drivers/net/mlx5/mlx5_ethdev.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2015 6WIND S.A.
   5  *   Copyright 2015 Mellanox.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of 6WIND S.A. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stddef.h>
  35 #include <assert.h>
  36 #include <unistd.h>
  37 #include <stdint.h>
  38 #include <stdio.h>
  39 #include <string.h>
  40 #include <stdlib.h>
  41 #include <errno.h>
  42 #include <dirent.h>
  43 #include <net/if.h>
  44 #include <sys/ioctl.h>
  45 #include <sys/socket.h>
  46 #include <netinet/in.h>
  47 #include <linux/ethtool.h>
  48 #include <linux/sockios.h>
  49 #include <fcntl.h>
  50
  51 /* DPDK headers don't like -pedantic. */
  52 #ifdef PEDANTIC
  53 #pragma GCC diagnostic ignored "-Wpedantic"
  54 #endif
  55 #include <rte_atomic.h>
  56 #include <rte_ethdev.h>
  57 #include <rte_mbuf.h>
  58 #include <rte_common.h>
  59 #include <rte_interrupts.h>
  60 #include <rte_alarm.h>
  61 #include <rte_malloc.h>
  62 #ifdef PEDANTIC
  63 #pragma GCC diagnostic error "-Wpedantic"
  64 #endif
  65
  66 #include "mlx5.h"
  67 #include "mlx5_rxtx.h"
  68 #include "mlx5_utils.h"
  69
  70 /**
  71  * Return private structure associated with an Ethernet device.
  72  *
  73  * @param dev
  74  *   Pointer to Ethernet device structure.
  75  *
  76  * @return
  77  *   Pointer to private structure.
  78  */
  79 struct priv *
  80 mlx5_get_priv(struct rte_eth_dev *dev)
  81 {
  82         struct mlx5_secondary_data *sd;
  83
  84         if (!mlx5_is_secondary())
  85                 return dev->data->dev_private;
  86         sd = &mlx5_secondary_data[dev->data->port_id];
  87         return sd->data.dev_private;
  88 }
  89
  90 /**
  91  * Check if running as a secondary process.
  92  *
  93  * @return
  94  *   Nonzero if running as a secondary process.
  95  */
  96 inline int
  97 mlx5_is_secondary(void)
  98 {
  99         return rte_eal_process_type() != RTE_PROC_PRIMARY;
 100 }
 101
 102 /**
 103  * Get interface name from private structure.
 104  *
 105  * @param[in] priv
 106  *   Pointer to private structure.
 107  * @param[out] ifname
 108  *   Interface name output buffer.
 109  *
 110  * @return
 111  *   0 on success, -1 on failure and errno is set.
 112  */
 113 int
 114 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
 115 {
 116         DIR *dir;
 117         struct dirent *dent;
 118         unsigned int dev_type = 0;
 119         unsigned int dev_port_prev = ~0u;
 120         char match[IF_NAMESIZE] = "";
 121
 122         {
 123                 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
 124
 125                 dir = opendir(path);
 126                 if (dir == NULL)
 127                         return -1;
 128         }
 129         while ((dent = readdir(dir)) != NULL) {
 130                 char *name = dent->d_name;
 131                 FILE *file;
 132                 unsigned int dev_port;
 133                 int r;
 134
 135                 if ((name[0] == '.') &&
 136                     ((name[1] == '\0') ||
 137                      ((name[1] == '.') && (name[2] == '\0'))))
 138                         continue;
 139
 140                 MKSTR(path, "%s/device/net/%s/%s",
 141                       priv->ctx->device->ibdev_path, name,
 142                       (dev_type ? "dev_id" : "dev_port"));
 143
 144                 file = fopen(path, "rb");
 145                 if (file == NULL) {
 146                         if (errno != ENOENT)
 147                                 continue;
 148                         /*
 149                          * Switch to dev_id when dev_port does not exist as
 150                          * is the case with Linux kernel versions < 3.15.
 151                          */
 152 try_dev_id:
 153                         match[0] = '\0';
 154                         if (dev_type)
 155                                 break;
 156                         dev_type = 1;
 157                         dev_port_prev = ~0u;
 158                         rewinddir(dir);
 159                         continue;
 160                 }
 161                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 162                 fclose(file);
 163                 if (r != 1)
 164                         continue;
 165                 /*
 166                  * Switch to dev_id when dev_port returns the same value for
 167                  * all ports. May happen when using a MOFED release older than
 168                  * 3.0 with a Linux kernel >= 3.15.
 169                  */
 170                 if (dev_port == dev_port_prev)
 171                         goto try_dev_id;
 172                 dev_port_prev = dev_port;
 173                 if (dev_port == (priv->port - 1u))
 174                         snprintf(match, sizeof(match), "%s", name);
 175         }
 176         closedir(dir);
 177         if (match[0] == '\0')
 178                 return -1;
 179         strncpy(*ifname, match, sizeof(*ifname));
 180         return 0;
 181 }
 182
 183 /**
 184  * Read from sysfs entry.
 185  *
 186  * @param[in] priv
 187  *   Pointer to private structure.
 188  * @param[in] entry
 189  *   Entry name relative to sysfs path.
 190  * @param[out] buf
 191  *   Data output buffer.
 192  * @param size
 193  *   Buffer size.
 194  *
 195  * @return
 196  *   0 on success, -1 on failure and errno is set.
 197  */
 198 static int
 199 priv_sysfs_read(const struct priv *priv, const char *entry,
 200                 char *buf, size_t size)
 201 {
 202         char ifname[IF_NAMESIZE];
 203         FILE *file;
 204         int ret;
 205         int err;
 206
 207         if (priv_get_ifname(priv, &ifname))
 208                 return -1;
 209
 210         MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
 211               ifname, entry);
 212
 213         file = fopen(path, "rb");
 214         if (file == NULL)
 215                 return -1;
 216         ret = fread(buf, 1, size, file);
 217         err = errno;
 218         if (((size_t)ret < size) && (ferror(file)))
 219                 ret = -1;
 220         else
 221                 ret = size;
 222         fclose(file);
 223         errno = err;
 224         return ret;
 225 }
 226
 227 /**
 228  * Write to sysfs entry.
 229  *
 230  * @param[in] priv
 231  *   Pointer to private structure.
