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) 2021-2023 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.h" 21 22 #include "nvmf_internal.h" 23 #include "transport.h" 24 25 #include "spdk_internal/trace_defs.h" 26 27 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 28 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 29 30 /* 31 RDMA Connection Resource Defaults 32 */ 33 #define NVMF_DEFAULT_MSDBD 16 34 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 35 #define NVMF_DEFAULT_RSP_SGE 1 36 #define NVMF_DEFAULT_RX_SGE 2 37 38 SPDK_STATIC_ASSERT(NVMF_DEFAULT_MSDBD <= SPDK_NVMF_MAX_SGL_ENTRIES, 39 "MSDBD must not exceed SPDK_NVMF_MAX_SGL_ENTRIES"); 40 41 /* The RDMA completion queue size */ 42 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 43 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 44 45 static int g_spdk_nvmf_ibv_query_mask = 46 IBV_QP_STATE | 47 IBV_QP_PKEY_INDEX | 48 IBV_QP_PORT | 49 IBV_QP_ACCESS_FLAGS | 50 IBV_QP_AV | 51 IBV_QP_PATH_MTU | 52 IBV_QP_DEST_QPN | 53 IBV_QP_RQ_PSN | 54 IBV_QP_MAX_DEST_RD_ATOMIC | 55 IBV_QP_MIN_RNR_TIMER | 56 IBV_QP_SQ_PSN | 57 IBV_QP_TIMEOUT | 58 IBV_QP_RETRY_CNT | 59 IBV_QP_RNR_RETRY | 60 IBV_QP_MAX_QP_RD_ATOMIC; 61 62 enum spdk_nvmf_rdma_request_state { 63 /* The request is not currently in use */ 64 RDMA_REQUEST_STATE_FREE = 0, 65 66 /* Initial state when request first received */ 67 RDMA_REQUEST_STATE_NEW, 68 69 /* The request is queued until a data buffer is available. */ 70 RDMA_REQUEST_STATE_NEED_BUFFER, 71 72 /* The request is waiting on RDMA queue depth availability 73 * to transfer data from the host to the controller. 74 */ 75 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 76 77 /* The request is currently transferring data from the host to the controller. */ 78 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 79 80 /* The request is ready to execute at the block device */ 81 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 82 83 /* The request is currently executing at the block device */ 84 RDMA_REQUEST_STATE_EXECUTING, 85 86 /* The request finished executing at the block device */ 87 RDMA_REQUEST_STATE_EXECUTED, 88 89 /* The request is waiting on RDMA queue depth availability 90 * to transfer data from the controller to the host. 91 */ 92 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 93 94 /* The request is ready to send a completion */ 95 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 96 97 /* The request is currently transferring data from the controller to the host. */ 98 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 99 100 /* The request currently has an outstanding completion without an 101 * associated data transfer. 102 */ 103 RDMA_REQUEST_STATE_COMPLETING, 104 105 /* The request completed and can be marked free. */ 106 RDMA_REQUEST_STATE_COMPLETED, 107 108 /* Terminator */ 109 RDMA_REQUEST_NUM_STATES, 110 }; 111 112 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 113 { 114 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 115 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 116 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 117 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 118 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 119 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 120 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 121 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 122 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 123 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 124 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 125 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 126 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 127 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 128 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 129 spdk_trace_register_description("RDMA_REQ_TX_H2C", 130 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 131 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 132 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 133 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 134 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 135 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 136 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 137 spdk_trace_register_description("RDMA_REQ_EXECUTING", 138 TRACE_RDMA_REQUEST_STATE_EXECUTING, 139 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 140 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 141 spdk_trace_register_description("RDMA_REQ_EXECUTED", 142 TRACE_RDMA_REQUEST_STATE_EXECUTED, 143 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 144 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 145 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 146 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 147 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 148 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 149 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 150 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 151 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 152 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 153 spdk_trace_register_description("RDMA_REQ_COMPLETING", 154 TRACE_RDMA_REQUEST_STATE_COMPLETING, 155 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 156 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 157 spdk_trace_register_description("RDMA_REQ_COMPLETED", 158 TRACE_RDMA_REQUEST_STATE_COMPLETED, 159 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 160 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 161 162 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 163 OWNER_NONE, OBJECT_NONE, 0, 164 SPDK_TRACE_ARG_TYPE_INT, ""); 165 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 166 OWNER_NONE, OBJECT_NONE, 0, 167 SPDK_TRACE_ARG_TYPE_INT, "type"); 168 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 169 OWNER_NONE, OBJECT_NONE, 0, 170 SPDK_TRACE_ARG_TYPE_INT, "type"); 171 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 172 OWNER_NONE, OBJECT_NONE, 0, 173 SPDK_TRACE_ARG_TYPE_PTR, "state"); 174 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 175 OWNER_NONE, OBJECT_NONE, 0, 176 SPDK_TRACE_ARG_TYPE_INT, ""); 177 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 178 OWNER_NONE, OBJECT_NONE, 0, 179 SPDK_TRACE_ARG_TYPE_INT, ""); 180 } 181 182 enum spdk_nvmf_rdma_wr_type { 183 RDMA_WR_TYPE_RECV, 184 RDMA_WR_TYPE_SEND, 185 RDMA_WR_TYPE_DATA, 186 }; 187 188 struct spdk_nvmf_rdma_wr { 189 /* Uses enum spdk_nvmf_rdma_wr_type */ 190 uint8_t type; 191 }; 192 193 /* This structure holds commands as they are received off the wire. 194 * It must be dynamically paired with a full request object 195 * (spdk_nvmf_rdma_request) to service a request. It is separate 196 * from the request because RDMA does not appear to order 197 * completions, so occasionally we'll get a new incoming 198 * command when there aren't any free request objects. 199 */ 200 struct spdk_nvmf_rdma_recv { 201 struct ibv_recv_wr wr; 202 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 203 204 struct spdk_nvmf_rdma_qpair *qpair; 205 206 /* In-capsule data buffer */ 207 uint8_t *buf; 208 209 struct spdk_nvmf_rdma_wr rdma_wr; 210 uint64_t receive_tsc; 211 212 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 213 }; 214 215 struct spdk_nvmf_rdma_request_data { 216 struct ibv_send_wr wr; 217 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 218 }; 219 220 struct spdk_nvmf_rdma_request { 221 struct spdk_nvmf_request req; 222 223 bool fused_failed; 224 225 struct spdk_nvmf_rdma_wr data_wr; 226 struct spdk_nvmf_rdma_wr rsp_wr; 227 228 /* Uses enum spdk_nvmf_rdma_request_state */ 229 uint8_t state; 230 231 /* Data offset in req.iov */ 232 uint32_t offset; 233 234 struct spdk_nvmf_rdma_recv *recv; 235 236 struct { 237 struct ibv_send_wr wr; 238 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 239 } rsp; 240 241 uint32_t iovpos; 242 243 uint32_t num_outstanding_data_wr; 244 uint64_t receive_tsc; 245 struct spdk_nvmf_rdma_request *fused_pair; 246 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 247 struct ibv_send_wr *transfer_wr; 248 struct spdk_nvmf_rdma_request_data data; 249 }; 250 251 struct spdk_nvmf_rdma_resource_opts { 252 struct spdk_nvmf_rdma_qpair *qpair; 253 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 254 void *qp; 255 struct spdk_rdma_mem_map *map; 256 uint32_t max_queue_depth; 257 uint32_t in_capsule_data_size; 258 bool shared; 259 }; 260 261 struct spdk_nvmf_rdma_resources { 262 /* Array of size "max_queue_depth" containing RDMA requests. */ 263 struct spdk_nvmf_rdma_request *reqs; 264 265 /* Array of size "max_queue_depth" containing RDMA recvs. */ 266 struct spdk_nvmf_rdma_recv *recvs; 267 268 /* Array of size "max_queue_depth" containing 64 byte capsules 269 * used for receive. 270 */ 271 union nvmf_h2c_msg *cmds; 272 273 /* Array of size "max_queue_depth" containing 16 byte completions 274 * to be sent back to the user. 275 */ 276 union nvmf_c2h_msg *cpls; 277 278 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 279 * buffers to be used for in capsule data. 280 */ 281 void *bufs; 282 283 /* Receives that are waiting for a request object */ 284 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 285 286 /* Queue to track free requests */ 287 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 288 }; 289 290 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 291 292 typedef void (*spdk_poller_destroy_cb)(void *ctx); 293 294 struct spdk_nvmf_rdma_ibv_event_ctx { 295 struct spdk_nvmf_rdma_qpair *rqpair; 296 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 297 /* Link to other ibv events associated with this qpair */ 298 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 299 }; 300 301 struct spdk_nvmf_rdma_qpair { 302 struct spdk_nvmf_qpair qpair; 303 304 struct spdk_nvmf_rdma_device *device; 305 struct spdk_nvmf_rdma_poller *poller; 306 307 struct spdk_rdma_qp *rdma_qp; 308 struct rdma_cm_id *cm_id; 309 struct spdk_rdma_srq *srq; 310 struct rdma_cm_id *listen_id; 311 312 /* Cache the QP number to improve QP search by RB tree. */ 313 uint32_t qp_num; 314 315 /* The maximum number of I/O outstanding on this connection at one time */ 316 uint16_t max_queue_depth; 317 318 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 319 uint16_t max_read_depth; 320 321 /* The maximum number of RDMA SEND operations at one time */ 322 uint32_t max_send_depth; 323 324 /* The current number of outstanding WRs from this qpair's 325 * recv queue. Should not exceed device->attr.max_queue_depth. 326 */ 327 uint16_t current_recv_depth; 328 329 /* The current number of active RDMA READ operations */ 330 uint16_t current_read_depth; 331 332 /* The current number of posted WRs from this qpair's 333 * send queue. Should not exceed max_send_depth. 334 */ 335 uint32_t current_send_depth; 336 337 /* The maximum number of SGEs per WR on the send queue */ 338 uint32_t max_send_sge; 339 340 /* The maximum number of SGEs per WR on the recv queue */ 341 uint32_t max_recv_sge; 342 343 struct spdk_nvmf_rdma_resources *resources; 344 345 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 346 347 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 348 349 /* Number of requests not in the free state */ 350 uint32_t qd; 351 352 RB_ENTRY(spdk_nvmf_rdma_qpair) node; 353 354 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 355 356 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 357 358 /* IBV queue pair attributes: they are used to manage 359 * qp state and recover from errors. 360 */ 361 enum ibv_qp_state ibv_state; 362 363 /* Points to the a request that has fuse bits set to 364 * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting 365 * for the request that has SPDK_NVME_CMD_FUSE_SECOND. 366 */ 367 struct spdk_nvmf_rdma_request *fused_first; 368 369 /* 370 * io_channel which is used to destroy qpair when it is removed from poll group 371 */ 372 struct spdk_io_channel *destruct_channel; 373 374 /* List of ibv async events */ 375 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 376 377 /* Lets us know that we have received the last_wqe event. */ 378 bool last_wqe_reached; 379 380 /* Indicate that nvmf_rdma_close_qpair is called */ 381 bool to_close; 382 }; 383 384 struct spdk_nvmf_rdma_poller_stat { 385 uint64_t completions; 386 uint64_t polls; 387 uint64_t idle_polls; 388 uint64_t requests; 389 uint64_t request_latency; 390 uint64_t pending_free_request; 391 uint64_t pending_rdma_read; 392 uint64_t pending_rdma_write; 393 struct spdk_rdma_qp_stats qp_stats; 394 }; 395 396 struct spdk_nvmf_rdma_poller { 397 struct spdk_nvmf_rdma_device *device; 398 struct spdk_nvmf_rdma_poll_group *group; 399 400 int num_cqe; 401 int required_num_wr; 402 struct ibv_cq *cq; 403 404 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 405 uint16_t max_srq_depth; 406 bool need_destroy; 407 408 /* Shared receive queue */ 409 struct spdk_rdma_srq *srq; 410 411 struct spdk_nvmf_rdma_resources *resources; 412 struct spdk_nvmf_rdma_poller_stat stat; 413 414 spdk_poller_destroy_cb destroy_cb; 415 void *destroy_cb_ctx; 416 417 RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs; 418 419 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 420 421 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 422 423 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 424 }; 425 426 struct spdk_nvmf_rdma_poll_group_stat { 427 uint64_t pending_data_buffer; 428 }; 429 430 struct spdk_nvmf_rdma_poll_group { 431 struct spdk_nvmf_transport_poll_group group; 432 struct spdk_nvmf_rdma_poll_group_stat stat; 433 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 434 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 435 }; 436 437 struct spdk_nvmf_rdma_conn_sched { 438 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 439 struct spdk_nvmf_rdma_poll_group *next_io_pg; 440 }; 441 442 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 443 struct spdk_nvmf_rdma_device { 444 struct ibv_device_attr attr; 445 struct ibv_context *context; 446 447 struct spdk_rdma_mem_map *map; 448 struct ibv_pd *pd; 449 450 int num_srq; 451 bool need_destroy; 452 bool ready_to_destroy; 453 bool is_ready; 454 455 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 456 }; 457 458 struct spdk_nvmf_rdma_port { 459 const struct spdk_nvme_transport_id *trid; 460 struct rdma_cm_id *id; 461 struct spdk_nvmf_rdma_device *device; 462 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 463 }; 464 465 struct rdma_transport_opts { 466 int num_cqe; 467 uint32_t max_srq_depth; 468 bool no_srq; 469 bool no_wr_batching; 470 int acceptor_backlog; 471 }; 472 473 struct spdk_nvmf_rdma_transport { 474 struct spdk_nvmf_transport transport; 475 struct rdma_transport_opts rdma_opts; 476 477 struct spdk_nvmf_rdma_conn_sched conn_sched; 478 479 struct rdma_event_channel *event_channel; 480 481 struct spdk_mempool *data_wr_pool; 482 483 struct spdk_poller *accept_poller; 484 485 /* fields used to poll RDMA/IB events */ 486 nfds_t npoll_fds; 487 struct pollfd *poll_fds; 488 489 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 490 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 491 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 492 493 /* ports that are removed unexpectedly and need retry listen */ 494 TAILQ_HEAD(, spdk_nvmf_rdma_port) retry_ports; 495 }; 496 497 struct poller_manage_ctx { 498 struct spdk_nvmf_rdma_transport *rtransport; 499 struct spdk_nvmf_rdma_poll_group *rgroup; 500 struct spdk_nvmf_rdma_poller *rpoller; 501 struct spdk_nvmf_rdma_device *device; 502 503 struct spdk_thread *thread; 504 volatile int *inflight_op_counter; 505 }; 506 507 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 508 { 509 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe), 510 spdk_json_decode_int32, true 511 }, 512 { 513 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 514 spdk_json_decode_uint32, true 515 }, 516 { 517 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 518 spdk_json_decode_bool, true 519 }, 520 { 521 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 522 spdk_json_decode_bool, true 523 }, 524 { 525 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 526 spdk_json_decode_int32, true 527 }, 528 }; 529 530 static int 531 nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2) 532 { 533 return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num; 534 } 535 536 RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare); 537 538 static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 539 struct spdk_nvmf_rdma_request *rdma_req); 540 541 static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 542 struct spdk_nvmf_rdma_poller *rpoller); 543 544 static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 545 struct spdk_nvmf_rdma_poller *rpoller); 546 547 static void _nvmf_rdma_remove_destroyed_device(void *c); 548 549 static inline int 550 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 551 { 552 switch (state) { 553 case IBV_QPS_RESET: 554 case IBV_QPS_INIT: 555 case IBV_QPS_RTR: 556 case IBV_QPS_RTS: 557 case IBV_QPS_SQD: 558 case IBV_QPS_SQE: 559 case IBV_QPS_ERR: 560 return 0; 561 default: 562 return -1; 563 } 564 } 565 566 static inline enum spdk_nvme_media_error_status_code 567 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 568 enum spdk_nvme_media_error_status_code result; 569 switch (err_type) 570 { 571 case SPDK_DIF_REFTAG_ERROR: 572 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 573 break; 574 case SPDK_DIF_APPTAG_ERROR: 575 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 576 break; 577 case SPDK_DIF_GUARD_ERROR: 578 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 579 break; 580 default: 581 SPDK_UNREACHABLE(); 582 } 583 584 return result; 585 } 586 587 static enum ibv_qp_state 588 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 589 enum ibv_qp_state old_state, new_state; 590 struct ibv_qp_attr qp_attr; 591 struct ibv_qp_init_attr init_attr; 592 int rc; 593 594 old_state = rqpair->ibv_state; 595 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 596 g_spdk_nvmf_ibv_query_mask, &init_attr); 597 598 if (rc) 599 { 600 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 601 return IBV_QPS_ERR + 1; 602 } 603 604 new_state = qp_attr.qp_state; 605 rqpair->ibv_state = new_state; 606 qp_attr.ah_attr.port_num = qp_attr.port_num; 607 608 rc = nvmf_rdma_check_ibv_state(new_state); 609 if (rc) 610 { 611 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 612 /* 613 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 614 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 615 */ 616 return IBV_QPS_ERR + 1; 617 } 618 619 if (old_state != new_state) 620 { 621 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, (uintptr_t)rqpair, new_state); 622 } 623 return new_state; 624 } 625 626 /* 627 * Return data_wrs to pool starting from \b data_wr 628 * Request's own response and data WR are excluded 629 */ 630 static void 631 _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 632 struct ibv_send_wr *data_wr, 633 struct spdk_mempool *pool) 634 { 635 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 636 struct spdk_nvmf_rdma_request_data *nvmf_data; 637 struct ibv_send_wr *next_send_wr; 638 uint64_t req_wrid = data_wr->wr_id; 639 uint32_t num_wrs = 0; 640 641 while (data_wr && data_wr->wr_id == req_wrid) { 642 nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr); 643 memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge); 644 data_wr->num_sge = 0; 645 next_send_wr = data_wr->next; 646 if (data_wr != &rdma_req->data.wr) { 647 data_wr->next = NULL; 648 assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES); 649 work_requests[num_wrs] = nvmf_data; 650 num_wrs++; 651 } 652 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr; 653 } 654 655 if (num_wrs) { 656 spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs); 657 } 658 } 659 660 static void 661 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 662 struct