1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (C) 2016 Intel Corporation. All rights reserved. 3 * Copyright (c) 2019-2021 Mellanox Technologies LTD. All rights reserved. 4 * Copyright (c) 2021-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. 5 */ 6 7 #include "spdk/stdinc.h" 8 9 #include "spdk/config.h" 10 #include "spdk/thread.h" 11 #include "spdk/likely.h" 12 #include "spdk/nvmf_transport.h" 13 #include "spdk/string.h" 14 #include "spdk/trace.h" 15 #include "spdk/tree.h" 16 #include "spdk/util.h" 17 18 #include "spdk_internal/assert.h" 19 #include "spdk/log.h" 20 #include "spdk_internal/rdma.h" 21 22 #include "nvmf_internal.h" 23 #include "transport.h" 24 25 #include "spdk_internal/trace_defs.h" 26 27 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 28 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 29 30 /* 31 RDMA Connection Resource Defaults 32 */ 33 #define NVMF_DEFAULT_MSDBD 16 34 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 35 #define NVMF_DEFAULT_RSP_SGE 1 36 #define NVMF_DEFAULT_RX_SGE 2 37 38 SPDK_STATIC_ASSERT(NVMF_DEFAULT_MSDBD <= SPDK_NVMF_MAX_SGL_ENTRIES, 39 "MSDBD must not exceed SPDK_NVMF_MAX_SGL_ENTRIES"); 40 41 /* The RDMA completion queue size */ 42 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 43 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 44 45 static int g_spdk_nvmf_ibv_query_mask = 46 IBV_QP_STATE | 47 IBV_QP_PKEY_INDEX | 48 IBV_QP_PORT | 49 IBV_QP_ACCESS_FLAGS | 50 IBV_QP_AV | 51 IBV_QP_PATH_MTU | 52 IBV_QP_DEST_QPN | 53 IBV_QP_RQ_PSN | 54 IBV_QP_MAX_DEST_RD_ATOMIC | 55 IBV_QP_MIN_RNR_TIMER | 56 IBV_QP_SQ_PSN | 57 IBV_QP_TIMEOUT | 58 IBV_QP_RETRY_CNT | 59 IBV_QP_RNR_RETRY | 60 IBV_QP_MAX_QP_RD_ATOMIC; 61 62 enum spdk_nvmf_rdma_request_state { 63 /* The request is not currently in use */ 64 RDMA_REQUEST_STATE_FREE = 0, 65 66 /* Initial state when request first received */ 67 RDMA_REQUEST_STATE_NEW, 68 69 /* The request is queued until a data buffer is available. */ 70 RDMA_REQUEST_STATE_NEED_BUFFER, 71 72 /* The request is waiting on RDMA queue depth availability 73 * to transfer data from the host to the controller. 74 */ 75 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 76 77 /* The request is currently transferring data from the host to the controller. */ 78 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 79 80 /* The request is ready to execute at the block device */ 81 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 82 83 /* The request is currently executing at the block device */ 84 RDMA_REQUEST_STATE_EXECUTING, 85 86 /* The request finished executing at the block device */ 87 RDMA_REQUEST_STATE_EXECUTED, 88 89 /* The request is waiting on RDMA queue depth availability 90 * to transfer data from the controller to the host. 91 */ 92 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 93 94 /* The request is ready to send a completion */ 95 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 96 97 /* The request is currently transferring data from the controller to the host. */ 98 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 99 100 /* The request currently has an outstanding completion without an 101 * associated data transfer. 102 */ 103 RDMA_REQUEST_STATE_COMPLETING, 104 105 /* The request completed and can be marked free. */ 106 RDMA_REQUEST_STATE_COMPLETED, 107 108 /* Terminator */ 109 RDMA_REQUEST_NUM_STATES, 110 }; 111 112 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 113 { 114 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 115 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 116 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 117 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 118 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 119 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 120 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 121 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 122 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 123 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 124 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 125 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 126 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 127 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 128 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 129 spdk_trace_register_description("RDMA_REQ_TX_H2C", 130 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 131 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 132 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 133 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 134 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 135 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 136 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 137 spdk_trace_register_description("RDMA_REQ_EXECUTING", 138 TRACE_RDMA_REQUEST_STATE_EXECUTING, 139 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 140 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 141 spdk_trace_register_description("RDMA_REQ_EXECUTED", 142 TRACE_RDMA_REQUEST_STATE_EXECUTED, 143 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 144 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 145 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 146 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 147 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 148 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 149 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 150 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 151 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 152 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 153 spdk_trace_register_description("RDMA_REQ_COMPLETING", 154 TRACE_RDMA_REQUEST_STATE_COMPLETING, 155 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 156 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 157 spdk_trace_register_description("RDMA_REQ_COMPLETED", 158 TRACE_RDMA_REQUEST_STATE_COMPLETED, 159 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 160 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 161 162 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 163 OWNER_NONE, OBJECT_NONE, 0, 164 SPDK_TRACE_ARG_TYPE_INT, ""); 165 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 166 OWNER_NONE, OBJECT_NONE, 0, 167 SPDK_TRACE_ARG_TYPE_INT, "type"); 168 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 169 OWNER_NONE, OBJECT_NONE, 0, 170 SPDK_TRACE_ARG_TYPE_INT, "type"); 171 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 172 OWNER_NONE, OBJECT_NONE, 0, 173 SPDK_TRACE_ARG_TYPE_PTR, "state"); 174 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 175 OWNER_NONE, OBJECT_NONE, 0, 176 SPDK_TRACE_ARG_TYPE_INT, ""); 177 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 178 OWNER_NONE, OBJECT_NONE, 0, 179 SPDK_TRACE_ARG_TYPE_INT, ""); 180 } 181 182 enum spdk_nvmf_rdma_wr_type { 183 RDMA_WR_TYPE_RECV, 184 RDMA_WR_TYPE_SEND, 185 RDMA_WR_TYPE_DATA, 186 }; 187 188 struct spdk_nvmf_rdma_wr { 189 /* Uses enum spdk_nvmf_rdma_wr_type */ 190 uint8_t type; 191 }; 192 193 /* This structure holds commands as they are received off the wire. 194 * It must be dynamically paired with a full request object 195 * (spdk_nvmf_rdma_request) to service a request. It is separate 196 * from the request because RDMA does not appear to order 197 * completions, so occasionally we'll get a new incoming 198 * command when there aren't any free request objects. 199 */ 200 struct spdk_nvmf_rdma_recv { 201 struct ibv_recv_wr wr; 202 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 203 204 struct spdk_nvmf_rdma_qpair *qpair; 205 206 /* In-capsule data buffer */ 207 uint8_t *buf; 208 209 struct spdk_nvmf_rdma_wr rdma_wr; 210 uint64_t receive_tsc; 211 212 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 213 }; 214 215 struct spdk_nvmf_rdma_request_data { 216 struct ibv_send_wr wr; 217 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 218 }; 219 220 struct spdk_nvmf_rdma_request { 221 struct spdk_nvmf_request req; 222 223 bool fused_failed; 224 225 struct spdk_nvmf_rdma_wr data_wr; 226 struct spdk_nvmf_rdma_wr rsp_wr; 227 228 /* Uses enum spdk_nvmf_rdma_request_state */ 229 uint8_t state; 230 231 /* Data offset in req.iov */ 232 uint32_t offset; 233 234 struct spdk_nvmf_rdma_recv *recv; 235 236 struct { 237 struct ibv_send_wr wr; 238 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 239 } rsp; 240 241 uint16_t iovpos; 242 uint16_t num_outstanding_data_wr; 243 /* Used to split Write IO with multi SGL payload */ 244 uint16_t num_remaining_data_wr; 245 uint64_t receive_tsc; 246 struct spdk_nvmf_rdma_request *fused_pair; 247 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 248 struct ibv_send_wr *remaining_tranfer_in_wrs; 249 struct ibv_send_wr *transfer_wr; 250 struct spdk_nvmf_rdma_request_data data; 251 }; 252 253 struct spdk_nvmf_rdma_resource_opts { 254 struct spdk_nvmf_rdma_qpair *qpair; 255 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 256 void *qp; 257 struct spdk_rdma_mem_map *map; 258 uint32_t max_queue_depth; 259 uint32_t in_capsule_data_size; 260 bool shared; 261 }; 262 263 struct spdk_nvmf_rdma_resources { 264 /* Array of size "max_queue_depth" containing RDMA requests. */ 265 struct spdk_nvmf_rdma_request *reqs; 266 267 /* Array of size "max_queue_depth" containing RDMA recvs. */ 268 struct spdk_nvmf_rdma_recv *recvs; 269 270 /* Array of size "max_queue_depth" containing 64 byte capsules 271 * used for receive. 272 */ 273 union nvmf_h2c_msg *cmds; 274 275 /* Array of size "max_queue_depth" containing 16 byte completions 276 * to be sent back to the user. 277 */ 278 union nvmf_c2h_msg *cpls; 279 280 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 281 * buffers to be used for in capsule data. 282 */ 283 void *bufs; 284 285 /* Receives that are waiting for a request object */ 286 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 287 288 /* Queue to track free requests */ 289 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 290 }; 291 292 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 293 294 typedef void (*spdk_poller_destroy_cb)(void *ctx); 295 296 struct spdk_nvmf_rdma_ibv_event_ctx { 297 struct spdk_nvmf_rdma_qpair *rqpair; 298 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 299 /* Link to other ibv events associated with this qpair */ 300 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 301 }; 302 303 struct spdk_nvmf_rdma_qpair { 304 struct spdk_nvmf_qpair qpair; 305 306 struct spdk_nvmf_rdma_device *device; 307 struct spdk_nvmf_rdma_poller *poller; 308 309 struct spdk_rdma_qp *rdma_qp; 310 struct rdma_cm_id *cm_id; 311 struct spdk_rdma_srq *srq; 312 struct rdma_cm_id *listen_id; 313 314 /* Cache the QP number to improve QP search by RB tree. */ 315 uint32_t qp_num; 316 317 /* The maximum number of I/O outstanding on this connection at one time */ 318 uint16_t max_queue_depth; 319 320 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 321 uint16_t max_read_depth; 322 323 /* The maximum number of RDMA SEND operations at one time */ 324 uint32_t max_send_depth; 325 326 /* The current number of outstanding WRs from this qpair's 327 * recv queue. Should not exceed device->attr.max_queue_depth. 328 */ 329 uint16_t current_recv_depth; 330 331 /* The current number of active RDMA READ operations */ 332 uint16_t current_read_depth; 333 334 /* The current number of posted WRs from this qpair's 335 * send queue. Should not exceed max_send_depth. 336 */ 337 uint32_t current_send_depth; 338 339 /* The maximum number of SGEs per WR on the send queue */ 340 uint32_t max_send_sge; 341 342 /* The maximum number of SGEs per WR on the recv queue */ 343 uint32_t max_recv_sge; 344 345 struct spdk_nvmf_rdma_resources *resources; 346 347 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 348 349 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 350 351 /* Number of requests not in the free state */ 352 uint32_t qd; 353 354 RB_ENTRY(spdk_nvmf_rdma_qpair) node; 355 356 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 357 358 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 359 360 /* IBV queue pair attributes: they are used to manage 361 * qp state and recover from errors. 362 */ 363 enum ibv_qp_state ibv_state; 364 365 /* Points to the a request that has fuse bits set to 366 * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting 367 * for the request that has SPDK_NVME_CMD_FUSE_SECOND. 368 */ 369 struct spdk_nvmf_rdma_request *fused_first; 370 371 /* 372 * io_channel which is used to destroy qpair when it is removed from poll group 373 */ 374 struct spdk_io_channel *destruct_channel; 375 376 /* List of ibv async events */ 377 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 378 379 /* Lets us know that we have received the last_wqe event. */ 380 bool last_wqe_reached; 381 382 /* Indicate that nvmf_rdma_close_qpair is called */ 383 bool to_close; 384 }; 385 386 struct spdk_nvmf_rdma_poller_stat { 387 uint64_t completions; 388 uint64_t polls; 389 uint64_t idle_polls; 390 uint64_t requests; 391 uint64_t request_latency; 392 uint64_t pending_free_request; 393 uint64_t pending_rdma_read; 394 uint64_t pending_rdma_write; 395 struct spdk_rdma_qp_stats qp_stats; 396 }; 397 398 struct spdk_nvmf_rdma_poller { 399 struct spdk_nvmf_rdma_device *device; 400 struct spdk_nvmf_rdma_poll_group *group; 401 402 int num_cqe; 403 int required_num_wr; 404 struct ibv_cq *cq; 405 406 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 407 uint16_t max_srq_depth; 408 bool need_destroy; 409 410 /* Shared receive queue */ 411 struct spdk_rdma_srq *srq; 412 413 struct spdk_nvmf_rdma_resources *resources; 414 struct spdk_nvmf_rdma_poller_stat stat; 415 416 spdk_poller_destroy_cb destroy_cb; 417 void *destroy_cb_ctx; 418 419 RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs; 420 421 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 422 423 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 424 425 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 426 }; 427 428 struct spdk_nvmf_rdma_poll_group_stat { 429 uint64_t pending_data_buffer; 430 }; 431 432 struct spdk_nvmf_rdma_poll_group { 433 struct spdk_nvmf_transport_poll_group group; 434 struct spdk_nvmf_rdma_poll_group_stat stat; 435 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 436 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 437 }; 438 439 struct spdk_nvmf_rdma_conn_sched { 440 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 441 struct spdk_nvmf_rdma_poll_group *next_io_pg; 442 }; 443 444 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 445 struct spdk_nvmf_rdma_device { 446 struct ibv_device_attr attr; 447 struct ibv_context *context; 448 449 struct spdk_rdma_mem_map *map; 450 struct ibv_pd *pd; 451 452 int num_srq; 453 bool need_destroy; 454 bool ready_to_destroy; 455 bool is_ready; 456 457 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 458 }; 459 460 struct spdk_nvmf_rdma_port { 461 const struct spdk_nvme_transport_id *trid; 462 struct rdma_cm_id *id; 463 struct spdk_nvmf_rdma_device *device; 464 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 465 }; 466 467 struct rdma_transport_opts { 468 int num_cqe; 469 uint32_t max_srq_depth; 470 bool no_srq; 471 bool no_wr_batching; 472 int acceptor_backlog; 473 }; 474 475 struct spdk_nvmf_rdma_transport { 476 struct spdk_nvmf_transport transport; 477 struct rdma_transport_opts rdma_opts; 478 479 struct spdk_nvmf_rdma_conn_sched conn_sched; 480 481 struct rdma_event_channel *event_channel; 482 483 struct spdk_mempool *data_wr_pool; 484 485 struct spdk_poller *accept_poller; 486 487 /* fields used to poll RDMA/IB events */ 488 nfds_t npoll_fds; 489 struct pollfd *poll_fds; 490 491 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 492 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 493 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 494 495 /* ports that are removed unexpectedly and need retry listen */ 496 TAILQ_HEAD(, spdk_nvmf_rdma_port) retry_ports; 497 }; 498 499 struct poller_manage_ctx { 500 struct spdk_nvmf_rdma_transport *rtransport; 501 struct spdk_nvmf_rdma_poll_group *rgroup; 502 struct spdk_nvmf_rdma_poller *rpoller; 503 struct spdk_nvmf_rdma_device *device; 504 505 struct spdk_thread *thread; 506 volatile int *inflight_op_counter; 507 }; 508 509 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 510 { 511 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe), 512 spdk_json_decode_int32, true 513 }, 514 { 515 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 516 spdk_json_decode_uint32, true 517 }, 518 { 519 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 520 spdk_json_decode_bool, true 521 }, 522 { 523 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 524 spdk_json_decode_bool, true 525 }, 526 { 527 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 528 spdk_json_decode_int32, true 529 }, 530 }; 531 532 static int 533 nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2) 534 { 535 return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num; 536 } 537 538 RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare); 539 540 static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 541 struct spdk_nvmf_rdma_request *rdma_req); 542 543 static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 544 struct spdk_nvmf_rdma_poller *rpoller); 545 546 static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 547 struct spdk_nvmf_rdma_poller *rpoller); 548 549 static void _nvmf_rdma_remove_destroyed_device(void *c); 550 551 static inline int 552 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 553 { 554 switch (state) { 555 case IBV_QPS_RESET: 556 case IBV_QPS_INIT: 557 case IBV_QPS_RTR: 558 case IBV_QPS_RTS: 559 case IBV_QPS_SQD: 560 case IBV_QPS_SQE: 561 case IBV_QPS_ERR: 562 return 0; 563 default: 564 return -1; 565 } 566 } 567 568 static inline enum spdk_nvme_media_error_status_code 569 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 570 enum spdk_nvme_media_error_status_code result; 571 switch (err_type) 572 { 573 case SPDK_DIF_REFTAG_ERROR: 574 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 575 break; 576 case SPDK_DIF_APPTAG_ERROR: 577 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 578 break; 579 case SPDK_DIF_GUARD_ERROR: 580 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 581 break; 582 default: 583 SPDK_UNREACHABLE(); 584 } 585 586 return result; 587 } 588 589 static enum ibv_qp_state 590 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 591 enum ibv_qp_state old_state, new_state; 592 struct ibv_qp_attr qp_attr; 593 struct ibv_qp_init_attr init_attr; 594 int rc; 595 596 old_state = rqpair->ibv_state; 597 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 598 g_spdk_nvmf_ibv_query_mask, &init_attr); 599 600 if (rc) 601 { 602 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 603 return IBV_QPS_ERR + 1; 604 } 605 606 new_state = qp_attr.qp_state; 607 rqpair->ibv_state = new_state; 608 qp_attr.ah_attr.port_num = qp_attr.port_num; 609 610 rc = nvmf_rdma_check_ibv_state(new_state); 611 if (rc) 612 { 613 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 614 /* 615 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 616 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 617 */ 618 return IBV_QPS_ERR + 1; 619 } 620 621 if (old_state != new_state) 622 { 623 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, (uintptr_t)rqpair, new_state); 624 } 625 return new_state; 626 } 627 628 /* 629 * Return data_wrs to pool starting from \b data_wr 630 * Request's own response and data WR are excluded 631 */ 632 static void 633 _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 634 struct ibv_send_wr *data_wr, 635 struct spdk_mempool *pool) 636 { 637 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 638 struct spdk_nvmf_rdma_request_data *nvmf_data; 639 struct ibv_send_wr *next_send_wr; 640 uint64_t req_wrid = (uint64_t)&rdma_req->data_wr; 641 uint32_t num_wrs = 0; 642 643 while (data_wr && data_wr->wr_id == req_wrid) { 644 nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr); 645 memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge); 646 data_wr->num_sge = 0; 647 next_send_wr = data_wr->next; 648 if (data_wr != &rdma_req->data.wr) { 649 data_wr->next = NULL; 650 assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES); 651 work_requests[num_wrs] = nvmf_data; 652 num_wrs++; 653 } 654 