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  * Retrieve integer value from environment variable.
  72  *
  73  * @param[in] name
  74  *   Environment variable name.
  75  *
  76  * @return
  77  *   Integer value, 0 if the variable is not set.
  78  */
  79 int
  80 mlx5_getenv_int(const char *name)
  81 {
  82         const char *val = getenv(name);
  83
  84         if (val == NULL)
  85                 return 0;
  86         return atoi(val);
  87 }
  88
  89 /**
  90  * DPDK callback to close the device.
  91  *
  92  * Destroy all queues and objects, free memory.
  93  *
  94  * @param dev
  95  *   Pointer to Ethernet device structure.
  96  */
  97 static void
  98 mlx5_dev_close(struct rte_eth_dev *dev)
  99 {
 100         struct priv *priv = mlx5_get_priv(dev);
 101         unsigned int i;
 102
 103         priv_lock(priv);
 104         DEBUG("%p: closing device \"%s\"",
 105               (void *)dev,
 106               ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
 107         /* In case mlx5_dev_stop() has not been called. */
 108         priv_dev_interrupt_handler_uninstall(priv, dev);
 109         priv_special_flow_disable_all(priv);
 110         priv_mac_addrs_disable(priv);
 111         priv_destroy_hash_rxqs(priv);
 112
 113         /* Remove flow director elements. */
 114         priv_fdir_disable(priv);
 115         priv_fdir_delete_filters_list(priv);
 116
 117         /* Prevent crashes when queues are still in use. */
 118         dev->rx_pkt_burst = removed_rx_burst;
 119         dev->tx_pkt_burst = removed_tx_burst;
 120         if (priv->rxqs != NULL) {
 121                 /* XXX race condition if mlx5_rx_burst() is still running. */
 122                 usleep(1000);
 123                 for (i = 0; (i != priv->rxqs_n); ++i) {
 124                         struct rxq *rxq = (*priv->rxqs)[i];
 125                         struct rxq_ctrl *rxq_ctrl;
 126
 127                         if (rxq == NULL)
 128                                 continue;
 129                         rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
 130                         (*priv->rxqs)[i] = NULL;
 131                         rxq_cleanup(rxq_ctrl);
 132                         rte_free(rxq_ctrl);
 133                 }
 134                 priv->rxqs_n = 0;
 135                 priv->rxqs = NULL;
 136         }
 137         if (priv->txqs != NULL) {
 138                 /* XXX race condition if mlx5_tx_burst() is still running. */
 139                 usleep(1000);
 140                 for (i = 0; (i != priv->txqs_n); ++i) {
 141                         struct txq *txq = (*priv->txqs)[i];
 142                         struct txq_ctrl *txq_ctrl;
 143
 144                         if (txq == NULL)
 145                                 continue;
 146                         txq_ctrl = container_of(txq, struct txq_ctrl, txq);
 147                         (*priv->txqs)[i] = NULL;
 148                         txq_cleanup(txq_ctrl);
 149                         rte_free(txq_ctrl);
 150                 }
 151                 priv->txqs_n = 0;
 152                 priv->txqs = NULL;
 153         }
 154         if (priv->pd != NULL) {
 155                 assert(priv->ctx != NULL);
 156                 claim_zero(ibv_dealloc_pd(priv->pd));
 157                 claim_zero(ibv_close_device(priv->ctx));
 158         } else
 159                 assert(priv->ctx == NULL);
 160         if (priv->rss_conf != NULL) {
 161                 for (i = 0; (i != hash_rxq_init_n); ++i)
 162                         rte_free((*priv->rss_conf)[i]);
 163                 rte_free(priv->rss_conf);
