drivers/net/mlx5/mlx5.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2015 6WIND S.A.
   5  *   Copyright 2015 Mellanox.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of 6WIND S.A. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stddef.h>
  35 #include <unistd.h>
  36 #include <string.h>
  37 #include <assert.h>
  38 #include <stdint.h>
  39 #include <stdlib.h>
  40 #include <net/if.h>
  41
  42 /* Verbs header. */
  43 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  44 #ifdef PEDANTIC
  45 #pragma GCC diagnostic ignored "-pedantic"
  46 #endif
  47 #include <infiniband/verbs.h>
  48 #ifdef PEDANTIC
  49 #pragma GCC diagnostic error "-pedantic"
  50 #endif
  51
  52 /* DPDK headers don't like -pedantic. */
  53 #ifdef PEDANTIC
  54 #pragma GCC diagnostic ignored "-pedantic"
  55 #endif
  56 #include <rte_malloc.h>
  57 #include <rte_ethdev.h>
  58 #include <rte_pci.h>
  59 #include <rte_common.h>
  60 #ifdef PEDANTIC
  61 #pragma GCC diagnostic error "-pedantic"
  62 #endif
  63
  64 #include "mlx5.h"
  65 #include "mlx5_utils.h"
  66 #include "mlx5_rxtx.h"
  67 #include "mlx5_autoconf.h"
  68 #include "mlx5_defs.h"
  69
  70 /**
  71  * DPDK callback to close the device.
  72  *
  73  * Destroy all queues and objects, free memory.
  74  *
  75  * @param dev
  76  *   Pointer to Ethernet device structure.
  77  */
  78 static void
  79 mlx5_dev_close(struct rte_eth_dev *dev)
  80 {
  81         struct priv *priv = mlx5_get_priv(dev);
  82         void *tmp;
  83         unsigned int i;
  84
  85         priv_lock(priv);
  86         DEBUG("%p: closing device \"%s\"",
  87               (void *)dev,
  88               ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
  89         /* In case mlx5_dev_stop() has not been called. */
  90         priv_dev_interrupt_handler_uninstall(priv, dev);
  91         priv_special_flow_disable_all(priv);
  92         priv_mac_addrs_disable(priv);
  93         priv_destroy_hash_rxqs(priv);
  94
  95         /* Remove flow director elements. */
  96         priv_fdir_disable(priv);
  97         priv_fdir_delete_filters_list(priv);
  98
  99         /* Prevent crashes when queues are still in use. */
 100         dev->rx_pkt_burst = removed_rx_burst;
 101         dev->tx_pkt_burst = removed_tx_burst;
 102         if (priv->rxqs != NULL) {
 103                 /* XXX race condition if mlx5_rx_burst() is still running. */
 104                 usleep(1000);
 105                 for (i = 0; (i != priv->rxqs_n); ++i) {
 106                         tmp = (*priv->rxqs)[i];
 107                         if (tmp == NULL)
 108                                 continue;
 109                         (*priv->rxqs)[i] = NULL;
 110                         rxq_cleanup(tmp);
 111                         rte_free(tmp);
 112                 }
 113                 priv->rxqs_n = 0;
 114                 priv->rxqs = NULL;
 115         }
 116         if (priv->txqs != NULL) {
 117                 /* XXX race condition if mlx5_tx_burst() is still running. */
 118                 usleep(1000);
 119                 for (i = 0; (i != priv->txqs_n); ++i) {
 120                         tmp = (*priv->txqs)[i];
 121                         if (tmp == NULL)
 122                                 continue;
 123                         (*priv->txqs)[i] = NULL;
 124                         txq_cleanup(tmp);
 125                         rte_free(tmp);
 126                 }
 127                 priv->txqs_n = 0;
 128                 priv->txqs = NULL;
 129         }
 130         if (priv->pd != NULL) {
 131                 assert(priv->ctx != NULL);
 132                 claim_zero(ibv_dealloc_pd(priv->pd));
 133                 claim_zero(ibv_close_device(priv->ctx));
 134         } else
 135                 assert(priv->ctx == NULL);
 136         if (priv->rss_conf != NULL) {
 137                 for (i = 0; (i != hash_rxq_init_n); ++i)
