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 * When using a shared receive queue on the target side, 1209 * we need to pass this value to the initiator to prevent the 1210 * initiator side NIC from completing SEND requests back to the 1211 * initiator with status rnr_retry_count_exceeded. */ 1212 if (rqpair->srq != NULL) { 1213 ctrlr_event_data.rnr_retry_count = 0x7; 1214 } 1215 1216 /* When qpair is created without use of rdma cm API, an additional 1217 * information must be provided to initiator in the connection response: 1218 * whether qpair is using SRQ and its qp_num 1219 * Fields below are ignored by rdma cm if qpair has been 1220 * created using rdma cm API. */ 1221 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1222 ctrlr_event_data.qp_num = rqpair->qp_num; 1223 1224 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1225 if (rc) { 1226 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1227 } else { 1228 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1229 } 1230 1231 return rc; 1232 } 1233 1234 static void 1235 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1236 { 1237 struct spdk_nvmf_rdma_reject_private_data rej_data; 1238 1239 rej_data.recfmt = 0; 1240 rej_data.sts = error; 1241 1242 rdma_reject(id, &rej_data, sizeof(rej_data)); 1243 } 1244 1245 static int 1246 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1247 { 1248 struct spdk_nvmf_rdma_transport *rtransport; 1249 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1250 struct spdk_nvmf_rdma_port *port; 1251 struct rdma_conn_param *rdma_param = NULL; 1252 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1253 uint16_t max_queue_depth; 1254 uint16_t max_read_depth; 1255 1256 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1257 1258 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1259 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1260 1261 rdma_param = &event->param.conn; 1262 if (rdma_param->private_data == NULL || 1263 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1264 SPDK_ERRLOG("connect request: no private data provided\n"); 1265 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1266 return -1; 1267 } 1268 1269 private_data = rdma_param->private_data; 1270 if (private_data->recfmt != 0) { 1271 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1272 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1273 return -1; 1274 } 1275 1276 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1277 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1278 1279 port = event->listen_id->context; 1280 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1281 event->listen_id, event->listen_id->verbs, port); 1282 1283 /* Figure out the supported queue depth. This is a multi-step process 1284 * that takes into account hardware maximums, host provided values, 1285 * and our target's internal memory limits */ 1286 1287 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1288 1289 /* Start with the maximum queue depth allowed by the target */ 1290 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1291 max_read_depth = rtransport->transport.opts.max_queue_depth; 1292 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1293 rtransport->transport.opts.max_queue_depth); 1294 1295 /* Next check the local NIC's hardware limitations */ 1296 SPDK_DEBUGLOG(rdma, 1297 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1298 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1299 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1300 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1301 1302 /* Next check the remote NIC's hardware limitations */ 1303 SPDK_DEBUGLOG(rdma, 1304 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1305 rdma_param->initiator_depth, rdma_param->responder_resources); 1306 if (rdma_param->initiator_depth > 0) { 1307 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1308 } 1309 1310 /* Finally check for the host software requested values, which are 1311 * optional. */ 1312 if (rdma_param->private_data != NULL && 1313 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1314 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1315 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1316 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1317 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1318 } 1319 1320 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1321 max_queue_depth, max_read_depth); 1322 1323 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1324 if (rqpair == NULL) { 1325 SPDK_ERRLOG("Could not allocate new connection.\n"); 1326 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1327 return -1; 1328 } 1329 1330 rqpair->device = port->device; 1331 rqpair->max_queue_depth = max_queue_depth; 1332 rqpair->max_read_depth = max_read_depth; 1333 rqpair->cm_id = event->id; 1334 rqpair->listen_id = event->listen_id; 1335 rqpair->qpair.transport = transport; 1336 STAILQ_INIT(&rqpair->ibv_events); 1337 /* use qid from the private data to determine the qpair type 1338 qid will be set to the appropriate value when the controller is created */ 1339 rqpair->qpair.qid = private_data->qid; 1340 1341 event->id->context = &rqpair->qpair; 1342 1343 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1344 1345 return 0; 1346 } 1347 1348 static inline void 1349 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1350 enum spdk_nvme_data_transfer xfer) 1351 { 1352 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1353 wr->opcode = IBV_WR_RDMA_WRITE; 1354 wr->send_flags = 0; 1355 wr->next = next; 1356 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1357 wr->opcode = IBV_WR_RDMA_READ; 1358 wr->send_flags = IBV_SEND_SIGNALED; 1359 wr->next = NULL; 1360 } else { 1361 assert(0); 1362 } 1363 } 1364 1365 static int 1366 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1367 struct spdk_nvmf_rdma_request *rdma_req, 1368 uint32_t num_sgl_descriptors) 1369 { 1370 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1371 struct spdk_nvmf_rdma_request_data *current_data_wr; 1372 uint32_t i; 1373 1374 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1375 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1376 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1377 return -EINVAL; 1378 } 1379 1380 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1381 return -ENOMEM; 1382 } 1383 1384 current_data_wr = &rdma_req->data; 1385 1386 for (i = 0; i < num_sgl_descriptors; i++) { 1387 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1388 current_data_wr->wr.next = &work_requests[i]->wr; 1389 current_data_wr = work_requests[i]; 1390 current_data_wr->wr.sg_list = current_data_wr->sgl; 1391 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1392 } 1393 1394 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1395 1396 return 0; 1397 } 1398 1399 static inline void 1400 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1401 { 1402 struct ibv_send_wr *wr = &rdma_req->data.wr; 1403 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1404 1405 wr->wr.rdma.rkey = sgl->keyed.key; 1406 wr->wr.rdma.remote_addr = sgl->address; 1407 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1408 } 1409 1410 static inline void 1411 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1412 { 1413 struct ibv_send_wr *wr = &rdma_req->data.wr; 1414 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1415 uint32_t i; 1416 int j; 1417 uint64_t remote_addr_offset = 0; 1418 1419 for (i = 0; i < num_wrs; ++i) { 1420 wr->wr.rdma.rkey = sgl->keyed.key; 1421 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1422 for (j = 0; j < wr->num_sge; ++j) { 1423 remote_addr_offset += wr->sg_list[j].length; 1424 } 1425 wr = wr->next; 1426 } 1427 } 1428 1429 static int 1430 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1431 struct spdk_nvmf_rdma_device *device, 1432 struct spdk_nvmf_rdma_request *rdma_req, 1433 struct ibv_send_wr *wr, 1434 uint32_t total_length, 1435 uint32_t num_extra_wrs) 1436 { 1437 struct spdk_rdma_memory_translation mem_translation; 1438 struct spdk_dif_ctx *dif_ctx = NULL; 1439 struct ibv_sge *sg_ele; 1440 struct iovec *iov; 1441 uint32_t remaining_data_block = 0; 1442 uint32_t lkey, remaining; 1443 int rc; 1444 1445 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 1446 dif_ctx = &rdma_req->req.dif.dif_ctx; 1447 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1448 } 1449 1450 wr->num_sge = 0; 1451 1452 while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1453 iov = &rdma_req->req.iov[rdma_req->iovpos]; 1454 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1455 if (spdk_unlikely(rc)) { 1456 return false; 1457 } 1458 1459 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1460 sg_ele = &wr->sg_list[wr->num_sge]; 1461 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1462 1463 if (spdk_likely(!dif_ctx)) { 1464 sg_ele->lkey = lkey; 1465 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset; 1466 sg_ele->length = remaining; 1467 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1468 sg_ele->length); 1469 rdma_req->offset += sg_ele->length; 1470 total_length -= sg_ele->length; 1471 wr->num_sge++; 1472 1473 if (rdma_req->offset == iov->iov_len) { 1474 rdma_req->offset = 0; 1475 rdma_req->iovpos++; 1476 } 1477 } else { 1478 uint32_t data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1479 uint32_t md_size = dif_ctx->md_size; 1480 uint32_t sge_len; 1481 1482 while (remaining) { 1483 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1484 if (num_extra_wrs > 0 && wr->next) { 1485 wr = wr->next; 1486 wr->num_sge = 0; 1487 sg_ele = &wr->sg_list[wr->num_sge]; 1488 num_extra_wrs--; 1489 } else { 1490 break; 1491 } 1492 } 1493 sg_ele->lkey = lkey; 1494 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset); 1495 sge_len = spdk_min(remaining, remaining_data_block); 1496 sg_ele->length = sge_len; 1497 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1498 sg_ele->length); 1499 remaining -= sge_len; 1500 remaining_data_block -= sge_len; 1501 rdma_req->offset += sge_len; 1502 total_length -= sge_len; 1503 1504 sg_ele++; 1505 wr->num_sge++; 1506 1507 if (remaining_data_block == 0) { 1508 /* skip metadata */ 1509 rdma_req->offset += md_size; 1510 total_length -= md_size; 1511 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1512 remaining -= spdk_min(remaining, md_size); 1513 remaining_data_block = data_block_size; 1514 } 1515 1516 if (remaining == 0) { 1517 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1518 the remaining metadata bits at the beginning of the next buffer */ 1519 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset); 1520 rdma_req->iovpos++; 1521 } 1522 } 1523 } 1524 } 1525 1526 if (total_length) { 1527 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1528 return -EINVAL; 1529 } 1530 1531 return 0; 1532 } 1533 1534 static inline uint32_t 1535 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1536 { 1537 /* estimate the number of SG entries and WRs needed to process the request */ 1538 uint32_t num_sge = 0; 1539 uint32_t i; 1540 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1541 1542 for (i = 0; i < num_buffers && length > 0; i++) { 1543 uint32_t buffer_len = spdk_min(length, io_unit_size); 1544 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1545 1546 if (num_sge_in_block * block_size > buffer_len) { 1547 ++num_sge_in_block; 1548 } 1549 num_sge += num_sge_in_block; 1550 length -= buffer_len; 1551 } 1552 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1553 } 1554 1555 static int 1556 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1557 struct spdk_nvmf_rdma_device *device, 1558 struct spdk_nvmf_rdma_request *rdma_req, 1559 uint32_t length) 1560 { 1561 struct spdk_nvmf_rdma_qpair *rqpair; 1562 struct spdk_nvmf_rdma_poll_group *rgroup; 1563 struct spdk_nvmf_request *req = &rdma_req->req; 1564 struct ibv_send_wr *wr = &rdma_req->data.wr; 1565 int rc; 1566 uint32_t num_wrs = 1; 1567 1568 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1569 rgroup = rqpair->poller->group; 1570 1571 /* rdma wr specifics */ 1572 nvmf_rdma_setup_request(rdma_req); 1573 1574 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1575 length); 1576 if (rc != 0) { 1577 return rc; 1578 } 1579 1580 assert(req->iovcnt <= rqpair->max_send_sge); 1581 1582 rdma_req->iovpos = 0; 1583 1584 if (spdk_unlikely(req->dif_enabled)) { 1585 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1586 req->dif.dif_ctx.block_size); 1587 if (num_wrs > 1) { 1588 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1589 if (rc != 0) { 1590 goto err_exit; 1591 } 1592 } 1593 } 1594 1595 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1596 if (spdk_unlikely(rc != 0)) { 1597 goto err_exit; 1598 } 1599 1600 if (spdk_unlikely(num_wrs > 1)) { 1601 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1602 } 1603 1604 /* set the number of outstanding data WRs for this request. */ 1605 rdma_req->num_outstanding_data_wr = num_wrs; 1606 1607 return rc; 1608 1609 err_exit: 1610 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1611 nvmf_rdma_request_free_data(rdma_req, rtransport); 1612 req->iovcnt = 0; 1613 return rc; 1614 } 1615 1616 static int 1617 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1618 struct spdk_nvmf_rdma_device *device, 1619 struct spdk_nvmf_rdma_request *rdma_req) 1620 { 1621 struct spdk_nvmf_rdma_qpair *rqpair; 1622 struct spdk_nvmf_rdma_poll_group *rgroup; 1623 struct ibv_send_wr *current_wr; 1624 struct spdk_nvmf_request *req = &rdma_req->req; 1625 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1626 uint32_t num_sgl_descriptors; 1627 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0; 1628 uint32_t i; 1629 int rc; 1630 1631 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1632 rgroup = rqpair->poller->group; 1633 1634 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1635 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1636 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1637 1638 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1639 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1640 1641 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1642 for (i = 0; i < num_sgl_descriptors; i++) { 1643 if (spdk_likely(!req->dif_enabled)) { 1644 lengths[i] = desc->keyed.length; 1645 } else { 1646 req->dif.orig_length += desc->keyed.length; 1647 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1648 req->dif.elba_length += lengths[i]; 1649 } 1650 total_length += lengths[i]; 1651 desc++; 1652 } 1653 1654 if (total_length > rtransport->transport.opts.max_io_size) { 1655 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n", 1656 total_length, rtransport->transport.opts.max_io_size); 1657 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1658 return -EINVAL; 1659 } 1660 1661 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1662 return -ENOMEM; 1663 } 1664 1665 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length); 1666 if (rc != 0) { 1667 nvmf_rdma_request_free_data(rdma_req, rtransport); 1668 return rc; 1669 } 1670 1671 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1672 current_wr = &rdma_req->data.wr; 1673 assert(current_wr != NULL); 1674 1675 req->length = 0; 1676 rdma_req->iovpos = 0; 1677 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1678 for (i = 0; i < num_sgl_descriptors; i++) { 1679 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1680 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1681 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1682 rc = -EINVAL; 1683 goto err_exit; 1684 } 1685 1686 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1687 if (rc != 0) { 1688 rc = -ENOMEM; 1689 goto err_exit; 1690 } 1691 1692 req->length += desc->keyed.length; 1693 current_wr->wr.rdma.rkey = desc->keyed.key; 1694 current_wr->wr.rdma.remote_addr = desc->address; 1695 current_wr = current_wr->next; 1696 desc++; 1697 } 1698 1699 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1700 /* Go back to the last descriptor in the list. */ 1701 desc--; 1702 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1703 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1704 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1705 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1706 } 1707 } 1708 #endif 1709 1710 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1711 1712 return 0; 1713 1714 err_exit: 1715 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1716 nvmf_rdma_request_free_data(rdma_req, rtransport); 1717 return rc; 1718 } 1719 1720 static int 1721 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1722 struct spdk_nvmf_rdma_device *device, 1723 struct spdk_nvmf_rdma_request *rdma_req) 1724 { 1725 struct spdk_nvmf_request *req = &rdma_req->req; 1726 struct spdk_nvme_cpl *rsp; 1727 struct spdk_nvme_sgl_descriptor *sgl; 1728 int rc; 1729 uint32_t length; 1730 1731 rsp = &req->rsp->nvme_cpl; 1732 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1733 1734 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1735 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1736 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1737 1738 length = sgl->keyed.length; 1739 if (length > rtransport->transport.opts.max_io_size) { 1740 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1741 length, rtransport->transport.opts.max_io_size); 1742 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1743 return -1; 1744 } 1745 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1746 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1747 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1748 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1749 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1750 } 1751 } 1752 #endif 1753 1754 /* fill request length and populate iovs */ 1755 req->length = length; 1756 1757 if (spdk_unlikely(req->dif_enabled)) { 1758 req->dif.orig_length = length; 1759 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1760 req->dif.elba_length = length; 1761 } 1762 1763 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1764 if (spdk_unlikely(rc < 0)) { 1765 if (rc == -EINVAL) { 1766 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1767 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1768 return -1; 1769 } 1770 /* No available buffers. Queue this request up. */ 1771 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1772 return 0; 1773 } 1774 1775 /* backward compatible */ 1776 req->data = req->iov[0].iov_base; 1777 1778 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1779 req->iovcnt); 1780 1781 return 0; 1782 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1783 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1784 uint64_t offset = sgl->address; 1785 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1786 1787 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1788 offset, sgl->unkeyed.length); 1789 1790 if (offset > max_len) { 1791 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1792 offset, max_len); 1793 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1794 return -1; 1795 } 1796 max_len -= (uint32_t)offset; 1797 1798 if (sgl->unkeyed.length > max_len) { 1799 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1800 sgl->unkeyed.length, max_len); 1801 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1802 return -1; 1803 } 1804 1805 rdma_req->num_outstanding_data_wr = 0; 1806 req->data = rdma_req->recv->buf + offset; 1807 req->data_from_pool = false; 1808 req->length = sgl->unkeyed.length; 1809 1810 req->iov[0].iov_base = req->data; 1811 req->iov[0].iov_len = req->length; 1812 req->iovcnt = 1; 1813 1814 return 0; 1815 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1816 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1817 1818 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1819 if (rc == -ENOMEM) { 1820 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1821 return 0; 1822 } else if (rc == -EINVAL) { 1823 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1824 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1825 return -1; 1826 } 1827 1828 /* backward compatible */ 1829 req->data = req->iov[0].iov_base; 1830 1831 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1832 req->iovcnt); 1833 1834 return 0; 1835 } 1836 1837 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1838 sgl->generic.type, sgl->generic.subtype); 1839 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1840 return -1; 1841 } 1842 1843 static void 1844 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1845 struct spdk_nvmf_rdma_transport *rtransport) 1846 { 1847 struct spdk_nvmf_rdma_qpair *rqpair; 1848 struct spdk_nvmf_rdma_poll_group *rgroup; 1849 1850 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1851 if (rdma_req->req.data_from_pool) { 1852 rgroup = rqpair->poller->group; 1853 1854 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1855 } 1856 nvmf_rdma_request_free_data(rdma_req, rtransport); 1857 rdma_req->req.length = 0; 1858 rdma_req->req.iovcnt = 0; 1859 rdma_req->req.data = NULL; 1860 rdma_req->rsp.wr.next = NULL; 1861 rdma_req->data.wr.next = NULL; 1862 rdma_req->offset = 0; 1863 rdma_req->req.dif_enabled = false; 1864 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1865 rqpair->qd--; 1866 1867 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1868 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1869 } 1870 1871 bool 1872 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1873 struct spdk_nvmf_rdma_request *rdma_req) 1874 { 1875 struct spdk_nvmf_rdma_qpair *rqpair; 1876 struct spdk_nvmf_rdma_device *device; 1877 struct spdk_nvmf_rdma_poll_group *rgroup; 1878 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1879 int rc; 1880 struct spdk_nvmf_rdma_recv *rdma_recv; 1881 enum spdk_nvmf_rdma_request_state prev_state; 1882 bool progress = false; 1883 int data_posted; 1884 uint32_t num_blocks; 1885 1886 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1887 device = rqpair->device; 1888 rgroup = rqpair->poller->group; 1889 1890 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1891 1892 /* If the queue pair is in an error state, force the request to the completed state 1893 * to release resources. */ 1894 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1895 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1896 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1897 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1898 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1899 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1900 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1901 } 1902 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1903 } 1904 1905 /* The loop here is to allow for several back-to-back state changes. */ 1906 do { 1907 prev_state = rdma_req->state; 1908 1909 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 1910 1911 switch (rdma_req->state) { 1912 case RDMA_REQUEST_STATE_FREE: 1913 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1914 * to escape this state. */ 1915 break; 1916 case RDMA_REQUEST_STATE_NEW: 1917 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1918 (uintptr_t)rdma_req, (uintptr_t)rqpair); 1919 rdma_recv = rdma_req->recv; 1920 1921 /* The first element of the SGL is the NVMe command */ 1922 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1923 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1924 1925 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1926 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1927 break; 1928 } 1929 1930 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 1931 rdma_req->req.dif_enabled = true; 1932 } 1933 1934 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1935 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1936 rdma_req->rsp.wr.imm_data = 0; 1937 #endif 1938 1939 /* The next state transition depends on the data transfer needs of this request. */ 1940 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 1941 1942 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 1943 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 1944 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 1945 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1946 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 1947 break; 1948 } 1949 1950 /* If no data to transfer, ready to execute. */ 1951 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1952 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1953 break; 1954 } 1955 1956 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1957 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 