1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright 2012 6WIND S.A. 3 * Copyright 2012 Mellanox Technologies, Ltd 4 */ 5 6 /** 7 * @file 8 * mlx4 driver initialization. 9 */ 10 11 #include <errno.h> 12 #include <inttypes.h> 13 #include <stddef.h> 14 #include <stdint.h> 15 #include <stdio.h> 16 #include <stdlib.h> 17 #include <string.h> 18 #include <sys/mman.h> 19 #include <unistd.h> 20 #ifdef RTE_IBVERBS_LINK_DLOPEN 21 #include <dlfcn.h> 22 #endif 23 24 /* Verbs headers do not support -pedantic. */ 25 #ifdef PEDANTIC 26 #pragma GCC diagnostic ignored "-Wpedantic" 27 #endif 28 #include <infiniband/verbs.h> 29 #ifdef PEDANTIC 30 #pragma GCC diagnostic error "-Wpedantic" 31 #endif 32 33 #include <rte_common.h> 34 #include <rte_dev.h> 35 #include <rte_errno.h> 36 #include <rte_ethdev_driver.h> 37 #include <rte_ethdev_pci.h> 38 #include <rte_ether.h> 39 #include <rte_flow.h> 40 #include <rte_interrupts.h> 41 #include <rte_kvargs.h> 42 #include <rte_malloc.h> 43 #include <rte_mbuf.h> 44 45 #include "mlx4.h" 46 #include "mlx4_glue.h" 47 #include "mlx4_flow.h" 48 #include "mlx4_mr.h" 49 #include "mlx4_rxtx.h" 50 #include "mlx4_utils.h" 51 52 #ifdef MLX4_GLUE 53 const struct mlx4_glue *mlx4_glue; 54 #endif 55 56 static const char *MZ_MLX4_PMD_SHARED_DATA = "mlx4_pmd_shared_data"; 57 58 /* Shared memory between primary and secondary processes. */ 59 struct mlx4_shared_data *mlx4_shared_data; 60 61 /* Spinlock for mlx4_shared_data allocation. */ 62 static rte_spinlock_t mlx4_shared_data_lock = RTE_SPINLOCK_INITIALIZER; 63 64 /* Process local data for secondary processes. */ 65 static struct mlx4_local_data mlx4_local_data; 66 67 /** Configuration structure for device arguments. */ 68 struct mlx4_conf { 69 struct { 70 uint32_t present; /**< Bit-field for existing ports. */ 71 uint32_t enabled; /**< Bit-field for user-enabled ports. */ 72 } ports; 73 int mr_ext_memseg_en; 74 /** Whether memseg should be extended for MR creation. */ 75 }; 76 77 /* Available parameters list. */ 78 const char *pmd_mlx4_init_params[] = { 79 MLX4_PMD_PORT_KVARG, 80 MLX4_MR_EXT_MEMSEG_EN_KVARG, 81 NULL, 82 }; 83 84 static int mlx4_dev_stop(struct rte_eth_dev *dev); 85 86 /** 87 * Initialize shared data between primary and secondary process. 88 * 89 * A memzone is reserved by primary process and secondary processes attach to 90 * the memzone. 91 * 92 * @return 93 * 0 on success, a negative errno value otherwise and rte_errno is set. 94 */ 95 static int 96 mlx4_init_shared_data(void) 97 { 98 const struct rte_memzone *mz; 99 int ret = 0; 100 101 rte_spinlock_lock(&mlx4_shared_data_lock); 102 if (mlx4_shared_data == NULL) { 103 if (rte_eal_process_type() == RTE_PROC_PRIMARY) { 104 /* Allocate shared memory. */ 105 mz = rte_memzone_reserve(MZ_MLX4_PMD_SHARED_DATA, 106 sizeof(*mlx4_shared_data), 107 SOCKET_ID_ANY, 0); 108 if (mz == NULL) { 109 ERROR("Cannot allocate mlx4 shared data\n"); 110 ret = -rte_errno; 111 goto error; 112 } 113 mlx4_shared_data = mz->addr; 114 memset(mlx4_shared_data, 0, sizeof(*mlx4_shared_data)); 115 rte_spinlock_init(&mlx4_shared_data->lock); 116 } else { 117 /* Lookup allocated shared memory. */ 118 mz = rte_memzone_lookup(MZ_MLX4_PMD_SHARED_DATA); 119 if (mz == NULL) { 120 ERROR("Cannot attach mlx4 shared data\n"); 121 ret = -rte_errno; 122 goto error; 123 } 124 mlx4_shared_data = mz->addr; 125 memset(&mlx4_local_data, 0, sizeof(mlx4_local_data)); 126 } 127 } 128 error: 129 rte_spinlock_unlock(&mlx4_shared_data_lock); 130 return ret; 131 } 132 133 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS 134 /** 135 * Verbs callback to allocate a memory. This function should allocate the space 136 * according to the size provided residing inside a huge page. 137 * Please note that all allocation must respect the alignment from libmlx4 138 * (i.e. currently sysconf(_SC_PAGESIZE)). 139 * 140 * @param[in] size 141 * The size in bytes of the memory to allocate. 142 * @param[in] data 143 * A pointer to the callback data. 144 * 145 * @return 146 * Allocated buffer, NULL otherwise and rte_errno is set. 147 */ 148 static void * 149 mlx4_alloc_verbs_buf(size_t size, void *data) 150 { 151 struct mlx4_priv *priv = data; 152 void *ret; 153 size_t alignment = sysconf(_SC_PAGESIZE); 154 unsigned int socket = SOCKET_ID_ANY; 155 156 if (priv->verbs_alloc_ctx.type == MLX4_VERBS_ALLOC_TYPE_TX_QUEUE) { 157 const struct txq *txq = priv->verbs_alloc_ctx.obj; 158 159 socket = txq->socket; 160 } else if (priv->verbs_alloc_ctx.type == 161 MLX4_VERBS_ALLOC_TYPE_RX_QUEUE) { 162 const struct rxq *rxq = priv->verbs_alloc_ctx.obj; 163 164 socket = rxq->socket; 165 } 166 MLX4_ASSERT(data != NULL); 167 ret = rte_malloc_socket(__func__, size, alignment, socket); 168 if (!ret && size) 169 rte_errno = ENOMEM; 170 return ret; 171 } 172 173 /** 174 * Verbs callback to free a memory. 