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, 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved. 5 */ 6 7 #include "spdk/stdinc.h" 8 9 #include "spdk/config.h" 10 #include "spdk/thread.h" 11 #include "spdk/likely.h" 12 #include "spdk/nvmf_transport.h" 13 #include "spdk/string.h" 14 #include "spdk/trace.h" 15 #include "spdk/tree.h" 16 #include "spdk/util.h" 17 18 #include "spdk_internal/assert.h" 19 #include "spdk/log.h" 20 #include "spdk_internal/rdma.h" 21 22 #include "nvmf_internal.h" 23 #include "transport.h" 24 25 #include "spdk_internal/trace_defs.h" 26 27 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 28 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 29 30 /* 31 RDMA Connection Resource Defaults 32 */ 33 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 34 #define NVMF_DEFAULT_RSP_SGE 1 35 #define NVMF_DEFAULT_RX_SGE 2 36 37 /* The RDMA completion queue size */ 38 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 39 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 40 41 static int g_spdk_nvmf_ibv_query_mask = 42 IBV_QP_STATE | 43 IBV_QP_PKEY_INDEX | 44 IBV_QP_PORT | 45 IBV_QP_ACCESS_FLAGS | 46 IBV_QP_AV | 47 IBV_QP_PATH_MTU | 48 IBV_QP_DEST_QPN | 49 IBV_QP_RQ_PSN | 50 IBV_QP_MAX_DEST_RD_ATOMIC | 51 IBV_QP_MIN_RNR_TIMER | 52 IBV_QP_SQ_PSN | 53 IBV_QP_TIMEOUT | 54 IBV_QP_RETRY_CNT | 55 IBV_QP_RNR_RETRY | 56 IBV_QP_MAX_QP_RD_ATOMIC; 57 58 enum spdk_nvmf_rdma_request_state { 59 /* The request is not currently in use */ 60 RDMA_REQUEST_STATE_FREE = 0, 61 62 /* Initial state when request first received */ 63 RDMA_REQUEST_STATE_NEW, 64 65 /* The request is queued until a data buffer is available. */ 66 RDMA_REQUEST_STATE_NEED_BUFFER, 67 68 /* The request is waiting on RDMA queue depth availability 69 * to transfer data from the host to the controller. 70 */ 71 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 72 73 /* The request is currently transferring data from the host to the controller. */ 74 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 75 76 /* The request is ready to execute at the block device */ 77 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 78 79 /* The request is currently executing at the block device */ 80 RDMA_REQUEST_STATE_EXECUTING, 81 82 /* The request finished executing at the block device */ 83 RDMA_REQUEST_STATE_EXECUTED, 84 85 /* The request is waiting on RDMA queue depth availability 86 * to transfer data from the controller to the host. 87 */ 88 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 89 90 /* The request is ready to send a completion */ 91 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 92 93 /* The request is currently transferring data from the controller to the host. */ 94 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 95 96 /* The request currently has an outstanding completion without an 97 * associated data transfer. 98 */ 99 RDMA_REQUEST_STATE_COMPLETING, 100 101 /* The request completed and can be marked free. */ 102 RDMA_REQUEST_STATE_COMPLETED, 103 104 /* Terminator */ 105 RDMA_REQUEST_NUM_STATES, 106 }; 107 108 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 109 { 110 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 111 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 112 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 113 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 114 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 115 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 116 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 117 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 118 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 119 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 120 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 121 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 122 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 123 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 124 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 125 spdk_trace_register_description("RDMA_REQ_TX_H2C", 126 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 127 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 128 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 129 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 130 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 131 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 132 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 133 spdk_trace_register_description("RDMA_REQ_EXECUTING", 134 TRACE_RDMA_REQUEST_STATE_EXECUTING, 135 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 136 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 137 spdk_trace_register_description("RDMA_REQ_EXECUTED", 138 TRACE_RDMA_REQUEST_STATE_EXECUTED, 139 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 140 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 141 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 142 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 143 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 144 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 145 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 146 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 147 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 148 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 149 spdk_trace_register_description("RDMA_REQ_COMPLETING", 150 TRACE_RDMA_REQUEST_STATE_COMPLETING, 151 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 152 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 153 spdk_trace_register_description("RDMA_REQ_COMPLETED", 154 TRACE_RDMA_REQUEST_STATE_COMPLETED, 155 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 156 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 157 158 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 159 OWNER_NONE, OBJECT_NONE, 0, 160 SPDK_TRACE_ARG_TYPE_INT, ""); 161 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 162 OWNER_NONE, OBJECT_NONE, 0, 163 SPDK_TRACE_ARG_TYPE_INT, "type"); 164 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 165 OWNER_NONE, OBJECT_NONE, 0, 166 SPDK_TRACE_ARG_TYPE_INT, "type"); 167 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 168 OWNER_NONE, OBJECT_NONE, 0, 169 SPDK_TRACE_ARG_TYPE_PTR, "state"); 170 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 171 OWNER_NONE, OBJECT_NONE, 0, 172 SPDK_TRACE_ARG_TYPE_INT, ""); 173 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 174 OWNER_NONE, OBJECT_NONE, 0, 175 SPDK_TRACE_ARG_TYPE_INT, ""); 176 } 177 178 enum spdk_nvmf_rdma_wr_type { 179 RDMA_WR_TYPE_RECV, 180 RDMA_WR_TYPE_SEND, 181 RDMA_WR_TYPE_DATA, 182 }; 183 184 struct spdk_nvmf_rdma_wr { 185 enum spdk_nvmf_rdma_wr_type type; 186 }; 187 188 /* This structure holds commands as they are received off the wire. 189 * It must be dynamically paired with a full request object 190 * (spdk_nvmf_rdma_request) to service a request. It is separate 191 * from the request because RDMA does not appear to order 192 * completions, so occasionally we'll get a new incoming 193 * command when there aren't any free request objects. 194 */ 195 struct spdk_nvmf_rdma_recv { 196 struct ibv_recv_wr wr; 197 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 198 199 struct spdk_nvmf_rdma_qpair *qpair; 200 201 /* In-capsule data buffer */ 202 uint8_t *buf; 203 204 struct spdk_nvmf_rdma_wr rdma_wr; 205 uint64_t receive_tsc; 206 207 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 208 }; 209 210 struct spdk_nvmf_rdma_request_data { 211 struct spdk_nvmf_rdma_wr rdma_wr; 212 struct ibv_send_wr wr; 213 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 214 }; 215 216 struct spdk_nvmf_rdma_request { 217 struct spdk_nvmf_request req; 218 219 enum spdk_nvmf_rdma_request_state state; 220 221 /* Data offset in req.iov */ 222 uint32_t offset; 223 224 struct spdk_nvmf_rdma_recv *recv; 225 226 struct { 227 struct spdk_nvmf_rdma_wr rdma_wr; 228 struct ibv_send_wr wr; 229 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 230 } rsp; 231 232 struct spdk_nvmf_rdma_request_data data; 233 234 uint32_t iovpos; 235 236 uint32_t num_outstanding_data_wr; 237 uint64_t receive_tsc; 238 239 bool fused_failed; 240 struct spdk_nvmf_rdma_request *fused_pair; 241 242 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 243 }; 244 245 struct spdk_nvmf_rdma_resource_opts { 246 struct spdk_nvmf_rdma_qpair *qpair; 247 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 248 void *qp; 249 struct spdk_rdma_mem_map *map; 250 uint32_t max_queue_depth; 251 uint32_t in_capsule_data_size; 252 bool shared; 253 }; 254 255 struct spdk_nvmf_rdma_resources { 256 /* Array of size "max_queue_depth" containing RDMA requests. */ 257 struct spdk_nvmf_rdma_request *reqs; 258 259 /* Array of size "max_queue_depth" containing RDMA recvs. */ 260 struct spdk_nvmf_rdma_recv *recvs; 261 262 /* Array of size "max_queue_depth" containing 64 byte capsules 263 * used for receive. 264 */ 265 union nvmf_h2c_msg *cmds; 266 267 /* Array of size "max_queue_depth" containing 16 byte completions 268 * to be sent back to the user. 269 */ 270 union nvmf_c2h_msg *cpls; 271 272 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 273 * buffers to be used for in capsule data. 274 */ 275 void *bufs; 276 277 /* Receives that are waiting for a request object */ 278 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 279 280 /* Queue to track free requests */ 281 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 282 }; 283 284 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 285 286 struct spdk_nvmf_rdma_ibv_event_ctx { 287 struct spdk_nvmf_rdma_qpair *rqpair; 288 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 289 /* Link to other ibv events associated with this qpair */ 290 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 291 }; 292 293 struct spdk_nvmf_rdma_qpair { 294 struct spdk_nvmf_qpair qpair; 295 296 struct spdk_nvmf_rdma_device *device; 297 struct spdk_nvmf_rdma_poller *poller; 298 299 struct spdk_rdma_qp *rdma_qp; 300 struct rdma_cm_id *cm_id; 301 struct spdk_rdma_srq *srq; 302 struct rdma_cm_id *listen_id; 303 304 /* Cache the QP number to improve QP search by RB tree. */ 305 uint32_t qp_num; 306 307 /* The maximum number of I/O outstanding on this connection at one time */ 308 uint16_t max_queue_depth; 309 310 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 311 uint16_t max_read_depth; 312 313 /* The maximum number of RDMA SEND operations at one time */ 314 uint32_t max_send_depth; 315 316 /* The current number of outstanding WRs from this qpair's 317 * recv queue. Should not exceed device->attr.max_queue_depth. 318 */ 319 uint16_t current_recv_depth; 320 321 /* The current number of active RDMA READ operations */ 322 uint16_t current_read_depth; 323 324 /* The current number of posted WRs from this qpair's 325 * send queue. Should not exceed max_send_depth. 326 */ 327 uint32_t current_send_depth; 328 329 /* The maximum number of SGEs per WR on the send queue */ 330 uint32_t max_send_sge; 331 332 /* The maximum number of SGEs per WR on the recv queue */ 333 uint32_t max_recv_sge; 334 335 struct spdk_nvmf_rdma_resources *resources; 336 337 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 338 339 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 340 341 /* Number of requests not in the free state */ 342 uint32_t qd; 343 344 RB_ENTRY(spdk_nvmf_rdma_qpair) node; 345 346 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 347 348 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 349 350 /* IBV queue pair attributes: they are used to manage 351 * qp state and recover from errors. 352 */ 353 enum ibv_qp_state ibv_state; 354 355 /* Points to the a request that has fuse bits set to 356 * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting 357 * for the request that has SPDK_NVME_CMD_FUSE_SECOND. 358 */ 359 struct spdk_nvmf_rdma_request *fused_first; 360 361 /* 362 * io_channel which is used to destroy qpair when it is removed from poll group 363 */ 364 struct spdk_io_channel *destruct_channel; 365 366 /* List of ibv async events */ 367 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 368 369 /* Lets us know that we have received the last_wqe event. */ 370 bool last_wqe_reached; 371 372 /* Indicate that nvmf_rdma_close_qpair is called */ 373 bool to_close; 374 }; 375 376 struct spdk_nvmf_rdma_poller_stat { 377 uint64_t completions; 378 uint64_t polls; 379 uint64_t idle_polls; 380 uint64_t requests; 381 uint64_t request_latency; 382 uint64_t pending_free_request; 383 uint64_t pending_rdma_read; 384 uint64_t pending_rdma_write; 385 struct spdk_rdma_qp_stats qp_stats; 386 }; 387 388 struct spdk_nvmf_rdma_poller { 389 struct spdk_nvmf_rdma_device *device; 390 struct spdk_nvmf_rdma_poll_group *group; 391 392 int num_cqe; 393 int required_num_wr; 394 struct ibv_cq *cq; 395 396 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 397 uint16_t max_srq_depth; 398 399 /* Shared receive queue */ 400 struct spdk_rdma_srq *srq; 401 402 struct spdk_nvmf_rdma_resources *resources; 403 struct spdk_nvmf_rdma_poller_stat stat; 404 405 RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs; 406 407 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 408 409 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 410 411 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 412 }; 413 414 struct spdk_nvmf_rdma_poll_group_stat { 415 uint64_t pending_data_buffer; 416 }; 417 418 struct spdk_nvmf_rdma_poll_group { 419 struct spdk_nvmf_transport_poll_group group; 420 struct spdk_nvmf_rdma_poll_group_stat stat; 421 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 422 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 423 }; 424 425 struct spdk_nvmf_rdma_conn_sched { 426 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 427 struct spdk_nvmf_rdma_poll_group *next_io_pg; 428 }; 429 430 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 431 struct spdk_nvmf_rdma_device { 432 struct ibv_device_attr attr; 433 struct ibv_context *context; 434 435 struct spdk_rdma_mem_map *map; 436 struct ibv_pd *pd; 437 438 int num_srq; 439 440 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 441 }; 442 443 struct spdk_nvmf_rdma_port { 444 const struct spdk_nvme_transport_id *trid; 445 struct rdma_cm_id *id; 446 struct spdk_nvmf_rdma_device *device; 447 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 448 }; 449 450 struct rdma_transport_opts { 451 int num_cqe; 452 uint32_t max_srq_depth; 453 bool no_srq; 454 bool no_wr_batching; 455 int acceptor_backlog; 456 }; 457 458 struct spdk_nvmf_rdma_transport { 459 struct spdk_nvmf_transport transport; 460 struct rdma_transport_opts rdma_opts; 461 462 struct spdk_nvmf_rdma_conn_sched conn_sched; 463 464 struct rdma_event_channel *event_channel; 465 466 struct spdk_mempool *data_wr_pool; 467 468 struct spdk_poller *accept_poller; 469 470 /* fields used to poll RDMA/IB events */ 471 nfds_t npoll_fds; 472 struct pollfd *poll_fds; 473 474 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 475 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 476 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 477 }; 478 479 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 480 { 481 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe), 482 spdk_json_decode_int32, true 483 }, 484 { 485 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 486 spdk_json_decode_uint32, true 487 }, 488 { 489 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 490 spdk_json_decode_bool, true 491 }, 492 { 493 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 494 spdk_json_decode_bool, true 495 }, 496 { 497 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 498 spdk_json_decode_int32, true 499 }, 500 }; 501 502 static int 503 nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2) 504 { 505 return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num; 506 } 507 508 RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare); 509 510 static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 511 struct spdk_nvmf_rdma_request *rdma_req); 512 513 static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 514 struct spdk_nvmf_rdma_poller *rpoller); 515 516 static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 517 struct spdk_nvmf_rdma_poller *rpoller); 518 519 static inline int 520 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 521 { 522 switch (state) { 523 case IBV_QPS_RESET: 524 case IBV_QPS_INIT: 525 case IBV_QPS_RTR: 526 case IBV_QPS_RTS: 527 case IBV_QPS_SQD: 528 case IBV_QPS_SQE: 529 case IBV_QPS_ERR: 530 return 0; 531 default: 532 return -1; 533 } 534 } 535 536 static inline enum spdk_nvme_media_error_status_code 537 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 538 enum spdk_nvme_media_error_status_code result; 539 switch (err_type) 540 { 541 case SPDK_DIF_REFTAG_ERROR: 542 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 543 break; 544 case SPDK_DIF_APPTAG_ERROR: 545 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 546 break; 547 case SPDK_DIF_GUARD_ERROR: 548 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 549 break; 550 default: 551 SPDK_UNREACHABLE(); 552 } 553 554 return result; 555 } 556 557 static enum ibv_qp_state 558 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 559 enum ibv_qp_state old_state, new_state; 560 struct ibv_qp_attr qp_attr; 561 struct ibv_qp_init_attr init_attr; 562 int rc; 563 564 old_state = rqpair->ibv_state; 565 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 566 g_spdk_nvmf_ibv_query_mask, &init_attr); 567 568 if (rc) 569 { 570 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 571 return IBV_QPS_ERR + 1; 572 } 573 574 new_state = qp_attr.qp_state; 575 rqpair->ibv_state = new_state; 576 qp_attr.ah_attr.port_num = qp_attr.port_num; 577 578 rc = nvmf_rdma_check_ibv_state(new_state); 579 if (rc) 580 { 581 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 582 /* 583 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 