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         void *tmp;
 102         unsigned int i;
 103
 104         priv_lock(priv);
 105         DEBUG("%p: closing device \"%s\"",
 106               (void *)dev,
 107               ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
 108         /* In case mlx5_dev_stop() has not been called. */
 109         priv_dev_interrupt_handler_uninstall(priv, dev);
 110         priv_special_flow_disable_all(priv);
 111         priv_mac_addrs_disable(priv);
 112         priv_destroy_hash_rxqs(priv);
 113
 114         /* Remove flow director elements. */
 115         priv_fdir_disable(priv);
 116         priv_fdir_delete_filters_list(priv);
 117
 118         /* Prevent crashes when queues are still in use. */
 119         dev->rx_pkt_burst = removed_rx_burst;
 120         dev->tx_pkt_burst = removed_tx_burst;
 121         if (priv->rxqs != NULL) {
 122                 /* XXX race condition if mlx5_rx_burst() is still running. */
 123                 usleep(1000);
 124                 for (i = 0; (i != priv->rxqs_n); ++i) {
 125                         tmp = (*priv->rxqs)[i];
 126                         if (tmp == NULL)
 127                                 continue;
 128                         (*priv->rxqs)[i] = NULL;
 129                         rxq_cleanup(tmp);
 130                         rte_free(tmp);
 131                 }
 132                 priv->rxqs_n = 0;
 133                 priv->rxqs = NULL;
 134         }
 135         if (priv->txqs != NULL) {
 136                 /* XXX race condition if mlx5_tx_burst() is still running. */
 137                 usleep(1000);
 138                 for (i = 0; (i != priv->txqs_n); ++i) {
 139                         tmp = (*priv->txqs)[i];
 140                         if (tmp == NULL)
 141                                 continue;
 142                         (*priv->txqs)[i] = NULL;
 143                         txq_cleanup(tmp);
 144                         rte_free(tmp);
 145                 }
 146                 priv->txqs_n = 0;
 147                 priv->txqs = NULL;
 148         }
 149         if (priv->pd != NULL) {
 150                 assert(priv->ctx != NULL);
 151                 claim_zero(ibv_dealloc_pd(priv->pd));
 152                 claim_zero(ibv_close_device(priv->ctx));
 153         } else
 154                 assert(priv->ctx == NULL);
 155         if (priv->rss_conf != NULL) {
 156                 for (i = 0; (i != hash_rxq_init_n); ++i)
 157                         rte_free((*priv->rss_conf)[i]);
 158                 rte_free(priv->rss_conf);
 159         }
 160         if (priv->reta_idx != NULL)
 161                 rte_free(priv->reta_idx);
 162         priv_unlock(priv);
 163         memset(priv, 0, sizeof(*priv));
 164 }
 165
 166 static const struct eth_dev_ops mlx5_dev_ops = {
 167         .dev_configure = mlx5_dev_configure,
 168         .dev_start = mlx5_dev_start,
 169         .dev_stop = mlx5_dev_stop,
 170         .dev_set_link_down = mlx5_set_link_down,
 171         .dev_set_link_up = mlx5_set_link_up,
 172         .dev_close = mlx5_dev_close,
 173         .promiscuous_enable = mlx5_promiscuous_enable,
 174         .promiscuous_disable = mlx5_promiscuous_disable,
 175         .allmulticast_enable = mlx5_allmulticast_enable,
 176         .allmulticast_disable = mlx5_allmulticast_disable,
 177         .link_update = mlx5_link_update,
 178         .stats_get = mlx5_stats_get,
 179         .stats_reset = mlx5_stats_reset,
 180         .dev_infos_get = mlx5_dev_infos_get,
 181         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 182         .vlan_filter_set = mlx5_vlan_filter_set,
 183         .rx_queue_setup = mlx5_rx_queue_setup,
 184         .tx_queue_setup = mlx5_tx_queue_setup,
 185         .rx_queue_release = mlx5_rx_queue_release,
 186         .tx_queue_release = mlx5_tx_queue_release,
 187         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 188         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 189         .mac_addr_remove = mlx5_mac_addr_remove,
 190         .mac_addr_add = mlx5_mac_addr_add,
 191         .mac_addr_set = mlx5_mac_addr_set,
 192         .mtu_set = mlx5_dev_set_mtu,
 193 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
 194         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 195         .vlan_offload_set = mlx5_vlan_offload_set,
 196 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
 197         .reta_update = mlx5_dev_rss_reta_update,
 198         .reta_query = mlx5_dev_rss_reta_query,
 199         .rss_hash_update = mlx5_rss_hash_update,
 200         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
 201 #ifdef MLX5_FDIR_SUPPORT
 202         .filter_ctrl = mlx5_dev_filter_ctrl,
 203 #endif /* MLX5_FDIR_SUPPORT */
 204 };
 205
 206 static struct {
 207         struct rte_pci_addr pci_addr; /* associated PCI address */
 208         uint32_t ports; /* physical ports bitfield. */
 209 } mlx5_dev[32];
 210
 211 /**
 212  * Get device index in mlx5_dev[] from PCI bus address.