175 * 176 * @param[in] ptr 177 * A pointer to the memory to free. 178 * @param[in] data 179 * A pointer to the callback data. 180 */ 181 static void 182 mlx4_free_verbs_buf(void *ptr, void *data __rte_unused) 183 { 184 MLX4_ASSERT(data != NULL); 185 rte_free(ptr); 186 } 187 #endif 188 189 /** 190 * Initialize process private data structure. 191 * 192 * @param dev 193 * Pointer to Ethernet device structure. 194 * 195 * @return 196 * 0 on success, a negative errno value otherwise and rte_errno is set. 197 */ 198 static int 199 mlx4_proc_priv_init(struct rte_eth_dev *dev) 200 { 201 struct mlx4_proc_priv *ppriv; 202 size_t ppriv_size; 203 204 /* 205 * UAR register table follows the process private structure. BlueFlame 206 * registers for Tx queues are stored in the table. 207 */ 208 ppriv_size = sizeof(struct mlx4_proc_priv) + 209 dev->data->nb_tx_queues * sizeof(void *); 210 ppriv = rte_malloc_socket("mlx4_proc_priv", ppriv_size, 211 RTE_CACHE_LINE_SIZE, dev->device->numa_node); 212 if (!ppriv) { 213 rte_errno = ENOMEM; 214 return -rte_errno; 215 } 216 ppriv->uar_table_sz = ppriv_size; 217 dev->process_private = ppriv; 218 return 0; 219 } 220 221 /** 222 * Un-initialize process private data structure. 223 * 224 * @param dev 225 * Pointer to Ethernet device structure. 226 */ 227 static void 228 mlx4_proc_priv_uninit(struct rte_eth_dev *dev) 229 { 230 if (!dev->process_private) 231 return; 232 rte_free(dev->process_private); 233 dev->process_private = NULL; 234 } 235 236 /** 237 * DPDK callback for Ethernet device configuration. 238 * 239 * @param dev 240 * Pointer to Ethernet device structure. 241 * 242 * @return 243 * 0 on success, negative errno value otherwise and rte_errno is set. 244 */ 245 static int 246 mlx4_dev_configure(struct rte_eth_dev *dev) 247 { 248 struct mlx4_priv *priv = dev->data->dev_private; 249 struct rte_flow_error error; 250 int ret; 251 252 if (dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG) 253 dev->data->dev_conf.rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH; 254 255 /* Prepare internal flow rules. */ 256 ret = mlx4_flow_sync(priv, &error); 257 if (ret) { 258 ERROR("cannot set up internal flow rules (code %d, \"%s\")," 259 " flow error type %d, cause %p, message: %s", 260 -ret, strerror(-ret), error.type, error.cause, 261 error.message ? error.message : "(unspecified)"); 262 goto exit; 263 } 264 ret = mlx4_intr_install(priv); 265 if (ret) { 266 ERROR("%p: interrupt handler installation failed", 267 (void *)dev); 268 goto exit; 269 } 270 ret = mlx4_proc_priv_init(dev); 271 if (ret) { 272 ERROR("%p: process private data allocation failed", 273 (void *)dev); 274 goto exit; 275 } 276 exit: 277 return ret; 278 } 279 280 /** 281 * DPDK callback to start the device. 282 * 283 * Simulate device start by initializing common RSS resources and attaching 284 * all configured flows. 285 * 286 * @param dev 287 * Pointer to Ethernet device structure. 288 * 289 * @return 290 * 0 on success, negative errno value otherwise and rte_errno is set. 291 */ 292 static int 293 mlx4_dev_start(struct rte_eth_dev *dev) 294 { 295 struct mlx4_priv *priv = dev->data->dev_private; 296 struct rte_flow_error error; 297 int ret; 298 299 if (priv->started) 300 return 0; 301 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev); 302 priv->started = 1; 303 ret = mlx4_rss_init(priv); 304 if (ret) { 305 ERROR("%p: cannot initialize RSS resources: %s", 306 (void *)dev, strerror(-ret)); 307 goto err; 308 } 309 #ifdef RTE_LIBRTE_MLX4_DEBUG 310 mlx4_mr_dump_dev(dev); 311 #endif 312 ret = mlx4_rxq_intr_enable(priv); 313 if (ret) { 314 ERROR("%p: interrupt handler installation failed", 315 (void *)dev); 316 goto err; 317 } 318 ret = mlx4_flow_sync(priv, &error); 319 if (ret) { 320 ERROR("%p: cannot attach flow rules (code %d, \"%s\")," 321 " flow error type %d, cause %p, message: %s", 322 (void *)dev, 323 -ret, strerror(-ret), error.type, error.cause, 324 error.message ? error.message : "(unspecified)"); 325 goto err; 326 } 327 rte_wmb(); 328 dev->tx_pkt_burst = mlx4_tx_burst; 329 dev->rx_pkt_burst = mlx4_rx_burst; 330 /* Enable datapath on secondary process. */ 331 mlx4_mp_req_start_rxtx(dev); 332 return 0; 333 err: 334 mlx4_dev_stop(dev); 335 return ret; 336 } 337 338 /** 339 * DPDK callback to stop the device. 340 * 341 * Simulate device stop by detaching all configured flows. 342 * 343 * @param dev 344 * Pointer to Ethernet device structure. 345 */ 346 static int 347 mlx4_dev_stop(struct rte_eth_dev *dev) 348 { 349 struct mlx4_priv *priv = dev->data->dev_private; 350 351 if (!priv->started) 352 return 0; 353 DEBUG("%p: detaching flows from all RX queues", (void *)dev); 354 priv->started = 0; 355 dev->tx_pkt_burst = mlx4_tx_burst_removed; 356 dev->rx_pkt_burst = mlx4_rx_burst_removed; 357 rte_wmb(); 358 /* Disable datapath on secondary process. */ 359 mlx4_mp_req_stop_rxtx(dev); 360 mlx4_flow_sync(priv, NULL); 361 mlx4_rxq_intr_disable(priv); 362 mlx4_rss_deinit(priv); 363 364 return 0; 365 } 366 367 /** 368 * DPDK callback to close the device. 