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         info->pci_dev = RTE_DEV_TO_PCI(dev->device);
 566
 567         priv_lock(priv);
 568         /* FIXME: we should ask the device for these values. */
 569         info->min_rx_bufsize = 32;
 570         info->max_rx_pktlen = 65536;
 571         /*
 572          * Since we need one CQ per QP, the limit is the minimum number
 573          * between the two values.
 574          */
 575         max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
 576                priv->device_attr.max_qp : priv->device_attr.max_cq);
 577         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
 578         if (max >= 65535)
 579                 max = 65535;
 580         info->max_rx_queues = max;
 581         info->max_tx_queues = max;
 582         info->max_mac_addrs = RTE_DIM(priv->mac);
 583         info->rx_offload_capa =
 584                 (priv->hw_csum ?
 585                  (DEV_RX_OFFLOAD_IPV4_CKSUM |
 586                   DEV_RX_OFFLOAD_UDP_CKSUM |
 587                   DEV_RX_OFFLOAD_TCP_CKSUM) :
 588                  0) |
 589                 (priv->hw_vlan_strip ? DEV_RX_OFFLOAD_VLAN_STRIP : 0);
 590         if (!priv->mps)
 591                 info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
 592         if (priv->hw_csum)
 593                 info->tx_offload_capa |=
 594                         (DEV_TX_OFFLOAD_IPV4_CKSUM |
 595                          DEV_TX_OFFLOAD_UDP_CKSUM |
 596                          DEV_TX_OFFLOAD_TCP_CKSUM);
 597         if (priv_get_ifname(priv, &ifname) == 0)
 598                 info->if_index = if_nametoindex(ifname);
 599         /* FIXME: RETA update/query API expects the callee to know the size of
 600          * the indirection table, for this PMD the size varies depending on
 601          * the number of RX queues, it becomes impossible to find the correct
 602          * size if it is not fixed.
 603          * The API should be updated to solve this problem. */
 604         info->reta_size = priv->ind_table_max_size;
 605         info->hash_key_size = ((*priv->rss_conf) ?
 606                                (*priv->rss_conf)[0]->rss_key_len :
 607                                0);
 608         info->speed_capa = priv->link_speed_capa;
 609         priv_unlock(priv);
 610 }
 611
 612 const uint32_t *
 613 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 614 {
 615         static const uint32_t ptypes[] = {
 616                 /* refers to rxq_cq_to_pkt_type() */
 617                 RTE_PTYPE_L3_IPV4,
 618                 RTE_PTYPE_L3_IPV6,
 619                 RTE_PTYPE_INNER_L3_IPV4,
 620                 RTE_PTYPE_INNER_L3_IPV6,
 621                 RTE_PTYPE_UNKNOWN
 622
 623         };
 624
 625         if (dev->rx_pkt_burst == mlx5_rx_burst)
 626                 return ptypes;
 627         return NULL;
 628 }
 629
 630 /**
 631  * DPDK callback to retrieve physical link information.
 632  *
 633  * @param dev
 634  *   Pointer to Ethernet device structure.
 635  * @param wait_to_complete
 636  *   Wait for request completion (ignored).
 637  */
 638 static int
 639 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
 640 {
 641         struct priv *priv = mlx5_get_priv(dev);
 642         struct ethtool_cmd edata = {
 643                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 644         };
 645         struct ifreq ifr;
 646         struct rte_eth_link dev_link;
 647         int link_speed = 0;
 648
 649         /* priv_lock() is not taken to allow concurrent calls. */
 650
 651         (void)wait_to_complete;
 652         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
 653                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
 654                 return -1;
 655         }
 656         memset(&dev_link, 0, sizeof(dev_link));
 657         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 658                                 (ifr.ifr_flags & IFF_RUNNING));
 659         ifr.ifr_data = (void *)&edata;
 660         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 661                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
 662                      strerror(errno));
 663                 return -1;
 664         }
 665         link_speed = ethtool_cmd_speed(&edata);
 666         if (link_speed == -1)
 667                 dev_link.link_speed = 0;
 668         else
 669                 dev_link.link_speed = link_speed;
 670         priv->link_speed_capa = 0;
 671         if (edata.supported & SUPPORTED_Autoneg)
 672                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 673         if (edata.supported & (SUPPORTED_1000baseT_Full |
 674                                SUPPORTED_1000baseKX_Full))
 675                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 676         if (edata.supported & SUPPORTED_10000baseKR_Full)
 677                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 678         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 679                                SUPPORTED_40000baseCR4_Full |
 680                                SUPPORTED_40000baseSR4_Full |
 681                                SUPPORTED_40000baseLR4_Full))
 682                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 683         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 684                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 685         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 686                         ETH_LINK_SPEED_FIXED);
 687         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 688                 /* Link status changed. */
 689                 dev->data->dev_link = dev_link;
 690                 return 0;
 691         }
 692         /* Link status is still the same. */
 693         return -1;
 694 }
 695
 696 /**
 697  * Retrieve physical link information (unlocked version using new ioctl from
 698  * Linux 4.5).