 213  *
 214  * @param[in] pci_addr
 215  *   PCI bus address to look for.
 216  *
 217  * @return
 218  *   mlx5_dev[] index on success, -1 on failure.
 219  */
 220 static int
 221 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
 222 {
 223         unsigned int i;
 224         int ret = -1;
 225
 226         assert(pci_addr != NULL);
 227         for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
 228                 if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
 229                     (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
 230                     (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
 231                     (mlx5_dev[i].pci_addr.function == pci_addr->function))
 232                         return i;
 233                 if ((mlx5_dev[i].ports == 0) && (ret == -1))
 234                         ret = i;
 235         }
 236         return ret;
 237 }
 238
 239 static struct eth_driver mlx5_driver;
 240
 241 /**
 242  * DPDK callback to register a PCI device.
 243  *
 244  * This function creates an Ethernet device for each port of a given
 245  * PCI device.
 246  *
 247  * @param[in] pci_drv
 248  *   PCI driver structure (mlx5_driver).
 249  * @param[in] pci_dev
 250  *   PCI device information.
 251  *
 252  * @return
 253  *   0 on success, negative errno value on failure.
 254  */
 255 static int
 256 mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 257 {
 258         struct ibv_device **list;
 259         struct ibv_device *ibv_dev;
 260         int err = 0;
 261         struct ibv_context *attr_ctx = NULL;
 262         struct ibv_device_attr device_attr;
 263         unsigned int sriov;
 264         unsigned int mps;
 265         int idx;
 266         int i;
 267
 268         (void)pci_drv;
 269         assert(pci_drv == &mlx5_driver.pci_drv);
 270         /* Get mlx5_dev[] index. */
 271         idx = mlx5_dev_idx(&pci_dev->addr);
 272         if (idx == -1) {
 273                 ERROR("this driver cannot support any more adapters");
 274                 return -ENOMEM;
 275         }
 276         DEBUG("using driver device index %d", idx);
 277
 278         /* Save PCI address. */
 279         mlx5_dev[idx].pci_addr = pci_dev->addr;
 280         list = ibv_get_device_list(&i);
 281         if (list == NULL) {
 282                 assert(errno);
 283                 if (errno == ENOSYS) {
 284                         WARN("cannot list devices, is ib_uverbs loaded?");
 285                         return 0;
 286                 }
 287                 return -errno;
 288         }
 289         assert(i >= 0);
 290         /*
 291          * For each listed device, check related sysfs entry against
 292          * the provided PCI ID.
