mlx5_proc_priv *ppriv;
 572         size_t ppriv_size;
 573
 574         /*
 575          * UAR register table follows the process private structure. BlueFlame
 576          * registers for Tx queues are stored in the table.
 577          */
 578         ppriv_size =
 579                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
 580         ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
 581                                   RTE_CACHE_LINE_SIZE, dev->device->numa_node);
 582         if (!ppriv) {
 583                 rte_errno = ENOMEM;
 584                 return -rte_errno;
 585         }
 586         ppriv->uar_table_sz = ppriv_size;
 587         dev->process_private = ppriv;
 588         return 0;
 589 }
 590
 591 /**
 592  * Un-initialize process private data structure.
 593  *
 594  * @param dev
 595  *   Pointer to Ethernet device structure.
 596  */
 597 static void
 598 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
 599 {
 600         if (!dev->process_private)
 601                 return;
 602         rte_free(dev->process_private);
 603         dev->process_private = NULL;
 604 }
 605
 606 /**
 607  * DPDK callback to close the device.
 608  *
 609  * Destroy all queues and objects, free memory.
 610  *
 611  * @param dev
 612  *   Pointer to Ethernet device structure.
 613  */
 614 static void
 615 mlx5_dev_close(struct rte_eth_dev *dev)
 616 {
 617         struct mlx5_priv *priv = dev->data->dev_private;
 618         unsigned int i;
 619         int ret;
 620
 621         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
 622                 dev->data->port_id,
 623                 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
 624         /* In case mlx5_dev_stop() has not been called. */
 625         mlx5_dev_interrupt_handler_uninstall(dev);
 626         mlx5_traffic_disable(dev);
 627         mlx5_flow_flush(dev, NULL);
 628         /* Prevent crashes when queues are still in use. */
 629         dev->rx_pkt_burst = removed_rx_burst;
 630         dev->tx_pkt_burst = removed_tx_burst;
 631         rte_wmb();
 632         /* Disable datapath on secondary process. */
 633         mlx5_mp_req_stop_rxtx(dev);
 634         if (priv->rxqs != NULL) {
 635                 /* XXX race condition if mlx5_rx_burst() is still running. */
 636                 usleep(1000);
 637                 for (i = 0; (i != priv->rxqs_n); ++i)
 638                         mlx5_rxq_release(dev, i);
 639                 priv->rxqs_n = 0;
 640                 priv->rxqs = NULL;
 641         }
 642         if (priv->txqs != NULL) {
 643                 /* XXX race condition if mlx5_tx_burst() is still running. */
 644                 usleep(1000);
 645                 for (i = 0; (i != priv->txqs_n); ++i)
 646                         mlx5_txq_release(dev, i);
 647                 priv->txqs_n = 0;
 648                 priv->txqs = NULL;
 649         }
 650         mlx5_proc_priv_uninit(dev);
 651         mlx5_mprq_free_mp(dev);
 652         mlx5_mr_release(dev);
 653         assert(priv->sh);
 654         mlx5_free_shared_dr(priv);
 655         if (priv->rss_conf.rss_key != NULL)
 656                 rte_free(priv->rss_conf.rss_key);
 657         if (priv->reta_idx != NULL)
 658                 rte_free(priv->reta_idx);
 659         if (priv->config.vf)
 660                 mlx5_nl_mac_addr_flush(dev);
 661         if (priv->nl_socket_route >= 0)
 662                 close(priv->nl_socket_route);
 663         if (priv->nl_socket_rdma >= 0)
 664                 close(priv->nl_socket_rdma);
 665         if (priv->tcf_context)
 666                 mlx5_flow_tcf_context_destroy(priv->tcf_context);
 667         if (priv->sh) {
 668                 /*
 669                  * Free the shared context in last turn, because the cleanup
 670                  * routines above may use some shared fields, like
 671                  * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
 672                  * ifindex if Netlink fails.
 673                  */
 674                 mlx5_free_shared_ibctx(priv->sh);
 675                 priv->sh = NULL;
 676         }
 677         ret = mlx5_hrxq_ibv_verify(dev);
 678         if (ret)
 679                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
 680                         dev->data->port_id);
 681         ret = mlx5_ind_table_ibv_verify(dev);
 682         if (ret)
 683                 DRV_LOG(WARNING, "port %u some indirection table still remain",
 684                         dev->data->port_id);
 685         ret = mlx5_rxq_ibv_verify(dev);
 686         if (ret)
 687                 DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
 688                         dev->data->port_id);
 689         ret = mlx5_rxq_verify(dev);
 690         if (ret)
 691                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
 692                         dev->data->port_id);
 693         ret = mlx5_txq_ibv_verify(dev);
 694         if (ret)
 695                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
 696                         dev->data->port_id);
 697         ret = mlx5_txq_verify(dev);
 698         if (ret)
 699                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
 700                         dev->data->port_id);
 701         ret = mlx5_flow_verify(dev);
 702         if (ret)
 703                 DRV_LOG(WARNING, "port %u some flows still remain",
 704                         dev->data->port_id);
 705         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 706                 unsigned int c = 0;
 707                 uint16_t port_id;
 708
 709                 RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
 710                         struct mlx5_priv *opriv =
 711                                 rte_eth_devices[port_id].data->dev_private;
 712
 713                         if (!opriv ||
 714                             opriv->domain_id != priv->domain_id ||
 715                             &rte_eth_devices[port_id] == dev)
 716                                 continue;
 717                         ++c;
 718                 }
 719                 if (!c)
 720                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
 721         }
 722         memset(priv, 0, sizeof(*priv));
 723         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 724         /*
 725          * Reset mac_addrs to NULL such that it is not freed as part of
 726          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
 727          * it is freed when dev_private is freed.
