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