drivers/net/mlx4/mlx4.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2012 6WIND S.A.
   3  * Copyright 2012 Mellanox Technologies, Ltd
   4  */
   5
   6 /**
   7  * @file
   8  * mlx4 driver initialization.
   9  */
  10
  11 #include <assert.h>
  12 #include <dlfcn.h>
  13 #include <errno.h>
  14 #include <inttypes.h>
  15 #include <stddef.h>
  16 #include <stdint.h>
  17 #include <stdio.h>
  18 #include <stdlib.h>
  19 #include <string.h>
  20 #include <unistd.h>
  21
  22 /* Verbs headers do not support -pedantic. */
  23 #ifdef PEDANTIC
  24 #pragma GCC diagnostic ignored "-Wpedantic"
  25 #endif
  26 #include <infiniband/verbs.h>
  27 #ifdef PEDANTIC
  28 #pragma GCC diagnostic error "-Wpedantic"
  29 #endif
  30
  31 #include <rte_common.h>
  32 #include <rte_config.h>
  33 #include <rte_dev.h>
  34 #include <rte_errno.h>
  35 #include <rte_ethdev_driver.h>
  36 #include <rte_ethdev_pci.h>
  37 #include <rte_ether.h>
  38 #include <rte_flow.h>
  39 #include <rte_interrupts.h>
  40 #include <rte_kvargs.h>
  41 #include <rte_malloc.h>
  42 #include <rte_mbuf.h>
  43
  44 #include "mlx4.h"
  45 #include "mlx4_glue.h"
  46 #include "mlx4_flow.h"
  47 #include "mlx4_rxtx.h"
  48 #include "mlx4_utils.h"
  49
  50 /** Configuration structure for device arguments. */
  51 struct mlx4_conf {
  52         struct {
  53                 uint32_t present; /**< Bit-field for existing ports. */
  54                 uint32_t enabled; /**< Bit-field for user-enabled ports. */
  55         } ports;
  56 };
  57
  58 /* Available parameters list. */
  59 const char *pmd_mlx4_init_params[] = {
  60         MLX4_PMD_PORT_KVARG,
  61         NULL,
  62 };
  63
  64 static void mlx4_dev_stop(struct rte_eth_dev *dev);
  65
  66 /**
  67  * DPDK callback for Ethernet device configuration.
  68  *
  69  * @param dev
  70  *   Pointer to Ethernet device structure.
  71  *
  72  * @return
  73  *   0 on success, negative errno value otherwise and rte_errno is set.
  74  */
  75 static int
  76 mlx4_dev_configure(struct rte_eth_dev *dev)
  77 {
  78         struct priv *priv = dev->data->dev_private;
  79         struct rte_flow_error error;
  80         int ret;
  81
  82         /* Prepare internal flow rules. */
  83         ret = mlx4_flow_sync(priv, &error);
  84         if (ret) {
  85                 ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
  86                       " flow error type %d, cause %p, message: %s",
  87                       -ret, strerror(-ret), error.type, error.cause,
  88                       error.message ? error.message : "(unspecified)");
  89                 goto exit;
  90         }
  91         ret = mlx4_intr_install(priv);
  92         if (ret)
  93                 ERROR("%p: interrupt handler installation failed",
  94                       (void *)dev);
  95 exit:
  96         return ret;
  97 }
  98
  99 /**
 100  * DPDK callback to start the device.
 101  *
 102  * Simulate device start by initializing common RSS resources and attaching
 103  * all configured flows.
 104  *
 105  * @param dev
 106  *   Pointer to Ethernet device structure.
 107  *
 108  * @return
 109  *   0 on success, negative errno value otherwise and rte_errno is set.
 110  */
 111 static int
 112 mlx4_dev_start(struct rte_eth_dev *dev)
 113 {
 114         struct priv *priv = dev->data->dev_private;
 115         struct rte_flow_error error;
 116         int ret;
 117
 118         if (priv->started)
 119                 return 0;
 120         DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
 121         priv->started = 1;
 122         ret = mlx4_rss_init(priv);
 123         if (ret) {
 124                 ERROR("%p: cannot initialize RSS resources: %s",
 125                       (void *)dev, strerror(-ret));
 126                 goto err;
 127         }
 128         ret = mlx4_rxq_intr_enable(priv);
 129         if (ret) {
 130                 ERROR("%p: interrupt handler installation failed",
 131                      (void *)dev);
 132                 goto err;
 133         }
 134         ret = mlx4_flow_sync(priv, &error);
 135         if (ret) {
 136                 ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
 137                       " flow error type %d, cause %p, message: %s",
 138                       (void *)dev,
 139                       -ret, strerror(-ret), error.type, error.cause,
 140                       error.message ? error.message : "(unspecified)");
 141                 goto err;
 142         }
 143         rte_wmb();
 144         dev->tx_pkt_burst = mlx4_tx_burst;
 145         dev->rx_pkt_burst = mlx4_rx_burst;
 146         return 0;
 147 err:
 148         mlx4_dev_stop(dev);
 149         return ret;
 150 }
 151
 152 /**
 153  * DPDK callback to stop the device.
