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 uint16_t num_entries; 202 203 bool delay_cmd_submit; 204 205 uint32_t num_completions; 206 uint32_t num_outstanding_reqs; 207 208 struct nvme_rdma_rsps *rsps; 209 210 /* 211 * Array of num_entries NVMe commands registered as RDMA message buffers. 212 * Indexed by rdma_req->id. 213 */ 214 struct spdk_nvmf_cmd *cmds; 215 216 struct spdk_rdma_utils_mem_map *mr_map; 217 218 TAILQ_HEAD(, spdk_nvme_rdma_req) free_reqs; 219 TAILQ_HEAD(, spdk_nvme_rdma_req) outstanding_reqs; 220 221 /* Count of outstanding send objects */ 222 uint16_t current_num_sends; 223 224 TAILQ_ENTRY(nvme_rdma_qpair) link_active; 225 226 /* Placed at the end of the struct since it is not used frequently */ 227 struct rdma_cm_event *evt; 228 struct nvme_rdma_poller *poller; 229 230 uint64_t evt_timeout_ticks; 231 nvme_rdma_cm_event_cb evt_cb; 232 enum rdma_cm_event_type expected_evt_type; 233 234 enum nvme_rdma_qpair_state state; 235 236 bool in_connect_poll; 237 238 uint8_t stale_conn_retry_count; 239 bool need_destroy; 240 TAILQ_ENTRY(nvme_rdma_qpair) link_connecting; 241 }; 242 243 enum NVME_RDMA_COMPLETION_FLAGS { 244 NVME_RDMA_SEND_COMPLETED = 1u << 0, 245 NVME_RDMA_RECV_COMPLETED = 1u << 1, 246 }; 247 248 struct spdk_nvme_rdma_req { 249 uint16_t id; 250 uint16_t completion_flags: 2; 251 uint16_t in_progress_accel: 1; 252 uint16_t reserved: 13; 253 /* if completion of RDMA_RECV received before RDMA_SEND, we will complete nvme request 254 * during processing of RDMA_SEND. To complete the request we must know the response 255 * received in RDMA_RECV, so store it in this field */ 256 struct spdk_nvme_rdma_rsp *rdma_rsp; 257 258 struct nvme_rdma_wr rdma_wr; 259 260 struct ibv_send_wr send_wr; 261 262 struct nvme_request *req; 263 264 struct ibv_sge send_sgl[NVME_RDMA_DEFAULT_TX_SGE]; 265 266 TAILQ_ENTRY(spdk_nvme_rdma_req) link; 267 268 /* Fields below are not used in regular IO path, keep them last */ 269 spdk_memory_domain_data_cpl_cb transfer_cpl_cb; 270 void *transfer_cpl_cb_arg; 271 /* Accel sequence API works with iovec pointer, we need to store result of next_sge callback */ 272 struct iovec iovs[NVME_RDMA_MAX_SGL_DESCRIPTORS]; 273 }; 274 275 struct spdk_nvme_rdma_rsp { 276 struct spdk_nvme_cpl cpl; 277 struct nvme_rdma_qpair *rqpair; 278 struct ibv_recv_wr *recv_wr; 279 struct nvme_rdma_wr rdma_wr; 280 }; 281 282 struct nvme_rdma_memory_translation_ctx { 283 void *addr; 284 size_t length; 285 uint32_t lkey; 286 uint32_t rkey; 287 }; 288 289 static const char *rdma_cm_event_str[] = { 290 "RDMA_CM_EVENT_ADDR_RESOLVED", 291 "RDMA_CM_EVENT_ADDR_ERROR", 292 "RDMA_CM_EVENT_ROUTE_RESOLVED", 293 "RDMA_CM_EVENT_ROUTE_ERROR", 294 "RDMA_CM_EVENT_CONNECT_REQUEST", 295 "RDMA_CM_EVENT_CONNECT_RESPONSE", 296 "RDMA_CM_EVENT_CONNECT_ERROR", 297 "RDMA_CM_EVENT_UNREACHABLE", 298 "RDMA_CM_EVENT_REJECTED", 299 "RDMA_CM_EVENT_ESTABLISHED", 300 "RDMA_CM_EVENT_DISCONNECTED", 301 "RDMA_CM_EVENT_DEVICE_REMOVAL", 302 "RDMA_CM_EVENT_MULTICAST_JOIN", 303 "RDMA_CM_EVENT_MULTICAST_ERROR", 304 "RDMA_CM_EVENT_ADDR_CHANGE", 305 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 306 }; 307 308 static struct nvme_rdma_poller *nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, 309 struct ibv_context *device); 310 static void nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, 311 struct nvme_rdma_poller *poller); 312 313 static int nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, 314 struct spdk_nvme_qpair *qpair); 315 316 static inline int nvme_rdma_memory_domain_transfer_data(struct spdk_memory_domain *dst_domain, 317 void *dst_domain_ctx, 318 struct iovec *dst_iov, uint32_t dst_iovcnt, 319 struct spdk_memory_domain *src_domain, void *src_domain_ctx, 320 struct iovec *src_iov, uint32_t src_iovcnt, 321 struct spdk_memory_domain_translation_result *translation, 322 spdk_memory_domain_data_cpl_cb cpl_cb, void *cpl_cb_arg); 323 324 static inline int _nvme_rdma_qpair_submit_request(struct nvme_rdma_qpair *rqpair, 325 struct spdk_nvme_rdma_req *rdma_req); 326 327 static inline struct nvme_rdma_qpair * 328 nvme_rdma_qpair(struct spdk_nvme_qpair *qpair) 329 { 330 assert(qpair->trtype == SPDK_NVME_TRANSPORT_RDMA); 331 return SPDK_CONTAINEROF(qpair, struct nvme_rdma_qpair, qpair); 332 } 333 334 static inline struct nvme_rdma_poll_group * 335 nvme_rdma_poll_group(struct spdk_nvme_transport_poll_group *group) 336 { 337 return (SPDK_CONTAINEROF(group, struct nvme_rdma_poll_group, group)); 338 } 339 340 static inline struct nvme_rdma_ctrlr * 341 nvme_rdma_ctrlr(struct spdk_nvme_ctrlr *ctrlr) 342 { 343 assert(ctrlr->trid.trtype == SPDK_NVME_TRANSPORT_RDMA); 344 return SPDK_CONTAINEROF(ctrlr, struct nvme_rdma_ctrlr, ctrlr); 345 } 346 347 static inline struct spdk_nvme_rdma_req * 348 nvme_rdma_req_get(struct nvme_rdma_qpair *rqpair) 349 { 350 struct spdk_nvme_rdma_req *rdma_req; 351 352 rdma_req = TAILQ_FIRST(&rqpair->free_reqs); 353 if (spdk_likely(rdma_req)) { 354 TAILQ_REMOVE(&rqpair->free_reqs, rdma_req, link); 355 } 356 357 return rdma_req; 358 } 359 360 static inline void 361 nvme_rdma_req_put(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req) 362 { 363 rdma_req->completion_flags = 0; 364 rdma_req->req = NULL; 365 rdma_req->rdma_rsp = NULL; 366 assert(rdma_req->transfer_cpl_cb == NULL); 367 TAILQ_INSERT_HEAD(&rqpair->free_reqs, rdma_req, link); 368 } 369 370 static inline void 371 nvme_rdma_finish_data_transfer(struct spdk_nvme_rdma_req *rdma_req, int rc) 372 { 373 spdk_memory_domain_data_cpl_cb cb = rdma_req->transfer_cpl_cb; 374 375 SPDK_DEBUGLOG(nvme, "req %p, finish data transfer, rc %d\n", rdma_req, rc); 376 rdma_req->transfer_cpl_cb = NULL; 377 assert(cb); 378 cb(rdma_req->transfer_cpl_cb_arg, rc); 379 } 380 381 static void 382 nvme_rdma_req_complete(struct spdk_nvme_rdma_req *rdma_req, 383 struct spdk_nvme_cpl *rsp, 384 bool print_on_error) 385 { 386 struct nvme_request *req = rdma_req->req; 387 struct nvme_rdma_qpair *rqpair; 388 struct spdk_nvme_qpair *qpair; 389 bool error, print_error; 390 391 assert(req != NULL); 392 393 qpair = req->qpair; 394 rqpair = nvme_rdma_qpair(qpair); 395 396 error = spdk_nvme_cpl_is_error(rsp); 397 print_error = error && print_on_error && !qpair->ctrlr->opts.disable_error_logging; 398 399 if (print_error) { 400 spdk_nvme_qpair_print_command(qpair, &req->cmd); 401 } 402 403 if (print_error || SPDK_DEBUGLOG_FLAG_ENABLED("nvme")) { 404 spdk_nvme_qpair_print_completion(qpair, rsp); 405 } 406 407 assert(rqpair->num_outstanding_reqs > 0); 408 rqpair->num_outstanding_reqs--; 409 410 TAILQ_REMOVE(&rqpair->outstanding_reqs, rdma_req, link); 411 412 nvme_complete_request(req->cb_fn, req->cb_arg, qpair, req, rsp); 413 nvme_rdma_req_put(rqpair, rdma_req); 414 } 415 416 static const char * 417 nvme_rdma_cm_event_str_get(uint32_t event) 418 { 419 if (event < SPDK_COUNTOF(rdma_cm_event_str)) { 420 return rdma_cm_event_str[event]; 421 } else { 422 return "Undefined"; 423 } 424 } 425 426 427 static int 428 nvme_rdma_qpair_process_cm_event(struct nvme_rdma_qpair *rqpair) 429 { 430 struct rdma_cm_event *event = rqpair->evt; 431 struct spdk_nvmf_rdma_accept_private_data *accept_data; 432 int rc = 0; 433 434 if (event) { 435 switch (event->event) { 436 case RDMA_CM_EVENT_ADDR_RESOLVED: 437 case RDMA_CM_EVENT_ADDR_ERROR: 438 case RDMA_CM_EVENT_ROUTE_RESOLVED: 439 case RDMA_CM_EVENT_ROUTE_ERROR: 440 break; 441 case RDMA_CM_EVENT_CONNECT_REQUEST: 442 break; 443 case RDMA_CM_EVENT_CONNECT_ERROR: 444 break; 445 case RDMA_CM_EVENT_UNREACHABLE: 446 case RDMA_CM_EVENT_REJECTED: 447 break; 448 case RDMA_CM_EVENT_CONNECT_RESPONSE: 449 rc = spdk_rdma_provider_qp_complete_connect(rqpair->rdma_qp); 450 /* fall through */ 451 case RDMA_CM_EVENT_ESTABLISHED: 452 accept_data = (struct spdk_nvmf_rdma_accept_private_data *)event->param.conn.private_data; 453 if (accept_data == NULL) { 454 rc = -1; 455 } else { 456 SPDK_DEBUGLOG(nvme, "Requested queue depth %d. Target receive queue depth %d.\n", 457 rqpair->num_entries + 1, accept_data->crqsize); 458 } 459 break; 460 case RDMA_CM_EVENT_DISCONNECTED: 461 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE; 462 break; 463 case RDMA_CM_EVENT_DEVICE_REMOVAL: 464 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL; 465 rqpair->need_destroy = true; 466 break; 467 case RDMA_CM_EVENT_MULTICAST_JOIN: 468 case RDMA_CM_EVENT_MULTICAST_ERROR: 469 break; 470 case RDMA_CM_EVENT_ADDR_CHANGE: 471 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL; 472 break; 473 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 474 break; 475 default: 476 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 477 break; 478 } 479 rqpair->evt = NULL; 480 rdma_ack_cm_event(event); 481 } 482 483 return rc; 484 } 485 486 /* 487 * This function must be called under the nvme controller's lock 488 * because it touches global controller variables. The lock is taken 489 * by the generic transport code before invoking a few of the functions 490 * in this file: nvme_rdma_ctrlr_connect_qpair, nvme_rdma_ctrlr_delete_io_qpair, 491 * and conditionally nvme_rdma_qpair_process_completions when it is calling 492 * completions on the admin qpair. When adding a new call to this function, please 493 * verify that it is in a situation where it falls under the lock. 494 */ 495 static int 496 nvme_rdma_poll_events(struct nvme_rdma_ctrlr *rctrlr) 497 { 498 struct nvme_rdma_cm_event_entry *entry, *tmp; 499 struct nvme_rdma_qpair *event_qpair; 500 struct rdma_cm_event *event; 501 struct rdma_event_channel *channel = rctrlr->cm_channel; 502 503 STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) { 504 event_qpair = entry->evt->id->context; 505 if (event_qpair->evt == NULL) { 506 event_qpair->evt = entry->evt; 507 STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link); 508 STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link); 509 } 510 } 511 512 while (rdma_get_cm_event(channel, &event) == 0) { 513 event_qpair = event->id->context; 514 if (event_qpair->evt == NULL) { 515 event_qpair->evt = event; 516 } else { 517 assert(rctrlr == nvme_rdma_ctrlr(event_qpair->qpair.ctrlr)); 518 entry = STAILQ_FIRST(&rctrlr->free_cm_events); 519 if (entry == NULL) { 520 rdma_ack_cm_event(event); 521 return -ENOMEM; 522 } 523 STAILQ_REMOVE(&rctrlr->free_cm_events, entry, nvme_rdma_cm_event_entry, link); 524 entry->evt = event; 525 STAILQ_INSERT_TAIL(&rctrlr->pending_cm_events, entry, link); 526 } 527 } 528 529 /* rdma_get_cm_event() returns -1 on error. If an error occurs, errno 530 * will be set to indicate the failure reason. So return negated errno here. 531 */ 532 return -errno; 533 } 534 535 static int 536 nvme_rdma_validate_cm_event(enum rdma_cm_event_type expected_evt_type, 537 struct rdma_cm_event *reaped_evt) 538 { 539 int rc = -EBADMSG; 540 541 if (expected_evt_type == reaped_evt->event) { 542 return 0; 543 } 544 545 switch (expected_evt_type) { 546 case RDMA_CM_EVENT_ESTABLISHED: 547 /* 548 * There is an enum ib_cm_rej_reason in the kernel headers that sets 10 as 549 * IB_CM_REJ_STALE_CONN. I can't find the corresponding userspace but we get 550 * the same values here. 551 */ 552 if (reaped_evt->event == RDMA_CM_EVENT_REJECTED && reaped_evt->status == 10) { 553 rc = -ESTALE; 554 } else if (reaped_evt->event == RDMA_CM_EVENT_CONNECT_RESPONSE) { 555 /* 556 * If we are using a qpair which is not created using rdma cm API 557 * then we will receive RDMA_CM_EVENT_CONNECT_RESPONSE instead of 558 * RDMA_CM_EVENT_ESTABLISHED. 559 */ 560 return 0; 561 } 562 break; 563 default: 564 break; 565 } 566 567 SPDK_ERRLOG("Expected %s but received %s (%d) from CM event channel (status = %d)\n", 568 nvme_rdma_cm_event_str_get(expected_evt_type), 569 nvme_rdma_cm_event_str_get(reaped_evt->event), reaped_evt->event, 570 reaped_evt->status); 571 return rc; 572 } 573 574 static int 575 nvme_rdma_process_event_start(struct nvme_rdma_qpair *rqpair, 576 enum rdma_cm_event_type evt, 577 nvme_rdma_cm_event_cb evt_cb) 578 { 579 int rc; 580 581 assert(evt_cb != NULL); 582 583 if (rqpair->evt != NULL) { 584 rc = nvme_rdma_qpair_process_cm_event(rqpair); 585 if (rc) { 586 return rc; 587 } 588 } 589 590 rqpair->expected_evt_type = evt; 591 rqpair->evt_cb = evt_cb; 592 rqpair->evt_timeout_ticks = (g_spdk_nvme_transport_opts.rdma_cm_event_timeout_ms * 1000 * 593 spdk_get_ticks_hz()) / SPDK_SEC_TO_USEC + spdk_get_ticks(); 594 595 return 0; 596 } 597 598 static int 599 nvme_rdma_process_event_poll(struct nvme_rdma_qpair *rqpair) 600 { 601 struct nvme_rdma_ctrlr *rctrlr; 602 int rc = 0, rc2; 603 604 rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr); 605 assert(rctrlr != NULL); 606 607 if (!rqpair->evt && spdk_get_ticks() < rqpair->evt_timeout_ticks) { 608 rc = nvme_rdma_poll_events(rctrlr); 609 if (rc == -EAGAIN || rc == -EWOULDBLOCK) { 610 return rc; 611 } 612 } 613 614 if (rqpair->evt == NULL) { 615 rc = -EADDRNOTAVAIL; 616 goto exit; 617 } 618 619 rc = nvme_rdma_validate_cm_event(rqpair->expected_evt_type, rqpair->evt); 620 621 rc2 = nvme_rdma_qpair_process_cm_event(rqpair); 622 /* bad message takes precedence over the other error codes from processing the event. */ 623 rc = rc == 0 ? rc2 : rc; 624 625 exit: 626 assert(rqpair->evt_cb != NULL); 627 return rqpair->evt_cb(rqpair, rc); 628 } 629 630 static int 631 nvme_rdma_resize_cq(struct nvme_rdma_qpair *rqpair, struct nvme_rdma_poller *poller) 632 { 633 int current_num_wc, required_num_wc; 634 int max_cq_size; 635 636 required_num_wc = poller->required_num_wc + WC_PER_QPAIR(rqpair->num_entries); 637 current_num_wc = poller->current_num_wc; 638 if (current_num_wc < required_num_wc) { 639 current_num_wc = spdk_max(current_num_wc * 2, required_num_wc); 640 } 641 642 max_cq_size = g_spdk_nvme_transport_opts.rdma_max_cq_size; 643 if (max_cq_size != 0 && current_num_wc > max_cq_size) { 644 current_num_wc = max_cq_size; 645 } 646 647 if (poller->current_num_wc != current_num_wc) { 648 SPDK_DEBUGLOG(nvme, "Resize RDMA CQ from %d to %d\n", poller->current_num_wc, 649 current_num_wc); 650 if (ibv_resize_cq(poller->cq, current_num_wc)) { 651 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 652 return -1; 653 } 654 655 poller->current_num_wc = current_num_wc; 656 } 657 658 poller->required_num_wc = required_num_wc; 659 return 0; 660 } 661 662 static int 663 nvme_rdma_qpair_set_poller(struct spdk_nvme_qpair *qpair) 664 { 665 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 666 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group); 667 struct nvme_rdma_poller *poller; 668 669 assert(rqpair->cq == NULL); 670 671 poller = nvme_rdma_poll_group_get_poller(group, rqpair->cm_id->verbs); 672 if (!poller) { 673 SPDK_ERRLOG("Unable to find a cq for qpair %p on poll group %p\n", qpair, qpair->poll_group); 674 return -EINVAL; 675 } 676 677 if (!poller->srq) { 678 if (nvme_rdma_resize_cq(rqpair, poller)) { 679 nvme_rdma_poll_group_put_poller(group, poller); 680 return -EPROTO; 681 } 682 } 683 684 rqpair->cq = poller->cq; 685 rqpair->srq = poller->srq; 686 if (rqpair->srq) { 687 rqpair->rsps = poller->rsps; 688 } 689 rqpair->poller = poller; 690 return 0; 691 } 692 693 static int 694 nvme_rdma_qpair_init(struct nvme_rdma_qpair *rqpair) 695 { 696 int rc; 697 struct spdk_rdma_provider_qp_init_attr attr = {}; 698 struct ibv_device_attr dev_attr; 699 struct nvme_rdma_ctrlr *rctrlr; 700 uint32_t num_cqe, max_num_cqe; 701 702 rc = ibv_query_device(rqpair->cm_id->verbs, &dev_attr); 703 if (rc != 0) { 704 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 705 return -1; 706 } 707 708 if (rqpair->qpair.poll_group) { 709 assert(!rqpair->cq); 710 rc = nvme_rdma_qpair_set_poller(&rqpair->qpair); 711 if (rc) { 712 SPDK_ERRLOG("Unable to activate the rdmaqpair.