 138                         rte_free((*priv->rss_conf)[i]);
 139                 rte_free(priv->rss_conf);
 140         }
 141         if (priv->reta_idx != NULL)
 142                 rte_free(priv->reta_idx);
 143         priv_unlock(priv);
 144         memset(priv, 0, sizeof(*priv));
 145 }
 146
 147 static const struct eth_dev_ops mlx5_dev_ops = {
 148         .dev_configure = mlx5_dev_configure,
 149         .dev_start = mlx5_dev_start,
 150         .dev_stop = mlx5_dev_stop,
 151         .dev_set_link_down = mlx5_set_link_down,
 152         .dev_set_link_up = mlx5_set_link_up,
 153         .dev_close = mlx5_dev_close,
 154         .promiscuous_enable = mlx5_promiscuous_enable,
 155         .promiscuous_disable = mlx5_promiscuous_disable,
 156         .allmulticast_enable = mlx5_allmulticast_enable,
 157         .allmulticast_disable = mlx5_allmulticast_disable,
 158         .link_update = mlx5_link_update,
 159         .stats_get = mlx5_stats_get,
 160         .stats_reset = mlx5_stats_reset,
 161         .dev_infos_get = mlx5_dev_infos_get,
 162         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 163         .vlan_filter_set = mlx5_vlan_filter_set,
 164         .rx_queue_setup = mlx5_rx_queue_setup,
 165         .tx_queue_setup = mlx5_tx_queue_setup,
 166         .rx_queue_release = mlx5_rx_queue_release,
 167         .tx_queue_release = mlx5_tx_queue_release,
 168         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 169         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 170         .mac_addr_remove = mlx5_mac_addr_remove,
 171         .mac_addr_add = mlx5_mac_addr_add,
 172         .mac_addr_set = mlx5_mac_addr_set,
 173         .mtu_set = mlx5_dev_set_mtu,
 174 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
 175         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 176         .vlan_offload_set = mlx5_vlan_offload_set,
 177 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
 178         .reta_update = mlx5_dev_rss_reta_update,
 179         .reta_query = mlx5_dev_rss_reta_query,
 180         .rss_hash_update = mlx5_rss_hash_update,
 181         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
 182 #ifdef MLX5_FDIR_SUPPORT
 183         .filter_ctrl = mlx5_dev_filter_ctrl,
 184 #endif /* MLX5_FDIR_SUPPORT */
 185 };
 186
 187 static struct {
 188         struct rte_pci_addr pci_addr; /* associated PCI address */
 189         uint32_t ports; /* physical ports bitfield. */
 190 } mlx5_dev[32];
 191
 192 /**
 193  * Get device index in mlx5_dev[] from PCI bus address.
 194  *
 195  * @param[in] pci_addr
 196  *   PCI bus address to look for.
 197  *
 198  * @return
 199  *   mlx5_dev[] index on success, -1 on failure.
 200  */
 201 static int
 202 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
 203 {
 204         unsigned int i;
 205         int ret = -1;
 206
 207         assert(pci_addr != NULL);
 208         for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
 209                 if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
 210                     (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
 211                     (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
 212                     (mlx5_dev[i].pci_addr.function == pci_addr->function))
 213                         return i;
 214                 if ((mlx5_dev[i].ports == 0) && (ret == -1))
 215                         ret = i;
 216         }
 217         return ret;
 218 }
 219
 220 static struct eth_driver mlx5_driver;
 221
 222 /**
 223  * DPDK callback to register a PCI device.
 224  *
 225  * This function creates an Ethernet device for each port of a given
 226  * PCI device.
 227  *
 228  * @param[in] pci_drv
 229  *   PCI driver structure (mlx5_driver).
 230  * @param[in] pci_dev
 231  *   PCI device information.
 232  *
 233  * @return
 234  *   0 on success, negative errno value on failure.
