1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (C) 2016 Intel Corporation. All rights reserved. 3 * Copyright (c) 2019-2021 Mellanox Technologies LTD. All rights reserved. 4 * Copyright (c) 2021-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved. 5 */ 6 7 /* 8 * NVMe over RDMA transport 9 */ 10 11 #include "spdk/stdinc.h" 12 13 #include "spdk/assert.h" 14 #include "spdk/dma.h" 15 #include "spdk/log.h" 16 #include "spdk/trace.h" 17 #include "spdk/queue.h" 18 #include "spdk/nvme.h" 19 #include "spdk/nvmf_spec.h" 20 #include "spdk/string.h" 21 #include "spdk/endian.h" 22 #include "spdk/likely.h" 23 #include "spdk/config.h" 24 25 #include "nvme_internal.h" 26 #include "spdk_internal/rdma.h" 27 28 #define NVME_RDMA_TIME_OUT_IN_MS 2000 29 #define NVME_RDMA_RW_BUFFER_SIZE 131072 30 31 /* 32 * NVME RDMA qpair Resource Defaults 33 */ 34 #define NVME_RDMA_DEFAULT_TX_SGE 2 35 #define NVME_RDMA_DEFAULT_RX_SGE 1 36 37 /* Max number of NVMe-oF SGL descriptors supported by the host */ 38 #define NVME_RDMA_MAX_SGL_DESCRIPTORS 16 39 40 /* number of STAILQ entries for holding pending RDMA CM events. */ 41 #define NVME_RDMA_NUM_CM_EVENTS 256 42 43 /* The default size for a shared rdma completion queue. */ 44 #define DEFAULT_NVME_RDMA_CQ_SIZE 4096 45 46 /* 47 * In the special case of a stale connection we don't expose a mechanism 48 * for the user to retry the connection so we need to handle it internally. 49 */ 50 #define NVME_RDMA_STALE_CONN_RETRY_MAX 5 51 #define NVME_RDMA_STALE_CONN_RETRY_DELAY_US 10000 52 53 /* 54 * Maximum value of transport_retry_count used by RDMA controller 55 */ 56 #define NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT 7 57 58 /* 59 * Maximum value of transport_ack_timeout used by RDMA controller 60 */ 61 #define NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT 31 62 63 /* 64 * Number of microseconds to wait until the lingering qpair becomes quiet. 65 */ 66 #define NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US 1000000ull 67 68 /* 69 * The max length of keyed SGL data block (3 bytes) 70 */ 71 #define NVME_RDMA_MAX_KEYED_SGL_LENGTH ((1u << 24u) - 1) 72 73 #define WC_PER_QPAIR(queue_depth) (queue_depth * 2) 74 75 #define NVME_RDMA_POLL_GROUP_CHECK_QPN(_rqpair, qpn) \ 76 ((_rqpair)->rdma_qp && (_rqpair)->rdma_qp->qp->qp_num == (qpn)) \ 77 78 struct nvme_rdma_memory_domain { 79 TAILQ_ENTRY(nvme_rdma_memory_domain) link; 80 uint32_t ref; 81 struct ibv_pd *pd; 82 struct spdk_memory_domain *domain; 83 struct spdk_memory_domain_rdma_ctx rdma_ctx; 84 }; 85 86 enum nvme_rdma_wr_type { 87 RDMA_WR_TYPE_RECV, 88 RDMA_WR_TYPE_SEND, 89 }; 90 91 struct nvme_rdma_wr { 92 /* Using this instead of the enum allows this struct to only occupy one byte. */ 93 uint8_t type; 94 }; 95 96 struct spdk_nvmf_cmd { 97 struct spdk_nvme_cmd cmd; 98 struct spdk_nvme_sgl_descriptor sgl[NVME_RDMA_MAX_SGL_DESCRIPTORS]; 99 }; 100 101 struct spdk_nvme_rdma_hooks g_nvme_hooks = {}; 102 103 /* STAILQ wrapper for cm events. */ 104 struct nvme_rdma_cm_event_entry { 105 struct rdma_cm_event *evt; 106 STAILQ_ENTRY(nvme_rdma_cm_event_entry) link; 107 }; 108 109 /* NVMe RDMA transport extensions for spdk_nvme_ctrlr */ 110 struct nvme_rdma_ctrlr { 111 struct spdk_nvme_ctrlr ctrlr; 112 113 uint16_t max_sge; 114 115 struct rdma_event_channel *cm_channel; 116 117 STAILQ_HEAD(, nvme_rdma_cm_event_entry) pending_cm_events; 118 119 STAILQ_HEAD(, nvme_rdma_cm_event_entry) free_cm_events; 120 121 struct nvme_rdma_cm_event_entry *cm_events; 122 }; 123 124 struct nvme_rdma_poller_stats { 125 uint64_t polls; 126 uint64_t idle_polls; 127 uint64_t queued_requests; 128 uint64_t completions; 129 struct spdk_rdma_qp_stats rdma_stats; 130 }; 131 132 struct nvme_rdma_poll_group; 133 struct nvme_rdma_rsps; 134 135 struct nvme_rdma_poller { 136 struct ibv_context *device; 137 struct ibv_cq *cq; 138 struct spdk_rdma_srq *srq; 139 struct nvme_rdma_rsps *rsps; 140 struct ibv_pd *pd; 141 struct spdk_rdma_mem_map *mr_map; 142 uint32_t refcnt; 143 int required_num_wc; 144 int current_num_wc; 145 struct nvme_rdma_poller_stats stats; 146 struct nvme_rdma_poll_group *group; 147 STAILQ_ENTRY(nvme_rdma_poller) link; 148 }; 149 150 struct nvme_rdma_qpair; 151 152 struct nvme_rdma_poll_group { 153 struct spdk_nvme_transport_poll_group group; 154 STAILQ_HEAD(, nvme_rdma_poller) pollers; 155 uint32_t num_pollers; 156 TAILQ_HEAD(, nvme_rdma_qpair) connecting_qpairs; 157 }; 158 159 enum nvme_rdma_qpair_state { 160 NVME_RDMA_QPAIR_STATE_INVALID = 0, 161 NVME_RDMA_QPAIR_STATE_STALE_CONN, 162 NVME_RDMA_QPAIR_STATE_INITIALIZING, 163 NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND, 164 NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL, 165 NVME_RDMA_QPAIR_STATE_RUNNING, 166 NVME_RDMA_QPAIR_STATE_EXITING, 167 NVME_RDMA_QPAIR_STATE_LINGERING, 168 NVME_RDMA_QPAIR_STATE_EXITED, 169 }; 170 171 typedef int (*nvme_rdma_cm_event_cb)(struct nvme_rdma_qpair *rqpair, int ret); 172 173 struct nvme_rdma_rsp_opts { 174 uint16_t num_entries; 175 struct nvme_rdma_qpair *rqpair; 176 struct spdk_rdma_srq *srq; 177 struct spdk_rdma_mem_map *mr_map; 178 }; 179 180 struct nvme_rdma_rsps { 181 /* Parallel arrays of response buffers + response SGLs of size num_entries */ 182 struct ibv_sge *rsp_sgls; 183 struct spdk_nvme_rdma_rsp *rsps; 184 185 struct ibv_recv_wr *rsp_recv_wrs; 186 187 /* Count of outstanding recv objects */ 188 uint16_t current_num_recvs; 189 190 uint16_t num_entries; 191 }; 192 193 /* NVMe RDMA qpair extensions for spdk_nvme_qpair */ 194 struct nvme_rdma_qpair { 195 struct spdk_nvme_qpair qpair; 196 197 struct spdk_rdma_qp *rdma_qp; 198 struct rdma_cm_id *cm_id; 199 struct ibv_cq *cq; 200 struct spdk_rdma_srq *srq; 201 202 struct spdk_nvme_rdma_req *rdma_reqs; 203 204 uint32_t max_send_sge; 205 206 uint32_t max_recv_sge; 207 208 uint16_t num_entries; 209 210 bool delay_cmd_submit; 211 212 uint32_t num_completions; 213 214 struct nvme_rdma_rsps *rsps; 215 216 /* 217 * Array of num_entries NVMe commands registered as RDMA message buffers. 218 * Indexed by rdma_req->id. 219 */ 220 struct spdk_nvmf_cmd *cmds; 221 222 struct spdk_rdma_mem_map *mr_map; 223 224 TAILQ_HEAD(, spdk_nvme_rdma_req) free_reqs; 225 TAILQ_HEAD(, spdk_nvme_rdma_req) outstanding_reqs; 226 227 struct nvme_rdma_memory_domain *memory_domain; 228 229 /* Count of outstanding send objects */ 230 uint16_t current_num_sends; 231 232 /* Placed at the end of the struct since it is not used frequently */ 233 struct rdma_cm_event *evt; 234 struct nvme_rdma_poller *poller; 235 236 uint64_t evt_timeout_ticks; 237 nvme_rdma_cm_event_cb evt_cb; 238 enum rdma_cm_event_type expected_evt_type; 239 240 enum nvme_rdma_qpair_state state; 241 242 bool in_connect_poll; 243 244 uint8_t stale_conn_retry_count; 245 bool need_destroy; 246 247 TAILQ_ENTRY(nvme_rdma_qpair) link_connecting; 248 }; 249 250 enum NVME_RDMA_COMPLETION_FLAGS { 251 NVME_RDMA_SEND_COMPLETED = 1u << 0, 252 NVME_RDMA_RECV_COMPLETED = 1u << 1, 253 }; 254 255 struct spdk_nvme_rdma_req { 256 uint16_t id; 257 uint16_t completion_flags: 2; 258 uint16_t reserved: 14; 259 /* if completion of RDMA_RECV received before RDMA_SEND, we will complete nvme request 260 * during processing of RDMA_SEND. To complete the request we must know the response 261 * received in RDMA_RECV, so store it in this field */ 262 struct spdk_nvme_rdma_rsp *rdma_rsp; 263 264 struct nvme_rdma_wr rdma_wr; 265 266 struct ibv_send_wr send_wr; 267 268 struct nvme_request *req; 269 270 struct ibv_sge send_sgl[NVME_RDMA_DEFAULT_TX_SGE]; 271 272 TAILQ_ENTRY(spdk_nvme_rdma_req) link; 273 }; 274 275 struct spdk_nvme_rdma_rsp { 276 struct spdk_nvme_cpl cpl; 277 struct nvme_rdma_qpair *rqpair; 278 struct ibv_recv_wr *recv_wr; 279 struct nvme_rdma_wr rdma_wr; 280 }; 281 282 struct nvme_rdma_memory_translation_ctx { 283 void *addr; 284 size_t length; 285 uint32_t lkey; 286 uint32_t rkey; 287 }; 288 289 static const char *rdma_cm_event_str[] = { 290 "RDMA_CM_EVENT_ADDR_RESOLVED", 291 "RDMA_CM_EVENT_ADDR_ERROR", 292 "RDMA_CM_EVENT_ROUTE_RESOLVED", 293 "RDMA_CM_EVENT_ROUTE_ERROR", 294 "RDMA_CM_EVENT_CONNECT_REQUEST", 295 "RDMA_CM_EVENT_CONNECT_RESPONSE", 296 "RDMA_CM_EVENT_CONNECT_ERROR", 297 "RDMA_CM_EVENT_UNREACHABLE", 298 "RDMA_CM_EVENT_REJECTED", 299 "RDMA_CM_EVENT_ESTABLISHED", 300 "RDMA_CM_EVENT_DISCONNECTED", 301 "RDMA_CM_EVENT_DEVICE_REMOVAL", 302 "RDMA_CM_EVENT_MULTICAST_JOIN", 303 "RDMA_CM_EVENT_MULTICAST_ERROR", 304 "RDMA_CM_EVENT_ADDR_CHANGE", 305 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 306 }; 307 308 static struct nvme_rdma_poller *nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, 309 struct ibv_context *device); 310 static void nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, 311 struct nvme_rdma_poller *poller); 312 313 static TAILQ_HEAD(, nvme_rdma_memory_domain) g_memory_domains = TAILQ_HEAD_INITIALIZER( 314 g_memory_domains); 315 static pthread_mutex_t g_memory_domains_lock = PTHREAD_MUTEX_INITIALIZER; 316 317 static struct nvme_rdma_memory_domain * 318 nvme_rdma_get_memory_domain(struct ibv_pd *pd) 319 { 320 struct nvme_rdma_memory_domain *domain = NULL; 321 struct spdk_memory_domain_ctx ctx; 322 int rc; 323 324 pthread_mutex_lock(&g_memory_domains_lock); 325 326 TAILQ_FOREACH(domain, &g_memory_domains, link) { 327 if (domain->pd == pd) { 328 domain->ref++; 329 pthread_mutex_unlock(&g_memory_domains_lock); 330 return domain; 331 } 332 } 333 334 domain = calloc(1, sizeof(*domain)); 335 if (!domain) { 336 SPDK_ERRLOG("Memory allocation failed\n"); 337 pthread_mutex_unlock(&g_memory_domains_lock); 338 return NULL; 339 } 340 341 domain->rdma_ctx.size = sizeof(domain->rdma_ctx); 342 domain->rdma_ctx.ibv_pd = pd; 343 ctx.size = sizeof(ctx); 344 ctx.user_ctx = &domain->rdma_ctx; 345 346 rc = spdk_memory_domain_create(&domain->domain, SPDK_DMA_DEVICE_TYPE_RDMA, &ctx, 347 SPDK_RDMA_DMA_DEVICE); 348 if (rc) { 349 SPDK_ERRLOG("Failed to create memory domain\n"); 350 free(domain); 351 pthread_mutex_unlock(&g_memory_domains_lock); 352 return NULL; 353 } 354 355 domain->pd = pd; 356 domain->ref = 1; 357 TAILQ_INSERT_TAIL(&g_memory_domains, domain, link); 358 359 pthread_mutex_unlock(&g_memory_domains_lock); 360 361 return domain; 362 } 363 364 static void 365 nvme_rdma_put_memory_domain(struct nvme_rdma_memory_domain *device) 366 { 367 if (!device) { 368 return; 369 } 370 371 pthread_mutex_lock(&g_memory_domains_lock); 372 373 assert(device->ref > 0); 374 375 device->ref--; 376 377 if (device->ref == 0) { 378 spdk_memory_domain_destroy(device->domain); 379 TAILQ_REMOVE(&g_memory_domains, device, link); 380 free(device); 381 } 382 383 pthread_mutex_unlock(&g_memory_domains_lock); 384 } 385 386 static int nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, 387 struct spdk_nvme_qpair *qpair); 388 389 static inline struct nvme_rdma_qpair * 390 nvme_rdma_qpair(struct spdk_nvme_qpair *qpair) 391 { 392 assert(qpair->trtype == SPDK_NVME_TRANSPORT_RDMA); 393 return SPDK_CONTAINEROF(qpair, struct nvme_rdma_qpair, qpair); 394 } 395 396 static inline struct nvme_rdma_poll_group * 397 nvme_rdma_poll_group(struct spdk_nvme_transport_poll_group *group) 398 { 399 return (SPDK_CONTAINEROF(group, struct nvme_rdma_poll_group, group)); 400 } 401 402 static inline struct nvme_rdma_ctrlr * 403 nvme_rdma_ctrlr(struct spdk_nvme_ctrlr *ctrlr) 404 { 405 assert(ctrlr->trid.trtype == SPDK_NVME_TRANSPORT_RDMA); 406 return SPDK_CONTAINEROF(ctrlr, struct nvme_rdma_ctrlr, ctrlr); 407 } 408 409 static struct spdk_nvme_rdma_req * 410 nvme_rdma_req_get(struct nvme_rdma_qpair *rqpair) 411 { 412 struct spdk_nvme_rdma_req *rdma_req; 413 414 rdma_req = TAILQ_FIRST(&rqpair->free_reqs); 415 if (rdma_req) { 416 TAILQ_REMOVE(&rqpair->free_reqs, rdma_req, link); 417 } 418 419 return rdma_req; 420 } 421 422 static void 423 nvme_rdma_req_put(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req) 424 { 425 rdma_req->completion_flags = 0; 426 rdma_req->req = NULL; 427 rdma_req->rdma_rsp = NULL; 428 TAILQ_INSERT_HEAD(&rqpair->free_reqs, rdma_req, link); 429 } 430 431 static void 432 nvme_rdma_req_complete(struct spdk_nvme_rdma_req *rdma_req, 433 struct spdk_nvme_cpl *rsp, 434 bool print_on_error) 435 { 436 struct nvme_request *req = rdma_req->req; 437 struct nvme_rdma_qpair *rqpair; 438 struct spdk_nvme_qpair *qpair; 439 bool error, print_error; 440 441 assert(req != NULL); 442 443 qpair = req->qpair; 444 rqpair = nvme_rdma_qpair(qpair); 445 446 error = spdk_nvme_cpl_is_error(rsp); 447 print_error = error && print_on_error && !qpair->ctrlr->opts.disable_error_logging; 448 449 if (print_error) { 450 spdk_nvme_qpair_print_command(qpair, &req->cmd); 451 } 452 453 if (print_error || SPDK_DEBUGLOG_FLAG_ENABLED("nvme")) { 454 spdk_nvme_qpair_print_completion(qpair, rsp); 455 } 456 457 TAILQ_REMOVE(&rqpair->outstanding_reqs, rdma_req, link); 458 459 nvme_complete_request(req->cb_fn, req->cb_arg, qpair, req, rsp); 460 nvme_rdma_req_put(rqpair, rdma_req); 461 } 462 463 static const char * 464 nvme_rdma_cm_event_str_get(uint32_t event) 465 { 466 if (event < SPDK_COUNTOF(rdma_cm_event_str)) { 467 return rdma_cm_event_str[event]; 468 } else { 469 return "Undefined"; 470 } 471 } 472 473 474 static int 475 nvme_rdma_qpair_process_cm_event(struct nvme_rdma_qpair *rqpair) 476 { 477 struct rdma_cm_event *event = rqpair->evt; 478 struct spdk_nvmf_rdma_accept_private_data *accept_data; 479 int rc = 0; 480 481 if (event) { 482 switch (event->event) { 483 case RDMA_CM_EVENT_ADDR_RESOLVED: 484 case RDMA_CM_EVENT_ADDR_ERROR: 485 case RDMA_CM_EVENT_ROUTE_RESOLVED: 486 case RDMA_CM_EVENT_ROUTE_ERROR: 487 break; 488 case RDMA_CM_EVENT_CONNECT_REQUEST: 489 break; 490 case RDMA_CM_EVENT_CONNECT_ERROR: 491 break; 492 case RDMA_CM_EVENT_UNREACHABLE: 493 case RDMA_CM_EVENT_REJECTED: 494 break; 495 case RDMA_CM_EVENT_CONNECT_RESPONSE: 496 rc = spdk_rdma_qp_complete_connect(rqpair->rdma_qp); 497 /* fall through */ 498 case RDMA_CM_EVENT_ESTABLISHED: 499 accept_data = (struct spdk_nvmf_rdma_accept_private_data *)event->param.conn.private_data; 500 if (accept_data == NULL) { 501 rc = -1; 502 } else { 503 SPDK_DEBUGLOG(nvme, "Requested queue depth %d. Target receive queue depth %d.