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 } 950 951 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 952 spdk_nvmf_rdma_update_ibv_state(rqpair); 953 954 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 955 956 return 0; 957 } 958 959 #ifdef DEBUG 960 static const char *CM_EVENT_STR[] = { 961 "RDMA_CM_EVENT_ADDR_RESOLVED", 962 "RDMA_CM_EVENT_ADDR_ERROR", 963 "RDMA_CM_EVENT_ROUTE_RESOLVED", 964 "RDMA_CM_EVENT_ROUTE_ERROR", 965 "RDMA_CM_EVENT_CONNECT_REQUEST", 966 "RDMA_CM_EVENT_CONNECT_RESPONSE", 967 "RDMA_CM_EVENT_CONNECT_ERROR", 968 "RDMA_CM_EVENT_UNREACHABLE", 969 "RDMA_CM_EVENT_REJECTED", 970 "RDMA_CM_EVENT_ESTABLISHED", 971 "RDMA_CM_EVENT_DISCONNECTED", 972 "RDMA_CM_EVENT_DEVICE_REMOVAL", 973 "RDMA_CM_EVENT_MULTICAST_JOIN", 974 "RDMA_CM_EVENT_MULTICAST_ERROR", 975 "RDMA_CM_EVENT_ADDR_CHANGE", 976 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 977 }; 978 #endif /* DEBUG */ 979 980 static void 981 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 982 { 983 struct spdk_nvmf_rdma_transport *rtransport; 984 struct rdma_cm_event *event; 985 int rc; 986 987 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 988 989 if (rtransport->event_channel == NULL) { 990 return; 991 } 992 993 while (1) { 994 rc = rdma_get_cm_event(rtransport->event_channel, &event); 995 if (rc == 0) { 996 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 997 998 switch (event->event) { 999 case RDMA_CM_EVENT_ADDR_RESOLVED: 1000 case RDMA_CM_EVENT_ADDR_ERROR: 1001 case RDMA_CM_EVENT_ROUTE_RESOLVED: 1002 case RDMA_CM_EVENT_ROUTE_ERROR: 1003 /* No action required. The target never attempts to resolve routes. */ 1004 break; 1005 case RDMA_CM_EVENT_CONNECT_REQUEST: 1006 rc = nvmf_rdma_connect(transport, event, cb_fn); 1007 if (rc < 0) { 1008 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 1009 break; 1010 } 1011 break; 1012 case RDMA_CM_EVENT_CONNECT_RESPONSE: 1013 /* The target never initiates a new connection. So this will not occur. */ 1014 break; 1015 case RDMA_CM_EVENT_CONNECT_ERROR: 1016 /* Can this happen? The docs say it can, but not sure what causes it. */ 1017 break; 1018 case RDMA_CM_EVENT_UNREACHABLE: 1019 case RDMA_CM_EVENT_REJECTED: 1020 /* These only occur on the client side. */ 1021 break; 1022 case RDMA_CM_EVENT_ESTABLISHED: 1023 /* TODO: Should we be waiting for this event anywhere? */ 1024 break; 1025 case RDMA_CM_EVENT_DISCONNECTED: 1026 case RDMA_CM_EVENT_DEVICE_REMOVAL: 1027 rc = nvmf_rdma_disconnect(event); 1028 if (rc < 0) { 1029 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 1030 break; 1031 } 1032 break; 1033 case RDMA_CM_EVENT_MULTICAST_JOIN: 1034 case RDMA_CM_EVENT_MULTICAST_ERROR: 1035 /* Multicast is not used */ 1036 break; 1037 case RDMA_CM_EVENT_ADDR_CHANGE: 1038 /* Not utilizing this event */ 1039 break; 1040 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 1041 /* For now, do nothing. The target never re-uses queue pairs. */ 1042 break; 1043 default: 1044 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 1045 break; 1046 } 1047 1048 rdma_ack_cm_event(event); 1049 } else { 1050 if (errno != EAGAIN && errno != EWOULDBLOCK) { 1051 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 1052 } 1053 break; 1054 } 1055 } 1056 } 1057 1058 static int 1059 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1060 enum spdk_mem_map_notify_action action, 1061 void *vaddr, size_t size) 1062 { 1063 struct spdk_nvmf_rdma_device *device = cb_ctx; 1064 struct ibv_pd *pd = device->pd; 1065 struct ibv_mr *mr; 1066 1067 switch (action) { 1068 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1069 mr = ibv_reg_mr(pd, vaddr, size, 1070 IBV_ACCESS_LOCAL_WRITE | 1071 IBV_ACCESS_REMOTE_READ | 1072 IBV_ACCESS_REMOTE_WRITE); 1073 if (mr == NULL) { 1074 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1075 return -1; 1076 } else { 1077 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1078 } 1079 break; 1080 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1081 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, size); 1082 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1083 if (mr) { 1084 ibv_dereg_mr(mr); 1085 } 1086 break; 1087 } 1088 1089 return 0; 1090 } 1091 1092 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1093 1094 static spdk_nvme_data_transfer_t 1095 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1096 { 1097 enum spdk_nvme_data_transfer xfer; 1098 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1099 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1100 1101 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1102 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1103 rdma_req->rsp.wr.imm_data = 0; 1104 #endif 1105 1106 /* Figure out data transfer direction */ 1107 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1108 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1109 } else { 1110 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1111 1112 /* Some admin commands are special cases */ 1113 if ((rdma_req->req.qpair->qid == 0) && 1114 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1115 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1116 switch (cmd->cdw10 & 0xff) { 1117 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1118 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1119 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1120 break; 1121 default: 1122 xfer = SPDK_NVME_DATA_NONE; 1123 } 1124 } 1125 } 1126 1127 if (xfer == SPDK_NVME_DATA_NONE) { 1128 return xfer; 1129 } 1130 1131 /* Even for commands that may transfer data, they could have specified 0 length. 1132 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1133 */ 1134 switch (sgl->generic.type) { 1135 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1136 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1137 case SPDK_NVME_SGL_TYPE_SEGMENT: 1138 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1139 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1140 if (sgl->unkeyed.length == 0) { 1141 xfer = SPDK_NVME_DATA_NONE; 1142 } 1143 break; 1144 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1145 if (sgl->keyed.length == 0) { 1146 xfer = SPDK_NVME_DATA_NONE; 1147 } 1148 break; 1149 } 1150 1151 return xfer; 1152 } 1153 1154 static int 1155 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1156 struct spdk_nvmf_rdma_device *device, 1157 struct spdk_nvmf_rdma_request *rdma_req) 1158 { 1159 void *buf = NULL; 1160 uint32_t length = rdma_req->req.length; 1161 uint32_t i = 0; 1162 1163 rdma_req->req.iovcnt = 0; 1164 while (length) { 1165 buf = spdk_mempool_get(rtransport->data_buf_pool); 1166 if (!buf) { 1167 goto nomem; 1168 } 1169 1170 rdma_req->req.iov[i].iov_base = (void *)((uintptr_t)(buf + NVMF_DATA_BUFFER_MASK) & 1171 ~NVMF_DATA_BUFFER_MASK); 1172 