spdk_nvmf_rdma_transport *rtransport) 663 { 664 rdma_req->num_outstanding_data_wr = 0; 665 666 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 667 668 rdma_req->data.wr.next = NULL; 669 rdma_req->rsp.wr.next = NULL; 670 } 671 672 static void 673 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 674 { 675 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 676 if (req->req.cmd) { 677 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 678 } 679 if (req->recv) { 680 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 681 } 682 } 683 684 static void 685 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 686 { 687 int i; 688 689 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 690 for (i = 0; i < rqpair->max_queue_depth; i++) { 691 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 692 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 693 } 694 } 695 } 696 697 static void 698 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 699 { 700 spdk_free(resources->cmds); 701 spdk_free(resources->cpls); 702 spdk_free(resources->bufs); 703 spdk_free(resources->reqs); 704 spdk_free(resources->recvs); 705 free(resources); 706 } 707 708 709 static struct spdk_nvmf_rdma_resources * 710 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 711 { 712 struct spdk_nvmf_rdma_resources *resources; 713 struct spdk_nvmf_rdma_request *rdma_req; 714 struct spdk_nvmf_rdma_recv *rdma_recv; 715 struct spdk_rdma_qp *qp = NULL; 716 struct spdk_rdma_srq *srq = NULL; 717 struct ibv_recv_wr *bad_wr = NULL; 718 struct spdk_rdma_memory_translation translation; 719 uint32_t i; 720 int rc = 0; 721 722 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 723 if (!resources) { 724 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 725 return NULL; 726 } 727 728 resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs), 729 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 730 resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs), 731 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 732 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 733 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 734 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 735 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 736 737 if (opts->in_capsule_data_size > 0) { 738 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 739 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 740 SPDK_MALLOC_DMA); 741 } 742 743 if (!resources->reqs || !resources->recvs || !resources->cmds || 744 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 745 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 746 goto cleanup; 747 } 748 749 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n", 750 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds)); 751 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n", 752 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls)); 753 if (resources->bufs) { 754 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n", 755 resources->bufs, opts->max_queue_depth * 756 opts->in_capsule_data_size); 757 } 758 759 /* Initialize queues */ 760 STAILQ_INIT(&resources->incoming_queue); 761 STAILQ_INIT(&resources->free_queue); 762 763 if (opts->shared) { 764 srq = (struct spdk_rdma_srq *)opts->qp; 765 } else { 766 qp = (struct spdk_rdma_qp *)opts->qp; 767 } 768 769 for (i = 0; i < opts->max_queue_depth; i++) { 770 rdma_recv = &resources->recvs[i]; 771 rdma_recv->qpair = opts->qpair; 772 773 /* Set up memory to receive commands */ 774 if (resources->bufs) { 775 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 776 opts->in_capsule_data_size)); 777 } 778 779 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 780 781 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 782 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 783 rc = spdk_rdma_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]), 784 &translation); 785 if (rc) { 786 goto cleanup; 787 } 788 rdma_recv->sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 789 rdma_recv->wr.num_sge = 1; 790 791 if (rdma_recv->buf) { 792 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 793 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 794 rc = spdk_rdma_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size, &translation); 795 if (rc) { 796 goto cleanup; 797 } 798 rdma_recv->sgl[1].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 799 rdma_recv->wr.num_sge++; 800 } 801 802 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 803 rdma_recv->wr.sg_list = rdma_recv->sgl; 804 if (srq) { 805 spdk_rdma_srq_queue_recv_wrs(srq, &rdma_recv->wr); 806 } else { 807 spdk_rdma_qp_queue_recv_wrs(qp, &rdma_recv->wr); 808 } 809 } 810 811 for (i = 0; i < opts->max_queue_depth; i++) { 812 rdma_req = &resources->reqs[i]; 813 814 if (opts->qpair != NULL) { 815 rdma_req->req.qpair = &opts->qpair->qpair; 816 } else { 817 rdma_req->req.qpair = NULL; 818 } 819 rdma_req->req.cmd = NULL; 820 rdma_req->req.iovcnt = 0; 821 rdma_req->req.stripped_data = NULL; 822 823 /* Set up memory to send responses */ 824 rdma_req->req.rsp = &resources->cpls[i]; 825 826 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 827 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 828 rc = spdk_rdma_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]), 829 &translation); 830 if (rc) { 831 goto cleanup; 832 } 833 rdma_req->rsp.sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 834 835 rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND; 836 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr; 837 rdma_req->rsp.wr.next = NULL; 838 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 839 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 840 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 841 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 842 843 /* Set up memory for data buffers */ 844 rdma_req->data_wr.type = RDMA_WR_TYPE_DATA; 845 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr; 846 rdma_req->data.wr.next = NULL; 847 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 848 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 849 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 850 851 /* Initialize request state to FREE */ 852 rdma_req->state = RDMA_REQUEST_STATE_FREE; 853 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 854 } 855 856 if (srq) { 857 rc = spdk_rdma_srq_flush_recv_wrs(srq, &bad_wr); 858 } else { 859 rc = spdk_rdma_qp_flush_recv_wrs(qp, &bad_wr); 860 } 861 862 if (rc) { 863 goto cleanup; 864 } 865 866 return resources; 867 868 cleanup: 869 nvmf_rdma_resources_destroy(resources); 870 return NULL; 871 } 872 873 static void 874 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 875 { 876 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 877 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 878 ctx->rqpair = NULL; 879 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 880 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 881 } 882 } 883 884 static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller); 885 886 static void 887 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 888 { 889 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 890 struct ibv_recv_wr *bad_recv_wr = NULL; 891 int rc; 892 893 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair); 894 895 if (rqpair->qd != 0) { 896 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 897 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 898 struct spdk_nvmf_rdma_transport, transport); 899 struct spdk_nvmf_rdma_request *req; 900 uint32_t i, max_req_count = 0; 901 902 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 903 904 if (rqpair->srq == NULL) { 905 nvmf_rdma_dump_qpair_contents(rqpair); 906 max_req_count = rqpair->max_queue_depth; 907 } else if (rqpair->poller && rqpair->resources) { 908 max_req_count = rqpair->poller->max_srq_depth; 909 } 910 911 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 912 for (i = 0; i < max_req_count; i++) { 913 req = &rqpair->resources->reqs[i]; 914 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 915 /* nvmf_rdma_request_process checks qpair ibv and internal state 916 * and completes a request */ 917 nvmf_rdma_request_process(rtransport, req); 918 } 919 } 920 assert(rqpair->qd == 0); 921 } 922 923 if (rqpair->poller) { 924 RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair); 925 926 if (rqpair->srq != NULL && rqpair->resources != NULL) { 927 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 928 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 929 if (rqpair == rdma_recv->qpair) { 930 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 931 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr); 932 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 933 if (rc) { 934 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 935 } 936 } 937 } 938 } 939 } 940 941 if (rqpair->cm_id) { 942 if (rqpair->rdma_qp != NULL) { 943 spdk_rdma_qp_destroy(rqpair->rdma_qp); 944 rqpair->rdma_qp = NULL; 945 } 946 947 if (rqpair->poller != NULL && rqpair->srq == NULL) { 948 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 949 } 950 } 951 952 if (rqpair->srq == NULL && rqpair->resources != NULL) { 953 nvmf_rdma_resources_destroy(rqpair->resources); 954 } 955 956 nvmf_rdma_qpair_clean_ibv_events(rqpair); 957 958 if (rqpair->destruct_channel) { 959 spdk_put_io_channel(rqpair->destruct_channel); 960 rqpair->destruct_channel = NULL; 961 } 962 963 if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) { 964 nvmf_rdma_poller_destroy(rqpair->poller); 965 } 966 967 /* destroy cm_id last so cma device will not be freed before we destroy the cq. */ 968 if (rqpair->cm_id) { 969 rdma_destroy_id(rqpair->cm_id); 970 } 971 972 free(rqpair); 973 } 974 975 static int 976 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 977 { 978 struct spdk_nvmf_rdma_poller *rpoller; 979 int rc, num_cqe, required_num_wr; 980 981 /* Enlarge CQ size dynamically */ 982 rpoller = rqpair->poller; 983 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 984 num_cqe = rpoller->num_cqe; 985 if (num_cqe < required_num_wr) { 986 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 987 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 988 } 989 990 if (rpoller->num_cqe != num_cqe) { 991 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) { 992 SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n" 993 "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe); 994 return -1; 995 } 996 if (required_num_wr > device->attr.max_cqe) { 997 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 998 required_num_wr, device->attr.max_cqe); 999 return -1; 1000 } 1001 1002 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 1003 rc = ibv_resize_cq(rpoller->cq, num_cqe); 1004 if (rc) { 1005 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 1006 return -1; 1007 } 1008 1009 rpoller->num_cqe = num_cqe; 1010 } 1011 1012 rpoller->required_num_wr = required_num_wr; 1013 return 0; 1014 } 1015 1016 static int 1017 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 1018 { 1019 struct spdk_nvmf_rdma_qpair *rqpair; 1020 struct spdk_nvmf_rdma_transport *rtransport; 1021 struct spdk_nvmf_transport *transport; 1022 struct spdk_nvmf_rdma_resource_opts opts; 1023 struct spdk_nvmf_rdma_device *device; 1024 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 1025 1026 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1027 device = rqpair->device; 1028 1029 qp_init_attr.qp_context = rqpair; 1030 qp_init_attr.pd = device->pd; 1031 qp_init_attr.send_cq = rqpair->poller->cq; 1032 qp_init_attr.recv_cq = rqpair->poller->cq; 1033 1034 if (rqpair->srq) { 1035 qp_init_attr.srq = rqpair->srq->srq; 1036 } else { 1037 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 1038 } 1039 1040 /* SEND, READ, and WRITE operations */ 1041 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 1042 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 1043 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1044 qp_init_attr.stats = &rqpair->poller->stat.qp_stats; 1045 1046 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1047 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1048 goto error; 1049 } 1050 1051 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 1052 if (!rqpair->rdma_qp) { 1053 goto error; 1054 } 1055 1056 rqpair->qp_num = rqpair->rdma_qp->qp->qp_num; 1057 1058 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1059 qp_init_attr.cap.max_send_wr); 1060 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1061 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1062 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair); 1063 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1064 1065 if (rqpair->poller->srq == NULL) { 1066 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1067 transport = &rtransport->transport; 1068 1069 opts.qp = rqpair->rdma_qp; 1070 opts.map = device->map; 1071 opts.qpair = rqpair; 1072 opts.shared = false; 1073 opts.max_queue_depth = rqpair->max_queue_depth; 1074 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1075 1076 rqpair->resources = nvmf_rdma_resources_create(&opts); 1077 1078 if (!rqpair->resources) { 1079 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1080 rdma_destroy_qp(rqpair->cm_id); 1081 goto error; 1082 } 1083 } else { 1084 rqpair->resources = rqpair->poller->resources; 1085 } 1086 1087 rqpair->current_recv_depth = 0; 1088 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1089 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1090 1091 return 0; 1092 1093 error: 1094 rdma_destroy_id(rqpair->cm_id); 1095 rqpair->cm_id = NULL; 1096 return -1; 1097 } 1098 1099 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1100 /* This function accepts either a single wr or the first wr in a linked list. */ 1101 static void 1102 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1103 { 1104 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1105 struct spdk_nvmf_rdma_transport, transport); 1106 1107 if (rqpair->srq != NULL) { 1108 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, first); 1109 } else { 1110 if (spdk_rdma_qp_queue_recv_wrs(rqpair->rdma_qp, first)) { 1111 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1112 } 1113 } 1114 1115 if (rtransport->rdma_opts.no_wr_batching) { 1116 _poller_submit_recvs(rtransport, rqpair->poller); 1117 } 1118 } 1119 1120 static int 1121 request_transfer_in(struct spdk_nvmf_request *req) 1122 { 1123 struct spdk_nvmf_rdma_request *rdma_req; 1124 struct spdk_nvmf_qpair *qpair; 1125 struct spdk_nvmf_rdma_qpair *rqpair; 1126 struct spdk_nvmf_rdma_transport *rtransport; 1127 1128 qpair = req->qpair; 1129 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1130 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1131 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1132 struct spdk_nvmf_rdma_transport, transport); 1133 1134 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1135 assert(rdma_req != NULL); 1136 1137 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) { 1138 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1139 } 1140 if (rtransport->rdma_opts.no_wr_batching) { 1141 _poller_submit_sends(rtransport, rqpair->poller); 1142 } 1143 1144 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1145 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1146 return 0; 1147 } 1148 1149 static int 1150 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1151 { 1152 int num_outstanding_data_wr = 0; 1153 struct spdk_nvmf_rdma_request *rdma_req; 1154 struct spdk_nvmf_qpair *qpair; 1155 struct spdk_nvmf_rdma_qpair *rqpair; 1156 struct spdk_nvme_cpl *rsp; 1157 struct ibv_send_wr *first = NULL; 1158 struct spdk_nvmf_rdma_transport *rtransport; 1159 1160 *data_posted = 0; 1161 qpair = req->qpair; 1162 rsp = &req->rsp->nvme_cpl; 1163 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1164 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1165 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1166 struct spdk_nvmf_rdma_transport, transport); 1167 1168 /* Advance our sq_head pointer */ 1169 if (qpair->sq_head == qpair->sq_head_max) { 1170 qpair->sq_head = 0; 1171 } else { 1172 qpair->sq_head++; 1173 } 1174 rsp->sqhd = qpair->sq_head; 1175 1176 /* queue the capsule for the recv buffer */ 1177 assert(rdma_req->recv != NULL); 1178 1179 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1180 1181 rdma_req->recv = NULL; 1182 assert(rqpair->current_recv_depth > 0); 1183 rqpair->current_recv_depth--; 1184 1185 /* Build the response which consists of optional 1186 * RDMA WRITEs to transfer data, plus an RDMA SEND 1187 * containing the response. 1188 */ 1189 first = &rdma_req->rsp.wr; 1190 1191 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1192 /* On failure, data was not read from the controller. So clear the 1193 * number of outstanding data WRs to zero. 1194 */ 1195 rdma_req->num_outstanding_data_wr = 0; 1196 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1197 first = rdma_req->transfer_wr; 1198 *data_posted = 1; 1199 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1200 } 1201 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1202 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1203 } 1204 if (rtransport->rdma_opts.no_wr_batching) { 1205 _poller_submit_sends(rtransport, rqpair->poller); 1206 } 1207 1208 /* +1 for the rsp wr */ 1209 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1210 1211 return 0; 1212 } 1213 1214 static int 1215 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1216 { 1217 struct spdk_nvmf_rdma_accept_private_data accept_data; 1218 struct rdma_conn_param ctrlr_event_data = {}; 1219 int rc; 1220 1221 accept_data.recfmt = 0; 1222 accept_data.crqsize = rqpair->max_queue_depth; 1223 1224 ctrlr_event_data.private_data = &accept_data; 1225 ctrlr_event_data.private_data_len = sizeof(accept_data); 1226 if (id->ps == RDMA_PS_TCP) { 1227 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1228 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1229 } 1230 1231 /* Configure infinite retries for the initiator side qpair. 1232 * We need to pass this value to the initiator to prevent the 1233 * initiator side NIC from completing SEND requests back to the 1234 * initiator with status rnr_retry_count_exceeded. */ 1235 ctrlr_event_data.rnr_retry_count = 0x7; 1236 1237 /* When qpair is created without use of rdma cm API, an additional 1238 * information must be provided to initiator in the connection response: 1239 * whether qpair is using SRQ and its qp_num 1240 * Fields below are ignored by rdma cm if qpair has been 1241 * created using rdma cm API. */ 1242 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1243 ctrlr_event_data.qp_num = rqpair->qp_num; 1244 1245 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1246 if (rc) { 1247 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1248 } else { 1249 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1250 } 1251 1252 return rc; 1253 } 1254 1255 static void 1256 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1257 { 1258 struct spdk_nvmf_rdma_reject_private_data rej_data; 1259 1260 rej_data.recfmt = 0; 1261 rej_data.sts = error; 1262 1263 rdma_reject(id, &rej_data, sizeof(rej_data)); 1264 } 1265 1266 static int 1267 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1268 { 1269 struct spdk_nvmf_rdma_transport *rtransport; 1270 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1271 struct spdk_nvmf_rdma_port *port; 1272 struct rdma_conn_param *rdma_param = NULL; 1273 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1274 uint16_t max_queue_depth; 1275 uint16_t max_read_depth; 1276 1277 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1278 1279 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1280 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1281 1282 rdma_param = &event->param.conn; 1283 if (rdma_param->private_data == NULL || 1284 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1285 SPDK_ERRLOG("connect request: no private data provided\n"); 1286 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1287 return -1; 1288 } 1289 1290 private_data = rdma_param->private_data; 1291 if (private_data->recfmt != 0) { 1292 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1293 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1294 return -1; 1295 } 1296 1297 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1298 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1299 1300 port = event->listen_id->context; 1301 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1302 event->listen_id, event->listen_id->verbs, port); 1303 1304 /* Figure out the supported queue depth. This is a multi-step process 1305 * that takes into account hardware maximums, host provided values, 1306 * and our target's internal memory limits */ 1307 1308 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1309 1310 /* Start with the maximum queue depth allowed by the target */ 1311 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1312 max_read_depth = rtransport->transport.opts.max_queue_depth; 1313 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1314 rtransport->transport.opts.max_queue_depth); 1315 1316 /* Next check the local NIC's hardware limitations */ 1317 SPDK_DEBUGLOG(rdma, 1318 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1319 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1320 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1321 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1322 1323 /* Next check the remote NIC's hardware limitations */ 1324 SPDK_DEBUGLOG(rdma, 1325 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1326 rdma_param->initiator_depth, rdma_param->responder_resources); 1327 /* from man3 rdma_get_cm_event 1328 * responder_resources - Specifies the number of responder resources that is requested by the recipient. 