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_provider.h" 21 #include "spdk_internal/rdma_utils.h" 22 23 #include "nvmf_internal.h" 24 #include "transport.h" 25 26 #include "spdk_internal/trace_defs.h" 27 28 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 29 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 30 31 /* 32 RDMA Connection Resource Defaults 33 */ 34 #define NVMF_DEFAULT_MSDBD 16 35 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 36 #define NVMF_DEFAULT_RSP_SGE 1 37 #define NVMF_DEFAULT_RX_SGE 2 38 39 #define NVMF_RDMA_MAX_EVENTS_PER_POLL 32 40 41 SPDK_STATIC_ASSERT(NVMF_DEFAULT_MSDBD <= SPDK_NVMF_MAX_SGL_ENTRIES, 42 "MSDBD must not exceed SPDK_NVMF_MAX_SGL_ENTRIES"); 43 44 /* The RDMA completion queue size */ 45 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 46 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 47 48 enum spdk_nvmf_rdma_request_state { 49 /* The request is not currently in use */ 50 RDMA_REQUEST_STATE_FREE = 0, 51 52 /* Initial state when request first received */ 53 RDMA_REQUEST_STATE_NEW, 54 55 /* The request is queued until a data buffer is available. */ 56 RDMA_REQUEST_STATE_NEED_BUFFER, 57 58 /* The request is waiting on RDMA queue depth availability 59 * to transfer data from the host to the controller. 60 */ 61 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 62 63 /* The request is currently transferring data from the host to the controller. */ 64 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 65 66 /* The request is ready to execute at the block device */ 67 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 68 69 /* The request is currently executing at the block device */ 70 RDMA_REQUEST_STATE_EXECUTING, 71 72 /* The request finished executing at the block device */ 73 RDMA_REQUEST_STATE_EXECUTED, 74 75 /* The request is waiting on RDMA queue depth availability 76 * to transfer data from the controller to the host. 77 */ 78 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 79 80 /* The request is waiting on RDMA queue depth availability 81 * to send response to the host. 82 */ 83 RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 84 85 /* The request is ready to send a completion */ 86 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 87 88 /* The request is currently transferring data from the controller to the host. */ 89 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 90 91 /* The request currently has an outstanding completion without an 92 * associated data transfer. 93 */ 94 RDMA_REQUEST_STATE_COMPLETING, 95 96 /* The request completed and can be marked free. */ 97 RDMA_REQUEST_STATE_COMPLETED, 98 99 /* Terminator */ 100 RDMA_REQUEST_NUM_STATES, 101 }; 102 103 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 104 { 105 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 106 107 struct spdk_trace_tpoint_opts opts[] = { 108 { 109 "RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 110 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 1, 111 { 112 { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 }, 113 { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 } 114 } 115 }, 116 { 117 "RDMA_REQ_COMPLETED", TRACE_RDMA_REQUEST_STATE_COMPLETED, 118 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 119 { 120 { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 }, 121 { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 } 122 } 123 }, 124 }; 125 126 spdk_trace_register_description_ext(opts, SPDK_COUNTOF(opts)); 127 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 128 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 129 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 130 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 131 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 132 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 133 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 134 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 135 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 136 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 137 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 138 spdk_trace_register_description("RDMA_REQ_TX_H2C", 139 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 140 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 141 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 142 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 143 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 144 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 145 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 146 spdk_trace_register_description("RDMA_REQ_EXECUTING", 147 TRACE_RDMA_REQUEST_STATE_EXECUTING, 148 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 149 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 150 spdk_trace_register_description("RDMA_REQ_EXECUTED", 151 TRACE_RDMA_REQUEST_STATE_EXECUTED, 152 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 153 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 154 spdk_trace_register_description("RDMA_REQ_RDY2COMPL_PEND", 155 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 156 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 157 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 158 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 159 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 160 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 161 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 162 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 163 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 164 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 165 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 166 spdk_trace_register_description("RDMA_REQ_COMPLETING", 167 TRACE_RDMA_REQUEST_STATE_COMPLETING, 168 OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0, 169 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 170 171 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 172 OWNER_TYPE_NONE, OBJECT_NONE, 0, 173 SPDK_TRACE_ARG_TYPE_INT, ""); 174 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 175 OWNER_TYPE_NONE, OBJECT_NONE, 0, 176 SPDK_TRACE_ARG_TYPE_INT, "type"); 177 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 178 OWNER_TYPE_NONE, OBJECT_NONE, 0, 179 SPDK_TRACE_ARG_TYPE_INT, "type"); 180 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 181 OWNER_TYPE_NONE, OBJECT_NONE, 0, 182 SPDK_TRACE_ARG_TYPE_INT, ""); 183 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 184 OWNER_TYPE_NONE, OBJECT_NONE, 0, 185 SPDK_TRACE_ARG_TYPE_INT, ""); 186 } 187 188 enum spdk_nvmf_rdma_wr_type { 189 RDMA_WR_TYPE_RECV, 190 RDMA_WR_TYPE_SEND, 191 RDMA_WR_TYPE_DATA, 192 }; 193 194 struct spdk_nvmf_rdma_wr { 195 /* Uses enum spdk_nvmf_rdma_wr_type */ 196 uint8_t type; 197 }; 198 199 /* This structure holds commands as they are received off the wire. 200 * It must be dynamically paired with a full request object 201 * (spdk_nvmf_rdma_request) to service a request. It is separate 202 * from the request because RDMA does not appear to order 203 * completions, so occasionally we'll get a new incoming 204 * command when there aren't any free request objects. 205 */ 206 struct spdk_nvmf_rdma_recv { 207 struct ibv_recv_wr wr; 208 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 209 210 struct spdk_nvmf_rdma_qpair *qpair; 211 212 /* In-capsule data buffer */ 213 uint8_t *buf; 214 215 struct spdk_nvmf_rdma_wr rdma_wr; 216 uint64_t receive_tsc; 217 218 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 219 }; 220 221 struct spdk_nvmf_rdma_request_data { 222 struct ibv_send_wr wr; 223 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 224 }; 225 226 struct spdk_nvmf_rdma_request { 227 struct spdk_nvmf_request req; 228 229 bool fused_failed; 230 231 struct spdk_nvmf_rdma_wr data_wr; 232 struct spdk_nvmf_rdma_wr rsp_wr; 233 234 /* Uses enum spdk_nvmf_rdma_request_state */ 235 uint8_t state; 236 237 /* Data offset in req.iov */ 238 uint32_t offset; 239 240 struct spdk_nvmf_rdma_recv *recv; 241 242 struct { 243 struct ibv_send_wr wr; 244 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 245 } rsp; 246 247 uint16_t iovpos; 248 uint16_t num_outstanding_data_wr; 249 /* Used to split Write IO with multi SGL payload */ 250 uint16_t num_remaining_data_wr; 251 uint64_t receive_tsc; 252 struct spdk_nvmf_rdma_request *fused_pair; 253 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 254 struct ibv_send_wr *remaining_tranfer_in_wrs; 255 struct ibv_send_wr *transfer_wr; 256 struct spdk_nvmf_rdma_request_data data; 257 }; 258 259 struct spdk_nvmf_rdma_resource_opts { 260 struct spdk_nvmf_rdma_qpair *qpair; 261 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 262 void *qp; 263 struct spdk_rdma_utils_mem_map *map; 264 uint32_t max_queue_depth; 265 uint32_t in_capsule_data_size; 266 bool shared; 267 }; 268 269 struct spdk_nvmf_rdma_resources { 270 /* Array of size "max_queue_depth" containing RDMA requests. */ 271 struct spdk_nvmf_rdma_request *reqs; 272 273 /* Array of size "max_queue_depth" containing RDMA recvs. */ 274 struct spdk_nvmf_rdma_recv *recvs; 275 276 /* Array of size "max_queue_depth" containing 64 byte capsules 277 * used for receive. 278 */ 279 union nvmf_h2c_msg *cmds; 280 281 /* Array of size "max_queue_depth" containing 16 byte completions 282 * to be sent back to the user. 283 */ 284 union nvmf_c2h_msg *cpls; 285 286 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 287 * buffers to be used for in capsule data. 288 */ 289 void *bufs; 290 291 /* Receives that are waiting for a request object */ 292 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 293 294 /* Queue to track free requests */ 295 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 296 }; 297 298 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 299 300 typedef void (*spdk_poller_destroy_cb)(void *ctx); 301 302 struct spdk_nvmf_rdma_ibv_event_ctx { 303 struct spdk_nvmf_rdma_qpair *rqpair; 304 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 305 /* Link to other ibv events associated with this qpair */ 306 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 307 }; 308 309 struct spdk_nvmf_rdma_qpair { 310 struct spdk_nvmf_qpair qpair; 311 312 struct spdk_nvmf_rdma_device *device; 313 struct spdk_nvmf_rdma_poller *poller; 314 315 struct spdk_rdma_provider_qp *rdma_qp; 316 struct rdma_cm_id *cm_id; 317 struct spdk_rdma_provider_srq *srq; 318 struct rdma_cm_id *listen_id; 319 320 /* Cache the QP number to improve QP search by RB tree. */ 321 uint32_t qp_num; 322 323 /* The maximum number of I/O outstanding on this connection at one time */ 324 uint16_t max_queue_depth; 325 326 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 327 uint16_t max_read_depth; 328 329 /* The maximum number of RDMA SEND operations at one time */ 330 uint32_t max_send_depth; 331 332 /* The current number of outstanding WRs from this qpair's 333 * recv queue. Should not exceed device->attr.max_queue_depth. 334 */ 335 uint16_t current_recv_depth; 336 337 /* The current number of active RDMA READ operations */ 338 uint16_t current_read_depth; 339 340 /* The current number of posted WRs from this qpair's 341 * send queue. Should not exceed max_send_depth. 342 */ 343 uint32_t current_send_depth; 344 345 /* The maximum number of SGEs per WR on the send queue */ 346 uint32_t max_send_sge; 347 348 /* The maximum number of SGEs per WR on the recv queue */ 349 uint32_t max_recv_sge; 350 351 struct spdk_nvmf_rdma_resources *resources; 352 353 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 354 355 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 356 357 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_send_queue; 358 359 /* Number of requests not in the free state */ 360 uint32_t qd; 361 362 bool ibv_in_error_state; 363 364 RB_ENTRY(spdk_nvmf_rdma_qpair) node; 365 366 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 367 368 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 369 370 /* Points to the a request that has fuse bits set to 371 * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting 372 * for the request that has SPDK_NVME_CMD_FUSE_SECOND. 373 */ 374 struct spdk_nvmf_rdma_request *fused_first; 375 376 /* 377 * io_channel which is used to destroy qpair when it is removed from poll group 378 */ 379 struct spdk_io_channel *destruct_channel; 380 381 /* List of ibv async events */ 382 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 383 384 /* Lets us know that we have received the last_wqe event. */ 385 bool last_wqe_reached; 386 387 /* Indicate that nvmf_rdma_close_qpair is called */ 388 bool to_close; 389 }; 390 391 struct spdk_nvmf_rdma_poller_stat { 392 uint64_t completions; 393 uint64_t polls; 394 uint64_t idle_polls; 395 uint64_t requests; 396 uint64_t request_latency; 397 uint64_t pending_free_request; 398 uint64_t pending_rdma_read; 399 uint64_t pending_rdma_write; 400 uint64_t pending_rdma_send; 401 struct spdk_rdma_provider_qp_stats qp_stats; 402 }; 403 404 struct spdk_nvmf_rdma_poller { 405 struct spdk_nvmf_rdma_device *device; 406 struct spdk_nvmf_rdma_poll_group *group; 407 408 int num_cqe; 409 int required_num_wr; 410 struct ibv_cq *cq; 411 412 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 413 uint16_t max_srq_depth; 414 bool need_destroy; 415 416 /* Shared receive queue */ 417 struct spdk_rdma_provider_srq *srq; 418 419 struct spdk_nvmf_rdma_resources *resources; 420 struct spdk_nvmf_rdma_poller_stat stat; 421 422 spdk_poller_destroy_cb destroy_cb; 423 void *destroy_cb_ctx; 424 425 RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs; 426 427 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 428 429 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 430 431 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 432 }; 433 434 struct spdk_nvmf_rdma_poll_group_stat { 435 uint64_t pending_data_buffer; 436 }; 437 438 struct spdk_nvmf_rdma_poll_group { 439 struct spdk_nvmf_transport_poll_group group; 440 struct spdk_nvmf_rdma_poll_group_stat stat; 441 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 442 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 443 }; 444 445 struct spdk_nvmf_rdma_conn_sched { 446 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 447 struct spdk_nvmf_rdma_poll_group *next_io_pg; 448 }; 449 450 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 451 struct spdk_nvmf_rdma_device { 452 struct ibv_device_attr attr; 453 struct ibv_context *context; 454 455 struct spdk_rdma_utils_mem_map *map; 456 struct ibv_pd *pd; 457 458 int num_srq; 459 bool need_destroy; 460 bool ready_to_destroy; 461 bool is_ready; 462 463 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 464 }; 465 466 struct spdk_nvmf_rdma_port { 467 const struct spdk_nvme_transport_id *trid; 468 struct rdma_cm_id *id; 469 struct spdk_nvmf_rdma_device *device; 470 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 471 }; 472 473 struct rdma_transport_opts { 474 int num_cqe; 475 uint32_t max_srq_depth; 476 bool no_srq; 477 bool no_wr_batching; 478 int acceptor_backlog; 479 }; 480 481 struct spdk_nvmf_rdma_transport { 482 struct spdk_nvmf_transport transport; 483 struct rdma_transport_opts rdma_opts; 484 485 struct spdk_nvmf_rdma_conn_sched conn_sched; 486 487 struct rdma_event_channel *event_channel; 488 489 struct spdk_mempool *data_wr_pool; 490 491 struct spdk_poller *accept_poller; 492 493 /* fields used to poll RDMA/IB events */ 494 nfds_t npoll_fds; 495 struct pollfd *poll_fds; 496 497 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 498 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 499 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 500 501 /* ports that are removed unexpectedly and need retry listen */ 502 TAILQ_HEAD(, spdk_nvmf_rdma_port) retry_ports; 503 }; 504 505 struct poller_manage_ctx { 506 struct spdk_nvmf_rdma_transport *rtransport; 507 struct spdk_nvmf_rdma_poll_group *rgroup; 508 struct spdk_nvmf_rdma_poller *rpoller; 509 struct spdk_nvmf_rdma_device *device; 510 511 struct spdk_thread *thread; 512 volatile int *inflight_op_counter; 513 }; 514 515 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 516 { 517 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe), 518 spdk_json_decode_int32, true 519 }, 520 { 521 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 522 spdk_json_decode_uint32, true 523 }, 524 { 525 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 526 spdk_json_decode_bool, true 527 }, 528 { 529 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 530 spdk_json_decode_bool, true 531 }, 532 { 533 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 534 spdk_json_decode_int32, true 535 }, 536 }; 537 538 static int 539 nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2) 540 { 541 return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num; 542 } 543 544 RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare); 545 546 static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 547 struct spdk_nvmf_rdma_request *rdma_req); 548 549 static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 550 struct spdk_nvmf_rdma_poller *rpoller); 551 552 static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 553 struct spdk_nvmf_rdma_poller *rpoller); 554 555 static void _nvmf_rdma_remove_destroyed_device(void *c); 556 557 static inline enum spdk_nvme_media_error_status_code 558 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 559 enum spdk_nvme_media_error_status_code result; 560 switch (err_type) 561 { 562 case SPDK_DIF_REFTAG_ERROR: 563 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 564 break; 565 case SPDK_DIF_APPTAG_ERROR: 566 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 567 break; 568 case SPDK_DIF_GUARD_ERROR: 569 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 570 break; 571 default: 572 SPDK_UNREACHABLE(); 573 } 574 575 return result; 576 } 577 578 /* 579 * Return data_wrs to pool starting from \b data_wr 580 * Request's own response and data WR are excluded 581 */ 582 static void 583 _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 584 struct ibv_send_wr *data_wr, 585 struct spdk_mempool *pool) 586 { 587 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 588 struct spdk_nvmf_rdma_request_data *nvmf_data; 589 struct ibv_send_wr *next_send_wr; 590 uint64_t req_wrid = (uint64_t)&rdma_req->data_wr; 591 uint32_t num_wrs = 0; 592 593 while (data_wr && data_wr->wr_id == req_wrid) { 594 nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr); 595 memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge); 596 data_wr->num_sge = 0; 597 next_send_wr = data_wr->next; 598 if (data_wr != &rdma_req->data.wr) { 599 data_wr->next = NULL; 600 assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES); 601 work_requests[num_wrs] = nvmf_data; 602 num_wrs++; 603 } 604 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr; 605 } 606 607 if (num_wrs) { 608 spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs); 609 } 610 } 611 612 static void 613 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 614 struct spdk_nvmf_rdma_transport *rtransport) 615 { 616 rdma_req->num_outstanding_data_wr = 0; 617 618 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 619 620 if (rdma_req->remaining_tranfer_in_wrs) { 621 _nvmf_rdma_request_free_data(rdma_req, rdma_req->remaining_tranfer_in_wrs, 622 rtransport->data_wr_pool); 623 rdma_req->remaining_tranfer_in_wrs = NULL; 624 } 625 626 rdma_req->data.wr.next = NULL; 627 rdma_req->rsp.wr.next = NULL; 628 } 629 630 static void 631 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 632 { 633 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 634 if (req->req.cmd) { 635 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 636 } 637 if (req->recv) { 638 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 639 } 640 } 641 642 static void 643 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 644 { 645 int i; 646 647 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 648 for (i = 0; i < rqpair->max_queue_depth; i++) { 649 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 650 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 651 } 652 } 653 } 654 655 static void 656 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 657 { 658 spdk_free(resources->cmds); 659 spdk_free(resources->cpls); 660 spdk_free(resources->bufs); 661 spdk_free(resources->reqs); 662 spdk_free(resources->recvs); 663 free(resources); 664 } 665 666 667 static struct spdk_nvmf_rdma_resources * 668 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 669 { 670 struct spdk_nvmf_rdma_resources *resources; 671 struct spdk_nvmf_rdma_request *rdma_req; 672 struct spdk_nvmf_rdma_recv *rdma_recv; 673 struct spdk_rdma_provider_qp *qp = NULL; 674 struct spdk_rdma_provider_srq *srq = NULL; 675 struct ibv_recv_wr *bad_wr = NULL; 676 struct spdk_rdma_utils_memory_translation translation; 677 uint32_t i; 678 int rc = 0; 679 680 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 681 if (!resources) { 682 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 683 return NULL; 684 } 685 686 resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs), 687 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 688 resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs), 689 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 690 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 691 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 692 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 693 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 694 695 if (opts->in_capsule_data_size > 0) { 696 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 697 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 698 SPDK_MALLOC_DMA); 699 } 700 701 if (!resources->reqs || !resources->recvs || !resources->cmds || 702 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 703 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 704 goto cleanup; 705 } 706 707 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n", 708 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds)); 709 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n", 710 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls)); 711 if (resources->bufs) { 712 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n", 713 resources->bufs, opts->max_queue_depth * 714 opts->in_capsule_data_size); 715 } 716 717 /* Initialize queues */ 718 STAILQ_INIT(&resources->incoming_queue); 719 STAILQ_INIT(&resources->free_queue); 720 721 if (opts->shared) { 722 srq = (struct spdk_rdma_provider_srq *)opts->qp; 723 } else { 724 qp = (struct spdk_rdma_provider_qp *)opts->qp; 725 } 726 727 for (i = 0; i < opts->max_queue_depth; i++) { 728 rdma_recv = &resources->recvs[i]; 729 rdma_recv->qpair = opts->qpair; 730 731 /* Set up memory to receive commands */ 732 if (resources->bufs) { 733 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 734 opts->in_capsule_data_size)); 735 } 736 737 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 738 739 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 740 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 741 rc = spdk_rdma_utils_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]), 742 &translation); 743 if (rc) { 744 goto cleanup; 745 } 746 rdma_recv->sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 747 rdma_recv->wr.num_sge = 1; 748 749 if (rdma_recv->buf) { 750 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 751 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 752 rc = spdk_rdma_utils_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size, 753 &translation); 754 if (rc) { 755 goto cleanup; 756 } 757 rdma_recv->sgl[1].