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 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_RDMA_IO, 1); 188 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_RDMA_IO, 0); 189 } 190 191 enum spdk_nvmf_rdma_wr_type { 192 RDMA_WR_TYPE_RECV, 193 RDMA_WR_TYPE_SEND, 194 RDMA_WR_TYPE_DATA, 195 }; 196 197 struct spdk_nvmf_rdma_wr { 198 /* Uses enum spdk_nvmf_rdma_wr_type */ 199 uint8_t type; 200 }; 201 202 /* This structure holds commands as they are received off the wire. 203 * It must be dynamically paired with a full request object 204 * (spdk_nvmf_rdma_request) to service a request. It is separate 205 * from the request because RDMA does not appear to order 206 * completions, so occasionally we'll get a new incoming 207 * command when there aren't any free request objects. 208 */ 209 struct spdk_nvmf_rdma_recv { 210 struct ibv_recv_wr wr; 211 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 212 213 struct spdk_nvmf_rdma_qpair *qpair; 214 215 /* In-capsule data buffer */ 216 uint8_t *buf; 217 218 struct spdk_nvmf_rdma_wr rdma_wr; 219 uint64_t receive_tsc; 220 221 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 222 }; 223 224 struct spdk_nvmf_rdma_request_data { 225 struct ibv_send_wr wr; 226 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 227 }; 228 229 struct spdk_nvmf_rdma_request { 230 struct spdk_nvmf_request req; 231 232 bool fused_failed; 233 234 struct spdk_nvmf_rdma_wr data_wr; 235 struct spdk_nvmf_rdma_wr rsp_wr; 236 237 /* Uses enum spdk_nvmf_rdma_request_state */ 238 uint8_t state; 239 240 /* Data offset in req.iov */ 241 uint32_t offset; 242 243 struct spdk_nvmf_rdma_recv *recv; 244 245 struct { 246 struct ibv_send_wr wr; 247 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 248 } rsp; 249 250 uint16_t iovpos; 251 uint16_t num_outstanding_data_wr; 252 /* Used to split Write IO with multi SGL payload */ 253 uint16_t num_remaining_data_wr; 254 uint64_t receive_tsc; 255 struct spdk_nvmf_rdma_request *fused_pair; 256 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 257 struct ibv_send_wr *remaining_tranfer_in_wrs; 258 struct ibv_send_wr *transfer_wr; 259 struct spdk_nvmf_rdma_request_data data; 260 }; 261 262 struct spdk_nvmf_rdma_resource_opts { 263 struct spdk_nvmf_rdma_qpair *qpair; 264 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 265 void *qp; 266 struct spdk_rdma_utils_mem_map *map; 267 uint32_t max_queue_depth; 268 uint32_t in_capsule_data_size; 269 bool shared; 270 }; 271 272 struct spdk_nvmf_rdma_resources { 273 /* Array of size "max_queue_depth" containing RDMA requests. */ 274 struct spdk_nvmf_rdma_request *reqs; 275 276 /* Array of size "max_queue_depth" containing RDMA recvs. */ 277 struct spdk_nvmf_rdma_recv *recvs; 278 279 /* Array of size "max_queue_depth" containing 64 byte capsules 280 * used for receive. 281 */ 282 union nvmf_h2c_msg *cmds; 283 284 /* Array of size "max_queue_depth" containing 16 byte completions 285 * to be sent back to the user. 286 */ 287 union nvmf_c2h_msg *cpls; 288 289 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 290 * buffers to be used for in capsule data. 291 */ 292 void *bufs; 293 294 /* Receives that are waiting for a request object */ 295 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 296 297 /* Queue to track free requests */ 298 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 299 }; 300 301 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 302 303 typedef void (*spdk_poller_destroy_cb)(void *ctx); 304 305 struct spdk_nvmf_rdma_ibv_event_ctx { 306 struct spdk_nvmf_rdma_qpair *rqpair; 307 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 308 /* Link to other ibv events associated with this qpair */ 309 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 310 }; 311 312 struct spdk_nvmf_rdma_qpair { 313 struct spdk_nvmf_qpair qpair; 314 315 struct spdk_nvmf_rdma_device *device; 316 struct spdk_nvmf_rdma_poller *poller; 317 318 struct spdk_rdma_provider_qp *rdma_qp; 319 struct rdma_cm_id *cm_id; 320 struct spdk_rdma_provider_srq *srq; 321 struct rdma_cm_id *listen_id; 322 323 /* Cache the QP number to improve QP search by RB tree. */ 324 uint32_t qp_num; 325 326 /* The maximum number of I/O outstanding on this connection at one time */ 327 uint16_t max_queue_depth; 328 329 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 330 uint16_t max_read_depth; 331 332 /* The maximum number of RDMA SEND operations at one time */ 333 uint32_t max_send_depth; 334 335 /* The current number of outstanding WRs from this qpair's 336 * recv queue. Should not exceed device->attr.max_queue_depth. 337 */ 338 uint16_t current_recv_depth; 339 340 /* The current number of active RDMA READ operations */ 341 uint16_t current_read_depth; 342 343 /* The current number of posted WRs from this qpair's 344 * send queue. Should not exceed max_send_depth. 345 */ 346 uint32_t current_send_depth; 347 348 /* The maximum number of SGEs per WR on the send queue */ 349 uint32_t max_send_sge; 350 351 /* The maximum number of SGEs per WR on the recv queue */ 352 uint32_t max_recv_sge; 353 354 struct spdk_nvmf_rdma_resources *resources; 355 356 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 357 358 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 359 360 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_send_queue; 361 362 /* Number of requests not in the free state */ 363 uint32_t qd; 364 365 bool ibv_in_error_state; 366 367 RB_ENTRY(spdk_nvmf_rdma_qpair) node; 368 369 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 370 371 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 372 373 /* Points to the a request that has fuse bits set to 374 * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting 375 * for the request that has SPDK_NVME_CMD_FUSE_SECOND. 376 */ 377 struct spdk_nvmf_rdma_request *fused_first; 378 379 /* 380 * io_channel which is used to destroy qpair when it is removed from poll group 381 */ 382 struct spdk_io_channel *destruct_channel; 383 384 /* List of ibv async events */ 385 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 386 387 /* Lets us know that we have received the last_wqe event. */ 388 bool last_wqe_reached; 389 390 /* Indicate that nvmf_rdma_close_qpair is called */ 391 bool to_close; 392 }; 393 394 struct spdk_nvmf_rdma_poller_stat { 395 uint64_t completions; 396 uint64_t polls; 397 uint64_t idle_polls; 398 uint64_t requests; 399 uint64_t request_latency; 400 uint64_t pending_free_request; 401 uint64_t pending_rdma_read; 402 uint64_t pending_rdma_write; 403 uint64_t pending_rdma_send; 404 struct spdk_rdma_provider_qp_stats qp_stats; 405 }; 406 407 struct spdk_nvmf_rdma_poller { 408 struct spdk_nvmf_rdma_device *device; 409 struct spdk_nvmf_rdma_poll_group *group; 410 411 int num_cqe; 412 int required_num_wr; 413 struct ibv_cq *cq; 414 415 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 416 uint16_t max_srq_depth; 417 bool need_destroy; 418 419 /* Shared receive queue */ 420 struct spdk_rdma_provider_srq *srq; 421 422 struct spdk_nvmf_rdma_resources *resources; 423 struct spdk_nvmf_rdma_poller_stat stat; 424 425 spdk_poller_destroy_cb destroy_cb; 426 void *destroy_cb_ctx; 427 428 RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs; 429 430 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 431 432 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 433 434 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 435 }; 436 437 struct spdk_nvmf_rdma_poll_group_stat { 438 uint64_t pending_data_buffer; 439 }; 440 441 struct spdk_nvmf_rdma_poll_group { 442 struct spdk_nvmf_transport_poll_group group; 443 struct spdk_nvmf_rdma_poll_group_stat stat; 444 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 445 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 446 }; 447 448 struct spdk_nvmf_rdma_conn_sched { 449 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 450 struct spdk_nvmf_rdma_poll_group *next_io_pg; 451 }; 452 453 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 454 struct spdk_nvmf_rdma_device { 455 struct ibv_device_attr attr; 456 struct ibv_context *context; 457 458 struct spdk_rdma_utils_mem_map *map; 459 struct ibv_pd *pd; 460 461 int num_srq; 462 bool need_destroy; 463 bool ready_to_destroy; 464 bool is_ready; 465 466 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 467 }; 468 469 struct spdk_nvmf_rdma_port { 470 const struct spdk_nvme_transport_id *trid; 471 struct rdma_cm_id *id; 472 struct spdk_nvmf_rdma_device *device; 473 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 474 }; 475 476 struct rdma_transport_opts { 477 int num_cqe; 478 uint32_t max_srq_depth; 479 bool no_srq; 480 bool no_wr_batching; 481 int acceptor_backlog; 482 }; 483 484 struct spdk_nvmf_rdma_transport { 485 struct spdk_nvmf_transport transport; 486 struct rdma_transport_opts rdma_opts; 487 488 struct spdk_nvmf_rdma_conn_sched conn_sched; 489 490 struct rdma_event_channel *event_channel; 491 492 struct spdk_mempool *data_wr_pool; 493 494 struct spdk_poller *accept_poller; 495 496 /* fields used to poll RDMA/IB events */ 497 nfds_t npoll_fds; 498 struct pollfd *poll_fds; 499 500 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 501 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 502 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 503 504 /* ports that are removed unexpectedly and need retry listen */ 505 TAILQ_HEAD(, spdk_nvmf_rdma_port) retry_ports; 506 }; 507 508 struct poller_manage_ctx { 509 struct spdk_nvmf_rdma_transport *rtransport; 510 struct spdk_nvmf_rdma_poll_group *rgroup; 511 struct spdk_nvmf_rdma_poller *rpoller; 512 struct spdk_nvmf_rdma_device *device; 513 514 struct spdk_thread *thread; 515 volatile int *inflight_op_counter; 516 }; 517 518 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 519 { 520 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe), 521 spdk_json_decode_int32, true 522 }, 523 { 524 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 525 spdk_json_decode_uint32, true 526 }, 527 { 528 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 529 spdk_json_decode_bool, true 530 }, 531 { 532 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 533 spdk_json_decode_bool, true 534 }, 535 { 536 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 537 spdk_json_decode_int32, true 538 }, 539 }; 540 541 static int 542 nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2) 543 { 544 return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num; 545 } 546 547 RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare); 548 549 static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 550 struct spdk_nvmf_rdma_request *rdma_req); 551 552 static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 553 struct spdk_nvmf_rdma_poller *rpoller); 554 555 static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 556 struct spdk_nvmf_rdma_poller *rpoller); 557 558 static void _nvmf_rdma_remove_destroyed_device(void *c); 559 560 static inline enum spdk_nvme_media_error_status_code 561 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 562 enum spdk_nvme_media_error_status_code result; 563 switch (err_type) 564 { 565 case SPDK_DIF_REFTAG_ERROR: 566 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 567 break; 568 case SPDK_DIF_APPTAG_ERROR: 569 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 570 break; 571 case SPDK_DIF_GUARD_ERROR: 572 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 573 break; 574 default: 575 SPDK_UNREACHABLE(); 576 } 577 578 return result; 579 } 580 581 /* 582 * Return data_wrs to pool starting from \b data_wr 583 * Request's own response and data WR are excluded 584 */ 585 static void 586 _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 587 struct ibv_send_wr *data_wr, 588 struct spdk_mempool *pool) 589 { 590 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 591 struct spdk_nvmf_rdma_request_data *nvmf_data; 592 struct ibv_send_wr *next_send_wr; 593 uint64_t req_wrid = (uint64_t)&rdma_req->data_wr; 594 uint32_t num_wrs = 0; 595 596 while (data_wr && data_wr->wr_id == req_wrid) { 597 nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr); 598 memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge); 599 data_wr->num_sge = 0; 600 next_send_wr = data_wr->next; 601 if (data_wr != &rdma_req->data.wr) { 602 data_wr->next = NULL; 603 assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES); 604 work_requests[num_wrs] = nvmf_data; 605 num_wrs++; 606 } 607 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr; 608 } 609 610 if (num_wrs) { 611 spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs); 612 } 613 } 614 615 static void 616 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 617 struct spdk_nvmf_rdma_transport *rtransport) 618 { 619 rdma_req->num_outstanding_data_wr = 0; 620 621 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 622 623 if (rdma_req->remaining_tranfer_in_wrs) { 624 _nvmf_rdma_request_free_data(rdma_req, rdma_req->remaining_tranfer_in_wrs, 625 rtransport->data_wr_pool); 626 rdma_req->remaining_tranfer_in_wrs = NULL; 627 } 628 629 rdma_req->data.wr.next = NULL; 630 rdma_req->rsp.wr.next = NULL; 631 } 632 633 static void 634 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 635 { 636 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 637 if (req->req.cmd) { 638 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 639 } 640 if (req->recv) { 641 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 642 } 643 } 644 645 static void 646 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 647 { 648 int i; 649 650 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 651 for (i = 0; i < rqpair->max_queue_depth; i++) { 652 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 653 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 654 } 655 } 656 } 657 658 static void 659 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 660 { 661 spdk_free(resources->cmds); 662 spdk_free(resources->cpls); 663 spdk_free(resources->bufs); 664 spdk_free(resources->reqs); 665 spdk_free(resources->recvs); 666 free(resources); 667 } 668 669 670 static struct spdk_nvmf_rdma_resources * 671 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 672 { 673 struct spdk_nvmf_rdma_resources *resources; 674 struct spdk_nvmf_rdma_request *rdma_req; 675 struct spdk_nvmf_rdma_recv *rdma_recv; 676 struct spdk_rdma_provider_qp *qp = NULL; 677 struct spdk_rdma_provider_srq *srq = NULL; 678 struct ibv_recv_wr *bad_wr = NULL; 679 struct spdk_rdma_utils_memory_translation translation; 680 uint32_t i; 681 int rc = 0; 682 683 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 684 if (!resources) { 685 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 686 return NULL; 687 } 688 689 resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs), 690 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 691 resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs), 692 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 693 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 694 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 695 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 696 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 697 698 if (opts->in_capsule_data_size > 0) { 699 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 700 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 701 SPDK_MALLOC_DMA); 702 } 703 704 if (!resources->reqs || !resources->recvs || !resources->cmds || 705 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 706 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 707 goto cleanup; 708 } 709 710 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n", 711 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds)); 712 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n", 713 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls)); 714 if (resources->bufs) { 715 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n", 716 resources->bufs, opts->max_queue_depth * 717 opts->in_capsule_data_size); 718 } 719 720 /* Initialize queues */ 721 STAILQ_INIT(&resources->incoming_queue); 722 STAILQ_INIT(&resources->free_queue); 723 724 if (opts->shared) { 725 srq = (struct spdk_rdma_provider_srq *)opts->qp; 726 } else { 727 qp = (struct spdk_rdma_provider_qp *)opts->qp; 728 } 729 730 for (i = 0; i < opts->max_queue_depth; i++) { 731 rdma_recv = &resources->recvs[i]; 732 rdma_recv->qpair = opts->qpair; 733 734 /* Set up memory to receive commands */ 735 if (resources->bufs) { 736 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 737 opts->in_capsule_data_size)); 738 } 739 740 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 741 742 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 743 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 744 rc = spdk_rdma_utils_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]), 745 &translation); 746 if (rc) { 747 goto cleanup; 748 } 749 rdma_recv->sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 750 rdma_recv->wr.num_sge = 1; 751 752 if (rdma_recv->buf) { 753 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 754 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 755 rc = spdk_rdma_utils_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size, 756 &translation); 757 if (rc) { 758 goto cleanup; 759 } 760 rdma_recv->sgl[1].