1329 * The responder_resources field must match the initiator depth specified by the remote node when running 1330 * the rdma_connect and rdma_accept functions. */ 1331 if (rdma_param->responder_resources != 0) { 1332 SPDK_ERRLOG("Host (Initiator) is not allowed to use RDMA operations (responder_resources %u)\n", 1333 rdma_param->responder_resources); 1334 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_ORD); 1335 return -1; 1336 } 1337 /* from man3 rdma_get_cm_event 1338 * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds. 1339 * The initiator_depth field must match the responder resources specified by the remote node when running 1340 * the rdma_connect and rdma_accept functions. */ 1341 if (rdma_param->initiator_depth == 0) { 1342 SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n"); 1343 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD); 1344 return -1; 1345 } 1346 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1347 1348 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1349 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1350 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1351 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1352 1353 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1354 max_queue_depth, max_read_depth); 1355 1356 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1357 if (rqpair == NULL) { 1358 SPDK_ERRLOG("Could not allocate new connection.\n"); 1359 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1360 return -1; 1361 } 1362 1363 rqpair->device = port->device; 1364 rqpair->max_queue_depth = max_queue_depth; 1365 rqpair->max_read_depth = max_read_depth; 1366 rqpair->cm_id = event->id; 1367 rqpair->listen_id = event->listen_id; 1368 rqpair->qpair.transport = transport; 1369 STAILQ_INIT(&rqpair->ibv_events); 1370 /* use qid from the private data to determine the qpair type 1371 qid will be set to the appropriate value when the controller is created */ 1372 rqpair->qpair.qid = private_data->qid; 1373 1374 event->id->context = &rqpair->qpair; 1375 1376 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1377 1378 return 0; 1379 } 1380 1381 static inline void 1382 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1383 enum spdk_nvme_data_transfer xfer) 1384 { 1385 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1386 wr->opcode = IBV_WR_RDMA_WRITE; 1387 wr->send_flags = 0; 1388 wr->next = next; 1389 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1390 wr->opcode = IBV_WR_RDMA_READ; 1391 wr->send_flags = IBV_SEND_SIGNALED; 1392 wr->next = NULL; 1393 } else { 1394 assert(0); 1395 } 1396 } 1397 1398 static int 1399 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1400 struct spdk_nvmf_rdma_request *rdma_req, 1401 uint32_t num_sgl_descriptors) 1402 { 1403 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1404 struct spdk_nvmf_rdma_request_data *current_data_wr; 1405 uint32_t i; 1406 1407 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1408 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1409 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1410 return -EINVAL; 1411 } 1412 1413 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1414 return -ENOMEM; 1415 } 1416 1417 current_data_wr = &rdma_req->data; 1418 1419 for (i = 0; i < num_sgl_descriptors; i++) { 1420 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1421 current_data_wr->wr.next = &work_requests[i]->wr; 1422 current_data_wr = work_requests[i]; 1423 current_data_wr->wr.sg_list = current_data_wr->sgl; 1424 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1425 } 1426 1427 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1428 1429 return 0; 1430 } 1431 1432 static inline void 1433 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1434 { 1435 struct ibv_send_wr *wr = &rdma_req->data.wr; 1436 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1437 1438 wr->wr.rdma.rkey = sgl->keyed.key; 1439 wr->wr.rdma.remote_addr = sgl->address; 1440 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1441 } 1442 1443 static inline void 1444 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1445 { 1446 struct ibv_send_wr *wr = &rdma_req->data.wr; 1447 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1448 uint32_t i; 1449 int j; 1450 uint64_t remote_addr_offset = 0; 1451 1452 for (i = 0; i < num_wrs; ++i) { 1453 wr->wr.rdma.rkey = sgl->keyed.key; 1454 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1455 for (j = 0; j < wr->num_sge; ++j) { 1456 remote_addr_offset += wr->sg_list[j].length; 1457 } 1458 wr = wr->next; 1459 } 1460 } 1461 1462 static int 1463 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1464 struct spdk_nvmf_rdma_device *device, 1465 struct spdk_nvmf_rdma_request *rdma_req, 1466 struct ibv_send_wr *wr, 1467 uint32_t total_length) 1468 { 1469 struct spdk_rdma_memory_translation mem_translation; 1470 struct ibv_sge *sg_ele; 1471 struct iovec *iov; 1472 uint32_t lkey, remaining; 1473 int rc; 1474 1475 wr->num_sge = 0; 1476 1477 while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) { 1478 iov = &rdma_req->req.iov[rdma_req->iovpos]; 1479 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1480 if (spdk_unlikely(rc)) { 1481 return rc; 1482 } 1483 1484 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1485 sg_ele = &wr->sg_list[wr->num_sge]; 1486 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1487 1488 sg_ele->lkey = lkey; 1489 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset; 1490 sg_ele->length = remaining; 1491 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1492 sg_ele->length); 1493 rdma_req->offset += sg_ele->length; 1494 total_length -= sg_ele->length; 1495 wr->num_sge++; 1496 1497 if (rdma_req->offset == iov->iov_len) { 1498 rdma_req->offset = 0; 1499 rdma_req->iovpos++; 1500 } 1501 } 1502 1503 if (total_length) { 1504 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1505 return -EINVAL; 1506 } 1507 1508 return 0; 1509 } 1510 1511 static int 1512 nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_poll_group *rgroup, 1513 struct spdk_nvmf_rdma_device *device, 1514 struct spdk_nvmf_rdma_request *rdma_req, 1515 struct ibv_send_wr *wr, 1516 uint32_t total_length, 1517 uint32_t num_extra_wrs) 1518 { 1519 struct spdk_rdma_memory_translation mem_translation; 1520 struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx; 1521 struct ibv_sge *sg_ele; 1522 struct iovec *iov; 1523 struct iovec *rdma_iov; 1524 uint32_t lkey, remaining; 1525 uint32_t remaining_data_block, data_block_size, md_size; 1526 uint32_t sge_len; 1527 int rc; 1528 1529 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1530 1531 if (spdk_likely(!rdma_req->req.stripped_data)) { 1532 rdma_iov = rdma_req->req.iov; 1533 remaining_data_block = data_block_size; 1534 md_size = dif_ctx->md_size; 1535 } else { 1536 rdma_iov = rdma_req->req.stripped_data->iov; 1537 total_length = total_length / dif_ctx->block_size * data_block_size; 1538 remaining_data_block = total_length; 1539 md_size = 0; 1540 } 1541 1542 wr->num_sge = 0; 1543 1544 while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1545 iov = rdma_iov + rdma_req->iovpos; 1546 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1547 if (spdk_unlikely(rc)) { 1548 return rc; 1549 } 1550 1551 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1552 sg_ele = &wr->sg_list[wr->num_sge]; 1553 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1554 1555 while (remaining) { 1556 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1557 if (num_extra_wrs > 0 && wr->next) { 1558 wr = wr->next; 1559 wr->num_sge = 0; 1560 sg_ele = &wr->sg_list[wr->num_sge]; 1561 num_extra_wrs--; 1562 } else { 1563 break; 1564 } 1565 } 1566 sg_ele->lkey = lkey; 1567 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset); 1568 sge_len = spdk_min(remaining, remaining_data_block); 1569 sg_ele->length = sge_len; 1570 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, 1571 sg_ele->addr, sg_ele->length); 1572 remaining -= sge_len; 1573 remaining_data_block -= sge_len; 1574 rdma_req->offset += sge_len; 1575 total_length -= sge_len; 1576 1577 sg_ele++; 1578 wr->num_sge++; 1579 1580 if (remaining_data_block == 0) { 1581 /* skip metadata */ 1582 rdma_req->offset += md_size; 1583 total_length -= md_size; 1584 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1585 remaining -= spdk_min(remaining, md_size); 1586 remaining_data_block = data_block_size; 1587 } 1588 1589 if (remaining == 0) { 1590 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1591 the remaining metadata bits at the beginning of the next buffer */ 1592 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset); 1593 rdma_req->iovpos++; 1594 } 1595 } 1596 } 1597 1598 if (total_length) { 1599 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1600 return -EINVAL; 1601 } 1602 1603 return 0; 1604 } 1605 1606 static inline uint32_t 1607 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1608 { 1609 /* estimate the number of SG entries and WRs needed to process the request */ 1610 uint32_t num_sge = 0; 1611 uint32_t i; 1612 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1613 1614 for (i = 0; i < num_buffers && length > 0; i++) { 1615 uint32_t buffer_len = spdk_min(length, io_unit_size); 1616 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1617 1618 if (num_sge_in_block * block_size > buffer_len) { 1619 ++num_sge_in_block; 1620 } 1621 num_sge += num_sge_in_block; 1622 length -= buffer_len; 1623 } 1624 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1625 } 1626 1627 static int 1628 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1629 struct spdk_nvmf_rdma_device *device, 1630 struct spdk_nvmf_rdma_request *rdma_req) 1631 { 1632 struct spdk_nvmf_rdma_qpair *rqpair; 1633 struct spdk_nvmf_rdma_poll_group *rgroup; 1634 struct spdk_nvmf_request *req = &rdma_req->req; 1635 struct ibv_send_wr *wr = &rdma_req->data.wr; 1636 int rc; 1637 uint32_t num_wrs = 1; 1638 uint32_t length; 1639 1640 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1641 rgroup = rqpair->poller->group; 1642 1643 /* rdma wr specifics */ 1644 nvmf_rdma_setup_request(rdma_req); 1645 1646 length = req->length; 1647 if (spdk_unlikely(req->dif_enabled)) { 1648 req->dif.orig_length = length; 1649 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1650 req->dif.elba_length = length; 1651 } 1652 1653 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1654 length); 1655 if (rc != 0) { 1656 return rc; 1657 } 1658 1659 assert(req->iovcnt <= rqpair->max_send_sge); 1660 1661 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1662 * the stripped_buffers are got for DIF stripping. */ 1663 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1664 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1665 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1666 &rtransport->transport, req->dif.orig_length); 1667 if (rc != 0) { 1668 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1669 } 1670 } 1671 1672 rdma_req->iovpos = 0; 1673 1674 if (spdk_unlikely(req->dif_enabled)) { 1675 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1676 req->dif.dif_ctx.block_size); 1677 if (num_wrs > 1) { 1678 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1679 if (rc != 0) { 1680 goto err_exit; 1681 } 1682 } 1683 1684 rc = nvmf_rdma_fill_wr_sgl_with_dif(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1685 if (spdk_unlikely(rc != 0)) { 1686 goto err_exit; 1687 } 1688 1689 if (num_wrs > 1) { 1690 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1691 } 1692 } else { 1693 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length); 1694 if (spdk_unlikely(rc != 0)) { 1695 goto err_exit; 1696 } 1697 } 1698 1699 /* set the number of outstanding data WRs for this request. */ 1700 rdma_req->num_outstanding_data_wr = num_wrs; 1701 1702 return rc; 1703 1704 err_exit: 1705 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1706 nvmf_rdma_request_free_data(rdma_req, rtransport); 1707 req->iovcnt = 0; 1708 return rc; 1709 } 1710 1711 static int 1712 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1713 struct spdk_nvmf_rdma_device *device, 1714 struct spdk_nvmf_rdma_request *rdma_req) 1715 { 1716 struct spdk_nvmf_rdma_qpair *rqpair; 1717 struct spdk_nvmf_rdma_poll_group *rgroup; 1718 struct ibv_send_wr *current_wr; 1719 struct spdk_nvmf_request *req = &rdma_req->req; 1720 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1721 uint32_t num_sgl_descriptors; 1722 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0; 1723 uint32_t i; 1724 int rc; 1725 1726 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1727 rgroup = rqpair->poller->group; 1728 1729 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1730 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1731 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1732 1733 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1734 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1735 1736 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1737 for (i = 0; i < num_sgl_descriptors; i++) { 1738 if (spdk_likely(!req->dif_enabled)) { 1739 lengths[i] = desc->keyed.length; 1740 } else { 1741 req->dif.orig_length += desc->keyed.length; 1742 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1743 req->dif.elba_length += lengths[i]; 1744 } 1745 total_length += lengths[i]; 1746 desc++; 1747 } 1748 1749 if (total_length > rtransport->transport.opts.max_io_size) { 1750 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n", 1751 total_length, rtransport->transport.opts.max_io_size); 1752 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1753 return -EINVAL; 1754 } 1755 1756 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1757 return -ENOMEM; 1758 } 1759 1760 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length); 1761 if (rc != 0) { 1762 nvmf_rdma_request_free_data(rdma_req, rtransport); 1763 return rc; 1764 } 1765 1766 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1767 * the stripped_buffers are got for DIF stripping. */ 1768 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1769 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1770 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1771 &rtransport->transport, req->dif.orig_length); 1772 if (rc != 0) { 1773 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1774 } 1775 } 1776 1777 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1778 current_wr = &rdma_req->data.wr; 1779 assert(current_wr != NULL); 1780 1781 req->length = 0; 1782 rdma_req->iovpos = 0; 1783 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1784 for (i = 0; i < num_sgl_descriptors; i++) { 1785 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1786 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1787 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1788 rc = -EINVAL; 1789 goto err_exit; 1790 } 1791 1792 if (spdk_likely(!req->dif_enabled)) { 1793 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i]); 1794 } else { 1795 rc = nvmf_rdma_fill_wr_sgl_with_dif(rgroup, device, rdma_req, current_wr, 1796 lengths[i], 0); 1797 } 1798 if (rc != 0) { 1799 rc = -ENOMEM; 1800 goto err_exit; 1801 } 1802 1803 req->length += desc->keyed.length; 1804 current_wr->wr.rdma.rkey = desc->keyed.key; 1805 current_wr->wr.rdma.remote_addr = desc->address; 1806 current_wr = current_wr->next; 1807 desc++; 1808 } 1809 1810 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1811 /* Go back to the last descriptor in the list. */ 1812 desc--; 1813 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1814 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1815 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1816 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1817 } 1818 } 1819 #endif 1820 1821 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1822 1823 return 0; 1824 1825 err_exit: 1826 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1827 nvmf_rdma_request_free_data(rdma_req, rtransport); 1828 return rc; 1829 } 1830 1831 static int 1832 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1833 struct spdk_nvmf_rdma_device *device, 1834 struct spdk_nvmf_rdma_request *rdma_req) 1835 { 1836 struct spdk_nvmf_request *req = &rdma_req->req; 1837 struct spdk_nvme_cpl *rsp; 1838 struct spdk_nvme_sgl_descriptor *sgl; 1839 int rc; 1840 uint32_t length; 1841 1842 rsp = &req->rsp->nvme_cpl; 1843 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1844 1845 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1846 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1847 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1848 1849 length = sgl->keyed.length; 1850 if (length > rtransport->transport.opts.max_io_size) { 1851 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1852 length, rtransport->transport.opts.max_io_size); 1853 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1854 return -1; 1855 } 1856 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1857 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1858 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1859 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1860 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1861 } 1862 } 1863 #endif 1864 1865 /* fill request length and populate iovs */ 1866 req->length = length; 1867 1868 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req); 1869 if (spdk_unlikely(rc < 0)) { 1870 if (rc == -EINVAL) { 1871 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1872 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1873 return -1; 1874 } 1875 /* No available buffers. Queue this request up. */ 1876 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1877 return 0; 1878 } 1879 1880 /* backward compatible */ 1881 req->data = req->iov[0].iov_base; 1882 1883 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1884 req->iovcnt); 1885 1886 return 0; 1887 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1888 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1889 uint64_t offset = sgl->address; 1890 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1891 1892 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1893 offset, sgl->unkeyed.length); 1894 1895 if (offset > max_len) { 1896 