rdma_req->req.iov[i].iov_len = spdk_min(length, rtransport->transport.opts.io_unit_size); 1173 rdma_req->req.iovcnt++; 1174 rdma_req->data.buffers[i] = buf; 1175 rdma_req->data.wr.sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[i].iov_base); 1176 rdma_req->data.wr.sg_list[i].length = rdma_req->req.iov[i].iov_len; 1177 rdma_req->data.wr.sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1178 (uint64_t)buf, rdma_req->req.iov[i].iov_len))->lkey; 1179 1180 length -= rdma_req->req.iov[i].iov_len; 1181 i++; 1182 } 1183 1184 rdma_req->data_from_pool = true; 1185 1186 return 0; 1187 1188 nomem: 1189 while (i) { 1190 i--; 1191 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->req.iov[i].iov_base); 1192 rdma_req->req.iov[i].iov_base = NULL; 1193 rdma_req->req.iov[i].iov_len = 0; 1194 1195 rdma_req->data.wr.sg_list[i].addr = 0; 1196 rdma_req->data.wr.sg_list[i].length = 0; 1197 rdma_req->data.wr.sg_list[i].lkey = 0; 1198 } 1199 rdma_req->req.iovcnt = 0; 1200 return -ENOMEM; 1201 } 1202 1203 static int 1204 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1205 struct spdk_nvmf_rdma_device *device, 1206 struct spdk_nvmf_rdma_request *rdma_req) 1207 { 1208 struct spdk_nvme_cmd *cmd; 1209 struct spdk_nvme_cpl *rsp; 1210 struct spdk_nvme_sgl_descriptor *sgl; 1211 1212 cmd = &rdma_req->req.cmd->nvme_cmd; 1213 rsp = &rdma_req->req.rsp->nvme_cpl; 1214 sgl = &cmd->dptr.sgl1; 1215 1216 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1217 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1218 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1219 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1220 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1221 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1222 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1223 return -1; 1224 } 1225 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1226 /** 1227 * These vendor IDs are assigned by the IEEE and an ID of 0 implies Soft-RoCE. 1228 * The Soft-RoCE RXE driver does not currently support send with invalidate. 1229 * There are changes making their way through the kernel now that will enable 1230 * this feature. When they are merged, we can conditionally enable this feature. 1231 * 1232 * todo: enable this for versions of the kernel rxe driver that support it. 1233 */ 1234 if (device->attr.vendor_id != 0) { 1235 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1236 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1237 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1238 } 1239 } 1240 #endif 1241 1242 /* fill request length and populate iovs */ 1243 rdma_req->req.length = sgl->keyed.length; 1244 1245 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1246 /* No available buffers. Queue this request up. */ 1247 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1248 return 0; 1249 } 1250 1251 /* backward compatible */ 1252 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1253 1254 /* rdma wr specifics */ 1255 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1256 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1257 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1258 1259 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1260 rdma_req->req.iovcnt); 1261 1262 return 0; 1263 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1264 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1265 uint64_t offset = sgl->address; 1266 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1267 1268 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1269 offset, sgl->unkeyed.length); 1270 1271 if (offset > max_len) { 1272 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1273 offset, max_len); 1274 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1275 return -1; 1276 } 1277 max_len -= (uint32_t)offset; 1278 1279 if (sgl->unkeyed.length > max_len) { 1280 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1281 sgl->unkeyed.length, max_len); 1282 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1283 return -1; 1284 } 1285 1286 rdma_req->req.data = rdma_req->recv->buf + offset; 1287 rdma_req->data_from_pool = false; 1288 rdma_req->req.length = sgl->unkeyed.length; 1289 1290 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1291 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1292 rdma_req->req.iovcnt = 1; 1293 1294 return 0; 1295 } 1296 1297 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1298 sgl->generic.type, sgl->generic.subtype); 1299 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1300 return -1; 1301 } 1302 1303 static bool 1304 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1305 struct spdk_nvmf_rdma_request *rdma_req) 1306 { 1307 struct spdk_nvmf_rdma_qpair *rqpair; 1308 struct spdk_nvmf_rdma_device *device; 1309 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1310 int rc; 1311 struct spdk_nvmf_rdma_recv *rdma_recv; 1312 enum spdk_nvmf_rdma_request_state prev_state; 1313 bool progress = false; 1314 int data_posted; 1315 int cur_rdma_rw_depth; 1316 1317 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1318 device = rqpair->port->device; 1319 1320 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1321 1322 /* If the queue pair is in an error state, force the request to the completed state 1323 * to release resources. */ 1324 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1325 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1326 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1327 } 1328 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1329 } 1330 1331 /* The loop here is to allow for several back-to-back state changes. */ 1332 do { 1333 prev_state = rdma_req->state; 1334 1335 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1336 1337 switch (rdma_req->state) { 1338 case RDMA_REQUEST_STATE_FREE: 1339 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1340 * to escape this state. */ 1341 break; 1342 case RDMA_REQUEST_STATE_NEW: 1343 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, (uintptr_t)rdma_req, 0); 1344 rdma_recv = rdma_req->recv; 1345 1346 /* The first element of the SGL is the NVMe command */ 1347 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1348 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1349 1350 TAILQ_REMOVE(&rqpair->incoming_queue, rdma_recv, link); 1351 1352 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 1353 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1354 break; 1355 } 1356 1357 /* The next state transition depends on the data transfer needs of this request. */ 1358 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1359 1360 /* If no data to transfer, ready to execute. */ 1361 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1362 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1363 