1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. 5 * 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 <infiniband/verbs.h> 37 #include <rdma/rdma_cma.h> 38 #include <rdma/rdma_verbs.h> 39 40 #include "nvmf_internal.h" 41 #include "transport.h" 42 43 #include "spdk/assert.h" 44 #include "spdk/thread.h" 45 #include "spdk/nvmf.h" 46 #include "spdk/nvmf_spec.h" 47 #include "spdk/string.h" 48 #include "spdk/trace.h" 49 #include "spdk/util.h" 50 51 #include "spdk_internal/log.h" 52 53 /* 54 RDMA Connection Resource Defaults 55 */ 56 #define NVMF_DEFAULT_TX_SGE 1 57 #define NVMF_DEFAULT_RX_SGE 2 58 #define NVMF_DEFAULT_DATA_SGE 16 59 60 /* The RDMA completion queue size */ 61 #define NVMF_RDMA_CQ_SIZE 4096 62 63 /* AIO backend requires block size aligned data buffers, 64 * extra 4KiB aligned data buffer should work for most devices. 65 */ 66 #define SHIFT_4KB 12 67 #define NVMF_DATA_BUFFER_ALIGNMENT (1 << SHIFT_4KB) 68 #define NVMF_DATA_BUFFER_MASK (NVMF_DATA_BUFFER_ALIGNMENT - 1) 69 70 enum spdk_nvmf_rdma_request_state { 71 /* The request is not currently in use */ 72 RDMA_REQUEST_STATE_FREE = 0, 73 74 /* Initial state when request first received */ 75 RDMA_REQUEST_STATE_NEW, 76 77 /* The request is queued until a data buffer is available. */ 78 RDMA_REQUEST_STATE_NEED_BUFFER, 79 80 /* The request is waiting on RDMA queue depth availability 81 * to transfer data between the host and the controller. 82 */ 83 RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 84 85 /* The request is currently transferring data from the host to the controller. */ 86 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 87 88 /* The request is ready to execute at the block device */ 89 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 90 91 /* The request is currently executing at the block device */ 92 RDMA_REQUEST_STATE_EXECUTING, 93 94 /* The request finished executing at the block device */ 95 RDMA_REQUEST_STATE_EXECUTED, 96 97 /* The request is ready to send a completion */ 98 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 99 100 /* The request is currently transferring data from the controller to the host. */ 101 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 102 103 /* The request currently has an outstanding completion without an 104 * associated data transfer. 105 */ 106 RDMA_REQUEST_STATE_COMPLETING, 107 108 /* The request completed and can be marked free. */ 109 RDMA_REQUEST_STATE_COMPLETED, 110 111 /* Terminator */ 112 RDMA_REQUEST_NUM_STATES, 113 }; 114 115 #define OBJECT_NVMF_RDMA_IO 0x40 116 117 #define TRACE_GROUP_NVMF_RDMA 0x4 118 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 119 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 120 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 121 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 122 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 123 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 124 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 125 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 126 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 127 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 128 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 129 130 SPDK_TRACE_REGISTER_FN(nvmf_trace) 131 { 132 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 133 spdk_trace_register_description("RDMA_REQ_NEW", "", 134 TRACE_RDMA_REQUEST_STATE_NEW, 135 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 0, 0, ""); 136 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", "", 137 TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 138 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 0, 0, ""); 139 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H_TO_C", "", 140 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 141 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 0, 0, ""); 142 spdk_trace_register_description("RDMA_REQ_TX_H_TO_C", "", 143 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 144 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 0, 0, ""); 145 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", "", 146 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 147 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 0, 0, ""); 148 spdk_trace_register_description("RDMA_REQ_EXECUTING", "", 149 TRACE_RDMA_REQUEST_STATE_EXECUTING, 150 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 0, 0, ""); 151 spdk_trace_register_description("RDMA_REQ_EXECUTED", "", 152 TRACE_RDMA_REQUEST_STATE_EXECUTED, 153 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 0, 0, ""); 154 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPLETE", "", 155 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 156 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 0, 0, ""); 157 spdk_trace_register_description("RDMA_REQ_COMPLETING_CONTROLLER_TO_HOST", "", 158 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 159 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 0, 0, ""); 160 spdk_trace_register_description("RDMA_REQ_COMPLETING_INCAPSULE", "", 161 TRACE_RDMA_REQUEST_STATE_COMPLETING, 162 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 0, 0, ""); 163 spdk_trace_register_description("RDMA_REQ_COMPLETED", "", 164 TRACE_RDMA_REQUEST_STATE_COMPLETED, 165 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 0, 0, ""); 166 } 167 168 /* This structure holds commands as they are received off the wire. 169 * It must be dynamically paired with a full request object 170 * (spdk_nvmf_rdma_request) to service a request. It is separate 171 * from the request because RDMA does not appear to order 172 * completions, so occasionally we'll get a new incoming 173 * command when there aren't any free request objects. 174 */ 175 struct spdk_nvmf_rdma_recv { 176 struct ibv_recv_wr wr; 177 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 178 179 struct spdk_nvmf_rdma_qpair *qpair; 180 181 /* In-capsule data buffer */ 182 uint8_t *buf; 183 184 TAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 185 }; 186 187 struct spdk_nvmf_rdma_request { 188 struct spdk_nvmf_request req; 189 bool data_from_pool; 190 191 enum spdk_nvmf_rdma_request_state state; 192 193 struct spdk_nvmf_rdma_recv *recv; 194 195 struct { 196 struct ibv_send_wr wr; 197 struct ibv_sge sgl[NVMF_DEFAULT_TX_SGE]; 198 } rsp; 199 200 struct { 201 struct ibv_send_wr wr; 202 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 203 void *buffers[SPDK_NVMF_MAX_SGL_ENTRIES]; 204 } data; 205 206 TAILQ_ENTRY(spdk_nvmf_rdma_request) link; 207 TAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 208 }; 209 210 struct spdk_nvmf_rdma_qpair { 211 struct spdk_nvmf_qpair qpair; 212 213 struct spdk_nvmf_rdma_port *port; 214 struct spdk_nvmf_rdma_poller *poller; 215 216 struct rdma_cm_id *cm_id; 217 218 /* The maximum number of I/O outstanding on this connection at one time */ 219 uint16_t max_queue_depth; 220 221 /* The maximum number of active RDMA READ and WRITE operations at one time */ 222 uint16_t max_rw_depth; 223 224 /* Receives that are waiting for a request object */ 225 TAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 226 227 /* Queues to track the requests in all states */ 228 TAILQ_HEAD(, spdk_nvmf_rdma_request) state_queue[RDMA_REQUEST_NUM_STATES]; 229 230 /* Number of requests in each state */ 231 uint32_t state_cntr[RDMA_REQUEST_NUM_STATES]; 232 233 int max_sge; 234 235 /* Array of size "max_queue_depth" containing RDMA requests. */ 236 struct spdk_nvmf_rdma_request *reqs; 237 238 /* Array of size "max_queue_depth" containing RDMA recvs. */ 239 struct spdk_nvmf_rdma_recv *recvs; 240 241 /* Array of size "max_queue_depth" containing 64 byte capsules 242 * used for receive. 243 */ 244 union nvmf_h2c_msg *cmds; 245 struct ibv_mr *cmds_mr; 246 247 /* Array of size "max_queue_depth" containing 16 byte completions 248 * to be sent back to the user. 249 */ 250 union nvmf_c2h_msg *cpls; 251 struct ibv_mr *cpls_mr; 252 253 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 254 * buffers to be used for in capsule data. 255 */ 256 void *bufs; 257 struct ibv_mr *bufs_mr; 258 259 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 260 261 /* Mgmt channel */ 262 struct spdk_io_channel *mgmt_channel; 263 struct spdk_nvmf_rdma_mgmt_channel *ch; 264 265 /* IBV queue pair attributes: they are used to manage 266 * qp state and recover from errors. 267 */ 268 struct ibv_qp_init_attr ibv_init_attr; 269 struct ibv_qp_attr ibv_attr; 270 271 bool qpair_disconnected; 272 }; 273 274 struct spdk_nvmf_rdma_poller { 275 struct spdk_nvmf_rdma_device *device; 276 struct spdk_nvmf_rdma_poll_group *group; 277 278 struct ibv_cq *cq; 279 280 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 281 282 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 283 }; 284 285 struct spdk_nvmf_rdma_poll_group { 286 struct spdk_nvmf_transport_poll_group group; 287 288 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 289 }; 290 291 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 292 struct spdk_nvmf_rdma_device { 293 struct ibv_device_attr attr; 294 struct ibv_context *context; 295 296 struct spdk_mem_map *map; 297 struct ibv_pd *pd; 298 299 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 300 }; 301 302 struct spdk_nvmf_rdma_port { 303 struct spdk_nvme_transport_id trid; 304 struct rdma_cm_id *id; 305 struct spdk_nvmf_rdma_device *device; 306 uint32_t ref; 307 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 308 }; 309 310 struct spdk_nvmf_rdma_transport { 311 struct spdk_nvmf_transport transport; 312 313 struct rdma_event_channel *event_channel; 314 315 struct spdk_mempool *data_buf_pool; 316 317 pthread_mutex_t lock; 318 319 /* fields used to poll RDMA/IB events */ 320 nfds_t npoll_fds; 321 struct pollfd *poll_fds; 322 323 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 324 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 325 }; 326 327 struct spdk_nvmf_rdma_mgmt_channel { 328 /* Requests that are waiting to obtain a data buffer */ 329 TAILQ_HEAD(, spdk_nvmf_rdma_request) pending_data_buf_queue; 330 }; 331 332 /* API to IBV QueuePair */ 333 static const char *str_ibv_qp_state[] = { 334 "IBV_QPS_RESET", 335 "IBV_QPS_INIT", 336 "IBV_QPS_RTR", 337 "IBV_QPS_RTS", 338 "IBV_QPS_SQD", 339 "IBV_QPS_SQE", 340 "IBV_QPS_ERR" 341 }; 342 343 static enum ibv_qp_state 344 