 699  *
 700  * @param dev
 701  *   Pointer to Ethernet device structure.
 702  * @param wait_to_complete
 703  *   Wait for request completion (ignored).
 704  */
 705 static int
 706 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
 707 {
 708 #ifdef ETHTOOL_GLINKSETTINGS
 709         struct priv *priv = mlx5_get_priv(dev);
 710         struct ethtool_link_settings edata = {
 711                 .cmd = ETHTOOL_GLINKSETTINGS,
 712         };
 713         struct ifreq ifr;
 714         struct rte_eth_link dev_link;
 715         uint64_t sc;
 716
 717         (void)wait_to_complete;
 718         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
 719                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
 720                 return -1;
 721         }
 722         memset(&dev_link, 0, sizeof(dev_link));
 723         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 724                                 (ifr.ifr_flags & IFF_RUNNING));
 725         ifr.ifr_data = (void *)&edata;
 726         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 727                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
 728                       strerror(errno));
 729                 return -1;
 730         }
 731         dev_link.link_speed = edata.speed;
 732         sc = edata.link_mode_masks[0] |
 733                 ((uint64_t)edata.link_mode_masks[1] << 32);
 734         priv->link_speed_capa = 0;
 735         /* Link speeds available in kernel v4.5. */
 736         if (sc & ETHTOOL_LINK_MODE_Autoneg_BIT)
 737                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 738         if (sc & (ETHTOOL_LINK_MODE_1000baseT_Full_BIT |
 739                   ETHTOOL_LINK_MODE_1000baseKX_Full_BIT))
 740                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 741         if (sc & (ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT |
 742                   ETHTOOL_LINK_MODE_10000baseKR_Full_BIT |
 743                   ETHTOOL_LINK_MODE_10000baseR_FEC_BIT))
 744                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 745         if (sc & (ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT |
 746                   ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT))
 747                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
 748         if (sc & (ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT |
 749                   ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT |
 750                   ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT |
 751                   ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT))
 752                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 753         if (sc & (ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT |
 754                   ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT |
 755                   ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT |
 756                   ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT))
 757                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
 758         /* Link speeds available in kernel v4.6. */
 759 #ifdef HAVE_ETHTOOL_LINK_MODE_25G
 760         if (sc & (ETHTOOL_LINK_MODE_25000baseCR_Full_BIT |
 761                   ETHTOOL_LINK_MODE_25000baseKR_Full_BIT |
 762                   ETHTOOL_LINK_MODE_25000baseSR_Full_BIT))
 763                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
 764 #endif
 765 #ifdef HAVE_ETHTOOL_LINK_MODE_50G
 766         if (sc & (ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT |
 767                   ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT))
 768                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
 769 #endif
 770 #ifdef HAVE_ETHTOOL_LINK_MODE_100G
 771         if (sc & (ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT |
 772                   ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT |
 773                   ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT |
 774                   ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT))
 775                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
 776 #endif
 777         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 778                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 779         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 780                                   ETH_LINK_SPEED_FIXED);
 781         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 782                 /* Link status changed. */
 783                 dev->data->dev_link = dev_link;
 784                 return 0;
 785         }
 786 #else
 787         (void)dev;
 788         (void)wait_to_complete;
 789 #endif
 790         /* Link status is still the same. */
 791         return -1;
 792 }
 793
 794 /**
 795  * DPDK callback to retrieve physical link information.
 796  *
 797  * @param dev
 798  *   Pointer to Ethernet device structure.
