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 /* DPDK headers don't like -pedantic. */ 54 #ifdef PEDANTIC 55 #pragma GCC diagnostic ignored "-Wpedantic" 56 #endif 57 #include <rte_atomic.h> 58 #include <rte_ethdev.h> 59 #include <rte_mbuf.h> 60 #include <rte_common.h> 61 #include <rte_interrupts.h> 62 #include <rte_alarm.h> 63 #include <rte_malloc.h> 64 #ifdef PEDANTIC 65 #pragma GCC diagnostic error "-Wpedantic" 66 #endif 67 68 #include "mlx5.h" 69 #include "mlx5_rxtx.h" 70 #include "mlx5_utils.h" 71 72 /* Add defines in case the running kernel is not the same as user headers. */ 73 #ifndef ETHTOOL_GLINKSETTINGS 74 struct ethtool_link_settings { 75 uint32_t cmd; 76 uint32_t speed; 77 uint8_t duplex; 78 uint8_t port; 79 uint8_t phy_address; 80 uint8_t autoneg; 81 uint8_t mdio_support; 82 uint8_t eth_to_mdix; 83 uint8_t eth_tp_mdix_ctrl; 84 int8_t link_mode_masks_nwords; 85 uint32_t reserved[8]; 86 uint32_t link_mode_masks[]; 87 }; 88 89 #define ETHTOOL_GLINKSETTINGS 0x0000004c 90 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5 91 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6 92 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17 93 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18 94 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19 95 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20 96 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21 97 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22 98 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23 99 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24 100 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25 101 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26 102 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27 103 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28 104 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29 105 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30 106 #endif 107 #ifndef HAVE_ETHTOOL_LINK_MODE_25G 108 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31 109 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32 110 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33 111 #endif 112 #ifndef HAVE_ETHTOOL_LINK_MODE_50G 113 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34 114 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35 115 #endif 116 #ifndef HAVE_ETHTOOL_LINK_MODE_100G 117 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36 118 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37 119 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38 120 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39 121 #endif 122 123 /** 124 * Return private structure associated with an Ethernet device. 125 * 126 * @param dev 127 * Pointer to Ethernet device structure. 128 * 129 * @return 130 * Pointer to private structure. 131 */ 132 struct priv * 133 mlx5_get_priv(struct rte_eth_dev *dev) 134 { 135 struct mlx5_secondary_data *sd; 136 137 if (!mlx5_is_secondary()) 138 return dev->data->dev_private; 139 sd = &mlx5_secondary_data[dev->data->port_id]; 140 return sd->data.dev_private; 141 } 142 143 /** 144 * Check if running as a secondary process. 145 * 146 * @return 147 * Nonzero if running as a secondary process. 148 */ 149 inline int 150 mlx5_is_secondary(void) 151 { 152 return rte_eal_process_type() != RTE_PROC_PRIMARY; 153 } 154 155 /** 156 * Get interface name from private structure. 157 * 158 * @param[in] priv 159 * Pointer to private structure. 160 * @param[out] ifname 161 * Interface name output buffer. 162 * 163 * @return 164 * 0 on success, -1 on failure and errno is set. 165 */ 166 int 167 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE]) 168 { 169 DIR *dir; 170 struct dirent *dent; 171 unsigned int dev_type = 0; 172 unsigned int dev_port_prev = ~0u; 173 char match[IF_NAMESIZE] = ""; 174 175 { 176 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path); 177 178 dir = opendir(path); 179 if (dir == NULL) 180 return -1; 181 } 182 while ((dent = readdir(dir)) != NULL) { 183 char *name = dent->d_name; 184 FILE *file; 185 unsigned int dev_port; 186 int r; 187 188 if ((name[0] == '.') && 189 ((name[1] == '\0') || 190 ((name[1] == '.') && (name[2] == '\0')))) 191 continue; 192 193 MKSTR(path, "%s/device/net/%s/%s", 194 priv->ctx->device->ibdev_path, name, 195 (dev_type ? "dev_id" : "dev_port")); 196 197 file = fopen(path, "rb"); 198 if (file == NULL) { 199 if (errno != ENOENT) 200 continue; 201 /* 202 * Switch to dev_id when dev_port does not exist as 203 * is the case with Linux kernel versions < 3.15. 204 */ 205 try_dev_id: 206 match[0] = '\0'; 207 if (dev_type) 208 break; 209 dev_type = 1; 210 dev_port_prev = ~0u; 211 rewinddir(dir); 212 continue; 213 } 214 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port); 215 fclose(file); 216 if (r != 1) 217 continue; 218 /* 219 * Switch to dev_id when dev_port returns the same value for 220 * all ports. May happen when using a MOFED release older than 221 * 3.0 with a Linux kernel >= 3.15. 222 */ 223 if (dev_port == dev_port_prev) 224 goto try_dev_id; 225 dev_port_prev = dev_port; 226 if (dev_port == (priv->port - 1u)) 227 snprintf(match, sizeof(match), "%s", name); 228 } 229 closedir(dir); 230 if (match[0] == '\0') 231 return -1; 232 strncpy(*ifname, match, sizeof(*ifname)); 233 return 0; 234 } 235 236 /** 237 * Check if the counter is located on ib counters file. 238 * 239 * @param[in] cntr 240 * Counter name. 241 * 242 * @return 243 * 1 if counter is located on ib counters file , 0 otherwise. 