1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2018 Mellanox Technologies LTD. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of Intel Corporation nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include "spdk/stdinc.h" 35 36 #include <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/config.h" 44 #include "spdk/assert.h" 45 #include "spdk/thread.h" 46 #include "spdk/nvmf.h" 47 #include "spdk/nvmf_spec.h" 48 #include "spdk/string.h" 49 #include "spdk/trace.h" 50 #include "spdk/util.h" 51 52 #include "spdk_internal/log.h" 53 54 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 55 56 /* 57 RDMA Connection Resource Defaults 58 */ 59 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 60 #define NVMF_DEFAULT_RSP_SGE 1 61 #define NVMF_DEFAULT_RX_SGE 2 62 63 /* The RDMA completion queue size */ 64 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 65 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 66 67 enum spdk_nvmf_rdma_request_state { 68 /* The request is not currently in use */ 69 RDMA_REQUEST_STATE_FREE = 0, 70 71 /* Initial state when request first received */ 72 RDMA_REQUEST_STATE_NEW, 73 74 /* The request is queued until a data buffer is available. */ 75 RDMA_REQUEST_STATE_NEED_BUFFER, 76 77 /* The request is waiting on RDMA queue depth availability 78 * to transfer data from the host to the controller. 79 */ 80 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 81 82 /* The request is currently transferring data from the host to the controller. */ 83 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 84 85 /* The request is ready to execute at the block device */ 86 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 87 88 /* The request is currently executing at the block device */ 89 RDMA_REQUEST_STATE_EXECUTING, 90 91 /* The request finished executing at the block device */ 92 RDMA_REQUEST_STATE_EXECUTED, 93 94 /* The request is waiting on RDMA queue depth availability 95 * to transfer data from the controller to the host. 96 */ 97 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 98 99 /* The request is ready to send a completion */ 100 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 101 102 /* The request is currently transferring data from the controller to the host. */ 103 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 104 105 /* The request currently has an outstanding completion without an 106 * associated data transfer. 107 */ 108 RDMA_REQUEST_STATE_COMPLETING, 109 110 /* The request completed and can be marked free. */ 111 RDMA_REQUEST_STATE_COMPLETED, 112 113 /* Terminator */ 114 RDMA_REQUEST_NUM_STATES, 115 }; 116 117 #define OBJECT_NVMF_RDMA_IO 0x40 118 119 #define TRACE_GROUP_NVMF_RDMA 0x4 120 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 121 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 122 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 123 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 124 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 125 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 126 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 127 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 128 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 129 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 130 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 131 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 132 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 133 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 134 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 135 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 136 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 137 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 138 139 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 140 { 141 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 142 spdk_trace_register_description("RDMA_REQ_NEW", "", 143 TRACE_RDMA_REQUEST_STATE_NEW, 144 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 145 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", "", 146 TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 147 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 148 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C_TO_H", "", 149 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 150 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 151 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H_TO_C", "", 152 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 153 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 154 spdk_trace_register_description("RDMA_REQ_TX_H_TO_C", "", 155 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 156 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 157 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", "", 158 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 159 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 160 spdk_trace_register_description("RDMA_REQ_EXECUTING", "", 161 TRACE_RDMA_REQUEST_STATE_EXECUTING, 162 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 163 spdk_trace_register_description("RDMA_REQ_EXECUTED", "", 164 TRACE_RDMA_REQUEST_STATE_EXECUTED, 165 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 166 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPLETE", "", 167 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 168 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 169 spdk_trace_register_description("RDMA_REQ_COMPLETING_CONTROLLER_TO_HOST", "", 170 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 171 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 172 spdk_trace_register_description("RDMA_REQ_COMPLETING_INCAPSULE", "", 173 TRACE_RDMA_REQUEST_STATE_COMPLETING, 174 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 175 spdk_trace_register_description("RDMA_REQ_COMPLETED", "", 176 TRACE_RDMA_REQUEST_STATE_COMPLETED, 177 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 178 179 spdk_trace_register_description("RDMA_QP_CREATE", "", TRACE_RDMA_QP_CREATE, 180 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 181 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", "", TRACE_RDMA_IBV_ASYNC_EVENT, 182 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 183 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", "", TRACE_RDMA_CM_ASYNC_EVENT, 184 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 185 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", "", TRACE_RDMA_QP_STATE_CHANGE, 186 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 187 spdk_trace_register_description("RDMA_QP_DISCONNECT", "", TRACE_RDMA_QP_DISCONNECT, 188 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 189 spdk_trace_register_description("RDMA_QP_DESTROY", "", TRACE_RDMA_QP_DESTROY, 190 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 191 } 192 193 enum spdk_nvmf_rdma_wr_type { 194 RDMA_WR_TYPE_RECV, 195 RDMA_WR_TYPE_SEND, 196 RDMA_WR_TYPE_DATA, 197 RDMA_WR_TYPE_DRAIN_SEND, 198 RDMA_WR_TYPE_DRAIN_RECV 199 }; 200 201 struct spdk_nvmf_rdma_wr { 202 enum spdk_nvmf_rdma_wr_type type; 203 }; 204 205 /* This structure holds commands as they are received off the wire. 206 * It must be dynamically paired with a full request object 207 * (spdk_nvmf_rdma_request) to service a request. It is separate 208 * from the request because RDMA does not appear to order 209 * completions, so occasionally we'll get a new incoming 210 * command when there aren't any free request objects. 211 */ 212 struct spdk_nvmf_rdma_recv { 213 struct ibv_recv_wr wr; 214 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 215 216 struct spdk_nvmf_rdma_qpair *qpair; 217 218 /* In-capsule data buffer */ 219 uint8_t *buf; 220 221 struct spdk_nvmf_rdma_wr rdma_wr; 222 223 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 224 }; 225 226 struct spdk_nvmf_rdma_request_data { 227 struct spdk_nvmf_rdma_wr rdma_wr; 228 struct ibv_send_wr wr; 229 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 230 void *buffers[SPDK_NVMF_MAX_SGL_ENTRIES]; 231 }; 232 233 struct spdk_nvmf_rdma_request { 234 struct spdk_nvmf_request req; 235 bool data_from_pool; 236 237 enum spdk_nvmf_rdma_request_state state; 238 239 struct spdk_nvmf_rdma_recv *recv; 240 241 struct { 242 struct spdk_nvmf_rdma_wr rdma_wr; 243 struct ibv_send_wr wr; 244 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 245 } rsp; 246 247 struct spdk_nvmf_rdma_request_data data; 248 249 uint32_t num_outstanding_data_wr; 250 251 TAILQ_ENTRY(spdk_nvmf_rdma_request) link; 252 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 253 }; 254 255 enum spdk_nvmf_rdma_qpair_disconnect_flags { 256 RDMA_QP_DISCONNECTING = 1, 257 RDMA_QP_RECV_DRAINED = 1 << 1, 258 RDMA_QP_SEND_DRAINED = 1 << 2 259 }; 260 261 struct spdk_nvmf_rdma_qpair { 262 struct spdk_nvmf_qpair qpair; 263 264 struct spdk_nvmf_rdma_port *port; 265 struct spdk_nvmf_rdma_poller *poller; 266 267 struct rdma_cm_id *cm_id; 268 struct rdma_cm_id *listen_id; 269 270 /* The maximum number of I/O outstanding on this connection at one time */ 271 uint16_t max_queue_depth; 272 273 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 274 uint16_t max_read_depth; 275 276 /* The maximum number of RDMA SEND operations at one time */ 277 uint32_t max_send_depth; 278 279 /* The current number of outstanding WRs from this qpair's 280 * recv queue. Should not exceed device->attr.max_queue_depth. 281 */ 282 uint16_t current_recv_depth; 283 284 /* The current number of posted WRs from this qpair's 285 * send queue. Should not exceed max_send_depth. 286 */ 287 uint32_t current_send_depth; 288 289 /* The current number of active RDMA READ operations */ 290 uint16_t current_read_depth; 291 292 /* The maximum number of SGEs per WR on the send queue */ 293 uint32_t max_send_sge; 294 295 /* The maximum number of SGEs per WR on the recv queue */ 296 uint32_t max_recv_sge; 297 298 /* Receives that are waiting for a request object */ 299 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 300 301 /* Queues to track requests in critical states */ 302 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 303 304 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 305 306 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 307 308 /* Number of requests not in the free state */ 309 uint32_t qd; 310 311 /* Array of size "max_queue_depth" containing RDMA requests. */ 312 struct spdk_nvmf_rdma_request *reqs; 313 314 /* Array of size "max_queue_depth" containing RDMA recvs. */ 315 struct spdk_nvmf_rdma_recv *recvs; 316 317 /* Array of size "max_queue_depth" containing 64 byte capsules 318 * used for receive. 319 */ 320 union nvmf_h2c_msg *cmds; 321 struct ibv_mr *cmds_mr; 322 323 /* Array of size "max_queue_depth" containing 16 byte completions 324 * to be sent back to the user. 325 */ 326 union nvmf_c2h_msg *cpls; 327 struct ibv_mr *cpls_mr; 328 329 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 330 * buffers to be used for in capsule data. 331 */ 332 void *bufs; 333 struct ibv_mr *bufs_mr; 334 335 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 336 337 /* IBV queue pair attributes: they are used to manage 338 * qp state and recover from errors. 339 */ 340 struct ibv_qp_attr ibv_attr; 341 342 uint32_t disconnect_flags; 343 struct spdk_nvmf_rdma_wr drain_send_wr; 344 struct spdk_nvmf_rdma_wr drain_recv_wr; 345 346 /* There are several ways a disconnect can start on a qpair 347 * and they are not all mutually exclusive. It is important 348 * that we only initialize one of these paths. 