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, 1, 0, "cmid: "); 136 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", "", 137 TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 138 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, 0, "cmid: "); 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, 1, 0, "cmid: "); 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, 1, 0, "cmid: "); 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, 1, 0, "cmid: "); 148 spdk_trace_register_description("RDMA_REQ_EXECUTING", "", 149 TRACE_RDMA_REQUEST_STATE_EXECUTING, 150 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, 0, "cmid: "); 151 spdk_trace_register_description("RDMA_REQ_EXECUTED", "", 152 TRACE_RDMA_REQUEST_STATE_EXECUTED, 153 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, 0, "cmid: "); 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, 1, 0, "cmid: "); 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, 1, 0, "cmid: "); 160 spdk_trace_register_description("RDMA_REQ_COMPLETING_INCAPSULE", "", 161 TRACE_RDMA_REQUEST_STATE_COMPLETING, 162 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, 0, "cmid: "); 163 spdk_trace_register_description("RDMA_REQ_COMPLETED", "", 164 TRACE_RDMA_REQUEST_STATE_COMPLETED, 165 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, 0, "cmid: "); 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, 1344 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1345 rdma_recv = rdma_req->recv; 1346 1347 /* The first element of the SGL is the NVMe command */ 1348 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1349 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1350 1351 TAILQ_REMOVE(&rqpair->incoming_queue, rdma_recv, link); 1352 1353 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 1354 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1355 break; 1356 } 1357 1358 /* The next state transition depends on the data transfer needs of this request. */ 1359 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1360 1361 /* If no data to transfer, ready to execute. */ 1362 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1363 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1364 break; 1365 } 1366 1367 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEED_BUFFER); 1368 TAILQ_INSERT_TAIL(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1369 break; 1370 case RDMA_REQUEST_STATE_NEED_BUFFER: 1371 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1372 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1373 1374 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1375 1376 if (rdma_req != TAILQ_FIRST(&rqpair->ch->pending_data_buf_queue)) { 1377 /* This request needs to wait in line to obtain a buffer */ 1378 break; 1379 } 1380 1381 /* Try to get a data buffer */ 1382 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1383 if (rc < 0) { 1384 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1385 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1386 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1387 break; 1388 } 1389 1390 if (!rdma_req->req.data) { 1391 /* No buffers available. */ 1392 break; 1393 } 1394 1395 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1396 1397 /* If data is transferring from host to controller and the data didn't 1398 * arrive using in capsule data, we need to do a transfer from the host. 1399 */ 1400 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1401 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1402 break; 1403 } 1404 1405 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1406 break; 1407 case RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING: 1408 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 0, 0, 1409 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1410 1411 if (rdma_req != TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING])) { 1412 /* This request needs to wait in line to perform RDMA */ 1413 break; 1414 } 1415 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1416 1417 if (cur_rdma_rw_depth >= rqpair->max_rw_depth) { 1418 /* R/W queue is full, need to wait */ 1419 break; 1420 } 1421 1422 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1423 rc = request_transfer_in(&rdma_req->req); 1424 if (!rc) { 1425 spdk_nvmf_rdma_request_set_state(rdma_req, 1426 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1427 } else { 1428 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1429 spdk_nvmf_rdma_request_set_state(rdma_req, 1430 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1431 } 1432 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1433 /* The data transfer will be kicked off from 1434 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1435 */ 1436 spdk_nvmf_rdma_request_set_state(rdma_req, 1437 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1438 } else { 1439 SPDK_ERRLOG("Cannot perform data transfer, unknown state: %u\n", 1440 rdma_req->req.xfer); 1441 assert(0); 1442 } 1443 break; 1444 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1445 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1446 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1447 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1448 * to escape this state. */ 1449 break; 1450 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1451 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1452 