1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2019, 2020 Mellanox Technologies LTD. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of Intel Corporation nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include "spdk/stdinc.h" 35 36 #include "spdk/config.h" 37 #include "spdk/thread.h" 38 #include "spdk/likely.h" 39 #include "spdk/nvmf_transport.h" 40 #include "spdk/string.h" 41 #include "spdk/trace.h" 42 #include "spdk/util.h" 43 44 #include "spdk_internal/assert.h" 45 #include "spdk/log.h" 46 #include "spdk_internal/rdma.h" 47 48 #include "nvmf_internal.h" 49 50 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 51 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 52 53 /* 54 RDMA Connection Resource Defaults 55 */ 56 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 57 #define NVMF_DEFAULT_RSP_SGE 1 58 #define NVMF_DEFAULT_RX_SGE 2 59 60 /* The RDMA completion queue size */ 61 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 62 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 63 64 static int g_spdk_nvmf_ibv_query_mask = 65 IBV_QP_STATE | 66 IBV_QP_PKEY_INDEX | 67 IBV_QP_PORT | 68 IBV_QP_ACCESS_FLAGS | 69 IBV_QP_AV | 70 IBV_QP_PATH_MTU | 71 IBV_QP_DEST_QPN | 72 IBV_QP_RQ_PSN | 73 IBV_QP_MAX_DEST_RD_ATOMIC | 74 IBV_QP_MIN_RNR_TIMER | 75 IBV_QP_SQ_PSN | 76 IBV_QP_TIMEOUT | 77 IBV_QP_RETRY_CNT | 78 IBV_QP_RNR_RETRY | 79 IBV_QP_MAX_QP_RD_ATOMIC; 80 81 enum spdk_nvmf_rdma_request_state { 82 /* The request is not currently in use */ 83 RDMA_REQUEST_STATE_FREE = 0, 84 85 /* Initial state when request first received */ 86 RDMA_REQUEST_STATE_NEW, 87 88 /* The request is queued until a data buffer is available. */ 89 RDMA_REQUEST_STATE_NEED_BUFFER, 90 91 /* The request is waiting on RDMA queue depth availability 92 * to transfer data from the host to the controller. 93 */ 94 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 95 96 /* The request is currently transferring data from the host to the controller. */ 97 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 98 99 /* The request is ready to execute at the block device */ 100 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 101 102 /* The request is currently executing at the block device */ 103 RDMA_REQUEST_STATE_EXECUTING, 104 105 /* The request finished executing at the block device */ 106 RDMA_REQUEST_STATE_EXECUTED, 107 108 /* The request is waiting on RDMA queue depth availability 109 * to transfer data from the controller to the host. 110 */ 111 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 112 113 /* The request is ready to send a completion */ 114 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 115 116 /* The request is currently transferring data from the controller to the host. */ 117 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 118 119 /* The request currently has an outstanding completion without an 120 * associated data transfer. 121 */ 122 RDMA_REQUEST_STATE_COMPLETING, 123 124 /* The request completed and can be marked free. */ 125 RDMA_REQUEST_STATE_COMPLETED, 126 127 /* Terminator */ 128 RDMA_REQUEST_NUM_STATES, 129 }; 130 131 #define OBJECT_NVMF_RDMA_IO 0x40 132 133 #define TRACE_GROUP_NVMF_RDMA 0x4 134 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 135 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 136 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 137 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 138 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 139 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 140 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 141 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 142 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 143 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 144 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 145 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 146 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 147 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 148 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 149 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 150 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 151 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 152 153 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 154 { 155 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 156 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 157 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 158 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 159 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 160 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 161 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 162 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 163 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 164 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 165 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 166 spdk_trace_register_description("RDMA_REQ_TX_H2C", 167 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 168 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 169 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 170 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 171 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 172 spdk_trace_register_description("RDMA_REQ_EXECUTING", 173 TRACE_RDMA_REQUEST_STATE_EXECUTING, 174 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 175 spdk_trace_register_description("RDMA_REQ_EXECUTED", 176 TRACE_RDMA_REQUEST_STATE_EXECUTED, 177 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 178 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 179 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 180 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 181 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 182 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 183 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 184 spdk_trace_register_description("RDMA_REQ_COMPLETING", 185 TRACE_RDMA_REQUEST_STATE_COMPLETING, 186 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 187 spdk_trace_register_description("RDMA_REQ_COMPLETED", 188 TRACE_RDMA_REQUEST_STATE_COMPLETED, 189 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 190 191 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 192 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 193 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 194 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 195 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 196 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 197 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 198 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 199 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 200 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 201 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 202 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 203 } 204 205 enum spdk_nvmf_rdma_wr_type { 206 RDMA_WR_TYPE_RECV, 207 RDMA_WR_TYPE_SEND, 208 RDMA_WR_TYPE_DATA, 209 }; 210 211 struct spdk_nvmf_rdma_wr { 212 enum spdk_nvmf_rdma_wr_type type; 213 }; 214 215 /* This structure holds commands as they are received off the wire. 216 * It must be dynamically paired with a full request object 217 * (spdk_nvmf_rdma_request) to service a request. It is separate 218 * from the request because RDMA does not appear to order 219 * completions, so occasionally we'll get a new incoming 220 * command when there aren't any free request objects. 221 */ 222 struct spdk_nvmf_rdma_recv { 223 struct ibv_recv_wr wr; 224 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 225 226 struct spdk_nvmf_rdma_qpair *qpair; 227 228 /* In-capsule data buffer */ 229 uint8_t *buf; 230 231 struct spdk_nvmf_rdma_wr rdma_wr; 232 uint64_t receive_tsc; 233 234 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 235 }; 236 237 struct spdk_nvmf_rdma_request_data { 238 struct spdk_nvmf_rdma_wr rdma_wr; 239 struct ibv_send_wr wr; 240 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 241 }; 242 243 struct spdk_nvmf_rdma_request { 244 struct spdk_nvmf_request req; 245 246 enum spdk_nvmf_rdma_request_state state; 247 248 struct spdk_nvmf_rdma_recv *recv; 249 250 struct { 251 struct spdk_nvmf_rdma_wr rdma_wr; 252 struct ibv_send_wr wr; 253 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 254 } rsp; 255 256 struct spdk_nvmf_rdma_request_data data; 257 258 uint32_t iovpos; 259 260 uint32_t num_outstanding_data_wr; 261 uint64_t receive_tsc; 262 263 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 264 }; 265 266 struct spdk_nvmf_rdma_resource_opts { 267 struct spdk_nvmf_rdma_qpair *qpair; 268 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 269 void *qp; 270 struct ibv_pd *pd; 271 uint32_t max_queue_depth; 272 uint32_t in_capsule_data_size; 273 bool shared; 274 }; 275 276 struct spdk_nvmf_recv_wr_list { 277 struct ibv_recv_wr *first; 278 struct ibv_recv_wr *last; 279 }; 280 281 struct spdk_nvmf_rdma_resources { 282 /* Array of size "max_queue_depth" containing RDMA requests. */ 283 struct spdk_nvmf_rdma_request *reqs; 284 285 /* Array of size "max_queue_depth" containing RDMA recvs. */ 286 struct spdk_nvmf_rdma_recv *recvs; 287 288 /* Array of size "max_queue_depth" containing 64 byte capsules 289 * used for receive. 290 */ 291 union nvmf_h2c_msg *cmds; 292 struct ibv_mr *cmds_mr; 293 294 /* Array of size "max_queue_depth" containing 16 byte completions 295 * to be sent back to the user. 296 */ 297 union nvmf_c2h_msg *cpls; 298 struct ibv_mr *cpls_mr; 299 300 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 301 * buffers to be used for in capsule data. 302 */ 303 void *bufs; 304 struct ibv_mr *bufs_mr; 305 306 /* The list of pending recvs to transfer */ 307 struct spdk_nvmf_recv_wr_list recvs_to_post; 308 309 /* Receives that are waiting for a request object */ 310 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 311 312 /* Queue to track free requests */ 313 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 314 }; 315 316 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 317 318 struct spdk_nvmf_rdma_ibv_event_ctx { 319 struct spdk_nvmf_rdma_qpair *rqpair; 320 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 321 /* Link to other ibv events associated with this qpair */ 322 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 323 }; 324 325 struct spdk_nvmf_rdma_qpair { 326 struct spdk_nvmf_qpair qpair; 327 328 struct spdk_nvmf_rdma_device *device; 329 struct spdk_nvmf_rdma_poller *poller; 330 331 struct spdk_rdma_qp *rdma_qp; 332 struct rdma_cm_id *cm_id; 333 struct ibv_srq *srq; 334 struct rdma_cm_id *listen_id; 335 336 /* The maximum number of I/O outstanding on this connection at one time */ 337 uint16_t max_queue_depth; 338 339 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 340 uint16_t max_read_depth; 341 342 /* The maximum number of RDMA SEND operations at one time */ 343 uint32_t max_send_depth; 344 345 /* The current number of outstanding WRs from this qpair's 346 * recv queue. Should not exceed device->attr.max_queue_depth. 347 */ 348 uint16_t current_recv_depth; 349 350 /* The current number of active RDMA READ operations */ 351 uint16_t current_read_depth; 352 353 /* The current number of posted WRs from this qpair's 354 * send queue. Should not exceed max_send_depth. 355 */ 356 uint32_t current_send_depth; 357 358 /* The maximum number of SGEs per WR on the send queue */ 359 uint32_t max_send_sge; 360 361 /* The maximum number of SGEs per WR on the recv queue */ 362 uint32_t max_recv_sge; 363 364 struct spdk_nvmf_rdma_resources *resources; 365 366 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 367 368 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 369 370 /* Number of requests not in the free state */ 371 uint32_t qd; 372 373 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 374 375 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 376 377 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 378 379 /* IBV queue pair attributes: they are used to manage 380 * qp state and recover from errors. 381 */ 382 enum ibv_qp_state ibv_state; 383 384 /* 385 * io_channel which is used to destroy qpair when it is removed from poll group 386 */ 387 struct spdk_io_channel *destruct_channel; 388 389 /* List of ibv async events */ 390 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 391 392 /* Lets us know that we have received the last_wqe event. */ 393 bool last_wqe_reached; 394 395 /* Indicate that nvmf_rdma_close_qpair is called */ 396 bool to_close; 397 }; 398 399 struct spdk_nvmf_rdma_poller_stat { 400 uint64_t completions; 401 uint64_t polls; 402 uint64_t requests; 403 uint64_t request_latency; 404 uint64_t pending_free_request; 405 uint64_t pending_rdma_read; 406 uint64_t pending_rdma_write; 407 }; 408 409 struct spdk_nvmf_rdma_poller { 410 struct spdk_nvmf_rdma_device *device; 411 struct spdk_nvmf_rdma_poll_group *group; 412 413 int num_cqe; 414 int required_num_wr; 415 struct ibv_cq *cq; 416 417 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 418 uint16_t max_srq_depth; 419 420 /* Shared receive queue */ 421 struct ibv_srq *srq; 422 423 struct spdk_nvmf_rdma_resources *resources; 424 struct spdk_nvmf_rdma_poller_stat stat; 425 426 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 427 428 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 429 430 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 431 432 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 433 }; 434 435 struct spdk_nvmf_rdma_poll_group_stat { 436 uint64_t pending_data_buffer; 437 }; 438 439 struct spdk_nvmf_rdma_poll_group { 440 struct spdk_nvmf_transport_poll_group group; 441 struct spdk_nvmf_rdma_poll_group_stat stat; 442 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 443 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 444 /* 445 * buffers which are split across multiple RDMA 446 * memory regions cannot be used by this transport. 447 */ 448 STAILQ_HEAD(, spdk_nvmf_transport_pg_cache_buf) retired_bufs; 449 }; 450 451 struct spdk_nvmf_rdma_conn_sched { 452 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 453 struct spdk_nvmf_rdma_poll_group *next_io_pg; 454 }; 455 456 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 457 struct spdk_nvmf_rdma_device { 458 struct ibv_device_attr attr; 459 struct ibv_context *context; 460 461 struct spdk_mem_map *map; 462 struct ibv_pd *pd; 463 464 int num_srq; 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 uint32_t max_srq_depth; 478 bool no_srq; 479 bool no_wr_batching; 480 int acceptor_backlog; 481 }; 482 483 struct spdk_nvmf_rdma_transport { 484 struct spdk_nvmf_transport transport; 485 struct rdma_transport_opts rdma_opts; 486 487 struct spdk_nvmf_rdma_conn_sched conn_sched; 488 489 struct rdma_event_channel *event_channel; 490 491 struct spdk_mempool *data_wr_pool; 492 493 pthread_mutex_t lock; 494 495 /* fields used to poll RDMA/IB events */ 496 nfds_t npoll_fds; 497 struct pollfd *poll_fds; 498 499 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 500 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 501 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 502 }; 503 504 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 505 { 506 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 507 spdk_json_decode_uint32, true 508 }, 509 { 510 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 511 spdk_json_decode_bool, true 512 }, 513 { 514 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 515 spdk_json_decode_bool, true 516 }, 517 { 518 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 519 spdk_json_decode_int32, true 520 }, 521 }; 522 523 static bool 524 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 525 struct spdk_nvmf_rdma_request *rdma_req); 526 527 static void 528 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 529 struct spdk_nvmf_rdma_poller *rpoller); 530 531 static void 532 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 533 struct spdk_nvmf_rdma_poller *rpoller); 534 535 static inline int 536 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 537 { 538 switch (state) { 539 case IBV_QPS_RESET: 540 case IBV_QPS_INIT: 541 case IBV_QPS_RTR: 542 case IBV_QPS_RTS: 543 case IBV_QPS_SQD: 544 case IBV_QPS_SQE: 545 case IBV_QPS_ERR: 546 return 0; 547 default: 548 return -1; 549 } 550 } 551 552 static inline enum spdk_nvme_media_error_status_code 553 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 554 enum spdk_nvme_media_error_status_code result; 555 switch (err_type) 556 { 557 case SPDK_DIF_REFTAG_ERROR: 558 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 559 break; 560 case SPDK_DIF_APPTAG_ERROR: 561 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 562 break; 563 case SPDK_DIF_GUARD_ERROR: 564 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 565 break; 566 default: 567 SPDK_UNREACHABLE(); 568 } 569 570 return result; 571 } 572 573 static enum ibv_qp_state 574 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 