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 #include <stddef.h>
   7 #include <assert.h>
   8 #include <inttypes.h>
   9 #include <unistd.h>
  10 #include <stdbool.h>
  11 #include <stdint.h>
  12 #include <stdio.h>
  13 #include <string.h>
  14 #include <stdlib.h>
  15 #include <errno.h>
  16 #include <dirent.h>
  17 #include <net/if.h>
  18 #include <sys/ioctl.h>
  19 #include <sys/socket.h>
  20 #include <netinet/in.h>
  21 #include <linux/ethtool.h>
  22 #include <linux/sockios.h>
  23 #include <fcntl.h>
  24 #include <stdalign.h>
  25 #include <sys/un.h>
  26 #include <time.h>
  27
  28 #include <rte_atomic.h>
  29 #include <rte_ethdev_driver.h>
  30 #include <rte_bus_pci.h>
  31 #include <rte_mbuf.h>
  32 #include <rte_common.h>
  33 #include <rte_interrupts.h>
  34 #include <rte_malloc.h>
  35 #include <rte_string_fns.h>
  36 #include <rte_rwlock.h>
  37 #include <rte_cycles.h>
  38
  39 #include "mlx5.h"
  40 #include "mlx5_glue.h"
  41 #include "mlx5_rxtx.h"
  42 #include "mlx5_utils.h"
  43
  44 /* Supported speed values found in /usr/include/linux/ethtool.h */
  45 #ifndef HAVE_SUPPORTED_40000baseKR4_Full
  46 #define SUPPORTED_40000baseKR4_Full (1 << 23)
  47 #endif
  48 #ifndef HAVE_SUPPORTED_40000baseCR4_Full
  49 #define SUPPORTED_40000baseCR4_Full (1 << 24)
  50 #endif
  51 #ifndef HAVE_SUPPORTED_40000baseSR4_Full
  52 #define SUPPORTED_40000baseSR4_Full (1 << 25)
  53 #endif
  54 #ifndef HAVE_SUPPORTED_40000baseLR4_Full
  55 #define SUPPORTED_40000baseLR4_Full (1 << 26)
  56 #endif
  57 #ifndef HAVE_SUPPORTED_56000baseKR4_Full
  58 #define SUPPORTED_56000baseKR4_Full (1 << 27)
  59 #endif
  60 #ifndef HAVE_SUPPORTED_56000baseCR4_Full
  61 #define SUPPORTED_56000baseCR4_Full (1 << 28)
  62 #endif
  63 #ifndef HAVE_SUPPORTED_56000baseSR4_Full
  64 #define SUPPORTED_56000baseSR4_Full (1 << 29)
  65 #endif
  66 #ifndef HAVE_SUPPORTED_56000baseLR4_Full
  67 #define SUPPORTED_56000baseLR4_Full (1 << 30)
  68 #endif
  69
  70 /* Add defines in case the running kernel is not the same as user headers. */
  71 #ifndef ETHTOOL_GLINKSETTINGS
  72 struct ethtool_link_settings {
  73         uint32_t cmd;
  74         uint32_t speed;
  75         uint8_t duplex;
  76         uint8_t port;
  77         uint8_t phy_address;
  78         uint8_t autoneg;
  79         uint8_t mdio_support;
  80         uint8_t eth_to_mdix;
  81         uint8_t eth_tp_mdix_ctrl;
  82         int8_t link_mode_masks_nwords;
  83         uint32_t reserved[8];
  84         uint32_t link_mode_masks[];
  85 };
  86
  87 #define ETHTOOL_GLINKSETTINGS 0x0000004c
  88 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
  89 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
  90 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
  91 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
  92 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
  93 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
  94 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
  95 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
  96 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
  97 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
  98 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
  99 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
 100 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
 101 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
 102 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
 103 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
 104 #endif
 105 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
 106 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
 107 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
 108 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
 109 #endif
 110 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
 111 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
 112 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
 113 #endif
 114 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
 115 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
 116 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
 117 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
 118 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
 119 #endif
 120
 121 /**
 122  * Get master interface name from private structure.
 123  *
 124  * @param[in] dev
 125  *   Pointer to Ethernet device.
 126  * @param[out] ifname
 127  *   Interface name output buffer.
 128  *
 129  * @return
 130  *   0 on success, a negative errno value otherwise and rte_errno is set.
 131  */
 132 int
 133 mlx5_get_master_ifname(const char *ibdev_path, char (*ifname)[IF_NAMESIZE])
 134 {
 135         DIR *dir;
 136         struct dirent *dent;
 137         unsigned int dev_type = 0;
 138         unsigned int dev_port_prev = ~0u;
 139         char match[IF_NAMESIZE] = "";
 140
 141         assert(ibdev_path);
 142         {
 143                 MKSTR(path, "%s/device/net", ibdev_path);
 144
 145                 dir = opendir(path);
 146                 if (dir == NULL) {
 147                         rte_errno = errno;
 148                         return -rte_errno;
 149                 }
 150         }
 151         while ((dent = readdir(dir)) != NULL) {
 152                 char *name = dent->d_name;
 153                 FILE *file;
 154                 unsigned int dev_port;
 155                 int r;
 156
 157                 if ((name[0] == '.') &&
 158                     ((name[1] == '\0') ||
 159                      ((name[1] == '.') && (name[2] == '\0'))))
 160                         continue;
 161
 162                 MKSTR(path, "%s/device/net/%s/%s",
 163                       ibdev_path, name,
 164                       (dev_type ? "dev_id" : "dev_port"));
 165
 166                 file = fopen(path, "rb");
 167                 if (file == NULL) {
 168                         if (errno != ENOENT)
 169                                 continue;
 170                         /*
 171                          * Switch to dev_id when dev_port does not exist as
 172                          * is the case with Linux kernel versions < 3.15.
 173                          */
 174 try_dev_id:
 175                         match[0] = '\0';
 176                         if (dev_type)
 177                                 break;
 178                         dev_type = 1;
 179                         dev_port_prev = ~0u;
 180                         rewinddir(dir);
 181                         continue;
 182                 }
 183                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 184                 fclose(file);
 185                 if (r != 1)
 186                         continue;
 187                 /*
 188                  * Switch to dev_id when dev_port returns the same value for
 189                  * all ports. May happen when using a MOFED release older than
 190                  * 3.0 with a Linux kernel >= 3.15.
 191                  */
 192                 if (dev_port == dev_port_prev)
 193                         goto try_dev_id;
 194                 dev_port_prev = dev_port;
 195                 if (dev_port == 0)
 196                         strlcpy(match, name, sizeof(match));
 197         }
 198         closedir(dir);
 199         if (match[0] == '\0') {
 200                 rte_errno = ENOENT;
 201                 return -rte_errno;
 202         }
 203         strncpy(*ifname, match, sizeof(*ifname));
 204         return 0;
 205 }
 206
 207 /**
 208  * Get interface name from private structure.
 209  *
 210  * This is a port representor-aware version of mlx5_get_master_ifname().
 211  *
 212  * @param[in] dev
 213  *   Pointer to Ethernet device.
 214  * @param[out] ifname
 215  *   Interface name output buffer.
 216  *
 217  * @return
 218  *   0 on success, a negative errno value otherwise and rte_errno is set.
 219  */
 220 int
 221 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
 222 {
 223         struct mlx5_priv *priv = dev->data->dev_private;
 224         unsigned int ifindex;
 225
 226         assert(priv);
 227         assert(priv->sh);
 228         ifindex = mlx5_ifindex(dev);
 229         if (!ifindex) {
 230                 if (!priv->representor)
 231                         return mlx5_get_master_ifname(priv->sh->ibdev_path,
 232                                                       ifname);
 233                 rte_errno = ENXIO;
 234                 return -rte_errno;
 235         }
 236         if (if_indextoname(ifindex, &(*ifname)[0]))
 237                 return 0;
 238         rte_errno = errno;
 239         return -rte_errno;
 240 }
 241
 242 /**
 243  * Get the interface index from device name.
 244  *
 245  * @param[in] dev
 246  *   Pointer to Ethernet device.
 247  *
 248  * @return
 249  *   Nonzero interface index on success, zero otherwise and rte_errno is set.
 250  */
 251 unsigned int
 252 mlx5_ifindex(const struct rte_eth_dev *dev)
 253 {
 254         struct mlx5_priv *priv = dev->data->dev_private;
 255         unsigned int ifindex;
 256
 257         assert(priv);
 258         assert(priv->if_index);
 259         ifindex = priv->if_index;
 260         if (!ifindex)
 261                 rte_errno = ENXIO;
 262         return ifindex;
 263 }
 264
 265 /**
 266  * Perform ifreq ioctl() on associated Ethernet device.
 267  *
 268  * @param[in] dev
 269  *   Pointer to Ethernet device.
 270  * @param req
 271  *   Request number to pass to ioctl().
 272  * @param[out] ifr
 273  *   Interface request structure output buffer.
 274  *
 275  * @return
 276  *   0 on success, a negative errno value otherwise and rte_errno is set.
