drivers/net/mlx5/mlx5_ethdev.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2015 6WIND S.A.
   3  * Copyright 2015 Mellanox Technologies, Ltd
   4  */
   5
   6 #define _GNU_SOURCE
   7
   8 #include <stddef.h>
   9 #include <assert.h>
  10 #include <inttypes.h>
  11 #include <unistd.h>
  12 #include <stdint.h>
  13 #include <stdio.h>
  14 #include <string.h>
  15 #include <stdlib.h>
  16 #include <errno.h>
  17 #include <dirent.h>
  18 #include <net/if.h>
  19 #include <sys/ioctl.h>
  20 #include <sys/socket.h>
  21 #include <netinet/in.h>
  22 #include <linux/ethtool.h>
  23 #include <linux/sockios.h>
  24 #include <fcntl.h>
  25 #include <stdalign.h>
  26 #include <sys/un.h>
  27 #include <time.h>
  28
  29 #include <rte_atomic.h>
  30 #include <rte_ethdev_driver.h>
  31 #include <rte_bus_pci.h>
  32 #include <rte_mbuf.h>
  33 #include <rte_common.h>
  34 #include <rte_interrupts.h>
  35 #include <rte_malloc.h>
  36 #include <rte_string_fns.h>
  37 #include <rte_rwlock.h>
  38
  39 #include "mlx5.h"
  40 #include "mlx5_glue.h"
  41 #include "mlx5_rxtx.h"
  42 #include "mlx5_utils.h"
  43
  44 /* Add defines in case the running kernel is not the same as user headers. */
  45 #ifndef ETHTOOL_GLINKSETTINGS
  46 struct ethtool_link_settings {
  47         uint32_t cmd;
  48         uint32_t speed;
  49         uint8_t duplex;
  50         uint8_t port;
  51         uint8_t phy_address;
  52         uint8_t autoneg;
  53         uint8_t mdio_support;
  54         uint8_t eth_to_mdix;
  55         uint8_t eth_tp_mdix_ctrl;
  56         int8_t link_mode_masks_nwords;
  57         uint32_t reserved[8];
  58         uint32_t link_mode_masks[];
  59 };
  60
  61 #define ETHTOOL_GLINKSETTINGS 0x0000004c
  62 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
  63 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
  64 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
  65 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
  66 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
  67 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
  68 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
  69 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
  70 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
  71 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
  72 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
  73 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
  74 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
  75 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
  76 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
  77 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
  78 #endif
  79 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
  80 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
  81 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
  82 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
  83 #endif
  84 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
  85 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
  86 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
  87 #endif
  88 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
  89 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
  90 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
  91 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
  92 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
  93 #endif
  94
  95 /**
  96  * Get interface name from private structure.
  97  *
  98  * @param[in] dev
  99  *   Pointer to Ethernet device.
 100  * @param[out] ifname
 101  *   Interface name output buffer.
 102  *
 103  * @return
 104  *   0 on success, a negative errno value otherwise and rte_errno is set.
 105  */
 106 int
 107 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
 108 {
 109         struct priv *priv = dev->data->dev_private;
 110         DIR *dir;
 111         struct dirent *dent;
 112         unsigned int dev_type = 0;
 113         unsigned int dev_port_prev = ~0u;
 114         char match[IF_NAMESIZE] = "";
 115
 116         {
 117                 MKSTR(path, "%s/device/net", priv->ibdev_path);
 118
 119                 dir = opendir(path);
 120                 if (dir == NULL) {
 121                         rte_errno = errno;
 122                         return -rte_errno;
 123                 }
 124         }
 125         while ((dent = readdir(dir)) != NULL) {
 126                 char *name = dent->d_name;
 127                 FILE *file;
 128                 unsigned int dev_port;
 129                 int r;
 130
 131                 if ((name[0] == '.') &&
 132                     ((name[1] == '\0') ||
 133                      ((name[1] == '.') && (name[2] == '\0'))))
 134                         continue;
 135
 136                 MKSTR(path, "%s/device/net/%s/%s",
 137                       priv->ibdev_path, name,
 138                       (dev_type ? "dev_id" : "dev_port"));
 139
 140                 file = fopen(path, "rb");
 141                 if (file == NULL) {
 142                         if (errno != ENOENT)
 143                                 continue;
 144                         /*
 145                          * Switch to dev_id when dev_port does not exist as
 146                          * is the case with Linux kernel versions < 3.15.
 147                          */
 148 try_dev_id:
 149                         match[0] = '\0';
 150                         if (dev_type)
 151                                 break;
 152                         dev_type = 1;
 153                         dev_port_prev = ~0u;
 154                         rewinddir(dir);
 155                         continue;
 156                 }
 157                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 158                 fclose(file);
 159                 if (r != 1)
 160                         continue;
 161                 /*
 162                  * Switch to dev_id when dev_port returns the same value for
 163                  * all ports. May happen when using a MOFED release older than
 164                  * 3.0 with a Linux kernel >= 3.15.