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 758 rdma_recv->wr.num_sge++; 759 } 760 761 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 762 rdma_recv->wr.sg_list = rdma_recv->sgl; 763 if (srq) { 764 spdk_rdma_provider_srq_queue_recv_wrs(srq, &rdma_recv->wr); 765 } else { 766 spdk_rdma_provider_qp_queue_recv_wrs(qp, &rdma_recv->wr); 767 } 768 } 769 770 for (i = 0; i < opts->max_queue_depth; i++) { 771 rdma_req = &resources->reqs[i]; 772 773 if (opts->qpair != NULL) { 774 rdma_req->req.qpair = &opts->qpair->qpair; 775 } else { 776 rdma_req->req.qpair = NULL; 777 } 778 rdma_req->req.cmd = NULL; 779 rdma_req->req.iovcnt = 0; 780 rdma_req->req.stripped_data = NULL; 781 782 /* Set up memory to send responses */ 783 rdma_req->req.rsp = &resources->cpls[i]; 784 785 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 786 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 787 rc = spdk_rdma_utils_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]), 788 &translation); 789 if (rc) { 790 goto cleanup; 791 } 792 rdma_req->rsp.sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 793 794 rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND; 795 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr; 796 rdma_req->rsp.wr.next = NULL; 797 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 798 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 799 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 800 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 801 802 /* Set up memory for data buffers */ 803 rdma_req->data_wr.type = RDMA_WR_TYPE_DATA; 804 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr; 805 rdma_req->data.wr.next = NULL; 806 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 807 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 808 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 809 810 /* Initialize request state to FREE */ 811 rdma_req->state = RDMA_REQUEST_STATE_FREE; 812 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 813 } 814 815 if (srq) { 816 rc = spdk_rdma_provider_srq_flush_recv_wrs(srq, &bad_wr); 817 } else { 818 rc = spdk_rdma_provider_qp_flush_recv_wrs(qp, &bad_wr); 819 } 820 821 if (rc) { 822 goto cleanup; 823 } 824 825 return resources; 826 827 cleanup: 828 nvmf_rdma_resources_destroy(resources); 829 return NULL; 830 } 831 832 static void 833 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 834 { 835 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 836 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 837 ctx->rqpair = NULL; 838 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 839 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 840 } 841 } 842 843 static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller); 844 845 static void 846 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 847 { 848 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 849 struct ibv_recv_wr *bad_recv_wr = NULL; 850 int rc; 851 852 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair); 853 854 if (rqpair->qd != 0) { 855 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 856 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 857 struct spdk_nvmf_rdma_transport, transport); 858 struct spdk_nvmf_rdma_request *req; 859 uint32_t i, max_req_count = 0; 860 861 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 862 863 if (rqpair->srq == NULL) { 864 nvmf_rdma_dump_qpair_contents(rqpair); 865 max_req_count = rqpair->max_queue_depth; 866 } else if (rqpair->poller && rqpair->resources) { 867 max_req_count = rqpair->poller->max_srq_depth; 868 } 869 870 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 871 for (i = 0; i < max_req_count; i++) { 872 req = &rqpair->resources->reqs[i]; 873 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 874 /* nvmf_rdma_request_process checks qpair ibv and internal state 875 * and completes a request */ 876 nvmf_rdma_request_process(rtransport, req); 877 } 878 } 879 assert(rqpair->qd == 0); 880 } 881 882 if (rqpair->poller) { 883 RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair); 884 885 if (rqpair->srq != NULL && rqpair->resources != NULL) { 886 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 887 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 888 if (rqpair == rdma_recv->qpair) { 889 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 890 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr); 891 rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 892 if (rc) { 893 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 894 } 895 } 896 } 897 } 898 } 899 900 if (rqpair->cm_id) { 901 if (rqpair->rdma_qp != NULL) { 902 spdk_rdma_provider_qp_destroy(rqpair->rdma_qp); 903 rqpair->rdma_qp = NULL; 904 } 905 906 if (rqpair->poller != NULL && rqpair->srq == NULL) { 907 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 908 } 909 } 910 911 if (rqpair->srq == NULL && rqpair->resources != NULL) { 912 nvmf_rdma_resources_destroy(rqpair->resources); 913 } 914 915 nvmf_rdma_qpair_clean_ibv_events(rqpair); 916 917 if (rqpair->destruct_channel) { 918 spdk_put_io_channel(rqpair->destruct_channel); 919 rqpair->destruct_channel = NULL; 920 } 921 922 if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) { 923 nvmf_rdma_poller_destroy(rqpair->poller); 924 } 925 926 /* destroy cm_id last so cma device will not be freed before we destroy the cq. */ 927 if (rqpair->cm_id) { 928 rdma_destroy_id(rqpair->cm_id); 929 } 930 931 free(rqpair); 932 } 933 934 static int 935 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 936 { 937 struct spdk_nvmf_rdma_poller *rpoller; 938 int rc, num_cqe, required_num_wr; 939 940 /* Enlarge CQ size dynamically */ 941 rpoller = rqpair->poller; 942 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 943 num_cqe = rpoller->num_cqe; 944 if (num_cqe < required_num_wr) { 945 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 946 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 947 } 948 949 if (rpoller->num_cqe != num_cqe) { 950 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) { 951 SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n" 952 "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe); 953 return -1; 954 } 955 if (required_num_wr > device->attr.max_cqe) { 956 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 957 required_num_wr, device->attr.max_cqe); 958 return -1; 959 } 960 961 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 962 rc = ibv_resize_cq(rpoller->cq, num_cqe); 963 if (rc) { 964 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 965 return -1; 966 } 967 968 rpoller->num_cqe = num_cqe; 969 } 970 971 rpoller->required_num_wr = required_num_wr; 972 return 0; 973 } 974 975 static int 976 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 977 { 978 struct spdk_nvmf_rdma_qpair *rqpair; 979 struct spdk_nvmf_rdma_transport *rtransport; 980 struct spdk_nvmf_transport *transport; 981 struct spdk_nvmf_rdma_resource_opts opts; 982 struct spdk_nvmf_rdma_device *device; 983 struct spdk_rdma_provider_qp_init_attr qp_init_attr = {}; 984 985 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 986 device = rqpair->device; 987 988 qp_init_attr.qp_context = rqpair; 989 qp_init_attr.pd = device->pd; 990 qp_init_attr.send_cq = rqpair->poller->cq; 991 qp_init_attr.recv_cq = rqpair->poller->cq; 992 993 if (rqpair->srq) { 994 qp_init_attr.srq = rqpair->srq->srq; 995 } else { 996 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 997 } 998 999 /* SEND, READ, and WRITE operations */ 1000 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 1001 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 1002 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1003 qp_init_attr.stats = &rqpair->poller->stat.qp_stats; 1004 1005 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1006 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1007 goto error; 1008 } 1009 1010 rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &qp_init_attr); 1011 if (!rqpair->rdma_qp) { 1012 goto error; 1013 } 1014 1015 rqpair->qp_num = rqpair->rdma_qp->qp->qp_num; 1016 1017 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1018 qp_init_attr.cap.max_send_wr); 1019 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1020 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1021 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair); 1022 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1023 1024 if (rqpair->poller->srq == NULL) { 1025 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1026 transport = &rtransport->transport; 1027 1028 opts.qp = rqpair->rdma_qp; 1029 opts.map = device->map; 1030 opts.qpair = rqpair; 1031 opts.shared = false; 1032 opts.max_queue_depth = rqpair->max_queue_depth; 1033 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1034 1035 rqpair->resources = nvmf_rdma_resources_create(&opts); 1036 1037 if (!rqpair->resources) { 1038 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1039 rdma_destroy_qp(rqpair->cm_id); 1040 goto error; 1041 } 1042 } else { 1043 rqpair->resources = rqpair->poller->resources; 1044 } 1045 1046 rqpair->current_recv_depth = 0; 1047 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1048 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1049 STAILQ_INIT(&rqpair->pending_rdma_send_queue); 1050 rqpair->qpair.queue_depth = 0; 1051 1052 return 0; 1053 1054 error: 1055 rdma_destroy_id(rqpair->cm_id); 1056 rqpair->cm_id = NULL; 1057 return -1; 1058 } 1059 1060 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1061 /* This function accepts either a single wr or the first wr in a linked list. */ 1062 static void 1063 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1064 { 1065 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1066 struct spdk_nvmf_rdma_transport, transport); 1067 1068 if (rqpair->srq != NULL) { 1069 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, first); 1070 } else { 1071 if (spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, first)) { 1072 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1073 } 1074 } 1075 1076 if (rtransport->rdma_opts.no_wr_batching) { 1077 _poller_submit_recvs(rtransport, rqpair->poller); 1078 } 1079 } 1080 1081 static int 1082 request_transfer_in(struct spdk_nvmf_request *req) 1083 { 1084 struct spdk_nvmf_rdma_request *rdma_req; 1085 struct spdk_nvmf_qpair *qpair; 1086 struct spdk_nvmf_rdma_qpair *rqpair; 1087 struct spdk_nvmf_rdma_transport *rtransport; 1088 1089 qpair = req->qpair; 1090 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1091 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1092 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1093 struct spdk_nvmf_rdma_transport, transport); 1094 1095 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1096 assert(rdma_req != NULL); 1097 1098 if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) { 1099 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1100 } 1101 if (rtransport->rdma_opts.no_wr_batching) { 1102 _poller_submit_sends(rtransport, rqpair->poller); 1103 } 1104 1105 assert(rqpair->current_read_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_read_depth); 1106 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1107 assert(rqpair->current_send_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_send_depth); 1108 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1109 return 0; 1110 } 1111 1112 static inline void 1113 nvmf_rdma_request_reset_transfer_in(struct spdk_nvmf_rdma_request *rdma_req, 1114 struct spdk_nvmf_rdma_transport *rtransport) 1115 { 1116 /* Put completed WRs back to pool and move transfer_wr pointer */ 1117 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 1118 rdma_req->transfer_wr = rdma_req->remaining_tranfer_in_wrs; 1119 rdma_req->remaining_tranfer_in_wrs = NULL; 1120 rdma_req->num_outstanding_data_wr = rdma_req->num_remaining_data_wr; 1121 rdma_req->num_remaining_data_wr = 0; 1122 } 1123 1124 static inline int 1125 request_prepare_transfer_in_part(struct spdk_nvmf_request *req, uint32_t num_reads_available) 1126 { 1127 struct spdk_nvmf_rdma_request *rdma_req; 1128 struct ibv_send_wr *wr; 1129 uint32_t i; 1130 1131 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1132 1133 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1134 assert(rdma_req != NULL); 1135 assert(num_reads_available > 0); 1136 assert(rdma_req->num_outstanding_data_wr > num_reads_available); 1137 wr = rdma_req->transfer_wr; 1138 1139 for (i = 0; i < num_reads_available - 1; i++) { 1140 wr = wr->next; 1141 } 1142 1143 rdma_req->remaining_tranfer_in_wrs = wr->next; 1144 rdma_req->num_remaining_data_wr = rdma_req->num_outstanding_data_wr - num_reads_available; 1145 rdma_req->num_outstanding_data_wr = num_reads_available; 1146 /* Break chain of WRs to send only part. Once this portion completes, we continue sending RDMA_READs */ 1147 wr->next = NULL; 1148 1149 return 0; 1150 } 1151 1152 static int 1153 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1154 { 1155 int num_outstanding_data_wr = 0; 1156 struct spdk_nvmf_rdma_request *rdma_req; 1157 struct spdk_nvmf_qpair *qpair; 1158 struct spdk_nvmf_rdma_qpair *rqpair; 1159 struct spdk_nvme_cpl *rsp; 1160 struct ibv_send_wr *first = NULL; 1161 struct spdk_nvmf_rdma_transport *rtransport; 1162 1163 *data_posted = 0; 1164 qpair = req->qpair; 1165 rsp = &req->rsp->nvme_cpl; 1166 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1167 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1168 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1169 struct spdk_nvmf_rdma_transport, transport); 1170 1171 /* Advance our sq_head pointer */ 1172 if (qpair->sq_head == qpair->sq_head_max) { 1173 qpair->sq_head = 0; 1174 } else { 1175 qpair->sq_head++; 1176 } 1177 rsp->sqhd = qpair->sq_head; 1178 1179 /* queue the capsule for the recv buffer */ 1180 assert(rdma_req->recv != NULL); 1181 1182 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1183 1184 rdma_req->recv = NULL; 1185 assert(rqpair->current_recv_depth > 0); 1186 rqpair->current_recv_depth--; 1187 1188 /* Build the response which consists of optional 1189 * RDMA WRITEs to transfer data, plus an RDMA SEND 1190 * containing the response. 1191 */ 1192 first = &rdma_req->rsp.wr; 1193 1194 if (spdk_unlikely(rsp->status.sc != SPDK_NVME_SC_SUCCESS)) { 1195 /* On failure, data was not read from the controller. So clear the 1196 * number of outstanding data WRs to zero. 1197 */ 1198 rdma_req->num_outstanding_data_wr = 0; 1199 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1200 first = rdma_req->transfer_wr; 1201 *data_posted = 1; 1202 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1203 } 1204 if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1205 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1206 } 1207 if (rtransport->rdma_opts.no_wr_batching) { 1208 _poller_submit_sends(rtransport, rqpair->poller); 1209 } 1210 1211 /* +1 for the rsp wr */ 1212 assert(rqpair->current_send_depth + num_outstanding_data_wr + 1 <= rqpair->max_send_depth); 1213 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1214 1215 return 0; 1216 } 1217 1218 static int 1219 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1220 { 1221 struct spdk_nvmf_rdma_accept_private_data accept_data; 1222 struct rdma_conn_param ctrlr_event_data = {}; 1223 int rc; 1224 1225 accept_data.recfmt = 0; 1226 accept_data.crqsize = rqpair->max_queue_depth; 1227 1228 ctrlr_event_data.private_data = &accept_data; 1229 ctrlr_event_data.private_data_len = sizeof(accept_data); 1230 if (id->ps == RDMA_PS_TCP) { 1231 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1232 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1233 } 1234 1235 /* Configure infinite retries for the initiator side qpair. 1236 * We need to pass this value to the initiator to prevent the 1237 * initiator side NIC from completing SEND requests back to the 1238 * initiator with status rnr_retry_count_exceeded. */ 1239 ctrlr_event_data.rnr_retry_count = 0x7; 1240 1241 /* When qpair is created without use of rdma cm API, an additional 1242 * information must be provided to initiator in the connection response: 1243 * whether qpair is using SRQ and its qp_num 1244 * Fields below are ignored by rdma cm if qpair has been 1245 * created using rdma cm API. */ 1246 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1247 ctrlr_event_data.qp_num = rqpair->qp_num; 1248 1249 rc = spdk_rdma_provider_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1250 if (rc) { 1251 SPDK_ERRLOG("Error %d on spdk_rdma_provider_qp_accept\n", errno); 1252 } else { 1253 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1254 } 1255 1256 return rc; 1257 } 1258 1259 static void 1260 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1261 { 1262 struct spdk_nvmf_rdma_reject_private_data rej_data; 1263 1264 rej_data.recfmt = 0; 1265 rej_data.sts = error; 1266 1267 rdma_reject(id, &rej_data, sizeof(rej_data)); 1268 } 1269 1270 static int 1271 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1272 { 1273 struct spdk_nvmf_rdma_transport *rtransport; 1274 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1275 struct spdk_nvmf_rdma_port *port; 1276 struct rdma_conn_param *rdma_param = NULL; 1277 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1278 uint16_t max_queue_depth; 1279 uint16_t max_read_depth; 1280 1281 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1282 1283 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1284 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1285 1286 rdma_param = &event->param.conn; 1287 if (rdma_param->private_data == NULL || 1288 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1289 SPDK_ERRLOG("connect request: no private data provided\n"); 1290 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1291 return -1; 1292 } 1293 1294 private_data = rdma_param->private_data; 1295 if (private_data->recfmt != 0) { 1296 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1297 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1298 return -1; 1299 } 1300 1301 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1302 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1303 1304 port = event->listen_id->context; 1305 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1306 event->listen_id, event->listen_id->verbs, port); 1307 1308 /* Figure out the supported queue depth. This is a multi-step process 1309 * that takes into account hardware maximums, host provided values, 1310 * and our target's internal memory limits */ 1311 1312 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1313 1314 /* Start with the maximum queue depth allowed by the target */ 1315 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1316 max_read_depth = rtransport->transport.opts.max_queue_depth; 1317 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1318 rtransport->transport.opts.max_queue_depth); 1319 1320 /* Next check the local NIC's hardware limitations */ 1321 SPDK_DEBUGLOG(rdma, 1322 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1323 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1324 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1325 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1326 1327 /* Next check the remote NIC's hardware limitations */ 1328 SPDK_DEBUGLOG(rdma, 1329 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1330 rdma_param->initiator_depth, rdma_param->responder_resources); 1331 /* from man3 rdma_get_cm_event 1332 * responder_resources - Specifies the number of responder resources that is requested by the recipient. 