584 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 585 */ 586 return IBV_QPS_ERR + 1; 587 } 588 589 if (old_state != new_state) 590 { 591 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, (uintptr_t)rqpair, new_state); 592 } 593 return new_state; 594 } 595 596 static void 597 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 598 struct spdk_nvmf_rdma_transport *rtransport) 599 { 600 struct spdk_nvmf_rdma_request_data *data_wr; 601 struct ibv_send_wr *next_send_wr; 602 uint64_t req_wrid; 603 604 rdma_req->num_outstanding_data_wr = 0; 605 data_wr = &rdma_req->data; 606 req_wrid = data_wr->wr.wr_id; 607 while (data_wr && data_wr->wr.wr_id == req_wrid) { 608 memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge); 609 data_wr->wr.num_sge = 0; 610 next_send_wr = data_wr->wr.next; 611 if (data_wr != &rdma_req->data) { 612 data_wr->wr.next = NULL; 613 spdk_mempool_put(rtransport->data_wr_pool, data_wr); 614 } 615 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : 616 SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr); 617 } 618 rdma_req->data.wr.next = NULL; 619 rdma_req->rsp.wr.next = NULL; 620 } 621 622 static void 623 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 624 { 625 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 626 if (req->req.cmd) { 627 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 628 } 629 if (req->recv) { 630 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 631 } 632 } 633 634 static void 635 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 636 { 637 int i; 638 639 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 640 for (i = 0; i < rqpair->max_queue_depth; i++) { 641 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 642 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 643 } 644 } 645 } 646 647 static void 648 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 649 { 650 spdk_free(resources->cmds); 651 spdk_free(resources->cpls); 652 spdk_free(resources->bufs); 653 spdk_free(resources->reqs); 654 spdk_free(resources->recvs); 655 free(resources); 656 } 657 658 659 static struct spdk_nvmf_rdma_resources * 660 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 661 { 662 struct spdk_nvmf_rdma_resources *resources; 663 struct spdk_nvmf_rdma_request *rdma_req; 664 struct spdk_nvmf_rdma_recv *rdma_recv; 665 struct spdk_rdma_qp *qp = NULL; 666 struct spdk_rdma_srq *srq = NULL; 667 struct ibv_recv_wr *bad_wr = NULL; 668 struct spdk_rdma_memory_translation translation; 669 uint32_t i; 670 int rc = 0; 671 672 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 673 if (!resources) { 674 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 675 return NULL; 676 } 677 678 resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs), 679 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 680 resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs), 681 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 682 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 683 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 684 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 685 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 686 687 if (opts->in_capsule_data_size > 0) { 688 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 689 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 690 SPDK_MALLOC_DMA); 691 } 692 693 if (!resources->reqs || !resources->recvs || !resources->cmds || 694 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 695 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 696 goto cleanup; 697 } 698 699 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n", 700 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds)); 701 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n", 702 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls)); 703 if (resources->bufs) { 704 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n", 705 resources->bufs, opts->max_queue_depth * 706 opts->in_capsule_data_size); 707 } 708 709 /* Initialize queues */ 710 STAILQ_INIT(&resources->incoming_queue); 711 STAILQ_INIT(&resources->free_queue); 712 713 if (opts->shared) { 714 srq = (struct spdk_rdma_srq *)opts->qp; 715 } else { 716 qp = (struct spdk_rdma_qp *)opts->qp; 717 } 718 719 for (i = 0; i < opts->max_queue_depth; i++) { 720 rdma_recv = &resources->recvs[i]; 721 rdma_recv->qpair = opts->qpair; 722 723 /* Set up memory to receive commands */ 724 if (resources->bufs) { 725 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 726 opts->in_capsule_data_size)); 727 } 728 729 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 730 731 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 732 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 733 rc = spdk_rdma_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]), 734 &translation); 735 if (rc) { 736 goto cleanup; 737 } 738 rdma_recv->sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 739 rdma_recv->wr.num_sge = 1; 740 741 if (rdma_recv->buf) { 742 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 743 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 744 rc = spdk_rdma_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size, &translation); 745 if (rc) { 746 goto cleanup; 747 } 748 rdma_recv->sgl[1].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 749 rdma_recv->wr.num_sge++; 750 } 751 752 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 753 rdma_recv->wr.sg_list = rdma_recv->sgl; 754 if (srq) { 755 spdk_rdma_srq_queue_recv_wrs(srq, &rdma_recv->wr); 756 } else { 757 spdk_rdma_qp_queue_recv_wrs(qp, &rdma_recv->wr); 758 } 759 } 760 761 for (i = 0; i < opts->max_queue_depth; i++) { 762 rdma_req = &resources->reqs[i]; 763 764 if (opts->qpair != NULL) { 765 rdma_req->req.qpair = &opts->qpair->qpair; 766 } else { 767 rdma_req->req.qpair = NULL; 768 } 769 rdma_req->req.cmd = NULL; 770 rdma_req->req.iovcnt = 0; 771 rdma_req->req.stripped_data = NULL; 772 773 /* Set up memory to send responses */ 774 rdma_req->req.rsp = &resources->cpls[i]; 775 776 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 777 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 778 rc = spdk_rdma_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]), 779 &translation); 780 if (rc) { 781 goto cleanup; 782 } 783 rdma_req->rsp.sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 784 785 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 786 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 787 rdma_req->rsp.wr.next = NULL; 788 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 789 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 790 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 791 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 792 793 /* Set up memory for data buffers */ 794 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 795 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 796 rdma_req->data.wr.next = NULL; 797 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 798 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 799 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 800 801 /* Initialize request state to FREE */ 802 rdma_req->state = RDMA_REQUEST_STATE_FREE; 803 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 804 } 805 806 if (srq) { 807 rc = spdk_rdma_srq_flush_recv_wrs(srq, &bad_wr); 808 } else { 809 rc = spdk_rdma_qp_flush_recv_wrs(qp, &bad_wr); 810 } 811 812 if (rc) { 813 goto cleanup; 814 } 815 816 return resources; 817 818 cleanup: 819 nvmf_rdma_resources_destroy(resources); 820 return NULL; 821 } 822 823 static void 824 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 825 { 826 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 827 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 828 ctx->rqpair = NULL; 829 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 830 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 831 } 832 } 833 834 static void 835 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 836 { 837 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 838 struct ibv_recv_wr *bad_recv_wr = NULL; 839 int rc; 840 841 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair); 842 843 if (rqpair->qd != 0) { 844 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 845 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 846 struct spdk_nvmf_rdma_transport, transport); 847 struct spdk_nvmf_rdma_request *req; 848 uint32_t i, max_req_count = 0; 849 850 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 851 852 if (rqpair->srq == NULL) { 853 nvmf_rdma_dump_qpair_contents(rqpair); 854 max_req_count = rqpair->max_queue_depth; 855 } else if (rqpair->poller && rqpair->resources) { 856 max_req_count = rqpair->poller->max_srq_depth; 857 } 858 859 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 860 for (i = 0; i < max_req_count; i++) { 861 req = &rqpair->resources->reqs[i]; 862 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 863 /* nvmf_rdma_request_process checks qpair ibv and internal state 864 * and completes a request */ 865 nvmf_rdma_request_process(rtransport, req); 866 } 867 } 868 assert(rqpair->qd == 0); 869 } 870 871 if (rqpair->poller) { 872 RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair); 873 874 if (rqpair->srq != NULL && rqpair->resources != NULL) { 875 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 876 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 877 if (rqpair == rdma_recv->qpair) { 878 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 879 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr); 880 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 881 if (rc) { 882 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 883 } 884 } 885 } 886 } 887 } 888 889 if (rqpair->cm_id) { 890 if (rqpair->rdma_qp != NULL) { 891 spdk_rdma_qp_destroy(rqpair->rdma_qp); 892 rqpair->rdma_qp = NULL; 893 } 894 rdma_destroy_id(rqpair->cm_id); 895 896 if (rqpair->poller != NULL && rqpair->srq == NULL) { 897 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 898 } 899 } 900 901 if (rqpair->srq == NULL && rqpair->resources != NULL) { 902 nvmf_rdma_resources_destroy(rqpair->resources); 903 } 904 905 nvmf_rdma_qpair_clean_ibv_events(rqpair); 906 907 if (rqpair->destruct_channel) { 908 spdk_put_io_channel(rqpair->destruct_channel); 909 rqpair->destruct_channel = NULL; 910 } 911 912 free(rqpair); 913 } 914 915 static int 916 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 917 { 918 struct spdk_nvmf_rdma_poller *rpoller; 919 int rc, num_cqe, required_num_wr; 920 921 /* Enlarge CQ size dynamically */ 922 rpoller = rqpair->poller; 923 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 924 num_cqe = rpoller->num_cqe; 925 if (num_cqe < required_num_wr) { 926 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 927 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 928 } 929 930 if (rpoller->num_cqe != num_cqe) { 931 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) { 932 SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n" 933 "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe); 934 return -1; 935 } 936 if (required_num_wr > device->attr.max_cqe) { 937 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 938 required_num_wr, device->attr.max_cqe); 939 return -1; 940 } 941 942 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 943 rc = ibv_resize_cq(rpoller->cq, num_cqe); 944 if (rc) { 945 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 946 return -1; 947 } 948 949 rpoller->num_cqe = num_cqe; 950 } 951 952 rpoller->required_num_wr = required_num_wr; 953 return 0; 954 } 955 956 static int 957 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 958 { 959 struct spdk_nvmf_rdma_qpair *rqpair; 960 struct spdk_nvmf_rdma_transport *rtransport; 961 struct spdk_nvmf_transport *transport; 962 struct spdk_nvmf_rdma_resource_opts opts; 963 struct spdk_nvmf_rdma_device *device; 964 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 965 966 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 967 device = rqpair->device; 968 969 qp_init_attr.qp_context = rqpair; 970 qp_init_attr.pd = device->pd; 971 qp_init_attr.send_cq = rqpair->poller->cq; 972 qp_init_attr.recv_cq = rqpair->poller->cq; 973 974 if (rqpair->srq) { 975 qp_init_attr.srq = rqpair->srq->srq; 976 } else { 977 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 978 } 979 980 /* SEND, READ, and WRITE operations */ 981 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 982 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 983 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 984 qp_init_attr.stats = &rqpair->poller->stat.qp_stats; 985 986 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 987 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 988 goto error; 989 } 990 991 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 992 if (!rqpair->rdma_qp) { 993 goto error; 994 } 995 996 rqpair->qp_num = rqpair->rdma_qp->qp->qp_num; 997 998 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 999 qp_init_attr.cap.max_send_wr); 1000 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1001 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1002 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair); 1003 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1004 1005 if (rqpair->poller->srq == NULL) { 1006 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1007 transport = &rtransport->transport; 1008 1009 opts.qp = rqpair->rdma_qp; 1010 opts.map = device->map; 1011 opts.qpair = rqpair; 1012 opts.shared = false; 1013 opts.max_queue_depth = rqpair->max_queue_depth; 1014 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1015 1016 rqpair->resources = nvmf_rdma_resources_create(&opts); 1017 1018 if (!rqpair->resources) { 1019 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1020 rdma_destroy_qp(rqpair->cm_id); 1021 goto error; 1022 } 1023 } else { 1024 rqpair->resources = rqpair->poller->resources; 1025 } 1026 1027 rqpair->current_recv_depth = 0; 1028 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1029 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1030 1031 return 0; 1032 1033 error: 1034 rdma_destroy_id(rqpair->cm_id); 1035 rqpair->cm_id = NULL; 1036 return -1; 1037 } 1038 1039 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1040 /* This function accepts either a single wr or the first wr in a linked list. */ 1041 static void 1042 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1043 { 1044 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1045 struct spdk_nvmf_rdma_transport, transport); 1046 1047 if (rqpair->srq != NULL) { 1048 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, first); 1049 } else { 1050 if (spdk_rdma_qp_queue_recv_wrs(rqpair->rdma_qp, first)) { 1051 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1052 } 1053 } 1054 1055 if (rtransport->rdma_opts.no_wr_batching) { 1056 _poller_submit_recvs(rtransport, rqpair->poller); 1057 } 1058 } 1059 1060 static int 1061 request_transfer_in(struct spdk_nvmf_request *req) 1062 { 1063 struct spdk_nvmf_rdma_request *rdma_req; 1064 struct spdk_nvmf_qpair *qpair; 1065 struct spdk_nvmf_rdma_qpair *rqpair; 1066 struct spdk_nvmf_rdma_transport *rtransport; 1067 1068 qpair = req->qpair; 1069 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1070 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1071 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1072 struct spdk_nvmf_rdma_transport, transport); 1073 1074 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1075 assert(rdma_req != NULL); 1076 1077 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1078 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1079 } 1080 if (rtransport->rdma_opts.no_wr_batching) { 1081 _poller_submit_sends(rtransport, rqpair->poller); 1082 } 1083 1084 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1085 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1086 return 0; 1087 } 1088 1089 static int 1090 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1091 { 1092 int num_outstanding_data_wr = 0; 1093 struct spdk_nvmf_rdma_request *rdma_req; 1094 struct spdk_nvmf_qpair *qpair; 1095 struct spdk_nvmf_rdma_qpair *rqpair; 1096 struct spdk_nvme_cpl *rsp; 1097 struct ibv_send_wr *first = NULL; 1098 struct spdk_nvmf_rdma_transport *rtransport; 1099 1100 *data_posted = 0; 1101 qpair = req->qpair; 1102 rsp = &req->rsp->nvme_cpl; 1103 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1104 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1105 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1106 struct spdk_nvmf_rdma_transport, transport); 1107 1108 /* Advance our sq_head pointer */ 1109 if (qpair->sq_head == qpair->sq_head_max) { 1110 qpair->sq_head = 0; 1111 } else { 1112 qpair->sq_head++; 1113 } 1114 rsp->sqhd = qpair->sq_head; 1115 1116 /* queue the capsule for the recv buffer */ 1117 assert(rdma_req->recv != NULL); 1118 1119 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1120 1121 rdma_req->recv = NULL; 1122 assert(rqpair->current_recv_depth > 0); 1123 rqpair->current_recv_depth--; 1124 1125 /* Build the response which consists of optional 1126 * RDMA WRITEs to transfer data, plus an RDMA SEND 1127 * containing the response. 1128 */ 1129 first = &rdma_req->rsp.wr; 1130 1131 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1132 /* On failure, data was not read from the controller. So clear the 1133 * number of outstanding data WRs to zero. 1134 */ 1135 rdma_req->num_outstanding_data_wr = 0; 1136 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1137 first = &rdma_req->data.wr; 1138 *data_posted = 1; 1139 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1140 } 1141 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1142 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1143 } 1144 if (rtransport->rdma_opts.no_wr_batching) { 1145 _poller_submit_sends(rtransport, rqpair->poller); 1146 } 1147 1148 /* +1 for the rsp wr */ 1149 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1150 1151 return 0; 1152 } 1153 1154 static int 1155 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1156 { 1157 struct spdk_nvmf_rdma_accept_private_data accept_data; 1158 struct rdma_conn_param ctrlr_event_data = {}; 1159 int rc; 1160 1161 accept_data.recfmt = 0; 1162 accept_data.crqsize = rqpair->max_queue_depth; 1163 1164 ctrlr_event_data.private_data = &accept_data; 1165 ctrlr_event_data.private_data_len = sizeof(accept_data); 1166 if (id->ps == RDMA_PS_TCP) { 1167 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1168 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1169 } 1170 1171 /* Configure infinite retries for the initiator side qpair. 