 165                  */
 166                 if (dev_port == dev_port_prev)
 167                         goto try_dev_id;
 168                 dev_port_prev = dev_port;
 169                 if (dev_port == (priv->port - 1u))
 170                         strlcpy(match, name, sizeof(match));
 171         }
 172         closedir(dir);
 173         if (match[0] == '\0') {
 174                 rte_errno = ENOENT;
 175                 return -rte_errno;
 176         }
 177         strncpy(*ifname, match, sizeof(*ifname));
 178         return 0;
 179 }
 180
 181 /**
 182  * Get the interface index from device name.
 183  *
 184  * @param[in] dev
 185  *   Pointer to Ethernet device.
 186  *
 187  * @return
 188  *   Interface index on success, a negative errno value otherwise and
 189  *   rte_errno is set.
 190  */
 191 int
 192 mlx5_ifindex(const struct rte_eth_dev *dev)
 193 {
 194         char ifname[IF_NAMESIZE];
 195         int ret;
 196
 197         ret = mlx5_get_ifname(dev, &ifname);
 198         if (ret)
 199                 return ret;
 200         ret = if_nametoindex(ifname);
 201         if (ret == -1) {
 202                 rte_errno = errno;
 203                 return -rte_errno;
 204         }
 205         return ret;
 206 }
 207
 208 /**
 209  * Perform ifreq ioctl() on associated Ethernet device.
 210  *
 211  * @param[in] dev
 212  *   Pointer to Ethernet device.
 213  * @param req
 214  *   Request number to pass to ioctl().
 215  * @param[out] ifr
 216  *   Interface request structure output buffer.
 217  *
 218  * @return
 219  *   0 on success, a negative errno value otherwise and rte_errno is set.
 220  */
 221 int
 222 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
 223 {
 224         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 225         int ret = 0;
 226
 227         if (sock == -1) {
 228                 rte_errno = errno;
 229                 return -rte_errno;
 230         }
 231         ret = mlx5_get_ifname(dev, &ifr->ifr_name);
 232         if (ret)
 233                 goto error;
 234         ret = ioctl(sock, req, ifr);
 235         if (ret == -1) {
 236                 rte_errno = errno;
 237                 goto error;
 238         }
 239         close(sock);
 240         return 0;
 241 error:
 242         close(sock);
 243         return -rte_errno;
 244 }
 245
 246 /**
 247  * Get device MTU.
 248  *
 249  * @param dev
 250  *   Pointer to Ethernet device.
 251  * @param[out] mtu
 252  *   MTU value output buffer.
 253  *
 254  * @return
 255  *   0 on success, a negative errno value otherwise and rte_errno is set.
 256  */
 257 int
 258 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
 259 {
 260         struct ifreq request;
 261         int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
 262
 263         if (ret)
 264                 return ret;
 265         *mtu = request.ifr_mtu;
 266         return 0;
 267 }
 268
 269 /**
 270  * Set device MTU.
 271  *
 272  * @param dev
 273  *   Pointer to Ethernet device.
 274  * @param mtu
 275  *   MTU value to set.
 276  *
 277  * @return
 278  *   0 on success, a negative errno value otherwise and rte_errno is set.
 279  */
 280 static int
 281 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 282 {
 283         struct ifreq request = { .ifr_mtu = mtu, };
 284
 285         return mlx5_ifreq(dev, SIOCSIFMTU, &request);
 286 }
 287
 288 /**
 289  * Set device flags.
 290  *
 291  * @param dev
 292  *   Pointer to Ethernet device.
 293  * @param keep
 294  *   Bitmask for flags that must remain untouched.
 295  * @param flags
 296  *   Bitmask for flags to modify.
 297  *
 298  * @return
 299  *   0 on success, a negative errno value otherwise and rte_errno is set.
 300  */
 301 int
 302 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
 303 {
 304         struct ifreq request;
 305         int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
 306
 307         if (ret)
 308                 return ret;
 309         request.ifr_flags &= keep;
 310         request.ifr_flags |= flags & ~keep;
 311         return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
 312 }
 313
 314 /**
 315  * DPDK callback for Ethernet device configuration.
 316  *
 317  * @param dev
 318  *   Pointer to Ethernet device structure.
 319  *
 320  * @return
 321  *   0 on success, a negative errno value otherwise and rte_errno is set.