 277  */
 278 int
 279 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
 280 {
 281         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 282         int ret = 0;
 283
 284         if (sock == -1) {
 285                 rte_errno = errno;
 286                 return -rte_errno;
 287         }
 288         ret = mlx5_get_ifname(dev, &ifr->ifr_name);
 289         if (ret)
 290                 goto error;
 291         ret = ioctl(sock, req, ifr);
 292         if (ret == -1) {
 293                 rte_errno = errno;
 294                 goto error;
 295         }
 296         close(sock);
 297         return 0;
 298 error:
 299         close(sock);
 300         return -rte_errno;
 301 }
 302
 303 /**
 304  * Get device MTU.
 305  *
 306  * @param dev
 307  *   Pointer to Ethernet device.
 308  * @param[out] mtu
 309  *   MTU value output buffer.
 310  *
 311  * @return
 312  *   0 on success, a negative errno value otherwise and rte_errno is set.
 313  */
 314 int
 315 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
 316 {
 317         struct ifreq request;
 318         int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
 319
 320         if (ret)
 321                 return ret;
 322         *mtu = request.ifr_mtu;
 323         return 0;
 324 }
 325
 326 /**
 327  * Set device MTU.
 328  *
 329  * @param dev
 330  *   Pointer to Ethernet device.
 331  * @param mtu
 332  *   MTU value to set.
 333  *
 334  * @return
 335  *   0 on success, a negative errno value otherwise and rte_errno is set.
 336  */
 337 static int
 338 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 339 {
 340         struct ifreq request = { .ifr_mtu = mtu, };
 341
 342         return mlx5_ifreq(dev, SIOCSIFMTU, &request);
 343 }
 344
 345 /**
 346  * Set device flags.
 347  *
 348  * @param dev
 349  *   Pointer to Ethernet device.
 350  * @param keep
 351  *   Bitmask for flags that must remain untouched.
 352  * @param flags
 353  *   Bitmask for flags to modify.
 354  *
 355  * @return
 356  *   0 on success, a negative errno value otherwise and rte_errno is set.
 357  */
 358 int
 359 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
 360 {
 361         struct ifreq request;
 362         int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
 363
 364         if (ret)
 365                 return ret;
 366         request.ifr_flags &= keep;
 367         request.ifr_flags |= flags & ~keep;
 368         return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
 369 }
 370
 371 /**
 372  * DPDK callback for Ethernet device configuration.
 373  *
 374  * @param dev
 375  *   Pointer to Ethernet device structure.
 376  *
 377  * @return
 378  *   0 on success, a negative errno value otherwise and rte_errno is set.
 379  */
 380 int
 381 mlx5_dev_configure(struct rte_eth_dev *dev)
 382 {
 383         struct mlx5_priv *priv = dev->data->dev_private;
 384         unsigned int rxqs_n = dev->data->nb_rx_queues;
 385         unsigned int txqs_n = dev->data->nb_tx_queues;
 386         const uint8_t use_app_rss_key =
 387                 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
 388         int ret = 0;
 389
 390         if (use_app_rss_key &&
 391             (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
 392              MLX5_RSS_HASH_KEY_LEN)) {
 393                 DRV_LOG(ERR, "port %u RSS key len must be %s Bytes long",
 394                         dev->data->port_id, RTE_STR(MLX5_RSS_HASH_KEY_LEN));
 395                 rte_errno = EINVAL;
 396                 return -rte_errno;
 397         }
 398         priv->rss_conf.rss_key =
 399                 rte_realloc(priv->rss_conf.rss_key,
 400                             MLX5_RSS_HASH_KEY_LEN, 0);
 401         if (!priv->rss_conf.rss_key) {
 402                 DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
 403                         dev->data->port_id, rxqs_n);
 404                 rte_errno = ENOMEM;
 405                 return -rte_errno;
 406         }
 407
 408         if (dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
 409                 dev->data->dev_conf.rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH;
 410
 411         memcpy(priv->rss_conf.rss_key,
 412                use_app_rss_key ?
 413                dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
 414                rss_hash_default_key,
 415                MLX5_RSS_HASH_KEY_LEN);
 416         priv->rss_conf.rss_key_len = MLX5_RSS_HASH_KEY_LEN;
 417         priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
 418         priv->rxqs = (void *)dev->data->rx_queues;
 419         priv->txqs = (void *)dev->data->tx_queues;
 420         if (txqs_n != priv->txqs_n) {
 421                 DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
 422                         dev->data->port_id, priv->txqs_n, txqs_n);
 423                 priv->txqs_n = txqs_n;
 424         }
 425         if (rxqs_n > priv->config.ind_table_max_size) {
 426                 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
 427                         dev->data->port_id, rxqs_n);
 428                 rte_errno = EINVAL;
 429                 return -rte_errno;
 430         }
 431         if (rxqs_n != priv->rxqs_n) {
 432                 DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
 433                         dev->data->port_id, priv->rxqs_n, rxqs_n);
 434                 priv->rxqs_n = rxqs_n;
 435         }
 436         priv->skip_default_rss_reta = 0;
 437         ret = mlx5_proc_priv_init(dev);
 438         if (ret)
 439                 return ret;
 440         return 0;
 441 }
 442
 443 /**
 444  * Configure default RSS reta.
 445  *
 446  * @param dev
 447  *   Pointer to Ethernet device structure.
 448  *
 449  * @return
 450  *   0 on success, a negative errno value otherwise and rte_errno is set.
 451  */
 452 int
 453 mlx5_dev_configure_rss_reta(struct rte_eth_dev *dev)
 454 {
 455         struct mlx5_priv *priv = dev->data->dev_private;
 456         unsigned int rxqs_n = dev->data->nb_rx_queues;
 457         unsigned int i;
 458         unsigned int j;
 459         unsigned int reta_idx_n;
 460         int ret = 0;
 461         unsigned int *rss_queue_arr = NULL;
 462         unsigned int rss_queue_n = 0;
 463
 464         if (priv->skip_default_rss_reta)
 465                 return ret;
 466         rss_queue_arr = rte_malloc("", rxqs_n * sizeof(unsigned int), 0);
 467         if (!rss_queue_arr) {
 468                 DRV_LOG(ERR, "port %u cannot allocate RSS queue list (%u)",
 469                         dev->data->port_id, rxqs_n);
 470                 rte_errno = ENOMEM;
 471                 return -rte_errno;
 472         }
 473         for (i = 0, j = 0; i < rxqs_n; i++) {
 474                 struct mlx5_rxq_data *rxq_data;
 475                 struct mlx5_rxq_ctrl *rxq_ctrl;
 476
 477                 rxq_data = (*priv->rxqs)[i];
 478                 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 479                 if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
 480                         rss_queue_arr[j++] = i;
 481         }
 482         rss_queue_n = j;
 483         if (rss_queue_n > priv->config.ind_table_max_size) {
 484                 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
 485                         dev->data->port_id, rss_queue_n);
 486                 rte_errno = EINVAL;
 487                 rte_free(rss_queue_arr);
 488                 return -rte_errno;
 489         }
 490         DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
 491                 dev->data->port_id, priv->rxqs_n, rxqs_n);
 492         priv->rxqs_n = rxqs_n;
 493         /*
 494          * If the requested number of RX queues is not a power of two,
 495          * use the maximum indirection table size for better balancing.
 496          * The result is always rounded to the next power of two.
 497          */
 498         reta_idx_n = (1 << log2above((rss_queue_n & (rss_queue_n - 1)) ?
 499                                 priv->config.ind_table_max_size :
 500                                 rss_queue_n));
 501         ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
 502         if (ret) {
 503                 rte_free(rss_queue_arr);
 504                 return ret;
 505         }
 506         /*
 507          * When the number of RX queues is not a power of two,
 508          * the remaining table entries are padded with reused WQs
 509          * and hashes are not spread uniformly.
 510          */
 511         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
 512                 (*priv->reta_idx)[i] = rss_queue_arr[j];
 513                 if (++j == rss_queue_n)
 514                         j = 0;
 515         }
 516         rte_free(rss_queue_arr);
 517         return ret;
 518 }
 519
 520 /**
 521  * Sets default tuning parameters.
 522  *
 523  * @param dev
 524  *   Pointer to Ethernet device.
 525  * @param[out] info
 526  *   Info structure output buffer.
 527  */
 528 static void
 529 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 530 {
 531         struct mlx5_priv *priv = dev->data->dev_private;
 532
 533         /* Minimum CPU utilization. */
 534         info->default_rxportconf.ring_size = 256;
 535         info->default_txportconf.ring_size = 256;
 536         info->default_rxportconf.burst_size = MLX5_RX_DEFAULT_BURST;
 537         info->default_txportconf.burst_size = MLX5_TX_DEFAULT_BURST;
 538         if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
 539                 info->default_rxportconf.nb_queues = 16;
 540                 info->default_txportconf.nb_queues = 16;
 541                 if (dev->data->nb_rx_queues > 2 ||
 542                     dev->data->nb_tx_queues > 2) {
 543                         /* Max Throughput. */
 544                         info->default_rxportconf.ring_size = 2048;
 545                         info->default_txportconf.ring_size = 2048;
 546                 }
 547         } else {
 548                 info->default_rxportconf.nb_queues = 8;
 549                 info->default_txportconf.nb_queues = 8;
 550                 if (dev->data->nb_rx_queues > 2 ||
 551                     dev->data->nb_tx_queues > 2) {
 552                         /* Max Throughput. */
 553                         info->default_rxportconf.ring_size = 4096;
 554                         info->default_txportconf.ring_size = 4096;
 555                 }
 556         }
 557 }
 558
 559 /**
 560  * Sets tx mbuf limiting parameters.