1333 * The responder_resources field must match the initiator depth specified by the remote node when running 1334 * the rdma_connect and rdma_accept functions. */ 1335 if (rdma_param->responder_resources != 0) { 1336 if (private_data->qid) { 1337 SPDK_DEBUGLOG(rdma, "Host (Initiator) is not allowed to use RDMA operations," 1338 " responder_resources must be 0 but set to %u\n", 1339 rdma_param->responder_resources); 1340 } else { 1341 SPDK_WARNLOG("Host (Initiator) is not allowed to use RDMA operations," 1342 " responder_resources must be 0 but set to %u\n", 1343 rdma_param->responder_resources); 1344 } 1345 } 1346 /* from man3 rdma_get_cm_event 1347 * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds. 1348 * The initiator_depth field must match the responder resources specified by the remote node when running 1349 * the rdma_connect and rdma_accept functions. */ 1350 if (rdma_param->initiator_depth == 0) { 1351 SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n"); 1352 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD); 1353 return -1; 1354 } 1355 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1356 1357 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1358 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1359 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1360 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1361 1362 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1363 max_queue_depth, max_read_depth); 1364 1365 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1366 if (rqpair == NULL) { 1367 SPDK_ERRLOG("Could not allocate new connection.\n"); 1368 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1369 return -1; 1370 } 1371 1372 rqpair->device = port->device; 1373 rqpair->max_queue_depth = max_queue_depth; 1374 rqpair->max_read_depth = max_read_depth; 1375 rqpair->cm_id = event->id; 1376 rqpair->listen_id = event->listen_id; 1377 rqpair->qpair.transport = transport; 1378 STAILQ_INIT(&rqpair->ibv_events); 1379 /* use qid from the private data to determine the qpair type 1380 qid will be set to the appropriate value when the controller is created */ 1381 rqpair->qpair.qid = private_data->qid; 1382 1383 event->id->context = &rqpair->qpair; 1384 1385 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1386 1387 return 0; 1388 } 1389 1390 static inline void 1391 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1392 enum spdk_nvme_data_transfer xfer) 1393 { 1394 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1395 wr->opcode = IBV_WR_RDMA_WRITE; 1396 wr->send_flags = 0; 1397 wr->next = next; 1398 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1399 wr->opcode = IBV_WR_RDMA_READ; 1400 wr->send_flags = IBV_SEND_SIGNALED; 1401 wr->next = NULL; 1402 } else { 1403 assert(0); 1404 } 1405 } 1406 1407 static int 1408 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1409 struct spdk_nvmf_rdma_request *rdma_req, 1410 uint32_t num_sgl_descriptors) 1411 { 1412 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1413 struct spdk_nvmf_rdma_request_data *current_data_wr; 1414 uint32_t i; 1415 1416 if (spdk_unlikely(num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES)) { 1417 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1418 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1419 return -EINVAL; 1420 } 1421 1422 if (spdk_unlikely(spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, 1423 num_sgl_descriptors))) { 1424 return -ENOMEM; 1425 } 1426 1427 current_data_wr = &rdma_req->data; 1428 1429 for (i = 0; i < num_sgl_descriptors; i++) { 1430 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1431 current_data_wr->wr.next = &work_requests[i]->wr; 1432 current_data_wr = work_requests[i]; 1433 current_data_wr->wr.sg_list = current_data_wr->sgl; 1434 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1435 } 1436 1437 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1438 1439 return 0; 1440 } 1441 1442 static inline void 1443 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1444 { 1445 struct ibv_send_wr *wr = &rdma_req->data.wr; 1446 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1447 1448 wr->wr.rdma.rkey = sgl->keyed.key; 1449 wr->wr.rdma.remote_addr = sgl->address; 1450 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1451 } 1452 1453 static inline void 1454 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1455 { 1456 struct ibv_send_wr *wr = &rdma_req->data.wr; 1457 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1458 uint32_t i; 1459 int j; 1460 uint64_t remote_addr_offset = 0; 1461 1462 for (i = 0; i < num_wrs; ++i) { 1463 wr->wr.rdma.rkey = sgl->keyed.key; 1464 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1465 for (j = 0; j < wr->num_sge; ++j) { 1466 remote_addr_offset += wr->sg_list[j].length; 1467 } 1468 wr = wr->next; 1469 } 1470 } 1471 1472 static int 1473 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_device *device, 1474 struct spdk_nvmf_rdma_request *rdma_req, 1475 struct ibv_send_wr *wr, 1476 uint32_t total_length) 1477 { 1478 struct spdk_rdma_utils_memory_translation mem_translation; 1479 struct ibv_sge *sg_ele; 1480 struct iovec *iov; 1481 uint32_t lkey, remaining; 1482 int rc; 1483 1484 wr->num_sge = 0; 1485 1486 while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) { 1487 iov = &rdma_req->req.iov[rdma_req->iovpos]; 1488 rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1489 if (spdk_unlikely(rc)) { 1490 return rc; 1491 } 1492 1493 lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation); 1494 sg_ele = &wr->sg_list[wr->num_sge]; 1495 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1496 1497 sg_ele->lkey = lkey; 1498 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset; 1499 sg_ele->length = remaining; 1500 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1501 sg_ele->length); 1502 rdma_req->offset += sg_ele->length; 1503 total_length -= sg_ele->length; 1504 wr->num_sge++; 1505 1506 if (rdma_req->offset == iov->iov_len) { 1507 rdma_req->offset = 0; 1508 rdma_req->iovpos++; 1509 } 1510 } 1511 1512 if (spdk_unlikely(total_length)) { 1513 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1514 return -EINVAL; 1515 } 1516 1517 return 0; 1518 } 1519 1520 static int 1521 nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_device *device, 1522 struct spdk_nvmf_rdma_request *rdma_req, 1523 struct ibv_send_wr *wr, 1524 uint32_t total_length, 1525 uint32_t num_extra_wrs) 1526 { 1527 struct spdk_rdma_utils_memory_translation mem_translation; 1528 struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx; 1529 struct ibv_sge *sg_ele; 1530 struct iovec *iov; 1531 struct iovec *rdma_iov; 1532 uint32_t lkey, remaining; 1533 uint32_t remaining_data_block, data_block_size, md_size; 1534 uint32_t sge_len; 1535 int rc; 1536 1537 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1538 1539 if (spdk_likely(!rdma_req->req.stripped_data)) { 1540 rdma_iov = rdma_req->req.iov; 1541 remaining_data_block = data_block_size; 1542 md_size = dif_ctx->md_size; 1543 } else { 1544 rdma_iov = rdma_req->req.stripped_data->iov; 1545 total_length = total_length / dif_ctx->block_size * data_block_size; 1546 remaining_data_block = total_length; 1547 md_size = 0; 1548 } 1549 1550 wr->num_sge = 0; 1551 1552 while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1553 iov = rdma_iov + rdma_req->iovpos; 1554 rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1555 if (spdk_unlikely(rc)) { 1556 return rc; 1557 } 1558 1559 lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation); 1560 sg_ele = &wr->sg_list[wr->num_sge]; 1561 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1562 1563 while (remaining) { 1564 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1565 if (num_extra_wrs > 0 && wr->next) { 1566 wr = wr->next; 1567 wr->num_sge = 0; 1568 sg_ele = &wr->sg_list[wr->num_sge]; 1569 num_extra_wrs--; 1570 } else { 1571 break; 1572 } 1573 } 1574 sg_ele->lkey = lkey; 1575 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset); 1576 sge_len = spdk_min(remaining, remaining_data_block); 1577 sg_ele->length = sge_len; 1578 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, 1579 sg_ele->addr, sg_ele->length); 1580 remaining -= sge_len; 1581 remaining_data_block -= sge_len; 1582 rdma_req->offset += sge_len; 1583 total_length -= sge_len; 1584 1585 sg_ele++; 1586 wr->num_sge++; 1587 1588 if (remaining_data_block == 0) { 1589 /* skip metadata */ 1590 rdma_req->offset += md_size; 1591 total_length -= md_size; 1592 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1593 remaining -= spdk_min(remaining, md_size); 1594 remaining_data_block = data_block_size; 1595 } 1596 1597 if (remaining == 0) { 1598 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1599 the remaining metadata bits at the beginning of the next buffer */ 1600 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset); 1601 rdma_req->iovpos++; 1602 } 1603 } 1604 } 1605 1606 if (spdk_unlikely(total_length)) { 1607 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1608 return -EINVAL; 1609 } 1610 1611 return 0; 1612 } 1613 1614 static inline uint32_t 1615 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1616 { 1617 /* estimate the number of SG entries and WRs needed to process the request */ 1618 uint32_t num_sge = 0; 1619 uint32_t i; 1620 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1621 1622 for (i = 0; i < num_buffers && length > 0; i++) { 1623 uint32_t buffer_len = spdk_min(length, io_unit_size); 1624 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1625 1626 if (num_sge_in_block * block_size > buffer_len) { 1627 ++num_sge_in_block; 1628 } 1629 num_sge += num_sge_in_block; 1630 length -= buffer_len; 1631 } 1632 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1633 } 1634 1635 static int 1636 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1637 struct spdk_nvmf_rdma_device *device, 1638 struct spdk_nvmf_rdma_request *rdma_req) 1639 { 1640 struct spdk_nvmf_rdma_qpair *rqpair; 1641 struct spdk_nvmf_rdma_poll_group *rgroup; 1642 struct spdk_nvmf_request *req = &rdma_req->req; 1643 struct ibv_send_wr *wr = &rdma_req->data.wr; 1644 int rc; 1645 uint32_t num_wrs = 1; 1646 uint32_t length; 1647 1648 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1649 rgroup = rqpair->poller->group; 1650 1651 /* rdma wr specifics */ 1652 nvmf_rdma_setup_request(rdma_req); 1653 1654 length = req->length; 1655 if (spdk_unlikely(req->dif_enabled)) { 1656 req->dif.orig_length = length; 1657 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1658 req->dif.elba_length = length; 1659 } 1660 1661 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1662 length); 1663 if (spdk_unlikely(rc != 0)) { 1664 return rc; 1665 } 1666 1667 assert(req->iovcnt <= rqpair->max_send_sge); 1668 1669 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1670 * the stripped_buffers are got for DIF stripping. */ 1671 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1672 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1673 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1674 &rtransport->transport, req->dif.orig_length); 1675 if (rc != 0) { 1676 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1677 } 1678 } 1679 1680 rdma_req->iovpos = 0; 1681 1682 if (spdk_unlikely(req->dif_enabled)) { 1683 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1684 req->dif.dif_ctx.block_size); 1685 if (num_wrs > 1) { 1686 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1687 if (spdk_unlikely(rc != 0)) { 1688 goto err_exit; 1689 } 1690 } 1691 1692 rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, wr, length, num_wrs - 1); 1693 if (spdk_unlikely(rc != 0)) { 1694 goto err_exit; 1695 } 1696 1697 if (num_wrs > 1) { 1698 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1699 } 1700 } else { 1701 rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, wr, length); 1702 if (spdk_unlikely(rc != 0)) { 1703 goto err_exit; 1704 } 1705 } 1706 1707 /* set the number of outstanding data WRs for this request. */ 1708 rdma_req->num_outstanding_data_wr = num_wrs; 1709 1710 return rc; 1711 1712 err_exit: 1713 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1714 nvmf_rdma_request_free_data(rdma_req, rtransport); 1715 req->iovcnt = 0; 1716 return rc; 1717 } 1718 1719 static int 1720 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1721 struct spdk_nvmf_rdma_device *device, 1722 struct spdk_nvmf_rdma_request *rdma_req) 1723 { 1724 struct spdk_nvmf_rdma_qpair *rqpair; 1725 struct spdk_nvmf_rdma_poll_group *rgroup; 1726 struct ibv_send_wr *current_wr; 1727 struct spdk_nvmf_request *req = &rdma_req->req; 1728 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1729 uint32_t num_sgl_descriptors; 1730 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0; 1731 uint32_t i; 1732 int rc; 1733 1734 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1735 rgroup = rqpair->poller->group; 1736 1737 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1738 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1739 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1740 1741 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1742 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1743 1744 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1745 for (i = 0; i < num_sgl_descriptors; i++) { 1746 if (spdk_likely(!req->dif_enabled)) { 1747 lengths[i] = desc->keyed.length; 1748 } else { 1749 req->dif.orig_length += desc->keyed.length; 1750 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1751 req->dif.elba_length += lengths[i]; 1752 } 1753 total_length += lengths[i]; 1754 desc++; 1755 } 1756 1757 if (spdk_unlikely(total_length > rtransport->transport.opts.max_io_size)) { 1758 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n", 1759 total_length, rtransport->transport.opts.max_io_size); 1760 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1761 return -EINVAL; 1762 } 1763 1764 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1); 1765 if (spdk_unlikely(rc != 0)) { 1766 return -ENOMEM; 1767 } 1768 1769 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length); 1770 if (spdk_unlikely(rc != 0)) { 1771 nvmf_rdma_request_free_data(rdma_req, rtransport); 1772 return rc; 1773 } 1774 1775 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1776 * the stripped_buffers are got for DIF stripping. */ 1777 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1778 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1779 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1780 &rtransport->transport, req->dif.orig_length); 1781 if (spdk_unlikely(rc != 0)) { 1782 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1783 } 1784 } 1785 1786 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1787 current_wr = &rdma_req->data.wr; 1788 assert(current_wr != NULL); 1789 1790 req->length = 0; 1791 rdma_req->iovpos = 0; 1792 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1793 for (i = 0; i < num_sgl_descriptors; i++) { 1794 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1795 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1796 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1797 rc = -EINVAL; 1798 goto err_exit; 1799 } 1800 1801 if (spdk_likely(!req->dif_enabled)) { 1802 rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, current_wr, lengths[i]); 1803 } else { 1804 rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, current_wr, 1805 lengths[i], 0); 1806 } 1807 if (spdk_unlikely(rc != 0)) { 1808 rc = -ENOMEM; 1809 goto err_exit; 1810 } 1811 1812 req->length += desc->keyed.length; 1813 current_wr->wr.rdma.rkey = desc->keyed.key; 1814 current_wr->wr.rdma.remote_addr = desc->address; 1815 current_wr = current_wr->next; 1816 desc++; 1817 } 1818 1819 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1820 /* Go back to the last descriptor in the list. */ 1821 desc--; 1822 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1823 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1824 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1825 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1826 } 1827 } 1828 #endif 1829 1830 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1831 1832 return 0; 1833 1834 err_exit: 1835 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1836 nvmf_rdma_request_free_data(rdma_req, rtransport); 1837 return rc; 1838 } 1839 1840 static int 1841 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1842 struct spdk_nvmf_rdma_device *device, 1843 struct spdk_nvmf_rdma_request *rdma_req) 1844 { 1845 struct spdk_nvmf_request *req = &rdma_req->req; 1846 struct spdk_nvme_cpl *rsp; 1847 struct spdk_nvme_sgl_descriptor *sgl; 1848 int rc; 1849 uint32_t length; 1850 1851 rsp = &req->rsp->nvme_cpl; 1852 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1853 1854 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1855 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1856 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1857 1858 length = sgl->keyed.length; 1859 if (spdk_unlikely(length > rtransport->transport.opts.max_io_size)) { 1860 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1861 length, rtransport->transport.opts.max_io_size); 1862 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1863 return -1; 1864 } 1865 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1866 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1867 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1868 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1869 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1870 } 1871 } 1872 #endif 1873 1874 /* fill request length and populate iovs */ 1875 req->length = length; 1876 1877 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req); 1878 if (spdk_unlikely(rc < 0)) { 1879 if (rc == -EINVAL) { 1880 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1881 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1882 return -1; 1883 } 1884 /* No available buffers. Queue this request up. */ 1885 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1886 return 0; 1887 } 1888 1889 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1890 req->iovcnt); 1891 1892 return 0; 1893 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1894 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1895 uint64_t offset = sgl->address; 1896 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1897 1898 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1899 offset, sgl->unkeyed.length); 1900 1901 if (spdk_unlikely(offset > max_len)) { 1902 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1903 offset, max_len); 1904 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1905 return -1; 1906 } 1907 max_len -= (uint32_t)offset; 1908 1909 if (spdk_unlikely(sgl->unkeyed.length > max_len)) { 1910 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1911 sgl->unkeyed.length, max_len); 1912 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1913 return -1; 1914 } 1915 1916 rdma_req->num_outstanding_data_wr = 0; 1917 req->data_from_pool = false; 1918 req->length = sgl->unkeyed.length; 1919 1920 req->iov[0].iov_base = rdma_req->recv->buf + offset; 1921 req->iov[0].iov_len = req->length; 1922 req->iovcnt = 1; 1923 1924 return 0; 1925 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1926 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1927 1928 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1929 if (spdk_unlikely(rc == -ENOMEM)) { 1930 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1931 return 0; 1932 } else if (spdk_unlikely(rc == -EINVAL)) { 1933 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1934 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1935 return -1; 1936 } 1937 1938 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1939 req->iovcnt); 1940 1941 return 0; 1942 } 1943 1944 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1945 sgl->generic.type, sgl->generic.subtype); 1946 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1947 return -1; 1948 } 1949 1950 static void 1951 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1952 struct spdk_nvmf_rdma_transport *rtransport) 1953 { 1954 struct spdk_nvmf_rdma_qpair *rqpair; 1955 struct spdk_nvmf_rdma_poll_group *rgroup; 1956 1957 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1958 if (rdma_req->req.data_from_pool) { 1959 rgroup = rqpair->poller->group; 1960 1961 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1962 } 1963 if (rdma_req->req.stripped_data) { 1964 nvmf_request_free_stripped_buffers(&rdma_req->req, 1965 &rqpair->poller->group->group, 1966 &rtransport->transport); 1967 } 1968 nvmf_rdma_request_free_data(rdma_req, rtransport); 1969 rdma_req->req.length = 0; 1970 rdma_req->req.iovcnt = 0; 1971 rdma_req->offset = 0; 1972 rdma_req->req.dif_enabled = false; 1973 rdma_req->fused_failed = false; 1974 rdma_req->transfer_wr = NULL; 1975 if (rdma_req->fused_pair) { 1976 /* This req was part of a valid fused pair, but failed before it got to 1977 * READ_TO_EXECUTE state. This means we need to fail the other request 1978 * in the pair, because it is no longer part of a valid pair. If the pair 1979 * already reached READY_TO_EXECUTE state, we need to kick it. 1980 */ 1981 rdma_req->fused_pair->fused_failed = true; 1982 if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 1983 nvmf_rdma_request_process(rtransport, rdma_req->fused_pair); 1984 } 1985 rdma_req->fused_pair = NULL; 1986 } 1987 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1988 rqpair->qd--; 1989 1990 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1991 rqpair->qpair.queue_depth--; 1992 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1993 } 1994 1995 static void 1996 nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport, 1997 struct spdk_nvmf_rdma_qpair *rqpair, 1998 struct spdk_nvmf_rdma_request *rdma_req) 1999 { 2000 enum spdk_nvme_cmd_fuse last, next; 2001 2002 last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2003 next = rdma_req->req.cmd->nvme_cmd.fuse; 2004 2005 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2006 2007 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2008 return; 2009 } 2010 2011 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2012 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2013 /* This is a valid pair of fused commands. Point them at each other 2014 * so they can be submitted consecutively once ready to be executed. 