 799  * @param wait_to_complete
 800  *   Wait for request completion (ignored).
 801  */
 802 int
 803 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
 804 {
 805         int ret;
 806
 807         ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
 808         if (ret < 0)
 809                 ret = mlx5_link_update_unlocked_gset(dev, wait_to_complete);
 810         return ret;
 811 }
 812
 813 /**
 814  * DPDK callback to change the MTU.
 815  *
 816  * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
 817  * received). Use this as a hint to enable/disable scattered packets support
 818  * and improve performance when not needed.
 819  * Since failure is not an option, reconfiguring queues on the fly is not
 820  * recommended.
 821  *
 822  * @param dev
 823  *   Pointer to Ethernet device structure.
 824  * @param in_mtu
 825  *   New MTU.
 826  *
 827  * @return
 828  *   0 on success, negative errno value on failure.
 829  */
 830 int
 831 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 832 {
 833         struct priv *priv = dev->data->dev_private;
 834         int ret = 0;
 835         unsigned int i;
 836         uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
 837                 mlx5_rx_burst;
 838         unsigned int max_frame_len;
 839         int rehash;
 840         int restart = priv->started;
 841
 842         if (mlx5_is_secondary())
 843                 return -E_RTE_SECONDARY;
 844
 845         priv_lock(priv);
 846         /* Set kernel interface MTU first. */
 847         if (priv_set_mtu(priv, mtu)) {
 848                 ret = errno;
 849                 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
 850                      strerror(ret));
 851                 goto out;
 852         } else
 853                 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
 854         /* Temporarily replace RX handler with a fake one, assuming it has not
 855          * been copied elsewhere. */
 856         dev->rx_pkt_burst = removed_rx_burst;
 857         /* Make sure everyone has left mlx5_rx_burst() and uses
 858          * removed_rx_burst() instead. */
 859         rte_wmb();
 860         usleep(1000);
 861         /* MTU does not include header and CRC. */
 862         max_frame_len = ETHER_HDR_LEN + mtu + ETHER_CRC_LEN;
 863         /* Check if at least one queue is going to need a SGE update. */
 864         for (i = 0; i != priv->rxqs_n; ++i) {
 865                 struct rxq *rxq = (*priv->rxqs)[i];
 866                 unsigned int mb_len;
 867                 unsigned int size = RTE_PKTMBUF_HEADROOM + max_frame_len;
 868                 unsigned int sges_n;
 869
 870                 if (rxq == NULL)
 871                         continue;
 872                 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
 873                 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
 874                 /*
 875                  * Determine the number of SGEs needed for a full packet
 876                  * and round it to the next power of two.
 877                  */
 878                 sges_n = log2above((size / mb_len) + !!(size % mb_len));
 879                 if (sges_n != rxq->sges_n)
 880                         break;
 881         }
 882         /*
 883          * If all queues have the right number of SGEs, a simple rehash
 884          * of their buffers is enough, otherwise SGE information can only
 885          * be updated in a queue by recreating it. All resources that depend
 886          * on queues (flows, indirection tables) must be recreated as well in
 887          * that case.
 888          */
 889         rehash = (i == priv->rxqs_n);
 890         if (!rehash) {
 891                 /* Clean up everything as with mlx5_dev_stop(). */
 892                 priv_special_flow_disable_all(priv);
 893                 priv_mac_addrs_disable(priv);
 894                 priv_destroy_hash_rxqs(priv);
 895                 priv_fdir_disable(priv);
 896                 priv_dev_interrupt_handler_uninstall(priv, dev);
 897         }
 898 recover:
 899         /* Reconfigure each RX queue. */
 900         for (i = 0; (i != priv->rxqs_n); ++i) {
 901                 struct rxq *rxq = (*priv->rxqs)[i];
 902                 struct rxq_ctrl *rxq_ctrl =
 903                         container_of(rxq, struct rxq_ctrl, rxq);
 904                 int sp;
 905                 unsigned int mb_len;
 906                 unsigned int tmp;
 907
 908                 if (rxq == NULL)
 909                         continue;
 910                 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
 911                 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
 912                 /* Toggle scattered support (sp) if necessary. */
 913                 sp = (max_frame_len > (mb_len - RTE_PKTMBUF_HEADROOM));
 914                 /* Provide new values to rxq_setup(). */
 915                 dev->data->dev_conf.rxmode.jumbo_frame = sp;
 916                 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
 917                 if (rehash)
 918                         ret = rxq_rehash(dev, rxq_ctrl);
 919                 else
 920                         ret = rxq_ctrl_setup(dev, rxq_ctrl, 1 << rxq->elts_n,
 921                                              rxq_ctrl->socket, NULL, rxq->mp);
 922                 if (!ret)
 923                         continue;
 924                 /* Attempt to roll back in case of error. */
 925                 tmp = (mb_len << rxq->sges_n) - RTE_PKTMBUF_HEADROOM;
 926                 if (max_frame_len != tmp) {
 927                         max_frame_len = tmp;
 928                         goto recover;
 929                 }
 930                 /* Double fault, disable RX. */
 931                 break;
 932         }
 933         /*
 934          * Use a safe RX burst function in case of error, otherwise mimic
 935          * mlx5_dev_start().
