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         unsigned int i;
 387         unsigned int j;
 388         unsigned int reta_idx_n;
 389         const uint8_t use_app_rss_key =
 390                 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
 391         int ret = 0;
 392
 393         if (use_app_rss_key &&
 394             (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
 395              MLX5_RSS_HASH_KEY_LEN)) {
 396                 DRV_LOG(ERR, "port %u RSS key len must be %s Bytes long",
 397                         dev->data->port_id, RTE_STR(MLX5_RSS_HASH_KEY_LEN));
 398                 rte_errno = EINVAL;
 399                 return -rte_errno;
 400         }
 401         priv->rss_conf.rss_key =
 402                 rte_realloc(priv->rss_conf.rss_key,
 403                             MLX5_RSS_HASH_KEY_LEN, 0);
 404         if (!priv->rss_conf.rss_key) {
 405                 DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
 406                         dev->data->port_id, rxqs_n);
 407                 rte_errno = ENOMEM;
 408                 return -rte_errno;
 409         }
 410         memcpy(priv->rss_conf.rss_key,
 411                use_app_rss_key ?
 412                dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
 413                rss_hash_default_key,
 414                MLX5_RSS_HASH_KEY_LEN);
 415         priv->rss_conf.rss_key_len = MLX5_RSS_HASH_KEY_LEN;
 416         priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
 417         priv->rxqs = (void *)dev->data->rx_queues;
 418         priv->txqs = (void *)dev->data->tx_queues;
 419         if (txqs_n != priv->txqs_n) {
 420                 DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
 421                         dev->data->port_id, priv->txqs_n, txqs_n);
 422                 priv->txqs_n = txqs_n;
 423         }
 424         if (rxqs_n > priv->config.ind_table_max_size) {
 425                 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
 426                         dev->data->port_id, rxqs_n);
 427                 rte_errno = EINVAL;
 428                 return -rte_errno;
 429         }
 430         if (rxqs_n != priv->rxqs_n) {
 431                 DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
 432                         dev->data->port_id, priv->rxqs_n, rxqs_n);
 433                 priv->rxqs_n = rxqs_n;
 434                 /*
 435                  * If the requested number of RX queues is not a power of two,
 436                  * use the maximum indirection table size for better balancing.
 437                  * The result is always rounded to the next power of two.
 438                  */
 439                 reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
 440                                              priv->config.ind_table_max_size :
 441                                              rxqs_n));
 442                 ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
 443                 if (ret)
 444                         return ret;
 445                 /*
 446                  * When the number of RX queues is not a power of two,
 447                  * the remaining table entries are padded with reused WQs
 448                  * and hashes are not spread uniformly.
 449                  */
 450                 for (i = 0, j = 0; (i != reta_idx_n); ++i) {
 451                         (*priv->reta_idx)[i] = j;
 452                         if (++j == rxqs_n)
 453                                 j = 0;
 454                 }
 455         }
 456         ret = mlx5_proc_priv_init(dev);
 457         if (ret)
 458                 return ret;
 459         return 0;
 460 }
 461
 462 /**
 463  * Sets default tuning parameters.
 464  *
 465  * @param dev
 466  *   Pointer to Ethernet device.
 467  * @param[out] info
 468  *   Info structure output buffer.
 469  */
 470 static void
 471 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 472 {
 473         struct mlx5_priv *priv = dev->data->dev_private;
 474
 475         /* Minimum CPU utilization. */
 476         info->default_rxportconf.ring_size = 256;
 477         info->default_txportconf.ring_size = 256;
 478         info->default_rxportconf.burst_size = 64;
 479         info->default_txportconf.burst_size = 64;
 480         if (priv->link_speed_capa & ETH_LINK_SPEED_100G) {
 481                 info->default_rxportconf.nb_queues = 16;
 482                 info->default_txportconf.nb_queues = 16;
 483                 if (dev->data->nb_rx_queues > 2 ||
 484                     dev->data->nb_tx_queues > 2) {
 485                         /* Max Throughput. */
 486                         info->default_rxportconf.ring_size = 2048;
 487                         info->default_txportconf.ring_size = 2048;
 488                 }
 489         } else {
 490                 info->default_rxportconf.nb_queues = 8;
 491                 info->default_txportconf.nb_queues = 8;
 492                 if (dev->data->nb_rx_queues > 2 ||
 493                     dev->data->nb_tx_queues > 2) {
 494                         /* Max Throughput. */
 495                         info->default_rxportconf.ring_size = 4096;
 496                         info->default_txportconf.ring_size = 4096;
 497                 }
 498         }
 499 }
 500
 501 /**
 502  * Sets tx mbuf limiting parameters.
 503  *
 504  * @param dev
 505  *   Pointer to Ethernet device.
 506  * @param[out] info
 507  *   Info structure output buffer.
 508  */
 509 static void
 510 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 511 {
 512         struct mlx5_priv *priv = dev->data->dev_private;
 513         struct mlx5_dev_config *config = &priv->config;
 514         unsigned int inlen;
 515         uint16_t nb_max;
 516
 517         inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
 518                 MLX5_SEND_DEF_INLINE_LEN :
 519                 (unsigned int)config->txq_inline_max;
 520         assert(config->txq_inline_min >= 0);
 521         inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
 522         inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
 523                                MLX5_ESEG_MIN_INLINE_SIZE -
 524                                MLX5_WQE_CSEG_SIZE -
 525                                MLX5_WQE_ESEG_SIZE -
 526                                MLX5_WQE_DSEG_SIZE * 2);
 527         nb_max = (MLX5_WQE_SIZE_MAX +
 528                   MLX5_ESEG_MIN_INLINE_SIZE -
 529                   MLX5_WQE_CSEG_SIZE -
 530                   MLX5_WQE_ESEG_SIZE -
 531                   MLX5_WQE_DSEG_SIZE -
 532                   inlen) / MLX5_WSEG_SIZE;
 533         info->tx_desc_lim.nb_seg_max = nb_max;
 534         info->tx_desc_lim.nb_mtu_seg_max = nb_max;
 535 }
 536
 537 /**
 538  * DPDK callback to get information about the device.
 539  *
 540  * @param dev
 541  *   Pointer to Ethernet device structure.
 542  * @param[out] info
 543  *   Info structure output buffer.
 544  */
 545 int
 546 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 547 {
 548         struct mlx5_priv *priv = dev->data->dev_private;
 549         struct mlx5_dev_config *config = &priv->config;
 550         unsigned int max;
 551
 552         /* FIXME: we should ask the device for these values. */
 553         info->min_rx_bufsize = 32;
 554         info->max_rx_pktlen = 65536;
 555         /*
 556          * Since we need one CQ per QP, the limit is the minimum number
 557          * between the two values.