 293          */
 294         while (i != 0) {
 295                 struct rte_pci_addr pci_addr;
 296
 297                 --i;
 298                 DEBUG("checking device \"%s\"", list[i]->name);
 299                 if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
 300                         continue;
 301                 if ((pci_dev->addr.domain != pci_addr.domain) ||
 302                     (pci_dev->addr.bus != pci_addr.bus) ||
 303                     (pci_dev->addr.devid != pci_addr.devid) ||
 304                     (pci_dev->addr.function != pci_addr.function))
 305                         continue;
 306                 sriov = ((pci_dev->id.device_id ==
 307                        PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
 308                       (pci_dev->id.device_id ==
 309                        PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF));
 310                 /* Multi-packet send is only supported by ConnectX-4 Lx PF. */
 311                 mps = (pci_dev->id.device_id ==
 312                        PCI_DEVICE_ID_MELLANOX_CONNECTX4LX);
 313                 INFO("PCI information matches, using device \"%s\""
 314                      " (SR-IOV: %s, MPS: %s)",
 315                      list[i]->name,
 316                      sriov ? "true" : "false",
 317                      mps ? "true" : "false");
 318                 attr_ctx = ibv_open_device(list[i]);
 319                 err = errno;
 320                 break;
 321         }
 322         if (attr_ctx == NULL) {
 323                 ibv_free_device_list(list);
 324                 switch (err) {
 325                 case 0:
 326                         WARN("cannot access device, is mlx5_ib loaded?");
 327                         return 0;
 328                 case EINVAL:
 329                         WARN("cannot use device, are drivers up to date?");
 330                         return 0;
 331                 }
 332                 assert(err > 0);
 333                 return -err;
 334         }
 335         ibv_dev = list[i];
 336
 337         DEBUG("device opened");
 338         if (ibv_query_device(attr_ctx, &device_attr))
 339                 goto error;
 340         INFO("%u port(s) detected", device_attr.phys_port_cnt);
 341
 342         for (i = 0; i < device_attr.phys_port_cnt; i++) {
 343                 uint32_t port = i + 1; /* ports are indexed from one */
 344                 uint32_t test = (1 << i);
 345                 struct ibv_context *ctx = NULL;
 346                 struct ibv_port_attr port_attr;
 347                 struct ibv_pd *pd = NULL;
 348                 struct priv *priv = NULL;
 349                 struct rte_eth_dev *eth_dev;
 350 #ifdef HAVE_EXP_QUERY_DEVICE
 351                 struct ibv_exp_device_attr exp_device_attr;
 352 #endif /* HAVE_EXP_QUERY_DEVICE */
 353                 struct ether_addr mac;
 354                 uint16_t num_vfs = 0;
 355
 356 #ifdef HAVE_EXP_QUERY_DEVICE
 357                 exp_device_attr.comp_mask =
 358                         IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
 359                         IBV_EXP_DEVICE_ATTR_RX_HASH |
 360 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
 361                         IBV_EXP_DEVICE_ATTR_VLAN_OFFLOADS |
 362 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
 363 #ifdef HAVE_VERBS_RX_END_PADDING
 364                         IBV_EXP_DEVICE_ATTR_RX_PAD_END_ALIGN |
 365 #endif /* HAVE_VERBS_RX_END_PADDING */
 366                         0;
 367 #endif /* HAVE_EXP_QUERY_DEVICE */
 368
 369                 DEBUG("using port %u (%08" PRIx32 ")", port, test);
 370
 371                 ctx = ibv_open_device(ibv_dev);
 372                 if (ctx == NULL)
 373                         goto port_error;
 374
 375                 /* Check port status. */
 376                 err = ibv_query_port(ctx, port, &port_attr);
 377                 if (err) {
 378                         ERROR("port query failed: %s", strerror(err));
 379                         goto port_error;
 380                 }
 381
 382                 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 383                         ERROR("port %d is not configured in Ethernet mode",
 384                               port);
 385                         goto port_error;
 386                 }
 387
 388                 if (port_attr.state != IBV_PORT_ACTIVE)
 389                         DEBUG("port %d is not active: \"%s\" (%d)",
 390                               port, ibv_port_state_str(port_attr.state),
 391                               port_attr.state);
 392
 393                 /* Allocate protection domain. */