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 761 rdma_recv->wr.num_sge++; 762 } 763 764 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 765 rdma_recv->wr.sg_list = rdma_recv->sgl; 766 if (srq) { 767 spdk_rdma_provider_srq_queue_recv_wrs(srq, &rdma_recv->wr); 768 } else { 769 spdk_rdma_provider_qp_queue_recv_wrs(qp, &rdma_recv->wr); 770 } 771 } 772 773 for (i = 0; i < opts->max_queue_depth; i++) { 774 rdma_req = &resources->reqs[i]; 775 776 if (opts->qpair != NULL) { 777 rdma_req->req.qpair = &opts->qpair->qpair; 778 } else { 779 rdma_req->req.qpair = NULL; 780 } 781 rdma_req->req.cmd = NULL; 782 rdma_req->req.iovcnt = 0; 783 rdma_req->req.stripped_data = NULL; 784 785 /* Set up memory to send responses */ 786 rdma_req->req.rsp = &resources->cpls[i]; 787 788 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 789 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 790 rc = spdk_rdma_utils_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]), 791 &translation); 792 if (rc) { 793 goto cleanup; 794 } 795 rdma_req->rsp.sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation); 796 797 rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND; 798 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr; 799 rdma_req->rsp.wr.next = NULL; 800 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 801 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 802 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 803 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 804 805 /* Set up memory for data buffers */ 806 rdma_req->data_wr.type = RDMA_WR_TYPE_DATA; 807 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr; 808 rdma_req->data.wr.next = NULL; 809 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 810 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 811 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 812 813 /* Initialize request state to FREE */ 814 rdma_req->state = RDMA_REQUEST_STATE_FREE; 815 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 816 } 817 818 if (srq) { 819 rc = spdk_rdma_provider_srq_flush_recv_wrs(srq, &bad_wr); 820 } else { 821 rc = spdk_rdma_provider_qp_flush_recv_wrs(qp, &bad_wr); 822 } 823 824 if (rc) { 825 goto cleanup; 826 } 827 828 return resources; 829 830 cleanup: 831 nvmf_rdma_resources_destroy(resources); 832 return NULL; 833 } 834 835 static void 836 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 837 { 838 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 839 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 840 ctx->rqpair = NULL; 841 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 842 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 843 } 844 } 845 846 static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller); 847 848 static void 849 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 850 { 851 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 852 struct ibv_recv_wr *bad_recv_wr = NULL; 853 int rc; 854 855 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair); 856 857 if (rqpair->qd != 0) { 858 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 859 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 860 struct spdk_nvmf_rdma_transport, transport); 861 struct spdk_nvmf_rdma_request *req; 862 uint32_t i, max_req_count = 0; 863 864 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 865 866 if (rqpair->srq == NULL) { 867 nvmf_rdma_dump_qpair_contents(rqpair); 868 max_req_count = rqpair->max_queue_depth; 869 } else if (rqpair->poller && rqpair->resources) { 870 max_req_count = rqpair->poller->max_srq_depth; 871 } 872 873 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 874 for (i = 0; i < max_req_count; i++) { 875 req = &rqpair->resources->reqs[i]; 876 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 877 /* nvmf_rdma_request_process checks qpair ibv and internal state 878 * and completes a request */ 879 nvmf_rdma_request_process(rtransport, req); 880 } 881 } 882 assert(rqpair->qd == 0); 883 } 884 885 if (rqpair->poller) { 886 RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair); 887 888 if (rqpair->srq != NULL && rqpair->resources != NULL) { 889 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 890 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 891 if (rqpair == rdma_recv->qpair) { 892 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 893 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr); 894 rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 895 if (rc) { 896 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 897 } 898 } 899 } 900 } 901 } 902 903 if (rqpair->cm_id) { 904 if (rqpair->rdma_qp != NULL) { 905 spdk_rdma_provider_qp_destroy(rqpair->rdma_qp); 906 rqpair->rdma_qp = NULL; 907 } 908 909 if (rqpair->poller != NULL && rqpair->srq == NULL) { 910 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 911 } 912 } 913 914 if (rqpair->srq == NULL && rqpair->resources != NULL) { 915 nvmf_rdma_resources_destroy(rqpair->resources); 916 } 917 918 nvmf_rdma_qpair_clean_ibv_events(rqpair); 919 920 if (rqpair->destruct_channel) { 921 spdk_put_io_channel(rqpair->destruct_channel); 922 rqpair->destruct_channel = NULL; 923 } 924 925 if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) { 926 nvmf_rdma_poller_destroy(rqpair->poller); 927 } 928 929 /* destroy cm_id last so cma device will not be freed before we destroy the cq. */ 930 if (rqpair->cm_id) { 931 rdma_destroy_id(rqpair->cm_id); 932 } 933 934 free(rqpair); 935 } 936 937 static int 938 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 939 { 940 struct spdk_nvmf_rdma_poller *rpoller; 941 int rc, num_cqe, required_num_wr; 942 943 /* Enlarge CQ size dynamically */ 944 rpoller = rqpair->poller; 945 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 946 num_cqe = rpoller->num_cqe; 947 if (num_cqe < required_num_wr) { 948 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 949 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 950 } 951 952 if (rpoller->num_cqe != num_cqe) { 953 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) { 954 SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n" 955 "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe); 956 return -1; 957 } 958 if (required_num_wr > device->attr.max_cqe) { 959 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 960 required_num_wr, device->attr.max_cqe); 961 return -1; 962 } 963 964 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 965 rc = ibv_resize_cq(rpoller->cq, num_cqe); 966 if (rc) { 967 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 968 return -1; 969 } 970 971 rpoller->num_cqe = num_cqe; 972 } 973 974 rpoller->required_num_wr = required_num_wr; 975 return 0; 976 } 977 978 static int 979 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 980 { 981 struct spdk_nvmf_rdma_qpair *rqpair; 982 struct spdk_nvmf_rdma_transport *rtransport; 983 struct spdk_nvmf_transport *transport; 984 struct spdk_nvmf_rdma_resource_opts opts; 985 struct spdk_nvmf_rdma_device *device; 986 struct spdk_rdma_provider_qp_init_attr qp_init_attr = {}; 987 988 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 989 device = rqpair->device; 990 991 qp_init_attr.qp_context = rqpair; 992 qp_init_attr.pd = device->pd; 993 qp_init_attr.send_cq = rqpair->poller->cq; 994 qp_init_attr.recv_cq = rqpair->poller->cq; 995 996 if (rqpair->srq) { 997 qp_init_attr.srq = rqpair->srq->srq; 998 } else { 999 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 1000 } 1001 1002 /* SEND, READ, and WRITE operations */ 1003 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 1004 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 1005 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1006 qp_init_attr.stats = &rqpair->poller->stat.qp_stats; 1007 1008 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1009 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1010 goto error; 1011 } 1012 1013 rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &qp_init_attr); 1014 if (!rqpair->rdma_qp) { 1015 goto error; 1016 } 1017 1018 rqpair->qp_num = rqpair->rdma_qp->qp->qp_num; 1019 1020 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1021 qp_init_attr.cap.max_send_wr); 1022 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1023 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1024 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair); 1025 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1026 1027 if (rqpair->poller->srq == NULL) { 1028 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1029 transport = &rtransport->transport; 1030 1031 opts.qp = rqpair->rdma_qp; 1032 opts.map = device->map; 1033 opts.qpair = rqpair; 1034 opts.shared = false; 1035 opts.max_queue_depth = rqpair->max_queue_depth; 1036 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1037 1038 rqpair->resources = nvmf_rdma_resources_create(&opts); 1039 1040 if (!rqpair->resources) { 1041 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1042 rdma_destroy_qp(rqpair->cm_id); 1043 goto error; 1044 } 1045 } else { 1046 rqpair->resources = rqpair->poller->resources; 1047 } 1048 1049 rqpair->current_recv_depth = 0; 1050 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1051 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1052 STAILQ_INIT(&rqpair->pending_rdma_send_queue); 1053 rqpair->qpair.queue_depth = 0; 1054 1055 return 0; 1056 1057 error: 1058 rdma_destroy_id(rqpair->cm_id); 1059 rqpair->cm_id = NULL; 1060 return -1; 1061 } 1062 1063 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1064 /* This function accepts either a single wr or the first wr in a linked list. */ 1065 static void 1066 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1067 { 1068 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1069 struct spdk_nvmf_rdma_transport, transport); 1070 1071 if (rqpair->srq != NULL) { 1072 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, first); 1073 } else { 1074 if (spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, first)) { 1075 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1076 } 1077 } 1078 1079 if (rtransport->rdma_opts.no_wr_batching) { 1080 _poller_submit_recvs(rtransport, rqpair->poller); 1081 } 1082 } 1083 1084 static int 1085 request_transfer_in(struct spdk_nvmf_request *req) 1086 { 1087 struct spdk_nvmf_rdma_request *rdma_req; 1088 struct spdk_nvmf_qpair *qpair; 1089 struct spdk_nvmf_rdma_qpair *rqpair; 1090 struct spdk_nvmf_rdma_transport *rtransport; 1091 1092 qpair = req->qpair; 1093 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1094 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1095 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1096 struct spdk_nvmf_rdma_transport, transport); 1097 1098 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1099 assert(rdma_req != NULL); 1100 1101 if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) { 1102 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1103 } 1104 if (rtransport->rdma_opts.no_wr_batching) { 1105 _poller_submit_sends(rtransport, rqpair->poller); 1106 } 1107 1108 assert(rqpair->current_read_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_read_depth); 1109 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1110 assert(rqpair->current_send_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_send_depth); 1111 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1112 return 0; 1113 } 1114 1115 static inline void 1116 nvmf_rdma_request_reset_transfer_in(struct spdk_nvmf_rdma_request *rdma_req, 1117 struct spdk_nvmf_rdma_transport *rtransport) 1118 { 1119 /* Put completed WRs back to pool and move transfer_wr pointer */ 1120 _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool); 1121 rdma_req->transfer_wr = rdma_req->remaining_tranfer_in_wrs; 1122 rdma_req->remaining_tranfer_in_wrs = NULL; 1123 rdma_req->num_outstanding_data_wr = rdma_req->num_remaining_data_wr; 1124 rdma_req->num_remaining_data_wr = 0; 1125 } 1126 1127 static inline int 1128 request_prepare_transfer_in_part(struct spdk_nvmf_request *req, uint32_t num_reads_available) 1129 { 1130 struct spdk_nvmf_rdma_request *rdma_req; 1131 struct ibv_send_wr *wr; 1132 uint32_t i; 1133 1134 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1135 1136 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1137 assert(rdma_req != NULL); 1138 assert(num_reads_available > 0); 1139 assert(rdma_req->num_outstanding_data_wr > num_reads_available); 1140 wr = rdma_req->transfer_wr; 1141 1142 for (i = 0; i < num_reads_available - 1; i++) { 1143 wr = wr->next; 1144 } 1145 1146 rdma_req->remaining_tranfer_in_wrs = wr->next; 1147 rdma_req->num_remaining_data_wr = rdma_req->num_outstanding_data_wr - num_reads_available; 1148 rdma_req->num_outstanding_data_wr = num_reads_available; 1149 /* Break chain of WRs to send only part. Once this portion completes, we continue sending RDMA_READs */ 1150 wr->next = NULL; 1151 1152 return 0; 1153 } 1154 1155 static int 1156 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1157 { 1158 int num_outstanding_data_wr = 0; 1159 struct spdk_nvmf_rdma_request *rdma_req; 1160 struct spdk_nvmf_qpair *qpair; 1161 struct spdk_nvmf_rdma_qpair *rqpair; 1162 struct spdk_nvme_cpl *rsp; 1163 struct ibv_send_wr *first = NULL; 1164 struct spdk_nvmf_rdma_transport *rtransport; 1165 1166 *data_posted = 0; 1167 qpair = req->qpair; 1168 rsp = &req->rsp->nvme_cpl; 1169 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1170 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1171 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1172 struct spdk_nvmf_rdma_transport, transport); 1173 1174 /* Advance our sq_head pointer */ 1175 if (qpair->sq_head == qpair->sq_head_max) { 1176 qpair->sq_head = 0; 1177 } else { 1178 qpair->sq_head++; 1179 } 1180 rsp->sqhd = qpair->sq_head; 1181 1182 /* queue the capsule for the recv buffer */ 1183 assert(rdma_req->recv != NULL); 1184 1185 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1186 1187 rdma_req->recv = NULL; 1188 assert(rqpair->current_recv_depth > 0); 1189 rqpair->current_recv_depth--; 1190 1191 /* Build the response which consists of optional 1192 * RDMA WRITEs to transfer data, plus an RDMA SEND 1193 * containing the response. 1194 */ 1195 first = &rdma_req->rsp.wr; 1196 1197 if (spdk_unlikely(rsp->status.sc != SPDK_NVME_SC_SUCCESS)) { 1198 /* On failure, data was not read from the controller. So clear the 1199 * number of outstanding data WRs to zero. 1200 */ 1201 rdma_req->num_outstanding_data_wr = 0; 1202 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1203 first = rdma_req->transfer_wr; 1204 *data_posted = 1; 1205 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1206 } 1207 if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1208 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1209 } 1210 if (rtransport->rdma_opts.no_wr_batching) { 1211 _poller_submit_sends(rtransport, rqpair->poller); 1212 } 1213 1214 /* +1 for the rsp wr */ 1215 assert(rqpair->current_send_depth + num_outstanding_data_wr + 1 <= rqpair->max_send_depth); 1216 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1217 1218 return 0; 1219 } 1220 1221 static int 1222 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1223 { 1224 struct spdk_nvmf_rdma_accept_private_data accept_data; 1225 struct rdma_conn_param ctrlr_event_data = {}; 1226 int rc; 1227 1228 accept_data.recfmt = 0; 1229 accept_data.crqsize = rqpair->max_queue_depth; 1230 1231 ctrlr_event_data.private_data = &accept_data; 1232 ctrlr_event_data.private_data_len = sizeof(accept_data); 1233 if (id->ps == RDMA_PS_TCP) { 1234 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1235 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1236 } 1237 1238 /* Configure infinite retries for the initiator side qpair. 1239 * We need to pass this value to the initiator to prevent the 1240 * initiator side NIC from completing SEND requests back to the 1241 * initiator with status rnr_retry_count_exceeded. */ 1242 ctrlr_event_data.rnr_retry_count = 0x7; 1243 1244 /* When qpair is created without use of rdma cm API, an additional 1245 * information must be provided to initiator in the connection response: 1246 * whether qpair is using SRQ and its qp_num 1247 * Fields below are ignored by rdma cm if qpair has been 1248 * created using rdma cm API. */ 1249 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1250 ctrlr_event_data.qp_num = rqpair->qp_num; 1251 1252 rc = spdk_rdma_provider_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1253 if (rc) { 1254 SPDK_ERRLOG("Error %d on spdk_rdma_provider_qp_accept\n", errno); 1255 } else { 1256 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1257 } 1258 1259 return rc; 1260 } 1261 1262 static void 1263 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1264 { 1265 struct spdk_nvmf_rdma_reject_private_data rej_data; 1266 1267 rej_data.recfmt = 0; 1268 rej_data.sts = error; 1269 1270 rdma_reject(id, &rej_data, sizeof(rej_data)); 1271 } 1272 1273 static int 1274 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1275 { 1276 struct spdk_nvmf_rdma_transport *rtransport; 1277 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1278 struct spdk_nvmf_rdma_port *port; 1279 struct rdma_conn_param *rdma_param = NULL; 1280 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1281 uint16_t max_queue_depth; 1282 uint16_t max_read_depth; 1283 1284 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1285 1286 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1287 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1288 1289 rdma_param = &event->param.conn; 1290 if (rdma_param->private_data == NULL || 1291 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1292 SPDK_ERRLOG("connect request: no private data provided\n"); 1293 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1294 return -1; 1295 } 1296 1297 private_data = rdma_param->private_data; 1298 if (private_data->recfmt != 0) { 1299 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1300 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1301 return -1; 1302 } 1303 1304 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1305 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1306 1307 port = event->listen_id->context; 1308 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1309 event->listen_id, event->listen_id->verbs, port); 1310 1311 /* Figure out the supported queue depth. This is a multi-step process 1312 * that takes into account hardware maximums, host provided values, 1313 * and our target's internal memory limits */ 1314 1315 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1316 1317 /* Start with the maximum queue depth allowed by the target */ 1318 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1319 max_read_depth = rtransport->transport.opts.max_queue_depth; 1320 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1321 rtransport->transport.opts.max_queue_depth); 1322 1323 /* Next check the local NIC's hardware limitations */ 1324 SPDK_DEBUGLOG(rdma, 1325 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1326 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1327 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1328 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1329 1330 /* Next check the remote NIC's hardware limitations */ 1331 SPDK_DEBUGLOG(rdma, 1332 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1333 rdma_param->initiator_depth, rdma_param->responder_resources); 1334 /* from man3 rdma_get_cm_event 1335 * responder_resources - Specifies the number of responder resources that is requested by the recipient. 1336 * The responder_resources field must match the initiator depth specified by the remote node when running 1337 * the rdma_connect and rdma_accept functions. */ 1338 if (rdma_param->responder_resources != 0) { 1339 if (private_data->qid) { 1340 SPDK_DEBUGLOG(rdma, "Host (Initiator) is not allowed to use RDMA operations," 1341 " responder_resources must be 0 but set to %u\n", 1342 rdma_param->responder_resources); 1343 } else { 1344 SPDK_WARNLOG("Host (Initiator) is not allowed to use RDMA operations," 1345 " responder_resources must be 0 but set to %u\n", 1346 rdma_param->responder_resources); 1347 } 1348 } 1349 /* from man3 rdma_get_cm_event 1350 * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds. 1351 * The initiator_depth field must match the responder resources specified by the remote node when running 1352 * the rdma_connect and rdma_accept functions. */ 1353 if (rdma_param->initiator_depth == 0) { 1354 SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n"); 1355 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD); 1356 return -1; 1357 } 1358 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1359 1360 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1361 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1362 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1363 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1364 1365 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1366 max_queue_depth, max_read_depth); 1367 1368 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1369 if (rqpair == NULL) { 1370 SPDK_ERRLOG("Could not allocate new connection.