369 * 370 * Destroy all queues and objects, free memory. 371 * 372 * @param dev 373 * Pointer to Ethernet device structure. 374 */ 375 static int 376 mlx4_dev_close(struct rte_eth_dev *dev) 377 { 378 struct mlx4_priv *priv = dev->data->dev_private; 379 unsigned int i; 380 381 if (rte_eal_process_type() != RTE_PROC_PRIMARY) 382 return 0; 383 DEBUG("%p: closing device \"%s\"", 384 (void *)dev, 385 ((priv->ctx != NULL) ? priv->ctx->device->name : "")); 386 dev->rx_pkt_burst = mlx4_rx_burst_removed; 387 dev->tx_pkt_burst = mlx4_tx_burst_removed; 388 rte_wmb(); 389 /* Disable datapath on secondary process. */ 390 mlx4_mp_req_stop_rxtx(dev); 391 mlx4_flow_clean(priv); 392 mlx4_rss_deinit(priv); 393 for (i = 0; i != dev->data->nb_rx_queues; ++i) 394 mlx4_rx_queue_release(dev->data->rx_queues[i]); 395 for (i = 0; i != dev->data->nb_tx_queues; ++i) 396 mlx4_tx_queue_release(dev->data->tx_queues[i]); 397 mlx4_proc_priv_uninit(dev); 398 mlx4_mr_release(dev); 399 if (priv->pd != NULL) { 400 MLX4_ASSERT(priv->ctx != NULL); 401 claim_zero(mlx4_glue->dealloc_pd(priv->pd)); 402 claim_zero(mlx4_glue->close_device(priv->ctx)); 403 } else 404 MLX4_ASSERT(priv->ctx == NULL); 405 mlx4_intr_uninstall(priv); 406 memset(priv, 0, sizeof(*priv)); 407 /* mac_addrs must not be freed because part of dev_private */ 408 dev->data->mac_addrs = NULL; 409 return 0; 410 } 411 412 static const struct eth_dev_ops mlx4_dev_ops = { 413 .dev_configure = mlx4_dev_configure, 414 .dev_start = mlx4_dev_start, 415 .dev_stop = mlx4_dev_stop, 416 .dev_set_link_down = mlx4_dev_set_link_down, 417 .dev_set_link_up = mlx4_dev_set_link_up, 418 .dev_close = mlx4_dev_close, 419 .link_update = mlx4_link_update, 420 .promiscuous_enable = mlx4_promiscuous_enable, 421 .promiscuous_disable = mlx4_promiscuous_disable, 422 .allmulticast_enable = mlx4_allmulticast_enable, 423 .allmulticast_disable = mlx4_allmulticast_disable, 424 .mac_addr_remove = mlx4_mac_addr_remove, 425 .mac_addr_add = mlx4_mac_addr_add, 426 .mac_addr_set = mlx4_mac_addr_set, 427 .set_mc_addr_list = mlx4_set_mc_addr_list, 428 .stats_get = mlx4_stats_get, 429 .stats_reset = mlx4_stats_reset, 430 .fw_version_get = mlx4_fw_version_get, 431 .dev_infos_get = mlx4_dev_infos_get, 432 .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get, 433 .vlan_filter_set = mlx4_vlan_filter_set, 434 .rx_queue_setup = mlx4_rx_queue_setup, 435 .tx_queue_setup = mlx4_tx_queue_setup, 436 .rx_queue_release = mlx4_rx_queue_release, 437 .tx_queue_release = mlx4_tx_queue_release, 438 .flow_ctrl_get = mlx4_flow_ctrl_get, 439 .flow_ctrl_set = mlx4_flow_ctrl_set, 440 .mtu_set = mlx4_mtu_set, 441 .filter_ctrl = mlx4_filter_ctrl, 442 .rx_queue_intr_enable = mlx4_rx_intr_enable, 443 .rx_queue_intr_disable = mlx4_rx_intr_disable, 444 .is_removed = mlx4_is_removed, 445 }; 446 447 /* Available operations from secondary process. */ 448 static const struct eth_dev_ops mlx4_dev_sec_ops = { 449 .stats_get = mlx4_stats_get, 450 .stats_reset = mlx4_stats_reset, 451 .fw_version_get = mlx4_fw_version_get, 452 .dev_infos_get = mlx4_dev_infos_get, 453 }; 454 455 /** 456 * Get PCI information from struct ibv_device. 457 * 458 * @param device 459 * Pointer to Ethernet device structure. 460 * @param[out] pci_addr 461 * PCI bus address output buffer. 462 * 463 * @return 464 * 0 on success, negative errno value otherwise and rte_errno is set. 465 */ 466 static int 467 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device, 468 struct rte_pci_addr *pci_addr) 469 { 470 FILE *file; 471 char line[32]; 472 MKSTR(path, "%s/device/uevent", device->ibdev_path); 473 474 file = fopen(path, "rb"); 475 if (file == NULL) { 476 rte_errno = errno; 477 return -rte_errno; 478 } 479 while (fgets(line, sizeof(line), file) == line) { 480 size_t len = strlen(line); 481 int ret; 482 483 /* Truncate long lines. */ 484 if (len == (sizeof(line) - 1)) 485 while (line[(len - 1)] != '\n') { 486 ret = fgetc(file); 487 if (ret == EOF) 488 break; 489 line[(len - 1)] = ret; 490 } 491 /* Extract information. */ 492 if (sscanf(line, 493 "PCI_SLOT_NAME=" 494 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n", 495 &pci_addr->domain, 496 &pci_addr->bus, 497 &pci_addr->devid, 498 &pci_addr->function) == 4) { 499 break; 500 } 501 } 502 fclose(file); 503 return 0; 504 } 505 506 /** 507 * Verify and store value for device argument. 508 * 509 * @param[in] key 510 * Key argument to verify. 