 394                 pd = ibv_alloc_pd(ctx);
 395                 if (pd == NULL) {
 396                         ERROR("PD allocation failure");
 397                         err = ENOMEM;
 398                         goto port_error;
 399                 }
 400
 401                 mlx5_dev[idx].ports |= test;
 402
 403                 /* from rte_ethdev.c */
 404                 priv = rte_zmalloc("ethdev private structure",
 405                                    sizeof(*priv),
 406                                    RTE_CACHE_LINE_SIZE);
 407                 if (priv == NULL) {
 408                         ERROR("priv allocation failure");
 409                         err = ENOMEM;
 410                         goto port_error;
 411                 }
 412
 413                 priv->ctx = ctx;
 414                 priv->device_attr = device_attr;
 415                 priv->port = port;
 416                 priv->pd = pd;
 417                 priv->mtu = ETHER_MTU;
 418 #ifdef HAVE_EXP_QUERY_DEVICE
 419                 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
 420                         ERROR("ibv_exp_query_device() failed");
 421                         goto port_error;
 422                 }
 423
 424                 priv->hw_csum =
 425                         ((exp_device_attr.exp_device_cap_flags &
 426                           IBV_EXP_DEVICE_RX_CSUM_TCP_UDP_PKT) &&
 427                          (exp_device_attr.exp_device_cap_flags &
 428                           IBV_EXP_DEVICE_RX_CSUM_IP_PKT));
 429                 DEBUG("checksum offloading is %ssupported",
 430                       (priv->hw_csum ? "" : "not "));
 431
 432                 priv->hw_csum_l2tun = !!(exp_device_attr.exp_device_cap_flags &
 433                                          IBV_EXP_DEVICE_VXLAN_SUPPORT);
 434                 DEBUG("L2 tunnel checksum offloads are %ssupported",
 435                       (priv->hw_csum_l2tun ? "" : "not "));
 436
 437                 priv->ind_table_max_size = exp_device_attr.rx_hash_caps.max_rwq_indirection_table_size;
 438                 /* Remove this check once DPDK supports larger/variable
 439                  * indirection tables. */
 440                 if (priv->ind_table_max_size > (unsigned int)RSS_INDIRECTION_TABLE_SIZE)
 441                         priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
 442                 DEBUG("maximum RX indirection table size is %u",
 443                       priv->ind_table_max_size);
 444 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
 445                 priv->hw_vlan_strip = !!(exp_device_attr.wq_vlan_offloads_cap &
 446                                          IBV_EXP_RECEIVE_WQ_CVLAN_STRIP);
 447 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
 448                 DEBUG("VLAN stripping is %ssupported",
 449                       (priv->hw_vlan_strip ? "" : "not "));
 450
 451 #ifdef HAVE_VERBS_FCS
 452                 priv->hw_fcs_strip = !!(exp_device_attr.exp_device_cap_flags &
 453                                         IBV_EXP_DEVICE_SCATTER_FCS);
 454 #endif /* HAVE_VERBS_FCS */
 455                 DEBUG("FCS stripping configuration is %ssupported",
 456                       (priv->hw_fcs_strip ? "" : "not "));
 457
 458 #ifdef HAVE_VERBS_RX_END_PADDING
 459                 priv->hw_padding = !!exp_device_attr.rx_pad_end_addr_align;
 460 #endif /* HAVE_VERBS_RX_END_PADDING */
 461                 DEBUG("hardware RX end alignment padding is %ssupported",
 462                       (priv->hw_padding ? "" : "not "));
 463
 464 #else /* HAVE_EXP_QUERY_DEVICE */
 465                 priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
 466 #endif /* HAVE_EXP_QUERY_DEVICE */
 467
 468                 priv_get_num_vfs(priv, &num_vfs);
 469                 priv->sriov = (num_vfs || sriov);
 470                 priv->mps = mps;
 471                 /* Allocate and register default RSS hash keys. */
 472                 priv->rss_conf = rte_calloc(__func__, hash_rxq_init_n,
 473                                             sizeof((*priv->rss_conf)[0]), 0);
 474                 if (priv->rss_conf == NULL) {
 475                         err = ENOMEM;
 476                         goto port_error;
 477                 }
 478                 err = rss_hash_rss_conf_new_key(priv,
 479                                                 rss_hash_default_key,
 480                                                 rss_hash_default_key_len,
 481                                                 ETH_RSS_PROTO_MASK);
 482                 if (err)
 483                         goto port_error;
