1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright 2015 6WIND S.A. 3 * Copyright 2015 Mellanox Technologies, Ltd 4 */ 5 6 #define _GNU_SOURCE 7 8 #include <stddef.h> 9 #include <assert.h> 10 #include <inttypes.h> 11 #include <unistd.h> 12 #include <stdint.h> 13 #include <stdio.h> 14 #include <string.h> 15 #include <stdlib.h> 16 #include <errno.h> 17 #include <dirent.h> 18 #include <net/if.h> 19 #include <sys/ioctl.h> 20 #include <sys/socket.h> 21 #include <netinet/in.h> 22 #include <linux/ethtool.h> 23 #include <linux/sockios.h> 24 #include <fcntl.h> 25 #include <stdalign.h> 26 #include <sys/un.h> 27 #include <time.h> 28 29 #include <rte_atomic.h> 30 #include <rte_ethdev_driver.h> 31 #include <rte_bus_pci.h> 32 #include <rte_mbuf.h> 33 #include <rte_common.h> 34 #include <rte_interrupts.h> 35 #include <rte_malloc.h> 36 #include <rte_string_fns.h> 37 #include <rte_rwlock.h> 38 39 #include "mlx5.h" 40 #include "mlx5_glue.h" 41 #include "mlx5_rxtx.h" 42 #include "mlx5_utils.h" 43 44 /* Add defines in case the running kernel is not the same as user headers. */ 45 #ifndef ETHTOOL_GLINKSETTINGS 46 struct ethtool_link_settings { 47 uint32_t cmd; 48 uint32_t speed; 49 uint8_t duplex; 50 uint8_t port; 51 uint8_t phy_address; 52 uint8_t autoneg; 53 uint8_t mdio_support; 54 uint8_t eth_to_mdix; 55 uint8_t eth_tp_mdix_ctrl; 56 int8_t link_mode_masks_nwords; 57 uint32_t reserved[8]; 58 uint32_t link_mode_masks[]; 59 }; 60 61 #define ETHTOOL_GLINKSETTINGS 0x0000004c 62 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5 63 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6 64 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17 65 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18 66 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19 67 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20 68 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21 69 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22 70 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23 71 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24 72 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25 73 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26 74 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27 75 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28 76 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29 77 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30 78 #endif 79 #ifndef HAVE_ETHTOOL_LINK_MODE_25G 80 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31 81 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32 82 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33 83 #endif 84 #ifndef HAVE_ETHTOOL_LINK_MODE_50G 85 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34 86 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35 87 #endif 88 #ifndef HAVE_ETHTOOL_LINK_MODE_100G 89 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36 90 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37 91 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38 92 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39 93 #endif 94 95 /** 96 * Get interface name from private structure. 97 * 98 * @param[in] dev 99 * Pointer to Ethernet device. 100 * @param[out] ifname 101 * Interface name output buffer. 102 * 103 * @return 104 * 0 on success, a negative errno value otherwise and rte_errno is set. 105 */ 106 int 107 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE]) 108 { 109 struct priv *priv = dev->data->dev_private; 110 DIR *dir; 111 struct dirent *dent; 112 unsigned int dev_type = 0; 113 unsigned int dev_port_prev = ~0u; 114 char match[IF_NAMESIZE] = ""; 115 116 { 117 MKSTR(path, "%s/device/net", priv->ibdev_path); 118 119 dir = opendir(path); 120 if (dir == NULL) { 121 rte_errno = errno; 122 return -rte_errno; 123 } 124 } 125 while ((dent = readdir(dir)) != NULL) { 126 char *name = dent->d_name; 127 FILE *file; 128 unsigned int dev_port; 129 int r; 130 131 if ((name[0] == '.') && 132 ((name[1] == '\0') || 133 ((name[1] == '.') && (name[2] == '\0')))) 134 continue; 135 136 MKSTR(path, "%s/device/net/%s/%s", 137 priv->ibdev_path, name, 138 (dev_type ? "dev_id" : "dev_port")); 139 140 file = fopen(path, "rb"); 141 if (file == NULL) { 142 if (errno != ENOENT) 143 continue; 144 /* 145 * Switch to dev_id when dev_port does not exist as 146 * is the case with Linux kernel versions < 3.15. 