1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (C) 2016 Intel Corporation. All rights reserved. 3 * Copyright (c) 2019-2021 Mellanox Technologies LTD. All rights reserved. 4 * Copyright (c) 2021-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved. 5 */ 6 7 #include "spdk/stdinc.h" 8 9 #include "spdk/config.h" 10 #include "spdk/thread.h" 11 #include "spdk/likely.h" 12 #include "spdk/nvmf_transport.h" 13 #include "spdk/string.h" 14 #include "spdk/trace.h" 15 #include "spdk/tree.h" 16 #include "spdk/util.h" 17 18 #include "spdk_internal/assert.h" 19 #include "spdk/log.h" 20 #include "spdk_internal/rdma_provider.h" 21 #include "spdk_internal/rdma_utils.h" 22 23 #include "nvmf_internal.h" 24 #include "transport.h" 25 26 #include "spdk_internal/trace_defs.h" 27 28 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 29 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 30 31 /* 32 RDMA Connection Resource Defaults 33 */ 34 #define NVMF_DEFAULT_MSDBD 16 35 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 36 #define NVMF_DEFAULT_RSP_SGE 1 37 #define NVMF_DEFAULT_RX_SGE 2 38 39 #define NVMF_RDMA_MAX_EVENTS_PER_POLL 32 40 41 SPDK_STATIC_ASSERT(NVMF_DEFAULT_MSDBD <= SPDK_NVMF_MAX_SGL_ENTRIES, 42 "MSDBD must not exceed SPDK_NVMF_MAX_SGL_ENTRIES"); 43 44 /* The RDMA completion queue size */ 45 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 46 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 47 48 enum spdk_nvmf_rdma_request_state { 49 /* The request is not currently in use */ 50 RDMA_REQUEST_STATE_FREE = 0, 51 52 /* Initial state when request first received */ 53 RDMA_REQUEST_STATE_NEW, 54 55 /* The request is queued until a data buffer is available. */ 56 RDMA_REQUEST_STATE_NEED_BUFFER, 57 58 /* The request is waiting on RDMA queue depth availability 59 * to transfer data from the host to the controller. 60 */ 61 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 62 63 /* The request is currently transferring data from the host to the controller. */ 64 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 65 66 /* The request is ready to execute at the block device */ 67 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 68 69 /* The request is currently executing at the block device */ 70 RDMA_REQUEST_STATE_EXECUTING, 71 72 /* The request finished executing at the block device */ 73 RDMA_REQUEST_STATE_EXECUTED, 74 75 /* The request is waiting on RDMA queue depth availability 76 * to transfer data from the controller to the host. 77 */ 78 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 79 80 /* The request is waiting on RDMA queue depth availability 81 * to send response to the host. 82 */ 83 RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 84 85 /* The request is ready to send a completion */ 86 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 87 88 /* The request is currently transferring data from the controller to the host. */ 89 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 90 91 /* The request currently has an outstanding completion without an 92 * associated data transfer. 93 */ 94 RDMA_REQUEST_STATE_COMPLETING, 95 96 /* The request completed and can be marked free. */ 97 RDMA_REQUEST_STATE_COMPLETED, 98 99 /* Terminator */ 100 RDMA_REQUEST_NUM_STATES, 101 }; 102 103 static void 104 nvmf_trace(void) 105 { 106 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 107 108 struct spdk_trace_tpoint_opts opts[] = { 109 { 110 "RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 111 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 1, 112 { 113 { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 }, 114 { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 } 115 } 116 }, 117 { 118 "RDMA_REQ_COMPLETED", TRACE_RDMA_REQUEST_STATE_COMPLETED, 119 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 120 { 121 { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 }, 122 { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 } 123 } 124 }, 125 }; 126 127 spdk_trace_register_description_ext(opts, SPDK_COUNTOF(opts)); 128 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 129 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 130 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 131 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 132 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 133 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 134 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 135 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 136 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 137 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 138 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 139 spdk_trace_register_description("RDMA_REQ_TX_H2C", 140 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 141 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 142 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 143 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 144 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 145 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 146 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 147 spdk_trace_register_description("RDMA_REQ_EXECUTING", 148 TRACE_RDMA_REQUEST_STATE_EXECUTING, 149 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 150 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 151 spdk_trace_register_description("RDMA_REQ_EXECUTED", 152 TRACE_RDMA_REQUEST_STATE_EXECUTED, 153 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 154 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 155 spdk_trace_register_description("RDMA_REQ_RDY2COMPL_PEND", 156 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 157 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 158 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 159 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 160 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 161 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 162 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 163 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 164 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 165 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 166 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 167 spdk_trace_register_description("RDMA_REQ_COMPLETING", 168 TRACE_RDMA_REQUEST_STATE_COMPLETING, 169 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 170 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 171 172 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 173 OWNER_TYPE_NONE, OBJECT_NONE, 0, 174 SPDK_TRACE_ARG_TYPE_INT, ""); 175 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 176 OWNER_TYPE_NONE, OBJECT_NONE, 0, 177 SPDK_TRACE_ARG_TYPE_INT, "type"); 178 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 179 OWNER_TYPE_NONE, OBJECT_NONE, 0, 180 SPDK_TRACE_ARG_TYPE_INT, "type"); 181 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 182 OWNER_TYPE_NONE, OBJECT_NONE, 0, 183 SPDK_TRACE_ARG_TYPE_INT, ""); 184 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 185 OWNER_TYPE_NONE, OBJECT_NONE, 0, 186 SPDK_TRACE_ARG_TYPE_INT, ""); 187 188 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_RDMA_IO, 1); 189 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_RDMA_IO, 0); 190 } 191 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 192 193 enum spdk_nvmf_rdma_wr_type { 194 RDMA_WR_TYPE_RECV, 195 RDMA_WR_TYPE_SEND, 196 RDMA_WR_TYPE_DATA, 197 }; 198 199 struct spdk_nvmf_rdma_wr { 200 /* Uses enum spdk_nvmf_rdma_wr_type */ 201 uint8_t type; 202 }; 203 204 /* This structure holds commands as they are received off the wire. 205 * It must be dynamically paired with a full request object 206 * (spdk_nvmf_rdma_request) to service a request. It is separate 207 * from the request because RDMA does not appear to order 208 * completions, so occasionally we'll get a new incoming 209 * command when there aren't any free request objects. 210 */ 211 struct spdk_nvmf_rdma_recv { 212 struct ibv_recv_wr wr; 213 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 214 215 struct spdk_nvmf_rdma_qpair *qpair; 216 217 /* In-capsule data buffer */ 218 uint8_t *buf; 219 220 struct spdk_nvmf_rdma_wr rdma_wr; 221 uint64_t receive_tsc; 222 223 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 224 }; 225 226 struct spdk_nvmf_rdma_request_data { 227 struct ibv_send_wr wr; 228 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 229 }; 230 231 struct spdk_nvmf_rdma_request { 232 struct spdk_nvmf_request req; 233 234 bool fused_failed; 235 236 struct spdk_nvmf_rdma_wr data_wr; 237 struct spdk_nvmf_rdma_wr rsp_wr; 238 239 /* Uses enum spdk_nvmf_rdma_request_state */ 240 uint8_t state; 241 242 /* Data offset in req.iov */ 243 uint32_t offset; 244 245 struct spdk_nvmf_rdma_recv *recv; 246 247 struct { 248 struct ibv_send_wr wr; 249 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 250 } rsp; 251 252 uint16_t iovpos; 253 uint16_t num_outstanding_data_wr; 254 /* Used to split Write IO with multi SGL payload */ 255 uint16_t num_remaining_data_wr; 256 uint64_t receive_tsc; 257 struct spdk_nvmf_rdma_request *fused_pair; 258 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 259 struct ibv_send_wr *remaining_tranfer_in_wrs; 260 struct ibv_send_wr *transfer_wr; 261 struct spdk_nvmf_rdma_request_data data; 262 }; 263 264 struct spdk_nvmf_rdma_resource_opts { 265 struct spdk_nvmf_rdma_qpair *qpair; 266 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 267 void *qp; 268 struct spdk_rdma_utils_mem_map *map; 269 uint32_t max_queue_depth; 270 uint32_t in_capsule_data_size; 271 bool shared; 272 }; 273 274 struct spdk_nvmf_rdma_resources { 275 /* Array of size "max_queue_depth" containing RDMA requests. */ 276 struct spdk_nvmf_rdma_request *reqs; 277 278 /* Array of size "max_queue_depth" containing RDMA recvs. */ 279 struct spdk_nvmf_rdma_recv *recvs; 280 281 /* Array of size "max_queue_depth" containing 64 byte capsules 282 * used for receive. 283 */ 284 union nvmf_h2c_msg *cmds; 285 286 /* Array of size "max_queue_depth" containing 16 byte completions 287 * to be sent back to the user. 288 */ 289 union nvmf_c2h_msg *cpls; 290 291 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 292 * buffers to be used for in capsule data. 293 */ 294 void *bufs; 295 296 /* Receives that are waiting for a request object */ 297 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 298 299 /* Queue to track free requests */ 300 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 301 }; 302 303 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 304 305 typedef void (*spdk_poller_destroy_cb)(void *ctx); 306 307 struct spdk_nvmf_rdma_ibv_event_ctx { 308 struct spdk_nvmf_rdma_qpair *rqpair; 309 }; 310 311 struct spdk_nvmf_rdma_qpair { 312 struct spdk_nvmf_qpair qpair; 313 314 struct spdk_nvmf_rdma_device *device; 315 struct spdk_nvmf_rdma_poller *poller; 316 317 struct spdk_rdma_provider_qp *rdma_qp; 318 struct rdma_cm_id *cm_id; 319 struct spdk_rdma_provider_srq *srq; 320 struct rdma_cm_id *listen_id; 321 322 /* Cache the QP number to improve QP search by RB tree. */ 323 uint32_t qp_num; 324 325 /* The maximum number of I/O outstanding on this connection at one time */ 326 uint16_t max_queue_depth; 327 328 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 329 uint16_t max_read_depth; 330 331 /* The maximum number of RDMA SEND operations at one time */ 332 uint32_t max_send_depth; 333 334 /* The current number of outstanding WRs from this qpair's 335 * recv queue. Should not exceed device->attr.max_queue_depth. 336 */ 337 uint16_t current_recv_depth; 338 339 /* The current number of active RDMA READ operations */ 340 uint16_t current_read_depth; 341 342 /* The current number of posted WRs from this qpair's 343 * send queue. Should not exceed max_send_depth. 344 */ 345 uint32_t current_send_depth; 346 347 /* The maximum number of SGEs per WR on the send queue */ 348 uint32_t max_send_sge; 349 350 /* The maximum number of SGEs per WR on the recv queue */ 351 uint32_t max_recv_sge; 352 353 struct spdk_nvmf_rdma_resources *resources; 354 355 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 356 357 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 358 359 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_send_queue; 360 361 /* Number of requests not in the free state */ 362 uint32_t qd; 363 364 bool ibv_in_error_state; 365 366 RB_ENTRY(spdk_nvmf_rdma_qpair) node; 367 368 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 369 370 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 371 372 /* Points to the a request that has fuse bits set to 373 * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting 374 * for the request that has SPDK_NVME_CMD_FUSE_SECOND. 375 */ 376 struct spdk_nvmf_rdma_request *fused_first; 377 378 /* 379 * io_channel which is used to destroy qpair when it is removed from poll group 380 */ 381 struct spdk_io_channel *destruct_channel; 382 383 /* ctx for async processing of last_wqe_reached event */ 384 struct spdk_nvmf_rdma_ibv_event_ctx *last_wqe_reached_ctx; 385 386 /* Lets us know that we have received the last_wqe event. */ 387 bool last_wqe_reached; 388 389 /* Indicate that nvmf_rdma_close_qpair is called */ 390 bool to_close; 391 }; 392 393 struct spdk_nvmf_rdma_poller_stat { 394 uint64_t completions; 395 uint64_t polls; 396 uint64_t idle_polls; 397 uint64_t requests; 398 uint64_t request_latency; 399 uint64_t pending_free_request; 400 uint64_t pending_rdma_read; 401 uint64_t pending_rdma_write; 402 uint64_t pending_rdma_send; 403 struct spdk_rdma_provider_qp_stats qp_stats; 404 }; 405 406 struct spdk_nvmf_rdma_poller { 407 struct spdk_nvmf_rdma_device *device; 408 struct spdk_nvmf_rdma_poll_group *group; 409 410 int num_cqe; 411 int required_num_wr; 412 struct ibv_cq *cq; 413 414 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 415 uint16_t max_srq_depth; 416 bool need_destroy; 417 418 /* Shared receive queue */ 419 struct spdk_rdma_provider_srq *srq; 420 421 struct spdk_nvmf_rdma_resources *resources; 422 struct spdk_nvmf_rdma_poller_stat stat; 423 424 spdk_poller_destroy_cb destroy_cb; 425 void *destroy_cb_ctx; 426 427 RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs; 428 429 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 430 431 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 432 433 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 434 }; 435 436 struct spdk_nvmf_rdma_poll_group_stat { 437 uint64_t pending_data_buffer; 438 }; 439 440 struct spdk_nvmf_rdma_poll_group { 441 struct spdk_nvmf_transport_poll_group group; 442 struct spdk_nvmf_rdma_poll_group_stat stat; 443 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 444 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 445 }; 446 447 struct spdk_nvmf_rdma_conn_sched { 448 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 449 struct spdk_nvmf_rdma_poll_group *next_io_pg; 450 }; 451 452 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 453 struct spdk_nvmf_rdma_device { 454 struct ibv_device_attr attr; 455 struct ibv_context *context; 456 457 struct spdk_rdma_utils_mem_map *map; 458 struct ibv_pd *pd; 459 460 int num_srq; 461 bool need_destroy; 462 bool ready_to_destroy; 463 bool is_ready; 464 465 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 466 }; 467 468 struct spdk_nvmf_rdma_port { 469 const struct spdk_nvme_transport_id *trid; 470 struct rdma_cm_id *id; 471 struct spdk_nvmf_rdma_device *device; 472 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 473 }; 474 475 struct rdma_transport_opts { 476 int num_cqe; 477 uint32_t max_srq_depth; 478 bool no_srq; 479 bool no_wr_batching; 480 int acceptor_backlog; 481 }; 482 483 struct spdk_nvmf_rdma_transport { 484 struct spdk_nvmf_transport transport; 485 struct rdma_transport_opts rdma_opts; 486 487 struct spdk_nvmf_rdma_conn_sched conn_sched; 488 489 struct rdma_event_channel *event_channel; 490 491 struct spdk_mempool *data_wr_pool; 492 493 struct spdk_poller *accept_poller; 494 495 /* fields used to poll RDMA/IB events */ 496 nfds_t npoll_fds; 497 struct pollfd *poll_fds; 498 499 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 500 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 501 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 502 503 /* ports that are removed unexpectedly and need retry listen */ 504 TAILQ_HEAD(, spdk_nvmf_rdma_port) retry_ports; 505 }; 506 507 struct poller_manage_ctx { 508 struct spdk_nvmf_rdma_transport *rtransport; 509 struct spdk_nvmf_rdma_poll_group *rgroup; 510 struct spdk_nvmf_rdma_poller *rpoller; 511 struct spdk_nvmf_rdma_device *device; 512 513 struct spdk_thread *thread; 514 volatile int *inflight_op_counter; 515 }; 516 517 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 518 { 519 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe), 520 spdk_json_decode_int32, true 521 }, 522 { 523 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 524 spdk_json_decode_uint32, true 525 }, 526 { 527 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 528 spdk_json_decode_bool, true 529 }, 530 { 531 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 532 spdk_json_decode_bool, true 533 }, 534 { 535 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 536 spdk_json_decode_int32, true 537 }, 538 }; 539 540 static int 541 nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2) 542 { 543 return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num; 544 } 545 546 RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare); 547 548 static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 549 struct spdk_nvmf_rdma_request *rdma_req); 550 551 static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 552 struct spdk_nvmf_rdma_poller *rpoller); 553 554 static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 555 struct spdk_nvmf_rdma_poller *rpoller); 556 557 static void _nvmf_rdma_remove_destroyed_device(void *c); 558 559 static inline enum spdk_nvme_media_error_status_code 560 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 561 enum spdk_nvme_media_error_status_code result; 562 switch (err_type) 563 { 564 case SPDK_DIF_REFTAG_ERROR: 565 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 566 break; 567 case SPDK_DIF_APPTAG_ERROR: 568 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 569 break; 570 case SPDK_DIF_GUARD_ERROR: 571 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 572 break; 573 default: 574 SPDK_UNREACHABLE(); 575 } 576 577 return result; 578 } 579 580 /* 581 * Return data_wrs to pool starting from \b data_wr 582 * Request's own response and data WR are excluded 583 */ 584 static void 585 _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 586 struct ibv_send_wr *data_wr, 587 struct spdk_mempool *pool) 588 { 589 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 590 struct spdk_nvmf_rdma_request_data *nvmf_data; 591 struct ibv_send_wr *next_send_wr; 592 uint64_t req_wrid = (uint64_t)&rdma_req->data_wr; 593 uint32_t num_wrs = 0; 594 595 while (data_wr && data_wr->wr_id == req_wrid) { 596 nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr); 597 memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge); 598 data_wr->num_sge = 0; 599 next_send_wr = data_wr->next; 600 if (data_wr != &rdma_req->data.wr) { 601 data_wr->next = NULL; 602 assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES); 603 work_requests[num_wrs] = nvmf_data; 604 num_wrs++; 605 } 606 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr; 607 } 608 609 if (num_wrs) { 610 spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs); 611 } 612 } 613 614 static void 615 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 616 struct spdk_nvmf_rdma_transport *rtransport) 617 { 618 rdma_req->num_outstanding_data_wr = 0; 619 620 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 621 622 if (rdma_req->remaining_tranfer_in_wrs) { 623 _nvmf_rdma_request_free_data(rdma_req, rdma_req->remaining_tranfer_in_wrs, 624 rtransport->data_wr_pool); 625 rdma_req->remaining_tranfer_in_wrs = NULL; 626 } 627 628 rdma_req->data.wr.next = NULL; 629 rdma_req->rsp.wr.next = NULL; 630 } 631 632 static void 633 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 634 { 635 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 636 if (req->req.cmd) { 637 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 638 } 639 if (req->recv) { 640 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 641 } 642 } 643 644 static void 645 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 646 { 647 int i; 648 649 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 650 for (i = 0; i < rqpair->max_queue_depth; i++) { 651 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 652 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 653 } 654 } 655 } 656 657 static void 658 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 659 { 660 spdk_free(resources->cmds); 661 spdk_free(resources->cpls); 662 spdk_free(resources->bufs); 663 spdk_free(resources->reqs); 664 spdk_free(resources->recvs); 665 free(resources); 666 } 667 668 669 static struct spdk_nvmf_rdma_resources * 670 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 671 { 672 struct spdk_nvmf_rdma_resources *resources; 673 struct spdk_nvmf_rdma_request *rdma_req; 674 struct spdk_nvmf_rdma_recv *rdma_recv; 675 struct spdk_rdma_provider_qp *qp = NULL; 676 struct spdk_rdma_provider_srq *srq = NULL; 677 struct ibv_recv_wr *bad_wr = NULL; 678 struct spdk_rdma_utils_memory_translation translation; 679 uint32_t i; 680 int rc = 0; 681 682 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 683 if (!resources) { 684 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 685 return NULL; 686 } 687 688 resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs), 689 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 690 resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs), 691 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 692 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 693 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 694 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 695 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 696 697 if (opts->in_capsule_data_size > 0) { 698 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 699 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 700 SPDK_MALLOC_DMA); 701 } 702 703 if (!resources->reqs || !resources->recvs || !resources->cmds || 704 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 705 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 706 goto cleanup; 707 } 708 709 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n", 710 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds)); 711 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n", 712 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls)); 713 if (resources->bufs) { 714 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n", 715 resources->bufs, opts->max_queue_depth * 716 opts->in_capsule_data_size); 717 } 718 719 /* Initialize queues */ 720 STAILQ_INIT(&resources->incoming_queue); 721 STAILQ_INIT(&resources->free_queue); 722 723 if (opts->shared) { 724 srq = (struct spdk_rdma_provider_srq *)opts->qp; 725 } else { 726 qp = (struct spdk_rdma_provider_qp *)opts->qp; 727 } 728 729 for (i = 0; i < opts->max_queue_depth; i++) { 730 rdma_recv = &resources->recvs[i]; 731 rdma_recv->qpair = opts->qpair; 732 733 /* Set up memory to receive commands */ 734 if (resources->bufs) { 735 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 736 opts->in_capsule_data_size)); 737 } 738 739 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 740 741 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 742 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 743 rc = spdk_rdma_utils_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]), 744 &translation); 745 if (rc) { 746 goto cleanup; 747 } 748 rdma_recv->sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 749 rdma_recv->wr.num_sge = 1; 750 751 if (rdma_recv->buf) { 752 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 753 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 754 rc = spdk_rdma_utils_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size, 755 &translation); 756 if (rc) { 757 goto cleanup; 758 } 759 rdma_recv->sgl[1].