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 #define _GNU_SOURCE
  35
  36 #include <stddef.h>
  37 #include <assert.h>
  38 #include <unistd.h>
  39 #include <stdint.h>
  40 #include <stdio.h>
  41 #include <string.h>
  42 #include <stdlib.h>
  43 #include <errno.h>
  44 #include <dirent.h>
  45 #include <net/if.h>
  46 #include <sys/ioctl.h>
  47 #include <sys/socket.h>
  48 #include <sys/utsname.h>
  49 #include <netinet/in.h>
  50 #include <linux/ethtool.h>
  51 #include <linux/sockios.h>
  52 #include <linux/version.h>
  53 #include <fcntl.h>
  54 #include <stdalign.h>
  55 #include <sys/un.h>
  56
  57 #include <rte_atomic.h>
  58 #include <rte_ethdev.h>
  59 #include <rte_bus_pci.h>
  60 #include <rte_mbuf.h>
  61 #include <rte_common.h>
  62 #include <rte_interrupts.h>
  63 #include <rte_alarm.h>
  64 #include <rte_malloc.h>
  65
  66 #include "mlx5.h"
  67 #include "mlx5_rxtx.h"
  68 #include "mlx5_utils.h"
  69
  70 /* Add defines in case the running kernel is not the same as user headers. */
  71 #ifndef ETHTOOL_GLINKSETTINGS
  72 struct ethtool_link_settings {
  73         uint32_t cmd;
  74         uint32_t speed;
  75         uint8_t duplex;
  76         uint8_t port;
  77         uint8_t phy_address;
  78         uint8_t autoneg;
  79         uint8_t mdio_support;
  80         uint8_t eth_to_mdix;
  81         uint8_t eth_tp_mdix_ctrl;
  82         int8_t link_mode_masks_nwords;
  83         uint32_t reserved[8];
  84         uint32_t link_mode_masks[];
  85 };
  86
  87 #define ETHTOOL_GLINKSETTINGS 0x0000004c
  88 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
  89 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
  90 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
  91 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
  92 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
  93 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
  94 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
  95 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
  96 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
  97 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
  98 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
  99 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
 100 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
 101 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
 102 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
 103 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
 104 #endif
 105 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
 106 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
 107 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
 108 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
 109 #endif
 110 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
 111 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
 112 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
 113 #endif
 114 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
 115 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
 116 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
 117 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
 118 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
 119 #endif
 120
 121 /**
 122  * Return private structure associated with an Ethernet device.
 123  *
 124  * @param dev
 125  *   Pointer to Ethernet device structure.
 126  *
 127  * @return
 128  *   Pointer to private structure.
 129  */
 130 struct priv *
 131 mlx5_get_priv(struct rte_eth_dev *dev)
 132 {
 133         return dev->data->dev_private;
 134 }
 135
 136 /**
 137  * Check if running as a secondary process.
 138  *
 139  * @return
 140  *   Nonzero if running as a secondary process.
 141  */
 142 inline int
 143 mlx5_is_secondary(void)
 144 {
 145         return rte_eal_process_type() == RTE_PROC_SECONDARY;
 146 }
 147
 148 /**
 149  * Get interface name from private structure.
 150  *
 151  * @param[in] priv
 152  *   Pointer to private structure.
 153  * @param[out] ifname
 154  *   Interface name output buffer.
 155  *
 156  * @return
 157  *   0 on success, -1 on failure and errno is set.
 158  */
 159 int
 160 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
 161 {
 162         DIR *dir;
 163         struct dirent *dent;
 164         unsigned int dev_type = 0;
 165         unsigned int dev_port_prev = ~0u;
 166         char match[IF_NAMESIZE] = "";
 167
 168         {
 169                 MKSTR(path, "%s/device/net", priv->ibdev_path);
 170
 171                 dir = opendir(path);
 172                 if (dir == NULL)
 173                         return -1;
 174         }
 175         while ((dent = readdir(dir)) != NULL) {
 176                 char *name = dent->d_name;
 177                 FILE *file;
 178                 unsigned int dev_port;
 179                 int r;
 180
 181                 if ((name[0] == '.') &&
 182                     ((name[1] == '\0') ||
 183                      ((name[1] == '.') && (name[2] == '\0'))))
 184                         continue;
 185
 186                 MKSTR(path, "%s/device/net/%s/%s",
 187                       priv->ibdev_path, name,
 188                       (dev_type ? "dev_id" : "dev_port"));
 189
 190                 file = fopen(path, "rb");
 191                 if (file == NULL) {
 192                         if (errno != ENOENT)
 193                                 continue;
 194                         /*
 195                          * Switch to dev_id when dev_port does not exist as
 196                          * is the case with Linux kernel versions < 3.15.
 197                          */
 198 try_dev_id:
 199                         match[0] = '\0';
 200                         if (dev_type)
 201                                 break;
 202                         dev_type = 1;
 203                         dev_port_prev = ~0u;
 204                         rewinddir(dir);
 205                         continue;
 206                 }
 207                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 208                 fclose(file);
 209                 if (r != 1)
 210                         continue;
 211                 /*
 212                  * Switch to dev_id when dev_port returns the same value for
 213                  * all ports. May happen when using a MOFED release older than
 214                  * 3.0 with a Linux kernel >= 3.15.
 215                  */
 216                 if (dev_port == dev_port_prev)
 217                         goto try_dev_id;
 218                 dev_port_prev = dev_port;
 219                 if (dev_port == (priv->port - 1u))
 220                         snprintf(match, sizeof(match), "%s", name);
 221         }
 222         closedir(dir);
 223         if (match[0] == '\0')
 224                 return -1;
 225         strncpy(*ifname, match, sizeof(*ifname));
 226         return 0;
 227 }
 228
 229 /**
 230  * Check if the counter is located on ib counters file.
