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 if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) { 2510 SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n"); 2511 SPDK_ERRLOG("Probably running in multiprocess environment, which is " 2512 "unsupported by the nvmf library\n"); 2513 } else { 2514 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2515 } 2516 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2517 return NULL; 2518 } 2519 2520 contexts = rdma_get_devices(NULL); 2521 if (contexts == NULL) { 2522 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2523 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2524 return NULL; 2525 } 2526 2527 i = 0; 2528 rc = 0; 2529 while (contexts[i] != NULL) { 2530 device = calloc(1, sizeof(*device)); 2531 if (!device) { 2532 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2533 rc = -ENOMEM; 2534 break; 2535 } 2536 device->context = contexts[i]; 2537 rc = ibv_query_device(device->context, &device->attr); 2538 if (rc < 0) { 2539 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2540 free(device); 2541 break; 2542 2543 } 2544 2545 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2546 2547 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2548 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2549 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2550 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2551 } 2552 2553 /** 2554 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2555 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2556 * but incorrectly reports that it does. There are changes making their way 2557 * through the kernel now that will enable this feature. When they are merged, 2558 * we can conditionally enable this feature. 2559 * 2560 * TODO: enable this for versions of the kernel rxe driver that support it. 2561 */ 2562 if (nvmf_rdma_is_rxe_device(device)) { 2563 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2564 } 2565 #endif 2566 2567 /* set up device context async ev fd as NON_BLOCKING */ 2568 flag = fcntl(device->context->async_fd, F_GETFL); 2569 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2570 if (rc < 0) { 2571 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2572 free(device); 2573 break; 2574 } 2575 2576 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2577 i++; 2578 2579 if (g_nvmf_hooks.get_ibv_pd) { 2580 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2581 } else { 2582 device->pd = ibv_alloc_pd(device->context); 2583 } 2584 2585 if (!device->pd) { 2586 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2587 rc = -ENOMEM; 2588 break; 2589 } 2590 2591 assert(device->map == NULL); 2592 2593 device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks, SPDK_RDMA_MEMORY_MAP_ROLE_TARGET); 2594 if (!device->map) { 2595 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2596 rc = -ENOMEM; 2597 break; 2598 } 2599 2600 assert(device->map != NULL); 2601 assert(device->pd != NULL); 2602 } 2603 rdma_free_devices(contexts); 2604 2605 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2606 /* divide and round up. */ 2607 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2608 2609 /* round up to the nearest 4k. */ 2610 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2611 2612 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2613 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2614 opts->io_unit_size); 2615 } 2616 2617 if (rc < 0) { 2618 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2619 return NULL; 2620 } 2621 2622 /* Set up poll descriptor array to monitor events from RDMA and IB 2623 * in a single poll syscall 2624 */ 2625 rtransport->npoll_fds = i + 1; 2626 i = 0; 2627 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2628 if (rtransport->poll_fds == NULL) { 2629 SPDK_ERRLOG("poll_fds allocation failed\n"); 2630 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2631 return NULL; 2632 } 2633 2634 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2635 rtransport->poll_fds[i++].events = POLLIN; 2636 2637 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2638 rtransport->poll_fds[i].fd = device->context->async_fd; 2639 rtransport->poll_fds[i++].events = POLLIN; 2640 } 2641 2642 rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport, 2643 opts->acceptor_poll_rate); 2644 if (!rtransport->accept_poller) { 2645 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2646 return NULL; 2647 } 2648 2649 return &rtransport->transport; 2650 } 2651 2652 static void 2653 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2654 { 2655 struct spdk_nvmf_rdma_transport *rtransport; 2656 assert(w != NULL); 2657 2658 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2659 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2660 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2661 if (rtransport->rdma_opts.no_srq == true) { 2662 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2663 } 2664 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2665 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2666 } 2667 2668 static int 2669 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2670 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2671 { 2672 struct spdk_nvmf_rdma_transport *rtransport; 2673 struct spdk_nvmf_rdma_port *port, *port_tmp; 2674 struct spdk_nvmf_rdma_device *device, *device_tmp; 2675 2676 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2677 2678 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2679 TAILQ_REMOVE(&rtransport->ports, port, link); 2680 rdma_destroy_id(port->id); 2681 free(port); 2682 } 2683 2684 if (rtransport->poll_fds != NULL) { 2685 free(rtransport->poll_fds); 2686 } 2687 2688 if (rtransport->event_channel != NULL) { 2689 rdma_destroy_event_channel(rtransport->event_channel); 2690 } 2691 2692 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2693 TAILQ_REMOVE(&rtransport->devices, device, link); 2694 spdk_rdma_free_mem_map(&device->map); 2695 if (device->pd) { 2696 if (!g_nvmf_hooks.get_ibv_pd) { 2697 ibv_dealloc_pd(device->pd); 2698 } 2699 } 2700 free(device); 2701 } 2702 2703 if (rtransport->data_wr_pool != NULL) { 2704 if (spdk_mempool_count(rtransport->data_wr_pool) != 2705 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2706 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2707 spdk_mempool_count(rtransport->data_wr_pool), 2708 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2709 } 2710 } 2711 2712 spdk_mempool_free(rtransport->data_wr_pool); 2713 2714 spdk_poller_unregister(&rtransport->accept_poller); 2715 free(rtransport); 2716 2717 if (cb_fn) { 2718 cb_fn(cb_arg); 2719 } 2720 return 0; 2721 } 2722 2723 static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2724 struct spdk_nvme_transport_id *trid, 2725 bool peer); 2726 2727 static int 2728 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2729 struct spdk_nvmf_listen_opts *listen_opts) 2730 { 2731 struct spdk_nvmf_rdma_transport *rtransport; 2732 struct spdk_nvmf_rdma_device *device; 2733 struct spdk_nvmf_rdma_port *port; 2734 struct addrinfo *res; 2735 struct addrinfo hints; 2736 int family; 2737 int rc; 2738 2739 if (!strlen(trid->trsvcid)) { 2740 SPDK_ERRLOG("Service id is required\n"); 2741 return -EINVAL; 2742 } 2743 2744 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2745 