\n"); 1371 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1372 return -1; 1373 } 1374 1375 rqpair->device = port->device; 1376 rqpair->max_queue_depth = max_queue_depth; 1377 rqpair->max_read_depth = max_read_depth; 1378 rqpair->cm_id = event->id; 1379 rqpair->listen_id = event->listen_id; 1380 rqpair->qpair.transport = transport; 1381 STAILQ_INIT(&rqpair->ibv_events); 1382 /* use qid from the private data to determine the qpair type 1383 qid will be set to the appropriate value when the controller is created */ 1384 rqpair->qpair.qid = private_data->qid; 1385 1386 event->id->context = &rqpair->qpair; 1387 1388 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1389 1390 return 0; 1391 } 1392 1393 static inline void 1394 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1395 enum spdk_nvme_data_transfer xfer) 1396 { 1397 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1398 wr->opcode = IBV_WR_RDMA_WRITE; 1399 wr->send_flags = 0; 1400 wr->next = next; 1401 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1402 wr->opcode = IBV_WR_RDMA_READ; 1403 wr->send_flags = IBV_SEND_SIGNALED; 1404 wr->next = NULL; 1405 } else { 1406 assert(0); 1407 } 1408 } 1409 1410 static int 1411 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1412 struct spdk_nvmf_rdma_request *rdma_req, 1413 uint32_t num_sgl_descriptors) 1414 { 1415 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1416 struct spdk_nvmf_rdma_request_data *current_data_wr; 1417 uint32_t i; 1418 1419 if (spdk_unlikely(num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES)) { 1420 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1421 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1422 return -EINVAL; 1423 } 1424 1425 if (spdk_unlikely(spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, 1426 num_sgl_descriptors))) { 1427 return -ENOMEM; 1428 } 1429 1430 current_data_wr = &rdma_req->data; 1431 1432 for (i = 0; i < num_sgl_descriptors; i++) { 1433 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1434 current_data_wr->wr.next = &work_requests[i]->wr; 1435 current_data_wr = work_requests[i]; 1436 current_data_wr->wr.sg_list = current_data_wr->sgl; 1437 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1438 } 1439 1440 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1441 1442 return 0; 1443 } 1444 1445 static inline void 1446 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1447 { 1448 struct ibv_send_wr *wr = &rdma_req->data.wr; 1449 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1450 1451 wr->wr.rdma.rkey = sgl->keyed.key; 1452 wr->wr.rdma.remote_addr = sgl->address; 1453 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1454 } 1455 1456 static inline void 1457 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1458 { 1459 struct ibv_send_wr *wr = &rdma_req->data.wr; 1460 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1461 uint32_t i; 1462 int j; 1463 uint64_t remote_addr_offset = 0; 1464 1465 for (i = 0; i < num_wrs; ++i) { 1466 wr->wr.rdma.rkey = sgl->keyed.key; 1467 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1468 for (j = 0; j < wr->num_sge; ++j) { 1469 remote_addr_offset += wr->sg_list[j].length; 1470 } 1471 wr = wr->next; 1472 } 1473 } 1474 1475 static int 1476 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_device *device, 1477 struct spdk_nvmf_rdma_request *rdma_req, 1478 struct ibv_send_wr *wr, 1479 uint32_t total_length) 1480 { 1481 struct spdk_rdma_utils_memory_translation mem_translation; 1482 struct ibv_sge *sg_ele; 1483 struct iovec *iov; 1484 uint32_t lkey, remaining; 1485 int rc; 1486 1487 wr->num_sge = 0; 1488 1489 while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) { 1490 iov = &rdma_req->req.iov[rdma_req->iovpos]; 1491 rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1492 if (spdk_unlikely(rc)) { 1493 return rc; 1494 } 1495 1496 lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation); 1497 sg_ele = &wr->sg_list[wr->num_sge]; 1498 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1499 1500 sg_ele->lkey = lkey; 1501 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset; 1502 sg_ele->length = remaining; 1503 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1504 sg_ele->length); 1505 rdma_req->offset += sg_ele->length; 1506 total_length -= sg_ele->length; 1507 wr->num_sge++; 1508 1509 if (rdma_req->offset == iov->iov_len) { 1510 rdma_req->offset = 0; 1511 rdma_req->iovpos++; 1512 } 1513 } 1514 1515 if (spdk_unlikely(total_length)) { 1516 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1517 return -EINVAL; 1518 } 1519 1520 return 0; 1521 } 1522 1523 static int 1524 nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_device *device, 1525 struct spdk_nvmf_rdma_request *rdma_req, 1526 struct ibv_send_wr *wr, 1527 uint32_t total_length, 1528 uint32_t num_extra_wrs) 1529 { 1530 struct spdk_rdma_utils_memory_translation mem_translation; 1531 struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx; 1532 struct ibv_sge *sg_ele; 1533 struct iovec *iov; 1534 struct iovec *rdma_iov; 1535 uint32_t lkey, remaining; 1536 uint32_t remaining_data_block, data_block_size, md_size; 1537 uint32_t sge_len; 1538 int rc; 1539 1540 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1541 1542 if (spdk_likely(!rdma_req->req.stripped_data)) { 1543 rdma_iov = rdma_req->req.iov; 1544 remaining_data_block = data_block_size; 1545 md_size = dif_ctx->md_size; 1546 } else { 1547 rdma_iov = rdma_req->req.stripped_data->iov; 1548 total_length = total_length / dif_ctx->block_size * data_block_size; 1549 remaining_data_block = total_length; 1550 md_size = 0; 1551 } 1552 1553 wr->num_sge = 0; 1554 1555 while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1556 iov = rdma_iov + rdma_req->iovpos; 1557 rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1558 if (spdk_unlikely(rc)) { 1559 return rc; 1560 } 1561 1562 lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation); 1563 sg_ele = &wr->sg_list[wr->num_sge]; 1564 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1565 1566 while (remaining) { 1567 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1568 if (num_extra_wrs > 0 && wr->next) { 1569 wr = wr->next; 1570 wr->num_sge = 0; 1571 sg_ele = &wr->sg_list[wr->num_sge]; 1572 num_extra_wrs--; 1573 } else { 1574 break; 1575 } 1576 } 1577 sg_ele->lkey = lkey; 1578 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset); 1579 sge_len = spdk_min(remaining, remaining_data_block); 1580 sg_ele->length = sge_len; 1581 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, 1582 sg_ele->addr, sg_ele->length); 1583 remaining -= sge_len; 1584 remaining_data_block -= sge_len; 1585 rdma_req->offset += sge_len; 1586 total_length -= sge_len; 1587 1588 sg_ele++; 1589 wr->num_sge++; 1590 1591 if (remaining_data_block == 0) { 1592 /* skip metadata */ 1593 rdma_req->offset += md_size; 1594 total_length -= md_size; 1595 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1596 remaining -= spdk_min(remaining, md_size); 1597 remaining_data_block = data_block_size; 1598 } 1599 1600 if (remaining == 0) { 1601 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1602 the remaining metadata bits at the beginning of the next buffer */ 1603 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset); 1604 rdma_req->iovpos++; 1605 } 1606 } 1607 } 1608 1609 if (spdk_unlikely(total_length)) { 1610 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1611 return -EINVAL; 1612 } 1613 1614 return 0; 1615 } 1616 1617 static inline uint32_t 1618 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1619 { 1620 /* estimate the number of SG entries and WRs needed to process the request */ 1621 uint32_t num_sge = 0; 1622 uint32_t i; 1623 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1624 1625 for (i = 0; i < num_buffers && length > 0; i++) { 1626 uint32_t buffer_len = spdk_min(length, io_unit_size); 1627 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1628 1629 if (num_sge_in_block * block_size > buffer_len) { 1630 ++num_sge_in_block; 1631 } 1632 num_sge += num_sge_in_block; 1633 length -= buffer_len; 1634 } 1635 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1636 } 1637 1638 static int 1639 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1640 struct spdk_nvmf_rdma_device *device, 1641 struct spdk_nvmf_rdma_request *rdma_req) 1642 { 1643 struct spdk_nvmf_rdma_qpair *rqpair; 1644 struct spdk_nvmf_rdma_poll_group *rgroup; 1645 struct spdk_nvmf_request *req = &rdma_req->req; 1646 struct ibv_send_wr *wr = &rdma_req->data.wr; 1647 int rc; 1648 uint32_t num_wrs = 1; 1649 uint32_t length; 1650 1651 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1652 rgroup = rqpair->poller->group; 1653 1654 /* rdma wr specifics */ 1655 nvmf_rdma_setup_request(rdma_req); 1656 1657 length = req->length; 1658 if (spdk_unlikely(req->dif_enabled)) { 1659 req->dif.orig_length = length; 1660 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1661 req->dif.elba_length = length; 1662 } 1663 1664 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1665 length); 1666 if (spdk_unlikely(rc != 0)) { 1667 return rc; 1668 } 1669 1670 assert(req->iovcnt <= rqpair->max_send_sge); 1671 1672 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1673 * the stripped_buffers are got for DIF stripping. */ 1674 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1675 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1676 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1677 &rtransport->transport, req->dif.orig_length); 1678 if (rc != 0) { 1679 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1680 } 1681 } 1682 1683 rdma_req->iovpos = 0; 1684 1685 if (spdk_unlikely(req->dif_enabled)) { 1686 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1687 req->dif.dif_ctx.block_size); 1688 if (num_wrs > 1) { 1689 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1690 if (spdk_unlikely(rc != 0)) { 1691 goto err_exit; 1692 } 1693 } 1694 1695 rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, wr, length, num_wrs - 1); 1696 if (spdk_unlikely(rc != 0)) { 1697 goto err_exit; 1698 } 1699 1700 if (num_wrs > 1) { 1701 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1702 } 1703 } else { 1704 rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, wr, length); 1705 if (spdk_unlikely(rc != 0)) { 1706 goto err_exit; 1707 } 1708 } 1709 1710 /* set the number of outstanding data WRs for this request. */ 1711 rdma_req->num_outstanding_data_wr = num_wrs; 1712 1713 return rc; 1714 1715 err_exit: 1716 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1717 nvmf_rdma_request_free_data(rdma_req, rtransport); 1718 req->iovcnt = 0; 1719 return rc; 1720 } 1721 1722 static int 1723 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1724 struct spdk_nvmf_rdma_device *device, 1725 struct spdk_nvmf_rdma_request *rdma_req) 1726 { 1727 struct spdk_nvmf_rdma_qpair *rqpair; 1728 struct spdk_nvmf_rdma_poll_group *rgroup; 1729 struct ibv_send_wr *current_wr; 1730 struct spdk_nvmf_request *req = &rdma_req->req; 1731 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1732 uint32_t num_sgl_descriptors; 1733 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0; 1734 uint32_t i; 1735 int rc; 1736 1737 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1738 rgroup = rqpair->poller->group; 1739 1740 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1741 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1742 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1743 1744 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1745 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1746 1747 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1748 for (i = 0; i < num_sgl_descriptors; i++) { 1749 if (spdk_likely(!req->dif_enabled)) { 1750 lengths[i] = desc->keyed.length; 1751 } else { 1752 req->dif.orig_length += desc->keyed.length; 1753 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1754 req->dif.elba_length += lengths[i]; 1755 } 1756 total_length += lengths[i]; 1757 desc++; 1758 } 1759 1760 if (spdk_unlikely(total_length > rtransport->transport.opts.max_io_size)) { 1761 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n", 1762 total_length, rtransport->transport.opts.max_io_size); 1763 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1764 return -EINVAL; 1765 } 1766 1767 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1); 1768 if (spdk_unlikely(rc != 0)) { 1769 return -ENOMEM; 1770 } 1771 1772 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length); 1773 if (spdk_unlikely(rc != 0)) { 1774 nvmf_rdma_request_free_data(rdma_req, rtransport); 1775 return rc; 1776 } 1777 1778 /* When dif_insert_or_strip is true and the I/O data length is greater than one block, 1779 * the stripped_buffers are got for DIF stripping. */ 1780 if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) 1781 && (req->dif.elba_length > req->dif.dif_ctx.block_size))) { 1782 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group, 1783 &rtransport->transport, req->dif.orig_length); 1784 if (spdk_unlikely(rc != 0)) { 1785 SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc); 1786 } 1787 } 1788 1789 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1790 current_wr = &rdma_req->data.wr; 1791 assert(current_wr != NULL); 1792 1793 req->length = 0; 1794 rdma_req->iovpos = 0; 1795 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1796 for (i = 0; i < num_sgl_descriptors; i++) { 1797 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1798 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1799 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1800 rc = -EINVAL; 1801 goto err_exit; 1802 } 1803 1804 if (spdk_likely(!req->dif_enabled)) { 1805 rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, current_wr, lengths[i]); 1806 } else { 1807 rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, current_wr, 1808 lengths[i], 0); 1809 } 1810 if (spdk_unlikely(rc != 0)) { 1811 rc = -ENOMEM; 1812 goto err_exit; 1813 } 1814 1815 req->length += desc->keyed.length; 1816 current_wr->wr.rdma.rkey = desc->keyed.key; 1817 current_wr->wr.rdma.remote_addr = desc->address; 1818 current_wr = current_wr->next; 1819 desc++; 1820 } 1821 1822 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1823 /* Go back to the last descriptor in the list. */ 1824 desc--; 1825 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1826 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1827 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1828 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1829 } 1830 } 1831 #endif 1832 1833 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1834 1835 return 0; 1836 1837 err_exit: 1838 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1839 nvmf_rdma_request_free_data(rdma_req, rtransport); 1840 return rc; 1841 } 1842 1843 static int 1844 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1845 struct spdk_nvmf_rdma_device *device, 1846 struct spdk_nvmf_rdma_request *rdma_req) 1847 { 1848 struct spdk_nvmf_request *req = &rdma_req->req; 1849 struct spdk_nvme_cpl *rsp; 1850 struct spdk_nvme_sgl_descriptor *sgl; 1851 int rc; 1852 uint32_t length; 1853 1854 rsp = &req->rsp->nvme_cpl; 1855 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1856 1857 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1858 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1859 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1860 1861 length = sgl->keyed.length; 1862 if (spdk_unlikely(length > rtransport->transport.opts.max_io_size)) { 1863 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1864 length, rtransport->transport.opts.max_io_size); 1865 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1866 return -1; 1867 } 1868 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1869 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1870 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1871 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1872 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1873 } 1874 } 1875 #endif 1876 1877 /* fill request length and populate iovs */ 1878 req->length = length; 1879 1880 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req); 1881 if (spdk_unlikely(rc < 0)) { 1882 if (rc == -EINVAL) { 1883 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1884 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1885 return -1; 1886 } 1887 /* No available buffers. Queue this request up. */ 1888 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1889 return 0; 1890 } 1891 1892 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1893 req->iovcnt); 1894 1895 return 0; 1896 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1897 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1898 uint64_t offset = sgl->address; 1899 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1900 1901 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1902 offset, sgl->unkeyed.length); 1903 1904 if (spdk_unlikely(offset > max_len)) { 1905 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1906 offset, max_len); 1907 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1908 return -1; 1909 } 1910 max_len -= (uint32_t)offset; 1911 1912 if (spdk_unlikely(sgl->unkeyed.length > max_len)) { 1913 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1914 sgl->unkeyed.length, max_len); 1915 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1916 return -1; 1917 } 1918 1919 rdma_req->num_outstanding_data_wr = 0; 1920 req->data_from_pool = false; 1921 req->length = sgl->unkeyed.length; 1922 1923 req->iov[0].iov_base = rdma_req->recv->buf + offset; 1924 req->iov[0].iov_len = req->length; 1925 req->iovcnt = 1; 1926 1927 return 0; 1928 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1929 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1930 1931 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1932 if (spdk_unlikely(rc == -ENOMEM)) { 1933 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1934 return 0; 1935 } else if (spdk_unlikely(rc == -EINVAL)) { 1936 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1937 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1938 return -1; 1939 } 1940 1941 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1942 req->iovcnt); 1943 1944 return 0; 1945 } 1946 1947 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1948 sgl->generic.type, sgl->generic.subtype); 1949 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1950 return -1; 1951 } 1952 1953 static void 1954 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1955 struct spdk_nvmf_rdma_transport *rtransport) 1956 { 1957 struct spdk_nvmf_rdma_qpair *rqpair; 1958 struct spdk_nvmf_rdma_poll_group *rgroup; 1959 1960 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1961 if (rdma_req->req.data_from_pool) { 1962 rgroup = rqpair->poller->group; 1963 1964 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1965 } 1966 if (rdma_req->req.stripped_data) { 1967 nvmf_request_free_stripped_buffers(&rdma_req->req, 1968 &rqpair->poller->group->group, 1969 &rtransport->transport); 1970 } 1971 nvmf_rdma_request_free_data(rdma_req, rtransport); 1972 rdma_req->req.length = 0; 1973 rdma_req->req.iovcnt = 0; 1974 rdma_req->offset = 0; 1975 rdma_req->req.dif_enabled = false; 1976 rdma_req->fused_failed = false; 1977 rdma_req->transfer_wr = NULL; 1978 if (rdma_req->fused_pair) { 1979 /* This req was part of a valid fused pair, but failed before it got to 1980 * READ_TO_EXECUTE state. This means we need to fail the other request 1981 * in the pair, because it is no longer part of a valid pair. If the pair 1982 * already reached READY_TO_EXECUTE state, we need to kick it. 1983 */ 1984 rdma_req->fused_pair->fused_failed = true; 1985 if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 1986 nvmf_rdma_request_process(rtransport, rdma_req->fused_pair); 1987 } 1988 rdma_req->fused_pair = NULL; 1989 } 1990 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1991 rqpair->qd--; 1992 1993 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1994 rqpair->qpair.queue_depth--; 1995 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1996 } 1997 1998 static void 1999 nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport, 2000 struct spdk_nvmf_rdma_qpair *rqpair, 2001 struct spdk_nvmf_rdma_request *rdma_req) 2002 { 2003 enum spdk_nvme_cmd_fuse last, next; 2004 2005 last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2006 next = rdma_req->req.cmd->nvme_cmd.fuse; 2007 2008 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2009 2010 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2011 return; 2012 } 2013 2014 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2015 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2016 /* This is a valid pair of fused commands. Point them at each other 2017 * so they can be submitted consecutively once ready to be executed. 2018 */ 2019 rqpair->fused_first->fused_pair = rdma_req; 2020 rdma_req->fused_pair = rqpair->fused_first; 2021 rqpair->fused_first = NULL; 2022 return; 2023 } else { 2024 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2025 rqpair->fused_first->fused_failed = true; 2026 2027 /* If the last req is in READY_TO_EXECUTE state, then call 2028 * nvmf_rdma_request_process(), otherwise nothing else will kick it. 2029 */ 2030 if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2031 nvmf_rdma_request_process(rtransport, rqpair->fused_first); 2032 } 2033 2034 rqpair->fused_first = NULL; 2035 } 2036 } 2037 2038 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2039 /* Set rqpair->fused_first here so that we know to check that the next request 2040 * is a SECOND (and to fail this one if it isn't). 