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1453 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTING); 1454 spdk_nvmf_request_exec(&rdma_req->req); 1455 break; 1456 case RDMA_REQUEST_STATE_EXECUTING: 1457 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1458 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1459 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1460 * to escape this state. */ 1461 break; 1462 case RDMA_REQUEST_STATE_EXECUTED: 1463 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1464 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1465 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1466 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1467 } else { 1468 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1469 } 1470 break; 1471 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1472 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 1473 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1474 rc = request_transfer_out(&rdma_req->req, &data_posted); 1475 assert(rc == 0); /* No good way to handle this currently */ 1476 spdk_nvmf_rdma_request_set_state(rdma_req, 1477 data_posted ? 1478 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1479 RDMA_REQUEST_STATE_COMPLETING); 1480 break; 1481 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1482 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 1483 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 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_COMPLETING: 1488 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 1489 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1490 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1491 * to escape this state. */ 1492 break; 1493 case RDMA_REQUEST_STATE_COMPLETED: 1494 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 1495 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1496 1497 if (rdma_req->data_from_pool) { 1498 /* Put the buffer/s back in the pool */ 1499 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 1500 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 1501 rdma_req->req.iov[i].iov_base = NULL; 1502 rdma_req->data.buffers[i] = NULL; 1503 } 1504 rdma_req->data_from_pool = false; 1505 } 1506 rdma_req->req.length = 0; 1507 rdma_req->req.iovcnt = 0; 1508 rdma_req->req.data = NULL; 1509 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 1510 break; 1511 case RDMA_REQUEST_NUM_STATES: 1512 default: 1513 assert(0); 1514 break; 1515 } 1516 1517 if (rdma_req->state != prev_state) { 1518 progress = true; 1519 } 1520 } while (rdma_req->state != prev_state); 1521 1522 return progress; 1523 } 1524 1525 /* Public API callbacks begin here */ 1526 1527 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 1528 1529 static struct spdk_nvmf_transport * 1530 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 1531 { 1532 int rc; 1533 struct spdk_nvmf_rdma_transport *rtransport; 1534 struct spdk_nvmf_rdma_device *device, *tmp; 1535 struct ibv_context **contexts; 1536 uint32_t i; 1537 int flag; 1538 uint32_t sge_count; 1539 1540 rtransport = calloc(1, sizeof(*rtransport)); 1541 if (!rtransport) { 1542 return NULL; 1543 } 1544 1545 if (pthread_mutex_init(&rtransport->lock, NULL)) { 1546 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 1547 free(rtransport); 1548 return NULL; 1549 } 1550 1551 spdk_io_device_register(rtransport, spdk_nvmf_rdma_mgmt_channel_create, 1552 spdk_nvmf_rdma_mgmt_channel_destroy, 1553 sizeof(struct spdk_nvmf_rdma_mgmt_channel), 1554 "rdma_transport"); 1555 1556 TAILQ_INIT(&rtransport->devices); 1557 TAILQ_INIT(&rtransport->ports); 1558 1559 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 1560 1561 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 1562 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 1563 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 1564 " in_capsule_data_size=%d, max_aq_depth=%d\n", 1565 opts->max_queue_depth, 1566 opts->max_io_size, 1567 opts->max_qpairs_per_ctrlr, 1568 opts->io_unit_size, 1569 opts->in_capsule_data_size, 1570 opts->max_aq_depth); 1571 1572 /* I/O unit size cannot be larger than max I/O size */ 1573 if (opts->io_unit_size > opts->max_io_size) { 1574 opts->io_unit_size = opts->max_io_size; 1575 } 1576 1577 sge_count = opts->max_io_size / opts->io_unit_size; 1578 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 1579 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 1580 spdk_nvmf_rdma_destroy(&rtransport->transport); 1581 return NULL; 1582 } 1583 1584 rtransport->event_channel = rdma_create_event_channel(); 1585 if (rtransport->event_channel == NULL) { 1586 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 1587 spdk_nvmf_rdma_destroy(&rtransport->transport); 1588 return NULL; 1589 } 1590 1591 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 1592 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 1593 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 1594 rtransport->event_channel->fd, spdk_strerror(errno)); 1595 spdk_nvmf_rdma_destroy(&rtransport->transport); 1596 return NULL; 1597 } 1598 1599 rtransport->data_buf_pool = spdk_mempool_create("spdk_nvmf_rdma", 1600 opts->max_queue_depth * 4, /* The 4 is arbitrarily chosen. Needs to be configurable. */ 1601 opts->max_io_size + NVMF_DATA_BUFFER_ALIGNMENT, 1602 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 