 231  *
 232  * @param[in] cntr
 233  *   Counter name.
 234  *
 235  * @return
 236  *   1 if counter is located on ib counters file , 0 otherwise.
 237  */
 238 int
 239 priv_is_ib_cntr(const char *cntr)
 240 {
 241         if (!strcmp(cntr, "out_of_buffer"))
 242                 return 1;
 243         return 0;
 244 }
 245
 246 /**
 247  * Read from sysfs entry.
 248  *
 249  * @param[in] priv
 250  *   Pointer to private structure.
 251  * @param[in] entry
 252  *   Entry name relative to sysfs path.
 253  * @param[out] buf
 254  *   Data output buffer.
 255  * @param size
 256  *   Buffer size.
 257  *
 258  * @return
 259  *   0 on success, -1 on failure and errno is set.
 260  */
 261 static int
 262 priv_sysfs_read(const struct priv *priv, const char *entry,
 263                 char *buf, size_t size)
 264 {
 265         char ifname[IF_NAMESIZE];
 266         FILE *file;
 267         int ret;
 268         int err;
 269
 270         if (priv_get_ifname(priv, &ifname))
 271                 return -1;
 272
 273         if (priv_is_ib_cntr(entry)) {
 274                 MKSTR(path, "%s/ports/1/hw_counters/%s",
 275                       priv->ibdev_path, entry);
 276                 file = fopen(path, "rb");
 277         } else {
 278                 MKSTR(path, "%s/device/net/%s/%s",
 279                       priv->ibdev_path, ifname, entry);
 280                 file = fopen(path, "rb");
 281         }
 282         if (file == NULL)
 283                 return -1;
 284         ret = fread(buf, 1, size, file);
 285         err = errno;
 286         if (((size_t)ret < size) && (ferror(file)))
 287                 ret = -1;
 288         else
 289                 ret = size;
 290         fclose(file);
 291         errno = err;
 292         return ret;
 293 }
 294
 295 /**
 296  * Write to sysfs entry.
 297  *
 298  * @param[in] priv
 299  *   Pointer to private structure.
 300  * @param[in] entry
 301  *   Entry name relative to sysfs path.
 302  * @param[in] buf
 303  *   Data buffer.
 304  * @param size
 305  *   Buffer size.
 306  *
 307  * @return
 308  *   0 on success, -1 on failure and errno is set.
 309  */
 310 static int
 311 priv_sysfs_write(const struct priv *priv, const char *entry,
 312                  char *buf, size_t size)
 313 {
 314         char ifname[IF_NAMESIZE];
 315         FILE *file;
 316         int ret;
 317         int err;
 318
 319         if (priv_get_ifname(priv, &ifname))
 320                 return -1;
 321
 322         MKSTR(path, "%s/device/net/%s/%s", priv->ibdev_path, ifname, entry);
 323
 324         file = fopen(path, "wb");
 325         if (file == NULL)
 326                 return -1;
 327         ret = fwrite(buf, 1, size, file);
 328         err = errno;
 329         if (((size_t)ret < size) || (ferror(file)))
 330                 ret = -1;
 331         else
 332                 ret = size;
 333         fclose(file);
 334         errno = err;
 335         return ret;
 336 }
 337
 338 /**
 339  * Get unsigned long sysfs property.
 340  *
 341  * @param priv
 342  *   Pointer to private structure.
 343  * @param[in] name
 344  *   Entry name relative to sysfs path.
 345  * @param[out] value
 346  *   Value output buffer.
 347  *
 348  * @return
 349  *   0 on success, -1 on failure and errno is set.
 350  */
 351 static int
 352 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
 353 {
 354         int ret;
 355         unsigned long value_ret;
 356         char value_str[32];
 357
 358         ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
 359         if (ret == -1) {
 360                 DEBUG("cannot read %s value from sysfs: %s",
 361                       name, strerror(errno));
 362                 return -1;
 363         }
 364         value_str[ret] = '\0';
 365         errno = 0;
 366         value_ret = strtoul(value_str, NULL, 0);
 367         if (errno) {
 368                 DEBUG("invalid %s value `%s': %s", name, value_str,
 369                       strerror(errno));
 370                 return -1;
 371         }
 372         *value = value_ret;
 373         return 0;
 374 }
 375
 376 /**
 377  * Set unsigned long sysfs property.
 378  *
 379  * @param priv
 380  *   Pointer to private structure.
 381  * @param[in] name
 382  *   Entry name relative to sysfs path.
 383  * @param value
 384  *   Value to set.
 385  *
 386  * @return
 387  *   0 on success, -1 on failure and errno is set.
 388  */
 389 static int
 390 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
 391 {
 392         int ret;
 393         MKSTR(value_str, "%lu", value);
 394
 395         ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
 396         if (ret == -1) {
 397                 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
 398                       name, value_str, value, strerror(errno));
 399                 return -1;
 400         }
 401         return 0;
 402 }
 403
 404 /**
 405  * Perform ifreq ioctl() on associated Ethernet device.
 406  *
 407  * @param[in] priv
 408  *   Pointer to private structure.
 409  * @param req
 410  *   Request number to pass to ioctl().
 411  * @param[out] ifr
 412  *   Interface request structure output buffer.
 413  *
 414  * @return
 415  *   0 on success, -1 on failure and errno is set.