1958 break; 1959 case RDMA_REQUEST_STATE_NEED_BUFFER: 1960 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1961 (uintptr_t)rdma_req, (uintptr_t)rqpair); 1962 1963 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1964 1965 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 1966 /* This request needs to wait in line to obtain a buffer */ 1967 break; 1968 } 1969 1970 /* Try to get a data buffer */ 1971 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1972 if (rc < 0) { 1973 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 1974 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1975 break; 1976 } 1977 1978 if (!rdma_req->req.data) { 1979 /* No buffers available. */ 1980 rgroup->stat.pending_data_buffer++; 1981 break; 1982 } 1983 1984 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 1985 1986 /* If data is transferring from host to controller and the data didn't 1987 * arrive using in capsule data, we need to do a transfer from the host. 1988 */ 1989 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 1990 rdma_req->req.data_from_pool) { 1991 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1992 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1993 break; 1994 } 1995 1996 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1997 break; 1998 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1999 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2000 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2001 2002 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2003 /* This request needs to wait in line to perform RDMA */ 2004 break; 2005 } 2006 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2007 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2008 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2009 rqpair->poller->stat.pending_rdma_read++; 2010 break; 2011 } 2012 2013 /* We have already verified that this request is the head of the queue. */ 2014 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2015 2016 rc = request_transfer_in(&rdma_req->req); 2017 if (!rc) { 2018 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2019 } else { 2020 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2021 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2022 } 2023 break; 2024 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2025 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2026 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2027 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2028 * to escape this state. */ 2029 break; 2030 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2031 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2032 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2033 2034 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2035 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2036 /* generate DIF for write operation */ 2037 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2038 assert(num_blocks > 0); 2039 2040 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2041 num_blocks, &rdma_req->req.dif.dif_ctx); 2042 if (rc != 0) { 2043 SPDK_ERRLOG("DIF generation failed\n"); 2044 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2045 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2046 break; 2047 } 2048 } 2049 2050 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2051 /* set extended length before IO operation */ 2052 rdma_req->req.length = rdma_req->req.dif.elba_length; 2053 } 2054 2055 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2056 spdk_nvmf_request_exec(&rdma_req->req); 2057 break; 2058 case RDMA_REQUEST_STATE_EXECUTING: 2059 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2060 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2061 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2062 * to escape this state. */ 2063 break; 2064 case RDMA_REQUEST_STATE_EXECUTED: 2065 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2066 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2067 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2068 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2069 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2070 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2071 } else { 2072 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2073 } 2074 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2075 /* restore the original length */ 2076 rdma_req->req.length = rdma_req->req.dif.orig_length; 2077 2078 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2079 struct spdk_dif_error error_blk; 2080 2081 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2082 2083 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2084 &rdma_req->req.dif.dif_ctx, &error_blk); 2085 if (rc) { 2086 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2087 2088 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2089 error_blk.err_offset); 2090 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2091 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2092 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2093 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2094 } 2095 } 2096 } 2097 break; 2098 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2099 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2100 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2101 2102 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2103 /* This request needs to wait in line to perform RDMA */ 2104 break; 2105 } 2106 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2107 rqpair->max_send_depth) { 2108 /* We can only have so many WRs outstanding. we have to wait until some finish. 2109 * +1 since each request has an additional wr in the resp. */ 2110 rqpair->poller->stat.pending_rdma_write++; 2111 break; 2112 } 2113 2114 /* We have already verified that this request is the head of the queue. */ 2115 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2116 2117 /* The data transfer will be kicked off from 2118 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2119 */ 2120 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2121 break; 2122 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2123 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2124 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2125 rc = request_transfer_out(&rdma_req->req, &data_posted); 2126 assert(rc == 0); /* No good way to handle this currently */ 2127 if (rc) { 2128 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2129 } else { 2130 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2131 RDMA_REQUEST_STATE_COMPLETING; 2132 } 2133 break; 2134 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2135 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2136 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2137 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2138 * to escape this state. */ 2139 break; 2140 case RDMA_REQUEST_STATE_COMPLETING: 2141 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2142 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2143 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2144 * to escape this state. */ 2145 break; 2146 case RDMA_REQUEST_STATE_COMPLETED: 2147 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2148 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2149 2150 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2151 _nvmf_rdma_request_free(rdma_req, rtransport); 2152 break; 2153 case RDMA_REQUEST_NUM_STATES: 2154 default: 2155 assert(0); 2156 break; 2157 } 2158 2159 if (rdma_req->state != prev_state) { 2160 progress = true; 2161 } 2162 } while (rdma_req->state != prev_state); 2163 2164 return progress; 2165 } 2166 2167 /* Public API callbacks begin here */ 2168 2169 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2170 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2171 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2172 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2173 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2174 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2175 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2176 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2177 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2178 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2179 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2180 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2181 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2182 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2183 2184 static void 2185 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2186 { 2187 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2188 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2189 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2190 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2191 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2192 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2193 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2194 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2195 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2196 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2197 opts->transport_specific = NULL; 2198 } 2199 2200 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2201 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2202 2203 static inline bool 2204 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2205 { 2206 return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD || 2207 device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW; 2208 } 2209 2210 static int 2211 nvmf_rdma_accept(void *ctx); 2212 2213 static struct spdk_nvmf_transport * 2214 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2215 { 2216 int rc; 2217 struct spdk_nvmf_rdma_transport *rtransport; 2218 struct spdk_nvmf_rdma_device *device, *tmp; 2219 struct ibv_context **contexts; 2220 uint32_t i; 2221 int flag; 2222 uint32_t sge_count; 2223 uint32_t min_shared_buffers; 2224 uint32_t min_in_capsule_data_size; 2225 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2226 pthread_mutexattr_t attr; 2227 2228 rtransport = calloc(1, sizeof(*rtransport)); 2229 if (!rtransport) { 2230 return NULL; 2231 } 2232 2233 if (pthread_mutexattr_init(&attr)) { 2234 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2235 free(rtransport); 2236 return NULL; 2237 } 2238 2239 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2240 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2241 pthread_mutexattr_destroy(&attr); 2242 free(rtransport); 2243 return NULL; 2244 } 2245 2246 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2247 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2248 pthread_mutexattr_destroy(&attr); 2249 free(rtransport); 2250 return NULL; 2251 } 2252 2253 pthread_mutexattr_destroy(&attr); 2254 2255 TAILQ_INIT(&rtransport->devices); 2256 TAILQ_INIT(&rtransport->ports); 2257 TAILQ_INIT(&rtransport->poll_groups); 2258 2259 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2260 rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2261 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2262 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2263 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2264 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2265 if (opts->transport_specific != NULL && 2266 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2267 SPDK_COUNTOF(rdma_transport_opts_decoder), 2268 &rtransport->rdma_opts)) { 2269 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2270 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2271 return NULL; 2272 } 2273 2274 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2275 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2276 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2277 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2278 " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d," 2279 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2280 opts->max_queue_depth, 2281 opts->max_io_size, 2282 opts->max_qpairs_per_ctrlr - 1, 2283 opts->io_unit_size, 2284 opts->in_capsule_data_size, 2285 opts->max_aq_depth, 2286 opts->num_shared_buffers, 2287 rtransport->rdma_opts.num_cqe, 2288 rtransport->rdma_opts.max_srq_depth, 2289 