\n"); 713 return -1; 714 } 715 assert(rqpair->cq); 716 } else { 717 num_cqe = rqpair->num_entries * 2; 718 max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size; 719 if (max_num_cqe != 0 && num_cqe > max_num_cqe) { 720 num_cqe = max_num_cqe; 721 } 722 rqpair->cq = ibv_create_cq(rqpair->cm_id->verbs, num_cqe, rqpair, NULL, 0); 723 if (!rqpair->cq) { 724 SPDK_ERRLOG("Unable to create completion queue: errno %d: %s\n", errno, spdk_strerror(errno)); 725 return -1; 726 } 727 } 728 729 rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr); 730 if (g_nvme_hooks.get_ibv_pd) { 731 attr.pd = g_nvme_hooks.get_ibv_pd(&rctrlr->ctrlr.trid, rqpair->cm_id->verbs); 732 } else { 733 attr.pd = spdk_rdma_utils_get_pd(rqpair->cm_id->verbs); 734 } 735 736 attr.stats = rqpair->poller ? &rqpair->poller->stats.rdma_stats : NULL; 737 attr.send_cq = rqpair->cq; 738 attr.recv_cq = rqpair->cq; 739 attr.cap.max_send_wr = rqpair->num_entries; /* SEND operations */ 740 if (rqpair->srq) { 741 attr.srq = rqpair->srq->srq; 742 } else { 743 attr.cap.max_recv_wr = rqpair->num_entries; /* RECV operations */ 744 } 745 attr.cap.max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, dev_attr.max_sge); 746 attr.cap.max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, dev_attr.max_sge); 747 attr.domain_transfer = spdk_rdma_provider_accel_sequence_supported() ? 748 nvme_rdma_memory_domain_transfer_data : NULL; 749 750 rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &attr); 751 752 if (!rqpair->rdma_qp) { 753 return -1; 754 } 755 756 /* ibv_create_qp will change the values in attr.cap. Make sure we store the proper value. */ 757 rqpair->max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, attr.cap.max_send_sge); 758 rqpair->current_num_sends = 0; 759 760 rqpair->cm_id->context = rqpair; 761 762 return 0; 763 } 764 765 static void 766 nvme_rdma_reset_failed_sends(struct nvme_rdma_qpair *rqpair, 767 struct ibv_send_wr *bad_send_wr) 768 { 769 while (bad_send_wr != NULL) { 770 assert(rqpair->current_num_sends > 0); 771 rqpair->current_num_sends--; 772 bad_send_wr = bad_send_wr->next; 773 } 774 } 775 776 static void 777 nvme_rdma_reset_failed_recvs(struct nvme_rdma_rsps *rsps, 778 struct ibv_recv_wr *bad_recv_wr, int rc) 779 { 780 SPDK_ERRLOG("Failed to post WRs on receive queue, errno %d (%s), bad_wr %p\n", 781 rc, spdk_strerror(rc), bad_recv_wr); 782 while (bad_recv_wr != NULL) { 783 assert(rsps->current_num_recvs > 0); 784 rsps->current_num_recvs--; 785 bad_recv_wr = bad_recv_wr->next; 786 } 787 } 788 789 static inline int 790 nvme_rdma_qpair_submit_sends(struct nvme_rdma_qpair *rqpair) 791 { 792 struct ibv_send_wr *bad_send_wr = NULL; 793 int rc; 794 795 rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_send_wr); 796 797 if (spdk_unlikely(rc)) { 798 SPDK_ERRLOG("Failed to post WRs on send queue, errno %d (%s), bad_wr %p\n", 799 rc, spdk_strerror(rc), bad_send_wr); 800 nvme_rdma_reset_failed_sends(rqpair, bad_send_wr); 801 } 802 803 return rc; 804 } 805 806 static inline int 807 nvme_rdma_qpair_submit_recvs(struct nvme_rdma_qpair *rqpair) 808 { 809 struct ibv_recv_wr *bad_recv_wr; 810 int rc = 0; 811 812 rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 813 if (spdk_unlikely(rc)) { 814 nvme_rdma_reset_failed_recvs(rqpair->rsps, bad_recv_wr, rc); 815 } 816 817 return rc; 818 } 819 820 static inline int 821 nvme_rdma_poller_submit_recvs(struct nvme_rdma_poller *poller) 822 { 823 struct ibv_recv_wr *bad_recv_wr; 824 int rc; 825 826 rc = spdk_rdma_provider_srq_flush_recv_wrs(poller->srq, &bad_recv_wr); 827 if (spdk_unlikely(rc)) { 828 nvme_rdma_reset_failed_recvs(poller->rsps, bad_recv_wr, rc); 829 } 830 831 return rc; 832 } 833 834 #define nvme_rdma_trace_ibv_sge(sg_list) \ 835 if (sg_list) { \ 836 SPDK_DEBUGLOG(nvme, "local addr %p length 0x%x lkey 0x%x\n", \ 837 (void *)(sg_list)->addr, (sg_list)->length, (sg_list)->lkey); \ 838 } 839 840 static void 841 nvme_rdma_free_rsps(struct nvme_rdma_rsps *rsps) 842 { 843 if (!rsps) { 844 return; 845 } 846 847 spdk_free(rsps->rsps); 848 spdk_free(rsps->rsp_sgls); 849 spdk_free(rsps->rsp_recv_wrs); 850 spdk_free(rsps); 851 } 852 853 static struct nvme_rdma_rsps * 854 nvme_rdma_create_rsps(struct nvme_rdma_rsp_opts *opts) 855 { 856 struct nvme_rdma_rsps *rsps; 857 struct spdk_rdma_utils_memory_translation translation; 858 uint16_t i; 859 int rc; 860 861 rsps = spdk_zmalloc(sizeof(*rsps), 0, NULL, SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 862 if (!rsps) { 863 SPDK_ERRLOG("Failed to allocate rsps object\n"); 864 return NULL; 865 } 866 867 rsps->rsp_sgls = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_sgls), 0, NULL, 868 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 869 if (!rsps->rsp_sgls) { 870 SPDK_ERRLOG("Failed to allocate rsp_sgls\n"); 871 goto fail; 872 } 873 874 rsps->rsp_recv_wrs = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_recv_wrs), 0, NULL, 875 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 876 if (!rsps->rsp_recv_wrs) { 877 SPDK_ERRLOG("Failed to allocate rsp_recv_wrs\n"); 878 goto fail; 879 } 880 881 rsps->rsps = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsps), 0, NULL, 882 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 883 if (!rsps->rsps) { 884 SPDK_ERRLOG("can not allocate rdma rsps\n"); 885 goto fail; 886 } 887 888 for (i = 0; i < opts->num_entries; i++) { 889 struct ibv_sge *rsp_sgl = &rsps->rsp_sgls[i]; 890 struct spdk_nvme_rdma_rsp *rsp = &rsps->rsps[i]; 891 struct ibv_recv_wr *recv_wr = &rsps->rsp_recv_wrs[i]; 892 893 rsp->rqpair = opts->rqpair; 894 rsp->rdma_wr.type = RDMA_WR_TYPE_RECV; 895 rsp->recv_wr = recv_wr; 896 rsp_sgl->addr = (uint64_t)rsp; 897 rsp_sgl->length = sizeof(struct spdk_nvme_cpl); 898 rc = spdk_rdma_utils_get_translation(opts->mr_map, rsp, sizeof(*rsp), &translation); 899 if (rc) { 900 goto fail; 901 } 902 rsp_sgl->lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 903 904 recv_wr->wr_id = (uint64_t)&rsp->rdma_wr; 905 recv_wr->next = NULL; 906 recv_wr->sg_list = rsp_sgl; 907 recv_wr->num_sge = 1; 908 909 nvme_rdma_trace_ibv_sge(recv_wr->sg_list); 910 911 if (opts->rqpair) { 912 spdk_rdma_provider_qp_queue_recv_wrs(opts->rqpair->rdma_qp, recv_wr); 913 } else { 914 spdk_rdma_provider_srq_queue_recv_wrs(opts->srq, recv_wr); 915 } 916 } 917 918 rsps->num_entries = opts->num_entries; 919 rsps->current_num_recvs = opts->num_entries; 920 921 return rsps; 922 fail: 923 nvme_rdma_free_rsps(rsps); 924 return NULL; 925 } 926 927 static void 928 nvme_rdma_free_reqs(struct nvme_rdma_qpair *rqpair) 929 { 930 if (!rqpair->rdma_reqs) { 931 return; 932 } 933 934 spdk_free(rqpair->cmds); 935 rqpair->cmds = NULL; 936 937 spdk_free(rqpair->rdma_reqs); 938 rqpair->rdma_reqs = NULL; 939 } 940 941 static int 942 nvme_rdma_create_reqs(struct nvme_rdma_qpair *rqpair) 943 { 944 struct spdk_rdma_utils_memory_translation translation; 945 uint16_t i; 946 int rc; 947 948 assert(!rqpair->rdma_reqs); 949 rqpair->rdma_reqs = spdk_zmalloc(rqpair->num_entries * sizeof(struct spdk_nvme_rdma_req), 0, NULL, 950 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 951 if (rqpair->rdma_reqs == NULL) { 952 SPDK_ERRLOG("Failed to allocate rdma_reqs\n"); 953 goto fail; 954 } 955 956 assert(!rqpair->cmds); 957 rqpair->cmds = spdk_zmalloc(rqpair->num_entries * sizeof(*rqpair->cmds), 0, NULL, 958 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 959 if (!rqpair->cmds) { 960 SPDK_ERRLOG("Failed to allocate RDMA cmds\n"); 961 goto fail; 962 } 963 964 TAILQ_INIT(&rqpair->free_reqs); 965 TAILQ_INIT(&rqpair->outstanding_reqs); 966 for (i = 0; i < rqpair->num_entries; i++) { 967 struct spdk_nvme_rdma_req *rdma_req; 968 struct spdk_nvmf_cmd *cmd; 969 970 rdma_req = &rqpair->rdma_reqs[i]; 971 rdma_req->rdma_wr.type = RDMA_WR_TYPE_SEND; 972 cmd = &rqpair->cmds[i]; 973 974 rdma_req->id = i; 975 976 rc = spdk_rdma_utils_get_translation(rqpair->mr_map, cmd, sizeof(*cmd), &translation); 977 if (rc) { 978 goto fail; 979 } 980 rdma_req->send_sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 981 982 /* The first RDMA sgl element will always point 983 * at this data structure. Depending on whether 984 * an NVMe-oF SGL is required, the length of 985 * this element may change. */ 986 rdma_req->send_sgl[0].addr = (uint64_t)cmd; 987 rdma_req->send_wr.wr_id = (uint64_t)&rdma_req->rdma_wr; 988 rdma_req->send_wr.next = NULL; 989 rdma_req->send_wr.opcode = IBV_WR_SEND; 990 rdma_req->send_wr.send_flags = IBV_SEND_SIGNALED; 991 rdma_req->send_wr.sg_list = rdma_req->send_sgl; 992 rdma_req->send_wr.imm_data = 0; 993 994 TAILQ_INSERT_TAIL(&rqpair->free_reqs, rdma_req, link); 995 } 996 997 return 0; 998 fail: 999 nvme_rdma_free_reqs(rqpair); 1000 return -ENOMEM; 1001 } 1002 1003 static int nvme_rdma_connect(struct nvme_rdma_qpair *rqpair); 1004 1005 static int 1006 nvme_rdma_route_resolved(struct nvme_rdma_qpair *rqpair, int ret) 1007 { 1008 if (ret) { 1009 SPDK_ERRLOG("RDMA route resolution error\n"); 1010 return -1; 1011 } 1012 1013 ret = nvme_rdma_qpair_init(rqpair); 1014 if (ret < 0) { 1015 SPDK_ERRLOG("nvme_rdma_qpair_init() failed\n"); 1016 return -1; 1017 } 1018 1019 return nvme_rdma_connect(rqpair); 1020 } 1021 1022 static int 1023 nvme_rdma_addr_resolved(struct nvme_rdma_qpair *rqpair, int ret) 1024 { 1025 if (ret) { 1026 SPDK_ERRLOG("RDMA address resolution error\n"); 1027 return -1; 1028 } 1029 1030 if (rqpair->qpair.ctrlr->opts.transport_ack_timeout != SPDK_NVME_TRANSPORT_ACK_TIMEOUT_DISABLED) { 1031 #ifdef SPDK_CONFIG_RDMA_SET_ACK_TIMEOUT 1032 uint8_t timeout = rqpair->qpair.ctrlr->opts.transport_ack_timeout; 1033 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, 1034 RDMA_OPTION_ID_ACK_TIMEOUT, 1035 &timeout, sizeof(timeout)); 1036 if (ret) { 1037 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_ACK_TIMEOUT %d, ret %d\n", timeout, ret); 1038 } 1039 #else 1040 SPDK_DEBUGLOG(nvme, "transport_ack_timeout is not supported\n"); 1041 #endif 1042 } 1043 1044 if (rqpair->qpair.ctrlr->opts.transport_tos != SPDK_NVME_TRANSPORT_TOS_DISABLED) { 1045 #ifdef SPDK_CONFIG_RDMA_SET_TOS 1046 uint8_t tos = rqpair->qpair.ctrlr->opts.transport_tos; 1047 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_TOS, &tos, sizeof(tos)); 1048 if (ret) { 1049 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_TOS %u, ret %d\n", tos, ret); 1050 } 1051 #else 1052 SPDK_DEBUGLOG(nvme, "transport_tos is not supported\n"); 1053 #endif 1054 } 1055 1056 ret = rdma_resolve_route(rqpair->cm_id, NVME_RDMA_TIME_OUT_IN_MS); 1057 if (ret) { 1058 SPDK_ERRLOG("rdma_resolve_route\n"); 1059 return ret; 1060 } 1061 1062 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ROUTE_RESOLVED, 1063 nvme_rdma_route_resolved); 1064 } 1065 1066 static int 1067 nvme_rdma_resolve_addr(struct nvme_rdma_qpair *rqpair, 1068 struct sockaddr *src_addr, 1069 struct sockaddr *dst_addr) 1070 { 1071 int ret; 1072 1073 if (src_addr) { 1074 int reuse = 1; 1075 1076 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_REUSEADDR, 1077 &reuse, sizeof(reuse)); 1078 if (ret) { 1079 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_REUSEADDR %d, ret %d\n", 1080 reuse, ret); 1081 /* It is likely that rdma_resolve_addr() returns -EADDRINUSE, but 1082 * we may missing something. We rely on rdma_resolve_addr(). 