147 */ 148 try_dev_id: 149 match[0] = '\0'; 150 if (dev_type) 151 break; 152 dev_type = 1; 153 dev_port_prev = ~0u; 154 rewinddir(dir); 155 continue; 156 } 157 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port); 158 fclose(file); 159 if (r != 1) 160 continue; 161 /* 162 * Switch to dev_id when dev_port returns the same value for 163 * all ports. May happen when using a MOFED release older than 164 * 3.0 with a Linux kernel >= 3.15. 165 */ 166 if (dev_port == dev_port_prev) 167 goto try_dev_id; 168 dev_port_prev = dev_port; 169 if (dev_port == (priv->port - 1u)) 170 strlcpy(match, name, sizeof(match)); 171 } 172 closedir(dir); 173 if (match[0] == '\0') { 174 rte_errno = ENOENT; 175 return -rte_errno; 176 } 177 strncpy(*ifname, match, sizeof(*ifname)); 178 return 0; 179 } 180 181 /** 182 * Get the interface index from device name. 183 * 184 * @param[in] dev 185 * Pointer to Ethernet device. 186 * 187 * @return 188 * Interface index on success, a negative errno value otherwise and 189 * rte_errno is set. 190 */ 191 int 192 mlx5_ifindex(const struct rte_eth_dev *dev) 193 { 194 char ifname[IF_NAMESIZE]; 195 int ret; 196 197 ret = mlx5_get_ifname(dev, &ifname); 198 if (ret) 199 return ret; 200 ret = if_nametoindex(ifname); 201 if (ret == -1) { 202 rte_errno = errno; 203 return -rte_errno; 204 } 205 return ret; 206 } 207 208 /** 209 * Perform ifreq ioctl() on associated Ethernet device. 210 * 211 * @param[in] dev 212 * Pointer to Ethernet device. 213 * @param req 214 * Request number to pass to ioctl(). 215 * @param[out] ifr 216 * Interface request structure output buffer. 217 * 218 * @return 219 * 0 on success, a negative errno value otherwise and rte_errno is set. 220 */ 221 int 222 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr) 223 { 224 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP); 225 int ret = 0; 226 227 if (sock == -1) { 228 rte_errno = errno; 229 return -rte_errno; 230 } 231 ret = mlx5_get_ifname(dev, &ifr->ifr_name); 232 if (ret) 233 goto error; 234 ret = ioctl(sock, req, ifr); 235 if (ret == -1) { 236 rte_errno = errno; 237 goto error; 238 } 239 close(sock); 240 return 0; 241 error: 242 close(sock); 243 return -rte_errno; 244 } 245 246 /** 247 * Get device MTU. 248 * 249 * @param dev 250 * Pointer to Ethernet device. 251 * @param[out] mtu 252 * MTU value output buffer. 253 * 254 * @return 255 * 0 on success, a negative errno value otherwise and rte_errno is set. 256 */ 257 int 258 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu) 259 { 260 struct ifreq request; 261 int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request); 262 263 if (ret) 264 return ret; 265 *mtu = request.ifr_mtu; 266 return 0; 267 } 268 269 /** 270 * Set device MTU. 271 * 272 * @param dev 273 * Pointer to Ethernet device. 274 * @param mtu 275 * MTU value to set. 276 * 277 * @return 278 * 0 on success, a negative errno value otherwise and rte_errno is set. 279 */ 280 static int 281 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu) 282 { 283 struct ifreq request = { .ifr_mtu = mtu, }; 284 285 return mlx5_ifreq(dev, SIOCSIFMTU, &request); 286 } 287 288 /** 289 * Set device flags. 290 * 291 * @param dev 292 * Pointer to Ethernet device. 293 * @param keep 294 * Bitmask for flags that must remain untouched. 295 * @param flags 296 * Bitmask for flags to modify. 297 * 298 * @return 299 * 0 on success, a negative errno value otherwise and rte_errno is set. 300 */ 301 int 302 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags) 303 { 304 struct ifreq request; 305 int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request); 306 307 if (ret) 308 return ret; 309 request.ifr_flags &= keep; 310 request.ifr_flags |= flags & ~keep; 311 return mlx5_ifreq(dev, SIOCSIFFLAGS, &request); 312 } 313 314 /** 315 * DPDK callback for Ethernet device configuration. 316 * 317 * @param dev 318 * Pointer to Ethernet device structure. 319 * 320 * @return 321 * 0 on success, a negative errno value otherwise and rte_errno is set. 322 */ 323 int 324 mlx5_dev_configure(struct rte_eth_dev *dev) 325 { 326 struct priv *priv = dev->data->dev_private; 327 unsigned int rxqs_n = dev->data->nb_rx_queues; 328 unsigned int txqs_n = dev->data->nb_tx_queues; 329 unsigned int i; 330 unsigned int j; 331 unsigned int reta_idx_n; 332 const uint8_t use_app_rss_key = 333 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key; 334 int ret = 0; 335 336 if (use_app_rss_key && 337 (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len != 338 rss_hash_default_key_len)) { 339 DRV_LOG(ERR, "port %u RSS key len must be %zu Bytes long", 340 dev->data->port_id, rss_hash_default_key_len); 341 rte_errno = EINVAL; 342 return -rte_errno; 343 } 344 priv->rss_conf.rss_key = 345 rte_realloc(priv->rss_conf.rss_key, 346 rss_hash_default_key_len, 0); 347 if (!priv->rss_conf.rss_key) { 348 DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)", 349 dev->data->port_id, rxqs_n); 350 rte_errno = ENOMEM; 351 return -rte_errno; 352 } 353 memcpy(priv->rss_conf.rss_key, 354 use_app_rss_key ? 