 728          */
 729         dev->data->mac_addrs = NULL;
 730 }
 731
 732 const struct eth_dev_ops mlx5_dev_ops = {
 733         .dev_configure = mlx5_dev_configure,
 734         .dev_start = mlx5_dev_start,
 735         .dev_stop = mlx5_dev_stop,
 736         .dev_set_link_down = mlx5_set_link_down,
 737         .dev_set_link_up = mlx5_set_link_up,
 738         .dev_close = mlx5_dev_close,
 739         .promiscuous_enable = mlx5_promiscuous_enable,
 740         .promiscuous_disable = mlx5_promiscuous_disable,
 741         .allmulticast_enable = mlx5_allmulticast_enable,
 742         .allmulticast_disable = mlx5_allmulticast_disable,
 743         .link_update = mlx5_link_update,
 744         .stats_get = mlx5_stats_get,
 745         .stats_reset = mlx5_stats_reset,
 746         .xstats_get = mlx5_xstats_get,
 747         .xstats_reset = mlx5_xstats_reset,
 748         .xstats_get_names = mlx5_xstats_get_names,
 749         .fw_version_get = mlx5_fw_version_get,
 750         .dev_infos_get = mlx5_dev_infos_get,
 751         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 752         .vlan_filter_set = mlx5_vlan_filter_set,
 753         .rx_queue_setup = mlx5_rx_queue_setup,
 754         .tx_queue_setup = mlx5_tx_queue_setup,
 755         .rx_queue_release = mlx5_rx_queue_release,
 756         .tx_queue_release = mlx5_tx_queue_release,
 757         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 758         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 759         .mac_addr_remove = mlx5_mac_addr_remove,
 760         .mac_addr_add = mlx5_mac_addr_add,
 761         .mac_addr_set = mlx5_mac_addr_set,
 762         .set_mc_addr_list = mlx5_set_mc_addr_list,
 763         .mtu_set = mlx5_dev_set_mtu,
 764         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 765         .vlan_offload_set = mlx5_vlan_offload_set,
 766         .reta_update = mlx5_dev_rss_reta_update,
 767         .reta_query = mlx5_dev_rss_reta_query,
 768         .rss_hash_update = mlx5_rss_hash_update,
 769         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
 770         .filter_ctrl = mlx5_dev_filter_ctrl,
 771         .rx_descriptor_status = mlx5_rx_descriptor_status,
 772         .tx_descriptor_status = mlx5_tx_descriptor_status,
 773         .rx_queue_count = mlx5_rx_queue_count,
 774         .rx_queue_intr_enable = mlx5_rx_intr_enable,
 775         .rx_queue_intr_disable = mlx5_rx_intr_disable,
 776         .is_removed = mlx5_is_removed,
 777 };
 778
 779 /* Available operations from secondary process. */
 780 static const struct eth_dev_ops mlx5_dev_sec_ops = {
 781         .stats_get = mlx5_stats_get,
 782         .stats_reset = mlx5_stats_reset,
 783         .xstats_get = mlx5_xstats_get,
 784         .xstats_reset = mlx5_xstats_reset,
 785         .xstats_get_names = mlx5_xstats_get_names,
 786         .fw_version_get = mlx5_fw_version_get,
 787         .dev_infos_get = mlx5_dev_infos_get,
 788         .rx_descriptor_status = mlx5_rx_descriptor_status,
 789         .tx_descriptor_status = mlx5_tx_descriptor_status,
 790 };
 791
 792 /* Available operations in flow isolated mode. */
 793 const struct eth_dev_ops mlx5_dev_ops_isolate = {
 794         .dev_configure = mlx5_dev_configure,
 795         .dev_start = mlx5_dev_start,
 796         .dev_stop = mlx5_dev_stop,
 797         .dev_set_link_down = mlx5_set_link_down,
 798         .dev_set_link_up = mlx5_set_link_up,
 799         .dev_close = mlx5_dev_close,
 800         .promiscuous_enable = mlx5_promiscuous_enable,
 801         .promiscuous_disable = mlx5_promiscuous_disable,
 802         .allmulticast_enable = mlx5_allmulticast_enable,
 803         .allmulticast_disable = mlx5_allmulticast_disable,
 804         .link_update = mlx5_link_update,
 805         .stats_get = mlx5_stats_get,
 806         .stats_reset = mlx5_stats_reset,
 807         .xstats_get = mlx5_xstats_get,
 808         .xstats_reset = mlx5_xstats_reset,
 809         .xstats_get_names = mlx5_xstats_get_names,
 810         .fw_version_get = mlx5_fw_version_get,
 811         .dev_infos_get = mlx5_dev_infos_get,
 812         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 813         .vlan_filter_set = mlx5_vlan_filter_set,
 814         .rx_queue_setup = mlx5_rx_queue_setup,
 815         .tx_queue_setup = mlx5_tx_queue_setup,
 816         .rx_queue_release = mlx5_rx_queue_release,
 817         .tx_queue_release = mlx5_tx_queue_release,
 818         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 819         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 820         .mac_addr_remove = mlx5_mac_addr_remove,
 821         .mac_addr_add = mlx5_mac_addr_add,
 822         .mac_addr_set = mlx5_mac_addr_set,
 823         .set_mc_addr_list = mlx5_set_mc_addr_list,
 824         .mtu_set = mlx5_dev_set_mtu,
 825         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 826         .vlan_offload_set = mlx5_vlan_offload_set,
 827         .filter_ctrl = mlx5_dev_filter_ctrl,
 828         .rx_descriptor_status = mlx5_rx_descriptor_status,
 829         .tx_descriptor_status = mlx5_tx_descriptor_status,
 830         .rx_queue_intr_enable = mlx5_rx_intr_enable,
 831         .rx_queue_intr_disable = mlx5_rx_intr_disable,
 832         .is_removed = mlx5_is_removed,
 833 };
 834
 835 /**
 836  * Verify and store value for device argument.
 837  *
 838  * @param[in] key
 839  *   Key argument to verify.
 840  * @param[in] val
 841  *   Value associated with key.
 842  * @param opaque
 843  *   User data.
 844  *
 845  * @return
 846  *   0 on success, a negative errno value otherwise and rte_errno is set.
 847  */
 848 static int
 849 mlx5_args_check(const char *key, const char *val, void *opaque)
 850 {
 851         struct mlx5_dev_config *config = opaque;
 852         unsigned long tmp;
 853
 854         /* No-op, port representors are processed in mlx5_dev_spawn(). */
 855         if (!strcmp(MLX5_REPRESENTOR, key))
 856                 return 0;
 857         errno = 0;
 858         tmp = strtoul(val, NULL, 0);
 859         if (errno) {
 860                 rte_errno = errno;
 861                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
 862                 return -rte_errno;
 863         }
 864         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
 865                 config->cqe_comp = !!tmp;
 866         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
 867                 config->cqe_pad = !!tmp;
 868         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
 869                 config->hw_padding = !!tmp;
 870         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
 871                 config->mprq.enabled = !!tmp;
 872         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
 873                 config->mprq.stride_num_n = tmp;
 874         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
 875                 config->mprq.max_memcpy_len = tmp;
 876         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
 877                 config->mprq.min_rxqs_num = tmp;
 878         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
 879                 config->txq_inline = tmp;
 880         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
 881                 config->txqs_inline = tmp;
 882         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
 883                 config->txqs_vec = tmp;
 884         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
 885                 config->mps = !!tmp;
 886         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
 887                 config->mpw_hdr_dseg = !!tmp;
 888         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
 889                 config->inline_max_packet_sz = tmp;
 890         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
 891                 config->tx_vec_en = !!tmp;
 892         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
 893                 config->rx_vec_en = !!tmp;
 894         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
 895                 config->l3_vxlan_en = !!tmp;
 896         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
 897                 config->vf_nl_en = !!tmp;
 898         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
 899                 config->dv_esw_en = !!tmp;
 900         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
 901                 config->dv_flow_en = !!tmp;
 902         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
 903                 config->mr_ext_memseg_en = !!tmp;
 904         } else {
 905                 DRV_LOG(WARNING, "%s: unknown parameter", key);
 906                 rte_errno = EINVAL;
 907                 return -rte_errno;
 908         }
 909         return 0;
 910 }
 911
 912 /**
 913  * Parse device parameters.
 914  *
 915  * @param config
 916  *   Pointer to device configuration structure.
 917  * @param devargs
 918  *   Device arguments structure.
 919  *
 920  * @return
 921  *   0 on success, a negative errno value otherwise and rte_errno is set.
