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 <inttypes.h>
   8 #include <unistd.h>
   9 #include <stdbool.h>
  10 #include <stdint.h>
  11 #include <stdio.h>
  12 #include <string.h>
  13 #include <stdlib.h>
  14 #include <errno.h>
  15 #include <dirent.h>
  16 #include <net/if.h>
  17 #include <sys/ioctl.h>
  18 #include <sys/socket.h>
  19 #include <netinet/in.h>
  20 #include <linux/ethtool.h>
  21 #include <linux/sockios.h>
  22 #include <fcntl.h>
  23 #include <stdalign.h>
  24 #include <sys/un.h>
  25 #include <time.h>
  26
  27 #include <rte_atomic.h>
  28 #include <rte_ethdev_driver.h>
  29 #include <rte_bus_pci.h>
  30 #include <rte_mbuf.h>
  31 #include <rte_common.h>
  32 #include <rte_interrupts.h>
  33 #include <rte_malloc.h>
  34 #include <rte_string_fns.h>
  35 #include <rte_rwlock.h>
  36 #include <rte_cycles.h>
  37
  38 #include <mlx5_glue.h>
  39 #include <mlx5_devx_cmds.h>
  40 #include <mlx5_common.h>
  41
  42 #include "mlx5.h"
  43 #include "mlx5_rxtx.h"
  44 #include "mlx5_utils.h"
  45
  46 /* Supported speed values found in /usr/include/linux/ethtool.h */
  47 #ifndef HAVE_SUPPORTED_40000baseKR4_Full
  48 #define SUPPORTED_40000baseKR4_Full (1 << 23)
  49 #endif
  50 #ifndef HAVE_SUPPORTED_40000baseCR4_Full
  51 #define SUPPORTED_40000baseCR4_Full (1 << 24)
  52 #endif
  53 #ifndef HAVE_SUPPORTED_40000baseSR4_Full
  54 #define SUPPORTED_40000baseSR4_Full (1 << 25)
  55 #endif
  56 #ifndef HAVE_SUPPORTED_40000baseLR4_Full
  57 #define SUPPORTED_40000baseLR4_Full (1 << 26)
  58 #endif
  59 #ifndef HAVE_SUPPORTED_56000baseKR4_Full
  60 #define SUPPORTED_56000baseKR4_Full (1 << 27)
  61 #endif
  62 #ifndef HAVE_SUPPORTED_56000baseCR4_Full
  63 #define SUPPORTED_56000baseCR4_Full (1 << 28)
  64 #endif
  65 #ifndef HAVE_SUPPORTED_56000baseSR4_Full
  66 #define SUPPORTED_56000baseSR4_Full (1 << 29)
  67 #endif
  68 #ifndef HAVE_SUPPORTED_56000baseLR4_Full
  69 #define SUPPORTED_56000baseLR4_Full (1 << 30)
  70 #endif
  71
  72 /* Add defines in case the running kernel is not the same as user headers. */
  73 #ifndef ETHTOOL_GLINKSETTINGS
  74 struct ethtool_link_settings {
  75         uint32_t cmd;
  76         uint32_t speed;
  77         uint8_t duplex;
  78         uint8_t port;
  79         uint8_t phy_address;
  80         uint8_t autoneg;
  81         uint8_t mdio_support;
  82         uint8_t eth_to_mdix;
  83         uint8_t eth_tp_mdix_ctrl;
  84         int8_t link_mode_masks_nwords;
  85         uint32_t reserved[8];
  86         uint32_t link_mode_masks[];
  87 };
  88
  89 /* The kernel values can be found in /include/uapi/linux/ethtool.h */
  90 #define ETHTOOL_GLINKSETTINGS 0x0000004c
  91 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
  92 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
  93 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
  94 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
  95 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
  96 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
  97 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
  98 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
  99 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
 100 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
 101 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
 102 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
 103 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
 104 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
 105 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
 106 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
 107 #endif
 108 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
 109 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
 110 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
 111 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
 112 #endif
 113 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
 114 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
 115 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
 116 #endif
 117 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
 118 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
 119 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
 120 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
 121 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
 122 #endif
 123 #ifndef HAVE_ETHTOOL_LINK_MODE_200G
 124 #define ETHTOOL_LINK_MODE_200000baseKR4_Full_BIT 62
 125 #define ETHTOOL_LINK_MODE_200000baseSR4_Full_BIT 63
 126 #define ETHTOOL_LINK_MODE_200000baseLR4_ER4_FR4_Full_BIT 0 /* 64 - 64 */
 127 #define ETHTOOL_LINK_MODE_200000baseDR4_Full_BIT 1 /* 65 - 64 */
 128 #define ETHTOOL_LINK_MODE_200000baseCR4_Full_BIT 2 /* 66 - 64 */
 129 #endif
 130
 131 /**
 132  * Get master interface name from private structure.
 133  *
 134  * @param[in] dev
 135  *   Pointer to Ethernet device.
 136  * @param[out] ifname
 137  *   Interface name output buffer.
 138  *
 139  * @return
 140  *   0 on success, a negative errno value otherwise and rte_errno is set.
 141  */
 142 int
 143 mlx5_get_master_ifname(const char *ibdev_path, char (*ifname)[IF_NAMESIZE])
 144 {
 145         DIR *dir;
 146         struct dirent *dent;
 147         unsigned int dev_type = 0;
 148         unsigned int dev_port_prev = ~0u;
 149         char match[IF_NAMESIZE] = "";
 150
 151         MLX5_ASSERT(ibdev_path);
 152         {
 153                 MKSTR(path, "%s/device/net", ibdev_path);
 154
 155                 dir = opendir(path);
 156                 if (dir == NULL) {
 157                         rte_errno = errno;
 158                         return -rte_errno;
 159                 }
 160         }
 161         while ((dent = readdir(dir)) != NULL) {
 162                 char *name = dent->d_name;
 163                 FILE *file;
 164                 unsigned int dev_port;
 165                 int r;
 166
 167                 if ((name[0] == '.') &&
 168                     ((name[1] == '\0') ||
 169                      ((name[1] == '.') && (name[2] == '\0'))))
 170                         continue;
 171
 172                 MKSTR(path, "%s/device/net/%s/%s",
 173                       ibdev_path, name,
 174                       (dev_type ? "dev_id" : "dev_port"));
 175
 176                 file = fopen(path, "rb");
 177                 if (file == NULL) {
 178                         if (errno != ENOENT)
 179                                 continue;
 180                         /*
 181                          * Switch to dev_id when dev_port does not exist as
 182                          * is the case with Linux kernel versions < 3.15.
 183                          */
 184 try_dev_id:
 185                         match[0] = '\0';
 186                         if (dev_type)
 187                                 break;
 188                         dev_type = 1;
 189                         dev_port_prev = ~0u;
 190                         rewinddir(dir);
 191                         continue;
 192                 }
 193                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 194                 fclose(file);
 195                 if (r != 1)
 196                         continue;
 197                 /*
 198                  * Switch to dev_id when dev_port returns the same value for
 199                  * all ports. May happen when using a MOFED release older than
 200                  * 3.0 with a Linux kernel >= 3.15.
 201                  */
 202                 if (dev_port == dev_port_prev)
 203                         goto try_dev_id;
 204                 dev_port_prev = dev_port;
 205                 if (dev_port == 0)
 206                         strlcpy(match, name, sizeof(match));
 207         }
 208         closedir(dir);
 209         if (match[0] == '\0') {
 210                 rte_errno = ENOENT;
 211                 return -rte_errno;
 212         }
 213         strncpy(*ifname, match, sizeof(*ifname));
 214         return 0;
 215 }
 216
 217 /**
 218  * Get interface name from private structure.
 219  *
 220  * This is a port representor-aware version of mlx5_get_master_ifname().
 221  *
 222  * @param[in] dev
 223  *   Pointer to Ethernet device.
 224  * @param[out] ifname
 225  *   Interface name output buffer.
 226  *
 227  * @return
 228  *   0 on success, a negative errno value otherwise and rte_errno is set.
 229  */
 230 int
 231 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
 232 {
 233         struct mlx5_priv *priv = dev->data->dev_private;
 234         unsigned int ifindex;
 235
 236         MLX5_ASSERT(priv);
 237         MLX5_ASSERT(priv->sh);
 238         ifindex = mlx5_ifindex(dev);
 239         if (!ifindex) {
 240                 if (!priv->representor)
 241                         return mlx5_get_master_ifname(priv->sh->ibdev_path,
 242                                                       ifname);
 243                 rte_errno = ENXIO;
 244                 return -rte_errno;
 245         }
 246         if (if_indextoname(ifindex, &(*ifname)[0]))
 247                 return 0;
 248         rte_errno = errno;
 249         return -rte_errno;
 250 }
 251
 252 /**
 253  * Get the interface index from device name.
 254  *
 255  * @param[in] dev
 256  *   Pointer to Ethernet device.
 257  *
 258  * @return
 259  *   Nonzero interface index on success, zero otherwise and rte_errno is set.