assert(rtransport->event_channel != NULL); 2746 2747 port = calloc(1, sizeof(*port)); 2748 if (!port) { 2749 SPDK_ERRLOG("Port allocation failed\n"); 2750 return -ENOMEM; 2751 } 2752 2753 port->trid = trid; 2754 2755 switch (trid->adrfam) { 2756 case SPDK_NVMF_ADRFAM_IPV4: 2757 family = AF_INET; 2758 break; 2759 case SPDK_NVMF_ADRFAM_IPV6: 2760 family = AF_INET6; 2761 break; 2762 default: 2763 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2764 free(port); 2765 return -EINVAL; 2766 } 2767 2768 memset(&hints, 0, sizeof(hints)); 2769 hints.ai_family = family; 2770 hints.ai_flags = AI_NUMERICSERV; 2771 hints.ai_socktype = SOCK_STREAM; 2772 hints.ai_protocol = 0; 2773 2774 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2775 if (rc) { 2776 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2777 free(port); 2778 return -EINVAL; 2779 } 2780 2781 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2782 if (rc < 0) { 2783 SPDK_ERRLOG("rdma_create_id() failed\n"); 2784 freeaddrinfo(res); 2785 free(port); 2786 return rc; 2787 } 2788 2789 rc = rdma_bind_addr(port->id, res->ai_addr); 2790 freeaddrinfo(res); 2791 2792 if (rc < 0) { 2793 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2794 rdma_destroy_id(port->id); 2795 free(port); 2796 return rc; 2797 } 2798 2799 if (!port->id->verbs) { 2800 SPDK_ERRLOG("ibv_context is null\n"); 2801 rdma_destroy_id(port->id); 2802 free(port); 2803 return -1; 2804 } 2805 2806 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2807 if (rc < 0) { 2808 SPDK_ERRLOG("rdma_listen() failed\n"); 2809 rdma_destroy_id(port->id); 2810 free(port); 2811 return rc; 2812 } 2813 2814 TAILQ_FOREACH(device, &rtransport->devices, link) { 2815 if (device->context == port->id->verbs) { 2816 port->device = device; 2817 break; 2818 } 2819 } 2820 if (!port->device) { 2821 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2822 port->id->verbs); 2823 rdma_destroy_id(port->id); 2824 free(port); 2825 return -EINVAL; 2826 } 2827 2828 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2829 trid->traddr, trid->trsvcid); 2830 2831 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2832 return 0; 2833 } 2834 2835 static void 2836 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2837 const struct spdk_nvme_transport_id *trid) 2838 { 2839 struct spdk_nvmf_rdma_transport *rtransport; 2840 struct spdk_nvmf_rdma_port *port, *tmp; 2841 2842 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2843 2844 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2845 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2846 TAILQ_REMOVE(&rtransport->ports, port, link); 2847 rdma_destroy_id(port->id); 2848 free(port); 2849 break; 2850 } 2851 } 2852 } 2853 2854 static void 2855 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2856 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2857 { 2858 struct spdk_nvmf_request *req, *tmp; 2859 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2860 struct spdk_nvmf_rdma_resources *resources; 2861 2862 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2863 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2864 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2865 break; 2866 } 2867 } 2868 2869 /* Then RDMA writes since reads have stronger restrictions than writes */ 2870 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2871 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2872 break; 2873 } 2874 } 2875 2876 /* Then we handle request waiting on memory buffers. */ 2877 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2878 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2879 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2880 break; 2881 } 2882 } 2883 2884 resources = rqpair->resources; 2885 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2886 rdma_req = STAILQ_FIRST(&resources->free_queue); 2887 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2888 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2889 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2890 2891 if (rqpair->srq != NULL) { 2892 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2893 rdma_req->recv->qpair->qd++; 2894 } else { 2895 rqpair->qd++; 2896 } 2897 2898 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2899 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2900 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2901 break; 2902 } 2903 } 2904 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2905 rqpair->poller->stat.pending_free_request++; 2906 } 2907 } 2908 2909 static inline bool 2910 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 2911 { 2912 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 2913 return nvmf_rdma_is_rxe_device(device) || 2914 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 2915 } 2916 2917 static void 2918 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2919 { 2920 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2921 struct spdk_nvmf_rdma_transport, transport); 2922 2923 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2924 2925 /* nvmr_rdma_close_qpair is not called */ 2926 if (!rqpair->to_close) { 2927 return; 2928 } 2929 2930 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2931 if (rqpair->current_send_depth != 0) { 2932 return; 2933 } 2934 2935 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2936 return; 2937 } 2938 2939 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 2940 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 2941 return; 2942 } 2943 2944 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2945 2946 nvmf_rdma_qpair_destroy(rqpair); 2947 } 2948 2949 static int 2950 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2951 { 2952 struct spdk_nvmf_qpair *qpair; 2953 struct spdk_nvmf_rdma_qpair *rqpair; 2954 2955 if (evt->id == NULL) { 2956 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2957 return -1; 2958 } 2959 2960 qpair = evt->id->context; 2961 if (qpair == NULL) { 2962 SPDK_ERRLOG("disconnect request: no active connection\n"); 2963 return -1; 2964 } 2965 2966 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2967 2968 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair); 2969 2970 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2971 2972 return 0; 2973 } 2974 2975 #ifdef DEBUG 2976 static const char *CM_EVENT_STR[] = { 2977 "RDMA_CM_EVENT_ADDR_RESOLVED", 2978 "RDMA_CM_EVENT_ADDR_ERROR", 2979 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2980 "RDMA_CM_EVENT_ROUTE_ERROR", 2981 "RDMA_CM_EVENT_CONNECT_REQUEST", 2982 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2983 "RDMA_CM_EVENT_CONNECT_ERROR", 2984 "RDMA_CM_EVENT_UNREACHABLE", 2985 "RDMA_CM_EVENT_REJECTED", 2986 "RDMA_CM_EVENT_ESTABLISHED", 2987 "RDMA_CM_EVENT_DISCONNECTED", 2988 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2989 "RDMA_CM_EVENT_MULTICAST_JOIN", 2990 "RDMA_CM_EVENT_MULTICAST_ERROR", 2991 "RDMA_CM_EVENT_ADDR_CHANGE", 2992 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2993 }; 2994 #endif /* DEBUG */ 2995 2996 static void 2997 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2998 struct spdk_nvmf_rdma_port *port) 2999 { 3000 struct spdk_nvmf_rdma_poll_group *rgroup; 3001 struct spdk_nvmf_rdma_poller *rpoller; 3002 struct spdk_nvmf_rdma_qpair *rqpair; 3003 3004 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3005 