 164         }
 165         if (priv->reta_idx != NULL)
 166                 rte_free(priv->reta_idx);
 167         priv_unlock(priv);
 168         memset(priv, 0, sizeof(*priv));
 169 }
 170
 171 static const struct eth_dev_ops mlx5_dev_ops = {
 172         .dev_configure = mlx5_dev_configure,
 173         .dev_start = mlx5_dev_start,
 174         .dev_stop = mlx5_dev_stop,
 175         .dev_set_link_down = mlx5_set_link_down,
 176         .dev_set_link_up = mlx5_set_link_up,
 177         .dev_close = mlx5_dev_close,
 178         .promiscuous_enable = mlx5_promiscuous_enable,
 179         .promiscuous_disable = mlx5_promiscuous_disable,
 180         .allmulticast_enable = mlx5_allmulticast_enable,
 181         .allmulticast_disable = mlx5_allmulticast_disable,
 182         .link_update = mlx5_link_update,
 183         .stats_get = mlx5_stats_get,
 184         .stats_reset = mlx5_stats_reset,
 185         .dev_infos_get = mlx5_dev_infos_get,
 186         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 187         .vlan_filter_set = mlx5_vlan_filter_set,
 188         .rx_queue_setup = mlx5_rx_queue_setup,
 189         .tx_queue_setup = mlx5_tx_queue_setup,
 190         .rx_queue_release = mlx5_rx_queue_release,
 191         .tx_queue_release = mlx5_tx_queue_release,
 192         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 193         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 194         .mac_addr_remove = mlx5_mac_addr_remove,
 195         .mac_addr_add = mlx5_mac_addr_add,
 196         .mac_addr_set = mlx5_mac_addr_set,
 197         .mtu_set = mlx5_dev_set_mtu,
 198         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 199         .vlan_offload_set = mlx5_vlan_offload_set,
 200         .reta_update = mlx5_dev_rss_reta_update,
 201         .reta_query = mlx5_dev_rss_reta_query,
 202         .rss_hash_update = mlx5_rss_hash_update,
 203         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
 204         .filter_ctrl = mlx5_dev_filter_ctrl,
 205 };
 206
 207 static struct {
 208         struct rte_pci_addr pci_addr; /* associated PCI address */
 209         uint32_t ports; /* physical ports bitfield. */
 210 } mlx5_dev[32];
 211
 212 /**
 213  * Get device index in mlx5_dev[] from PCI bus address.
 214  *
 215  * @param[in] pci_addr
 216  *   PCI bus address to look for.
 217  *
 218  * @return
 219  *   mlx5_dev[] index on success, -1 on failure.
 220  */
 221 static int
 222 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
 223 {
 224         unsigned int i;
 225         int ret = -1;
 226
 227         assert(pci_addr != NULL);
 228         for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
 229                 if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
 230                     (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
 231                     (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
 232                     (mlx5_dev[i].pci_addr.function == pci_addr->function))
 233                         return i;
 234                 if ((mlx5_dev[i].ports == 0) && (ret == -1))
 235                         ret = i;
 236         }
 237         return ret;
 238 }
 239
 240 static struct eth_driver mlx5_driver;
 241
 242 /**
 243  * DPDK callback to register a PCI device.
 244  *
 245  * This function creates an Ethernet device for each port of a given
 246  * PCI device.
 247  *
 248  * @param[in] pci_drv
 249  *   PCI driver structure (mlx5_driver).
 250  * @param[in] pci_dev
 251  *   PCI device information.
 252  *
 253  * @return
 254  *   0 on success, negative errno value on failure.