 260  */
 261 unsigned int
 262 mlx5_ifindex(const struct rte_eth_dev *dev)
 263 {
 264         struct mlx5_priv *priv = dev->data->dev_private;
 265         unsigned int ifindex;
 266
 267         MLX5_ASSERT(priv);
 268         MLX5_ASSERT(priv->if_index);
 269         ifindex = priv->if_index;
 270         if (!ifindex)
 271                 rte_errno = ENXIO;
 272         return ifindex;
 273 }
 274
 275 /**
 276  * Perform ifreq ioctl() on associated Ethernet device.
 277  *
 278  * @param[in] dev
 279  *   Pointer to Ethernet device.
 280  * @param req
 281  *   Request number to pass to ioctl().
 282  * @param[out] ifr
 283  *   Interface request structure output buffer.
 284  *
 285  * @return
 286  *   0 on success, a negative errno value otherwise and rte_errno is set.
 287  */
 288 int
 289 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
 290 {
 291         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 292         int ret = 0;
 293
 294         if (sock == -1) {
 295                 rte_errno = errno;
 296                 return -rte_errno;
 297         }
 298         ret = mlx5_get_ifname(dev, &ifr->ifr_name);
 299         if (ret)
 300                 goto error;
 301         ret = ioctl(sock, req, ifr);
 302         if (ret == -1) {
 303                 rte_errno = errno;
 304                 goto error;
 305         }
 306         close(sock);
 307         return 0;
 308 error:
 309         close(sock);
 310         return -rte_errno;
 311 }
 312
 313 /**
 314  * Get device MTU.
 315  *
 316  * @param dev
 317  *   Pointer to Ethernet device.
 318  * @param[out] mtu
 319  *   MTU value output buffer.
 320  *
 321  * @return
 322  *   0 on success, a negative errno value otherwise and rte_errno is set.
 323  */
 324 int
 325 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
 326 {
 327         struct ifreq request;
 328         int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
 329
 330         if (ret)
 331                 return ret;
 332         *mtu = request.ifr_mtu;
 333         return 0;
 334 }
 335
 336 /**
 337  * Set device MTU.
 338  *
 339  * @param dev
 340  *   Pointer to Ethernet device.
 341  * @param mtu
 342  *   MTU value to set.
 343  *
 344  * @return
 345  *   0 on success, a negative errno value otherwise and rte_errno is set.
 346  */
 347 static int
 348 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 349 {
 350         struct ifreq request = { .ifr_mtu = mtu, };
 351
 352         return mlx5_ifreq(dev, SIOCSIFMTU, &request);
 353 }
 354
 355 /**
 356  * Set device flags.
 357  *
 358  * @param dev
 359  *   Pointer to Ethernet device.
 360  * @param keep
 361  *   Bitmask for flags that must remain untouched.
 362  * @param flags
 363  *   Bitmask for flags to modify.
 364  *
 365  * @return
 366  *   0 on success, a negative errno value otherwise and rte_errno is set.
 367  */
 368 int
 369 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
 370 {
 371         struct ifreq request;
 372         int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
 373
 374         if (ret)
 375                 return ret;
 376         request.ifr_flags &= keep;
 377         request.ifr_flags |= flags & ~keep;
 378         return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
 379 }
 380
 381 /**
 382  * DPDK callback for Ethernet device configuration.
 383  *
 384  * @param dev
 385  *   Pointer to Ethernet device structure.
 386  *
 387  * @return
 388  *   0 on success, a negative errno value otherwise and rte_errno is set.
 389  */
 390 int
 391 mlx5_dev_configure(struct rte_eth_dev *dev)
 392 {
 393         struct mlx5_priv *priv = dev->data->dev_private;
 394         unsigned int rxqs_n = dev->data->nb_rx_queues;
 395         unsigned int txqs_n = dev->data->nb_tx_queues;
 396         const uint8_t use_app_rss_key =
 397                 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
 398         int ret = 0;
 399
 400         if (use_app_rss_key &&
 401             (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
 402              MLX5_RSS_HASH_KEY_LEN)) {
 403                 DRV_LOG(ERR, "port %u RSS key len must be %s Bytes long",
 404                         dev->data->port_id, RTE_STR(MLX5_RSS_HASH_KEY_LEN));
 405                 rte_errno = EINVAL;
 406                 return -rte_errno;
 407         }
 408         priv->rss_conf.rss_key =
 409                 rte_realloc(priv->rss_conf.rss_key,
 410                             MLX5_RSS_HASH_KEY_LEN, 0);
 411         if (!priv->rss_conf.rss_key) {
 412                 DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
 413                         dev->data->port_id, rxqs_n);
 414                 rte_errno = ENOMEM;
 415                 return -rte_errno;
 416         }
 417
 418         if (dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
 419                 dev->data->dev_conf.rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH;
 420
 421         memcpy(priv->rss_conf.rss_key,
 422                use_app_rss_key ?
 423                dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
 424                rss_hash_default_key,
 425                MLX5_RSS_HASH_KEY_LEN);
 426         priv->rss_conf.rss_key_len = MLX5_RSS_HASH_KEY_LEN;
 427         priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
 428         priv->rxqs = (void *)dev->data->rx_queues;
 429         priv->txqs = (void *)dev->data->tx_queues;
 430         if (txqs_n != priv->txqs_n) {
 431                 DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
 432                         dev->data->port_id, priv->txqs_n, txqs_n);
 433                 priv->txqs_n = txqs_n;
 434         }
 435         if (rxqs_n > priv->config.ind_table_max_size) {
 436                 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
 437                         dev->data->port_id, rxqs_n);
 438                 rte_errno = EINVAL;
 439                 return -rte_errno;
 440         }
 441         if (rxqs_n != priv->rxqs_n) {
 442                 DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
 443                         dev->data->port_id, priv->rxqs_n, rxqs_n);
 444                 priv->rxqs_n = rxqs_n;
 445         }
 446         priv->skip_default_rss_reta = 0;
 447         ret = mlx5_proc_priv_init(dev);
 448         if (ret)
 449                 return ret;
 450         return 0;
 451 }
 452
 453 /**
 454  * Configure default RSS reta.
 455  *
 456  * @param dev
 457  *   Pointer to Ethernet device structure.
 458  *
 459  * @return
 460  *   0 on success, a negative errno value otherwise and rte_errno is set.
 461  */
 462 int
 463 mlx5_dev_configure_rss_reta(struct rte_eth_dev *dev)
 464 {
 465         struct mlx5_priv *priv = dev->data->dev_private;
 466         unsigned int rxqs_n = dev->data->nb_rx_queues;
 467         unsigned int i;
 468         unsigned int j;
 469         unsigned int reta_idx_n;
 470         int ret = 0;
 471         unsigned int *rss_queue_arr = NULL;
 472         unsigned int rss_queue_n = 0;
 473
 474         if (priv->skip_default_rss_reta)
 475                 return ret;
 476         rss_queue_arr = rte_malloc("", rxqs_n * sizeof(unsigned int), 0);
 477         if (!rss_queue_arr) {
 478                 DRV_LOG(ERR, "port %u cannot allocate RSS queue list (%u)",
 479                         dev->data->port_id, rxqs_n);
 480                 rte_errno = ENOMEM;
 481                 return -rte_errno;
 482         }
 483         for (i = 0, j = 0; i < rxqs_n; i++) {
 484                 struct mlx5_rxq_data *rxq_data;
 485                 struct mlx5_rxq_ctrl *rxq_ctrl;
 486
 487                 rxq_data = (*priv->rxqs)[i];
 488                 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 489                 if (rxq_ctrl && rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
 490                         rss_queue_arr[j++] = i;
 491         }
 492         rss_queue_n = j;
 493         if (rss_queue_n > priv->config.ind_table_max_size) {
 494                 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
 495                         dev->data->port_id, rss_queue_n);
 496                 rte_errno = EINVAL;
 497                 rte_free(rss_queue_arr);
 498                 return -rte_errno;
 499         }
 500         DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
 501                 dev->data->port_id, priv->rxqs_n, rxqs_n);
 502         priv->rxqs_n = rxqs_n;
 503         /*
 504          * If the requested number of RX queues is not a power of two,
 505          * use the maximum indirection table size for better balancing.
 506          * The result is always rounded to the next power of two.
 507          */
 508         reta_idx_n = (1 << log2above((rss_queue_n & (rss_queue_n - 1)) ?
 509                                 priv->config.ind_table_max_size :
 510                                 rss_queue_n));
 511         ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
 512         if (ret) {
 513                 rte_free(rss_queue_arr);
 514                 return ret;
 515         }
 516         /*
 517          * When the number of RX queues is not a power of two,
 518          * the remaining table entries are padded with reused WQs
 519          * and hashes are not spread uniformly.
 520          */
 521         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
 522                 (*priv->reta_idx)[i] = rss_queue_arr[j];
 523                 if (++j == rss_queue_n)
 524                         j = 0;
 525         }
 526         rte_free(rss_queue_arr);
 527         return ret;
 528 }
 529
 530 /**
 531  * Sets default tuning parameters.
 532  *
 533  * @param dev
 534  *   Pointer to Ethernet device.
 535  * @param[out] info
 536  *   Info structure output buffer.