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3006 RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) { 3007 if (rqpair->listen_id == port->id) { 3008 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3009 } 3010 } 3011 } 3012 } 3013 } 3014 3015 static bool 3016 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 3017 struct rdma_cm_event *event) 3018 { 3019 const struct spdk_nvme_transport_id *trid; 3020 struct spdk_nvmf_rdma_port *port; 3021 struct spdk_nvmf_rdma_transport *rtransport; 3022 bool event_acked = false; 3023 3024 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3025 TAILQ_FOREACH(port, &rtransport->ports, link) { 3026 if (port->id == event->id) { 3027 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 3028 rdma_ack_cm_event(event); 3029 event_acked = true; 3030 trid = port->trid; 3031 break; 3032 } 3033 } 3034 3035 if (event_acked) { 3036 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 3037 3038 nvmf_rdma_stop_listen(transport, trid); 3039 nvmf_rdma_listen(transport, trid, NULL); 3040 } 3041 3042 return event_acked; 3043 } 3044 3045 static void 3046 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 3047 struct rdma_cm_event *event) 3048 { 3049 struct spdk_nvmf_rdma_port *port; 3050 struct spdk_nvmf_rdma_transport *rtransport; 3051 3052 port = event->id->context; 3053 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3054 3055 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 3056 3057 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 3058 3059 rdma_ack_cm_event(event); 3060 3061 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 3062 ; 3063 } 3064 } 3065 3066 static void 3067 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 3068 { 3069 struct spdk_nvmf_rdma_transport *rtransport; 3070 struct rdma_cm_event *event; 3071 int rc; 3072 bool event_acked; 3073 3074 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3075 3076 if (rtransport->event_channel == NULL) { 3077 return; 3078 } 3079 3080 while (1) { 3081 event_acked = false; 3082 rc = rdma_get_cm_event(rtransport->event_channel, &event); 3083 if (rc) { 3084 if (errno != EAGAIN && errno != EWOULDBLOCK) { 3085 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 3086 } 3087 break; 3088 } 3089 3090 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 3091 3092 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 3093 3094 switch (event->event) { 3095 case RDMA_CM_EVENT_ADDR_RESOLVED: 3096 case RDMA_CM_EVENT_ADDR_ERROR: 3097 case RDMA_CM_EVENT_ROUTE_RESOLVED: 3098 case RDMA_CM_EVENT_ROUTE_ERROR: 3099 /* No action required. The target never attempts to resolve routes. */ 3100 break; 3101 case RDMA_CM_EVENT_CONNECT_REQUEST: 3102 rc = nvmf_rdma_connect(transport, event); 3103 if (rc < 0) { 3104 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 3105 break; 3106 } 3107 break; 3108 case RDMA_CM_EVENT_CONNECT_RESPONSE: 3109 /* The target never initiates a new connection. So this will not occur. */ 3110 break; 3111 case RDMA_CM_EVENT_CONNECT_ERROR: 3112 /* Can this happen? The docs say it can, but not sure what causes it. */ 3113 break; 3114 case RDMA_CM_EVENT_UNREACHABLE: 3115 case RDMA_CM_EVENT_REJECTED: 3116 /* These only occur on the client side. */ 3117 break; 3118 case RDMA_CM_EVENT_ESTABLISHED: 3119 /* TODO: Should we be waiting for this event anywhere? */ 3120 break; 3121 case RDMA_CM_EVENT_DISCONNECTED: 3122 rc = nvmf_rdma_disconnect(event); 3123 if (rc < 0) { 3124 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3125 break; 3126 } 3127 break; 3128 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3129 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3130 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3131 * Once these events are sent to SPDK, we should release all IB resources and 3132 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3133 * ibv_destroy* functions to release user space memory allocated by IB. All kernel 3134 * resources are already cleaned. */ 3135 if (event->id->qp) { 3136 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3137 * corresponding qpair. Otherwise the event refers to a listening device */ 3138 rc = nvmf_rdma_disconnect(event); 3139 if (rc < 0) { 3140 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3141 break; 3142 } 3143 } else { 3144 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3145 event_acked = true; 3146 } 3147 break; 3148 case RDMA_CM_EVENT_MULTICAST_JOIN: 3149 case RDMA_CM_EVENT_MULTICAST_ERROR: 3150 /* Multicast is not used */ 3151 break; 3152 case RDMA_CM_EVENT_ADDR_CHANGE: 3153 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3154 break; 3155 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3156 /* For now, do nothing. The target never re-uses queue pairs. */ 3157 break; 3158 default: 3159 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3160 break; 3161 } 3162 if (!event_acked) { 3163 rdma_ack_cm_event(event); 3164 } 3165 } 3166 } 3167 3168 static void 3169 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3170 { 3171 rqpair->last_wqe_reached = true; 3172 nvmf_rdma_destroy_drained_qpair(rqpair); 3173 } 3174 3175 static void 3176 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3177 { 3178 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3179 3180 if (event_ctx->rqpair) { 3181 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3182 if (event_ctx->cb_fn) { 3183 event_ctx->cb_fn(event_ctx->rqpair); 3184 } 3185 } 3186 free(event_ctx); 3187 } 3188 3189 static int 3190 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3191 spdk_nvmf_rdma_qpair_ibv_event fn) 3192 { 3193 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3194 struct spdk_thread *thr = NULL; 3195 int rc; 3196 3197 if (rqpair->qpair.group) { 3198 thr = rqpair->qpair.group->thread; 3199 } else if (rqpair->destruct_channel) { 3200 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3201 } 3202 3203 if (!thr) { 3204 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3205 return -EINVAL; 3206 } 3207 3208 ctx = calloc(1, sizeof(*ctx)); 3209 if (!ctx) { 3210 return -ENOMEM; 3211 } 3212 3213 ctx->rqpair = rqpair; 3214 ctx->cb_fn = fn; 3215 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3216 3217 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3218 if (rc) { 3219 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3220 free(ctx); 3221 } 3222 3223 return rc; 3224 } 3225 3226 static int 3227 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3228 { 3229 int rc; 3230 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3231 struct ibv_async_event event; 3232 3233 rc = ibv_get_async_event(device->context, &event); 3234 3235 if (rc) { 3236 /* In non-blocking mode -1 means there are no events available */ 3237 return rc; 3238 } 3239 3240 switch (event.event_type) { 3241 case IBV_EVENT_QP_FATAL: 3242 rqpair = event.element.qp->qp_context; 3243 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3244 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3245 (uintptr_t)rqpair, event.event_type); 3246 nvmf_rdma_update_ibv_state(rqpair); 3247 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3248 break; 3249 case IBV_EVENT_QP_LAST_WQE_REACHED: 3250 /* This event only occurs for shared receive queues. */ 3251 rqpair = event.element.qp->qp_context; 3252 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3253 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3254 if (rc) { 3255 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3256 rqpair->last_wqe_reached = true; 3257 } 3258 break; 3259 case IBV_EVENT_SQ_DRAINED: 3260 /* This event occurs frequently in both error and non-error states. 