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr; 655 } 656 657 if (num_wrs) { 658 spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs); 659 } 660 } 661 662 static void 663 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 664 struct spdk_nvmf_rdma_transport *rtransport) 665 { 666 rdma_req->num_outstanding_data_wr = 0; 667 668 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 669 670 rdma_req->data.wr.next = NULL; 671 rdma_req->rsp.wr.next = NULL; 672 } 673 674 static void 675 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 676 { 677 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 678 if (req->req.cmd) { 679 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 680 } 681 if (req->recv) { 682 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 683 } 684 } 685 686 static void 687 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 688 { 689 int i; 690 691 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 692 for (i = 0; i < rqpair->max_queue_depth; i++) { 693 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 694 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 695 } 696 } 697 } 698 699 static void 700 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 701 { 702 spdk_free(resources->cmds); 703 spdk_free(resources->cpls); 704 spdk_free(resources->bufs); 705 spdk_free(resources->reqs); 706 spdk_free(resources->recvs); 707 free(resources); 708 } 709 710 711 static struct spdk_nvmf_rdma_resources * 712 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 713 { 714 struct spdk_nvmf_rdma_resources *resources; 715 struct spdk_nvmf_rdma_request *rdma_req; 716 struct spdk_nvmf_rdma_recv *rdma_recv; 717 struct spdk_rdma_qp *qp = NULL; 718 struct spdk_rdma_srq *srq = NULL; 719 struct ibv_recv_wr *bad_wr = NULL; 720 struct spdk_rdma_memory_translation translation; 721 uint32_t i; 722 int rc = 0; 723 724 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 725 if (!resources) { 726 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 727 return NULL; 728 } 729 730 resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs), 731 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 732 resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs), 733 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 734 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 735 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 736 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 737 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 738 739 if (opts->in_capsule_data_size > 0) { 740 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 741 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 742 SPDK_MALLOC_DMA); 743 } 744 745 if (!resources->reqs || !resources->recvs || !resources->cmds || 746 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 747 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 748 goto cleanup; 749 } 750 751 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n", 752 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds)); 753 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n", 754 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls)); 755 if (resources->bufs) { 756 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n", 757 resources->bufs, opts->max_queue_depth * 758 opts->in_capsule_data_size); 759 } 760 761 /* Initialize queues */ 762 STAILQ_INIT(&resources->incoming_queue); 763 STAILQ_INIT(&resources->free_queue); 764 765 if (opts->shared) { 766 srq = (struct spdk_rdma_srq *)opts->qp; 767 } else { 768 qp = (struct spdk_rdma_qp *)opts->qp; 769 } 770 771 for (i = 0; i < opts->max_queue_depth; i++) { 772 rdma_recv = &resources->recvs[i]; 773 rdma_recv->qpair = opts->qpair; 774 775 /* Set up memory to receive commands */ 776 if (resources->bufs) { 777 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 778 opts->in_capsule_data_size)); 779 } 780 781 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 782 783 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 784 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 785 rc = spdk_rdma_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]), 786 &translation); 787 if (rc) { 788 goto cleanup; 789 } 790 rdma_recv->sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 791 rdma_recv->wr.num_sge = 1; 792 793 if (rdma_recv->buf) { 794 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 795 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 796 rc = spdk_rdma_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size, &translation); 797 if (rc) { 798 goto cleanup; 799 } 800 rdma_recv->sgl[1].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 801 rdma_recv->wr.num_sge++; 802 } 803 804 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 805 rdma_recv->wr.sg_list = rdma_recv->sgl; 806 if (srq) { 807 spdk_rdma_srq_queue_recv_wrs(srq, &rdma_recv->wr); 808 } else { 809 spdk_rdma_qp_queue_recv_wrs(qp, &rdma_recv->wr); 810 } 811 } 812 813 for (i = 0; i < opts->max_queue_depth; i++) { 814 rdma_req = &resources->reqs[i]; 815 816 if (opts->qpair != NULL) { 817 rdma_req->req.qpair = &opts->qpair->qpair; 818 } else { 819 rdma_req->req.qpair = NULL; 820 } 821 rdma_req->req.cmd = NULL; 822 rdma_req->req.iovcnt = 0; 823 rdma_req->req.stripped_data = NULL; 824 825 /* Set up memory to send responses */ 826 rdma_req->req.rsp = &resources->cpls[i]; 827 828 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 829 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 830 rc = spdk_rdma_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]), 831 &translation); 832 if (rc) { 833 goto cleanup; 834 } 835 rdma_req->rsp.sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 836 837 rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND; 838 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr; 839 rdma_req->rsp.wr.next = NULL; 840 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 841 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 842 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 843 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 844 845 /* Set up memory for data buffers */ 846 rdma_req->data_wr.type = RDMA_WR_TYPE_DATA; 847 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr; 848 rdma_req->data.wr.next = NULL; 849 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 850 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 851 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 852 853 /* Initialize request state to FREE */ 854 rdma_req->state = RDMA_REQUEST_STATE_FREE; 855 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 856 } 857 858 if (srq) { 859 rc = spdk_rdma_srq_flush_recv_wrs(srq, &bad_wr); 860 } else { 861 rc = spdk_rdma_qp_flush_recv_wrs(qp, &bad_wr); 862 } 863 864 if (rc) { 865 goto cleanup; 866 } 867 868 return resources; 869 870 cleanup: 871 nvmf_rdma_resources_destroy(resources); 872 return NULL; 873 } 874 875 static void 876 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 877 { 878 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 879 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 880 ctx->rqpair = NULL; 881 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 882 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 883 } 884 } 885 886 static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller); 887 888 static void 889 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 890 { 891 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 892 struct ibv_recv_wr *bad_recv_wr = NULL; 893 int rc; 894 895 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair); 896 897 if (rqpair->qd != 0) { 898 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 899 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 900 struct spdk_nvmf_rdma_transport, transport); 901 struct spdk_nvmf_rdma_request *req; 902 uint32_t i, max_req_count = 0; 903 904 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 905 906 if (rqpair->srq == NULL) { 907 nvmf_rdma_dump_qpair_contents(rqpair); 908 max_req_count = rqpair->max_queue_depth; 909 } else if (rqpair->poller && rqpair->resources) { 910 max_req_count = rqpair->poller->max_srq_depth; 911 } 912 913 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 914 for (i = 0; i < max_req_count; i++) { 915 req = &rqpair->resources->reqs[i]; 916 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 917 /* nvmf_rdma_request_process checks qpair ibv and internal state 918 * and completes a request */ 919 nvmf_rdma_request_process(rtransport, req); 920 } 921 } 922 assert(rqpair->qd == 0); 923 } 924 925 if (rqpair->poller) { 926 RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair); 927 928 if (rqpair->srq != NULL && rqpair->resources != NULL) { 929 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 930 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 931 if (rqpair == rdma_recv->qpair) { 932 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 933 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr); 934 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 935 if (rc) { 936 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 937 } 938 } 939 } 940 } 941 } 942 943 if (rqpair->cm_id) { 944 if (rqpair->rdma_qp != NULL) { 945 spdk_rdma_qp_destroy(rqpair->rdma_qp); 946 rqpair->rdma_qp = NULL; 947 } 948 949 if (rqpair->poller != NULL && rqpair->srq == NULL) { 950 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 951 } 952 } 953 954 if (rqpair->srq == NULL && rqpair->resources != NULL) { 955 nvmf_rdma_resources_destroy(rqpair->resources); 956 } 957 958 nvmf_rdma_qpair_clean_ibv_events(rqpair); 959 960 if (rqpair->destruct_channel) { 961 spdk_put_io_channel(rqpair->destruct_channel); 962 rqpair->destruct_channel = NULL; 963 } 964 965 if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) { 966 nvmf_rdma_poller_destroy(rqpair->poller); 967 } 968 969 /* destroy cm_id last so cma device will not be freed before we destroy the cq. */ 970 if (rqpair->cm_id) { 971 rdma_destroy_id(rqpair->cm_id); 972 } 973 974 free(rqpair); 975 } 976 977 static int 978 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 979 { 980 struct spdk_nvmf_rdma_poller *rpoller; 981 int rc, num_cqe, required_num_wr; 982 983 /* Enlarge CQ size dynamically */ 984 rpoller = rqpair->poller; 985 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 986 num_cqe = rpoller->num_cqe; 987 if (num_cqe < required_num_wr) { 988 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 989 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 990 } 991 992 if (rpoller->num_cqe != num_cqe) { 993 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) { 994 SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n" 995 "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe); 996 return -1; 997 } 998 if (required_num_wr > device->attr.max_cqe) { 999 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 1000 required_num_wr, device->attr.max_cqe); 1001 return -1; 1002 } 1003 1004 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 1005 rc = ibv_resize_cq(rpoller->cq, num_cqe); 1006 if (rc) { 1007 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 1008 return -1; 1009 } 1010 1011 rpoller->num_cqe = num_cqe; 1012 } 1013 1014 rpoller->required_num_wr = required_num_wr; 1015 return 0; 1016 } 1017 1018 static int 1019 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 1020 { 1021 struct spdk_nvmf_rdma_qpair *rqpair; 1022 struct spdk_nvmf_rdma_transport *rtransport; 1023 struct spdk_nvmf_transport *transport; 1024 struct spdk_nvmf_rdma_resource_opts opts; 1025 struct spdk_nvmf_rdma_device *device; 1026 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 1027 1028 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1029 device = rqpair->device; 1030 1031 qp_init_attr.qp_context = rqpair; 1032 qp_init_attr.pd = device->pd; 1033 qp_init_attr.send_cq = rqpair->poller->cq; 1034 qp_init_attr.recv_cq = rqpair->poller->cq; 1035 1036 if (rqpair->srq) { 1037 qp_init_attr.srq = rqpair->srq->srq; 1038 } else { 1039 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 1040 } 1041 1042 /* SEND, READ, and WRITE operations */ 1043 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 1044 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 1045 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1046 qp_init_attr.stats = &rqpair->poller->stat.qp_stats; 1047 1048 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1049 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1050 goto error; 1051 } 1052 1053 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 1054 if (!rqpair->rdma_qp) { 1055 goto error; 1056 } 1057 1058 rqpair->qp_num = rqpair->rdma_qp->qp->qp_num; 1059 1060 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1061 qp_init_attr.cap.max_send_wr); 1062 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1063 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1064 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair); 1065 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1066 1067 if (rqpair->poller->srq == NULL) { 1068 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1069 transport = &rtransport->transport; 1070 1071 opts.qp = rqpair->rdma_qp; 1072 opts.map = device->map; 1073 opts.qpair = rqpair; 1074 opts.shared = false; 1075 opts.max_queue_depth = rqpair->max_queue_depth; 1076 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1077 1078 rqpair->resources = nvmf_rdma_resources_create(&opts); 1079 1080 if (!rqpair->resources) { 1081 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1082 rdma_destroy_qp(rqpair->cm_id); 1083 goto error; 1084 } 1085 } else { 1086 rqpair->resources = rqpair->poller->resources; 1087 } 1088 1089 rqpair->current_recv_depth = 0; 1090 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1091 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1092 1093 return 0; 1094 1095 error: 1096 rdma_destroy_id(rqpair->cm_id); 1097 rqpair->cm_id = NULL; 1098 return -1; 1099 } 1100 1101 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1102 /* This function accepts either a single wr or the first wr in a linked list. */ 1103 static void 1104 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1105 { 1106 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1107 struct spdk_nvmf_rdma_transport, transport); 1108 1109 if (rqpair->srq != NULL) { 1110 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, first); 1111 } else { 1112 if (spdk_rdma_qp_queue_recv_wrs(rqpair->rdma_qp, first)) { 1113 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1114 } 1115 } 1116 1117 if (rtransport->rdma_opts.no_wr_batching) { 1118 _poller_submit_recvs(rtransport, rqpair->poller); 1119 } 1120 } 1121 1122 static int 1123 request_transfer_in(struct spdk_nvmf_request *req) 1124 { 1125 struct spdk_nvmf_rdma_request *rdma_req; 1126 struct spdk_nvmf_qpair *qpair; 1127 struct spdk_nvmf_rdma_qpair *rqpair; 1128 struct spdk_nvmf_rdma_transport *rtransport; 1129 1130 qpair = req->qpair; 1131 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1132 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1133 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1134 struct spdk_nvmf_rdma_transport, transport); 1135 1136 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1137 assert(rdma_req != NULL); 1138 1139 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) { 1140 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1141 } 1142 if (rtransport->rdma_opts.no_wr_batching) { 1143 _poller_submit_sends(rtransport, rqpair->poller); 1144 } 1145 1146 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1147 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1148 return 0; 1149 } 1150 1151 static inline int 1152 nvmf_rdma_request_reset_transfer_in(struct spdk_nvmf_rdma_request *rdma_req, 1153 struct spdk_nvmf_rdma_transport *rtransport) 1154 { 1155 /* Put completed WRs back to pool and move transfer_wr pointer */ 1156 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 1157 rdma_req->transfer_wr = rdma_req->remaining_tranfer_in_wrs; 1158 rdma_req->remaining_tranfer_in_wrs = NULL; 1159 rdma_req->num_outstanding_data_wr = rdma_req->num_remaining_data_wr; 1160 rdma_req->num_remaining_data_wr = 0; 1161 1162 return 0; 1163 } 1164 1165 static inline int 1166 request_prepare_transfer_in_part(struct spdk_nvmf_request *req, uint32_t num_reads_available) 1167 { 1168 struct spdk_nvmf_rdma_request *rdma_req; 1169 struct ibv_send_wr *wr; 1170 uint32_t i; 1171 1172 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1173 1174 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1175 assert(rdma_req != NULL); 1176 assert(num_reads_available > 0); 1177 assert(rdma_req->num_outstanding_data_wr > num_reads_available); 1178 wr = rdma_req->transfer_wr; 1179 1180 for (i = 0; i < num_reads_available - 1; i++) { 1181 wr = wr->next; 1182 } 1183 1184 rdma_req->remaining_tranfer_in_wrs = wr->next; 1185 rdma_req->num_remaining_data_wr = rdma_req->num_outstanding_data_wr - num_reads_available; 1186 rdma_req->num_outstanding_data_wr = num_reads_available; 1187 /* Break chain of WRs to send only part. Once this portion completes, we continue sending RDMA_READs */ 1188 wr->next = NULL; 1189 1190 return 0; 1191 } 1192 1193 static int 1194 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1195 { 1196 int num_outstanding_data_wr = 0; 1197 struct spdk_nvmf_rdma_request *rdma_req; 1198 struct spdk_nvmf_qpair *qpair; 1199 struct spdk_nvmf_rdma_qpair *rqpair; 1200 struct spdk_nvme_cpl *rsp; 1201 struct ibv_send_wr *first = NULL; 1202 struct spdk_nvmf_rdma_transport *rtransport; 1203 1204 *data_posted = 0; 1205 qpair = req->qpair; 1206 rsp = &req->rsp->nvme_cpl; 1207 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1208 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1209 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1210 struct spdk_nvmf_rdma_transport, transport); 1211 1212 /* Advance our sq_head pointer */ 1213 if (qpair->sq_head == qpair->sq_head_max) { 1214 qpair->sq_head = 0; 1215 } else { 1216 qpair->sq_head++; 1217 } 1218 rsp->sqhd = qpair->sq_head; 1219 1220 /* queue the capsule for the recv buffer */ 1221 assert(rdma_req->recv != NULL); 1222 1223 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1224 1225 rdma_req->recv = NULL; 1226 assert(rqpair->current_recv_depth > 0); 1227 rqpair->current_recv_depth--; 1228 1229 /* Build the response which consists of optional 1230 * RDMA WRITEs to transfer data, plus an RDMA SEND 1231 * containing the response. 1232 */ 1233 first = &rdma_req->rsp.wr; 1234 1235 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1236 /* On failure, data was not read from the controller. So clear the 1237 * number of outstanding data WRs to zero. 1238 */ 1239 rdma_req->num_outstanding_data_wr = 0; 1240 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1241 first = rdma_req->transfer_wr; 1242 *data_posted = 1; 1243 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1244 } 1245 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1246 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1247 } 1248 if (rtransport->rdma_opts.no_wr_batching) { 1249 _poller_submit_sends(rtransport, rqpair->poller); 1250 } 1251 1252 /* +1 for the rsp wr */ 1253 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1254 1255 return 0; 1256 } 1257 1258 static int 1259 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1260 { 1261 struct spdk_nvmf_rdma_accept_private_data accept_data; 1262 struct rdma_conn_param ctrlr_event_data = {}; 1263 int rc; 1264 1265 accept_data.recfmt = 0; 1266 accept_data.crqsize = rqpair->max_queue_depth; 1267 1268 ctrlr_event_data.private_data = &accept_data; 1269 ctrlr_event_data.private_data_len = sizeof(accept_data); 1270 if (id->ps == RDMA_PS_TCP) { 1271 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1272 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1273 } 1274 1275 /* Configure infinite retries for the initiator side qpair. 