break; 1364 } 1365 1366 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEED_BUFFER); 1367 TAILQ_INSERT_TAIL(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1368 break; 1369 case RDMA_REQUEST_STATE_NEED_BUFFER: 1370 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, (uintptr_t)rdma_req, 0); 1371 1372 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1373 1374 if (rdma_req != TAILQ_FIRST(&rqpair->ch->pending_data_buf_queue)) { 1375 /* This request needs to wait in line to obtain a buffer */ 1376 break; 1377 } 1378 1379 /* Try to get a data buffer */ 1380 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1381 if (rc < 0) { 1382 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1383 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1384 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1385 break; 1386 } 1387 1388 if (!rdma_req->req.data) { 1389 /* No buffers available. */ 1390 break; 1391 } 1392 1393 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1394 1395 /* If data is transferring from host to controller and the data didn't 1396 * arrive using in capsule data, we need to do a transfer from the host. 1397 */ 1398 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1399 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1400 break; 1401 } 1402 1403 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1404 break; 1405 case RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING: 1406 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 0, 0, 1407 (uintptr_t)rdma_req, 0); 1408 1409 if (rdma_req != TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING])) { 1410 /* This request needs to wait in line to perform RDMA */ 1411 break; 1412 } 1413 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1414 1415 if (cur_rdma_rw_depth >= rqpair->max_rw_depth) { 1416 /* R/W queue is full, need to wait */ 1417 break; 1418 } 1419 1420 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1421 rc = request_transfer_in(&rdma_req->req); 1422 if (!rc) { 1423 spdk_nvmf_rdma_request_set_state(rdma_req, 1424 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1425 } else { 1426 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1427 spdk_nvmf_rdma_request_set_state(rdma_req, 1428 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1429 } 1430 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1431 /* The data transfer will be kicked off from 1432 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1433 */ 1434 spdk_nvmf_rdma_request_set_state(rdma_req, 1435 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1436 } else { 1437 SPDK_ERRLOG("Cannot perform data transfer, unknown state: %u\n", 1438 rdma_req->req.xfer); 1439 assert(0); 1440 } 1441 break; 1442 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1443 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1444 (uintptr_t)rdma_req, 0); 1445 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1446 * to escape this state. */ 1447 break; 1448 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1449 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, (uintptr_t)rdma_req, 0); 1450 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTING); 1451 spdk_nvmf_request_exec(&rdma_req->req); 1452 break; 1453 case RDMA_REQUEST_STATE_EXECUTING: 1454 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, (uintptr_t)rdma_req, 0); 1455 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1456 * to escape this state. */ 1457 break; 1458 case RDMA_REQUEST_STATE_EXECUTED: 1459 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, (uintptr_t)rdma_req, 0); 1460 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1461 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1462 } else { 1463 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1464 } 1465 break; 1466 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1467 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, (uintptr_t)rdma_req, 0); 1468 rc = request_transfer_out(&rdma_req->req, &data_posted); 1469 assert(rc == 0); /* No good way to handle this currently */ 1470 spdk_nvmf_rdma_request_set_state(rdma_req, 1471 data_posted ? 1472 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1473 RDMA_REQUEST_STATE_COMPLETING); 1474 break; 1475 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1476 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 1477 (uintptr_t)rdma_req, 1478 0); 1479 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1480 * to escape this state. */ 1481 break; 1482 case RDMA_REQUEST_STATE_COMPLETING: 1483 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, (uintptr_t)rdma_req, 0); 1484 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1485 * to escape this state. */ 1486 break; 1487 case RDMA_REQUEST_STATE_COMPLETED: 1488 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, (uintptr_t)rdma_req, 0); 1489 1490 if (rdma_req->data_from_pool) { 1491 /* Put the buffer/s back in the pool */ 1492 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 1493 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 1494 rdma_req->req.iov[i].iov_base = NULL; 1495 rdma_req->data.buffers[i] = NULL; 1496 } 1497 rdma_req->data_from_pool = false; 1498 } 1499 rdma_req->req.length = 0; 1500 rdma_req->req.iovcnt = 0; 1501 rdma_req->req.data = NULL; 1502 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 1503 break; 1504 case RDMA_REQUEST_NUM_STATES: 1505 default: 1506 assert(0); 1507 break; 1508 } 1509 1510 if (rdma_req->state != prev_state) { 1511 progress = true; 1512 } 1513 } while (rdma_req->state != prev_state); 1514 1515 return progress; 1516 } 1517 1518 /* Public API callbacks begin here */ 1519 1520 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 1521 1522 static struct spdk_nvmf_transport * 1523 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 1524 { 1525 int rc; 1526 struct spdk_nvmf_rdma_transport *rtransport; 1527 struct spdk_nvmf_rdma_device *device, *tmp; 1528 struct ibv_context **contexts; 1529 uint32_t i; 1530 int flag; 1531 uint32_t sge_count; 1532 1533 rtransport = calloc(1, sizeof(*rtransport)); 1534 if (!rtransport) { 1535 return NULL; 1536 } 1537 1538 if (pthread_mutex_init(&rtransport->lock, NULL)) { 1539 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 1540 free(rtransport); 1541 return NULL; 1542 } 1543 1544 spdk_io_device_register(rtransport, spdk_nvmf_rdma_mgmt_channel_create, 1545 spdk_nvmf_rdma_mgmt_channel_destroy, 1546 sizeof(struct spdk_nvmf_rdma_mgmt_channel)); 1547 1548 TAILQ_INIT(&rtransport->devices); 1549 TAILQ_INIT(&rtransport->ports); 1550 1551 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 1552 1553 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 1554 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 1555 