3261 * Check if the qpair is in an error state before sending a message. */ 3262 rqpair = event.element.qp->qp_context; 3263 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3264 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3265 (uintptr_t)rqpair, event.event_type); 3266 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3267 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3268 } 3269 break; 3270 case IBV_EVENT_QP_REQ_ERR: 3271 case IBV_EVENT_QP_ACCESS_ERR: 3272 case IBV_EVENT_COMM_EST: 3273 case IBV_EVENT_PATH_MIG: 3274 case IBV_EVENT_PATH_MIG_ERR: 3275 SPDK_NOTICELOG("Async event: %s\n", 3276 ibv_event_type_str(event.event_type)); 3277 rqpair = event.element.qp->qp_context; 3278 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3279 (uintptr_t)rqpair, event.event_type); 3280 nvmf_rdma_update_ibv_state(rqpair); 3281 break; 3282 case IBV_EVENT_CQ_ERR: 3283 case IBV_EVENT_DEVICE_FATAL: 3284 case IBV_EVENT_PORT_ACTIVE: 3285 case IBV_EVENT_PORT_ERR: 3286 case IBV_EVENT_LID_CHANGE: 3287 case IBV_EVENT_PKEY_CHANGE: 3288 case IBV_EVENT_SM_CHANGE: 3289 case IBV_EVENT_SRQ_ERR: 3290 case IBV_EVENT_SRQ_LIMIT_REACHED: 3291 case IBV_EVENT_CLIENT_REREGISTER: 3292 case IBV_EVENT_GID_CHANGE: 3293 default: 3294 SPDK_NOTICELOG("Async event: %s\n", 3295 ibv_event_type_str(event.event_type)); 3296 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3297 break; 3298 } 3299 ibv_ack_async_event(&event); 3300 3301 return 0; 3302 } 3303 3304 static void 3305 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3306 { 3307 int rc = 0; 3308 uint32_t i = 0; 3309 3310 for (i = 0; i < max_events; i++) { 3311 rc = nvmf_process_ib_event(device); 3312 if (rc) { 3313 break; 3314 } 3315 } 3316 3317 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3318 } 3319 3320 static int 3321 nvmf_rdma_accept(void *ctx) 3322 { 3323 int nfds, i = 0; 3324 struct spdk_nvmf_transport *transport = ctx; 3325 struct spdk_nvmf_rdma_transport *rtransport; 3326 struct spdk_nvmf_rdma_device *device, *tmp; 3327 uint32_t count; 3328 3329 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3330 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3331 3332 if (nfds <= 0) { 3333 return SPDK_POLLER_IDLE; 3334 } 3335 3336 /* The first poll descriptor is RDMA CM event */ 3337 if (rtransport->poll_fds[i++].revents & POLLIN) { 3338 nvmf_process_cm_event(transport); 3339 nfds--; 3340 } 3341 3342 if (nfds == 0) { 3343 return SPDK_POLLER_BUSY; 3344 } 3345 3346 /* Second and subsequent poll descriptors are IB async events */ 3347 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3348 if (rtransport->poll_fds[i++].revents & POLLIN) { 3349 nvmf_process_ib_events(device, 32); 3350 nfds--; 3351 } 3352 } 3353 /* check all flagged fd's have been served */ 3354 assert(nfds == 0); 3355 3356 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3357 } 3358 3359 static void 3360 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3361 struct spdk_nvmf_ctrlr_data *cdata) 3362 { 3363 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3364 3365 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3366 since in-capsule data only works with NVME drives that support SGL memory layout */ 3367 if (transport->opts.dif_insert_or_strip) { 3368 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3369 } 3370 3371 if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) { 3372 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule" 3373 " data is greater than 4KiB.\n"); 3374 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux " 3375 "kernel between versions 5.4 and 5.12 data corruption may occur for " 3376 "writes that are not a multiple of 4KiB in size.\n"); 3377 } 3378 } 3379 3380 static void 3381 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3382 struct spdk_nvme_transport_id *trid, 3383 struct spdk_nvmf_discovery_log_page_entry *entry) 3384 { 3385 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3386 entry->adrfam = trid->adrfam; 3387 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3388 3389 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3390 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3391 3392 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3393 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3394 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3395 } 3396 3397 static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3398 3399 static struct spdk_nvmf_transport_poll_group * 3400 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport, 3401 struct spdk_nvmf_poll_group *group) 3402 { 3403 struct spdk_nvmf_rdma_transport *rtransport; 3404 struct spdk_nvmf_rdma_poll_group *rgroup; 3405 struct spdk_nvmf_rdma_poller *poller; 3406 struct spdk_nvmf_rdma_device *device; 3407 struct spdk_rdma_srq_init_attr srq_init_attr; 3408 struct spdk_nvmf_rdma_resource_opts opts; 3409 int num_cqe; 3410 3411 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3412 3413 rgroup = calloc(1, sizeof(*rgroup)); 3414 if (!rgroup) { 3415 return NULL; 3416 } 3417 3418 TAILQ_INIT(&rgroup->pollers); 3419 3420 TAILQ_FOREACH(device, &rtransport->devices, link) { 3421 poller = calloc(1, sizeof(*poller)); 3422 if (!poller) { 3423 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3424 nvmf_rdma_poll_group_destroy(&rgroup->group); 3425 return NULL; 3426 } 3427 3428 poller->device = device; 3429 poller->group = rgroup; 3430 3431 RB_INIT(&poller->qpairs); 3432 STAILQ_INIT(&poller->qpairs_pending_send); 3433 STAILQ_INIT(&poller->qpairs_pending_recv); 3434 3435 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3436 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3437 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) { 3438 SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n", 3439 rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr); 3440 } 3441 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr); 3442 3443 device->num_srq++; 3444 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 3445 srq_init_attr.pd = device->pd; 3446 srq_init_attr.stats = &poller->stat.qp_stats.recv; 3447 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 3448 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3449 poller->srq = spdk_rdma_srq_create(&srq_init_attr); 3450 if (!poller->srq) { 3451 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3452 nvmf_rdma_poll_group_destroy(&rgroup->group); 3453 return NULL; 3454 } 3455 3456 opts.qp = poller->srq; 3457 opts.pd = device->pd; 3458 opts.qpair = NULL; 3459 opts.shared = true; 3460 opts.max_queue_depth = poller->max_srq_depth; 3461 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3462 3463 poller->resources = nvmf_rdma_resources_create(&opts); 3464 if (!poller->resources) { 3465 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3466 nvmf_rdma_poll_group_destroy(&rgroup->group); 3467 return NULL; 3468 } 3469 } 3470 3471 /* 3472 * When using an srq, we can limit the completion queue at startup. 