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 760 rdma_recv->wr.num_sge++; 761 } 762 763 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 764 rdma_recv->wr.sg_list = rdma_recv->sgl; 765 if (srq) { 766 spdk_rdma_provider_srq_queue_recv_wrs(srq, &rdma_recv->wr); 767 } else { 768 spdk_rdma_provider_qp_queue_recv_wrs(qp, &rdma_recv->wr); 769 } 770 } 771 772 for (i = 0; i < opts->max_queue_depth; i++) { 773 rdma_req = &resources->reqs[i]; 774 775 if (opts->qpair != NULL) { 776 rdma_req->req.qpair = &opts->qpair->qpair; 777 } else { 778 rdma_req->req.qpair = NULL; 779 } 780 rdma_req->req.cmd = NULL; 781 rdma_req->req.iovcnt = 0; 782 rdma_req->req.stripped_data = NULL; 783 784 /* Set up memory to send responses */ 785 rdma_req->req.rsp = &resources->cpls[i]; 786 787 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 788 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 789 rc = spdk_rdma_utils_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]), 790 &translation); 791 if (rc) { 792 goto cleanup; 793 } 794 rdma_req->rsp.sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 795 796 rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND; 797 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr; 798 rdma_req->rsp.wr.next = NULL; 799 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 800 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 801 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 802 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 803 804 /* Set up memory for data buffers */ 805 rdma_req->data_wr.type = RDMA_WR_TYPE_DATA; 806 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr; 807 rdma_req->data.wr.next = NULL; 808 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 809 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 810 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 811 812 /* Initialize request state to FREE */ 813 rdma_req->state = RDMA_REQUEST_STATE_FREE; 814 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 815 } 816 817 if (srq) { 818 rc = spdk_rdma_provider_srq_flush_recv_wrs(srq, &bad_wr); 819 } else { 820 rc = spdk_rdma_provider_qp_flush_recv_wrs(qp, &bad_wr); 821 } 822 823 if (rc) { 824 goto cleanup; 825 } 826 827 return resources; 828 829 cleanup: 830 nvmf_rdma_resources_destroy(resources); 831 return NULL; 832 } 833 834 static void 835 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 836 { 837 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 838 839 ctx = rqpair->last_wqe_reached_ctx; 840 if (ctx) { 841 ctx->rqpair = NULL; 842 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_last_wqe_event */ 843 rqpair->last_wqe_reached_ctx = NULL; 844 } 845 } 846 847 static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller); 848 849 static void 850 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 851 { 852 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 853 struct ibv_recv_wr *bad_recv_wr = NULL; 854 int rc; 855 856 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair); 857 858 if (rqpair->qd != 0) { 859 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 860 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 861 struct spdk_nvmf_rdma_transport, transport); 862 struct spdk_nvmf_rdma_request *req; 863 uint32_t i, max_req_count = 0; 864 865 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 866 867 if (rqpair->srq == NULL) { 868 nvmf_rdma_dump_qpair_contents(rqpair); 869 max_req_count = rqpair->max_queue_depth; 870 } else if (rqpair->poller && rqpair->resources) { 871 max_req_count = rqpair->poller->max_srq_depth; 872 } 873 874 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 875 for (i = 0; i < max_req_count; i++) { 876 req = &rqpair->resources->reqs[i]; 877 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 878 /* nvmf_rdma_request_process checks qpair ibv and internal state 879 * and completes a request */ 880 nvmf_rdma_request_process(rtransport, req); 881 } 882 } 883 assert(rqpair->qd == 0); 884 } 885 886 if (rqpair->poller) { 887 RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair); 888 889 if (rqpair->srq != NULL && rqpair->resources != NULL) { 890 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 891 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 892 if (rqpair == rdma_recv->qpair) { 893 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 894 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr); 895 rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 896 if (rc) { 897 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 898 } 899 } 900 } 901 } 902 } 903 904 if (rqpair->cm_id) { 905 if (rqpair->rdma_qp != NULL) { 906 spdk_rdma_provider_qp_destroy(rqpair->rdma_qp); 907 rqpair->rdma_qp = NULL; 908 } 909 910 if (rqpair->poller != NULL && rqpair->srq == NULL) { 911 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 912 } 913 } 914 915 if (rqpair->srq == NULL && rqpair->resources != NULL) { 916 nvmf_rdma_resources_destroy(rqpair->resources); 917 } 918 919 nvmf_rdma_qpair_clean_ibv_events(rqpair); 920 921 if (rqpair->destruct_channel) { 922 spdk_put_io_channel(rqpair->destruct_channel); 923 rqpair->destruct_channel = NULL; 924 } 925 926 if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) { 927 nvmf_rdma_poller_destroy(rqpair->poller); 928 } 929 930 /* destroy cm_id last so cma device will not be freed before we destroy the cq. */ 931 if (rqpair->cm_id) { 932 rdma_destroy_id(rqpair->cm_id); 933 } 934 935 free(rqpair); 936 } 937 938 static int 939 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 940 { 941 struct spdk_nvmf_rdma_poller *rpoller; 942 int rc, num_cqe, required_num_wr; 943 944 /* Enlarge CQ size dynamically */ 945 rpoller = rqpair->poller; 946 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 947 num_cqe = rpoller->num_cqe; 948 if (num_cqe < required_num_wr) { 949 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 950 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 951 } 952 953 if (rpoller->num_cqe != num_cqe) { 954 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) { 955 SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n" 956 "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe); 957 return -1; 958 } 959 if (required_num_wr > device->attr.max_cqe) { 960 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 961 required_num_wr, device->attr.max_cqe); 962 return -1; 963 } 964 965 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 966 rc = ibv_resize_cq(rpoller->cq, num_cqe); 967 if (rc) { 968 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 969 return -1; 970 } 971 972 rpoller->num_cqe = num_cqe; 973 } 974 975 rpoller->required_num_wr = required_num_wr; 976 return 0; 977 } 978 979 static int 980 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 981 { 982 struct spdk_nvmf_rdma_qpair *rqpair; 983 struct spdk_nvmf_rdma_transport *rtransport; 984 struct spdk_nvmf_transport *transport; 985 struct spdk_nvmf_rdma_resource_opts opts; 986 struct spdk_nvmf_rdma_device *device; 987 struct spdk_rdma_provider_qp_init_attr qp_init_attr = {}; 988 989 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 990 device = rqpair->device; 991 992 qp_init_attr.qp_context = rqpair; 993 qp_init_attr.pd = device->pd; 994 qp_init_attr.send_cq = rqpair->poller->cq; 995 qp_init_attr.recv_cq = rqpair->poller->cq; 996 997 if (rqpair->srq) { 998 qp_init_attr.srq = rqpair->srq->srq; 999 } else { 1000 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 1001 } 1002 1003 /* SEND, READ, and WRITE operations */ 1004 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 1005 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 1006 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1007 qp_init_attr.stats = &rqpair->poller->stat.qp_stats; 1008 1009 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1010 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1011 goto error; 1012 } 1013 1014 rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &qp_init_attr); 1015 if (!rqpair->rdma_qp) { 1016 goto error; 1017 } 1018 1019 rqpair->qp_num = rqpair->rdma_qp->qp->qp_num; 1020 1021 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1022 qp_init_attr.cap.max_send_wr); 1023 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1024 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1025 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair); 1026 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1027 1028 if (rqpair->poller->srq == NULL) { 1029 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1030 transport = &rtransport->transport; 1031 1032 opts.qp = rqpair->rdma_qp; 1033 opts.map = device->map; 1034 opts.qpair = rqpair; 1035 opts.shared = false; 1036 opts.max_queue_depth = rqpair->max_queue_depth; 1037 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1038 1039 rqpair->resources = nvmf_rdma_resources_create(&opts); 1040 1041 if (!rqpair->resources) { 1042 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1043 rdma_destroy_qp(rqpair->cm_id); 1044 goto error; 1045 } 1046 } else { 1047 rqpair->resources = rqpair->poller->resources; 1048 } 1049 1050 rqpair->current_recv_depth = 0; 1051 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1052 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1053 STAILQ_INIT(&rqpair->pending_rdma_send_queue); 1054 rqpair->qpair.queue_depth = 0; 1055 1056 return 0; 1057 1058 error: 1059 rdma_destroy_id(rqpair->cm_id); 1060 rqpair->cm_id = NULL; 1061 return -1; 1062 } 1063 1064 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1065 /* This function accepts either a single wr or the first wr in a linked list. */ 1066 static void 1067 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1068 { 1069 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1070 struct spdk_nvmf_rdma_transport, transport); 1071 1072 if (rqpair->srq != NULL) { 1073 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, first); 1074 } else { 1075 if (spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, first)) { 1076 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1077 } 1078 } 1079 1080 if (rtransport->rdma_opts.no_wr_batching) { 1081 _poller_submit_recvs(rtransport, rqpair->poller); 1082 } 1083 } 1084 1085 static inline void 1086 request_transfer_in(struct spdk_nvmf_request *req) 1087 { 1088 struct spdk_nvmf_rdma_request *rdma_req; 1089 struct spdk_nvmf_qpair *qpair; 1090 struct spdk_nvmf_rdma_qpair *rqpair; 1091 struct spdk_nvmf_rdma_transport *rtransport; 1092 1093 qpair = req->qpair; 1094 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1095 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1096 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1097 struct spdk_nvmf_rdma_transport, transport); 1098 1099 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1100 assert(rdma_req != NULL); 1101 1102 if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) { 1103 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1104 } 1105 if (rtransport->rdma_opts.no_wr_batching) { 1106 _poller_submit_sends(rtransport, rqpair->poller); 1107 } 1108 1109 assert(rqpair->current_read_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_read_depth); 1110 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1111 assert(rqpair->current_send_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_send_depth); 1112 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1113 } 1114 1115 static inline void 1116 nvmf_rdma_request_reset_transfer_in(struct spdk_nvmf_rdma_request *rdma_req, 1117 struct spdk_nvmf_rdma_transport *rtransport) 1118 { 1119 /* Put completed WRs back to pool and move transfer_wr pointer */ 1120 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 1121 rdma_req->transfer_wr = rdma_req->remaining_tranfer_in_wrs; 1122 rdma_req->remaining_tranfer_in_wrs = NULL; 1123 rdma_req->num_outstanding_data_wr = rdma_req->num_remaining_data_wr; 1124 rdma_req->num_remaining_data_wr = 0; 1125 } 1126 1127 static inline int 1128 request_prepare_transfer_in_part(struct spdk_nvmf_request *req, uint32_t num_reads_available) 1129 { 1130 struct spdk_nvmf_rdma_request *rdma_req; 1131 struct ibv_send_wr *wr; 1132 uint32_t i; 1133 1134 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1135 1136 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1137 assert(rdma_req != NULL); 1138 assert(num_reads_available > 0); 1139 assert(rdma_req->num_outstanding_data_wr > num_reads_available); 1140 wr = rdma_req->transfer_wr; 1141 1142 for (i = 0; i < num_reads_available - 1; i++) { 1143 wr = wr->next; 1144 } 1145 1146 rdma_req->remaining_tranfer_in_wrs = wr->next; 1147 rdma_req->num_remaining_data_wr = rdma_req->num_outstanding_data_wr - num_reads_available; 1148 rdma_req->num_outstanding_data_wr = num_reads_available; 1149 /* Break chain of WRs to send only part. Once this portion completes, we continue sending RDMA_READs */ 1150 wr->next = NULL; 1151 1152 return 0; 1153 } 1154 1155 static int 1156 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1157 { 1158 int num_outstanding_data_wr = 0; 1159 struct spdk_nvmf_rdma_request *rdma_req; 1160 struct spdk_nvmf_qpair *qpair; 1161 struct spdk_nvmf_rdma_qpair *rqpair; 1162 struct spdk_nvme_cpl *rsp; 1163 struct ibv_send_wr *first = NULL; 1164 struct spdk_nvmf_rdma_transport *rtransport; 1165 1166 *data_posted = 0; 1167 qpair = req->qpair; 1168 rsp = &req->rsp->nvme_cpl; 1169 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1170 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1171 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1172 struct spdk_nvmf_rdma_transport, transport); 1173 1174 /* Advance our sq_head pointer */ 1175 if (qpair->sq_head == qpair->sq_head_max) { 1176 qpair->sq_head = 0; 1177 } else { 1178 qpair->sq_head++; 1179 } 1180 rsp->sqhd = qpair->sq_head; 1181 1182 /* queue the capsule for the recv buffer */ 1183 assert(rdma_req->recv != NULL); 1184 1185 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1186 1187 rdma_req->recv = NULL; 1188 assert(rqpair->current_recv_depth > 0); 1189 rqpair->current_recv_depth--; 1190 1191 /* Build the response which consists of optional 1192 * RDMA WRITEs to transfer data, plus an RDMA SEND 1193 * containing the response. 1194 */ 1195 first = &rdma_req->rsp.wr; 1196 1197 if (spdk_unlikely(rsp->status.sc != SPDK_NVME_SC_SUCCESS)) { 1198 /* On failure, data was not read from the controller. So clear the 1199 * number of outstanding data WRs to zero. 1200 */ 1201 rdma_req->num_outstanding_data_wr = 0; 1202 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1203 first = rdma_req->transfer_wr; 1204 *data_posted = 1; 1205 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1206 } 1207 if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1208 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1209 } 1210 if (rtransport->rdma_opts.no_wr_batching) { 1211 _poller_submit_sends(rtransport, rqpair->poller); 1212 } 1213 1214 /* +1 for the rsp wr */ 1215 assert(rqpair->current_send_depth + num_outstanding_data_wr + 1 <= rqpair->max_send_depth); 1216 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1217 1218 return 0; 1219 } 1220 1221 static int 1222 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1223 { 1224 struct spdk_nvmf_rdma_accept_private_data accept_data; 1225 struct rdma_conn_param ctrlr_event_data = {}; 1226 int rc; 1227 1228 accept_data.recfmt = 0; 1229 accept_data.crqsize = rqpair->max_queue_depth; 1230 1231 ctrlr_event_data.private_data = &accept_data; 1232 ctrlr_event_data.private_data_len = sizeof(accept_data); 1233 if (id->ps == RDMA_PS_TCP) { 1234 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1235 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1236 } 1237 1238 /* Configure infinite retries for the initiator side qpair. 1239 * We need to pass this value to the initiator to prevent the 1240 * initiator side NIC from completing SEND requests back to the 1241 * initiator with status rnr_retry_count_exceeded. */ 1242 ctrlr_event_data.rnr_retry_count = 0x7; 1243 1244 /* When qpair is created without use of rdma cm API, an additional 1245 * information must be provided to initiator in the connection response: 1246 * whether qpair is using SRQ and its qp_num 1247 * Fields below are ignored by rdma cm if qpair has been 1248 * created using rdma cm API. */ 1249 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1250 ctrlr_event_data.qp_num = rqpair->qp_num; 1251 1252 rc = spdk_rdma_provider_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1253 if (rc) { 1254 SPDK_ERRLOG("Error %d on spdk_rdma_provider_qp_accept\n", errno); 1255 } else { 1256 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1257 } 1258 1259 return rc; 1260 } 1261 1262 static void 1263 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1264 { 1265 struct spdk_nvmf_rdma_reject_private_data rej_data; 1266 1267 rej_data.recfmt = 0; 1268 rej_data.sts = error; 1269 1270 rdma_reject(id, &rej_data, sizeof(rej_data)); 1271 } 1272 1273 static int 1274 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1275 { 1276 struct spdk_nvmf_rdma_transport *rtransport; 1277 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1278 struct spdk_nvmf_rdma_port *port; 1279 struct rdma_conn_param *rdma_param = NULL; 1280 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1281 uint16_t max_queue_depth; 1282 uint16_t max_read_depth; 1283 1284 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1285 1286 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1287 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1288 1289 rdma_param = &event->param.conn; 1290 if (rdma_param->private_data == NULL || 1291 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1292 SPDK_ERRLOG("connect request: no private data provided\n"); 1293 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1294 return -1; 1295 } 1296 1297 private_data = rdma_param->private_data; 1298 if (private_data->recfmt != 0) { 1299 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1300 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1301 return -1; 1302 } 1303 1304 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1305 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1306 1307 port = event->listen_id->context; 1308 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1309 event->listen_id, event->listen_id->verbs, port); 1310 1311 /* Figure out the supported queue depth. This is a multi-step process 1312 * that takes into account hardware maximums, host provided values, 1313 * and our target's internal memory limits */ 1314 1315 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1316 1317 /* Start with the maximum queue depth allowed by the target */ 1318 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1319 max_read_depth = rtransport->transport.opts.max_queue_depth; 1320 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1321 rtransport->transport.opts.max_queue_depth); 1322 1323 /* Next check the local NIC's hardware limitations */ 1324 SPDK_DEBUGLOG(rdma, 1325 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1326 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1327 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1328 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1329 1330 /* Next check the remote NIC's hardware limitations */ 1331 SPDK_DEBUGLOG(rdma, 1332 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1333 rdma_param->initiator_depth, rdma_param->responder_resources); 1334 /* from man3 rdma_get_cm_event 1335 * responder_resources - Specifies the number of responder resources that is requested by the recipient. 