575 enum ibv_qp_state old_state, new_state; 576 struct ibv_qp_attr qp_attr; 577 struct ibv_qp_init_attr init_attr; 578 int rc; 579 580 old_state = rqpair->ibv_state; 581 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 582 g_spdk_nvmf_ibv_query_mask, &init_attr); 583 584 if (rc) 585 { 586 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 587 return IBV_QPS_ERR + 1; 588 } 589 590 new_state = qp_attr.qp_state; 591 rqpair->ibv_state = new_state; 592 qp_attr.ah_attr.port_num = qp_attr.port_num; 593 594 rc = nvmf_rdma_check_ibv_state(new_state); 595 if (rc) 596 { 597 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 598 /* 599 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 600 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 601 */ 602 return IBV_QPS_ERR + 1; 603 } 604 605 if (old_state != new_state) 606 { 607 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 608 (uintptr_t)rqpair->cm_id, new_state); 609 } 610 return new_state; 611 } 612 613 static void 614 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 615 struct spdk_nvmf_rdma_transport *rtransport) 616 { 617 struct spdk_nvmf_rdma_request_data *data_wr; 618 struct ibv_send_wr *next_send_wr; 619 uint64_t req_wrid; 620 621 rdma_req->num_outstanding_data_wr = 0; 622 data_wr = &rdma_req->data; 623 req_wrid = data_wr->wr.wr_id; 624 while (data_wr && data_wr->wr.wr_id == req_wrid) { 625 memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge); 626 data_wr->wr.num_sge = 0; 627 next_send_wr = data_wr->wr.next; 628 if (data_wr != &rdma_req->data) { 629 spdk_mempool_put(rtransport->data_wr_pool, data_wr); 630 } 631 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : 632 SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr); 633 } 634 } 635 636 static void 637 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 638 { 639 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 640 if (req->req.cmd) { 641 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 642 } 643 if (req->recv) { 644 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 645 } 646 } 647 648 static void 649 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 650 { 651 int i; 652 653 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 654 for (i = 0; i < rqpair->max_queue_depth; i++) { 655 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 656 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 657 } 658 } 659 } 660 661 static void 662 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 663 { 664 if (resources->cmds_mr) { 665 ibv_dereg_mr(resources->cmds_mr); 666 } 667 668 if (resources->cpls_mr) { 669 ibv_dereg_mr(resources->cpls_mr); 670 } 671 672 if (resources->bufs_mr) { 673 ibv_dereg_mr(resources->bufs_mr); 674 } 675 676 spdk_free(resources->cmds); 677 spdk_free(resources->cpls); 678 spdk_free(resources->bufs); 679 free(resources->reqs); 680 free(resources->recvs); 681 free(resources); 682 } 683 684 685 static struct spdk_nvmf_rdma_resources * 686 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 687 { 688 struct spdk_nvmf_rdma_resources *resources; 689 struct spdk_nvmf_rdma_request *rdma_req; 690 struct spdk_nvmf_rdma_recv *rdma_recv; 691 struct ibv_qp *qp; 692 struct ibv_srq *srq; 693 uint32_t i; 694 int rc; 695 696 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 697 if (!resources) { 698 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 699 return NULL; 700 } 701 702 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 703 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 704 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 705 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 706 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 707 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 708 709 if (opts->in_capsule_data_size > 0) { 710 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 711 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 712 SPDK_MALLOC_DMA); 713 } 714 715 if (!resources->reqs || !resources->recvs || !resources->cmds || 716 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 717 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 718 goto cleanup; 719 } 720 721 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 722 opts->max_queue_depth * sizeof(*resources->cmds), 723 IBV_ACCESS_LOCAL_WRITE); 724 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 725 opts->max_queue_depth * sizeof(*resources->cpls), 726 0); 727 728 if (opts->in_capsule_data_size) { 729 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 730 opts->max_queue_depth * 731 opts->in_capsule_data_size, 732 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 733 } 734 735 if (!resources->cmds_mr || !resources->cpls_mr || 736 (opts->in_capsule_data_size && 737 !resources->bufs_mr)) { 738 goto cleanup; 739 } 740 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx LKey: %x\n", 741 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 742 resources->cmds_mr->lkey); 743 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx LKey: %x\n", 744 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 745 resources->cpls_mr->lkey); 746 if (resources->bufs && resources->bufs_mr) { 747 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x LKey: %x\n", 748 resources->bufs, opts->max_queue_depth * 749 opts->in_capsule_data_size, resources->bufs_mr->lkey); 750 } 751 752 /* Initialize queues */ 753 STAILQ_INIT(&resources->incoming_queue); 754 STAILQ_INIT(&resources->free_queue); 755 756 for (i = 0; i < opts->max_queue_depth; i++) { 757 struct ibv_recv_wr *bad_wr = NULL; 758 759 rdma_recv = &resources->recvs[i]; 760 rdma_recv->qpair = opts->qpair; 761 762 /* Set up memory to receive commands */ 763 if (resources->bufs) { 764 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 765 opts->in_capsule_data_size)); 766 } 767 768 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 769 770 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 771 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 772 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 773 rdma_recv->wr.num_sge = 1; 774 775 if (rdma_recv->buf && resources->bufs_mr) { 776 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 777 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 778 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 779 rdma_recv->wr.num_sge++; 780 } 781 782 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 783 rdma_recv->wr.sg_list = rdma_recv->sgl; 784 if (opts->shared) { 785 srq = (struct ibv_srq *)opts->qp; 786 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 787 } else { 788 qp = (struct ibv_qp *)opts->qp; 789 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 790 } 791 if (rc) { 792 goto cleanup; 793 } 794 } 795 796 for (i = 0; i < opts->max_queue_depth; i++) { 797 rdma_req = &resources->reqs[i]; 798 799 if (opts->qpair != NULL) { 800 rdma_req->req.qpair = &opts->qpair->qpair; 801 } else { 802 rdma_req->req.qpair = NULL; 803 } 804 rdma_req->req.cmd = NULL; 805 806 /* Set up memory to send responses */ 807 rdma_req->req.rsp = &resources->cpls[i]; 808 809 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 810 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 811 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 812 813 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 814 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 815 rdma_req->rsp.wr.next = NULL; 816 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 817 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 818 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 819 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 820 821 /* Set up memory for data buffers */ 822 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 823 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 824 rdma_req->data.wr.next = NULL; 825 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 826 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 827 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 828 829 /* Initialize request state to FREE */ 830 rdma_req->state = RDMA_REQUEST_STATE_FREE; 831 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 832 } 833 834 return resources; 835 836 cleanup: 837 nvmf_rdma_resources_destroy(resources); 838 return NULL; 839 } 840 841 static void 842 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 843 { 844 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 845 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 846 ctx->rqpair = NULL; 847 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 848 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 849 } 850 } 851 852 static void 853 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 854 { 855 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 856 struct ibv_recv_wr *bad_recv_wr = NULL; 857 int rc; 858 859 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 860 861 if (rqpair->qd != 0) { 862 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 863 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 864 struct spdk_nvmf_rdma_transport, transport); 865 struct spdk_nvmf_rdma_request *req; 866 uint32_t i, max_req_count = 0; 867 868 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 869 870 if (rqpair->srq == NULL) { 871 nvmf_rdma_dump_qpair_contents(rqpair); 872 max_req_count = rqpair->max_queue_depth; 873 } else if (rqpair->poller && rqpair->resources) { 874 max_req_count = rqpair->poller->max_srq_depth; 875 } 876 877 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 878 for (i = 0; i < max_req_count; i++) { 879 req = &rqpair->resources->reqs[i]; 880 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 881 /* nvmf_rdma_request_process checks qpair ibv and internal state 882 * and completes a request */ 883 nvmf_rdma_request_process(rtransport, req); 884 } 885 } 886 assert(rqpair->qd == 0); 887 } 888 889 if (rqpair->poller) { 890 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 891 892 if (rqpair->srq != NULL && rqpair->resources != NULL) { 893 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 894 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 895 if (rqpair == rdma_recv->qpair) { 896 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 897 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 898 if (rc) { 899 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 900 } 901 } 902 } 903 } 904 } 905 906 if (rqpair->cm_id) { 907 if (rqpair->rdma_qp != NULL) { 908 spdk_rdma_qp_destroy(rqpair->rdma_qp); 909 rqpair->rdma_qp = NULL; 910 } 911 rdma_destroy_id(rqpair->cm_id); 912 913 if (rqpair->poller != NULL && rqpair->srq == NULL) { 914 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 915 } 916 } 917 918 if (rqpair->srq == NULL && rqpair->resources != NULL) { 919 nvmf_rdma_resources_destroy(rqpair->resources); 920 } 921 922 nvmf_rdma_qpair_clean_ibv_events(rqpair); 923 924 if (rqpair->destruct_channel) { 925 spdk_put_io_channel(rqpair->destruct_channel); 926 rqpair->destruct_channel = NULL; 927 } 928 929 free(rqpair); 930 } 931 932 static int 933 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 934 { 935 struct spdk_nvmf_rdma_poller *rpoller; 936 int rc, num_cqe, required_num_wr; 937 938 /* Enlarge CQ size dynamically */ 939 rpoller = rqpair->poller; 940 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 941 num_cqe = rpoller->num_cqe; 942 if (num_cqe < required_num_wr) { 943 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 944 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 945 } 946 947 if (rpoller->num_cqe != num_cqe) { 948 if (required_num_wr > device->attr.max_cqe) { 949 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 950 required_num_wr, device->attr.max_cqe); 951 return -1; 952 } 953 954 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 955 rc = ibv_resize_cq(rpoller->cq, num_cqe); 956 if (rc) { 957 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 958 return -1; 959 } 960 961 rpoller->num_cqe = num_cqe; 962 } 963 964 rpoller->required_num_wr = required_num_wr; 965 return 0; 966 } 967 968 static int 969 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 970 { 971 struct spdk_nvmf_rdma_qpair *rqpair; 972 struct spdk_nvmf_rdma_transport *rtransport; 973 struct spdk_nvmf_transport *transport; 974 struct spdk_nvmf_rdma_resource_opts opts; 975 struct spdk_nvmf_rdma_device *device; 976 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 977 978 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 979 device = rqpair->device; 980 981 qp_init_attr.qp_context = rqpair; 982 qp_init_attr.pd = device->pd; 983 qp_init_attr.send_cq = rqpair->poller->cq; 984 qp_init_attr.recv_cq = rqpair->poller->cq; 985 986 if (rqpair->srq) { 987 qp_init_attr.srq = rqpair->srq; 988 } else { 989 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 990 } 991 992 /* SEND, READ, and WRITE operations */ 993 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 994 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 995 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 996 997 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 998 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 999 goto error; 1000 } 1001 1002 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 1003 if (!rqpair->rdma_qp) { 1004 goto error; 1005 } 1006 1007 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1008 qp_init_attr.cap.max_send_wr); 1009 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1010 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1011 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 1012 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1013 1014 if (rqpair->poller->srq == NULL) { 1015 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1016 transport = &rtransport->transport; 1017 1018 opts.qp = rqpair->rdma_qp->qp; 1019 opts.pd = rqpair->cm_id->pd; 1020 opts.qpair = rqpair; 1021 opts.shared = false; 1022 opts.max_queue_depth = rqpair->max_queue_depth; 1023 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1024 1025 rqpair->resources = nvmf_rdma_resources_create(&opts); 1026 1027 if (!rqpair->resources) { 1028 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1029 rdma_destroy_qp(rqpair->cm_id); 1030 goto error; 1031 } 1032 } else { 1033 rqpair->resources = rqpair->poller->resources; 1034 } 1035 1036 rqpair->current_recv_depth = 0; 1037 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1038 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1039 1040 return 0; 1041 1042 error: 1043 rdma_destroy_id(rqpair->cm_id); 1044 rqpair->cm_id = NULL; 1045 return -1; 1046 } 1047 1048 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1049 /* This function accepts either a single wr or the first wr in a linked list. */ 1050 static void 1051 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1052 { 1053 struct ibv_recv_wr *last; 1054 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1055 struct spdk_nvmf_rdma_transport, transport); 1056 1057 last = first; 1058 while (last->next != NULL) { 1059 last = last->next; 1060 } 1061 1062 if (rqpair->resources->recvs_to_post.first == NULL) { 1063 rqpair->resources->recvs_to_post.first = first; 1064 rqpair->resources->recvs_to_post.last = last; 1065 if (rqpair->srq == NULL) { 1066 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1067 } 1068 } else { 1069 rqpair->resources->recvs_to_post.last->next = first; 1070 rqpair->resources->recvs_to_post.last = last; 1071 } 1072 1073 if (rtransport->rdma_opts.no_wr_batching) { 1074 _poller_submit_recvs(rtransport, rqpair->poller); 1075 } 1076 } 1077 1078 static int 1079 request_transfer_in(struct spdk_nvmf_request *req) 1080 { 1081 struct spdk_nvmf_rdma_request *rdma_req; 1082 struct spdk_nvmf_qpair *qpair; 1083 struct spdk_nvmf_rdma_qpair *rqpair; 1084 struct spdk_nvmf_rdma_transport *rtransport; 1085 1086 qpair = req->qpair; 1087 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1088 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1089 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1090 struct spdk_nvmf_rdma_transport, transport); 1091 1092 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1093 assert(rdma_req != NULL); 1094 1095 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1096 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1097 } 1098 if (rtransport->rdma_opts.no_wr_batching) { 1099 _poller_submit_sends(rtransport, rqpair->poller); 1100 } 1101 1102 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1103 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1104 return 0; 1105 } 1106 1107 static int 1108 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1109 { 1110 int num_outstanding_data_wr = 0; 1111 struct spdk_nvmf_rdma_request *rdma_req; 1112 struct spdk_nvmf_qpair *qpair; 1113 struct spdk_nvmf_rdma_qpair *rqpair; 1114 struct spdk_nvme_cpl *rsp; 1115 struct ibv_send_wr *first = NULL; 1116 struct spdk_nvmf_rdma_transport *rtransport; 1117 1118 *data_posted = 0; 1119 qpair = req->qpair; 1120 rsp = &req->rsp->nvme_cpl; 1121 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1122 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1123 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1124 struct spdk_nvmf_rdma_transport, transport); 1125 1126 /* Advance our sq_head pointer */ 1127 if (qpair->sq_head == qpair->sq_head_max) { 1128 qpair->sq_head = 0; 1129 } else { 1130 qpair->sq_head++; 1131 } 1132 rsp->sqhd = qpair->sq_head; 1133 1134 /* queue the capsule for the recv buffer */ 1135 assert(rdma_req->recv != NULL); 1136 1137 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1138 1139 rdma_req->recv = NULL; 1140 assert(rqpair->current_recv_depth > 0); 1141 rqpair->current_recv_depth--; 1142 1143 /* Build the response which consists of optional 1144 * RDMA WRITEs to transfer data, plus an RDMA SEND 1145 * containing the response. 