1276 * We need to pass this value to the initiator to prevent the 1277 * initiator side NIC from completing SEND requests back to the 1278 * initiator with status rnr_retry_count_exceeded. */ 1279 ctrlr_event_data.rnr_retry_count = 0x7; 1280 1281 /* When qpair is created without use of rdma cm API, an additional 1282 * information must be provided to initiator in the connection response: 1283 * whether qpair is using SRQ and its qp_num 1284 * Fields below are ignored by rdma cm if qpair has been 1285 * created using rdma cm API. */ 1286 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1287 ctrlr_event_data.qp_num = rqpair->qp_num; 1288 1289 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1290 if (rc) { 1291 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1292 } else { 1293 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1294 } 1295 1296 return rc; 1297 } 1298 1299 static void 1300 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1301 { 1302 struct spdk_nvmf_rdma_reject_private_data rej_data; 1303 1304 rej_data.recfmt = 0; 1305 rej_data.sts = error; 1306 1307 rdma_reject(id, &rej_data, sizeof(rej_data)); 1308 } 1309 1310 static int 1311 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1312 { 1313 struct spdk_nvmf_rdma_transport *rtransport; 1314 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1315 struct spdk_nvmf_rdma_port *port; 1316 struct rdma_conn_param *rdma_param = NULL; 1317 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1318 uint16_t max_queue_depth; 1319 uint16_t max_read_depth; 1320 1321 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1322 1323 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1324 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1325 1326 rdma_param = &event->param.conn; 1327 if (rdma_param->private_data == NULL || 1328 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1329 SPDK_ERRLOG("connect request: no private data provided\n"); 1330 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1331 return -1; 1332 } 1333 1334 private_data = rdma_param->private_data; 1335 if (private_data->recfmt != 0) { 1336 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1337 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1338 return -1; 1339 } 1340 1341 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1342 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1343 1344 port = event->listen_id->context; 1345 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1346 event->listen_id, event->listen_id->verbs, port); 1347 1348 /* Figure out the supported queue depth. This is a multi-step process 1349 * that takes into account hardware maximums, host provided values, 1350 * and our target's internal memory limits */ 1351 1352 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1353 1354 /* Start with the maximum queue depth allowed by the target */ 1355 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1356 max_read_depth = rtransport->transport.opts.max_queue_depth; 1357 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1358 rtransport->transport.opts.max_queue_depth); 1359 1360 /* Next check the local NIC's hardware limitations */ 1361 SPDK_DEBUGLOG(rdma, 1362 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1363 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1364 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1365 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1366 1367 /* Next check the remote NIC's hardware limitations */ 1368 SPDK_DEBUGLOG(rdma, 1369 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1370 rdma_param->initiator_depth, rdma_param->responder_resources); 1371 /* from man3 rdma_get_cm_event 1372 * responder_resources - Specifies the number of responder resources that is requested by the recipient. 1373 * The responder_resources field must match the initiator depth specified by the remote node when running 1374 * the rdma_connect and rdma_accept functions. */ 1375 if (rdma_param->responder_resources != 0) { 1376 if (private_data->qid) { 1377 SPDK_DEBUGLOG(rdma, "Host (Initiator) is not allowed to use RDMA operations," 1378 " responder_resources must be 0 but set to %u\n", 1379 rdma_param->responder_resources); 1380 } else { 1381 SPDK_WARNLOG("Host (Initiator) is not allowed to use RDMA operations," 1382 " responder_resources must be 0 but set to %u\n", 1383 rdma_param->responder_resources); 1384 } 1385 } 1386 /* from man3 rdma_get_cm_event 1387 * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds. 1388 * The initiator_depth field must match the responder resources specified by the remote node when running 1389 * the rdma_connect and rdma_accept functions. */ 1390 if (rdma_param->initiator_depth == 0) { 1391 SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n"); 1392 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD); 1393 return -1; 1394 } 1395 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1396 1397 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1398 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1399 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1400 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1401 1402 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1403 max_queue_depth, max_read_depth); 1404 1405 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1406 if (rqpair == NULL) { 1407 SPDK_ERRLOG("Could not allocate new connection.\n"); 1408 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1409 return -1; 1410 } 1411 1412 rqpair->device = port->device; 1413 rqpair->max_queue_depth = max_queue_depth; 1414 rqpair->max_read_depth = max_read_depth; 1415 rqpair->cm_id = event->id; 1416 rqpair->listen_id = event->listen_id; 1417 rqpair->qpair.transport = transport; 1418 STAILQ_INIT(&rqpair->ibv_events); 1419 /* use qid from the private data to determine the qpair type 1420 qid will be set to the appropriate value when the controller is created */ 1421 rqpair->qpair.qid = private_data->qid; 1422 1423 event->id->context = &rqpair->qpair; 1424 1425 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1426 1427 return 0; 1428 } 1429 1430 static inline void 1431 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1432 enum spdk_nvme_data_transfer xfer) 1433 { 1434 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1435 wr->opcode = IBV_WR_RDMA_WRITE; 1436 wr->send_flags = 0; 1437 wr->next = next; 1438 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1439 wr->opcode = IBV_WR_RDMA_READ; 1440 wr->send_flags = IBV_SEND_SIGNALED; 1441 wr->next = NULL; 1442 } else { 1443 assert(0); 1444 } 1445 } 1446 1447 static int 1448 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1449 struct spdk_nvmf_rdma_request *rdma_req, 1450 uint32_t num_sgl_descriptors) 1451 { 1452 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1453 struct spdk_nvmf_rdma_request_data *current_data_wr; 1454 uint32_t i; 1455 1456 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1457 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1458 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1459 return -EINVAL; 1460 } 1461 1462 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1463 return -ENOMEM; 1464 } 1465 1466 current_data_wr = &rdma_req->data; 1467 1468 for (i = 0; i < num_sgl_descriptors; i++) { 1469 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1470 current_data_wr->wr.next = &work_requests[i]->wr; 1471 current_data_wr = work_requests[i]; 1472 current_data_wr->wr.sg_list = current_data_wr->sgl; 1473 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1474 } 1475 1476 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1477 1478 return 0; 1479 } 1480 1481 static inline void 1482 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1483 { 1484 struct ibv_send_wr *wr = &rdma_req->data.wr; 1485 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1486 1487 wr->wr.rdma.rkey = sgl->keyed.key; 1488 wr->wr.rdma.remote_addr = sgl->address; 1489 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1490 } 1491 1492 static inline void 1493 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1494 { 1495 struct ibv_send_wr *wr = &rdma_req->data.wr; 1496 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1497 uint32_t i; 1498 int j; 1499 uint64_t remote_addr_offset = 0; 1500 1501 for (i = 0; i < num_wrs; ++i) { 1502 wr->wr.rdma.rkey = sgl->keyed.key; 1503 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1504 for (j = 0; j < wr->num_sge; ++j) { 1505 remote_addr_offset += wr->sg_list[j].length; 1506 } 1507 wr = wr->next; 1508 } 1509 } 1510 1511 static int 1512 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1513 struct spdk_nvmf_rdma_device *device, 1514 struct spdk_nvmf_rdma_request *rdma_req, 1515 struct ibv_send_wr *wr, 1516 uint32_t total_length) 1517 { 1518 struct spdk_rdma_memory_translation mem_translation; 1519 struct ibv_sge *sg_ele; 1520 struct iovec *iov; 1521 uint32_t lkey, remaining; 1522 int rc; 1523 1524 wr->num_sge = 0; 1525 1526 while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) { 1527 iov = &rdma_req->req.iov[rdma_req->iovpos]; 1528 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1529 if (spdk_unlikely(rc)) { 1530 return rc; 1531 } 1532 1533 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1534 sg_ele = &wr->sg_list[wr->num_sge]; 1535 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1536 1537 sg_ele->lkey = lkey; 1538 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset; 1539 sg_ele->length = remaining; 1540 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1541 sg_ele->length); 1542 rdma_req->offset += sg_ele->length; 1543 total_length -= sg_ele->length; 1544 wr->num_sge++; 1545 1546 if (rdma_req->offset == iov->iov_len) { 1547 rdma_req->offset = 0; 1548 rdma_req->iovpos++; 1549 } 1550 } 1551 1552 if (total_length) { 1553 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1554 return -EINVAL; 1555 } 1556 1557 return 0; 1558 } 1559 1560 static int 1561 nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_poll_group *rgroup, 1562 struct spdk_nvmf_rdma_device *device, 1563 struct spdk_nvmf_rdma_request *rdma_req, 1564 struct ibv_send_wr *wr, 1565 uint32_t total_length, 1566 uint32_t num_extra_wrs) 1567 { 1568 struct spdk_rdma_memory_translation mem_translation; 1569 struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx; 1570 struct ibv_sge *sg_ele; 1571 struct iovec *iov; 1572 struct iovec *rdma_iov; 1573 uint32_t lkey, remaining; 1574 uint32_t remaining_data_block, data_block_size, md_size; 1575 uint32_t sge_len; 1576 int rc; 1577 1578 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1579 1580 if (spdk_likely(!rdma_req->req.stripped_data)) { 1581 rdma_iov = rdma_req->req.iov; 1582 remaining_data_block = data_block_size; 1583 md_size = dif_ctx->md_size; 1584 } else { 1585 rdma_iov = rdma_req->req.stripped_data->iov; 1586 total_length = total_length / dif_ctx->block_size * data_block_size; 1587 remaining_data_block = total_length; 1588 md_size = 0; 1589 } 1590 1591 wr->num_sge = 0; 1592 1593 while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1594 iov = rdma_iov + rdma_req->iovpos; 1595 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1596 if (spdk_unlikely(rc)) { 1597 return rc; 1598 } 1599 1600 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1601 sg_ele = &wr->sg_list[wr->num_sge]; 1602 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1603 1604 while (remaining) { 1605 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1606 if (num_extra_wrs > 0 && wr->next) { 1607 wr = wr->next; 1608 wr->num_sge = 0; 1609 sg_ele = &wr->sg_list[wr->num_sge]; 1610 num_extra_wrs--; 1611 } else { 1612 break; 1613 } 1614 } 1615 sg_ele->lkey = lkey; 1616 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset); 1617 sge_len = spdk_min(remaining, remaining_data_block); 1618 sg_ele->length = sge_len; 1619 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, 1620 sg_ele->addr, sg_ele->length); 1621 remaining -= sge_len; 1622 remaining_data_block -= sge_len; 1623 rdma_req->offset += sge_len; 1624 total_length -= sge_len; 1625 1626 sg_ele++; 1627 wr->num_sge++; 1628 1629 if (remaining_data_block == 0) { 1630 /* skip metadata */ 1631 rdma_req->offset += md_size; 1632 total_length -= md_size; 1633 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1634 remaining -= spdk_min(remaining, md_size); 1635 remaining_data_block = data_block_size; 1636 } 1637 1638 if (remaining == 0) { 1639 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1640 the remaining metadata bits at the beginning of the next buffer */ 1641 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset); 1642 rdma_req->iovpos++; 1643 } 1644 } 1645 } 1646 1647 if (total_length) { 1648 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1649 return -EINVAL; 1650 } 1651 1652 return 0; 1653 } 1654 1655 static inline uint32_t 1656 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1657 { 1658 /* estimate the number of SG entries and WRs needed to process the request */ 1659 uint32_t num_sge = 0; 1660 uint32_t i; 1661 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1662 1663 for (i = 0; i < num_buffers && length > 0; i++) { 1664 uint32_t buffer_len = spdk_min(length, io_unit_size); 1665 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1666 1667 if (num_sge_in_block * block_size > buffer_len) { 1668 ++num_sge_in_block; 1669 } 1670 num_sge += num_sge_in_block; 1671 length -= buffer_len; 1672 } 1673 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1674 } 1675 1676 static int 1677 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1678 struct spdk_nvmf_rdma_device *device, 1679 struct spdk_nvmf_rdma_request *rdma_req) 1680 { 1681 struct spdk_nvmf_rdma_qpair *rqpair; 1682 struct spdk_nvmf_rdma_poll_group *rgroup; 1683 struct spdk_nvmf_request *req = &rdma_req->req; 1684 struct ibv_send_wr *wr = &rdma_req->data.wr; 1685 int rc; 1686 uint32_t num_wrs = 1; 1687 uint32_t length; 1688 1689 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1690 rgroup = rqpair->poller->group; 1691 1692 /* rdma wr specifics */ 1693 nvmf_rdma_setup_request(rdma_req); 1694 1695 length = req->length; 1696 if (spdk_unlikely(req->dif_enabled)) { 1697 req->dif.orig_length = length; 1698 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1699 req->dif.elba_length = length; 1700 } 1701 1702 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1703 length); 1704 if (rc != 0) { 1705 return rc; 1706 } 1707 1708 assert(req->iovcnt <= rqpair->max_send_sge); 1709 1710 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1711 * the stripped_buffers are got for DIF stripping. */ 1712 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1713 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1714 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1715 &rtransport->transport, req->dif.orig_length); 1716 if (rc != 0) { 1717 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1718 } 1719 } 1720 1721 rdma_req->iovpos = 0; 1722 1723 if (spdk_unlikely(req->dif_enabled)) { 1724 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1725 req->dif.dif_ctx.block_size); 1726 if (num_wrs > 1) { 1727 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1728 if (rc != 0) { 1729 goto err_exit; 1730 } 1731 } 1732 1733 rc = nvmf_rdma_fill_wr_sgl_with_dif(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1734 if (spdk_unlikely(rc != 0)) { 1735 goto err_exit; 1736 } 1737 1738 if (num_wrs > 1) { 1739 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1740 } 1741 } else { 1742 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length); 1743 if (spdk_unlikely(rc != 0)) { 1744 goto err_exit; 1745 } 1746 } 1747 1748 /* set the number of outstanding data WRs for this request. */ 1749 rdma_req->num_outstanding_data_wr = num_wrs; 1750 1751 return rc; 1752 1753 err_exit: 1754 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1755 nvmf_rdma_request_free_data(rdma_req, rtransport); 1756 req->iovcnt = 0; 1757 return rc; 1758 } 1759 1760 static int 1761 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1762 struct spdk_nvmf_rdma_device *device, 1763 struct spdk_nvmf_rdma_request *rdma_req) 1764 { 1765 struct spdk_nvmf_rdma_qpair *rqpair; 1766 struct spdk_nvmf_rdma_poll_group *rgroup; 1767 struct ibv_send_wr *current_wr; 1768 struct spdk_nvmf_request *req = &rdma_req->req; 1769 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1770 uint32_t num_sgl_descriptors; 1771 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0; 1772 uint32_t i; 1773 int rc; 1774 1775 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1776 rgroup = rqpair->poller->group; 1777 1778 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1779 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1780 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1781 1782 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1783 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1784 1785 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1786 for (i = 0; i < num_sgl_descriptors; i++) { 1787 if (spdk_likely(!req->dif_enabled)) { 1788 lengths[i] = desc->keyed.length; 1789 } else { 1790 req->dif.orig_length += desc->keyed.length; 1791 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1792 req->dif.elba_length += lengths[i]; 1793 } 1794 total_length += lengths[i]; 1795 desc++; 1796 } 1797 1798 if (total_length > rtransport->transport.opts.max_io_size) { 1799 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n", 1800 total_length, rtransport->transport.opts.max_io_size); 1801 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1802 return -EINVAL; 1803 } 1804 1805 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1806 return -ENOMEM; 1807 } 1808 1809 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length); 1810 if (rc != 0) { 1811 nvmf_rdma_request_free_data(rdma_req, rtransport); 1812 return rc; 1813 } 1814 1815 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1816 * the stripped_buffers are got for DIF stripping. */ 1817 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1818 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1819 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1820 &rtransport->transport, req->dif.orig_length); 1821 if (rc != 0) { 1822 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1823 } 1824 } 1825 1826 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1827 current_wr = &rdma_req->data.wr; 1828 assert(current_wr != NULL); 1829 1830 req->length = 0; 1831 rdma_req->iovpos = 0; 1832 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1833 for (i = 0; i < num_sgl_descriptors; i++) { 1834 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1835 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1836 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1837 rc = -EINVAL; 1838 goto err_exit; 1839 } 1840 1841 if (spdk_likely(!req->dif_enabled)) { 1842 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i]); 1843 } else { 1844 rc = nvmf_rdma_fill_wr_sgl_with_dif(rgroup, device, rdma_req, current_wr, 1845 lengths[i], 0); 1846 } 1847 if (rc != 0) { 1848 rc = -ENOMEM; 1849 goto err_exit; 1850 } 1851 1852 req->length += desc->keyed.length; 1853 current_wr->wr.rdma.rkey = desc->keyed.key; 1854 current_wr->wr.rdma.remote_addr = desc->address; 1855 current_wr = current_wr->next; 1856 desc++; 1857 } 1858 1859 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1860 /* Go back to the last descriptor in the list. */ 1861 desc--; 1862 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1863 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1864 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1865 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1866 } 1867 } 1868 #endif 1869 1870 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1871 1872 return 0; 1873 1874 err_exit: 1875 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1876 nvmf_rdma_request_free_data(rdma_req, rtransport); 1877 return rc; 1878 } 1879 1880 static int 1881 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1882 struct spdk_nvmf_rdma_device *device, 1883 struct spdk_nvmf_rdma_request *rdma_req) 1884 { 1885 struct spdk_nvmf_request *req = &rdma_req->req; 1886 struct spdk_nvme_cpl *rsp; 1887 struct spdk_nvme_sgl_descriptor *sgl; 1888 int rc; 1889 uint32_t length; 1890 1891 rsp = &req->rsp->nvme_cpl; 1892 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1893 1894 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1895 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1896 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1897 1898 length = sgl->keyed.length; 1899 if (length > rtransport->transport.opts.max_io_size) { 1900 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1901 length, rtransport->transport.opts.max_io_size); 1902 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1903 return -1; 1904 } 1905 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1906 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1907 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1908 