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 345 int rc; 346 347 /* All the attributes needed for recovery */ 348 static int spdk_nvmf_ibv_attr_mask = 349 IBV_QP_STATE | 350 IBV_QP_PKEY_INDEX | 351 IBV_QP_PORT | 352 IBV_QP_ACCESS_FLAGS | 353 IBV_QP_AV | 354 IBV_QP_PATH_MTU | 355 IBV_QP_DEST_QPN | 356 IBV_QP_RQ_PSN | 357 IBV_QP_MAX_DEST_RD_ATOMIC | 358 IBV_QP_MIN_RNR_TIMER | 359 IBV_QP_SQ_PSN | 360 IBV_QP_TIMEOUT | 361 IBV_QP_RETRY_CNT | 362 IBV_QP_RNR_RETRY | 363 IBV_QP_MAX_QP_RD_ATOMIC; 364 365 rc = ibv_query_qp(rqpair->cm_id->qp, &rqpair->ibv_attr, 366 spdk_nvmf_ibv_attr_mask, &rqpair->ibv_init_attr); 367 368 if (rc) 369 { 370 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 371 assert(false); 372 } 373 374 return rqpair->ibv_attr.qp_state; 375 } 376 377 static int 378 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 379 enum ibv_qp_state new_state) 380 { 381 int rc; 382 enum ibv_qp_state state; 383 static int attr_mask_rc[] = { 384 [IBV_QPS_RESET] = IBV_QP_STATE, 385 [IBV_QPS_INIT] = (IBV_QP_STATE | 386 IBV_QP_PKEY_INDEX | 387 IBV_QP_PORT | 388 IBV_QP_ACCESS_FLAGS), 389 [IBV_QPS_RTR] = (IBV_QP_STATE | 390 IBV_QP_AV | 391 IBV_QP_PATH_MTU | 392 IBV_QP_DEST_QPN | 393 IBV_QP_RQ_PSN | 394 IBV_QP_MAX_DEST_RD_ATOMIC | 395 IBV_QP_MIN_RNR_TIMER), 396 [IBV_QPS_RTS] = (IBV_QP_STATE | 397 IBV_QP_SQ_PSN | 398 IBV_QP_TIMEOUT | 399 IBV_QP_RETRY_CNT | 400 IBV_QP_RNR_RETRY | 401 IBV_QP_MAX_QP_RD_ATOMIC), 402 [IBV_QPS_SQD] = IBV_QP_STATE, 403 [IBV_QPS_SQE] = IBV_QP_STATE, 404 [IBV_QPS_ERR] = IBV_QP_STATE, 405 }; 406 407 switch (new_state) { 408 case IBV_QPS_RESET: 409 case IBV_QPS_INIT: 410 case IBV_QPS_RTR: 411 case IBV_QPS_RTS: 412 case IBV_QPS_SQD: 413 case IBV_QPS_SQE: 414 case IBV_QPS_ERR: 415 break; 416 default: 417 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 418 rqpair->qpair.qid, new_state); 419 return -1; 420 } 421 rqpair->ibv_attr.cur_qp_state = rqpair->ibv_attr.qp_state; 422 rqpair->ibv_attr.qp_state = new_state; 423 rqpair->ibv_attr.ah_attr.port_num = rqpair->ibv_attr.port_num; 424 425 rc = ibv_modify_qp(rqpair->cm_id->qp, &rqpair->ibv_attr, 426 attr_mask_rc[new_state]); 427 428 if (rc) { 429 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 430 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 431 return rc; 432 } 433 434 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 435 436 if (state != new_state) { 437 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 438 rqpair->qpair.qid, str_ibv_qp_state[new_state], 439 str_ibv_qp_state[state]); 440 return -1; 441 } 442 SPDK_NOTICELOG("IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 443 str_ibv_qp_state[state]); 444 return 0; 445 } 446 447 static void 448 spdk_nvmf_rdma_request_set_state(struct spdk_nvmf_rdma_request *rdma_req, 449 enum spdk_nvmf_rdma_request_state state) 450 { 451 struct spdk_nvmf_qpair *qpair; 452 struct spdk_nvmf_rdma_qpair *rqpair; 453 454 qpair = rdma_req->req.qpair; 455 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 456 457 TAILQ_REMOVE(&rqpair->state_queue[rdma_req->state], rdma_req, state_link); 458 rqpair->state_cntr[rdma_req->state]--; 459 460 rdma_req->state = state; 461 462 TAILQ_INSERT_TAIL(&rqpair->state_queue[rdma_req->state], rdma_req, state_link); 463 rqpair->state_cntr[rdma_req->state]++; 464 } 465 466 static int 467 spdk_nvmf_rdma_mgmt_channel_create(void *io_device, void *ctx_buf) 468 { 469 struct spdk_nvmf_rdma_mgmt_channel *ch = ctx_buf; 470 471 TAILQ_INIT(&ch->pending_data_buf_queue); 472 return 0; 473 } 474 475 static void 476 spdk_nvmf_rdma_mgmt_channel_destroy(void *io_device, void *ctx_buf) 477 { 478 struct spdk_nvmf_rdma_mgmt_channel *ch = ctx_buf; 479 480 if (!TAILQ_EMPTY(&ch->pending_data_buf_queue)) { 481 SPDK_ERRLOG("Pending I/O list wasn't empty on channel destruction\n"); 482 } 483 } 484 485 static int 486 spdk_nvmf_rdma_cur_rw_depth(struct spdk_nvmf_rdma_qpair *rqpair) 487 { 488 return rqpair->state_cntr[RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER] + 489 rqpair->state_cntr[RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST]; 490 } 491 492 static int 493 spdk_nvmf_rdma_cur_queue_depth(struct spdk_nvmf_rdma_qpair *rqpair) 494 { 495 return rqpair->max_queue_depth - 496 rqpair->state_cntr[RDMA_REQUEST_STATE_FREE]; 497 } 498 499 static void 500 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 501 { 502 if (spdk_nvmf_rdma_cur_queue_depth(rqpair)) { 503 rqpair->qpair_disconnected = true; 504 return; 505 } 506 507 if (rqpair->poller) { 508 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 509 } 510 511 if (rqpair->cmds_mr) { 512 ibv_dereg_mr(rqpair->cmds_mr); 513 } 514 515 if (rqpair->cpls_mr) { 516 ibv_dereg_mr(rqpair->cpls_mr); 517 } 518 519 if (rqpair->bufs_mr) { 520 ibv_dereg_mr(rqpair->bufs_mr); 521 } 522 523 if (rqpair->cm_id) { 524 rdma_destroy_qp(rqpair->cm_id); 525 rdma_destroy_id(rqpair->cm_id); 526 } 527 528 if (rqpair->mgmt_channel) { 529 spdk_put_io_channel(rqpair->mgmt_channel); 530 } 531 532 /* Free all memory */ 533 spdk_dma_free(rqpair->cmds); 534 spdk_dma_free(rqpair->cpls); 535 spdk_dma_free(rqpair->bufs); 536 free(rqpair->reqs); 537 free(rqpair->recvs); 538 free(rqpair); 539 } 540 541 static int 542 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 543 { 544 struct spdk_nvmf_rdma_transport *rtransport; 545 struct spdk_nvmf_rdma_qpair *rqpair; 546 int rc, i; 547 struct spdk_nvmf_rdma_recv *rdma_recv; 548 struct spdk_nvmf_rdma_request *rdma_req; 549 struct spdk_nvmf_transport *transport; 550 551 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 552 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 553 transport = &rtransport->transport; 554 555 memset(&rqpair->ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 556 rqpair->ibv_init_attr.qp_context = rqpair; 557 rqpair->ibv_init_attr.qp_type = IBV_QPT_RC; 558 rqpair->ibv_init_attr.send_cq = rqpair->poller->cq; 559 rqpair->ibv_init_attr.recv_cq = rqpair->poller->cq; 560 rqpair->ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 561 2; /* SEND, READ, and WRITE operations */ 562 rqpair->ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; /* RECV operations */ 563 rqpair->ibv_init_attr.cap.max_send_sge = rqpair->max_sge; 564 rqpair->ibv_init_attr.cap.max_recv_sge = NVMF_DEFAULT_RX_SGE; 565 566 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &rqpair->ibv_init_attr); 567 if (rc) { 568 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 569 rdma_destroy_id(rqpair->cm_id); 570 rqpair->cm_id = NULL; 571 spdk_nvmf_rdma_qpair_destroy(rqpair); 572 return -1; 573 } 574 575 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 576 577 rqpair->reqs = calloc(rqpair->max_queue_depth, sizeof(*rqpair->reqs)); 578 rqpair->recvs = calloc(rqpair->max_queue_depth, sizeof(*rqpair->recvs)); 579 rqpair->cmds = spdk_dma_zmalloc(rqpair->max_queue_depth * sizeof(*rqpair->cmds), 580 0x1000, NULL); 581 rqpair->cpls = spdk_dma_zmalloc(rqpair->max_queue_depth * sizeof(*rqpair->cpls), 582 0x1000, NULL); 583 584 585 if (transport->opts.in_capsule_data_size > 0) { 586 rqpair->bufs = spdk_dma_zmalloc(rqpair->max_queue_depth * 587 transport->opts.in_capsule_data_size, 588 0x1000, NULL); 589 } 590 591 if (!rqpair->reqs || !rqpair->recvs || !rqpair->cmds || 592 !rqpair->cpls || (transport->opts.in_capsule_data_size && !rqpair->bufs)) { 593 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 594 spdk_nvmf_rdma_qpair_destroy(rqpair); 595 return -1; 596 } 597 598 rqpair->cmds_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->cmds, 599 rqpair->max_queue_depth * sizeof(*rqpair->cmds), 600 IBV_ACCESS_LOCAL_WRITE); 601 rqpair->cpls_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->cpls, 602 rqpair->max_queue_depth * sizeof(*rqpair->cpls), 603 0); 604 605 if (transport->opts.in_capsule_data_size) { 606 rqpair->bufs_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->bufs, 607 rqpair->max_queue_depth * 608 transport->opts.in_capsule_data_size, 609 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 610 } 611 612 if (!rqpair->cmds_mr || !rqpair->cpls_mr || (transport->opts.in_capsule_data_size && 613 !rqpair->bufs_mr)) { 614 SPDK_ERRLOG("Unable to register required memory for RDMA queue.\n"); 615 spdk_nvmf_rdma_qpair_destroy(rqpair); 616 return -1; 617 } 618 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 619 rqpair->cmds, rqpair->max_queue_depth * sizeof(*rqpair->cmds), rqpair->cmds_mr->lkey); 620 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 621 rqpair->cpls, rqpair->max_queue_depth * sizeof(*rqpair->cpls), rqpair->cpls_mr->lkey); 622 if (rqpair->bufs && rqpair->bufs_mr) { 623 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 624 rqpair->bufs, rqpair->max_queue_depth * 625 transport->opts.in_capsule_data_size, rqpair->bufs_mr->lkey); 626 } 627 628 /* Initialise request state queues and counters of the queue pair */ 629 for (i = RDMA_REQUEST_STATE_FREE; i < RDMA_REQUEST_NUM_STATES; i++) { 630 TAILQ_INIT(&rqpair->state_queue[i]); 631 rqpair->state_cntr[i] = 0; 632 } 633 634 for (i = 0; i < rqpair->max_queue_depth; i++) { 635 struct ibv_recv_wr *bad_wr = NULL; 636 637 rdma_recv = &rqpair->recvs[i]; 638 rdma_recv->qpair = rqpair; 639 640 /* Set up memory to receive commands */ 641 if (rqpair->bufs) { 642 rdma_recv->buf = (void *)((uintptr_t)rqpair->bufs + (i * 643 transport->opts.in_capsule_data_size)); 644 } 645 646 rdma_recv->sgl[0].addr = (uintptr_t)&rqpair->cmds[i]; 647 rdma_recv->sgl[0].length = sizeof(rqpair->cmds[i]); 648 rdma_recv->sgl[0].lkey = rqpair->cmds_mr->lkey; 649 rdma_recv->wr.num_sge = 1; 650 651 if (rdma_recv->buf && rqpair->bufs_mr) { 652 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 653 rdma_recv->sgl[1].length = transport->opts.in_capsule_data_size; 654 rdma_recv->sgl[1].lkey = rqpair->bufs_mr->lkey; 655 rdma_recv->wr.num_sge++; 656 } 657 658 rdma_recv->wr.wr_id = (uintptr_t)rdma_recv; 659 rdma_recv->wr.sg_list = rdma_recv->sgl; 660 661 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_recv->wr, &bad_wr); 662 if (rc) { 663 SPDK_ERRLOG("Unable to post capsule for RDMA RECV\n"); 664 spdk_nvmf_rdma_qpair_destroy(rqpair); 665 return -1; 666 } 667 } 668 669 for (i = 0; i < rqpair->max_queue_depth; i++) { 670 rdma_req = &rqpair->reqs[i]; 671 672 rdma_req->req.qpair = &rqpair->qpair; 673 rdma_req->req.cmd = NULL; 674 675 /* Set up memory to send responses */ 676 rdma_req->req.rsp = &rqpair->cpls[i]; 677 