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_AUTHENTICATING, 160 NVME_RDMA_QPAIR_STATE_RUNNING, 161 NVME_RDMA_QPAIR_STATE_EXITING, 162 NVME_RDMA_QPAIR_STATE_LINGERING, 163 NVME_RDMA_QPAIR_STATE_EXITED, 164 }; 165 166 typedef int (*nvme_rdma_cm_event_cb)(struct nvme_rdma_qpair *rqpair, int ret); 167 168 struct nvme_rdma_rsp_opts { 169 uint16_t num_entries; 170 struct nvme_rdma_qpair *rqpair; 171 struct spdk_rdma_provider_srq *srq; 172 struct spdk_rdma_utils_mem_map *mr_map; 173 }; 174 175 struct nvme_rdma_rsps { 176 /* Parallel arrays of response buffers + response SGLs of size num_entries */ 177 struct ibv_sge *rsp_sgls; 178 struct spdk_nvme_rdma_rsp *rsps; 179 180 struct ibv_recv_wr *rsp_recv_wrs; 181 182 /* Count of outstanding recv objects */ 183 uint16_t current_num_recvs; 184 185 uint16_t num_entries; 186 }; 187 188 /* NVMe RDMA qpair extensions for spdk_nvme_qpair */ 189 struct nvme_rdma_qpair { 190 struct spdk_nvme_qpair qpair; 191 192 struct spdk_rdma_provider_qp *rdma_qp; 193 struct rdma_cm_id *cm_id; 194 struct ibv_cq *cq; 195 struct spdk_rdma_provider_srq *srq; 196 197 struct spdk_nvme_rdma_req *rdma_reqs; 198 199 uint32_t max_send_sge; 200 201 uint32_t max_recv_sge; 202 203 uint16_t num_entries; 204 205 bool delay_cmd_submit; 206 207 uint32_t num_completions; 208 uint32_t num_outstanding_reqs; 209 210 struct nvme_rdma_rsps *rsps; 211 212 /* 213 * Array of num_entries NVMe commands registered as RDMA message buffers. 214 * Indexed by rdma_req->id. 215 */ 216 struct spdk_nvmf_cmd *cmds; 217 218 struct spdk_rdma_utils_mem_map *mr_map; 219 220 TAILQ_HEAD(, spdk_nvme_rdma_req) free_reqs; 221 TAILQ_HEAD(, spdk_nvme_rdma_req) outstanding_reqs; 222 223 struct spdk_memory_domain *memory_domain; 224 225 /* Count of outstanding send objects */ 226 uint16_t current_num_sends; 227 228 TAILQ_ENTRY(nvme_rdma_qpair) link_active; 229 230 /* Placed at the end of the struct since it is not used frequently */ 231 struct rdma_cm_event *evt; 232 struct nvme_rdma_poller *poller; 233 234 uint64_t evt_timeout_ticks; 235 nvme_rdma_cm_event_cb evt_cb; 236 enum rdma_cm_event_type expected_evt_type; 237 238 enum nvme_rdma_qpair_state state; 239 240 bool in_connect_poll; 241 242 uint8_t stale_conn_retry_count; 243 bool need_destroy; 244 245 TAILQ_ENTRY(nvme_rdma_qpair) link_connecting; 246 }; 247 248 enum NVME_RDMA_COMPLETION_FLAGS { 249 NVME_RDMA_SEND_COMPLETED = 1u << 0, 250 NVME_RDMA_RECV_COMPLETED = 1u << 1, 251 }; 252 253 struct spdk_nvme_rdma_req { 254 uint16_t id; 255 uint16_t completion_flags: 2; 256 uint16_t reserved: 14; 257 /* if completion of RDMA_RECV received before RDMA_SEND, we will complete nvme request 258 * during processing of RDMA_SEND. To complete the request we must know the response 259 * received in RDMA_RECV, so store it in this field */ 260 struct spdk_nvme_rdma_rsp *rdma_rsp; 261 262 struct nvme_rdma_wr rdma_wr; 263 264 struct ibv_send_wr send_wr; 265 266 struct nvme_request *req; 267 268 struct ibv_sge send_sgl[NVME_RDMA_DEFAULT_TX_SGE]; 269 270 TAILQ_ENTRY(spdk_nvme_rdma_req) link; 271 }; 272 273 struct spdk_nvme_rdma_rsp { 274 struct spdk_nvme_cpl cpl; 275 struct nvme_rdma_qpair *rqpair; 276 struct ibv_recv_wr *recv_wr; 277 struct nvme_rdma_wr rdma_wr; 278 }; 279 280 struct nvme_rdma_memory_translation_ctx { 281 void *addr; 282 size_t length; 283 uint32_t lkey; 284 uint32_t rkey; 285 }; 286 287 static const char *rdma_cm_event_str[] = { 288 "RDMA_CM_EVENT_ADDR_RESOLVED", 289 "RDMA_CM_EVENT_ADDR_ERROR", 290 "RDMA_CM_EVENT_ROUTE_RESOLVED", 291 "RDMA_CM_EVENT_ROUTE_ERROR", 292 "RDMA_CM_EVENT_CONNECT_REQUEST", 293 "RDMA_CM_EVENT_CONNECT_RESPONSE", 294 "RDMA_CM_EVENT_CONNECT_ERROR", 295 "RDMA_CM_EVENT_UNREACHABLE", 296 "RDMA_CM_EVENT_REJECTED", 297 "RDMA_CM_EVENT_ESTABLISHED", 298 "RDMA_CM_EVENT_DISCONNECTED", 299 "RDMA_CM_EVENT_DEVICE_REMOVAL", 300 "RDMA_CM_EVENT_MULTICAST_JOIN", 301 "RDMA_CM_EVENT_MULTICAST_ERROR", 302 "RDMA_CM_EVENT_ADDR_CHANGE", 303 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 304 }; 305 306 static struct nvme_rdma_poller *nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, 307 struct ibv_context *device); 308 static void nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, 309 struct nvme_rdma_poller *poller); 310 311 static int nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, 312 struct spdk_nvme_qpair *qpair); 313 314 static inline struct nvme_rdma_qpair * 315 nvme_rdma_qpair(struct spdk_nvme_qpair *qpair) 316 { 317 assert(qpair->trtype == SPDK_NVME_TRANSPORT_RDMA); 318 return SPDK_CONTAINEROF(qpair, struct nvme_rdma_qpair, qpair); 319 } 320 321 static inline struct nvme_rdma_poll_group * 322 nvme_rdma_poll_group(struct spdk_nvme_transport_poll_group *group) 323 { 324 return (SPDK_CONTAINEROF(group, struct nvme_rdma_poll_group, group)); 325 } 326 327 static inline struct nvme_rdma_ctrlr * 328 nvme_rdma_ctrlr(struct spdk_nvme_ctrlr *ctrlr) 329 { 330 assert(ctrlr->trid.trtype == SPDK_NVME_TRANSPORT_RDMA); 331 return SPDK_CONTAINEROF(ctrlr, struct nvme_rdma_ctrlr, ctrlr); 332 } 333 334 static struct spdk_nvme_rdma_req * 335 nvme_rdma_req_get(struct nvme_rdma_qpair *rqpair) 336 { 337 struct spdk_nvme_rdma_req *rdma_req; 338 339 rdma_req = TAILQ_FIRST(&rqpair->free_reqs); 340 if (rdma_req) { 341 TAILQ_REMOVE(&rqpair->free_reqs, rdma_req, link); 342 } 343 344 return rdma_req; 345 } 346 347 static void 348 nvme_rdma_req_put(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req) 349 { 350 rdma_req->completion_flags = 0; 351 rdma_req->req = NULL; 352 rdma_req->rdma_rsp = NULL; 353 TAILQ_INSERT_HEAD(&rqpair->free_reqs, rdma_req, link); 354 } 355 356 static void 357 nvme_rdma_req_complete(struct spdk_nvme_rdma_req *rdma_req, 358 struct spdk_nvme_cpl *rsp, 359 bool print_on_error) 360 { 361 struct nvme_request *req = rdma_req->req; 362 struct nvme_rdma_qpair *rqpair; 363 struct spdk_nvme_qpair *qpair; 364 bool error, print_error; 365 366 assert(req != NULL); 367 368 qpair = req->qpair; 369 rqpair = nvme_rdma_qpair(qpair); 370 371 error = spdk_nvme_cpl_is_error(rsp); 372 print_error = error && print_on_error && !qpair->ctrlr->opts.disable_error_logging; 373 374 if (print_error) { 375 spdk_nvme_qpair_print_command(qpair, &req->cmd); 376 } 377 378 if (print_error || SPDK_DEBUGLOG_FLAG_ENABLED("nvme")) { 379 spdk_nvme_qpair_print_completion(qpair, rsp); 380 } 381 382 assert(rqpair->num_outstanding_reqs > 0); 383 rqpair->num_outstanding_reqs--; 384 385 TAILQ_REMOVE(&rqpair->outstanding_reqs, rdma_req, link); 386 387 nvme_complete_request(req->cb_fn, req->cb_arg, qpair, req, rsp); 388 nvme_rdma_req_put(rqpair, rdma_req); 389 } 390 391 static const char * 392 nvme_rdma_cm_event_str_get(uint32_t event) 393 { 394 if (event < SPDK_COUNTOF(rdma_cm_event_str)) { 395 return rdma_cm_event_str[event]; 396 } else { 397 return "Undefined"; 398 } 399 } 400 401 402 static int 403 nvme_rdma_qpair_process_cm_event(struct nvme_rdma_qpair *rqpair) 404 { 405 struct rdma_cm_event *event = rqpair->evt; 406 struct spdk_nvmf_rdma_accept_private_data *accept_data; 407 int rc = 0; 408 409 if (event) { 410 switch (event->event) { 411 case RDMA_CM_EVENT_ADDR_RESOLVED: 412 case RDMA_CM_EVENT_ADDR_ERROR: 413 case RDMA_CM_EVENT_ROUTE_RESOLVED: 414 case RDMA_CM_EVENT_ROUTE_ERROR: 415 break; 416 case RDMA_CM_EVENT_CONNECT_REQUEST: 417 break; 418 case RDMA_CM_EVENT_CONNECT_ERROR: 419 break; 420 case RDMA_CM_EVENT_UNREACHABLE: 421 case RDMA_CM_EVENT_REJECTED: 422 break; 423 case RDMA_CM_EVENT_CONNECT_RESPONSE: 424 rc = spdk_rdma_provider_qp_complete_connect(rqpair->rdma_qp); 425 /* fall through */ 426 case RDMA_CM_EVENT_ESTABLISHED: 427 accept_data = (struct spdk_nvmf_rdma_accept_private_data *)event->param.conn.private_data; 428 if (accept_data == NULL) { 429 rc = -1; 430 } else { 431 SPDK_DEBUGLOG(nvme, "Requested queue depth %d. Target receive queue depth %d.\n", 432 rqpair->num_entries + 1, accept_data->crqsize); 433 } 434 break; 435 case RDMA_CM_EVENT_DISCONNECTED: 436 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE; 437 break; 438 case RDMA_CM_EVENT_DEVICE_REMOVAL: 439 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL; 440 rqpair->need_destroy = true; 441 break; 442 case RDMA_CM_EVENT_MULTICAST_JOIN: 443 case RDMA_CM_EVENT_MULTICAST_ERROR: 444 break; 445 case RDMA_CM_EVENT_ADDR_CHANGE: 446 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL; 447 break; 448 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 449 break; 450 default: 451 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 452 break; 453 } 454 rqpair->evt = NULL; 455 rdma_ack_cm_event(event); 456 } 457 458 return rc; 459 } 460 461 /* 462 * This function must be called under the nvme controller's lock 463 * because it touches global controller variables. The lock is taken 464 * by the generic transport code before invoking a few of the functions 465 * in this file: nvme_rdma_ctrlr_connect_qpair, nvme_rdma_ctrlr_delete_io_qpair, 466 * and conditionally nvme_rdma_qpair_process_completions when it is calling 467 * completions on the admin qpair. When adding a new call to this function, please 468 * verify that it is in a situation where it falls under the lock. 469 */ 470 static int 471 nvme_rdma_poll_events(struct nvme_rdma_ctrlr *rctrlr) 472 { 473 struct nvme_rdma_cm_event_entry *entry, *tmp; 474 struct nvme_rdma_qpair *event_qpair; 475 struct rdma_cm_event *event; 476 struct rdma_event_channel *channel = rctrlr->cm_channel; 477 478 STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) { 479 event_qpair = entry->evt->id->context; 480 if (event_qpair->evt == NULL) { 481 event_qpair->evt = entry->evt; 482 STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link); 483 STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link); 484 } 485 } 486 487 while (rdma_get_cm_event(channel, &event) == 0) { 488 event_qpair = event->id->context; 489 if (event_qpair->evt == NULL) { 490 event_qpair->evt = event; 491 } else { 492 assert(rctrlr == nvme_rdma_ctrlr(event_qpair->qpair.ctrlr)); 493 entry = STAILQ_FIRST(&rctrlr->free_cm_events); 494 if (entry == NULL) { 495 rdma_ack_cm_event(event); 496 return -ENOMEM; 497 } 498 STAILQ_REMOVE(&rctrlr->free_cm_events, entry, nvme_rdma_cm_event_entry, link); 499 entry->evt = event; 500 STAILQ_INSERT_TAIL(&rctrlr->pending_cm_events, entry, link); 501 } 502 } 503 504 /* rdma_get_cm_event() returns -1 on error. If an error occurs, errno 505 * will be set to indicate the failure reason. So return negated errno here. 506 */ 507 return -errno; 508 } 509 510 static int 511 nvme_rdma_validate_cm_event(enum rdma_cm_event_type expected_evt_type, 512 struct rdma_cm_event *reaped_evt) 513 { 514 int rc = -EBADMSG; 515 516 if (expected_evt_type == reaped_evt->event) { 517 return 0; 518 } 519 520 switch (expected_evt_type) { 521 case RDMA_CM_EVENT_ESTABLISHED: 522 /* 523 * There is an enum ib_cm_rej_reason in the kernel headers that sets 10 as 524 * IB_CM_REJ_STALE_CONN. I can't find the corresponding userspace but we get 525 * the same values here. 526 */ 527 if (reaped_evt->event == RDMA_CM_EVENT_REJECTED && reaped_evt->status == 10) { 528 rc = -ESTALE; 529 } else if (reaped_evt->event == RDMA_CM_EVENT_CONNECT_RESPONSE) { 530 /* 531 * If we are using a qpair which is not created using rdma cm API 532 * then we will receive RDMA_CM_EVENT_CONNECT_RESPONSE instead of 533 * RDMA_CM_EVENT_ESTABLISHED. 