511 * @param[in] val 512 * Value associated with key. 513 * @param[in, out] conf 514 * Shared configuration data. 515 * 516 * @return 517 * 0 on success, negative errno value otherwise and rte_errno is set. 518 */ 519 static int 520 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf) 521 { 522 unsigned long tmp; 523 524 errno = 0; 525 tmp = strtoul(val, NULL, 0); 526 if (errno) { 527 rte_errno = errno; 528 WARN("%s: \"%s\" is not a valid integer", key, val); 529 return -rte_errno; 530 } 531 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) { 532 uint32_t ports = rte_log2_u32(conf->ports.present + 1); 533 534 if (tmp >= ports) { 535 ERROR("port index %lu outside range [0,%" PRIu32 ")", 536 tmp, ports); 537 return -EINVAL; 538 } 539 if (!(conf->ports.present & (1 << tmp))) { 540 rte_errno = EINVAL; 541 ERROR("invalid port index %lu", tmp); 542 return -rte_errno; 543 } 544 conf->ports.enabled |= 1 << tmp; 545 } else if (strcmp(MLX4_MR_EXT_MEMSEG_EN_KVARG, key) == 0) { 546 conf->mr_ext_memseg_en = !!tmp; 547 } else { 548 rte_errno = EINVAL; 549 WARN("%s: unknown parameter", key); 550 return -rte_errno; 551 } 552 return 0; 553 } 554 555 /** 556 * Parse device parameters. 557 * 558 * @param devargs 559 * Device arguments structure. 560 * 561 * @return 562 * 0 on success, negative errno value otherwise and rte_errno is set. 563 */ 564 static int 565 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf) 566 { 567 struct rte_kvargs *kvlist; 568 unsigned int arg_count; 569 int ret = 0; 570 int i; 571 572 if (devargs == NULL) 573 return 0; 574 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params); 575 if (kvlist == NULL) { 576 rte_errno = EINVAL; 577 ERROR("failed to parse kvargs"); 578 return -rte_errno; 579 } 580 /* Process parameters. */ 581 for (i = 0; pmd_mlx4_init_params[i]; ++i) { 582 arg_count = rte_kvargs_count(kvlist, pmd_mlx4_init_params[i]); 583 while (arg_count-- > 0) { 584 ret = rte_kvargs_process(kvlist, 585 pmd_mlx4_init_params[i], 586 (int (*)(const char *, 587 const char *, 588 void *)) 589 mlx4_arg_parse, 590 conf); 591 if (ret != 0) 592 goto free_kvlist; 593 } 594 } 595 free_kvlist: 596 rte_kvargs_free(kvlist); 597 return ret; 598 } 599 600 /** 601 * Interpret RSS capabilities reported by device. 602 * 603 * This function returns the set of usable Verbs RSS hash fields, kernel 604 * quirks taken into account. 605 * 606 * @param ctx 607 * Verbs context. 608 * @param pd 609 * Verbs protection domain. 610 * @param device_attr_ex 611 * Extended device attributes to interpret. 612 * 613 * @return 614 * Usable RSS hash fields mask in Verbs format. 615 */ 616 static uint64_t 617 mlx4_hw_rss_sup(struct ibv_context *ctx, struct ibv_pd *pd, 618 struct ibv_device_attr_ex *device_attr_ex) 619 { 620 uint64_t hw_rss_sup = device_attr_ex->rss_caps.rx_hash_fields_mask; 621 struct ibv_cq *cq = NULL; 622 struct ibv_wq *wq = NULL; 623 struct ibv_rwq_ind_table *ind = NULL; 624 struct ibv_qp *qp = NULL; 625 626 if (!hw_rss_sup) { 627 WARN("no RSS capabilities reported; disabling support for UDP" 628 " RSS and inner VXLAN RSS"); 629 return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 | 630 IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 | 631 IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP; 632 } 633 if (!(hw_rss_sup & IBV_RX_HASH_INNER)) 634 return hw_rss_sup; 635 /* 636 * Although reported as supported, missing code in some Linux 637 * versions (v4.15, v4.16) prevents the creation of hash QPs with 638 * inner capability. 639 * 640 * There is no choice but to attempt to instantiate a temporary RSS 641 * context in order to confirm its support. 642 */ 643 cq = mlx4_glue->create_cq(ctx, 1, NULL, NULL, 0); 644 wq = cq ? mlx4_glue->create_wq 645 (ctx, 646 &(struct ibv_wq_init_attr){ 647 .wq_type = IBV_WQT_RQ, 648 .max_wr = 1, 649 .max_sge = 1, 650 .pd = pd, 651 .cq = cq, 652 }) : NULL; 653 ind = wq ? mlx4_glue->create_rwq_ind_table 654 (ctx, 655 &(struct ibv_rwq_ind_table_init_attr){ 656 .log_ind_tbl_size = 0, 657 .ind_tbl = &wq, 658 .comp_mask = 0, 659 }) : NULL; 660 qp = ind ? mlx4_glue->create_qp_ex 661 (ctx, 662 &(struct ibv_qp_init_attr_ex){ 663 .comp_mask = 664 (IBV_QP_INIT_ATTR_PD | 665 IBV_QP_INIT_ATTR_RX_HASH | 666 IBV_QP_INIT_ATTR_IND_TABLE), 667 .qp_type = IBV_QPT_RAW_PACKET, 668 .pd = pd, 669 .rwq_ind_tbl = ind, 670 .rx_hash_conf = { 671 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ, 672 .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE, 673 .rx_hash_key = mlx4_rss_hash_key_default, 674 .rx_hash_fields_mask = hw_rss_sup, 675 }, 676 }) : NULL; 677 if (!qp) { 678 WARN("disabling unusable inner RSS capability due to kernel" 679 " quirk"); 680 hw_rss_sup &= ~IBV_RX_HASH_INNER; 681 } else { 682 claim_zero(mlx4_glue->destroy_qp(qp)); 683 } 684 if (ind) 685 claim_zero(mlx4_glue->destroy_rwq_ind_table(ind)); 686 if (wq) 687 claim_zero(mlx4_glue->destroy_wq(wq)); 688 if (cq) 689 claim_zero(mlx4_glue->destroy_cq(cq)); 690 return hw_rss_sup; 691 } 692 693 static struct rte_pci_driver mlx4_driver; 694 695 /** 696 * PMD global initialization. 697 * 698 * Independent from individual device, this function initializes global 699 * per-PMD data structures distinguishing primary and secondary processes. 