rtransport->rdma_opts.no_srq, 2290 rtransport->rdma_opts.acceptor_backlog, 2291 rtransport->rdma_opts.no_wr_batching, 2292 opts->abort_timeout_sec); 2293 2294 /* I/O unit size cannot be larger than max I/O size */ 2295 if (opts->io_unit_size > opts->max_io_size) { 2296 opts->io_unit_size = opts->max_io_size; 2297 } 2298 2299 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2300 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2301 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2302 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2303 } 2304 2305 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2306 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2307 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2308 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2309 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2310 return NULL; 2311 } 2312 2313 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2314 if (min_shared_buffers > opts->num_shared_buffers) { 2315 SPDK_ERRLOG("There are not enough buffers to satisfy" 2316 "per-poll group caches for each thread. (%" PRIu32 ")" 2317 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2318 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2319 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2320 return NULL; 2321 } 2322 2323 sge_count = opts->max_io_size / opts->io_unit_size; 2324 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2325 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2326 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2327 return NULL; 2328 } 2329 2330 min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES; 2331 if (opts->in_capsule_data_size < min_in_capsule_data_size) { 2332 SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n", 2333 min_in_capsule_data_size); 2334 opts->in_capsule_data_size = min_in_capsule_data_size; 2335 } 2336 2337 rtransport->event_channel = rdma_create_event_channel(); 2338 if (rtransport->event_channel == NULL) { 2339 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2340 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2341 return NULL; 2342 } 2343 2344 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2345 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2346 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2347 rtransport->event_channel->fd, spdk_strerror(errno)); 2348 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2349 return NULL; 2350 } 2351 2352 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2353 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2354 sizeof(struct spdk_nvmf_rdma_request_data), 2355 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2356 SPDK_ENV_SOCKET_ID_ANY); 2357 if (!rtransport->data_wr_pool) { 2358 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2359 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2360 return NULL; 2361 } 2362 2363 contexts = rdma_get_devices(NULL); 2364 if (contexts == NULL) { 2365 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2366 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2367 return NULL; 2368 } 2369 2370 i = 0; 2371 rc = 0; 2372 while (contexts[i] != NULL) { 2373 device = calloc(1, sizeof(*device)); 2374 if (!device) { 2375 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2376 rc = -ENOMEM; 2377 break; 2378 } 2379 device->context = contexts[i]; 2380 rc = ibv_query_device(device->context, &device->attr); 2381 if (rc < 0) { 2382 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2383 free(device); 2384 break; 2385 2386 } 2387 2388 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2389 2390 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2391 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2392 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2393 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2394 } 2395 2396 /** 2397 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2398 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2399 * but incorrectly reports that it does. There are changes making their way 2400 * through the kernel now that will enable this feature. When they are merged, 2401 * we can conditionally enable this feature. 2402 * 2403 * TODO: enable this for versions of the kernel rxe driver that support it. 2404 */ 2405 if (nvmf_rdma_is_rxe_device(device)) { 2406 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2407 } 2408 #endif 2409 2410 /* set up device context async ev fd as NON_BLOCKING */ 2411 flag = fcntl(device->context->async_fd, F_GETFL); 2412 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2413 if (rc < 0) { 2414 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2415 free(device); 2416 break; 2417 } 2418 2419 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2420 i++; 2421 2422 if (g_nvmf_hooks.get_ibv_pd) { 2423 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2424 } else { 2425 device->pd = ibv_alloc_pd(device->context); 2426 } 2427 2428 if (!device->pd) { 2429 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2430 rc = -ENOMEM; 2431 break; 2432 } 2433 2434 assert(device->map == NULL); 2435 2436 device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks, SPDK_RDMA_MEMORY_MAP_ROLE_TARGET); 2437 if (!device->map) { 2438 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2439 rc = -ENOMEM; 2440 break; 2441 } 2442 2443 assert(device->map != NULL); 2444 assert(device->pd != NULL); 2445 } 2446 rdma_free_devices(contexts); 2447 2448 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2449 /* divide and round up. */ 2450 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2451 2452 /* round up to the nearest 4k. */ 2453 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2454 2455 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2456 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2457 opts->io_unit_size); 2458 } 2459 2460 if (rc < 0) { 2461 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2462 return NULL; 2463 } 2464 2465 /* Set up poll descriptor array to monitor events from RDMA and IB 2466 * in a single poll syscall 2467 */ 2468 rtransport->npoll_fds = i + 1; 2469 i = 0; 2470 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2471 if (rtransport->poll_fds == NULL) { 2472 SPDK_ERRLOG("poll_fds allocation failed\n"); 2473 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2474 return NULL; 2475 } 2476 2477 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2478 rtransport->poll_fds[i++].events = POLLIN; 2479 2480 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2481 rtransport->poll_fds[i].fd = device->context->async_fd; 2482 rtransport->poll_fds[i++].events = POLLIN; 2483 } 2484 2485 rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport, 2486 opts->acceptor_poll_rate); 2487 if (!rtransport->accept_poller) { 2488 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2489 return NULL; 2490 } 2491 2492 return &rtransport->transport; 2493 } 2494 2495 static void 2496 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2497 { 2498 struct spdk_nvmf_rdma_transport *rtransport; 2499 assert(w != NULL); 2500 2501 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2502 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2503 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2504 if (rtransport->rdma_opts.no_srq == true) { 2505 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2506 } 2507 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2508 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2509 } 2510 2511 static int 2512 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2513 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2514 { 2515 struct spdk_nvmf_rdma_transport *rtransport; 2516 struct spdk_nvmf_rdma_port *port, *port_tmp; 2517 struct spdk_nvmf_rdma_device *device, *device_tmp; 2518 2519 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2520 2521 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2522 TAILQ_REMOVE(&rtransport->ports, port, link); 2523 rdma_destroy_id(port->id); 2524 free(port); 2525 } 2526 2527 if (rtransport->poll_fds != NULL) { 2528 free(rtransport->poll_fds); 2529 } 2530 2531 if (rtransport->event_channel != NULL) { 2532 rdma_destroy_event_channel(rtransport->event_channel); 2533 } 2534 2535 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2536 TAILQ_REMOVE(&rtransport->devices, device, link); 2537 spdk_rdma_free_mem_map(&device->map); 2538 if (device->pd) { 2539 if (!g_nvmf_hooks.get_ibv_pd) { 2540 ibv_dealloc_pd(device->pd); 2541 } 2542 } 2543 free(device); 2544 } 2545 2546 if (rtransport->data_wr_pool != NULL) { 2547 if (spdk_mempool_count(rtransport->data_wr_pool) != 2548 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2549 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2550 spdk_mempool_count(rtransport->data_wr_pool), 2551 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2552 } 2553 } 2554 2555 spdk_mempool_free(rtransport->data_wr_pool); 2556 2557 spdk_poller_unregister(&rtransport->accept_poller); 2558 pthread_mutex_destroy(&rtransport->lock); 2559 free(rtransport); 2560 2561 if (cb_fn) { 2562 cb_fn(cb_arg); 2563 } 2564 return 0; 2565 } 2566 2567 static int 2568 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2569 struct spdk_nvme_transport_id *trid, 2570 bool peer); 2571 2572 static int 2573 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2574 struct spdk_nvmf_listen_opts *listen_opts) 2575 { 2576 struct spdk_nvmf_rdma_transport *rtransport; 2577 struct spdk_nvmf_rdma_device *device; 2578 struct spdk_nvmf_rdma_port *port; 2579 struct addrinfo *res; 2580 struct addrinfo hints; 2581 int family; 2582 int rc; 2583 2584 if (!strlen(trid->trsvcid)) { 2585 SPDK_ERRLOG("Service id is required\n"); 2586 return -EINVAL; 2587 } 2588 2589 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2590 assert(rtransport->event_channel != NULL); 2591 2592 pthread_mutex_lock(&rtransport->lock); 2593 port = calloc(1, sizeof(*port)); 2594 if (!port) { 2595 SPDK_ERRLOG("Port allocation failed\n"); 2596 pthread_mutex_unlock(&rtransport->lock); 2597 return -ENOMEM; 2598 } 2599 2600 port->trid = trid; 2601 2602 switch (trid->adrfam) { 2603 case SPDK_NVMF_ADRFAM_IPV4: 2604 family = AF_INET; 2605 break; 2606 case SPDK_NVMF_ADRFAM_IPV6: 2607 family = AF_INET6; 2608 break; 2609 default: 2610 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2611 free(port); 2612 pthread_mutex_unlock(&rtransport->lock); 2613 return -EINVAL; 2614 } 2615 2616 memset(&hints, 0, sizeof(hints)); 2617 hints.ai_family = family; 2618 hints.ai_flags = AI_NUMERICSERV; 2619 hints.ai_socktype = SOCK_STREAM; 2620 hints.ai_protocol = 0; 2621 2622 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2623 if (rc) { 2624 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2625 free(port); 2626 pthread_mutex_unlock(&rtransport->lock); 2627 return -EINVAL; 2628 } 2629 2630 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2631 if (rc < 0) { 2632 SPDK_ERRLOG("rdma_create_id() failed\n"); 2633 freeaddrinfo(res); 2634 free(port); 2635 pthread_mutex_unlock(&rtransport->lock); 2636 return rc; 2637 } 2638 2639 rc = rdma_bind_addr(port->id, res->ai_addr); 2640 freeaddrinfo(res); 2641 2642 if (rc < 0) { 2643 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2644 rdma_destroy_id(port->id); 2645 free(port); 2646 pthread_mutex_unlock(&rtransport->lock); 2647 return rc; 2648 } 2649 2650 if (!port->id->verbs) { 2651 SPDK_ERRLOG("ibv_context is null\n"); 2652 rdma_destroy_id(port->id); 2653 free(port); 2654 pthread_mutex_unlock(&rtransport->lock); 2655 return -1; 2656 } 2657 2658 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2659 if (rc < 0) { 2660 SPDK_ERRLOG("rdma_listen() failed\n"); 2661 rdma_destroy_id(port->id); 2662 free(port); 2663 pthread_mutex_unlock(&rtransport->lock); 2664 return rc; 2665 } 2666 2667 TAILQ_FOREACH(device, &rtransport->devices, link) { 2668 if (device->context == port->id->verbs) { 2669 port->device = device; 2670 break; 2671 } 2672 } 2673 if (!port->device) { 2674 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2675 port->id->verbs); 2676 rdma_destroy_id(port->id); 2677 free(port); 2678 pthread_mutex_unlock(&rtransport->lock); 2679 return -EINVAL; 2680 } 2681 2682 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2683 trid->traddr, trid->trsvcid); 2684 2685 