 936          */
 937         if (ret) {
 938                 ERROR("unable to reconfigure RX queues, RX disabled");
 939                 rx_func = removed_rx_burst;
 940         } else if (restart &&
 941                  !rehash &&
 942                  !priv_create_hash_rxqs(priv) &&
 943                  !priv_rehash_flows(priv)) {
 944                 if (dev->data->dev_conf.fdir_conf.mode == RTE_FDIR_MODE_NONE)
 945                         priv_fdir_enable(priv);
 946                 priv_dev_interrupt_handler_install(priv, dev);
 947         }
 948         priv->mtu = mtu;
 949         /* Burst functions can now be called again. */
 950         rte_wmb();
 951         dev->rx_pkt_burst = rx_func;
 952 out:
 953         priv_unlock(priv);
 954         assert(ret >= 0);
 955         return -ret;
 956 }
 957
 958 /**
 959  * DPDK callback to get flow control status.
 960  *
 961  * @param dev
 962  *   Pointer to Ethernet device structure.
 963  * @param[out] fc_conf
 964  *   Flow control output buffer.
 965  *
 966  * @return
 967  *   0 on success, negative errno value on failure.
 968  */
 969 int
 970 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 971 {
 972         struct priv *priv = dev->data->dev_private;
 973         struct ifreq ifr;
 974         struct ethtool_pauseparam ethpause = {
 975                 .cmd = ETHTOOL_GPAUSEPARAM
 976         };
 977         int ret;
 978
 979         if (mlx5_is_secondary())
 980                 return -E_RTE_SECONDARY;
 981
 982         ifr.ifr_data = (void *)&ethpause;
 983         priv_lock(priv);
 984         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 985                 ret = errno;
 986                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
 987                      " failed: %s",
 988                      strerror(ret));
 989                 goto out;
 990         }
 991
 992         fc_conf->autoneg = ethpause.autoneg;
 993         if (ethpause.rx_pause && ethpause.tx_pause)
 994                 fc_conf->mode = RTE_FC_FULL;
 995         else if (ethpause.rx_pause)
 996                 fc_conf->mode = RTE_FC_RX_PAUSE;
 997         else if (ethpause.tx_pause)
 998                 fc_conf->mode = RTE_FC_TX_PAUSE;
 999         else
1000                 fc_conf->mode = RTE_FC_NONE;
1001         ret = 0;
1002
1003 out:
1004         priv_unlock(priv);
1005         assert(ret >= 0);
1006         return -ret;
1007 }
1008
1009 /**
1010  * DPDK callback to modify flow control parameters.
1011  *
1012  * @param dev
1013  *   Pointer to Ethernet device structure.
1014  * @param[in] fc_conf
1015  *   Flow control parameters.
1016  *
1017  * @return
1018  *   0 on success, negative errno value on failure.
