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 /* Append copy task even if no accel sequence is attached to IO. 205 * Result is UMR configured per IO data buffer */ 206 bool append_copy; 207 208 uint32_t num_completions; 209 uint32_t num_outstanding_reqs; 210 211 struct nvme_rdma_rsps *rsps; 212 213 /* 214 * Array of num_entries NVMe commands registered as RDMA message buffers. 215 * Indexed by rdma_req->id. 216 */ 217 struct spdk_nvmf_cmd *cmds; 218 219 struct spdk_rdma_utils_mem_map *mr_map; 220 221 TAILQ_HEAD(, spdk_nvme_rdma_req) free_reqs; 222 TAILQ_HEAD(, spdk_nvme_rdma_req) outstanding_reqs; 223 224 /* Count of outstanding send objects */ 225 uint16_t current_num_sends; 226 227 TAILQ_ENTRY(nvme_rdma_qpair) link_active; 228 229 /* Placed at the end of the struct since it is not used frequently */ 230 struct rdma_cm_event *evt; 231 struct nvme_rdma_poller *poller; 232 233 uint64_t evt_timeout_ticks; 234 nvme_rdma_cm_event_cb evt_cb; 235 enum rdma_cm_event_type expected_evt_type; 236 237 enum nvme_rdma_qpair_state state; 238 239 bool in_connect_poll; 240 241 uint8_t stale_conn_retry_count; 242 bool need_destroy; 243 TAILQ_ENTRY(nvme_rdma_qpair) link_connecting; 244 }; 245 246 enum NVME_RDMA_COMPLETION_FLAGS { 247 NVME_RDMA_SEND_COMPLETED = 1u << 0, 248 NVME_RDMA_RECV_COMPLETED = 1u << 1, 249 }; 250 251 struct spdk_nvme_rdma_req { 252 uint16_t id; 253 uint16_t completion_flags: 2; 254 uint16_t in_progress_accel: 1; 255 uint16_t reserved: 13; 256 /* if completion of RDMA_RECV received before RDMA_SEND, we will complete nvme request 257 * during processing of RDMA_SEND. To complete the request we must know the response 258 * received in RDMA_RECV, so store it in this field */ 259 struct spdk_nvme_rdma_rsp *rdma_rsp; 260 261 struct nvme_rdma_wr rdma_wr; 262 263 struct ibv_send_wr send_wr; 264 265 struct nvme_request *req; 266 267 struct ibv_sge send_sgl[NVME_RDMA_DEFAULT_TX_SGE]; 268 269 TAILQ_ENTRY(spdk_nvme_rdma_req) link; 270 271 /* Fields below are not used in regular IO path, keep them last */ 272 spdk_memory_domain_data_cpl_cb transfer_cpl_cb; 273 void *transfer_cpl_cb_arg; 274 /* Accel sequence API works with iovec pointer, we need to store result of next_sge callback */ 275 struct iovec iovs[NVME_RDMA_MAX_SGL_DESCRIPTORS]; 276 }; 277 278 struct spdk_nvme_rdma_rsp { 279 struct spdk_nvme_cpl cpl; 280 struct nvme_rdma_qpair *rqpair; 281 struct ibv_recv_wr *recv_wr; 282 struct nvme_rdma_wr rdma_wr; 283 }; 284 285 struct nvme_rdma_memory_translation_ctx { 286 void *addr; 287 size_t length; 288 uint32_t lkey; 289 uint32_t rkey; 290 }; 291 292 static const char *rdma_cm_event_str[] = { 293 "RDMA_CM_EVENT_ADDR_RESOLVED", 294 "RDMA_CM_EVENT_ADDR_ERROR", 295 "RDMA_CM_EVENT_ROUTE_RESOLVED", 296 "RDMA_CM_EVENT_ROUTE_ERROR", 297 "RDMA_CM_EVENT_CONNECT_REQUEST", 298 "RDMA_CM_EVENT_CONNECT_RESPONSE", 299 "RDMA_CM_EVENT_CONNECT_ERROR", 300 "RDMA_CM_EVENT_UNREACHABLE", 301 "RDMA_CM_EVENT_REJECTED", 302 "RDMA_CM_EVENT_ESTABLISHED", 303 "RDMA_CM_EVENT_DISCONNECTED", 304 "RDMA_CM_EVENT_DEVICE_REMOVAL", 305 "RDMA_CM_EVENT_MULTICAST_JOIN", 306 "RDMA_CM_EVENT_MULTICAST_ERROR", 307 "RDMA_CM_EVENT_ADDR_CHANGE", 308 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 309 }; 310 311 static struct nvme_rdma_poller *nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, 312 struct ibv_context *device); 313 static void nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, 314 struct nvme_rdma_poller *poller); 315 316 static int nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, 317 struct spdk_nvme_qpair *qpair); 318 319 static inline int nvme_rdma_memory_domain_transfer_data(struct spdk_memory_domain *dst_domain, 320 void *dst_domain_ctx, 321 struct iovec *dst_iov, uint32_t dst_iovcnt, 322 struct spdk_memory_domain *src_domain, void *src_domain_ctx, 323 struct iovec *src_iov, uint32_t src_iovcnt, 324 struct spdk_memory_domain_translation_result *translation, 325 spdk_memory_domain_data_cpl_cb cpl_cb, void *cpl_cb_arg); 326 327 static inline int _nvme_rdma_qpair_submit_request(struct nvme_rdma_qpair *rqpair, 328 struct spdk_nvme_rdma_req *rdma_req); 329 330 static inline struct nvme_rdma_qpair * 331 nvme_rdma_qpair(struct spdk_nvme_qpair *qpair) 332 { 333 assert(qpair->trtype == SPDK_NVME_TRANSPORT_RDMA); 334 return SPDK_CONTAINEROF(qpair, struct nvme_rdma_qpair, qpair); 335 } 336 337 static inline struct nvme_rdma_poll_group * 338 nvme_rdma_poll_group(struct spdk_nvme_transport_poll_group *group) 339 { 340 return (SPDK_CONTAINEROF(group, struct nvme_rdma_poll_group, group)); 341 } 342 343 static inline struct nvme_rdma_ctrlr * 344 nvme_rdma_ctrlr(struct spdk_nvme_ctrlr *ctrlr) 345 { 346 assert(ctrlr->trid.trtype == SPDK_NVME_TRANSPORT_RDMA); 347 return SPDK_CONTAINEROF(ctrlr, struct nvme_rdma_ctrlr, ctrlr); 348 } 349 350 static inline struct spdk_nvme_rdma_req * 351 nvme_rdma_req_get(struct nvme_rdma_qpair *rqpair) 352 { 353 struct spdk_nvme_rdma_req *rdma_req; 354 355 rdma_req = TAILQ_FIRST(&rqpair->free_reqs); 356 if (spdk_likely(rdma_req)) { 357 TAILQ_REMOVE(&rqpair->free_reqs, rdma_req, link); 358 } 359 360 return rdma_req; 361 } 362 363 static inline void 364 nvme_rdma_req_put(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req) 365 { 366 rdma_req->completion_flags = 0; 367 rdma_req->req = NULL; 368 rdma_req->rdma_rsp = NULL; 369 assert(rdma_req->transfer_cpl_cb == NULL); 370 TAILQ_INSERT_HEAD(&rqpair->free_reqs, rdma_req, link); 371 } 372 373 static inline void 374 nvme_rdma_finish_data_transfer(struct spdk_nvme_rdma_req *rdma_req, int rc) 375 { 376 spdk_memory_domain_data_cpl_cb cb = rdma_req->transfer_cpl_cb; 377 378 SPDK_DEBUGLOG(nvme, "req %p, finish data transfer, rc %d\n", rdma_req, rc); 379 rdma_req->transfer_cpl_cb = NULL; 380 assert(cb); 381 cb(rdma_req->transfer_cpl_cb_arg, rc); 382 } 383 384 static void 385 nvme_rdma_req_complete(struct spdk_nvme_rdma_req *rdma_req, 386 struct spdk_nvme_cpl *rsp, 387 bool print_on_error) 388 { 389 struct nvme_request *req = rdma_req->req; 390 struct nvme_rdma_qpair *rqpair; 391 struct spdk_nvme_qpair *qpair; 392 bool error, print_error; 393 394 assert(req != NULL); 395 396 qpair = req->qpair; 397 rqpair = nvme_rdma_qpair(qpair); 398 399 error = spdk_nvme_cpl_is_error(rsp); 400 print_error = error && print_on_error && !qpair->ctrlr->opts.disable_error_logging; 401 402 if (print_error) { 403 spdk_nvme_qpair_print_command(qpair, &req->cmd); 404 } 405 406 if (print_error || SPDK_DEBUGLOG_FLAG_ENABLED("nvme")) { 407 spdk_nvme_qpair_print_completion(qpair, rsp); 408 } 409 410 assert(rqpair->num_outstanding_reqs > 0); 411 rqpair->num_outstanding_reqs--; 412 413 TAILQ_REMOVE(&rqpair->outstanding_reqs, rdma_req, link); 414 415 nvme_complete_request(req->cb_fn, req->cb_arg, qpair, req, rsp); 416 nvme_rdma_req_put(rqpair, rdma_req); 417 } 418 419 static const char * 420 nvme_rdma_cm_event_str_get(uint32_t event) 421 { 422 if (event < SPDK_COUNTOF(rdma_cm_event_str)) { 423 return rdma_cm_event_str[event]; 424 } else { 425 return "Undefined"; 426 } 427 } 428 429 430 static int 431 nvme_rdma_qpair_process_cm_event(struct nvme_rdma_qpair *rqpair) 432 { 433 struct rdma_cm_event *event = rqpair->evt; 434 struct spdk_nvmf_rdma_accept_private_data *accept_data; 435 int rc = 0; 436 437 if (event) { 438 switch (event->event) { 439 case RDMA_CM_EVENT_ADDR_RESOLVED: 440 case RDMA_CM_EVENT_ADDR_ERROR: 441 case RDMA_CM_EVENT_ROUTE_RESOLVED: 442 case RDMA_CM_EVENT_ROUTE_ERROR: 443 break; 444 case RDMA_CM_EVENT_CONNECT_REQUEST: 445 break; 446 case RDMA_CM_EVENT_CONNECT_ERROR: 447 break; 448 case RDMA_CM_EVENT_UNREACHABLE: 449 case RDMA_CM_EVENT_REJECTED: 450 break; 451 case RDMA_CM_EVENT_CONNECT_RESPONSE: 452 rc = spdk_rdma_provider_qp_complete_connect(rqpair->rdma_qp); 453 /* fall through */ 454 case RDMA_CM_EVENT_ESTABLISHED: 455 accept_data = (struct spdk_nvmf_rdma_accept_private_data *)event->param.conn.private_data; 456 if (accept_data == NULL) { 457 rc = -1; 458 } else { 459 SPDK_DEBUGLOG(nvme, "Requested queue depth %d. Target receive queue depth %d.\n", 460 rqpair->num_entries + 1, accept_data->crqsize); 461 } 462 break; 463 case RDMA_CM_EVENT_DISCONNECTED: 464 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE; 465 break; 466 case RDMA_CM_EVENT_DEVICE_REMOVAL: 467 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL; 468 rqpair->need_destroy = true; 469 break; 470 case RDMA_CM_EVENT_MULTICAST_JOIN: 471 case RDMA_CM_EVENT_MULTICAST_ERROR: 472 break; 473 case RDMA_CM_EVENT_ADDR_CHANGE: 474 rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL; 475 break; 476 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 477 break; 478 default: 479 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 480 break; 481 } 482 rqpair->evt = NULL; 483 rdma_ack_cm_event(event); 484 } 485 486 return rc; 487 } 488 489 /* 490 * This function must be called under the nvme controller's lock 491 * because it touches global controller variables. The lock is taken 492 * by the generic transport code before invoking a few of the functions 493 * in this file: nvme_rdma_ctrlr_connect_qpair, nvme_rdma_ctrlr_delete_io_qpair, 494 * and conditionally nvme_rdma_qpair_process_completions when it is calling 495 * completions on the admin qpair. When adding a new call to this function, please 496 * verify that it is in a situation where it falls under the lock. 497 */ 498 static int 499 nvme_rdma_poll_events(struct nvme_rdma_ctrlr *rctrlr) 500 { 501 struct nvme_rdma_cm_event_entry *entry, *tmp; 502 struct nvme_rdma_qpair *event_qpair; 503 struct rdma_cm_event *event; 504 struct rdma_event_channel *channel = rctrlr->cm_channel; 505 506 STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) { 507 event_qpair = entry->evt->id->context; 508 if (event_qpair->evt == NULL) { 509 event_qpair->evt = entry->evt; 510 STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link); 511 STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link); 512 } 513 } 514 515 while (rdma_get_cm_event(channel, &event) == 0) { 516 event_qpair = event->id->context; 517 if (event_qpair->evt == NULL) { 518 event_qpair->evt = event; 519 } else { 520 assert(rctrlr == nvme_rdma_ctrlr(event_qpair->qpair.ctrlr)); 521 entry = STAILQ_FIRST(&rctrlr->free_cm_events); 522 if (entry == NULL) { 523 rdma_ack_cm_event(event); 524 return -ENOMEM; 525 } 526 STAILQ_REMOVE(&rctrlr->free_cm_events, entry, nvme_rdma_cm_event_entry, link); 527 entry->evt = event; 528 STAILQ_INSERT_TAIL(&rctrlr->pending_cm_events, entry, link); 529 } 530 } 531 532 /* rdma_get_cm_event() returns -1 on error. If an error occurs, errno 533 * will be set to indicate the failure reason. So return negated errno here. 534 */ 535 return -errno; 536 } 537 538 static int 539 nvme_rdma_validate_cm_event(enum rdma_cm_event_type expected_evt_type, 540 struct rdma_cm_event *reaped_evt) 541 { 542 int rc = -EBADMSG; 543 544 if (expected_evt_type == reaped_evt->event) { 545 return 0; 546 } 547 548 switch (expected_evt_type) { 549 case RDMA_CM_EVENT_ESTABLISHED: 550 /* 551 * There is an enum ib_cm_rej_reason in the kernel headers that sets 10 as 552 * IB_CM_REJ_STALE_CONN. I can't find the corresponding userspace but we get 553 * the same values here. 554 */ 555 if (reaped_evt->event == RDMA_CM_EVENT_REJECTED && reaped_evt->status == 10) { 556 rc = -ESTALE; 557 } else if (reaped_evt->event == RDMA_CM_EVENT_CONNECT_RESPONSE) { 558 /* 559 * If we are using a qpair which is not created using rdma cm API 560 * then we will receive RDMA_CM_EVENT_CONNECT_RESPONSE instead of 561 * RDMA_CM_EVENT_ESTABLISHED. 562 */ 563 return 0; 564 } 565 break; 566 default: 567 break; 568 } 569 570 SPDK_ERRLOG("Expected %s but received %s (%d) from CM event channel (status = %d)\n", 571 nvme_rdma_cm_event_str_get(expected_evt_type), 572 nvme_rdma_cm_event_str_get(reaped_evt->event), reaped_evt->event, 573 reaped_evt->status); 574 return rc; 575 } 576 577 static int 578 nvme_rdma_process_event_start(struct nvme_rdma_qpair *rqpair, 579 enum rdma_cm_event_type evt, 580 nvme_rdma_cm_event_cb evt_cb) 581 { 582 int rc; 583 584 assert(evt_cb != NULL); 585 586 if (rqpair->evt != NULL) { 587 rc = nvme_rdma_qpair_process_cm_event(rqpair); 588 if (rc) { 589 return rc; 590 } 591 } 592 593 rqpair->expected_evt_type = evt; 594 rqpair->evt_cb = evt_cb; 595 rqpair->evt_timeout_ticks = (g_spdk_nvme_transport_opts.rdma_cm_event_timeout_ms * 1000 * 596 spdk_get_ticks_hz()) / SPDK_SEC_TO_USEC + spdk_get_ticks(); 597 598 return 0; 599 } 600 601 static int 602 nvme_rdma_process_event_poll(struct nvme_rdma_qpair *rqpair) 603 { 604 struct nvme_rdma_ctrlr *rctrlr; 605 int rc = 0, rc2; 606 607 rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr); 608 assert(rctrlr != NULL); 609 610 if (!rqpair->evt && spdk_get_ticks() < rqpair->evt_timeout_ticks) { 611 rc = nvme_rdma_poll_events(rctrlr); 612 if (rc == -EAGAIN || rc == -EWOULDBLOCK) { 613 return rc; 614 } 615 } 616 617 if (rqpair->evt == NULL) { 618 rc = -EADDRNOTAVAIL; 619 goto exit; 620 } 621 622 rc = nvme_rdma_validate_cm_event(rqpair->expected_evt_type, rqpair->evt); 623 624 rc2 = nvme_rdma_qpair_process_cm_event(rqpair); 625 /* bad message takes precedence over the other error codes from processing the event. */ 626 rc = rc == 0 ? rc2 : rc; 627 628 exit: 629 assert(rqpair->evt_cb != NULL); 630 return rqpair->evt_cb(rqpair, rc); 631 } 632 633 static int 634 nvme_rdma_resize_cq(struct nvme_rdma_qpair *rqpair, struct nvme_rdma_poller *poller) 635 { 636 int current_num_wc, required_num_wc; 637 int max_cq_size; 638 639 required_num_wc = poller->required_num_wc + WC_PER_QPAIR(rqpair->num_entries); 640 current_num_wc = poller->current_num_wc; 641 if (current_num_wc < required_num_wc) { 642 current_num_wc = spdk_max(current_num_wc * 2, required_num_wc); 643 } 644 645 max_cq_size = g_spdk_nvme_transport_opts.rdma_max_cq_size; 646 if (max_cq_size != 0 && current_num_wc > max_cq_size) { 647 current_num_wc = max_cq_size; 648 } 649 650 if (poller->current_num_wc != current_num_wc) { 651 SPDK_DEBUGLOG(nvme, "Resize RDMA CQ from %d to %d\n", poller->current_num_wc, 652 current_num_wc); 653 if (ibv_resize_cq(poller->cq, current_num_wc)) { 654 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 655 return -1; 656 } 657 658 poller->current_num_wc = current_num_wc; 659 } 660 661 poller->required_num_wc = required_num_wc; 662 return 0; 663 } 664 665 static int 666 nvme_rdma_qpair_set_poller(struct spdk_nvme_qpair *qpair) 667 { 668 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 669 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group); 670 struct nvme_rdma_poller *poller; 671 672 assert(rqpair->cq == NULL); 673 674 poller = nvme_rdma_poll_group_get_poller(group, rqpair->cm_id->verbs); 675 if (!poller) { 676 SPDK_ERRLOG("Unable to find a cq for qpair %p on poll group %p\n", qpair, qpair->poll_group); 677 return -EINVAL; 678 } 679 680 if (!poller->srq) { 681 if (nvme_rdma_resize_cq(rqpair, poller)) { 682 nvme_rdma_poll_group_put_poller(group, poller); 683 return -EPROTO; 684 } 685 } 686 687 rqpair->cq = poller->cq; 688 rqpair->srq = poller->srq; 689 if (rqpair->srq) { 690 rqpair->rsps = poller->rsps; 691 } 692 rqpair->poller = poller; 693 return 0; 694 } 695 696 static int 697 nvme_rdma_qpair_init(struct nvme_rdma_qpair *rqpair) 698 { 699 int rc; 700 struct spdk_rdma_provider_qp_init_attr attr = {}; 701 struct ibv_device_attr dev_attr; 702 struct nvme_rdma_ctrlr *rctrlr; 703 uint32_t num_cqe, max_num_cqe; 704 705 rc = ibv_query_device(rqpair->cm_id->verbs, &dev_attr); 706 if (rc != 0) { 707 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 708 return -1; 709 } 710 711 if (rqpair->qpair.poll_group) { 712 assert(!rqpair->cq); 713 rc = nvme_rdma_qpair_set_poller(&rqpair->qpair); 714 if (rc) { 715 SPDK_ERRLOG("Unable to activate the rdmaqpair.