 232  * @param[in] entry
 233  *   Entry name relative to sysfs path.
 234  * @param[in] buf
 235  *   Data buffer.
 236  * @param size
 237  *   Buffer size.
 238  *
 239  * @return
 240  *   0 on success, -1 on failure and errno is set.
 241  */
 242 static int
 243 priv_sysfs_write(const struct priv *priv, const char *entry,
 244                  char *buf, size_t size)
 245 {
 246         char ifname[IF_NAMESIZE];
 247         FILE *file;
 248         int ret;
 249         int err;
 250
 251         if (priv_get_ifname(priv, &ifname))
 252                 return -1;
 253
 254         MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
 255               ifname, entry);
 256
 257         file = fopen(path, "wb");
 258         if (file == NULL)
 259                 return -1;
 260         ret = fwrite(buf, 1, size, file);
 261         err = errno;
 262         if (((size_t)ret < size) || (ferror(file)))
 263                 ret = -1;
 264         else
 265                 ret = size;
 266         fclose(file);
 267         errno = err;
 268         return ret;
 269 }
 270
 271 /**
 272  * Get unsigned long sysfs property.
 273  *
 274  * @param priv
 275  *   Pointer to private structure.
 276  * @param[in] name
 277  *   Entry name relative to sysfs path.
 278  * @param[out] value
 279  *   Value output buffer.
 280  *
 281  * @return
 282  *   0 on success, -1 on failure and errno is set.
 283  */
 284 static int
 285 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
 286 {
 287         int ret;
 288         unsigned long value_ret;
 289         char value_str[32];
 290
 291         ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
 292         if (ret == -1) {
 293                 DEBUG("cannot read %s value from sysfs: %s",
 294                       name, strerror(errno));
 295                 return -1;
 296         }
 297         value_str[ret] = '\0';
 298         errno = 0;
 299         value_ret = strtoul(value_str, NULL, 0);
 300         if (errno) {
 301                 DEBUG("invalid %s value `%s': %s", name, value_str,
 302                       strerror(errno));
 303                 return -1;
 304         }
 305         *value = value_ret;
 306         return 0;
 307 }
 308
 309 /**
 310  * Set unsigned long sysfs property.
 311  *
 312  * @param priv
 313  *   Pointer to private structure.
 314  * @param[in] name
 315  *   Entry name relative to sysfs path.
 316  * @param value
 317  *   Value to set.
 318  *
 319  * @return
 320  *   0 on success, -1 on failure and errno is set.
 321  */
 322 static int
 323 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
 324 {
 325         int ret;
 326         MKSTR(value_str, "%lu", value);
 327
 328         ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
 329         if (ret == -1) {
 330                 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
 331                       name, value_str, value, strerror(errno));
 332                 return -1;
 333         }
 334         return 0;
 335 }
 336
 337 /**
 338  * Perform ifreq ioctl() on associated Ethernet device.
 339  *
 340  * @param[in] priv
 341  *   Pointer to private structure.
 342  * @param req
 343  *   Request number to pass to ioctl().
 344  * @param[out] ifr
 345  *   Interface request structure output buffer.
 346  *
 347  * @return
 348  *   0 on success, -1 on failure and errno is set.
 349  */
 350 int
 351 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
 352 {
 353         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 354         int ret = -1;
 355
 356         if (sock == -1)
 357                 return ret;
 358         if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
 359                 ret = ioctl(sock, req, ifr);
 360         close(sock);
 361         return ret;
 362 }
 363
 364 /**
 365  * Return the number of active VFs for the current device.
 366  *
 367  * @param[in] priv
 368  *   Pointer to private structure.
 369  * @param[out] num_vfs
 370  *   Number of active VFs.
 371  *
 372  * @return
 373  *   0 on success, -1 on failure and errno is set.
 374  */
 375 int
 376 priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs)
 377 {
 378         /* The sysfs entry name depends on the operating system. */
 379         const char **name = (const char *[]){
 380                 "device/sriov_numvfs",
 381                 "device/mlx5_num_vfs",
 382                 NULL,
 383         };
 384         int ret;
 385
 386         do {
 387                 unsigned long ulong_num_vfs;
 388
 389                 ret = priv_get_sysfs_ulong(priv, *name, &ulong_num_vfs);
 390                 if (!ret)
 391                         *num_vfs = ulong_num_vfs;
 392         } while (*(++name) && ret);
 393         return ret;
 394 }
 395
 396 /**
 397  * Get device MTU.
 398  *
 399  * @param priv
 400  *   Pointer to private structure.
 401  * @param[out] mtu
 402  *   MTU value output buffer.
 403  *
 404  * @return
 405  *   0 on success, -1 on failure and errno is set.
 406  */
 407 int
 408 priv_get_mtu(struct priv *priv, uint16_t *mtu)
 409 {
 410         unsigned long ulong_mtu;
 411
 412         if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
 413                 return -1;
 414         *mtu = ulong_mtu;
 415         return 0;
 416 }
 417
 418 /**
 419  * Set device MTU.
 420  *
 421  * @param priv
 422  *   Pointer to private structure.
 423  * @param mtu
 424  *   MTU value to set.
 425  *
 426  * @return
 427  *   0 on success, -1 on failure and errno is set.
 428  */
 429 static int
 430 priv_set_mtu(struct priv *priv, uint16_t mtu)
 431 {
 432         uint16_t new_mtu;
 433
 434         if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
 435             priv_get_mtu(priv, &new_mtu))
 436                 return -1;
 437         if (new_mtu == mtu)
 438                 return 0;
 439         errno = EINVAL;
 440         return -1;
 441 }
 442
 443 /**
 444  * Set device flags.
 445  *
 446  * @param priv
 447  *   Pointer to private structure.
 448  * @param keep
 449  *   Bitmask for flags that must remain untouched.
 450  * @param flags
 451  *   Bitmask for flags to modify.
 452  *
 453  * @return
 454  *   0 on success, -1 on failure and errno is set.
 455  */
 456 int
 457 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
 458 {
 459         unsigned long tmp;
 460
 461         if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
 462                 return -1;
 463         tmp &= keep;
 464         tmp |= (flags & (~keep));
 465         return priv_set_sysfs_ulong(priv, "flags", tmp);
 466 }
 467
 468 /**
 469  * Ethernet device configuration.