1336 * The responder_resources field must match the initiator depth specified by the remote node when running 1337 * the rdma_connect and rdma_accept functions. */ 1338 if (rdma_param->responder_resources != 0) { 1339 if (private_data->qid) { 1340 SPDK_DEBUGLOG(rdma, "Host (Initiator) is not allowed to use RDMA operations," 1341 " responder_resources must be 0 but set to %u\n", 1342 rdma_param->responder_resources); 1343 } else { 1344 SPDK_WARNLOG("Host (Initiator) is not allowed to use RDMA operations," 1345 " responder_resources must be 0 but set to %u\n", 1346 rdma_param->responder_resources); 1347 } 1348 } 1349 /* from man3 rdma_get_cm_event 1350 * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds. 1351 * The initiator_depth field must match the responder resources specified by the remote node when running 1352 * the rdma_connect and rdma_accept functions. */ 1353 if (rdma_param->initiator_depth == 0) { 1354 SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n"); 1355 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD); 1356 return -1; 1357 } 1358 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1359 1360 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1361 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1362 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1363 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1364 1365 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1366 max_queue_depth, max_read_depth); 1367 1368 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1369 if (rqpair == NULL) { 1370 SPDK_ERRLOG("Could not allocate new connection.\n"); 1371 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1372 return -1; 1373 } 1374 1375 rqpair->device = port->device; 1376 rqpair->max_queue_depth = max_queue_depth; 1377 rqpair->max_read_depth = max_read_depth; 1378 rqpair->cm_id = event->id; 1379 rqpair->listen_id = event->listen_id; 1380 rqpair->qpair.transport = transport; 1381 /* use qid from the private data to determine the qpair type 1382 qid will be set to the appropriate value when the controller is created */ 1383 rqpair->qpair.qid = private_data->qid; 1384 rqpair->qpair.numa.id_valid = 1; 1385 rqpair->qpair.numa.id = spdk_rdma_cm_id_get_numa_id(rqpair->cm_id); 1386 1387 event->id->context = &rqpair->qpair; 1388 1389 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1390 1391 return 0; 1392 } 1393 1394 static inline void 1395 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1396 enum spdk_nvme_data_transfer xfer) 1397 { 1398 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1399 wr->opcode = IBV_WR_RDMA_WRITE; 1400 wr->send_flags = 0; 1401 wr->next = next; 1402 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1403 wr->opcode = IBV_WR_RDMA_READ; 1404 wr->send_flags = IBV_SEND_SIGNALED; 1405 wr->next = NULL; 1406 } else { 1407 assert(0); 1408 } 1409 } 1410 1411 static int 1412 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1413 struct spdk_nvmf_rdma_request *rdma_req, 1414 uint32_t num_sgl_descriptors) 1415 { 1416 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1417 struct spdk_nvmf_rdma_request_data *current_data_wr; 1418 uint32_t i; 1419 1420 if (spdk_unlikely(num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES)) { 1421 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1422 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1423 return -EINVAL; 1424 } 1425 1426 if (spdk_unlikely(spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, 1427 num_sgl_descriptors))) { 1428 return -ENOMEM; 1429 } 1430 1431 current_data_wr = &rdma_req->data; 1432 1433 for (i = 0; i < num_sgl_descriptors; i++) { 1434 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1435 current_data_wr->wr.next = &work_requests[i]->wr; 1436 current_data_wr = work_requests[i]; 1437 current_data_wr->wr.sg_list = current_data_wr->sgl; 1438 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1439 } 1440 1441 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1442 1443 return 0; 1444 } 1445 1446 static inline void 1447 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1448 { 1449 struct ibv_send_wr *wr = &rdma_req->data.wr; 1450 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1451 1452 wr->wr.rdma.rkey = sgl->keyed.key; 1453 wr->wr.rdma.remote_addr = sgl->address; 1454 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1455 } 1456 1457 static inline void 1458 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1459 { 1460 struct ibv_send_wr *wr = &rdma_req->data.wr; 1461 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1462 uint32_t i; 1463 int j; 1464 uint64_t remote_addr_offset = 0; 1465 1466 for (i = 0; i < num_wrs; ++i) { 1467 wr->wr.rdma.rkey = sgl->keyed.key; 1468 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1469 for (j = 0; j < wr->num_sge; ++j) { 1470 remote_addr_offset += wr->sg_list[j].length; 1471 } 1472 wr = wr->next; 1473 } 1474 } 1475 1476 static int 1477 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_device *device, 1478 struct spdk_nvmf_rdma_request *rdma_req, 1479 struct ibv_send_wr *wr, 1480 uint32_t total_length) 1481 { 1482 struct spdk_rdma_utils_memory_translation mem_translation; 1483 struct ibv_sge *sg_ele; 1484 struct iovec *iov; 1485 uint32_t lkey, remaining; 1486 int rc; 1487 1488 wr->num_sge = 0; 1489 1490 while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) { 1491 iov = &rdma_req->req.iov[rdma_req->iovpos]; 1492 rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1493 if (spdk_unlikely(rc)) { 1494 return rc; 1495 } 1496 1497 lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation); 1498 sg_ele = &wr->sg_list[wr->num_sge]; 1499 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1500 1501 sg_ele->lkey = lkey; 1502 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset; 1503 sg_ele->length = remaining; 1504 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1505 sg_ele->length); 1506 rdma_req->offset += sg_ele->length; 1507 total_length -= sg_ele->length; 1508 wr->num_sge++; 1509 1510 if (rdma_req->offset == iov->iov_len) { 1511 rdma_req->offset = 0; 1512 rdma_req->iovpos++; 1513 } 1514 } 1515 1516 if (spdk_unlikely(total_length)) { 1517 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1518 return -EINVAL; 1519 } 1520 1521 return 0; 1522 } 1523 1524 static int 1525 nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_device *device, 1526 struct spdk_nvmf_rdma_request *rdma_req, 1527 struct ibv_send_wr *wr, 1528 uint32_t total_length, 1529 uint32_t num_extra_wrs) 1530 { 1531 struct spdk_rdma_utils_memory_translation mem_translation; 1532 struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx; 1533 struct ibv_sge *sg_ele; 1534 struct iovec *iov; 1535 struct iovec *rdma_iov; 1536 uint32_t lkey, remaining; 1537 uint32_t remaining_data_block, data_block_size, md_size; 1538 uint32_t sge_len; 1539 int rc; 1540 1541 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1542 1543 if (spdk_likely(!rdma_req->req.stripped_data)) { 1544 rdma_iov = rdma_req->req.iov; 1545 remaining_data_block = data_block_size; 1546 md_size = dif_ctx->md_size; 1547 } else { 1548 rdma_iov = rdma_req->req.stripped_data->iov; 1549 total_length = total_length / dif_ctx->block_size * data_block_size; 1550 remaining_data_block = total_length; 1551 md_size = 0; 1552 } 1553 1554 wr->num_sge = 0; 1555 1556 while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1557 iov = rdma_iov + rdma_req->iovpos; 1558 rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1559 if (spdk_unlikely(rc)) { 1560 return rc; 1561 } 1562 1563 lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation); 1564 sg_ele = &wr->sg_list[wr->num_sge]; 1565 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1566 1567 while (remaining) { 1568 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1569 if (num_extra_wrs > 0 && wr->next) { 1570 wr = wr->next; 1571 wr->num_sge = 0; 1572 sg_ele = &wr->sg_list[wr->num_sge]; 1573 num_extra_wrs--; 1574 } else { 1575 break; 1576 } 1577 } 1578 sg_ele->lkey = lkey; 1579 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset); 1580 sge_len = spdk_min(remaining, remaining_data_block); 1581 sg_ele->length = sge_len; 1582 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, 1583 sg_ele->addr, sg_ele->length); 1584 remaining -= sge_len; 1585 remaining_data_block -= sge_len; 1586 rdma_req->offset += sge_len; 1587 total_length -= sge_len; 1588 1589 sg_ele++; 1590 wr->num_sge++; 1591 1592 if (remaining_data_block == 0) { 1593 /* skip metadata */ 1594 rdma_req->offset += md_size; 1595 total_length -= md_size; 1596 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1597 remaining -= spdk_min(remaining, md_size); 1598 remaining_data_block = data_block_size; 1599 } 1600 1601 if (remaining == 0) { 1602 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1603 the remaining metadata bits at the beginning of the next buffer */ 1604 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset); 1605 rdma_req->iovpos++; 1606 } 1607 } 1608 } 1609 1610 if (spdk_unlikely(total_length)) { 1611 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1612 return -EINVAL; 1613 } 1614 1615 return 0; 1616 } 1617 1618 static inline uint32_t 1619 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1620 { 1621 /* estimate the number of SG entries and WRs needed to process the request */ 1622 uint32_t num_sge = 0; 1623 uint32_t i; 1624 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1625 1626 for (i = 0; i < num_buffers && length > 0; i++) { 1627 uint32_t buffer_len = spdk_min(length, io_unit_size); 1628 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1629 1630 if (num_sge_in_block * block_size > buffer_len) { 1631 ++num_sge_in_block; 1632 } 1633 num_sge += num_sge_in_block; 1634 length -= buffer_len; 1635 } 1636 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1637 } 1638 1639 static int 1640 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1641 struct spdk_nvmf_rdma_device *device, 1642 struct spdk_nvmf_rdma_request *rdma_req) 1643 { 1644 struct spdk_nvmf_rdma_qpair *rqpair; 1645 struct spdk_nvmf_rdma_poll_group *rgroup; 1646 struct spdk_nvmf_request *req = &rdma_req->req; 1647 struct ibv_send_wr *wr = &rdma_req->data.wr; 1648 int rc; 1649 uint32_t num_wrs = 1; 1650 uint32_t length; 1651 1652 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1653 rgroup = rqpair->poller->group; 1654 1655 /* rdma wr specifics */ 1656 nvmf_rdma_setup_request(rdma_req); 1657 1658 length = req->length; 1659 if (spdk_unlikely(req->dif_enabled)) { 1660 req->dif.orig_length = length; 1661 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1662 req->dif.elba_length = length; 1663 } 1664 1665 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1666 length); 1667 if (spdk_unlikely(rc != 0)) { 1668 return rc; 1669 } 1670 1671 assert(req->iovcnt <= rqpair->max_send_sge); 1672 1673 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1674 * the stripped_buffers are got for DIF stripping. */ 1675 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1676 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1677 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1678 &rtransport->transport, req->dif.orig_length); 1679 if (rc != 0) { 1680 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1681 } 1682 } 1683 1684 rdma_req->iovpos = 0; 1685 1686 if (spdk_unlikely(req->dif_enabled)) { 1687 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1688 req->dif.dif_ctx.block_size); 1689 if (num_wrs > 1) { 1690 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1691 if (spdk_unlikely(rc != 0)) { 1692 goto err_exit; 1693 } 1694 } 1695 1696 rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, wr, length, num_wrs - 1); 1697 if (spdk_unlikely(rc != 0)) { 1698 goto err_exit; 1699 } 1700 1701 if (num_wrs > 1) { 1702 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1703 } 1704 } else { 1705 rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, wr, length); 1706 if (spdk_unlikely(rc != 0)) { 1707 goto err_exit; 1708 } 1709 } 1710 1711 /* set the number of outstanding data WRs for this request. */ 1712 rdma_req->num_outstanding_data_wr = num_wrs; 1713 1714 return rc; 1715 1716 err_exit: 1717 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1718 nvmf_rdma_request_free_data(rdma_req, rtransport); 1719 req->iovcnt = 0; 1720 return rc; 1721 } 1722 1723 static int 1724 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1725 struct spdk_nvmf_rdma_device *device, 1726 struct spdk_nvmf_rdma_request *rdma_req) 1727 { 1728 struct spdk_nvmf_rdma_qpair *rqpair; 1729 struct spdk_nvmf_rdma_poll_group *rgroup; 1730 struct ibv_send_wr *current_wr; 1731 struct spdk_nvmf_request *req = &rdma_req->req; 1732 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1733 uint32_t num_sgl_descriptors; 1734 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0; 1735 uint32_t i; 1736 int rc; 1737 1738 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1739 rgroup = rqpair->poller->group; 1740 1741 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1742 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1743 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1744 1745 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1746 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1747 1748 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1749 for (i = 0; i < num_sgl_descriptors; i++) { 1750 if (spdk_likely(!req->dif_enabled)) { 1751 lengths[i] = desc->keyed.length; 1752 } else { 1753 req->dif.orig_length += desc->keyed.length; 1754 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1755 req->dif.elba_length += lengths[i]; 1756 } 1757 total_length += lengths[i]; 1758 desc++; 1759 } 1760 1761 if (spdk_unlikely(total_length > rtransport->transport.opts.max_io_size)) { 1762 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n", 1763 total_length, rtransport->transport.opts.max_io_size); 1764 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1765 return -EINVAL; 1766 } 1767 1768 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1); 1769 if (spdk_unlikely(rc != 0)) { 1770 return -ENOMEM; 1771 } 1772 1773 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length); 1774 if (spdk_unlikely(rc != 0)) { 1775 nvmf_rdma_request_free_data(rdma_req, rtransport); 1776 return rc; 1777 } 1778 1779 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1780 * the stripped_buffers are got for DIF stripping. */ 1781 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1782 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1783 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1784 &rtransport->transport, req->dif.orig_length); 1785 if (spdk_unlikely(rc != 0)) { 1786 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1787 } 1788 } 1789 1790 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1791 current_wr = &rdma_req->data.wr; 1792 assert(current_wr != NULL); 1793 1794 req->length = 0; 1795 rdma_req->iovpos = 0; 1796 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1797 for (i = 0; i < num_sgl_descriptors; i++) { 1798 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1799 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1800 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1801 rc = -EINVAL; 1802 goto err_exit; 1803 } 1804 1805 if (spdk_likely(!req->dif_enabled)) { 1806 rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, current_wr, lengths[i]); 1807 } else { 1808 rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, current_wr, 1809 lengths[i], 0); 1810 } 1811 if (spdk_unlikely(rc != 0)) { 1812 rc = -ENOMEM; 1813 goto err_exit; 1814 } 1815 1816 req->length += desc->keyed.length; 1817 current_wr->wr.rdma.rkey = desc->keyed.key; 1818 current_wr->wr.rdma.remote_addr = desc->address; 1819 current_wr = current_wr->next; 1820 desc++; 1821 } 1822 1823 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1824 /* Go back to the last descriptor in the list. */ 1825 desc--; 1826 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1827 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1828 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1829 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1830 } 1831 } 1832 #endif 1833 1834 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1835 1836 return 0; 1837 1838 err_exit: 1839 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1840 nvmf_rdma_request_free_data(rdma_req, rtransport); 1841 return rc; 1842 } 1843 1844 static int 1845 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1846 struct spdk_nvmf_rdma_device *device, 1847 struct spdk_nvmf_rdma_request *rdma_req) 1848 { 1849 struct spdk_nvmf_request *req = &rdma_req->req; 1850 struct spdk_nvme_cpl *rsp; 1851 struct spdk_nvme_sgl_descriptor *sgl; 1852 int rc; 1853 uint32_t length; 1854 1855 rsp = &req->rsp->nvme_cpl; 1856 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1857 1858 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1859 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1860 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1861 1862 length = sgl->keyed.length; 1863 if (spdk_unlikely(length > rtransport->transport.opts.max_io_size)) { 1864 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1865 length, rtransport->transport.opts.max_io_size); 1866 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1867 return -1; 1868 } 1869 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1870 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1871 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1872 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1873 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1874 } 1875 } 1876 #endif 1877 1878 /* fill request length and populate iovs */ 1879 req->length = length; 1880 1881 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req); 1882 if (spdk_unlikely(rc < 0)) { 1883 if (rc == -EINVAL) { 1884 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1885 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1886 return -1; 1887 } 1888 /* No available buffers. Queue this request up. */ 1889 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1890 return 0; 1891 } 1892 1893 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1894 req->iovcnt); 1895 1896 return 0; 1897 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1898 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1899 uint64_t offset = sgl->address; 1900 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1901 1902 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1903 offset, sgl->unkeyed.length); 1904 1905 if (spdk_unlikely(offset > max_len)) { 1906 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1907 offset, max_len); 1908 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1909 return -1; 1910 } 1911 max_len -= (uint32_t)offset; 1912 1913 if (spdk_unlikely(sgl->unkeyed.length > max_len)) { 1914 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1915 sgl->unkeyed.length, max_len); 1916 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1917 return -1; 1918 } 1919 1920 rdma_req->num_outstanding_data_wr = 0; 1921 req->data_from_pool = false; 1922 req->length = sgl->unkeyed.length; 1923 1924 req->iov[0].iov_base = rdma_req->recv->buf + offset; 1925 req->iov[0].iov_len = req->length; 1926 req->iovcnt = 1; 1927 1928 return 0; 1929 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1930 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1931 1932 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1933 if (spdk_unlikely(rc == -ENOMEM)) { 1934 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1935 return 0; 1936 } else if (spdk_unlikely(rc == -EINVAL)) { 1937 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1938 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1939 return -1; 1940 } 1941 1942 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1943 req->iovcnt); 1944 1945 return 0; 1946 } 1947 1948 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1949 sgl->generic.type, sgl->generic.subtype); 1950 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1951 return -1; 1952 } 1953 1954 static void 1955 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1956 struct spdk_nvmf_rdma_transport *rtransport) 1957 { 1958 struct spdk_nvmf_rdma_qpair *rqpair; 1959 struct spdk_nvmf_rdma_poll_group *rgroup; 1960 1961 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1962 if (rdma_req->req.data_from_pool) { 1963 rgroup = rqpair->poller->group; 1964 1965 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1966 } 1967 if (rdma_req->req.stripped_data) { 1968 nvmf_request_free_stripped_buffers(&rdma_req->req, 1969 &rqpair->poller->group->group, 1970 &rtransport->transport); 1971 } 1972 nvmf_rdma_request_free_data(rdma_req, rtransport); 1973 rdma_req->req.length = 0; 1974 rdma_req->req.iovcnt = 0; 1975 rdma_req->offset = 0; 1976 rdma_req->req.dif_enabled = false; 1977 rdma_req->fused_failed = false; 1978 rdma_req->transfer_wr = NULL; 1979 if (rdma_req->fused_pair) { 1980 /* This req was part of a valid fused pair, but failed before it got to 1981 * READ_TO_EXECUTE state. This means we need to fail the other request 1982 * in the pair, because it is no longer part of a valid pair. If the pair 1983 * already reached READY_TO_EXECUTE state, we need to kick it. 1984 */ 1985 rdma_req->fused_pair->fused_failed = true; 1986 if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 1987 nvmf_rdma_request_process(rtransport, rdma_req->fused_pair); 1988 } 1989 rdma_req->fused_pair = NULL; 1990 } 1991 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1992 rqpair->qd--; 1993 1994 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1995 rqpair->qpair.queue_depth--; 1996 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1997 } 1998 1999 static void 2000 nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport, 2001 struct spdk_nvmf_rdma_qpair *rqpair, 2002 struct spdk_nvmf_rdma_request *rdma_req) 2003 { 2004 enum spdk_nvme_cmd_fuse last, next; 2005 2006 last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2007 next = rdma_req->req.cmd->nvme_cmd.fuse; 2008 2009 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2010 2011 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2012 return; 2013 } 2014 2015 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2016 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2017 /* This is a valid pair of fused commands. Point them at each other 2018 * so they can be submitted consecutively once ready to be executed. 