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