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 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 130 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 131 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 132 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 133 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 134 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 135 136 SPDK_TRACE_REGISTER_FN(nvmf_trace) 137 { 138 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 139 spdk_trace_register_description("RDMA_REQ_NEW", "", 140 TRACE_RDMA_REQUEST_STATE_NEW, 141 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 142 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", "", 143 TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 144 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 145 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H_TO_C", "", 146 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 147 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 148 spdk_trace_register_description("RDMA_REQ_TX_H_TO_C", "", 149 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 150 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 151 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", "", 152 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 153 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 154 spdk_trace_register_description("RDMA_REQ_EXECUTING", "", 155 TRACE_RDMA_REQUEST_STATE_EXECUTING, 156 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 157 spdk_trace_register_description("RDMA_REQ_EXECUTED", "", 158 TRACE_RDMA_REQUEST_STATE_EXECUTED, 159 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 160 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPLETE", "", 161 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 162 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 163 spdk_trace_register_description("RDMA_REQ_COMPLETING_CONTROLLER_TO_HOST", "", 164 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 165 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 166 spdk_trace_register_description("RDMA_REQ_COMPLETING_INCAPSULE", "", 167 TRACE_RDMA_REQUEST_STATE_COMPLETING, 168 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 169 spdk_trace_register_description("RDMA_REQ_COMPLETED", "", 170 TRACE_RDMA_REQUEST_STATE_COMPLETED, 171 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 172 173 spdk_trace_register_description("RDMA_QP_CREATE", "", TRACE_RDMA_QP_CREATE, 174 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 175 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", "", TRACE_RDMA_IBV_ASYNC_EVENT, 176 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 177 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", "", TRACE_RDMA_CM_ASYNC_EVENT, 178 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 179 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", "", TRACE_RDMA_QP_STATE_CHANGE, 180 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 181 spdk_trace_register_description("RDMA_QP_DISCONNECT", "", TRACE_RDMA_QP_DISCONNECT, 182 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 183 spdk_trace_register_description("RDMA_QP_DESTROY", "", TRACE_RDMA_QP_DESTROY, 184 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 185 } 186 187 /* This structure holds commands as they are received off the wire. 188 * It must be dynamically paired with a full request object 189 * (spdk_nvmf_rdma_request) to service a request. It is separate 190 * from the request because RDMA does not appear to order 191 * completions, so occasionally we'll get a new incoming 192 * command when there aren't any free request objects. 193 */ 194 struct spdk_nvmf_rdma_recv { 195 struct ibv_recv_wr wr; 196 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 197 198 struct spdk_nvmf_rdma_qpair *qpair; 199 200 /* In-capsule data buffer */ 201 uint8_t *buf; 202 203 TAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 204 }; 205 206 struct spdk_nvmf_rdma_request { 207 struct spdk_nvmf_request req; 208 bool data_from_pool; 209 210 enum spdk_nvmf_rdma_request_state state; 211 212 struct spdk_nvmf_rdma_recv *recv; 213 214 struct { 215 struct ibv_send_wr wr; 216 struct ibv_sge sgl[NVMF_DEFAULT_TX_SGE]; 217 } rsp; 218 219 struct { 220 struct ibv_send_wr wr; 221 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 222 void *buffers[SPDK_NVMF_MAX_SGL_ENTRIES]; 223 } data; 224 225 TAILQ_ENTRY(spdk_nvmf_rdma_request) link; 226 TAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 227 }; 228 229 struct spdk_nvmf_rdma_qpair { 230 struct spdk_nvmf_qpair qpair; 231 232 struct spdk_nvmf_rdma_port *port; 233 struct spdk_nvmf_rdma_poller *poller; 234 235 struct rdma_cm_id *cm_id; 236 struct rdma_cm_id *listen_id; 237 238 /* The maximum number of I/O outstanding on this connection at one time */ 239 uint16_t max_queue_depth; 240 241 /* The maximum number of active RDMA READ and WRITE operations at one time */ 242 uint16_t max_rw_depth; 243 244 /* Receives that are waiting for a request object */ 245 TAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 246 247 /* Queues to track the requests in all states */ 248 TAILQ_HEAD(, spdk_nvmf_rdma_request) state_queue[RDMA_REQUEST_NUM_STATES]; 249 250 /* Number of requests in each state */ 251 uint32_t state_cntr[RDMA_REQUEST_NUM_STATES]; 252 253 int max_sge; 254 255 /* Array of size "max_queue_depth" containing RDMA requests. */ 256 struct spdk_nvmf_rdma_request *reqs; 257 258 /* Array of size "max_queue_depth" containing RDMA recvs. */ 259 struct spdk_nvmf_rdma_recv *recvs; 260 261 /* Array of size "max_queue_depth" containing 64 byte capsules 262 * used for receive. 263 */ 264 union nvmf_h2c_msg *cmds; 265 struct ibv_mr *cmds_mr; 266 267 /* Array of size "max_queue_depth" containing 16 byte completions 268 * to be sent back to the user. 269 */ 270 union nvmf_c2h_msg *cpls; 271 struct ibv_mr *cpls_mr; 272 273 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 274 * buffers to be used for in capsule data. 275 */ 276 void *bufs; 277 struct ibv_mr *bufs_mr; 278 279 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 280 281 /* Mgmt channel */ 282 struct spdk_io_channel *mgmt_channel; 283 struct spdk_nvmf_rdma_mgmt_channel *ch; 284 285 /* IBV queue pair attributes: they are used to manage 286 * qp state and recover from errors. 287 */ 288 struct ibv_qp_init_attr ibv_init_attr; 289 struct ibv_qp_attr ibv_attr; 290 291 bool qpair_disconnected; 292 293 /* Reference counter for how many unprocessed messages 294 * from other threads are currently outstanding. The 295 * qpair cannot be destroyed until this is 0. This is 296 * atomically incremented from any thread, but only 297 * decremented and read from the thread that owns this 298 * qpair. 299 */ 300 uint32_t refcnt; 301 }; 302 303 struct spdk_nvmf_rdma_poller { 304 struct spdk_nvmf_rdma_device *device; 305 struct spdk_nvmf_rdma_poll_group *group; 306 307 struct ibv_cq *cq; 308 309 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 310 311 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 312 }; 313 314 struct spdk_nvmf_rdma_poll_group { 315 struct spdk_nvmf_transport_poll_group group; 316 317 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 318 }; 319 320 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 321 struct spdk_nvmf_rdma_device { 322 struct ibv_device_attr attr; 323 struct ibv_context *context; 324 325 struct spdk_mem_map *map; 326 struct ibv_pd *pd; 327 328 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 329 }; 330 331 struct spdk_nvmf_rdma_port { 332 struct spdk_nvme_transport_id trid; 333 struct rdma_cm_id *id; 334 struct spdk_nvmf_rdma_device *device; 335 uint32_t ref; 336 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 337 }; 338 339 struct spdk_nvmf_rdma_transport { 340 struct spdk_nvmf_transport transport; 341 342 struct rdma_event_channel *event_channel; 343 344 struct spdk_mempool *data_buf_pool; 345 346 pthread_mutex_t lock; 347 348 /* fields used to poll RDMA/IB events */ 349 nfds_t npoll_fds; 350 struct pollfd *poll_fds; 351 352 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 353 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 354 }; 355 356 struct spdk_nvmf_rdma_mgmt_channel { 357 /* Requests that are waiting to obtain a data buffer */ 358 TAILQ_HEAD(, spdk_nvmf_rdma_request) pending_data_buf_queue; 359 }; 360 361 static inline void 362 spdk_nvmf_rdma_qpair_inc_refcnt(struct spdk_nvmf_rdma_qpair *rqpair) 363 { 364 __sync_fetch_and_add(&rqpair->refcnt, 1); 365 } 366 367 static inline uint32_t 368 spdk_nvmf_rdma_qpair_dec_refcnt(struct spdk_nvmf_rdma_qpair *rqpair) 369 { 370 uint32_t old_refcnt, new_refcnt; 371 372 do { 373 old_refcnt = rqpair->refcnt; 374 assert(old_refcnt > 0); 375 new_refcnt = old_refcnt - 1; 376 } while (__sync_bool_compare_and_swap(&rqpair->refcnt, old_refcnt, new_refcnt) == false); 377 378 return new_refcnt; 379 } 380 381 /* API to IBV QueuePair */ 382 static const char *str_ibv_qp_state[] = { 383 "IBV_QPS_RESET", 384 "IBV_QPS_INIT", 385 "IBV_QPS_RTR", 386 "IBV_QPS_RTS", 387 "IBV_QPS_SQD", 388 "IBV_QPS_SQE", 389 "IBV_QPS_ERR" 390 }; 391 392 static enum ibv_qp_state 393 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 394 enum ibv_qp_state old_state, new_state; 395 int rc; 396 397 /* All the attributes needed for recovery */ 398 static int spdk_nvmf_ibv_attr_mask = 399 IBV_QP_STATE | 400 IBV_QP_PKEY_INDEX | 401 IBV_QP_PORT | 402 IBV_QP_ACCESS_FLAGS | 403 IBV_QP_AV | 404 IBV_QP_PATH_MTU | 405 IBV_QP_DEST_QPN | 406 IBV_QP_RQ_PSN | 407 IBV_QP_MAX_DEST_RD_ATOMIC | 408 IBV_QP_MIN_RNR_TIMER | 409 IBV_QP_SQ_PSN | 410 IBV_QP_TIMEOUT | 411 IBV_QP_RETRY_CNT | 412 IBV_QP_RNR_RETRY | 413 IBV_QP_MAX_QP_RD_ATOMIC; 414 415 old_state = rqpair->ibv_attr.qp_state; 416 rc = ibv_query_qp(rqpair->cm_id->qp, &rqpair->ibv_attr, 417 spdk_nvmf_ibv_attr_mask, &rqpair->ibv_init_attr); 418 419 if (rc) 420 { 421 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 422 assert(false); 423 } 424 425 new_state = rqpair->ibv_attr.qp_state; 426 if (old_state != new_state) 427 { 428 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 429 (uintptr_t)rqpair->cm_id, new_state); 430 } 431 return new_state; 432 } 433 434 static int 435 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 436 enum ibv_qp_state new_state) 437 { 438 int rc; 439 enum ibv_qp_state state; 440 static int attr_mask_rc[] = { 441 [IBV_QPS_RESET] = IBV_QP_STATE, 442 [IBV_QPS_INIT] = (IBV_QP_STATE | 443 IBV_QP_PKEY_INDEX | 444 IBV_QP_PORT | 445 IBV_QP_ACCESS_FLAGS), 446 [IBV_QPS_RTR] = (IBV_QP_STATE | 447 IBV_QP_AV | 448 IBV_QP_PATH_MTU | 449 IBV_QP_DEST_QPN | 450 IBV_QP_RQ_PSN | 451 IBV_QP_MAX_DEST_RD_ATOMIC | 452 IBV_QP_MIN_RNR_TIMER), 453 [IBV_QPS_RTS] = (IBV_QP_STATE | 454 IBV_QP_SQ_PSN | 455 IBV_QP_TIMEOUT | 456 IBV_QP_RETRY_CNT | 457 IBV_QP_RNR_RETRY | 458 IBV_QP_MAX_QP_RD_ATOMIC), 459 [IBV_QPS_SQD] = IBV_QP_STATE, 460 [IBV_QPS_SQE] = IBV_QP_STATE, 461 [IBV_QPS_ERR] = IBV_QP_STATE, 462 }; 463 464 switch (new_state) { 465 case IBV_QPS_RESET: 466 case IBV_QPS_INIT: 467 case IBV_QPS_RTR: 468 case IBV_QPS_RTS: 469 case IBV_QPS_SQD: 470 case IBV_QPS_SQE: 471 case IBV_QPS_ERR: 