 235  */
 236 static int
 237 mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 238 {
 239         struct ibv_device **list;
 240         struct ibv_device *ibv_dev;
 241         int err = 0;
 242         struct ibv_context *attr_ctx = NULL;
 243         struct ibv_device_attr device_attr;
 244         unsigned int vf;
 245         int idx;
 246         int i;
 247
 248         (void)pci_drv;
 249         assert(pci_drv == &mlx5_driver.pci_drv);
 250         /* Get mlx5_dev[] index. */
 251         idx = mlx5_dev_idx(&pci_dev->addr);
 252         if (idx == -1) {
 253                 ERROR("this driver cannot support any more adapters");
 254                 return -ENOMEM;
 255         }
 256         DEBUG("using driver device index %d", idx);
 257
 258         /* Save PCI address. */
 259         mlx5_dev[idx].pci_addr = pci_dev->addr;
 260         list = ibv_get_device_list(&i);
 261         if (list == NULL) {
 262                 assert(errno);
 263                 if (errno == ENOSYS) {
 264                         WARN("cannot list devices, is ib_uverbs loaded?");
 265                         return 0;
 266                 }
 267                 return -errno;
 268         }
 269         assert(i >= 0);
 270         /*
 271          * For each listed device, check related sysfs entry against
 272          * the provided PCI ID.
 273          */
 274         while (i != 0) {
 275                 struct rte_pci_addr pci_addr;
 276
 277                 --i;
 278                 DEBUG("checking device \"%s\"", list[i]->name);
 279                 if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
 280                         continue;
 281                 if ((pci_dev->addr.domain != pci_addr.domain) ||
 282                     (pci_dev->addr.bus != pci_addr.bus) ||
 283                     (pci_dev->addr.devid != pci_addr.devid) ||
 284                     (pci_dev->addr.function != pci_addr.function))
 285                         continue;
 286                 vf = ((pci_dev->id.device_id ==
 287                        PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
 288                       (pci_dev->id.device_id ==
 289                        PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF));
 290                 INFO("PCI information matches, using device \"%s\" (VF: %s)",
 291                      list[i]->name, (vf ? "true" : "false"));
 292                 attr_ctx = ibv_open_device(list[i]);
 293                 err = errno;
 294                 break;
 295         }
 296         if (attr_ctx == NULL) {
 297                 ibv_free_device_list(list);
 298                 switch (err) {
 299                 case 0:
 300                         WARN("cannot access device, is mlx5_ib loaded?");
 301                         return 0;
 302                 case EINVAL:
 303                         WARN("cannot use device, are drivers up to date?");
 304                         return 0;
 305                 }
 306                 assert(err > 0);
 307                 return -err;
 308         }
 309         ibv_dev = list[i];
 310
 311         DEBUG("device opened");
 312         if (ibv_query_device(attr_ctx, &device_attr))
 313                 goto error;
 314         INFO("%u port(s) detected", device_attr.phys_port_cnt);
 315
 316         for (i = 0; i < device_attr.phys_port_cnt; i++) {
 317                 uint32_t port = i + 1; /* ports are indexed from one */
 318                 uint32_t test = (1 << i);
 319                 struct ibv_context *ctx = NULL;
 320                 struct ibv_port_attr port_attr;
 321                 struct ibv_pd *pd = NULL;
 322                 struct priv *priv = NULL;
 323                 struct rte_eth_dev *eth_dev;
 324 #ifdef HAVE_EXP_QUERY_DEVICE
 325                 struct ibv_exp_device_attr exp_device_attr;
 326 #endif /* HAVE_EXP_QUERY_DEVICE */
 327                 struct ether_addr mac;
 328
 329 #ifdef HAVE_EXP_QUERY_DEVICE
 330                 exp_device_attr.comp_mask =
 331                         IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
 332                         IBV_EXP_DEVICE_ATTR_RX_HASH |
 333 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
 334                         IBV_EXP_DEVICE_ATTR_VLAN_OFFLOADS |
 335 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
 336                         0;
 337 #endif /* HAVE_EXP_QUERY_DEVICE */
 338
 339                 DEBUG("using port %u (%08" PRIx32 ")", port, test);
 340
 341                 ctx = ibv_open_device(ibv_dev);
 342                 if (ctx == NULL)
 343                         goto port_error;
 344
 345                 /* Check port status. */