\n", 504 rqpair->num_entries + 1, accept_data->crqsize); 505 } 506 break; 507 case RDMA_CM_EVENT_DISCONNECTED: 508 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE; 509 break; 510 case RDMA_CM_EVENT_DEVICE_REMOVAL: 511 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL; 512 rqpair->need_destroy = true; 513 break; 514 case RDMA_CM_EVENT_MULTICAST_JOIN: 515 case RDMA_CM_EVENT_MULTICAST_ERROR: 516 break; 517 case RDMA_CM_EVENT_ADDR_CHANGE: 518 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL; 519 break; 520 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 521 break; 522 default: 523 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 524 break; 525 } 526 rqpair->evt = NULL; 527 rdma_ack_cm_event(event); 528 } 529 530 return rc; 531 } 532 533 /* 534 * This function must be called under the nvme controller's lock 535 * because it touches global controller variables. The lock is taken 536 * by the generic transport code before invoking a few of the functions 537 * in this file: nvme_rdma_ctrlr_connect_qpair, nvme_rdma_ctrlr_delete_io_qpair, 538 * and conditionally nvme_rdma_qpair_process_completions when it is calling 539 * completions on the admin qpair. When adding a new call to this function, please 540 * verify that it is in a situation where it falls under the lock. 541 */ 542 static int 543 nvme_rdma_poll_events(struct nvme_rdma_ctrlr *rctrlr) 544 { 545 struct nvme_rdma_cm_event_entry *entry, *tmp; 546 struct nvme_rdma_qpair *event_qpair; 547 struct rdma_cm_event *event; 548 struct rdma_event_channel *channel = rctrlr->cm_channel; 549 550 STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) { 551 event_qpair = entry->evt->id->context; 552 if (event_qpair->evt == NULL) { 553 event_qpair->evt = entry->evt; 554 STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link); 555 STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link); 556 } 557 } 558 559 while (rdma_get_cm_event(channel, &event) == 0) { 560 event_qpair = event->id->context; 561 if (event_qpair->evt == NULL) { 562 event_qpair->evt = event; 563 } else { 564 assert(rctrlr == nvme_rdma_ctrlr(event_qpair->qpair.ctrlr)); 565 entry = STAILQ_FIRST(&rctrlr->free_cm_events); 566 if (entry == NULL) { 567 rdma_ack_cm_event(event); 568 return -ENOMEM; 569 } 570 STAILQ_REMOVE(&rctrlr->free_cm_events, entry, nvme_rdma_cm_event_entry, link); 571 entry->evt = event; 572 STAILQ_INSERT_TAIL(&rctrlr->pending_cm_events, entry, link); 573 } 574 } 575 576 /* rdma_get_cm_event() returns -1 on error. If an error occurs, errno 577 * will be set to indicate the failure reason. So return negated errno here. 578 */ 579 return -errno; 580 } 581 582 static int 583 nvme_rdma_validate_cm_event(enum rdma_cm_event_type expected_evt_type, 584 struct rdma_cm_event *reaped_evt) 585 { 586 int rc = -EBADMSG; 587 588 if (expected_evt_type == reaped_evt->event) { 589 return 0; 590 } 591 592 switch (expected_evt_type) { 593 case RDMA_CM_EVENT_ESTABLISHED: 594 /* 595 * There is an enum ib_cm_rej_reason in the kernel headers that sets 10 as 596 * IB_CM_REJ_STALE_CONN. I can't find the corresponding userspace but we get 597 * the same values here. 598 */ 599 if (reaped_evt->event == RDMA_CM_EVENT_REJECTED && reaped_evt->status == 10) { 600 rc = -ESTALE; 601 } else if (reaped_evt->event == RDMA_CM_EVENT_CONNECT_RESPONSE) { 602 /* 603 * If we are using a qpair which is not created using rdma cm API 604 * then we will receive RDMA_CM_EVENT_CONNECT_RESPONSE instead of 605 * RDMA_CM_EVENT_ESTABLISHED. 606 */ 607 return 0; 608 } 609 break; 610 default: 611 break; 612 } 613 614 SPDK_ERRLOG("Expected %s but received %s (%d) from CM event channel (status = %d)\n", 615 nvme_rdma_cm_event_str_get(expected_evt_type), 616 nvme_rdma_cm_event_str_get(reaped_evt->event), reaped_evt->event, 617 reaped_evt->status); 618 return rc; 619 } 620 621 static int 622 nvme_rdma_process_event_start(struct nvme_rdma_qpair *rqpair, 623 enum rdma_cm_event_type evt, 624 nvme_rdma_cm_event_cb evt_cb) 625 { 626 int rc; 627 628 assert(evt_cb != NULL); 629 630 if (rqpair->evt != NULL) { 631 rc = nvme_rdma_qpair_process_cm_event(rqpair); 632 if (rc) { 633 return rc; 634 } 635 } 636 637 rqpair->expected_evt_type = evt; 638 rqpair->evt_cb = evt_cb; 639 rqpair->evt_timeout_ticks = (g_spdk_nvme_transport_opts.rdma_cm_event_timeout_ms * 1000 * 640 spdk_get_ticks_hz()) / SPDK_SEC_TO_USEC + spdk_get_ticks(); 641 642 return 0; 643 } 644 645 static int 646 nvme_rdma_process_event_poll(struct nvme_rdma_qpair *rqpair) 647 { 648 struct nvme_rdma_ctrlr *rctrlr; 649 int rc = 0, rc2; 650 651 rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr); 652 assert(rctrlr != NULL); 653 654 if (!rqpair->evt && spdk_get_ticks() < rqpair->evt_timeout_ticks) { 655 rc = nvme_rdma_poll_events(rctrlr); 656 if (rc == -EAGAIN || rc == -EWOULDBLOCK) { 657 return rc; 658 } 659 } 660 661 if (rqpair->evt == NULL) { 662 rc = -EADDRNOTAVAIL; 663 goto exit; 664 } 665 666 rc = nvme_rdma_validate_cm_event(rqpair->expected_evt_type, rqpair->evt); 667 668 rc2 = nvme_rdma_qpair_process_cm_event(rqpair); 669 /* bad message takes precedence over the other error codes from processing the event. */ 670 rc = rc == 0 ? rc2 : rc; 671 672 exit: 673 assert(rqpair->evt_cb != NULL); 674 return rqpair->evt_cb(rqpair, rc); 675 } 676 677 static int 678 nvme_rdma_resize_cq(struct nvme_rdma_qpair *rqpair, struct nvme_rdma_poller *poller) 679 { 680 int current_num_wc, required_num_wc; 681 int max_cq_size; 682 683 required_num_wc = poller->required_num_wc + WC_PER_QPAIR(rqpair->num_entries); 684 current_num_wc = poller->current_num_wc; 685 if (current_num_wc < required_num_wc) { 686 current_num_wc = spdk_max(current_num_wc * 2, required_num_wc); 687 } 688 689 max_cq_size = g_spdk_nvme_transport_opts.rdma_max_cq_size; 690 if (max_cq_size != 0 && current_num_wc > max_cq_size) { 691 current_num_wc = max_cq_size; 692 } 693 694 if (poller->current_num_wc != current_num_wc) { 695 SPDK_DEBUGLOG(nvme, "Resize RDMA CQ from %d to %d\n", poller->current_num_wc, 696 current_num_wc); 697 if (ibv_resize_cq(poller->cq, current_num_wc)) { 698 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 699 return -1; 700 } 701 702 poller->current_num_wc = current_num_wc; 703 } 704 705 poller->required_num_wc = required_num_wc; 706 return 0; 707 } 708 709 static int 710 nvme_rdma_qpair_set_poller(struct spdk_nvme_qpair *qpair) 711 { 712 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 713 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group); 714 struct nvme_rdma_poller *poller; 715 716 assert(rqpair->cq == NULL); 717 718 poller = nvme_rdma_poll_group_get_poller(group, rqpair->cm_id->verbs); 719 if (!poller) { 720 SPDK_ERRLOG("Unable to find a cq for qpair %p on poll group %p\n", qpair, qpair->poll_group); 721 return -EINVAL; 722 } 723 724 if (!poller->srq) { 725 if (nvme_rdma_resize_cq(rqpair, poller)) { 726 nvme_rdma_poll_group_put_poller(group, poller); 727 return -EPROTO; 728 } 729 } 730 731 rqpair->cq = poller->cq; 732 rqpair->srq = poller->srq; 733 if (rqpair->srq) { 734 rqpair->rsps = poller->rsps; 735 } 736 rqpair->poller = poller; 737 return 0; 738 } 739 740 static int 741 nvme_rdma_qpair_init(struct nvme_rdma_qpair *rqpair) 742 { 743 int rc; 744 struct spdk_rdma_qp_init_attr attr = {}; 745 struct ibv_device_attr dev_attr; 746 struct nvme_rdma_ctrlr *rctrlr; 747 uint32_t num_cqe, max_num_cqe; 748 749 rc = ibv_query_device(rqpair->cm_id->verbs, &dev_attr); 750 if (rc != 0) { 751 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 752 return -1; 753 } 754 755 if (rqpair->qpair.poll_group) { 756 assert(!rqpair->cq); 757 rc = nvme_rdma_qpair_set_poller(&rqpair->qpair); 758 if (rc) { 759 SPDK_ERRLOG("Unable to activate the rdmaqpair.\n"); 760 return -1; 761 } 762 assert(rqpair->cq); 763 } else { 764 num_cqe = rqpair->num_entries * 2; 765 max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size; 766 if (max_num_cqe != 0 && num_cqe > max_num_cqe) { 767 num_cqe = max_num_cqe; 768 } 769 rqpair->cq = ibv_create_cq(rqpair->cm_id->verbs, num_cqe, rqpair, NULL, 0); 770 if (!rqpair->cq) { 771 SPDK_ERRLOG("Unable to create completion queue: errno %d: %s\n", errno, spdk_strerror(errno)); 772 return -1; 773 } 774 } 775 776 rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr); 777 if (g_nvme_hooks.get_ibv_pd) { 778 attr.pd = g_nvme_hooks.get_ibv_pd(&rctrlr->ctrlr.trid, rqpair->cm_id->verbs); 779 } else { 780 attr.pd = spdk_rdma_get_pd(rqpair->cm_id->verbs); 781 } 782 783 attr.stats = rqpair->poller ? &rqpair->poller->stats.rdma_stats : NULL; 784 attr.send_cq = rqpair->cq; 785 attr.recv_cq = rqpair->cq; 786 attr.cap.max_send_wr = rqpair->num_entries; /* SEND operations */ 787 if (rqpair->srq) { 788 attr.srq = rqpair->srq->srq; 789 } else { 790 attr.cap.max_recv_wr = rqpair->num_entries; /* RECV operations */ 791 } 792 attr.cap.max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, dev_attr.max_sge); 793 attr.cap.max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, dev_attr.max_sge); 794 795 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &attr); 796 797 if (!rqpair->rdma_qp) { 798 return -1; 799 } 800 801 rqpair->memory_domain = nvme_rdma_get_memory_domain(rqpair->rdma_qp->qp->pd); 802 if (!rqpair->memory_domain) { 803 SPDK_ERRLOG("Failed to get memory domain\n"); 804 return -1; 805 } 806 807 /* ibv_create_qp will change the values in attr.cap. Make sure we store the proper value. */ 808 rqpair->max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, attr.cap.max_send_sge); 809 rqpair->max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, attr.cap.max_recv_sge); 810 rqpair->current_num_sends = 0; 811 812 rqpair->cm_id->context = rqpair; 813 814 return 0; 815 } 816 817 static void 818 nvme_rdma_reset_failed_sends(struct nvme_rdma_qpair *rqpair, 819 struct ibv_send_wr *bad_send_wr, int rc) 820 { 821 SPDK_ERRLOG("Failed to post WRs on send queue, errno %d (%s), bad_wr %p\n", 822 rc, spdk_strerror(rc), bad_send_wr); 823 while (bad_send_wr != NULL) { 824 assert(rqpair->current_num_sends > 0); 825 rqpair->current_num_sends--; 826 bad_send_wr = bad_send_wr->next; 827 } 828 } 829 830 static void 831 nvme_rdma_reset_failed_recvs(struct nvme_rdma_rsps *rsps, 832 struct ibv_recv_wr *bad_recv_wr, int rc) 833 { 834 SPDK_ERRLOG("Failed to post WRs on receive queue, errno %d (%s), bad_wr %p\n", 835 rc, spdk_strerror(rc), bad_recv_wr); 836 while (bad_recv_wr != NULL) { 837 assert(rsps->current_num_recvs > 0); 838 rsps->current_num_recvs--; 839 bad_recv_wr = bad_recv_wr->next; 840 } 841 } 842 843 static inline int 844 nvme_rdma_qpair_submit_sends(struct nvme_rdma_qpair *rqpair) 845 { 846 struct ibv_send_wr *bad_send_wr = NULL; 847 int rc; 848 849 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_send_wr); 850 851 if (spdk_unlikely(rc)) { 852 nvme_rdma_reset_failed_sends(rqpair, bad_send_wr, rc); 853 } 854 855 return rc; 856 } 857 858 static inline int 859 nvme_rdma_qpair_submit_recvs(struct nvme_rdma_qpair *rqpair) 860 { 861 struct ibv_recv_wr *bad_recv_wr; 862 int rc = 0; 863 864 rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 865 if (spdk_unlikely(rc)) { 866 nvme_rdma_reset_failed_recvs(rqpair->rsps, bad_recv_wr, rc); 867 } 868 869 return rc; 870 } 871 872 static inline int 873 nvme_rdma_poller_submit_recvs(struct nvme_rdma_poller *poller) 874 { 875 struct ibv_recv_wr *bad_recv_wr; 876 int rc; 877 878 rc = spdk_rdma_srq_flush_recv_wrs(poller->srq, &bad_recv_wr); 879 if (spdk_unlikely(rc)) { 880 nvme_rdma_reset_failed_recvs(poller->rsps, bad_recv_wr, rc); 881 } 882 883 return rc; 884 } 885 886 #define nvme_rdma_trace_ibv_sge(sg_list) \ 887 if (sg_list) { \ 888 SPDK_DEBUGLOG(nvme, "local addr %p length 0x%x lkey 0x%x\n", \ 889 (void *)(sg_list)->addr, (sg_list)->length, (sg_list)->lkey); \ 