 416  */
 417 int
 418 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
 419 {
 420         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 421         int ret = -1;
 422
 423         if (sock == -1)
 424                 return ret;
 425         if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
 426                 ret = ioctl(sock, req, ifr);
 427         close(sock);
 428         return ret;
 429 }
 430
 431 /**
 432  * Return the number of active VFs for the current device.
 433  *
 434  * @param[in] priv
 435  *   Pointer to private structure.
 436  * @param[out] num_vfs
 437  *   Number of active VFs.
 438  *
 439  * @return
 440  *   0 on success, -1 on failure and errno is set.
 441  */
 442 int
 443 priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs)
 444 {
 445         /* The sysfs entry name depends on the operating system. */
 446         const char **name = (const char *[]){
 447                 "device/sriov_numvfs",
 448                 "device/mlx5_num_vfs",
 449                 NULL,
 450         };
 451         int ret;
 452
 453         do {
 454                 unsigned long ulong_num_vfs;
 455
 456                 ret = priv_get_sysfs_ulong(priv, *name, &ulong_num_vfs);
 457                 if (!ret)
 458                         *num_vfs = ulong_num_vfs;
 459         } while (*(++name) && ret);
 460         return ret;
 461 }
 462
 463 /**
 464  * Get device MTU.
 465  *
 466  * @param priv
 467  *   Pointer to private structure.
 468  * @param[out] mtu
 469  *   MTU value output buffer.
 470  *
 471  * @return
 472  *   0 on success, -1 on failure and errno is set.
 473  */
 474 int
 475 priv_get_mtu(struct priv *priv, uint16_t *mtu)
 476 {
 477         unsigned long ulong_mtu;
 478
 479         if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
 480                 return -1;
 481         *mtu = ulong_mtu;
 482         return 0;
 483 }
 484
 485 /**
 486  * Read device counter from sysfs.
 487  *
 488  * @param priv
 489  *   Pointer to private structure.
 490  * @param name
 491  *   Counter name.
 492  * @param[out] cntr
 493  *   Counter output buffer.
 494  *
 495  * @return
 496  *   0 on success, -1 on failure and errno is set.
 497  */
 498 int
 499 priv_get_cntr_sysfs(struct priv *priv, const char *name, uint64_t *cntr)
 500 {
 501         unsigned long ulong_ctr;
 502
 503         if (priv_get_sysfs_ulong(priv, name, &ulong_ctr) == -1)
 504                 return -1;
 505         *cntr = ulong_ctr;
 506         return 0;
 507 }
 508
 509 /**
 510  * Set device MTU.
 511  *
 512  * @param priv
 513  *   Pointer to private structure.
 514  * @param mtu
 515  *   MTU value to set.
 516  *
 517  * @return
 518  *   0 on success, -1 on failure and errno is set.
 519  */
 520 static int
 521 priv_set_mtu(struct priv *priv, uint16_t mtu)
 522 {
 523         uint16_t new_mtu;
 524
 525         if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
 526             priv_get_mtu(priv, &new_mtu))
 527                 return -1;
 528         if (new_mtu == mtu)
 529                 return 0;
 530         errno = EINVAL;
 531         return -1;
 532 }
 533
 534 /**
 535  * Set device flags.
 536  *
 537  * @param priv
 538  *   Pointer to private structure.
 539  * @param keep
 540  *   Bitmask for flags that must remain untouched.
 541  * @param flags
 542  *   Bitmask for flags to modify.
 543  *
 544  * @return
 545  *   0 on success, -1 on failure and errno is set.
 546  */
 547 int
 548 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
 549 {
 550         unsigned long tmp;
 551
 552         if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
 553                 return -1;
 554         tmp &= keep;
 555         tmp |= (flags & (~keep));
 556         return priv_set_sysfs_ulong(priv, "flags", tmp);
 557 }
 558
 559 /**
 560  * Ethernet device configuration.
 561  *
 562  * Prepare the driver for a given number of TX and RX queues.
 563  *
 564  * @param dev
 565  *   Pointer to Ethernet device structure.
 566  *
 567  * @return
 568  *   0 on success, errno value on failure.
 569  */
 570 static int
 571 dev_configure(struct rte_eth_dev *dev)
 572 {
 573         struct priv *priv = dev->data->dev_private;
 574         unsigned int rxqs_n = dev->data->nb_rx_queues;
 575         unsigned int txqs_n = dev->data->nb_tx_queues;
 576         unsigned int i;
 577         unsigned int j;
 578         unsigned int reta_idx_n;
 579         const uint8_t use_app_rss_key =
 580                 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len;
 581
 582         if (use_app_rss_key &&
 583             (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
 584              rss_hash_default_key_len)) {
 585                 /* MLX5 RSS only support 40bytes key. */
 586                 return EINVAL;
 587         }
 588         priv->rss_conf.rss_key =
 589                 rte_realloc(priv->rss_conf.rss_key,
 590                             rss_hash_default_key_len, 0);
 591         if (!priv->rss_conf.rss_key) {
 592                 ERROR("cannot allocate RSS hash key memory (%u)", rxqs_n);
 593                 return ENOMEM;
 594         }
 595         memcpy(priv->rss_conf.rss_key,
 596                use_app_rss_key ?