700 * Hence, each initialization is called once per a process. 701 * 702 * @return 703 * 0 on success, a negative errno value otherwise and rte_errno is set. 704 */ 705 static int 706 mlx4_init_once(void) 707 { 708 struct mlx4_shared_data *sd; 709 struct mlx4_local_data *ld = &mlx4_local_data; 710 int ret = 0; 711 712 if (mlx4_init_shared_data()) 713 return -rte_errno; 714 sd = mlx4_shared_data; 715 MLX4_ASSERT(sd); 716 rte_spinlock_lock(&sd->lock); 717 switch (rte_eal_process_type()) { 718 case RTE_PROC_PRIMARY: 719 if (sd->init_done) 720 break; 721 LIST_INIT(&sd->mem_event_cb_list); 722 rte_rwlock_init(&sd->mem_event_rwlock); 723 rte_mem_event_callback_register("MLX4_MEM_EVENT_CB", 724 mlx4_mr_mem_event_cb, NULL); 725 ret = mlx4_mp_init_primary(); 726 if (ret) 727 goto out; 728 sd->init_done = 1; 729 break; 730 case RTE_PROC_SECONDARY: 731 if (ld->init_done) 732 break; 733 ret = mlx4_mp_init_secondary(); 734 if (ret) 735 goto out; 736 ++sd->secondary_cnt; 737 ld->init_done = 1; 738 break; 739 default: 740 break; 741 } 742 out: 743 rte_spinlock_unlock(&sd->lock); 744 return ret; 745 } 746 747 /** 748 * DPDK callback to register a PCI device. 749 * 750 * This function creates an Ethernet device for each port of a given 751 * PCI device. 752 * 753 * @param[in] pci_drv 754 * PCI driver structure (mlx4_driver). 755 * @param[in] pci_dev 756 * PCI device information. 757 * 758 * @return 759 * 0 on success, negative errno value otherwise and rte_errno is set. 760 */ 761 static int 762 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev) 763 { 764 struct ibv_device **list; 765 struct ibv_device *ibv_dev; 766 int err = 0; 767 struct ibv_context *attr_ctx = NULL; 768 struct ibv_device_attr device_attr; 769 struct ibv_device_attr_ex device_attr_ex; 770 struct mlx4_conf conf = { 771 .ports.present = 0, 772 .mr_ext_memseg_en = 1, 773 }; 774 unsigned int vf; 775 int i; 776 char ifname[IF_NAMESIZE]; 777 778 (void)pci_drv; 779 err = mlx4_init_once(); 780 if (err) { 781 ERROR("unable to init PMD global data: %s", 782 strerror(rte_errno)); 783 return -rte_errno; 784 } 785 MLX4_ASSERT(pci_drv == &mlx4_driver); 786 list = mlx4_glue->get_device_list(&i); 787 if (list == NULL) { 788 rte_errno = errno; 789 MLX4_ASSERT(rte_errno); 790 if (rte_errno == ENOSYS) 791 ERROR("cannot list devices, is ib_uverbs loaded?"); 792 return -rte_errno; 793 } 794 MLX4_ASSERT(i >= 0); 795 /* 796 * For each listed device, check related sysfs entry against 797 * the provided PCI ID. 798 */ 799 while (i != 0) { 800 struct rte_pci_addr pci_addr; 801 802 --i; 803 DEBUG("checking device \"%s\"", list[i]->name); 804 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr)) 805 continue; 806 if ((pci_dev->addr.domain != pci_addr.domain) || 807 (pci_dev->addr.bus != pci_addr.bus) || 808 (pci_dev->addr.devid != pci_addr.devid) || 809 (pci_dev->addr.function != pci_addr.function)) 810 continue; 811 vf = (pci_dev->id.device_id == 812 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF); 813 INFO("PCI information matches, using device \"%s\" (VF: %s)", 814 list[i]->name, (vf ? "true" : "false")); 815 attr_ctx = mlx4_glue->open_device(list[i]); 816 err = errno; 817 break; 818 } 819 if (attr_ctx == NULL) { 820 mlx4_glue->free_device_list(list); 821 switch (err) { 822 case 0: 823 rte_errno = ENODEV; 824 ERROR("cannot access device, is mlx4_ib loaded?"); 825 return -rte_errno; 826 case EINVAL: 827 rte_errno = EINVAL; 828 ERROR("cannot use device, are drivers up to date?"); 829 return -rte_errno; 830 } 831 MLX4_ASSERT(err > 0); 832 rte_errno = err; 833 return -rte_errno; 834 } 835 ibv_dev = list[i]; 836 DEBUG("device opened"); 837 if (mlx4_glue->query_device(attr_ctx, &device_attr)) { 838 err = ENODEV; 839 goto error; 840 } 841 INFO("%u port(s) detected", device_attr.phys_port_cnt); 842 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1; 843 if (mlx4_args(pci_dev->device.devargs, &conf)) { 844 ERROR("failed to process device arguments"); 845 err = EINVAL; 846 goto error; 847 } 848 /* Use all ports when none are defined */ 849 if (!conf.ports.enabled) 850 conf.ports.enabled = conf.ports.present; 851 /* Retrieve extended device attributes. */ 852 if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) { 853 err = ENODEV; 854 goto error; 855 } 856 MLX4_ASSERT(device_attr.max_sge >= MLX4_MAX_SGE); 857 for (i = 0; i < device_attr.phys_port_cnt; i++) { 858 uint32_t port = i + 1; /* ports are indexed from one */ 859 struct ibv_context *ctx = NULL; 860 struct ibv_port_attr port_attr; 861 struct ibv_pd *pd = NULL; 862 struct mlx4_priv *priv = NULL; 863 struct rte_eth_dev *eth_dev = NULL; 864 struct rte_ether_addr mac; 865 char name[RTE_ETH_NAME_MAX_LEN]; 866 867 /* If port is not enabled, skip. */ 868 if (!