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 1556 " in_capsule_data_size=%d, max_aq_depth=%d\n", 1557 opts->max_queue_depth, 1558 opts->max_io_size, 1559 opts->max_qpairs_per_ctrlr, 1560 opts->io_unit_size, 1561 opts->in_capsule_data_size, 1562 opts->max_aq_depth); 1563 1564 /* I/O unit size cannot be larger than max I/O size */ 1565 if (opts->io_unit_size > opts->max_io_size) { 1566 opts->io_unit_size = opts->max_io_size; 1567 } 1568 1569 sge_count = opts->max_io_size / opts->io_unit_size; 1570 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 1571 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 1572 spdk_nvmf_rdma_destroy(&rtransport->transport); 1573 return NULL; 1574 } 1575 1576 rtransport->event_channel = rdma_create_event_channel(); 1577 if (rtransport->event_channel == NULL) { 1578 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 1579 spdk_nvmf_rdma_destroy(&rtransport->transport); 1580 return NULL; 1581 } 1582 1583 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 1584 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 1585 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 1586 rtransport->event_channel->fd, spdk_strerror(errno)); 1587 spdk_nvmf_rdma_destroy(&rtransport->transport); 1588 return NULL; 1589 } 1590 1591 rtransport->data_buf_pool = spdk_mempool_create("spdk_nvmf_rdma", 1592 opts->max_queue_depth * 4, /* The 4 is arbitrarily chosen. Needs to be configurable. */ 1593 opts->max_io_size + NVMF_DATA_BUFFER_ALIGNMENT, 1594 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 1595 SPDK_ENV_SOCKET_ID_ANY); 1596 if (!rtransport->data_buf_pool) { 1597 SPDK_ERRLOG("Unable to allocate buffer pool for poll group\n"); 1598 spdk_nvmf_rdma_destroy(&rtransport->transport); 1599 return NULL; 1600 } 1601 1602 contexts = rdma_get_devices(NULL); 1603 if (contexts == NULL) { 1604 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 1605 spdk_nvmf_rdma_destroy(&rtransport->transport); 1606 return NULL; 1607 } 1608 1609 i = 0; 1610 rc = 0; 1611 while (contexts[i] != NULL) { 1612 device = calloc(1, sizeof(*device)); 1613 if (!device) { 1614 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 1615 rc = -ENOMEM; 1616 break; 1617 } 1618 device->context = contexts[i]; 1619 rc = ibv_query_device(device->context, &device->attr); 1620 if (rc < 0) { 1621 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1622 free(device); 1623 break; 1624 1625 } 1626 /* set up device context async ev fd as NON_BLOCKING */ 1627 flag = fcntl(device->context->async_fd, F_GETFL); 1628 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 1629 if (rc < 0) { 1630 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 1631 free(device); 1632 break; 1633 } 1634 1635 device->pd = ibv_alloc_pd(device->context); 1636 if (!device->pd) { 1637 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 1638 free(device); 1639 rc = -1; 1640 break; 1641 } 1642 1643 device->map = spdk_mem_map_alloc(0, spdk_nvmf_rdma_mem_notify, device); 1644 if (!device->map) { 1645 SPDK_ERRLOG("Unable to allocate memory map for new poll group\n"); 1646 ibv_dealloc_pd(device->pd); 1647 free(device); 1648 rc = -1; 1649 break; 1650 } 1651 1652 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 1653 i++; 1654 } 1655 rdma_free_devices(contexts); 1656 1657 if (rc < 0) { 1658 spdk_nvmf_rdma_destroy(&rtransport->transport); 1659 return NULL; 1660 } 1661 1662 /* Set up poll descriptor array to monitor events from RDMA and IB 1663 * in a single poll syscall 1664 */ 1665 rtransport->npoll_fds = i + 1; 1666 i = 0; 1667 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 1668 if (rtransport->poll_fds == NULL) { 1669 SPDK_ERRLOG("poll_fds allocation failed\n"); 1670 spdk_nvmf_rdma_destroy(&rtransport->transport); 1671 return NULL; 1672 } 1673 1674 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 1675 rtransport->poll_fds[i++].events = POLLIN; 1676 1677 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 1678 rtransport->poll_fds[i].fd = device->context->async_fd; 1679 rtransport->poll_fds[i++].events = POLLIN; 1680 } 1681 1682 return &rtransport->transport; 1683 } 1684 1685 static int 1686 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 1687 { 1688 struct spdk_nvmf_rdma_transport *rtransport; 1689 struct spdk_nvmf_rdma_port *port, *port_tmp; 1690 struct spdk_nvmf_rdma_device *device, *device_tmp; 1691 1692 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1693 1694 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 1695 TAILQ_REMOVE(&rtransport->ports, port, link); 1696 rdma_destroy_id(port->id); 1697 free(port); 1698 } 1699 1700 if (rtransport->poll_fds != NULL) { 1701 free(rtransport->poll_fds); 1702 } 1703 1704 if (rtransport->event_channel != NULL) { 1705 rdma_destroy_event_channel(rtransport->event_channel); 1706 } 1707 1708 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 1709 TAILQ_REMOVE(&rtransport->devices, device, link); 1710 if (device->map) { 1711 spdk_mem_map_free(&device->map); 1712 } 1713 if (device->pd) { 1714 ibv_dealloc_pd(device->pd); 1715 } 1716 free(device); 1717 } 1718 1719 if (rtransport->data_buf_pool != NULL) { 1720 if (spdk_mempool_count(rtransport->data_buf_pool) != 1721 (transport->opts.max_queue_depth * 4)) { 1722 SPDK_ERRLOG("transport buffer pool count is %zu but should be %u\n", 1723 spdk_mempool_count(rtransport->data_buf_pool), 1724 transport->opts.max_queue_depth * 4); 1725 } 1726 } 1727 1728 spdk_mempool_free(rtransport->data_buf_pool); 1729 spdk_io_device_unregister(rtransport, NULL); 1730 pthread_mutex_destroy(&rtransport->lock); 1731 free(rtransport); 1732 1733 return 0; 1734 } 1735 1736 static int 1737 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 1738 const struct spdk_nvme_transport_id *trid) 1739 { 1740 struct spdk_nvmf_rdma_transport *rtransport; 1741 struct spdk_nvmf_rdma_device *device; 1742 struct spdk_nvmf_rdma_port *port_tmp, *port; 1743 struct addrinfo *res; 1744 struct addrinfo hints; 1745 int family; 1746 int rc; 1747 1748 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1749 1750 port = calloc(1, sizeof(*port)); 1751 if (!port) { 1752 return -ENOMEM; 1753 } 1754 1755 /* Selectively copy the trid. Things like NQN don't matter here - that 1756 * mapping is enforced elsewhere. 