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