244 */ 245 int 246 priv_is_ib_cntr(const char *cntr) 247 { 248 if (!strcmp(cntr, "out_of_buffer")) 249 return 1; 250 return 0; 251 } 252 253 /** 254 * Read from sysfs entry. 255 * 256 * @param[in] priv 257 * Pointer to private structure. 258 * @param[in] entry 259 * Entry name relative to sysfs path. 260 * @param[out] buf 261 * Data output buffer. 262 * @param size 263 * Buffer size. 264 * 265 * @return 266 * 0 on success, -1 on failure and errno is set. 267 */ 268 static int 269 priv_sysfs_read(const struct priv *priv, const char *entry, 270 char *buf, size_t size) 271 { 272 char ifname[IF_NAMESIZE]; 273 FILE *file; 274 int ret; 275 int err; 276 277 if (priv_get_ifname(priv, &ifname)) 278 return -1; 279 280 if (priv_is_ib_cntr(entry)) { 281 MKSTR(path, "%s/ports/1/hw_counters/%s", 282 priv->ctx->device->ibdev_path, entry); 283 file = fopen(path, "rb"); 284 } else { 285 MKSTR(path, "%s/device/net/%s/%s", 286 priv->ctx->device->ibdev_path, ifname, entry); 287 file = fopen(path, "rb"); 288 } 289 if (file == NULL) 290 return -1; 291 ret = fread(buf, 1, size, file); 292 err = errno; 293 if (((size_t)ret < size) && (ferror(file))) 294 ret = -1; 295 else 296 ret = size; 297 fclose(file); 298 errno = err; 299 return ret; 300 } 301 302 /** 303 * Write to sysfs entry. 304 * 305 * @param[in] priv 306 * Pointer to private structure. 307 * @param[in] entry 308 * Entry name relative to sysfs path. 309 * @param[in] buf 310 * Data buffer. 311 * @param size 312 * Buffer size. 313 * 314 * @return 315 * 0 on success, -1 on failure and errno is set. 316 */ 317 static int 318 priv_sysfs_write(const struct priv *priv, const char *entry, 319 char *buf, size_t size) 320 { 321 char ifname[IF_NAMESIZE]; 322 FILE *file; 323 int ret; 324 int err; 325 326 if (priv_get_ifname(priv, &ifname)) 327 return -1; 328 329 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path, 330 ifname, entry); 331 332 file = fopen(path, "wb"); 333 if (file == NULL) 334 return -1; 335 ret = fwrite(buf, 1, size, file); 336 err = errno; 337 if (((size_t)ret < size) || (ferror(file))) 338 ret = -1; 339 else 340 ret = size; 341 fclose(file); 342 errno = err; 343 return ret; 344 } 345 346 /** 347 * Get unsigned long sysfs property. 348 * 349 * @param priv 350 * Pointer to private structure. 351 * @param[in] name 352 * Entry name relative to sysfs path. 353 * @param[out] value 354 * Value output buffer. 355 * 356 * @return 357 * 0 on success, -1 on failure and errno is set. 358 */ 359 static int 360 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value) 361 { 362 int ret; 363 unsigned long value_ret; 364 char value_str[32]; 365 366 ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1)); 367 if (ret == -1) { 368 DEBUG("cannot read %s value from sysfs: %s", 369 name, strerror(errno)); 370 return -1; 371 } 372 value_str[ret] = '\0'; 373 errno = 0; 374 value_ret = strtoul(value_str, NULL, 0); 375 if (errno) { 376 DEBUG("invalid %s value `%s': %s", name, value_str, 377 strerror(errno)); 378 return -1; 379 } 380 *value = value_ret; 381 return 0; 382 } 383 384 /** 385 * Set unsigned long sysfs property. 386 * 387 * @param priv 388 * Pointer to private structure. 389 * @param[in] name 390 * Entry name relative to sysfs path. 391 * @param value 392 * Value to set. 393 * 394 * @return 395 * 0 on success, -1 on failure and errno is set. 396 */ 397 static int 398 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value) 399 { 400 int ret; 401 MKSTR(value_str, "%lu", value); 402 403 ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1)); 404 if (ret == -1) { 405 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s", 406 name, value_str, value, strerror(errno)); 407 return -1; 408 } 409 return 0; 410 } 411 412 /** 413 * Perform ifreq ioctl() on associated Ethernet device. 414 * 415 * @param[in] priv 416 * Pointer to private structure. 417 * @param req 418 * Request number to pass to ioctl(). 419 * @param[out] ifr 420 * Interface request structure output buffer. 421 * 422 * @return 423 * 0 on success, -1 on failure and errno is set. 424 */ 425 int 426 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr) 427 { 428 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP); 429 int ret = -1; 430 431 if (sock == -1) 432 return ret; 433 if (priv_get_ifname(priv, &ifr->ifr_name) == 0) 434 ret = ioctl(sock, req, ifr); 435 close(sock); 436 return ret; 437 } 438 439 /** 440 * Return the number of active VFs for the current device. 441 * 442 * @param[in] priv 443 * Pointer to private structure. 444 * @param[out] num_vfs 445 * Number of active VFs. 446 * 447 * @return 448 * 0 on success, -1 on failure and errno is set. 449 */ 450 int 451 priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs) 452 { 453 /* The sysfs entry name depends on the operating system. */ 454 const char **name = (const char *[]){ 455 "device/sriov_numvfs", 456 "device/mlx5_num_vfs", 457 NULL, 458 }; 459 int ret; 460 461 do { 462 unsigned long ulong_num_vfs; 463 464 ret = priv_get_sysfs_ulong(priv, *name, &ulong_num_vfs); 465 if (!ret) 466 *num_vfs = ulong_num_vfs; 467 } while (*(++name) && ret); 468 return ret; 469 } 470 471 /** 472 * Get device MTU. 473 * 474 * @param priv 475 * Pointer to private structure. 476 * @param[out] mtu 477 * MTU value output buffer. 478 * 479 * @return 480 * 0 on success, -1 on failure and errno is set. 481 */ 482 int 483 priv_get_mtu(struct priv *priv, uint16_t *mtu) 484 { 485 unsigned long ulong_mtu; 486 487 if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1) 488 return -1; 489 *mtu = ulong_mtu; 490 return 0; 491 } 492 493 /** 494 * Read device counter from sysfs. 495 * 496 * @param priv 497 * Pointer to private structure. 