1757 */ 1758 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1759 port->trid.adrfam = trid->adrfam; 1760 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 1761 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 1762 1763 pthread_mutex_lock(&rtransport->lock); 1764 assert(rtransport->event_channel != NULL); 1765 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 1766 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 1767 port_tmp->ref++; 1768 free(port); 1769 /* Already listening at this address */ 1770 pthread_mutex_unlock(&rtransport->lock); 1771 return 0; 1772 } 1773 } 1774 1775 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 1776 if (rc < 0) { 1777 SPDK_ERRLOG("rdma_create_id() failed\n"); 1778 free(port); 1779 pthread_mutex_unlock(&rtransport->lock); 1780 return rc; 1781 } 1782 1783 switch (port->trid.adrfam) { 1784 case SPDK_NVMF_ADRFAM_IPV4: 1785 family = AF_INET; 1786 break; 1787 case SPDK_NVMF_ADRFAM_IPV6: 1788 family = AF_INET6; 1789 break; 1790 default: 1791 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 1792 free(port); 1793 pthread_mutex_unlock(&rtransport->lock); 1794 return -EINVAL; 1795 } 1796 1797 memset(&hints, 0, sizeof(hints)); 1798 hints.ai_family = family; 1799 hints.ai_socktype = SOCK_STREAM; 1800 hints.ai_protocol = 0; 1801 1802 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 1803 if (rc) { 1804 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 1805 free(port); 1806 pthread_mutex_unlock(&rtransport->lock); 1807 return -EINVAL; 1808 } 1809 1810 rc = rdma_bind_addr(port->id, res->ai_addr); 1811 freeaddrinfo(res); 1812 1813 if (rc < 0) { 1814 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 1815 rdma_destroy_id(port->id); 1816 free(port); 1817 pthread_mutex_unlock(&rtransport->lock); 1818 return rc; 1819 } 1820 1821 if (!port->id->verbs) { 1822 SPDK_ERRLOG("ibv_context is null\n"); 1823 rdma_destroy_id(port->id); 1824 free(port); 1825 pthread_mutex_unlock(&rtransport->lock); 1826 return -1; 1827 } 1828 1829 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 1830 if (rc < 0) { 1831 SPDK_ERRLOG("rdma_listen() failed\n"); 1832 rdma_destroy_id(port->id); 1833 free(port); 1834 pthread_mutex_unlock(&rtransport->lock); 1835 return rc; 1836 } 1837 1838 TAILQ_FOREACH(device, &rtransport->devices, link) { 1839 if (device->context == port->id->verbs) { 1840 port->device = device; 1841 break; 1842 } 1843 } 1844 if (!port->device) { 1845 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 1846 port->id->verbs); 1847 rdma_destroy_id(port->id); 1848 free(port); 1849 pthread_mutex_unlock(&rtransport->lock); 1850 return -EINVAL; 1851 } 1852 1853 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 1854 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 1855 1856 port->ref = 1; 1857 1858 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 1859 pthread_mutex_unlock(&rtransport->lock); 1860 1861 return 0; 1862 } 1863 1864 static int 1865 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 1866 const struct spdk_nvme_transport_id *_trid) 1867 { 1868 struct spdk_nvmf_rdma_transport *rtransport; 1869 struct spdk_nvmf_rdma_port *port, *tmp; 1870 struct spdk_nvme_transport_id trid = {}; 1871 1872 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1873 1874 /* Selectively copy the trid. Things like NQN don't matter here - that 1875 * mapping is enforced elsewhere. 1876 */ 1877 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1878 trid.adrfam = _trid->adrfam; 1879 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 1880 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 1881 1882 pthread_mutex_lock(&rtransport->lock); 1883 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 1884 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 1885 assert(port->ref > 0); 1886 port->ref--; 1887 if (port->ref == 0) { 1888 TAILQ_REMOVE(&rtransport->ports, port, link); 1889 rdma_destroy_id(port->id); 1890 free(port); 1891 } 1892 break; 1893 } 1894 } 1895 1896 pthread_mutex_unlock(&rtransport->lock); 1897 return 0; 1898 } 1899 1900 static bool 1901 spdk_nvmf_rdma_qpair_is_idle(struct spdk_nvmf_qpair *qpair) 1902 { 1903 int cur_queue_depth, cur_rdma_rw_depth; 1904 struct spdk_nvmf_rdma_qpair *rqpair; 1905 1906 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1907 cur_queue_depth = spdk_nvmf_rdma_cur_queue_depth(rqpair); 1908 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1909 1910 if (cur_queue_depth == 0 && cur_rdma_rw_depth == 0) { 1911 return true; 1912 } 1913 return false; 1914 } 1915 1916 static void 1917 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 1918 struct spdk_nvmf_rdma_qpair *rqpair) 1919 { 1920 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 1921 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 1922 1923 /* We process I/O in the data transfer pending queue at the highest priority. */ 1924 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING], 1925 state_link, req_tmp) { 1926 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1927 break; 1928 } 1929 } 1930 1931 /* The second highest priority is I/O waiting on memory buffers. */ 1932 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->ch->pending_data_buf_queue, link, 1933 req_tmp) { 1934 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1935 break; 1936 } 1937 } 1938 1939 if (rqpair->qpair_disconnected) { 1940 spdk_nvmf_rdma_qpair_destroy(rqpair); 1941 return; 1942 } 1943 1944 /* Do not process newly received commands if qp is in ERROR state, 1945 * wait till the recovery is complete. 1946 */ 1947 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 1948 return; 1949 } 1950 1951 /* The lowest priority is processing newly received commands */ 1952 TAILQ_FOREACH_SAFE(rdma_recv, &rqpair->incoming_queue, link, recv_tmp) { 1953 if (TAILQ_EMPTY(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE])) { 1954 break; 1955 } 1956 1957 rdma_req = TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE]); 1958 rdma_req->recv = rdma_recv; 1959 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEW); 1960 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1961 break; 1962 } 1963 } 1964 } 1965 1966 static void 1967 spdk_nvmf_rdma_drain_state_queue(struct spdk_nvmf_rdma_qpair *rqpair, 1968 enum spdk_nvmf_rdma_request_state state) 1969 { 1970 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 1971 struct spdk_nvmf_rdma_transport *rtransport; 1972 1973 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[state], state_link, req_tmp) { 1974 rtransport = SPDK_CONTAINEROF(rdma_req->req.qpair->transport, 1975 struct spdk_nvmf_rdma_transport, transport); 1976 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1977 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 1978 } 1979 } 1980 1981 static void 1982 spdk_nvmf_rdma_qpair_recover(struct spdk_nvmf_rdma_qpair *rqpair) 1983 { 1984 enum ibv_qp_state state, next_state; 1985 int recovered; 1986 struct spdk_nvmf_rdma_transport *rtransport; 1987 1988 if (!spdk_nvmf_rdma_qpair_is_idle(&rqpair->qpair)) { 1989 /* There must be outstanding requests down to media. 1990 * If so, wait till they're complete. 1991 */ 1992 assert(!TAILQ_EMPTY(&rqpair->qpair.outstanding)); 1993 return; 1994 } 1995 1996 state = rqpair->ibv_attr.qp_state; 1997 next_state = state; 1998 1999 SPDK_NOTICELOG("RDMA qpair %u is in state: %s\n", 2000 rqpair->qpair.qid, 2001 str_ibv_qp_state[state]); 2002 2003 if (!