890 } 891 892 static void 893 nvme_rdma_free_rsps(struct nvme_rdma_rsps *rsps) 894 { 895 if (!rsps) { 896 return; 897 } 898 899 spdk_free(rsps->rsps); 900 spdk_free(rsps->rsp_sgls); 901 spdk_free(rsps->rsp_recv_wrs); 902 spdk_free(rsps); 903 } 904 905 static struct nvme_rdma_rsps * 906 nvme_rdma_create_rsps(struct nvme_rdma_rsp_opts *opts) 907 { 908 struct nvme_rdma_rsps *rsps; 909 struct spdk_rdma_memory_translation translation; 910 uint16_t i; 911 int rc; 912 913 rsps = spdk_zmalloc(sizeof(*rsps), 0, NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 914 if (!rsps) { 915 SPDK_ERRLOG("Failed to allocate rsps object\n"); 916 return NULL; 917 } 918 919 rsps->rsp_sgls = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_sgls), 0, NULL, 920 SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 921 if (!rsps->rsp_sgls) { 922 SPDK_ERRLOG("Failed to allocate rsp_sgls\n"); 923 goto fail; 924 } 925 926 rsps->rsp_recv_wrs = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_recv_wrs), 0, NULL, 927 SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 928 if (!rsps->rsp_recv_wrs) { 929 SPDK_ERRLOG("Failed to allocate rsp_recv_wrs\n"); 930 goto fail; 931 } 932 933 rsps->rsps = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsps), 0, NULL, 934 SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 935 if (!rsps->rsps) { 936 SPDK_ERRLOG("can not allocate rdma rsps\n"); 937 goto fail; 938 } 939 940 for (i = 0; i < opts->num_entries; i++) { 941 struct ibv_sge *rsp_sgl = &rsps->rsp_sgls[i]; 942 struct spdk_nvme_rdma_rsp *rsp = &rsps->rsps[i]; 943 struct ibv_recv_wr *recv_wr = &rsps->rsp_recv_wrs[i]; 944 945 rsp->rqpair = opts->rqpair; 946 rsp->rdma_wr.type = RDMA_WR_TYPE_RECV; 947 rsp->recv_wr = recv_wr; 948 rsp_sgl->addr = (uint64_t)rsp; 949 rsp_sgl->length = sizeof(struct spdk_nvme_cpl); 950 rc = spdk_rdma_get_translation(opts->mr_map, rsp, sizeof(*rsp), &translation); 951 if (rc) { 952 goto fail; 953 } 954 rsp_sgl->lkey = spdk_rdma_memory_translation_get_lkey(&translation); 955 956 recv_wr->wr_id = (uint64_t)&rsp->rdma_wr; 957 recv_wr->next = NULL; 958 recv_wr->sg_list = rsp_sgl; 959 recv_wr->num_sge = 1; 960 961 nvme_rdma_trace_ibv_sge(recv_wr->sg_list); 962 963 if (opts->rqpair) { 964 spdk_rdma_qp_queue_recv_wrs(opts->rqpair->rdma_qp, recv_wr); 965 } else { 966 spdk_rdma_srq_queue_recv_wrs(opts->srq, recv_wr); 967 } 968 } 969 970 rsps->num_entries = opts->num_entries; 971 rsps->current_num_recvs = opts->num_entries; 972 973 return rsps; 974 fail: 975 nvme_rdma_free_rsps(rsps); 976 return NULL; 977 } 978 979 static void 980 nvme_rdma_free_reqs(struct nvme_rdma_qpair *rqpair) 981 { 982 if (!rqpair->rdma_reqs) { 983 return; 984 } 985 986 spdk_free(rqpair->cmds); 987 rqpair->cmds = NULL; 988 989 spdk_free(rqpair->rdma_reqs); 990 rqpair->rdma_reqs = NULL; 991 } 992 993 static int 994 nvme_rdma_create_reqs(struct nvme_rdma_qpair *rqpair) 995 { 996 struct spdk_rdma_memory_translation translation; 997 uint16_t i; 998 int rc; 999 1000 assert(!rqpair->rdma_reqs); 1001 rqpair->rdma_reqs = spdk_zmalloc(rqpair->num_entries * sizeof(struct spdk_nvme_rdma_req), 0, NULL, 1002 SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 1003 if (rqpair->rdma_reqs == NULL) { 1004 SPDK_ERRLOG("Failed to allocate rdma_reqs\n"); 1005 goto fail; 1006 } 1007 1008 assert(!rqpair->cmds); 1009 rqpair->cmds = spdk_zmalloc(rqpair->num_entries * sizeof(*rqpair->cmds), 0, NULL, 1010 SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 1011 if (!rqpair->cmds) { 1012 SPDK_ERRLOG("Failed to allocate RDMA cmds\n"); 1013 goto fail; 1014 } 1015 1016 TAILQ_INIT(&rqpair->free_reqs); 1017 TAILQ_INIT(&rqpair->outstanding_reqs); 1018 for (i = 0; i < rqpair->num_entries; i++) { 1019 struct spdk_nvme_rdma_req *rdma_req; 1020 struct spdk_nvmf_cmd *cmd; 1021 1022 rdma_req = &rqpair->rdma_reqs[i]; 1023 rdma_req->rdma_wr.type = RDMA_WR_TYPE_SEND; 1024 cmd = &rqpair->cmds[i]; 1025 1026 rdma_req->id = i; 1027 1028 rc = spdk_rdma_get_translation(rqpair->mr_map, cmd, sizeof(*cmd), &translation); 1029 if (rc) { 1030 goto fail; 1031 } 1032 rdma_req->send_sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 1033 1034 /* The first RDMA sgl element will always point 1035 * at this data structure. Depending on whether 1036 * an NVMe-oF SGL is required, the length of 1037 * this element may change. */ 1038 rdma_req->send_sgl[0].addr = (uint64_t)cmd; 1039 rdma_req->send_wr.wr_id = (uint64_t)&rdma_req->rdma_wr; 1040 rdma_req->send_wr.next = NULL; 1041 rdma_req->send_wr.opcode = IBV_WR_SEND; 1042 rdma_req->send_wr.send_flags = IBV_SEND_SIGNALED; 1043 rdma_req->send_wr.sg_list = rdma_req->send_sgl; 1044 rdma_req->send_wr.imm_data = 0; 1045 1046 TAILQ_INSERT_TAIL(&rqpair->free_reqs, rdma_req, link); 1047 } 1048 1049 return 0; 1050 fail: 1051 nvme_rdma_free_reqs(rqpair); 1052 return -ENOMEM; 1053 } 1054 1055 static int nvme_rdma_connect(struct nvme_rdma_qpair *rqpair); 1056 1057 static int 1058 nvme_rdma_route_resolved(struct nvme_rdma_qpair *rqpair, int ret) 1059 { 1060 if (ret) { 1061 SPDK_ERRLOG("RDMA route resolution error\n"); 1062 return -1; 1063 } 1064 1065 ret = nvme_rdma_qpair_init(rqpair); 1066 if (ret < 0) { 1067 SPDK_ERRLOG("nvme_rdma_qpair_init() failed\n"); 1068 return -1; 1069 } 1070 1071 return nvme_rdma_connect(rqpair); 1072 } 1073 1074 static int 1075 nvme_rdma_addr_resolved(struct nvme_rdma_qpair *rqpair, int ret) 1076 { 1077 if (ret) { 1078 SPDK_ERRLOG("RDMA address resolution error\n"); 1079 return -1; 1080 } 1081 1082 if (rqpair->qpair.ctrlr->opts.transport_ack_timeout != SPDK_NVME_TRANSPORT_ACK_TIMEOUT_DISABLED) { 1083 #ifdef SPDK_CONFIG_RDMA_SET_ACK_TIMEOUT 1084 uint8_t timeout = rqpair->qpair.ctrlr->opts.transport_ack_timeout; 1085 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, 1086 RDMA_OPTION_ID_ACK_TIMEOUT, 1087 &timeout, sizeof(timeout)); 1088 if (ret) { 1089 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_ACK_TIMEOUT %d, ret %d\n", timeout, ret); 1090 } 1091 #else 1092 SPDK_DEBUGLOG(nvme, "transport_ack_timeout is not supported\n"); 1093 #endif 1094 } 1095 1096 if (rqpair->qpair.ctrlr->opts.transport_tos != SPDK_NVME_TRANSPORT_TOS_DISABLED) { 1097 #ifdef SPDK_CONFIG_RDMA_SET_TOS 1098 uint8_t tos = rqpair->qpair.ctrlr->opts.transport_tos; 1099 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_TOS, &tos, sizeof(tos)); 1100 if (ret) { 1101 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_TOS %u, ret %d\n", tos, ret); 1102 } 1103 #else 1104 SPDK_DEBUGLOG(nvme, "transport_tos is not supported\n"); 1105 #endif 1106 } 1107 1108 ret = rdma_resolve_route(rqpair->cm_id, NVME_RDMA_TIME_OUT_IN_MS); 1109 if (ret) { 1110 SPDK_ERRLOG("rdma_resolve_route\n"); 1111 return ret; 1112 } 1113 1114 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ROUTE_RESOLVED, 1115 nvme_rdma_route_resolved); 1116 } 1117 1118 static int 1119 nvme_rdma_resolve_addr(struct nvme_rdma_qpair *rqpair, 1120 struct sockaddr *src_addr, 1121 struct sockaddr *dst_addr) 1122 { 1123 int ret; 1124 1125 if (src_addr) { 1126 int reuse = 1; 1127 1128 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_REUSEADDR, 1129 &reuse, sizeof(reuse)); 1130 if (ret) { 1131 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_REUSEADDR %d, ret %d\n", 1132 reuse, ret); 1133 /* It is likely that rdma_resolve_addr() returns -EADDRINUSE, but 1134 * we may missing something. We rely on rdma_resolve_addr(). 1135 */ 1136 } 1137 } 1138 1139 ret = rdma_resolve_addr(rqpair->cm_id, src_addr, dst_addr, 1140 NVME_RDMA_TIME_OUT_IN_MS); 1141 if (ret) { 1142 SPDK_ERRLOG("rdma_resolve_addr, %d\n", errno); 1143 return ret; 1144 } 1145 1146 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ADDR_RESOLVED, 1147 nvme_rdma_addr_resolved); 1148 } 1149 1150 static int nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair); 1151 1152 static int 1153 nvme_rdma_connect_established(struct nvme_rdma_qpair *rqpair, int ret) 1154 { 1155 struct nvme_rdma_rsp_opts opts = {}; 1156 1157 if (ret == -ESTALE) { 1158 return nvme_rdma_stale_conn_retry(rqpair); 1159 } else if (ret) { 1160 SPDK_ERRLOG("RDMA connect error %d\n", ret); 1161 return ret; 1162 } 1163 1164 assert(!rqpair->mr_map); 1165 rqpair->mr_map = spdk_rdma_create_mem_map(rqpair->rdma_qp->qp->pd, &g_nvme_hooks, 1166 SPDK_RDMA_MEMORY_MAP_ROLE_INITIATOR); 1167 if (!rqpair->mr_map) { 1168 SPDK_ERRLOG("Unable to register RDMA memory translation map\n"); 1169 return -1; 1170 } 1171 1172 ret = nvme_rdma_create_reqs(rqpair); 1173 SPDK_DEBUGLOG(nvme, "rc =%d\n", ret); 1174 if (ret) { 1175 SPDK_ERRLOG("Unable to create rqpair RDMA requests\n"); 1176 return -1; 1177 } 1178 SPDK_DEBUGLOG(nvme, "RDMA requests created\n"); 1179 1180 if (!rqpair->srq) { 1181 opts.num_entries = rqpair->num_entries; 1182 opts.rqpair = rqpair; 1183 opts.srq = NULL; 1184 opts.mr_map = rqpair->mr_map; 1185 1186 assert(!rqpair->rsps); 1187 rqpair->rsps = nvme_rdma_create_rsps(&opts); 1188 if (!rqpair->rsps) { 1189 SPDK_ERRLOG("Unable to create rqpair RDMA responses\n"); 1190 return -1; 1191 } 1192 SPDK_DEBUGLOG(nvme, "RDMA responses created\n"); 1193 1194 ret = nvme_rdma_qpair_submit_recvs(rqpair); 1195 SPDK_DEBUGLOG(nvme, "rc =%d\n", ret); 1196 if (ret) { 1197 SPDK_ERRLOG("Unable to submit rqpair RDMA responses\n"); 1198 return -1; 1199 } 1200 SPDK_DEBUGLOG(nvme, "RDMA responses submitted\n"); 1201 } 1202 1203 rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND; 1204 1205 return 0; 1206 } 1207 1208 static int 1209 nvme_rdma_connect(struct nvme_rdma_qpair *rqpair) 1210 { 1211 struct rdma_conn_param param = {}; 1212 struct spdk_nvmf_rdma_request_private_data request_data = {}; 1213 struct ibv_device_attr attr; 1214 int ret; 1215 struct spdk_nvme_ctrlr *ctrlr; 1216 1217 ret = ibv_query_device(rqpair->cm_id->verbs, &attr); 1218 if (ret != 0) { 1219 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1220 return ret; 1221 } 1222 1223 param.responder_resources = attr.max_qp_rd_atom; 1224 1225 ctrlr = rqpair->qpair.ctrlr; 1226 if (!ctrlr) { 1227 return -1; 1228 } 1229 1230 request_data.qid = rqpair->qpair.id; 1231 request_data.hrqsize = rqpair->num_entries + 1; 1232 request_data.hsqsize = rqpair->num_entries; 1233 request_data.cntlid = ctrlr->cntlid; 1234 1235 param.private_data = &request_data; 1236 param.private_data_len = sizeof(request_data); 1237 param.retry_count = ctrlr->opts.transport_retry_count; 1238 param.rnr_retry_count = 7; 1239 1240 /* Fields below are ignored by rdma cm if qpair has been 1241 * created using rdma cm API. */ 1242 param.srq = 0; 1243 param.qp_num = rqpair->rdma_qp->qp->qp_num; 1244 1245 ret = rdma_connect(rqpair->cm_id, ¶m); 1246 if (ret) { 1247 SPDK_ERRLOG("nvme rdma connect error\n"); 1248 return ret; 1249 } 1250 1251 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ESTABLISHED, 1252 nvme_rdma_connect_established); 1253 } 1254 1255 static int 1256 nvme_rdma_ctrlr_connect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 1257 { 1258 struct sockaddr_storage dst_addr; 1259 struct sockaddr_storage src_addr; 1260 bool src_addr_specified; 1261 long int port, src_port; 1262 int rc; 1263 struct nvme_rdma_ctrlr *rctrlr; 1264 struct nvme_rdma_qpair *rqpair; 1265 struct nvme_rdma_poll_group *group; 1266 int family; 1267 1268 rqpair = nvme_rdma_qpair(qpair); 1269 rctrlr = nvme_rdma_ctrlr(ctrlr); 1270 assert(rctrlr != NULL); 1271 1272 switch (ctrlr->trid.adrfam) { 1273 case SPDK_NVMF_ADRFAM_IPV4: 1274 family = AF_INET; 1275 break; 1276 case SPDK_NVMF_ADRFAM_IPV6: 1277 family = AF_INET6; 1278 break; 1279 default: 1280 SPDK_ERRLOG("Unhandled ADRFAM %d\n", ctrlr->trid.adrfam); 1281 return -1; 1282 } 1283 1284 SPDK_DEBUGLOG(nvme, "adrfam %d ai_family %d\n", ctrlr->trid.adrfam, family); 1285 1286 memset(&dst_addr, 0, sizeof(dst_addr)); 1287 1288 SPDK_DEBUGLOG(nvme, "trsvcid is %s\n", ctrlr->trid.trsvcid); 1289 rc = nvme_parse_addr(&dst_addr, family, ctrlr->trid.traddr, ctrlr->trid.trsvcid, &port); 1290 if (rc != 0) { 1291 SPDK_ERRLOG("dst_addr nvme_parse_addr() failed\n"); 1292 return -1; 1293 } 1294 1295 if (ctrlr->opts.src_addr[0] || ctrlr->opts.src_svcid[0]) { 1296 memset(&src_addr, 0, sizeof(src_addr)); 1297 rc = nvme_parse_addr(&src_addr, family, ctrlr->opts.src_addr, ctrlr->opts.src_svcid, &src_port); 1298 if (rc != 0) { 1299 SPDK_ERRLOG("src_addr nvme_parse_addr() failed\n"); 1300 return -1; 1301 } 1302 src_addr_specified = true; 1303 } else { 1304 src_addr_specified = false; 1305 } 1306 1307 rc = rdma_create_id(rctrlr->cm_channel, &rqpair->cm_id, rqpair, RDMA_PS_TCP); 1308 if (rc < 0) { 1309 SPDK_ERRLOG("rdma_create_id() failed\n"); 1310 return -1; 1311 } 1312 1313 rc = nvme_rdma_resolve_addr(rqpair, 1314 src_addr_specified ? (struct sockaddr *)&src_addr : NULL, 1315 (struct sockaddr *)&dst_addr); 1316 if (rc < 0) { 1317 SPDK_ERRLOG("nvme_rdma_resolve_addr() failed\n"); 1318 return -1; 1319 } 1320 1321 rqpair->state = NVME_RDMA_QPAIR_STATE_INITIALIZING; 1322 1323 if (qpair->poll_group != NULL) { 1324 group = nvme_rdma_poll_group(qpair->poll_group); 1325 TAILQ_INSERT_TAIL(&group->connecting_qpairs, rqpair, link_connecting); 1326 } 1327 1328 return 0; 1329 } 1330 1331 static int 1332 nvme_rdma_stale_conn_reconnect(struct nvme_rdma_qpair *rqpair) 1333 { 1334 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 1335 1336 if (spdk_get_ticks() < rqpair->evt_timeout_ticks) { 1337 return -EAGAIN; 1338 } 1339 1340 return nvme_rdma_ctrlr_connect_qpair(qpair->ctrlr, qpair); 1341 } 1342 1343 static int 1344 nvme_rdma_ctrlr_connect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, 1345 struct spdk_nvme_qpair *qpair) 1346 { 1347 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 1348 int rc; 1349 1350 if (rqpair->in_connect_poll) { 1351 return -EAGAIN; 1352 } 1353 1354 rqpair->in_connect_poll = true; 1355 1356 switch (rqpair->state) { 1357 case NVME_RDMA_QPAIR_STATE_INVALID: 1358 rc = -EAGAIN; 1359 break; 1360 1361 case NVME_RDMA_QPAIR_STATE_INITIALIZING: 1362 case NVME_RDMA_QPAIR_STATE_EXITING: 1363 if (!nvme_qpair_is_admin_queue(qpair)) { 1364 nvme_robust_mutex_lock(&ctrlr->ctrlr_lock); 1365 } 1366 1367 rc = nvme_rdma_process_event_poll(rqpair); 1368 1369 if (!nvme_qpair_is_admin_queue(qpair)) { 1370 nvme_robust_mutex_unlock(&ctrlr->ctrlr_lock); 1371 } 1372 1373 if (rc == 0) { 1374 rc = -EAGAIN; 1375 } 1376 rqpair->in_connect_poll = false; 1377 1378 return rc; 1379 1380 case NVME_RDMA_QPAIR_STATE_STALE_CONN: 1381 rc = nvme_rdma_stale_conn_reconnect(rqpair); 1382 if (rc == 0) { 1383 rc = -EAGAIN; 1384 } 1385 break; 1386 case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND: 1387 rc = nvme_fabric_qpair_connect_async(qpair, rqpair->num_entries + 1); 1388 if (rc == 0) { 1389 rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL; 1390 rc = -EAGAIN; 1391 } else { 1392 SPDK_ERRLOG("Failed to send an NVMe-oF Fabric CONNECT command\n"); 1393 } 1394 break; 1395 case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL: 1396 rc = nvme_fabric_qpair_connect_poll(qpair); 1397 if (rc == 0) { 1398 rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING; 1399 nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED); 1400 } else if (rc != -EAGAIN) { 1401 SPDK_ERRLOG("Failed to poll NVMe-oF Fabric CONNECT command\n"); 1402 } 1403 break; 1404 case NVME_RDMA_QPAIR_STATE_RUNNING: 1405 rc = 0; 1406 break; 1407 default: 1408 assert(false); 1409 rc = -EINVAL; 1410 break; 1411 } 1412 1413 rqpair->in_connect_poll = false; 1414 1415 return rc; 1416 } 1417 1418 static inline int 1419 nvme_rdma_get_memory_translation(struct nvme_request *req, struct nvme_rdma_qpair *rqpair, 1420 struct nvme_rdma_memory_translation_ctx *_ctx) 1421 { 1422 struct spdk_memory_domain_translation_ctx ctx; 1423 struct spdk_memory_domain_translation_result dma_translation = {.iov_count = 0}; 1424 struct spdk_rdma_memory_translation rdma_translation; 1425 int rc; 1426 1427 assert(req); 1428 assert(rqpair); 1429 assert(_ctx); 1430 1431 if (req->payload.opts && req->payload.opts->memory_domain) { 1432 ctx.size = sizeof(struct spdk_memory_domain_translation_ctx); 1433 ctx.rdma.ibv_qp = rqpair->rdma_qp->qp; 1434 dma_translation.size = sizeof(struct spdk_memory_domain_translation_result); 1435 1436 rc = spdk_memory_domain_translate_data(req->payload.opts->memory_domain, 1437 req->payload.opts->memory_domain_ctx, 1438 rqpair->memory_domain->domain, &ctx, _ctx->addr, 1439 _ctx->length, &dma_translation); 1440 if (spdk_unlikely(rc) || dma_translation.iov_count != 1) { 1441 SPDK_ERRLOG("DMA memory translation failed, rc %d, iov count %u\n", rc, dma_translation.iov_count); 1442 return rc; 1443 } 1444 1445 _ctx->lkey = dma_translation.rdma.lkey; 1446 _ctx->rkey = dma_translation.rdma.rkey; 1447 _ctx->addr = dma_translation.iov.iov_base; 1448 _ctx->length = dma_translation.iov.iov_len; 1449 } else { 1450 rc = spdk_rdma_get_translation(rqpair->mr_map, _ctx->addr, _ctx->length, &rdma_translation); 1451 if (spdk_unlikely(rc)) { 1452 SPDK_ERRLOG("RDMA memory translation failed, rc %d\n", rc); 1453 return rc; 1454 } 1455 if (rdma_translation.translation_type == SPDK_RDMA_TRANSLATION_MR) { 1456 _ctx->lkey = rdma_translation.mr_or_key.mr->lkey; 1457 _ctx->rkey = rdma_translation.mr_or_key.mr->rkey; 1458 } else { 1459 _ctx->lkey = _ctx->rkey = (uint32_t)rdma_translation.mr_or_key.key; 1460 } 1461 } 1462 1463 return 0; 1464 } 1465 1466 1467 /* 1468 * Build SGL describing empty payload. 1469 */ 1470 static int 1471 nvme_rdma_build_null_request(struct spdk_nvme_rdma_req *rdma_req) 1472 { 1473 struct nvme_request *req = rdma_req->req; 1474 1475 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1476 1477 /* The first element of this SGL is pointing at an 1478 * spdk_nvmf_cmd object. For this particular command, 1479 * we only need the first 64 bytes corresponding to 1480 * the NVMe command. */ 1481 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1482 1483 /* The RDMA SGL needs one element describing the NVMe command. */ 1484 rdma_req->send_wr.num_sge = 1; 1485 1486 req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1487 req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1488 req->cmd.dptr.sgl1.keyed.length = 0; 1489 req->cmd.dptr.sgl1.keyed.key = 0; 1490 req->cmd.dptr.sgl1.address = 0; 1491 1492 return 0; 1493 } 1494 1495 /* 1496 * Build inline SGL describing contiguous payload buffer. 1497 */ 1498 static int 1499 nvme_rdma_build_contig_inline_request(struct nvme_rdma_qpair *rqpair, 1500 struct spdk_nvme_rdma_req *rdma_req) 1501 { 1502 struct nvme_request *req = rdma_req->req; 1503 struct nvme_rdma_memory_translation_ctx ctx = { 1504 .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset, 1505 .length = req->payload_size 1506 }; 1507 int rc; 1508 1509 assert(ctx.length != 0); 1510 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG); 1511 1512 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1513 if (spdk_unlikely(rc)) { 1514 return -1; 1515 } 1516 1517 rdma_req->send_sgl[1].lkey = ctx.lkey; 1518 1519 /* The first element of this SGL is pointing at an 1520 * spdk_nvmf_cmd object. For this particular command, 1521 * we only need the first 64 bytes corresponding to 1522 * the NVMe command. */ 1523 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1524 1525 rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr; 1526 rdma_req->send_sgl[1].length = (uint32_t)ctx.length; 1527 1528 /* The RDMA SGL contains two elements. The first describes 1529 * the NVMe command and the second describes the data 1530 * payload. */ 1531 rdma_req->send_wr.num_sge = 2; 1532 1533 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1534 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK; 1535 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1536 req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length; 1537 /* Inline only supported for icdoff == 0 currently. This function will 1538 * not get called for controllers with other values. */ 1539 req->cmd.dptr.sgl1.address = (uint64_t)0; 1540 1541 return 0; 1542 } 1543 1544 /* 1545 * Build SGL describing contiguous payload buffer. 1546 */ 1547 static int 1548 nvme_rdma_build_contig_request(struct nvme_rdma_qpair *rqpair, 1549 struct spdk_nvme_rdma_req *rdma_req) 1550 { 1551 struct nvme_request *req = rdma_req->req; 1552 struct nvme_rdma_memory_translation_ctx ctx = { 1553 .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset, 1554 .length = req->payload_size 1555 }; 1556 int rc; 1557 1558 assert(req->payload_size != 0); 1559 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG); 1560 1561 if (spdk_unlikely(req->payload_size > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) { 1562 SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n", 1563 req->payload_size, NVME_RDMA_MAX_KEYED_SGL_LENGTH); 1564 return -1; 1565 } 1566 1567 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1568 if (spdk_unlikely(rc)) { 1569 return -1; 1570 } 1571 1572 req->cmd.dptr.sgl1.keyed.key = ctx.rkey; 1573 1574 /* The first element of this SGL is pointing at an 1575 * spdk_nvmf_cmd object. For this particular command, 1576 * we only need the first 64 bytes corresponding to 1577 * the NVMe command. */ 1578 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1579 1580 /* The RDMA SGL needs one element describing the NVMe command. */ 1581 rdma_req->send_wr.num_sge = 1; 1582 1583 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1584 req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1585 req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1586 req->cmd.dptr.sgl1.keyed.length = (uint32_t)ctx.length; 1587 req->cmd.dptr.sgl1.address = (uint64_t)ctx.addr; 1588 1589 return 0; 1590 } 1591 1592 /* 1593 * Build SGL describing scattered payload buffer. 1594 */ 1595 static int 1596 nvme_rdma_build_sgl_request(struct nvme_rdma_qpair *rqpair, 1597 struct spdk_nvme_rdma_req *rdma_req) 1598 { 1599 struct nvme_request *req = rdma_req->req; 1600 struct spdk_nvmf_cmd *cmd = &rqpair->cmds[rdma_req->id]; 1601 struct nvme_rdma_memory_translation_ctx ctx; 1602 uint32_t remaining_size; 1603 uint32_t sge_length; 1604 int rc, max_num_sgl, num_sgl_desc; 1605 1606 assert(req->payload_size != 0); 1607 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL); 1608 assert(req->payload.reset_sgl_fn != NULL); 1609 assert(req->payload.next_sge_fn != NULL); 1610 req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset); 1611 1612 max_num_sgl = req->qpair->ctrlr->max_sges; 1613 1614 remaining_size = req->payload_size; 1615 num_sgl_desc = 0; 1616 do { 1617 rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &sge_length); 1618 if (rc) { 1619 return -1; 1620 } 1621 1622 sge_length = spdk_min(remaining_size, sge_length); 1623 1624 if (spdk_unlikely(sge_length > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) { 1625 SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n", 1626 sge_length, NVME_RDMA_MAX_KEYED_SGL_LENGTH); 1627 return -1; 1628 } 1629 ctx.length = sge_length; 1630 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1631 if (spdk_unlikely(rc)) { 1632 return -1; 1633 } 1634 1635 cmd->sgl[num_sgl_desc].keyed.key = ctx.rkey; 1636 cmd->sgl[num_sgl_desc].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1637 cmd->sgl[num_sgl_desc].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1638 cmd->sgl[num_sgl_desc].keyed.length = (uint32_t)ctx.length; 1639 cmd->sgl[num_sgl_desc].address = (uint64_t)ctx.addr; 1640 1641 remaining_size -= ctx.length; 1642 num_sgl_desc++; 1643 } while (remaining_size > 0 && num_sgl_desc < max_num_sgl); 1644 1645 1646 /* Should be impossible if we did our sgl checks properly up the stack, but do a sanity check here. */ 1647 if (remaining_size > 0) { 1648 return -1; 1649 } 1650 1651 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1652 1653 /* The RDMA SGL needs one element describing some portion 1654 * of the spdk_nvmf_cmd structure. */ 1655 rdma_req->send_wr.num_sge = 1; 1656 1657 /* 1658 * If only one SGL descriptor is required, it can be embedded directly in the command 1659 * as a data block descriptor. 1660 */ 1661 if (num_sgl_desc == 1) { 1662 /* The first element of this SGL is pointing at an 1663 * spdk_nvmf_cmd object. For this particular command, 1664 * we only need the first 64 bytes corresponding to 1665 * the NVMe command. */ 1666 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1667 1668 req->cmd.dptr.sgl1.keyed.type = cmd->sgl[0].keyed.type; 1669 req->cmd.dptr.sgl1.keyed.subtype = cmd->sgl[0].keyed.subtype; 1670 req->cmd.dptr.sgl1.keyed.length = cmd->sgl[0].keyed.length; 1671 req->cmd.dptr.sgl1.keyed.key = cmd->sgl[0].keyed.key; 1672 req->cmd.dptr.sgl1.address = cmd->sgl[0].address; 1673 } else { 1674 /* 1675 * Otherwise, The SGL descriptor embedded in the command must point to the list of 1676 * SGL descriptors used to describe the operation. In that case it is a last segment descriptor. 1677 */ 1678 uint32_t descriptors_size = sizeof(struct spdk_nvme_sgl_descriptor) * num_sgl_desc; 1679 1680 if (spdk_unlikely(descriptors_size > rqpair->qpair.ctrlr->ioccsz_bytes)) { 1681 SPDK_ERRLOG("Size of SGL descriptors (%u) exceeds ICD (%u)\n", 1682 descriptors_size, rqpair->qpair.ctrlr->ioccsz_bytes); 1683 return -1; 1684 } 1685 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd) + descriptors_size; 1686 1687 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT; 1688 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1689 req->cmd.dptr.sgl1.unkeyed.length = descriptors_size; 1690 req->cmd.dptr.sgl1.address = (uint64_t)0; 1691 } 1692 1693 return 0; 1694 } 1695 1696 /* 1697 * Build inline SGL describing sgl payload buffer. 