 922  */
 923 static int
 924 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
 925 {
 926         const char **params = (const char *[]){
 927                 MLX5_RXQ_CQE_COMP_EN,
 928                 MLX5_RXQ_CQE_PAD_EN,
 929                 MLX5_RXQ_PKT_PAD_EN,
 930                 MLX5_RX_MPRQ_EN,
 931                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
 932                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
 933                 MLX5_RXQS_MIN_MPRQ,
 934                 MLX5_TXQ_INLINE,
 935                 MLX5_TXQS_MIN_INLINE,
 936                 MLX5_TXQS_MAX_VEC,
 937                 MLX5_TXQ_MPW_EN,
 938                 MLX5_TXQ_MPW_HDR_DSEG_EN,
 939                 MLX5_TXQ_MAX_INLINE_LEN,
 940                 MLX5_TX_VEC_EN,
 941                 MLX5_RX_VEC_EN,
 942                 MLX5_L3_VXLAN_EN,
 943                 MLX5_VF_NL_EN,
 944                 MLX5_DV_ESW_EN,
 945                 MLX5_DV_FLOW_EN,
 946                 MLX5_MR_EXT_MEMSEG_EN,
 947                 MLX5_REPRESENTOR,
 948                 NULL,
 949         };
 950         struct rte_kvargs *kvlist;
 951         int ret = 0;
 952         int i;
 953
 954         if (devargs == NULL)
 955                 return 0;
 956         /* Following UGLY cast is done to pass checkpatch. */
 957         kvlist = rte_kvargs_parse(devargs->args, params);
 958         if (kvlist == NULL)
 959                 return 0;
 960         /* Process parameters. */
 961         for (i = 0; (params[i] != NULL); ++i) {
 962                 if (rte_kvargs_count(kvlist, params[i])) {
 963                         ret = rte_kvargs_process(kvlist, params[i],
 964                                                  mlx5_args_check, config);
 965                         if (ret) {
 966                                 rte_errno = EINVAL;
 967                                 rte_kvargs_free(kvlist);
 968                                 return -rte_errno;
 969                         }
 970                 }
 971         }
 972         rte_kvargs_free(kvlist);
 973         return 0;
 974 }
 975
 976 static struct rte_pci_driver mlx5_driver;
 977
 978 /**
 979  * PMD global initialization.
 980  *
 981  * Independent from individual device, this function initializes global
 982  * per-PMD data structures distinguishing primary and secondary processes.
 983  * Hence, each initialization is called once per a process.
 984  *
 985  * @return
 986  *   0 on success, a negative errno value otherwise and rte_errno is set.
 987  */
 988 static int
 989 mlx5_init_once(void)
 990 {
 991         struct mlx5_shared_data *sd;
 992         struct mlx5_local_data *ld = &mlx5_local_data;
 993
 994         if (mlx5_init_shared_data())
 995                 return -rte_errno;
 996         sd = mlx5_shared_data;
 997         assert(sd);
 998         rte_spinlock_lock(&sd->lock);
 999         switch (rte_eal_process_type()) {
1000         case RTE_PROC_PRIMARY:
1001                 if (sd->init_done)
1002                         break;
1003                 LIST_INIT(&sd->mem_event_cb_list);
1004                 rte_rwlock_init(&sd->mem_event_rwlock);
1005                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1006                                                 mlx5_mr_mem_event_cb, NULL);
1007                 mlx5_mp_init_primary();
1008                 sd->init_done = true;
1009                 break;
1010         case RTE_PROC_SECONDARY:
1011                 if (ld->init_done)
1012                         break;
1013                 mlx5_mp_init_secondary();
1014                 ++sd->secondary_cnt;
1015                 ld->init_done = true;
1016                 break;
1017         default:
1018                 break;
1019         }
1020         rte_spinlock_unlock(&sd->lock);
1021         return 0;
1022 }
1023
1024 /**
1025  * Spawn an Ethernet device from Verbs information.
1026  *
1027  * @param dpdk_dev
1028  *   Backing DPDK device.
1029  * @param spawn
1030  *   Verbs device parameters (name, port, switch_info) to spawn.
1031  * @param config
1032  *   Device configuration parameters.
1033  *
1034  * @return
1035  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1036  *   is set. The following errors are defined:
1037  *
1038  *   EBUSY: device is not supposed to be spawned.
1039  *   EEXIST: device is already spawned
1040  */
1041 static struct rte_eth_dev *
1042 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1043                struct mlx5_dev_spawn_data *spawn,
1044                struct mlx5_dev_config config)
1045 {
1046         const struct mlx5_switch_info *switch_info = &spawn->info;
1047         struct mlx5_ibv_shared *sh = NULL;
1048         struct ibv_port_attr port_attr;
1049         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1050         struct rte_eth_dev *eth_dev = NULL;
1051         struct mlx5_priv *priv = NULL;
1052         int err = 0;
1053         unsigned int hw_padding = 0;
1054         unsigned int mps;
1055         unsigned int cqe_comp;
1056         unsigned int cqe_pad = 0;
1057         unsigned int tunnel_en = 0;
1058         unsigned int mpls_en = 0;
1059         unsigned int swp = 0;
1060         unsigned int mprq = 0;
1061         unsigned int mprq_min_stride_size_n = 0;
1062         unsigned int mprq_max_stride_size_n = 0;
1063         unsigned int mprq_min_stride_num_n = 0;
1064         unsigned int mprq_max_stride_num_n = 0;
1065         struct ether_addr mac;
1066         char name[RTE_ETH_NAME_MAX_LEN];
1067         int own_domain_id = 0;
1068         uint16_t port_id;
1069         unsigned int i;
1070
1071         /* Determine if this port representor is supposed to be spawned. */
1072         if (switch_info->representor && dpdk_dev->devargs) {
1073                 struct rte_eth_devargs eth_da;
1074
1075                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
1076                 if (err) {
1077                         rte_errno = -err;
1078                         DRV_LOG(ERR, "failed to process device arguments: %s",
1079                                 strerror(rte_errno));
1080                         return NULL;
1081                 }
1082                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
1083                         if (eth_da.representor_ports[i] ==
1084                             (uint16_t)switch_info->port_name)
1085                                 break;
1086                 if (i == eth_da.nb_representor_ports) {
1087                         rte_errno = EBUSY;
1088                         return NULL;
1089                 }
1090         }
1091         /* Build device name. */
1092         if (!switch_info->representor)
1093                 strlcpy(name, dpdk_dev->name, sizeof(name));
1094         else
1095                 snprintf(name, sizeof(name), "%s_representor_%u",
1096                          dpdk_dev->name, switch_info->port_name);
1097         /* check if the device is already spawned */
1098         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1099                 rte_errno = EEXIST;
1100                 return NULL;
1101         }
1102         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1103         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1104                 eth_dev = rte_eth_dev_attach_secondary(name);
1105                 if (eth_dev == NULL) {
1106                         DRV_LOG(ERR, "can not attach rte ethdev");
1107                         rte_errno = ENOMEM;
1108                         return NULL;
1109                 }
1110                 eth_dev->device = dpdk_dev;
1111                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
1112                 err = mlx5_proc_priv_init(eth_dev);
1113                 if (err)
1114                         return NULL;
1115                 /* Receive command fd from primary process */
1116                 err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
1117                 if (err < 0)
1118                         return NULL;
1119                 /* Remap UAR for Tx queues. */
1120                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
1121                 if (err)
1122                         return NULL;
1123                 /*
1124                  * Ethdev pointer is still required as input since
1125                  * the primary device is not accessible from the
1126                  * secondary process.
1127                  */
1128                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1129                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1130                 return eth_dev;
1131         }
1132         sh = mlx5_alloc_shared_ibctx(spawn);
1133         if (!sh)
1134                 return NULL;
1135         config.devx = sh->devx;
1136 #ifdef HAVE_IBV_MLX5_MOD_SWP
1137         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1138 #endif
1139         /*
1140          * Multi-packet send is supported by ConnectX-4 Lx PF as well
1141          * as all ConnectX-5 devices.