3473 * The following formula represents the calculation: 3474 * num_cqe = num_recv + num_data_wr + num_send_wr. 3475 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3476 */ 3477 if (poller->srq) { 3478 num_cqe = poller->max_srq_depth * 3; 3479 } else { 3480 num_cqe = rtransport->rdma_opts.num_cqe; 3481 } 3482 3483 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3484 if (!poller->cq) { 3485 SPDK_ERRLOG("Unable to create completion queue\n"); 3486 nvmf_rdma_poll_group_destroy(&rgroup->group); 3487 return NULL; 3488 } 3489 poller->num_cqe = num_cqe; 3490 } 3491 3492 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3493 if (rtransport->conn_sched.next_admin_pg == NULL) { 3494 rtransport->conn_sched.next_admin_pg = rgroup; 3495 rtransport->conn_sched.next_io_pg = rgroup; 3496 } 3497 3498 return &rgroup->group; 3499 } 3500 3501 static struct spdk_nvmf_transport_poll_group * 3502 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3503 { 3504 struct spdk_nvmf_rdma_transport *rtransport; 3505 struct spdk_nvmf_rdma_poll_group **pg; 3506 struct spdk_nvmf_transport_poll_group *result; 3507 3508 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3509 3510 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3511 return NULL; 3512 } 3513 3514 if (qpair->qid == 0) { 3515 pg = &rtransport->conn_sched.next_admin_pg; 3516 } else { 3517 struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current; 3518 uint32_t min_value; 3519 3520 pg = &rtransport->conn_sched.next_io_pg; 3521 pg_min = *pg; 3522 pg_start = *pg; 3523 pg_current = *pg; 3524 min_value = (*pg)->group.group->stat.current_io_qpairs; 3525 3526 while (pg_current->group.group->stat.current_io_qpairs) { 3527 pg_current = TAILQ_NEXT(pg_current, link); 3528 if (pg_current == NULL) { 3529 pg_current = TAILQ_FIRST(&rtransport->poll_groups); 3530 } 3531 3532 if (pg_current->group.group->stat.current_io_qpairs < min_value) { 3533 min_value = pg_current->group.group->stat.current_io_qpairs; 3534 pg_min = pg_current; 3535 } 3536 3537 if (pg_current == pg_start) { 3538 break; 3539 } 3540 } 3541 *pg = pg_min; 3542 } 3543 3544 assert(*pg != NULL); 3545 3546 result = &(*pg)->group; 3547 3548 *pg = TAILQ_NEXT(*pg, link); 3549 if (*pg == NULL) { 3550 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3551 } 3552 3553 return result; 3554 } 3555 3556 static void 3557 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3558 { 3559 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3560 struct spdk_nvmf_rdma_poller *poller, *tmp; 3561 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3562 struct spdk_nvmf_rdma_transport *rtransport; 3563 3564 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3565 if (!rgroup) { 3566 return; 3567 } 3568 3569 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3570 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3571 3572 RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) { 3573 nvmf_rdma_qpair_destroy(qpair); 3574 } 3575 3576 if (poller->srq) { 3577 if (poller->resources) { 3578 nvmf_rdma_resources_destroy(poller->resources); 3579 } 3580 spdk_rdma_srq_destroy(poller->srq); 3581 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 3582 } 3583 3584 if (poller->cq) { 3585 ibv_destroy_cq(poller->cq); 3586 } 3587 3588 free(poller); 3589 } 3590 3591 if (rgroup->group.transport == NULL) { 3592 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3593 * calls this function directly in a failure path. */ 3594 free(rgroup); 3595 return; 3596 } 3597 3598 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3599 3600 next_rgroup = TAILQ_NEXT(rgroup, link); 3601 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3602 if (next_rgroup == NULL) { 3603 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3604 } 3605 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3606 rtransport->conn_sched.next_admin_pg = next_rgroup; 3607 } 3608 if (rtransport->conn_sched.next_io_pg == rgroup) { 3609 rtransport->conn_sched.next_io_pg = next_rgroup; 3610 } 3611 3612 free(rgroup); 3613 } 3614 3615 static void 3616 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3617 { 3618 if (rqpair->cm_id != NULL) { 3619 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3620 } 3621 } 3622 3623 static int 3624 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3625 struct spdk_nvmf_qpair *qpair) 3626 { 3627 struct spdk_nvmf_rdma_poll_group *rgroup; 3628 struct spdk_nvmf_rdma_qpair *rqpair; 3629 struct spdk_nvmf_rdma_device *device; 3630 struct spdk_nvmf_rdma_poller *poller; 3631 int rc; 3632 3633 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3634 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3635 3636 device = rqpair->device; 3637 3638 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3639 if (poller->device == device) { 3640 break; 3641 } 3642 } 3643 3644 if (!poller) { 3645 SPDK_ERRLOG("No poller found for device.\n"); 3646 return -1; 3647 } 3648 3649 rqpair->poller = poller; 3650 rqpair->srq = rqpair->poller->srq; 3651 3652 rc = nvmf_rdma_qpair_initialize(qpair); 3653 if (rc < 0) { 3654 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3655 rqpair->poller = NULL; 3656 rqpair->srq = NULL; 3657 return -1; 3658 } 3659 3660 RB_INSERT(qpairs_tree, &poller->qpairs, rqpair); 3661 3662 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3663 if (rc) { 3664 /* Try to reject, but we probably can't */ 3665 nvmf_rdma_qpair_reject_connection(rqpair); 3666 return -1; 3667 } 3668 3669 nvmf_rdma_update_ibv_state(rqpair); 3670 3671 return 0; 3672 } 3673 3674 static int 3675 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3676 struct spdk_nvmf_qpair *qpair) 3677 { 3678 struct spdk_nvmf_rdma_qpair *rqpair; 3679 3680 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3681 assert(group->transport->tgt != NULL); 3682 3683 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3684 3685 if (!rqpair->destruct_channel) { 3686 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3687 return 0; 3688 } 3689 3690 /* Sanity check that we get io_channel on the correct thread */ 3691 if (qpair->group) { 3692 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3693 } 3694 3695 return 0; 3696 } 3697 3698 static int 3699 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3700 { 3701 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3702 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3703 struct spdk_nvmf_rdma_transport, transport); 3704 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3705 struct spdk_nvmf_rdma_qpair, qpair); 3706 3707 /* 3708 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3709 * needs to be returned to the shared receive queue or the poll group will eventually be 3710 * starved of RECV structures. 