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) 2021-2024 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 #define NVMF_RDMA_MAX_EVENTS_PER_POLL 32 39 40 SPDK_STATIC_ASSERT(NVMF_DEFAULT_MSDBD <= SPDK_NVMF_MAX_SGL_ENTRIES, 41 "MSDBD must not exceed SPDK_NVMF_MAX_SGL_ENTRIES"); 42 43 /* The RDMA completion queue size */ 44 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 45 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 46 47 enum spdk_nvmf_rdma_request_state { 48 /* The request is not currently in use */ 49 RDMA_REQUEST_STATE_FREE = 0, 50 51 /* Initial state when request first received */ 52 RDMA_REQUEST_STATE_NEW, 53 54 /* The request is queued until a data buffer is available. */ 55 RDMA_REQUEST_STATE_NEED_BUFFER, 56 57 /* The request is waiting on RDMA queue depth availability 58 * to transfer data from the host to the controller. 59 */ 60 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 61 62 /* The request is currently transferring data from the host to the controller. */ 63 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 64 65 /* The request is ready to execute at the block device */ 66 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 67 68 /* The request is currently executing at the block device */ 69 RDMA_REQUEST_STATE_EXECUTING, 70 71 /* The request finished executing at the block device */ 72 RDMA_REQUEST_STATE_EXECUTED, 73 74 /* The request is waiting on RDMA queue depth availability 75 * to transfer data from the controller to the host. 76 */ 77 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 78 79 /* The request is waiting on RDMA queue depth availability 80 * to send response to the host. 81 */ 82 RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 83 84 /* The request is ready to send a completion */ 85 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 86 87 /* The request is currently transferring data from the controller to the host. */ 88 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 89 90 /* The request currently has an outstanding completion without an 91 * associated data transfer. 92 */ 93 RDMA_REQUEST_STATE_COMPLETING, 94 95 /* The request completed and can be marked free. */ 96 RDMA_REQUEST_STATE_COMPLETED, 97 98 /* Terminator */ 99 RDMA_REQUEST_NUM_STATES, 100 }; 101 102 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 103 { 104 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 105 106 struct spdk_trace_tpoint_opts opts[] = { 107 { 108 "RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 109 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 1, 110 { 111 { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 }, 112 { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 } 113 } 114 }, 115 { 116 "RDMA_REQ_COMPLETED", TRACE_RDMA_REQUEST_STATE_COMPLETED, 117 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 118 { 119 { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 }, 120 { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 } 121 } 122 }, 123 }; 124 125 spdk_trace_register_description_ext(opts, SPDK_COUNTOF(opts)); 126 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 127 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 128 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 129 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 130 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 131 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 132 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 133 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 134 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 135 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 136 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 137 spdk_trace_register_description("RDMA_REQ_TX_H2C", 138 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 139 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 140 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 141 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 142 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 143 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 144 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 145 spdk_trace_register_description("RDMA_REQ_EXECUTING", 146 TRACE_RDMA_REQUEST_STATE_EXECUTING, 147 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 148 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 149 spdk_trace_register_description("RDMA_REQ_EXECUTED", 150 TRACE_RDMA_REQUEST_STATE_EXECUTED, 151 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 152 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 153 spdk_trace_register_description("RDMA_REQ_RDY2COMPL_PEND", 154 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 155 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 156 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 157 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 158 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 159 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 160 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 161 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 162 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 163 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 164 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 165 spdk_trace_register_description("RDMA_REQ_COMPLETING", 166 TRACE_RDMA_REQUEST_STATE_COMPLETING, 167 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 168 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 169 170 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 171 OWNER_TYPE_NONE, OBJECT_NONE, 0, 172 SPDK_TRACE_ARG_TYPE_INT, ""); 173 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 174 OWNER_TYPE_NONE, OBJECT_NONE, 0, 175 SPDK_TRACE_ARG_TYPE_INT, "type"); 176 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 177 OWNER_TYPE_NONE, OBJECT_NONE, 0, 178 SPDK_TRACE_ARG_TYPE_INT, "type"); 179 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 180 OWNER_TYPE_NONE, OBJECT_NONE, 0, 181 SPDK_TRACE_ARG_TYPE_INT, ""); 182 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 183 OWNER_TYPE_NONE, OBJECT_NONE, 0, 184 SPDK_TRACE_ARG_TYPE_INT, ""); 185 } 186 187 enum spdk_nvmf_rdma_wr_type { 188 RDMA_WR_TYPE_RECV, 189 RDMA_WR_TYPE_SEND, 190 RDMA_WR_TYPE_DATA, 191 }; 192 193 struct spdk_nvmf_rdma_wr { 194 /* Uses enum spdk_nvmf_rdma_wr_type */ 195 uint8_t type; 196 }; 197 198 /* This structure holds commands as they are received off the wire. 199 * It must be dynamically paired with a full request object 200 * (spdk_nvmf_rdma_request) to service a request. It is separate 201 * from the request because RDMA does not appear to order 202 * completions, so occasionally we'll get a new incoming 203 * command when there aren't any free request objects. 204 */ 205 struct spdk_nvmf_rdma_recv { 206 struct ibv_recv_wr wr; 207 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 208 209 struct spdk_nvmf_rdma_qpair *qpair; 210 211 /* In-capsule data buffer */ 212 uint8_t *buf; 213 214 struct spdk_nvmf_rdma_wr rdma_wr; 215 uint64_t receive_tsc; 216 217 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 218 }; 219 220 struct spdk_nvmf_rdma_request_data { 221 struct ibv_send_wr wr; 222 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 223 }; 224 225 struct spdk_nvmf_rdma_request { 226 struct spdk_nvmf_request req; 227 228 bool fused_failed; 229 230 struct spdk_nvmf_rdma_wr data_wr; 231 struct spdk_nvmf_rdma_wr rsp_wr; 232 233 /* Uses enum spdk_nvmf_rdma_request_state */ 234 uint8_t state; 235 236 /* Data offset in req.iov */ 237 uint32_t offset; 238 239 struct spdk_nvmf_rdma_recv *recv; 240 241 struct { 242 struct ibv_send_wr wr; 243 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 244 } rsp; 245 246 uint16_t iovpos; 247 uint16_t num_outstanding_data_wr; 248 /* Used to split Write IO with multi SGL payload */ 249 uint16_t num_remaining_data_wr; 250 uint64_t receive_tsc; 251 struct spdk_nvmf_rdma_request *fused_pair; 252 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 253 struct ibv_send_wr *remaining_tranfer_in_wrs; 254 struct ibv_send_wr *transfer_wr; 255 struct spdk_nvmf_rdma_request_data data; 256 }; 257 258 struct spdk_nvmf_rdma_resource_opts { 259 struct spdk_nvmf_rdma_qpair *qpair; 260 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 261 void *qp; 262 struct spdk_rdma_mem_map *map; 263 uint32_t max_queue_depth; 264 uint32_t in_capsule_data_size; 265 bool shared; 266 }; 267 268 struct spdk_nvmf_rdma_resources { 269 /* Array of size "max_queue_depth" containing RDMA requests. */ 270 struct spdk_nvmf_rdma_request *reqs; 271 272 /* Array of size "max_queue_depth" containing RDMA recvs. */ 273 struct spdk_nvmf_rdma_recv *recvs; 274 275 /* Array of size "max_queue_depth" containing 64 byte capsules 276 * used for receive. 277 */ 278 union nvmf_h2c_msg *cmds; 279 280 /* Array of size "max_queue_depth" containing 16 byte completions 281 * to be sent back to the user. 282 */ 283 union nvmf_c2h_msg *cpls; 284 285 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 286 * buffers to be used for in capsule data. 287 */ 288 void *bufs; 289 290 /* Receives that are waiting for a request object */ 291 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 292 293 /* Queue to track free requests */ 294 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 295 }; 296 297 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 298 299 typedef void (*spdk_poller_destroy_cb)(void *ctx); 300 301 struct spdk_nvmf_rdma_ibv_event_ctx { 302 struct spdk_nvmf_rdma_qpair *rqpair; 303 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 304 /* Link to other ibv events associated with this qpair */ 305 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 306 }; 307 308 struct spdk_nvmf_rdma_qpair { 309 struct spdk_nvmf_qpair qpair; 310 311 struct spdk_nvmf_rdma_device *device; 312 struct spdk_nvmf_rdma_poller *poller; 313 314 struct spdk_rdma_qp *rdma_qp; 315 struct rdma_cm_id *cm_id; 316 struct spdk_rdma_srq *srq; 317 struct rdma_cm_id *listen_id; 318 319 /* Cache the QP number to improve QP search by RB tree. */ 320 uint32_t qp_num; 321 322 /* The maximum number of I/O outstanding on this connection at one time */ 323 uint16_t max_queue_depth; 324 325 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 326 uint16_t max_read_depth; 327 328 /* The maximum number of RDMA SEND operations at one time */ 329 uint32_t max_send_depth; 330 331 /* The current number of outstanding WRs from this qpair's 332 * recv queue. Should not exceed device->attr.max_queue_depth. 333 */ 334 uint16_t current_recv_depth; 335 336 /* The current number of active RDMA READ operations */ 337 uint16_t current_read_depth; 338 339 /* The current number of posted WRs from this qpair's 340 * send queue. Should not exceed max_send_depth. 341 */ 342 uint32_t current_send_depth; 343 344 /* The maximum number of SGEs per WR on the send queue */ 345 uint32_t max_send_sge; 346 347 /* The maximum number of SGEs per WR on the recv queue */ 348 uint32_t max_recv_sge; 349 350 struct spdk_nvmf_rdma_resources *resources; 351 352 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 353 354 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 355 356 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_send_queue; 357 358 /* Number of requests not in the free state */ 359 uint32_t qd; 360 361 bool ibv_in_error_state; 362 363 RB_ENTRY(spdk_nvmf_rdma_qpair) node; 364 365 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 366 367 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 368 369 /* Points to the a request that has fuse bits set to 370 * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting 371 * for the request that has SPDK_NVME_CMD_FUSE_SECOND. 372 */ 373 struct spdk_nvmf_rdma_request *fused_first; 374 375 /* 376 * io_channel which is used to destroy qpair when it is removed from poll group 377 */ 378 struct spdk_io_channel *destruct_channel; 379 380 /* List of ibv async events */ 381 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 382 383 /* Lets us know that we have received the last_wqe event. */ 384 bool last_wqe_reached; 385 386 /* Indicate that nvmf_rdma_close_qpair is called */ 387 bool to_close; 388 }; 389 390 struct spdk_nvmf_rdma_poller_stat { 391 uint64_t completions; 392 uint64_t polls; 393 uint64_t idle_polls; 394 uint64_t requests; 395 uint64_t request_latency; 396 uint64_t pending_free_request; 397 uint64_t pending_rdma_read; 398 uint64_t pending_rdma_write; 399 uint64_t pending_rdma_send; 400 struct spdk_rdma_qp_stats qp_stats; 401 }; 402 403 struct spdk_nvmf_rdma_poller { 404 struct spdk_nvmf_rdma_device *device; 405 struct spdk_nvmf_rdma_poll_group *group; 406 407 int num_cqe; 408 int required_num_wr; 409 struct ibv_cq *cq; 410 411 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 412 uint16_t max_srq_depth; 413 bool need_destroy; 414 415 /* Shared receive queue */ 416 struct spdk_rdma_srq *srq; 417 418 struct spdk_nvmf_rdma_resources *resources; 419 struct spdk_nvmf_rdma_poller_stat stat; 420 421 spdk_poller_destroy_cb destroy_cb; 422 void *destroy_cb_ctx; 423 424 RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs; 425 426 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 427 428 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 429 430 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 431 }; 432 433 struct spdk_nvmf_rdma_poll_group_stat { 434 uint64_t pending_data_buffer; 435 }; 436 437 struct spdk_nvmf_rdma_poll_group { 438 struct spdk_nvmf_transport_poll_group group; 439 struct spdk_nvmf_rdma_poll_group_stat stat; 440 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 441 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 442 }; 443 444 struct spdk_nvmf_rdma_conn_sched { 445 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 446 struct spdk_nvmf_rdma_poll_group *next_io_pg; 447 }; 448 449 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 450 struct spdk_nvmf_rdma_device { 451 struct ibv_device_attr attr; 452 struct ibv_context *context; 453 454 struct spdk_rdma_mem_map *map; 455 struct ibv_pd *pd; 456 457 int num_srq; 458 bool need_destroy; 459 bool ready_to_destroy; 460 bool is_ready; 461 462 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 463 }; 464 465 struct spdk_nvmf_rdma_port { 466 const struct spdk_nvme_transport_id *trid; 467 struct rdma_cm_id *id; 468 struct spdk_nvmf_rdma_device *device; 469 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 470 }; 471 472 struct rdma_transport_opts { 473 int num_cqe; 474 uint32_t max_srq_depth; 475 bool no_srq; 476 bool no_wr_batching; 477 int acceptor_backlog; 478 }; 479 480 struct spdk_nvmf_rdma_transport { 481 struct spdk_nvmf_transport transport; 482 struct rdma_transport_opts rdma_opts; 483 484 struct spdk_nvmf_rdma_conn_sched conn_sched; 485 486 struct rdma_event_channel *event_channel; 487 488 struct spdk_mempool *data_wr_pool; 489 490 struct spdk_poller *accept_poller; 491 492 /* fields used to poll RDMA/IB events */ 493 nfds_t npoll_fds; 494 struct pollfd *poll_fds; 495 496 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 497 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 498 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 499 500 /* ports that are removed unexpectedly and need retry listen */ 501 TAILQ_HEAD(, spdk_nvmf_rdma_port) retry_ports; 502 }; 503 504 struct poller_manage_ctx { 505 struct spdk_nvmf_rdma_transport *rtransport; 506 struct spdk_nvmf_rdma_poll_group *rgroup; 507 struct spdk_nvmf_rdma_poller *rpoller; 508 struct spdk_nvmf_rdma_device *device; 509 510 struct spdk_thread *thread; 511 volatile int *inflight_op_counter; 512 }; 513 514 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 515 { 516 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe), 517 spdk_json_decode_int32, true 518 }, 519 { 520 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 521 spdk_json_decode_uint32, true 522 }, 523 { 524 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 525 spdk_json_decode_bool, true 526 }, 527 { 528 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 529 spdk_json_decode_bool, true 530 }, 531 { 532 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 533 spdk_json_decode_int32, true 534 }, 535 }; 536 537 static int 538 nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2) 539 { 540 return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num; 541 } 542 543 RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare); 544 545 static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 546 struct spdk_nvmf_rdma_request *rdma_req); 547 548 static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 549 struct spdk_nvmf_rdma_poller *rpoller); 550 551 static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 552 struct spdk_nvmf_rdma_poller *rpoller); 553 554 static void _nvmf_rdma_remove_destroyed_device(void *c); 555 556 static inline enum spdk_nvme_media_error_status_code 557 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 558 enum spdk_nvme_media_error_status_code result; 559 switch (err_type) 560 { 561 case SPDK_DIF_REFTAG_ERROR: 562 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 563 break; 564 case SPDK_DIF_APPTAG_ERROR: 565 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 566 break; 567 case SPDK_DIF_GUARD_ERROR: 568 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 569 break; 570 default: 571 SPDK_UNREACHABLE(); 572 } 573 574 return result; 575 } 576 577 /* 578 * Return data_wrs to pool starting from \b data_wr 579 * Request's own response and data WR are excluded 580 */ 581 static void 582 _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 583 struct ibv_send_wr *data_wr, 584 struct spdk_mempool *pool) 585 { 586 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 587 struct spdk_nvmf_rdma_request_data *nvmf_data; 588 struct ibv_send_wr *next_send_wr; 589 uint64_t req_wrid = (uint64_t)&rdma_req->data_wr; 590 uint32_t num_wrs = 0; 591 592 while (data_wr && data_wr->wr_id == req_wrid) { 593 nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr); 594 memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge); 595 data_wr->num_sge = 0; 596 next_send_wr = data_wr->next; 597 if (data_wr != &rdma_req->data.wr) { 598 data_wr->next = NULL; 599 assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES); 600 work_requests[num_wrs] = nvmf_data; 601 num_wrs++; 602 } 603 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr; 604 } 605 606 if (num_wrs) { 607 spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs); 608 } 609 } 610 611 static void 612 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 613 struct spdk_nvmf_rdma_transport *rtransport) 614 { 615 rdma_req->num_outstanding_data_wr = 0; 616 617 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 618 619 if (rdma_req->remaining_tranfer_in_wrs) { 620 _nvmf_rdma_request_free_data(rdma_req, rdma_req->remaining_tranfer_in_wrs, 621 rtransport->data_wr_pool); 622 rdma_req->remaining_tranfer_in_wrs = NULL; 623 } 624 625 rdma_req->data.wr.next = NULL; 626 rdma_req->rsp.wr.next = NULL; 627 } 628 629 static void 630 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 631 { 632 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 633 if (req->req.cmd) { 634 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 635 } 636 if (req->recv) { 637 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 638 } 639 } 640 641 static void 642 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 643 { 644 int i; 645 646 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 647 for (i = 0; i < rqpair->max_queue_depth; i++) { 648 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 649 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 650 } 651 } 652 } 653 654 static void 655 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 656 { 657 spdk_free(resources->cmds); 658 spdk_free(resources->cpls); 659 spdk_free(resources->bufs); 660 spdk_free(resources->reqs); 661 spdk_free(resources->recvs); 662 free(resources); 663 } 664 665 666 static struct spdk_nvmf_rdma_resources * 667 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 668 { 669 struct spdk_nvmf_rdma_resources *resources; 670 struct spdk_nvmf_rdma_request *rdma_req; 671 struct spdk_nvmf_rdma_recv *rdma_recv; 672 struct spdk_rdma_qp *qp = NULL; 673 struct spdk_rdma_srq *srq = NULL; 674 struct ibv_recv_wr *bad_wr = NULL; 675 struct spdk_rdma_memory_translation translation; 676 uint32_t i; 677 int rc = 0; 678 679 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 680 if (!resources) { 681 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 682 return NULL; 683 } 684 685 resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs), 686 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 687 resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs), 688 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 689 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 690 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 691 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 692 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 693 694 if (opts->in_capsule_data_size > 0) { 695 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 696 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 697 SPDK_MALLOC_DMA); 698 } 699 700 if (!resources->reqs || !resources->recvs || !resources->cmds || 701 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 702 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 703 goto cleanup; 704 } 705 706 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n", 707 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds)); 708 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n", 709 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls)); 710 if (resources->bufs) { 711 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n", 712 resources->bufs, opts->max_queue_depth * 713 opts->in_capsule_data_size); 714 } 715 716 /* Initialize queues */ 717 STAILQ_INIT(&resources->incoming_queue); 718 STAILQ_INIT(&resources->free_queue); 719 720 if (opts->shared) { 721 srq = (struct spdk_rdma_srq *)opts->qp; 722 } else { 723 qp = (struct spdk_rdma_qp *)opts->qp; 724 } 725 726 for (i = 0; i < opts->max_queue_depth; i++) { 727 rdma_recv = &resources->recvs[i]; 728 rdma_recv->qpair = opts->qpair; 729 730 /* Set up memory to receive commands */ 731 if (resources->bufs) { 732 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 733 opts->in_capsule_data_size)); 734 } 735 736 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 737 738 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 739 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 740 rc = spdk_rdma_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]), 741 &translation); 742 if (rc) { 743 goto cleanup; 744 } 745 rdma_recv->sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 746 rdma_recv->wr.num_sge = 1; 747 748 if (rdma_recv->buf) { 749 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 750 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 751 rc = spdk_rdma_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size, &translation); 752 if (rc) { 753 goto cleanup; 754 } 755 rdma_recv->sgl[1].