 470  *
 471  * Prepare the driver for a given number of TX and RX queues.
 472  *
 473  * @param dev
 474  *   Pointer to Ethernet device structure.
 475  *
 476  * @return
 477  *   0 on success, errno value on failure.
 478  */
 479 static int
 480 dev_configure(struct rte_eth_dev *dev)
 481 {
 482         struct priv *priv = dev->data->dev_private;
 483         unsigned int rxqs_n = dev->data->nb_rx_queues;
 484         unsigned int txqs_n = dev->data->nb_tx_queues;
 485         unsigned int i;
 486         unsigned int j;
 487         unsigned int reta_idx_n;
 488
 489         priv->rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
 490         priv->rxqs = (void *)dev->data->rx_queues;
 491         priv->txqs = (void *)dev->data->tx_queues;
 492         if (txqs_n != priv->txqs_n) {
 493                 INFO("%p: TX queues number update: %u -> %u",
 494                      (void *)dev, priv->txqs_n, txqs_n);
 495                 priv->txqs_n = txqs_n;
 496         }
 497         if (rxqs_n > priv->ind_table_max_size) {
 498                 ERROR("cannot handle this many RX queues (%u)", rxqs_n);
 499                 return EINVAL;
 500         }
 501         if (rxqs_n == priv->rxqs_n)
 502                 return 0;
 503         INFO("%p: RX queues number update: %u -> %u",
 504              (void *)dev, priv->rxqs_n, rxqs_n);
 505         priv->rxqs_n = rxqs_n;
 506         /* If the requested number of RX queues is not a power of two, use the
 507          * maximum indirection table size for better balancing.
 508          * The result is always rounded to the next power of two. */
 509         reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
 510                                      priv->ind_table_max_size :
 511                                      rxqs_n));
 512         if (priv_rss_reta_index_resize(priv, reta_idx_n))
 513                 return ENOMEM;
 514         /* When the number of RX queues is not a power of two, the remaining
 515          * table entries are padded with reused WQs and hashes are not spread
 516          * uniformly. */
 517         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
 518                 (*priv->reta_idx)[i] = j;
 519                 if (++j == rxqs_n)
 520                         j = 0;
 521         }
 522         return 0;
 523 }
 524
 525 /**
 526  * DPDK callback for Ethernet device configuration.
 527  *
 528  * @param dev
 529  *   Pointer to Ethernet device structure.
 530  *
 531  * @return
 532  *   0 on success, negative errno value on failure.
 533  */
 534 int
 535 mlx5_dev_configure(struct rte_eth_dev *dev)
 536 {
 537         struct priv *priv = dev->data->dev_private;
 538         int ret;
 539
 540         if (mlx5_is_secondary())
 541                 return -E_RTE_SECONDARY;
 542
 543         priv_lock(priv);
 544         ret = dev_configure(dev);
 545         assert(ret >= 0);
 546         priv_unlock(priv);
 547         return -ret;
 548 }
 549
 550 /**
 551  * DPDK callback to get information about the device.
 552  *
 553  * @param dev
 554  *   Pointer to Ethernet device structure.
 555  * @param[out] info
 556  *   Info structure output buffer.
 557  */
 558 void
 559 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 560 {
 561         struct priv *priv = mlx5_get_priv(dev);
 562         unsigned int max;
 563         char ifname[IF_NAMESIZE];
 564
 565         priv_lock(priv);
 566         /* FIXME: we should ask the device for these values. */
 567         info->min_rx_bufsize = 32;
 568         info->max_rx_pktlen = 65536;
 569         /*
 570          * Since we need one CQ per QP, the limit is the minimum number
 571          * between the two values.
 572          */
 573         max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
 574                priv->device_attr.max_qp : priv->device_attr.max_cq);
 575         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
 576         if (max >= 65535)
 577                 max = 65535;
 578         info->max_rx_queues = max;
 579         info->max_tx_queues = max;
 580         info->max_mac_addrs = RTE_DIM(priv->mac);
 581         info->rx_offload_capa =
 582                 (priv->hw_csum ?
 583                  (DEV_RX_OFFLOAD_IPV4_CKSUM |
 584                   DEV_RX_OFFLOAD_UDP_CKSUM |
 585                   DEV_RX_OFFLOAD_TCP_CKSUM) :
 586                  0) |
 587                 (priv->hw_vlan_strip ? DEV_RX_OFFLOAD_VLAN_STRIP : 0);
 588         if (!priv->mps)
 589                 info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
 590         if (priv->hw_csum)
 591                 info->tx_offload_capa |=
 592                         (DEV_TX_OFFLOAD_IPV4_CKSUM |
 593                          DEV_TX_OFFLOAD_UDP_CKSUM |
 594                          DEV_TX_OFFLOAD_TCP_CKSUM);
 595         if (priv_get_ifname(priv, &ifname) == 0)
 596                 info->if_index = if_nametoindex(ifname);
 597         /* FIXME: RETA update/query API expects the callee to know the size of
 598          * the indirection table, for this PMD the size varies depending on
 599          * the number of RX queues, it becomes impossible to find the correct
 600          * size if it is not fixed.
 601          * The API should be updated to solve this problem. */
 602         info->reta_size = priv->ind_table_max_size;
 603         info->hash_key_size = ((*priv->rss_conf) ?
 604                                (*priv->rss_conf)[0]->rss_key_len :
 605                                0);
 606         info->speed_capa = priv->link_speed_capa;
 607         priv_unlock(priv);
 608 }
 609
 610 const uint32_t *
 611 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 612 {
 613         static const uint32_t ptypes[] = {
 614                 /* refers to rxq_cq_to_pkt_type() */
 615                 RTE_PTYPE_L3_IPV4,
 616                 RTE_PTYPE_L3_IPV6,
 617                 RTE_PTYPE_INNER_L3_IPV4,
 618                 RTE_PTYPE_INNER_L3_IPV6,
 619                 RTE_PTYPE_UNKNOWN
 620
 621         };
 622
 623         if (dev->rx_pkt_burst == mlx5_rx_burst)
 624                 return ptypes;
 625         return NULL;
 626 }
 627
 628 /**
 629  * Retrieve physical link information (unlocked version using legacy ioctl).
 630  *
 631  * @param dev
 632  *   Pointer to Ethernet device structure.