3711 */ 3712 if (rqpair->srq && rdma_req->recv) { 3713 int rc; 3714 struct ibv_recv_wr *bad_recv_wr; 3715 3716 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 3717 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 3718 if (rc) { 3719 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3720 } 3721 } 3722 3723 _nvmf_rdma_request_free(rdma_req, rtransport); 3724 return 0; 3725 } 3726 3727 static int 3728 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3729 { 3730 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3731 struct spdk_nvmf_rdma_transport, transport); 3732 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3733 struct spdk_nvmf_rdma_request, req); 3734 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3735 struct spdk_nvmf_rdma_qpair, qpair); 3736 3737 if (rqpair->ibv_state != IBV_QPS_ERR) { 3738 /* The connection is alive, so process the request as normal */ 3739 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3740 } else { 3741 /* The connection is dead. Move the request directly to the completed state. */ 3742 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3743 } 3744 3745 nvmf_rdma_request_process(rtransport, rdma_req); 3746 3747 return 0; 3748 } 3749 3750 static void 3751 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 3752 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3753 { 3754 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3755 3756 rqpair->to_close = true; 3757 3758 /* This happens only when the qpair is disconnected before 3759 * it is added to the poll group. Since there is no poll group, 3760 * the RDMA qp has not been initialized yet and the RDMA CM 3761 * event has not yet been acknowledged, so we need to reject it. 3762 */ 3763 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3764 nvmf_rdma_qpair_reject_connection(rqpair); 3765 nvmf_rdma_qpair_destroy(rqpair); 3766 return; 3767 } 3768 3769 if (rqpair->rdma_qp) { 3770 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3771 } 3772 3773 nvmf_rdma_destroy_drained_qpair(rqpair); 3774 3775 if (cb_fn) { 3776 cb_fn(cb_arg); 3777 } 3778 } 3779 3780 static struct spdk_nvmf_rdma_qpair * 3781 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3782 { 3783 struct spdk_nvmf_rdma_qpair find; 3784 3785 find.qp_num = wc->qp_num; 3786 3787 return RB_FIND(qpairs_tree, &rpoller->qpairs, &find); 3788 } 3789 3790 #ifdef DEBUG 3791 static int 3792 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3793 { 3794 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3795 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3796 } 3797 #endif 3798 3799 static void 3800 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3801 int rc) 3802 { 3803 struct spdk_nvmf_rdma_recv *rdma_recv; 3804 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3805 3806 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3807 while (bad_recv_wr != NULL) { 3808 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3809 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3810 3811 rdma_recv->qpair->current_recv_depth++; 3812 bad_recv_wr = bad_recv_wr->next; 3813 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3814 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3815 } 3816 } 3817 3818 static void 3819 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3820 { 3821 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3822 while (bad_recv_wr != NULL) { 3823 bad_recv_wr = bad_recv_wr->next; 3824 rqpair->current_recv_depth++; 3825 } 3826 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3827 } 3828 3829 static void 3830 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3831 struct spdk_nvmf_rdma_poller *rpoller) 3832 { 3833 struct spdk_nvmf_rdma_qpair *rqpair; 3834 struct ibv_recv_wr *bad_recv_wr; 3835 int rc; 3836 3837 if (rpoller->srq) { 3838 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 3839 if (rc) { 3840 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3841 } 3842 } else { 3843 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3844 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3845 rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 3846 if (rc) { 3847 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3848 } 3849 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3850 } 3851 } 3852 } 3853 3854 static void 3855 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3856 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3857 { 3858 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3859 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3860 3861 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3862 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3863 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3864 assert(rqpair->current_send_depth > 0); 3865 rqpair->current_send_depth--; 3866 switch (bad_rdma_wr->type) { 3867 case RDMA_WR_TYPE_DATA: 3868 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3869 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3870 assert(rqpair->current_read_depth > 0); 3871 rqpair->current_read_depth--; 3872 } 3873 break; 3874 case RDMA_WR_TYPE_SEND: 3875 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3876 break; 3877 default: 3878 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3879 prev_rdma_req = cur_rdma_req; 3880 continue; 3881 } 3882 3883 if (prev_rdma_req == cur_rdma_req) { 3884 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3885 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3886 continue; 3887 } 3888 3889 switch (cur_rdma_req->state) { 3890 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3891 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3892 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3893 break; 3894 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3895 case RDMA_REQUEST_STATE_COMPLETING: 3896 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3897 break; 3898 default: 3899 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3900 cur_rdma_req->state, rqpair); 3901 continue; 3902 } 3903 3904 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3905 prev_rdma_req = cur_rdma_req; 3906 } 3907 3908 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3909 /* Disconnect the connection. */ 3910 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3911 } 3912 3913 } 3914 3915 static void 3916 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3917 struct spdk_nvmf_rdma_poller *rpoller) 3918 { 3919 struct spdk_nvmf_rdma_qpair *rqpair; 3920 struct ibv_send_wr *bad_wr = NULL; 3921 int rc; 3922 3923 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3924 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3925 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3926 3927 /* bad wr always points to the first wr that failed. */ 3928 if (rc) { 3929 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3930 } 3931 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3932 } 3933 } 3934 3935 static const char * 3936 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 3937 { 3938 switch (wr_type) { 3939 case RDMA_WR_TYPE_RECV: 3940 return "RECV"; 3941 case RDMA_WR_TYPE_SEND: 3942 return "SEND"; 3943 case RDMA_WR_TYPE_DATA: 3944 return "DATA"; 3945 default: 3946 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 3947 SPDK_UNREACHABLE(); 3948 } 3949 } 3950 3951 static inline void 3952 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 3953 { 3954 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 3955 3956 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 3957 /* If qpair is in ERR state, we will receive completions for all posted and not completed 3958 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 3959 SPDK_DEBUGLOG(rdma, 3960 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3961 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3962 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3963 } else { 3964 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3965 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3966 