2041 */ 2042 rqpair->fused_first = rdma_req; 2043 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2044 /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */ 2045 rdma_req->fused_failed = true; 2046 } 2047 } 2048 2049 bool 2050 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 2051 struct spdk_nvmf_rdma_request *rdma_req) 2052 { 2053 struct spdk_nvmf_rdma_qpair *rqpair; 2054 struct spdk_nvmf_rdma_device *device; 2055 struct spdk_nvmf_rdma_poll_group *rgroup; 2056 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2057 int rc; 2058 struct spdk_nvmf_rdma_recv *rdma_recv; 2059 enum spdk_nvmf_rdma_request_state prev_state; 2060 bool progress = false; 2061 int data_posted; 2062 uint32_t num_blocks, num_rdma_reads_available, qdepth; 2063 2064 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2065 device = rqpair->device; 2066 rgroup = rqpair->poller->group; 2067 2068 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 2069 2070 /* If the queue pair is in an error state, force the request to the completed state 2071 * to release resources. */ 2072 if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 2073 switch (rdma_req->state) { 2074 case RDMA_REQUEST_STATE_NEED_BUFFER: 2075 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 2076 break; 2077 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2078 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2079 break; 2080 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2081 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2082 break; 2083 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 2084 STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2085 break; 2086 default: 2087 break; 2088 } 2089 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2090 } 2091 2092 /* The loop here is to allow for several back-to-back state changes. */ 2093 do { 2094 prev_state = rdma_req->state; 2095 2096 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 2097 2098 switch (rdma_req->state) { 2099 case RDMA_REQUEST_STATE_FREE: 2100 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 2101 * to escape this state. */ 2102 break; 2103 case RDMA_REQUEST_STATE_NEW: 2104 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 2105 (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth); 2106 rdma_recv = rdma_req->recv; 2107 2108 /* The first element of the SGL is the NVMe command */ 2109 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 2110 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 2111 rdma_req->transfer_wr = &rdma_req->data.wr; 2112 2113 if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 2114 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2115 break; 2116 } 2117 2118 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 2119 rdma_req->req.dif_enabled = true; 2120 } 2121 2122 nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req); 2123 2124 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2125 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 2126 rdma_req->rsp.wr.imm_data = 0; 2127 #endif 2128 2129 /* The next state transition depends on the data transfer needs of this request. */ 2130 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2131 2132 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2133 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2134 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2135 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2136 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2137 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 2138 break; 2139 } 2140 2141 /* If no data to transfer, ready to execute. */ 2142 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2143 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2144 break; 2145 } 2146 2147 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2148 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2149 break; 2150 case RDMA_REQUEST_STATE_NEED_BUFFER: 2151 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2152 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2153 2154 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2155 2156 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2157 /* This request needs to wait in line to obtain a buffer */ 2158 break; 2159 } 2160 2161 /* Try to get a data buffer */ 2162 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2163 if (spdk_unlikely(rc < 0)) { 2164 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2165 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2166 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2167 break; 2168 } 2169 2170 if (rdma_req->req.iovcnt == 0) { 2171 /* No buffers available. */ 2172 rgroup->stat.pending_data_buffer++; 2173 break; 2174 } 2175 2176 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2177 2178 /* If data is transferring from host to controller and the data didn't 2179 * arrive using in capsule data, we need to do a transfer from the host. 2180 */ 2181 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2182 rdma_req->req.data_from_pool) { 2183 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2184 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2185 break; 2186 } 2187 2188 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2189 break; 2190 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2191 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2192 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2193 2194 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2195 /* This request needs to wait in line to perform RDMA */ 2196 break; 2197 } 2198 assert(rqpair->max_send_depth >= rqpair->current_send_depth); 2199 qdepth = rqpair->max_send_depth - rqpair->current_send_depth; 2200 assert(rqpair->max_read_depth >= rqpair->current_read_depth); 2201 num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth; 2202 if (rdma_req->num_outstanding_data_wr > qdepth || 2203 rdma_req->num_outstanding_data_wr > num_rdma_reads_available) { 2204 if (num_rdma_reads_available && qdepth) { 2205 /* Send as much as we can */ 2206 request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth)); 2207 } else { 2208 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2209 rqpair->poller->stat.pending_rdma_read++; 2210 break; 2211 } 2212 } 2213 2214 /* We have already verified that this request is the head of the queue. */ 2215 if (rdma_req->num_remaining_data_wr == 0) { 2216 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2217 } 2218 2219 rc = request_transfer_in(&rdma_req->req); 2220 if (spdk_likely(rc == 0)) { 2221 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2222 } else { 2223 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2224 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2225 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2226 } 2227 break; 2228 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2229 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2230 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2231 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2232 * to escape this state. */ 2233 break; 2234 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2235 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2236 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2237 2238 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2239 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2240 /* generate DIF for write operation */ 2241 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2242 assert(num_blocks > 0); 2243 2244 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2245 num_blocks, &rdma_req->req.dif.dif_ctx); 2246 if (rc != 0) { 2247 SPDK_ERRLOG("DIF generation failed\n"); 2248 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2249 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 2250 break; 2251 } 2252 } 2253 2254 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2255 /* set extended length before IO operation */ 2256 rdma_req->req.length = rdma_req->req.dif.elba_length; 2257 } 2258 2259 if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 2260 if (rdma_req->fused_failed) { 2261 /* This request failed FUSED semantics. Fail it immediately, without 2262 * even sending it to the target layer. 2263 */ 2264 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2265 rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED; 2266 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2267 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2268 break; 2269 } 2270 2271 if (rdma_req->fused_pair == NULL || 2272 rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) { 2273 /* This request is ready to execute, but either we don't know yet if it's 2274 * valid - i.e. this is a FIRST but we haven't received the next 2275 * request yet or the other request of this fused pair isn't ready to 2276 * execute. So break here and this request will get processed later either 2277 * when the other request is ready or we find that this request isn't valid. 2278 */ 2279 break; 2280 } 2281 } 2282 2283 /* If we get to this point, and this request is a fused command, we know that 2284 * it is part of valid sequence (FIRST followed by a SECOND) and that both 2285 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 2286 * request, and the other request of the fused pair, in the correct order. 2287 * Also clear the ->fused_pair pointers on both requests, since after this point 2288 * we no longer need to maintain the relationship between these two requests. 2289 */ 2290 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 2291 assert(rdma_req->fused_pair != NULL); 2292 assert(rdma_req->fused_pair->fused_pair != NULL); 2293 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2294 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2295 rdma_req->fused_pair->fused_pair = NULL; 2296 rdma_req->fused_pair = NULL; 2297 } 2298 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2299 spdk_nvmf_request_exec(&rdma_req->req); 2300 if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 2301 assert(rdma_req->fused_pair != NULL); 2302 assert(rdma_req->fused_pair->fused_pair != NULL); 2303 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING; 2304 spdk_nvmf_request_exec(&rdma_req->fused_pair->req); 2305 rdma_req->fused_pair->fused_pair = NULL; 2306 rdma_req->fused_pair = NULL; 2307 } 2308 break; 2309 case RDMA_REQUEST_STATE_EXECUTING: 2310 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2311 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2312 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2313 * to escape this state. */ 2314 break; 2315 case RDMA_REQUEST_STATE_EXECUTED: 2316 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2317 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2318 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2319 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2320 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2321 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2322 } else { 2323 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2324 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2325 } 2326 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2327 /* restore the original length */ 2328 rdma_req->req.length = rdma_req->req.dif.orig_length; 2329 2330 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2331 struct spdk_dif_error error_blk; 2332 2333 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2334 if (!rdma_req->req.stripped_data) { 2335 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2336 &rdma_req->req.dif.dif_ctx, &error_blk); 2337 } else { 2338 rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov, 2339 rdma_req->req.stripped_data->iovcnt, 2340 rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2341 &rdma_req->req.dif.dif_ctx, &error_blk); 2342 } 2343 if (rc) { 2344 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2345 2346 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2347 error_blk.err_offset); 2348 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2349 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2350 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2351 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link); 2352 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 2353 } 2354 } 2355 } 2356 break; 2357 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2358 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2359 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2360 2361 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2362 /* This request needs to wait in line to perform RDMA */ 2363 break; 2364 } 2365 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2366 rqpair->max_send_depth) { 2367 /* We can only have so many WRs outstanding. we have to wait until some finish. 2368 * +1 since each request has an additional wr in the resp. */ 2369 rqpair->poller->stat.pending_rdma_write++; 2370 break; 2371 } 2372 2373 /* We have already verified that this request is the head of the queue. */ 2374 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2375 2376 /* The data transfer will be kicked off from 2377 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2378 * We verified that data + response fit into send queue, so we can go to the next state directly 2379 */ 2380 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2381 break; 2382 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING: 2383 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 0, 0, 2384 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2385 2386 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_send_queue)) { 2387 /* This request needs to wait in line to send the completion */ 2388 break; 2389 } 2390 2391 assert(rqpair->current_send_depth <= rqpair->max_send_depth); 2392 if (rqpair->current_send_depth == rqpair->max_send_depth) { 2393 /* We can only have so many WRs outstanding. we have to wait until some finish */ 2394 rqpair->poller->stat.pending_rdma_send++; 2395 break; 2396 } 2397 2398 /* We have already verified that this request is the head of the queue. */ 2399 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_send_queue, state_link); 2400 2401 /* The response sending will be kicked off from 2402 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2403 */ 2404 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2405 break; 2406 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2407 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2408 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2409 rc = request_transfer_out(&rdma_req->req, &data_posted); 2410 assert(rc == 0); /* No good way to handle this currently */ 2411 if (spdk_unlikely(rc)) { 2412 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2413 } else { 2414 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2415 RDMA_REQUEST_STATE_COMPLETING; 2416 } 2417 break; 2418 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2419 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2420 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2421 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2422 * to escape this state. */ 2423 break; 2424 case RDMA_REQUEST_STATE_COMPLETING: 2425 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2426 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2427 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2428 * to escape this state. */ 2429 break; 2430 case RDMA_REQUEST_STATE_COMPLETED: 2431 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2432 (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth); 2433 2434 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2435 _nvmf_rdma_request_free(rdma_req, rtransport); 2436 break; 2437 case RDMA_REQUEST_NUM_STATES: 2438 default: 2439 assert(0); 2440 break; 2441 } 2442 2443 if (rdma_req->state != prev_state) { 2444 progress = true; 2445 } 2446 } while (rdma_req->state != prev_state); 2447 2448 return progress; 2449 } 2450 2451 /* Public API callbacks begin here */ 2452 2453 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2454 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2455 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2456 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2457 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2458 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2459 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2460 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2461 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX 2462 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2463 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2464 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2465 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2466 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2467 #define SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE 4095 2468 2469 static void 2470 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2471 { 2472 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2473 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2474 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2475 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2476 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2477 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2478 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2479 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2480 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2481 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2482 opts->transport_specific = NULL; 2483 opts->data_wr_pool_size = SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE; 2484 } 2485 2486 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2487 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2488 2489 static inline bool 2490 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2491 { 2492 return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD || 2493 device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW; 2494 } 2495 2496 static int nvmf_rdma_accept(void *ctx); 2497 static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport); 2498 static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2499 struct spdk_nvmf_rdma_device *device); 2500 2501 static int 2502 create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context, 2503 struct spdk_nvmf_rdma_device **new_device) 2504 { 2505 struct spdk_nvmf_rdma_device *device; 2506 int flag = 0; 2507 int rc = 0; 2508 2509 device = calloc(1, sizeof(*device)); 2510 if (!device) { 2511 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2512 return -ENOMEM; 2513 } 2514 device->context = context; 2515 rc = ibv_query_device(device->context, &device->attr); 2516 if (rc < 0) { 2517 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2518 free(device); 2519 return rc; 2520 } 2521 2522 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2523 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2524 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2525 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2526 } 2527 2528 /** 2529 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2530 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2531 * but incorrectly reports that it does. There are changes making their way 2532 * through the kernel now that will enable this feature. When they are merged, 2533 * we can conditionally enable this feature. 