 633  * @param wait_to_complete
 634  *   Wait for request completion (ignored).
 635  */
 636 static int
 637 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
 638 {
 639         struct priv *priv = mlx5_get_priv(dev);
 640         struct ethtool_cmd edata = {
 641                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 642         };
 643         struct ifreq ifr;
 644         struct rte_eth_link dev_link;
 645         int link_speed = 0;
 646
 647         (void)wait_to_complete;
 648         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
 649                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
 650                 return -1;
 651         }
 652         memset(&dev_link, 0, sizeof(dev_link));
 653         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 654                                 (ifr.ifr_flags & IFF_RUNNING));
 655         ifr.ifr_data = (void *)&edata;
 656         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 657                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
 658                      strerror(errno));
 659                 return -1;
 660         }
 661         link_speed = ethtool_cmd_speed(&edata);
 662         if (link_speed == -1)
 663                 dev_link.link_speed = 0;
 664         else
 665                 dev_link.link_speed = link_speed;
 666         priv->link_speed_capa = 0;
 667         if (edata.supported & SUPPORTED_Autoneg)
 668                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 669         if (edata.supported & (SUPPORTED_1000baseT_Full |
 670                                SUPPORTED_1000baseKX_Full))
 671                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 672         if (edata.supported & SUPPORTED_10000baseKR_Full)
 673                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 674         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 675                                SUPPORTED_40000baseCR4_Full |
 676                                SUPPORTED_40000baseSR4_Full |
 677                                SUPPORTED_40000baseLR4_Full))
 678                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 679         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 680                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 681         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 682                         ETH_LINK_SPEED_FIXED);
 683         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 684                 /* Link status changed. */
 685                 dev->data->dev_link = dev_link;
 686                 return 0;
 687         }
 688         /* Link status is still the same. */
 689         return -1;
 690 }
 691
 692 /**
 693  * Retrieve physical link information (unlocked version using new ioctl from
 694  * Linux 4.5).
 695  *
 696  * @param dev
 697  *   Pointer to Ethernet device structure.
 698  * @param wait_to_complete
 699  *   Wait for request completion (ignored).
 700  */
 701 static int
 702 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
 703 {
 704 #ifdef ETHTOOL_GLINKSETTINGS
 705         struct priv *priv = mlx5_get_priv(dev);
 706         struct ethtool_link_settings edata = {
 707                 .cmd = ETHTOOL_GLINKSETTINGS,
 708         };
 709         struct ifreq ifr;
 710         struct rte_eth_link dev_link;
 711         uint64_t sc;
 712
 713         (void)wait_to_complete;
 714         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
 715                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
 716                 return -1;
 717         }
 718         memset(&dev_link, 0, sizeof(dev_link));
 719         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 720                                 (ifr.ifr_flags & IFF_RUNNING));
 721         ifr.ifr_data = (void *)&edata;
 722         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 723                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
 724                       strerror(errno));
 725                 return -1;
 726         }
 727         dev_link.link_speed = edata.speed;
 728         sc = edata.link_mode_masks[0] |
 729                 ((uint64_t)edata.link_mode_masks[1] << 32);
 730         priv->link_speed_capa = 0;
 731         /* Link speeds available in kernel v4.5. */
 732         if (sc & ETHTOOL_LINK_MODE_Autoneg_BIT)
 733                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 734         if (sc & (ETHTOOL_LINK_MODE_1000baseT_Full_BIT |
 735                   ETHTOOL_LINK_MODE_1000baseKX_Full_BIT))
 736                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 737         if (sc & (ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT |
 738                   ETHTOOL_LINK_MODE_10000baseKR_Full_BIT |
 739                   ETHTOOL_LINK_MODE_10000baseR_FEC_BIT))
 740                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 741         if (sc & (ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT |
 742                   ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT))
 743                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
 744         if (sc & (ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT |
 745                   ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT |
 746                   ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT |
 747                   ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT))
 748                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 749         if (sc & (ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT |
 750                   ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT |
 751                   ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT |
 752                   ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT))
 753                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
 754         /* Link speeds available in kernel v4.6. */
 755 #ifdef HAVE_ETHTOOL_LINK_MODE_25G
 756         if (sc & (ETHTOOL_LINK_MODE_25000baseCR_Full_BIT |
 757                   ETHTOOL_LINK_MODE_25000baseKR_Full_BIT |
 758                   ETHTOOL_LINK_MODE_25000baseSR_Full_BIT))
 759                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
 760 #endif
 761 #ifdef HAVE_ETHTOOL_LINK_MODE_50G
 762         if (sc & (ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT |
 763                   ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT))
 764                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
 765 #endif
 766 #ifdef HAVE_ETHTOOL_LINK_MODE_100G
 767         if (sc & (ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT |
 768                   ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT |
 769                   ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT |
 770                   ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT))
 771                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
 772 #endif
 773         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 774                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 775         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 776                                   ETH_LINK_SPEED_FIXED);
 777         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 778                 /* Link status changed. */
 779                 dev->data->dev_link = dev_link;
 780                 return 0;
 781         }
 782 #else
 783         (void)dev;
 784         (void)wait_to_complete;
 785 #endif
 786         /* Link status is still the same. */
 787         return -1;
 788 }
 789
 790 /**
 791  * DPDK callback to retrieve physical link information (unlocked version).
 792  *
 793  * @param dev
 794  *   Pointer to Ethernet device structure.
 795  * @param wait_to_complete
 796  *   Wait for request completion (ignored).
 797  */
 798 int
 799 mlx5_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
 800 {
 801         int ret;
 802
 803         ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
 804         if (ret < 0)
 805                 ret = mlx5_link_update_unlocked_gset(dev, wait_to_complete);
 806         return ret;
 807 }
 808
 809 /**
 810  * DPDK callback to retrieve physical link information.
 811  *
 812  * @param dev
 813  *   Pointer to Ethernet device structure.
 814  * @param wait_to_complete
 815  *   Wait for request completion (ignored).
 816  */
 817 int
 818 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
 819 {
 820         struct priv *priv = mlx5_get_priv(dev);
 821         int ret;
 822
 823         priv_lock(priv);
 824         ret = mlx5_link_update_unlocked(dev, wait_to_complete);
 825         priv_unlock(priv);
 826         return ret;
 827 }
 828
 829 /**
 830  * DPDK callback to change the MTU.