 558          */
 559         max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
 560                       priv->sh->device_attr.orig_attr.max_qp);
 561         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
 562         if (max >= 65535)
 563                 max = 65535;
 564         info->max_rx_queues = max;
 565         info->max_tx_queues = max;
 566         info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
 567         info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
 568         info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
 569                                  info->rx_queue_offload_capa);
 570         info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
 571         info->if_index = mlx5_ifindex(dev);
 572         info->reta_size = priv->reta_idx_n ?
 573                 priv->reta_idx_n : config->ind_table_max_size;
 574         info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
 575         info->speed_capa = priv->link_speed_capa;
 576         info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
 577         mlx5_set_default_params(dev, info);
 578         mlx5_set_txlimit_params(dev, info);
 579         info->switch_info.name = dev->data->name;
 580         info->switch_info.domain_id = priv->domain_id;
 581         info->switch_info.port_id = priv->representor_id;
 582         if (priv->representor) {
 583                 uint16_t port_id;
 584
 585                 if (priv->pf_bond >= 0) {
 586                         /*
 587                          * Switch port ID is opaque value with driver defined
 588                          * format. Push the PF index in bonding configurations
 589                          * in upper four bits of port ID. If we get too many
 590                          * representors (more than 4K) or PFs (more than 15)
 591                          * this approach must be reconsidered.
 592                          */
 593                         if ((info->switch_info.port_id >>
 594                                 MLX5_PORT_ID_BONDING_PF_SHIFT) ||
 595                             priv->pf_bond > MLX5_PORT_ID_BONDING_PF_MASK) {
 596                                 DRV_LOG(ERR, "can't update switch port ID"
 597                                              " for bonding device");
 598                                 assert(false);
 599                                 return -ENODEV;
 600                         }
 601                         info->switch_info.port_id |=
 602                                 priv->pf_bond << MLX5_PORT_ID_BONDING_PF_SHIFT;
 603                 }
 604                 MLX5_ETH_FOREACH_DEV(port_id) {
 605                         struct mlx5_priv *opriv =
 606                                 rte_eth_devices[port_id].data->dev_private;
 607
 608                         if (!opriv ||
 609                             opriv->representor ||
 610                             opriv->sh != priv->sh ||
 611                             opriv->domain_id != priv->domain_id)
 612                                 continue;
 613                         /*
 614                          * Override switch name with that of the master
 615                          * device.
 616                          */
 617                         info->switch_info.name = opriv->dev_data->name;
 618                         break;
 619                 }
 620         }
 621         return 0;
 622 }
 623
 624 /**
 625  * Get device current raw clock counter
 626  *
 627  * @param dev
 628  *   Pointer to Ethernet device structure.
 629  * @param[out] time
 630  *   Current raw clock counter of the device.
 631  *
 632  * @return
 633  *   0 if the clock has correctly been read
 634  *   The value of errno in case of error
 635  */
 636 int
 637 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
 638 {
 639         struct mlx5_priv *priv = dev->data->dev_private;
 640         struct ibv_context *ctx = priv->sh->ctx;
 641         struct ibv_values_ex values;
 642         int err = 0;
 643
 644         values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
 645         err = mlx5_glue->query_rt_values_ex(ctx, &values);
 646         if (err != 0) {
 647                 DRV_LOG(WARNING, "Could not query the clock !");
 648                 return err;
 649         }
 650         *clock = values.raw_clock.tv_nsec;
 651         return 0;
 652 }
 653
 654 /**
 655  * Get firmware version of a device.
 656  *
 657  * @param dev
 658  *   Ethernet device port.
 659  * @param fw_ver
 660  *   String output allocated by caller.
 661  * @param fw_size
 662  *   Size of the output string, including terminating null byte.
 663  *
 664  * @return
 665  *   0 on success, or the size of the non truncated string if too big.
 666  */
 667 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
 668 {
 669         struct mlx5_priv *priv = dev->data->dev_private;
 670         struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
 671         size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
 672
 673         if (fw_size < size)
 674                 return size;
 675         if (fw_ver != NULL)
 676                 strlcpy(fw_ver, attr->fw_ver, fw_size);
 677         return 0;
 678 }
 679
 680 /**
 681  * Get supported packet types.
 682  *
 683  * @param dev
 684  *   Pointer to Ethernet device structure.
 685  *
 686  * @return
 687  *   A pointer to the supported Packet types array.
 688  */
 689 const uint32_t *
 690 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 691 {
 692         static const uint32_t ptypes[] = {
 693                 /* refers to rxq_cq_to_pkt_type() */
 694                 RTE_PTYPE_L2_ETHER,
 695                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 696                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 697                 RTE_PTYPE_L4_NONFRAG,
 698                 RTE_PTYPE_L4_FRAG,
 699                 RTE_PTYPE_L4_TCP,
 700                 RTE_PTYPE_L4_UDP,
 701                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 702                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 703                 RTE_PTYPE_INNER_L4_NONFRAG,
 704                 RTE_PTYPE_INNER_L4_FRAG,
 705                 RTE_PTYPE_INNER_L4_TCP,
 706                 RTE_PTYPE_INNER_L4_UDP,
 707                 RTE_PTYPE_UNKNOWN
 708         };
 709
 710         if (dev->rx_pkt_burst == mlx5_rx_burst ||
 711             dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
 712             dev->rx_pkt_burst == mlx5_rx_burst_vec)
 713                 return ptypes;
 714         return NULL;
 715 }
 716
 717 /**
 718  * Retrieve the master device for representor in the same switch domain.
 719  *
 720  * @param dev
 721  *   Pointer to representor Ethernet device structure.
 722  *
 723  * @return
 724  *   Master device structure  on success, NULL otherwise.
 725  */
 726
 727 static struct rte_eth_dev *
 728 mlx5_find_master_dev(struct rte_eth_dev *dev)
 729 {
 730         struct mlx5_priv *priv;
 731         uint16_t port_id;
 732         uint16_t domain_id;
 733
 734         priv = dev->data->dev_private;
 735         domain_id = priv->domain_id;
 736         assert(priv->representor);
 737         MLX5_ETH_FOREACH_DEV(port_id) {
 738                 struct mlx5_priv *opriv =
 739                         rte_eth_devices[port_id].data->dev_private;
 740                 if (opriv &&
 741                     opriv->master &&
 742                     opriv->domain_id == domain_id &&
 743                     opriv->sh == priv->sh)
 744                         return &rte_eth_devices[port_id];
 745         }
 746         return NULL;
 747 }
 748
 749 /**
 750  * DPDK callback to retrieve physical link information.
 751  *
 752  * @param dev
 753  *   Pointer to Ethernet device structure.
 754  * @param[out] link
 755  *   Storage for current link status.
 756  *
 757  * @return
 758  *   0 on success, a negative errno value otherwise and rte_errno is set.