\n"); 716 return -1; 717 } 718 assert(rqpair->cq); 719 } else { 720 num_cqe = rqpair->num_entries * 2; 721 max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size; 722 if (max_num_cqe != 0 && num_cqe > max_num_cqe) { 723 num_cqe = max_num_cqe; 724 } 725 rqpair->cq = ibv_create_cq(rqpair->cm_id->verbs, num_cqe, rqpair, NULL, 0); 726 if (!rqpair->cq) { 727 SPDK_ERRLOG("Unable to create completion queue: errno %d: %s\n", errno, spdk_strerror(errno)); 728 return -1; 729 } 730 } 731 732 rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr); 733 if (g_nvme_hooks.get_ibv_pd) { 734 attr.pd = g_nvme_hooks.get_ibv_pd(&rctrlr->ctrlr.trid, rqpair->cm_id->verbs); 735 } else { 736 attr.pd = spdk_rdma_utils_get_pd(rqpair->cm_id->verbs); 737 } 738 739 attr.stats = rqpair->poller ? &rqpair->poller->stats.rdma_stats : NULL; 740 attr.send_cq = rqpair->cq; 741 attr.recv_cq = rqpair->cq; 742 attr.cap.max_send_wr = rqpair->num_entries; /* SEND operations */ 743 if (rqpair->srq) { 744 attr.srq = rqpair->srq->srq; 745 } else { 746 attr.cap.max_recv_wr = rqpair->num_entries; /* RECV operations */ 747 } 748 attr.cap.max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, dev_attr.max_sge); 749 attr.cap.max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, dev_attr.max_sge); 750 attr.domain_transfer = spdk_rdma_provider_accel_sequence_supported() ? 751 nvme_rdma_memory_domain_transfer_data : NULL; 752 753 rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &attr); 754 755 if (!rqpair->rdma_qp) { 756 return -1; 757 } 758 759 /* ibv_create_qp will change the values in attr.cap. Make sure we store the proper value. */ 760 rqpair->max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, attr.cap.max_send_sge); 761 rqpair->current_num_sends = 0; 762 763 rqpair->cm_id->context = rqpair; 764 765 return 0; 766 } 767 768 static void 769 nvme_rdma_reset_failed_sends(struct nvme_rdma_qpair *rqpair, 770 struct ibv_send_wr *bad_send_wr) 771 { 772 while (bad_send_wr != NULL) { 773 assert(rqpair->current_num_sends > 0); 774 rqpair->current_num_sends--; 775 bad_send_wr = bad_send_wr->next; 776 } 777 } 778 779 static void 780 nvme_rdma_reset_failed_recvs(struct nvme_rdma_rsps *rsps, 781 struct ibv_recv_wr *bad_recv_wr, int rc) 782 { 783 SPDK_ERRLOG("Failed to post WRs on receive queue, errno %d (%s), bad_wr %p\n", 784 rc, spdk_strerror(rc), bad_recv_wr); 785 while (bad_recv_wr != NULL) { 786 assert(rsps->current_num_recvs > 0); 787 rsps->current_num_recvs--; 788 bad_recv_wr = bad_recv_wr->next; 789 } 790 } 791 792 static inline int 793 nvme_rdma_qpair_submit_sends(struct nvme_rdma_qpair *rqpair) 794 { 795 struct ibv_send_wr *bad_send_wr = NULL; 796 int rc; 797 798 rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_send_wr); 799 800 if (spdk_unlikely(rc)) { 801 SPDK_ERRLOG("Failed to post WRs on send queue, errno %d (%s), bad_wr %p\n", 802 rc, spdk_strerror(rc), bad_send_wr); 803 nvme_rdma_reset_failed_sends(rqpair, bad_send_wr); 804 } 805 806 return rc; 807 } 808 809 static inline int 810 nvme_rdma_qpair_submit_recvs(struct nvme_rdma_qpair *rqpair) 811 { 812 struct ibv_recv_wr *bad_recv_wr; 813 int rc = 0; 814 815 rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 816 if (spdk_unlikely(rc)) { 817 nvme_rdma_reset_failed_recvs(rqpair->rsps, bad_recv_wr, rc); 818 } 819 820 return rc; 821 } 822 823 static inline int 824 nvme_rdma_poller_submit_recvs(struct nvme_rdma_poller *poller) 825 { 826 struct ibv_recv_wr *bad_recv_wr; 827 int rc; 828 829 rc = spdk_rdma_provider_srq_flush_recv_wrs(poller->srq, &bad_recv_wr); 830 if (spdk_unlikely(rc)) { 831 nvme_rdma_reset_failed_recvs(poller->rsps, bad_recv_wr, rc); 832 } 833 834 return rc; 835 } 836 837 #define nvme_rdma_trace_ibv_sge(sg_list) \ 838 if (sg_list) { \ 839 SPDK_DEBUGLOG(nvme, "local addr %p length 0x%x lkey 0x%x\n", \ 840 (void *)(sg_list)->addr, (sg_list)->length, (sg_list)->lkey); \ 841 } 842 843 static void 844 nvme_rdma_free_rsps(struct nvme_rdma_rsps *rsps) 845 { 846 if (!rsps) { 847 return; 848 } 849 850 spdk_free(rsps->rsps); 851 spdk_free(rsps->rsp_sgls); 852 spdk_free(rsps->rsp_recv_wrs); 853 spdk_free(rsps); 854 } 855 856 static struct nvme_rdma_rsps * 857 nvme_rdma_create_rsps(struct nvme_rdma_rsp_opts *opts) 858 { 859 struct nvme_rdma_rsps *rsps; 860 struct spdk_rdma_utils_memory_translation translation; 861 uint16_t i; 862 int rc; 863 864 rsps = spdk_zmalloc(sizeof(*rsps), 0, NULL, SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 865 if (!rsps) { 866 SPDK_ERRLOG("Failed to allocate rsps object\n"); 867 return NULL; 868 } 869 870 rsps->rsp_sgls = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_sgls), 0, NULL, 871 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 872 if (!rsps->rsp_sgls) { 873 SPDK_ERRLOG("Failed to allocate rsp_sgls\n"); 874 goto fail; 875 } 876 877 rsps->rsp_recv_wrs = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_recv_wrs), 0, NULL, 878 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 879 if (!rsps->rsp_recv_wrs) { 880 SPDK_ERRLOG("Failed to allocate rsp_recv_wrs\n"); 881 goto fail; 882 } 883 884 rsps->rsps = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsps), 0, NULL, 885 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 886 if (!rsps->rsps) { 887 SPDK_ERRLOG("can not allocate rdma rsps\n"); 888 goto fail; 889 } 890 891 for (i = 0; i < opts->num_entries; i++) { 892 struct ibv_sge *rsp_sgl = &rsps->rsp_sgls[i]; 893 struct spdk_nvme_rdma_rsp *rsp = &rsps->rsps[i]; 894 struct ibv_recv_wr *recv_wr = &rsps->rsp_recv_wrs[i]; 895 896 rsp->rqpair = opts->rqpair; 897 rsp->rdma_wr.type = RDMA_WR_TYPE_RECV; 898 rsp->recv_wr = recv_wr; 899 rsp_sgl->addr = (uint64_t)rsp; 900 rsp_sgl->length = sizeof(struct spdk_nvme_cpl); 901 rc = spdk_rdma_utils_get_translation(opts->mr_map, rsp, sizeof(*rsp), &translation); 902 if (rc) { 903 goto fail; 904 } 905 rsp_sgl->lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 906 907 recv_wr->wr_id = (uint64_t)&rsp->rdma_wr; 908 recv_wr->next = NULL; 909 recv_wr->sg_list = rsp_sgl; 910 recv_wr->num_sge = 1; 911 912 nvme_rdma_trace_ibv_sge(recv_wr->sg_list); 913 914 if (opts->rqpair) { 915 spdk_rdma_provider_qp_queue_recv_wrs(opts->rqpair->rdma_qp, recv_wr); 916 } else { 917 spdk_rdma_provider_srq_queue_recv_wrs(opts->srq, recv_wr); 918 } 919 } 920 921 rsps->num_entries = opts->num_entries; 922 rsps->current_num_recvs = opts->num_entries; 923 924 return rsps; 925 fail: 926 nvme_rdma_free_rsps(rsps); 927 return NULL; 928 } 929 930 static void 931 nvme_rdma_free_reqs(struct nvme_rdma_qpair *rqpair) 932 { 933 if (!rqpair->rdma_reqs) { 934 return; 935 } 936 937 spdk_free(rqpair->cmds); 938 rqpair->cmds = NULL; 939 940 spdk_free(rqpair->rdma_reqs); 941 rqpair->rdma_reqs = NULL; 942 } 943 944 static int 945 nvme_rdma_create_reqs(struct nvme_rdma_qpair *rqpair) 946 { 947 struct spdk_rdma_utils_memory_translation translation; 948 uint16_t i; 949 int rc; 950 951 assert(!rqpair->rdma_reqs); 952 rqpair->rdma_reqs = spdk_zmalloc(rqpair->num_entries * sizeof(struct spdk_nvme_rdma_req), 0, NULL, 953 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 954 if (rqpair->rdma_reqs == NULL) { 955 SPDK_ERRLOG("Failed to allocate rdma_reqs\n"); 956 goto fail; 957 } 958 959 assert(!rqpair->cmds); 960 rqpair->cmds = spdk_zmalloc(rqpair->num_entries * sizeof(*rqpair->cmds), 0, NULL, 961 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 962 if (!rqpair->cmds) { 963 SPDK_ERRLOG("Failed to allocate RDMA cmds\n"); 964 goto fail; 965 } 966 967 TAILQ_INIT(&rqpair->free_reqs); 968 TAILQ_INIT(&rqpair->outstanding_reqs); 969 for (i = 0; i < rqpair->num_entries; i++) { 970 struct spdk_nvme_rdma_req *rdma_req; 971 struct spdk_nvmf_cmd *cmd; 972 973 rdma_req = &rqpair->rdma_reqs[i]; 974 rdma_req->rdma_wr.type = RDMA_WR_TYPE_SEND; 975 cmd = &rqpair->cmds[i]; 976 977 rdma_req->id = i; 978 979 rc = spdk_rdma_utils_get_translation(rqpair->mr_map, cmd, sizeof(*cmd), &translation); 980 if (rc) { 981 goto fail; 982 } 983 rdma_req->send_sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 984 985 /* The first RDMA sgl element will always point 986 * at this data structure. Depending on whether 987 * an NVMe-oF SGL is required, the length of 988 * this element may change. */ 989 rdma_req->send_sgl[0].addr = (uint64_t)cmd; 990 rdma_req->send_wr.wr_id = (uint64_t)&rdma_req->rdma_wr; 991 rdma_req->send_wr.next = NULL; 992 rdma_req->send_wr.opcode = IBV_WR_SEND; 993 rdma_req->send_wr.send_flags = IBV_SEND_SIGNALED; 994 rdma_req->send_wr.sg_list = rdma_req->send_sgl; 995 rdma_req->send_wr.imm_data = 0; 996 997 TAILQ_INSERT_TAIL(&rqpair->free_reqs, rdma_req, link); 998 } 999 1000 return 0; 1001 fail: 1002 nvme_rdma_free_reqs(rqpair); 1003 return -ENOMEM; 1004 } 1005 1006 static int nvme_rdma_connect(struct nvme_rdma_qpair *rqpair); 1007 1008 static int 1009 nvme_rdma_route_resolved(struct nvme_rdma_qpair *rqpair, int ret) 1010 { 1011 if (ret) { 1012 SPDK_ERRLOG("RDMA route resolution error\n"); 1013 return -1; 1014 } 1015 1016 ret = nvme_rdma_qpair_init(rqpair); 1017 if (ret < 0) { 1018 SPDK_ERRLOG("nvme_rdma_qpair_init() failed\n"); 1019 return -1; 1020 } 1021 1022 return nvme_rdma_connect(rqpair); 1023 } 1024 1025 static int 1026 nvme_rdma_addr_resolved(struct nvme_rdma_qpair *rqpair, int ret) 1027 { 1028 if (ret) { 1029 SPDK_ERRLOG("RDMA address resolution error\n"); 1030 return -1; 1031 } 1032 1033 if (rqpair->qpair.ctrlr->opts.transport_ack_timeout != SPDK_NVME_TRANSPORT_ACK_TIMEOUT_DISABLED) { 1034 #ifdef SPDK_CONFIG_RDMA_SET_ACK_TIMEOUT 1035 uint8_t timeout = rqpair->qpair.ctrlr->opts.transport_ack_timeout; 1036 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, 1037 RDMA_OPTION_ID_ACK_TIMEOUT, 1038 &timeout, sizeof(timeout)); 1039 if (ret) { 1040 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_ACK_TIMEOUT %d, ret %d\n", timeout, ret); 1041 } 1042 #else 1043 SPDK_DEBUGLOG(nvme, "transport_ack_timeout is not supported\n"); 1044 #endif 1045 } 1046 1047 if (rqpair->qpair.ctrlr->opts.transport_tos != SPDK_NVME_TRANSPORT_TOS_DISABLED) { 1048 #ifdef SPDK_CONFIG_RDMA_SET_TOS 1049 uint8_t tos = rqpair->qpair.ctrlr->opts.transport_tos; 1050 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_TOS, &tos, sizeof(tos)); 1051 if (ret) { 1052 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_TOS %u, ret %d\n", tos, ret); 1053 } 1054 #else 1055 SPDK_DEBUGLOG(nvme, "transport_tos is not supported\n"); 1056 #endif 1057 } 1058 1059 ret = rdma_resolve_route(rqpair->cm_id, NVME_RDMA_TIME_OUT_IN_MS); 1060 if (ret) { 1061 SPDK_ERRLOG("rdma_resolve_route\n"); 1062 return ret; 1063 } 1064 1065 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ROUTE_RESOLVED, 1066 nvme_rdma_route_resolved); 1067 } 1068 1069 static int 1070 nvme_rdma_resolve_addr(struct nvme_rdma_qpair *rqpair, 1071 struct sockaddr *src_addr, 1072 struct sockaddr *dst_addr) 1073 { 1074 int ret; 1075 1076 if (src_addr) { 1077 int reuse = 1; 1078 1079 ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_REUSEADDR, 1080 &reuse, sizeof(reuse)); 1081 if (ret) { 1082 SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_REUSEADDR %d, ret %d\n", 1083 reuse, ret); 1084 /* It is likely that rdma_resolve_addr() returns -EADDRINUSE, but 1085 * we may missing something. We rely on rdma_resolve_addr(). 