 346                 err = ibv_query_port(ctx, port, &port_attr);
 347                 if (err) {
 348                         ERROR("port query failed: %s", strerror(err));
 349                         goto port_error;
 350                 }
 351
 352                 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 353                         ERROR("port %d is not configured in Ethernet mode",
 354                               port);
 355                         goto port_error;
 356                 }
 357
 358                 if (port_attr.state != IBV_PORT_ACTIVE)
 359                         DEBUG("port %d is not active: \"%s\" (%d)",
 360                               port, ibv_port_state_str(port_attr.state),
 361                               port_attr.state);
 362
 363                 /* Allocate protection domain. */
 364                 pd = ibv_alloc_pd(ctx);
 365                 if (pd == NULL) {
 366                         ERROR("PD allocation failure");
 367                         err = ENOMEM;
 368                         goto port_error;
 369                 }
 370
 371                 mlx5_dev[idx].ports |= test;
 372
 373                 /* from rte_ethdev.c */
 374                 priv = rte_zmalloc("ethdev private structure",
 375                                    sizeof(*priv),
 376                                    RTE_CACHE_LINE_SIZE);
 377                 if (priv == NULL) {
 378                         ERROR("priv allocation failure");
 379                         err = ENOMEM;
 380                         goto port_error;
 381                 }
 382
 383                 priv->ctx = ctx;
 384                 priv->device_attr = device_attr;
 385                 priv->port = port;
 386                 priv->pd = pd;
 387                 priv->mtu = ETHER_MTU;
 388 #ifdef HAVE_EXP_QUERY_DEVICE
 389                 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
 390                         ERROR("ibv_exp_query_device() failed");
 391                         goto port_error;
 392                 }
 393
 394                 priv->hw_csum =
 395                         ((exp_device_attr.exp_device_cap_flags &
 396                           IBV_EXP_DEVICE_RX_CSUM_TCP_UDP_PKT) &&
 397                          (exp_device_attr.exp_device_cap_flags &
 398                           IBV_EXP_DEVICE_RX_CSUM_IP_PKT));
 399                 DEBUG("checksum offloading is %ssupported",
 400                       (priv->hw_csum ? "" : "not "));
 401
 402                 priv->hw_csum_l2tun = !!(exp_device_attr.exp_device_cap_flags &
 403                                          IBV_EXP_DEVICE_VXLAN_SUPPORT);
 404                 DEBUG("L2 tunnel checksum offloads are %ssupported",
 405                       (priv->hw_csum_l2tun ? "" : "not "));
 406
 407                 priv->ind_table_max_size = exp_device_attr.rx_hash_caps.max_rwq_indirection_table_size;
 408                 /* Remove this check once DPDK supports larger/variable
 409                  * indirection tables. */
 410                 if (priv->ind_table_max_size > (unsigned int)RSS_INDIRECTION_TABLE_SIZE)
 411                         priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
 412                 DEBUG("maximum RX indirection table size is %u",
 413                       priv->ind_table_max_size);
 414 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
 415                 priv->hw_vlan_strip = !!(exp_device_attr.wq_vlan_offloads_cap &
 416                                          IBV_EXP_RECEIVE_WQ_CVLAN_STRIP);
 417 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
 418                 DEBUG("VLAN stripping is %ssupported",
 419                       (priv->hw_vlan_strip ? "" : "not "));
 420
 421 #else /* HAVE_EXP_QUERY_DEVICE */
 422                 priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
 423 #endif /* HAVE_EXP_QUERY_DEVICE */
 424
 425                 priv->vf = vf;
 426                 /* Allocate and register default RSS hash keys. */
 427                 priv->rss_conf = rte_calloc(__func__, hash_rxq_init_n,
 428                                             sizeof((*priv->rss_conf)[0]), 0);
 429                 if (priv->rss_conf == NULL) {
 430                         err = ENOMEM;
 431                         goto port_error;
 432                 }
 433                 err = rss_hash_rss_conf_new_key(priv,
 434                                                 rss_hash_default_key,
 435                                                 rss_hash_default_key_len,
 436                                                 ETH_RSS_PROTO_MASK);
 437                 if (err)
 438                         goto port_error;