472 break; 473 default: 474 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 475 rqpair->qpair.qid, new_state); 476 return -1; 477 } 478 rqpair->ibv_attr.cur_qp_state = rqpair->ibv_attr.qp_state; 479 rqpair->ibv_attr.qp_state = new_state; 480 rqpair->ibv_attr.ah_attr.port_num = rqpair->ibv_attr.port_num; 481 482 rc = ibv_modify_qp(rqpair->cm_id->qp, &rqpair->ibv_attr, 483 attr_mask_rc[new_state]); 484 485 if (rc) { 486 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 487 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 488 return rc; 489 } 490 491 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 492 493 if (state != new_state) { 494 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 495 rqpair->qpair.qid, str_ibv_qp_state[new_state], 496 str_ibv_qp_state[state]); 497 return -1; 498 } 499 SPDK_NOTICELOG("IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 500 str_ibv_qp_state[state]); 501 return 0; 502 } 503 504 static void 505 spdk_nvmf_rdma_request_set_state(struct spdk_nvmf_rdma_request *rdma_req, 506 enum spdk_nvmf_rdma_request_state state) 507 { 508 struct spdk_nvmf_qpair *qpair; 509 struct spdk_nvmf_rdma_qpair *rqpair; 510 511 qpair = rdma_req->req.qpair; 512 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 513 514 TAILQ_REMOVE(&rqpair->state_queue[rdma_req->state], rdma_req, state_link); 515 rqpair->state_cntr[rdma_req->state]--; 516 517 rdma_req->state = state; 518 519 TAILQ_INSERT_TAIL(&rqpair->state_queue[rdma_req->state], rdma_req, state_link); 520 rqpair->state_cntr[rdma_req->state]++; 521 } 522 523 static int 524 spdk_nvmf_rdma_mgmt_channel_create(void *io_device, void *ctx_buf) 525 { 526 struct spdk_nvmf_rdma_mgmt_channel *ch = ctx_buf; 527 528 TAILQ_INIT(&ch->pending_data_buf_queue); 529 return 0; 530 } 531 532 static void 533 spdk_nvmf_rdma_mgmt_channel_destroy(void *io_device, void *ctx_buf) 534 { 535 struct spdk_nvmf_rdma_mgmt_channel *ch = ctx_buf; 536 537 if (!TAILQ_EMPTY(&ch->pending_data_buf_queue)) { 538 SPDK_ERRLOG("Pending I/O list wasn't empty on channel destruction\n"); 539 } 540 } 541 542 static int 543 spdk_nvmf_rdma_cur_rw_depth(struct spdk_nvmf_rdma_qpair *rqpair) 544 { 545 return rqpair->state_cntr[RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER] + 546 rqpair->state_cntr[RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST]; 547 } 548 549 static int 550 spdk_nvmf_rdma_cur_queue_depth(struct spdk_nvmf_rdma_qpair *rqpair) 551 { 552 return rqpair->max_queue_depth - 553 rqpair->state_cntr[RDMA_REQUEST_STATE_FREE]; 554 } 555 556 static void 557 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 558 { 559 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 560 561 if (spdk_nvmf_rdma_cur_queue_depth(rqpair)) { 562 rqpair->qpair_disconnected = true; 563 return; 564 } 565 566 if (rqpair->refcnt > 0) { 567 return; 568 } 569 570 if (rqpair->poller) { 571 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 572 } 573 574 if (rqpair->cmds_mr) { 575 ibv_dereg_mr(rqpair->cmds_mr); 576 } 577 578 if (rqpair->cpls_mr) { 579 ibv_dereg_mr(rqpair->cpls_mr); 580 } 581 582 if (rqpair->bufs_mr) { 583 ibv_dereg_mr(rqpair->bufs_mr); 584 } 585 586 if (rqpair->cm_id) { 587 rdma_destroy_qp(rqpair->cm_id); 588 rdma_destroy_id(rqpair->cm_id); 589 } 590 591 if (rqpair->mgmt_channel) { 592 spdk_put_io_channel(rqpair->mgmt_channel); 593 } 594 595 /* Free all memory */ 596 spdk_dma_free(rqpair->cmds); 597 spdk_dma_free(rqpair->cpls); 598 spdk_dma_free(rqpair->bufs); 599 free(rqpair->reqs); 600 free(rqpair->recvs); 601 free(rqpair); 602 } 603 604 static int 605 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 606 { 607 struct spdk_nvmf_rdma_transport *rtransport; 608 struct spdk_nvmf_rdma_qpair *rqpair; 609 int rc, i; 610 struct spdk_nvmf_rdma_recv *rdma_recv; 611 struct spdk_nvmf_rdma_request *rdma_req; 612 struct spdk_nvmf_transport *transport; 613 614 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 615 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 616 transport = &rtransport->transport; 617 618 memset(&rqpair->ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 619 rqpair->ibv_init_attr.qp_context = rqpair; 620 rqpair->ibv_init_attr.qp_type = IBV_QPT_RC; 621 rqpair->ibv_init_attr.send_cq = rqpair->poller->cq; 622 rqpair->ibv_init_attr.recv_cq = rqpair->poller->cq; 623 rqpair->ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 624 2; /* SEND, READ, and WRITE operations */ 625 rqpair->ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; /* RECV operations */ 626 rqpair->ibv_init_attr.cap.max_send_sge = rqpair->max_sge; 627 rqpair->ibv_init_attr.cap.max_recv_sge = NVMF_DEFAULT_RX_SGE; 628 629 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &rqpair->ibv_init_attr); 630 if (rc) { 631 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 632 rdma_destroy_id(rqpair->cm_id); 633 rqpair->cm_id = NULL; 634 spdk_nvmf_rdma_qpair_destroy(rqpair); 635 return -1; 636 } 637 638 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 639 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 640 641 rqpair->reqs = calloc(rqpair->max_queue_depth, sizeof(*rqpair->reqs)); 642 rqpair->recvs = calloc(rqpair->max_queue_depth, sizeof(*rqpair->recvs)); 643 rqpair->cmds = spdk_dma_zmalloc(rqpair->max_queue_depth * sizeof(*rqpair->cmds), 644 0x1000, NULL); 645 rqpair->cpls = spdk_dma_zmalloc(rqpair->max_queue_depth * sizeof(*rqpair->cpls), 646 0x1000, NULL); 647 648 649 if (transport->opts.in_capsule_data_size > 0) { 650 rqpair->bufs = spdk_dma_zmalloc(rqpair->max_queue_depth * 651 transport->opts.in_capsule_data_size, 652 0x1000, NULL); 653 } 654 655 if (!rqpair->reqs || !rqpair->recvs || !rqpair->cmds || 656 !rqpair->cpls || (transport->opts.in_capsule_data_size && !rqpair->bufs)) { 657 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 658 spdk_nvmf_rdma_qpair_destroy(rqpair); 659 return -1; 660 } 661 662 rqpair->cmds_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->cmds, 663 rqpair->max_queue_depth * sizeof(*rqpair->cmds), 664 IBV_ACCESS_LOCAL_WRITE); 665 rqpair->cpls_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->cpls, 666 rqpair->max_queue_depth * sizeof(*rqpair->cpls), 667 0); 668 669 if (transport->opts.in_capsule_data_size) { 670 rqpair->bufs_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->bufs, 671 rqpair->max_queue_depth * 672 transport->opts.in_capsule_data_size, 673 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 674 } 675 676 if (!rqpair->cmds_mr || !rqpair->cpls_mr || (transport->opts.in_capsule_data_size && 677 !rqpair->bufs_mr)) { 678 SPDK_ERRLOG("Unable to register required memory for RDMA queue.\n"); 679 spdk_nvmf_rdma_qpair_destroy(rqpair); 680 return -1; 681 } 682 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 683 rqpair->cmds, rqpair->max_queue_depth * sizeof(*rqpair->cmds), rqpair->cmds_mr->lkey); 684 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 685 rqpair->cpls, rqpair->max_queue_depth * sizeof(*rqpair->cpls), rqpair->cpls_mr->lkey); 686 if (rqpair->bufs && rqpair->bufs_mr) { 687 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 688 rqpair->bufs, rqpair->max_queue_depth * 689 transport->opts.in_capsule_data_size, rqpair->bufs_mr->lkey); 690 } 691 692 /* Initialise request state queues and counters of the queue pair */ 693 for (i = RDMA_REQUEST_STATE_FREE; i < RDMA_REQUEST_NUM_STATES; i++) { 694 TAILQ_INIT(&rqpair->state_queue[i]); 695 rqpair->state_cntr[i] = 0; 696 } 697 698 for (i = 0; i < rqpair->max_queue_depth; i++) { 699 struct ibv_recv_wr *bad_wr = NULL; 700 701 rdma_recv = &rqpair->recvs[i]; 702 rdma_recv->qpair = rqpair; 703 704 /* Set up memory to receive commands */ 705 if (rqpair->bufs) { 706 rdma_recv->buf = (void *)((uintptr_t)rqpair->bufs + (i * 707 transport->opts.in_capsule_data_size)); 708 } 709 710 rdma_recv->sgl[0].addr = (uintptr_t)&rqpair->cmds[i]; 711 rdma_recv->sgl[0].length = sizeof(rqpair->cmds[i]); 712 rdma_recv->sgl[0].lkey = rqpair->cmds_mr->lkey; 713 rdma_recv->wr.num_sge = 1; 714 715 if (rdma_recv->buf && rqpair->bufs_mr) { 716 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 717 rdma_recv->sgl[1].length = transport->opts.in_capsule_data_size; 718 rdma_recv->sgl[1].lkey = rqpair->bufs_mr->lkey; 719 rdma_recv->wr.num_sge++; 720 } 721 722 rdma_recv->wr.wr_id = (uintptr_t)rdma_recv; 723 rdma_recv->wr.sg_list = rdma_recv->sgl; 724 725 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_recv->wr, &bad_wr); 726 if (rc) { 727 SPDK_ERRLOG("Unable to post capsule for RDMA RECV\n"); 728 spdk_nvmf_rdma_qpair_destroy(rqpair); 729 return -1; 730 } 731 } 732 733 for (i = 0; i < rqpair->max_queue_depth; i++) { 734 rdma_req = &rqpair->reqs[i]; 735 736 rdma_req->req.qpair = &rqpair->qpair; 737 rdma_req->req.cmd = NULL; 738 739 /* Set up memory to send responses */ 740 rdma_req->req.rsp = &rqpair->cpls[i]; 741 742 rdma_req->rsp.sgl[0].addr = (uintptr_t)&rqpair->cpls[i]; 743 rdma_req->rsp.sgl[0].length = sizeof(rqpair->cpls[i]); 744 rdma_req->rsp.sgl[0].lkey = rqpair->cpls_mr->lkey; 745 746 rdma_req->rsp.wr.wr_id = (uintptr_t)rdma_req; 747 rdma_req->rsp.wr.next = NULL; 748 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 749 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 750 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 751 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 752 753 /* Set up memory for data buffers */ 754 rdma_req->data.wr.wr_id = (uint64_t)rdma_req; 755 rdma_req->data.wr.next = NULL; 756 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 757 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 758 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 759 760 /* Initialize request state to FREE */ 761 rdma_req->state = RDMA_REQUEST_STATE_FREE; 762 TAILQ_INSERT_TAIL(&rqpair->state_queue[rdma_req->state], rdma_req, state_link); 763 rqpair->state_cntr[rdma_req->state]++; 764 } 765 766 return 0; 767 } 768 769 static int 770 request_transfer_in(struct spdk_nvmf_request *req) 771 { 772 int rc; 773 struct spdk_nvmf_rdma_request *rdma_req; 774 struct spdk_nvmf_qpair *qpair; 775 struct spdk_nvmf_rdma_qpair *rqpair; 776 struct ibv_send_wr *bad_wr = NULL; 777 778 qpair = req->qpair; 779 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 780 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 781 782 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 783 784 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA READ POSTED. Request: %p Connection: %p\n", req, qpair); 785 786 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 787 rdma_req->data.wr.next = NULL; 788 rc = ibv_post_send(rqpair->cm_id->qp, &rdma_req->data.wr, &bad_wr); 789 if (rc) { 790 SPDK_ERRLOG("Unable to transfer data from host to target\n"); 791 return -1; 792 } 793 return 0; 794 } 795 796 static int 797 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 798 { 799 int rc; 800 struct spdk_nvmf_rdma_request *rdma_req; 801 struct spdk_nvmf_qpair *qpair; 802 struct spdk_nvmf_rdma_qpair *rqpair; 803 struct spdk_nvme_cpl *rsp; 804 struct ibv_recv_wr *bad_recv_wr = NULL; 805 struct ibv_send_wr *send_wr, *bad_send_wr = NULL; 806 807 *data_posted = 0; 808 qpair = req->qpair; 809 rsp = &req->rsp->nvme_cpl; 810 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 811 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 812 813 /* Advance our sq_head pointer */ 814 if (qpair->sq_head == qpair->sq_head_max) { 815 qpair->sq_head = 0; 816 } else { 817 qpair->sq_head++; 818 } 819 rsp->sqhd = qpair->sq_head; 820 821 /* Post the capsule to the recv buffer */ 822 assert(rdma_req->recv != NULL); 823 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA RECV POSTED. Recv: %p Connection: %p\n", rdma_req->recv, 824 rqpair); 825 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_req->recv->wr, &bad_recv_wr); 826 if (rc) { 827 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 828 return rc; 829 } 830 rdma_req->recv = NULL; 831 832 /* Build the response which consists of an optional 833 * RDMA WRITE to transfer data, plus an RDMA SEND 834 * containing the response. 