1083 */ 1084 } 1085 } 1086 1087 ret = rdma_resolve_addr(rqpair->cm_id, src_addr, dst_addr, 1088 NVME_RDMA_TIME_OUT_IN_MS); 1089 if (ret) { 1090 SPDK_ERRLOG("rdma_resolve_addr, %d\n", errno); 1091 return ret; 1092 } 1093 1094 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ADDR_RESOLVED, 1095 nvme_rdma_addr_resolved); 1096 } 1097 1098 static int nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair); 1099 1100 static int 1101 nvme_rdma_connect_established(struct nvme_rdma_qpair *rqpair, int ret) 1102 { 1103 struct nvme_rdma_rsp_opts opts = {}; 1104 1105 if (ret == -ESTALE) { 1106 return nvme_rdma_stale_conn_retry(rqpair); 1107 } else if (ret) { 1108 SPDK_ERRLOG("RDMA connect error %d\n", ret); 1109 return ret; 1110 } 1111 1112 assert(!rqpair->mr_map); 1113 rqpair->mr_map = spdk_rdma_utils_create_mem_map(rqpair->rdma_qp->qp->pd, &g_nvme_hooks, 1114 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE); 1115 if (!rqpair->mr_map) { 1116 SPDK_ERRLOG("Unable to register RDMA memory translation map\n"); 1117 return -1; 1118 } 1119 1120 ret = nvme_rdma_create_reqs(rqpair); 1121 SPDK_DEBUGLOG(nvme, "rc =%d\n", ret); 1122 if (ret) { 1123 SPDK_ERRLOG("Unable to create rqpair RDMA requests\n"); 1124 return -1; 1125 } 1126 SPDK_DEBUGLOG(nvme, "RDMA requests created\n"); 1127 1128 if (!rqpair->srq) { 1129 opts.num_entries = rqpair->num_entries; 1130 opts.rqpair = rqpair; 1131 opts.srq = NULL; 1132 opts.mr_map = rqpair->mr_map; 1133 1134 assert(!rqpair->rsps); 1135 rqpair->rsps = nvme_rdma_create_rsps(&opts); 1136 if (!rqpair->rsps) { 1137 SPDK_ERRLOG("Unable to create rqpair RDMA responses\n"); 1138 return -1; 1139 } 1140 SPDK_DEBUGLOG(nvme, "RDMA responses created\n"); 1141 1142 ret = nvme_rdma_qpair_submit_recvs(rqpair); 1143 SPDK_DEBUGLOG(nvme, "rc =%d\n", ret); 1144 if (ret) { 1145 SPDK_ERRLOG("Unable to submit rqpair RDMA responses\n"); 1146 return -1; 1147 } 1148 SPDK_DEBUGLOG(nvme, "RDMA responses submitted\n"); 1149 } 1150 1151 rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND; 1152 1153 return 0; 1154 } 1155 1156 static int 1157 nvme_rdma_connect(struct nvme_rdma_qpair *rqpair) 1158 { 1159 struct rdma_conn_param param = {}; 1160 struct spdk_nvmf_rdma_request_private_data request_data = {}; 1161 struct ibv_device_attr attr; 1162 int ret; 1163 struct spdk_nvme_ctrlr *ctrlr; 1164 1165 ret = ibv_query_device(rqpair->cm_id->verbs, &attr); 1166 if (ret != 0) { 1167 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1168 return ret; 1169 } 1170 1171 param.responder_resources = attr.max_qp_rd_atom; 1172 1173 ctrlr = rqpair->qpair.ctrlr; 1174 if (!ctrlr) { 1175 return -1; 1176 } 1177 1178 request_data.qid = rqpair->qpair.id; 1179 request_data.hrqsize = rqpair->num_entries + 1; 1180 request_data.hsqsize = rqpair->num_entries; 1181 request_data.cntlid = ctrlr->cntlid; 1182 1183 param.private_data = &request_data; 1184 param.private_data_len = sizeof(request_data); 1185 param.retry_count = ctrlr->opts.transport_retry_count; 1186 param.rnr_retry_count = 7; 1187 1188 /* Fields below are ignored by rdma cm if qpair has been 1189 * created using rdma cm API. */ 1190 param.srq = 0; 1191 param.qp_num = rqpair->rdma_qp->qp->qp_num; 1192 1193 ret = rdma_connect(rqpair->cm_id, ¶m); 1194 if (ret) { 1195 SPDK_ERRLOG("nvme rdma connect error\n"); 1196 return ret; 1197 } 1198 1199 ctrlr->numa.id_valid = 1; 1200 ctrlr->numa.id = spdk_rdma_cm_id_get_numa_id(rqpair->cm_id); 1201 1202 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ESTABLISHED, 1203 nvme_rdma_connect_established); 1204 } 1205 1206 static int 1207 nvme_rdma_ctrlr_connect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 1208 { 1209 struct sockaddr_storage dst_addr; 1210 struct sockaddr_storage src_addr; 1211 bool src_addr_specified; 1212 long int port, src_port = 0; 1213 int rc; 1214 struct nvme_rdma_ctrlr *rctrlr; 1215 struct nvme_rdma_qpair *rqpair; 1216 struct nvme_rdma_poll_group *group; 1217 int family; 1218 1219 rqpair = nvme_rdma_qpair(qpair); 1220 rctrlr = nvme_rdma_ctrlr(ctrlr); 1221 assert(rctrlr != NULL); 1222 1223 switch (ctrlr->trid.adrfam) { 1224 case SPDK_NVMF_ADRFAM_IPV4: 1225 family = AF_INET; 1226 break; 1227 case SPDK_NVMF_ADRFAM_IPV6: 1228 family = AF_INET6; 1229 break; 1230 default: 1231 SPDK_ERRLOG("Unhandled ADRFAM %d\n", ctrlr->trid.adrfam); 1232 return -1; 1233 } 1234 1235 SPDK_DEBUGLOG(nvme, "adrfam %d ai_family %d\n", ctrlr->trid.adrfam, family); 1236 1237 memset(&dst_addr, 0, sizeof(dst_addr)); 1238 1239 SPDK_DEBUGLOG(nvme, "trsvcid is %s\n", ctrlr->trid.trsvcid); 1240 rc = nvme_parse_addr(&dst_addr, family, ctrlr->trid.traddr, ctrlr->trid.trsvcid, &port); 1241 if (rc != 0) { 1242 SPDK_ERRLOG("dst_addr nvme_parse_addr() failed\n"); 1243 return -1; 1244 } 1245 1246 if (ctrlr->opts.src_addr[0] || ctrlr->opts.src_svcid[0]) { 1247 memset(&src_addr, 0, sizeof(src_addr)); 1248 rc = nvme_parse_addr(&src_addr, family, 1249 ctrlr->opts.src_addr[0] ? ctrlr->opts.src_addr : NULL, 1250 ctrlr->opts.src_svcid[0] ? ctrlr->opts.src_svcid : NULL, 1251 &src_port); 1252 if (rc != 0) { 1253 SPDK_ERRLOG("src_addr nvme_parse_addr() failed\n"); 1254 return -1; 1255 } 1256 src_addr_specified = true; 1257 } else { 1258 src_addr_specified = false; 1259 } 1260 1261 rc = rdma_create_id(rctrlr->cm_channel, &rqpair->cm_id, rqpair, RDMA_PS_TCP); 1262 if (rc < 0) { 1263 SPDK_ERRLOG("rdma_create_id() failed\n"); 1264 return -1; 1265 } 1266 1267 rc = nvme_rdma_resolve_addr(rqpair, 1268 src_addr_specified ? (struct sockaddr *)&src_addr : NULL, 1269 (struct sockaddr *)&dst_addr); 1270 if (rc < 0) { 1271 SPDK_ERRLOG("nvme_rdma_resolve_addr() failed\n"); 1272 return -1; 1273 } 1274 1275 rqpair->state = NVME_RDMA_QPAIR_STATE_INITIALIZING; 1276 1277 if (qpair->poll_group != NULL) { 1278 group = nvme_rdma_poll_group(qpair->poll_group); 1279 TAILQ_INSERT_TAIL(&group->connecting_qpairs, rqpair, link_connecting); 1280 } 1281 1282 return 0; 1283 } 1284 1285 static int 1286 nvme_rdma_stale_conn_reconnect(struct nvme_rdma_qpair *rqpair) 1287 { 1288 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 1289 1290 if (spdk_get_ticks() < rqpair->evt_timeout_ticks) { 1291 return -EAGAIN; 1292 } 1293 1294 return nvme_rdma_ctrlr_connect_qpair(qpair->ctrlr, qpair); 1295 } 1296 1297 static int 1298 nvme_rdma_ctrlr_connect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, 1299 struct spdk_nvme_qpair *qpair) 1300 { 1301 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 1302 int rc; 1303 1304 if (rqpair->in_connect_poll) { 1305 return -EAGAIN; 1306 } 1307 1308 rqpair->in_connect_poll = true; 1309 1310 switch (rqpair->state) { 1311 case NVME_RDMA_QPAIR_STATE_INVALID: 1312 rc = -EAGAIN; 1313 break; 1314 1315 case NVME_RDMA_QPAIR_STATE_INITIALIZING: 1316 case NVME_RDMA_QPAIR_STATE_EXITING: 1317 if (!nvme_qpair_is_admin_queue(qpair)) { 1318 nvme_ctrlr_lock(ctrlr); 1319 } 1320 1321 rc = nvme_rdma_process_event_poll(rqpair); 1322 1323 if (!nvme_qpair_is_admin_queue(qpair)) { 1324 nvme_ctrlr_unlock(ctrlr); 1325 } 1326 1327 if (rc == 0) { 1328 rc = -EAGAIN; 1329 } 1330 rqpair->in_connect_poll = false; 1331 1332 return rc; 1333 1334 case NVME_RDMA_QPAIR_STATE_STALE_CONN: 1335 rc = nvme_rdma_stale_conn_reconnect(rqpair); 1336 if (rc == 0) { 1337 rc = -EAGAIN; 1338 } 1339 break; 1340 case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND: 1341 rc = nvme_fabric_qpair_connect_async(qpair, rqpair->num_entries + 1); 1342 if (rc == 0) { 1343 rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL; 1344 rc = -EAGAIN; 1345 } else { 1346 SPDK_ERRLOG("Failed to send an NVMe-oF Fabric CONNECT command\n"); 1347 } 1348 break; 1349 case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL: 1350 rc = nvme_fabric_qpair_connect_poll(qpair); 1351 if (rc == 0) { 1352 if (nvme_fabric_qpair_auth_required(qpair)) { 1353 rc = nvme_fabric_qpair_authenticate_async(qpair); 1354 if (rc == 0) { 1355 rqpair->state = NVME_RDMA_QPAIR_STATE_AUTHENTICATING; 1356 rc = -EAGAIN; 1357 } 1358 } else { 1359 rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING; 1360 nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED); 1361 } 1362 } else if (rc != -EAGAIN) { 1363 SPDK_ERRLOG("Failed to poll NVMe-oF Fabric CONNECT command\n"); 1364 } 1365 break; 1366 case NVME_RDMA_QPAIR_STATE_AUTHENTICATING: 1367 rc = nvme_fabric_qpair_authenticate_poll(qpair); 1368 if (rc == 0) { 1369 rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING; 1370 nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED); 1371 } 1372 break; 1373 case NVME_RDMA_QPAIR_STATE_RUNNING: 1374 rc = 0; 1375 break; 1376 default: 1377 assert(false); 1378 rc = -EINVAL; 1379 break; 1380 } 1381 1382 rqpair->in_connect_poll = false; 1383 1384 return rc; 1385 } 1386 1387 static inline int 1388 nvme_rdma_get_memory_translation(struct nvme_request *req, struct nvme_rdma_qpair *rqpair, 1389 struct nvme_rdma_memory_translation_ctx *_ctx) 1390 { 1391 struct spdk_memory_domain_translation_ctx ctx; 1392 struct spdk_memory_domain_translation_result dma_translation = {.iov_count = 0}; 1393 struct spdk_rdma_utils_memory_translation rdma_translation; 1394 int rc; 1395 1396 assert(req); 1397 assert(rqpair); 1398 assert(_ctx); 1399 1400 if (req->payload.opts && req->payload.opts->memory_domain) { 1401 ctx.size = sizeof(struct spdk_memory_domain_translation_ctx); 1402 ctx.rdma.ibv_qp = rqpair->rdma_qp->qp; 1403 dma_translation.size = sizeof(struct spdk_memory_domain_translation_result); 1404 1405 rc = spdk_memory_domain_translate_data(req->payload.opts->memory_domain, 1406 req->payload.opts->memory_domain_ctx, 1407 rqpair->rdma_qp->domain, &ctx, _ctx->addr, 1408 _ctx->length, &dma_translation); 1409 if (spdk_unlikely(rc) || dma_translation.iov_count != 1) { 1410 SPDK_ERRLOG("DMA memory translation failed, rc %d, iov count %u\n", rc, dma_translation.iov_count); 1411 return rc; 1412 } 1413 1414 _ctx->lkey = dma_translation.rdma.lkey; 1415 _ctx->rkey = dma_translation.rdma.rkey; 1416 _ctx->addr = dma_translation.iov.iov_base; 1417 _ctx->length = dma_translation.iov.iov_len; 1418 } else { 1419 rc = spdk_rdma_utils_get_translation(rqpair->mr_map, _ctx->addr, _ctx->length, &rdma_translation); 1420 if (spdk_unlikely(rc)) { 1421 SPDK_ERRLOG("RDMA memory translation failed, rc %d\n", rc); 1422 return rc; 1423 } 1424 if (rdma_translation.translation_type == SPDK_RDMA_UTILS_TRANSLATION_MR) { 1425 _ctx->lkey = rdma_translation.mr_or_key.mr->lkey; 1426 _ctx->rkey = rdma_translation.mr_or_key.mr->rkey; 1427 } else { 1428 _ctx->lkey = _ctx->rkey = (uint32_t)rdma_translation.mr_or_key.key; 1429 } 1430 } 1431 1432 return 0; 1433 } 1434 1435 1436 /* 1437 * Build SGL describing empty payload. 1438 */ 1439 static int 1440 nvme_rdma_build_null_request(struct spdk_nvme_rdma_req *rdma_req) 1441 { 1442 struct nvme_request *req = rdma_req->req; 1443 1444 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1445 1446 /* The first element of this SGL is pointing at an 1447 * spdk_nvmf_cmd object. For this particular command, 1448 * we only need the first 64 bytes corresponding to 1449 * the NVMe command. */ 1450 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1451 1452 /* The RDMA SGL needs one element describing the NVMe command. */ 1453 rdma_req->send_wr.num_sge = 1; 1454 1455 req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1456 req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1457 req->cmd.dptr.sgl1.keyed.length = 0; 1458 req->cmd.dptr.sgl1.keyed.key = 0; 1459 req->cmd.dptr.sgl1.address = 0; 1460 1461 return 0; 1462 } 1463 1464 static inline void 1465 nvme_rdma_configure_contig_inline_request(struct spdk_nvme_rdma_req *rdma_req, 1466 struct nvme_request *req, struct nvme_rdma_memory_translation_ctx *ctx) 1467 { 1468 rdma_req->send_sgl[1].lkey = ctx->lkey; 1469 1470 /* The first element of this SGL is pointing at an 1471 * spdk_nvmf_cmd object. For this particular command, 1472 * we only need the first 64 bytes corresponding to 1473 * the NVMe command. */ 1474 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1475 1476 rdma_req->send_sgl[1].addr = (uint64_t)ctx->addr; 1477 rdma_req->send_sgl[1].length = (uint32_t)ctx->length; 1478 1479 /* The RDMA SGL contains two elements. The first describes 1480 * the NVMe command and the second describes the data 1481 * payload. */ 1482 rdma_req->send_wr.num_sge = 2; 1483 1484 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1485 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK; 1486 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1487 req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx->length; 1488 /* Inline only supported for icdoff == 0 currently. This function will 1489 * not get called for controllers with other values. */ 1490 req->cmd.dptr.sgl1.address = (uint64_t)0; 1491 } 1492 1493 /* 1494 * Build inline SGL describing contiguous payload buffer. 1495 */ 1496 static inline int 1497 nvme_rdma_build_contig_inline_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(ctx.length != 0); 1508 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG); 1509 1510 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1511 if (spdk_unlikely(rc)) { 1512 return -1; 1513 } 1514 1515 nvme_rdma_configure_contig_inline_request(rdma_req, req, &ctx); 1516 1517 return 0; 1518 } 1519 1520 static inline void 1521 nvme_rdma_configure_contig_request(struct spdk_nvme_rdma_req *rdma_req, struct nvme_request *req, 1522 struct nvme_rdma_memory_translation_ctx *ctx) 1523 { 1524 req->cmd.dptr.sgl1.keyed.key = ctx->rkey; 1525 1526 /* The first element of this SGL is pointing at an 1527 * spdk_nvmf_cmd object. For this particular command, 1528 * we only need the first 64 bytes corresponding to 1529 * the NVMe command. */ 1530 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1531 1532 /* The RDMA SGL needs one element describing the NVMe command. */ 1533 rdma_req->send_wr.num_sge = 1; 1534 1535 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1536 req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1537 req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1538 req->cmd.dptr.sgl1.keyed.length = (uint32_t)ctx->length; 1539 req->cmd.dptr.sgl1.address = (uint64_t)ctx->addr; 1540 } 1541 1542 /* 1543 * Build SGL describing contiguous payload buffer. 1544 */ 1545 static inline int 1546 nvme_rdma_build_contig_request(struct nvme_rdma_qpair *rqpair, 1547 struct spdk_nvme_rdma_req *rdma_req) 1548 { 1549 struct nvme_request *req = rdma_req->req; 1550 struct nvme_rdma_memory_translation_ctx ctx = { 1551 .