1698 */ 1699 static int 1700 nvme_rdma_build_sgl_inline_request(struct nvme_rdma_qpair *rqpair, 1701 struct spdk_nvme_rdma_req *rdma_req) 1702 { 1703 struct nvme_request *req = rdma_req->req; 1704 struct nvme_rdma_memory_translation_ctx ctx; 1705 uint32_t length; 1706 int rc; 1707 1708 assert(req->payload_size != 0); 1709 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL); 1710 assert(req->payload.reset_sgl_fn != NULL); 1711 assert(req->payload.next_sge_fn != NULL); 1712 req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset); 1713 1714 rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &length); 1715 if (rc) { 1716 return -1; 1717 } 1718 1719 if (length < req->payload_size) { 1720 SPDK_DEBUGLOG(nvme, "Inline SGL request split so sending separately.\n"); 1721 return nvme_rdma_build_sgl_request(rqpair, rdma_req); 1722 } 1723 1724 if (length > req->payload_size) { 1725 length = req->payload_size; 1726 } 1727 1728 ctx.length = length; 1729 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1730 if (spdk_unlikely(rc)) { 1731 return -1; 1732 } 1733 1734 rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr; 1735 rdma_req->send_sgl[1].length = (uint32_t)ctx.length; 1736 rdma_req->send_sgl[1].lkey = ctx.lkey; 1737 1738 rdma_req->send_wr.num_sge = 2; 1739 1740 /* The first element of this SGL is pointing at an 1741 * spdk_nvmf_cmd object. For this particular command, 1742 * we only need the first 64 bytes corresponding to 1743 * the NVMe command. */ 1744 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1745 1746 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1747 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK; 1748 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1749 req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length; 1750 /* Inline only supported for icdoff == 0 currently. This function will 1751 * not get called for controllers with other values. */ 1752 req->cmd.dptr.sgl1.address = (uint64_t)0; 1753 1754 return 0; 1755 } 1756 1757 static int 1758 nvme_rdma_req_init(struct nvme_rdma_qpair *rqpair, struct nvme_request *req, 1759 struct spdk_nvme_rdma_req *rdma_req) 1760 { 1761 struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr; 1762 enum nvme_payload_type payload_type; 1763 bool icd_supported; 1764 int rc; 1765 1766 assert(rdma_req->req == NULL); 1767 rdma_req->req = req; 1768 req->cmd.cid = rdma_req->id; 1769 payload_type = nvme_payload_type(&req->payload); 1770 /* 1771 * Check if icdoff is non zero, to avoid interop conflicts with 1772 * targets with non-zero icdoff. Both SPDK and the Linux kernel 1773 * targets use icdoff = 0. For targets with non-zero icdoff, we 1774 * will currently just not use inline data for now. 1775 */ 1776 icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER 1777 && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0; 1778 1779 if (req->payload_size == 0) { 1780 rc = nvme_rdma_build_null_request(rdma_req); 1781 } else if (payload_type == NVME_PAYLOAD_TYPE_CONTIG) { 1782 if (icd_supported) { 1783 rc = nvme_rdma_build_contig_inline_request(rqpair, rdma_req); 1784 } else { 1785 rc = nvme_rdma_build_contig_request(rqpair, rdma_req); 1786 } 1787 } else if (payload_type == NVME_PAYLOAD_TYPE_SGL) { 1788 if (icd_supported) { 1789 rc = nvme_rdma_build_sgl_inline_request(rqpair, rdma_req); 1790 } else { 1791 rc = nvme_rdma_build_sgl_request(rqpair, rdma_req); 1792 } 1793 } else { 1794 rc = -1; 1795 } 1796 1797 if (rc) { 1798 rdma_req->req = NULL; 1799 return rc; 1800 } 1801 1802 memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd)); 1803 return 0; 1804 } 1805 1806 static struct spdk_nvme_qpair * 1807 nvme_rdma_ctrlr_create_qpair(struct spdk_nvme_ctrlr *ctrlr, 1808 uint16_t qid, uint32_t qsize, 1809 enum spdk_nvme_qprio qprio, 1810 uint32_t num_requests, 1811 bool delay_cmd_submit, 1812 bool async) 1813 { 1814 struct nvme_rdma_qpair *rqpair; 1815 struct spdk_nvme_qpair *qpair; 1816 int rc; 1817 1818 if (qsize < SPDK_NVME_QUEUE_MIN_ENTRIES) { 1819 SPDK_ERRLOG("Failed to create qpair with size %u. Minimum queue size is %d.\n", 1820 qsize, SPDK_NVME_QUEUE_MIN_ENTRIES); 1821 return NULL; 1822 } 1823 1824 rqpair = spdk_zmalloc(sizeof(struct nvme_rdma_qpair), 0, NULL, SPDK_ENV_SOCKET_ID_ANY, 1825 SPDK_MALLOC_DMA); 1826 if (!rqpair) { 1827 SPDK_ERRLOG("failed to get create rqpair\n"); 1828 return NULL; 1829 } 1830 1831 /* Set num_entries one less than queue size. According to NVMe 1832 * and NVMe-oF specs we can not submit queue size requests, 1833 * one slot shall always remain empty. 1834 */ 1835 rqpair->num_entries = qsize - 1; 1836 rqpair->delay_cmd_submit = delay_cmd_submit; 1837 rqpair->state = NVME_RDMA_QPAIR_STATE_INVALID; 1838 qpair = &rqpair->qpair; 1839 rc = nvme_qpair_init(qpair, qid, ctrlr, qprio, num_requests, async); 1840 if (rc != 0) { 1841 spdk_free(rqpair); 1842 return NULL; 1843 } 1844 1845 return qpair; 1846 } 1847 1848 static void 1849 nvme_rdma_qpair_destroy(struct nvme_rdma_qpair *rqpair) 1850 { 1851 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 1852 struct nvme_rdma_ctrlr *rctrlr; 1853 struct nvme_rdma_cm_event_entry *entry, *tmp; 1854 1855 spdk_rdma_free_mem_map(&rqpair->mr_map); 1856 1857 if (rqpair->evt) { 1858 rdma_ack_cm_event(rqpair->evt); 1859 rqpair->evt = NULL; 1860 } 1861 1862 /* 1863 * This works because we have the controller lock both in 1864 * this function and in the function where we add new events. 1865 */ 1866 if (qpair->ctrlr != NULL) { 1867 rctrlr = nvme_rdma_ctrlr(qpair->ctrlr); 1868 STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) { 1869 if (entry->evt->id->context == rqpair) { 1870 STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link); 1871 rdma_ack_cm_event(entry->evt); 1872 STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link); 1873 } 1874 } 1875 } 1876 1877 if (rqpair->cm_id) { 1878 if (rqpair->rdma_qp) { 1879 spdk_rdma_put_pd(rqpair->rdma_qp->qp->pd); 1880 spdk_rdma_qp_destroy(rqpair->rdma_qp); 1881 rqpair->rdma_qp = NULL; 1882 } 1883 } 1884 1885 if (rqpair->poller) { 1886 struct nvme_rdma_poll_group *group; 1887 1888 assert(qpair->poll_group); 1889 group = nvme_rdma_poll_group(qpair->poll_group); 1890 1891 nvme_rdma_poll_group_put_poller(group, rqpair->poller); 1892 1893 rqpair->poller = NULL; 1894 rqpair->cq = NULL; 1895 if (rqpair->srq) { 1896 rqpair->srq = NULL; 1897 rqpair->rsps = NULL; 1898 } 1899 } else if (rqpair->cq) { 1900 ibv_destroy_cq(rqpair->cq); 1901 rqpair->cq = NULL; 1902 } 1903 1904 nvme_rdma_free_reqs(rqpair); 1905 nvme_rdma_free_rsps(rqpair->rsps); 1906 rqpair->rsps = NULL; 1907 1908 /* destroy cm_id last so cma device will not be freed before we destroy the cq. */ 1909 if (rqpair->cm_id) { 1910 rdma_destroy_id(rqpair->cm_id); 1911 rqpair->cm_id = NULL; 1912 } 1913 } 1914 1915 static void nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr); 1916 1917 static int 1918 nvme_rdma_qpair_disconnected(struct nvme_rdma_qpair *rqpair, int ret) 1919 { 1920 if (ret) { 1921 SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n"); 1922 goto quiet; 1923 } 1924 1925 if (rqpair->poller == NULL) { 1926 /* If poller is not used, cq is not shared. 1927 * So complete disconnecting qpair immediately. 1928 */ 1929 goto quiet; 1930 } 1931 1932 if (rqpair->rsps == NULL) { 1933 goto quiet; 1934 } 1935 1936 if (rqpair->need_destroy || 1937 (rqpair->current_num_sends != 0 || 1938 (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) { 1939 rqpair->state = NVME_RDMA_QPAIR_STATE_LINGERING; 1940 rqpair->evt_timeout_ticks = (NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US * spdk_get_ticks_hz()) / 1941 SPDK_SEC_TO_USEC + spdk_get_ticks(); 1942 1943 return -EAGAIN; 1944 } 1945 1946 quiet: 1947 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED; 1948 1949 nvme_rdma_qpair_abort_reqs(&rqpair->qpair, 0); 1950 nvme_rdma_qpair_destroy(rqpair); 1951 nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair); 1952 1953 return 0; 1954 } 1955 1956 static int 1957 nvme_rdma_qpair_wait_until_quiet(struct nvme_rdma_qpair *rqpair) 1958 { 1959 if (spdk_get_ticks() < rqpair->evt_timeout_ticks && 1960 (rqpair->current_num_sends != 0 || 1961 (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) { 1962 return -EAGAIN; 1963 } 1964 1965 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED; 1966 1967 nvme_rdma_qpair_abort_reqs(&rqpair->qpair, 0); 1968 nvme_rdma_qpair_destroy(rqpair); 1969 nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair); 1970 1971 return 0; 1972 } 1973 1974 static void 1975 _nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair, 1976 nvme_rdma_cm_event_cb disconnected_qpair_cb) 1977 { 1978 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 1979 int rc; 1980 1981 assert(disconnected_qpair_cb != NULL); 1982 1983 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITING; 1984 1985 if (rqpair->cm_id) { 1986 if (rqpair->rdma_qp) { 1987 rc = spdk_rdma_qp_disconnect(rqpair->rdma_qp); 1988 if ((qpair->ctrlr != NULL) && (rc == 0)) { 1989 rc = nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_DISCONNECTED, 1990 disconnected_qpair_cb); 1991 if (rc == 0) { 1992 return; 1993 } 1994 } 1995 } 1996 } 1997 1998 disconnected_qpair_cb(rqpair, 0); 1999 } 2000 2001 static int 2002 nvme_rdma_ctrlr_disconnect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 2003 { 2004 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2005 int rc; 2006 2007 switch (rqpair->state) { 2008 case NVME_RDMA_QPAIR_STATE_EXITING: 2009 if (!nvme_qpair_is_admin_queue(qpair)) { 2010 nvme_robust_mutex_lock(&ctrlr->ctrlr_lock); 2011 } 2012 2013 rc = nvme_rdma_process_event_poll(rqpair); 2014 2015 if (!nvme_qpair_is_admin_queue(qpair)) { 2016 nvme_robust_mutex_unlock(&ctrlr->ctrlr_lock); 2017 } 2018 break; 2019 2020 case NVME_RDMA_QPAIR_STATE_LINGERING: 2021 rc = nvme_rdma_qpair_wait_until_quiet(rqpair); 2022 break; 2023 case NVME_RDMA_QPAIR_STATE_EXITED: 2024 rc = 0; 2025 break; 2026 2027 default: 2028 assert(false); 2029 rc = -EAGAIN; 2030 break; 2031 } 2032 2033 return rc; 2034 } 2035 2036 static void 2037 nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 2038 { 2039 int rc; 2040 2041 _nvme_rdma_ctrlr_disconnect_qpair(ctrlr, qpair, nvme_rdma_qpair_disconnected); 2042 2043 /* If the async mode is disabled, poll the qpair until it is actually disconnected. 2044 * It is ensured that poll_group_process_completions() calls disconnected_qpair_cb 2045 * for any disconnected qpair. Hence, we do not have to check if the qpair is in 2046 * a poll group or not. 2047 * At the same time, if the qpair is being destroyed, i.e. this function is called by 2048 * spdk_nvme_ctrlr_free_io_qpair then we need to wait until qpair is disconnected, otherwise 2049 * we may leak some resources. 2050 */ 2051 if (qpair->async && !qpair->destroy_in_progress) { 2052 return; 2053 } 2054 2055 while (1) { 2056 rc = nvme_rdma_ctrlr_disconnect_qpair_poll(ctrlr, qpair); 2057 if (rc != -EAGAIN) { 2058 break; 2059 } 2060 } 2061 } 2062 2063 static int 2064 nvme_rdma_stale_conn_disconnected(struct nvme_rdma_qpair *rqpair, int ret) 2065 { 2066 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2067 2068 if (ret) { 2069 SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n"); 2070 } 2071 2072 nvme_rdma_qpair_destroy(rqpair); 2073 2074 qpair->last_transport_failure_reason = qpair->transport_failure_reason; 2075 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_NONE; 2076 2077 rqpair->state = NVME_RDMA_QPAIR_STATE_STALE_CONN; 2078 rqpair->evt_timeout_ticks = (NVME_RDMA_STALE_CONN_RETRY_DELAY_US * spdk_get_ticks_hz()) / 2079 SPDK_SEC_TO_USEC + spdk_get_ticks(); 2080 2081 return 0; 2082 } 2083 2084 static int 2085 nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair) 2086 { 2087 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2088 2089 if (rqpair->stale_conn_retry_count >= NVME_RDMA_STALE_CONN_RETRY_MAX) { 2090 SPDK_ERRLOG("Retry failed %d times, give up stale connection to qpair (cntlid:%u, qid:%u).\n", 2091 NVME_RDMA_STALE_CONN_RETRY_MAX, qpair->ctrlr->cntlid, qpair->id); 2092 return -ESTALE; 2093 } 2094 2095 rqpair->stale_conn_retry_count++; 2096 2097 SPDK_NOTICELOG("%d times, retry stale connection to qpair (cntlid:%u, qid:%u).\n", 2098 rqpair->stale_conn_retry_count, qpair->ctrlr->cntlid, qpair->id); 2099 2100 _nvme_rdma_ctrlr_disconnect_qpair(qpair->ctrlr, qpair, nvme_rdma_stale_conn_disconnected); 2101 2102 return 0; 2103 } 2104 2105 static int 2106 nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 2107 { 2108 struct nvme_rdma_qpair *rqpair; 2109 2110 assert(qpair != NULL); 2111 rqpair = nvme_rdma_qpair(qpair); 2112 2113 if (rqpair->state != NVME_RDMA_QPAIR_STATE_EXITED) { 2114 int rc __attribute__((unused)); 2115 2116 /* qpair was removed from the poll group while the disconnect is not finished. 