534 */ 535 return 0; 536 } 537 break; 538 default: 539 break; 540 } 541 542 SPDK_ERRLOG("Expected %s but received %s (%d) from CM event channel (status = %d)\n", 543 nvme_rdma_cm_event_str_get(expected_evt_type), 544 nvme_rdma_cm_event_str_get(reaped_evt->event), reaped_evt->event, 545 reaped_evt->status); 546 return rc; 547 } 548 549 static int 550 nvme_rdma_process_event_start(struct nvme_rdma_qpair *rqpair, 551 enum rdma_cm_event_type evt, 552 nvme_rdma_cm_event_cb evt_cb) 553 { 554 int rc; 555 556 assert(evt_cb != NULL); 557 558 if (rqpair->evt != NULL) { 559 rc = nvme_rdma_qpair_process_cm_event(rqpair); 560 if (rc) { 561 return rc; 562 } 563 } 564 565 rqpair->expected_evt_type = evt; 566 rqpair->evt_cb = evt_cb; 567 rqpair->evt_timeout_ticks = (g_spdk_nvme_transport_opts.rdma_cm_event_timeout_ms * 1000 * 568 spdk_get_ticks_hz()) / SPDK_SEC_TO_USEC + spdk_get_ticks(); 569 570 return 0; 571 } 572 573 static int 574 nvme_rdma_process_event_poll(struct nvme_rdma_qpair *rqpair) 575 { 576 struct nvme_rdma_ctrlr *rctrlr; 577 int rc = 0, rc2; 578 579 rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr); 580 assert(rctrlr != NULL); 581 582 if (!rqpair->evt && spdk_get_ticks() < rqpair->evt_timeout_ticks) { 583 rc = nvme_rdma_poll_events(rctrlr); 584 if (rc == -EAGAIN || rc == -EWOULDBLOCK) { 585 return rc; 586 } 587 } 588 589 if (rqpair->evt == NULL) { 590 rc = -EADDRNOTAVAIL; 591 goto exit; 592 } 593 594 rc = nvme_rdma_validate_cm_event(rqpair->expected_evt_type, rqpair->evt); 595 596 rc2 = nvme_rdma_qpair_process_cm_event(rqpair); 597 /* bad message takes precedence over the other error codes from processing the event. */ 598 rc = rc == 0 ? rc2 : rc; 599 600 exit: 601 assert(rqpair->evt_cb != NULL); 602 return rqpair->evt_cb(rqpair, rc); 603 } 604 605 static int 606 nvme_rdma_resize_cq(struct nvme_rdma_qpair *rqpair, struct nvme_rdma_poller *poller) 607 { 608 int current_num_wc, required_num_wc; 609 int max_cq_size; 610 611 required_num_wc = poller->required_num_wc + WC_PER_QPAIR(rqpair->num_entries); 612 current_num_wc = poller->current_num_wc; 613 if (current_num_wc < required_num_wc) { 614 current_num_wc = spdk_max(current_num_wc * 2, required_num_wc); 615 } 616 617 max_cq_size = g_spdk_nvme_transport_opts.rdma_max_cq_size; 618 if (max_cq_size != 0 && current_num_wc > max_cq_size) { 619 current_num_wc = max_cq_size; 620 } 621 622 if (poller->current_num_wc != current_num_wc) { 623 SPDK_DEBUGLOG(nvme, "Resize RDMA CQ from %d to %d\n", poller->current_num_wc, 624 current_num_wc); 625 if (ibv_resize_cq(poller->cq, current_num_wc)) { 626 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 627 return -1; 628 } 629 630 poller->current_num_wc = current_num_wc; 631 } 632 633 poller->required_num_wc = required_num_wc; 634 return 0; 635 } 636 637 static int 638 nvme_rdma_qpair_set_poller(struct spdk_nvme_qpair *qpair) 639 { 640 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 641 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group); 642 struct nvme_rdma_poller *poller; 643 644 assert(rqpair->cq == NULL); 645 646 poller = nvme_rdma_poll_group_get_poller(group, rqpair->cm_id->verbs); 647 if (!poller) { 648 SPDK_ERRLOG("Unable to find a cq for qpair %p on poll group %p\n", qpair, qpair->poll_group); 649 return -EINVAL; 650 } 651 652 if (!poller->srq) { 653 if (nvme_rdma_resize_cq(rqpair, poller)) { 654 nvme_rdma_poll_group_put_poller(group, poller); 655 return -EPROTO; 656 } 657 } 658 659 rqpair->cq = poller->cq; 660 rqpair->srq = poller->srq; 661 if (rqpair->srq) { 662 rqpair->rsps = poller->rsps; 663 } 664 rqpair->poller = poller; 665 return 0; 666 } 667 668 static int 669 nvme_rdma_qpair_init(struct nvme_rdma_qpair *rqpair) 670 { 671 int rc; 672 struct spdk_rdma_provider_qp_init_attr attr = {}; 673 struct ibv_device_attr dev_attr; 674 struct nvme_rdma_ctrlr *rctrlr; 675 uint32_t num_cqe, max_num_cqe; 676 677 rc = ibv_query_device(rqpair->cm_id->verbs, &dev_attr); 678 if (rc != 0) { 679 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 680 return -1; 681 } 682 683 if (rqpair->qpair.poll_group) { 684 assert(!rqpair->cq); 685 rc = nvme_rdma_qpair_set_poller(&rqpair->qpair); 686 if (rc) { 687 SPDK_ERRLOG("Unable to activate the rdmaqpair.\n"); 688 return -1; 689 } 690 assert(rqpair->cq); 691 } else { 692 num_cqe = rqpair->num_entries * 2; 693 max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size; 694 if (max_num_cqe != 0 && num_cqe > max_num_cqe) { 695 num_cqe = max_num_cqe; 696 } 697 rqpair->cq = ibv_create_cq(rqpair->cm_id->verbs, num_cqe, rqpair, NULL, 0); 698 if (!rqpair->cq) { 699 SPDK_ERRLOG("Unable to create completion queue: errno %d: %s\n", errno, spdk_strerror(errno)); 700 return -1; 701 } 702 } 703 704 rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr); 705 if (g_nvme_hooks.get_ibv_pd) { 706 attr.pd = g_nvme_hooks.get_ibv_pd(&rctrlr->ctrlr.trid, rqpair->cm_id->verbs); 707 } else { 708 attr.pd = spdk_rdma_utils_get_pd(rqpair->cm_id->verbs); 709 } 710 711 attr.stats = rqpair->poller ? &rqpair->poller->stats.rdma_stats : NULL; 712 attr.send_cq = rqpair->cq; 713 attr.recv_cq = rqpair->cq; 714 attr.cap.max_send_wr = rqpair->num_entries; /* SEND operations */ 715 if (rqpair->srq) { 716 attr.srq = rqpair->srq->srq; 717 } else { 718 attr.cap.max_recv_wr = rqpair->num_entries; /* RECV operations */ 719 } 720 attr.cap.max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, dev_attr.max_sge); 721 attr.cap.max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, dev_attr.max_sge); 722 723 rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &attr); 724 725 if (!rqpair->rdma_qp) { 726 return -1; 727 } 728 729 rqpair->memory_domain = spdk_rdma_utils_get_memory_domain(rqpair->rdma_qp->qp->pd); 730 if (!rqpair->memory_domain) { 731 SPDK_ERRLOG("Failed to get memory domain\n"); 732 return -1; 733 } 734 735 /* ibv_create_qp will change the values in attr.cap. Make sure we store the proper value. */ 736 rqpair->max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, attr.cap.max_send_sge); 737 rqpair->max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, attr.cap.max_recv_sge); 738 rqpair->current_num_sends = 0; 739 740 rqpair->cm_id->context = rqpair; 741 742 return 0; 743 } 744 745 static void 746 nvme_rdma_reset_failed_sends(struct nvme_rdma_qpair *rqpair, 747 struct ibv_send_wr *bad_send_wr, int rc) 748 { 749 SPDK_ERRLOG("Failed to post WRs on send queue, errno %d (%s), bad_wr %p\n", 750 rc, spdk_strerror(rc), bad_send_wr); 751 while (bad_send_wr != NULL) { 752 assert(rqpair->current_num_sends > 0); 753 rqpair->current_num_sends--; 754 bad_send_wr = bad_send_wr->next; 755 } 756 } 757 758 static void 759 nvme_rdma_reset_failed_recvs(struct nvme_rdma_rsps *rsps, 760 struct ibv_recv_wr *bad_recv_wr, int rc) 761 { 762 SPDK_ERRLOG("Failed to post WRs on receive queue, errno %d (%s), bad_wr %p\n", 763 rc, spdk_strerror(rc), bad_recv_wr); 764 while (bad_recv_wr != NULL) { 765 assert(rsps->current_num_recvs > 0); 766 rsps->current_num_recvs--; 767 bad_recv_wr = bad_recv_wr->next; 768 } 769 } 770 771 static inline int 772 nvme_rdma_qpair_submit_sends(struct nvme_rdma_qpair *rqpair) 773 { 774 struct ibv_send_wr *bad_send_wr = NULL; 775 int rc; 776 777 rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_send_wr); 778 779 if (spdk_unlikely(rc)) { 780 nvme_rdma_reset_failed_sends(rqpair, bad_send_wr, rc); 781 } 782 783 return rc; 784 } 785 786 static inline int 787 nvme_rdma_qpair_submit_recvs(struct nvme_rdma_qpair *rqpair) 788 { 789 struct ibv_recv_wr *bad_recv_wr; 790 int rc = 0; 791 792 rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 793 if (spdk_unlikely(rc)) { 794 nvme_rdma_reset_failed_recvs(rqpair->rsps, bad_recv_wr, rc); 795 } 796 797 return rc; 798 } 799 800 static inline int 801 nvme_rdma_poller_submit_recvs(struct nvme_rdma_poller *poller) 802 { 803 struct ibv_recv_wr *bad_recv_wr; 804 int rc; 805 806 rc = spdk_rdma_provider_srq_flush_recv_wrs(poller->srq, &bad_recv_wr); 807 if (spdk_unlikely(rc)) { 808 nvme_rdma_reset_failed_recvs(poller->rsps, bad_recv_wr, rc); 809 } 810 811 return rc; 812 } 813 814 #define nvme_rdma_trace_ibv_sge(sg_list) \ 815 if (sg_list) { \ 816 SPDK_DEBUGLOG(nvme, "local addr %p length 0x%x lkey 0x%x\n", \ 817 (void *)(sg_list)->addr, (sg_list)->length, (sg_list)->lkey); \ 818 } 819 820 static void 821 nvme_rdma_free_rsps(struct nvme_rdma_rsps *rsps) 822 { 823 if (!rsps) { 824 return; 825 } 826 827 spdk_free(rsps->rsps); 828 spdk_free(rsps->rsp_sgls); 829 spdk_free(rsps->rsp_recv_wrs); 830 spdk_free(rsps); 831 } 832 833 static struct nvme_rdma_rsps * 834 nvme_rdma_create_rsps(struct nvme_rdma_rsp_opts *opts) 835 { 836 struct nvme_rdma_rsps *rsps; 837 struct spdk_rdma_utils_memory_translation translation; 838 uint16_t i; 839 int rc; 840 841 rsps = spdk_zmalloc(sizeof(*rsps), 0, NULL, SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 842 if (!rsps) { 843 SPDK_ERRLOG("Failed to allocate rsps object\n"); 844 return NULL; 845 } 846 847 rsps->rsp_sgls = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_sgls), 0, NULL, 848 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 849 if (!rsps->rsp_sgls) { 850 SPDK_ERRLOG("Failed to allocate rsp_sgls\n"); 851 goto fail; 852 } 853 854 rsps->rsp_recv_wrs = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_recv_wrs), 0, NULL, 855 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 856 if (!rsps->rsp_recv_wrs) { 857 SPDK_ERRLOG("Failed to allocate rsp_recv_wrs\n"); 858 goto fail; 859 } 860 861 rsps->rsps = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsps), 0, NULL, 862 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 863 if (!rsps->rsps) { 864 SPDK_ERRLOG("can not allocate rdma rsps\n"); 865 goto fail; 866 } 867 868 for (i = 0; i < opts->num_entries; i++) { 869 struct ibv_sge *rsp_sgl = &rsps->rsp_sgls[i]; 870 struct spdk_nvme_rdma_rsp *rsp = &rsps->rsps[i]; 871 struct ibv_recv_wr *recv_wr = &rsps->rsp_recv_wrs[i]; 872 873 rsp->rqpair = opts->rqpair; 874 rsp->rdma_wr.type = RDMA_WR_TYPE_RECV; 875 rsp->recv_wr = recv_wr; 876 rsp_sgl->addr = (uint64_t)rsp; 877 rsp_sgl->length = sizeof(struct spdk_nvme_cpl); 878 rc = spdk_rdma_utils_get_translation(opts->mr_map, rsp, sizeof(*rsp), &translation); 879 if (rc) { 880 goto fail; 881 } 882 rsp_sgl->lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 883 884 recv_wr->wr_id = (uint64_t)&rsp->rdma_wr; 885 recv_wr->next = NULL; 886 recv_wr->sg_list = rsp_sgl; 887 recv_wr->num_sge = 1; 888 889 nvme_rdma_trace_ibv_sge(recv_wr->sg_list); 890 891 if (opts->rqpair) { 892 spdk_rdma_provider_qp_queue_recv_wrs(opts->rqpair->rdma_qp, recv_wr); 893 } else { 894 spdk_rdma_provider_srq_queue_recv_wrs(opts->srq, recv_wr); 895 } 896 } 897 898 rsps->num_entries = opts->num_entries; 899 rsps->current_num_recvs = opts->num_entries; 900 901 return rsps; 902 fail: 903 nvme_rdma_free_rsps(rsps); 904 return NULL; 905 } 906 907 static void 908 nvme_rdma_free_reqs(struct nvme_rdma_qpair *rqpair) 909 { 910 if (!rqpair->rdma_reqs) { 911 return; 912 } 913 914 spdk_free(rqpair->cmds); 915 rqpair->cmds = NULL; 916 917 spdk_free(rqpair->rdma_reqs); 918 rqpair->rdma_reqs = NULL; 919 } 920 921 static int 922 nvme_rdma_create_reqs(struct nvme_rdma_qpair *rqpair) 923 { 924 struct spdk_rdma_utils_memory_translation translation; 925 uint16_t i; 926 int rc; 927 928 assert(!rqpair->rdma_reqs); 929 rqpair->rdma_reqs = spdk_zmalloc(rqpair->num_entries * sizeof(struct spdk_nvme_rdma_req), 0, NULL, 930 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 931 if (rqpair->rdma_reqs == NULL) { 932 SPDK_ERRLOG("Failed to allocate rdma_reqs\n"); 933 goto fail; 934 } 935 936 assert(!rqpair->cmds); 937 rqpair->cmds = spdk_zmalloc(rqpair->num_entries * sizeof(*rqpair->cmds), 0, NULL, 938 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 939 if (!rqpair->cmds) { 940 SPDK_ERRLOG("Failed to allocate RDMA cmds\n"); 941 goto fail; 942 } 943 944 TAILQ_INIT(&rqpair->free_reqs); 945 TAILQ_INIT(&rqpair->outstanding_reqs); 946 for (i = 0; i < rqpair->num_entries; i++) { 947 struct spdk_nvme_rdma_req *rdma_req; 948 struct spdk_nvmf_cmd *cmd; 949 950 rdma_req = &rqpair->rdma_reqs[i]; 951 rdma_req->rdma_wr.type = RDMA_WR_TYPE_SEND; 952 cmd = &rqpair->cmds[i]; 953 954 rdma_req->id = i; 955 956 rc = spdk_rdma_utils_get_translation(rqpair->mr_map, cmd, sizeof(*cmd), &translation); 957 if (rc) { 958 goto fail; 959 } 960 rdma_req->send_sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 961 962 /* The first RDMA sgl element will always point 963 * at this data structure. Depending on whether 964 * an NVMe-oF SGL is required, the length of 965 * this element may change. */ 966 rdma_req->send_sgl[0].addr = (uint64_t)cmd; 967 rdma_req->send_wr.wr_id = (uint64_t)&rdma_req->rdma_wr; 968 rdma_req->send_wr.next = NULL; 969 rdma_req->send_wr.opcode = IBV_WR_SEND; 970 rdma_req->send_wr.send_flags = IBV_SEND_SIGNALED; 971 rdma_req->send_wr.sg_list = rdma_req->send_sgl; 972 rdma_req->send_wr.imm_data = 0; 973 974 TAILQ_INSERT_TAIL(&rqpair->free_reqs, rdma_req, link); 975 } 976 977 return 0; 978 fail: 979 nvme_rdma_free_reqs(rqpair); 980 return -ENOMEM; 981 } 982 983 static int nvme_rdma_connect(struct nvme_rdma_qpair *rqpair); 984 985 static int 986 nvme_rdma_route_resolved(struct nvme_rdma_qpair *rqpair, int ret) 987 { 988 if (ret) { 989 SPDK_ERRLOG("RDMA route resolution error\n"); 990 return -1; 991 } 992 993 ret = nvme_rdma_qpair_init(rqpair); 994 if (ret < 0) { 995 SPDK_ERRLOG("nvme_rdma_qpair_init() failed\n"); 996 return -1; 997 } 998 999 return nvme_rdma_connect(rqpair); 1000 } 1001 1002 static int 1003 nvme_rdma_addr_resolved(struct nvme_rdma_qpair *rqpair, int ret) 1004 { 1005 if (ret) { 1006 SPDK_ERRLOG("RDMA address resolution error\n"); 1007 return -1; 1008 } 1009 1010 if (rqpair->qpair.ctrlr->opts.transport_ack_timeout != SPDK_NVME_TRANSPORT_ACK_TIMEOUT_DISABLED) { 1011 #ifdef SPDK_CONFIG_RDMA_SET_ACK_TIMEOUT 1012 uint8_t timeout = rqpair->qpair.ctrlr->opts.transport_ack_timeout; 1013 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, 1014 RDMA_OPTION_ID_ACK_TIMEOUT, 1015 &timeout, sizeof(timeout)); 1016 if (ret) { 1017 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_ACK_TIMEOUT %d, ret %d\n", timeout, ret); 1018 } 1019 #else 1020 SPDK_DEBUGLOG(nvme, "transport_ack_timeout is not supported\n"); 1021 #endif 1022 } 1023 1024 if (rqpair->qpair.ctrlr->opts.transport_tos != SPDK_NVME_TRANSPORT_TOS_DISABLED) { 1025 #ifdef SPDK_CONFIG_RDMA_SET_TOS 1026 uint8_t tos = rqpair->qpair.ctrlr->opts.transport_tos; 1027 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_TOS, &tos, sizeof(tos)); 1028 if (ret) { 1029 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_TOS %u, ret %d\n", tos, ret); 1030 } 1031 #else 1032 SPDK_DEBUGLOG(nvme, "transport_tos is not supported\n"); 1033 #endif 1034 } 1035 1036 ret = rdma_resolve_route(rqpair->cm_id, NVME_RDMA_TIME_OUT_IN_MS); 1037 if (ret) { 1038 SPDK_ERRLOG("rdma_resolve_route\n"); 1039 return ret; 1040 } 1041 1042 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ROUTE_RESOLVED, 1043 nvme_rdma_route_resolved); 1044 } 1045 1046 static int 1047 nvme_rdma_resolve_addr(struct nvme_rdma_qpair *rqpair, 1048 struct sockaddr *src_addr, 1049 struct sockaddr *dst_addr) 1050 { 1051 int ret; 1052 1053 if (src_addr) { 1054 int reuse = 1; 1055 1056 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_REUSEADDR, 1057 &reuse, sizeof(reuse)); 1058 if (ret) { 1059 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_REUSEADDR %d, ret %d\n", 1060 reuse, ret); 1061 /* It is likely that rdma_resolve_addr() returns -EADDRINUSE, but 1062 * we may missing something. We rely on rdma_resolve_addr(). 