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 /* Timeout for destroying defunct rqpairs */ 68 #define NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US 4000000 69 70 enum spdk_nvmf_rdma_request_state { 71 /* The request is not currently in use */ 72 RDMA_REQUEST_STATE_FREE = 0, 73 74 /* Initial state when request first received */ 75 RDMA_REQUEST_STATE_NEW, 76 77 /* The request is queued until a data buffer is available. */ 78 RDMA_REQUEST_STATE_NEED_BUFFER, 79 80 /* The request is waiting on RDMA queue depth availability 81 * to transfer data from the host to the controller. 82 */ 83 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 84 85 /* The request is currently transferring data from the host to the controller. */ 86 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 87 88 /* The request is ready to execute at the block device */ 89 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 90 91 /* The request is currently executing at the block device */ 92 RDMA_REQUEST_STATE_EXECUTING, 93 94 /* The request finished executing at the block device */ 95 RDMA_REQUEST_STATE_EXECUTED, 96 97 /* The request is waiting on RDMA queue depth availability 98 * to transfer data from the controller to the host. 99 */ 100 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 101 102 /* The request is ready to send a completion */ 103 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 104 105 /* The request is currently transferring data from the controller to the host. */ 106 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 107 108 /* The request currently has an outstanding completion without an 109 * associated data transfer. 110 */ 111 RDMA_REQUEST_STATE_COMPLETING, 112 113 /* The request completed and can be marked free. */ 114 RDMA_REQUEST_STATE_COMPLETED, 115 116 /* Terminator */ 117 RDMA_REQUEST_NUM_STATES, 118 }; 119 120 #define OBJECT_NVMF_RDMA_IO 0x40 121 122 #define TRACE_GROUP_NVMF_RDMA 0x4 123 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 124 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 125 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 126 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 127 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 128 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 129 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 130 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 131 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 132 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 133 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 134 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 135 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 136 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 137 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 138 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 139 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 140 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 141 142 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 143 { 144 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 145 spdk_trace_register_description("RDMA_REQ_NEW", "", 146 TRACE_RDMA_REQUEST_STATE_NEW, 147 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 148 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", "", 149 TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 150 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 151 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C_TO_H", "", 152 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 153 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 154 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H_TO_C", "", 155 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 156 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 157 spdk_trace_register_description("RDMA_REQ_TX_H_TO_C", "", 158 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 159 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 160 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", "", 161 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 162 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 163 spdk_trace_register_description("RDMA_REQ_EXECUTING", "", 164 TRACE_RDMA_REQUEST_STATE_EXECUTING, 165 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 166 spdk_trace_register_description("RDMA_REQ_EXECUTED", "", 167 TRACE_RDMA_REQUEST_STATE_EXECUTED, 168 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 169 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPLETE", "", 170 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 171 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 172 spdk_trace_register_description("RDMA_REQ_COMPLETING_CONTROLLER_TO_HOST", "", 173 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 174 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 175 spdk_trace_register_description("RDMA_REQ_COMPLETING_INCAPSULE", "", 176 TRACE_RDMA_REQUEST_STATE_COMPLETING, 177 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 178 spdk_trace_register_description("RDMA_REQ_COMPLETED", "", 179 TRACE_RDMA_REQUEST_STATE_COMPLETED, 180 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 181 182 spdk_trace_register_description("RDMA_QP_CREATE", "", TRACE_RDMA_QP_CREATE, 183 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 184 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", "", TRACE_RDMA_IBV_ASYNC_EVENT, 185 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 186 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", "", TRACE_RDMA_CM_ASYNC_EVENT, 187 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 188 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", "", TRACE_RDMA_QP_STATE_CHANGE, 