349 */ 350 bool disconnect_started; 351 }; 352 353 struct spdk_nvmf_rdma_poller { 354 struct spdk_nvmf_rdma_device *device; 355 struct spdk_nvmf_rdma_poll_group *group; 356 357 int num_cqe; 358 int required_num_wr; 359 struct ibv_cq *cq; 360 361 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 362 363 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 364 }; 365 366 struct spdk_nvmf_rdma_poll_group { 367 struct spdk_nvmf_transport_poll_group group; 368 369 /* Requests that are waiting to obtain a data buffer */ 370 TAILQ_HEAD(, spdk_nvmf_rdma_request) pending_data_buf_queue; 371 372 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 373 }; 374 375 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 376 struct spdk_nvmf_rdma_device { 377 struct ibv_device_attr attr; 378 struct ibv_context *context; 379 380 struct spdk_mem_map *map; 381 struct ibv_pd *pd; 382 383 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 384 }; 385 386 struct spdk_nvmf_rdma_port { 387 struct spdk_nvme_transport_id trid; 388 struct rdma_cm_id *id; 389 struct spdk_nvmf_rdma_device *device; 390 uint32_t ref; 391 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 392 }; 393 394 struct spdk_nvmf_rdma_transport { 395 struct spdk_nvmf_transport transport; 396 397 struct rdma_event_channel *event_channel; 398 399 struct spdk_mempool *data_wr_pool; 400 401 pthread_mutex_t lock; 402 403 /* fields used to poll RDMA/IB events */ 404 nfds_t npoll_fds; 405 struct pollfd *poll_fds; 406 407 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 408 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 409 }; 410 411 static inline int 412 spdk_nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 413 { 414 switch (state) { 415 case IBV_QPS_RESET: 416 case IBV_QPS_INIT: 417 case IBV_QPS_RTR: 418 case IBV_QPS_RTS: 419 case IBV_QPS_SQD: 420 case IBV_QPS_SQE: 421 case IBV_QPS_ERR: 422 return 0; 423 default: 424 return -1; 425 } 426 } 427 428 static enum ibv_qp_state 429 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 430 enum ibv_qp_state old_state, new_state; 431 struct ibv_qp_init_attr init_attr; 432 int rc; 433 434 /* All the attributes needed for recovery */ 435 static int spdk_nvmf_ibv_attr_mask = 436 IBV_QP_STATE | 437 IBV_QP_PKEY_INDEX | 438 IBV_QP_PORT | 439 IBV_QP_ACCESS_FLAGS | 440 IBV_QP_AV | 441 IBV_QP_PATH_MTU | 442 IBV_QP_DEST_QPN | 443 IBV_QP_RQ_PSN | 444 IBV_QP_MAX_DEST_RD_ATOMIC | 445 IBV_QP_MIN_RNR_TIMER | 446 IBV_QP_SQ_PSN | 447 IBV_QP_TIMEOUT | 448 IBV_QP_RETRY_CNT | 449 IBV_QP_RNR_RETRY | 450 IBV_QP_MAX_QP_RD_ATOMIC; 451 452 old_state = rqpair->ibv_attr.qp_state; 453 rc = ibv_query_qp(rqpair->cm_id->qp, &rqpair->ibv_attr, 454 spdk_nvmf_ibv_attr_mask, &init_attr); 455 456 if (rc) 457 { 458 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 459 assert(false); 460 } 461 462 new_state = rqpair->ibv_attr.qp_state; 463 464 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 465 if (rc) 466 { 467 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 468 /* 469 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 470 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 471 */ 472 return IBV_QPS_ERR + 1; 473 } 474 475 if (old_state != new_state) 476 { 477 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 478 (uintptr_t)rqpair->cm_id, new_state); 479 } 480 return new_state; 481 } 482 483 static const char *str_ibv_qp_state[] = { 484 "IBV_QPS_RESET", 485 "IBV_QPS_INIT", 486 "IBV_QPS_RTR", 487 "IBV_QPS_RTS", 488 "IBV_QPS_SQD", 489 "IBV_QPS_SQE", 490 "IBV_QPS_ERR", 491 "IBV_QPS_UNKNOWN" 492 }; 493 494 static int 495 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 496 enum ibv_qp_state new_state) 497 { 498 int rc; 499 enum ibv_qp_state state; 500 static int attr_mask_rc[] = { 501 [IBV_QPS_RESET] = IBV_QP_STATE, 502 [IBV_QPS_INIT] = (IBV_QP_STATE | 503 IBV_QP_PKEY_INDEX | 504 IBV_QP_PORT | 505 IBV_QP_ACCESS_FLAGS), 506 [IBV_QPS_RTR] = (IBV_QP_STATE | 507 IBV_QP_AV | 508 IBV_QP_PATH_MTU | 509 IBV_QP_DEST_QPN | 510 IBV_QP_RQ_PSN | 511 IBV_QP_MAX_DEST_RD_ATOMIC | 512 IBV_QP_MIN_RNR_TIMER), 513 [IBV_QPS_RTS] = (IBV_QP_STATE | 514 IBV_QP_SQ_PSN | 515 IBV_QP_TIMEOUT | 516 IBV_QP_RETRY_CNT | 517 IBV_QP_RNR_RETRY | 518 IBV_QP_MAX_QP_RD_ATOMIC), 519 [IBV_QPS_SQD] = IBV_QP_STATE, 520 [IBV_QPS_SQE] = IBV_QP_STATE, 521 [IBV_QPS_ERR] = IBV_QP_STATE, 522 }; 523 524 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 525 if (rc) { 526 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 527 rqpair->qpair.qid, new_state); 528 return rc; 529 } 530 531 rqpair->ibv_attr.cur_qp_state = rqpair->ibv_attr.qp_state; 532 rqpair->ibv_attr.qp_state = new_state; 533 rqpair->ibv_attr.ah_attr.port_num = rqpair->ibv_attr.port_num; 534 535 rc = ibv_modify_qp(rqpair->cm_id->qp, &rqpair->ibv_attr, 536 attr_mask_rc[new_state]); 537 538 if (rc) { 539 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 540 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 541 return rc; 542 } 543 544 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 545 546 if (state != new_state) { 547 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 548 rqpair->qpair.qid, str_ibv_qp_state[new_state], 549 str_ibv_qp_state[state]); 550 return -1; 551 } 552 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 553 str_ibv_qp_state[state]); 554 return 0; 555 } 556 557 static void 558 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 559 { 560 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->data_from_pool); 561 if (req->req.cmd) { 562 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 563 } 564 if (req->recv) { 565 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 566 } 567 } 568 569 static void 570 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 571 { 572 int i; 573 574 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 575 for (i = 0; i < rqpair->max_queue_depth; i++) { 576 if (rqpair->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 577 nvmf_rdma_dump_request(&rqpair->reqs[i]); 578 } 579 } 580 } 581 582 static void 583 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 584 { 585 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 586 587 if (rqpair->qd != 0) { 588 nvmf_rdma_dump_qpair_contents(rqpair); 589 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 590 } 591 592 if (rqpair->poller) { 593 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 594 } 595 596 if (rqpair->cmds_mr) { 597 ibv_dereg_mr(rqpair->cmds_mr); 598 } 599 600 if (rqpair->cpls_mr) { 601 ibv_dereg_mr(rqpair->cpls_mr); 602 } 603 604 if (rqpair->bufs_mr) { 605 ibv_dereg_mr(rqpair->bufs_mr); 606 } 607 608 if (rqpair->cm_id) { 609 rdma_destroy_qp(rqpair->cm_id); 610 rdma_destroy_id(rqpair->cm_id); 611 612 if (rqpair->poller) { 613 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 614 } 615 } 616 617 /* Free all memory */ 618 spdk_dma_free(rqpair->cmds); 619 spdk_dma_free(rqpair->cpls); 620 spdk_dma_free(rqpair->bufs); 621 free(rqpair->reqs); 622 free(rqpair->recvs); 623 free(rqpair); 624 } 625 626 static int 627 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 628 { 629 struct spdk_nvmf_rdma_transport *rtransport; 630 struct spdk_nvmf_rdma_qpair *rqpair; 631 struct spdk_nvmf_rdma_poller *rpoller; 632 int rc, i, num_cqe, required_num_wr;; 633 struct spdk_nvmf_rdma_recv *rdma_recv; 634 struct spdk_nvmf_rdma_request *rdma_req; 635 struct spdk_nvmf_transport *transport; 636 struct spdk_nvmf_rdma_device *device; 637 struct ibv_qp_init_attr ibv_init_attr; 638 639 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 640 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 641 transport = &rtransport->transport; 642 device = rqpair->port->device; 643 644 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 645 ibv_init_attr.qp_context = rqpair; 646 ibv_init_attr.qp_type = IBV_QPT_RC; 647 ibv_init_attr.send_cq = rqpair->poller->cq; 648 ibv_init_attr.recv_cq = rqpair->poller->cq; 649 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 650 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 651 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 652 1; /* RECV operations + dummy drain WR */ 653 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 654 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 655 656 /* Enlarge CQ size dynamically */ 657 rpoller = rqpair->poller; 658 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 659 num_cqe = rpoller->num_cqe; 660 if (num_cqe < required_num_wr) { 661 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 662 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 663 } 664 665 if (rpoller->num_cqe != num_cqe) { 666 if (required_num_wr > device->attr.max_cqe) { 667 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 668 required_num_wr, device->attr.max_cqe); 669 rdma_destroy_id(rqpair->cm_id); 670 rqpair->cm_id = NULL; 671 spdk_nvmf_rdma_qpair_destroy(rqpair); 672 return -1; 673 } 674 675 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 676 rc = ibv_resize_cq(rpoller->cq, num_cqe); 677 if (rc) { 678 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 679 rdma_destroy_id(rqpair->cm_id); 680 rqpair->cm_id = NULL; 681 spdk_nvmf_rdma_qpair_destroy(rqpair); 682 return -1; 683 } 684 685 rpoller->num_cqe = num_cqe; 686 } 687 688 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 689 if (rc) { 690 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 691 rdma_destroy_id(rqpair->cm_id); 692 rqpair->cm_id = NULL; 693 spdk_nvmf_rdma_qpair_destroy(rqpair); 694 return -1; 695 } 696 697 rpoller->required_num_wr = required_num_wr; 698 699 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2 + 1), 700 ibv_init_attr.cap.max_send_wr); 701 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 702 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 703 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 704 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 705 706 rqpair->reqs = calloc(rqpair->max_queue_depth, sizeof(*rqpair->reqs)); 707 rqpair->recvs = calloc(rqpair->max_queue_depth, sizeof(*rqpair->recvs)); 708 rqpair->cmds = spdk_dma_zmalloc(rqpair->max_queue_depth * sizeof(*rqpair->cmds), 709 0x1000, NULL); 710 rqpair->cpls = spdk_dma_zmalloc(rqpair->max_queue_depth * sizeof(*rqpair->cpls), 711 0x1000, NULL); 712 713 714 if (transport->opts.in_capsule_data_size > 0) { 715 rqpair->bufs = spdk_dma_zmalloc(rqpair->max_queue_depth * 716 transport->opts.in_capsule_data_size, 717 0x1000, NULL); 718 } 719 720 if (!rqpair->reqs || !rqpair->recvs || !rqpair->cmds || 721 !rqpair->cpls || (transport->opts.in_capsule_data_size && !rqpair->bufs)) { 722 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 723 spdk_nvmf_rdma_qpair_destroy(rqpair); 724 return -1; 725 } 726 727 rqpair->cmds_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->cmds, 728 rqpair->max_queue_depth * sizeof(*rqpair->cmds), 729 IBV_ACCESS_LOCAL_WRITE); 730 rqpair->cpls_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->cpls, 731 rqpair->max_queue_depth * sizeof(*rqpair->cpls), 732 0); 733 734 if (transport->opts.in_capsule_data_size) { 735 rqpair->bufs_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->bufs, 736 rqpair->max_queue_depth * 737 transport->opts.in_capsule_data_size, 738 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 739 } 740 741 if (!rqpair->cmds_mr || !rqpair->cpls_mr || (transport->opts.in_capsule_data_size && 742 !rqpair->bufs_mr)) { 743 SPDK_ERRLOG("Unable to register required memory for RDMA queue.