1142          */
1143 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1144         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1145 #endif
1146 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1147         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1148 #endif
1149         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1150         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1151                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1152                         DRV_LOG(DEBUG, "enhanced MPW is supported");
1153                         mps = MLX5_MPW_ENHANCED;
1154                 } else {
1155                         DRV_LOG(DEBUG, "MPW is supported");
1156                         mps = MLX5_MPW;
1157                 }
1158         } else {
1159                 DRV_LOG(DEBUG, "MPW isn't supported");
1160                 mps = MLX5_MPW_DISABLED;
1161         }
1162 #ifdef HAVE_IBV_MLX5_MOD_SWP
1163         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1164                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1165         DRV_LOG(DEBUG, "SWP support: %u", swp);
1166 #endif
1167         config.swp = !!swp;
1168 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1169         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1170                 struct mlx5dv_striding_rq_caps mprq_caps =
1171                         dv_attr.striding_rq_caps;
1172
1173                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1174                         mprq_caps.min_single_stride_log_num_of_bytes);
1175                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1176                         mprq_caps.max_single_stride_log_num_of_bytes);
1177                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1178                         mprq_caps.min_single_wqe_log_num_of_strides);
1179                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1180                         mprq_caps.max_single_wqe_log_num_of_strides);
1181                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1182                         mprq_caps.supported_qpts);
1183                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1184                 mprq = 1;
1185                 mprq_min_stride_size_n =
1186                         mprq_caps.min_single_stride_log_num_of_bytes;
1187                 mprq_max_stride_size_n =
1188                         mprq_caps.max_single_stride_log_num_of_bytes;
1189                 mprq_min_stride_num_n =
1190                         mprq_caps.min_single_wqe_log_num_of_strides;
1191                 mprq_max_stride_num_n =
1192                         mprq_caps.max_single_wqe_log_num_of_strides;
1193                 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1194                                                    mprq_min_stride_num_n);
1195         }
1196 #endif
1197         if (RTE_CACHE_LINE_SIZE == 128 &&
1198             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1199                 cqe_comp = 0;
1200         else
1201                 cqe_comp = 1;
1202         config.cqe_comp = cqe_comp;
1203 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1204         /* Whether device supports 128B Rx CQE padding. */
1205         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1206                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1207 #endif
1208 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1209         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1210                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
1211                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1212                              (dv_attr.tunnel_offloads_caps &
1213                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1214         }
1215         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1216                 tunnel_en ? "" : "not ");
1217 #else
1218         DRV_LOG(WARNING,
1219                 "tunnel offloading disabled due to old OFED/rdma-core version");
1220 #endif
1221         config.tunnel_en = tunnel_en;
1222 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1223         mpls_en = ((dv_attr.tunnel_offloads_caps &
1224                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1225                    (dv_attr.tunnel_offloads_caps &
1226                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1227         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1228                 mpls_en ? "" : "not ");
1229 #else
1230         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1231                 " old OFED/rdma-core version or firmware configuration");
1232 #endif
1233         config.mpls_en = mpls_en;
1234         /* Check port status. */
1235         err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1236         if (err) {
1237                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1238                 goto error;
1239         }
1240         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1241                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1242                 err = EINVAL;
1243                 goto error;
1244         }
1245         if (port_attr.state != IBV_PORT_ACTIVE)
1246                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1247                         mlx5_glue->port_state_str(port_attr.state),
1248                         port_attr.state);
1249         /* Allocate private eth device data. */
1250         priv = rte_zmalloc("ethdev private structure",
1251                            sizeof(*priv),
1252                            RTE_CACHE_LINE_SIZE);
1253         if (priv == NULL) {
1254                 DRV_LOG(ERR, "priv allocation failure");
1255                 err = ENOMEM;
1256                 goto error;
1257         }
1258         priv->sh = sh;
1259         priv->ibv_port = spawn->ibv_port;
1260         priv->mtu = ETHER_MTU;
1261 #ifndef RTE_ARCH_64
1262         /* Initialize UAR access locks for 32bit implementations. */
1263         rte_spinlock_init(&priv->uar_lock_cq);
1264         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1265                 rte_spinlock_init(&priv->uar_lock[i]);
1266 #endif
1267         /* Some internal functions rely on Netlink sockets, open them now. */
1268         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1269         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1270         priv->nl_sn = 0;
1271         priv->representor = !!switch_info->representor;
1272         priv->master = !!switch_info->master;
1273         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1274         /*
1275          * Currently we support single E-Switch per PF configurations
1276          * only and vport_id field contains the vport index for
1277          * associated VF, which is deduced from representor port name.
1278          * For example, let's have the IB device port 10, it has
1279          * attached network device eth0, which has port name attribute
1280          * pf0vf2, we can deduce the VF number as 2, and set vport index
1281          * as 3 (2+1). This assigning schema should be changed if the
1282          * multiple E-Switch instances per PF configurations or/and PCI
1283          * subfunctions are added.
1284          */
1285         priv->vport_id = switch_info->representor ?
1286                          switch_info->port_name + 1 : -1;
1287         /* representor_id field keeps the unmodified port/VF index. */
1288         priv->representor_id = switch_info->representor ?
1289                                switch_info->port_name : -1;
1290         /*
1291          * Look for sibling devices in order to reuse their switch domain
1292          * if any, otherwise allocate one.
1293          */
1294         RTE_ETH_FOREACH_DEV_OF(port_id, dpdk_dev) {
1295                 const struct mlx5_priv *opriv =
1296                         rte_eth_devices[port_id].data->dev_private;
1297
1298                 if (!opriv ||
1299                         opriv->domain_id ==
1300                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1301                         continue;
1302                 priv->domain_id = opriv->domain_id;
1303                 break;
1304         }
1305         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1306                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1307                 if (err) {
1308                         err = rte_errno;
1309                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
1310                                 strerror(rte_errno));
1311                         goto error;
1312                 }
1313                 own_domain_id = 1;
1314         }
1315         err = mlx5_args(&config, dpdk_dev->devargs);
1316         if (err) {
1317                 err = rte_errno;
1318                 DRV_LOG(ERR, "failed to process device arguments: %s",
1319                         strerror(rte_errno));
1320                 goto error;
1321         }
1322         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1323                             IBV_DEVICE_RAW_IP_CSUM);
1324         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1325                 (config.hw_csum ? "" : "not "));
1326 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1327         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1328         DRV_LOG(DEBUG, "counters are not supported");
1329 #endif
1330 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1331         if (config.dv_flow_en) {
1332                 DRV_LOG(WARNING, "DV flow is not supported");
1333                 config.dv_flow_en = 0;
1334         }
1335 #endif
1336         config.ind_table_max_size =
1337                 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1338         /*
1339          * Remove this check once DPDK supports larger/variable
1340          * indirection tables.
1341          */
1342         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1343                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1344         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1345                 config.ind_table_max_size);
1346         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1347                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1348         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1349                 (config.hw_vlan_strip ? "" : "not "));
1350         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1351                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
1352         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1353                 (config.hw_fcs_strip ? "" : "not "));
1354 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1355         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1356 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1357         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1358                         IBV_DEVICE_PCI_WRITE_END_PADDING);
1359 #endif
1360         if (config.hw_padding && !hw_padding) {
1361                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1362                 config.hw_padding = 0;
1363         } else if (config.hw_padding) {
1364                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1365         }
1366         config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1367                       (sh->device_attr.tso_caps.supported_qpts &
1368                        (1 << IBV_QPT_RAW_PACKET)));
1369         if (config.tso)
1370                 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1371         /*
1372          * MPW is disabled by default, while the Enhanced MPW is enabled
1373          * by default.