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1909 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1910 } 1911 } 1912 #endif 1913 1914 /* fill request length and populate iovs */ 1915 req->length = length; 1916 1917 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req); 1918 if (spdk_unlikely(rc < 0)) { 1919 if (rc == -EINVAL) { 1920 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1921 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1922 return -1; 1923 } 1924 /* No available buffers. Queue this request up. */ 1925 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1926 return 0; 1927 } 1928 1929 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1930 req->iovcnt); 1931 1932 return 0; 1933 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1934 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1935 uint64_t offset = sgl->address; 1936 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1937 1938 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1939 offset, sgl->unkeyed.length); 1940 1941 if (offset > max_len) { 1942 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1943 offset, max_len); 1944 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1945 return -1; 1946 } 1947 max_len -= (uint32_t)offset; 1948 1949 if (sgl->unkeyed.length > max_len) { 1950 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1951 sgl->unkeyed.length, max_len); 1952 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1953 return -1; 1954 } 1955 1956 rdma_req->num_outstanding_data_wr = 0; 1957 req->data_from_pool = false; 1958 req->length = sgl->unkeyed.length; 1959 1960 req->iov[0].iov_base = rdma_req->recv->buf + offset; 1961 req->iov[0].iov_len = req->length; 1962 req->iovcnt = 1; 1963 1964 return 0; 1965 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1966 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1967 1968 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1969 if (rc == -ENOMEM) { 1970 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1971 return 0; 1972 } else if (rc == -EINVAL) { 1973 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1974 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1975 return -1; 1976 } 1977 1978 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1979 req->iovcnt); 1980 1981 return 0; 1982 } 1983 1984 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1985 sgl->generic.type, sgl->generic.subtype); 1986 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1987 return -1; 1988 } 1989 1990 static void 1991 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1992 struct spdk_nvmf_rdma_transport *rtransport) 1993 { 1994 struct spdk_nvmf_rdma_qpair *rqpair; 1995 struct spdk_nvmf_rdma_poll_group *rgroup; 1996 1997 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1998 if (rdma_req->req.data_from_pool) { 1999 rgroup = rqpair->poller->group; 2000 2001 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 2002 } 2003 if (rdma_req->req.stripped_data) { 2004 nvmf_request_free_stripped_buffers(&rdma_req->req, 2005 &rqpair->poller->group->group, 2006 &rtransport->transport); 2007 } 2008 nvmf_rdma_request_free_data(rdma_req, rtransport); 2009 rdma_req->req.length = 0; 2010 rdma_req->req.iovcnt = 0; 2011 rdma_req->offset = 0; 2012 rdma_req->req.dif_enabled = false; 2013 rdma_req->fused_failed = false; 2014 if (rdma_req->fused_pair) { 2015 /* This req was part of a valid fused pair, but failed before it got to 2016 * READ_TO_EXECUTE state. This means we need to fail the other request 2017 * in the pair, because it is no longer part of a valid pair. If the pair 2018 * already reached READY_TO_EXECUTE state, we need to kick it. 2019 */ 2020 rdma_req->fused_pair->fused_failed = true; 2021 if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2022 nvmf_rdma_request_process(rtransport, rdma_req->fused_pair); 2023 } 2024 rdma_req->fused_pair = NULL; 2025 } 2026 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 2027 rqpair->qd--; 2028 2029 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 2030 rdma_req->state = RDMA_REQUEST_STATE_FREE; 2031 } 2032 2033 static void 2034 nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport, 2035 struct spdk_nvmf_rdma_qpair *rqpair, 2036 struct spdk_nvmf_rdma_request *rdma_req) 2037 { 2038 enum spdk_nvme_cmd_fuse last, next; 2039 2040 last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2041 next = rdma_req->req.cmd->nvme_cmd.fuse; 2042 2043 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2044 2045 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2046 return; 2047 } 2048 2049 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2050 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2051 /* This is a valid pair of fused commands. Point them at each other 2052 * so they can be submitted consecutively once ready to be executed. 2053 */ 2054 rqpair->fused_first->fused_pair = rdma_req; 2055 rdma_req->fused_pair = rqpair->fused_first; 2056 rqpair->fused_first = NULL; 2057 return; 2058 } else { 2059 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2060 rqpair->fused_first->fused_failed = true; 2061 2062 /* If the last req is in READY_TO_EXECUTE state, then call 2063 * nvmf_rdma_request_process(), otherwise nothing else will kick it. 2064 */ 2065 if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2066 nvmf_rdma_request_process(rtransport, rqpair->fused_first); 2067 } 2068 2069 rqpair->fused_first = NULL; 2070 } 2071 } 2072 2073 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2074 /* Set rqpair->fused_first here so that we know to check that the next request 2075 * is a SECOND (and to fail this one if it isn't). 2076 */ 2077 rqpair->fused_first = rdma_req; 2078 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2079 /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */ 2080 rdma_req->fused_failed = true; 2081 } 2082 } 2083 2084 bool 2085 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 2086 struct spdk_nvmf_rdma_request *rdma_req) 2087 { 2088 struct spdk_nvmf_rdma_qpair *rqpair; 2089 struct spdk_nvmf_rdma_device *device; 2090 struct spdk_nvmf_rdma_poll_group *rgroup; 2091 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2092 int rc; 2093 struct spdk_nvmf_rdma_recv *rdma_recv; 2094 enum spdk_nvmf_rdma_request_state prev_state; 2095 bool progress = false; 2096 int data_posted; 2097 uint32_t num_blocks, num_rdma_reads_available, qdepth; 2098 2099 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2100 device = rqpair->device; 2101 rgroup = rqpair->poller->group; 2102 2103 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 2104 2105 /* If the queue pair is in an error state, force the request to the completed state 2106 * to release resources. */ 2107 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2108 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 2109 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 2110 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 2111 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2112 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 2113 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2114 } 2115 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2116 } 2117 2118 /* The loop here is to allow for several back-to-back state changes. */ 2119 do { 2120 prev_state = rdma_req->state; 2121 2122 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 2123 2124 switch (rdma_req->state) { 2125 case RDMA_REQUEST_STATE_FREE: 2126 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 2127 * to escape this state. */ 2128 break; 2129 case RDMA_REQUEST_STATE_NEW: 2130 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 2131 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2132 rdma_recv = rdma_req->recv; 2133 2134 /* The first element of the SGL is the NVMe command */ 2135 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 2136 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 2137 rdma_req->transfer_wr = &rdma_req->data.wr; 2138 2139 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2140 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2141 break; 2142 } 2143 2144 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 2145 rdma_req->req.dif_enabled = true; 2146 } 2147 2148 nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req); 2149 2150 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2151 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 2152 rdma_req->rsp.wr.imm_data = 0; 2153 #endif 2154 2155 /* The next state transition depends on the data transfer needs of this request. */ 2156 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2157 2158 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2159 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2160 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2161 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2162 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 2163 break; 2164 } 2165 2166 /* If no data to transfer, ready to execute. */ 2167 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2168 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2169 break; 2170 } 2171 2172 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2173 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2174 break; 2175 case RDMA_REQUEST_STATE_NEED_BUFFER: 2176 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2177 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2178 2179 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2180 2181 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2182 /* This request needs to wait in line to obtain a buffer */ 2183 break; 2184 } 2185 2186 /* Try to get a data buffer */ 2187 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2188 if (rc < 0) { 2189 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2190 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2191 break; 2192 } 2193 2194 if (rdma_req->req.iovcnt == 0) { 2195 /* No buffers available. */ 2196 rgroup->stat.pending_data_buffer++; 2197 break; 2198 } 2199 2200 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2201 2202 /* If data is transferring from host to controller and the data didn't 2203 * arrive using in capsule data, we need to do a transfer from the host. 2204 */ 2205 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2206 rdma_req->req.data_from_pool) { 2207 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2208 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2209 break; 2210 } 2211 2212 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2213 break; 2214 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2215 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2216 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2217 2218 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2219 /* This request needs to wait in line to perform RDMA */ 2220 break; 2221 } 2222 qdepth = rqpair->max_send_depth - rqpair->current_send_depth; 2223 num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth; 2224 if (rdma_req->num_outstanding_data_wr > qdepth || 2225 rdma_req->num_outstanding_data_wr > num_rdma_reads_available) { 2226 if (num_rdma_reads_available && qdepth) { 2227 /* Send as much as we can */ 2228 request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth)); 2229 } else { 2230 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2231 rqpair->poller->stat.pending_rdma_read++; 2232 break; 2233 } 2234 } 2235 2236 /* We have already verified that this request is the head of the queue. */ 2237 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2238 2239 rc = request_transfer_in(&rdma_req->req); 2240 if (!rc) { 2241 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2242 } else { 2243 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2244 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2245 } 2246 break; 2247 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2248 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2249 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2250 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2251 * to escape this state. */ 2252 break; 2253 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2254 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2255 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2256 2257 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2258 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2259 /* generate DIF for write operation */ 2260 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2261 assert(num_blocks > 0); 2262 2263 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2264 num_blocks, &rdma_req->req.dif.dif_ctx); 2265 if (rc != 0) { 2266 SPDK_ERRLOG("DIF generation failed\n"); 2267 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2268 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2269 break; 2270 } 2271 } 2272 2273 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2274 /* set extended length before IO operation */ 2275 rdma_req->req.length = rdma_req->req.dif.elba_length; 2276 } 2277 2278 if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 2279 if (rdma_req->fused_failed) { 2280 /* This request failed FUSED semantics. Fail it immediately, without 2281 * even sending it to the target layer. 2282 */ 2283 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2284 rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED; 2285 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2286 break; 2287 } 2288 2289 if (rdma_req->fused_pair == NULL || 2290 rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2291 /* This request is ready to execute, but either we don't know yet if it's 2292 * valid - i.e. this is a FIRST but we haven't received the next 2293 * request yet or the other request of this fused pair isn't ready to 2294 * execute. So break here and this request will get processed later either 2295 * when the other request is ready or we find that this request isn't valid. 2296 */ 2297 break; 2298 } 2299 } 2300 2301 /* If we get to this point, and this request is a fused command, we know that 2302 * it is part of valid sequence (FIRST followed by a SECOND) and that both 2303 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 2304 * request, and the other request of the fused pair, in the correct order. 2305 * Also clear the ->fused_pair pointers on both requests, since after this point 2306 * we no longer need to maintain the relationship between these two requests. 2307 */ 2308 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 2309 assert(rdma_req->fused_pair != NULL); 2310 assert(rdma_req->fused_pair->fused_pair != NULL); 2311 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2312 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2313 rdma_req->fused_pair->fused_pair = NULL; 2314 rdma_req->fused_pair = NULL; 2315 } 2316 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2317 spdk_nvmf_request_exec(&rdma_req->req); 2318 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 2319 assert(rdma_req->fused_pair != NULL); 2320 assert(rdma_req->fused_pair->fused_pair != NULL); 2321 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2322 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2323 rdma_req->fused_pair->fused_pair = NULL; 2324 rdma_req->fused_pair = NULL; 2325 } 2326 break; 2327 case RDMA_REQUEST_STATE_EXECUTING: 2328 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2329 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2330 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2331 * to escape this state. */ 2332 break; 2333 case RDMA_REQUEST_STATE_EXECUTED: 2334 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2335 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2336 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2337 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2338 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2339 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2340 } else { 2341 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2342 } 2343 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2344 /* restore the original length */ 2345 rdma_req->req.length = rdma_req->req.dif.orig_length; 2346 2347 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2348 struct spdk_dif_error error_blk; 2349 2350 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2351 if (!rdma_req->req.stripped_data) { 2352 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2353 &rdma_req->req.dif.dif_ctx, &error_blk); 2354 } else { 2355 rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov, 2356 rdma_req->req.stripped_data->iovcnt, 2357 rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2358 &rdma_req->req.dif.dif_ctx, &error_blk); 2359 } 2360 if (rc) { 2361 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2362 2363 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2364 error_blk.err_offset); 2365 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2366 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2367 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2368 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2369 } 2370 } 2371 } 2372 break; 2373 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2374 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2375 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2376 2377 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2378 /* This request needs to wait in line to perform RDMA */ 2379 break; 2380 } 2381 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2382 rqpair->max_send_depth) { 2383 /* We can only have so many WRs outstanding. we have to wait until some finish. 2384 * +1 since each request has an additional wr in the resp. */ 2385 rqpair->poller->stat.pending_rdma_write++; 2386 break; 2387 } 2388 2389 /* We have already verified that this request is the head of the queue. */ 2390 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2391 2392 /* The data transfer will be kicked off from 2393 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2394 */ 2395 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2396 break; 2397 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2398 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2399 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2400 rc = request_transfer_out(&rdma_req->req, &data_posted); 2401 assert(rc == 0); /* No good way to handle this currently */ 2402 if (rc) { 2403 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2404 } else { 2405 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2406 RDMA_REQUEST_STATE_COMPLETING; 2407 } 2408 break; 2409 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2410 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2411 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2412 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2413 * to escape this state. */ 2414 break; 2415 case RDMA_REQUEST_STATE_COMPLETING: 2416 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2417 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2418 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2419 * to escape this state. */ 2420 break; 2421 case RDMA_REQUEST_STATE_COMPLETED: 2422 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2423 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2424 2425 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2426 _nvmf_rdma_request_free(rdma_req, rtransport); 2427 break; 2428 case RDMA_REQUEST_NUM_STATES: 2429 default: 2430 assert(0); 2431 break; 2432 } 2433 2434 if (rdma_req->state != prev_state) { 2435 progress = true; 2436 } 2437 } while (rdma_req->state != prev_state); 2438 2439 return progress; 2440 } 2441 2442 /* Public API callbacks begin here */ 2443 2444 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2445 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2446 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2447 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2448 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2449 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2450 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2451 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2452 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX 2453 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2454 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2455 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2456 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2457 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2458 2459 static void 2460 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2461 { 2462 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2463 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2464 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2465 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2466 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2467 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2468 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2469 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2470 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2471 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2472 opts->transport_specific = NULL; 2473 } 2474 2475 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2476 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2477 2478 static inline bool 2479 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2480 { 2481 return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD || 2482 device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW; 2483 } 2484 2485 static int nvmf_rdma_accept(void *ctx); 2486 static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport); 2487 static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2488 struct spdk_nvmf_rdma_device *device); 2489 2490 static int 2491 create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context, 2492 struct spdk_nvmf_rdma_device **new_device) 2493 { 2494 struct spdk_nvmf_rdma_device *device; 2495 int flag = 0; 2496 int rc = 0; 2497 2498 device = calloc(1, sizeof(*device)); 2499 if (!device) { 2500 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2501 return -ENOMEM; 2502 } 2503 device->context = context; 2504 rc = ibv_query_device(device->context, &device->attr); 2505 if (rc < 0) { 2506 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2507 free(device); 2508 return rc; 2509 } 2510 2511 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2512 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2513 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2514 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2515 } 2516 2517 /** 2518 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2519 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2520 * but incorrectly reports that it does. There are changes making their way 2521 * through the kernel now that will enable this feature. When they are merged, 2522 * we can conditionally enable this feature. 