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 <sys/utsname.h>
  47 #include <netinet/in.h>
  48 #include <linux/ethtool.h>
  49 #include <linux/sockios.h>
  50 #include <linux/version.h>
  51 #include <fcntl.h>
  52
  53 #include <rte_atomic.h>
  54 #include <rte_ethdev.h>
  55 #include <rte_mbuf.h>
  56 #include <rte_common.h>
  57 #include <rte_interrupts.h>
  58 #include <rte_alarm.h>
  59 #include <rte_malloc.h>
  60
  61 #include "mlx5.h"
  62 #include "mlx5_rxtx.h"
  63 #include "mlx5_utils.h"
  64
  65 /* Add defines in case the running kernel is not the same as user headers. */
  66 #ifndef ETHTOOL_GLINKSETTINGS
  67 struct ethtool_link_settings {
  68         uint32_t cmd;
  69         uint32_t speed;
  70         uint8_t duplex;
  71         uint8_t port;
  72         uint8_t phy_address;
  73         uint8_t autoneg;
  74         uint8_t mdio_support;
  75         uint8_t eth_to_mdix;
  76         uint8_t eth_tp_mdix_ctrl;
  77         int8_t link_mode_masks_nwords;
  78         uint32_t reserved[8];
  79         uint32_t link_mode_masks[];
  80 };
  81
  82 #define ETHTOOL_GLINKSETTINGS 0x0000004c
  83 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
  84 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
  85 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
  86 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
  87 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
  88 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
  89 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
  90 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
  91 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
  92 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
  93 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
  94 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
  95 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
  96 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
  97 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
  98 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
  99 #endif
 100 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
 101 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
 102 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
 103 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
 104 #endif
 105 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
 106 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
 107 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
 108 #endif
 109 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
 110 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
 111 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
 112 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
 113 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
 114 #endif
 115 #define ETHTOOL_LINK_MODE_MASK_MAX_KERNEL_NU32 (SCHAR_MAX)
 116
 117 /**
 118  * Return private structure associated with an Ethernet device.
 119  *
 120  * @param dev
 121  *   Pointer to Ethernet device structure.
 122  *
 123  * @return
 124  *   Pointer to private structure.
 125  */
 126 struct priv *
 127 mlx5_get_priv(struct rte_eth_dev *dev)
 128 {
 129         struct mlx5_secondary_data *sd;
 130
 131         if (!mlx5_is_secondary())
 132                 return dev->data->dev_private;
 133         sd = &mlx5_secondary_data[dev->data->port_id];
 134         return sd->data.dev_private;
 135 }
 136
 137 /**
 138  * Check if running as a secondary process.
 139  *
 140  * @return
 141  *   Nonzero if running as a secondary process.
 142  */
 143 inline int
 144 mlx5_is_secondary(void)
 145 {
 146         return rte_eal_process_type() != RTE_PROC_PRIMARY;
 147 }
 148
 149 /**
 150  * Get interface name from private structure.
 151  *
 152  * @param[in] priv
 153  *   Pointer to private structure.
 154  * @param[out] ifname
 155  *   Interface name output buffer.
 156  *
 157  * @return
 158  *   0 on success, -1 on failure and errno is set.
 159  */
 160 int
 161 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
 162 {
 163         DIR *dir;
 164         struct dirent *dent;
 165         unsigned int dev_type = 0;
 166         unsigned int dev_port_prev = ~0u;
 167         char match[IF_NAMESIZE] = "";
 168
 169         {
 170                 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
 171
 172                 dir = opendir(path);
 173                 if (dir == NULL)
 174                         return -1;
 175         }
 176         while ((dent = readdir(dir)) != NULL) {
 177                 char *name = dent->d_name;
 178                 FILE *file;
 179                 unsigned int dev_port;
 180                 int r;
 181
 182                 if ((name[0] == '.') &&
 183                     ((name[1] == '\0') ||
 184                      ((name[1] == '.') && (name[2] == '\0'))))
 185                         continue;
 186
 187                 MKSTR(path, "%s/device/net/%s/%s",
 188                       priv->ctx->device->ibdev_path, name,
 189                       (dev_type ? "dev_id" : "dev_port"));
 190
 191                 file = fopen(path, "rb");
 192                 if (file == NULL) {
 193                         if (errno != ENOENT)
 194                                 continue;
 195                         /*
 196                          * Switch to dev_id when dev_port does not exist as
 197                          * is the case with Linux kernel versions < 3.15.
 198                          */
 199 try_dev_id:
 200                         match[0] = '\0';
 201                         if (dev_type)
 202                                 break;
 203                         dev_type = 1;
 204                         dev_port_prev = ~0u;
 205                         rewinddir(dir);
 206                         continue;
 207                 }
 208                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 209                 fclose(file);
 210                 if (r != 1)
 211                         continue;
 212                 /*
 213                  * Switch to dev_id when dev_port returns the same value for
 214                  * all ports. May happen when using a MOFED release older than
 215                  * 3.0 with a Linux kernel >= 3.15.
 216                  */
 217                 if (dev_port == dev_port_prev)
 218                         goto try_dev_id;
 219                 dev_port_prev = dev_port;
 220                 if (dev_port == (priv->port - 1u))
 221                         snprintf(match, sizeof(match), "%s", name);
 222         }
 223         closedir(dir);
 224         if (match[0] == '\0')
 225                 return -1;
 226         strncpy(*ifname, match, sizeof(*ifname));
 227         return 0;
 228 }
 229
 230 /**
 231  * Check if the counter is located on ib counters file.
 232  *
 233  * @param[in] cntr
 234  *   Counter name.
 235  *
 236  * @return
 237  *   1 if counter is located on ib counters file , 0 otherwise.
 238  */
 239 int
 240 priv_is_ib_cntr(const char *cntr)
 241 {
 242         if (!strcmp(cntr, "out_of_buffer"))
 243                 return 1;
 244         return 0;
 245 }
 246
 247 /**
 248  * Read from sysfs entry.
 249  *
 250  * @param[in] priv
 251  *   Pointer to private structure.
 252  * @param[in] entry
 253  *   Entry name relative to sysfs path.