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 756 rdma_recv->wr.num_sge++; 757 } 758 759 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 760 rdma_recv->wr.sg_list = rdma_recv->sgl; 761 if (srq) { 762 spdk_rdma_srq_queue_recv_wrs(srq, &rdma_recv->wr); 763 } else { 764 spdk_rdma_qp_queue_recv_wrs(qp, &rdma_recv->wr); 765 } 766 } 767 768 for (i = 0; i < opts->max_queue_depth; i++) { 769 rdma_req = &resources->reqs[i]; 770 771 if (opts->qpair != NULL) { 772 rdma_req->req.qpair = &opts->qpair->qpair; 773 } else { 774 rdma_req->req.qpair = NULL; 775 } 776 rdma_req->req.cmd = NULL; 777 rdma_req->req.iovcnt = 0; 778 rdma_req->req.stripped_data = NULL; 779 780 /* Set up memory to send responses */ 781 rdma_req->req.rsp = &resources->cpls[i]; 782 783 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 784 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 785 rc = spdk_rdma_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]), 786 &translation); 787 if (rc) { 788 goto cleanup; 789 } 790 rdma_req->rsp.sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation); 791 792 rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND; 793 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr; 794 rdma_req->rsp.wr.next = NULL; 795 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 796 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 797 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 798 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 799 800 /* Set up memory for data buffers */ 801 rdma_req->data_wr.type = RDMA_WR_TYPE_DATA; 802 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr; 803 rdma_req->data.wr.next = NULL; 804 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 805 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 806 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 807 808 /* Initialize request state to FREE */ 809 rdma_req->state = RDMA_REQUEST_STATE_FREE; 810 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 811 } 812 813 if (srq) { 814 rc = spdk_rdma_srq_flush_recv_wrs(srq, &bad_wr); 815 } else { 816 rc = spdk_rdma_qp_flush_recv_wrs(qp, &bad_wr); 817 } 818 819 if (rc) { 820 goto cleanup; 821 } 822 823 return resources; 824 825 cleanup: 826 nvmf_rdma_resources_destroy(resources); 827 return NULL; 828 } 829 830 static void 831 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 832 { 833 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 834 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 835 ctx->rqpair = NULL; 836 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 837 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 838 } 839 } 840 841 static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller); 842 843 static void 844 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 845 { 846 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 847 struct ibv_recv_wr *bad_recv_wr = NULL; 848 int rc; 849 850 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair); 851 852 if (rqpair->qd != 0) { 853 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 854 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 855 struct spdk_nvmf_rdma_transport, transport); 856 struct spdk_nvmf_rdma_request *req; 857 uint32_t i, max_req_count = 0; 858 859 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 860 861 if (rqpair->srq == NULL) { 862 nvmf_rdma_dump_qpair_contents(rqpair); 863 max_req_count = rqpair->max_queue_depth; 864 } else if (rqpair->poller && rqpair->resources) { 865 max_req_count = rqpair->poller->max_srq_depth; 866 } 867 868 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 869 for (i = 0; i < max_req_count; i++) { 870 req = &rqpair->resources->reqs[i]; 871 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 872 /* nvmf_rdma_request_process checks qpair ibv and internal state 873 * and completes a request */ 874 nvmf_rdma_request_process(rtransport, req); 875 } 876 } 877 assert(rqpair->qd == 0); 878 } 879 880 if (rqpair->poller) { 881 RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair); 882 883 if (rqpair->srq != NULL && rqpair->resources != NULL) { 884 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 885 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 886 if (rqpair == rdma_recv->qpair) { 887 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 888 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr); 889 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 890 if (rc) { 891 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 892 } 893 } 894 } 895 } 896 } 897 898 if (rqpair->cm_id) { 899 if (rqpair->rdma_qp != NULL) { 900 spdk_rdma_qp_destroy(rqpair->rdma_qp); 901 rqpair->rdma_qp = NULL; 902 } 903 904 if (rqpair->poller != NULL && rqpair->srq == NULL) { 905 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 906 } 907 } 908 909 if (rqpair->srq == NULL && rqpair->resources != NULL) { 910 nvmf_rdma_resources_destroy(rqpair->resources); 911 } 912 913 nvmf_rdma_qpair_clean_ibv_events(rqpair); 914 915 if (rqpair->destruct_channel) { 916 spdk_put_io_channel(rqpair->destruct_channel); 917 rqpair->destruct_channel = NULL; 918 } 919 920 if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) { 921 nvmf_rdma_poller_destroy(rqpair->poller); 922 } 923 924 /* destroy cm_id last so cma device will not be freed before we destroy the cq. */ 925 if (rqpair->cm_id) { 926 rdma_destroy_id(rqpair->cm_id); 927 } 928 929 free(rqpair); 930 } 931 932 static int 933 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 934 { 935 struct spdk_nvmf_rdma_poller *rpoller; 936 int rc, num_cqe, required_num_wr; 937 938 /* Enlarge CQ size dynamically */ 939 rpoller = rqpair->poller; 940 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 941 num_cqe = rpoller->num_cqe; 942 if (num_cqe < required_num_wr) { 943 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 944 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 945 } 946 947 if (rpoller->num_cqe != num_cqe) { 948 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) { 949 SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n" 950 "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe); 951 return -1; 952 } 953 if (required_num_wr > device->attr.max_cqe) { 954 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 955 required_num_wr, device->attr.max_cqe); 956 return -1; 957 } 958 959 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 960 rc = ibv_resize_cq(rpoller->cq, num_cqe); 961 if (rc) { 962 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 963 return -1; 964 } 965 966 rpoller->num_cqe = num_cqe; 967 } 968 969 rpoller->required_num_wr = required_num_wr; 970 return 0; 971 } 972 973 static int 974 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 975 { 976 struct spdk_nvmf_rdma_qpair *rqpair; 977 struct spdk_nvmf_rdma_transport *rtransport; 978 struct spdk_nvmf_transport *transport; 979 struct spdk_nvmf_rdma_resource_opts opts; 980 struct spdk_nvmf_rdma_device *device; 981 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 982 983 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 984 device = rqpair->device; 985 986 qp_init_attr.qp_context = rqpair; 987 qp_init_attr.pd = device->pd; 988 qp_init_attr.send_cq = rqpair->poller->cq; 989 qp_init_attr.recv_cq = rqpair->poller->cq; 990 991 if (rqpair->srq) { 992 qp_init_attr.srq = rqpair->srq->srq; 993 } else { 994 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 995 } 996 997 /* SEND, READ, and WRITE operations */ 998 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 999 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 1000 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1001 qp_init_attr.stats = &rqpair->poller->stat.qp_stats; 1002 1003 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1004 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1005 goto error; 1006 } 1007 1008 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 1009 if (!rqpair->rdma_qp) { 1010 goto error; 1011 } 1012 1013 rqpair->qp_num = rqpair->rdma_qp->qp->qp_num; 1014 1015 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1016 qp_init_attr.cap.max_send_wr); 1017 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1018 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1019 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair); 1020 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1021 1022 if (rqpair->poller->srq == NULL) { 1023 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1024 transport = &rtransport->transport; 1025 1026 opts.qp = rqpair->rdma_qp; 1027 opts.map = device->map; 1028 opts.qpair = rqpair; 1029 opts.shared = false; 1030 opts.max_queue_depth = rqpair->max_queue_depth; 1031 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1032 1033 rqpair->resources = nvmf_rdma_resources_create(&opts); 1034 1035 if (!rqpair->resources) { 1036 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1037 rdma_destroy_qp(rqpair->cm_id); 1038 goto error; 1039 } 1040 } else { 1041 rqpair->resources = rqpair->poller->resources; 1042 } 1043 1044 rqpair->current_recv_depth = 0; 1045 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1046 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1047 STAILQ_INIT(&rqpair->pending_rdma_send_queue); 1048 rqpair->qpair.queue_depth = 0; 1049 1050 return 0; 1051 1052 error: 1053 rdma_destroy_id(rqpair->cm_id); 1054 rqpair->cm_id = NULL; 1055 return -1; 1056 } 1057 1058 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1059 /* This function accepts either a single wr or the first wr in a linked list. */ 1060 static void 1061 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1062 { 1063 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1064 struct spdk_nvmf_rdma_transport, transport); 1065 1066 if (rqpair->srq != NULL) { 1067 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, first); 1068 } else { 1069 if (spdk_rdma_qp_queue_recv_wrs(rqpair->rdma_qp, first)) { 1070 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1071 } 1072 } 1073 1074 if (rtransport->rdma_opts.no_wr_batching) { 1075 _poller_submit_recvs(rtransport, rqpair->poller); 1076 } 1077 } 1078 1079 static int 1080 request_transfer_in(struct spdk_nvmf_request *req) 1081 { 1082 struct spdk_nvmf_rdma_request *rdma_req; 1083 struct spdk_nvmf_qpair *qpair; 1084 struct spdk_nvmf_rdma_qpair *rqpair; 1085 struct spdk_nvmf_rdma_transport *rtransport; 1086 1087 qpair = req->qpair; 1088 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1089 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1090 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1091 struct spdk_nvmf_rdma_transport, transport); 1092 1093 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1094 assert(rdma_req != NULL); 1095 1096 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) { 1097 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1098 } 1099 if (rtransport->rdma_opts.no_wr_batching) { 1100 _poller_submit_sends(rtransport, rqpair->poller); 1101 } 1102 1103 assert(rqpair->current_read_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_read_depth); 1104 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1105 assert(rqpair->current_send_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_send_depth); 1106 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1107 return 0; 1108 } 1109 1110 static inline int 1111 nvmf_rdma_request_reset_transfer_in(struct spdk_nvmf_rdma_request *rdma_req, 1112 struct spdk_nvmf_rdma_transport *rtransport) 1113 { 1114 /* Put completed WRs back to pool and move transfer_wr pointer */ 1115 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 1116 rdma_req->transfer_wr = rdma_req->remaining_tranfer_in_wrs; 1117 rdma_req->remaining_tranfer_in_wrs = NULL; 1118 rdma_req->num_outstanding_data_wr = rdma_req->num_remaining_data_wr; 1119 rdma_req->num_remaining_data_wr = 0; 1120 1121 return 0; 1122 } 1123 1124 static inline int 1125 request_prepare_transfer_in_part(struct spdk_nvmf_request *req, uint32_t num_reads_available) 1126 { 1127 struct spdk_nvmf_rdma_request *rdma_req; 1128 struct ibv_send_wr *wr; 1129 uint32_t i; 1130 1131 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1132 1133 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1134 assert(rdma_req != NULL); 1135 assert(num_reads_available > 0); 1136 assert(rdma_req->num_outstanding_data_wr > num_reads_available); 1137 wr = rdma_req->transfer_wr; 1138 1139 for (i = 0; i < num_reads_available - 1; i++) { 1140 wr = wr->next; 1141 } 1142 1143 rdma_req->remaining_tranfer_in_wrs = wr->next; 1144 rdma_req->num_remaining_data_wr = rdma_req->num_outstanding_data_wr - num_reads_available; 1145 rdma_req->num_outstanding_data_wr = num_reads_available; 1146 /* Break chain of WRs to send only part. Once this portion completes, we continue sending RDMA_READs */ 1147 wr->next = NULL; 1148 1149 return 0; 1150 } 1151 1152 static int 1153 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1154 { 1155 int num_outstanding_data_wr = 0; 1156 struct spdk_nvmf_rdma_request *rdma_req; 1157 struct spdk_nvmf_qpair *qpair; 1158 struct spdk_nvmf_rdma_qpair *rqpair; 1159 struct spdk_nvme_cpl *rsp; 1160 struct ibv_send_wr *first = NULL; 1161 struct spdk_nvmf_rdma_transport *rtransport; 1162 1163 *data_posted = 0; 1164 qpair = req->qpair; 1165 rsp = &req->rsp->nvme_cpl; 1166 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1167 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1168 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1169 struct spdk_nvmf_rdma_transport, transport); 1170 1171 /* Advance our sq_head pointer */ 1172 if (qpair->sq_head == qpair->sq_head_max) { 1173 qpair->sq_head = 0; 1174 } else { 1175 qpair->sq_head++; 1176 } 1177 rsp->sqhd = qpair->sq_head; 1178 1179 /* queue the capsule for the recv buffer */ 1180 assert(rdma_req->recv != NULL); 1181 1182 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1183 1184 rdma_req->recv = NULL; 1185 assert(rqpair->current_recv_depth > 0); 1186 rqpair->current_recv_depth--; 1187 1188 /* Build the response which consists of optional 1189 * RDMA WRITEs to transfer data, plus an RDMA SEND 1190 * containing the response. 1191 */ 1192 first = &rdma_req->rsp.wr; 1193 1194 if (spdk_unlikely(rsp->status.sc != SPDK_NVME_SC_SUCCESS)) { 1195 /* On failure, data was not read from the controller. So clear the 1196 * number of outstanding data WRs to zero. 1197 */ 1198 rdma_req->num_outstanding_data_wr = 0; 1199 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1200 first = rdma_req->transfer_wr; 1201 *data_posted = 1; 1202 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1203 } 1204 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1205 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1206 } 1207 if (rtransport->rdma_opts.no_wr_batching) { 1208 _poller_submit_sends(rtransport, rqpair->poller); 1209 } 1210 1211 /* +1 for the rsp wr */ 1212 assert(rqpair->current_send_depth + num_outstanding_data_wr + 1 <= rqpair->max_send_depth); 1213 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1214 1215 return 0; 1216 } 1217 1218 static int 1219 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1220 { 1221 struct spdk_nvmf_rdma_accept_private_data accept_data; 1222 struct rdma_conn_param ctrlr_event_data = {}; 1223 int rc; 1224 1225 accept_data.recfmt = 0; 1226 accept_data.crqsize = rqpair->max_queue_depth; 1227 1228 ctrlr_event_data.private_data = &accept_data; 1229 ctrlr_event_data.private_data_len = sizeof(accept_data); 1230 if (id->ps == RDMA_PS_TCP) { 1231 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1232 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1233 } 1234 1235 /* Configure infinite retries for the initiator side qpair. 1236 * We need to pass this value to the initiator to prevent the 1237 * initiator side NIC from completing SEND requests back to the 1238 * initiator with status rnr_retry_count_exceeded. */ 1239 ctrlr_event_data.rnr_retry_count = 0x7; 1240 1241 /* When qpair is created without use of rdma cm API, an additional 1242 * information must be provided to initiator in the connection response: 1243 * whether qpair is using SRQ and its qp_num 1244 * Fields below are ignored by rdma cm if qpair has been 1245 * created using rdma cm API. */ 1246 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1247 ctrlr_event_data.qp_num = rqpair->qp_num; 1248 1249 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1250 if (rc) { 1251 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1252 } else { 1253 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1254 } 1255 1256 return rc; 1257 } 1258 1259 static void 1260 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1261 { 1262 struct spdk_nvmf_rdma_reject_private_data rej_data; 1263 1264 rej_data.recfmt = 0; 1265 rej_data.sts = error; 1266 1267 rdma_reject(id, &rej_data, sizeof(rej_data)); 1268 } 1269 1270 static int 1271 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1272 { 1273 struct spdk_nvmf_rdma_transport *rtransport; 1274 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1275 struct spdk_nvmf_rdma_port *port; 1276 struct rdma_conn_param *rdma_param = NULL; 1277 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1278 uint16_t max_queue_depth; 1279 uint16_t max_read_depth; 1280 1281 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1282 1283 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1284 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1285 1286 rdma_param = &event->param.conn; 1287 if (rdma_param->private_data == NULL || 1288 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1289 SPDK_ERRLOG("connect request: no private data provided\n"); 1290 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1291 return -1; 1292 } 1293 1294 private_data = rdma_param->private_data; 1295 if (private_data->recfmt != 0) { 1296 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1297 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1298 return -1; 1299 } 1300 1301 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1302 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1303 1304 port = event->listen_id->context; 1305 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1306 event->listen_id, event->listen_id->verbs, port); 1307 1308 /* Figure out the supported queue depth. This is a multi-step process 1309 * that takes into account hardware maximums, host provided values, 1310 * and our target's internal memory limits */ 1311 1312 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1313 1314 /* Start with the maximum queue depth allowed by the target */ 1315 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1316 max_read_depth = rtransport->transport.opts.max_queue_depth; 1317 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1318 rtransport->transport.opts.max_queue_depth); 1319 1320 /* Next check the local NIC's hardware limitations */ 1321 SPDK_DEBUGLOG(rdma, 1322 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1323 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1324 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1325 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1326 1327 /* Next check the remote NIC's hardware limitations */ 1328 SPDK_DEBUGLOG(rdma, 1329 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1330 rdma_param->initiator_depth, rdma_param->responder_resources); 1331 /* from man3 rdma_get_cm_event 1332 * responder_resources - Specifies the number of responder resources that is requested by the recipient. 