 255  */
 256 static int
 257 mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 258 {
 259         struct ibv_device **list;
 260         struct ibv_device *ibv_dev;
 261         int err = 0;
 262         struct ibv_context *attr_ctx = NULL;
 263         struct ibv_device_attr device_attr;
 264         unsigned int sriov;
 265         unsigned int mps;
 266         int idx;
 267         int i;
 268
 269         (void)pci_drv;
 270         assert(pci_drv == &mlx5_driver.pci_drv);
 271         /* Get mlx5_dev[] index. */
 272         idx = mlx5_dev_idx(&pci_dev->addr);
 273         if (idx == -1) {
 274                 ERROR("this driver cannot support any more adapters");
 275                 return -ENOMEM;
 276         }
 277         DEBUG("using driver device index %d", idx);
 278
 279         /* Save PCI address. */
 280         mlx5_dev[idx].pci_addr = pci_dev->addr;
 281         list = ibv_get_device_list(&i);
 282         if (list == NULL) {
 283                 assert(errno);
 284                 if (errno == ENOSYS) {
 285                         WARN("cannot list devices, is ib_uverbs loaded?");
 286                         return 0;
 287                 }
 288                 return -errno;
 289         }
 290         assert(i >= 0);
 291         /*
 292          * For each listed device, check related sysfs entry against
 293          * the provided PCI ID.
 294          */
 295         while (i != 0) {
 296                 struct rte_pci_addr pci_addr;
 297
 298                 --i;
 299                 DEBUG("checking device \"%s\"", list[i]->name);
 300                 if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
 301                         continue;
 302                 if ((pci_dev->addr.domain != pci_addr.domain) ||
 303                     (pci_dev->addr.bus != pci_addr.bus) ||
 304                     (pci_dev->addr.devid != pci_addr.devid) ||
 305                     (pci_dev->addr.function != pci_addr.function))
 306                         continue;
 307                 sriov = ((pci_dev->id.device_id ==
 308                        PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
 309                       (pci_dev->id.device_id ==
 310                        PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF));
 311                 /* Multi-packet send is only supported by ConnectX-4 Lx PF. */
 312                 mps = (pci_dev->id.device_id ==
 313                        PCI_DEVICE_ID_MELLANOX_CONNECTX4LX);
 314                 INFO("PCI information matches, using device \"%s\""
 315                      " (SR-IOV: %s, MPS: %s)",
 316                      list[i]->name,
 317                      sriov ? "true" : "false",
 318                      mps ? "true" : "false");
 319                 attr_ctx = ibv_open_device(list[i]);
 320                 err = errno;
 321                 break;
 322         }
 323         if (attr_ctx == NULL) {
 324                 ibv_free_device_list(list);
 325                 switch (err) {
 326                 case 0:
 327                         WARN("cannot access device, is mlx5_ib loaded?");
 328                         return 0;
 329                 case EINVAL:
 330                         WARN("cannot use device, are drivers up to date?");
 331                         return 0;
 332                 }
 333                 assert(err > 0);
 334                 return -err;
 335         }
 336         ibv_dev = list[i];
 337
 338         DEBUG("device opened");
 339         if (ibv_query_device(attr_ctx, &device_attr))
 340                 goto error;
 341         INFO("%u port(s) detected", device_attr.phys_port_cnt);
 342
 343         for (i = 0; i < device_attr.phys_port_cnt; i++) {
 344                 uint32_t port = i + 1; /* ports are indexed from one */
 345                 uint32_t test = (1 << i);
 346                 struct ibv_context *ctx = NULL;
 347                 struct ibv_port_attr port_attr;
 348                 struct ibv_pd *pd = NULL;
 349                 struct priv *priv = NULL;
 350                 struct rte_eth_dev *eth_dev;
 351                 struct ibv_exp_device_attr exp_device_attr;
 352                 struct ether_addr mac;
 353                 uint16_t num_vfs = 0;
 354
 355                 exp_device_attr.comp_mask =
 356                         IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
 357                         IBV_EXP_DEVICE_ATTR_RX_HASH |
 358                         IBV_EXP_DEVICE_ATTR_VLAN_OFFLOADS |
 359                         IBV_EXP_DEVICE_ATTR_RX_PAD_END_ALIGN |
 360                         0;
 361
 362                 DEBUG("using port %u (%08" PRIx32 ")", port, test);
 363
 364                 ctx = ibv_open_device(ibv_dev);
 365                 if (ctx == NULL)
 366                         goto port_error;
 367
 368                 /* Check port status. */
 369                 err = ibv_query_port(ctx, port, &port_attr);
