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