1333 * The responder_resources field must match the initiator depth specified by the remote node when running 1334 * the rdma_connect and rdma_accept functions. */ 1335 if (rdma_param->responder_resources != 0) { 1336 if (private_data->qid) { 1337 SPDK_DEBUGLOG(rdma, "Host (Initiator) is not allowed to use RDMA operations," 1338 " responder_resources must be 0 but set to %u\n", 1339 rdma_param->responder_resources); 1340 } else { 1341 SPDK_WARNLOG("Host (Initiator) is not allowed to use RDMA operations," 1342 " responder_resources must be 0 but set to %u\n", 1343 rdma_param->responder_resources); 1344 } 1345 } 1346 /* from man3 rdma_get_cm_event 1347 * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds. 1348 * The initiator_depth field must match the responder resources specified by the remote node when running 1349 * the rdma_connect and rdma_accept functions. */ 1350 if (rdma_param->initiator_depth == 0) { 1351 SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n"); 1352 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD); 1353 return -1; 1354 } 1355 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1356 1357 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1358 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1359 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1360 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1361 1362 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1363 max_queue_depth, max_read_depth); 1364 1365 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1366 if (rqpair == NULL) { 1367 SPDK_ERRLOG("Could not allocate new connection.\n"); 1368 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1369 return -1; 1370 } 1371 1372 rqpair->device = port->device; 1373 rqpair->max_queue_depth = max_queue_depth; 1374 rqpair->max_read_depth = max_read_depth; 1375 rqpair->cm_id = event->id; 1376 rqpair->listen_id = event->listen_id; 1377 rqpair->qpair.transport = transport; 1378 STAILQ_INIT(&rqpair->ibv_events); 1379 /* use qid from the private data to determine the qpair type 1380 qid will be set to the appropriate value when the controller is created */ 1381 rqpair->qpair.qid = private_data->qid; 1382 1383 event->id->context = &rqpair->qpair; 1384 1385 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1386 1387 return 0; 1388 } 1389 1390 static inline void 1391 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1392 enum spdk_nvme_data_transfer xfer) 1393 { 1394 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1395 wr->opcode = IBV_WR_RDMA_WRITE; 1396 wr->send_flags = 0; 1397 wr->next = next; 1398 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1399 wr->opcode = IBV_WR_RDMA_READ; 1400 wr->send_flags = IBV_SEND_SIGNALED; 1401 wr->next = NULL; 1402 } else { 1403 assert(0); 1404 } 1405 } 1406 1407 static int 1408 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1409 struct spdk_nvmf_rdma_request *rdma_req, 1410 uint32_t num_sgl_descriptors) 1411 { 1412 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1413 struct spdk_nvmf_rdma_request_data *current_data_wr; 1414 uint32_t i; 1415 1416 if (spdk_unlikely(num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES)) { 1417 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1418 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1419 return -EINVAL; 1420 } 1421 1422 if (spdk_unlikely(spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, 1423 num_sgl_descriptors))) { 1424 return -ENOMEM; 1425 } 1426 1427 current_data_wr = &rdma_req->data; 1428 1429 for (i = 0; i < num_sgl_descriptors; i++) { 1430 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1431 current_data_wr->wr.next = &work_requests[i]->wr; 1432 current_data_wr = work_requests[i]; 1433 current_data_wr->wr.sg_list = current_data_wr->sgl; 1434 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1435 } 1436 1437 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1438 1439 return 0; 1440 } 1441 1442 static inline void 1443 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1444 { 1445 struct ibv_send_wr *wr = &rdma_req->data.wr; 1446 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1447 1448 wr->wr.rdma.rkey = sgl->keyed.key; 1449 wr->wr.rdma.remote_addr = sgl->address; 1450 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1451 } 1452 1453 static inline void 1454 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1455 { 1456 struct ibv_send_wr *wr = &rdma_req->data.wr; 1457 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1458 uint32_t i; 1459 int j; 1460 uint64_t remote_addr_offset = 0; 1461 1462 for (i = 0; i < num_wrs; ++i) { 1463 wr->wr.rdma.rkey = sgl->keyed.key; 1464 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1465 for (j = 0; j < wr->num_sge; ++j) { 1466 remote_addr_offset += wr->sg_list[j].length; 1467 } 1468 wr = wr->next; 1469 } 1470 } 1471 1472 static int 1473 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_device *device, 1474 struct spdk_nvmf_rdma_request *rdma_req, 1475 struct ibv_send_wr *wr, 1476 uint32_t total_length) 1477 { 1478 struct spdk_rdma_memory_translation mem_translation; 1479 struct ibv_sge *sg_ele; 1480 struct iovec *iov; 1481 uint32_t lkey, remaining; 1482 int rc; 1483 1484 wr->num_sge = 0; 1485 1486 while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) { 1487 iov = &rdma_req->req.iov[rdma_req->iovpos]; 1488 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1489 if (spdk_unlikely(rc)) { 1490 return rc; 1491 } 1492 1493 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1494 sg_ele = &wr->sg_list[wr->num_sge]; 1495 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1496 1497 sg_ele->lkey = lkey; 1498 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset; 1499 sg_ele->length = remaining; 1500 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1501 sg_ele->length); 1502 rdma_req->offset += sg_ele->length; 1503 total_length -= sg_ele->length; 1504 wr->num_sge++; 1505 1506 if (rdma_req->offset == iov->iov_len) { 1507 rdma_req->offset = 0; 1508 rdma_req->iovpos++; 1509 } 1510 } 1511 1512 if (spdk_unlikely(total_length)) { 1513 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1514 return -EINVAL; 1515 } 1516 1517 return 0; 1518 } 1519 1520 static int 1521 nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_device *device, 1522 struct spdk_nvmf_rdma_request *rdma_req, 1523 struct ibv_send_wr *wr, 1524 uint32_t total_length, 1525 uint32_t num_extra_wrs) 1526 { 1527 struct spdk_rdma_memory_translation mem_translation; 1528 struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx; 1529 struct ibv_sge *sg_ele; 1530 struct iovec *iov; 1531 struct iovec *rdma_iov; 1532 uint32_t lkey, remaining; 1533 uint32_t remaining_data_block, data_block_size, md_size; 1534 uint32_t sge_len; 1535 int rc; 1536 1537 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1538 1539 if (spdk_likely(!rdma_req->req.stripped_data)) { 1540 rdma_iov = rdma_req->req.iov; 1541 remaining_data_block = data_block_size; 1542 md_size = dif_ctx->md_size; 1543 } else { 1544 rdma_iov = rdma_req->req.stripped_data->iov; 1545 total_length = total_length / dif_ctx->block_size * data_block_size; 1546 remaining_data_block = total_length; 1547 md_size = 0; 1548 } 1549 1550 wr->num_sge = 0; 1551 1552 while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1553 iov = rdma_iov + rdma_req->iovpos; 1554 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1555 if (spdk_unlikely(rc)) { 1556 return rc; 1557 } 1558 1559 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1560 sg_ele = &wr->sg_list[wr->num_sge]; 1561 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1562 1563 while (remaining) { 1564 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1565 if (num_extra_wrs > 0 && wr->next) { 1566 wr = wr->next; 1567 wr->num_sge = 0; 1568 sg_ele = &wr->sg_list[wr->num_sge]; 1569 num_extra_wrs--; 1570 } else { 1571 break; 1572 } 1573 } 1574 sg_ele->lkey = lkey; 1575 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset); 1576 sge_len = spdk_min(remaining, remaining_data_block); 1577 sg_ele->length = sge_len; 1578 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, 1579 sg_ele->addr, sg_ele->length); 1580 remaining -= sge_len; 1581 remaining_data_block -= sge_len; 1582 rdma_req->offset += sge_len; 1583 total_length -= sge_len; 1584 1585 sg_ele++; 1586 wr->num_sge++; 1587 1588 if (remaining_data_block == 0) { 1589 /* skip metadata */ 1590 rdma_req->offset += md_size; 1591 total_length -= md_size; 1592 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1593 remaining -= spdk_min(remaining, md_size); 1594 remaining_data_block = data_block_size; 1595 } 1596 1597 if (remaining == 0) { 1598 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1599 the remaining metadata bits at the beginning of the next buffer */ 1600 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset); 1601 rdma_req->iovpos++; 1602 } 1603 } 1604 } 1605 1606 if (spdk_unlikely(total_length)) { 1607 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1608 return -EINVAL; 1609 } 1610 1611 return 0; 1612 } 1613 1614 static inline uint32_t 1615 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1616 { 1617 /* estimate the number of SG entries and WRs needed to process the request */ 1618 uint32_t num_sge = 0; 1619 uint32_t i; 1620 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1621 1622 for (i = 0; i < num_buffers && length > 0; i++) { 1623 uint32_t buffer_len = spdk_min(length, io_unit_size); 1624 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1625 1626 if (num_sge_in_block * block_size > buffer_len) { 1627 ++num_sge_in_block; 1628 } 1629 num_sge += num_sge_in_block; 1630 length -= buffer_len; 1631 } 1632 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1633 } 1634 1635 static int 1636 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1637 struct spdk_nvmf_rdma_device *device, 1638 struct spdk_nvmf_rdma_request *rdma_req) 1639 { 1640 struct spdk_nvmf_rdma_qpair *rqpair; 1641 struct spdk_nvmf_rdma_poll_group *rgroup; 1642 struct spdk_nvmf_request *req = &rdma_req->req; 1643 struct ibv_send_wr *wr = &rdma_req->data.wr; 1644 int rc; 1645 uint32_t num_wrs = 1; 1646 uint32_t length; 1647 1648 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1649 rgroup = rqpair->poller->group; 1650 1651 /* rdma wr specifics */ 1652 nvmf_rdma_setup_request(rdma_req); 1653 1654 length = req->length; 1655 if (spdk_unlikely(req->dif_enabled)) { 1656 req->dif.orig_length = length; 1657 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1658 req->dif.elba_length = length; 1659 } 1660 1661 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1662 length); 1663 if (spdk_unlikely(rc != 0)) { 1664 return rc; 1665 } 1666 1667 assert(req->iovcnt <= rqpair->max_send_sge); 1668 1669 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1670 * the stripped_buffers are got for DIF stripping. */ 1671 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1672 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1673 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1674 &rtransport->transport, req->dif.orig_length); 1675 if (rc != 0) { 1676 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1677 } 1678 } 1679 1680 rdma_req->iovpos = 0; 1681 1682 if (spdk_unlikely(req->dif_enabled)) { 1683 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1684 req->dif.dif_ctx.block_size); 1685 if (num_wrs > 1) { 1686 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1687 if (spdk_unlikely(rc != 0)) { 1688 goto err_exit; 1689 } 1690 } 1691 1692 rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, wr, length, num_wrs - 1); 1693 if (spdk_unlikely(rc != 0)) { 1694 goto err_exit; 1695 } 1696 1697 if (num_wrs > 1) { 1698 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1699 } 1700 } else { 1701 rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, wr, length); 1702 if (spdk_unlikely(rc != 0)) { 1703 goto err_exit; 1704 } 1705 } 1706 1707 /* set the number of outstanding data WRs for this request. */ 1708 rdma_req->num_outstanding_data_wr = num_wrs; 1709 1710 return rc; 1711 1712 err_exit: 1713 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1714 nvmf_rdma_request_free_data(rdma_req, rtransport); 1715 req->iovcnt = 0; 1716 return rc; 1717 } 1718 1719 static int 1720 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1721 struct spdk_nvmf_rdma_device *device, 1722 struct spdk_nvmf_rdma_request *rdma_req) 1723 { 1724 struct spdk_nvmf_rdma_qpair *rqpair; 1725 struct spdk_nvmf_rdma_poll_group *rgroup; 1726 struct ibv_send_wr *current_wr; 1727 struct spdk_nvmf_request *req = &rdma_req->req; 1728 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1729 uint32_t num_sgl_descriptors; 1730 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0; 1731 uint32_t i; 1732 int rc; 1733 1734 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1735 rgroup = rqpair->poller->group; 1736 1737 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1738 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1739 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1740 1741 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1742 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1743 1744 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1745 for (i = 0; i < num_sgl_descriptors; i++) { 1746 if (spdk_likely(!req->dif_enabled)) { 1747 lengths[i] = desc->keyed.length; 1748 } else { 1749 req->dif.orig_length += desc->keyed.length; 1750 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1751 req->dif.elba_length += lengths[i]; 1752 } 1753 total_length += lengths[i]; 1754 desc++; 1755 } 1756 1757 if (spdk_unlikely(total_length > rtransport->transport.opts.max_io_size)) { 1758 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n", 1759 total_length, rtransport->transport.opts.max_io_size); 1760 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1761 return -EINVAL; 1762 } 1763 1764 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1); 1765 if (spdk_unlikely(rc != 0)) { 1766 return -ENOMEM; 1767 } 1768 1769 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length); 1770 if (spdk_unlikely(rc != 0)) { 1771 nvmf_rdma_request_free_data(rdma_req, rtransport); 1772 return rc; 1773 } 1774 1775 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1776 * the stripped_buffers are got for DIF stripping. */ 1777 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1778 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1779 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1780 &rtransport->transport, req->dif.orig_length); 1781 if (spdk_unlikely(rc != 0)) { 1782 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1783 } 1784 } 1785 1786 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1787 current_wr = &rdma_req->data.wr; 1788 assert(current_wr != NULL); 1789 1790 req->length = 0; 1791 rdma_req->iovpos = 0; 1792 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1793 for (i = 0; i < num_sgl_descriptors; i++) { 1794 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1795 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1796 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1797 rc = -EINVAL; 1798 goto err_exit; 1799 } 1800 1801 if (spdk_likely(!req->dif_enabled)) { 1802 rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, current_wr, lengths[i]); 1803 } else { 1804 rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, current_wr, 1805 lengths[i], 0); 1806 } 1807 if (spdk_unlikely(rc != 0)) { 1808 rc = -ENOMEM; 1809 goto err_exit; 1810 } 1811 1812 req->length += desc->keyed.length; 1813 current_wr->wr.rdma.rkey = desc->keyed.key; 1814 current_wr->wr.rdma.remote_addr = desc->address; 1815 current_wr = current_wr->next; 1816 desc++; 1817 } 1818 1819 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1820 /* Go back to the last descriptor in the list. */ 1821 desc--; 1822 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1823 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1824 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1825 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1826 } 1827 } 1828 #endif 1829 1830 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1831 1832 return 0; 1833 1834 err_exit: 1835 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1836 nvmf_rdma_request_free_data(rdma_req, rtransport); 1837 return rc; 1838 } 1839 1840 static int 1841 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1842 struct spdk_nvmf_rdma_device *device, 1843 struct spdk_nvmf_rdma_request *rdma_req) 1844 { 1845 struct spdk_nvmf_request *req = &rdma_req->req; 1846 struct spdk_nvme_cpl *rsp; 1847 struct spdk_nvme_sgl_descriptor *sgl; 1848 int rc; 1849 uint32_t length; 1850 1851 rsp = &req->rsp->nvme_cpl; 1852 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1853 1854 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1855 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1856 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1857 1858 length = sgl->keyed.length; 1859 if (spdk_unlikely(length > rtransport->transport.opts.max_io_size)) { 1860 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1861 length, rtransport->transport.opts.max_io_size); 1862 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1863 return -1; 1864 } 1865 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1866 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1867 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1868 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1869 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1870 } 1871 } 1872 #endif 1873 1874 /* fill request length and populate iovs */ 1875 req->length = length; 1876 1877 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req); 1878 if (spdk_unlikely(rc < 0)) { 1879 if (rc == -EINVAL) { 1880 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1881 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1882 return -1; 1883 } 1884 /* No available buffers. Queue this request up. */ 1885 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1886 return 0; 1887 } 1888 1889 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1890 req->iovcnt); 1891 1892 return 0; 1893 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1894 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1895 uint64_t offset = sgl->address; 1896 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1897 1898 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1899 offset, sgl->unkeyed.length); 1900 1901 if (spdk_unlikely(offset > max_len)) { 1902 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1903 offset, max_len); 1904 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1905 return -1; 1906 } 1907 max_len -= (uint32_t)offset; 1908 1909 if (spdk_unlikely(sgl->unkeyed.length > max_len)) { 1910 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1911 sgl->unkeyed.length, max_len); 1912 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1913 return -1; 1914 } 1915 1916 rdma_req->num_outstanding_data_wr = 0; 1917 req->data_from_pool = false; 1918 req->length = sgl->unkeyed.length; 1919 1920 req->iov[0].iov_base = rdma_req->recv->buf + offset; 1921 req->iov[0].iov_len = req->length; 1922 req->iovcnt = 1; 1923 1924 return 0; 1925 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1926 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1927 1928 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1929 if (spdk_unlikely(rc == -ENOMEM)) { 1930 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1931 return 0; 1932 } else if (spdk_unlikely(rc == -EINVAL)) { 1933 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1934 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1935 return -1; 1936 } 1937 1938 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1939 req->iovcnt); 1940 1941 return 0; 1942 } 1943 1944 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1945 sgl->generic.type, sgl->generic.subtype); 1946 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1947 return -1; 1948 } 1949 1950 static void 1951 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1952 struct spdk_nvmf_rdma_transport *rtransport) 1953 { 1954 struct spdk_nvmf_rdma_qpair *rqpair; 1955 struct spdk_nvmf_rdma_poll_group *rgroup; 1956 1957 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1958 if (rdma_req->req.data_from_pool) { 1959 rgroup = rqpair->poller->group; 1960 1961 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1962 } 1963 if (rdma_req->req.stripped_data) { 1964 nvmf_request_free_stripped_buffers(&rdma_req->req, 1965 &rqpair->poller->group->group, 1966 &rtransport->transport); 1967 } 1968 nvmf_rdma_request_free_data(rdma_req, rtransport); 1969 rdma_req->req.length = 0; 1970 rdma_req->req.iovcnt = 0; 1971 rdma_req->offset = 0; 1972 rdma_req->req.dif_enabled = false; 1973 rdma_req->fused_failed = false; 1974 rdma_req->transfer_wr = NULL; 1975 if (rdma_req->fused_pair) { 1976 /* This req was part of a valid fused pair, but failed before it got to 1977 * READ_TO_EXECUTE state. This means we need to fail the other request 1978 * in the pair, because it is no longer part of a valid pair. If the pair 1979 * already reached READY_TO_EXECUTE state, we need to kick it. 1980 */ 1981 rdma_req->fused_pair->fused_failed = true; 1982 if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 1983 nvmf_rdma_request_process(rtransport, rdma_req->fused_pair); 1984 } 1985 rdma_req->fused_pair = NULL; 1986 } 1987 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1988 rqpair->qd--; 1989 1990 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1991 rqpair->qpair.queue_depth--; 1992 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1993 } 1994 1995 static void 1996 nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport, 1997 struct spdk_nvmf_rdma_qpair *rqpair, 1998 struct spdk_nvmf_rdma_request *rdma_req) 1999 { 2000 enum spdk_nvme_cmd_fuse last, next; 2001 2002 last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2003 next = rdma_req->req.cmd->nvme_cmd.fuse; 2004 2005 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2006 2007 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2008 return; 2009 } 2010 2011 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2012 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2013 /* This is a valid pair of fused commands. Point them at each other 2014 * so they can be submitted consecutively once ready to be executed. 