 154  *
 155  * Simulate device stop by detaching all configured flows.
 156  *
 157  * @param dev
 158  *   Pointer to Ethernet device structure.
 159  */
 160 static void
 161 mlx4_dev_stop(struct rte_eth_dev *dev)
 162 {
 163         struct priv *priv = dev->data->dev_private;
 164
 165         if (!priv->started)
 166                 return;
 167         DEBUG("%p: detaching flows from all RX queues", (void *)dev);
 168         priv->started = 0;
 169         dev->tx_pkt_burst = mlx4_tx_burst_removed;
 170         dev->rx_pkt_burst = mlx4_rx_burst_removed;
 171         rte_wmb();
 172         mlx4_flow_sync(priv, NULL);
 173         mlx4_rxq_intr_disable(priv);
 174         mlx4_rss_deinit(priv);
 175 }
 176
 177 /**
 178  * DPDK callback to close the device.
 179  *
 180  * Destroy all queues and objects, free memory.
 181  *
 182  * @param dev
 183  *   Pointer to Ethernet device structure.
 184  */
 185 static void
 186 mlx4_dev_close(struct rte_eth_dev *dev)
 187 {
 188         struct priv *priv = dev->data->dev_private;
 189         unsigned int i;
 190
 191         DEBUG("%p: closing device \"%s\"",
 192               (void *)dev,
 193               ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
 194         dev->rx_pkt_burst = mlx4_rx_burst_removed;
 195         dev->tx_pkt_burst = mlx4_tx_burst_removed;
 196         rte_wmb();
 197         mlx4_flow_clean(priv);
 198         mlx4_rss_deinit(priv);
 199         for (i = 0; i != dev->data->nb_rx_queues; ++i)
 200                 mlx4_rx_queue_release(dev->data->rx_queues[i]);
 201         for (i = 0; i != dev->data->nb_tx_queues; ++i)
 202                 mlx4_tx_queue_release(dev->data->tx_queues[i]);
 203         if (priv->pd != NULL) {
 204                 assert(priv->ctx != NULL);
 205                 claim_zero(mlx4_glue->dealloc_pd(priv->pd));
 206                 claim_zero(mlx4_glue->close_device(priv->ctx));
 207         } else
 208                 assert(priv->ctx == NULL);
 209         mlx4_intr_uninstall(priv);
 210         memset(priv, 0, sizeof(*priv));
 211 }
 212
 213 static const struct eth_dev_ops mlx4_dev_ops = {
 214         .dev_configure = mlx4_dev_configure,
 215         .dev_start = mlx4_dev_start,
 216         .dev_stop = mlx4_dev_stop,
 217         .dev_set_link_down = mlx4_dev_set_link_down,
 218         .dev_set_link_up = mlx4_dev_set_link_up,
 219         .dev_close = mlx4_dev_close,
 220         .link_update = mlx4_link_update,
 221         .promiscuous_enable = mlx4_promiscuous_enable,
 222         .promiscuous_disable = mlx4_promiscuous_disable,
 223         .allmulticast_enable = mlx4_allmulticast_enable,
 224         .allmulticast_disable = mlx4_allmulticast_disable,
 225         .mac_addr_remove = mlx4_mac_addr_remove,
 226         .mac_addr_add = mlx4_mac_addr_add,
 227         .mac_addr_set = mlx4_mac_addr_set,
 228         .stats_get = mlx4_stats_get,
 229         .stats_reset = mlx4_stats_reset,
 230         .dev_infos_get = mlx4_dev_infos_get,
 231         .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
 232         .vlan_filter_set = mlx4_vlan_filter_set,
 233         .rx_queue_setup = mlx4_rx_queue_setup,
 234         .tx_queue_setup = mlx4_tx_queue_setup,
 235         .rx_queue_release = mlx4_rx_queue_release,
 236         .tx_queue_release = mlx4_tx_queue_release,
 237         .flow_ctrl_get = mlx4_flow_ctrl_get,
 238         .flow_ctrl_set = mlx4_flow_ctrl_set,
 239         .mtu_set = mlx4_mtu_set,
 240         .filter_ctrl = mlx4_filter_ctrl,
 241         .rx_queue_intr_enable = mlx4_rx_intr_enable,
 242         .rx_queue_intr_disable = mlx4_rx_intr_disable,
 243         .is_removed = mlx4_is_removed,
 244 };
 245
 246 /**
 247  * Get PCI information from struct ibv_device.