 561  *
 562  * @param dev
 563  *   Pointer to Ethernet device.
 564  * @param[out] info
 565  *   Info structure output buffer.
 566  */
 567 static void
 568 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 569 {
 570         struct mlx5_priv *priv = dev->data->dev_private;
 571         struct mlx5_dev_config *config = &priv->config;
 572         unsigned int inlen;
 573         uint16_t nb_max;
 574
 575         inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
 576                 MLX5_SEND_DEF_INLINE_LEN :
 577                 (unsigned int)config->txq_inline_max;
 578         assert(config->txq_inline_min >= 0);
 579         inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
 580         inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
 581                                MLX5_ESEG_MIN_INLINE_SIZE -
 582                                MLX5_WQE_CSEG_SIZE -
 583                                MLX5_WQE_ESEG_SIZE -
 584                                MLX5_WQE_DSEG_SIZE * 2);
 585         nb_max = (MLX5_WQE_SIZE_MAX +
 586                   MLX5_ESEG_MIN_INLINE_SIZE -
 587                   MLX5_WQE_CSEG_SIZE -
 588                   MLX5_WQE_ESEG_SIZE -
 589                   MLX5_WQE_DSEG_SIZE -
 590                   inlen) / MLX5_WSEG_SIZE;
 591         info->tx_desc_lim.nb_seg_max = nb_max;
 592         info->tx_desc_lim.nb_mtu_seg_max = nb_max;
 593 }
 594
 595 /**
 596  * DPDK callback to get information about the device.
 597  *
 598  * @param dev
 599  *   Pointer to Ethernet device structure.
 600  * @param[out] info
 601  *   Info structure output buffer.
 602  */
 603 int
 604 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 605 {
 606         struct mlx5_priv *priv = dev->data->dev_private;
 607         struct mlx5_dev_config *config = &priv->config;
 608         unsigned int max;
 609
 610         /* FIXME: we should ask the device for these values. */
 611         info->min_rx_bufsize = 32;
 612         info->max_rx_pktlen = 65536;
 613         info->max_lro_pkt_size = MLX5_MAX_LRO_SIZE;
 614         /*
 615          * Since we need one CQ per QP, the limit is the minimum number
 616          * between the two values.
 617          */
 618         max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
 619                       priv->sh->device_attr.orig_attr.max_qp);
 620         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
 621         if (max >= 65535)
 622                 max = 65535;
 623         info->max_rx_queues = max;
 624         info->max_tx_queues = max;
 625         info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
 626         info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
 627         info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
 628                                  info->rx_queue_offload_capa);
 629         info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
 630         info->if_index = mlx5_ifindex(dev);
 631         info->reta_size = priv->reta_idx_n ?
 632                 priv->reta_idx_n : config->ind_table_max_size;
 633         info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
 634         info->speed_capa = priv->link_speed_capa;
 635         info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
 636         mlx5_set_default_params(dev, info);
 637         mlx5_set_txlimit_params(dev, info);
 638         info->switch_info.name = dev->data->name;
 639         info->switch_info.domain_id = priv->domain_id;
 640         info->switch_info.port_id = priv->representor_id;
 641         if (priv->representor) {
 642                 uint16_t port_id;
 643
 644                 if (priv->pf_bond >= 0) {
 645                         /*
 646                          * Switch port ID is opaque value with driver defined
 647                          * format. Push the PF index in bonding configurations
 648                          * in upper four bits of port ID. If we get too many
 649                          * representors (more than 4K) or PFs (more than 15)
 650                          * this approach must be reconsidered.
 651                          */
 652                         if ((info->switch_info.port_id >>
 653                                 MLX5_PORT_ID_BONDING_PF_SHIFT) ||
 654                             priv->pf_bond > MLX5_PORT_ID_BONDING_PF_MASK) {
 655                                 DRV_LOG(ERR, "can't update switch port ID"
 656                                              " for bonding device");
 657                                 assert(false);
 658                                 return -ENODEV;
 659                         }
 660                         info->switch_info.port_id |=
 661                                 priv->pf_bond << MLX5_PORT_ID_BONDING_PF_SHIFT;
 662                 }
 663                 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 664                         struct mlx5_priv *opriv =
 665                                 rte_eth_devices[port_id].data->dev_private;
 666
 667                         if (!opriv ||
 668                             opriv->representor ||
 669                             opriv->sh != priv->sh ||
 670                             opriv->domain_id != priv->domain_id)
 671                                 continue;
 672                         /*
 673                          * Override switch name with that of the master
 674                          * device.
 675                          */
 676                         info->switch_info.name = opriv->dev_data->name;
 677                         break;
 678                 }
 679         }
 680         return 0;
 681 }
 682
 683 /**
 684  * Get device current raw clock counter
 685  *
 686  * @param dev
 687  *   Pointer to Ethernet device structure.
 688  * @param[out] time
 689  *   Current raw clock counter of the device.
 690  *
 691  * @return
 692  *   0 if the clock has correctly been read
 693  *   The value of errno in case of error
 694  */
 695 int
 696 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
 697 {
 698         struct mlx5_priv *priv = dev->data->dev_private;
 699         struct ibv_context *ctx = priv->sh->ctx;
 700         struct ibv_values_ex values;
 701         int err = 0;
 702
 703         values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
 704         err = mlx5_glue->query_rt_values_ex(ctx, &values);
 705         if (err != 0) {
 706                 DRV_LOG(WARNING, "Could not query the clock !");
 707                 return err;
 708         }
 709         *clock = values.raw_clock.tv_nsec;
 710         return 0;
 711 }
 712
 713 /**
 714  * Get firmware version of a device.
 715  *
 716  * @param dev
 717  *   Ethernet device port.
 718  * @param fw_ver
 719  *   String output allocated by caller.
 720  * @param fw_size
 721  *   Size of the output string, including terminating null byte.
 722  *
 723  * @return
 724  *   0 on success, or the size of the non truncated string if too big.
 725  */
 726 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
 727 {
 728         struct mlx5_priv *priv = dev->data->dev_private;
 729         struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
 730         size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
 731
 732         if (fw_size < size)
 733                 return size;
 734         if (fw_ver != NULL)
 735                 strlcpy(fw_ver, attr->fw_ver, fw_size);
 736         return 0;
 737 }
 738
 739 /**
 740  * Get supported packet types.
 741  *
 742  * @param dev
 743  *   Pointer to Ethernet device structure.
 744  *
 745  * @return
 746  *   A pointer to the supported Packet types array.
 747  */
 748 const uint32_t *
 749 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 750 {
 751         static const uint32_t ptypes[] = {
 752                 /* refers to rxq_cq_to_pkt_type() */
 753                 RTE_PTYPE_L2_ETHER,
 754                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 755                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 756                 RTE_PTYPE_L4_NONFRAG,
 757                 RTE_PTYPE_L4_FRAG,
 758                 RTE_PTYPE_L4_TCP,
 759                 RTE_PTYPE_L4_UDP,
 760                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 761                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 762                 RTE_PTYPE_INNER_L4_NONFRAG,
 763                 RTE_PTYPE_INNER_L4_FRAG,
 764                 RTE_PTYPE_INNER_L4_TCP,
 765                 RTE_PTYPE_INNER_L4_UDP,
 766                 RTE_PTYPE_UNKNOWN
 767         };
 768
 769         if (dev->rx_pkt_burst == mlx5_rx_burst ||
 770             dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
 771             dev->rx_pkt_burst == mlx5_rx_burst_vec)
 772                 return ptypes;
 773         return NULL;
 774 }
 775
 776 /**
 777  * Retrieve the master device for representor in the same switch domain.
 778  *
 779  * @param dev
 780  *   Pointer to representor Ethernet device structure.
 781  *
 782  * @return
 783  *   Master device structure  on success, NULL otherwise.
 784  */
 785
 786 static struct rte_eth_dev *
 787 mlx5_find_master_dev(struct rte_eth_dev *dev)
 788 {
 789         struct mlx5_priv *priv;
 790         uint16_t port_id;
 791         uint16_t domain_id;
 792
 793         priv = dev->data->dev_private;
 794         domain_id = priv->domain_id;
 795         assert(priv->representor);
 796         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 797                 struct mlx5_priv *opriv =
 798                         rte_eth_devices[port_id].data->dev_private;
 799                 if (opriv &&
 800                     opriv->master &&
 801                     opriv->domain_id == domain_id &&
 802                     opriv->sh == priv->sh)
 803                         return &rte_eth_devices[port_id];
 804         }
 805         return NULL;
 806 }
 807
 808 /**
 809  * DPDK callback to retrieve physical link information.
 810  *
 811  * @param dev
 812  *   Pointer to Ethernet device structure.
 813  * @param[out] link
 814  *   Storage for current link status.
 815  *
 816  * @return
 817  *   0 on success, a negative errno value otherwise and rte_errno is set.