(conf.ports.enabled & (1 << i))) 869 continue; 870 DEBUG("using port %u", port); 871 ctx = mlx4_glue->open_device(ibv_dev); 872 if (ctx == NULL) { 873 err = ENODEV; 874 goto port_error; 875 } 876 snprintf(name, sizeof(name), "%s port %u", 877 mlx4_glue->get_device_name(ibv_dev), port); 878 if (rte_eal_process_type() == RTE_PROC_SECONDARY) { 879 eth_dev = rte_eth_dev_attach_secondary(name); 880 if (eth_dev == NULL) { 881 ERROR("can not attach rte ethdev"); 882 rte_errno = ENOMEM; 883 err = rte_errno; 884 goto error; 885 } 886 priv = eth_dev->data->dev_private; 887 if (!priv->verbs_alloc_ctx.enabled) { 888 ERROR("secondary process is not supported" 889 " due to lack of external allocator" 890 " from Verbs"); 891 rte_errno = ENOTSUP; 892 err = rte_errno; 893 goto error; 894 } 895 eth_dev->device = &pci_dev->device; 896 eth_dev->dev_ops = &mlx4_dev_sec_ops; 897 err = mlx4_proc_priv_init(eth_dev); 898 if (err) 899 goto error; 900 /* Receive command fd from primary process. */ 901 err = mlx4_mp_req_verbs_cmd_fd(eth_dev); 902 if (err < 0) { 903 err = rte_errno; 904 goto error; 905 } 906 /* Remap UAR for Tx queues. */ 907 err = mlx4_tx_uar_init_secondary(eth_dev, err); 908 if (err) { 909 err = rte_errno; 910 goto error; 911 } 912 /* 913 * Ethdev pointer is still required as input since 914 * the primary device is not accessible from the 915 * secondary process. 916 */ 917 eth_dev->tx_pkt_burst = mlx4_tx_burst; 918 eth_dev->rx_pkt_burst = mlx4_rx_burst; 919 claim_zero(mlx4_glue->close_device(ctx)); 920 rte_eth_copy_pci_info(eth_dev, pci_dev); 921 rte_eth_dev_probing_finish(eth_dev); 922 continue; 923 } 924 /* Check port status. */ 925 err = mlx4_glue->query_port(ctx, port, &port_attr); 926 if (err) { 927 err = ENODEV; 928 ERROR("port query failed: %s", strerror(err)); 929 goto port_error; 930 } 931 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) { 932 err = ENOTSUP; 933 ERROR("port %d is not configured in Ethernet mode", 934 port); 935 goto port_error; 936 } 937 if (port_attr.state != IBV_PORT_ACTIVE) 938 DEBUG("port %d is not active: \"%s\" (%d)", 939 port, mlx4_glue->port_state_str(port_attr.state), 940 port_attr.state); 941 /* Make asynchronous FD non-blocking to handle interrupts. */ 942 err = mlx4_fd_set_non_blocking(ctx->async_fd); 943 if (err) { 944 ERROR("cannot make asynchronous FD non-blocking: %s", 945 strerror(err)); 946 goto port_error; 947 } 948 /* Allocate protection domain. */ 949 pd = mlx4_glue->alloc_pd(ctx); 950 if (pd == NULL) { 951 err = ENOMEM; 952 ERROR("PD allocation failure"); 953 goto port_error; 954 } 955 /* from rte_ethdev.c */ 956 priv = rte_zmalloc("ethdev private structure", 957 sizeof(*priv), 958 RTE_CACHE_LINE_SIZE); 959 if (priv == NULL) { 960 err = ENOMEM; 961 ERROR("priv allocation failure"); 962 goto port_error; 963 } 964 priv->ctx = ctx; 965 priv->device_attr = device_attr; 966 priv->port = port; 967 priv->pd = pd; 968 priv->mtu = RTE_ETHER_MTU; 969 priv->vf = vf; 970 priv->hw_csum = !!(device_attr.device_cap_flags & 971 IBV_DEVICE_RAW_IP_CSUM); 972 DEBUG("checksum offloading is %ssupported", 973 (priv->hw_csum ? "" : "not ")); 974 /* Only ConnectX-3 Pro supports tunneling. */ 975 priv->hw_csum_l2tun = 976 priv->hw_csum && 977 (device_attr.vendor_part_id == 978 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO); 979 DEBUG("L2 tunnel checksum offloads are %ssupported", 980 priv->hw_csum_l2tun ? "" : "not "); 981 priv->hw_rss_sup = mlx4_hw_rss_sup(priv->ctx, priv->pd, 982 &device_attr_ex); 983 DEBUG("supported RSS hash fields mask: %016" PRIx64, 984 priv->hw_rss_sup); 985 priv->hw_rss_max_qps = 986 device_attr_ex.rss_caps.max_rwq_indirection_table_size; 987 DEBUG("MAX RSS queues %d", priv->hw_rss_max_qps); 988 priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps & 989 IBV_RAW_PACKET_CAP_SCATTER_FCS); 990 DEBUG("FCS stripping toggling is %ssupported", 991 priv->hw_fcs_strip ? "" : "not "); 992 priv->tso = 993 ((device_attr_ex.tso_caps.max_tso > 0) && 994 (device_attr_ex.tso_caps.supported_qpts & 995 (1 << IBV_QPT_RAW_PACKET))); 996 if (priv->tso) 997 priv->tso_max_payload_sz = 998 device_attr_ex.tso_caps.max_tso; 999 DEBUG("TSO is %ssupported", 1000 priv->tso ? "" : "not "); 1001 priv->mr_ext_memseg_en = conf.mr_ext_memseg_en; 1002 /* Configure the first MAC address by default. */ 1003 err = mlx4_get_mac(priv, &mac.addr_bytes); 1004 if (err) { 1005 ERROR("cannot get MAC address, is mlx4_en loaded?" 