2019 */ 2020 rqpair->fused_first->fused_pair = rdma_req; 2021 rdma_req->fused_pair = rqpair->fused_first; 2022 rqpair->fused_first = NULL; 2023 return; 2024 } else { 2025 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2026 rqpair->fused_first->fused_failed = true; 2027 2028 /* If the last req is in READY_TO_EXECUTE state, then call 2029 * nvmf_rdma_request_process(), otherwise nothing else will kick it. 2030 */ 2031 if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2032 nvmf_rdma_request_process(rtransport, rqpair->fused_first); 2033 } 2034 2035 rqpair->fused_first = NULL; 2036 } 2037 } 2038 2039 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2040 /* Set rqpair->fused_first here so that we know to check that the next request 2041 * is a SECOND (and to fail this one if it isn't). 2042 */ 2043 rqpair->fused_first = rdma_req; 2044 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2045 /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */ 2046 rdma_req->fused_failed = true; 2047 } 2048 } 2049 2050 static void 2051 nvmf_rdma_poll_group_insert_need_buffer_req(struct spdk_nvmf_rdma_poll_group *rgroup, 2052 struct spdk_nvmf_rdma_request *rdma_req) 2053 { 2054 struct spdk_nvmf_request *r; 2055 2056 /* CONNECT commands have a timeout, so we need to avoid a CONNECT command 2057 * from getting buried behind a long list of other non-FABRIC requests 2058 * waiting for a buffer. Note that even though the CONNECT command's data is 2059 * in-capsule, the request still goes to this STAILQ. 2060 */ 2061 if (spdk_likely(rdma_req->req.cmd->nvme_cmd.opc != SPDK_NVME_OPC_FABRIC)) { 2062 /* This is the most likely case. */ 2063 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2064 return; 2065 } else { 2066 /* STAILQ doesn't have INSERT_BEFORE, so we need to either INSERT_HEAD 2067 * or INSERT_AFTER. Put it after any other FABRIC commands that are 2068 * already in the queue. 2069 */ 2070 r = STAILQ_FIRST(&rgroup->group.pending_buf_queue); 2071 if (r == NULL || r->cmd->nvme_cmd.opc != SPDK_NVME_OPC_FABRIC) { 2072 STAILQ_INSERT_HEAD(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2073 return; 2074 } 2075 while (true) { 2076 struct spdk_nvmf_request *next; 2077 2078 next = STAILQ_NEXT(r, buf_link); 2079 if (next == NULL || next->cmd->nvme_cmd.opc != SPDK_NVME_OPC_FABRIC) { 2080 STAILQ_INSERT_AFTER(&rgroup->group.pending_buf_queue, r, &rdma_req->req, buf_link); 2081 return; 2082 } 2083 r = next; 2084 } 2085 } 2086 } 2087 2088 bool 2089 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 2090 struct spdk_nvmf_rdma_request *rdma_req) 2091 { 2092 struct spdk_nvmf_rdma_qpair *rqpair; 2093 struct spdk_nvmf_rdma_device *device; 2094 struct spdk_nvmf_rdma_poll_group *rgroup; 2095 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2096 int rc; 2097 struct spdk_nvmf_rdma_recv *rdma_recv; 2098 enum spdk_nvmf_rdma_request_state prev_state; 2099 bool progress = false; 2100 int data_posted; 2101 uint32_t num_blocks, num_rdma_reads_available, qdepth; 2102 2103 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2104 device = rqpair->device; 2105 rgroup = rqpair->poller->group; 2106 2107 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 2108 2109 /* If the queue pair is in an error state, force the request to the completed state 2110 * to release resources. */ 2111 if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 2112 switch (rdma_req->state) { 2113 case RDMA_REQUEST_STATE_NEED_BUFFER: 2114 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 2115 break; 2116 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2117 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2118 break; 2119 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2120 if (rdma_req->num_remaining_data_wr) { 2121 /* Partially sent request is still in the pending_rdma_read_queue, 2122 * remove it before completing */ 2123 rdma_req->num_remaining_data_wr = 0; 2124 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2125 } 2126 break; 2127 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2128 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2129 break; 2130 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 2131 STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2132 break; 2133 default: 2134 break; 2135 } 2136 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2137 } 2138 2139 /* The loop here is to allow for several back-to-back state changes. */ 2140 do { 2141 prev_state = rdma_req->state; 2142 2143 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 2144 2145 switch (rdma_req->state) { 2146 case RDMA_REQUEST_STATE_FREE: 2147 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 2148 * to escape this state. */ 2149 break; 2150 case RDMA_REQUEST_STATE_NEW: 2151 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 2152 (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth); 2153 rdma_recv = rdma_req->recv; 2154 2155 /* The first element of the SGL is the NVMe command */ 2156 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 2157 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 2158 rdma_req->transfer_wr = &rdma_req->data.wr; 2159 2160 if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 2161 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2162 break; 2163 } 2164 2165 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 2166 rdma_req->req.dif_enabled = true; 2167 } 2168 2169 nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req); 2170 2171 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2172 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 2173 rdma_req->rsp.wr.imm_data = 0; 2174 #endif 2175 2176 /* The next state transition depends on the data transfer needs of this request. */ 2177 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2178 2179 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2180 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2181 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2182 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2183 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2184 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 2185 break; 2186 } 2187 2188 /* If no data to transfer, ready to execute. */ 2189 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2190 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2191 break; 2192 } 2193 2194 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2195 nvmf_rdma_poll_group_insert_need_buffer_req(rgroup, rdma_req); 2196 break; 2197 case RDMA_REQUEST_STATE_NEED_BUFFER: 2198 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2199 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2200 2201 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2202 2203 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2204 /* This request needs to wait in line to obtain a buffer */ 2205 break; 2206 } 2207 2208 /* Try to get a data buffer */ 2209 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2210 if (spdk_unlikely(rc < 0)) { 2211 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2212 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2213 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2214 break; 2215 } 2216 2217 if (rdma_req->req.iovcnt == 0) { 2218 /* No buffers available. */ 2219 rgroup->stat.pending_data_buffer++; 2220 break; 2221 } 2222 2223 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2224 2225 /* If data is transferring from host to controller and the data didn't 2226 * arrive using in capsule data, we need to do a transfer from the host. 2227 */ 2228 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2229 rdma_req->req.data_from_pool) { 2230 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2231 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2232 break; 2233 } 2234 2235 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2236 break; 2237 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2238 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2239 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2240 2241 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2242 /* This request needs to wait in line to perform RDMA */ 2243 break; 2244 } 2245 assert(rqpair->max_send_depth >= rqpair->current_send_depth); 2246 qdepth = rqpair->max_send_depth - rqpair->current_send_depth; 2247 assert(rqpair->max_read_depth >= rqpair->current_read_depth); 2248 num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth; 2249 if (rdma_req->num_outstanding_data_wr > qdepth || 2250 rdma_req->num_outstanding_data_wr > num_rdma_reads_available) { 2251 if (num_rdma_reads_available && qdepth) { 2252 /* Send as much as we can */ 2253 request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth)); 2254 } else { 2255 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2256 rqpair->poller->stat.pending_rdma_read++; 2257 break; 2258 } 2259 } 2260 2261 /* We have already verified that this request is the head of the queue. */ 2262 if (rdma_req->num_remaining_data_wr == 0) { 2263 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2264 } 2265 2266 request_transfer_in(&rdma_req->req); 2267 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2268 2269 break; 2270 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2271 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2272 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2273 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2274 * to escape this state. */ 2275 break; 2276 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2277 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2278 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2279 2280 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2281 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2282 /* generate DIF for write operation */ 2283 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2284 assert(num_blocks > 0); 2285 2286 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2287 num_blocks, &rdma_req->req.dif.dif_ctx); 2288 if (rc != 0) { 2289 SPDK_ERRLOG("DIF generation failed\n"); 2290 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2291 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 2292 break; 2293 } 2294 } 2295 2296 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2297 /* set extended length before IO operation */ 2298 rdma_req->req.length = rdma_req->req.dif.elba_length; 2299 } 2300 2301 if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 2302 if (rdma_req->fused_failed) { 2303 /* This request failed FUSED semantics. Fail it immediately, without 2304 * even sending it to the target layer. 2305 */ 2306 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2307 rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED; 2308 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2309 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2310 break; 2311 } 2312 2313 if (rdma_req->fused_pair == NULL || 2314 rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2315 /* This request is ready to execute, but either we don't know yet if it's 2316 * valid - i.e. this is a FIRST but we haven't received the next 2317 * request yet or the other request of this fused pair isn't ready to 2318 * execute. So break here and this request will get processed later either 2319 * when the other request is ready or we find that this request isn't valid. 2320 */ 2321 break; 2322 } 2323 } 2324 2325 /* If we get to this point, and this request is a fused command, we know that 2326 * it is part of valid sequence (FIRST followed by a SECOND) and that both 2327 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 2328 * request, and the other request of the fused pair, in the correct order. 2329 * Also clear the ->fused_pair pointers on both requests, since after this point 2330 * we no longer need to maintain the relationship between these two requests. 2331 */ 2332 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 2333 assert(rdma_req->fused_pair != NULL); 2334 assert(rdma_req->fused_pair->fused_pair != NULL); 2335 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2336 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2337 rdma_req->fused_pair->fused_pair = NULL; 2338 rdma_req->fused_pair = NULL; 2339 } 2340 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2341 spdk_nvmf_request_exec(&rdma_req->req); 2342 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 2343 assert(rdma_req->fused_pair != NULL); 2344 assert(rdma_req->fused_pair->fused_pair != NULL); 2345 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2346 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2347 rdma_req->fused_pair->fused_pair = NULL; 2348 rdma_req->fused_pair = NULL; 2349 } 2350 break; 2351 case RDMA_REQUEST_STATE_EXECUTING: 2352 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2353 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2354 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2355 * to escape this state. */ 2356 break; 2357 case RDMA_REQUEST_STATE_EXECUTED: 2358 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2359 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2360 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2361 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2362 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2363 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2364 } else { 2365 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2366 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2367 } 2368 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2369 /* restore the original length */ 2370 rdma_req->req.length = rdma_req->req.dif.orig_length; 2371 2372 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2373 struct spdk_dif_error error_blk; 2374 2375 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2376 if (!rdma_req->req.stripped_data) { 2377 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2378 &rdma_req->req.dif.dif_ctx, &error_blk); 2379 } else { 2380 rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov, 2381 rdma_req->req.stripped_data->iovcnt, 2382 rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2383 &rdma_req->req.dif.dif_ctx, &error_blk); 2384 } 2385 if (rc) { 2386 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2387 2388 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2389 error_blk.err_offset); 2390 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2391 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2392 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2393 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2394 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2395 } 2396 } 2397 } 2398 break; 2399 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2400 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2401 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2402 2403 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2404 /* This request needs to wait in line to perform RDMA */ 2405 break; 2406 } 2407 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2408 rqpair->max_send_depth) { 2409 /* We can only have so many WRs outstanding. we have to wait until some finish. 2410 * +1 since each request has an additional wr in the resp. */ 2411 rqpair->poller->stat.pending_rdma_write++; 2412 break; 2413 } 2414 2415 /* We have already verified that this request is the head of the queue. */ 2416 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2417 2418 /* The data transfer will be kicked off from 2419 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2420 * We verified that data + response fit into send queue, so we can go to the next state directly 2421 */ 2422 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2423 break; 2424 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 2425 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 0, 0, 2426 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2427 2428 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_send_queue)) { 2429 /* This request needs to wait in line to send the completion */ 2430 break; 2431 } 2432 2433 assert(rqpair->current_send_depth <= rqpair->max_send_depth); 2434 if (rqpair->current_send_depth == rqpair->max_send_depth) { 2435 /* We can only have so many WRs outstanding. we have to wait until some finish */ 2436 rqpair->poller->stat.pending_rdma_send++; 2437 break; 2438 } 2439 2440 /* We have already verified that this request is the head of the queue. */ 2441 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_send_queue, state_link); 2442 2443 /* The response sending will be kicked off from 2444 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2445 */ 2446 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2447 break; 2448 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2449 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2450 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2451 rc = request_transfer_out(&rdma_req->req, &data_posted); 2452 assert(rc == 0); /* No good way to handle this currently */ 2453 if (spdk_unlikely(rc)) { 2454 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2455 } else { 2456 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2457 RDMA_REQUEST_STATE_COMPLETING; 2458 } 2459 break; 2460 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2461 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2462 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2463 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2464 * to escape this state. */ 2465 break; 2466 case RDMA_REQUEST_STATE_COMPLETING: 2467 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2468 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2469 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2470 * to escape this state. */ 2471 break; 2472 case RDMA_REQUEST_STATE_COMPLETED: 2473 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2474 (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth); 2475 2476 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2477 _nvmf_rdma_request_free(rdma_req, rtransport); 2478 break; 2479 case RDMA_REQUEST_NUM_STATES: 2480 default: 2481 assert(0); 2482 break; 2483 } 2484 2485 if (rdma_req->state != prev_state) { 2486 progress = true; 2487 } 2488 } while (rdma_req->state != prev_state); 2489 2490 return progress; 2491 } 2492 2493 /* Public API callbacks begin here */ 2494 2495 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2496 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2497 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2498 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2499 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2500 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2501 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2502 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2503 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX 2504 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2505 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2506 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2507 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2508 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2509 #define SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE 4095 2510 2511 static void 2512 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2513 { 2514 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2515 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2516 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2517 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2518 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2519 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2520 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2521 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2522 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2523 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2524 opts->transport_specific = NULL; 2525 opts->data_wr_pool_size = SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE; 2526 } 2527 2528 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2529 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2530 2531 static inline bool 2532 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2533 { 2534 return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD || 2535 device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW; 2536 } 2537 2538 static int nvmf_rdma_accept(void *ctx); 2539 static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport); 2540 static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2541 struct spdk_nvmf_rdma_device *device); 2542 2543 static int 2544 create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context, 2545 struct spdk_nvmf_rdma_device **new_device) 2546 { 2547 struct spdk_nvmf_rdma_device *device; 2548 int flag = 0; 2549 int rc = 0; 2550 2551 device = calloc(1, sizeof(*device)); 2552 if (!device) { 2553 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2554 return -ENOMEM; 2555 } 2556 device->context = context; 2557 rc = ibv_query_device(device->context, &device->attr); 2558 if (rc < 0) { 2559 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2560 free(device); 2561 return rc; 2562 } 2563 2564 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2565 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2566 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2567 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2568 } 2569 2570 /** 2571 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2572 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2573 * but incorrectly reports that it does. There are changes making their way 2574 * through the kernel now that will enable this feature. When they are merged, 2575 * we can conditionally enable this feature. 