835 */ 836 send_wr = &rdma_req->rsp.wr; 837 838 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 839 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 840 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA WRITE POSTED. Request: %p Connection: %p\n", req, qpair); 841 842 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 843 844 rdma_req->data.wr.next = send_wr; 845 *data_posted = 1; 846 send_wr = &rdma_req->data.wr; 847 } 848 849 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA SEND POSTED. Request: %p Connection: %p\n", req, qpair); 850 851 /* Send the completion */ 852 rc = ibv_post_send(rqpair->cm_id->qp, send_wr, &bad_send_wr); 853 if (rc) { 854 SPDK_ERRLOG("Unable to send response capsule\n"); 855 } 856 857 return rc; 858 } 859 860 static int 861 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 862 { 863 struct spdk_nvmf_rdma_accept_private_data accept_data; 864 struct rdma_conn_param ctrlr_event_data = {}; 865 int rc; 866 867 accept_data.recfmt = 0; 868 accept_data.crqsize = rqpair->max_queue_depth; 869 870 ctrlr_event_data.private_data = &accept_data; 871 ctrlr_event_data.private_data_len = sizeof(accept_data); 872 if (id->ps == RDMA_PS_TCP) { 873 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 874 ctrlr_event_data.initiator_depth = rqpair->max_rw_depth; 875 } 876 877 rc = rdma_accept(id, &ctrlr_event_data); 878 if (rc) { 879 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 880 } else { 881 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 882 } 883 884 return rc; 885 } 886 887 static void 888 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 889 { 890 struct spdk_nvmf_rdma_reject_private_data rej_data; 891 892 rej_data.recfmt = 0; 893 rej_data.sts = error; 894 895 rdma_reject(id, &rej_data, sizeof(rej_data)); 896 } 897 898 static int 899 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 900 new_qpair_fn cb_fn) 901 { 902 struct spdk_nvmf_rdma_transport *rtransport; 903 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 904 struct spdk_nvmf_rdma_port *port; 905 struct rdma_conn_param *rdma_param = NULL; 906 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 907 uint16_t max_queue_depth; 908 uint16_t max_rw_depth; 909 910 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 911 912 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 913 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 914 915 rdma_param = &event->param.conn; 916 if (rdma_param->private_data == NULL || 917 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 918 SPDK_ERRLOG("connect request: no private data provided\n"); 919 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 920 return -1; 921 } 922 923 private_data = rdma_param->private_data; 924 if (private_data->recfmt != 0) { 925 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 926 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 927 return -1; 928 } 929 930 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 931 event->id->verbs->device->name, event->id->verbs->device->dev_name); 932 933 port = event->listen_id->context; 934 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 935 event->listen_id, event->listen_id->verbs, port); 936 937 /* Figure out the supported queue depth. This is a multi-step process 938 * that takes into account hardware maximums, host provided values, 939 * and our target's internal memory limits */ 940 941 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 942 943 /* Start with the maximum queue depth allowed by the target */ 944 max_queue_depth = rtransport->transport.opts.max_queue_depth; 945 max_rw_depth = rtransport->transport.opts.max_queue_depth; 946 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 947 rtransport->transport.opts.max_queue_depth); 948 949 /* Next check the local NIC's hardware limitations */ 950 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 951 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 952 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 953 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 954 max_rw_depth = spdk_min(max_rw_depth, port->device->attr.max_qp_rd_atom); 955 956 /* Next check the remote NIC's hardware limitations */ 957 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 958 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 959 rdma_param->initiator_depth, rdma_param->responder_resources); 960 if (rdma_param->initiator_depth > 0) { 961 max_rw_depth = spdk_min(max_rw_depth, rdma_param->initiator_depth); 962 } 963 964 /* Finally check for the host software requested values, which are 965 * optional. */ 966 if (rdma_param->private_data != NULL && 967 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 968 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 969 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 970 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 971 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 972 } 973 974 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 975 max_queue_depth, max_rw_depth); 976 977 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 978 if (rqpair == NULL) { 979 SPDK_ERRLOG("Could not allocate new connection.\n"); 980 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 981 return -1; 982 } 983 984 rqpair->port = port; 985 rqpair->max_queue_depth = max_queue_depth; 986 rqpair->max_rw_depth = max_rw_depth; 987 rqpair->cm_id = event->id; 988 rqpair->listen_id = event->listen_id; 989 rqpair->qpair.transport = transport; 990 rqpair->max_sge = spdk_min(port->device->attr.max_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 991 TAILQ_INIT(&rqpair->incoming_queue); 992 event->id->context = &rqpair->qpair; 993 994 cb_fn(&rqpair->qpair); 995 996 return 0; 997 } 998 999 static void 1000 _nvmf_rdma_disconnect(void *ctx) 1001 { 1002 struct spdk_nvmf_qpair *qpair = ctx; 1003 struct spdk_nvmf_rdma_qpair *rqpair; 1004 1005 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1006 1007 spdk_nvmf_rdma_qpair_dec_refcnt(rqpair); 1008 1009 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 1010 } 1011 1012 static int 1013 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 1014 { 1015 struct spdk_nvmf_qpair *qpair; 1016 struct spdk_nvmf_rdma_qpair *rqpair; 1017 1018 if (evt->id == NULL) { 1019 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 1020 return -1; 1021 } 1022 1023 qpair = evt->id->context; 1024 if (qpair == NULL) { 1025 SPDK_ERRLOG("disconnect request: no active connection\n"); 1026 return -1; 1027 } 1028 1029 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1030 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 1031 spdk_nvmf_rdma_update_ibv_state(rqpair); 1032 1033 spdk_nvmf_rdma_qpair_inc_refcnt(rqpair); 1034 1035 spdk_thread_send_msg(qpair->group->thread, _nvmf_rdma_disconnect, qpair); 1036 1037 return 0; 1038 } 1039 1040 #ifdef DEBUG 1041 static const char *CM_EVENT_STR[] = { 1042 "RDMA_CM_EVENT_ADDR_RESOLVED", 1043 "RDMA_CM_EVENT_ADDR_ERROR", 1044 "RDMA_CM_EVENT_ROUTE_RESOLVED", 1045 "RDMA_CM_EVENT_ROUTE_ERROR", 1046 "RDMA_CM_EVENT_CONNECT_REQUEST", 1047 "RDMA_CM_EVENT_CONNECT_RESPONSE", 1048 "RDMA_CM_EVENT_CONNECT_ERROR", 1049 "RDMA_CM_EVENT_UNREACHABLE", 1050 "RDMA_CM_EVENT_REJECTED", 1051 "RDMA_CM_EVENT_ESTABLISHED", 1052 "RDMA_CM_EVENT_DISCONNECTED", 1053 "RDMA_CM_EVENT_DEVICE_REMOVAL", 1054 "RDMA_CM_EVENT_MULTICAST_JOIN", 1055 "RDMA_CM_EVENT_MULTICAST_ERROR", 1056 "RDMA_CM_EVENT_ADDR_CHANGE", 1057 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 1058 }; 1059 #endif /* DEBUG */ 1060 1061 static void 1062 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 1063 { 1064 struct spdk_nvmf_rdma_transport *rtransport; 1065 struct rdma_cm_event *event; 1066 int rc; 1067 1068 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1069 1070 if (rtransport->event_channel == NULL) { 1071 return; 1072 } 1073 1074 while (1) { 1075 rc = rdma_get_cm_event(rtransport->event_channel, &event); 1076 if (rc == 0) { 1077 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 1078 1079 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 1080 1081 switch (event->event) { 1082 case RDMA_CM_EVENT_ADDR_RESOLVED: 1083 case RDMA_CM_EVENT_ADDR_ERROR: 1084 case RDMA_CM_EVENT_ROUTE_RESOLVED: 1085 case RDMA_CM_EVENT_ROUTE_ERROR: 1086 /* No action required. The target never attempts to resolve routes. */ 1087 break; 1088 case RDMA_CM_EVENT_CONNECT_REQUEST: 1089 rc = nvmf_rdma_connect(transport, event, cb_fn); 1090 if (rc < 0) { 1091 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 1092 break; 1093 } 1094 break; 1095 case RDMA_CM_EVENT_CONNECT_RESPONSE: 1096 /* The target never initiates a new connection. So this will not occur. */ 1097 break; 1098 case RDMA_CM_EVENT_CONNECT_ERROR: 1099 /* Can this happen? The docs say it can, but not sure what causes it. */ 1100 break; 1101 case RDMA_CM_EVENT_UNREACHABLE: 1102 case RDMA_CM_EVENT_REJECTED: 1103 /* These only occur on the client side. */ 1104 break; 1105 case RDMA_CM_EVENT_ESTABLISHED: 1106 /* TODO: Should we be waiting for this event anywhere? */ 1107 break; 1108 case RDMA_CM_EVENT_DISCONNECTED: 1109 case RDMA_CM_EVENT_DEVICE_REMOVAL: 1110 rc = nvmf_rdma_disconnect(event); 1111 if (rc < 0) { 1112 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 1113 break; 1114 } 1115 break; 1116 case RDMA_CM_EVENT_MULTICAST_JOIN: 1117 case RDMA_CM_EVENT_MULTICAST_ERROR: 1118 /* Multicast is not used */ 1119 break; 1120 case RDMA_CM_EVENT_ADDR_CHANGE: 1121 /* Not utilizing this event */ 1122 break; 1123 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 1124 /* For now, do nothing. The target never re-uses queue pairs. */ 1125 break; 1126 default: 1127 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 1128 break; 1129 } 1130 1131 rdma_ack_cm_event(event); 1132 } else { 1133 if (errno != EAGAIN && errno != EWOULDBLOCK) { 1134 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 1135 } 1136 break; 1137 } 1138 } 1139 } 1140 1141 static int 1142 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1143 enum spdk_mem_map_notify_action action, 1144 void *vaddr, size_t size) 1145 { 1146 struct spdk_nvmf_rdma_device *device = cb_ctx; 1147 struct ibv_pd *pd = device->pd; 1148 struct ibv_mr *mr; 1149 1150 switch (action) { 1151 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1152 mr = ibv_reg_mr(pd, vaddr, size, 