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2686 pthread_mutex_unlock(&rtransport->lock); 2687 return 0; 2688 } 2689 2690 static void 2691 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2692 const struct spdk_nvme_transport_id *trid) 2693 { 2694 struct spdk_nvmf_rdma_transport *rtransport; 2695 struct spdk_nvmf_rdma_port *port, *tmp; 2696 2697 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2698 2699 pthread_mutex_lock(&rtransport->lock); 2700 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2701 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2702 TAILQ_REMOVE(&rtransport->ports, port, link); 2703 rdma_destroy_id(port->id); 2704 free(port); 2705 break; 2706 } 2707 } 2708 2709 pthread_mutex_unlock(&rtransport->lock); 2710 } 2711 2712 static void 2713 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2714 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2715 { 2716 struct spdk_nvmf_request *req, *tmp; 2717 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2718 struct spdk_nvmf_rdma_resources *resources; 2719 2720 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2721 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2722 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2723 break; 2724 } 2725 } 2726 2727 /* Then RDMA writes since reads have stronger restrictions than writes */ 2728 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2729 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2730 break; 2731 } 2732 } 2733 2734 /* Then we handle request waiting on memory buffers. */ 2735 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2736 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2737 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2738 break; 2739 } 2740 } 2741 2742 resources = rqpair->resources; 2743 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2744 rdma_req = STAILQ_FIRST(&resources->free_queue); 2745 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2746 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2747 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2748 2749 if (rqpair->srq != NULL) { 2750 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2751 rdma_req->recv->qpair->qd++; 2752 } else { 2753 rqpair->qd++; 2754 } 2755 2756 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2757 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2758 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2759 break; 2760 } 2761 } 2762 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2763 rqpair->poller->stat.pending_free_request++; 2764 } 2765 } 2766 2767 static inline bool 2768 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 2769 { 2770 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 2771 return nvmf_rdma_is_rxe_device(device) || 2772 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 2773 } 2774 2775 static void 2776 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2777 { 2778 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2779 struct spdk_nvmf_rdma_transport, transport); 2780 2781 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2782 2783 /* nvmr_rdma_close_qpair is not called */ 2784 if (!rqpair->to_close) { 2785 return; 2786 } 2787 2788 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2789 if (rqpair->current_send_depth != 0) { 2790 return; 2791 } 2792 2793 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2794 return; 2795 } 2796 2797 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 2798 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 2799 return; 2800 } 2801 2802 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2803 2804 nvmf_rdma_qpair_destroy(rqpair); 2805 } 2806 2807 static int 2808 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2809 { 2810 struct spdk_nvmf_qpair *qpair; 2811 struct spdk_nvmf_rdma_qpair *rqpair; 2812 2813 if (evt->id == NULL) { 2814 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2815 return -1; 2816 } 2817 2818 qpair = evt->id->context; 2819 if (qpair == NULL) { 2820 SPDK_ERRLOG("disconnect request: no active connection\n"); 2821 return -1; 2822 } 2823 2824 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2825 2826 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair); 2827 2828 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2829 2830 return 0; 2831 } 2832 2833 #ifdef DEBUG 2834 static const char *CM_EVENT_STR[] = { 2835 "RDMA_CM_EVENT_ADDR_RESOLVED", 2836 "RDMA_CM_EVENT_ADDR_ERROR", 2837 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2838 "RDMA_CM_EVENT_ROUTE_ERROR", 2839 "RDMA_CM_EVENT_CONNECT_REQUEST", 2840 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2841 "RDMA_CM_EVENT_CONNECT_ERROR", 2842 "RDMA_CM_EVENT_UNREACHABLE", 2843 "RDMA_CM_EVENT_REJECTED", 2844 "RDMA_CM_EVENT_ESTABLISHED", 2845 "RDMA_CM_EVENT_DISCONNECTED", 2846 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2847 "RDMA_CM_EVENT_MULTICAST_JOIN", 2848 "RDMA_CM_EVENT_MULTICAST_ERROR", 2849 "RDMA_CM_EVENT_ADDR_CHANGE", 2850 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2851 }; 2852 #endif /* DEBUG */ 2853 2854 static void 2855 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2856 struct spdk_nvmf_rdma_port *port) 2857 { 2858 struct spdk_nvmf_rdma_poll_group *rgroup; 2859 struct spdk_nvmf_rdma_poller *rpoller; 2860 struct spdk_nvmf_rdma_qpair *rqpair; 2861 2862 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2863 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2864 RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) { 2865 if (rqpair->listen_id == port->id) { 2866 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2867 } 2868 } 2869 } 2870 } 2871 } 2872 2873 static bool 2874 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2875 struct rdma_cm_event *event) 2876 { 2877 const struct spdk_nvme_transport_id *trid; 2878 struct spdk_nvmf_rdma_port *port; 2879 struct spdk_nvmf_rdma_transport *rtransport; 2880 bool event_acked = false; 2881 2882 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2883 TAILQ_FOREACH(port, &rtransport->ports, link) { 2884 if (port->id == event->id) { 2885 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2886 rdma_ack_cm_event(event); 2887 event_acked = true; 2888 trid = port->trid; 2889 break; 2890 } 2891 } 2892 2893 if (event_acked) { 2894 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2895 2896 nvmf_rdma_stop_listen(transport, trid); 2897 nvmf_rdma_listen(transport, trid, NULL); 2898 } 2899 2900 return event_acked; 2901 } 2902 2903 static void 2904 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2905 struct rdma_cm_event *event) 2906 { 2907 struct spdk_nvmf_rdma_port *port; 2908 struct spdk_nvmf_rdma_transport *rtransport; 2909 2910 port = event->id->context; 2911 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2912 2913 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2914 2915 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2916 2917 rdma_ack_cm_event(event); 2918 2919 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2920 ; 2921 } 2922 } 2923 2924 static void 2925 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2926 { 2927 struct spdk_nvmf_rdma_transport *rtransport; 2928 struct rdma_cm_event *event; 2929 int rc; 2930 bool event_acked; 2931 2932 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2933 2934 if (rtransport->event_channel == NULL) { 2935 return; 2936 } 2937 2938 while (1) { 2939 event_acked = false; 2940 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2941 if (rc) { 2942 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2943 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2944 } 2945 break; 2946 } 2947 2948 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2949 2950 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2951 2952 switch (event->event) { 2953 case RDMA_CM_EVENT_ADDR_RESOLVED: 2954 case RDMA_CM_EVENT_ADDR_ERROR: 2955 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2956 case RDMA_CM_EVENT_ROUTE_ERROR: 2957 /* No action required. The target never attempts to resolve routes. */ 2958 break; 2959 case RDMA_CM_EVENT_CONNECT_REQUEST: 2960 rc = nvmf_rdma_connect(transport, event); 2961 if (rc < 0) { 2962 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2963 break; 2964 } 2965 break; 2966 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2967 /* The target never initiates a new connection. So this will not occur. */ 2968 break; 2969 case RDMA_CM_EVENT_CONNECT_ERROR: 2970 /* Can this happen? The docs say it can, but not sure what causes it. */ 2971 break; 2972 case RDMA_CM_EVENT_UNREACHABLE: 2973 case RDMA_CM_EVENT_REJECTED: 2974 /* These only occur on the client side. */ 2975 break; 2976 case RDMA_CM_EVENT_ESTABLISHED: 2977 /* TODO: Should we be waiting for this event anywhere? */ 2978 break; 2979 case RDMA_CM_EVENT_DISCONNECTED: 2980 rc = nvmf_rdma_disconnect(event); 2981 if (rc < 0) { 2982 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2983 break; 2984 } 2985 break; 2986 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2987 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 2988 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 2989 * Once these events are sent to SPDK, we should release all IB resources and 2990 * don't make attempts to call any ibv_query/modify/create functions. We can only call 2991 * ibv_destroy* functions to release user space memory allocated by IB. All kernel 2992 * resources are already cleaned. */ 2993 if (event->id->qp) { 2994 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 2995 * corresponding qpair. Otherwise the event refers to a listening device */ 2996 rc = nvmf_rdma_disconnect(event); 2997 if (rc < 0) { 2998 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2999 break; 3000 } 3001 } else { 3002 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3003 event_acked = true; 3004 } 3005 break; 3006 case RDMA_CM_EVENT_MULTICAST_JOIN: 3007 case RDMA_CM_EVENT_MULTICAST_ERROR: 3008 /* Multicast is not used */ 3009 break; 3010 case RDMA_CM_EVENT_ADDR_CHANGE: 3011 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3012 break; 3013 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3014 /* For now, do nothing. The target never re-uses queue pairs. */ 3015 break; 3016 default: 3017 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3018 break; 3019 } 3020 if (!event_acked) { 3021 rdma_ack_cm_event(event); 3022 } 3023 } 3024 } 3025 3026 static void 3027 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3028 { 3029 rqpair->last_wqe_reached = true; 3030 nvmf_rdma_destroy_drained_qpair(rqpair); 3031 } 3032 3033 static void 3034 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3035 { 3036 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3037 3038 if (event_ctx->rqpair) { 3039 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3040 if (event_ctx->cb_fn) { 3041 event_ctx->cb_fn(event_ctx->rqpair); 3042 } 3043 } 3044 free(event_ctx); 3045 } 3046 3047 static int 3048 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3049 spdk_nvmf_rdma_qpair_ibv_event fn) 3050 { 3051 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3052 struct spdk_thread *thr = NULL; 3053 int rc; 3054 3055 if (rqpair->qpair.group) { 3056 thr = rqpair->qpair.group->thread; 3057 } else if (rqpair->destruct_channel) { 3058 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3059 } 3060 3061 if (!thr) { 3062 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3063 return -EINVAL; 3064 } 3065 3066 ctx = calloc(1, sizeof(*ctx)); 3067 if (!ctx) { 3068 return -ENOMEM; 3069 } 3070 3071 ctx->rqpair = rqpair; 3072 ctx->cb_fn = fn; 3073 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3074 3075 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3076 if (rc) { 3077 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3078 free(ctx); 3079 } 3080 3081 return rc; 3082 } 3083 3084 static int 3085 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3086 { 3087 int rc; 3088 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3089 struct ibv_async_event event; 3090 3091 rc = ibv_get_async_event(device->context, &event); 3092 3093 if (rc) { 3094 /* In non-blocking mode -1 means there are no events available */ 3095 return rc; 3096 } 3097 3098 switch (event.event_type) { 3099 case IBV_EVENT_QP_FATAL: 3100 rqpair = event.element.qp->qp_context; 3101 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3102 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3103 (uintptr_t)rqpair, event.event_type); 3104 nvmf_rdma_update_ibv_state(rqpair); 3105 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3106 break; 3107 case IBV_EVENT_QP_LAST_WQE_REACHED: 3108 /* This event only occurs for shared receive queues. */ 3109 rqpair = event.element.qp->qp_context; 3110 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3111 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3112 if (rc) { 3113 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3114 rqpair->last_wqe_reached = true; 3115 } 3116 break; 3117 case IBV_EVENT_SQ_DRAINED: 3118 /* This event occurs frequently in both error and non-error states. 