2534 * 2535 * TODO: enable this for versions of the kernel rxe driver that support it. 2536 */ 2537 if (nvmf_rdma_is_rxe_device(device)) { 2538 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2539 } 2540 #endif 2541 2542 /* set up device context async ev fd as NON_BLOCKING */ 2543 flag = fcntl(device->context->async_fd, F_GETFL); 2544 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2545 if (rc < 0) { 2546 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2547 free(device); 2548 return rc; 2549 } 2550 2551 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2552 SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA transport\n", device); 2553 2554 if (g_nvmf_hooks.get_ibv_pd) { 2555 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2556 } else { 2557 device->pd = ibv_alloc_pd(device->context); 2558 } 2559 2560 if (!device->pd) { 2561 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2562 destroy_ib_device(rtransport, device); 2563 return -ENOMEM; 2564 } 2565 2566 assert(device->map == NULL); 2567 2568 device->map = spdk_rdma_utils_create_mem_map(device->pd, &g_nvmf_hooks, IBV_ACCESS_LOCAL_WRITE); 2569 if (!device->map) { 2570 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2571 destroy_ib_device(rtransport, device); 2572 return -ENOMEM; 2573 } 2574 2575 assert(device->map != NULL); 2576 assert(device->pd != NULL); 2577 2578 if (new_device) { 2579 *new_device = device; 2580 } 2581 SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device), 2582 device, context); 2583 2584 return 0; 2585 } 2586 2587 static void 2588 free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2589 { 2590 if (rtransport->poll_fds) { 2591 free(rtransport->poll_fds); 2592 rtransport->poll_fds = NULL; 2593 } 2594 rtransport->npoll_fds = 0; 2595 } 2596 2597 static int 2598 generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport) 2599 { 2600 /* Set up poll descriptor array to monitor events from RDMA and IB 2601 * in a single poll syscall 2602 */ 2603 int device_count = 0; 2604 int i = 0; 2605 struct spdk_nvmf_rdma_device *device, *tmp; 2606 2607 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2608 device_count++; 2609 } 2610 2611 rtransport->npoll_fds = device_count + 1; 2612 2613 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2614 if (rtransport->poll_fds == NULL) { 2615 SPDK_ERRLOG("poll_fds allocation failed\n"); 2616 return -ENOMEM; 2617 } 2618 2619 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2620 rtransport->poll_fds[i++].events = POLLIN; 2621 2622 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2623 rtransport->poll_fds[i].fd = device->context->async_fd; 2624 rtransport->poll_fds[i++].events = POLLIN; 2625 } 2626 2627 return 0; 2628 } 2629 2630 static struct spdk_nvmf_transport * 2631 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2632 { 2633 int rc; 2634 struct spdk_nvmf_rdma_transport *rtransport; 2635 struct spdk_nvmf_rdma_device *device; 2636 struct ibv_context **contexts; 2637 size_t data_wr_pool_size; 2638 uint32_t i; 2639 int flag; 2640 uint32_t sge_count; 2641 uint32_t min_shared_buffers; 2642 uint32_t min_in_capsule_data_size; 2643 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2644 2645 rtransport = calloc(1, sizeof(*rtransport)); 2646 if (!rtransport) { 2647 return NULL; 2648 } 2649 2650 TAILQ_INIT(&rtransport->devices); 2651 TAILQ_INIT(&rtransport->ports); 2652 TAILQ_INIT(&rtransport->poll_groups); 2653 TAILQ_INIT(&rtransport->retry_ports); 2654 2655 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2656 rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2657 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2658 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2659 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2660 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2661 if (opts->transport_specific != NULL && 2662 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2663 SPDK_COUNTOF(rdma_transport_opts_decoder), 2664 &rtransport->rdma_opts)) { 2665 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2666 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2667 return NULL; 2668 } 2669 2670 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2671 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2672 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2673 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2674 " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d," 2675 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2676 opts->max_queue_depth, 2677 opts->max_io_size, 2678 opts->max_qpairs_per_ctrlr - 1, 2679 opts->io_unit_size, 2680 opts->in_capsule_data_size, 2681 opts->max_aq_depth, 2682 opts->num_shared_buffers, 2683 rtransport->rdma_opts.num_cqe, 2684 rtransport->rdma_opts.max_srq_depth, 2685 rtransport->rdma_opts.no_srq, 2686 rtransport->rdma_opts.acceptor_backlog, 2687 rtransport->rdma_opts.no_wr_batching, 2688 opts->abort_timeout_sec); 2689 2690 /* I/O unit size cannot be larger than max I/O size */ 2691 if (opts->io_unit_size > opts->max_io_size) { 2692 opts->io_unit_size = opts->max_io_size; 2693 } 2694 2695 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2696 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2697 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2698 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2699 } 2700 2701 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2702 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2703 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2704 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2705 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2706 return NULL; 2707 } 2708 2709 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 2710 if (opts->buf_cache_size < UINT32_MAX) { 2711 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2712 if (min_shared_buffers > opts->num_shared_buffers) { 2713 SPDK_ERRLOG("There are not enough buffers to satisfy" 2714 "per-poll group caches for each thread. (%" PRIu32 ")" 2715 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2716 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2717 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2718 return NULL; 2719 } 2720 } 2721 2722 sge_count = opts->max_io_size / opts->io_unit_size; 2723 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2724 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2725 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2726 return NULL; 2727 } 2728 2729 min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES; 2730 if (opts->in_capsule_data_size < min_in_capsule_data_size) { 2731 SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n", 2732 min_in_capsule_data_size); 2733 opts->in_capsule_data_size = min_in_capsule_data_size; 2734 } 2735 2736 rtransport->event_channel = rdma_create_event_channel(); 2737 if (rtransport->event_channel == NULL) { 2738 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2739 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2740 return NULL; 2741 } 2742 2743 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2744 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2745 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2746 rtransport->event_channel->fd, spdk_strerror(errno)); 2747 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2748 return NULL; 2749 } 2750 2751 data_wr_pool_size = opts->data_wr_pool_size; 2752 if (data_wr_pool_size < SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count()) { 2753 data_wr_pool_size = SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count(); 2754 SPDK_NOTICELOG("data_wr_pool_size is changed to %zu to guarantee enough cache for handling " 2755 "at least one IO in each core\n", data_wr_pool_size); 2756 } 2757 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", data_wr_pool_size, 2758 sizeof(struct spdk_nvmf_rdma_request_data), SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2759 SPDK_ENV_SOCKET_ID_ANY); 2760 if (!rtransport->data_wr_pool) { 2761 if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) { 2762 SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n"); 2763 SPDK_ERRLOG("Probably running in multiprocess environment, which is " 2764 "unsupported by the nvmf library\n"); 2765 } else { 2766 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2767 } 2768 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2769 return NULL; 2770 } 2771 2772 contexts = rdma_get_devices(NULL); 2773 if (contexts == NULL) { 2774 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2775 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2776 return NULL; 2777 } 2778 2779 i = 0; 2780 rc = 0; 2781 while (contexts[i] != NULL) { 2782 rc = create_ib_device(rtransport, contexts[i], &device); 2783 if (rc < 0) { 2784 break; 2785 } 2786 i++; 2787 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2788 device->is_ready = true; 2789 } 2790 rdma_free_devices(contexts); 2791 2792 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2793 /* divide and round up. */ 2794 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2795 2796 /* round up to the nearest 4k. */ 2797 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2798 2799 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2800 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2801 opts->io_unit_size); 2802 } 2803 2804 if (rc < 0) { 2805 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2806 return NULL; 2807 } 2808 2809 rc = generate_poll_fds(rtransport); 2810 if (rc < 0) { 2811 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2812 return NULL; 2813 } 2814 2815 rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport, 2816 opts->acceptor_poll_rate); 2817 if (!rtransport->accept_poller) { 2818 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2819 return NULL; 2820 } 2821 2822 return &rtransport->transport; 2823 } 2824 2825 static void 2826 destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 2827 struct spdk_nvmf_rdma_device *device) 2828 { 2829 TAILQ_REMOVE(&rtransport->devices, device, link); 2830 spdk_rdma_utils_free_mem_map(&device->map); 2831 if (device->pd) { 2832 if (!g_nvmf_hooks.get_ibv_pd) { 2833 ibv_dealloc_pd(device->pd); 2834 } 2835 } 2836 SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device); 2837 free(device); 2838 } 2839 2840 static void 2841 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2842 { 2843 struct spdk_nvmf_rdma_transport *rtransport; 2844 assert(w != NULL); 2845 2846 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2847 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2848 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2849 if (rtransport->rdma_opts.no_srq == true) { 2850 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2851 } 2852 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2853 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2854 } 2855 2856 static int 2857 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2858 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2859 { 2860 struct spdk_nvmf_rdma_transport *rtransport; 2861 struct spdk_nvmf_rdma_port *port, *port_tmp; 2862 struct spdk_nvmf_rdma_device *device, *device_tmp; 2863 2864 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2865 2866 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) { 2867 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 2868 free(port); 2869 } 2870 2871 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2872 TAILQ_REMOVE(&rtransport->ports, port, link); 2873 rdma_destroy_id(port->id); 2874 free(port); 2875 } 2876 2877 free_poll_fds(rtransport); 2878 2879 if (rtransport->event_channel != NULL) { 2880 rdma_destroy_event_channel(rtransport->event_channel); 2881 } 2882 2883 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2884 destroy_ib_device(rtransport, device); 2885 } 2886 2887 if (rtransport->data_wr_pool != NULL) { 2888 if (spdk_mempool_count(rtransport->data_wr_pool) != transport->opts.data_wr_pool_size) { 2889 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2890 spdk_mempool_count(rtransport->data_wr_pool), 2891 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2892 } 2893 } 2894 2895 spdk_mempool_free(rtransport->data_wr_pool); 2896 2897 spdk_poller_unregister(&rtransport->accept_poller); 2898 free(rtransport); 2899 2900 if (cb_fn) { 2901 cb_fn(cb_arg); 2902 } 2903 return 0; 2904 } 2905 2906 static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2907 struct spdk_nvme_transport_id *trid, 2908 bool peer); 2909 2910 static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport); 2911 2912 static int 2913 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2914 struct spdk_nvmf_listen_opts *listen_opts) 2915 { 2916 struct spdk_nvmf_rdma_transport *rtransport; 2917 struct spdk_nvmf_rdma_device *device; 2918 struct spdk_nvmf_rdma_port *port, *tmp_port; 2919 struct addrinfo *res; 2920 struct addrinfo hints; 2921 int family; 2922 int rc; 2923 long int port_val; 2924 bool is_retry = false; 2925 2926 if (!strlen(trid->trsvcid)) { 2927 SPDK_ERRLOG("Service id is required\n"); 2928 return -EINVAL; 2929 } 2930 2931 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2932 assert(rtransport->event_channel != NULL); 2933 2934 port = calloc(1, sizeof(*port)); 2935 if (!port) { 2936 SPDK_ERRLOG("Port allocation failed\n"); 2937 return -ENOMEM; 2938 } 2939 2940 port->trid = trid; 2941 2942 switch (trid->adrfam) { 2943 case SPDK_NVMF_ADRFAM_IPV4: 2944 family = AF_INET; 2945 break; 2946 case SPDK_NVMF_ADRFAM_IPV6: 2947 family = AF_INET6; 2948 break; 2949 default: 2950 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2951 free(port); 2952 return -EINVAL; 2953 } 2954 2955 memset(&hints, 0, sizeof(hints)); 2956 hints.ai_family = family; 2957 hints.ai_flags = AI_NUMERICSERV; 2958 hints.ai_socktype = SOCK_STREAM; 2959 hints.ai_protocol = 0; 2960 2961 /* Range check the trsvcid. Fail in 3 cases: 2962 * < 0: means that spdk_strtol hit an error 2963 * 0: this results in ephemeral port which we don't want 2964 * > 65535: port too high 2965 */ 2966 port_val = spdk_strtol(trid->trsvcid, 10); 2967 if (port_val <= 0 || port_val > 65535) { 2968 SPDK_ERRLOG("invalid trsvcid %s\n", trid->trsvcid); 2969 free(port); 2970 return -EINVAL; 2971 } 2972 2973 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2974 if (rc) { 2975 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2976 free(port); 2977 return -(abs(rc)); 2978 } 2979 2980 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2981 if (rc < 0) { 2982 SPDK_ERRLOG("rdma_create_id() failed\n"); 2983 freeaddrinfo(res); 2984 free(port); 2985 return rc; 2986 } 2987 2988 rc = rdma_bind_addr(port->id, res->ai_addr); 2989 freeaddrinfo(res); 2990 2991 if (rc < 0) { 2992 TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) { 2993 if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) { 2994 is_retry = true; 2995 break; 2996 } 2997 } 2998 if (!is_retry) { 2999 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 3000 } 3001 rdma_destroy_id(port->id); 3002 free(port); 3003 return rc; 3004 } 3005 3006 if (!port->id->verbs) { 3007 SPDK_ERRLOG("ibv_context is null\n"); 3008 rdma_destroy_id(port->id); 3009 free(port); 3010 return -1; 3011 } 3012 3013 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 3014 if (rc < 0) { 3015 SPDK_ERRLOG("rdma_listen() failed\n"); 3016 rdma_destroy_id(port->id); 3017 free(port); 3018 return rc; 3019 } 3020 3021 TAILQ_FOREACH(device, &rtransport->devices, link) { 3022 if (device->context == port->id->verbs && device->is_ready) { 3023 port->device = device; 3024 break; 3025 } 3026 } 3027 if (!port->device) { 3028 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 3029 port->id->verbs); 3030 rdma_destroy_id(port->id); 3031 free(port); 3032 nvmf_rdma_rescan_devices(rtransport); 3033 return -EINVAL; 3034 } 3035 3036 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 3037 trid->traddr, trid->trsvcid); 3038 3039 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 3040 return 0; 3041 } 3042 3043 static void 3044 nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport, 3045 const struct spdk_nvme_transport_id *trid, bool need_retry) 3046 { 3047 struct spdk_nvmf_rdma_transport *rtransport; 3048 struct spdk_nvmf_rdma_port *port, *tmp; 3049 3050 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3051 3052 if (!need_retry) { 3053 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) { 3054 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 3055 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3056 free(port); 3057 } 3058 } 3059 } 3060 3061 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 3062 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 3063 SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n", 3064 port->trid->traddr, port->trid->trsvcid, need_retry); 3065 TAILQ_REMOVE(&rtransport->ports, port, link); 3066 rdma_destroy_id(port->id); 3067 port->id = NULL; 3068 port->device = NULL; 3069 if (need_retry) { 3070 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3071 } else { 3072 free(port); 3073 } 3074 break; 3075 } 3076 } 3077 } 3078 3079 static void 3080 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 3081 const struct spdk_nvme_transport_id *trid) 3082 { 3083 nvmf_rdma_stop_listen_ex(transport, trid, false); 3084 } 3085 3086 static void _nvmf_rdma_register_poller_in_group(void *c); 3087 static void _nvmf_rdma_remove_poller_in_group(void *c); 3088 3089 static bool 3090 nvmf_rdma_all_pollers_management_done(void *c) 3091 { 3092 struct poller_manage_ctx *ctx = c; 3093 int counter; 3094 3095 counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST); 3096 SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n", 3097 counter, ctx->rpoller); 3098 3099 if (counter == 0) { 3100 free((void *)ctx->inflight_op_counter); 3101 } 3102 free(ctx); 3103 3104 return counter == 0; 3105 } 3106 3107 static int 3108 nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport, 3109 struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add) 3110 { 3111 struct spdk_nvmf_rdma_poll_group *rgroup; 3112 struct spdk_nvmf_rdma_poller *rpoller; 3113 struct spdk_nvmf_poll_group *poll_group; 3114 struct poller_manage_ctx *ctx; 3115 bool found; 3116 int *inflight_counter; 3117 spdk_msg_fn do_fn; 3118 3119 *has_inflight = false; 3120 do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group; 3121 inflight_counter = calloc(1, sizeof(int)); 3122 if (!inflight_counter) { 3123 SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n"); 3124 return -ENOMEM; 3125 } 3126 3127 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3128 (*inflight_counter)++; 3129 } 3130 3131 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3132 found = false; 3133 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3134 if (rpoller->device == device) { 3135 found = true; 3136 break; 3137 } 3138 } 3139 if (found == is_add) { 3140 __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST); 3141 continue; 3142 } 3143 3144 ctx = calloc(1, sizeof(struct poller_manage_ctx)); 3145 if (!ctx) { 3146 SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n"); 3147 if (!*has_inflight) { 3148 free(inflight_counter); 3149 } 3150 return -ENOMEM; 3151 } 3152 3153 ctx->rtransport = rtransport; 3154 ctx->rgroup = rgroup; 3155 ctx->rpoller = rpoller; 3156 ctx->device = device; 3157 ctx->thread = spdk_get_thread(); 3158 ctx->inflight_op_counter = inflight_counter; 3159 *has_inflight = true; 3160 3161 poll_group = rgroup->group.group; 3162 if (poll_group->thread != spdk_get_thread()) { 3163 spdk_thread_send_msg(poll_group->thread, do_fn, ctx); 3164 } else { 3165 do_fn(ctx); 3166 } 3167 } 3168 3169 if (!