 254  * @param[out] buf
 255  *   Data output buffer.
 256  * @param size
 257  *   Buffer size.
 258  *
 259  * @return
 260  *   0 on success, -1 on failure and errno is set.
 261  */
 262 static int
 263 priv_sysfs_read(const struct priv *priv, const char *entry,
 264                 char *buf, size_t size)
 265 {
 266         char ifname[IF_NAMESIZE];
 267         FILE *file;
 268         int ret;
 269         int err;
 270
 271         if (priv_get_ifname(priv, &ifname))
 272                 return -1;
 273
 274         if (priv_is_ib_cntr(entry)) {
 275                 MKSTR(path, "%s/ports/1/hw_counters/%s",
 276                       priv->ctx->device->ibdev_path, entry);
 277                 file = fopen(path, "rb");
 278         } else {
 279                 MKSTR(path, "%s/device/net/%s/%s",
 280                       priv->ctx->device->ibdev_path, ifname, entry);
 281                 file = fopen(path, "rb");
 282         }
 283         if (file == NULL)
 284                 return -1;
 285         ret = fread(buf, 1, size, file);
 286         err = errno;
 287         if (((size_t)ret < size) && (ferror(file)))
 288                 ret = -1;
 289         else
 290                 ret = size;
 291         fclose(file);
 292         errno = err;
 293         return ret;
 294 }
 295
 296 /**
 297  * Write to sysfs entry.
 298  *
 299  * @param[in] priv
 300  *   Pointer to private structure.
 301  * @param[in] entry
 302  *   Entry name relative to sysfs path.
 303  * @param[in] buf
 304  *   Data buffer.
 305  * @param size
 306  *   Buffer size.
 307  *
 308  * @return
 309  *   0 on success, -1 on failure and errno is set.
 310  */
 311 static int
 312 priv_sysfs_write(const struct priv *priv, const char *entry,
 313                  char *buf, size_t size)
 314 {
 315         char ifname[IF_NAMESIZE];
 316         FILE *file;
 317         int ret;
 318         int err;
 319
 320         if (priv_get_ifname(priv, &ifname))
 321                 return -1;
 322
 323         MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
 324               ifname, entry);
 325
 326         file = fopen(path, "wb");
 327         if (file == NULL)
 328                 return -1;
 329         ret = fwrite(buf, 1, size, file);
 330         err = errno;
 331         if (((size_t)ret < size) || (ferror(file)))
 332                 ret = -1;
 333         else
 334                 ret = size;
 335         fclose(file);
 336         errno = err;
 337         return ret;
 338 }
 339
 340 /**
 341  * Get unsigned long sysfs property.
 342  *
 343  * @param priv
 344  *   Pointer to private structure.
 345  * @param[in] name
 346  *   Entry name relative to sysfs path.
 347  * @param[out] value
 348  *   Value output buffer.
 349  *
 350  * @return
 351  *   0 on success, -1 on failure and errno is set.
 352  */
 353 static int
 354 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
 355 {
 356         int ret;
 357         unsigned long value_ret;
 358         char value_str[32];
 359
 360         ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
 361         if (ret == -1) {
 362                 DEBUG("cannot read %s value from sysfs: %s",
 363                       name, strerror(errno));
 364                 return -1;
 365         }
 366         value_str[ret] = '\0';
 367         errno = 0;
 368         value_ret = strtoul(value_str, NULL, 0);
 369         if (errno) {
 370                 DEBUG("invalid %s value `%s': %s", name, value_str,
 371                       strerror(errno));
 372                 return -1;
 373         }
 374         *value = value_ret;
 375         return 0;
 376 }
 377
 378 /**
 379  * Set unsigned long sysfs property.
 380  *
 381  * @param priv
 382  *   Pointer to private structure.
 383  * @param[in] name
 384  *   Entry name relative to sysfs path.
 385  * @param value
 386  *   Value to set.
 387  *
 388  * @return
 389  *   0 on success, -1 on failure and errno is set.
 390  */
 391 static int
 392 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
 393 {
 394         int ret;
 395         MKSTR(value_str, "%lu", value);
 396
 397         ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
 398         if (ret == -1) {
 399                 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
 400                       name, value_str, value, strerror(errno));
 401                 return -1;
 402         }
 403         return 0;
 404 }
 405
 406 /**
 407  * Perform ifreq ioctl() on associated Ethernet device.
 408  *
 409  * @param[in] priv
 410  *   Pointer to private structure.
 411  * @param req
 412  *   Request number to pass to ioctl().
 413  * @param[out] ifr
 414  *   Interface request structure output buffer.
 415  *
 416  * @return
 417  *   0 on success, -1 on failure and errno is set.
 418  */
 419 int
 420 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
 421 {
 422         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 423         int ret = -1;
 424
 425         if (sock == -1)
 426                 return ret;
 427         if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
 428                 ret = ioctl(sock, req, ifr);
 429         close(sock);
 430         return ret;
 431 }
 432
 433 /**
 434  * Return the number of active VFs for the current device.
 435  *
 436  * @param[in] priv
 437  *   Pointer to private structure.
 438  * @param[out] num_vfs
 439  *   Number of active VFs.
 440  *
 441  * @return
 442  *   0 on success, -1 on failure and errno is set.
 443  */
 444 int
 445 priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs)
 446 {
 447         /* The sysfs entry name depends on the operating system. */
 448         const char **name = (const char *[]){
 449                 "device/sriov_numvfs",
 450                 "device/mlx5_num_vfs",
 451                 NULL,
 452         };
 453         int ret;
 454
 455         do {
 456                 unsigned long ulong_num_vfs;
 457
 458                 ret = priv_get_sysfs_ulong(priv, *name, &ulong_num_vfs);
 459                 if (!ret)
 460                         *num_vfs = ulong_num_vfs;
 461         } while (*(++name) && ret);
 462         return ret;
 463 }
 464
 465 /**
 466  * Get device MTU.
 467  *
 468  * @param priv
 469  *   Pointer to private structure.
 470  * @param[out] mtu
 471  *   MTU value output buffer.
 472  *
 473  * @return
 474  *   0 on success, -1 on failure and errno is set.