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset, 1552 .length = req->payload_size 1553 }; 1554 int rc; 1555 1556 assert(req->payload_size != 0); 1557 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG); 1558 1559 if (spdk_unlikely(req->payload_size > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) { 1560 SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n", 1561 req->payload_size, NVME_RDMA_MAX_KEYED_SGL_LENGTH); 1562 return -1; 1563 } 1564 1565 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1566 if (spdk_unlikely(rc)) { 1567 return -1; 1568 } 1569 1570 nvme_rdma_configure_contig_request(rdma_req, req, &ctx); 1571 1572 return 0; 1573 } 1574 1575 /* 1576 * Build SGL describing scattered payload buffer. 1577 */ 1578 static inline int 1579 nvme_rdma_build_sgl_request(struct nvme_rdma_qpair *rqpair, 1580 struct spdk_nvme_rdma_req *rdma_req) 1581 { 1582 struct nvme_request *req = rdma_req->req; 1583 struct spdk_nvmf_cmd *cmd = &rqpair->cmds[rdma_req->id]; 1584 struct nvme_rdma_memory_translation_ctx ctx; 1585 uint32_t remaining_size; 1586 uint32_t sge_length; 1587 int rc, max_num_sgl, num_sgl_desc; 1588 1589 assert(req->payload_size != 0); 1590 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL); 1591 assert(req->payload.reset_sgl_fn != NULL); 1592 assert(req->payload.next_sge_fn != NULL); 1593 req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset); 1594 1595 max_num_sgl = req->qpair->ctrlr->max_sges; 1596 1597 remaining_size = req->payload_size; 1598 num_sgl_desc = 0; 1599 do { 1600 rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &sge_length); 1601 if (spdk_unlikely(rc)) { 1602 return -1; 1603 } 1604 1605 sge_length = spdk_min(remaining_size, sge_length); 1606 1607 if (spdk_unlikely(sge_length > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) { 1608 SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n", 1609 sge_length, NVME_RDMA_MAX_KEYED_SGL_LENGTH); 1610 return -1; 1611 } 1612 ctx.length = sge_length; 1613 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1614 if (spdk_unlikely(rc)) { 1615 return -1; 1616 } 1617 1618 cmd->sgl[num_sgl_desc].keyed.key = ctx.rkey; 1619 cmd->sgl[num_sgl_desc].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1620 cmd->sgl[num_sgl_desc].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1621 cmd->sgl[num_sgl_desc].keyed.length = (uint32_t)ctx.length; 1622 cmd->sgl[num_sgl_desc].address = (uint64_t)ctx.addr; 1623 1624 remaining_size -= ctx.length; 1625 num_sgl_desc++; 1626 } while (remaining_size > 0 && num_sgl_desc < max_num_sgl); 1627 1628 1629 /* Should be impossible if we did our sgl checks properly up the stack, but do a sanity check here. */ 1630 if (spdk_unlikely(remaining_size > 0)) { 1631 return -1; 1632 } 1633 1634 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1635 1636 /* The RDMA SGL needs one element describing some portion 1637 * of the spdk_nvmf_cmd structure. */ 1638 rdma_req->send_wr.num_sge = 1; 1639 1640 /* 1641 * If only one SGL descriptor is required, it can be embedded directly in the command 1642 * as a data block descriptor. 1643 */ 1644 if (num_sgl_desc == 1) { 1645 /* The first element of this SGL is pointing at an 1646 * spdk_nvmf_cmd object. For this particular command, 1647 * we only need the first 64 bytes corresponding to 1648 * the NVMe command. */ 1649 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1650 1651 req->cmd.dptr.sgl1.keyed.type = cmd->sgl[0].keyed.type; 1652 req->cmd.dptr.sgl1.keyed.subtype = cmd->sgl[0].keyed.subtype; 1653 req->cmd.dptr.sgl1.keyed.length = cmd->sgl[0].keyed.length; 1654 req->cmd.dptr.sgl1.keyed.key = cmd->sgl[0].keyed.key; 1655 req->cmd.dptr.sgl1.address = cmd->sgl[0].address; 1656 } else { 1657 /* 1658 * Otherwise, The SGL descriptor embedded in the command must point to the list of 1659 * SGL descriptors used to describe the operation. In that case it is a last segment descriptor. 1660 */ 1661 uint32_t descriptors_size = sizeof(struct spdk_nvme_sgl_descriptor) * num_sgl_desc; 1662 1663 if (spdk_unlikely(descriptors_size > rqpair->qpair.ctrlr->ioccsz_bytes)) { 1664 SPDK_ERRLOG("Size of SGL descriptors (%u) exceeds ICD (%u)\n", 1665 descriptors_size, rqpair->qpair.ctrlr->ioccsz_bytes); 1666 return -1; 1667 } 1668 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd) + descriptors_size; 1669 1670 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT; 1671 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1672 req->cmd.dptr.sgl1.unkeyed.length = descriptors_size; 1673 req->cmd.dptr.sgl1.address = (uint64_t)0; 1674 } 1675 1676 return 0; 1677 } 1678 1679 /* 1680 * Build inline SGL describing sgl payload buffer. 1681 */ 1682 static inline int 1683 nvme_rdma_build_sgl_inline_request(struct nvme_rdma_qpair *rqpair, 1684 struct spdk_nvme_rdma_req *rdma_req) 1685 { 1686 struct nvme_request *req = rdma_req->req; 1687 struct nvme_rdma_memory_translation_ctx ctx; 1688 uint32_t length; 1689 int rc; 1690 1691 assert(req->payload_size != 0); 1692 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL); 1693 assert(req->payload.reset_sgl_fn != NULL); 1694 assert(req->payload.next_sge_fn != NULL); 1695 req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset); 1696 1697 rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &length); 1698 if (spdk_unlikely(rc)) { 1699 return -1; 1700 } 1701 1702 if (length < req->payload_size) { 1703 SPDK_DEBUGLOG(nvme, "Inline SGL request split so sending separately.\n"); 1704 return nvme_rdma_build_sgl_request(rqpair, rdma_req); 1705 } 1706 1707 if (length > req->payload_size) { 1708 length = req->payload_size; 1709 } 1710 1711 ctx.length = length; 1712 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1713 if (spdk_unlikely(rc)) { 1714 return -1; 1715 } 1716 1717 rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr; 1718 rdma_req->send_sgl[1].length = (uint32_t)ctx.length; 1719 rdma_req->send_sgl[1].lkey = ctx.lkey; 1720 1721 rdma_req->send_wr.num_sge = 2; 1722 1723 /* The first element of this SGL is pointing at an 1724 * spdk_nvmf_cmd object. For this particular command, 1725 * we only need the first 64 bytes corresponding to 1726 * the NVMe command. */ 1727 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1728 1729 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1730 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK; 1731 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1732 req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length; 1733 /* Inline only supported for icdoff == 0 currently. This function will 1734 * not get called for controllers with other values. */ 1735 req->cmd.dptr.sgl1.address = (uint64_t)0; 1736 1737 return 0; 1738 } 1739 1740 static inline int 1741 nvme_rdma_accel_append_copy(struct spdk_nvme_poll_group *pg, void **seq, 1742 struct spdk_memory_domain *rdma_domain, struct spdk_nvme_rdma_req *rdma_req, 1743 struct iovec *iovs, uint32_t iovcnt, 1744 struct spdk_memory_domain *src_domain, void *src_domain_ctx) 1745 { 1746 return pg->accel_fn_table.append_copy(pg->ctx, seq, iovs, iovcnt, rdma_domain, rdma_req, iovs, 1747 iovcnt, src_domain, src_domain_ctx, NULL, NULL); 1748 } 1749 1750 static inline void 1751 nvme_rdma_accel_reverse(struct spdk_nvme_poll_group *pg, void *seq) 1752 { 1753 pg->accel_fn_table.reverse_sequence(seq); 1754 } 1755 1756 static inline void 1757 nvme_rdma_accel_finish(struct spdk_nvme_poll_group *pg, void *seq, 1758 spdk_nvme_accel_completion_cb cb_fn, void *cb_arg) 1759 { 1760 pg->accel_fn_table.finish_sequence(seq, cb_fn, cb_arg); 1761 } 1762 1763 static inline void 1764 nvme_rdma_accel_completion_cb(void *cb_arg, int status) 1765 { 1766 struct spdk_nvme_rdma_req *rdma_req = cb_arg; 1767 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(rdma_req->req->qpair); 1768 struct spdk_nvme_cpl cpl; 1769 enum spdk_nvme_generic_command_status_code sc; 1770 uint16_t dnr = 0; 1771 1772 rdma_req->in_progress_accel = 0; 1773 rdma_req->req->accel_sequence = NULL; 1774 SPDK_DEBUGLOG(nvme, "rdma_req %p qpair %p, accel completion rc %d\n", rdma_req, rqpair, status); 1775 1776 /* nvme_rdma driver may fail data transfer on WC_FLUSH error completion which is expected. 1777 * To prevent false errors from accel, first check if qpair is in the process of disconnect */ 1778 if (spdk_unlikely(!spdk_nvme_qpair_is_connected(&rqpair->qpair))) { 1779 SPDK_DEBUGLOG(nvme, "qpair %p, req %p accel cpl in disconnecting, outstanding %u\n", 1780 rqpair, rdma_req, rqpair->qpair.num_outstanding_reqs); 1781 sc = SPDK_NVME_SC_ABORTED_SQ_DELETION; 1782 goto fail_req; 1783 } 1784 if (spdk_unlikely(status)) { 1785 SPDK_ERRLOG("qpair %p, req %p, accel sequence status %d\n", rdma_req->req->qpair, rdma_req, status); 1786 sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1787 /* Something wrong happened, let the upper layer know that retry is no desired */ 1788 dnr = 1; 1789 goto fail_req; 1790 } 1791 1792 nvme_rdma_req_complete(rdma_req, &rdma_req->rdma_rsp->cpl, true); 1793 return; 1794 1795 fail_req: 1796 memset(&cpl, 0, sizeof(cpl)); 1797 cpl.status.sc = sc; 1798 cpl.status.sct = SPDK_NVME_SCT_GENERIC; 1799 cpl.status.dnr = dnr; 1800 nvme_rdma_req_complete(rdma_req, &cpl, true); 1801 } 1802 1803 static inline int 1804 nvme_rdma_apply_accel_sequence(struct nvme_rdma_qpair *rqpair, struct nvme_request *req, 1805 struct spdk_nvme_rdma_req *rdma_req) 1806 { 1807 struct spdk_nvme_poll_group *pg = rqpair->qpair.poll_group->group; 1808 struct spdk_memory_domain *src_domain; 1809 void *src_domain_ctx; 1810 void *accel_seq = req->accel_sequence; 1811 uint32_t iovcnt = 0; 1812 int rc; 1813 1814 SPDK_DEBUGLOG(nvme, "req %p, start accel seq %p\n", rdma_req, accel_seq); 1815 if (nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL) { 1816 void *addr; 1817 uint32_t sge_length, payload_size; 1818 1819 payload_size = req->payload_size; 1820 assert(payload_size); 1821 req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset); 1822 do { 1823 rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &addr, &sge_length); 1824 if (spdk_unlikely(rc)) { 1825 return -1; 1826 } 1827 sge_length = spdk_min(payload_size, sge_length); 1828 rdma_req->iovs[iovcnt].iov_base = addr; 1829 rdma_req->iovs[iovcnt].iov_len = sge_length; 1830 iovcnt++; 1831 payload_size -= sge_length; 1832 } while (payload_size && iovcnt < NVME_RDMA_MAX_SGL_DESCRIPTORS); 1833 1834 if (spdk_unlikely(payload_size)) { 1835 SPDK_ERRLOG("not enough iovs to handle req %p, remaining len %u\n", rdma_req, payload_size); 1836 return -E2BIG; 1837 } 1838 } else { 1839 rdma_req->iovs[iovcnt].iov_base = req->payload.contig_or_cb_arg; 1840 rdma_req->iovs[iovcnt].iov_len = req->payload_size; 1841 iovcnt = 1; 1842 } 1843 if (req->payload.opts && req->payload.opts->memory_domain) { 1844 if (accel_seq) { 1845 src_domain = rqpair->rdma_qp->domain; 1846 src_domain_ctx = rdma_req; 1847 } else { 1848 src_domain = req->payload.opts->memory_domain; 1849 src_domain_ctx = req->payload.opts->memory_domain_ctx; 1850 } 1851 } else { 1852 src_domain = NULL; 1853 src_domain_ctx = NULL; 1854 } 1855 1856 rc = nvme_rdma_accel_append_copy(pg, &accel_seq, rqpair->rdma_qp->domain, rdma_req, rdma_req->iovs, 1857 iovcnt, src_domain, src_domain_ctx); 1858 if (spdk_unlikely(rc)) { 1859 return rc; 1860 } 1861 1862 if (spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1863 nvme_rdma_accel_reverse(pg, accel_seq); 1864 } 1865 1866 rdma_req->in_progress_accel = 1; 1867 TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link); 1868 rqpair->num_outstanding_reqs++; 1869 1870 SPDK_DEBUGLOG(nvme, "req %p, finish accel seq %p\n", rdma_req, accel_seq); 1871 nvme_rdma_accel_finish(pg, accel_seq, nvme_rdma_accel_completion_cb, rdma_req); 1872 1873 return 0; 1874 } 1875 1876 static inline int 1877 nvme_rdma_memory_domain_transfer_data(struct spdk_memory_domain *dst_domain, void *dst_domain_ctx, 1878 struct iovec *dst_iov, uint32_t dst_iovcnt, 1879 struct spdk_memory_domain *src_domain, void *src_domain_ctx, 1880 struct iovec *src_iov, uint32_t src_iovcnt, 1881 struct spdk_memory_domain_translation_result *translation, 1882 spdk_memory_domain_data_cpl_cb cpl_cb, void *cpl_cb_arg) 1883 { 1884 struct nvme_rdma_memory_translation_ctx ctx; 1885 struct spdk_nvme_rdma_req *rdma_req = dst_domain_ctx; 1886 struct nvme_request *req = rdma_req->req; 1887 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(rdma_req->req->qpair); 1888 struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr; 1889 bool icd_supported; 1890 1891 assert(dst_domain == rqpair->rdma_qp->domain); 1892 assert(src_domain); 1893 assert(spdk_memory_domain_get_dma_device_type(src_domain) == SPDK_DMA_DEVICE_TYPE_RDMA); 1894 /* We expect "inplace" operation */ 1895 assert(dst_iov == src_iov); 1896 assert(dst_iovcnt == src_iovcnt); 1897 1898 if (spdk_unlikely(!src_domain || 1899 spdk_memory_domain_get_dma_device_type(src_domain) != SPDK_DMA_DEVICE_TYPE_RDMA)) { 1900 SPDK_ERRLOG("Unexpected source memory domain %p, type %d\n", src_domain, 1901 src_domain ? (int)spdk_memory_domain_get_dma_device_type(src_domain) : -1); 1902 return -ENOTSUP; 1903 } 1904 if (spdk_unlikely(dst_iovcnt != 1 || !translation || translation->iov_count != 1)) { 1905 SPDK_ERRLOG("Unexpected iovcnt %u or missed translation, rdma_req %p\n", dst_iovcnt, rdma_req); 1906 return -ENOTSUP; 1907 } 1908 ctx.addr = translation->iov.iov_base; 1909 ctx.length = translation->iov.iov_len; 1910 ctx.lkey = translation->rdma.lkey; 1911 ctx.rkey = translation->rdma.rkey; 1912 1913 SPDK_DEBUGLOG(nvme, "req %p, addr %p, len %zu, key %u\n", rdma_req, ctx.addr, ctx.length, ctx.rkey); 1914 icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER 1915 && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0; 1916 1917 /* We expect that result of accel sequence is a Memory Key which describes a virtually contig address space. 