 322  */
 323 int
 324 mlx5_dev_configure(struct rte_eth_dev *dev)
 325 {
 326         struct priv *priv = dev->data->dev_private;
 327         unsigned int rxqs_n = dev->data->nb_rx_queues;
 328         unsigned int txqs_n = dev->data->nb_tx_queues;
 329         unsigned int i;
 330         unsigned int j;
 331         unsigned int reta_idx_n;
 332         const uint8_t use_app_rss_key =
 333                 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
 334         int ret = 0;
 335
 336         if (use_app_rss_key &&
 337             (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
 338              rss_hash_default_key_len)) {
 339                 DRV_LOG(ERR, "port %u RSS key len must be %zu Bytes long",
 340                         dev->data->port_id, rss_hash_default_key_len);
 341                 rte_errno = EINVAL;
 342                 return -rte_errno;
 343         }
 344         priv->rss_conf.rss_key =
 345                 rte_realloc(priv->rss_conf.rss_key,
 346                             rss_hash_default_key_len, 0);
 347         if (!priv->rss_conf.rss_key) {
 348                 DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
 349                         dev->data->port_id, rxqs_n);
 350                 rte_errno = ENOMEM;
 351                 return -rte_errno;
 352         }
 353         memcpy(priv->rss_conf.rss_key,
 354                use_app_rss_key ?
 355                dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
 356                rss_hash_default_key,
 357                rss_hash_default_key_len);
 358         priv->rss_conf.rss_key_len = rss_hash_default_key_len;
 359         priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
 360         priv->rxqs = (void *)dev->data->rx_queues;
 361         priv->txqs = (void *)dev->data->tx_queues;
 362         if (txqs_n != priv->txqs_n) {
 363                 DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
 364                         dev->data->port_id, priv->txqs_n, txqs_n);
 365                 priv->txqs_n = txqs_n;
 366         }
 367         if (rxqs_n > priv->config.ind_table_max_size) {
 368                 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
 369                         dev->data->port_id, rxqs_n);
 370                 rte_errno = EINVAL;
 371                 return -rte_errno;
 372         }
 373         if (rxqs_n == priv->rxqs_n)
 374                 return 0;
 375         DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
 376                 dev->data->port_id, priv->rxqs_n, rxqs_n);
 377         priv->rxqs_n = rxqs_n;
 378         /* If the requested number of RX queues is not a power of two, use the
 379          * maximum indirection table size for better balancing.
 380          * The result is always rounded to the next power of two. */
 381         reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
 382                                      priv->config.ind_table_max_size :
 383                                      rxqs_n));
 384         ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
 385         if (ret)
 386                 return ret;
 387         /* When the number of RX queues is not a power of two, the remaining
 388          * table entries are padded with reused WQs and hashes are not spread
 389          * uniformly. */
 390         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
 391                 (*priv->reta_idx)[i] = j;
 392                 if (++j == rxqs_n)
 393                         j = 0;
 394         }
 395         /*
 396          * Once the device is added to the list of memory event callback, its
 397          * global MR cache table cannot be expanded on the fly because of
 398          * deadlock. If it overflows, lookup should be done by searching MR list
 399          * linearly, which is slow.
 400          */
 401         if (mlx5_mr_btree_init(&priv->mr.cache, MLX5_MR_BTREE_CACHE_N * 2,
 402                                dev->device->numa_node)) {
 403                 /* rte_errno is already set. */
 404                 return -rte_errno;
 405         }
 406         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
 407         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
 408                          priv, mem_event_cb);
 409         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
 410         return 0;
 411 }
 412
 413 /**
 414  * Sets default tuning parameters.
 415  *
 416  * @param dev
 417  *   Pointer to Ethernet device.
 418  * @param[out] info
 419  *   Info structure output buffer.
 420  */
 421 static void
 422 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 423 {
 424         struct priv *priv = dev->data->dev_private;
 425
 426         /* Minimum CPU utilization. */
 427         info->default_rxportconf.ring_size = 256;
 428         info->default_txportconf.ring_size = 256;
 429         info->default_rxportconf.burst_size = 64;
 430         info->default_txportconf.burst_size = 64;
 431         if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
 432                 info->default_rxportconf.nb_queues = 16;
 433                 info->default_txportconf.nb_queues = 16;
 434                 if (dev->data->nb_rx_queues > 2 ||
 435                     dev->data->nb_tx_queues > 2) {
 436                         /* Max Throughput. */
 437                         info->default_rxportconf.ring_size = 2048;
 438                         info->default_txportconf.ring_size = 2048;
 439                 }
 440         } else {
 441                 info->default_rxportconf.nb_queues = 8;
 442                 info->default_txportconf.nb_queues = 8;
 443                 if (dev->data->nb_rx_queues > 2 ||
 444                     dev->data->nb_tx_queues > 2) {
 445                         /* Max Throughput. */
 446                         info->default_rxportconf.ring_size = 4096;
 447                         info->default_txportconf.ring_size = 4096;
 448                 }
 449         }
 450 }
 451
 452 /**
 453  * DPDK callback to get information about the device.
