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