1918 * That means we prepare a contig request even if original payload was scattered */ 1919 if (icd_supported) { 1920 nvme_rdma_configure_contig_inline_request(rdma_req, req, &ctx); 1921 } else { 1922 nvme_rdma_configure_contig_request(rdma_req, req, &ctx); 1923 } 1924 rdma_req->transfer_cpl_cb = cpl_cb; 1925 rdma_req->transfer_cpl_cb_arg = cpl_cb_arg; 1926 1927 memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd)); 1928 1929 return _nvme_rdma_qpair_submit_request(rqpair, rdma_req); 1930 } 1931 1932 static inline int 1933 nvme_rdma_req_init(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req) 1934 { 1935 struct nvme_request *req = rdma_req->req; 1936 struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr; 1937 enum nvme_payload_type payload_type; 1938 bool icd_supported; 1939 int rc = -1; 1940 1941 payload_type = nvme_payload_type(&req->payload); 1942 /* 1943 * Check if icdoff is non zero, to avoid interop conflicts with 1944 * targets with non-zero icdoff. Both SPDK and the Linux kernel 1945 * targets use icdoff = 0. For targets with non-zero icdoff, we 1946 * will currently just not use inline data for now. 1947 */ 1948 icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER 1949 && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0; 1950 1951 if (spdk_unlikely(req->payload_size == 0)) { 1952 rc = nvme_rdma_build_null_request(rdma_req); 1953 } else if (payload_type == NVME_PAYLOAD_TYPE_CONTIG) { 1954 if (icd_supported) { 1955 rc = nvme_rdma_build_contig_inline_request(rqpair, rdma_req); 1956 } else { 1957 rc = nvme_rdma_build_contig_request(rqpair, rdma_req); 1958 } 1959 } else if (payload_type == NVME_PAYLOAD_TYPE_SGL) { 1960 if (icd_supported) { 1961 rc = nvme_rdma_build_sgl_inline_request(rqpair, rdma_req); 1962 } else { 1963 rc = nvme_rdma_build_sgl_request(rqpair, rdma_req); 1964 } 1965 } 1966 1967 if (spdk_unlikely(rc)) { 1968 return rc; 1969 } 1970 1971 memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd)); 1972 return 0; 1973 } 1974 1975 static struct spdk_nvme_qpair * 1976 nvme_rdma_ctrlr_create_qpair(struct spdk_nvme_ctrlr *ctrlr, 1977 uint16_t qid, uint32_t qsize, 1978 enum spdk_nvme_qprio qprio, 1979 uint32_t num_requests, 1980 bool delay_cmd_submit, 1981 bool async) 1982 { 1983 struct nvme_rdma_qpair *rqpair; 1984 struct spdk_nvme_qpair *qpair; 1985 int rc; 1986 1987 if (qsize < SPDK_NVME_QUEUE_MIN_ENTRIES) { 1988 SPDK_ERRLOG("Failed to create qpair with size %u. Minimum queue size is %d.\n", 1989 qsize, SPDK_NVME_QUEUE_MIN_ENTRIES); 1990 return NULL; 1991 } 1992 1993 rqpair = spdk_zmalloc(sizeof(struct nvme_rdma_qpair), 0, NULL, SPDK_ENV_NUMA_ID_ANY, 1994 SPDK_MALLOC_DMA); 1995 if (!rqpair) { 1996 SPDK_ERRLOG("failed to get create rqpair\n"); 1997 return NULL; 1998 } 1999 2000 /* Set num_entries one less than queue size. According to NVMe 2001 * and NVMe-oF specs we can not submit queue size requests, 2002 * one slot shall always remain empty. 2003 */ 2004 rqpair->num_entries = qsize - 1; 2005 rqpair->delay_cmd_submit = delay_cmd_submit; 2006 rqpair->state = NVME_RDMA_QPAIR_STATE_INVALID; 2007 qpair = &rqpair->qpair; 2008 rc = nvme_qpair_init(qpair, qid, ctrlr, qprio, num_requests, async); 2009 if (rc != 0) { 2010 spdk_free(rqpair); 2011 return NULL; 2012 } 2013 2014 return qpair; 2015 } 2016 2017 static void 2018 nvme_rdma_qpair_destroy(struct nvme_rdma_qpair *rqpair) 2019 { 2020 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2021 struct nvme_rdma_ctrlr *rctrlr; 2022 struct nvme_rdma_cm_event_entry *entry, *tmp; 2023 2024 spdk_rdma_utils_free_mem_map(&rqpair->mr_map); 2025 2026 if (rqpair->evt) { 2027 rdma_ack_cm_event(rqpair->evt); 2028 rqpair->evt = NULL; 2029 } 2030 2031 /* 2032 * This works because we have the controller lock both in 2033 * this function and in the function where we add new events. 2034 */ 2035 if (qpair->ctrlr != NULL) { 2036 rctrlr = nvme_rdma_ctrlr(qpair->ctrlr); 2037 STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) { 2038 if (entry->evt->id->context == rqpair) { 2039 STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link); 2040 rdma_ack_cm_event(entry->evt); 2041 STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link); 2042 } 2043 } 2044 } 2045 2046 if (rqpair->cm_id) { 2047 if (rqpair->rdma_qp) { 2048 spdk_rdma_utils_put_pd(rqpair->rdma_qp->qp->pd); 2049 spdk_rdma_provider_qp_destroy(rqpair->rdma_qp); 2050 rqpair->rdma_qp = NULL; 2051 } 2052 } 2053 2054 if (rqpair->poller) { 2055 struct nvme_rdma_poll_group *group; 2056 2057 assert(qpair->poll_group); 2058 group = nvme_rdma_poll_group(qpair->poll_group); 2059 2060 nvme_rdma_poll_group_put_poller(group, rqpair->poller); 2061 2062 rqpair->poller = NULL; 2063 rqpair->cq = NULL; 2064 if (rqpair->srq) { 2065 rqpair->srq = NULL; 2066 rqpair->rsps = NULL; 2067 } 2068 } else if (rqpair->cq) { 2069 ibv_destroy_cq(rqpair->cq); 2070 rqpair->cq = NULL; 2071 } 2072 2073 nvme_rdma_free_reqs(rqpair); 2074 nvme_rdma_free_rsps(rqpair->rsps); 2075 rqpair->rsps = NULL; 2076 2077 /* destroy cm_id last so cma device will not be freed before we destroy the cq. */ 2078 if (rqpair->cm_id) { 2079 rdma_destroy_id(rqpair->cm_id); 2080 rqpair->cm_id = NULL; 2081 } 2082 } 2083 2084 static void nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr); 2085 2086 static void 2087 nvme_rdma_qpair_flush_send_wrs(struct nvme_rdma_qpair *rqpair) 2088 { 2089 struct ibv_send_wr *bad_wr = NULL; 2090 int rc; 2091 2092 rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 2093 if (rc) { 2094 nvme_rdma_reset_failed_sends(rqpair, bad_wr); 2095 } 2096 } 2097 2098 static int 2099 nvme_rdma_qpair_disconnected(struct nvme_rdma_qpair *rqpair, int ret) 2100 { 2101 if (ret) { 2102 SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n"); 2103 goto quiet; 2104 } 2105 2106 if (rqpair->poller == NULL) { 2107 /* If poller is not used, cq is not shared. 2108 * So complete disconnecting qpair immediately. 2109 */ 2110 goto quiet; 2111 } 2112 2113 if (rqpair->rsps == NULL) { 2114 goto quiet; 2115 } 2116 2117 nvme_rdma_qpair_flush_send_wrs(rqpair); 2118 2119 if (rqpair->need_destroy || 2120 (rqpair->current_num_sends != 0 || 2121 (!rqpair->srq && rqpair->rsps->current_num_recvs != 0)) || 2122 ((rqpair->qpair.ctrlr->flags & SPDK_NVME_CTRLR_ACCEL_SEQUENCE_SUPPORTED) && 2123 (!TAILQ_EMPTY(&rqpair->outstanding_reqs)))) { 2124 rqpair->state = NVME_RDMA_QPAIR_STATE_LINGERING; 2125 rqpair->evt_timeout_ticks = (NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US * spdk_get_ticks_hz()) / 2126 SPDK_SEC_TO_USEC + spdk_get_ticks(); 2127 2128 return -EAGAIN; 2129 } 2130 2131 quiet: 2132 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED; 2133 2134 nvme_rdma_qpair_abort_reqs(&rqpair->qpair, rqpair->qpair.abort_dnr); 2135 nvme_rdma_qpair_destroy(rqpair); 2136 nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair); 2137 2138 return 0; 2139 } 2140 2141 static int 2142 nvme_rdma_qpair_wait_until_quiet(struct nvme_rdma_qpair *rqpair) 2143 { 2144 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2145 struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr; 2146 2147 if (spdk_get_ticks() < rqpair->evt_timeout_ticks && 2148 (rqpair->current_num_sends != 0 || 2149 (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) { 2150 return -EAGAIN; 2151 } 2152 2153 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED; 2154 nvme_rdma_qpair_abort_reqs(qpair, qpair->abort_dnr); 2155 if (!nvme_qpair_is_admin_queue(qpair)) { 2156 nvme_robust_mutex_lock(&ctrlr->ctrlr_lock); 2157 } 2158 nvme_rdma_qpair_destroy(rqpair); 2159 if (!nvme_qpair_is_admin_queue(qpair)) { 2160 nvme_robust_mutex_unlock(&ctrlr->ctrlr_lock); 2161 } 2162 nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair); 2163 2164 return 0; 2165 } 2166 2167 static void 2168 _nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair, 2169 nvme_rdma_cm_event_cb disconnected_qpair_cb) 2170 { 2171 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2172 int rc; 2173 2174 assert(disconnected_qpair_cb != NULL); 2175 2176 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITING; 2177 2178 if (rqpair->cm_id) { 2179 if (rqpair->rdma_qp) { 2180 rc = spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp); 2181 if ((qpair->ctrlr != NULL) && (rc == 0)) { 2182 rc = nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_DISCONNECTED, 2183 disconnected_qpair_cb); 2184 if (rc == 0) { 2185 return; 2186 } 2187 } 2188 } 2189 } 2190 2191 disconnected_qpair_cb(rqpair, 0); 2192 } 2193 2194 static int 2195 nvme_rdma_ctrlr_disconnect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 2196 { 2197 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2198 int rc; 2199 2200 switch (rqpair->state) { 2201 case NVME_RDMA_QPAIR_STATE_EXITING: 2202 if (!nvme_qpair_is_admin_queue(qpair)) { 2203 nvme_ctrlr_lock(ctrlr); 2204 } 2205 2206 rc = nvme_rdma_process_event_poll(rqpair); 2207 2208 if (!nvme_qpair_is_admin_queue(qpair)) { 2209 nvme_ctrlr_unlock(ctrlr); 2210 } 2211 break; 2212 2213 case NVME_RDMA_QPAIR_STATE_LINGERING: 2214 rc = nvme_rdma_qpair_wait_until_quiet(rqpair); 2215 break; 2216 case NVME_RDMA_QPAIR_STATE_EXITED: 2217 rc = 0; 2218 break; 2219 2220 default: 2221 assert(false); 2222 rc = -EAGAIN; 2223 break; 2224 } 2225 2226 return rc; 2227 } 2228 2229 static void 2230 nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 2231 { 2232 int rc; 2233 2234 _nvme_rdma_ctrlr_disconnect_qpair(ctrlr, qpair, nvme_rdma_qpair_disconnected); 2235 2236 /* If the async mode is disabled, poll the qpair until it is actually disconnected. 2237 * It is ensured that poll_group_process_completions() calls disconnected_qpair_cb 2238 * for any disconnected qpair. Hence, we do not have to check if the qpair is in 2239 * a poll group or not. 2240 * At the same time, if the qpair is being destroyed, i.e. this function is called by 2241 * spdk_nvme_ctrlr_free_io_qpair then we need to wait until qpair is disconnected, otherwise 2242 * we may leak some resources. 2243 */ 2244 if (qpair->async && !qpair->destroy_in_progress) { 2245 return; 2246 } 2247 2248 while (1) { 2249 rc = nvme_rdma_ctrlr_disconnect_qpair_poll(ctrlr, qpair); 2250 if (rc != -EAGAIN) { 2251 break; 2252 } 2253 } 2254 } 2255 2256 static int 2257 nvme_rdma_stale_conn_disconnected(struct nvme_rdma_qpair *rqpair, int ret) 2258 { 2259 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2260 2261 if (ret) { 2262 SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n"); 2263 } 2264 2265 nvme_rdma_qpair_destroy(rqpair); 2266 2267 qpair->last_transport_failure_reason = qpair->transport_failure_reason; 2268 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_NONE; 2269 2270 rqpair->state = NVME_RDMA_QPAIR_STATE_STALE_CONN; 2271 rqpair->evt_timeout_ticks = (NVME_RDMA_STALE_CONN_RETRY_DELAY_US * spdk_get_ticks_hz()) / 2272 SPDK_SEC_TO_USEC + spdk_get_ticks(); 2273 2274 return 0; 2275 } 2276 2277 static int 2278 nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair) 2279 { 2280 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2281 2282 if (rqpair->stale_conn_retry_count >= NVME_RDMA_STALE_CONN_RETRY_MAX) { 2283 SPDK_ERRLOG("Retry failed %d times, give up stale connection to qpair (cntlid:%u, qid:%u).\n", 2284 NVME_RDMA_STALE_CONN_RETRY_MAX, qpair->ctrlr->cntlid, qpair->id); 2285 return -ESTALE; 2286 } 2287 2288 rqpair->stale_conn_retry_count++; 2289 2290 SPDK_NOTICELOG("%d times, retry stale connection to qpair (cntlid:%u, qid:%u).\n", 2291 rqpair->stale_conn_retry_count, qpair->ctrlr->cntlid, qpair->id); 2292 2293 _nvme_rdma_ctrlr_disconnect_qpair(qpair->ctrlr, qpair, nvme_rdma_stale_conn_disconnected); 2294 2295 return 0; 2296 } 2297 2298 static int 2299 nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 2300 { 2301 struct nvme_rdma_qpair *rqpair; 2302 2303 assert(qpair != NULL); 2304 rqpair = nvme_rdma_qpair(qpair); 2305 2306 if (rqpair->state != NVME_RDMA_QPAIR_STATE_EXITED) { 2307 int rc __attribute__((unused)); 2308 2309 /* qpair was removed from the poll group while the disconnect is not finished. 2310 * Destroy rdma resources forcefully. */ 2311 rc = nvme_rdma_qpair_disconnected(rqpair, 0); 2312 assert(rc == 0); 2313 } 2314 2315 nvme_rdma_qpair_abort_reqs(qpair, qpair->abort_dnr); 2316 nvme_qpair_deinit(qpair); 2317 2318 spdk_free(rqpair); 2319 2320 return 0; 2321 } 2322 2323 static struct spdk_nvme_qpair * 2324 nvme_rdma_ctrlr_create_io_qpair(struct spdk_nvme_ctrlr *ctrlr, uint16_t qid, 2325 const struct spdk_nvme_io_qpair_opts *opts) 2326 { 2327 return nvme_rdma_ctrlr_create_qpair(ctrlr, qid, opts->io_queue_size, opts->qprio, 2328 opts->io_queue_requests, 2329 opts->delay_cmd_submit, 2330 opts->async_mode); 2331 } 2332 2333 static int 2334 nvme_rdma_ctrlr_enable(struct spdk_nvme_ctrlr *ctrlr) 2335 { 2336 /* do nothing here */ 2337 return 0; 2338 } 2339 2340 static int nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr); 2341 2342 /* We have to use the typedef in the function declaration to appease astyle. */ 2343 typedef struct spdk_nvme_ctrlr spdk_nvme_ctrlr_t; 2344 2345 static spdk_nvme_ctrlr_t * 2346 nvme_rdma_ctrlr_construct(const struct spdk_nvme_transport_id *trid, 2347 const struct spdk_nvme_ctrlr_opts *opts, 2348 void *devhandle) 2349 { 2350 struct nvme_rdma_ctrlr *rctrlr; 2351 struct ibv_context **contexts; 2352 struct ibv_device_attr dev_attr; 2353 int i, flag, rc; 2354 2355 rctrlr = spdk_zmalloc(sizeof(struct nvme_rdma_ctrlr), 0, NULL, SPDK_ENV_NUMA_ID_ANY, 2356 SPDK_MALLOC_DMA); 2357 if (rctrlr == NULL) { 2358 SPDK_ERRLOG("could not allocate ctrlr\n"); 2359 return NULL; 2360 } 2361 2362 rctrlr->ctrlr.opts = *opts; 2363 rctrlr->ctrlr.trid = *trid; 2364 2365 if (opts->transport_retry_count > NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT) { 2366 SPDK_NOTICELOG("transport_retry_count exceeds max value %d, use max value\n", 2367 NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT); 2368 rctrlr->ctrlr.opts.transport_retry_count = NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT; 2369 } 2370 2371 if (opts->transport_ack_timeout > NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT) { 2372 SPDK_NOTICELOG("transport_ack_timeout exceeds max value %d, use max value\n", 2373 NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT); 2374 rctrlr->ctrlr.opts.transport_ack_timeout = NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT; 2375 } 2376 2377 contexts = rdma_get_devices(NULL); 2378 if (contexts == NULL) { 2379 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2380 spdk_free(rctrlr); 2381 return NULL; 2382 } 2383 2384 i = 0; 2385 rctrlr->max_sge = NVME_RDMA_MAX_SGL_DESCRIPTORS; 2386 2387 while (contexts[i] != NULL) { 2388 rc = ibv_query_device(contexts[i], &dev_attr); 2389 if (rc < 0) { 2390 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2391 rdma_free_devices(contexts); 2392 spdk_free(rctrlr); 2393 return NULL; 2394 } 2395 rctrlr->max_sge = spdk_min(rctrlr->max_sge, (uint16_t)dev_attr.max_sge); 2396 i++; 2397 } 2398 2399 rdma_free_devices(contexts); 2400 2401 rc = nvme_ctrlr_construct(&rctrlr->ctrlr); 2402 if (rc != 0) { 2403 spdk_free(rctrlr); 2404 return NULL; 2405 } 2406 2407 STAILQ_INIT(&rctrlr->pending_cm_events); 2408 STAILQ_INIT(&rctrlr->free_cm_events); 2409 rctrlr->cm_events = spdk_zmalloc(NVME_RDMA_NUM_CM_EVENTS * sizeof(*rctrlr->cm_events), 0, NULL, 2410 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 2411 if (rctrlr->cm_events == NULL) { 2412 SPDK_ERRLOG("unable to allocate buffers to hold CM events.