1063 */ 1064 } 1065 } 1066 1067 ret = rdma_resolve_addr(rqpair->cm_id, src_addr, dst_addr, 1068 NVME_RDMA_TIME_OUT_IN_MS); 1069 if (ret) { 1070 SPDK_ERRLOG("rdma_resolve_addr, %d\n", errno); 1071 return ret; 1072 } 1073 1074 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ADDR_RESOLVED, 1075 nvme_rdma_addr_resolved); 1076 } 1077 1078 static int nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair); 1079 1080 static int 1081 nvme_rdma_connect_established(struct nvme_rdma_qpair *rqpair, int ret) 1082 { 1083 struct nvme_rdma_rsp_opts opts = {}; 1084 1085 if (ret == -ESTALE) { 1086 return nvme_rdma_stale_conn_retry(rqpair); 1087 } else if (ret) { 1088 SPDK_ERRLOG("RDMA connect error %d\n", ret); 1089 return ret; 1090 } 1091 1092 assert(!rqpair->mr_map); 1093 rqpair->mr_map = spdk_rdma_utils_create_mem_map(rqpair->rdma_qp->qp->pd, &g_nvme_hooks, 1094 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE); 1095 if (!rqpair->mr_map) { 1096 SPDK_ERRLOG("Unable to register RDMA memory translation map\n"); 1097 return -1; 1098 } 1099 1100 ret = nvme_rdma_create_reqs(rqpair); 1101 SPDK_DEBUGLOG(nvme, "rc =%d\n", ret); 1102 if (ret) { 1103 SPDK_ERRLOG("Unable to create rqpair RDMA requests\n"); 1104 return -1; 1105 } 1106 SPDK_DEBUGLOG(nvme, "RDMA requests created\n"); 1107 1108 if (!rqpair->srq) { 1109 opts.num_entries = rqpair->num_entries; 1110 opts.rqpair = rqpair; 1111 opts.srq = NULL; 1112 opts.mr_map = rqpair->mr_map; 1113 1114 assert(!rqpair->rsps); 1115 rqpair->rsps = nvme_rdma_create_rsps(&opts); 1116 if (!rqpair->rsps) { 1117 SPDK_ERRLOG("Unable to create rqpair RDMA responses\n"); 1118 return -1; 1119 } 1120 SPDK_DEBUGLOG(nvme, "RDMA responses created\n"); 1121 1122 ret = nvme_rdma_qpair_submit_recvs(rqpair); 1123 SPDK_DEBUGLOG(nvme, "rc =%d\n", ret); 1124 if (ret) { 1125 SPDK_ERRLOG("Unable to submit rqpair RDMA responses\n"); 1126 return -1; 1127 } 1128 SPDK_DEBUGLOG(nvme, "RDMA responses submitted\n"); 1129 } 1130 1131 rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND; 1132 1133 return 0; 1134 } 1135 1136 static int 1137 nvme_rdma_connect(struct nvme_rdma_qpair *rqpair) 1138 { 1139 struct rdma_conn_param param = {}; 1140 struct spdk_nvmf_rdma_request_private_data request_data = {}; 1141 struct ibv_device_attr attr; 1142 int ret; 1143 struct spdk_nvme_ctrlr *ctrlr; 1144 1145 ret = ibv_query_device(rqpair->cm_id->verbs, &attr); 1146 if (ret != 0) { 1147 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1148 return ret; 1149 } 1150 1151 param.responder_resources = attr.max_qp_rd_atom; 1152 1153 ctrlr = rqpair->qpair.ctrlr; 1154 if (!ctrlr) { 1155 return -1; 1156 } 1157 1158 request_data.qid = rqpair->qpair.id; 1159 request_data.hrqsize = rqpair->num_entries + 1; 1160 request_data.hsqsize = rqpair->num_entries; 1161 request_data.cntlid = ctrlr->cntlid; 1162 1163 param.private_data = &request_data; 1164 param.private_data_len = sizeof(request_data); 1165 param.retry_count = ctrlr->opts.transport_retry_count; 1166 param.rnr_retry_count = 7; 1167 1168 /* Fields below are ignored by rdma cm if qpair has been 1169 * created using rdma cm API. */ 1170 param.srq = 0; 1171 param.qp_num = rqpair->rdma_qp->qp->qp_num; 1172 1173 ret = rdma_connect(rqpair->cm_id, ¶m); 1174 if (ret) { 1175 SPDK_ERRLOG("nvme rdma connect error\n"); 1176 return ret; 1177 } 1178 1179 ctrlr->numa.id_valid = 1; 1180 ctrlr->numa.id = spdk_rdma_cm_id_get_numa_id(rqpair->cm_id); 1181 1182 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ESTABLISHED, 1183 nvme_rdma_connect_established); 1184 } 1185 1186 static int 1187 nvme_rdma_ctrlr_connect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 1188 { 1189 struct sockaddr_storage dst_addr; 1190 struct sockaddr_storage src_addr; 1191 bool src_addr_specified; 1192 long int port, src_port = 0; 1193 int rc; 1194 struct nvme_rdma_ctrlr *rctrlr; 1195 struct nvme_rdma_qpair *rqpair; 1196 struct nvme_rdma_poll_group *group; 1197 int family; 1198 1199 rqpair = nvme_rdma_qpair(qpair); 1200 rctrlr = nvme_rdma_ctrlr(ctrlr); 1201 assert(rctrlr != NULL); 1202 1203 switch (ctrlr->trid.adrfam) { 1204 case SPDK_NVMF_ADRFAM_IPV4: 1205 family = AF_INET; 1206 break; 1207 case SPDK_NVMF_ADRFAM_IPV6: 1208 family = AF_INET6; 1209 break; 1210 default: 1211 SPDK_ERRLOG("Unhandled ADRFAM %d\n", ctrlr->trid.adrfam); 1212 return -1; 1213 } 1214 1215 SPDK_DEBUGLOG(nvme, "adrfam %d ai_family %d\n", ctrlr->trid.adrfam, family); 1216 1217 memset(&dst_addr, 0, sizeof(dst_addr)); 1218 1219 SPDK_DEBUGLOG(nvme, "trsvcid is %s\n", ctrlr->trid.trsvcid); 1220 rc = nvme_parse_addr(&dst_addr, family, ctrlr->trid.traddr, ctrlr->trid.trsvcid, &port); 1221 if (rc != 0) { 1222 SPDK_ERRLOG("dst_addr nvme_parse_addr() failed\n"); 1223 return -1; 1224 } 1225 1226 if (ctrlr->opts.src_addr[0] || ctrlr->opts.src_svcid[0]) { 1227 memset(&src_addr, 0, sizeof(src_addr)); 1228 rc = nvme_parse_addr(&src_addr, family, 1229 ctrlr->opts.src_addr[0] ? ctrlr->opts.src_addr : NULL, 1230 ctrlr->opts.src_svcid[0] ? ctrlr->opts.src_svcid : NULL, 1231 &src_port); 1232 if (rc != 0) { 1233 SPDK_ERRLOG("src_addr nvme_parse_addr() failed\n"); 1234 return -1; 1235 } 1236 src_addr_specified = true; 1237 } else { 1238 src_addr_specified = false; 1239 } 1240 1241 rc = rdma_create_id(rctrlr->cm_channel, &rqpair->cm_id, rqpair, RDMA_PS_TCP); 1242 if (rc < 0) { 1243 SPDK_ERRLOG("rdma_create_id() failed\n"); 1244 return -1; 1245 } 1246 1247 rc = nvme_rdma_resolve_addr(rqpair, 1248 src_addr_specified ? (struct sockaddr *)&src_addr : NULL, 1249 (struct sockaddr *)&dst_addr); 1250 if (rc < 0) { 1251 SPDK_ERRLOG("nvme_rdma_resolve_addr() failed\n"); 1252 return -1; 1253 } 1254 1255 rqpair->state = NVME_RDMA_QPAIR_STATE_INITIALIZING; 1256 1257 if (qpair->poll_group != NULL) { 1258 group = nvme_rdma_poll_group(qpair->poll_group); 1259 TAILQ_INSERT_TAIL(&group->connecting_qpairs, rqpair, link_connecting); 1260 } 1261 1262 return 0; 1263 } 1264 1265 static int 1266 nvme_rdma_stale_conn_reconnect(struct nvme_rdma_qpair *rqpair) 1267 { 1268 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 1269 1270 if (spdk_get_ticks() < rqpair->evt_timeout_ticks) { 1271 return -EAGAIN; 1272 } 1273 1274 return nvme_rdma_ctrlr_connect_qpair(qpair->ctrlr, qpair); 1275 } 1276 1277 static int 1278 nvme_rdma_ctrlr_connect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, 1279 struct spdk_nvme_qpair *qpair) 1280 { 1281 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 1282 int rc; 1283 1284 if (rqpair->in_connect_poll) { 1285 return -EAGAIN; 1286 } 1287 1288 rqpair->in_connect_poll = true; 1289 1290 switch (rqpair->state) { 1291 case NVME_RDMA_QPAIR_STATE_INVALID: 1292 rc = -EAGAIN; 1293 break; 1294 1295 case NVME_RDMA_QPAIR_STATE_INITIALIZING: 1296 case NVME_RDMA_QPAIR_STATE_EXITING: 1297 if (!nvme_qpair_is_admin_queue(qpair)) { 1298 nvme_ctrlr_lock(ctrlr); 1299 } 1300 1301 rc = nvme_rdma_process_event_poll(rqpair); 1302 1303 if (!nvme_qpair_is_admin_queue(qpair)) { 1304 nvme_ctrlr_unlock(ctrlr); 1305 } 1306 1307 if (rc == 0) { 1308 rc = -EAGAIN; 1309 } 1310 rqpair->in_connect_poll = false; 1311 1312 return rc; 1313 1314 case NVME_RDMA_QPAIR_STATE_STALE_CONN: 1315 rc = nvme_rdma_stale_conn_reconnect(rqpair); 1316 if (rc == 0) { 1317 rc = -EAGAIN; 1318 } 1319 break; 1320 case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND: 1321 rc = nvme_fabric_qpair_connect_async(qpair, rqpair->num_entries + 1); 1322 if (rc == 0) { 1323 rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL; 1324 rc = -EAGAIN; 1325 } else { 1326 SPDK_ERRLOG("Failed to send an NVMe-oF Fabric CONNECT command\n"); 1327 } 1328 break; 1329 case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL: 1330 rc = nvme_fabric_qpair_connect_poll(qpair); 1331 if (rc == 0) { 1332 if (nvme_fabric_qpair_auth_required(qpair)) { 1333 rc = nvme_fabric_qpair_authenticate_async(qpair); 1334 if (rc == 0) { 1335 rqpair->state = NVME_RDMA_QPAIR_STATE_AUTHENTICATING; 1336 rc = -EAGAIN; 1337 } 1338 } else { 1339 rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING; 1340 nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED); 1341 } 1342 } else if (rc != -EAGAIN) { 1343 SPDK_ERRLOG("Failed to poll NVMe-oF Fabric CONNECT command\n"); 1344 } 1345 break; 1346 case NVME_RDMA_QPAIR_STATE_AUTHENTICATING: 1347 rc = nvme_fabric_qpair_authenticate_poll(qpair); 1348 if (rc == 0) { 1349 rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING; 1350 nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED); 1351 } 1352 break; 1353 case NVME_RDMA_QPAIR_STATE_RUNNING: 1354 rc = 0; 1355 break; 1356 default: 1357 assert(false); 1358 rc = -EINVAL; 1359 break; 1360 } 1361 1362 rqpair->in_connect_poll = false; 1363 1364 return rc; 1365 } 1366 1367 static inline int 1368 nvme_rdma_get_memory_translation(struct nvme_request *req, struct nvme_rdma_qpair *rqpair, 1369 struct nvme_rdma_memory_translation_ctx *_ctx) 1370 { 1371 struct spdk_memory_domain_translation_ctx ctx; 1372 struct spdk_memory_domain_translation_result dma_translation = {.iov_count = 0}; 1373 struct spdk_rdma_utils_memory_translation rdma_translation; 1374 int rc; 1375 1376 assert(req); 1377 assert(rqpair); 1378 assert(_ctx); 1379 1380 if (req->payload.opts && req->payload.opts->memory_domain) { 1381 ctx.size = sizeof(struct spdk_memory_domain_translation_ctx); 1382 ctx.rdma.ibv_qp = rqpair->rdma_qp->qp; 1383 dma_translation.size = sizeof(struct spdk_memory_domain_translation_result); 1384 1385 rc = spdk_memory_domain_translate_data(req->payload.opts->memory_domain, 1386 req->payload.opts->memory_domain_ctx, 1387 rqpair->memory_domain, &ctx, _ctx->addr, 1388 _ctx->length, &dma_translation); 1389 if (spdk_unlikely(rc) || dma_translation.iov_count != 1) { 1390 SPDK_ERRLOG("DMA memory translation failed, rc %d, iov count %u\n", rc, dma_translation.iov_count); 1391 return rc; 1392 } 1393 1394 _ctx->lkey = dma_translation.rdma.lkey; 1395 _ctx->rkey = dma_translation.rdma.rkey; 1396 _ctx->addr = dma_translation.iov.iov_base; 1397 _ctx->length = dma_translation.iov.iov_len; 1398 } else { 1399 rc = spdk_rdma_utils_get_translation(rqpair->mr_map, _ctx->addr, _ctx->length, &rdma_translation); 1400 if (spdk_unlikely(rc)) { 1401 SPDK_ERRLOG("RDMA memory translation failed, rc %d\n", rc); 1402 return rc; 1403 } 1404 if (rdma_translation.translation_type == SPDK_RDMA_UTILS_TRANSLATION_MR) { 1405 _ctx->lkey = rdma_translation.mr_or_key.mr->lkey; 1406 _ctx->rkey = rdma_translation.mr_or_key.mr->rkey; 1407 } else { 1408 _ctx->lkey = _ctx->rkey = (uint32_t)rdma_translation.mr_or_key.key; 1409 } 1410 } 1411 1412 return 0; 1413 } 1414 1415 1416 /* 1417 * Build SGL describing empty payload. 