498 * @param name 499 * Counter name. 500 * @param[out] cntr 501 * Counter output buffer. 502 * 503 * @return 504 * 0 on success, -1 on failure and errno is set. 505 */ 506 int 507 priv_get_cntr_sysfs(struct priv *priv, const char *name, uint64_t *cntr) 508 { 509 unsigned long ulong_ctr; 510 511 if (priv_get_sysfs_ulong(priv, name, &ulong_ctr) == -1) 512 return -1; 513 *cntr = ulong_ctr; 514 return 0; 515 } 516 517 /** 518 * Set device MTU. 519 * 520 * @param priv 521 * Pointer to private structure. 522 * @param mtu 523 * MTU value to set. 524 * 525 * @return 526 * 0 on success, -1 on failure and errno is set. 527 */ 528 static int 529 priv_set_mtu(struct priv *priv, uint16_t mtu) 530 { 531 uint16_t new_mtu; 532 533 if (priv_set_sysfs_ulong(priv, "mtu", mtu) || 534 priv_get_mtu(priv, &new_mtu)) 535 return -1; 536 if (new_mtu == mtu) 537 return 0; 538 errno = EINVAL; 539 return -1; 540 } 541 542 /** 543 * Set device flags. 544 * 545 * @param priv 546 * Pointer to private structure. 547 * @param keep 548 * Bitmask for flags that must remain untouched. 549 * @param flags 550 * Bitmask for flags to modify. 551 * 552 * @return 553 * 0 on success, -1 on failure and errno is set. 554 */ 555 int 556 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags) 557 { 558 unsigned long tmp; 559 560 if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1) 561 return -1; 562 tmp &= keep; 563 tmp |= (flags & (~keep)); 564 return priv_set_sysfs_ulong(priv, "flags", tmp); 565 } 566 567 /** 568 * Ethernet device configuration. 569 * 570 * Prepare the driver for a given number of TX and RX queues. 571 * 572 * @param dev 573 * Pointer to Ethernet device structure. 574 * 575 * @return 576 * 0 on success, errno value on failure. 577 */ 578 static int 579 dev_configure(struct rte_eth_dev *dev) 580 { 581 struct priv *priv = dev->data->dev_private; 582 unsigned int rxqs_n = dev->data->nb_rx_queues; 583 unsigned int txqs_n = dev->data->nb_tx_queues; 584 unsigned int i; 585 unsigned int j; 586 unsigned int reta_idx_n; 587 588 priv->rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf; 589 priv->rxqs = (void *)dev->data->rx_queues; 590 priv->txqs = (void *)dev->data->tx_queues; 591 if (txqs_n != priv->txqs_n) { 592 INFO("%p: TX queues number update: %u -> %u", 593 (void *)dev, priv->txqs_n, txqs_n); 594 priv->txqs_n = txqs_n; 595 } 596 if (rxqs_n > priv->ind_table_max_size) { 597 ERROR("cannot handle this many RX queues (%u)", rxqs_n); 598 return EINVAL; 599 } 600 if (rxqs_n == priv->rxqs_n) 601 return 0; 602 INFO("%p: RX queues number update: %u -> %u", 603 (void *)dev, priv->rxqs_n, rxqs_n); 604 priv->rxqs_n = rxqs_n; 605 /* If the requested number of RX queues is not a power of two, use the 606 * maximum indirection table size for better balancing. 607 * The result is always rounded to the next power of two. */ 608 reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ? 609 priv->ind_table_max_size : 610 rxqs_n)); 611 if (priv_rss_reta_index_resize(priv, reta_idx_n)) 612 return ENOMEM; 613 /* When the number of RX queues is not a power of two, the remaining 614 * table entries are padded with reused WQs and hashes are not spread 615 * uniformly. */ 616 for (i = 0, j = 0; (i != reta_idx_n); ++i) { 617 (*priv->reta_idx)[i] = j; 618 if (++j == rxqs_n) 619 j = 0; 620 } 621 return 0; 622 } 623 624 /** 625 * DPDK callback for Ethernet device configuration. 626 * 627 * @param dev 628 * Pointer to Ethernet device structure. 629 * 630 * @return 631 * 0 on success, negative errno value on failure. 632 */ 633 int 634 mlx5_dev_configure(struct rte_eth_dev *dev) 635 { 636 struct priv *priv = dev->data->dev_private; 637 int ret; 638 639 if (mlx5_is_secondary()) 640 return -E_RTE_SECONDARY; 641 642 priv_lock(priv); 643 ret = dev_configure(dev); 644 assert(ret >= 0); 645 priv_unlock(priv); 646 return -ret; 647 } 648 649 /** 650 * DPDK callback to get information about the device. 651 * 652 * @param dev 653 * Pointer to Ethernet device structure. 654 * @param[out] info 655 * Info structure output buffer. 656 */ 657 void 658 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info) 659 { 660 struct priv *priv = mlx5_get_priv(dev); 661 unsigned int max; 662 char ifname[IF_NAMESIZE]; 663 664 info->pci_dev = RTE_DEV_TO_PCI(dev->device); 665 666 priv_lock(priv); 667 /* FIXME: we should ask the device for these values. */ 668 info->min_rx_bufsize = 32; 669 info->max_rx_pktlen = 65536; 670 /* 671 * Since we need one CQ per QP, the limit is the minimum number 672 * between the two values. 673 */ 674 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ? 675 priv->device_attr.max_qp : priv->device_attr.max_cq); 676 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */ 677 if (max >= 65535) 678 max = 65535; 679 info->max_rx_queues = max; 680 info->max_tx_queues = max; 681 info->max_mac_addrs = RTE_DIM(priv->mac); 682 info->rx_offload_capa = 683 (priv->hw_csum ? 684 (DEV_RX_OFFLOAD_IPV4_CKSUM | 685 DEV_RX_OFFLOAD_UDP_CKSUM | 686 DEV_RX_OFFLOAD_TCP_CKSUM) : 687 0) | 688 (priv->hw_vlan_strip ? DEV_RX_OFFLOAD_VLAN_STRIP : 0); 689 if (!priv->mps) 690 info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT; 691 if (priv->hw_csum) 692 info->tx_offload_capa |= 693 (DEV_TX_OFFLOAD_IPV4_CKSUM | 694 DEV_TX_OFFLOAD_UDP_CKSUM | 695 DEV_TX_OFFLOAD_TCP_CKSUM); 696 if (priv->tso) 697 info->tx_offload_capa |= DEV_TX_OFFLOAD_TCP_TSO; 698 if (priv->tunnel_en) 699 info->tx_offload_capa |= (DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM | 700 DEV_TX_OFFLOAD_VXLAN_TNL_TSO | 701 DEV_TX_OFFLOAD_GRE_TNL_TSO); 702 if (priv_get_ifname(priv, &ifname) == 0) 703 info->if_index = if_nametoindex(ifname); 704 /* FIXME: RETA update/query API expects the callee to know the size of 705 * the indirection table, for this PMD the size varies depending on 706 * the number of RX queues, it becomes impossible to find the correct 707 * size if it is not fixed. 