*has_inflight) { 3170 free(inflight_counter); 3171 } 3172 3173 return 0; 3174 } 3175 3176 static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3177 struct spdk_nvmf_rdma_device *device); 3178 3179 static struct spdk_nvmf_rdma_device * 3180 nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport, 3181 struct ibv_context *context) 3182 { 3183 struct spdk_nvmf_rdma_device *device, *tmp_device; 3184 3185 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) { 3186 if (device->need_destroy) { 3187 continue; 3188 } 3189 3190 if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) { 3191 return device; 3192 } 3193 } 3194 3195 return NULL; 3196 } 3197 3198 static bool 3199 nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport, 3200 struct ibv_context *context) 3201 { 3202 struct spdk_nvmf_rdma_device *old_device, *new_device; 3203 int rc = 0; 3204 bool has_inflight; 3205 3206 old_device = nvmf_rdma_find_ib_device(rtransport, context); 3207 3208 if (old_device) { 3209 if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) { 3210 /* context may not have time to be cleaned when rescan. exactly one context 3211 * is valid for a device so this context must be invalid and just remove it. */ 3212 SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context); 3213 old_device->need_destroy = true; 3214 nvmf_rdma_handle_device_removal(rtransport, old_device); 3215 } 3216 return false; 3217 } 3218 3219 rc = create_ib_device(rtransport, context, &new_device); 3220 /* TODO: update transport opts. */ 3221 if (rc < 0) { 3222 SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n", 3223 ibv_get_device_name(context->device), context); 3224 return false; 3225 } 3226 3227 rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true); 3228 if (rc < 0) { 3229 SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n", 3230 ibv_get_device_name(context->device), context); 3231 return false; 3232 } 3233 3234 if (has_inflight) { 3235 new_device->is_ready = true; 3236 } 3237 3238 return true; 3239 } 3240 3241 static bool 3242 nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport) 3243 { 3244 struct spdk_nvmf_rdma_device *device; 3245 struct ibv_device **ibv_device_list = NULL; 3246 struct ibv_context **contexts = NULL; 3247 int i = 0; 3248 int num_dev = 0; 3249 bool new_create = false, has_new_device = false; 3250 struct ibv_context *tmp_verbs = NULL; 3251 3252 /* do not rescan when any device is destroying, or context may be freed when 3253 * regenerating the poll fds. 3254 */ 3255 TAILQ_FOREACH(device, &rtransport->devices, link) { 3256 if (device->need_destroy) { 3257 return false; 3258 } 3259 } 3260 3261 ibv_device_list = ibv_get_device_list(&num_dev); 3262 3263 /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be 3264 * marked as dead verbs and never be init again. So we need to make sure the 3265 * verbs is available before we call rdma_get_devices. */ 3266 if (num_dev >= 0) { 3267 for (i = 0; i < num_dev; i++) { 3268 tmp_verbs = ibv_open_device(ibv_device_list[i]); 3269 if (!tmp_verbs) { 3270 SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno); 3271 break; 3272 } 3273 if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) { 3274 SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i], 3275 tmp_verbs->device->dev_name); 3276 has_new_device = true; 3277 } 3278 ibv_close_device(tmp_verbs); 3279 } 3280 ibv_free_device_list(ibv_device_list); 3281 if (!tmp_verbs || !has_new_device) { 3282 return false; 3283 } 3284 } 3285 3286 contexts = rdma_get_devices(NULL); 3287 3288 for (i = 0; contexts && contexts[i] != NULL; i++) { 3289 new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]); 3290 } 3291 3292 if (new_create) { 3293 free_poll_fds(rtransport); 3294 generate_poll_fds(rtransport); 3295 } 3296 3297 if (contexts) { 3298 rdma_free_devices(contexts); 3299 } 3300 3301 return new_create; 3302 } 3303 3304 static bool 3305 nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport) 3306 { 3307 struct spdk_nvmf_rdma_port *port, *tmp_port; 3308 int rc = 0; 3309 bool new_create = false; 3310 3311 if (TAILQ_EMPTY(&rtransport->retry_ports)) { 3312 return false; 3313 } 3314 3315 new_create = nvmf_rdma_rescan_devices(rtransport); 3316 3317 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) { 3318 rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL); 3319 3320 TAILQ_REMOVE(&rtransport->retry_ports, port, link); 3321 if (rc) { 3322 if (new_create) { 3323 SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n", 3324 port->trid->traddr, port->trid->trsvcid, rc); 3325 } 3326 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link); 3327 break; 3328 } else { 3329 SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid); 3330 free(port); 3331 } 3332 } 3333 3334 return true; 3335 } 3336 3337 static void 3338 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 3339 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 3340 { 3341 struct spdk_nvmf_request *req, *tmp; 3342 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 3343 struct spdk_nvmf_rdma_resources *resources; 3344 3345 /* First process requests which are waiting for response to be sent */ 3346 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_send_queue, state_link, req_tmp) { 3347 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3348 break; 3349 } 3350 } 3351 3352 /* We process I/O in the data transfer pending queue at the highest priority. */ 3353 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 3354 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3355 break; 3356 } 3357 } 3358 3359 /* Then RDMA writes since reads have stronger restrictions than writes */ 3360 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 3361 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3362 break; 3363 } 3364 } 3365 3366 /* Then we handle request waiting on memory buffers. */ 3367 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 3368 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3369 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 3370 break; 3371 } 3372 } 3373 3374 resources = rqpair->resources; 3375 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 3376 rdma_req = STAILQ_FIRST(&resources->free_queue); 3377 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 3378 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 3379 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 3380 3381 if (rqpair->srq != NULL) { 3382 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 3383 rdma_req->recv->qpair->qd++; 3384 } else { 3385 rqpair->qd++; 3386 } 3387 3388 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 3389 rdma_req->state = RDMA_REQUEST_STATE_NEW; 3390 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 3391 break; 3392 } 3393 } 3394 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 3395 rqpair->poller->stat.pending_free_request++; 3396 } 3397 } 3398 3399 static void 3400 nvmf_rdma_poller_process_pending_buf_queue(struct spdk_nvmf_rdma_transport *rtransport, 3401 struct spdk_nvmf_rdma_poller *rpoller) 3402 { 3403 struct spdk_nvmf_request *req, *tmp; 3404 struct spdk_nvmf_rdma_request *rdma_req; 3405 3406 STAILQ_FOREACH_SAFE(req, &rpoller->group->group.pending_buf_queue, buf_link, tmp) { 3407 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3408 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 3409 break; 3410 } 3411 } 3412 } 3413 3414 static inline bool 3415 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 3416 { 3417 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 3418 return nvmf_rdma_is_rxe_device(device) || 3419 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 3420 } 3421 3422 static void 3423 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 3424 { 3425 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3426 struct spdk_nvmf_rdma_transport, transport); 3427 3428 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3429 3430 /* nvmf_rdma_close_qpair is not called */ 3431 if (!rqpair->to_close) { 3432 return; 3433 } 3434 3435 /* device is already destroyed and we should force destroy this qpair. */ 3436 if (rqpair->poller && rqpair->poller->need_destroy) { 3437 nvmf_rdma_qpair_destroy(rqpair); 3438 return; 3439 } 3440 3441 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 3442 if (rqpair->current_send_depth != 0) { 3443 return; 3444 } 3445 3446 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 3447 return; 3448 } 3449 3450 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 3451 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 3452 return; 3453 } 3454 3455 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 3456 3457 nvmf_rdma_qpair_destroy(rqpair); 3458 } 3459 3460 static int 3461 nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked) 3462 { 3463 struct spdk_nvmf_qpair *qpair; 3464 struct spdk_nvmf_rdma_qpair *rqpair; 3465 3466 if (evt->id == NULL) { 3467 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 3468 return -1; 3469 } 3470 3471 qpair = evt->id->context; 3472 if (qpair == NULL) { 3473 SPDK_ERRLOG("disconnect request: no active connection\n"); 3474 return -1; 3475 } 3476 3477 rdma_ack_cm_event(evt); 3478 *event_acked = true; 3479 3480 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3481 3482 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair); 3483 3484 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3485 3486 return 0; 3487 } 3488 3489 #ifdef DEBUG 3490 static const char *CM_EVENT_STR[] = { 3491 "RDMA_CM_EVENT_ADDR_RESOLVED", 3492 "RDMA_CM_EVENT_ADDR_ERROR", 3493 "RDMA_CM_EVENT_ROUTE_RESOLVED", 3494 "RDMA_CM_EVENT_ROUTE_ERROR", 3495 "RDMA_CM_EVENT_CONNECT_REQUEST", 3496 "RDMA_CM_EVENT_CONNECT_RESPONSE", 3497 "RDMA_CM_EVENT_CONNECT_ERROR", 3498 "RDMA_CM_EVENT_UNREACHABLE", 3499 "RDMA_CM_EVENT_REJECTED", 3500 "RDMA_CM_EVENT_ESTABLISHED", 3501 "RDMA_CM_EVENT_DISCONNECTED", 3502 "RDMA_CM_EVENT_DEVICE_REMOVAL", 3503 "RDMA_CM_EVENT_MULTICAST_JOIN", 3504 "RDMA_CM_EVENT_MULTICAST_ERROR", 3505 "RDMA_CM_EVENT_ADDR_CHANGE", 3506 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 3507 }; 3508 #endif /* DEBUG */ 3509 3510 static void 3511 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 3512 struct spdk_nvmf_rdma_port *port) 3513 { 3514 struct spdk_nvmf_rdma_poll_group *rgroup; 3515 struct spdk_nvmf_rdma_poller *rpoller; 3516 struct spdk_nvmf_rdma_qpair *rqpair; 3517 3518 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 3519 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3520 RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) { 3521 if (rqpair->listen_id == port->id) { 3522 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3523 } 3524 } 3525 } 3526 } 3527 } 3528 3529 static bool 3530 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 3531 struct rdma_cm_event *event) 3532 { 3533 const struct spdk_nvme_transport_id *trid; 3534 struct spdk_nvmf_rdma_port *port; 3535 struct spdk_nvmf_rdma_transport *rtransport; 3536 bool event_acked = false; 3537 3538 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3539 TAILQ_FOREACH(port, &rtransport->ports, link) { 3540 if (port->id == event->id) { 3541 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 3542 rdma_ack_cm_event(event); 3543 event_acked = true; 3544 trid = port->trid; 3545 break; 3546 } 3547 } 3548 3549 if (event_acked) { 3550 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 3551 3552 nvmf_rdma_stop_listen(transport, trid); 3553 nvmf_rdma_listen(transport, trid, NULL); 3554 } 3555 3556 return event_acked; 3557 } 3558 3559 static void 3560 nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport, 3561 struct spdk_nvmf_rdma_device *device) 3562 { 3563 struct spdk_nvmf_rdma_port *port, *port_tmp; 3564 int rc; 3565 bool has_inflight; 3566 3567 rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false); 3568 if (rc) { 3569 SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc); 3570 return; 3571 } 3572 3573 if (!has_inflight) { 3574 /* no pollers, destroy the device */ 3575 device->ready_to_destroy = true; 3576 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport); 3577 } 3578 3579 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 3580 if (port->device == device) { 3581 SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n", 3582 port->trid->traddr, 3583 port->trid->trsvcid, 3584 ibv_get_device_name(port->device->context->device)); 3585 3586 /* keep NVMF listener and only destroy structures of the 3587 * RDMA transport. when the device comes back we can retry listening 3588 * and the application's workflow will not be interrupted. 3589 */ 3590 nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true); 3591 } 3592 } 3593 } 3594 3595 static void 3596 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 3597 struct rdma_cm_event *event) 3598 { 3599 struct spdk_nvmf_rdma_port *port, *tmp_port; 3600 struct spdk_nvmf_rdma_transport *rtransport; 3601 3602 port = event->id->context; 3603 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3604 3605 rdma_ack_cm_event(event); 3606 3607 /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive 3608 * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that 3609 * we are handling a port event here. 3610 */ 3611 TAILQ_FOREACH(tmp_port, &rtransport->ports, link) { 3612 if (port == tmp_port && port->device && !port->device->need_destroy) { 3613 port->device->need_destroy = true; 3614 nvmf_rdma_handle_device_removal(rtransport, port->device); 3615 } 3616 } 3617 } 3618 3619 static void 3620 nvmf_process_cm_events(struct spdk_nvmf_transport *transport, uint32_t max_events) 3621 { 3622 struct spdk_nvmf_rdma_transport *rtransport; 3623 struct rdma_cm_event *event; 3624 uint32_t i; 3625 int rc; 3626 bool event_acked; 3627 3628 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3629 3630 if (rtransport->event_channel == NULL) { 3631 return; 3632 } 3633 3634 for (i = 0; i < max_events; i++) { 3635 event_acked = false; 3636 rc = rdma_get_cm_event(rtransport->event_channel, &event); 3637 if (rc) { 3638 if (errno != EAGAIN && errno != EWOULDBLOCK) { 3639 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 3640 } 3641 break; 3642 } 3643 3644 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 3645 3646 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 3647 3648 switch (event->event) { 3649 case RDMA_CM_EVENT_ADDR_RESOLVED: 3650 case RDMA_CM_EVENT_ADDR_ERROR: 3651 case RDMA_CM_EVENT_ROUTE_RESOLVED: 3652 case RDMA_CM_EVENT_ROUTE_ERROR: 3653 /* No action required. The target never attempts to resolve routes. */ 3654 break; 3655 case RDMA_CM_EVENT_CONNECT_REQUEST: 3656 rc = nvmf_rdma_connect(transport, event); 3657 if (rc < 0) { 3658 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 3659 break; 3660 } 3661 break; 3662 case RDMA_CM_EVENT_CONNECT_RESPONSE: 3663 /* The target never initiates a new connection. So this will not occur. */ 3664 break; 3665 case RDMA_CM_EVENT_CONNECT_ERROR: 3666 /* Can this happen? The docs say it can, but not sure what causes it. */ 3667 break; 3668 case RDMA_CM_EVENT_UNREACHABLE: 3669 case RDMA_CM_EVENT_REJECTED: 3670 /* These only occur on the client side. */ 3671 break; 3672 case RDMA_CM_EVENT_ESTABLISHED: 3673 /* TODO: Should we be waiting for this event anywhere? */ 3674 break; 3675 case RDMA_CM_EVENT_DISCONNECTED: 3676 rc = nvmf_rdma_disconnect(event, &event_acked); 3677 if (rc < 0) { 3678 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3679 break; 3680 } 3681 break; 3682 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3683 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3684 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3685 * Once these events are sent to SPDK, we should release all IB resources and 3686 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3687 * ibv_destroy* functions to release user space memory allocated by IB. All kernel 3688 * resources are already cleaned. */ 3689 if (event->id->qp) { 3690 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3691 * corresponding qpair. Otherwise the event refers to a listening device. */ 3692 rc = nvmf_rdma_disconnect(event, &event_acked); 3693 if (rc < 0) { 3694 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3695 break; 3696 } 3697 } else { 3698 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3699 event_acked = true; 3700 } 3701 break; 3702 case RDMA_CM_EVENT_MULTICAST_JOIN: 3703 case RDMA_CM_EVENT_MULTICAST_ERROR: 3704 /* Multicast is not used */ 3705 break; 3706 case RDMA_CM_EVENT_ADDR_CHANGE: 3707 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3708 break; 3709 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3710 /* For now, do nothing. The target never re-uses queue pairs. */ 3711 break; 3712 default: 3713 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3714 break; 3715 } 3716 if (!event_acked) { 3717 rdma_ack_cm_event(event); 3718 } 3719 } 3720 } 3721 3722 static void 3723 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3724 { 3725 rqpair->last_wqe_reached = true; 3726 nvmf_rdma_destroy_drained_qpair(rqpair); 3727 } 3728 3729 static void 3730 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3731 { 3732 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3733 3734 if (event_ctx->rqpair) { 3735 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3736 if (event_ctx->cb_fn) { 3737 event_ctx->cb_fn(event_ctx->rqpair); 3738 } 3739 } 3740 free(event_ctx); 3741 } 3742 3743 static int 3744 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3745 spdk_nvmf_rdma_qpair_ibv_event fn) 3746 { 3747 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3748 struct spdk_thread *thr = NULL; 3749 int rc; 3750 3751 if (rqpair->qpair.group) { 3752 thr = rqpair->qpair.group->thread; 3753 } else if (rqpair->destruct_channel) { 3754 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3755 } 3756 3757 if (!thr) { 3758 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3759 return -EINVAL; 3760 } 3761 3762 ctx = calloc(1, sizeof(*ctx)); 3763 if (!ctx) { 3764 return -ENOMEM; 3765 } 3766 3767 ctx->rqpair = rqpair; 3768 ctx->cb_fn = fn; 3769 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3770 3771 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3772 if (rc) { 3773 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3774 free(ctx); 3775 } 3776 3777 return rc; 3778 } 3779 3780 static int 3781 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3782 { 3783 int rc; 3784 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3785 struct ibv_async_event event; 3786 3787 rc = ibv_get_async_event(device->context, &event); 3788 3789 if (rc) { 3790 /* In non-blocking mode -1 means there are no events available */ 3791 return rc; 3792 } 3793 3794 switch (event.event_type) { 3795 case IBV_EVENT_QP_FATAL: 3796 case IBV_EVENT_QP_LAST_WQE_REACHED: 3797 case IBV_EVENT_QP_REQ_ERR: 3798 case IBV_EVENT_QP_ACCESS_ERR: 3799 case IBV_EVENT_COMM_EST: 3800 case IBV_EVENT_PATH_MIG: 3801 case IBV_EVENT_PATH_MIG_ERR: 3802 rqpair = event.element.qp->qp_context; 3803 if (!rqpair) { 3804 /* Any QP event for NVMe-RDMA initiator may be returned. */ 3805 SPDK_NOTICELOG("Async QP event for unknown QP: %s\n", 3806 ibv_event_type_str(event.event_type)); 3807 break; 3808 } 3809 3810 switch (event.event_type) { 3811 case IBV_EVENT_QP_FATAL: 3812 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3813 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3814 (uintptr_t)rqpair, event.event_type); 3815 rqpair->ibv_in_error_state = true; 3816 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 3817 break; 3818 case IBV_EVENT_QP_LAST_WQE_REACHED: 3819 /* This event only occurs for shared receive queues. */ 3820 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3821 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3822 if (rc) { 3823 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3824 rqpair->last_wqe_reached = true; 3825 } 3826 break; 3827 case IBV_EVENT_QP_REQ_ERR: 3828 case IBV_EVENT_QP_ACCESS_ERR: 3829 case IBV_EVENT_COMM_EST: 3830 case IBV_EVENT_PATH_MIG: 3831 case IBV_EVENT_PATH_MIG_ERR: 3832 SPDK_NOTICELOG("Async QP event: %s\n", 3833 ibv_event_type_str(event.event_type)); 3834 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3835 (uintptr_t)rqpair, event.event_type); 3836 rqpair->ibv_in_error_state = true; 3837 break; 3838 default: 3839 break; 3840 } 3841 break; 3842 case IBV_EVENT_DEVICE_FATAL: 3843 SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n", 3844 ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device); 3845 device->need_destroy = true; 3846 break; 3847 case IBV_EVENT_CQ_ERR: 3848 case IBV_EVENT_PORT_ACTIVE: 3849 case IBV_EVENT_PORT_ERR: 3850 case IBV_EVENT_LID_CHANGE: 3851 case IBV_EVENT_PKEY_CHANGE: 3852 case IBV_EVENT_SM_CHANGE: 3853 case IBV_EVENT_SRQ_ERR: 3854 case IBV_EVENT_SRQ_LIMIT_REACHED: 3855 case IBV_EVENT_CLIENT_REREGISTER: 3856 case IBV_EVENT_GID_CHANGE: 3857 case IBV_EVENT_SQ_DRAINED: 3858 default: 3859 SPDK_NOTICELOG("Async event: %s\n", 3860 ibv_event_type_str(event.event_type)); 3861 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3862 break; 3863 } 3864 ibv_ack_async_event(&event); 3865 3866 return 0; 3867 } 3868 3869 static void 3870 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3871 { 3872 int rc = 0; 3873 uint32_t i = 0; 3874 3875 for (i = 0; i < max_events; i++) { 3876 rc = nvmf_process_ib_event(device); 3877 if (rc) { 3878 break; 3879 } 3880 } 3881 3882 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3883 } 3884 3885 static int 3886 nvmf_rdma_accept(void *ctx) 3887 { 3888 int nfds, i = 0; 3889 struct spdk_nvmf_transport *transport = ctx; 3890 struct spdk_nvmf_rdma_transport *rtransport; 3891 struct spdk_nvmf_rdma_device *device, *tmp; 3892 uint32_t count; 3893 short revents; 3894 bool do_retry; 3895 3896 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3897 do_retry = nvmf_rdma_retry_listen_port(rtransport); 3898 3899 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3900 3901 if (nfds <= 0) { 3902 return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3903 } 3904 3905 /* The first poll descriptor is RDMA CM event */ 3906 if (rtransport->poll_fds[i++].revents & POLLIN) { 3907 nvmf_process_cm_events(transport, NVMF_RDMA_MAX_EVENTS_PER_POLL); 3908 nfds--; 3909 } 3910 3911 if (nfds == 0) { 3912 return SPDK_POLLER_BUSY; 3913 } 3914 3915 /* Second and subsequent poll descriptors are IB async events */ 3916 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3917 revents = rtransport->poll_fds[i++].revents; 3918 if (revents & POLLIN) { 3919 if (spdk_likely(!device->need_destroy)) { 3920 nvmf_process_ib_events(device, NVMF_RDMA_MAX_EVENTS_PER_POLL); 3921 if (spdk_unlikely(device->need_destroy)) { 3922 nvmf_rdma_handle_device_removal(rtransport, device); 3923 } 3924 } 3925 nfds--; 3926 } else if (revents & POLLNVAL || revents & POLLHUP) { 3927 SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device); 3928 nfds--; 3929 } 3930 } 3931 /* check all flagged fd's have been served */ 3932 assert(nfds == 0); 3933 3934 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 3935 } 3936 3937 static void 3938 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3939 struct spdk_nvmf_ctrlr_data *cdata) 3940 { 3941 cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD; 3942 3943 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3944 since in-capsule data only works with NVME drives that support SGL memory layout */ 3945 if (transport->opts.dif_insert_or_strip) { 3946 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3947 } 3948 3949 if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) { 3950 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule" 3951 " data is greater than 4KiB.