2523 * 2524 * TODO: enable this for versions of the kernel rxe driver that support it. 2525 */ 2526 if (nvmf_rdma_is_rxe_device(device)) { 2527 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2528 } 2529 #endif 2530 2531 /* set up device context async ev fd as NON_BLOCKING */ 2532 flag = fcntl(device->context->async_fd, F_GETFL); 2533 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2534 if (rc < 0) { 2535 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2536 free(device); 2537 return rc; 2538 } 2539 2540 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2541 SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA trasport\n", device); 2542 2543 if (g_nvmf_hooks.get_ibv_pd) { 2544 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2545 } else { 2546 device->pd = ibv_alloc_pd(device->context); 2547 } 2548 2549 if (!device->pd) { 2550 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2551 destroy_ib_device(rtransport, device); 2552 return -ENOMEM; 2553 } 2554 2555 assert(device->map == NULL); 2556 2557 device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks, SPDK_RDMA_MEMORY_MAP_ROLE_TARGET); 2558 if (!device->map) { 2559 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2560 destroy_ib_device(rtransport, device); 2561 return -ENOMEM; 2562 } 2563 2564 assert(device->map != NULL); 2565 assert(device->pd != NULL); 2566 2567 if (new_device) { 2568 *new_device = device; 2569 } 2570 SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device), 2571 device, context); 2572 2573 return 0; 2574 } 2575 2576 static void 2577 free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2578 { 2579 if (rtransport->poll_fds) { 2580 free(rtransport->poll_fds); 2581 rtransport->poll_fds = NULL; 2582 } 2583 rtransport->npoll_fds = 0; 2584 } 2585 2586 static int 2587 generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2588 { 2589 /* Set up poll descriptor array to monitor events from RDMA and IB 2590 * in a single poll syscall 2591 */ 2592 int device_count = 0; 2593 int i = 0; 2594 struct spdk_nvmf_rdma_device *device, *tmp; 2595 2596 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2597 device_count++; 2598 } 2599 2600 rtransport->npoll_fds = device_count + 1; 2601 2602 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2603 if (rtransport->poll_fds == NULL) { 2604 SPDK_ERRLOG("poll_fds allocation failed\n"); 2605 return -ENOMEM; 2606 } 2607 2608 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2609 rtransport->poll_fds[i++].events = POLLIN; 2610 2611 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2612 rtransport->poll_fds[i].fd = device->context->async_fd; 2613 rtransport->poll_fds[i++].events = POLLIN; 2614 } 2615 2616 return 0; 2617 } 2618 2619 static struct spdk_nvmf_transport * 2620 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2621 { 2622 int rc; 2623 struct spdk_nvmf_rdma_transport *rtransport; 2624 struct spdk_nvmf_rdma_device *device; 2625 struct ibv_context **contexts; 2626 uint32_t i; 2627 int flag; 2628 uint32_t sge_count; 2629 uint32_t min_shared_buffers; 2630 uint32_t min_in_capsule_data_size; 2631 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2632 2633 rtransport = calloc(1, sizeof(*rtransport)); 2634 if (!rtransport) { 2635 return NULL; 2636 } 2637 2638 TAILQ_INIT(&rtransport->devices); 2639 TAILQ_INIT(&rtransport->ports); 2640 TAILQ_INIT(&rtransport->poll_groups); 2641 TAILQ_INIT(&rtransport->retry_ports); 2642 2643 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2644 rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2645 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2646 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2647 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2648 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2649 if (opts->transport_specific != NULL && 2650 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2651 SPDK_COUNTOF(rdma_transport_opts_decoder), 2652 &rtransport->rdma_opts)) { 2653 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2654 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2655 return NULL; 2656 } 2657 2658 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2659 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2660 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2661 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2662 " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d," 2663 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2664 opts->max_queue_depth, 2665 opts->max_io_size, 2666 opts->max_qpairs_per_ctrlr - 1, 2667 opts->io_unit_size, 2668 opts->in_capsule_data_size, 2669 opts->max_aq_depth, 2670 opts->num_shared_buffers, 2671 rtransport->rdma_opts.num_cqe, 2672 rtransport->rdma_opts.max_srq_depth, 2673 rtransport->rdma_opts.no_srq, 2674 rtransport->rdma_opts.acceptor_backlog, 2675 rtransport->rdma_opts.no_wr_batching, 2676 opts->abort_timeout_sec); 2677 2678 /* I/O unit size cannot be larger than max I/O size */ 2679 if (opts->io_unit_size > opts->max_io_size) { 2680 opts->io_unit_size = opts->max_io_size; 2681 } 2682 2683 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2684 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2685 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2686 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2687 } 2688 2689 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2690 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2691 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2692 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2693 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2694 return NULL; 2695 } 2696 2697 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 2698 if (opts->buf_cache_size < UINT32_MAX) { 2699 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2700 if (min_shared_buffers > opts->num_shared_buffers) { 2701 SPDK_ERRLOG("There are not enough buffers to satisfy" 2702 "per-poll group caches for each thread. (%" PRIu32 ")" 2703 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2704 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2705 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2706 return NULL; 2707 } 2708 } 2709 2710 sge_count = opts->max_io_size / opts->io_unit_size; 2711 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2712 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2713 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2714 return NULL; 2715 } 2716 2717 min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES; 2718 if (opts->in_capsule_data_size < min_in_capsule_data_size) { 2719 SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n", 2720 min_in_capsule_data_size); 2721 opts->in_capsule_data_size = min_in_capsule_data_size; 2722 } 2723 2724 rtransport->event_channel = rdma_create_event_channel(); 2725 if (rtransport->event_channel == NULL) { 2726 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2727 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2728 return NULL; 2729 } 2730 2731 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2732 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2733 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2734 rtransport->event_channel->fd, spdk_strerror(errno)); 2735 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2736 return NULL; 2737 } 2738 2739 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2740 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2741 sizeof(struct spdk_nvmf_rdma_request_data), 2742 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2743 SPDK_ENV_SOCKET_ID_ANY); 2744 if (!rtransport->data_wr_pool) { 2745 if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) { 2746 SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n"); 2747 SPDK_ERRLOG("Probably running in multiprocess environment, which is " 2748 "unsupported by the nvmf library\n"); 2749 } else { 2750 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2751 } 2752 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2753 return NULL; 2754 } 2755 2756 contexts = rdma_get_devices(NULL); 2757 if (contexts == NULL) { 2758 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2759 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2760 return NULL; 2761 } 2762 2763 i = 0; 2764 rc = 0; 2765 while (contexts[i] != NULL) { 2766 rc = create_ib_device(rtransport, contexts[i], &device); 2767 if (rc < 0) { 2768 break; 2769 } 2770 i++; 2771 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2772 device->is_ready = true; 2773 } 2774 rdma_free_devices(contexts); 2775 2776 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2777 /* divide and round up. */ 2778 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2779 2780 /* round up to the nearest 4k. */ 2781 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2782 2783 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2784 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2785 opts->io_unit_size); 2786 } 2787 2788 if (rc < 0) { 2789 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2790 return NULL; 2791 } 2792 2793 rc = generate_poll_fds(rtransport); 2794 if (rc < 0) { 2795 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2796 return NULL; 2797 } 2798 2799 rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport, 2800 opts->acceptor_poll_rate); 2801 if (!rtransport->accept_poller) { 2802 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2803 return NULL; 2804 } 2805 2806 return &rtransport->transport; 2807 } 2808 2809 static void 2810 destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2811 struct spdk_nvmf_rdma_device *device) 2812 { 2813 TAILQ_REMOVE(&rtransport->devices, device, link); 2814 spdk_rdma_free_mem_map(&device->map); 2815 if (device->pd) { 2816 if (!g_nvmf_hooks.get_ibv_pd) { 2817 ibv_dealloc_pd(device->pd); 2818 } 2819 } 2820 SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device); 2821 free(device); 2822 } 2823 2824 static void 2825 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2826 { 2827 struct spdk_nvmf_rdma_transport *rtransport; 2828 assert(w != NULL); 2829 2830 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2831 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2832 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2833 if (rtransport->rdma_opts.no_srq == true) { 2834 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2835 } 2836 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2837 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2838 } 2839 2840 static int 2841 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2842 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2843 { 2844 struct spdk_nvmf_rdma_transport *rtransport; 2845 struct spdk_nvmf_rdma_port *port, *port_tmp; 2846 struct spdk_nvmf_rdma_device *device, *device_tmp; 2847 2848 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2849 2850 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) { 2851 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 2852 free(port); 2853 } 2854 2855 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2856 TAILQ_REMOVE(&rtransport->ports, port, link); 2857 rdma_destroy_id(port->id); 2858 free(port); 2859 } 2860 2861 free_poll_fds(rtransport); 2862 2863 if (rtransport->event_channel != NULL) { 2864 rdma_destroy_event_channel(rtransport->event_channel); 2865 } 2866 2867 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2868 destroy_ib_device(rtransport, device); 2869 } 2870 2871 if (rtransport->data_wr_pool != NULL) { 2872 if (spdk_mempool_count(rtransport->data_wr_pool) != 2873 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2874 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2875 spdk_mempool_count(rtransport->data_wr_pool), 2876 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2877 } 2878 } 2879 2880 spdk_mempool_free(rtransport->data_wr_pool); 2881 2882 spdk_poller_unregister(&rtransport->accept_poller); 2883 free(rtransport); 2884 2885 if (cb_fn) { 2886 cb_fn(cb_arg); 2887 } 2888 return 0; 2889 } 2890 2891 static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2892 struct spdk_nvme_transport_id *trid, 2893 bool peer); 2894 2895 static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport); 2896 2897 static int 2898 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2899 struct spdk_nvmf_listen_opts *listen_opts) 2900 { 2901 struct spdk_nvmf_rdma_transport *rtransport; 2902 struct spdk_nvmf_rdma_device *device; 2903 struct spdk_nvmf_rdma_port *port, *tmp_port; 2904 struct addrinfo *res; 2905 struct addrinfo hints; 2906 int family; 2907 int rc; 2908 bool is_retry = false; 2909 2910 if (!strlen(trid->trsvcid)) { 2911 SPDK_ERRLOG("Service id is required\n"); 2912 return -EINVAL; 2913 } 2914 2915 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2916 assert(rtransport->event_channel != NULL); 2917 2918 port = calloc(1, sizeof(*port)); 2919 if (!port) { 2920 SPDK_ERRLOG("Port allocation failed\n"); 2921 return -ENOMEM; 2922 } 2923 2924 port->trid = trid; 2925 2926 switch (trid->adrfam) { 2927 case SPDK_NVMF_ADRFAM_IPV4: 2928 family = AF_INET; 2929 break; 2930 case SPDK_NVMF_ADRFAM_IPV6: 2931 family = AF_INET6; 2932 break; 2933 default: 2934 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2935 free(port); 2936 return -EINVAL; 2937 } 2938 2939 memset(&hints, 0, sizeof(hints)); 2940 hints.ai_family = family; 2941 hints.ai_flags = AI_NUMERICSERV; 2942 hints.ai_socktype = SOCK_STREAM; 2943 hints.ai_protocol = 0; 2944 2945 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2946 if (rc) { 2947 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2948 free(port); 2949 return -(abs(rc)); 2950 } 2951 2952 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2953 if (rc < 0) { 2954 SPDK_ERRLOG("rdma_create_id() failed\n"); 2955 freeaddrinfo(res); 2956 free(port); 2957 return rc; 2958 } 2959 2960 rc = rdma_bind_addr(port->id, res->ai_addr); 2961 freeaddrinfo(res); 2962 2963 if (rc < 0) { 2964 TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) { 2965 if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) { 2966 is_retry = true; 2967 break; 2968 } 2969 } 2970 if (!is_retry) { 2971 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2972 } 2973 rdma_destroy_id(port->id); 2974 free(port); 2975 return rc; 2976 } 2977 2978 if (!port->id->verbs) { 2979 SPDK_ERRLOG("ibv_context is null\n"); 2980 rdma_destroy_id(port->id); 2981 free(port); 2982 return -1; 2983 } 2984 2985 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2986 if (rc < 0) { 2987 SPDK_ERRLOG("rdma_listen() failed\n"); 2988 rdma_destroy_id(port->id); 2989 free(port); 2990 return rc; 2991 } 2992 2993 TAILQ_FOREACH(device, &rtransport->devices, link) { 2994 if (device->context == port->id->verbs && device->is_ready) { 2995 port->device = device; 2996 break; 2997 } 2998 } 2999 if (!port->device) { 3000 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 3001 port->id->verbs); 3002 rdma_destroy_id(port->id); 3003 free(port); 3004 nvmf_rdma_rescan_devices(rtransport); 3005 return -EINVAL; 3006 } 3007 3008 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 3009 trid->traddr, trid->trsvcid); 3010 3011 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 3012 return 0; 3013 } 3014 3015 static void 3016 nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport, 3017 const struct spdk_nvme_transport_id *trid, bool need_retry) 3018 { 3019 struct spdk_nvmf_rdma_transport *rtransport; 3020 struct spdk_nvmf_rdma_port *port, *tmp; 3021 3022 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3023 3024 if (!need_retry) { 3025 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) { 3026 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 3027 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3028 free(port); 3029 } 3030 } 3031 } 3032 3033 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 3034 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 3035 SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n", 3036 port->trid->traddr, port->trid->trsvcid, need_retry); 3037 TAILQ_REMOVE(&rtransport->ports, port, link); 3038 rdma_destroy_id(port->id); 3039 port->id = NULL; 3040 port->device = NULL; 3041 if (need_retry) { 3042 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3043 } else { 3044 free(port); 3045 } 3046 break; 3047 } 3048 } 3049 } 3050 3051 static void 3052 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 3053 const struct spdk_nvme_transport_id *trid) 3054 { 3055 nvmf_rdma_stop_listen_ex(transport, trid, false); 3056 } 3057 3058 static void _nvmf_rdma_register_poller_in_group(void *c); 3059 static void _nvmf_rdma_remove_poller_in_group(void *c); 3060 3061 static bool 3062 nvmf_rdma_all_pollers_management_done(void *c) 3063 { 3064 struct poller_manage_ctx *ctx = c; 3065 int counter; 3066 3067 counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST); 3068 SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n", 3069 counter, ctx->rpoller); 3070 3071 if (counter == 0) { 3072 free((void *)ctx->inflight_op_counter); 3073 } 3074 free(ctx); 3075 3076 return counter == 0; 3077 } 3078 3079 static int 3080 nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport, 3081 struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add) 3082 { 3083 struct spdk_nvmf_rdma_poll_group *rgroup; 3084 struct spdk_nvmf_rdma_poller *rpoller; 3085 struct spdk_nvmf_poll_group *poll_group; 3086 struct poller_manage_ctx *ctx; 3087 bool found; 3088 int *inflight_counter; 3089 spdk_msg_fn do_fn; 3090 3091 *has_inflight = false; 3092 do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group; 3093 inflight_counter = calloc(1, sizeof(int)); 3094 if (!inflight_counter) { 3095 SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n"); 3096 return -ENOMEM; 3097 } 3098 3099 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3100 (*inflight_counter)++; 3101 } 3102 3103 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3104 found = false; 3105 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3106 if (rpoller->device == device) { 3107 found = true; 3108 break; 3109 } 3110 } 3111 if (found == is_add) { 3112 __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST); 3113 continue; 3114 } 3115 3116 ctx = calloc(1, sizeof(struct poller_manage_ctx)); 3117 if (!ctx) { 3118 SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n"); 3119 if (!