 475  */
 476 int
 477 priv_get_mtu(struct priv *priv, uint16_t *mtu)
 478 {
 479         unsigned long ulong_mtu;
 480
 481         if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
 482                 return -1;
 483         *mtu = ulong_mtu;
 484         return 0;
 485 }
 486
 487 /**
 488  * Read device counter from sysfs.
 489  *
 490  * @param priv
 491  *   Pointer to private structure.
 492  * @param name
 493  *   Counter name.
 494  * @param[out] cntr
 495  *   Counter output buffer.
 496  *
 497  * @return
 498  *   0 on success, -1 on failure and errno is set.
 499  */
 500 int
 501 priv_get_cntr_sysfs(struct priv *priv, const char *name, uint64_t *cntr)
 502 {
 503         unsigned long ulong_ctr;
 504
 505         if (priv_get_sysfs_ulong(priv, name, &ulong_ctr) == -1)
 506                 return -1;
 507         *cntr = ulong_ctr;
 508         return 0;
 509 }
 510
 511 /**
 512  * Set device MTU.
 513  *
 514  * @param priv
 515  *   Pointer to private structure.
 516  * @param mtu
 517  *   MTU value to set.
 518  *
 519  * @return
 520  *   0 on success, -1 on failure and errno is set.
 521  */
 522 static int
 523 priv_set_mtu(struct priv *priv, uint16_t mtu)
 524 {
 525         uint16_t new_mtu;
 526
 527         if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
 528             priv_get_mtu(priv, &new_mtu))
 529                 return -1;
 530         if (new_mtu == mtu)
 531                 return 0;
 532         errno = EINVAL;
 533         return -1;
 534 }
 535
 536 /**
 537  * Set device flags.
 538  *
 539  * @param priv
 540  *   Pointer to private structure.
 541  * @param keep
 542  *   Bitmask for flags that must remain untouched.
 543  * @param flags
 544  *   Bitmask for flags to modify.
 545  *
 546  * @return
 547  *   0 on success, -1 on failure and errno is set.
 548  */
 549 int
 550 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
 551 {
 552         unsigned long tmp;
 553
 554         if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
 555                 return -1;
 556         tmp &= keep;
 557         tmp |= (flags & (~keep));
 558         return priv_set_sysfs_ulong(priv, "flags", tmp);
 559 }
 560
 561 /**
 562  * Ethernet device configuration.
 563  *
 564  * Prepare the driver for a given number of TX and RX queues.
 565  *
 566  * @param dev
 567  *   Pointer to Ethernet device structure.
 568  *
 569  * @return
 570  *   0 on success, errno value on failure.
 571  */
 572 static int
 573 dev_configure(struct rte_eth_dev *dev)
 574 {
 575         struct priv *priv = dev->data->dev_private;
 576         unsigned int rxqs_n = dev->data->nb_rx_queues;
 577         unsigned int txqs_n = dev->data->nb_tx_queues;
 578         unsigned int i;
 579         unsigned int j;
 580         unsigned int reta_idx_n;
 581
 582         priv->rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
 583         priv->rxqs = (void *)dev->data->rx_queues;
 584         priv->txqs = (void *)dev->data->tx_queues;
 585         if (txqs_n != priv->txqs_n) {
 586                 INFO("%p: TX queues number update: %u -> %u",
 587                      (void *)dev, priv->txqs_n, txqs_n);
 588                 priv->txqs_n = txqs_n;
 589         }
 590         if (rxqs_n > priv->ind_table_max_size) {
 591                 ERROR("cannot handle this many RX queues (%u)", rxqs_n);
 592                 return EINVAL;
 593         }
 594         if (rxqs_n == priv->rxqs_n)
 595                 return 0;
 596         INFO("%p: RX queues number update: %u -> %u",
 597              (void *)dev, priv->rxqs_n, rxqs_n);
 598         priv->rxqs_n = rxqs_n;
 599         /* If the requested number of RX queues is not a power of two, use the
 600          * maximum indirection table size for better balancing.
 601          * The result is always rounded to the next power of two. */
 602         reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
 603                                      priv->ind_table_max_size :
 604                                      rxqs_n));
 605         if (priv_rss_reta_index_resize(priv, reta_idx_n))
 606                 return ENOMEM;
 607         /* When the number of RX queues is not a power of two, the remaining
 608          * table entries are padded with reused WQs and hashes are not spread
 609          * uniformly. */
 610         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
 611                 (*priv->reta_idx)[i] = j;
 612                 if (++j == rxqs_n)
 613                         j = 0;
 614         }
 615         return 0;
 616 }
 617
 618 /**
 619  * DPDK callback for Ethernet device configuration.
 620  *
 621  * @param dev
 622  *   Pointer to Ethernet device structure.
 623  *
 624  * @return
 625  *   0 on success, negative errno value on failure.
 626  */
 627 int
 628 mlx5_dev_configure(struct rte_eth_dev *dev)
 629 {
 630         struct priv *priv = dev->data->dev_private;
 631         int ret;
 632
 633         if (mlx5_is_secondary())
 634                 return -E_RTE_SECONDARY;
 635
 636         priv_lock(priv);
 637         ret = dev_configure(dev);
 638         assert(ret >= 0);
 639         priv_unlock(priv);
 640         return -ret;
 641 }
 642
 643 /**
 644  * DPDK callback to get information about the device.
 645  *
 646  * @param dev
 647  *   Pointer to Ethernet device structure.
 648  * @param[out] info
 649  *   Info structure output buffer.
 650  */
 651 void
 652 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 653 {
 654         struct priv *priv = mlx5_get_priv(dev);
 655         unsigned int max;
 656         char ifname[IF_NAMESIZE];
 657
 658         info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
 659
 660         priv_lock(priv);
 661         /* FIXME: we should ask the device for these values. */
 662         info->min_rx_bufsize = 32;
 663         info->max_rx_pktlen = 65536;
 664         /*
 665          * Since we need one CQ per QP, the limit is the minimum number
 666          * between the two values.