1086 */ 1087 } 1088 } 1089 1090 ret = rdma_resolve_addr(rqpair->cm_id, src_addr, dst_addr, 1091 NVME_RDMA_TIME_OUT_IN_MS); 1092 if (ret) { 1093 SPDK_ERRLOG("rdma_resolve_addr, %d\n", errno); 1094 return ret; 1095 } 1096 1097 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ADDR_RESOLVED, 1098 nvme_rdma_addr_resolved); 1099 } 1100 1101 static int nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair); 1102 1103 static int 1104 nvme_rdma_connect_established(struct nvme_rdma_qpair *rqpair, int ret) 1105 { 1106 struct nvme_rdma_rsp_opts opts = {}; 1107 1108 if (ret == -ESTALE) { 1109 return nvme_rdma_stale_conn_retry(rqpair); 1110 } else if (ret) { 1111 SPDK_ERRLOG("RDMA connect error %d\n", ret); 1112 return ret; 1113 } 1114 1115 assert(!rqpair->mr_map); 1116 rqpair->mr_map = spdk_rdma_utils_create_mem_map(rqpair->rdma_qp->qp->pd, &g_nvme_hooks, 1117 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE); 1118 if (!rqpair->mr_map) { 1119 SPDK_ERRLOG("Unable to register RDMA memory translation map\n"); 1120 return -1; 1121 } 1122 1123 ret = nvme_rdma_create_reqs(rqpair); 1124 SPDK_DEBUGLOG(nvme, "rc =%d\n", ret); 1125 if (ret) { 1126 SPDK_ERRLOG("Unable to create rqpair RDMA requests\n"); 1127 return -1; 1128 } 1129 SPDK_DEBUGLOG(nvme, "RDMA requests created\n"); 1130 1131 if (!rqpair->srq) { 1132 opts.num_entries = rqpair->num_entries; 1133 opts.rqpair = rqpair; 1134 opts.srq = NULL; 1135 opts.mr_map = rqpair->mr_map; 1136 1137 assert(!rqpair->rsps); 1138 rqpair->rsps = nvme_rdma_create_rsps(&opts); 1139 if (!rqpair->rsps) { 1140 SPDK_ERRLOG("Unable to create rqpair RDMA responses\n"); 1141 return -1; 1142 } 1143 SPDK_DEBUGLOG(nvme, "RDMA responses created\n"); 1144 1145 ret = nvme_rdma_qpair_submit_recvs(rqpair); 1146 SPDK_DEBUGLOG(nvme, "rc =%d\n", ret); 1147 if (ret) { 1148 SPDK_ERRLOG("Unable to submit rqpair RDMA responses\n"); 1149 return -1; 1150 } 1151 SPDK_DEBUGLOG(nvme, "RDMA responses submitted\n"); 1152 } 1153 1154 rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND; 1155 1156 return 0; 1157 } 1158 1159 static int 1160 nvme_rdma_connect(struct nvme_rdma_qpair *rqpair) 1161 { 1162 struct rdma_conn_param param = {}; 1163 struct spdk_nvmf_rdma_request_private_data request_data = {}; 1164 struct ibv_device_attr attr; 1165 int ret; 1166 struct spdk_nvme_ctrlr *ctrlr; 1167 1168 ret = ibv_query_device(rqpair->cm_id->verbs, &attr); 1169 if (ret != 0) { 1170 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1171 return ret; 1172 } 1173 1174 param.responder_resources = attr.max_qp_rd_atom; 1175 1176 ctrlr = rqpair->qpair.ctrlr; 1177 if (!ctrlr) { 1178 return -1; 1179 } 1180 1181 request_data.qid = rqpair->qpair.id; 1182 request_data.hrqsize = rqpair->num_entries + 1; 1183 request_data.hsqsize = rqpair->num_entries; 1184 request_data.cntlid = ctrlr->cntlid; 1185 1186 param.private_data = &request_data; 1187 param.private_data_len = sizeof(request_data); 1188 param.retry_count = ctrlr->opts.transport_retry_count; 1189 param.rnr_retry_count = 7; 1190 1191 /* Fields below are ignored by rdma cm if qpair has been 1192 * created using rdma cm API. */ 1193 param.srq = 0; 1194 param.qp_num = rqpair->rdma_qp->qp->qp_num; 1195 1196 ret = rdma_connect(rqpair->cm_id, ¶m); 1197 if (ret) { 1198 SPDK_ERRLOG("nvme rdma connect error\n"); 1199 return ret; 1200 } 1201 1202 ctrlr->numa.id_valid = 1; 1203 ctrlr->numa.id = spdk_rdma_cm_id_get_numa_id(rqpair->cm_id); 1204 1205 return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ESTABLISHED, 1206 nvme_rdma_connect_established); 1207 } 1208 1209 static int 1210 nvme_rdma_ctrlr_connect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 1211 { 1212 struct sockaddr_storage dst_addr; 1213 struct sockaddr_storage src_addr; 1214 bool src_addr_specified; 1215 long int port, src_port = 0; 1216 int rc; 1217 struct nvme_rdma_ctrlr *rctrlr; 1218 struct nvme_rdma_qpair *rqpair; 1219 struct nvme_rdma_poll_group *group; 1220 int family; 1221 1222 rqpair = nvme_rdma_qpair(qpair); 1223 rctrlr = nvme_rdma_ctrlr(ctrlr); 1224 assert(rctrlr != NULL); 1225 1226 switch (ctrlr->trid.adrfam) { 1227 case SPDK_NVMF_ADRFAM_IPV4: 1228 family = AF_INET; 1229 break; 1230 case SPDK_NVMF_ADRFAM_IPV6: 1231 family = AF_INET6; 1232 break; 1233 default: 1234 SPDK_ERRLOG("Unhandled ADRFAM %d\n", ctrlr->trid.adrfam); 1235 return -1; 1236 } 1237 1238 SPDK_DEBUGLOG(nvme, "adrfam %d ai_family %d\n", ctrlr->trid.adrfam, family); 1239 1240 memset(&dst_addr, 0, sizeof(dst_addr)); 1241 1242 SPDK_DEBUGLOG(nvme, "trsvcid is %s\n", ctrlr->trid.trsvcid); 1243 rc = nvme_parse_addr(&dst_addr, family, ctrlr->trid.traddr, ctrlr->trid.trsvcid, &port); 1244 if (rc != 0) { 1245 SPDK_ERRLOG("dst_addr nvme_parse_addr() failed\n"); 1246 return -1; 1247 } 1248 1249 if (ctrlr->opts.src_addr[0] || ctrlr->opts.src_svcid[0]) { 1250 memset(&src_addr, 0, sizeof(src_addr)); 1251 rc = nvme_parse_addr(&src_addr, family, 1252 ctrlr->opts.src_addr[0] ? ctrlr->opts.src_addr : NULL, 1253 ctrlr->opts.src_svcid[0] ? ctrlr->opts.src_svcid : NULL, 1254 &src_port); 1255 if (rc != 0) { 1256 SPDK_ERRLOG("src_addr nvme_parse_addr() failed\n"); 1257 return -1; 1258 } 1259 src_addr_specified = true; 1260 } else { 1261 src_addr_specified = false; 1262 } 1263 1264 rc = rdma_create_id(rctrlr->cm_channel, &rqpair->cm_id, rqpair, RDMA_PS_TCP); 1265 if (rc < 0) { 1266 SPDK_ERRLOG("rdma_create_id() failed\n"); 1267 return -1; 1268 } 1269 1270 rc = nvme_rdma_resolve_addr(rqpair, 1271 src_addr_specified ? (struct sockaddr *)&src_addr : NULL, 1272 (struct sockaddr *)&dst_addr); 1273 if (rc < 0) { 1274 SPDK_ERRLOG("nvme_rdma_resolve_addr() failed\n"); 1275 return -1; 1276 } 1277 1278 rqpair->state = NVME_RDMA_QPAIR_STATE_INITIALIZING; 1279 1280 if (qpair->poll_group != NULL && rqpair->link_connecting.tqe_prev == NULL) { 1281 group = nvme_rdma_poll_group(qpair->poll_group); 1282 TAILQ_INSERT_TAIL(&group->connecting_qpairs, rqpair, link_connecting); 1283 } 1284 1285 return 0; 1286 } 1287 1288 static int 1289 nvme_rdma_stale_conn_reconnect(struct nvme_rdma_qpair *rqpair) 1290 { 1291 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 1292 1293 if (spdk_get_ticks() < rqpair->evt_timeout_ticks) { 1294 return -EAGAIN; 1295 } 1296 1297 return nvme_rdma_ctrlr_connect_qpair(qpair->ctrlr, qpair); 1298 } 1299 1300 static int 1301 nvme_rdma_ctrlr_connect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, 1302 struct spdk_nvme_qpair *qpair) 1303 { 1304 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 1305 int rc; 1306 1307 if (rqpair->in_connect_poll) { 1308 return -EAGAIN; 1309 } 1310 1311 rqpair->in_connect_poll = true; 1312 1313 switch (rqpair->state) { 1314 case NVME_RDMA_QPAIR_STATE_INVALID: 1315 rc = -EAGAIN; 1316 break; 1317 1318 case NVME_RDMA_QPAIR_STATE_INITIALIZING: 1319 case NVME_RDMA_QPAIR_STATE_EXITING: 1320 if (!nvme_qpair_is_admin_queue(qpair)) { 1321 nvme_ctrlr_lock(ctrlr); 1322 } 1323 1324 rc = nvme_rdma_process_event_poll(rqpair); 1325 1326 if (!nvme_qpair_is_admin_queue(qpair)) { 1327 nvme_ctrlr_unlock(ctrlr); 1328 } 1329 1330 if (rc == 0) { 1331 rc = -EAGAIN; 1332 } 1333 rqpair->in_connect_poll = false; 1334 1335 return rc; 1336 1337 case NVME_RDMA_QPAIR_STATE_STALE_CONN: 1338 rc = nvme_rdma_stale_conn_reconnect(rqpair); 1339 if (rc == 0) { 1340 rc = -EAGAIN; 1341 } 1342 break; 1343 case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND: 1344 rc = nvme_fabric_qpair_connect_async(qpair, rqpair->num_entries + 1); 1345 if (rc == 0) { 1346 rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL; 1347 rc = -EAGAIN; 1348 } else { 1349 SPDK_ERRLOG("Failed to send an NVMe-oF Fabric CONNECT command\n"); 1350 } 1351 break; 1352 case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL: 1353 rc = nvme_fabric_qpair_connect_poll(qpair); 1354 if (rc == 0) { 1355 if (nvme_fabric_qpair_auth_required(qpair)) { 1356 rc = nvme_fabric_qpair_authenticate_async(qpair); 1357 if (rc == 0) { 1358 rqpair->state = NVME_RDMA_QPAIR_STATE_AUTHENTICATING; 1359 rc = -EAGAIN; 1360 } 1361 } else { 1362 rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING; 1363 nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED); 1364 } 1365 } else if (rc != -EAGAIN) { 1366 SPDK_ERRLOG("Failed to poll NVMe-oF Fabric CONNECT command\n"); 1367 } 1368 break; 1369 case NVME_RDMA_QPAIR_STATE_AUTHENTICATING: 1370 rc = nvme_fabric_qpair_authenticate_poll(qpair); 1371 if (rc == 0) { 1372 rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING; 1373 nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED); 1374 } 1375 break; 1376 case NVME_RDMA_QPAIR_STATE_RUNNING: 1377 rc = 0; 1378 break; 1379 default: 1380 assert(false); 1381 rc = -EINVAL; 1382 break; 1383 } 1384 1385 rqpair->in_connect_poll = false; 1386 1387 return rc; 1388 } 1389 1390 static inline int 1391 nvme_rdma_get_memory_translation(struct nvme_request *req, struct nvme_rdma_qpair *rqpair, 1392 struct nvme_rdma_memory_translation_ctx *_ctx) 1393 { 1394 struct spdk_memory_domain_translation_ctx ctx; 1395 struct spdk_memory_domain_translation_result dma_translation = {.iov_count = 0}; 1396 struct spdk_rdma_utils_memory_translation rdma_translation; 1397 int rc; 1398 1399 assert(req); 1400 assert(rqpair); 1401 assert(_ctx); 1402 1403 if (req->payload.opts && req->payload.opts->memory_domain) { 1404 ctx.size = sizeof(struct spdk_memory_domain_translation_ctx); 1405 ctx.rdma.ibv_qp = rqpair->rdma_qp->qp; 1406 dma_translation.size = sizeof(struct spdk_memory_domain_translation_result); 1407 1408 rc = spdk_memory_domain_translate_data(req->payload.opts->memory_domain, 1409 req->payload.opts->memory_domain_ctx, 1410 rqpair->rdma_qp->domain, &ctx, _ctx->addr, 1411 _ctx->length, &dma_translation); 1412 if (spdk_unlikely(rc) || dma_translation.iov_count != 1) { 1413 SPDK_ERRLOG("DMA memory translation failed, rc %d, iov count %u\n", rc, dma_translation.iov_count); 1414 return rc; 1415 } 1416 1417 _ctx->lkey = dma_translation.rdma.lkey; 1418 _ctx->rkey = dma_translation.rdma.rkey; 1419 _ctx->addr = dma_translation.iov.iov_base; 1420 _ctx->length = dma_translation.iov.iov_len; 1421 } else { 1422 rc = spdk_rdma_utils_get_translation(rqpair->mr_map, _ctx->addr, _ctx->length, &rdma_translation); 1423 if (spdk_unlikely(rc)) { 1424 SPDK_ERRLOG("RDMA memory translation failed, rc %d\n", rc); 1425 return rc; 1426 } 1427 if (rdma_translation.translation_type == SPDK_RDMA_UTILS_TRANSLATION_MR) { 1428 _ctx->lkey = rdma_translation.mr_or_key.mr->lkey; 1429 _ctx->rkey = rdma_translation.mr_or_key.mr->rkey; 1430 } else { 1431 _ctx->lkey = _ctx->rkey = (uint32_t)rdma_translation.mr_or_key.key; 1432 } 1433 } 1434 1435 return 0; 1436 } 1437 1438 1439 /* 1440 * Build SGL describing empty payload. 1441 */ 1442 static int 1443 nvme_rdma_build_null_request(struct spdk_nvme_rdma_req *rdma_req) 1444 { 1445 struct nvme_request *req = rdma_req->req; 1446 1447 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1448 1449 /* The first element of this SGL is pointing at an 1450 * spdk_nvmf_cmd object. For this particular command, 1451 * we only need the first 64 bytes corresponding to 1452 * the NVMe command. */ 1453 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1454 1455 /* The RDMA SGL needs one element describing the NVMe command. */ 1456 rdma_req->send_wr.num_sge = 1; 1457 1458 req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1459 req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1460 req->cmd.dptr.sgl1.keyed.length = 0; 1461 req->cmd.dptr.sgl1.keyed.key = 0; 1462 req->cmd.dptr.sgl1.address = 0; 1463 1464 return 0; 1465 } 1466 1467 static inline void 1468 nvme_rdma_configure_contig_inline_request(struct spdk_nvme_rdma_req *rdma_req, 1469 struct nvme_request *req, struct nvme_rdma_memory_translation_ctx *ctx) 1470 { 1471 rdma_req->send_sgl[1].lkey = ctx->lkey; 1472 1473 /* The first element of this SGL is pointing at an 1474 * spdk_nvmf_cmd object. For this particular command, 1475 * we only need the first 64 bytes corresponding to 1476 * the NVMe command. */ 1477 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1478 1479 rdma_req->send_sgl[1].addr = (uint64_t)ctx->addr; 1480 rdma_req->send_sgl[1].length = (uint32_t)ctx->length; 1481 1482 /* The RDMA SGL contains two elements. The first describes 1483 * the NVMe command and the second describes the data 1484 * payload. */ 1485 rdma_req->send_wr.num_sge = 2; 1486 1487 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1488 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK; 1489 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1490 req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx->length; 1491 /* Inline only supported for icdoff == 0 currently. This function will 1492 * not get called for controllers with other values. */ 1493 req->cmd.dptr.sgl1.address = (uint64_t)0; 1494 } 1495 1496 /* 1497 * Build inline SGL describing contiguous payload buffer. 1498 */ 1499 static inline int 1500 nvme_rdma_build_contig_inline_request(struct nvme_rdma_qpair *rqpair, 1501 struct spdk_nvme_rdma_req *rdma_req) 1502 { 1503 struct nvme_request *req = rdma_req->req; 1504 struct nvme_rdma_memory_translation_ctx ctx = { 1505 .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset, 1506 .length = req->payload_size 1507 }; 1508 int rc; 1509 1510 assert(ctx.length != 0); 1511 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG); 1512 1513 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1514 if (spdk_unlikely(rc)) { 1515 return -1; 1516 } 1517 1518 nvme_rdma_configure_contig_inline_request(rdma_req, req, &ctx); 1519 1520 return 0; 1521 } 1522 1523 static inline void 1524 nvme_rdma_configure_contig_request(struct spdk_nvme_rdma_req *rdma_req, struct nvme_request *req, 1525 struct nvme_rdma_memory_translation_ctx *ctx) 1526 { 1527 req->cmd.dptr.sgl1.keyed.key = ctx->rkey; 1528 1529 /* The first element of this SGL is pointing at an 1530 * spdk_nvmf_cmd object. For this particular command, 1531 * we only need the first 64 bytes corresponding to 1532 * the NVMe command. */ 1533 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1534 1535 /* The RDMA SGL needs one element describing the NVMe command. */ 1536 rdma_req->send_wr.num_sge = 1; 1537 1538 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1539 req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1540 req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1541 req->cmd.dptr.sgl1.keyed.length = (uint32_t)ctx->length; 1542 req->cmd.dptr.sgl1.address = (uint64_t)ctx->addr; 1543 } 1544 1545 /* 1546 * Build SGL describing contiguous payload buffer. 1547 */ 1548 static inline int 1549 nvme_rdma_build_contig_request(struct nvme_rdma_qpair *rqpair, 1550 struct spdk_nvme_rdma_req *rdma_req) 1551 { 1552 struct nvme_request *req = rdma_req->req; 1553 struct nvme_rdma_memory_translation_ctx ctx = { 1554 .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset, 1555 .length = req->payload_size 1556 }; 1557 int rc; 1558 1559 assert(req->payload_size != 0); 1560 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG); 1561 1562 if (spdk_unlikely(req->payload_size > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) { 1563 SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n", 1564 req->payload_size, NVME_RDMA_MAX_KEYED_SGL_LENGTH); 1565 return -1; 1566 } 1567 1568 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1569 if (spdk_unlikely(rc)) { 1570 return -1; 1571 } 1572 1573 nvme_rdma_configure_contig_request(rdma_req, req, &ctx); 1574 1575 return 0; 1576 } 1577 1578 /* 1579 * Build SGL describing scattered payload buffer. 