1146 */ 1147 first = &rdma_req->rsp.wr; 1148 1149 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1150 /* On failure, data was not read from the controller. So clear the 1151 * number of outstanding data WRs to zero. 1152 */ 1153 rdma_req->num_outstanding_data_wr = 0; 1154 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1155 first = &rdma_req->data.wr; 1156 *data_posted = 1; 1157 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1158 } 1159 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1160 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1161 } 1162 if (rtransport->rdma_opts.no_wr_batching) { 1163 _poller_submit_sends(rtransport, rqpair->poller); 1164 } 1165 1166 /* +1 for the rsp wr */ 1167 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1168 1169 return 0; 1170 } 1171 1172 static int 1173 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1174 { 1175 struct spdk_nvmf_rdma_accept_private_data accept_data; 1176 struct rdma_conn_param ctrlr_event_data = {}; 1177 int rc; 1178 1179 accept_data.recfmt = 0; 1180 accept_data.crqsize = rqpair->max_queue_depth; 1181 1182 ctrlr_event_data.private_data = &accept_data; 1183 ctrlr_event_data.private_data_len = sizeof(accept_data); 1184 if (id->ps == RDMA_PS_TCP) { 1185 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1186 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1187 } 1188 1189 /* Configure infinite retries for the initiator side qpair. 1190 * When using a shared receive queue on the target side, 1191 * we need to pass this value to the initiator to prevent the 1192 * initiator side NIC from completing SEND requests back to the 1193 * initiator with status rnr_retry_count_exceeded. */ 1194 if (rqpair->srq != NULL) { 1195 ctrlr_event_data.rnr_retry_count = 0x7; 1196 } 1197 1198 /* When qpair is created without use of rdma cm API, an additional 1199 * information must be provided to initiator in the connection response: 1200 * whether qpair is using SRQ and its qp_num 1201 * Fields below are ignored by rdma cm if qpair has been 1202 * created using rdma cm API. */ 1203 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1204 ctrlr_event_data.qp_num = rqpair->rdma_qp->qp->qp_num; 1205 1206 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1207 if (rc) { 1208 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1209 } else { 1210 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1211 } 1212 1213 return rc; 1214 } 1215 1216 static void 1217 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1218 { 1219 struct spdk_nvmf_rdma_reject_private_data rej_data; 1220 1221 rej_data.recfmt = 0; 1222 rej_data.sts = error; 1223 1224 rdma_reject(id, &rej_data, sizeof(rej_data)); 1225 } 1226 1227 static int 1228 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1229 { 1230 struct spdk_nvmf_rdma_transport *rtransport; 1231 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1232 struct spdk_nvmf_rdma_port *port; 1233 struct rdma_conn_param *rdma_param = NULL; 1234 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1235 uint16_t max_queue_depth; 1236 uint16_t max_read_depth; 1237 1238 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1239 1240 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1241 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1242 1243 rdma_param = &event->param.conn; 1244 if (rdma_param->private_data == NULL || 1245 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1246 SPDK_ERRLOG("connect request: no private data provided\n"); 1247 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1248 return -1; 1249 } 1250 1251 private_data = rdma_param->private_data; 1252 if (private_data->recfmt != 0) { 1253 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1254 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1255 return -1; 1256 } 1257 1258 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1259 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1260 1261 port = event->listen_id->context; 1262 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1263 event->listen_id, event->listen_id->verbs, port); 1264 1265 /* Figure out the supported queue depth. This is a multi-step process 1266 * that takes into account hardware maximums, host provided values, 1267 * and our target's internal memory limits */ 1268 1269 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1270 1271 /* Start with the maximum queue depth allowed by the target */ 1272 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1273 max_read_depth = rtransport->transport.opts.max_queue_depth; 1274 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1275 rtransport->transport.opts.max_queue_depth); 1276 1277 /* Next check the local NIC's hardware limitations */ 1278 SPDK_DEBUGLOG(rdma, 1279 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1280 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1281 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1282 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1283 1284 /* Next check the remote NIC's hardware limitations */ 1285 SPDK_DEBUGLOG(rdma, 1286 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1287 rdma_param->initiator_depth, rdma_param->responder_resources); 1288 if (rdma_param->initiator_depth > 0) { 1289 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1290 } 1291 1292 /* Finally check for the host software requested values, which are 1293 * optional. */ 1294 if (rdma_param->private_data != NULL && 1295 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1296 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1297 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1298 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1299 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1300 } 1301 1302 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1303 max_queue_depth, max_read_depth); 1304 1305 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1306 if (rqpair == NULL) { 1307 SPDK_ERRLOG("Could not allocate new connection.\n"); 1308 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1309 return -1; 1310 } 1311 1312 rqpair->device = port->device; 1313 rqpair->max_queue_depth = max_queue_depth; 1314 rqpair->max_read_depth = max_read_depth; 1315 rqpair->cm_id = event->id; 1316 rqpair->listen_id = event->listen_id; 1317 rqpair->qpair.transport = transport; 1318 rqpair->qpair.trid = port->trid; 1319 STAILQ_INIT(&rqpair->ibv_events); 1320 /* use qid from the private data to determine the qpair type 1321 qid will be set to the appropriate value when the controller is created */ 1322 rqpair->qpair.qid = private_data->qid; 1323 1324 event->id->context = &rqpair->qpair; 1325 1326 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1327 1328 return 0; 1329 } 1330 1331 static int 1332 nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1333 enum spdk_mem_map_notify_action action, 1334 void *vaddr, size_t size) 1335 { 1336 struct ibv_pd *pd = cb_ctx; 1337 struct ibv_mr *mr; 1338 int rc; 1339 1340 switch (action) { 1341 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1342 if (!g_nvmf_hooks.get_rkey) { 1343 mr = ibv_reg_mr(pd, vaddr, size, 1344 IBV_ACCESS_LOCAL_WRITE | 1345 IBV_ACCESS_REMOTE_READ | 1346 IBV_ACCESS_REMOTE_WRITE); 1347 if (mr == NULL) { 1348 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1349 return -1; 1350 } else { 1351 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1352 } 1353 } else { 1354 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1355 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1356 } 1357 break; 1358 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1359 if (!g_nvmf_hooks.get_rkey) { 1360 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1361 if (mr) { 1362 ibv_dereg_mr(mr); 1363 } 1364 } 1365 rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1366 break; 1367 default: 1368 SPDK_UNREACHABLE(); 1369 } 1370 1371 return rc; 1372 } 1373 1374 static int 1375 nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1376 { 1377 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1378 return addr_1 == addr_2; 1379 } 1380 1381 static inline void 1382 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1383 enum spdk_nvme_data_transfer xfer) 1384 { 1385 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1386 wr->opcode = IBV_WR_RDMA_WRITE; 1387 wr->send_flags = 0; 1388 wr->next = next; 1389 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1390 wr->opcode = IBV_WR_RDMA_READ; 1391 wr->send_flags = IBV_SEND_SIGNALED; 1392 wr->next = NULL; 1393 } else { 1394 assert(0); 1395 } 1396 } 1397 1398 static int 1399 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1400 struct spdk_nvmf_rdma_request *rdma_req, 1401 uint32_t num_sgl_descriptors) 1402 { 1403 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1404 struct spdk_nvmf_rdma_request_data *current_data_wr; 1405 uint32_t i; 1406 1407 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1408 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1409 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1410 return -EINVAL; 1411 } 1412 1413 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1414 return -ENOMEM; 1415 } 1416 1417 current_data_wr = &rdma_req->data; 1418 1419 for (i = 0; i < num_sgl_descriptors; i++) { 1420 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1421 current_data_wr->wr.next = &work_requests[i]->wr; 1422 current_data_wr = work_requests[i]; 1423 current_data_wr->wr.sg_list = current_data_wr->sgl; 1424 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1425 } 1426 1427 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1428 1429 return 0; 1430 } 1431 1432 static inline void 1433 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1434 { 1435 struct ibv_send_wr *wr = &rdma_req->data.wr; 1436 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1437 1438 wr->wr.rdma.rkey = sgl->keyed.key; 1439 wr->wr.rdma.remote_addr = sgl->address; 1440 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1441 } 1442 1443 static inline void 1444 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1445 { 1446 struct ibv_send_wr *wr = &rdma_req->data.wr; 1447 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1448 uint32_t i; 1449 int j; 1450 uint64_t remote_addr_offset = 0; 1451 1452 for (i = 0; i < num_wrs; ++i) { 1453 wr->wr.rdma.rkey = sgl->keyed.key; 1454 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1455 for (j = 0; j < wr->num_sge; ++j) { 1456 remote_addr_offset += wr->sg_list[j].length; 1457 } 1458 wr = wr->next; 1459 } 1460 } 1461 1462 /* This function is used in the rare case that we have a buffer split over multiple memory regions. */ 1463 static int 1464 nvmf_rdma_replace_buffer(struct spdk_nvmf_rdma_poll_group *rgroup, void **buf) 1465 { 1466 struct spdk_nvmf_transport_poll_group *group = &rgroup->group; 1467 struct spdk_nvmf_transport *transport = group->transport; 1468 struct spdk_nvmf_transport_pg_cache_buf *old_buf; 1469 void *new_buf; 1470 1471 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1472 group->buf_cache_count--; 1473 new_buf = STAILQ_FIRST(&group->buf_cache); 1474 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1475 assert(*buf != NULL); 1476 } else { 1477 new_buf = spdk_mempool_get(transport->data_buf_pool); 1478 } 1479 1480 if (*buf == NULL) { 1481 return -ENOMEM; 1482 } 1483 1484 old_buf = *buf; 1485 STAILQ_INSERT_HEAD(&rgroup->retired_bufs, old_buf, link); 1486 *buf = new_buf; 1487 return 0; 1488 } 1489 1490 static bool 1491 nvmf_rdma_get_lkey(struct spdk_nvmf_rdma_device *device, struct iovec *iov, 1492 uint32_t *_lkey) 1493 { 1494 uint64_t translation_len; 1495 uint32_t lkey; 1496 1497 translation_len = iov->iov_len; 1498 1499 if (!g_nvmf_hooks.get_rkey) { 1500 lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1501 (uint64_t)iov->iov_base, &translation_len))->lkey; 1502 } else { 1503 lkey = spdk_mem_map_translate(device->map, 1504 (uint64_t)iov->iov_base, &translation_len); 1505 } 1506 1507 if (spdk_unlikely(translation_len < iov->iov_len)) { 1508 return false; 1509 } 1510 1511 *_lkey = lkey; 1512 return true; 1513 } 1514 1515 static bool 1516 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device, 1517 struct iovec *iov, struct ibv_send_wr **_wr, 1518 uint32_t *_remaining_data_block, uint32_t *_offset, 1519 uint32_t *_num_extra_wrs, 1520 const struct spdk_dif_ctx *dif_ctx) 1521 { 1522 struct ibv_send_wr *wr = *_wr; 1523 struct ibv_sge *sg_ele = &wr->sg_list[wr->num_sge]; 1524 uint32_t lkey = 0; 1525 uint32_t remaining, data_block_size, md_size, sge_len; 1526 1527 if (spdk_unlikely(!nvmf_rdma_get_lkey(device, iov, &lkey))) { 1528 /* This is a very rare case that can occur when using DPDK version < 19.05 */ 1529 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions. Removing it from circulation.\n"); 1530 return false; 1531 } 1532 1533 if (spdk_likely(!dif_ctx)) { 1534 sg_ele->lkey = lkey; 1535 sg_ele->addr = (uintptr_t)(iov->iov_base); 1536 sg_ele->length = iov->iov_len; 1537 wr->num_sge++; 1538 } else { 1539 remaining = iov->iov_len - *_offset; 1540 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1541 md_size = dif_ctx->md_size; 1542 1543 while (remaining) { 1544 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1545 if (*_num_extra_wrs > 0 && wr->next) { 1546 *_wr = wr->next; 1547 wr = *_wr; 1548 wr->num_sge = 0; 1549 sg_ele = &wr->sg_list[wr->num_sge]; 1550 (*_num_extra_wrs)--; 1551 } else { 1552 break; 1553 } 1554 } 1555 sg_ele->lkey = lkey; 1556 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset); 1557 sge_len = spdk_min(remaining, *_remaining_data_block); 1558 sg_ele->length = sge_len; 1559 remaining -= sge_len; 1560 *_remaining_data_block -= sge_len; 1561 *_offset += sge_len; 1562 1563 sg_ele++; 1564 wr->num_sge++; 1565 1566 if (*_remaining_data_block == 0) { 1567 /* skip metadata */ 1568 *_offset += md_size; 1569 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1570 remaining -= spdk_min(remaining, md_size); 1571 *_remaining_data_block = data_block_size; 1572 } 1573 1574 if (remaining == 0) { 1575 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1576 the remaining metadata bits at the beginning of the next buffer */ 1577 *_offset -= iov->iov_len; 1578 } 1579 } 1580 } 1581 1582 return true; 1583 } 1584 1585 static int 1586 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1587 struct spdk_nvmf_rdma_device *device, 1588 struct spdk_nvmf_rdma_request *rdma_req, 1589 struct ibv_send_wr *wr, 1590 uint32_t length, 1591 uint32_t num_extra_wrs) 1592 { 1593 struct spdk_nvmf_request *req = &rdma_req->req; 1594 struct spdk_dif_ctx *dif_ctx = NULL; 1595 uint32_t remaining_data_block = 0; 1596 uint32_t offset = 0; 1597 1598 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 1599 dif_ctx = &rdma_req->req.dif.dif_ctx; 1600 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1601 } 1602 1603 wr->num_sge = 0; 1604 1605 while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1606 while (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr, 1607 &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) { 1608 if (nvmf_rdma_replace_buffer(rgroup, &req->buffers[rdma_req->iovpos]) == -ENOMEM) { 1609 return -ENOMEM; 1610 } 1611 req->iov[rdma_req->iovpos].iov_base = (void *)((uintptr_t)(req->buffers[rdma_req->iovpos] + 1612 NVMF_DATA_BUFFER_MASK) & 1613 ~NVMF_DATA_BUFFER_MASK); 1614 } 1615 1616 length -= req->iov[rdma_req->iovpos].iov_len; 1617 rdma_req->iovpos++; 1618 } 1619 1620 if (length) { 1621 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1622 return -EINVAL; 1623 } 1624 1625 return 0; 1626 } 1627 1628 static inline uint32_t 1629 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1630 { 1631 /* estimate the number of SG entries and WRs needed to process the request */ 1632 uint32_t num_sge = 0; 1633 uint32_t i; 1634 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1635 1636 for (i = 0; i < num_buffers && length > 0; i++) { 1637 uint32_t buffer_len = spdk_min(length, io_unit_size); 1638 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1639 1640 if (num_sge_in_block * block_size > buffer_len) { 1641 ++num_sge_in_block; 1642 } 1643 num_sge += num_sge_in_block; 1644 length -= buffer_len; 1645 } 1646 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1647 } 1648 1649 static int 1650 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1651 