drivers/net/mlx5/mlx5_ethdev.c

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