\n"); 744 spdk_nvmf_rdma_qpair_destroy(rqpair); 745 return -1; 746 } 747 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 748 rqpair->cmds, rqpair->max_queue_depth * sizeof(*rqpair->cmds), rqpair->cmds_mr->lkey); 749 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 750 rqpair->cpls, rqpair->max_queue_depth * sizeof(*rqpair->cpls), rqpair->cpls_mr->lkey); 751 if (rqpair->bufs && rqpair->bufs_mr) { 752 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 753 rqpair->bufs, rqpair->max_queue_depth * 754 transport->opts.in_capsule_data_size, rqpair->bufs_mr->lkey); 755 } 756 757 STAILQ_INIT(&rqpair->free_queue); 758 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 759 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 760 761 rqpair->current_recv_depth = rqpair->max_queue_depth; 762 for (i = 0; i < rqpair->max_queue_depth; i++) { 763 struct ibv_recv_wr *bad_wr = NULL; 764 765 rdma_recv = &rqpair->recvs[i]; 766 rdma_recv->qpair = rqpair; 767 768 /* Set up memory to receive commands */ 769 if (rqpair->bufs) { 770 rdma_recv->buf = (void *)((uintptr_t)rqpair->bufs + (i * 771 transport->opts.in_capsule_data_size)); 772 } 773 774 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 775 776 rdma_recv->sgl[0].addr = (uintptr_t)&rqpair->cmds[i]; 777 rdma_recv->sgl[0].length = sizeof(rqpair->cmds[i]); 778 rdma_recv->sgl[0].lkey = rqpair->cmds_mr->lkey; 779 rdma_recv->wr.num_sge = 1; 780 781 if (rdma_recv->buf && rqpair->bufs_mr) { 782 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 783 rdma_recv->sgl[1].length = transport->opts.in_capsule_data_size; 784 rdma_recv->sgl[1].lkey = rqpair->bufs_mr->lkey; 785 rdma_recv->wr.num_sge++; 786 } 787 788 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 789 rdma_recv->wr.sg_list = rdma_recv->sgl; 790 791 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_recv->wr, &bad_wr); 792 assert(rqpair->current_recv_depth > 0); 793 rqpair->current_recv_depth--; 794 if (rc) { 795 SPDK_ERRLOG("Unable to post capsule for RDMA RECV\n"); 796 spdk_nvmf_rdma_qpair_destroy(rqpair); 797 return -1; 798 } 799 } 800 assert(rqpair->current_recv_depth == 0); 801 802 for (i = 0; i < rqpair->max_queue_depth; i++) { 803 rdma_req = &rqpair->reqs[i]; 804 805 rdma_req->req.qpair = &rqpair->qpair; 806 rdma_req->req.cmd = NULL; 807 808 /* Set up memory to send responses */ 809 rdma_req->req.rsp = &rqpair->cpls[i]; 810 811 rdma_req->rsp.sgl[0].addr = (uintptr_t)&rqpair->cpls[i]; 812 rdma_req->rsp.sgl[0].length = sizeof(rqpair->cpls[i]); 813 rdma_req->rsp.sgl[0].lkey = rqpair->cpls_mr->lkey; 814 815 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 816 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 817 rdma_req->rsp.wr.next = NULL; 818 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 819 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 820 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 821 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 822 823 /* Set up memory for data buffers */ 824 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 825 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 826 rdma_req->data.wr.next = NULL; 827 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 828 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 829 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 830 831 /* Initialize request state to FREE */ 832 rdma_req->state = RDMA_REQUEST_STATE_FREE; 833 STAILQ_INSERT_HEAD(&rqpair->free_queue, rdma_req, state_link); 834 } 835 836 return 0; 837 } 838 839 static int 840 request_transfer_in(struct spdk_nvmf_request *req) 841 { 842 int rc; 843 struct spdk_nvmf_rdma_request *rdma_req; 844 struct spdk_nvmf_qpair *qpair; 845 struct spdk_nvmf_rdma_qpair *rqpair; 846 struct ibv_send_wr *bad_wr = NULL; 847 848 qpair = req->qpair; 849 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 850 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 851 852 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 853 assert(rdma_req != NULL); 854 855 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA READ POSTED. Request: %p Connection: %p\n", req, qpair); 856 857 rc = ibv_post_send(rqpair->cm_id->qp, &rdma_req->data.wr, &bad_wr); 858 if (rc) { 859 SPDK_ERRLOG("Unable to transfer data from host to target\n"); 860 return -1; 861 } 862 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 863 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 864 return 0; 865 } 866 867 static int 868 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 869 { 870 int rc; 871 struct spdk_nvmf_rdma_request *rdma_req; 872 struct spdk_nvmf_qpair *qpair; 873 struct spdk_nvmf_rdma_qpair *rqpair; 874 struct spdk_nvme_cpl *rsp; 875 struct ibv_recv_wr *bad_recv_wr = NULL; 876 struct ibv_send_wr *send_wr, *bad_send_wr = NULL; 877 878 *data_posted = 0; 879 qpair = req->qpair; 880 rsp = &req->rsp->nvme_cpl; 881 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 882 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 883 884 /* Advance our sq_head pointer */ 885 if (qpair->sq_head == qpair->sq_head_max) { 886 qpair->sq_head = 0; 887 } else { 888 qpair->sq_head++; 889 } 890 rsp->sqhd = qpair->sq_head; 891 892 /* Post the capsule to the recv buffer */ 893 assert(rdma_req->recv != NULL); 894 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA RECV POSTED. Recv: %p Connection: %p\n", rdma_req->recv, 895 rqpair); 896 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_req->recv->wr, &bad_recv_wr); 897 if (rc) { 898 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 899 return rc; 900 } 901 rdma_req->recv = NULL; 902 assert(rqpair->current_recv_depth > 0); 903 rqpair->current_recv_depth--; 904 905 /* Build the response which consists of an optional 906 * RDMA WRITE to transfer data, plus an RDMA SEND 907 * containing the response. 908 */ 909 send_wr = &rdma_req->rsp.wr; 910 911 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 912 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 913 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA WRITE POSTED. Request: %p Connection: %p\n", req, qpair); 914 send_wr = &rdma_req->data.wr; 915 *data_posted = 1; 916 } 917 918 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA SEND POSTED. Request: %p Connection: %p\n", req, qpair); 919 920 /* Send the completion */ 921 rc = ibv_post_send(rqpair->cm_id->qp, send_wr, &bad_send_wr); 922 if (rc) { 923 SPDK_ERRLOG("Unable to send response capsule\n"); 924 return rc; 925 } 926 /* +1 for the rsp wr */ 927 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr + 1; 928 929 return 0; 930 } 931 932 static int 933 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 934 { 935 struct spdk_nvmf_rdma_accept_private_data accept_data; 936 struct rdma_conn_param ctrlr_event_data = {}; 937 int rc; 938 939 accept_data.recfmt = 0; 940 accept_data.crqsize = rqpair->max_queue_depth; 941 942 ctrlr_event_data.private_data = &accept_data; 943 ctrlr_event_data.private_data_len = sizeof(accept_data); 944 if (id->ps == RDMA_PS_TCP) { 945 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 946 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 947 } 948 949 rc = rdma_accept(id, &ctrlr_event_data); 950 if (rc) { 951 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 952 } else { 953 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 954 } 955 956 return rc; 957 } 958 959 static void 960 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 961 { 962 struct spdk_nvmf_rdma_reject_private_data rej_data; 963 964 rej_data.recfmt = 0; 965 rej_data.sts = error; 966 967 rdma_reject(id, &rej_data, sizeof(rej_data)); 968 } 969 970 static int 971 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 972 new_qpair_fn cb_fn) 973 { 974 struct spdk_nvmf_rdma_transport *rtransport; 975 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 976 struct spdk_nvmf_rdma_port *port; 977 struct rdma_conn_param *rdma_param = NULL; 978 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 979 uint16_t max_queue_depth; 980 uint16_t max_read_depth; 981 982 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 983 984 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 985 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 986 987 rdma_param = &event->param.conn; 988 if (rdma_param->private_data == NULL || 989 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 990 SPDK_ERRLOG("connect request: no private data provided\n"); 991 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 992 return -1; 993 } 994 995 private_data = rdma_param->private_data; 996 if (private_data->recfmt != 0) { 997 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 998 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 999 return -1; 1000 } 1001 1002 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1003 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1004 1005 port = event->listen_id->context; 1006 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1007 event->listen_id, event->listen_id->verbs, port); 1008 1009 /* Figure out the supported queue depth. This is a multi-step process 1010 * that takes into account hardware maximums, host provided values, 1011 * and our target's internal memory limits */ 1012 1013 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1014 1015 /* Start with the maximum queue depth allowed by the target */ 1016 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1017 max_read_depth = rtransport->transport.opts.max_queue_depth; 1018 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1019 rtransport->transport.opts.max_queue_depth); 1020 1021 /* Next check the local NIC's hardware limitations */ 1022 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1023 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1024 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1025 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1026 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1027 1028 /* Next check the remote NIC's hardware limitations */ 1029 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1030 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1031 rdma_param->initiator_depth, rdma_param->responder_resources); 1032 if (rdma_param->initiator_depth > 0) { 1033 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1034 } 1035 1036 /* Finally check for the host software requested values, which are 1037 * optional. */ 1038 if (rdma_param->private_data != NULL && 1039 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1040 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1041 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1042 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1043 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1044 } 1045 1046 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1047 max_queue_depth, max_read_depth); 1048 1049 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1050 if (rqpair == NULL) { 1051 SPDK_ERRLOG("Could not allocate new connection.\n"); 1052 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1053 return -1; 1054 } 1055 1056 rqpair->port = port; 1057 rqpair->max_queue_depth = max_queue_depth; 1058 rqpair->max_read_depth = max_read_depth; 1059 rqpair->cm_id = event->id; 1060 rqpair->listen_id = event->listen_id; 1061 rqpair->qpair.transport = transport; 1062 STAILQ_INIT(&rqpair->incoming_queue); 1063 event->id->context = &rqpair->qpair; 1064 1065 cb_fn(&rqpair->qpair); 1066 1067 return 0; 1068 } 1069 1070 static int 1071 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1072 enum spdk_mem_map_notify_action action, 1073 void *vaddr, size_t size) 1074 { 1075 struct ibv_pd *pd = cb_ctx; 1076 struct ibv_mr *mr; 1077 1078 switch (action) { 1079 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1080 if (!g_nvmf_hooks.get_rkey) { 1081 mr = ibv_reg_mr(pd, vaddr, size, 1082 IBV_ACCESS_LOCAL_WRITE | 1083 IBV_ACCESS_REMOTE_READ | 1084 IBV_ACCESS_REMOTE_WRITE); 1085 if (mr == NULL) { 1086 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1087 return -1; 1088 } else { 1089 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1090 } 1091 } else { 1092 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1093 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1094 } 1095 break; 1096 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1097 if (!g_nvmf_hooks.get_rkey) { 1098 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1099 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1100 if (mr) { 1101 ibv_dereg_mr(mr); 1102 } 1103 } 1104 break; 1105 } 1106 1107 return 0; 1108 } 1109 1110 static int 1111 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1112 { 1113 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1114 return addr_1 == addr_2; 1115 } 1116 1117 static void 1118 spdk_nvmf_rdma_request_free_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1119 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport) 1120 { 1121 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 1122 if (group->buf_cache_count < group->buf_cache_size) { 1123 STAILQ_INSERT_HEAD(&group->buf_cache, 1124 (struct spdk_nvmf_transport_pg_cache_buf *)rdma_req->data.buffers[i], link); 1125 group->buf_cache_count++; 1126 } else { 1127 spdk_mempool_put(transport->data_buf_pool, rdma_req->data.buffers[i]); 1128 } 1129 rdma_req->req.iov[i].iov_base = NULL; 1130 rdma_req->data.buffers[i] = NULL; 1131 rdma_req->req.iov[i].iov_len = 0; 1132 1133 } 1134 rdma_req->data_from_pool = false; 1135 } 1136 1137 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1138 1139 static spdk_nvme_data_transfer_t 1140 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1141 { 1142 enum spdk_nvme_data_transfer xfer; 1143 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1144 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1145 1146 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1147 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1148 rdma_req->rsp.wr.imm_data = 0; 1149 #endif 1150 1151 /* Figure out data transfer direction */ 1152 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1153 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1154 } else { 1155 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1156 1157 /* Some admin commands are special cases */ 1158 if ((rdma_req->req.qpair->qid == 0) && 1159 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1160 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1161 switch (cmd->cdw10 & 0xff) { 1162 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1163 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1164 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1165 break; 1166 default: 1167 xfer = SPDK_NVME_DATA_NONE; 1168 } 1169 } 1170 } 1171 1172 if (xfer == SPDK_NVME_DATA_NONE) { 1173 return xfer; 1174 } 1175 1176 /* Even for commands that may transfer data, they could have specified 0 length. 