 248  *
 249  * @param device
 250  *   Pointer to Ethernet device structure.
 251  * @param[out] pci_addr
 252  *   PCI bus address output buffer.
 253  *
 254  * @return
 255  *   0 on success, negative errno value otherwise and rte_errno is set.
 256  */
 257 static int
 258 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
 259                             struct rte_pci_addr *pci_addr)
 260 {
 261         FILE *file;
 262         char line[32];
 263         MKSTR(path, "%s/device/uevent", device->ibdev_path);
 264
 265         file = fopen(path, "rb");
 266         if (file == NULL) {
 267                 rte_errno = errno;
 268                 return -rte_errno;
 269         }
 270         while (fgets(line, sizeof(line), file) == line) {
 271                 size_t len = strlen(line);
 272                 int ret;
 273
 274                 /* Truncate long lines. */
 275                 if (len == (sizeof(line) - 1))
 276                         while (line[(len - 1)] != '\n') {
 277                                 ret = fgetc(file);
 278                                 if (ret == EOF)
 279                                         break;
 280                                 line[(len - 1)] = ret;
 281                         }
 282                 /* Extract information. */
 283                 if (sscanf(line,
 284                            "PCI_SLOT_NAME="
 285                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
 286                            &pci_addr->domain,
 287                            &pci_addr->bus,
 288                            &pci_addr->devid,
 289                            &pci_addr->function) == 4) {
 290                         ret = 0;
 291                         break;
 292                 }
 293         }
 294         fclose(file);
 295         return 0;
 296 }
 297
 298 /**
 299  * Verify and store value for device argument.
 300  *
 301  * @param[in] key
 302  *   Key argument to verify.
 303  * @param[in] val
 304  *   Value associated with key.
 305  * @param[in, out] conf
 306  *   Shared configuration data.
 307  *
 308  * @return
 309  *   0 on success, negative errno value otherwise and rte_errno is set.
 310  */
 311 static int
 312 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
 313 {
 314         unsigned long tmp;
 315
 316         errno = 0;
 317         tmp = strtoul(val, NULL, 0);
 318         if (errno) {
 319                 rte_errno = errno;
 320                 WARN("%s: \"%s\" is not a valid integer", key, val);
 321                 return -rte_errno;
 322         }
 323         if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
 324                 uint32_t ports = rte_log2_u32(conf->ports.present + 1);
 325
 326                 if (tmp >= ports) {
 327                         ERROR("port index %lu outside range [0,%" PRIu32 ")",
 328                               tmp, ports);
 329                         return -EINVAL;
 330                 }
 331                 if (!(conf->ports.present & (1 << tmp))) {
 332                         rte_errno = EINVAL;
 333                         ERROR("invalid port index %lu", tmp);
 334                         return -rte_errno;
 335                 }
 336                 conf->ports.enabled |= 1 << tmp;
 337         } else {
 338                 rte_errno = EINVAL;
 339                 WARN("%s: unknown parameter", key);
 340                 return -rte_errno;
 341         }
 342         return 0;
 343 }
 344
 345 /**
 346  * Parse device parameters.
 347  *
 348  * @param devargs
 349  *   Device arguments structure.
 350  *
 351  * @return
 352  *   0 on success, negative errno value otherwise and rte_errno is set.
 353  */
 354 static int
 355 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
 356 {
 357         struct rte_kvargs *kvlist;
 358         unsigned int arg_count;
 359         int ret = 0;
 360         int i;
 361
 362         if (devargs == NULL)
 363                 return 0;
 364         kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
 365         if (kvlist == NULL) {
 366                 rte_errno = EINVAL;
 367                 ERROR("failed to parse kvargs");
 368                 return -rte_errno;
 369         }
 370         /* Process parameters. */
 371         for (i = 0; pmd_mlx4_init_params[i]; ++i) {
 372                 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
 373                 while (arg_count-- > 0) {
 374                         ret = rte_kvargs_process(kvlist,
 375                                                  MLX4_PMD_PORT_KVARG,
 376                                                  (int (*)(const char *,
 377                                                           const char *,
 378                                                           void *))
 379                                                  mlx4_arg_parse,
 380                                                  conf);
 381                         if (ret != 0)
 382                                 goto free_kvlist;
 383                 }
 384         }
 385 free_kvlist:
 386         rte_kvargs_free(kvlist);
 387         return ret;
 388 }
 389
 390 /**
 391  * Interpret RSS capabilities reported by device.