 667          */
 668         max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
 669                priv->device_attr.max_qp : priv->device_attr.max_cq);
 670         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
 671         if (max >= 65535)
 672                 max = 65535;
 673         info->max_rx_queues = max;
 674         info->max_tx_queues = max;
 675         info->max_mac_addrs = RTE_DIM(priv->mac);
 676         info->rx_offload_capa =
 677                 (priv->hw_csum ?
 678                  (DEV_RX_OFFLOAD_IPV4_CKSUM |
 679                   DEV_RX_OFFLOAD_UDP_CKSUM |
 680                   DEV_RX_OFFLOAD_TCP_CKSUM) :
 681                  0) |
 682                 (priv->hw_vlan_strip ? DEV_RX_OFFLOAD_VLAN_STRIP : 0);
 683         if (!priv->mps)
 684                 info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
 685         if (priv->hw_csum)
 686                 info->tx_offload_capa |=
 687                         (DEV_TX_OFFLOAD_IPV4_CKSUM |
 688                          DEV_TX_OFFLOAD_UDP_CKSUM |
 689                          DEV_TX_OFFLOAD_TCP_CKSUM);
 690         if (priv->tso)
 691                 info->tx_offload_capa |= DEV_TX_OFFLOAD_TCP_TSO;
 692         if (priv->tunnel_en)
 693                 info->tx_offload_capa |= (DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM |
 694                                           DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
 695                                           DEV_TX_OFFLOAD_GRE_TNL_TSO);
 696         if (priv_get_ifname(priv, &ifname) == 0)
 697                 info->if_index = if_nametoindex(ifname);
 698         info->reta_size = priv->reta_idx_n ?
 699                 priv->reta_idx_n : priv->ind_table_max_size;
 700         info->hash_key_size = ((*priv->rss_conf) ?
 701                                (*priv->rss_conf)[0]->rss_key_len :
 702                                0);
 703         info->speed_capa = priv->link_speed_capa;
 704         priv_unlock(priv);
 705 }
 706
 707 const uint32_t *
 708 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 709 {
 710         static const uint32_t ptypes[] = {
 711                 /* refers to rxq_cq_to_pkt_type() */
 712                 RTE_PTYPE_L2_ETHER,
 713                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 714                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 715                 RTE_PTYPE_L4_NONFRAG,
 716                 RTE_PTYPE_L4_FRAG,
 717                 RTE_PTYPE_L4_TCP,
 718                 RTE_PTYPE_L4_UDP,
 719                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 720                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 721                 RTE_PTYPE_INNER_L4_NONFRAG,
 722                 RTE_PTYPE_INNER_L4_FRAG,
 723                 RTE_PTYPE_INNER_L4_TCP,
 724                 RTE_PTYPE_INNER_L4_UDP,
 725                 RTE_PTYPE_UNKNOWN
 726         };
 727
 728         if (dev->rx_pkt_burst == mlx5_rx_burst ||
 729             dev->rx_pkt_burst == mlx5_rx_burst_vec)
 730                 return ptypes;
 731         return NULL;
 732 }
 733
 734 /**
 735  * DPDK callback to retrieve physical link information.
 736  *
 737  * @param dev
 738  *   Pointer to Ethernet device structure.
 739  * @param wait_to_complete
 740  *   Wait for request completion (ignored).
 741  */
 742 static int
 743 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
 744 {
 745         struct priv *priv = mlx5_get_priv(dev);
 746         struct ethtool_cmd edata = {
 747                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
 748         };
 749         struct ifreq ifr;
 750         struct rte_eth_link dev_link;
 751         int link_speed = 0;
 752
 753         /* priv_lock() is not taken to allow concurrent calls. */
 754
 755         (void)wait_to_complete;
 756         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
 757                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
 758                 return -1;
 759         }
 760         memset(&dev_link, 0, sizeof(dev_link));
 761         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 762                                 (ifr.ifr_flags & IFF_RUNNING));
 763         ifr.ifr_data = (void *)&edata;
 764         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 765                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
 766                      strerror(errno));
 767                 return -1;
 768         }
 769         link_speed = ethtool_cmd_speed(&edata);
 770         if (link_speed == -1)
 771                 dev_link.link_speed = 0;
 772         else
 773                 dev_link.link_speed = link_speed;
 774         priv->link_speed_capa = 0;
 775         if (edata.supported & SUPPORTED_Autoneg)
 776                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 777         if (edata.supported & (SUPPORTED_1000baseT_Full |
 778                                SUPPORTED_1000baseKX_Full))
 779                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 780         if (edata.supported & SUPPORTED_10000baseKR_Full)
 781                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 782         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
 783                                SUPPORTED_40000baseCR4_Full |
 784                                SUPPORTED_40000baseSR4_Full |
 785                                SUPPORTED_40000baseLR4_Full))
 786                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 787         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
 788                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 789         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 790                         ETH_LINK_SPEED_FIXED);
 791         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 792                 /* Link status changed. */
 793                 dev->data->dev_link = dev_link;
 794                 return 0;
 795         }
 796         /* Link status is still the same. */
 797         return -1;
 798 }
 799
 800 /**
 801  * Retrieve physical link information (unlocked version using new ioctl).
 802  *
 803  * @param dev
 804  *   Pointer to Ethernet device structure.
 805  * @param wait_to_complete
 806  *   Wait for request completion (ignored).