1374          */
1375         if (config.mps == MLX5_ARG_UNSET)
1376                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1377                                                           MLX5_MPW_DISABLED;
1378         else
1379                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1380         DRV_LOG(INFO, "%sMPS is %s",
1381                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1382                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1383         if (config.cqe_comp && !cqe_comp) {
1384                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1385                 config.cqe_comp = 0;
1386         }
1387         if (config.cqe_pad && !cqe_pad) {
1388                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1389                 config.cqe_pad = 0;
1390         } else if (config.cqe_pad) {
1391                 DRV_LOG(INFO, "Rx CQE padding is enabled");
1392         }
1393         if (config.mprq.enabled && mprq) {
1394                 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1395                     config.mprq.stride_num_n < mprq_min_stride_num_n) {
1396                         config.mprq.stride_num_n =
1397                                 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1398                                         mprq_min_stride_num_n);
1399                         DRV_LOG(WARNING,
1400                                 "the number of strides"
1401                                 " for Multi-Packet RQ is out of range,"
1402                                 " setting default value (%u)",
1403                                 1 << config.mprq.stride_num_n);
1404                 }
1405                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1406                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1407         } else if (config.mprq.enabled && !mprq) {
1408                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1409                 config.mprq.enabled = 0;
1410         }
1411         eth_dev = rte_eth_dev_allocate(name);
1412         if (eth_dev == NULL) {
1413                 DRV_LOG(ERR, "can not allocate rte ethdev");
1414                 err = ENOMEM;
1415                 goto error;
1416         }
1417         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1418         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1419         if (priv->representor) {
1420                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1421                 eth_dev->data->representor_id = priv->representor_id;
1422         }
1423         eth_dev->data->dev_private = priv;
1424         priv->dev_data = eth_dev->data;
1425         eth_dev->data->mac_addrs = priv->mac;
1426         eth_dev->device = dpdk_dev;
1427         /* Configure the first MAC address by default. */
1428         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1429                 DRV_LOG(ERR,
1430                         "port %u cannot get MAC address, is mlx5_en"
1431                         " loaded? (errno: %s)",
1432                         eth_dev->data->port_id, strerror(rte_errno));
1433                 err = ENODEV;
1434                 goto error;
1435         }
1436         DRV_LOG(INFO,
1437                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1438                 eth_dev->data->port_id,
1439                 mac.addr_bytes[0], mac.addr_bytes[1],
1440                 mac.addr_bytes[2], mac.addr_bytes[3],
1441                 mac.addr_bytes[4], mac.addr_bytes[5]);
1442 #ifndef NDEBUG
1443         {
1444                 char ifname[IF_NAMESIZE];
1445
1446                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1447                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1448                                 eth_dev->data->port_id, ifname);
1449                 else
1450                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1451                                 eth_dev->data->port_id);
1452         }
1453 #endif
1454         /* Get actual MTU if possible. */
1455         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1456         if (err) {
1457                 err = rte_errno;
1458                 goto error;
1459         }
1460         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1461                 priv->mtu);
1462         /* Initialize burst functions to prevent crashes before link-up. */
1463         eth_dev->rx_pkt_burst = removed_rx_burst;
1464         eth_dev->tx_pkt_burst = removed_tx_burst;
1465         eth_dev->dev_ops = &mlx5_dev_ops;
1466         /* Register MAC address. */
1467         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1468         if (config.vf && config.vf_nl_en)
1469                 mlx5_nl_mac_addr_sync(eth_dev);
1470         priv->tcf_context = mlx5_flow_tcf_context_create();
1471         if (!priv->tcf_context) {
1472                 err = -rte_errno;
1473                 DRV_LOG(WARNING,
1474                         "flow rules relying on switch offloads will not be"
1475                         " supported: cannot open libmnl socket: %s",
1476                         strerror(rte_errno));
1477         } else {
1478                 struct rte_flow_error error;
1479                 unsigned int ifindex = mlx5_ifindex(eth_dev);
1480
1481                 if (!ifindex) {
1482                         err = -rte_errno;
1483                         error.message =
1484                                 "cannot retrieve network interface index";
1485                 } else {
1486                         err = mlx5_flow_tcf_init(priv->tcf_context,
1487                                                  ifindex, &error);
1488                 }
1489                 if (err) {
1490                         DRV_LOG(WARNING,
1491                                 "flow rules relying on switch offloads will"
1492                                 " not be supported: %s: %s",
1493                                 error.message, strerror(rte_errno));
1494                         mlx5_flow_tcf_context_destroy(priv->tcf_context);
1495                         priv->tcf_context = NULL;
1496                 }
1497         }
1498         TAILQ_INIT(&priv->flows);
1499         TAILQ_INIT(&priv->ctrl_flows);
1500         /* Hint libmlx5 to use PMD allocator for data plane resources */
1501         struct mlx5dv_ctx_allocators alctr = {
1502                 .alloc = &mlx5_alloc_verbs_buf,
1503                 .free = &mlx5_free_verbs_buf,
1504                 .data = priv,
1505         };
1506         mlx5_glue->dv_set_context_attr(sh->ctx,
1507                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1508                                        (void *)((uintptr_t)&alctr));
1509         /* Bring Ethernet device up. */
1510         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1511                 eth_dev->data->port_id);
1512         mlx5_set_link_up(eth_dev);
1513         /*
1514          * Even though the interrupt handler is not installed yet,
1515          * interrupts will still trigger on the async_fd from
1516          * Verbs context returned by ibv_open_device().
1517          */
1518         mlx5_link_update(eth_dev, 0);
1519 #ifdef HAVE_IBV_DEVX_OBJ
1520         err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
1521         if (err) {
1522                 err = -err;
1523                 goto error;
1524         }
1525 #endif
1526 #ifdef HAVE_MLX5DV_DR_ESWITCH
1527         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1528               (switch_info->representor || switch_info->master)))
1529                 config.dv_esw_en = 0;
1530 #else
1531         config.dv_esw_en = 0;
1532 #endif
1533         /* Store device configuration on private structure. */
1534         priv->config = config;
1535         if (config.dv_flow_en) {
1536                 err = mlx5_alloc_shared_dr(priv);
1537                 if (err)
1538                         goto error;
1539         }
1540         /* Supported Verbs flow priority number detection. */
1541         err = mlx5_flow_discover_priorities(eth_dev);
1542         if (err < 0) {
1543                 err = -err;
1544                 goto error;
1545         }
1546         priv->config.flow_prio = err;
1547         /*
1548          * Once the device is added to the list of memory event
1549          * callback, its global MR cache table cannot be expanded
1550          * on the fly because of deadlock. If it overflows, lookup
1551          * should be done by searching MR list linearly, which is slow.
1552          */
1553         err = mlx5_mr_btree_init(&priv->mr.cache,
1554                                  MLX5_MR_BTREE_CACHE_N * 2,
1555                                  eth_dev->device->numa_node);
1556         if (err) {
1557                 err = rte_errno;
1558                 goto error;
1559         }
1560         /* Add device to memory callback list. */
1561         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1562         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1563                          priv, mem_event_cb);
1564         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1565         return eth_dev;
1566 error:
1567         if (priv) {
1568                 if (priv->sh)
1569                         mlx5_free_shared_dr(priv);
1570                 if (priv->nl_socket_route >= 0)
1571                         close(priv->nl_socket_route);
1572                 if (priv->nl_socket_rdma >= 0)
1573                         close(priv->nl_socket_rdma);
1574                 if (priv->tcf_context)
1575                         mlx5_flow_tcf_context_destroy(priv->tcf_context);
1576                 if (own_domain_id)
1577                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1578                 rte_free(priv);
1579                 if (eth_dev != NULL)
1580                         eth_dev->data->dev_private = NULL;
1581         }
1582         if (eth_dev != NULL) {
1583                 /* mac_addrs must not be freed alone because part of dev_private */
1584                 eth_dev->data->mac_addrs = NULL;
1585                 rte_eth_dev_release_port(eth_dev);
1586         }
1587         if (sh)
1588                 mlx5_free_shared_ibctx(sh);
1589         assert(err > 0);
1590         rte_errno = err;
1591         return NULL;
1592 }
1593
1594 /**
1595  * Comparison callback to sort device data.