 392  *
 393  * This function returns the set of usable Verbs RSS hash fields, kernel
 394  * quirks taken into account.
 395  *
 396  * @param ctx
 397  *   Verbs context.
 398  * @param pd
 399  *   Verbs protection domain.
 400  * @param device_attr_ex
 401  *   Extended device attributes to interpret.
 402  *
 403  * @return
 404  *   Usable RSS hash fields mask in Verbs format.
 405  */
 406 static uint64_t
 407 mlx4_hw_rss_sup(struct ibv_context *ctx, struct ibv_pd *pd,
 408                 struct ibv_device_attr_ex *device_attr_ex)
 409 {
 410         uint64_t hw_rss_sup = device_attr_ex->rss_caps.rx_hash_fields_mask;
 411         struct ibv_cq *cq = NULL;
 412         struct ibv_wq *wq = NULL;
 413         struct ibv_rwq_ind_table *ind = NULL;
 414         struct ibv_qp *qp = NULL;
 415
 416         if (!hw_rss_sup) {
 417                 WARN("no RSS capabilities reported; disabling support for UDP"
 418                      " RSS and inner VXLAN RSS");
 419                 return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 |
 420                         IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 |
 421                         IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP;
 422         }
 423         if (!(hw_rss_sup & IBV_RX_HASH_INNER))
 424                 return hw_rss_sup;
 425         /*
 426          * Although reported as supported, missing code in some Linux
 427          * versions (v4.15, v4.16) prevents the creation of hash QPs with
 428          * inner capability.
 429          *
 430          * There is no choice but to attempt to instantiate a temporary RSS
 431          * context in order to confirm its support.
 432          */
 433         cq = mlx4_glue->create_cq(ctx, 1, NULL, NULL, 0);
 434         wq = cq ? mlx4_glue->create_wq
 435                 (ctx,
 436                  &(struct ibv_wq_init_attr){
 437                         .wq_type = IBV_WQT_RQ,
 438                         .max_wr = 1,
 439                         .max_sge = 1,
 440                         .pd = pd,
 441                         .cq = cq,
 442                  }) : NULL;
 443         ind = wq ? mlx4_glue->create_rwq_ind_table
 444                 (ctx,
 445                  &(struct ibv_rwq_ind_table_init_attr){
 446                         .log_ind_tbl_size = 0,
 447                         .ind_tbl = &wq,
 448                         .comp_mask = 0,
 449                  }) : NULL;
 450         qp = ind ? mlx4_glue->create_qp_ex
 451                 (ctx,
 452                  &(struct ibv_qp_init_attr_ex){
 453                         .comp_mask =
 454                                 (IBV_QP_INIT_ATTR_PD |
 455                                  IBV_QP_INIT_ATTR_RX_HASH |
 456                                  IBV_QP_INIT_ATTR_IND_TABLE),
 457                         .qp_type = IBV_QPT_RAW_PACKET,
 458                         .pd = pd,
 459                         .rwq_ind_tbl = ind,
 460                         .rx_hash_conf = {
 461                                 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
 462                                 .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
 463                                 .rx_hash_key = mlx4_rss_hash_key_default,
 464                                 .rx_hash_fields_mask = hw_rss_sup,
 465                         },
 466                  }) : NULL;
 467         if (!qp) {
 468                 WARN("disabling unusable inner RSS capability due to kernel"
 469                      " quirk");
 470                 hw_rss_sup &= ~IBV_RX_HASH_INNER;
 471         } else {
 472                 claim_zero(mlx4_glue->destroy_qp(qp));
 473         }
 474         if (ind)
 475                 claim_zero(mlx4_glue->destroy_rwq_ind_table(ind));
 476         if (wq)
 477                 claim_zero(mlx4_glue->destroy_wq(wq));
 478         if (cq)
 479                 claim_zero(mlx4_glue->destroy_cq(cq));
 480         return hw_rss_sup;
 481 }
 482
 483 static struct rte_pci_driver mlx4_driver;
 484
 485 /**
 486  * DPDK callback to register a PCI device.
 487  *
 488  * This function creates an Ethernet device for each port of a given
 489  * PCI device.
 490  *
 491  * @param[in] pci_drv
 492  *   PCI driver structure (mlx4_driver).
 493  * @param[in] pci_dev
 494  *   PCI device information.
 495  *
 496  * @return
 497  *   0 on success, negative errno value otherwise and rte_errno is set.