 370                 if (err) {
 371                         ERROR("port query failed: %s", strerror(err));
 372                         goto port_error;
 373                 }
 374
 375                 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 376                         ERROR("port %d is not configured in Ethernet mode",
 377                               port);
 378                         goto port_error;
 379                 }
 380
 381                 if (port_attr.state != IBV_PORT_ACTIVE)
 382                         DEBUG("port %d is not active: \"%s\" (%d)",
 383                               port, ibv_port_state_str(port_attr.state),
 384                               port_attr.state);
 385
 386                 /* Allocate protection domain. */
 387                 pd = ibv_alloc_pd(ctx);
 388                 if (pd == NULL) {
 389                         ERROR("PD allocation failure");
 390                         err = ENOMEM;
 391                         goto port_error;
 392                 }
 393
 394                 mlx5_dev[idx].ports |= test;
 395
 396                 /* from rte_ethdev.c */
 397                 priv = rte_zmalloc("ethdev private structure",
 398                                    sizeof(*priv),
 399                                    RTE_CACHE_LINE_SIZE);
 400                 if (priv == NULL) {
 401                         ERROR("priv allocation failure");
 402                         err = ENOMEM;
 403                         goto port_error;
 404                 }
 405
 406                 priv->ctx = ctx;
 407                 priv->device_attr = device_attr;
 408                 priv->port = port;
 409                 priv->pd = pd;
 410                 priv->mtu = ETHER_MTU;
 411                 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
 412                         ERROR("ibv_exp_query_device() failed");
 413                         goto port_error;
 414                 }
 415
 416                 priv->hw_csum =
 417                         ((exp_device_attr.exp_device_cap_flags &
 418                           IBV_EXP_DEVICE_RX_CSUM_TCP_UDP_PKT) &&
 419                          (exp_device_attr.exp_device_cap_flags &
 420                           IBV_EXP_DEVICE_RX_CSUM_IP_PKT));
 421                 DEBUG("checksum offloading is %ssupported",
 422                       (priv->hw_csum ? "" : "not "));
 423
 424                 priv->hw_csum_l2tun = !!(exp_device_attr.exp_device_cap_flags &
 425                                          IBV_EXP_DEVICE_VXLAN_SUPPORT);
 426                 DEBUG("L2 tunnel checksum offloads are %ssupported",
 427                       (priv->hw_csum_l2tun ? "" : "not "));
 428
 429                 priv->ind_table_max_size = exp_device_attr.rx_hash_caps.max_rwq_indirection_table_size;
 430                 /* Remove this check once DPDK supports larger/variable
 431                  * indirection tables. */
 432                 if (priv->ind_table_max_size > (unsigned int)RSS_INDIRECTION_TABLE_SIZE)
 433                         priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
 434                 DEBUG("maximum RX indirection table size is %u",
 435                       priv->ind_table_max_size);
 436                 priv->hw_vlan_strip = !!(exp_device_attr.wq_vlan_offloads_cap &
 437                                          IBV_EXP_RECEIVE_WQ_CVLAN_STRIP);
 438                 DEBUG("VLAN stripping is %ssupported",
 439                       (priv->hw_vlan_strip ? "" : "not "));
 440
 441                 priv->hw_fcs_strip = !!(exp_device_attr.exp_device_cap_flags &
 442                                         IBV_EXP_DEVICE_SCATTER_FCS);
 443                 DEBUG("FCS stripping configuration is %ssupported",
 444                       (priv->hw_fcs_strip ? "" : "not "));
 445
 446                 priv->hw_padding = !!exp_device_attr.rx_pad_end_addr_align;
 447                 DEBUG("hardware RX end alignment padding is %ssupported",
 448                       (priv->hw_padding ? "" : "not "));
 449
 450                 priv_get_num_vfs(priv, &num_vfs);
 451                 priv->sriov = (num_vfs || sriov);
 452                 priv->mps = mps;
 453                 /* Allocate and register default RSS hash keys. */
 454                 priv->rss_conf = rte_calloc(__func__, hash_rxq_init_n,
 455                                             sizeof((*priv->rss_conf)[0]), 0);
 456                 if (priv->rss_conf == NULL) {
 457                         err = ENOMEM;
 458                         goto port_error;
 459                 }
 460                 err = rss_hash_rss_conf_new_key(priv,
 461                                                 rss_hash_default_key,
 462                                                 rss_hash_default_key_len,
 463                                                 ETH_RSS_PROTO_MASK);
 464                 if (err)
 465                         goto port_error;