1172 * We need to pass this value to the initiator to prevent the 1173 * initiator side NIC from completing SEND requests back to the 1174 * initiator with status rnr_retry_count_exceeded. */ 1175 ctrlr_event_data.rnr_retry_count = 0x7; 1176 1177 /* When qpair is created without use of rdma cm API, an additional 1178 * information must be provided to initiator in the connection response: 1179 * whether qpair is using SRQ and its qp_num 1180 * Fields below are ignored by rdma cm if qpair has been 1181 * created using rdma cm API. */ 1182 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1183 ctrlr_event_data.qp_num = rqpair->qp_num; 1184 1185 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1186 if (rc) { 1187 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1188 } else { 1189 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1190 } 1191 1192 return rc; 1193 } 1194 1195 static void 1196 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1197 { 1198 struct spdk_nvmf_rdma_reject_private_data rej_data; 1199 1200 rej_data.recfmt = 0; 1201 rej_data.sts = error; 1202 1203 rdma_reject(id, &rej_data, sizeof(rej_data)); 1204 } 1205 1206 static int 1207 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1208 { 1209 struct spdk_nvmf_rdma_transport *rtransport; 1210 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1211 struct spdk_nvmf_rdma_port *port; 1212 struct rdma_conn_param *rdma_param = NULL; 1213 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1214 uint16_t max_queue_depth; 1215 uint16_t max_read_depth; 1216 1217 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1218 1219 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1220 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1221 1222 rdma_param = &event->param.conn; 1223 if (rdma_param->private_data == NULL || 1224 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1225 SPDK_ERRLOG("connect request: no private data provided\n"); 1226 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1227 return -1; 1228 } 1229 1230 private_data = rdma_param->private_data; 1231 if (private_data->recfmt != 0) { 1232 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1233 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1234 return -1; 1235 } 1236 1237 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1238 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1239 1240 port = event->listen_id->context; 1241 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1242 event->listen_id, event->listen_id->verbs, port); 1243 1244 /* Figure out the supported queue depth. This is a multi-step process 1245 * that takes into account hardware maximums, host provided values, 1246 * and our target's internal memory limits */ 1247 1248 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1249 1250 /* Start with the maximum queue depth allowed by the target */ 1251 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1252 max_read_depth = rtransport->transport.opts.max_queue_depth; 1253 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1254 rtransport->transport.opts.max_queue_depth); 1255 1256 /* Next check the local NIC's hardware limitations */ 1257 SPDK_DEBUGLOG(rdma, 1258 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1259 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1260 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1261 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1262 1263 /* Next check the remote NIC's hardware limitations */ 1264 SPDK_DEBUGLOG(rdma, 1265 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1266 rdma_param->initiator_depth, rdma_param->responder_resources); 1267 if (rdma_param->initiator_depth > 0) { 1268 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1269 } 1270 1271 /* Finally check for the host software requested values, which are 1272 * optional. */ 1273 if (rdma_param->private_data != NULL && 1274 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1275 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1276 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1277 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1278 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1279 } 1280 1281 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1282 max_queue_depth, max_read_depth); 1283 1284 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1285 if (rqpair == NULL) { 1286 SPDK_ERRLOG("Could not allocate new connection.\n"); 1287 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1288 return -1; 1289 } 1290 1291 rqpair->device = port->device; 1292 rqpair->max_queue_depth = max_queue_depth; 1293 rqpair->max_read_depth = max_read_depth; 1294 rqpair->cm_id = event->id; 1295 rqpair->listen_id = event->listen_id; 1296 rqpair->qpair.transport = transport; 1297 STAILQ_INIT(&rqpair->ibv_events); 1298 /* use qid from the private data to determine the qpair type 1299 qid will be set to the appropriate value when the controller is created */ 1300 rqpair->qpair.qid = private_data->qid; 1301 1302 event->id->context = &rqpair->qpair; 1303 1304 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1305 1306 return 0; 1307 } 1308 1309 static inline void 1310 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1311 enum spdk_nvme_data_transfer xfer) 1312 { 1313 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1314 wr->opcode = IBV_WR_RDMA_WRITE; 1315 wr->send_flags = 0; 1316 wr->next = next; 1317 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1318 wr->opcode = IBV_WR_RDMA_READ; 1319 wr->send_flags = IBV_SEND_SIGNALED; 1320 wr->next = NULL; 1321 } else { 1322 assert(0); 1323 } 1324 } 1325 1326 static int 1327 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1328 struct spdk_nvmf_rdma_request *rdma_req, 1329 uint32_t num_sgl_descriptors) 1330 { 1331 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1332 struct spdk_nvmf_rdma_request_data *current_data_wr; 1333 uint32_t i; 1334 1335 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1336 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1337 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1338 return -EINVAL; 1339 } 1340 1341 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1342 return -ENOMEM; 1343 } 1344 1345 current_data_wr = &rdma_req->data; 1346 1347 for (i = 0; i < num_sgl_descriptors; i++) { 1348 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1349 current_data_wr->wr.next = &work_requests[i]->wr; 1350 current_data_wr = work_requests[i]; 1351 current_data_wr->wr.sg_list = current_data_wr->sgl; 1352 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1353 } 1354 1355 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1356 1357 return 0; 1358 } 1359 1360 static inline void 1361 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1362 { 1363 struct ibv_send_wr *wr = &rdma_req->data.wr; 1364 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1365 1366 wr->wr.rdma.rkey = sgl->keyed.key; 1367 wr->wr.rdma.remote_addr = sgl->address; 1368 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1369 } 1370 1371 static inline void 1372 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1373 { 1374 struct ibv_send_wr *wr = &rdma_req->data.wr; 1375 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1376 uint32_t i; 1377 int j; 1378 uint64_t remote_addr_offset = 0; 1379 1380 for (i = 0; i < num_wrs; ++i) { 1381 wr->wr.rdma.rkey = sgl->keyed.key; 1382 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1383 for (j = 0; j < wr->num_sge; ++j) { 1384 remote_addr_offset += wr->sg_list[j].length; 1385 } 1386 wr = wr->next; 1387 } 1388 } 1389 1390 static int 1391 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1392 struct spdk_nvmf_rdma_device *device, 1393 struct spdk_nvmf_rdma_request *rdma_req, 1394 struct ibv_send_wr *wr, 1395 uint32_t total_length) 1396 { 1397 struct spdk_rdma_memory_translation mem_translation; 1398 struct ibv_sge *sg_ele; 1399 struct iovec *iov; 1400 uint32_t lkey, remaining; 1401 int rc; 1402 1403 wr->num_sge = 0; 1404 1405 while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) { 1406 iov = &rdma_req->req.iov[rdma_req->iovpos]; 1407 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1408 if (spdk_unlikely(rc)) { 1409 return rc; 1410 } 1411 1412 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1413 sg_ele = &wr->sg_list[wr->num_sge]; 1414 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1415 1416 sg_ele->lkey = lkey; 1417 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset; 1418 sg_ele->length = remaining; 1419 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1420 sg_ele->length); 1421 rdma_req->offset += sg_ele->length; 1422 total_length -= sg_ele->length; 1423 wr->num_sge++; 1424 1425 if (rdma_req->offset == iov->iov_len) { 1426 rdma_req->offset = 0; 1427 rdma_req->iovpos++; 1428 } 1429 } 1430 1431 if (total_length) { 1432 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1433 return -EINVAL; 1434 } 1435 1436 return 0; 1437 } 1438 1439 static int 1440 nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_poll_group *rgroup, 1441 struct spdk_nvmf_rdma_device *device, 1442 struct spdk_nvmf_rdma_request *rdma_req, 1443 struct ibv_send_wr *wr, 1444 uint32_t total_length, 1445 uint32_t num_extra_wrs) 1446 { 1447 struct spdk_rdma_memory_translation mem_translation; 1448 struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx; 1449 struct ibv_sge *sg_ele; 1450 struct iovec *iov; 1451 struct iovec *rdma_iov; 1452 uint32_t lkey, remaining; 1453 uint32_t remaining_data_block, data_block_size, md_size; 1454 uint32_t sge_len; 1455 int rc; 1456 1457 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1458 1459 if (spdk_likely(!rdma_req->req.stripped_data)) { 1460 rdma_iov = rdma_req->req.iov; 1461 remaining_data_block = data_block_size; 1462 md_size = dif_ctx->md_size; 1463 } else { 1464 rdma_iov = rdma_req->req.stripped_data->iov; 1465 total_length = total_length / dif_ctx->block_size * data_block_size; 1466 remaining_data_block = total_length; 1467 md_size = 0; 1468 } 1469 1470 wr->num_sge = 0; 1471 1472 while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1473 iov = rdma_iov + rdma_req->iovpos; 1474 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1475 if (spdk_unlikely(rc)) { 1476 return rc; 1477 } 1478 1479 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1480 sg_ele = &wr->sg_list[wr->num_sge]; 1481 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1482 1483 while (remaining) { 1484 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1485 if (num_extra_wrs > 0 && wr->next) { 1486 wr = wr->next; 1487 wr->num_sge = 0; 1488 sg_ele = &wr->sg_list[wr->num_sge]; 1489 num_extra_wrs--; 1490 } else { 1491 break; 1492 } 1493 } 1494 sg_ele->lkey = lkey; 1495 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset); 1496 sge_len = spdk_min(remaining, remaining_data_block); 1497 sg_ele->length = sge_len; 1498 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, 1499 sg_ele->addr, sg_ele->length); 1500 remaining -= sge_len; 1501 remaining_data_block -= sge_len; 1502 rdma_req->offset += sge_len; 1503 total_length -= sge_len; 1504 1505 sg_ele++; 1506 wr->num_sge++; 1507 1508 if (remaining_data_block == 0) { 1509 /* skip metadata */ 1510 rdma_req->offset += md_size; 1511 total_length -= md_size; 1512 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1513 remaining -= spdk_min(remaining, md_size); 1514 remaining_data_block = data_block_size; 1515 } 1516 1517 if (remaining == 0) { 1518 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1519 the remaining metadata bits at the beginning of the next buffer */ 1520 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset); 1521 rdma_req->iovpos++; 1522 } 1523 } 1524 } 1525 1526 if (total_length) { 1527 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1528 return -EINVAL; 1529 } 1530 1531 return 0; 1532 } 1533 1534 static inline uint32_t 1535 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1536 { 1537 /* estimate the number of SG entries and WRs needed to process the request */ 1538 uint32_t num_sge = 0; 1539 uint32_t i; 1540 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1541 1542 for (i = 0; i < num_buffers && length > 0; i++) { 1543 uint32_t buffer_len = spdk_min(length, io_unit_size); 1544 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1545 1546 if (num_sge_in_block * block_size > buffer_len) { 1547 ++num_sge_in_block; 1548 } 1549 num_sge += num_sge_in_block; 1550 length -= buffer_len; 1551 } 1552 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1553 } 1554 1555 static int 1556 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1557 struct spdk_nvmf_rdma_device *device, 1558 struct spdk_nvmf_rdma_request *rdma_req) 1559 { 1560 struct spdk_nvmf_rdma_qpair *rqpair; 1561 struct spdk_nvmf_rdma_poll_group *rgroup; 1562 struct spdk_nvmf_request *req = &rdma_req->req; 1563 struct ibv_send_wr *wr = &rdma_req->data.wr; 1564 int rc; 1565 uint32_t num_wrs = 1; 1566 uint32_t length; 1567 1568 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1569 rgroup = rqpair->poller->group; 1570 1571 /* rdma wr specifics */ 1572 nvmf_rdma_setup_request(rdma_req); 1573 1574 length = req->length; 1575 if (spdk_unlikely(req->dif_enabled)) { 1576 req->dif.orig_length = length; 1577 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1578 req->dif.elba_length = length; 1579 } 1580 1581 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1582 length); 1583 if (rc != 0) { 1584 return rc; 1585 } 1586 1587 assert(req->iovcnt <= rqpair->max_send_sge); 1588 1589 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1590 * the stripped_buffers are got for DIF stripping. */ 1591 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1592 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1593 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1594 &rtransport->transport, req->dif.orig_length); 1595 if (rc != 0) { 1596 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1597 } 1598 } 1599 1600 rdma_req->iovpos = 0; 1601 1602 if (spdk_unlikely(req->dif_enabled)) { 1603 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1604 req->dif.dif_ctx.block_size); 1605 if (num_wrs > 1) { 1606 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1607 if (rc != 0) { 1608 goto err_exit; 1609 } 1610 } 1611 1612 rc = nvmf_rdma_fill_wr_sgl_with_dif(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1613 if (spdk_unlikely(rc != 0)) { 1614 goto err_exit; 1615 } 1616 1617 if (num_wrs > 1) { 1618 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1619 } 1620 } else { 1621 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length); 1622 if (spdk_unlikely(rc != 0)) { 1623 goto err_exit; 1624 } 1625 } 1626 1627 /* set the number of outstanding data WRs for this request. */ 1628 rdma_req->num_outstanding_data_wr = num_wrs; 1629 1630 return rc; 1631 1632 err_exit: 1633 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1634 nvmf_rdma_request_free_data(rdma_req, rtransport); 1635 req->iovcnt = 0; 1636 return rc; 1637 } 1638 1639 static int 1640 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1641 struct spdk_nvmf_rdma_device *device, 1642 struct spdk_nvmf_rdma_request *rdma_req) 1643 { 1644 struct spdk_nvmf_rdma_qpair *rqpair; 1645 struct spdk_nvmf_rdma_poll_group *rgroup; 1646 struct ibv_send_wr *current_wr; 1647 struct spdk_nvmf_request *req = &rdma_req->req; 1648 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1649 uint32_t num_sgl_descriptors; 1650 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0; 1651 uint32_t i; 1652 int rc; 1653 1654 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1655 rgroup = rqpair->poller->group; 1656 1657 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1658 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1659 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1660 1661 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1662 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1663 1664 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1665 for (i = 0; i < num_sgl_descriptors; i++) { 1666 if (spdk_likely(!req->dif_enabled)) { 1667 lengths[i] = desc->keyed.length; 1668 } else { 1669 req->dif.orig_length += desc->keyed.length; 1670 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1671 req->dif.elba_length += lengths[i]; 1672 } 1673 total_length += lengths[i]; 1674 desc++; 1675 } 1676 1677 if (total_length > rtransport->transport.opts.max_io_size) { 1678 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n", 1679 total_length, rtransport->transport.opts.max_io_size); 1680 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1681 return -EINVAL; 1682 } 1683 1684 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1685 return -ENOMEM; 1686 } 1687 1688 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length); 1689 if (rc != 0) { 1690 nvmf_rdma_request_free_data(rdma_req, rtransport); 1691 return rc; 1692 } 1693 1694 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1695 * the stripped_buffers are got for DIF stripping. */ 1696 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1697 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1698 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1699 &rtransport->transport, req->dif.orig_length); 1700 if (rc != 0) { 1701 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1702 } 1703 } 1704 1705 