 498  */
 499 static int
 500 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 501 {
 502         struct ibv_device **list;
 503         struct ibv_device *ibv_dev;
 504         int err = 0;
 505         struct ibv_context *attr_ctx = NULL;
 506         struct ibv_device_attr device_attr;
 507         struct ibv_device_attr_ex device_attr_ex;
 508         struct mlx4_conf conf = {
 509                 .ports.present = 0,
 510         };
 511         unsigned int vf;
 512         int i;
 513
 514         (void)pci_drv;
 515         assert(pci_drv == &mlx4_driver);
 516         list = mlx4_glue->get_device_list(&i);
 517         if (list == NULL) {
 518                 rte_errno = errno;
 519                 assert(rte_errno);
 520                 if (rte_errno == ENOSYS)
 521                         ERROR("cannot list devices, is ib_uverbs loaded?");
 522                 return -rte_errno;
 523         }
 524         assert(i >= 0);
 525         /*
 526          * For each listed device, check related sysfs entry against
 527          * the provided PCI ID.
 528          */
 529         while (i != 0) {
 530                 struct rte_pci_addr pci_addr;
 531
 532                 --i;
 533                 DEBUG("checking device \"%s\"", list[i]->name);
 534                 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
 535                         continue;
 536                 if ((pci_dev->addr.domain != pci_addr.domain) ||
 537                     (pci_dev->addr.bus != pci_addr.bus) ||
 538                     (pci_dev->addr.devid != pci_addr.devid) ||
 539                     (pci_dev->addr.function != pci_addr.function))
 540                         continue;
 541                 vf = (pci_dev->id.device_id ==
 542                       PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
 543                 INFO("PCI information matches, using device \"%s\" (VF: %s)",
 544                      list[i]->name, (vf ? "true" : "false"));
 545                 attr_ctx = mlx4_glue->open_device(list[i]);
 546                 err = errno;
 547                 break;
 548         }
 549         if (attr_ctx == NULL) {
 550                 mlx4_glue->free_device_list(list);
 551                 switch (err) {
 552                 case 0:
 553                         rte_errno = ENODEV;
 554                         ERROR("cannot access device, is mlx4_ib loaded?");
 555                         return -rte_errno;
 556                 case EINVAL:
 557                         rte_errno = EINVAL;
 558                         ERROR("cannot use device, are drivers up to date?");
 559                         return -rte_errno;
 560                 }
 561                 assert(err > 0);
 562                 rte_errno = err;
 563                 return -rte_errno;
 564         }
 565         ibv_dev = list[i];
 566         DEBUG("device opened");
 567         if (mlx4_glue->query_device(attr_ctx, &device_attr)) {
 568                 rte_errno = ENODEV;
 569                 goto error;
 570         }
 571         INFO("%u port(s) detected", device_attr.phys_port_cnt);
 572         conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
 573         if (mlx4_args(pci_dev->device.devargs, &conf)) {
 574                 ERROR("failed to process device arguments");
 575                 rte_errno = EINVAL;
 576                 goto error;
 577         }
 578         /* Use all ports when none are defined */
 579         if (!conf.ports.enabled)
 580                 conf.ports.enabled = conf.ports.present;
 581         /* Retrieve extended device attributes. */
 582         if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
 583                 rte_errno = ENODEV;
 584                 goto error;
 585         }
 586         assert(device_attr.max_sge >= MLX4_MAX_SGE);
 587         for (i = 0; i < device_attr.phys_port_cnt; i++) {
 588                 uint32_t port = i + 1; /* ports are indexed from one */
 589                 struct ibv_context *ctx = NULL;
 590                 struct ibv_port_attr port_attr;
 591                 struct ibv_pd *pd = NULL;
 592                 struct priv *priv = NULL;
 593                 struct rte_eth_dev *eth_dev = NULL;
 594                 struct ether_addr mac;
 595
 596                 /* If port is not enabled, skip. */
 597                 if (!(conf.ports.enabled & (1 << i)))
 598                         continue;
 599                 DEBUG("using port %u", port);
 600                 ctx = mlx4_glue->open_device(ibv_dev);
 601                 if (ctx == NULL) {
 602                         rte_errno = ENODEV;
 603                         goto port_error;
 604                 }
 605                 /* Check port status. */
 606                 err = mlx4_glue->query_port(ctx, port, &port_attr);
 607                 if (err) {
 608                         rte_errno = err;
 609                         ERROR("port query failed: %s", strerror(rte_errno));
 610                         goto port_error;
 611                 }
 612                 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 613                         rte_errno = ENOTSUP;
 614                         ERROR("port %d is not configured in Ethernet mode",
 615                               port);
 616                         goto port_error;
 617                 }
 618                 if (port_attr.state != IBV_PORT_ACTIVE)
 619                         DEBUG("port %d is not active: \"%s\" (%d)",
 620                               port, mlx4_glue->port_state_str(port_attr.state),
 621                               port_attr.state);