1603 SPDK_ENV_SOCKET_ID_ANY); 1604 if (!rtransport->data_buf_pool) { 1605 SPDK_ERRLOG("Unable to allocate buffer pool for poll group\n"); 1606 spdk_nvmf_rdma_destroy(&rtransport->transport); 1607 return NULL; 1608 } 1609 1610 contexts = rdma_get_devices(NULL); 1611 if (contexts == NULL) { 1612 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 1613 spdk_nvmf_rdma_destroy(&rtransport->transport); 1614 return NULL; 1615 } 1616 1617 i = 0; 1618 rc = 0; 1619 while (contexts[i] != NULL) { 1620 device = calloc(1, sizeof(*device)); 1621 if (!device) { 1622 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 1623 rc = -ENOMEM; 1624 break; 1625 } 1626 device->context = contexts[i]; 1627 rc = ibv_query_device(device->context, &device->attr); 1628 if (rc < 0) { 1629 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1630 free(device); 1631 break; 1632 1633 } 1634 /* set up device context async ev fd as NON_BLOCKING */ 1635 flag = fcntl(device->context->async_fd, F_GETFL); 1636 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 1637 if (rc < 0) { 1638 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 1639 free(device); 1640 break; 1641 } 1642 1643 device->pd = ibv_alloc_pd(device->context); 1644 if (!device->pd) { 1645 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 1646 free(device); 1647 rc = -1; 1648 break; 1649 } 1650 1651 device->map = spdk_mem_map_alloc(0, spdk_nvmf_rdma_mem_notify, device); 1652 if (!device->map) { 1653 SPDK_ERRLOG("Unable to allocate memory map for new poll group\n"); 1654 ibv_dealloc_pd(device->pd); 1655 free(device); 1656 rc = -1; 1657 break; 1658 } 1659 1660 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 1661 i++; 1662 } 1663 rdma_free_devices(contexts); 1664 1665 if (rc < 0) { 1666 spdk_nvmf_rdma_destroy(&rtransport->transport); 1667 return NULL; 1668 } 1669 1670 /* Set up poll descriptor array to monitor events from RDMA and IB 1671 * in a single poll syscall 1672 */ 1673 rtransport->npoll_fds = i + 1; 1674 i = 0; 1675 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 1676 if (rtransport->poll_fds == NULL) { 1677 SPDK_ERRLOG("poll_fds allocation failed\n"); 1678 spdk_nvmf_rdma_destroy(&rtransport->transport); 1679 return NULL; 1680 } 1681 1682 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 1683 rtransport->poll_fds[i++].events = POLLIN; 1684 1685 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 1686 rtransport->poll_fds[i].fd = device->context->async_fd; 1687 rtransport->poll_fds[i++].events = POLLIN; 1688 } 1689 1690 return &rtransport->transport; 1691 } 1692 1693 static int 1694 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 1695 { 1696 struct spdk_nvmf_rdma_transport *rtransport; 1697 struct spdk_nvmf_rdma_port *port, *port_tmp; 1698 struct spdk_nvmf_rdma_device *device, *device_tmp; 1699 1700 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1701 1702 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 1703 TAILQ_REMOVE(&rtransport->ports, port, link); 1704 rdma_destroy_id(port->id); 1705 free(port); 1706 } 1707 1708 if (rtransport->poll_fds != NULL) { 1709 free(rtransport->poll_fds); 1710 } 1711 1712 if (rtransport->event_channel != NULL) { 1713 rdma_destroy_event_channel(rtransport->event_channel); 1714 } 1715 1716 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 1717 TAILQ_REMOVE(&rtransport->devices, device, link); 1718 if (device->map) { 1719 spdk_mem_map_free(&device->map); 1720 } 1721 if (device->pd) { 1722 ibv_dealloc_pd(device->pd); 1723 } 1724 free(device); 1725 } 1726 1727 if (rtransport->data_buf_pool != NULL) { 1728 if (spdk_mempool_count(rtransport->data_buf_pool) != 1729 (transport->opts.max_queue_depth * 4)) { 1730 SPDK_ERRLOG("transport buffer pool count is %zu but should be %u\n", 1731 spdk_mempool_count(rtransport->data_buf_pool), 1732 transport->opts.max_queue_depth * 4); 1733 } 1734 } 1735 1736 spdk_mempool_free(rtransport->data_buf_pool); 1737 spdk_io_device_unregister(rtransport, NULL); 1738 pthread_mutex_destroy(&rtransport->lock); 1739 free(rtransport); 1740 1741 return 0; 1742 } 1743 1744 static int 1745 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 1746 const struct spdk_nvme_transport_id *trid) 1747 { 1748 struct spdk_nvmf_rdma_transport *rtransport; 1749 struct spdk_nvmf_rdma_device *device; 1750 struct spdk_nvmf_rdma_port *port_tmp, *port; 1751 struct addrinfo *res; 1752 struct addrinfo hints; 1753 int family; 1754 int rc; 1755 1756 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1757 1758 port = calloc(1, sizeof(*port)); 1759 if (!port) { 1760 return -ENOMEM; 1761 } 1762 1763 /* Selectively copy the trid. Things like NQN don't matter here - that 1764 * mapping is enforced elsewhere. 