 831  *
 832  * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
 833  * received). Use this as a hint to enable/disable scattered packets support
 834  * and improve performance when not needed.
 835  * Since failure is not an option, reconfiguring queues on the fly is not
 836  * recommended.
 837  *
 838  * @param dev
 839  *   Pointer to Ethernet device structure.
 840  * @param in_mtu
 841  *   New MTU.
 842  *
 843  * @return
 844  *   0 on success, negative errno value on failure.
 845  */
 846 int
 847 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 848 {
 849         struct priv *priv = dev->data->dev_private;
 850         int ret = 0;
 851         unsigned int i;
 852         uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
 853                 mlx5_rx_burst;
 854         unsigned int max_frame_len;
 855         int rehash;
 856         int restart = priv->started;
 857
 858         if (mlx5_is_secondary())
 859                 return -E_RTE_SECONDARY;
 860
 861         priv_lock(priv);
 862         /* Set kernel interface MTU first. */
 863         if (priv_set_mtu(priv, mtu)) {
 864                 ret = errno;
 865                 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
 866                      strerror(ret));
 867                 goto out;
 868         } else
 869                 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
 870         /* Temporarily replace RX handler with a fake one, assuming it has not
 871          * been copied elsewhere. */
 872         dev->rx_pkt_burst = removed_rx_burst;
 873         /* Make sure everyone has left mlx5_rx_burst() and uses
 874          * removed_rx_burst() instead. */
 875         rte_wmb();
 876         usleep(1000);
 877         /* MTU does not include header and CRC. */
 878         max_frame_len = ETHER_HDR_LEN + mtu + ETHER_CRC_LEN;
 879         /* Check if at least one queue is going to need a SGE update. */
 880         for (i = 0; i != priv->rxqs_n; ++i) {
 881                 struct rxq *rxq = (*priv->rxqs)[i];
 882                 unsigned int mb_len;
 883                 unsigned int size = RTE_PKTMBUF_HEADROOM + max_frame_len;
 884                 unsigned int sges_n;
 885
 886                 if (rxq == NULL)
 887                         continue;
 888                 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
 889                 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
 890                 /*
 891                  * Determine the number of SGEs needed for a full packet
 892                  * and round it to the next power of two.
 893                  */
 894                 sges_n = log2above((size / mb_len) + !!(size % mb_len));
 895                 if (sges_n != rxq->sges_n)
 896                         break;
 897         }
 898         /*
 899          * If all queues have the right number of SGEs, a simple rehash
 900          * of their buffers is enough, otherwise SGE information can only
 901          * be updated in a queue by recreating it. All resources that depend
 902          * on queues (flows, indirection tables) must be recreated as well in
 903          * that case.
 904          */
 905         rehash = (i == priv->rxqs_n);
 906         if (!rehash) {
 907                 /* Clean up everything as with mlx5_dev_stop(). */
 908                 priv_special_flow_disable_all(priv);
 909                 priv_mac_addrs_disable(priv);
 910                 priv_destroy_hash_rxqs(priv);
 911                 priv_fdir_disable(priv);
 912                 priv_dev_interrupt_handler_uninstall(priv, dev);
 913         }
 914 recover:
 915         /* Reconfigure each RX queue. */
 916         for (i = 0; (i != priv->rxqs_n); ++i) {
 917                 struct rxq *rxq = (*priv->rxqs)[i];
 918                 struct rxq_ctrl *rxq_ctrl =
 919                         container_of(rxq, struct rxq_ctrl, rxq);
 920                 int sp;
 921                 unsigned int mb_len;
 922                 unsigned int tmp;
 923
 924                 if (rxq == NULL)
 925                         continue;
 926                 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
 927                 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
 928                 /* Toggle scattered support (sp) if necessary. */
 929                 sp = (max_frame_len > (mb_len - RTE_PKTMBUF_HEADROOM));
 930                 /* Provide new values to rxq_setup(). */
 931                 dev->data->dev_conf.rxmode.jumbo_frame = sp;
 932                 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
 933                 if (rehash)
 934                         ret = rxq_rehash(dev, rxq_ctrl);
 935                 else
 936                         ret = rxq_ctrl_setup(dev, rxq_ctrl, 1 << rxq->elts_n,
 937                                              rxq_ctrl->socket, NULL, rxq->mp);
 938                 if (!ret)
 939                         continue;
 940                 /* Attempt to roll back in case of error. */
 941                 tmp = (mb_len << rxq->sges_n) - RTE_PKTMBUF_HEADROOM;
 942                 if (max_frame_len != tmp) {
 943                         max_frame_len = tmp;
 944                         goto recover;
 945                 }
 946                 /* Double fault, disable RX. */
 947                 break;
 948         }
 949         /*
 950          * Use a safe RX burst function in case of error, otherwise mimic
 951          * mlx5_dev_start().
 952          */
 953         if (ret) {
 954                 ERROR("unable to reconfigure RX queues, RX disabled");
 955                 rx_func = removed_rx_burst;
 956         } else if (restart &&
 957                  !rehash &&
 958                  !priv_create_hash_rxqs(priv) &&
 959                  !priv_rehash_flows(priv)) {
 960                 if (dev->data->dev_conf.fdir_conf.mode == RTE_FDIR_MODE_NONE)
 961                         priv_fdir_enable(priv);
 962                 priv_dev_interrupt_handler_install(priv, dev);
 963         }
 964         priv->mtu = mtu;
 965         /* Burst functions can now be called again. */
 966         rte_wmb();
 967         dev->rx_pkt_burst = rx_func;
 968 out:
 969         priv_unlock(priv);
 970         assert(ret >= 0);
 971         return -ret;
 972 }
 973
 974 /**
 975  * DPDK callback to get flow control status.
 976  *
 977  * @param dev
 978  *   Pointer to Ethernet device structure.
 979  * @param[out] fc_conf
 980  *   Flow control output buffer.
 981  *
 982  * @return
 983  *   0 on success, negative errno value on failure.