 466                 /* Configure the first MAC address by default. */
 467                 if (priv_get_mac(priv, &mac.addr_bytes)) {
 468                         ERROR("cannot get MAC address, is mlx5_en loaded?"
 469                               " (errno: %s)", strerror(errno));
 470                         goto port_error;
 471                 }
 472                 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
 473                      priv->port,
 474                      mac.addr_bytes[0], mac.addr_bytes[1],
 475                      mac.addr_bytes[2], mac.addr_bytes[3],
 476                      mac.addr_bytes[4], mac.addr_bytes[5]);
 477                 /* Register MAC address. */
 478                 claim_zero(priv_mac_addr_add(priv, 0,
 479                                              (const uint8_t (*)[ETHER_ADDR_LEN])
 480                                              mac.addr_bytes));
 481                 /* Initialize FD filters list. */
 482                 err = fdir_init_filters_list(priv);
 483                 if (err)
 484                         goto port_error;
 485 #ifndef NDEBUG
 486                 {
 487                         char ifname[IF_NAMESIZE];
 488
 489                         if (priv_get_ifname(priv, &ifname) == 0)
 490                                 DEBUG("port %u ifname is \"%s\"",
 491                                       priv->port, ifname);
 492                         else
 493                                 DEBUG("port %u ifname is unknown", priv->port);
 494                 }
 495 #endif
 496                 /* Get actual MTU if possible. */
 497                 priv_get_mtu(priv, &priv->mtu);
 498                 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
 499
 500                 /* from rte_ethdev.c */
 501                 {
 502                         char name[RTE_ETH_NAME_MAX_LEN];
 503
 504                         snprintf(name, sizeof(name), "%s port %u",
 505                                  ibv_get_device_name(ibv_dev), port);
 506                         eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
 507                 }
 508                 if (eth_dev == NULL) {
 509                         ERROR("can not allocate rte ethdev");
 510                         err = ENOMEM;
 511                         goto port_error;
 512                 }
 513
 514                 /* Secondary processes have to use local storage for their
 515                  * private data as well as a copy of eth_dev->data, but this
 516                  * pointer must not be modified before burst functions are
 517                  * actually called. */
 518                 if (mlx5_is_secondary()) {
 519                         struct mlx5_secondary_data *sd =
 520                                 &mlx5_secondary_data[eth_dev->data->port_id];
 521                         sd->primary_priv = eth_dev->data->dev_private;
 522                         if (sd->primary_priv == NULL) {
 523                                 ERROR("no private data for port %u",
 524                                                 eth_dev->data->port_id);
 525                                 err = EINVAL;
 526                                 goto port_error;
 527                         }
 528                         sd->shared_dev_data = eth_dev->data;
 529                         rte_spinlock_init(&sd->lock);
 530                         memcpy(sd->data.name, sd->shared_dev_data->name,
 531                                    sizeof(sd->data.name));
 532                         sd->data.dev_private = priv;
 533                         sd->data.rx_mbuf_alloc_failed = 0;
 534                         sd->data.mtu = ETHER_MTU;
 535                         sd->data.port_id = sd->shared_dev_data->port_id;
 536                         sd->data.mac_addrs = priv->mac;
 537                         eth_dev->tx_pkt_burst = mlx5_tx_burst_secondary_setup;
 538                         eth_dev->rx_pkt_burst = mlx5_rx_burst_secondary_setup;
 539                 } else {