1580 */ 1581 static inline int 1582 nvme_rdma_build_sgl_request(struct nvme_rdma_qpair *rqpair, 1583 struct spdk_nvme_rdma_req *rdma_req) 1584 { 1585 struct nvme_request *req = rdma_req->req; 1586 struct spdk_nvmf_cmd *cmd = &rqpair->cmds[rdma_req->id]; 1587 struct nvme_rdma_memory_translation_ctx ctx; 1588 uint32_t remaining_size; 1589 uint32_t sge_length; 1590 int rc, max_num_sgl, num_sgl_desc; 1591 1592 assert(req->payload_size != 0); 1593 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL); 1594 assert(req->payload.reset_sgl_fn != NULL); 1595 assert(req->payload.next_sge_fn != NULL); 1596 req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset); 1597 1598 max_num_sgl = req->qpair->ctrlr->max_sges; 1599 1600 remaining_size = req->payload_size; 1601 num_sgl_desc = 0; 1602 do { 1603 rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &sge_length); 1604 if (spdk_unlikely(rc)) { 1605 return -1; 1606 } 1607 1608 sge_length = spdk_min(remaining_size, sge_length); 1609 1610 if (spdk_unlikely(sge_length > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) { 1611 SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n", 1612 sge_length, NVME_RDMA_MAX_KEYED_SGL_LENGTH); 1613 return -1; 1614 } 1615 ctx.length = sge_length; 1616 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1617 if (spdk_unlikely(rc)) { 1618 return -1; 1619 } 1620 1621 cmd->sgl[num_sgl_desc].keyed.key = ctx.rkey; 1622 cmd->sgl[num_sgl_desc].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1623 cmd->sgl[num_sgl_desc].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1624 cmd->sgl[num_sgl_desc].keyed.length = (uint32_t)ctx.length; 1625 cmd->sgl[num_sgl_desc].address = (uint64_t)ctx.addr; 1626 1627 remaining_size -= ctx.length; 1628 num_sgl_desc++; 1629 } while (remaining_size > 0 && num_sgl_desc < max_num_sgl); 1630 1631 1632 /* Should be impossible if we did our sgl checks properly up the stack, but do a sanity check here. */ 1633 if (spdk_unlikely(remaining_size > 0)) { 1634 return -1; 1635 } 1636 1637 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1638 1639 /* The RDMA SGL needs one element describing some portion 1640 * of the spdk_nvmf_cmd structure. */ 1641 rdma_req->send_wr.num_sge = 1; 1642 1643 /* 1644 * If only one SGL descriptor is required, it can be embedded directly in the command 1645 * as a data block descriptor. 1646 */ 1647 if (num_sgl_desc == 1) { 1648 /* The first element of this SGL is pointing at an 1649 * spdk_nvmf_cmd object. For this particular command, 1650 * we only need the first 64 bytes corresponding to 1651 * the NVMe command. */ 1652 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1653 1654 req->cmd.dptr.sgl1.keyed.type = cmd->sgl[0].keyed.type; 1655 req->cmd.dptr.sgl1.keyed.subtype = cmd->sgl[0].keyed.subtype; 1656 req->cmd.dptr.sgl1.keyed.length = cmd->sgl[0].keyed.length; 1657 req->cmd.dptr.sgl1.keyed.key = cmd->sgl[0].keyed.key; 1658 req->cmd.dptr.sgl1.address = cmd->sgl[0].address; 1659 } else { 1660 /* 1661 * Otherwise, The SGL descriptor embedded in the command must point to the list of 1662 * SGL descriptors used to describe the operation. In that case it is a last segment descriptor. 1663 */ 1664 uint32_t descriptors_size = sizeof(struct spdk_nvme_sgl_descriptor) * num_sgl_desc; 1665 1666 if (spdk_unlikely(descriptors_size > rqpair->qpair.ctrlr->ioccsz_bytes)) { 1667 SPDK_ERRLOG("Size of SGL descriptors (%u) exceeds ICD (%u)\n", 1668 descriptors_size, rqpair->qpair.ctrlr->ioccsz_bytes); 1669 return -1; 1670 } 1671 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd) + descriptors_size; 1672 1673 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT; 1674 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1675 req->cmd.dptr.sgl1.unkeyed.length = descriptors_size; 1676 req->cmd.dptr.sgl1.address = (uint64_t)0; 1677 } 1678 1679 return 0; 1680 } 1681 1682 /* 1683 * Build inline SGL describing sgl payload buffer. 1684 */ 1685 static inline int 1686 nvme_rdma_build_sgl_inline_request(struct nvme_rdma_qpair *rqpair, 1687 struct spdk_nvme_rdma_req *rdma_req) 1688 { 1689 struct nvme_request *req = rdma_req->req; 1690 struct nvme_rdma_memory_translation_ctx ctx; 1691 uint32_t length; 1692 int rc; 1693 1694 assert(req->payload_size != 0); 1695 assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL); 1696 assert(req->payload.reset_sgl_fn != NULL); 1697 assert(req->payload.next_sge_fn != NULL); 1698 req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset); 1699 1700 rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &length); 1701 if (spdk_unlikely(rc)) { 1702 return -1; 1703 } 1704 1705 if (length < req->payload_size) { 1706 SPDK_DEBUGLOG(nvme, "Inline SGL request split so sending separately.\n"); 1707 return nvme_rdma_build_sgl_request(rqpair, rdma_req); 1708 } 1709 1710 if (length > req->payload_size) { 1711 length = req->payload_size; 1712 } 1713 1714 ctx.length = length; 1715 rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx); 1716 if (spdk_unlikely(rc)) { 1717 return -1; 1718 } 1719 1720 rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr; 1721 rdma_req->send_sgl[1].length = (uint32_t)ctx.length; 1722 rdma_req->send_sgl[1].lkey = ctx.lkey; 1723 1724 rdma_req->send_wr.num_sge = 2; 1725 1726 /* The first element of this SGL is pointing at an 1727 * spdk_nvmf_cmd object. For this particular command, 1728 * we only need the first 64 bytes corresponding to 1729 * the NVMe command. */ 1730 rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd); 1731 1732 req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG; 1733 req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK; 1734 req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1735 req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length; 1736 /* Inline only supported for icdoff == 0 currently. This function will 1737 * not get called for controllers with other values. */ 1738 req->cmd.dptr.sgl1.address = (uint64_t)0; 1739 1740 return 0; 1741 } 1742 1743 static inline int 1744 nvme_rdma_accel_append_copy(struct spdk_nvme_poll_group *pg, void **seq, 1745 struct spdk_memory_domain *rdma_domain, struct spdk_nvme_rdma_req *rdma_req, 1746 struct iovec *iovs, uint32_t iovcnt, 1747 struct spdk_memory_domain *src_domain, void *src_domain_ctx) 1748 { 1749 return pg->accel_fn_table.append_copy(pg->ctx, seq, iovs, iovcnt, rdma_domain, rdma_req, iovs, 1750 iovcnt, src_domain, src_domain_ctx, NULL, NULL); 1751 } 1752 1753 static inline void 1754 nvme_rdma_accel_reverse(struct spdk_nvme_poll_group *pg, void *seq) 1755 { 1756 pg->accel_fn_table.reverse_sequence(seq); 1757 } 1758 1759 static inline void 1760 nvme_rdma_accel_finish(struct spdk_nvme_poll_group *pg, void *seq, 1761 spdk_nvme_accel_completion_cb cb_fn, void *cb_arg) 1762 { 1763 pg->accel_fn_table.finish_sequence(seq, cb_fn, cb_arg); 1764 } 1765 1766 static inline void 1767 nvme_rdma_accel_completion_cb(void *cb_arg, int status) 1768 { 1769 struct spdk_nvme_rdma_req *rdma_req = cb_arg; 1770 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(rdma_req->req->qpair); 1771 struct spdk_nvme_cpl cpl; 1772 enum spdk_nvme_generic_command_status_code sc; 1773 uint16_t dnr = 0; 1774 1775 rdma_req->in_progress_accel = 0; 1776 rdma_req->req->accel_sequence = NULL; 1777 SPDK_DEBUGLOG(nvme, "rdma_req %p qpair %p, accel completion rc %d\n", rdma_req, rqpair, status); 1778 1779 /* nvme_rdma driver may fail data transfer on WC_FLUSH error completion which is expected. 1780 * To prevent false errors from accel, first check if qpair is in the process of disconnect */ 1781 if (spdk_unlikely(!spdk_nvme_qpair_is_connected(&rqpair->qpair))) { 1782 struct spdk_nvmf_fabric_connect_cmd *cmd = (struct spdk_nvmf_fabric_connect_cmd *) 1783 &rdma_req->req->cmd; 1784 1785 if (cmd->opcode != SPDK_NVME_OPC_FABRIC && cmd->fctype != SPDK_NVMF_FABRIC_COMMAND_CONNECT) { 1786 SPDK_DEBUGLOG(nvme, "qpair %p, req %p accel cpl in disconnecting, outstanding %u\n", 1787 rqpair, rdma_req, rqpair->qpair.num_outstanding_reqs); 1788 sc = SPDK_NVME_SC_ABORTED_SQ_DELETION; 1789 goto fail_req; 1790 } 1791 } 1792 if (spdk_unlikely(status)) { 1793 SPDK_ERRLOG("qpair %p, req %p, accel sequence status %d\n", rdma_req->req->qpair, rdma_req, status); 1794 sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1795 /* Something wrong happened, let the upper layer know that retry is no desired */ 1796 dnr = 1; 1797 goto fail_req; 1798 } 1799 1800 nvme_rdma_req_complete(rdma_req, &rdma_req->rdma_rsp->cpl, true); 1801 return; 1802 1803 fail_req: 1804 memset(&cpl, 0, sizeof(cpl)); 1805 cpl.status.sc = sc; 1806 cpl.status.sct = SPDK_NVME_SCT_GENERIC; 1807 cpl.status.dnr = dnr; 1808 nvme_rdma_req_complete(rdma_req, &cpl, true); 1809 } 1810 1811 static inline int 1812 nvme_rdma_apply_accel_sequence(struct nvme_rdma_qpair *rqpair, struct nvme_request *req, 1813 struct spdk_nvme_rdma_req *rdma_req) 1814 { 1815 struct spdk_nvme_poll_group *pg = rqpair->qpair.poll_group->group; 1816 struct spdk_memory_domain *src_domain; 1817 void *src_domain_ctx; 1818 void *accel_seq = req->accel_sequence; 1819 uint32_t iovcnt = 0; 1820 int rc; 1821 1822 SPDK_DEBUGLOG(nvme, "req %p, start accel seq %p\n", rdma_req, accel_seq); 1823 if (nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL) { 1824 void *addr; 1825 uint32_t sge_length, payload_size; 1826 1827 payload_size = req->payload_size; 1828 assert(payload_size); 1829 req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset); 1830 do { 1831 rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &addr, &sge_length); 1832 if (spdk_unlikely(rc)) { 1833 return -1; 1834 } 1835 sge_length = spdk_min(payload_size, sge_length); 1836 rdma_req->iovs[iovcnt].iov_base = addr; 1837 rdma_req->iovs[iovcnt].iov_len = sge_length; 1838 iovcnt++; 1839 payload_size -= sge_length; 1840 } while (payload_size && iovcnt < NVME_RDMA_MAX_SGL_DESCRIPTORS); 1841 1842 if (spdk_unlikely(payload_size)) { 1843 SPDK_ERRLOG("not enough iovs to handle req %p, remaining len %u\n", rdma_req, payload_size); 1844 return -E2BIG; 1845 } 1846 } else { 1847 rdma_req->iovs[iovcnt].iov_base = req->payload.contig_or_cb_arg; 1848 rdma_req->iovs[iovcnt].iov_len = req->payload_size; 1849 iovcnt = 1; 1850 } 1851 if (req->payload.opts && req->payload.opts->memory_domain) { 1852 if (accel_seq) { 1853 src_domain = rqpair->rdma_qp->domain; 1854 src_domain_ctx = rdma_req; 1855 } else { 1856 src_domain = req->payload.opts->memory_domain; 1857 src_domain_ctx = req->payload.opts->memory_domain_ctx; 1858 } 1859 } else { 1860 src_domain = NULL; 1861 src_domain_ctx = NULL; 1862 } 1863 1864 rc = nvme_rdma_accel_append_copy(pg, &accel_seq, rqpair->rdma_qp->domain, rdma_req, rdma_req->iovs, 1865 iovcnt, src_domain, src_domain_ctx); 1866 if (spdk_unlikely(rc)) { 1867 return rc; 1868 } 1869 1870 if (spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1871 nvme_rdma_accel_reverse(pg, accel_seq); 1872 } 1873 1874 rdma_req->in_progress_accel = 1; 1875 TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link); 1876 rqpair->num_outstanding_reqs++; 1877 1878 SPDK_DEBUGLOG(nvme, "req %p, finish accel seq %p\n", rdma_req, accel_seq); 1879 nvme_rdma_accel_finish(pg, accel_seq, nvme_rdma_accel_completion_cb, rdma_req); 1880 1881 return 0; 1882 } 1883 1884 static inline int 1885 nvme_rdma_memory_domain_transfer_data(struct spdk_memory_domain *dst_domain, void *dst_domain_ctx, 1886 struct iovec *dst_iov, uint32_t dst_iovcnt, 1887 struct spdk_memory_domain *src_domain, void *src_domain_ctx, 1888 struct iovec *src_iov, uint32_t src_iovcnt, 1889 struct spdk_memory_domain_translation_result *translation, 1890 spdk_memory_domain_data_cpl_cb cpl_cb, void *cpl_cb_arg) 1891 { 1892 struct nvme_rdma_memory_translation_ctx ctx; 1893 struct spdk_nvme_rdma_req *rdma_req = dst_domain_ctx; 1894 struct nvme_request *req = rdma_req->req; 1895 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(rdma_req->req->qpair); 1896 struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr; 1897 bool icd_supported; 1898 1899 assert(dst_domain == rqpair->rdma_qp->domain); 1900 assert(src_domain); 1901 assert(spdk_memory_domain_get_dma_device_type(src_domain) == SPDK_DMA_DEVICE_TYPE_RDMA); 1902 /* We expect "inplace" operation */ 1903 assert(dst_iov == src_iov); 1904 assert(dst_iovcnt == src_iovcnt); 1905 1906 if (spdk_unlikely(!src_domain || 1907 spdk_memory_domain_get_dma_device_type(src_domain) != SPDK_DMA_DEVICE_TYPE_RDMA)) { 1908 SPDK_ERRLOG("Unexpected source memory domain %p, type %d\n", src_domain, 1909 src_domain ? (int)spdk_memory_domain_get_dma_device_type(src_domain) : -1); 1910 return -ENOTSUP; 1911 } 1912 if (spdk_unlikely(dst_iovcnt != 1 || !translation || translation->iov_count != 1)) { 1913 SPDK_ERRLOG("Unexpected iovcnt %u or missed translation, rdma_req %p\n", dst_iovcnt, rdma_req); 1914 return -ENOTSUP; 1915 } 1916 ctx.addr = translation->iov.iov_base; 1917 ctx.length = translation->iov.iov_len; 1918 ctx.lkey = translation->rdma.lkey; 1919 ctx.rkey = translation->rdma.rkey; 1920 1921 SPDK_DEBUGLOG(nvme, "req %p, addr %p, len %zu, key %u\n", rdma_req, ctx.addr, ctx.length, ctx.rkey); 1922 icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER 1923 && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0; 1924 1925 /* We expect that result of accel sequence is a Memory Key which describes a virtually contig address space. 