 597                dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
 598                rss_hash_default_key,
 599                rss_hash_default_key_len);
 600         priv->rss_conf.rss_key_len = rss_hash_default_key_len;
 601         priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
 602         priv->rxqs = (void *)dev->data->rx_queues;
 603         priv->txqs = (void *)dev->data->tx_queues;
 604         if (txqs_n != priv->txqs_n) {
 605                 INFO("%p: TX queues number update: %u -> %u",
 606                      (void *)dev, priv->txqs_n, txqs_n);
 607                 priv->txqs_n = txqs_n;
 608         }
 609         if (rxqs_n > priv->ind_table_max_size) {
 610                 ERROR("cannot handle this many RX queues (%u)", rxqs_n);
 611                 return EINVAL;
 612         }
 613         if (rxqs_n == priv->rxqs_n)
 614                 return 0;
 615         INFO("%p: RX queues number update: %u -> %u",
 616              (void *)dev, priv->rxqs_n, rxqs_n);
 617         priv->rxqs_n = rxqs_n;
 618         /* If the requested number of RX queues is not a power of two, use the
 619          * maximum indirection table size for better balancing.
 620          * The result is always rounded to the next power of two. */
 621         reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
 622                                      priv->ind_table_max_size :
 623                                      rxqs_n));
 624         if (priv_rss_reta_index_resize(priv, reta_idx_n))
 625                 return ENOMEM;
 626         /* When the number of RX queues is not a power of two, the remaining
 627          * table entries are padded with reused WQs and hashes are not spread
 628          * uniformly. */
 629         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
 630                 (*priv->reta_idx)[i] = j;
 631                 if (++j == rxqs_n)
 632                         j = 0;
 633         }
 634         return 0;
 635 }
 636
 637 /**
 638  * DPDK callback for Ethernet device configuration.
 639  *
 640  * @param dev
 641  *   Pointer to Ethernet device structure.
 642  *
 643  * @return
 644  *   0 on success, negative errno value on failure.
 645  */
 646 int
 647 mlx5_dev_configure(struct rte_eth_dev *dev)
 648 {
 649         struct priv *priv = dev->data->dev_private;
 650         int ret;
 651
 652         if (mlx5_is_secondary())
 653                 return -E_RTE_SECONDARY;
 654
 655         priv_lock(priv);
 656         ret = dev_configure(dev);
 657         assert(ret >= 0);
 658         priv_unlock(priv);
 659         return -ret;
 660 }
 661
 662 /**
 663  * DPDK callback to get information about the device.
 664  *
 665  * @param dev
 666  *   Pointer to Ethernet device structure.
 667  * @param[out] info
 668  *   Info structure output buffer.
 669  */
 670 void
 671 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 672 {
 673         struct priv *priv = mlx5_get_priv(dev);
 674         unsigned int max;
 675         char ifname[IF_NAMESIZE];
 676
 677         info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
 678
 679         priv_lock(priv);
 680         /* FIXME: we should ask the device for these values. */
 681         info->min_rx_bufsize = 32;
 682         info->max_rx_pktlen = 65536;
 683         /*
 684          * Since we need one CQ per QP, the limit is the minimum number
 685          * between the two values.
 686          */
 687         max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
 688                       priv->device_attr.orig_attr.max_qp);
 689         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
 690         if (max >= 65535)
 691                 max = 65535;
 692         info->max_rx_queues = max;
 693         info->max_tx_queues = max;
 694         info->max_mac_addrs = RTE_DIM(priv->mac);
 695         info->rx_offload_capa =
 696                 (priv->hw_csum ?
 697                  (DEV_RX_OFFLOAD_IPV4_CKSUM |
 698                   DEV_RX_OFFLOAD_UDP_CKSUM |
 699                   DEV_RX_OFFLOAD_TCP_CKSUM) :
 700                  0) |
 701                 (priv->hw_vlan_strip ? DEV_RX_OFFLOAD_VLAN_STRIP : 0) |
 702                 DEV_RX_OFFLOAD_TIMESTAMP;
 703
 704         if (!priv->mps)
 705                 info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
 706         if (priv->hw_csum)
 707                 info->tx_offload_capa |=
 708                         (DEV_TX_OFFLOAD_IPV4_CKSUM |
 709                          DEV_TX_OFFLOAD_UDP_CKSUM |
 710                          DEV_TX_OFFLOAD_TCP_CKSUM);
 711         if (priv->tso)
 712                 info->tx_offload_capa |= DEV_TX_OFFLOAD_TCP_TSO;
 713         if (priv->tunnel_en)
 714                 info->tx_offload_capa |= (DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM |
 715                                           DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
 716                                           DEV_TX_OFFLOAD_GRE_TNL_TSO);
 717         if (priv_get_ifname(priv, &ifname) == 0)
 718                 info->if_index = if_nametoindex(ifname);
 719         info->reta_size = priv->reta_idx_n ?
 720                 priv->reta_idx_n : priv->ind_table_max_size;
 721         info->hash_key_size = priv->rss_conf.rss_key_len;
 722         info->speed_capa = priv->link_speed_capa;
 723         priv_unlock(priv);
 724 }
 725
 726 const uint32_t *
 727 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 728 {
 729         static const uint32_t ptypes[] = {
 730                 /* refers to rxq_cq_to_pkt_type() */
 731                 RTE_PTYPE_L2_ETHER,
 732                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 733                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 734                 RTE_PTYPE_L4_NONFRAG,
 735                 RTE_PTYPE_L4_FRAG,
 736                 RTE_PTYPE_L4_TCP,
 737                 RTE_PTYPE_L4_UDP,
 738                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 739                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 740                 RTE_PTYPE_INNER_L4_NONFRAG,
 741                 RTE_PTYPE_INNER_L4_FRAG,
 742                 RTE_PTYPE_INNER_L4_TCP,
 743                 RTE_PTYPE_INNER_L4_UDP,
 744                 RTE_PTYPE_UNKNOWN
 745         };
 746
 747         if (dev->rx_pkt_burst == mlx5_rx_burst ||
 748             dev->rx_pkt_burst == mlx5_rx_burst_vec)
 749                 return ptypes;
 750         return NULL;
 751 }
 752
 753 /**
 754  * DPDK callback to retrieve physical link information.