 807  */
 808 static int
 809 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
 810 {
 811         struct priv *priv = mlx5_get_priv(dev);
 812         __extension__ struct {
 813                 struct ethtool_link_settings edata;
 814                 uint32_t link_mode_data[3 *
 815                                         ETHTOOL_LINK_MODE_MASK_MAX_KERNEL_NU32];
 816         } ecmd;
 817
 818         struct ifreq ifr;
 819         struct rte_eth_link dev_link;
 820         uint64_t sc;
 821
 822         (void)wait_to_complete;
 823         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
 824                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
 825                 return -1;
 826         }
 827         memset(&dev_link, 0, sizeof(dev_link));
 828         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
 829                                 (ifr.ifr_flags & IFF_RUNNING));
 830         memset(&ecmd, 0, sizeof(ecmd));
 831         ecmd.edata.cmd = ETHTOOL_GLINKSETTINGS;
 832         ifr.ifr_data = (void *)&ecmd;
 833         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 834                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
 835                       strerror(errno));
 836                 return -1;
 837         }
 838         ecmd.edata.link_mode_masks_nwords = -ecmd.edata.link_mode_masks_nwords;
 839         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 840                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
 841                       strerror(errno));
 842                 return -1;
 843         }
 844         dev_link.link_speed = ecmd.edata.speed;
 845         sc = ecmd.edata.link_mode_masks[0] |
 846                 ((uint64_t)ecmd.edata.link_mode_masks[1] << 32);
 847         priv->link_speed_capa = 0;
 848         if (sc & ETHTOOL_LINK_MODE_Autoneg_BIT)
 849                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
 850         if (sc & (ETHTOOL_LINK_MODE_1000baseT_Full_BIT |
 851                   ETHTOOL_LINK_MODE_1000baseKX_Full_BIT))
 852                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
 853         if (sc & (ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT |
 854                   ETHTOOL_LINK_MODE_10000baseKR_Full_BIT |
 855                   ETHTOOL_LINK_MODE_10000baseR_FEC_BIT))
 856                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
 857         if (sc & (ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT |
 858                   ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT))
 859                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
 860         if (sc & (ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT |
 861                   ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT |
 862                   ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT |
 863                   ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT))
 864                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
 865         if (sc & (ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT |
 866                   ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT |
 867                   ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT |
 868                   ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT))
 869                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
 870         if (sc & (ETHTOOL_LINK_MODE_25000baseCR_Full_BIT |
 871                   ETHTOOL_LINK_MODE_25000baseKR_Full_BIT |
 872                   ETHTOOL_LINK_MODE_25000baseSR_Full_BIT))
 873                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
 874         if (sc & (ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT |
 875                   ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT))
 876                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
 877         if (sc & (ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT |
 878                   ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT |
 879                   ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT |
 880                   ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT))
 881                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
 882         dev_link.link_duplex = ((ecmd.edata.duplex == DUPLEX_HALF) ?
 883                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
 884         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
 885                                   ETH_LINK_SPEED_FIXED);
 886         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
 887                 /* Link status changed. */
 888                 dev->data->dev_link = dev_link;
 889                 return 0;
 890         }
 891         /* Link status is still the same. */
 892         return -1;
 893 }
 894
 895 /**
 896  * DPDK callback to retrieve physical link information.
 897  *
 898  * @param dev
 899  *   Pointer to Ethernet device structure.
 900  * @param wait_to_complete
 901  *   Wait for request completion (ignored).
 902  */
 903 int
 904 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
 905 {
 906         struct utsname utsname;
 907         int ver[3];
 908
 909         if (uname(&utsname) == -1 ||
 910             sscanf(utsname.release, "%d.%d.%d",
 911                    &ver[0], &ver[1], &ver[2]) != 3 ||
 912             KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
 913                 return mlx5_link_update_unlocked_gset(dev, wait_to_complete);
 914         return mlx5_link_update_unlocked_gs(dev, wait_to_complete);
 915 }
 916
 917 /**
 918  * DPDK callback to change the MTU.
 919  *
 920  * @param dev
 921  *   Pointer to Ethernet device structure.
 922  * @param in_mtu
 923  *   New MTU.
 924  *
 925  * @return
 926  *   0 on success, negative errno value on failure.
 927  */
 928 int
 929 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 930 {
 931         struct priv *priv = dev->data->dev_private;
 932         uint16_t kern_mtu;
 933         int ret = 0;
 934
 935         if (mlx5_is_secondary())
 936                 return -E_RTE_SECONDARY;
 937
 938         priv_lock(priv);
 939         ret = priv_get_mtu(priv, &kern_mtu);
 940         if (ret)
 941                 goto out;
 942         /* Set kernel interface MTU first. */
 943         ret = priv_set_mtu(priv, mtu);
 944         if (ret)
 945                 goto out;
 946         ret = priv_get_mtu(priv, &kern_mtu);
 947         if (ret)
 948                 goto out;
 949         if (kern_mtu == mtu) {
 950                 priv->mtu = mtu;
 951                 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
 952         }
 953         priv_unlock(priv);
 954         return 0;
 955 out:
 956         ret = errno;
 957         WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
 958              strerror(ret));
 959         priv_unlock(priv);
 960         assert(ret >= 0);
 961         return -ret;
 962 }
 963
 964 /**
 965  * DPDK callback to get flow control status.
 966  *
 967  * @param dev
 968  *   Pointer to Ethernet device structure.
 969  * @param[out] fc_conf
 970  *   Flow control output buffer.
 971  *
 972  * @return
 973  *   0 on success, negative errno value on failure.
 974  */
 975 int
 976 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
 977 {
 978         struct priv *priv = dev->data->dev_private;
 979         struct ifreq ifr;
 980         struct ethtool_pauseparam ethpause = {
 981                 .cmd = ETHTOOL_GPAUSEPARAM
 982         };
 983         int ret;
 984
 985         if (mlx5_is_secondary())
 986                 return -E_RTE_SECONDARY;
 987
 988         ifr.ifr_data = (void *)&ethpause;
 989         priv_lock(priv);
 990         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
 991                 ret = errno;
 992                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
 993                      " failed: %s",
 994                      strerror(ret));
 995                 goto out;
 996         }
 997
 998         fc_conf->autoneg = ethpause.autoneg;
 999         if (ethpause.rx_pause && ethpause.tx_pause)
1000                 fc_conf->mode = RTE_FC_FULL;
1001         else if (ethpause.rx_pause)
1002                 fc_conf->mode = RTE_FC_RX_PAUSE;
1003         else if (ethpause.tx_pause)
1004                 fc_conf->mode = RTE_FC_TX_PAUSE;
1005         else
1006                 fc_conf->mode = RTE_FC_NONE;
1007         ret = 0;
1008
1009 out:
1010         priv_unlock(priv);
1011         assert(ret >= 0);
1012         return -ret;
1013 }
1014
1015 /**
1016  * DPDK callback to modify flow control parameters.
1017  *
1018  * @param dev
1019  *   Pointer to Ethernet device structure.