1926 * That means we prepare a contig request even if original payload was scattered */ 1927 if (icd_supported) { 1928 nvme_rdma_configure_contig_inline_request(rdma_req, req, &ctx); 1929 } else { 1930 nvme_rdma_configure_contig_request(rdma_req, req, &ctx); 1931 } 1932 rdma_req->transfer_cpl_cb = cpl_cb; 1933 rdma_req->transfer_cpl_cb_arg = cpl_cb_arg; 1934 1935 memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd)); 1936 1937 return _nvme_rdma_qpair_submit_request(rqpair, rdma_req); 1938 } 1939 1940 static inline int 1941 nvme_rdma_req_init(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req) 1942 { 1943 struct nvme_request *req = rdma_req->req; 1944 struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr; 1945 enum nvme_payload_type payload_type; 1946 bool icd_supported; 1947 int rc = -1; 1948 1949 payload_type = nvme_payload_type(&req->payload); 1950 /* 1951 * Check if icdoff is non zero, to avoid interop conflicts with 1952 * targets with non-zero icdoff. Both SPDK and the Linux kernel 1953 * targets use icdoff = 0. For targets with non-zero icdoff, we 1954 * will currently just not use inline data for now. 1955 */ 1956 icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER 1957 && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0; 1958 1959 if (spdk_unlikely(req->payload_size == 0)) { 1960 rc = nvme_rdma_build_null_request(rdma_req); 1961 } else if (payload_type == NVME_PAYLOAD_TYPE_CONTIG) { 1962 if (icd_supported) { 1963 rc = nvme_rdma_build_contig_inline_request(rqpair, rdma_req); 1964 } else { 1965 rc = nvme_rdma_build_contig_request(rqpair, rdma_req); 1966 } 1967 } else if (payload_type == NVME_PAYLOAD_TYPE_SGL) { 1968 if (icd_supported) { 1969 rc = nvme_rdma_build_sgl_inline_request(rqpair, rdma_req); 1970 } else { 1971 rc = nvme_rdma_build_sgl_request(rqpair, rdma_req); 1972 } 1973 } 1974 1975 if (spdk_unlikely(rc)) { 1976 return rc; 1977 } 1978 1979 memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd)); 1980 return 0; 1981 } 1982 1983 static struct spdk_nvme_qpair * 1984 nvme_rdma_ctrlr_create_qpair(struct spdk_nvme_ctrlr *ctrlr, 1985 uint16_t qid, uint32_t qsize, 1986 enum spdk_nvme_qprio qprio, 1987 uint32_t num_requests, 1988 bool delay_cmd_submit, 1989 bool async) 1990 { 1991 struct nvme_rdma_qpair *rqpair; 1992 struct spdk_nvme_qpair *qpair; 1993 int rc; 1994 1995 if (qsize < SPDK_NVME_QUEUE_MIN_ENTRIES) { 1996 SPDK_ERRLOG("Failed to create qpair with size %u. Minimum queue size is %d.\n", 1997 qsize, SPDK_NVME_QUEUE_MIN_ENTRIES); 1998 return NULL; 1999 } 2000 2001 rqpair = spdk_zmalloc(sizeof(struct nvme_rdma_qpair), 0, NULL, SPDK_ENV_NUMA_ID_ANY, 2002 SPDK_MALLOC_DMA); 2003 if (!rqpair) { 2004 SPDK_ERRLOG("failed to get create rqpair\n"); 2005 return NULL; 2006 } 2007 2008 /* Set num_entries one less than queue size. According to NVMe 2009 * and NVMe-oF specs we can not submit queue size requests, 2010 * one slot shall always remain empty. 2011 */ 2012 rqpair->num_entries = qsize - 1; 2013 rqpair->delay_cmd_submit = delay_cmd_submit; 2014 rqpair->state = NVME_RDMA_QPAIR_STATE_INVALID; 2015 rqpair->append_copy = g_spdk_nvme_transport_opts.rdma_umr_per_io && 2016 spdk_rdma_provider_accel_sequence_supported() && qid != 0; 2017 SPDK_DEBUGLOG(nvme, "rqpair %p, append_copy %s\n", rqpair, 2018 rqpair->append_copy ? "enabled" : "diabled"); 2019 qpair = &rqpair->qpair; 2020 rc = nvme_qpair_init(qpair, qid, ctrlr, qprio, num_requests, async); 2021 if (rc != 0) { 2022 spdk_free(rqpair); 2023 return NULL; 2024 } 2025 2026 return qpair; 2027 } 2028 2029 static void 2030 nvme_rdma_qpair_destroy(struct nvme_rdma_qpair *rqpair) 2031 { 2032 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2033 struct nvme_rdma_ctrlr *rctrlr; 2034 struct nvme_rdma_cm_event_entry *entry, *tmp; 2035 2036 spdk_rdma_utils_free_mem_map(&rqpair->mr_map); 2037 2038 if (rqpair->evt) { 2039 rdma_ack_cm_event(rqpair->evt); 2040 rqpair->evt = NULL; 2041 } 2042 2043 /* 2044 * This works because we have the controller lock both in 2045 * this function and in the function where we add new events. 2046 */ 2047 if (qpair->ctrlr != NULL) { 2048 rctrlr = nvme_rdma_ctrlr(qpair->ctrlr); 2049 STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) { 2050 if (entry->evt->id->context == rqpair) { 2051 STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link); 2052 rdma_ack_cm_event(entry->evt); 2053 STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link); 2054 } 2055 } 2056 } 2057 2058 if (rqpair->cm_id) { 2059 if (rqpair->rdma_qp) { 2060 spdk_rdma_utils_put_pd(rqpair->rdma_qp->qp->pd); 2061 spdk_rdma_provider_qp_destroy(rqpair->rdma_qp); 2062 rqpair->rdma_qp = NULL; 2063 } 2064 } 2065 2066 if (rqpair->poller) { 2067 struct nvme_rdma_poll_group *group; 2068 2069 assert(qpair->poll_group); 2070 group = nvme_rdma_poll_group(qpair->poll_group); 2071 2072 nvme_rdma_poll_group_put_poller(group, rqpair->poller); 2073 2074 rqpair->poller = NULL; 2075 rqpair->cq = NULL; 2076 if (rqpair->srq) { 2077 rqpair->srq = NULL; 2078 rqpair->rsps = NULL; 2079 } 2080 } else if (rqpair->cq) { 2081 ibv_destroy_cq(rqpair->cq); 2082 rqpair->cq = NULL; 2083 } 2084 2085 nvme_rdma_free_reqs(rqpair); 2086 nvme_rdma_free_rsps(rqpair->rsps); 2087 rqpair->rsps = NULL; 2088 2089 /* destroy cm_id last so cma device will not be freed before we destroy the cq. */ 2090 if (rqpair->cm_id) { 2091 rdma_destroy_id(rqpair->cm_id); 2092 rqpair->cm_id = NULL; 2093 } 2094 } 2095 2096 static void nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr); 2097 2098 static void 2099 nvme_rdma_qpair_flush_send_wrs(struct nvme_rdma_qpair *rqpair) 2100 { 2101 struct ibv_send_wr *bad_wr = NULL; 2102 int rc; 2103 2104 rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 2105 if (rc) { 2106 nvme_rdma_reset_failed_sends(rqpair, bad_wr); 2107 } 2108 } 2109 2110 static int 2111 nvme_rdma_qpair_disconnected(struct nvme_rdma_qpair *rqpair, int ret) 2112 { 2113 if (ret) { 2114 SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n"); 2115 goto quiet; 2116 } 2117 2118 if (rqpair->poller == NULL) { 2119 /* If poller is not used, cq is not shared. 2120 * So complete disconnecting qpair immediately. 2121 */ 2122 goto quiet; 2123 } 2124 2125 if (rqpair->rsps == NULL) { 2126 goto quiet; 2127 } 2128 2129 nvme_rdma_qpair_flush_send_wrs(rqpair); 2130 2131 if (rqpair->need_destroy || 2132 (rqpair->current_num_sends != 0 || 2133 (!rqpair->srq && rqpair->rsps->current_num_recvs != 0)) || 2134 ((rqpair->qpair.ctrlr->flags & SPDK_NVME_CTRLR_ACCEL_SEQUENCE_SUPPORTED) && 2135 (!TAILQ_EMPTY(&rqpair->outstanding_reqs)))) { 2136 rqpair->state = NVME_RDMA_QPAIR_STATE_LINGERING; 2137 rqpair->evt_timeout_ticks = (NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US * spdk_get_ticks_hz()) / 2138 SPDK_SEC_TO_USEC + spdk_get_ticks(); 2139 2140 return -EAGAIN; 2141 } 2142 2143 quiet: 2144 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED; 2145 2146 nvme_rdma_qpair_abort_reqs(&rqpair->qpair, rqpair->qpair.abort_dnr); 2147 nvme_rdma_qpair_destroy(rqpair); 2148 nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair); 2149 2150 return 0; 2151 } 2152 2153 static int 2154 nvme_rdma_qpair_wait_until_quiet(struct nvme_rdma_qpair *rqpair) 2155 { 2156 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2157 struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr; 2158 2159 if (spdk_get_ticks() < rqpair->evt_timeout_ticks && 2160 (rqpair->current_num_sends != 0 || 2161 (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) { 2162 return -EAGAIN; 2163 } 2164 2165 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED; 2166 nvme_rdma_qpair_abort_reqs(qpair, qpair->abort_dnr); 2167 if (!nvme_qpair_is_admin_queue(qpair)) { 2168 nvme_robust_mutex_lock(&ctrlr->ctrlr_lock); 2169 } 2170 nvme_rdma_qpair_destroy(rqpair); 2171 if (!nvme_qpair_is_admin_queue(qpair)) { 2172 nvme_robust_mutex_unlock(&ctrlr->ctrlr_lock); 2173 } 2174 nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair); 2175 2176 return 0; 2177 } 2178 2179 static void 2180 _nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair, 2181 nvme_rdma_cm_event_cb disconnected_qpair_cb) 2182 { 2183 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2184 int rc; 2185 2186 assert(disconnected_qpair_cb != NULL); 2187 2188 rqpair->state = NVME_RDMA_QPAIR_STATE_EXITING; 2189 2190 if (rqpair->cm_id) { 2191 if (rqpair->rdma_qp) { 2192 rc = spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp); 2193 if ((qpair->ctrlr != NULL) && (rc == 0)) { 2194 rc = nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_DISCONNECTED, 2195 disconnected_qpair_cb); 2196 if (rc == 0) { 2197 return; 2198 } 2199 } 2200 } 2201 } 2202 2203 disconnected_qpair_cb(rqpair, 0); 2204 } 2205 2206 static int 2207 nvme_rdma_ctrlr_disconnect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 2208 { 2209 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2210 int rc; 2211 2212 switch (rqpair->state) { 2213 case NVME_RDMA_QPAIR_STATE_EXITING: 2214 if (!nvme_qpair_is_admin_queue(qpair)) { 2215 nvme_ctrlr_lock(ctrlr); 2216 } 2217 2218 rc = nvme_rdma_process_event_poll(rqpair); 2219 2220 if (!nvme_qpair_is_admin_queue(qpair)) { 2221 nvme_ctrlr_unlock(ctrlr); 2222 } 2223 break; 2224 2225 case NVME_RDMA_QPAIR_STATE_LINGERING: 2226 rc = nvme_rdma_qpair_wait_until_quiet(rqpair); 2227 break; 2228 case NVME_RDMA_QPAIR_STATE_EXITED: 2229 rc = 0; 2230 break; 2231 2232 default: 2233 assert(false); 2234 rc = -EAGAIN; 2235 break; 2236 } 2237 2238 return rc; 2239 } 2240 2241 static void 2242 nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 2243 { 2244 int rc; 2245 2246 _nvme_rdma_ctrlr_disconnect_qpair(ctrlr, qpair, nvme_rdma_qpair_disconnected); 2247 2248 /* If the async mode is disabled, poll the qpair until it is actually disconnected. 2249 * It is ensured that poll_group_process_completions() calls disconnected_qpair_cb 2250 * for any disconnected qpair. Hence, we do not have to check if the qpair is in 2251 * a poll group or not. 2252 * At the same time, if the qpair is being destroyed, i.e. this function is called by 2253 * spdk_nvme_ctrlr_free_io_qpair then we need to wait until qpair is disconnected, otherwise 2254 * we may leak some resources. 2255 */ 2256 if (qpair->async && !qpair->destroy_in_progress) { 2257 return; 2258 } 2259 2260 while (1) { 2261 rc = nvme_rdma_ctrlr_disconnect_qpair_poll(ctrlr, qpair); 2262 if (rc != -EAGAIN) { 2263 break; 2264 } 2265 } 2266 } 2267 2268 static int 2269 nvme_rdma_stale_conn_disconnected(struct nvme_rdma_qpair *rqpair, int ret) 2270 { 2271 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2272 2273 if (ret) { 2274 SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n"); 2275 } 2276 2277 nvme_rdma_qpair_destroy(rqpair); 2278 2279 qpair->last_transport_failure_reason = qpair->transport_failure_reason; 2280 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_NONE; 2281 2282 rqpair->state = NVME_RDMA_QPAIR_STATE_STALE_CONN; 2283 rqpair->evt_timeout_ticks = (NVME_RDMA_STALE_CONN_RETRY_DELAY_US * spdk_get_ticks_hz()) / 2284 SPDK_SEC_TO_USEC + spdk_get_ticks(); 2285 2286 return 0; 2287 } 2288 2289 static int 2290 nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair) 2291 { 2292 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2293 2294 if (rqpair->stale_conn_retry_count >= NVME_RDMA_STALE_CONN_RETRY_MAX) { 2295 SPDK_ERRLOG("Retry failed %d times, give up stale connection to qpair (cntlid:%u, qid:%u).\n", 2296 NVME_RDMA_STALE_CONN_RETRY_MAX, qpair->ctrlr->cntlid, qpair->id); 2297 return -ESTALE; 2298 } 2299 2300 rqpair->stale_conn_retry_count++; 2301 2302 SPDK_NOTICELOG("%d times, retry stale connection to qpair (cntlid:%u, qid:%u).\n", 2303 rqpair->stale_conn_retry_count, qpair->ctrlr->cntlid, qpair->id); 2304 2305 _nvme_rdma_ctrlr_disconnect_qpair(qpair->ctrlr, qpair, nvme_rdma_stale_conn_disconnected); 2306 2307 return 0; 2308 } 2309 2310 static int 2311 nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair) 2312 { 2313 struct nvme_rdma_qpair *rqpair; 2314 2315 assert(qpair != NULL); 2316 rqpair = nvme_rdma_qpair(qpair); 2317 2318 if (rqpair->state != NVME_RDMA_QPAIR_STATE_EXITED) { 2319 int rc __attribute__((unused)); 2320 2321 /* qpair was removed from the poll group while the disconnect is not finished. 2322 * Destroy rdma resources forcefully. */ 2323 rc = nvme_rdma_qpair_disconnected(rqpair, 0); 2324 assert(rc == 0); 2325 } 2326 2327 nvme_rdma_qpair_abort_reqs(qpair, qpair->abort_dnr); 2328 nvme_qpair_deinit(qpair); 2329 2330 spdk_free(rqpair); 2331 2332 return 0; 2333 } 2334 2335 static struct spdk_nvme_qpair * 2336 nvme_rdma_ctrlr_create_io_qpair(struct spdk_nvme_ctrlr *ctrlr, uint16_t qid, 2337 const struct spdk_nvme_io_qpair_opts *opts) 2338 { 2339 return nvme_rdma_ctrlr_create_qpair(ctrlr, qid, opts->io_queue_size, opts->qprio, 2340 opts->io_queue_requests, 2341 opts->delay_cmd_submit, 2342 opts->async_mode); 2343 } 2344 2345 static int 2346 nvme_rdma_ctrlr_enable(struct spdk_nvme_ctrlr *ctrlr) 2347 { 2348 /* do nothing here */ 2349 return 0; 2350 } 2351 2352 static int nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr); 2353 2354 /* We have to use the typedef in the function declaration to appease astyle. */ 2355 typedef struct spdk_nvme_ctrlr spdk_nvme_ctrlr_t; 2356 2357 static spdk_nvme_ctrlr_t * 2358 nvme_rdma_ctrlr_construct(const struct spdk_nvme_transport_id *trid, 2359 const struct spdk_nvme_ctrlr_opts *opts, 2360 void *devhandle) 2361 { 2362 struct nvme_rdma_ctrlr *rctrlr; 2363 struct ibv_context **contexts; 2364 struct ibv_device_attr dev_attr; 2365 int i, flag, rc; 2366 2367 rctrlr = spdk_zmalloc(sizeof(struct nvme_rdma_ctrlr), 0, NULL, SPDK_ENV_NUMA_ID_ANY, 2368 SPDK_MALLOC_DMA); 2369 if (rctrlr == NULL) { 2370 SPDK_ERRLOG("could not allocate ctrlr\n"); 2371 return NULL; 2372 } 2373 2374 rctrlr->ctrlr.opts = *opts; 2375 rctrlr->ctrlr.trid = *trid; 2376 2377 if (opts->transport_retry_count > NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT) { 2378 SPDK_NOTICELOG("transport_retry_count exceeds max value %d, use max value\n", 2379 NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT); 2380 rctrlr->ctrlr.opts.transport_retry_count = NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT; 2381 } 2382 2383 if (opts->transport_ack_timeout > NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT) { 2384 SPDK_NOTICELOG("transport_ack_timeout exceeds max value %d, use max value\n", 2385 NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT); 2386 rctrlr->ctrlr.opts.transport_ack_timeout = NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT; 2387 } 2388 2389 contexts = rdma_get_devices(NULL); 2390 if (contexts == NULL) { 2391 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2392 spdk_free(rctrlr); 2393 return NULL; 2394 } 2395 2396 i = 0; 2397 rctrlr->max_sge = NVME_RDMA_MAX_SGL_DESCRIPTORS; 2398 2399 while (contexts[i] != NULL) { 2400 rc = ibv_query_device(contexts[i], &dev_attr); 2401 if (rc < 0) { 2402 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2403 rdma_free_devices(contexts); 2404 spdk_free(rctrlr); 2405 return NULL; 2406 } 2407 rctrlr->max_sge = spdk_min(rctrlr->max_sge, (uint16_t)dev_attr.max_sge); 2408 i++; 2409 } 2410 2411 rdma_free_devices(contexts); 2412 2413 rc = nvme_ctrlr_construct(&rctrlr->ctrlr); 2414 if (rc != 0) { 2415 spdk_free(rctrlr); 2416 return NULL; 2417 } 2418 2419 STAILQ_INIT(&rctrlr->pending_cm_events); 2420 STAILQ_INIT(&rctrlr->free_cm_events); 2421 rctrlr->cm_events = spdk_zmalloc(NVME_RDMA_NUM_CM_EVENTS * sizeof(*rctrlr->cm_events), 0, NULL, 2422 SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 2423 if (rctrlr->cm_events == NULL) { 2424 SPDK_ERRLOG("unable to allocate buffers to hold CM events.