189 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 190 spdk_trace_register_description("RDMA_QP_DISCONNECT", "", TRACE_RDMA_QP_DISCONNECT, 191 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 192 spdk_trace_register_description("RDMA_QP_DESTROY", "", TRACE_RDMA_QP_DESTROY, 193 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 194 } 195 196 enum spdk_nvmf_rdma_wr_type { 197 RDMA_WR_TYPE_RECV, 198 RDMA_WR_TYPE_SEND, 199 RDMA_WR_TYPE_DATA, 200 RDMA_WR_TYPE_DRAIN_SEND, 201 RDMA_WR_TYPE_DRAIN_RECV 202 }; 203 204 struct spdk_nvmf_rdma_wr { 205 enum spdk_nvmf_rdma_wr_type type; 206 }; 207 208 /* This structure holds commands as they are received off the wire. 209 * It must be dynamically paired with a full request object 210 * (spdk_nvmf_rdma_request) to service a request. It is separate 211 * from the request because RDMA does not appear to order 212 * completions, so occasionally we'll get a new incoming 213 * command when there aren't any free request objects. 214 */ 215 struct spdk_nvmf_rdma_recv { 216 struct ibv_recv_wr wr; 217 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 218 219 struct spdk_nvmf_rdma_qpair *qpair; 220 221 /* In-capsule data buffer */ 222 uint8_t *buf; 223 224 struct spdk_nvmf_rdma_wr rdma_wr; 225 226 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 227 }; 228 229 struct spdk_nvmf_rdma_request_data { 230 struct spdk_nvmf_rdma_wr rdma_wr; 231 struct ibv_send_wr wr; 232 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 233 void *buffers[SPDK_NVMF_MAX_SGL_ENTRIES]; 234 }; 235 236 struct spdk_nvmf_rdma_request { 237 struct spdk_nvmf_request req; 238 bool data_from_pool; 239 240 enum spdk_nvmf_rdma_request_state state; 241 242 struct spdk_nvmf_rdma_recv *recv; 243 244 struct { 245 struct spdk_nvmf_rdma_wr rdma_wr; 246 struct ibv_send_wr wr; 247 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 248 } rsp; 249 250 struct spdk_nvmf_rdma_request_data data; 251 252 uint32_t num_outstanding_data_wr; 253 254 TAILQ_ENTRY(spdk_nvmf_rdma_request) link; 255 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 256 }; 257 258 enum spdk_nvmf_rdma_qpair_disconnect_flags { 259 RDMA_QP_DISCONNECTING = 1, 260 RDMA_QP_RECV_DRAINED = 1 << 1, 261 RDMA_QP_SEND_DRAINED = 1 << 2 262 }; 263 264 struct spdk_nvmf_rdma_qpair { 265 struct spdk_nvmf_qpair qpair; 266 267 struct spdk_nvmf_rdma_port *port; 268 struct spdk_nvmf_rdma_poller *poller; 269 270 struct rdma_cm_id *cm_id; 271 struct rdma_cm_id *listen_id; 272 273 /* The maximum number of I/O outstanding on this connection at one time */ 274 uint16_t max_queue_depth; 275 276 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 277 uint16_t max_read_depth; 278 279 /* The maximum number of RDMA SEND operations at one time */ 280 uint32_t max_send_depth; 281 282 /* The current number of outstanding WRs from this qpair's 283 * recv queue. Should not exceed device->attr.max_queue_depth. 284 */ 285 uint16_t current_recv_depth; 286 287 /* The current number of posted WRs from this qpair's 288 * send queue. Should not exceed max_send_depth. 289 */ 290 uint32_t current_send_depth; 291 292 /* The current number of active RDMA READ operations */ 293 uint16_t current_read_depth; 294 295 /* The maximum number of SGEs per WR on the send queue */ 296 uint32_t max_send_sge; 297 298 /* The maximum number of SGEs per WR on the recv queue */ 299 uint32_t max_recv_sge; 300 301 /* Receives that are waiting for a request object */ 302 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 303 304 /* Queues to track requests in critical states */ 305 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 306 307 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 308 309 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 310 311 /* Number of requests not in the free state */ 312 uint32_t qd; 313 314 /* Array of size "max_queue_depth" containing RDMA requests. */ 315 struct spdk_nvmf_rdma_request *reqs; 316 317 /* Array of size "max_queue_depth" containing RDMA recvs. */ 318 struct spdk_nvmf_rdma_recv *recvs; 319 320 /* Array of size "max_queue_depth" containing 64 byte capsules 321 * used for receive. 322 */ 323 union nvmf_h2c_msg *cmds; 324 struct ibv_mr *cmds_mr; 325 326 /* Array of size "max_queue_depth" containing 16 byte completions 327 * to be sent back to the user. 328 */ 329 union nvmf_c2h_msg *cpls; 330 struct ibv_mr *cpls_mr; 331 332 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 333 * buffers to be used for in capsule data. 334 */ 335 void *bufs; 336 struct ibv_mr *bufs_mr; 337 338 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 339 340 /* IBV queue pair attributes: they are used to manage 341 * qp state and recover from errors. 342 */ 343 struct ibv_qp_attr ibv_attr; 344 345 uint32_t disconnect_flags; 346 struct spdk_nvmf_rdma_wr drain_send_wr; 347 struct spdk_nvmf_rdma_wr drain_recv_wr; 348 349 /* Poller registered in case the qpair doesn't properly 350 * complete the qpair destruct process and becomes defunct. 351 */ 352 353 struct spdk_poller *destruct_poller; 354 355 /* There are several ways a disconnect can start on a qpair 356 * and they are not all mutually exclusive. It is important 357 * that we only initialize one of these paths. 