 454  *
 455  * @param dev
 456  *   Pointer to Ethernet device structure.
 457  * @param[out] info
 458  *   Info structure output buffer.
 459  */
 460 void
 461 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 462 {
 463         struct priv *priv = dev->data->dev_private;
 464         struct mlx5_dev_config *config = &priv->config;
 465         unsigned int max;
 466         char ifname[IF_NAMESIZE];
 467
 468         /* FIXME: we should ask the device for these values. */
 469         info->min_rx_bufsize = 32;
 470         info->max_rx_pktlen = 65536;
 471         /*
 472          * Since we need one CQ per QP, the limit is the minimum number
 473          * between the two values.
 474          */
 475         max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
 476                       priv->device_attr.orig_attr.max_qp);
 477         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
 478         if (max >= 65535)
 479                 max = 65535;
 480         info->max_rx_queues = max;
 481         info->max_tx_queues = max;
 482         info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
 483         info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
 484         info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
 485                                  info->rx_queue_offload_capa);
 486         info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
 487         if (mlx5_get_ifname(dev, &ifname) == 0)
 488                 info->if_index = if_nametoindex(ifname);
 489         info->reta_size = priv->reta_idx_n ?
 490                 priv->reta_idx_n : config->ind_table_max_size;
 491         info->hash_key_size = rss_hash_default_key_len;
 492         info->speed_capa = priv->link_speed_capa;
 493         info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
 494         mlx5_set_default_params(dev, info);
 495 }
 496
 497 /**
 498  * Get supported packet types.
 499  *
 500  * @param dev
 501  *   Pointer to Ethernet device structure.
 502  *
 503  * @return
 504  *   A pointer to the supported Packet types array.
 505  */
 506 const uint32_t *
 507 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 508 {
 509         static const uint32_t ptypes[] = {
 510                 /* refers to rxq_cq_to_pkt_type() */
 511                 RTE_PTYPE_L2_ETHER,
 512                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 513                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 514                 RTE_PTYPE_L4_NONFRAG,
 515                 RTE_PTYPE_L4_FRAG,
 516                 RTE_PTYPE_L4_TCP,
 517                 RTE_PTYPE_L4_UDP,
 518                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 519                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 520                 RTE_PTYPE_INNER_L4_NONFRAG,
 521                 RTE_PTYPE_INNER_L4_FRAG,
 522                 RTE_PTYPE_INNER_L4_TCP,
 523                 RTE_PTYPE_INNER_L4_UDP,
 524                 RTE_PTYPE_UNKNOWN
 525         };
 526
 527         if (dev->rx_pkt_burst == mlx5_rx_burst ||
 528             dev->rx_pkt_burst == mlx5_rx_burst_vec)
 529                 return ptypes;
 530         return NULL;
 531 }
 532
 533 /**
 534  * DPDK callback to retrieve physical link information.
 535  *
 536  * @param dev
 537  *   Pointer to Ethernet device structure.
 538  * @param[out] link
 539  *   Storage for current link status.
 540  *
 541  * @return
 542  *   0 on success, a negative errno value otherwise and rte_errno is set.
 543  */
 544 static int
 545 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
 546                                struct rte_eth_link *link)
 547 {
 548         struct priv *priv = dev->data->dev_private;
 549         struct ethtool_cmd edata = {
 550                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 551         };
 552         struct ifreq ifr;
 553         struct rte_eth_link dev_link;
 554         int link_speed = 0;
 555         int ret;
 556
 557         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 558         if (ret) {
 559                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 560                         dev->data->port_id, strerror(rte_errno));
 561                 return ret;
 562         }
 563         memset(&dev_link, 0, sizeof(dev_link));
 564         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 565                                 (ifr.ifr_flags & IFF_RUNNING));
 566         ifr.ifr_data = (void *)&edata;
 567         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 568         if (ret) {
 569                 DRV_LOG(WARNING,
 570                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
 571                         dev->data->port_id, strerror(rte_errno));
 572                 return ret;
 573         }
 574         link_speed = ethtool_cmd_speed(&edata);
 575         if (link_speed == -1)
 576                 dev_link.link_speed = ETH_SPEED_NUM_NONE;
 577         else
 578                 dev_link.link_speed = link_speed;
 579         priv->link_speed_capa = 0;
 580         if (edata.supported & SUPPORTED_Autoneg)
 581                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 582         if (edata.supported & (SUPPORTED_1000baseT_Full |
 583                                SUPPORTED_1000baseKX_Full))
 584                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 585         if (edata.supported & SUPPORTED_10000baseKR_Full)
 586                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 587         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 588                                SUPPORTED_40000baseCR4_Full |
 589                                SUPPORTED_40000baseSR4_Full |
 590                                SUPPORTED_40000baseLR4_Full))
 591                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 592         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 593                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 594         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 595                         ETH_LINK_SPEED_FIXED);
 596         if ((dev_link.link_speed && !dev_link.link_status) ||
 597             (!dev_link.link_speed && dev_link.link_status)) {
 598                 rte_errno = EAGAIN;
 599                 return -rte_errno;
 600         }
 601         *link = dev_link;
 602         return 0;
 603 }
 604
 605 /**
 606  * Retrieve physical link information (unlocked version using new ioctl).