\n"); 3952 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux " 3953 "kernel between versions 5.4 and 5.12 data corruption may occur for " 3954 "writes that are not a multiple of 4KiB in size.\n"); 3955 } 3956 } 3957 3958 static void 3959 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3960 struct spdk_nvme_transport_id *trid, 3961 struct spdk_nvmf_discovery_log_page_entry *entry) 3962 { 3963 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3964 entry->adrfam = trid->adrfam; 3965 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3966 3967 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3968 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3969 3970 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3971 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3972 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3973 } 3974 3975 static int 3976 nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport, 3977 struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device, 3978 struct spdk_nvmf_rdma_poller **out_poller) 3979 { 3980 struct spdk_nvmf_rdma_poller *poller; 3981 struct spdk_rdma_provider_srq_init_attr srq_init_attr; 3982 struct spdk_nvmf_rdma_resource_opts opts; 3983 int num_cqe; 3984 3985 poller = calloc(1, sizeof(*poller)); 3986 if (!poller) { 3987 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3988 return -1; 3989 } 3990 3991 poller->device = device; 3992 poller->group = rgroup; 3993 *out_poller = poller; 3994 3995 RB_INIT(&poller->qpairs); 3996 STAILQ_INIT(&poller->qpairs_pending_send); 3997 STAILQ_INIT(&poller->qpairs_pending_recv); 3998 3999 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 4000 SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup); 4001 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 4002 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) { 4003 SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n", 4004 rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr); 4005 } 4006 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr); 4007 4008 device->num_srq++; 4009 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 4010 srq_init_attr.pd = device->pd; 4011 srq_init_attr.stats = &poller->stat.qp_stats.recv; 4012 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 4013 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 4014 poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr); 4015 if (!poller->srq) { 4016 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 4017 return -1; 4018 } 4019 4020 opts.qp = poller->srq; 4021 opts.map = device->map; 4022 opts.qpair = NULL; 4023 opts.shared = true; 4024 opts.max_queue_depth = poller->max_srq_depth; 4025 opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size; 4026 4027 poller->resources = nvmf_rdma_resources_create(&opts); 4028 if (!poller->resources) { 4029 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 4030 return -1; 4031 } 4032 } 4033 4034 /* 4035 * When using an srq, we can limit the completion queue at startup. 4036 * The following formula represents the calculation: 4037 * num_cqe = num_recv + num_data_wr + num_send_wr. 4038 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 4039 */ 4040 if (poller->srq) { 4041 num_cqe = poller->max_srq_depth * 3; 4042 } else { 4043 num_cqe = rtransport->rdma_opts.num_cqe; 4044 } 4045 4046 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 4047 if (!poller->cq) { 4048 SPDK_ERRLOG("Unable to create completion queue\n"); 4049 return -1; 4050 } 4051 poller->num_cqe = num_cqe; 4052 return 0; 4053 } 4054 4055 static void 4056 _nvmf_rdma_register_poller_in_group(void *c) 4057 { 4058 struct spdk_nvmf_rdma_poller *poller; 4059 struct poller_manage_ctx *ctx = c; 4060 struct spdk_nvmf_rdma_device *device; 4061 int rc; 4062 4063 rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller); 4064 if (rc < 0 && poller) { 4065 nvmf_rdma_poller_destroy(poller); 4066 } 4067 4068 device = ctx->device; 4069 if (nvmf_rdma_all_pollers_management_done(ctx)) { 4070 device->is_ready = true; 4071 } 4072 } 4073 4074 static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 4075 4076 static struct spdk_nvmf_transport_poll_group * 4077 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport, 4078 struct spdk_nvmf_poll_group *group) 4079 { 4080 struct spdk_nvmf_rdma_transport *rtransport; 4081 struct spdk_nvmf_rdma_poll_group *rgroup; 4082 struct spdk_nvmf_rdma_poller *poller; 4083 struct spdk_nvmf_rdma_device *device; 4084 int rc; 4085 4086 if (spdk_interrupt_mode_is_enabled()) { 4087 SPDK_ERRLOG("RDMA transport does not support interrupt mode\n"); 4088 return NULL; 4089 } 4090 4091 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 4092 4093 rgroup = calloc(1, sizeof(*rgroup)); 4094 if (!rgroup) { 4095 return NULL; 4096 } 4097 4098 TAILQ_INIT(&rgroup->pollers); 4099 4100 TAILQ_FOREACH(device, &rtransport->devices, link) { 4101 rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller); 4102 if (rc < 0) { 4103 nvmf_rdma_poll_group_destroy(&rgroup->group); 4104 return NULL; 4105 } 4106 } 4107 4108 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 4109 if (rtransport->conn_sched.next_admin_pg == NULL) { 4110 rtransport->conn_sched.next_admin_pg = rgroup; 4111 rtransport->conn_sched.next_io_pg = rgroup; 4112 } 4113 4114 return &rgroup->group; 4115 } 4116 4117 static uint32_t 4118 nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg) 4119 { 4120 uint32_t count; 4121 4122 /* Just assume that unassociated qpairs will eventually be io 4123 * qpairs. This is close enough for the use cases for this 4124 * function. 4125 */ 4126 pthread_mutex_lock(&pg->mutex); 4127 count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs; 4128 pthread_mutex_unlock(&pg->mutex); 4129 4130 return count; 4131 } 4132 4133 static struct spdk_nvmf_transport_poll_group * 4134 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 4135 { 4136 struct spdk_nvmf_rdma_transport *rtransport; 4137 struct spdk_nvmf_rdma_poll_group **pg; 4138 struct spdk_nvmf_transport_poll_group *result; 4139 uint32_t count; 4140 4141 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4142 4143 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 4144 return NULL; 4145 } 4146 4147 if (qpair->qid == 0) { 4148 pg = &rtransport->conn_sched.next_admin_pg; 4149 } else { 4150 struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current; 4151 uint32_t min_value; 4152 4153 pg = &rtransport->conn_sched.next_io_pg; 4154 pg_min = *pg; 4155 pg_start = *pg; 4156 pg_current = *pg; 4157 min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 4158 4159 while (1) { 4160 count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group); 4161 4162 if (count < min_value) { 4163 min_value = count; 4164 pg_min = pg_current; 4165 } 4166 4167 pg_current = TAILQ_NEXT(pg_current, link); 4168 if (pg_current == NULL) { 4169 pg_current = TAILQ_FIRST(&rtransport->poll_groups); 4170 } 4171 4172 if (pg_current == pg_start || min_value == 0) { 4173 break; 4174 } 4175 } 4176 *pg = pg_min; 4177 } 4178 4179 assert(*pg != NULL); 4180 4181 result = &(*pg)->group; 4182 4183 *pg = TAILQ_NEXT(*pg, link); 4184 if (*pg == NULL) { 4185 *pg = TAILQ_FIRST(&rtransport->poll_groups); 4186 } 4187 4188 return result; 4189 } 4190 4191 static void 4192 nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller) 4193 { 4194 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 4195 int rc; 4196 4197 TAILQ_REMOVE(&poller->group->pollers, poller, link); 4198 RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) { 4199 nvmf_rdma_qpair_destroy(qpair); 4200 } 4201 4202 if (poller->srq) { 4203 if (poller->resources) { 4204 nvmf_rdma_resources_destroy(poller->resources); 4205 } 4206 spdk_rdma_provider_srq_destroy(poller->srq); 4207 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 4208 } 4209 4210 if (poller->cq) { 4211 rc = ibv_destroy_cq(poller->cq); 4212 if (rc != 0) { 4213 SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno)); 4214 } 4215 } 4216 4217 if (poller->destroy_cb) { 4218 poller->destroy_cb(poller->destroy_cb_ctx); 4219 poller->destroy_cb = NULL; 4220 } 4221 4222 free(poller); 4223 } 4224 4225 static void 4226 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 4227 { 4228 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 4229 struct spdk_nvmf_rdma_poller *poller, *tmp; 4230 struct spdk_nvmf_rdma_transport *rtransport; 4231 4232 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4233 if (!rgroup) { 4234 return; 4235 } 4236 4237 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 4238 nvmf_rdma_poller_destroy(poller); 4239 } 4240 4241 if (rgroup->group.transport == NULL) { 4242 /* Transport can be NULL when nvmf_rdma_poll_group_create() 4243 * calls this function directly in a failure path. */ 4244 free(rgroup); 4245 return; 4246 } 4247 4248 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 4249 4250 next_rgroup = TAILQ_NEXT(rgroup, link); 4251 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 4252 if (next_rgroup == NULL) { 4253 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 4254 } 4255 if (rtransport->conn_sched.next_admin_pg == rgroup) { 4256 rtransport->conn_sched.next_admin_pg = next_rgroup; 4257 } 4258 if (rtransport->conn_sched.next_io_pg == rgroup) { 4259 rtransport->conn_sched.next_io_pg = next_rgroup; 4260 } 4261 4262 free(rgroup); 4263 } 4264 4265 static void 4266 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 4267 { 4268 if (rqpair->cm_id != NULL) { 4269 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 4270 } 4271 } 4272 4273 static int 4274 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 4275 struct spdk_nvmf_qpair *qpair) 4276 { 4277 struct spdk_nvmf_rdma_poll_group *rgroup; 4278 struct spdk_nvmf_rdma_qpair *rqpair; 4279 struct spdk_nvmf_rdma_device *device; 4280 struct spdk_nvmf_rdma_poller *poller; 4281 int rc; 4282 4283 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4284 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4285 4286 device = rqpair->device; 4287 4288 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 4289 if (poller->device == device) { 4290 break; 4291 } 4292 } 4293 4294 if (!poller) { 4295 SPDK_ERRLOG("No poller found for device.\n"); 4296 return -1; 4297 } 4298 4299 if (poller->need_destroy) { 4300 SPDK_ERRLOG("Poller is destroying.\n"); 4301 return -1; 4302 } 4303 4304 rqpair->poller = poller; 4305 rqpair->srq = rqpair->poller->srq; 4306 4307 rc = nvmf_rdma_qpair_initialize(qpair); 4308 if (rc < 0) { 4309 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 4310 rqpair->poller = NULL; 4311 rqpair->srq = NULL; 4312 return -1; 4313 } 4314 4315 RB_INSERT(qpairs_tree, &poller->qpairs, rqpair); 4316 4317 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 4318 if (rc) { 4319 /* Try to reject, but we probably can't */ 4320 nvmf_rdma_qpair_reject_connection(rqpair); 4321 return -1; 4322 } 4323 4324 return 0; 4325 } 4326 4327 static int 4328 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 4329 struct spdk_nvmf_qpair *qpair) 4330 { 4331 struct spdk_nvmf_rdma_qpair *rqpair; 4332 4333 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4334 assert(group->transport->tgt != NULL); 4335 4336 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 4337 4338 if (!rqpair->destruct_channel) { 4339 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 4340 return 0; 4341 } 4342 4343 /* Sanity check that we get io_channel on the correct thread */ 4344 if (qpair->group) { 4345 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 4346 } 4347 4348 return 0; 4349 } 4350 4351 static int 4352 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 4353 { 4354 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 4355 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4356 struct spdk_nvmf_rdma_transport, transport); 4357 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4358 struct spdk_nvmf_rdma_qpair, qpair); 4359 4360 /* 4361 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 4362 * needs to be returned to the shared receive queue or the poll group will eventually be 4363 * starved of RECV structures. 4364 */ 4365 if (rqpair->srq && rdma_req->recv) { 4366 int rc; 4367 struct ibv_recv_wr *bad_recv_wr; 4368 4369 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 4370 rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 4371 if (rc) { 4372 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 4373 } 4374 } 4375 4376 _nvmf_rdma_request_free(rdma_req, rtransport); 4377 return 0; 4378 } 4379 4380 static int 4381 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 4382 { 4383 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 4384 struct spdk_nvmf_rdma_transport, transport); 4385 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 4386 struct spdk_nvmf_rdma_request, req); 4387 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 4388 struct spdk_nvmf_rdma_qpair, qpair); 4389 4390 if (spdk_unlikely(rqpair->ibv_in_error_state)) { 4391 /* The connection is dead. Move the request directly to the completed state. */ 4392 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4393 } else { 4394 /* The connection is alive, so process the request as normal */ 4395 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 4396 } 4397 4398 nvmf_rdma_request_process(rtransport, rdma_req); 4399 4400 return 0; 4401 } 4402 4403 static void 4404 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 4405 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 4406 { 4407 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4408 4409 rqpair->to_close = true; 4410 4411 /* This happens only when the qpair is disconnected before 4412 * it is added to the poll group. Since there is no poll group, 4413 * the RDMA qp has not been initialized yet and the RDMA CM 4414 * event has not yet been acknowledged, so we need to reject it. 4415 */ 4416 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 4417 nvmf_rdma_qpair_reject_connection(rqpair); 4418 nvmf_rdma_qpair_destroy(rqpair); 4419 return; 4420 } 4421 4422 if (rqpair->rdma_qp) { 4423 spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp); 4424 } 4425 4426 nvmf_rdma_destroy_drained_qpair(rqpair); 4427 4428 if (cb_fn) { 4429 cb_fn(cb_arg); 4430 } 4431 } 4432 4433 static struct spdk_nvmf_rdma_qpair * 4434 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 4435 { 4436 struct spdk_nvmf_rdma_qpair find; 4437 4438 find.qp_num = wc->qp_num; 4439 4440 return RB_FIND(qpairs_tree, &rpoller->qpairs, &find); 4441 } 4442 4443 #ifdef DEBUG 4444 static int 4445 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 4446 { 4447 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 4448 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 4449 } 4450 #endif 4451 4452 static void 4453 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 4454 int rc) 4455 { 4456 struct spdk_nvmf_rdma_recv *rdma_recv; 4457 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4458 4459 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 4460 while (bad_recv_wr != NULL) { 4461 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 4462 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4463 4464 rdma_recv->qpair->current_recv_depth++; 4465 bad_recv_wr = bad_recv_wr->next; 4466 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 4467 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair); 4468 } 4469 } 4470 4471 static void 4472 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 4473 { 4474 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 4475 while (bad_recv_wr != NULL) { 4476 bad_recv_wr = bad_recv_wr->next; 4477 rqpair->current_recv_depth++; 4478 } 4479 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4480 } 4481 4482 static void 4483 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 4484 struct spdk_nvmf_rdma_poller *rpoller) 4485 { 4486 struct spdk_nvmf_rdma_qpair *rqpair; 4487 struct ibv_recv_wr *bad_recv_wr; 4488 int rc; 4489 4490 if (rpoller->srq) { 4491 rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 4492 if (spdk_unlikely(rc)) { 4493 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 4494 } 4495 } else { 4496 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 4497 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 4498 rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 4499 if (spdk_unlikely(rc)) { 4500 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 4501 } 4502 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 4503 } 4504 } 4505 } 4506 4507 static void 4508 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 4509 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 4510 { 4511 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 4512 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 4513 4514 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 4515 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 4516 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 4517 assert(rqpair->current_send_depth > 0); 4518 rqpair->current_send_depth--; 4519 switch (bad_rdma_wr->type) { 4520 case RDMA_WR_TYPE_DATA: 4521 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4522 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 4523 assert(rqpair->current_read_depth > 0); 4524 rqpair->current_read_depth--; 4525 } 4526 break; 4527 case RDMA_WR_TYPE_SEND: 4528 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4529 break; 4530 default: 4531 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 4532 prev_rdma_req = cur_rdma_req; 4533 continue; 4534 } 4535 4536 if (prev_rdma_req == cur_rdma_req) { 4537 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 4538 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 4539 continue; 4540 } 4541 4542 switch (cur_rdma_req->state) { 4543 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4544 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 4545 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, cur_rdma_req, state_link); 4546 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 4547 break; 4548 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 4549 case RDMA_REQUEST_STATE_COMPLETING: 4550 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4551 break; 4552 default: 4553 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 