1765 */ 1766 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1767 port->trid.adrfam = trid->adrfam; 1768 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 1769 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 1770 1771 pthread_mutex_lock(&rtransport->lock); 1772 assert(rtransport->event_channel != NULL); 1773 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 1774 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 1775 port_tmp->ref++; 1776 free(port); 1777 /* Already listening at this address */ 1778 pthread_mutex_unlock(&rtransport->lock); 1779 return 0; 1780 } 1781 } 1782 1783 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 1784 if (rc < 0) { 1785 SPDK_ERRLOG("rdma_create_id() failed\n"); 1786 free(port); 1787 pthread_mutex_unlock(&rtransport->lock); 1788 return rc; 1789 } 1790 1791 switch (port->trid.adrfam) { 1792 case SPDK_NVMF_ADRFAM_IPV4: 1793 family = AF_INET; 1794 break; 1795 case SPDK_NVMF_ADRFAM_IPV6: 1796 family = AF_INET6; 1797 break; 1798 default: 1799 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 1800 free(port); 1801 pthread_mutex_unlock(&rtransport->lock); 1802 return -EINVAL; 1803 } 1804 1805 memset(&hints, 0, sizeof(hints)); 1806 hints.ai_family = family; 1807 hints.ai_socktype = SOCK_STREAM; 1808 hints.ai_protocol = 0; 1809 1810 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 1811 if (rc) { 1812 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 1813 free(port); 1814 pthread_mutex_unlock(&rtransport->lock); 1815 return -EINVAL; 1816 } 1817 1818 rc = rdma_bind_addr(port->id, res->ai_addr); 1819 freeaddrinfo(res); 1820 1821 if (rc < 0) { 1822 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 1823 rdma_destroy_id(port->id); 1824 free(port); 1825 pthread_mutex_unlock(&rtransport->lock); 1826 return rc; 1827 } 1828 1829 if (!port->id->verbs) { 1830 SPDK_ERRLOG("ibv_context is null\n"); 1831 rdma_destroy_id(port->id); 1832 free(port); 1833 pthread_mutex_unlock(&rtransport->lock); 1834 return -1; 1835 } 1836 1837 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 1838 if (rc < 0) { 1839 SPDK_ERRLOG("rdma_listen() failed\n"); 1840 rdma_destroy_id(port->id); 1841 free(port); 1842 pthread_mutex_unlock(&rtransport->lock); 1843 return rc; 1844 } 1845 1846 TAILQ_FOREACH(device, &rtransport->devices, link) { 1847 if (device->context == port->id->verbs) { 1848 port->device = device; 1849 break; 1850 } 1851 } 1852 if (!port->device) { 1853 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 1854 port->id->verbs); 1855 rdma_destroy_id(port->id); 1856 free(port); 1857 pthread_mutex_unlock(&rtransport->lock); 1858 return -EINVAL; 1859 } 1860 1861 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 1862 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 1863 1864 port->ref = 1; 1865 1866 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 1867 pthread_mutex_unlock(&rtransport->lock); 1868 1869 return 0; 1870 } 1871 1872 static int 1873 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 1874 const struct spdk_nvme_transport_id *_trid) 1875 { 1876 struct spdk_nvmf_rdma_transport *rtransport; 1877 struct spdk_nvmf_rdma_port *port, *tmp; 1878 struct spdk_nvme_transport_id trid = {}; 1879 1880 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1881 1882 /* Selectively copy the trid. Things like NQN don't matter here - that 1883 * mapping is enforced elsewhere. 1884 */ 1885 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1886 trid.adrfam = _trid->adrfam; 1887 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 1888 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 1889 1890 pthread_mutex_lock(&rtransport->lock); 1891 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 1892 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 1893 assert(port->ref > 0); 1894 port->ref--; 1895 if (port->ref == 0) { 1896 TAILQ_REMOVE(&rtransport->ports, port, link); 1897 rdma_destroy_id(port->id); 1898 free(port); 1899 } 1900 break; 1901 } 1902 } 1903 1904 pthread_mutex_unlock(&rtransport->lock); 1905 return 0; 1906 } 1907 1908 static bool 1909 spdk_nvmf_rdma_qpair_is_idle(struct spdk_nvmf_qpair *qpair) 1910 { 1911 int cur_queue_depth, cur_rdma_rw_depth; 1912 struct spdk_nvmf_rdma_qpair *rqpair; 1913 1914 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1915 cur_queue_depth = spdk_nvmf_rdma_cur_queue_depth(rqpair); 1916 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1917 1918 if (cur_queue_depth == 0 && cur_rdma_rw_depth == 0) { 1919 return true; 1920 } 1921 return false; 1922 } 1923 1924 static void 1925 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 1926 struct spdk_nvmf_rdma_qpair *rqpair) 1927 { 1928 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 1929 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 1930 1931 /* We process I/O in the data transfer pending queue at the highest priority. */ 1932 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING], 1933 state_link, req_tmp) { 1934 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1935 break; 1936 } 1937 } 1938 1939 /* The second highest priority is I/O waiting on memory buffers. */ 1940 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->ch->pending_data_buf_queue, link, 1941 req_tmp) { 1942 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1943 break; 1944 } 1945 } 1946 1947 if (rqpair->qpair_disconnected) { 1948 spdk_nvmf_rdma_qpair_destroy(rqpair); 1949 return; 1950 } 1951 1952 /* Do not process newly received commands if qp is in ERROR state, 1953 * wait till the recovery is complete. 