3119 * Check if the qpair is in an error state before sending a message. */ 3120 rqpair = event.element.qp->qp_context; 3121 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3122 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3123 (uintptr_t)rqpair, event.event_type); 3124 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3125 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3126 } 3127 break; 3128 case IBV_EVENT_QP_REQ_ERR: 3129 case IBV_EVENT_QP_ACCESS_ERR: 3130 case IBV_EVENT_COMM_EST: 3131 case IBV_EVENT_PATH_MIG: 3132 case IBV_EVENT_PATH_MIG_ERR: 3133 SPDK_NOTICELOG("Async event: %s\n", 3134 ibv_event_type_str(event.event_type)); 3135 rqpair = event.element.qp->qp_context; 3136 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3137 (uintptr_t)rqpair, event.event_type); 3138 nvmf_rdma_update_ibv_state(rqpair); 3139 break; 3140 case IBV_EVENT_CQ_ERR: 3141 case IBV_EVENT_DEVICE_FATAL: 3142 case IBV_EVENT_PORT_ACTIVE: 3143 case IBV_EVENT_PORT_ERR: 3144 case IBV_EVENT_LID_CHANGE: 3145 case IBV_EVENT_PKEY_CHANGE: 3146 case IBV_EVENT_SM_CHANGE: 3147 case IBV_EVENT_SRQ_ERR: 3148 case IBV_EVENT_SRQ_LIMIT_REACHED: 3149 case IBV_EVENT_CLIENT_REREGISTER: 3150 case IBV_EVENT_GID_CHANGE: 3151 default: 3152 SPDK_NOTICELOG("Async event: %s\n", 3153 ibv_event_type_str(event.event_type)); 3154 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3155 break; 3156 } 3157 ibv_ack_async_event(&event); 3158 3159 return 0; 3160 } 3161 3162 static void 3163 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3164 { 3165 int rc = 0; 3166 uint32_t i = 0; 3167 3168 for (i = 0; i < max_events; i++) { 3169 rc = nvmf_process_ib_event(device); 3170 if (rc) { 3171 break; 3172 } 3173 } 3174 3175 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3176 } 3177 3178 static int 3179 nvmf_rdma_accept(void *ctx) 3180 { 3181 int nfds, i = 0; 3182 struct spdk_nvmf_transport *transport = ctx; 3183 struct spdk_nvmf_rdma_transport *rtransport; 3184 struct spdk_nvmf_rdma_device *device, *tmp; 3185 uint32_t count; 3186 3187 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3188 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3189 3190 if (nfds <= 0) { 3191 return SPDK_POLLER_IDLE; 3192 } 3193 3194 /* The first poll descriptor is RDMA CM event */ 3195 if (rtransport->poll_fds[i++].revents & POLLIN) { 3196 nvmf_process_cm_event(transport); 3197 nfds--; 3198 } 3199 3200 if (nfds == 0) { 3201 return SPDK_POLLER_BUSY; 3202 } 3203 3204 /* Second and subsequent poll descriptors are IB async events */ 3205 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3206 if (rtransport->poll_fds[i++].revents & POLLIN) { 3207 nvmf_process_ib_events(device, 32); 3208 nfds--; 3209 } 3210 } 3211 /* check all flagged fd's have been served */ 3212 assert(nfds == 0); 3213 3214 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3215 } 3216 3217 static void 3218 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3219 struct spdk_nvmf_ctrlr_data *cdata) 3220 { 3221 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3222 3223 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3224 since in-capsule data only works with NVME drives that support SGL memory layout */ 3225 if (transport->opts.dif_insert_or_strip) { 3226 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3227 } 3228 3229 if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) { 3230 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule" 3231 " data is greater than 4KiB.\n"); 3232 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux " 3233 "kernel between versions 5.4 and 5.12 data corruption may occur for " 3234 "writes that are not a multiple of 4KiB in size.\n"); 3235 } 3236 } 3237 3238 static void 3239 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3240 struct spdk_nvme_transport_id *trid, 3241 struct spdk_nvmf_discovery_log_page_entry *entry) 3242 { 3243 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3244 entry->adrfam = trid->adrfam; 3245 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3246 3247 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3248 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3249 3250 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3251 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3252 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3253 } 3254 3255 static void 3256 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3257 3258 static struct spdk_nvmf_transport_poll_group * 3259 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport, 3260 struct spdk_nvmf_poll_group *group) 3261 { 3262 struct spdk_nvmf_rdma_transport *rtransport; 3263 struct spdk_nvmf_rdma_poll_group *rgroup; 3264 struct spdk_nvmf_rdma_poller *poller; 3265 struct spdk_nvmf_rdma_device *device; 3266 struct spdk_rdma_srq_init_attr srq_init_attr; 3267 struct spdk_nvmf_rdma_resource_opts opts; 3268 int num_cqe; 3269 3270 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3271 3272 rgroup = calloc(1, sizeof(*rgroup)); 3273 if (!rgroup) { 3274 return NULL; 3275 } 3276 3277 TAILQ_INIT(&rgroup->pollers); 3278 3279 pthread_mutex_lock(&rtransport->lock); 3280 TAILQ_FOREACH(device, &rtransport->devices, link) { 3281 poller = calloc(1, sizeof(*poller)); 3282 if (!poller) { 3283 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3284 nvmf_rdma_poll_group_destroy(&rgroup->group); 3285 pthread_mutex_unlock(&rtransport->lock); 3286 return NULL; 3287 } 3288 3289 poller->device = device; 3290 poller->group = rgroup; 3291 3292 RB_INIT(&poller->qpairs); 3293 STAILQ_INIT(&poller->qpairs_pending_send); 3294 STAILQ_INIT(&poller->qpairs_pending_recv); 3295 3296 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3297 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3298 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) { 3299 SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n", 3300 rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr); 3301 } 3302 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr); 3303 3304 device->num_srq++; 3305 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 3306 srq_init_attr.pd = device->pd; 3307 srq_init_attr.stats = &poller->stat.qp_stats.recv; 3308 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 3309 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3310 poller->srq = spdk_rdma_srq_create(&srq_init_attr); 3311 if (!poller->srq) { 3312 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3313 nvmf_rdma_poll_group_destroy(&rgroup->group); 3314 pthread_mutex_unlock(&rtransport->lock); 3315 return NULL; 3316 } 3317 3318 opts.qp = poller->srq; 3319 opts.pd = device->pd; 3320 opts.qpair = NULL; 3321 opts.shared = true; 3322 opts.max_queue_depth = poller->max_srq_depth; 3323 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3324 3325 poller->resources = nvmf_rdma_resources_create(&opts); 3326 if (!poller->resources) { 3327 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3328 nvmf_rdma_poll_group_destroy(&rgroup->group); 3329 pthread_mutex_unlock(&rtransport->lock); 3330 return NULL; 3331 } 3332 } 3333 3334 /* 3335 * When using an srq, we can limit the completion queue at startup. 3336 * The following formula represents the calculation: 3337 * num_cqe = num_recv + num_data_wr + num_send_wr. 3338 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3339 */ 3340 if (poller->srq) { 3341 num_cqe = poller->max_srq_depth * 3; 3342 } else { 3343 num_cqe = rtransport->rdma_opts.num_cqe; 3344 } 3345 3346 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3347 if (!poller->cq) { 3348 SPDK_ERRLOG("Unable to create completion queue\n"); 3349 nvmf_rdma_poll_group_destroy(&rgroup->group); 3350 pthread_mutex_unlock(&rtransport->lock); 3351 return NULL; 3352 } 3353 poller->num_cqe = num_cqe; 3354 } 3355 3356 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3357 if (rtransport->conn_sched.next_admin_pg == NULL) { 3358 rtransport->conn_sched.next_admin_pg = rgroup; 3359 rtransport->conn_sched.next_io_pg = rgroup; 3360 } 3361 3362 pthread_mutex_unlock(&rtransport->lock); 3363 return &rgroup->group; 3364 } 3365 3366 static struct spdk_nvmf_transport_poll_group * 3367 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3368 { 3369 struct spdk_nvmf_rdma_transport *rtransport; 3370 struct spdk_nvmf_rdma_poll_group **pg; 3371 struct spdk_nvmf_transport_poll_group *result; 3372 3373 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3374 3375 pthread_mutex_lock(&rtransport->lock); 3376 3377 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3378 pthread_mutex_unlock(&rtransport->lock); 3379 return NULL; 3380 } 3381 3382 if (qpair->qid == 0) { 3383 pg = &rtransport->conn_sched.next_admin_pg; 3384 } else { 3385 pg = &rtransport->conn_sched.next_io_pg; 3386 } 3387 3388 assert(*pg != NULL); 3389 3390 result = &(*pg)->group; 3391 3392 *pg = TAILQ_NEXT(*pg, link); 3393 if (*pg == NULL) { 3394 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3395 } 3396 3397 pthread_mutex_unlock(&rtransport->lock); 3398 3399 return result; 3400 } 3401 3402 static void 3403 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3404 { 3405 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3406 struct spdk_nvmf_rdma_poller *poller, *tmp; 3407 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3408 struct spdk_nvmf_rdma_transport *rtransport; 3409 3410 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3411 if (!rgroup) { 3412 return; 3413 } 3414 3415 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3416 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3417 3418 RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) { 3419 nvmf_rdma_qpair_destroy(qpair); 3420 } 3421 3422 if (poller->srq) { 3423 if (poller->resources) { 3424 nvmf_rdma_resources_destroy(poller->resources); 3425 } 3426 spdk_rdma_srq_destroy(poller->srq); 3427 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 3428 } 3429 3430 if (poller->cq) { 3431 ibv_destroy_cq(poller->cq); 3432 } 3433 3434 free(poller); 3435 } 3436 3437 if (rgroup->group.transport == NULL) { 3438 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3439 * calls this function directly in a failure path. */ 3440 free(rgroup); 3441 return; 3442 } 3443 3444 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3445 3446 pthread_mutex_lock(&rtransport->lock); 3447 next_rgroup = TAILQ_NEXT(rgroup, link); 3448 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3449 if (next_rgroup == NULL) { 3450 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3451 } 3452 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3453 rtransport->conn_sched.next_admin_pg = next_rgroup; 3454 } 3455 if (rtransport->conn_sched.next_io_pg == rgroup) { 3456 rtransport->conn_sched.next_io_pg = next_rgroup; 3457 } 3458 pthread_mutex_unlock(&rtransport->lock); 3459 3460 free(rgroup); 3461 } 3462 3463 static void 3464 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3465 { 3466 if (rqpair->cm_id != NULL) { 3467 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3468 } 3469 } 3470 3471 static int 3472 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3473 struct spdk_nvmf_qpair *qpair) 3474 { 3475 struct spdk_nvmf_rdma_poll_group *rgroup; 3476 struct spdk_nvmf_rdma_qpair *rqpair; 3477 struct spdk_nvmf_rdma_device *device; 3478 struct spdk_nvmf_rdma_poller *poller; 3479 int rc; 3480 3481 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3482 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3483 3484 device = rqpair->device; 3485 3486 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3487 if (poller->device == device) { 3488 break; 3489 } 3490 } 3491 3492 if (!poller) { 3493 SPDK_ERRLOG("No poller found for device.