2576 * 2577 * TODO: enable this for versions of the kernel rxe driver that support it. 2578 */ 2579 if (nvmf_rdma_is_rxe_device(device)) { 2580 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2581 } 2582 #endif 2583 2584 /* set up device context async ev fd as NON_BLOCKING */ 2585 flag = fcntl(device->context->async_fd, F_GETFL); 2586 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2587 if (rc < 0) { 2588 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2589 free(device); 2590 return rc; 2591 } 2592 2593 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2594 SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA transport\n", device); 2595 2596 if (g_nvmf_hooks.get_ibv_pd) { 2597 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2598 } else { 2599 device->pd = ibv_alloc_pd(device->context); 2600 } 2601 2602 if (!device->pd) { 2603 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2604 destroy_ib_device(rtransport, device); 2605 return -ENOMEM; 2606 } 2607 2608 assert(device->map == NULL); 2609 2610 device->map = spdk_rdma_utils_create_mem_map(device->pd, &g_nvmf_hooks, IBV_ACCESS_LOCAL_WRITE); 2611 if (!device->map) { 2612 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2613 destroy_ib_device(rtransport, device); 2614 return -ENOMEM; 2615 } 2616 2617 assert(device->map != NULL); 2618 assert(device->pd != NULL); 2619 2620 if (new_device) { 2621 *new_device = device; 2622 } 2623 SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device), 2624 device, context); 2625 2626 return 0; 2627 } 2628 2629 static void 2630 free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2631 { 2632 if (rtransport->poll_fds) { 2633 free(rtransport->poll_fds); 2634 rtransport->poll_fds = NULL; 2635 } 2636 rtransport->npoll_fds = 0; 2637 } 2638 2639 static int 2640 generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2641 { 2642 /* Set up poll descriptor array to monitor events from RDMA and IB 2643 * in a single poll syscall 2644 */ 2645 int device_count = 0; 2646 int i = 0; 2647 struct spdk_nvmf_rdma_device *device, *tmp; 2648 2649 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2650 device_count++; 2651 } 2652 2653 rtransport->npoll_fds = device_count + 1; 2654 2655 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2656 if (rtransport->poll_fds == NULL) { 2657 SPDK_ERRLOG("poll_fds allocation failed\n"); 2658 return -ENOMEM; 2659 } 2660 2661 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2662 rtransport->poll_fds[i++].events = POLLIN; 2663 2664 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2665 rtransport->poll_fds[i].fd = device->context->async_fd; 2666 rtransport->poll_fds[i++].events = POLLIN; 2667 } 2668 2669 return 0; 2670 } 2671 2672 static struct spdk_nvmf_transport * 2673 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2674 { 2675 int rc; 2676 struct spdk_nvmf_rdma_transport *rtransport; 2677 struct spdk_nvmf_rdma_device *device; 2678 struct ibv_context **contexts; 2679 size_t data_wr_pool_size; 2680 uint32_t i; 2681 int flag; 2682 uint32_t sge_count; 2683 uint32_t min_shared_buffers; 2684 uint32_t min_in_capsule_data_size; 2685 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2686 2687 rtransport = calloc(1, sizeof(*rtransport)); 2688 if (!rtransport) { 2689 return NULL; 2690 } 2691 2692 TAILQ_INIT(&rtransport->devices); 2693 TAILQ_INIT(&rtransport->ports); 2694 TAILQ_INIT(&rtransport->poll_groups); 2695 TAILQ_INIT(&rtransport->retry_ports); 2696 2697 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2698 rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2699 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2700 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2701 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2702 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2703 if (opts->transport_specific != NULL && 2704 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2705 SPDK_COUNTOF(rdma_transport_opts_decoder), 2706 &rtransport->rdma_opts)) { 2707 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2708 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2709 return NULL; 2710 } 2711 2712 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2713 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2714 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2715 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2716 " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d," 2717 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2718 opts->max_queue_depth, 2719 opts->max_io_size, 2720 opts->max_qpairs_per_ctrlr - 1, 2721 opts->io_unit_size, 2722 opts->in_capsule_data_size, 2723 opts->max_aq_depth, 2724 opts->num_shared_buffers, 2725 rtransport->rdma_opts.num_cqe, 2726 rtransport->rdma_opts.max_srq_depth, 2727 rtransport->rdma_opts.no_srq, 2728 rtransport->rdma_opts.acceptor_backlog, 2729 rtransport->rdma_opts.no_wr_batching, 2730 opts->abort_timeout_sec); 2731 2732 /* I/O unit size cannot be larger than max I/O size */ 2733 if (opts->io_unit_size > opts->max_io_size) { 2734 opts->io_unit_size = opts->max_io_size; 2735 } 2736 2737 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2738 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2739 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2740 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2741 } 2742 2743 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2744 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2745 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2746 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2747 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2748 return NULL; 2749 } 2750 2751 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 2752 if (opts->buf_cache_size < UINT32_MAX) { 2753 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2754 if (min_shared_buffers > opts->num_shared_buffers) { 2755 SPDK_ERRLOG("There are not enough buffers to satisfy" 2756 "per-poll group caches for each thread. (%" PRIu32 ")" 2757 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2758 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2759 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2760 return NULL; 2761 } 2762 } 2763 2764 sge_count = opts->max_io_size / opts->io_unit_size; 2765 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2766 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2767 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2768 return NULL; 2769 } 2770 2771 min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES; 2772 if (opts->in_capsule_data_size < min_in_capsule_data_size) { 2773 SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n", 2774 min_in_capsule_data_size); 2775 opts->in_capsule_data_size = min_in_capsule_data_size; 2776 } 2777 2778 rtransport->event_channel = rdma_create_event_channel(); 2779 if (rtransport->event_channel == NULL) { 2780 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2781 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2782 return NULL; 2783 } 2784 2785 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2786 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2787 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2788 rtransport->event_channel->fd, spdk_strerror(errno)); 2789 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2790 return NULL; 2791 } 2792 2793 data_wr_pool_size = opts->data_wr_pool_size; 2794 if (data_wr_pool_size < SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count()) { 2795 data_wr_pool_size = SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count(); 2796 SPDK_NOTICELOG("data_wr_pool_size is changed to %zu to guarantee enough cache for handling " 2797 "at least one IO in each core\n", data_wr_pool_size); 2798 } 2799 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", data_wr_pool_size, 2800 sizeof(struct spdk_nvmf_rdma_request_data), SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2801 SPDK_ENV_NUMA_ID_ANY); 2802 if (!rtransport->data_wr_pool) { 2803 if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) { 2804 SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n"); 2805 SPDK_ERRLOG("Probably running in multiprocess environment, which is " 2806 "unsupported by the nvmf library\n"); 2807 } else { 2808 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2809 } 2810 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2811 return NULL; 2812 } 2813 2814 contexts = rdma_get_devices(NULL); 2815 if (contexts == NULL) { 2816 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2817 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2818 return NULL; 2819 } 2820 2821 i = 0; 2822 rc = 0; 2823 while (contexts[i] != NULL) { 2824 rc = create_ib_device(rtransport, contexts[i], &device); 2825 if (rc < 0) { 2826 break; 2827 } 2828 i++; 2829 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2830 device->is_ready = true; 2831 } 2832 rdma_free_devices(contexts); 2833 2834 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2835 /* divide and round up. */ 2836 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2837 2838 /* round up to the nearest 4k. */ 2839 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2840 2841 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2842 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2843 opts->io_unit_size); 2844 } 2845 2846 if (rc < 0) { 2847 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2848 return NULL; 2849 } 2850 2851 rc = generate_poll_fds(rtransport); 2852 if (rc < 0) { 2853 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2854 return NULL; 2855 } 2856 2857 rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport, 2858 opts->acceptor_poll_rate); 2859 if (!rtransport->accept_poller) { 2860 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2861 return NULL; 2862 } 2863 2864 return &rtransport->transport; 2865 } 2866 2867 static void 2868 destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2869 struct spdk_nvmf_rdma_device *device) 2870 { 2871 TAILQ_REMOVE(&rtransport->devices, device, link); 2872 spdk_rdma_utils_free_mem_map(&device->map); 2873 if (device->pd) { 2874 if (!g_nvmf_hooks.get_ibv_pd) { 2875 ibv_dealloc_pd(device->pd); 2876 } 2877 } 2878 SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device); 2879 free(device); 2880 } 2881 2882 static void 2883 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2884 { 2885 struct spdk_nvmf_rdma_transport *rtransport; 2886 assert(w != NULL); 2887 2888 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2889 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2890 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2891 if (rtransport->rdma_opts.no_srq == true) { 2892 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2893 } 2894 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2895 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2896 } 2897 2898 static int 2899 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2900 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2901 { 2902 struct spdk_nvmf_rdma_transport *rtransport; 2903 struct spdk_nvmf_rdma_port *port, *port_tmp; 2904 struct spdk_nvmf_rdma_device *device, *device_tmp; 2905 2906 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2907 2908 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) { 2909 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 2910 free(port); 2911 } 2912 2913 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2914 TAILQ_REMOVE(&rtransport->ports, port, link); 2915 rdma_destroy_id(port->id); 2916 free(port); 2917 } 2918 2919 free_poll_fds(rtransport); 2920 2921 if (rtransport->event_channel != NULL) { 2922 rdma_destroy_event_channel(rtransport->event_channel); 2923 } 2924 2925 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2926 destroy_ib_device(rtransport, device); 2927 } 2928 2929 if (rtransport->data_wr_pool != NULL) { 2930 if (spdk_mempool_count(rtransport->data_wr_pool) != transport->opts.data_wr_pool_size) { 2931 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2932 spdk_mempool_count(rtransport->data_wr_pool), 2933 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2934 } 2935 } 2936 2937 spdk_mempool_free(rtransport->data_wr_pool); 2938 2939 spdk_poller_unregister(&rtransport->accept_poller); 2940 free(rtransport); 2941 2942 if (cb_fn) { 2943 cb_fn(cb_arg); 2944 } 2945 return 0; 2946 } 2947 2948 static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2949 struct spdk_nvme_transport_id *trid, 2950 bool peer); 2951 2952 static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport); 2953 2954 static int 2955 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2956 struct spdk_nvmf_listen_opts *listen_opts) 2957 { 2958 struct spdk_nvmf_rdma_transport *rtransport; 2959 struct spdk_nvmf_rdma_device *device; 2960 struct spdk_nvmf_rdma_port *port, *tmp_port; 2961 struct addrinfo *res; 2962 struct addrinfo hints; 2963 int family; 2964 int rc; 2965 long int port_val; 2966 bool is_retry = false; 2967 2968 if (!strlen(trid->trsvcid)) { 2969 SPDK_ERRLOG("Service id is required\n"); 2970 return -EINVAL; 2971 } 2972 2973 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2974 assert(rtransport->event_channel != NULL); 2975 2976 port = calloc(1, sizeof(*port)); 2977 if (!port) { 2978 SPDK_ERRLOG("Port allocation failed\n"); 2979 return -ENOMEM; 2980 } 2981 2982 port->trid = trid; 2983 2984 switch (trid->adrfam) { 2985 case SPDK_NVMF_ADRFAM_IPV4: 2986 family = AF_INET; 2987 break; 2988 case SPDK_NVMF_ADRFAM_IPV6: 2989 family = AF_INET6; 2990 break; 2991 default: 2992 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2993 free(port); 2994 return -EINVAL; 2995 } 2996 2997 memset(&hints, 0, sizeof(hints)); 2998 hints.ai_family = family; 2999 hints.ai_flags = AI_NUMERICSERV; 3000 hints.ai_socktype = SOCK_STREAM; 3001 hints.ai_protocol = 0; 3002 3003 /* Range check the trsvcid. Fail in 3 cases: 3004 * < 0: means that spdk_strtol hit an error 3005 * 0: this results in ephemeral port which we don't want 3006 * > 65535: port too high 3007 */ 3008 port_val = spdk_strtol(trid->trsvcid, 10); 3009 if (port_val <= 0 || port_val > 65535) { 3010 SPDK_ERRLOG("invalid trsvcid %s\n", trid->trsvcid); 3011 free(port); 3012 return -EINVAL; 3013 } 3014 3015 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 3016 if (rc) { 3017 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 3018 free(port); 3019 return -(abs(rc)); 3020 } 3021 3022 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 3023 if (rc < 0) { 3024 SPDK_ERRLOG("rdma_create_id() failed\n"); 3025 freeaddrinfo(res); 3026 free(port); 3027 return rc; 3028 } 3029 3030 rc = rdma_bind_addr(port->id, res->ai_addr); 3031 freeaddrinfo(res); 3032 3033 if (rc < 0) { 3034 TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) { 3035 if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) { 3036 is_retry = true; 3037 break; 3038 } 3039 } 3040 if (!is_retry) { 3041 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 3042 } 3043 rdma_destroy_id(port->id); 3044 free(port); 3045 return rc; 3046 } 3047 3048 if (!port->id->verbs) { 3049 SPDK_ERRLOG("ibv_context is null\n"); 3050 rdma_destroy_id(port->id); 3051 free(port); 3052 return -1; 3053 } 3054 3055 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 3056 if (rc < 0) { 3057 SPDK_ERRLOG("rdma_listen() failed\n"); 3058 rdma_destroy_id(port->id); 3059 free(port); 3060 return rc; 3061 } 3062 3063 TAILQ_FOREACH(device, &rtransport->devices, link) { 3064 if (device->context == port->id->verbs && device->is_ready) { 3065 port->device = device; 3066 break; 3067 } 3068 } 3069 if (!port->device) { 3070 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 3071 port->id->verbs); 3072 rdma_destroy_id(port->id); 3073 free(port); 3074 nvmf_rdma_rescan_devices(rtransport); 3075 return -EINVAL; 3076 } 3077 3078 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 3079 trid->traddr, trid->trsvcid); 3080 3081 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 3082 return 0; 3083 } 3084 3085 static void 3086 nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport, 3087 const struct spdk_nvme_transport_id *trid, bool need_retry) 3088 { 3089 struct spdk_nvmf_rdma_transport *rtransport; 3090 struct spdk_nvmf_rdma_port *port, *tmp; 3091 3092 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3093 3094 if (!need_retry) { 3095 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) { 3096 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 3097 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3098 free(port); 3099 } 3100 } 3101 } 3102 3103 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 3104 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 3105 SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n", 3106 port->trid->traddr, port->trid->trsvcid, need_retry); 3107 TAILQ_REMOVE(&rtransport->ports, port, link); 3108 rdma_destroy_id(port->id); 3109 port->id = NULL; 3110 port->device = NULL; 3111 if (need_retry) { 3112 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3113 } else { 3114 free(port); 3115 } 3116 break; 3117 } 3118 } 3119 } 3120 3121 static void 3122 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 3123 const struct spdk_nvme_transport_id *trid) 3124 { 3125 nvmf_rdma_stop_listen_ex(transport, trid, false); 3126 } 3127 3128 static void _nvmf_rdma_register_poller_in_group(void *c); 3129 static void _nvmf_rdma_remove_poller_in_group(void *c); 3130 3131 static bool 3132 nvmf_rdma_all_pollers_management_done(void *c) 3133 { 3134 struct poller_manage_ctx *ctx = c; 3135 int counter; 3136 3137 counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST); 3138 SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n", 3139 counter, ctx->rpoller); 3140 3141 if (counter == 0) { 3142 free((void *)ctx->inflight_op_counter); 3143 } 3144 free(ctx); 3145 3146 return counter == 0; 3147 } 3148 3149 static int 3150 nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport, 3151 struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add) 3152 { 3153 struct spdk_nvmf_rdma_poll_group *rgroup; 3154 struct spdk_nvmf_rdma_poller *rpoller; 3155 struct spdk_nvmf_poll_group *poll_group; 3156 struct poller_manage_ctx *ctx; 3157 bool found; 3158 int *inflight_counter; 3159 spdk_msg_fn do_fn; 3160 3161 *has_inflight = false; 3162 do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group; 3163 inflight_counter = calloc(1, sizeof(int)); 3164 if (!inflight_counter) { 3165 SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n"); 3166 return -ENOMEM; 3167 } 3168 3169 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3170 (*inflight_counter)++; 3171 } 3172 3173 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3174 found = false; 3175 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3176 if (rpoller->device == device) { 3177 found = true; 3178 break; 3179 } 3180 } 3181 if (found == is_add) { 3182 __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST); 3183 continue; 3184 } 3185 3186 ctx = calloc(1, sizeof(struct poller_manage_ctx)); 3187 if (!ctx) { 3188 SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n"); 3189 if (!