 759  */
 760 static int
 761 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
 762                                struct rte_eth_link *link)
 763 {
 764         struct mlx5_priv *priv = dev->data->dev_private;
 765         struct ethtool_cmd edata = {
 766                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 767         };
 768         struct ifreq ifr;
 769         struct rte_eth_link dev_link;
 770         int link_speed = 0;
 771         int ret;
 772
 773         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 774         if (ret) {
 775                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 776                         dev->data->port_id, strerror(rte_errno));
 777                 return ret;
 778         }
 779         dev_link = (struct rte_eth_link) {
 780                 .link_status = ((ifr.ifr_flags & IFF_UP) &&
 781                                 (ifr.ifr_flags & IFF_RUNNING)),
 782         };
 783         ifr = (struct ifreq) {
 784                 .ifr_data = (void *)&edata,
 785         };
 786         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 787         if (ret) {
 788                 if (ret == -ENOTSUP && priv->representor) {
 789                         struct rte_eth_dev *master;
 790
 791                         /*
 792                          * For representors we can try to inherit link
 793                          * settings from the master device. Actually
 794                          * link settings do not make a lot of sense
 795                          * for representors due to missing physical
 796                          * link. The old kernel drivers supported
 797                          * emulated settings query for representors,
 798                          * the new ones do not, so we have to add
 799                          * this code for compatibility issues.
 800                          */
 801                         master = mlx5_find_master_dev(dev);
 802                         if (master) {
 803                                 ifr = (struct ifreq) {
 804                                         .ifr_data = (void *)&edata,
 805                                 };
 806                                 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
 807                         }
 808                 }
 809                 if (ret) {
 810                         DRV_LOG(WARNING,
 811                                 "port %u ioctl(SIOCETHTOOL,"
 812                                 " ETHTOOL_GSET) failed: %s",
 813                                 dev->data->port_id, strerror(rte_errno));
 814                         return ret;
 815                 }
 816         }
 817         link_speed = ethtool_cmd_speed(&edata);
 818         if (link_speed == -1)
 819                 dev_link.link_speed = ETH_SPEED_NUM_NONE;
 820         else
 821                 dev_link.link_speed = link_speed;
 822         priv->link_speed_capa = 0;
 823         if (edata.supported & SUPPORTED_Autoneg)
 824                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 825         if (edata.supported & (SUPPORTED_1000baseT_Full |
 826                                SUPPORTED_1000baseKX_Full))
 827                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 828         if (edata.supported & SUPPORTED_10000baseKR_Full)
 829                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 830         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 831                                SUPPORTED_40000baseCR4_Full |
 832                                SUPPORTED_40000baseSR4_Full |
 833                                SUPPORTED_40000baseLR4_Full))
 834                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 835         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 836                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 837         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 838                         ETH_LINK_SPEED_FIXED);
 839         if (((dev_link.link_speed && !dev_link.link_status) ||
 840              (!dev_link.link_speed && dev_link.link_status))) {
 841                 rte_errno = EAGAIN;
 842                 return -rte_errno;
 843         }
 844         *link = dev_link;
 845         return 0;
 846 }
 847
 848 /**
 849  * Retrieve physical link information (unlocked version using new ioctl).
 850  *
 851  * @param dev
 852  *   Pointer to Ethernet device structure.
 853  * @param[out] link
 854  *   Storage for current link status.
 855  *
 856  * @return
 857  *   0 on success, a negative errno value otherwise and rte_errno is set.
 858  */
 859 static int
 860 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
 861                              struct rte_eth_link *link)
 862
 863 {
 864         struct mlx5_priv *priv = dev->data->dev_private;
 865         struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
 866         struct ifreq ifr;
 867         struct rte_eth_link dev_link;
 868         struct rte_eth_dev *master = NULL;
 869         uint64_t sc;
 870         int ret;
 871
 872         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 873         if (ret) {
 874                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 875                         dev->data->port_id, strerror(rte_errno));
 876                 return ret;
 877         }
 878         dev_link = (struct rte_eth_link) {
 879                 .link_status = ((ifr.ifr_flags & IFF_UP) &&
 880                                 (ifr.ifr_flags & IFF_RUNNING)),
 881         };
 882         ifr = (struct ifreq) {
 883                 .ifr_data = (void *)&gcmd,
 884         };
 885         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 886         if (ret) {
 887                 if (ret == -ENOTSUP && priv->representor) {
 888                         /*
 889                          * For representors we can try to inherit link
 890                          * settings from the master device. Actually
 891                          * link settings do not make a lot of sense
 892                          * for representors due to missing physical
 893                          * link. The old kernel drivers supported
 894                          * emulated settings query for representors,
 895                          * the new ones do not, so we have to add
 896                          * this code for compatibility issues.
 897                          */
 898                         master = mlx5_find_master_dev(dev);
 899                         if (master) {
 900                                 ifr = (struct ifreq) {
 901                                         .ifr_data = (void *)&gcmd,
 902                                 };
 903                                 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
 904                         }
 905                 }
 906                 if (ret) {
 907                         DRV_LOG(DEBUG,
 908                                 "port %u ioctl(SIOCETHTOOL,"
 909                                 " ETHTOOL_GLINKSETTINGS) failed: %s",
 910                                 dev->data->port_id, strerror(rte_errno));
 911                         return ret;
 912                 }
 913
 914         }
 915         gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
 916
 917         alignas(struct ethtool_link_settings)
 918         uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
 919                      sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
 920         struct ethtool_link_settings *ecmd = (void *)data;
 921
 922         *ecmd = gcmd;
 923         ifr.ifr_data = (void *)ecmd;
 924         ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
 925         if (ret) {
 926                 DRV_LOG(DEBUG,
 927                         "port %u ioctl(SIOCETHTOOL,"
 928                         "ETHTOOL_GLINKSETTINGS) failed: %s",
 929                         dev->data->port_id, strerror(rte_errno));
 930                 return ret;
 931         }
 932         dev_link.link_speed = (ecmd->speed == UINT32_MAX) ? ETH_SPEED_NUM_NONE :
 933                                                             ecmd->speed;
 934         sc = ecmd->link_mode_masks[0] |
 935                 ((uint64_t)ecmd->link_mode_masks[1] << 32);
 936         priv->link_speed_capa = 0;
 937         if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
 938                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 939         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
 940                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
 941                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 942         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
 943                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
 944                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
 945                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 946         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
 947                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
 948                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
 949         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
 950                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
 951                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
 952                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
 953                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 954         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
 955                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
 956                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
 957                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
 958                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
 959         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
 960                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
 961                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
 962                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
 963         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
 964                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
 965                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
 966         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
 967                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
 968                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
 969                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
 970                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
 971         dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
 972                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 973         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 974                                   ETH_LINK_SPEED_FIXED);
 975         if (((dev_link.link_speed && !dev_link.link_status) ||
 976              (!dev_link.link_speed && dev_link.link_status))) {
 977                 rte_errno = EAGAIN;
 978                 return -rte_errno;
 979         }
 980         *link = dev_link;
 981         return 0;
 982 }
 983
 984 /**
 985  * DPDK callback to retrieve physical link information.
 986  *
 987  * @param dev
 988  *   Pointer to Ethernet device structure.