 818  */
 819 static int
 820 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
 821                                struct rte_eth_link *link)
 822 {
 823         struct mlx5_priv *priv = dev->data->dev_private;
 824         struct ethtool_cmd edata = {
 825                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 826         };
 827         struct ifreq ifr;
 828         struct rte_eth_link dev_link;
 829         int link_speed = 0;
 830         int ret;
 831
 832         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 833         if (ret) {
 834                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 835                         dev->data->port_id, strerror(rte_errno));
 836                 return ret;
 837         }
 838         dev_link = (struct rte_eth_link) {
 839                 .link_status = ((ifr.ifr_flags & IFF_UP) &&
 840                                 (ifr.ifr_flags & IFF_RUNNING)),
 841         };
 842         ifr = (struct ifreq) {
 843                 .ifr_data = (void *)&edata,
 844         };
 845         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 846         if (ret) {
 847                 if (ret == -ENOTSUP && priv->representor) {
 848                         struct rte_eth_dev *master;
 849
 850                         /*
 851                          * For representors we can try to inherit link
 852                          * settings from the master device. Actually
 853                          * link settings do not make a lot of sense
 854                          * for representors due to missing physical
 855                          * link. The old kernel drivers supported
 856                          * emulated settings query for representors,
 857                          * the new ones do not, so we have to add
 858                          * this code for compatibility issues.
 859                          */
 860                         master = mlx5_find_master_dev(dev);
 861                         if (master) {
 862                                 ifr = (struct ifreq) {
 863                                         .ifr_data = (void *)&edata,
 864                                 };
 865                                 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
 866                         }
 867                 }
 868                 if (ret) {
 869                         DRV_LOG(WARNING,
 870                                 "port %u ioctl(SIOCETHTOOL,"
 871                                 " ETHTOOL_GSET) failed: %s",
 872                                 dev->data->port_id, strerror(rte_errno));
 873                         return ret;
 874                 }
 875         }
 876         link_speed = ethtool_cmd_speed(&edata);
 877         if (link_speed == -1)
 878                 dev_link.link_speed = ETH_SPEED_NUM_NONE;
 879         else
 880                 dev_link.link_speed = link_speed;
 881         priv->link_speed_capa = 0;
 882         if (edata.supported & SUPPORTED_Autoneg)
 883                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 884         if (edata.supported & (SUPPORTED_1000baseT_Full |
 885                                SUPPORTED_1000baseKX_Full))
 886                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 887         if (edata.supported & SUPPORTED_10000baseKR_Full)
 888                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 889         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 890                                SUPPORTED_40000baseCR4_Full |
 891                                SUPPORTED_40000baseSR4_Full |
 892                                SUPPORTED_40000baseLR4_Full))
 893                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 894         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 895                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 896         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 897                         ETH_LINK_SPEED_FIXED);
 898         if (((dev_link.link_speed && !dev_link.link_status) ||
 899              (!dev_link.link_speed && dev_link.link_status))) {
 900                 rte_errno = EAGAIN;
 901                 return -rte_errno;
 902         }
 903         *link = dev_link;
 904         return 0;
 905 }
 906
 907 /**
 908  * Retrieve physical link information (unlocked version using new ioctl).
 909  *
 910  * @param dev
 911  *   Pointer to Ethernet device structure.
 912  * @param[out] link
 913  *   Storage for current link status.
 914  *
 915  * @return
 916  *   0 on success, a negative errno value otherwise and rte_errno is set.
 917  */
 918 static int
 919 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
 920                              struct rte_eth_link *link)
 921
 922 {
 923         struct mlx5_priv *priv = dev->data->dev_private;
 924         struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
 925         struct ifreq ifr;
 926         struct rte_eth_link dev_link;
 927         struct rte_eth_dev *master = NULL;
 928         uint64_t sc;
 929         int ret;
 930
 931         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 932         if (ret) {
 933                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 934                         dev->data->port_id, strerror(rte_errno));
 935                 return ret;
 936         }
 937         dev_link = (struct rte_eth_link) {
 938                 .link_status = ((ifr.ifr_flags & IFF_UP) &&
 939                                 (ifr.ifr_flags & IFF_RUNNING)),
 940         };
 941         ifr = (struct ifreq) {
 942                 .ifr_data = (void *)&gcmd,
 943         };
 944         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 945         if (ret) {
 946                 if (ret == -ENOTSUP && priv->representor) {
 947                         /*
 948                          * For representors we can try to inherit link
 949                          * settings from the master device. Actually
 950                          * link settings do not make a lot of sense
 951                          * for representors due to missing physical
 952                          * link. The old kernel drivers supported
 953                          * emulated settings query for representors,
 954                          * the new ones do not, so we have to add
 955                          * this code for compatibility issues.
 956                          */
 957                         master = mlx5_find_master_dev(dev);
 958                         if (master) {
 959                                 ifr = (struct ifreq) {
 960                                         .ifr_data = (void *)&gcmd,
 961                                 };
 962                                 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
 963                         }
 964                 }
 965                 if (ret) {
 966                         DRV_LOG(DEBUG,
 967                                 "port %u ioctl(SIOCETHTOOL,"
 968                                 " ETHTOOL_GLINKSETTINGS) failed: %s",
 969                                 dev->data->port_id, strerror(rte_errno));
 970                         return ret;
 971                 }
 972
 973         }
 974         gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
 975
 976         alignas(struct ethtool_link_settings)
 977         uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
 978                      sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
 979         struct ethtool_link_settings *ecmd = (void *)data;
 980
 981         *ecmd = gcmd;
 982         ifr.ifr_data = (void *)ecmd;
 983         ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
 984         if (ret) {
 985                 DRV_LOG(DEBUG,
 986                         "port %u ioctl(SIOCETHTOOL,"
 987                         "ETHTOOL_GLINKSETTINGS) failed: %s",
 988                         dev->data->port_id, strerror(rte_errno));
 989                 return ret;
 990         }
 991         dev_link.link_speed = (ecmd->speed == UINT32_MAX) ? ETH_SPEED_NUM_NONE :
 992                                                             ecmd->speed;
 993         sc = ecmd->link_mode_masks[0] |
 994                 ((uint64_t)ecmd->link_mode_masks[1] << 32);
 995         priv->link_speed_capa = 0;
 996         if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
 997                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 998         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
 999                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
1000                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
1001         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
1002                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
1003                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
1004                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
1005         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
1006                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
1007                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
1008         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
1009                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
1010                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
1011                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
1012                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
1013         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
1014                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
1015                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
1016                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
1017                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
1018         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
1019                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
1020                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
1021                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
1022         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
1023                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
1024                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
1025         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
1026                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
1027                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
1028                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
1029                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
1030         dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
1031                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
1032         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
1033                                   ETH_LINK_SPEED_FIXED);
1034         if (((dev_link.link_speed && !dev_link.link_status) ||
1035              (!dev_link.link_speed && dev_link.link_status))) {
1036                 rte_errno = EAGAIN;
1037                 return -rte_errno;
1038         }
1039         *link = dev_link;
1040         return 0;
1041 }
1042
1043 /**
1044  * DPDK callback to retrieve physical link information.
1045  *
1046  * @param dev
1047  *   Pointer to Ethernet device structure.
1048  * @param wait_to_complete
1049  *   Wait for request completion.
1050  *
1051  * @return
1052  *   0 if link status was not updated, positive if it was, a negative errno
1053  *   value otherwise and rte_errno is set.
1054  */
1055 int
1056 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
1057 {
1058         int ret;
1059         struct rte_eth_link dev_link;
1060         time_t start_time = time(NULL);
1061         int retry = MLX5_GET_LINK_STATUS_RETRY_COUNT;
1062
1063         do {
1064                 ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
1065                 if (ret == -ENOTSUP)
1066                         ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
1067                 if (ret == 0)
1068                         break;
1069                 /* Handle wait to complete situation. */
1070                 if ((wait_to_complete || retry) && ret == -EAGAIN) {
1071                         if (abs((int)difftime(time(NULL), start_time)) <
1072                             MLX5_LINK_STATUS_TIMEOUT) {
1073                                 usleep(0);
1074                                 continue;
1075                         } else {
1076                                 rte_errno = EBUSY;
1077                                 return -rte_errno;
1078                         }
1079                 } else if (ret < 0) {
1080                         return ret;
1081                 }
1082         } while (wait_to_complete || retry-- > 0);
1083         ret = !!memcmp(&dev->data->dev_link, &dev_link,
1084                        sizeof(struct rte_eth_link));
1085         dev->data->dev_link = dev_link;
1086         return ret;
1087 }
1088
1089 /**
1090  * DPDK callback to change the MTU.
1091  *
1092  * @param dev
1093  *   Pointer to Ethernet device structure.
1094  * @param in_mtu
1095  *   New MTU.
1096  *
1097  * @return
1098  *   0 on success, a negative errno value otherwise and rte_errno is set.
1099  */
1100 int
1101 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1102 {
1103         struct mlx5_priv *priv = dev->data->dev_private;
1104         uint16_t kern_mtu = 0;
1105         int ret;
1106
1107         ret = mlx5_get_mtu(dev, &kern_mtu);
1108         if (ret)
1109                 return ret;
1110         /* Set kernel interface MTU first. */
1111         ret = mlx5_set_mtu(dev, mtu);
1112         if (ret)
1113                 return ret;
1114         ret = mlx5_get_mtu(dev, &kern_mtu);
1115         if (ret)
1116                 return ret;
1117         if (kern_mtu == mtu) {
1118                 priv->mtu = mtu;
1119                 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1120                         dev->data->port_id, mtu);
1121                 return 0;
1122         }
1123         rte_errno = EAGAIN;
1124         return -rte_errno;
1125 }
1126
1127 /**
1128  * DPDK callback to get flow control status.