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3967 } 3968 } 3969 3970 static int 3971 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3972 struct spdk_nvmf_rdma_poller *rpoller) 3973 { 3974 struct ibv_wc wc[32]; 3975 struct spdk_nvmf_rdma_wr *rdma_wr; 3976 struct spdk_nvmf_rdma_request *rdma_req; 3977 struct spdk_nvmf_rdma_recv *rdma_recv; 3978 struct spdk_nvmf_rdma_qpair *rqpair; 3979 int reaped, i; 3980 int count = 0; 3981 bool error = false; 3982 uint64_t poll_tsc = spdk_get_ticks(); 3983 3984 /* Poll for completing operations. */ 3985 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3986 if (reaped < 0) { 3987 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3988 errno, spdk_strerror(errno)); 3989 return -1; 3990 } else if (reaped == 0) { 3991 rpoller->stat.idle_polls++; 3992 } 3993 3994 rpoller->stat.polls++; 3995 rpoller->stat.completions += reaped; 3996 3997 for (i = 0; i < reaped; i++) { 3998 3999 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 4000 4001 switch (rdma_wr->type) { 4002 case RDMA_WR_TYPE_SEND: 4003 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 4004 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4005 4006 if (!wc[i].status) { 4007 count++; 4008 assert(wc[i].opcode == IBV_WC_SEND); 4009 assert(nvmf_rdma_req_is_completing(rdma_req)); 4010 } 4011 4012 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4013 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 4014 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 4015 rdma_req->num_outstanding_data_wr = 0; 4016 4017 nvmf_rdma_request_process(rtransport, rdma_req); 4018 break; 4019 case RDMA_WR_TYPE_RECV: 4020 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 4021 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4022 if (rpoller->srq != NULL) { 4023 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 4024 /* It is possible that there are still some completions for destroyed QP 4025 * associated with SRQ. We just ignore these late completions and re-post 4026 * receive WRs back to SRQ. 4027 */ 4028 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 4029 struct ibv_recv_wr *bad_wr; 4030 int rc; 4031 4032 rdma_recv->wr.next = NULL; 4033 spdk_rdma_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 4034 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 4035 if (rc) { 4036 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 4037 } 4038 continue; 4039 } 4040 } 4041 rqpair = rdma_recv->qpair; 4042 4043 assert(rqpair != NULL); 4044 if (!wc[i].status) { 4045 assert(wc[i].opcode == IBV_WC_RECV); 4046 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 4047 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4048 break; 4049 } 4050 } 4051 4052 rdma_recv->wr.next = NULL; 4053 rqpair->current_recv_depth++; 4054 rdma_recv->receive_tsc = poll_tsc; 4055 rpoller->stat.requests++; 4056 STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link); 4057 break; 4058 case RDMA_WR_TYPE_DATA: 4059 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 4060 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4061 4062 assert(rdma_req->num_outstanding_data_wr > 0); 4063 4064 rqpair->current_send_depth--; 4065 rdma_req->num_outstanding_data_wr--; 4066 if (!wc[i].status) { 4067 assert(wc[i].opcode == IBV_WC_RDMA_READ); 4068 rqpair->current_read_depth--; 4069 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 4070 if (rdma_req->num_outstanding_data_wr == 0) { 4071 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 4072 nvmf_rdma_request_process(rtransport, rdma_req); 4073 } 4074 } else { 4075 /* If the data transfer fails still force the queue into the error state, 4076 * if we were performing an RDMA_READ, we need to force the request into a 4077 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 4078 * case, we should wait for the SEND to complete. */ 4079 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 4080 rqpair->current_read_depth--; 4081 if (rdma_req->num_outstanding_data_wr == 0) { 4082 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4083 } 4084 } 4085 } 4086 break; 4087 default: 4088 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 4089 continue; 4090 } 4091 4092 /* Handle error conditions */ 4093 if (wc[i].status) { 4094 nvmf_rdma_update_ibv_state(rqpair); 4095 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 4096 4097 error = true; 4098 4099 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 4100 /* Disconnect the connection. */ 4101 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4102 } else { 4103 nvmf_rdma_destroy_drained_qpair(rqpair); 4104 } 4105 continue; 4106 } 4107 4108 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 4109 4110 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 4111 nvmf_rdma_destroy_drained_qpair(rqpair); 4112 } 4113 } 4114 4115 if (error == true) { 4116 return -1; 4117 } 4118 4119 /* submit outstanding work requests. */ 4120 _poller_submit_recvs(rtransport, rpoller); 4121 _poller_submit_sends(rtransport, rpoller); 4122 4123 return count; 4124 } 4125 4126 static int 4127 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 4128 { 4129 struct spdk_nvmf_rdma_transport *rtransport; 4130 struct spdk_nvmf_rdma_poll_group *rgroup; 4131 struct spdk_nvmf_rdma_poller *rpoller; 4132 int count, rc; 4133 4134 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 4135 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4136 4137 count = 0; 4138 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4139 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4140 if (rc < 0) { 4141 return rc; 4142 } 4143 count += rc; 4144 } 4145 4146 return count; 4147 } 4148 4149 static int 4150 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4151 struct spdk_nvme_transport_id *trid, 4152 bool peer) 4153 { 4154 struct sockaddr *saddr; 4155 uint16_t port; 4156 4157 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4158 4159 if (peer) { 4160 saddr = rdma_get_peer_addr(id); 4161 } else { 4162 saddr = rdma_get_local_addr(id); 4163 } 4164 switch (saddr->sa_family) { 4165 case AF_INET: { 4166 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4167 4168 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4169 inet_ntop(AF_INET, &saddr_in->sin_addr, 4170 trid->traddr, sizeof(trid->traddr)); 4171 if (peer) { 4172 port = ntohs(rdma_get_dst_port(id)); 4173 } else { 4174 port = ntohs(rdma_get_src_port(id)); 4175 } 4176 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4177 break; 4178 } 4179 case AF_INET6: { 4180 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4181 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4182 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4183 trid->traddr, sizeof(trid->traddr)); 4184 if (peer) { 4185 port = ntohs(rdma_get_dst_port(id)); 4186 } else { 4187 port = ntohs(rdma_get_src_port(id)); 4188 } 4189 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4190 break; 4191 } 4192 default: 4193 return -1; 4194 4195 } 4196 4197 return 0; 4198 } 4199 4200 static int 4201 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4202 struct spdk_nvme_transport_id *trid) 4203 { 4204 struct spdk_nvmf_rdma_qpair *rqpair; 4205 4206 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4207 4208 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4209 } 4210 4211 static int 4212 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4213 