1020  * @param[in] fc_conf
1021  *   Flow control parameters.
1022  *
1023  * @return
1024  *   0 on success, negative errno value on failure.
1025  */
1026 int
1027 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1028 {
1029         struct priv *priv = dev->data->dev_private;
1030         struct ifreq ifr;
1031         struct ethtool_pauseparam ethpause = {
1032                 .cmd = ETHTOOL_SPAUSEPARAM
1033         };
1034         int ret;
1035
1036         if (mlx5_is_secondary())
1037                 return -E_RTE_SECONDARY;
1038
1039         ifr.ifr_data = (void *)&ethpause;
1040         ethpause.autoneg = fc_conf->autoneg;
1041         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1042             (fc_conf->mode & RTE_FC_RX_PAUSE))
1043                 ethpause.rx_pause = 1;
1044         else
1045                 ethpause.rx_pause = 0;
1046
1047         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1048             (fc_conf->mode & RTE_FC_TX_PAUSE))
1049                 ethpause.tx_pause = 1;
1050         else
1051                 ethpause.tx_pause = 0;
1052
1053         priv_lock(priv);
1054         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1055                 ret = errno;
1056                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1057                      " failed: %s",
1058                      strerror(ret));
1059                 goto out;
1060         }
1061         ret = 0;
1062
1063 out:
1064         priv_unlock(priv);
1065         assert(ret >= 0);
1066         return -ret;
1067 }
1068
1069 /**
1070  * Get PCI information from struct ibv_device.
1071  *
1072  * @param device
1073  *   Pointer to Ethernet device structure.
1074  * @param[out] pci_addr
1075  *   PCI bus address output buffer.
1076  *
1077  * @return
1078  *   0 on success, -1 on failure and errno is set.
1079  */
1080 int
1081 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1082                             struct rte_pci_addr *pci_addr)
1083 {
1084         FILE *file;
1085         char line[32];
1086         MKSTR(path, "%s/device/uevent", device->ibdev_path);
1087
1088         file = fopen(path, "rb");
1089         if (file == NULL)
1090                 return -1;
1091         while (fgets(line, sizeof(line), file) == line) {
1092                 size_t len = strlen(line);
1093                 int ret;
1094
1095                 /* Truncate long lines. */
1096                 if (len == (sizeof(line) - 1))
1097                         while (line[(len - 1)] != '\n') {
1098                                 ret = fgetc(file);
1099                                 if (ret == EOF)
1100                                         break;
1101                                 line[(len - 1)] = ret;
1102                         }
1103                 /* Extract information. */
1104                 if (sscanf(line,
1105                            "PCI_SLOT_NAME="
1106                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1107                            &pci_addr->domain,
1108                            &pci_addr->bus,
1109                            &pci_addr->devid,
1110                            &pci_addr->function) == 4) {
1111                         ret = 0;
1112                         break;
1113                 }
1114         }
1115         fclose(file);
1116         return 0;
1117 }
1118
1119 /**
1120  * Link status handler.
1121  *
1122  * @param priv
1123  *   Pointer to private structure.
1124  * @param dev
1125  *   Pointer to the rte_eth_dev structure.
1126  *
1127  * @return
1128  *   Nonzero if the callback process can be called immediately.
1129  */
1130 static int
1131 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
1132 {
1133         struct ibv_async_event event;
1134         struct rte_eth_link *link = &dev->data->dev_link;
1135         int ret = 0;
1136
1137         /* Read all message and acknowledge them. */
1138         for (;;) {
1139                 if (ibv_get_async_event(priv->ctx, &event))
1140                         break;
1141
1142                 if (event.event_type != IBV_EVENT_PORT_ACTIVE &&
1143                     event.event_type != IBV_EVENT_PORT_ERR)
1144                         DEBUG("event type %d on port %d not handled",
1145                               event.event_type, event.element.port_num);
1146                 ibv_ack_async_event(&event);
1147         }
1148         mlx5_link_update(dev, 0);
1149         if (((link->link_speed == 0) && link->link_status) ||
1150             ((link->link_speed != 0) && !link->link_status)) {
1151                 if (!priv->pending_alarm) {
1152                         /* Inconsistent status, check again later. */
1153                         priv->pending_alarm = 1;
1154                         rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1155                                           mlx5_dev_link_status_handler,
1156                                           dev);
1157                 }
1158         } else {
1159                 ret = 1;
1160         }
1161         return ret;
1162 }
1163
1164 /**
1165  * Handle delayed link status event.
1166  *
1167  * @param arg
1168  *   Registered argument.
1169  */
1170 void
1171 mlx5_dev_link_status_handler(void *arg)
1172 {
1173         struct rte_eth_dev *dev = arg;
1174         struct priv *priv = dev->data->dev_private;
1175         int ret;
1176
1177         priv_lock(priv);
1178         assert(priv->pending_alarm == 1);
1179         priv->pending_alarm = 0;
1180         ret = priv_dev_link_status_handler(priv, dev);
1181         priv_unlock(priv);
1182         if (ret)
1183                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL,
1184                                               NULL);
1185 }
1186
1187 /**
1188  * Handle interrupts from the NIC.
1189  *
1190  * @param[in] intr_handle
1191  *   Interrupt handler.
1192  * @param cb_arg
1193  *   Callback argument.
1194  */
1195 void
1196 mlx5_dev_interrupt_handler(void *cb_arg)
1197 {
1198         struct rte_eth_dev *dev = cb_arg;
1199         struct priv *priv = dev->data->dev_private;
1200         int ret;
1201
1202         priv_lock(priv);
1203         ret = priv_dev_link_status_handler(priv, dev);
1204         priv_unlock(priv);
1205         if (ret)
1206                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL,
1207                                               NULL);
1208 }
1209
1210 /**
1211  * Uninstall interrupt handler.
1212  *
1213  * @param priv
1214  *   Pointer to private structure.
1215  * @param dev
1216  *   Pointer to the rte_eth_dev structure.