(state == IBV_QPS_ERR || state == IBV_QPS_RESET)) { 2004 SPDK_ERRLOG("Can't recover RDMA qpair %u from the state: %s\n", 2005 rqpair->qpair.qid, 2006 str_ibv_qp_state[state]); 2007 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2008 return; 2009 } 2010 2011 recovered = 0; 2012 while (!recovered) { 2013 switch (state) { 2014 case IBV_QPS_ERR: 2015 next_state = IBV_QPS_RESET; 2016 break; 2017 case IBV_QPS_RESET: 2018 next_state = IBV_QPS_INIT; 2019 break; 2020 case IBV_QPS_INIT: 2021 next_state = IBV_QPS_RTR; 2022 break; 2023 case IBV_QPS_RTR: 2024 next_state = IBV_QPS_RTS; 2025 break; 2026 case IBV_QPS_RTS: 2027 recovered = 1; 2028 break; 2029 default: 2030 SPDK_ERRLOG("RDMA qpair %u unexpected state for recovery: %u\n", 2031 rqpair->qpair.qid, state); 2032 goto error; 2033 } 2034 /* Do not transition into same state */ 2035 if (next_state == state) { 2036 break; 2037 } 2038 2039 if (spdk_nvmf_rdma_set_ibv_state(rqpair, next_state)) { 2040 goto error; 2041 } 2042 2043 state = next_state; 2044 } 2045 2046 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2047 struct spdk_nvmf_rdma_transport, 2048 transport); 2049 2050 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2051 2052 return; 2053 error: 2054 SPDK_NOTICELOG("RDMA qpair %u: recovery failed, disconnecting...\n", 2055 rqpair->qpair.qid); 2056 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2057 } 2058 2059 /* Clean up only the states that can be aborted at any time */ 2060 static void 2061 _spdk_nvmf_rdma_qp_cleanup_safe_states(struct spdk_nvmf_rdma_qpair *rqpair) 2062 { 2063 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2064 2065 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_NEW); 2066 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[RDMA_REQUEST_STATE_NEED_BUFFER], link, req_tmp) { 2067 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 2068 } 2069 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_NEED_BUFFER); 2070 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 2071 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 2072 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_EXECUTED); 2073 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 2074 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_COMPLETED); 2075 } 2076 2077 /* This cleans up all memory. It is only safe to use if the rest of the software stack 2078 * has been shut down */ 2079 static void 2080 _spdk_nvmf_rdma_qp_cleanup_all_states(struct spdk_nvmf_rdma_qpair *rqpair) 2081 { 2082 _spdk_nvmf_rdma_qp_cleanup_safe_states(rqpair); 2083 2084 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_EXECUTING); 2085 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2086 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2087 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_COMPLETING); 2088 } 2089 2090 static void 2091 _spdk_nvmf_rdma_qp_error(void *arg) 2092 { 2093 struct spdk_nvmf_rdma_qpair *rqpair = arg; 2094 enum ibv_qp_state state; 2095 2096 state = rqpair->ibv_attr.qp_state; 2097 if (state != IBV_QPS_ERR) { 2098 /* Error was already recovered */ 2099 return; 2100 } 2101 2102 if (spdk_nvmf_qpair_is_admin_queue(&rqpair->qpair)) { 2103 spdk_nvmf_ctrlr_abort_aer(rqpair->qpair.ctrlr); 2104 } 2105 2106 _spdk_nvmf_rdma_qp_cleanup_safe_states(rqpair); 2107 2108 /* Attempt recovery. This will exit without recovering if I/O requests 2109 * are still outstanding */ 2110 spdk_nvmf_rdma_qpair_recover(rqpair); 2111 } 2112 2113 static void 2114 _spdk_nvmf_rdma_qp_last_wqe(void *arg) 2115 { 2116 struct spdk_nvmf_rdma_qpair *rqpair = arg; 2117 enum ibv_qp_state state; 2118 2119 state = rqpair->ibv_attr.qp_state; 2120 if (state != IBV_QPS_ERR) { 2121 /* Error was already recovered */ 2122 return; 2123 } 2124 2125 /* Clear out the states that are safe to clear any time, plus the 2126 * RDMA data transfer states. */ 2127 _spdk_nvmf_rdma_qp_cleanup_safe_states(rqpair); 2128 2129 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2130 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2131 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_COMPLETING); 2132 2133 spdk_nvmf_rdma_qpair_recover(rqpair); 2134 } 2135 2136 static void 2137 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2138 { 2139 int rc; 2140 struct spdk_nvmf_rdma_qpair *rqpair; 2141 struct ibv_async_event event; 2142 enum ibv_qp_state state; 2143 2144 rc = ibv_get_async_event(device->context, &event); 2145 2146 if (rc) { 2147 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2148 errno, spdk_strerror(errno)); 2149 return; 2150 } 2151 2152 SPDK_NOTICELOG("Async event: %s\n", 2153 ibv_event_type_str(event.event_type)); 2154 2155 rqpair = event.element.qp->qp_context; 2156 2157 switch (event.event_type) { 2158 case IBV_EVENT_QP_FATAL: 2159 spdk_nvmf_rdma_update_ibv_state(rqpair); 2160 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_error, rqpair); 2161 break; 2162 case IBV_EVENT_QP_LAST_WQE_REACHED: 2163 spdk_nvmf_rdma_update_ibv_state(rqpair); 2164 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_last_wqe, rqpair); 2165 break; 2166 case IBV_EVENT_SQ_DRAINED: 2167 /* This event occurs frequently in both error and non-error states. 2168 * Check if the qpair is in an error state before sending a message. 2169 * Note that we're not on the correct thread to access the qpair, but 2170 * the operations that the below calls make all happen to be thread 2171 * safe. */ 2172 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2173 if (state == IBV_QPS_ERR) { 2174 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_error, rqpair); 2175 } 2176 break; 2177 case IBV_EVENT_QP_REQ_ERR: 2178 case IBV_EVENT_QP_ACCESS_ERR: 2179 case IBV_EVENT_COMM_EST: 2180 case IBV_EVENT_PATH_MIG: 2181 case IBV_EVENT_PATH_MIG_ERR: 2182 spdk_nvmf_rdma_update_ibv_state(rqpair); 2183 break; 2184 case IBV_EVENT_CQ_ERR: 2185 case IBV_EVENT_DEVICE_FATAL: 2186 case IBV_EVENT_PORT_ACTIVE: 2187 case IBV_EVENT_PORT_ERR: 2188 case IBV_EVENT_LID_CHANGE: 2189 case IBV_EVENT_PKEY_CHANGE: 2190 case IBV_EVENT_SM_CHANGE: 2191 case IBV_EVENT_SRQ_ERR: 2192 case IBV_EVENT_SRQ_LIMIT_REACHED: 2193 case IBV_EVENT_CLIENT_REREGISTER: 2194 case IBV_EVENT_GID_CHANGE: 2195 default: 2196 break; 2197 } 2198 ibv_ack_async_event(&event); 2199 } 2200 2201 static void 2202 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2203 { 2204 int nfds, i = 0; 2205 struct spdk_nvmf_rdma_transport *rtransport; 2206 struct spdk_nvmf_rdma_device *device, *tmp; 2207 2208 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2209 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2210 2211 if (nfds <= 0) { 2212 return; 2213 } 2214 2215 /* The first poll descriptor is RDMA CM event */ 2216 