1596  *
1597  * This is meant to be used with qsort().
1598  *
1599  * @param a[in]
1600  *   Pointer to pointer to first data object.
1601  * @param b[in]
1602  *   Pointer to pointer to second data object.
1603  *
1604  * @return
1605  *   0 if both objects are equal, less than 0 if the first argument is less
1606  *   than the second, greater than 0 otherwise.
1607  */
1608 static int
1609 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1610 {
1611         const struct mlx5_switch_info *si_a =
1612                 &((const struct mlx5_dev_spawn_data *)a)->info;
1613         const struct mlx5_switch_info *si_b =
1614                 &((const struct mlx5_dev_spawn_data *)b)->info;
1615         int ret;
1616
1617         /* Master device first. */
1618         ret = si_b->master - si_a->master;
1619         if (ret)
1620                 return ret;
1621         /* Then representor devices. */
1622         ret = si_b->representor - si_a->representor;
1623         if (ret)
1624                 return ret;
1625         /* Unidentified devices come last in no specific order. */
1626         if (!si_a->representor)
1627                 return 0;
1628         /* Order representors by name. */
1629         return si_a->port_name - si_b->port_name;
1630 }
1631
1632 /**
1633  * DPDK callback to register a PCI device.
1634  *
1635  * This function spawns Ethernet devices out of a given PCI device.
1636  *
1637  * @param[in] pci_drv
1638  *   PCI driver structure (mlx5_driver).
1639  * @param[in] pci_dev
1640  *   PCI device information.
1641  *
1642  * @return
1643  *   0 on success, a negative errno value otherwise and rte_errno is set.
1644  */
1645 static int
1646 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1647                struct rte_pci_device *pci_dev)
1648 {
1649         struct ibv_device **ibv_list;
1650         /*
1651          * Number of found IB Devices matching with requested PCI BDF.
1652          * nd != 1 means there are multiple IB devices over the same
1653          * PCI device and we have representors and master.
1654          */
1655         unsigned int nd = 0;
1656         /*
1657          * Number of found IB device Ports. nd = 1 and np = 1..n means
1658          * we have the single multiport IB device, and there may be
1659          * representors attached to some of found ports.
1660          */
1661         unsigned int np = 0;
1662         /*
1663          * Number of DPDK ethernet devices to Spawn - either over
1664          * multiple IB devices or multiple ports of single IB device.
1665          * Actually this is the number of iterations to spawn.
1666          */
1667         unsigned int ns = 0;
1668         struct mlx5_dev_config dev_config;
1669         int ret;
1670
1671         ret = mlx5_init_once();
1672         if (ret) {
1673                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1674                         strerror(rte_errno));
1675                 return -rte_errno;
1676         }
1677         assert(pci_drv == &mlx5_driver);
1678         errno = 0;
1679         ibv_list = mlx5_glue->get_device_list(&ret);
1680         if (!ibv_list) {
1681                 rte_errno = errno ? errno : ENOSYS;
1682                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1683                 return -rte_errno;
1684         }
1685         /*
1686          * First scan the list of all Infiniband devices to find
1687          * matching ones, gathering into the list.
1688          */
1689         struct ibv_device *ibv_match[ret + 1];
1690         int nl_route = -1;
1691         int nl_rdma = -1;
1692         unsigned int i;
1693
1694         while (ret-- > 0) {
1695                 struct rte_pci_addr pci_addr;
1696
1697                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1698                 if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
1699                         continue;
1700                 if (pci_dev->addr.domain != pci_addr.domain ||
1701                     pci_dev->addr.bus != pci_addr.bus ||
1702                     pci_dev->addr.devid != pci_addr.devid ||
1703                     pci_dev->addr.function != pci_addr.function)
1704                         continue;
1705                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1706                         ibv_list[ret]->name);
1707                 ibv_match[nd++] = ibv_list[ret];
1708         }
1709         ibv_match[nd] = NULL;
1710         if (!nd) {
1711                 /* No device matches, just complain and bail out. */
1712                 mlx5_glue->free_device_list(ibv_list);
1713                 DRV_LOG(WARNING,
1714                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1715                         " are kernel drivers loaded?",
1716                         pci_dev->addr.domain, pci_dev->addr.bus,
1717                         pci_dev->addr.devid, pci_dev->addr.function);
1718                 rte_errno = ENOENT;
1719                 ret = -rte_errno;
1720                 return ret;
1721         }
1722         nl_route = mlx5_nl_init(NETLINK_ROUTE);
1723         nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1724         if (nd == 1) {
1725                 /*
1726                  * Found single matching device may have multiple ports.
1727                  * Each port may be representor, we have to check the port
1728                  * number and check the representors existence.
1729                  */
1730                 if (nl_rdma >= 0)
1731                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1732                 if (!np)
1733                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1734                                          " ports number", ibv_match[0]->name);
1735         }
1736         /*
1737          * Now we can determine the maximal
1738          * amount of devices to be spawned.
1739          */
1740         struct mlx5_dev_spawn_data list[np ? np : nd];
1741
1742         if (np > 1) {
1743                 /*
1744                  * Single IB device with multiple ports found,
1745                  * it may be E-Switch master device and representors.
1746                  * We have to perform identification trough the ports.
1747                  */
1748                 assert(nl_rdma >= 0);
1749                 assert(ns == 0);
1750                 assert(nd == 1);
1751                 for (i = 1; i <= np; ++i) {
1752                         list[ns].max_port = np;
1753                         list[ns].ibv_port = i;
1754                         list[ns].ibv_dev = ibv_match[0];
1755                         list[ns].eth_dev = NULL;
1756                         list[ns].ifindex = mlx5_nl_ifindex
1757                                         (nl_rdma, list[ns].ibv_dev->name, i);
1758                         if (!list[ns].ifindex) {
1759                                 /*
1760                                  * No network interface index found for the
1761                                  * specified port, it means there is no
1762                                  * representor on this port. It's OK,
1763                                  * there can be disabled ports, for example
1764                                  * if sriov_numvfs < sriov_totalvfs.
1765                                  */
1766                                 continue;
1767                         }
1768                         ret = -1;
1769                         if (nl_route >= 0)
1770                                 ret = mlx5_nl_switch_info
1771                                                (nl_route,
1772                                                 list[ns].ifindex,
1773                                                 &list[ns].info);
1774                         if (ret || (!list[ns].info.representor &&
1775                                     !list[ns].info.master)) {
1776                                 /*
1777                                  * We failed to recognize representors with
1778                                  * Netlink, let's try to perform the task
1779                                  * with sysfs.
1780                                  */
1781                                 ret =  mlx5_sysfs_switch_info
1782                                                 (list[ns].ifindex,
1783                                                  &list[ns].info);
1784                         }
1785                         if (!ret && (list[ns].info.representor ^
1786                                      list[ns].info.master))
1787                                 ns++;
1788                 }
1789                 if (!ns) {
1790                         DRV_LOG(ERR,
1791                                 "unable to recognize master/representors"
1792                                 " on the IB device with multiple ports");
1793                         rte_errno = ENOENT;
1794                         ret = -rte_errno;
1795                         goto exit;
1796                 }
1797         } else {
1798                 /*
1799                  * The existence of several matching entries (nd > 1) means
1800                  * port representors have been instantiated. No existing Verbs
1801                  * call nor sysfs entries can tell them apart, this can only
1802                  * be done through Netlink calls assuming kernel drivers are
1803                  * recent enough to support them.