1177 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1178 */ 1179 switch (sgl->generic.type) { 1180 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1181 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1182 case SPDK_NVME_SGL_TYPE_SEGMENT: 1183 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1184 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1185 if (sgl->unkeyed.length == 0) { 1186 xfer = SPDK_NVME_DATA_NONE; 1187 } 1188 break; 1189 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1190 if (sgl->keyed.length == 0) { 1191 xfer = SPDK_NVME_DATA_NONE; 1192 } 1193 break; 1194 } 1195 1196 return xfer; 1197 } 1198 1199 static int 1200 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1201 struct spdk_nvmf_rdma_device *device, 1202 struct spdk_nvmf_rdma_request *rdma_req) 1203 { 1204 struct spdk_nvmf_rdma_qpair *rqpair; 1205 struct spdk_nvmf_rdma_poll_group *rgroup; 1206 void *buf = NULL; 1207 uint32_t length = rdma_req->req.length; 1208 uint64_t translation_len; 1209 uint32_t i = 0; 1210 int rc = 0; 1211 1212 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1213 rgroup = rqpair->poller->group; 1214 rdma_req->req.iovcnt = 0; 1215 while (length) { 1216 if (!(STAILQ_EMPTY(&rgroup->group.buf_cache))) { 1217 rgroup->group.buf_cache_count--; 1218 buf = STAILQ_FIRST(&rgroup->group.buf_cache); 1219 STAILQ_REMOVE_HEAD(&rgroup->group.buf_cache, link); 1220 assert(buf != NULL); 1221 } else { 1222 buf = spdk_mempool_get(rtransport->transport.data_buf_pool); 1223 if (!buf) { 1224 rc = -ENOMEM; 1225 goto err_exit; 1226 } 1227 } 1228 1229 rdma_req->req.iov[i].iov_base = (void *)((uintptr_t)(buf + NVMF_DATA_BUFFER_MASK) & 1230 ~NVMF_DATA_BUFFER_MASK); 1231 rdma_req->req.iov[i].iov_len = spdk_min(length, rtransport->transport.opts.io_unit_size); 1232 rdma_req->req.iovcnt++; 1233 rdma_req->data.buffers[i] = buf; 1234 rdma_req->data.wr.sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[i].iov_base); 1235 rdma_req->data.wr.sg_list[i].length = rdma_req->req.iov[i].iov_len; 1236 translation_len = rdma_req->req.iov[i].iov_len; 1237 1238 if (!g_nvmf_hooks.get_rkey) { 1239 rdma_req->data.wr.sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1240 (uint64_t)buf, &translation_len))->lkey; 1241 } else { 1242 rdma_req->data.wr.sg_list[i].lkey = spdk_mem_map_translate(device->map, 1243 (uint64_t)buf, &translation_len); 1244 } 1245 1246 length -= rdma_req->req.iov[i].iov_len; 1247 1248 if (translation_len < rdma_req->req.iov[i].iov_len) { 1249 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions\n"); 1250 rc = -EINVAL; 1251 goto err_exit; 1252 } 1253 i++; 1254 } 1255 1256 assert(rdma_req->req.iovcnt <= rqpair->max_send_sge); 1257 1258 rdma_req->data_from_pool = true; 1259 1260 return rc; 1261 1262 err_exit: 1263 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport); 1264 while (i) { 1265 i--; 1266 rdma_req->data.wr.sg_list[i].addr = 0; 1267 rdma_req->data.wr.sg_list[i].length = 0; 1268 rdma_req->data.wr.sg_list[i].lkey = 0; 1269 } 1270 rdma_req->req.iovcnt = 0; 1271 return rc; 1272 } 1273 1274 static int 1275 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1276 struct spdk_nvmf_rdma_device *device, 1277 struct spdk_nvmf_rdma_request *rdma_req) 1278 { 1279 struct spdk_nvme_cmd *cmd; 1280 struct spdk_nvme_cpl *rsp; 1281 struct spdk_nvme_sgl_descriptor *sgl; 1282 1283 cmd = &rdma_req->req.cmd->nvme_cmd; 1284 rsp = &rdma_req->req.rsp->nvme_cpl; 1285 sgl = &cmd->dptr.sgl1; 1286 1287 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1288 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1289 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1290 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1291 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1292 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1293 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1294 return -1; 1295 } 1296 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1297 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1298 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1299 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1300 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1301 } 1302 } 1303 #endif 1304 1305 /* fill request length and populate iovs */ 1306 rdma_req->req.length = sgl->keyed.length; 1307 1308 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1309 /* No available buffers. Queue this request up. */ 1310 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1311 return 0; 1312 } 1313 1314 /* backward compatible */ 1315 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1316 1317 /* rdma wr specifics */ 1318 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1319 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1320 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1321 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1322 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 1323 rdma_req->data.wr.next = &rdma_req->rsp.wr; 1324 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1325 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 1326 rdma_req->data.wr.next = NULL; 1327 } 1328 1329 /* set the number of outstanding data WRs for this request. */ 1330 rdma_req->num_outstanding_data_wr = 1; 1331 1332 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1333 rdma_req->req.iovcnt); 1334 1335 return 0; 1336 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1337 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1338 uint64_t offset = sgl->address; 1339 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1340 1341 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1342 offset, sgl->unkeyed.length); 1343 1344 if (offset > max_len) { 1345 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1346 offset, max_len); 1347 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1348 return -1; 1349 } 1350 max_len -= (uint32_t)offset; 1351 1352 if (sgl->unkeyed.length > max_len) { 1353 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1354 sgl->unkeyed.length, max_len); 1355 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1356 return -1; 1357 } 1358 1359 rdma_req->num_outstanding_data_wr = 0; 1360 rdma_req->req.data = rdma_req->recv->buf + offset; 1361 rdma_req->data_from_pool = false; 1362 rdma_req->req.length = sgl->unkeyed.length; 1363 1364 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1365 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1366 rdma_req->req.iovcnt = 1; 1367 1368 return 0; 1369 } 1370 1371 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1372 sgl->generic.type, sgl->generic.subtype); 1373 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1374 return -1; 1375 } 1376 1377 static void 1378 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1379 struct spdk_nvmf_rdma_transport *rtransport) 1380 { 1381 struct spdk_nvmf_rdma_qpair *rqpair; 1382 struct spdk_nvmf_rdma_poll_group *rgroup; 1383 1384 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1385 if (rdma_req->data_from_pool) { 1386 rgroup = rqpair->poller->group; 1387 1388 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport); 1389 } 1390 rdma_req->num_outstanding_data_wr = 0; 1391 rdma_req->req.length = 0; 1392 rdma_req->req.iovcnt = 0; 1393 rdma_req->req.data = NULL; 1394 rqpair->qd--; 1395 STAILQ_INSERT_HEAD(&rqpair->free_queue, rdma_req, state_link); 1396 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1397 } 1398 1399 static bool 1400 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1401 struct spdk_nvmf_rdma_request *rdma_req) 1402 { 1403 struct spdk_nvmf_rdma_qpair *rqpair; 1404 struct spdk_nvmf_rdma_device *device; 1405 struct spdk_nvmf_rdma_poll_group *rgroup; 1406 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1407 int rc; 1408 struct spdk_nvmf_rdma_recv *rdma_recv; 1409 enum spdk_nvmf_rdma_request_state prev_state; 1410 bool progress = false; 1411 int data_posted; 1412 1413 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1414 device = rqpair->port->device; 1415 rgroup = rqpair->poller->group; 1416 1417 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1418 1419 /* If the queue pair is in an error state, force the request to the completed state 1420 * to release resources. */ 1421 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1422 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1423 TAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, link); 1424 } 1425 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1426 } 1427 1428 /* The loop here is to allow for several back-to-back state changes. */ 1429 do { 1430 prev_state = rdma_req->state; 1431 1432 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1433 1434 switch (rdma_req->state) { 1435 case RDMA_REQUEST_STATE_FREE: 1436 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1437 * to escape this state. */ 1438 break; 1439 case RDMA_REQUEST_STATE_NEW: 1440 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1441 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1442 rdma_recv = rdma_req->recv; 1443 1444 /* The first element of the SGL is the NVMe command */ 1445 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1446 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1447 1448 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1449 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1450 break; 1451 } 1452 1453 /* The next state transition depends on the data transfer needs of this request. */ 1454 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1455 1456 /* If no data to transfer, ready to execute. */ 1457 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1458 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1459 break; 1460 } 1461 1462 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1463 TAILQ_INSERT_TAIL(&rgroup->pending_data_buf_queue, rdma_req, link); 1464 break; 1465 case RDMA_REQUEST_STATE_NEED_BUFFER: 1466 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1467 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1468 1469 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1470 1471 if (rdma_req != TAILQ_FIRST(&rgroup->pending_data_buf_queue)) { 1472 /* This request needs to wait in line to obtain a buffer */ 1473 break; 1474 } 1475 1476 /* Try to get a data buffer */ 1477 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1478 if (rc < 0) { 1479 TAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, link); 1480 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1481 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1482 break; 1483 } 1484 1485 if (!rdma_req->req.data) { 1486 /* No buffers available. */ 1487 break; 1488 } 1489 1490 TAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, link); 1491 1492 /* If data is transferring from host to controller and the data didn't 1493 * arrive using in capsule data, we need to do a transfer from the host. 1494 */ 1495 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1496 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1497 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1498 break; 1499 } 1500 1501 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1502 break; 1503 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1504 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1505 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1506 1507 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1508 /* This request needs to wait in line to perform RDMA */ 1509 break; 1510 } 1511 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1512 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1513 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1514 break; 1515 } 1516 1517 /* We have already verified that this request is the head of the queue. */ 1518 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 1519 1520 rc = request_transfer_in(&rdma_req->req); 1521 if (!rc) { 1522 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 1523 } else { 1524 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1525 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1526 } 1527 break; 1528 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1529 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1530 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1531 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1532 * to escape this state. */ 1533 break; 1534 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1535 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1536 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1537 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 1538 spdk_nvmf_request_exec(&rdma_req->req); 1539 break; 1540 case RDMA_REQUEST_STATE_EXECUTING: 1541 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1542 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1543 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1544 * to escape this state. */ 1545 break; 1546 case RDMA_REQUEST_STATE_EXECUTED: 1547 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1548 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1549 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1550 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 1551 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 1552 } else { 1553 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1554 } 1555 break; 1556 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 1557 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 1558 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1559 1560 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 1561 /* This request needs to wait in line to perform RDMA */ 1562 break; 1563 } 1564 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 1565 rqpair->max_send_depth) { 1566 /* We can only have so many WRs outstanding. we have to wait until some finish. 1567 * +1 since each request has an additional wr in the resp. */ 1568 break; 1569 } 1570 1571 /* We have already verified that this request is the head of the queue. */ 1572 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 1573 1574 /* The data transfer will be kicked off from 1575 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1576 */ 1577 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1578 break; 1579 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1580 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 1581 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1582 rc = request_transfer_out(&rdma_req->req, &data_posted); 1583 assert(rc == 0); /* No good way to handle this currently */ 1584 if (rc) { 1585 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1586 } else { 1587 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1588 RDMA_REQUEST_STATE_COMPLETING; 1589 } 1590 break; 1591 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1592 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 1593 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1594 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1595 * to escape this state. */ 1596 break; 1597 case RDMA_REQUEST_STATE_COMPLETING: 1598 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 1599 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1600 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1601 * to escape this state. */ 1602 break; 1603 case RDMA_REQUEST_STATE_COMPLETED: 1604 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 1605 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1606 1607 nvmf_rdma_request_free(rdma_req, rtransport); 1608 break; 1609 case RDMA_REQUEST_NUM_STATES: 1610 default: 1611 assert(0); 1612 break; 1613 } 1614 1615 if (rdma_req->state != prev_state) { 1616 progress = true; 1617 } 1618 } while (rdma_req->state != prev_state); 1619 1620 return progress; 1621 } 1622 1623 /* Public API callbacks begin here */ 1624 1625 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 1626 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 1627 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 1628 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 1629 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 1630 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 1631 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 512 1632 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 1633 1634 static void 1635 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 1636 { 1637 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 1638 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 1639 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 1640 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 1641 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 1642 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 1643 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 1644 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 1645 } 1646 1647 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 1648 1649 static struct spdk_nvmf_transport * 1650 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 1651 { 1652 int rc; 1653 struct spdk_nvmf_rdma_transport *rtransport; 1654 struct spdk_nvmf_rdma_device *device, *tmp; 1655 struct ibv_context **contexts; 1656 uint32_t i; 1657 int flag; 1658 uint32_t sge_count; 1659 uint32_t min_shared_buffers; 1660 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 1661 1662 rtransport = calloc(1, sizeof(*rtransport)); 1663 if (!rtransport) { 1664 return NULL; 1665 } 1666 1667 if (pthread_mutex_init(&rtransport->lock, NULL)) { 1668 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 1669 free(rtransport); 1670 return NULL; 1671 } 1672 1673 TAILQ_INIT(&rtransport->devices); 1674 TAILQ_INIT(&rtransport->ports); 1675 1676 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 1677 1678 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 1679 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 1680 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 1681 " in_capsule_data_size=%d, max_aq_depth=%d\n" 1682 " num_shared_buffers=%d\n", 1683 opts->max_queue_depth, 1684 opts->max_io_size, 1685 opts->max_qpairs_per_ctrlr, 1686 opts->io_unit_size, 1687 opts->in_capsule_data_size, 1688 opts->max_aq_depth, 1689 opts->num_shared_buffers); 1690 1691 /* I/O unit size cannot be larger than max I/O size */ 1692 if (opts->io_unit_size > opts->max_io_size) { 1693 opts->io_unit_size = opts->max_io_size; 1694 } 1695 1696 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 1697 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 1698 "the minimum number required to guarantee that forward progress can be made (%d)\n", 1699 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 1700 spdk_nvmf_rdma_destroy(&rtransport->transport); 1701 return NULL; 1702 } 1703 1704 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 1705 if (min_shared_buffers > opts->num_shared_buffers) { 1706 SPDK_ERRLOG("There are not enough buffers to satisfy" 1707 "per-poll group caches for each thread. (%" PRIu32 ")" 1708 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 1709 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 1710 spdk_nvmf_rdma_destroy(&rtransport->transport); 1711 return NULL; 1712 } 1713 1714 sge_count = opts->max_io_size / opts->io_unit_size; 1715 if (sge_count > NVMF_DEFAULT_TX_SGE) { 1716 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 1717 spdk_nvmf_rdma_destroy(&rtransport->transport); 1718 return NULL; 1719 } 1720 1721 rtransport->event_channel = rdma_create_event_channel(); 1722 if (rtransport->event_channel == NULL) { 1723 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 1724 spdk_nvmf_rdma_destroy(&rtransport->transport); 1725 return NULL; 1726 } 1727 1728 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 1729 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 1730 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 1731 rtransport->event_channel->fd, spdk_strerror(errno)); 1732 spdk_nvmf_rdma_destroy(&rtransport->transport); 1733 return NULL; 1734 } 1735 1736 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 1737 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 1738 sizeof(struct spdk_nvmf_rdma_request_data), 1739 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 1740 SPDK_ENV_SOCKET_ID_ANY); 1741 if (!rtransport->data_wr_pool) { 1742 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 1743 spdk_nvmf_rdma_destroy(&rtransport->transport); 1744 return NULL; 1745 } 1746 1747 contexts = rdma_get_devices(NULL); 1748 if (contexts == NULL) { 1749 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 1750 spdk_nvmf_rdma_destroy(&rtransport->transport); 1751 return NULL; 1752 } 1753 1754 i = 0; 1755 rc = 0; 1756 while (contexts[i] != NULL) { 1757 device = calloc(1, sizeof(*device)); 1758 if (!device) { 1759 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 1760 rc = -ENOMEM; 1761 break; 1762 } 1763 device->context = contexts[i]; 1764 rc = ibv_query_device(device->context, &device->attr); 1765 if (rc < 0) { 1766 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1767 free(device); 1768 break; 1769 1770 } 1771 1772 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 1773 1774 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1775 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 1776 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 1777 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 1778 } 1779 1780 /** 1781 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 1782 * The Soft-RoCE RXE driver does not currently support send with invalidate, 1783 * but incorrectly reports that it does. There are changes making their way 1784 * through the kernel now that will enable this feature. When they are merged, 1785 * we can conditionally enable this feature. 1786 * 1787 * TODO: enable this for versions of the kernel rxe driver that support it. 1788 */ 1789 if (device->attr.vendor_id == 0) { 1790 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 1791 } 1792 #endif 1793 1794 /* set up device context async ev fd as NON_BLOCKING */ 1795 flag = fcntl(device->context->async_fd, F_GETFL); 1796 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 1797 if (rc < 0) { 1798 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 1799 free(device); 1800 break; 1801 } 1802 1803 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 1804 i++; 1805 } 1806 rdma_free_devices(contexts); 1807 1808 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 1809 /* divide and round up. */ 1810 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 1811 1812 /* round up to the nearest 4k. */ 1813 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 1814 1815 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 1816 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 1817 opts->io_unit_size); 1818 } 1819 1820 if (rc < 0) { 1821 spdk_nvmf_rdma_destroy(&rtransport->transport); 1822 return NULL; 1823 } 1824 1825 /* Set up poll descriptor array to monitor events from RDMA and IB 1826 * in a single poll syscall 1827 */ 1828 rtransport->npoll_fds = i + 1; 1829 i = 0; 1830 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 1831 if (rtransport->poll_fds == NULL) { 1832 SPDK_ERRLOG("poll_fds allocation failed\n"); 1833 spdk_nvmf_rdma_destroy(&rtransport->transport); 1834 return NULL; 1835 } 1836 1837 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 1838 rtransport->poll_fds[i++].events = POLLIN; 1839 1840 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 1841 rtransport->poll_fds[i].fd = device->context->async_fd; 1842 rtransport->poll_fds[i++].events = POLLIN; 1843 } 1844 1845 return &rtransport->transport; 1846 } 1847 1848 static int 1849 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 1850 { 1851 struct spdk_nvmf_rdma_transport *rtransport; 1852 struct spdk_nvmf_rdma_port *port, *port_tmp; 1853 struct spdk_nvmf_rdma_device *device, *device_tmp; 1854 1855 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1856 1857 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 1858 TAILQ_REMOVE(&rtransport->ports, port, link); 