 755  *
 756  * @param dev
 757  *   Pointer to Ethernet device structure.
 758  * @param wait_to_complete
 759  *   Wait for request completion (ignored).
 760  */
 761 static int
 762 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
 763 {
 764         struct priv *priv = mlx5_get_priv(dev);
 765         struct ethtool_cmd edata = {
 766                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 767         };
 768         struct ifreq ifr;
 769         struct rte_eth_link dev_link;
 770         int link_speed = 0;
 771
 772         /* priv_lock() is not taken to allow concurrent calls. */
 773
 774         (void)wait_to_complete;
 775         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
 776                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
 777                 return -1;
 778         }
 779         memset(&dev_link, 0, sizeof(dev_link));
 780         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 781                                 (ifr.ifr_flags & IFF_RUNNING));
 782         ifr.ifr_data = (void *)&edata;
 783         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 784                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
 785                      strerror(errno));
 786                 return -1;
 787         }
 788         link_speed = ethtool_cmd_speed(&edata);
 789         if (link_speed == -1)
 790                 dev_link.link_speed = 0;
 791         else
 792                 dev_link.link_speed = link_speed;
 793         priv->link_speed_capa = 0;
 794         if (edata.supported & SUPPORTED_Autoneg)
 795                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 796         if (edata.supported & (SUPPORTED_1000baseT_Full |
 797                                SUPPORTED_1000baseKX_Full))
 798                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 799         if (edata.supported & SUPPORTED_10000baseKR_Full)
 800                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 801         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 802                                SUPPORTED_40000baseCR4_Full |
 803                                SUPPORTED_40000baseSR4_Full |
 804                                SUPPORTED_40000baseLR4_Full))
 805                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 806         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 807                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 808         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 809                         ETH_LINK_SPEED_FIXED);
 810         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 811                 /* Link status changed. */
 812                 dev->data->dev_link = dev_link;
 813                 return 0;
 814         }
 815         /* Link status is still the same. */
 816         return -1;
 817 }
 818
 819 /**
 820  * Retrieve physical link information (unlocked version using new ioctl).
 821  *
 822  * @param dev
 823  *   Pointer to Ethernet device structure.
 824  * @param wait_to_complete
 825  *   Wait for request completion (ignored).
 826  */
 827 static int
 828 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
 829 {
 830         struct priv *priv = mlx5_get_priv(dev);
 831         struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
 832         struct ifreq ifr;
 833         struct rte_eth_link dev_link;
 834         uint64_t sc;
 835
 836         (void)wait_to_complete;
 837         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
 838                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
 839                 return -1;
 840         }
 841         memset(&dev_link, 0, sizeof(dev_link));
 842         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 843                                 (ifr.ifr_flags & IFF_RUNNING));
 844         ifr.ifr_data = (void *)&gcmd;
 845         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 846                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
 847                       strerror(errno));
 848                 return -1;
 849         }
 850         gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
 851
 852         alignas(struct ethtool_link_settings)
 853         uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
 854                      sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
 855         struct ethtool_link_settings *ecmd = (void *)data;
 856
 857         *ecmd = gcmd;
 858         ifr.ifr_data = (void *)ecmd;
 859         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 860                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
 861                       strerror(errno));
 862                 return -1;
 863         }
 864         dev_link.link_speed = ecmd->speed;
 865         sc = ecmd->link_mode_masks[0] |
 866                 ((uint64_t)ecmd->link_mode_masks[1] << 32);
 867         priv->link_speed_capa = 0;
 868         if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
 869                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 870         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
 871                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
 872                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 873         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
 874                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
 875                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
 876                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 877         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
 878                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
 879                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
 880         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
 881                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
 882                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
 883                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
 884                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 885         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
 886                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
 887                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
 888                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
 889                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
 890         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
 891                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
 892                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
 893                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
 894         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
 895                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
 896                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
 897         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
 898                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
 899                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
 900                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
 901                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
 902         dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
 903                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 904         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 905                                   ETH_LINK_SPEED_FIXED);
 906         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 907                 /* Link status changed. */
 908                 dev->data->dev_link = dev_link;
 909                 return 0;
 910         }
 911         /* Link status is still the same. */
 912         return -1;
 913 }
 914
 915 /**
 916  * DPDK callback to retrieve physical link information.
 917  *
 918  * @param dev
 919  *   Pointer to Ethernet device structure.
 920  * @param wait_to_complete
 921  *   Wait for request completion (ignored).
 922  */
 923 int
 924 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
 925 {
 926         struct utsname utsname;
 927         int ver[3];
 928
 929         if (uname(&utsname) == -1 ||
 930             sscanf(utsname.release, "%d.%d.%d",
 931                    &ver[0], &ver[1], &ver[2]) != 3 ||
 932             KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
 933                 return mlx5_link_update_unlocked_gset(dev, wait_to_complete);
 934         return mlx5_link_update_unlocked_gs(dev, wait_to_complete);
 935 }
 936
 937 /**
 938  * DPDK callback to change the MTU.
 939  *
 940  * @param dev
 941  *   Pointer to Ethernet device structure.
 942  * @param in_mtu
 943  *   New MTU.
 944  *
 945  * @return
 946  *   0 on success, negative errno value on failure.