*has_inflight) { 3190 free(inflight_counter); 3191 } 3192 return -ENOMEM; 3193 } 3194 3195 ctx->rtransport = rtransport; 3196 ctx->rgroup = rgroup; 3197 ctx->rpoller = rpoller; 3198 ctx->device = device; 3199 ctx->thread = spdk_get_thread(); 3200 ctx->inflight_op_counter = inflight_counter; 3201 *has_inflight = true; 3202 3203 poll_group = rgroup->group.group; 3204 if (poll_group->thread != spdk_get_thread()) { 3205 spdk_thread_send_msg(poll_group->thread, do_fn, ctx); 3206 } else { 3207 do_fn(ctx); 3208 } 3209 } 3210 3211 if (!*has_inflight) { 3212 free(inflight_counter); 3213 } 3214 3215 return 0; 3216 } 3217 3218 static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3219 struct spdk_nvmf_rdma_device *device); 3220 3221 static struct spdk_nvmf_rdma_device * 3222 nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 3223 struct ibv_context *context) 3224 { 3225 struct spdk_nvmf_rdma_device *device, *tmp_device; 3226 3227 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) { 3228 if (device->need_destroy) { 3229 continue; 3230 } 3231 3232 if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) { 3233 return device; 3234 } 3235 } 3236 3237 return NULL; 3238 } 3239 3240 static bool 3241 nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport, 3242 struct ibv_context *context) 3243 { 3244 struct spdk_nvmf_rdma_device *old_device, *new_device; 3245 int rc = 0; 3246 bool has_inflight; 3247 3248 old_device = nvmf_rdma_find_ib_device(rtransport, context); 3249 3250 if (old_device) { 3251 if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) { 3252 /* context may not have time to be cleaned when rescan. exactly one context 3253 * is valid for a device so this context must be invalid and just remove it. */ 3254 SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context); 3255 old_device->need_destroy = true; 3256 nvmf_rdma_handle_device_removal(rtransport, old_device); 3257 } 3258 return false; 3259 } 3260 3261 rc = create_ib_device(rtransport, context, &new_device); 3262 /* TODO: update transport opts. */ 3263 if (rc < 0) { 3264 SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n", 3265 ibv_get_device_name(context->device), context); 3266 return false; 3267 } 3268 3269 rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true); 3270 if (rc < 0) { 3271 SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n", 3272 ibv_get_device_name(context->device), context); 3273 return false; 3274 } 3275 3276 if (has_inflight) { 3277 new_device->is_ready = true; 3278 } 3279 3280 return true; 3281 } 3282 3283 static bool 3284 nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport) 3285 { 3286 struct spdk_nvmf_rdma_device *device; 3287 struct ibv_device **ibv_device_list = NULL; 3288 struct ibv_context **contexts = NULL; 3289 int i = 0; 3290 int num_dev = 0; 3291 bool new_create = false, has_new_device = false; 3292 struct ibv_context *tmp_verbs = NULL; 3293 3294 /* do not rescan when any device is destroying, or context may be freed when 3295 * regenerating the poll fds. 3296 */ 3297 TAILQ_FOREACH(device, &rtransport->devices, link) { 3298 if (device->need_destroy) { 3299 return false; 3300 } 3301 } 3302 3303 ibv_device_list = ibv_get_device_list(&num_dev); 3304 3305 /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be 3306 * marked as dead verbs and never be init again. So we need to make sure the 3307 * verbs is available before we call rdma_get_devices. */ 3308 if (num_dev >= 0) { 3309 for (i = 0; i < num_dev; i++) { 3310 tmp_verbs = ibv_open_device(ibv_device_list[i]); 3311 if (!tmp_verbs) { 3312 SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno); 3313 break; 3314 } 3315 if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) { 3316 SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i], 3317 tmp_verbs->device->dev_name); 3318 has_new_device = true; 3319 } 3320 ibv_close_device(tmp_verbs); 3321 } 3322 ibv_free_device_list(ibv_device_list); 3323 if (!tmp_verbs || !has_new_device) { 3324 return false; 3325 } 3326 } 3327 3328 contexts = rdma_get_devices(NULL); 3329 3330 for (i = 0; contexts && contexts[i] != NULL; i++) { 3331 new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]); 3332 } 3333 3334 if (new_create) { 3335 free_poll_fds(rtransport); 3336 generate_poll_fds(rtransport); 3337 } 3338 3339 if (contexts) { 3340 rdma_free_devices(contexts); 3341 } 3342 3343 return new_create; 3344 } 3345 3346 static bool 3347 nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport) 3348 { 3349 struct spdk_nvmf_rdma_port *port, *tmp_port; 3350 int rc = 0; 3351 bool new_create = false; 3352 3353 if (TAILQ_EMPTY(&rtransport->retry_ports)) { 3354 return false; 3355 } 3356 3357 new_create = nvmf_rdma_rescan_devices(rtransport); 3358 3359 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) { 3360 rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL); 3361 3362 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3363 if (rc) { 3364 if (new_create) { 3365 SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n", 3366 port->trid->traddr, port->trid->trsvcid, rc); 3367 } 3368 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3369 break; 3370 } else { 3371 SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid); 3372 free(port); 3373 } 3374 } 3375 3376 return true; 3377 } 3378 3379 static void 3380 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 3381 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 3382 { 3383 struct spdk_nvmf_request *req, *tmp; 3384 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 3385 struct spdk_nvmf_rdma_resources *resources; 3386 3387 /* First process requests which are waiting for response to be sent */ 3388 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_send_queue, state_link, req_tmp) { 3389 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3390 break; 3391 } 3392 } 3393 3394 /* We process I/O in the data transfer pending queue at the highest priority. */ 3395 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 3396 if (rdma_req->state != RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 3397 /* Requests in this queue might be in state RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 3398 * they are transmitting data over network but we keep them in the list to guarantee 3399 * fair processing. */ 3400 continue; 3401 } 3402 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3403 break; 3404 } 3405 } 3406 3407 /* Then RDMA writes since reads have stronger restrictions than writes */ 3408 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 3409 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3410 break; 3411 } 3412 } 3413 3414 /* Then we handle request waiting on memory buffers. */ 3415 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 3416 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3417 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3418 break; 3419 } 3420 } 3421 3422 resources = rqpair->resources; 3423 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 3424 rdma_req = STAILQ_FIRST(&resources->free_queue); 3425 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 3426 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 3427 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 3428 3429 if (rqpair->srq != NULL) { 3430 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 3431 rdma_req->recv->qpair->qd++; 3432 } else { 3433 rqpair->qd++; 3434 } 3435 3436 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 3437 rdma_req->state = RDMA_REQUEST_STATE_NEW; 3438 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 3439 break; 3440 } 3441 } 3442 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 3443 rqpair->poller->stat.pending_free_request++; 3444 } 3445 } 3446 3447 static void 3448 nvmf_rdma_poller_process_pending_buf_queue(struct spdk_nvmf_rdma_transport *rtransport, 3449 struct spdk_nvmf_rdma_poller *rpoller) 3450 { 3451 struct spdk_nvmf_request *req, *tmp; 3452 struct spdk_nvmf_rdma_request *rdma_req; 3453 3454 STAILQ_FOREACH_SAFE(req, &rpoller->group->group.pending_buf_queue, buf_link, tmp) { 3455 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3456 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 3457 break; 3458 } 3459 } 3460 } 3461 3462 static inline bool 3463 nvmf_rdma_device_supports_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 3464 { 3465 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 3466 return !nvmf_rdma_is_rxe_device(device) && 3467 device->context->device->transport_type != IBV_TRANSPORT_IWARP; 3468 } 3469 3470 static void 3471 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 3472 { 3473 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3474 struct spdk_nvmf_rdma_transport, transport); 3475 3476 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3477 3478 /* nvmf_rdma_close_qpair is not called */ 3479 if (!rqpair->to_close) { 3480 return; 3481 } 3482 3483 /* device is already destroyed and we should force destroy this qpair. */ 3484 if (rqpair->poller && rqpair->poller->need_destroy) { 3485 nvmf_rdma_qpair_destroy(rqpair); 3486 return; 3487 } 3488 3489 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 3490 if (rqpair->current_send_depth != 0) { 3491 return; 3492 } 3493 3494 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 3495 return; 3496 } 3497 3498 /* For devices that support LAST_WQE_REACHED with srq, we need to 3499 * wait to destroy the qpair until that event has been received. 3500 */ 3501 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 3502 nvmf_rdma_device_supports_last_wqe_reached(rqpair->device)) { 3503 return; 3504 } 3505 3506 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED || 3507 rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 3508 3509 nvmf_rdma_qpair_destroy(rqpair); 3510 } 3511 3512 static int 3513 nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked) 3514 { 3515 struct spdk_nvmf_qpair *qpair; 3516 struct spdk_nvmf_rdma_qpair *rqpair; 3517 3518 if (evt->id == NULL) { 3519 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 3520 return -1; 3521 } 3522 3523 qpair = evt->id->context; 3524 if (qpair == NULL) { 3525 SPDK_ERRLOG("disconnect request: no active connection\n"); 3526 return -1; 3527 } 3528 3529 rdma_ack_cm_event(evt); 3530 *event_acked = true; 3531 3532 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3533 3534 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair); 3535 3536 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3537 3538 return 0; 3539 } 3540 3541 #ifdef DEBUG 3542 static const char *CM_EVENT_STR[] = { 3543 "RDMA_CM_EVENT_ADDR_RESOLVED", 3544 "RDMA_CM_EVENT_ADDR_ERROR", 3545 "RDMA_CM_EVENT_ROUTE_RESOLVED", 3546 "RDMA_CM_EVENT_ROUTE_ERROR", 3547 "RDMA_CM_EVENT_CONNECT_REQUEST", 3548 "RDMA_CM_EVENT_CONNECT_RESPONSE", 3549 "RDMA_CM_EVENT_CONNECT_ERROR", 3550 "RDMA_CM_EVENT_UNREACHABLE", 3551 "RDMA_CM_EVENT_REJECTED", 3552 "RDMA_CM_EVENT_ESTABLISHED", 3553 "RDMA_CM_EVENT_DISCONNECTED", 3554 "RDMA_CM_EVENT_DEVICE_REMOVAL", 3555 "RDMA_CM_EVENT_MULTICAST_JOIN", 3556 "RDMA_CM_EVENT_MULTICAST_ERROR", 3557 "RDMA_CM_EVENT_ADDR_CHANGE", 3558 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 3559 }; 3560 #endif /* DEBUG */ 3561 3562 static void 3563 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 3564 struct spdk_nvmf_rdma_port *port) 3565 { 3566 struct spdk_nvmf_rdma_poll_group *rgroup; 3567 struct spdk_nvmf_rdma_poller *rpoller; 3568 struct spdk_nvmf_rdma_qpair *rqpair; 3569 3570 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3571 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3572 RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) { 3573 if (rqpair->listen_id == port->id) { 3574 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3575 } 3576 } 3577 } 3578 } 3579 } 3580 3581 static bool 3582 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 3583 struct rdma_cm_event *event) 3584 { 3585 const struct spdk_nvme_transport_id *trid; 3586 struct spdk_nvmf_rdma_port *port; 3587 struct spdk_nvmf_rdma_transport *rtransport; 3588 bool event_acked = false; 3589 3590 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3591 TAILQ_FOREACH(port, &rtransport->ports, link) { 3592 if (port->id == event->id) { 3593 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 3594 rdma_ack_cm_event(event); 3595 event_acked = true; 3596 trid = port->trid; 3597 break; 3598 } 3599 } 3600 3601 if (event_acked) { 3602 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 3603 3604 nvmf_rdma_stop_listen(transport, trid); 3605 nvmf_rdma_listen(transport, trid, NULL); 3606 } 3607 3608 return event_acked; 3609 } 3610 3611 static void 3612 nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3613 struct spdk_nvmf_rdma_device *device) 3614 { 3615 struct spdk_nvmf_rdma_port *port, *port_tmp; 3616 int rc; 3617 bool has_inflight; 3618 3619 rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false); 3620 if (rc) { 3621 SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc); 3622 return; 3623 } 3624 3625 if (!has_inflight) { 3626 /* no pollers, destroy the device */ 3627 device->ready_to_destroy = true; 3628 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport); 3629 } 3630 3631 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 3632 if (port->device == device) { 3633 SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n", 3634 port->trid->traddr, 3635 port->trid->trsvcid, 3636 ibv_get_device_name(port->device->context->device)); 3637 3638 /* keep NVMF listener and only destroy structures of the 3639 * RDMA transport. when the device comes back we can retry listening 3640 * and the application's workflow will not be interrupted. 3641 */ 3642 nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true); 3643 } 3644 } 3645 } 3646 3647 static void 3648 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 3649 struct rdma_cm_event *event) 3650 { 3651 struct spdk_nvmf_rdma_port *port, *tmp_port; 3652 struct spdk_nvmf_rdma_transport *rtransport; 3653 3654 port = event->id->context; 3655 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3656 3657 rdma_ack_cm_event(event); 3658 3659 /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive 3660 * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that 3661 * we are handling a port event here. 3662 */ 3663 TAILQ_FOREACH(tmp_port, &rtransport->ports, link) { 3664 if (port == tmp_port && port->device && !port->device->need_destroy) { 3665 port->device->need_destroy = true; 3666 nvmf_rdma_handle_device_removal(rtransport, port->device); 3667 } 3668 } 3669 } 3670 3671 static void 3672 nvmf_process_cm_events(struct spdk_nvmf_transport *transport, uint32_t max_events) 3673 { 3674 struct spdk_nvmf_rdma_transport *rtransport; 3675 struct rdma_cm_event *event; 3676 uint32_t i; 3677 int rc; 3678 bool event_acked; 3679 3680 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3681 3682 if (rtransport->event_channel == NULL) { 3683 return; 3684 } 3685 3686 for (i = 0; i < max_events; i++) { 3687 event_acked = false; 3688 rc = rdma_get_cm_event(rtransport->event_channel, &event); 3689 if (rc) { 3690 if (errno != EAGAIN && errno != EWOULDBLOCK) { 3691 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 3692 } 3693 break; 3694 } 3695 3696 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 3697 3698 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 3699 3700 switch (event->event) { 3701 case RDMA_CM_EVENT_ADDR_RESOLVED: 3702 case RDMA_CM_EVENT_ADDR_ERROR: 3703 case RDMA_CM_EVENT_ROUTE_RESOLVED: 3704 case RDMA_CM_EVENT_ROUTE_ERROR: 3705 /* No action required. The target never attempts to resolve routes. */ 3706 break; 3707 case RDMA_CM_EVENT_CONNECT_REQUEST: 3708 rc = nvmf_rdma_connect(transport, event); 3709 if (rc < 0) { 3710 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 3711 break; 3712 } 3713 break; 3714 case RDMA_CM_EVENT_CONNECT_RESPONSE: 3715 /* The target never initiates a new connection. So this will not occur. */ 3716 break; 3717 case RDMA_CM_EVENT_CONNECT_ERROR: 3718 /* Can this happen? The docs say it can, but not sure what causes it. */ 3719 break; 3720 case RDMA_CM_EVENT_UNREACHABLE: 3721 case RDMA_CM_EVENT_REJECTED: 3722 /* These only occur on the client side. */ 3723 break; 3724 case RDMA_CM_EVENT_ESTABLISHED: 3725 /* TODO: Should we be waiting for this event anywhere? */ 3726 break; 3727 case RDMA_CM_EVENT_DISCONNECTED: 3728 rc = nvmf_rdma_disconnect(event, &event_acked); 3729 if (rc < 0) { 3730 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3731 break; 3732 } 3733 break; 3734 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3735 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3736 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3737 * Once these events are sent to SPDK, we should release all IB resources and 3738 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3739 * ibv_destroy* functions to release user space memory allocated by IB. All kernel 3740 * resources are already cleaned. */ 3741 if (event->id->qp) { 3742 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3743 * corresponding qpair. Otherwise the event refers to a listening device. */ 3744 rc = nvmf_rdma_disconnect(event, &event_acked); 3745 if (rc < 0) { 3746 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3747 break; 3748 } 3749 } else { 3750 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3751 event_acked = true; 3752 } 3753 break; 3754 case RDMA_CM_EVENT_MULTICAST_JOIN: 3755 case RDMA_CM_EVENT_MULTICAST_ERROR: 3756 /* Multicast is not used */ 3757 break; 3758 case RDMA_CM_EVENT_ADDR_CHANGE: 3759 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3760 break; 3761 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3762 /* For now, do nothing. The target never re-uses queue pairs. */ 3763 break; 3764 default: 3765 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3766 break; 3767 } 3768 if (!event_acked) { 3769 rdma_ack_cm_event(event); 3770 } 3771 } 3772 } 3773 3774 static void 3775 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3776 { 3777 rqpair->last_wqe_reached = true; 3778 nvmf_rdma_destroy_drained_qpair(rqpair); 3779 } 3780 3781 static void 3782 nvmf_rdma_qpair_process_last_wqe_event(void *ctx) 3783 { 3784 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3785 struct spdk_nvmf_rdma_qpair *rqpair; 3786 3787 rqpair = event_ctx->rqpair; 3788 3789 if (rqpair) { 3790 assert(event_ctx == rqpair->last_wqe_reached_ctx); 3791 rqpair->last_wqe_reached_ctx = NULL; 3792 nvmf_rdma_handle_last_wqe_reached(rqpair); 3793 } 3794 free(event_ctx); 3795 } 3796 3797 static int 3798 nvmf_rdma_send_qpair_last_wqe_event(struct spdk_nvmf_rdma_qpair *rqpair) 3799 { 3800 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3801 struct spdk_thread *thr = NULL; 3802 int rc; 3803 3804 if (rqpair->qpair.group) { 3805 thr = rqpair->qpair.group->thread; 3806 } else if (rqpair->destruct_channel) { 3807 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3808 } 3809 3810 if (!thr) { 3811 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3812 return -EINVAL; 3813 } 3814 3815 if (rqpair->last_wqe_reached || rqpair->last_wqe_reached_ctx != NULL) { 3816 SPDK_ERRLOG("LAST_WQE_REACHED already received for rqpair %p\n", rqpair); 3817 return -EALREADY; 3818 } 3819 3820 ctx = calloc(1, sizeof(*ctx)); 3821 if (!ctx) { 3822 return -ENOMEM; 3823 } 3824 3825 ctx->rqpair = rqpair; 3826 rqpair->last_wqe_reached_ctx = ctx; 3827 3828 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_last_wqe_event, ctx); 3829 if (rc) { 3830 rqpair->last_wqe_reached_ctx = NULL; 3831 free(ctx); 3832 } 3833 3834 return rc; 3835 } 3836 3837 static int 3838 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3839 { 3840 int rc; 3841 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3842 struct ibv_async_event event; 3843 3844 rc = ibv_get_async_event(device->context, &event); 3845 3846 if (rc) { 3847 /* In non-blocking mode -1 means there are no events available */ 3848 return rc; 3849 } 3850 3851 switch (event.event_type) { 3852 case IBV_EVENT_QP_FATAL: 3853 case IBV_EVENT_QP_LAST_WQE_REACHED: 3854 case IBV_EVENT_QP_REQ_ERR: 3855 case IBV_EVENT_QP_ACCESS_ERR: 3856 case IBV_EVENT_COMM_EST: 3857 case IBV_EVENT_PATH_MIG: 3858 case IBV_EVENT_PATH_MIG_ERR: 3859 rqpair = event.element.qp->qp_context; 3860 if (!rqpair) { 3861 /* Any QP event for NVMe-RDMA initiator may be returned. */ 3862 SPDK_NOTICELOG("Async QP event for unknown QP: %s\n", 3863 ibv_event_type_str(event.event_type)); 3864 break; 3865 } 3866 3867 switch (event.event_type) { 3868 case IBV_EVENT_QP_FATAL: 3869 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3870 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3871 (uintptr_t)rqpair, event.event_type); 3872 rqpair->ibv_in_error_state = true; 3873 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3874 