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1897 offset, max_len); 1898 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1899 return -1; 1900 } 1901 max_len -= (uint32_t)offset; 1902 1903 if (sgl->unkeyed.length > max_len) { 1904 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1905 sgl->unkeyed.length, max_len); 1906 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1907 return -1; 1908 } 1909 1910 rdma_req->num_outstanding_data_wr = 0; 1911 req->data_from_pool = false; 1912 req->length = sgl->unkeyed.length; 1913 1914 req->iov[0].iov_base = rdma_req->recv->buf + offset; 1915 req->iov[0].iov_len = req->length; 1916 req->iovcnt = 1; 1917 req->data = req->iov[0].iov_base; 1918 1919 return 0; 1920 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1921 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1922 1923 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1924 if (rc == -ENOMEM) { 1925 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1926 return 0; 1927 } else if (rc == -EINVAL) { 1928 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1929 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1930 return -1; 1931 } 1932 1933 /* backward compatible */ 1934 req->data = req->iov[0].iov_base; 1935 1936 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1937 req->iovcnt); 1938 1939 return 0; 1940 } 1941 1942 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1943 sgl->generic.type, sgl->generic.subtype); 1944 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1945 return -1; 1946 } 1947 1948 static void 1949 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1950 struct spdk_nvmf_rdma_transport *rtransport) 1951 { 1952 struct spdk_nvmf_rdma_qpair *rqpair; 1953 struct spdk_nvmf_rdma_poll_group *rgroup; 1954 1955 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1956 if (rdma_req->req.data_from_pool) { 1957 rgroup = rqpair->poller->group; 1958 1959 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1960 } 1961 if (rdma_req->req.stripped_data) { 1962 nvmf_request_free_stripped_buffers(&rdma_req->req, 1963 &rqpair->poller->group->group, 1964 &rtransport->transport); 1965 } 1966 nvmf_rdma_request_free_data(rdma_req, rtransport); 1967 rdma_req->req.length = 0; 1968 rdma_req->req.iovcnt = 0; 1969 rdma_req->req.data = NULL; 1970 rdma_req->offset = 0; 1971 rdma_req->req.dif_enabled = false; 1972 rdma_req->fused_failed = false; 1973 if (rdma_req->fused_pair) { 1974 /* This req was part of a valid fused pair, but failed before it got to 1975 * READ_TO_EXECUTE state. This means we need to fail the other request 1976 * in the pair, because it is no longer part of a valid pair. If the pair 1977 * already reached READY_TO_EXECUTE state, we need to kick it. 1978 */ 1979 rdma_req->fused_pair->fused_failed = true; 1980 if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 1981 nvmf_rdma_request_process(rtransport, rdma_req->fused_pair); 1982 } 1983 rdma_req->fused_pair = NULL; 1984 } 1985 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1986 rqpair->qd--; 1987 1988 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1989 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1990 } 1991 1992 static void 1993 nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport, 1994 struct spdk_nvmf_rdma_qpair *rqpair, 1995 struct spdk_nvmf_rdma_request *rdma_req) 1996 { 1997 enum spdk_nvme_cmd_fuse last, next; 1998 1999 last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2000 next = rdma_req->req.cmd->nvme_cmd.fuse; 2001 2002 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2003 2004 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2005 return; 2006 } 2007 2008 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2009 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2010 /* This is a valid pair of fused commands. Point them at each other 2011 * so they can be submitted consecutively once ready to be executed. 2012 */ 2013 rqpair->fused_first->fused_pair = rdma_req; 2014 rdma_req->fused_pair = rqpair->fused_first; 2015 rqpair->fused_first = NULL; 2016 return; 2017 } else { 2018 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2019 rqpair->fused_first->fused_failed = true; 2020 2021 /* If the last req is in READY_TO_EXECUTE state, then call 2022 * nvmf_rdma_request_process(), otherwise nothing else will kick it. 2023 */ 2024 if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2025 nvmf_rdma_request_process(rtransport, rqpair->fused_first); 2026 } 2027 2028 rqpair->fused_first = NULL; 2029 } 2030 } 2031 2032 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2033 /* Set rqpair->fused_first here so that we know to check that the next request 2034 * is a SECOND (and to fail this one if it isn't). 2035 */ 2036 rqpair->fused_first = rdma_req; 2037 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2038 /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */ 2039 rdma_req->fused_failed = true; 2040 } 2041 } 2042 2043 bool 2044 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 2045 struct spdk_nvmf_rdma_request *rdma_req) 2046 { 2047 struct spdk_nvmf_rdma_qpair *rqpair; 2048 struct spdk_nvmf_rdma_device *device; 2049 struct spdk_nvmf_rdma_poll_group *rgroup; 2050 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2051 int rc; 2052 struct spdk_nvmf_rdma_recv *rdma_recv; 2053 enum spdk_nvmf_rdma_request_state prev_state; 2054 bool progress = false; 2055 int data_posted; 2056 uint32_t num_blocks; 2057 2058 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2059 device = rqpair->device; 2060 rgroup = rqpair->poller->group; 2061 2062 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 2063 2064 /* If the queue pair is in an error state, force the request to the completed state 2065 * to release resources. */ 2066 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2067 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 2068 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 2069 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 2070 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2071 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 2072 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2073 } 2074 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2075 } 2076 2077 /* The loop here is to allow for several back-to-back state changes. */ 2078 do { 2079 prev_state = rdma_req->state; 2080 2081 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 2082 2083 switch (rdma_req->state) { 2084 case RDMA_REQUEST_STATE_FREE: 2085 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 2086 * to escape this state. */ 2087 break; 2088 case RDMA_REQUEST_STATE_NEW: 2089 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 2090 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2091 rdma_recv = rdma_req->recv; 2092 2093 /* The first element of the SGL is the NVMe command */ 2094 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 2095 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 2096 rdma_req->transfer_wr = &rdma_req->data.wr; 2097 2098 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2099 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2100 break; 2101 } 2102 2103 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 2104 rdma_req->req.dif_enabled = true; 2105 } 2106 2107 nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req); 2108 2109 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2110 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 2111 rdma_req->rsp.wr.imm_data = 0; 2112 #endif 2113 2114 /* The next state transition depends on the data transfer needs of this request. */ 2115 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2116 2117 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2118 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2119 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2120 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2121 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 2122 break; 2123 } 2124 2125 /* If no data to transfer, ready to execute. */ 2126 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2127 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2128 break; 2129 } 2130 2131 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2132 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2133 break; 2134 case RDMA_REQUEST_STATE_NEED_BUFFER: 2135 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2136 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2137 2138 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2139 2140 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2141 /* This request needs to wait in line to obtain a buffer */ 2142 break; 2143 } 2144 2145 /* Try to get a data buffer */ 2146 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2147 if (rc < 0) { 2148 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2149 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2150 break; 2151 } 2152 2153 if (rdma_req->req.iovcnt == 0) { 2154 /* No buffers available. */ 2155 rgroup->stat.pending_data_buffer++; 2156 break; 2157 } 2158 2159 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2160 2161 /* If data is transferring from host to controller and the data didn't 2162 * arrive using in capsule data, we need to do a transfer from the host. 2163 */ 2164 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2165 rdma_req->req.data_from_pool) { 2166 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2167 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2168 break; 2169 } 2170 2171 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2172 break; 2173 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2174 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2175 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2176 2177 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2178 /* This request needs to wait in line to perform RDMA */ 2179 break; 2180 } 2181 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2182 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2183 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2184 rqpair->poller->stat.pending_rdma_read++; 2185 break; 2186 } 2187 2188 /* We have already verified that this request is the head of the queue. */ 2189 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2190 2191 rc = request_transfer_in(&rdma_req->req); 2192 if (!rc) { 2193 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2194 } else { 2195 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2196 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2197 } 2198 break; 2199 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2200 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2201 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2202 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2203 * to escape this state. */ 2204 break; 2205 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2206 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2207 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2208 2209 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2210 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2211 /* generate DIF for write operation */ 2212 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2213 assert(num_blocks > 0); 2214 2215 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2216 num_blocks, &rdma_req->req.dif.dif_ctx); 2217 if (rc != 0) { 2218 SPDK_ERRLOG("DIF generation failed\n"); 2219 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2220 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2221 break; 2222 } 2223 } 2224 2225 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2226 /* set extended length before IO operation */ 2227 rdma_req->req.length = rdma_req->req.dif.elba_length; 2228 } 2229 2230 if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 2231 if (rdma_req->fused_failed) { 2232 /* This request failed FUSED semantics. Fail it immediately, without 2233 * even sending it to the target layer. 2234 */ 2235 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2236 rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED; 2237 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2238 break; 2239 } 2240 2241 if (rdma_req->fused_pair == NULL || 2242 rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2243 /* This request is ready to execute, but either we don't know yet if it's 2244 * valid - i.e. this is a FIRST but we haven't received the next 2245 * request yet or the other request of this fused pair isn't ready to 2246 * execute. So break here and this request will get processed later either 2247 * when the other request is ready or we find that this request isn't valid. 2248 */ 2249 break; 2250 } 2251 } 2252 2253 /* If we get to this point, and this request is a fused command, we know that 2254 * it is part of valid sequence (FIRST followed by a SECOND) and that both 2255 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 2256 * request, and the other request of the fused pair, in the correct order. 2257 * Also clear the ->fused_pair pointers on both requests, since after this point 2258 * we no longer need to maintain the relationship between these two requests. 2259 */ 2260 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 2261 assert(rdma_req->fused_pair != NULL); 2262 assert(rdma_req->fused_pair->fused_pair != NULL); 2263 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2264 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2265 rdma_req->fused_pair->fused_pair = NULL; 2266 rdma_req->fused_pair = NULL; 2267 } 2268 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2269 spdk_nvmf_request_exec(&rdma_req->req); 2270 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 2271 assert(rdma_req->fused_pair != NULL); 2272 assert(rdma_req->fused_pair->fused_pair != NULL); 2273 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2274 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2275 rdma_req->fused_pair->fused_pair = NULL; 2276 rdma_req->fused_pair = NULL; 2277 } 2278 break; 2279 case RDMA_REQUEST_STATE_EXECUTING: 2280 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2281 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2282 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2283 * to escape this state. */ 2284 break; 2285 case RDMA_REQUEST_STATE_EXECUTED: 2286 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2287 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2288 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2289 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2290 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2291 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2292 } else { 2293 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2294 } 2295 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2296 /* restore the original length */ 2297 rdma_req->req.length = rdma_req->req.dif.orig_length; 2298 2299 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2300 struct spdk_dif_error error_blk; 2301 2302 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2303 if (!rdma_req->req.stripped_data) { 2304 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2305 &rdma_req->req.dif.dif_ctx, &error_blk); 2306 } else { 2307 rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov, 2308 rdma_req->req.stripped_data->iovcnt, 2309 rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2310 &rdma_req->req.dif.dif_ctx, &error_blk); 2311 } 2312 if (rc) { 2313 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2314 2315 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2316 error_blk.err_offset); 2317 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2318 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2319 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2320 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2321 } 2322 } 2323 } 2324 break; 2325 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2326 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2327 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2328 2329 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2330 /* This request needs to wait in line to perform RDMA */ 2331 break; 2332 } 2333 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2334 rqpair->max_send_depth) { 2335 /* We can only have so many WRs outstanding. we have to wait until some finish. 2336 * +1 since each request has an additional wr in the resp. */ 2337 rqpair->poller->stat.pending_rdma_write++; 2338 break; 2339 } 2340 2341 /* We have already verified that this request is the head of the queue. */ 2342 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2343 2344 /* The data transfer will be kicked off from 2345 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2346 */ 2347 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2348 break; 2349 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2350 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2351 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2352 rc = request_transfer_out(&rdma_req->req, &data_posted); 2353 assert(rc == 0); /* No good way to handle this currently */ 2354 if (rc) { 2355 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2356 } else { 2357 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2358 RDMA_REQUEST_STATE_COMPLETING; 2359 } 2360 break; 2361 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2362 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2363 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2364 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2365 * to escape this state. */ 2366 break; 2367 case RDMA_REQUEST_STATE_COMPLETING: 2368 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2369 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2370 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2371 * to escape this state. */ 2372 break; 2373 case RDMA_REQUEST_STATE_COMPLETED: 2374 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2375 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2376 2377 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2378 _nvmf_rdma_request_free(rdma_req, rtransport); 2379 break; 2380 case RDMA_REQUEST_NUM_STATES: 2381 default: 2382 assert(0); 2383 break; 2384 } 2385 2386 if (rdma_req->state != prev_state) { 2387 progress = true; 2388 } 2389 } while (rdma_req->state != prev_state); 2390 2391 return progress; 2392 } 2393 2394 /* Public API callbacks begin here */ 2395 2396 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2397 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2398 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2399 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2400 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2401 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2402 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2403 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2404 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX 2405 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2406 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2407 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2408 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2409 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2410 2411 static void 2412 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2413 { 2414 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2415 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2416 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2417 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2418 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2419 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2420 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2421 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2422 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2423 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2424 opts->transport_specific = NULL; 2425 } 2426 2427 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2428 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2429 2430 static inline bool 2431 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2432 { 2433 return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD || 2434 device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW; 2435 } 2436 2437 static int nvmf_rdma_accept(void *ctx); 2438 static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport); 2439 static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2440 struct spdk_nvmf_rdma_device *device); 2441 2442 static int 2443 create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context, 2444 struct spdk_nvmf_rdma_device **new_device) 2445 { 2446 struct spdk_nvmf_rdma_device *device; 2447 int flag = 0; 2448 int rc = 0; 2449 2450 device = calloc(1, sizeof(*device)); 2451 if (!device) { 2452 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2453 return -ENOMEM; 2454 } 2455 device->context = context; 2456 rc = ibv_query_device(device->context, &device->attr); 2457 if (rc < 0) { 2458 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2459 free(device); 2460 return rc; 2461 } 2462 2463 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2464 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2465 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2466 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2467 } 2468 2469 /** 2470 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2471 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2472 * but incorrectly reports that it does. There are changes making their way 2473 * through the kernel now that will enable this feature. When they are merged, 2474 * we can conditionally enable this feature. 