2015 */ 2016 rqpair->fused_first->fused_pair = rdma_req; 2017 rdma_req->fused_pair = rqpair->fused_first; 2018 rqpair->fused_first = NULL; 2019 return; 2020 } else { 2021 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2022 rqpair->fused_first->fused_failed = true; 2023 2024 /* If the last req is in READY_TO_EXECUTE state, then call 2025 * nvmf_rdma_request_process(), otherwise nothing else will kick it. 2026 */ 2027 if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2028 nvmf_rdma_request_process(rtransport, rqpair->fused_first); 2029 } 2030 2031 rqpair->fused_first = NULL; 2032 } 2033 } 2034 2035 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2036 /* Set rqpair->fused_first here so that we know to check that the next request 2037 * is a SECOND (and to fail this one if it isn't). 2038 */ 2039 rqpair->fused_first = rdma_req; 2040 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2041 /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */ 2042 rdma_req->fused_failed = true; 2043 } 2044 } 2045 2046 bool 2047 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 2048 struct spdk_nvmf_rdma_request *rdma_req) 2049 { 2050 struct spdk_nvmf_rdma_qpair *rqpair; 2051 struct spdk_nvmf_rdma_device *device; 2052 struct spdk_nvmf_rdma_poll_group *rgroup; 2053 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2054 int rc; 2055 struct spdk_nvmf_rdma_recv *rdma_recv; 2056 enum spdk_nvmf_rdma_request_state prev_state; 2057 bool progress = false; 2058 int data_posted; 2059 uint32_t num_blocks, num_rdma_reads_available, qdepth; 2060 2061 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2062 device = rqpair->device; 2063 rgroup = rqpair->poller->group; 2064 2065 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 2066 2067 /* If the queue pair is in an error state, force the request to the completed state 2068 * to release resources. */ 2069 if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 2070 switch (rdma_req->state) { 2071 case RDMA_REQUEST_STATE_NEED_BUFFER: 2072 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 2073 break; 2074 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2075 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2076 break; 2077 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2078 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2079 break; 2080 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 2081 STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2082 break; 2083 default: 2084 break; 2085 } 2086 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2087 } 2088 2089 /* The loop here is to allow for several back-to-back state changes. */ 2090 do { 2091 prev_state = rdma_req->state; 2092 2093 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 2094 2095 switch (rdma_req->state) { 2096 case RDMA_REQUEST_STATE_FREE: 2097 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 2098 * to escape this state. */ 2099 break; 2100 case RDMA_REQUEST_STATE_NEW: 2101 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 2102 (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth); 2103 rdma_recv = rdma_req->recv; 2104 2105 /* The first element of the SGL is the NVMe command */ 2106 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 2107 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 2108 rdma_req->transfer_wr = &rdma_req->data.wr; 2109 2110 if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 2111 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2112 break; 2113 } 2114 2115 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 2116 rdma_req->req.dif_enabled = true; 2117 } 2118 2119 nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req); 2120 2121 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2122 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 2123 rdma_req->rsp.wr.imm_data = 0; 2124 #endif 2125 2126 /* The next state transition depends on the data transfer needs of this request. */ 2127 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2128 2129 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2130 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2131 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2132 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2133 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2134 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 2135 break; 2136 } 2137 2138 /* If no data to transfer, ready to execute. */ 2139 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2140 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2141 break; 2142 } 2143 2144 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2145 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2146 break; 2147 case RDMA_REQUEST_STATE_NEED_BUFFER: 2148 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2149 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2150 2151 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2152 2153 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2154 /* This request needs to wait in line to obtain a buffer */ 2155 break; 2156 } 2157 2158 /* Try to get a data buffer */ 2159 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2160 if (spdk_unlikely(rc < 0)) { 2161 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2162 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2163 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2164 break; 2165 } 2166 2167 if (rdma_req->req.iovcnt == 0) { 2168 /* No buffers available. */ 2169 rgroup->stat.pending_data_buffer++; 2170 break; 2171 } 2172 2173 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2174 2175 /* If data is transferring from host to controller and the data didn't 2176 * arrive using in capsule data, we need to do a transfer from the host. 2177 */ 2178 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2179 rdma_req->req.data_from_pool) { 2180 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2181 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2182 break; 2183 } 2184 2185 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2186 break; 2187 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2188 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2189 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2190 2191 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2192 /* This request needs to wait in line to perform RDMA */ 2193 break; 2194 } 2195 assert(rqpair->max_send_depth >= rqpair->current_send_depth); 2196 qdepth = rqpair->max_send_depth - rqpair->current_send_depth; 2197 assert(rqpair->max_read_depth >= rqpair->current_read_depth); 2198 num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth; 2199 if (rdma_req->num_outstanding_data_wr > qdepth || 2200 rdma_req->num_outstanding_data_wr > num_rdma_reads_available) { 2201 if (num_rdma_reads_available && qdepth) { 2202 /* Send as much as we can */ 2203 request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth)); 2204 } else { 2205 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2206 rqpair->poller->stat.pending_rdma_read++; 2207 break; 2208 } 2209 } 2210 2211 /* We have already verified that this request is the head of the queue. */ 2212 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2213 2214 rc = request_transfer_in(&rdma_req->req); 2215 if (spdk_likely(rc == 0)) { 2216 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2217 } else { 2218 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2219 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2220 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2221 } 2222 break; 2223 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2224 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2225 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2226 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2227 * to escape this state. */ 2228 break; 2229 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2230 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2231 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2232 2233 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2234 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2235 /* generate DIF for write operation */ 2236 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2237 assert(num_blocks > 0); 2238 2239 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2240 num_blocks, &rdma_req->req.dif.dif_ctx); 2241 if (rc != 0) { 2242 SPDK_ERRLOG("DIF generation failed\n"); 2243 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2244 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 2245 break; 2246 } 2247 } 2248 2249 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2250 /* set extended length before IO operation */ 2251 rdma_req->req.length = rdma_req->req.dif.elba_length; 2252 } 2253 2254 if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 2255 if (rdma_req->fused_failed) { 2256 /* This request failed FUSED semantics. Fail it immediately, without 2257 * even sending it to the target layer. 2258 */ 2259 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2260 rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED; 2261 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2262 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2263 break; 2264 } 2265 2266 if (rdma_req->fused_pair == NULL || 2267 rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2268 /* This request is ready to execute, but either we don't know yet if it's 2269 * valid - i.e. this is a FIRST but we haven't received the next 2270 * request yet or the other request of this fused pair isn't ready to 2271 * execute. So break here and this request will get processed later either 2272 * when the other request is ready or we find that this request isn't valid. 2273 */ 2274 break; 2275 } 2276 } 2277 2278 /* If we get to this point, and this request is a fused command, we know that 2279 * it is part of valid sequence (FIRST followed by a SECOND) and that both 2280 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 2281 * request, and the other request of the fused pair, in the correct order. 2282 * Also clear the ->fused_pair pointers on both requests, since after this point 2283 * we no longer need to maintain the relationship between these two requests. 2284 */ 2285 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 2286 assert(rdma_req->fused_pair != NULL); 2287 assert(rdma_req->fused_pair->fused_pair != NULL); 2288 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2289 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2290 rdma_req->fused_pair->fused_pair = NULL; 2291 rdma_req->fused_pair = NULL; 2292 } 2293 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2294 spdk_nvmf_request_exec(&rdma_req->req); 2295 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 2296 assert(rdma_req->fused_pair != NULL); 2297 assert(rdma_req->fused_pair->fused_pair != NULL); 2298 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2299 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2300 rdma_req->fused_pair->fused_pair = NULL; 2301 rdma_req->fused_pair = NULL; 2302 } 2303 break; 2304 case RDMA_REQUEST_STATE_EXECUTING: 2305 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2306 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2307 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2308 * to escape this state. */ 2309 break; 2310 case RDMA_REQUEST_STATE_EXECUTED: 2311 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2312 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2313 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2314 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2315 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2316 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2317 } else { 2318 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2319 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2320 } 2321 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2322 /* restore the original length */ 2323 rdma_req->req.length = rdma_req->req.dif.orig_length; 2324 2325 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2326 struct spdk_dif_error error_blk; 2327 2328 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2329 if (!rdma_req->req.stripped_data) { 2330 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2331 &rdma_req->req.dif.dif_ctx, &error_blk); 2332 } else { 2333 rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov, 2334 rdma_req->req.stripped_data->iovcnt, 2335 rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2336 &rdma_req->req.dif.dif_ctx, &error_blk); 2337 } 2338 if (rc) { 2339 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2340 2341 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2342 error_blk.err_offset); 2343 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2344 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2345 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2346 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2347 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2348 } 2349 } 2350 } 2351 break; 2352 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2353 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2354 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2355 2356 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2357 /* This request needs to wait in line to perform RDMA */ 2358 break; 2359 } 2360 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2361 rqpair->max_send_depth) { 2362 /* We can only have so many WRs outstanding. we have to wait until some finish. 2363 * +1 since each request has an additional wr in the resp. */ 2364 rqpair->poller->stat.pending_rdma_write++; 2365 break; 2366 } 2367 2368 /* We have already verified that this request is the head of the queue. */ 2369 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2370 2371 /* The data transfer will be kicked off from 2372 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2373 * We verified that data + response fit into send queue, so we can go to the next state directly 2374 */ 2375 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2376 break; 2377 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 2378 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 0, 0, 2379 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2380 2381 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_send_queue)) { 2382 /* This request needs to wait in line to send the completion */ 2383 break; 2384 } 2385 2386 assert(rqpair->current_send_depth <= rqpair->max_send_depth); 2387 if (rqpair->current_send_depth == rqpair->max_send_depth) { 2388 /* We can only have so many WRs outstanding. we have to wait until some finish */ 2389 rqpair->poller->stat.pending_rdma_send++; 2390 break; 2391 } 2392 2393 /* We have already verified that this request is the head of the queue. */ 2394 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_send_queue, state_link); 2395 2396 /* The response sending will be kicked off from 2397 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2398 */ 2399 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2400 break; 2401 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2402 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2403 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2404 rc = request_transfer_out(&rdma_req->req, &data_posted); 2405 assert(rc == 0); /* No good way to handle this currently */ 2406 if (spdk_unlikely(rc)) { 2407 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2408 } else { 2409 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2410 RDMA_REQUEST_STATE_COMPLETING; 2411 } 2412 break; 2413 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2414 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2415 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2416 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2417 * to escape this state. */ 2418 break; 2419 case RDMA_REQUEST_STATE_COMPLETING: 2420 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2421 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2422 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2423 * to escape this state. */ 2424 break; 2425 case RDMA_REQUEST_STATE_COMPLETED: 2426 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2427 (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth); 2428 2429 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2430 _nvmf_rdma_request_free(rdma_req, rtransport); 2431 break; 2432 case RDMA_REQUEST_NUM_STATES: 2433 default: 2434 assert(0); 2435 break; 2436 } 2437 2438 if (rdma_req->state != prev_state) { 2439 progress = true; 2440 } 2441 } while (rdma_req->state != prev_state); 2442 2443 return progress; 2444 } 2445 2446 /* Public API callbacks begin here */ 2447 2448 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2449 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2450 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2451 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2452 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2453 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2454 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2455 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2456 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX 2457 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2458 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2459 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2460 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2461 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2462 #define SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE 4095 2463 2464 static void 2465 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2466 { 2467 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2468 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2469 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2470 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2471 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2472 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2473 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2474 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2475 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2476 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2477 opts->transport_specific = NULL; 2478 opts->data_wr_pool_size = SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE; 2479 } 2480 2481 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2482 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2483 2484 static inline bool 2485 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2486 { 2487 return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD || 2488 device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW; 2489 } 2490 2491 static int nvmf_rdma_accept(void *ctx); 2492 static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport); 2493 static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2494 struct spdk_nvmf_rdma_device *device); 2495 2496 static int 2497 create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context, 2498 struct spdk_nvmf_rdma_device **new_device) 2499 { 2500 struct spdk_nvmf_rdma_device *device; 2501 int flag = 0; 2502 int rc = 0; 2503 2504 device = calloc(1, sizeof(*device)); 2505 if (!device) { 2506 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2507 return -ENOMEM; 2508 } 2509 device->context = context; 2510 rc = ibv_query_device(device->context, &device->attr); 2511 if (rc < 0) { 2512 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2513 free(device); 2514 return rc; 2515 } 2516 2517 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2518 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2519 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2520 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2521 } 2522 2523 /** 2524 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2525 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2526 * but incorrectly reports that it does. There are changes making their way 2527 * through the kernel now that will enable this feature. When they are merged, 2528 * we can conditionally enable this feature. 