 622                 /* Make asynchronous FD non-blocking to handle interrupts. */
 623                 if (mlx4_fd_set_non_blocking(ctx->async_fd) < 0) {
 624                         ERROR("cannot make asynchronous FD non-blocking: %s",
 625                               strerror(rte_errno));
 626                         goto port_error;
 627                 }
 628                 /* Allocate protection domain. */
 629                 pd = mlx4_glue->alloc_pd(ctx);
 630                 if (pd == NULL) {
 631                         rte_errno = ENOMEM;
 632                         ERROR("PD allocation failure");
 633                         goto port_error;
 634                 }
 635                 /* from rte_ethdev.c */
 636                 priv = rte_zmalloc("ethdev private structure",
 637                                    sizeof(*priv),
 638                                    RTE_CACHE_LINE_SIZE);
 639                 if (priv == NULL) {
 640                         rte_errno = ENOMEM;
 641                         ERROR("priv allocation failure");
 642                         goto port_error;
 643                 }
 644                 priv->ctx = ctx;
 645                 priv->device_attr = device_attr;
 646                 priv->port = port;
 647                 priv->pd = pd;
 648                 priv->mtu = ETHER_MTU;
 649                 priv->vf = vf;
 650                 priv->hw_csum = !!(device_attr.device_cap_flags &
 651                                    IBV_DEVICE_RAW_IP_CSUM);
 652                 DEBUG("checksum offloading is %ssupported",
 653                       (priv->hw_csum ? "" : "not "));
 654                 /* Only ConnectX-3 Pro supports tunneling. */
 655                 priv->hw_csum_l2tun =
 656                         priv->hw_csum &&
 657                         (device_attr.vendor_part_id ==
 658                          PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
 659                 DEBUG("L2 tunnel checksum offloads are %ssupported",
 660                       priv->hw_csum_l2tun ? "" : "not ");
 661                 priv->hw_rss_sup = mlx4_hw_rss_sup(priv->ctx, priv->pd,
 662                                                    &device_attr_ex);
 663                 DEBUG("supported RSS hash fields mask: %016" PRIx64,
 664                       priv->hw_rss_sup);
 665                 priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
 666                                         IBV_RAW_PACKET_CAP_SCATTER_FCS);
 667                 DEBUG("FCS stripping toggling is %ssupported",
 668                       priv->hw_fcs_strip ? "" : "not ");
 669                 /* Configure the first MAC address by default. */
 670                 if (mlx4_get_mac(priv, &mac.addr_bytes)) {
 671                         ERROR("cannot get MAC address, is mlx4_en loaded?"
 672                               " (rte_errno: %s)", strerror(rte_errno));
 673                         goto port_error;
 674                 }
 675                 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
 676                      priv->port,
 677                      mac.addr_bytes[0], mac.addr_bytes[1],
 678                      mac.addr_bytes[2], mac.addr_bytes[3],
 679                      mac.addr_bytes[4], mac.addr_bytes[5]);
 680                 /* Register MAC address. */
 681                 priv->mac[0] = mac;
 682 #ifndef NDEBUG
 683                 {
 684                         char ifname[IF_NAMESIZE];
 685
 686                         if (mlx4_get_ifname(priv, &ifname) == 0)
 687                                 DEBUG("port %u ifname is \"%s\"",
 688                                       priv->port, ifname);
 689                         else
 690                                 DEBUG("port %u ifname is unknown", priv->port);
 691                 }
 692 #endif
 693                 /* Get actual MTU if possible. */
 694                 mlx4_mtu_get(priv, &priv->mtu);
 695                 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
 696                 /* from rte_ethdev.c */
 697                 {
 698                         char name[RTE_ETH_NAME_MAX_LEN];
 699
 700                         snprintf(name, sizeof(name), "%s port %u",
 701                                  mlx4_glue->get_device_name(ibv_dev), port);
 702                         eth_dev = rte_eth_dev_allocate(name);
 703                 }
 704                 if (eth_dev == NULL) {
 705                         ERROR("can not allocate rte ethdev");
 706                         rte_errno = ENOMEM;
 707                         goto port_error;
 708                 }
 709                 eth_dev->data->dev_private = priv;
 710                 eth_dev->data->mac_addrs = priv->mac;
 711                 eth_dev->device = &pci_dev->device;
 712                 rte_eth_copy_pci_info(eth_dev, pci_dev);
 713                 eth_dev->device->driver = &mlx4_driver.driver;
 714                 /* Initialize local interrupt handle for current port. */
 715                 priv->intr_handle = (struct rte_intr_handle){
 716                         .fd = -1,
 717                         .type = RTE_INTR_HANDLE_EXT,
 718                 };
 719                 /*
 720                  * Override ethdev interrupt handle pointer with private
 721                  * handle instead of that of the parent PCI device used by
 722                  * default. This prevents it from being shared between all
 723                  * ports of the same PCI device since each of them is
 724                  * associated its own Verbs context.