2475 * 2476 * TODO: enable this for versions of the kernel rxe driver that support it. 2477 */ 2478 if (nvmf_rdma_is_rxe_device(device)) { 2479 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2480 } 2481 #endif 2482 2483 /* set up device context async ev fd as NON_BLOCKING */ 2484 flag = fcntl(device->context->async_fd, F_GETFL); 2485 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2486 if (rc < 0) { 2487 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2488 free(device); 2489 return rc; 2490 } 2491 2492 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2493 SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA trasport\n", device); 2494 2495 if (g_nvmf_hooks.get_ibv_pd) { 2496 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2497 } else { 2498 device->pd = ibv_alloc_pd(device->context); 2499 } 2500 2501 if (!device->pd) { 2502 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2503 destroy_ib_device(rtransport, device); 2504 return -ENOMEM; 2505 } 2506 2507 assert(device->map == NULL); 2508 2509 device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks, SPDK_RDMA_MEMORY_MAP_ROLE_TARGET); 2510 if (!device->map) { 2511 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2512 destroy_ib_device(rtransport, device); 2513 return -ENOMEM; 2514 } 2515 2516 assert(device->map != NULL); 2517 assert(device->pd != NULL); 2518 2519 if (new_device) { 2520 *new_device = device; 2521 } 2522 SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device), 2523 device, context); 2524 2525 return 0; 2526 } 2527 2528 static void 2529 free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2530 { 2531 if (rtransport->poll_fds) { 2532 free(rtransport->poll_fds); 2533 rtransport->poll_fds = NULL; 2534 } 2535 rtransport->npoll_fds = 0; 2536 } 2537 2538 static int 2539 generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2540 { 2541 /* Set up poll descriptor array to monitor events from RDMA and IB 2542 * in a single poll syscall 2543 */ 2544 int device_count = 0; 2545 int i = 0; 2546 struct spdk_nvmf_rdma_device *device, *tmp; 2547 2548 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2549 device_count++; 2550 } 2551 2552 rtransport->npoll_fds = device_count + 1; 2553 2554 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2555 if (rtransport->poll_fds == NULL) { 2556 SPDK_ERRLOG("poll_fds allocation failed\n"); 2557 return -ENOMEM; 2558 } 2559 2560 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2561 rtransport->poll_fds[i++].events = POLLIN; 2562 2563 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2564 rtransport->poll_fds[i].fd = device->context->async_fd; 2565 rtransport->poll_fds[i++].events = POLLIN; 2566 } 2567 2568 return 0; 2569 } 2570 2571 static struct spdk_nvmf_transport * 2572 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2573 { 2574 int rc; 2575 struct spdk_nvmf_rdma_transport *rtransport; 2576 struct spdk_nvmf_rdma_device *device; 2577 struct ibv_context **contexts; 2578 uint32_t i; 2579 int flag; 2580 uint32_t sge_count; 2581 uint32_t min_shared_buffers; 2582 uint32_t min_in_capsule_data_size; 2583 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2584 2585 rtransport = calloc(1, sizeof(*rtransport)); 2586 if (!rtransport) { 2587 return NULL; 2588 } 2589 2590 TAILQ_INIT(&rtransport->devices); 2591 TAILQ_INIT(&rtransport->ports); 2592 TAILQ_INIT(&rtransport->poll_groups); 2593 TAILQ_INIT(&rtransport->retry_ports); 2594 2595 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2596 rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2597 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2598 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2599 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2600 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2601 if (opts->transport_specific != NULL && 2602 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2603 SPDK_COUNTOF(rdma_transport_opts_decoder), 2604 &rtransport->rdma_opts)) { 2605 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2606 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2607 return NULL; 2608 } 2609 2610 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2611 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2612 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2613 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2614 " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d," 2615 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2616 opts->max_queue_depth, 2617 opts->max_io_size, 2618 opts->max_qpairs_per_ctrlr - 1, 2619 opts->io_unit_size, 2620 opts->in_capsule_data_size, 2621 opts->max_aq_depth, 2622 opts->num_shared_buffers, 2623 rtransport->rdma_opts.num_cqe, 2624 rtransport->rdma_opts.max_srq_depth, 2625 rtransport->rdma_opts.no_srq, 2626 rtransport->rdma_opts.acceptor_backlog, 2627 rtransport->rdma_opts.no_wr_batching, 2628 opts->abort_timeout_sec); 2629 2630 /* I/O unit size cannot be larger than max I/O size */ 2631 if (opts->io_unit_size > opts->max_io_size) { 2632 opts->io_unit_size = opts->max_io_size; 2633 } 2634 2635 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2636 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2637 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2638 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2639 } 2640 2641 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2642 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2643 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2644 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2645 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2646 return NULL; 2647 } 2648 2649 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 2650 if (opts->buf_cache_size < UINT32_MAX) { 2651 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2652 if (min_shared_buffers > opts->num_shared_buffers) { 2653 SPDK_ERRLOG("There are not enough buffers to satisfy" 2654 "per-poll group caches for each thread. (%" PRIu32 ")" 2655 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2656 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2657 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2658 return NULL; 2659 } 2660 } 2661 2662 sge_count = opts->max_io_size / opts->io_unit_size; 2663 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2664 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2665 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2666 return NULL; 2667 } 2668 2669 min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES; 2670 if (opts->in_capsule_data_size < min_in_capsule_data_size) { 2671 SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n", 2672 min_in_capsule_data_size); 2673 opts->in_capsule_data_size = min_in_capsule_data_size; 2674 } 2675 2676 rtransport->event_channel = rdma_create_event_channel(); 2677 if (rtransport->event_channel == NULL) { 2678 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2679 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2680 return NULL; 2681 } 2682 2683 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2684 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2685 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2686 rtransport->event_channel->fd, spdk_strerror(errno)); 2687 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2688 return NULL; 2689 } 2690 2691 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2692 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2693 sizeof(struct spdk_nvmf_rdma_request_data), 2694 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2695 SPDK_ENV_SOCKET_ID_ANY); 2696 if (!rtransport->data_wr_pool) { 2697 if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) { 2698 SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n"); 2699 SPDK_ERRLOG("Probably running in multiprocess environment, which is " 2700 "unsupported by the nvmf library\n"); 2701 } else { 2702 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2703 } 2704 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2705 return NULL; 2706 } 2707 2708 contexts = rdma_get_devices(NULL); 2709 if (contexts == NULL) { 2710 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2711 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2712 return NULL; 2713 } 2714 2715 i = 0; 2716 rc = 0; 2717 while (contexts[i] != NULL) { 2718 rc = create_ib_device(rtransport, contexts[i], &device); 2719 if (rc < 0) { 2720 break; 2721 } 2722 i++; 2723 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2724 device->is_ready = true; 2725 } 2726 rdma_free_devices(contexts); 2727 2728 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2729 /* divide and round up. */ 2730 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2731 2732 /* round up to the nearest 4k. */ 2733 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2734 2735 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2736 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2737 opts->io_unit_size); 2738 } 2739 2740 if (rc < 0) { 2741 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2742 return NULL; 2743 } 2744 2745 rc = generate_poll_fds(rtransport); 2746 if (rc < 0) { 2747 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2748 return NULL; 2749 } 2750 2751 rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport, 2752 opts->acceptor_poll_rate); 2753 if (!rtransport->accept_poller) { 2754 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2755 return NULL; 2756 } 2757 2758 return &rtransport->transport; 2759 } 2760 2761 static void 2762 destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2763 struct spdk_nvmf_rdma_device *device) 2764 { 2765 TAILQ_REMOVE(&rtransport->devices, device, link); 2766 spdk_rdma_free_mem_map(&device->map); 2767 if (device->pd) { 2768 if (!g_nvmf_hooks.get_ibv_pd) { 2769 ibv_dealloc_pd(device->pd); 2770 } 2771 } 2772 SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device); 2773 free(device); 2774 } 2775 2776 static void 2777 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2778 { 2779 struct spdk_nvmf_rdma_transport *rtransport; 2780 assert(w != NULL); 2781 2782 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2783 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2784 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2785 if (rtransport->rdma_opts.no_srq == true) { 2786 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2787 } 2788 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2789 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2790 } 2791 2792 static int 2793 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2794 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2795 { 2796 struct spdk_nvmf_rdma_transport *rtransport; 2797 struct spdk_nvmf_rdma_port *port, *port_tmp; 2798 struct spdk_nvmf_rdma_device *device, *device_tmp; 2799 2800 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2801 2802 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) { 2803 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 2804 free(port); 2805 } 2806 2807 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2808 TAILQ_REMOVE(&rtransport->ports, port, link); 2809 rdma_destroy_id(port->id); 2810 free(port); 2811 } 2812 2813 free_poll_fds(rtransport); 2814 2815 if (rtransport->event_channel != NULL) { 2816 rdma_destroy_event_channel(rtransport->event_channel); 2817 } 2818 2819 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2820 destroy_ib_device(rtransport, device); 2821 } 2822 2823 if (rtransport->data_wr_pool != NULL) { 2824 if (spdk_mempool_count(rtransport->data_wr_pool) != 2825 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2826 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2827 spdk_mempool_count(rtransport->data_wr_pool), 2828 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2829 } 2830 } 2831 2832 spdk_mempool_free(rtransport->data_wr_pool); 2833 2834 spdk_poller_unregister(&rtransport->accept_poller); 2835 free(rtransport); 2836 2837 if (cb_fn) { 2838 cb_fn(cb_arg); 2839 } 2840 return 0; 2841 } 2842 2843 static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2844 struct spdk_nvme_transport_id *trid, 2845 bool peer); 2846 2847 static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport); 2848 2849 static int 2850 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2851 struct spdk_nvmf_listen_opts *listen_opts) 2852 { 2853 struct spdk_nvmf_rdma_transport *rtransport; 2854 struct spdk_nvmf_rdma_device *device; 2855 struct spdk_nvmf_rdma_port *port, *tmp_port; 2856 struct addrinfo *res; 2857 struct addrinfo hints; 2858 int family; 2859 int rc; 2860 bool is_retry = false; 2861 2862 if (!strlen(trid->trsvcid)) { 2863 SPDK_ERRLOG("Service id is required\n"); 2864 return -EINVAL; 2865 } 2866 2867 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2868 assert(rtransport->event_channel != NULL); 2869 2870 port = calloc(1, sizeof(*port)); 2871 if (!port) { 2872 SPDK_ERRLOG("Port allocation failed\n"); 2873 return -ENOMEM; 2874 } 2875 2876 port->trid = trid; 2877 2878 switch (trid->adrfam) { 2879 case SPDK_NVMF_ADRFAM_IPV4: 2880 family = AF_INET; 2881 break; 2882 case SPDK_NVMF_ADRFAM_IPV6: 2883 family = AF_INET6; 2884 break; 2885 default: 2886 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2887 free(port); 2888 return -EINVAL; 2889 } 2890 2891 memset(&hints, 0, sizeof(hints)); 2892 hints.ai_family = family; 2893 hints.ai_flags = AI_NUMERICSERV; 2894 hints.ai_socktype = SOCK_STREAM; 2895 hints.ai_protocol = 0; 2896 2897 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2898 if (rc) { 2899 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2900 free(port); 2901 return -(abs(rc)); 2902 } 2903 2904 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2905 if (rc < 0) { 2906 SPDK_ERRLOG("rdma_create_id() failed\n"); 2907 freeaddrinfo(res); 2908 free(port); 2909 return rc; 2910 } 2911 2912 rc = rdma_bind_addr(port->id, res->ai_addr); 2913 freeaddrinfo(res); 2914 2915 if (rc < 0) { 2916 TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) { 2917 if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) { 2918 is_retry = true; 2919 break; 2920 } 2921 } 2922 if (!is_retry) { 2923 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2924 } 2925 rdma_destroy_id(port->id); 2926 free(port); 2927 return rc; 2928 } 2929 2930 if (!port->id->verbs) { 2931 SPDK_ERRLOG("ibv_context is null\n"); 2932 rdma_destroy_id(port->id); 2933 free(port); 2934 return -1; 2935 } 2936 2937 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2938 if (rc < 0) { 2939 SPDK_ERRLOG("rdma_listen() failed\n"); 2940 rdma_destroy_id(port->id); 2941 free(port); 2942 return rc; 2943 } 2944 2945 TAILQ_FOREACH(device, &rtransport->devices, link) { 2946 if (device->context == port->id->verbs && device->is_ready) { 2947 port->device = device; 2948 break; 2949 } 2950 } 2951 if (!port->device) { 2952 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2953 port->id->verbs); 2954 rdma_destroy_id(port->id); 2955 free(port); 2956 return -EINVAL; 2957 } 2958 2959 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2960 trid->traddr, trid->trsvcid); 2961 2962 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2963 return 0; 2964 } 2965 2966 static void 2967 nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport, 2968 const struct spdk_nvme_transport_id *trid, bool need_retry) 2969 { 2970 struct spdk_nvmf_rdma_transport *rtransport; 2971 struct spdk_nvmf_rdma_port *port, *tmp; 2972 2973 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2974 2975 if (!need_retry) { 2976 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) { 2977 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2978 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 2979 free(port); 2980 } 2981 } 2982 } 2983 2984 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2985 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2986 SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n", 2987 port->trid->traddr, port->trid->trsvcid, need_retry); 2988 TAILQ_REMOVE(&rtransport->ports, port, link); 2989 rdma_destroy_id(port->id); 2990 port->id = NULL; 2991 port->device = NULL; 2992 if (need_retry) { 2993 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 2994 } else { 2995 free(port); 2996 } 2997 break; 2998 } 2999 } 3000 } 3001 3002 static void 3003 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 3004 const struct spdk_nvme_transport_id *trid) 3005 { 3006 nvmf_rdma_stop_listen_ex(transport, trid, false); 3007 } 3008 3009 static void _nvmf_rdma_register_poller_in_group(void *c); 3010 static void _nvmf_rdma_remove_poller_in_group(void *c); 3011 3012 static bool 3013 nvmf_rdma_all_pollers_management_done(void *c) 3014 { 3015 struct poller_manage_ctx *ctx = c; 3016 int counter; 3017 3018 counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST); 3019 SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n", 3020 counter, ctx->rpoller); 3021 3022 if (counter == 0) { 3023 free((void *)ctx->inflight_op_counter); 3024 } 3025 free(ctx); 3026 3027 return counter == 0; 3028 } 3029 3030 static int 3031 nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport, 3032 struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add) 3033 { 3034 struct spdk_nvmf_rdma_poll_group *rgroup; 3035 struct spdk_nvmf_rdma_poller *rpoller; 3036 struct spdk_nvmf_poll_group *poll_group; 3037 struct poller_manage_ctx *ctx; 3038 bool found; 3039 int *inflight_counter; 3040 spdk_msg_fn do_fn; 3041 3042 *has_inflight = false; 3043 do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group; 3044 inflight_counter = calloc(1, sizeof(int)); 3045 if (!inflight_counter) { 3046 SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n"); 3047 return -ENOMEM; 3048 } 3049 3050 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3051 (*inflight_counter)++; 3052 } 3053 3054 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3055 found = false; 3056 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3057 if (rpoller->device == device) { 3058 found = true; 3059 break; 3060 } 3061 } 3062 if (found == is_add) { 3063 __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST); 3064 continue; 3065 } 3066 3067 ctx = calloc(1, sizeof(struct poller_manage_ctx)); 3068 if (!ctx) { 3069 SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n"); 3070 if (!