358 */ 359 bool disconnect_started; 360 }; 361 362 struct spdk_nvmf_rdma_poller { 363 struct spdk_nvmf_rdma_device *device; 364 struct spdk_nvmf_rdma_poll_group *group; 365 366 int num_cqe; 367 int required_num_wr; 368 struct ibv_cq *cq; 369 370 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 371 372 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 373 }; 374 375 struct spdk_nvmf_rdma_poll_group { 376 struct spdk_nvmf_transport_poll_group group; 377 378 /* Requests that are waiting to obtain a data buffer */ 379 TAILQ_HEAD(, spdk_nvmf_rdma_request) pending_data_buf_queue; 380 381 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 382 }; 383 384 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 385 struct spdk_nvmf_rdma_device { 386 struct ibv_device_attr attr; 387 struct ibv_context *context; 388 389 struct spdk_mem_map *map; 390 struct ibv_pd *pd; 391 392 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 393 }; 394 395 struct spdk_nvmf_rdma_port { 396 struct spdk_nvme_transport_id trid; 397 struct rdma_cm_id *id; 398 struct spdk_nvmf_rdma_device *device; 399 uint32_t ref; 400 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 401 }; 402 403 struct spdk_nvmf_rdma_transport { 404 struct spdk_nvmf_transport transport; 405 406 struct rdma_event_channel *event_channel; 407 408 struct spdk_mempool *data_wr_pool; 409 410 pthread_mutex_t lock; 411 412 /* fields used to poll RDMA/IB events */ 413 nfds_t npoll_fds; 414 struct pollfd *poll_fds; 415 416 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 417 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 418 }; 419 420 static inline int 421 spdk_nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 422 { 423 switch (state) { 424 case IBV_QPS_RESET: 425 case IBV_QPS_INIT: 426 case IBV_QPS_RTR: 427 case IBV_QPS_RTS: 428 case IBV_QPS_SQD: 429 case IBV_QPS_SQE: 430 case IBV_QPS_ERR: 431 return 0; 432 default: 433 return -1; 434 } 435 } 436 437 static enum ibv_qp_state 438 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 439 enum ibv_qp_state old_state, new_state; 440 struct ibv_qp_init_attr init_attr; 441 int rc; 442 443 /* All the attributes needed for recovery */ 444 static int spdk_nvmf_ibv_attr_mask = 445 IBV_QP_STATE | 446 IBV_QP_PKEY_INDEX | 447 IBV_QP_PORT | 448 IBV_QP_ACCESS_FLAGS | 449 IBV_QP_AV | 450 IBV_QP_PATH_MTU | 451 IBV_QP_DEST_QPN | 452 IBV_QP_RQ_PSN | 453 IBV_QP_MAX_DEST_RD_ATOMIC | 454 IBV_QP_MIN_RNR_TIMER | 455 IBV_QP_SQ_PSN | 456 IBV_QP_TIMEOUT | 457 IBV_QP_RETRY_CNT | 458 IBV_QP_RNR_RETRY | 459 IBV_QP_MAX_QP_RD_ATOMIC; 460 461 old_state = rqpair->ibv_attr.qp_state; 462 rc = ibv_query_qp(rqpair->cm_id->qp, &rqpair->ibv_attr, 463 spdk_nvmf_ibv_attr_mask, &init_attr); 464 465 if (rc) 466 { 467 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 468 assert(false); 469 } 470 471 new_state = rqpair->ibv_attr.qp_state; 472 473 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 474 if (rc) 475 { 476 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 477 /* 478 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 479 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 480 */ 481 return IBV_QPS_ERR + 1; 482 } 483 484 if (old_state != new_state) 485 { 486 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 487 (uintptr_t)rqpair->cm_id, new_state); 488 } 489 return new_state; 490 } 491 492 static const char *str_ibv_qp_state[] = { 493 "IBV_QPS_RESET", 494 "IBV_QPS_INIT", 495 "IBV_QPS_RTR", 496 "IBV_QPS_RTS", 497 "IBV_QPS_SQD", 498 "IBV_QPS_SQE", 499 "IBV_QPS_ERR", 500 "IBV_QPS_UNKNOWN" 501 }; 502 503 static int 504 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 505 enum ibv_qp_state new_state) 506 { 507 int rc; 508 enum ibv_qp_state state; 509 static int attr_mask_rc[] = { 510 [IBV_QPS_RESET] = IBV_QP_STATE, 511 [IBV_QPS_INIT] = (IBV_QP_STATE | 512 IBV_QP_PKEY_INDEX | 513 IBV_QP_PORT | 514 IBV_QP_ACCESS_FLAGS), 515 [IBV_QPS_RTR] = (IBV_QP_STATE | 516 IBV_QP_AV | 517 IBV_QP_PATH_MTU | 518 IBV_QP_DEST_QPN | 519 IBV_QP_RQ_PSN | 520 IBV_QP_MAX_DEST_RD_ATOMIC | 521 IBV_QP_MIN_RNR_TIMER), 522 [IBV_QPS_RTS] = (IBV_QP_STATE | 523 IBV_QP_SQ_PSN | 524 IBV_QP_TIMEOUT | 525 IBV_QP_RETRY_CNT | 526 IBV_QP_RNR_RETRY | 527 IBV_QP_MAX_QP_RD_ATOMIC), 528 [IBV_QPS_SQD] = IBV_QP_STATE, 529 [IBV_QPS_SQE] = IBV_QP_STATE, 530 [IBV_QPS_ERR] = IBV_QP_STATE, 531 }; 532 533 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 534 if (rc) { 535 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 536 rqpair->qpair.qid, new_state); 537 return rc; 538 } 539 540 rqpair->ibv_attr.cur_qp_state = rqpair->ibv_attr.qp_state; 541 rqpair->ibv_attr.qp_state = new_state; 542 rqpair->ibv_attr.ah_attr.port_num = rqpair->ibv_attr.port_num; 543 544 rc = ibv_modify_qp(rqpair->cm_id->qp, &rqpair->ibv_attr, 545 attr_mask_rc[new_state]); 546 547 if (rc) { 548 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 549 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 550 return rc; 551 } 552 553 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 554 555 if (state != new_state) { 556 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 557 rqpair->qpair.qid, str_ibv_qp_state[new_state], 558 str_ibv_qp_state[state]); 559 return -1; 560 } 561 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 562 str_ibv_qp_state[state]); 563 return 0; 564 } 565 566 static void 567 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 568 { 569 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->data_from_pool); 570 if (req->req.cmd) { 571 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 572 } 573 if (req->recv) { 574 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 575 } 576 } 577 578 static void 579 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 580 { 581 int i; 582 583 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 584 for (i = 0; i < rqpair->max_queue_depth; i++) { 585 if (rqpair->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 586 nvmf_rdma_dump_request(&rqpair->reqs[i]); 587 } 588 } 589 } 590 591 static void 592 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 593 { 594 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 595 596 spdk_poller_unregister(&rqpair->destruct_poller); 597 598 if (rqpair->qd != 0) { 599 nvmf_rdma_dump_qpair_contents(rqpair); 600 