 725                  *
 726                  * Rx interrupts in particular require this as the PMD has
 727                  * no control over the registration of queue interrupts
 728                  * besides setting up eth_dev->intr_handle, the rest is
 729                  * handled by rte_intr_rx_ctl().
 730                  */
 731                 eth_dev->intr_handle = &priv->intr_handle;
 732                 priv->dev = eth_dev;
 733                 eth_dev->dev_ops = &mlx4_dev_ops;
 734                 /* Bring Ethernet device up. */
 735                 DEBUG("forcing Ethernet interface up");
 736                 mlx4_dev_set_link_up(priv->dev);
 737                 /* Update link status once if waiting for LSC. */
 738                 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
 739                         mlx4_link_update(eth_dev, 0);
 740                 continue;
 741 port_error:
 742                 rte_free(priv);
 743                 if (pd)
 744                         claim_zero(mlx4_glue->dealloc_pd(pd));
 745                 if (ctx)
 746                         claim_zero(mlx4_glue->close_device(ctx));
 747                 if (eth_dev)
 748                         rte_eth_dev_release_port(eth_dev);
 749                 break;
 750         }
 751         if (i == device_attr.phys_port_cnt)
 752                 return 0;
 753         /*
 754          * XXX if something went wrong in the loop above, there is a resource
 755          * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
 756          * long as the dpdk does not provide a way to deallocate a ethdev and a
 757          * way to enumerate the registered ethdevs to free the previous ones.
 758          */
 759 error:
 760         if (attr_ctx)
 761                 claim_zero(mlx4_glue->close_device(attr_ctx));
 762         if (list)
 763                 mlx4_glue->free_device_list(list);
 764         assert(rte_errno >= 0);
 765         return -rte_errno;
 766 }
 767
 768 static const struct rte_pci_id mlx4_pci_id_map[] = {
 769         {
 770                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 771                                PCI_DEVICE_ID_MELLANOX_CONNECTX3)
 772         },
 773         {
 774                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 775                                PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
 776         },
 777         {
 778                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 779                                PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
 780         },
 781         {
 782                 .vendor_id = 0
 783         }
 784 };
 785
 786 static struct rte_pci_driver mlx4_driver = {
 787         .driver = {
 788                 .name = MLX4_DRIVER_NAME
 789         },
 790         .id_table = mlx4_pci_id_map,
 791         .probe = mlx4_pci_probe,
 792         .drv_flags = RTE_PCI_DRV_INTR_LSC |
 793                      RTE_PCI_DRV_INTR_RMV,
 794 };
 795
 796 #ifdef RTE_LIBRTE_MLX4_DLOPEN_DEPS
 797
 798 /**
 799  * Suffix RTE_EAL_PMD_PATH with "-glue".
 800  *
 801  * This function performs a sanity check on RTE_EAL_PMD_PATH before
 802  * suffixing its last component.
 803  *
 804  * @param buf[out]
 805  *   Output buffer, should be large enough otherwise NULL is returned.
 806  * @param size
 807  *   Size of @p out.
 808  *
 809  * @return
 810  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
 811  */
 812 static char *
 813 mlx4_glue_path(char *buf, size_t size)
 814 {
 815         static const char *const bad[] = { "/", ".", "..", NULL };
 816         const char *path = RTE_EAL_PMD_PATH;
 817         size_t len = strlen(path);
 818         size_t off;
 819         int i;
 820
 821         while (len && path[len - 1] == '/')
 822                 --len;
 823         for (off = len; off && path[off - 1] != '/'; --off)
 824                 ;
 825         for (i = 0; bad[i]; ++i)
 826                 if (!strncmp(path + off, bad[i], (int)(len - off)))
 827                         goto error;
 828         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
 829         if (i == -1 || (size_t)i >= size)
 830                 goto error;
 831         return buf;
 832 error:
 833         ERROR("unable to append \"-glue\" to last component of"
 834               " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
 835               " please re-configure DPDK");
 836         return NULL;
 837 }
 838
 839 /**
 840  * Initialization routine for run-time dependency on rdma-core.