1019  */
1020 int
1021 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1022 {
1023         struct priv *priv = dev->data->dev_private;
1024         struct ifreq ifr;
1025         struct ethtool_pauseparam ethpause = {
1026                 .cmd = ETHTOOL_SPAUSEPARAM
1027         };
1028         int ret;
1029
1030         if (mlx5_is_secondary())
1031                 return -E_RTE_SECONDARY;
1032
1033         ifr.ifr_data = (void *)&ethpause;
1034         ethpause.autoneg = fc_conf->autoneg;
1035         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1036             (fc_conf->mode & RTE_FC_RX_PAUSE))
1037                 ethpause.rx_pause = 1;
1038         else
1039                 ethpause.rx_pause = 0;
1040
1041         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1042             (fc_conf->mode & RTE_FC_TX_PAUSE))
1043                 ethpause.tx_pause = 1;
1044         else
1045                 ethpause.tx_pause = 0;
1046
1047         priv_lock(priv);
1048         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1049                 ret = errno;
1050                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1051                      " failed: %s",
1052                      strerror(ret));
1053                 goto out;
1054         }
1055         ret = 0;
1056
1057 out:
1058         priv_unlock(priv);
1059         assert(ret >= 0);
1060         return -ret;
1061 }
1062
1063 /**
1064  * Get PCI information from struct ibv_device.
1065  *
1066  * @param device
1067  *   Pointer to Ethernet device structure.
1068  * @param[out] pci_addr
1069  *   PCI bus address output buffer.
1070  *
1071  * @return
1072  *   0 on success, -1 on failure and errno is set.
1073  */
1074 int
1075 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1076                             struct rte_pci_addr *pci_addr)
1077 {
1078         FILE *file;
1079         char line[32];
1080         MKSTR(path, "%s/device/uevent", device->ibdev_path);
1081
1082         file = fopen(path, "rb");
1083         if (file == NULL)
1084                 return -1;
1085         while (fgets(line, sizeof(line), file) == line) {
1086                 size_t len = strlen(line);
1087                 int ret;
1088
1089                 /* Truncate long lines. */
1090                 if (len == (sizeof(line) - 1))
1091                         while (line[(len - 1)] != '\n') {
1092                                 ret = fgetc(file);
1093                                 if (ret == EOF)
1094                                         break;
1095                                 line[(len - 1)] = ret;
1096                         }
1097                 /* Extract information. */
1098                 if (sscanf(line,
1099                            "PCI_SLOT_NAME="
1100                            "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1101                            &pci_addr->domain,
1102                            &pci_addr->bus,
1103                            &pci_addr->devid,
1104                            &pci_addr->function) == 4) {
1105                         ret = 0;
1106                         break;
1107                 }
1108         }
1109         fclose(file);
1110         return 0;
1111 }
1112
1113 /**
1114  * Link status handler.
1115  *
1116  * @param priv
1117  *   Pointer to private structure.
1118  * @param dev
1119  *   Pointer to the rte_eth_dev structure.
1120  *
1121  * @return
1122  *   Nonzero if the callback process can be called immediately.
1123  */
1124 static int
1125 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
1126 {
1127         struct ibv_async_event event;
1128         struct rte_eth_link *link = &dev->data->dev_link;
1129         int ret = 0;
1130
1131         /* Read all message and acknowledge them. */
1132         for (;;) {
1133                 if (ibv_get_async_event(priv->ctx, &event))
1134                         break;
1135
1136                 if (event.event_type != IBV_EVENT_PORT_ACTIVE &&
1137                     event.event_type != IBV_EVENT_PORT_ERR)
1138                         DEBUG("event type %d on port %d not handled",
1139                               event.event_type, event.element.port_num);
1140                 ibv_ack_async_event(&event);
1141         }
1142         mlx5_link_update(dev, 0);
1143         if (((link->link_speed == 0) && link->link_status) ||
1144             ((link->link_speed != 0) && !link->link_status)) {
1145                 if (!priv->pending_alarm) {
1146                         /* Inconsistent status, check again later. */
1147                         priv->pending_alarm = 1;
1148                         rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1149                                           mlx5_dev_link_status_handler,
1150                                           dev);
1151                 }
1152         } else {
1153                 ret = 1;
1154         }
1155         return ret;
1156 }
1157
1158 /**
1159  * Handle delayed link status event.
1160  *
1161  * @param arg
1162  *   Registered argument.