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 601 } 602 603 if (rqpair->poller) { 604 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 605 } 606 607 if (rqpair->cmds_mr) { 608 ibv_dereg_mr(rqpair->cmds_mr); 609 } 610 611 if (rqpair->cpls_mr) { 612 ibv_dereg_mr(rqpair->cpls_mr); 613 } 614 615 if (rqpair->bufs_mr) { 616 ibv_dereg_mr(rqpair->bufs_mr); 617 } 618 619 if (rqpair->cm_id) { 620 rdma_destroy_qp(rqpair->cm_id); 621 rdma_destroy_id(rqpair->cm_id); 622 623 if (rqpair->poller) { 624 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 625 } 626 } 627 628 /* Free all memory */ 629 spdk_dma_free(rqpair->cmds); 630 spdk_dma_free(rqpair->cpls); 631 spdk_dma_free(rqpair->bufs); 632 free(rqpair->reqs); 633 free(rqpair->recvs); 634 free(rqpair); 635 } 636 637 static int 638 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 639 { 640 struct spdk_nvmf_rdma_transport *rtransport; 641 struct spdk_nvmf_rdma_qpair *rqpair; 642 struct spdk_nvmf_rdma_poller *rpoller; 643 int rc, i, num_cqe, required_num_wr;; 644 struct spdk_nvmf_rdma_recv *rdma_recv; 645 struct spdk_nvmf_rdma_request *rdma_req; 646 struct spdk_nvmf_transport *transport; 647 struct spdk_nvmf_rdma_device *device; 648 struct ibv_qp_init_attr ibv_init_attr; 649 650 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 651 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 652 transport = &rtransport->transport; 653 device = rqpair->port->device; 654 655 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 656 ibv_init_attr.qp_context = rqpair; 657 ibv_init_attr.qp_type = IBV_QPT_RC; 658 ibv_init_attr.send_cq = rqpair->poller->cq; 659 ibv_init_attr.recv_cq = rqpair->poller->cq; 660 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 661 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 662 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 663 1; /* RECV operations + dummy drain WR */ 664 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 665 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 666 667 /* Enlarge CQ size dynamically */ 668 rpoller = rqpair->poller; 669 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 670 num_cqe = rpoller->num_cqe; 671 if (num_cqe < required_num_wr) { 672 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 673 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 674 } 675 676 if (rpoller->num_cqe != num_cqe) { 677 if (required_num_wr > device->attr.max_cqe) { 678 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 679 required_num_wr, device->attr.max_cqe); 680 rdma_destroy_id(rqpair->cm_id); 681 rqpair->cm_id = NULL; 682 spdk_nvmf_rdma_qpair_destroy(rqpair); 683 return -1; 684 } 685 686 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 687 rc = ibv_resize_cq(rpoller->cq, num_cqe); 688 if (rc) { 689 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 690 rdma_destroy_id(rqpair->cm_id); 691 rqpair->cm_id = NULL; 692 spdk_nvmf_rdma_qpair_destroy(rqpair); 693 return -1; 694 } 695 696 rpoller->num_cqe = num_cqe; 697 } 698 699 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 700 if (rc) { 701 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 702 rdma_destroy_id(rqpair->cm_id); 703 rqpair->cm_id = NULL; 704 spdk_nvmf_rdma_qpair_destroy(rqpair); 705 return -1; 706 } 707 708 rpoller->required_num_wr = required_num_wr; 709 710 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2 + 1), 711 ibv_init_attr.cap.max_send_wr); 712 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 713 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 714 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 715 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 716 717 rqpair->reqs = calloc(rqpair->max_queue_depth, sizeof(*rqpair->reqs)); 718 rqpair->recvs = calloc(rqpair->max_queue_depth, sizeof(*rqpair->recvs)); 719 rqpair->cmds = spdk_dma_zmalloc(rqpair->max_queue_depth * sizeof(*rqpair->cmds), 720 0x1000, NULL); 721 rqpair->cpls = spdk_dma_zmalloc(rqpair->max_queue_depth * sizeof(*rqpair->cpls), 722 0x1000, NULL); 723 724 725 if (transport->opts.in_capsule_data_size > 0) { 726 rqpair->bufs = spdk_dma_zmalloc(rqpair->max_queue_depth * 727 transport->opts.in_capsule_data_size, 728 0x1000, NULL); 729 } 730 731 if (!rqpair->reqs || !rqpair->recvs || !rqpair->cmds || 732 !rqpair->cpls || (transport->opts.in_capsule_data_size && !rqpair->bufs)) { 733 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 734 spdk_nvmf_rdma_qpair_destroy(rqpair); 735 return -1; 736 } 737 738 rqpair->cmds_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->cmds, 739 rqpair->max_queue_depth * sizeof(*rqpair->cmds), 740 IBV_ACCESS_LOCAL_WRITE); 741 rqpair->cpls_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->cpls, 742 rqpair->max_queue_depth * sizeof(*rqpair->cpls), 743 0); 744 745 if (transport->opts.in_capsule_data_size) { 746 rqpair->bufs_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->bufs, 747 rqpair->max_queue_depth * 748 transport->opts.in_capsule_data_size, 749 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 750 } 751 752 if (!rqpair->cmds_mr || !rqpair->cpls_mr || (transport->opts.in_capsule_data_size && 753 !rqpair->bufs_mr)) { 754 SPDK_ERRLOG("Unable to register required memory for RDMA queue.