struct spdk_nvmf_rdma_device *device, 1652 struct spdk_nvmf_rdma_request *rdma_req, 1653 uint32_t length) 1654 { 1655 struct spdk_nvmf_rdma_qpair *rqpair; 1656 struct spdk_nvmf_rdma_poll_group *rgroup; 1657 struct spdk_nvmf_request *req = &rdma_req->req; 1658 struct ibv_send_wr *wr = &rdma_req->data.wr; 1659 int rc; 1660 uint32_t num_wrs = 1; 1661 1662 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1663 rgroup = rqpair->poller->group; 1664 1665 /* rdma wr specifics */ 1666 nvmf_rdma_setup_request(rdma_req); 1667 1668 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1669 length); 1670 if (rc != 0) { 1671 return rc; 1672 } 1673 1674 assert(req->iovcnt <= rqpair->max_send_sge); 1675 1676 rdma_req->iovpos = 0; 1677 1678 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1679 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1680 req->dif.dif_ctx.block_size); 1681 if (num_wrs > 1) { 1682 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1683 if (rc != 0) { 1684 goto err_exit; 1685 } 1686 } 1687 } 1688 1689 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1690 if (spdk_unlikely(rc != 0)) { 1691 goto err_exit; 1692 } 1693 1694 if (spdk_unlikely(num_wrs > 1)) { 1695 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1696 } 1697 1698 /* set the number of outstanding data WRs for this request. */ 1699 rdma_req->num_outstanding_data_wr = num_wrs; 1700 1701 return rc; 1702 1703 err_exit: 1704 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1705 nvmf_rdma_request_free_data(rdma_req, rtransport); 1706 req->iovcnt = 0; 1707 return rc; 1708 } 1709 1710 static int 1711 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1712 struct spdk_nvmf_rdma_device *device, 1713 struct spdk_nvmf_rdma_request *rdma_req) 1714 { 1715 struct spdk_nvmf_rdma_qpair *rqpair; 1716 struct spdk_nvmf_rdma_poll_group *rgroup; 1717 struct ibv_send_wr *current_wr; 1718 struct spdk_nvmf_request *req = &rdma_req->req; 1719 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1720 uint32_t num_sgl_descriptors; 1721 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES]; 1722 uint32_t i; 1723 int rc; 1724 1725 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1726 rgroup = rqpair->poller->group; 1727 1728 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1729 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1730 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1731 1732 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1733 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1734 1735 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1736 return -ENOMEM; 1737 } 1738 1739 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1740 for (i = 0; i < num_sgl_descriptors; i++) { 1741 if (spdk_likely(!req->dif.dif_insert_or_strip)) { 1742 lengths[i] = desc->keyed.length; 1743 } else { 1744 req->dif.orig_length += desc->keyed.length; 1745 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1746 req->dif.elba_length += lengths[i]; 1747 } 1748 desc++; 1749 } 1750 1751 rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport, 1752 lengths, num_sgl_descriptors); 1753 if (rc != 0) { 1754 nvmf_rdma_request_free_data(rdma_req, rtransport); 1755 return rc; 1756 } 1757 1758 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1759 current_wr = &rdma_req->data.wr; 1760 assert(current_wr != NULL); 1761 1762 req->length = 0; 1763 rdma_req->iovpos = 0; 1764 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1765 for (i = 0; i < num_sgl_descriptors; i++) { 1766 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1767 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1768 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1769 rc = -EINVAL; 1770 goto err_exit; 1771 } 1772 1773 current_wr->num_sge = 0; 1774 1775 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1776 if (rc != 0) { 1777 rc = -ENOMEM; 1778 goto err_exit; 1779 } 1780 1781 req->length += desc->keyed.length; 1782 current_wr->wr.rdma.rkey = desc->keyed.key; 1783 current_wr->wr.rdma.remote_addr = desc->address; 1784 current_wr = current_wr->next; 1785 desc++; 1786 } 1787 1788 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1789 /* Go back to the last descriptor in the list. */ 1790 desc--; 1791 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1792 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1793 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1794 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1795 } 1796 } 1797 #endif 1798 1799 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1800 1801 return 0; 1802 1803 err_exit: 1804 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1805 nvmf_rdma_request_free_data(rdma_req, rtransport); 1806 return rc; 1807 } 1808 1809 static int 1810 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1811 struct spdk_nvmf_rdma_device *device, 1812 struct spdk_nvmf_rdma_request *rdma_req) 1813 { 1814 struct spdk_nvmf_request *req = &rdma_req->req; 1815 struct spdk_nvme_cpl *rsp; 1816 struct spdk_nvme_sgl_descriptor *sgl; 1817 int rc; 1818 uint32_t length; 1819 1820 rsp = &req->rsp->nvme_cpl; 1821 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1822 1823 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1824 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1825 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1826 1827 length = sgl->keyed.length; 1828 if (length > rtransport->transport.opts.max_io_size) { 1829 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1830 length, rtransport->transport.opts.max_io_size); 1831 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1832 return -1; 1833 } 1834 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1835 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1836 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1837 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1838 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1839 } 1840 } 1841 #endif 1842 1843 /* fill request length and populate iovs */ 1844 req->length = length; 1845 1846 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1847 req->dif.orig_length = length; 1848 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1849 req->dif.elba_length = length; 1850 } 1851 1852 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1853 if (spdk_unlikely(rc < 0)) { 1854 if (rc == -EINVAL) { 1855 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1856 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1857 return -1; 1858 } 1859 /* No available buffers. Queue this request up. */ 1860 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1861 return 0; 1862 } 1863 1864 /* backward compatible */ 1865 req->data = req->iov[0].iov_base; 1866 1867 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1868 req->iovcnt); 1869 1870 return 0; 1871 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1872 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1873 uint64_t offset = sgl->address; 1874 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1875 1876 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1877 offset, sgl->unkeyed.length); 1878 1879 if (offset > max_len) { 1880 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1881 offset, max_len); 1882 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1883 return -1; 1884 } 1885 max_len -= (uint32_t)offset; 1886 1887 if (sgl->unkeyed.length > max_len) { 1888 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1889 sgl->unkeyed.length, max_len); 1890 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1891 return -1; 1892 } 1893 1894 rdma_req->num_outstanding_data_wr = 0; 1895 req->data = rdma_req->recv->buf + offset; 1896 req->data_from_pool = false; 1897 req->length = sgl->unkeyed.length; 1898 1899 req->iov[0].iov_base = req->data; 1900 req->iov[0].iov_len = req->length; 1901 req->iovcnt = 1; 1902 1903 return 0; 1904 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1905 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1906 1907 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1908 if (rc == -ENOMEM) { 1909 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1910 return 0; 1911 } else if (rc == -EINVAL) { 1912 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1913 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1914 return -1; 1915 } 1916 1917 /* backward compatible */ 1918 req->data = req->iov[0].iov_base; 1919 1920 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1921 req->iovcnt); 1922 1923 return 0; 1924 } 1925 1926 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1927 sgl->generic.type, sgl->generic.subtype); 1928 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1929 return -1; 1930 } 1931 1932 static void 1933 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1934 struct spdk_nvmf_rdma_transport *rtransport) 1935 { 1936 struct spdk_nvmf_rdma_qpair *rqpair; 1937 struct spdk_nvmf_rdma_poll_group *rgroup; 1938 1939 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1940 if (rdma_req->req.data_from_pool) { 1941 rgroup = rqpair->poller->group; 1942 1943 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1944 } 1945 nvmf_rdma_request_free_data(rdma_req, rtransport); 1946 rdma_req->req.length = 0; 1947 rdma_req->req.iovcnt = 0; 1948 rdma_req->req.data = NULL; 1949 rdma_req->rsp.wr.next = NULL; 1950 rdma_req->data.wr.next = NULL; 1951 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1952 rqpair->qd--; 1953 1954 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1955 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1956 } 1957 1958 bool 1959 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1960 struct spdk_nvmf_rdma_request *rdma_req) 1961 { 1962 struct spdk_nvmf_rdma_qpair *rqpair; 1963 struct spdk_nvmf_rdma_device *device; 1964 struct spdk_nvmf_rdma_poll_group *rgroup; 1965 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1966 int rc; 1967 struct spdk_nvmf_rdma_recv *rdma_recv; 1968 enum spdk_nvmf_rdma_request_state prev_state; 1969 bool progress = false; 1970 int data_posted; 1971 uint32_t num_blocks; 1972 1973 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1974 device = rqpair->device; 1975 rgroup = rqpair->poller->group; 1976 1977 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1978 1979 /* If the queue pair is in an error state, force the request to the completed state 1980 * to release resources. */ 1981 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1982 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1983 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1984 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1985 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1986 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1987 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1988 } 1989 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1990 } 1991 1992 /* The loop here is to allow for several back-to-back state changes. */ 1993 do { 1994 prev_state = rdma_req->state; 1995 1996 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 1997 1998 switch (rdma_req->state) { 1999 case RDMA_REQUEST_STATE_FREE: 2000 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 2001 * to escape this state. */ 2002 break; 2003 case RDMA_REQUEST_STATE_NEW: 2004 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 2005 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2006 rdma_recv = rdma_req->recv; 2007 2008 /* The first element of the SGL is the NVMe command */ 2009 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 2010 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 2011 2012 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2013 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2014 break; 2015 } 2016 2017 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 2018 rdma_req->req.dif.dif_insert_or_strip = true; 2019 } 2020 2021 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2022 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 2023 rdma_req->rsp.wr.imm_data = 0; 2024 #endif 2025 2026 /* The next state transition depends on the data transfer needs of this request. */ 2027 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2028 2029 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2030 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2031 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2032 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2033 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 2034 break; 2035 } 2036 2037 /* If no data to transfer, ready to execute. */ 2038 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2039 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2040 break; 2041 } 2042 2043 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2044 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2045 break; 2046 case RDMA_REQUEST_STATE_NEED_BUFFER: 2047 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2048 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2049 2050 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2051 2052 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2053 /* This request needs to wait in line to obtain a buffer */ 2054 break; 2055 } 2056 2057 /* Try to get a data buffer */ 2058 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2059 if (rc < 0) { 2060 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2061 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2062 break; 2063 } 2064 2065 if (!rdma_req->req.data) { 2066 /* No buffers available. */ 2067 rgroup->stat.pending_data_buffer++; 2068 break; 2069 } 2070 2071 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2072 2073 /* If data is transferring from host to controller and the data didn't 2074 * arrive using in capsule data, we need to do a transfer from the host. 2075 */ 2076 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2077 rdma_req->req.data_from_pool) { 2078 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2079 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2080 break; 2081 } 2082 2083 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2084 break; 2085 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2086 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2087 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2088 2089 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2090 /* This request needs to wait in line to perform RDMA */ 2091 break; 2092 } 2093 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2094 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2095 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2096 rqpair->poller->stat.pending_rdma_read++; 2097 break; 2098 } 2099 2100 /* We have already verified that this request is the head of the queue. */ 2101 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2102 2103 rc = request_transfer_in(&rdma_req->req); 2104 if (!rc) { 2105 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2106 } else { 2107 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2108 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2109 } 2110 break; 2111 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2112 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2113 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2114 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2115 * to escape this state. */ 2116 break; 2117 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2118 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2119 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2120 2121 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2122 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2123 /* generate DIF for write operation */ 2124 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2125 assert(num_blocks > 0); 2126 2127 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2128 num_blocks, &rdma_req->req.dif.dif_ctx); 2129 if (rc != 0) { 2130 SPDK_ERRLOG("DIF generation failed\n"); 2131 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2132 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2133 break; 2134 } 2135 } 2136 2137 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2138 /* set extended length before IO operation */ 2139 rdma_req->req.length = rdma_req->req.dif.elba_length; 2140 } 2141 2142 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2143 spdk_nvmf_request_exec(&rdma_req->req); 2144 break; 2145 case RDMA_REQUEST_STATE_EXECUTING: 2146 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2147 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2148 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2149 * to escape this state. */ 2150 break; 2151 case RDMA_REQUEST_STATE_EXECUTED: 2152 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2153 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2154 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2155 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2156 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2157 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2158 } else { 2159 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2160 } 2161 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2162 /* restore the original length */ 2163 rdma_req->req.length = rdma_req->req.dif.orig_length; 2164 2165 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2166 struct spdk_dif_error error_blk; 2167 2168 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2169 2170 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2171 &rdma_req->req.dif.dif_ctx, &error_blk); 2172 if (rc) { 2173 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2174 2175 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2176 error_blk.err_offset); 2177 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2178 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2179 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2180 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2181 } 2182 } 2183 } 2184 break; 2185 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2186 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2187 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2188 2189 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2190 /* This request needs to wait in line to perform RDMA */ 2191 break; 2192 } 2193 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2194 rqpair->max_send_depth) { 2195 /* We can only have so many WRs outstanding. we have to wait until some finish. 