1163  */
1164 void
1165 mlx5_dev_link_status_handler(void *arg)
1166 {
1167         struct rte_eth_dev *dev = arg;
1168         struct priv *priv = dev->data->dev_private;
1169         int ret;
1170
1171         priv_lock(priv);
1172         assert(priv->pending_alarm == 1);
1173         priv->pending_alarm = 0;
1174         ret = priv_dev_link_status_handler(priv, dev);
1175         priv_unlock(priv);
1176         if (ret)
1177                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1178 }
1179
1180 /**
1181  * Handle interrupts from the NIC.
1182  *
1183  * @param[in] intr_handle
1184  *   Interrupt handler.
1185  * @param cb_arg
1186  *   Callback argument.
1187  */
1188 void
1189 mlx5_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
1190 {
1191         struct rte_eth_dev *dev = cb_arg;
1192         struct priv *priv = dev->data->dev_private;
1193         int ret;
1194
1195         (void)intr_handle;
1196         priv_lock(priv);
1197         ret = priv_dev_link_status_handler(priv, dev);
1198         priv_unlock(priv);
1199         if (ret)
1200                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1201 }
1202
1203 /**
1204  * Uninstall interrupt handler.
1205  *
1206  * @param priv
1207  *   Pointer to private structure.
1208  * @param dev
1209  *   Pointer to the rte_eth_dev structure.
1210  */
1211 void
1212 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1213 {
1214         if (!dev->data->dev_conf.intr_conf.lsc)
1215                 return;
1216         rte_intr_callback_unregister(&priv->intr_handle,
1217                                      mlx5_dev_interrupt_handler,
1218                                      dev);
1219         if (priv->pending_alarm)
1220                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1221         priv->pending_alarm = 0;
1222         priv->intr_handle.fd = 0;
1223         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1224 }
1225
1226 /**
1227  * Install interrupt handler.
1228  *
1229  * @param priv
1230  *   Pointer to private structure.
1231  * @param dev
1232  *   Pointer to the rte_eth_dev structure.
1233  */
1234 void
1235 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1236 {
1237         int rc, flags;
1238
1239         if (!dev->data->dev_conf.intr_conf.lsc)
1240                 return;
1241         assert(priv->ctx->async_fd > 0);
1242         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1243         rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1244         if (rc < 0) {
1245                 INFO("failed to change file descriptor async event queue");
1246                 dev->data->dev_conf.intr_conf.lsc = 0;
1247         } else {
1248                 priv->intr_handle.fd = priv->ctx->async_fd;
1249                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1250                 rte_intr_callback_register(&priv->intr_handle,
1251                                            mlx5_dev_interrupt_handler,
1252                                            dev);
1253         }
1254 }
1255
1256 /**
1257  * Change the link state (UP / DOWN).
1258  *
1259  * @param priv
1260  *   Pointer to Ethernet device structure.
1261  * @param up
1262  *   Nonzero for link up, otherwise link down.
1263  *
1264  * @return
1265  *   0 on success, errno value on failure.
1266  */
1267 static int
1268 priv_set_link(struct priv *priv, int up)
1269 {
1270         struct rte_eth_dev *dev = priv->dev;
1271         int err;
1272
1273         if (up) {
1274                 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
1275                 if (err)
1276                         return err;
1277                 priv_select_tx_function(priv);
1278                 priv_select_rx_function(priv);
1279         } else {
1280                 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
1281                 if (err)
1282                         return err;
1283                 dev->rx_pkt_burst = removed_rx_burst;
1284                 dev->tx_pkt_burst = removed_tx_burst;
1285         }
1286         return 0;
1287 }
1288
1289 /**
1290  * DPDK callback to bring the link DOWN.
1291  *
1292  * @param dev
1293  *   Pointer to Ethernet device structure.
1294  *
1295  * @return
1296  *   0 on success, errno value on failure.
1297  */
1298 int
1299 mlx5_set_link_down(struct rte_eth_dev *dev)
1300 {
1301         struct priv *priv = dev->data->dev_private;
1302         int err;
1303
1304         priv_lock(priv);
1305         err = priv_set_link(priv, 0);
1306         priv_unlock(priv);
1307         return err;
1308 }
1309
1310 /**
1311  * DPDK callback to bring the link UP.
1312  *
1313  * @param dev
1314  *   Pointer to Ethernet device structure.
1315  *
1316  * @return
1317  *   0 on success, errno value on failure.