678 rdma_req->rsp.sgl[0].addr = (uintptr_t)&rqpair->cpls[i]; 679 rdma_req->rsp.sgl[0].length = sizeof(rqpair->cpls[i]); 680 rdma_req->rsp.sgl[0].lkey = rqpair->cpls_mr->lkey; 681 682 rdma_req->rsp.wr.wr_id = (uintptr_t)rdma_req; 683 rdma_req->rsp.wr.next = NULL; 684 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 685 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 686 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 687 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 688 689 /* Set up memory for data buffers */ 690 rdma_req->data.wr.wr_id = (uint64_t)rdma_req; 691 rdma_req->data.wr.next = NULL; 692 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 693 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 694 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 695 696 /* Initialize request state to FREE */ 697 rdma_req->state = RDMA_REQUEST_STATE_FREE; 698 TAILQ_INSERT_TAIL(&rqpair->state_queue[rdma_req->state], rdma_req, state_link); 699 rqpair->state_cntr[rdma_req->state]++; 700 } 701 702 return 0; 703 } 704 705 static int 706 request_transfer_in(struct spdk_nvmf_request *req) 707 { 708 int rc; 709 struct spdk_nvmf_rdma_request *rdma_req; 710 struct spdk_nvmf_qpair *qpair; 711 struct spdk_nvmf_rdma_qpair *rqpair; 712 struct ibv_send_wr *bad_wr = NULL; 713 714 qpair = req->qpair; 715 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 716 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 717 718 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 719 720 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA READ POSTED. Request: %p Connection: %p\n", req, qpair); 721 722 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 723 rdma_req->data.wr.next = NULL; 724 rc = ibv_post_send(rqpair->cm_id->qp, &rdma_req->data.wr, &bad_wr); 725 if (rc) { 726 SPDK_ERRLOG("Unable to transfer data from host to target\n"); 727 return -1; 728 } 729 return 0; 730 } 731 732 static int 733 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 734 { 735 int rc; 736 struct spdk_nvmf_rdma_request *rdma_req; 737 struct spdk_nvmf_qpair *qpair; 738 struct spdk_nvmf_rdma_qpair *rqpair; 739 struct spdk_nvme_cpl *rsp; 740 struct ibv_recv_wr *bad_recv_wr = NULL; 741 struct ibv_send_wr *send_wr, *bad_send_wr = NULL; 742 743 *data_posted = 0; 744 qpair = req->qpair; 745 rsp = &req->rsp->nvme_cpl; 746 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 747 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 748 749 /* Advance our sq_head pointer */ 750 if (qpair->sq_head == qpair->sq_head_max) { 751 qpair->sq_head = 0; 752 } else { 753 qpair->sq_head++; 754 } 755 rsp->sqhd = qpair->sq_head; 756 757 /* Post the capsule to the recv buffer */ 758 assert(rdma_req->recv != NULL); 759 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA RECV POSTED. Recv: %p Connection: %p\n", rdma_req->recv, 760 rqpair); 761 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_req->recv->wr, &bad_recv_wr); 762 if (rc) { 763 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 764 return rc; 765 } 766 rdma_req->recv = NULL; 767 768 /* Build the response which consists of an optional 769 * RDMA WRITE to transfer data, plus an RDMA SEND 770 * containing the response. 771 */ 772 send_wr = &rdma_req->rsp.wr; 773 774 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 775 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 776 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA WRITE POSTED. Request: %p Connection: %p\n", req, qpair); 777 778 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 779 780 rdma_req->data.wr.next = send_wr; 781 *data_posted = 1; 782 send_wr = &rdma_req->data.wr; 783 } 784 785 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA SEND POSTED. Request: %p Connection: %p\n", req, qpair); 786 787 /* Send the completion */ 788 rc = ibv_post_send(rqpair->cm_id->qp, send_wr, &bad_send_wr); 789 if (rc) { 790 SPDK_ERRLOG("Unable to send response capsule\n"); 791 } 792 793 return rc; 794 } 795 796 static int 797 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 798 { 799 struct spdk_nvmf_rdma_accept_private_data accept_data; 800 struct rdma_conn_param ctrlr_event_data = {}; 801 int rc; 802 803 accept_data.recfmt = 0; 804 accept_data.crqsize = rqpair->max_queue_depth; 805 806 ctrlr_event_data.private_data = &accept_data; 807 ctrlr_event_data.private_data_len = sizeof(accept_data); 808 if (id->ps == RDMA_PS_TCP) { 809 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 810 ctrlr_event_data.initiator_depth = rqpair->max_rw_depth; 811 } 812 813 rc = rdma_accept(id, &ctrlr_event_data); 814 if (rc) { 815 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 816 } else { 817 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 818 } 819 820 return rc; 821 } 822 823 static void 824 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 825 { 826 struct spdk_nvmf_rdma_reject_private_data rej_data; 827 828 rej_data.recfmt = 0; 829 rej_data.sts = error; 830 831 rdma_reject(id, &rej_data, sizeof(rej_data)); 832 } 833 834 static int 835 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 836 new_qpair_fn cb_fn) 837 { 838 struct spdk_nvmf_rdma_transport *rtransport; 839 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 840 struct spdk_nvmf_rdma_port *port; 841 struct rdma_conn_param *rdma_param = NULL; 842 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 843 uint16_t max_queue_depth; 844 uint16_t max_rw_depth; 845 846 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 847 848 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 849 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 850 851 rdma_param = &event->param.conn; 852 if (rdma_param->private_data == NULL || 853 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 854 SPDK_ERRLOG("connect request: no private data provided\n"); 855 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 856 return -1; 857 } 858 859 private_data = rdma_param->private_data; 860 if (private_data->recfmt != 0) { 861 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 862 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 863 return -1; 864 } 865 866 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 867 event->id->verbs->device->name, event->id->verbs->device->dev_name); 868 869 port = event->listen_id->context; 870 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 871 event->listen_id, event->listen_id->verbs, port); 872 873 /* Figure out the supported queue depth. This is a multi-step process 874 * that takes into account hardware maximums, host provided values, 875 * and our target's internal memory limits */ 876 877 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 878 879 /* Start with the maximum queue depth allowed by the target */ 880 max_queue_depth = rtransport->transport.opts.max_queue_depth; 881 max_rw_depth = rtransport->transport.opts.max_queue_depth; 882 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 883 rtransport->transport.opts.max_queue_depth); 884 885 /* Next check the local NIC's hardware limitations */ 886 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 887 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 888 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 889 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 890 max_rw_depth = spdk_min(max_rw_depth, port->device->attr.max_qp_rd_atom); 891 892 /* Next check the remote NIC's hardware limitations */ 893 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 894 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 895 rdma_param->initiator_depth, rdma_param->responder_resources); 896 if (rdma_param->initiator_depth > 0) { 897 max_rw_depth = spdk_min(max_rw_depth, rdma_param->initiator_depth); 898 } 899 900 /* Finally check for the host software requested values, which are 901 * optional. */ 902 if (rdma_param->private_data != NULL && 903 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 904 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 905 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 906 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 907 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 908 } 909 910 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 911 max_queue_depth, max_rw_depth); 912 913 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 914 if (rqpair == NULL) { 915 SPDK_ERRLOG("Could not allocate new connection.\n"); 916 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 917 return -1; 918 } 919 920 rqpair->port = port; 921 rqpair->max_queue_depth = max_queue_depth; 922 rqpair->max_rw_depth = max_rw_depth; 923 rqpair->cm_id = event->id; 924 rqpair->qpair.transport = transport; 925 rqpair->max_sge = spdk_min(port->device->attr.max_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 926 TAILQ_INIT(&rqpair->incoming_queue); 927 event->id->context = &rqpair->qpair; 928 929 cb_fn(&rqpair->qpair); 930 931 return 0; 932 } 933 934 static int 935 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 936 { 937 struct spdk_nvmf_qpair *qpair; 938 struct spdk_nvmf_rdma_qpair *rqpair; 939 940 if (evt->id == NULL) { 941 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 942 return -1; 943 } 944 945 qpair = evt->id->context; 946 if (qpair == NULL) { 947 SPDK_ERRLOG("disconnect request: no active connection\n"); 948 return -1; 949 } 950 /* ack the disconnect event before rdma_destroy_id */ 951 rdma_ack_cm_event(evt); 952 953 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 954 spdk_nvmf_rdma_update_ibv_state(rqpair); 955 956 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 957 958 return 0; 959 } 960 961 #ifdef DEBUG 962 static const char *CM_EVENT_STR[] = { 963 "RDMA_CM_EVENT_ADDR_RESOLVED", 964 "RDMA_CM_EVENT_ADDR_ERROR", 965 "RDMA_CM_EVENT_ROUTE_RESOLVED", 966 "RDMA_CM_EVENT_ROUTE_ERROR", 967 "RDMA_CM_EVENT_CONNECT_REQUEST", 968 "RDMA_CM_EVENT_CONNECT_RESPONSE", 969 "RDMA_CM_EVENT_CONNECT_ERROR", 970 "RDMA_CM_EVENT_UNREACHABLE", 971 "RDMA_CM_EVENT_REJECTED", 972 "RDMA_CM_EVENT_ESTABLISHED", 973 "RDMA_CM_EVENT_DISCONNECTED", 974 "RDMA_CM_EVENT_DEVICE_REMOVAL", 975 "RDMA_CM_EVENT_MULTICAST_JOIN", 976 "RDMA_CM_EVENT_MULTICAST_ERROR", 977 "RDMA_CM_EVENT_ADDR_CHANGE", 978 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 979 }; 980 #endif /* DEBUG */ 981 982 static int 983 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 984 enum spdk_mem_map_notify_action