 439                 /* Configure the first MAC address by default. */
 440                 if (priv_get_mac(priv, &mac.addr_bytes)) {
 441                         ERROR("cannot get MAC address, is mlx5_en loaded?"
 442                               " (errno: %s)", strerror(errno));
 443                         goto port_error;
 444                 }
 445                 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
 446                      priv->port,
 447                      mac.addr_bytes[0], mac.addr_bytes[1],
 448                      mac.addr_bytes[2], mac.addr_bytes[3],
 449                      mac.addr_bytes[4], mac.addr_bytes[5]);
 450                 /* Register MAC address. */
 451                 claim_zero(priv_mac_addr_add(priv, 0,
 452                                              (const uint8_t (*)[ETHER_ADDR_LEN])
 453                                              mac.addr_bytes));
 454                 /* Initialize FD filters list. */
 455                 err = fdir_init_filters_list(priv);
 456                 if (err)
 457                         goto port_error;
 458 #ifndef NDEBUG
 459                 {
 460                         char ifname[IF_NAMESIZE];
 461
 462                         if (priv_get_ifname(priv, &ifname) == 0)
 463                                 DEBUG("port %u ifname is \"%s\"",
 464                                       priv->port, ifname);
 465                         else
 466                                 DEBUG("port %u ifname is unknown", priv->port);
 467                 }
 468 #endif
 469                 /* Get actual MTU if possible. */
 470                 priv_get_mtu(priv, &priv->mtu);
 471                 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
 472
 473                 /* from rte_ethdev.c */
 474                 {
 475                         char name[RTE_ETH_NAME_MAX_LEN];
 476
 477                         snprintf(name, sizeof(name), "%s port %u",
 478                                  ibv_get_device_name(ibv_dev), port);
 479                         eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
 480                 }
 481                 if (eth_dev == NULL) {
 482                         ERROR("can not allocate rte ethdev");
 483                         err = ENOMEM;
 484                         goto port_error;
 485                 }
 486
 487                 /* Secondary processes have to use local storage for their
 488                  * private data as well as a copy of eth_dev->data, but this
 489                  * pointer must not be modified before burst functions are
 490                  * actually called. */
 491                 if (mlx5_is_secondary()) {
 492                         struct mlx5_secondary_data *sd =
 493                                 &mlx5_secondary_data[eth_dev->data->port_id];
 494                         sd->primary_priv = eth_dev->data->dev_private;
 495                         if (sd->primary_priv == NULL) {
 496                                 ERROR("no private data for port %u",
 497                                                 eth_dev->data->port_id);
 498                                 err = EINVAL;
 499                                 goto port_error;
 500                         }
 501                         sd->shared_dev_data = eth_dev->data;
 502                         rte_spinlock_init(&sd->lock);
 503                         memcpy(sd->data.name, sd->shared_dev_data->name,
 504                                    sizeof(sd->data.name));
 505                         sd->data.dev_private = priv;
 506                         sd->data.rx_mbuf_alloc_failed = 0;
 507                         sd->data.mtu = ETHER_MTU;
 508                         sd->data.port_id = sd->shared_dev_data->port_id;
 509                         sd->data.mac_addrs = priv->mac;
 510                         eth_dev->tx_pkt_burst = mlx5_tx_burst_secondary_setup;
 511                         eth_dev->rx_pkt_burst = mlx5_rx_burst_secondary_setup;
 512                 } else {
 513                         eth_dev->data->dev_private = priv;
 514                         eth_dev->data->rx_mbuf_alloc_failed = 0;
 515                         eth_dev->data->mtu = ETHER_MTU;
 516                         eth_dev->data->mac_addrs = priv->mac;
 517                 }
 518
 519                 eth_dev->pci_dev = pci_dev;
 520                 rte_eth_copy_pci_info(eth_dev, pci_dev);
 521                 eth_dev->driver = &mlx5_driver;
 522                 priv->dev = eth_dev;
 523                 eth_dev->dev_ops = &mlx5_dev_ops;
 524
 525                 TAILQ_INIT(&eth_dev->link_intr_cbs);
 526
 527                 /* Bring Ethernet device up. */
 528                 DEBUG("forcing Ethernet interface up");
 529                 priv_set_flags(priv, ~IFF_UP, IFF_UP);
 530                 continue;
 531
 532 port_error:
 533                 if (priv) {
 534                         rte_free(priv->rss_conf);
 535                         rte_free(priv);
 536                 }
 537                 if (pd)
 538                         claim_zero(ibv_dealloc_pd(pd));
 539                 if (ctx)
 540                         claim_zero(ibv_close_device(ctx));
 541                 break;
 542         }
 543
 544         /*
 545          * XXX if something went wrong in the loop above, there is a resource
 546          * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
 547          * long as the dpdk does not provide a way to deallocate a ethdev and a
 548          * way to enumerate the registered ethdevs to free the previous ones.