*has_inflight) { 3120 free(inflight_counter); 3121 } 3122 return -ENOMEM; 3123 } 3124 3125 ctx->rtransport = rtransport; 3126 ctx->rgroup = rgroup; 3127 ctx->rpoller = rpoller; 3128 ctx->device = device; 3129 ctx->thread = spdk_get_thread(); 3130 ctx->inflight_op_counter = inflight_counter; 3131 *has_inflight = true; 3132 3133 poll_group = rgroup->group.group; 3134 if (poll_group->thread != spdk_get_thread()) { 3135 spdk_thread_send_msg(poll_group->thread, do_fn, ctx); 3136 } else { 3137 do_fn(ctx); 3138 } 3139 } 3140 3141 if (!*has_inflight) { 3142 free(inflight_counter); 3143 } 3144 3145 return 0; 3146 } 3147 3148 static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3149 struct spdk_nvmf_rdma_device *device); 3150 3151 static struct spdk_nvmf_rdma_device * 3152 nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 3153 struct ibv_context *context) 3154 { 3155 struct spdk_nvmf_rdma_device *device, *tmp_device; 3156 3157 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) { 3158 if (device->need_destroy) { 3159 continue; 3160 } 3161 3162 if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) { 3163 return device; 3164 } 3165 } 3166 3167 return NULL; 3168 } 3169 3170 static bool 3171 nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport, 3172 struct ibv_context *context) 3173 { 3174 struct spdk_nvmf_rdma_device *old_device, *new_device; 3175 int rc = 0; 3176 bool has_inflight; 3177 3178 old_device = nvmf_rdma_find_ib_device(rtransport, context); 3179 3180 if (old_device) { 3181 if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) { 3182 /* context may not have time to be cleaned when rescan. exactly one context 3183 * is valid for a device so this context must be invalid and just remove it. */ 3184 SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context); 3185 old_device->need_destroy = true; 3186 nvmf_rdma_handle_device_removal(rtransport, old_device); 3187 } 3188 return false; 3189 } 3190 3191 rc = create_ib_device(rtransport, context, &new_device); 3192 /* TODO: update transport opts. */ 3193 if (rc < 0) { 3194 SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n", 3195 ibv_get_device_name(context->device), context); 3196 return false; 3197 } 3198 3199 rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true); 3200 if (rc < 0) { 3201 SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n", 3202 ibv_get_device_name(context->device), context); 3203 return false; 3204 } 3205 3206 if (has_inflight) { 3207 new_device->is_ready = true; 3208 } 3209 3210 return true; 3211 } 3212 3213 static bool 3214 nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport) 3215 { 3216 struct spdk_nvmf_rdma_device *device; 3217 struct ibv_device **ibv_device_list = NULL; 3218 struct ibv_context **contexts = NULL; 3219 int i = 0; 3220 int num_dev = 0; 3221 bool new_create = false, has_new_device = false; 3222 struct ibv_context *tmp_verbs = NULL; 3223 3224 /* do not rescan when any device is destroying, or context may be freed when 3225 * regenerating the poll fds. 3226 */ 3227 TAILQ_FOREACH(device, &rtransport->devices, link) { 3228 if (device->need_destroy) { 3229 return false; 3230 } 3231 } 3232 3233 ibv_device_list = ibv_get_device_list(&num_dev); 3234 3235 /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be 3236 * marked as dead verbs and never be init again. So we need to make sure the 3237 * verbs is available before we call rdma_get_devices. */ 3238 if (num_dev >= 0) { 3239 for (i = 0; i < num_dev; i++) { 3240 tmp_verbs = ibv_open_device(ibv_device_list[i]); 3241 if (!tmp_verbs) { 3242 SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno); 3243 break; 3244 } 3245 if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) { 3246 SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i], 3247 tmp_verbs->device->dev_name); 3248 has_new_device = true; 3249 } 3250 ibv_close_device(tmp_verbs); 3251 } 3252 ibv_free_device_list(ibv_device_list); 3253 if (!tmp_verbs || !has_new_device) { 3254 return false; 3255 } 3256 } 3257 3258 contexts = rdma_get_devices(NULL); 3259 3260 for (i = 0; contexts && contexts[i] != NULL; i++) { 3261 new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]); 3262 } 3263 3264 if (new_create) { 3265 free_poll_fds(rtransport); 3266 generate_poll_fds(rtransport); 3267 } 3268 3269 if (contexts) { 3270 rdma_free_devices(contexts); 3271 } 3272 3273 return new_create; 3274 } 3275 3276 static bool 3277 nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport) 3278 { 3279 struct spdk_nvmf_rdma_port *port, *tmp_port; 3280 int rc = 0; 3281 bool new_create = false; 3282 3283 if (TAILQ_EMPTY(&rtransport->retry_ports)) { 3284 return false; 3285 } 3286 3287 new_create = nvmf_rdma_rescan_devices(rtransport); 3288 3289 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) { 3290 rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL); 3291 3292 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3293 if (rc) { 3294 if (new_create) { 3295 SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n", 3296 port->trid->traddr, port->trid->trsvcid, rc); 3297 } 3298 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3299 break; 3300 } else { 3301 SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid); 3302 free(port); 3303 } 3304 } 3305 3306 return true; 3307 } 3308 3309 static void 3310 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 3311 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 3312 { 3313 struct spdk_nvmf_request *req, *tmp; 3314 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 3315 struct spdk_nvmf_rdma_resources *resources; 3316 3317 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 3318 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 3319 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3320 break; 3321 } 3322 } 3323 3324 /* Then RDMA writes since reads have stronger restrictions than writes */ 3325 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 3326 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3327 break; 3328 } 3329 } 3330 3331 /* Then we handle request waiting on memory buffers. */ 3332 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 3333 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3334 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3335 break; 3336 } 3337 } 3338 3339 resources = rqpair->resources; 3340 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 3341 rdma_req = STAILQ_FIRST(&resources->free_queue); 3342 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 3343 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 3344 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 3345 3346 if (rqpair->srq != NULL) { 3347 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 3348 rdma_req->recv->qpair->qd++; 3349 } else { 3350 rqpair->qd++; 3351 } 3352 3353 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 3354 rdma_req->state = RDMA_REQUEST_STATE_NEW; 3355 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 3356 break; 3357 } 3358 } 3359 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 3360 rqpair->poller->stat.pending_free_request++; 3361 } 3362 } 3363 3364 static inline bool 3365 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 3366 { 3367 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 3368 return nvmf_rdma_is_rxe_device(device) || 3369 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 3370 } 3371 3372 static void 3373 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 3374 { 3375 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3376 struct spdk_nvmf_rdma_transport, transport); 3377 3378 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3379 3380 /* nvmf_rdma_close_qpair is not called */ 3381 if (!rqpair->to_close) { 3382 return; 3383 } 3384 3385 /* device is already destroyed and we should force destroy this qpair. */ 3386 if (rqpair->poller && rqpair->poller->need_destroy) { 3387 nvmf_rdma_qpair_destroy(rqpair); 3388 return; 3389 } 3390 3391 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 3392 if (rqpair->current_send_depth != 0) { 3393 return; 3394 } 3395 3396 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 3397 return; 3398 } 3399 3400 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 3401 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 3402 return; 3403 } 3404 3405 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 3406 3407 nvmf_rdma_qpair_destroy(rqpair); 3408 } 3409 3410 static int 3411 nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked) 3412 { 3413 struct spdk_nvmf_qpair *qpair; 3414 struct spdk_nvmf_rdma_qpair *rqpair; 3415 3416 if (evt->id == NULL) { 3417 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 3418 return -1; 3419 } 3420 3421 qpair = evt->id->context; 3422 if (qpair == NULL) { 3423 SPDK_ERRLOG("disconnect request: no active connection\n"); 3424 return -1; 3425 } 3426 3427 rdma_ack_cm_event(evt); 3428 *event_acked = true; 3429 3430 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3431 3432 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair); 3433 3434 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3435 3436 return 0; 3437 } 3438 3439 #ifdef DEBUG 3440 static const char *CM_EVENT_STR[] = { 3441 "RDMA_CM_EVENT_ADDR_RESOLVED", 3442 "RDMA_CM_EVENT_ADDR_ERROR", 3443 "RDMA_CM_EVENT_ROUTE_RESOLVED", 3444 "RDMA_CM_EVENT_ROUTE_ERROR", 3445 "RDMA_CM_EVENT_CONNECT_REQUEST", 3446 "RDMA_CM_EVENT_CONNECT_RESPONSE", 3447 "RDMA_CM_EVENT_CONNECT_ERROR", 3448 "RDMA_CM_EVENT_UNREACHABLE", 3449 "RDMA_CM_EVENT_REJECTED", 3450 "RDMA_CM_EVENT_ESTABLISHED", 3451 "RDMA_CM_EVENT_DISCONNECTED", 3452 "RDMA_CM_EVENT_DEVICE_REMOVAL", 3453 "RDMA_CM_EVENT_MULTICAST_JOIN", 3454 "RDMA_CM_EVENT_MULTICAST_ERROR", 3455 "RDMA_CM_EVENT_ADDR_CHANGE", 3456 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 3457 }; 3458 #endif /* DEBUG */ 3459 3460 static void 3461 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 3462 struct spdk_nvmf_rdma_port *port) 3463 { 3464 struct spdk_nvmf_rdma_poll_group *rgroup; 3465 struct spdk_nvmf_rdma_poller *rpoller; 3466 struct spdk_nvmf_rdma_qpair *rqpair; 3467 3468 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3469 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3470 RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) { 3471 if (rqpair->listen_id == port->id) { 3472 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3473 } 3474 } 3475 } 3476 } 3477 } 3478 3479 static bool 3480 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 3481 struct rdma_cm_event *event) 3482 { 3483 const struct spdk_nvme_transport_id *trid; 3484 struct spdk_nvmf_rdma_port *port; 3485 struct spdk_nvmf_rdma_transport *rtransport; 3486 bool event_acked = false; 3487 3488 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3489 TAILQ_FOREACH(port, &rtransport->ports, link) { 3490 if (port->id == event->id) { 3491 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 3492 rdma_ack_cm_event(event); 3493 event_acked = true; 3494 trid = port->trid; 3495 break; 3496 } 3497 } 3498 3499 if (event_acked) { 3500 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 3501 3502 nvmf_rdma_stop_listen(transport, trid); 3503 nvmf_rdma_listen(transport, trid, NULL); 3504 } 3505 3506 return event_acked; 3507 } 3508 3509 static void 3510 nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3511 struct spdk_nvmf_rdma_device *device) 3512 { 3513 struct spdk_nvmf_rdma_port *port, *port_tmp; 3514 int rc; 3515 bool has_inflight; 3516 3517 rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false); 3518 if (rc) { 3519 SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc); 3520 return; 3521 } 3522 3523 if (!has_inflight) { 3524 /* no pollers, destroy the device */ 3525 device->ready_to_destroy = true; 3526 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport); 3527 } 3528 3529 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 3530 if (port->device == device) { 3531 SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n", 3532 port->trid->traddr, 3533 port->trid->trsvcid, 3534 ibv_get_device_name(port->device->context->device)); 3535 3536 /* keep NVMF listener and only destroy structures of the 3537 * RDMA transport. when the device comes back we can retry listening 3538 * and the application's workflow will not be interrupted. 3539 */ 3540 nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true); 3541 } 3542 } 3543 } 3544 3545 static void 3546 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 3547 struct rdma_cm_event *event) 3548 { 3549 struct spdk_nvmf_rdma_port *port, *tmp_port; 3550 struct spdk_nvmf_rdma_transport *rtransport; 3551 3552 port = event->id->context; 3553 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3554 3555 rdma_ack_cm_event(event); 3556 3557 /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive 3558 * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that 3559 * we are handling a port event here. 3560 */ 3561 TAILQ_FOREACH(tmp_port, &rtransport->ports, link) { 3562 if (port == tmp_port && port->device && !port->device->need_destroy) { 3563 port->device->need_destroy = true; 3564 nvmf_rdma_handle_device_removal(rtransport, port->device); 3565 } 3566 } 3567 } 3568 3569 static void 3570 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 3571 { 3572 struct spdk_nvmf_rdma_transport *rtransport; 3573 struct rdma_cm_event *event; 3574 int rc; 3575 bool event_acked; 3576 3577 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3578 3579 if (rtransport->event_channel == NULL) { 3580 return; 3581 } 3582 3583 while (1) { 3584 event_acked = false; 3585 rc = rdma_get_cm_event(rtransport->event_channel, &event); 3586 if (rc) { 3587 if (errno != EAGAIN && errno != EWOULDBLOCK) { 3588 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 3589 } 3590 break; 3591 } 3592 3593 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 3594 3595 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 3596 3597 switch (event->event) { 3598 case RDMA_CM_EVENT_ADDR_RESOLVED: 3599 case RDMA_CM_EVENT_ADDR_ERROR: 3600 case RDMA_CM_EVENT_ROUTE_RESOLVED: 3601 case RDMA_CM_EVENT_ROUTE_ERROR: 3602 /* No action required. The target never attempts to resolve routes. */ 3603 break; 3604 case RDMA_CM_EVENT_CONNECT_REQUEST: 3605 rc = nvmf_rdma_connect(transport, event); 3606 if (rc < 0) { 3607 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 3608 break; 3609 } 3610 break; 3611 case RDMA_CM_EVENT_CONNECT_RESPONSE: 3612 /* The target never initiates a new connection. So this will not occur. */ 3613 break; 3614 case RDMA_CM_EVENT_CONNECT_ERROR: 3615 /* Can this happen? The docs say it can, but not sure what causes it. */ 3616 break; 3617 case RDMA_CM_EVENT_UNREACHABLE: 3618 case RDMA_CM_EVENT_REJECTED: 3619 /* These only occur on the client side. */ 3620 break; 3621 case RDMA_CM_EVENT_ESTABLISHED: 3622 /* TODO: Should we be waiting for this event anywhere? */ 3623 break; 3624 case RDMA_CM_EVENT_DISCONNECTED: 3625 rc = nvmf_rdma_disconnect(event, &event_acked); 3626 if (rc < 0) { 3627 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3628 break; 3629 } 3630 break; 3631 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3632 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3633 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3634 * Once these events are sent to SPDK, we should release all IB resources and 3635 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3636 * ibv_destroy* functions to release user space memory allocated by IB. All kernel 3637 * resources are already cleaned. */ 3638 if (event->id->qp) { 3639 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3640 * corresponding qpair. Otherwise the event refers to a listening device. */ 3641 rc = nvmf_rdma_disconnect(event, &event_acked); 3642 if (rc < 0) { 3643 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3644 break; 3645 } 3646 } else { 3647 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3648 event_acked = true; 3649 } 3650 break; 3651 case RDMA_CM_EVENT_MULTICAST_JOIN: 3652 case RDMA_CM_EVENT_MULTICAST_ERROR: 3653 /* Multicast is not used */ 3654 break; 3655 case RDMA_CM_EVENT_ADDR_CHANGE: 3656 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3657 break; 3658 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3659 /* For now, do nothing. The target never re-uses queue pairs. */ 3660 break; 3661 default: 3662 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3663 break; 3664 } 3665 if (!event_acked) { 3666 rdma_ack_cm_event(event); 3667 } 3668 } 3669 } 3670 3671 static void 3672 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3673 { 3674 rqpair->last_wqe_reached = true; 3675 nvmf_rdma_destroy_drained_qpair(rqpair); 3676 } 3677 3678 static void 3679 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3680 { 3681 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3682 3683 if (event_ctx->rqpair) { 3684 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3685 if (event_ctx->cb_fn) { 3686 event_ctx->cb_fn(event_ctx->rqpair); 3687 } 3688 } 3689 free(event_ctx); 3690 } 3691 3692 static int 3693 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3694 spdk_nvmf_rdma_qpair_ibv_event fn) 3695 { 3696 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3697 struct spdk_thread *thr = NULL; 3698 int rc; 3699 3700 if (rqpair->qpair.group) { 3701 thr = rqpair->qpair.group->thread; 3702 } else if (rqpair->destruct_channel) { 3703 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3704 } 3705 3706 if (!thr) { 3707 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3708 return -EINVAL; 3709 } 3710 3711 ctx = calloc(1, sizeof(*ctx)); 3712 if (!ctx) { 3713 return -ENOMEM; 3714 } 3715 3716 ctx->rqpair = rqpair; 3717 ctx->cb_fn = fn; 3718 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3719 3720 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3721 if (rc) { 3722 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3723 free(ctx); 3724 } 3725 3726 return rc; 3727 } 3728 3729 static int 3730 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3731 { 3732 int rc; 3733 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3734 struct ibv_async_event event; 3735 3736 rc = ibv_get_async_event(device->context, &event); 3737 3738 if (rc) { 3739 /* In non-blocking mode -1 means there are no events available */ 3740 return rc; 3741 } 3742 3743 switch (event.event_type) { 3744 case IBV_EVENT_QP_FATAL: 3745 case IBV_EVENT_QP_LAST_WQE_REACHED: 3746 case IBV_EVENT_SQ_DRAINED: 3747 case IBV_EVENT_QP_REQ_ERR: 3748 case IBV_EVENT_QP_ACCESS_ERR: 3749 case IBV_EVENT_COMM_EST: 3750 case IBV_EVENT_PATH_MIG: 3751 case IBV_EVENT_PATH_MIG_ERR: 3752 rqpair = event.element.qp->qp_context; 3753 if (!rqpair) { 3754 /* Any QP event for NVMe-RDMA initiator may be returned. */ 3755 SPDK_NOTICELOG("Async QP event for unknown QP: %s\n", 3756 ibv_event_type_str(event.event_type)); 3757 break; 3758 } 3759 3760 switch (event.event_type) { 3761 case IBV_EVENT_QP_FATAL: 3762 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3763 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3764 (uintptr_t)rqpair, event.event_type); 3765 nvmf_rdma_update_ibv_state(rqpair); 3766 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3767 break; 3768 case IBV_EVENT_QP_LAST_WQE_REACHED: 3769 /* This event only occurs for shared receive queues. */ 3770 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3771 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3772 if (rc) { 3773 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3774 rqpair->last_wqe_reached = true; 3775 } 3776 break; 3777 case IBV_EVENT_SQ_DRAINED: 3778 /* This event occurs frequently in both error and non-error states. 