\n"); 3494 return -1; 3495 } 3496 3497 rqpair->poller = poller; 3498 rqpair->srq = rqpair->poller->srq; 3499 3500 rc = nvmf_rdma_qpair_initialize(qpair); 3501 if (rc < 0) { 3502 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3503 rqpair->poller = NULL; 3504 rqpair->srq = NULL; 3505 return -1; 3506 } 3507 3508 RB_INSERT(qpairs_tree, &poller->qpairs, rqpair); 3509 3510 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3511 if (rc) { 3512 /* Try to reject, but we probably can't */ 3513 nvmf_rdma_qpair_reject_connection(rqpair); 3514 return -1; 3515 } 3516 3517 nvmf_rdma_update_ibv_state(rqpair); 3518 3519 return 0; 3520 } 3521 3522 static int 3523 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3524 struct spdk_nvmf_qpair *qpair) 3525 { 3526 struct spdk_nvmf_rdma_qpair *rqpair; 3527 3528 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3529 assert(group->transport->tgt != NULL); 3530 3531 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3532 3533 if (!rqpair->destruct_channel) { 3534 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3535 return 0; 3536 } 3537 3538 /* Sanity check that we get io_channel on the correct thread */ 3539 if (qpair->group) { 3540 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3541 } 3542 3543 return 0; 3544 } 3545 3546 static int 3547 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3548 { 3549 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3550 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3551 struct spdk_nvmf_rdma_transport, transport); 3552 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3553 struct spdk_nvmf_rdma_qpair, qpair); 3554 3555 /* 3556 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3557 * needs to be returned to the shared receive queue or the poll group will eventually be 3558 * starved of RECV structures. 3559 */ 3560 if (rqpair->srq && rdma_req->recv) { 3561 int rc; 3562 struct ibv_recv_wr *bad_recv_wr; 3563 3564 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 3565 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 3566 if (rc) { 3567 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3568 } 3569 } 3570 3571 _nvmf_rdma_request_free(rdma_req, rtransport); 3572 return 0; 3573 } 3574 3575 static int 3576 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3577 { 3578 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3579 struct spdk_nvmf_rdma_transport, transport); 3580 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3581 struct spdk_nvmf_rdma_request, req); 3582 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3583 struct spdk_nvmf_rdma_qpair, qpair); 3584 3585 if (rqpair->ibv_state != IBV_QPS_ERR) { 3586 /* The connection is alive, so process the request as normal */ 3587 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3588 } else { 3589 /* The connection is dead. Move the request directly to the completed state. */ 3590 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3591 } 3592 3593 nvmf_rdma_request_process(rtransport, rdma_req); 3594 3595 return 0; 3596 } 3597 3598 static void 3599 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 3600 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3601 { 3602 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3603 3604 rqpair->to_close = true; 3605 3606 /* This happens only when the qpair is disconnected before 3607 * it is added to the poll group. Since there is no poll group, 3608 * the RDMA qp has not been initialized yet and the RDMA CM 3609 * event has not yet been acknowledged, so we need to reject it. 3610 */ 3611 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3612 nvmf_rdma_qpair_reject_connection(rqpair); 3613 nvmf_rdma_qpair_destroy(rqpair); 3614 return; 3615 } 3616 3617 if (rqpair->rdma_qp) { 3618 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3619 } 3620 3621 nvmf_rdma_destroy_drained_qpair(rqpair); 3622 3623 if (cb_fn) { 3624 cb_fn(cb_arg); 3625 } 3626 } 3627 3628 static struct spdk_nvmf_rdma_qpair * 3629 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3630 { 3631 struct spdk_nvmf_rdma_qpair find; 3632 3633 find.qp_num = wc->qp_num; 3634 3635 return RB_FIND(qpairs_tree, &rpoller->qpairs, &find); 3636 } 3637 3638 #ifdef DEBUG 3639 static int 3640 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3641 { 3642 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3643 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3644 } 3645 #endif 3646 3647 static void 3648 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3649 int rc) 3650 { 3651 struct spdk_nvmf_rdma_recv *rdma_recv; 3652 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3653 3654 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3655 while (bad_recv_wr != NULL) { 3656 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3657 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3658 3659 rdma_recv->qpair->current_recv_depth++; 3660 bad_recv_wr = bad_recv_wr->next; 3661 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3662 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3663 } 3664 } 3665 3666 static void 3667 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3668 { 3669 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3670 while (bad_recv_wr != NULL) { 3671 bad_recv_wr = bad_recv_wr->next; 3672 rqpair->current_recv_depth++; 3673 } 3674 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3675 } 3676 3677 static void 3678 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3679 struct spdk_nvmf_rdma_poller *rpoller) 3680 { 3681 struct spdk_nvmf_rdma_qpair *rqpair; 3682 struct ibv_recv_wr *bad_recv_wr; 3683 int rc; 3684 3685 if (rpoller->srq) { 3686 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 3687 if (rc) { 3688 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3689 } 3690 } else { 3691 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3692 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3693 rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 3694 if (rc) { 3695 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3696 } 3697 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3698 } 3699 } 3700 } 3701 3702 static void 3703 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3704 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3705 { 3706 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3707 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3708 3709 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3710 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3711 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3712 assert(rqpair->current_send_depth > 0); 3713 rqpair->current_send_depth--; 3714 switch (bad_rdma_wr->type) { 3715 case RDMA_WR_TYPE_DATA: 3716 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3717 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3718 assert(rqpair->current_read_depth > 0); 3719 rqpair->current_read_depth--; 3720 } 3721 break; 3722 case RDMA_WR_TYPE_SEND: 3723 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3724 break; 3725 default: 3726 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3727 prev_rdma_req = cur_rdma_req; 3728 continue; 3729 } 3730 3731 if (prev_rdma_req == cur_rdma_req) { 3732 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3733 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3734 continue; 3735 } 3736 3737 switch (cur_rdma_req->state) { 3738 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3739 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3740 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3741 break; 3742 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3743 case RDMA_REQUEST_STATE_COMPLETING: 3744 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3745 break; 3746 default: 3747 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3748 cur_rdma_req->state, rqpair); 3749 continue; 3750 } 3751 3752 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3753 prev_rdma_req = cur_rdma_req; 3754 } 3755 3756 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3757 /* Disconnect the connection. */ 3758 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3759 } 3760 3761 } 3762 3763 static void 3764 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3765 struct spdk_nvmf_rdma_poller *rpoller) 3766 { 3767 struct spdk_nvmf_rdma_qpair *rqpair; 3768 struct ibv_send_wr *bad_wr = NULL; 3769 int rc; 3770 3771 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3772 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3773 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3774 3775 /* bad wr always points to the first wr that failed. */ 3776 if (rc) { 3777 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3778 } 3779 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3780 } 3781 } 3782 3783 static const char * 3784 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 3785 { 3786 switch (wr_type) { 3787 case RDMA_WR_TYPE_RECV: 3788 return "RECV"; 3789 case RDMA_WR_TYPE_SEND: 3790 return "SEND"; 3791 case RDMA_WR_TYPE_DATA: 3792 return "DATA"; 3793 default: 3794 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 3795 SPDK_UNREACHABLE(); 3796 } 3797 } 3798 3799 static inline void 3800 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 3801 { 3802 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 3803 3804 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 3805 /* If qpair is in ERR state, we will receive completions for all posted and not completed 3806 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 3807 SPDK_DEBUGLOG(rdma, 3808 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3809 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3810 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3811 } else { 3812 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3813 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3814 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3815 } 3816 } 3817 3818 static int 3819 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3820 struct spdk_nvmf_rdma_poller *rpoller) 3821 { 3822 struct ibv_wc wc[32]; 3823 struct spdk_nvmf_rdma_wr *rdma_wr; 3824 struct spdk_nvmf_rdma_request *rdma_req; 3825 struct spdk_nvmf_rdma_recv *rdma_recv; 3826 struct spdk_nvmf_rdma_qpair *rqpair; 3827 int reaped, i; 3828 int count = 0; 3829 bool error = false; 3830 uint64_t poll_tsc = spdk_get_ticks(); 3831 3832 /* Poll for completing operations. */ 3833 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3834 if (reaped < 0) { 3835 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3836 errno, spdk_strerror(errno)); 3837 return -1; 3838 } else if (reaped == 0) { 3839 rpoller->stat.idle_polls++; 3840 } 3841 3842 rpoller->stat.polls++; 3843 rpoller->stat.completions += reaped; 3844 3845 for (i = 0; i < reaped; i++) { 3846 3847 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3848 3849 switch (rdma_wr->type) { 3850 case RDMA_WR_TYPE_SEND: 3851 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3852 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3853 3854 if (!wc[i].status) { 3855 count++; 3856 assert(wc[i].opcode == IBV_WC_SEND); 3857 assert(nvmf_rdma_req_is_completing(rdma_req)); 3858 } 3859 3860 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3861 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3862 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3863 rdma_req->num_outstanding_data_wr = 0; 3864 3865 nvmf_rdma_request_process(rtransport, rdma_req); 3866 break; 3867 case RDMA_WR_TYPE_RECV: 3868 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3869 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3870 if (rpoller->srq != NULL) { 3871 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3872 /* It is possible that there are still some completions for destroyed QP 3873 * associated with SRQ. We just ignore these late completions and re-post 3874 * receive WRs back to SRQ. 3875 */ 3876 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3877 struct ibv_recv_wr *bad_wr; 3878 int rc; 3879 3880 rdma_recv->wr.next = NULL; 3881 spdk_rdma_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 3882 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 3883 if (rc) { 3884 