708 * The API should be updated to solve this problem. */ 709 info->reta_size = priv->ind_table_max_size; 710 info->hash_key_size = ((*priv->rss_conf) ? 711 (*priv->rss_conf)[0]->rss_key_len : 712 0); 713 info->speed_capa = priv->link_speed_capa; 714 priv_unlock(priv); 715 } 716 717 const uint32_t * 718 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev) 719 { 720 static const uint32_t ptypes[] = { 721 /* refers to rxq_cq_to_pkt_type() */ 722 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN, 723 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN, 724 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN, 725 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN, 726 RTE_PTYPE_UNKNOWN 727 728 }; 729 730 if (dev->rx_pkt_burst == mlx5_rx_burst) 731 return ptypes; 732 return NULL; 733 } 734 735 /** 736 * DPDK callback to retrieve physical link information. 737 * 738 * @param dev 739 * Pointer to Ethernet device structure. 740 * @param wait_to_complete 741 * Wait for request completion (ignored). 742 */ 743 static int 744 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete) 745 { 746 struct priv *priv = mlx5_get_priv(dev); 747 struct ethtool_cmd edata = { 748 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */ 749 }; 750 struct ifreq ifr; 751 struct rte_eth_link dev_link; 752 int link_speed = 0; 753 754 /* priv_lock() is not taken to allow concurrent calls. */ 755 756 (void)wait_to_complete; 757 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) { 758 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno)); 759 return -1; 760 } 761 memset(&dev_link, 0, sizeof(dev_link)); 762 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) && 763 (ifr.ifr_flags & IFF_RUNNING)); 764 ifr.ifr_data = (void *)&edata; 765 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) { 766 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s", 767 strerror(errno)); 768 return -1; 769 } 770 link_speed = ethtool_cmd_speed(&edata); 771 if (link_speed == -1) 772 dev_link.link_speed = 0; 773 else 774 dev_link.link_speed = link_speed; 775 priv->link_speed_capa = 0; 776 if (edata.supported & SUPPORTED_Autoneg) 777 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG; 778 if (edata.supported & (SUPPORTED_1000baseT_Full | 779 SUPPORTED_1000baseKX_Full)) 780 priv->link_speed_capa |= ETH_LINK_SPEED_1G; 781 if (edata.supported & SUPPORTED_10000baseKR_Full) 782 priv->link_speed_capa |= ETH_LINK_SPEED_10G; 783 if (edata.supported & (SUPPORTED_40000baseKR4_Full | 784 SUPPORTED_40000baseCR4_Full | 785 SUPPORTED_40000baseSR4_Full | 786 SUPPORTED_40000baseLR4_Full)) 787 priv->link_speed_capa |= ETH_LINK_SPEED_40G; 788 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ? 789 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX); 790 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds & 791 ETH_LINK_SPEED_FIXED); 792 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) { 793 /* Link status changed. */ 794 dev->data->dev_link = dev_link; 795 return 0; 796 } 797 /* Link status is still the same. */ 798 return -1; 799 } 800 801 /** 802 * Retrieve physical link information (unlocked version using new ioctl). 803 * 804 * @param dev 805 * Pointer to Ethernet device structure. 806 * @param wait_to_complete 807 * Wait for request completion (ignored). 808 */ 809 static int 810 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete) 811 { 812 struct priv *priv = mlx5_get_priv(dev); 813 struct ethtool_link_settings edata = { 814 .cmd = ETHTOOL_GLINKSETTINGS, 815 }; 816 struct ifreq ifr; 817 struct rte_eth_link dev_link; 818 uint64_t sc; 819 820 (void)wait_to_complete; 821 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) { 822 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno)); 823 return -1; 824 } 825 memset(&dev_link, 0, sizeof(dev_link)); 826 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) && 827 (ifr.ifr_flags & IFF_RUNNING)); 828 ifr.ifr_data = (void *)&edata; 829 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) { 830 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s", 831 strerror(errno)); 832 return -1; 833 } 834 dev_link.link_speed = edata.speed; 835 sc = edata.link_mode_masks[0] | 836 ((uint64_t)edata.link_mode_masks[1] << 32); 837 priv->link_speed_capa = 0; 838 if (sc & ETHTOOL_LINK_MODE_Autoneg_BIT) 839 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG; 840 if (sc & (ETHTOOL_LINK_MODE_1000baseT_Full_BIT | 841 ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)) 842 priv->link_speed_capa |= ETH_LINK_SPEED_1G; 843 if (sc & (ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT | 844 ETHTOOL_LINK_MODE_10000baseKR_Full_BIT | 845 ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)) 846 priv->link_speed_capa |= ETH_LINK_SPEED_10G; 847 if (sc & (ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT | 848 ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)) 849 priv->link_speed_capa |= ETH_LINK_SPEED_20G; 850 if (sc & (ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT | 851 ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT | 852 ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT | 853 ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)) 854 priv->link_speed_capa |= ETH_LINK_SPEED_40G; 855 if (sc & (ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT | 856 ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT | 857 ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT | 858 ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)) 859 priv->link_speed_capa |= ETH_LINK_SPEED_56G; 860 if (sc & (ETHTOOL_LINK_MODE_25000baseCR_Full_BIT | 861 ETHTOOL_LINK_MODE_25000baseKR_Full_BIT | 862 ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)) 863 priv->link_speed_capa |= ETH_LINK_SPEED_25G; 864 if (sc & (ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT | 865 ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)) 866 priv->link_speed_capa |= ETH_LINK_SPEED_50G; 867 if (sc & (ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT | 868 ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT | 869 ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT | 870 ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)) 871 priv->link_speed_capa |= ETH_LINK_SPEED_100G; 872 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ? 873 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX); 874 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds & 875 ETH_LINK_SPEED_FIXED); 876 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) { 877 /* Link status changed. */ 878 dev->data->dev_link = dev_link; 879 return 0; 880 } 881 /* Link status is still the same. */ 882 return -1; 883 } 884 885 /** 886 * DPDK callback to retrieve physical link information. 887 * 888 * @param dev 889 * Pointer to Ethernet device structure. 890 * @param wait_to_complete 891 * Wait for request completion (ignored). 892 */ 893 int 894 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete) 895 { 896 struct utsname utsname; 897 int ver[3]; 898 899 if (uname(&utsname) == -1 || 900 sscanf(utsname.release, "%d.%d.%d", 901 &ver[0], &ver[1], &ver[2]) != 3 || 902 KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0)) 903 return mlx5_link_update_unlocked_gset(dev, wait_to_complete); 904 return mlx5_link_update_unlocked_gs(dev, wait_to_complete); 905 } 906 907 /** 908 * DPDK callback to change the MTU. 909 * 910 * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be 911 * received). Use this as a hint to enable/disable scattered packets support 912 * and improve performance when not needed. 913 * Since failure is not an option, reconfiguring queues on the fly is not 914 * recommended. 915 * 916 * @param dev 917 * Pointer to Ethernet device structure. 918 * @param in_mtu 919 * New MTU. 920 * 921 * @return 922 * 0 on success, negative errno value on failure. 923 */ 924 int 925 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu) 926 { 927 struct priv *priv = dev->data->dev_private; 928 int ret = 0; 929 unsigned int i; 930 uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) = 931 mlx5_rx_burst; 932 unsigned int max_frame_len; 933 int rehash; 934 int restart = priv->started; 935 936 if (mlx5_is_secondary()) 937 return -E_RTE_SECONDARY; 938 939 priv_lock(priv); 940 /* Set kernel interface MTU first. */ 941 if (priv_set_mtu(priv, mtu)) { 942 ret = errno; 943 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu, 944 strerror(ret)); 945 goto out; 946 } else 947 DEBUG("adapter port %u MTU set to %u", priv->port, mtu); 948 /* Temporarily replace RX handler with a fake one, assuming it has not 949 * been copied elsewhere. */ 950 dev->rx_pkt_burst = removed_rx_burst; 951 /* Make sure everyone has left mlx5_rx_burst() and uses 952 * removed_rx_burst() instead. */ 953 rte_wmb(); 954 usleep(1000); 955 /* MTU does not include header and CRC. */ 956 max_frame_len = ETHER_HDR_LEN + mtu + ETHER_CRC_LEN; 957 /* Check if at least one queue is going to need a SGE update. */ 958 for (i = 0; i != priv->rxqs_n; ++i) { 959 struct rxq *rxq = (*priv->rxqs)[i]; 960 unsigned int mb_len; 961 unsigned int size = RTE_PKTMBUF_HEADROOM + max_frame_len; 962 unsigned int sges_n; 963 964 if (rxq == NULL) 965 continue; 966 mb_len = rte_pktmbuf_data_room_size(rxq->mp); 967 assert(mb_len >= RTE_PKTMBUF_HEADROOM); 968 /* 969 * Determine the number of SGEs needed for a full packet 970 * and round it to the next power of two. 971 */ 972 sges_n = log2above((size / mb_len) + !!(size % mb_len)); 973 if (sges_n != rxq->sges_n) 974 break; 975 } 976 /* 977 * If all queues have the right number of SGEs, a simple rehash 978 * of their buffers is enough, otherwise SGE information can only 979 * be updated in a queue by recreating it. All resources that depend 980 * on queues (flows, indirection tables) must be recreated as well in 981 * that case. 982 */ 983 rehash = (i == priv->rxqs_n); 984 if (!rehash) { 985 /* Clean up everything as with mlx5_dev_stop(). */ 986 priv_special_flow_disable_all(priv); 987 priv_mac_addrs_disable(priv); 988 priv_destroy_hash_rxqs(priv); 989 priv_fdir_disable(priv); 990 priv_dev_interrupt_handler_uninstall(priv, dev); 991 } 992 recover: 993 /* Reconfigure each RX queue. */ 994 for (i = 0; (i != priv->rxqs_n); ++i) { 995 struct rxq *rxq = (*priv->rxqs)[i]; 996 struct rxq_ctrl *rxq_ctrl = 997 container_of(rxq, struct rxq_ctrl, rxq); 998 int sp; 999 unsigned int mb_len; 1000 unsigned int tmp; 1001 1002 if (rxq == NULL) 1003 continue; 1004 mb_len = rte_pktmbuf_data_room_size(rxq->mp); 1005 assert(mb_len >= RTE_PKTMBUF_HEADROOM); 1006 /* Toggle scattered support (sp) if necessary. */ 1007 sp = (max_frame_len > (mb_len - RTE_PKTMBUF_HEADROOM)); 1008 /* Provide new values to rxq_setup(). */ 1009 dev->data->dev_conf.rxmode.jumbo_frame = sp; 1010 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len; 1011 if (rehash) 1012 ret = rxq_rehash(dev, rxq_ctrl); 1013 else 1014 ret = rxq_ctrl_setup(dev, rxq_ctrl, 1 << rxq->elts_n, 1015 rxq_ctrl->socket, NULL, rxq->mp); 1016 if (!ret) 1017 continue; 1018 /* Attempt to roll back in case of error. */ 1019 tmp = (mb_len << rxq->sges_n) - RTE_PKTMBUF_HEADROOM; 1020 if (max_frame_len != tmp) { 1021 max_frame_len = tmp; 1022 goto recover; 1023 } 1024 /* Double fault, disable RX. */ 1025 break; 1026 } 1027 /* 1028 * Use a safe RX burst function in case of error, otherwise mimic 1029 * mlx5_dev_start(). 