2117 * Destroy rdma resources forcefully. */ 2118 rc = nvme_rdma_qpair_disconnected(rqpair, 0); 2119 assert(rc == 0); 2120 } 2121 2122 nvme_rdma_qpair_abort_reqs(qpair, 0); 2123 nvme_qpair_deinit(qpair); 2124 2125 nvme_rdma_put_memory_domain(rqpair->memory_domain); 2126 2127 spdk_free(rqpair); 2128 2129 return 0; 2130 } 2131 2132 static struct spdk_nvme_qpair * 2133 nvme_rdma_ctrlr_create_io_qpair(struct spdk_nvme_ctrlr *ctrlr, uint16_t qid, 2134 const struct spdk_nvme_io_qpair_opts *opts) 2135 { 2136 return nvme_rdma_ctrlr_create_qpair(ctrlr, qid, opts->io_queue_size, opts->qprio, 2137 opts->io_queue_requests, 2138 opts->delay_cmd_submit, 2139 opts->async_mode); 2140 } 2141 2142 static int 2143 nvme_rdma_ctrlr_enable(struct spdk_nvme_ctrlr *ctrlr) 2144 { 2145 /* do nothing here */ 2146 return 0; 2147 } 2148 2149 static int nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr); 2150 2151 /* We have to use the typedef in the function declaration to appease astyle. */ 2152 typedef struct spdk_nvme_ctrlr spdk_nvme_ctrlr_t; 2153 2154 static spdk_nvme_ctrlr_t * 2155 nvme_rdma_ctrlr_construct(const struct spdk_nvme_transport_id *trid, 2156 const struct spdk_nvme_ctrlr_opts *opts, 2157 void *devhandle) 2158 { 2159 struct nvme_rdma_ctrlr *rctrlr; 2160 struct ibv_context **contexts; 2161 struct ibv_device_attr dev_attr; 2162 int i, flag, rc; 2163 2164 rctrlr = spdk_zmalloc(sizeof(struct nvme_rdma_ctrlr), 0, NULL, SPDK_ENV_SOCKET_ID_ANY, 2165 SPDK_MALLOC_DMA); 2166 if (rctrlr == NULL) { 2167 SPDK_ERRLOG("could not allocate ctrlr\n"); 2168 return NULL; 2169 } 2170 2171 rctrlr->ctrlr.opts = *opts; 2172 rctrlr->ctrlr.trid = *trid; 2173 2174 if (opts->transport_retry_count > NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT) { 2175 SPDK_NOTICELOG("transport_retry_count exceeds max value %d, use max value\n", 2176 NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT); 2177 rctrlr->ctrlr.opts.transport_retry_count = NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT; 2178 } 2179 2180 if (opts->transport_ack_timeout > NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT) { 2181 SPDK_NOTICELOG("transport_ack_timeout exceeds max value %d, use max value\n", 2182 NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT); 2183 rctrlr->ctrlr.opts.transport_ack_timeout = NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT; 2184 } 2185 2186 contexts = rdma_get_devices(NULL); 2187 if (contexts == NULL) { 2188 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2189 spdk_free(rctrlr); 2190 return NULL; 2191 } 2192 2193 i = 0; 2194 rctrlr->max_sge = NVME_RDMA_MAX_SGL_DESCRIPTORS; 2195 2196 while (contexts[i] != NULL) { 2197 rc = ibv_query_device(contexts[i], &dev_attr); 2198 if (rc < 0) { 2199 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2200 rdma_free_devices(contexts); 2201 spdk_free(rctrlr); 2202 return NULL; 2203 } 2204 rctrlr->max_sge = spdk_min(rctrlr->max_sge, (uint16_t)dev_attr.max_sge); 2205 i++; 2206 } 2207 2208 rdma_free_devices(contexts); 2209 2210 rc = nvme_ctrlr_construct(&rctrlr->ctrlr); 2211 if (rc != 0) { 2212 spdk_free(rctrlr); 2213 return NULL; 2214 } 2215 2216 STAILQ_INIT(&rctrlr->pending_cm_events); 2217 STAILQ_INIT(&rctrlr->free_cm_events); 2218 rctrlr->cm_events = spdk_zmalloc(NVME_RDMA_NUM_CM_EVENTS * sizeof(*rctrlr->cm_events), 0, NULL, 2219 SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 2220 if (rctrlr->cm_events == NULL) { 2221 SPDK_ERRLOG("unable to allocate buffers to hold CM events.\n"); 2222 goto destruct_ctrlr; 2223 } 2224 2225 for (i = 0; i < NVME_RDMA_NUM_CM_EVENTS; i++) { 2226 STAILQ_INSERT_TAIL(&rctrlr->free_cm_events, &rctrlr->cm_events[i], link); 2227 } 2228 2229 rctrlr->cm_channel = rdma_create_event_channel(); 2230 if (rctrlr->cm_channel == NULL) { 2231 SPDK_ERRLOG("rdma_create_event_channel() failed\n"); 2232 goto destruct_ctrlr; 2233 } 2234 2235 flag = fcntl(rctrlr->cm_channel->fd, F_GETFL); 2236 if (fcntl(rctrlr->cm_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2237 SPDK_ERRLOG("Cannot set event channel to non blocking\n"); 2238 goto destruct_ctrlr; 2239 } 2240 2241 rctrlr->ctrlr.adminq = nvme_rdma_ctrlr_create_qpair(&rctrlr->ctrlr, 0, 2242 rctrlr->ctrlr.opts.admin_queue_size, 0, 2243 rctrlr->ctrlr.opts.admin_queue_size, false, true); 2244 if (!rctrlr->ctrlr.adminq) { 2245 SPDK_ERRLOG("failed to create admin qpair\n"); 2246 goto destruct_ctrlr; 2247 } 2248 2249 if (nvme_ctrlr_add_process(&rctrlr->ctrlr, 0) != 0) { 2250 SPDK_ERRLOG("nvme_ctrlr_add_process() failed\n"); 2251 goto destruct_ctrlr; 2252 } 2253 2254 SPDK_DEBUGLOG(nvme, "successfully initialized the nvmf ctrlr\n"); 2255 return &rctrlr->ctrlr; 2256 2257 destruct_ctrlr: 2258 nvme_ctrlr_destruct(&rctrlr->ctrlr); 2259 return NULL; 2260 } 2261 2262 static int 2263 nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr) 2264 { 2265 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr); 2266 struct nvme_rdma_cm_event_entry *entry; 2267 2268 if (ctrlr->adminq) { 2269 nvme_rdma_ctrlr_delete_io_qpair(ctrlr, ctrlr->adminq); 2270 } 2271 2272 STAILQ_FOREACH(entry, &rctrlr->pending_cm_events, link) { 2273 rdma_ack_cm_event(entry->evt); 2274 } 2275 2276 STAILQ_INIT(&rctrlr->free_cm_events); 2277 STAILQ_INIT(&rctrlr->pending_cm_events); 2278 spdk_free(rctrlr->cm_events); 2279 2280 if (rctrlr->cm_channel) { 2281 rdma_destroy_event_channel(rctrlr->cm_channel); 2282 rctrlr->cm_channel = NULL; 2283 } 2284 2285 nvme_ctrlr_destruct_finish(ctrlr); 2286 2287 spdk_free(rctrlr); 2288 2289 return 0; 2290 } 2291 2292 static int 2293 nvme_rdma_qpair_submit_request(struct spdk_nvme_qpair *qpair, 2294 struct nvme_request *req) 2295 { 2296 struct nvme_rdma_qpair *rqpair; 2297 struct spdk_nvme_rdma_req *rdma_req; 2298 struct ibv_send_wr *wr; 2299 2300 rqpair = nvme_rdma_qpair(qpair); 2301 assert(rqpair != NULL); 2302 assert(req != NULL); 2303 2304 rdma_req = nvme_rdma_req_get(rqpair); 2305 if (spdk_unlikely(!rdma_req)) { 2306 if (rqpair->poller) { 2307 rqpair->poller->stats.queued_requests++; 2308 } 2309 /* Inform the upper layer to try again later. */ 2310 return -EAGAIN; 2311 } 2312 2313 if (nvme_rdma_req_init(rqpair, req, rdma_req)) { 2314 SPDK_ERRLOG("nvme_rdma_req_init() failed\n"); 2315 nvme_rdma_req_put(rqpair, rdma_req); 2316 return -1; 2317 } 2318 2319 TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link); 2320 2321 assert(rqpair->current_num_sends < rqpair->num_entries); 2322 rqpair->current_num_sends++; 2323 2324 wr = &rdma_req->send_wr; 2325 wr->next = NULL; 2326 nvme_rdma_trace_ibv_sge(wr->sg_list); 2327 2328 spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, wr); 2329 2330 if (!rqpair->delay_cmd_submit) { 2331 return nvme_rdma_qpair_submit_sends(rqpair); 2332 } 2333 2334 return 0; 2335 } 2336 2337 static int 2338 nvme_rdma_qpair_reset(struct spdk_nvme_qpair *qpair) 2339 { 2340 /* Currently, doing nothing here */ 2341 return 0; 2342 } 2343 2344 static void 2345 nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr) 2346 { 2347 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2348 struct spdk_nvme_cpl cpl; 2349 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2350 2351 cpl.sqid = qpair->id; 2352 cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION; 2353 cpl.status.sct = SPDK_NVME_SCT_GENERIC; 2354 cpl.status.dnr = dnr; 2355 2356 /* 2357 * We cannot abort requests at the RDMA layer without 2358 * unregistering them. If we do, we can still get error 2359 * free completions on the shared completion queue. 2360 */ 2361 if (nvme_qpair_get_state(qpair) > NVME_QPAIR_DISCONNECTING && 2362 nvme_qpair_get_state(qpair) != NVME_QPAIR_DESTROYING) { 2363 nvme_ctrlr_disconnect_qpair(qpair); 2364 } 2365 2366 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2367 nvme_rdma_req_complete(rdma_req, &cpl, true); 2368 } 2369 } 2370 2371 static void 2372 nvme_rdma_qpair_check_timeout(struct spdk_nvme_qpair *qpair) 2373 { 2374 uint64_t t02; 2375 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2376 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2377 struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr; 2378 struct spdk_nvme_ctrlr_process *active_proc; 2379 2380 /* Don't check timeouts during controller initialization. */ 2381 if (ctrlr->state != NVME_CTRLR_STATE_READY) { 2382 return; 2383 } 2384 2385 if (nvme_qpair_is_admin_queue(qpair)) { 2386 active_proc = nvme_ctrlr_get_current_process(ctrlr); 2387 } else { 2388 active_proc = qpair->active_proc; 2389 } 2390 2391 /* Only check timeouts if the current process has a timeout callback. */ 2392 if (active_proc == NULL || active_proc->timeout_cb_fn == NULL) { 2393 return; 2394 } 2395 2396 t02 = spdk_get_ticks(); 2397 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2398 assert(rdma_req->req != NULL); 2399 2400 if (nvme_request_check_timeout(rdma_req->req, rdma_req->id, active_proc, t02)) { 2401 /* 2402 * The requests are in order, so as soon as one has not timed out, 2403 * stop iterating. 2404 */ 2405 break; 2406 } 2407 } 2408 } 2409 2410 static inline void 2411 nvme_rdma_request_ready(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req) 2412 { 2413 struct spdk_nvme_rdma_rsp *rdma_rsp = rdma_req->rdma_rsp; 2414 struct ibv_recv_wr *recv_wr = rdma_rsp->recv_wr; 2415 2416 nvme_rdma_req_complete(rdma_req, &rdma_rsp->cpl, true); 2417 2418 assert(rqpair->rsps->current_num_recvs < rqpair->rsps->num_entries); 2419 rqpair->rsps->current_num_recvs++; 2420 2421 recv_wr->next = NULL; 2422 nvme_rdma_trace_ibv_sge(recv_wr->sg_list); 2423 2424 if (!rqpair->srq) { 2425 spdk_rdma_qp_queue_recv_wrs(rqpair->rdma_qp, recv_wr); 2426 } else { 2427 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, recv_wr); 2428 } 2429 } 2430 2431 #define MAX_COMPLETIONS_PER_POLL 128 2432 2433 static void 2434 nvme_rdma_fail_qpair(struct spdk_nvme_qpair *qpair, int failure_reason) 2435 { 2436 if (failure_reason == IBV_WC_RETRY_EXC_ERR) { 2437 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE; 2438 } else if (qpair->transport_failure_reason == SPDK_NVME_QPAIR_FAILURE_NONE) { 2439 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_UNKNOWN; 2440 } 2441 2442 nvme_ctrlr_disconnect_qpair(qpair); 2443 } 2444 2445 static struct nvme_rdma_qpair * 2446 get_rdma_qpair_from_wc(struct nvme_rdma_poll_group *group, struct ibv_wc *wc) 2447 { 2448 struct spdk_nvme_qpair *qpair; 2449 struct nvme_rdma_qpair *rqpair; 2450 2451 STAILQ_FOREACH(qpair, &group->group.connected_qpairs, poll_group_stailq) { 2452 rqpair = nvme_rdma_qpair(qpair); 2453 if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) { 2454 return rqpair; 2455 } 2456 } 2457 2458 STAILQ_FOREACH(qpair, &group->group.disconnected_qpairs, poll_group_stailq) { 2459 rqpair = nvme_rdma_qpair(qpair); 2460 if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) { 2461 return rqpair; 2462 } 2463 } 2464 2465 return NULL; 2466 } 2467 2468 static inline void 2469 nvme_rdma_log_wc_status(struct nvme_rdma_qpair *rqpair, struct ibv_wc *wc) 2470 { 2471 struct nvme_rdma_wr *rdma_wr = (struct nvme_rdma_wr *)wc->wr_id; 2472 2473 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 2474 /* If qpair is in ERR state, we will receive completions for all posted and not completed 2475 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 2476 SPDK_DEBUGLOG(nvme, "WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n", 2477 rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status, 2478 ibv_wc_status_str(wc->status)); 2479 } else { 2480 SPDK_ERRLOG("WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n", 2481 rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status, 2482 ibv_wc_status_str(wc->status)); 2483 } 2484 } 2485 2486 static inline int 2487 nvme_rdma_process_recv_completion(struct nvme_rdma_poller *poller, struct ibv_wc *wc, 2488 struct nvme_rdma_wr *rdma_wr) 2489 { 2490 struct nvme_rdma_qpair *rqpair; 2491 struct spdk_nvme_rdma_req *rdma_req; 2492 struct spdk_nvme_rdma_rsp *rdma_rsp; 2493 2494 rdma_rsp = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_rsp, rdma_wr); 2495 2496 if (poller && poller->srq) { 2497 rqpair = get_rdma_qpair_from_wc(poller->group, wc); 2498 if (spdk_unlikely(!rqpair)) { 2499 /* Since we do not handle the LAST_WQE_REACHED event, we do not know when 2500 * a Receive Queue in a QP, that is associated with an SRQ, is flushed. 2501 * We may get a WC for a already destroyed QP. 2502 * 2503 * However, for the SRQ, this is not any error. Hence, just re-post the 2504 * receive request to the SRQ to reuse for other QPs, and return 0. 2505 */ 2506 spdk_rdma_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr); 2507 return 0; 2508 } 2509 } else { 2510 rqpair = rdma_rsp->rqpair; 2511 if (spdk_unlikely(!rqpair)) { 2512 /* TODO: Fix forceful QP destroy when it is not async mode. 2513 * CQ itself did not cause any error. Hence, return 0 for now. 2514 */ 2515 SPDK_WARNLOG("QP might be already destroyed.