355 dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key : 356 rss_hash_default_key, 357 rss_hash_default_key_len); 358 priv->rss_conf.rss_key_len = rss_hash_default_key_len; 359 priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf; 360 priv->rxqs = (void *)dev->data->rx_queues; 361 priv->txqs = (void *)dev->data->tx_queues; 362 if (txqs_n != priv->txqs_n) { 363 DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u", 364 dev->data->port_id, priv->txqs_n, txqs_n); 365 priv->txqs_n = txqs_n; 366 } 367 if (rxqs_n > priv->config.ind_table_max_size) { 368 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)", 369 dev->data->port_id, rxqs_n); 370 rte_errno = EINVAL; 371 return -rte_errno; 372 } 373 if (rxqs_n == priv->rxqs_n) 374 return 0; 375 DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u", 376 dev->data->port_id, priv->rxqs_n, rxqs_n); 377 priv->rxqs_n = rxqs_n; 378 /* If the requested number of RX queues is not a power of two, use the 379 * maximum indirection table size for better balancing. 380 * The result is always rounded to the next power of two. */ 381 reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ? 382 priv->config.ind_table_max_size : 383 rxqs_n)); 384 ret = mlx5_rss_reta_index_resize(dev, reta_idx_n); 385 if (ret) 386 return ret; 387 /* When the number of RX queues is not a power of two, the remaining 388 * table entries are padded with reused WQs and hashes are not spread 389 * uniformly. */ 390 for (i = 0, j = 0; (i != reta_idx_n); ++i) { 391 (*priv->reta_idx)[i] = j; 392 if (++j == rxqs_n) 393 j = 0; 394 } 395 /* 396 * Once the device is added to the list of memory event callback, its 397 * global MR cache table cannot be expanded on the fly because of 398 * deadlock. If it overflows, lookup should be done by searching MR list 399 * linearly, which is slow. 400 */ 401 if (mlx5_mr_btree_init(&priv->mr.cache, MLX5_MR_BTREE_CACHE_N * 2, 402 dev->device->numa_node)) { 403 /* rte_errno is already set. */ 404 return -rte_errno; 405 } 406 rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock); 407 LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list, 408 priv, mem_event_cb); 409 rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock); 410 return 0; 411 } 412 413 /** 414 * Sets default tuning parameters. 415 * 416 * @param dev 417 * Pointer to Ethernet device. 418 * @param[out] info 419 * Info structure output buffer. 420 */ 421 static void 422 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info) 423 { 424 struct priv *priv = dev->data->dev_private; 425 426 /* Minimum CPU utilization. */ 427 info->default_rxportconf.ring_size = 256; 428 info->default_txportconf.ring_size = 256; 429 info->default_rxportconf.burst_size = 64; 430 info->default_txportconf.burst_size = 64; 431 if (priv->link_speed_capa & ETH_LINK_SPEED_100G) { 432 info->default_rxportconf.nb_queues = 16; 433 info->default_txportconf.nb_queues = 16; 434 if (dev->data->nb_rx_queues > 2 || 435 dev->data->nb_tx_queues > 2) { 436 /* Max Throughput. */ 437 info->default_rxportconf.ring_size = 2048; 438 info->default_txportconf.ring_size = 2048; 439 } 440 } else { 441 info->default_rxportconf.nb_queues = 8; 442 info->default_txportconf.nb_queues = 8; 443 if (dev->data->nb_rx_queues > 2 || 444 dev->data->nb_tx_queues > 2) { 445 /* Max Throughput. */ 446 info->default_rxportconf.ring_size = 4096; 447 info->default_txportconf.ring_size = 4096; 448 } 449 } 450 } 451 452 /** 453 * DPDK callback to get information about the device. 454 * 455 * @param dev 456 * Pointer to Ethernet device structure. 457 * @param[out] info 458 * Info structure output buffer. 459 */ 460 void 461 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info) 462 { 463 struct priv *priv = dev->data->dev_private; 464 struct mlx5_dev_config *config = &priv->config; 465 unsigned int max; 466 char ifname[IF_NAMESIZE]; 467 468 /* FIXME: we should ask the device for these values. */ 469 info->min_rx_bufsize = 32; 470 info->max_rx_pktlen = 65536; 471 /* 472 * Since we need one CQ per QP, the limit is the minimum number 473 * between the two values. 