1006 " (error: %s)", strerror(err)); 1007 goto port_error; 1008 } 1009 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x", 1010 priv->port, 1011 mac.addr_bytes[0], mac.addr_bytes[1], 1012 mac.addr_bytes[2], mac.addr_bytes[3], 1013 mac.addr_bytes[4], mac.addr_bytes[5]); 1014 /* Register MAC address. */ 1015 priv->mac[0] = mac; 1016 1017 if (mlx4_get_ifname(priv, &ifname) == 0) { 1018 DEBUG("port %u ifname is \"%s\"", 1019 priv->port, ifname); 1020 priv->if_index = if_nametoindex(ifname); 1021 } else { 1022 DEBUG("port %u ifname is unknown", priv->port); 1023 } 1024 1025 /* Get actual MTU if possible. */ 1026 mlx4_mtu_get(priv, &priv->mtu); 1027 DEBUG("port %u MTU is %u", priv->port, priv->mtu); 1028 eth_dev = rte_eth_dev_allocate(name); 1029 if (eth_dev == NULL) { 1030 err = ENOMEM; 1031 ERROR("can not allocate rte ethdev"); 1032 goto port_error; 1033 } 1034 eth_dev->data->dev_private = priv; 1035 eth_dev->data->mac_addrs = priv->mac; 1036 eth_dev->device = &pci_dev->device; 1037 rte_eth_copy_pci_info(eth_dev, pci_dev); 1038 eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS; 1039 /* Initialize local interrupt handle for current port. */ 1040 memset(&priv->intr_handle, 0, sizeof(struct rte_intr_handle)); 1041 priv->intr_handle.fd = -1; 1042 priv->intr_handle.type = RTE_INTR_HANDLE_EXT; 1043 /* 1044 * Override ethdev interrupt handle pointer with private 1045 * handle instead of that of the parent PCI device used by 1046 * default. This prevents it from being shared between all 1047 * ports of the same PCI device since each of them is 1048 * associated its own Verbs context. 1049 * 1050 * Rx interrupts in particular require this as the PMD has 1051 * no control over the registration of queue interrupts 1052 * besides setting up eth_dev->intr_handle, the rest is 1053 * handled by rte_intr_rx_ctl(). 1054 */ 1055 eth_dev->intr_handle = &priv->intr_handle; 1056 priv->dev_data = eth_dev->data; 1057 eth_dev->dev_ops = &mlx4_dev_ops; 1058 #ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS 1059 /* Hint libmlx4 to use PMD allocator for data plane resources */ 1060 err = mlx4_glue->dv_set_context_attr 1061 (ctx, MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS, 1062 (void *)((uintptr_t)&(struct mlx4dv_ctx_allocators){ 1063 .alloc = &mlx4_alloc_verbs_buf, 1064 .free = &mlx4_free_verbs_buf, 1065 .data = priv, 1066 })); 1067 if (err) 1068 WARN("Verbs external allocator is not supported"); 1069 else 1070 priv->verbs_alloc_ctx.enabled = 1; 1071 #endif 1072 /* Bring Ethernet device up. */ 1073 DEBUG("forcing Ethernet interface up"); 1074 mlx4_dev_set_link_up(eth_dev); 1075 /* Update link status once if waiting for LSC. */ 1076 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC) 1077 mlx4_link_update(eth_dev, 0); 1078 /* 1079 * Once the device is added to the list of memory event 1080 * callback, its global MR cache table cannot be expanded 1081 * on the fly because of deadlock. If it overflows, lookup 1082 * should be done by searching MR list linearly, which is slow. 1083 */ 1084 err = mlx4_mr_btree_init(&priv->mr.cache, 1085 MLX4_MR_BTREE_CACHE_N * 2, 1086 eth_dev->device->numa_node); 1087 if (err) { 1088 /* rte_errno is already set. */ 1089 goto port_error; 1090 } 1091 /* Add device to memory callback list. */ 1092 rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock); 1093 LIST_INSERT_HEAD(&mlx4_shared_data->mem_event_cb_list, 1094 priv, mem_event_cb); 1095 rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock); 1096 rte_eth_dev_probing_finish(eth_dev); 1097 continue; 1098 port_error: 1099 rte_free(priv); 1100 if (eth_dev != NULL) 1101 eth_dev->data->dev_private = NULL; 1102 if (pd) 1103 claim_zero(mlx4_glue->dealloc_pd(pd)); 1104 if (ctx) 1105 claim_zero(mlx4_glue->close_device(ctx)); 1106 if (eth_dev != NULL) { 1107 /* mac_addrs must not be freed because part of dev_private */ 1108 eth_dev->data->mac_addrs = NULL; 1109 rte_eth_dev_release_port(eth_dev); 1110 } 1111 break; 1112 } 1113 /* 1114 * XXX if something went wrong in the loop above, there is a resource 1115 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as 1116 * long as the dpdk does not provide a way to deallocate a ethdev and a 1117 * way to enumerate the registered ethdevs to free the previous ones. 1118 */ 1119 error: 1120 if (attr_ctx) 1121 claim_zero(mlx4_glue->close_device(attr_ctx)); 1122 if (list) 1123 mlx4_glue->free_device_list(list); 1124 if (err) 1125 rte_errno = err; 1126 return -err; 1127 } 1128 1129 static const struct rte_pci_id mlx4_pci_id_map[] = { 1130 { 1131 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, 1132 PCI_DEVICE_ID_MELLANOX_CONNECTX3) 1133 }, 1134 { 1135 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, 1136 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO) 1137 }, 1138 { 1139 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, 1140 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF) 1141 }, 1142 { 1143 .vendor_id = 0 1144 } 1145 }; 1146 1147 static struct rte_pci_driver mlx4_driver = { 1148 .driver = { 1149 .name = MLX4_DRIVER_NAME 1150 }, 1151 .id_table = mlx4_pci_id_map, 1152 .probe = mlx4_pci_probe, 1153 .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV, 1154 }; 1155 1156 #ifdef RTE_IBVERBS_LINK_DLOPEN 1157 1158 /** 1159 * Suffix RTE_EAL_PMD_PATH with "-glue". 1160 * 1161 * This function performs a sanity check on RTE_EAL_PMD_PATH before 1162 * suffixing its last component. 