2015 */ 2016 rqpair->fused_first->fused_pair = rdma_req; 2017 rdma_req->fused_pair = rqpair->fused_first; 2018 rqpair->fused_first = NULL; 2019 return; 2020 } else { 2021 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2022 rqpair->fused_first->fused_failed = true; 2023 2024 /* If the last req is in READY_TO_EXECUTE state, then call 2025 * nvmf_rdma_request_process(), otherwise nothing else will kick it. 2026 */ 2027 if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2028 nvmf_rdma_request_process(rtransport, rqpair->fused_first); 2029 } 2030 2031 rqpair->fused_first = NULL; 2032 } 2033 } 2034 2035 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2036 /* Set rqpair->fused_first here so that we know to check that the next request 2037 * is a SECOND (and to fail this one if it isn't). 2038 */ 2039 rqpair->fused_first = rdma_req; 2040 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2041 /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */ 2042 rdma_req->fused_failed = true; 2043 } 2044 } 2045 2046 bool 2047 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 2048 struct spdk_nvmf_rdma_request *rdma_req) 2049 { 2050 struct spdk_nvmf_rdma_qpair *rqpair; 2051 struct spdk_nvmf_rdma_device *device; 2052 struct spdk_nvmf_rdma_poll_group *rgroup; 2053 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2054 int rc; 2055 struct spdk_nvmf_rdma_recv *rdma_recv; 2056 enum spdk_nvmf_rdma_request_state prev_state; 2057 bool progress = false; 2058 int data_posted; 2059 uint32_t num_blocks, num_rdma_reads_available, qdepth; 2060 2061 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2062 device = rqpair->device; 2063 rgroup = rqpair->poller->group; 2064 2065 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 2066 2067 /* If the queue pair is in an error state, force the request to the completed state 2068 * to release resources. */ 2069 if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 2070 switch (rdma_req->state) { 2071 case RDMA_REQUEST_STATE_NEED_BUFFER: 2072 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 2073 break; 2074 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2075 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2076 break; 2077 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2078 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2079 break; 2080 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 2081 STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2082 break; 2083 default: 2084 break; 2085 } 2086 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2087 } 2088 2089 /* The loop here is to allow for several back-to-back state changes. */ 2090 do { 2091 prev_state = rdma_req->state; 2092 2093 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 2094 2095 switch (rdma_req->state) { 2096 case RDMA_REQUEST_STATE_FREE: 2097 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 2098 * to escape this state. */ 2099 break; 2100 case RDMA_REQUEST_STATE_NEW: 2101 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 2102 (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth); 2103 rdma_recv = rdma_req->recv; 2104 2105 /* The first element of the SGL is the NVMe command */ 2106 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 2107 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 2108 rdma_req->transfer_wr = &rdma_req->data.wr; 2109 2110 if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 2111 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2112 break; 2113 } 2114 2115 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 2116 rdma_req->req.dif_enabled = true; 2117 } 2118 2119 nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req); 2120 2121 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2122 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 2123 rdma_req->rsp.wr.imm_data = 0; 2124 #endif 2125 2126 /* The next state transition depends on the data transfer needs of this request. */ 2127 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2128 2129 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2130 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2131 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2132 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2133 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2134 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 2135 break; 2136 } 2137 2138 /* If no data to transfer, ready to execute. */ 2139 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2140 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2141 break; 2142 } 2143 2144 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2145 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2146 break; 2147 case RDMA_REQUEST_STATE_NEED_BUFFER: 2148 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2149 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2150 2151 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2152 2153 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2154 /* This request needs to wait in line to obtain a buffer */ 2155 break; 2156 } 2157 2158 /* Try to get a data buffer */ 2159 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2160 if (spdk_unlikely(rc < 0)) { 2161 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2162 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2163 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2164 break; 2165 } 2166 2167 if (rdma_req->req.iovcnt == 0) { 2168 /* No buffers available. */ 2169 rgroup->stat.pending_data_buffer++; 2170 break; 2171 } 2172 2173 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2174 2175 /* If data is transferring from host to controller and the data didn't 2176 * arrive using in capsule data, we need to do a transfer from the host. 2177 */ 2178 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2179 rdma_req->req.data_from_pool) { 2180 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2181 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2182 break; 2183 } 2184 2185 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2186 break; 2187 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2188 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2189 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2190 2191 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2192 /* This request needs to wait in line to perform RDMA */ 2193 break; 2194 } 2195 assert(rqpair->max_send_depth >= rqpair->current_send_depth); 2196 qdepth = rqpair->max_send_depth - rqpair->current_send_depth; 2197 assert(rqpair->max_read_depth >= rqpair->current_read_depth); 2198 num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth; 2199 if (rdma_req->num_outstanding_data_wr > qdepth || 2200 rdma_req->num_outstanding_data_wr > num_rdma_reads_available) { 2201 if (num_rdma_reads_available && qdepth) { 2202 /* Send as much as we can */ 2203 request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth)); 2204 } else { 2205 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2206 rqpair->poller->stat.pending_rdma_read++; 2207 break; 2208 } 2209 } 2210 2211 /* We have already verified that this request is the head of the queue. */ 2212 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2213 2214 rc = request_transfer_in(&rdma_req->req); 2215 if (spdk_likely(rc == 0)) { 2216 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2217 } else { 2218 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2219 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2220 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2221 } 2222 break; 2223 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2224 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2225 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2226 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2227 * to escape this state. */ 2228 break; 2229 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2230 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2231 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2232 2233 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2234 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2235 /* generate DIF for write operation */ 2236 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2237 assert(num_blocks > 0); 2238 2239 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2240 num_blocks, &rdma_req->req.dif.dif_ctx); 2241 if (rc != 0) { 2242 SPDK_ERRLOG("DIF generation failed\n"); 2243 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2244 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 2245 break; 2246 } 2247 } 2248 2249 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2250 /* set extended length before IO operation */ 2251 rdma_req->req.length = rdma_req->req.dif.elba_length; 2252 } 2253 2254 if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 2255 if (rdma_req->fused_failed) { 2256 /* This request failed FUSED semantics. Fail it immediately, without 2257 * even sending it to the target layer. 2258 */ 2259 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2260 rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED; 2261 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2262 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2263 break; 2264 } 2265 2266 if (rdma_req->fused_pair == NULL || 2267 rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2268 /* This request is ready to execute, but either we don't know yet if it's 2269 * valid - i.e. this is a FIRST but we haven't received the next 2270 * request yet or the other request of this fused pair isn't ready to 2271 * execute. So break here and this request will get processed later either 2272 * when the other request is ready or we find that this request isn't valid. 2273 */ 2274 break; 2275 } 2276 } 2277 2278 /* If we get to this point, and this request is a fused command, we know that 2279 * it is part of valid sequence (FIRST followed by a SECOND) and that both 2280 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 2281 * request, and the other request of the fused pair, in the correct order. 2282 * Also clear the ->fused_pair pointers on both requests, since after this point 2283 * we no longer need to maintain the relationship between these two requests. 2284 */ 2285 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 2286 assert(rdma_req->fused_pair != NULL); 2287 assert(rdma_req->fused_pair->fused_pair != NULL); 2288 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2289 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2290 rdma_req->fused_pair->fused_pair = NULL; 2291 rdma_req->fused_pair = NULL; 2292 } 2293 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2294 spdk_nvmf_request_exec(&rdma_req->req); 2295 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 2296 assert(rdma_req->fused_pair != NULL); 2297 assert(rdma_req->fused_pair->fused_pair != NULL); 2298 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2299 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2300 rdma_req->fused_pair->fused_pair = NULL; 2301 rdma_req->fused_pair = NULL; 2302 } 2303 break; 2304 case RDMA_REQUEST_STATE_EXECUTING: 2305 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2306 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2307 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2308 * to escape this state. */ 2309 break; 2310 case RDMA_REQUEST_STATE_EXECUTED: 2311 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2312 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2313 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2314 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2315 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2316 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2317 } else { 2318 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2319 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2320 } 2321 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2322 /* restore the original length */ 2323 rdma_req->req.length = rdma_req->req.dif.orig_length; 2324 2325 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2326 struct spdk_dif_error error_blk; 2327 2328 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2329 if (!rdma_req->req.stripped_data) { 2330 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2331 &rdma_req->req.dif.dif_ctx, &error_blk); 2332 } else { 2333 rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov, 2334 rdma_req->req.stripped_data->iovcnt, 2335 rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2336 &rdma_req->req.dif.dif_ctx, &error_blk); 2337 } 2338 if (rc) { 2339 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2340 2341 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2342 error_blk.err_offset); 2343 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2344 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2345 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2346 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2347 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2348 } 2349 } 2350 } 2351 break; 2352 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2353 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2354 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2355 2356 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2357 /* This request needs to wait in line to perform RDMA */ 2358 break; 2359 } 2360 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2361 rqpair->max_send_depth) { 2362 /* We can only have so many WRs outstanding. we have to wait until some finish. 2363 * +1 since each request has an additional wr in the resp. */ 2364 rqpair->poller->stat.pending_rdma_write++; 2365 break; 2366 } 2367 2368 /* We have already verified that this request is the head of the queue. */ 2369 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2370 2371 /* The data transfer will be kicked off from 2372 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2373 * We verified that data + response fit into send queue, so we can go to the next state directly 2374 */ 2375 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2376 break; 2377 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 2378 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 0, 0, 2379 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2380 2381 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_send_queue)) { 2382 /* This request needs to wait in line to send the completion */ 2383 break; 2384 } 2385 2386 assert(rqpair->current_send_depth <= rqpair->max_send_depth); 2387 if (rqpair->current_send_depth == rqpair->max_send_depth) { 2388 /* We can only have so many WRs outstanding. we have to wait until some finish */ 2389 rqpair->poller->stat.pending_rdma_send++; 2390 break; 2391 } 2392 2393 /* We have already verified that this request is the head of the queue. */ 2394 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_send_queue, state_link); 2395 2396 /* The response sending will be kicked off from 2397 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2398 */ 2399 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2400 break; 2401 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2402 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2403 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2404 rc = request_transfer_out(&rdma_req->req, &data_posted); 2405 assert(rc == 0); /* No good way to handle this currently */ 2406 if (spdk_unlikely(rc)) { 2407 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2408 } else { 2409 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2410 RDMA_REQUEST_STATE_COMPLETING; 2411 } 2412 break; 2413 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2414 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2415 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2416 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2417 * to escape this state. */ 2418 break; 2419 case RDMA_REQUEST_STATE_COMPLETING: 2420 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2421 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2422 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2423 * to escape this state. */ 2424 break; 2425 case RDMA_REQUEST_STATE_COMPLETED: 2426 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2427 (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth); 2428 2429 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2430 _nvmf_rdma_request_free(rdma_req, rtransport); 2431 break; 2432 case RDMA_REQUEST_NUM_STATES: 2433 default: 2434 assert(0); 2435 break; 2436 } 2437 2438 if (rdma_req->state != prev_state) { 2439 progress = true; 2440 } 2441 } while (rdma_req->state != prev_state); 2442 2443 return progress; 2444 } 2445 2446 /* Public API callbacks begin here */ 2447 2448 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2449 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2450 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2451 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2452 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2453 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2454 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2455 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2456 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX 2457 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2458 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2459 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2460 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2461 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2462 #define SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE 4095 2463 2464 static void 2465 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2466 { 2467 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2468 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2469 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2470 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2471 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2472 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2473 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2474 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2475 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2476 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2477 opts->transport_specific = NULL; 2478 opts->data_wr_pool_size = SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE; 2479 } 2480 2481 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2482 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2483 2484 static inline bool 2485 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2486 { 2487 return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD || 2488 device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW; 2489 } 2490 2491 static int nvmf_rdma_accept(void *ctx); 2492 static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport); 2493 static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2494 struct spdk_nvmf_rdma_device *device); 2495 2496 static int 2497 create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context, 2498 struct spdk_nvmf_rdma_device **new_device) 2499 { 2500 struct spdk_nvmf_rdma_device *device; 2501 int flag = 0; 2502 int rc = 0; 2503 2504 device = calloc(1, sizeof(*device)); 2505 if (!device) { 2506 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2507 return -ENOMEM; 2508 } 2509 device->context = context; 2510 rc = ibv_query_device(device->context, &device->attr); 2511 if (rc < 0) { 2512 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2513 free(device); 2514 return rc; 2515 } 2516 2517 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2518 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2519 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2520 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2521 } 2522 2523 /** 2524 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2525 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2526 * but incorrectly reports that it does. There are changes making their way 2527 * through the kernel now that will enable this feature. When they are merged, 2528 * we can conditionally enable this feature. 