 984  */
 985 int
 986 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 987 {
 988         struct priv *priv = dev->data->dev_private;
 989         struct ifreq ifr;
 990         struct ethtool_pauseparam ethpause = {
 991                 .cmd = ETHTOOL_GPAUSEPARAM
 992         };
 993         int ret;
 994
 995         if (mlx5_is_secondary())
 996                 return -E_RTE_SECONDARY;
 997
 998         ifr.ifr_data = (void *)&ethpause;
 999         priv_lock(priv);
1000         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1001                 ret = errno;
1002                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
1003                      " failed: %s",
1004                      strerror(ret));
1005                 goto out;
1006         }
1007
1008         fc_conf->autoneg = ethpause.autoneg;
1009         if (ethpause.rx_pause && ethpause.tx_pause)
1010                 fc_conf->mode = RTE_FC_FULL;
1011         else if (ethpause.rx_pause)
1012                 fc_conf->mode = RTE_FC_RX_PAUSE;
1013         else if (ethpause.tx_pause)
1014                 fc_conf->mode = RTE_FC_TX_PAUSE;
1015         else
1016                 fc_conf->mode = RTE_FC_NONE;
1017         ret = 0;
1018
1019 out:
1020         priv_unlock(priv);
1021         assert(ret >= 0);
1022         return -ret;
1023 }
1024
1025 /**
1026  * DPDK callback to modify flow control parameters.
1027  *
1028  * @param dev
1029  *   Pointer to Ethernet device structure.
1030  * @param[in] fc_conf
1031  *   Flow control parameters.
1032  *
1033  * @return
1034  *   0 on success, negative errno value on failure.
1035  */
1036 int
1037 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1038 {
1039         struct priv *priv = dev->data->dev_private;
1040         struct ifreq ifr;
1041         struct ethtool_pauseparam ethpause = {
1042                 .cmd = ETHTOOL_SPAUSEPARAM
1043         };
1044         int ret;
1045
1046         if (mlx5_is_secondary())
1047                 return -E_RTE_SECONDARY;
1048
1049         ifr.ifr_data = (void *)&ethpause;
1050         ethpause.autoneg = fc_conf->autoneg;
1051         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1052             (fc_conf->mode & RTE_FC_RX_PAUSE))
1053                 ethpause.rx_pause = 1;
1054         else
1055                 ethpause.rx_pause = 0;
1056
1057         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1058             (fc_conf->mode & RTE_FC_TX_PAUSE))
1059                 ethpause.tx_pause = 1;
1060         else
1061                 ethpause.tx_pause = 0;
1062
1063         priv_lock(priv);
1064         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1065                 ret = errno;
1066                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1067                      " failed: %s",
1068                      strerror(ret));
1069                 goto out;
1070         }
1071         ret = 0;
1072
1073 out:
1074         priv_unlock(priv);
1075         assert(ret >= 0);
1076         return -ret;
1077 }
1078
1079 /**
1080  * Get PCI information from struct ibv_device.
1081  *
1082  * @param device
1083  *   Pointer to Ethernet device structure.
1084  * @param[out] pci_addr
1085  *   PCI bus address output buffer.
1086  *
1087  * @return
1088  *   0 on success, -1 on failure and errno is set.
1089  */
1090 int
1091 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1092                             struct rte_pci_addr *pci_addr)
1093 {
1094         FILE *file;
1095         char line[32];
1096         MKSTR(path, "%s/device/uevent", device->ibdev_path);
1097
1098         file = fopen(path, "rb");
1099         if (file == NULL)
1100                 return -1;
1101         while (fgets(line, sizeof(line), file) == line) {
1102                 size_t len = strlen(line);
1103                 int ret;
1104
1105                 /* Truncate long lines. */
1106                 if (len == (sizeof(line) - 1))
1107                         while (line[(len - 1)] != '\n') {
1108                                 ret = fgetc(file);
1109                                 if (ret == EOF)
1110                                         break;
1111                                 line[(len - 1)] = ret;
1112                         }
1113                 /* Extract information. */
1114                 if (sscanf(line,
1115                            "PCI_SLOT_NAME="
1116                            "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1117                            &pci_addr->domain,
1118                            &pci_addr->bus,
1119                            &pci_addr->devid,
1120                            &pci_addr->function) == 4) {
1121                         ret = 0;
1122                         break;
1123                 }
1124         }
1125         fclose(file);
1126         return 0;
1127 }
1128
1129 /**
1130  * Link status handler.
1131  *
1132  * @param priv
1133  *   Pointer to private structure.
1134  * @param dev
1135  *   Pointer to the rte_eth_dev structure.
1136  *
1137  * @return
1138  *   Nonzero if the callback process can be called immediately.
1139  */
1140 static int
1141 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
1142 {
1143         struct ibv_async_event event;
1144         int port_change = 0;
1145         int ret = 0;
1146
1147         /* Read all message and acknowledge them. */
1148         for (;;) {
1149                 if (ibv_get_async_event(priv->ctx, &event))
1150                         break;
1151
1152                 if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
1153                     event.event_type == IBV_EVENT_PORT_ERR)
1154                         port_change = 1;
1155                 else
1156                         DEBUG("event type %d on port %d not handled",
1157                               event.event_type, event.element.port_num);
1158                 ibv_ack_async_event(&event);
1159         }
1160
1161         if (port_change ^ priv->pending_alarm) {
1162                 struct rte_eth_link *link = &dev->data->dev_link;
1163
1164                 priv->pending_alarm = 0;
1165                 mlx5_link_update_unlocked(dev, 0);
1166                 if (((link->link_speed == 0) && link->link_status) ||
1167                     ((link->link_speed != 0) && !link->link_status)) {
1168                         /* Inconsistent status, check again later. */
1169                         priv->pending_alarm = 1;
1170                         rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1171                                           mlx5_dev_link_status_handler,
1172                                           dev);
1173                 } else
1174                         ret = 1;
1175         }
1176         return ret;
1177 }
1178
1179 /**
1180  * Handle delayed link status event.
1181  *
1182  * @param arg
1183  *   Registered argument.
1184  */
1185 void
1186 mlx5_dev_link_status_handler(void *arg)
1187 {
1188         struct rte_eth_dev *dev = arg;
1189         struct priv *priv = dev->data->dev_private;
1190         int ret;
1191
1192         priv_lock(priv);
1193         assert(priv->pending_alarm == 1);
1194         ret = priv_dev_link_status_handler(priv, dev);
1195         priv_unlock(priv);
1196         if (ret)
1197                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1198 }
1199
1200 /**
1201  * Handle interrupts from the NIC.