1318  */
1319 int
1320 mlx5_set_link_up(struct rte_eth_dev *dev)
1321 {
1322         struct priv *priv = dev->data->dev_private;
1323         int err;
1324
1325         priv_lock(priv);
1326         err = priv_set_link(priv, 1);
1327         priv_unlock(priv);
1328         return err;
1329 }
1330
1331 /**
1332  * Configure secondary process queues from a private data pointer (primary
1333  * or secondary) and update burst callbacks. Can take place only once.
1334  *
1335  * All queues must have been previously created by the primary process to
1336  * avoid undefined behavior.
1337  *
1338  * @param priv
1339  *   Private data pointer from either primary or secondary process.
1340  *
1341  * @return
1342  *   Private data pointer from secondary process, NULL in case of error.
1343  */
1344 struct priv *
1345 mlx5_secondary_data_setup(struct priv *priv)
1346 {
1347         unsigned int port_id = 0;
1348         struct mlx5_secondary_data *sd;
1349         void **tx_queues;
1350         void **rx_queues;
1351         unsigned int nb_tx_queues;
1352         unsigned int nb_rx_queues;
1353         unsigned int i;
1354
1355         /* priv must be valid at this point. */
1356         assert(priv != NULL);
1357         /* priv->dev must also be valid but may point to local memory from
1358          * another process, possibly with the same address and must not
1359          * be dereferenced yet. */
1360         assert(priv->dev != NULL);
1361         /* Determine port ID by finding out where priv comes from. */
1362         while (1) {
1363                 sd = &mlx5_secondary_data[port_id];
1364                 rte_spinlock_lock(&sd->lock);
1365                 /* Primary process? */
1366                 if (sd->primary_priv == priv)
1367                         break;
1368                 /* Secondary process? */
1369                 if (sd->data.dev_private == priv)
1370                         break;
1371                 rte_spinlock_unlock(&sd->lock);
1372                 if (++port_id == RTE_DIM(mlx5_secondary_data))
1373                         port_id = 0;
1374         }
1375         /* Switch to secondary private structure. If private data has already
1376          * been updated by another thread, there is nothing else to do. */
1377         priv = sd->data.dev_private;
1378         if (priv->dev->data == &sd->data)
1379                 goto end;
1380         /* Sanity checks. Secondary private structure is supposed to point
1381          * to local eth_dev, itself still pointing to the shared device data
1382          * structure allocated by the primary process. */
1383         assert(sd->shared_dev_data != &sd->data);
1384         assert(sd->data.nb_tx_queues == 0);
1385         assert(sd->data.tx_queues == NULL);
1386         assert(sd->data.nb_rx_queues == 0);
1387         assert(sd->data.rx_queues == NULL);
1388         assert(priv != sd->primary_priv);
1389         assert(priv->dev->data == sd->shared_dev_data);
1390         assert(priv->txqs_n == 0);
1391         assert(priv->txqs == NULL);
1392         assert(priv->rxqs_n == 0);
1393         assert(priv->rxqs == NULL);
1394         nb_tx_queues = sd->shared_dev_data->nb_tx_queues;
1395         nb_rx_queues = sd->shared_dev_data->nb_rx_queues;
1396         /* Allocate local storage for queues. */
1397         tx_queues = rte_zmalloc("secondary ethdev->tx_queues",
1398                                 sizeof(sd->data.tx_queues[0]) * nb_tx_queues,
1399                                 RTE_CACHE_LINE_SIZE);
1400         rx_queues = rte_zmalloc("secondary ethdev->rx_queues",
1401                                 sizeof(sd->data.rx_queues[0]) * nb_rx_queues,
1402                                 RTE_CACHE_LINE_SIZE);
1403         if (tx_queues == NULL || rx_queues == NULL)
1404                 goto error;
1405         /* Lock to prevent control operations during setup. */
1406         priv_lock(priv);
1407         /* TX queues. */
1408         for (i = 0; i != nb_tx_queues; ++i) {
1409                 struct txq *primary_txq = (*sd->primary_priv->txqs)[i];
1410                 struct txq_ctrl *primary_txq_ctrl;
1411                 struct txq_ctrl *txq_ctrl;
1412
1413                 if (primary_txq == NULL)
1414                         continue;