2529 * 2530 * TODO: enable this for versions of the kernel rxe driver that support it. 2531 */ 2532 if (nvmf_rdma_is_rxe_device(device)) { 2533 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2534 } 2535 #endif 2536 2537 /* set up device context async ev fd as NON_BLOCKING */ 2538 flag = fcntl(device->context->async_fd, F_GETFL); 2539 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2540 if (rc < 0) { 2541 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2542 free(device); 2543 return rc; 2544 } 2545 2546 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2547 SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA trasport\n", device); 2548 2549 if (g_nvmf_hooks.get_ibv_pd) { 2550 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2551 } else { 2552 device->pd = ibv_alloc_pd(device->context); 2553 } 2554 2555 if (!device->pd) { 2556 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2557 destroy_ib_device(rtransport, device); 2558 return -ENOMEM; 2559 } 2560 2561 assert(device->map == NULL); 2562 2563 device->map = spdk_rdma_utils_create_mem_map(device->pd, &g_nvmf_hooks, IBV_ACCESS_LOCAL_WRITE); 2564 if (!device->map) { 2565 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2566 destroy_ib_device(rtransport, device); 2567 return -ENOMEM; 2568 } 2569 2570 assert(device->map != NULL); 2571 assert(device->pd != NULL); 2572 2573 if (new_device) { 2574 *new_device = device; 2575 } 2576 SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device), 2577 device, context); 2578 2579 return 0; 2580 } 2581 2582 static void 2583 free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2584 { 2585 if (rtransport->poll_fds) { 2586 free(rtransport->poll_fds); 2587 rtransport->poll_fds = NULL; 2588 } 2589 rtransport->npoll_fds = 0; 2590 } 2591 2592 static int 2593 generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2594 { 2595 /* Set up poll descriptor array to monitor events from RDMA and IB 2596 * in a single poll syscall 2597 */ 2598 int device_count = 0; 2599 int i = 0; 2600 struct spdk_nvmf_rdma_device *device, *tmp; 2601 2602 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2603 device_count++; 2604 } 2605 2606 rtransport->npoll_fds = device_count + 1; 2607 2608 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2609 if (rtransport->poll_fds == NULL) { 2610 SPDK_ERRLOG("poll_fds allocation failed\n"); 2611 return -ENOMEM; 2612 } 2613 2614 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2615 rtransport->poll_fds[i++].events = POLLIN; 2616 2617 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2618 rtransport->poll_fds[i].fd = device->context->async_fd; 2619 rtransport->poll_fds[i++].events = POLLIN; 2620 } 2621 2622 return 0; 2623 } 2624 2625 static struct spdk_nvmf_transport * 2626 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2627 { 2628 int rc; 2629 struct spdk_nvmf_rdma_transport *rtransport; 2630 struct spdk_nvmf_rdma_device *device; 2631 struct ibv_context **contexts; 2632 size_t data_wr_pool_size; 2633 uint32_t i; 2634 int flag; 2635 uint32_t sge_count; 2636 uint32_t min_shared_buffers; 2637 uint32_t min_in_capsule_data_size; 2638 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2639 2640 rtransport = calloc(1, sizeof(*rtransport)); 2641 if (!rtransport) { 2642 return NULL; 2643 } 2644 2645 TAILQ_INIT(&rtransport->devices); 2646 TAILQ_INIT(&rtransport->ports); 2647 TAILQ_INIT(&rtransport->poll_groups); 2648 TAILQ_INIT(&rtransport->retry_ports); 2649 2650 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2651 rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2652 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2653 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2654 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2655 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2656 if (opts->transport_specific != NULL && 2657 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2658 SPDK_COUNTOF(rdma_transport_opts_decoder), 2659 &rtransport->rdma_opts)) { 2660 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2661 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2662 return NULL; 2663 } 2664 2665 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2666 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2667 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2668 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2669 " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d," 2670 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2671 opts->max_queue_depth, 2672 opts->max_io_size, 2673 opts->max_qpairs_per_ctrlr - 1, 2674 opts->io_unit_size, 2675 opts->in_capsule_data_size, 2676 opts->max_aq_depth, 2677 opts->num_shared_buffers, 2678 rtransport->rdma_opts.num_cqe, 2679 rtransport->rdma_opts.max_srq_depth, 2680 rtransport->rdma_opts.no_srq, 2681 rtransport->rdma_opts.acceptor_backlog, 2682 rtransport->rdma_opts.no_wr_batching, 2683 opts->abort_timeout_sec); 2684 2685 /* I/O unit size cannot be larger than max I/O size */ 2686 if (opts->io_unit_size > opts->max_io_size) { 2687 opts->io_unit_size = opts->max_io_size; 2688 } 2689 2690 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2691 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2692 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2693 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2694 } 2695 2696 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2697 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2698 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2699 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2700 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2701 return NULL; 2702 } 2703 2704 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 2705 if (opts->buf_cache_size < UINT32_MAX) { 2706 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2707 if (min_shared_buffers > opts->num_shared_buffers) { 2708 SPDK_ERRLOG("There are not enough buffers to satisfy" 2709 "per-poll group caches for each thread. (%" PRIu32 ")" 2710 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2711 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2712 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2713 return NULL; 2714 } 2715 } 2716 2717 sge_count = opts->max_io_size / opts->io_unit_size; 2718 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2719 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2720 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2721 return NULL; 2722 } 2723 2724 min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES; 2725 if (opts->in_capsule_data_size < min_in_capsule_data_size) { 2726 SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n", 2727 min_in_capsule_data_size); 2728 opts->in_capsule_data_size = min_in_capsule_data_size; 2729 } 2730 2731 rtransport->event_channel = rdma_create_event_channel(); 2732 if (rtransport->event_channel == NULL) { 2733 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2734 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2735 return NULL; 2736 } 2737 2738 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2739 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2740 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2741 rtransport->event_channel->fd, spdk_strerror(errno)); 2742 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2743 return NULL; 2744 } 2745 2746 data_wr_pool_size = opts->data_wr_pool_size; 2747 if (data_wr_pool_size < SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count()) { 2748 data_wr_pool_size = SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count(); 2749 SPDK_NOTICELOG("data_wr_pool_size is changed to %zu to guarantee enough cache for handling " 2750 "at least one IO in each core\n", data_wr_pool_size); 2751 } 2752 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", data_wr_pool_size, 2753 sizeof(struct spdk_nvmf_rdma_request_data), SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2754 SPDK_ENV_SOCKET_ID_ANY); 2755 if (!rtransport->data_wr_pool) { 2756 if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) { 2757 SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n"); 2758 SPDK_ERRLOG("Probably running in multiprocess environment, which is " 2759 "unsupported by the nvmf library\n"); 2760 } else { 2761 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2762 } 2763 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2764 return NULL; 2765 } 2766 2767 contexts = rdma_get_devices(NULL); 2768 if (contexts == NULL) { 2769 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2770 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2771 return NULL; 2772 } 2773 2774 i = 0; 2775 rc = 0; 2776 while (contexts[i] != NULL) { 2777 rc = create_ib_device(rtransport, contexts[i], &device); 2778 if (rc < 0) { 2779 break; 2780 } 2781 i++; 2782 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2783 device->is_ready = true; 2784 } 2785 rdma_free_devices(contexts); 2786 2787 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2788 /* divide and round up. */ 2789 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2790 2791 /* round up to the nearest 4k. */ 2792 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2793 2794 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2795 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2796 opts->io_unit_size); 2797 } 2798 2799 if (rc < 0) { 2800 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2801 return NULL; 2802 } 2803 2804 rc = generate_poll_fds(rtransport); 2805 if (rc < 0) { 2806 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2807 return NULL; 2808 } 2809 2810 rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport, 2811 opts->acceptor_poll_rate); 2812 if (!rtransport->accept_poller) { 2813 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2814 return NULL; 2815 } 2816 2817 return &rtransport->transport; 2818 } 2819 2820 static void 2821 destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2822 struct spdk_nvmf_rdma_device *device) 2823 { 2824 TAILQ_REMOVE(&rtransport->devices, device, link); 2825 spdk_rdma_utils_free_mem_map(&device->map); 2826 if (device->pd) { 2827 if (!g_nvmf_hooks.get_ibv_pd) { 2828 ibv_dealloc_pd(device->pd); 2829 } 2830 } 2831 SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device); 2832 free(device); 2833 } 2834 2835 static void 2836 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2837 { 2838 struct spdk_nvmf_rdma_transport *rtransport; 2839 assert(w != NULL); 2840 2841 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2842 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2843 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2844 if (rtransport->rdma_opts.no_srq == true) { 2845 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2846 } 2847 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2848 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2849 } 2850 2851 static int 2852 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2853 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2854 { 2855 struct spdk_nvmf_rdma_transport *rtransport; 2856 struct spdk_nvmf_rdma_port *port, *port_tmp; 2857 struct spdk_nvmf_rdma_device *device, *device_tmp; 2858 2859 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2860 2861 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) { 2862 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 2863 free(port); 2864 } 2865 2866 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2867 TAILQ_REMOVE(&rtransport->ports, port, link); 2868 rdma_destroy_id(port->id); 2869 free(port); 2870 } 2871 2872 free_poll_fds(rtransport); 2873 2874 if (rtransport->event_channel != NULL) { 2875 rdma_destroy_event_channel(rtransport->event_channel); 2876 } 2877 2878 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2879 destroy_ib_device(rtransport, device); 2880 } 2881 2882 if (rtransport->data_wr_pool != NULL) { 2883 if (spdk_mempool_count(rtransport->data_wr_pool) != transport->opts.data_wr_pool_size) { 2884 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2885 spdk_mempool_count(rtransport->data_wr_pool), 2886 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2887 } 2888 } 2889 2890 spdk_mempool_free(rtransport->data_wr_pool); 2891 2892 spdk_poller_unregister(&rtransport->accept_poller); 2893 free(rtransport); 2894 2895 if (cb_fn) { 2896 cb_fn(cb_arg); 2897 } 2898 return 0; 2899 } 2900 2901 static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2902 struct spdk_nvme_transport_id *trid, 2903 bool peer); 2904 2905 static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport); 2906 2907 static int 2908 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2909 struct spdk_nvmf_listen_opts *listen_opts) 2910 { 2911 struct spdk_nvmf_rdma_transport *rtransport; 2912 struct spdk_nvmf_rdma_device *device; 2913 struct spdk_nvmf_rdma_port *port, *tmp_port; 2914 struct addrinfo *res; 2915 struct addrinfo hints; 2916 int family; 2917 int rc; 2918 long int port_val; 2919 bool is_retry = false; 2920 2921 if (!strlen(trid->trsvcid)) { 2922 SPDK_ERRLOG("Service id is required\n"); 2923 return -EINVAL; 2924 } 2925 2926 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2927 assert(rtransport->event_channel != NULL); 2928 2929 port = calloc(1, sizeof(*port)); 2930 if (!port) { 2931 SPDK_ERRLOG("Port allocation failed\n"); 2932 return -ENOMEM; 2933 } 2934 2935 port->trid = trid; 2936 2937 switch (trid->adrfam) { 2938 case SPDK_NVMF_ADRFAM_IPV4: 2939 family = AF_INET; 2940 break; 2941 case SPDK_NVMF_ADRFAM_IPV6: 2942 family = AF_INET6; 2943 break; 2944 default: 2945 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2946 free(port); 2947 return -EINVAL; 2948 } 2949 2950 memset(&hints, 0, sizeof(hints)); 2951 hints.ai_family = family; 2952 hints.ai_flags = AI_NUMERICSERV; 2953 hints.ai_socktype = SOCK_STREAM; 2954 hints.ai_protocol = 0; 2955 2956 /* Range check the trsvcid. Fail in 3 cases: 2957 * < 0: means that spdk_strtol hit an error 2958 * 0: this results in ephemeral port which we don't want 2959 * > 65535: port too high 2960 */ 2961 port_val = spdk_strtol(trid->trsvcid, 10); 2962 if (port_val <= 0 || port_val > 65535) { 2963 SPDK_ERRLOG("invalid trsvcid %s\n", trid->trsvcid); 2964 free(port); 2965 return -EINVAL; 2966 } 2967 2968 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2969 if (rc) { 2970 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2971 free(port); 2972 return -(abs(rc)); 2973 } 2974 2975 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2976 if (rc < 0) { 2977 SPDK_ERRLOG("rdma_create_id() failed\n"); 2978 freeaddrinfo(res); 2979 free(port); 2980 return rc; 2981 } 2982 2983 rc = rdma_bind_addr(port->id, res->ai_addr); 2984 freeaddrinfo(res); 2985 2986 if (rc < 0) { 2987 TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) { 2988 if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) { 2989 is_retry = true; 2990 break; 2991 } 2992 } 2993 if (!is_retry) { 2994 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2995 } 2996 rdma_destroy_id(port->id); 2997 free(port); 2998 return rc; 2999 } 3000 3001 if (!port->id->verbs) { 3002 SPDK_ERRLOG("ibv_context is null\n"); 3003 rdma_destroy_id(port->id); 3004 free(port); 3005 return -1; 3006 } 3007 3008 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 3009 if (rc < 0) { 3010 SPDK_ERRLOG("rdma_listen() failed\n"); 3011 rdma_destroy_id(port->id); 3012 free(port); 3013 return rc; 3014 } 3015 3016 TAILQ_FOREACH(device, &rtransport->devices, link) { 3017 if (device->context == port->id->verbs && device->is_ready) { 3018 port->device = device; 3019 break; 3020 } 3021 } 3022 if (!port->device) { 3023 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 3024 port->id->verbs); 3025 rdma_destroy_id(port->id); 3026 free(port); 3027 nvmf_rdma_rescan_devices(rtransport); 3028 return -EINVAL; 3029 } 3030 3031 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 3032 trid->traddr, trid->trsvcid); 3033 3034 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 3035 return 0; 3036 } 3037 3038 static void 3039 nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport, 3040 const struct spdk_nvme_transport_id *trid, bool need_retry) 3041 { 3042 struct spdk_nvmf_rdma_transport *rtransport; 3043 struct spdk_nvmf_rdma_port *port, *tmp; 3044 3045 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3046 3047 if (!need_retry) { 3048 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) { 3049 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 3050 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3051 free(port); 3052 } 3053 } 3054 } 3055 3056 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 3057 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 3058 SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n", 3059 port->trid->traddr, port->trid->trsvcid, need_retry); 3060 TAILQ_REMOVE(&rtransport->ports, port, link); 3061 rdma_destroy_id(port->id); 3062 port->id = NULL; 3063 port->device = NULL; 3064 if (need_retry) { 3065 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3066 } else { 3067 free(port); 3068 } 3069 break; 3070 } 3071 } 3072 } 3073 3074 static void 3075 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 3076 const struct spdk_nvme_transport_id *trid) 3077 { 3078 nvmf_rdma_stop_listen_ex(transport, trid, false); 3079 } 3080 3081 static void _nvmf_rdma_register_poller_in_group(void *c); 3082 static void _nvmf_rdma_remove_poller_in_group(void *c); 3083 3084 static bool 3085 nvmf_rdma_all_pollers_management_done(void *c) 3086 { 3087 struct poller_manage_ctx *ctx = c; 3088 int counter; 3089 3090 counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST); 3091 SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n", 3092 counter, ctx->rpoller); 3093 3094 if (counter == 0) { 3095 free((void *)ctx->inflight_op_counter); 3096 } 3097 free(ctx); 3098 3099 return counter == 0; 3100 } 3101 3102 static int 3103 nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport, 3104 struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add) 3105 { 3106 struct spdk_nvmf_rdma_poll_group *rgroup; 3107 struct spdk_nvmf_rdma_poller *rpoller; 3108 struct spdk_nvmf_poll_group *poll_group; 3109 struct poller_manage_ctx *ctx; 3110 bool found; 3111 int *inflight_counter; 3112 spdk_msg_fn do_fn; 3113 3114 *has_inflight = false; 3115 do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group; 3116 inflight_counter = calloc(1, sizeof(int)); 3117 if (!inflight_counter) { 3118 SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n"); 3119 return -ENOMEM; 3120 } 3121 3122 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3123 (*inflight_counter)++; 3124 } 3125 3126 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3127 found = false; 3128 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3129 if (rpoller->device == device) { 3130 found = true; 3131 break; 3132 } 3133 } 3134 if (found == is_add) { 3135 __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST); 3136 continue; 3137 } 3138 3139 ctx = calloc(1, sizeof(struct poller_manage_ctx)); 3140 if (!ctx) { 3141 SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n"); 3142 if (!