 989  * @param wait_to_complete
 990  *   Wait for request completion.
 991  *
 992  * @return
 993  *   0 if link status was not updated, positive if it was, a negative errno
 994  *   value otherwise and rte_errno is set.
 995  */
 996 int
 997 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
 998 {
 999         int ret;
1000         struct rte_eth_link dev_link;
1001         time_t start_time = time(NULL);
1002
1003         do {
1004                 ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
1005                 if (ret == -ENOTSUP)
1006                         ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
1007                 if (ret == 0)
1008                         break;
1009                 /* Handle wait to complete situation. */
1010                 if (wait_to_complete && ret == -EAGAIN) {
1011                         if (abs((int)difftime(time(NULL), start_time)) <
1012                             MLX5_LINK_STATUS_TIMEOUT) {
1013                                 usleep(0);
1014                                 continue;
1015                         } else {
1016                                 rte_errno = EBUSY;
1017                                 return -rte_errno;
1018                         }
1019                 } else if (ret < 0) {
1020                         return ret;
1021                 }
1022         } while (wait_to_complete);
1023         ret = !!memcmp(&dev->data->dev_link, &dev_link,
1024                        sizeof(struct rte_eth_link));
1025         dev->data->dev_link = dev_link;
1026         return ret;
1027 }
1028
1029 /**
1030  * DPDK callback to change the MTU.
1031  *
1032  * @param dev
1033  *   Pointer to Ethernet device structure.
1034  * @param in_mtu
1035  *   New MTU.
1036  *
1037  * @return
1038  *   0 on success, a negative errno value otherwise and rte_errno is set.
1039  */
1040 int
1041 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1042 {
1043         struct mlx5_priv *priv = dev->data->dev_private;
1044         uint16_t kern_mtu = 0;
1045         int ret;
1046
1047         ret = mlx5_get_mtu(dev, &kern_mtu);
1048         if (ret)
1049                 return ret;
1050         /* Set kernel interface MTU first. */
1051         ret = mlx5_set_mtu(dev, mtu);
1052         if (ret)
1053                 return ret;
1054         ret = mlx5_get_mtu(dev, &kern_mtu);
1055         if (ret)
1056                 return ret;
1057         if (kern_mtu == mtu) {
1058                 priv->mtu = mtu;
1059                 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1060                         dev->data->port_id, mtu);
1061                 return 0;
1062         }
1063         rte_errno = EAGAIN;
1064         return -rte_errno;
1065 }
1066
1067 /**
1068  * DPDK callback to get flow control status.
1069  *
1070  * @param dev
1071  *   Pointer to Ethernet device structure.
1072  * @param[out] fc_conf
1073  *   Flow control output buffer.
1074  *
1075  * @return
1076  *   0 on success, a negative errno value otherwise and rte_errno is set.
1077  */
1078 int
1079 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1080 {
1081         struct ifreq ifr;
1082         struct ethtool_pauseparam ethpause = {
1083                 .cmd = ETHTOOL_GPAUSEPARAM
1084         };
1085         int ret;
1086
1087         ifr.ifr_data = (void *)&ethpause;
1088         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1089         if (ret) {
1090                 DRV_LOG(WARNING,
1091                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1092                         " %s",
1093                         dev->data->port_id, strerror(rte_errno));
1094                 return ret;
1095         }
1096         fc_conf->autoneg = ethpause.autoneg;
1097         if (ethpause.rx_pause && ethpause.tx_pause)
1098                 fc_conf->mode = RTE_FC_FULL;
1099         else if (ethpause.rx_pause)
1100                 fc_conf->mode = RTE_FC_RX_PAUSE;
1101         else if (ethpause.tx_pause)
1102                 fc_conf->mode = RTE_FC_TX_PAUSE;
1103         else
1104                 fc_conf->mode = RTE_FC_NONE;
1105         return 0;
1106 }
1107
1108 /**
1109  * DPDK callback to modify flow control parameters.
1110  *
1111  * @param dev
1112  *   Pointer to Ethernet device structure.
1113  * @param[in] fc_conf
1114  *   Flow control parameters.
1115  *
1116  * @return
1117  *   0 on success, a negative errno value otherwise and rte_errno is set.
1118  */
1119 int
1120 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1121 {
1122         struct ifreq ifr;
1123         struct ethtool_pauseparam ethpause = {
1124                 .cmd = ETHTOOL_SPAUSEPARAM
1125         };
1126         int ret;
1127
1128         ifr.ifr_data = (void *)&ethpause;
1129         ethpause.autoneg = fc_conf->autoneg;
1130         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1131             (fc_conf->mode & RTE_FC_RX_PAUSE))
1132                 ethpause.rx_pause = 1;
1133         else
1134                 ethpause.rx_pause = 0;
1135
1136         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1137             (fc_conf->mode & RTE_FC_TX_PAUSE))
1138                 ethpause.tx_pause = 1;
1139         else
1140                 ethpause.tx_pause = 0;
1141         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1142         if (ret) {
1143                 DRV_LOG(WARNING,
1144                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1145                         " failed: %s",
1146                         dev->data->port_id, strerror(rte_errno));
1147                 return ret;
1148         }
1149         return 0;
1150 }
1151
1152 /**
1153  * Get PCI information by sysfs device path.
1154  *
1155  * @param dev_path
1156  *   Pointer to device sysfs folder name.
1157  * @param[out] pci_addr
1158  *   PCI bus address output buffer.
1159  *
1160  * @return
1161  *   0 on success, a negative errno value otherwise and rte_errno is set.
1162  */
1163 int
1164 mlx5_dev_to_pci_addr(const char *dev_path,
1165                      struct rte_pci_addr *pci_addr)
1166 {
1167         FILE *file;
1168         char line[32];
1169         MKSTR(path, "%s/device/uevent", dev_path);
1170
1171         file = fopen(path, "rb");
1172         if (file == NULL) {
1173                 rte_errno = errno;
1174                 return -rte_errno;
1175         }
1176         while (fgets(line, sizeof(line), file) == line) {
1177                 size_t len = strlen(line);
1178                 int ret;
1179
1180                 /* Truncate long lines. */
1181                 if (len == (sizeof(line) - 1))
1182                         while (line[(len - 1)] != '\n') {
1183                                 ret = fgetc(file);
1184                                 if (ret == EOF)
1185                                         break;
1186                                 line[(len - 1)] = ret;
1187                         }
1188                 /* Extract information. */
1189                 if (sscanf(line,
1190                            "PCI_SLOT_NAME="
1191                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1192                            &pci_addr->domain,
1193                            &pci_addr->bus,
1194                            &pci_addr->devid,
1195                            &pci_addr->function) == 4) {
1196                         ret = 0;
1197                         break;
1198                 }
1199         }
1200         fclose(file);
1201         return 0;
1202 }
1203
1204 /**
1205  * Handle asynchronous removal event for entire multiport device.
1206  *
1207  * @param sh
1208  *   Infiniband device shared context.