\n"); 755 spdk_nvmf_rdma_qpair_destroy(rqpair); 756 return -1; 757 } 758 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 759 rqpair->cmds, rqpair->max_queue_depth * sizeof(*rqpair->cmds), rqpair->cmds_mr->lkey); 760 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 761 rqpair->cpls, rqpair->max_queue_depth * sizeof(*rqpair->cpls), rqpair->cpls_mr->lkey); 762 if (rqpair->bufs && rqpair->bufs_mr) { 763 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 764 rqpair->bufs, rqpair->max_queue_depth * 765 transport->opts.in_capsule_data_size, rqpair->bufs_mr->lkey); 766 } 767 768 STAILQ_INIT(&rqpair->free_queue); 769 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 770 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 771 772 rqpair->current_recv_depth = rqpair->max_queue_depth; 773 for (i = 0; i < rqpair->max_queue_depth; i++) { 774 struct ibv_recv_wr *bad_wr = NULL; 775 776 rdma_recv = &rqpair->recvs[i]; 777 rdma_recv->qpair = rqpair; 778 779 /* Set up memory to receive commands */ 780 if (rqpair->bufs) { 781 rdma_recv->buf = (void *)((uintptr_t)rqpair->bufs + (i * 782 transport->opts.in_capsule_data_size)); 783 } 784 785 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 786 787 rdma_recv->sgl[0].addr = (uintptr_t)&rqpair->cmds[i]; 788 rdma_recv->sgl[0].length = sizeof(rqpair->cmds[i]); 789 rdma_recv->sgl[0].lkey = rqpair->cmds_mr->lkey; 790 rdma_recv->wr.num_sge = 1; 791 792 if (rdma_recv->buf && rqpair->bufs_mr) { 793 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 794 rdma_recv->sgl[1].length = transport->opts.in_capsule_data_size; 795 rdma_recv->sgl[1].lkey = rqpair->bufs_mr->lkey; 796 rdma_recv->wr.num_sge++; 797 } 798 799 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 800 rdma_recv->wr.sg_list = rdma_recv->sgl; 801 802 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_recv->wr, &bad_wr); 803 assert(rqpair->current_recv_depth > 0); 804 rqpair->current_recv_depth--; 805 if (rc) { 806 SPDK_ERRLOG("Unable to post capsule for RDMA RECV\n"); 807 spdk_nvmf_rdma_qpair_destroy(rqpair); 808 return -1; 809 } 810 } 811 assert(rqpair->current_recv_depth == 0); 812 813 for (i = 0; i < rqpair->max_queue_depth; i++) { 814 rdma_req = &rqpair->reqs[i]; 815 816 rdma_req->req.qpair = &rqpair->qpair; 817 rdma_req->req.cmd = NULL; 818 819 /* Set up memory to send responses */ 820 rdma_req->req.rsp = &rqpair->cpls[i]; 821 822 rdma_req->rsp.sgl[0].addr = (uintptr_t)&rqpair->cpls[i]; 823 rdma_req->rsp.sgl[0].length = sizeof(rqpair->cpls[i]); 824 rdma_req->rsp.sgl[0].lkey = rqpair->cpls_mr->lkey; 825 826 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 827 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 828 rdma_req->rsp.wr.next = NULL; 829 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 830 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 831 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 832 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 833 834 /* Set up memory for data buffers */ 835 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 836 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 837 rdma_req->data.wr.next = NULL; 838 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 839 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 840 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 841 842 /* Initialize request state to FREE */ 843 rdma_req->state = RDMA_REQUEST_STATE_FREE; 844 STAILQ_INSERT_HEAD(&rqpair->free_queue, rdma_req, state_link); 845 } 846 847 return 0; 848 } 849 850 static int 851 request_transfer_in(struct spdk_nvmf_request *req) 852 { 853 int rc; 854 struct spdk_nvmf_rdma_request *rdma_req; 855 struct spdk_nvmf_qpair *qpair; 856 struct spdk_nvmf_rdma_qpair *rqpair; 857 struct ibv_send_wr *bad_wr = NULL; 858 859 qpair = req->qpair; 860 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 861 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 862 863 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 864 assert(rdma_req != NULL); 865 866 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA READ POSTED. Request: %p Connection: %p\n", req, qpair); 867 868 rc = ibv_post_send(rqpair->cm_id->qp, &rdma_req->data.wr, &bad_wr); 869 if (rc) { 870 SPDK_ERRLOG("Unable to transfer data from host to target\n"); 871 return -1; 872 } 873 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 874 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 875 return 0; 876 } 877 878 static int 879 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 880 { 881 int rc; 882 struct spdk_nvmf_rdma_request *rdma_req; 883 struct spdk_nvmf_qpair *qpair; 884 struct spdk_nvmf_rdma_qpair *rqpair; 885 struct spdk_nvme_cpl *rsp; 886 struct ibv_recv_wr *bad_recv_wr = NULL; 887 struct ibv_send_wr *send_wr, *bad_send_wr = NULL; 888 889 *data_posted = 0; 890 qpair = req->qpair; 891 rsp = &req->rsp->nvme_cpl; 892 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 893 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 894 895 /* Advance our sq_head pointer */ 896 if (qpair->sq_head == qpair->sq_head_max) { 897 qpair->sq_head = 0; 898 } else { 899 qpair->sq_head++; 900 } 901 rsp->sqhd = qpair->sq_head; 902 903 /* Post the capsule to the recv buffer */ 904 assert(rdma_req->recv != NULL); 905 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA RECV POSTED. Recv: %p Connection: %p\n", rdma_req->recv, 906 rqpair); 907 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_req->recv->wr, &bad_recv_wr); 908 if (rc) { 909 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 910 return rc; 911 } 912 rdma_req->recv = NULL; 913 assert(rqpair->current_recv_depth > 0); 914 rqpair->current_recv_depth--; 915 916 /* Build the response which consists of an optional 917 * RDMA WRITE to transfer data, plus an RDMA SEND 918 * containing the response. 