 947  */
 948 int
 949 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 950 {
 951         struct priv *priv = dev->data->dev_private;
 952         uint16_t kern_mtu;
 953         int ret = 0;
 954
 955         if (mlx5_is_secondary())
 956                 return -E_RTE_SECONDARY;
 957
 958         priv_lock(priv);
 959         ret = priv_get_mtu(priv, &kern_mtu);
 960         if (ret)
 961                 goto out;
 962         /* Set kernel interface MTU first. */
 963         ret = priv_set_mtu(priv, mtu);
 964         if (ret)
 965                 goto out;
 966         ret = priv_get_mtu(priv, &kern_mtu);
 967         if (ret)
 968                 goto out;
 969         if (kern_mtu == mtu) {
 970                 priv->mtu = mtu;
 971                 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
 972         }
 973         priv_unlock(priv);
 974         return 0;
 975 out:
 976         ret = errno;
 977         WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
 978              strerror(ret));
 979         priv_unlock(priv);
 980         assert(ret >= 0);
 981         return -ret;
 982 }
 983
 984 /**
 985  * DPDK callback to get flow control status.
 986  *
 987  * @param dev
 988  *   Pointer to Ethernet device structure.
 989  * @param[out] fc_conf
 990  *   Flow control output buffer.
 991  *
 992  * @return
 993  *   0 on success, negative errno value on failure.
 994  */
 995 int
 996 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 997 {
 998         struct priv *priv = dev->data->dev_private;
 999         struct ifreq ifr;
1000         struct ethtool_pauseparam ethpause = {
1001                 .cmd = ETHTOOL_GPAUSEPARAM
1002         };
1003         int ret;
1004
1005         if (mlx5_is_secondary())
1006                 return -E_RTE_SECONDARY;
1007
1008         ifr.ifr_data = (void *)&ethpause;
1009         priv_lock(priv);
1010         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1011                 ret = errno;
1012                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
1013                      " failed: %s",
1014                      strerror(ret));
1015                 goto out;
1016         }
1017
1018         fc_conf->autoneg = ethpause.autoneg;
1019         if (ethpause.rx_pause && ethpause.tx_pause)
1020                 fc_conf->mode = RTE_FC_FULL;
1021         else if (ethpause.rx_pause)
1022                 fc_conf->mode = RTE_FC_RX_PAUSE;
1023         else if (ethpause.tx_pause)
1024                 fc_conf->mode = RTE_FC_TX_PAUSE;
1025         else
1026                 fc_conf->mode = RTE_FC_NONE;
1027         ret = 0;
1028
1029 out:
1030         priv_unlock(priv);
1031         assert(ret >= 0);
1032         return -ret;
1033 }
1034
1035 /**
1036  * DPDK callback to modify flow control parameters.
1037  *
1038  * @param dev
1039  *   Pointer to Ethernet device structure.
1040  * @param[in] fc_conf
1041  *   Flow control parameters.
1042  *
1043  * @return
1044  *   0 on success, negative errno value on failure.
1045  */
1046 int
1047 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1048 {
1049         struct priv *priv = dev->data->dev_private;
1050         struct ifreq ifr;
1051         struct ethtool_pauseparam ethpause = {
1052                 .cmd = ETHTOOL_SPAUSEPARAM
1053         };
1054         int ret;
1055
1056         if (mlx5_is_secondary())
1057                 return -E_RTE_SECONDARY;
1058
1059         ifr.ifr_data = (void *)&ethpause;
1060         ethpause.autoneg = fc_conf->autoneg;
1061         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1062             (fc_conf->mode & RTE_FC_RX_PAUSE))
1063                 ethpause.rx_pause = 1;
1064         else
1065                 ethpause.rx_pause = 0;
1066
1067         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1068             (fc_conf->mode & RTE_FC_TX_PAUSE))
1069                 ethpause.tx_pause = 1;
1070         else
1071                 ethpause.tx_pause = 0;
1072
1073         priv_lock(priv);
1074         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1075                 ret = errno;
1076                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1077                      " failed: %s",
1078                      strerror(ret));
1079                 goto out;
1080         }
1081         ret = 0;
1082
1083 out:
1084         priv_unlock(priv);
1085         assert(ret >= 0);
1086         return -ret;
1087 }
1088
1089 /**
1090  * Get PCI information from struct ibv_device.
1091  *
1092  * @param device
1093  *   Pointer to Ethernet device structure.
1094  * @param[out] pci_addr
1095  *   PCI bus address output buffer.
1096  *
1097  * @return
1098  *   0 on success, -1 on failure and errno is set.
1099  */
1100 int
1101 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1102                             struct rte_pci_addr *pci_addr)
1103 {
1104         FILE *file;
1105         char line[32];
1106         MKSTR(path, "%s/device/uevent", device->ibdev_path);
1107
1108         file = fopen(path, "rb");
1109         if (file == NULL)
1110                 return -1;
1111         while (fgets(line, sizeof(line), file) == line) {
1112                 size_t len = strlen(line);
1113                 int ret;
1114
1115                 /* Truncate long lines. */
1116                 if (len == (sizeof(line) - 1))
1117                         while (line[(len - 1)] != '\n') {
1118                                 ret = fgetc(file);
1119                                 if (ret == EOF)
1120                                         break;
1121                                 line[(len - 1)] = ret;
1122                         }
1123                 /* Extract information. */
1124                 if (sscanf(line,
1125                            "PCI_SLOT_NAME="
1126                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1127                            &pci_addr->domain,
1128                            &pci_addr->bus,
1129                            &pci_addr->devid,
1130                            &pci_addr->function) == 4) {
1131                         ret = 0;
1132                         break;
1133                 }
1134         }
1135         fclose(file);
1136         return 0;
1137 }
1138
1139 /**
1140  * Update the link status.