\n"); 2413 goto destruct_ctrlr; 2414 } 2415 2416 for (i = 0; i < NVME_RDMA_NUM_CM_EVENTS; i++) { 2417 STAILQ_INSERT_TAIL(&rctrlr->free_cm_events, &rctrlr->cm_events[i], link); 2418 } 2419 2420 rctrlr->cm_channel = rdma_create_event_channel(); 2421 if (rctrlr->cm_channel == NULL) { 2422 SPDK_ERRLOG("rdma_create_event_channel() failed\n"); 2423 goto destruct_ctrlr; 2424 } 2425 2426 flag = fcntl(rctrlr->cm_channel->fd, F_GETFL); 2427 if (fcntl(rctrlr->cm_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2428 SPDK_ERRLOG("Cannot set event channel to non blocking\n"); 2429 goto destruct_ctrlr; 2430 } 2431 2432 rctrlr->ctrlr.adminq = nvme_rdma_ctrlr_create_qpair(&rctrlr->ctrlr, 0, 2433 rctrlr->ctrlr.opts.admin_queue_size, 0, 2434 rctrlr->ctrlr.opts.admin_queue_size, false, true); 2435 if (!rctrlr->ctrlr.adminq) { 2436 SPDK_ERRLOG("failed to create admin qpair\n"); 2437 goto destruct_ctrlr; 2438 } 2439 if (spdk_rdma_provider_accel_sequence_supported()) { 2440 rctrlr->ctrlr.flags |= SPDK_NVME_CTRLR_ACCEL_SEQUENCE_SUPPORTED; 2441 } 2442 2443 if (nvme_ctrlr_add_process(&rctrlr->ctrlr, 0) != 0) { 2444 SPDK_ERRLOG("nvme_ctrlr_add_process() failed\n"); 2445 goto destruct_ctrlr; 2446 } 2447 2448 SPDK_DEBUGLOG(nvme, "successfully initialized the nvmf ctrlr\n"); 2449 return &rctrlr->ctrlr; 2450 2451 destruct_ctrlr: 2452 nvme_ctrlr_destruct(&rctrlr->ctrlr); 2453 return NULL; 2454 } 2455 2456 static int 2457 nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr) 2458 { 2459 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr); 2460 struct nvme_rdma_cm_event_entry *entry; 2461 2462 if (ctrlr->adminq) { 2463 nvme_rdma_ctrlr_delete_io_qpair(ctrlr, ctrlr->adminq); 2464 } 2465 2466 STAILQ_FOREACH(entry, &rctrlr->pending_cm_events, link) { 2467 rdma_ack_cm_event(entry->evt); 2468 } 2469 2470 STAILQ_INIT(&rctrlr->free_cm_events); 2471 STAILQ_INIT(&rctrlr->pending_cm_events); 2472 spdk_free(rctrlr->cm_events); 2473 2474 if (rctrlr->cm_channel) { 2475 rdma_destroy_event_channel(rctrlr->cm_channel); 2476 rctrlr->cm_channel = NULL; 2477 } 2478 2479 nvme_ctrlr_destruct_finish(ctrlr); 2480 2481 spdk_free(rctrlr); 2482 2483 return 0; 2484 } 2485 2486 static inline int 2487 _nvme_rdma_qpair_submit_request(struct nvme_rdma_qpair *rqpair, 2488 struct spdk_nvme_rdma_req *rdma_req) 2489 { 2490 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2491 struct ibv_send_wr *wr; 2492 struct nvme_rdma_poll_group *group; 2493 2494 if (!rqpair->link_active.tqe_prev && qpair->poll_group) { 2495 group = nvme_rdma_poll_group(qpair->poll_group); 2496 TAILQ_INSERT_TAIL(&group->active_qpairs, rqpair, link_active); 2497 } 2498 assert(rqpair->current_num_sends < rqpair->num_entries); 2499 rqpair->current_num_sends++; 2500 2501 wr = &rdma_req->send_wr; 2502 wr->next = NULL; 2503 nvme_rdma_trace_ibv_sge(wr->sg_list); 2504 2505 spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, wr); 2506 2507 if (!rqpair->delay_cmd_submit) { 2508 return nvme_rdma_qpair_submit_sends(rqpair); 2509 } 2510 2511 return 0; 2512 } 2513 2514 static int 2515 nvme_rdma_qpair_submit_request(struct spdk_nvme_qpair *qpair, 2516 struct nvme_request *req) 2517 { 2518 struct nvme_rdma_qpair *rqpair; 2519 struct spdk_nvme_rdma_req *rdma_req; 2520 int rc; 2521 2522 rqpair = nvme_rdma_qpair(qpair); 2523 assert(rqpair != NULL); 2524 assert(req != NULL); 2525 2526 rdma_req = nvme_rdma_req_get(rqpair); 2527 if (spdk_unlikely(!rdma_req)) { 2528 if (rqpair->poller) { 2529 rqpair->poller->stats.queued_requests++; 2530 } 2531 /* Inform the upper layer to try again later. */ 2532 return -EAGAIN; 2533 } 2534 2535 assert(rdma_req->req == NULL); 2536 rdma_req->req = req; 2537 req->cmd.cid = rdma_req->id; 2538 if (req->accel_sequence) { 2539 assert(spdk_rdma_provider_accel_sequence_supported()); 2540 assert(rqpair->qpair.poll_group->group); 2541 assert(rqpair->qpair.poll_group->group->accel_fn_table.append_copy); 2542 assert(rqpair->qpair.poll_group->group->accel_fn_table.reverse_sequence); 2543 assert(rqpair->qpair.poll_group->group->accel_fn_table.finish_sequence); 2544 2545 rc = nvme_rdma_apply_accel_sequence(rqpair, req, rdma_req); 2546 if (spdk_unlikely(rc)) { 2547 SPDK_ERRLOG("failed to apply accel seq, rqpair %p, req %p, rc %d\n", rqpair, rdma_req, rc); 2548 nvme_rdma_req_put(rqpair, rdma_req); 2549 return rc; 2550 } 2551 /* Capsule will be sent in data_transfer callback */ 2552 return 0; 2553 } 2554 2555 rc = nvme_rdma_req_init(rqpair, rdma_req); 2556 if (spdk_unlikely(rc)) { 2557 SPDK_ERRLOG("nvme_rdma_req_init() failed\n"); 2558 nvme_rdma_req_put(rqpair, rdma_req); 2559 return -1; 2560 } 2561 2562 TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link); 2563 rqpair->num_outstanding_reqs++; 2564 2565 return _nvme_rdma_qpair_submit_request(rqpair, rdma_req); 2566 } 2567 2568 static int 2569 nvme_rdma_qpair_reset(struct spdk_nvme_qpair *qpair) 2570 { 2571 /* Currently, doing nothing here */ 2572 return 0; 2573 } 2574 2575 static void 2576 nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr) 2577 { 2578 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2579 struct spdk_nvme_cpl cpl; 2580 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2581 2582 cpl.sqid = qpair->id; 2583 cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION; 2584 cpl.status.sct = SPDK_NVME_SCT_GENERIC; 2585 cpl.status.dnr = dnr; 2586 2587 /* 2588 * We cannot abort requests at the RDMA layer without 2589 * unregistering them. If we do, we can still get error 2590 * free completions on the shared completion queue. 2591 */ 2592 if (nvme_qpair_get_state(qpair) > NVME_QPAIR_DISCONNECTING && 2593 nvme_qpair_get_state(qpair) != NVME_QPAIR_DESTROYING) { 2594 nvme_ctrlr_disconnect_qpair(qpair); 2595 } 2596 2597 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2598 if (rdma_req->in_progress_accel) { 2599 /* We should wait for accel completion */ 2600 continue; 2601 } 2602 nvme_rdma_req_complete(rdma_req, &cpl, true); 2603 } 2604 } 2605 2606 static void 2607 nvme_rdma_qpair_check_timeout(struct spdk_nvme_qpair *qpair) 2608 { 2609 uint64_t t02; 2610 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2611 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2612 struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr; 2613 struct spdk_nvme_ctrlr_process *active_proc; 2614 2615 /* Don't check timeouts during controller initialization. */ 2616 if (ctrlr->state != NVME_CTRLR_STATE_READY) { 2617 return; 2618 } 2619 2620 if (nvme_qpair_is_admin_queue(qpair)) { 2621 active_proc = nvme_ctrlr_get_current_process(ctrlr); 2622 } else { 2623 active_proc = qpair->active_proc; 2624 } 2625 2626 /* Only check timeouts if the current process has a timeout callback. */ 2627 if (active_proc == NULL || active_proc->timeout_cb_fn == NULL) { 2628 return; 2629 } 2630 2631 t02 = spdk_get_ticks(); 2632 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2633 assert(rdma_req->req != NULL); 2634 2635 if (nvme_request_check_timeout(rdma_req->req, rdma_req->id, active_proc, t02)) { 2636 /* 2637 * The requests are in order, so as soon as one has not timed out, 2638 * stop iterating. 2639 */ 2640 break; 2641 } 2642 } 2643 } 2644 2645 static inline void 2646 nvme_rdma_request_ready(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req) 2647 { 2648 struct spdk_nvme_rdma_rsp *rdma_rsp = rdma_req->rdma_rsp; 2649 struct ibv_recv_wr *recv_wr = rdma_rsp->recv_wr; 2650 2651 if (rdma_req->transfer_cpl_cb) { 2652 int rc = 0; 2653 2654 if (spdk_unlikely(spdk_nvme_cpl_is_error(&rdma_rsp->cpl))) { 2655 SPDK_WARNLOG("req %p, error cpl sct %d, sc %d\n", rdma_req, rdma_rsp->cpl.status.sct, 2656 rdma_rsp->cpl.status.sc); 2657 rc = -EIO; 2658 } 2659 nvme_rdma_finish_data_transfer(rdma_req, rc); 2660 } else { 2661 nvme_rdma_req_complete(rdma_req, &rdma_rsp->cpl, true); 2662 } 2663 2664 if (spdk_unlikely(rqpair->state >= NVME_RDMA_QPAIR_STATE_EXITING && !rqpair->srq)) { 2665 /* Skip posting back recv wr if we are in a disconnection process. We may never get 2666 * a WC and we may end up stuck in LINGERING state until the timeout. */ 2667 return; 2668 } 2669 2670 assert(rqpair->rsps->current_num_recvs < rqpair->rsps->num_entries); 2671 rqpair->rsps->current_num_recvs++; 2672 2673 recv_wr->next = NULL; 2674 nvme_rdma_trace_ibv_sge(recv_wr->sg_list); 2675 2676 if (!rqpair->srq) { 2677 spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, recv_wr); 2678 } else { 2679 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, recv_wr); 2680 } 2681 } 2682 2683 #define MAX_COMPLETIONS_PER_POLL 128 2684 2685 static void 2686 nvme_rdma_fail_qpair(struct spdk_nvme_qpair *qpair, int failure_reason) 2687 { 2688 if (failure_reason == IBV_WC_RETRY_EXC_ERR) { 2689 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE; 2690 } else if (qpair->transport_failure_reason == SPDK_NVME_QPAIR_FAILURE_NONE) { 2691 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_UNKNOWN; 2692 } 2693 2694 nvme_ctrlr_disconnect_qpair(qpair); 2695 } 2696 2697 static struct nvme_rdma_qpair * 2698 get_rdma_qpair_from_wc(struct nvme_rdma_poll_group *group, struct ibv_wc *wc) 2699 { 2700 struct spdk_nvme_qpair *qpair; 2701 struct nvme_rdma_qpair *rqpair; 2702 2703 STAILQ_FOREACH(qpair, &group->group.connected_qpairs, poll_group_stailq) { 2704 rqpair = nvme_rdma_qpair(qpair); 2705 if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) { 2706 return rqpair; 2707 } 2708 } 2709 2710 STAILQ_FOREACH(qpair, &group->group.disconnected_qpairs, poll_group_stailq) { 2711 rqpair = nvme_rdma_qpair(qpair); 2712 if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) { 2713 return rqpair; 2714 } 2715 } 2716 2717 return NULL; 2718 } 2719 2720 static inline void 2721 nvme_rdma_log_wc_status(struct nvme_rdma_qpair *rqpair, struct ibv_wc *wc) 2722 { 2723 struct nvme_rdma_wr *rdma_wr = (struct nvme_rdma_wr *)wc->wr_id; 2724 2725 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 2726 /* If qpair is in ERR state, we will receive completions for all posted and not completed 2727 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 2728 SPDK_DEBUGLOG(nvme, "WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n", 2729 rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status, 2730 ibv_wc_status_str(wc->status)); 2731 } else { 2732 SPDK_ERRLOG("WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n", 2733 rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status, 2734 ibv_wc_status_str(wc->status)); 2735 } 2736 } 2737 2738 static inline int 2739 nvme_rdma_process_recv_completion(struct nvme_rdma_poller *poller, struct ibv_wc *wc, 2740 struct nvme_rdma_wr *rdma_wr) 2741 { 2742 struct nvme_rdma_qpair *rqpair; 2743 struct spdk_nvme_rdma_req *rdma_req; 2744 struct spdk_nvme_rdma_rsp *rdma_rsp; 2745 2746 rdma_rsp = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_rsp, rdma_wr); 2747 2748 if (poller && poller->srq) { 2749 rqpair = get_rdma_qpair_from_wc(poller->group, wc); 2750 if (spdk_unlikely(!rqpair)) { 2751 /* Since we do not handle the LAST_WQE_REACHED event, we do not know when 2752 * a Receive Queue in a QP, that is associated with an SRQ, is flushed. 2753 * We may get a WC for a already destroyed QP. 2754 * 2755 * However, for the SRQ, this is not any error. Hence, just re-post the 2756 * receive request to the SRQ to reuse for other QPs, and return 0. 2757 */ 2758 spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr); 2759 return 0; 2760 } 2761 } else { 2762 rqpair = rdma_rsp->rqpair; 2763 if (spdk_unlikely(!rqpair)) { 2764 /* TODO: Fix forceful QP destroy when it is not async mode. 2765 * CQ itself did not cause any error. Hence, return 0 for now. 2766 */ 2767 SPDK_WARNLOG("QP might be already destroyed.