1153 IBV_ACCESS_LOCAL_WRITE | 1154 IBV_ACCESS_REMOTE_READ | 1155 IBV_ACCESS_REMOTE_WRITE); 1156 if (mr == NULL) { 1157 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1158 return -1; 1159 } else { 1160 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1161 } 1162 break; 1163 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1164 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1165 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1166 if (mr) { 1167 ibv_dereg_mr(mr); 1168 } 1169 break; 1170 } 1171 1172 return 0; 1173 } 1174 1175 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1176 1177 static spdk_nvme_data_transfer_t 1178 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1179 { 1180 enum spdk_nvme_data_transfer xfer; 1181 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1182 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1183 1184 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1185 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1186 rdma_req->rsp.wr.imm_data = 0; 1187 #endif 1188 1189 /* Figure out data transfer direction */ 1190 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1191 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1192 } else { 1193 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1194 1195 /* Some admin commands are special cases */ 1196 if ((rdma_req->req.qpair->qid == 0) && 1197 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1198 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1199 switch (cmd->cdw10 & 0xff) { 1200 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1201 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1202 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1203 break; 1204 default: 1205 xfer = SPDK_NVME_DATA_NONE; 1206 } 1207 } 1208 } 1209 1210 if (xfer == SPDK_NVME_DATA_NONE) { 1211 return xfer; 1212 } 1213 1214 /* Even for commands that may transfer data, they could have specified 0 length. 1215 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1216 */ 1217 switch (sgl->generic.type) { 1218 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1219 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1220 case SPDK_NVME_SGL_TYPE_SEGMENT: 1221 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1222 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1223 if (sgl->unkeyed.length == 0) { 1224 xfer = SPDK_NVME_DATA_NONE; 1225 } 1226 break; 1227 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1228 if (sgl->keyed.length == 0) { 1229 xfer = SPDK_NVME_DATA_NONE; 1230 } 1231 break; 1232 } 1233 1234 return xfer; 1235 } 1236 1237 static int 1238 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1239 struct spdk_nvmf_rdma_device *device, 1240 struct spdk_nvmf_rdma_request *rdma_req) 1241 { 1242 void *buf = NULL; 1243 uint32_t length = rdma_req->req.length; 1244 uint32_t i = 0; 1245 1246 rdma_req->req.iovcnt = 0; 1247 while (length) { 1248 buf = spdk_mempool_get(rtransport->data_buf_pool); 1249 if (!buf) { 1250 goto nomem; 1251 } 1252 1253 rdma_req->req.iov[i].iov_base = (void *)((uintptr_t)(buf + NVMF_DATA_BUFFER_MASK) & 1254 ~NVMF_DATA_BUFFER_MASK); 1255 rdma_req->req.iov[i].iov_len = spdk_min(length, rtransport->transport.opts.io_unit_size); 1256 rdma_req->req.iovcnt++; 1257 rdma_req->data.buffers[i] = buf; 1258 rdma_req->data.wr.sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[i].iov_base); 1259 rdma_req->data.wr.sg_list[i].length = rdma_req->req.iov[i].iov_len; 1260 rdma_req->data.wr.sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1261 (uint64_t)buf, NULL))->lkey; 1262 1263 length -= rdma_req->req.iov[i].iov_len; 1264 i++; 1265 } 1266 1267 rdma_req->data_from_pool = true; 1268 1269 return 0; 1270 1271 nomem: 1272 while (i) { 1273 i--; 1274 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->req.iov[i].iov_base); 1275 rdma_req->req.iov[i].iov_base = NULL; 1276 rdma_req->req.iov[i].iov_len = 0; 1277 1278 rdma_req->data.wr.sg_list[i].addr = 0; 1279 rdma_req->data.wr.sg_list[i].length = 0; 1280 rdma_req->data.wr.sg_list[i].lkey = 0; 1281 } 1282 rdma_req->req.iovcnt = 0; 1283 return -ENOMEM; 1284 } 1285 1286 static int 1287 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1288 struct spdk_nvmf_rdma_device *device, 1289 struct spdk_nvmf_rdma_request *rdma_req) 1290 { 1291 struct spdk_nvme_cmd *cmd; 1292 struct spdk_nvme_cpl *rsp; 1293 struct spdk_nvme_sgl_descriptor *sgl; 1294 1295 cmd = &rdma_req->req.cmd->nvme_cmd; 1296 rsp = &rdma_req->req.rsp->nvme_cpl; 1297 sgl = &cmd->dptr.sgl1; 1298 1299 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1300 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1301 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1302 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1303 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1304 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1305 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1306 return -1; 1307 } 1308 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1309 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1310 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1311 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1312 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1313 } 1314 } 1315 #endif 1316 1317 /* fill request length and populate iovs */ 1318 rdma_req->req.length = sgl->keyed.length; 1319 1320 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1321 /* No available buffers. Queue this request up. */ 1322 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1323 return 0; 1324 } 1325 1326 /* backward compatible */ 1327 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1328 1329 /* rdma wr specifics */ 1330 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1331 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1332 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1333 1334 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1335 rdma_req->req.iovcnt); 1336 1337 return 0; 1338 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1339 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1340 uint64_t offset = sgl->address; 1341 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1342 1343 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1344 offset, sgl->unkeyed.length); 1345 1346 if (offset > max_len) { 1347 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1348 offset, max_len); 1349 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1350 return -1; 1351 } 1352 max_len -= (uint32_t)offset; 1353 1354 if (sgl->unkeyed.length > max_len) { 1355 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1356 sgl->unkeyed.length, max_len); 1357 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1358 return -1; 1359 } 1360 1361 rdma_req->req.data = rdma_req->recv->buf + offset; 1362 rdma_req->data_from_pool = false; 1363 rdma_req->req.length = sgl->unkeyed.length; 1364 1365 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1366 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1367 rdma_req->req.iovcnt = 1; 1368 1369 return 0; 1370 } 1371 1372 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1373 sgl->generic.type, sgl->generic.subtype); 1374 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1375 return -1; 1376 } 1377 1378 static bool 1379 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1380 struct spdk_nvmf_rdma_request *rdma_req) 1381 { 1382 struct spdk_nvmf_rdma_qpair *rqpair; 1383 struct spdk_nvmf_rdma_device *device; 1384 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1385 int rc; 1386 struct spdk_nvmf_rdma_recv *rdma_recv; 1387 enum spdk_nvmf_rdma_request_state prev_state; 1388 bool progress = false; 1389 int data_posted; 1390 int cur_rdma_rw_depth; 1391 1392 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1393 device = rqpair->port->device; 1394 1395 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1396 1397 /* If the queue pair is in an error state, force the request to the completed state 1398 * to release resources. */ 1399 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1400 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1401 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1402 } 1403 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1404 } 1405 1406 /* The loop here is to allow for several back-to-back state changes. */ 1407 do { 1408 prev_state = rdma_req->state; 1409 1410 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1411 1412 switch (rdma_req->state) { 1413 case RDMA_REQUEST_STATE_FREE: 1414 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1415 * to escape this state. */ 1416 break; 1417 case RDMA_REQUEST_STATE_NEW: 1418 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1419 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1420 rdma_recv = rdma_req->recv; 1421 1422 /* The first element of the SGL is the NVMe command */ 1423 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1424 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1425 1426 TAILQ_REMOVE(&rqpair->incoming_queue, rdma_recv, link); 1427 1428 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 1429 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1430 break; 1431 } 1432 1433 /* The next state transition depends on the data transfer needs of this request. */ 1434 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1435 1436 /* If no data to transfer, ready to execute. */ 1437 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1438 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1439 break; 1440 } 1441 1442 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEED_BUFFER); 1443 TAILQ_INSERT_TAIL(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1444 break; 1445 case RDMA_REQUEST_STATE_NEED_BUFFER: 1446 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1447 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1448 1449 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1450 1451 if (rdma_req != TAILQ_FIRST(&rqpair->ch->pending_data_buf_queue)) { 1452 /* This request needs to wait in line to obtain a buffer */ 1453 break; 1454 } 1455 1456 /* Try to get a data buffer */ 1457 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1458 if (rc < 0) { 1459 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1460 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1461 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1462 break; 1463 } 1464 1465 if (!rdma_req->req.data) { 1466 /* No buffers available. */ 1467 break; 1468 } 1469 1470 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1471 1472 /* If data is transferring from host to controller and the data didn't 1473 * arrive using in capsule data, we need to do a transfer from the host. 1474 */ 1475 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1476 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1477 break; 1478 } 1479 1480 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1481 break; 1482 case RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING: 1483 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 0, 0, 1484 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1485 1486 if (rdma_req != TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING])) { 