struct spdk_nvme_transport_id *trid) 4214 { 4215 struct spdk_nvmf_rdma_qpair *rqpair; 4216 4217 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4218 4219 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4220 } 4221 4222 static int 4223 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4224 struct spdk_nvme_transport_id *trid) 4225 { 4226 struct spdk_nvmf_rdma_qpair *rqpair; 4227 4228 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4229 4230 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4231 } 4232 4233 void 4234 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4235 { 4236 g_nvmf_hooks = *hooks; 4237 } 4238 4239 static void 4240 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4241 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4242 { 4243 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4244 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4245 4246 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4247 4248 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4249 } 4250 4251 static int 4252 _nvmf_rdma_qpair_abort_request(void *ctx) 4253 { 4254 struct spdk_nvmf_request *req = ctx; 4255 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4256 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4257 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4258 struct spdk_nvmf_rdma_qpair, qpair); 4259 int rc; 4260 4261 spdk_poller_unregister(&req->poller); 4262 4263 switch (rdma_req_to_abort->state) { 4264 case RDMA_REQUEST_STATE_EXECUTING: 4265 rc = nvmf_ctrlr_abort_request(req); 4266 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4267 return SPDK_POLLER_BUSY; 4268 } 4269 break; 4270 4271 case RDMA_REQUEST_STATE_NEED_BUFFER: 4272 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4273 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4274 4275 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4276 break; 4277 4278 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4279 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4280 spdk_nvmf_rdma_request, state_link); 4281 4282 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4283 break; 4284 4285 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4286 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4287 spdk_nvmf_rdma_request, state_link); 4288 4289 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4290 break; 4291 4292 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4293 if (spdk_get_ticks() < req->timeout_tsc) { 4294 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4295 return SPDK_POLLER_BUSY; 4296 } 4297 break; 4298 4299 default: 4300 break; 4301 } 4302 4303 spdk_nvmf_request_complete(req); 4304 return SPDK_POLLER_BUSY; 4305 } 4306 4307 static void 4308 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4309 struct spdk_nvmf_request *req) 4310 { 4311 struct spdk_nvmf_rdma_qpair *rqpair; 4312 struct spdk_nvmf_rdma_transport *rtransport; 4313 struct spdk_nvmf_transport *transport; 4314 uint16_t cid; 4315 uint32_t i, max_req_count; 4316 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 4317 4318 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4319 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4320 transport = &rtransport->transport; 4321 4322 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4323 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 4324 4325 for (i = 0; i < max_req_count; i++) { 4326 rdma_req = &rqpair->resources->reqs[i]; 4327 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 4328 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 4329 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 4330 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 4331 rdma_req->req.qpair == qpair) { 4332 rdma_req_to_abort = rdma_req; 4333 break; 4334 } 4335 } 4336 4337 if (rdma_req_to_abort == NULL) { 4338 spdk_nvmf_request_complete(req); 4339 return; 4340 } 4341 4342 req->req_to_abort = &rdma_req_to_abort->req; 4343 req->timeout_tsc = spdk_get_ticks() + 4344 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4345 req->poller = NULL; 4346 4347 _nvmf_rdma_qpair_abort_request(req); 4348 } 4349 4350 static void 4351 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 4352 struct spdk_json_write_ctx *w) 4353 { 4354 struct spdk_nvmf_rdma_poll_group *rgroup; 4355 struct spdk_nvmf_rdma_poller *rpoller; 4356 4357 assert(w != NULL); 4358 4359 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4360 4361 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 4362 4363 spdk_json_write_named_array_begin(w, "devices"); 4364 4365 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4366 spdk_json_write_object_begin(w); 4367 spdk_json_write_named_string(w, "name", 4368 ibv_get_device_name(rpoller->device->context->device)); 4369 spdk_json_write_named_uint64(w, "polls", 4370 rpoller->stat.polls); 4371 spdk_json_write_named_uint64(w, "idle_polls", 4372 rpoller->stat.idle_polls); 4373 spdk_json_write_named_uint64(w, "completions", 4374 rpoller->stat.completions); 4375 spdk_json_write_named_uint64(w, "requests", 4376 rpoller->stat.requests); 4377 spdk_json_write_named_uint64(w, "request_latency", 4378 rpoller->stat.request_latency); 4379 spdk_json_write_named_uint64(w, "pending_free_request", 4380 rpoller->stat.pending_free_request); 4381 spdk_json_write_named_uint64(w, "pending_rdma_read", 4382 rpoller->stat.pending_rdma_read); 4383 spdk_json_write_named_uint64(w, "pending_rdma_write", 4384 rpoller->stat.pending_rdma_write); 4385 spdk_json_write_named_uint64(w, "total_send_wrs", 4386 rpoller->stat.qp_stats.send.num_submitted_wrs); 4387 spdk_json_write_named_uint64(w, "send_doorbell_updates", 4388 rpoller->stat.qp_stats.send.doorbell_updates); 4389 spdk_json_write_named_uint64(w, "total_recv_wrs", 4390 rpoller->stat.qp_stats.recv.num_submitted_wrs); 4391 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 4392 rpoller->stat.qp_stats.recv.doorbell_updates); 4393 spdk_json_write_object_end(w); 4394 } 4395 4396 spdk_json_write_array_end(w); 4397 } 4398 4399 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4400 .name = "RDMA", 4401 .type = SPDK_NVME_TRANSPORT_RDMA, 4402 .opts_init = nvmf_rdma_opts_init, 4403 .create = nvmf_rdma_create, 4404 .dump_opts = nvmf_rdma_dump_opts, 4405 .destroy = nvmf_rdma_destroy, 4406 4407 .listen = nvmf_rdma_listen, 4408 .stop_listen = nvmf_rdma_stop_listen, 4409 .cdata_init = nvmf_rdma_cdata_init, 4410 4411 .listener_discover = nvmf_rdma_discover, 4412 4413 .poll_group_create = nvmf_rdma_poll_group_create, 4414 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4415 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4416 .poll_group_add = nvmf_rdma_poll_group_add, 4417 .poll_group_remove = nvmf_rdma_poll_group_remove, 4418 .poll_group_poll = nvmf_rdma_poll_group_poll, 4419 4420 .req_free = nvmf_rdma_request_free, 4421 .req_complete = nvmf_rdma_request_complete, 4422 4423 .qpair_fini = nvmf_rdma_close_qpair, 4424 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4425 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4426 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4427 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4428 4429 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 4430 }; 4431 4432 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4433 SPDK_LOG_REGISTER_COMPONENT(rdma) 4434