919 */ 920 send_wr = &rdma_req->rsp.wr; 921 922 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 923 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 924 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA WRITE POSTED. Request: %p Connection: %p\n", req, qpair); 925 send_wr = &rdma_req->data.wr; 926 *data_posted = 1; 927 } 928 929 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA SEND POSTED. Request: %p Connection: %p\n", req, qpair); 930 931 /* Send the completion */ 932 rc = ibv_post_send(rqpair->cm_id->qp, send_wr, &bad_send_wr); 933 if (rc) { 934 SPDK_ERRLOG("Unable to send response capsule\n"); 935 return rc; 936 } 937 /* +1 for the rsp wr */ 938 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr + 1; 939 940 return 0; 941 } 942 943 static int 944 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 945 { 946 struct spdk_nvmf_rdma_accept_private_data accept_data; 947 struct rdma_conn_param ctrlr_event_data = {}; 948 int rc; 949 950 accept_data.recfmt = 0; 951 accept_data.crqsize = rqpair->max_queue_depth; 952 953 ctrlr_event_data.private_data = &accept_data; 954 ctrlr_event_data.private_data_len = sizeof(accept_data); 955 if (id->ps == RDMA_PS_TCP) { 956 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 957 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 958 } 959 960 rc = rdma_accept(id, &ctrlr_event_data); 961 if (rc) { 962 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 963 } else { 964 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 965 } 966 967 return rc; 968 } 969 970 static void 971 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 972 { 973 struct spdk_nvmf_rdma_reject_private_data rej_data; 974 975 rej_data.recfmt = 0; 976 rej_data.sts = error; 977 978 rdma_reject(id, &rej_data, sizeof(rej_data)); 979 } 980 981 static int 982 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 983 new_qpair_fn cb_fn) 984 { 985 struct spdk_nvmf_rdma_transport *rtransport; 986 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 987 struct spdk_nvmf_rdma_port *port; 988 struct rdma_conn_param *rdma_param = NULL; 989 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 990 uint16_t max_queue_depth; 991 uint16_t max_read_depth; 992 993 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 994 995 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 996 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 997 998 rdma_param = &event->param.conn; 999 if (rdma_param->private_data == NULL || 1000 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1001 SPDK_ERRLOG("connect request: no private data provided\n"); 1002 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1003 return -1; 1004 } 1005 1006 private_data = rdma_param->private_data; 1007 if (private_data->recfmt != 0) { 1008 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1009 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1010 return -1; 1011 } 1012 1013 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1014 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1015 1016 port = event->listen_id->context; 1017 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1018 event->listen_id, event->listen_id->verbs, port); 1019 1020 /* Figure out the supported queue depth. This is a multi-step process 1021 * that takes into account hardware maximums, host provided values, 1022 * and our target's internal memory limits */ 1023 1024 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1025 1026 /* Start with the maximum queue depth allowed by the target */ 1027 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1028 max_read_depth = rtransport->transport.opts.max_queue_depth; 1029 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1030 rtransport->transport.opts.max_queue_depth); 1031 1032 /* Next check the local NIC's hardware limitations */ 1033 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1034 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1035 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1036 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1037 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1038 1039 /* Next check the remote NIC's hardware limitations */ 1040 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1041 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1042 rdma_param->initiator_depth, rdma_param->responder_resources); 1043 if (rdma_param->initiator_depth > 0) { 1044 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1045 } 1046 1047 /* Finally check for the host software requested values, which are 1048 * optional. */ 1049 if (rdma_param->private_data != NULL && 1050 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1051 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1052 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1053 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1054 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1055 } 1056 1057 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1058 max_queue_depth, max_read_depth); 1059 1060 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1061 if (rqpair == NULL) { 1062 SPDK_ERRLOG("Could not allocate new connection.\n"); 1063 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1064 return -1; 1065 } 1066 1067 rqpair->port = port; 1068 rqpair->max_queue_depth = max_queue_depth; 1069 rqpair->max_read_depth = max_read_depth; 1070 rqpair->cm_id = event->id; 1071 rqpair->listen_id = event->listen_id; 1072 rqpair->qpair.transport = transport; 1073 STAILQ_INIT(&rqpair->incoming_queue); 1074 event->id->context = &rqpair->qpair; 1075 1076 cb_fn(&rqpair->qpair); 1077 1078 return 0; 1079 } 1080 1081 static int 1082 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1083 enum spdk_mem_map_notify_action action, 1084 void *vaddr, size_t size) 1085 { 1086 struct ibv_pd *pd = cb_ctx; 1087 struct ibv_mr *mr; 1088 1089 switch (action) { 1090 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1091 if (!g_nvmf_hooks.get_rkey) { 1092 mr = ibv_reg_mr(pd, vaddr, size, 1093 IBV_ACCESS_LOCAL_WRITE | 1094 IBV_ACCESS_REMOTE_READ | 1095 IBV_ACCESS_REMOTE_WRITE); 1096 if (mr == NULL) { 1097 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1098 return -1; 1099 } else { 1100 