1487 /* This request needs to wait in line to perform RDMA */ 1488 break; 1489 } 1490 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1491 1492 if (cur_rdma_rw_depth >= rqpair->max_rw_depth) { 1493 /* R/W queue is full, need to wait */ 1494 break; 1495 } 1496 1497 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1498 rc = request_transfer_in(&rdma_req->req); 1499 if (!rc) { 1500 spdk_nvmf_rdma_request_set_state(rdma_req, 1501 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1502 } else { 1503 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1504 spdk_nvmf_rdma_request_set_state(rdma_req, 1505 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1506 } 1507 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1508 /* The data transfer will be kicked off from 1509 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1510 */ 1511 spdk_nvmf_rdma_request_set_state(rdma_req, 1512 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1513 } else { 1514 SPDK_ERRLOG("Cannot perform data transfer, unknown state: %u\n", 1515 rdma_req->req.xfer); 1516 assert(0); 1517 } 1518 break; 1519 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1520 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1521 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1522 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1523 * to escape this state. */ 1524 break; 1525 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1526 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1527 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1528 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTING); 1529 spdk_nvmf_request_exec(&rdma_req->req); 1530 break; 1531 case RDMA_REQUEST_STATE_EXECUTING: 1532 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1533 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1534 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1535 * to escape this state. */ 1536 break; 1537 case RDMA_REQUEST_STATE_EXECUTED: 1538 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1539 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1540 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1541 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1542 } else { 1543 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1544 } 1545 break; 1546 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1547 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 1548 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1549 rc = request_transfer_out(&rdma_req->req, &data_posted); 1550 assert(rc == 0); /* No good way to handle this currently */ 1551 if (rc) { 1552 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1553 } else { 1554 spdk_nvmf_rdma_request_set_state(rdma_req, 1555 data_posted ? 1556 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1557 RDMA_REQUEST_STATE_COMPLETING); 1558 } 1559 break; 1560 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1561 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 1562 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1563 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1564 * to escape this state. */ 1565 break; 1566 case RDMA_REQUEST_STATE_COMPLETING: 1567 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 1568 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1569 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1570 * to escape this state. */ 1571 break; 1572 case RDMA_REQUEST_STATE_COMPLETED: 1573 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 1574 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1575 1576 if (rdma_req->data_from_pool) { 1577 /* Put the buffer/s back in the pool */ 1578 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 1579 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 1580 rdma_req->req.iov[i].iov_base = NULL; 1581 rdma_req->data.buffers[i] = NULL; 1582 } 1583 rdma_req->data_from_pool = false; 1584 } 1585 rdma_req->req.length = 0; 1586 rdma_req->req.iovcnt = 0; 1587 rdma_req->req.data = NULL; 1588 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 1589 break; 1590 case RDMA_REQUEST_NUM_STATES: 1591 default: 1592 assert(0); 1593 break; 1594 } 1595 1596 if (rdma_req->state != prev_state) { 1597 progress = true; 1598 } 1599 } while (rdma_req->state != prev_state); 1600 1601 return progress; 1602 } 1603 1604 /* Public API callbacks begin here */ 1605 1606 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 1607 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 1608 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 1609 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 1610 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 1611 #define SPDK_NVMF_RDMA_DEFAULT_IO_UNIT_SIZE 131072 1612 1613 static void 1614 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 1615 { 1616 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 1617 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 1618 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 1619 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 1620 opts->io_unit_size = SPDK_NVMF_RDMA_DEFAULT_IO_UNIT_SIZE; 1621 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 1622 } 1623 1624 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 1625 1626 static struct spdk_nvmf_transport * 1627 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 1628 { 1629 int rc; 1630 struct spdk_nvmf_rdma_transport *rtransport; 1631 struct spdk_nvmf_rdma_device *device, *tmp; 1632 struct ibv_context **contexts; 1633 uint32_t i; 1634 int flag; 1635 uint32_t sge_count; 1636 1637 rtransport = calloc(1, sizeof(*rtransport)); 1638 if (!rtransport) { 1639 return NULL; 1640 } 1641 1642 if (pthread_mutex_init(&rtransport->lock, NULL)) { 1643 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 1644 free(rtransport); 1645 return NULL; 1646 } 1647 1648 spdk_io_device_register(rtransport, spdk_nvmf_rdma_mgmt_channel_create, 1649 spdk_nvmf_rdma_mgmt_channel_destroy, 1650 sizeof(struct spdk_nvmf_rdma_mgmt_channel), 1651 "rdma_transport"); 1652 1653 TAILQ_INIT(&rtransport->devices); 1654 TAILQ_INIT(&rtransport->ports); 1655 1656 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 1657 1658 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 1659 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 1660 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 1661 " in_capsule_data_size=%d, max_aq_depth=%d\n", 1662 opts->max_queue_depth, 1663 opts->max_io_size, 1664 opts->max_qpairs_per_ctrlr, 1665 opts->io_unit_size, 1666 opts->in_capsule_data_size, 1667 opts->max_aq_depth); 1668 1669 /* I/O unit size cannot be larger than max I/O size */ 1670 if (opts->io_unit_size > opts->max_io_size) { 1671 opts->io_unit_size = opts->max_io_size; 1672 } 1673 1674 sge_count = opts->max_io_size / opts->io_unit_size; 1675 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 1676 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 1677 spdk_nvmf_rdma_destroy(&rtransport->transport); 1678 return NULL; 1679 } 1680 1681 rtransport->event_channel = rdma_create_event_channel(); 1682 if (rtransport->event_channel == NULL) { 1683 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 1684 spdk_nvmf_rdma_destroy(&rtransport->transport); 1685 return NULL; 1686 } 1687 1688 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 1689 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 1690 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 1691 rtransport->event_channel->fd, spdk_strerror(errno)); 1692 spdk_nvmf_rdma_destroy(&rtransport->transport); 1693 return NULL; 1694 } 1695 1696 rtransport->data_buf_pool = spdk_mempool_create("spdk_nvmf_rdma", 1697 opts->max_queue_depth * 4, /* The 4 is arbitrarily chosen. Needs to be configurable. */ 1698 opts->max_io_size + NVMF_DATA_BUFFER_ALIGNMENT, 1699 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 1700 SPDK_ENV_SOCKET_ID_ANY); 1701 if (!rtransport->data_buf_pool) { 1702 SPDK_ERRLOG("Unable to allocate buffer pool for poll group\n"); 1703 spdk_nvmf_rdma_destroy(&rtransport->transport); 1704 return NULL; 1705 } 1706 1707 contexts = rdma_get_devices(NULL); 1708 if (contexts == NULL) { 1709 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 1710 spdk_nvmf_rdma_destroy(&rtransport->transport); 1711 return NULL; 1712 } 1713 1714 i = 0; 1715 rc = 0; 1716 while (contexts[i] != NULL) { 1717 device = calloc(1, sizeof(*device)); 1718 if (!device) { 1719 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 1720 rc = -ENOMEM; 1721 break; 1722 } 1723 device->context = contexts[i]; 1724 rc = ibv_query_device(device->context, &device->attr); 1725 if (rc < 0) { 1726 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1727 free(device); 1728 break; 1729 1730 } 1731 1732 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1733 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 1734 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 1735 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 1736 } 1737 1738 /** 1739 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 1740 * The Soft-RoCE RXE driver does not currently support send with invalidate, 1741 * but incorrectly reports that it does. There are changes making their way 1742 * through the kernel now that will enable this feature. When they are merged, 1743 * we can conditionally enable this feature. 1744 * 1745 * TODO: enable this for versions of the kernel rxe driver that support it. 1746 */ 1747 if (device->attr.vendor_id == 0) { 1748 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 1749 } 1750 #endif 1751 1752 /* set up device context async ev fd as NON_BLOCKING */ 1753 flag = fcntl(device->context->async_fd, F_GETFL); 1754 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 1755 if (rc < 0) { 1756 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 1757 free(device); 1758 break; 1759 } 1760 1761 device->pd = ibv_alloc_pd(device->context); 1762 if (!device->pd) { 1763 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 1764 free(device); 1765 rc = -1; 1766 break; 1767 } 1768 1769 device->map = spdk_mem_map_alloc(0, spdk_nvmf_rdma_mem_notify, device); 1770 if (!device->map) { 1771 SPDK_ERRLOG("Unable to allocate memory map for new poll group\n"); 1772 ibv_dealloc_pd(device->pd); 1773 free(device); 1774 rc = -1; 1775 break; 1776 } 1777 1778 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 1779 i++; 1780 } 1781 rdma_free_devices(contexts); 1782 1783 if (rc < 0) { 1784 spdk_nvmf_rdma_destroy(&rtransport->transport); 1785 return NULL; 1786 } 1787 1788 /* Set up poll descriptor array to monitor events from RDMA and IB 1789 * in a single poll syscall 1790 */ 1791 rtransport->npoll_fds = i + 1; 1792 i = 0; 1793 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 1794 if (rtransport->poll_fds == NULL) { 1795 SPDK_ERRLOG("poll_fds allocation failed\n"); 1796 spdk_nvmf_rdma_destroy(&rtransport->transport); 1797 return NULL; 1798 } 1799 1800 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 1801 rtransport->poll_fds[i++].events = POLLIN; 1802 1803 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 1804 rtransport->poll_fds[i].fd = device->context->async_fd; 1805 