 607  *
 608  * @param dev
 609  *   Pointer to Ethernet device structure.
 610  * @param[out] link
 611  *   Storage for current link status.
 612  *
 613  * @return
 614  *   0 on success, a negative errno value otherwise and rte_errno is set.
 615  */
 616 static int
 617 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
 618                              struct rte_eth_link *link)
 619
 620 {
 621         struct priv *priv = dev->data->dev_private;
 622         struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
 623         struct ifreq ifr;
 624         struct rte_eth_link dev_link;
 625         uint64_t sc;
 626         int ret;
 627
 628         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 629         if (ret) {
 630                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 631                         dev->data->port_id, strerror(rte_errno));
 632                 return ret;
 633         }
 634         memset(&dev_link, 0, sizeof(dev_link));
 635         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 636                                 (ifr.ifr_flags & IFF_RUNNING));
 637         ifr.ifr_data = (void *)&gcmd;
 638         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 639         if (ret) {
 640                 DRV_LOG(DEBUG,
 641                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
 642                         " failed: %s",
 643                         dev->data->port_id, strerror(rte_errno));
 644                 return ret;
 645         }
 646         gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
 647
 648         alignas(struct ethtool_link_settings)
 649         uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
 650                      sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
 651         struct ethtool_link_settings *ecmd = (void *)data;
 652
 653         *ecmd = gcmd;
 654         ifr.ifr_data = (void *)ecmd;
 655         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 656         if (ret) {
 657                 DRV_LOG(DEBUG,
 658                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)"
 659                         " failed: %s",
 660                         dev->data->port_id, strerror(rte_errno));
 661                 return ret;
 662         }
 663         dev_link.link_speed = ecmd->speed;
 664         sc = ecmd->link_mode_masks[0] |
 665                 ((uint64_t)ecmd->link_mode_masks[1] << 32);
 666         priv->link_speed_capa = 0;
 667         if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
 668                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 669         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
 670                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
 671                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 672         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
 673                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
 674                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
 675                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 676         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
 677                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
 678                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
 679         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
 680                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
 681                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
 682                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
 683                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 684         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
 685                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
 686                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
 687                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
 688                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
 689         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
 690                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
 691                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
 692                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
 693         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
 694                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
 695                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
 696         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
 697                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
 698                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
 699                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
 700                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
 701         dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
 702                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 703         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 704                                   ETH_LINK_SPEED_FIXED);
 705         if ((dev_link.link_speed && !dev_link.link_status) ||
 706             (!dev_link.link_speed && dev_link.link_status)) {
 707                 rte_errno = EAGAIN;
 708                 return -rte_errno;
 709         }
 710         *link = dev_link;
 711         return 0;
 712 }
 713
 714 /**
 715  * DPDK callback to retrieve physical link information.
 716  *
 717  * @param dev
 718  *   Pointer to Ethernet device structure.
 719  * @param wait_to_complete
 720  *   Wait for request completion.
 721  *
 722  * @return
 723  *   0 if link status was not updated, positive if it was, a negative errno
 724  *   value otherwise and rte_errno is set.
 725  */
 726 int
 727 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
 728 {
 729         int ret;
 730         struct rte_eth_link dev_link;
 731         time_t start_time = time(NULL);
 732
 733         do {
 734                 ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
 735                 if (ret)
 736                         ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
 737                 if (ret == 0)
 738                         break;
 739                 /* Handle wait to complete situation. */
 740                 if (wait_to_complete && ret == -EAGAIN) {
 741                         if (abs((int)difftime(time(NULL), start_time)) <
 742                             MLX5_LINK_STATUS_TIMEOUT) {
 743                                 usleep(0);
 744                                 continue;
 745                         } else {
 746                                 rte_errno = EBUSY;
 747                                 return -rte_errno;
 748                         }
 749                 } else if (ret < 0) {
 750                         return ret;
 751                 }
 752         } while (wait_to_complete);
 753         ret = !!memcmp(&dev->data->dev_link, &dev_link,
 754                        sizeof(struct rte_eth_link));
 755         dev->data->dev_link = dev_link;
 756         return ret;
 757 }
 758
 759 /**
 760  * DPDK callback to change the MTU.
 761  *
 762  * @param dev
 763  *   Pointer to Ethernet device structure.
 764  * @param in_mtu
 765  *   New MTU.
 766  *
 767  * @return
 768  *   0 on success, a negative errno value otherwise and rte_errno is set.