 537  */
 538 static void
 539 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 540 {
 541         struct mlx5_priv *priv = dev->data->dev_private;
 542
 543         /* Minimum CPU utilization. */
 544         info->default_rxportconf.ring_size = 256;
 545         info->default_txportconf.ring_size = 256;
 546         info->default_rxportconf.burst_size = MLX5_RX_DEFAULT_BURST;
 547         info->default_txportconf.burst_size = MLX5_TX_DEFAULT_BURST;
 548         if ((priv->link_speed_capa & ETH_LINK_SPEED_200G) |
 549                 (priv->link_speed_capa & ETH_LINK_SPEED_100G)) {
 550                 info->default_rxportconf.nb_queues = 16;
 551                 info->default_txportconf.nb_queues = 16;
 552                 if (dev->data->nb_rx_queues > 2 ||
 553                     dev->data->nb_tx_queues > 2) {
 554                         /* Max Throughput. */
 555                         info->default_rxportconf.ring_size = 2048;
 556                         info->default_txportconf.ring_size = 2048;
 557                 }
 558         } else {
 559                 info->default_rxportconf.nb_queues = 8;
 560                 info->default_txportconf.nb_queues = 8;
 561                 if (dev->data->nb_rx_queues > 2 ||
 562                     dev->data->nb_tx_queues > 2) {
 563                         /* Max Throughput. */
 564                         info->default_rxportconf.ring_size = 4096;
 565                         info->default_txportconf.ring_size = 4096;
 566                 }
 567         }
 568 }
 569
 570 /**
 571  * Sets tx mbuf limiting parameters.
 572  *
 573  * @param dev
 574  *   Pointer to Ethernet device.
 575  * @param[out] info
 576  *   Info structure output buffer.
 577  */
 578 static void
 579 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 580 {
 581         struct mlx5_priv *priv = dev->data->dev_private;
 582         struct mlx5_dev_config *config = &priv->config;
 583         unsigned int inlen;
 584         uint16_t nb_max;
 585
 586         inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
 587                 MLX5_SEND_DEF_INLINE_LEN :
 588                 (unsigned int)config->txq_inline_max;
 589         MLX5_ASSERT(config->txq_inline_min >= 0);
 590         inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
 591         inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
 592                                MLX5_ESEG_MIN_INLINE_SIZE -
 593                                MLX5_WQE_CSEG_SIZE -
 594                                MLX5_WQE_ESEG_SIZE -
 595                                MLX5_WQE_DSEG_SIZE * 2);
 596         nb_max = (MLX5_WQE_SIZE_MAX +
 597                   MLX5_ESEG_MIN_INLINE_SIZE -
 598                   MLX5_WQE_CSEG_SIZE -
 599                   MLX5_WQE_ESEG_SIZE -
 600                   MLX5_WQE_DSEG_SIZE -
 601                   inlen) / MLX5_WSEG_SIZE;
 602         info->tx_desc_lim.nb_seg_max = nb_max;
 603         info->tx_desc_lim.nb_mtu_seg_max = nb_max;
 604 }
 605
 606 /**
 607  * DPDK callback to get information about the device.
 608  *
 609  * @param dev
 610  *   Pointer to Ethernet device structure.
 611  * @param[out] info
 612  *   Info structure output buffer.
 613  */
 614 int
 615 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 616 {
 617         struct mlx5_priv *priv = dev->data->dev_private;
 618         struct mlx5_dev_config *config = &priv->config;
 619         unsigned int max;
 620
 621         /* FIXME: we should ask the device for these values. */
 622         info->min_rx_bufsize = 32;
 623         info->max_rx_pktlen = 65536;
 624         info->max_lro_pkt_size = MLX5_MAX_LRO_SIZE;
 625         /*
 626          * Since we need one CQ per QP, the limit is the minimum number
 627          * between the two values.
 628          */
 629         max = RTE_MIN(priv->sh->device_attr.orig_attr.max_cq,
 630                       priv->sh->device_attr.orig_attr.max_qp);
 631         /* max_rx_queues is uint16_t. */
 632         max = RTE_MIN(max, (unsigned int)UINT16_MAX);
 633         info->max_rx_queues = max;
 634         info->max_tx_queues = max;
 635         info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
 636         info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
 637         info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
 638                                  info->rx_queue_offload_capa);
 639         info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
 640         info->if_index = mlx5_ifindex(dev);
 641         info->reta_size = priv->reta_idx_n ?
 642                 priv->reta_idx_n : config->ind_table_max_size;
 643         info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
 644         info->speed_capa = priv->link_speed_capa;
 645         info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
 646         mlx5_set_default_params(dev, info);
 647         mlx5_set_txlimit_params(dev, info);
 648         info->switch_info.name = dev->data->name;
 649         info->switch_info.domain_id = priv->domain_id;
 650         info->switch_info.port_id = priv->representor_id;
 651         if (priv->representor) {
 652                 uint16_t port_id;
 653
 654                 if (priv->pf_bond >= 0) {
 655                         /*
 656                          * Switch port ID is opaque value with driver defined
 657                          * format. Push the PF index in bonding configurations
 658                          * in upper four bits of port ID. If we get too many
 659                          * representors (more than 4K) or PFs (more than 15)
 660                          * this approach must be reconsidered.
 661                          */
 662                         if ((info->switch_info.port_id >>
 663                                 MLX5_PORT_ID_BONDING_PF_SHIFT) ||
 664                             priv->pf_bond > MLX5_PORT_ID_BONDING_PF_MASK) {
 665                                 DRV_LOG(ERR, "can't update switch port ID"
 666                                              " for bonding device");
 667                                 MLX5_ASSERT(false);
 668                                 return -ENODEV;
 669                         }
 670                         info->switch_info.port_id |=
 671                                 priv->pf_bond << MLX5_PORT_ID_BONDING_PF_SHIFT;
 672                 }
 673                 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 674                         struct mlx5_priv *opriv =
 675                                 rte_eth_devices[port_id].data->dev_private;
 676
 677                         if (!opriv ||
 678                             opriv->representor ||
 679                             opriv->sh != priv->sh ||
 680                             opriv->domain_id != priv->domain_id)
 681                                 continue;
 682                         /*
 683                          * Override switch name with that of the master
 684                          * device.
 685                          */
 686                         info->switch_info.name = opriv->dev_data->name;
 687                         break;
 688                 }
 689         }
 690         return 0;
 691 }
 692
 693 /**
 694  * Get device current raw clock counter
 695  *
 696  * @param dev
 697  *   Pointer to Ethernet device structure.
 698  * @param[out] time
 699  *   Current raw clock counter of the device.
 700  *
 701  * @return
 702  *   0 if the clock has correctly been read
 703  *   The value of errno in case of error
 704  */
 705 int
 706 mlx5_read_clock(struct rte_eth_dev *dev, uint64_t *clock)
 707 {
 708         struct mlx5_priv *priv = dev->data->dev_private;
 709         struct ibv_context *ctx = priv->sh->ctx;
 710         struct ibv_values_ex values;
 711         int err = 0;
 712
 713         values.comp_mask = IBV_VALUES_MASK_RAW_CLOCK;
 714         err = mlx5_glue->query_rt_values_ex(ctx, &values);
 715         if (err != 0) {
 716                 DRV_LOG(WARNING, "Could not query the clock !");
 717                 return err;
 718         }
 719         *clock = values.raw_clock.tv_nsec;
 720         return 0;
 721 }
 722
 723 /**
 724  * Get firmware version of a device.
 725  *
 726  * @param dev
 727  *   Ethernet device port.
 728  * @param fw_ver
 729  *   String output allocated by caller.
 730  * @param fw_size
 731  *   Size of the output string, including terminating null byte.
 732  *
 733  * @return
 734  *   0 on success, or the size of the non truncated string if too big.
 735  */
 736 int mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
 737 {
 738         struct mlx5_priv *priv = dev->data->dev_private;
 739         struct ibv_device_attr *attr = &priv->sh->device_attr.orig_attr;
 740         size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
 741
 742         if (fw_size < size)
 743                 return size;
 744         if (fw_ver != NULL)
 745                 strlcpy(fw_ver, attr->fw_ver, fw_size);
 746         return 0;
 747 }
 748
 749 /**
 750  * Get supported packet types.
 751  *
 752  * @param dev
 753  *   Pointer to Ethernet device structure.
 754  *
 755  * @return
 756  *   A pointer to the supported Packet types array.
 757  */
 758 const uint32_t *
 759 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 760 {
 761         static const uint32_t ptypes[] = {
 762                 /* refers to rxq_cq_to_pkt_type() */
 763                 RTE_PTYPE_L2_ETHER,
 764                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 765                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 766                 RTE_PTYPE_L4_NONFRAG,
 767                 RTE_PTYPE_L4_FRAG,
 768                 RTE_PTYPE_L4_TCP,
 769                 RTE_PTYPE_L4_UDP,
 770                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 771                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 772                 RTE_PTYPE_INNER_L4_NONFRAG,
 773                 RTE_PTYPE_INNER_L4_FRAG,
 774                 RTE_PTYPE_INNER_L4_TCP,
 775                 RTE_PTYPE_INNER_L4_UDP,
 776                 RTE_PTYPE_UNKNOWN
 777         };
 778
 779         if (dev->rx_pkt_burst == mlx5_rx_burst ||
 780             dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
 781             dev->rx_pkt_burst == mlx5_rx_burst_vec)
 782                 return ptypes;
 783         return NULL;
 784 }
 785
 786 /**
 787  * Retrieve the master device for representor in the same switch domain.