rtransport->poll_fds[i++].events = POLLIN; 1806 } 1807 1808 return &rtransport->transport; 1809 } 1810 1811 static int 1812 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 1813 { 1814 struct spdk_nvmf_rdma_transport *rtransport; 1815 struct spdk_nvmf_rdma_port *port, *port_tmp; 1816 struct spdk_nvmf_rdma_device *device, *device_tmp; 1817 1818 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1819 1820 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 1821 TAILQ_REMOVE(&rtransport->ports, port, link); 1822 rdma_destroy_id(port->id); 1823 free(port); 1824 } 1825 1826 if (rtransport->poll_fds != NULL) { 1827 free(rtransport->poll_fds); 1828 } 1829 1830 if (rtransport->event_channel != NULL) { 1831 rdma_destroy_event_channel(rtransport->event_channel); 1832 } 1833 1834 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 1835 TAILQ_REMOVE(&rtransport->devices, device, link); 1836 if (device->map) { 1837 spdk_mem_map_free(&device->map); 1838 } 1839 if (device->pd) { 1840 ibv_dealloc_pd(device->pd); 1841 } 1842 free(device); 1843 } 1844 1845 if (rtransport->data_buf_pool != NULL) { 1846 if (spdk_mempool_count(rtransport->data_buf_pool) != 1847 (transport->opts.max_queue_depth * 4)) { 1848 SPDK_ERRLOG("transport buffer pool count is %zu but should be %u\n", 1849 spdk_mempool_count(rtransport->data_buf_pool), 1850 transport->opts.max_queue_depth * 4); 1851 } 1852 } 1853 1854 spdk_mempool_free(rtransport->data_buf_pool); 1855 spdk_io_device_unregister(rtransport, NULL); 1856 pthread_mutex_destroy(&rtransport->lock); 1857 free(rtransport); 1858 1859 return 0; 1860 } 1861 1862 static int 1863 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 1864 const struct spdk_nvme_transport_id *trid) 1865 { 1866 struct spdk_nvmf_rdma_transport *rtransport; 1867 struct spdk_nvmf_rdma_device *device; 1868 struct spdk_nvmf_rdma_port *port_tmp, *port; 1869 struct addrinfo *res; 1870 struct addrinfo hints; 1871 int family; 1872 int rc; 1873 1874 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1875 1876 port = calloc(1, sizeof(*port)); 1877 if (!port) { 1878 return -ENOMEM; 1879 } 1880 1881 /* Selectively copy the trid. Things like NQN don't matter here - that 1882 * mapping is enforced elsewhere. 1883 */ 1884 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1885 port->trid.adrfam = trid->adrfam; 1886 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 1887 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 1888 1889 pthread_mutex_lock(&rtransport->lock); 1890 assert(rtransport->event_channel != NULL); 1891 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 1892 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 1893 port_tmp->ref++; 1894 free(port); 1895 /* Already listening at this address */ 1896 pthread_mutex_unlock(&rtransport->lock); 1897 return 0; 1898 } 1899 } 1900 1901 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 1902 if (rc < 0) { 1903 SPDK_ERRLOG("rdma_create_id() failed\n"); 1904 free(port); 1905 pthread_mutex_unlock(&rtransport->lock); 1906 return rc; 1907 } 1908 1909 switch (port->trid.adrfam) { 1910 case SPDK_NVMF_ADRFAM_IPV4: 1911 family = AF_INET; 1912 break; 1913 case SPDK_NVMF_ADRFAM_IPV6: 1914 family = AF_INET6; 1915 break; 1916 default: 1917 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 1918 free(port); 1919 pthread_mutex_unlock(&rtransport->lock); 1920 return -EINVAL; 1921 } 1922 1923 memset(&hints, 0, sizeof(hints)); 1924 hints.ai_family = family; 1925 hints.ai_flags = AI_NUMERICSERV; 1926 hints.ai_socktype = SOCK_STREAM; 1927 hints.ai_protocol = 0; 1928 1929 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 1930 if (rc) { 1931 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 1932 free(port); 1933 pthread_mutex_unlock(&rtransport->lock); 1934 return -EINVAL; 1935 } 1936 1937 rc = rdma_bind_addr(port->id, res->ai_addr); 1938 freeaddrinfo(res); 1939 1940 if (rc < 0) { 1941 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 1942 rdma_destroy_id(port->id); 1943 free(port); 1944 pthread_mutex_unlock(&rtransport->lock); 1945 return rc; 1946 } 1947 1948 if (!port->id->verbs) { 1949 SPDK_ERRLOG("ibv_context is null\n"); 1950 rdma_destroy_id(port->id); 1951 free(port); 1952 pthread_mutex_unlock(&rtransport->lock); 1953 return -1; 1954 } 1955 1956 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 1957 if (rc < 0) { 1958 SPDK_ERRLOG("rdma_listen() failed\n"); 1959 rdma_destroy_id(port->id); 1960 free(port); 1961 pthread_mutex_unlock(&rtransport->lock); 1962 return rc; 1963 } 1964 1965 TAILQ_FOREACH(device, &rtransport->devices, link) { 1966 if (device->context == port->id->verbs) { 1967 port->device = device; 1968 break; 1969 } 1970 } 1971 if (!port->device) { 1972 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 1973 port->id->verbs); 1974 rdma_destroy_id(port->id); 1975 free(port); 1976 pthread_mutex_unlock(&rtransport->lock); 1977 return -EINVAL; 1978 } 1979 1980 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 1981 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 1982 1983 port->ref = 1; 1984 1985 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 1986 pthread_mutex_unlock(&rtransport->lock); 1987 1988 return 0; 1989 } 1990 1991 static int 1992 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 1993 const struct spdk_nvme_transport_id *_trid) 1994 { 1995 struct spdk_nvmf_rdma_transport *rtransport; 1996 struct spdk_nvmf_rdma_port *port, *tmp; 1997 struct spdk_nvme_transport_id trid = {}; 1998 1999 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2000 2001 /* Selectively copy the trid. Things like NQN don't matter here - that 2002 * mapping is enforced elsewhere. 2003 */ 2004 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2005 trid.adrfam = _trid->adrfam; 2006 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2007 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2008 2009 pthread_mutex_lock(&rtransport->lock); 2010 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2011 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2012 assert(port->ref > 0); 2013 port->ref--; 2014 if (port->ref == 0) { 2015 TAILQ_REMOVE(&rtransport->ports, port, link); 2016 rdma_destroy_id(port->id); 2017 free(port); 2018 } 2019 break; 2020 } 2021 } 2022 2023 pthread_mutex_unlock(&rtransport->lock); 2024 return 0; 2025 } 2026 2027 static bool 2028 spdk_nvmf_rdma_qpair_is_idle(struct spdk_nvmf_qpair *qpair) 2029 { 2030 int cur_queue_depth, cur_rdma_rw_depth; 2031 struct spdk_nvmf_rdma_qpair *rqpair; 2032 2033 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2034 cur_queue_depth = spdk_nvmf_rdma_cur_queue_depth(rqpair); 2035 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 2036 2037 if (cur_queue_depth == 0 && cur_rdma_rw_depth == 0) { 2038 return true; 2039 } 2040 return false; 2041 } 2042 2043 static void 2044 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2045 struct spdk_nvmf_rdma_qpair *rqpair) 2046 { 2047 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 2048 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2049 2050 /* We process I/O in the data transfer pending queue at the highest priority. */ 2051 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING], 2052 state_link, req_tmp) { 2053 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2054 break; 2055 } 2056 } 2057 2058 /* The second highest priority is I/O waiting on memory buffers. */ 2059 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->ch->pending_data_buf_queue, link, 2060 req_tmp) { 2061 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2062 break; 2063 } 2064 } 2065 2066 if (rqpair->qpair_disconnected) { 2067 spdk_nvmf_rdma_qpair_destroy(rqpair); 2068 return; 2069 } 2070 2071 /* Do not process newly received commands if qp is in ERROR state, 2072 * wait till the recovery is complete. 2073 */ 2074 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 2075 return; 2076 } 2077 2078 /* The lowest priority is processing newly received commands */ 2079 TAILQ_FOREACH_SAFE(rdma_recv, &rqpair->incoming_queue, link, recv_tmp) { 2080 if (TAILQ_EMPTY(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE])) { 2081 break; 2082 } 2083 2084 rdma_req = TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE]); 2085 rdma_req->recv = rdma_recv; 2086 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEW); 2087 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2088 break; 2089 } 2090 } 2091 } 2092 2093 static void 2094 spdk_nvmf_rdma_drain_state_queue(struct spdk_nvmf_rdma_qpair *rqpair, 2095 enum spdk_nvmf_rdma_request_state state) 2096 { 2097 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2098 struct spdk_nvmf_rdma_transport *rtransport; 2099 2100 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[state], state_link, req_tmp) { 2101 rtransport = SPDK_CONTAINEROF(rdma_req->req.qpair->transport, 2102 struct spdk_nvmf_rdma_transport, transport); 2103 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2104 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2105 } 2106 } 2107 2108 static void 2109 spdk_nvmf_rdma_qpair_recover(struct spdk_nvmf_rdma_qpair *rqpair) 2110 { 2111 enum ibv_qp_state state, next_state; 2112 int recovered; 2113 struct spdk_nvmf_rdma_transport *rtransport; 2114 2115 if (!spdk_nvmf_rdma_qpair_is_idle(&rqpair->qpair)) { 2116 /* There must be outstanding requests down to media. 2117 * If so, wait till they're complete. 2118 */ 2119 assert(!TAILQ_EMPTY(&rqpair->qpair.outstanding)); 2120 return; 2121 } 2122 2123 state = rqpair->ibv_attr.qp_state; 2124 next_state = state; 2125 2126 SPDK_NOTICELOG("RDMA qpair %u is in state: %s\n", 2127 rqpair->qpair.qid, 2128 str_ibv_qp_state[state]); 2129 2130 if (!(state == IBV_QPS_ERR || state == IBV_QPS_RESET)) { 2131 SPDK_ERRLOG("Can't recover RDMA qpair %u from the state: %s\n", 2132 rqpair->qpair.qid, 2133 str_ibv_qp_state[state]); 2134 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2135 return; 2136 } 2137 2138 recovered = 0; 2139 while (!recovered) { 2140 switch (state) { 2141 case IBV_QPS_ERR: 2142 next_state = IBV_QPS_RESET; 2143 break; 2144 case IBV_QPS_RESET: 2145 next_state = IBV_QPS_INIT; 2146 break; 2147 case IBV_QPS_INIT: 2148 next_state = IBV_QPS_RTR; 2149 break; 2150 case IBV_QPS_RTR: 2151 next_state = IBV_QPS_RTS; 2152 break; 2153 case IBV_QPS_RTS: 2154 recovered = 1; 2155 break; 2156 default: 2157 SPDK_ERRLOG("RDMA qpair %u unexpected state for recovery: %u\n", 2158 rqpair->qpair.qid, state); 2159 goto error; 2160 } 2161 /* Do not transition into same state */ 2162 if (next_state == state) { 2163 break; 2164 } 2165 2166 if (spdk_nvmf_rdma_set_ibv_state(rqpair, next_state)) { 2167 goto error; 2168 } 2169 2170 state = next_state; 2171 } 2172 2173 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2174 struct spdk_nvmf_rdma_transport, 2175 transport); 2176 2177 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2178 2179 return; 2180 error: 2181 SPDK_NOTICELOG("RDMA qpair %u: recovery failed, disconnecting...