\n"); 2425 goto destruct_ctrlr; 2426 } 2427 2428 for (i = 0; i < NVME_RDMA_NUM_CM_EVENTS; i++) { 2429 STAILQ_INSERT_TAIL(&rctrlr->free_cm_events, &rctrlr->cm_events[i], link); 2430 } 2431 2432 rctrlr->cm_channel = rdma_create_event_channel(); 2433 if (rctrlr->cm_channel == NULL) { 2434 SPDK_ERRLOG("rdma_create_event_channel() failed\n"); 2435 goto destruct_ctrlr; 2436 } 2437 2438 flag = fcntl(rctrlr->cm_channel->fd, F_GETFL); 2439 if (fcntl(rctrlr->cm_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2440 SPDK_ERRLOG("Cannot set event channel to non blocking\n"); 2441 goto destruct_ctrlr; 2442 } 2443 2444 rctrlr->ctrlr.adminq = nvme_rdma_ctrlr_create_qpair(&rctrlr->ctrlr, 0, 2445 rctrlr->ctrlr.opts.admin_queue_size, 0, 2446 rctrlr->ctrlr.opts.admin_queue_size, false, true); 2447 if (!rctrlr->ctrlr.adminq) { 2448 SPDK_ERRLOG("failed to create admin qpair\n"); 2449 goto destruct_ctrlr; 2450 } 2451 if (spdk_rdma_provider_accel_sequence_supported()) { 2452 rctrlr->ctrlr.flags |= SPDK_NVME_CTRLR_ACCEL_SEQUENCE_SUPPORTED; 2453 } 2454 2455 if (nvme_ctrlr_add_process(&rctrlr->ctrlr, 0) != 0) { 2456 SPDK_ERRLOG("nvme_ctrlr_add_process() failed\n"); 2457 goto destruct_ctrlr; 2458 } 2459 2460 SPDK_DEBUGLOG(nvme, "successfully initialized the nvmf ctrlr\n"); 2461 return &rctrlr->ctrlr; 2462 2463 destruct_ctrlr: 2464 nvme_ctrlr_destruct(&rctrlr->ctrlr); 2465 return NULL; 2466 } 2467 2468 static int 2469 nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr) 2470 { 2471 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr); 2472 struct nvme_rdma_cm_event_entry *entry; 2473 2474 if (ctrlr->adminq) { 2475 nvme_rdma_ctrlr_delete_io_qpair(ctrlr, ctrlr->adminq); 2476 } 2477 2478 STAILQ_FOREACH(entry, &rctrlr->pending_cm_events, link) { 2479 rdma_ack_cm_event(entry->evt); 2480 } 2481 2482 STAILQ_INIT(&rctrlr->free_cm_events); 2483 STAILQ_INIT(&rctrlr->pending_cm_events); 2484 spdk_free(rctrlr->cm_events); 2485 2486 if (rctrlr->cm_channel) { 2487 rdma_destroy_event_channel(rctrlr->cm_channel); 2488 rctrlr->cm_channel = NULL; 2489 } 2490 2491 nvme_ctrlr_destruct_finish(ctrlr); 2492 2493 spdk_free(rctrlr); 2494 2495 return 0; 2496 } 2497 2498 static inline int 2499 _nvme_rdma_qpair_submit_request(struct nvme_rdma_qpair *rqpair, 2500 struct spdk_nvme_rdma_req *rdma_req) 2501 { 2502 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 2503 struct ibv_send_wr *wr; 2504 struct nvme_rdma_poll_group *group; 2505 2506 if (!rqpair->link_active.tqe_prev && qpair->poll_group) { 2507 group = nvme_rdma_poll_group(qpair->poll_group); 2508 TAILQ_INSERT_TAIL(&group->active_qpairs, rqpair, link_active); 2509 } 2510 assert(rqpair->current_num_sends < rqpair->num_entries); 2511 rqpair->current_num_sends++; 2512 2513 wr = &rdma_req->send_wr; 2514 wr->next = NULL; 2515 nvme_rdma_trace_ibv_sge(wr->sg_list); 2516 2517 spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, wr); 2518 2519 if (!rqpair->delay_cmd_submit) { 2520 return nvme_rdma_qpair_submit_sends(rqpair); 2521 } 2522 2523 return 0; 2524 } 2525 2526 static int 2527 nvme_rdma_qpair_submit_request(struct spdk_nvme_qpair *qpair, 2528 struct nvme_request *req) 2529 { 2530 struct nvme_rdma_qpair *rqpair; 2531 struct spdk_nvme_rdma_req *rdma_req; 2532 int rc; 2533 2534 rqpair = nvme_rdma_qpair(qpair); 2535 assert(rqpair != NULL); 2536 assert(req != NULL); 2537 2538 rdma_req = nvme_rdma_req_get(rqpair); 2539 if (spdk_unlikely(!rdma_req)) { 2540 if (rqpair->poller) { 2541 rqpair->poller->stats.queued_requests++; 2542 } 2543 /* Inform the upper layer to try again later. */ 2544 return -EAGAIN; 2545 } 2546 2547 assert(rdma_req->req == NULL); 2548 rdma_req->req = req; 2549 req->cmd.cid = rdma_req->id; 2550 if (req->accel_sequence || rqpair->append_copy) { 2551 assert(spdk_rdma_provider_accel_sequence_supported()); 2552 assert(rqpair->qpair.poll_group->group); 2553 assert(rqpair->qpair.poll_group->group->accel_fn_table.append_copy); 2554 assert(rqpair->qpair.poll_group->group->accel_fn_table.reverse_sequence); 2555 assert(rqpair->qpair.poll_group->group->accel_fn_table.finish_sequence); 2556 2557 rc = nvme_rdma_apply_accel_sequence(rqpair, req, rdma_req); 2558 if (spdk_unlikely(rc)) { 2559 SPDK_ERRLOG("failed to apply accel seq, rqpair %p, req %p, rc %d\n", rqpair, rdma_req, rc); 2560 nvme_rdma_req_put(rqpair, rdma_req); 2561 return rc; 2562 } 2563 /* Capsule will be sent in data_transfer callback */ 2564 return 0; 2565 } 2566 2567 rc = nvme_rdma_req_init(rqpair, rdma_req); 2568 if (spdk_unlikely(rc)) { 2569 SPDK_ERRLOG("nvme_rdma_req_init() failed\n"); 2570 nvme_rdma_req_put(rqpair, rdma_req); 2571 return -1; 2572 } 2573 2574 TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link); 2575 rqpair->num_outstanding_reqs++; 2576 2577 return _nvme_rdma_qpair_submit_request(rqpair, rdma_req); 2578 } 2579 2580 static int 2581 nvme_rdma_qpair_reset(struct spdk_nvme_qpair *qpair) 2582 { 2583 /* Currently, doing nothing here */ 2584 return 0; 2585 } 2586 2587 static void 2588 nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr) 2589 { 2590 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2591 struct spdk_nvme_cpl cpl; 2592 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2593 2594 cpl.sqid = qpair->id; 2595 cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION; 2596 cpl.status.sct = SPDK_NVME_SCT_GENERIC; 2597 cpl.status.dnr = dnr; 2598 2599 /* 2600 * We cannot abort requests at the RDMA layer without 2601 * unregistering them. If we do, we can still get error 2602 * free completions on the shared completion queue. 2603 */ 2604 if (nvme_qpair_get_state(qpair) > NVME_QPAIR_DISCONNECTING && 2605 nvme_qpair_get_state(qpair) != NVME_QPAIR_DESTROYING) { 2606 nvme_ctrlr_disconnect_qpair(qpair); 2607 } 2608 2609 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2610 if (rdma_req->in_progress_accel) { 2611 /* We should wait for accel completion */ 2612 continue; 2613 } 2614 nvme_rdma_req_complete(rdma_req, &cpl, true); 2615 } 2616 } 2617 2618 static void 2619 nvme_rdma_qpair_check_timeout(struct spdk_nvme_qpair *qpair) 2620 { 2621 uint64_t t02; 2622 struct spdk_nvme_rdma_req *rdma_req, *tmp; 2623 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2624 struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr; 2625 struct spdk_nvme_ctrlr_process *active_proc; 2626 2627 /* Don't check timeouts during controller initialization. */ 2628 if (ctrlr->state != NVME_CTRLR_STATE_READY) { 2629 return; 2630 } 2631 2632 if (nvme_qpair_is_admin_queue(qpair)) { 2633 active_proc = nvme_ctrlr_get_current_process(ctrlr); 2634 } else { 2635 active_proc = qpair->active_proc; 2636 } 2637 2638 /* Only check timeouts if the current process has a timeout callback. */ 2639 if (active_proc == NULL || active_proc->timeout_cb_fn == NULL) { 2640 return; 2641 } 2642 2643 t02 = spdk_get_ticks(); 2644 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 2645 assert(rdma_req->req != NULL); 2646 2647 if (nvme_request_check_timeout(rdma_req->req, rdma_req->id, active_proc, t02)) { 2648 /* 2649 * The requests are in order, so as soon as one has not timed out, 2650 * stop iterating. 2651 */ 2652 break; 2653 } 2654 } 2655 } 2656 2657 static inline void 2658 nvme_rdma_request_ready(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req) 2659 { 2660 struct spdk_nvme_rdma_rsp *rdma_rsp = rdma_req->rdma_rsp; 2661 struct ibv_recv_wr *recv_wr = rdma_rsp->recv_wr; 2662 2663 if (rdma_req->transfer_cpl_cb) { 2664 int rc = 0; 2665 2666 if (spdk_unlikely(spdk_nvme_cpl_is_error(&rdma_rsp->cpl))) { 2667 SPDK_WARNLOG("req %p, error cpl sct %d, sc %d\n", rdma_req, rdma_rsp->cpl.status.sct, 2668 rdma_rsp->cpl.status.sc); 2669 rc = -EIO; 2670 } 2671 nvme_rdma_finish_data_transfer(rdma_req, rc); 2672 } else { 2673 nvme_rdma_req_complete(rdma_req, &rdma_rsp->cpl, true); 2674 } 2675 2676 if (spdk_unlikely(rqpair->state >= NVME_RDMA_QPAIR_STATE_EXITING && !rqpair->srq)) { 2677 /* Skip posting back recv wr if we are in a disconnection process. We may never get 2678 * a WC and we may end up stuck in LINGERING state until the timeout. */ 2679 return; 2680 } 2681 2682 assert(rqpair->rsps->current_num_recvs < rqpair->rsps->num_entries); 2683 rqpair->rsps->current_num_recvs++; 2684 2685 recv_wr->next = NULL; 2686 nvme_rdma_trace_ibv_sge(recv_wr->sg_list); 2687 2688 if (!rqpair->srq) { 2689 spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, recv_wr); 2690 } else { 2691 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, recv_wr); 2692 } 2693 } 2694 2695 #define MAX_COMPLETIONS_PER_POLL 128 2696 2697 static void 2698 nvme_rdma_fail_qpair(struct spdk_nvme_qpair *qpair, int failure_reason) 2699 { 2700 if (failure_reason == IBV_WC_RETRY_EXC_ERR) { 2701 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE; 2702 } else if (qpair->transport_failure_reason == SPDK_NVME_QPAIR_FAILURE_NONE) { 2703 qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_UNKNOWN; 2704 } 2705 2706 nvme_ctrlr_disconnect_qpair(qpair); 2707 } 2708 2709 static struct nvme_rdma_qpair * 2710 get_rdma_qpair_from_wc(struct nvme_rdma_poll_group *group, struct ibv_wc *wc) 2711 { 2712 struct spdk_nvme_qpair *qpair; 2713 struct nvme_rdma_qpair *rqpair; 2714 2715 STAILQ_FOREACH(qpair, &group->group.connected_qpairs, poll_group_stailq) { 2716 rqpair = nvme_rdma_qpair(qpair); 2717 if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) { 2718 return rqpair; 2719 } 2720 } 2721 2722 STAILQ_FOREACH(qpair, &group->group.disconnected_qpairs, poll_group_stailq) { 2723 rqpair = nvme_rdma_qpair(qpair); 2724 if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) { 2725 return rqpair; 2726 } 2727 } 2728 2729 return NULL; 2730 } 2731 2732 static inline void 2733 nvme_rdma_log_wc_status(struct nvme_rdma_qpair *rqpair, struct ibv_wc *wc) 2734 { 2735 struct nvme_rdma_wr *rdma_wr = (struct nvme_rdma_wr *)wc->wr_id; 2736 2737 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 2738 /* If qpair is in ERR state, we will receive completions for all posted and not completed 2739 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 2740 SPDK_DEBUGLOG(nvme, "WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n", 2741 rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status, 2742 ibv_wc_status_str(wc->status)); 2743 } else { 2744 SPDK_ERRLOG("WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n", 2745 rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status, 2746 ibv_wc_status_str(wc->status)); 2747 } 2748 } 2749 2750 static inline int 2751 nvme_rdma_process_recv_completion(struct nvme_rdma_poller *poller, struct ibv_wc *wc, 2752 struct nvme_rdma_wr *rdma_wr) 2753 { 2754 struct nvme_rdma_qpair *rqpair; 2755 struct spdk_nvme_rdma_req *rdma_req; 2756 struct spdk_nvme_rdma_rsp *rdma_rsp; 2757 2758 rdma_rsp = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_rsp, rdma_wr); 2759 2760 if (poller && poller->srq) { 2761 rqpair = get_rdma_qpair_from_wc(poller->group, wc); 2762 if (spdk_unlikely(!rqpair)) { 2763 /* Since we do not handle the LAST_WQE_REACHED event, we do not know when 2764 * a Receive Queue in a QP, that is associated with an SRQ, is flushed. 2765 * We may get a WC for a already destroyed QP. 2766 * 2767 * However, for the SRQ, this is not any error. Hence, just re-post the 2768 * receive request to the SRQ to reuse for other QPs, and return 0. 2769 */ 2770 spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr); 2771 return 0; 2772 } 2773 } else { 2774 rqpair = rdma_rsp->rqpair; 2775 if (spdk_unlikely(!rqpair)) { 2776 /* TODO: Fix forceful QP destroy when it is not async mode. 2777 * CQ itself did not cause any error. Hence, return 0 for now. 2778 */ 2779 SPDK_WARNLOG("QP might be already destroyed.