 788  *
 789  * @param dev
 790  *   Pointer to representor Ethernet device structure.
 791  *
 792  * @return
 793  *   Master device structure  on success, NULL otherwise.
 794  */
 795
 796 static struct rte_eth_dev *
 797 mlx5_find_master_dev(struct rte_eth_dev *dev)
 798 {
 799         struct mlx5_priv *priv;
 800         uint16_t port_id;
 801         uint16_t domain_id;
 802
 803         priv = dev->data->dev_private;
 804         domain_id = priv->domain_id;
 805         MLX5_ASSERT(priv->representor);
 806         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 807                 struct mlx5_priv *opriv =
 808                         rte_eth_devices[port_id].data->dev_private;
 809                 if (opriv &&
 810                     opriv->master &&
 811                     opriv->domain_id == domain_id &&
 812                     opriv->sh == priv->sh)
 813                         return &rte_eth_devices[port_id];
 814         }
 815         return NULL;
 816 }
 817
 818 /**
 819  * DPDK callback to retrieve physical link information.
 820  *
 821  * @param dev
 822  *   Pointer to Ethernet device structure.
 823  * @param[out] link
 824  *   Storage for current link status.
 825  *
 826  * @return
 827  *   0 on success, a negative errno value otherwise and rte_errno is set.
 828  */
 829 static int
 830 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev,
 831                                struct rte_eth_link *link)
 832 {
 833         struct mlx5_priv *priv = dev->data->dev_private;
 834         struct ethtool_cmd edata = {
 835                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 836         };
 837         struct ifreq ifr;
 838         struct rte_eth_link dev_link;
 839         int link_speed = 0;
 840         int ret;
 841
 842         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 843         if (ret) {
 844                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 845                         dev->data->port_id, strerror(rte_errno));
 846                 return ret;
 847         }
 848         dev_link = (struct rte_eth_link) {
 849                 .link_status = ((ifr.ifr_flags & IFF_UP) &&
 850                                 (ifr.ifr_flags & IFF_RUNNING)),
 851         };
 852         ifr = (struct ifreq) {
 853                 .ifr_data = (void *)&edata,
 854         };
 855         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 856         if (ret) {
 857                 if (ret == -ENOTSUP && priv->representor) {
 858                         struct rte_eth_dev *master;
 859
 860                         /*
 861                          * For representors we can try to inherit link
 862                          * settings from the master device. Actually
 863                          * link settings do not make a lot of sense
 864                          * for representors due to missing physical
 865                          * link. The old kernel drivers supported
 866                          * emulated settings query for representors,
 867                          * the new ones do not, so we have to add
 868                          * this code for compatibility issues.
 869                          */
 870                         master = mlx5_find_master_dev(dev);
 871                         if (master) {
 872                                 ifr = (struct ifreq) {
 873                                         .ifr_data = (void *)&edata,
 874                                 };
 875                                 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
 876                         }
 877                 }
 878                 if (ret) {
 879                         DRV_LOG(WARNING,
 880                                 "port %u ioctl(SIOCETHTOOL,"
 881                                 " ETHTOOL_GSET) failed: %s",
 882                                 dev->data->port_id, strerror(rte_errno));
 883                         return ret;
 884                 }
 885         }
 886         link_speed = ethtool_cmd_speed(&edata);
 887         if (link_speed == -1)
 888                 dev_link.link_speed = ETH_SPEED_NUM_NONE;
 889         else
 890                 dev_link.link_speed = link_speed;
 891         priv->link_speed_capa = 0;
 892         if (edata.supported & SUPPORTED_Autoneg)
 893                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 894         if (edata.supported & (SUPPORTED_1000baseT_Full |
 895                                SUPPORTED_1000baseKX_Full))
 896                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 897         if (edata.supported & SUPPORTED_10000baseKR_Full)
 898                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 899         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 900                                SUPPORTED_40000baseCR4_Full |
 901                                SUPPORTED_40000baseSR4_Full |
 902                                SUPPORTED_40000baseLR4_Full))
 903                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 904         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 905                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 906         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 907                         ETH_LINK_SPEED_FIXED);
 908         if (((dev_link.link_speed && !dev_link.link_status) ||
 909              (!dev_link.link_speed && dev_link.link_status))) {
 910                 rte_errno = EAGAIN;
 911                 return -rte_errno;
 912         }
 913         *link = dev_link;
 914         return 0;
 915 }
 916
 917 /**
 918  * Retrieve physical link information (unlocked version using new ioctl).
 919  *
 920  * @param dev
 921  *   Pointer to Ethernet device structure.
 922  * @param[out] link
 923  *   Storage for current link status.
 924  *
 925  * @return
 926  *   0 on success, a negative errno value otherwise and rte_errno is set.
 927  */
 928 static int
 929 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev,
 930                              struct rte_eth_link *link)
 931
 932 {
 933         struct mlx5_priv *priv = dev->data->dev_private;
 934         struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
 935         struct ifreq ifr;
 936         struct rte_eth_link dev_link;
 937         struct rte_eth_dev *master = NULL;
 938         uint64_t sc;
 939         int ret;
 940
 941         ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr);
 942         if (ret) {
 943                 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s",
 944                         dev->data->port_id, strerror(rte_errno));
 945                 return ret;
 946         }
 947         dev_link = (struct rte_eth_link) {
 948                 .link_status = ((ifr.ifr_flags & IFF_UP) &&
 949                                 (ifr.ifr_flags & IFF_RUNNING)),
 950         };
 951         ifr = (struct ifreq) {
 952                 .ifr_data = (void *)&gcmd,
 953         };
 954         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
 955         if (ret) {
 956                 if (ret == -ENOTSUP && priv->representor) {
 957                         /*
 958                          * For representors we can try to inherit link
 959                          * settings from the master device. Actually
 960                          * link settings do not make a lot of sense
 961                          * for representors due to missing physical
 962                          * link. The old kernel drivers supported
 963                          * emulated settings query for representors,
 964                          * the new ones do not, so we have to add
 965                          * this code for compatibility issues.
 966                          */
 967                         master = mlx5_find_master_dev(dev);
 968                         if (master) {
 969                                 ifr = (struct ifreq) {
 970                                         .ifr_data = (void *)&gcmd,
 971                                 };
 972                                 ret = mlx5_ifreq(master, SIOCETHTOOL, &ifr);
 973                         }
 974                 }
 975                 if (ret) {
 976                         DRV_LOG(DEBUG,
 977                                 "port %u ioctl(SIOCETHTOOL,"
 978                                 " ETHTOOL_GLINKSETTINGS) failed: %s",
 979                                 dev->data->port_id, strerror(rte_errno));
 980                         return ret;
 981                 }
 982
 983         }
 984         gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
 985
 986         alignas(struct ethtool_link_settings)
 987         uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
 988                      sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
 989         struct ethtool_link_settings *ecmd = (void *)data;
 990
 991         *ecmd = gcmd;
 992         ifr.ifr_data = (void *)ecmd;
 993         ret = mlx5_ifreq(master ? master : dev, SIOCETHTOOL, &ifr);
 994         if (ret) {
 995                 DRV_LOG(DEBUG,
 996                         "port %u ioctl(SIOCETHTOOL,"
 997                         "ETHTOOL_GLINKSETTINGS) failed: %s",
 998                         dev->data->port_id, strerror(rte_errno));
 999                 return ret;
1000         }
1001         dev_link.link_speed = (ecmd->speed == UINT32_MAX) ? ETH_SPEED_NUM_NONE :
1002                                                             ecmd->speed;
1003         sc = ecmd->link_mode_masks[0] |
1004                 ((uint64_t)ecmd->link_mode_masks[1] << 32);
1005         priv->link_speed_capa = 0;
1006         if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
1007                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
1008         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
1009                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
1010                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
1011         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
1012                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
1013                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
1014                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
1015         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
1016                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
1017                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
1018         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
1019                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
1020                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
1021                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
1022                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
1023         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
1024                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
1025                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
1026                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
1027                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
1028         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
1029                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
1030                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
1031                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
1032         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
1033                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
1034                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
1035         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
1036                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
1037                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
1038                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
1039                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
1040         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_200000baseKR4_Full_BIT) |
1041                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_200000baseSR4_Full_BIT)))
1042                 priv->link_speed_capa |= ETH_LINK_SPEED_200G;
1043
1044         sc = ecmd->link_mode_masks[2] |
1045                 ((uint64_t)ecmd->link_mode_masks[3] << 32);
1046         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_200000baseCR4_Full_BIT) |
1047                   MLX5_BITSHIFT(
1048                         ETHTOOL_LINK_MODE_200000baseLR4_ER4_FR4_Full_BIT) |
1049                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_200000baseDR4_Full_BIT)))
1050                 priv->link_speed_capa |= ETH_LINK_SPEED_200G;
1051         dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
1052                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
1053         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
1054                                   ETH_LINK_SPEED_FIXED);
1055         if (((dev_link.link_speed && !dev_link.link_status) ||
1056              (!dev_link.link_speed && dev_link.link_status))) {
1057                 rte_errno = EAGAIN;
1058                 return -rte_errno;
1059         }
1060         *link = dev_link;
1061         return 0;
1062 }
1063
1064 /**
1065  * DPDK callback to retrieve physical link information.