action, 985 void *vaddr, size_t size) 986 { 987 struct spdk_nvmf_rdma_device *device = cb_ctx; 988 struct ibv_pd *pd = device->pd; 989 struct ibv_mr *mr; 990 991 switch (action) { 992 case SPDK_MEM_MAP_NOTIFY_REGISTER: 993 mr = ibv_reg_mr(pd, vaddr, size, 994 IBV_ACCESS_LOCAL_WRITE | 995 IBV_ACCESS_REMOTE_READ | 996 IBV_ACCESS_REMOTE_WRITE); 997 if (mr == NULL) { 998 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 999 return -1; 1000 } else { 1001 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1002 } 1003 break; 1004 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1005 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, size); 1006 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1007 if (mr) { 1008 ibv_dereg_mr(mr); 1009 } 1010 break; 1011 } 1012 1013 return 0; 1014 } 1015 1016 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1017 1018 static spdk_nvme_data_transfer_t 1019 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1020 { 1021 enum spdk_nvme_data_transfer xfer; 1022 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1023 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1024 1025 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1026 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1027 rdma_req->rsp.wr.imm_data = 0; 1028 #endif 1029 1030 /* Figure out data transfer direction */ 1031 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1032 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1033 } else { 1034 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1035 1036 /* Some admin commands are special cases */ 1037 if ((rdma_req->req.qpair->qid == 0) && 1038 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1039 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1040 switch (cmd->cdw10 & 0xff) { 1041 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1042 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1043 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1044 break; 1045 default: 1046 xfer = SPDK_NVME_DATA_NONE; 1047 } 1048 } 1049 } 1050 1051 if (xfer == SPDK_NVME_DATA_NONE) { 1052 return xfer; 1053 } 1054 1055 /* Even for commands that may transfer data, they could have specified 0 length. 1056 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1057 */ 1058 switch (sgl->generic.type) { 1059 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1060 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1061 case SPDK_NVME_SGL_TYPE_SEGMENT: 1062 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1063 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1064 if (sgl->unkeyed.length == 0) { 1065 xfer = SPDK_NVME_DATA_NONE; 1066 } 1067 break; 1068 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1069 if (sgl->keyed.length == 0) { 1070 xfer = SPDK_NVME_DATA_NONE; 1071 } 1072 break; 1073 } 1074 1075 return xfer; 1076 } 1077 1078 static int 1079 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1080 struct spdk_nvmf_rdma_device *device, 1081 struct spdk_nvmf_rdma_request *rdma_req) 1082 { 1083 void *buf = NULL; 1084 uint32_t length = rdma_req->req.length; 1085 uint32_t i = 0; 1086 1087 rdma_req->req.iovcnt = 0; 1088 while (length) { 1089 buf = spdk_mempool_get(rtransport->data_buf_pool); 1090 if (!buf) { 1091 goto nomem; 1092 } 1093 1094 rdma_req->req.iov[i].iov_base = (void *)((uintptr_t)(buf + NVMF_DATA_BUFFER_MASK) & 1095 ~NVMF_DATA_BUFFER_MASK); 1096 rdma_req->req.iov[i].iov_len = spdk_min(length, rtransport->transport.opts.io_unit_size); 1097 rdma_req->req.iovcnt++; 1098 rdma_req->data.buffers[i] = buf; 1099 rdma_req->data.wr.sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[i].iov_base); 1100 rdma_req->data.wr.sg_list[i].length = rdma_req->req.iov[i].iov_len; 1101 rdma_req->data.wr.sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1102 (uint64_t)buf, rdma_req->req.iov[i].iov_len))->lkey; 1103 1104 length -= rdma_req->req.iov[i].iov_len; 1105 i++; 1106 } 1107 1108 rdma_req->data_from_pool = true; 1109 1110 return 0; 1111 1112 nomem: 1113 while (i) { 1114 i--; 1115 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->req.iov[i].iov_base); 1116 rdma_req->req.iov[i].iov_base = NULL; 1117 rdma_req->req.iov[i].iov_len = 0; 1118 1119 rdma_req->data.wr.sg_list[i].addr = 0; 1120 rdma_req->data.wr.sg_list[i].length = 0; 1121 rdma_req->data.wr.sg_list[i].lkey = 0; 1122 } 1123 rdma_req->req.iovcnt = 0; 1124 return -ENOMEM; 1125 } 1126 1127 static int 1128 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1129 struct spdk_nvmf_rdma_device *device, 1130 struct spdk_nvmf_rdma_request *rdma_req) 1131 { 1132 struct spdk_nvme_cmd *cmd; 1133 struct spdk_nvme_cpl *rsp; 1134 struct spdk_nvme_sgl_descriptor *sgl; 1135 1136 cmd = &rdma_req->req.cmd->nvme_cmd; 1137 rsp = &rdma_req->req.rsp->nvme_cpl; 1138 sgl = &cmd->dptr.sgl1; 1139 1140 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1141 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1142 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1143 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1144 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1145 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1146 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1147 return -1; 1148 } 1149 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1150 /** 1151 * These vendor IDs are assigned by the IEEE and an ID of 0 implies Soft-RoCE. 1152 * The Soft-RoCE RXE driver does not currently support send with invalidate. 1153 * There are changes making their way through the kernel now that will enable 1154 * this feature. When they are merged, we can conditionally enable this feature. 1155 * 1156 * todo: enable this for versions of the kernel rxe driver that support it. 1157 */ 1158 if (device->attr.vendor_id != 0) { 1159 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1160 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1161 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1162 } 1163 } 1164 #endif 1165 1166 /* fill request length and populate iovs */ 1167 rdma_req->req.length = sgl->keyed.length; 1168 1169 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1170 /* No available buffers. Queue this request up. */ 1171 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1172 return 0; 1173 } 1174 1175 /* backward compatible */ 1176 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1177 1178 /* rdma wr specifics */ 1179 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1180 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1181 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1182 1183 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1184 rdma_req->req.iovcnt); 1185 1186 return 0; 1187 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1188 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1189 uint64_t offset = sgl->address; 1190 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1191 1192 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1193 offset, sgl->unkeyed.length); 1194 1195 if (offset > max_len) { 1196 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1197 offset, max_len); 1198 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1199 return -1; 1200 } 1201 max_len -= (uint32_t)offset; 1202 1203 if (sgl->unkeyed.length > max_len) { 1204 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1205 sgl->unkeyed.length, max_len); 1206 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1207 return -1; 1208 } 1209 1210 rdma_req->req.data = rdma_req->recv->buf + offset; 1211 rdma_req->data_from_pool = false; 1212 rdma_req->req.length = sgl->unkeyed.length; 1213 1214 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1215 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1216 rdma_req->req.iovcnt = 1; 1217 1218 return 0; 1219 } 1220 1221 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1222 sgl->generic.type, sgl->generic.subtype); 1223 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1224 return -1; 1225 } 1226 1227 static bool 1228 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1229 struct spdk_nvmf_rdma_request *rdma_req) 1230 { 1231 struct spdk_nvmf_rdma_qpair *rqpair; 1232 struct spdk_nvmf_rdma_device *device; 1233 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1234 int rc; 1235 struct spdk_nvmf_rdma_recv *rdma_recv; 1236 enum spdk_nvmf_rdma_request_state prev_state; 1237 bool progress = false; 1238 int data_posted; 1239 int cur_rdma_rw_depth; 1240 1241 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1242 device = rqpair->port->device; 1243 1244 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1245 1246 /* If the queue pair is in an error state, force the request to the completed state 1247 * to release resources. */ 1248 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 1249 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1250 } 1251 1252 /* The loop here is to allow for several back-to-back state changes. */ 1253 do { 1254 prev_state = rdma_req->state; 1255 1256 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1257 1258 switch (rdma_req->state) { 1259 case RDMA_REQUEST_STATE_FREE: 1260 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1261 * to escape this state. */ 1262 break; 1263 case RDMA_REQUEST_STATE_NEW: 1264 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, (uintptr_t)rdma_req, 0); 1265 rdma_recv = rdma_req->recv; 1266 1267 /* The first element of the SGL is the NVMe command */ 1268 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1269 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1270 1271 TAILQ_REMOVE(&rqpair->incoming_queue, rdma_recv, link); 1272 1273 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1274 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1275 break; 1276 } 1277 1278 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 1279 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1280 break; 1281 } 1282 1283 /* The next state transition depends on the data transfer needs of this request. */ 1284 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1285 1286 /* If no data to transfer, ready to execute. */ 1287 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1288 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1289 break; 1290 } 1291 1292 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEED_BUFFER); 1293 TAILQ_INSERT_TAIL(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1294 