2529 * 2530 * TODO: enable this for versions of the kernel rxe driver that support it. 2531 */ 2532 if (nvmf_rdma_is_rxe_device(device)) { 2533 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2534 } 2535 #endif 2536 2537 /* set up device context async ev fd as NON_BLOCKING */ 2538 flag = fcntl(device->context->async_fd, F_GETFL); 2539 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2540 if (rc < 0) { 2541 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2542 free(device); 2543 return rc; 2544 } 2545 2546 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2547 SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA trasport\n", device); 2548 2549 if (g_nvmf_hooks.get_ibv_pd) { 2550 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2551 } else { 2552 device->pd = ibv_alloc_pd(device->context); 2553 } 2554 2555 if (!device->pd) { 2556 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2557 destroy_ib_device(rtransport, device); 2558 return -ENOMEM; 2559 } 2560 2561 assert(device->map == NULL); 2562 2563 device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks, SPDK_RDMA_MEMORY_MAP_ROLE_TARGET); 2564 if (!device->map) { 2565 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2566 destroy_ib_device(rtransport, device); 2567 return -ENOMEM; 2568 } 2569 2570 assert(device->map != NULL); 2571 assert(device->pd != NULL); 2572 2573 if (new_device) { 2574 *new_device = device; 2575 } 2576 SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device), 2577 device, context); 2578 2579 return 0; 2580 } 2581 2582 static void 2583 free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2584 { 2585 if (rtransport->poll_fds) { 2586 free(rtransport->poll_fds); 2587 rtransport->poll_fds = NULL; 2588 } 2589 rtransport->npoll_fds = 0; 2590 } 2591 2592 static int 2593 generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2594 { 2595 /* Set up poll descriptor array to monitor events from RDMA and IB 2596 * in a single poll syscall 2597 */ 2598 int device_count = 0; 2599 int i = 0; 2600 struct spdk_nvmf_rdma_device *device, *tmp; 2601 2602 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2603 device_count++; 2604 } 2605 2606 rtransport->npoll_fds = device_count + 1; 2607 2608 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2609 if (rtransport->poll_fds == NULL) { 2610 SPDK_ERRLOG("poll_fds allocation failed\n"); 2611 return -ENOMEM; 2612 } 2613 2614 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2615 rtransport->poll_fds[i++].events = POLLIN; 2616 2617 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2618 rtransport->poll_fds[i].fd = device->context->async_fd; 2619 rtransport->poll_fds[i++].events = POLLIN; 2620 } 2621 2622 return 0; 2623 } 2624 2625 static struct spdk_nvmf_transport * 2626 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2627 { 2628 int rc; 2629 struct spdk_nvmf_rdma_transport *rtransport; 2630 struct spdk_nvmf_rdma_device *device; 2631 struct ibv_context **contexts; 2632 size_t data_wr_pool_size; 2633 uint32_t i; 2634 int flag; 2635 uint32_t sge_count; 2636 uint32_t min_shared_buffers; 2637 uint32_t min_in_capsule_data_size; 2638 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2639 2640 rtransport = calloc(1, sizeof(*rtransport)); 2641 if (!rtransport) { 2642 return NULL; 2643 } 2644 2645 TAILQ_INIT(&rtransport->devices); 2646 TAILQ_INIT(&rtransport->ports); 2647 TAILQ_INIT(&rtransport->poll_groups); 2648 TAILQ_INIT(&rtransport->retry_ports); 2649 2650 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2651 rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2652 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2653 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2654 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2655 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2656 if (opts->transport_specific != NULL && 2657 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2658 SPDK_COUNTOF(rdma_transport_opts_decoder), 2659 &rtransport->rdma_opts)) { 2660 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2661 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2662 return NULL; 2663 } 2664 2665 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2666 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2667 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2668 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2669 " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d," 2670 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2671 opts->max_queue_depth, 2672 opts->max_io_size, 2673 opts->max_qpairs_per_ctrlr - 1, 2674 opts->io_unit_size, 2675 opts->in_capsule_data_size, 2676 opts->max_aq_depth, 2677 opts->num_shared_buffers, 2678 rtransport->rdma_opts.num_cqe, 2679 rtransport->rdma_opts.max_srq_depth, 2680 rtransport->rdma_opts.no_srq, 2681 rtransport->rdma_opts.acceptor_backlog, 2682 rtransport->rdma_opts.no_wr_batching, 2683 opts->abort_timeout_sec); 2684 2685 /* I/O unit size cannot be larger than max I/O size */ 2686 if (opts->io_unit_size > opts->max_io_size) { 2687 opts->io_unit_size = opts->max_io_size; 2688 } 2689 2690 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2691 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2692 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2693 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2694 } 2695 2696 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2697 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2698 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2699 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2700 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2701 return NULL; 2702 } 2703 2704 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 2705 if (opts->buf_cache_size < UINT32_MAX) { 2706 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2707 if (min_shared_buffers > opts->num_shared_buffers) { 2708 SPDK_ERRLOG("There are not enough buffers to satisfy" 2709 "per-poll group caches for each thread. (%" PRIu32 ")" 2710 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2711 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2712 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2713 return NULL; 2714 } 2715 } 2716 2717 sge_count = opts->max_io_size / opts->io_unit_size; 2718 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2719 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2720 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2721 return NULL; 2722 } 2723 2724 min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES; 2725 if (opts->in_capsule_data_size < min_in_capsule_data_size) { 2726 SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n", 2727 min_in_capsule_data_size); 2728 opts->in_capsule_data_size = min_in_capsule_data_size; 2729 } 2730 2731 rtransport->event_channel = rdma_create_event_channel(); 2732 if (rtransport->event_channel == NULL) { 2733 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2734 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2735 return NULL; 2736 } 2737 2738 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2739 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2740 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2741 rtransport->event_channel->fd, spdk_strerror(errno)); 2742 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2743 return NULL; 2744 } 2745 2746 data_wr_pool_size = opts->data_wr_pool_size; 2747 if (data_wr_pool_size < SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count()) { 2748 data_wr_pool_size = SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count(); 2749 SPDK_NOTICELOG("data_wr_pool_size is changed to %zu to guarantee enough cache for handling " 2750 "at least one IO in each core\n", data_wr_pool_size); 2751 } 2752 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", data_wr_pool_size, 2753 sizeof(struct spdk_nvmf_rdma_request_data), SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2754 SPDK_ENV_SOCKET_ID_ANY); 2755 if (!rtransport->data_wr_pool) { 2756 if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) { 2757 SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n"); 2758 SPDK_ERRLOG("Probably running in multiprocess environment, which is " 2759 "unsupported by the nvmf library\n"); 2760 } else { 2761 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2762 } 2763 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2764 return NULL; 2765 } 2766 2767 contexts = rdma_get_devices(NULL); 2768 if (contexts == NULL) { 2769 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2770 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2771 return NULL; 2772 } 2773 2774 i = 0; 2775 rc = 0; 2776 while (contexts[i] != NULL) { 2777 rc = create_ib_device(rtransport, contexts[i], &device); 2778 if (rc < 0) { 2779 break; 2780 } 2781 i++; 2782 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2783 device->is_ready = true; 2784 } 2785 rdma_free_devices(contexts); 2786 2787 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2788 /* divide and round up. */ 2789 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2790 2791 /* round up to the nearest 4k. */ 2792 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2793 2794 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2795 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2796 opts->io_unit_size); 2797 } 2798 2799 if (rc < 0) { 2800 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2801 return NULL; 2802 } 2803 2804 rc = generate_poll_fds(rtransport); 2805 if (rc < 0) { 2806 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2807 return NULL; 2808 } 2809 2810 rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport, 2811 opts->acceptor_poll_rate); 2812 if (!rtransport->accept_poller) { 2813 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2814 return NULL; 2815 } 2816 2817 return &rtransport->transport; 2818 } 2819 2820 static void 2821 destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2822 struct spdk_nvmf_rdma_device *device) 2823 { 2824 TAILQ_REMOVE(&rtransport->devices, device, link); 2825 spdk_rdma_free_mem_map(&device->map); 2826 if (device->pd) { 2827 if (!g_nvmf_hooks.get_ibv_pd) { 2828 ibv_dealloc_pd(device->pd); 2829 } 2830 } 2831 SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device); 2832 free(device); 2833 } 2834 2835 static void 2836 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2837 { 2838 struct spdk_nvmf_rdma_transport *rtransport; 2839 assert(w != NULL); 2840 2841 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2842 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2843 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2844 if (rtransport->rdma_opts.no_srq == true) { 2845 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2846 } 2847 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2848 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2849 } 2850 2851 static int 2852 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2853 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2854 { 2855 struct spdk_nvmf_rdma_transport *rtransport; 2856 struct spdk_nvmf_rdma_port *port, *port_tmp; 2857 struct spdk_nvmf_rdma_device *device, *device_tmp; 2858 2859 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2860 2861 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) { 2862 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 2863 free(port); 2864 } 2865 2866 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2867 TAILQ_REMOVE(&rtransport->ports, port, link); 2868 rdma_destroy_id(port->id); 2869 free(port); 2870 } 2871 2872 free_poll_fds(rtransport); 2873 2874 if (rtransport->event_channel != NULL) { 2875 rdma_destroy_event_channel(rtransport->event_channel); 2876 } 2877 2878 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2879 destroy_ib_device(rtransport, device); 2880 } 2881 2882 if (rtransport->data_wr_pool != NULL) { 2883 if (spdk_mempool_count(rtransport->data_wr_pool) != 2884 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2885 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2886 spdk_mempool_count(rtransport->data_wr_pool), 2887 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2888 } 2889 } 2890 2891 spdk_mempool_free(rtransport->data_wr_pool); 2892 2893 spdk_poller_unregister(&rtransport->accept_poller); 2894 free(rtransport); 2895 2896 if (cb_fn) { 2897 cb_fn(cb_arg); 2898 } 2899 return 0; 2900 } 2901 2902 static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2903 struct spdk_nvme_transport_id *trid, 2904 bool peer); 2905 2906 static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport); 2907 2908 static int 2909 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2910 struct spdk_nvmf_listen_opts *listen_opts) 2911 { 2912 struct spdk_nvmf_rdma_transport *rtransport; 2913 struct spdk_nvmf_rdma_device *device; 2914 struct spdk_nvmf_rdma_port *port, *tmp_port; 2915 struct addrinfo *res; 2916 struct addrinfo hints; 2917 int family; 2918 int rc; 2919 long int port_val; 2920 bool is_retry = false; 2921 2922 if (!strlen(trid->trsvcid)) { 2923 SPDK_ERRLOG("Service id is required\n"); 2924 return -EINVAL; 2925 } 2926 2927 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2928 assert(rtransport->event_channel != NULL); 2929 2930 port = calloc(1, sizeof(*port)); 2931 if (!port) { 2932 SPDK_ERRLOG("Port allocation failed\n"); 2933 return -ENOMEM; 2934 } 2935 2936 port->trid = trid; 2937 2938 switch (trid->adrfam) { 2939 case SPDK_NVMF_ADRFAM_IPV4: 2940 family = AF_INET; 2941 break; 2942 case SPDK_NVMF_ADRFAM_IPV6: 2943 family = AF_INET6; 2944 break; 2945 default: 2946 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2947 free(port); 2948 return -EINVAL; 2949 } 2950 2951 memset(&hints, 0, sizeof(hints)); 2952 hints.ai_family = family; 2953 hints.ai_flags = AI_NUMERICSERV; 2954 hints.ai_socktype = SOCK_STREAM; 2955 hints.ai_protocol = 0; 2956 2957 /* Range check the trsvcid. Fail in 3 cases: 2958 * < 0: means that spdk_strtol hit an error 2959 * 0: this results in ephemeral port which we don't want 2960 * > 65535: port too high 2961 */ 2962 port_val = spdk_strtol(trid->trsvcid, 10); 2963 if (port_val <= 0 || port_val > 65535) { 2964 SPDK_ERRLOG("invalid trsvcid %s\n", trid->trsvcid); 2965 free(port); 2966 return -EINVAL; 2967 } 2968 2969 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2970 if (rc) { 2971 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2972 free(port); 2973 return -(abs(rc)); 2974 } 2975 2976 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2977 if (rc < 0) { 2978 SPDK_ERRLOG("rdma_create_id() failed\n"); 2979 freeaddrinfo(res); 2980 free(port); 2981 return rc; 2982 } 2983 2984 rc = rdma_bind_addr(port->id, res->ai_addr); 2985 freeaddrinfo(res); 2986 2987 if (rc < 0) { 2988 TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) { 2989 if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) { 2990 is_retry = true; 2991 break; 2992 } 2993 } 2994 if (!is_retry) { 2995 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2996 } 2997 rdma_destroy_id(port->id); 2998 free(port); 2999 return rc; 3000 } 3001 3002 if (!port->id->verbs) { 3003 SPDK_ERRLOG("ibv_context is null\n"); 3004 rdma_destroy_id(port->id); 3005 free(port); 3006 return -1; 3007 } 3008 3009 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 3010 if (rc < 0) { 3011 SPDK_ERRLOG("rdma_listen() failed\n"); 3012 rdma_destroy_id(port->id); 3013 free(port); 3014 return rc; 3015 } 3016 3017 TAILQ_FOREACH(device, &rtransport->devices, link) { 3018 if (device->context == port->id->verbs && device->is_ready) { 3019 port->device = device; 3020 break; 3021 } 3022 } 3023 if (!port->device) { 3024 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 3025 port->id->verbs); 3026 rdma_destroy_id(port->id); 3027 free(port); 3028 nvmf_rdma_rescan_devices(rtransport); 3029 return -EINVAL; 3030 } 3031 3032 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 3033 trid->traddr, trid->trsvcid); 3034 3035 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 3036 return 0; 3037 } 3038 3039 static void 3040 nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport, 3041 const struct spdk_nvme_transport_id *trid, bool need_retry) 3042 { 3043 struct spdk_nvmf_rdma_transport *rtransport; 3044 struct spdk_nvmf_rdma_port *port, *tmp; 3045 3046 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3047 3048 if (!need_retry) { 3049 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) { 3050 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 3051 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3052 free(port); 3053 } 3054 } 3055 } 3056 3057 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 3058 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 3059 SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n", 3060 port->trid->traddr, port->trid->trsvcid, need_retry); 3061 TAILQ_REMOVE(&rtransport->ports, port, link); 3062 rdma_destroy_id(port->id); 3063 port->id = NULL; 3064 port->device = NULL; 3065 if (need_retry) { 3066 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3067 } else { 3068 free(port); 3069 } 3070 break; 3071 } 3072 } 3073 } 3074 3075 static void 3076 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 3077 const struct spdk_nvme_transport_id *trid) 3078 { 3079 nvmf_rdma_stop_listen_ex(transport, trid, false); 3080 } 3081 3082 static void _nvmf_rdma_register_poller_in_group(void *c); 3083 static void _nvmf_rdma_remove_poller_in_group(void *c); 3084 3085 static bool 3086 nvmf_rdma_all_pollers_management_done(void *c) 3087 { 3088 struct poller_manage_ctx *ctx = c; 3089 int counter; 3090 3091 counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST); 3092 SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n", 3093 counter, ctx->rpoller); 3094 3095 if (counter == 0) { 3096 free((void *)ctx->inflight_op_counter); 3097 } 3098 free(ctx); 3099 3100 return counter == 0; 3101 } 3102 3103 static int 3104 nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport, 3105 struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add) 3106 { 3107 struct spdk_nvmf_rdma_poll_group *rgroup; 3108 struct spdk_nvmf_rdma_poller *rpoller; 3109 struct spdk_nvmf_poll_group *poll_group; 3110 struct poller_manage_ctx *ctx; 3111 bool found; 3112 int *inflight_counter; 3113 spdk_msg_fn do_fn; 3114 3115 *has_inflight = false; 3116 do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group; 3117 inflight_counter = calloc(1, sizeof(int)); 3118 if (!inflight_counter) { 3119 SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n"); 3120 return -ENOMEM; 3121 } 3122 3123 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3124 (*inflight_counter)++; 3125 } 3126 3127 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3128 found = false; 3129 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3130 if (rpoller->device == device) { 3131 found = true; 3132 break; 3133 } 3134 } 3135 if (found == is_add) { 3136 __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST); 3137 continue; 3138 } 3139 3140 ctx = calloc(1, sizeof(struct poller_manage_ctx)); 3141 if (!ctx) { 3142 SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n"); 3143 if (!