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1706 current_wr = &rdma_req->data.wr; 1707 assert(current_wr != NULL); 1708 1709 req->length = 0; 1710 rdma_req->iovpos = 0; 1711 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1712 for (i = 0; i < num_sgl_descriptors; i++) { 1713 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1714 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1715 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1716 rc = -EINVAL; 1717 goto err_exit; 1718 } 1719 1720 if (spdk_likely(!req->dif_enabled)) { 1721 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i]); 1722 } else { 1723 rc = nvmf_rdma_fill_wr_sgl_with_dif(rgroup, device, rdma_req, current_wr, 1724 lengths[i], 0); 1725 } 1726 if (rc != 0) { 1727 rc = -ENOMEM; 1728 goto err_exit; 1729 } 1730 1731 req->length += desc->keyed.length; 1732 current_wr->wr.rdma.rkey = desc->keyed.key; 1733 current_wr->wr.rdma.remote_addr = desc->address; 1734 current_wr = current_wr->next; 1735 desc++; 1736 } 1737 1738 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1739 /* Go back to the last descriptor in the list. */ 1740 desc--; 1741 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1742 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1743 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1744 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1745 } 1746 } 1747 #endif 1748 1749 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1750 1751 return 0; 1752 1753 err_exit: 1754 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1755 nvmf_rdma_request_free_data(rdma_req, rtransport); 1756 return rc; 1757 } 1758 1759 static int 1760 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1761 struct spdk_nvmf_rdma_device *device, 1762 struct spdk_nvmf_rdma_request *rdma_req) 1763 { 1764 struct spdk_nvmf_request *req = &rdma_req->req; 1765 struct spdk_nvme_cpl *rsp; 1766 struct spdk_nvme_sgl_descriptor *sgl; 1767 int rc; 1768 uint32_t length; 1769 1770 rsp = &req->rsp->nvme_cpl; 1771 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1772 1773 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1774 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1775 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1776 1777 length = sgl->keyed.length; 1778 if (length > rtransport->transport.opts.max_io_size) { 1779 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1780 length, rtransport->transport.opts.max_io_size); 1781 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1782 return -1; 1783 } 1784 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1785 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1786 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1787 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1788 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1789 } 1790 } 1791 #endif 1792 1793 /* fill request length and populate iovs */ 1794 req->length = length; 1795 1796 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req); 1797 if (spdk_unlikely(rc < 0)) { 1798 if (rc == -EINVAL) { 1799 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1800 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1801 return -1; 1802 } 1803 /* No available buffers. Queue this request up. */ 1804 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1805 return 0; 1806 } 1807 1808 /* backward compatible */ 1809 req->data = req->iov[0].iov_base; 1810 1811 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1812 req->iovcnt); 1813 1814 return 0; 1815 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1816 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1817 uint64_t offset = sgl->address; 1818 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1819 1820 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1821 offset, sgl->unkeyed.length); 1822 1823 if (offset > max_len) { 1824 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1825 offset, max_len); 1826 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1827 return -1; 1828 } 1829 max_len -= (uint32_t)offset; 1830 1831 if (sgl->unkeyed.length > max_len) { 1832 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1833 sgl->unkeyed.length, max_len); 1834 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1835 return -1; 1836 } 1837 1838 rdma_req->num_outstanding_data_wr = 0; 1839 req->data = rdma_req->recv->buf + offset; 1840 req->data_from_pool = false; 1841 req->length = sgl->unkeyed.length; 1842 1843 req->iov[0].iov_base = req->data; 1844 req->iov[0].iov_len = req->length; 1845 req->iovcnt = 1; 1846 1847 return 0; 1848 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1849 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1850 1851 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1852 if (rc == -ENOMEM) { 1853 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1854 return 0; 1855 } else if (rc == -EINVAL) { 1856 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1857 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1858 return -1; 1859 } 1860 1861 /* backward compatible */ 1862 req->data = req->iov[0].iov_base; 1863 1864 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1865 req->iovcnt); 1866 1867 return 0; 1868 } 1869 1870 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1871 sgl->generic.type, sgl->generic.subtype); 1872 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1873 return -1; 1874 } 1875 1876 static void 1877 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1878 struct spdk_nvmf_rdma_transport *rtransport) 1879 { 1880 struct spdk_nvmf_rdma_qpair *rqpair; 1881 struct spdk_nvmf_rdma_poll_group *rgroup; 1882 1883 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1884 if (rdma_req->req.data_from_pool) { 1885 rgroup = rqpair->poller->group; 1886 1887 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1888 } 1889 if (rdma_req->req.stripped_data) { 1890 nvmf_request_free_stripped_buffers(&rdma_req->req, 1891 &rqpair->poller->group->group, 1892 &rtransport->transport); 1893 } 1894 nvmf_rdma_request_free_data(rdma_req, rtransport); 1895 rdma_req->req.length = 0; 1896 rdma_req->req.iovcnt = 0; 1897 rdma_req->req.data = NULL; 1898 rdma_req->offset = 0; 1899 rdma_req->req.dif_enabled = false; 1900 rdma_req->fused_failed = false; 1901 if (rdma_req->fused_pair) { 1902 /* This req was part of a valid fused pair, but failed before it got to 1903 * READ_TO_EXECUTE state. This means we need to fail the other request 1904 * in the pair, because it is no longer part of a valid pair. If the pair 1905 * already reached READY_TO_EXECUTE state, we need to kick it. 1906 */ 1907 rdma_req->fused_pair->fused_failed = true; 1908 if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 1909 nvmf_rdma_request_process(rtransport, rdma_req->fused_pair); 1910 } 1911 rdma_req->fused_pair = NULL; 1912 } 1913 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1914 rqpair->qd--; 1915 1916 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1917 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1918 } 1919 1920 static void 1921 nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport, 1922 struct spdk_nvmf_rdma_qpair *rqpair, 1923 struct spdk_nvmf_rdma_request *rdma_req) 1924 { 1925 enum spdk_nvme_cmd_fuse last, next; 1926 1927 last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 1928 next = rdma_req->req.cmd->nvme_cmd.fuse; 1929 1930 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 1931 1932 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 1933 return; 1934 } 1935 1936 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 1937 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 1938 /* This is a valid pair of fused commands. Point them at each other 1939 * so they can be submitted consecutively once ready to be executed. 1940 */ 1941 rqpair->fused_first->fused_pair = rdma_req; 1942 rdma_req->fused_pair = rqpair->fused_first; 1943 rqpair->fused_first = NULL; 1944 return; 1945 } else { 1946 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 1947 rqpair->fused_first->fused_failed = true; 1948 1949 /* If the last req is in READY_TO_EXECUTE state, then call 1950 * nvmf_rdma_request_process(), otherwise nothing else will kick it. 1951 */ 1952 if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 1953 nvmf_rdma_request_process(rtransport, rqpair->fused_first); 1954 } 1955 1956 rqpair->fused_first = NULL; 1957 } 1958 } 1959 1960 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 1961 /* Set rqpair->fused_first here so that we know to check that the next request 1962 * is a SECOND (and to fail this one if it isn't). 1963 */ 1964 rqpair->fused_first = rdma_req; 1965 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 1966 /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */ 1967 rdma_req->fused_failed = true; 1968 } 1969 } 1970 1971 bool 1972 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1973 struct spdk_nvmf_rdma_request *rdma_req) 1974 { 1975 struct spdk_nvmf_rdma_qpair *rqpair; 1976 struct spdk_nvmf_rdma_device *device; 1977 struct spdk_nvmf_rdma_poll_group *rgroup; 1978 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1979 int rc; 1980 struct spdk_nvmf_rdma_recv *rdma_recv; 1981 enum spdk_nvmf_rdma_request_state prev_state; 1982 bool progress = false; 1983 int data_posted; 1984 uint32_t num_blocks; 1985 1986 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1987 device = rqpair->device; 1988 rgroup = rqpair->poller->group; 1989 1990 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1991 1992 /* If the queue pair is in an error state, force the request to the completed state 1993 * to release resources. */ 1994 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1995 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1996 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1997 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1998 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1999 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 2000 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2001 } 2002 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2003 } 2004 2005 /* The loop here is to allow for several back-to-back state changes. */ 2006 do { 2007 prev_state = rdma_req->state; 2008 2009 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 2010 2011 switch (rdma_req->state) { 2012 case RDMA_REQUEST_STATE_FREE: 2013 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 2014 * to escape this state. */ 2015 break; 2016 case RDMA_REQUEST_STATE_NEW: 2017 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 2018 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2019 rdma_recv = rdma_req->recv; 2020 2021 /* The first element of the SGL is the NVMe command */ 2022 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 2023 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 2024 2025 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2026 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2027 break; 2028 } 2029 2030 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 2031 rdma_req->req.dif_enabled = true; 2032 } 2033 2034 nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req); 2035 2036 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2037 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 2038 rdma_req->rsp.wr.imm_data = 0; 2039 #endif 2040 2041 /* The next state transition depends on the data transfer needs of this request. */ 2042 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2043 2044 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2045 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2046 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2047 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2048 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 2049 break; 2050 } 2051 2052 /* If no data to transfer, ready to execute. */ 2053 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2054 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2055 break; 2056 } 2057 2058 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2059 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2060 break; 2061 case RDMA_REQUEST_STATE_NEED_BUFFER: 2062 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2063 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2064 2065 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2066 2067 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2068 /* This request needs to wait in line to obtain a buffer */ 2069 break; 2070 } 2071 2072 /* Try to get a data buffer */ 2073 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2074 if (rc < 0) { 2075 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2076 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2077 break; 2078 } 2079 2080 if (!rdma_req->req.data) { 2081 /* No buffers available. */ 2082 rgroup->stat.pending_data_buffer++; 2083 break; 2084 } 2085 2086 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2087 2088 /* If data is transferring from host to controller and the data didn't 2089 * arrive using in capsule data, we need to do a transfer from the host. 2090 */ 2091 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2092 rdma_req->req.data_from_pool) { 2093 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2094 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2095 break; 2096 } 2097 2098 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2099 break; 2100 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2101 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2102 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2103 2104 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2105 /* This request needs to wait in line to perform RDMA */ 2106 break; 2107 } 2108 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2109 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2110 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2111 rqpair->poller->stat.pending_rdma_read++; 2112 break; 2113 } 2114 2115 /* We have already verified that this request is the head of the queue. */ 2116 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2117 2118 rc = request_transfer_in(&rdma_req->req); 2119 if (!rc) { 2120 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2121 } else { 2122 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2123 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2124 } 2125 break; 2126 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2127 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2128 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2129 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2130 * to escape this state. */ 2131 break; 2132 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2133 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2134 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2135 2136 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2137 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2138 /* generate DIF for write operation */ 2139 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2140 assert(num_blocks > 0); 2141 2142 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2143 num_blocks, &rdma_req->req.dif.dif_ctx); 2144 if (rc != 0) { 2145 SPDK_ERRLOG("DIF generation failed\n"); 2146 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2147 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2148 break; 2149 } 2150 } 2151 2152 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2153 /* set extended length before IO operation */ 2154 rdma_req->req.length = rdma_req->req.dif.elba_length; 2155 } 2156 2157 if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 2158 if (rdma_req->fused_failed) { 2159 /* This request failed FUSED semantics. Fail it immediately, without 2160 * even sending it to the target layer. 2161 */ 2162 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2163 rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED; 2164 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2165 break; 2166 } 2167 2168 if (rdma_req->fused_pair == NULL || 2169 rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2170 /* This request is ready to execute, but either we don't know yet if it's 2171 * valid - i.e. this is a FIRST but we haven't received the next 2172 * request yet or the other request of this fused pair isn't ready to 2173 * execute. So break here and this request will get processed later either 2174 * when the other request is ready or we find that this request isn't valid. 2175 */ 2176 break; 2177 } 2178 } 2179 2180 /* If we get to this point, and this request is a fused command, we know that 2181 * it is part of valid sequence (FIRST followed by a SECOND) and that both 2182 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 2183 * request, and the other request of the fused pair, in the correct order. 2184 * Also clear the ->fused_pair pointers on both requests, since after this point 2185 * we no longer need to maintain the relationship between these two requests. 