474 */ 475 max = RTE_MIN(priv->device_attr.orig_attr.max_cq, 476 priv->device_attr.orig_attr.max_qp); 477 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */ 478 if (max >= 65535) 479 max = 65535; 480 info->max_rx_queues = max; 481 info->max_tx_queues = max; 482 info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES; 483 info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev); 484 info->rx_offload_capa = (mlx5_get_rx_port_offloads() | 485 info->rx_queue_offload_capa); 486 info->tx_offload_capa = mlx5_get_tx_port_offloads(dev); 487 if (mlx5_get_ifname(dev, &ifname) == 0) 488 info->if_index = if_nametoindex(ifname); 489 info->reta_size = priv->reta_idx_n ? 490 priv->reta_idx_n : config->ind_table_max_size; 491 info->hash_key_size = rss_hash_default_key_len; 492 info->speed_capa = priv->link_speed_capa; 493 info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK; 494 mlx5_set_default_params(dev, info); 495 } 496 497 /** 498 * Get supported packet types. 499 * 500 * @param dev 501 * Pointer to Ethernet device structure. 502 * 503 * @return 504 * A pointer to the supported Packet types array. 505 */ 506 const uint32_t * 507 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev) 508 { 509 static const uint32_t ptypes[] = { 510 /* refers to rxq_cq_to_pkt_type() */ 511 RTE_PTYPE_L2_ETHER, 512 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN, 513 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN, 514 RTE_PTYPE_L4_NONFRAG, 515 RTE_PTYPE_L4_FRAG, 516 RTE_PTYPE_L4_TCP, 517 RTE_PTYPE_L4_UDP, 518 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN, 519 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN, 520 RTE_PTYPE_INNER_L4_NONFRAG, 521 RTE_PTYPE_INNER_L4_FRAG, 522 RTE_PTYPE_INNER_L4_TCP, 523 RTE_PTYPE_INNER_L4_UDP, 524 RTE_PTYPE_UNKNOWN 525 }; 526 527 if (dev->rx_pkt_burst == mlx5_rx_burst || 528 dev->rx_pkt_burst == mlx5_rx_burst_mprq || 529 dev->rx_pkt_burst == mlx5_rx_burst_vec) 530 return ptypes; 531 return NULL; 532 } 533 534 /** 535 * DPDK callback to retrieve physical link information. 536 * 537 * @param dev 538 * Pointer to Ethernet device structure. 539 * @param[out] link 540 * Storage for current link status. 541 * 542 * @return 543 * 0 on success, a negative errno value otherwise and rte_errno is set. 544 */ 545 static int 546 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, 547 struct rte_eth_link *link) 548 { 549 struct priv *priv = dev->data->dev_private; 550 struct ethtool_cmd edata = { 551 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */ 552 }; 553 struct ifreq ifr; 554 struct rte_eth_link dev_link; 555 int link_speed = 0; 556 int ret; 557 558 ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr); 559 if (ret) { 560 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s", 561 dev->data->port_id, strerror(rte_errno)); 562 return ret; 563 } 564 memset(&dev_link, 0, sizeof(dev_link)); 565 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) && 566 (ifr.ifr_flags & IFF_RUNNING)); 567 ifr.ifr_data = (void *)&edata; 568 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr); 569 if (ret) { 570 DRV_LOG(WARNING, 571 "port %u ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s", 572 dev->data->port_id, strerror(rte_errno)); 573 return ret; 574 } 575 link_speed = ethtool_cmd_speed(&edata); 576 if (link_speed == -1) 577 dev_link.link_speed = ETH_SPEED_NUM_NONE; 578 else 579 dev_link.link_speed = link_speed; 580 priv->link_speed_capa = 0; 581 if (edata.supported & SUPPORTED_Autoneg) 582 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG; 583 if (edata.supported & (SUPPORTED_1000baseT_Full | 584 SUPPORTED_1000baseKX_Full)) 585 priv->link_speed_capa |= ETH_LINK_SPEED_1G; 586 if (edata.supported & SUPPORTED_10000baseKR_Full) 587 priv->link_speed_capa |= ETH_LINK_SPEED_10G; 588 if (edata.supported & (SUPPORTED_40000baseKR4_Full | 589 SUPPORTED_40000baseCR4_Full | 590 SUPPORTED_40000baseSR4_Full | 591 SUPPORTED_40000baseLR4_Full)) 592 priv->link_speed_capa |= ETH_LINK_SPEED_40G; 593 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ? 594 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX); 595 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds & 596 ETH_LINK_SPEED_FIXED); 597 if ((dev_link.link_speed && !dev_link.link_status) || 598 (!dev_link.link_speed && dev_link.link_status)) { 599 rte_errno = EAGAIN; 600 return -rte_errno; 601 } 602 *link = dev_link; 603 return 0; 604 } 605 606 /** 607 * Retrieve physical link information (unlocked version using new ioctl). 