3779 * Check if the qpair is in an error state before sending a message. */ 3780 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3781 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3782 (uintptr_t)rqpair, event.event_type); 3783 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3784 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3785 } 3786 break; 3787 case IBV_EVENT_QP_REQ_ERR: 3788 case IBV_EVENT_QP_ACCESS_ERR: 3789 case IBV_EVENT_COMM_EST: 3790 case IBV_EVENT_PATH_MIG: 3791 case IBV_EVENT_PATH_MIG_ERR: 3792 SPDK_NOTICELOG("Async QP event: %s\n", 3793 ibv_event_type_str(event.event_type)); 3794 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3795 (uintptr_t)rqpair, event.event_type); 3796 nvmf_rdma_update_ibv_state(rqpair); 3797 break; 3798 default: 3799 break; 3800 } 3801 break; 3802 case IBV_EVENT_DEVICE_FATAL: 3803 SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n", 3804 ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device); 3805 device->need_destroy = true; 3806 break; 3807 case IBV_EVENT_CQ_ERR: 3808 case IBV_EVENT_PORT_ACTIVE: 3809 case IBV_EVENT_PORT_ERR: 3810 case IBV_EVENT_LID_CHANGE: 3811 case IBV_EVENT_PKEY_CHANGE: 3812 case IBV_EVENT_SM_CHANGE: 3813 case IBV_EVENT_SRQ_ERR: 3814 case IBV_EVENT_SRQ_LIMIT_REACHED: 3815 case IBV_EVENT_CLIENT_REREGISTER: 3816 case IBV_EVENT_GID_CHANGE: 3817 default: 3818 SPDK_NOTICELOG("Async event: %s\n", 3819 ibv_event_type_str(event.event_type)); 3820 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3821 break; 3822 } 3823 ibv_ack_async_event(&event); 3824 3825 return 0; 3826 } 3827 3828 static void 3829 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3830 { 3831 int rc = 0; 3832 uint32_t i = 0; 3833 3834 for (i = 0; i < max_events; i++) { 3835 rc = nvmf_process_ib_event(device); 3836 if (rc) { 3837 break; 3838 } 3839 } 3840 3841 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3842 } 3843 3844 static int 3845 nvmf_rdma_accept(void *ctx) 3846 { 3847 int nfds, i = 0; 3848 struct spdk_nvmf_transport *transport = ctx; 3849 struct spdk_nvmf_rdma_transport *rtransport; 3850 struct spdk_nvmf_rdma_device *device, *tmp; 3851 uint32_t count; 3852 short revents; 3853 bool do_retry; 3854 3855 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3856 do_retry = nvmf_rdma_retry_listen_port(rtransport); 3857 3858 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3859 3860 if (nfds <= 0) { 3861 return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3862 } 3863 3864 /* The first poll descriptor is RDMA CM event */ 3865 if (rtransport->poll_fds[i++].revents & POLLIN) { 3866 nvmf_process_cm_event(transport); 3867 nfds--; 3868 } 3869 3870 if (nfds == 0) { 3871 return SPDK_POLLER_BUSY; 3872 } 3873 3874 /* Second and subsequent poll descriptors are IB async events */ 3875 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3876 revents = rtransport->poll_fds[i++].revents; 3877 if (revents & POLLIN) { 3878 if (spdk_likely(!device->need_destroy)) { 3879 nvmf_process_ib_events(device, 32); 3880 if (spdk_unlikely(device->need_destroy)) { 3881 nvmf_rdma_handle_device_removal(rtransport, device); 3882 } 3883 } 3884 nfds--; 3885 } else if (revents & POLLNVAL || revents & POLLHUP) { 3886 SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device); 3887 nfds--; 3888 } 3889 } 3890 /* check all flagged fd's have been served */ 3891 assert(nfds == 0); 3892 3893 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3894 } 3895 3896 static void 3897 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3898 struct spdk_nvmf_ctrlr_data *cdata) 3899 { 3900 cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD; 3901 3902 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3903 since in-capsule data only works with NVME drives that support SGL memory layout */ 3904 if (transport->opts.dif_insert_or_strip) { 3905 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3906 } 3907 3908 if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) { 3909 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule" 3910 " data is greater than 4KiB.\n"); 3911 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux " 3912 "kernel between versions 5.4 and 5.12 data corruption may occur for " 3913 "writes that are not a multiple of 4KiB in size.\n"); 3914 } 3915 } 3916 3917 static void 3918 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3919 struct spdk_nvme_transport_id *trid, 3920 struct spdk_nvmf_discovery_log_page_entry *entry) 3921 { 3922 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3923 entry->adrfam = trid->adrfam; 3924 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3925 3926 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3927 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3928 3929 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3930 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3931 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3932 } 3933 3934 static int 3935 nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport, 3936 struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device, 3937 struct spdk_nvmf_rdma_poller **out_poller) 3938 { 3939 struct spdk_nvmf_rdma_poller *poller; 3940 struct spdk_rdma_srq_init_attr srq_init_attr; 3941 struct spdk_nvmf_rdma_resource_opts opts; 3942 int num_cqe; 3943 3944 poller = calloc(1, sizeof(*poller)); 3945 if (!poller) { 3946 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3947 return -1; 3948 } 3949 3950 poller->device = device; 3951 poller->group = rgroup; 3952 *out_poller = poller; 3953 3954 RB_INIT(&poller->qpairs); 3955 STAILQ_INIT(&poller->qpairs_pending_send); 3956 STAILQ_INIT(&poller->qpairs_pending_recv); 3957 3958 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3959 SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup); 3960 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3961 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) { 3962 SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n", 3963 rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr); 3964 } 3965 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr); 3966 3967 device->num_srq++; 3968 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 3969 srq_init_attr.pd = device->pd; 3970 srq_init_attr.stats = &poller->stat.qp_stats.recv; 3971 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 3972 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3973 poller->srq = spdk_rdma_srq_create(&srq_init_attr); 3974 if (!poller->srq) { 3975 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3976 return -1; 3977 } 3978 3979 opts.qp = poller->srq; 3980 opts.map = device->map; 3981 opts.qpair = NULL; 3982 opts.shared = true; 3983 opts.max_queue_depth = poller->max_srq_depth; 3984 opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size; 3985 3986 poller->resources = nvmf_rdma_resources_create(&opts); 3987 if (!poller->resources) { 3988 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3989 return -1; 3990 } 3991 } 3992 3993 /* 3994 * When using an srq, we can limit the completion queue at startup. 3995 * The following formula represents the calculation: 3996 * num_cqe = num_recv + num_data_wr + num_send_wr. 3997 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3998 */ 3999 if (poller->srq) { 4000 num_cqe = poller->max_srq_depth * 3; 4001 } else { 4002 num_cqe = rtransport->rdma_opts.num_cqe; 4003 } 4004 4005 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 4006 if (!poller->cq) { 4007 SPDK_ERRLOG("Unable to create completion queue\n"); 4008 return -1; 4009 } 4010 poller->num_cqe = num_cqe; 4011 return 0; 4012 } 4013 4014 static void 4015 _nvmf_rdma_register_poller_in_group(void *c) 4016 { 4017 struct spdk_nvmf_rdma_poller *poller; 4018 struct poller_manage_ctx *ctx = c; 4019 struct spdk_nvmf_rdma_device *device; 4020 int rc; 4021 4022 rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller); 4023 if (rc < 0 && poller) { 4024 nvmf_rdma_poller_destroy(poller); 4025 } 4026 4027 device = ctx->device; 4028 if (nvmf_rdma_all_pollers_management_done(ctx)) { 4029 device->is_ready = true; 4030 } 4031 } 4032 4033 static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 4034 4035 static struct spdk_nvmf_transport_poll_group * 4036 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport, 4037 struct spdk_nvmf_poll_group *group) 4038 { 4039 struct spdk_nvmf_rdma_transport *rtransport; 4040 struct spdk_nvmf_rdma_poll_group *rgroup; 4041 struct spdk_nvmf_rdma_poller *poller; 4042 struct spdk_nvmf_rdma_device *device; 4043 int rc; 4044 4045 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 4046 4047 rgroup = calloc(1, sizeof(*rgroup)); 4048 if (!rgroup) { 4049 return NULL; 4050 } 4051 4052 TAILQ_INIT(&rgroup->pollers); 4053 4054 TAILQ_FOREACH(device, &rtransport->devices, link) { 4055 rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller); 4056 if (rc < 0) { 4057 nvmf_rdma_poll_group_destroy(&rgroup->group); 4058 return NULL; 4059 } 4060 } 4061 4062 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 4063 if (rtransport->conn_sched.next_admin_pg == NULL) { 4064 rtransport->conn_sched.next_admin_pg = rgroup; 4065 rtransport->conn_sched.next_io_pg = rgroup; 4066 } 4067 4068 return &rgroup->group; 4069 } 4070 4071 static uint32_t 4072 nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg) 4073 { 4074 uint32_t count; 4075 4076 /* Just assume that unassociated qpairs will eventually be io 4077 * qpairs. This is close enough for the use cases for this 4078 * function. 4079 */ 4080 pthread_mutex_lock(&pg->mutex); 4081 count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs; 4082 pthread_mutex_unlock(&pg->mutex); 4083 4084 return count; 4085 } 4086 4087 static struct spdk_nvmf_transport_poll_group * 4088 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 4089 { 4090 struct spdk_nvmf_rdma_transport *rtransport; 4091 struct spdk_nvmf_rdma_poll_group **pg; 4092 struct spdk_nvmf_transport_poll_group *result; 4093 uint32_t count; 4094 4095 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4096 4097 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 4098 return NULL; 4099 } 4100 4101 if (qpair->qid == 0) { 4102 pg = &rtransport->conn_sched.next_admin_pg; 4103 } else { 4104 struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current; 4105 uint32_t min_value; 4106 4107 pg = &rtransport->conn_sched.next_io_pg; 4108 pg_min = *pg; 4109 pg_start = *pg; 4110 pg_current = *pg; 4111 min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 4112 4113 while ((count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group)) > 0) { 4114 pg_current = TAILQ_NEXT(pg_current, link); 4115 if (pg_current == NULL) { 4116 pg_current = TAILQ_FIRST(&rtransport->poll_groups); 4117 } 4118 4119 if (count < min_value) { 4120 min_value = count; 4121 pg_min = pg_current; 4122 } 4123 4124 if (pg_current == pg_start) { 4125 break; 4126 } 4127 } 4128 *pg = pg_min; 4129 } 4130 4131 assert(*pg != NULL); 4132 4133 result = &(*pg)->group; 4134 4135 *pg = TAILQ_NEXT(*pg, link); 4136 if (*pg == NULL) { 4137 *pg = TAILQ_FIRST(&rtransport->poll_groups); 4138 } 4139 4140 return result; 4141 } 4142 4143 static void 4144 nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller) 4145 { 4146 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 4147 int rc; 4148 4149 TAILQ_REMOVE(&poller->group->pollers, poller, link); 4150 RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) { 4151 nvmf_rdma_qpair_destroy(qpair); 4152 } 4153 4154 if (poller->srq) { 4155 if (poller->resources) { 4156 nvmf_rdma_resources_destroy(poller->resources); 4157 } 4158 spdk_rdma_srq_destroy(poller->srq); 4159 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 4160 } 4161 4162 if (poller->cq) { 4163 rc = ibv_destroy_cq(poller->cq); 4164 if (rc != 0) { 4165 SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno)); 4166 } 4167 } 4168 4169 if (poller->destroy_cb) { 4170 poller->destroy_cb(poller->destroy_cb_ctx); 4171 poller->destroy_cb = NULL; 4172 } 4173 4174 free(poller); 4175 } 4176 4177 static void 4178 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 4179 { 4180 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 4181 struct spdk_nvmf_rdma_poller *poller, *tmp; 4182 struct spdk_nvmf_rdma_transport *rtransport; 4183 4184 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4185 if (!rgroup) { 4186 return; 4187 } 4188 4189 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 4190 nvmf_rdma_poller_destroy(poller); 4191 } 4192 4193 if (rgroup->group.transport == NULL) { 4194 /* Transport can be NULL when nvmf_rdma_poll_group_create() 4195 * calls this function directly in a failure path. */ 4196 free(rgroup); 4197 return; 4198 } 4199 4200 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 4201 4202 next_rgroup = TAILQ_NEXT(rgroup, link); 4203 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 4204 if (next_rgroup == NULL) { 4205 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 4206 } 4207 if (rtransport->conn_sched.next_admin_pg == rgroup) { 4208 rtransport->conn_sched.next_admin_pg = next_rgroup; 4209 } 4210 if (rtransport->conn_sched.next_io_pg == rgroup) { 4211 rtransport->conn_sched.next_io_pg = next_rgroup; 4212 } 4213 4214 free(rgroup); 4215 } 4216 4217 static void 4218 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 4219 { 4220 if (rqpair->cm_id != NULL) { 4221 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 4222 } 4223 } 4224 4225 static int 4226 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 4227 struct spdk_nvmf_qpair *qpair) 4228 { 4229 struct spdk_nvmf_rdma_poll_group *rgroup; 4230 struct spdk_nvmf_rdma_qpair *rqpair; 4231 struct spdk_nvmf_rdma_device *device; 4232 struct spdk_nvmf_rdma_poller *poller; 4233 int rc; 4234 4235 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4236 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4237 4238 device = rqpair->device; 4239 4240 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 4241 if (poller->device == device) { 4242 break; 4243 } 4244 } 4245 4246 if (!poller) { 4247 SPDK_ERRLOG("No poller found for device.\n"); 4248 return -1; 4249 } 4250 4251 if (poller->need_destroy) { 4252 SPDK_ERRLOG("Poller is destroying.\n"); 4253 return -1; 4254 } 4255 4256 rqpair->poller = poller; 4257 rqpair->srq = rqpair->poller->srq; 4258 4259 rc = nvmf_rdma_qpair_initialize(qpair); 4260 if (rc < 0) { 4261 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 4262 rqpair->poller = NULL; 4263 rqpair->srq = NULL; 4264 return -1; 4265 } 4266 4267 RB_INSERT(qpairs_tree, &poller->qpairs, rqpair); 4268 4269 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 4270 if (rc) { 4271 /* Try to reject, but we probably can't */ 4272 nvmf_rdma_qpair_reject_connection(rqpair); 4273 return -1; 4274 } 4275 4276 nvmf_rdma_update_ibv_state(rqpair); 4277 4278 return 0; 4279 } 4280 4281 static int 4282 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 4283 struct spdk_nvmf_qpair *qpair) 4284 { 4285 struct spdk_nvmf_rdma_qpair *rqpair; 4286 4287 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4288 assert(group->transport->tgt != NULL); 4289 4290 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 4291 4292 if (!rqpair->destruct_channel) { 4293 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 4294 return 0; 4295 } 4296 4297 /* Sanity check that we get io_channel on the correct thread */ 4298 if (qpair->group) { 4299 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 4300 } 4301 4302 return 0; 4303 } 4304 4305 static int 4306 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 4307 { 4308 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 4309 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4310 struct spdk_nvmf_rdma_transport, transport); 4311 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4312 struct spdk_nvmf_rdma_qpair, qpair); 4313 4314 /* 4315 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 4316 * needs to be returned to the shared receive queue or the poll group will eventually be 4317 * starved of RECV structures. 4318 */ 4319 if (rqpair->srq && rdma_req->recv) { 4320 int rc; 4321 struct ibv_recv_wr *bad_recv_wr; 4322 4323 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 4324 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 4325 if (rc) { 4326 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 4327 } 4328 } 4329 4330 _nvmf_rdma_request_free(rdma_req, rtransport); 4331 return 0; 4332 } 4333 4334 static int 4335 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 4336 { 4337 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4338 struct spdk_nvmf_rdma_transport, transport); 4339 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 4340 struct spdk_nvmf_rdma_request, req); 4341 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4342 struct spdk_nvmf_rdma_qpair, qpair); 4343 4344 if (rqpair->ibv_state != IBV_QPS_ERR) { 4345 /* The connection is alive, so process the request as normal */ 4346 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 4347 } else { 4348 /* The connection is dead. Move the request directly to the completed state. */ 4349 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4350 } 4351 4352 nvmf_rdma_request_process(rtransport, rdma_req); 4353 4354 return 0; 4355 } 4356 4357 static void 4358 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 4359 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 4360 { 4361 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4362 4363 rqpair->to_close = true; 4364 4365 /* This happens only when the qpair is disconnected before 4366 * it is added to the poll group. Since there is no poll group, 4367 * the RDMA qp has not been initialized yet and the RDMA CM 4368 * event has not yet been acknowledged, so we need to reject it. 4369 */ 4370 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 4371 nvmf_rdma_qpair_reject_connection(rqpair); 4372 nvmf_rdma_qpair_destroy(rqpair); 4373 return; 4374 } 4375 4376 if (rqpair->rdma_qp) { 4377 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 4378 } 4379 4380 nvmf_rdma_destroy_drained_qpair(rqpair); 4381 4382 if (cb_fn) { 4383 cb_fn(cb_arg); 4384 } 4385 } 4386 4387 static struct spdk_nvmf_rdma_qpair * 4388 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 4389 { 4390 struct spdk_nvmf_rdma_qpair find; 4391 4392 find.qp_num = wc->qp_num; 4393 4394 return RB_FIND(qpairs_tree, &rpoller->qpairs, &find); 4395 } 4396 4397 #ifdef DEBUG 4398 static int 4399 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 4400 { 4401 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 4402 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 4403 } 4404 #endif 4405 4406 static void 4407 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 4408 int rc) 4409 { 4410 struct spdk_nvmf_rdma_recv *rdma_recv; 4411 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4412 4413 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 4414 while (bad_recv_wr != NULL) { 4415 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 4416 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4417 4418 rdma_recv->qpair->current_recv_depth++; 4419 bad_recv_wr = bad_recv_wr->next; 4420 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 4421 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 4422 } 4423 } 4424 4425 static void 4426 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 4427 { 4428 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 4429 while (bad_recv_wr != NULL) { 4430 bad_recv_wr = bad_recv_wr->next; 4431 rqpair->current_recv_depth++; 4432 } 4433 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4434 } 4435 4436 static void 4437 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 4438 struct spdk_nvmf_rdma_poller *rpoller) 4439 { 4440 struct spdk_nvmf_rdma_qpair *rqpair; 4441 struct ibv_recv_wr *bad_recv_wr; 4442 int rc; 4443 4444 if (rpoller->srq) { 4445 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 4446 if (rc) { 4447 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 4448 } 4449 } else { 4450 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 4451 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 4452 rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 4453 if (rc) { 4454 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 4455 } 4456 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 4457 } 4458 } 4459 } 4460 4461 static void 4462 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 4463 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 4464 { 4465 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4466 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 4467 4468 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 4469 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 4470 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 4471 assert(rqpair->current_send_depth > 0); 4472 rqpair->current_send_depth--; 4473 