*has_inflight) { 3144 free(inflight_counter); 3145 } 3146 return -ENOMEM; 3147 } 3148 3149 ctx->rtransport = rtransport; 3150 ctx->rgroup = rgroup; 3151 ctx->rpoller = rpoller; 3152 ctx->device = device; 3153 ctx->thread = spdk_get_thread(); 3154 ctx->inflight_op_counter = inflight_counter; 3155 *has_inflight = true; 3156 3157 poll_group = rgroup->group.group; 3158 if (poll_group->thread != spdk_get_thread()) { 3159 spdk_thread_send_msg(poll_group->thread, do_fn, ctx); 3160 } else { 3161 do_fn(ctx); 3162 } 3163 } 3164 3165 if (!*has_inflight) { 3166 free(inflight_counter); 3167 } 3168 3169 return 0; 3170 } 3171 3172 static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3173 struct spdk_nvmf_rdma_device *device); 3174 3175 static struct spdk_nvmf_rdma_device * 3176 nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 3177 struct ibv_context *context) 3178 { 3179 struct spdk_nvmf_rdma_device *device, *tmp_device; 3180 3181 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) { 3182 if (device->need_destroy) { 3183 continue; 3184 } 3185 3186 if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) { 3187 return device; 3188 } 3189 } 3190 3191 return NULL; 3192 } 3193 3194 static bool 3195 nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport, 3196 struct ibv_context *context) 3197 { 3198 struct spdk_nvmf_rdma_device *old_device, *new_device; 3199 int rc = 0; 3200 bool has_inflight; 3201 3202 old_device = nvmf_rdma_find_ib_device(rtransport, context); 3203 3204 if (old_device) { 3205 if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) { 3206 /* context may not have time to be cleaned when rescan. exactly one context 3207 * is valid for a device so this context must be invalid and just remove it. */ 3208 SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context); 3209 old_device->need_destroy = true; 3210 nvmf_rdma_handle_device_removal(rtransport, old_device); 3211 } 3212 return false; 3213 } 3214 3215 rc = create_ib_device(rtransport, context, &new_device); 3216 /* TODO: update transport opts. */ 3217 if (rc < 0) { 3218 SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n", 3219 ibv_get_device_name(context->device), context); 3220 return false; 3221 } 3222 3223 rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true); 3224 if (rc < 0) { 3225 SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n", 3226 ibv_get_device_name(context->device), context); 3227 return false; 3228 } 3229 3230 if (has_inflight) { 3231 new_device->is_ready = true; 3232 } 3233 3234 return true; 3235 } 3236 3237 static bool 3238 nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport) 3239 { 3240 struct spdk_nvmf_rdma_device *device; 3241 struct ibv_device **ibv_device_list = NULL; 3242 struct ibv_context **contexts = NULL; 3243 int i = 0; 3244 int num_dev = 0; 3245 bool new_create = false, has_new_device = false; 3246 struct ibv_context *tmp_verbs = NULL; 3247 3248 /* do not rescan when any device is destroying, or context may be freed when 3249 * regenerating the poll fds. 3250 */ 3251 TAILQ_FOREACH(device, &rtransport->devices, link) { 3252 if (device->need_destroy) { 3253 return false; 3254 } 3255 } 3256 3257 ibv_device_list = ibv_get_device_list(&num_dev); 3258 3259 /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be 3260 * marked as dead verbs and never be init again. So we need to make sure the 3261 * verbs is available before we call rdma_get_devices. */ 3262 if (num_dev >= 0) { 3263 for (i = 0; i < num_dev; i++) { 3264 tmp_verbs = ibv_open_device(ibv_device_list[i]); 3265 if (!tmp_verbs) { 3266 SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno); 3267 break; 3268 } 3269 if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) { 3270 SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i], 3271 tmp_verbs->device->dev_name); 3272 has_new_device = true; 3273 } 3274 ibv_close_device(tmp_verbs); 3275 } 3276 ibv_free_device_list(ibv_device_list); 3277 if (!tmp_verbs || !has_new_device) { 3278 return false; 3279 } 3280 } 3281 3282 contexts = rdma_get_devices(NULL); 3283 3284 for (i = 0; contexts && contexts[i] != NULL; i++) { 3285 new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]); 3286 } 3287 3288 if (new_create) { 3289 free_poll_fds(rtransport); 3290 generate_poll_fds(rtransport); 3291 } 3292 3293 if (contexts) { 3294 rdma_free_devices(contexts); 3295 } 3296 3297 return new_create; 3298 } 3299 3300 static bool 3301 nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport) 3302 { 3303 struct spdk_nvmf_rdma_port *port, *tmp_port; 3304 int rc = 0; 3305 bool new_create = false; 3306 3307 if (TAILQ_EMPTY(&rtransport->retry_ports)) { 3308 return false; 3309 } 3310 3311 new_create = nvmf_rdma_rescan_devices(rtransport); 3312 3313 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) { 3314 rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL); 3315 3316 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3317 if (rc) { 3318 if (new_create) { 3319 SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n", 3320 port->trid->traddr, port->trid->trsvcid, rc); 3321 } 3322 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3323 break; 3324 } else { 3325 SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid); 3326 free(port); 3327 } 3328 } 3329 3330 return true; 3331 } 3332 3333 static void 3334 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 3335 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 3336 { 3337 struct spdk_nvmf_request *req, *tmp; 3338 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 3339 struct spdk_nvmf_rdma_resources *resources; 3340 3341 /* First process requests which are waiting for response to be sent */ 3342 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_send_queue, state_link, req_tmp) { 3343 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3344 break; 3345 } 3346 } 3347 3348 /* We process I/O in the data transfer pending queue at the highest priority. */ 3349 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 3350 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3351 break; 3352 } 3353 } 3354 3355 /* Then RDMA writes since reads have stronger restrictions than writes */ 3356 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 3357 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3358 break; 3359 } 3360 } 3361 3362 /* Then we handle request waiting on memory buffers. */ 3363 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 3364 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3365 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3366 break; 3367 } 3368 } 3369 3370 resources = rqpair->resources; 3371 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 3372 rdma_req = STAILQ_FIRST(&resources->free_queue); 3373 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 3374 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 3375 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 3376 3377 if (rqpair->srq != NULL) { 3378 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 3379 rdma_req->recv->qpair->qd++; 3380 } else { 3381 rqpair->qd++; 3382 } 3383 3384 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 3385 rdma_req->state = RDMA_REQUEST_STATE_NEW; 3386 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 3387 break; 3388 } 3389 } 3390 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 3391 rqpair->poller->stat.pending_free_request++; 3392 } 3393 } 3394 3395 static void 3396 nvmf_rdma_poller_process_pending_buf_queue(struct spdk_nvmf_rdma_transport *rtransport, 3397 struct spdk_nvmf_rdma_poller *rpoller) 3398 { 3399 struct spdk_nvmf_request *req, *tmp; 3400 struct spdk_nvmf_rdma_request *rdma_req; 3401 3402 STAILQ_FOREACH_SAFE(req, &rpoller->group->group.pending_buf_queue, buf_link, tmp) { 3403 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3404 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 3405 break; 3406 } 3407 } 3408 } 3409 3410 static inline bool 3411 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 3412 { 3413 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 3414 return nvmf_rdma_is_rxe_device(device) || 3415 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 3416 } 3417 3418 static void 3419 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 3420 { 3421 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3422 struct spdk_nvmf_rdma_transport, transport); 3423 3424 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3425 3426 /* nvmf_rdma_close_qpair is not called */ 3427 if (!rqpair->to_close) { 3428 return; 3429 } 3430 3431 /* device is already destroyed and we should force destroy this qpair. */ 3432 if (rqpair->poller && rqpair->poller->need_destroy) { 3433 nvmf_rdma_qpair_destroy(rqpair); 3434 return; 3435 } 3436 3437 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 3438 if (rqpair->current_send_depth != 0) { 3439 return; 3440 } 3441 3442 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 3443 return; 3444 } 3445 3446 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 3447 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 3448 return; 3449 } 3450 3451 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 3452 3453 nvmf_rdma_qpair_destroy(rqpair); 3454 } 3455 3456 static int 3457 nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked) 3458 { 3459 struct spdk_nvmf_qpair *qpair; 3460 struct spdk_nvmf_rdma_qpair *rqpair; 3461 3462 if (evt->id == NULL) { 3463 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 3464 return -1; 3465 } 3466 3467 qpair = evt->id->context; 3468 if (qpair == NULL) { 3469 SPDK_ERRLOG("disconnect request: no active connection\n"); 3470 return -1; 3471 } 3472 3473 rdma_ack_cm_event(evt); 3474 *event_acked = true; 3475 3476 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3477 3478 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair); 3479 3480 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3481 3482 return 0; 3483 } 3484 3485 #ifdef DEBUG 3486 static const char *CM_EVENT_STR[] = { 3487 "RDMA_CM_EVENT_ADDR_RESOLVED", 3488 "RDMA_CM_EVENT_ADDR_ERROR", 3489 "RDMA_CM_EVENT_ROUTE_RESOLVED", 3490 "RDMA_CM_EVENT_ROUTE_ERROR", 3491 "RDMA_CM_EVENT_CONNECT_REQUEST", 3492 "RDMA_CM_EVENT_CONNECT_RESPONSE", 3493 "RDMA_CM_EVENT_CONNECT_ERROR", 3494 "RDMA_CM_EVENT_UNREACHABLE", 3495 "RDMA_CM_EVENT_REJECTED", 3496 "RDMA_CM_EVENT_ESTABLISHED", 3497 "RDMA_CM_EVENT_DISCONNECTED", 3498 "RDMA_CM_EVENT_DEVICE_REMOVAL", 3499 "RDMA_CM_EVENT_MULTICAST_JOIN", 3500 "RDMA_CM_EVENT_MULTICAST_ERROR", 3501 "RDMA_CM_EVENT_ADDR_CHANGE", 3502 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 3503 }; 3504 #endif /* DEBUG */ 3505 3506 static void 3507 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 3508 struct spdk_nvmf_rdma_port *port) 3509 { 3510 struct spdk_nvmf_rdma_poll_group *rgroup; 3511 struct spdk_nvmf_rdma_poller *rpoller; 3512 struct spdk_nvmf_rdma_qpair *rqpair; 3513 3514 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3515 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3516 RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) { 3517 if (rqpair->listen_id == port->id) { 3518 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3519 } 3520 } 3521 } 3522 } 3523 } 3524 3525 static bool 3526 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 3527 struct rdma_cm_event *event) 3528 { 3529 const struct spdk_nvme_transport_id *trid; 3530 struct spdk_nvmf_rdma_port *port; 3531 struct spdk_nvmf_rdma_transport *rtransport; 3532 bool event_acked = false; 3533 3534 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3535 TAILQ_FOREACH(port, &rtransport->ports, link) { 3536 if (port->id == event->id) { 3537 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 3538 rdma_ack_cm_event(event); 3539 event_acked = true; 3540 trid = port->trid; 3541 break; 3542 } 3543 } 3544 3545 if (event_acked) { 3546 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 3547 3548 nvmf_rdma_stop_listen(transport, trid); 3549 nvmf_rdma_listen(transport, trid, NULL); 3550 } 3551 3552 return event_acked; 3553 } 3554 3555 static void 3556 nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3557 struct spdk_nvmf_rdma_device *device) 3558 { 3559 struct spdk_nvmf_rdma_port *port, *port_tmp; 3560 int rc; 3561 bool has_inflight; 3562 3563 rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false); 3564 if (rc) { 3565 SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc); 3566 return; 3567 } 3568 3569 if (!has_inflight) { 3570 /* no pollers, destroy the device */ 3571 device->ready_to_destroy = true; 3572 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport); 3573 } 3574 3575 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 3576 if (port->device == device) { 3577 SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n", 3578 port->trid->traddr, 3579 port->trid->trsvcid, 3580 ibv_get_device_name(port->device->context->device)); 3581 3582 /* keep NVMF listener and only destroy structures of the 3583 * RDMA transport. when the device comes back we can retry listening 3584 * and the application's workflow will not be interrupted. 3585 */ 3586 nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true); 3587 } 3588 } 3589 } 3590 3591 static void 3592 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 3593 struct rdma_cm_event *event) 3594 { 3595 struct spdk_nvmf_rdma_port *port, *tmp_port; 3596 struct spdk_nvmf_rdma_transport *rtransport; 3597 3598 port = event->id->context; 3599 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3600 3601 rdma_ack_cm_event(event); 3602 3603 /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive 3604 * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that 3605 * we are handling a port event here. 3606 */ 3607 TAILQ_FOREACH(tmp_port, &rtransport->ports, link) { 3608 if (port == tmp_port && port->device && !port->device->need_destroy) { 3609 port->device->need_destroy = true; 3610 nvmf_rdma_handle_device_removal(rtransport, port->device); 3611 } 3612 } 3613 } 3614 3615 static void 3616 nvmf_process_cm_events(struct spdk_nvmf_transport *transport, uint32_t max_events) 3617 { 3618 struct spdk_nvmf_rdma_transport *rtransport; 3619 struct rdma_cm_event *event; 3620 uint32_t i; 3621 int rc; 3622 bool event_acked; 3623 3624 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3625 3626 if (rtransport->event_channel == NULL) { 3627 return; 3628 } 3629 3630 for (i = 0; i < max_events; i++) { 3631 event_acked = false; 3632 rc = rdma_get_cm_event(rtransport->event_channel, &event); 3633 if (rc) { 3634 if (errno != EAGAIN && errno != EWOULDBLOCK) { 3635 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 3636 } 3637 break; 3638 } 3639 3640 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 3641 3642 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 3643 3644 switch (event->event) { 3645 case RDMA_CM_EVENT_ADDR_RESOLVED: 3646 case RDMA_CM_EVENT_ADDR_ERROR: 3647 case RDMA_CM_EVENT_ROUTE_RESOLVED: 3648 case RDMA_CM_EVENT_ROUTE_ERROR: 3649 /* No action required. The target never attempts to resolve routes. */ 3650 break; 3651 case RDMA_CM_EVENT_CONNECT_REQUEST: 3652 rc = nvmf_rdma_connect(transport, event); 3653 if (rc < 0) { 3654 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 3655 break; 3656 } 3657 break; 3658 case RDMA_CM_EVENT_CONNECT_RESPONSE: 3659 /* The target never initiates a new connection. So this will not occur. */ 3660 break; 3661 case RDMA_CM_EVENT_CONNECT_ERROR: 3662 /* Can this happen? The docs say it can, but not sure what causes it. */ 3663 break; 3664 case RDMA_CM_EVENT_UNREACHABLE: 3665 case RDMA_CM_EVENT_REJECTED: 3666 /* These only occur on the client side. */ 3667 break; 3668 case RDMA_CM_EVENT_ESTABLISHED: 3669 /* TODO: Should we be waiting for this event anywhere? */ 3670 break; 3671 case RDMA_CM_EVENT_DISCONNECTED: 3672 rc = nvmf_rdma_disconnect(event, &event_acked); 3673 if (rc < 0) { 3674 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3675 break; 3676 } 3677 break; 3678 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3679 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3680 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3681 * Once these events are sent to SPDK, we should release all IB resources and 3682 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3683 * ibv_destroy* functions to release user space memory allocated by IB. All kernel 3684 * resources are already cleaned. */ 3685 if (event->id->qp) { 3686 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3687 * corresponding qpair. Otherwise the event refers to a listening device. */ 3688 rc = nvmf_rdma_disconnect(event, &event_acked); 3689 if (rc < 0) { 3690 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3691 break; 3692 } 3693 } else { 3694 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3695 event_acked = true; 3696 } 3697 break; 3698 case RDMA_CM_EVENT_MULTICAST_JOIN: 3699 case RDMA_CM_EVENT_MULTICAST_ERROR: 3700 /* Multicast is not used */ 3701 break; 3702 case RDMA_CM_EVENT_ADDR_CHANGE: 3703 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3704 break; 3705 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3706 /* For now, do nothing. The target never re-uses queue pairs. */ 3707 break; 3708 default: 3709 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3710 break; 3711 } 3712 if (!event_acked) { 3713 rdma_ack_cm_event(event); 3714 } 3715 } 3716 } 3717 3718 static void 3719 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3720 { 3721 rqpair->last_wqe_reached = true; 3722 nvmf_rdma_destroy_drained_qpair(rqpair); 3723 } 3724 3725 static void 3726 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3727 { 3728 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3729 3730 if (event_ctx->rqpair) { 3731 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3732 if (event_ctx->cb_fn) { 3733 event_ctx->cb_fn(event_ctx->rqpair); 3734 } 3735 } 3736 free(event_ctx); 3737 } 3738 3739 static int 3740 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3741 spdk_nvmf_rdma_qpair_ibv_event fn) 3742 { 3743 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3744 struct spdk_thread *thr = NULL; 3745 int rc; 3746 3747 if (rqpair->qpair.group) { 3748 thr = rqpair->qpair.group->thread; 3749 } else if (rqpair->destruct_channel) { 3750 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3751 } 3752 3753 if (!thr) { 3754 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3755 return -EINVAL; 3756 } 3757 3758 ctx = calloc(1, sizeof(*ctx)); 3759 if (!ctx) { 3760 return -ENOMEM; 3761 } 3762 3763 ctx->rqpair = rqpair; 3764 ctx->cb_fn = fn; 3765 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3766 3767 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3768 if (rc) { 3769 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3770 free(ctx); 3771 } 3772 3773 return rc; 3774 } 3775 3776 static int 3777 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3778 { 3779 int rc; 3780 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3781 struct ibv_async_event event; 3782 3783 rc = ibv_get_async_event(device->context, &event); 3784 3785 if (rc) { 3786 /* In non-blocking mode -1 means there are no events available */ 3787 return rc; 3788 } 3789 3790 switch (event.event_type) { 3791 case IBV_EVENT_QP_FATAL: 3792 case IBV_EVENT_QP_LAST_WQE_REACHED: 3793 case IBV_EVENT_QP_REQ_ERR: 3794 case IBV_EVENT_QP_ACCESS_ERR: 3795 case IBV_EVENT_COMM_EST: 3796 case IBV_EVENT_PATH_MIG: 3797 case IBV_EVENT_PATH_MIG_ERR: 3798 rqpair = event.element.qp->qp_context; 3799 if (!rqpair) { 3800 /* Any QP event for NVMe-RDMA initiator may be returned. */ 3801 SPDK_NOTICELOG("Async QP event for unknown QP: %s\n", 3802 ibv_event_type_str(event.event_type)); 3803 break; 3804 } 3805 3806 switch (event.event_type) { 3807 case IBV_EVENT_QP_FATAL: 3808 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3809 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3810 (uintptr_t)rqpair, event.event_type); 3811 rqpair->ibv_in_error_state = true; 3812 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3813 break; 3814 case IBV_EVENT_QP_LAST_WQE_REACHED: 3815 /* This event only occurs for shared receive queues. */ 3816 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3817 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3818 if (rc) { 3819 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3820 rqpair->last_wqe_reached = true; 3821 } 3822 break; 3823 case IBV_EVENT_QP_REQ_ERR: 3824 case IBV_EVENT_QP_ACCESS_ERR: 3825 case IBV_EVENT_COMM_EST: 3826 case IBV_EVENT_PATH_MIG: 3827 case IBV_EVENT_PATH_MIG_ERR: 3828 SPDK_NOTICELOG("Async QP event: %s\n", 3829 ibv_event_type_str(event.event_type)); 3830 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3831 (uintptr_t)rqpair, event.event_type); 3832 rqpair->ibv_in_error_state = true; 3833 break; 3834 default: 3835 break; 3836 } 3837 break; 3838 case IBV_EVENT_DEVICE_FATAL: 3839 SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n", 3840 ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device); 3841 device->need_destroy = true; 3842 