 484                 /* Configure the first MAC address by default. */
 485                 if (priv_get_mac(priv, &mac.addr_bytes)) {
 486                         ERROR("cannot get MAC address, is mlx5_en loaded?"
 487                               " (errno: %s)", strerror(errno));
 488                         goto port_error;
 489                 }
 490                 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
 491                      priv->port,
 492                      mac.addr_bytes[0], mac.addr_bytes[1],
 493                      mac.addr_bytes[2], mac.addr_bytes[3],
 494                      mac.addr_bytes[4], mac.addr_bytes[5]);
 495                 /* Register MAC address. */
 496                 claim_zero(priv_mac_addr_add(priv, 0,
 497                                              (const uint8_t (*)[ETHER_ADDR_LEN])
 498                                              mac.addr_bytes));
 499                 /* Initialize FD filters list. */
 500                 err = fdir_init_filters_list(priv);
 501                 if (err)
 502                         goto port_error;
 503 #ifndef NDEBUG
 504                 {
 505                         char ifname[IF_NAMESIZE];
 506
 507                         if (priv_get_ifname(priv, &ifname) == 0)
 508                                 DEBUG("port %u ifname is \"%s\"",
 509                                       priv->port, ifname);
 510                         else
 511                                 DEBUG("port %u ifname is unknown", priv->port);
 512                 }
 513 #endif
 514                 /* Get actual MTU if possible. */
 515                 priv_get_mtu(priv, &priv->mtu);
 516                 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
 517
 518                 /* from rte_ethdev.c */
 519                 {
 520                         char name[RTE_ETH_NAME_MAX_LEN];
 521
 522                         snprintf(name, sizeof(name), "%s port %u",
 523                                  ibv_get_device_name(ibv_dev), port);
 524                         eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
 525                 }
 526                 if (eth_dev == NULL) {
 527                         ERROR("can not allocate rte ethdev");
 528                         err = ENOMEM;
 529                         goto port_error;
 530                 }
 531
 532                 /* Secondary processes have to use local storage for their
 533                  * private data as well as a copy of eth_dev->data, but this
 534                  * pointer must not be modified before burst functions are
 535                  * actually called. */
 536                 if (mlx5_is_secondary()) {
 537                         struct mlx5_secondary_data *sd =
 538                                 &mlx5_secondary_data[eth_dev->data->port_id];
 539                         sd->primary_priv = eth_dev->data->dev_private;
 540                         if (sd->primary_priv == NULL) {
 541                                 ERROR("no private data for port %u",
 542                                                 eth_dev->data->port_id);
 543                                 err = EINVAL;
 544                                 goto port_error;
 545                         }
 546                         sd->shared_dev_data = eth_dev->data;
 547                         rte_spinlock_init(&sd->lock);
 548                         memcpy(sd->data.name, sd->shared_dev_data->name,
 549                                    sizeof(sd->data.name));
 550                         sd->data.dev_private = priv;
 551                         sd->data.rx_mbuf_alloc_failed = 0;
 552                         sd->data.mtu = ETHER_MTU;
 553                         sd->data.port_id = sd->shared_dev_data->port_id;
 554                         sd->data.mac_addrs = priv->mac;
 555                         eth_dev->tx_pkt_burst = mlx5_tx_burst_secondary_setup;
 556                         eth_dev->rx_pkt_burst = mlx5_rx_burst_secondary_setup;
 557                 } else {
 558                         eth_dev->data->dev_private = priv;
 559                         eth_dev->data->rx_mbuf_alloc_failed = 0;
 560                         eth_dev->data->mtu = ETHER_MTU;
 561                         eth_dev->data->mac_addrs = priv->mac;
 562                 }
 563
 564                 eth_dev->pci_dev = pci_dev;
 565                 rte_eth_copy_pci_info(eth_dev, pci_dev);
 566                 eth_dev->driver = &mlx5_driver;
 567                 priv->dev = eth_dev;
 568                 eth_dev->dev_ops = &mlx5_dev_ops;
 569
 570                 TAILQ_INIT(&eth_dev->link_intr_cbs);
 571
 572                 /* Bring Ethernet device up. */
 573                 DEBUG("forcing Ethernet interface up");
 574                 priv_set_flags(priv, ~IFF_UP, IFF_UP);
 575                 continue;
 576
 577 port_error:
 578                 if (priv) {
 579                         rte_free(priv->rss_conf);
 580                         rte_free(priv);
 581                 }
 582                 if (pd)
 583                         claim_zero(ibv_dealloc_pd(pd));
 584                 if (ctx)
 585                         claim_zero(ibv_close_device(ctx));
 586                 break;
 587         }
 588
 589         /*
 590          * XXX if something went wrong in the loop above, there is a resource
 591          * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
 592          * long as the dpdk does not provide a way to deallocate a ethdev and a
 593          * way to enumerate the registered ethdevs to free the previous ones.