1954 */ 1955 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 1956 return; 1957 } 1958 1959 /* The lowest priority is processing newly received commands */ 1960 TAILQ_FOREACH_SAFE(rdma_recv, &rqpair->incoming_queue, link, recv_tmp) { 1961 if (TAILQ_EMPTY(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE])) { 1962 break; 1963 } 1964 1965 rdma_req = TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE]); 1966 rdma_req->recv = rdma_recv; 1967 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEW); 1968 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1969 break; 1970 } 1971 } 1972 } 1973 1974 static void 1975 spdk_nvmf_rdma_drain_state_queue(struct spdk_nvmf_rdma_qpair *rqpair, 1976 enum spdk_nvmf_rdma_request_state state) 1977 { 1978 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 1979 struct spdk_nvmf_rdma_transport *rtransport; 1980 1981 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[state], state_link, req_tmp) { 1982 rtransport = SPDK_CONTAINEROF(rdma_req->req.qpair->transport, 1983 struct spdk_nvmf_rdma_transport, transport); 1984 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1985 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 1986 } 1987 } 1988 1989 static void 1990 spdk_nvmf_rdma_qpair_recover(struct spdk_nvmf_rdma_qpair *rqpair) 1991 { 1992 enum ibv_qp_state state, next_state; 1993 int recovered; 1994 struct spdk_nvmf_rdma_transport *rtransport; 1995 1996 if (!spdk_nvmf_rdma_qpair_is_idle(&rqpair->qpair)) { 1997 /* There must be outstanding requests down to media. 1998 * If so, wait till they're complete. 1999 */ 2000 assert(!TAILQ_EMPTY(&rqpair->qpair.outstanding)); 2001 return; 2002 } 2003 2004 state = rqpair->ibv_attr.qp_state; 2005 next_state = state; 2006 2007 SPDK_NOTICELOG("RDMA qpair %u is in state: %s\n", 2008 rqpair->qpair.qid, 2009 str_ibv_qp_state[state]); 2010 2011 if (!(state == IBV_QPS_ERR || state == IBV_QPS_RESET)) { 2012 SPDK_ERRLOG("Can't recover RDMA qpair %u from the state: %s\n", 2013 rqpair->qpair.qid, 2014 str_ibv_qp_state[state]); 2015 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2016 return; 2017 } 2018 2019 recovered = 0; 2020 while (!recovered) { 2021 switch (state) { 2022 case IBV_QPS_ERR: 2023 next_state = IBV_QPS_RESET; 2024 break; 2025 case IBV_QPS_RESET: 2026 next_state = IBV_QPS_INIT; 2027 break; 2028 case IBV_QPS_INIT: 2029 next_state = IBV_QPS_RTR; 2030 break; 2031 case IBV_QPS_RTR: 2032 next_state = IBV_QPS_RTS; 2033 break; 2034 case IBV_QPS_RTS: 2035 recovered = 1; 2036 break; 2037 default: 2038 SPDK_ERRLOG("RDMA qpair %u unexpected state for recovery: %u\n", 2039 rqpair->qpair.qid, state); 2040 goto error; 2041 } 2042 /* Do not transition into same state */ 2043 if (next_state == state) { 2044 break; 2045 } 2046 2047 if (spdk_nvmf_rdma_set_ibv_state(rqpair, next_state)) { 2048 goto error; 2049 } 2050 2051 state = next_state; 2052 } 2053 2054 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2055 struct spdk_nvmf_rdma_transport, 2056 transport); 2057 2058 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2059 2060 return; 2061 error: 2062 SPDK_NOTICELOG("RDMA qpair %u: recovery failed, disconnecting...\n", 2063 rqpair->qpair.qid); 2064 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2065 } 2066 2067 /* Clean up only the states that can be aborted at any time */ 2068 static void 2069 _spdk_nvmf_rdma_qp_cleanup_safe_states(struct spdk_nvmf_rdma_qpair *rqpair) 2070 { 2071 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2072 2073 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_NEW); 2074 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[RDMA_REQUEST_STATE_NEED_BUFFER], link, req_tmp) { 2075 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 2076 } 2077 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_NEED_BUFFER); 2078 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 2079 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 2080 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_EXECUTED); 2081 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 2082 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_COMPLETED); 2083 } 2084 2085 /* This cleans up all memory. It is only safe to use if the rest of the software stack 2086 * has been shut down */ 2087 static void 2088 _spdk_nvmf_rdma_qp_cleanup_all_states(struct spdk_nvmf_rdma_qpair *rqpair) 2089 { 2090 _spdk_nvmf_rdma_qp_cleanup_safe_states(rqpair); 2091 2092 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_EXECUTING); 2093 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2094 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2095 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_COMPLETING); 2096 } 2097 2098 static void 2099 _spdk_nvmf_rdma_qp_error(void *arg) 2100 { 2101 struct spdk_nvmf_rdma_qpair *rqpair = arg; 2102 enum ibv_qp_state state; 2103 2104 state = rqpair->ibv_attr.qp_state; 2105 if (state != IBV_QPS_ERR) { 2106 /* Error was already recovered */ 2107 return; 2108 } 2109 2110 if (spdk_nvmf_qpair_is_admin_queue(&rqpair->qpair)) { 2111 spdk_nvmf_ctrlr_abort_aer(rqpair->qpair.ctrlr); 2112 } 2113 2114 _spdk_nvmf_rdma_qp_cleanup_safe_states(rqpair); 2115 2116 /* Attempt recovery. This will exit without recovering if I/O requests 2117 * are still outstanding */ 2118 spdk_nvmf_rdma_qpair_recover(rqpair); 2119 } 2120 2121 static void 2122 _spdk_nvmf_rdma_qp_last_wqe(void *arg) 2123 { 2124 struct spdk_nvmf_rdma_qpair *rqpair = arg; 2125 enum ibv_qp_state state; 2126 2127 state = rqpair->ibv_attr.qp_state; 2128 if (state != IBV_QPS_ERR) { 2129 /* Error was already recovered */ 2130 return; 2131 } 2132 2133 /* Clear out the states that are safe to clear any time, plus the 2134 * RDMA data transfer states. */ 2135 _spdk_nvmf_rdma_qp_cleanup_safe_states(rqpair); 2136 2137 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2138 