2186 */ 2187 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 2188 assert(rdma_req->fused_pair != NULL); 2189 assert(rdma_req->fused_pair->fused_pair != NULL); 2190 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2191 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2192 rdma_req->fused_pair->fused_pair = NULL; 2193 rdma_req->fused_pair = NULL; 2194 } 2195 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2196 spdk_nvmf_request_exec(&rdma_req->req); 2197 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 2198 assert(rdma_req->fused_pair != NULL); 2199 assert(rdma_req->fused_pair->fused_pair != NULL); 2200 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2201 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2202 rdma_req->fused_pair->fused_pair = NULL; 2203 rdma_req->fused_pair = NULL; 2204 } 2205 break; 2206 case RDMA_REQUEST_STATE_EXECUTING: 2207 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2208 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2209 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2210 * to escape this state. */ 2211 break; 2212 case RDMA_REQUEST_STATE_EXECUTED: 2213 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2214 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2215 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2216 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2217 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2218 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2219 } else { 2220 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2221 } 2222 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2223 /* restore the original length */ 2224 rdma_req->req.length = rdma_req->req.dif.orig_length; 2225 2226 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2227 struct spdk_dif_error error_blk; 2228 2229 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2230 if (!rdma_req->req.stripped_data) { 2231 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2232 &rdma_req->req.dif.dif_ctx, &error_blk); 2233 } else { 2234 rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov, 2235 rdma_req->req.stripped_data->iovcnt, 2236 rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2237 &rdma_req->req.dif.dif_ctx, &error_blk); 2238 } 2239 if (rc) { 2240 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2241 2242 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2243 error_blk.err_offset); 2244 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2245 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2246 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2247 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2248 } 2249 } 2250 } 2251 break; 2252 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2253 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2254 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2255 2256 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2257 /* This request needs to wait in line to perform RDMA */ 2258 break; 2259 } 2260 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2261 rqpair->max_send_depth) { 2262 /* We can only have so many WRs outstanding. we have to wait until some finish. 2263 * +1 since each request has an additional wr in the resp. */ 2264 rqpair->poller->stat.pending_rdma_write++; 2265 break; 2266 } 2267 2268 /* We have already verified that this request is the head of the queue. */ 2269 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2270 2271 /* The data transfer will be kicked off from 2272 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2273 */ 2274 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2275 break; 2276 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2277 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2278 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2279 rc = request_transfer_out(&rdma_req->req, &data_posted); 2280 assert(rc == 0); /* No good way to handle this currently */ 2281 if (rc) { 2282 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2283 } else { 2284 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2285 RDMA_REQUEST_STATE_COMPLETING; 2286 } 2287 break; 2288 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2289 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2290 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2291 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2292 * to escape this state. */ 2293 break; 2294 case RDMA_REQUEST_STATE_COMPLETING: 2295 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2296 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2297 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2298 * to escape this state. */ 2299 break; 2300 case RDMA_REQUEST_STATE_COMPLETED: 2301 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2302 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2303 2304 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2305 _nvmf_rdma_request_free(rdma_req, rtransport); 2306 break; 2307 case RDMA_REQUEST_NUM_STATES: 2308 default: 2309 assert(0); 2310 break; 2311 } 2312 2313 if (rdma_req->state != prev_state) { 2314 progress = true; 2315 } 2316 } while (rdma_req->state != prev_state); 2317 2318 return progress; 2319 } 2320 2321 /* Public API callbacks begin here */ 2322 2323 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2324 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2325 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2326 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2327 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2328 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2329 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2330 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2331 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2332 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2333 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2334 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2335 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2336 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2337 2338 static void 2339 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2340 { 2341 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2342 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2343 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2344 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2345 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2346 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2347 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2348 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2349 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2350 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2351 opts->transport_specific = NULL; 2352 } 2353 2354 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2355 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2356 2357 static inline bool 2358 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2359 { 2360 return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD || 2361 device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW; 2362 } 2363 2364 static int nvmf_rdma_accept(void *ctx); 2365 static int 2366 create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context, 2367 struct spdk_nvmf_rdma_device **new_device) 2368 { 2369 struct spdk_nvmf_rdma_device *device; 2370 int flag = 0; 2371 int rc = 0; 2372 2373 device = calloc(1, sizeof(*device)); 2374 if (!device) { 2375 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2376 return -ENOMEM; 2377 } 2378 device->context = context; 2379 rc = ibv_query_device(device->context, &device->attr); 2380 if (rc < 0) { 2381 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2382 free(device); 2383 return rc; 2384 } 2385 2386 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2387 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2388 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2389 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2390 } 2391 2392 /** 2393 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2394 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2395 * but incorrectly reports that it does. There are changes making their way 2396 * through the kernel now that will enable this feature. When they are merged, 2397 * we can conditionally enable this feature. 2398 * 2399 * TODO: enable this for versions of the kernel rxe driver that support it. 2400 */ 2401 if (nvmf_rdma_is_rxe_device(device)) { 2402 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2403 } 2404 #endif 2405 2406 /* set up device context async ev fd as NON_BLOCKING */ 2407 flag = fcntl(device->context->async_fd, F_GETFL); 2408 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2409 if (rc < 0) { 2410 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2411 free(device); 2412 return rc; 2413 } 2414 2415 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2416 SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA trasport\n", device); 2417 2418 if (g_nvmf_hooks.get_ibv_pd) { 2419 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2420 } else { 2421 device->pd = ibv_alloc_pd(device->context); 2422 } 2423 2424 if (!device->pd) { 2425 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2426 return -ENOMEM; 2427 } 2428 2429 assert(device->map == NULL); 2430 2431 device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks, SPDK_RDMA_MEMORY_MAP_ROLE_TARGET); 2432 if (!device->map) { 2433 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2434 return -ENOMEM; 2435 } 2436 2437 assert(device->map != NULL); 2438 assert(device->pd != NULL); 2439 2440 if (new_device) { 2441 *new_device = device; 2442 } 2443 return 0; 2444 } 2445 2446 static void 2447 free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2448 { 2449 if (rtransport->poll_fds) { 2450 free(rtransport->poll_fds); 2451 rtransport->poll_fds = NULL; 2452 } 2453 rtransport->npoll_fds = 0; 2454 } 2455 2456 static int 2457 generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2458 { 2459 /* Set up poll descriptor array to monitor events from RDMA and IB 2460 * in a single poll syscall 2461 */ 2462 int device_count = 0; 2463 int i = 0; 2464 struct spdk_nvmf_rdma_device *device, *tmp; 2465 2466 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2467 device_count++; 2468 } 2469 2470 rtransport->npoll_fds = device_count + 1; 2471 2472 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2473 if (rtransport->poll_fds == NULL) { 2474 SPDK_ERRLOG("poll_fds allocation failed\n"); 2475 return -ENOMEM; 2476 } 2477 2478 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2479 rtransport->poll_fds[i++].events = POLLIN; 2480 2481 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2482 rtransport->poll_fds[i].fd = device->context->async_fd; 2483 rtransport->poll_fds[i++].events = POLLIN; 2484 } 2485 2486 return 0; 2487 } 2488 2489 static struct spdk_nvmf_transport * 2490 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2491 { 2492 int rc; 2493 struct spdk_nvmf_rdma_transport *rtransport; 2494 struct spdk_nvmf_rdma_device *device; 2495 struct ibv_context **contexts; 2496 uint32_t i; 2497 int flag; 2498 uint32_t sge_count; 2499 uint32_t min_shared_buffers; 2500 uint32_t min_in_capsule_data_size; 2501 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2502 2503 rtransport = calloc(1, sizeof(*rtransport)); 2504 if (!rtransport) { 2505 return NULL; 2506 } 2507 2508 TAILQ_INIT(&rtransport->devices); 2509 TAILQ_INIT(&rtransport->ports); 2510 TAILQ_INIT(&rtransport->poll_groups); 2511 2512 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2513 rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2514 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2515 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2516 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2517 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2518 if (opts->transport_specific != NULL && 2519 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2520 SPDK_COUNTOF(rdma_transport_opts_decoder), 2521 &rtransport->rdma_opts)) { 2522 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2523 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2524 return NULL; 2525 } 2526 2527 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2528 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2529 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2530 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2531 " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d," 2532 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2533 opts->max_queue_depth, 2534 opts->max_io_size, 2535 opts->max_qpairs_per_ctrlr - 1, 2536 opts->io_unit_size, 2537 opts->in_capsule_data_size, 2538 opts->max_aq_depth, 2539 opts->num_shared_buffers, 2540 rtransport->rdma_opts.num_cqe, 2541 rtransport->rdma_opts.max_srq_depth, 2542 rtransport->rdma_opts.no_srq, 2543 rtransport->rdma_opts.acceptor_backlog, 2544 rtransport->rdma_opts.no_wr_batching, 2545 opts->abort_timeout_sec); 2546 2547 /* I/O unit size cannot be larger than max I/O size */ 2548 if (opts->io_unit_size > opts->max_io_size) { 2549 opts->io_unit_size = opts->max_io_size; 2550 } 2551 2552 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2553 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2554 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2555 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2556 } 2557 2558 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2559 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2560 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2561 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2562 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2563 return NULL; 2564 } 2565 2566 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2567 if (min_shared_buffers > opts->num_shared_buffers) { 2568 SPDK_ERRLOG("There are not enough buffers to satisfy" 2569 "per-poll group caches for each thread. (%" PRIu32 ")" 2570 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2571 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2572 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2573 return NULL; 2574 } 2575 2576 sge_count = opts->max_io_size / opts->io_unit_size; 2577 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2578 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2579 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2580 return NULL; 2581 } 2582 2583 min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES; 2584 if (opts->in_capsule_data_size < min_in_capsule_data_size) { 2585 SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n", 2586 min_in_capsule_data_size); 2587 opts->in_capsule_data_size = min_in_capsule_data_size; 2588 } 2589 2590 rtransport->event_channel = rdma_create_event_channel(); 2591 if (rtransport->event_channel == NULL) { 2592 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2593 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2594 return NULL; 2595 } 2596 2597 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2598 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2599 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2600 rtransport->event_channel->fd, spdk_strerror(errno)); 2601 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2602 return NULL; 2603 } 2604 2605 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2606 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2607 sizeof(struct spdk_nvmf_rdma_request_data), 2608 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2609 SPDK_ENV_SOCKET_ID_ANY); 2610 if (!rtransport->data_wr_pool) { 2611 if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) { 2612 SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n"); 2613 SPDK_ERRLOG("Probably running in multiprocess environment, which is " 2614 "unsupported by the nvmf library\n"); 2615 } else { 2616 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2617 } 2618 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2619 return NULL; 2620 } 2621 2622 contexts = rdma_get_devices(NULL); 2623 if (contexts == NULL) { 2624 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2625 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2626 return NULL; 2627 } 2628 2629 i = 0; 2630 rc = 0; 2631 while (contexts[i] != NULL) { 2632 rc = create_ib_device(rtransport, contexts[i], &device); 2633 if (rc < 0) { 2634 break; 2635 } 2636 i++; 2637 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2638 } 2639 rdma_free_devices(contexts); 2640 2641 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2642 /* divide and round up. */ 2643 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2644 2645 /* round up to the nearest 4k. */ 2646 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2647 2648 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2649 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2650 opts->io_unit_size); 2651 } 2652 2653 if (rc < 0) { 2654 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2655 return NULL; 2656 } 2657 2658 rc = generate_poll_fds(rtransport); 2659 if (rc < 0) { 2660 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2661 return NULL; 2662 } 2663 2664 rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport, 2665 opts->acceptor_poll_rate); 2666 if (!rtransport->accept_poller) { 2667 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2668 return NULL; 2669 } 2670 2671 return &rtransport->transport; 2672 } 2673 2674 static void 2675 destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2676 struct spdk_nvmf_rdma_device *device) 2677 { 2678 TAILQ_REMOVE(&rtransport->devices, device, link); 2679 spdk_rdma_free_mem_map(&device->map); 2680 if (device->pd) { 2681 if (!g_nvmf_hooks.get_ibv_pd) { 2682 ibv_dealloc_pd(device->pd); 2683 } 2684 } 2685 free(device); 2686 } 2687 2688 static void 2689 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2690 { 2691 struct spdk_nvmf_rdma_transport *rtransport; 2692 assert(w != NULL); 2693 2694 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2695 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2696 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2697 if (rtransport->rdma_opts.no_srq == true) { 2698 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2699 } 2700 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2701 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2702 } 2703 2704 static int 2705 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2706 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2707 { 2708 struct spdk_nvmf_rdma_transport *rtransport; 2709 struct spdk_nvmf_rdma_port *port, *port_tmp; 2710 struct spdk_nvmf_rdma_device *device, *device_tmp; 2711 2712 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2713 2714 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2715 TAILQ_REMOVE(&rtransport->ports, port, link); 2716 rdma_destroy_id(port->id); 2717 free(port); 2718 } 2719 2720 free_poll_fds(rtransport); 2721 2722 if (rtransport->event_channel != NULL) { 2723 rdma_destroy_event_channel(rtransport->event_channel); 2724 } 2725 2726 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2727 destroy_ib_device(rtransport, device); 2728 } 2729 2730 if (rtransport->data_wr_pool != NULL) { 2731 if (spdk_mempool_count(rtransport->data_wr_pool) != 2732 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2733 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2734 spdk_mempool_count(rtransport->data_wr_pool), 2735 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2736 } 2737 } 2738 2739 spdk_mempool_free(rtransport->data_wr_pool); 2740 2741 spdk_poller_unregister(&rtransport->accept_poller); 2742 free(rtransport); 2743 2744 if (cb_fn) { 2745 cb_fn(cb_arg); 2746 } 2747 return 0; 2748 } 2749 2750 static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2751 struct spdk_nvme_transport_id *trid, 2752 bool peer); 2753 2754 static int 2755 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2756 struct spdk_nvmf_listen_opts *listen_opts) 2757 { 2758 struct spdk_nvmf_rdma_transport *rtransport; 2759 struct spdk_nvmf_rdma_device *device; 2760 struct spdk_nvmf_rdma_port *port; 2761 struct addrinfo *res; 2762 struct addrinfo hints; 2763 int family; 2764 int rc; 2765 2766 if (!strlen(trid->trsvcid)) { 2767 SPDK_ERRLOG("Service id is required\n"); 2768 return -EINVAL; 2769 } 2770 2771 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2772 assert(rtransport->event_channel != NULL); 2773 2774 port = calloc(1, sizeof(*port)); 2775 if (!port) { 2776 SPDK_ERRLOG("Port allocation failed\n"); 2777 return -ENOMEM; 2778 } 2779 2780 port->trid = trid; 2781 2782 switch (trid->adrfam) { 2783 case SPDK_NVMF_ADRFAM_IPV4: 2784 family = AF_INET; 2785 break; 2786 case SPDK_NVMF_ADRFAM_IPV6: 2787 family = AF_INET6; 2788 break; 2789 default: 2790 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2791 free(port); 2792 return -EINVAL; 2793 } 2794 2795 memset(&hints, 0, sizeof(hints)); 2796 hints.ai_family = family; 2797 hints.ai_flags = AI_NUMERICSERV; 2798 hints.ai_socktype = SOCK_STREAM; 2799 hints.ai_protocol = 0; 2800 2801 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2802 if (rc) { 2803 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2804 free(port); 2805 return -EINVAL; 2806 } 2807 2808 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2809 if (rc < 0) { 2810 SPDK_ERRLOG("rdma_create_id() failed\n"); 2811 freeaddrinfo(res); 2812 free(port); 2813 return rc; 2814 } 2815 2816 rc = rdma_bind_addr(port->id, res->ai_addr); 2817 freeaddrinfo(res); 2818 2819 if (rc < 0) { 2820 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2821 rdma_destroy_id(port->id); 2822 free(port); 2823 return rc; 2824 } 2825 2826 if (!port->id->verbs) { 2827 SPDK_ERRLOG("ibv_context is null\n"); 2828 rdma_destroy_id(port->id); 2829 free(port); 2830 return -1; 2831 } 2832 2833 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2834 if (rc < 0) { 2835 SPDK_ERRLOG("rdma_listen() failed\n"); 2836 rdma_destroy_id(port->id); 2837 free(port); 2838 return rc; 2839 } 2840 2841 TAILQ_FOREACH(device, &rtransport->devices, link) { 2842 if (device->context == port->id->verbs) { 2843 port->device = device; 2844 break; 2845 } 2846 } 2847 if (!port->device) { 2848 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2849 port->id->verbs); 2850 rdma_destroy_id(port->id); 2851 free(port); 2852 return -EINVAL; 2853 } 2854 2855 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2856 trid->traddr, trid->trsvcid); 2857 2858 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2859 return 0; 2860 } 2861 2862 static void 2863 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2864 const struct spdk_nvme_transport_id *trid) 2865 { 2866 struct spdk_nvmf_rdma_transport *rtransport; 2867 struct spdk_nvmf_rdma_port *port, *tmp; 2868 2869 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2870 2871 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2872 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2873 TAILQ_REMOVE(&rtransport->ports, port, link); 2874 rdma_destroy_id(port->id); 2875 free(port); 2876 break; 2877 } 2878 } 2879 } 2880 2881 static void 2882 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2883 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2884 { 2885 struct spdk_nvmf_request *req, *tmp; 2886 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2887 struct spdk_nvmf_rdma_resources *resources; 2888 2889 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2890 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2891 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2892 break; 2893 } 2894 } 2895 2896 /* Then RDMA writes since reads have stronger restrictions than writes */ 2897 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2898 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2899 break; 2900 } 2901 } 2902 2903 /* Then we handle request waiting on memory buffers. */ 2904 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2905 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2906 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2907 break; 2908 } 2909 } 2910 2911 resources = rqpair->resources; 2912 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2913 rdma_req = STAILQ_FIRST(&resources->free_queue); 2914 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2915 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2916 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2917 2918 if (rqpair->srq != NULL) { 2919 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2920 rdma_req->recv->qpair->qd++; 2921 } else { 2922 rqpair->qd++; 2923 } 2924 2925 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2926 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2927 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2928 break; 2929 } 2930 } 2931 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2932 rqpair->poller->stat.pending_free_request++; 2933 } 2934 } 2935 2936 static inline bool 2937 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 2938 { 2939 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 2940 return nvmf_rdma_is_rxe_device(device) || 2941 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 2942 } 2943 2944 static void 2945 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2946 { 2947 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2948 struct spdk_nvmf_rdma_transport, transport); 2949 2950 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2951 2952 /* nvmf_rdma_close_qpair is not called */ 2953 if (!rqpair->to_close) { 2954 return; 2955 } 2956 2957 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2958 if (rqpair->current_send_depth != 0) { 2959 return; 2960 } 2961 2962 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2963 return; 2964 } 2965 2966 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 2967 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 2968 return; 2969 } 2970 2971 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2972 2973 nvmf_rdma_qpair_destroy(rqpair); 2974 } 2975 2976 static int 2977 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2978 { 2979 struct spdk_nvmf_qpair *qpair; 2980 struct spdk_nvmf_rdma_qpair *rqpair; 2981 2982 if (evt->id == NULL) { 2983 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2984 return -1; 2985 } 2986 2987 qpair = evt->id->context; 2988 if (qpair == NULL) { 2989 SPDK_ERRLOG("disconnect request: no active connection\n"); 2990 return -1; 2991 } 2992 2993 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2994 2995 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair); 2996 2997 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2998 2999 return 0; 3000 } 3001 3002 #ifdef DEBUG 3003 static const char *CM_EVENT_STR[] = { 3004 "RDMA_CM_EVENT_ADDR_RESOLVED", 3005 "RDMA_CM_EVENT_ADDR_ERROR", 3006 "RDMA_CM_EVENT_ROUTE_RESOLVED", 3007 "RDMA_CM_EVENT_ROUTE_ERROR", 3008 "RDMA_CM_EVENT_CONNECT_REQUEST", 3009 "RDMA_CM_EVENT_CONNECT_RESPONSE", 3010 "RDMA_CM_EVENT_CONNECT_ERROR", 3011 "RDMA_CM_EVENT_UNREACHABLE", 3012 "RDMA_CM_EVENT_REJECTED", 3013 "RDMA_CM_EVENT_ESTABLISHED", 3014 "RDMA_CM_EVENT_DISCONNECTED", 3015 "RDMA_CM_EVENT_DEVICE_REMOVAL", 3016 "RDMA_CM_EVENT_MULTICAST_JOIN", 3017 "RDMA_CM_EVENT_MULTICAST_ERROR", 3018 "RDMA_CM_EVENT_ADDR_CHANGE", 3019 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 3020 }; 3021 #endif /* DEBUG */ 3022 3023 static void 3024 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 3025 struct spdk_nvmf_rdma_port *port) 3026 { 3027 struct spdk_nvmf_rdma_poll_group *rgroup; 3028 struct spdk_nvmf_rdma_poller *rpoller; 3029 struct spdk_nvmf_rdma_qpair *rqpair; 3030 3031 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3032 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3033 RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) { 3034 if (rqpair->listen_id == port->id) { 3035 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3036 } 3037 } 3038 } 3039 } 3040 } 3041 3042 static bool 3043 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 3044 struct rdma_cm_event *event) 3045 { 3046 const struct spdk_nvme_transport_id *trid; 3047 struct spdk_nvmf_rdma_port *port; 3048 struct spdk_nvmf_rdma_transport *rtransport; 3049 bool event_acked = false; 3050 3051 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3052 TAILQ_FOREACH(port, &rtransport->ports, link) { 3053 if (port->id == event->id) { 3054 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 3055 rdma_ack_cm_event(event); 3056 event_acked = true; 3057 trid = port->trid; 3058 break; 3059 } 3060 } 3061 3062 if (event_acked) { 3063 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 3064 3065 nvmf_rdma_stop_listen(transport, trid); 3066 nvmf_rdma_listen(transport, trid, NULL); 3067 } 3068 3069 return event_acked; 3070 } 3071 3072 static void 3073 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 3074 struct rdma_cm_event *event) 3075 { 3076 struct spdk_nvmf_rdma_port *port; 3077 struct spdk_nvmf_rdma_transport *rtransport; 3078 3079 port = event->id->context; 3080 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3081 3082 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 3083 3084 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 3085 3086 rdma_ack_cm_event(event); 3087 3088 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 3089 ; 3090 } 3091 } 3092 3093 static void 3094 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 3095 { 3096 struct spdk_nvmf_rdma_transport *rtransport; 3097 struct rdma_cm_event *event; 3098 int rc; 3099 bool event_acked; 3100 3101 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3102 3103 if (rtransport->event_channel == NULL) { 3104 return; 3105 } 3106 3107 while (1) { 3108 event_acked = false; 3109 rc = rdma_get_cm_event(rtransport->event_channel, &event); 3110 if (rc) { 3111 if (errno != EAGAIN && errno != EWOULDBLOCK) { 3112 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 3113 } 3114 break; 3115 } 3116 3117 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 3118 3119 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 3120 3121 switch (event->event) { 3122 case RDMA_CM_EVENT_ADDR_RESOLVED: 3123 case RDMA_CM_EVENT_ADDR_ERROR: 3124 case RDMA_CM_EVENT_ROUTE_RESOLVED: 3125 case RDMA_CM_EVENT_ROUTE_ERROR: 3126 /* No action required. The target never attempts to resolve routes. */ 3127 break; 3128 case RDMA_CM_EVENT_CONNECT_REQUEST: 3129 rc = nvmf_rdma_connect(transport, event); 3130 if (rc < 0) { 3131 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 3132 break; 3133 } 3134 break; 3135 case RDMA_CM_EVENT_CONNECT_RESPONSE: 3136 /* The target never initiates a new connection. So this will not occur. */ 3137 break; 3138 case RDMA_CM_EVENT_CONNECT_ERROR: 3139 /* Can this happen? The docs say it can, but not sure what causes it. */ 3140 break; 3141 case RDMA_CM_EVENT_UNREACHABLE: 3142 case RDMA_CM_EVENT_REJECTED: 3143 /* These only occur on the client side. */ 3144 break; 3145 case RDMA_CM_EVENT_ESTABLISHED: 3146 /* TODO: Should we be waiting for this event anywhere? */ 3147 break; 3148 case RDMA_CM_EVENT_DISCONNECTED: 3149 rc = nvmf_rdma_disconnect(event); 3150 if (rc < 0) { 3151 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3152 break; 3153 } 3154 break; 3155 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3156 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3157 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3158 * Once these events are sent to SPDK, we should release all IB resources and 3159 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3160 * ibv_destroy* functions to release user space memory allocated by IB. All kernel 3161 * resources are already cleaned. */ 3162 if (event->id->qp) { 3163 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3164 * corresponding qpair. Otherwise the event refers to a listening device */ 3165 rc = nvmf_rdma_disconnect(event); 3166 if (rc < 0) { 3167 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3168 break; 3169 } 3170 } else { 3171 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3172 event_acked = true; 3173 } 3174 break; 3175 case RDMA_CM_EVENT_MULTICAST_JOIN: 3176 case RDMA_CM_EVENT_MULTICAST_ERROR: 3177 /* Multicast is not used */ 3178 break; 3179 case RDMA_CM_EVENT_ADDR_CHANGE: 3180 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3181 break; 3182 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3183 /* For now, do nothing. The target never re-uses queue pairs. */ 3184 break; 3185 default: 3186 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3187 break; 3188 } 3189 if (!event_acked) { 3190 rdma_ack_cm_event(event); 3191 } 3192 } 3193 } 3194 3195 static void 3196 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3197 { 3198 rqpair->last_wqe_reached = true; 3199 nvmf_rdma_destroy_drained_qpair(rqpair); 3200 } 3201 3202 static void 3203 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3204 { 3205 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3206 3207 if (event_ctx->rqpair) { 3208 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3209 if (event_ctx->cb_fn) { 3210 event_ctx->cb_fn(event_ctx->rqpair); 3211 } 3212 } 3213 free(event_ctx); 3214 } 3215 3216 static int 3217 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3218 spdk_nvmf_rdma_qpair_ibv_event fn) 3219 { 3220 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3221 struct spdk_thread *thr = NULL; 3222 int rc; 3223 3224 if (rqpair->qpair.group) { 3225 thr = rqpair->qpair.group->thread; 3226 } else if (rqpair->destruct_channel) { 3227 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3228 } 3229 3230 if (!thr) { 3231 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3232 return -EINVAL; 3233 } 3234 3235 ctx = calloc(1, sizeof(*ctx)); 3236 if (!ctx) { 3237 return -ENOMEM; 3238 } 3239 3240 ctx->rqpair = rqpair; 3241 ctx->cb_fn = fn; 3242 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3243 3244 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3245 if (rc) { 3246 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3247 free(ctx); 3248 } 3249 3250 return rc; 3251 } 3252 3253 static int 3254 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3255 { 3256 int rc; 3257 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3258 struct ibv_async_event event; 3259 3260 rc = ibv_get_async_event(device->context, &event); 3261 3262 if (rc) { 3263 /* In non-blocking mode -1 means there are no events available */ 3264 return rc; 3265 } 3266 3267 switch (event.event_type) { 3268 case IBV_EVENT_QP_FATAL: 3269 case IBV_EVENT_QP_LAST_WQE_REACHED: 3270 case IBV_EVENT_SQ_DRAINED: 3271 case IBV_EVENT_QP_REQ_ERR: 3272 case IBV_EVENT_QP_ACCESS_ERR: 3273 case IBV_EVENT_COMM_EST: 3274 case IBV_EVENT_PATH_MIG: 3275 case IBV_EVENT_PATH_MIG_ERR: 3276 rqpair = event.element.qp->qp_context; 3277 if (!rqpair) { 3278 /* Any QP event for NVMe-RDMA initiator may be returned. */ 3279 SPDK_NOTICELOG("Async QP event for unknown QP: %s\n", 3280 ibv_event_type_str(event.event_type)); 3281 break; 3282 } 3283 3284 switch (event.event_type) { 3285 case IBV_EVENT_QP_FATAL: 3286 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3287 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3288 (uintptr_t)rqpair, event.event_type); 3289 nvmf_rdma_update_ibv_state(rqpair); 3290 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3291 break; 3292 case IBV_EVENT_QP_LAST_WQE_REACHED: 3293 /* This event only occurs for shared receive queues. */ 3294 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3295 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3296 if (rc) { 3297 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3298 rqpair->last_wqe_reached = true; 3299 } 3300 break; 3301 case IBV_EVENT_SQ_DRAINED: 3302 /* This event occurs frequently in both error and non-error states. 