1217  */
1218 void
1219 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1220 {
1221         if (!dev->data->dev_conf.intr_conf.lsc)
1222                 return;
1223         rte_intr_callback_unregister(&priv->intr_handle,
1224                                      mlx5_dev_interrupt_handler,
1225                                      dev);
1226         if (priv->pending_alarm)
1227                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1228         priv->pending_alarm = 0;
1229         priv->intr_handle.fd = 0;
1230         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1231 }
1232
1233 /**
1234  * Install interrupt handler.
1235  *
1236  * @param priv
1237  *   Pointer to private structure.
1238  * @param dev
1239  *   Pointer to the rte_eth_dev structure.
1240  */
1241 void
1242 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1243 {
1244         int rc, flags;
1245
1246         if (!dev->data->dev_conf.intr_conf.lsc)
1247                 return;
1248         assert(priv->ctx->async_fd > 0);
1249         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1250         rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1251         if (rc < 0) {
1252                 INFO("failed to change file descriptor async event queue");
1253                 dev->data->dev_conf.intr_conf.lsc = 0;
1254         } else {
1255                 priv->intr_handle.fd = priv->ctx->async_fd;
1256                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1257                 rte_intr_callback_register(&priv->intr_handle,
1258                                            mlx5_dev_interrupt_handler,
1259                                            dev);
1260         }
1261 }
1262
1263 /**
1264  * Change the link state (UP / DOWN).
1265  *
1266  * @param priv
1267  *   Pointer to Ethernet device structure.
1268  * @param up
1269  *   Nonzero for link up, otherwise link down.
1270  *
1271  * @return
1272  *   0 on success, errno value on failure.
1273  */
1274 static int
1275 priv_set_link(struct priv *priv, int up)
1276 {
1277         struct rte_eth_dev *dev = priv->dev;
1278         int err;
1279
1280         if (up) {
1281                 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
1282                 if (err)
1283                         return err;
1284                 priv_select_tx_function(priv);
1285                 priv_select_rx_function(priv);
1286         } else {
1287                 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
1288                 if (err)
1289                         return err;
1290                 dev->rx_pkt_burst = removed_rx_burst;
1291                 dev->tx_pkt_burst = removed_tx_burst;
1292         }
1293         return 0;
1294 }
1295
1296 /**
1297  * DPDK callback to bring the link DOWN.
1298  *
1299  * @param dev
1300  *   Pointer to Ethernet device structure.
1301  *
1302  * @return
1303  *   0 on success, errno value on failure.
1304  */
1305 int
1306 mlx5_set_link_down(struct rte_eth_dev *dev)
1307 {
1308         struct priv *priv = dev->data->dev_private;
1309         int err;
1310
1311         priv_lock(priv);
1312         err = priv_set_link(priv, 0);
1313         priv_unlock(priv);
1314         return err;
1315 }
1316
1317 /**
1318  * DPDK callback to bring the link UP.
1319  *
1320  * @param dev
1321  *   Pointer to Ethernet device structure.
1322  *
1323  * @return
1324  *   0 on success, errno value on failure.
1325  */
1326 int
1327 mlx5_set_link_up(struct rte_eth_dev *dev)
1328 {
1329         struct priv *priv = dev->data->dev_private;
1330         int err;
1331
1332         priv_lock(priv);
1333         err = priv_set_link(priv, 1);
1334         priv_unlock(priv);
1335         return err;
1336 }
1337
1338 /**
1339  * Configure secondary process queues from a private data pointer (primary
1340  * or secondary) and update burst callbacks. Can take place only once.
1341  *
1342  * All queues must have been previously created by the primary process to
1343  * avoid undefined behavior.
1344  *
1345  * @param priv
1346  *   Private data pointer from either primary or secondary process.
1347  *
1348  * @return
1349  *   Private data pointer from secondary process, NULL in case of error.
1350  */
1351 struct priv *
1352 mlx5_secondary_data_setup(struct priv *priv)
1353 {
1354         unsigned int port_id = 0;
1355         struct mlx5_secondary_data *sd;
1356         void **tx_queues;
1357         void **rx_queues;
1358         unsigned int nb_tx_queues;
1359         unsigned int nb_rx_queues;
1360         unsigned int i;
1361
1362         /* priv must be valid at this point. */
1363         assert(priv != NULL);
1364         /* priv->dev must also be valid but may point to local memory from
1365          * another process, possibly with the same address and must not
1366          * be dereferenced yet. */
1367         assert(priv->dev != NULL);
1368         /* Determine port ID by finding out where priv comes from. */
1369         while (1) {
1370                 sd = &mlx5_secondary_data[port_id];
1371                 rte_spinlock_lock(&sd->lock);
1372                 /* Primary process? */
1373                 if (sd->primary_priv == priv)
1374                         break;
1375                 /* Secondary process? */
1376                 if (sd->data.dev_private == priv)
1377                         break;
1378                 rte_spinlock_unlock(&sd->lock);
1379                 if (++port_id == RTE_DIM(mlx5_secondary_data))
1380                         port_id = 0;
1381         }
1382         /* Switch to secondary private structure. If private data has already
1383          * been updated by another thread, there is nothing else to do. */
1384         priv = sd->data.dev_private;
1385         if (priv->dev->data == &sd->data)
1386                 goto end;
1387         /* Sanity checks. Secondary private structure is supposed to point
1388          * to local eth_dev, itself still pointing to the shared device data
1389          * structure allocated by the primary process. */
1390         assert(sd->shared_dev_data != &sd->data);
1391         assert(sd->data.nb_tx_queues == 0);
1392         assert(sd->data.tx_queues == NULL);
1393         assert(sd->data.nb_rx_queues == 0);
1394         assert(sd->data.rx_queues == NULL);
1395         assert(priv != sd->primary_priv);
1396         assert(priv->dev->data == sd->shared_dev_data);
1397         assert(priv->txqs_n == 0);