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1101 } 1102 } else { 1103 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1104 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1105 } 1106 break; 1107 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1108 if (!g_nvmf_hooks.get_rkey) { 1109 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1110 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1111 if (mr) { 1112 ibv_dereg_mr(mr); 1113 } 1114 } 1115 break; 1116 } 1117 1118 return 0; 1119 } 1120 1121 static int 1122 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1123 { 1124 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1125 return addr_1 == addr_2; 1126 } 1127 1128 static void 1129 spdk_nvmf_rdma_request_free_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1130 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport) 1131 { 1132 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 1133 if (group->buf_cache_count < group->buf_cache_size) { 1134 STAILQ_INSERT_HEAD(&group->buf_cache, 1135 (struct spdk_nvmf_transport_pg_cache_buf *)rdma_req->data.buffers[i], link); 1136 group->buf_cache_count++; 1137 } else { 1138 spdk_mempool_put(transport->data_buf_pool, rdma_req->data.buffers[i]); 1139 } 1140 rdma_req->req.iov[i].iov_base = NULL; 1141 rdma_req->data.buffers[i] = NULL; 1142 rdma_req->req.iov[i].iov_len = 0; 1143 1144 } 1145 rdma_req->data_from_pool = false; 1146 } 1147 1148 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1149 1150 static spdk_nvme_data_transfer_t 1151 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1152 { 1153 enum spdk_nvme_data_transfer xfer; 1154 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1155 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1156 1157 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1158 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1159 rdma_req->rsp.wr.imm_data = 0; 1160 #endif 1161 1162 /* Figure out data transfer direction */ 1163 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1164 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1165 } else { 1166 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1167 1168 /* Some admin commands are special cases */ 1169 if ((rdma_req->req.qpair->qid == 0) && 1170 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1171 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1172 switch (cmd->cdw10 & 0xff) { 1173 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1174 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1175 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1176 break; 1177 default: 1178 xfer = SPDK_NVME_DATA_NONE; 1179 } 1180 } 1181 } 1182 1183 if (xfer == SPDK_NVME_DATA_NONE) { 1184 return xfer; 1185 } 1186 1187 /* Even for commands that may transfer data, they could have specified 0 length. 1188 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1189 */ 1190 switch (sgl->generic.type) { 1191 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1192 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1193 case SPDK_NVME_SGL_TYPE_SEGMENT: 1194 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1195 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1196 if (sgl->unkeyed.length == 0) { 1197 xfer = SPDK_NVME_DATA_NONE; 1198 } 1199 break; 1200 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1201 if (sgl->keyed.length == 0) { 1202 xfer = SPDK_NVME_DATA_NONE; 1203 } 1204 break; 1205 } 1206 1207 return xfer; 1208 } 1209 1210 static int 1211 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1212 struct spdk_nvmf_rdma_device *device, 1213 struct spdk_nvmf_rdma_request *rdma_req) 1214 { 1215 struct spdk_nvmf_rdma_qpair *rqpair; 1216 struct spdk_nvmf_rdma_poll_group *rgroup; 1217 void *buf = NULL; 1218 uint32_t length = rdma_req->req.length; 1219 uint64_t translation_len; 1220 uint32_t i = 0; 1221 int rc = 0; 1222 1223 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1224 rgroup = rqpair->poller->group; 1225 rdma_req->req.iovcnt = 0; 1226 while (length) { 1227 if (!(STAILQ_EMPTY(&rgroup->group.buf_cache))) { 1228 rgroup->group.buf_cache_count--; 1229 buf = STAILQ_FIRST(&rgroup->group.buf_cache); 1230 STAILQ_REMOVE_HEAD(&rgroup->group.buf_cache, link); 1231 assert(buf != NULL); 1232 } else { 1233 buf = spdk_mempool_get(rtransport->transport.data_buf_pool); 1234 if (!buf) { 1235 rc = -ENOMEM; 1236 goto err_exit; 1237 } 1238 } 1239 1240 rdma_req->req.iov[i].iov_base = (void *)((uintptr_t)(buf + NVMF_DATA_BUFFER_MASK) & 1241 ~NVMF_DATA_BUFFER_MASK); 1242 rdma_req->req.iov[i].iov_len = spdk_min(length, rtransport->transport.opts.io_unit_size); 1243 rdma_req->req.iovcnt++; 1244 rdma_req->data.buffers[i] = buf; 1245 rdma_req->data.wr.sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[i].iov_base); 1246 rdma_req->data.wr.sg_list[i].length = rdma_req->req.iov[i].iov_len; 1247 translation_len = rdma_req->req.iov[i].iov_len; 1248 1249 if (!g_nvmf_hooks.get_rkey) { 1250 rdma_req->data.wr.sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1251 (uint64_t)buf, &translation_len))->lkey; 1252 } else { 1253 rdma_req->data.wr.sg_list[i].lkey = spdk_mem_map_translate(device->map, 1254 (uint64_t)buf, &translation_len); 1255 } 1256 1257 length -= rdma_req->req.iov[i].iov_len; 1258 1259 if (translation_len < rdma_req->req.iov[i].iov_len) { 1260 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions\n"); 1261 rc = -EINVAL; 1262 goto err_exit; 1263 } 1264 i++; 1265 } 1266 1267 assert(rdma_req->req.iovcnt <= rqpair->max_send_sge); 1268 1269 rdma_req->data_from_pool = true; 1270 1271 return rc; 1272 1273 err_exit: 1274 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport); 1275 while (i) { 1276 i--; 1277 rdma_req->data.wr.sg_list[i].addr = 0; 1278 rdma_req->data.wr.sg_list[i].length = 0; 