1859 rdma_destroy_id(port->id); 1860 free(port); 1861 } 1862 1863 if (rtransport->poll_fds != NULL) { 1864 free(rtransport->poll_fds); 1865 } 1866 1867 if (rtransport->event_channel != NULL) { 1868 rdma_destroy_event_channel(rtransport->event_channel); 1869 } 1870 1871 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 1872 TAILQ_REMOVE(&rtransport->devices, device, link); 1873 if (device->map) { 1874 spdk_mem_map_free(&device->map); 1875 } 1876 if (device->pd) { 1877 if (!g_nvmf_hooks.get_ibv_pd) { 1878 ibv_dealloc_pd(device->pd); 1879 } 1880 } 1881 free(device); 1882 } 1883 1884 if (rtransport->data_wr_pool != NULL) { 1885 if (spdk_mempool_count(rtransport->data_wr_pool) != 1886 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 1887 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 1888 spdk_mempool_count(rtransport->data_wr_pool), 1889 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 1890 } 1891 } 1892 1893 spdk_mempool_free(rtransport->data_wr_pool); 1894 pthread_mutex_destroy(&rtransport->lock); 1895 free(rtransport); 1896 1897 return 0; 1898 } 1899 1900 static int 1901 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 1902 struct spdk_nvme_transport_id *trid, 1903 bool peer); 1904 1905 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 1906 .notify_cb = spdk_nvmf_rdma_mem_notify, 1907 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 1908 }; 1909 1910 static int 1911 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 1912 const struct spdk_nvme_transport_id *trid) 1913 { 1914 struct spdk_nvmf_rdma_transport *rtransport; 1915 struct spdk_nvmf_rdma_device *device; 1916 struct spdk_nvmf_rdma_port *port_tmp, *port; 1917 struct ibv_pd *pd; 1918 struct addrinfo *res; 1919 struct addrinfo hints; 1920 int family; 1921 int rc; 1922 1923 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1924 1925 port = calloc(1, sizeof(*port)); 1926 if (!port) { 1927 return -ENOMEM; 1928 } 1929 1930 /* Selectively copy the trid. Things like NQN don't matter here - that 1931 * mapping is enforced elsewhere. 1932 */ 1933 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1934 port->trid.adrfam = trid->adrfam; 1935 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 1936 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 1937 1938 pthread_mutex_lock(&rtransport->lock); 1939 assert(rtransport->event_channel != NULL); 1940 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 1941 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 1942 port_tmp->ref++; 1943 free(port); 1944 /* Already listening at this address */ 1945 pthread_mutex_unlock(&rtransport->lock); 1946 return 0; 1947 } 1948 } 1949 1950 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 1951 if (rc < 0) { 1952 SPDK_ERRLOG("rdma_create_id() failed\n"); 1953 free(port); 1954 pthread_mutex_unlock(&rtransport->lock); 1955 return rc; 1956 } 1957 1958 switch (port->trid.adrfam) { 1959 case SPDK_NVMF_ADRFAM_IPV4: 1960 family = AF_INET; 1961 break; 1962 case SPDK_NVMF_ADRFAM_IPV6: 1963 family = AF_INET6; 1964 break; 1965 default: 1966 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 1967 free(port); 1968 pthread_mutex_unlock(&rtransport->lock); 1969 return -EINVAL; 1970 } 1971 1972 memset(&hints, 0, sizeof(hints)); 1973 hints.ai_family = family; 1974 hints.ai_flags = AI_NUMERICSERV; 1975 hints.ai_socktype = SOCK_STREAM; 1976 hints.ai_protocol = 0; 1977 1978 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 1979 if (rc) { 1980 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 1981 free(port); 1982 pthread_mutex_unlock(&rtransport->lock); 1983 return -EINVAL; 1984 } 1985 1986 rc = rdma_bind_addr(port->id, res->ai_addr); 1987 freeaddrinfo(res); 1988 1989 if (rc < 0) { 1990 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 1991 rdma_destroy_id(port->id); 1992 free(port); 1993 pthread_mutex_unlock(&rtransport->lock); 1994 return rc; 1995 } 1996 1997 if (!port->id->verbs) { 1998 SPDK_ERRLOG("ibv_context is null\n"); 1999 rdma_destroy_id(port->id); 2000 free(port); 2001 pthread_mutex_unlock(&rtransport->lock); 2002 return -1; 2003 } 2004 2005 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 2006 if (rc < 0) { 2007 SPDK_ERRLOG("rdma_listen() failed\n"); 2008 rdma_destroy_id(port->id); 2009 free(port); 2010 pthread_mutex_unlock(&rtransport->lock); 2011 return rc; 2012 } 2013 2014 TAILQ_FOREACH(device, &rtransport->devices, link) { 2015 if (device->context == port->id->verbs) { 2016 port->device = device; 2017 break; 2018 } 2019 } 2020 if (!port->device) { 2021 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2022 port->id->verbs); 2023 rdma_destroy_id(port->id); 2024 free(port); 2025 pthread_mutex_unlock(&rtransport->lock); 2026 return -EINVAL; 2027 } 2028 2029 pd = NULL; 2030 if (g_nvmf_hooks.get_ibv_pd) { 2031 if (spdk_nvmf_rdma_trid_from_cm_id(port->id, &port->trid, 1) < 0) { 2032 rdma_destroy_id(port->id); 2033 free(port); 2034 pthread_mutex_unlock(&rtransport->lock); 2035 return -EINVAL; 2036 } 2037 2038 pd = g_nvmf_hooks.get_ibv_pd(&port->trid, port->id->verbs); 2039 } 2040 2041 if (device->pd == NULL) { 2042 /* Haven't created a protection domain yet. */ 2043 2044 if (!g_nvmf_hooks.get_ibv_pd) { 2045 device->pd = ibv_alloc_pd(device->context); 2046 if (!device->pd) { 2047 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2048 rdma_destroy_id(port->id); 2049 free(port); 2050 pthread_mutex_unlock(&rtransport->lock); 2051 return -ENOMEM; 2052 } 2053 } else { 2054 device->pd = pd; 2055 } 2056 2057 assert(device->map == NULL); 2058 2059 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2060 if (!device->map) { 2061 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2062 if (!g_nvmf_hooks.get_ibv_pd) { 2063 ibv_dealloc_pd(device->pd); 2064 } 2065 rdma_destroy_id(port->id); 2066 free(port); 2067 pthread_mutex_unlock(&rtransport->lock); 2068 return -ENOMEM; 2069 } 2070 } else if (g_nvmf_hooks.get_ibv_pd) { 2071 /* A protection domain exists for this device, but the user has 2072 * enabled hooks. Verify that they only supply one pd per device. */ 2073 if (device->pd != pd) { 2074 SPDK_ERRLOG("The NVMe-oF target only supports one protection domain per device.\n"); 2075 rdma_destroy_id(port->id); 2076 free(port); 2077 pthread_mutex_unlock(&rtransport->lock); 2078 return -EINVAL; 2079 } 2080 } 2081 2082 assert(device->map != NULL); 2083 assert(device->pd != NULL); 2084 2085 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 2086 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 2087 2088 port->ref = 1; 2089 2090 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2091 pthread_mutex_unlock(&rtransport->lock); 2092 2093 return 0; 2094 } 2095 2096 static int 2097 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2098 const struct spdk_nvme_transport_id *_trid) 2099 { 2100 struct spdk_nvmf_rdma_transport *rtransport; 2101 struct spdk_nvmf_rdma_port *port, *tmp; 2102 struct spdk_nvme_transport_id trid = {}; 2103 2104 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2105 2106 /* Selectively copy the trid. Things like NQN don't matter here - that 2107 * mapping is enforced elsewhere. 2108 */ 2109 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2110 trid.adrfam = _trid->adrfam; 2111 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2112 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2113 2114 pthread_mutex_lock(&rtransport->lock); 2115 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2116 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2117 assert(port->ref > 0); 2118 port->ref--; 2119 if (port->ref == 0) { 2120 TAILQ_REMOVE(&rtransport->ports, port, link); 2121 rdma_destroy_id(port->id); 2122 free(port); 2123 } 2124 break; 2125 } 2126 } 2127 2128 pthread_mutex_unlock(&rtransport->lock); 2129 return 0; 2130 } 2131 2132 static bool 2133 spdk_nvmf_rdma_qpair_is_idle(struct spdk_nvmf_qpair *qpair) 2134 { 2135 struct spdk_nvmf_rdma_qpair *rqpair; 2136 2137 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2138 2139 if (rqpair->qd == 0) { 2140 return true; 2141 } 2142 return false; 2143 } 2144 2145 static void 2146 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2147 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2148 { 2149 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2150 2151 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2152 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2153 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2154 break; 2155 } 2156 } 2157 2158 /* Then RDMA writes since reads have stronger restrictions than writes */ 2159 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2160 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2161 break; 2162 } 2163 } 2164 2165 /* The second highest priority is I/O waiting on memory buffers. */ 2166 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->poller->group->pending_data_buf_queue, link, 2167 req_tmp) { 2168 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2169 break; 2170 } 2171 } 2172 2173 while (!STAILQ_EMPTY(&rqpair->free_queue) && !STAILQ_EMPTY(&rqpair->incoming_queue)) { 2174 2175 rdma_req = STAILQ_FIRST(&rqpair->free_queue); 2176 STAILQ_REMOVE_HEAD(&rqpair->free_queue, state_link); 2177 rdma_req->recv = STAILQ_FIRST(&rqpair->incoming_queue); 2178 STAILQ_REMOVE_HEAD(&rqpair->incoming_queue, link); 2179 2180 rqpair->qd++; 2181 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2182 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2183 break; 2184 } 2185 } 2186 } 2187 2188 static void 2189 _nvmf_rdma_qpair_disconnect(void *ctx) 2190 { 2191 struct spdk_nvmf_qpair *qpair = ctx; 2192 2193 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2194 } 2195 2196 static void 2197 _nvmf_rdma_try_disconnect(void *ctx) 2198 { 2199 struct spdk_nvmf_qpair *qpair = ctx; 2200 struct spdk_nvmf_poll_group *group; 2201 2202 /* Read the group out of the qpair. This is normally set and accessed only from 2203 * the thread that created the group. Here, we're not on that thread necessarily. 2204 * The data member qpair->group begins it's life as NULL and then is assigned to 2205 * a pointer and never changes. So fortunately reading this and checking for 2206 * non-NULL is thread safe in the x86_64 memory model. */ 2207 group = qpair->group; 2208 2209 if (group == NULL) { 2210 /* The qpair hasn't been assigned to a group yet, so we can't 2211 * process a disconnect. Send a message to ourself and try again. */ 2212 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2213 return; 2214 } 2215 2216 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2217 } 2218 2219 static inline void 2220 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2221 { 2222 if (__sync_bool_compare_and_swap(&rqpair->disconnect_started, false, true)) { 2223 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2224 } 2225 } 2226 2227 2228 static int 2229 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2230 { 2231 struct spdk_nvmf_qpair *qpair; 2232 struct spdk_nvmf_rdma_qpair *rqpair; 2233 2234 if (evt->id == NULL) { 2235 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2236 return -1; 2237 } 2238 2239 qpair = evt->id->context; 2240 if (qpair == NULL) { 2241 SPDK_ERRLOG("disconnect request: no active connection\n"); 2242 return -1; 2243 } 2244 2245 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2246 2247 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2248 2249 spdk_nvmf_rdma_update_ibv_state(rqpair); 2250 2251 spdk_nvmf_rdma_start_disconnect(rqpair); 2252 2253 return 0; 2254 } 2255 2256 #ifdef DEBUG 2257 static const char *CM_EVENT_STR[] = { 2258 "RDMA_CM_EVENT_ADDR_RESOLVED", 2259 "RDMA_CM_EVENT_ADDR_ERROR", 2260 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2261 "RDMA_CM_EVENT_ROUTE_ERROR", 2262 "RDMA_CM_EVENT_CONNECT_REQUEST", 2263 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2264 "RDMA_CM_EVENT_CONNECT_ERROR", 2265 "RDMA_CM_EVENT_UNREACHABLE", 2266 "RDMA_CM_EVENT_REJECTED", 2267 "RDMA_CM_EVENT_ESTABLISHED", 2268 "RDMA_CM_EVENT_DISCONNECTED", 2269 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2270 "RDMA_CM_EVENT_MULTICAST_JOIN", 2271 "RDMA_CM_EVENT_MULTICAST_ERROR", 2272 "RDMA_CM_EVENT_ADDR_CHANGE", 2273 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2274 }; 2275 #endif /* DEBUG */ 2276 2277 static void 2278 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2279 { 2280 struct