608 * 609 * @param dev 610 * Pointer to Ethernet device structure. 611 * @param[out] link 612 * Storage for current link status. 613 * 614 * @return 615 * 0 on success, a negative errno value otherwise and rte_errno is set. 616 */ 617 static int 618 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, 619 struct rte_eth_link *link) 620 621 { 622 struct priv *priv = dev->data->dev_private; 623 struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS }; 624 struct ifreq ifr; 625 struct rte_eth_link dev_link; 626 uint64_t sc; 627 int ret; 628 629 ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr); 630 if (ret) { 631 DRV_LOG(WARNING, "port %u ioctl(SIOCGIFFLAGS) failed: %s", 632 dev->data->port_id, strerror(rte_errno)); 633 return ret; 634 } 635 memset(&dev_link, 0, sizeof(dev_link)); 636 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) && 637 (ifr.ifr_flags & IFF_RUNNING)); 638 ifr.ifr_data = (void *)&gcmd; 639 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr); 640 if (ret) { 641 DRV_LOG(DEBUG, 642 "port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)" 643 " failed: %s", 644 dev->data->port_id, strerror(rte_errno)); 645 return ret; 646 } 647 gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords; 648 649 alignas(struct ethtool_link_settings) 650 uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) + 651 sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3]; 652 struct ethtool_link_settings *ecmd = (void *)data; 653 654 *ecmd = gcmd; 655 ifr.ifr_data = (void *)ecmd; 656 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr); 657 if (ret) { 658 DRV_LOG(DEBUG, 659 "port %u ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS)" 660 " failed: %s", 661 dev->data->port_id, strerror(rte_errno)); 662 return ret; 663 } 664 dev_link.link_speed = ecmd->speed; 665 sc = ecmd->link_mode_masks[0] | 666 ((uint64_t)ecmd->link_mode_masks[1] << 32); 667 priv->link_speed_capa = 0; 668 if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT)) 669 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG; 670 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) | 671 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT))) 672 priv->link_speed_capa |= ETH_LINK_SPEED_1G; 673 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) | 674 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) | 675 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT))) 676 priv->link_speed_capa |= ETH_LINK_SPEED_10G; 677 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) | 678 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT))) 679 priv->link_speed_capa |= ETH_LINK_SPEED_20G; 680 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) | 681 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) | 682 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) | 683 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT))) 684 priv->link_speed_capa |= ETH_LINK_SPEED_40G; 685 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) | 686 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) | 687 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) | 688 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT))) 689 priv->link_speed_capa |= ETH_LINK_SPEED_56G; 690 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) | 691 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) | 692 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT))) 693 priv->link_speed_capa |= ETH_LINK_SPEED_25G; 694 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) | 695 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT))) 696 priv->link_speed_capa |= ETH_LINK_SPEED_50G; 697 if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) | 698 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) | 699 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) | 700 MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT))) 701 priv->link_speed_capa |= ETH_LINK_SPEED_100G; 702 dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ? 703 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX); 704 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds & 705 ETH_LINK_SPEED_FIXED); 706 if ((dev_link.link_speed && !dev_link.link_status) || 707 (!dev_link.link_speed && dev_link.link_status)) { 708 rte_errno = EAGAIN; 709 return -rte_errno; 710 } 711 *link = dev_link; 712 return 0; 713 } 714 715 /** 716 * DPDK callback to retrieve physical link information. 