\n"); 2516 return 0; 2517 } 2518 } 2519 2520 2521 assert(rqpair->rsps->current_num_recvs > 0); 2522 rqpair->rsps->current_num_recvs--; 2523 2524 if (wc->status) { 2525 nvme_rdma_log_wc_status(rqpair, wc); 2526 goto err_wc; 2527 } 2528 2529 SPDK_DEBUGLOG(nvme, "CQ recv completion\n"); 2530 2531 if (wc->byte_len < sizeof(struct spdk_nvme_cpl)) { 2532 SPDK_ERRLOG("recv length %u less than expected response size\n", wc->byte_len); 2533 goto err_wc; 2534 } 2535 rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid]; 2536 rdma_req->completion_flags |= NVME_RDMA_RECV_COMPLETED; 2537 rdma_req->rdma_rsp = rdma_rsp; 2538 2539 if ((rdma_req->completion_flags & NVME_RDMA_SEND_COMPLETED) == 0) { 2540 return 0; 2541 } 2542 2543 nvme_rdma_request_ready(rqpair, rdma_req); 2544 2545 if (!rqpair->delay_cmd_submit) { 2546 if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) { 2547 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 2548 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2549 return -ENXIO; 2550 } 2551 } 2552 2553 rqpair->num_completions++; 2554 return 1; 2555 2556 err_wc: 2557 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2558 if (poller && poller->srq) { 2559 spdk_rdma_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr); 2560 } 2561 return -ENXIO; 2562 } 2563 2564 static inline int 2565 nvme_rdma_process_send_completion(struct nvme_rdma_poller *poller, 2566 struct nvme_rdma_qpair *rdma_qpair, 2567 struct ibv_wc *wc, struct nvme_rdma_wr *rdma_wr) 2568 { 2569 struct nvme_rdma_qpair *rqpair; 2570 struct spdk_nvme_rdma_req *rdma_req; 2571 2572 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_req, rdma_wr); 2573 rqpair = rdma_req->req ? nvme_rdma_qpair(rdma_req->req->qpair) : NULL; 2574 if (!rqpair) { 2575 rqpair = rdma_qpair != NULL ? rdma_qpair : get_rdma_qpair_from_wc(poller->group, wc); 2576 } 2577 2578 /* If we are flushing I/O */ 2579 if (wc->status) { 2580 if (!rqpair) { 2581 /* When poll_group is used, several qpairs share the same CQ and it is possible to 2582 * receive a completion with error (e.g. IBV_WC_WR_FLUSH_ERR) for already disconnected qpair 2583 * That happens due to qpair is destroyed while there are submitted but not completed send/receive 2584 * Work Requests */ 2585 assert(poller); 2586 return 0; 2587 } 2588 assert(rqpair->current_num_sends > 0); 2589 rqpair->current_num_sends--; 2590 nvme_rdma_log_wc_status(rqpair, wc); 2591 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2592 if (rdma_req->rdma_rsp && poller && poller->srq) { 2593 spdk_rdma_srq_queue_recv_wrs(poller->srq, rdma_req->rdma_rsp->recv_wr); 2594 } 2595 return -ENXIO; 2596 } 2597 2598 /* We do not support Soft Roce anymore. Other than Soft Roce's bug, we should not 2599 * receive a completion without error status after qpair is disconnected/destroyed. 2600 */ 2601 if (spdk_unlikely(rdma_req->req == NULL)) { 2602 /* 2603 * Some infiniband drivers do not guarantee the previous assumption after we 2604 * received a RDMA_CM_EVENT_DEVICE_REMOVAL event. 2605 */ 2606 SPDK_ERRLOG("Received malformed completion: request 0x%"PRIx64" type %d\n", wc->wr_id, 2607 rdma_wr->type); 2608 if (!rqpair || !rqpair->need_destroy) { 2609 assert(0); 2610 } 2611 return -ENXIO; 2612 } 2613 2614 rdma_req->completion_flags |= NVME_RDMA_SEND_COMPLETED; 2615 assert(rqpair->current_num_sends > 0); 2616 rqpair->current_num_sends--; 2617 2618 if ((rdma_req->completion_flags & NVME_RDMA_RECV_COMPLETED) == 0) { 2619 return 0; 2620 } 2621 2622 nvme_rdma_request_ready(rqpair, rdma_req); 2623 2624 if (!rqpair->delay_cmd_submit) { 2625 if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) { 2626 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 2627 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2628 return -ENXIO; 2629 } 2630 } 2631 2632 rqpair->num_completions++; 2633 return 1; 2634 } 2635 2636 static int 2637 nvme_rdma_cq_process_completions(struct ibv_cq *cq, uint32_t batch_size, 2638 struct nvme_rdma_poller *poller, 2639 struct nvme_rdma_qpair *rdma_qpair, 2640 uint64_t *rdma_completions) 2641 { 2642 struct ibv_wc wc[MAX_COMPLETIONS_PER_POLL]; 2643 struct nvme_rdma_wr *rdma_wr; 2644 uint32_t reaped = 0; 2645 int completion_rc = 0; 2646 int rc, _rc, i; 2647 2648 rc = ibv_poll_cq(cq, batch_size, wc); 2649 if (rc < 0) { 2650 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2651 errno, spdk_strerror(errno)); 2652 return -ECANCELED; 2653 } else if (rc == 0) { 2654 return 0; 2655 } 2656 2657 for (i = 0; i < rc; i++) { 2658 rdma_wr = (struct nvme_rdma_wr *)wc[i].wr_id; 2659 switch (rdma_wr->type) { 2660 case RDMA_WR_TYPE_RECV: 2661 _rc = nvme_rdma_process_recv_completion(poller, &wc[i], rdma_wr); 2662 break; 2663 2664 case RDMA_WR_TYPE_SEND: 2665 _rc = nvme_rdma_process_send_completion(poller, rdma_qpair, &wc[i], rdma_wr); 2666 break; 2667 2668 default: 2669 SPDK_ERRLOG("Received an unexpected opcode on the CQ: %d\n", rdma_wr->type); 2670 return -ECANCELED; 2671 } 2672 if (spdk_likely(_rc >= 0)) { 2673 reaped += _rc; 2674 } else { 2675 completion_rc = _rc; 2676 } 2677 } 2678 2679 *rdma_completions += rc; 2680 2681 if (completion_rc) { 2682 return completion_rc; 2683 } 2684 2685 return reaped; 2686 } 2687 2688 static void 2689 dummy_disconnected_qpair_cb(struct spdk_nvme_qpair *qpair, void *poll_group_ctx) 2690 { 2691 2692 } 2693 2694 static int 2695 nvme_rdma_qpair_process_completions(struct spdk_nvme_qpair *qpair, 2696 uint32_t max_completions) 2697 { 2698 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2699 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(qpair->ctrlr); 2700 int rc = 0, batch_size; 2701 struct ibv_cq *cq; 2702 uint64_t rdma_completions = 0; 2703 2704 /* 2705 * This is used during the connection phase. It's possible that we are still reaping error completions 2706 * from other qpairs so we need to call the poll group function. Also, it's more correct since the cq 2707 * is shared. 2708 */ 2709 if (qpair->poll_group != NULL) { 2710 return spdk_nvme_poll_group_process_completions(qpair->poll_group->group, max_completions, 2711 dummy_disconnected_qpair_cb); 2712 } 2713 2714 if (max_completions == 0) { 2715 max_completions = rqpair->num_entries; 2716 } else { 2717 max_completions = spdk_min(max_completions, rqpair->num_entries); 2718 } 2719 2720 switch (nvme_qpair_get_state(qpair)) { 2721 case NVME_QPAIR_CONNECTING: 2722 rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair); 2723 if (rc == 0) { 2724 /* Once the connection is completed, we can submit queued requests */ 2725 nvme_qpair_resubmit_requests(qpair, rqpair->num_entries); 2726 } else if (rc != -EAGAIN) { 2727 SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair); 2728 goto failed; 2729 } else if (rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING) { 2730 return 0; 2731 } 2732 break; 2733 2734 case NVME_QPAIR_DISCONNECTING: 2735 nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair); 2736 return -ENXIO; 2737 2738 default: 2739 if (nvme_qpair_is_admin_queue(qpair)) { 2740 nvme_rdma_poll_events(rctrlr); 2741 } 2742 nvme_rdma_qpair_process_cm_event(rqpair); 2743 break; 2744 } 2745 2746 if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) { 2747 goto failed; 2748 } 2749 2750 cq = rqpair->cq; 2751 2752 rqpair->num_completions = 0; 2753 do { 2754 batch_size = spdk_min((max_completions - rqpair->num_completions), MAX_COMPLETIONS_PER_POLL); 2755 rc = nvme_rdma_cq_process_completions(cq, batch_size, NULL, rqpair, &rdma_completions); 2756 2757 if (rc == 0) { 2758 break; 2759 /* Handle the case where we fail to poll the cq. */ 2760 } else if (rc == -ECANCELED) { 2761 goto failed; 2762 } else if (rc == -ENXIO) { 2763 return rc; 2764 } 2765 } while (rqpair->num_completions < max_completions); 2766 2767 if (spdk_unlikely(nvme_rdma_qpair_submit_sends(rqpair) || 2768 nvme_rdma_qpair_submit_recvs(rqpair))) { 2769 goto failed; 2770 } 2771 2772 if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) { 2773 nvme_rdma_qpair_check_timeout(qpair); 2774 } 2775 2776 return rqpair->num_completions; 2777 2778 failed: 2779 nvme_rdma_fail_qpair(qpair, 0); 2780 return -ENXIO; 2781 } 2782 2783 static uint32_t 2784 nvme_rdma_ctrlr_get_max_xfer_size(struct spdk_nvme_ctrlr *ctrlr) 2785 { 2786 /* max_mr_size by ibv_query_device indicates the largest value that we can 2787 * set for a registered memory region. It is independent from the actual 2788 * I/O size and is very likely to be larger than 2 MiB which is the 2789 * granularity we currently register memory regions. Hence return 2790 * UINT32_MAX here and let the generic layer use the controller data to 2791 * moderate this value. 2792 */ 2793 return UINT32_MAX; 2794 } 2795 2796 static uint16_t 2797 nvme_rdma_ctrlr_get_max_sges(struct spdk_nvme_ctrlr *ctrlr) 2798 { 2799 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr); 2800 uint32_t max_sge = rctrlr->max_sge; 2801 uint32_t max_in_capsule_sge = (ctrlr->cdata.nvmf_specific.ioccsz * 16 - 2802 sizeof(struct spdk_nvme_cmd)) / 2803 sizeof(struct spdk_nvme_sgl_descriptor); 2804 2805 /* Max SGE is limited by capsule size */ 2806 max_sge = spdk_min(max_sge, max_in_capsule_sge); 2807 /* Max SGE may be limited by MSDBD */ 2808 if (ctrlr->cdata.nvmf_specific.msdbd != 0) { 2809 max_sge = spdk_min(max_sge, ctrlr->cdata.nvmf_specific.msdbd); 2810 } 2811 2812 /* Max SGE can't be less than 1 */ 2813 max_sge = spdk_max(1, max_sge); 2814 return max_sge; 2815 } 2816 2817 static int 2818 nvme_rdma_qpair_iterate_requests(struct spdk_nvme_qpair *qpair, 2819 int (*iter_fn)(struct nvme_request *req, void *arg), 2820 void *arg) 2821 { 2822 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2823 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2824 int rc; 2825 2826 assert(iter_fn != NULL); 2827 2828 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2829 assert(rdma_req->req != NULL); 2830 2831 rc = iter_fn(rdma_req->req, arg); 2832 if (rc != 0) { 2833 return rc; 2834 } 2835 } 2836 2837 return 0; 2838 } 2839 2840 static void 2841 nvme_rdma_admin_qpair_abort_aers(struct spdk_nvme_qpair *qpair) 2842 { 2843 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2844 struct spdk_nvme_cpl cpl; 2845 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2846 2847 cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION; 2848 cpl.status.sct = SPDK_NVME_SCT_GENERIC; 2849 2850 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2851 assert(rdma_req->req != NULL); 2852 2853 if (rdma_req->req->cmd.opc != SPDK_NVME_OPC_ASYNC_EVENT_REQUEST) { 2854 continue; 2855 } 2856 2857 nvme_rdma_req_complete(rdma_req, &cpl, false); 2858 } 2859 } 2860 2861 static void 2862 nvme_rdma_poller_destroy(struct nvme_rdma_poller *poller) 2863 { 2864 if (poller->cq) { 2865 ibv_destroy_cq(poller->cq); 2866 } 2867 if (poller->rsps) { 2868 nvme_rdma_free_rsps(poller->rsps); 2869 } 2870 if (poller->srq) { 2871 spdk_rdma_srq_destroy(poller->srq); 2872 } 2873 if (poller->mr_map) { 2874 spdk_rdma_free_mem_map(&poller->mr_map); 2875 } 2876 if (poller->pd) { 2877 spdk_rdma_put_pd(poller->pd); 2878 } 2879 free(poller); 2880 } 2881 2882 static struct nvme_rdma_poller * 2883 nvme_rdma_poller_create(struct nvme_rdma_poll_group *group, struct ibv_context *ctx) 2884 { 2885 struct nvme_rdma_poller *poller; 2886 struct ibv_device_attr dev_attr; 2887 struct spdk_rdma_srq_init_attr srq_init_attr = {}; 2888 struct nvme_rdma_rsp_opts opts; 2889 int num_cqe, max_num_cqe; 2890 int rc; 2891 2892 poller = calloc(1, sizeof(*poller)); 2893 if (poller == NULL) { 2894 SPDK_ERRLOG("Unable to allocate poller.\n"); 2895 return NULL; 2896 } 2897 2898 poller->group = group; 2899 poller->device = ctx; 2900 2901 if (g_spdk_nvme_transport_opts.rdma_srq_size != 0) { 2902 rc = ibv_query_device(ctx, &dev_attr); 2903 if (rc) { 2904 SPDK_ERRLOG("Unable to query RDMA device.\n"); 2905 goto fail; 2906 } 2907 2908 poller->pd = spdk_rdma_get_pd(ctx); 2909 if (poller->pd == NULL) { 2910 SPDK_ERRLOG("Unable to get PD.\n"); 2911 goto fail; 2912 } 2913 2914 poller->mr_map = spdk_rdma_create_mem_map(poller->pd, &g_nvme_hooks, 2915 SPDK_RDMA_MEMORY_MAP_ROLE_INITIATOR); 2916 if (poller->mr_map == NULL) { 2917 SPDK_ERRLOG("Unable to create memory map.\n"); 2918 goto fail; 2919 } 2920 2921 srq_init_attr.stats = &poller->stats.rdma_stats.recv; 2922 srq_init_attr.pd = poller->pd; 2923 srq_init_attr.srq_init_attr.attr.max_wr = spdk_min((uint32_t)dev_attr.max_srq_wr, 2924 g_spdk_nvme_transport_opts.rdma_srq_size); 2925 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(dev_attr.max_sge, 2926 NVME_RDMA_DEFAULT_RX_SGE); 2927 2928 poller->srq = spdk_rdma_srq_create(&srq_init_attr); 2929 if (poller->srq == NULL) { 2930 SPDK_ERRLOG("Unable to create SRQ.\n"); 2931 goto fail; 2932 } 2933 2934 opts.num_entries = g_spdk_nvme_transport_opts.rdma_srq_size; 2935 opts.rqpair = NULL; 2936 opts.srq = poller->srq; 2937 opts.mr_map = poller->mr_map; 2938 2939 poller->rsps = nvme_rdma_create_rsps(&opts); 2940 if (poller->rsps == NULL) { 2941 SPDK_ERRLOG("Unable to create poller RDMA responses.\n"); 2942 goto fail; 2943 } 2944 2945 rc = nvme_rdma_poller_submit_recvs(poller); 2946 if (rc) { 2947 SPDK_ERRLOG("Unable to submit poller RDMA responses.\n"); 2948 goto fail; 2949 } 2950 2951 /* 2952 * When using an srq, fix the size of the completion queue at startup. 2953 * The initiator sends only send and recv WRs. Hence, the multiplier is 2. 2954 * (The target sends also data WRs. Hence, the multiplier is 3.) 