 594          */
 595
 596         /* no port found, complain */
 597         if (!mlx5_dev[idx].ports) {
 598                 err = ENODEV;
 599                 goto error;
 600         }
 601
 602 error:
 603         if (attr_ctx)
 604                 claim_zero(ibv_close_device(attr_ctx));
 605         if (list)
 606                 ibv_free_device_list(list);
 607         assert(err >= 0);
 608         return -err;
 609 }
 610
 611 static const struct rte_pci_id mlx5_pci_id_map[] = {
 612         {
 613                 .vendor_id = PCI_VENDOR_ID_MELLANOX,
 614                 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4,
 615                 .subsystem_vendor_id = PCI_ANY_ID,
 616                 .subsystem_device_id = PCI_ANY_ID
 617         },
 618         {
 619                 .vendor_id = PCI_VENDOR_ID_MELLANOX,
 620                 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4VF,
 621                 .subsystem_vendor_id = PCI_ANY_ID,
 622                 .subsystem_device_id = PCI_ANY_ID
 623         },
 624         {
 625                 .vendor_id = PCI_VENDOR_ID_MELLANOX,
 626                 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4LX,
 627                 .subsystem_vendor_id = PCI_ANY_ID,
 628                 .subsystem_device_id = PCI_ANY_ID
 629         },
 630         {
 631                 .vendor_id = PCI_VENDOR_ID_MELLANOX,
 632                 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF,
 633                 .subsystem_vendor_id = PCI_ANY_ID,
 634                 .subsystem_device_id = PCI_ANY_ID
 635         },
 636         {
 637                 .vendor_id = 0
 638         }
 639 };
 640
 641 static struct eth_driver mlx5_driver = {
 642         .pci_drv = {
 643                 .name = MLX5_DRIVER_NAME,
 644                 .id_table = mlx5_pci_id_map,
 645                 .devinit = mlx5_pci_devinit,
 646                 .drv_flags = RTE_PCI_DRV_INTR_LSC,
 647         },
 648         .dev_private_size = sizeof(struct priv)
 649 };
 650
 651 /**
 652  * Driver initialization routine.
 653  */
 654 static int
 655 rte_mlx5_pmd_init(const char *name, const char *args)
 656 {
 657         (void)name;
 658         (void)args;
 659         /*
 660          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
 661          * huge pages. Calling ibv_fork_init() during init allows
 662          * applications to use fork() safely for purposes other than
 663          * using this PMD, which is not supported in forked processes.
 664          */
 665         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
 666         ibv_fork_init();
 667         rte_eal_pci_register(&mlx5_driver.pci_drv);
 668         return 0;
 669 }
 670
 671 static struct rte_driver rte_mlx5_driver = {
 672         .type = PMD_PDEV,
 673         .name = MLX5_DRIVER_NAME,
 674         .init = rte_mlx5_pmd_init,
 675 };
 676
 677 PMD_REGISTER_DRIVER(rte_mlx5_driver)