\n", 2182 rqpair->qpair.qid); 2183 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2184 } 2185 2186 /* Clean up only the states that can be aborted at any time */ 2187 static void 2188 _spdk_nvmf_rdma_qp_cleanup_safe_states(struct spdk_nvmf_rdma_qpair *rqpair) 2189 { 2190 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2191 2192 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_NEW); 2193 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[RDMA_REQUEST_STATE_NEED_BUFFER], link, req_tmp) { 2194 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 2195 } 2196 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_NEED_BUFFER); 2197 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 2198 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 2199 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_EXECUTED); 2200 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 2201 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_COMPLETED); 2202 } 2203 2204 /* This cleans up all memory. It is only safe to use if the rest of the software stack 2205 * has been shut down */ 2206 static void 2207 _spdk_nvmf_rdma_qp_cleanup_all_states(struct spdk_nvmf_rdma_qpair *rqpair) 2208 { 2209 _spdk_nvmf_rdma_qp_cleanup_safe_states(rqpair); 2210 2211 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_EXECUTING); 2212 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2213 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2214 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_COMPLETING); 2215 } 2216 2217 static void 2218 _spdk_nvmf_rdma_qp_error(void *arg) 2219 { 2220 struct spdk_nvmf_rdma_qpair *rqpair = arg; 2221 enum ibv_qp_state state; 2222 2223 spdk_nvmf_rdma_qpair_dec_refcnt(rqpair); 2224 2225 state = rqpair->ibv_attr.qp_state; 2226 if (state != IBV_QPS_ERR) { 2227 /* Error was already recovered */ 2228 return; 2229 } 2230 2231 if (spdk_nvmf_qpair_is_admin_queue(&rqpair->qpair)) { 2232 spdk_nvmf_ctrlr_abort_aer(rqpair->qpair.ctrlr); 2233 } 2234 2235 _spdk_nvmf_rdma_qp_cleanup_safe_states(rqpair); 2236 2237 /* Attempt recovery. This will exit without recovering if I/O requests 2238 * are still outstanding */ 2239 spdk_nvmf_rdma_qpair_recover(rqpair); 2240 } 2241 2242 static void 2243 _spdk_nvmf_rdma_qp_last_wqe(void *arg) 2244 { 2245 struct spdk_nvmf_rdma_qpair *rqpair = arg; 2246 enum ibv_qp_state state; 2247 2248 spdk_nvmf_rdma_qpair_dec_refcnt(rqpair); 2249 2250 state = rqpair->ibv_attr.qp_state; 2251 if (state != IBV_QPS_ERR) { 2252 /* Error was already recovered */ 2253 return; 2254 } 2255 2256 /* Clear out the states that are safe to clear any time, plus the 2257 * RDMA data transfer states. */ 2258 _spdk_nvmf_rdma_qp_cleanup_safe_states(rqpair); 2259 2260 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2261 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2262 spdk_nvmf_rdma_drain_state_queue(rqpair, RDMA_REQUEST_STATE_COMPLETING); 2263 2264 spdk_nvmf_rdma_qpair_recover(rqpair); 2265 } 2266 2267 static void 2268 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2269 { 2270 int rc; 2271 struct spdk_nvmf_rdma_qpair *rqpair; 2272 struct ibv_async_event event; 2273 enum ibv_qp_state state; 2274 2275 rc = ibv_get_async_event(device->context, &event); 2276 2277 if (rc) { 2278 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2279 errno, spdk_strerror(errno)); 2280 return; 2281 } 2282 2283 SPDK_NOTICELOG("Async event: %s\n", 2284 ibv_event_type_str(event.event_type)); 2285 2286 switch (event.event_type) { 2287 case IBV_EVENT_QP_FATAL: 2288 rqpair = event.element.qp->qp_context; 2289 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2290 (uintptr_t)rqpair->cm_id, event.event_type); 2291 spdk_nvmf_rdma_update_ibv_state(rqpair); 2292 spdk_nvmf_rdma_qpair_inc_refcnt(rqpair); 2293 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_error, rqpair); 2294 break; 2295 case IBV_EVENT_QP_LAST_WQE_REACHED: 2296 rqpair = event.element.qp->qp_context; 2297 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2298 (uintptr_t)rqpair->cm_id, event.event_type); 2299 spdk_nvmf_rdma_update_ibv_state(rqpair); 2300 spdk_nvmf_rdma_qpair_inc_refcnt(rqpair); 2301 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_last_wqe, rqpair); 2302 break; 2303 case IBV_EVENT_SQ_DRAINED: 2304 /* This event occurs frequently in both error and non-error states. 2305 * Check if the qpair is in an error state before sending a message. 2306 * Note that we're not on the correct thread to access the qpair, but 2307 * the operations that the below calls make all happen to be thread 2308 * safe. */ 2309 rqpair = event.element.qp->qp_context; 2310 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2311 (uintptr_t)rqpair->cm_id, event.event_type); 2312 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2313 if (state == IBV_QPS_ERR) { 2314 spdk_nvmf_rdma_qpair_inc_refcnt(rqpair); 2315 spdk_thread_send_msg(rqpair->qpair.group->thread, _spdk_nvmf_rdma_qp_error, rqpair); 2316 } 2317 break; 2318 case IBV_EVENT_QP_REQ_ERR: 2319 case IBV_EVENT_QP_ACCESS_ERR: 2320 case IBV_EVENT_COMM_EST: 2321 case IBV_EVENT_PATH_MIG: 2322 case IBV_EVENT_PATH_MIG_ERR: 2323 rqpair = event.element.qp->qp_context; 2324 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2325 (uintptr_t)rqpair->cm_id, event.event_type); 2326 spdk_nvmf_rdma_update_ibv_state(rqpair); 2327 break; 2328 case IBV_EVENT_CQ_ERR: 2329 case IBV_EVENT_DEVICE_FATAL: 2330 case IBV_EVENT_PORT_ACTIVE: 2331 case IBV_EVENT_PORT_ERR: 2332 case IBV_EVENT_LID_CHANGE: 2333 case IBV_EVENT_PKEY_CHANGE: 2334 case IBV_EVENT_SM_CHANGE: 2335 case IBV_EVENT_SRQ_ERR: 2336 case IBV_EVENT_SRQ_LIMIT_REACHED: 2337 case IBV_EVENT_CLIENT_REREGISTER: 2338 case IBV_EVENT_GID_CHANGE: 2339 default: 2340 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2341 break; 2342 } 2343 ibv_ack_async_event(&event); 2344 } 2345 2346 static void 2347 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2348 { 2349 int nfds, i = 0; 2350 struct spdk_nvmf_rdma_transport *rtransport; 2351 struct spdk_nvmf_rdma_device *device, *tmp; 2352 2353 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2354 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2355 2356 if (nfds <= 0) { 2357 return; 2358 } 2359 2360 /* The first poll descriptor is RDMA CM event */ 2361 if (rtransport->poll_fds[i++].revents & POLLIN) { 2362 spdk_nvmf_process_cm_event(transport, cb_fn); 2363 nfds--; 2364 } 2365 2366 if (nfds == 0) { 2367 return; 2368 } 2369 2370 /* Second and subsequent poll descriptors are IB async events */ 2371 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2372 if (rtransport->poll_fds[i++].revents & POLLIN) { 2373 spdk_nvmf_process_ib_event(device); 2374 nfds--; 2375 } 2376 } 2377 /* check all flagged fd's have been served */ 2378 assert(nfds == 0); 2379 } 2380 2381 static void 2382 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2383 struct spdk_nvme_transport_id *trid, 2384 struct spdk_nvmf_discovery_log_page_entry *entry) 2385 { 2386 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2387 entry->adrfam = trid->adrfam; 2388 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2389 2390 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2391 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2392 2393 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2394 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2395 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2396 } 2397 2398 static struct spdk_nvmf_transport_poll_group * 2399 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2400 { 2401 struct spdk_nvmf_rdma_transport *rtransport; 2402 struct spdk_nvmf_rdma_poll_group *rgroup; 2403 struct spdk_nvmf_rdma_poller *poller; 2404 struct spdk_nvmf_rdma_device *device; 2405 2406 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2407 2408 rgroup = calloc(1, sizeof(*rgroup)); 2409 if (!rgroup) { 2410 return NULL; 2411 } 2412 2413 TAILQ_INIT(&rgroup->pollers); 2414 2415 pthread_mutex_lock(&rtransport->lock); 2416 TAILQ_FOREACH(device, &rtransport->devices, link) { 2417 poller = calloc(1, sizeof(*poller)); 2418 if (!poller) { 2419 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2420 free(rgroup); 2421 pthread_mutex_unlock(&rtransport->lock); 2422 return NULL; 2423 } 2424 2425 poller->device = device; 2426 poller->group = rgroup; 2427 2428 TAILQ_INIT(&poller->qpairs); 2429 2430 poller->cq = ibv_create_cq(device->context, NVMF_RDMA_CQ_SIZE, poller, NULL, 0); 2431 if (!poller->cq) { 2432 SPDK_ERRLOG("Unable to create completion queue\n"); 2433 free(poller); 2434 free(rgroup); 2435 pthread_mutex_unlock(&rtransport->lock); 2436 return NULL; 2437 } 2438 2439 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2440 } 2441 2442 pthread_mutex_unlock(&rtransport->lock); 2443 return &rgroup->group; 2444 } 2445 2446 static void 2447 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2448 { 2449 struct spdk_nvmf_rdma_poll_group *rgroup; 2450 struct spdk_nvmf_rdma_poller *poller, *tmp; 2451 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2452 2453 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2454 2455 if (!rgroup) { 2456 return; 2457 } 2458 2459 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2460 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2461 2462 if (poller->cq) { 2463 ibv_destroy_cq(poller->cq); 2464 } 2465 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 2466 _spdk_nvmf_rdma_qp_cleanup_all_states(qpair); 2467 spdk_nvmf_rdma_qpair_destroy(qpair); 2468 } 2469 2470 free(poller); 2471 } 2472 2473 free(rgroup); 2474 } 2475 2476 static int 2477 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 2478 struct spdk_nvmf_qpair *qpair) 2479 { 2480 struct spdk_nvmf_rdma_transport *rtransport; 2481 struct spdk_nvmf_rdma_poll_group *rgroup; 2482 struct spdk_nvmf_rdma_qpair *rqpair; 2483 struct spdk_nvmf_rdma_device *device; 2484 struct spdk_nvmf_rdma_poller *poller; 2485 int rc; 2486 2487 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 2488 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2489 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2490 2491 device = rqpair->port->device; 2492 2493 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 2494 if (poller->device == device) { 2495 break; 2496 } 2497 } 2498 2499 if (!poller) { 2500 SPDK_ERRLOG("No poller found for device.