1129  *
1130  * @param dev
1131  *   Pointer to Ethernet device structure.
1132  * @param[out] fc_conf
1133  *   Flow control output buffer.
1134  *
1135  * @return
1136  *   0 on success, a negative errno value otherwise and rte_errno is set.
1137  */
1138 int
1139 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1140 {
1141         struct ifreq ifr;
1142         struct ethtool_pauseparam ethpause = {
1143                 .cmd = ETHTOOL_GPAUSEPARAM
1144         };
1145         int ret;
1146
1147         ifr.ifr_data = (void *)&ethpause;
1148         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1149         if (ret) {
1150                 DRV_LOG(WARNING,
1151                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1152                         " %s",
1153                         dev->data->port_id, strerror(rte_errno));
1154                 return ret;
1155         }
1156         fc_conf->autoneg = ethpause.autoneg;
1157         if (ethpause.rx_pause && ethpause.tx_pause)
1158                 fc_conf->mode = RTE_FC_FULL;
1159         else if (ethpause.rx_pause)
1160                 fc_conf->mode = RTE_FC_RX_PAUSE;
1161         else if (ethpause.tx_pause)
1162                 fc_conf->mode = RTE_FC_TX_PAUSE;
1163         else
1164                 fc_conf->mode = RTE_FC_NONE;
1165         return 0;
1166 }
1167
1168 /**
1169  * DPDK callback to modify flow control parameters.
1170  *
1171  * @param dev
1172  *   Pointer to Ethernet device structure.
1173  * @param[in] fc_conf
1174  *   Flow control parameters.
1175  *
1176  * @return
1177  *   0 on success, a negative errno value otherwise and rte_errno is set.
1178  */
1179 int
1180 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1181 {
1182         struct ifreq ifr;
1183         struct ethtool_pauseparam ethpause = {
1184                 .cmd = ETHTOOL_SPAUSEPARAM
1185         };
1186         int ret;
1187
1188         ifr.ifr_data = (void *)&ethpause;
1189         ethpause.autoneg = fc_conf->autoneg;
1190         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1191             (fc_conf->mode & RTE_FC_RX_PAUSE))
1192                 ethpause.rx_pause = 1;
1193         else
1194                 ethpause.rx_pause = 0;
1195
1196         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1197             (fc_conf->mode & RTE_FC_TX_PAUSE))
1198                 ethpause.tx_pause = 1;
1199         else
1200                 ethpause.tx_pause = 0;
1201         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1202         if (ret) {
1203                 DRV_LOG(WARNING,
1204                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1205                         " failed: %s",
1206                         dev->data->port_id, strerror(rte_errno));
1207                 return ret;
1208         }
1209         return 0;
1210 }
1211
1212 /**
1213  * Get PCI information by sysfs device path.
1214  *
1215  * @param dev_path
1216  *   Pointer to device sysfs folder name.
1217  * @param[out] pci_addr
1218  *   PCI bus address output buffer.
1219  *
1220  * @return
1221  *   0 on success, a negative errno value otherwise and rte_errno is set.
1222  */
1223 int
1224 mlx5_dev_to_pci_addr(const char *dev_path,
1225                      struct rte_pci_addr *pci_addr)
1226 {
1227         FILE *file;
1228         char line[32];
1229         MKSTR(path, "%s/device/uevent", dev_path);
1230
1231         file = fopen(path, "rb");
1232         if (file == NULL) {
1233                 rte_errno = errno;
1234                 return -rte_errno;
1235         }
1236         while (fgets(line, sizeof(line), file) == line) {
1237                 size_t len = strlen(line);
1238                 int ret;
1239
1240                 /* Truncate long lines. */
1241                 if (len == (sizeof(line) - 1))
1242                         while (line[(len - 1)] != '\n') {
1243                                 ret = fgetc(file);
1244                                 if (ret == EOF)
1245                                         break;
1246                                 line[(len - 1)] = ret;
1247                         }
1248                 /* Extract information. */
1249                 if (sscanf(line,
1250                            "PCI_SLOT_NAME="
1251                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1252                            &pci_addr->domain,
1253                            &pci_addr->bus,
1254                            &pci_addr->devid,
1255                            &pci_addr->function) == 4) {
1256                         ret = 0;
1257                         break;
1258                 }
1259         }
1260         fclose(file);
1261         return 0;
1262 }
1263
1264 /**
1265  * Handle asynchronous removal event for entire multiport device.
1266  *
1267  * @param sh
1268  *   Infiniband device shared context.
1269  */
1270 static void
1271 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1272 {
1273         uint32_t i;
1274
1275         for (i = 0; i < sh->max_port; ++i) {
1276                 struct rte_eth_dev *dev;
1277
1278                 if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1279                         /*
1280                          * Or not existing port either no
1281                          * handler installed for this port.
1282                          */
1283                         continue;
1284                 }
1285                 dev = &rte_eth_devices[sh->port[i].ih_port_id];
1286                 assert(dev);
1287                 if (dev->data->dev_conf.intr_conf.rmv)
1288                         _rte_eth_dev_callback_process
1289                                 (dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1290         }
1291 }
1292
1293 /**
1294  * Handle shared asynchronous events the NIC (removal event
1295  * and link status change). Supports multiport IB device.
1296  *
1297  * @param cb_arg
1298  *   Callback argument.
1299  */
1300 void
1301 mlx5_dev_interrupt_handler(void *cb_arg)
1302 {
1303         struct mlx5_ibv_shared *sh = cb_arg;
1304         struct ibv_async_event event;
1305
1306         /* Read all message from the IB device and acknowledge them. */
1307         for (;;) {
1308                 struct rte_eth_dev *dev;
1309                 uint32_t tmp;
1310
1311                 if (mlx5_glue->get_async_event(sh->ctx, &event))
1312                         break;
1313                 /* Retrieve and check IB port index. */
1314                 tmp = (uint32_t)event.element.port_num;
1315                 if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1316                         /*
1317                          * The DEVICE_FATAL event is called once for
1318                          * entire device without port specifying.
1319                          * We should notify all existing ports.
1320                          */
1321                         mlx5_glue->ack_async_event(&event);
1322                         mlx5_dev_interrupt_device_fatal(sh);
1323                         continue;
1324                 }
1325                 assert(tmp && (tmp <= sh->max_port));
1326                 if (!tmp) {
1327                         /* Unsupported devive level event. */
1328                         mlx5_glue->ack_async_event(&event);
1329                         DRV_LOG(DEBUG,
1330                                 "unsupported common event (type %d)",
1331                                 event.event_type);
1332                         continue;
1333                 }
1334                 if (tmp > sh->max_port) {
1335                         /* Invalid IB port index. */
1336                         mlx5_glue->ack_async_event(&event);
1337                         DRV_LOG(DEBUG,
1338                                 "cannot handle an event (type %d)"
1339                                 "due to invalid IB port index (%u)",
1340                                 event.event_type, tmp);
1341                         continue;
1342                 }
1343                 if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1344                         /* No handler installed. */
1345                         mlx5_glue->ack_async_event(&event);
1346                         DRV_LOG(DEBUG,
1347                                 "cannot handle an event (type %d)"
1348                                 "due to no handler installed for port %u",
1349                                 event.event_type, tmp);
1350                         continue;
1351                 }
1352                 /* Retrieve ethernet device descriptor. */
1353                 tmp = sh->port[tmp - 1].ih_port_id;
1354                 dev = &rte_eth_devices[tmp];
1355                 assert(dev);
1356                 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1357                      event.event_type == IBV_EVENT_PORT_ERR) &&
1358                         dev->data->dev_conf.intr_conf.lsc) {
1359                         mlx5_glue->ack_async_event(&event);
1360                         if (mlx5_link_update(dev, 0) == -EAGAIN) {
1361                                 usleep(0);
1362                                 continue;
1363                         }
1364                         _rte_eth_dev_callback_process
1365                                 (dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1366                         continue;
1367                 }
1368                 DRV_LOG(DEBUG,
1369                         "port %u cannot handle an unknown event (type %d)",
1370                         dev->data->port_id, event.event_type);
1371                 mlx5_glue->ack_async_event(&event);
1372         }
1373 }
1374
1375 /*
1376  * Unregister callback handler safely. The handler may be active
1377  * while we are trying to unregister it, in this case code -EAGAIN
1378  * is returned by rte_intr_callback_unregister(). This routine checks
1379  * the return code and tries to unregister handler again.
1380  *
1381  * @param handle
1382  *   interrupt handle
1383  * @param cb_fn
1384  *   pointer to callback routine
1385  * @cb_arg
1386  *   opaque callback parameter
1387  */
1388 void
1389 mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
1390                               rte_intr_callback_fn cb_fn, void *cb_arg)
1391 {
1392         /*
1393          * Try to reduce timeout management overhead by not calling
1394          * the timer related routines on the first iteration. If the
1395          * unregistering succeeds on first call there will be no
1396          * timer calls at all.