*has_inflight) { 3143 free(inflight_counter); 3144 } 3145 return -ENOMEM; 3146 } 3147 3148 ctx->rtransport = rtransport; 3149 ctx->rgroup = rgroup; 3150 ctx->rpoller = rpoller; 3151 ctx->device = device; 3152 ctx->thread = spdk_get_thread(); 3153 ctx->inflight_op_counter = inflight_counter; 3154 *has_inflight = true; 3155 3156 poll_group = rgroup->group.group; 3157 if (poll_group->thread != spdk_get_thread()) { 3158 spdk_thread_send_msg(poll_group->thread, do_fn, ctx); 3159 } else { 3160 do_fn(ctx); 3161 } 3162 } 3163 3164 if (!*has_inflight) { 3165 free(inflight_counter); 3166 } 3167 3168 return 0; 3169 } 3170 3171 static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3172 struct spdk_nvmf_rdma_device *device); 3173 3174 static struct spdk_nvmf_rdma_device * 3175 nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 3176 struct ibv_context *context) 3177 { 3178 struct spdk_nvmf_rdma_device *device, *tmp_device; 3179 3180 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) { 3181 if (device->need_destroy) { 3182 continue; 3183 } 3184 3185 if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) { 3186 return device; 3187 } 3188 } 3189 3190 return NULL; 3191 } 3192 3193 static bool 3194 nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport, 3195 struct ibv_context *context) 3196 { 3197 struct spdk_nvmf_rdma_device *old_device, *new_device; 3198 int rc = 0; 3199 bool has_inflight; 3200 3201 old_device = nvmf_rdma_find_ib_device(rtransport, context); 3202 3203 if (old_device) { 3204 if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) { 3205 /* context may not have time to be cleaned when rescan. exactly one context 3206 * is valid for a device so this context must be invalid and just remove it. */ 3207 SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context); 3208 old_device->need_destroy = true; 3209 nvmf_rdma_handle_device_removal(rtransport, old_device); 3210 } 3211 return false; 3212 } 3213 3214 rc = create_ib_device(rtransport, context, &new_device); 3215 /* TODO: update transport opts. */ 3216 if (rc < 0) { 3217 SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n", 3218 ibv_get_device_name(context->device), context); 3219 return false; 3220 } 3221 3222 rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true); 3223 if (rc < 0) { 3224 SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n", 3225 ibv_get_device_name(context->device), context); 3226 return false; 3227 } 3228 3229 if (has_inflight) { 3230 new_device->is_ready = true; 3231 } 3232 3233 return true; 3234 } 3235 3236 static bool 3237 nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport) 3238 { 3239 struct spdk_nvmf_rdma_device *device; 3240 struct ibv_device **ibv_device_list = NULL; 3241 struct ibv_context **contexts = NULL; 3242 int i = 0; 3243 int num_dev = 0; 3244 bool new_create = false, has_new_device = false; 3245 struct ibv_context *tmp_verbs = NULL; 3246 3247 /* do not rescan when any device is destroying, or context may be freed when 3248 * regenerating the poll fds. 3249 */ 3250 TAILQ_FOREACH(device, &rtransport->devices, link) { 3251 if (device->need_destroy) { 3252 return false; 3253 } 3254 } 3255 3256 ibv_device_list = ibv_get_device_list(&num_dev); 3257 3258 /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be 3259 * marked as dead verbs and never be init again. So we need to make sure the 3260 * verbs is available before we call rdma_get_devices. */ 3261 if (num_dev >= 0) { 3262 for (i = 0; i < num_dev; i++) { 3263 tmp_verbs = ibv_open_device(ibv_device_list[i]); 3264 if (!tmp_verbs) { 3265 SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno); 3266 break; 3267 } 3268 if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) { 3269 SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i], 3270 tmp_verbs->device->dev_name); 3271 has_new_device = true; 3272 } 3273 ibv_close_device(tmp_verbs); 3274 } 3275 ibv_free_device_list(ibv_device_list); 3276 if (!tmp_verbs || !has_new_device) { 3277 return false; 3278 } 3279 } 3280 3281 contexts = rdma_get_devices(NULL); 3282 3283 for (i = 0; contexts && contexts[i] != NULL; i++) { 3284 new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]); 3285 } 3286 3287 if (new_create) { 3288 free_poll_fds(rtransport); 3289 generate_poll_fds(rtransport); 3290 } 3291 3292 if (contexts) { 3293 rdma_free_devices(contexts); 3294 } 3295 3296 return new_create; 3297 } 3298 3299 static bool 3300 nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport) 3301 { 3302 struct spdk_nvmf_rdma_port *port, *tmp_port; 3303 int rc = 0; 3304 bool new_create = false; 3305 3306 if (TAILQ_EMPTY(&rtransport->retry_ports)) { 3307 return false; 3308 } 3309 3310 new_create = nvmf_rdma_rescan_devices(rtransport); 3311 3312 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) { 3313 rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL); 3314 3315 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3316 if (rc) { 3317 if (new_create) { 3318 SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n", 3319 port->trid->traddr, port->trid->trsvcid, rc); 3320 } 3321 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3322 break; 3323 } else { 3324 SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid); 3325 free(port); 3326 } 3327 } 3328 3329 return true; 3330 } 3331 3332 static void 3333 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 3334 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 3335 { 3336 struct spdk_nvmf_request *req, *tmp; 3337 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 3338 struct spdk_nvmf_rdma_resources *resources; 3339 3340 /* First process requests which are waiting for response to be sent */ 3341 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_send_queue, state_link, req_tmp) { 3342 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3343 break; 3344 } 3345 } 3346 3347 /* We process I/O in the data transfer pending queue at the highest priority. */ 3348 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 3349 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3350 break; 3351 } 3352 } 3353 3354 /* Then RDMA writes since reads have stronger restrictions than writes */ 3355 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 3356 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3357 break; 3358 } 3359 } 3360 3361 /* Then we handle request waiting on memory buffers. */ 3362 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 3363 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3364 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3365 break; 3366 } 3367 } 3368 3369 resources = rqpair->resources; 3370 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 3371 rdma_req = STAILQ_FIRST(&resources->free_queue); 3372 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 3373 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 3374 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 3375 3376 if (rqpair->srq != NULL) { 3377 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 3378 rdma_req->recv->qpair->qd++; 3379 } else { 3380 rqpair->qd++; 3381 } 3382 3383 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 3384 rdma_req->state = RDMA_REQUEST_STATE_NEW; 3385 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 3386 break; 3387 } 3388 } 3389 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 3390 rqpair->poller->stat.pending_free_request++; 3391 } 3392 } 3393 3394 static void 3395 nvmf_rdma_poller_process_pending_buf_queue(struct spdk_nvmf_rdma_transport *rtransport, 3396 struct spdk_nvmf_rdma_poller *rpoller) 3397 { 3398 struct spdk_nvmf_request *req, *tmp; 3399 struct spdk_nvmf_rdma_request *rdma_req; 3400 3401 STAILQ_FOREACH_SAFE(req, &rpoller->group->group.pending_buf_queue, buf_link, tmp) { 3402 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3403 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 3404 break; 3405 } 3406 } 3407 } 3408 3409 static inline bool 3410 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 3411 { 3412 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 3413 return nvmf_rdma_is_rxe_device(device) || 3414 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 3415 } 3416 3417 static void 3418 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 3419 { 3420 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3421 struct spdk_nvmf_rdma_transport, transport); 3422 3423 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3424 3425 /* nvmf_rdma_close_qpair is not called */ 3426 if (!rqpair->to_close) { 3427 return; 3428 } 3429 3430 /* device is already destroyed and we should force destroy this qpair. */ 3431 if (rqpair->poller && rqpair->poller->need_destroy) { 3432 nvmf_rdma_qpair_destroy(rqpair); 3433 return; 3434 } 3435 3436 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 3437 if (rqpair->current_send_depth != 0) { 3438 return; 3439 } 3440 3441 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 3442 return; 3443 } 3444 3445 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 3446 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 3447 return; 3448 } 3449 3450 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 3451 3452 nvmf_rdma_qpair_destroy(rqpair); 3453 } 3454 3455 static int 3456 nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked) 3457 { 3458 struct spdk_nvmf_qpair *qpair; 3459 struct spdk_nvmf_rdma_qpair *rqpair; 3460 3461 if (evt->id == NULL) { 3462 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 3463 return -1; 3464 } 3465 3466 qpair = evt->id->context; 3467 if (qpair == NULL) { 3468 SPDK_ERRLOG("disconnect request: no active connection\n"); 3469 return -1; 3470 } 3471 3472 rdma_ack_cm_event(evt); 3473 *event_acked = true; 3474 3475 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3476 3477 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair); 3478 3479 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3480 3481 return 0; 3482 } 3483 3484 #ifdef DEBUG 3485 static const char *CM_EVENT_STR[] = { 3486 "RDMA_CM_EVENT_ADDR_RESOLVED", 3487 "RDMA_CM_EVENT_ADDR_ERROR", 3488 "RDMA_CM_EVENT_ROUTE_RESOLVED", 3489 "RDMA_CM_EVENT_ROUTE_ERROR", 3490 "RDMA_CM_EVENT_CONNECT_REQUEST", 3491 "RDMA_CM_EVENT_CONNECT_RESPONSE", 3492 "RDMA_CM_EVENT_CONNECT_ERROR", 3493 "RDMA_CM_EVENT_UNREACHABLE", 3494 "RDMA_CM_EVENT_REJECTED", 3495 "RDMA_CM_EVENT_ESTABLISHED", 3496 "RDMA_CM_EVENT_DISCONNECTED", 3497 "RDMA_CM_EVENT_DEVICE_REMOVAL", 3498 "RDMA_CM_EVENT_MULTICAST_JOIN", 3499 "RDMA_CM_EVENT_MULTICAST_ERROR", 3500 "RDMA_CM_EVENT_ADDR_CHANGE", 3501 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 3502 }; 3503 #endif /* DEBUG */ 3504 3505 static void 3506 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 3507 struct spdk_nvmf_rdma_port *port) 3508 { 3509 struct spdk_nvmf_rdma_poll_group *rgroup; 3510 struct spdk_nvmf_rdma_poller *rpoller; 3511 struct spdk_nvmf_rdma_qpair *rqpair; 3512 3513 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3514 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3515 RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) { 3516 if (rqpair->listen_id == port->id) { 3517 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3518 } 3519 } 3520 } 3521 } 3522 } 3523 3524 static bool 3525 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 3526 struct rdma_cm_event *event) 3527 { 3528 const struct spdk_nvme_transport_id *trid; 3529 struct spdk_nvmf_rdma_port *port; 3530 struct spdk_nvmf_rdma_transport *rtransport; 3531 bool event_acked = false; 3532 3533 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3534 TAILQ_FOREACH(port, &rtransport->ports, link) { 3535 if (port->id == event->id) { 3536 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 3537 rdma_ack_cm_event(event); 3538 event_acked = true; 3539 trid = port->trid; 3540 break; 3541 } 3542 } 3543 3544 if (event_acked) { 3545 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 3546 3547 nvmf_rdma_stop_listen(transport, trid); 3548 nvmf_rdma_listen(transport, trid, NULL); 3549 } 3550 3551 return event_acked; 3552 } 3553 3554 static void 3555 nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3556 struct spdk_nvmf_rdma_device *device) 3557 { 3558 struct spdk_nvmf_rdma_port *port, *port_tmp; 3559 int rc; 3560 bool has_inflight; 3561 3562 rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false); 3563 if (rc) { 3564 SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc); 3565 return; 3566 } 3567 3568 if (!has_inflight) { 3569 /* no pollers, destroy the device */ 3570 device->ready_to_destroy = true; 3571 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport); 3572 } 3573 3574 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 3575 if (port->device == device) { 3576 SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n", 3577 port->trid->traddr, 3578 port->trid->trsvcid, 3579 ibv_get_device_name(port->device->context->device)); 3580 3581 /* keep NVMF listener and only destroy structures of the 3582 * RDMA transport. when the device comes back we can retry listening 3583 * and the application's workflow will not be interrupted. 3584 */ 3585 nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true); 3586 } 3587 } 3588 } 3589 3590 static void 3591 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 3592 struct rdma_cm_event *event) 3593 { 3594 struct spdk_nvmf_rdma_port *port, *tmp_port; 3595 struct spdk_nvmf_rdma_transport *rtransport; 3596 3597 port = event->id->context; 3598 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3599 3600 rdma_ack_cm_event(event); 3601 3602 /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive 3603 * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that 3604 * we are handling a port event here. 3605 */ 3606 TAILQ_FOREACH(tmp_port, &rtransport->ports, link) { 3607 if (port == tmp_port && port->device && !port->device->need_destroy) { 3608 port->device->need_destroy = true; 3609 nvmf_rdma_handle_device_removal(rtransport, port->device); 3610 } 3611 } 3612 } 3613 3614 static void 3615 nvmf_process_cm_events(struct spdk_nvmf_transport *transport, uint32_t max_events) 3616 { 3617 struct spdk_nvmf_rdma_transport *rtransport; 3618 struct rdma_cm_event *event; 3619 uint32_t i; 3620 int rc; 3621 bool event_acked; 3622 3623 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3624 3625 if (rtransport->event_channel == NULL) { 3626 return; 3627 } 3628 3629 for (i = 0; i < max_events; i++) { 3630 event_acked = false; 3631 rc = rdma_get_cm_event(rtransport->event_channel, &event); 3632 if (rc) { 3633 if (errno != EAGAIN && errno != EWOULDBLOCK) { 3634 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 3635 } 3636 break; 3637 } 3638 3639 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 3640 3641 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 3642 3643 switch (event->event) { 3644 case RDMA_CM_EVENT_ADDR_RESOLVED: 3645 case RDMA_CM_EVENT_ADDR_ERROR: 3646 case RDMA_CM_EVENT_ROUTE_RESOLVED: 3647 case RDMA_CM_EVENT_ROUTE_ERROR: 3648 /* No action required. The target never attempts to resolve routes. */ 3649 break; 3650 case RDMA_CM_EVENT_CONNECT_REQUEST: 3651 rc = nvmf_rdma_connect(transport, event); 3652 if (rc < 0) { 3653 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 3654 break; 3655 } 3656 break; 3657 case RDMA_CM_EVENT_CONNECT_RESPONSE: 3658 /* The target never initiates a new connection. So this will not occur. */ 3659 break; 3660 case RDMA_CM_EVENT_CONNECT_ERROR: 3661 /* Can this happen? The docs say it can, but not sure what causes it. */ 3662 break; 3663 case RDMA_CM_EVENT_UNREACHABLE: 3664 case RDMA_CM_EVENT_REJECTED: 3665 /* These only occur on the client side. */ 3666 break; 3667 case RDMA_CM_EVENT_ESTABLISHED: 3668 /* TODO: Should we be waiting for this event anywhere? */ 3669 break; 3670 case RDMA_CM_EVENT_DISCONNECTED: 3671 rc = nvmf_rdma_disconnect(event, &event_acked); 3672 if (rc < 0) { 3673 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3674 break; 3675 } 3676 break; 3677 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3678 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3679 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3680 * Once these events are sent to SPDK, we should release all IB resources and 3681 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3682 * ibv_destroy* functions to release user space memory allocated by IB. All kernel 3683 * resources are already cleaned. */ 3684 if (event->id->qp) { 3685 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3686 * corresponding qpair. Otherwise the event refers to a listening device. */ 3687 rc = nvmf_rdma_disconnect(event, &event_acked); 3688 if (rc < 0) { 3689 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3690 break; 3691 } 3692 } else { 3693 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3694 event_acked = true; 3695 } 3696 break; 3697 case RDMA_CM_EVENT_MULTICAST_JOIN: 3698 case RDMA_CM_EVENT_MULTICAST_ERROR: 3699 /* Multicast is not used */ 3700 break; 3701 case RDMA_CM_EVENT_ADDR_CHANGE: 3702 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3703 break; 3704 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3705 /* For now, do nothing. The target never re-uses queue pairs. */ 3706 break; 3707 default: 3708 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3709 break; 3710 } 3711 if (!event_acked) { 3712 rdma_ack_cm_event(event); 3713 } 3714 } 3715 } 3716 3717 static void 3718 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3719 { 3720 rqpair->last_wqe_reached = true; 3721 nvmf_rdma_destroy_drained_qpair(rqpair); 3722 } 3723 3724 static void 3725 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3726 { 3727 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3728 3729 if (event_ctx->rqpair) { 3730 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3731 if (event_ctx->cb_fn) { 3732 event_ctx->cb_fn(event_ctx->rqpair); 3733 } 3734 } 3735 free(event_ctx); 3736 } 3737 3738 static int 3739 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3740 spdk_nvmf_rdma_qpair_ibv_event fn) 3741 { 3742 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3743 struct spdk_thread *thr = NULL; 3744 int rc; 3745 3746 if (rqpair->qpair.group) { 3747 thr = rqpair->qpair.group->thread; 3748 } else if (rqpair->destruct_channel) { 3749 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3750 } 3751 3752 if (!thr) { 3753 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3754 return -EINVAL; 3755 } 3756 3757 ctx = calloc(1, sizeof(*ctx)); 3758 if (!ctx) { 3759 return -ENOMEM; 3760 } 3761 3762 ctx->rqpair = rqpair; 3763 ctx->cb_fn = fn; 3764 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3765 3766 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3767 if (rc) { 3768 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3769 free(ctx); 3770 } 3771 3772 return rc; 3773 } 3774 3775 static int 3776 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3777 { 3778 int rc; 3779 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3780 struct ibv_async_event event; 3781 3782 rc = ibv_get_async_event(device->context, &event); 3783 3784 if (rc) { 3785 /* In non-blocking mode -1 means there are no events available */ 3786 return rc; 3787 } 3788 3789 switch (event.event_type) { 3790 case IBV_EVENT_QP_FATAL: 3791 case IBV_EVENT_QP_LAST_WQE_REACHED: 3792 case IBV_EVENT_QP_REQ_ERR: 3793 case IBV_EVENT_QP_ACCESS_ERR: 3794 case IBV_EVENT_COMM_EST: 3795 case IBV_EVENT_PATH_MIG: 3796 case IBV_EVENT_PATH_MIG_ERR: 3797 rqpair = event.element.qp->qp_context; 3798 if (!rqpair) { 3799 /* Any QP event for NVMe-RDMA initiator may be returned. */ 3800 SPDK_NOTICELOG("Async QP event for unknown QP: %s\n", 3801 ibv_event_type_str(event.event_type)); 3802 break; 3803 } 3804 3805 switch (event.event_type) { 3806 case IBV_EVENT_QP_FATAL: 3807 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3808 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3809 (uintptr_t)rqpair, event.event_type); 3810 rqpair->ibv_in_error_state = true; 3811 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3812 break; 3813 case IBV_EVENT_QP_LAST_WQE_REACHED: 3814 /* This event only occurs for shared receive queues. */ 3815 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3816 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3817 if (rc) { 3818 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3819 rqpair->last_wqe_reached = true; 3820 } 3821 break; 3822 case IBV_EVENT_QP_REQ_ERR: 3823 case IBV_EVENT_QP_ACCESS_ERR: 3824 case IBV_EVENT_COMM_EST: 3825 case IBV_EVENT_PATH_MIG: 3826 case IBV_EVENT_PATH_MIG_ERR: 3827 SPDK_NOTICELOG("Async QP event: %s\n", 3828 ibv_event_type_str(event.event_type)); 3829 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3830 (uintptr_t)rqpair, event.event_type); 3831 rqpair->ibv_in_error_state = true; 3832 break; 3833 default: 