break; 1295 case RDMA_REQUEST_STATE_NEED_BUFFER: 1296 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, (uintptr_t)rdma_req, 0); 1297 1298 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1299 1300 if (rdma_req != TAILQ_FIRST(&rqpair->ch->pending_data_buf_queue)) { 1301 /* This request needs to wait in line to obtain a buffer */ 1302 break; 1303 } 1304 1305 /* Try to get a data buffer */ 1306 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1307 if (rc < 0) { 1308 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1309 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1310 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1311 break; 1312 } 1313 1314 if (!rdma_req->req.data) { 1315 /* No buffers available. */ 1316 break; 1317 } 1318 1319 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1320 1321 /* If data is transferring from host to controller and the data didn't 1322 * arrive using in capsule data, we need to do a transfer from the host. 1323 */ 1324 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1325 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1326 break; 1327 } 1328 1329 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1330 break; 1331 case RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING: 1332 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 0, 0, 1333 (uintptr_t)rdma_req, 0); 1334 1335 if (rdma_req != TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING])) { 1336 /* This request needs to wait in line to perform RDMA */ 1337 break; 1338 } 1339 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1340 1341 if (cur_rdma_rw_depth >= rqpair->max_rw_depth) { 1342 /* R/W queue is full, need to wait */ 1343 break; 1344 } 1345 1346 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1347 rc = request_transfer_in(&rdma_req->req); 1348 if (!rc) { 1349 spdk_nvmf_rdma_request_set_state(rdma_req, 1350 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1351 } else { 1352 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1353 spdk_nvmf_rdma_request_set_state(rdma_req, 1354 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1355 } 1356 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1357 /* The data transfer will be kicked off from 1358 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1359 */ 1360 spdk_nvmf_rdma_request_set_state(rdma_req, 1361 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1362 } else { 1363 SPDK_ERRLOG("Cannot perform data transfer, unknown state: %u\n", 1364 rdma_req->req.xfer); 1365 assert(0); 1366 } 1367 break; 1368 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1369 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1370 (uintptr_t)rdma_req, 0); 1371 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1372 * to escape this state. */ 1373 break; 1374 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1375 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, (uintptr_t)rdma_req, 0); 1376 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTING); 1377 spdk_nvmf_request_exec(&rdma_req->req); 1378 break; 1379 case RDMA_REQUEST_STATE_EXECUTING: 1380 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, (uintptr_t)rdma_req, 0); 1381 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1382 * to escape this state. */ 1383 break; 1384 case RDMA_REQUEST_STATE_EXECUTED: 1385 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, (uintptr_t)rdma_req, 0); 1386 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1387 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1388 } else { 1389 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1390 } 1391 break; 1392 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1393 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, (uintptr_t)rdma_req, 0); 1394 rc = request_transfer_out(&rdma_req->req, &data_posted); 1395 assert(rc == 0); /* No good way to handle this currently */ 1396 spdk_nvmf_rdma_request_set_state(rdma_req, 1397 data_posted ? 1398 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1399 RDMA_REQUEST_STATE_COMPLETING); 1400 break; 1401 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1402 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 1403 (uintptr_t)rdma_req, 1404 0); 1405 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1406 * to escape this state. */ 1407 break; 1408 case RDMA_REQUEST_STATE_COMPLETING: 1409 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, (uintptr_t)rdma_req, 0); 1410 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1411 * to escape this state. */ 1412 break; 1413 case RDMA_REQUEST_STATE_COMPLETED: 1414 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, (uintptr_t)rdma_req, 0); 1415 1416 if (rdma_req->data_from_pool) { 1417 /* Put the buffer/s back in the pool */ 1418 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 1419 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 1420 rdma_req->req.iov[i].iov_base = NULL; 1421 rdma_req->data.buffers[i] = NULL; 1422 } 1423 rdma_req->data_from_pool = false; 1424 } 1425 rdma_req->req.length = 0; 1426 rdma_req->req.iovcnt = 0; 1427 rdma_req->req.data = NULL; 1428 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 1429 break; 1430 case RDMA_REQUEST_NUM_STATES: 1431 default: 1432 assert(0); 1433 break; 1434 } 1435 1436 if (rdma_req->state != prev_state) { 1437 progress = true; 1438 } 1439 } while (rdma_req->state != prev_state); 1440 1441 return progress; 1442 } 1443 1444 /* Public API callbacks begin here */ 1445 1446 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 1447 1448 static struct spdk_nvmf_transport * 1449 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 1450 { 1451 int rc; 1452 struct spdk_nvmf_rdma_transport *rtransport; 1453 struct spdk_nvmf_rdma_device *device, *tmp; 1454 struct ibv_context **contexts; 1455 uint32_t i; 1456 int flag; 1457 uint32_t sge_count; 1458 1459 rtransport = calloc(1, sizeof(*rtransport)); 1460 if (!rtransport) { 1461 return NULL; 1462 } 1463 1464 if (pthread_mutex_init(&rtransport->lock, NULL)) { 1465 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 1466 free(rtransport); 1467 return NULL; 1468 } 1469 1470 spdk_io_device_register(rtransport, spdk_nvmf_rdma_mgmt_channel_create, 1471 spdk_nvmf_rdma_mgmt_channel_destroy, 1472 sizeof(struct spdk_nvmf_rdma_mgmt_channel)); 1473 1474 TAILQ_INIT(&rtransport->devices); 1475 TAILQ_INIT(&rtransport->ports); 1476 1477 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 1478 1479 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 1480 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 1481 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 1482 " in_capsule_data_size=%d, max_aq_depth=%d\n", 1483 opts->max_queue_depth, 1484 opts->max_io_size, 1485 opts->max_qpairs_per_ctrlr, 1486 opts->io_unit_size, 1487 opts->in_capsule_data_size, 1488 opts->max_aq_depth); 1489 1490 /* I/O unit size cannot be larger than max I/O size */ 1491 if (opts->io_unit_size > opts->max_io_size) { 1492 opts->io_unit_size = opts->max_io_size; 1493 } 1494 1495 sge_count = opts->max_io_size / opts->io_unit_size; 1496 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 1497 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 1498 spdk_nvmf_rdma_destroy(&rtransport->transport); 1499 return NULL; 1500 } 1501 1502 rtransport->event_channel = rdma_create_event_channel(); 1503 if (rtransport->event_channel == NULL) { 1504 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 1505 spdk_nvmf_rdma_destroy(&rtransport->transport); 1506 return NULL; 1507 } 1508 1509 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 1510 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 1511 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 1512 rtransport->event_channel->fd, spdk_strerror(errno)); 1513 spdk_nvmf_rdma_destroy(&rtransport->transport); 1514 return NULL; 1515 } 1516 1517 rtransport->data_buf_pool = spdk_mempool_create("spdk_nvmf_rdma", 1518 opts->max_queue_depth * 4, /* The 4 is arbitrarily chosen. Needs to be configurable. */ 1519 opts->max_io_size + NVMF_DATA_BUFFER_ALIGNMENT, 1520 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 1521 SPDK_ENV_SOCKET_ID_ANY); 1522 if (!rtransport->data_buf_pool) { 1523 SPDK_ERRLOG("Unable to allocate buffer pool for poll group\n"); 1524 spdk_nvmf_rdma_destroy(&rtransport->transport); 1525 return NULL; 1526 } 1527 1528 contexts = rdma_get_devices(NULL); 1529 if (contexts == NULL) { 1530 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 1531 spdk_nvmf_rdma_destroy(&rtransport->transport); 1532 return NULL; 1533 } 1534 1535 i = 0; 1536 rc = 0; 1537 while (contexts[i] != NULL) { 1538 device = calloc(1, sizeof(*device)); 1539 if (!device) { 1540 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 1541 rc = -ENOMEM; 1542 break; 1543 } 1544 device->context = contexts[i]; 1545 rc = ibv_query_device(device->context, &device->attr); 1546 if (rc < 0) { 1547 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1548 free(device); 1549 break; 1550 1551 } 1552 /* set up device context async ev fd as NON_BLOCKING */ 1553 flag = fcntl(device->context->async_fd, F_GETFL); 1554 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 1555 if (rc < 0) { 1556 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 1557 free(device); 1558 break; 1559 } 1560 1561 device->pd = ibv_alloc_pd(device->context); 1562 if (!device->pd) { 1563 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 1564 free(device); 1565 rc = -1; 1566 break; 1567 } 1568 1569 device->map = spdk_mem_map_alloc(0, spdk_nvmf_rdma_mem_notify, device); 1570 if (!device->map) { 1571 SPDK_ERRLOG("Unable to allocate memory map for new poll group\n"); 1572 ibv_dealloc_pd(device->pd); 1573 free(device); 1574 rc = -1; 1575 break; 1576 } 1577 1578 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 1579 i++; 1580 } 1581 rdma_free_devices(contexts); 1582 1583 if (rc < 0) { 1584 spdk_nvmf_rdma_destroy(&rtransport->transport); 1585 return NULL; 1586 } 1587 1588 /* Set up poll descriptor array to monitor events from RDMA and IB 1589 * in a single poll syscall 1590 */ 1591 rtransport->npoll_fds = i + 1; 1592 i = 0; 1593 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 1594 if (rtransport->poll_fds == NULL) { 1595 SPDK_ERRLOG("poll_fds allocation failed\n"); 1596 spdk_nvmf_rdma_destroy(&rtransport->transport); 1597 return NULL; 1598 } 1599 1600 