2196 * +1 since each request has an additional wr in the resp. */ 2197 rqpair->poller->stat.pending_rdma_write++; 2198 break; 2199 } 2200 2201 /* We have already verified that this request is the head of the queue. */ 2202 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2203 2204 /* The data transfer will be kicked off from 2205 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2206 */ 2207 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2208 break; 2209 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2210 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2211 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2212 rc = request_transfer_out(&rdma_req->req, &data_posted); 2213 assert(rc == 0); /* No good way to handle this currently */ 2214 if (rc) { 2215 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2216 } else { 2217 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2218 RDMA_REQUEST_STATE_COMPLETING; 2219 } 2220 break; 2221 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2222 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2223 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2224 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2225 * to escape this state. */ 2226 break; 2227 case RDMA_REQUEST_STATE_COMPLETING: 2228 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2229 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2230 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2231 * to escape this state. */ 2232 break; 2233 case RDMA_REQUEST_STATE_COMPLETED: 2234 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2235 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2236 2237 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2238 _nvmf_rdma_request_free(rdma_req, rtransport); 2239 break; 2240 case RDMA_REQUEST_NUM_STATES: 2241 default: 2242 assert(0); 2243 break; 2244 } 2245 2246 if (rdma_req->state != prev_state) { 2247 progress = true; 2248 } 2249 } while (rdma_req->state != prev_state); 2250 2251 return progress; 2252 } 2253 2254 /* Public API callbacks begin here */ 2255 2256 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2257 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2258 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2259 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2260 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2261 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2262 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2263 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2264 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2265 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2266 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2267 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2268 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2269 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2270 2271 static void 2272 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2273 { 2274 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2275 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2276 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2277 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2278 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2279 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2280 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2281 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2282 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2283 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2284 opts->transport_specific = NULL; 2285 } 2286 2287 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2288 .notify_cb = nvmf_rdma_mem_notify, 2289 .are_contiguous = nvmf_rdma_check_contiguous_entries 2290 }; 2291 2292 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2293 2294 static struct spdk_nvmf_transport * 2295 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2296 { 2297 int rc; 2298 struct spdk_nvmf_rdma_transport *rtransport; 2299 struct spdk_nvmf_rdma_device *device, *tmp; 2300 struct ibv_context **contexts; 2301 uint32_t i; 2302 int flag; 2303 uint32_t sge_count; 2304 uint32_t min_shared_buffers; 2305 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2306 pthread_mutexattr_t attr; 2307 2308 rtransport = calloc(1, sizeof(*rtransport)); 2309 if (!rtransport) { 2310 return NULL; 2311 } 2312 2313 if (pthread_mutexattr_init(&attr)) { 2314 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2315 free(rtransport); 2316 return NULL; 2317 } 2318 2319 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2320 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2321 pthread_mutexattr_destroy(&attr); 2322 free(rtransport); 2323 return NULL; 2324 } 2325 2326 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2327 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2328 pthread_mutexattr_destroy(&attr); 2329 free(rtransport); 2330 return NULL; 2331 } 2332 2333 pthread_mutexattr_destroy(&attr); 2334 2335 TAILQ_INIT(&rtransport->devices); 2336 TAILQ_INIT(&rtransport->ports); 2337 TAILQ_INIT(&rtransport->poll_groups); 2338 2339 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2340 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2341 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2342 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2343 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2344 if (opts->transport_specific != NULL && 2345 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2346 SPDK_COUNTOF(rdma_transport_opts_decoder), 2347 &rtransport->rdma_opts)) { 2348 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2349 nvmf_rdma_destroy(&rtransport->transport); 2350 return NULL; 2351 } 2352 2353 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2354 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2355 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2356 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2357 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d," 2358 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2359 opts->max_queue_depth, 2360 opts->max_io_size, 2361 opts->max_qpairs_per_ctrlr - 1, 2362 opts->io_unit_size, 2363 opts->in_capsule_data_size, 2364 opts->max_aq_depth, 2365 opts->num_shared_buffers, 2366 rtransport->rdma_opts.max_srq_depth, 2367 rtransport->rdma_opts.no_srq, 2368 rtransport->rdma_opts.acceptor_backlog, 2369 rtransport->rdma_opts.no_wr_batching, 2370 opts->abort_timeout_sec); 2371 2372 /* I/O unit size cannot be larger than max I/O size */ 2373 if (opts->io_unit_size > opts->max_io_size) { 2374 opts->io_unit_size = opts->max_io_size; 2375 } 2376 2377 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2378 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2379 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2380 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2381 } 2382 2383 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2384 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2385 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2386 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2387 nvmf_rdma_destroy(&rtransport->transport); 2388 return NULL; 2389 } 2390 2391 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2392 if (min_shared_buffers > opts->num_shared_buffers) { 2393 SPDK_ERRLOG("There are not enough buffers to satisfy" 2394 "per-poll group caches for each thread. (%" PRIu32 ")" 2395 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2396 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2397 nvmf_rdma_destroy(&rtransport->transport); 2398 return NULL; 2399 } 2400 2401 sge_count = opts->max_io_size / opts->io_unit_size; 2402 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2403 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2404 nvmf_rdma_destroy(&rtransport->transport); 2405 return NULL; 2406 } 2407 2408 rtransport->event_channel = rdma_create_event_channel(); 2409 if (rtransport->event_channel == NULL) { 2410 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2411 nvmf_rdma_destroy(&rtransport->transport); 2412 return NULL; 2413 } 2414 2415 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2416 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2417 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2418 rtransport->event_channel->fd, spdk_strerror(errno)); 2419 nvmf_rdma_destroy(&rtransport->transport); 2420 return NULL; 2421 } 2422 2423 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2424 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2425 sizeof(struct spdk_nvmf_rdma_request_data), 2426 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2427 SPDK_ENV_SOCKET_ID_ANY); 2428 if (!rtransport->data_wr_pool) { 2429 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2430 nvmf_rdma_destroy(&rtransport->transport); 2431 return NULL; 2432 } 2433 2434 contexts = rdma_get_devices(NULL); 2435 if (contexts == NULL) { 2436 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2437 nvmf_rdma_destroy(&rtransport->transport); 2438 return NULL; 2439 } 2440 2441 i = 0; 2442 rc = 0; 2443 while (contexts[i] != NULL) { 2444 device = calloc(1, sizeof(*device)); 2445 if (!device) { 2446 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2447 rc = -ENOMEM; 2448 break; 2449 } 2450 device->context = contexts[i]; 2451 rc = ibv_query_device(device->context, &device->attr); 2452 if (rc < 0) { 2453 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2454 free(device); 2455 break; 2456 2457 } 2458 2459 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2460 2461 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2462 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2463 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2464 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2465 } 2466 2467 /** 2468 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2469 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2470 * but incorrectly reports that it does. There are changes making their way 2471 * through the kernel now that will enable this feature. When they are merged, 2472 * we can conditionally enable this feature. 2473 * 2474 * TODO: enable this for versions of the kernel rxe driver that support it. 2475 */ 2476 if (device->attr.vendor_id == 0) { 2477 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2478 } 2479 #endif 2480 2481 /* set up device context async ev fd as NON_BLOCKING */ 2482 flag = fcntl(device->context->async_fd, F_GETFL); 2483 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2484 if (rc < 0) { 2485 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2486 free(device); 2487 break; 2488 } 2489 2490 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2491 i++; 2492 2493 if (g_nvmf_hooks.get_ibv_pd) { 2494 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2495 } else { 2496 device->pd = ibv_alloc_pd(device->context); 2497 } 2498 2499 if (!device->pd) { 2500 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2501 rc = -ENOMEM; 2502 break; 2503 } 2504 2505 assert(device->map == NULL); 2506 2507 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2508 if (!device->map) { 2509 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2510 rc = -ENOMEM; 2511 break; 2512 } 2513 2514 assert(device->map != NULL); 2515 assert(device->pd != NULL); 2516 } 2517 rdma_free_devices(contexts); 2518 2519 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2520 /* divide and round up. */ 2521 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2522 2523 /* round up to the nearest 4k. */ 2524 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2525 2526 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2527 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2528 opts->io_unit_size); 2529 } 2530 2531 if (rc < 0) { 2532 nvmf_rdma_destroy(&rtransport->transport); 2533 return NULL; 2534 } 2535 2536 /* Set up poll descriptor array to monitor events from RDMA and IB 2537 * in a single poll syscall 2538 */ 2539 rtransport->npoll_fds = i + 1; 2540 i = 0; 2541 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2542 if (rtransport->poll_fds == NULL) { 2543 SPDK_ERRLOG("poll_fds allocation failed\n"); 2544 nvmf_rdma_destroy(&rtransport->transport); 2545 return NULL; 2546 } 2547 2548 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2549 rtransport->poll_fds[i++].events = POLLIN; 2550 2551 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2552 rtransport->poll_fds[i].fd = device->context->async_fd; 2553 rtransport->poll_fds[i++].events = POLLIN; 2554 } 2555 2556 return &rtransport->transport; 2557 } 2558 2559 static void 2560 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2561 { 2562 struct spdk_nvmf_rdma_transport *rtransport; 2563 assert(w != NULL); 2564 2565 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2566 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2567 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2568 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2569 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2570 } 2571 2572 static int 2573 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2574 { 2575 struct spdk_nvmf_rdma_transport *rtransport; 2576 struct spdk_nvmf_rdma_port *port, *port_tmp; 2577 struct spdk_nvmf_rdma_device *device, *device_tmp; 2578 2579 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2580 2581 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2582 TAILQ_REMOVE(&rtransport->ports, port, link); 2583 rdma_destroy_id(port->id); 2584 free(port); 2585 } 2586 2587 if (rtransport->poll_fds != NULL) { 2588 free(rtransport->poll_fds); 2589 } 2590 2591 if (rtransport->event_channel != NULL) { 2592 rdma_destroy_event_channel(rtransport->event_channel); 2593 } 2594 2595 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2596 TAILQ_REMOVE(&rtransport->devices, device, link); 2597 if (device->map) { 2598 spdk_mem_map_free(&device->map); 2599 } 2600 if (device->pd) { 2601 if (!g_nvmf_hooks.get_ibv_pd) { 2602 ibv_dealloc_pd(device->pd); 2603 } 2604 } 2605 free(device); 2606 } 2607 2608 if (rtransport->data_wr_pool != NULL) { 2609 if (spdk_mempool_count(rtransport->data_wr_pool) != 2610 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2611 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2612 spdk_mempool_count(rtransport->data_wr_pool), 2613 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2614 } 2615 } 2616 2617 spdk_mempool_free(rtransport->data_wr_pool); 2618 2619 pthread_mutex_destroy(&rtransport->lock); 2620 free(rtransport); 2621 2622 return 0; 2623 } 2624 2625 static int 2626 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2627 struct spdk_nvme_transport_id *trid, 2628 bool peer); 2629 2630 static int 2631 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2632 const struct spdk_nvme_transport_id *trid) 2633 { 2634 struct spdk_nvmf_rdma_transport *rtransport; 2635 struct spdk_nvmf_rdma_device *device; 2636 struct spdk_nvmf_rdma_port *port; 2637 struct addrinfo *res; 2638 struct addrinfo hints; 2639 int family; 2640 int rc; 2641 2642 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2643 assert(rtransport->event_channel != NULL); 2644 2645 pthread_mutex_lock(&rtransport->lock); 2646 port = calloc(1, sizeof(*port)); 2647 if (!port) { 2648 SPDK_ERRLOG("Port allocation failed\n"); 2649 pthread_mutex_unlock(&rtransport->lock); 2650 return -ENOMEM; 2651 } 2652 2653 port->trid = trid; 2654 2655 switch (trid->adrfam) { 2656 case SPDK_NVMF_ADRFAM_IPV4: 2657 family = AF_INET; 2658 break; 2659 case SPDK_NVMF_ADRFAM_IPV6: 2660 family = AF_INET6; 2661 break; 2662 default: 2663 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2664 