1141  *
1142  * @param priv
1143  *   Pointer to private structure.
1144  *
1145  * @return
1146  *   Zero if the callback process can be called immediately.
1147  */
1148 static int
1149 priv_link_status_update(struct priv *priv)
1150 {
1151         struct rte_eth_link *link = &priv->dev->data->dev_link;
1152
1153         mlx5_link_update(priv->dev, 0);
1154         if (((link->link_speed == 0) && link->link_status) ||
1155                 ((link->link_speed != 0) && !link->link_status)) {
1156                 /*
1157                  * Inconsistent status. Event likely occurred before the
1158                  * kernel netdevice exposes the new status.
1159                  */
1160                 if (!priv->pending_alarm) {
1161                         priv->pending_alarm = 1;
1162                         rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1163                                           mlx5_dev_link_status_handler,
1164                                           priv->dev);
1165                 }
1166                 return 1;
1167         } else if (unlikely(priv->pending_alarm)) {
1168                 /* Link interrupt occurred while alarm is already scheduled. */
1169                 priv->pending_alarm = 0;
1170                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
1171         }
1172         return 0;
1173 }
1174
1175 /**
1176  * Device status handler.
1177  *
1178  * @param priv
1179  *   Pointer to private structure.
1180  * @param events
1181  *   Pointer to event flags holder.
1182  *
1183  * @return
1184  *   Events bitmap of callback process which can be called immediately.
1185  */
1186 static uint32_t
1187 priv_dev_status_handler(struct priv *priv)
1188 {
1189         struct ibv_async_event event;
1190         uint32_t ret = 0;
1191
1192         /* Read all message and acknowledge them. */
1193         for (;;) {
1194                 if (ibv_get_async_event(priv->ctx, &event))
1195                         break;
1196                 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
1197                         event.event_type == IBV_EVENT_PORT_ERR) &&
1198                         (priv->dev->data->dev_conf.intr_conf.lsc == 1))
1199                         ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
1200                 else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
1201                         priv->dev->data->dev_conf.intr_conf.rmv == 1)
1202                         ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
1203                 else
1204                         DEBUG("event type %d on port %d not handled",
1205                               event.event_type, event.element.port_num);
1206                 ibv_ack_async_event(&event);
1207         }
1208         if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
1209                 if (priv_link_status_update(priv))
1210                         ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
1211         return ret;
1212 }
1213
1214 /**
1215  * Handle delayed link status event.
1216  *
1217  * @param arg
1218  *   Registered argument.
1219  */
1220 void
1221 mlx5_dev_link_status_handler(void *arg)
1222 {
1223         struct rte_eth_dev *dev = arg;
1224         struct priv *priv = dev->data->dev_private;
1225         int ret;
1226
1227         priv_lock(priv);
1228         assert(priv->pending_alarm == 1);
1229         priv->pending_alarm = 0;
1230         ret = priv_link_status_update(priv);
1231         priv_unlock(priv);
1232         if (!ret)
1233                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL,
1234                                               NULL);
1235 }
1236
1237 /**
1238  * Handle interrupts from the NIC.
1239  *
1240  * @param[in] intr_handle
1241  *   Interrupt handler.
1242  * @param cb_arg
1243  *   Callback argument.
1244  */
1245 void
1246 mlx5_dev_interrupt_handler(void *cb_arg)
1247 {
1248         struct rte_eth_dev *dev = cb_arg;
1249         struct priv *priv = dev->data->dev_private;
1250         uint32_t events;
1251
1252         priv_lock(priv);
1253         events = priv_dev_status_handler(priv);
1254         priv_unlock(priv);
1255         if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1256                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL,
1257                                               NULL);
1258         if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1259                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL,
1260                                               NULL);
1261 }
1262
1263 /**
1264  * Handle interrupts from the socket.
1265  *
1266  * @param cb_arg
1267  *   Callback argument.
1268  */
1269 static void
1270 mlx5_dev_handler_socket(void *cb_arg)
1271 {
1272         struct rte_eth_dev *dev = cb_arg;
1273         struct priv *priv = dev->data->dev_private;
1274
1275         priv_lock(priv);
1276         priv_socket_handle(priv);
1277         priv_unlock(priv);
1278 }
1279
1280 /**
1281  * Uninstall interrupt handler.
1282  *
1283  * @param priv
1284  *   Pointer to private structure.
1285  * @param dev
1286  *   Pointer to the rte_eth_dev structure.
1287  */
1288 void
1289 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1290 {
1291         if (dev->data->dev_conf.intr_conf.lsc ||
1292             dev->data->dev_conf.intr_conf.rmv)
1293                 rte_intr_callback_unregister(&priv->intr_handle,
1294                                              mlx5_dev_interrupt_handler, dev);
1295         if (priv->primary_socket)
1296                 rte_intr_callback_unregister(&priv->intr_handle_socket,
1297                                              mlx5_dev_handler_socket, dev);
1298         if (priv->pending_alarm)
1299                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1300         priv->pending_alarm = 0;
1301         priv->intr_handle.fd = 0;
1302         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1303         priv->intr_handle_socket.fd = 0;
1304         priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1305 }
1306
1307 /**
1308  * Install interrupt handler.
1309  *
1310  * @param priv
1311  *   Pointer to private structure.
1312  * @param dev
1313  *   Pointer to the rte_eth_dev structure.