1418 */ 1419 static int 1420 nvme_rdma_build_null_request(struct spdk_nvme_rdma_req *rdma_req) 1421 { 1422 struct nvme_request *req = rdma_req->req; 1423 1424 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1425 1426 /* The first element of this SGL is pointing at an 1427 * spdk_nvmf_cmd object. For this particular command, 1428 * we only need the first 64 bytes corresponding to 1429 * the NVMe command. */ 1430 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1431 1432 /* The RDMA SGL needs one element describing the NVMe command. */ 1433 rdma_req->send_wr.num_sge = 1; 1434 1435 req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1436 req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1437 req->cmd.dptr.sgl1.keyed.length = 0; 1438 req->cmd.dptr.sgl1.keyed.key = 0; 1439 req->cmd.dptr.sgl1.address = 0; 1440 1441 return 0; 1442 } 1443 1444 /* 1445 * Build inline SGL describing contiguous payload buffer. 1446 */ 1447 static int 1448 nvme_rdma_build_contig_inline_request(struct nvme_rdma_qpair *rqpair, 1449 struct spdk_nvme_rdma_req *rdma_req) 1450 { 1451 struct nvme_request *req = rdma_req->req; 1452 struct nvme_rdma_memory_translation_ctx ctx = { 1453 .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset, 1454 .length = req->payload_size 1455 }; 1456 int rc; 1457 1458 assert(ctx.length != 0); 1459 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG); 1460 1461 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1462 if (spdk_unlikely(rc)) { 1463 return -1; 1464 } 1465 1466 rdma_req->send_sgl[1].lkey = ctx.lkey; 1467 1468 /* The first element of this SGL is pointing at an 1469 * spdk_nvmf_cmd object. For this particular command, 1470 * we only need the first 64 bytes corresponding to 1471 * the NVMe command. */ 1472 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1473 1474 rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr; 1475 rdma_req->send_sgl[1].length = (uint32_t)ctx.length; 1476 1477 /* The RDMA SGL contains two elements. The first describes 1478 * the NVMe command and the second describes the data 1479 * payload. */ 1480 rdma_req->send_wr.num_sge = 2; 1481 1482 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1483 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK; 1484 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1485 req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length; 1486 /* Inline only supported for icdoff == 0 currently. This function will 1487 * not get called for controllers with other values. */ 1488 req->cmd.dptr.sgl1.address = (uint64_t)0; 1489 1490 return 0; 1491 } 1492 1493 /* 1494 * Build SGL describing contiguous payload buffer. 1495 */ 1496 static int 1497 nvme_rdma_build_contig_request(struct nvme_rdma_qpair *rqpair, 1498 struct spdk_nvme_rdma_req *rdma_req) 1499 { 1500 struct nvme_request *req = rdma_req->req; 1501 struct nvme_rdma_memory_translation_ctx ctx = { 1502 .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset, 1503 .length = req->payload_size 1504 }; 1505 int rc; 1506 1507 assert(req->payload_size != 0); 1508 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG); 1509 1510 if (spdk_unlikely(req->payload_size > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) { 1511 SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n", 1512 req->payload_size, NVME_RDMA_MAX_KEYED_SGL_LENGTH); 1513 return -1; 1514 } 1515 1516 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1517 if (spdk_unlikely(rc)) { 1518 return -1; 1519 } 1520 1521 req->cmd.dptr.sgl1.keyed.key = ctx.rkey; 1522 1523 /* The first element of this SGL is pointing at an 1524 * spdk_nvmf_cmd object. For this particular command, 1525 * we only need the first 64 bytes corresponding to 1526 * the NVMe command. */ 1527 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1528 1529 /* The RDMA SGL needs one element describing the NVMe command. */ 1530 rdma_req->send_wr.num_sge = 1; 1531 1532 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1533 req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1534 req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1535 req->cmd.dptr.sgl1.keyed.length = (uint32_t)ctx.length; 1536 req->cmd.dptr.sgl1.address = (uint64_t)ctx.addr; 1537 1538 return 0; 1539 } 1540 1541 /* 1542 * Build SGL describing scattered payload buffer. 1543 */ 1544 static int 1545 nvme_rdma_build_sgl_request(struct nvme_rdma_qpair *rqpair, 1546 struct spdk_nvme_rdma_req *rdma_req) 1547 { 1548 struct nvme_request *req = rdma_req->req; 1549 struct spdk_nvmf_cmd *cmd = &rqpair->cmds[rdma_req->id]; 1550 struct nvme_rdma_memory_translation_ctx ctx; 1551 uint32_t remaining_size; 1552 uint32_t sge_length; 1553 int rc, max_num_sgl, num_sgl_desc; 1554 1555 assert(req->payload_size != 0); 1556 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL); 1557 assert(req->payload.reset_sgl_fn != NULL); 1558 assert(req->payload.next_sge_fn != NULL); 1559 req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset); 1560 1561 max_num_sgl = req->qpair->ctrlr->max_sges; 1562 1563 remaining_size = req->payload_size; 1564 num_sgl_desc = 0; 1565 do { 1566 rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &sge_length); 1567 if (rc) { 1568 return -1; 1569 } 1570 1571 sge_length = spdk_min(remaining_size, sge_length); 1572 1573 if (spdk_unlikely(sge_length > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) { 1574 SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n", 1575 sge_length, NVME_RDMA_MAX_KEYED_SGL_LENGTH); 1576 return -1; 1577 } 1578 ctx.length = sge_length; 1579 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1580 if (spdk_unlikely(rc)) { 1581 return -1; 1582 } 1583 1584 cmd->sgl[num_sgl_desc].keyed.key = ctx.rkey; 1585 cmd->sgl[num_sgl_desc].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1586 cmd->sgl[num_sgl_desc].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1587 cmd->sgl[num_sgl_desc].keyed.length = (uint32_t)ctx.length; 1588 cmd->sgl[num_sgl_desc].address = (uint64_t)ctx.addr; 1589 1590 remaining_size -= ctx.length; 1591 num_sgl_desc++; 1592 } while (remaining_size > 0 && num_sgl_desc < max_num_sgl); 1593 1594 1595 /* Should be impossible if we did our sgl checks properly up the stack, but do a sanity check here. */ 1596 if (remaining_size > 0) { 1597 return -1; 1598 } 1599 1600 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1601 1602 /* The RDMA SGL needs one element describing some portion 1603 * of the spdk_nvmf_cmd structure. */ 1604 rdma_req->send_wr.num_sge = 1; 1605 1606 /* 1607 * If only one SGL descriptor is required, it can be embedded directly in the command 1608 * as a data block descriptor. 1609 */ 1610 if (num_sgl_desc == 1) { 1611 /* The first element of this SGL is pointing at an 1612 * spdk_nvmf_cmd object. For this particular command, 1613 * we only need the first 64 bytes corresponding to 1614 * the NVMe command. */ 1615 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1616 1617 req->cmd.dptr.sgl1.keyed.type = cmd->sgl[0].keyed.type; 1618 req->cmd.dptr.sgl1.keyed.subtype = cmd->sgl[0].keyed.subtype; 1619 req->cmd.dptr.sgl1.keyed.length = cmd->sgl[0].keyed.length; 1620 req->cmd.dptr.sgl1.keyed.key = cmd->sgl[0].keyed.key; 1621 req->cmd.dptr.sgl1.address = cmd->sgl[0].address; 1622 } else { 1623 /* 1624 * Otherwise, The SGL descriptor embedded in the command must point to the list of 1625 * SGL descriptors used to describe the operation. In that case it is a last segment descriptor. 1626 */ 1627 uint32_t descriptors_size = sizeof(struct spdk_nvme_sgl_descriptor) * num_sgl_desc; 1628 1629 if (spdk_unlikely(descriptors_size > rqpair->qpair.ctrlr->ioccsz_bytes)) { 1630 SPDK_ERRLOG("Size of SGL descriptors (%u) exceeds ICD (%u)\n", 1631 descriptors_size, rqpair->qpair.ctrlr->ioccsz_bytes); 1632 return -1; 1633 } 1634 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd) + descriptors_size; 1635 1636 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT; 1637 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1638 req->cmd.dptr.sgl1.unkeyed.length = descriptors_size; 1639 req->cmd.dptr.sgl1.address = (uint64_t)0; 1640 } 1641 1642 return 0; 1643 } 1644 1645 /* 1646 * Build inline SGL describing sgl payload buffer. 1647 */ 1648 static int 1649 nvme_rdma_build_sgl_inline_request(struct nvme_rdma_qpair *rqpair, 1650 struct spdk_nvme_rdma_req *rdma_req) 1651 { 1652 struct nvme_request *req = rdma_req->req; 1653 struct nvme_rdma_memory_translation_ctx ctx; 1654 uint32_t length; 1655 int rc; 1656 1657 assert(req->payload_size != 0); 1658 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL); 1659 assert(req->payload.reset_sgl_fn != NULL); 1660 assert(req->payload.next_sge_fn != NULL); 1661 req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset); 1662 1663 rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &length); 1664 if (rc) { 1665 return -1; 1666 } 1667 1668 if (length < req->payload_size) { 1669 SPDK_DEBUGLOG(nvme, "Inline SGL request split so sending separately.\n"); 1670 return nvme_rdma_build_sgl_request(rqpair, rdma_req); 1671 } 1672 1673 if (length > req->payload_size) { 1674 length = req->payload_size; 1675 } 1676 1677 ctx.length = length; 1678 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1679 if (spdk_unlikely(rc)) { 1680 return -1; 1681 } 1682 1683 rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr; 1684 rdma_req->send_sgl[1].length = (uint32_t)ctx.length; 1685 rdma_req->send_sgl[1].lkey = ctx.lkey; 1686 1687 rdma_req->send_wr.num_sge = 2; 1688 1689 /* The first element of this SGL is pointing at an 1690 * spdk_nvmf_cmd object. For this particular command, 1691 * we only need the first 64 bytes corresponding to 1692 * the NVMe command. */ 1693 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1694 1695 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1696 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK; 1697 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1698 req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length; 1699 /* Inline only supported for icdoff == 0 currently. This function will 1700 * not get called for controllers with other values. */ 1701 req->cmd.dptr.sgl1.address = (uint64_t)0; 1702 1703 return 0; 1704 } 1705 1706 static int 1707 nvme_rdma_req_init(struct nvme_rdma_qpair *rqpair, struct nvme_request *req, 1708 struct spdk_nvme_rdma_req *rdma_req) 1709 { 1710 struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr; 1711 enum nvme_payload_type payload_type; 1712 bool icd_supported; 1713 int rc; 1714 1715 assert(rdma_req->req == NULL); 1716 rdma_req->req = req; 1717 req->cmd.cid = rdma_req->id; 1718 payload_type = nvme_payload_type(&req->payload); 1719 /* 1720 * Check if icdoff is non zero, to avoid interop conflicts with 1721 * targets with non-zero icdoff. Both SPDK and the Linux kernel 1722 * targets use icdoff = 0. For targets with non-zero icdoff, we 1723 * will currently just not use inline data for now. 1724 */ 1725 icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER 1726 && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0; 1727 1728 if (req->payload_size == 0) { 1729 rc = nvme_rdma_build_null_request(rdma_req); 1730 } else if (payload_type == NVME_PAYLOAD_TYPE_CONTIG) { 1731 if (icd_supported) { 1732 rc = nvme_rdma_build_contig_inline_request(rqpair, rdma_req); 1733 } else { 1734 rc = nvme_rdma_build_contig_request(rqpair, rdma_req); 1735 } 1736 } else if (payload_type == NVME_PAYLOAD_TYPE_SGL) { 1737 if (icd_supported) { 1738 rc = nvme_rdma_build_sgl_inline_request(rqpair, rdma_req); 1739 } else { 1740 rc = nvme_rdma_build_sgl_request(rqpair, rdma_req); 1741 } 1742 } else { 1743 rc = -1; 1744 } 1745 1746 if (rc) { 1747 rdma_req->req = NULL; 1748 return rc; 1749 } 1750 1751 memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd)); 1752 return 0; 1753 } 1754 1755 static struct spdk_nvme_qpair * 1756 nvme_rdma_ctrlr_create_qpair(struct spdk_nvme_ctrlr *ctrlr, 1757 uint16_t qid, uint32_t qsize, 1758 enum spdk_nvme_qprio qprio, 1759 uint32_t num_requests, 1760 bool delay_cmd_submit, 1761 bool async) 1762 { 1763 struct nvme_rdma_qpair *rqpair; 1764 struct spdk_nvme_qpair *qpair; 1765 int rc; 1766 1767 if (qsize < SPDK_NVME_QUEUE_MIN_ENTRIES) { 1768 SPDK_ERRLOG("Failed to create qpair with size %u. Minimum queue size is %d.\n", 1769 qsize, SPDK_NVME_QUEUE_MIN_ENTRIES); 1770 return NULL; 1771 } 1772 1773 rqpair = spdk_zmalloc(sizeof(struct nvme_rdma_qpair), 0, NULL, SPDK_ENV_NUMA_ID_ANY, 1774 SPDK_MALLOC_DMA); 1775 if (!rqpair) { 1776 SPDK_ERRLOG("failed to get create rqpair\n"); 1777 return NULL; 1778 } 1779 1780 /* Set num_entries one less than queue size. According to NVMe 1781 * and NVMe-oF specs we can not submit queue size requests, 1782 * one slot shall always remain empty. 