1415                 primary_txq_ctrl = container_of(primary_txq,
1416                                                 struct txq_ctrl, txq);
1417                 txq_ctrl = rte_calloc_socket("TXQ", 1, sizeof(*txq_ctrl) +
1418                                              (1 << primary_txq->elts_n) *
1419                                              sizeof(struct rte_mbuf *), 0,
1420                                              primary_txq_ctrl->socket);
1421                 if (txq_ctrl != NULL) {
1422                         if (txq_ctrl_setup(priv->dev,
1423                                            txq_ctrl,
1424                                            1 << primary_txq->elts_n,
1425                                            primary_txq_ctrl->socket,
1426                                            NULL) == 0) {
1427                                 txq_ctrl->txq.stats.idx =
1428                                         primary_txq->stats.idx;
1429                                 tx_queues[i] = &txq_ctrl->txq;
1430                                 continue;
1431                         }
1432                         rte_free(txq_ctrl);
1433                 }
1434                 while (i) {
1435                         txq_ctrl = tx_queues[--i];
1436                         txq_cleanup(txq_ctrl);
1437                         rte_free(txq_ctrl);
1438                 }
1439                 goto error;
1440         }
1441         /* RX queues. */
1442         for (i = 0; i != nb_rx_queues; ++i) {
1443                 struct rxq_ctrl *primary_rxq =
1444                         container_of((*sd->primary_priv->rxqs)[i],
1445                                      struct rxq_ctrl, rxq);
1446
1447                 if (primary_rxq == NULL)
1448                         continue;
1449                 /* Not supported yet. */
1450                 rx_queues[i] = NULL;
1451         }
1452         /* Update everything. */
1453         priv->txqs = (void *)tx_queues;
1454         priv->txqs_n = nb_tx_queues;
1455         priv->rxqs = (void *)rx_queues;
1456         priv->rxqs_n = nb_rx_queues;
1457         sd->data.rx_queues = rx_queues;
1458         sd->data.tx_queues = tx_queues;
1459         sd->data.nb_rx_queues = nb_rx_queues;
1460         sd->data.nb_tx_queues = nb_tx_queues;
1461         sd->data.dev_link = sd->shared_dev_data->dev_link;
1462         sd->data.mtu = sd->shared_dev_data->mtu;
1463         memcpy(sd->data.rx_queue_state, sd->shared_dev_data->rx_queue_state,
1464                sizeof(sd->data.rx_queue_state));
1465         memcpy(sd->data.tx_queue_state, sd->shared_dev_data->tx_queue_state,
1466                sizeof(sd->data.tx_queue_state));
1467         sd->data.dev_flags = sd->shared_dev_data->dev_flags;
1468         /* Use local data from now on. */
1469         rte_mb();
1470         priv->dev->data = &sd->data;
1471         rte_mb();
1472         priv_select_tx_function(priv);
1473         priv_select_rx_function(priv);
1474         priv_unlock(priv);
1475 end:
1476         /* More sanity checks. */
1477         assert(priv->dev->data == &sd->data);
1478         rte_spinlock_unlock(&sd->lock);
1479         return priv;
1480 error:
1481         priv_unlock(priv);
1482         rte_free(tx_queues);
1483         rte_free(rx_queues);
1484         rte_spinlock_unlock(&sd->lock);
1485         return NULL;
1486 }
1487
1488 /**
1489  * Configure the TX function to use.
1490  *
1491  * @param priv
1492  *   Pointer to private structure.
1493  */
1494 void
1495 priv_select_tx_function(struct priv *priv)
1496 {
1497         priv->dev->tx_pkt_burst = mlx5_tx_burst;
1498         /* Select appropriate TX function. */
1499         if (priv->mps && priv->txq_inline) {
1500                 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1501                 DEBUG("selected MPW inline TX function");
1502         } else if (priv->mps) {
1503                 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw;
1504                 DEBUG("selected MPW TX function");
1505         }
1506 }
1507
1508 /**
1509  * Configure the RX function to use.
1510  *
1511  * @param priv
1512  *   Pointer to private structure.
1513  */
1514 void
1515 priv_select_rx_function(struct priv *priv)
1516 {
1517         priv->dev->rx_pkt_burst = mlx5_rx_burst;
1518 }