1397          */
1398         uint64_t twait = 0;
1399         uint64_t start = 0;
1400
1401         do {
1402                 int ret;
1403
1404                 ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
1405                 if (ret >= 0)
1406                         return;
1407                 if (ret != -EAGAIN) {
1408                         DRV_LOG(INFO, "failed to unregister interrupt"
1409                                       " handler (error: %d)", ret);
1410                         assert(false);
1411                         return;
1412                 }
1413                 if (twait) {
1414                         struct timespec onems;
1415
1416                         /* Wait one millisecond and try again. */
1417                         onems.tv_sec = 0;
1418                         onems.tv_nsec = NS_PER_S / MS_PER_S;
1419                         nanosleep(&onems, 0);
1420                         /* Check whether one second elapsed. */
1421                         if ((rte_get_timer_cycles() - start) <= twait)
1422                                 continue;
1423                 } else {
1424                         /*
1425                          * We get the amount of timer ticks for one second.
1426                          * If this amount elapsed it means we spent one
1427                          * second in waiting. This branch is executed once
1428                          * on first iteration.
1429                          */
1430                         twait = rte_get_timer_hz();
1431                         assert(twait);
1432                 }
1433                 /*
1434                  * Timeout elapsed, show message (once a second) and retry.
1435                  * We have no other acceptable option here, if we ignore
1436                  * the unregistering return code the handler will not
1437                  * be unregistered, fd will be closed and we may get the
1438                  * crush. Hanging and messaging in the loop seems not to be
1439                  * the worst choice.
1440                  */
1441                 DRV_LOG(INFO, "Retrying to unregister interrupt handler");
1442                 start = rte_get_timer_cycles();
1443         } while (true);
1444 }
1445
1446 /**
1447  * Handle DEVX interrupts from the NIC.
1448  * This function is probably called from the DPDK host thread.
1449  *
1450  * @param cb_arg
1451  *   Callback argument.
1452  */
1453 void
1454 mlx5_dev_interrupt_handler_devx(void *cb_arg)
1455 {
1456 #ifndef HAVE_IBV_DEVX_ASYNC
1457         (void)cb_arg;
1458         return;
1459 #else
1460         struct mlx5_ibv_shared *sh = cb_arg;
1461         union {
1462                 struct mlx5dv_devx_async_cmd_hdr cmd_resp;
1463                 uint8_t buf[MLX5_ST_SZ_BYTES(query_flow_counter_out) +
1464                             MLX5_ST_SZ_BYTES(traffic_counter) +
1465                             sizeof(struct mlx5dv_devx_async_cmd_hdr)];
1466         } out;
1467         uint8_t *buf = out.buf + sizeof(out.cmd_resp);
1468
1469         while (!mlx5_glue->devx_get_async_cmd_comp(sh->devx_comp,
1470                                                    &out.cmd_resp,
1471                                                    sizeof(out.buf)))
1472                 mlx5_flow_async_pool_query_handle
1473                         (sh, (uint64_t)out.cmd_resp.wr_id,
1474                          mlx5_devx_get_out_command_status(buf));
1475 #endif /* HAVE_IBV_DEVX_ASYNC */
1476 }
1477
1478 /**
1479  * Uninstall shared asynchronous device events handler.
1480  * This function is implemented to support event sharing
1481  * between multiple ports of single IB device.
1482  *
1483  * @param dev
1484  *   Pointer to Ethernet device.
1485  */
1486 static void
1487 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1488 {
1489         struct mlx5_priv *priv = dev->data->dev_private;
1490         struct mlx5_ibv_shared *sh = priv->sh;
1491
1492         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1493                 return;
1494         pthread_mutex_lock(&sh->intr_mutex);
1495         assert(priv->ibv_port);
1496         assert(priv->ibv_port <= sh->max_port);
1497         assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1498         if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1499                 goto exit;
1500         assert(sh->port[priv->ibv_port - 1].ih_port_id ==
1501                                         (uint32_t)dev->data->port_id);
1502         assert(sh->intr_cnt);
1503         sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1504         if (!sh->intr_cnt || --sh->intr_cnt)
1505                 goto exit;
1506         mlx5_intr_callback_unregister(&sh->intr_handle,
1507                                      mlx5_dev_interrupt_handler, sh);
1508         sh->intr_handle.fd = 0;
1509         sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1510 exit:
1511         pthread_mutex_unlock(&sh->intr_mutex);
1512 }
1513
1514 /**
1515  * Uninstall devx shared asynchronous device events handler.
1516  * This function is implemeted to support event sharing
1517  * between multiple ports of single IB device.
1518  *
1519  * @param dev
1520  *   Pointer to Ethernet device.
1521  */
1522 static void
1523 mlx5_dev_shared_handler_devx_uninstall(struct rte_eth_dev *dev)
1524 {
1525         struct mlx5_priv *priv = dev->data->dev_private;
1526         struct mlx5_ibv_shared *sh = priv->sh;
1527
1528         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1529                 return;
1530         pthread_mutex_lock(&sh->intr_mutex);
1531         assert(priv->ibv_port);
1532         assert(priv->ibv_port <= sh->max_port);
1533         assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1534         if (sh->port[priv->ibv_port - 1].devx_ih_port_id >= RTE_MAX_ETHPORTS)
1535                 goto exit;
1536         assert(sh->port[priv->ibv_port - 1].devx_ih_port_id ==
1537                                         (uint32_t)dev->data->port_id);
1538         sh->port[priv->ibv_port - 1].devx_ih_port_id = RTE_MAX_ETHPORTS;
1539         if (!sh->devx_intr_cnt || --sh->devx_intr_cnt)
1540                 goto exit;
1541         if (sh->intr_handle_devx.fd) {
1542                 rte_intr_callback_unregister(&sh->intr_handle_devx,
1543                                              mlx5_dev_interrupt_handler_devx,
1544                                              sh);
1545                 sh->intr_handle_devx.fd = 0;
1546                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_UNKNOWN;
1547         }
1548         if (sh->devx_comp) {
1549                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1550                 sh->devx_comp = NULL;
1551         }
1552 exit:
1553         pthread_mutex_unlock(&sh->intr_mutex);
1554 }
1555
1556 /**
1557  * Install shared asynchronous device events handler.
1558  * This function is implemented to support event sharing
1559  * between multiple ports of single IB device.
1560  *
1561  * @param dev
1562  *   Pointer to Ethernet device.
1563  */
1564 static void
1565 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1566 {
1567         struct mlx5_priv *priv = dev->data->dev_private;
1568         struct mlx5_ibv_shared *sh = priv->sh;
1569         int ret;
1570         int flags;
1571
1572         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1573                 return;
1574         pthread_mutex_lock(&sh->intr_mutex);
1575         assert(priv->ibv_port);
1576         assert(priv->ibv_port <= sh->max_port);
1577         assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1578         if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1579                 /* The handler is already installed for this port. */
1580                 assert(sh->intr_cnt);
1581                 goto exit;
1582         }
1583         if (sh->intr_cnt) {
1584                 sh->port[priv->ibv_port - 1].ih_port_id =
1585                                                 (uint32_t)dev->data->port_id;
1586                 sh->intr_cnt++;
1587                 goto exit;
1588         }
1589         /* No shared handler installed. */
1590         assert(sh->ctx->async_fd > 0);
1591         flags = fcntl(sh->ctx->async_fd, F_GETFL);
1592         ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1593         if (ret) {
1594                 DRV_LOG(INFO, "failed to change file descriptor async event"
1595                         " queue");
1596                 /* Indicate there will be no interrupts. */
1597                 dev->data->dev_conf.intr_conf.lsc = 0;
1598                 dev->data->dev_conf.intr_conf.rmv = 0;
1599         } else {
1600                 sh->intr_handle.fd = sh->ctx->async_fd;
1601                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1602                 rte_intr_callback_register(&sh->intr_handle,
1603                                            mlx5_dev_interrupt_handler, sh);
1604                 sh->intr_cnt++;
1605                 sh->port[priv->ibv_port - 1].ih_port_id =
1606                                                 (uint32_t)dev->data->port_id;
1607         }
1608 exit:
1609         pthread_mutex_unlock(&sh->intr_mutex);
1610 }
1611
1612 /**
1613  * Install devx shared asyncronous device events handler.
1614  * This function is implemeted to support event sharing
1615  * between multiple ports of single IB device.
1616  *
1617  * @param dev
1618  *   Pointer to Ethernet device.
