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