3834 break; 3835 } 3836 break; 3837 case IBV_EVENT_DEVICE_FATAL: 3838 SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n", 3839 ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device); 3840 device->need_destroy = true; 3841 break; 3842 case IBV_EVENT_CQ_ERR: 3843 case IBV_EVENT_PORT_ACTIVE: 3844 case IBV_EVENT_PORT_ERR: 3845 case IBV_EVENT_LID_CHANGE: 3846 case IBV_EVENT_PKEY_CHANGE: 3847 case IBV_EVENT_SM_CHANGE: 3848 case IBV_EVENT_SRQ_ERR: 3849 case IBV_EVENT_SRQ_LIMIT_REACHED: 3850 case IBV_EVENT_CLIENT_REREGISTER: 3851 case IBV_EVENT_GID_CHANGE: 3852 case IBV_EVENT_SQ_DRAINED: 3853 default: 3854 SPDK_NOTICELOG("Async event: %s\n", 3855 ibv_event_type_str(event.event_type)); 3856 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3857 break; 3858 } 3859 ibv_ack_async_event(&event); 3860 3861 return 0; 3862 } 3863 3864 static void 3865 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3866 { 3867 int rc = 0; 3868 uint32_t i = 0; 3869 3870 for (i = 0; i < max_events; i++) { 3871 rc = nvmf_process_ib_event(device); 3872 if (rc) { 3873 break; 3874 } 3875 } 3876 3877 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3878 } 3879 3880 static int 3881 nvmf_rdma_accept(void *ctx) 3882 { 3883 int nfds, i = 0; 3884 struct spdk_nvmf_transport *transport = ctx; 3885 struct spdk_nvmf_rdma_transport *rtransport; 3886 struct spdk_nvmf_rdma_device *device, *tmp; 3887 uint32_t count; 3888 short revents; 3889 bool do_retry; 3890 3891 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3892 do_retry = nvmf_rdma_retry_listen_port(rtransport); 3893 3894 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3895 3896 if (nfds <= 0) { 3897 return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3898 } 3899 3900 /* The first poll descriptor is RDMA CM event */ 3901 if (rtransport->poll_fds[i++].revents & POLLIN) { 3902 nvmf_process_cm_events(transport, NVMF_RDMA_MAX_EVENTS_PER_POLL); 3903 nfds--; 3904 } 3905 3906 if (nfds == 0) { 3907 return SPDK_POLLER_BUSY; 3908 } 3909 3910 /* Second and subsequent poll descriptors are IB async events */ 3911 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3912 revents = rtransport->poll_fds[i++].revents; 3913 if (revents & POLLIN) { 3914 if (spdk_likely(!device->need_destroy)) { 3915 nvmf_process_ib_events(device, NVMF_RDMA_MAX_EVENTS_PER_POLL); 3916 if (spdk_unlikely(device->need_destroy)) { 3917 nvmf_rdma_handle_device_removal(rtransport, device); 3918 } 3919 } 3920 nfds--; 3921 } else if (revents & POLLNVAL || revents & POLLHUP) { 3922 SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device); 3923 nfds--; 3924 } 3925 } 3926 /* check all flagged fd's have been served */ 3927 assert(nfds == 0); 3928 3929 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3930 } 3931 3932 static void 3933 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3934 struct spdk_nvmf_ctrlr_data *cdata) 3935 { 3936 cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD; 3937 3938 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3939 since in-capsule data only works with NVME drives that support SGL memory layout */ 3940 if (transport->opts.dif_insert_or_strip) { 3941 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3942 } 3943 3944 if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) { 3945 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule" 3946 " data is greater than 4KiB.\n"); 3947 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux " 3948 "kernel between versions 5.4 and 5.12 data corruption may occur for " 3949 "writes that are not a multiple of 4KiB in size.\n"); 3950 } 3951 } 3952 3953 static void 3954 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3955 struct spdk_nvme_transport_id *trid, 3956 struct spdk_nvmf_discovery_log_page_entry *entry) 3957 { 3958 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3959 entry->adrfam = trid->adrfam; 3960 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3961 3962 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3963 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3964 3965 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3966 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3967 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3968 } 3969 3970 static int 3971 nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport, 3972 struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device, 3973 struct spdk_nvmf_rdma_poller **out_poller) 3974 { 3975 struct spdk_nvmf_rdma_poller *poller; 3976 struct spdk_rdma_provider_srq_init_attr srq_init_attr; 3977 struct spdk_nvmf_rdma_resource_opts opts; 3978 int num_cqe; 3979 3980 poller = calloc(1, sizeof(*poller)); 3981 if (!poller) { 3982 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3983 return -1; 3984 } 3985 3986 poller->device = device; 3987 poller->group = rgroup; 3988 *out_poller = poller; 3989 3990 RB_INIT(&poller->qpairs); 3991 STAILQ_INIT(&poller->qpairs_pending_send); 3992 STAILQ_INIT(&poller->qpairs_pending_recv); 3993 3994 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3995 SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup); 3996 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3997 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) { 3998 SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n", 3999 rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr); 4000 } 4001 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr); 4002 4003 device->num_srq++; 4004 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 4005 srq_init_attr.pd = device->pd; 4006 srq_init_attr.stats = &poller->stat.qp_stats.recv; 4007 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 4008 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 4009 poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr); 4010 if (!poller->srq) { 4011 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 4012 return -1; 4013 } 4014 4015 opts.qp = poller->srq; 4016 opts.map = device->map; 4017 opts.qpair = NULL; 4018 opts.shared = true; 4019 opts.max_queue_depth = poller->max_srq_depth; 4020 opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size; 4021 4022 poller->resources = nvmf_rdma_resources_create(&opts); 4023 if (!poller->resources) { 4024 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 4025 return -1; 4026 } 4027 } 4028 4029 /* 4030 * When using an srq, we can limit the completion queue at startup. 4031 * The following formula represents the calculation: 4032 * num_cqe = num_recv + num_data_wr + num_send_wr. 4033 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 4034 */ 4035 if (poller->srq) { 4036 num_cqe = poller->max_srq_depth * 3; 4037 } else { 4038 num_cqe = rtransport->rdma_opts.num_cqe; 4039 } 4040 4041 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 4042 if (!poller->cq) { 4043 SPDK_ERRLOG("Unable to create completion queue\n"); 4044 return -1; 4045 } 4046 poller->num_cqe = num_cqe; 4047 return 0; 4048 } 4049 4050 static void 4051 _nvmf_rdma_register_poller_in_group(void *c) 4052 { 4053 struct spdk_nvmf_rdma_poller *poller; 4054 struct poller_manage_ctx *ctx = c; 4055 struct spdk_nvmf_rdma_device *device; 4056 int rc; 4057 4058 rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller); 4059 if (rc < 0 && poller) { 4060 nvmf_rdma_poller_destroy(poller); 4061 } 4062 4063 device = ctx->device; 4064 if (nvmf_rdma_all_pollers_management_done(ctx)) { 4065 device->is_ready = true; 4066 } 4067 } 4068 4069 static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 4070 4071 static struct spdk_nvmf_transport_poll_group * 4072 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport, 4073 struct spdk_nvmf_poll_group *group) 4074 { 4075 struct spdk_nvmf_rdma_transport *rtransport; 4076 struct spdk_nvmf_rdma_poll_group *rgroup; 4077 struct spdk_nvmf_rdma_poller *poller; 4078 struct spdk_nvmf_rdma_device *device; 4079 int rc; 4080 4081 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 4082 4083 rgroup = calloc(1, sizeof(*rgroup)); 4084 if (!rgroup) { 4085 return NULL; 4086 } 4087 4088 TAILQ_INIT(&rgroup->pollers); 4089 4090 TAILQ_FOREACH(device, &rtransport->devices, link) { 4091 rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller); 4092 if (rc < 0) { 4093 nvmf_rdma_poll_group_destroy(&rgroup->group); 4094 return NULL; 4095 } 4096 } 4097 4098 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 4099 if (rtransport->conn_sched.next_admin_pg == NULL) { 4100 rtransport->conn_sched.next_admin_pg = rgroup; 4101 rtransport->conn_sched.next_io_pg = rgroup; 4102 } 4103 4104 return &rgroup->group; 4105 } 4106 4107 static uint32_t 4108 nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg) 4109 { 4110 uint32_t count; 4111 4112 /* Just assume that unassociated qpairs will eventually be io 4113 * qpairs. This is close enough for the use cases for this 4114 * function. 4115 */ 4116 pthread_mutex_lock(&pg->mutex); 4117 count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs; 4118 pthread_mutex_unlock(&pg->mutex); 4119 4120 return count; 4121 } 4122 4123 static struct spdk_nvmf_transport_poll_group * 4124 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 4125 { 4126 struct spdk_nvmf_rdma_transport *rtransport; 4127 struct spdk_nvmf_rdma_poll_group **pg; 4128 struct spdk_nvmf_transport_poll_group *result; 4129 uint32_t count; 4130 4131 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4132 4133 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 4134 return NULL; 4135 } 4136 4137 if (qpair->qid == 0) { 4138 pg = &rtransport->conn_sched.next_admin_pg; 4139 } else { 4140 struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current; 4141 uint32_t min_value; 4142 4143 pg = &rtransport->conn_sched.next_io_pg; 4144 pg_min = *pg; 4145 pg_start = *pg; 4146 pg_current = *pg; 4147 min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 4148 4149 while (1) { 4150 count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 4151 4152 if (count < min_value) { 4153 min_value = count; 4154 pg_min = pg_current; 4155 } 4156 4157 pg_current = TAILQ_NEXT(pg_current, link); 4158 if (pg_current == NULL) { 4159 pg_current = TAILQ_FIRST(&rtransport->poll_groups); 4160 } 4161 4162 if (pg_current == pg_start || min_value == 0) { 4163 break; 4164 } 4165 } 4166 *pg = pg_min; 4167 } 4168 4169 assert(*pg != NULL); 4170 4171 result = &(*pg)->group; 4172 4173 *pg = TAILQ_NEXT(*pg, link); 4174 if (*pg == NULL) { 4175 *pg = TAILQ_FIRST(&rtransport->poll_groups); 4176 } 4177 4178 return result; 4179 } 4180 4181 static void 4182 nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller) 4183 { 4184 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 4185 int rc; 4186 4187 TAILQ_REMOVE(&poller->group->pollers, poller, link); 4188 RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) { 4189 nvmf_rdma_qpair_destroy(qpair); 4190 } 4191 4192 if (poller->srq) { 4193 if (poller->resources) { 4194 nvmf_rdma_resources_destroy(poller->resources); 4195 } 4196 spdk_rdma_provider_srq_destroy(poller->srq); 4197 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 4198 } 4199 4200 if (poller->cq) { 4201 rc = ibv_destroy_cq(poller->cq); 4202 if (rc != 0) { 4203 SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno)); 4204 } 4205 } 4206 4207 if (poller->destroy_cb) { 4208 poller->destroy_cb(poller->destroy_cb_ctx); 4209 poller->destroy_cb = NULL; 4210 } 4211 4212 free(poller); 4213 } 4214 4215 static void 4216 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 4217 { 4218 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 4219 struct spdk_nvmf_rdma_poller *poller, *tmp; 4220 struct spdk_nvmf_rdma_transport *rtransport; 4221 4222 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4223 if (!rgroup) { 4224 return; 4225 } 4226 4227 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 4228 nvmf_rdma_poller_destroy(poller); 4229 } 4230 4231 if (rgroup->group.transport == NULL) { 4232 /* Transport can be NULL when nvmf_rdma_poll_group_create() 4233 * calls this function directly in a failure path. */ 4234 free(rgroup); 4235 return; 4236 } 4237 4238 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 4239 4240 next_rgroup = TAILQ_NEXT(rgroup, link); 4241 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 4242 if (next_rgroup == NULL) { 4243 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 4244 } 4245 if (rtransport->conn_sched.next_admin_pg == rgroup) { 4246 rtransport->conn_sched.next_admin_pg = next_rgroup; 4247 } 4248 if (rtransport->conn_sched.next_io_pg == rgroup) { 4249 rtransport->conn_sched.next_io_pg = next_rgroup; 4250 } 4251 4252 free(rgroup); 4253 } 4254 4255 static void 4256 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 4257 { 4258 if (rqpair->cm_id != NULL) { 4259 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 4260 } 4261 } 4262 4263 static int 4264 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 4265 struct spdk_nvmf_qpair *qpair) 4266 { 4267 struct spdk_nvmf_rdma_poll_group *rgroup; 4268 struct spdk_nvmf_rdma_qpair *rqpair; 4269 struct spdk_nvmf_rdma_device *device; 4270 struct spdk_nvmf_rdma_poller *poller; 4271 int rc; 4272 4273 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4274 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4275 4276 device = rqpair->device; 4277 4278 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 4279 if (poller->device == device) { 4280 break; 4281 } 4282 } 4283 4284 if (!poller) { 4285 SPDK_ERRLOG("No poller found for device.\n"); 4286 return -1; 4287 } 4288 4289 if (poller->need_destroy) { 4290 SPDK_ERRLOG("Poller is destroying.\n"); 4291 return -1; 4292 } 4293 4294 rqpair->poller = poller; 4295 rqpair->srq = rqpair->poller->srq; 4296 4297 rc = nvmf_rdma_qpair_initialize(qpair); 4298 if (rc < 0) { 4299 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 4300 rqpair->poller = NULL; 4301 rqpair->srq = NULL; 4302 return -1; 4303 } 4304 4305 RB_INSERT(qpairs_tree, &poller->qpairs, rqpair); 4306 4307 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 4308 if (rc) { 4309 /* Try to reject, but we probably can't */ 4310 nvmf_rdma_qpair_reject_connection(rqpair); 4311 return -1; 4312 } 4313 4314 return 0; 4315 } 4316 4317 static int 4318 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 4319 struct spdk_nvmf_qpair *qpair) 4320 { 4321 struct spdk_nvmf_rdma_qpair *rqpair; 4322 4323 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4324 assert(group->transport->tgt != NULL); 4325 4326 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 4327 4328 if (!rqpair->destruct_channel) { 4329 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 4330 return 0; 4331 } 4332 4333 /* Sanity check that we get io_channel on the correct thread */ 4334 if (qpair->group) { 4335 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 4336 } 4337 4338 return 0; 4339 } 4340 4341 static int 4342 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 4343 { 4344 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 4345 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4346 struct spdk_nvmf_rdma_transport, transport); 4347 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4348 struct spdk_nvmf_rdma_qpair, qpair); 4349 4350 /* 4351 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 4352 * needs to be returned to the shared receive queue or the poll group will eventually be 4353 * starved of RECV structures. 4354 */ 4355 if (rqpair->srq && rdma_req->recv) { 4356 int rc; 4357 struct ibv_recv_wr *bad_recv_wr; 4358 4359 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 4360 rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 4361 if (rc) { 4362 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 4363 } 4364 } 4365 4366 _nvmf_rdma_request_free(rdma_req, rtransport); 4367 return 0; 4368 } 4369 4370 static int 4371 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 4372 { 4373 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4374 struct spdk_nvmf_rdma_transport, transport); 4375 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 4376 struct spdk_nvmf_rdma_request, req); 4377 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4378 struct spdk_nvmf_rdma_qpair, qpair); 4379 4380 if (spdk_unlikely(rqpair->ibv_in_error_state)) { 4381 /* The connection is dead. Move the request directly to the completed state. */ 4382 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4383 } else { 4384 /* The connection is alive, so process the request as normal */ 4385 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 4386 } 4387 4388 nvmf_rdma_request_process(rtransport, rdma_req); 4389 4390 return 0; 4391 } 4392 4393 static void 4394 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 4395 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 4396 { 4397 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4398 4399 rqpair->to_close = true; 4400 4401 /* This happens only when the qpair is disconnected before 4402 * it is added to the poll group. Since there is no poll group, 4403 * the RDMA qp has not been initialized yet and the RDMA CM 4404 * event has not yet been acknowledged, so we need to reject it. 4405 */ 4406 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 4407 nvmf_rdma_qpair_reject_connection(rqpair); 4408 nvmf_rdma_qpair_destroy(rqpair); 4409 return; 4410 } 4411 4412 if (rqpair->rdma_qp) { 4413 spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp); 4414 } 4415 4416 nvmf_rdma_destroy_drained_qpair(rqpair); 4417 4418 if (cb_fn) { 4419 cb_fn(cb_arg); 4420 } 4421 } 4422 4423 static struct spdk_nvmf_rdma_qpair * 4424 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 4425 { 4426 struct spdk_nvmf_rdma_qpair find; 4427 4428 find.qp_num = wc->qp_num; 4429 4430 return RB_FIND(qpairs_tree, &rpoller->qpairs, &find); 4431 } 4432 4433 #ifdef DEBUG 4434 static int 4435 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 4436 { 4437 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 4438 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 4439 } 4440 #endif 4441 4442 static void 4443 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 4444 int rc) 4445 { 4446 struct spdk_nvmf_rdma_recv *rdma_recv; 4447 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4448 4449 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 4450 while (bad_recv_wr != NULL) { 4451 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 4452 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4453 4454 rdma_recv->qpair->current_recv_depth++; 4455 bad_recv_wr = bad_recv_wr->next; 4456 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 4457 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair); 4458 } 4459 } 4460 4461 static void 4462 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 4463 { 4464 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 4465 while (bad_recv_wr != NULL) { 4466 bad_recv_wr = bad_recv_wr->next; 4467 rqpair->current_recv_depth++; 4468 } 4469 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4470 } 4471 4472 static void 4473 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 4474 struct spdk_nvmf_rdma_poller *rpoller) 4475 { 4476 struct spdk_nvmf_rdma_qpair *rqpair; 4477 struct ibv_recv_wr *bad_recv_wr; 4478 int rc; 4479 4480 if (rpoller->srq) { 4481 rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 4482 if (spdk_unlikely(rc)) { 4483 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 4484 } 4485 } else { 4486 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 4487 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 4488 rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 4489 if (spdk_unlikely(rc)) { 4490 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 4491 } 4492 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 4493 } 4494 } 4495 } 4496 4497 static void 4498 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 4499 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 4500 { 4501 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4502 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 4503 4504 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 4505 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 4506 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 4507 assert(rqpair->current_send_depth > 0); 4508 rqpair->current_send_depth--; 4509 switch (bad_rdma_wr->type) { 4510 case RDMA_WR_TYPE_DATA: 4511 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4512 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 4513 assert(rqpair->current_read_depth > 0); 4514 rqpair->current_read_depth--; 4515 } 4516 break; 4517 case RDMA_WR_TYPE_SEND: 4518 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4519 break; 4520 default: 4521 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 4522 prev_rdma_req = cur_rdma_req; 4523 continue; 4524 } 4525 4526 if (prev_rdma_req == cur_rdma_req) { 4527 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 4528 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 4529 continue; 4530 } 4531 4532 switch (cur_rdma_req->state) { 