\n"); 2768 return 0; 2769 } 2770 } 2771 2772 2773 assert(rqpair->rsps->current_num_recvs > 0); 2774 rqpair->rsps->current_num_recvs--; 2775 2776 if (spdk_unlikely(wc->status)) { 2777 nvme_rdma_log_wc_status(rqpair, wc); 2778 goto err_wc; 2779 } 2780 2781 SPDK_DEBUGLOG(nvme, "CQ recv completion\n"); 2782 2783 if (spdk_unlikely(wc->byte_len < sizeof(struct spdk_nvme_cpl))) { 2784 SPDK_ERRLOG("recv length %u less than expected response size\n", wc->byte_len); 2785 goto err_wc; 2786 } 2787 rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid]; 2788 rdma_req->completion_flags |= NVME_RDMA_RECV_COMPLETED; 2789 rdma_req->rdma_rsp = rdma_rsp; 2790 2791 if ((rdma_req->completion_flags & NVME_RDMA_SEND_COMPLETED) == 0) { 2792 return 0; 2793 } 2794 2795 rqpair->num_completions++; 2796 2797 nvme_rdma_request_ready(rqpair, rdma_req); 2798 2799 if (!rqpair->delay_cmd_submit) { 2800 if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) { 2801 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 2802 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2803 return -ENXIO; 2804 } 2805 } 2806 2807 return 1; 2808 2809 err_wc: 2810 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2811 if (poller && poller->srq) { 2812 spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr); 2813 } 2814 rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid]; 2815 if (rdma_req->transfer_cpl_cb) { 2816 nvme_rdma_finish_data_transfer(rdma_req, -ENXIO); 2817 } 2818 return -ENXIO; 2819 } 2820 2821 static inline int 2822 nvme_rdma_process_send_completion(struct nvme_rdma_poller *poller, 2823 struct nvme_rdma_qpair *rdma_qpair, 2824 struct ibv_wc *wc, struct nvme_rdma_wr *rdma_wr) 2825 { 2826 struct nvme_rdma_qpair *rqpair; 2827 struct spdk_nvme_rdma_req *rdma_req; 2828 2829 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_req, rdma_wr); 2830 rqpair = rdma_req->req ? nvme_rdma_qpair(rdma_req->req->qpair) : NULL; 2831 if (spdk_unlikely(!rqpair)) { 2832 rqpair = rdma_qpair != NULL ? rdma_qpair : get_rdma_qpair_from_wc(poller->group, wc); 2833 } 2834 2835 /* If we are flushing I/O */ 2836 if (spdk_unlikely(wc->status)) { 2837 if (!rqpair) { 2838 /* When poll_group is used, several qpairs share the same CQ and it is possible to 2839 * receive a completion with error (e.g. IBV_WC_WR_FLUSH_ERR) for already disconnected qpair 2840 * That happens due to qpair is destroyed while there are submitted but not completed send/receive 2841 * Work Requests */ 2842 assert(poller); 2843 return 0; 2844 } 2845 assert(rqpair->current_num_sends > 0); 2846 rqpair->current_num_sends--; 2847 nvme_rdma_log_wc_status(rqpair, wc); 2848 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2849 if (rdma_req->rdma_rsp && poller && poller->srq) { 2850 spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_req->rdma_rsp->recv_wr); 2851 } 2852 if (rdma_req->transfer_cpl_cb) { 2853 nvme_rdma_finish_data_transfer(rdma_req, -ENXIO); 2854 } 2855 return -ENXIO; 2856 } 2857 2858 /* We do not support Soft Roce anymore. Other than Soft Roce's bug, we should not 2859 * receive a completion without error status after qpair is disconnected/destroyed. 2860 */ 2861 if (spdk_unlikely(rdma_req->req == NULL)) { 2862 /* 2863 * Some infiniband drivers do not guarantee the previous assumption after we 2864 * received a RDMA_CM_EVENT_DEVICE_REMOVAL event. 2865 */ 2866 SPDK_ERRLOG("Received malformed completion: request 0x%"PRIx64" type %d\n", wc->wr_id, 2867 rdma_wr->type); 2868 if (!rqpair || !rqpair->need_destroy) { 2869 assert(0); 2870 } 2871 return -ENXIO; 2872 } 2873 2874 rdma_req->completion_flags |= NVME_RDMA_SEND_COMPLETED; 2875 assert(rqpair->current_num_sends > 0); 2876 rqpair->current_num_sends--; 2877 2878 if ((rdma_req->completion_flags & NVME_RDMA_RECV_COMPLETED) == 0) { 2879 return 0; 2880 } 2881 2882 rqpair->num_completions++; 2883 2884 nvme_rdma_request_ready(rqpair, rdma_req); 2885 2886 if (!rqpair->delay_cmd_submit) { 2887 if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) { 2888 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 2889 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2890 return -ENXIO; 2891 } 2892 } 2893 2894 return 1; 2895 } 2896 2897 static inline int 2898 nvme_rdma_cq_process_completions(struct ibv_cq *cq, uint32_t batch_size, 2899 struct nvme_rdma_poller *poller, 2900 struct nvme_rdma_qpair *rdma_qpair, 2901 uint64_t *rdma_completions) 2902 { 2903 struct ibv_wc wc[MAX_COMPLETIONS_PER_POLL]; 2904 struct nvme_rdma_wr *rdma_wr; 2905 uint32_t reaped = 0; 2906 int completion_rc = 0; 2907 int rc, _rc, i; 2908 2909 rc = ibv_poll_cq(cq, batch_size, wc); 2910 if (spdk_unlikely(rc < 0)) { 2911 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2912 errno, spdk_strerror(errno)); 2913 return -ECANCELED; 2914 } else if (rc == 0) { 2915 return 0; 2916 } 2917 2918 for (i = 0; i < rc; i++) { 2919 rdma_wr = (struct nvme_rdma_wr *)wc[i].wr_id; 2920 switch (rdma_wr->type) { 2921 case RDMA_WR_TYPE_RECV: 2922 _rc = nvme_rdma_process_recv_completion(poller, &wc[i], rdma_wr); 2923 break; 2924 2925 case RDMA_WR_TYPE_SEND: 2926 _rc = nvme_rdma_process_send_completion(poller, rdma_qpair, &wc[i], rdma_wr); 2927 break; 2928 2929 default: 2930 SPDK_ERRLOG("Received an unexpected opcode on the CQ: %d\n", rdma_wr->type); 2931 return -ECANCELED; 2932 } 2933 if (spdk_likely(_rc >= 0)) { 2934 reaped += _rc; 2935 } else { 2936 completion_rc = _rc; 2937 } 2938 } 2939 2940 *rdma_completions += rc; 2941 2942 if (spdk_unlikely(completion_rc)) { 2943 return completion_rc; 2944 } 2945 2946 return reaped; 2947 } 2948 2949 static void 2950 dummy_disconnected_qpair_cb(struct spdk_nvme_qpair *qpair, void *poll_group_ctx) 2951 { 2952 2953 } 2954 2955 static int 2956 nvme_rdma_qpair_process_completions(struct spdk_nvme_qpair *qpair, 2957 uint32_t max_completions) 2958 { 2959 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2960 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(qpair->ctrlr); 2961 int rc = 0, batch_size; 2962 struct ibv_cq *cq; 2963 uint64_t rdma_completions = 0; 2964 2965 /* 2966 * This is used during the connection phase. It's possible that we are still reaping error completions 2967 * from other qpairs so we need to call the poll group function. Also, it's more correct since the cq 2968 * is shared. 2969 */ 2970 if (qpair->poll_group != NULL) { 2971 return spdk_nvme_poll_group_process_completions(qpair->poll_group->group, max_completions, 2972 dummy_disconnected_qpair_cb); 2973 } 2974 2975 if (max_completions == 0) { 2976 max_completions = rqpair->num_entries; 2977 } else { 2978 max_completions = spdk_min(max_completions, rqpair->num_entries); 2979 } 2980 2981 switch (nvme_qpair_get_state(qpair)) { 2982 case NVME_QPAIR_CONNECTING: 2983 rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair); 2984 if (rc == 0) { 2985 /* Once the connection is completed, we can submit queued requests */ 2986 nvme_qpair_resubmit_requests(qpair, rqpair->num_entries); 2987 } else if (rc != -EAGAIN) { 2988 SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair); 2989 goto failed; 2990 } else if (rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING) { 2991 return 0; 2992 } 2993 break; 2994 2995 case NVME_QPAIR_DISCONNECTING: 2996 nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair); 2997 return -ENXIO; 2998 2999 default: 3000 if (nvme_qpair_is_admin_queue(qpair)) { 3001 nvme_rdma_poll_events(rctrlr); 3002 } 3003 nvme_rdma_qpair_process_cm_event(rqpair); 3004 break; 3005 } 3006 3007 if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) { 3008 goto failed; 3009 } 3010 3011 cq = rqpair->cq; 3012 3013 rqpair->num_completions = 0; 3014 do { 3015 batch_size = spdk_min((max_completions - rqpair->num_completions), MAX_COMPLETIONS_PER_POLL); 3016 rc = nvme_rdma_cq_process_completions(cq, batch_size, NULL, rqpair, &rdma_completions); 3017 3018 if (rc == 0) { 3019 break; 3020 /* Handle the case where we fail to poll the cq. */ 3021 } else if (rc == -ECANCELED) { 3022 goto failed; 3023 } else if (rc == -ENXIO) { 3024 return rc; 3025 } 3026 } while (rqpair->num_completions < max_completions); 3027 3028 if (spdk_unlikely(nvme_rdma_qpair_submit_sends(rqpair) || 3029 nvme_rdma_qpair_submit_recvs(rqpair))) { 3030 goto failed; 3031 } 3032 3033 if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) { 3034 nvme_rdma_qpair_check_timeout(qpair); 3035 } 3036 3037 return rqpair->num_completions; 3038 3039 failed: 3040 nvme_rdma_fail_qpair(qpair, 0); 3041 return -ENXIO; 3042 } 3043 3044 static uint32_t 3045 nvme_rdma_ctrlr_get_max_xfer_size(struct spdk_nvme_ctrlr *ctrlr) 3046 { 3047 /* max_mr_size by ibv_query_device indicates the largest value that we can 3048 * set for a registered memory region. It is independent from the actual 3049 * I/O size and is very likely to be larger than 2 MiB which is the 3050 * granularity we currently register memory regions. Hence return 3051 * UINT32_MAX here and let the generic layer use the controller data to 3052 * moderate this value. 3053 */ 3054 return UINT32_MAX; 3055 } 3056 3057 static uint16_t 3058 nvme_rdma_ctrlr_get_max_sges(struct spdk_nvme_ctrlr *ctrlr) 3059 { 3060 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr); 3061 uint32_t max_sge = rctrlr->max_sge; 3062 uint32_t max_in_capsule_sge = (ctrlr->cdata.nvmf_specific.ioccsz * 16 - 3063 sizeof(struct spdk_nvme_cmd)) / 3064 sizeof(struct spdk_nvme_sgl_descriptor); 3065 3066 /* Max SGE is limited by capsule size */ 3067 max_sge = spdk_min(max_sge, max_in_capsule_sge); 3068 /* Max SGE may be limited by MSDBD */ 3069 if (ctrlr->cdata.nvmf_specific.msdbd != 0) { 3070 max_sge = spdk_min(max_sge, ctrlr->cdata.nvmf_specific.msdbd); 3071 } 3072 3073 /* Max SGE can't be less than 1 */ 3074 max_sge = spdk_max(1, max_sge); 3075 return max_sge; 3076 } 3077 3078 static int 3079 nvme_rdma_qpair_iterate_requests(struct spdk_nvme_qpair *qpair, 3080 int (*iter_fn)(struct nvme_request *req, void *arg), 3081 void *arg) 3082 { 3083 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3084 struct spdk_nvme_rdma_req *rdma_req, *tmp; 3085 int rc; 3086 3087 assert(iter_fn != NULL); 3088 3089 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 3090 assert(rdma_req->req != NULL); 3091 3092 rc = iter_fn(rdma_req->req, arg); 3093 if (rc != 0) { 3094 return rc; 3095 } 3096 } 3097 3098 return 0; 3099 } 3100 3101 static int 3102 nvme_rdma_qpair_authenticate(struct spdk_nvme_qpair *qpair) 3103 { 3104 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3105 int rc; 3106 3107 /* If the qpair is still connecting, it'll be forced to authenticate later on */ 3108 if (rqpair->state < NVME_RDMA_QPAIR_STATE_RUNNING) { 3109 return 0; 3110 } else if (rqpair->state != NVME_RDMA_QPAIR_STATE_RUNNING) { 3111 return -ENOTCONN; 3112 } 3113 3114 rc = nvme_fabric_qpair_authenticate_async(qpair); 3115 if (rc == 0) { 3116 nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTING); 3117 rqpair->state = NVME_RDMA_QPAIR_STATE_AUTHENTICATING; 3118 } 3119 3120 return rc; 3121 } 3122 3123 static void 3124 nvme_rdma_admin_qpair_abort_aers(struct spdk_nvme_qpair *qpair) 3125 { 3126 struct spdk_nvme_rdma_req *rdma_req, *tmp; 3127 struct spdk_nvme_cpl cpl; 3128 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3129 3130 cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION; 3131 cpl.status.sct = SPDK_NVME_SCT_GENERIC; 3132 3133 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 3134 assert(rdma_req->req != NULL); 3135 3136 if (rdma_req->req->cmd.opc != SPDK_NVME_OPC_ASYNC_EVENT_REQUEST) { 3137 continue; 3138 } 3139 3140 nvme_rdma_req_complete(rdma_req, &cpl, false); 3141 } 3142 } 3143 3144 static void 3145 nvme_rdma_poller_destroy(struct nvme_rdma_poller *poller) 3146 { 3147 if (poller->cq) { 3148 ibv_destroy_cq(poller->cq); 3149 } 3150 if (poller->rsps) { 3151 nvme_rdma_free_rsps(poller->rsps); 3152 } 3153 if (poller->srq) { 3154 spdk_rdma_provider_srq_destroy(poller->srq); 3155 } 3156 if (poller->mr_map) { 3157 spdk_rdma_utils_free_mem_map(&poller->mr_map); 3158 } 3159 if (poller->pd) { 3160 spdk_rdma_utils_put_pd(poller->pd); 3161 } 3162 free(poller); 3163 } 3164 3165 static struct nvme_rdma_poller * 3166 nvme_rdma_poller_create(struct nvme_rdma_poll_group *group, struct ibv_context *ctx) 3167 { 3168 struct nvme_rdma_poller *poller; 3169 struct ibv_device_attr dev_attr; 3170 struct spdk_rdma_provider_srq_init_attr srq_init_attr = {}; 3171 struct nvme_rdma_rsp_opts opts; 3172 int num_cqe, max_num_cqe; 3173 int rc; 3174 3175 poller = calloc(1, sizeof(*poller)); 3176 if (poller == NULL) { 3177 SPDK_ERRLOG("Unable to allocate poller.\n"); 3178 return NULL; 3179 } 3180 3181 poller->group = group; 3182 poller->device = ctx; 3183 3184 if (g_spdk_nvme_transport_opts.rdma_srq_size != 0) { 3185 rc = ibv_query_device(ctx, &dev_attr); 3186 if (rc) { 3187 SPDK_ERRLOG("Unable to query RDMA device.\n"); 3188 goto fail; 3189 } 3190 3191 poller->pd = spdk_rdma_utils_get_pd(ctx); 3192 if (poller->pd == NULL) { 3193 SPDK_ERRLOG("Unable to get PD.\n"); 3194 goto fail; 3195 } 3196 3197 poller->mr_map = spdk_rdma_utils_create_mem_map(poller->pd, &g_nvme_hooks, 3198 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE); 3199 if (poller->mr_map == NULL) { 3200 SPDK_ERRLOG("Unable to create memory map.\n"); 3201 goto fail; 3202 } 3203 3204 srq_init_attr.stats = &poller->stats.rdma_stats.recv; 3205 srq_init_attr.pd = poller->pd; 3206 srq_init_attr.srq_init_attr.attr.max_wr = spdk_min((uint32_t)dev_attr.max_srq_wr, 3207 g_spdk_nvme_transport_opts.rdma_srq_size); 3208 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(dev_attr.max_sge, 3209 NVME_RDMA_DEFAULT_RX_SGE); 3210 3211 poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr); 3212 if (poller->srq == NULL) { 3213 SPDK_ERRLOG("Unable to create SRQ.\n"); 3214 goto fail; 3215 } 3216 3217 opts.num_entries = g_spdk_nvme_transport_opts.rdma_srq_size; 3218 opts.rqpair = NULL; 3219 opts.srq = poller->srq; 3220 opts.mr_map = poller->mr_map; 3221 3222 poller->rsps = nvme_rdma_create_rsps(&opts); 3223 if (poller->rsps == NULL) { 3224 SPDK_ERRLOG("Unable to create poller RDMA responses.\n"); 3225 goto fail; 3226 } 3227 3228 rc = nvme_rdma_poller_submit_recvs(poller); 3229 if (rc) { 3230 SPDK_ERRLOG("Unable to submit poller RDMA responses.\n"); 3231 goto fail; 3232 } 3233 3234 /* 3235 * When using an srq, fix the size of the completion queue at startup. 3236 * The initiator sends only send and recv WRs. Hence, the multiplier is 2. 3237 * (The target sends also data WRs. Hence, the multiplier is 3.) 3238 */ 3239 num_cqe = g_spdk_nvme_transport_opts.rdma_srq_size * 2; 3240 } else { 3241 num_cqe = DEFAULT_NVME_RDMA_CQ_SIZE; 3242 } 3243 3244 max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size; 3245 if (max_num_cqe != 0 && num_cqe > max_num_cqe) { 3246 num_cqe = max_num_cqe; 3247 } 3248 3249 poller->cq = ibv_create_cq(poller->device, num_cqe, group, NULL, 0); 3250 3251 if (poller->cq == NULL) { 3252 SPDK_ERRLOG("Unable to create CQ, errno %d.