1066  *
1067  * @param dev
1068  *   Pointer to Ethernet device structure.
1069  * @param wait_to_complete
1070  *   Wait for request completion.
1071  *
1072  * @return
1073  *   0 if link status was not updated, positive if it was, a negative errno
1074  *   value otherwise and rte_errno is set.
1075  */
1076 int
1077 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
1078 {
1079         int ret;
1080         struct rte_eth_link dev_link;
1081         time_t start_time = time(NULL);
1082         int retry = MLX5_GET_LINK_STATUS_RETRY_COUNT;
1083
1084         do {
1085                 ret = mlx5_link_update_unlocked_gs(dev, &dev_link);
1086                 if (ret == -ENOTSUP)
1087                         ret = mlx5_link_update_unlocked_gset(dev, &dev_link);
1088                 if (ret == 0)
1089                         break;
1090                 /* Handle wait to complete situation. */
1091                 if ((wait_to_complete || retry) && ret == -EAGAIN) {
1092                         if (abs((int)difftime(time(NULL), start_time)) <
1093                             MLX5_LINK_STATUS_TIMEOUT) {
1094                                 usleep(0);
1095                                 continue;
1096                         } else {
1097                                 rte_errno = EBUSY;
1098                                 return -rte_errno;
1099                         }
1100                 } else if (ret < 0) {
1101                         return ret;
1102                 }
1103         } while (wait_to_complete || retry-- > 0);
1104         ret = !!memcmp(&dev->data->dev_link, &dev_link,
1105                        sizeof(struct rte_eth_link));
1106         dev->data->dev_link = dev_link;
1107         return ret;
1108 }
1109
1110 /**
1111  * DPDK callback to change the MTU.
1112  *
1113  * @param dev
1114  *   Pointer to Ethernet device structure.
1115  * @param in_mtu
1116  *   New MTU.
1117  *
1118  * @return
1119  *   0 on success, a negative errno value otherwise and rte_errno is set.
1120  */
1121 int
1122 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
1123 {
1124         struct mlx5_priv *priv = dev->data->dev_private;
1125         uint16_t kern_mtu = 0;
1126         int ret;
1127
1128         ret = mlx5_get_mtu(dev, &kern_mtu);
1129         if (ret)
1130                 return ret;
1131         /* Set kernel interface MTU first. */
1132         ret = mlx5_set_mtu(dev, mtu);
1133         if (ret)
1134                 return ret;
1135         ret = mlx5_get_mtu(dev, &kern_mtu);
1136         if (ret)
1137                 return ret;
1138         if (kern_mtu == mtu) {
1139                 priv->mtu = mtu;
1140                 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
1141                         dev->data->port_id, mtu);
1142                 return 0;
1143         }
1144         rte_errno = EAGAIN;
1145         return -rte_errno;
1146 }
1147
1148 /**
1149  * DPDK callback to get flow control status.
1150  *
1151  * @param dev
1152  *   Pointer to Ethernet device structure.
1153  * @param[out] fc_conf
1154  *   Flow control output buffer.
1155  *
1156  * @return
1157  *   0 on success, a negative errno value otherwise and rte_errno is set.
1158  */
1159 int
1160 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1161 {
1162         struct ifreq ifr;
1163         struct ethtool_pauseparam ethpause = {
1164                 .cmd = ETHTOOL_GPAUSEPARAM
1165         };
1166         int ret;
1167
1168         ifr.ifr_data = (void *)&ethpause;
1169         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1170         if (ret) {
1171                 DRV_LOG(WARNING,
1172                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:"
1173                         " %s",
1174                         dev->data->port_id, strerror(rte_errno));
1175                 return ret;
1176         }
1177         fc_conf->autoneg = ethpause.autoneg;
1178         if (ethpause.rx_pause && ethpause.tx_pause)
1179                 fc_conf->mode = RTE_FC_FULL;
1180         else if (ethpause.rx_pause)
1181                 fc_conf->mode = RTE_FC_RX_PAUSE;
1182         else if (ethpause.tx_pause)
1183                 fc_conf->mode = RTE_FC_TX_PAUSE;
1184         else
1185                 fc_conf->mode = RTE_FC_NONE;
1186         return 0;
1187 }
1188
1189 /**
1190  * DPDK callback to modify flow control parameters.
1191  *
1192  * @param dev
1193  *   Pointer to Ethernet device structure.
1194  * @param[in] fc_conf
1195  *   Flow control parameters.
1196  *
1197  * @return
1198  *   0 on success, a negative errno value otherwise and rte_errno is set.
1199  */
1200 int
1201 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1202 {
1203         struct ifreq ifr;
1204         struct ethtool_pauseparam ethpause = {
1205                 .cmd = ETHTOOL_SPAUSEPARAM
1206         };
1207         int ret;
1208
1209         ifr.ifr_data = (void *)&ethpause;
1210         ethpause.autoneg = fc_conf->autoneg;
1211         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1212             (fc_conf->mode & RTE_FC_RX_PAUSE))
1213                 ethpause.rx_pause = 1;
1214         else
1215                 ethpause.rx_pause = 0;
1216
1217         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1218             (fc_conf->mode & RTE_FC_TX_PAUSE))
1219                 ethpause.tx_pause = 1;
1220         else
1221                 ethpause.tx_pause = 0;
1222         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1223         if (ret) {
1224                 DRV_LOG(WARNING,
1225                         "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1226                         " failed: %s",
1227                         dev->data->port_id, strerror(rte_errno));
1228                 return ret;
1229         }
1230         return 0;
1231 }
1232
1233 /**
1234  * Handle asynchronous removal event for entire multiport device.
1235  *
1236  * @param sh
1237  *   Infiniband device shared context.
1238  */
1239 static void
1240 mlx5_dev_interrupt_device_fatal(struct mlx5_ibv_shared *sh)
1241 {
1242         uint32_t i;
1243
1244         for (i = 0; i < sh->max_port; ++i) {
1245                 struct rte_eth_dev *dev;
1246
1247                 if (sh->port[i].ih_port_id >= RTE_MAX_ETHPORTS) {
1248                         /*
1249                          * Or not existing port either no
1250                          * handler installed for this port.
1251                          */
1252                         continue;
1253                 }
1254                 dev = &rte_eth_devices[sh->port[i].ih_port_id];
1255                 MLX5_ASSERT(dev);
1256                 if (dev->data->dev_conf.intr_conf.rmv)
1257                         _rte_eth_dev_callback_process
1258                                 (dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1259         }
1260 }
1261
1262 /**
1263  * Handle shared asynchronous events the NIC (removal event
1264  * and link status change). Supports multiport IB device.
1265  *
1266  * @param cb_arg
1267  *   Callback argument.
1268  */
1269 void
1270 mlx5_dev_interrupt_handler(void *cb_arg)
1271 {
1272         struct mlx5_ibv_shared *sh = cb_arg;
1273         struct ibv_async_event event;
1274
1275         /* Read all message from the IB device and acknowledge them. */
1276         for (;;) {
1277                 struct rte_eth_dev *dev;
1278                 uint32_t tmp;
1279
1280                 if (mlx5_glue->get_async_event(sh->ctx, &event))
1281                         break;
1282                 /* Retrieve and check IB port index. */
1283                 tmp = (uint32_t)event.element.port_num;
1284                 if (!tmp && event.event_type == IBV_EVENT_DEVICE_FATAL) {
1285                         /*
1286                          * The DEVICE_FATAL event is called once for
1287                          * entire device without port specifying.
1288                          * We should notify all existing ports.
1289                          */
1290                         mlx5_glue->ack_async_event(&event);
1291                         mlx5_dev_interrupt_device_fatal(sh);
1292                         continue;
1293                 }
1294                 MLX5_ASSERT(tmp && (tmp <= sh->max_port));
1295                 if (!tmp) {
1296                         /* Unsupported devive level event. */
1297                         mlx5_glue->ack_async_event(&event);
1298                         DRV_LOG(DEBUG,
1299                                 "unsupported common event (type %d)",
1300                                 event.event_type);
1301                         continue;
1302                 }
1303                 if (tmp > sh->max_port) {
1304                         /* Invalid IB port index. */
1305                         mlx5_glue->ack_async_event(&event);
1306                         DRV_LOG(DEBUG,
1307                                 "cannot handle an event (type %d)"
1308                                 "due to invalid IB port index (%u)",
1309                                 event.event_type, tmp);
1310                         continue;
1311                 }
1312                 if (sh->port[tmp - 1].ih_port_id >= RTE_MAX_ETHPORTS) {
1313                         /* No handler installed. */
1314                         mlx5_glue->ack_async_event(&event);
1315                         DRV_LOG(DEBUG,
1316                                 "cannot handle an event (type %d)"
1317                                 "due to no handler installed for port %u",
1318                                 event.event_type, tmp);
1319                         continue;
1320                 }
1321                 /* Retrieve ethernet device descriptor. */
1322                 tmp = sh->port[tmp - 1].ih_port_id;
1323                 dev = &rte_eth_devices[tmp];
1324                 MLX5_ASSERT(dev);
1325                 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1326                      event.event_type == IBV_EVENT_PORT_ERR) &&
1327                         dev->data->dev_conf.intr_conf.lsc) {
1328                         mlx5_glue->ack_async_event(&event);
1329                         if (mlx5_link_update(dev, 0) == -EAGAIN) {
1330                                 usleep(0);
1331                                 continue;
1332                         }
1333                         _rte_eth_dev_callback_process
1334                                 (dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1335                         continue;
1336                 }
1337                 DRV_LOG(DEBUG,
1338                         "port %u cannot handle an unknown event (type %d)",
1339                         dev->data->port_id, event.event_type);
1340                 mlx5_glue->ack_async_event(&event);
1341         }
1342 }
1343
1344 /*
1345  * Unregister callback handler safely. The handler may be active
1346  * while we are trying to unregister it, in this case code -EAGAIN
1347  * is returned by rte_intr_callback_unregister(). This routine checks
1348  * the return code and tries to unregister handler again.