1209  */
1210 static void
1211 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1212 {
1213         uint32_t i;
1214
1215         for (i = 0; i < sh->max_port; ++i) {
1216                 struct rte_eth_dev *dev;
1217
1218                 if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1219                         /*
1220                          * Or not existing port either no
1221                          * handler installed for this port.
1222                          */
1223                         continue;
1224                 }
1225                 dev = &rte_eth_devices[sh->port[i].ih_port_id];
1226                 assert(dev);
1227                 if (dev->data->dev_conf.intr_conf.rmv)
1228                         _rte_eth_dev_callback_process
1229                                 (dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1230         }
1231 }
1232
1233 /**
1234  * Handle shared asynchronous events the NIC (removal event
1235  * and link status change). Supports multiport IB device.
1236  *
1237  * @param cb_arg
1238  *   Callback argument.
1239  */
1240 void
1241 mlx5_dev_interrupt_handler(void *cb_arg)
1242 {
1243         struct mlx5_ibv_shared *sh = cb_arg;
1244         struct ibv_async_event event;
1245
1246         /* Read all message from the IB device and acknowledge them. */
1247         for (;;) {
1248                 struct rte_eth_dev *dev;
1249                 uint32_t tmp;
1250
1251                 if (mlx5_glue->get_async_event(sh->ctx, &event))
1252                         break;
1253                 /* Retrieve and check IB port index. */
1254                 tmp = (uint32_t)event.element.port_num;
1255                 if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1256                         /*
1257                          * The DEVICE_FATAL event is called once for
1258                          * entire device without port specifying.
1259                          * We should notify all existing ports.
1260                          */
1261                         mlx5_glue->ack_async_event(&event);
1262                         mlx5_dev_interrupt_device_fatal(sh);
1263                         continue;
1264                 }
1265                 assert(tmp && (tmp <= sh->max_port));
1266                 if (!tmp) {
1267                         /* Unsupported devive level event. */
1268                         mlx5_glue->ack_async_event(&event);
1269                         DRV_LOG(DEBUG,
1270                                 "unsupported common event (type %d)",
1271                                 event.event_type);
1272                         continue;
1273                 }
1274                 if (tmp > sh->max_port) {
1275                         /* Invalid IB port index. */
1276                         mlx5_glue->ack_async_event(&event);
1277                         DRV_LOG(DEBUG,
1278                                 "cannot handle an event (type %d)"
1279                                 "due to invalid IB port index (%u)",
1280                                 event.event_type, tmp);
1281                         continue;
1282                 }
1283                 if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1284                         /* No handler installed. */
1285                         mlx5_glue->ack_async_event(&event);
1286                         DRV_LOG(DEBUG,
1287                                 "cannot handle an event (type %d)"
1288                                 "due to no handler installed for port %u",
1289                                 event.event_type, tmp);
1290                         continue;
1291                 }
1292                 /* Retrieve ethernet device descriptor. */
1293                 tmp = sh->port[tmp - 1].ih_port_id;
1294                 dev = &rte_eth_devices[tmp];
1295                 assert(dev);
1296                 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1297                      event.event_type == IBV_EVENT_PORT_ERR) &&
1298                         dev->data->dev_conf.intr_conf.lsc) {
1299                         mlx5_glue->ack_async_event(&event);
1300                         if (mlx5_link_update(dev, 0) == -EAGAIN) {
1301                                 usleep(0);
1302                                 continue;
1303                         }
1304                         _rte_eth_dev_callback_process
1305                                 (dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1306                         continue;
1307                 }
1308                 DRV_LOG(DEBUG,
1309                         "port %u cannot handle an unknown event (type %d)",
1310                         dev->data->port_id, event.event_type);
1311                 mlx5_glue->ack_async_event(&event);
1312         }
1313 }
1314
1315 /*
1316  * Unregister callback handler safely. The handler may be active
1317  * while we are trying to unregister it, in this case code -EAGAIN
1318  * is returned by rte_intr_callback_unregister(). This routine checks
1319  * the return code and tries to unregister handler again.
1320  *
1321  * @param handle
1322  *   interrupt handle
1323  * @param cb_fn
1324  *   pointer to callback routine
1325  * @cb_arg
1326  *   opaque callback parameter
1327  */
1328 void
1329 mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
1330                               rte_intr_callback_fn cb_fn, void *cb_arg)
1331 {
1332         /*
1333          * Try to reduce timeout management overhead by not calling
1334          * the timer related routines on the first iteration. If the
1335          * unregistering succeeds on first call there will be no
1336          * timer calls at all.
1337          */
1338         uint64_t twait = 0;
1339         uint64_t start = 0;
1340
1341         do {
1342                 int ret;
1343
1344                 ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
1345                 if (ret >= 0)
1346                         return;
1347                 if (ret != -EAGAIN) {
1348                         DRV_LOG(INFO, "failed to unregister interrupt"
1349                                       " handler (error: %d)", ret);
1350                         assert(false);
1351                         return;
1352                 }
1353                 if (twait) {
1354                         struct timespec onems;
1355
1356                         /* Wait one millisecond and try again. */
1357                         onems.tv_sec = 0;
1358                         onems.tv_nsec = NS_PER_S / MS_PER_S;
1359                         nanosleep(&onems, 0);
1360                         /* Check whether one second elapsed. */
1361                         if ((rte_get_timer_cycles() - start) <= twait)
1362                                 continue;
1363                 } else {
1364                         /*
1365                          * We get the amount of timer ticks for one second.
1366                          * If this amount elapsed it means we spent one
1367                          * second in waiting. This branch is executed once
1368                          * on first iteration.
1369                          */
1370                         twait = rte_get_timer_hz();
1371                         assert(twait);
1372                 }
1373                 /*
1374                  * Timeout elapsed, show message (once a second) and retry.
1375                  * We have no other acceptable option here, if we ignore
1376                  * the unregistering return code the handler will not
1377                  * be unregistered, fd will be closed and we may get the
1378                  * crush. Hanging and messaging in the loop seems not to be
1379                  * the worst choice.
1380                  */
1381                 DRV_LOG(INFO, "Retrying to unregister interrupt handler");
1382                 start = rte_get_timer_cycles();
1383         } while (true);
1384 }
1385
1386 /**
1387  * Handle DEVX interrupts from the NIC.
1388  * This function is probably called from the DPDK host thread.
1389  *
1390  * @param cb_arg
1391  *   Callback argument.