1783 */ 1784 rqpair->num_entries = qsize - 1; 1785 rqpair->delay_cmd_submit = delay_cmd_submit; 1786 rqpair->state = NVME_RDMA_QPAIR_STATE_INVALID; 1787 qpair = &rqpair->qpair; 1788 rc = nvme_qpair_init(qpair, qid, ctrlr, qprio, num_requests, async); 1789 if (rc != 0) { 1790 spdk_free(rqpair); 1791 return NULL; 1792 } 1793 1794 return qpair; 1795 } 1796 1797 static void 1798 nvme_rdma_qpair_destroy(struct nvme_rdma_qpair *rqpair) 1799 { 1800 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 1801 struct nvme_rdma_ctrlr *rctrlr; 1802 struct nvme_rdma_cm_event_entry *entry, *tmp; 1803 1804 spdk_rdma_utils_free_mem_map(&rqpair->mr_map); 1805 1806 if (rqpair->evt) { 1807 rdma_ack_cm_event(rqpair->evt); 1808 rqpair->evt = NULL; 1809 } 1810 1811 /* 1812 * This works because we have the controller lock both in 1813 * this function and in the function where we add new events. 1814 */ 1815 if (qpair->ctrlr != NULL) { 1816 rctrlr = nvme_rdma_ctrlr(qpair->ctrlr); 1817 STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) { 1818 if (entry->evt->id->context == rqpair) { 1819 STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link); 1820 rdma_ack_cm_event(entry->evt); 1821 STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link); 1822 } 1823 } 1824 } 1825 1826 if (rqpair->cm_id) { 1827 if (rqpair->rdma_qp) { 1828 spdk_rdma_utils_put_pd(rqpair->rdma_qp->qp->pd); 1829 spdk_rdma_provider_qp_destroy(rqpair->rdma_qp); 1830 rqpair->rdma_qp = NULL; 1831 } 1832 } 1833 1834 if (rqpair->poller) { 1835 struct nvme_rdma_poll_group *group; 1836 1837 assert(qpair->poll_group); 1838 group = nvme_rdma_poll_group(qpair->poll_group); 1839 1840 nvme_rdma_poll_group_put_poller(group, rqpair->poller); 1841 1842 rqpair->poller = NULL; 1843 rqpair->cq = NULL; 1844 if (rqpair->srq) { 1845 rqpair->srq = NULL; 1846 rqpair->rsps = NULL; 1847 } 1848 } else if (rqpair->cq) { 1849 ibv_destroy_cq(rqpair->cq); 1850 rqpair->cq = NULL; 1851 } 1852 1853 nvme_rdma_free_reqs(rqpair); 1854 nvme_rdma_free_rsps(rqpair->rsps); 1855 rqpair->rsps = NULL; 1856 1857 /* destroy cm_id last so cma device will not be freed before we destroy the cq. */ 1858 if (rqpair->cm_id) { 1859 rdma_destroy_id(rqpair->cm_id); 1860 rqpair->cm_id = NULL; 1861 } 1862 } 1863 1864 static void nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr); 1865 1866 static int 1867 nvme_rdma_qpair_disconnected(struct nvme_rdma_qpair *rqpair, int ret) 1868 { 1869 if (ret) { 1870 SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n"); 1871 goto quiet; 1872 } 1873 1874 if (rqpair->poller == NULL) { 1875 /* If poller is not used, cq is not shared. 1876 * So complete disconnecting qpair immediately. 1877 */ 1878 goto quiet; 1879 } 1880 1881 if (rqpair->rsps == NULL) { 1882 goto quiet; 1883 } 1884 1885 if (rqpair->need_destroy || 1886 (rqpair->current_num_sends != 0 || 1887 (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) { 1888 rqpair->state = NVME_RDMA_QPAIR_STATE_LINGERING; 1889 rqpair->evt_timeout_ticks = (NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US * spdk_get_ticks_hz()) / 1890 SPDK_SEC_TO_USEC + spdk_get_ticks(); 1891 1892 return -EAGAIN; 1893 } 1894 1895 quiet: 1896 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED; 1897 1898 nvme_rdma_qpair_abort_reqs(&rqpair->qpair, rqpair->qpair.abort_dnr); 1899 nvme_rdma_qpair_destroy(rqpair); 1900 nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair); 1901 1902 return 0; 1903 } 1904 1905 static int 1906 nvme_rdma_qpair_wait_until_quiet(struct nvme_rdma_qpair *rqpair) 1907 { 1908 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 1909 struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr; 1910 1911 if (spdk_get_ticks() < rqpair->evt_timeout_ticks && 1912 (rqpair->current_num_sends != 0 || 1913 (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) { 1914 return -EAGAIN; 1915 } 1916 1917 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED; 1918 nvme_rdma_qpair_abort_reqs(qpair, qpair->abort_dnr); 1919 if (!nvme_qpair_is_admin_queue(qpair)) { 1920 nvme_robust_mutex_lock(&ctrlr->ctrlr_lock); 1921 } 1922 nvme_rdma_qpair_destroy(rqpair); 1923 if (!nvme_qpair_is_admin_queue(qpair)) { 1924 nvme_robust_mutex_unlock(&ctrlr->ctrlr_lock); 1925 } 1926 nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair); 1927 1928 return 0; 1929 } 1930 1931 static void 1932 _nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair, 1933 nvme_rdma_cm_event_cb disconnected_qpair_cb) 1934 { 1935 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 1936 int rc; 1937 1938 assert(disconnected_qpair_cb != NULL); 1939 1940 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITING; 1941 1942 if (rqpair->cm_id) { 1943 if (rqpair->rdma_qp) { 1944 rc = spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp); 1945 if ((qpair->ctrlr != NULL) && (rc == 0)) { 1946 rc = nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_DISCONNECTED, 1947 disconnected_qpair_cb); 1948 if (rc == 0) { 1949 return; 1950 } 1951 } 1952 } 1953 } 1954 1955 disconnected_qpair_cb(rqpair, 0); 1956 } 1957 1958 static int 1959 nvme_rdma_ctrlr_disconnect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 1960 { 1961 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 1962 int rc; 1963 1964 switch (rqpair->state) { 1965 case NVME_RDMA_QPAIR_STATE_EXITING: 1966 if (!nvme_qpair_is_admin_queue(qpair)) { 1967 nvme_ctrlr_lock(ctrlr); 1968 } 1969 1970 rc = nvme_rdma_process_event_poll(rqpair); 1971 1972 if (!nvme_qpair_is_admin_queue(qpair)) { 1973 nvme_ctrlr_unlock(ctrlr); 1974 } 1975 break; 1976 1977 case NVME_RDMA_QPAIR_STATE_LINGERING: 1978 rc = nvme_rdma_qpair_wait_until_quiet(rqpair); 1979 break; 1980 case NVME_RDMA_QPAIR_STATE_EXITED: 1981 rc = 0; 1982 break; 1983 1984 default: 1985 assert(false); 1986 rc = -EAGAIN; 1987 break; 1988 } 1989 1990 return rc; 1991 } 1992 1993 static void 1994 nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 1995 { 1996 int rc; 1997 1998 _nvme_rdma_ctrlr_disconnect_qpair(ctrlr, qpair, nvme_rdma_qpair_disconnected); 1999 2000 /* If the async mode is disabled, poll the qpair until it is actually disconnected. 2001 * It is ensured that poll_group_process_completions() calls disconnected_qpair_cb 2002 * for any disconnected qpair. Hence, we do not have to check if the qpair is in 2003 * a poll group or not. 2004 * At the same time, if the qpair is being destroyed, i.e. this function is called by 2005 * spdk_nvme_ctrlr_free_io_qpair then we need to wait until qpair is disconnected, otherwise 2006 * we may leak some resources. 2007 */ 2008 if (qpair->async && !qpair->destroy_in_progress) { 2009 return; 2010 } 2011 2012 while (1) { 2013 rc = nvme_rdma_ctrlr_disconnect_qpair_poll(ctrlr, qpair); 2014 if (rc != -EAGAIN) { 2015 break; 2016 } 2017 } 2018 } 2019 2020 static int 2021 nvme_rdma_stale_conn_disconnected(struct nvme_rdma_qpair *rqpair, int ret) 2022 { 2023 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2024 2025 if (ret) { 2026 SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n"); 2027 } 2028 2029 nvme_rdma_qpair_destroy(rqpair); 2030 2031 qpair->last_transport_failure_reason = qpair->transport_failure_reason; 2032 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_NONE; 2033 2034 rqpair->state = NVME_RDMA_QPAIR_STATE_STALE_CONN; 2035 rqpair->evt_timeout_ticks = (NVME_RDMA_STALE_CONN_RETRY_DELAY_US * spdk_get_ticks_hz()) / 2036 SPDK_SEC_TO_USEC + spdk_get_ticks(); 2037 2038 return 0; 2039 } 2040 2041 static int 2042 nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair) 2043 { 2044 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2045 2046 if (rqpair->stale_conn_retry_count >= NVME_RDMA_STALE_CONN_RETRY_MAX) { 2047 SPDK_ERRLOG("Retry failed %d times, give up stale connection to qpair (cntlid:%u, qid:%u).\n", 2048 NVME_RDMA_STALE_CONN_RETRY_MAX, qpair->ctrlr->cntlid, qpair->id); 2049 return -ESTALE; 2050 } 2051 2052 rqpair->stale_conn_retry_count++; 2053 2054 SPDK_NOTICELOG("%d times, retry stale connection to qpair (cntlid:%u, qid:%u).\n", 2055 rqpair->stale_conn_retry_count, qpair->ctrlr->cntlid, qpair->id); 2056 2057 _nvme_rdma_ctrlr_disconnect_qpair(qpair->ctrlr, qpair, nvme_rdma_stale_conn_disconnected); 2058 2059 return 0; 2060 } 2061 2062 static int 2063 nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 2064 { 2065 struct nvme_rdma_qpair *rqpair; 2066 2067 assert(qpair != NULL); 2068 rqpair = nvme_rdma_qpair(qpair); 2069 2070 if (rqpair->state != NVME_RDMA_QPAIR_STATE_EXITED) { 2071 int rc __attribute__((unused)); 2072 2073 /* qpair was removed from the poll group while the disconnect is not finished. 2074 * Destroy rdma resources forcefully. */ 2075 rc = nvme_rdma_qpair_disconnected(rqpair, 0); 2076 assert(rc == 0); 2077 } 2078 2079 nvme_rdma_qpair_abort_reqs(qpair, qpair->abort_dnr); 2080 nvme_qpair_deinit(qpair); 2081 2082 if (spdk_rdma_utils_put_memory_domain(rqpair->memory_domain) != 0) { 2083 SPDK_ERRLOG("Failed to release memory domain\n"); 2084 assert(0); 2085 } 2086 2087 spdk_free(rqpair); 2088 2089 return 0; 2090 } 2091 2092 static struct spdk_nvme_qpair * 2093 nvme_rdma_ctrlr_create_io_qpair(struct spdk_nvme_ctrlr *ctrlr, uint16_t qid, 2094 const struct spdk_nvme_io_qpair_opts *opts) 2095 { 2096 return nvme_rdma_ctrlr_create_qpair(ctrlr, qid, opts->io_queue_size, opts->qprio, 2097 opts->io_queue_requests, 2098 opts->delay_cmd_submit, 2099 opts->async_mode); 2100 } 2101 2102 static int 2103 nvme_rdma_ctrlr_enable(struct spdk_nvme_ctrlr *ctrlr) 2104 { 2105 /* do nothing here */ 2106 return 0; 2107 } 2108 2109 static int nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr); 2110 2111 /* We have to use the typedef in the function declaration to appease astyle. */ 2112 typedef struct spdk_nvme_ctrlr spdk_nvme_ctrlr_t; 2113 2114 static spdk_nvme_ctrlr_t * 2115 nvme_rdma_ctrlr_construct(const struct spdk_nvme_transport_id *trid, 2116 const struct spdk_nvme_ctrlr_opts *opts, 2117 void *devhandle) 2118 { 2119 struct nvme_rdma_ctrlr *rctrlr; 2120 struct ibv_context **contexts; 2121 struct ibv_device_attr dev_attr; 2122 int i, flag, rc; 2123 2124 rctrlr = spdk_zmalloc(sizeof(struct nvme_rdma_ctrlr), 0, NULL, SPDK_ENV_NUMA_ID_ANY, 2125 SPDK_MALLOC_DMA); 2126 if (rctrlr == NULL) { 2127 SPDK_ERRLOG("could not allocate ctrlr\n"); 2128 return NULL; 2129 } 2130 2131 rctrlr->ctrlr.opts = *opts; 2132 rctrlr->ctrlr.trid = *trid; 2133 2134 if (opts->transport_retry_count > NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT) { 2135 SPDK_NOTICELOG("transport_retry_count exceeds max value %d, use max value\n", 2136 NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT); 2137 rctrlr->ctrlr.opts.transport_retry_count = NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT; 2138 } 2139 2140 if (opts->transport_ack_timeout > NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT) { 2141 SPDK_NOTICELOG("transport_ack_timeout exceeds max value %d, use max value\n", 2142 NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT); 2143 rctrlr->ctrlr.opts.transport_ack_timeout = NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT; 2144 } 2145 2146 contexts = rdma_get_devices(NULL); 2147 if (contexts == NULL) { 2148 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2149 spdk_free(rctrlr); 2150 return NULL; 2151 } 2152 2153 i = 0; 2154 rctrlr->max_sge = NVME_RDMA_MAX_SGL_DESCRIPTORS; 2155 2156 while (contexts[i] != NULL) { 2157 rc = ibv_query_device(contexts[i], &dev_attr); 2158 if (rc < 0) { 2159 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2160 rdma_free_devices(contexts); 2161 spdk_free(rctrlr); 2162 return NULL; 2163 } 2164 rctrlr->max_sge = spdk_min(rctrlr->max_sge, (uint16_t)dev_attr.max_sge); 2165 i++; 2166 } 2167 2168 rdma_free_devices(contexts); 2169 2170 rc = nvme_ctrlr_construct(&rctrlr->ctrlr); 2171 if (rc != 0) { 2172 spdk_free(rctrlr); 2173 return NULL; 2174 } 2175 2176 STAILQ_INIT(&rctrlr->pending_cm_events); 2177 STAILQ_INIT(&rctrlr->free_cm_events); 2178 rctrlr->cm_events = spdk_zmalloc(NVME_RDMA_NUM_CM_EVENTS * sizeof(*rctrlr->cm_events), 0, NULL, 2179 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 2180 if (rctrlr->cm_events == NULL) { 2181 SPDK_ERRLOG("unable to allocate buffers to hold CM events.