1314  */
1315 void
1316 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1317 {
1318         int rc, flags;
1319
1320         assert(!mlx5_is_secondary());
1321         assert(priv->ctx->async_fd > 0);
1322         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1323         rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1324         if (rc < 0) {
1325                 INFO("failed to change file descriptor async event queue");
1326                 dev->data->dev_conf.intr_conf.lsc = 0;
1327                 dev->data->dev_conf.intr_conf.rmv = 0;
1328         }
1329         if (dev->data->dev_conf.intr_conf.lsc ||
1330             dev->data->dev_conf.intr_conf.rmv) {
1331                 priv->intr_handle.fd = priv->ctx->async_fd;
1332                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1333                 rte_intr_callback_register(&priv->intr_handle,
1334                                            mlx5_dev_interrupt_handler, dev);
1335         }
1336
1337         rc = priv_socket_init(priv);
1338         if (!rc && priv->primary_socket) {
1339                 priv->intr_handle_socket.fd = priv->primary_socket;
1340                 priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1341                 rte_intr_callback_register(&priv->intr_handle_socket,
1342                                            mlx5_dev_handler_socket, dev);
1343         }
1344 }
1345
1346 /**
1347  * Change the link state (UP / DOWN).
1348  *
1349  * @param priv
1350  *   Pointer to private data structure.
1351  * @param dev
1352  *   Pointer to rte_eth_dev structure.
1353  * @param up
1354  *   Nonzero for link up, otherwise link down.
1355  *
1356  * @return
1357  *   0 on success, errno value on failure.
1358  */
1359 static int
1360 priv_dev_set_link(struct priv *priv, struct rte_eth_dev *dev, int up)
1361 {
1362         int err;
1363
1364         if (up) {
1365                 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
1366                 if (err)
1367                         return err;
1368                 priv_dev_select_tx_function(priv, dev);
1369                 priv_dev_select_rx_function(priv, dev);
1370         } else {
1371                 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
1372                 if (err)
1373                         return err;
1374                 dev->rx_pkt_burst = removed_rx_burst;
1375                 dev->tx_pkt_burst = removed_tx_burst;
1376         }
1377         return 0;
1378 }
1379
1380 /**
1381  * DPDK callback to bring the link DOWN.
1382  *
1383  * @param dev
1384  *   Pointer to Ethernet device structure.
1385  *
1386  * @return
1387  *   0 on success, errno value on failure.
1388  */
1389 int
1390 mlx5_set_link_down(struct rte_eth_dev *dev)
1391 {
1392         struct priv *priv = dev->data->dev_private;
1393         int err;
1394
1395         priv_lock(priv);
1396         err = priv_dev_set_link(priv, dev, 0);
1397         priv_unlock(priv);
1398         return err;
1399 }
1400
1401 /**
1402  * DPDK callback to bring the link UP.
1403  *
1404  * @param dev
1405  *   Pointer to Ethernet device structure.
1406  *
1407  * @return
1408  *   0 on success, errno value on failure.
1409  */
1410 int
1411 mlx5_set_link_up(struct rte_eth_dev *dev)
1412 {
1413         struct priv *priv = dev->data->dev_private;
1414         int err;
1415
1416         priv_lock(priv);
1417         err = priv_dev_set_link(priv, dev, 1);
1418         priv_unlock(priv);
1419         return err;
1420 }
1421
1422 /**
1423  * Configure the TX function to use.
1424  *
1425  * @param priv
1426  *   Pointer to private data structure.
1427  * @param dev
1428  *   Pointer to rte_eth_dev structure.
1429  */
1430 void
1431 priv_dev_select_tx_function(struct priv *priv, struct rte_eth_dev *dev)
1432 {
1433         assert(priv != NULL);
1434         assert(dev != NULL);
1435         dev->tx_pkt_burst = mlx5_tx_burst;
1436         /* Select appropriate TX function. */
1437         if (priv->mps == MLX5_MPW_ENHANCED) {
1438                 if (priv_check_vec_tx_support(priv) > 0) {
1439                         if (priv_check_raw_vec_tx_support(priv) > 0)
1440                                 dev->tx_pkt_burst = mlx5_tx_burst_raw_vec;
1441                         else
1442                                 dev->tx_pkt_burst = mlx5_tx_burst_vec;
1443                         DEBUG("selected Enhanced MPW TX vectorized function");
1444                 } else {
1445                         dev->tx_pkt_burst = mlx5_tx_burst_empw;
1446                         DEBUG("selected Enhanced MPW TX function");
1447                 }
1448         } else if (priv->mps && priv->txq_inline) {
1449                 dev->tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1450                 DEBUG("selected MPW inline TX function");
1451         } else if (priv->mps) {
1452                 dev->tx_pkt_burst = mlx5_tx_burst_mpw;
1453                 DEBUG("selected MPW TX function");
1454         }
1455 }
1456
1457 /**
1458  * Configure the RX function to use.
1459  *
1460  * @param priv
1461  *   Pointer to private data structure.
1462  * @param dev
1463  *   Pointer to rte_eth_dev structure.
1464  */
1465 void
1466 priv_dev_select_rx_function(struct priv *priv, struct rte_eth_dev *dev)
1467 {
1468         assert(priv != NULL);
1469         assert(dev != NULL);
1470         if (priv_check_vec_rx_support(priv) > 0) {
1471                 dev->rx_pkt_burst = mlx5_rx_burst_vec;
1472                 DEBUG("selected RX vectorized function");
1473         } else {
1474                 dev->rx_pkt_burst = mlx5_rx_burst;
1475         }
1476 }