spdk_nvmf_rdma_transport *rtransport; 2281 struct rdma_cm_event *event; 2282 int rc; 2283 2284 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2285 2286 if (rtransport->event_channel == NULL) { 2287 return; 2288 } 2289 2290 while (1) { 2291 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2292 if (rc == 0) { 2293 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2294 2295 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2296 2297 switch (event->event) { 2298 case RDMA_CM_EVENT_ADDR_RESOLVED: 2299 case RDMA_CM_EVENT_ADDR_ERROR: 2300 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2301 case RDMA_CM_EVENT_ROUTE_ERROR: 2302 /* No action required. The target never attempts to resolve routes. */ 2303 break; 2304 case RDMA_CM_EVENT_CONNECT_REQUEST: 2305 rc = nvmf_rdma_connect(transport, event, cb_fn); 2306 if (rc < 0) { 2307 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2308 break; 2309 } 2310 break; 2311 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2312 /* The target never initiates a new connection. So this will not occur. */ 2313 break; 2314 case RDMA_CM_EVENT_CONNECT_ERROR: 2315 /* Can this happen? The docs say it can, but not sure what causes it. */ 2316 break; 2317 case RDMA_CM_EVENT_UNREACHABLE: 2318 case RDMA_CM_EVENT_REJECTED: 2319 /* These only occur on the client side. */ 2320 break; 2321 case RDMA_CM_EVENT_ESTABLISHED: 2322 /* TODO: Should we be waiting for this event anywhere? */ 2323 break; 2324 case RDMA_CM_EVENT_DISCONNECTED: 2325 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2326 rc = nvmf_rdma_disconnect(event); 2327 if (rc < 0) { 2328 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2329 break; 2330 } 2331 break; 2332 case RDMA_CM_EVENT_MULTICAST_JOIN: 2333 case RDMA_CM_EVENT_MULTICAST_ERROR: 2334 /* Multicast is not used */ 2335 break; 2336 case RDMA_CM_EVENT_ADDR_CHANGE: 2337 /* Not utilizing this event */ 2338 break; 2339 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2340 /* For now, do nothing. The target never re-uses queue pairs. */ 2341 break; 2342 default: 2343 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2344 break; 2345 } 2346 2347 rdma_ack_cm_event(event); 2348 } else { 2349 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2350 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2351 } 2352 break; 2353 } 2354 } 2355 } 2356 2357 static void 2358 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2359 { 2360 int rc; 2361 struct spdk_nvmf_rdma_qpair *rqpair; 2362 struct ibv_async_event event; 2363 enum ibv_qp_state state; 2364 2365 rc = ibv_get_async_event(device->context, &event); 2366 2367 if (rc) { 2368 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2369 errno, spdk_strerror(errno)); 2370 return; 2371 } 2372 2373 SPDK_NOTICELOG("Async event: %s\n", 2374 ibv_event_type_str(event.event_type)); 2375 2376 switch (event.event_type) { 2377 case IBV_EVENT_QP_FATAL: 2378 rqpair = event.element.qp->qp_context; 2379 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2380 (uintptr_t)rqpair->cm_id, event.event_type); 2381 spdk_nvmf_rdma_update_ibv_state(rqpair); 2382 spdk_nvmf_rdma_start_disconnect(rqpair); 2383 break; 2384 case IBV_EVENT_QP_LAST_WQE_REACHED: 2385 /* This event only occurs for shared receive queues, which are not currently supported. */ 2386 break; 2387 case IBV_EVENT_SQ_DRAINED: 2388 /* This event occurs frequently in both error and non-error states. 2389 * Check if the qpair is in an error state before sending a message. 2390 * Note that we're not on the correct thread to access the qpair, but 2391 * the operations that the below calls make all happen to be thread 2392 * safe. */ 2393 rqpair = event.element.qp->qp_context; 2394 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2395 (uintptr_t)rqpair->cm_id, event.event_type); 2396 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2397 if (state == IBV_QPS_ERR) { 2398 spdk_nvmf_rdma_start_disconnect(rqpair); 2399 } 2400 break; 2401 case IBV_EVENT_QP_REQ_ERR: 2402 case IBV_EVENT_QP_ACCESS_ERR: 2403 case IBV_EVENT_COMM_EST: 2404 case IBV_EVENT_PATH_MIG: 2405 case IBV_EVENT_PATH_MIG_ERR: 2406 rqpair = event.element.qp->qp_context; 2407 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2408 (uintptr_t)rqpair->cm_id, event.event_type); 2409 spdk_nvmf_rdma_update_ibv_state(rqpair); 2410 break; 2411 case IBV_EVENT_CQ_ERR: 2412 case IBV_EVENT_DEVICE_FATAL: 2413 case IBV_EVENT_PORT_ACTIVE: 2414 case IBV_EVENT_PORT_ERR: 2415 case IBV_EVENT_LID_CHANGE: 2416 case IBV_EVENT_PKEY_CHANGE: 2417 case IBV_EVENT_SM_CHANGE: 2418 case IBV_EVENT_SRQ_ERR: 2419 case IBV_EVENT_SRQ_LIMIT_REACHED: 2420 case IBV_EVENT_CLIENT_REREGISTER: 2421 case IBV_EVENT_GID_CHANGE: 2422 default: 2423 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2424 break; 2425 } 2426 ibv_ack_async_event(&event); 2427 } 2428 2429 static void 2430 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2431 { 2432 int nfds, i = 0; 2433 struct spdk_nvmf_rdma_transport *rtransport; 2434 struct spdk_nvmf_rdma_device *device, *tmp; 2435 2436 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2437 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2438 2439 if (nfds <= 0) { 2440 return; 2441 } 2442 2443 /* The first poll descriptor is RDMA CM event */ 2444 if (rtransport->poll_fds[i++].revents & POLLIN) { 2445 spdk_nvmf_process_cm_event(transport, cb_fn); 2446 nfds--; 2447 } 2448 2449 if (nfds == 0) { 2450 return; 2451 } 2452 2453 /* Second and subsequent poll descriptors are IB async events */ 2454 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2455 if (rtransport->poll_fds[i++].revents & POLLIN) { 2456 spdk_nvmf_process_ib_event(device); 2457 nfds--; 2458 } 2459 } 2460 /* check all flagged fd's have been served */ 2461 assert(nfds == 0); 2462 } 2463 2464 static void 2465 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2466 struct spdk_nvme_transport_id *trid, 2467 struct spdk_nvmf_discovery_log_page_entry *entry) 2468 { 2469 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2470 entry->adrfam = trid->adrfam; 2471 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2472 2473 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2474 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2475 2476 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2477 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2478 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2479 } 2480 2481 static struct spdk_nvmf_transport_poll_group * 2482 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2483 { 2484 struct spdk_nvmf_rdma_transport *rtransport; 2485 struct spdk_nvmf_rdma_poll_group *rgroup; 2486 struct spdk_nvmf_rdma_poller *poller, *tpoller; 2487 struct spdk_nvmf_rdma_device *device; 2488 2489 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2490 2491 rgroup = calloc(1, sizeof(*rgroup)); 2492 if (!rgroup) { 2493 return NULL; 2494 } 2495 2496 TAILQ_INIT(&rgroup->pollers); 2497 TAILQ_INIT(&rgroup->pending_data_buf_queue); 2498 2499 pthread_mutex_lock(&rtransport->lock); 2500 TAILQ_FOREACH(device, &rtransport->devices, link) { 2501 poller = calloc(1, sizeof(*poller)); 2502 if (!poller) { 2503 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2504 goto err_exit; 2505 } 2506 2507 poller->device = device; 2508 poller->group = rgroup; 2509 2510 TAILQ_INIT(&poller->qpairs); 2511 2512 poller->cq = ibv_create_cq(device->context, DEFAULT_NVMF_RDMA_CQ_SIZE, poller, NULL, 0); 2513 if (!poller->cq) { 2514 SPDK_ERRLOG("Unable to create completion queue\n"); 2515 free(poller); 2516 goto err_exit; 2517 } 2518 poller->num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2519 2520 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2521 } 2522 2523 pthread_mutex_unlock(&rtransport->lock); 2524 return &rgroup->group; 2525 2526 err_exit: 2527 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tpoller) { 2528 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2529 if (poller->cq) { 2530 ibv_destroy_cq(poller->cq); 2531 } 2532 free(poller); 2533 } 2534 2535 free(rgroup); 2536 pthread_mutex_unlock(&rtransport->lock); 2537 return NULL; 2538 } 2539 2540 static void 2541 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2542 { 2543 struct spdk_nvmf_rdma_poll_group *rgroup; 2544 struct spdk_nvmf_rdma_poller *poller, *tmp; 2545 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2546 2547 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2548 2549 if (!rgroup) { 2550 return; 2551 } 2552 2553 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2554 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2555 2556 if (poller->cq) { 2557 ibv_destroy_cq(poller->cq); 2558 } 2559 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 2560 spdk_nvmf_rdma_qpair_destroy(qpair); 2561 } 2562 2563 free(poller); 2564 } 2565 2566 if (!TAILQ_EMPTY(&rgroup->pending_data_buf_queue)) { 2567 SPDK_ERRLOG("Pending I/O list wasn't empty on poll group destruction\n"); 2568 } 2569 2570 free(rgroup); 2571 } 2572 2573 static void 2574 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 2575 { 2576 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2577 spdk_nvmf_rdma_qpair_destroy(rqpair); 2578 } 2579 2580 static int 2581 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 2582 struct spdk_nvmf_qpair *qpair) 2583 { 2584 struct spdk_nvmf_rdma_poll_group *rgroup; 2585 struct spdk_nvmf_rdma_qpair *rqpair; 2586 struct spdk_nvmf_rdma_device *device; 2587 struct spdk_nvmf_rdma_poller *poller; 2588 int rc; 2589 2590 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2591 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2592 2593 device = rqpair->port->device; 2594 2595 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 2596 if (poller->device == device) { 2597 break; 2598 } 2599 } 2600 2601 if (!poller) { 2602 SPDK_ERRLOG("No poller found for device.\n"); 2603 return -1; 2604 } 2605 2606 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 2607 rqpair->poller = poller; 2608 2609 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 2610 if (rc < 0) { 2611 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 2612 return -1; 2613 } 2614 2615 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 2616 if (rc) { 2617 /* Try to reject, but we probably can't */ 2618 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 2619 return -1; 2620 } 2621 2622 spdk_nvmf_rdma_update_ibv_state(rqpair); 2623 2624 return 0; 2625 } 2626 2627 static int 2628 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 2629 { 2630 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2631 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2632 struct spdk_nvmf_rdma_transport, transport); 2633 2634 nvmf_rdma_request_free(rdma_req, rtransport); 2635 return 0; 2636 } 2637 2638 static int 2639 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 2640 { 2641 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2642 struct spdk_nvmf_rdma_transport, transport); 2643 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 2644 struct spdk_nvmf_rdma_request, req); 2645 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 2646 struct spdk_nvmf_rdma_qpair, qpair); 2647 2648 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2649 /* The connection is alive, so process the request as normal */ 2650 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 2651 } else { 2652 /* The connection is dead. Move the request directly to the completed state. */ 2653 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2654 } 2655 2656 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2657 2658 return 0; 2659 } 2660 2661 static void 2662 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 2663 { 2664 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2665 struct ibv_recv_wr recv_wr = {}; 2666 struct ibv_recv_wr *bad_recv_wr; 2667 struct ibv_send_wr send_wr = {}; 2668 struct ibv_send_wr *bad_send_wr; 2669 int rc; 2670 2671 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 2672 return; 2673 } 2674 2675 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 2676 2677 /* This happens only when the qpair is disconnected before 2678 * it is added to the poll group. Since there is no poll group, 2679 * the RDMA qp has not been initialized yet and the RDMA CM 2680 * event has not yet been acknowledged, so we need to reject it. 2681 */ 2682 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 2683 