if (rtransport->poll_fds[i++].revents & POLLIN) { 2217 spdk_nvmf_process_cm_event(transport, cb_fn); 2218 nfds--; 2219 } 2220 2221 if (nfds == 0) { 2222 return; 2223 } 2224 2225 /* Second and subsequent poll descriptors are IB async events */ 2226 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2227 if (rtransport->poll_fds[i++].revents & POLLIN) { 2228 spdk_nvmf_process_ib_event(device); 2229 nfds--; 2230 } 2231 } 2232 /* check all flagged fd's have been served */ 2233 assert(nfds == 0); 2234 } 2235 2236 static void 2237 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2238 struct spdk_nvme_transport_id *trid, 2239 struct spdk_nvmf_discovery_log_page_entry *entry) 2240 { 2241 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2242 entry->adrfam = trid->adrfam; 2243 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2244 2245 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2246 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2247 2248 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2249 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2250 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2251 } 2252 2253 static struct spdk_nvmf_transport_poll_group * 2254 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2255 { 2256 struct spdk_nvmf_rdma_transport *rtransport; 2257 struct spdk_nvmf_rdma_poll_group *rgroup; 2258 struct spdk_nvmf_rdma_poller *poller; 2259 struct spdk_nvmf_rdma_device *device; 2260 2261 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2262 2263 rgroup = calloc(1, sizeof(*rgroup)); 2264 if (!rgroup) { 2265 return NULL; 2266 } 2267 2268 TAILQ_INIT(&rgroup->pollers); 2269 2270 pthread_mutex_lock(&rtransport->lock); 2271 TAILQ_FOREACH(device, &rtransport->devices, link) { 2272 poller = calloc(1, sizeof(*poller)); 2273 if (!poller) { 2274 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2275 free(rgroup); 2276 pthread_mutex_unlock(&rtransport->lock); 2277 return NULL; 2278 } 2279 2280 poller->device = device; 2281 poller->group = rgroup; 2282 2283 TAILQ_INIT(&poller->qpairs); 2284 2285 poller->cq = ibv_create_cq(device->context, NVMF_RDMA_CQ_SIZE, poller, NULL, 0); 2286 if (!poller->cq) { 2287 SPDK_ERRLOG("Unable to create completion queue\n"); 2288 free(poller); 2289 free(rgroup); 2290 pthread_mutex_unlock(&rtransport->lock); 2291 return NULL; 2292 } 2293 2294 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2295 } 2296 2297 pthread_mutex_unlock(&rtransport->lock); 2298 return &rgroup->group; 2299 } 2300 2301 static void 2302 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2303 { 2304 struct spdk_nvmf_rdma_poll_group *rgroup; 2305 struct spdk_nvmf_rdma_poller *poller, *tmp; 2306 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2307 2308 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2309 2310 if (!rgroup) { 2311 return; 2312 } 2313 2314 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2315 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2316 2317 if (poller->cq) { 2318 ibv_destroy_cq(poller->cq); 2319 } 2320 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 2321 _spdk_nvmf_rdma_qp_cleanup_all_states(qpair); 2322 spdk_nvmf_rdma_qpair_destroy(qpair); 2323 } 2324 2325 free(poller); 2326 } 2327 2328 free(rgroup); 2329 } 2330 2331 static int 2332 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 2333 struct spdk_nvmf_qpair *qpair) 2334 { 2335 struct spdk_nvmf_rdma_transport *rtransport; 2336 struct spdk_nvmf_rdma_poll_group *rgroup; 2337 struct spdk_nvmf_rdma_qpair *rqpair; 2338 struct spdk_nvmf_rdma_device *device; 2339 struct spdk_nvmf_rdma_poller *poller; 2340 int rc; 2341 2342 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 2343 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2344 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2345 2346 device = rqpair->port->device; 2347 2348 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 2349 if (poller->device == device) { 2350 break; 2351 } 2352 } 2353 2354 if (!poller) { 2355 SPDK_ERRLOG("No poller found for device.\n"); 2356 return -1; 2357 } 2358 2359 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 2360 rqpair->poller = poller; 2361 2362 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 2363 if (rc < 0) { 2364 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 2365 return -1; 2366 } 2367 2368 rqpair->mgmt_channel = spdk_get_io_channel(rtransport); 2369 if (!rqpair->mgmt_channel) { 2370 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2371 spdk_nvmf_rdma_qpair_destroy(rqpair); 2372 return -1; 2373 } 2374 2375 rqpair->ch = spdk_io_channel_get_ctx(rqpair->mgmt_channel); 2376 assert(rqpair->ch != NULL); 2377 2378 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 2379 if (rc) { 2380 /* Try to reject, but we probably can't */ 2381 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2382 spdk_nvmf_rdma_qpair_destroy(rqpair); 2383 return -1; 2384 } 2385 2386 spdk_nvmf_rdma_update_ibv_state(rqpair); 2387 2388 return 0; 2389 } 2390 2391 static int 2392 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 2393 { 2394 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2395 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2396 struct spdk_nvmf_rdma_transport, transport); 2397 2398 if (rdma_req->data_from_pool) { 2399 /* Put the buffer/s back in the pool */ 2400 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 2401 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 2402 rdma_req->req.iov[i].iov_base = NULL; 2403 rdma_req->data.buffers[i] = NULL; 2404 } 2405 rdma_req->data_from_pool = false; 2406 } 2407 rdma_req->req.length = 0; 2408 rdma_req->req.iovcnt = 0; 2409 rdma_req->req.data = NULL; 2410 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 2411 return 0; 2412 } 2413 2414 static int 2415 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 2416 { 2417 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2418 struct spdk_nvmf_rdma_transport, transport); 2419 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 2420 struct spdk_nvmf_rdma_request, req); 2421 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 2422 struct spdk_nvmf_rdma_qpair, qpair); 2423 2424 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2425 /* The connection is alive, so process the request as normal */ 2426 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTED); 2427 } else { 2428 /* The connection is dead. Move the request directly to the completed state. */ 2429 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2430 } 2431 2432 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2433 2434 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE && rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 2435 /* If the NVMe-oF layer thinks the connection is active, but the RDMA