1619  */
1620 static void
1621 mlx5_dev_shared_handler_devx_install(struct rte_eth_dev *dev)
1622 {
1623         struct mlx5_priv *priv = dev->data->dev_private;
1624         struct mlx5_ibv_shared *sh = priv->sh;
1625
1626         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1627                 return;
1628         pthread_mutex_lock(&sh->intr_mutex);
1629         assert(priv->ibv_port);
1630         assert(priv->ibv_port <= sh->max_port);
1631         assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1632         if (sh->port[priv->ibv_port - 1].devx_ih_port_id < RTE_MAX_ETHPORTS) {
1633                 /* The handler is already installed for this port. */
1634                 assert(sh->devx_intr_cnt);
1635                 goto exit;
1636         }
1637         if (sh->devx_intr_cnt) {
1638                 sh->devx_intr_cnt++;
1639                 sh->port[priv->ibv_port - 1].devx_ih_port_id =
1640                                         (uint32_t)dev->data->port_id;
1641                 goto exit;
1642         }
1643         if (priv->config.devx) {
1644 #ifndef HAVE_IBV_DEVX_ASYNC
1645                 goto exit;
1646 #else
1647                 sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1648                 if (sh->devx_comp) {
1649                         int flags = fcntl(sh->devx_comp->fd, F_GETFL);
1650                         int ret = fcntl(sh->devx_comp->fd, F_SETFL,
1651                                     flags | O_NONBLOCK);
1652
1653                         if (ret) {
1654                                 DRV_LOG(INFO, "failed to change file descriptor"
1655                                         " devx async event queue");
1656                         } else {
1657                                 sh->intr_handle_devx.fd = sh->devx_comp->fd;
1658                                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1659                                 rte_intr_callback_register
1660                                         (&sh->intr_handle_devx,
1661                                          mlx5_dev_interrupt_handler_devx, sh);
1662                                 sh->devx_intr_cnt++;
1663                                 sh->port[priv->ibv_port - 1].devx_ih_port_id =
1664                                                 (uint32_t)dev->data->port_id;
1665                         }
1666                 }
1667 #endif /* HAVE_IBV_DEVX_ASYNC */
1668         }
1669 exit:
1670         pthread_mutex_unlock(&sh->intr_mutex);
1671 }
1672
1673 /**
1674  * Uninstall interrupt handler.
1675  *
1676  * @param dev
1677  *   Pointer to Ethernet device.
1678  */
1679 void
1680 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1681 {
1682         mlx5_dev_shared_handler_uninstall(dev);
1683 }
1684
1685 /**
1686  * Install interrupt handler.
1687  *
1688  * @param dev
1689  *   Pointer to Ethernet device.
1690  */
1691 void
1692 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1693 {
1694         mlx5_dev_shared_handler_install(dev);
1695 }
1696
1697 /**
1698  * Devx uninstall interrupt handler.
1699  *
1700  * @param dev
1701  *   Pointer to Ethernet device.
1702  */
1703 void
1704 mlx5_dev_interrupt_handler_devx_uninstall(struct rte_eth_dev *dev)
1705 {
1706         mlx5_dev_shared_handler_devx_uninstall(dev);
1707 }
1708
1709 /**
1710  * Devx install interrupt handler.
1711  *
1712  * @param dev
1713  *   Pointer to Ethernet device.
1714  */
1715 void
1716 mlx5_dev_interrupt_handler_devx_install(struct rte_eth_dev *dev)
1717 {
1718         mlx5_dev_shared_handler_devx_install(dev);
1719 }
1720
1721 /**
1722  * DPDK callback to bring the link DOWN.
1723  *
1724  * @param dev
1725  *   Pointer to Ethernet device structure.
1726  *
1727  * @return
1728  *   0 on success, a negative errno value otherwise and rte_errno is set.
1729  */
1730 int
1731 mlx5_set_link_down(struct rte_eth_dev *dev)
1732 {
1733         return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1734 }
1735
1736 /**
1737  * DPDK callback to bring the link UP.
1738  *
1739  * @param dev
1740  *   Pointer to Ethernet device structure.
1741  *
1742  * @return
1743  *   0 on success, a negative errno value otherwise and rte_errno is set.
1744  */
1745 int
1746 mlx5_set_link_up(struct rte_eth_dev *dev)
1747 {
1748         return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1749 }
1750
1751 /**
1752  * Configure the RX function to use.
1753  *
1754  * @param dev
1755  *   Pointer to private data structure.
1756  *
1757  * @return
1758  *   Pointer to selected Rx burst function.
1759  */
1760 eth_rx_burst_t
1761 mlx5_select_rx_function(struct rte_eth_dev *dev)
1762 {
1763         eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1764
1765         assert(dev != NULL);
1766         if (mlx5_check_vec_rx_support(dev) > 0) {
1767                 rx_pkt_burst = mlx5_rx_burst_vec;
1768                 DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1769                         dev->data->port_id);
1770         } else if (mlx5_mprq_enabled(dev)) {
1771                 rx_pkt_burst = mlx5_rx_burst_mprq;
1772         }
1773         return rx_pkt_burst;
1774 }
1775
1776 /**
1777  * Check if mlx5 device was removed.
1778  *
1779  * @param dev
1780  *   Pointer to Ethernet device structure.
1781  *
1782  * @return
1783  *   1 when device is removed, otherwise 0.
1784  */
1785 int
1786 mlx5_is_removed(struct rte_eth_dev *dev)
1787 {
1788         struct ibv_device_attr device_attr;
1789         struct mlx5_priv *priv = dev->data->dev_private;
1790
1791         if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1792                 return 1;
1793         return 0;
1794 }
1795
1796 /**
1797  * Get the E-Switch parameters by port id.
1798  *
1799  * @param[in] port
1800  *   Device port id.
1801  * @param[in] valid
1802  *   Device port id is valid, skip check. This flag is useful
1803  *   when trials are performed from probing and device is not
1804  *   flagged as valid yet (in attaching process).
1805  * @param[out] es_domain_id
1806  *   E-Switch domain id.
1807  * @param[out] es_port_id
1808  *   The port id of the port in the E-Switch.
1809  *
1810  * @return
1811  *   pointer to device private data structure containing data needed
1812  *   on success, NULL otherwise and rte_errno is set.
1813  */
1814 struct mlx5_priv *
1815 mlx5_port_to_eswitch_info(uint16_t port, bool valid)
1816 {
1817         struct rte_eth_dev *dev;
1818         struct mlx5_priv *priv;
1819
1820         if (port >= RTE_MAX_ETHPORTS) {
1821                 rte_errno = EINVAL;
1822                 return NULL;
1823         }
1824         if (!valid && !rte_eth_dev_is_valid_port(port)) {
1825                 rte_errno = ENODEV;
1826                 return NULL;
1827         }
1828         dev = &rte_eth_devices[port];
1829         priv = dev->data->dev_private;
1830         if (!(priv->representor || priv->master)) {
1831                 rte_errno = EINVAL;
1832                 return NULL;
1833         }
1834         return priv;
1835 }
1836
1837 /**
1838  * Get the E-Switch parameters by device instance.
1839  *
1840  * @param[in] port
1841  *   Device port id.
1842  * @param[out] es_domain_id
1843  *   E-Switch domain id.
1844  * @param[out] es_port_id
1845  *   The port id of the port in the E-Switch.
1846  *
1847  * @return
1848  *   pointer to device private data structure containing data needed
1849  *   on success, NULL otherwise and rte_errno is set.
1850  */
1851 struct mlx5_priv *
1852 mlx5_dev_to_eswitch_info(struct rte_eth_dev *dev)
1853 {
1854         struct mlx5_priv *priv;
1855
1856         priv = dev->data->dev_private;
1857         if (!(priv->representor || priv->master)) {
1858                 rte_errno = EINVAL;
1859                 return NULL;
1860         }
1861         return priv;
1862 }
1863
1864 /**
1865  * Get switch information associated with network interface.
1866  *
1867  * @param ifindex
1868  *   Network interface index.
1869  * @param[out] info
1870  *   Switch information object, populated in case of success.
1871  *
1872  * @return
1873  *   0 on success, a negative errno value otherwise and rte_errno is set.
1874  */
1875 int
1876 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1877 {
1878         char ifname[IF_NAMESIZE];
1879         char port_name[IF_NAMESIZE];
1880         FILE *file;
1881         struct mlx5_switch_info data = {
1882                 .master = 0,
1883                 .representor = 0,
1884                 .name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1885                 .port_name = 0,
1886                 .switch_id = 0,
1887         };
1888         DIR *dir;
1889         bool port_switch_id_set = false;
1890         bool device_dir = false;
1891         char c;
1892         int ret;
1893
1894         if (!if_indextoname(ifindex, ifname)) {
1895                 rte_errno = errno;
1896                 return -rte_errno;
1897         }
1898
1899         MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1900               ifname);
1901         MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1902               ifname);
1903         MKSTR(pci_device, "/sys/class/net/%s/device",
1904               ifname);
1905
1906         file = fopen(phys_port_name, "rb");
1907         if (file != NULL) {
1908                 ret = fscanf(file, "%s", port_name);
1909                 fclose(file);
1910                 if (ret == 1)
1911                         mlx5_translate_port_name(port_name, &data);
1912         }
1913         file = fopen(phys_switch_id, "rb");
1914         if (file == NULL) {
1915                 rte_errno = errno;
1916                 return -rte_errno;
1917         }
1918         port_switch_id_set =
1919                 fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1920                 c == '\n';
1921         fclose(file);
1922         dir = opendir(pci_device);
1923         if (dir != NULL) {
1924                 closedir(dir);
1925                 device_dir = true;
1926         }
1927         if (port_switch_id_set) {
1928                 /* We have some E-Switch configuration. */
1929                 mlx5_sysfs_check_switch_info(device_dir, &data);
1930         }
1931         *info = data;
1932         assert(!(data.master && data.representor));
1933         if (data.master && data.representor) {
1934                 DRV_LOG(ERR, "ifindex %u device is recognized as master"
1935                              " and as representor", ifindex);
1936                 rte_errno = ENODEV;
1937                 return -rte_errno;
1938         }
1939         return 0;
1940 }
1941
1942 /**
1943  * Analyze gathered port parameters via Netlink to recognize master
1944  * and representor devices for E-Switch configuration.