1349  *
1350  * @param handle
1351  *   interrupt handle
1352  * @param cb_fn
1353  *   pointer to callback routine
1354  * @cb_arg
1355  *   opaque callback parameter
1356  */
1357 void
1358 mlx5_intr_callback_unregister(const struct rte_intr_handle *handle,
1359                               rte_intr_callback_fn cb_fn, void *cb_arg)
1360 {
1361         /*
1362          * Try to reduce timeout management overhead by not calling
1363          * the timer related routines on the first iteration. If the
1364          * unregistering succeeds on first call there will be no
1365          * timer calls at all.
1366          */
1367         uint64_t twait = 0;
1368         uint64_t start = 0;
1369
1370         do {
1371                 int ret;
1372
1373                 ret = rte_intr_callback_unregister(handle, cb_fn, cb_arg);
1374                 if (ret >= 0)
1375                         return;
1376                 if (ret != -EAGAIN) {
1377                         DRV_LOG(INFO, "failed to unregister interrupt"
1378                                       " handler (error: %d)", ret);
1379                         MLX5_ASSERT(false);
1380                         return;
1381                 }
1382                 if (twait) {
1383                         struct timespec onems;
1384
1385                         /* Wait one millisecond and try again. */
1386                         onems.tv_sec = 0;
1387                         onems.tv_nsec = NS_PER_S / MS_PER_S;
1388                         nanosleep(&onems, 0);
1389                         /* Check whether one second elapsed. */
1390                         if ((rte_get_timer_cycles() - start) <= twait)
1391                                 continue;
1392                 } else {
1393                         /*
1394                          * We get the amount of timer ticks for one second.
1395                          * If this amount elapsed it means we spent one
1396                          * second in waiting. This branch is executed once
1397                          * on first iteration.
1398                          */
1399                         twait = rte_get_timer_hz();
1400                         MLX5_ASSERT(twait);
1401                 }
1402                 /*
1403                  * Timeout elapsed, show message (once a second) and retry.
1404                  * We have no other acceptable option here, if we ignore
1405                  * the unregistering return code the handler will not
1406                  * be unregistered, fd will be closed and we may get the
1407                  * crush. Hanging and messaging in the loop seems not to be
1408                  * the worst choice.
1409                  */
1410                 DRV_LOG(INFO, "Retrying to unregister interrupt handler");
1411                 start = rte_get_timer_cycles();
1412         } while (true);
1413 }
1414
1415 /**
1416  * Handle DEVX interrupts from the NIC.
1417  * This function is probably called from the DPDK host thread.
1418  *
1419  * @param cb_arg
1420  *   Callback argument.
1421  */
1422 void
1423 mlx5_dev_interrupt_handler_devx(void *cb_arg)
1424 {
1425 #ifndef HAVE_IBV_DEVX_ASYNC
1426         (void)cb_arg;
1427         return;
1428 #else
1429         struct mlx5_ibv_shared *sh = cb_arg;
1430         union {
1431                 struct mlx5dv_devx_async_cmd_hdr cmd_resp;
1432                 uint8_t buf[MLX5_ST_SZ_BYTES(query_flow_counter_out) +
1433                             MLX5_ST_SZ_BYTES(traffic_counter) +
1434                             sizeof(struct mlx5dv_devx_async_cmd_hdr)];
1435         } out;
1436         uint8_t *buf = out.buf + sizeof(out.cmd_resp);
1437
1438         while (!mlx5_glue->devx_get_async_cmd_comp(sh->devx_comp,
1439                                                    &out.cmd_resp,
1440                                                    sizeof(out.buf)))
1441                 mlx5_flow_async_pool_query_handle
1442                         (sh, (uint64_t)out.cmd_resp.wr_id,
1443                          mlx5_devx_get_out_command_status(buf));
1444 #endif /* HAVE_IBV_DEVX_ASYNC */
1445 }
1446
1447 /**
1448  * Uninstall shared asynchronous device events handler.
1449  * This function is implemented to support event sharing
1450  * between multiple ports of single IB device.
1451  *
1452  * @param dev
1453  *   Pointer to Ethernet device.
1454  */
1455 static void
1456 mlx5_dev_shared_handler_uninstall(struct rte_eth_dev *dev)
1457 {
1458         struct mlx5_priv *priv = dev->data->dev_private;
1459         struct mlx5_ibv_shared *sh = priv->sh;
1460
1461         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1462                 return;
1463         pthread_mutex_lock(&sh->intr_mutex);
1464         MLX5_ASSERT(priv->ibv_port);
1465         MLX5_ASSERT(priv->ibv_port <= sh->max_port);
1466         MLX5_ASSERT(dev->data->port_id < RTE_MAX_ETHPORTS);
1467         if (sh->port[priv->ibv_port - 1].ih_port_id >= RTE_MAX_ETHPORTS)
1468                 goto exit;
1469         MLX5_ASSERT(sh->port[priv->ibv_port - 1].ih_port_id ==
1470                                         (uint32_t)dev->data->port_id);
1471         MLX5_ASSERT(sh->intr_cnt);
1472         sh->port[priv->ibv_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1473         if (!sh->intr_cnt || --sh->intr_cnt)
1474                 goto exit;
1475         mlx5_intr_callback_unregister(&sh->intr_handle,
1476                                      mlx5_dev_interrupt_handler, sh);
1477         sh->intr_handle.fd = 0;
1478         sh->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1479 exit:
1480         pthread_mutex_unlock(&sh->intr_mutex);
1481 }
1482
1483 /**
1484  * Uninstall devx shared asynchronous device events handler.
1485  * This function is implemeted to support event sharing
1486  * between multiple ports of single IB device.
1487  *
1488  * @param dev
1489  *   Pointer to Ethernet device.
1490  */
1491 static void
1492 mlx5_dev_shared_handler_devx_uninstall(struct rte_eth_dev *dev)
1493 {
1494         struct mlx5_priv *priv = dev->data->dev_private;
1495         struct mlx5_ibv_shared *sh = priv->sh;
1496
1497         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1498                 return;
1499         pthread_mutex_lock(&sh->intr_mutex);
1500         MLX5_ASSERT(priv->ibv_port);
1501         MLX5_ASSERT(priv->ibv_port <= sh->max_port);
1502         MLX5_ASSERT(dev->data->port_id < RTE_MAX_ETHPORTS);
1503         if (sh->port[priv->ibv_port - 1].devx_ih_port_id >= RTE_MAX_ETHPORTS)
1504                 goto exit;
1505         MLX5_ASSERT(sh->port[priv->ibv_port - 1].devx_ih_port_id ==
1506                     (uint32_t)dev->data->port_id);
1507         sh->port[priv->ibv_port - 1].devx_ih_port_id = RTE_MAX_ETHPORTS;
1508         if (!sh->devx_intr_cnt || --sh->devx_intr_cnt)
1509                 goto exit;
1510         if (sh->intr_handle_devx.fd) {
1511                 rte_intr_callback_unregister(&sh->intr_handle_devx,
1512                                              mlx5_dev_interrupt_handler_devx,
1513                                              sh);
1514                 sh->intr_handle_devx.fd = 0;
1515                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_UNKNOWN;
1516         }
1517         if (sh->devx_comp) {
1518                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1519                 sh->devx_comp = NULL;
1520         }
1521 exit:
1522         pthread_mutex_unlock(&sh->intr_mutex);
1523 }
1524
1525 /**
1526  * Install shared asynchronous device events handler.
1527  * This function is implemented to support event sharing
1528  * between multiple ports of single IB device.
1529  *
1530  * @param dev
1531  *   Pointer to Ethernet device.