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 2684 return; 2685 } 2686 2687 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2688 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 2689 } 2690 2691 rqpair->drain_recv_wr.type = RDMA_WR_TYPE_DRAIN_RECV; 2692 recv_wr.wr_id = (uintptr_t)&rqpair->drain_recv_wr; 2693 rc = ibv_post_recv(rqpair->cm_id->qp, &recv_wr, &bad_recv_wr); 2694 if (rc) { 2695 SPDK_ERRLOG("Failed to post dummy receive WR, errno %d\n", errno); 2696 assert(false); 2697 return; 2698 } 2699 2700 rqpair->drain_send_wr.type = RDMA_WR_TYPE_DRAIN_SEND; 2701 send_wr.wr_id = (uintptr_t)&rqpair->drain_send_wr; 2702 send_wr.opcode = IBV_WR_SEND; 2703 rc = ibv_post_send(rqpair->cm_id->qp, &send_wr, &bad_send_wr); 2704 if (rc) { 2705 SPDK_ERRLOG("Failed to post dummy send WR, errno %d\n", errno); 2706 assert(false); 2707 return; 2708 } 2709 rqpair->current_send_depth++; 2710 } 2711 2712 #ifdef DEBUG 2713 static int 2714 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 2715 { 2716 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 2717 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 2718 } 2719 #endif 2720 2721 static int 2722 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 2723 struct spdk_nvmf_rdma_poller *rpoller) 2724 { 2725 struct ibv_wc wc[32]; 2726 struct spdk_nvmf_rdma_wr *rdma_wr; 2727 struct spdk_nvmf_rdma_request *rdma_req; 2728 struct spdk_nvmf_rdma_recv *rdma_recv; 2729 struct spdk_nvmf_rdma_qpair *rqpair; 2730 int reaped, i; 2731 int count = 0; 2732 bool error = false; 2733 2734 /* Poll for completing operations. */ 2735 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 2736 if (reaped < 0) { 2737 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2738 errno, spdk_strerror(errno)); 2739 return -1; 2740 } 2741 2742 for (i = 0; i < reaped; i++) { 2743 2744 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 2745 2746 /* Handle error conditions */ 2747 if (wc[i].status) { 2748 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 2749 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 2750 2751 error = true; 2752 2753 switch (rdma_wr->type) { 2754 case RDMA_WR_TYPE_SEND: 2755 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 2756 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2757 2758 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 2759 /* We're going to attempt an error recovery, so force the request into 2760 * the completed state. */ 2761 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2762 rqpair->current_send_depth--; 2763 2764 assert(rdma_req->num_outstanding_data_wr == 0); 2765 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2766 break; 2767 case RDMA_WR_TYPE_RECV: 2768 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 2769 rqpair = rdma_recv->qpair; 2770 2771 /* Dump this into the incoming queue. This gets cleaned up when 2772 * the queue pair disconnects or recovers. */ 2773 STAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2774 rqpair->current_recv_depth++; 2775 2776 /* Don't worry about responding to recv overflow, we are disconnecting anyways */ 2777 break; 2778 case RDMA_WR_TYPE_DATA: 2779 /* If the data transfer fails still force the queue into the error state, 2780 * if we were performing an RDMA_READ, we need to force the request into a 2781 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 2782 * case, we should wait for the SEND to complete. */ 2783 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2784 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2785 2786 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 2787 rdma_req->num_outstanding_data_wr--; 2788 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 2789 assert(rdma_req->num_outstanding_data_wr > 0); 2790 rqpair->current_read_depth--; 2791 if (rdma_req->num_outstanding_data_wr == 0) { 2792 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2793 } 2794 } 2795 rqpair->current_send_depth--; 2796 break; 2797 case RDMA_WR_TYPE_DRAIN_RECV: 2798 rqpair = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_qpair, drain_recv_wr); 2799 assert(rqpair->disconnect_flags & RDMA_QP_DISCONNECTING); 2800 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Drained QP RECV %u (%p)\n", rqpair->qpair.qid, rqpair); 2801 rqpair->disconnect_flags |= RDMA_QP_RECV_DRAINED; 2802 assert(rqpair->current_recv_depth == rqpair->max_queue_depth); 2803 /* Don't worry about responding to recv overflow, we are disconnecting anyways */ 2804 if (rqpair->disconnect_flags & RDMA_QP_SEND_DRAINED) { 2805 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2806 spdk_nvmf_rdma_qpair_destroy(rqpair); 2807 } 2808 /* Continue so that this does not trigger the disconnect path below. */ 2809 continue; 2810 case RDMA_WR_TYPE_DRAIN_SEND: 2811 rqpair = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_qpair, drain_send_wr); 2812 assert(rqpair->disconnect_flags & RDMA_QP_DISCONNECTING); 2813 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Drained QP SEND %u (%p)\n", rqpair->qpair.qid, rqpair); 2814 rqpair->disconnect_flags |= RDMA_QP_SEND_DRAINED; 2815 rqpair->current_send_depth--; 2816 if (rqpair->disconnect_flags & RDMA_QP_RECV_DRAINED) { 2817 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2818 spdk_nvmf_rdma_qpair_destroy(rqpair); 2819 } 2820 /* Continue so that this does not trigger the disconnect path below. */ 2821 continue; 2822 default: 2823 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2824 continue; 2825 } 2826 2827 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 2828 /* Disconnect the connection. */ 2829 spdk_nvmf_rdma_start_disconnect(rqpair); 2830 } 2831 continue; 2832 } 2833 2834 switch (wc[i].opcode) { 2835 case IBV_WC_SEND: 2836 assert(rdma_wr->type == RDMA_WR_TYPE_SEND); 2837 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 2838 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2839 2840 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 2841 2842 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2843 rqpair->current_send_depth--; 2844 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2845 2846 count++; 2847 2848 assert(rdma_req->num_outstanding_data_wr == 0); 2849 /* Try to process other queued requests */ 2850 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 2851 break; 2852 2853 case IBV_WC_RDMA_WRITE: 2854 assert(rdma_wr->type == RDMA_WR_TYPE_DATA); 2855 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2856 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2857 rqpair->current_send_depth--; 2858 rdma_req->num_outstanding_data_wr--; 2859 2860 /* Try to process other queued requests */ 2861 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 2862 break; 2863 2864 case IBV_WC_RDMA_READ: 2865 assert(rdma_wr->type == RDMA_WR_TYPE_DATA); 2866 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2867 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2868 rqpair->current_send_depth--; 2869 2870 assert(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2871 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 2872 assert(rdma_req->num_outstanding_data_wr > 0); 2873 rqpair->current_read_depth--; 2874 rdma_req->num_outstanding_data_wr--; 2875 if (rdma_req->num_outstanding_data_wr == 0) { 2876 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2877 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2878 } 2879 2880 /* Try to process other queued requests */ 2881 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 2882 break; 2883 2884 case IBV_WC_RECV: 2885 assert(rdma_wr->type == RDMA_WR_TYPE_RECV); 2886 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 2887 rqpair = rdma_recv->qpair; 2888 /* The qpair should not send more requests than are allowed per qpair. */ 2889 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 2890 spdk_nvmf_rdma_start_disconnect(rqpair); 2891 } else { 2892 rqpair->current_recv_depth++; 2893 } 2894 STAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2895 /* Try to process other queued requests */ 2896 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 2897 break; 2898 2899 default: 2900 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2901 continue; 2902 } 2903 } 2904 2905 if (error == true) { 2906 return -1; 2907 } 2908 2909 return count; 2910 } 2911 2912 static int 2913 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 2914 { 2915 struct spdk_nvmf_rdma_transport *rtransport; 2916 struct spdk_nvmf_rdma_poll_group *rgroup; 2917 struct spdk_nvmf_rdma_poller *rpoller; 2918 int count, rc; 2919 2920 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 2921 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2922 2923 count = 0; 2924 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2925 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 2926 if (rc < 0) { 2927 return rc; 2928 } 2929 count += rc; 2930 } 2931 2932 return count; 2933 } 2934 2935 static int 2936 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2937 struct spdk_nvme_transport_id *trid, 2938 bool peer) 2939 { 2940 struct sockaddr *saddr; 2941 uint16_t port; 2942 2943 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 2944 2945 if (peer) { 2946 saddr = rdma_get_peer_addr(id); 2947 } else { 2948 saddr = rdma_get_local_addr(id); 2949 } 2950 switch (saddr->sa_family) { 2951 case AF_INET: { 2952 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 2953 2954 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 2955 inet_ntop(AF_INET, &saddr_in->sin_addr, 2956 trid->traddr, sizeof(trid->traddr)); 2957 if (peer) { 2958 port = ntohs(rdma_get_dst_port(id)); 2959 } else { 2960 port = ntohs(rdma_get_src_port(id)); 2961 } 2962 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 2963 break; 2964 } 2965 case AF_INET6: { 2966 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 2967 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 2968 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 2969 trid->traddr, sizeof(trid->traddr)); 2970 if (peer) { 2971 port = ntohs(rdma_get_dst_port(id)); 2972 } else { 2973 port = ntohs(rdma_get_src_port(id)); 2974 } 2975 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 2976 break; 2977 } 2978 default: 2979 return -1; 2980 2981 } 2982 2983 return 0; 2984 } 2985 2986 static int 2987 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 2988 struct spdk_nvme_transport_id *trid) 2989 { 2990 struct spdk_nvmf_rdma_qpair *rqpair; 2991 2992 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2993 2994 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 2995 } 2996 2997 static int 2998 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 2999 struct spdk_nvme_transport_id *trid) 3000 { 3001 struct spdk_nvmf_rdma_qpair *rqpair; 3002 3003 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3004 3005 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 3006 } 3007 3008 static int 3009 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3010 struct spdk_nvme_transport_id *trid) 3011 { 3012 struct spdk_nvmf_rdma_qpair *rqpair; 3013 3014 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3015 3016 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 3017 } 3018 3019 void 3020 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3021 { 3022 g_nvmf_hooks = *hooks; 3023 } 3024 3025 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 3026 .type = SPDK_NVME_TRANSPORT_RDMA, 3027 .opts_init = spdk_nvmf_rdma_opts_init, 3028 .create = spdk_nvmf_rdma_create, 3029 .destroy = spdk_nvmf_rdma_destroy, 3030 3031 .listen = spdk_nvmf_rdma_listen, 3032 .stop_listen = spdk_nvmf_rdma_stop_listen, 3033 .accept = spdk_nvmf_rdma_accept, 3034 3035 .listener_discover = spdk_nvmf_rdma_discover, 3036 3037 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 3038 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 3039 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 3040 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 3041 3042 .req_free = spdk_nvmf_rdma_request_free, 3043 .req_complete = spdk_nvmf_rdma_request_complete, 3044 3045 .qpair_fini = spdk_nvmf_rdma_close_qpair, 3046 .qpair_is_idle = spdk_nvmf_rdma_qpair_is_idle, 3047 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 3048 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 3049 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 3050 3051 }; 3052 3053 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 3054