 769  */
 770 int
 771 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 772 {
 773         struct priv *priv = dev->data->dev_private;
 774         uint16_t kern_mtu = 0;
 775         int ret;
 776
 777         ret = mlx5_get_mtu(dev, &kern_mtu);
 778         if (ret)
 779                 return ret;
 780         /* Set kernel interface MTU first. */
 781         ret = mlx5_set_mtu(dev, mtu);
 782         if (ret)
 783                 return ret;
 784         ret = mlx5_get_mtu(dev, &kern_mtu);
 785         if (ret)
 786                 return ret;
 787         if (kern_mtu == mtu) {
 788                 priv->mtu = mtu;
 789                 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
 790                         dev->data->port_id, mtu);
 791                 return 0;
 792         }
 793         rte_errno = EAGAIN;
 794         return -rte_errno;
 795 }
 796
 797 /**
 798  * DPDK callback to get flow control status.
 799  *
 800  * @param dev
 801  *   Pointer to Ethernet device structure.
 802  * @param[out] fc_conf
 803  *   Flow control output buffer.
 804  *
 805  * @return
 806  *   0 on success, a negative errno value otherwise and rte_errno is set.
 807  */
 808 int
 809 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 810 {
 811         struct ifreq ifr;
 812         struct ethtool_pauseparam ethpause = {
 813                 .cmd = ETHTOOL_GPAUSEPARAM
 814         };
 815         int ret;
 816
 817         ifr.ifr_data = (void *)&ethpause;
 818         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 819         if (ret) {
 820                 DRV_LOG(WARNING,
 821                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
 822                         " %s",
 823                         dev->data->port_id, strerror(rte_errno));
 824                 return ret;
 825         }
 826         fc_conf->autoneg = ethpause.autoneg;
 827         if (ethpause.rx_pause && ethpause.tx_pause)
 828                 fc_conf->mode = RTE_FC_FULL;
 829         else if (ethpause.rx_pause)
 830                 fc_conf->mode = RTE_FC_RX_PAUSE;
 831         else if (ethpause.tx_pause)
 832                 fc_conf->mode = RTE_FC_TX_PAUSE;
 833         else
 834                 fc_conf->mode = RTE_FC_NONE;
 835         return 0;
 836 }
 837
 838 /**
 839  * DPDK callback to modify flow control parameters.
 840  *
 841  * @param dev
 842  *   Pointer to Ethernet device structure.
 843  * @param[in] fc_conf
 844  *   Flow control parameters.
 845  *
 846  * @return
 847  *   0 on success, a negative errno value otherwise and rte_errno is set.
 848  */
 849 int
 850 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 851 {
 852         struct ifreq ifr;
 853         struct ethtool_pauseparam ethpause = {
 854                 .cmd = ETHTOOL_SPAUSEPARAM
 855         };
 856         int ret;
 857
 858         ifr.ifr_data = (void *)&ethpause;
 859         ethpause.autoneg = fc_conf->autoneg;
 860         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
 861             (fc_conf->mode & RTE_FC_RX_PAUSE))
 862                 ethpause.rx_pause = 1;
 863         else
 864                 ethpause.rx_pause = 0;
 865
 866         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
 867             (fc_conf->mode & RTE_FC_TX_PAUSE))
 868                 ethpause.tx_pause = 1;
 869         else
 870                 ethpause.tx_pause = 0;
 871         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 872         if (ret) {
 873                 DRV_LOG(WARNING,
 874                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
 875                         " failed: %s",
 876                         dev->data->port_id, strerror(rte_errno));
 877                 return ret;
 878         }
 879         return 0;
 880 }
 881
 882 /**
 883  * Get PCI information from struct ibv_device.
 884  *
 885  * @param device
 886  *   Pointer to Ethernet device structure.
 887  * @param[out] pci_addr
 888  *   PCI bus address output buffer.
 889  *
 890  * @return
 891  *   0 on success, a negative errno value otherwise and rte_errno is set.
 892  */
 893 int
 894 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
 895                             struct rte_pci_addr *pci_addr)
 896 {
 897         FILE *file;
 898         char line[32];
 899         MKSTR(path, "%s/device/uevent", device->ibdev_path);
 900
 901         file = fopen(path, "rb");
 902         if (file == NULL) {
 903                 rte_errno = errno;
 904                 return -rte_errno;
 905         }
 906         while (fgets(line, sizeof(line), file) == line) {
 907                 size_t len = strlen(line);
 908                 int ret;
 909
 910                 /* Truncate long lines. */
 911                 if (len == (sizeof(line) - 1))
 912                         while (line[(len - 1)] != '\n') {
 913                                 ret = fgetc(file);
 914                                 if (ret == EOF)
 915                                         break;
 916                                 line[(len - 1)] = ret;
 917                         }
 918                 /* Extract information. */
 919                 if (sscanf(line,
 920                            "PCI_SLOT_NAME="
 921                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
 922                            &pci_addr->domain,
 923                            &pci_addr->bus,
 924                            &pci_addr->devid,
 925                            &pci_addr->function) == 4) {
 926                         ret = 0;
 927                         break;
 928                 }
 929         }
 930         fclose(file);
 931         return 0;
 932 }
 933
 934 /**
 935  * Device status handler.