1532  */
1533 static void
1534 mlx5_dev_shared_handler_install(struct rte_eth_dev *dev)
1535 {
1536         struct mlx5_priv *priv = dev->data->dev_private;
1537         struct mlx5_ibv_shared *sh = priv->sh;
1538         int ret;
1539         int flags;
1540
1541         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1542                 return;
1543         pthread_mutex_lock(&sh->intr_mutex);
1544         MLX5_ASSERT(priv->ibv_port);
1545         MLX5_ASSERT(priv->ibv_port <= sh->max_port);
1546         MLX5_ASSERT(dev->data->port_id < RTE_MAX_ETHPORTS);
1547         if (sh->port[priv->ibv_port - 1].ih_port_id < RTE_MAX_ETHPORTS) {
1548                 /* The handler is already installed for this port. */
1549                 MLX5_ASSERT(sh->intr_cnt);
1550                 goto exit;
1551         }
1552         if (sh->intr_cnt) {
1553                 sh->port[priv->ibv_port - 1].ih_port_id =
1554                                                 (uint32_t)dev->data->port_id;
1555                 sh->intr_cnt++;
1556                 goto exit;
1557         }
1558         /* No shared handler installed. */
1559         MLX5_ASSERT(sh->ctx->async_fd > 0);
1560         flags = fcntl(sh->ctx->async_fd, F_GETFL);
1561         ret = fcntl(sh->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1562         if (ret) {
1563                 DRV_LOG(INFO, "failed to change file descriptor async event"
1564                         " queue");
1565                 /* Indicate there will be no interrupts. */
1566                 dev->data->dev_conf.intr_conf.lsc = 0;
1567                 dev->data->dev_conf.intr_conf.rmv = 0;
1568         } else {
1569                 sh->intr_handle.fd = sh->ctx->async_fd;
1570                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1571                 rte_intr_callback_register(&sh->intr_handle,
1572                                            mlx5_dev_interrupt_handler, sh);
1573                 sh->intr_cnt++;
1574                 sh->port[priv->ibv_port - 1].ih_port_id =
1575                                                 (uint32_t)dev->data->port_id;
1576         }
1577 exit:
1578         pthread_mutex_unlock(&sh->intr_mutex);
1579 }
1580
1581 /**
1582  * Install devx shared asyncronous device events handler.
1583  * This function is implemeted to support event sharing
1584  * between multiple ports of single IB device.
1585  *
1586  * @param dev
1587  *   Pointer to Ethernet device.
1588  */
1589 static void
1590 mlx5_dev_shared_handler_devx_install(struct rte_eth_dev *dev)
1591 {
1592         struct mlx5_priv *priv = dev->data->dev_private;
1593         struct mlx5_ibv_shared *sh = priv->sh;
1594
1595         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1596                 return;
1597         pthread_mutex_lock(&sh->intr_mutex);
1598         MLX5_ASSERT(priv->ibv_port);
1599         MLX5_ASSERT(priv->ibv_port <= sh->max_port);
1600         MLX5_ASSERT(dev->data->port_id < RTE_MAX_ETHPORTS);
1601         if (sh->port[priv->ibv_port - 1].devx_ih_port_id < RTE_MAX_ETHPORTS) {
1602                 /* The handler is already installed for this port. */
1603                 MLX5_ASSERT(sh->devx_intr_cnt);
1604                 goto exit;
1605         }
1606         if (sh->devx_intr_cnt) {
1607                 sh->devx_intr_cnt++;
1608                 sh->port[priv->ibv_port - 1].devx_ih_port_id =
1609                                         (uint32_t)dev->data->port_id;
1610                 goto exit;
1611         }
1612         if (priv->config.devx) {
1613 #ifndef HAVE_IBV_DEVX_ASYNC
1614                 goto exit;
1615 #else
1616                 sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1617                 if (sh->devx_comp) {
1618                         int flags = fcntl(sh->devx_comp->fd, F_GETFL);
1619                         int ret = fcntl(sh->devx_comp->fd, F_SETFL,
1620                                     flags | O_NONBLOCK);
1621
1622                         if (ret) {
1623                                 DRV_LOG(INFO, "failed to change file descriptor"
1624                                         " devx async event queue");
1625                         } else {
1626                                 sh->intr_handle_devx.fd = sh->devx_comp->fd;
1627                                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1628                                 rte_intr_callback_register
1629                                         (&sh->intr_handle_devx,
1630                                          mlx5_dev_interrupt_handler_devx, sh);
1631                                 sh->devx_intr_cnt++;
1632                                 sh->port[priv->ibv_port - 1].devx_ih_port_id =
1633                                                 (uint32_t)dev->data->port_id;
1634                         }
1635                 }
1636 #endif /* HAVE_IBV_DEVX_ASYNC */
1637         }
1638 exit:
1639         pthread_mutex_unlock(&sh->intr_mutex);
1640 }
1641
1642 /**
1643  * Uninstall interrupt handler.
1644  *
1645  * @param dev
1646  *   Pointer to Ethernet device.
1647  */
1648 void
1649 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1650 {
1651         mlx5_dev_shared_handler_uninstall(dev);
1652 }
1653
1654 /**
1655  * Install interrupt handler.
1656  *
1657  * @param dev
1658  *   Pointer to Ethernet device.
1659  */
1660 void
1661 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1662 {
1663         mlx5_dev_shared_handler_install(dev);
1664 }
1665
1666 /**
1667  * Devx uninstall interrupt handler.
1668  *
1669  * @param dev
1670  *   Pointer to Ethernet device.
1671  */
1672 void
1673 mlx5_dev_interrupt_handler_devx_uninstall(struct rte_eth_dev *dev)
1674 {
1675         mlx5_dev_shared_handler_devx_uninstall(dev);
1676 }
1677
1678 /**
1679  * Devx install interrupt handler.
1680  *
1681  * @param dev
1682  *   Pointer to Ethernet device.
1683  */
1684 void
1685 mlx5_dev_interrupt_handler_devx_install(struct rte_eth_dev *dev)
1686 {
1687         mlx5_dev_shared_handler_devx_install(dev);
1688 }
1689
1690 /**
1691  * DPDK callback to bring the link DOWN.
1692  *
1693  * @param dev
1694  *   Pointer to Ethernet device structure.
1695  *
1696  * @return
1697  *   0 on success, a negative errno value otherwise and rte_errno is set.
1698  */
1699 int
1700 mlx5_set_link_down(struct rte_eth_dev *dev)
1701 {
1702         return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1703 }
1704
1705 /**
1706  * DPDK callback to bring the link UP.
1707  *
1708  * @param dev
1709  *   Pointer to Ethernet device structure.
1710  *
1711  * @return
1712  *   0 on success, a negative errno value otherwise and rte_errno is set.
1713  */
1714 int
1715 mlx5_set_link_up(struct rte_eth_dev *dev)
1716 {
1717         return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1718 }
1719
1720 /**
1721  * Configure the RX function to use.
1722  *
1723  * @param dev
1724  *   Pointer to private data structure.
1725  *
1726  * @return
1727  *   Pointer to selected Rx burst function.
1728  */
1729 eth_rx_burst_t
1730 mlx5_select_rx_function(struct rte_eth_dev *dev)
1731 {
1732         eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1733
1734         MLX5_ASSERT(dev != NULL);
1735         if (mlx5_check_vec_rx_support(dev) > 0) {
1736                 rx_pkt_burst = mlx5_rx_burst_vec;
1737                 DRV_LOG(DEBUG, "port %u selected Rx vectorized function",
1738                         dev->data->port_id);
1739         } else if (mlx5_mprq_enabled(dev)) {
1740                 rx_pkt_burst = mlx5_rx_burst_mprq;
1741         }
1742         return rx_pkt_burst;
1743 }
1744
1745 /**
1746  * Check if mlx5 device was removed.
1747  *
1748  * @param dev
1749  *   Pointer to Ethernet device structure.
1750  *
1751  * @return
1752  *   1 when device is removed, otherwise 0.
1753  */
1754 int
1755 mlx5_is_removed(struct rte_eth_dev *dev)
1756 {
1757         struct ibv_device_attr device_attr;
1758         struct mlx5_priv *priv = dev->data->dev_private;
1759
1760         if (mlx5_glue->query_device(priv->sh->ctx, &device_attr) == EIO)
1761                 return 1;
1762         return 0;
1763 }
1764
1765 /**
1766  * Get the E-Switch parameters by port id.
1767  *
1768  * @param[in] port
1769  *   Device port id.
1770  * @param[in] valid
1771  *   Device port id is valid, skip check. This flag is useful
1772  *   when trials are performed from probing and device is not
1773  *   flagged as valid yet (in attaching process).
1774  * @param[out] es_domain_id
1775  *   E-Switch domain id.
1776  * @param[out] es_port_id
1777  *   The port id of the port in the E-Switch.
1778  *
1779  * @return
1780  *   pointer to device private data structure containing data needed
1781  *   on success, NULL otherwise and rte_errno is set.
1782  */
1783 struct mlx5_priv *
1784 mlx5_port_to_eswitch_info(uint16_t port, bool valid)
1785 {
1786         struct rte_eth_dev *dev;
1787         struct mlx5_priv *priv;
1788
1789         if (port >= RTE_MAX_ETHPORTS) {
1790                 rte_errno = EINVAL;
1791                 return NULL;
1792         }
1793         if (!valid && !rte_eth_dev_is_valid_port(port)) {
1794                 rte_errno = ENODEV;
1795                 return NULL;
1796         }
1797         dev = &rte_eth_devices[port];
1798         priv = dev->data->dev_private;
1799         if (!(priv->representor || priv->master)) {
1800                 rte_errno = EINVAL;
1801                 return NULL;
1802         }
1803         return priv;
1804 }
1805
1806 /**
1807  * Get the E-Switch parameters by device instance.