2955 */ 2956 num_cqe = g_spdk_nvme_transport_opts.rdma_srq_size * 2; 2957 } else { 2958 num_cqe = DEFAULT_NVME_RDMA_CQ_SIZE; 2959 } 2960 2961 max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size; 2962 if (max_num_cqe != 0 && num_cqe > max_num_cqe) { 2963 num_cqe = max_num_cqe; 2964 } 2965 2966 poller->cq = ibv_create_cq(poller->device, num_cqe, group, NULL, 0); 2967 2968 if (poller->cq == NULL) { 2969 SPDK_ERRLOG("Unable to create CQ, errno %d.\n", errno); 2970 goto fail; 2971 } 2972 2973 STAILQ_INSERT_HEAD(&group->pollers, poller, link); 2974 group->num_pollers++; 2975 poller->current_num_wc = num_cqe; 2976 poller->required_num_wc = 0; 2977 return poller; 2978 2979 fail: 2980 nvme_rdma_poller_destroy(poller); 2981 return NULL; 2982 } 2983 2984 static void 2985 nvme_rdma_poll_group_free_pollers(struct nvme_rdma_poll_group *group) 2986 { 2987 struct nvme_rdma_poller *poller, *tmp_poller; 2988 2989 STAILQ_FOREACH_SAFE(poller, &group->pollers, link, tmp_poller) { 2990 assert(poller->refcnt == 0); 2991 if (poller->refcnt) { 2992 SPDK_WARNLOG("Destroying poller with non-zero ref count: poller %p, refcnt %d\n", 2993 poller, poller->refcnt); 2994 } 2995 2996 STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link); 2997 nvme_rdma_poller_destroy(poller); 2998 } 2999 } 3000 3001 static struct nvme_rdma_poller * 3002 nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, struct ibv_context *device) 3003 { 3004 struct nvme_rdma_poller *poller = NULL; 3005 3006 STAILQ_FOREACH(poller, &group->pollers, link) { 3007 if (poller->device == device) { 3008 break; 3009 } 3010 } 3011 3012 if (!poller) { 3013 poller = nvme_rdma_poller_create(group, device); 3014 if (!poller) { 3015 SPDK_ERRLOG("Failed to create a poller for device %p\n", device); 3016 return NULL; 3017 } 3018 } 3019 3020 poller->refcnt++; 3021 return poller; 3022 } 3023 3024 static void 3025 nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, struct nvme_rdma_poller *poller) 3026 { 3027 assert(poller->refcnt > 0); 3028 if (--poller->refcnt == 0) { 3029 STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link); 3030 group->num_pollers--; 3031 nvme_rdma_poller_destroy(poller); 3032 } 3033 } 3034 3035 static struct spdk_nvme_transport_poll_group * 3036 nvme_rdma_poll_group_create(void) 3037 { 3038 struct nvme_rdma_poll_group *group; 3039 3040 group = calloc(1, sizeof(*group)); 3041 if (group == NULL) { 3042 SPDK_ERRLOG("Unable to allocate poll group.\n"); 3043 return NULL; 3044 } 3045 3046 STAILQ_INIT(&group->pollers); 3047 TAILQ_INIT(&group->connecting_qpairs); 3048 return &group->group; 3049 } 3050 3051 static int 3052 nvme_rdma_poll_group_connect_qpair(struct spdk_nvme_qpair *qpair) 3053 { 3054 return 0; 3055 } 3056 3057 static int 3058 nvme_rdma_poll_group_disconnect_qpair(struct spdk_nvme_qpair *qpair) 3059 { 3060 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3061 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);; 3062 3063 if (rqpair->link_connecting.tqe_prev) { 3064 TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting); 3065 /* We use prev pointer to check if qpair is in connecting list or not . 3066 * TAILQ_REMOVE doesn't do it. So, we do it manually. 3067 */ 3068 rqpair->link_connecting.tqe_prev = NULL; 3069 } 3070 3071 return 0; 3072 } 3073 3074 static int 3075 nvme_rdma_poll_group_add(struct spdk_nvme_transport_poll_group *tgroup, 3076 struct spdk_nvme_qpair *qpair) 3077 { 3078 return 0; 3079 } 3080 3081 static int 3082 nvme_rdma_poll_group_remove(struct spdk_nvme_transport_poll_group *tgroup, 3083 struct spdk_nvme_qpair *qpair) 3084 { 3085 return 0; 3086 } 3087 3088 static inline void 3089 nvme_rdma_qpair_process_submits(struct spdk_nvme_qpair *qpair) 3090 { 3091 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3092 3093 if (spdk_unlikely(rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING)) { 3094 return; 3095 } 3096 3097 if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) { 3098 nvme_rdma_qpair_check_timeout(qpair); 3099 } 3100 3101 nvme_rdma_qpair_submit_sends(rqpair); 3102 if (!rqpair->srq) { 3103 nvme_rdma_qpair_submit_recvs(rqpair); 3104 } 3105 if (rqpair->num_completions > 0) { 3106 nvme_qpair_resubmit_requests(qpair, rqpair->num_completions); 3107 rqpair->num_completions = 0; 3108 } 3109 } 3110 3111 static int64_t 3112 nvme_rdma_poll_group_process_completions(struct spdk_nvme_transport_poll_group *tgroup, 3113 uint32_t completions_per_qpair, spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb) 3114 { 3115 struct spdk_nvme_qpair *qpair, *tmp_qpair; 3116 struct nvme_rdma_qpair *rqpair, *tmp_rqpair; 3117 struct nvme_rdma_poll_group *group; 3118 struct nvme_rdma_poller *poller; 3119 int batch_size, rc, rc2 = 0; 3120 int64_t total_completions = 0; 3121 uint64_t completions_allowed = 0; 3122 uint64_t completions_per_poller = 0; 3123 uint64_t poller_completions = 0; 3124 uint64_t rdma_completions; 3125 3126 if (completions_per_qpair == 0) { 3127 completions_per_qpair = MAX_COMPLETIONS_PER_POLL; 3128 } 3129 3130 group = nvme_rdma_poll_group(tgroup); 3131 3132 STAILQ_FOREACH_SAFE(qpair, &tgroup->disconnected_qpairs, poll_group_stailq, tmp_qpair) { 3133 rc = nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair); 3134 if (rc == 0) { 3135 disconnected_qpair_cb(qpair, tgroup->group->ctx); 3136 } 3137 } 3138 3139 TAILQ_FOREACH_SAFE(rqpair, &group->connecting_qpairs, link_connecting, tmp_rqpair) { 3140 qpair = &rqpair->qpair; 3141 3142 rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair); 3143 if (rc == 0 || rc != -EAGAIN) { 3144 TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting); 3145 /* We use prev pointer to check if qpair is in connecting list or not. 3146 * TAILQ_REMOVE does not do it. So, we do it manually. 3147 */ 3148 rqpair->link_connecting.tqe_prev = NULL; 3149 3150 if (rc == 0) { 3151 /* Once the connection is completed, we can submit queued requests */ 3152 nvme_qpair_resubmit_requests(qpair, rqpair->num_entries); 3153 } else if (rc != -EAGAIN) { 3154 SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair); 3155 nvme_rdma_fail_qpair(qpair, 0); 3156 } 3157 } 3158 } 3159 3160 STAILQ_FOREACH_SAFE(qpair, &tgroup->connected_qpairs, poll_group_stailq, tmp_qpair) { 3161 rqpair = nvme_rdma_qpair(qpair); 3162 3163 if (spdk_likely(nvme_qpair_get_state(qpair) != NVME_QPAIR_CONNECTING)) { 3164 nvme_rdma_qpair_process_cm_event(rqpair); 3165 } 3166 3167 if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) { 3168 rc2 = -ENXIO; 3169 nvme_rdma_fail_qpair(qpair, 0); 3170 } 3171 } 3172 3173 completions_allowed = completions_per_qpair * tgroup->num_connected_qpairs; 3174 if (group->num_pollers) { 3175 completions_per_poller = spdk_max(completions_allowed / group->num_pollers, 1); 3176 } 3177 3178 STAILQ_FOREACH(poller, &group->pollers, link) { 3179 poller_completions = 0; 3180 rdma_completions = 0; 3181 do { 3182 poller->stats.polls++; 3183 batch_size = spdk_min((completions_per_poller - poller_completions), MAX_COMPLETIONS_PER_POLL); 3184 rc = nvme_rdma_cq_process_completions(poller->cq, batch_size, poller, NULL, &rdma_completions); 3185 if (rc <= 0) { 3186 if (rc == -ECANCELED) { 3187 return -EIO; 3188 } else if (rc == 0) { 3189 poller->stats.idle_polls++; 3190 } 3191 break; 3192 } 3193 3194 poller_completions += rc; 3195 } while (poller_completions < completions_per_poller); 3196 total_completions += poller_completions; 3197 poller->stats.completions += rdma_completions; 3198 if (poller->srq) { 3199 nvme_rdma_poller_submit_recvs(poller); 3200 } 3201 } 3202 3203 STAILQ_FOREACH_SAFE(qpair, &tgroup->connected_qpairs, poll_group_stailq, tmp_qpair) { 3204 nvme_rdma_qpair_process_submits(qpair); 3205 } 3206 3207 return rc2 != 0 ? rc2 : total_completions; 3208 } 3209 3210 static int 3211 nvme_rdma_poll_group_destroy(struct spdk_nvme_transport_poll_group *tgroup) 3212 { 3213 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(tgroup); 3214 3215 if (!STAILQ_EMPTY(&tgroup->connected_qpairs) || !STAILQ_EMPTY(&tgroup->disconnected_qpairs)) { 3216 return -EBUSY; 3217 } 3218 3219 nvme_rdma_poll_group_free_pollers(group); 3220 free(group); 3221 3222 return 0; 3223 } 3224 3225 static int 3226 nvme_rdma_poll_group_get_stats(struct spdk_nvme_transport_poll_group *tgroup, 3227 struct spdk_nvme_transport_poll_group_stat **_stats) 3228 { 3229 struct nvme_rdma_poll_group *group; 3230 struct spdk_nvme_transport_poll_group_stat *stats; 3231 struct spdk_nvme_rdma_device_stat *device_stat; 3232 struct nvme_rdma_poller *poller; 3233 uint32_t i = 0; 3234 3235 if (tgroup == NULL || _stats == NULL) { 3236 SPDK_ERRLOG("Invalid stats or group pointer\n"); 3237 return -EINVAL; 3238 } 3239 3240 group = nvme_rdma_poll_group(tgroup); 3241 stats = calloc(1, sizeof(*stats)); 3242 if (!stats) { 3243 SPDK_ERRLOG("Can't allocate memory for RDMA stats\n"); 3244 return -ENOMEM; 3245 } 3246 stats->trtype = SPDK_NVME_TRANSPORT_RDMA; 3247 stats->rdma.num_devices = group->num_pollers; 3248 3249 if (stats->rdma.num_devices == 0) { 3250 *_stats = stats; 3251 return 0; 3252 } 3253 3254 stats->rdma.device_stats = calloc(stats->rdma.num_devices, sizeof(*stats->rdma.device_stats)); 3255 if (!stats->rdma.device_stats) { 3256 SPDK_ERRLOG("Can't allocate memory for RDMA device stats\n"); 3257 free(stats); 3258 return -ENOMEM; 3259 } 3260 3261 STAILQ_FOREACH(poller, &group->pollers, link) { 3262 device_stat = &stats->rdma.device_stats[i]; 3263 device_stat->name = poller->device->device->name; 3264 device_stat->polls = poller->stats.polls; 3265 device_stat->idle_polls = poller->stats.idle_polls; 3266 device_stat->completions = poller->stats.completions; 3267 device_stat->queued_requests = poller->stats.queued_requests; 3268 device_stat->total_send_wrs = poller->stats.rdma_stats.send.num_submitted_wrs; 3269 device_stat->send_doorbell_updates = poller->stats.rdma_stats.send.doorbell_updates; 3270 device_stat->total_recv_wrs = poller->stats.rdma_stats.recv.num_submitted_wrs; 3271 device_stat->recv_doorbell_updates = poller->stats.rdma_stats.recv.doorbell_updates; 3272 i++; 3273 } 3274 3275 *_stats = stats; 3276 3277 return 0; 3278 } 3279 3280 static void 3281 nvme_rdma_poll_group_free_stats(struct spdk_nvme_transport_poll_group *tgroup, 3282 struct spdk_nvme_transport_poll_group_stat *stats) 3283 { 3284 if (stats) { 3285 free(stats->rdma.device_stats); 3286 } 3287 free(stats); 3288 } 3289 3290 static int 3291 nvme_rdma_ctrlr_get_memory_domains(const struct spdk_nvme_ctrlr *ctrlr, 3292 struct spdk_memory_domain **domains, int array_size) 3293 { 3294 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(ctrlr->adminq); 3295 3296 if (domains && array_size > 0) { 3297 domains[0] = rqpair->memory_domain->domain; 3298 } 3299 3300 return 1; 3301 } 3302 3303 void 3304 spdk_nvme_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3305 { 3306 g_nvme_hooks = *hooks; 3307 } 3308 3309 const struct spdk_nvme_transport_ops rdma_ops = { 3310 .name = "RDMA", 3311 .type = SPDK_NVME_TRANSPORT_RDMA, 3312 .ctrlr_construct = nvme_rdma_ctrlr_construct, 3313 .ctrlr_scan = nvme_fabric_ctrlr_scan, 3314 .ctrlr_destruct = nvme_rdma_ctrlr_destruct, 3315 .ctrlr_enable = nvme_rdma_ctrlr_enable, 3316 3317 .ctrlr_set_reg_4 = nvme_fabric_ctrlr_set_reg_4, 3318 .ctrlr_set_reg_8 = nvme_fabric_ctrlr_set_reg_8, 3319 .ctrlr_get_reg_4 = nvme_fabric_ctrlr_get_reg_4, 3320 .ctrlr_get_reg_8 = nvme_fabric_ctrlr_get_reg_8, 3321 .ctrlr_set_reg_4_async = nvme_fabric_ctrlr_set_reg_4_async, 3322 .ctrlr_set_reg_8_async = nvme_fabric_ctrlr_set_reg_8_async, 3323 .ctrlr_get_reg_4_async = nvme_fabric_ctrlr_get_reg_4_async, 3324 .ctrlr_get_reg_8_async = nvme_fabric_ctrlr_get_reg_8_async, 3325 3326 .ctrlr_get_max_xfer_size = nvme_rdma_ctrlr_get_max_xfer_size, 3327 .ctrlr_get_max_sges = nvme_rdma_ctrlr_get_max_sges, 3328 3329 .ctrlr_create_io_qpair = nvme_rdma_ctrlr_create_io_qpair, 3330 .ctrlr_delete_io_qpair = nvme_rdma_ctrlr_delete_io_qpair, 3331 .ctrlr_connect_qpair = nvme_rdma_ctrlr_connect_qpair, 3332 .ctrlr_disconnect_qpair = nvme_rdma_ctrlr_disconnect_qpair, 3333 3334 .ctrlr_get_memory_domains = nvme_rdma_ctrlr_get_memory_domains, 3335 3336 .qpair_abort_reqs = nvme_rdma_qpair_abort_reqs, 3337 .qpair_reset = nvme_rdma_qpair_reset, 3338 .qpair_submit_request = nvme_rdma_qpair_submit_request, 3339 .qpair_process_completions = nvme_rdma_qpair_process_completions, 3340 .qpair_iterate_requests = nvme_rdma_qpair_iterate_requests, 3341 .admin_qpair_abort_aers = nvme_rdma_admin_qpair_abort_aers, 3342 3343 .poll_group_create = nvme_rdma_poll_group_create, 3344 .poll_group_connect_qpair = nvme_rdma_poll_group_connect_qpair, 3345 .poll_group_disconnect_qpair = nvme_rdma_poll_group_disconnect_qpair, 3346 .poll_group_add = nvme_rdma_poll_group_add, 3347 .poll_group_remove = nvme_rdma_poll_group_remove, 3348 .poll_group_process_completions = nvme_rdma_poll_group_process_completions, 3349 .poll_group_destroy = nvme_rdma_poll_group_destroy, 3350 .poll_group_get_stats = nvme_rdma_poll_group_get_stats, 3351 .poll_group_free_stats = nvme_rdma_poll_group_free_stats, 3352 }; 3353 3354 SPDK_NVME_TRANSPORT_REGISTER(rdma, &rdma_ops); 3355