\n", errno); 3253 goto fail; 3254 } 3255 3256 STAILQ_INSERT_HEAD(&group->pollers, poller, link); 3257 group->num_pollers++; 3258 poller->current_num_wc = num_cqe; 3259 poller->required_num_wc = 0; 3260 return poller; 3261 3262 fail: 3263 nvme_rdma_poller_destroy(poller); 3264 return NULL; 3265 } 3266 3267 static void 3268 nvme_rdma_poll_group_free_pollers(struct nvme_rdma_poll_group *group) 3269 { 3270 struct nvme_rdma_poller *poller, *tmp_poller; 3271 3272 STAILQ_FOREACH_SAFE(poller, &group->pollers, link, tmp_poller) { 3273 assert(poller->refcnt == 0); 3274 if (poller->refcnt) { 3275 SPDK_WARNLOG("Destroying poller with non-zero ref count: poller %p, refcnt %d\n", 3276 poller, poller->refcnt); 3277 } 3278 3279 STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link); 3280 nvme_rdma_poller_destroy(poller); 3281 } 3282 } 3283 3284 static struct nvme_rdma_poller * 3285 nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, struct ibv_context *device) 3286 { 3287 struct nvme_rdma_poller *poller = NULL; 3288 3289 STAILQ_FOREACH(poller, &group->pollers, link) { 3290 if (poller->device == device) { 3291 break; 3292 } 3293 } 3294 3295 if (!poller) { 3296 poller = nvme_rdma_poller_create(group, device); 3297 if (!poller) { 3298 SPDK_ERRLOG("Failed to create a poller for device %p\n", device); 3299 return NULL; 3300 } 3301 } 3302 3303 poller->refcnt++; 3304 return poller; 3305 } 3306 3307 static void 3308 nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, struct nvme_rdma_poller *poller) 3309 { 3310 assert(poller->refcnt > 0); 3311 if (--poller->refcnt == 0) { 3312 STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link); 3313 group->num_pollers--; 3314 nvme_rdma_poller_destroy(poller); 3315 } 3316 } 3317 3318 static struct spdk_nvme_transport_poll_group * 3319 nvme_rdma_poll_group_create(void) 3320 { 3321 struct nvme_rdma_poll_group *group; 3322 3323 group = calloc(1, sizeof(*group)); 3324 if (group == NULL) { 3325 SPDK_ERRLOG("Unable to allocate poll group.\n"); 3326 return NULL; 3327 } 3328 3329 STAILQ_INIT(&group->pollers); 3330 TAILQ_INIT(&group->connecting_qpairs); 3331 TAILQ_INIT(&group->active_qpairs); 3332 return &group->group; 3333 } 3334 3335 static int 3336 nvme_rdma_poll_group_connect_qpair(struct spdk_nvme_qpair *qpair) 3337 { 3338 return 0; 3339 } 3340 3341 static int 3342 nvme_rdma_poll_group_disconnect_qpair(struct spdk_nvme_qpair *qpair) 3343 { 3344 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3345 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group); 3346 3347 if (rqpair->link_connecting.tqe_prev) { 3348 TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting); 3349 /* We use prev pointer to check if qpair is in connecting list or not . 3350 * TAILQ_REMOVE doesn't do it. So, we do it manually. 3351 */ 3352 rqpair->link_connecting.tqe_prev = NULL; 3353 } 3354 3355 return 0; 3356 } 3357 3358 static int 3359 nvme_rdma_poll_group_add(struct spdk_nvme_transport_poll_group *tgroup, 3360 struct spdk_nvme_qpair *qpair) 3361 { 3362 return 0; 3363 } 3364 3365 static int 3366 nvme_rdma_poll_group_remove(struct spdk_nvme_transport_poll_group *tgroup, 3367 struct spdk_nvme_qpair *qpair) 3368 { 3369 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3370 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group); 3371 3372 if (rqpair->link_active.tqe_prev) { 3373 TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active); 3374 rqpair->link_active.tqe_prev = NULL; 3375 } 3376 3377 return 0; 3378 } 3379 3380 static inline void 3381 nvme_rdma_qpair_process_submits(struct nvme_rdma_poll_group *group, 3382 struct nvme_rdma_qpair *rqpair) 3383 { 3384 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 3385 3386 assert(rqpair->link_active.tqe_prev != NULL); 3387 3388 if (spdk_unlikely(rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING || 3389 rqpair->state >= NVME_RDMA_QPAIR_STATE_EXITING)) { 3390 return; 3391 } 3392 3393 if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) { 3394 nvme_rdma_qpair_check_timeout(qpair); 3395 } 3396 3397 nvme_rdma_qpair_submit_sends(rqpair); 3398 if (!rqpair->srq) { 3399 nvme_rdma_qpair_submit_recvs(rqpair); 3400 } 3401 if (rqpair->num_completions > 0) { 3402 nvme_qpair_resubmit_requests(qpair, rqpair->num_completions); 3403 rqpair->num_completions = 0; 3404 } 3405 3406 if (rqpair->num_outstanding_reqs == 0 && STAILQ_EMPTY(&qpair->queued_req)) { 3407 TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active); 3408 /* We use prev pointer to check if qpair is in active list or not. 3409 * TAILQ_REMOVE doesn't do it. So, we do it manually. 3410 */ 3411 rqpair->link_active.tqe_prev = NULL; 3412 } 3413 } 3414 3415 static int64_t 3416 nvme_rdma_poll_group_process_completions(struct spdk_nvme_transport_poll_group *tgroup, 3417 uint32_t completions_per_qpair, spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb) 3418 { 3419 struct spdk_nvme_qpair *qpair, *tmp_qpair; 3420 struct nvme_rdma_qpair *rqpair, *tmp_rqpair; 3421 struct nvme_rdma_poll_group *group; 3422 struct nvme_rdma_poller *poller; 3423 int batch_size, rc, rc2 = 0; 3424 int64_t total_completions = 0; 3425 uint64_t completions_allowed = 0; 3426 uint64_t completions_per_poller = 0; 3427 uint64_t poller_completions = 0; 3428 uint64_t rdma_completions; 3429 3430 if (completions_per_qpair == 0) { 3431 completions_per_qpair = MAX_COMPLETIONS_PER_POLL; 3432 } 3433 3434 group = nvme_rdma_poll_group(tgroup); 3435 3436 STAILQ_FOREACH_SAFE(qpair, &tgroup->disconnected_qpairs, poll_group_stailq, tmp_qpair) { 3437 rc = nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair); 3438 if (rc == 0) { 3439 disconnected_qpair_cb(qpair, tgroup->group->ctx); 3440 } 3441 } 3442 3443 TAILQ_FOREACH_SAFE(rqpair, &group->connecting_qpairs, link_connecting, tmp_rqpair) { 3444 qpair = &rqpair->qpair; 3445 3446 rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair); 3447 if (rc == 0 || rc != -EAGAIN) { 3448 TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting); 3449 /* We use prev pointer to check if qpair is in connecting list or not. 3450 * TAILQ_REMOVE does not do it. So, we do it manually. 3451 */ 3452 rqpair->link_connecting.tqe_prev = NULL; 3453 3454 if (rc == 0) { 3455 /* Once the connection is completed, we can submit queued requests */ 3456 nvme_qpair_resubmit_requests(qpair, rqpair->num_entries); 3457 } else if (rc != -EAGAIN) { 3458 SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair); 3459 nvme_rdma_fail_qpair(qpair, 0); 3460 } 3461 } 3462 } 3463 3464 STAILQ_FOREACH_SAFE(qpair, &tgroup->connected_qpairs, poll_group_stailq, tmp_qpair) { 3465 rqpair = nvme_rdma_qpair(qpair); 3466 3467 if (spdk_likely(nvme_qpair_get_state(qpair) != NVME_QPAIR_CONNECTING)) { 3468 nvme_rdma_qpair_process_cm_event(rqpair); 3469 } 3470 3471 if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) { 3472 rc2 = -ENXIO; 3473 nvme_rdma_fail_qpair(qpair, 0); 3474 } 3475 } 3476 3477 completions_allowed = completions_per_qpair * tgroup->num_connected_qpairs; 3478 if (spdk_likely(group->num_pollers)) { 3479 completions_per_poller = spdk_max(completions_allowed / group->num_pollers, 1); 3480 } 3481 3482 STAILQ_FOREACH(poller, &group->pollers, link) { 3483 poller_completions = 0; 3484 rdma_completions = 0; 3485 do { 3486 poller->stats.polls++; 3487 batch_size = spdk_min((completions_per_poller - poller_completions), MAX_COMPLETIONS_PER_POLL); 3488 rc = nvme_rdma_cq_process_completions(poller->cq, batch_size, poller, NULL, &rdma_completions); 3489 if (rc <= 0) { 3490 if (rc == -ECANCELED) { 3491 return -EIO; 3492 } else if (rc == 0) { 3493 poller->stats.idle_polls++; 3494 } 3495 break; 3496 } 3497 3498 poller_completions += rc; 3499 } while (poller_completions < completions_per_poller); 3500 total_completions += poller_completions; 3501 poller->stats.completions += rdma_completions; 3502 if (poller->srq) { 3503 nvme_rdma_poller_submit_recvs(poller); 3504 } 3505 } 3506 3507 TAILQ_FOREACH_SAFE(rqpair, &group->active_qpairs, link_active, tmp_rqpair) { 3508 nvme_rdma_qpair_process_submits(group, rqpair); 3509 } 3510 3511 return rc2 != 0 ? rc2 : total_completions; 3512 } 3513 3514 /* 3515 * Handle disconnected qpairs when interrupt support gets added. 3516 */ 3517 static void 3518 nvme_rdma_poll_group_check_disconnected_qpairs(struct spdk_nvme_transport_poll_group *tgroup, 3519 spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb) 3520 { 3521 } 3522 3523 static int 3524 nvme_rdma_poll_group_destroy(struct spdk_nvme_transport_poll_group *tgroup) 3525 { 3526 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(tgroup); 3527 3528 if (!STAILQ_EMPTY(&tgroup->connected_qpairs) || !STAILQ_EMPTY(&tgroup->disconnected_qpairs)) { 3529 return -EBUSY; 3530 } 3531 3532 nvme_rdma_poll_group_free_pollers(group); 3533 free(group); 3534 3535 return 0; 3536 } 3537 3538 static int 3539 nvme_rdma_poll_group_get_stats(struct spdk_nvme_transport_poll_group *tgroup, 3540 struct spdk_nvme_transport_poll_group_stat **_stats) 3541 { 3542 struct nvme_rdma_poll_group *group; 3543 struct spdk_nvme_transport_poll_group_stat *stats; 3544 struct spdk_nvme_rdma_device_stat *device_stat; 3545 struct nvme_rdma_poller *poller; 3546 uint32_t i = 0; 3547 3548 if (tgroup == NULL || _stats == NULL) { 3549 SPDK_ERRLOG("Invalid stats or group pointer\n"); 3550 return -EINVAL; 3551 } 3552 3553 group = nvme_rdma_poll_group(tgroup); 3554 stats = calloc(1, sizeof(*stats)); 3555 if (!stats) { 3556 SPDK_ERRLOG("Can't allocate memory for RDMA stats\n"); 3557 return -ENOMEM; 3558 } 3559 stats->trtype = SPDK_NVME_TRANSPORT_RDMA; 3560 stats->rdma.num_devices = group->num_pollers; 3561 3562 if (stats->rdma.num_devices == 0) { 3563 *_stats = stats; 3564 return 0; 3565 } 3566 3567 stats->rdma.device_stats = calloc(stats->rdma.num_devices, sizeof(*stats->rdma.device_stats)); 3568 if (!stats->rdma.device_stats) { 3569 SPDK_ERRLOG("Can't allocate memory for RDMA device stats\n"); 3570 free(stats); 3571 return -ENOMEM; 3572 } 3573 3574 STAILQ_FOREACH(poller, &group->pollers, link) { 3575 device_stat = &stats->rdma.device_stats[i]; 3576 device_stat->name = poller->device->device->name; 3577 device_stat->polls = poller->stats.polls; 3578 device_stat->idle_polls = poller->stats.idle_polls; 3579 device_stat->completions = poller->stats.completions; 3580 device_stat->queued_requests = poller->stats.queued_requests; 3581 device_stat->total_send_wrs = poller->stats.rdma_stats.send.num_submitted_wrs; 3582 device_stat->send_doorbell_updates = poller->stats.rdma_stats.send.doorbell_updates; 3583 device_stat->total_recv_wrs = poller->stats.rdma_stats.recv.num_submitted_wrs; 3584 device_stat->recv_doorbell_updates = poller->stats.rdma_stats.recv.doorbell_updates; 3585 i++; 3586 } 3587 3588 *_stats = stats; 3589 3590 return 0; 3591 } 3592 3593 static void 3594 nvme_rdma_poll_group_free_stats(struct spdk_nvme_transport_poll_group *tgroup, 3595 struct spdk_nvme_transport_poll_group_stat *stats) 3596 { 3597 if (stats) { 3598 free(stats->rdma.device_stats); 3599 } 3600 free(stats); 3601 } 3602 3603 static int 3604 nvme_rdma_ctrlr_get_memory_domains(const struct spdk_nvme_ctrlr *ctrlr, 3605 struct spdk_memory_domain **domains, int array_size) 3606 { 3607 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(ctrlr->adminq); 3608 3609 if (domains && array_size > 0) { 3610 domains[0] = rqpair->rdma_qp->domain; 3611 } 3612 3613 return 1; 3614 } 3615 3616 void 3617 spdk_nvme_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3618 { 3619 g_nvme_hooks = *hooks; 3620 } 3621 3622 const struct spdk_nvme_transport_ops rdma_ops = { 3623 .name = "RDMA", 3624 .type = SPDK_NVME_TRANSPORT_RDMA, 3625 .ctrlr_construct = nvme_rdma_ctrlr_construct, 3626 .ctrlr_scan = nvme_fabric_ctrlr_scan, 3627 .ctrlr_destruct = nvme_rdma_ctrlr_destruct, 3628 .ctrlr_enable = nvme_rdma_ctrlr_enable, 3629 3630 .ctrlr_set_reg_4 = nvme_fabric_ctrlr_set_reg_4, 3631 .ctrlr_set_reg_8 = nvme_fabric_ctrlr_set_reg_8, 3632 .ctrlr_get_reg_4 = nvme_fabric_ctrlr_get_reg_4, 3633 .ctrlr_get_reg_8 = nvme_fabric_ctrlr_get_reg_8, 3634 .ctrlr_set_reg_4_async = nvme_fabric_ctrlr_set_reg_4_async, 3635 .ctrlr_set_reg_8_async = nvme_fabric_ctrlr_set_reg_8_async, 3636 .ctrlr_get_reg_4_async = nvme_fabric_ctrlr_get_reg_4_async, 3637 .ctrlr_get_reg_8_async = nvme_fabric_ctrlr_get_reg_8_async, 3638 3639 .ctrlr_get_max_xfer_size = nvme_rdma_ctrlr_get_max_xfer_size, 3640 .ctrlr_get_max_sges = nvme_rdma_ctrlr_get_max_sges, 3641 3642 .ctrlr_create_io_qpair = nvme_rdma_ctrlr_create_io_qpair, 3643 .ctrlr_delete_io_qpair = nvme_rdma_ctrlr_delete_io_qpair, 3644 .ctrlr_connect_qpair = nvme_rdma_ctrlr_connect_qpair, 3645 .ctrlr_disconnect_qpair = nvme_rdma_ctrlr_disconnect_qpair, 3646 3647 .ctrlr_get_memory_domains = nvme_rdma_ctrlr_get_memory_domains, 3648 3649 .qpair_abort_reqs = nvme_rdma_qpair_abort_reqs, 3650 .qpair_reset = nvme_rdma_qpair_reset, 3651 .qpair_submit_request = nvme_rdma_qpair_submit_request, 3652 .qpair_process_completions = nvme_rdma_qpair_process_completions, 3653 .qpair_iterate_requests = nvme_rdma_qpair_iterate_requests, 3654 .qpair_authenticate = nvme_rdma_qpair_authenticate, 3655 .admin_qpair_abort_aers = nvme_rdma_admin_qpair_abort_aers, 3656 3657 .poll_group_create = nvme_rdma_poll_group_create, 3658 .poll_group_connect_qpair = nvme_rdma_poll_group_connect_qpair, 3659 .poll_group_disconnect_qpair = nvme_rdma_poll_group_disconnect_qpair, 3660 .poll_group_add = nvme_rdma_poll_group_add, 3661 .poll_group_remove = nvme_rdma_poll_group_remove, 3662 .poll_group_process_completions = nvme_rdma_poll_group_process_completions, 3663 .poll_group_check_disconnected_qpairs = nvme_rdma_poll_group_check_disconnected_qpairs, 3664 .poll_group_destroy = nvme_rdma_poll_group_destroy, 3665 .poll_group_get_stats = nvme_rdma_poll_group_get_stats, 3666 .poll_group_free_stats = nvme_rdma_poll_group_free_stats, 3667 }; 3668 3669 SPDK_NVME_TRANSPORT_REGISTER(rdma, &rdma_ops); 3670