1202  *
1203  * @param[in] intr_handle
1204  *   Interrupt handler.
1205  * @param cb_arg
1206  *   Callback argument.
1207  */
1208 void
1209 mlx5_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
1210 {
1211         struct rte_eth_dev *dev = cb_arg;
1212         struct priv *priv = dev->data->dev_private;
1213         int ret;
1214
1215         (void)intr_handle;
1216         priv_lock(priv);
1217         ret = priv_dev_link_status_handler(priv, dev);
1218         priv_unlock(priv);
1219         if (ret)
1220                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1221 }
1222
1223 /**
1224  * Uninstall interrupt handler.
1225  *
1226  * @param priv
1227  *   Pointer to private structure.
1228  * @param dev
1229  *   Pointer to the rte_eth_dev structure.
1230  */
1231 void
1232 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1233 {
1234         if (!dev->data->dev_conf.intr_conf.lsc)
1235                 return;
1236         rte_intr_callback_unregister(&priv->intr_handle,
1237                                      mlx5_dev_interrupt_handler,
1238                                      dev);
1239         if (priv->pending_alarm)
1240                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1241         priv->pending_alarm = 0;
1242         priv->intr_handle.fd = 0;
1243         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1244 }
1245
1246 /**
1247  * Install interrupt handler.
1248  *
1249  * @param priv
1250  *   Pointer to private structure.
1251  * @param dev
1252  *   Pointer to the rte_eth_dev structure.
1253  */
1254 void
1255 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1256 {
1257         int rc, flags;
1258
1259         if (!dev->data->dev_conf.intr_conf.lsc)
1260                 return;
1261         assert(priv->ctx->async_fd > 0);
1262         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1263         rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1264         if (rc < 0) {
1265                 INFO("failed to change file descriptor async event queue");
1266                 dev->data->dev_conf.intr_conf.lsc = 0;
1267         } else {
1268                 priv->intr_handle.fd = priv->ctx->async_fd;
1269                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1270                 rte_intr_callback_register(&priv->intr_handle,
1271                                            mlx5_dev_interrupt_handler,
1272                                            dev);
1273         }
1274 }
1275
1276 /**
1277  * Change the link state (UP / DOWN).
1278  *
1279  * @param priv
1280  *   Pointer to Ethernet device structure.
1281  * @param up
1282  *   Nonzero for link up, otherwise link down.
1283  *
1284  * @return
1285  *   0 on success, errno value on failure.
1286  */
1287 static int
1288 priv_set_link(struct priv *priv, int up)
1289 {
1290         struct rte_eth_dev *dev = priv->dev;
1291         int err;
1292
1293         if (up) {
1294                 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
1295                 if (err)
1296                         return err;
1297                 priv_select_tx_function(priv);
1298                 priv_select_rx_function(priv);
1299         } else {
1300                 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
1301                 if (err)
1302                         return err;
1303                 dev->rx_pkt_burst = removed_rx_burst;
1304                 dev->tx_pkt_burst = removed_tx_burst;
1305         }
1306         return 0;
1307 }
1308
1309 /**
1310  * DPDK callback to bring the link DOWN.
1311  *
1312  * @param dev
1313  *   Pointer to Ethernet device structure.
1314  *
1315  * @return
1316  *   0 on success, errno value on failure.
1317  */
1318 int
1319 mlx5_set_link_down(struct rte_eth_dev *dev)
1320 {
1321         struct priv *priv = dev->data->dev_private;
1322         int err;
1323
1324         priv_lock(priv);
1325         err = priv_set_link(priv, 0);
1326         priv_unlock(priv);
1327         return err;
1328 }
1329
1330 /**
1331  * DPDK callback to bring the link UP.
1332  *
1333  * @param dev
1334  *   Pointer to Ethernet device structure.
1335  *
1336  * @return
1337  *   0 on success, errno value on failure.
1338  */
1339 int
1340 mlx5_set_link_up(struct rte_eth_dev *dev)
1341 {
1342         struct priv *priv = dev->data->dev_private;
1343         int err;
1344
1345         priv_lock(priv);
1346         err = priv_set_link(priv, 1);
1347         priv_unlock(priv);
1348         return err;
1349 }
1350
1351 /**
1352  * Configure secondary process queues from a private data pointer (primary
1353  * or secondary) and update burst callbacks. Can take place only once.
1354  *
1355  * All queues must have been previously created by the primary process to
1356  * avoid undefined behavior.
1357  *
1358  * @param priv
1359  *   Private data pointer from either primary or secondary process.
1360  *
1361  * @return
1362  *   Private data pointer from secondary process, NULL in case of error.