717 * 718 * @param dev 719 * Pointer to Ethernet device structure. 720 * @param wait_to_complete 721 * Wait for request completion. 722 * 723 * @return 724 * 0 if link status was not updated, positive if it was, a negative errno 725 * value otherwise and rte_errno is set. 726 */ 727 int 728 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete) 729 { 730 int ret; 731 struct rte_eth_link dev_link; 732 time_t start_time = time(NULL); 733 734 do { 735 ret = mlx5_link_update_unlocked_gs(dev, &dev_link); 736 if (ret) 737 ret = mlx5_link_update_unlocked_gset(dev, &dev_link); 738 if (ret == 0) 739 break; 740 /* Handle wait to complete situation. */ 741 if (wait_to_complete && ret == -EAGAIN) { 742 if (abs((int)difftime(time(NULL), start_time)) < 743 MLX5_LINK_STATUS_TIMEOUT) { 744 usleep(0); 745 continue; 746 } else { 747 rte_errno = EBUSY; 748 return -rte_errno; 749 } 750 } else if (ret < 0) { 751 return ret; 752 } 753 } while (wait_to_complete); 754 ret = !!memcmp(&dev->data->dev_link, &dev_link, 755 sizeof(struct rte_eth_link)); 756 dev->data->dev_link = dev_link; 757 return ret; 758 } 759 760 /** 761 * DPDK callback to change the MTU. 762 * 763 * @param dev 764 * Pointer to Ethernet device structure. 765 * @param in_mtu 766 * New MTU. 767 * 768 * @return 769 * 0 on success, a negative errno value otherwise and rte_errno is set. 770 */ 771 int 772 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu) 773 { 774 struct priv *priv = dev->data->dev_private; 775 uint16_t kern_mtu = 0; 776 int ret; 777 778 ret = mlx5_get_mtu(dev, &kern_mtu); 779 if (ret) 780 return ret; 781 /* Set kernel interface MTU first. */ 782 ret = mlx5_set_mtu(dev, mtu); 783 if (ret) 784 return ret; 785 ret = mlx5_get_mtu(dev, &kern_mtu); 786 if (ret) 787 return ret; 788 if (kern_mtu == mtu) { 789 priv->mtu = mtu; 790 DRV_LOG(DEBUG, "port %u adapter MTU set to %u", 791 dev->data->port_id, mtu); 792 return 0; 793 } 794 rte_errno = EAGAIN; 795 return -rte_errno; 796 } 797 798 /** 799 * DPDK callback to get flow control status. 800 * 801 * @param dev 802 * Pointer to Ethernet device structure. 803 * @param[out] fc_conf 804 * Flow control output buffer. 805 * 806 * @return 807 * 0 on success, a negative errno value otherwise and rte_errno is set. 808 */ 809 int 810 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf) 811 { 812 struct ifreq ifr; 813 struct ethtool_pauseparam ethpause = { 814 .cmd = ETHTOOL_GPAUSEPARAM 815 }; 816 int ret; 817 818 ifr.ifr_data = (void *)ðpause; 819 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr); 820 if (ret) { 821 DRV_LOG(WARNING, 822 "port %u ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed:" 823 " %s", 824 dev->data->port_id, strerror(rte_errno)); 825 return ret; 826 } 827 fc_conf->autoneg = ethpause.autoneg; 828 if (ethpause.rx_pause && ethpause.tx_pause) 829 fc_conf->mode = RTE_FC_FULL; 830 else if (ethpause.rx_pause) 831 fc_conf->mode = RTE_FC_RX_PAUSE; 832 else if (ethpause.tx_pause) 833 fc_conf->mode = RTE_FC_TX_PAUSE; 834 else 835 fc_conf->mode = RTE_FC_NONE; 836 return 0; 837 } 838 839 /** 840 * DPDK callback to modify flow control parameters. 841 * 842 * @param dev 843 * Pointer to Ethernet device structure. 844 * @param[in] fc_conf 845 * Flow control parameters. 846 * 847 * @return 848 * 0 on success, a negative errno value otherwise and rte_errno is set. 849 */ 850 int 851 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf) 852 { 853 struct ifreq ifr; 854 struct ethtool_pauseparam ethpause = { 855 .cmd = ETHTOOL_SPAUSEPARAM 856 }; 857 int ret; 858 859 ifr.ifr_data = (void *)ðpause; 860 ethpause.autoneg = fc_conf->autoneg; 861 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) || 862 (fc_conf->mode & RTE_FC_RX_PAUSE)) 863 ethpause.rx_pause = 1; 864 else 865 ethpause.rx_pause = 0; 866 867 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) || 868 (fc_conf->mode & RTE_FC_TX_PAUSE)) 869 ethpause.tx_pause = 1; 870 else 871 ethpause.tx_pause = 0; 872 ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr); 873 if (ret) { 874 DRV_LOG(WARNING, 875 "port %u ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)" 876 " failed: %s", 877 dev->data->port_id, strerror(rte_errno)); 878 return ret; 879 } 880 return 0; 881 } 882 883 /** 884 * Get PCI information from struct ibv_device. 885 * 886 * @param device 887 * Pointer to Ethernet device structure. 888 * @param[out] pci_addr 889 * PCI bus address output buffer. 