 936  *
 937  * @param dev
 938  *   Pointer to Ethernet device.
 939  * @param events
 940  *   Pointer to event flags holder.
 941  *
 942  * @return
 943  *   Events bitmap of callback process which can be called immediately.
 944  */
 945 static uint32_t
 946 mlx5_dev_status_handler(struct rte_eth_dev *dev)
 947 {
 948         struct priv *priv = dev->data->dev_private;
 949         struct ibv_async_event event;
 950         uint32_t ret = 0;
 951
 952         if (mlx5_link_update(dev, 0) == -EAGAIN) {
 953                 usleep(0);
 954                 return 0;
 955         }
 956         /* Read all message and acknowledge them. */
 957         for (;;) {
 958                 if (mlx5_glue->get_async_event(priv->ctx, &event))
 959                         break;
 960                 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
 961                         event.event_type == IBV_EVENT_PORT_ERR) &&
 962                         (dev->data->dev_conf.intr_conf.lsc == 1))
 963                         ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
 964                 else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
 965                         dev->data->dev_conf.intr_conf.rmv == 1)
 966                         ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
 967                 else
 968                         DRV_LOG(DEBUG,
 969                                 "port %u event type %d on not handled",
 970                                 dev->data->port_id, event.event_type);
 971                 mlx5_glue->ack_async_event(&event);
 972         }
 973         return ret;
 974 }
 975
 976 /**
 977  * Handle interrupts from the NIC.
 978  *
 979  * @param[in] intr_handle
 980  *   Interrupt handler.
 981  * @param cb_arg
 982  *   Callback argument.
 983  */
 984 void
 985 mlx5_dev_interrupt_handler(void *cb_arg)
 986 {
 987         struct rte_eth_dev *dev = cb_arg;
 988         uint32_t events;
 989
 990         events = mlx5_dev_status_handler(dev);
 991         if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
 992                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
 993         if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
 994                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
 995 }
 996
 997 /**
 998  * Handle interrupts from the socket.
 999  *
1000  * @param cb_arg
1001  *   Callback argument.
1002  */
1003 static void
1004 mlx5_dev_handler_socket(void *cb_arg)
1005 {
1006         struct rte_eth_dev *dev = cb_arg;
1007
1008         mlx5_socket_handle(dev);
1009 }
1010
1011 /**
1012  * Uninstall interrupt handler.
1013  *
1014  * @param dev
1015  *   Pointer to Ethernet device.
1016  */
1017 void
1018 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1019 {
1020         struct priv *priv = dev->data->dev_private;
1021
1022         if (dev->data->dev_conf.intr_conf.lsc ||
1023             dev->data->dev_conf.intr_conf.rmv)
1024                 rte_intr_callback_unregister(&priv->intr_handle,
1025                                              mlx5_dev_interrupt_handler, dev);
1026         if (priv->primary_socket)
1027                 rte_intr_callback_unregister(&priv->intr_handle_socket,
1028                                              mlx5_dev_handler_socket, dev);
1029         priv->intr_handle.fd = 0;
1030         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1031         priv->intr_handle_socket.fd = 0;
1032         priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1033 }
1034
1035 /**
1036  * Install interrupt handler.
1037  *
1038  * @param dev
1039  *   Pointer to Ethernet device.
1040  */
1041 void
1042 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1043 {
1044         struct priv *priv = dev->data->dev_private;
1045         int ret;
1046         int flags;
1047
1048         assert(priv->ctx->async_fd > 0);
1049         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1050         ret = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1051         if (ret) {
1052                 DRV_LOG(INFO,
1053                         "port %u failed to change file descriptor async event"
1054                         " queue",
1055                         dev->data->port_id);
1056                 dev->data->dev_conf.intr_conf.lsc = 0;
1057                 dev->data->dev_conf.intr_conf.rmv = 0;
1058         }
1059         if (dev->data->dev_conf.intr_conf.lsc ||
1060             dev->data->dev_conf.intr_conf.rmv) {
1061                 priv->intr_handle.fd = priv->ctx->async_fd;
1062                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1063                 rte_intr_callback_register(&priv->intr_handle,
1064                                            mlx5_dev_interrupt_handler, dev);
1065         }
1066         ret = mlx5_socket_init(dev);
1067         if (ret)
1068                 DRV_LOG(ERR, "port %u cannot initialise socket: %s",
1069                         dev->data->port_id, strerror(rte_errno));
1070         else if (priv->primary_socket) {
1071                 priv->intr_handle_socket.fd = priv->primary_socket;
1072                 priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1073                 rte_intr_callback_register(&priv->intr_handle_socket,
1074                                            mlx5_dev_handler_socket, dev);
1075         }
1076 }
1077
1078 /**
1079  * DPDK callback to bring the link DOWN.