free(port); 2665 pthread_mutex_unlock(&rtransport->lock); 2666 return -EINVAL; 2667 } 2668 2669 memset(&hints, 0, sizeof(hints)); 2670 hints.ai_family = family; 2671 hints.ai_flags = AI_NUMERICSERV; 2672 hints.ai_socktype = SOCK_STREAM; 2673 hints.ai_protocol = 0; 2674 2675 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2676 if (rc) { 2677 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2678 free(port); 2679 pthread_mutex_unlock(&rtransport->lock); 2680 return -EINVAL; 2681 } 2682 2683 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2684 if (rc < 0) { 2685 SPDK_ERRLOG("rdma_create_id() failed\n"); 2686 freeaddrinfo(res); 2687 free(port); 2688 pthread_mutex_unlock(&rtransport->lock); 2689 return rc; 2690 } 2691 2692 rc = rdma_bind_addr(port->id, res->ai_addr); 2693 freeaddrinfo(res); 2694 2695 if (rc < 0) { 2696 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2697 rdma_destroy_id(port->id); 2698 free(port); 2699 pthread_mutex_unlock(&rtransport->lock); 2700 return rc; 2701 } 2702 2703 if (!port->id->verbs) { 2704 SPDK_ERRLOG("ibv_context is null\n"); 2705 rdma_destroy_id(port->id); 2706 free(port); 2707 pthread_mutex_unlock(&rtransport->lock); 2708 return -1; 2709 } 2710 2711 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2712 if (rc < 0) { 2713 SPDK_ERRLOG("rdma_listen() failed\n"); 2714 rdma_destroy_id(port->id); 2715 free(port); 2716 pthread_mutex_unlock(&rtransport->lock); 2717 return rc; 2718 } 2719 2720 TAILQ_FOREACH(device, &rtransport->devices, link) { 2721 if (device->context == port->id->verbs) { 2722 port->device = device; 2723 break; 2724 } 2725 } 2726 if (!port->device) { 2727 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2728 port->id->verbs); 2729 rdma_destroy_id(port->id); 2730 free(port); 2731 pthread_mutex_unlock(&rtransport->lock); 2732 return -EINVAL; 2733 } 2734 2735 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2736 trid->traddr, trid->trsvcid); 2737 2738 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2739 pthread_mutex_unlock(&rtransport->lock); 2740 return 0; 2741 } 2742 2743 static void 2744 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2745 const struct spdk_nvme_transport_id *trid) 2746 { 2747 struct spdk_nvmf_rdma_transport *rtransport; 2748 struct spdk_nvmf_rdma_port *port, *tmp; 2749 2750 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2751 2752 pthread_mutex_lock(&rtransport->lock); 2753 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2754 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2755 TAILQ_REMOVE(&rtransport->ports, port, link); 2756 rdma_destroy_id(port->id); 2757 free(port); 2758 break; 2759 } 2760 } 2761 2762 pthread_mutex_unlock(&rtransport->lock); 2763 } 2764 2765 static void 2766 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2767 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2768 { 2769 struct spdk_nvmf_request *req, *tmp; 2770 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2771 struct spdk_nvmf_rdma_resources *resources; 2772 2773 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2774 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2775 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2776 break; 2777 } 2778 } 2779 2780 /* Then RDMA writes since reads have stronger restrictions than writes */ 2781 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2782 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2783 break; 2784 } 2785 } 2786 2787 /* The second highest priority is I/O waiting on memory buffers. */ 2788 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2789 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2790 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2791 break; 2792 } 2793 } 2794 2795 resources = rqpair->resources; 2796 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2797 rdma_req = STAILQ_FIRST(&resources->free_queue); 2798 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2799 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2800 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2801 2802 if (rqpair->srq != NULL) { 2803 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2804 rdma_req->recv->qpair->qd++; 2805 } else { 2806 rqpair->qd++; 2807 } 2808 2809 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2810 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2811 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2812 break; 2813 } 2814 } 2815 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2816 rqpair->poller->stat.pending_free_request++; 2817 } 2818 } 2819 2820 static void 2821 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2822 { 2823 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2824 struct spdk_nvmf_rdma_transport, transport); 2825 2826 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2827 2828 /* nvmr_rdma_close_qpair is not called */ 2829 if (!rqpair->to_close) { 2830 return; 2831 } 2832 2833 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2834 if (rqpair->current_send_depth != 0) { 2835 return; 2836 } 2837 2838 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2839 return; 2840 } 2841 2842 /* Judge whether the device is emulated by Software RoCE. 2843 * And it will not send last_wqe event 2844 */ 2845 if (rqpair->srq != NULL && rqpair->device->attr.vendor_id != 0 && 2846 rqpair->last_wqe_reached == false) { 2847 return; 2848 } 2849 2850 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2851 2852 nvmf_rdma_qpair_destroy(rqpair); 2853 } 2854 2855 static int 2856 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2857 { 2858 struct spdk_nvmf_qpair *qpair; 2859 struct spdk_nvmf_rdma_qpair *rqpair; 2860 2861 if (evt->id == NULL) { 2862 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2863 return -1; 2864 } 2865 2866 qpair = evt->id->context; 2867 if (qpair == NULL) { 2868 SPDK_ERRLOG("disconnect request: no active connection\n"); 2869 return -1; 2870 } 2871 2872 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2873 2874 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2875 2876 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2877 2878 return 0; 2879 } 2880 2881 #ifdef DEBUG 2882 static const char *CM_EVENT_STR[] = { 2883 "RDMA_CM_EVENT_ADDR_RESOLVED", 2884 "RDMA_CM_EVENT_ADDR_ERROR", 2885 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2886 "RDMA_CM_EVENT_ROUTE_ERROR", 2887 "RDMA_CM_EVENT_CONNECT_REQUEST", 2888 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2889 "RDMA_CM_EVENT_CONNECT_ERROR", 2890 "RDMA_CM_EVENT_UNREACHABLE", 2891 "RDMA_CM_EVENT_REJECTED", 2892 "RDMA_CM_EVENT_ESTABLISHED", 2893 "RDMA_CM_EVENT_DISCONNECTED", 2894 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2895 "RDMA_CM_EVENT_MULTICAST_JOIN", 2896 "RDMA_CM_EVENT_MULTICAST_ERROR", 2897 "RDMA_CM_EVENT_ADDR_CHANGE", 2898 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2899 }; 2900 #endif /* DEBUG */ 2901 2902 static void 2903 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2904 struct spdk_nvmf_rdma_port *port) 2905 { 2906 struct spdk_nvmf_rdma_poll_group *rgroup; 2907 struct spdk_nvmf_rdma_poller *rpoller; 2908 struct spdk_nvmf_rdma_qpair *rqpair; 2909 2910 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2911 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2912 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 2913 if (rqpair->listen_id == port->id) { 2914 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2915 } 2916 } 2917 } 2918 } 2919 } 2920 2921 static bool 2922 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2923 struct rdma_cm_event *event) 2924 { 2925 const struct spdk_nvme_transport_id *trid; 2926 struct spdk_nvmf_rdma_port *port; 2927 struct spdk_nvmf_rdma_transport *rtransport; 2928 bool event_acked = false; 2929 2930 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2931 TAILQ_FOREACH(port, &rtransport->ports, link) { 2932 if (port->id == event->id) { 2933 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2934 rdma_ack_cm_event(event); 2935 event_acked = true; 2936 trid = port->trid; 2937 break; 2938 } 2939 } 2940 2941 if (event_acked) { 2942 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2943 2944 nvmf_rdma_stop_listen(transport, trid); 2945 nvmf_rdma_listen(transport, trid); 2946 } 2947 2948 return event_acked; 2949 } 2950 2951 static void 2952 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2953 struct rdma_cm_event *event) 2954 { 2955 struct spdk_nvmf_rdma_port *port; 2956 struct spdk_nvmf_rdma_transport *rtransport; 2957 2958 port = event->id->context; 2959 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2960 2961 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2962 2963 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2964 2965 rdma_ack_cm_event(event); 2966 2967 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2968 ; 2969 } 2970 } 2971 2972 static void 2973 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2974 { 2975 struct spdk_nvmf_rdma_transport *rtransport; 2976 struct rdma_cm_event *event; 2977 int rc; 2978 bool event_acked; 2979 2980 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2981 2982 if (rtransport->event_channel == NULL) { 2983 return; 2984 } 2985 2986 while (1) { 2987 event_acked = false; 2988 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2989 if (rc) { 2990 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2991 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2992 } 2993 break; 2994 } 2995 2996 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2997 2998 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2999 3000 switch (event->event) { 3001 case RDMA_CM_EVENT_ADDR_RESOLVED: 3002 case RDMA_CM_EVENT_ADDR_ERROR: 3003 case RDMA_CM_EVENT_ROUTE_RESOLVED: 3004 case RDMA_CM_EVENT_ROUTE_ERROR: 3005 /* No action required. The target never attempts to resolve routes. */ 3006 break; 3007 case RDMA_CM_EVENT_CONNECT_REQUEST: 3008 rc = nvmf_rdma_connect(transport, event); 3009 if (rc < 0) { 3010 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 3011 break; 3012 } 3013 break; 3014 case RDMA_CM_EVENT_CONNECT_RESPONSE: 3015 /* The target never initiates a new connection. So this will not occur. */ 3016 break; 3017 case RDMA_CM_EVENT_CONNECT_ERROR: 3018 /* Can this happen? The docs say it can, but not sure what causes it. */ 3019 break; 3020 case RDMA_CM_EVENT_UNREACHABLE: 3021 case RDMA_CM_EVENT_REJECTED: 3022 /* These only occur on the client side. */ 3023 break; 3024 case RDMA_CM_EVENT_ESTABLISHED: 3025 /* TODO: Should we be waiting for this event anywhere? */ 3026 break; 3027 case RDMA_CM_EVENT_DISCONNECTED: 3028 rc = nvmf_rdma_disconnect(event); 3029 if (rc < 0) { 3030 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3031 break; 3032 } 3033 break; 3034 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3035 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3036 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3037 * Once these events are sent to SPDK, we should release all IB resources and 3038 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3039 * ibv_destory* functions to release user space memory allocated by IB. All kernel 3040 * resources are already cleaned. */ 3041 if (event->id->qp) { 3042 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3043 * corresponding qpair. Otherwise the event refers to a listening device */ 3044 rc = nvmf_rdma_disconnect(event); 3045 if (rc < 0) { 3046 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3047 break; 3048 } 3049 } else { 3050 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3051 event_acked = true; 3052 } 3053 break; 3054 case RDMA_CM_EVENT_MULTICAST_JOIN: 3055 case RDMA_CM_EVENT_MULTICAST_ERROR: 3056 /* Multicast is not used */ 3057 break; 3058 case RDMA_CM_EVENT_ADDR_CHANGE: 3059 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3060 break; 3061 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3062 /* For now, do nothing. The target never re-uses queue pairs. */ 3063 break; 3064 default: 3065 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3066 break; 3067 } 3068 if (!event_acked) { 3069 rdma_ack_cm_event(event); 3070 } 3071 } 3072 } 3073 3074 static void 3075 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3076 { 3077 rqpair->last_wqe_reached = true; 3078 nvmf_rdma_destroy_drained_qpair(rqpair); 3079 } 3080 3081 static void 3082 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3083 { 3084 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3085 3086 if (event_ctx->rqpair) { 3087 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3088 if (event_ctx->cb_fn) { 3089 event_ctx->cb_fn(event_ctx->rqpair); 3090 } 3091 } 3092 free(event_ctx); 3093 } 3094 3095 static int 3096 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3097 spdk_nvmf_rdma_qpair_ibv_event fn) 3098 { 3099 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3100 struct spdk_thread *thr = NULL; 3101 int rc; 3102 3103 if (rqpair->qpair.group) { 3104 thr = rqpair->qpair.group->thread; 3105 } else if (rqpair->destruct_channel) { 3106 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3107 } 3108 3109 if (!thr) { 3110 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3111 return -EINVAL; 3112 } 3113 3114 ctx = calloc(1, sizeof(*ctx)); 3115 if (!ctx) { 3116 return -ENOMEM; 3117 } 3118 3119 ctx->rqpair = rqpair; 3120 ctx->cb_fn = fn; 3121 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3122 3123 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3124 if (rc) { 3125 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3126 free(ctx); 3127 } 3128 3129 return rc; 3130 } 3131 3132 static int 3133 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3134 { 3135 int rc; 3136 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3137 struct ibv_async_event event; 3138 3139 rc = ibv_get_async_event(device->context, &event); 3140 3141 if (rc) { 3142 /* In non-blocking mode -1 means there are no events available */ 3143 return rc; 3144 } 3145 3146 switch (event.event_type) { 3147 case IBV_EVENT_QP_FATAL: 3148 rqpair = event.element.qp->qp_context; 3149 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3150 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3151 (uintptr_t)rqpair->cm_id, event.event_type); 3152 nvmf_rdma_update_ibv_state(rqpair); 3153 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3154 break; 3155 case IBV_EVENT_QP_LAST_WQE_REACHED: 3156 /* This event only occurs for shared receive queues. */ 3157 rqpair = event.element.qp->qp_context; 3158 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3159 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3160 if (rc) { 3161 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3162 rqpair->last_wqe_reached = true; 3163 } 3164 break; 3165 case IBV_EVENT_SQ_DRAINED: 3166 /* This event occurs frequently in both error and non-error states. 