1030 */ 1031 if (ret) { 1032 ERROR("unable to reconfigure RX queues, RX disabled"); 1033 rx_func = removed_rx_burst; 1034 } else if (restart && 1035 !rehash && 1036 !priv_create_hash_rxqs(priv) && 1037 !priv_rehash_flows(priv)) { 1038 if (dev->data->dev_conf.fdir_conf.mode == RTE_FDIR_MODE_NONE) 1039 priv_fdir_enable(priv); 1040 priv_dev_interrupt_handler_install(priv, dev); 1041 } 1042 priv->mtu = mtu; 1043 /* Burst functions can now be called again. */ 1044 rte_wmb(); 1045 dev->rx_pkt_burst = rx_func; 1046 out: 1047 priv_unlock(priv); 1048 assert(ret >= 0); 1049 return -ret; 1050 } 1051 1052 /** 1053 * DPDK callback to get flow control status. 1054 * 1055 * @param dev 1056 * Pointer to Ethernet device structure. 1057 * @param[out] fc_conf 1058 * Flow control output buffer. 1059 * 1060 * @return 1061 * 0 on success, negative errno value on failure. 1062 */ 1063 int 1064 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf) 1065 { 1066 struct priv *priv = dev->data->dev_private; 1067 struct ifreq ifr; 1068 struct ethtool_pauseparam ethpause = { 1069 .cmd = ETHTOOL_GPAUSEPARAM 1070 }; 1071 int ret; 1072 1073 if (mlx5_is_secondary()) 1074 return -E_RTE_SECONDARY; 1075 1076 ifr.ifr_data = (void *)ðpause; 1077 priv_lock(priv); 1078 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) { 1079 ret = errno; 1080 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)" 1081 " failed: %s", 1082 strerror(ret)); 1083 goto out; 1084 } 1085 1086 fc_conf->autoneg = ethpause.autoneg; 1087 if (ethpause.rx_pause && ethpause.tx_pause) 1088 fc_conf->mode = RTE_FC_FULL; 1089 else if (ethpause.rx_pause) 1090 fc_conf->mode = RTE_FC_RX_PAUSE; 1091 else if (ethpause.tx_pause) 1092 fc_conf->mode = RTE_FC_TX_PAUSE; 1093 else 1094 fc_conf->mode = RTE_FC_NONE; 1095 ret = 0; 1096 1097 out: 1098 priv_unlock(priv); 1099 assert(ret >= 0); 1100 return -ret; 1101 } 1102 1103 /** 1104 * DPDK callback to modify flow control parameters. 1105 * 1106 * @param dev 1107 * Pointer to Ethernet device structure. 1108 * @param[in] fc_conf 1109 * Flow control parameters. 1110 * 1111 * @return 1112 * 0 on success, negative errno value on failure. 1113 */ 1114 int 1115 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf) 1116 { 1117 struct priv *priv = dev->data->dev_private; 1118 struct ifreq ifr; 1119 struct ethtool_pauseparam ethpause = { 1120 .cmd = ETHTOOL_SPAUSEPARAM 1121 }; 1122 int ret; 1123 1124 if (mlx5_is_secondary()) 1125 return -E_RTE_SECONDARY; 1126 1127 ifr.ifr_data = (void *)ðpause; 1128 ethpause.autoneg = fc_conf->autoneg; 1129 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) || 1130 (fc_conf->mode & RTE_FC_RX_PAUSE)) 1131 ethpause.rx_pause = 1; 1132 else 1133 ethpause.rx_pause = 0; 1134 1135 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) || 1136 (fc_conf->mode & RTE_FC_TX_PAUSE)) 1137 ethpause.tx_pause = 1; 1138 else 1139 ethpause.tx_pause = 0; 1140 1141 priv_lock(priv); 1142 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) { 1143 ret = errno; 1144 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)" 1145 " failed: %s", 1146 strerror(ret)); 1147 goto out; 1148 } 1149 ret = 0; 1150 1151 out: 1152 priv_unlock(priv); 1153 assert(ret >= 0); 1154 return -ret; 1155 } 1156 1157 /** 1158 * Get PCI information from struct ibv_device. 1159 * 1160 * @param device 1161 * Pointer to Ethernet device structure. 1162 * @param[out] pci_addr 1163 * PCI bus address output buffer. 1164 * 1165 * @return 1166 * 0 on success, -1 on failure and errno is set. 1167 */ 1168 int 1169 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device, 1170 struct rte_pci_addr *pci_addr) 1171 { 1172 FILE *file; 1173 char line[32]; 1174 MKSTR(path, "%s/device/uevent", device->ibdev_path); 1175 1176 file = fopen(path, "rb"); 1177 if (file == NULL) 1178 return -1; 1179 while (fgets(line, sizeof(line), file) == line) { 1180 size_t len = strlen(line); 1181 int ret; 1182 1183 /* Truncate long lines. */ 1184 if (len == (sizeof(line) - 1)) 1185 while (line[(len - 1)] != '\n') { 1186 ret = fgetc(file); 1187 if (ret == EOF) 1188 break; 1189 line[(len - 1)] = ret; 1190 } 1191 /* Extract information. */ 1192 if (sscanf(line, 1193 "PCI_SLOT_NAME=" 1194 "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n", 1195 &pci_addr->domain, 1196 &pci_addr->bus, 1197 &pci_addr->devid, 1198 &pci_addr->function) == 4) { 1199 ret = 0; 1200 break; 1201 } 1202 } 1203 fclose(file); 1204 return 0; 1205 } 1206 1207 /** 1208 * Link status handler. 1209 * 1210 * @param priv 1211 * Pointer to private structure. 1212 * @param dev 1213 * Pointer to the rte_eth_dev structure. 1214 * 1215 * @return 1216 * Nonzero if the callback process can be called immediately. 