\n"); 2182 goto destruct_ctrlr; 2183 } 2184 2185 for (i = 0; i < NVME_RDMA_NUM_CM_EVENTS; i++) { 2186 STAILQ_INSERT_TAIL(&rctrlr->free_cm_events, &rctrlr->cm_events[i], link); 2187 } 2188 2189 rctrlr->cm_channel = rdma_create_event_channel(); 2190 if (rctrlr->cm_channel == NULL) { 2191 SPDK_ERRLOG("rdma_create_event_channel() failed\n"); 2192 goto destruct_ctrlr; 2193 } 2194 2195 flag = fcntl(rctrlr->cm_channel->fd, F_GETFL); 2196 if (fcntl(rctrlr->cm_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2197 SPDK_ERRLOG("Cannot set event channel to non blocking\n"); 2198 goto destruct_ctrlr; 2199 } 2200 2201 rctrlr->ctrlr.adminq = nvme_rdma_ctrlr_create_qpair(&rctrlr->ctrlr, 0, 2202 rctrlr->ctrlr.opts.admin_queue_size, 0, 2203 rctrlr->ctrlr.opts.admin_queue_size, false, true); 2204 if (!rctrlr->ctrlr.adminq) { 2205 SPDK_ERRLOG("failed to create admin qpair\n"); 2206 goto destruct_ctrlr; 2207 } 2208 2209 if (nvme_ctrlr_add_process(&rctrlr->ctrlr, 0) != 0) { 2210 SPDK_ERRLOG("nvme_ctrlr_add_process() failed\n"); 2211 goto destruct_ctrlr; 2212 } 2213 2214 SPDK_DEBUGLOG(nvme, "successfully initialized the nvmf ctrlr\n"); 2215 return &rctrlr->ctrlr; 2216 2217 destruct_ctrlr: 2218 nvme_ctrlr_destruct(&rctrlr->ctrlr); 2219 return NULL; 2220 } 2221 2222 static int 2223 nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr) 2224 { 2225 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr); 2226 struct nvme_rdma_cm_event_entry *entry; 2227 2228 if (ctrlr->adminq) { 2229 nvme_rdma_ctrlr_delete_io_qpair(ctrlr, ctrlr->adminq); 2230 } 2231 2232 STAILQ_FOREACH(entry, &rctrlr->pending_cm_events, link) { 2233 rdma_ack_cm_event(entry->evt); 2234 } 2235 2236 STAILQ_INIT(&rctrlr->free_cm_events); 2237 STAILQ_INIT(&rctrlr->pending_cm_events); 2238 spdk_free(rctrlr->cm_events); 2239 2240 if (rctrlr->cm_channel) { 2241 rdma_destroy_event_channel(rctrlr->cm_channel); 2242 rctrlr->cm_channel = NULL; 2243 } 2244 2245 nvme_ctrlr_destruct_finish(ctrlr); 2246 2247 spdk_free(rctrlr); 2248 2249 return 0; 2250 } 2251 2252 static int 2253 nvme_rdma_qpair_submit_request(struct spdk_nvme_qpair *qpair, 2254 struct nvme_request *req) 2255 { 2256 struct nvme_rdma_qpair *rqpair; 2257 struct spdk_nvme_rdma_req *rdma_req; 2258 struct ibv_send_wr *wr; 2259 struct nvme_rdma_poll_group *group; 2260 2261 rqpair = nvme_rdma_qpair(qpair); 2262 assert(rqpair != NULL); 2263 assert(req != NULL); 2264 2265 rdma_req = nvme_rdma_req_get(rqpair); 2266 if (spdk_unlikely(!rdma_req)) { 2267 if (rqpair->poller) { 2268 rqpair->poller->stats.queued_requests++; 2269 } 2270 /* Inform the upper layer to try again later. */ 2271 return -EAGAIN; 2272 } 2273 2274 if (nvme_rdma_req_init(rqpair, req, rdma_req)) { 2275 SPDK_ERRLOG("nvme_rdma_req_init() failed\n"); 2276 nvme_rdma_req_put(rqpair, rdma_req); 2277 return -1; 2278 } 2279 2280 TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link); 2281 2282 if (!rqpair->link_active.tqe_prev && qpair->poll_group) { 2283 group = nvme_rdma_poll_group(qpair->poll_group); 2284 TAILQ_INSERT_TAIL(&group->active_qpairs, rqpair, link_active); 2285 } 2286 rqpair->num_outstanding_reqs++; 2287 2288 assert(rqpair->current_num_sends < rqpair->num_entries); 2289 rqpair->current_num_sends++; 2290 2291 wr = &rdma_req->send_wr; 2292 wr->next = NULL; 2293 nvme_rdma_trace_ibv_sge(wr->sg_list); 2294 2295 spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, wr); 2296 2297 if (!rqpair->delay_cmd_submit) { 2298 return nvme_rdma_qpair_submit_sends(rqpair); 2299 } 2300 2301 return 0; 2302 } 2303 2304 static int 2305 nvme_rdma_qpair_reset(struct spdk_nvme_qpair *qpair) 2306 { 2307 /* Currently, doing nothing here */ 2308 return 0; 2309 } 2310 2311 static void 2312 nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr) 2313 { 2314 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2315 struct spdk_nvme_cpl cpl; 2316 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2317 2318 cpl.sqid = qpair->id; 2319 cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION; 2320 cpl.status.sct = SPDK_NVME_SCT_GENERIC; 2321 cpl.status.dnr = dnr; 2322 2323 /* 2324 * We cannot abort requests at the RDMA layer without 2325 * unregistering them. If we do, we can still get error 2326 * free completions on the shared completion queue. 2327 */ 2328 if (nvme_qpair_get_state(qpair) > NVME_QPAIR_DISCONNECTING && 2329 nvme_qpair_get_state(qpair) != NVME_QPAIR_DESTROYING) { 2330 nvme_ctrlr_disconnect_qpair(qpair); 2331 } 2332 2333 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2334 nvme_rdma_req_complete(rdma_req, &cpl, true); 2335 } 2336 } 2337 2338 static void 2339 nvme_rdma_qpair_check_timeout(struct spdk_nvme_qpair *qpair) 2340 { 2341 uint64_t t02; 2342 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2343 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2344 struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr; 2345 struct spdk_nvme_ctrlr_process *active_proc; 2346 2347 /* Don't check timeouts during controller initialization. */ 2348 if (ctrlr->state != NVME_CTRLR_STATE_READY) { 2349 return; 2350 } 2351 2352 if (nvme_qpair_is_admin_queue(qpair)) { 2353 active_proc = nvme_ctrlr_get_current_process(ctrlr); 2354 } else { 2355 active_proc = qpair->active_proc; 2356 } 2357 2358 /* Only check timeouts if the current process has a timeout callback. */ 2359 if (active_proc == NULL || active_proc->timeout_cb_fn == NULL) { 2360 return; 2361 } 2362 2363 t02 = spdk_get_ticks(); 2364 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2365 assert(rdma_req->req != NULL); 2366 2367 if (nvme_request_check_timeout(rdma_req->req, rdma_req->id, active_proc, t02)) { 2368 /* 2369 * The requests are in order, so as soon as one has not timed out, 2370 * stop iterating. 2371 */ 2372 break; 2373 } 2374 } 2375 } 2376 2377 static inline void 2378 nvme_rdma_request_ready(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req) 2379 { 2380 struct spdk_nvme_rdma_rsp *rdma_rsp = rdma_req->rdma_rsp; 2381 struct ibv_recv_wr *recv_wr = rdma_rsp->recv_wr; 2382 2383 nvme_rdma_req_complete(rdma_req, &rdma_rsp->cpl, true); 2384 2385 assert(rqpair->rsps->current_num_recvs < rqpair->rsps->num_entries); 2386 rqpair->rsps->current_num_recvs++; 2387 2388 recv_wr->next = NULL; 2389 nvme_rdma_trace_ibv_sge(recv_wr->sg_list); 2390 2391 if (!rqpair->srq) { 2392 spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, recv_wr); 2393 } else { 2394 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, recv_wr); 2395 } 2396 } 2397 2398 #define MAX_COMPLETIONS_PER_POLL 128 2399 2400 static void 2401 nvme_rdma_fail_qpair(struct spdk_nvme_qpair *qpair, int failure_reason) 2402 { 2403 if (failure_reason == IBV_WC_RETRY_EXC_ERR) { 2404 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE; 2405 } else if (qpair->transport_failure_reason == SPDK_NVME_QPAIR_FAILURE_NONE) { 2406 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_UNKNOWN; 2407 } 2408 2409 nvme_ctrlr_disconnect_qpair(qpair); 2410 } 2411 2412 static struct nvme_rdma_qpair * 2413 get_rdma_qpair_from_wc(struct nvme_rdma_poll_group *group, struct ibv_wc *wc) 2414 { 2415 struct spdk_nvme_qpair *qpair; 2416 struct nvme_rdma_qpair *rqpair; 2417 2418 STAILQ_FOREACH(qpair, &group->group.connected_qpairs, poll_group_stailq) { 2419 rqpair = nvme_rdma_qpair(qpair); 2420 if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) { 2421 return rqpair; 2422 } 2423 } 2424 2425 STAILQ_FOREACH(qpair, &group->group.disconnected_qpairs, poll_group_stailq) { 2426 rqpair = nvme_rdma_qpair(qpair); 2427 if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) { 2428 return rqpair; 2429 } 2430 } 2431 2432 return NULL; 2433 } 2434 2435 static inline void 2436 nvme_rdma_log_wc_status(struct nvme_rdma_qpair *rqpair, struct ibv_wc *wc) 2437 { 2438 struct nvme_rdma_wr *rdma_wr = (struct nvme_rdma_wr *)wc->wr_id; 2439 2440 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 2441 /* If qpair is in ERR state, we will receive completions for all posted and not completed 2442 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 2443 SPDK_DEBUGLOG(nvme, "WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n", 2444 rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status, 2445 ibv_wc_status_str(wc->status)); 2446 } else { 2447 SPDK_ERRLOG("WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n", 2448 rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status, 2449 ibv_wc_status_str(wc->status)); 2450 } 2451 } 2452 2453 static inline int 2454 nvme_rdma_process_recv_completion(struct nvme_rdma_poller *poller, struct ibv_wc *wc, 2455 struct nvme_rdma_wr *rdma_wr) 2456 { 2457 struct nvme_rdma_qpair *rqpair; 2458 struct spdk_nvme_rdma_req *rdma_req; 2459 struct spdk_nvme_rdma_rsp *rdma_rsp; 2460 2461 rdma_rsp = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_rsp, rdma_wr); 2462 2463 if (poller && poller->srq) { 2464 rqpair = get_rdma_qpair_from_wc(poller->group, wc); 2465 if (spdk_unlikely(!rqpair)) { 2466 /* Since we do not handle the LAST_WQE_REACHED event, we do not know when 2467 * a Receive Queue in a QP, that is associated with an SRQ, is flushed. 2468 * We may get a WC for a already destroyed QP. 2469 * 2470 * However, for the SRQ, this is not any error. Hence, just re-post the 2471 * receive request to the SRQ to reuse for other QPs, and return 0. 2472 */ 2473 spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr); 2474 return 0; 2475 } 2476 } else { 2477 rqpair = rdma_rsp->rqpair; 2478 if (spdk_unlikely(!rqpair)) { 2479 /* TODO: Fix forceful QP destroy when it is not async mode. 2480 * CQ itself did not cause any error. Hence, return 0 for now. 2481 */ 2482 SPDK_WARNLOG("QP might be already destroyed.\n"); 2483 return 0; 2484 } 2485 } 2486 2487 2488 assert(rqpair->rsps->current_num_recvs > 0); 2489 rqpair->rsps->current_num_recvs--; 2490 2491 if (wc->status) { 2492 nvme_rdma_log_wc_status(rqpair, wc); 2493 goto err_wc; 2494 } 2495 2496 SPDK_DEBUGLOG(nvme, "CQ recv completion\n"); 2497 2498 if (wc->byte_len < sizeof(struct spdk_nvme_cpl)) { 2499 SPDK_ERRLOG("recv length %u less than expected response size\n", wc->byte_len); 2500 goto err_wc; 2501 } 2502 rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid]; 2503 rdma_req->completion_flags |= NVME_RDMA_RECV_COMPLETED; 2504 rdma_req->rdma_rsp = rdma_rsp; 2505 2506 if ((rdma_req->completion_flags & NVME_RDMA_SEND_COMPLETED) == 0) { 2507 return 0; 2508 } 2509 2510 rqpair->num_completions++; 2511 2512 nvme_rdma_request_ready(rqpair, rdma_req); 2513 2514 if (!rqpair->delay_cmd_submit) { 2515 if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) { 2516 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 2517 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2518 return -ENXIO; 2519 } 2520 } 2521 2522 return 1; 2523 2524 err_wc: 2525 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2526 if (poller && poller->srq) { 2527 spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr); 2528 } 2529 return -ENXIO; 2530 } 2531 2532 static inline int 2533 nvme_rdma_process_send_completion(struct nvme_rdma_poller *poller, 2534 struct nvme_rdma_qpair *rdma_qpair, 2535 struct ibv_wc *wc, struct nvme_rdma_wr *rdma_wr) 2536 { 2537 struct nvme_rdma_qpair *rqpair; 2538 struct spdk_nvme_rdma_req *rdma_req; 2539 2540 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_req, rdma_wr); 2541 rqpair = rdma_req->req ? nvme_rdma_qpair(rdma_req->req->qpair) : NULL; 2542 if (!rqpair) { 2543 rqpair = rdma_qpair != NULL ? rdma_qpair : get_rdma_qpair_from_wc(poller->group, wc); 2544 } 2545 2546 /* If we are flushing I/O */ 2547 if (wc->status) { 2548 if (!rqpair) { 2549 /* When poll_group is used, several qpairs share the same CQ and it is possible to 2550 * receive a completion with error (e.g. IBV_WC_WR_FLUSH_ERR) for already disconnected qpair 2551 * That happens due to qpair is destroyed while there are submitted but not completed send/receive 2552 * Work Requests */ 2553 assert(poller); 2554 return 0; 2555 } 2556 assert(rqpair->current_num_sends > 0); 2557 rqpair->current_num_sends--; 2558 nvme_rdma_log_wc_status(rqpair, wc); 2559 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2560 if (rdma_req->rdma_rsp && poller && poller->srq) { 2561 spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_req->rdma_rsp->recv_wr); 2562 } 2563 return -ENXIO; 2564 } 2565 2566 /* We do not support Soft Roce anymore. Other than Soft Roce's bug, we should not 2567 * receive a completion without error status after qpair is disconnected/destroyed. 2568 */ 2569 if (spdk_unlikely(rdma_req->req == NULL)) { 2570 /* 2571 * Some infiniband drivers do not guarantee the previous assumption after we 2572 * received a RDMA_CM_EVENT_DEVICE_REMOVAL event. 