 549          */
 550
 551         /* no port found, complain */
 552         if (!mlx5_dev[idx].ports) {
 553                 err = ENODEV;
 554                 goto error;
 555         }
 556
 557 error:
 558         if (attr_ctx)
 559                 claim_zero(ibv_close_device(attr_ctx));
 560         if (list)
 561                 ibv_free_device_list(list);
 562         assert(err >= 0);
 563         return -err;
 564 }
 565
 566 static const struct rte_pci_id mlx5_pci_id_map[] = {
 567         {
 568                 .vendor_id = PCI_VENDOR_ID_MELLANOX,
 569                 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4,
 570                 .subsystem_vendor_id = PCI_ANY_ID,
 571                 .subsystem_device_id = PCI_ANY_ID
 572         },
 573         {
 574                 .vendor_id = PCI_VENDOR_ID_MELLANOX,
 575                 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4VF,
 576                 .subsystem_vendor_id = PCI_ANY_ID,
 577                 .subsystem_device_id = PCI_ANY_ID
 578         },
 579         {
 580                 .vendor_id = PCI_VENDOR_ID_MELLANOX,
 581                 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4LX,
 582                 .subsystem_vendor_id = PCI_ANY_ID,
 583                 .subsystem_device_id = PCI_ANY_ID
 584         },
 585         {
 586                 .vendor_id = PCI_VENDOR_ID_MELLANOX,
 587                 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF,
 588                 .subsystem_vendor_id = PCI_ANY_ID,
 589                 .subsystem_device_id = PCI_ANY_ID
 590         },
 591         {
 592                 .vendor_id = 0
 593         }
 594 };
 595
 596 static struct eth_driver mlx5_driver = {
 597         .pci_drv = {
 598                 .name = MLX5_DRIVER_NAME,
 599                 .id_table = mlx5_pci_id_map,
 600                 .devinit = mlx5_pci_devinit,
 601                 .drv_flags = RTE_PCI_DRV_INTR_LSC,
 602         },
 603         .dev_private_size = sizeof(struct priv)
 604 };
 605
 606 /**
 607  * Driver initialization routine.
 608  */
 609 static int
 610 rte_mlx5_pmd_init(const char *name, const char *args)
 611 {
 612         (void)name;
 613         (void)args;
 614         /*
 615          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
 616          * huge pages. Calling ibv_fork_init() during init allows
 617          * applications to use fork() safely for purposes other than
 618          * using this PMD, which is not supported in forked processes.
 619          */
 620         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
 621         ibv_fork_init();
 622         rte_eal_pci_register(&mlx5_driver.pci_drv);
 623         return 0;
 624 }
 625
 626 static struct rte_driver rte_mlx5_driver = {
 627         .type = PMD_PDEV,
 628         .name = MLX5_DRIVER_NAME,
 629         .init = rte_mlx5_pmd_init,
 630 };
 631
 632 PMD_REGISTER_DRIVER(rte_mlx5_driver)