 540                         eth_dev->data->dev_private = priv;
 541                         eth_dev->data->rx_mbuf_alloc_failed = 0;
 542                         eth_dev->data->mtu = ETHER_MTU;
 543                         eth_dev->data->mac_addrs = priv->mac;
 544                 }
 545
 546                 eth_dev->pci_dev = pci_dev;
 547                 rte_eth_copy_pci_info(eth_dev, pci_dev);
 548                 eth_dev->driver = &mlx5_driver;
 549                 priv->dev = eth_dev;
 550                 eth_dev->dev_ops = &mlx5_dev_ops;
 551
 552                 TAILQ_INIT(&eth_dev->link_intr_cbs);
 553
 554                 /* Bring Ethernet device up. */
 555                 DEBUG("forcing Ethernet interface up");
 556                 priv_set_flags(priv, ~IFF_UP, IFF_UP);
 557                 continue;
 558
 559 port_error:
 560                 if (priv) {
 561                         rte_free(priv->rss_conf);
 562                         rte_free(priv);
 563                 }
 564                 if (pd)
 565                         claim_zero(ibv_dealloc_pd(pd));
 566                 if (ctx)
 567                         claim_zero(ibv_close_device(ctx));
 568                 break;
 569         }
 570
 571         /*
 572          * XXX if something went wrong in the loop above, there is a resource
 573          * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
 574          * long as the dpdk does not provide a way to deallocate a ethdev and a
 575          * way to enumerate the registered ethdevs to free the previous ones.
 576          */
 577
 578         /* no port found, complain */
 579         if (!mlx5_dev[idx].ports) {
 580                 err = ENODEV;
 581                 goto error;
 582         }
 583
 584 error:
 585         if (attr_ctx)
 586                 claim_zero(ibv_close_device(attr_ctx));
 587         if (list)
 588                 ibv_free_device_list(list);
 589         assert(err >= 0);
 590         return -err;
 591 }
 592
 593 static const struct rte_pci_id mlx5_pci_id_map[] = {
 594         {
 595                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 596                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
 597         },
 598         {
 599                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 600                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
 601         },
 602         {
 603                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 604                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
 605         },
 606         {
 607                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 608                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
 609         },
 610         {
 611                 .vendor_id = 0
 612         }
 613 };
 614
 615 static struct eth_driver mlx5_driver = {
 616         .pci_drv = {
 617                 .name = MLX5_DRIVER_NAME,
 618                 .id_table = mlx5_pci_id_map,
 619                 .devinit = mlx5_pci_devinit,
 620                 .drv_flags = RTE_PCI_DRV_INTR_LSC,
 621         },
 622         .dev_private_size = sizeof(struct priv)
 623 };
 624
 625 /**
 626  * Driver initialization routine.
 627  */
 628 static int
 629 rte_mlx5_pmd_init(const char *name, const char *args)
 630 {
 631         (void)name;
 632         (void)args;
 633         /*
 634          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
 635          * huge pages. Calling ibv_fork_init() during init allows
 636          * applications to use fork() safely for purposes other than
 637          * using this PMD, which is not supported in forked processes.
 638          */
 639         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
 640         ibv_fork_init();
 641         rte_eal_pci_register(&mlx5_driver.pci_drv);
 642         return 0;
 643 }
 644
 645 static struct rte_driver rte_mlx5_driver = {
 646         .type = PMD_PDEV,
 647         .name = MLX5_DRIVER_NAME,
 648         .init = rte_mlx5_pmd_init,
 649 };
 650
 651 PMD_REGISTER_DRIVER(rte_mlx5_driver)