1163 * 1164 * @param buf[out] 1165 * Output buffer, should be large enough otherwise NULL is returned. 1166 * @param size 1167 * Size of @p out. 1168 * 1169 * @return 1170 * Pointer to @p buf or @p NULL in case suffix cannot be appended. 1171 */ 1172 static char * 1173 mlx4_glue_path(char *buf, size_t size) 1174 { 1175 static const char *const bad[] = { "/", ".", "..", NULL }; 1176 const char *path = RTE_EAL_PMD_PATH; 1177 size_t len = strlen(path); 1178 size_t off; 1179 int i; 1180 1181 while (len && path[len - 1] == '/') 1182 --len; 1183 for (off = len; off && path[off - 1] != '/'; --off) 1184 ; 1185 for (i = 0; bad[i]; ++i) 1186 if (!strncmp(path + off, bad[i], (int)(len - off))) 1187 goto error; 1188 i = snprintf(buf, size, "%.*s-glue", (int)len, path); 1189 if (i == -1 || (size_t)i >= size) 1190 goto error; 1191 return buf; 1192 error: 1193 ERROR("unable to append \"-glue\" to last component of" 1194 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\")," 1195 " please re-configure DPDK"); 1196 return NULL; 1197 } 1198 1199 /** 1200 * Initialization routine for run-time dependency on rdma-core. 1201 */ 1202 static int 1203 mlx4_glue_init(void) 1204 { 1205 char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")]; 1206 const char *path[] = { 1207 /* 1208 * A basic security check is necessary before trusting 1209 * MLX4_GLUE_PATH, which may override RTE_EAL_PMD_PATH. 1210 */ 1211 (geteuid() == getuid() && getegid() == getgid() ? 1212 getenv("MLX4_GLUE_PATH") : NULL), 1213 /* 1214 * When RTE_EAL_PMD_PATH is set, use its glue-suffixed 1215 * variant, otherwise let dlopen() look up libraries on its 1216 * own. 1217 */ 1218 (*RTE_EAL_PMD_PATH ? 1219 mlx4_glue_path(glue_path, sizeof(glue_path)) : ""), 1220 }; 1221 unsigned int i = 0; 1222 void *handle = NULL; 1223 void **sym; 1224 const char *dlmsg; 1225 1226 while (!handle && i != RTE_DIM(path)) { 1227 const char *end; 1228 size_t len; 1229 int ret; 1230 1231 if (!path[i]) { 1232 ++i; 1233 continue; 1234 } 1235 end = strpbrk(path[i], ":;"); 1236 if (!end) 1237 end = path[i] + strlen(path[i]); 1238 len = end - path[i]; 1239 ret = 0; 1240 do { 1241 char name[ret + 1]; 1242 1243 ret = snprintf(name, sizeof(name), "%.*s%s" MLX4_GLUE, 1244 (int)len, path[i], 1245 (!len || *(end - 1) == '/') ? "" : "/"); 1246 if (ret == -1) 1247 break; 1248 if (sizeof(name) != (size_t)ret + 1) 1249 continue; 1250 DEBUG("looking for rdma-core glue as \"%s\"", name); 1251 handle = dlopen(name, RTLD_LAZY); 1252 break; 1253 } while (1); 1254 path[i] = end + 1; 1255 if (!*end) 1256 ++i; 1257 } 1258 if (!handle) { 1259 rte_errno = EINVAL; 1260 dlmsg = dlerror(); 1261 if (dlmsg) 1262 WARN("cannot load glue library: %s", dlmsg); 1263 goto glue_error; 1264 } 1265 sym = dlsym(handle, "mlx4_glue"); 1266 if (!sym || !*sym) { 1267 rte_errno = EINVAL; 1268 dlmsg = dlerror(); 1269 if (dlmsg) 1270 ERROR("cannot resolve glue symbol: %s", dlmsg); 1271 goto glue_error; 1272 } 1273 mlx4_glue = *sym; 1274 return 0; 1275 glue_error: 1276 if (handle) 1277 dlclose(handle); 1278 WARN("cannot initialize PMD due to missing run-time" 1279 " dependency on rdma-core libraries (libibverbs," 1280 " libmlx4)"); 1281 return -rte_errno; 1282 } 1283 1284 #endif 1285 1286 /* Initialize driver log type. */ 1287 RTE_LOG_REGISTER(mlx4_logtype, pmd.net.mlx4, NOTICE) 1288 1289 /** 1290 * Driver initialization routine. 1291 */ 1292 RTE_INIT(rte_mlx4_pmd_init) 1293 { 1294 /* 1295 * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we 1296 * want to get success errno value in case of calling them 1297 * when the device was removed. 1298 */ 1299 setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1); 1300 /* 1301 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use 1302 * huge pages. Calling ibv_fork_init() during init allows 1303 * applications to use fork() safely for purposes other than 1304 * using this PMD, which is not supported in forked processes. 1305 */ 1306 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1); 1307 #ifdef RTE_IBVERBS_LINK_DLOPEN 1308 if (mlx4_glue_init()) 1309 return; 1310 MLX4_ASSERT(mlx4_glue); 1311 #endif 1312 #ifdef RTE_LIBRTE_MLX4_DEBUG 1313 /* Glue structure must not contain any NULL pointers. */ 1314 { 1315 unsigned int i; 1316 1317 for (i = 0; i != sizeof(*mlx4_glue) / sizeof(void *); ++i) 1318 MLX4_ASSERT(((const void *const *)mlx4_glue)[i]); 1319 } 1320 #endif 1321 if (strcmp(mlx4_glue->version, MLX4_GLUE_VERSION)) { 1322 ERROR("rdma-core glue \"%s\" mismatch: \"%s\" is required", 1323 mlx4_glue->version, MLX4_GLUE_VERSION); 1324 return; 1325 } 1326 mlx4_glue->fork_init(); 1327 rte_pci_register(&mlx4_driver); 1328 } 1329 1330 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__); 1331 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map); 1332 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4, 1333 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib"); 1334