layer thinks 2436 * the connection is dead, perform error recovery. */ 2437 spdk_nvmf_rdma_qpair_recover(rqpair); 2438 } 2439 2440 return 0; 2441 } 2442 2443 static void 2444 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 2445 { 2446 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2447 2448 spdk_nvmf_rdma_qpair_destroy(rqpair); 2449 } 2450 2451 static struct spdk_nvmf_rdma_request * 2452 get_rdma_req_from_wc(struct ibv_wc *wc) 2453 { 2454 struct spdk_nvmf_rdma_request *rdma_req; 2455 2456 rdma_req = (struct spdk_nvmf_rdma_request *)wc->wr_id; 2457 assert(rdma_req != NULL); 2458 2459 #ifdef DEBUG 2460 struct spdk_nvmf_rdma_qpair *rqpair; 2461 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2462 2463 assert(rdma_req - rqpair->reqs >= 0); 2464 assert(rdma_req - rqpair->reqs < (ptrdiff_t)rqpair->max_queue_depth); 2465 #endif 2466 2467 return rdma_req; 2468 } 2469 2470 static struct spdk_nvmf_rdma_recv * 2471 get_rdma_recv_from_wc(struct ibv_wc *wc) 2472 { 2473 struct spdk_nvmf_rdma_recv *rdma_recv; 2474 2475 assert(wc->byte_len >= sizeof(struct spdk_nvmf_capsule_cmd)); 2476 2477 rdma_recv = (struct spdk_nvmf_rdma_recv *)wc->wr_id; 2478 assert(rdma_recv != NULL); 2479 2480 #ifdef DEBUG 2481 struct spdk_nvmf_rdma_qpair *rqpair = rdma_recv->qpair; 2482 2483 assert(rdma_recv - rqpair->recvs >= 0); 2484 assert(rdma_recv - rqpair->recvs < (ptrdiff_t)rqpair->max_queue_depth); 2485 #endif 2486 2487 return rdma_recv; 2488 } 2489 2490 #ifdef DEBUG 2491 static int 2492 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 2493 { 2494 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 2495 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 2496 } 2497 #endif 2498 2499 static int 2500 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 2501 struct spdk_nvmf_rdma_poller *rpoller) 2502 { 2503 struct ibv_wc wc[32]; 2504 struct spdk_nvmf_rdma_request *rdma_req; 2505 struct spdk_nvmf_rdma_recv *rdma_recv; 2506 struct spdk_nvmf_rdma_qpair *rqpair; 2507 int reaped, i; 2508 int count = 0; 2509 bool error = false; 2510 2511 /* Poll for completing operations. */ 2512 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 2513 if (reaped < 0) { 2514 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2515 errno, spdk_strerror(errno)); 2516 return -1; 2517 } 2518 2519 for (i = 0; i < reaped; i++) { 2520 /* Handle error conditions */ 2521 if (wc[i].status) { 2522 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 2523 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 2524 error = true; 2525 2526 switch (wc[i].opcode) { 2527 case IBV_WC_SEND: 2528 case IBV_WC_RDMA_WRITE: 2529 case IBV_WC_RDMA_READ: 2530 rdma_req = get_rdma_req_from_wc(&wc[i]); 2531 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2532 2533 /* We're going to kill the connection, so force the request into 2534 * the completed state. */ 2535 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2536 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2537 break; 2538 case IBV_WC_RECV: 2539 rdma_recv = get_rdma_recv_from_wc(&wc[i]); 2540 rqpair = rdma_recv->qpair; 2541 2542 /* Dump this into the incoming queue. This gets cleaned up when 2543 * the queue pair disconnects. */ 2544 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2545 default: 2546 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2547 continue; 2548 } 2549 2550 /* Begin disconnecting the qpair. This is ok to call multiple times if lots of 2551 * errors occur on the same qpair in the same ibv_poll_cq batch. */ 2552 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2553 2554 continue; 2555 } 2556 2557 switch (wc[i].opcode) { 2558 case IBV_WC_SEND: 2559 rdma_req = get_rdma_req_from_wc(&wc[i]); 2560 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2561 2562 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 2563 2564 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2565 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2566 2567 count++; 2568 2569 /* Try to process other queued requests */ 2570 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2571 break; 2572 2573 case IBV_WC_RDMA_WRITE: 2574 rdma_req = get_rdma_req_from_wc(&wc[i]); 2575 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2576 2577 /* Try to process other queued requests */ 2578 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2579 break; 2580 2581 case IBV_WC_RDMA_READ: 2582 rdma_req = get_rdma_req_from_wc(&wc[i]); 2583 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2584 2585 assert(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2586 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 2587 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2588 2589 /* Try to process other queued requests */ 2590 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2591 break; 2592 2593 case IBV_WC_RECV: 2594 rdma_recv = get_rdma_recv_from_wc(&wc[i]); 2595 rqpair = rdma_recv->qpair; 2596 2597 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2598 /* Try to process other queued requests */ 2599 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2600 break; 2601 2602 default: 2603 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2604 continue; 2605 } 2606 } 2607 2608 if (error == true) { 2609 return -1; 2610 } 2611 2612 return count; 2613 } 2614 2615 static int 2616 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 2617 { 2618 struct spdk_nvmf_rdma_transport *rtransport; 2619 struct spdk_nvmf_rdma_poll_group *rgroup; 2620 struct spdk_nvmf_rdma_poller *rpoller; 2621 int count, rc; 2622 2623 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 2624 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2625 2626 count = 0; 2627 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2628 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 2629 if (rc < 0) { 2630 return rc; 2631 } 2632 count += rc; 2633 } 2634 2635 return count; 2636 } 2637 2638 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 2639 .type = SPDK_NVME_TRANSPORT_RDMA, 2640 .create = spdk_nvmf_rdma_create, 2641 .destroy = spdk_nvmf_rdma_destroy, 2642 2643 .listen = spdk_nvmf_rdma_listen, 2644 .stop_listen = spdk_nvmf_rdma_stop_listen, 2645 .accept = spdk_nvmf_rdma_accept, 2646 2647 .listener_discover = spdk_nvmf_rdma_discover, 2648 2649 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 2650 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 2651 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 2652 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 2653 2654 .req_free = spdk_nvmf_rdma_request_free, 2655 .req_complete = spdk_nvmf_rdma_request_complete, 2656 2657 .qpair_fini = spdk_nvmf_rdma_close_qpair, 2658 .qpair_is_idle = spdk_nvmf_rdma_qpair_is_idle, 2659 2660 }; 2661 2662 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 2663