1808  *
1809  * @param[in] port
1810  *   Device port id.
1811  * @param[out] es_domain_id
1812  *   E-Switch domain id.
1813  * @param[out] es_port_id
1814  *   The port id of the port in the E-Switch.
1815  *
1816  * @return
1817  *   pointer to device private data structure containing data needed
1818  *   on success, NULL otherwise and rte_errno is set.
1819  */
1820 struct mlx5_priv *
1821 mlx5_dev_to_eswitch_info(struct rte_eth_dev *dev)
1822 {
1823         struct mlx5_priv *priv;
1824
1825         priv = dev->data->dev_private;
1826         if (!(priv->representor || priv->master)) {
1827                 rte_errno = EINVAL;
1828                 return NULL;
1829         }
1830         return priv;
1831 }
1832
1833 /**
1834  * Get switch information associated with network interface.
1835  *
1836  * @param ifindex
1837  *   Network interface index.
1838  * @param[out] info
1839  *   Switch information object, populated in case of success.
1840  *
1841  * @return
1842  *   0 on success, a negative errno value otherwise and rte_errno is set.
1843  */
1844 int
1845 mlx5_sysfs_switch_info(unsigned int ifindex, struct mlx5_switch_info *info)
1846 {
1847         char ifname[IF_NAMESIZE];
1848         char port_name[IF_NAMESIZE];
1849         FILE *file;
1850         struct mlx5_switch_info data = {
1851                 .master = 0,
1852                 .representor = 0,
1853                 .name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET,
1854                 .port_name = 0,
1855                 .switch_id = 0,
1856         };
1857         DIR *dir;
1858         bool port_switch_id_set = false;
1859         bool device_dir = false;
1860         char c;
1861         int ret;
1862
1863         if (!if_indextoname(ifindex, ifname)) {
1864                 rte_errno = errno;
1865                 return -rte_errno;
1866         }
1867
1868         MKSTR(phys_port_name, "/sys/class/net/%s/phys_port_name",
1869               ifname);
1870         MKSTR(phys_switch_id, "/sys/class/net/%s/phys_switch_id",
1871               ifname);
1872         MKSTR(pci_device, "/sys/class/net/%s/device",
1873               ifname);
1874
1875         file = fopen(phys_port_name, "rb");
1876         if (file != NULL) {
1877                 ret = fscanf(file, "%s", port_name);
1878                 fclose(file);
1879                 if (ret == 1)
1880                         mlx5_translate_port_name(port_name, &data);
1881         }
1882         file = fopen(phys_switch_id, "rb");
1883         if (file == NULL) {
1884                 rte_errno = errno;
1885                 return -rte_errno;
1886         }
1887         port_switch_id_set =
1888                 fscanf(file, "%" SCNx64 "%c", &data.switch_id, &c) == 2 &&
1889                 c == '\n';
1890         fclose(file);
1891         dir = opendir(pci_device);
1892         if (dir != NULL) {
1893                 closedir(dir);
1894                 device_dir = true;
1895         }
1896         if (port_switch_id_set) {
1897                 /* We have some E-Switch configuration. */
1898                 mlx5_sysfs_check_switch_info(device_dir, &data);
1899         }
1900         *info = data;
1901         MLX5_ASSERT(!(data.master && data.representor));
1902         if (data.master && data.representor) {
1903                 DRV_LOG(ERR, "ifindex %u device is recognized as master"
1904                              " and as representor", ifindex);
1905                 rte_errno = ENODEV;
1906                 return -rte_errno;
1907         }
1908         return 0;
1909 }
1910
1911 /**
1912  * Analyze gathered port parameters via sysfs to recognize master
1913  * and representor devices for E-Switch configuration.
1914  *
1915  * @param[in] device_dir
1916  *   flag of presence of "device" directory under port device key.
1917  * @param[inout] switch_info
1918  *   Port information, including port name as a number and port name
1919  *   type if recognized
1920  *
1921  * @return
1922  *   master and representor flags are set in switch_info according to
1923  *   recognized parameters (if any).
1924  */
1925 void
1926 mlx5_sysfs_check_switch_info(bool device_dir,
1927                              struct mlx5_switch_info *switch_info)
1928 {
1929         switch (switch_info->name_type) {
1930         case MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN:
1931                 /*
1932                  * Name is not recognized, assume the master,
1933                  * check the device directory presence.
1934                  */
1935                 switch_info->master = device_dir;
1936                 break;
1937         case MLX5_PHYS_PORT_NAME_TYPE_NOTSET:
1938                 /*
1939                  * Name is not set, this assumes the legacy naming
1940                  * schema for master, just check if there is
1941                  * a device directory.
1942                  */
1943                 switch_info->master = device_dir;
1944                 break;
1945         case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1946                 /* New uplink naming schema recognized. */
1947                 switch_info->master = 1;
1948                 break;
1949         case MLX5_PHYS_PORT_NAME_TYPE_LEGACY:
1950                 /* Legacy representors naming schema. */
1951                 switch_info->representor = !device_dir;
1952                 break;
1953         case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1954                 /* New representors naming schema. */
1955                 switch_info->representor = 1;
1956                 break;
1957         }
1958 }
1959
1960 /**
1961  * DPDK callback to retrieve plug-in module EEPROM information (type and size).
1962  *
1963  * @param dev
1964  *   Pointer to Ethernet device structure.
1965  * @param[out] modinfo
1966  *   Storage for plug-in module EEPROM information.
1967  *
1968  * @return
1969  *   0 on success, a negative errno value otherwise and rte_errno is set.
1970  */
1971 int
1972 mlx5_get_module_info(struct rte_eth_dev *dev,
1973                      struct rte_eth_dev_module_info *modinfo)
1974 {
1975         struct ethtool_modinfo info = {
1976                 .cmd = ETHTOOL_GMODULEINFO,
1977         };
1978         struct ifreq ifr = (struct ifreq) {
1979                 .ifr_data = (void *)&info,
1980         };
1981         int ret = 0;
1982
1983         if (!dev || !modinfo) {
1984                 DRV_LOG(WARNING, "missing argument, cannot get module info");
1985                 rte_errno = EINVAL;
1986                 return -rte_errno;
1987         }
1988         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
1989         if (ret) {
1990                 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
1991                         dev->data->port_id, strerror(rte_errno));
1992                 return ret;
1993         }
1994         modinfo->type = info.type;
1995         modinfo->eeprom_len = info.eeprom_len;
1996         return ret;
1997 }
1998
1999 /**
2000  * DPDK callback to retrieve plug-in module EEPROM data.
2001  *
2002  * @param dev
2003  *   Pointer to Ethernet device structure.
2004  * @param[out] info
2005  *   Storage for plug-in module EEPROM data.
2006  *
2007  * @return
2008  *   0 on success, a negative errno value otherwise and rte_errno is set.
2009  */
2010 int mlx5_get_module_eeprom(struct rte_eth_dev *dev,
2011                            struct rte_dev_eeprom_info *info)
2012 {
2013         struct ethtool_eeprom *eeprom;
2014         struct ifreq ifr;
2015         int ret = 0;
2016
2017         if (!dev || !info) {
2018                 DRV_LOG(WARNING, "missing argument, cannot get module eeprom");
2019                 rte_errno = EINVAL;
2020                 return -rte_errno;
2021         }
2022         eeprom = rte_calloc(__func__, 1,
2023                             (sizeof(struct ethtool_eeprom) + info->length), 0);
2024         if (!eeprom) {
2025                 DRV_LOG(WARNING, "port %u cannot allocate memory for "
2026                         "eeprom data", dev->data->port_id);
2027                 rte_errno = ENOMEM;
2028                 return -rte_errno;
2029         }
2030         eeprom->cmd = ETHTOOL_GMODULEEEPROM;
2031         eeprom->offset = info->offset;
2032         eeprom->len = info->length;
2033         ifr = (struct ifreq) {
2034                 .ifr_data = (void *)eeprom,
2035         };
2036         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2037         if (ret)
2038                 DRV_LOG(WARNING, "port %u ioctl(SIOCETHTOOL) failed: %s",
2039                         dev->data->port_id, strerror(rte_errno));
2040         else
2041                 rte_memcpy(info->data, eeprom->data, info->length);
2042         rte_free(eeprom);
2043         return ret;
2044 }
2045
2046 /**
2047  * DPDK callback to retrieve hairpin capabilities.
2048  *
2049  * @param dev
2050  *   Pointer to Ethernet device structure.
2051  * @param[out] cap
2052  *   Storage for hairpin capability data.
2053  *
2054  * @return
2055  *   0 on success, a negative errno value otherwise and rte_errno is set.
2056  */
2057 int mlx5_hairpin_cap_get(struct rte_eth_dev *dev,
2058                          struct rte_eth_hairpin_cap *cap)
2059 {
2060         struct mlx5_priv *priv = dev->data->dev_private;
2061
2062         if (priv->sh->devx == 0) {
2063                 rte_errno = ENOTSUP;
2064                 return -rte_errno;
2065         }
2066         cap->max_nb_queues = UINT16_MAX;
2067         cap->max_rx_2_tx = 1;
2068         cap->max_tx_2_rx = 1;
2069         cap->max_nb_desc = 8192;
2070         return 0;
2071 }