1945  *
1946  * @param[in] num_vf_set
1947  *   flag of presence of number of VFs port attribute.
1948  * @param[inout] switch_info
1949  *   Port information, including port name as a number and port name
1950  *   type if recognized
1951  *
1952  * @return
1953  *   master and representor flags are set in switch_info according to
1954  *   recognized parameters (if any).
1955  */
1956 void
1957 mlx5_nl_check_switch_info(bool num_vf_set,
1958                           struct mlx5_switch_info *switch_info)
1959 {
1960         switch (switch_info->name_type) {
1961         case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1962                 /*
1963                  * Name is not recognized, assume the master,
1964                  * check the number of VFs key presence.
1965                  */
1966                 switch_info->master = num_vf_set;
1967                 break;
1968         case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1969                 /*
1970                  * Name is not set, this assumes the legacy naming
1971                  * schema for master, just check if there is a
1972                  * number of VFs key.
1973                  */
1974                 switch_info->master = num_vf_set;
1975                 break;
1976         case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1977                 /* New uplink naming schema recognized. */
1978                 switch_info->master = 1;
1979                 break;
1980         case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1981                 /* Legacy representors naming schema. */
1982                 switch_info->representor = !num_vf_set;
1983                 break;
1984         case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1985                 /* New representors naming schema. */
1986                 switch_info->representor = 1;
1987                 break;
1988         }
1989 }
1990
1991 /**
1992  * Analyze gathered port parameters via sysfs to recognize master
1993  * and representor devices for E-Switch configuration.
1994  *
1995  * @param[in] device_dir
1996  *   flag of presence of "device" directory under port device key.
1997  * @param[inout] switch_info
1998  *   Port information, including port name as a number and port name
1999  *   type if recognized
2000  *
2001  * @return
2002  *   master and representor flags are set in switch_info according to
2003  *   recognized parameters (if any).
2004  */
2005 void
2006 mlx5_sysfs_check_switch_info(bool device_dir,
2007                              struct mlx5_switch_info *switch_info)
2008 {
2009         switch (switch_info->name_type) {
2010         case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
2011                 /*
2012                  * Name is not recognized, assume the master,
2013                  * check the device directory presence.
2014                  */
2015                 switch_info->master = device_dir;
2016                 break;
2017         case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
2018                 /*
2019                  * Name is not set, this assumes the legacy naming
2020                  * schema for master, just check if there is
2021                  * a device directory.
2022                  */
2023                 switch_info->master = device_dir;
2024                 break;
2025         case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2026                 /* New uplink naming schema recognized. */
2027                 switch_info->master = 1;
2028                 break;
2029         case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
2030                 /* Legacy representors naming schema. */
2031                 switch_info->representor = !device_dir;
2032                 break;
2033         case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2034                 /* New representors naming schema. */
2035                 switch_info->representor = 1;
2036                 break;
2037         }
2038 }
2039
2040 /**
2041  * Extract port name, as a number, from sysfs or netlink information.
2042  *
2043  * @param[in] port_name_in
2044  *   String representing the port name.
2045  * @param[out] port_info_out
2046  *   Port information, including port name as a number and port name
2047  *   type if recognized
2048  *
2049  * @return
2050  *   port_name field set according to recognized name format.
2051  */
2052 void
2053 mlx5_translate_port_name(const char *port_name_in,
2054                          struct mlx5_switch_info *port_info_out)
2055 {
2056         char pf_c1, pf_c2, vf_c1, vf_c2;
2057         char *end;
2058         int sc_items;
2059
2060         /*
2061          * Check for port-name as a string of the form pf0vf0
2062          * (support kernel ver >= 5.0 or OFED ver >= 4.6).
2063          */
2064         sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
2065                           &pf_c1, &pf_c2, &port_info_out->pf_num,
2066                           &vf_c1, &vf_c2, &port_info_out->port_name);
2067         if (sc_items == 6 &&
2068             pf_c1 == 'p' && pf_c2 == 'f' &&
2069             vf_c1 == 'v' && vf_c2 == 'f') {
2070                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
2071                 return;
2072         }
2073         /*
2074          * Check for port-name as a string of the form p0
2075          * (support kernel ver >= 5.0, or OFED ver >= 4.6).
2076          */
2077         sc_items = sscanf(port_name_in, "%c%d",
2078                           &pf_c1, &port_info_out->port_name);
2079         if (sc_items == 2 && pf_c1 == 'p') {
2080                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
2081                 return;
2082         }
2083         /* Check for port-name as a number (support kernel ver < 5.0 */
2084         errno = 0;
2085         port_info_out->port_name = strtol(port_name_in, &end, 0);
2086         if (!errno &&
2087             (size_t)(end - port_name_in) == strlen(port_name_in)) {
2088                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
2089                 return;
2090         }
2091         port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
2092         return;
2093 }
2094
2095 /**
2096  * DPDK callback to retrieve plug-in module EEPROM information (type and size).
2097  *
2098  * @param dev
2099  *   Pointer to Ethernet device structure.
2100  * @param[out] modinfo
2101  *   Storage for plug-in module EEPROM information.
2102  *
2103  * @return
2104  *   0 on success, a negative errno value otherwise and rte_errno is set.
2105  */
2106 int
2107 mlx5_get_module_info(struct rte_eth_dev *dev,
2108                      struct rte_eth_dev_module_info *modinfo)
2109 {
2110         struct ethtool_modinfo info = {
2111                 .cmd = ETHTOOL_GMODULEINFO,
2112         };
2113         struct ifreq ifr = (struct ifreq) {
2114                 .ifr_data = (void *)&info,
2115         };
2116         int ret = 0;
2117
2118         if (!dev || !modinfo) {
2119                 DRV_LOG(WARNING, "missing argument, cannot get module info");
2120                 rte_errno = EINVAL;
2121                 return -rte_errno;
2122         }
2123         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2124         if (ret) {
2125                 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2126                         dev->data->port_id, strerror(rte_errno));
2127                 return ret;
2128         }
2129         modinfo->type = info.type;
2130         modinfo->eeprom_len = info.eeprom_len;
2131         return ret;
2132 }
2133
2134 /**
2135  * DPDK callback to retrieve plug-in module EEPROM data.
2136  *
2137  * @param dev
2138  *   Pointer to Ethernet device structure.
2139  * @param[out] info
2140  *   Storage for plug-in module EEPROM data.
2141  *
2142  * @return
2143  *   0 on success, a negative errno value otherwise and rte_errno is set.
2144  */
2145 int mlx5_get_module_eeprom(struct rte_eth_dev *dev,
2146                            struct rte_dev_eeprom_info *info)
2147 {
2148         struct ethtool_eeprom *eeprom;
2149         struct ifreq ifr;
2150         int ret = 0;
2151
2152         if (!dev || !info) {
2153                 DRV_LOG(WARNING, "missing argument, cannot get module eeprom");
2154                 rte_errno = EINVAL;
2155                 return -rte_errno;
2156         }
2157         eeprom = rte_calloc(__func__, 1,
2158                             (sizeof(struct ethtool_eeprom) + info->length), 0);
2159         if (!eeprom) {
2160                 DRV_LOG(WARNING, "port %u cannot allocate memory for "
2161                         "eeprom data", dev->data->port_id);
2162                 rte_errno = ENOMEM;
2163                 return -rte_errno;
2164         }
2165         eeprom->cmd = ETHTOOL_GMODULEEEPROM;
2166         eeprom->offset = info->offset;
2167         eeprom->len = info->length;
2168         ifr = (struct ifreq) {
2169                 .ifr_data = (void *)eeprom,
2170         };
2171         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2172         if (ret)
2173                 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2174                         dev->data->port_id, strerror(rte_errno));
2175         else
2176                 rte_memcpy(info->data, eeprom->data, info->length);
2177         rte_free(eeprom);
2178         return ret;
2179 }
2180
2181 /**
2182  * DPDK callback to retrieve hairpin capabilities.
2183  *
2184  * @param dev
2185  *   Pointer to Ethernet device structure.
2186  * @param[out] cap
2187  *   Storage for hairpin capability data.
2188  *
2189  * @return
2190  *   0 on success, a negative errno value otherwise and rte_errno is set.
2191  */
2192 int mlx5_hairpin_cap_get(struct rte_eth_dev *dev,
2193                          struct rte_eth_hairpin_cap *cap)
2194 {
2195         struct mlx5_priv *priv = dev->data->dev_private;
2196
2197         if (priv->sh->devx == 0) {
2198                 rte_errno = ENOTSUP;
2199                 return -rte_errno;
2200         }
2201         cap->max_nb_queues = UINT16_MAX;
2202         cap->max_rx_2_tx = 1;
2203         cap->max_tx_2_rx = 1;
2204         cap->max_nb_desc = 8192;
2205         return 0;
2206 }