3303 * Check if the qpair is in an error state before sending a message. */ 3304 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3305 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3306 (uintptr_t)rqpair, event.event_type); 3307 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3308 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3309 } 3310 break; 3311 case IBV_EVENT_QP_REQ_ERR: 3312 case IBV_EVENT_QP_ACCESS_ERR: 3313 case IBV_EVENT_COMM_EST: 3314 case IBV_EVENT_PATH_MIG: 3315 case IBV_EVENT_PATH_MIG_ERR: 3316 SPDK_NOTICELOG("Async QP event: %s\n", 3317 ibv_event_type_str(event.event_type)); 3318 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3319 (uintptr_t)rqpair, event.event_type); 3320 nvmf_rdma_update_ibv_state(rqpair); 3321 break; 3322 default: 3323 break; 3324 } 3325 break; 3326 case IBV_EVENT_CQ_ERR: 3327 case IBV_EVENT_DEVICE_FATAL: 3328 case IBV_EVENT_PORT_ACTIVE: 3329 case IBV_EVENT_PORT_ERR: 3330 case IBV_EVENT_LID_CHANGE: 3331 case IBV_EVENT_PKEY_CHANGE: 3332 case IBV_EVENT_SM_CHANGE: 3333 case IBV_EVENT_SRQ_ERR: 3334 case IBV_EVENT_SRQ_LIMIT_REACHED: 3335 case IBV_EVENT_CLIENT_REREGISTER: 3336 case IBV_EVENT_GID_CHANGE: 3337 default: 3338 SPDK_NOTICELOG("Async event: %s\n", 3339 ibv_event_type_str(event.event_type)); 3340 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3341 break; 3342 } 3343 ibv_ack_async_event(&event); 3344 3345 return 0; 3346 } 3347 3348 static void 3349 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3350 { 3351 int rc = 0; 3352 uint32_t i = 0; 3353 3354 for (i = 0; i < max_events; i++) { 3355 rc = nvmf_process_ib_event(device); 3356 if (rc) { 3357 break; 3358 } 3359 } 3360 3361 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3362 } 3363 3364 static int 3365 nvmf_rdma_accept(void *ctx) 3366 { 3367 int nfds, i = 0; 3368 struct spdk_nvmf_transport *transport = ctx; 3369 struct spdk_nvmf_rdma_transport *rtransport; 3370 struct spdk_nvmf_rdma_device *device, *tmp; 3371 uint32_t count; 3372 3373 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3374 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3375 3376 if (nfds <= 0) { 3377 return SPDK_POLLER_IDLE; 3378 } 3379 3380 /* The first poll descriptor is RDMA CM event */ 3381 if (rtransport->poll_fds[i++].revents & POLLIN) { 3382 nvmf_process_cm_event(transport); 3383 nfds--; 3384 } 3385 3386 if (nfds == 0) { 3387 return SPDK_POLLER_BUSY; 3388 } 3389 3390 /* Second and subsequent poll descriptors are IB async events */ 3391 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3392 if (rtransport->poll_fds[i++].revents & POLLIN) { 3393 nvmf_process_ib_events(device, 32); 3394 nfds--; 3395 } 3396 } 3397 /* check all flagged fd's have been served */ 3398 assert(nfds == 0); 3399 3400 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3401 } 3402 3403 static void 3404 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3405 struct spdk_nvmf_ctrlr_data *cdata) 3406 { 3407 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3408 3409 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3410 since in-capsule data only works with NVME drives that support SGL memory layout */ 3411 if (transport->opts.dif_insert_or_strip) { 3412 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3413 } 3414 3415 if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) { 3416 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule" 3417 " data is greater than 4KiB.\n"); 3418 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux " 3419 "kernel between versions 5.4 and 5.12 data corruption may occur for " 3420 "writes that are not a multiple of 4KiB in size.\n"); 3421 } 3422 } 3423 3424 static void 3425 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3426 struct spdk_nvme_transport_id *trid, 3427 struct spdk_nvmf_discovery_log_page_entry *entry) 3428 { 3429 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3430 entry->adrfam = trid->adrfam; 3431 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3432 3433 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3434 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3435 3436 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3437 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3438 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3439 } 3440 3441 static int 3442 nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport, 3443 struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device, 3444 struct spdk_nvmf_rdma_poller **out_poller) 3445 { 3446 struct spdk_nvmf_rdma_poller *poller; 3447 struct spdk_rdma_srq_init_attr srq_init_attr; 3448 struct spdk_nvmf_rdma_resource_opts opts; 3449 int num_cqe; 3450 3451 poller = calloc(1, sizeof(*poller)); 3452 if (!poller) { 3453 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3454 return -1; 3455 } 3456 3457 poller->device = device; 3458 poller->group = rgroup; 3459 *out_poller = poller; 3460 3461 RB_INIT(&poller->qpairs); 3462 STAILQ_INIT(&poller->qpairs_pending_send); 3463 STAILQ_INIT(&poller->qpairs_pending_recv); 3464 3465 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3466 SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup); 3467 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3468 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) { 3469 SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n", 3470 rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr); 3471 } 3472 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr); 3473 3474 device->num_srq++; 3475 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 3476 srq_init_attr.pd = device->pd; 3477 srq_init_attr.stats = &poller->stat.qp_stats.recv; 3478 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 3479 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3480 poller->srq = spdk_rdma_srq_create(&srq_init_attr); 3481 if (!poller->srq) { 3482 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3483 return -1; 3484 } 3485 3486 opts.qp = poller->srq; 3487 opts.map = device->map; 3488 opts.qpair = NULL; 3489 opts.shared = true; 3490 opts.max_queue_depth = poller->max_srq_depth; 3491 opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size; 3492 3493 poller->resources = nvmf_rdma_resources_create(&opts); 3494 if (!poller->resources) { 3495 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3496 return -1; 3497 } 3498 } 3499 3500 /* 3501 * When using an srq, we can limit the completion queue at startup. 3502 * The following formula represents the calculation: 3503 * num_cqe = num_recv + num_data_wr + num_send_wr. 3504 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3505 */ 3506 if (poller->srq) { 3507 num_cqe = poller->max_srq_depth * 3; 3508 } else { 3509 num_cqe = rtransport->rdma_opts.num_cqe; 3510 } 3511 3512 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3513 if (!poller->cq) { 3514 SPDK_ERRLOG("Unable to create completion queue\n"); 3515 return -1; 3516 } 3517 poller->num_cqe = num_cqe; 3518 return 0; 3519 } 3520 3521 static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3522 3523 static struct spdk_nvmf_transport_poll_group * 3524 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport, 3525 struct spdk_nvmf_poll_group *group) 3526 { 3527 struct spdk_nvmf_rdma_transport *rtransport; 3528 struct spdk_nvmf_rdma_poll_group *rgroup; 3529 struct spdk_nvmf_rdma_poller *poller; 3530 struct spdk_nvmf_rdma_device *device; 3531 int rc; 3532 3533 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3534 3535 rgroup = calloc(1, sizeof(*rgroup)); 3536 if (!rgroup) { 3537 return NULL; 3538 } 3539 3540 TAILQ_INIT(&rgroup->pollers); 3541 3542 TAILQ_FOREACH(device, &rtransport->devices, link) { 3543 rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller); 3544 if (rc < 0) { 3545 nvmf_rdma_poll_group_destroy(&rgroup->group); 3546 return NULL; 3547 } 3548 } 3549 3550 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3551 if (rtransport->conn_sched.next_admin_pg == NULL) { 3552 rtransport->conn_sched.next_admin_pg = rgroup; 3553 rtransport->conn_sched.next_io_pg = rgroup; 3554 } 3555 3556 return &rgroup->group; 3557 } 3558 3559 static uint32_t 3560 nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg) 3561 { 3562 uint32_t count; 3563 3564 /* Just assume that unassociated qpairs will eventually be io 3565 * qpairs. This is close enough for the use cases for this 3566 * function. 3567 */ 3568 pthread_mutex_lock(&pg->mutex); 3569 count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs; 3570 pthread_mutex_unlock(&pg->mutex); 3571 3572 return count; 3573 } 3574 3575 static struct spdk_nvmf_transport_poll_group * 3576 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3577 { 3578 struct spdk_nvmf_rdma_transport *rtransport; 3579 struct spdk_nvmf_rdma_poll_group **pg; 3580 struct spdk_nvmf_transport_poll_group *result; 3581 uint32_t count; 3582 3583 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3584 3585 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3586 return NULL; 3587 } 3588 3589 if (qpair->qid == 0) { 3590 pg = &rtransport->conn_sched.next_admin_pg; 3591 } else { 3592 struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current; 3593 uint32_t min_value; 3594 3595 pg = &rtransport->conn_sched.next_io_pg; 3596 pg_min = *pg; 3597 pg_start = *pg; 3598 pg_current = *pg; 3599 min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 3600 3601 while ((count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group)) > 0) { 3602 pg_current = TAILQ_NEXT(pg_current, link); 3603 if (pg_current == NULL) { 3604 pg_current = TAILQ_FIRST(&rtransport->poll_groups); 3605 } 3606 3607 if (count < min_value) { 3608 min_value = count; 3609 pg_min = pg_current; 3610 } 3611 3612 if (pg_current == pg_start) { 3613 break; 3614 } 3615 } 3616 *pg = pg_min; 3617 } 3618 3619 assert(*pg != NULL); 3620 3621 result = &(*pg)->group; 3622 3623 *pg = TAILQ_NEXT(*pg, link); 3624 if (*pg == NULL) { 3625 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3626 } 3627 3628 return result; 3629 } 3630 3631 static void 3632 nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller) 3633 { 3634 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3635 RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) { 3636 nvmf_rdma_qpair_destroy(qpair); 3637 } 3638 3639 if (poller->srq) { 3640 if (poller->resources) { 3641 nvmf_rdma_resources_destroy(poller->resources); 3642 } 3643 spdk_rdma_srq_destroy(poller->srq); 3644 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 3645 } 3646 3647 if (poller->cq) { 3648 ibv_destroy_cq(poller->cq); 3649 } 3650 3651 free(poller); 3652 } 3653 3654 static void 3655 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3656 { 3657 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3658 struct spdk_nvmf_rdma_poller *poller, *tmp; 3659 struct spdk_nvmf_rdma_transport *rtransport; 3660 3661 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3662 if (!rgroup) { 3663 return; 3664 } 3665 3666 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3667 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3668 nvmf_rdma_poller_destroy(poller); 3669 } 3670 3671 if (rgroup->group.transport == NULL) { 3672 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3673 * calls this function directly in a failure path. */ 3674 free(rgroup); 3675 return; 3676 } 3677 3678 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3679 3680 next_rgroup = TAILQ_NEXT(rgroup, link); 3681 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3682 if (next_rgroup == NULL) { 3683 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3684 } 3685 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3686 rtransport->conn_sched.next_admin_pg = next_rgroup; 3687 } 3688 if (rtransport->conn_sched.next_io_pg == rgroup) { 3689 rtransport->conn_sched.next_io_pg = next_rgroup; 3690 } 3691 3692 free(rgroup); 3693 } 3694 3695 static void 3696 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3697 { 3698 if (rqpair->cm_id != NULL) { 3699 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3700 } 3701 } 3702 3703 static int 3704 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3705 struct spdk_nvmf_qpair *qpair) 3706 { 3707 struct spdk_nvmf_rdma_poll_group *rgroup; 3708 struct spdk_nvmf_rdma_qpair *rqpair; 3709 struct spdk_nvmf_rdma_device *device; 3710 struct spdk_nvmf_rdma_poller *poller; 3711 int rc; 3712 3713 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3714 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3715 3716 device = rqpair->device; 3717 3718 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3719 if (poller->device == device) { 3720 break; 3721 } 3722 } 3723 3724 if (!poller) { 3725 SPDK_ERRLOG("No poller found for device.\n"); 3726 return -1; 3727 } 3728 3729 rqpair->poller = poller; 3730 rqpair->srq = rqpair->poller->srq; 3731 3732 rc = nvmf_rdma_qpair_initialize(qpair); 3733 if (rc < 0) { 3734 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3735 rqpair->poller = NULL; 3736 rqpair->srq = NULL; 3737 return -1; 3738 } 3739 3740 RB_INSERT(qpairs_tree, &poller->qpairs, rqpair); 3741 3742 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3743 if (rc) { 3744 /* Try to reject, but we probably can't */ 3745 nvmf_rdma_qpair_reject_connection(rqpair); 3746 return -1; 3747 } 3748 3749 nvmf_rdma_update_ibv_state(rqpair); 3750 3751 return 0; 3752 } 3753 3754 static int 3755 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3756 struct spdk_nvmf_qpair *qpair) 3757 { 3758 struct spdk_nvmf_rdma_qpair *rqpair; 3759 3760 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3761 assert(group->transport->tgt != NULL); 3762 3763 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3764 3765 if (!rqpair->destruct_channel) { 3766 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3767 return 0; 3768 } 3769 3770 /* Sanity check that we get io_channel on the correct thread */ 3771 if (qpair->group) { 3772 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3773 } 3774 3775 return 0; 3776 } 3777 3778 static int 3779 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3780 { 3781 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3782 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3783 struct spdk_nvmf_rdma_transport, transport); 3784 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3785 struct spdk_nvmf_rdma_qpair, qpair); 3786 3787 /* 3788 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3789 * needs to be returned to the shared receive queue or the poll group will eventually be 3790 * starved of RECV structures. 3791 */ 3792 if (rqpair->srq && rdma_req->recv) { 3793 int rc; 3794 struct ibv_recv_wr *bad_recv_wr; 3795 3796 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 3797 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 3798 if (rc) { 3799 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3800 } 3801 } 3802 3803 _nvmf_rdma_request_free(rdma_req, rtransport); 3804 return 0; 3805 } 3806 3807 static int 3808 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3809 { 3810 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3811 struct spdk_nvmf_rdma_transport, transport); 3812 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3813 struct spdk_nvmf_rdma_request, req); 3814 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3815 struct spdk_nvmf_rdma_qpair, qpair); 3816 3817 if (rqpair->ibv_state != IBV_QPS_ERR) { 3818 /* The connection is alive, so process the request as normal */ 3819 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3820 } else { 3821 /* The connection is dead. Move the request directly to the completed state. */ 3822 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3823 } 3824 3825 nvmf_rdma_request_process(rtransport, rdma_req); 3826 3827 return 0; 3828 } 3829 3830 static void 3831 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 3832 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3833 { 3834 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3835 3836 rqpair->to_close = true; 3837 3838 /* This happens only when the qpair is disconnected before 3839 * it is added to the poll group. Since there is no poll group, 3840 * the RDMA qp has not been initialized yet and the RDMA CM 3841 * event has not yet been acknowledged, so we need to reject it. 3842 */ 3843 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3844 nvmf_rdma_qpair_reject_connection(rqpair); 3845 nvmf_rdma_qpair_destroy(rqpair); 3846 return; 3847 } 3848 3849 if (rqpair->rdma_qp) { 3850 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3851 } 3852 3853 nvmf_rdma_destroy_drained_qpair(rqpair); 3854 3855 if (cb_fn) { 3856 cb_fn(cb_arg); 3857 } 3858 } 3859 3860 static struct spdk_nvmf_rdma_qpair * 3861 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3862 { 3863 struct spdk_nvmf_rdma_qpair find; 3864 3865 find.qp_num = wc->qp_num; 3866 3867 return RB_FIND(qpairs_tree, &rpoller->qpairs, &find); 3868 } 3869 3870 #ifdef DEBUG 3871 static int 3872 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3873 { 3874 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3875 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3876 } 3877 #endif 3878 3879 static void 3880 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3881 int rc) 3882 { 3883 struct spdk_nvmf_rdma_recv *rdma_recv; 3884 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3885 3886 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3887 while (bad_recv_wr != NULL) { 3888 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3889 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3890 3891 rdma_recv->qpair->current_recv_depth++; 3892 bad_recv_wr = bad_recv_wr->next; 3893 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3894 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3895 } 3896 } 3897 3898 static void 3899 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3900 { 3901 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3902 while (bad_recv_wr != NULL) { 3903 bad_recv_wr = bad_recv_wr->next; 3904 rqpair->current_recv_depth++; 3905 } 3906 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3907 } 3908 3909 static void 3910 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3911 struct spdk_nvmf_rdma_poller *rpoller) 3912 { 3913 struct spdk_nvmf_rdma_qpair *rqpair; 3914 struct ibv_recv_wr *bad_recv_wr; 3915 int rc; 3916 3917 if (rpoller->srq) { 3918 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 3919 if (rc) { 3920 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3921 } 3922 } else { 3923 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3924 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3925 rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 3926 if (rc) { 3927 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3928 } 3929 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3930 } 3931 } 3932 } 3933 3934 static void 3935 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3936 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3937 { 3938 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3939 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3940 3941 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3942 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3943 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3944 