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 1601 rtransport->poll_fds[i++].events = POLLIN; 1602 1603 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 1604 rtransport->poll_fds[i].fd = device->context->async_fd; 1605 rtransport->poll_fds[i++].events = POLLIN; 1606 } 1607 1608 return &rtransport->transport; 1609 } 1610 1611 static int 1612 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 1613 { 1614 struct spdk_nvmf_rdma_transport *rtransport; 1615 struct spdk_nvmf_rdma_port *port, *port_tmp; 1616 struct spdk_nvmf_rdma_device *device, *device_tmp; 1617 1618 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1619 1620 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 1621 TAILQ_REMOVE(&rtransport->ports, port, link); 1622 rdma_destroy_id(port->id); 1623 free(port); 1624 } 1625 1626 if (rtransport->poll_fds != NULL) { 1627 free(rtransport->poll_fds); 1628 } 1629 1630 if (rtransport->event_channel != NULL) { 1631 rdma_destroy_event_channel(rtransport->event_channel); 1632 } 1633 1634 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 1635 TAILQ_REMOVE(&rtransport->devices, device, link); 1636 if (device->map) { 1637 spdk_mem_map_free(&device->map); 1638 } 1639 if (device->pd) { 1640 ibv_dealloc_pd(device->pd); 1641 } 1642 free(device); 1643 } 1644 1645 if (rtransport->data_buf_pool != NULL) { 1646 if (spdk_mempool_count(rtransport->data_buf_pool) != 1647 (transport->opts.max_queue_depth * 4)) { 1648 SPDK_ERRLOG("transport buffer pool count is %zu but should be %u\n", 1649 spdk_mempool_count(rtransport->data_buf_pool), 1650 transport->opts.max_queue_depth * 4); 1651 } 1652 } 1653 1654 spdk_mempool_free(rtransport->data_buf_pool); 1655 spdk_io_device_unregister(rtransport, NULL); 1656 pthread_mutex_destroy(&rtransport->lock); 1657 free(rtransport); 1658 1659 return 0; 1660 } 1661 1662 static int 1663 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 1664 const struct spdk_nvme_transport_id *trid) 1665 { 1666 struct spdk_nvmf_rdma_transport *rtransport; 1667 struct spdk_nvmf_rdma_device *device; 1668 struct spdk_nvmf_rdma_port *port_tmp, *port; 1669 struct addrinfo *res; 1670 struct addrinfo hints; 1671 int family; 1672 int rc; 1673 1674 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1675 1676 port = calloc(1, sizeof(*port)); 1677 if (!port) { 1678 return -ENOMEM; 1679 } 1680 1681 /* Selectively copy the trid. Things like NQN don't matter here - that 1682 * mapping is enforced elsewhere. 1683 */ 1684 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1685 port->trid.adrfam = trid->adrfam; 1686 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 1687 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 1688 1689 pthread_mutex_lock(&rtransport->lock); 1690 assert(rtransport->event_channel != NULL); 1691 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 1692 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 1693 port_tmp->ref++; 1694 free(port); 1695 /* Already listening at this address */ 1696 pthread_mutex_unlock(&rtransport->lock); 1697 return 0; 1698 } 1699 } 1700 1701 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 1702 if (rc < 0) { 1703 SPDK_ERRLOG("rdma_create_id() failed\n"); 1704 free(port); 1705 pthread_mutex_unlock(&rtransport->lock); 1706 return rc; 1707 } 1708 1709 switch (port->trid.adrfam) { 1710 case SPDK_NVMF_ADRFAM_IPV4: 1711 family = AF_INET; 1712 break; 1713 case SPDK_NVMF_ADRFAM_IPV6: 1714 family = AF_INET6; 1715 break; 1716 default: 1717 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 1718 free(port); 1719 pthread_mutex_unlock(&rtransport->lock); 1720 return -EINVAL; 1721 } 1722 1723 memset(&hints, 0, sizeof(hints)); 1724 hints.ai_family = family; 1725 hints.ai_socktype = SOCK_STREAM; 1726 hints.ai_protocol = 0; 1727 1728 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 1729 if (rc) { 1730 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 1731 free(port); 1732 pthread_mutex_unlock(&rtransport->lock); 1733 return -EINVAL; 1734 } 1735 1736 rc = rdma_bind_addr(port->id, res->ai_addr); 1737 freeaddrinfo(res); 1738 1739 if (rc < 0) { 1740 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 1741 rdma_destroy_id(port->id); 1742 free(port); 1743 pthread_mutex_unlock(&rtransport->lock); 1744 return rc; 1745 } 1746 1747 if (!port->id->verbs) { 1748 SPDK_ERRLOG("ibv_context is null\n"); 1749 rdma_destroy_id(port->id); 1750 free(port); 1751 pthread_mutex_unlock(&rtransport->lock); 1752 return -1; 1753 } 1754 1755 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 1756 if (rc < 0) { 1757 SPDK_ERRLOG("rdma_listen() failed\n"); 1758 rdma_destroy_id(port->id); 1759 free(port); 1760 pthread_mutex_unlock(&rtransport->lock); 1761 return rc; 1762 } 1763 1764 TAILQ_FOREACH(device, &rtransport->devices, link) { 1765 if (device->context == port->id->verbs) { 1766 port->device = device; 1767 break; 1768 } 1769 } 1770 if (!port->device) { 1771 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 1772 port->id->verbs); 1773 rdma_destroy_id(port->id); 1774 free(port); 1775 pthread_mutex_unlock(&rtransport->lock); 1776 return -EINVAL; 1777 } 1778 1779 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 1780 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 1781 1782 port->ref = 1; 1783 1784 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 1785 pthread_mutex_unlock(&rtransport->lock); 1786 1787 return 0; 1788 } 1789 1790 static int 1791 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 1792 const struct spdk_nvme_transport_id *_trid) 1793 { 1794 struct spdk_nvmf_rdma_transport *rtransport; 1795 struct spdk_nvmf_rdma_port *port, *tmp; 1796 struct spdk_nvme_transport_id trid = {}; 1797 1798 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1799 1800 /* Selectively copy the trid. Things like NQN don't matter here - that 1801 * mapping is enforced elsewhere. 1802 */ 1803 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1804 trid.adrfam = _trid->adrfam; 1805 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 1806 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 1807 1808 pthread_mutex_lock(&rtransport->lock); 1809 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 1810 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 1811 assert(port->ref > 0); 1812 port->ref--; 1813 if (port->ref == 0) { 1814 TAILQ_REMOVE(&rtransport->ports, port, link); 1815 rdma_destroy_id(port->id); 1816 free(port); 1817 } 1818 break; 1819 } 1820 } 1821 1822 pthread_mutex_unlock(&rtransport->lock); 1823 return 0; 1824 } 1825 1826 static void 1827 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 1828 { 1829 struct spdk_nvmf_rdma_transport *rtransport; 1830 struct rdma_cm_event *event; 1831 int rc; 1832 1833 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1834 1835 if (rtransport->event_channel == NULL) { 1836 return; 1837 } 1838 1839 while (1) { 1840 rc = rdma_get_cm_event(rtransport->event_channel, &event); 1841 if (rc == 0) { 1842 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 1843 1844 switch (event->event) { 1845 case RDMA_CM_EVENT_ADDR_RESOLVED: 1846 case RDMA_CM_EVENT_ADDR_ERROR: 1847 case RDMA_CM_EVENT_ROUTE_RESOLVED: 1848 case RDMA_CM_EVENT_ROUTE_ERROR: 1849 /* No action required. The target never attempts to resolve routes. */ 1850 break; 1851 case RDMA_CM_EVENT_CONNECT_REQUEST: 1852 rc = nvmf_rdma_connect(transport, event, cb_fn); 1853 if (rc < 0) { 1854 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 1855 break; 1856 } 1857 break; 1858 case RDMA_CM_EVENT_CONNECT_RESPONSE: 1859 /* The target never initiates a new connection. So this will not occur. */ 1860 break; 1861 case RDMA_CM_EVENT_CONNECT_ERROR: 1862 /* Can this happen? The docs say it can, but not sure what causes it. */ 1863 break; 1864 case RDMA_CM_EVENT_UNREACHABLE: 1865 case RDMA_CM_EVENT_REJECTED: 1866 /* These only occur on the client side. */ 1867 break; 1868 case RDMA_CM_EVENT_ESTABLISHED: 1869 /* TODO: Should we be waiting for this event anywhere? */ 1870 break; 1871 case RDMA_CM_EVENT_DISCONNECTED: 1872 case RDMA_CM_EVENT_DEVICE_REMOVAL: 1873 rc = nvmf_rdma_disconnect(event); 1874 if (rc < 0) { 1875 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 1876 break; 1877 } 1878 continue; 1879 case RDMA_CM_EVENT_MULTICAST_JOIN: 1880 case RDMA_CM_EVENT_MULTICAST_ERROR: 1881 /* Multicast is not used */ 1882 break; 1883 case RDMA_CM_EVENT_ADDR_CHANGE: 1884 /* Not utilizing this event */ 1885 break; 1886 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 1887 /* For now, do nothing. The target never re-uses queue pairs. */ 1888 break; 1889 default: 1890 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 1891 break; 1892 } 1893 1894 rdma_ack_cm_event(event); 1895 } else { 1896 if (errno != EAGAIN && errno != EWOULDBLOCK) { 1897 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 1898 } 1899 break; 1900 } 1901 } 1902 } 1903 1904 static bool 1905 spdk_nvmf_rdma_qpair_is_idle(struct spdk_nvmf_qpair *qpair) 1906 { 1907 int cur_queue_depth, cur_rdma_rw_depth; 1908 struct spdk_nvmf_rdma_qpair *rqpair; 1909 1910 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1911 cur_queue_depth = spdk_nvmf_rdma_cur_queue_depth(rqpair); 1912 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1913 1914 if (cur_queue_depth == 0 && cur_rdma_rw_depth == 0) { 1915 return true; 1916 } 1917 return false; 1918 } 1919 1920 static void 1921 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 1922 struct spdk_nvmf_rdma_qpair *rqpair) 1923 { 1924 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 1925 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 1926 1927 /* We process I/O in the data transfer pending queue at the highest priority. */ 1928 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING], 1929 state_link, req_tmp) { 1930 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1931 break; 1932 } 1933 } 1934 1935 /* The second highest priority is I/O waiting on memory buffers. */ 1936 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->ch->pending_data_buf_queue, link, 1937 req_tmp) { 1938 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1939 break; 1940 } 1941 } 1942 1943 if (rqpair->qpair_disconnected) { 1944 spdk_nvmf_rdma_qpair_destroy(rqpair); 1945 return; 1946 } 1947 1948 /* Do not process newly received commands if qp is in ERROR state, 1949 * wait till the recovery is complete. 