1392  */
1393 void
1394 mlx5_dev_interrupt_handler_devx(void *cb_arg)
1395 {
1396 #ifndef HAVE_IBV_DEVX_ASYNC
1397         (void)cb_arg;
1398         return;
1399 #else
1400         struct mlx5_ibv_shared *sh = cb_arg;
1401         union {
1402                 struct mlx5dv_devx_async_cmd_hdr cmd_resp;
1403                 uint8_t buf[MLX5_ST_SZ_BYTES(query_flow_counter_out) +
1404                             MLX5_ST_SZ_BYTES(traffic_counter) +
1405                             sizeof(struct mlx5dv_devx_async_cmd_hdr)];
1406         } out;
1407         uint8_t *buf = out.buf + sizeof(out.cmd_resp);
1408
1409         while (!mlx5_glue->devx_get_async_cmd_comp(sh->devx_comp,
1410                                                    &out.cmd_resp,
1411                                                    sizeof(out.buf)))
1412                 mlx5_flow_async_pool_query_handle
1413                         (sh, (uint64_t)out.cmd_resp.wr_id,
1414                          mlx5_devx_get_out_command_status(buf));
1415 #endif /* HAVE_IBV_DEVX_ASYNC */
1416 }
1417
1418 /**
1419  * Uninstall shared asynchronous device events handler.
1420  * This function is implemented to support event sharing
1421  * between multiple ports of single IB device.
1422  *
1423  * @param dev
1424  *   Pointer to Ethernet device.
1425  */
1426 static void
1427 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1428 {
1429         struct mlx5_priv *priv = dev->data->dev_private;
1430         struct mlx5_ibv_shared *sh = priv->sh;
1431
1432         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1433                 return;
1434         pthread_mutex_lock(&sh->intr_mutex);
1435         assert(priv->ibv_port);
1436         assert(priv->ibv_port <= sh->max_port);
1437         assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1438         if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1439                 goto exit;
1440         assert(sh->port[priv->ibv_port - 1].ih_port_id ==
1441                                         (uint32_t)dev->data->port_id);
1442         assert(sh->intr_cnt);
1443         sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1444         if (!sh->intr_cnt || --sh->intr_cnt)
1445                 goto exit;
1446         mlx5_intr_callback_unregister(&sh->intr_handle,
1447                                      mlx5_dev_interrupt_handler, sh);
1448         sh->intr_handle.fd = 0;
1449         sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1450         if (sh->intr_handle_devx.fd) {
1451                 rte_intr_callback_unregister(&sh->intr_handle_devx,
1452                                              mlx5_dev_interrupt_handler_devx,
1453                                              sh);
1454                 sh->intr_handle_devx.fd = 0;
1455                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_UNKNOWN;
1456         }
1457         if (sh->devx_comp) {
1458                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1459                 sh->devx_comp = NULL;
1460         }
1461 exit:
1462         pthread_mutex_unlock(&sh->intr_mutex);
1463 }
1464
1465 /**
1466  * Install shared asynchronous device events handler.
1467  * This function is implemented to support event sharing
1468  * between multiple ports of single IB device.
1469  *
1470  * @param dev
1471  *   Pointer to Ethernet device.
1472  */
1473 static void
1474 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1475 {
1476         struct mlx5_priv *priv = dev->data->dev_private;
1477         struct mlx5_ibv_shared *sh = priv->sh;
1478         int ret;
1479         int flags;
1480
1481         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1482                 return;
1483         pthread_mutex_lock(&sh->intr_mutex);
1484         assert(priv->ibv_port);
1485         assert(priv->ibv_port <= sh->max_port);
1486         assert(dev->data->port_id < RTE_MAX_ETHPORTS);
1487         if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1488                 /* The handler is already installed for this port. */
1489                 assert(sh->intr_cnt);
1490                 goto exit;
1491         }
1492         sh->port[priv->ibv_port - 1].ih_port_id = (uint32_t)dev->data->port_id;
1493         if (sh->intr_cnt) {
1494                 sh->intr_cnt++;
1495                 goto exit;
1496         }
1497         /* No shared handler installed. */
1498         assert(sh->ctx->async_fd > 0);
1499         flags = fcntl(sh->ctx->async_fd, F_GETFL);
1500         ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1501         if (ret) {
1502                 DRV_LOG(INFO, "failed to change file descriptor"
1503                               " async event queue");
1504                 goto error;
1505         }
1506         sh->intr_handle.fd = sh->ctx->async_fd;
1507         sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1508         rte_intr_callback_register(&sh->intr_handle,
1509                                    mlx5_dev_interrupt_handler, sh);
1510         if (priv->config.devx) {
1511 #ifndef HAVE_IBV_DEVX_ASYNC
1512                 goto error_unregister;
1513 #else
1514                 sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1515                 if (sh->devx_comp) {
1516                         flags = fcntl(sh->devx_comp->fd, F_GETFL);
1517                         ret = fcntl(sh->devx_comp->fd, F_SETFL,
1518                                     flags | O_NONBLOCK);
1519                         if (ret) {
1520                                 DRV_LOG(INFO, "failed to change file descriptor"
1521                                               " devx async event queue");
1522                                 goto error_unregister;
1523                         }
1524                         sh->intr_handle_devx.fd = sh->devx_comp->fd;
1525                         sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1526                         rte_intr_callback_register
1527                                 (&sh->intr_handle_devx,
1528                                  mlx5_dev_interrupt_handler_devx, sh);
1529                 } else {
1530                         DRV_LOG(INFO, "failed to create devx async command "
1531                                 "completion");
1532                         goto error_unregister;
1533                 }
1534 #endif /* HAVE_IBV_DEVX_ASYNC */
1535         }
1536         sh->intr_cnt++;
1537         goto exit;
1538 error_unregister:
1539         rte_intr_callback_unregister(&sh->intr_handle,
1540                                      mlx5_dev_interrupt_handler, sh);
1541 error:
1542         /* Indicate there will be no interrupts. */
1543         dev->data->dev_conf.intr_conf.lsc = 0;
1544         dev->data->dev_conf.intr_conf.rmv = 0;
1545         sh->intr_handle.fd = 0;
1546         sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1547         sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1548 exit:
1549         pthread_mutex_unlock(&sh->intr_mutex);
1550 }
1551
1552 /**
1553  * Uninstall interrupt handler.
1554  *
1555  * @param dev
1556  *   Pointer to Ethernet device.
1557  */
1558 void
1559 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1560 {
1561         mlx5_dev_shared_handler_uninstall(dev);
1562 }
1563
1564 /**
1565  * Install interrupt handler.
1566  *
1567  * @param dev
1568  *   Pointer to Ethernet device.
1569  */
1570 void
1571 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1572 {
1573         mlx5_dev_shared_handler_install(dev);
1574 }
1575
1576 /**
1577  * DPDK callback to bring the link DOWN.
1578  *
1579  * @param dev
1580  *   Pointer to Ethernet device structure.
1581  *
1582  * @return
1583  *   0 on success, a negative errno value otherwise and rte_errno is set.
1584  */
1585 int
1586 mlx5_set_link_down(struct rte_eth_dev *dev)
1587 {
1588         return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1589 }
1590
1591 /**
1592  * DPDK callback to bring the link UP.
1593  *
1594  * @param dev
1595  *   Pointer to Ethernet device structure.
1596  *
1597  * @return
1598  *   0 on success, a negative errno value otherwise and rte_errno is set.