1804                  *
1805                  * In the event of identification failure through Netlink,
1806                  * try again through sysfs, then:
1807                  *
1808                  * 1. A single IB device matches (nd == 1) with single
1809                  *    port (np=0/1) and is not a representor, assume
1810                  *    no switch support.
1811                  *
1812                  * 2. Otherwise no safe assumptions can be made;
1813                  *    complain louder and bail out.
1814                  */
1815                 np = 1;
1816                 for (i = 0; i != nd; ++i) {
1817                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1818                         list[ns].max_port = 1;
1819                         list[ns].ibv_port = 1;
1820                         list[ns].ibv_dev = ibv_match[i];
1821                         list[ns].eth_dev = NULL;
1822                         list[ns].ifindex = 0;
1823                         if (nl_rdma >= 0)
1824                                 list[ns].ifindex = mlx5_nl_ifindex
1825                                         (nl_rdma, list[ns].ibv_dev->name, 1);
1826                         if (!list[ns].ifindex) {
1827                                 char ifname[IF_NAMESIZE];
1828
1829                                 /*
1830                                  * Netlink failed, it may happen with old
1831                                  * ib_core kernel driver (before 4.16).
1832                                  * We can assume there is old driver because
1833                                  * here we are processing single ports IB
1834                                  * devices. Let's try sysfs to retrieve
1835                                  * the ifindex. The method works for
1836                                  * master device only.
1837                                  */
1838                                 if (nd > 1) {
1839                                         /*
1840                                          * Multiple devices found, assume
1841                                          * representors, can not distinguish
1842                                          * master/representor and retrieve
1843                                          * ifindex via sysfs.
1844                                          */
1845                                         continue;
1846                                 }
1847                                 ret = mlx5_get_master_ifname
1848                                         (ibv_match[i]->ibdev_path, &ifname);
1849                                 if (!ret)
1850                                         list[ns].ifindex =
1851                                                 if_nametoindex(ifname);
1852                                 if (!list[ns].ifindex) {
1853                                         /*
1854                                          * No network interface index found
1855                                          * for the specified device, it means
1856                                          * there it is neither representor
1857                                          * nor master.
1858                                          */
1859                                         continue;
1860                                 }
1861                         }
1862                         ret = -1;
1863                         if (nl_route >= 0)
1864                                 ret = mlx5_nl_switch_info
1865                                                (nl_route,
1866                                                 list[ns].ifindex,
1867                                                 &list[ns].info);
1868                         if (ret || (!list[ns].info.representor &&
1869                                     !list[ns].info.master)) {
1870                                 /*
1871                                  * We failed to recognize representors with
1872                                  * Netlink, let's try to perform the task
1873                                  * with sysfs.
1874                                  */
1875                                 ret =  mlx5_sysfs_switch_info
1876                                                 (list[ns].ifindex,
1877                                                  &list[ns].info);
1878                         }
1879                         if (!ret && (list[ns].info.representor ^
1880                                      list[ns].info.master)) {
1881                                 ns++;
1882                         } else if ((nd == 1) &&
1883                                    !list[ns].info.representor &&
1884                                    !list[ns].info.master) {
1885                                 /*
1886                                  * Single IB device with
1887                                  * one physical port and
1888                                  * attached network device.
1889                                  * May be SRIOV is not enabled
1890                                  * or there is no representors.
1891                                  */
1892                                 DRV_LOG(INFO, "no E-Switch support detected");
1893                                 ns++;
1894                                 break;
1895                         }
1896                 }
1897                 if (!ns) {
1898                         DRV_LOG(ERR,
1899                                 "unable to recognize master/representors"
1900                                 " on the multiple IB devices");
1901                         rte_errno = ENOENT;
1902                         ret = -rte_errno;
1903                         goto exit;
1904                 }
1905         }
1906         assert(ns);
1907         /*
1908          * Sort list to probe devices in natural order for users convenience
1909          * (i.e. master first, then representors from lowest to highest ID).
1910          */
1911         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1912         /* Default configuration. */
1913         dev_config = (struct mlx5_dev_config){
1914                 .hw_padding = 0,
1915                 .mps = MLX5_ARG_UNSET,
1916                 .tx_vec_en = 1,
1917                 .rx_vec_en = 1,
1918                 .txq_inline = MLX5_ARG_UNSET,
1919                 .txqs_inline = MLX5_ARG_UNSET,
1920                 .txqs_vec = MLX5_ARG_UNSET,
1921                 .inline_max_packet_sz = MLX5_ARG_UNSET,
1922                 .vf_nl_en = 1,
1923                 .mr_ext_memseg_en = 1,
1924                 .mprq = {
1925                         .enabled = 0, /* Disabled by default. */
1926                         .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
1927                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1928                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1929                 },
1930                 .dv_esw_en = 1,
1931         };
1932         /* Device specific configuration. */
1933         switch (pci_dev->id.device_id) {
1934         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BF:
1935                 dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS_BLUEFIELD;
1936                 break;
1937         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1938         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1939         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1940         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1941                 dev_config.vf = 1;
1942                 break;
1943         default:
1944                 break;
1945         }
1946         /* Set architecture-dependent default value if unset. */
1947         if (dev_config.txqs_vec == MLX5_ARG_UNSET)
1948                 dev_config.txqs_vec = MLX5_VPMD_MAX_TXQS;
1949         for (i = 0; i != ns; ++i) {
1950                 uint32_t restore;
1951
1952                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1953                                                  &list[i],
1954                                                  dev_config);
1955                 if (!list[i].eth_dev) {
1956                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1957                                 break;
1958                         /* Device is disabled or already spawned. Ignore it. */
1959                         continue;
1960                 }
1961                 restore = list[i].eth_dev->data->dev_flags;
1962                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1963                 /* Restore non-PCI flags cleared by the above call. */
1964                 list[i].eth_dev->data->dev_flags |= restore;
1965                 rte_eth_dev_probing_finish(list[i].eth_dev);
1966         }
1967         if (i != ns) {
1968                 DRV_LOG(ERR,
1969                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1970                         " encountering an error: %s",
1971                         pci_dev->addr.domain, pci_dev->addr.bus,
1972                         pci_dev->addr.devid, pci_dev->addr.function,
1973                         strerror(rte_errno));
1974                 ret = -rte_errno;
1975                 /* Roll back. */
1976                 while (i--) {
1977                         if (!list[i].eth_dev)
1978                                 continue;
1979                         mlx5_dev_close(list[i].eth_dev);
1980                         /* mac_addrs must not be freed because in dev_private */
1981                         list[i].eth_dev->data->mac_addrs = NULL;
1982                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1983                 }
1984                 /* Restore original error. */
1985                 rte_errno = -ret;
1986         } else {
1987                 ret = 0;
1988         }
1989 exit:
1990         /*
1991          * Do the routine cleanup:
1992          * - close opened Netlink sockets
1993          * - free the Infiniband device list
1994          */
1995         if (nl_rdma >= 0)
1996                 close(nl_rdma);
1997         if (nl_route >= 0)
1998                 close(nl_route);
1999         assert(ibv_list);
2000         mlx5_glue->free_device_list(ibv_list);
2001         return ret;
2002 }
2003
2004 /**
2005  * DPDK callback to remove a PCI device.
2006  *
2007  * This function removes all Ethernet devices belong to a given PCI device.