\n"); 2780 return 0; 2781 } 2782 } 2783 2784 2785 assert(rqpair->rsps->current_num_recvs > 0); 2786 rqpair->rsps->current_num_recvs--; 2787 2788 if (spdk_unlikely(wc->status)) { 2789 nvme_rdma_log_wc_status(rqpair, wc); 2790 goto err_wc; 2791 } 2792 2793 SPDK_DEBUGLOG(nvme, "CQ recv completion\n"); 2794 2795 if (spdk_unlikely(wc->byte_len < sizeof(struct spdk_nvme_cpl))) { 2796 SPDK_ERRLOG("recv length %u less than expected response size\n", wc->byte_len); 2797 goto err_wc; 2798 } 2799 rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid]; 2800 rdma_req->completion_flags |= NVME_RDMA_RECV_COMPLETED; 2801 rdma_req->rdma_rsp = rdma_rsp; 2802 2803 if ((rdma_req->completion_flags & NVME_RDMA_SEND_COMPLETED) == 0) { 2804 return 0; 2805 } 2806 2807 rqpair->num_completions++; 2808 2809 nvme_rdma_request_ready(rqpair, rdma_req); 2810 2811 if (!rqpair->delay_cmd_submit) { 2812 if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) { 2813 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 2814 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2815 return -ENXIO; 2816 } 2817 } 2818 2819 return 1; 2820 2821 err_wc: 2822 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2823 if (poller && poller->srq) { 2824 spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr); 2825 } 2826 rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid]; 2827 if (rdma_req->transfer_cpl_cb) { 2828 nvme_rdma_finish_data_transfer(rdma_req, -ENXIO); 2829 } 2830 return -ENXIO; 2831 } 2832 2833 static inline int 2834 nvme_rdma_process_send_completion(struct nvme_rdma_poller *poller, 2835 struct nvme_rdma_qpair *rdma_qpair, 2836 struct ibv_wc *wc, struct nvme_rdma_wr *rdma_wr) 2837 { 2838 struct nvme_rdma_qpair *rqpair; 2839 struct spdk_nvme_rdma_req *rdma_req; 2840 2841 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_req, rdma_wr); 2842 rqpair = rdma_req->req ? nvme_rdma_qpair(rdma_req->req->qpair) : NULL; 2843 if (spdk_unlikely(!rqpair)) { 2844 rqpair = rdma_qpair != NULL ? rdma_qpair : get_rdma_qpair_from_wc(poller->group, wc); 2845 } 2846 2847 /* If we are flushing I/O */ 2848 if (spdk_unlikely(wc->status)) { 2849 if (!rqpair) { 2850 /* When poll_group is used, several qpairs share the same CQ and it is possible to 2851 * receive a completion with error (e.g. IBV_WC_WR_FLUSH_ERR) for already disconnected qpair 2852 * That happens due to qpair is destroyed while there are submitted but not completed send/receive 2853 * Work Requests */ 2854 assert(poller); 2855 return 0; 2856 } 2857 assert(rqpair->current_num_sends > 0); 2858 rqpair->current_num_sends--; 2859 nvme_rdma_log_wc_status(rqpair, wc); 2860 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2861 if (rdma_req->rdma_rsp && poller && poller->srq) { 2862 spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_req->rdma_rsp->recv_wr); 2863 } 2864 if (rdma_req->transfer_cpl_cb) { 2865 nvme_rdma_finish_data_transfer(rdma_req, -ENXIO); 2866 } 2867 return -ENXIO; 2868 } 2869 2870 /* We do not support Soft Roce anymore. Other than Soft Roce's bug, we should not 2871 * receive a completion without error status after qpair is disconnected/destroyed. 2872 */ 2873 if (spdk_unlikely(rdma_req->req == NULL)) { 2874 /* 2875 * Some infiniband drivers do not guarantee the previous assumption after we 2876 * received a RDMA_CM_EVENT_DEVICE_REMOVAL event. 2877 */ 2878 SPDK_ERRLOG("Received malformed completion: request 0x%"PRIx64" type %d\n", wc->wr_id, 2879 rdma_wr->type); 2880 if (!rqpair || !rqpair->need_destroy) { 2881 assert(0); 2882 } 2883 return -ENXIO; 2884 } 2885 2886 rdma_req->completion_flags |= NVME_RDMA_SEND_COMPLETED; 2887 assert(rqpair->current_num_sends > 0); 2888 rqpair->current_num_sends--; 2889 2890 if ((rdma_req->completion_flags & NVME_RDMA_RECV_COMPLETED) == 0) { 2891 return 0; 2892 } 2893 2894 rqpair->num_completions++; 2895 2896 nvme_rdma_request_ready(rqpair, rdma_req); 2897 2898 if (!rqpair->delay_cmd_submit) { 2899 if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) { 2900 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 2901 nvme_rdma_fail_qpair(&rqpair->qpair, 0); 2902 return -ENXIO; 2903 } 2904 } 2905 2906 return 1; 2907 } 2908 2909 static inline int 2910 nvme_rdma_cq_process_completions(struct ibv_cq *cq, uint32_t batch_size, 2911 struct nvme_rdma_poller *poller, 2912 struct nvme_rdma_qpair *rdma_qpair, 2913 uint64_t *rdma_completions) 2914 { 2915 struct ibv_wc wc[MAX_COMPLETIONS_PER_POLL]; 2916 struct nvme_rdma_wr *rdma_wr; 2917 uint32_t reaped = 0; 2918 int completion_rc = 0; 2919 int rc, _rc, i; 2920 2921 rc = ibv_poll_cq(cq, batch_size, wc); 2922 if (spdk_unlikely(rc < 0)) { 2923 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2924 errno, spdk_strerror(errno)); 2925 return -ECANCELED; 2926 } else if (rc == 0) { 2927 return 0; 2928 } 2929 2930 for (i = 0; i < rc; i++) { 2931 rdma_wr = (struct nvme_rdma_wr *)wc[i].wr_id; 2932 switch (rdma_wr->type) { 2933 case RDMA_WR_TYPE_RECV: 2934 _rc = nvme_rdma_process_recv_completion(poller, &wc[i], rdma_wr); 2935 break; 2936 2937 case RDMA_WR_TYPE_SEND: 2938 _rc = nvme_rdma_process_send_completion(poller, rdma_qpair, &wc[i], rdma_wr); 2939 break; 2940 2941 default: 2942 SPDK_ERRLOG("Received an unexpected opcode on the CQ: %d\n", rdma_wr->type); 2943 return -ECANCELED; 2944 } 2945 if (spdk_likely(_rc >= 0)) { 2946 reaped += _rc; 2947 } else { 2948 completion_rc = _rc; 2949 } 2950 } 2951 2952 *rdma_completions += rc; 2953 2954 if (spdk_unlikely(completion_rc)) { 2955 return completion_rc; 2956 } 2957 2958 return reaped; 2959 } 2960 2961 static void 2962 dummy_disconnected_qpair_cb(struct spdk_nvme_qpair *qpair, void *poll_group_ctx) 2963 { 2964 2965 } 2966 2967 static int 2968 nvme_rdma_qpair_process_completions(struct spdk_nvme_qpair *qpair, 2969 uint32_t max_completions) 2970 { 2971 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 2972 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(qpair->ctrlr); 2973 int rc = 0, batch_size; 2974 struct ibv_cq *cq; 2975 uint64_t rdma_completions = 0; 2976 2977 /* 2978 * This is used during the connection phase. It's possible that we are still reaping error completions 2979 * from other qpairs so we need to call the poll group function. Also, it's more correct since the cq 2980 * is shared. 2981 */ 2982 if (qpair->poll_group != NULL) { 2983 return spdk_nvme_poll_group_process_completions(qpair->poll_group->group, max_completions, 2984 dummy_disconnected_qpair_cb); 2985 } 2986 2987 if (max_completions == 0) { 2988 max_completions = rqpair->num_entries; 2989 } else { 2990 max_completions = spdk_min(max_completions, rqpair->num_entries); 2991 } 2992 2993 switch (nvme_qpair_get_state(qpair)) { 2994 case NVME_QPAIR_CONNECTING: 2995 rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair); 2996 if (rc == 0) { 2997 /* Once the connection is completed, we can submit queued requests */ 2998 nvme_qpair_resubmit_requests(qpair, rqpair->num_entries); 2999 } else if (rc != -EAGAIN) { 3000 SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair); 3001 goto failed; 3002 } else if (rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING) { 3003 return 0; 3004 } 3005 break; 3006 3007 case NVME_QPAIR_DISCONNECTING: 3008 nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair); 3009 return -ENXIO; 3010 3011 default: 3012 if (nvme_qpair_is_admin_queue(qpair)) { 3013 nvme_rdma_poll_events(rctrlr); 3014 } 3015 nvme_rdma_qpair_process_cm_event(rqpair); 3016 break; 3017 } 3018 3019 if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) { 3020 goto failed; 3021 } 3022 3023 cq = rqpair->cq; 3024 3025 rqpair->num_completions = 0; 3026 do { 3027 batch_size = spdk_min((max_completions - rqpair->num_completions), MAX_COMPLETIONS_PER_POLL); 3028 rc = nvme_rdma_cq_process_completions(cq, batch_size, NULL, rqpair, &rdma_completions); 3029 3030 if (rc == 0) { 3031 break; 3032 /* Handle the case where we fail to poll the cq. */ 3033 } else if (rc == -ECANCELED) { 3034 goto failed; 3035 } else if (rc == -ENXIO) { 3036 return rc; 3037 } 3038 } while (rqpair->num_completions < max_completions); 3039 3040 if (spdk_unlikely(nvme_rdma_qpair_submit_sends(rqpair) || 3041 nvme_rdma_qpair_submit_recvs(rqpair))) { 3042 goto failed; 3043 } 3044 3045 if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) { 3046 nvme_rdma_qpair_check_timeout(qpair); 3047 } 3048 3049 return rqpair->num_completions; 3050 3051 failed: 3052 nvme_rdma_fail_qpair(qpair, 0); 3053 return -ENXIO; 3054 } 3055 3056 static uint32_t 3057 nvme_rdma_ctrlr_get_max_xfer_size(struct spdk_nvme_ctrlr *ctrlr) 3058 { 3059 /* max_mr_size by ibv_query_device indicates the largest value that we can 3060 * set for a registered memory region. It is independent from the actual 3061 * I/O size and is very likely to be larger than 2 MiB which is the 3062 * granularity we currently register memory regions. Hence return 3063 * UINT32_MAX here and let the generic layer use the controller data to 3064 * moderate this value. 3065 */ 3066 return UINT32_MAX; 3067 } 3068 3069 static uint16_t 3070 nvme_rdma_ctrlr_get_max_sges(struct spdk_nvme_ctrlr *ctrlr) 3071 { 3072 struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr); 3073 uint32_t max_sge = rctrlr->max_sge; 3074 uint32_t max_in_capsule_sge = (ctrlr->cdata.nvmf_specific.ioccsz * 16 - 3075 sizeof(struct spdk_nvme_cmd)) / 3076 sizeof(struct spdk_nvme_sgl_descriptor); 3077 3078 /* Max SGE is limited by capsule size */ 3079 max_sge = spdk_min(max_sge, max_in_capsule_sge); 3080 /* Max SGE may be limited by MSDBD. 3081 * If umr_per_io is enabled and supported, we always use virtually contig buffer, we don't limit max_sge by 3082 * MSDBD in that case */ 3083 if (!(g_spdk_nvme_transport_opts.rdma_umr_per_io && 3084 spdk_rdma_provider_accel_sequence_supported()) && 3085 ctrlr->cdata.nvmf_specific.msdbd != 0) { 3086 max_sge = spdk_min(max_sge, ctrlr->cdata.nvmf_specific.msdbd); 3087 } 3088 3089 /* Max SGE can't be less than 1 */ 3090 max_sge = spdk_max(1, max_sge); 3091 return max_sge; 3092 } 3093 3094 static int 3095 nvme_rdma_qpair_iterate_requests(struct spdk_nvme_qpair *qpair, 3096 int (*iter_fn)(struct nvme_request *req, void *arg), 3097 void *arg) 3098 { 3099 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3100 struct spdk_nvme_rdma_req *rdma_req, *tmp; 3101 int rc; 3102 3103 assert(iter_fn != NULL); 3104 3105 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 3106 assert(rdma_req->req != NULL); 3107 3108 rc = iter_fn(rdma_req->req, arg); 3109 if (rc != 0) { 3110 return rc; 3111 } 3112 } 3113 3114 return 0; 3115 } 3116 3117 static int 3118 nvme_rdma_qpair_authenticate(struct spdk_nvme_qpair *qpair) 3119 { 3120 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3121 int rc; 3122 3123 /* If the qpair is still connecting, it'll be forced to authenticate later on */ 3124 if (rqpair->state < NVME_RDMA_QPAIR_STATE_RUNNING) { 3125 return 0; 3126 } else if (rqpair->state != NVME_RDMA_QPAIR_STATE_RUNNING) { 3127 return -ENOTCONN; 3128 } 3129 3130 rc = nvme_fabric_qpair_authenticate_async(qpair); 3131 if (rc == 0) { 3132 nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTING); 3133 rqpair->state = NVME_RDMA_QPAIR_STATE_AUTHENTICATING; 3134 } 3135 3136 return rc; 3137 } 3138 3139 static void 3140 nvme_rdma_admin_qpair_abort_aers(struct spdk_nvme_qpair *qpair) 3141 { 3142 struct spdk_nvme_rdma_req *rdma_req, *tmp; 3143 struct spdk_nvme_cpl cpl; 3144 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3145 3146 cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION; 3147 cpl.status.sct = SPDK_NVME_SCT_GENERIC; 3148 3149 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) { 3150 assert(rdma_req->req != NULL); 3151 3152 if (rdma_req->req->cmd.opc != SPDK_NVME_OPC_ASYNC_EVENT_REQUEST) { 3153 continue; 3154 } 3155 3156 nvme_rdma_req_complete(rdma_req, &cpl, false); 3157 } 3158 } 3159 3160 static void 3161 nvme_rdma_poller_destroy(struct nvme_rdma_poller *poller) 3162 { 3163 if (poller->cq) { 3164 ibv_destroy_cq(poller->cq); 3165 } 3166 if (poller->rsps) { 3167 nvme_rdma_free_rsps(poller->rsps); 3168 } 3169 if (poller->srq) { 3170 spdk_rdma_provider_srq_destroy(poller->srq); 3171 } 3172 if (poller->mr_map) { 3173 spdk_rdma_utils_free_mem_map(&poller->mr_map); 3174 } 3175 if (poller->pd) { 3176 spdk_rdma_utils_put_pd(poller->pd); 3177 } 3178 free(poller); 3179 } 3180 3181 static struct nvme_rdma_poller * 3182 nvme_rdma_poller_create(struct nvme_rdma_poll_group *group, struct ibv_context *ctx) 3183 { 3184 struct nvme_rdma_poller *poller; 3185 struct ibv_device_attr dev_attr; 3186 struct spdk_rdma_provider_srq_init_attr srq_init_attr = {}; 3187 struct nvme_rdma_rsp_opts opts; 3188 int num_cqe, max_num_cqe; 3189 int rc; 3190 3191 poller = calloc(1, sizeof(*poller)); 3192 if (poller == NULL) { 3193 SPDK_ERRLOG("Unable to allocate poller.\n"); 3194 return NULL; 3195 } 3196 3197 poller->group = group; 3198 poller->device = ctx; 3199 3200 if (g_spdk_nvme_transport_opts.rdma_srq_size != 0) { 3201 rc = ibv_query_device(ctx, &dev_attr); 3202 if (rc) { 3203 SPDK_ERRLOG("Unable to query RDMA device.\n"); 3204 goto fail; 3205 } 3206 3207 poller->pd = spdk_rdma_utils_get_pd(ctx); 3208 if (poller->pd == NULL) { 3209 SPDK_ERRLOG("Unable to get PD.\n"); 3210 goto fail; 3211 } 3212 3213 poller->mr_map = spdk_rdma_utils_create_mem_map(poller->pd, &g_nvme_hooks, 3214 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE); 3215 if (poller->mr_map == NULL) { 3216 SPDK_ERRLOG("Unable to create memory map.\n"); 3217 goto fail; 3218 } 3219 3220 srq_init_attr.stats = &poller->stats.rdma_stats.recv; 3221 srq_init_attr.pd = poller->pd; 3222 srq_init_attr.srq_init_attr.attr.max_wr = spdk_min((uint32_t)dev_attr.max_srq_wr, 3223 g_spdk_nvme_transport_opts.rdma_srq_size); 3224 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(dev_attr.max_sge, 3225 NVME_RDMA_DEFAULT_RX_SGE); 3226 3227 poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr); 3228 if (poller->srq == NULL) { 3229 SPDK_ERRLOG("Unable to create SRQ.\n"); 3230 goto fail; 3231 } 3232 3233 opts.num_entries = g_spdk_nvme_transport_opts.rdma_srq_size; 3234 opts.rqpair = NULL; 3235 opts.srq = poller->srq; 3236 opts.mr_map = poller->mr_map; 3237 3238 poller->rsps = nvme_rdma_create_rsps(&opts); 3239 if (poller->rsps == NULL) { 3240 SPDK_ERRLOG("Unable to create poller RDMA responses.\n"); 3241 goto fail; 3242 } 3243 3244 rc = nvme_rdma_poller_submit_recvs(poller); 3245 if (rc) { 3246 SPDK_ERRLOG("Unable to submit poller RDMA responses.\n"); 3247 goto fail; 3248 } 3249 3250 /* 3251 * When using an srq, fix the size of the completion queue at startup. 3252 * The initiator sends only send and recv WRs. Hence, the multiplier is 2. 3253 * (The target sends also data WRs. Hence, the multiplier is 3.) 3254 */ 3255 num_cqe = g_spdk_nvme_transport_opts.rdma_srq_size * 2; 3256 } else { 3257 num_cqe = DEFAULT_NVME_RDMA_CQ_SIZE; 3258 } 3259 3260 max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size; 3261 if (max_num_cqe != 0 && num_cqe > max_num_cqe) { 3262 num_cqe = max_num_cqe; 3263 } 3264 3265 poller->cq = ibv_create_cq(poller->device, num_cqe, group, NULL, 0); 3266 3267 if (poller->cq == NULL) { 3268 SPDK_ERRLOG("Unable to create CQ, errno %d.