890 * 891 * @return 892 * 0 on success, a negative errno value otherwise and rte_errno is set. 893 */ 894 int 895 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device, 896 struct rte_pci_addr *pci_addr) 897 { 898 FILE *file; 899 char line[32]; 900 MKSTR(path, "%s/device/uevent", device->ibdev_path); 901 902 file = fopen(path, "rb"); 903 if (file == NULL) { 904 rte_errno = errno; 905 return -rte_errno; 906 } 907 while (fgets(line, sizeof(line), file) == line) { 908 size_t len = strlen(line); 909 int ret; 910 911 /* Truncate long lines. */ 912 if (len == (sizeof(line) - 1)) 913 while (line[(len - 1)] != '\n') { 914 ret = fgetc(file); 915 if (ret == EOF) 916 break; 917 line[(len - 1)] = ret; 918 } 919 /* Extract information. */ 920 if (sscanf(line, 921 "PCI_SLOT_NAME=" 922 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n", 923 &pci_addr->domain, 924 &pci_addr->bus, 925 &pci_addr->devid, 926 &pci_addr->function) == 4) { 927 ret = 0; 928 break; 929 } 930 } 931 fclose(file); 932 return 0; 933 } 934 935 /** 936 * Device status handler. 937 * 938 * @param dev 939 * Pointer to Ethernet device. 940 * @param events 941 * Pointer to event flags holder. 942 * 943 * @return 944 * Events bitmap of callback process which can be called immediately. 945 */ 946 static uint32_t 947 mlx5_dev_status_handler(struct rte_eth_dev *dev) 948 { 949 struct priv *priv = dev->data->dev_private; 950 struct ibv_async_event event; 951 uint32_t ret = 0; 952 953 if (mlx5_link_update(dev, 0) == -EAGAIN) { 954 usleep(0); 955 return 0; 956 } 957 /* Read all message and acknowledge them. */ 958 for (;;) { 959 if (mlx5_glue->get_async_event(priv->ctx, &event)) 960 break; 961 if ((event.event_type == IBV_EVENT_PORT_ACTIVE || 962 event.event_type == IBV_EVENT_PORT_ERR) && 963 (dev->data->dev_conf.intr_conf.lsc == 1)) 964 ret |= (1 << RTE_ETH_EVENT_INTR_LSC); 965 else if (event.event_type == IBV_EVENT_DEVICE_FATAL && 966 dev->data->dev_conf.intr_conf.rmv == 1) 967 ret |= (1 << RTE_ETH_EVENT_INTR_RMV); 968 else 969 DRV_LOG(DEBUG, 970 "port %u event type %d on not handled", 971 dev->data->port_id, event.event_type); 972 mlx5_glue->ack_async_event(&event); 973 } 974 return ret; 975 } 976 977 /** 978 * Handle interrupts from the NIC. 979 * 980 * @param[in] intr_handle 981 * Interrupt handler. 982 * @param cb_arg 983 * Callback argument. 984 */ 985 void 986 mlx5_dev_interrupt_handler(void *cb_arg) 987 { 988 struct rte_eth_dev *dev = cb_arg; 989 uint32_t events; 990 991 events = mlx5_dev_status_handler(dev); 992 if (events & (1 << RTE_ETH_EVENT_INTR_LSC)) 993 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL); 994 if (events & (1 << RTE_ETH_EVENT_INTR_RMV)) 995 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL); 996 } 997 998 /** 999 * Handle interrupts from the socket. 1000 * 1001 * @param cb_arg 1002 * Callback argument. 1003 */ 1004 static void 1005 mlx5_dev_handler_socket(void *cb_arg) 1006 { 1007 struct rte_eth_dev *dev = cb_arg; 1008 1009 mlx5_socket_handle(dev); 1010 } 1011 1012 /** 1013 * Uninstall interrupt handler. 1014 * 1015 * @param dev 1016 * Pointer to Ethernet device. 1017 */ 1018 void 1019 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev) 1020 { 1021 struct priv *priv = dev->data->dev_private; 1022 1023 if (dev->data->dev_conf.intr_conf.lsc || 1024 dev->data->dev_conf.intr_conf.rmv) 1025 rte_intr_callback_unregister(&priv->intr_handle, 1026 mlx5_dev_interrupt_handler, dev); 1027 if (priv->primary_socket) 1028 rte_intr_callback_unregister(&priv->intr_handle_socket, 1029 mlx5_dev_handler_socket, dev); 1030 priv->intr_handle.fd = 0; 1031 priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN; 1032 priv->intr_handle_socket.fd = 0; 1033 priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN; 1034 } 1035 1036 /** 1037 * Install interrupt handler. 1038 * 1039 * @param dev 1040 * Pointer to Ethernet device. 