*has_inflight) { 3071 free(inflight_counter); 3072 } 3073 return -ENOMEM; 3074 } 3075 3076 ctx->rtransport = rtransport; 3077 ctx->rgroup = rgroup; 3078 ctx->rpoller = rpoller; 3079 ctx->device = device; 3080 ctx->thread = spdk_get_thread(); 3081 ctx->inflight_op_counter = inflight_counter; 3082 *has_inflight = true; 3083 3084 poll_group = rgroup->group.group; 3085 if (poll_group->thread != spdk_get_thread()) { 3086 spdk_thread_send_msg(poll_group->thread, do_fn, ctx); 3087 } else { 3088 do_fn(ctx); 3089 } 3090 } 3091 3092 if (!*has_inflight) { 3093 free(inflight_counter); 3094 } 3095 3096 return 0; 3097 } 3098 3099 static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3100 struct spdk_nvmf_rdma_device *device); 3101 3102 static struct spdk_nvmf_rdma_device * 3103 nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 3104 struct ibv_context *context) 3105 { 3106 struct spdk_nvmf_rdma_device *device, *tmp_device; 3107 3108 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) { 3109 if (device->need_destroy) { 3110 continue; 3111 } 3112 3113 if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) { 3114 return device; 3115 } 3116 } 3117 3118 return NULL; 3119 } 3120 3121 static bool 3122 nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport, 3123 struct ibv_context *context) 3124 { 3125 struct spdk_nvmf_rdma_device *old_device, *new_device; 3126 int rc = 0; 3127 bool has_inflight; 3128 3129 old_device = nvmf_rdma_find_ib_device(rtransport, context); 3130 3131 if (old_device) { 3132 if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) { 3133 /* context may not have time to be cleaned when rescan. exactly one context 3134 * is valid for a device so this context must be invalid and just remove it. */ 3135 SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context); 3136 old_device->need_destroy = true; 3137 nvmf_rdma_handle_device_removal(rtransport, old_device); 3138 } 3139 return false; 3140 } 3141 3142 rc = create_ib_device(rtransport, context, &new_device); 3143 /* TODO: update transport opts. */ 3144 if (rc < 0) { 3145 SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n", 3146 ibv_get_device_name(context->device), context); 3147 return false; 3148 } 3149 3150 rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true); 3151 if (rc < 0) { 3152 SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n", 3153 ibv_get_device_name(context->device), context); 3154 return false; 3155 } 3156 3157 if (has_inflight) { 3158 new_device->is_ready = true; 3159 } 3160 3161 return true; 3162 } 3163 3164 static bool 3165 nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport) 3166 { 3167 struct spdk_nvmf_rdma_device *device; 3168 struct ibv_device **ibv_device_list = NULL; 3169 struct ibv_context **contexts = NULL; 3170 int i = 0; 3171 int num_dev = 0; 3172 bool new_create = false, has_new_device = false; 3173 struct ibv_context *tmp_verbs = NULL; 3174 3175 /* do not rescan when any device is destroying, or context may be freed when 3176 * regenerating the poll fds. 3177 */ 3178 TAILQ_FOREACH(device, &rtransport->devices, link) { 3179 if (device->need_destroy) { 3180 return false; 3181 } 3182 } 3183 3184 ibv_device_list = ibv_get_device_list(&num_dev); 3185 3186 /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be 3187 * marked as dead verbs and never be init again. So we need to make sure the 3188 * verbs is available before we call rdma_get_devices. */ 3189 if (num_dev >= 0) { 3190 for (i = 0; i < num_dev; i++) { 3191 tmp_verbs = ibv_open_device(ibv_device_list[i]); 3192 if (!tmp_verbs) { 3193 SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno); 3194 break; 3195 } 3196 if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) { 3197 SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i], 3198 tmp_verbs->device->dev_name); 3199 has_new_device = true; 3200 } 3201 ibv_close_device(tmp_verbs); 3202 } 3203 ibv_free_device_list(ibv_device_list); 3204 if (!tmp_verbs || !has_new_device) { 3205 return false; 3206 } 3207 } 3208 3209 contexts = rdma_get_devices(NULL); 3210 3211 for (i = 0; contexts && contexts[i] != NULL; i++) { 3212 new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]); 3213 } 3214 3215 if (new_create) { 3216 free_poll_fds(rtransport); 3217 generate_poll_fds(rtransport); 3218 } 3219 3220 if (contexts) { 3221 rdma_free_devices(contexts); 3222 } 3223 3224 return new_create; 3225 } 3226 3227 static bool 3228 nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport) 3229 { 3230 struct spdk_nvmf_rdma_port *port, *tmp_port; 3231 int rc = 0; 3232 bool new_create = false; 3233 3234 if (TAILQ_EMPTY(&rtransport->retry_ports)) { 3235 return false; 3236 } 3237 3238 new_create = nvmf_rdma_rescan_devices(rtransport); 3239 3240 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) { 3241 rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL); 3242 3243 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3244 if (rc) { 3245 if (new_create) { 3246 SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n", 3247 port->trid->traddr, port->trid->trsvcid, rc); 3248 } 3249 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3250 break; 3251 } else { 3252 SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid); 3253 free(port); 3254 } 3255 } 3256 3257 return true; 3258 } 3259 3260 static void 3261 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 3262 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 3263 { 3264 struct spdk_nvmf_request *req, *tmp; 3265 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 3266 struct spdk_nvmf_rdma_resources *resources; 3267 3268 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 3269 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 3270 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3271 break; 3272 } 3273 } 3274 3275 /* Then RDMA writes since reads have stronger restrictions than writes */ 3276 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 3277 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3278 break; 3279 } 3280 } 3281 3282 /* Then we handle request waiting on memory buffers. */ 3283 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 3284 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3285 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3286 break; 3287 } 3288 } 3289 3290 resources = rqpair->resources; 3291 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 3292 rdma_req = STAILQ_FIRST(&resources->free_queue); 3293 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 3294 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 3295 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 3296 3297 if (rqpair->srq != NULL) { 3298 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 3299 rdma_req->recv->qpair->qd++; 3300 } else { 3301 rqpair->qd++; 3302 } 3303 3304 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 3305 rdma_req->state = RDMA_REQUEST_STATE_NEW; 3306 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 3307 break; 3308 } 3309 } 3310 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 3311 rqpair->poller->stat.pending_free_request++; 3312 } 3313 } 3314 3315 static inline bool 3316 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 3317 { 3318 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 3319 return nvmf_rdma_is_rxe_device(device) || 3320 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 3321 } 3322 3323 static void 3324 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 3325 { 3326 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3327 struct spdk_nvmf_rdma_transport, transport); 3328 3329 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3330 3331 /* nvmf_rdma_close_qpair is not called */ 3332 if (!rqpair->to_close) { 3333 return; 3334 } 3335 3336 /* device is already destroyed and we should force destroy this qpair. */ 3337 if (rqpair->poller && rqpair->poller->need_destroy) { 3338 nvmf_rdma_qpair_destroy(rqpair); 3339 return; 3340 } 3341 3342 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 3343 if (rqpair->current_send_depth != 0) { 3344 return; 3345 } 3346 3347 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 3348 return; 3349 } 3350 3351 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 3352 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 3353 return; 3354 } 3355 3356 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 3357 3358 nvmf_rdma_qpair_destroy(rqpair); 3359 } 3360 3361 static int 3362 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 3363 { 3364 struct spdk_nvmf_qpair *qpair; 3365 struct spdk_nvmf_rdma_qpair *rqpair; 3366 3367 if (evt->id == NULL) { 3368 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 3369 return -1; 3370 } 3371 3372 qpair = evt->id->context; 3373 if (qpair == NULL) { 3374 SPDK_ERRLOG("disconnect request: no active connection\n"); 3375 return -1; 3376 } 3377 3378 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3379 3380 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair); 3381 3382 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3383 3384 return 0; 3385 } 3386 3387 #ifdef DEBUG 3388 static const char *CM_EVENT_STR[] = { 3389 "RDMA_CM_EVENT_ADDR_RESOLVED", 3390 "RDMA_CM_EVENT_ADDR_ERROR", 3391 "RDMA_CM_EVENT_ROUTE_RESOLVED", 3392 "RDMA_CM_EVENT_ROUTE_ERROR", 3393 "RDMA_CM_EVENT_CONNECT_REQUEST", 3394 "RDMA_CM_EVENT_CONNECT_RESPONSE", 3395 "RDMA_CM_EVENT_CONNECT_ERROR", 3396 "RDMA_CM_EVENT_UNREACHABLE", 3397 "RDMA_CM_EVENT_REJECTED", 3398 "RDMA_CM_EVENT_ESTABLISHED", 3399 "RDMA_CM_EVENT_DISCONNECTED", 3400 "RDMA_CM_EVENT_DEVICE_REMOVAL", 3401 "RDMA_CM_EVENT_MULTICAST_JOIN", 3402 "RDMA_CM_EVENT_MULTICAST_ERROR", 3403 "RDMA_CM_EVENT_ADDR_CHANGE", 3404 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 3405 }; 3406 #endif /* DEBUG */ 3407 3408 static void 3409 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 3410 struct spdk_nvmf_rdma_port *port) 3411 { 3412 struct spdk_nvmf_rdma_poll_group *rgroup; 3413 struct spdk_nvmf_rdma_poller *rpoller; 3414 struct spdk_nvmf_rdma_qpair *rqpair; 3415 3416 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3417 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3418 RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) { 3419 if (rqpair->listen_id == port->id) { 3420 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3421 } 3422 } 3423 } 3424 } 3425 } 3426 3427 static bool 3428 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 3429 struct rdma_cm_event *event) 3430 { 3431 const struct spdk_nvme_transport_id *trid; 3432 struct spdk_nvmf_rdma_port *port; 3433 struct spdk_nvmf_rdma_transport *rtransport; 3434 bool event_acked = false; 3435 3436 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3437 TAILQ_FOREACH(port, &rtransport->ports, link) { 3438 if (port->id == event->id) { 3439 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 3440 rdma_ack_cm_event(event); 3441 event_acked = true; 3442 trid = port->trid; 3443 break; 3444 } 3445 } 3446 3447 if (event_acked) { 3448 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 3449 3450 nvmf_rdma_stop_listen(transport, trid); 3451 nvmf_rdma_listen(transport, trid, NULL); 3452 } 3453 3454 return event_acked; 3455 } 3456 3457 static void 3458 nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3459 struct spdk_nvmf_rdma_device *device) 3460 { 3461 struct spdk_nvmf_rdma_port *port, *port_tmp; 3462 int rc; 3463 bool has_inflight; 3464 3465 rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false); 3466 if (rc) { 3467 SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc); 3468 return; 3469 } 3470 3471 if (!has_inflight) { 3472 /* no pollers, destroy the device */ 3473 device->ready_to_destroy = true; 3474 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport); 3475 } 3476 3477 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 3478 if (port->device == device) { 3479 SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n", 3480 port->trid->traddr, 3481 port->trid->trsvcid, 3482 ibv_get_device_name(port->device->context->device)); 3483 3484 /* keep NVMF listener and only destroy structures of the 3485 * RDMA transport. when the device comes back we can retry listening 3486 * and the application's workflow will not be interrupted. 3487 */ 3488 nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true); 3489 } 3490 } 3491 } 3492 3493 static void 3494 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 3495 struct rdma_cm_event *event) 3496 { 3497 struct spdk_nvmf_rdma_port *port, *tmp_port; 3498 struct spdk_nvmf_rdma_transport *rtransport; 3499 3500 port = event->id->context; 3501 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3502 3503 rdma_ack_cm_event(event); 3504 3505 /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive 3506 * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that 3507 * we are handling a port event here. 3508 */ 3509 TAILQ_FOREACH(tmp_port, &rtransport->ports, link) { 3510 if (port == tmp_port && port->device && !port->device->need_destroy) { 3511 port->device->need_destroy = true; 3512 nvmf_rdma_handle_device_removal(rtransport, port->device); 3513 } 3514 } 3515 } 3516 3517 static void 3518 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 3519 { 3520 struct spdk_nvmf_rdma_transport *rtransport; 3521 struct rdma_cm_event *event; 3522 int rc; 3523 bool event_acked; 3524 3525 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3526 3527 if (rtransport->event_channel == NULL) { 3528 return; 3529 } 3530 3531 while (1) { 3532 event_acked = false; 3533 rc = rdma_get_cm_event(rtransport->event_channel, &event); 3534 if (rc) { 3535 if (errno != EAGAIN && errno != EWOULDBLOCK) { 3536 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 3537 } 3538 break; 3539 } 3540 3541 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 3542 3543 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 3544 3545 switch (event->event) { 3546 case RDMA_CM_EVENT_ADDR_RESOLVED: 3547 case RDMA_CM_EVENT_ADDR_ERROR: 3548 case RDMA_CM_EVENT_ROUTE_RESOLVED: 3549 case RDMA_CM_EVENT_ROUTE_ERROR: 3550 /* No action required. The target never attempts to resolve routes. */ 3551 break; 3552 case RDMA_CM_EVENT_CONNECT_REQUEST: 3553 rc = nvmf_rdma_connect(transport, event); 3554 if (rc < 0) { 3555 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 3556 break; 3557 } 3558 break; 3559 case RDMA_CM_EVENT_CONNECT_RESPONSE: 3560 /* The target never initiates a new connection. So this will not occur. */ 3561 break; 3562 case RDMA_CM_EVENT_CONNECT_ERROR: 3563 /* Can this happen? The docs say it can, but not sure what causes it. */ 3564 break; 3565 case RDMA_CM_EVENT_UNREACHABLE: 3566 case RDMA_CM_EVENT_REJECTED: 3567 /* These only occur on the client side. */ 3568 break; 3569 case RDMA_CM_EVENT_ESTABLISHED: 3570 /* TODO: Should we be waiting for this event anywhere? */ 3571 break; 3572 case RDMA_CM_EVENT_DISCONNECTED: 3573 rc = nvmf_rdma_disconnect(event); 3574 if (rc < 0) { 3575 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3576 break; 3577 } 3578 break; 3579 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3580 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3581 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3582 * Once these events are sent to SPDK, we should release all IB resources and 3583 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3584 * ibv_destroy* functions to release user space memory allocated by IB. All kernel 3585 * resources are already cleaned. */ 3586 if (event->id->qp) { 3587 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3588 * corresponding qpair. Otherwise the event refers to a listening device. */ 3589 rc = nvmf_rdma_disconnect(event); 3590 if (rc < 0) { 3591 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3592 break; 3593 } 3594 } else { 3595 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3596 event_acked = true; 3597 } 3598 break; 3599 case RDMA_CM_EVENT_MULTICAST_JOIN: 3600 case RDMA_CM_EVENT_MULTICAST_ERROR: 3601 /* Multicast is not used */ 3602 break; 3603 case RDMA_CM_EVENT_ADDR_CHANGE: 3604 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3605 break; 3606 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3607 /* For now, do nothing. The target never re-uses queue pairs. */ 3608 break; 3609 default: 3610 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3611 break; 3612 } 3613 if (!event_acked) { 3614 rdma_ack_cm_event(event); 3615 } 3616 } 3617 } 3618 3619 static void 3620 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3621 { 3622 rqpair->last_wqe_reached = true; 3623 nvmf_rdma_destroy_drained_qpair(rqpair); 3624 } 3625 3626 static void 3627 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3628 { 3629 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3630 3631 if (event_ctx->rqpair) { 3632 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3633 if (event_ctx->cb_fn) { 3634 event_ctx->cb_fn(event_ctx->rqpair); 3635 } 3636 } 3637 free(event_ctx); 3638 } 3639 3640 static int 3641 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3642 spdk_nvmf_rdma_qpair_ibv_event fn) 3643 { 3644 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3645 struct spdk_thread *thr = NULL; 3646 int rc; 3647 3648 if (rqpair->qpair.group) { 3649 thr = rqpair->qpair.group->thread; 3650 } else if (rqpair->destruct_channel) { 3651 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3652 } 3653 3654 if (!thr) { 3655 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3656 return -EINVAL; 3657 } 3658 3659 ctx = calloc(1, sizeof(*ctx)); 3660 if (!ctx) { 3661 return -ENOMEM; 3662 } 3663 3664 ctx->rqpair = rqpair; 3665 ctx->cb_fn = fn; 3666 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3667 3668 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3669 if (rc) { 3670 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3671 free(ctx); 3672 } 3673 3674 return rc; 3675 } 3676 3677 static int 3678 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3679 { 3680 int rc; 3681 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3682 struct ibv_async_event event; 3683 3684 rc = ibv_get_async_event(device->context, &event); 3685 3686 if (rc) { 3687 /* In non-blocking mode -1 means there are no events available */ 3688 return rc; 3689 } 3690 3691 switch (event.event_type) { 3692 case IBV_EVENT_QP_FATAL: 3693 case IBV_EVENT_QP_LAST_WQE_REACHED: 3694 case IBV_EVENT_SQ_DRAINED: 3695 case IBV_EVENT_QP_REQ_ERR: 3696 case IBV_EVENT_QP_ACCESS_ERR: 3697 case IBV_EVENT_COMM_EST: 3698 case IBV_EVENT_PATH_MIG: 3699 case IBV_EVENT_PATH_MIG_ERR: 3700 rqpair = event.element.qp->qp_context; 3701 if (!rqpair) { 3702 /* Any QP event for NVMe-RDMA initiator may be returned. */ 3703 SPDK_NOTICELOG("Async QP event for unknown QP: %s\n", 3704 ibv_event_type_str(event.event_type)); 3705 break; 3706 } 3707 3708 switch (event.event_type) { 3709 case IBV_EVENT_QP_FATAL: 3710 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3711 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3712 (uintptr_t)rqpair, event.event_type); 3713 nvmf_rdma_update_ibv_state(rqpair); 3714 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3715 break; 3716 case IBV_EVENT_QP_LAST_WQE_REACHED: 3717 /* This event only occurs for shared receive queues. */ 3718 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3719 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3720 if (rc) { 3721 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3722 rqpair->last_wqe_reached = true; 3723 } 3724 break; 3725 case IBV_EVENT_SQ_DRAINED: 3726 /* This event occurs frequently in both error and non-error states. 