switch (bad_rdma_wr->type) { 4474 case RDMA_WR_TYPE_DATA: 4475 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4476 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 4477 assert(rqpair->current_read_depth > 0); 4478 rqpair->current_read_depth--; 4479 } 4480 break; 4481 case RDMA_WR_TYPE_SEND: 4482 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4483 break; 4484 default: 4485 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 4486 prev_rdma_req = cur_rdma_req; 4487 continue; 4488 } 4489 4490 if (prev_rdma_req == cur_rdma_req) { 4491 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 4492 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 4493 continue; 4494 } 4495 4496 switch (cur_rdma_req->state) { 4497 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4498 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 4499 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4500 break; 4501 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 4502 case RDMA_REQUEST_STATE_COMPLETING: 4503 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4504 break; 4505 default: 4506 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 4507 cur_rdma_req->state, rqpair); 4508 continue; 4509 } 4510 4511 nvmf_rdma_request_process(rtransport, cur_rdma_req); 4512 prev_rdma_req = cur_rdma_req; 4513 } 4514 4515 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 4516 /* Disconnect the connection. */ 4517 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4518 } 4519 4520 } 4521 4522 static void 4523 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 4524 struct spdk_nvmf_rdma_poller *rpoller) 4525 { 4526 struct spdk_nvmf_rdma_qpair *rqpair; 4527 struct ibv_send_wr *bad_wr = NULL; 4528 int rc; 4529 4530 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 4531 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 4532 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 4533 4534 /* bad wr always points to the first wr that failed. */ 4535 if (rc) { 4536 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 4537 } 4538 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 4539 } 4540 } 4541 4542 static const char * 4543 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 4544 { 4545 switch (wr_type) { 4546 case RDMA_WR_TYPE_RECV: 4547 return "RECV"; 4548 case RDMA_WR_TYPE_SEND: 4549 return "SEND"; 4550 case RDMA_WR_TYPE_DATA: 4551 return "DATA"; 4552 default: 4553 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 4554 SPDK_UNREACHABLE(); 4555 } 4556 } 4557 4558 static inline void 4559 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 4560 { 4561 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 4562 4563 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 4564 /* If qpair is in ERR state, we will receive completions for all posted and not completed 4565 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 4566 SPDK_DEBUGLOG(rdma, 4567 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 4568 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 4569 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4570 } else { 4571 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 4572 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 4573 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4574 } 4575 } 4576 4577 static int 4578 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 4579 struct spdk_nvmf_rdma_poller *rpoller) 4580 { 4581 struct ibv_wc wc[32]; 4582 struct spdk_nvmf_rdma_wr *rdma_wr; 4583 struct spdk_nvmf_rdma_request *rdma_req; 4584 struct spdk_nvmf_rdma_recv *rdma_recv; 4585 struct spdk_nvmf_rdma_qpair *rqpair, *tmp_rqpair; 4586 int reaped, i; 4587 int count = 0; 4588 int rc; 4589 bool error = false; 4590 uint64_t poll_tsc = spdk_get_ticks(); 4591 4592 if (spdk_unlikely(rpoller->need_destroy)) { 4593 /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not 4594 * be called because we cannot poll anything from cq. So we call that here to force 4595 * destroy the qpair after to_close turning true. 4596 */ 4597 RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) { 4598 nvmf_rdma_destroy_drained_qpair(rqpair); 4599 } 4600 return 0; 4601 } 4602 4603 /* Poll for completing operations. */ 4604 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 4605 if (reaped < 0) { 4606 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 4607 errno, spdk_strerror(errno)); 4608 return -1; 4609 } else if (reaped == 0) { 4610 rpoller->stat.idle_polls++; 4611 } 4612 4613 rpoller->stat.polls++; 4614 rpoller->stat.completions += reaped; 4615 4616 for (i = 0; i < reaped; i++) { 4617 4618 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 4619 4620 switch (rdma_wr->type) { 4621 case RDMA_WR_TYPE_SEND: 4622 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4623 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4624 4625 if (!wc[i].status) { 4626 count++; 4627 assert(wc[i].opcode == IBV_WC_SEND); 4628 assert(nvmf_rdma_req_is_completing(rdma_req)); 4629 } 4630 4631 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4632 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 4633 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 4634 rdma_req->num_outstanding_data_wr = 0; 4635 4636 nvmf_rdma_request_process(rtransport, rdma_req); 4637 break; 4638 case RDMA_WR_TYPE_RECV: 4639 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 4640 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4641 if (rpoller->srq != NULL) { 4642 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 4643 /* It is possible that there are still some completions for destroyed QP 4644 * associated with SRQ. We just ignore these late completions and re-post 4645 * receive WRs back to SRQ. 4646 */ 4647 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 4648 struct ibv_recv_wr *bad_wr; 4649 4650 rdma_recv->wr.next = NULL; 4651 spdk_rdma_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 4652 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 4653 if (rc) { 4654 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 4655 } 4656 continue; 4657 } 4658 } 4659 rqpair = rdma_recv->qpair; 4660 4661 assert(rqpair != NULL); 4662 if (!wc[i].status) { 4663 assert(wc[i].opcode == IBV_WC_RECV); 4664 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 4665 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4666 break; 4667 } 4668 } 4669 4670 rdma_recv->wr.next = NULL; 4671 rqpair->current_recv_depth++; 4672 rdma_recv->receive_tsc = poll_tsc; 4673 rpoller->stat.requests++; 4674 STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link); 4675 break; 4676 case RDMA_WR_TYPE_DATA: 4677 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4678 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4679 4680 assert(rdma_req->num_outstanding_data_wr > 0); 4681 4682 rqpair->current_send_depth--; 4683 rdma_req->num_outstanding_data_wr--; 4684 if (!wc[i].status) { 4685 assert(wc[i].opcode == IBV_WC_RDMA_READ); 4686 rqpair->current_read_depth--; 4687 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 4688 if (rdma_req->num_outstanding_data_wr == 0) { 4689 if (spdk_unlikely(rdma_req->num_remaining_data_wr)) { 4690 /* Only part of RDMA_READ operations was submitted, process the rest */ 4691 rc = nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport); 4692 if (spdk_likely(!rc)) { 4693 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 4694 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 4695 } else { 4696 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4697 rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 4698 } 4699 nvmf_rdma_request_process(rtransport, rdma_req); 4700 break; 4701 } 4702 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 4703 nvmf_rdma_request_process(rtransport, rdma_req); 4704 } 4705 } else { 4706 /* If the data transfer fails still force the queue into the error state, 4707 * if we were performing an RDMA_READ, we need to force the request into a 4708 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 4709 * case, we should wait for the SEND to complete. */ 4710 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 4711 rqpair->current_read_depth--; 4712 if (rdma_req->num_outstanding_data_wr == 0) { 4713 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4714 } 4715 } 4716 } 4717 break; 4718 default: 4719 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 4720 continue; 4721 } 4722 4723 /* Handle error conditions */ 4724 if (wc[i].status) { 4725 nvmf_rdma_update_ibv_state(rqpair); 4726 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 4727 4728 error = true; 4729 4730 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 4731 /* Disconnect the connection. */ 4732 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4733 } else { 4734 nvmf_rdma_destroy_drained_qpair(rqpair); 4735 } 4736 continue; 4737 } 4738 4739 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 4740 4741 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 4742 nvmf_rdma_destroy_drained_qpair(rqpair); 4743 } 4744 } 4745 4746 if (error == true) { 4747 return -1; 4748 } 4749 4750 /* submit outstanding work requests. */ 4751 _poller_submit_recvs(rtransport, rpoller); 4752 _poller_submit_sends(rtransport, rpoller); 4753 4754 return count; 4755 } 4756 4757 static void 4758 _nvmf_rdma_remove_destroyed_device(void *c) 4759 { 4760 struct spdk_nvmf_rdma_transport *rtransport = c; 4761 struct spdk_nvmf_rdma_device *device, *device_tmp; 4762 int rc; 4763 4764 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 4765 if (device->ready_to_destroy) { 4766 destroy_ib_device(rtransport, device); 4767 } 4768 } 4769 4770 free_poll_fds(rtransport); 4771 rc = generate_poll_fds(rtransport); 4772 /* cannot handle fd allocation error here */ 4773 if (rc != 0) { 4774 SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n"); 4775 } 4776 } 4777 4778 static void 4779 _nvmf_rdma_remove_poller_in_group_cb(void *c) 4780 { 4781 struct poller_manage_ctx *ctx = c; 4782 struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport; 4783 struct spdk_nvmf_rdma_device *device = ctx->device; 4784 struct spdk_thread *thread = ctx->thread; 4785 4786 if (nvmf_rdma_all_pollers_management_done(c)) { 4787 /* destroy device when last poller is destroyed */ 4788 device->ready_to_destroy = true; 4789 spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport); 4790 } 4791 } 4792 4793 static void 4794 _nvmf_rdma_remove_poller_in_group(void *c) 4795 { 4796 struct poller_manage_ctx *ctx = c; 4797 4798 ctx->rpoller->need_destroy = true; 4799 ctx->rpoller->destroy_cb_ctx = ctx; 4800 ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb; 4801 4802 /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */ 4803 if (RB_EMPTY(&ctx->rpoller->qpairs)) { 4804 nvmf_rdma_poller_destroy(ctx->rpoller); 4805 } 4806 } 4807 4808 static int 4809 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 4810 { 4811 struct spdk_nvmf_rdma_transport *rtransport; 4812 struct spdk_nvmf_rdma_poll_group *rgroup; 4813 struct spdk_nvmf_rdma_poller *rpoller, *tmp; 4814 int count, rc; 4815 4816 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 4817 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4818 4819 count = 0; 4820 TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) { 4821 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4822 if (rc < 0) { 4823 return rc; 4824 } 4825 count += rc; 4826 } 4827 4828 return count; 4829 } 4830 4831 static int 4832 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4833 struct spdk_nvme_transport_id *trid, 4834 bool peer) 4835 { 4836 struct sockaddr *saddr; 4837 uint16_t port; 4838 4839 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4840 4841 if (peer) { 4842 saddr = rdma_get_peer_addr(id); 4843 } else { 4844 saddr = rdma_get_local_addr(id); 4845 } 4846 switch (saddr->sa_family) { 4847 case AF_INET: { 4848 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4849 4850 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4851 inet_ntop(AF_INET, &saddr_in->sin_addr, 4852 trid->traddr, sizeof(trid->traddr)); 4853 if (peer) { 4854 port = ntohs(rdma_get_dst_port(id)); 4855 } else { 4856 port = ntohs(rdma_get_src_port(id)); 4857 } 4858 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4859 break; 4860 } 4861 case AF_INET6: { 4862 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4863 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4864 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4865 trid->traddr, sizeof(trid->traddr)); 4866 if (peer) { 4867 port = ntohs(rdma_get_dst_port(id)); 4868 } else { 4869 port = ntohs(rdma_get_src_port(id)); 4870 } 4871 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4872 break; 4873 } 4874 default: 4875 return -1; 4876 4877 } 4878 4879 return 0; 4880 } 4881 4882 static int 4883 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4884 struct spdk_nvme_transport_id *trid) 4885 { 4886 struct spdk_nvmf_rdma_qpair *rqpair; 4887 4888 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4889 4890 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4891 } 4892 4893 static int 4894 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4895 struct spdk_nvme_transport_id *trid) 4896 { 4897 struct spdk_nvmf_rdma_qpair *rqpair; 4898 4899 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4900 4901 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4902 } 4903 4904 static int 4905 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4906 struct spdk_nvme_transport_id *trid) 4907 { 4908 struct spdk_nvmf_rdma_qpair *rqpair; 4909 4910 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4911 4912 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4913 } 4914 4915 void 4916 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4917 { 4918 g_nvmf_hooks = *hooks; 4919 } 4920 4921 static void 4922 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4923 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4924 { 4925 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4926 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4927 4928 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4929 4930 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4931 } 4932 4933 static int 4934 _nvmf_rdma_qpair_abort_request(void *ctx) 4935 { 4936 struct spdk_nvmf_request *req = ctx; 4937 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4938 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4939 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4940 struct spdk_nvmf_rdma_qpair, qpair); 4941 int rc; 4942 4943 spdk_poller_unregister(&req->poller); 4944 4945 switch (rdma_req_to_abort->state) { 4946 case RDMA_REQUEST_STATE_EXECUTING: 4947 rc = nvmf_ctrlr_abort_request(req); 4948 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4949 return SPDK_POLLER_BUSY; 4950 } 4951 break; 4952 4953 case RDMA_REQUEST_STATE_NEED_BUFFER: 4954 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4955 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4956 4957 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4958 break; 4959 4960 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4961 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4962 spdk_nvmf_rdma_request, state_link); 4963 4964 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4965 break; 4966 4967 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4968 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4969 spdk_nvmf_rdma_request, state_link); 4970 4971 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4972 break; 4973 4974 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4975 if (spdk_get_ticks() < req->timeout_tsc) { 4976 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4977 return SPDK_POLLER_BUSY; 4978 } 4979 break; 4980 4981 default: 4982 break; 4983 } 4984 4985 spdk_nvmf_request_complete(req); 4986 return SPDK_POLLER_BUSY; 4987 } 4988 4989 static void 4990 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4991 struct spdk_nvmf_request *req) 4992 { 4993 struct spdk_nvmf_rdma_qpair *rqpair; 4994 struct spdk_nvmf_rdma_transport *rtransport; 4995 struct spdk_nvmf_transport *transport; 4996 uint16_t cid; 4997 uint32_t i, max_req_count; 4998 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 4999 5000 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 5001 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 5002 transport = &rtransport->transport; 5003 5004 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 5005 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 5006 5007 for (i = 0; i < max_req_count; i++) { 5008 rdma_req = &rqpair->resources->reqs[i]; 5009 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 5010 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 5011 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 5012 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 5013 rdma_req->req.qpair == qpair) { 5014 rdma_req_to_abort = rdma_req; 5015 break; 5016 } 5017 } 5018 5019 if (rdma_req_to_abort == NULL) { 5020 spdk_nvmf_request_complete(req); 5021 return; 5022 } 5023 5024 req->req_to_abort = &rdma_req_to_abort->req; 5025 req->timeout_tsc = spdk_get_ticks() + 5026 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 5027 req->poller = NULL; 5028 5029 _nvmf_rdma_qpair_abort_request(req); 5030 } 5031 5032 static void 5033 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 5034 struct spdk_json_write_ctx *w) 5035 { 5036 struct spdk_nvmf_rdma_poll_group *rgroup; 5037 struct spdk_nvmf_rdma_poller *rpoller; 5038 5039 assert(w != NULL); 5040 5041 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 5042 5043 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 5044 5045 spdk_json_write_named_array_begin(w, "devices"); 5046 5047 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 5048 spdk_json_write_object_begin(w); 5049 spdk_json_write_named_string(w, "name", 5050 ibv_get_device_name(rpoller->device->context->device)); 5051 spdk_json_write_named_uint64(w, "polls", 5052 rpoller->stat.polls); 5053 spdk_json_write_named_uint64(w, "idle_polls", 5054 rpoller->stat.idle_polls); 5055 spdk_json_write_named_uint64(w, "completions", 5056 rpoller->stat.completions); 5057 spdk_json_write_named_uint64(w, "requests", 5058 rpoller->stat.requests); 5059 spdk_json_write_named_uint64(w, "request_latency", 5060 rpoller->stat.request_latency); 5061 spdk_json_write_named_uint64(w, "pending_free_request", 5062 rpoller->stat.pending_free_request); 5063 spdk_json_write_named_uint64(w, "pending_rdma_read", 5064 rpoller->stat.pending_rdma_read); 5065 spdk_json_write_named_uint64(w, "pending_rdma_write", 5066 rpoller->stat.pending_rdma_write); 5067 spdk_json_write_named_uint64(w, "total_send_wrs", 5068 rpoller->stat.qp_stats.send.num_submitted_wrs); 5069 spdk_json_write_named_uint64(w, "send_doorbell_updates", 5070 rpoller->stat.qp_stats.send.doorbell_updates); 5071 spdk_json_write_named_uint64(w, "total_recv_wrs", 5072 rpoller->stat.qp_stats.recv.num_submitted_wrs); 5073 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 5074 rpoller->stat.qp_stats.recv.doorbell_updates); 5075 spdk_json_write_object_end(w); 5076 } 5077 5078 spdk_json_write_array_end(w); 5079 } 5080 5081 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 5082 .name = "RDMA", 5083 .type = SPDK_NVME_TRANSPORT_RDMA, 5084 .opts_init = nvmf_rdma_opts_init, 5085 .create = nvmf_rdma_create, 5086 .dump_opts = nvmf_rdma_dump_opts, 5087 .destroy = nvmf_rdma_destroy, 5088 5089 .listen = nvmf_rdma_listen, 5090 .stop_listen = nvmf_rdma_stop_listen, 5091 .cdata_init = nvmf_rdma_cdata_init, 5092 5093 .listener_discover = nvmf_rdma_discover, 5094 5095 .poll_group_create = nvmf_rdma_poll_group_create, 5096 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 5097 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 5098 .poll_group_add = nvmf_rdma_poll_group_add, 5099 .poll_group_remove = nvmf_rdma_poll_group_remove, 5100 .poll_group_poll = nvmf_rdma_poll_group_poll, 5101 5102 .req_free = nvmf_rdma_request_free, 5103 .req_complete = nvmf_rdma_request_complete, 5104 5105 .qpair_fini = nvmf_rdma_close_qpair, 5106 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 5107 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 5108 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 5109 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 5110 5111 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 5112 }; 5113 5114 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 5115 SPDK_LOG_REGISTER_COMPONENT(rdma) 5116