1041 */ 1042 void 1043 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev) 1044 { 1045 struct priv *priv = dev->data->dev_private; 1046 int ret; 1047 int flags; 1048 1049 assert(priv->ctx->async_fd > 0); 1050 flags = fcntl(priv->ctx->async_fd, F_GETFL); 1051 ret = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK); 1052 if (ret) { 1053 DRV_LOG(INFO, 1054 "port %u failed to change file descriptor async event" 1055 " queue", 1056 dev->data->port_id); 1057 dev->data->dev_conf.intr_conf.lsc = 0; 1058 dev->data->dev_conf.intr_conf.rmv = 0; 1059 } 1060 if (dev->data->dev_conf.intr_conf.lsc || 1061 dev->data->dev_conf.intr_conf.rmv) { 1062 priv->intr_handle.fd = priv->ctx->async_fd; 1063 priv->intr_handle.type = RTE_INTR_HANDLE_EXT; 1064 rte_intr_callback_register(&priv->intr_handle, 1065 mlx5_dev_interrupt_handler, dev); 1066 } 1067 ret = mlx5_socket_init(dev); 1068 if (ret) 1069 DRV_LOG(ERR, "port %u cannot initialise socket: %s", 1070 dev->data->port_id, strerror(rte_errno)); 1071 else if (priv->primary_socket) { 1072 priv->intr_handle_socket.fd = priv->primary_socket; 1073 priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT; 1074 rte_intr_callback_register(&priv->intr_handle_socket, 1075 mlx5_dev_handler_socket, dev); 1076 } 1077 } 1078 1079 /** 1080 * DPDK callback to bring the link DOWN. 1081 * 1082 * @param dev 1083 * Pointer to Ethernet device structure. 1084 * 1085 * @return 1086 * 0 on success, a negative errno value otherwise and rte_errno is set. 1087 */ 1088 int 1089 mlx5_set_link_down(struct rte_eth_dev *dev) 1090 { 1091 return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP); 1092 } 1093 1094 /** 1095 * DPDK callback to bring the link UP. 1096 * 1097 * @param dev 1098 * Pointer to Ethernet device structure. 1099 * 1100 * @return 1101 * 0 on success, a negative errno value otherwise and rte_errno is set. 1102 */ 1103 int 1104 mlx5_set_link_up(struct rte_eth_dev *dev) 1105 { 1106 return mlx5_set_flags(dev, ~IFF_UP, IFF_UP); 1107 } 1108 1109 /** 1110 * Configure the TX function to use. 1111 * 1112 * @param dev 1113 * Pointer to private data structure. 1114 * 1115 * @return 1116 * Pointer to selected Tx burst function. 1117 */ 1118 eth_tx_burst_t 1119 mlx5_select_tx_function(struct rte_eth_dev *dev) 1120 { 1121 struct priv *priv = dev->data->dev_private; 1122 eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst; 1123 struct mlx5_dev_config *config = &priv->config; 1124 uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads; 1125 int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO | 1126 DEV_TX_OFFLOAD_VXLAN_TNL_TSO | 1127 DEV_TX_OFFLOAD_GRE_TNL_TSO | 1128 DEV_TX_OFFLOAD_IP_TNL_TSO | 1129 DEV_TX_OFFLOAD_UDP_TNL_TSO)); 1130 int swp = !!(tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO | 1131 DEV_TX_OFFLOAD_UDP_TNL_TSO | 1132 DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)); 1133 int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT); 1134 1135 assert(priv != NULL); 1136 /* Select appropriate TX function. */ 1137 if (vlan_insert || tso || swp) 1138 return tx_pkt_burst; 1139 if (config->mps == MLX5_MPW_ENHANCED) { 1140 if (mlx5_check_vec_tx_support(dev) > 0) { 1141 if (mlx5_check_raw_vec_tx_support(dev) > 0) 1142 tx_pkt_burst = mlx5_tx_burst_raw_vec; 1143 else 1144 tx_pkt_burst = mlx5_tx_burst_vec; 1145 DRV_LOG(DEBUG, 1146 "port %u selected enhanced MPW Tx vectorized" 1147 " function", 1148 dev->data->port_id); 1149 } else { 1150 tx_pkt_burst = mlx5_tx_burst_empw; 1151 DRV_LOG(DEBUG, 1152 "port %u selected enhanced MPW Tx function", 1153 dev->data->port_id); 1154 } 1155 } else if (config->mps && (config->txq_inline > 0)) { 1156 tx_pkt_burst = mlx5_tx_burst_mpw_inline; 1157 DRV_LOG(DEBUG, "port %u selected MPW inline Tx function", 1158 dev->data->port_id); 1159 } else if (config->mps) { 1160 tx_pkt_burst = mlx5_tx_burst_mpw; 1161 DRV_LOG(DEBUG, "port %u selected MPW Tx function", 1162 dev->data->port_id); 1163 } 1164 return tx_pkt_burst; 1165 } 1166 1167 /** 1168 * Configure the RX function to use. 1169 * 1170 * @param dev 1171 * Pointer to private data structure. 1172 * 1173 * @return 1174 * Pointer to selected Rx burst function. 1175 */ 1176 eth_rx_burst_t 1177 mlx5_select_rx_function(struct rte_eth_dev *dev) 1178 { 1179 eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst; 1180 1181 assert(dev != NULL); 1182 if (mlx5_check_vec_rx_support(dev) > 0) { 1183 rx_pkt_burst = mlx5_rx_burst_vec; 1184 DRV_LOG(DEBUG, "port %u selected Rx vectorized function", 1185 dev->data->port_id); 1186 } else if (mlx5_mprq_enabled(dev)) { 1187 rx_pkt_burst = mlx5_rx_burst_mprq; 1188 } 1189 return rx_pkt_burst; 1190 } 1191 1192 /** 1193 * Check if mlx5 device was removed. 1194 * 1195 * @param dev 1196 * Pointer to Ethernet device structure. 1197 * 1198 * @return 1199 * 1 when device is removed, otherwise 0. 1200 */ 1201 int 1202 mlx5_is_removed(struct rte_eth_dev *dev) 1203 { 1204 struct ibv_device_attr device_attr; 1205 struct priv *priv = dev->data->dev_private; 1206 1207 if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO) 1208 return 1; 1209 return 0; 1210 } 1211