3727 * Check if the qpair is in an error state before sending a message. */ 3728 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3729 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3730 (uintptr_t)rqpair, event.event_type); 3731 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3732 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3733 } 3734 break; 3735 case IBV_EVENT_QP_REQ_ERR: 3736 case IBV_EVENT_QP_ACCESS_ERR: 3737 case IBV_EVENT_COMM_EST: 3738 case IBV_EVENT_PATH_MIG: 3739 case IBV_EVENT_PATH_MIG_ERR: 3740 SPDK_NOTICELOG("Async QP event: %s\n", 3741 ibv_event_type_str(event.event_type)); 3742 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3743 (uintptr_t)rqpair, event.event_type); 3744 nvmf_rdma_update_ibv_state(rqpair); 3745 break; 3746 default: 3747 break; 3748 } 3749 break; 3750 case IBV_EVENT_DEVICE_FATAL: 3751 SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n", 3752 ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device); 3753 device->need_destroy = true; 3754 break; 3755 case IBV_EVENT_CQ_ERR: 3756 case IBV_EVENT_PORT_ACTIVE: 3757 case IBV_EVENT_PORT_ERR: 3758 case IBV_EVENT_LID_CHANGE: 3759 case IBV_EVENT_PKEY_CHANGE: 3760 case IBV_EVENT_SM_CHANGE: 3761 case IBV_EVENT_SRQ_ERR: 3762 case IBV_EVENT_SRQ_LIMIT_REACHED: 3763 case IBV_EVENT_CLIENT_REREGISTER: 3764 case IBV_EVENT_GID_CHANGE: 3765 default: 3766 SPDK_NOTICELOG("Async event: %s\n", 3767 ibv_event_type_str(event.event_type)); 3768 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3769 break; 3770 } 3771 ibv_ack_async_event(&event); 3772 3773 return 0; 3774 } 3775 3776 static void 3777 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3778 { 3779 int rc = 0; 3780 uint32_t i = 0; 3781 3782 for (i = 0; i < max_events; i++) { 3783 rc = nvmf_process_ib_event(device); 3784 if (rc) { 3785 break; 3786 } 3787 } 3788 3789 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3790 } 3791 3792 static int 3793 nvmf_rdma_accept(void *ctx) 3794 { 3795 int nfds, i = 0; 3796 struct spdk_nvmf_transport *transport = ctx; 3797 struct spdk_nvmf_rdma_transport *rtransport; 3798 struct spdk_nvmf_rdma_device *device, *tmp; 3799 uint32_t count; 3800 short revents; 3801 bool do_retry; 3802 3803 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3804 do_retry = nvmf_rdma_retry_listen_port(rtransport); 3805 3806 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3807 3808 if (nfds <= 0) { 3809 return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3810 } 3811 3812 /* The first poll descriptor is RDMA CM event */ 3813 if (rtransport->poll_fds[i++].revents & POLLIN) { 3814 nvmf_process_cm_event(transport); 3815 nfds--; 3816 } 3817 3818 if (nfds == 0) { 3819 return SPDK_POLLER_BUSY; 3820 } 3821 3822 /* Second and subsequent poll descriptors are IB async events */ 3823 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3824 revents = rtransport->poll_fds[i++].revents; 3825 if (revents & POLLIN) { 3826 if (spdk_likely(!device->need_destroy)) { 3827 nvmf_process_ib_events(device, 32); 3828 if (spdk_unlikely(device->need_destroy)) { 3829 nvmf_rdma_handle_device_removal(rtransport, device); 3830 } 3831 } 3832 nfds--; 3833 } else if (revents & POLLNVAL || revents & POLLHUP) { 3834 SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device); 3835 nfds--; 3836 } 3837 } 3838 /* check all flagged fd's have been served */ 3839 assert(nfds == 0); 3840 3841 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3842 } 3843 3844 static void 3845 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3846 struct spdk_nvmf_ctrlr_data *cdata) 3847 { 3848 cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD; 3849 3850 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3851 since in-capsule data only works with NVME drives that support SGL memory layout */ 3852 if (transport->opts.dif_insert_or_strip) { 3853 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3854 } 3855 3856 if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) { 3857 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule" 3858 " data is greater than 4KiB.\n"); 3859 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux " 3860 "kernel between versions 5.4 and 5.12 data corruption may occur for " 3861 "writes that are not a multiple of 4KiB in size.\n"); 3862 } 3863 } 3864 3865 static void 3866 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3867 struct spdk_nvme_transport_id *trid, 3868 struct spdk_nvmf_discovery_log_page_entry *entry) 3869 { 3870 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3871 entry->adrfam = trid->adrfam; 3872 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3873 3874 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3875 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3876 3877 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3878 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3879 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3880 } 3881 3882 static int 3883 nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport, 3884 struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device, 3885 struct spdk_nvmf_rdma_poller **out_poller) 3886 { 3887 struct spdk_nvmf_rdma_poller *poller; 3888 struct spdk_rdma_srq_init_attr srq_init_attr; 3889 struct spdk_nvmf_rdma_resource_opts opts; 3890 int num_cqe; 3891 3892 poller = calloc(1, sizeof(*poller)); 3893 if (!poller) { 3894 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3895 return -1; 3896 } 3897 3898 poller->device = device; 3899 poller->group = rgroup; 3900 *out_poller = poller; 3901 3902 RB_INIT(&poller->qpairs); 3903 STAILQ_INIT(&poller->qpairs_pending_send); 3904 STAILQ_INIT(&poller->qpairs_pending_recv); 3905 3906 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3907 SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup); 3908 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3909 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) { 3910 SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n", 3911 rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr); 3912 } 3913 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr); 3914 3915 device->num_srq++; 3916 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 3917 srq_init_attr.pd = device->pd; 3918 srq_init_attr.stats = &poller->stat.qp_stats.recv; 3919 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 3920 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3921 poller->srq = spdk_rdma_srq_create(&srq_init_attr); 3922 if (!poller->srq) { 3923 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3924 return -1; 3925 } 3926 3927 opts.qp = poller->srq; 3928 opts.map = device->map; 3929 opts.qpair = NULL; 3930 opts.shared = true; 3931 opts.max_queue_depth = poller->max_srq_depth; 3932 opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size; 3933 3934 poller->resources = nvmf_rdma_resources_create(&opts); 3935 if (!poller->resources) { 3936 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3937 return -1; 3938 } 3939 } 3940 3941 /* 3942 * When using an srq, we can limit the completion queue at startup. 3943 * The following formula represents the calculation: 3944 * num_cqe = num_recv + num_data_wr + num_send_wr. 3945 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3946 */ 3947 if (poller->srq) { 3948 num_cqe = poller->max_srq_depth * 3; 3949 } else { 3950 num_cqe = rtransport->rdma_opts.num_cqe; 3951 } 3952 3953 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3954 if (!poller->cq) { 3955 SPDK_ERRLOG("Unable to create completion queue\n"); 3956 return -1; 3957 } 3958 poller->num_cqe = num_cqe; 3959 return 0; 3960 } 3961 3962 static void 3963 _nvmf_rdma_register_poller_in_group(void *c) 3964 { 3965 struct spdk_nvmf_rdma_poller *poller; 3966 struct poller_manage_ctx *ctx = c; 3967 struct spdk_nvmf_rdma_device *device; 3968 int rc; 3969 3970 rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller); 3971 if (rc < 0 && poller) { 3972 nvmf_rdma_poller_destroy(poller); 3973 } 3974 3975 device = ctx->device; 3976 if (nvmf_rdma_all_pollers_management_done(ctx)) { 3977 device->is_ready = true; 3978 } 3979 } 3980 3981 static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3982 3983 static struct spdk_nvmf_transport_poll_group * 3984 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport, 3985 struct spdk_nvmf_poll_group *group) 3986 { 3987 struct spdk_nvmf_rdma_transport *rtransport; 3988 struct spdk_nvmf_rdma_poll_group *rgroup; 3989 struct spdk_nvmf_rdma_poller *poller; 3990 struct spdk_nvmf_rdma_device *device; 3991 int rc; 3992 3993 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3994 3995 rgroup = calloc(1, sizeof(*rgroup)); 3996 if (!rgroup) { 3997 return NULL; 3998 } 3999 4000 TAILQ_INIT(&rgroup->pollers); 4001 4002 TAILQ_FOREACH(device, &rtransport->devices, link) { 4003 rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller); 4004 if (rc < 0) { 4005 nvmf_rdma_poll_group_destroy(&rgroup->group); 4006 return NULL; 4007 } 4008 } 4009 4010 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 4011 if (rtransport->conn_sched.next_admin_pg == NULL) { 4012 rtransport->conn_sched.next_admin_pg = rgroup; 4013 rtransport->conn_sched.next_io_pg = rgroup; 4014 } 4015 4016 return &rgroup->group; 4017 } 4018 4019 static uint32_t 4020 nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg) 4021 { 4022 uint32_t count; 4023 4024 /* Just assume that unassociated qpairs will eventually be io 4025 * qpairs. This is close enough for the use cases for this 4026 * function. 4027 */ 4028 pthread_mutex_lock(&pg->mutex); 4029 count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs; 4030 pthread_mutex_unlock(&pg->mutex); 4031 4032 return count; 4033 } 4034 4035 static struct spdk_nvmf_transport_poll_group * 4036 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 4037 { 4038 struct spdk_nvmf_rdma_transport *rtransport; 4039 struct spdk_nvmf_rdma_poll_group **pg; 4040 struct spdk_nvmf_transport_poll_group *result; 4041 uint32_t count; 4042 4043 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4044 4045 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 4046 return NULL; 4047 } 4048 4049 if (qpair->qid == 0) { 4050 pg = &rtransport->conn_sched.next_admin_pg; 4051 } else { 4052 struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current; 4053 uint32_t min_value; 4054 4055 pg = &rtransport->conn_sched.next_io_pg; 4056 pg_min = *pg; 4057 pg_start = *pg; 4058 pg_current = *pg; 4059 min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 4060 4061 while ((count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group)) > 0) { 4062 pg_current = TAILQ_NEXT(pg_current, link); 4063 if (pg_current == NULL) { 4064 pg_current = TAILQ_FIRST(&rtransport->poll_groups); 4065 } 4066 4067 if (count < min_value) { 4068 min_value = count; 4069 pg_min = pg_current; 4070 } 4071 4072 if (pg_current == pg_start) { 4073 break; 4074 } 4075 } 4076 *pg = pg_min; 4077 } 4078 4079 assert(*pg != NULL); 4080 4081 result = &(*pg)->group; 4082 4083 *pg = TAILQ_NEXT(*pg, link); 4084 if (*pg == NULL) { 4085 *pg = TAILQ_FIRST(&rtransport->poll_groups); 4086 } 4087 4088 return result; 4089 } 4090 4091 static void 4092 nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller) 4093 { 4094 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 4095 int rc; 4096 4097 TAILQ_REMOVE(&poller->group->pollers, poller, link); 4098 RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) { 4099 nvmf_rdma_qpair_destroy(qpair); 4100 } 4101 4102 if (poller->srq) { 4103 if (poller->resources) { 4104 nvmf_rdma_resources_destroy(poller->resources); 4105 } 4106 spdk_rdma_srq_destroy(poller->srq); 4107 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 4108 } 4109 4110 if (poller->cq) { 4111 rc = ibv_destroy_cq(poller->cq); 4112 if (rc != 0) { 4113 SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno)); 4114 } 4115 } 4116 4117 if (poller->destroy_cb) { 4118 poller->destroy_cb(poller->destroy_cb_ctx); 4119 poller->destroy_cb = NULL; 4120 } 4121 4122 free(poller); 4123 } 4124 4125 static void 4126 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 4127 { 4128 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 4129 struct spdk_nvmf_rdma_poller *poller, *tmp; 4130 struct spdk_nvmf_rdma_transport *rtransport; 4131 4132 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4133 if (!rgroup) { 4134 return; 4135 } 4136 4137 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 4138 nvmf_rdma_poller_destroy(poller); 4139 } 4140 4141 if (rgroup->group.transport == NULL) { 4142 /* Transport can be NULL when nvmf_rdma_poll_group_create() 4143 * calls this function directly in a failure path. */ 4144 free(rgroup); 4145 return; 4146 } 4147 4148 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 4149 4150 next_rgroup = TAILQ_NEXT(rgroup, link); 4151 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 4152 if (next_rgroup == NULL) { 4153 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 4154 } 4155 if (rtransport->conn_sched.next_admin_pg == rgroup) { 4156 rtransport->conn_sched.next_admin_pg = next_rgroup; 4157 } 4158 if (rtransport->conn_sched.next_io_pg == rgroup) { 4159 rtransport->conn_sched.next_io_pg = next_rgroup; 4160 } 4161 4162 free(rgroup); 4163 } 4164 4165 static void 4166 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 4167 { 4168 if (rqpair->cm_id != NULL) { 4169 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 4170 } 4171 } 4172 4173 static int 4174 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 4175 struct spdk_nvmf_qpair *qpair) 4176 { 4177 struct spdk_nvmf_rdma_poll_group *rgroup; 4178 struct spdk_nvmf_rdma_qpair *rqpair; 4179 struct spdk_nvmf_rdma_device *device; 4180 struct spdk_nvmf_rdma_poller *poller; 4181 int rc; 4182 4183 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4184 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4185 4186 device = rqpair->device; 4187 4188 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 4189 if (poller->device == device) { 4190 break; 4191 } 4192 } 4193 4194 if (!poller) { 4195 SPDK_ERRLOG("No poller found for device.\n"); 4196 return -1; 4197 } 4198 4199 if (poller->need_destroy) { 4200 SPDK_ERRLOG("Poller is destroying.\n"); 4201 return -1; 4202 } 4203 4204 rqpair->poller = poller; 4205 rqpair->srq = rqpair->poller->srq; 4206 4207 rc = nvmf_rdma_qpair_initialize(qpair); 4208 if (rc < 0) { 4209 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 4210 rqpair->poller = NULL; 4211 rqpair->srq = NULL; 4212 return -1; 4213 } 4214 4215 RB_INSERT(qpairs_tree, &poller->qpairs, rqpair); 4216 4217 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 4218 if (rc) { 4219 /* Try to reject, but we probably can't */ 4220 nvmf_rdma_qpair_reject_connection(rqpair); 4221 return -1; 4222 } 4223 4224 nvmf_rdma_update_ibv_state(rqpair); 4225 4226 return 0; 4227 } 4228 4229 static int 4230 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 4231 struct spdk_nvmf_qpair *qpair) 4232 { 4233 struct spdk_nvmf_rdma_qpair *rqpair; 4234 4235 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4236 assert(group->transport->tgt != NULL); 4237 4238 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 4239 4240 if (!rqpair->destruct_channel) { 4241 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 4242 return 0; 4243 } 4244 4245 /* Sanity check that we get io_channel on the correct thread */ 4246 if (qpair->group) { 4247 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 4248 } 4249 4250 return 0; 4251 } 4252 4253 static int 4254 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 4255 { 4256 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 4257 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4258 struct spdk_nvmf_rdma_transport, transport); 4259 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4260 struct spdk_nvmf_rdma_qpair, qpair); 4261 4262 /* 4263 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 4264 * needs to be returned to the shared receive queue or the poll group will eventually be 4265 * starved of RECV structures. 4266 */ 4267 if (rqpair->srq && rdma_req->recv) { 4268 int rc; 4269 struct ibv_recv_wr *bad_recv_wr; 4270 4271 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 4272 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 4273 if (rc) { 4274 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 4275 } 4276 } 4277 4278 _nvmf_rdma_request_free(rdma_req, rtransport); 4279 return 0; 4280 } 4281 4282 static int 4283 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 4284 { 4285 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4286 struct spdk_nvmf_rdma_transport, transport); 4287 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 4288 struct spdk_nvmf_rdma_request, req); 4289 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4290 struct spdk_nvmf_rdma_qpair, qpair); 4291 4292 if (rqpair->ibv_state != IBV_QPS_ERR) { 4293 /* The connection is alive, so process the request as normal */ 4294 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 4295 } else { 4296 /* The connection is dead. Move the request directly to the completed state. */ 4297 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4298 } 4299 4300 nvmf_rdma_request_process(rtransport, rdma_req); 4301 4302 return 0; 4303 } 4304 4305 static void 4306 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 4307 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 4308 { 4309 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4310 4311 rqpair->to_close = true; 4312 4313 /* This happens only when the qpair is disconnected before 4314 * it is added to the poll group. Since there is no poll group, 4315 * the RDMA qp has not been initialized yet and the RDMA CM 4316 * event has not yet been acknowledged, so we need to reject it. 4317 */ 4318 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 4319 nvmf_rdma_qpair_reject_connection(rqpair); 4320 nvmf_rdma_qpair_destroy(rqpair); 4321 return; 4322 } 4323 4324 if (rqpair->rdma_qp) { 4325 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 4326 } 4327 4328 nvmf_rdma_destroy_drained_qpair(rqpair); 4329 4330 if (cb_fn) { 4331 cb_fn(cb_arg); 4332 } 4333 } 4334 4335 static struct spdk_nvmf_rdma_qpair * 4336 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 4337 { 4338 struct spdk_nvmf_rdma_qpair find; 4339 4340 find.qp_num = wc->qp_num; 4341 4342 return RB_FIND(qpairs_tree, &rpoller->qpairs, &find); 4343 } 4344 4345 #ifdef DEBUG 4346 static int 4347 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 4348 { 4349 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 4350 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 4351 } 4352 #endif 4353 4354 static void 4355 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 4356 int rc) 4357 { 4358 struct spdk_nvmf_rdma_recv *rdma_recv; 4359 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4360 4361 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 4362 while (bad_recv_wr != NULL) { 4363 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 4364 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4365 4366 rdma_recv->qpair->current_recv_depth++; 4367 bad_recv_wr = bad_recv_wr->next; 4368 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 4369 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 4370 } 4371 } 4372 4373 static void 4374 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 4375 { 4376 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 4377 while (bad_recv_wr != NULL) { 4378 bad_recv_wr = bad_recv_wr->next; 4379 rqpair->current_recv_depth++; 4380 } 4381 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4382 } 4383 4384 static void 4385 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 4386 struct spdk_nvmf_rdma_poller *rpoller) 4387 { 4388 struct spdk_nvmf_rdma_qpair *rqpair; 4389 struct ibv_recv_wr *bad_recv_wr; 4390 int rc; 4391 4392 if (rpoller->srq) { 4393 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 4394 if (rc) { 4395 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 4396 } 4397 } else { 4398 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 4399 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 4400 rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 4401 if (rc) { 4402 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 4403 } 4404 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 4405 } 4406 } 4407 } 4408 4409 static void 4410 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 4411 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 4412 { 4413 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4414 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 4415 4416 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 4417 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 