3167 * Check if the qpair is in an error state before sending a message. */ 3168 rqpair = event.element.qp->qp_context; 3169 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3170 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3171 (uintptr_t)rqpair->cm_id, event.event_type); 3172 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3173 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3174 } 3175 break; 3176 case IBV_EVENT_QP_REQ_ERR: 3177 case IBV_EVENT_QP_ACCESS_ERR: 3178 case IBV_EVENT_COMM_EST: 3179 case IBV_EVENT_PATH_MIG: 3180 case IBV_EVENT_PATH_MIG_ERR: 3181 SPDK_NOTICELOG("Async event: %s\n", 3182 ibv_event_type_str(event.event_type)); 3183 rqpair = event.element.qp->qp_context; 3184 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3185 (uintptr_t)rqpair->cm_id, event.event_type); 3186 nvmf_rdma_update_ibv_state(rqpair); 3187 break; 3188 case IBV_EVENT_CQ_ERR: 3189 case IBV_EVENT_DEVICE_FATAL: 3190 case IBV_EVENT_PORT_ACTIVE: 3191 case IBV_EVENT_PORT_ERR: 3192 case IBV_EVENT_LID_CHANGE: 3193 case IBV_EVENT_PKEY_CHANGE: 3194 case IBV_EVENT_SM_CHANGE: 3195 case IBV_EVENT_SRQ_ERR: 3196 case IBV_EVENT_SRQ_LIMIT_REACHED: 3197 case IBV_EVENT_CLIENT_REREGISTER: 3198 case IBV_EVENT_GID_CHANGE: 3199 default: 3200 SPDK_NOTICELOG("Async event: %s\n", 3201 ibv_event_type_str(event.event_type)); 3202 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3203 break; 3204 } 3205 ibv_ack_async_event(&event); 3206 3207 return 0; 3208 } 3209 3210 static void 3211 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3212 { 3213 int rc = 0; 3214 uint32_t i = 0; 3215 3216 for (i = 0; i < max_events; i++) { 3217 rc = nvmf_process_ib_event(device); 3218 if (rc) { 3219 break; 3220 } 3221 } 3222 3223 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3224 } 3225 3226 static uint32_t 3227 nvmf_rdma_accept(struct spdk_nvmf_transport *transport) 3228 { 3229 int nfds, i = 0; 3230 struct spdk_nvmf_rdma_transport *rtransport; 3231 struct spdk_nvmf_rdma_device *device, *tmp; 3232 uint32_t count; 3233 3234 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3235 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3236 3237 if (nfds <= 0) { 3238 return 0; 3239 } 3240 3241 /* The first poll descriptor is RDMA CM event */ 3242 if (rtransport->poll_fds[i++].revents & POLLIN) { 3243 nvmf_process_cm_event(transport); 3244 nfds--; 3245 } 3246 3247 if (nfds == 0) { 3248 return count; 3249 } 3250 3251 /* Second and subsequent poll descriptors are IB async events */ 3252 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3253 if (rtransport->poll_fds[i++].revents & POLLIN) { 3254 nvmf_process_ib_events(device, 32); 3255 nfds--; 3256 } 3257 } 3258 /* check all flagged fd's have been served */ 3259 assert(nfds == 0); 3260 3261 return count; 3262 } 3263 3264 static void 3265 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3266 struct spdk_nvmf_ctrlr_data *cdata) 3267 { 3268 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3269 3270 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3271 since in-capsule data only works with NVME drives that support SGL memory layout */ 3272 if (transport->opts.dif_insert_or_strip) { 3273 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3274 } 3275 } 3276 3277 static void 3278 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3279 struct spdk_nvme_transport_id *trid, 3280 struct spdk_nvmf_discovery_log_page_entry *entry) 3281 { 3282 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3283 entry->adrfam = trid->adrfam; 3284 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3285 3286 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3287 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3288 3289 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3290 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3291 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3292 } 3293 3294 static void 3295 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3296 3297 static struct spdk_nvmf_transport_poll_group * 3298 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3299 { 3300 struct spdk_nvmf_rdma_transport *rtransport; 3301 struct spdk_nvmf_rdma_poll_group *rgroup; 3302 struct spdk_nvmf_rdma_poller *poller; 3303 struct spdk_nvmf_rdma_device *device; 3304 struct ibv_srq_init_attr srq_init_attr; 3305 struct spdk_nvmf_rdma_resource_opts opts; 3306 int num_cqe; 3307 3308 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3309 3310 rgroup = calloc(1, sizeof(*rgroup)); 3311 if (!rgroup) { 3312 return NULL; 3313 } 3314 3315 TAILQ_INIT(&rgroup->pollers); 3316 STAILQ_INIT(&rgroup->retired_bufs); 3317 3318 pthread_mutex_lock(&rtransport->lock); 3319 TAILQ_FOREACH(device, &rtransport->devices, link) { 3320 poller = calloc(1, sizeof(*poller)); 3321 if (!poller) { 3322 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3323 nvmf_rdma_poll_group_destroy(&rgroup->group); 3324 pthread_mutex_unlock(&rtransport->lock); 3325 return NULL; 3326 } 3327 3328 poller->device = device; 3329 poller->group = rgroup; 3330 3331 TAILQ_INIT(&poller->qpairs); 3332 STAILQ_INIT(&poller->qpairs_pending_send); 3333 STAILQ_INIT(&poller->qpairs_pending_recv); 3334 3335 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3336 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3337 poller->max_srq_depth = rtransport->rdma_opts.max_srq_depth; 3338 3339 device->num_srq++; 3340 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 3341 srq_init_attr.attr.max_wr = poller->max_srq_depth; 3342 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3343 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 3344 if (!poller->srq) { 3345 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3346 nvmf_rdma_poll_group_destroy(&rgroup->group); 3347 pthread_mutex_unlock(&rtransport->lock); 3348 return NULL; 3349 } 3350 3351 opts.qp = poller->srq; 3352 opts.pd = device->pd; 3353 opts.qpair = NULL; 3354 opts.shared = true; 3355 opts.max_queue_depth = poller->max_srq_depth; 3356 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3357 3358 poller->resources = nvmf_rdma_resources_create(&opts); 3359 if (!poller->resources) { 3360 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3361 nvmf_rdma_poll_group_destroy(&rgroup->group); 3362 pthread_mutex_unlock(&rtransport->lock); 3363 return NULL; 3364 } 3365 } 3366 3367 /* 3368 * When using an srq, we can limit the completion queue at startup. 3369 * The following formula represents the calculation: 3370 * num_cqe = num_recv + num_data_wr + num_send_wr. 3371 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3372 */ 3373 if (poller->srq) { 3374 num_cqe = poller->max_srq_depth * 3; 3375 } else { 3376 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3377 } 3378 3379 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3380 if (!poller->cq) { 3381 SPDK_ERRLOG("Unable to create completion queue\n"); 3382 nvmf_rdma_poll_group_destroy(&rgroup->group); 3383 pthread_mutex_unlock(&rtransport->lock); 3384 return NULL; 3385 } 3386 poller->num_cqe = num_cqe; 3387 } 3388 3389 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3390 if (rtransport->conn_sched.next_admin_pg == NULL) { 3391 rtransport->conn_sched.next_admin_pg = rgroup; 3392 rtransport->conn_sched.next_io_pg = rgroup; 3393 } 3394 3395 pthread_mutex_unlock(&rtransport->lock); 3396 return &rgroup->group; 3397 } 3398 3399 static struct spdk_nvmf_transport_poll_group * 3400 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3401 { 3402 struct spdk_nvmf_rdma_transport *rtransport; 3403 struct spdk_nvmf_rdma_poll_group **pg; 3404 struct spdk_nvmf_transport_poll_group *result; 3405 3406 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3407 3408 pthread_mutex_lock(&rtransport->lock); 3409 3410 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3411 pthread_mutex_unlock(&rtransport->lock); 3412 return NULL; 3413 } 3414 3415 if (qpair->qid == 0) { 3416 pg = &rtransport->conn_sched.next_admin_pg; 3417 } else { 3418 pg = &rtransport->conn_sched.next_io_pg; 3419 } 3420 3421 assert(*pg != NULL); 3422 3423 result = &(*pg)->group; 3424 3425 *pg = TAILQ_NEXT(*pg, link); 3426 if (*pg == NULL) { 3427 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3428 } 3429 3430 pthread_mutex_unlock(&rtransport->lock); 3431 3432 return result; 3433 } 3434 3435 static void 3436 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3437 { 3438 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3439 struct spdk_nvmf_rdma_poller *poller, *tmp; 3440 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3441 struct spdk_nvmf_transport_pg_cache_buf *buf, *tmp_buf; 3442 struct spdk_nvmf_rdma_transport *rtransport; 3443 3444 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3445 if (!rgroup) { 3446 return; 3447 } 3448 3449 /* free all retired buffers back to the transport so we don't short the mempool. */ 3450 STAILQ_FOREACH_SAFE(buf, &rgroup->retired_bufs, link, tmp_buf) { 3451 STAILQ_REMOVE(&rgroup->retired_bufs, buf, spdk_nvmf_transport_pg_cache_buf, link); 3452 assert(group->transport != NULL); 3453 spdk_mempool_put(group->transport->data_buf_pool, buf); 3454 } 3455 3456 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3457 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3458 3459 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3460 nvmf_rdma_qpair_destroy(qpair); 3461 } 3462 3463 if (poller->srq) { 3464 if (poller->resources) { 3465 nvmf_rdma_resources_destroy(poller->resources); 3466 } 3467 ibv_destroy_srq(poller->srq); 3468 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 3469 } 3470 3471 if (poller->cq) { 3472 ibv_destroy_cq(poller->cq); 3473 } 3474 3475 free(poller); 3476 } 3477 3478 if (rgroup->group.transport == NULL) { 3479 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3480 * calls this function directly in a failure path. */ 3481 free(rgroup); 3482 return; 3483 } 3484 3485 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3486 3487 pthread_mutex_lock(&rtransport->lock); 3488 next_rgroup = TAILQ_NEXT(rgroup, link); 3489 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3490 if (next_rgroup == NULL) { 3491 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3492 } 3493 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3494 rtransport->conn_sched.next_admin_pg = next_rgroup; 3495 } 3496 if (rtransport->conn_sched.next_io_pg == rgroup) { 3497 rtransport->conn_sched.next_io_pg = next_rgroup; 3498 } 3499 pthread_mutex_unlock(&rtransport->lock); 3500 3501 free(rgroup); 3502 } 3503 3504 static void 3505 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3506 { 3507 if (rqpair->cm_id != NULL) { 3508 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3509 } 3510 } 3511 3512 static int 3513 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3514 struct spdk_nvmf_qpair *qpair) 3515 { 3516 struct spdk_nvmf_rdma_poll_group *rgroup; 3517 struct spdk_nvmf_rdma_qpair *rqpair; 3518 struct spdk_nvmf_rdma_device *device; 3519 struct spdk_nvmf_rdma_poller *poller; 3520 int rc; 3521 3522 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3523 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3524 3525 device = rqpair->device; 3526 3527 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3528 if (poller->device == device) { 3529 break; 3530 } 3531 } 3532 3533 if (!poller) { 3534 SPDK_ERRLOG("No poller found for device.\n"); 3535 return -1; 3536 } 3537 3538 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3539 rqpair->poller = poller; 3540 rqpair->srq = rqpair->poller->srq; 3541 3542 rc = nvmf_rdma_qpair_initialize(qpair); 3543 if (rc < 0) { 3544 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3545 return -1; 3546 } 3547 3548 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3549 if (rc) { 3550 /* Try to reject, but we probably can't */ 3551 nvmf_rdma_qpair_reject_connection(rqpair); 3552 return -1; 3553 } 3554 3555 nvmf_rdma_update_ibv_state(rqpair); 3556 3557 return 0; 3558 } 3559 3560 static int 3561 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3562 struct spdk_nvmf_qpair *qpair) 3563 { 3564 struct spdk_nvmf_rdma_qpair *rqpair; 3565 3566 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3567 assert(group->transport->tgt != NULL); 3568 3569 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3570 3571 if (!rqpair->destruct_channel) { 3572 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3573 return 0; 3574 } 3575 3576 /* Sanity check that we get io_channel on the correct thread */ 3577 if (qpair->group) { 3578 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3579 } 3580 3581 return 0; 3582 } 3583 3584 static int 3585 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3586 { 3587 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3588 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3589 struct spdk_nvmf_rdma_transport, transport); 3590 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3591 struct spdk_nvmf_rdma_qpair, qpair); 3592 3593 /* 3594 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3595 * needs to be returned to the shared receive queue or the poll group will eventually be 3596 * starved of RECV structures. 3597 */ 3598 if (rqpair->srq && rdma_req->recv) { 3599 int rc; 3600 struct ibv_recv_wr *bad_recv_wr; 3601 3602 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 3603 if (rc) { 3604 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3605 } 3606 } 3607 3608 _nvmf_rdma_request_free(rdma_req, rtransport); 3609 return 0; 3610 } 3611 3612 static int 3613 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3614 { 3615 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3616 struct spdk_nvmf_rdma_transport, transport); 3617 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3618 struct spdk_nvmf_rdma_request, req); 3619 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3620 struct spdk_nvmf_rdma_qpair, qpair); 3621 3622 if (rqpair->ibv_state != IBV_QPS_ERR) { 3623 /* The connection is alive, so process the request as normal */ 3624 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3625 } else { 3626 /* The connection is dead. Move the request directly to the completed state. */ 3627 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3628 } 3629 3630 nvmf_rdma_request_process(rtransport, rdma_req); 3631 3632 return 0; 3633 } 3634 3635 static void 3636 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3637 { 3638 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3639 3640 rqpair->to_close = true; 3641 3642 /* This happens only when the qpair is disconnected before 3643 * it is added to the poll group. Since there is no poll group, 3644 * the RDMA qp has not been initialized yet and the RDMA CM 3645 * event has not yet been acknowledged, so we need to reject it. 3646 */ 3647 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3648 nvmf_rdma_qpair_reject_connection(rqpair); 3649 nvmf_rdma_qpair_destroy(rqpair); 3650 return; 3651 } 3652 3653 if (rqpair->rdma_qp) { 3654 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3655 } 3656 3657 nvmf_rdma_destroy_drained_qpair(rqpair); 3658 } 3659 3660 static struct spdk_nvmf_rdma_qpair * 3661 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3662 { 3663 struct spdk_nvmf_rdma_qpair *rqpair; 3664 /* @todo: improve QP search */ 3665 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3666 if (wc->qp_num == rqpair->rdma_qp->qp->qp_num) { 3667 return rqpair; 3668 } 3669 } 3670 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3671 return NULL; 3672 } 3673 3674 #ifdef DEBUG 3675 static int 3676 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3677 { 3678 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3679 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3680 } 3681 #endif 3682 3683 static void 3684 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3685 int rc) 3686 { 3687 struct spdk_nvmf_rdma_recv *rdma_recv; 3688 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3689 3690 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3691 while (bad_recv_wr != NULL) { 3692 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3693 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3694 3695 rdma_recv->qpair->current_recv_depth++; 3696 bad_recv_wr = bad_recv_wr->next; 3697 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3698 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3699 } 3700 } 3701 3702 static void 3703 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3704 { 3705 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3706 while (bad_recv_wr != NULL) { 3707 bad_recv_wr = bad_recv_wr->next; 3708 rqpair->current_recv_depth++; 3709 } 3710 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3711 } 3712 3713 static void 3714 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3715 struct spdk_nvmf_rdma_poller *rpoller) 3716 { 3717 struct spdk_nvmf_rdma_qpair *rqpair; 3718 struct ibv_recv_wr *bad_recv_wr; 3719 int rc; 3720 3721 if (rpoller->srq) { 3722 if (rpoller->resources->recvs_to_post.first != NULL) { 3723 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3724 if (rc) { 3725 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3726 } 3727 rpoller->resources->recvs_to_post.first = NULL; 3728 rpoller->resources->recvs_to_post.last = NULL; 3729 } 3730 } else { 3731 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3732 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3733 assert(rqpair->resources->recvs_to_post.first != NULL); 3734 rc = ibv_post_recv(rqpair->rdma_qp->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3735 if (rc) { 3736 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3737 } 3738 rqpair->resources->recvs_to_post.first = NULL; 3739 rqpair->resources->recvs_to_post.last = NULL; 3740 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3741 } 3742 } 3743 } 3744 3745 static void 3746 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3747 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3748 { 3749 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3750 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3751 3752 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3753 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3754 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3755 assert(rqpair->current_send_depth > 0); 3756 rqpair->current_send_depth--; 3757 switch (bad_rdma_wr->type) { 3758 case RDMA_WR_TYPE_DATA: 3759 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3760 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3761 assert(rqpair->current_read_depth > 0); 3762 rqpair->current_read_depth--; 3763 } 3764 break; 3765 case RDMA_WR_TYPE_SEND: 3766 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3767 break; 3768 default: 3769 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3770 prev_rdma_req = cur_rdma_req; 3771 continue; 3772 } 3773 3774 if (prev_rdma_req == cur_rdma_req) { 3775 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3776 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3777 continue; 3778 } 3779 3780 switch (cur_rdma_req->state) { 3781 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3782 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3783 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3784 break; 3785 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3786 case