1398         assert(priv->txqs == NULL);
1399         assert(priv->rxqs_n == 0);
1400         assert(priv->rxqs == NULL);
1401         nb_tx_queues = sd->shared_dev_data->nb_tx_queues;
1402         nb_rx_queues = sd->shared_dev_data->nb_rx_queues;
1403         /* Allocate local storage for queues. */
1404         tx_queues = rte_zmalloc("secondary ethdev->tx_queues",
1405                                 sizeof(sd->data.tx_queues[0]) * nb_tx_queues,
1406                                 RTE_CACHE_LINE_SIZE);
1407         rx_queues = rte_zmalloc("secondary ethdev->rx_queues",
1408                                 sizeof(sd->data.rx_queues[0]) * nb_rx_queues,
1409                                 RTE_CACHE_LINE_SIZE);
1410         if (tx_queues == NULL || rx_queues == NULL)
1411                 goto error;
1412         /* Lock to prevent control operations during setup. */
1413         priv_lock(priv);
1414         /* TX queues. */
1415         for (i = 0; i != nb_tx_queues; ++i) {
1416                 struct txq *primary_txq = (*sd->primary_priv->txqs)[i];
1417                 struct txq_ctrl *primary_txq_ctrl;
1418                 struct txq_ctrl *txq_ctrl;
1419
1420                 if (primary_txq == NULL)
1421                         continue;
1422                 primary_txq_ctrl = container_of(primary_txq,
1423                                                 struct txq_ctrl, txq);
1424                 txq_ctrl = rte_calloc_socket("TXQ", 1, sizeof(*txq_ctrl) +
1425                                              (1 << primary_txq->elts_n) *
1426                                              sizeof(struct rte_mbuf *), 0,
1427                                              primary_txq_ctrl->socket);
1428                 if (txq_ctrl != NULL) {
1429                         if (txq_ctrl_setup(priv->dev,
1430                                            txq_ctrl,
1431                                            1 << primary_txq->elts_n,
1432                                            primary_txq_ctrl->socket,
1433                                            NULL) == 0) {
1434                                 txq_ctrl->txq.stats.idx =
1435                                         primary_txq->stats.idx;
1436                                 tx_queues[i] = &txq_ctrl->txq;
1437                                 continue;
1438                         }
1439                         rte_free(txq_ctrl);
1440                 }
1441                 while (i) {
1442                         txq_ctrl = tx_queues[--i];
1443                         txq_cleanup(txq_ctrl);
1444                         rte_free(txq_ctrl);
1445                 }
1446                 goto error;
1447         }
1448         /* RX queues. */
1449         for (i = 0; i != nb_rx_queues; ++i) {
1450                 struct rxq_ctrl *primary_rxq =
1451                         container_of((*sd->primary_priv->rxqs)[i],
1452                                      struct rxq_ctrl, rxq);
1453
1454                 if (primary_rxq == NULL)
1455                         continue;
1456                 /* Not supported yet. */
1457                 rx_queues[i] = NULL;
1458         }
1459         /* Update everything. */
1460         priv->txqs = (void *)tx_queues;
1461         priv->txqs_n = nb_tx_queues;
1462         priv->rxqs = (void *)rx_queues;
1463         priv->rxqs_n = nb_rx_queues;
1464         sd->data.rx_queues = rx_queues;
1465         sd->data.tx_queues = tx_queues;
1466         sd->data.nb_rx_queues = nb_rx_queues;
1467         sd->data.nb_tx_queues = nb_tx_queues;
1468         sd->data.dev_link = sd->shared_dev_data->dev_link;
1469         sd->data.mtu = sd->shared_dev_data->mtu;
1470         memcpy(sd->data.rx_queue_state, sd->shared_dev_data->rx_queue_state,
1471                sizeof(sd->data.rx_queue_state));
1472         memcpy(sd->data.tx_queue_state, sd->shared_dev_data->tx_queue_state,
1473                sizeof(sd->data.tx_queue_state));
1474         sd->data.dev_flags = sd->shared_dev_data->dev_flags;
1475         /* Use local data from now on. */
1476         rte_mb();
1477         priv->dev->data = &sd->data;
1478         rte_mb();
1479         priv_select_tx_function(priv);
1480         priv_select_rx_function(priv);
1481         priv_unlock(priv);
1482 end:
1483         /* More sanity checks. */
1484         assert(priv->dev->data == &sd->data);
1485         rte_spinlock_unlock(&sd->lock);
1486         return priv;
1487 error:
1488         priv_unlock(priv);
1489         rte_free(tx_queues);
1490         rte_free(rx_queues);
1491         rte_spinlock_unlock(&sd->lock);
1492         return NULL;
1493 }
1494
1495 /**
1496  * Configure the TX function to use.
1497  *
1498  * @param priv
1499  *   Pointer to private structure.
1500  */
1501 void
1502 priv_select_tx_function(struct priv *priv)
1503 {
1504         priv->dev->tx_pkt_burst = mlx5_tx_burst;
1505         /* Select appropriate TX function. */
1506         if (priv->mps == MLX5_MPW_ENHANCED) {
1507                 if (priv_check_vec_tx_support(priv) > 0) {
1508                         if (priv_check_raw_vec_tx_support(priv) > 0)
1509                                 priv->dev->tx_pkt_burst = mlx5_tx_burst_raw_vec;
1510                         else
1511                                 priv->dev->tx_pkt_burst = mlx5_tx_burst_vec;
1512                         DEBUG("selected Enhanced MPW TX vectorized function");
1513                 } else {
1514                         priv->dev->tx_pkt_burst = mlx5_tx_burst_empw;
1515                         DEBUG("selected Enhanced MPW TX function");
1516                 }
1517         } else if (priv->mps && priv->txq_inline) {
1518                 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1519                 DEBUG("selected MPW inline TX function");
1520         } else if (priv->mps) {
1521                 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw;
1522                 DEBUG("selected MPW TX function");
1523         }
1524 }
1525
1526 /**
1527  * Configure the RX function to use.
1528  *
1529  * @param priv
1530  *   Pointer to private structure.
1531  */
1532 void
1533 priv_select_rx_function(struct priv *priv)
1534 {
1535         if (priv_check_vec_rx_support(priv) > 0) {
1536                 priv->dev->rx_pkt_burst = mlx5_rx_burst_vec;
1537                 DEBUG("selected RX vectorized function");
1538         } else {
1539                 priv->dev->rx_pkt_burst = mlx5_rx_burst;
1540         }
1541 }