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