break; 3843 case IBV_EVENT_CQ_ERR: 3844 case IBV_EVENT_PORT_ACTIVE: 3845 case IBV_EVENT_PORT_ERR: 3846 case IBV_EVENT_LID_CHANGE: 3847 case IBV_EVENT_PKEY_CHANGE: 3848 case IBV_EVENT_SM_CHANGE: 3849 case IBV_EVENT_SRQ_ERR: 3850 case IBV_EVENT_SRQ_LIMIT_REACHED: 3851 case IBV_EVENT_CLIENT_REREGISTER: 3852 case IBV_EVENT_GID_CHANGE: 3853 case IBV_EVENT_SQ_DRAINED: 3854 default: 3855 SPDK_NOTICELOG("Async event: %s\n", 3856 ibv_event_type_str(event.event_type)); 3857 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3858 break; 3859 } 3860 ibv_ack_async_event(&event); 3861 3862 return 0; 3863 } 3864 3865 static void 3866 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3867 { 3868 int rc = 0; 3869 uint32_t i = 0; 3870 3871 for (i = 0; i < max_events; i++) { 3872 rc = nvmf_process_ib_event(device); 3873 if (rc) { 3874 break; 3875 } 3876 } 3877 3878 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3879 } 3880 3881 static int 3882 nvmf_rdma_accept(void *ctx) 3883 { 3884 int nfds, i = 0; 3885 struct spdk_nvmf_transport *transport = ctx; 3886 struct spdk_nvmf_rdma_transport *rtransport; 3887 struct spdk_nvmf_rdma_device *device, *tmp; 3888 uint32_t count; 3889 short revents; 3890 bool do_retry; 3891 3892 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3893 do_retry = nvmf_rdma_retry_listen_port(rtransport); 3894 3895 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3896 3897 if (nfds <= 0) { 3898 return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3899 } 3900 3901 /* The first poll descriptor is RDMA CM event */ 3902 if (rtransport->poll_fds[i++].revents & POLLIN) { 3903 nvmf_process_cm_events(transport, NVMF_RDMA_MAX_EVENTS_PER_POLL); 3904 nfds--; 3905 } 3906 3907 if (nfds == 0) { 3908 return SPDK_POLLER_BUSY; 3909 } 3910 3911 /* Second and subsequent poll descriptors are IB async events */ 3912 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3913 revents = rtransport->poll_fds[i++].revents; 3914 if (revents & POLLIN) { 3915 if (spdk_likely(!device->need_destroy)) { 3916 nvmf_process_ib_events(device, NVMF_RDMA_MAX_EVENTS_PER_POLL); 3917 if (spdk_unlikely(device->need_destroy)) { 3918 nvmf_rdma_handle_device_removal(rtransport, device); 3919 } 3920 } 3921 nfds--; 3922 } else if (revents & POLLNVAL || revents & POLLHUP) { 3923 SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device); 3924 nfds--; 3925 } 3926 } 3927 /* check all flagged fd's have been served */ 3928 assert(nfds == 0); 3929 3930 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3931 } 3932 3933 static void 3934 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3935 struct spdk_nvmf_ctrlr_data *cdata) 3936 { 3937 cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD; 3938 3939 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3940 since in-capsule data only works with NVME drives that support SGL memory layout */ 3941 if (transport->opts.dif_insert_or_strip) { 3942 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3943 } 3944 3945 if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) { 3946 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule" 3947 " data is greater than 4KiB.\n"); 3948 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux " 3949 "kernel between versions 5.4 and 5.12 data corruption may occur for " 3950 "writes that are not a multiple of 4KiB in size.\n"); 3951 } 3952 } 3953 3954 static void 3955 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3956 struct spdk_nvme_transport_id *trid, 3957 struct spdk_nvmf_discovery_log_page_entry *entry) 3958 { 3959 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3960 entry->adrfam = trid->adrfam; 3961 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3962 3963 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3964 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3965 3966 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3967 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3968 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3969 } 3970 3971 static int 3972 nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport, 3973 struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device, 3974 struct spdk_nvmf_rdma_poller **out_poller) 3975 { 3976 struct spdk_nvmf_rdma_poller *poller; 3977 struct spdk_rdma_srq_init_attr srq_init_attr; 3978 struct spdk_nvmf_rdma_resource_opts opts; 3979 int num_cqe; 3980 3981 poller = calloc(1, sizeof(*poller)); 3982 if (!poller) { 3983 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3984 return -1; 3985 } 3986 3987 poller->device = device; 3988 poller->group = rgroup; 3989 *out_poller = poller; 3990 3991 RB_INIT(&poller->qpairs); 3992 STAILQ_INIT(&poller->qpairs_pending_send); 3993 STAILQ_INIT(&poller->qpairs_pending_recv); 3994 3995 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3996 SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup); 3997 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3998 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) { 3999 SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n", 4000 rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr); 4001 } 4002 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr); 4003 4004 device->num_srq++; 4005 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 4006 srq_init_attr.pd = device->pd; 4007 srq_init_attr.stats = &poller->stat.qp_stats.recv; 4008 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 4009 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 4010 poller->srq = spdk_rdma_srq_create(&srq_init_attr); 4011 if (!poller->srq) { 4012 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 4013 return -1; 4014 } 4015 4016 opts.qp = poller->srq; 4017 opts.map = device->map; 4018 opts.qpair = NULL; 4019 opts.shared = true; 4020 opts.max_queue_depth = poller->max_srq_depth; 4021 opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size; 4022 4023 poller->resources = nvmf_rdma_resources_create(&opts); 4024 if (!poller->resources) { 4025 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 4026 return -1; 4027 } 4028 } 4029 4030 /* 4031 * When using an srq, we can limit the completion queue at startup. 4032 * The following formula represents the calculation: 4033 * num_cqe = num_recv + num_data_wr + num_send_wr. 4034 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 4035 */ 4036 if (poller->srq) { 4037 num_cqe = poller->max_srq_depth * 3; 4038 } else { 4039 num_cqe = rtransport->rdma_opts.num_cqe; 4040 } 4041 4042 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 4043 if (!poller->cq) { 4044 SPDK_ERRLOG("Unable to create completion queue\n"); 4045 return -1; 4046 } 4047 poller->num_cqe = num_cqe; 4048 return 0; 4049 } 4050 4051 static void 4052 _nvmf_rdma_register_poller_in_group(void *c) 4053 { 4054 struct spdk_nvmf_rdma_poller *poller; 4055 struct poller_manage_ctx *ctx = c; 4056 struct spdk_nvmf_rdma_device *device; 4057 int rc; 4058 4059 rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller); 4060 if (rc < 0 && poller) { 4061 nvmf_rdma_poller_destroy(poller); 4062 } 4063 4064 device = ctx->device; 4065 if (nvmf_rdma_all_pollers_management_done(ctx)) { 4066 device->is_ready = true; 4067 } 4068 } 4069 4070 static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 4071 4072 static struct spdk_nvmf_transport_poll_group * 4073 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport, 4074 struct spdk_nvmf_poll_group *group) 4075 { 4076 struct spdk_nvmf_rdma_transport *rtransport; 4077 struct spdk_nvmf_rdma_poll_group *rgroup; 4078 struct spdk_nvmf_rdma_poller *poller; 4079 struct spdk_nvmf_rdma_device *device; 4080 int rc; 4081 4082 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 4083 4084 rgroup = calloc(1, sizeof(*rgroup)); 4085 if (!rgroup) { 4086 return NULL; 4087 } 4088 4089 TAILQ_INIT(&rgroup->pollers); 4090 4091 TAILQ_FOREACH(device, &rtransport->devices, link) { 4092 rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller); 4093 if (rc < 0) { 4094 nvmf_rdma_poll_group_destroy(&rgroup->group); 4095 return NULL; 4096 } 4097 } 4098 4099 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 4100 if (rtransport->conn_sched.next_admin_pg == NULL) { 4101 rtransport->conn_sched.next_admin_pg = rgroup; 4102 rtransport->conn_sched.next_io_pg = rgroup; 4103 } 4104 4105 return &rgroup->group; 4106 } 4107 4108 static uint32_t 4109 nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg) 4110 { 4111 uint32_t count; 4112 4113 /* Just assume that unassociated qpairs will eventually be io 4114 * qpairs. This is close enough for the use cases for this 4115 * function. 4116 */ 4117 pthread_mutex_lock(&pg->mutex); 4118 count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs; 4119 pthread_mutex_unlock(&pg->mutex); 4120 4121 return count; 4122 } 4123 4124 static struct spdk_nvmf_transport_poll_group * 4125 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 4126 { 4127 struct spdk_nvmf_rdma_transport *rtransport; 4128 struct spdk_nvmf_rdma_poll_group **pg; 4129 struct spdk_nvmf_transport_poll_group *result; 4130 uint32_t count; 4131 4132 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4133 4134 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 4135 return NULL; 4136 } 4137 4138 if (qpair->qid == 0) { 4139 pg = &rtransport->conn_sched.next_admin_pg; 4140 } else { 4141 struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current; 4142 uint32_t min_value; 4143 4144 pg = &rtransport->conn_sched.next_io_pg; 4145 pg_min = *pg; 4146 pg_start = *pg; 4147 pg_current = *pg; 4148 min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 4149 4150 while (1) { 4151 count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 4152 4153 if (count < min_value) { 4154 min_value = count; 4155 pg_min = pg_current; 4156 } 4157 4158 pg_current = TAILQ_NEXT(pg_current, link); 4159 if (pg_current == NULL) { 4160 pg_current = TAILQ_FIRST(&rtransport->poll_groups); 4161 } 4162 4163 if (pg_current == pg_start || min_value == 0) { 4164 break; 4165 } 4166 } 4167 *pg = pg_min; 4168 } 4169 4170 assert(*pg != NULL); 4171 4172 result = &(*pg)->group; 4173 4174 *pg = TAILQ_NEXT(*pg, link); 4175 if (*pg == NULL) { 4176 *pg = TAILQ_FIRST(&rtransport->poll_groups); 4177 } 4178 4179 return result; 4180 } 4181 4182 static void 4183 nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller) 4184 { 4185 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 4186 int rc; 4187 4188 TAILQ_REMOVE(&poller->group->pollers, poller, link); 4189 RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) { 4190 nvmf_rdma_qpair_destroy(qpair); 4191 } 4192 4193 if (poller->srq) { 4194 if (poller->resources) { 4195 nvmf_rdma_resources_destroy(poller->resources); 4196 } 4197 spdk_rdma_srq_destroy(poller->srq); 4198 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 4199 } 4200 4201 if (poller->cq) { 4202 rc = ibv_destroy_cq(poller->cq); 4203 if (rc != 0) { 4204 SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno)); 4205 } 4206 } 4207 4208 if (poller->destroy_cb) { 4209 poller->destroy_cb(poller->destroy_cb_ctx); 4210 poller->destroy_cb = NULL; 4211 } 4212 4213 free(poller); 4214 } 4215 4216 static void 4217 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 4218 { 4219 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 4220 struct spdk_nvmf_rdma_poller *poller, *tmp; 4221 struct spdk_nvmf_rdma_transport *rtransport; 4222 4223 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4224 if (!rgroup) { 4225 return; 4226 } 4227 4228 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 4229 nvmf_rdma_poller_destroy(poller); 4230 } 4231 4232 if (rgroup->group.transport == NULL) { 4233 /* Transport can be NULL when nvmf_rdma_poll_group_create() 4234 * calls this function directly in a failure path. */ 4235 free(rgroup); 4236 return; 4237 } 4238 4239 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 4240 4241 next_rgroup = TAILQ_NEXT(rgroup, link); 4242 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 4243 if (next_rgroup == NULL) { 4244 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 4245 } 4246 if (rtransport->conn_sched.next_admin_pg == rgroup) { 4247 rtransport->conn_sched.next_admin_pg = next_rgroup; 4248 } 4249 if (rtransport->conn_sched.next_io_pg == rgroup) { 4250 rtransport->conn_sched.next_io_pg = next_rgroup; 4251 } 4252 4253 free(rgroup); 4254 } 4255 4256 static void 4257 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 4258 { 4259 if (rqpair->cm_id != NULL) { 4260 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 4261 } 4262 } 4263 4264 static int 4265 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 4266 struct spdk_nvmf_qpair *qpair) 4267 { 4268 struct spdk_nvmf_rdma_poll_group *rgroup; 4269 struct spdk_nvmf_rdma_qpair *rqpair; 4270 struct spdk_nvmf_rdma_device *device; 4271 struct spdk_nvmf_rdma_poller *poller; 4272 int rc; 4273 4274 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4275 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4276 4277 device = rqpair->device; 4278 4279 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 4280 if (poller->device == device) { 4281 break; 4282 } 4283 } 4284 4285 if (!poller) { 4286 SPDK_ERRLOG("No poller found for device.\n"); 4287 return -1; 4288 } 4289 4290 if (poller->need_destroy) { 4291 SPDK_ERRLOG("Poller is destroying.\n"); 4292 return -1; 4293 } 4294 4295 rqpair->poller = poller; 4296 rqpair->srq = rqpair->poller->srq; 4297 4298 rc = nvmf_rdma_qpair_initialize(qpair); 4299 if (rc < 0) { 4300 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 4301 rqpair->poller = NULL; 4302 rqpair->srq = NULL; 4303 return -1; 4304 } 4305 4306 RB_INSERT(qpairs_tree, &poller->qpairs, rqpair); 4307 4308 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 4309 if (rc) { 4310 /* Try to reject, but we probably can't */ 4311 nvmf_rdma_qpair_reject_connection(rqpair); 4312 return -1; 4313 } 4314 4315 return 0; 4316 } 4317 4318 static int 4319 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 4320 struct spdk_nvmf_qpair *qpair) 4321 { 4322 struct spdk_nvmf_rdma_qpair *rqpair; 4323 4324 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4325 assert(group->transport->tgt != NULL); 4326 4327 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 4328 4329 if (!rqpair->destruct_channel) { 4330 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 4331 return 0; 4332 } 4333 4334 /* Sanity check that we get io_channel on the correct thread */ 4335 if (qpair->group) { 4336 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 4337 } 4338 4339 return 0; 4340 } 4341 4342 static int 4343 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 4344 { 4345 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 4346 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4347 struct spdk_nvmf_rdma_transport, transport); 4348 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4349 struct spdk_nvmf_rdma_qpair, qpair); 4350 4351 /* 4352 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 4353 * needs to be returned to the shared receive queue or the poll group will eventually be 4354 * starved of RECV structures. 4355 */ 4356 if (rqpair->srq && rdma_req->recv) { 4357 int rc; 4358 struct ibv_recv_wr *bad_recv_wr; 4359 4360 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 4361 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 4362 if (rc) { 4363 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 4364 } 4365 } 4366 4367 _nvmf_rdma_request_free(rdma_req, rtransport); 4368 return 0; 4369 } 4370 4371 static int 4372 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 4373 { 4374 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4375 struct spdk_nvmf_rdma_transport, transport); 4376 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 4377 struct spdk_nvmf_rdma_request, req); 4378 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4379 struct spdk_nvmf_rdma_qpair, qpair); 4380 4381 if (spdk_unlikely(rqpair->ibv_in_error_state)) { 4382 /* The connection is dead. Move the request directly to the completed state. */ 4383 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4384 } else { 4385 /* The connection is alive, so process the request as normal */ 4386 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 4387 } 4388 4389 nvmf_rdma_request_process(rtransport, rdma_req); 4390 4391 return 0; 4392 } 4393 4394 static void 4395 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 4396 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 4397 { 4398 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4399 4400 rqpair->to_close = true; 4401 4402 /* This happens only when the qpair is disconnected before 4403 * it is added to the poll group. Since there is no poll group, 4404 * the RDMA qp has not been initialized yet and the RDMA CM 4405 * event has not yet been acknowledged, so we need to reject it. 4406 */ 4407 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 4408 nvmf_rdma_qpair_reject_connection(rqpair); 4409 nvmf_rdma_qpair_destroy(rqpair); 4410 return; 4411 } 4412 4413 if (rqpair->rdma_qp) { 4414 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 4415 } 4416 4417 nvmf_rdma_destroy_drained_qpair(rqpair); 4418 4419 if (cb_fn) { 4420 cb_fn(cb_arg); 4421 } 4422 } 4423 4424 static struct spdk_nvmf_rdma_qpair * 4425 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 4426 { 4427 struct spdk_nvmf_rdma_qpair find; 4428 4429 find.qp_num = wc->qp_num; 4430 4431 return RB_FIND(qpairs_tree, &rpoller->qpairs, &find); 4432 } 4433 4434 #ifdef DEBUG 4435 static int 4436 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 4437 { 4438 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 4439 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 4440 } 4441 #endif 4442 4443 static void 4444 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 4445 int rc) 4446 { 4447 struct spdk_nvmf_rdma_recv *rdma_recv; 4448 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4449 4450 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 4451 while (bad_recv_wr != NULL) { 4452 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 4453 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4454 4455 rdma_recv->qpair->current_recv_depth++; 4456 bad_recv_wr = bad_recv_wr->next; 4457 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 4458 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair); 4459 } 4460 } 4461 4462 static void 4463 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 4464 { 4465 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 4466 while (bad_recv_wr != NULL) { 4467 bad_recv_wr = bad_recv_wr->next; 4468 rqpair->current_recv_depth++; 4469 } 4470 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4471 } 4472 4473 static void 4474 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 4475 struct spdk_nvmf_rdma_poller *rpoller) 4476 { 4477 struct spdk_nvmf_rdma_qpair *rqpair; 4478 struct ibv_recv_wr *bad_recv_wr; 4479 int rc; 4480 4481 if (rpoller->srq) { 4482 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 4483 if (spdk_unlikely(rc)) { 4484 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 4485 } 4486 } else { 4487 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 4488 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 4489 rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 4490 if (spdk_unlikely(rc)) { 4491 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 4492 } 4493 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 4494 } 4495 } 4496 } 4497 4498 static void 4499 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 4500 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 4501 { 4502 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4503 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 4504 4505 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 4506 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 4507 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 4508 assert(rqpair->current_send_depth > 0); 4509 rqpair->current_send_depth--; 4510 switch (bad_rdma_wr->type) { 4511 case RDMA_WR_TYPE_DATA: 4512 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4513 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 4514 assert(rqpair->current_read_depth > 0); 4515 rqpair->current_read_depth--; 4516 } 4517 break; 4518 case RDMA_WR_TYPE_SEND: 4519 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4520 break; 4521 default: 4522 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 4523 prev_rdma_req = cur_rdma_req; 4524 continue; 4525 } 4526 4527 if (prev_rdma_req == cur_rdma_req) { 4528 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 4529 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 4530 continue; 4531 } 4532 4533 switch (cur_rdma_req->state) { 4534 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4535 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 4536 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, cur_rdma_req, state_link); 