\n", errno); 3269 goto fail; 3270 } 3271 3272 STAILQ_INSERT_HEAD(&group->pollers, poller, link); 3273 group->num_pollers++; 3274 poller->current_num_wc = num_cqe; 3275 poller->required_num_wc = 0; 3276 return poller; 3277 3278 fail: 3279 nvme_rdma_poller_destroy(poller); 3280 return NULL; 3281 } 3282 3283 static void 3284 nvme_rdma_poll_group_free_pollers(struct nvme_rdma_poll_group *group) 3285 { 3286 struct nvme_rdma_poller *poller, *tmp_poller; 3287 3288 STAILQ_FOREACH_SAFE(poller, &group->pollers, link, tmp_poller) { 3289 assert(poller->refcnt == 0); 3290 if (poller->refcnt) { 3291 SPDK_WARNLOG("Destroying poller with non-zero ref count: poller %p, refcnt %d\n", 3292 poller, poller->refcnt); 3293 } 3294 3295 STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link); 3296 nvme_rdma_poller_destroy(poller); 3297 } 3298 } 3299 3300 static struct nvme_rdma_poller * 3301 nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, struct ibv_context *device) 3302 { 3303 struct nvme_rdma_poller *poller = NULL; 3304 3305 STAILQ_FOREACH(poller, &group->pollers, link) { 3306 if (poller->device == device) { 3307 break; 3308 } 3309 } 3310 3311 if (!poller) { 3312 poller = nvme_rdma_poller_create(group, device); 3313 if (!poller) { 3314 SPDK_ERRLOG("Failed to create a poller for device %p\n", device); 3315 return NULL; 3316 } 3317 } 3318 3319 poller->refcnt++; 3320 return poller; 3321 } 3322 3323 static void 3324 nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, struct nvme_rdma_poller *poller) 3325 { 3326 assert(poller->refcnt > 0); 3327 if (--poller->refcnt == 0) { 3328 STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link); 3329 group->num_pollers--; 3330 nvme_rdma_poller_destroy(poller); 3331 } 3332 } 3333 3334 static struct spdk_nvme_transport_poll_group * 3335 nvme_rdma_poll_group_create(void) 3336 { 3337 struct nvme_rdma_poll_group *group; 3338 3339 group = calloc(1, sizeof(*group)); 3340 if (group == NULL) { 3341 SPDK_ERRLOG("Unable to allocate poll group.\n"); 3342 return NULL; 3343 } 3344 3345 STAILQ_INIT(&group->pollers); 3346 TAILQ_INIT(&group->connecting_qpairs); 3347 TAILQ_INIT(&group->active_qpairs); 3348 return &group->group; 3349 } 3350 3351 static int 3352 nvme_rdma_poll_group_connect_qpair(struct spdk_nvme_qpair *qpair) 3353 { 3354 return 0; 3355 } 3356 3357 static int 3358 nvme_rdma_poll_group_disconnect_qpair(struct spdk_nvme_qpair *qpair) 3359 { 3360 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3361 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group); 3362 3363 if (rqpair->link_connecting.tqe_prev) { 3364 TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting); 3365 /* We use prev pointer to check if qpair is in connecting list or not . 3366 * TAILQ_REMOVE doesn't do it. So, we do it manually. 3367 */ 3368 rqpair->link_connecting.tqe_prev = NULL; 3369 } 3370 3371 return 0; 3372 } 3373 3374 static int 3375 nvme_rdma_poll_group_add(struct spdk_nvme_transport_poll_group *tgroup, 3376 struct spdk_nvme_qpair *qpair) 3377 { 3378 return 0; 3379 } 3380 3381 static int 3382 nvme_rdma_poll_group_remove(struct spdk_nvme_transport_poll_group *tgroup, 3383 struct spdk_nvme_qpair *qpair) 3384 { 3385 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair); 3386 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group); 3387 3388 if (rqpair->poller) { 3389 /* A qpair may skip transport disconnect part if it was already disconnecting. But on RDMA level a qpair 3390 * may still have a poller reference. In that case we should continue transport disconnect here 3391 * because a poller depends on the poll group reference which is going to be removed */ 3392 SPDK_INFOLOG(nvme, "qpair %p, id %u, nvme state %d, rdma state %d, force disconnect\n", 3393 qpair, qpair->id, qpair->state, rqpair->state); 3394 nvme_rdma_ctrlr_disconnect_qpair(qpair->ctrlr, qpair); 3395 } 3396 3397 if (rqpair->link_active.tqe_prev) { 3398 TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active); 3399 rqpair->link_active.tqe_prev = NULL; 3400 } 3401 3402 return 0; 3403 } 3404 3405 static inline void 3406 nvme_rdma_qpair_process_submits(struct nvme_rdma_poll_group *group, 3407 struct nvme_rdma_qpair *rqpair) 3408 { 3409 struct spdk_nvme_qpair *qpair = &rqpair->qpair; 3410 3411 assert(rqpair->link_active.tqe_prev != NULL); 3412 3413 if (spdk_unlikely(rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING || 3414 rqpair->state >= NVME_RDMA_QPAIR_STATE_EXITING)) { 3415 return; 3416 } 3417 3418 if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) { 3419 nvme_rdma_qpair_check_timeout(qpair); 3420 } 3421 3422 nvme_rdma_qpair_submit_sends(rqpair); 3423 if (!rqpair->srq) { 3424 nvme_rdma_qpair_submit_recvs(rqpair); 3425 } 3426 if (rqpair->num_completions > 0) { 3427 nvme_qpair_resubmit_requests(qpair, rqpair->num_completions); 3428 rqpair->num_completions = 0; 3429 } 3430 3431 if (rqpair->num_outstanding_reqs == 0 && STAILQ_EMPTY(&qpair->queued_req)) { 3432 TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active); 3433 /* We use prev pointer to check if qpair is in active list or not. 3434 * TAILQ_REMOVE doesn't do it. So, we do it manually. 3435 */ 3436 rqpair->link_active.tqe_prev = NULL; 3437 } 3438 } 3439 3440 static int64_t 3441 nvme_rdma_poll_group_process_completions(struct spdk_nvme_transport_poll_group *tgroup, 3442 uint32_t completions_per_qpair, spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb) 3443 { 3444 struct spdk_nvme_qpair *qpair, *tmp_qpair; 3445 struct nvme_rdma_qpair *rqpair, *tmp_rqpair; 3446 struct nvme_rdma_poll_group *group; 3447 struct nvme_rdma_poller *poller; 3448 int batch_size, rc, rc2 = 0; 3449 int64_t total_completions = 0; 3450 uint64_t completions_allowed = 0; 3451 uint64_t completions_per_poller = 0; 3452 uint64_t poller_completions = 0; 3453 uint64_t rdma_completions; 3454 3455 if (completions_per_qpair == 0) { 3456 completions_per_qpair = MAX_COMPLETIONS_PER_POLL; 3457 } 3458 3459 group = nvme_rdma_poll_group(tgroup); 3460 3461 STAILQ_FOREACH_SAFE(qpair, &tgroup->disconnected_qpairs, poll_group_stailq, tmp_qpair) { 3462 rc = nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair); 3463 if (rc == 0) { 3464 disconnected_qpair_cb(qpair, tgroup->group->ctx); 3465 } 3466 } 3467 3468 TAILQ_FOREACH_SAFE(rqpair, &group->connecting_qpairs, link_connecting, tmp_rqpair) { 3469 qpair = &rqpair->qpair; 3470 3471 rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair); 3472 if (rc == 0 || rc != -EAGAIN) { 3473 TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting); 3474 /* We use prev pointer to check if qpair is in connecting list or not. 3475 * TAILQ_REMOVE does not do it. So, we do it manually. 3476 */ 3477 rqpair->link_connecting.tqe_prev = NULL; 3478 3479 if (rc == 0) { 3480 /* Once the connection is completed, we can submit queued requests */ 3481 nvme_qpair_resubmit_requests(qpair, rqpair->num_entries); 3482 } else if (rc != -EAGAIN) { 3483 SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair); 3484 nvme_rdma_fail_qpair(qpair, 0); 3485 } 3486 } 3487 } 3488 3489 STAILQ_FOREACH_SAFE(qpair, &tgroup->connected_qpairs, poll_group_stailq, tmp_qpair) { 3490 rqpair = nvme_rdma_qpair(qpair); 3491 3492 if (spdk_likely(nvme_qpair_get_state(qpair) != NVME_QPAIR_CONNECTING)) { 3493 nvme_rdma_qpair_process_cm_event(rqpair); 3494 } 3495 3496 if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) { 3497 rc2 = -ENXIO; 3498 nvme_rdma_fail_qpair(qpair, 0); 3499 } 3500 } 3501 3502 completions_allowed = completions_per_qpair * tgroup->num_connected_qpairs; 3503 if (spdk_likely(group->num_pollers)) { 3504 completions_per_poller = spdk_max(completions_allowed / group->num_pollers, 1); 3505 } 3506 3507 STAILQ_FOREACH(poller, &group->pollers, link) { 3508 poller_completions = 0; 3509 rdma_completions = 0; 3510 do { 3511 poller->stats.polls++; 3512 batch_size = spdk_min((completions_per_poller - poller_completions), MAX_COMPLETIONS_PER_POLL); 3513 rc = nvme_rdma_cq_process_completions(poller->cq, batch_size, poller, NULL, &rdma_completions); 3514 if (rc <= 0) { 3515 if (rc == -ECANCELED) { 3516 return -EIO; 3517 } else if (rc == 0) { 3518 poller->stats.idle_polls++; 3519 } 3520 break; 3521 } 3522 3523 poller_completions += rc; 3524 } while (poller_completions < completions_per_poller); 3525 total_completions += poller_completions; 3526 poller->stats.completions += rdma_completions; 3527 if (poller->srq) { 3528 nvme_rdma_poller_submit_recvs(poller); 3529 } 3530 } 3531 3532 TAILQ_FOREACH_SAFE(rqpair, &group->active_qpairs, link_active, tmp_rqpair) { 3533 nvme_rdma_qpair_process_submits(group, rqpair); 3534 } 3535 3536 return rc2 != 0 ? rc2 : total_completions; 3537 } 3538 3539 /* 3540 * Handle disconnected qpairs when interrupt support gets added. 3541 */ 3542 static void 3543 nvme_rdma_poll_group_check_disconnected_qpairs(struct spdk_nvme_transport_poll_group *tgroup, 3544 spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb) 3545 { 3546 } 3547 3548 static int 3549 nvme_rdma_poll_group_destroy(struct spdk_nvme_transport_poll_group *tgroup) 3550 { 3551 struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(tgroup); 3552 3553 if (!STAILQ_EMPTY(&tgroup->connected_qpairs) || !STAILQ_EMPTY(&tgroup->disconnected_qpairs)) { 3554 return -EBUSY; 3555 } 3556 3557 nvme_rdma_poll_group_free_pollers(group); 3558 free(group); 3559 3560 return 0; 3561 } 3562 3563 static int 3564 nvme_rdma_poll_group_get_stats(struct spdk_nvme_transport_poll_group *tgroup, 3565 struct spdk_nvme_transport_poll_group_stat **_stats) 3566 { 3567 struct nvme_rdma_poll_group *group; 3568 struct spdk_nvme_transport_poll_group_stat *stats; 3569 struct spdk_nvme_rdma_device_stat *device_stat; 3570 struct nvme_rdma_poller *poller; 3571 uint32_t i = 0; 3572 3573 if (tgroup == NULL || _stats == NULL) { 3574 SPDK_ERRLOG("Invalid stats or group pointer\n"); 3575 return -EINVAL; 3576 } 3577 3578 group = nvme_rdma_poll_group(tgroup); 3579 stats = calloc(1, sizeof(*stats)); 3580 if (!stats) { 3581 SPDK_ERRLOG("Can't allocate memory for RDMA stats\n"); 3582 return -ENOMEM; 3583 } 3584 stats->trtype = SPDK_NVME_TRANSPORT_RDMA; 3585 stats->rdma.num_devices = group->num_pollers; 3586 3587 if (stats->rdma.num_devices == 0) { 3588 *_stats = stats; 3589 return 0; 3590 } 3591 3592 stats->rdma.device_stats = calloc(stats->rdma.num_devices, sizeof(*stats->rdma.device_stats)); 3593 if (!stats->rdma.device_stats) { 3594 SPDK_ERRLOG("Can't allocate memory for RDMA device stats\n"); 3595 free(stats); 3596 return -ENOMEM; 3597 } 3598 3599 STAILQ_FOREACH(poller, &group->pollers, link) { 3600 device_stat = &stats->rdma.device_stats[i]; 3601 device_stat->name = poller->device->device->name; 3602 device_stat->polls = poller->stats.polls; 3603 device_stat->idle_polls = poller->stats.idle_polls; 3604 device_stat->completions = poller->stats.completions; 3605 device_stat->queued_requests = poller->stats.queued_requests; 3606 device_stat->total_send_wrs = poller->stats.rdma_stats.send.num_submitted_wrs; 3607 device_stat->send_doorbell_updates = poller->stats.rdma_stats.send.doorbell_updates; 3608 device_stat->total_recv_wrs = poller->stats.rdma_stats.recv.num_submitted_wrs; 3609 device_stat->recv_doorbell_updates = poller->stats.rdma_stats.recv.doorbell_updates; 3610 i++; 3611 } 3612 3613 *_stats = stats; 3614 3615 return 0; 3616 } 3617 3618 static void 3619 nvme_rdma_poll_group_free_stats(struct spdk_nvme_transport_poll_group *tgroup, 3620 struct spdk_nvme_transport_poll_group_stat *stats) 3621 { 3622 if (stats) { 3623 free(stats->rdma.device_stats); 3624 } 3625 free(stats); 3626 } 3627 3628 static int 3629 nvme_rdma_ctrlr_get_memory_domains(const struct spdk_nvme_ctrlr *ctrlr, 3630 struct spdk_memory_domain **domains, int array_size) 3631 { 3632 struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(ctrlr->adminq); 3633 3634 if (domains && array_size > 0) { 3635 domains[0] = rqpair->rdma_qp->domain; 3636 } 3637 3638 return 1; 3639 } 3640 3641 void 3642 spdk_nvme_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3643 { 3644 g_nvme_hooks = *hooks; 3645 } 3646 3647 const struct spdk_nvme_transport_ops rdma_ops = { 3648 .name = "RDMA", 3649 .type = SPDK_NVME_TRANSPORT_RDMA, 3650 .ctrlr_construct = nvme_rdma_ctrlr_construct, 3651 .ctrlr_scan = nvme_fabric_ctrlr_scan, 3652 .ctrlr_destruct = nvme_rdma_ctrlr_destruct, 3653 .ctrlr_enable = nvme_rdma_ctrlr_enable, 3654 3655 .ctrlr_set_reg_4 = nvme_fabric_ctrlr_set_reg_4, 3656 .ctrlr_set_reg_8 = nvme_fabric_ctrlr_set_reg_8, 3657 .ctrlr_get_reg_4 = nvme_fabric_ctrlr_get_reg_4, 3658 .ctrlr_get_reg_8 = nvme_fabric_ctrlr_get_reg_8, 3659 .ctrlr_set_reg_4_async = nvme_fabric_ctrlr_set_reg_4_async, 3660 .ctrlr_set_reg_8_async = nvme_fabric_ctrlr_set_reg_8_async, 3661 .ctrlr_get_reg_4_async = nvme_fabric_ctrlr_get_reg_4_async, 3662 .ctrlr_get_reg_8_async = nvme_fabric_ctrlr_get_reg_8_async, 3663 3664 .ctrlr_get_max_xfer_size = nvme_rdma_ctrlr_get_max_xfer_size, 3665 .ctrlr_get_max_sges = nvme_rdma_ctrlr_get_max_sges, 3666 3667 .ctrlr_create_io_qpair = nvme_rdma_ctrlr_create_io_qpair, 3668 .ctrlr_delete_io_qpair = nvme_rdma_ctrlr_delete_io_qpair, 3669 .ctrlr_connect_qpair = nvme_rdma_ctrlr_connect_qpair, 3670 .ctrlr_disconnect_qpair = nvme_rdma_ctrlr_disconnect_qpair, 3671 3672 .ctrlr_get_memory_domains = nvme_rdma_ctrlr_get_memory_domains, 3673 3674 .qpair_abort_reqs = nvme_rdma_qpair_abort_reqs, 3675 .qpair_reset = nvme_rdma_qpair_reset, 3676 .qpair_submit_request = nvme_rdma_qpair_submit_request, 3677 .qpair_process_completions = nvme_rdma_qpair_process_completions, 3678 .qpair_iterate_requests = nvme_rdma_qpair_iterate_requests, 3679 .qpair_authenticate = nvme_rdma_qpair_authenticate, 3680 .admin_qpair_abort_aers = nvme_rdma_admin_qpair_abort_aers, 3681 3682 .poll_group_create = nvme_rdma_poll_group_create, 3683 .poll_group_connect_qpair = nvme_rdma_poll_group_connect_qpair, 3684 .poll_group_disconnect_qpair = nvme_rdma_poll_group_disconnect_qpair, 3685 .poll_group_add = nvme_rdma_poll_group_add, 3686 .poll_group_remove = nvme_rdma_poll_group_remove, 3687 .poll_group_process_completions = nvme_rdma_poll_group_process_completions, 3688 .poll_group_check_disconnected_qpairs = nvme_rdma_poll_group_check_disconnected_qpairs, 3689 .poll_group_destroy = nvme_rdma_poll_group_destroy, 3690 .poll_group_get_stats = nvme_rdma_poll_group_get_stats, 3691 .poll_group_free_stats = nvme_rdma_poll_group_free_stats, 3692 }; 3693 3694 SPDK_NVME_TRANSPORT_REGISTER(rdma, &rdma_ops); 3695