assert(rqpair->current_send_depth > 0); 3945 rqpair->current_send_depth--; 3946 switch (bad_rdma_wr->type) { 3947 case RDMA_WR_TYPE_DATA: 3948 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3949 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3950 assert(rqpair->current_read_depth > 0); 3951 rqpair->current_read_depth--; 3952 } 3953 break; 3954 case RDMA_WR_TYPE_SEND: 3955 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3956 break; 3957 default: 3958 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3959 prev_rdma_req = cur_rdma_req; 3960 continue; 3961 } 3962 3963 if (prev_rdma_req == cur_rdma_req) { 3964 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3965 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3966 continue; 3967 } 3968 3969 switch (cur_rdma_req->state) { 3970 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3971 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3972 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3973 break; 3974 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3975 case RDMA_REQUEST_STATE_COMPLETING: 3976 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3977 break; 3978 default: 3979 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3980 cur_rdma_req->state, rqpair); 3981 continue; 3982 } 3983 3984 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3985 prev_rdma_req = cur_rdma_req; 3986 } 3987 3988 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3989 /* Disconnect the connection. */ 3990 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3991 } 3992 3993 } 3994 3995 static void 3996 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3997 struct spdk_nvmf_rdma_poller *rpoller) 3998 { 3999 struct spdk_nvmf_rdma_qpair *rqpair; 4000 struct ibv_send_wr *bad_wr = NULL; 4001 int rc; 4002 4003 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 4004 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 4005 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 4006 4007 /* bad wr always points to the first wr that failed. */ 4008 if (rc) { 4009 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 4010 } 4011 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 4012 } 4013 } 4014 4015 static const char * 4016 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 4017 { 4018 switch (wr_type) { 4019 case RDMA_WR_TYPE_RECV: 4020 return "RECV"; 4021 case RDMA_WR_TYPE_SEND: 4022 return "SEND"; 4023 case RDMA_WR_TYPE_DATA: 4024 return "DATA"; 4025 default: 4026 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 4027 SPDK_UNREACHABLE(); 4028 } 4029 } 4030 4031 static inline void 4032 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 4033 { 4034 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 4035 4036 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 4037 /* If qpair is in ERR state, we will receive completions for all posted and not completed 4038 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 4039 SPDK_DEBUGLOG(rdma, 4040 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 4041 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 4042 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4043 } else { 4044 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 4045 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 4046 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4047 } 4048 } 4049 4050 static int 4051 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 4052 struct spdk_nvmf_rdma_poller *rpoller) 4053 { 4054 struct ibv_wc wc[32]; 4055 struct spdk_nvmf_rdma_wr *rdma_wr; 4056 struct spdk_nvmf_rdma_request *rdma_req; 4057 struct spdk_nvmf_rdma_recv *rdma_recv; 4058 struct spdk_nvmf_rdma_qpair *rqpair; 4059 int reaped, i; 4060 int count = 0; 4061 bool error = false; 4062 uint64_t poll_tsc = spdk_get_ticks(); 4063 4064 /* Poll for completing operations. */ 4065 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 4066 if (reaped < 0) { 4067 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 4068 errno, spdk_strerror(errno)); 4069 return -1; 4070 } else if (reaped == 0) { 4071 rpoller->stat.idle_polls++; 4072 } 4073 4074 rpoller->stat.polls++; 4075 rpoller->stat.completions += reaped; 4076 4077 for (i = 0; i < reaped; i++) { 4078 4079 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 4080 4081 switch (rdma_wr->type) { 4082 case RDMA_WR_TYPE_SEND: 4083 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 4084 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4085 4086 if (!wc[i].status) { 4087 count++; 4088 assert(wc[i].opcode == IBV_WC_SEND); 4089 assert(nvmf_rdma_req_is_completing(rdma_req)); 4090 } 4091 4092 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4093 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 4094 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 4095 rdma_req->num_outstanding_data_wr = 0; 4096 4097 nvmf_rdma_request_process(rtransport, rdma_req); 4098 break; 4099 case RDMA_WR_TYPE_RECV: 4100 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 4101 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4102 if (rpoller->srq != NULL) { 4103 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 4104 /* It is possible that there are still some completions for destroyed QP 4105 * associated with SRQ. We just ignore these late completions and re-post 4106 * receive WRs back to SRQ. 4107 */ 4108 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 4109 struct ibv_recv_wr *bad_wr; 4110 int rc; 4111 4112 rdma_recv->wr.next = NULL; 4113 spdk_rdma_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 4114 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 4115 if (rc) { 4116 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 4117 } 4118 continue; 4119 } 4120 } 4121 rqpair = rdma_recv->qpair; 4122 4123 assert(rqpair != NULL); 4124 if (!wc[i].status) { 4125 assert(wc[i].opcode == IBV_WC_RECV); 4126 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 4127 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4128 break; 4129 } 4130 } 4131 4132 rdma_recv->wr.next = NULL; 4133 rqpair->current_recv_depth++; 4134 rdma_recv->receive_tsc = poll_tsc; 4135 rpoller->stat.requests++; 4136 STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link); 4137 break; 4138 case RDMA_WR_TYPE_DATA: 4139 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 4140 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4141 4142 assert(rdma_req->num_outstanding_data_wr > 0); 4143 4144 rqpair->current_send_depth--; 4145 rdma_req->num_outstanding_data_wr--; 4146 if (!wc[i].status) { 4147 assert(wc[i].opcode == IBV_WC_RDMA_READ); 4148 rqpair->current_read_depth--; 4149 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 4150 if (rdma_req->num_outstanding_data_wr == 0) { 4151 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 4152 nvmf_rdma_request_process(rtransport, rdma_req); 4153 } 4154 } else { 4155 /* If the data transfer fails still force the queue into the error state, 4156 * if we were performing an RDMA_READ, we need to force the request into a 4157 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 4158 * case, we should wait for the SEND to complete. */ 4159 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 4160 rqpair->current_read_depth--; 4161 if (rdma_req->num_outstanding_data_wr == 0) { 4162 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4163 } 4164 } 4165 } 4166 break; 4167 default: 4168 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 4169 continue; 4170 } 4171 4172 /* Handle error conditions */ 4173 if (wc[i].status) { 4174 nvmf_rdma_update_ibv_state(rqpair); 4175 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 4176 4177 error = true; 4178 4179 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 4180 /* Disconnect the connection. */ 4181 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4182 } else { 4183 nvmf_rdma_destroy_drained_qpair(rqpair); 4184 } 4185 continue; 4186 } 4187 4188 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 4189 4190 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 4191 nvmf_rdma_destroy_drained_qpair(rqpair); 4192 } 4193 } 4194 4195 if (error == true) { 4196 return -1; 4197 } 4198 4199 /* submit outstanding work requests. */ 4200 _poller_submit_recvs(rtransport, rpoller); 4201 _poller_submit_sends(rtransport, rpoller); 4202 4203 return count; 4204 } 4205 4206 static int 4207 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 4208 { 4209 struct spdk_nvmf_rdma_transport *rtransport; 4210 struct spdk_nvmf_rdma_poll_group *rgroup; 4211 struct spdk_nvmf_rdma_poller *rpoller; 4212 int count, rc; 4213 4214 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 4215 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4216 4217 count = 0; 4218 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4219 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4220 if (rc < 0) { 4221 return rc; 4222 } 4223 count += rc; 4224 } 4225 4226 return count; 4227 } 4228 4229 static int 4230 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4231 struct spdk_nvme_transport_id *trid, 4232 bool peer) 4233 { 4234 struct sockaddr *saddr; 4235 uint16_t port; 4236 4237 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4238 4239 if (peer) { 4240 saddr = rdma_get_peer_addr(id); 4241 } else { 4242 saddr = rdma_get_local_addr(id); 4243 } 4244 switch (saddr->sa_family) { 4245 case AF_INET: { 4246 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4247 4248 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4249 inet_ntop(AF_INET, &saddr_in->sin_addr, 4250 trid->traddr, sizeof(trid->traddr)); 4251 if (peer) { 4252 port = ntohs(rdma_get_dst_port(id)); 4253 } else { 4254 port = ntohs(rdma_get_src_port(id)); 4255 } 4256 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4257 break; 4258 } 4259 case AF_INET6: { 4260 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4261 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4262 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4263 trid->traddr, sizeof(trid->traddr)); 4264 if (peer) { 4265 port = ntohs(rdma_get_dst_port(id)); 4266 } else { 4267 port = ntohs(rdma_get_src_port(id)); 4268 } 4269 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4270 break; 4271 } 4272 default: 4273 return -1; 4274 4275 } 4276 4277 return 0; 4278 } 4279 4280 static int 4281 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4282 struct spdk_nvme_transport_id *trid) 4283 { 4284 struct spdk_nvmf_rdma_qpair *rqpair; 4285 4286 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4287 4288 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4289 } 4290 4291 static int 4292 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4293 struct spdk_nvme_transport_id *trid) 4294 { 4295 struct spdk_nvmf_rdma_qpair *rqpair; 4296 4297 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4298 4299 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4300 } 4301 4302 static int 4303 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4304 struct spdk_nvme_transport_id *trid) 4305 { 4306 struct spdk_nvmf_rdma_qpair *rqpair; 4307 4308 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4309 4310 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4311 } 4312 4313 void 4314 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4315 { 4316 g_nvmf_hooks = *hooks; 4317 } 4318 4319 static void 4320 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4321 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4322 { 4323 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4324 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4325 4326 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4327 4328 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4329 } 4330 4331 static int 4332 _nvmf_rdma_qpair_abort_request(void *ctx) 4333 { 4334 struct spdk_nvmf_request *req = ctx; 4335 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4336 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4337 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4338 struct spdk_nvmf_rdma_qpair, qpair); 4339 int rc; 4340 4341 spdk_poller_unregister(&req->poller); 4342 4343 switch (rdma_req_to_abort->state) { 4344 case RDMA_REQUEST_STATE_EXECUTING: 4345 rc = nvmf_ctrlr_abort_request(req); 4346 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4347 return SPDK_POLLER_BUSY; 4348 } 4349 break; 4350 4351 case RDMA_REQUEST_STATE_NEED_BUFFER: 4352 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4353 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4354 4355 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4356 break; 4357 4358 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4359 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4360 spdk_nvmf_rdma_request, state_link); 4361 4362 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4363 break; 4364 4365 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4366 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4367 spdk_nvmf_rdma_request, state_link); 4368 4369 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4370 break; 4371 4372 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4373 if (spdk_get_ticks() < req->timeout_tsc) { 4374 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4375 return SPDK_POLLER_BUSY; 4376 } 4377 break; 4378 4379 default: 4380 break; 4381 } 4382 4383 spdk_nvmf_request_complete(req); 4384 return SPDK_POLLER_BUSY; 4385 } 4386 4387 static void 4388 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4389 struct spdk_nvmf_request *req) 4390 { 4391 struct spdk_nvmf_rdma_qpair *rqpair; 4392 struct spdk_nvmf_rdma_transport *rtransport; 4393 struct spdk_nvmf_transport *transport; 4394 uint16_t cid; 4395 uint32_t i, max_req_count; 4396 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 4397 4398 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4399 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4400 transport = &rtransport->transport; 4401 4402 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4403 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 4404 4405 for (i = 0; i < max_req_count; i++) { 4406 rdma_req = &rqpair->resources->reqs[i]; 4407 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 4408 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 4409 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 4410 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 4411 rdma_req->req.qpair == qpair) { 4412 rdma_req_to_abort = rdma_req; 4413 break; 4414 } 4415 } 4416 4417 if (rdma_req_to_abort == NULL) { 4418 spdk_nvmf_request_complete(req); 4419 return; 4420 } 4421 4422 req->req_to_abort = &rdma_req_to_abort->req; 4423 req->timeout_tsc = spdk_get_ticks() + 4424 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4425 req->poller = NULL; 4426 4427 _nvmf_rdma_qpair_abort_request(req); 4428 } 4429 4430 static void 4431 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 4432 struct spdk_json_write_ctx *w) 4433 { 4434 struct spdk_nvmf_rdma_poll_group *rgroup; 4435 struct spdk_nvmf_rdma_poller *rpoller; 4436 4437 assert(w != NULL); 4438 4439 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4440 4441 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 4442 4443 spdk_json_write_named_array_begin(w, "devices"); 4444 4445 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4446 spdk_json_write_object_begin(w); 4447 spdk_json_write_named_string(w, "name", 4448 ibv_get_device_name(rpoller->device->context->device)); 4449 spdk_json_write_named_uint64(w, "polls", 4450 rpoller->stat.polls); 4451 spdk_json_write_named_uint64(w, "idle_polls", 4452 rpoller->stat.idle_polls); 4453 spdk_json_write_named_uint64(w, "completions", 4454 rpoller->stat.completions); 4455 spdk_json_write_named_uint64(w, "requests", 4456 rpoller->stat.requests); 4457 spdk_json_write_named_uint64(w, "request_latency", 4458 rpoller->stat.request_latency); 4459 spdk_json_write_named_uint64(w, "pending_free_request", 4460 rpoller->stat.pending_free_request); 4461 spdk_json_write_named_uint64(w, "pending_rdma_read", 4462 rpoller->stat.pending_rdma_read); 4463 spdk_json_write_named_uint64(w, "pending_rdma_write", 4464 rpoller->stat.pending_rdma_write); 4465 spdk_json_write_named_uint64(w, "total_send_wrs", 4466 rpoller->stat.qp_stats.send.num_submitted_wrs); 4467 spdk_json_write_named_uint64(w, "send_doorbell_updates", 4468 rpoller->stat.qp_stats.send.doorbell_updates); 4469 spdk_json_write_named_uint64(w, "total_recv_wrs", 4470 rpoller->stat.qp_stats.recv.num_submitted_wrs); 4471 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 4472 rpoller->stat.qp_stats.recv.doorbell_updates); 4473 spdk_json_write_object_end(w); 4474 } 4475 4476 spdk_json_write_array_end(w); 4477 } 4478 4479 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4480 .name = "RDMA", 4481 .type = SPDK_NVME_TRANSPORT_RDMA, 4482 .opts_init = nvmf_rdma_opts_init, 4483 .create = nvmf_rdma_create, 4484 .dump_opts = nvmf_rdma_dump_opts, 4485 .destroy = nvmf_rdma_destroy, 4486 4487 .listen = nvmf_rdma_listen, 4488 .stop_listen = nvmf_rdma_stop_listen, 4489 .cdata_init = nvmf_rdma_cdata_init, 4490 4491 .listener_discover = nvmf_rdma_discover, 4492 4493 .poll_group_create = nvmf_rdma_poll_group_create, 4494 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4495 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4496 .poll_group_add = nvmf_rdma_poll_group_add, 4497 .poll_group_remove = nvmf_rdma_poll_group_remove, 4498 .poll_group_poll = nvmf_rdma_poll_group_poll, 4499 4500 .req_free = nvmf_rdma_request_free, 4501 .req_complete = nvmf_rdma_request_complete, 4502 4503 .qpair_fini = nvmf_rdma_close_qpair, 4504 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4505 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4506 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4507 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4508 4509 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 4510 }; 4511 4512 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4513 SPDK_LOG_REGISTER_COMPONENT(rdma) 4514