 841  */
 842 static int
 843 mlx4_glue_init(void)
 844 {
 845         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
 846         const char *path[] = {
 847                 /*
 848                  * A basic security check is necessary before trusting
 849                  * MLX4_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
 850                  */
 851                 (geteuid() == getuid() && getegid() == getgid() ?
 852                  getenv("MLX4_GLUE_PATH") : NULL),
 853                 /*
 854                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
 855                  * variant, otherwise let dlopen() look up libraries on its
 856                  * own.
 857                  */
 858                 (*RTE_EAL_PMD_PATH ?
 859                  mlx4_glue_path(glue_path, sizeof(glue_path)) : ""),
 860         };
 861         unsigned int i = 0;
 862         void *handle = NULL;
 863         void **sym;
 864         const char *dlmsg;
 865
 866         while (!handle && i != RTE_DIM(path)) {
 867                 const char *end;
 868                 size_t len;
 869                 int ret;
 870
 871                 if (!path[i]) {
 872                         ++i;
 873                         continue;
 874                 }
 875                 end = strpbrk(path[i], ":;");
 876                 if (!end)
 877                         end = path[i] + strlen(path[i]);
 878                 len = end - path[i];
 879                 ret = 0;
 880                 do {
 881                         char name[ret + 1];
 882
 883                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX4_GLUE,
 884                                        (int)len, path[i],
 885                                        (!len || *(end - 1) == '/') ? "" : "/");
 886                         if (ret == -1)
 887                                 break;
 888                         if (sizeof(name) != (size_t)ret + 1)
 889                                 continue;
 890                         DEBUG("looking for rdma-core glue as \"%s\"", name);
 891                         handle = dlopen(name, RTLD_LAZY);
 892                         break;
 893                 } while (1);
 894                 path[i] = end + 1;
 895                 if (!*end)
 896                         ++i;
 897         }
 898         if (!handle) {
 899                 rte_errno = EINVAL;
 900                 dlmsg = dlerror();
 901                 if (dlmsg)
 902                         WARN("cannot load glue library: %s", dlmsg);
 903                 goto glue_error;
 904         }
 905         sym = dlsym(handle, "mlx4_glue");
 906         if (!sym || !*sym) {
 907                 rte_errno = EINVAL;
 908                 dlmsg = dlerror();
 909                 if (dlmsg)
 910                         ERROR("cannot resolve glue symbol: %s", dlmsg);
 911                 goto glue_error;
 912         }
 913         mlx4_glue = *sym;
 914         return 0;
 915 glue_error:
 916         if (handle)
 917                 dlclose(handle);
 918         WARN("cannot initialize PMD due to missing run-time"
 919              " dependency on rdma-core libraries (libibverbs,"
 920              " libmlx4)");
 921         return -rte_errno;
 922 }
 923
 924 #endif
 925
 926 /**
 927  * Driver initialization routine.
 928  */
 929 RTE_INIT(rte_mlx4_pmd_init);
 930 static void
 931 rte_mlx4_pmd_init(void)
 932 {
 933         /*
 934          * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we
 935          * want to get success errno value in case of calling them
 936          * when the device was removed.
 937          */
 938         setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1);
 939         /*
 940          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
 941          * huge pages. Calling ibv_fork_init() during init allows
 942          * applications to use fork() safely for purposes other than
 943          * using this PMD, which is not supported in forked processes.
 944          */
 945         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
 946 #ifdef RTE_LIBRTE_MLX4_DLOPEN_DEPS
 947         if (mlx4_glue_init())
 948                 return;
 949         assert(mlx4_glue);
 950 #endif
 951 #ifndef NDEBUG
 952         /* Glue structure must not contain any NULL pointers. */
 953         {
 954                 unsigned int i;
 955
 956                 for (i = 0; i != sizeof(*mlx4_glue) / sizeof(void *); ++i)
 957                         assert(((const void *const *)mlx4_glue)[i]);
 958         }
 959 #endif
 960         if (strcmp(mlx4_glue->version, MLX4_GLUE_VERSION)) {
 961                 ERROR("rdma-core glue \"%s\" mismatch: \"%s\" is required",
 962                       mlx4_glue->version, MLX4_GLUE_VERSION);
 963                 return;
 964         }
 965         mlx4_glue->fork_init();
 966         rte_pci_register(&mlx4_driver);
 967 }
 968
 969 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
 970 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
 971 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
 972         "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");