1950 */ 1951 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 1952 return; 1953 } 1954 1955 /* The lowest priority is processing newly received commands */ 1956 TAILQ_FOREACH_SAFE(rdma_recv, &rqpair->incoming_queue, link, recv_tmp) { 1957 if (TAILQ_EMPTY(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE])) { 1958 break; 1959 } 1960 1961 rdma_req = TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE]); 1962 rdma_req->recv = rdma_recv; 1963 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEW); 1964 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1965 break; 1966 } 1967 } 1968 } 1969 1970 static void 1971 spdk_nvmf_rdma_drain_state_queue(struct spdk_nvmf_rdma_qpair *rqpair, 1972 enum spdk_nvmf_rdma_request_state state) 1973 { 1974 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 1975 struct spdk_nvmf_rdma_transport *rtransport; 1976 1977 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[state], state_link, req_tmp) { 1978 rtransport = SPDK_CONTAINEROF(rdma_req->req.qpair->transport, 1979 struct spdk_nvmf_rdma_transport, transport); 1980 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1981 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 1982 } 1983 } 1984 1985 static void 1986 spdk_nvmf_rdma_qpair_recover(struct spdk_nvmf_rdma_qpair *rqpair) 1987 { 1988 enum ibv_qp_state state, next_state; 1989 int recovered; 1990 struct spdk_nvmf_rdma_transport *rtransport; 1991 1992 if (!spdk_nvmf_rdma_qpair_is_idle(&rqpair->qpair)) { 1993 /* There must be outstanding requests down to media. 1994 * If so, wait till they're complete. 1995 */ 1996 assert(!TAILQ_EMPTY(&rqpair->qpair.outstanding)); 1997 return; 1998 } 1999 2000 state = rqpair->ibv_attr.qp_state; 2001 next_state = state; 2002 2003 SPDK_NOTICELOG("RDMA qpair %u is in state: %s\n", 2004 rqpair->qpair.qid, 2005 str_ibv_qp_state[state]); 2006 2007 if (!(state == IBV_QPS_ERR || state == IBV_QPS_RESET)) { 2008 SPDK_ERRLOG("Can't recover RDMA qpair %u from the state: %s\n", 2009 rqpair->qpair.qid, 2010 str_ibv_qp_state[state]); 2011 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2012 return; 2013 } 2014 2015 recovered = 0; 2016 while (!recovered) { 2017 switch (state) { 2018 case IBV_QPS_ERR: 2019 next_state = IBV_QPS_RESET; 2020 break; 2021 case IBV_QPS_RESET: 2022 next_state = IBV_QPS_INIT; 2023 break; 2024 case IBV_QPS_INIT: 2025 next_state = IBV_QPS_RTR; 2026 break; 2027 case IBV_QPS_RTR: 2028 next_state = IBV_QPS_RTS; 2029 break; 2030 case IBV_QPS_RTS: 2031 recovered = 1; 2032 break; 2033 default: 2034 SPDK_ERRLOG("RDMA qpair %u unexpected state for recovery: %u\n", 2035 rqpair->qpair.qid, state); 2036 goto error; 2037 } 2038 /* Do not transition into same state */ 2039 if (next_state == state) { 2040 break; 2041 } 2042 2043 if (spdk_nvmf_rdma_set_ibv_state(rqpair, next_state)) { 2044 goto error; 2045 } 2046 2047 state = next_state; 2048 } 2049 2050 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2051 struct spdk_nvmf_rdma_transport, 2052 transport); 2053 2054 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2055 2056 return; 2057 error: 2058 SPDK_NOTICELOG("RDMA qpair %u: recovery failed, disconnecting...\n", 2059 rqpair->qpair.qid); 2060 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2061 } 2062 2063 static void 2064 _spdk_nvmf_rdma_qp_cleanup_all_states(struct spdk_nvmf_rdma_qpair *rqpair) 2065 { 2066 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2067 2068 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_NEW); 2069 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 2070 2071 /* First wipe the requests waiting for buffer from the global list */ 2072 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[RDMA_REQUEST_STATE_NEED_BUFFER], link, req_tmp) { 2073 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 2074 } 2075 /* Then drain the requests through the rdma queue */ 2076 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_NEED_BUFFER); 2077 2078 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_EXECUTING); 2079 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2080 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2081 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_COMPLETING); 2082 } 2083 2084 static void 2085 _spdk_nvmf_rdma_qp_error(void *arg) 2086 { 2087 struct spdk_nvmf_rdma_qpair *rqpair = arg; 2088 enum ibv_qp_state state; 2089 2090 state = rqpair->ibv_attr.qp_state; 2091 if (state != IBV_QPS_ERR) { 2092 /* Error was already recovered */ 2093 return; 2094 } 2095 2096 if (spdk_nvmf_qpair_is_admin_queue(&rqpair->qpair)) { 2097 spdk_nvmf_ctrlr_abort_aer(rqpair->qpair.ctrlr); 2098 } 2099 _spdk_nvmf_rdma_qp_cleanup_all_states(rqpair); 2100 spdk_nvmf_rdma_qpair_recover(rqpair); 2101 } 2102 2103 static void 2104 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2105 { 2106 int rc; 2107 struct spdk_nvmf_rdma_qpair *rqpair; 2108 struct ibv_async_event event; 2109 enum ibv_qp_state state; 2110 2111 rc = ibv_get_async_event(device->context, &event); 2112 2113 if (rc) { 2114 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2115 errno, spdk_strerror(errno)); 2116 return; 2117 } 2118 2119 SPDK_NOTICELOG("Async event: %s\n", 2120 ibv_event_type_str(event.event_type)); 2121 2122 rqpair = event.element.qp->qp_context; 2123 2124 switch (event.event_type) { 2125 case IBV_EVENT_QP_FATAL: 2126 case IBV_EVENT_QP_LAST_WQE_REACHED: 2127 /* This call is thread-safe. Immediately update the IBV state on error notification. */ 2128 spdk_nvmf_rdma_update_ibv_state(rqpair); 2129 2130 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_error, rqpair); 2131 break; 2132 case IBV_EVENT_SQ_DRAINED: 2133 /* This event occurs frequently in both error and non-error states. 2134 * Check if the qpair is in an error state before sending a message. 2135 * Note that we're not on the correct thread to access the qpair, but 2136 * the operations that the below calls make all happen to be thread 2137 * safe. */ 2138 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2139 if (state == IBV_QPS_ERR) { 2140 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_error, rqpair); 2141 } 2142 break; 2143 case IBV_EVENT_QP_REQ_ERR: 2144 case IBV_EVENT_QP_ACCESS_ERR: 2145 case IBV_EVENT_COMM_EST: 2146 case IBV_EVENT_PATH_MIG: 2147 case IBV_EVENT_PATH_MIG_ERR: 2148 spdk_nvmf_rdma_update_ibv_state(rqpair); 2149 break; 2150 case IBV_EVENT_CQ_ERR: 2151 case IBV_EVENT_DEVICE_FATAL: 2152 case IBV_EVENT_PORT_ACTIVE: 2153 case IBV_EVENT_PORT_ERR: 2154 case IBV_EVENT_LID_CHANGE: 2155 case IBV_EVENT_PKEY_CHANGE: 2156 case IBV_EVENT_SM_CHANGE: 2157 case IBV_EVENT_SRQ_ERR: 2158 case IBV_EVENT_SRQ_LIMIT_REACHED: 2159 case IBV_EVENT_CLIENT_REREGISTER: 2160 case IBV_EVENT_GID_CHANGE: 2161 default: 2162 break; 2163 } 2164 ibv_ack_async_event(&event); 2165 } 2166 2167 static void 2168 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2169 { 2170 int nfds, i = 0; 2171 struct spdk_nvmf_rdma_transport *rtransport; 2172 struct spdk_nvmf_rdma_device *device, *tmp; 2173 2174 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2175 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2176 2177 if (nfds <= 0) { 2178 return; 2179 } 2180 2181 /* The first poll descriptor is RDMA CM event */ 2182 if (rtransport->poll_fds[i++].revents & POLLIN) { 2183 spdk_nvmf_process_cm_event(transport, cb_fn); 2184 nfds--; 2185 } 2186 2187 if (nfds == 0) { 2188 return; 2189 } 2190 2191 /* Second and subsequent poll descriptors are IB async events */ 2192 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2193 if (rtransport->poll_fds[i++].revents & POLLIN) { 2194 spdk_nvmf_process_ib_event(device); 2195 nfds--; 2196 } 2197 } 2198 /* check all flagged fd's have been served */ 2199 assert(nfds == 0); 2200 } 2201 2202 static void 2203 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2204 struct spdk_nvme_transport_id *trid, 2205 struct spdk_nvmf_discovery_log_page_entry *entry) 2206 { 2207 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2208 entry->adrfam = trid->adrfam; 2209 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2210 2211 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2212 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2213 2214 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2215 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2216 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2217 } 2218 2219 static struct spdk_nvmf_transport_poll_group * 2220 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2221 { 2222 struct spdk_nvmf_rdma_transport *rtransport; 2223 struct spdk_nvmf_rdma_poll_group *rgroup; 2224 struct spdk_nvmf_rdma_poller *poller; 2225 struct spdk_nvmf_rdma_device *device; 2226 2227 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2228 2229 rgroup = calloc(1, sizeof(*rgroup)); 2230 if (!rgroup) { 2231 return NULL; 2232 } 2233 2234 TAILQ_INIT(&rgroup->pollers); 2235 2236 pthread_mutex_lock(&rtransport->lock); 2237 TAILQ_FOREACH(device, &rtransport->devices, link) { 2238 poller = calloc(1, sizeof(*poller)); 2239 if (!poller) { 2240 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2241 free(rgroup); 2242 pthread_mutex_unlock(&rtransport->lock); 2243 return NULL; 2244 } 2245 2246 poller->device = device; 2247 poller->group = rgroup; 2248 2249 TAILQ_INIT(&poller->qpairs); 2250 2251 poller->cq = ibv_create_cq(device->context, NVMF_RDMA_CQ_SIZE, poller, NULL, 0); 2252 if (!poller->cq) { 2253 SPDK_ERRLOG("Unable to create completion queue\n"); 2254 free(poller); 2255 free(rgroup); 2256 pthread_mutex_unlock(&rtransport->lock); 2257 return NULL; 2258 } 2259 2260 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2261 } 2262 2263 pthread_mutex_unlock(&rtransport->lock); 2264 return &rgroup->group; 2265 } 2266 2267 static void 2268 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2269 { 2270 struct spdk_nvmf_rdma_poll_group *rgroup; 2271 struct spdk_nvmf_rdma_poller *poller, *tmp; 2272 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2273 2274 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2275 2276 if (!rgroup) { 2277 return; 2278 } 2279 2280 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2281 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2282 2283 if (poller->cq) { 2284 ibv_destroy_cq(poller->cq); 2285 } 2286 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 2287 _spdk_nvmf_rdma_qp_cleanup_all_states(qpair); 2288 spdk_nvmf_rdma_qpair_destroy(qpair); 2289 } 2290 2291 free(poller); 2292 } 2293 2294 free(rgroup); 2295 } 2296 2297 static int 2298 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 2299 struct spdk_nvmf_qpair *qpair) 2300 { 2301 struct spdk_nvmf_rdma_transport *rtransport; 2302 struct spdk_nvmf_rdma_poll_group *rgroup; 2303 struct spdk_nvmf_rdma_qpair *rqpair; 2304 struct spdk_nvmf_rdma_device *device; 2305 struct spdk_nvmf_rdma_poller *poller; 2306 int rc; 2307 2308 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 2309 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2310 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2311 2312 device = rqpair->port->device; 2313 2314 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 2315 if (poller->device == device) { 2316 break; 2317 } 2318 } 2319 2320 if (!poller) { 2321 SPDK_ERRLOG("No poller found for device.