2573 */ 2574 SPDK_ERRLOG("Received malformed completion: request 0x%"PRIx64" type %d\n", wc->wr_id, 2575 rdma_wr->type); 2576 if (!rqpair || !rqpair->need_destroy) { 2577 assert(0); 2578 } 2579 return -ENXIO; 2580 } 2581 2582 rdma_req->completion_flags |= NVME_RDMA_SEND_COMPLETED; 2583 assert(rqpair->current_num_sends > 0); 2584 rqpair->current_num_sends--; 2585 2586 if ((rdma_req->completion_flags & NVME_RDMA_RECV_COMPLETED) == 0) { 2587 return 0; 2588 } 2589 2590 rqpair->num_completions++; 2591 2592 nvme_rdma_request_ready(rqpair, rdma_req); 2593 2594 if (!rqpair->delay_cmd_submit) { 2595 if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) { 2596 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 2597 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2598 return -ENXIO; 2599 } 2600 } 2601 2602 return 1; 2603 } 2604 2605 static int 2606 nvme_rdma_cq_process_completions(struct ibv_cq *cq, uint32_t batch_size, 2607 struct nvme_rdma_poller *poller, 2608 struct nvme_rdma_qpair *rdma_qpair, 2609 uint64_t *rdma_completions) 2610 { 2611 struct ibv_wc wc[MAX_COMPLETIONS_PER_POLL]; 2612 struct nvme_rdma_wr *rdma_wr; 2613 uint32_t reaped = 0; 2614 int completion_rc = 0; 2615 int rc, _rc, i; 2616 2617 rc = ibv_poll_cq(cq, batch_size, wc); 2618 if (rc < 0) { 2619 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2620 errno, spdk_strerror(errno)); 2621 return -ECANCELED; 2622 } else if (rc == 0) { 2623 return 0; 2624 } 2625 2626 for (i = 0; i < rc; i++) { 2627 rdma_wr = (struct nvme_rdma_wr *)wc[i].wr_id; 2628 switch (rdma_wr->type) { 2629 case RDMA_WR_TYPE_RECV: 2630 _rc = nvme_rdma_process_recv_completion(poller, &wc[i], rdma_wr); 2631 break; 2632 2633 case RDMA_WR_TYPE_SEND: 2634 _rc = nvme_rdma_process_send_completion(poller, rdma_qpair, &wc[i], rdma_wr); 2635 break; 2636 2637 default: 2638 SPDK_ERRLOG("Received an unexpected opcode on the CQ: %d\n", rdma_wr->type); 2639 return -ECANCELED; 2640 } 2641 if (spdk_likely(_rc >= 0)) { 2642 reaped += _rc; 2643 } else { 2644 completion_rc = _rc; 2645 } 2646 } 2647 2648 *rdma_completions += rc; 2649 2650 if (completion_rc) { 2651 return completion_rc; 2652 } 2653 2654 return reaped; 2655 } 2656 2657 static void 2658 dummy_disconnected_qpair_cb(struct spdk_nvme_qpair *qpair, void *poll_group_ctx) 2659 { 2660 2661 } 2662 2663 static int 2664 nvme_rdma_qpair_process_completions(struct spdk_nvme_qpair *qpair, 2665 uint32_t max_completions) 2666 { 2667 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2668 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(qpair->ctrlr); 2669 int rc = 0, batch_size; 2670 struct ibv_cq *cq; 2671 uint64_t rdma_completions = 0; 2672 2673 /* 2674 * This is used during the connection phase. It's possible that we are still reaping error completions 2675 * from other qpairs so we need to call the poll group function. Also, it's more correct since the cq 2676 * is shared. 2677 */ 2678 if (qpair->poll_group != NULL) { 2679 return spdk_nvme_poll_group_process_completions(qpair->poll_group->group, max_completions, 2680 dummy_disconnected_qpair_cb); 2681 } 2682 2683 if (max_completions == 0) { 2684 max_completions = rqpair->num_entries; 2685 } else { 2686 max_completions = spdk_min(max_completions, rqpair->num_entries); 2687 } 2688 2689 switch (nvme_qpair_get_state(qpair)) { 2690 case NVME_QPAIR_CONNECTING: 2691 rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair); 2692 if (rc == 0) { 2693 /* Once the connection is completed, we can submit queued requests */ 2694 nvme_qpair_resubmit_requests(qpair, rqpair->num_entries); 2695 } else if (rc != -EAGAIN) { 2696 SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair); 2697 goto failed; 2698 } else if (rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING) { 2699 return 0; 2700 } 2701 break; 2702 2703 case NVME_QPAIR_DISCONNECTING: 2704 nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair); 2705 return -ENXIO; 2706 2707 default: 2708 if (nvme_qpair_is_admin_queue(qpair)) { 2709 nvme_rdma_poll_events(rctrlr); 2710 } 2711 nvme_rdma_qpair_process_cm_event(rqpair); 2712 break; 2713 } 2714 2715 if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) { 2716 goto failed; 2717 } 2718 2719 cq = rqpair->cq; 2720 2721 rqpair->num_completions = 0; 2722 do { 2723 batch_size = spdk_min((max_completions - rqpair->num_completions), MAX_COMPLETIONS_PER_POLL); 2724 rc = nvme_rdma_cq_process_completions(cq, batch_size, NULL, rqpair, &rdma_completions); 2725 2726 if (rc == 0) { 2727 break; 2728 /* Handle the case where we fail to poll the cq. */ 2729 } else if (rc == -ECANCELED) { 2730 goto failed; 2731 } else if (rc == -ENXIO) { 2732 return rc; 2733 } 2734 } while (rqpair->num_completions < max_completions); 2735 2736 if (spdk_unlikely(nvme_rdma_qpair_submit_sends(rqpair) || 2737 nvme_rdma_qpair_submit_recvs(rqpair))) { 2738 goto failed; 2739 } 2740 2741 if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) { 2742 nvme_rdma_qpair_check_timeout(qpair); 2743 } 2744 2745 return rqpair->num_completions; 2746 2747 failed: 2748 nvme_rdma_fail_qpair(qpair, 0); 2749 return -ENXIO; 2750 } 2751 2752 static uint32_t 2753 nvme_rdma_ctrlr_get_max_xfer_size(struct spdk_nvme_ctrlr *ctrlr) 2754 { 2755 /* max_mr_size by ibv_query_device indicates the largest value that we can 2756 * set for a registered memory region. It is independent from the actual 2757 * I/O size and is very likely to be larger than 2 MiB which is the 2758 * granularity we currently register memory regions. Hence return 2759 * UINT32_MAX here and let the generic layer use the controller data to 2760 * moderate this value. 2761 */ 2762 return UINT32_MAX; 2763 } 2764 2765 static uint16_t 2766 nvme_rdma_ctrlr_get_max_sges(struct spdk_nvme_ctrlr *ctrlr) 2767 { 2768 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr); 2769 uint32_t max_sge = rctrlr->max_sge; 2770 uint32_t max_in_capsule_sge = (ctrlr->cdata.nvmf_specific.ioccsz * 16 - 2771 sizeof(struct spdk_nvme_cmd)) / 2772 sizeof(struct spdk_nvme_sgl_descriptor); 2773 2774 /* Max SGE is limited by capsule size */ 2775 max_sge = spdk_min(max_sge, max_in_capsule_sge); 2776 /* Max SGE may be limited by MSDBD */ 2777 if (ctrlr->cdata.nvmf_specific.msdbd != 0) { 2778 max_sge = spdk_min(max_sge, ctrlr->cdata.nvmf_specific.msdbd); 2779 } 2780 2781 /* Max SGE can't be less than 1 */ 2782 max_sge = spdk_max(1, max_sge); 2783 return max_sge; 2784 } 2785 2786 static int 2787 nvme_rdma_qpair_iterate_requests(struct spdk_nvme_qpair *qpair, 2788 int (*iter_fn)(struct nvme_request *req, void *arg), 2789 void *arg) 2790 { 2791 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2792 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2793 int rc; 2794 2795 assert(iter_fn != NULL); 2796 2797 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2798 assert(rdma_req->req != NULL); 2799 2800 rc = iter_fn(rdma_req->req, arg); 2801 if (rc != 0) { 2802 return rc; 2803 } 2804 } 2805 2806 return 0; 2807 } 2808 2809 static int 2810 nvme_rdma_qpair_authenticate(struct spdk_nvme_qpair *qpair) 2811 { 2812 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2813 int rc; 2814 2815 /* If the qpair is still connecting, it'll be forced to authenticate later on */ 2816 if (rqpair->state < NVME_RDMA_QPAIR_STATE_RUNNING) { 2817 return 0; 2818 } else if (rqpair->state != NVME_RDMA_QPAIR_STATE_RUNNING) { 2819 return -ENOTCONN; 2820 } 2821 2822 rc = nvme_fabric_qpair_authenticate_async(qpair); 2823 if (rc == 0) { 2824 nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTING); 2825 rqpair->state = NVME_RDMA_QPAIR_STATE_AUTHENTICATING; 2826 } 2827 2828 return rc; 2829 } 2830 2831 static void 2832 nvme_rdma_admin_qpair_abort_aers(struct spdk_nvme_qpair *qpair) 2833 { 2834 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2835 struct spdk_nvme_cpl cpl; 2836 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2837 2838 cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION; 2839 cpl.status.sct = SPDK_NVME_SCT_GENERIC; 2840 2841 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2842 assert(rdma_req->req != NULL); 2843 2844 if (rdma_req->req->cmd.opc != SPDK_NVME_OPC_ASYNC_EVENT_REQUEST) { 2845 continue; 2846 } 2847 2848 nvme_rdma_req_complete(rdma_req, &cpl, false); 2849 } 2850 } 2851 2852 static void 2853 nvme_rdma_poller_destroy(struct nvme_rdma_poller *poller) 2854 { 2855 if (poller->cq) { 2856 ibv_destroy_cq(poller->cq); 2857 } 2858 if (poller->rsps) { 2859 nvme_rdma_free_rsps(poller->rsps); 2860 } 2861 if (poller->srq) { 2862 spdk_rdma_provider_srq_destroy(poller->srq); 2863 } 2864 if (poller->mr_map) { 2865 spdk_rdma_utils_free_mem_map(&poller->mr_map); 2866 } 2867 if (poller->pd) { 2868 spdk_rdma_utils_put_pd(poller->pd); 2869 } 2870 free(poller); 2871 } 2872 2873 static struct nvme_rdma_poller * 2874 nvme_rdma_poller_create(struct nvme_rdma_poll_group *group, struct ibv_context *ctx) 2875 { 2876 struct nvme_rdma_poller *poller; 2877 struct ibv_device_attr dev_attr; 2878 struct spdk_rdma_provider_srq_init_attr srq_init_attr = {}; 2879 struct nvme_rdma_rsp_opts opts; 2880 int num_cqe, max_num_cqe; 2881 int rc; 2882 2883 poller = calloc(1, sizeof(*poller)); 2884 if (poller == NULL) { 2885 SPDK_ERRLOG("Unable to allocate poller.\n"); 2886 return NULL; 2887 } 2888 2889 poller->group = group; 2890 poller->device = ctx; 2891 2892 if (g_spdk_nvme_transport_opts.rdma_srq_size != 0) { 2893 rc = ibv_query_device(ctx, &dev_attr); 2894 if (rc) { 2895 SPDK_ERRLOG("Unable to query RDMA device.\n"); 2896 goto fail; 2897 } 2898 2899 poller->pd = spdk_rdma_utils_get_pd(ctx); 2900 if (poller->pd == NULL) { 2901 SPDK_ERRLOG("Unable to get PD.\n"); 2902 goto fail; 2903 } 2904 2905 poller->mr_map = spdk_rdma_utils_create_mem_map(poller->pd, &g_nvme_hooks, 2906 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE); 2907 if (poller->mr_map == NULL) { 2908 SPDK_ERRLOG("Unable to create memory map.\n"); 2909 goto fail; 2910 } 2911 2912 srq_init_attr.stats = &poller->stats.rdma_stats.recv; 2913 srq_init_attr.pd = poller->pd; 2914 srq_init_attr.srq_init_attr.attr.max_wr = spdk_min((uint32_t)dev_attr.max_srq_wr, 2915 g_spdk_nvme_transport_opts.rdma_srq_size); 2916 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(dev_attr.max_sge, 2917 NVME_RDMA_DEFAULT_RX_SGE); 2918 2919 poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr); 2920 if (poller->srq == NULL) { 2921 SPDK_ERRLOG("Unable to create SRQ.\n"); 2922 goto fail; 2923 } 2924 2925 opts.num_entries = g_spdk_nvme_transport_opts.rdma_srq_size; 2926 opts.rqpair = NULL; 2927 opts.srq = poller->srq; 2928 opts.mr_map = poller->mr_map; 2929 2930 poller->rsps = nvme_rdma_create_rsps(&opts); 2931 if (poller->rsps == NULL) { 2932 SPDK_ERRLOG("Unable to create poller RDMA responses.\n"); 2933 goto fail; 2934 } 2935 2936 rc = nvme_rdma_poller_submit_recvs(poller); 2937 if (rc) { 2938 SPDK_ERRLOG("Unable to submit poller RDMA responses.\n"); 2939 goto fail; 2940 } 2941 2942 /* 2943 * When using an srq, fix the size of the completion queue at startup. 2944 * The initiator sends only send and recv WRs. Hence, the multiplier is 2. 2945 * (The target sends also data WRs. Hence, the multiplier is 3.) 2946 */ 2947 num_cqe = g_spdk_nvme_transport_opts.rdma_srq_size * 2; 2948 } else { 2949 num_cqe = DEFAULT_NVME_RDMA_CQ_SIZE; 2950 } 2951 2952 max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size; 2953 if (max_num_cqe != 0 && num_cqe > max_num_cqe) { 2954 num_cqe = max_num_cqe; 2955 } 2956 2957 poller->cq = ibv_create_cq(poller->device, num_cqe, group, NULL, 0); 2958 2959 if (poller->cq == NULL) { 2960 SPDK_ERRLOG("Unable to create CQ, errno %d.\n", errno); 2961 goto fail; 2962 } 2963 2964 STAILQ_INSERT_HEAD(&group->pollers, poller, link); 2965 group->num_pollers++; 2966 poller->current_num_wc = num_cqe; 2967 poller->required_num_wc = 0; 2968 return poller; 2969 2970 fail: 2971 nvme_rdma_poller_destroy(poller); 2972 return NULL; 2973 } 2974 2975 static void 2976 nvme_rdma_poll_group_free_pollers(struct nvme_rdma_poll_group *group) 2977 { 2978 struct nvme_rdma_poller *poller, *tmp_poller; 2979 2980 STAILQ_FOREACH_SAFE(poller, &group->pollers, link, tmp_poller) { 2981 assert(poller->refcnt == 0); 2982 if (poller->refcnt) { 2983 SPDK_WARNLOG("Destroying poller with non-zero ref count: poller %p, refcnt %d\n", 2984 poller, poller->refcnt); 2985 } 2986 2987 STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link); 2988 nvme_rdma_poller_destroy(poller); 2989 } 2990 } 2991 2992 static struct nvme_rdma_poller * 2993 nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, struct ibv_context *device) 2994 { 2995 struct nvme_rdma_poller *poller = NULL; 2996 2997 STAILQ_FOREACH(poller, &group->pollers, link) { 2998 if (poller->device == device) { 2999 break; 3000 } 3001 } 3002 3003 if (!poller) { 3004 poller = nvme_rdma_poller_create(group, device); 3005 if (!poller) { 3006 SPDK_ERRLOG("Failed to create a poller for device %p\n", device); 3007 return NULL; 3008 } 3009 } 3010 3011 poller->refcnt++; 3012 return poller; 3013 } 3014 3015 static void 3016 nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, struct nvme_rdma_poller *poller) 3017 { 3018 assert(poller->refcnt > 0); 3019 if (--poller->refcnt == 0) { 3020 STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link); 3021 group->num_pollers--; 3022 nvme_rdma_poller_destroy(poller); 3023 } 3024 } 3025 3026 static struct spdk_nvme_transport_poll_group * 3027 nvme_rdma_poll_group_create(void) 3028 { 3029 struct nvme_rdma_poll_group *group; 3030 3031 group = calloc(1, sizeof(*group)); 3032 if (group == NULL) { 3033 SPDK_ERRLOG("Unable to allocate poll group.