4537 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 4538 break; 4539 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 4540 case RDMA_REQUEST_STATE_COMPLETING: 4541 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4542 break; 4543 default: 4544 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 4545 cur_rdma_req->state, rqpair); 4546 continue; 4547 } 4548 4549 nvmf_rdma_request_process(rtransport, cur_rdma_req); 4550 prev_rdma_req = cur_rdma_req; 4551 } 4552 4553 if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) { 4554 /* Disconnect the connection. */ 4555 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4556 } 4557 4558 } 4559 4560 static void 4561 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 4562 struct spdk_nvmf_rdma_poller *rpoller) 4563 { 4564 struct spdk_nvmf_rdma_qpair *rqpair; 4565 struct ibv_send_wr *bad_wr = NULL; 4566 int rc; 4567 4568 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 4569 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 4570 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 4571 4572 /* bad wr always points to the first wr that failed. */ 4573 if (spdk_unlikely(rc)) { 4574 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 4575 } 4576 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 4577 } 4578 } 4579 4580 static const char * 4581 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 4582 { 4583 switch (wr_type) { 4584 case RDMA_WR_TYPE_RECV: 4585 return "RECV"; 4586 case RDMA_WR_TYPE_SEND: 4587 return "SEND"; 4588 case RDMA_WR_TYPE_DATA: 4589 return "DATA"; 4590 default: 4591 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 4592 SPDK_UNREACHABLE(); 4593 } 4594 } 4595 4596 static inline void 4597 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 4598 { 4599 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 4600 4601 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 4602 /* If qpair is in ERR state, we will receive completions for all posted and not completed 4603 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 4604 SPDK_DEBUGLOG(rdma, 4605 "Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n", 4606 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id, 4607 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4608 } else { 4609 SPDK_ERRLOG("Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n", 4610 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id, 4611 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4612 } 4613 } 4614 4615 static int 4616 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 4617 struct spdk_nvmf_rdma_poller *rpoller) 4618 { 4619 struct ibv_wc wc[32]; 4620 struct spdk_nvmf_rdma_wr *rdma_wr; 4621 struct spdk_nvmf_rdma_request *rdma_req; 4622 struct spdk_nvmf_rdma_recv *rdma_recv; 4623 struct spdk_nvmf_rdma_qpair *rqpair, *tmp_rqpair; 4624 int reaped, i; 4625 int count = 0; 4626 int rc; 4627 bool error = false; 4628 uint64_t poll_tsc = spdk_get_ticks(); 4629 4630 if (spdk_unlikely(rpoller->need_destroy)) { 4631 /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not 4632 * be called because we cannot poll anything from cq. So we call that here to force 4633 * destroy the qpair after to_close turning true. 4634 */ 4635 RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) { 4636 nvmf_rdma_destroy_drained_qpair(rqpair); 4637 } 4638 return 0; 4639 } 4640 4641 /* Poll for completing operations. */ 4642 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 4643 if (spdk_unlikely(reaped < 0)) { 4644 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 4645 errno, spdk_strerror(errno)); 4646 return -1; 4647 } else if (reaped == 0) { 4648 rpoller->stat.idle_polls++; 4649 } 4650 4651 rpoller->stat.polls++; 4652 rpoller->stat.completions += reaped; 4653 4654 for (i = 0; i < reaped; i++) { 4655 4656 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 4657 4658 switch (rdma_wr->type) { 4659 case RDMA_WR_TYPE_SEND: 4660 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4661 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4662 4663 if (spdk_likely(!wc[i].status)) { 4664 count++; 4665 assert(wc[i].opcode == IBV_WC_SEND); 4666 assert(nvmf_rdma_req_is_completing(rdma_req)); 4667 } 4668 4669 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4670 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 4671 assert(rqpair->current_send_depth >= (uint32_t)rdma_req->num_outstanding_data_wr + 1); 4672 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 4673 rdma_req->num_outstanding_data_wr = 0; 4674 4675 nvmf_rdma_request_process(rtransport, rdma_req); 4676 break; 4677 case RDMA_WR_TYPE_RECV: 4678 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 4679 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4680 if (rpoller->srq != NULL) { 4681 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 4682 /* It is possible that there are still some completions for destroyed QP 4683 * associated with SRQ. We just ignore these late completions and re-post 4684 * receive WRs back to SRQ. 4685 */ 4686 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 4687 struct ibv_recv_wr *bad_wr; 4688 4689 rdma_recv->wr.next = NULL; 4690 spdk_rdma_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 4691 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 4692 if (rc) { 4693 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 4694 } 4695 continue; 4696 } 4697 } 4698 rqpair = rdma_recv->qpair; 4699 4700 assert(rqpair != NULL); 4701 if (spdk_likely(!wc[i].status)) { 4702 assert(wc[i].opcode == IBV_WC_RECV); 4703 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 4704 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4705 break; 4706 } 4707 } 4708 4709 rdma_recv->wr.next = NULL; 4710 rqpair->current_recv_depth++; 4711 rdma_recv->receive_tsc = poll_tsc; 4712 rpoller->stat.requests++; 4713 STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link); 4714 rqpair->qpair.queue_depth++; 4715 break; 4716 case RDMA_WR_TYPE_DATA: 4717 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4718 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4719 4720 assert(rdma_req->num_outstanding_data_wr > 0); 4721 4722 rqpair->current_send_depth--; 4723 rdma_req->num_outstanding_data_wr--; 4724 if (spdk_likely(!wc[i].status)) { 4725 assert(wc[i].opcode == IBV_WC_RDMA_READ); 4726 rqpair->current_read_depth--; 4727 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 4728 if (rdma_req->num_outstanding_data_wr == 0) { 4729 if (spdk_unlikely(rdma_req->num_remaining_data_wr)) { 4730 /* Only part of RDMA_READ operations was submitted, process the rest */ 4731 rc = nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport); 4732 if (spdk_likely(!rc)) { 4733 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 4734 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 4735 } else { 4736 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 4737 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 4738 rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 4739 } 4740 nvmf_rdma_request_process(rtransport, rdma_req); 4741 break; 4742 } 4743 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 4744 nvmf_rdma_request_process(rtransport, rdma_req); 4745 } 4746 } else { 4747 /* If the data transfer fails still force the queue into the error state, 4748 * if we were performing an RDMA_READ, we need to force the request into a 4749 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 4750 * case, we should wait for the SEND to complete. */ 4751 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 4752 rqpair->current_read_depth--; 4753 if (rdma_req->num_outstanding_data_wr == 0) { 4754 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4755 } 4756 } 4757 } 4758 break; 4759 default: 4760 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 4761 continue; 4762 } 4763 4764 /* Handle error conditions */ 4765 if (spdk_unlikely(wc[i].status)) { 4766 rqpair->ibv_in_error_state = true; 4767 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 4768 4769 error = true; 4770 4771 if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) { 4772 /* Disconnect the connection. */ 4773 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4774 } else { 4775 nvmf_rdma_destroy_drained_qpair(rqpair); 4776 } 4777 continue; 4778 } 4779 4780 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 4781 4782 if (spdk_unlikely(!spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 4783 nvmf_rdma_destroy_drained_qpair(rqpair); 4784 } 4785 } 4786 4787 if (spdk_unlikely(error == true)) { 4788 return -1; 4789 } 4790 4791 if (reaped == 0) { 4792 /* In some cases we may not receive any CQE but we still may have pending IO requests waiting for 4793 * a resource (e.g. a WR from the data_wr_pool). 4794 * We need to start processing of such requests if no CQE reaped */ 4795 nvmf_rdma_poller_process_pending_buf_queue(rtransport, rpoller); 4796 } 4797 4798 /* submit outstanding work requests. */ 4799 _poller_submit_recvs(rtransport, rpoller); 4800 _poller_submit_sends(rtransport, rpoller); 4801 4802 return count; 4803 } 4804 4805 static void 4806 _nvmf_rdma_remove_destroyed_device(void *c) 4807 { 4808 struct spdk_nvmf_rdma_transport *rtransport = c; 4809 struct spdk_nvmf_rdma_device *device, *device_tmp; 4810 int rc; 4811 4812 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 4813 if (device->ready_to_destroy) { 4814 destroy_ib_device(rtransport, device); 4815 } 4816 } 4817 4818 free_poll_fds(rtransport); 4819 rc = generate_poll_fds(rtransport); 4820 /* cannot handle fd allocation error here */ 4821 if (rc != 0) { 4822 SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n"); 4823 } 4824 } 4825 4826 static void 4827 _nvmf_rdma_remove_poller_in_group_cb(void *c) 4828 { 4829 struct poller_manage_ctx *ctx = c; 4830 struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport; 4831 struct spdk_nvmf_rdma_device *device = ctx->device; 4832 struct spdk_thread *thread = ctx->thread; 4833 4834 if (nvmf_rdma_all_pollers_management_done(c)) { 4835 /* destroy device when last poller is destroyed */ 4836 device->ready_to_destroy = true; 4837 spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport); 4838 } 4839 } 4840 4841 static void 4842 _nvmf_rdma_remove_poller_in_group(void *c) 4843 { 4844 struct poller_manage_ctx *ctx = c; 4845 4846 ctx->rpoller->need_destroy = true; 4847 ctx->rpoller->destroy_cb_ctx = ctx; 4848 ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb; 4849 4850 /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */ 4851 if (RB_EMPTY(&ctx->rpoller->qpairs)) { 4852 nvmf_rdma_poller_destroy(ctx->rpoller); 4853 } 4854 } 4855 4856 static int 4857 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 4858 { 4859 struct spdk_nvmf_rdma_transport *rtransport; 4860 struct spdk_nvmf_rdma_poll_group *rgroup; 4861 struct spdk_nvmf_rdma_poller *rpoller, *tmp; 4862 int count = 0, rc, rc2 = 0; 4863 4864 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 4865 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4866 4867 TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) { 4868 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4869 if (spdk_unlikely(rc < 0)) { 4870 if (rc2 == 0) { 4871 rc2 = rc; 4872 } 4873 continue; 4874 } 4875 count += rc; 4876 } 4877 4878 return rc2 ? rc2 : count; 4879 } 4880 4881 static int 4882 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4883 struct spdk_nvme_transport_id *trid, 4884 bool peer) 4885 { 4886 struct sockaddr *saddr; 4887 uint16_t port; 4888 4889 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4890 4891 if (peer) { 4892 saddr = rdma_get_peer_addr(id); 4893 } else { 4894 saddr = rdma_get_local_addr(id); 4895 } 4896 switch (saddr->sa_family) { 4897 case AF_INET: { 4898 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4899 4900 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4901 inet_ntop(AF_INET, &saddr_in->sin_addr, 4902 trid->traddr, sizeof(trid->traddr)); 4903 if (peer) { 4904 port = ntohs(rdma_get_dst_port(id)); 4905 } else { 4906 port = ntohs(rdma_get_src_port(id)); 4907 } 4908 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4909 break; 4910 } 4911 case AF_INET6: { 4912 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4913 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4914 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4915 trid->traddr, sizeof(trid->traddr)); 4916 if (peer) { 4917 port = ntohs(rdma_get_dst_port(id)); 4918 } else { 4919 port = ntohs(rdma_get_src_port(id)); 4920 } 4921 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4922 break; 4923 } 4924 default: 4925 return -1; 4926 4927 } 4928 4929 return 0; 4930 } 4931 4932 static int 4933 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4934 struct spdk_nvme_transport_id *trid) 4935 { 4936 struct spdk_nvmf_rdma_qpair *rqpair; 4937 4938 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4939 4940 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4941 } 4942 4943 static int 4944 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4945 struct spdk_nvme_transport_id *trid) 4946 { 4947 struct spdk_nvmf_rdma_qpair *rqpair; 4948 4949 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4950 4951 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4952 } 4953 4954 static int 4955 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4956 struct spdk_nvme_transport_id *trid) 4957 { 4958 struct spdk_nvmf_rdma_qpair *rqpair; 4959 4960 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4961 4962 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4963 } 4964 4965 void 4966 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4967 { 4968 g_nvmf_hooks = *hooks; 4969 } 4970 4971 static void 4972 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4973 struct spdk_nvmf_rdma_request *rdma_req_to_abort, 4974 struct spdk_nvmf_rdma_qpair *rqpair) 4975 { 4976 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4977 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4978 4979 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, state_link); 4980 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 4981 4982 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4983 } 4984 4985 static int 4986 _nvmf_rdma_qpair_abort_request(void *ctx) 4987 { 4988 struct spdk_nvmf_request *req = ctx; 4989 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4990 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4991 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4992 struct spdk_nvmf_rdma_qpair, qpair); 4993 int rc; 4994 4995 spdk_poller_unregister(&req->poller); 4996 4997 switch (rdma_req_to_abort->state) { 4998 case RDMA_REQUEST_STATE_EXECUTING: 4999 rc = nvmf_ctrlr_abort_request(req); 5000 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 5001 return SPDK_POLLER_BUSY; 5002 } 5003 break; 5004 5005 case RDMA_REQUEST_STATE_NEED_BUFFER: 5006 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 5007 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 5008 5009 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5010 break; 5011 5012 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 5013 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 5014 spdk_nvmf_rdma_request, state_link); 5015 5016 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5017 break; 5018 5019 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 5020 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 5021 spdk_nvmf_rdma_request, state_link); 5022 5023 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5024 break; 5025 5026 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 5027 /* Remove req from the list here to re-use common function */ 5028 STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, 5029 spdk_nvmf_rdma_request, state_link); 5030 5031 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5032 break; 5033 5034 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 5035 if (spdk_get_ticks() < req->timeout_tsc) { 5036 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 5037 return SPDK_POLLER_BUSY; 5038 } 5039 break; 5040 5041 default: 5042 break; 5043 } 5044 5045 spdk_nvmf_request_complete(req); 5046 return SPDK_POLLER_BUSY; 5047 } 5048 5049 static void 5050 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 5051 struct spdk_nvmf_request *req) 5052 { 5053 struct spdk_nvmf_rdma_qpair *rqpair; 5054 struct spdk_nvmf_rdma_transport *rtransport; 5055 struct spdk_nvmf_transport *transport; 5056 uint16_t cid; 5057 uint32_t i, max_req_count; 5058 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 5059 5060 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 5061 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 5062 transport = &rtransport->transport; 5063 5064 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 5065 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 5066 5067 for (i = 0; i < max_req_count; i++) { 5068 rdma_req = &rqpair->resources->reqs[i]; 5069 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 5070 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 5071 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 5072 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 5073 rdma_req->req.qpair == qpair) { 5074 rdma_req_to_abort = rdma_req; 5075 break; 5076 } 5077 } 5078 5079 if (rdma_req_to_abort == NULL) { 5080 spdk_nvmf_request_complete(req); 5081 return; 5082 } 5083 5084 req->req_to_abort = &rdma_req_to_abort->req; 5085 req->timeout_tsc = spdk_get_ticks() + 5086 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 5087 req->poller = NULL; 5088 5089 _nvmf_rdma_qpair_abort_request(req); 5090 } 5091 5092 static void 5093 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 5094 struct spdk_json_write_ctx *w) 5095 { 5096 struct spdk_nvmf_rdma_poll_group *rgroup; 5097 struct spdk_nvmf_rdma_poller *rpoller; 5098 5099 assert(w != NULL); 5100 5101 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 5102 5103 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 5104 5105 spdk_json_write_named_array_begin(w, "devices"); 5106 5107 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 5108 spdk_json_write_object_begin(w); 5109 spdk_json_write_named_string(w, "name", 5110 ibv_get_device_name(rpoller->device->context->device)); 5111 spdk_json_write_named_uint64(w, "polls", 5112 rpoller->stat.polls); 5113 spdk_json_write_named_uint64(w, "idle_polls", 5114 rpoller->stat.idle_polls); 5115 spdk_json_write_named_uint64(w, "completions", 5116 rpoller->stat.completions); 5117 spdk_json_write_named_uint64(w, "requests", 5118 rpoller->stat.requests); 5119 spdk_json_write_named_uint64(w, "request_latency", 5120 rpoller->stat.request_latency); 5121 spdk_json_write_named_uint64(w, "pending_free_request", 5122 rpoller->stat.pending_free_request); 5123 spdk_json_write_named_uint64(w, "pending_rdma_read", 5124 rpoller->stat.pending_rdma_read); 5125 spdk_json_write_named_uint64(w, "pending_rdma_write", 5126 rpoller->stat.pending_rdma_write); 5127 spdk_json_write_named_uint64(w, "pending_rdma_send", 5128 rpoller->stat.pending_rdma_send); 5129 spdk_json_write_named_uint64(w, "total_send_wrs", 5130 rpoller->stat.qp_stats.send.num_submitted_wrs); 5131 spdk_json_write_named_uint64(w, "send_doorbell_updates", 5132 rpoller->stat.qp_stats.send.doorbell_updates); 5133 spdk_json_write_named_uint64(w, "total_recv_wrs", 5134 rpoller->stat.qp_stats.recv.num_submitted_wrs); 5135 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 5136 rpoller->stat.qp_stats.recv.doorbell_updates); 5137 spdk_json_write_object_end(w); 5138 } 5139 5140 spdk_json_write_array_end(w); 5141 } 5142 5143 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 5144 .name = "RDMA", 5145 .type = SPDK_NVME_TRANSPORT_RDMA, 5146 .opts_init = nvmf_rdma_opts_init, 5147 .create = nvmf_rdma_create, 5148 .dump_opts = nvmf_rdma_dump_opts, 5149 .destroy = nvmf_rdma_destroy, 5150 5151 .listen = nvmf_rdma_listen, 5152 .stop_listen = nvmf_rdma_stop_listen, 5153 .cdata_init = nvmf_rdma_cdata_init, 5154 5155 .listener_discover = nvmf_rdma_discover, 5156 5157 .poll_group_create = nvmf_rdma_poll_group_create, 5158 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 5159 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 5160 .poll_group_add = nvmf_rdma_poll_group_add, 5161 .poll_group_remove = nvmf_rdma_poll_group_remove, 5162 .poll_group_poll = nvmf_rdma_poll_group_poll, 5163 5164 .req_free = nvmf_rdma_request_free, 5165 .req_complete = nvmf_rdma_request_complete, 5166 5167 .qpair_fini = nvmf_rdma_close_qpair, 5168 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 5169 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 5170 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 5171 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 5172 5173 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 5174 }; 5175 5176 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 5177 SPDK_LOG_REGISTER_COMPONENT(rdma) 5178