4418 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 4419 assert(rqpair->current_send_depth > 0); 4420 rqpair->current_send_depth--; 4421 switch (bad_rdma_wr->type) { 4422 case RDMA_WR_TYPE_DATA: 4423 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4424 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 4425 assert(rqpair->current_read_depth > 0); 4426 rqpair->current_read_depth--; 4427 } 4428 break; 4429 case RDMA_WR_TYPE_SEND: 4430 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4431 break; 4432 default: 4433 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 4434 prev_rdma_req = cur_rdma_req; 4435 continue; 4436 } 4437 4438 if (prev_rdma_req == cur_rdma_req) { 4439 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 4440 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 4441 continue; 4442 } 4443 4444 switch (cur_rdma_req->state) { 4445 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4446 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 4447 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4448 break; 4449 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 4450 case RDMA_REQUEST_STATE_COMPLETING: 4451 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4452 break; 4453 default: 4454 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 4455 cur_rdma_req->state, rqpair); 4456 continue; 4457 } 4458 4459 nvmf_rdma_request_process(rtransport, cur_rdma_req); 4460 prev_rdma_req = cur_rdma_req; 4461 } 4462 4463 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 4464 /* Disconnect the connection. */ 4465 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4466 } 4467 4468 } 4469 4470 static void 4471 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 4472 struct spdk_nvmf_rdma_poller *rpoller) 4473 { 4474 struct spdk_nvmf_rdma_qpair *rqpair; 4475 struct ibv_send_wr *bad_wr = NULL; 4476 int rc; 4477 4478 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 4479 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 4480 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 4481 4482 /* bad wr always points to the first wr that failed. */ 4483 if (rc) { 4484 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 4485 } 4486 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 4487 } 4488 } 4489 4490 static const char * 4491 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 4492 { 4493 switch (wr_type) { 4494 case RDMA_WR_TYPE_RECV: 4495 return "RECV"; 4496 case RDMA_WR_TYPE_SEND: 4497 return "SEND"; 4498 case RDMA_WR_TYPE_DATA: 4499 return "DATA"; 4500 default: 4501 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 4502 SPDK_UNREACHABLE(); 4503 } 4504 } 4505 4506 static inline void 4507 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 4508 { 4509 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 4510 4511 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 4512 /* If qpair is in ERR state, we will receive completions for all posted and not completed 4513 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 4514 SPDK_DEBUGLOG(rdma, 4515 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 4516 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 4517 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4518 } else { 4519 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 4520 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 4521 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4522 } 4523 } 4524 4525 static int 4526 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 4527 struct spdk_nvmf_rdma_poller *rpoller) 4528 { 4529 struct ibv_wc wc[32]; 4530 struct spdk_nvmf_rdma_wr *rdma_wr; 4531 struct spdk_nvmf_rdma_request *rdma_req; 4532 struct spdk_nvmf_rdma_recv *rdma_recv; 4533 struct spdk_nvmf_rdma_qpair *rqpair, *tmp_rqpair; 4534 int reaped, i; 4535 int count = 0; 4536 bool error = false; 4537 uint64_t poll_tsc = spdk_get_ticks(); 4538 4539 if (spdk_unlikely(rpoller->need_destroy)) { 4540 /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not 4541 * be called because we cannot poll anything from cq. So we call that here to force 4542 * destroy the qpair after to_close turning true. 4543 */ 4544 RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) { 4545 nvmf_rdma_destroy_drained_qpair(rqpair); 4546 } 4547 return 0; 4548 } 4549 4550 /* Poll for completing operations. */ 4551 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 4552 if (reaped < 0) { 4553 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 4554 errno, spdk_strerror(errno)); 4555 return -1; 4556 } else if (reaped == 0) { 4557 rpoller->stat.idle_polls++; 4558 } 4559 4560 rpoller->stat.polls++; 4561 rpoller->stat.completions += reaped; 4562 4563 for (i = 0; i < reaped; i++) { 4564 4565 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 4566 4567 switch (rdma_wr->type) { 4568 case RDMA_WR_TYPE_SEND: 4569 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4570 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4571 4572 if (!wc[i].status) { 4573 count++; 4574 assert(wc[i].opcode == IBV_WC_SEND); 4575 assert(nvmf_rdma_req_is_completing(rdma_req)); 4576 } 4577 4578 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4579 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 4580 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 4581 rdma_req->num_outstanding_data_wr = 0; 4582 4583 nvmf_rdma_request_process(rtransport, rdma_req); 4584 break; 4585 case RDMA_WR_TYPE_RECV: 4586 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 4587 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4588 if (rpoller->srq != NULL) { 4589 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 4590 /* It is possible that there are still some completions for destroyed QP 4591 * associated with SRQ. We just ignore these late completions and re-post 4592 * receive WRs back to SRQ. 4593 */ 4594 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 4595 struct ibv_recv_wr *bad_wr; 4596 int rc; 4597 4598 rdma_recv->wr.next = NULL; 4599 spdk_rdma_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 4600 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 4601 if (rc) { 4602 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 4603 } 4604 continue; 4605 } 4606 } 4607 rqpair = rdma_recv->qpair; 4608 4609 assert(rqpair != NULL); 4610 if (!wc[i].status) { 4611 assert(wc[i].opcode == IBV_WC_RECV); 4612 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 4613 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4614 break; 4615 } 4616 } 4617 4618 rdma_recv->wr.next = NULL; 4619 rqpair->current_recv_depth++; 4620 rdma_recv->receive_tsc = poll_tsc; 4621 rpoller->stat.requests++; 4622 STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link); 4623 break; 4624 case RDMA_WR_TYPE_DATA: 4625 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4626 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4627 4628 assert(rdma_req->num_outstanding_data_wr > 0); 4629 4630 rqpair->current_send_depth--; 4631 rdma_req->num_outstanding_data_wr--; 4632 if (!wc[i].status) { 4633 assert(wc[i].opcode == IBV_WC_RDMA_READ); 4634 rqpair->current_read_depth--; 4635 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 4636 if (rdma_req->num_outstanding_data_wr == 0) { 4637 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 4638 nvmf_rdma_request_process(rtransport, rdma_req); 4639 } 4640 } else { 4641 /* If the data transfer fails still force the queue into the error state, 4642 * if we were performing an RDMA_READ, we need to force the request into a 4643 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 4644 * case, we should wait for the SEND to complete. */ 4645 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 4646 rqpair->current_read_depth--; 4647 if (rdma_req->num_outstanding_data_wr == 0) { 4648 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4649 } 4650 } 4651 } 4652 break; 4653 default: 4654 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 4655 continue; 4656 } 4657 4658 /* Handle error conditions */ 4659 if (wc[i].status) { 4660 nvmf_rdma_update_ibv_state(rqpair); 4661 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 4662 4663 error = true; 4664 4665 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 4666 /* Disconnect the connection. */ 4667 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4668 } else { 4669 nvmf_rdma_destroy_drained_qpair(rqpair); 4670 } 4671 continue; 4672 } 4673 4674 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 4675 4676 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 4677 nvmf_rdma_destroy_drained_qpair(rqpair); 4678 } 4679 } 4680 4681 if (error == true) { 4682 return -1; 4683 } 4684 4685 /* submit outstanding work requests. */ 4686 _poller_submit_recvs(rtransport, rpoller); 4687 _poller_submit_sends(rtransport, rpoller); 4688 4689 return count; 4690 } 4691 4692 static void 4693 _nvmf_rdma_remove_destroyed_device(void *c) 4694 { 4695 struct spdk_nvmf_rdma_transport *rtransport = c; 4696 struct spdk_nvmf_rdma_device *device, *device_tmp; 4697 int rc; 4698 4699 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 4700 if (device->ready_to_destroy) { 4701 destroy_ib_device(rtransport, device); 4702 } 4703 } 4704 4705 free_poll_fds(rtransport); 4706 rc = generate_poll_fds(rtransport); 4707 /* cannot handle fd allocation error here */ 4708 if (rc != 0) { 4709 SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n"); 4710 } 4711 } 4712 4713 static void 4714 _nvmf_rdma_remove_poller_in_group_cb(void *c) 4715 { 4716 struct poller_manage_ctx *ctx = c; 4717 struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport; 4718 struct spdk_nvmf_rdma_device *device = ctx->device; 4719 struct spdk_thread *thread = ctx->thread; 4720 4721 if (nvmf_rdma_all_pollers_management_done(c)) { 4722 /* destroy device when last poller is destroyed */ 4723 device->ready_to_destroy = true; 4724 spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport); 4725 } 4726 } 4727 4728 static void 4729 _nvmf_rdma_remove_poller_in_group(void *c) 4730 { 4731 struct poller_manage_ctx *ctx = c; 4732 4733 ctx->rpoller->need_destroy = true; 4734 ctx->rpoller->destroy_cb_ctx = ctx; 4735 ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb; 4736 4737 /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */ 4738 if (RB_EMPTY(&ctx->rpoller->qpairs)) { 4739 nvmf_rdma_poller_destroy(ctx->rpoller); 4740 } 4741 } 4742 4743 static int 4744 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 4745 { 4746 struct spdk_nvmf_rdma_transport *rtransport; 4747 struct spdk_nvmf_rdma_poll_group *rgroup; 4748 struct spdk_nvmf_rdma_poller *rpoller, *tmp; 4749 int count, rc; 4750 4751 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 4752 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4753 4754 count = 0; 4755 TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) { 4756 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4757 if (rc < 0) { 4758 return rc; 4759 } 4760 count += rc; 4761 } 4762 4763 return count; 4764 } 4765 4766 static int 4767 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4768 struct spdk_nvme_transport_id *trid, 4769 bool peer) 4770 { 4771 struct sockaddr *saddr; 4772 uint16_t port; 4773 4774 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4775 4776 if (peer) { 4777 saddr = rdma_get_peer_addr(id); 4778 } else { 4779 saddr = rdma_get_local_addr(id); 4780 } 4781 switch (saddr->sa_family) { 4782 case AF_INET: { 4783 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4784 4785 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4786 inet_ntop(AF_INET, &saddr_in->sin_addr, 4787 trid->traddr, sizeof(trid->traddr)); 4788 if (peer) { 4789 port = ntohs(rdma_get_dst_port(id)); 4790 } else { 4791 port = ntohs(rdma_get_src_port(id)); 4792 } 4793 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4794 break; 4795 } 4796 case AF_INET6: { 4797 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4798 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4799 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4800 trid->traddr, sizeof(trid->traddr)); 4801 if (peer) { 4802 port = ntohs(rdma_get_dst_port(id)); 4803 } else { 4804 port = ntohs(rdma_get_src_port(id)); 4805 } 4806 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4807 break; 4808 } 4809 default: 4810 return -1; 4811 4812 } 4813 4814 return 0; 4815 } 4816 4817 static int 4818 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4819 struct spdk_nvme_transport_id *trid) 4820 { 4821 struct spdk_nvmf_rdma_qpair *rqpair; 4822 4823 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4824 4825 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4826 } 4827 4828 static int 4829 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4830 struct spdk_nvme_transport_id *trid) 4831 { 4832 struct spdk_nvmf_rdma_qpair *rqpair; 4833 4834 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4835 4836 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4837 } 4838 4839 static int 4840 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4841 struct spdk_nvme_transport_id *trid) 4842 { 4843 struct spdk_nvmf_rdma_qpair *rqpair; 4844 4845 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4846 4847 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4848 } 4849 4850 void 4851 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4852 { 4853 g_nvmf_hooks = *hooks; 4854 } 4855 4856 static void 4857 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4858 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4859 { 4860 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4861 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4862 4863 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4864 4865 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4866 } 4867 4868 static int 4869 _nvmf_rdma_qpair_abort_request(void *ctx) 4870 { 4871 struct spdk_nvmf_request *req = ctx; 4872 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4873 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4874 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4875 struct spdk_nvmf_rdma_qpair, qpair); 4876 int rc; 4877 4878 spdk_poller_unregister(&req->poller); 4879 4880 switch (rdma_req_to_abort->state) { 4881 case RDMA_REQUEST_STATE_EXECUTING: 4882 rc = nvmf_ctrlr_abort_request(req); 4883 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4884 return SPDK_POLLER_BUSY; 4885 } 4886 break; 4887 4888 case RDMA_REQUEST_STATE_NEED_BUFFER: 4889 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4890 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4891 4892 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4893 break; 4894 4895 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4896 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4897 spdk_nvmf_rdma_request, state_link); 4898 4899 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4900 break; 4901 4902 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4903 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4904 spdk_nvmf_rdma_request, state_link); 4905 4906 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4907 break; 4908 4909 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4910 if (spdk_get_ticks() < req->timeout_tsc) { 4911 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4912 return SPDK_POLLER_BUSY; 4913 } 4914 break; 4915 4916 default: 4917 break; 4918 } 4919 4920 spdk_nvmf_request_complete(req); 4921 return SPDK_POLLER_BUSY; 4922 } 4923 4924 static void 4925 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4926 struct spdk_nvmf_request *req) 4927 { 4928 struct spdk_nvmf_rdma_qpair *rqpair; 4929 struct spdk_nvmf_rdma_transport *rtransport; 4930 struct spdk_nvmf_transport *transport; 4931 uint16_t cid; 4932 uint32_t i, max_req_count; 4933 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 4934 4935 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4936 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4937 transport = &rtransport->transport; 4938 4939 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4940 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 4941 4942 for (i = 0; i < max_req_count; i++) { 4943 rdma_req = &rqpair->resources->reqs[i]; 4944 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 4945 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 4946 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 4947 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 4948 rdma_req->req.qpair == qpair) { 4949 rdma_req_to_abort = rdma_req; 4950 break; 4951 } 4952 } 4953 4954 if (rdma_req_to_abort == NULL) { 4955 spdk_nvmf_request_complete(req); 4956 return; 4957 } 4958 4959 req->req_to_abort = &rdma_req_to_abort->req; 4960 req->timeout_tsc = spdk_get_ticks() + 4961 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4962 req->poller = NULL; 4963 4964 _nvmf_rdma_qpair_abort_request(req); 4965 } 4966 4967 static void 4968 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 4969 struct spdk_json_write_ctx *w) 4970 { 4971 struct spdk_nvmf_rdma_poll_group *rgroup; 4972 struct spdk_nvmf_rdma_poller *rpoller; 4973 4974 assert(w != NULL); 4975 4976 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4977 4978 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 4979 4980 spdk_json_write_named_array_begin(w, "devices"); 4981 4982 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4983 spdk_json_write_object_begin(w); 4984 spdk_json_write_named_string(w, "name", 4985 ibv_get_device_name(rpoller->device->context->device)); 4986 spdk_json_write_named_uint64(w, "polls", 4987 rpoller->stat.polls); 4988 spdk_json_write_named_uint64(w, "idle_polls", 4989 rpoller->stat.idle_polls); 4990 spdk_json_write_named_uint64(w, "completions", 4991 rpoller->stat.completions); 4992 spdk_json_write_named_uint64(w, "requests", 4993 rpoller->stat.requests); 4994 spdk_json_write_named_uint64(w, "request_latency", 4995 rpoller->stat.request_latency); 4996 spdk_json_write_named_uint64(w, "pending_free_request", 4997 rpoller->stat.pending_free_request); 4998 spdk_json_write_named_uint64(w, "pending_rdma_read", 4999 rpoller->stat.pending_rdma_read); 5000 spdk_json_write_named_uint64(w, "pending_rdma_write", 5001 rpoller->stat.pending_rdma_write); 5002 spdk_json_write_named_uint64(w, "total_send_wrs", 5003 rpoller->stat.qp_stats.send.num_submitted_wrs); 5004 spdk_json_write_named_uint64(w, "send_doorbell_updates", 5005 rpoller->stat.qp_stats.send.doorbell_updates); 5006 spdk_json_write_named_uint64(w, "total_recv_wrs", 5007 rpoller->stat.qp_stats.recv.num_submitted_wrs); 5008 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 5009 rpoller->stat.qp_stats.recv.doorbell_updates); 5010 spdk_json_write_object_end(w); 5011 } 5012 5013 spdk_json_write_array_end(w); 5014 } 5015 5016 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 5017 .name = "RDMA", 5018 .type = SPDK_NVME_TRANSPORT_RDMA, 5019 .opts_init = nvmf_rdma_opts_init, 5020 .create = nvmf_rdma_create, 5021 .dump_opts = nvmf_rdma_dump_opts, 5022 .destroy = nvmf_rdma_destroy, 5023 5024 .listen = nvmf_rdma_listen, 5025 .stop_listen = nvmf_rdma_stop_listen, 5026 .cdata_init = nvmf_rdma_cdata_init, 5027 5028 .listener_discover = nvmf_rdma_discover, 5029 5030 .poll_group_create = nvmf_rdma_poll_group_create, 5031 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 5032 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 5033 .poll_group_add = nvmf_rdma_poll_group_add, 5034 .poll_group_remove = nvmf_rdma_poll_group_remove, 5035 .poll_group_poll = nvmf_rdma_poll_group_poll, 5036 5037 .req_free = nvmf_rdma_request_free, 5038 .req_complete = nvmf_rdma_request_complete, 5039 5040 .qpair_fini = nvmf_rdma_close_qpair, 5041 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 5042 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 5043 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 5044 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 5045 5046 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 5047 }; 5048 5049 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 5050 SPDK_LOG_REGISTER_COMPONENT(rdma) 5051