1599  */
1600 int
1601 mlx5_set_link_up(struct rte_eth_dev *dev)
1602 {
1603         return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1604 }
1605
1606 /**
1607  * Configure the RX function to use.
1608  *
1609  * @param dev
1610  *   Pointer to private data structure.
1611  *
1612  * @return
1613  *   Pointer to selected Rx burst function.
1614  */
1615 eth_rx_burst_t
1616 mlx5_select_rx_function(struct rte_eth_dev *dev)
1617 {
1618         eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1619
1620         assert(dev != NULL);
1621         if (mlx5_check_vec_rx_support(dev) > 0) {
1622                 rx_pkt_burst = mlx5_rx_burst_vec;
1623                 DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1624                         dev->data->port_id);
1625         } else if (mlx5_mprq_enabled(dev)) {
1626                 rx_pkt_burst = mlx5_rx_burst_mprq;
1627         }
1628         return rx_pkt_burst;
1629 }
1630
1631 /**
1632  * Check if mlx5 device was removed.
1633  *
1634  * @param dev
1635  *   Pointer to Ethernet device structure.
1636  *
1637  * @return
1638  *   1 when device is removed, otherwise 0.
1639  */
1640 int
1641 mlx5_is_removed(struct rte_eth_dev *dev)
1642 {
1643         struct ibv_device_attr device_attr;
1644         struct mlx5_priv *priv = dev->data->dev_private;
1645
1646         if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1647                 return 1;
1648         return 0;
1649 }
1650
1651 /**
1652  * Get the E-Switch parameters by port id.
1653  *
1654  * @param[in] port
1655  *   Device port id.
1656  * @param[out] es_domain_id
1657  *   E-Switch domain id.
1658  * @param[out] es_port_id
1659  *   The port id of the port in the E-Switch.
1660  *
1661  * @return
1662  *   pointer to device private data structure containing data needed
1663  *   on success, NULL otherwise and rte_errno is set.
1664  */
1665 struct mlx5_priv *
1666 mlx5_port_to_eswitch_info(uint16_t port)
1667 {
1668         struct rte_eth_dev *dev;
1669         struct mlx5_priv *priv;
1670
1671         if (port >= RTE_MAX_ETHPORTS) {
1672                 rte_errno = EINVAL;
1673                 return NULL;
1674         }
1675         if (!rte_eth_dev_is_valid_port(port)) {
1676                 rte_errno = ENODEV;
1677                 return NULL;
1678         }
1679         dev = &rte_eth_devices[port];
1680         priv = dev->data->dev_private;
1681         if (!(priv->representor || priv->master)) {
1682                 rte_errno = EINVAL;
1683                 return NULL;
1684         }
1685         return priv;
1686 }
1687
1688 /**
1689  * Get the E-Switch parameters by device instance.
1690  *
1691  * @param[in] port
1692  *   Device port id.
1693  * @param[out] es_domain_id
1694  *   E-Switch domain id.
1695  * @param[out] es_port_id
1696  *   The port id of the port in the E-Switch.
1697  *
1698  * @return
1699  *   pointer to device private data structure containing data needed
1700  *   on success, NULL otherwise and rte_errno is set.
1701  */
1702 struct mlx5_priv *
1703 mlx5_dev_to_eswitch_info(struct rte_eth_dev *dev)
1704 {
1705         struct mlx5_priv *priv;
1706
1707         priv = dev->data->dev_private;
1708         if (!(priv->representor || priv->master)) {
1709                 rte_errno = EINVAL;
1710                 return NULL;
1711         }
1712         return priv;
1713 }
1714
1715 /**
1716  * Get switch information associated with network interface.
1717  *
1718  * @param ifindex
1719  *   Network interface index.
1720  * @param[out] info
1721  *   Switch information object, populated in case of success.
1722  *
1723  * @return
1724  *   0 on success, a negative errno value otherwise and rte_errno is set.
1725  */
1726 int
1727 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1728 {
1729         char ifname[IF_NAMESIZE];
1730         char port_name[IF_NAMESIZE];
1731         FILE *file;
1732         struct mlx5_switch_info data = {
1733                 .master = 0,
1734                 .representor = 0,
1735                 .name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1736                 .port_name = 0,
1737                 .switch_id = 0,
1738         };
1739         DIR *dir;
1740         bool port_switch_id_set = false;
1741         bool device_dir = false;
1742         char c;
1743         int ret;
1744
1745         if (!if_indextoname(ifindex, ifname)) {
1746                 rte_errno = errno;
1747                 return -rte_errno;
1748         }
1749
1750         MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1751               ifname);
1752         MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1753               ifname);
1754         MKSTR(pci_device, "/sys/class/net/%s/device",
1755               ifname);
1756
1757         file = fopen(phys_port_name, "rb");
1758         if (file != NULL) {
1759                 ret = fscanf(file, "%s", port_name);
1760                 fclose(file);
1761                 if (ret == 1)
1762                         mlx5_translate_port_name(port_name, &data);
1763         }
1764         file = fopen(phys_switch_id, "rb");
1765         if (file == NULL) {
1766                 rte_errno = errno;
1767                 return -rte_errno;
1768         }
1769         port_switch_id_set =
1770                 fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1771                 c == '\n';
1772         fclose(file);
1773         dir = opendir(pci_device);
1774         if (dir != NULL) {
1775                 closedir(dir);
1776                 device_dir = true;
1777         }
1778         if (port_switch_id_set) {
1779                 /* We have some E-Switch configuration. */
1780                 mlx5_sysfs_check_switch_info(device_dir, &data);
1781         }
1782         *info = data;
1783         assert(!(data.master && data.representor));
1784         if (data.master && data.representor) {
1785                 DRV_LOG(ERR, "ifindex %u device is recognized as master"
1786                              " and as representor", ifindex);
1787                 rte_errno = ENODEV;
1788                 return -rte_errno;
1789         }
1790         return 0;
1791 }
1792
1793 /**
1794  * Analyze gathered port parameters via Netlink to recognize master
1795  * and representor devices for E-Switch configuration.
1796  *
1797  * @param[in] num_vf_set
1798  *   flag of presence of number of VFs port attribute.
1799  * @param[inout] switch_info
1800  *   Port information, including port name as a number and port name
1801  *   type if recognized
1802  *
1803  * @return
1804  *   master and representor flags are set in switch_info according to
1805  *   recognized parameters (if any).
1806  */
1807 void
1808 mlx5_nl_check_switch_info(bool num_vf_set,
1809                           struct mlx5_switch_info *switch_info)
1810 {
1811         switch (switch_info->name_type) {
1812         case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1813                 /*
1814                  * Name is not recognized, assume the master,
1815                  * check the number of VFs key presence.
1816                  */
1817                 switch_info->master = num_vf_set;
1818                 break;
1819         case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1820                 /*
1821                  * Name is not set, this assumes the legacy naming
1822                  * schema for master, just check if there is a
1823                  * number of VFs key.