\n"); 3034 return NULL; 3035 } 3036 3037 STAILQ_INIT(&group->pollers); 3038 TAILQ_INIT(&group->connecting_qpairs); 3039 TAILQ_INIT(&group->active_qpairs); 3040 return &group->group; 3041 } 3042 3043 static int 3044 nvme_rdma_poll_group_connect_qpair(struct spdk_nvme_qpair *qpair) 3045 { 3046 return 0; 3047 } 3048 3049 static int 3050 nvme_rdma_poll_group_disconnect_qpair(struct spdk_nvme_qpair *qpair) 3051 { 3052 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3053 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group); 3054 3055 if (rqpair->link_connecting.tqe_prev) { 3056 TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting); 3057 /* We use prev pointer to check if qpair is in connecting list or not . 3058 * TAILQ_REMOVE doesn't do it. So, we do it manually. 3059 */ 3060 rqpair->link_connecting.tqe_prev = NULL; 3061 } 3062 3063 return 0; 3064 } 3065 3066 static int 3067 nvme_rdma_poll_group_add(struct spdk_nvme_transport_poll_group *tgroup, 3068 struct spdk_nvme_qpair *qpair) 3069 { 3070 return 0; 3071 } 3072 3073 static int 3074 nvme_rdma_poll_group_remove(struct spdk_nvme_transport_poll_group *tgroup, 3075 struct spdk_nvme_qpair *qpair) 3076 { 3077 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3078 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group); 3079 3080 if (rqpair->link_active.tqe_prev) { 3081 TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active); 3082 rqpair->link_active.tqe_prev = NULL; 3083 } 3084 3085 return 0; 3086 } 3087 3088 static inline void 3089 nvme_rdma_qpair_process_submits(struct nvme_rdma_poll_group *group, 3090 struct nvme_rdma_qpair *rqpair) 3091 { 3092 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 3093 3094 assert(rqpair->link_active.tqe_prev != NULL); 3095 3096 if (spdk_unlikely(rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING || 3097 rqpair->state >= NVME_RDMA_QPAIR_STATE_EXITING)) { 3098 return; 3099 } 3100 3101 if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) { 3102 nvme_rdma_qpair_check_timeout(qpair); 3103 } 3104 3105 nvme_rdma_qpair_submit_sends(rqpair); 3106 if (!rqpair->srq) { 3107 nvme_rdma_qpair_submit_recvs(rqpair); 3108 } 3109 if (rqpair->num_completions > 0) { 3110 nvme_qpair_resubmit_requests(qpair, rqpair->num_completions); 3111 rqpair->num_completions = 0; 3112 } 3113 3114 if (rqpair->num_outstanding_reqs == 0 && STAILQ_EMPTY(&qpair->queued_req)) { 3115 TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active); 3116 /* We use prev pointer to check if qpair is in active list or not. 3117 * TAILQ_REMOVE doesn't do it. So, we do it manually. 3118 */ 3119 rqpair->link_active.tqe_prev = NULL; 3120 } 3121 } 3122 3123 static int64_t 3124 nvme_rdma_poll_group_process_completions(struct spdk_nvme_transport_poll_group *tgroup, 3125 uint32_t completions_per_qpair, spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb) 3126 { 3127 struct spdk_nvme_qpair *qpair, *tmp_qpair; 3128 struct nvme_rdma_qpair *rqpair, *tmp_rqpair; 3129 struct nvme_rdma_poll_group *group; 3130 struct nvme_rdma_poller *poller; 3131 int batch_size, rc, rc2 = 0; 3132 int64_t total_completions = 0; 3133 uint64_t completions_allowed = 0; 3134 uint64_t completions_per_poller = 0; 3135 uint64_t poller_completions = 0; 3136 uint64_t rdma_completions; 3137 3138 if (completions_per_qpair == 0) { 3139 completions_per_qpair = MAX_COMPLETIONS_PER_POLL; 3140 } 3141 3142 group = nvme_rdma_poll_group(tgroup); 3143 3144 STAILQ_FOREACH_SAFE(qpair, &tgroup->disconnected_qpairs, poll_group_stailq, tmp_qpair) { 3145 rc = nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair); 3146 if (rc == 0) { 3147 disconnected_qpair_cb(qpair, tgroup->group->ctx); 3148 } 3149 } 3150 3151 TAILQ_FOREACH_SAFE(rqpair, &group->connecting_qpairs, link_connecting, tmp_rqpair) { 3152 qpair = &rqpair->qpair; 3153 3154 rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair); 3155 if (rc == 0 || rc != -EAGAIN) { 3156 TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting); 3157 /* We use prev pointer to check if qpair is in connecting list or not. 3158 * TAILQ_REMOVE does not do it. So, we do it manually. 3159 */ 3160 rqpair->link_connecting.tqe_prev = NULL; 3161 3162 if (rc == 0) { 3163 /* Once the connection is completed, we can submit queued requests */ 3164 nvme_qpair_resubmit_requests(qpair, rqpair->num_entries); 3165 } else if (rc != -EAGAIN) { 3166 SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair); 3167 nvme_rdma_fail_qpair(qpair, 0); 3168 } 3169 } 3170 } 3171 3172 STAILQ_FOREACH_SAFE(qpair, &tgroup->connected_qpairs, poll_group_stailq, tmp_qpair) { 3173 rqpair = nvme_rdma_qpair(qpair); 3174 3175 if (spdk_likely(nvme_qpair_get_state(qpair) != NVME_QPAIR_CONNECTING)) { 3176 nvme_rdma_qpair_process_cm_event(rqpair); 3177 } 3178 3179 if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) { 3180 rc2 = -ENXIO; 3181 nvme_rdma_fail_qpair(qpair, 0); 3182 } 3183 } 3184 3185 completions_allowed = completions_per_qpair * tgroup->num_connected_qpairs; 3186 if (group->num_pollers) { 3187 completions_per_poller = spdk_max(completions_allowed / group->num_pollers, 1); 3188 } 3189 3190 STAILQ_FOREACH(poller, &group->pollers, link) { 3191 poller_completions = 0; 3192 rdma_completions = 0; 3193 do { 3194 poller->stats.polls++; 3195 batch_size = spdk_min((completions_per_poller - poller_completions), MAX_COMPLETIONS_PER_POLL); 3196 rc = nvme_rdma_cq_process_completions(poller->cq, batch_size, poller, NULL, &rdma_completions); 3197 if (rc <= 0) { 3198 if (rc == -ECANCELED) { 3199 return -EIO; 3200 } else if (rc == 0) { 3201 poller->stats.idle_polls++; 3202 } 3203 break; 3204 } 3205 3206 poller_completions += rc; 3207 } while (poller_completions < completions_per_poller); 3208 total_completions += poller_completions; 3209 poller->stats.completions += rdma_completions; 3210 if (poller->srq) { 3211 nvme_rdma_poller_submit_recvs(poller); 3212 } 3213 } 3214 3215 TAILQ_FOREACH_SAFE(rqpair, &group->active_qpairs, link_active, tmp_rqpair) { 3216 nvme_rdma_qpair_process_submits(group, rqpair); 3217 } 3218 3219 return rc2 != 0 ? rc2 : total_completions; 3220 } 3221 3222 /* 3223 * Handle disconnected qpairs when interrupt support gets added. 3224 */ 3225 static void 3226 nvme_rdma_poll_group_check_disconnected_qpairs(struct spdk_nvme_transport_poll_group *tgroup, 3227 spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb) 3228 { 3229 } 3230 3231 static int 3232 nvme_rdma_poll_group_destroy(struct spdk_nvme_transport_poll_group *tgroup) 3233 { 3234 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(tgroup); 3235 3236 if (!STAILQ_EMPTY(&tgroup->connected_qpairs) || !STAILQ_EMPTY(&tgroup->disconnected_qpairs)) { 3237 return -EBUSY; 3238 } 3239 3240 nvme_rdma_poll_group_free_pollers(group); 3241 free(group); 3242 3243 return 0; 3244 } 3245 3246 static int 3247 nvme_rdma_poll_group_get_stats(struct spdk_nvme_transport_poll_group *tgroup, 3248 struct spdk_nvme_transport_poll_group_stat **_stats) 3249 { 3250 struct nvme_rdma_poll_group *group; 3251 struct spdk_nvme_transport_poll_group_stat *stats; 3252 struct spdk_nvme_rdma_device_stat *device_stat; 3253 struct nvme_rdma_poller *poller; 3254 uint32_t i = 0; 3255 3256 if (tgroup == NULL || _stats == NULL) { 3257 SPDK_ERRLOG("Invalid stats or group pointer\n"); 3258 return -EINVAL; 3259 } 3260 3261 group = nvme_rdma_poll_group(tgroup); 3262 stats = calloc(1, sizeof(*stats)); 3263 if (!stats) { 3264 SPDK_ERRLOG("Can't allocate memory for RDMA stats\n"); 3265 return -ENOMEM; 3266 } 3267 stats->trtype = SPDK_NVME_TRANSPORT_RDMA; 3268 stats->rdma.num_devices = group->num_pollers; 3269 3270 if (stats->rdma.num_devices == 0) { 3271 *_stats = stats; 3272 return 0; 3273 } 3274 3275 stats->rdma.device_stats = calloc(stats->rdma.num_devices, sizeof(*stats->rdma.device_stats)); 3276 if (!stats->rdma.device_stats) { 3277 SPDK_ERRLOG("Can't allocate memory for RDMA device stats\n"); 3278 free(stats); 3279 return -ENOMEM; 3280 } 3281 3282 STAILQ_FOREACH(poller, &group->pollers, link) { 3283 device_stat = &stats->rdma.device_stats[i]; 3284 device_stat->name = poller->device->device->name; 3285 device_stat->polls = poller->stats.polls; 3286 device_stat->idle_polls = poller->stats.idle_polls; 3287 device_stat->completions = poller->stats.completions; 3288 device_stat->queued_requests = poller->stats.queued_requests; 3289 device_stat->total_send_wrs = poller->stats.rdma_stats.send.num_submitted_wrs; 3290 device_stat->send_doorbell_updates = poller->stats.rdma_stats.send.doorbell_updates; 3291 device_stat->total_recv_wrs = poller->stats.rdma_stats.recv.num_submitted_wrs; 3292 device_stat->recv_doorbell_updates = poller->stats.rdma_stats.recv.doorbell_updates; 3293 i++; 3294 } 3295 3296 *_stats = stats; 3297 3298 return 0; 3299 } 3300 3301 static void 3302 nvme_rdma_poll_group_free_stats(struct spdk_nvme_transport_poll_group *tgroup, 3303 struct spdk_nvme_transport_poll_group_stat *stats) 3304 { 3305 if (stats) { 3306 free(stats->rdma.device_stats); 3307 } 3308 free(stats); 3309 } 3310 3311 static int 3312 nvme_rdma_ctrlr_get_memory_domains(const struct spdk_nvme_ctrlr *ctrlr, 3313 struct spdk_memory_domain **domains, int array_size) 3314 { 3315 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(ctrlr->adminq); 3316 3317 if (domains && array_size > 0) { 3318 domains[0] = rqpair->memory_domain; 3319 } 3320 3321 return 1; 3322 } 3323 3324 void 3325 spdk_nvme_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3326 { 3327 g_nvme_hooks = *hooks; 3328 } 3329 3330 const struct spdk_nvme_transport_ops rdma_ops = { 3331 .name = "RDMA", 3332 .type = SPDK_NVME_TRANSPORT_RDMA, 3333 .ctrlr_construct = nvme_rdma_ctrlr_construct, 3334 .ctrlr_scan = nvme_fabric_ctrlr_scan, 3335 .ctrlr_destruct = nvme_rdma_ctrlr_destruct, 3336 .ctrlr_enable = nvme_rdma_ctrlr_enable, 3337 3338 .ctrlr_set_reg_4 = nvme_fabric_ctrlr_set_reg_4, 3339 .ctrlr_set_reg_8 = nvme_fabric_ctrlr_set_reg_8, 3340 .ctrlr_get_reg_4 = nvme_fabric_ctrlr_get_reg_4, 3341 .ctrlr_get_reg_8 = nvme_fabric_ctrlr_get_reg_8, 3342 .ctrlr_set_reg_4_async = nvme_fabric_ctrlr_set_reg_4_async, 3343 .ctrlr_set_reg_8_async = nvme_fabric_ctrlr_set_reg_8_async, 3344 .ctrlr_get_reg_4_async = nvme_fabric_ctrlr_get_reg_4_async, 3345 .ctrlr_get_reg_8_async = nvme_fabric_ctrlr_get_reg_8_async, 3346 3347 .ctrlr_get_max_xfer_size = nvme_rdma_ctrlr_get_max_xfer_size, 3348 .ctrlr_get_max_sges = nvme_rdma_ctrlr_get_max_sges, 3349 3350 .ctrlr_create_io_qpair = nvme_rdma_ctrlr_create_io_qpair, 3351 .ctrlr_delete_io_qpair = nvme_rdma_ctrlr_delete_io_qpair, 3352 .ctrlr_connect_qpair = nvme_rdma_ctrlr_connect_qpair, 3353 .ctrlr_disconnect_qpair = nvme_rdma_ctrlr_disconnect_qpair, 3354 3355 .ctrlr_get_memory_domains = nvme_rdma_ctrlr_get_memory_domains, 3356 3357 .qpair_abort_reqs = nvme_rdma_qpair_abort_reqs, 3358 .qpair_reset = nvme_rdma_qpair_reset, 3359 .qpair_submit_request = nvme_rdma_qpair_submit_request, 3360 .qpair_process_completions = nvme_rdma_qpair_process_completions, 3361 .qpair_iterate_requests = nvme_rdma_qpair_iterate_requests, 3362 .qpair_authenticate = nvme_rdma_qpair_authenticate, 3363 .admin_qpair_abort_aers = nvme_rdma_admin_qpair_abort_aers, 3364 3365 .poll_group_create = nvme_rdma_poll_group_create, 3366 .poll_group_connect_qpair = nvme_rdma_poll_group_connect_qpair, 3367 .poll_group_disconnect_qpair = nvme_rdma_poll_group_disconnect_qpair, 3368 .poll_group_add = nvme_rdma_poll_group_add, 3369 .poll_group_remove = nvme_rdma_poll_group_remove, 3370 .poll_group_process_completions = nvme_rdma_poll_group_process_completions, 3371 .poll_group_check_disconnected_qpairs = nvme_rdma_poll_group_check_disconnected_qpairs, 3372 .poll_group_destroy = nvme_rdma_poll_group_destroy, 3373 .poll_group_get_stats = nvme_rdma_poll_group_get_stats, 3374 .poll_group_free_stats = nvme_rdma_poll_group_free_stats, 3375 }; 3376 3377 SPDK_NVME_TRANSPORT_REGISTER(rdma, &rdma_ops); 3378