1080  *
1081  * @param dev
1082  *   Pointer to Ethernet device structure.
1083  *
1084  * @return
1085  *   0 on success, a negative errno value otherwise and rte_errno is set.
1086  */
1087 int
1088 mlx5_set_link_down(struct rte_eth_dev *dev)
1089 {
1090         return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1091 }
1092
1093 /**
1094  * DPDK callback to bring the link UP.
1095  *
1096  * @param dev
1097  *   Pointer to Ethernet device structure.
1098  *
1099  * @return
1100  *   0 on success, a negative errno value otherwise and rte_errno is set.
1101  */
1102 int
1103 mlx5_set_link_up(struct rte_eth_dev *dev)
1104 {
1105         return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1106 }
1107
1108 /**
1109  * Configure the TX function to use.
1110  *
1111  * @param dev
1112  *   Pointer to private data structure.
1113  *
1114  * @return
1115  *   Pointer to selected Tx burst function.
1116  */
1117 eth_tx_burst_t
1118 mlx5_select_tx_function(struct rte_eth_dev *dev)
1119 {
1120         struct priv *priv = dev->data->dev_private;
1121         eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1122         struct mlx5_dev_config *config = &priv->config;
1123         uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1124         int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1125                                     DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1126                                     DEV_TX_OFFLOAD_GRE_TNL_TSO |
1127                                     DEV_TX_OFFLOAD_IP_TNL_TSO |
1128                                     DEV_TX_OFFLOAD_UDP_TNL_TSO));
1129         int swp = !!(tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO |
1130                                     DEV_TX_OFFLOAD_UDP_TNL_TSO |
1131                                     DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM));
1132         int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1133
1134         assert(priv != NULL);
1135         /* Select appropriate TX function. */
1136         if (vlan_insert || tso || swp)
1137                 return tx_pkt_burst;
1138         if (config->mps == MLX5_MPW_ENHANCED) {
1139                 if (mlx5_check_vec_tx_support(dev) > 0) {
1140                         if (mlx5_check_raw_vec_tx_support(dev) > 0)
1141                                 tx_pkt_burst = mlx5_tx_burst_raw_vec;
1142                         else
1143                                 tx_pkt_burst = mlx5_tx_burst_vec;
1144                         DRV_LOG(DEBUG,
1145                                 "port %u selected enhanced MPW Tx vectorized"
1146                                 " function",
1147                                 dev->data->port_id);
1148                 } else {
1149                         tx_pkt_burst = mlx5_tx_burst_empw;
1150                         DRV_LOG(DEBUG,
1151                                 "port %u selected enhanced MPW Tx function",
1152                                 dev->data->port_id);
1153                 }
1154         } else if (config->mps && (config->txq_inline > 0)) {
1155                 tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1156                 DRV_LOG(DEBUG, "port %u selected MPW inline Tx function",
1157                         dev->data->port_id);
1158         } else if (config->mps) {
1159                 tx_pkt_burst = mlx5_tx_burst_mpw;
1160                 DRV_LOG(DEBUG, "port %u selected MPW Tx function",
1161                         dev->data->port_id);
1162         }
1163         return tx_pkt_burst;
1164 }
1165
1166 /**
1167  * Configure the RX function to use.
1168  *
1169  * @param dev
1170  *   Pointer to private data structure.
1171  *
1172  * @return
1173  *   Pointer to selected Rx burst function.
1174  */
1175 eth_rx_burst_t
1176 mlx5_select_rx_function(struct rte_eth_dev *dev)
1177 {
1178         eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1179
1180         assert(dev != NULL);
1181         if (mlx5_check_vec_rx_support(dev) > 0) {
1182                 rx_pkt_burst = mlx5_rx_burst_vec;
1183                 DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1184                         dev->data->port_id);
1185         }
1186         return rx_pkt_burst;
1187 }
1188
1189 /**
1190  * Check if mlx5 device was removed.
1191  *
1192  * @param dev
1193  *   Pointer to Ethernet device structure.
1194  *
1195  * @return
1196  *   1 when device is removed, otherwise 0.
1197  */
1198 int
1199 mlx5_is_removed(struct rte_eth_dev *dev)
1200 {
1201         struct ibv_device_attr device_attr;
1202         struct priv *priv = dev->data->dev_private;
1203
1204         if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1205                 return 1;
1206         return 0;
1207 }