1824                  */
1825                 switch_info->master = num_vf_set;
1826                 break;
1827         case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1828                 /* New uplink naming schema recognized. */
1829                 switch_info->master = 1;
1830                 break;
1831         case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1832                 /* Legacy representors naming schema. */
1833                 switch_info->representor = !num_vf_set;
1834                 break;
1835         case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1836                 /* New representors naming schema. */
1837                 switch_info->representor = 1;
1838                 break;
1839         }
1840 }
1841
1842 /**
1843  * Analyze gathered port parameters via sysfs to recognize master
1844  * and representor devices for E-Switch configuration.
1845  *
1846  * @param[in] device_dir
1847  *   flag of presence of "device" directory under port device key.
1848  * @param[inout] switch_info
1849  *   Port information, including port name as a number and port name
1850  *   type if recognized
1851  *
1852  * @return
1853  *   master and representor flags are set in switch_info according to
1854  *   recognized parameters (if any).
1855  */
1856 void
1857 mlx5_sysfs_check_switch_info(bool device_dir,
1858                              struct mlx5_switch_info *switch_info)
1859 {
1860         switch (switch_info->name_type) {
1861         case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1862                 /*
1863                  * Name is not recognized, assume the master,
1864                  * check the device directory presence.
1865                  */
1866                 switch_info->master = device_dir;
1867                 break;
1868         case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1869                 /*
1870                  * Name is not set, this assumes the legacy naming
1871                  * schema for master, just check if there is
1872                  * a device directory.
1873                  */
1874                 switch_info->master = device_dir;
1875                 break;
1876         case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1877                 /* New uplink naming schema recognized. */
1878                 switch_info->master = 1;
1879                 break;
1880         case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1881                 /* Legacy representors naming schema. */
1882                 switch_info->representor = !device_dir;
1883                 break;
1884         case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1885                 /* New representors naming schema. */
1886                 switch_info->representor = 1;
1887                 break;
1888         }
1889 }
1890
1891 /**
1892  * Extract port name, as a number, from sysfs or netlink information.
1893  *
1894  * @param[in] port_name_in
1895  *   String representing the port name.
1896  * @param[out] port_info_out
1897  *   Port information, including port name as a number and port name
1898  *   type if recognized
1899  *
1900  * @return
1901  *   port_name field set according to recognized name format.
1902  */
1903 void
1904 mlx5_translate_port_name(const char *port_name_in,
1905                          struct mlx5_switch_info *port_info_out)
1906 {
1907         char pf_c1, pf_c2, vf_c1, vf_c2;
1908         char *end;
1909         int sc_items;
1910
1911         /*
1912          * Check for port-name as a string of the form pf0vf0
1913          * (support kernel ver >= 5.0 or OFED ver >= 4.6).
1914          */
1915         sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
1916                           &pf_c1, &pf_c2, &port_info_out->pf_num,
1917                           &vf_c1, &vf_c2, &port_info_out->port_name);
1918         if (sc_items == 6 &&
1919             pf_c1 == 'p' && pf_c2 == 'f' &&
1920             vf_c1 == 'v' && vf_c2 == 'f') {
1921                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
1922                 return;
1923         }
1924         /*
1925          * Check for port-name as a string of the form p0
1926          * (support kernel ver >= 5.0, or OFED ver >= 4.6).
1927          */
1928         sc_items = sscanf(port_name_in, "%c%d",
1929                           &pf_c1, &port_info_out->port_name);
1930         if (sc_items == 2 && pf_c1 == 'p') {
1931                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
1932                 return;
1933         }
1934         /* Check for port-name as a number (support kernel ver < 5.0 */
1935         errno = 0;
1936         port_info_out->port_name = strtol(port_name_in, &end, 0);
1937         if (!errno &&
1938             (size_t)(end - port_name_in) == strlen(port_name_in)) {
1939                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
1940                 return;
1941         }
1942         port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
1943         return;
1944 }
1945
1946 /**
1947  * DPDK callback to retrieve plug-in module EEPROM information (type and size).
1948  *
1949  * @param dev
1950  *   Pointer to Ethernet device structure.
1951  * @param[out] modinfo
1952  *   Storage for plug-in module EEPROM information.
1953  *
1954  * @return
1955  *   0 on success, a negative errno value otherwise and rte_errno is set.
1956  */
1957 int
1958 mlx5_get_module_info(struct rte_eth_dev *dev,
1959                      struct rte_eth_dev_module_info *modinfo)
1960 {
1961         struct ethtool_modinfo info = {
1962                 .cmd = ETHTOOL_GMODULEINFO,
1963         };
1964         struct ifreq ifr = (struct ifreq) {
1965                 .ifr_data = (void *)&info,
1966         };
1967         int ret = 0;
1968
1969         if (!dev || !modinfo) {
1970                 DRV_LOG(WARNING, "missing argument, cannot get module info");
1971                 rte_errno = EINVAL;
1972                 return -rte_errno;
1973         }
1974         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1975         if (ret) {
1976                 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
1977                         dev->data->port_id, strerror(rte_errno));
1978                 return ret;
1979         }
1980         modinfo->type = info.type;
1981         modinfo->eeprom_len = info.eeprom_len;
1982         return ret;
1983 }
1984
1985 /**
1986  * DPDK callback to retrieve plug-in module EEPROM data.
1987  *
1988  * @param dev
1989  *   Pointer to Ethernet device structure.
1990  * @param[out] info
1991  *   Storage for plug-in module EEPROM data.
1992  *
1993  * @return
1994  *   0 on success, a negative errno value otherwise and rte_errno is set.
1995  */
1996 int mlx5_get_module_eeprom(struct rte_eth_dev *dev,
1997                            struct rte_dev_eeprom_info *info)
1998 {
1999         struct ethtool_eeprom *eeprom;
2000         struct ifreq ifr;
2001         int ret = 0;
2002
2003         if (!dev || !info) {
2004                 DRV_LOG(WARNING, "missing argument, cannot get module eeprom");
2005                 rte_errno = EINVAL;
2006                 return -rte_errno;
2007         }
2008         eeprom = rte_calloc(__func__, 1,
2009                             (sizeof(struct ethtool_eeprom) + info->length), 0);
2010         if (!eeprom) {
2011                 DRV_LOG(WARNING, "port %u cannot allocate memory for "
2012                         "eeprom data", dev->data->port_id);
2013                 rte_errno = ENOMEM;
2014                 return -rte_errno;
2015         }
2016         eeprom->cmd = ETHTOOL_GMODULEEEPROM;
2017         eeprom->offset = info->offset;
2018         eeprom->len = info->length;
2019         ifr = (struct ifreq) {
2020                 .ifr_data = (void *)eeprom,
2021         };
2022         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2023         if (ret)
2024                 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2025                         dev->data->port_id, strerror(rte_errno));
2026         else
2027                 rte_memcpy(info->data, eeprom->data, info->length);
2028         rte_free(eeprom);
2029         return ret;
2030 }