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2139 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_COMPLETING); 2140 2141 spdk_nvmf_rdma_qpair_recover(rqpair); 2142 } 2143 2144 static void 2145 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2146 { 2147 int rc; 2148 struct spdk_nvmf_rdma_qpair *rqpair; 2149 struct ibv_async_event event; 2150 enum ibv_qp_state state; 2151 2152 rc = ibv_get_async_event(device->context, &event); 2153 2154 if (rc) { 2155 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2156 errno, spdk_strerror(errno)); 2157 return; 2158 } 2159 2160 SPDK_NOTICELOG("Async event: %s\n", 2161 ibv_event_type_str(event.event_type)); 2162 2163 rqpair = event.element.qp->qp_context; 2164 2165 switch (event.event_type) { 2166 case IBV_EVENT_QP_FATAL: 2167 spdk_nvmf_rdma_update_ibv_state(rqpair); 2168 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_error, rqpair); 2169 break; 2170 case IBV_EVENT_QP_LAST_WQE_REACHED: 2171 spdk_nvmf_rdma_update_ibv_state(rqpair); 2172 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_last_wqe, rqpair); 2173 break; 2174 case IBV_EVENT_SQ_DRAINED: 2175 /* This event occurs frequently in both error and non-error states. 2176 * Check if the qpair is in an error state before sending a message. 2177 * Note that we're not on the correct thread to access the qpair, but 2178 * the operations that the below calls make all happen to be thread 2179 * safe. */ 2180 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2181 if (state == IBV_QPS_ERR) { 2182 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_error, rqpair); 2183 } 2184 break; 2185 case IBV_EVENT_QP_REQ_ERR: 2186 case IBV_EVENT_QP_ACCESS_ERR: 2187 case IBV_EVENT_COMM_EST: 2188 case IBV_EVENT_PATH_MIG: 2189 case IBV_EVENT_PATH_MIG_ERR: 2190 spdk_nvmf_rdma_update_ibv_state(rqpair); 2191 break; 2192 case IBV_EVENT_CQ_ERR: 2193 case IBV_EVENT_DEVICE_FATAL: 2194 case IBV_EVENT_PORT_ACTIVE: 2195 case IBV_EVENT_PORT_ERR: 2196 case IBV_EVENT_LID_CHANGE: 2197 case IBV_EVENT_PKEY_CHANGE: 2198 case IBV_EVENT_SM_CHANGE: 2199 case IBV_EVENT_SRQ_ERR: 2200 case IBV_EVENT_SRQ_LIMIT_REACHED: 2201 case IBV_EVENT_CLIENT_REREGISTER: 2202 case IBV_EVENT_GID_CHANGE: 2203 default: 2204 break; 2205 } 2206 ibv_ack_async_event(&event); 2207 } 2208 2209 static void 2210 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2211 { 2212 int nfds, i = 0; 2213 struct spdk_nvmf_rdma_transport *rtransport; 2214 struct spdk_nvmf_rdma_device *device, *tmp; 2215 2216 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2217 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2218 2219 if (nfds <= 0) { 2220 return; 2221 } 2222 2223 /* The first poll descriptor is RDMA CM event */ 2224 if (rtransport->poll_fds[i++].revents & POLLIN) { 2225 spdk_nvmf_process_cm_event(transport, cb_fn); 2226 nfds--; 2227 } 2228 2229 if (nfds == 0) { 2230 return; 2231 } 2232 2233 /* Second and subsequent poll descriptors are IB async events */ 2234 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2235 if (rtransport->poll_fds[i++].revents & POLLIN) { 2236 spdk_nvmf_process_ib_event(device); 2237 nfds--; 2238 } 2239 } 2240 /* check all flagged fd's have been served */ 2241 assert(nfds == 0); 2242 } 2243 2244 static void 2245 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2246 struct spdk_nvme_transport_id *trid, 2247 struct spdk_nvmf_discovery_log_page_entry *entry) 2248 { 2249 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2250 entry->adrfam = trid->adrfam; 2251 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2252 2253 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2254 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2255 2256 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2257 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2258 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2259 } 2260 2261 static struct spdk_nvmf_transport_poll_group * 2262 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2263 { 2264 struct spdk_nvmf_rdma_transport *rtransport; 2265 struct spdk_nvmf_rdma_poll_group *rgroup; 2266 struct spdk_nvmf_rdma_poller *poller; 2267 struct spdk_nvmf_rdma_device *device; 2268 2269 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2270 2271 rgroup = calloc(1, sizeof(*rgroup)); 2272 if (!rgroup) { 2273 return NULL; 2274 } 2275 2276 TAILQ_INIT(&rgroup->pollers); 2277 2278 pthread_mutex_lock(&rtransport->lock); 2279 TAILQ_FOREACH(device, &rtransport->devices, link) { 2280 poller = calloc(1, sizeof(*poller)); 2281 if (!poller) { 2282 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2283 free(rgroup); 2284 pthread_mutex_unlock(&rtransport->lock); 2285 return NULL; 2286 } 2287 2288 poller->device = device; 2289 poller->group = rgroup; 2290 2291 TAILQ_INIT(&poller->qpairs); 2292 2293 poller->cq = ibv_create_cq(device->context, NVMF_RDMA_CQ_SIZE, poller, NULL, 0); 2294 if (!poller->cq) { 2295 SPDK_ERRLOG("Unable to create completion queue\n"); 2296 free(poller); 2297 free(rgroup); 2298 pthread_mutex_unlock(&rtransport->lock); 2299 return NULL; 2300 } 2301 2302 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2303 } 2304 2305 pthread_mutex_unlock(&rtransport->lock); 2306 return &rgroup->group; 2307 } 2308 2309 static void 2310 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2311 { 2312 struct spdk_nvmf_rdma_poll_group *rgroup; 2313 struct spdk_nvmf_rdma_poller *poller, *tmp; 2314 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2315 2316 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2317 2318 if (!rgroup) { 2319 return; 2320 } 2321 2322 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2323 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2324 2325 if (poller->cq) { 2326 ibv_destroy_cq(poller->cq); 2327 } 2328 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 2329 _spdk_nvmf_rdma_qp_cleanup_all_states(qpair); 2330 spdk_nvmf_rdma_qpair_destroy(qpair); 2331 } 2332 2333 free(poller); 2334 } 2335 2336 free(rgroup); 2337 } 2338 2339 static int 2340 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 2341 struct spdk_nvmf_qpair *qpair) 2342 { 2343 struct spdk_nvmf_rdma_transport *rtransport; 2344 struct spdk_nvmf_rdma_poll_group *rgroup; 2345 struct spdk_nvmf_rdma_qpair *rqpair; 2346 struct spdk_nvmf_rdma_device *device; 2347 struct spdk_nvmf_rdma_poller *poller; 2348 int rc; 2349 2350 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 2351 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2352 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2353 2354 device = rqpair->port->device; 2355 2356 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 2357 if (poller->device == device) { 2358 break; 2359 } 2360 } 2361 2362 if (!poller) { 2363 SPDK_ERRLOG("No poller found for device.\n"); 2364 return -1; 2365 } 2366 2367 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 2368 rqpair->poller = poller; 2369 2370 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 2371 if (rc < 0) { 2372 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 2373 return -1; 2374 } 2375 2376 rqpair->mgmt_channel = spdk_get_io_channel(rtransport); 2377 if (!rqpair->mgmt_channel) { 2378 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2379 spdk_nvmf_rdma_qpair_destroy(rqpair); 2380 return -1; 2381 } 2382 2383 rqpair->ch = spdk_io_channel_get_ctx(rqpair->mgmt_channel); 2384 assert(rqpair->ch != NULL); 2385 2386 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 2387 if (rc) { 2388 /* Try to reject, but we probably can't */ 2389 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2390 spdk_nvmf_rdma_qpair_destroy(rqpair); 2391 return -1; 2392 } 2393 2394 spdk_nvmf_rdma_update_ibv_state(rqpair); 2395 2396 return 0; 2397 } 2398 2399 static int 2400 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 2401 { 2402 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2403 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2404 struct spdk_nvmf_rdma_transport, transport); 2405 2406 if (rdma_req->data_from_pool) { 2407 /* Put the buffer/s back in the pool */ 2408 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 2409 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 2410 rdma_req->req.iov[i].iov_base = NULL; 2411 rdma_req->data.buffers[i] = NULL; 2412 } 2413 rdma_req->data_from_pool = false; 2414 } 2415 rdma_req->req.length = 0; 2416 rdma_req->req.iovcnt = 0; 2417 rdma_req->req.data = NULL; 2418 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 2419 return 0; 2420 } 2421 2422 static int 2423 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 2424 { 2425 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2426 struct spdk_nvmf_rdma_transport, transport); 2427 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 2428 struct spdk_nvmf_rdma_request, req); 2429 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 2430 struct spdk_nvmf_rdma_qpair, qpair); 2431 2432 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2433 /* The connection is alive, so process the request as normal */ 2434 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTED); 2435 } else { 2436 /* The connection is dead. Move the request directly to the completed state. */ 2437 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2438 } 2439 2440 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2441 2442 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE && rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 2443 /* If the NVMe-oF layer thinks the connection is active, but the RDMA layer thinks 2444 * the connection is dead, perform error recovery. */ 2445 spdk_nvmf_rdma_qpair_recover(rqpair); 2446 } 2447 2448 return 0; 2449 } 2450 2451 static void 2452 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 2453 { 2454 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2455 2456 spdk_nvmf_rdma_qpair_destroy(rqpair); 2457 } 2458 2459 static struct spdk_nvmf_rdma_request * 2460 get_rdma_req_from_wc(struct ibv_wc *wc) 2461 { 2462 struct spdk_nvmf_rdma_request *rdma_req; 2463 2464 rdma_req = (struct spdk_nvmf_rdma_request *)wc->wr_id; 2465 assert(rdma_req != NULL); 2466 2467 #ifdef DEBUG 2468 struct spdk_nvmf_rdma_qpair *rqpair; 2469 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2470 2471 assert(rdma_req - rqpair->reqs >= 0); 2472 assert(rdma_req - rqpair->reqs < (ptrdiff_t)rqpair->max_queue_depth); 2473 #endif 2474 2475 return rdma_req; 2476 } 2477 2478 static struct spdk_nvmf_rdma_recv * 2479 get_rdma_recv_from_wc(struct ibv_wc *wc) 2480 { 2481 struct spdk_nvmf_rdma_recv *rdma_recv; 2482 2483 assert(wc->byte_len >= sizeof(struct spdk_nvmf_capsule_cmd)); 2484 2485 rdma_recv = (struct spdk_nvmf_rdma_recv *)wc->wr_id; 2486 assert(rdma_recv != NULL); 2487 2488 #ifdef DEBUG 2489 struct spdk_nvmf_rdma_qpair *rqpair = rdma_recv->qpair; 2490 2491 assert(rdma_recv - rqpair->recvs >= 0); 2492 assert(rdma_recv - rqpair->recvs < (ptrdiff_t)rqpair->max_queue_depth); 2493 #endif 2494 2495 return rdma_recv; 2496 } 2497 2498 #ifdef DEBUG 2499 static int 2500 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 2501 { 2502 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 2503 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 2504 } 2505 #endif 2506 2507 static int 2508 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 2509 struct spdk_nvmf_rdma_poller *rpoller) 2510 { 2511 struct ibv_wc wc[32]; 2512 struct spdk_nvmf_rdma_request *rdma_req; 2513 struct spdk_nvmf_rdma_recv *rdma_recv; 2514 struct spdk_nvmf_rdma_qpair *rqpair; 2515 int reaped, i; 2516 int count = 0; 2517 bool error = false; 2518 2519 /* Poll for completing operations. */ 2520 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 2521 if (reaped < 0) { 2522 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2523 errno, spdk_strerror(errno)); 2524 return -1; 2525 } 2526 2527 for (i = 0; i < reaped; i++) { 2528 /* Handle error conditions */ 2529 if (wc[i].status) { 2530 SPDK_WARNLOG("CQ error on CQ %p, Request 0x%lu (%d): %s\n", 2531 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 2532 error = true; 2533 2534 switch (wc[i].opcode) { 2535 case IBV_WC_SEND: 2536 case IBV_WC_RDMA_WRITE: 2537 case IBV_WC_RDMA_READ: 2538 rdma_req = get_rdma_req_from_wc(&wc[i]); 2539 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2540 2541 /* We're going to kill the connection, so force the request into 2542 * the completed state. */ 2543 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2544 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2545 break; 2546 case IBV_WC_RECV: 2547 rdma_recv = get_rdma_recv_from_wc(&wc[i]); 2548 rqpair = rdma_recv->qpair; 2549 2550 /* Dump this into the incoming queue. This gets cleaned up when 2551 * the queue pair disconnects. */ 2552 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2553 default: 2554 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2555 continue; 2556 } 2557 2558 /* Begin disconnecting the qpair. This is ok to call multiple times if lots of 2559 * errors occur on the same qpair in the same ibv_poll_cq batch. */ 2560 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2561 2562 continue; 2563 } 2564 2565 switch (wc[i].opcode) { 2566 case IBV_WC_SEND: 2567 rdma_req = get_rdma_req_from_wc(&wc[i]); 2568 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2569 2570 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 2571 2572 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2573 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2574 2575 count++; 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_WRITE: 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 /* Try to process other queued requests */ 2586 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2587 break; 2588 2589 case IBV_WC_RDMA_READ: 2590 rdma_req = get_rdma_req_from_wc(&wc[i]); 2591 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2592 2593 assert(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2594 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 2595 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2596 2597 /* Try to process other queued requests */ 2598 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2599 break; 2600 2601 case IBV_WC_RECV: 2602 rdma_recv = get_rdma_recv_from_wc(&wc[i]); 2603 rqpair = rdma_recv->qpair; 2604 2605 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2606 /* Try to process other queued requests */ 2607 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2608 break; 2609 2610 default: 2611 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2612 continue; 2613 } 2614 } 2615 2616 if (error == true) { 2617 return -1; 2618 } 2619 2620 return count; 2621 } 2622 2623 static int 2624 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 2625 { 2626 struct spdk_nvmf_rdma_transport *rtransport; 2627 struct spdk_nvmf_rdma_poll_group *rgroup; 2628 struct spdk_nvmf_rdma_poller *rpoller; 2629 int count, rc; 2630 2631 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 2632 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2633 2634 count = 0; 2635 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2636 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 2637 if (rc < 0) { 2638 return rc; 2639 } 2640 count += rc; 2641 } 2642 2643 return count; 2644 } 2645 2646 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 2647 .type = SPDK_NVME_TRANSPORT_RDMA, 2648 .create = spdk_nvmf_rdma_create, 2649 .destroy = spdk_nvmf_rdma_destroy, 2650 2651 .listen = spdk_nvmf_rdma_listen, 2652 .stop_listen = spdk_nvmf_rdma_stop_listen, 2653 .accept = spdk_nvmf_rdma_accept, 2654 2655 .listener_discover = spdk_nvmf_rdma_discover, 2656 2657 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 2658 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 2659 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 2660 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 2661 2662 .req_free = spdk_nvmf_rdma_request_free, 2663 .req_complete = spdk_nvmf_rdma_request_complete, 2664 2665 .qpair_fini = spdk_nvmf_rdma_close_qpair, 2666 .qpair_is_idle = spdk_nvmf_rdma_qpair_is_idle, 2667 2668 }; 2669 2670 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 2671