2008  *
2009  * @param[in] pci_dev
2010  *   Pointer to the PCI device.
2011  *
2012  * @return
2013  *   0 on success, the function cannot fail.
2014  */
2015 static int
2016 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2017 {
2018         uint16_t port_id;
2019
2020         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
2021                 rte_eth_dev_close(port_id);
2022         return 0;
2023 }
2024
2025 static const struct rte_pci_id mlx5_pci_id_map[] = {
2026         {
2027                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2028                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2029         },
2030         {
2031                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2032                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2033         },
2034         {
2035                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2036                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2037         },
2038         {
2039                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2040                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2041         },
2042         {
2043                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2044                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2045         },
2046         {
2047                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2048                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2049         },
2050         {
2051                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2052                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2053         },
2054         {
2055                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2056                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2057         },
2058         {
2059                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2060                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2061         },
2062         {
2063                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2064                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2065         },
2066         {
2067                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2068                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2069         },
2070         {
2071                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2072                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2073         },
2074         {
2075                 .vendor_id = 0
2076         }
2077 };
2078
2079 static struct rte_pci_driver mlx5_driver = {
2080         .driver = {
2081                 .name = MLX5_DRIVER_NAME
2082         },
2083         .id_table = mlx5_pci_id_map,
2084         .probe = mlx5_pci_probe,
2085         .remove = mlx5_pci_remove,
2086         .dma_map = mlx5_dma_map,
2087         .dma_unmap = mlx5_dma_unmap,
2088         .drv_flags = (RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
2089                       RTE_PCI_DRV_PROBE_AGAIN),
2090 };
2091
2092 #ifdef RTE_IBVERBS_LINK_DLOPEN
2093
2094 /**
2095  * Suffix RTE_EAL_PMD_PATH with "-glue".
2096  *
2097  * This function performs a sanity check on RTE_EAL_PMD_PATH before
2098  * suffixing its last component.
2099  *
2100  * @param buf[out]
2101  *   Output buffer, should be large enough otherwise NULL is returned.
2102  * @param size
2103  *   Size of @p out.
2104  *
2105  * @return
2106  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
2107  */
2108 static char *
2109 mlx5_glue_path(char *buf, size_t size)
2110 {
2111         static const char *const bad[] = { "/", ".", "..", NULL };
2112         const char *path = RTE_EAL_PMD_PATH;
2113         size_t len = strlen(path);
2114         size_t off;
2115         int i;
2116
2117         while (len && path[len - 1] == '/')
2118                 --len;
2119         for (off = len; off && path[off - 1] != '/'; --off)
2120                 ;
2121         for (i = 0; bad[i]; ++i)
2122                 if (!strncmp(path + off, bad[i], (int)(len - off)))
2123                         goto error;
2124         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
2125         if (i == -1 || (size_t)i >= size)
2126                 goto error;
2127         return buf;
2128 error:
2129         DRV_LOG(ERR,
2130                 "unable to append \"-glue\" to last component of"
2131                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
2132                 " please re-configure DPDK");
2133         return NULL;
2134 }
2135
2136 /**
2137  * Initialization routine for run-time dependency on rdma-core.
2138  */
2139 static int
2140 mlx5_glue_init(void)
2141 {
2142         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
2143         const char *path[] = {
2144                 /*
2145                  * A basic security check is necessary before trusting
2146                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
2147                  */
2148                 (geteuid() == getuid() && getegid() == getgid() ?
2149                  getenv("MLX5_GLUE_PATH") : NULL),
2150                 /*
2151                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
2152                  * variant, otherwise let dlopen() look up libraries on its
2153                  * own.
2154                  */
2155                 (*RTE_EAL_PMD_PATH ?
2156                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
2157         };
2158         unsigned int i = 0;
2159         void *handle = NULL;
2160         void **sym;
2161         const char *dlmsg;
2162
2163         while (!handle && i != RTE_DIM(path)) {
2164                 const char *end;
2165                 size_t len;
2166                 int ret;
2167
2168                 if (!path[i]) {
2169                         ++i;
2170                         continue;
2171                 }
2172                 end = strpbrk(path[i], ":;");
2173                 if (!end)
2174                         end = path[i] + strlen(path[i]);
2175                 len = end - path[i];
2176                 ret = 0;
2177                 do {
2178                         char name[ret + 1];
2179
2180                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
2181                                        (int)len, path[i],
2182                                        (!len || *(end - 1) == '/') ? "" : "/");
2183                         if (ret == -1)
2184                                 break;
2185                         if (sizeof(name) != (size_t)ret + 1)
2186                                 continue;
2187                         DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
2188                                 name);
2189                         handle = dlopen(name, RTLD_LAZY);
2190                         break;
2191                 } while (1);
2192                 path[i] = end + 1;
2193                 if (!*end)
2194                         ++i;
2195         }
2196         if (!handle) {
2197                 rte_errno = EINVAL;
2198                 dlmsg = dlerror();
2199                 if (dlmsg)
2200                         DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
2201                 goto glue_error;
2202         }
2203         sym = dlsym(handle, "mlx5_glue");
2204         if (!sym || !*sym) {
2205                 rte_errno = EINVAL;
2206                 dlmsg = dlerror();
2207                 if (dlmsg)
2208                         DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
2209                 goto glue_error;
2210         }
2211         mlx5_glue = *sym;
2212         return 0;
2213 glue_error:
2214         if (handle)
2215                 dlclose(handle);
2216         DRV_LOG(WARNING,
2217                 "cannot initialize PMD due to missing run-time dependency on"
2218                 " rdma-core libraries (libibverbs, libmlx5)");
2219         return -rte_errno;
2220 }
2221
2222 #endif
2223
2224 /**
2225  * Driver initialization routine.
2226  */
2227 RTE_INIT(rte_mlx5_pmd_init)
2228 {
2229         /* Initialize driver log type. */
2230         mlx5_logtype = rte_log_register("pmd.net.mlx5");
2231         if (mlx5_logtype >= 0)
2232                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
2233
2234         /* Build the static tables for Verbs conversion. */
2235         mlx5_set_ptype_table();
2236         mlx5_set_cksum_table();
2237         mlx5_set_swp_types_table();
2238         /*
2239          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
2240          * huge pages. Calling ibv_fork_init() during init allows
2241          * applications to use fork() safely for purposes other than
2242          * using this PMD, which is not supported in forked processes.
2243          */
2244         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
2245         /* Match the size of Rx completion entry to the size of a cacheline. */
2246         if (RTE_CACHE_LINE_SIZE == 128)
2247                 setenv("MLX5_CQE_SIZE", "128", 0);
2248         /*
2249          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
2250          * cleanup all the Verbs resources even when the device was removed.
2251          */
2252         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
2253 #ifdef RTE_IBVERBS_LINK_DLOPEN
2254         if (mlx5_glue_init())
2255                 return;
2256         assert(mlx5_glue);
2257 #endif
2258 #ifndef NDEBUG
2259         /* Glue structure must not contain any NULL pointers. */
2260         {
2261                 unsigned int i;
2262
2263                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
2264                         assert(((const void *const *)mlx5_glue)[i]);
2265         }
2266 #endif
2267         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
2268                 DRV_LOG(ERR,
2269                         "rdma-core glue \"%s\" mismatch: \"%s\" is required",
2270                         mlx5_glue->version, MLX5_GLUE_VERSION);
2271                 return;
2272         }
2273         mlx5_glue->fork_init();
2274         rte_pci_register(&mlx5_driver);
2275 }
2276
2277 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2278 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2279 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");