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