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3885 } 3886 continue; 3887 } 3888 } 3889 rqpair = rdma_recv->qpair; 3890 3891 assert(rqpair != NULL); 3892 if (!wc[i].status) { 3893 assert(wc[i].opcode == IBV_WC_RECV); 3894 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3895 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3896 break; 3897 } 3898 } 3899 3900 rdma_recv->wr.next = NULL; 3901 rqpair->current_recv_depth++; 3902 rdma_recv->receive_tsc = poll_tsc; 3903 rpoller->stat.requests++; 3904 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3905 break; 3906 case RDMA_WR_TYPE_DATA: 3907 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3908 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3909 3910 assert(rdma_req->num_outstanding_data_wr > 0); 3911 3912 rqpair->current_send_depth--; 3913 rdma_req->num_outstanding_data_wr--; 3914 if (!wc[i].status) { 3915 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3916 rqpair->current_read_depth--; 3917 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3918 if (rdma_req->num_outstanding_data_wr == 0) { 3919 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3920 nvmf_rdma_request_process(rtransport, rdma_req); 3921 } 3922 } else { 3923 /* If the data transfer fails still force the queue into the error state, 3924 * if we were performing an RDMA_READ, we need to force the request into a 3925 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3926 * case, we should wait for the SEND to complete. */ 3927 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3928 rqpair->current_read_depth--; 3929 if (rdma_req->num_outstanding_data_wr == 0) { 3930 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3931 } 3932 } 3933 } 3934 break; 3935 default: 3936 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3937 continue; 3938 } 3939 3940 /* Handle error conditions */ 3941 if (wc[i].status) { 3942 nvmf_rdma_update_ibv_state(rqpair); 3943 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 3944 3945 error = true; 3946 3947 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3948 /* Disconnect the connection. */ 3949 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3950 } else { 3951 nvmf_rdma_destroy_drained_qpair(rqpair); 3952 } 3953 continue; 3954 } 3955 3956 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3957 3958 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3959 nvmf_rdma_destroy_drained_qpair(rqpair); 3960 } 3961 } 3962 3963 if (error == true) { 3964 return -1; 3965 } 3966 3967 /* submit outstanding work requests. */ 3968 _poller_submit_recvs(rtransport, rpoller); 3969 _poller_submit_sends(rtransport, rpoller); 3970 3971 return count; 3972 } 3973 3974 static int 3975 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3976 { 3977 struct spdk_nvmf_rdma_transport *rtransport; 3978 struct spdk_nvmf_rdma_poll_group *rgroup; 3979 struct spdk_nvmf_rdma_poller *rpoller; 3980 int count, rc; 3981 3982 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3983 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3984 3985 count = 0; 3986 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3987 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 3988 if (rc < 0) { 3989 return rc; 3990 } 3991 count += rc; 3992 } 3993 3994 return count; 3995 } 3996 3997 static int 3998 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3999 struct spdk_nvme_transport_id *trid, 4000 bool peer) 4001 { 4002 struct sockaddr *saddr; 4003 uint16_t port; 4004 4005 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4006 4007 if (peer) { 4008 saddr = rdma_get_peer_addr(id); 4009 } else { 4010 saddr = rdma_get_local_addr(id); 4011 } 4012 switch (saddr->sa_family) { 4013 case AF_INET: { 4014 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4015 4016 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4017 inet_ntop(AF_INET, &saddr_in->sin_addr, 4018 trid->traddr, sizeof(trid->traddr)); 4019 if (peer) { 4020 port = ntohs(rdma_get_dst_port(id)); 4021 } else { 4022 port = ntohs(rdma_get_src_port(id)); 4023 } 4024 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4025 break; 4026 } 4027 case AF_INET6: { 4028 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4029 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4030 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4031 trid->traddr, sizeof(trid->traddr)); 4032 if (peer) { 4033 port = ntohs(rdma_get_dst_port(id)); 4034 } else { 4035 port = ntohs(rdma_get_src_port(id)); 4036 } 4037 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4038 break; 4039 } 4040 default: 4041 return -1; 4042 4043 } 4044 4045 return 0; 4046 } 4047 4048 static int 4049 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4050 struct spdk_nvme_transport_id *trid) 4051 { 4052 struct spdk_nvmf_rdma_qpair *rqpair; 4053 4054 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4055 4056 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4057 } 4058 4059 static int 4060 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4061 struct spdk_nvme_transport_id *trid) 4062 { 4063 struct spdk_nvmf_rdma_qpair *rqpair; 4064 4065 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4066 4067 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4068 } 4069 4070 static int 4071 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4072 struct spdk_nvme_transport_id *trid) 4073 { 4074 struct spdk_nvmf_rdma_qpair *rqpair; 4075 4076 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4077 4078 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4079 } 4080 4081 void 4082 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4083 { 4084 g_nvmf_hooks = *hooks; 4085 } 4086 4087 static void 4088 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4089 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4090 { 4091 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4092 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4093 4094 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4095 4096 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4097 } 4098 4099 static int 4100 _nvmf_rdma_qpair_abort_request(void *ctx) 4101 { 4102 struct spdk_nvmf_request *req = ctx; 4103 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4104 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4105 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4106 struct spdk_nvmf_rdma_qpair, qpair); 4107 int rc; 4108 4109 spdk_poller_unregister(&req->poller); 4110 4111 switch (rdma_req_to_abort->state) { 4112 case RDMA_REQUEST_STATE_EXECUTING: 4113 rc = nvmf_ctrlr_abort_request(req); 4114 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4115 return SPDK_POLLER_BUSY; 4116 } 4117 break; 4118 4119 case RDMA_REQUEST_STATE_NEED_BUFFER: 4120 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4121 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4122 4123 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4124 break; 4125 4126 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4127 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4128 spdk_nvmf_rdma_request, state_link); 4129 4130 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4131 break; 4132 4133 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4134 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4135 spdk_nvmf_rdma_request, state_link); 4136 4137 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4138 break; 4139 4140 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4141 if (spdk_get_ticks() < req->timeout_tsc) { 4142 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4143 return SPDK_POLLER_BUSY; 4144 } 4145 break; 4146 4147 default: 4148 break; 4149 } 4150 4151 spdk_nvmf_request_complete(req); 4152 return SPDK_POLLER_BUSY; 4153 } 4154 4155 static void 4156 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4157 struct spdk_nvmf_request *req) 4158 { 4159 struct spdk_nvmf_rdma_qpair *rqpair; 4160 struct spdk_nvmf_rdma_transport *rtransport; 4161 struct spdk_nvmf_transport *transport; 4162 uint16_t cid; 4163 uint32_t i, max_req_count; 4164 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 4165 4166 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4167 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4168 transport = &rtransport->transport; 4169 4170 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4171 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 4172 4173 for (i = 0; i < max_req_count; i++) { 4174 rdma_req = &rqpair->resources->reqs[i]; 4175 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 4176 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 4177 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 4178 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 4179 rdma_req->req.qpair == qpair) { 4180 rdma_req_to_abort = rdma_req; 4181 break; 4182 } 4183 } 4184 4185 if (rdma_req_to_abort == NULL) { 4186 spdk_nvmf_request_complete(req); 4187 return; 4188 } 4189 4190 req->req_to_abort = &rdma_req_to_abort->req; 4191 req->timeout_tsc = spdk_get_ticks() + 4192 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4193 req->poller = NULL; 4194 4195 _nvmf_rdma_qpair_abort_request(req); 4196 } 4197 4198 static void 4199 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 4200 struct spdk_json_write_ctx *w) 4201 { 4202 struct spdk_nvmf_rdma_poll_group *rgroup; 4203 struct spdk_nvmf_rdma_poller *rpoller; 4204 4205 assert(w != NULL); 4206 4207 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4208 4209 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 4210 4211 spdk_json_write_named_array_begin(w, "devices"); 4212 4213 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4214 spdk_json_write_object_begin(w); 4215 spdk_json_write_named_string(w, "name", 4216 ibv_get_device_name(rpoller->device->context->device)); 4217 spdk_json_write_named_uint64(w, "polls", 4218 rpoller->stat.polls); 4219 spdk_json_write_named_uint64(w, "idle_polls", 4220 rpoller->stat.idle_polls); 4221 spdk_json_write_named_uint64(w, "completions", 4222 rpoller->stat.completions); 4223 spdk_json_write_named_uint64(w, "requests", 4224 rpoller->stat.requests); 4225 spdk_json_write_named_uint64(w, "request_latency", 4226 rpoller->stat.request_latency); 4227 spdk_json_write_named_uint64(w, "pending_free_request", 4228 rpoller->stat.pending_free_request); 4229 spdk_json_write_named_uint64(w, "pending_rdma_read", 4230 rpoller->stat.pending_rdma_read); 4231 spdk_json_write_named_uint64(w, "pending_rdma_write", 4232 rpoller->stat.pending_rdma_write); 4233 spdk_json_write_named_uint64(w, "total_send_wrs", 4234 rpoller->stat.qp_stats.send.num_submitted_wrs); 4235 spdk_json_write_named_uint64(w, "send_doorbell_updates", 4236 rpoller->stat.qp_stats.send.doorbell_updates); 4237 spdk_json_write_named_uint64(w, "total_recv_wrs", 4238 rpoller->stat.qp_stats.recv.num_submitted_wrs); 4239 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 4240 rpoller->stat.qp_stats.recv.doorbell_updates); 4241 spdk_json_write_object_end(w); 4242 } 4243 4244 spdk_json_write_array_end(w); 4245 } 4246 4247 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4248 .name = "RDMA", 4249 .type = SPDK_NVME_TRANSPORT_RDMA, 4250 .opts_init = nvmf_rdma_opts_init, 4251 .create = nvmf_rdma_create, 4252 .dump_opts = nvmf_rdma_dump_opts, 4253 .destroy = nvmf_rdma_destroy, 4254 4255 .listen = nvmf_rdma_listen, 4256 .stop_listen = nvmf_rdma_stop_listen, 4257 .cdata_init = nvmf_rdma_cdata_init, 4258 4259 .listener_discover = nvmf_rdma_discover, 4260 4261 .poll_group_create = nvmf_rdma_poll_group_create, 4262 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4263 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4264 .poll_group_add = nvmf_rdma_poll_group_add, 4265 .poll_group_remove = nvmf_rdma_poll_group_remove, 4266 .poll_group_poll = nvmf_rdma_poll_group_poll, 4267 4268 .req_free = nvmf_rdma_request_free, 4269 .req_complete = nvmf_rdma_request_complete, 4270 4271 .qpair_fini = nvmf_rdma_close_qpair, 4272 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4273 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4274 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4275 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4276 4277 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 4278 }; 4279 4280 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4281 SPDK_LOG_REGISTER_COMPONENT(rdma) 4282