break; 3875 case IBV_EVENT_QP_LAST_WQE_REACHED: 3876 /* This event only occurs for shared receive queues. */ 3877 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3878 rc = nvmf_rdma_send_qpair_last_wqe_event(rqpair); 3879 if (rc) { 3880 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3881 rqpair->last_wqe_reached = true; 3882 } 3883 break; 3884 case IBV_EVENT_QP_REQ_ERR: 3885 case IBV_EVENT_QP_ACCESS_ERR: 3886 case IBV_EVENT_COMM_EST: 3887 case IBV_EVENT_PATH_MIG: 3888 case IBV_EVENT_PATH_MIG_ERR: 3889 SPDK_NOTICELOG("Async QP event: %s\n", 3890 ibv_event_type_str(event.event_type)); 3891 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3892 (uintptr_t)rqpair, event.event_type); 3893 rqpair->ibv_in_error_state = true; 3894 break; 3895 default: 3896 break; 3897 } 3898 break; 3899 case IBV_EVENT_DEVICE_FATAL: 3900 SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n", 3901 ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device); 3902 device->need_destroy = true; 3903 break; 3904 case IBV_EVENT_CQ_ERR: 3905 case IBV_EVENT_PORT_ACTIVE: 3906 case IBV_EVENT_PORT_ERR: 3907 case IBV_EVENT_LID_CHANGE: 3908 case IBV_EVENT_PKEY_CHANGE: 3909 case IBV_EVENT_SM_CHANGE: 3910 case IBV_EVENT_SRQ_ERR: 3911 case IBV_EVENT_SRQ_LIMIT_REACHED: 3912 case IBV_EVENT_CLIENT_REREGISTER: 3913 case IBV_EVENT_GID_CHANGE: 3914 case IBV_EVENT_SQ_DRAINED: 3915 default: 3916 SPDK_NOTICELOG("Async event: %s\n", 3917 ibv_event_type_str(event.event_type)); 3918 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3919 break; 3920 } 3921 ibv_ack_async_event(&event); 3922 3923 return 0; 3924 } 3925 3926 static void 3927 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3928 { 3929 int rc = 0; 3930 uint32_t i = 0; 3931 3932 for (i = 0; i < max_events; i++) { 3933 rc = nvmf_process_ib_event(device); 3934 if (rc) { 3935 break; 3936 } 3937 } 3938 3939 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3940 } 3941 3942 static int 3943 nvmf_rdma_accept(void *ctx) 3944 { 3945 int nfds, i = 0; 3946 struct spdk_nvmf_transport *transport = ctx; 3947 struct spdk_nvmf_rdma_transport *rtransport; 3948 struct spdk_nvmf_rdma_device *device, *tmp; 3949 uint32_t count; 3950 short revents; 3951 bool do_retry; 3952 3953 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3954 do_retry = nvmf_rdma_retry_listen_port(rtransport); 3955 3956 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3957 3958 if (nfds <= 0) { 3959 return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3960 } 3961 3962 /* The first poll descriptor is RDMA CM event */ 3963 if (rtransport->poll_fds[i++].revents & POLLIN) { 3964 nvmf_process_cm_events(transport, NVMF_RDMA_MAX_EVENTS_PER_POLL); 3965 nfds--; 3966 } 3967 3968 if (nfds == 0) { 3969 return SPDK_POLLER_BUSY; 3970 } 3971 3972 /* Second and subsequent poll descriptors are IB async events */ 3973 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3974 revents = rtransport->poll_fds[i++].revents; 3975 if (revents & POLLIN) { 3976 if (spdk_likely(!device->need_destroy)) { 3977 nvmf_process_ib_events(device, NVMF_RDMA_MAX_EVENTS_PER_POLL); 3978 if (spdk_unlikely(device->need_destroy)) { 3979 nvmf_rdma_handle_device_removal(rtransport, device); 3980 } 3981 } 3982 nfds--; 3983 } else if (revents & POLLNVAL || revents & POLLHUP) { 3984 SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device); 3985 nfds--; 3986 } 3987 } 3988 /* check all flagged fd's have been served */ 3989 assert(nfds == 0); 3990 3991 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3992 } 3993 3994 static void 3995 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3996 struct spdk_nvmf_ctrlr_data *cdata) 3997 { 3998 cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD; 3999 4000 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 4001 since in-capsule data only works with NVME drives that support SGL memory layout */ 4002 if (transport->opts.dif_insert_or_strip) { 4003 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 4004 } 4005 4006 if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) { 4007 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule" 4008 " data is greater than 4KiB.\n"); 4009 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux " 4010 "kernel between versions 5.4 and 5.12 data corruption may occur for " 4011 "writes that are not a multiple of 4KiB in size.\n"); 4012 } 4013 } 4014 4015 static void 4016 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 4017 struct spdk_nvme_transport_id *trid, 4018 struct spdk_nvmf_discovery_log_page_entry *entry) 4019 { 4020 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 4021 entry->adrfam = trid->adrfam; 4022 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 4023 4024 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 4025 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 4026 4027 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 4028 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 4029 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 4030 } 4031 4032 static int 4033 nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport, 4034 struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device, 4035 struct spdk_nvmf_rdma_poller **out_poller) 4036 { 4037 struct spdk_nvmf_rdma_poller *poller; 4038 struct spdk_rdma_provider_srq_init_attr srq_init_attr; 4039 struct spdk_nvmf_rdma_resource_opts opts; 4040 int num_cqe; 4041 4042 poller = calloc(1, sizeof(*poller)); 4043 if (!poller) { 4044 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 4045 return -1; 4046 } 4047 4048 poller->device = device; 4049 poller->group = rgroup; 4050 *out_poller = poller; 4051 4052 RB_INIT(&poller->qpairs); 4053 STAILQ_INIT(&poller->qpairs_pending_send); 4054 STAILQ_INIT(&poller->qpairs_pending_recv); 4055 4056 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 4057 SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup); 4058 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 4059 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) { 4060 SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n", 4061 rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr); 4062 } 4063 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr); 4064 4065 device->num_srq++; 4066 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 4067 srq_init_attr.pd = device->pd; 4068 srq_init_attr.stats = &poller->stat.qp_stats.recv; 4069 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 4070 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 4071 poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr); 4072 if (!poller->srq) { 4073 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 4074 return -1; 4075 } 4076 4077 opts.qp = poller->srq; 4078 opts.map = device->map; 4079 opts.qpair = NULL; 4080 opts.shared = true; 4081 opts.max_queue_depth = poller->max_srq_depth; 4082 opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size; 4083 4084 poller->resources = nvmf_rdma_resources_create(&opts); 4085 if (!poller->resources) { 4086 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 4087 return -1; 4088 } 4089 } 4090 4091 /* 4092 * When using an srq, we can limit the completion queue at startup. 4093 * The following formula represents the calculation: 4094 * num_cqe = num_recv + num_data_wr + num_send_wr. 4095 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 4096 */ 4097 if (poller->srq) { 4098 num_cqe = poller->max_srq_depth * 3; 4099 } else { 4100 num_cqe = rtransport->rdma_opts.num_cqe; 4101 } 4102 4103 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 4104 if (!poller->cq) { 4105 SPDK_ERRLOG("Unable to create completion queue\n"); 4106 return -1; 4107 } 4108 poller->num_cqe = num_cqe; 4109 return 0; 4110 } 4111 4112 static void 4113 _nvmf_rdma_register_poller_in_group(void *c) 4114 { 4115 struct spdk_nvmf_rdma_poller *poller; 4116 struct poller_manage_ctx *ctx = c; 4117 struct spdk_nvmf_rdma_device *device; 4118 int rc; 4119 4120 rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller); 4121 if (rc < 0 && poller) { 4122 nvmf_rdma_poller_destroy(poller); 4123 } 4124 4125 device = ctx->device; 4126 if (nvmf_rdma_all_pollers_management_done(ctx)) { 4127 device->is_ready = true; 4128 } 4129 } 4130 4131 static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 4132 4133 static struct spdk_nvmf_transport_poll_group * 4134 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport, 4135 struct spdk_nvmf_poll_group *group) 4136 { 4137 struct spdk_nvmf_rdma_transport *rtransport; 4138 struct spdk_nvmf_rdma_poll_group *rgroup; 4139 struct spdk_nvmf_rdma_poller *poller; 4140 struct spdk_nvmf_rdma_device *device; 4141 int rc; 4142 4143 if (spdk_interrupt_mode_is_enabled()) { 4144 SPDK_ERRLOG("RDMA transport does not support interrupt mode\n"); 4145 return NULL; 4146 } 4147 4148 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 4149 4150 rgroup = calloc(1, sizeof(*rgroup)); 4151 if (!rgroup) { 4152 return NULL; 4153 } 4154 4155 TAILQ_INIT(&rgroup->pollers); 4156 4157 TAILQ_FOREACH(device, &rtransport->devices, link) { 4158 rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller); 4159 if (rc < 0) { 4160 nvmf_rdma_poll_group_destroy(&rgroup->group); 4161 return NULL; 4162 } 4163 } 4164 4165 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 4166 if (rtransport->conn_sched.next_admin_pg == NULL) { 4167 rtransport->conn_sched.next_admin_pg = rgroup; 4168 rtransport->conn_sched.next_io_pg = rgroup; 4169 } 4170 4171 return &rgroup->group; 4172 } 4173 4174 static uint32_t 4175 nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg) 4176 { 4177 uint32_t count; 4178 4179 /* Just assume that unassociated qpairs will eventually be io 4180 * qpairs. This is close enough for the use cases for this 4181 * function. 4182 */ 4183 pthread_mutex_lock(&pg->mutex); 4184 count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs; 4185 pthread_mutex_unlock(&pg->mutex); 4186 4187 return count; 4188 } 4189 4190 static struct spdk_nvmf_transport_poll_group * 4191 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 4192 { 4193 struct spdk_nvmf_rdma_transport *rtransport; 4194 struct spdk_nvmf_rdma_poll_group **pg; 4195 struct spdk_nvmf_transport_poll_group *result; 4196 uint32_t count; 4197 4198 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4199 4200 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 4201 return NULL; 4202 } 4203 4204 if (qpair->qid == 0) { 4205 pg = &rtransport->conn_sched.next_admin_pg; 4206 } else { 4207 struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current; 4208 uint32_t min_value; 4209 4210 pg = &rtransport->conn_sched.next_io_pg; 4211 pg_min = *pg; 4212 pg_start = *pg; 4213 pg_current = *pg; 4214 min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 4215 4216 while (1) { 4217 count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 4218 4219 if (count < min_value) { 4220 min_value = count; 4221 pg_min = pg_current; 4222 } 4223 4224 pg_current = TAILQ_NEXT(pg_current, link); 4225 if (pg_current == NULL) { 4226 pg_current = TAILQ_FIRST(&rtransport->poll_groups); 4227 } 4228 4229 if (pg_current == pg_start || min_value == 0) { 4230 break; 4231 } 4232 } 4233 *pg = pg_min; 4234 } 4235 4236 assert(*pg != NULL); 4237 4238 result = &(*pg)->group; 4239 4240 *pg = TAILQ_NEXT(*pg, link); 4241 if (*pg == NULL) { 4242 *pg = TAILQ_FIRST(&rtransport->poll_groups); 4243 } 4244 4245 return result; 4246 } 4247 4248 static void 4249 nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller) 4250 { 4251 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 4252 int rc; 4253 4254 TAILQ_REMOVE(&poller->group->pollers, poller, link); 4255 RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) { 4256 nvmf_rdma_qpair_destroy(qpair); 4257 } 4258 4259 if (poller->srq) { 4260 if (poller->resources) { 4261 nvmf_rdma_resources_destroy(poller->resources); 4262 } 4263 spdk_rdma_provider_srq_destroy(poller->srq); 4264 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 4265 } 4266 4267 if (poller->cq) { 4268 rc = ibv_destroy_cq(poller->cq); 4269 if (rc != 0) { 4270 SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno)); 4271 } 4272 } 4273 4274 if (poller->destroy_cb) { 4275 poller->destroy_cb(poller->destroy_cb_ctx); 4276 poller->destroy_cb = NULL; 4277 } 4278 4279 free(poller); 4280 } 4281 4282 static void 4283 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 4284 { 4285 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 4286 struct spdk_nvmf_rdma_poller *poller, *tmp; 4287 struct spdk_nvmf_rdma_transport *rtransport; 4288 4289 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4290 if (!rgroup) { 4291 return; 4292 } 4293 4294 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 4295 nvmf_rdma_poller_destroy(poller); 4296 } 4297 4298 if (rgroup->group.transport == NULL) { 4299 /* Transport can be NULL when nvmf_rdma_poll_group_create() 4300 * calls this function directly in a failure path. */ 4301 free(rgroup); 4302 return; 4303 } 4304 4305 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 4306 4307 next_rgroup = TAILQ_NEXT(rgroup, link); 4308 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 4309 if (next_rgroup == NULL) { 4310 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 4311 } 4312 if (rtransport->conn_sched.next_admin_pg == rgroup) { 4313 rtransport->conn_sched.next_admin_pg = next_rgroup; 4314 } 4315 if (rtransport->conn_sched.next_io_pg == rgroup) { 4316 rtransport->conn_sched.next_io_pg = next_rgroup; 4317 } 4318 4319 free(rgroup); 4320 } 4321 4322 static void 4323 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 4324 { 4325 if (rqpair->cm_id != NULL) { 4326 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 4327 } 4328 } 4329 4330 static int 4331 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 4332 struct spdk_nvmf_qpair *qpair) 4333 { 4334 struct spdk_nvmf_rdma_poll_group *rgroup; 4335 struct spdk_nvmf_rdma_qpair *rqpair; 4336 struct spdk_nvmf_rdma_device *device; 4337 struct spdk_nvmf_rdma_poller *poller; 4338 int rc; 4339 4340 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4341 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4342 4343 device = rqpair->device; 4344 4345 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 4346 if (poller->device == device) { 4347 break; 4348 } 4349 } 4350 4351 if (!poller) { 4352 SPDK_ERRLOG("No poller found for device.\n"); 4353 return -1; 4354 } 4355 4356 if (poller->need_destroy) { 4357 SPDK_ERRLOG("Poller is destroying.\n"); 4358 return -1; 4359 } 4360 4361 rqpair->poller = poller; 4362 rqpair->srq = rqpair->poller->srq; 4363 4364 rc = nvmf_rdma_qpair_initialize(qpair); 4365 if (rc < 0) { 4366 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 4367 rqpair->poller = NULL; 4368 rqpair->srq = NULL; 4369 return -1; 4370 } 4371 4372 RB_INSERT(qpairs_tree, &poller->qpairs, rqpair); 4373 4374 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 4375 if (rc) { 4376 /* Try to reject, but we probably can't */ 4377 nvmf_rdma_qpair_reject_connection(rqpair); 4378 return -1; 4379 } 4380 4381 return 0; 4382 } 4383 4384 static int 4385 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 4386 struct spdk_nvmf_qpair *qpair) 4387 { 4388 struct spdk_nvmf_rdma_qpair *rqpair; 4389 4390 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4391 assert(group->transport->tgt != NULL); 4392 4393 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 4394 4395 if (!rqpair->destruct_channel) { 4396 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 4397 return 0; 4398 } 4399 4400 /* Sanity check that we get io_channel on the correct thread */ 4401 if (qpair->group) { 4402 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 4403 } 4404 4405 return 0; 4406 } 4407 4408 static int 4409 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 4410 { 4411 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 4412 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4413 struct spdk_nvmf_rdma_transport, transport); 4414 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4415 struct spdk_nvmf_rdma_qpair, qpair); 4416 4417 /* 4418 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 4419 * needs to be returned to the shared receive queue or the poll group will eventually be 4420 * starved of RECV structures. 4421 */ 4422 if (rqpair->srq && rdma_req->recv) { 4423 int rc; 4424 struct ibv_recv_wr *bad_recv_wr; 4425 4426 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 4427 rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 4428 if (rc) { 4429 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 4430 } 4431 } 4432 4433 _nvmf_rdma_request_free(rdma_req, rtransport); 4434 return 0; 4435 } 4436 4437 static int 4438 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 4439 { 4440 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4441 struct spdk_nvmf_rdma_transport, transport); 4442 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 4443 struct spdk_nvmf_rdma_request, req); 4444 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4445 struct spdk_nvmf_rdma_qpair, qpair); 4446 4447 if (spdk_unlikely(rqpair->ibv_in_error_state)) { 4448 /* The connection is dead. Move the request directly to the completed state. */ 4449 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4450 } else { 4451 /* The connection is alive, so process the request as normal */ 4452 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 4453 } 4454 4455 nvmf_rdma_request_process(rtransport, rdma_req); 4456 4457 return 0; 4458 } 4459 4460 static void 4461 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 4462 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 4463 { 4464 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4465 4466 rqpair->to_close = true; 4467 4468 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 4469 nvmf_rdma_qpair_reject_connection(rqpair); 4470 } 4471 if (rqpair->rdma_qp) { 4472 spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp); 4473 } 4474 4475 nvmf_rdma_destroy_drained_qpair(rqpair); 4476 4477 if (cb_fn) { 4478 cb_fn(cb_arg); 4479 } 4480 } 4481 4482 static struct spdk_nvmf_rdma_qpair * 4483 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 4484 { 4485 struct spdk_nvmf_rdma_qpair find; 4486 4487 find.qp_num = wc->qp_num; 4488 4489 return RB_FIND(qpairs_tree, &rpoller->qpairs, &find); 4490 } 4491 4492 #ifdef DEBUG 4493 static int 4494 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 4495 { 4496 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 4497 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 4498 } 4499 #endif 4500 4501 static void 4502 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 4503 int rc) 4504 { 4505 struct spdk_nvmf_rdma_recv *rdma_recv; 4506 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4507 4508 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 4509 while (bad_recv_wr != NULL) { 4510 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 4511 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4512 4513 rdma_recv->qpair->current_recv_depth++; 4514 bad_recv_wr = bad_recv_wr->next; 4515 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 4516 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair); 4517 } 4518 } 4519 4520 static void 4521 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 4522 { 4523 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 4524 while (bad_recv_wr != NULL) { 4525 bad_recv_wr = bad_recv_wr->next; 4526 rqpair->current_recv_depth++; 4527 } 4528 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4529 } 4530 4531 static void 4532 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 4533 struct spdk_nvmf_rdma_poller *rpoller) 4534 { 4535 struct spdk_nvmf_rdma_qpair *rqpair; 4536 struct ibv_recv_wr *bad_recv_wr; 4537 int rc; 4538 4539 if (rpoller->srq) { 4540 rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 4541 if (spdk_unlikely(rc)) { 4542 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 4543 } 4544 } else { 4545 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 4546 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 4547 rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 4548 if (spdk_unlikely(rc)) { 4549 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 4550 } 4551 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 4552 } 4553 } 4554 } 4555 4556 static void 4557 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 4558 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 4559 { 4560 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4561 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 4562 4563 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 4564 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 4565 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 4566 assert(rqpair->current_send_depth > 0); 4567 rqpair->current_send_depth--; 4568 switch (bad_rdma_wr->type) { 4569 case RDMA_WR_TYPE_DATA: 4570 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4571 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 4572 assert(rqpair->current_read_depth > 0); 4573 rqpair->current_read_depth--; 4574 } 4575 break; 4576 case RDMA_WR_TYPE_SEND: 4577 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4578 break; 4579 default: 4580 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 4581 prev_rdma_req = cur_rdma_req; 4582 continue; 4583 } 4584 4585 