1363  */
1364 struct priv *
1365 mlx5_secondary_data_setup(struct priv *priv)
1366 {
1367         unsigned int port_id = 0;
1368         struct mlx5_secondary_data *sd;
1369         void **tx_queues;
1370         void **rx_queues;
1371         unsigned int nb_tx_queues;
1372         unsigned int nb_rx_queues;
1373         unsigned int i;
1374
1375         /* priv must be valid at this point. */
1376         assert(priv != NULL);
1377         /* priv->dev must also be valid but may point to local memory from
1378          * another process, possibly with the same address and must not
1379          * be dereferenced yet. */
1380         assert(priv->dev != NULL);
1381         /* Determine port ID by finding out where priv comes from. */
1382         while (1) {
1383                 sd = &mlx5_secondary_data[port_id];
1384                 rte_spinlock_lock(&sd->lock);
1385                 /* Primary process? */
1386                 if (sd->primary_priv == priv)
1387                         break;
1388                 /* Secondary process? */
1389                 if (sd->data.dev_private == priv)
1390                         break;
1391                 rte_spinlock_unlock(&sd->lock);
1392                 if (++port_id == RTE_DIM(mlx5_secondary_data))
1393                         port_id = 0;
1394         }
1395         /* Switch to secondary private structure. If private data has already
1396          * been updated by another thread, there is nothing else to do. */
1397         priv = sd->data.dev_private;
1398         if (priv->dev->data == &sd->data)
1399                 goto end;
1400         /* Sanity checks. Secondary private structure is supposed to point
1401          * to local eth_dev, itself still pointing to the shared device data
1402          * structure allocated by the primary process. */
1403         assert(sd->shared_dev_data != &sd->data);
1404         assert(sd->data.nb_tx_queues == 0);
1405         assert(sd->data.tx_queues == NULL);
1406         assert(sd->data.nb_rx_queues == 0);
1407         assert(sd->data.rx_queues == NULL);
1408         assert(priv != sd->primary_priv);
1409         assert(priv->dev->data == sd->shared_dev_data);
1410         assert(priv->txqs_n == 0);
1411         assert(priv->txqs == NULL);
1412         assert(priv->rxqs_n == 0);
1413         assert(priv->rxqs == NULL);
1414         nb_tx_queues = sd->shared_dev_data->nb_tx_queues;
1415         nb_rx_queues = sd->shared_dev_data->nb_rx_queues;
1416         /* Allocate local storage for queues. */
1417         tx_queues = rte_zmalloc("secondary ethdev->tx_queues",
1418                                 sizeof(sd->data.tx_queues[0]) * nb_tx_queues,
1419                                 RTE_CACHE_LINE_SIZE);
1420         rx_queues = rte_zmalloc("secondary ethdev->rx_queues",
1421                                 sizeof(sd->data.rx_queues[0]) * nb_rx_queues,
1422                                 RTE_CACHE_LINE_SIZE);
1423         if (tx_queues == NULL || rx_queues == NULL)
1424                 goto error;
1425         /* Lock to prevent control operations during setup. */
1426         priv_lock(priv);
1427         /* TX queues. */
1428         for (i = 0; i != nb_tx_queues; ++i) {
1429                 struct txq *primary_txq = (*sd->primary_priv->txqs)[i];
1430                 struct txq_ctrl *primary_txq_ctrl;
1431                 struct txq_ctrl *txq_ctrl;
1432
1433                 if (primary_txq == NULL)
1434                         continue;
1435                 primary_txq_ctrl = container_of(primary_txq,
1436                                                 struct txq_ctrl, txq);
1437                 txq_ctrl = rte_calloc_socket("TXQ", 1, sizeof(*txq_ctrl) +
1438                                              (1 << primary_txq->elts_n) *
1439                                              sizeof(struct rte_mbuf *), 0,
1440                                              primary_txq_ctrl->socket);
1441                 if (txq_ctrl != NULL) {
1442                         if (txq_ctrl_setup(priv->dev,
1443                                            txq_ctrl,
1444                                            1 << primary_txq->elts_n,
1445                                            primary_txq_ctrl->socket,
1446                                            NULL) == 0) {
1447                                 txq_ctrl->txq.stats.idx =
1448                                         primary_txq->stats.idx;
1449                                 tx_queues[i] = &txq_ctrl->txq;
1450                                 continue;
1451                         }
1452                         rte_free(txq_ctrl);
1453                 }
1454                 while (i) {
1455                         txq_ctrl = tx_queues[--i];
1456                         txq_cleanup(txq_ctrl);
1457                         rte_free(txq_ctrl);
1458                 }
1459                 goto error;
1460         }
1461         /* RX queues. */
1462         for (i = 0; i != nb_rx_queues; ++i) {
1463                 struct rxq_ctrl *primary_rxq =
1464                         container_of((*sd->primary_priv->rxqs)[i],
1465                                      struct rxq_ctrl, rxq);
1466
1467                 if (primary_rxq == NULL)
1468                         continue;
1469                 /* Not supported yet. */
1470                 rx_queues[i] = NULL;
1471         }
1472         /* Update everything. */
1473         priv->txqs = (void *)tx_queues;
1474         priv->txqs_n = nb_tx_queues;
1475         priv->rxqs = (void *)rx_queues;
1476         priv->rxqs_n = nb_rx_queues;
1477         sd->data.rx_queues = rx_queues;
1478         sd->data.tx_queues = tx_queues;
1479         sd->data.nb_rx_queues = nb_rx_queues;
1480         sd->data.nb_tx_queues = nb_tx_queues;
1481         sd->data.dev_link = sd->shared_dev_data->dev_link;
1482         sd->data.mtu = sd->shared_dev_data->mtu;
1483         memcpy(sd->data.rx_queue_state, sd->shared_dev_data->rx_queue_state,
1484                sizeof(sd->data.rx_queue_state));
1485         memcpy(sd->data.tx_queue_state, sd->shared_dev_data->tx_queue_state,
1486                sizeof(sd->data.tx_queue_state));
1487         sd->data.dev_flags = sd->shared_dev_data->dev_flags;
1488         /* Use local data from now on. */
1489         rte_mb();
1490         priv->dev->data = &sd->data;
1491         rte_mb();
1492         priv_select_tx_function(priv);
1493         priv_select_rx_function(priv);
1494         priv_unlock(priv);
1495 end:
1496         /* More sanity checks. */
1497         assert(priv->dev->data == &sd->data);
1498         rte_spinlock_unlock(&sd->lock);
1499         return priv;
1500 error:
1501         priv_unlock(priv);
1502         rte_free(tx_queues);
1503         rte_free(rx_queues);
1504         rte_spinlock_unlock(&sd->lock);
1505         return NULL;
1506 }
1507
1508 /**
1509  * Configure the TX function to use.
1510  *
1511  * @param priv
1512  *   Pointer to private structure.
1513  */
1514 void
1515 priv_select_tx_function(struct priv *priv)
1516 {
1517         priv->dev->tx_pkt_burst = mlx5_tx_burst;
1518         /* Display warning for unsupported configurations. */
1519         if (priv->sriov && priv->mps)
1520                 WARN("multi-packet send WQE cannot be used on a SR-IOV setup");
1521         /* Select appropriate TX function. */
1522         if ((priv->sriov == 0) && priv->mps && priv->txq_inline) {
1523                 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1524                 DEBUG("selected MPW inline TX function");
1525         } else if ((priv->sriov == 0) && priv->mps) {
1526                 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw;
1527                 DEBUG("selected MPW TX function");
1528         }
1529 }
1530
1531 /**
1532  * Configure the RX function to use.
1533  *
1534  * @param priv
1535  *   Pointer to private structure.
1536  */
1537 void
1538 priv_select_rx_function(struct priv *priv)
1539 {
1540         priv->dev->rx_pkt_burst = mlx5_rx_burst;
1541 }