4554 cur_rdma_req->state, rqpair); 4555 continue; 4556 } 4557 4558 nvmf_rdma_request_process(rtransport, cur_rdma_req); 4559 prev_rdma_req = cur_rdma_req; 4560 } 4561 4562 if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) { 4563 /* Disconnect the connection. */ 4564 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4565 } 4566 4567 } 4568 4569 static void 4570 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 4571 struct spdk_nvmf_rdma_poller *rpoller) 4572 { 4573 struct spdk_nvmf_rdma_qpair *rqpair; 4574 struct ibv_send_wr *bad_wr = NULL; 4575 int rc; 4576 4577 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 4578 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 4579 rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 4580 4581 /* bad wr always points to the first wr that failed. */ 4582 if (spdk_unlikely(rc)) { 4583 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 4584 } 4585 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 4586 } 4587 } 4588 4589 static const char * 4590 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 4591 { 4592 switch (wr_type) { 4593 case RDMA_WR_TYPE_RECV: 4594 return "RECV"; 4595 case RDMA_WR_TYPE_SEND: 4596 return "SEND"; 4597 case RDMA_WR_TYPE_DATA: 4598 return "DATA"; 4599 default: 4600 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 4601 SPDK_UNREACHABLE(); 4602 } 4603 } 4604 4605 static inline void 4606 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 4607 { 4608 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 4609 4610 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 4611 /* If qpair is in ERR state, we will receive completions for all posted and not completed 4612 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 4613 SPDK_DEBUGLOG(rdma, 4614 "Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n", 4615 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id, 4616 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4617 } else { 4618 SPDK_ERRLOG("Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n", 4619 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id, 4620 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 4621 } 4622 } 4623 4624 static int 4625 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 4626 struct spdk_nvmf_rdma_poller *rpoller) 4627 { 4628 struct ibv_wc wc[32]; 4629 struct spdk_nvmf_rdma_wr *rdma_wr; 4630 struct spdk_nvmf_rdma_request *rdma_req; 4631 struct spdk_nvmf_rdma_recv *rdma_recv; 4632 struct spdk_nvmf_rdma_qpair *rqpair, *tmp_rqpair; 4633 int reaped, i; 4634 int count = 0; 4635 int rc; 4636 bool error = false; 4637 uint64_t poll_tsc = spdk_get_ticks(); 4638 4639 if (spdk_unlikely(rpoller->need_destroy)) { 4640 /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not 4641 * be called because we cannot poll anything from cq. So we call that here to force 4642 * destroy the qpair after to_close turning true. 4643 */ 4644 RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) { 4645 nvmf_rdma_destroy_drained_qpair(rqpair); 4646 } 4647 return 0; 4648 } 4649 4650 /* Poll for completing operations. */ 4651 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 4652 if (spdk_unlikely(reaped < 0)) { 4653 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 4654 errno, spdk_strerror(errno)); 4655 return -1; 4656 } else if (reaped == 0) { 4657 rpoller->stat.idle_polls++; 4658 } 4659 4660 rpoller->stat.polls++; 4661 rpoller->stat.completions += reaped; 4662 4663 for (i = 0; i < reaped; i++) { 4664 4665 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 4666 4667 switch (rdma_wr->type) { 4668 case RDMA_WR_TYPE_SEND: 4669 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr); 4670 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4671 4672 if (spdk_likely(!wc[i].status)) { 4673 count++; 4674 assert(wc[i].opcode == IBV_WC_SEND); 4675 assert(nvmf_rdma_req_is_completing(rdma_req)); 4676 } 4677 4678 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4679 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 4680 assert(rqpair->current_send_depth >= (uint32_t)rdma_req->num_outstanding_data_wr + 1); 4681 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 4682 rdma_req->num_outstanding_data_wr = 0; 4683 4684 nvmf_rdma_request_process(rtransport, rdma_req); 4685 break; 4686 case RDMA_WR_TYPE_RECV: 4687 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 4688 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 4689 if (rpoller->srq != NULL) { 4690 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 4691 /* It is possible that there are still some completions for destroyed QP 4692 * associated with SRQ. We just ignore these late completions and re-post 4693 * receive WRs back to SRQ. 4694 */ 4695 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 4696 struct ibv_recv_wr *bad_wr; 4697 4698 rdma_recv->wr.next = NULL; 4699 spdk_rdma_provider_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 4700 rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 4701 if (rc) { 4702 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 4703 } 4704 continue; 4705 } 4706 } 4707 rqpair = rdma_recv->qpair; 4708 4709 assert(rqpair != NULL); 4710 if (spdk_likely(!wc[i].status)) { 4711 assert(wc[i].opcode == IBV_WC_RECV); 4712 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 4713 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4714 break; 4715 } 4716 } 4717 4718 rdma_recv->wr.next = NULL; 4719 rqpair->current_recv_depth++; 4720 rdma_recv->receive_tsc = poll_tsc; 4721 rpoller->stat.requests++; 4722 STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link); 4723 rqpair->qpair.queue_depth++; 4724 break; 4725 case RDMA_WR_TYPE_DATA: 4726 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr); 4727 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 4728 4729 assert(rdma_req->num_outstanding_data_wr > 0); 4730 4731 rqpair->current_send_depth--; 4732 rdma_req->num_outstanding_data_wr--; 4733 if (spdk_likely(!wc[i].status)) { 4734 assert(wc[i].opcode == IBV_WC_RDMA_READ); 4735 rqpair->current_read_depth--; 4736 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 4737 if (rdma_req->num_outstanding_data_wr == 0) { 4738 if (rdma_req->num_remaining_data_wr) { 4739 /* Only part of RDMA_READ operations was submitted, process the rest */ 4740 nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport); 4741 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 4742 nvmf_rdma_request_process(rtransport, rdma_req); 4743 break; 4744 } 4745 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 4746 nvmf_rdma_request_process(rtransport, rdma_req); 4747 } 4748 } else { 4749 /* If the data transfer fails still force the queue into the error state, 4750 * if we were performing an RDMA_READ, we need to force the request into a 4751 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 4752 * case, we should wait for the SEND to complete. */ 4753 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 4754 rqpair->current_read_depth--; 4755 if (rdma_req->num_outstanding_data_wr == 0) { 4756 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 4757 } 4758 } 4759 } 4760 break; 4761 default: 4762 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 4763 continue; 4764 } 4765 4766 /* Handle error conditions */ 4767 if (spdk_unlikely(wc[i].status)) { 4768 rqpair->ibv_in_error_state = true; 4769 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 4770 4771 error = true; 4772 4773 if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) { 4774 /* Disconnect the connection. */ 4775 spdk_nvmf_qpair_disconnect(&rqpair->qpair); 4776 } else { 4777 nvmf_rdma_destroy_drained_qpair(rqpair); 4778 } 4779 continue; 4780 } 4781 4782 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 4783 4784 if (spdk_unlikely(!spdk_nvmf_qpair_is_active(&rqpair->qpair))) { 4785 nvmf_rdma_destroy_drained_qpair(rqpair); 4786 } 4787 } 4788 4789 if (spdk_unlikely(error == true)) { 4790 return -1; 4791 } 4792 4793 if (reaped == 0) { 4794 /* In some cases we may not receive any CQE but we still may have pending IO requests waiting for 4795 * a resource (e.g. a WR from the data_wr_pool). 4796 * We need to start processing of such requests if no CQE reaped */ 4797 nvmf_rdma_poller_process_pending_buf_queue(rtransport, rpoller); 4798 } 4799 4800 /* submit outstanding work requests. */ 4801 _poller_submit_recvs(rtransport, rpoller); 4802 _poller_submit_sends(rtransport, rpoller); 4803 4804 return count; 4805 } 4806 4807 static void 4808 _nvmf_rdma_remove_destroyed_device(void *c) 4809 { 4810 struct spdk_nvmf_rdma_transport *rtransport = c; 4811 struct spdk_nvmf_rdma_device *device, *device_tmp; 4812 int rc; 4813 4814 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 4815 if (device->ready_to_destroy) { 4816 destroy_ib_device(rtransport, device); 4817 } 4818 } 4819 4820 free_poll_fds(rtransport); 4821 rc = generate_poll_fds(rtransport); 4822 /* cannot handle fd allocation error here */ 4823 if (rc != 0) { 4824 SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n"); 4825 } 4826 } 4827 4828 static void 4829 _nvmf_rdma_remove_poller_in_group_cb(void *c) 4830 { 4831 struct poller_manage_ctx *ctx = c; 4832 struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport; 4833 struct spdk_nvmf_rdma_device *device = ctx->device; 4834 struct spdk_thread *thread = ctx->thread; 4835 4836 if (nvmf_rdma_all_pollers_management_done(c)) { 4837 /* destroy device when last poller is destroyed */ 4838 device->ready_to_destroy = true; 4839 spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport); 4840 } 4841 } 4842 4843 static void 4844 _nvmf_rdma_remove_poller_in_group(void *c) 4845 { 4846 struct poller_manage_ctx *ctx = c; 4847 4848 ctx->rpoller->need_destroy = true; 4849 ctx->rpoller->destroy_cb_ctx = ctx; 4850 ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb; 4851 4852 /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */ 4853 if (RB_EMPTY(&ctx->rpoller->qpairs)) { 4854 nvmf_rdma_poller_destroy(ctx->rpoller); 4855 } 4856 } 4857 4858 static int 4859 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 4860 { 4861 struct spdk_nvmf_rdma_transport *rtransport; 4862 struct spdk_nvmf_rdma_poll_group *rgroup; 4863 struct spdk_nvmf_rdma_poller *rpoller, *tmp; 4864 int count = 0, rc, rc2 = 0; 4865 4866 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 4867 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4868 4869 TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) { 4870 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4871 if (spdk_unlikely(rc < 0)) { 4872 if (rc2 == 0) { 4873 rc2 = rc; 4874 } 4875 continue; 4876 } 4877 count += rc; 4878 } 4879 4880 return rc2 ? rc2 : count; 4881 } 4882 4883 static int 4884 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4885 struct spdk_nvme_transport_id *trid, 4886 bool peer) 4887 { 4888 struct sockaddr *saddr; 4889 uint16_t port; 4890 4891 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4892 4893 if (peer) { 4894 saddr = rdma_get_peer_addr(id); 4895 } else { 4896 saddr = rdma_get_local_addr(id); 4897 } 4898 switch (saddr->sa_family) { 4899 case AF_INET: { 4900 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4901 4902 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4903 inet_ntop(AF_INET, &saddr_in->sin_addr, 4904 trid->traddr, sizeof(trid->traddr)); 4905 if (peer) { 4906 port = ntohs(rdma_get_dst_port(id)); 4907 } else { 4908 port = ntohs(rdma_get_src_port(id)); 4909 } 4910 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4911 break; 4912 } 4913 case AF_INET6: { 4914 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4915 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4916 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4917 trid->traddr, sizeof(trid->traddr)); 4918 if (peer) { 4919 port = ntohs(rdma_get_dst_port(id)); 4920 } else { 4921 port = ntohs(rdma_get_src_port(id)); 4922 } 4923 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4924 break; 4925 } 4926 default: 4927 return -1; 4928 4929 } 4930 4931 return 0; 4932 } 4933 4934 static int 4935 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4936 struct spdk_nvme_transport_id *trid) 4937 { 4938 struct spdk_nvmf_rdma_qpair *rqpair; 4939 4940 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4941 4942 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4943 } 4944 4945 static int 4946 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4947 struct spdk_nvme_transport_id *trid) 4948 { 4949 struct spdk_nvmf_rdma_qpair *rqpair; 4950 4951 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4952 4953 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4954 } 4955 4956 static int 4957 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4958 struct spdk_nvme_transport_id *trid) 4959 { 4960 struct spdk_nvmf_rdma_qpair *rqpair; 4961 4962 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4963 4964 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4965 } 4966 4967 void 4968 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4969 { 4970 g_nvmf_hooks = *hooks; 4971 } 4972 4973 static void 4974 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4975 struct spdk_nvmf_rdma_request *rdma_req_to_abort, 4976 struct spdk_nvmf_rdma_qpair *rqpair) 4977 { 4978 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4979 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4980 4981 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, state_link); 4982 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING; 4983 4984 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4985 } 4986 4987 static int 4988 _nvmf_rdma_qpair_abort_request(void *ctx) 4989 { 4990 struct spdk_nvmf_request *req = ctx; 4991 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4992 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4993 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4994 struct spdk_nvmf_rdma_qpair, qpair); 4995 int rc; 4996 4997 spdk_poller_unregister(&req->poller); 4998 4999 switch (rdma_req_to_abort->state) { 5000 case RDMA_REQUEST_STATE_EXECUTING: 5001 rc = nvmf_ctrlr_abort_request(req); 5002 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 5003 return SPDK_POLLER_BUSY; 5004 } 5005 break; 5006 5007 case RDMA_REQUEST_STATE_NEED_BUFFER: 5008 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 5009 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 5010 5011 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5012 break; 5013 5014 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 5015 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 5016 spdk_nvmf_rdma_request, state_link); 5017 5018 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5019 break; 5020 5021 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 5022 STAILQ_REMOVE(&rqpair->pending_rdma_write_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_READY_TO_COMPLETE_PENDING: 5029 /* Remove req from the list here to re-use common function */ 5030 STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, 5031 spdk_nvmf_rdma_request, state_link); 5032 5033 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair); 5034 break; 5035 5036 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 5037 if (spdk_get_ticks() < req->timeout_tsc) { 5038 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 5039 return SPDK_POLLER_BUSY; 5040 } 5041 break; 5042 5043 default: 5044 break; 5045 } 5046 5047 spdk_nvmf_request_complete(req); 5048 return SPDK_POLLER_BUSY; 5049 } 5050 5051 static void 5052 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 5053 struct spdk_nvmf_request *req) 5054 { 5055 struct spdk_nvmf_rdma_qpair *rqpair; 5056 struct spdk_nvmf_rdma_transport *rtransport; 5057 struct spdk_nvmf_transport *transport; 5058 uint16_t cid; 5059 uint32_t i, max_req_count; 5060 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 5061 5062 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 5063 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 5064 transport = &rtransport->transport; 5065 5066 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 5067 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 5068 5069 for (i = 0; i < max_req_count; i++) { 5070 rdma_req = &rqpair->resources->reqs[i]; 5071 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 5072 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 5073 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 5074 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 5075 rdma_req->req.qpair == qpair) { 5076 rdma_req_to_abort = rdma_req; 5077 break; 5078 } 5079 } 5080 5081 if (rdma_req_to_abort == NULL) { 5082 spdk_nvmf_request_complete(req); 5083 return; 5084 } 5085 5086 req->req_to_abort = &rdma_req_to_abort->req; 5087 req->timeout_tsc = spdk_get_ticks() + 5088 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 5089 req->poller = NULL; 5090 5091 _nvmf_rdma_qpair_abort_request(req); 5092 } 5093 5094 static void 5095 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 5096 struct spdk_json_write_ctx *w) 5097 { 5098 struct spdk_nvmf_rdma_poll_group *rgroup; 5099 struct spdk_nvmf_rdma_poller *rpoller; 5100 5101 assert(w != NULL); 5102 5103 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 5104 5105 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 5106 5107 spdk_json_write_named_array_begin(w, "devices"); 5108 5109 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 5110 spdk_json_write_object_begin(w); 5111 spdk_json_write_named_string(w, "name", 5112 ibv_get_device_name(rpoller->device->context->device)); 5113 spdk_json_write_named_uint64(w, "polls", 5114 rpoller->stat.polls); 5115 spdk_json_write_named_uint64(w, "idle_polls", 5116 rpoller->stat.idle_polls); 5117 spdk_json_write_named_uint64(w, "completions", 5118 rpoller->stat.completions); 5119 spdk_json_write_named_uint64(w, "requests", 5120 rpoller->stat.requests); 5121 spdk_json_write_named_uint64(w, "request_latency", 5122 rpoller->stat.request_latency); 5123 spdk_json_write_named_uint64(w, "pending_free_request", 5124 rpoller->stat.pending_free_request); 5125 spdk_json_write_named_uint64(w, "pending_rdma_read", 5126 rpoller->stat.pending_rdma_read); 5127 spdk_json_write_named_uint64(w, "pending_rdma_write", 5128 rpoller->stat.pending_rdma_write); 5129 spdk_json_write_named_uint64(w, "pending_rdma_send", 5130 rpoller->stat.pending_rdma_send); 5131 spdk_json_write_named_uint64(w, "total_send_wrs", 5132 rpoller->stat.qp_stats.send.num_submitted_wrs); 5133 spdk_json_write_named_uint64(w, "send_doorbell_updates", 5134 rpoller->stat.qp_stats.send.doorbell_updates); 5135 spdk_json_write_named_uint64(w, "total_recv_wrs", 5136 rpoller->stat.qp_stats.recv.num_submitted_wrs); 5137 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 5138 rpoller->stat.qp_stats.recv.doorbell_updates); 5139 spdk_json_write_object_end(w); 5140 } 5141 5142 spdk_json_write_array_end(w); 5143 } 5144 5145 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 5146 .name = "RDMA", 5147 .type = SPDK_NVME_TRANSPORT_RDMA, 5148 .opts_init = nvmf_rdma_opts_init, 5149 .create = nvmf_rdma_create, 5150 .dump_opts = nvmf_rdma_dump_opts, 5151 .destroy = nvmf_rdma_destroy, 5152 5153 .listen = nvmf_rdma_listen, 5154 .stop_listen = nvmf_rdma_stop_listen, 5155 .cdata_init = nvmf_rdma_cdata_init, 5156 5157 .listener_discover = nvmf_rdma_discover, 5158 5159 .poll_group_create = nvmf_rdma_poll_group_create, 5160 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 5161 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 5162 .poll_group_add = nvmf_rdma_poll_group_add, 5163 .poll_group_remove = nvmf_rdma_poll_group_remove, 5164 .poll_group_poll = nvmf_rdma_poll_group_poll, 5165 5166 .req_free = nvmf_rdma_request_free, 5167 .req_complete = nvmf_rdma_request_complete, 5168 5169 .qpair_fini = nvmf_rdma_close_qpair, 5170 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 5171 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 5172 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 5173 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 5174 5175 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 5176 }; 5177 5178 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 5179 SPDK_LOG_REGISTER_COMPONENT(rdma) 5180