if (prev_rdma_req == cur_rdma_req) { 4586 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 4587 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 4588 continue; 4589 } 4590 4591 switch (cur_rdma_req->state) { 4592 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4593 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 4594 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, cur_rdma_req, state_link); 4595 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 4596 break; 4597 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 4598 case RDMA_REQUEST_STATE_COMPLETING: 4599 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4600 break; 4601 default: 4602 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 4603 cur_rdma_req->state, rqpair); 4604 continue; 4605 } 4606 4607 nvmf_rdma_request_process(rtransport, cur_rdma_req); 4608 prev_rdma_req = cur_rdma_req; 4609 } 4610 4611 if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) { 4612 /* Disconnect the connection. */ 4613 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4614 } 4615 4616 } 4617 4618 static void 4619 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 4620 struct spdk_nvmf_rdma_poller *rpoller) 4621 { 4622 struct spdk_nvmf_rdma_qpair *rqpair; 4623 struct ibv_send_wr *bad_wr = NULL; 4624 int rc; 4625 4626 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 4627 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 4628 rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 4629 4630 /* bad wr always points to the first wr that failed. */ 4631 if (spdk_unlikely(rc)) { 4632 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 4633 } 4634 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 4635 } 4636 } 4637 4638 static const char * 4639 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 4640 { 4641 switch (wr_type) { 4642 case RDMA_WR_TYPE_RECV: 4643 return "RECV"; 4644 case RDMA_WR_TYPE_SEND: 4645 return "SEND"; 4646 case RDMA_WR_TYPE_DATA: 4647 return "DATA"; 4648 default: 4649 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 4650 SPDK_UNREACHABLE(); 4651 } 4652 } 4653 4654 static inline void 4655 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 4656 { 4657 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 4658 4659 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 4660 /* If qpair is in ERR state, we will receive completions for all posted and not completed 4661 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 4662 SPDK_DEBUGLOG(rdma, 4663 "Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n", 4664 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id, 4665 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4666 } else { 4667 SPDK_ERRLOG("Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n", 4668 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id, 4669 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4670 } 4671 } 4672 4673 static int 4674 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 4675 struct spdk_nvmf_rdma_poller *rpoller) 4676 { 4677 struct ibv_wc wc[32]; 4678 struct spdk_nvmf_rdma_wr *rdma_wr; 4679 struct spdk_nvmf_rdma_request *rdma_req; 4680 struct spdk_nvmf_rdma_recv *rdma_recv; 4681 struct spdk_nvmf_rdma_qpair *rqpair, *tmp_rqpair; 4682 int reaped, i; 4683 int count = 0; 4684 int rc; 4685 bool error = false; 4686 uint64_t poll_tsc = spdk_get_ticks(); 4687 4688 if (spdk_unlikely(rpoller->need_destroy)) { 4689 /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not 4690 * be called because we cannot poll anything from cq. So we call that here to force 4691 * destroy the qpair after to_close turning true. 4692 */ 4693 RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) { 4694 nvmf_rdma_destroy_drained_qpair(rqpair); 4695 } 4696 return 0; 4697 } 4698 4699 /* Poll for completing operations. */ 4700 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 4701 if (spdk_unlikely(reaped < 0)) { 4702 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 4703 errno, spdk_strerror(errno)); 4704 return -1; 4705 } else if (reaped == 0) { 4706 rpoller->stat.idle_polls++; 4707 } 4708 4709 rpoller->stat.polls++; 4710 rpoller->stat.completions += reaped; 4711 4712 for (i = 0; i < reaped; i++) { 4713 4714 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 4715 4716 switch (rdma_wr->type) { 4717 case RDMA_WR_TYPE_SEND: 4718 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4719 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4720 4721 if (spdk_likely(!wc[i].status)) { 4722 count++; 4723 assert(wc[i].opcode == IBV_WC_SEND); 4724 assert(nvmf_rdma_req_is_completing(rdma_req)); 4725 } 4726 4727 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4728 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 4729 assert(rqpair->current_send_depth >= (uint32_t)rdma_req->num_outstanding_data_wr + 1); 4730 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 4731 rdma_req->num_outstanding_data_wr = 0; 4732 4733 nvmf_rdma_request_process(rtransport, rdma_req); 4734 break; 4735 case RDMA_WR_TYPE_RECV: 4736 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 4737 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4738 if (rpoller->srq != NULL) { 4739 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 4740 /* It is possible that there are still some completions for destroyed QP 4741 * associated with SRQ. We just ignore these late completions and re-post 4742 * receive WRs back to SRQ. 4743 */ 4744 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 4745 struct ibv_recv_wr *bad_wr; 4746 4747 rdma_recv->wr.next = NULL; 4748 spdk_rdma_provider_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 4749 rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 4750 if (rc) { 4751 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 4752 } 4753 continue; 4754 } 4755 } 4756 rqpair = rdma_recv->qpair; 4757 4758 assert(rqpair != NULL); 4759 if (spdk_likely(!wc[i].status)) { 4760 assert(wc[i].opcode == IBV_WC_RECV); 4761 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 4762 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4763 break; 4764 } 4765 } 4766 4767 rdma_recv->wr.next = NULL; 4768 rqpair->current_recv_depth++; 4769 rdma_recv->receive_tsc = poll_tsc; 4770 rpoller->stat.requests++; 4771 STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link); 4772 rqpair->qpair.queue_depth++; 4773 break; 4774 case RDMA_WR_TYPE_DATA: 4775 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4776 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4777 4778 assert(rdma_req->num_outstanding_data_wr > 0); 4779 4780 rqpair->current_send_depth--; 4781 rdma_req->num_outstanding_data_wr--; 4782 if (spdk_likely(!wc[i].status)) { 4783 assert(wc[i].opcode == IBV_WC_RDMA_READ); 4784 rqpair->current_read_depth--; 4785 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 4786 if (rdma_req->num_outstanding_data_wr == 0) { 4787 if (rdma_req->num_remaining_data_wr) { 4788 /* Only part of RDMA_READ operations was submitted, process the rest */ 4789 nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport); 4790 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 4791 nvmf_rdma_request_process(rtransport, rdma_req); 4792 break; 4793 } 4794 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 4795 nvmf_rdma_request_process(rtransport, rdma_req); 4796 } 4797 } else { 4798 /* If the data transfer fails still force the queue into the error state, 4799 * if we were performing an RDMA_READ, we need to force the request into a 4800 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 4801 * case, we should wait for the SEND to complete. */ 4802 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 4803 rqpair->current_read_depth--; 4804 if (rdma_req->num_outstanding_data_wr == 0) { 4805 if (rdma_req->num_remaining_data_wr) { 4806 /* Partially sent request is still in the pending_rdma_read_queue, 4807 * remove it now before completing */ 4808 rdma_req->num_remaining_data_wr = 0; 4809 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 4810 } 4811 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4812 nvmf_rdma_request_process(rtransport, rdma_req); 4813 } 4814 } 4815 } 4816 break; 4817 default: 4818 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 4819 continue; 4820 } 4821 4822 /* Handle error conditions */ 4823 if (spdk_unlikely(wc[i].status)) { 4824 rqpair->ibv_in_error_state = true; 4825 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 4826 4827 error = true; 4828 4829 if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) { 4830 /* Disconnect the connection. */ 4831 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4832 } else { 4833 nvmf_rdma_destroy_drained_qpair(rqpair); 4834 } 4835 continue; 4836 } 4837 4838 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 4839 4840 if (spdk_unlikely(!spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 4841 nvmf_rdma_destroy_drained_qpair(rqpair); 4842 } 4843 } 4844 4845 if (spdk_unlikely(error == true)) { 4846 return -1; 4847 } 4848 4849 if (reaped == 0) { 4850 /* In some cases we may not receive any CQE but we still may have pending IO requests waiting for 4851 * a resource (e.g. a WR from the data_wr_pool). 4852 * We need to start processing of such requests if no CQE reaped */ 4853 nvmf_rdma_poller_process_pending_buf_queue(rtransport, rpoller); 4854 } 4855 4856 /* submit outstanding work requests. */ 4857 _poller_submit_recvs(rtransport, rpoller); 4858 _poller_submit_sends(rtransport, rpoller); 4859 4860 return count; 4861 } 4862 4863 static void 4864 _nvmf_rdma_remove_destroyed_device(void *c) 4865 { 4866 struct spdk_nvmf_rdma_transport *rtransport = c; 4867 struct spdk_nvmf_rdma_device *device, *device_tmp; 4868 int rc; 4869 4870 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 4871 if (device->ready_to_destroy) { 4872 destroy_ib_device(rtransport, device); 4873 } 4874 } 4875 4876 free_poll_fds(rtransport); 4877 rc = generate_poll_fds(rtransport); 4878 /* cannot handle fd allocation error here */ 4879 if (rc != 0) { 4880 SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n"); 4881 } 4882 } 4883 4884 static void 4885 _nvmf_rdma_remove_poller_in_group_cb(void *c) 4886 { 4887 struct poller_manage_ctx *ctx = c; 4888 struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport; 4889 struct spdk_nvmf_rdma_device *device = ctx->device; 4890 struct spdk_thread *thread = ctx->thread; 4891 4892 if (nvmf_rdma_all_pollers_management_done(c)) { 4893 /* destroy device when last poller is destroyed */ 4894 device->ready_to_destroy = true; 4895 spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport); 4896 } 4897 } 4898 4899 static void 4900 _nvmf_rdma_remove_poller_in_group(void *c) 4901 { 4902 struct poller_manage_ctx *ctx = c; 4903 4904 ctx->rpoller->need_destroy = true; 4905 ctx->rpoller->destroy_cb_ctx = ctx; 4906 ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb; 4907 4908 /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */ 4909 if (RB_EMPTY(&ctx->rpoller->qpairs)) { 4910 nvmf_rdma_poller_destroy(ctx->rpoller); 4911 } 4912 } 4913 4914 static int 4915 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 4916 { 4917 struct spdk_nvmf_rdma_transport *rtransport; 4918 struct spdk_nvmf_rdma_poll_group *rgroup; 4919 struct spdk_nvmf_rdma_poller *rpoller, *tmp; 4920 int count = 0, rc, rc2 = 0; 4921 4922 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 4923 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4924 4925 TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) { 4926 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4927 if (spdk_unlikely(rc < 0)) { 4928 if (rc2 == 0) { 4929 rc2 = rc; 4930 } 4931 continue; 4932 } 4933 count += rc; 4934 } 4935 4936 return rc2 ? rc2 : count; 4937 } 4938 4939 static int 4940 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4941 struct spdk_nvme_transport_id *trid, 4942 bool peer) 4943 { 4944 struct sockaddr *saddr; 4945 uint16_t port; 4946 4947 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4948 4949 if (peer) { 4950 saddr = rdma_get_peer_addr(id); 4951 } else { 4952 saddr = rdma_get_local_addr(id); 4953 } 4954 switch (saddr->sa_family) { 4955 case AF_INET: { 4956 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4957 4958 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4959 inet_ntop(AF_INET, &saddr_in->sin_addr, 4960 trid->traddr, sizeof(trid->traddr)); 4961 if (peer) { 4962 port = ntohs(rdma_get_dst_port(id)); 4963 } else { 4964 port = ntohs(rdma_get_src_port(id)); 4965 } 4966 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4967 break; 4968 } 4969 case AF_INET6: { 4970 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4971 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4972 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4973 trid->traddr, sizeof(trid->traddr)); 4974 if (peer) { 4975 port = ntohs(rdma_get_dst_port(id)); 4976 } else { 4977 port = ntohs(rdma_get_src_port(id)); 4978 } 4979 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4980 break; 4981 } 4982 default: 4983 return -1; 4984 4985 } 4986 4987 return 0; 4988 } 4989 4990 static int 4991 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4992 struct spdk_nvme_transport_id *trid) 4993 { 4994 struct spdk_nvmf_rdma_qpair *rqpair; 4995 4996 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4997 4998 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4999 } 5000 5001 static int 5002 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 5003 struct spdk_nvme_transport_id *trid) 5004 { 5005 struct spdk_nvmf_rdma_qpair *rqpair; 5006 5007 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 5008 5009 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 5010 } 5011 5012 static int 5013 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 5014 struct spdk_nvme_transport_id *trid) 5015 { 5016 struct spdk_nvmf_rdma_qpair *rqpair; 5017 5018 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 5019 5020 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 5021 } 5022 5023 void 5024 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 5025 { 5026 g_nvmf_hooks = *hooks; 5027 } 5028 5029 static void 5030 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 5031 struct spdk_nvmf_rdma_request *rdma_req_to_abort, 5032 struct spdk_nvmf_rdma_qpair *rqpair) 5033 { 5034 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 5035 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 5036 5037 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, state_link); 5038 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 5039 5040 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 5041 } 5042 5043 static int 5044 _nvmf_rdma_qpair_abort_request(void *ctx) 5045 { 5046 struct spdk_nvmf_request *req = ctx; 5047 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 5048 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 5049 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 5050 struct spdk_nvmf_rdma_qpair, qpair); 5051 int rc; 5052 5053 spdk_poller_unregister(&req->poller); 5054 5055 switch (rdma_req_to_abort->state) { 5056 case RDMA_REQUEST_STATE_EXECUTING: 5057 rc = nvmf_ctrlr_abort_request(req); 5058 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 5059 return SPDK_POLLER_BUSY; 5060 } 5061 break; 5062 5063 case RDMA_REQUEST_STATE_NEED_BUFFER: 5064 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 5065 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 5066 5067 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5068 break; 5069 5070 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 5071 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 5072 spdk_nvmf_rdma_request, state_link); 5073 5074 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5075 break; 5076 5077 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 5078 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 5079 spdk_nvmf_rdma_request, state_link); 5080 5081 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5082 break; 5083 5084 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 5085 /* Remove req from the list here to re-use common function */ 5086 STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, 5087 spdk_nvmf_rdma_request, state_link); 5088 5089 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5090 break; 5091 5092 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 5093 if (spdk_get_ticks() < req->timeout_tsc) { 5094 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 5095 return SPDK_POLLER_BUSY; 5096 } 5097 break; 5098 5099 default: 5100 break; 5101 } 5102 5103 spdk_nvmf_request_complete(req); 5104 return SPDK_POLLER_BUSY; 5105 } 5106 5107 static void 5108 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 5109 struct spdk_nvmf_request *req) 5110 { 5111 struct spdk_nvmf_rdma_qpair *rqpair; 5112 struct spdk_nvmf_rdma_transport *rtransport; 5113 struct spdk_nvmf_transport *transport; 5114 uint16_t cid; 5115 uint32_t i, max_req_count; 5116 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 5117 5118 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 5119 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 5120 transport = &rtransport->transport; 5121 5122 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 5123 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 5124 5125 for (i = 0; i < max_req_count; i++) { 5126 rdma_req = &rqpair->resources->reqs[i]; 5127 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 5128 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 5129 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 5130 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 5131 rdma_req->req.qpair == qpair) { 5132 rdma_req_to_abort = rdma_req; 5133 break; 5134 } 5135 } 5136 5137 if (rdma_req_to_abort == NULL) { 5138 spdk_nvmf_request_complete(req); 5139 return; 5140 } 5141 5142 req->req_to_abort = &rdma_req_to_abort->req; 5143 req->timeout_tsc = spdk_get_ticks() + 5144 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 5145 req->poller = NULL; 5146 5147 _nvmf_rdma_qpair_abort_request(req); 5148 } 5149 5150 static void 5151 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 5152 struct spdk_json_write_ctx *w) 5153 { 5154 struct spdk_nvmf_rdma_poll_group *rgroup; 5155 struct spdk_nvmf_rdma_poller *rpoller; 5156 5157 assert(w != NULL); 5158 5159 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 5160 5161 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 5162 5163 spdk_json_write_named_array_begin(w, "devices"); 5164 5165 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 5166 spdk_json_write_object_begin(w); 5167 spdk_json_write_named_string(w, "name", 5168 ibv_get_device_name(rpoller->device->context->device)); 5169 spdk_json_write_named_uint64(w, "polls", 5170 rpoller->stat.polls); 5171 spdk_json_write_named_uint64(w, "idle_polls", 5172 rpoller->stat.idle_polls); 5173 spdk_json_write_named_uint64(w, "completions", 5174 rpoller->stat.completions); 5175 spdk_json_write_named_uint64(w, "requests", 5176 rpoller->stat.requests); 5177 spdk_json_write_named_uint64(w, "request_latency", 5178 rpoller->stat.request_latency); 5179 spdk_json_write_named_uint64(w, "pending_free_request", 5180 rpoller->stat.pending_free_request); 5181 spdk_json_write_named_uint64(w, "pending_rdma_read", 5182 rpoller->stat.pending_rdma_read); 5183 spdk_json_write_named_uint64(w, "pending_rdma_write", 5184 rpoller->stat.pending_rdma_write); 5185 spdk_json_write_named_uint64(w, "pending_rdma_send", 5186 rpoller->stat.pending_rdma_send); 5187 spdk_json_write_named_uint64(w, "total_send_wrs", 5188 rpoller->stat.qp_stats.send.num_submitted_wrs); 5189 spdk_json_write_named_uint64(w, "send_doorbell_updates", 5190 rpoller->stat.qp_stats.send.doorbell_updates); 5191 spdk_json_write_named_uint64(w, "total_recv_wrs", 5192 rpoller->stat.qp_stats.recv.num_submitted_wrs); 5193 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 5194 rpoller->stat.qp_stats.recv.doorbell_updates); 5195 spdk_json_write_object_end(w); 5196 } 5197 5198 spdk_json_write_array_end(w); 5199 } 5200 5201 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 5202 .name = "RDMA", 5203 .type = SPDK_NVME_TRANSPORT_RDMA, 5204 .opts_init = nvmf_rdma_opts_init, 5205 .create = nvmf_rdma_create, 5206 .dump_opts = nvmf_rdma_dump_opts, 5207 .destroy = nvmf_rdma_destroy, 5208 5209 .listen = nvmf_rdma_listen, 5210 .stop_listen = nvmf_rdma_stop_listen, 5211 .cdata_init = nvmf_rdma_cdata_init, 5212 5213 .listener_discover = nvmf_rdma_discover, 5214 5215 .poll_group_create = nvmf_rdma_poll_group_create, 5216 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 5217 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 5218 .poll_group_add = nvmf_rdma_poll_group_add, 5219 .poll_group_remove = nvmf_rdma_poll_group_remove, 5220 .poll_group_poll = nvmf_rdma_poll_group_poll, 5221 5222 .req_free = nvmf_rdma_request_free, 5223 .req_complete = nvmf_rdma_request_complete, 5224 5225 .qpair_fini = nvmf_rdma_close_qpair, 5226 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 5227 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 5228 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 5229 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 5230 5231 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 5232 }; 5233 5234 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 5235 SPDK_LOG_REGISTER_COMPONENT(rdma) 5236