4533 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4534 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 4535 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, cur_rdma_req, state_link); 4536 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 4537 break; 4538 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 4539 case RDMA_REQUEST_STATE_COMPLETING: 4540 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4541 break; 4542 default: 4543 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 4544 cur_rdma_req->state, rqpair); 4545 continue; 4546 } 4547 4548 nvmf_rdma_request_process(rtransport, cur_rdma_req); 4549 prev_rdma_req = cur_rdma_req; 4550 } 4551 4552 if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) { 4553 /* Disconnect the connection. */ 4554 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4555 } 4556 4557 } 4558 4559 static void 4560 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 4561 struct spdk_nvmf_rdma_poller *rpoller) 4562 { 4563 struct spdk_nvmf_rdma_qpair *rqpair; 4564 struct ibv_send_wr *bad_wr = NULL; 4565 int rc; 4566 4567 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 4568 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 4569 rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 4570 4571 /* bad wr always points to the first wr that failed. */ 4572 if (spdk_unlikely(rc)) { 4573 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 4574 } 4575 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 4576 } 4577 } 4578 4579 static const char * 4580 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 4581 { 4582 switch (wr_type) { 4583 case RDMA_WR_TYPE_RECV: 4584 return "RECV"; 4585 case RDMA_WR_TYPE_SEND: 4586 return "SEND"; 4587 case RDMA_WR_TYPE_DATA: 4588 return "DATA"; 4589 default: 4590 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 4591 SPDK_UNREACHABLE(); 4592 } 4593 } 4594 4595 static inline void 4596 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 4597 { 4598 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 4599 4600 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 4601 /* If qpair is in ERR state, we will receive completions for all posted and not completed 4602 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 4603 SPDK_DEBUGLOG(rdma, 4604 "Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n", 4605 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id, 4606 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4607 } else { 4608 SPDK_ERRLOG("Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n", 4609 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id, 4610 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4611 } 4612 } 4613 4614 static int 4615 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 4616 struct spdk_nvmf_rdma_poller *rpoller) 4617 { 4618 struct ibv_wc wc[32]; 4619 struct spdk_nvmf_rdma_wr *rdma_wr; 4620 struct spdk_nvmf_rdma_request *rdma_req; 4621 struct spdk_nvmf_rdma_recv *rdma_recv; 4622 struct spdk_nvmf_rdma_qpair *rqpair, *tmp_rqpair; 4623 int reaped, i; 4624 int count = 0; 4625 int rc; 4626 bool error = false; 4627 uint64_t poll_tsc = spdk_get_ticks(); 4628 4629 if (spdk_unlikely(rpoller->need_destroy)) { 4630 /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not 4631 * be called because we cannot poll anything from cq. So we call that here to force 4632 * destroy the qpair after to_close turning true. 4633 */ 4634 RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) { 4635 nvmf_rdma_destroy_drained_qpair(rqpair); 4636 } 4637 return 0; 4638 } 4639 4640 /* Poll for completing operations. */ 4641 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 4642 if (spdk_unlikely(reaped < 0)) { 4643 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 4644 errno, spdk_strerror(errno)); 4645 return -1; 4646 } else if (reaped == 0) { 4647 rpoller->stat.idle_polls++; 4648 } 4649 4650 rpoller->stat.polls++; 4651 rpoller->stat.completions += reaped; 4652 4653 for (i = 0; i < reaped; i++) { 4654 4655 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 4656 4657 switch (rdma_wr->type) { 4658 case RDMA_WR_TYPE_SEND: 4659 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4660 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4661 4662 if (spdk_likely(!wc[i].status)) { 4663 count++; 4664 assert(wc[i].opcode == IBV_WC_SEND); 4665 assert(nvmf_rdma_req_is_completing(rdma_req)); 4666 } 4667 4668 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4669 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 4670 assert(rqpair->current_send_depth >= (uint32_t)rdma_req->num_outstanding_data_wr + 1); 4671 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 4672 rdma_req->num_outstanding_data_wr = 0; 4673 4674 nvmf_rdma_request_process(rtransport, rdma_req); 4675 break; 4676 case RDMA_WR_TYPE_RECV: 4677 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 4678 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4679 if (rpoller->srq != NULL) { 4680 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 4681 /* It is possible that there are still some completions for destroyed QP 4682 * associated with SRQ. We just ignore these late completions and re-post 4683 * receive WRs back to SRQ. 4684 */ 4685 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 4686 struct ibv_recv_wr *bad_wr; 4687 4688 rdma_recv->wr.next = NULL; 4689 spdk_rdma_provider_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 4690 rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 4691 if (rc) { 4692 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 4693 } 4694 continue; 4695 } 4696 } 4697 rqpair = rdma_recv->qpair; 4698 4699 assert(rqpair != NULL); 4700 if (spdk_likely(!wc[i].status)) { 4701 assert(wc[i].opcode == IBV_WC_RECV); 4702 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 4703 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4704 break; 4705 } 4706 } 4707 4708 rdma_recv->wr.next = NULL; 4709 rqpair->current_recv_depth++; 4710 rdma_recv->receive_tsc = poll_tsc; 4711 rpoller->stat.requests++; 4712 STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link); 4713 rqpair->qpair.queue_depth++; 4714 break; 4715 case RDMA_WR_TYPE_DATA: 4716 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4717 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4718 4719 assert(rdma_req->num_outstanding_data_wr > 0); 4720 4721 rqpair->current_send_depth--; 4722 rdma_req->num_outstanding_data_wr--; 4723 if (spdk_likely(!wc[i].status)) { 4724 assert(wc[i].opcode == IBV_WC_RDMA_READ); 4725 rqpair->current_read_depth--; 4726 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 4727 if (rdma_req->num_outstanding_data_wr == 0) { 4728 if (rdma_req->num_remaining_data_wr) { 4729 /* Only part of RDMA_READ operations was submitted, process the rest */ 4730 nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport); 4731 /* Prioritize partially handled request over others to avoid latency increase */ 4732 STAILQ_INSERT_HEAD(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 4733 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 4734 nvmf_rdma_request_process(rtransport, rdma_req); 4735 break; 4736 } 4737 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 4738 nvmf_rdma_request_process(rtransport, rdma_req); 4739 } 4740 } else { 4741 /* If the data transfer fails still force the queue into the error state, 4742 * if we were performing an RDMA_READ, we need to force the request into a 4743 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 4744 * case, we should wait for the SEND to complete. */ 4745 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 4746 rqpair->current_read_depth--; 4747 if (rdma_req->num_outstanding_data_wr == 0) { 4748 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4749 } 4750 } 4751 } 4752 break; 4753 default: 4754 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 4755 continue; 4756 } 4757 4758 /* Handle error conditions */ 4759 if (spdk_unlikely(wc[i].status)) { 4760 rqpair->ibv_in_error_state = true; 4761 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 4762 4763 error = true; 4764 4765 if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) { 4766 /* Disconnect the connection. */ 4767 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4768 } else { 4769 nvmf_rdma_destroy_drained_qpair(rqpair); 4770 } 4771 continue; 4772 } 4773 4774 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 4775 4776 if (spdk_unlikely(!spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 4777 nvmf_rdma_destroy_drained_qpair(rqpair); 4778 } 4779 } 4780 4781 if (spdk_unlikely(error == true)) { 4782 return -1; 4783 } 4784 4785 if (reaped == 0) { 4786 /* In some cases we may not receive any CQE but we still may have pending IO requests waiting for 4787 * a resource (e.g. a WR from the data_wr_pool). 4788 * We need to start processing of such requests if no CQE reaped */ 4789 nvmf_rdma_poller_process_pending_buf_queue(rtransport, rpoller); 4790 } 4791 4792 /* submit outstanding work requests. */ 4793 _poller_submit_recvs(rtransport, rpoller); 4794 _poller_submit_sends(rtransport, rpoller); 4795 4796 return count; 4797 } 4798 4799 static void 4800 _nvmf_rdma_remove_destroyed_device(void *c) 4801 { 4802 struct spdk_nvmf_rdma_transport *rtransport = c; 4803 struct spdk_nvmf_rdma_device *device, *device_tmp; 4804 int rc; 4805 4806 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 4807 if (device->ready_to_destroy) { 4808 destroy_ib_device(rtransport, device); 4809 } 4810 } 4811 4812 free_poll_fds(rtransport); 4813 rc = generate_poll_fds(rtransport); 4814 /* cannot handle fd allocation error here */ 4815 if (rc != 0) { 4816 SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n"); 4817 } 4818 } 4819 4820 static void 4821 _nvmf_rdma_remove_poller_in_group_cb(void *c) 4822 { 4823 struct poller_manage_ctx *ctx = c; 4824 struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport; 4825 struct spdk_nvmf_rdma_device *device = ctx->device; 4826 struct spdk_thread *thread = ctx->thread; 4827 4828 if (nvmf_rdma_all_pollers_management_done(c)) { 4829 /* destroy device when last poller is destroyed */ 4830 device->ready_to_destroy = true; 4831 spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport); 4832 } 4833 } 4834 4835 static void 4836 _nvmf_rdma_remove_poller_in_group(void *c) 4837 { 4838 struct poller_manage_ctx *ctx = c; 4839 4840 ctx->rpoller->need_destroy = true; 4841 ctx->rpoller->destroy_cb_ctx = ctx; 4842 ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb; 4843 4844 /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */ 4845 if (RB_EMPTY(&ctx->rpoller->qpairs)) { 4846 nvmf_rdma_poller_destroy(ctx->rpoller); 4847 } 4848 } 4849 4850 static int 4851 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 4852 { 4853 struct spdk_nvmf_rdma_transport *rtransport; 4854 struct spdk_nvmf_rdma_poll_group *rgroup; 4855 struct spdk_nvmf_rdma_poller *rpoller, *tmp; 4856 int count = 0, rc, rc2 = 0; 4857 4858 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 4859 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4860 4861 TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) { 4862 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4863 if (spdk_unlikely(rc < 0)) { 4864 if (rc2 == 0) { 4865 rc2 = rc; 4866 } 4867 continue; 4868 } 4869 count += rc; 4870 } 4871 4872 return rc2 ? rc2 : count; 4873 } 4874 4875 static int 4876 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4877 struct spdk_nvme_transport_id *trid, 4878 bool peer) 4879 { 4880 struct sockaddr *saddr; 4881 uint16_t port; 4882 4883 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4884 4885 if (peer) { 4886 saddr = rdma_get_peer_addr(id); 4887 } else { 4888 saddr = rdma_get_local_addr(id); 4889 } 4890 switch (saddr->sa_family) { 4891 case AF_INET: { 4892 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4893 4894 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4895 inet_ntop(AF_INET, &saddr_in->sin_addr, 4896 trid->traddr, sizeof(trid->traddr)); 4897 if (peer) { 4898 port = ntohs(rdma_get_dst_port(id)); 4899 } else { 4900 port = ntohs(rdma_get_src_port(id)); 4901 } 4902 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4903 break; 4904 } 4905 case AF_INET6: { 4906 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4907 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4908 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4909 trid->traddr, sizeof(trid->traddr)); 4910 if (peer) { 4911 port = ntohs(rdma_get_dst_port(id)); 4912 } else { 4913 port = ntohs(rdma_get_src_port(id)); 4914 } 4915 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4916 break; 4917 } 4918 default: 4919 return -1; 4920 4921 } 4922 4923 return 0; 4924 } 4925 4926 static int 4927 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4928 struct spdk_nvme_transport_id *trid) 4929 { 4930 struct spdk_nvmf_rdma_qpair *rqpair; 4931 4932 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4933 4934 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4935 } 4936 4937 static int 4938 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4939 struct spdk_nvme_transport_id *trid) 4940 { 4941 struct spdk_nvmf_rdma_qpair *rqpair; 4942 4943 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4944 4945 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4946 } 4947 4948 static int 4949 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4950 struct spdk_nvme_transport_id *trid) 4951 { 4952 struct spdk_nvmf_rdma_qpair *rqpair; 4953 4954 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4955 4956 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4957 } 4958 4959 void 4960 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4961 { 4962 g_nvmf_hooks = *hooks; 4963 } 4964 4965 static void 4966 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4967 struct spdk_nvmf_rdma_request *rdma_req_to_abort, 4968 struct spdk_nvmf_rdma_qpair *rqpair) 4969 { 4970 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4971 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4972 4973 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, state_link); 4974 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 4975 4976 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4977 } 4978 4979 static int 4980 _nvmf_rdma_qpair_abort_request(void *ctx) 4981 { 4982 struct spdk_nvmf_request *req = ctx; 4983 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4984 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4985 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4986 struct spdk_nvmf_rdma_qpair, qpair); 4987 int rc; 4988 4989 spdk_poller_unregister(&req->poller); 4990 4991 switch (rdma_req_to_abort->state) { 4992 case RDMA_REQUEST_STATE_EXECUTING: 4993 rc = nvmf_ctrlr_abort_request(req); 4994 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4995 return SPDK_POLLER_BUSY; 4996 } 4997 break; 4998 4999 case RDMA_REQUEST_STATE_NEED_BUFFER: 5000 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 5001 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 5002 5003 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5004 break; 5005 5006 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 5007 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 5008 spdk_nvmf_rdma_request, state_link); 5009 5010 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5011 break; 5012 5013 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 5014 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 5015 spdk_nvmf_rdma_request, state_link); 5016 5017 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5018 break; 5019 5020 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 5021 /* Remove req from the list here to re-use common function */ 5022 STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, 5023 spdk_nvmf_rdma_request, state_link); 5024 5025 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5026 break; 5027 5028 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 5029 if (spdk_get_ticks() < req->timeout_tsc) { 5030 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 5031 return SPDK_POLLER_BUSY; 5032 } 5033 break; 5034 5035 default: 5036 break; 5037 } 5038 5039 spdk_nvmf_request_complete(req); 5040 return SPDK_POLLER_BUSY; 5041 } 5042 5043 static void 5044 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 5045 struct spdk_nvmf_request *req) 5046 { 5047 struct spdk_nvmf_rdma_qpair *rqpair; 5048 struct spdk_nvmf_rdma_transport *rtransport; 5049 struct spdk_nvmf_transport *transport; 5050 uint16_t cid; 5051 uint32_t i, max_req_count; 5052 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 5053 5054 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 5055 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 5056 transport = &rtransport->transport; 5057 5058 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 5059 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 5060 5061 for (i = 0; i < max_req_count; i++) { 5062 rdma_req = &rqpair->resources->reqs[i]; 5063 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 5064 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 5065 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 5066 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 5067 rdma_req->req.qpair == qpair) { 5068 rdma_req_to_abort = rdma_req; 5069 break; 5070 } 5071 } 5072 5073 if (rdma_req_to_abort == NULL) { 5074 spdk_nvmf_request_complete(req); 5075 return; 5076 } 5077 5078 req->req_to_abort = &rdma_req_to_abort->req; 5079 req->timeout_tsc = spdk_get_ticks() + 5080 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 5081 req->poller = NULL; 5082 5083 _nvmf_rdma_qpair_abort_request(req); 5084 } 5085 5086 static void 5087 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 5088 struct spdk_json_write_ctx *w) 5089 { 5090 struct spdk_nvmf_rdma_poll_group *rgroup; 5091 struct spdk_nvmf_rdma_poller *rpoller; 5092 5093 assert(w != NULL); 5094 5095 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 5096 5097 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 5098 5099 spdk_json_write_named_array_begin(w, "devices"); 5100 5101 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 5102 spdk_json_write_object_begin(w); 5103 spdk_json_write_named_string(w, "name", 5104 ibv_get_device_name(rpoller->device->context->device)); 5105 spdk_json_write_named_uint64(w, "polls", 5106 rpoller->stat.polls); 5107 spdk_json_write_named_uint64(w, "idle_polls", 5108 rpoller->stat.idle_polls); 5109 spdk_json_write_named_uint64(w, "completions", 5110 rpoller->stat.completions); 5111 spdk_json_write_named_uint64(w, "requests", 5112 rpoller->stat.requests); 5113 spdk_json_write_named_uint64(w, "request_latency", 5114 rpoller->stat.request_latency); 5115 spdk_json_write_named_uint64(w, "pending_free_request", 5116 rpoller->stat.pending_free_request); 5117 spdk_json_write_named_uint64(w, "pending_rdma_read", 5118 rpoller->stat.pending_rdma_read); 5119 spdk_json_write_named_uint64(w, "pending_rdma_write", 5120 rpoller->stat.pending_rdma_write); 5121 spdk_json_write_named_uint64(w, "pending_rdma_send", 5122 rpoller->stat.pending_rdma_send); 5123 spdk_json_write_named_uint64(w, "total_send_wrs", 5124 rpoller->stat.qp_stats.send.num_submitted_wrs); 5125 spdk_json_write_named_uint64(w, "send_doorbell_updates", 5126 rpoller->stat.qp_stats.send.doorbell_updates); 5127 spdk_json_write_named_uint64(w, "total_recv_wrs", 5128 rpoller->stat.qp_stats.recv.num_submitted_wrs); 5129 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 5130 rpoller->stat.qp_stats.recv.doorbell_updates); 5131 spdk_json_write_object_end(w); 5132 } 5133 5134 spdk_json_write_array_end(w); 5135 } 5136 5137 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 5138 .name = "RDMA", 5139 .type = SPDK_NVME_TRANSPORT_RDMA, 5140 .opts_init = nvmf_rdma_opts_init, 5141 .create = nvmf_rdma_create, 5142 .dump_opts = nvmf_rdma_dump_opts, 5143 .destroy = nvmf_rdma_destroy, 5144 5145 .listen = nvmf_rdma_listen, 5146 .stop_listen = nvmf_rdma_stop_listen, 5147 .cdata_init = nvmf_rdma_cdata_init, 5148 5149 .listener_discover = nvmf_rdma_discover, 5150 5151 .poll_group_create = nvmf_rdma_poll_group_create, 5152 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 5153 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 5154 .poll_group_add = nvmf_rdma_poll_group_add, 5155 .poll_group_remove = nvmf_rdma_poll_group_remove, 5156 .poll_group_poll = nvmf_rdma_poll_group_poll, 5157 5158 .req_free = nvmf_rdma_request_free, 5159 .req_complete = nvmf_rdma_request_complete, 5160 5161 .qpair_fini = nvmf_rdma_close_qpair, 5162 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 5163 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 5164 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 5165 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 5166 5167 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 5168 }; 5169 5170 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 5171 SPDK_LOG_REGISTER_COMPONENT(rdma) 5172