\n"); 2501 return -1; 2502 } 2503 2504 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 2505 rqpair->poller = poller; 2506 2507 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 2508 if (rc < 0) { 2509 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 2510 return -1; 2511 } 2512 2513 rqpair->mgmt_channel = spdk_get_io_channel(rtransport); 2514 if (!rqpair->mgmt_channel) { 2515 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2516 spdk_nvmf_rdma_qpair_destroy(rqpair); 2517 return -1; 2518 } 2519 2520 rqpair->ch = spdk_io_channel_get_ctx(rqpair->mgmt_channel); 2521 assert(rqpair->ch != NULL); 2522 2523 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 2524 if (rc) { 2525 /* Try to reject, but we probably can't */ 2526 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2527 spdk_nvmf_rdma_qpair_destroy(rqpair); 2528 return -1; 2529 } 2530 2531 spdk_nvmf_rdma_update_ibv_state(rqpair); 2532 2533 return 0; 2534 } 2535 2536 static int 2537 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 2538 { 2539 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2540 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2541 struct spdk_nvmf_rdma_transport, transport); 2542 2543 if (rdma_req->data_from_pool) { 2544 /* Put the buffer/s back in the pool */ 2545 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 2546 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 2547 rdma_req->req.iov[i].iov_base = NULL; 2548 rdma_req->data.buffers[i] = NULL; 2549 } 2550 rdma_req->data_from_pool = false; 2551 } 2552 rdma_req->req.length = 0; 2553 rdma_req->req.iovcnt = 0; 2554 rdma_req->req.data = NULL; 2555 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 2556 return 0; 2557 } 2558 2559 static int 2560 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 2561 { 2562 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2563 struct spdk_nvmf_rdma_transport, transport); 2564 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 2565 struct spdk_nvmf_rdma_request, req); 2566 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 2567 struct spdk_nvmf_rdma_qpair, qpair); 2568 2569 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2570 /* The connection is alive, so process the request as normal */ 2571 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTED); 2572 } else { 2573 /* The connection is dead. Move the request directly to the completed state. */ 2574 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2575 } 2576 2577 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2578 2579 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE && rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 2580 /* If the NVMe-oF layer thinks the connection is active, but the RDMA layer thinks 2581 * the connection is dead, perform error recovery. */ 2582 spdk_nvmf_rdma_qpair_recover(rqpair); 2583 } 2584 2585 return 0; 2586 } 2587 2588 static void 2589 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 2590 { 2591 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2592 2593 spdk_nvmf_rdma_qpair_destroy(rqpair); 2594 } 2595 2596 static struct spdk_nvmf_rdma_request * 2597 get_rdma_req_from_wc(struct ibv_wc *wc) 2598 { 2599 struct spdk_nvmf_rdma_request *rdma_req; 2600 2601 rdma_req = (struct spdk_nvmf_rdma_request *)wc->wr_id; 2602 assert(rdma_req != NULL); 2603 2604 #ifdef DEBUG 2605 struct spdk_nvmf_rdma_qpair *rqpair; 2606 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2607 2608 assert(rdma_req - rqpair->reqs >= 0); 2609 assert(rdma_req - rqpair->reqs < (ptrdiff_t)rqpair->max_queue_depth); 2610 #endif 2611 2612 return rdma_req; 2613 } 2614 2615 static struct spdk_nvmf_rdma_recv * 2616 get_rdma_recv_from_wc(struct ibv_wc *wc) 2617 { 2618 struct spdk_nvmf_rdma_recv *rdma_recv; 2619 2620 assert(wc->byte_len >= sizeof(struct spdk_nvmf_capsule_cmd)); 2621 2622 rdma_recv = (struct spdk_nvmf_rdma_recv *)wc->wr_id; 2623 assert(rdma_recv != NULL); 2624 2625 #ifdef DEBUG 2626 struct spdk_nvmf_rdma_qpair *rqpair = rdma_recv->qpair; 2627 2628 assert(rdma_recv - rqpair->recvs >= 0); 2629 assert(rdma_recv - rqpair->recvs < (ptrdiff_t)rqpair->max_queue_depth); 2630 #endif 2631 2632 return rdma_recv; 2633 } 2634 2635 #ifdef DEBUG 2636 static int 2637 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 2638 { 2639 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 2640 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 2641 } 2642 #endif 2643 2644 static int 2645 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 2646 struct spdk_nvmf_rdma_poller *rpoller) 2647 { 2648 struct ibv_wc wc[32]; 2649 struct spdk_nvmf_rdma_request *rdma_req; 2650 struct spdk_nvmf_rdma_recv *rdma_recv; 2651 struct spdk_nvmf_rdma_qpair *rqpair; 2652 int reaped, i; 2653 int count = 0; 2654 bool error = false; 2655 2656 /* Poll for completing operations. */ 2657 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 2658 if (reaped < 0) { 2659 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2660 errno, spdk_strerror(errno)); 2661 return -1; 2662 } 2663 2664 for (i = 0; i < reaped; i++) { 2665 /* Handle error conditions */ 2666 if (wc[i].status) { 2667 SPDK_WARNLOG("CQ error on CQ %p, Request 0x%lu (%d): %s\n", 2668 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 2669 error = true; 2670 2671 switch (wc[i].opcode) { 2672 case IBV_WC_SEND: 2673 rdma_req = get_rdma_req_from_wc(&wc[i]); 2674 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2675 2676 /* We're going to attempt an error recovery, so force the request into 2677 * the completed state. */ 2678 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2679 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2680 break; 2681 case IBV_WC_RECV: 2682 rdma_recv = get_rdma_recv_from_wc(&wc[i]); 2683 rqpair = rdma_recv->qpair; 2684 2685 /* Dump this into the incoming queue. This gets cleaned up when 2686 * the queue pair disconnects or recovers. */ 2687 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2688 break; 2689 case IBV_WC_RDMA_WRITE: 2690 case IBV_WC_RDMA_READ: 2691 /* If the data transfer fails still force the queue into the error state, 2692 * but the rdma_req objects should only be manipulated in response to 2693 * SEND and RECV operations. */ 2694 rdma_req = get_rdma_req_from_wc(&wc[i]); 2695 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2696 break; 2697 default: 2698 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2699 continue; 2700 } 2701 2702 /* Set the qpair to the error state. This will initiate a recovery. */ 2703 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 2704 continue; 2705 } 2706 2707 switch (wc[i].opcode) { 2708 case IBV_WC_SEND: 2709 rdma_req = get_rdma_req_from_wc(&wc[i]); 2710 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2711 2712 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 2713 2714 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2715 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2716 2717 count++; 2718 2719 /* Try to process other queued requests */ 2720 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2721 break; 2722 2723 case IBV_WC_RDMA_WRITE: 2724 rdma_req = get_rdma_req_from_wc(&wc[i]); 2725 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2726 2727 /* Try to process other queued requests */ 2728 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2729 break; 2730 2731 case IBV_WC_RDMA_READ: 2732 rdma_req = get_rdma_req_from_wc(&wc[i]); 2733 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2734 2735 assert(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2736 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 2737 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2738 2739 /* Try to process other queued requests */ 2740 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2741 break; 2742 2743 case IBV_WC_RECV: 2744 rdma_recv = get_rdma_recv_from_wc(&wc[i]); 2745 rqpair = rdma_recv->qpair; 2746 2747 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2748 /* Try to process other queued requests */ 2749 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2750 break; 2751 2752 default: 2753 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2754 continue; 2755 } 2756 } 2757 2758 if (error == true) { 2759 return -1; 2760 } 2761 2762 return count; 2763 } 2764 2765 static int 2766 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 2767 { 2768 struct spdk_nvmf_rdma_transport *rtransport; 2769 struct spdk_nvmf_rdma_poll_group *rgroup; 2770 struct spdk_nvmf_rdma_poller *rpoller; 2771 int count, rc; 2772 2773 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 2774 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2775 2776 count = 0; 2777 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2778 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 2779 if (rc < 0) { 2780 return rc; 2781 } 2782 count += rc; 2783 } 2784 2785 return count; 2786 } 2787 2788 static int 2789 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2790 struct spdk_nvme_transport_id *trid, 2791 bool peer) 2792 { 2793 struct sockaddr *saddr; 2794 uint16_t port; 2795 2796 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 2797 2798 if (peer) { 2799 saddr = rdma_get_peer_addr(id); 2800 } else { 2801 saddr = rdma_get_local_addr(id); 2802 } 2803 switch (saddr->sa_family) { 2804 case AF_INET: { 2805 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 2806 2807 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 2808 inet_ntop(AF_INET, &saddr_in->sin_addr, 2809 trid->traddr, sizeof(trid->traddr)); 2810 if (peer) { 2811 port = ntohs(rdma_get_dst_port(id)); 2812 } else { 2813 port = ntohs(rdma_get_src_port(id)); 2814 } 2815 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 2816 break; 2817 } 2818 case AF_INET6: { 2819 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 2820 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 2821 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 2822 trid->traddr, sizeof(trid->traddr)); 2823 if (peer) { 2824 port = ntohs(rdma_get_dst_port(id)); 2825 } else { 2826 port = ntohs(rdma_get_src_port(id)); 2827 } 2828 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 2829 break; 2830 } 2831 default: 2832 return -1; 2833 2834 } 2835 2836 return 0; 2837 } 2838 2839 static int 2840 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 2841 struct spdk_nvme_transport_id *trid) 2842 { 2843 struct spdk_nvmf_rdma_qpair *rqpair; 2844 2845 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2846 2847 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 2848 } 2849 2850 static int 2851 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 2852 struct spdk_nvme_transport_id *trid) 2853 { 2854 struct spdk_nvmf_rdma_qpair *rqpair; 2855 2856 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2857 2858 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 2859 } 2860 2861 static int 2862 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 2863 struct spdk_nvme_transport_id *trid) 2864 { 2865 struct spdk_nvmf_rdma_qpair *rqpair; 2866 2867 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2868 2869 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 2870 } 2871 2872 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 2873 .type = SPDK_NVME_TRANSPORT_RDMA, 2874 .opts_init = spdk_nvmf_rdma_opts_init, 2875 .create = spdk_nvmf_rdma_create, 2876 .destroy = spdk_nvmf_rdma_destroy, 2877 2878 .listen = spdk_nvmf_rdma_listen, 2879 .stop_listen = spdk_nvmf_rdma_stop_listen, 2880 .accept = spdk_nvmf_rdma_accept, 2881 2882 .listener_discover = spdk_nvmf_rdma_discover, 2883 2884 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 2885 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 2886 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 2887 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 2888 2889 .req_free = spdk_nvmf_rdma_request_free, 2890 .req_complete = spdk_nvmf_rdma_request_complete, 2891 2892 .qpair_fini = spdk_nvmf_rdma_close_qpair, 2893 .qpair_is_idle = spdk_nvmf_rdma_qpair_is_idle, 2894 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 2895 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 2896 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 2897 2898 }; 2899 2900 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 2901