\n"); 2322 return -1; 2323 } 2324 2325 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 2326 rqpair->poller = poller; 2327 2328 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 2329 if (rc < 0) { 2330 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 2331 return -1; 2332 } 2333 2334 rqpair->mgmt_channel = spdk_get_io_channel(rtransport); 2335 if (!rqpair->mgmt_channel) { 2336 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2337 spdk_nvmf_rdma_qpair_destroy(rqpair); 2338 return -1; 2339 } 2340 2341 rqpair->ch = spdk_io_channel_get_ctx(rqpair->mgmt_channel); 2342 assert(rqpair->ch != NULL); 2343 2344 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 2345 if (rc) { 2346 /* Try to reject, but we probably can't */ 2347 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2348 spdk_nvmf_rdma_qpair_destroy(rqpair); 2349 return -1; 2350 } 2351 2352 spdk_nvmf_rdma_update_ibv_state(rqpair); 2353 2354 return 0; 2355 } 2356 2357 static int 2358 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 2359 { 2360 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2361 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2362 struct spdk_nvmf_rdma_transport, transport); 2363 2364 if (rdma_req->data_from_pool) { 2365 /* Put the buffer/s back in the pool */ 2366 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 2367 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 2368 rdma_req->req.iov[i].iov_base = NULL; 2369 rdma_req->data.buffers[i] = NULL; 2370 } 2371 rdma_req->data_from_pool = false; 2372 } 2373 rdma_req->req.length = 0; 2374 rdma_req->req.iovcnt = 0; 2375 rdma_req->req.data = NULL; 2376 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 2377 return 0; 2378 } 2379 2380 static int 2381 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 2382 { 2383 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2384 struct spdk_nvmf_rdma_transport, transport); 2385 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 2386 struct spdk_nvmf_rdma_request, req); 2387 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 2388 struct spdk_nvmf_rdma_qpair, qpair); 2389 2390 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2391 /* The connection is alive, so process the request as normal */ 2392 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTED); 2393 } else { 2394 /* The connection is dead. Move the request directly to the completed state. */ 2395 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2396 } 2397 2398 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2399 2400 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE && rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 2401 /* If the NVMe-oF layer thinks the connection is active, but the RDMA layer thinks 2402 * the connection is dead, perform error recovery. */ 2403 spdk_nvmf_rdma_qpair_recover(rqpair); 2404 } 2405 2406 return 0; 2407 } 2408 2409 static void 2410 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 2411 { 2412 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2413 2414 spdk_nvmf_rdma_qpair_destroy(rqpair); 2415 } 2416 2417 static struct spdk_nvmf_rdma_request * 2418 get_rdma_req_from_wc(struct ibv_wc *wc) 2419 { 2420 struct spdk_nvmf_rdma_request *rdma_req; 2421 2422 rdma_req = (struct spdk_nvmf_rdma_request *)wc->wr_id; 2423 assert(rdma_req != NULL); 2424 2425 #ifdef DEBUG 2426 struct spdk_nvmf_rdma_qpair *rqpair; 2427 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2428 2429 assert(rdma_req - rqpair->reqs >= 0); 2430 assert(rdma_req - rqpair->reqs < (ptrdiff_t)rqpair->max_queue_depth); 2431 #endif 2432 2433 return rdma_req; 2434 } 2435 2436 static struct spdk_nvmf_rdma_recv * 2437 get_rdma_recv_from_wc(struct ibv_wc *wc) 2438 { 2439 struct spdk_nvmf_rdma_recv *rdma_recv; 2440 2441 assert(wc->byte_len >= sizeof(struct spdk_nvmf_capsule_cmd)); 2442 2443 rdma_recv = (struct spdk_nvmf_rdma_recv *)wc->wr_id; 2444 assert(rdma_recv != NULL); 2445 2446 #ifdef DEBUG 2447 struct spdk_nvmf_rdma_qpair *rqpair = rdma_recv->qpair; 2448 2449 assert(rdma_recv - rqpair->recvs >= 0); 2450 assert(rdma_recv - rqpair->recvs < (ptrdiff_t)rqpair->max_queue_depth); 2451 #endif 2452 2453 return rdma_recv; 2454 } 2455 2456 #ifdef DEBUG 2457 static int 2458 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 2459 { 2460 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 2461 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 2462 } 2463 #endif 2464 2465 static int 2466 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 2467 struct spdk_nvmf_rdma_poller *rpoller) 2468 { 2469 struct ibv_wc wc[32]; 2470 struct spdk_nvmf_rdma_request *rdma_req; 2471 struct spdk_nvmf_rdma_recv *rdma_recv; 2472 struct spdk_nvmf_rdma_qpair *rqpair; 2473 int reaped, i; 2474 int count = 0; 2475 bool error = false; 2476 2477 /* Poll for completing operations. */ 2478 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 2479 if (reaped < 0) { 2480 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2481 errno, spdk_strerror(errno)); 2482 return -1; 2483 } 2484 2485 for (i = 0; i < reaped; i++) { 2486 /* Handle error conditions */ 2487 if (wc[i].status) { 2488 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 2489 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 2490 error = true; 2491 2492 switch (wc[i].opcode) { 2493 case IBV_WC_SEND: 2494 case IBV_WC_RDMA_WRITE: 2495 case IBV_WC_RDMA_READ: 2496 rdma_req = get_rdma_req_from_wc(&wc[i]); 2497 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2498 2499 /* We're going to kill the connection, so force the request into 2500 * the completed state. */ 2501 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2502 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2503 break; 2504 case IBV_WC_RECV: 2505 rdma_recv = get_rdma_recv_from_wc(&wc[i]); 2506 rqpair = rdma_recv->qpair; 2507 2508 /* Dump this into the incoming queue. This gets cleaned up when 2509 * the queue pair disconnects. */ 2510 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2511 default: 2512 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2513 continue; 2514 } 2515 2516 /* Begin disconnecting the qpair. This is ok to call multiple times if lots of 2517 * errors occur on the same qpair in the same ibv_poll_cq batch. */ 2518 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2519 2520 continue; 2521 } 2522 2523 switch (wc[i].opcode) { 2524 case IBV_WC_SEND: 2525 rdma_req = get_rdma_req_from_wc(&wc[i]); 2526 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2527 2528 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 2529 2530 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2531 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2532 2533 count++; 2534 2535 /* Try to process other queued requests */ 2536 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2537 break; 2538 2539 case IBV_WC_RDMA_WRITE: 2540 rdma_req = get_rdma_req_from_wc(&wc[i]); 2541 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2542 2543 /* Try to process other queued requests */ 2544 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2545 break; 2546 2547 case IBV_WC_RDMA_READ: 2548 rdma_req = get_rdma_req_from_wc(&wc[i]); 2549 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2550 2551 assert(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2552 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 2553 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2554 2555 /* Try to process other queued requests */ 2556 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2557 break; 2558 2559 case IBV_WC_RECV: 2560 rdma_recv = get_rdma_recv_from_wc(&wc[i]); 2561 rqpair = rdma_recv->qpair; 2562 2563 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2564 /* Try to process other queued requests */ 2565 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2566 break; 2567 2568 default: 2569 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2570 continue; 2571 } 2572 } 2573 2574 if (error == true) { 2575 return -1; 2576 } 2577 2578 return count; 2579 } 2580 2581 static int 2582 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 2583 { 2584 struct spdk_nvmf_rdma_transport *rtransport; 2585 struct spdk_nvmf_rdma_poll_group *rgroup; 2586 struct spdk_nvmf_rdma_poller *rpoller; 2587 int count, rc; 2588 2589 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 2590 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2591 2592 count = 0; 2593 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2594 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 2595 if (rc < 0) { 2596 return rc; 2597 } 2598 count += rc; 2599 } 2600 2601 return count; 2602 } 2603 2604 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 2605 .type = SPDK_NVME_TRANSPORT_RDMA, 2606 .create = spdk_nvmf_rdma_create, 2607 .destroy = spdk_nvmf_rdma_destroy, 2608 2609 .listen = spdk_nvmf_rdma_listen, 2610 .stop_listen = spdk_nvmf_rdma_stop_listen, 2611 .accept = spdk_nvmf_rdma_accept, 2612 2613 .listener_discover = spdk_nvmf_rdma_discover, 2614 2615 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 2616 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 2617 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 2618 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 2619 2620 .req_free = spdk_nvmf_rdma_request_free, 2621 .req_complete = spdk_nvmf_rdma_request_complete, 2622 2623 .qpair_fini = spdk_nvmf_rdma_close_qpair, 2624 .qpair_is_idle = spdk_nvmf_rdma_qpair_is_idle, 2625 2626 }; 2627 2628 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 2629