1217 */ 1218 static int 1219 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev) 1220 { 1221 struct ibv_async_event event; 1222 struct rte_eth_link *link = &dev->data->dev_link; 1223 int ret = 0; 1224 1225 /* Read all message and acknowledge them. */ 1226 for (;;) { 1227 if (ibv_get_async_event(priv->ctx, &event)) 1228 break; 1229 1230 if (event.event_type != IBV_EVENT_PORT_ACTIVE && 1231 event.event_type != IBV_EVENT_PORT_ERR) 1232 DEBUG("event type %d on port %d not handled", 1233 event.event_type, event.element.port_num); 1234 ibv_ack_async_event(&event); 1235 } 1236 mlx5_link_update(dev, 0); 1237 if (((link->link_speed == 0) && link->link_status) || 1238 ((link->link_speed != 0) && !link->link_status)) { 1239 if (!priv->pending_alarm) { 1240 /* Inconsistent status, check again later. */ 1241 priv->pending_alarm = 1; 1242 rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US, 1243 mlx5_dev_link_status_handler, 1244 dev); 1245 } 1246 } else { 1247 ret = 1; 1248 } 1249 return ret; 1250 } 1251 1252 /** 1253 * Handle delayed link status event. 1254 * 1255 * @param arg 1256 * Registered argument. 1257 */ 1258 void 1259 mlx5_dev_link_status_handler(void *arg) 1260 { 1261 struct rte_eth_dev *dev = arg; 1262 struct priv *priv = dev->data->dev_private; 1263 int ret; 1264 1265 priv_lock(priv); 1266 assert(priv->pending_alarm == 1); 1267 priv->pending_alarm = 0; 1268 ret = priv_dev_link_status_handler(priv, dev); 1269 priv_unlock(priv); 1270 if (ret) 1271 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL); 1272 } 1273 1274 /** 1275 * Handle interrupts from the NIC. 1276 * 1277 * @param[in] intr_handle 1278 * Interrupt handler. 1279 * @param cb_arg 1280 * Callback argument. 1281 */ 1282 void 1283 mlx5_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg) 1284 { 1285 struct rte_eth_dev *dev = cb_arg; 1286 struct priv *priv = dev->data->dev_private; 1287 int ret; 1288 1289 (void)intr_handle; 1290 priv_lock(priv); 1291 ret = priv_dev_link_status_handler(priv, dev); 1292 priv_unlock(priv); 1293 if (ret) 1294 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL); 1295 } 1296 1297 /** 1298 * Uninstall interrupt handler. 1299 * 1300 * @param priv 1301 * Pointer to private structure. 1302 * @param dev 1303 * Pointer to the rte_eth_dev structure. 1304 */ 1305 void 1306 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev) 1307 { 1308 if (!dev->data->dev_conf.intr_conf.lsc) 1309 return; 1310 rte_intr_callback_unregister(&priv->intr_handle, 1311 mlx5_dev_interrupt_handler, 1312 dev); 1313 if (priv->pending_alarm) 1314 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev); 1315 priv->pending_alarm = 0; 1316 priv->intr_handle.fd = 0; 1317 priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN; 1318 } 1319 1320 /** 1321 * Install interrupt handler. 1322 * 1323 * @param priv 1324 * Pointer to private structure. 1325 * @param dev 1326 * Pointer to the rte_eth_dev structure. 1327 */ 1328 void 1329 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev) 1330 { 1331 int rc, flags; 1332 1333 if (!dev->data->dev_conf.intr_conf.lsc) 1334 return; 1335 assert(priv->ctx->async_fd > 0); 1336 flags = fcntl(priv->ctx->async_fd, F_GETFL); 1337 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK); 1338 if (rc < 0) { 1339 INFO("failed to change file descriptor async event queue"); 1340 dev->data->dev_conf.intr_conf.lsc = 0; 1341 } else { 1342 priv->intr_handle.fd = priv->ctx->async_fd; 1343 priv->intr_handle.type = RTE_INTR_HANDLE_EXT; 1344 rte_intr_callback_register(&priv->intr_handle, 1345 mlx5_dev_interrupt_handler, 1346 dev); 1347 } 1348 } 1349 1350 /** 1351 * Change the link state (UP / DOWN). 1352 * 1353 * @param priv 1354 * Pointer to Ethernet device structure. 1355 * @param up 1356 * Nonzero for link up, otherwise link down. 1357 * 1358 * @return 1359 * 0 on success, errno value on failure. 1360 */ 1361 static int 1362 priv_set_link(struct priv *priv, int up) 1363 { 1364 struct rte_eth_dev *dev = priv->dev; 1365 int err; 1366 1367 if (up) { 1368 err = priv_set_flags(priv, ~IFF_UP, IFF_UP); 1369 if (err) 1370 return err; 1371 priv_select_tx_function(priv); 1372 priv_select_rx_function(priv); 1373 } else { 1374 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP); 1375 if (err) 1376 return err; 1377 dev->rx_pkt_burst = removed_rx_burst; 1378 dev->tx_pkt_burst = removed_tx_burst; 1379 } 1380 return 0; 1381 } 1382 1383 /** 1384 * DPDK callback to bring the link DOWN. 1385 * 1386 * @param dev 1387 * Pointer to Ethernet device structure. 1388 * 1389 * @return 1390 * 0 on success, errno value on failure. 1391 */ 1392 int 1393 mlx5_set_link_down(struct rte_eth_dev *dev) 1394 { 1395 struct priv *priv = dev->data->dev_private; 1396 int err; 1397 1398 priv_lock(priv); 1399 err = priv_set_link(priv, 0); 1400 priv_unlock(priv); 1401 return err; 1402 } 1403 1404 /** 1405 * DPDK callback to bring the link UP. 1406 * 1407 * @param dev 1408 * Pointer to Ethernet device structure. 1409 * 1410 * @return 1411 * 0 on success, errno value on failure. 1412 */ 1413 int 1414 mlx5_set_link_up(struct rte_eth_dev *dev) 1415 { 1416 struct priv *priv = dev->data->dev_private; 1417 int err; 1418 1419 priv_lock(priv); 1420 err = priv_set_link(priv, 1); 1421 priv_unlock(priv); 1422 return err; 1423 } 1424 1425 /** 1426 * Configure secondary process queues from a private data pointer (primary 1427 * or secondary) and update burst callbacks. Can take place only once. 1428 * 1429 * All queues must have been previously created by the primary process to 1430 * avoid undefined behavior. 1431 * 1432 * @param priv 1433 * Private data pointer from either primary or secondary process. 1434 * 1435 * @return 1436 * Private data pointer from secondary process, NULL in case of error. 