RDMA_REQUEST_STATE_COMPLETING: 3787 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3788 break; 3789 default: 3790 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3791 cur_rdma_req->state, rqpair); 3792 continue; 3793 } 3794 3795 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3796 prev_rdma_req = cur_rdma_req; 3797 } 3798 3799 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3800 /* Disconnect the connection. */ 3801 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3802 } 3803 3804 } 3805 3806 static void 3807 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3808 struct spdk_nvmf_rdma_poller *rpoller) 3809 { 3810 struct spdk_nvmf_rdma_qpair *rqpair; 3811 struct ibv_send_wr *bad_wr = NULL; 3812 int rc; 3813 3814 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3815 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3816 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3817 3818 /* bad wr always points to the first wr that failed. */ 3819 if (rc) { 3820 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3821 } 3822 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3823 } 3824 } 3825 3826 static const char * 3827 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 3828 { 3829 switch (wr_type) { 3830 case RDMA_WR_TYPE_RECV: 3831 return "RECV"; 3832 case RDMA_WR_TYPE_SEND: 3833 return "SEND"; 3834 case RDMA_WR_TYPE_DATA: 3835 return "DATA"; 3836 default: 3837 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 3838 SPDK_UNREACHABLE(); 3839 } 3840 } 3841 3842 static inline void 3843 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 3844 { 3845 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 3846 3847 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 3848 /* If qpair is in ERR state, we will receive completions for all posted and not completed 3849 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 3850 SPDK_DEBUGLOG(rdma, 3851 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3852 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3853 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3854 } else { 3855 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3856 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3857 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3858 } 3859 } 3860 3861 static int 3862 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3863 struct spdk_nvmf_rdma_poller *rpoller) 3864 { 3865 struct ibv_wc wc[32]; 3866 struct spdk_nvmf_rdma_wr *rdma_wr; 3867 struct spdk_nvmf_rdma_request *rdma_req; 3868 struct spdk_nvmf_rdma_recv *rdma_recv; 3869 struct spdk_nvmf_rdma_qpair *rqpair; 3870 int reaped, i; 3871 int count = 0; 3872 bool error = false; 3873 uint64_t poll_tsc = spdk_get_ticks(); 3874 3875 /* Poll for completing operations. */ 3876 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3877 if (reaped < 0) { 3878 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3879 errno, spdk_strerror(errno)); 3880 return -1; 3881 } 3882 3883 rpoller->stat.polls++; 3884 rpoller->stat.completions += reaped; 3885 3886 for (i = 0; i < reaped; i++) { 3887 3888 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3889 3890 switch (rdma_wr->type) { 3891 case RDMA_WR_TYPE_SEND: 3892 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3893 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3894 3895 if (!wc[i].status) { 3896 count++; 3897 assert(wc[i].opcode == IBV_WC_SEND); 3898 assert(nvmf_rdma_req_is_completing(rdma_req)); 3899 } 3900 3901 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3902 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3903 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3904 rdma_req->num_outstanding_data_wr = 0; 3905 3906 nvmf_rdma_request_process(rtransport, rdma_req); 3907 break; 3908 case RDMA_WR_TYPE_RECV: 3909 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3910 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3911 if (rpoller->srq != NULL) { 3912 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3913 /* It is possible that there are still some completions for destroyed QP 3914 * associated with SRQ. We just ignore these late completions and re-post 3915 * receive WRs back to SRQ. 3916 */ 3917 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3918 struct ibv_recv_wr *bad_wr; 3919 int rc; 3920 3921 rdma_recv->wr.next = NULL; 3922 rc = ibv_post_srq_recv(rpoller->srq, 3923 &rdma_recv->wr, 3924 &bad_wr); 3925 if (rc) { 3926 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3927 } 3928 continue; 3929 } 3930 } 3931 rqpair = rdma_recv->qpair; 3932 3933 assert(rqpair != NULL); 3934 if (!wc[i].status) { 3935 assert(wc[i].opcode == IBV_WC_RECV); 3936 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3937 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3938 break; 3939 } 3940 } 3941 3942 rdma_recv->wr.next = NULL; 3943 rqpair->current_recv_depth++; 3944 rdma_recv->receive_tsc = poll_tsc; 3945 rpoller->stat.requests++; 3946 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3947 break; 3948 case RDMA_WR_TYPE_DATA: 3949 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3950 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3951 3952 assert(rdma_req->num_outstanding_data_wr > 0); 3953 3954 rqpair->current_send_depth--; 3955 rdma_req->num_outstanding_data_wr--; 3956 if (!wc[i].status) { 3957 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3958 rqpair->current_read_depth--; 3959 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3960 if (rdma_req->num_outstanding_data_wr == 0) { 3961 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3962 nvmf_rdma_request_process(rtransport, rdma_req); 3963 } 3964 } else { 3965 /* If the data transfer fails still force the queue into the error state, 3966 * if we were performing an RDMA_READ, we need to force the request into a 3967 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3968 * case, we should wait for the SEND to complete. */ 3969 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3970 rqpair->current_read_depth--; 3971 if (rdma_req->num_outstanding_data_wr == 0) { 3972 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3973 } 3974 } 3975 } 3976 break; 3977 default: 3978 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3979 continue; 3980 } 3981 3982 /* Handle error conditions */ 3983 if (wc[i].status) { 3984 nvmf_rdma_update_ibv_state(rqpair); 3985 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 3986 3987 error = true; 3988 3989 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3990 /* Disconnect the connection. */ 3991 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3992 } else { 3993 nvmf_rdma_destroy_drained_qpair(rqpair); 3994 } 3995 continue; 3996 } 3997 3998 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3999 4000 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 4001 nvmf_rdma_destroy_drained_qpair(rqpair); 4002 } 4003 } 4004 4005 if (error == true) { 4006 return -1; 4007 } 4008 4009 /* submit outstanding work requests. */ 4010 _poller_submit_recvs(rtransport, rpoller); 4011 _poller_submit_sends(rtransport, rpoller); 4012 4013 return count; 4014 } 4015 4016 static int 4017 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 4018 { 4019 struct spdk_nvmf_rdma_transport *rtransport; 4020 struct spdk_nvmf_rdma_poll_group *rgroup; 4021 struct spdk_nvmf_rdma_poller *rpoller; 4022 int count, rc; 4023 4024 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 4025 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4026 4027 count = 0; 4028 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4029 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4030 if (rc < 0) { 4031 return rc; 4032 } 4033 count += rc; 4034 } 4035 4036 return count; 4037 } 4038 4039 static int 4040 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4041 struct spdk_nvme_transport_id *trid, 4042 bool peer) 4043 { 4044 struct sockaddr *saddr; 4045 uint16_t port; 4046 4047 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4048 4049 if (peer) { 4050 saddr = rdma_get_peer_addr(id); 4051 } else { 4052 saddr = rdma_get_local_addr(id); 4053 } 4054 switch (saddr->sa_family) { 4055 case AF_INET: { 4056 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4057 4058 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4059 inet_ntop(AF_INET, &saddr_in->sin_addr, 4060 trid->traddr, sizeof(trid->traddr)); 4061 if (peer) { 4062 port = ntohs(rdma_get_dst_port(id)); 4063 } else { 4064 port = ntohs(rdma_get_src_port(id)); 4065 } 4066 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4067 break; 4068 } 4069 case AF_INET6: { 4070 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4071 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4072 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4073 trid->traddr, sizeof(trid->traddr)); 4074 if (peer) { 4075 port = ntohs(rdma_get_dst_port(id)); 4076 } else { 4077 port = ntohs(rdma_get_src_port(id)); 4078 } 4079 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4080 break; 4081 } 4082 default: 4083 return -1; 4084 4085 } 4086 4087 return 0; 4088 } 4089 4090 static int 4091 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4092 struct spdk_nvme_transport_id *trid) 4093 { 4094 struct spdk_nvmf_rdma_qpair *rqpair; 4095 4096 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4097 4098 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4099 } 4100 4101 static int 4102 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4103 struct spdk_nvme_transport_id *trid) 4104 { 4105 struct spdk_nvmf_rdma_qpair *rqpair; 4106 4107 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4108 4109 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4110 } 4111 4112 static int 4113 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4114 struct spdk_nvme_transport_id *trid) 4115 { 4116 struct spdk_nvmf_rdma_qpair *rqpair; 4117 4118 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4119 4120 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4121 } 4122 4123 void 4124 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4125 { 4126 g_nvmf_hooks = *hooks; 4127 } 4128 4129 static void 4130 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4131 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4132 { 4133 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4134 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4135 4136 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4137 4138 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4139 } 4140 4141 static int 4142 _nvmf_rdma_qpair_abort_request(void *ctx) 4143 { 4144 struct spdk_nvmf_request *req = ctx; 4145 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4146 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4147 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4148 struct spdk_nvmf_rdma_qpair, qpair); 4149 int rc; 4150 4151 spdk_poller_unregister(&req->poller); 4152 4153 switch (rdma_req_to_abort->state) { 4154 case RDMA_REQUEST_STATE_EXECUTING: 4155 rc = nvmf_ctrlr_abort_request(req); 4156 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4157 return SPDK_POLLER_BUSY; 4158 } 4159 break; 4160 4161 case RDMA_REQUEST_STATE_NEED_BUFFER: 4162 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4163 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4164 4165 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4166 break; 4167 4168 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4169 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4170 spdk_nvmf_rdma_request, state_link); 4171 4172 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4173 break; 4174 4175 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4176 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4177 spdk_nvmf_rdma_request, state_link); 4178 4179 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4180 break; 4181 4182 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4183 if (spdk_get_ticks() < req->timeout_tsc) { 4184 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4185 return SPDK_POLLER_BUSY; 4186 } 4187 break; 4188 4189 default: 4190 break; 4191 } 4192 4193 spdk_nvmf_request_complete(req); 4194 return SPDK_POLLER_BUSY; 4195 } 4196 4197 static void 4198 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4199 struct spdk_nvmf_request *req) 4200 { 4201 struct spdk_nvmf_rdma_qpair *rqpair; 4202 struct spdk_nvmf_rdma_transport *rtransport; 4203 struct spdk_nvmf_transport *transport; 4204 uint16_t cid; 4205 uint32_t i; 4206 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL; 4207 4208 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4209 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4210 transport = &rtransport->transport; 4211 4212 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4213 4214 for (i = 0; i < rqpair->max_queue_depth; i++) { 4215 rdma_req_to_abort = &rqpair->resources->reqs[i]; 4216 4217 if (rdma_req_to_abort->state != RDMA_REQUEST_STATE_FREE && 4218 rdma_req_to_abort->req.cmd->nvme_cmd.cid == cid) { 4219 break; 4220 } 4221 } 4222 4223 if (rdma_req_to_abort == NULL) { 4224 spdk_nvmf_request_complete(req); 4225 return; 4226 } 4227 4228 req->req_to_abort = &rdma_req_to_abort->req; 4229 req->timeout_tsc = spdk_get_ticks() + 4230 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4231 req->poller = NULL; 4232 4233 _nvmf_rdma_qpair_abort_request(req); 4234 } 4235 4236 static int 4237 nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4238 struct spdk_nvmf_transport_poll_group_stat **stat) 4239 { 4240 struct spdk_io_channel *ch; 4241 struct spdk_nvmf_poll_group *group; 4242 struct spdk_nvmf_transport_poll_group *tgroup; 4243 struct spdk_nvmf_rdma_poll_group *rgroup; 4244 struct spdk_nvmf_rdma_poller *rpoller; 4245 struct spdk_nvmf_rdma_device_stat *device_stat; 4246 uint64_t num_devices = 0; 4247 4248 if (tgt == NULL || stat == NULL) { 4249 return -EINVAL; 4250 } 4251 4252 ch = spdk_get_io_channel(tgt); 4253 group = spdk_io_channel_get_ctx(ch);; 4254 spdk_put_io_channel(ch); 4255 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4256 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4257 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4258 if (!*stat) { 4259 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4260 return -ENOMEM; 4261 } 4262 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4263 4264 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4265 /* Count devices to allocate enough memory */ 4266 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4267 ++num_devices; 4268 } 4269 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4270 if (!(*stat)->rdma.devices) { 4271 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4272 free(*stat); 4273 return -ENOMEM; 4274 } 4275 4276 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4277 (*stat)->rdma.num_devices = num_devices; 4278 num_devices = 0; 4279 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4280 device_stat = &(*stat)->rdma.devices[num_devices++]; 4281 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4282 device_stat->polls = rpoller->stat.polls; 4283 device_stat->completions = rpoller->stat.completions; 4284 device_stat->requests = rpoller->stat.requests; 4285 device_stat->request_latency = rpoller->stat.request_latency; 4286 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4287 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4288 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4289 } 4290 return 0; 4291 } 4292 } 4293 return -ENOENT; 4294 } 4295 4296 static void 4297 nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4298 { 4299 if (stat) { 4300 free(stat->rdma.devices); 4301 } 4302 free(stat); 4303 } 4304 4305 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4306 .name = "RDMA", 4307 .type = SPDK_NVME_TRANSPORT_RDMA, 4308 .opts_init = nvmf_rdma_opts_init, 4309 .create = nvmf_rdma_create, 4310 .dump_opts = nvmf_rdma_dump_opts, 4311 .destroy = nvmf_rdma_destroy, 4312 4313 .listen = nvmf_rdma_listen, 4314 .stop_listen = nvmf_rdma_stop_listen, 4315 .accept = nvmf_rdma_accept, 4316 .cdata_init = nvmf_rdma_cdata_init, 4317 4318 .listener_discover = nvmf_rdma_discover, 4319 4320 .poll_group_create = nvmf_rdma_poll_group_create, 4321 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4322 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4323 .poll_group_add = nvmf_rdma_poll_group_add, 4324 .poll_group_remove = nvmf_rdma_poll_group_remove, 4325 .poll_group_poll = nvmf_rdma_poll_group_poll, 4326 4327 .req_free = nvmf_rdma_request_free, 4328 .req_complete = nvmf_rdma_request_complete, 4329 4330 .qpair_fini = nvmf_rdma_close_qpair, 4331 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4332 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4333 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4334 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4335 4336 .poll_group_get_stat = nvmf_rdma_poll_group_get_stat, 4337 .poll_group_free_stat = nvmf_rdma_poll_group_free_stat, 4338 }; 4339 4340 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4341 SPDK_LOG_REGISTER_COMPONENT(rdma) 4342