1437 */ 1438 struct priv * 1439 mlx5_secondary_data_setup(struct priv *priv) 1440 { 1441 unsigned int port_id = 0; 1442 struct mlx5_secondary_data *sd; 1443 void **tx_queues; 1444 void **rx_queues; 1445 unsigned int nb_tx_queues; 1446 unsigned int nb_rx_queues; 1447 unsigned int i; 1448 1449 /* priv must be valid at this point. */ 1450 assert(priv != NULL); 1451 /* priv->dev must also be valid but may point to local memory from 1452 * another process, possibly with the same address and must not 1453 * be dereferenced yet. */ 1454 assert(priv->dev != NULL); 1455 /* Determine port ID by finding out where priv comes from. */ 1456 while (1) { 1457 sd = &mlx5_secondary_data[port_id]; 1458 rte_spinlock_lock(&sd->lock); 1459 /* Primary process? */ 1460 if (sd->primary_priv == priv) 1461 break; 1462 /* Secondary process? */ 1463 if (sd->data.dev_private == priv) 1464 break; 1465 rte_spinlock_unlock(&sd->lock); 1466 if (++port_id == RTE_DIM(mlx5_secondary_data)) 1467 port_id = 0; 1468 } 1469 /* Switch to secondary private structure. If private data has already 1470 * been updated by another thread, there is nothing else to do. */ 1471 priv = sd->data.dev_private; 1472 if (priv->dev->data == &sd->data) 1473 goto end; 1474 /* Sanity checks. Secondary private structure is supposed to point 1475 * to local eth_dev, itself still pointing to the shared device data 1476 * structure allocated by the primary process. */ 1477 assert(sd->shared_dev_data != &sd->data); 1478 assert(sd->data.nb_tx_queues == 0); 1479 assert(sd->data.tx_queues == NULL); 1480 assert(sd->data.nb_rx_queues == 0); 1481 assert(sd->data.rx_queues == NULL); 1482 assert(priv != sd->primary_priv); 1483 assert(priv->dev->data == sd->shared_dev_data); 1484 assert(priv->txqs_n == 0); 1485 assert(priv->txqs == NULL); 1486 assert(priv->rxqs_n == 0); 1487 assert(priv->rxqs == NULL); 1488 nb_tx_queues = sd->shared_dev_data->nb_tx_queues; 1489 nb_rx_queues = sd->shared_dev_data->nb_rx_queues; 1490 /* Allocate local storage for queues. */ 1491 tx_queues = rte_zmalloc("secondary ethdev->tx_queues", 1492 sizeof(sd->data.tx_queues[0]) * nb_tx_queues, 1493 RTE_CACHE_LINE_SIZE); 1494 rx_queues = rte_zmalloc("secondary ethdev->rx_queues", 1495 sizeof(sd->data.rx_queues[0]) * nb_rx_queues, 1496 RTE_CACHE_LINE_SIZE); 1497 if (tx_queues == NULL || rx_queues == NULL) 1498 goto error; 1499 /* Lock to prevent control operations during setup. */ 1500 priv_lock(priv); 1501 /* TX queues. */ 1502 for (i = 0; i != nb_tx_queues; ++i) { 1503 struct txq *primary_txq = (*sd->primary_priv->txqs)[i]; 1504 struct txq_ctrl *primary_txq_ctrl; 1505 struct txq_ctrl *txq_ctrl; 1506 1507 if (primary_txq == NULL) 1508 continue; 1509 primary_txq_ctrl = container_of(primary_txq, 1510 struct txq_ctrl, txq); 1511 txq_ctrl = rte_calloc_socket("TXQ", 1, sizeof(*txq_ctrl) + 1512 (1 << primary_txq->elts_n) * 1513 sizeof(struct rte_mbuf *), 0, 1514 primary_txq_ctrl->socket); 1515 if (txq_ctrl != NULL) { 1516 if (txq_ctrl_setup(priv->dev, 1517 txq_ctrl, 1518 1 << primary_txq->elts_n, 1519 primary_txq_ctrl->socket, 1520 NULL) == 0) { 1521 txq_ctrl->txq.stats.idx = 1522 primary_txq->stats.idx; 1523 tx_queues[i] = &txq_ctrl->txq; 1524 continue; 1525 } 1526 rte_free(txq_ctrl); 1527 } 1528 while (i) { 1529 txq_ctrl = tx_queues[--i]; 1530 txq_cleanup(txq_ctrl); 1531 rte_free(txq_ctrl); 1532 } 1533 goto error; 1534 } 1535 /* RX queues. */ 1536 for (i = 0; i != nb_rx_queues; ++i) { 1537 struct rxq_ctrl *primary_rxq = 1538 container_of((*sd->primary_priv->rxqs)[i], 1539 struct rxq_ctrl, rxq); 1540 1541 if (primary_rxq == NULL) 1542 continue; 1543 /* Not supported yet. */ 1544 rx_queues[i] = NULL; 1545 } 1546 /* Update everything. */ 1547 priv->txqs = (void *)tx_queues; 1548 priv->txqs_n = nb_tx_queues; 1549 priv->rxqs = (void *)rx_queues; 1550 priv->rxqs_n = nb_rx_queues; 1551 sd->data.rx_queues = rx_queues; 1552 sd->data.tx_queues = tx_queues; 1553 sd->data.nb_rx_queues = nb_rx_queues; 1554 sd->data.nb_tx_queues = nb_tx_queues; 1555 sd->data.dev_link = sd->shared_dev_data->dev_link; 1556 sd->data.mtu = sd->shared_dev_data->mtu; 1557 memcpy(sd->data.rx_queue_state, sd->shared_dev_data->rx_queue_state, 1558 sizeof(sd->data.rx_queue_state)); 1559 memcpy(sd->data.tx_queue_state, sd->shared_dev_data->tx_queue_state, 1560 sizeof(sd->data.tx_queue_state)); 1561 sd->data.dev_flags = sd->shared_dev_data->dev_flags; 1562 /* Use local data from now on. */ 1563 rte_mb(); 1564 priv->dev->data = &sd->data; 1565 rte_mb(); 1566 priv_select_tx_function(priv); 1567 priv_select_rx_function(priv); 1568 priv_unlock(priv); 1569 end: 1570 /* More sanity checks. */ 1571 assert(priv->dev->data == &sd->data); 1572 rte_spinlock_unlock(&sd->lock); 1573 return priv; 1574 error: 1575 priv_unlock(priv); 1576 rte_free(tx_queues); 1577 rte_free(rx_queues); 1578 rte_spinlock_unlock(&sd->lock); 1579 return NULL; 1580 } 1581 1582 /** 1583 * Configure the TX function to use. 1584 * 1585 * @param priv 1586 * Pointer to private structure. 1587 */ 1588 void 1589 priv_select_tx_function(struct priv *priv) 1590 { 1591 priv->dev->tx_pkt_burst = mlx5_tx_burst; 1592 /* Select appropriate TX function. */ 1593 if (priv->mps == MLX5_MPW_ENHANCED) { 1594 priv->dev->tx_pkt_burst = 1595 mlx5_tx_burst_empw; 1596 DEBUG("selected Enhanced MPW TX function"); 1597 } else if (priv->mps && priv->txq_inline) { 1598 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw_inline; 1599 DEBUG("selected MPW inline TX function"); 1600 } else if (priv->mps) { 1601 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw; 1602 DEBUG("selected MPW TX function"); 1603 } 1604 } 1605 1606 /** 1607 * Configure the RX function to use. 1608 * 1609 * @param priv 1610 * Pointer to private structure. 1611 */ 1612 void 1613 priv_select_rx_function(struct priv *priv) 1614 { 1615 priv->dev->rx_pkt_burst = mlx5_rx_burst; 1616 } 1617