1279 rdma_req->data.wr.sg_list[i].lkey = 0; 1280 } 1281 rdma_req->req.iovcnt = 0; 1282 return rc; 1283 } 1284 1285 static int 1286 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1287 struct spdk_nvmf_rdma_device *device, 1288 struct spdk_nvmf_rdma_request *rdma_req) 1289 { 1290 struct spdk_nvme_cmd *cmd; 1291 struct spdk_nvme_cpl *rsp; 1292 struct spdk_nvme_sgl_descriptor *sgl; 1293 1294 cmd = &rdma_req->req.cmd->nvme_cmd; 1295 rsp = &rdma_req->req.rsp->nvme_cpl; 1296 sgl = &cmd->dptr.sgl1; 1297 1298 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1299 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1300 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1301 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1302 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1303 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1304 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1305 return -1; 1306 } 1307 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1308 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1309 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1310 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1311 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1312 } 1313 } 1314 #endif 1315 1316 /* fill request length and populate iovs */ 1317 rdma_req->req.length = sgl->keyed.length; 1318 1319 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1320 /* No available buffers. Queue this request up. */ 1321 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1322 return 0; 1323 } 1324 1325 /* backward compatible */ 1326 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1327 1328 /* rdma wr specifics */ 1329 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1330 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1331 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1332 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1333 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 1334 rdma_req->data.wr.next = &rdma_req->rsp.wr; 1335 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1336 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 1337 rdma_req->data.wr.next = NULL; 1338 } 1339 1340 /* set the number of outstanding data WRs for this request. */ 1341 rdma_req->num_outstanding_data_wr = 1; 1342 1343 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1344 rdma_req->req.iovcnt); 1345 1346 return 0; 1347 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1348 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1349 uint64_t offset = sgl->address; 1350 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1351 1352 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1353 offset, sgl->unkeyed.length); 1354 1355 if (offset > max_len) { 1356 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1357 offset, max_len); 1358 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1359 return -1; 1360 } 1361 max_len -= (uint32_t)offset; 1362 1363 if (sgl->unkeyed.length > max_len) { 1364 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1365 sgl->unkeyed.length, max_len); 1366 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1367 return -1; 1368 } 1369 1370 rdma_req->num_outstanding_data_wr = 0; 1371 rdma_req->req.data = rdma_req->recv->buf + offset; 1372 rdma_req->data_from_pool = false; 1373 rdma_req->req.length = sgl->unkeyed.length; 1374 1375 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1376 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1377 rdma_req->req.iovcnt = 1; 1378 1379 return 0; 1380 } 1381 1382 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1383 sgl->generic.type, sgl->generic.subtype); 1384 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1385 return -1; 1386 } 1387 1388 static void 1389 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1390 struct spdk_nvmf_rdma_transport *rtransport) 1391 { 1392 struct spdk_nvmf_rdma_qpair *rqpair; 1393 struct spdk_nvmf_rdma_poll_group *rgroup; 1394 1395 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1396 if (rdma_req->data_from_pool) { 1397 rgroup = rqpair->poller->group; 1398 1399 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport); 1400 } 1401 rdma_req->num_outstanding_data_wr = 0; 1402 rdma_req->req.length = 0; 1403 rdma_req->req.iovcnt = 0; 1404 rdma_req->req.data = NULL; 1405 rqpair->qd--; 1406 STAILQ_INSERT_HEAD(&rqpair->free_queue, rdma_req, state_link); 1407 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1408 } 1409 1410 static bool 1411 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1412 struct spdk_nvmf_rdma_request *rdma_req) 1413 { 1414 struct spdk_nvmf_rdma_qpair *rqpair; 1415 struct spdk_nvmf_rdma_device *device; 1416 struct spdk_nvmf_rdma_poll_group *rgroup; 1417 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1418 int rc; 1419 struct spdk_nvmf_rdma_recv *rdma_recv; 1420 enum spdk_nvmf_rdma_request_state prev_state; 1421 bool progress = false; 1422 int data_posted; 1423 1424 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1425 device = rqpair->port->device; 1426 rgroup = rqpair->poller->group; 1427 1428 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1429 1430 /* If the queue pair is in an error state, force the request to the completed state 1431 * to release resources. */ 1432 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1433 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1434 TAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, link); 1435 } 1436 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1437 } 1438 1439 /* The loop here is to allow for several back-to-back state changes. */ 1440 do { 1441 prev_state = rdma_req->state; 1442 1443 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1444 1445 switch (rdma_req->state) { 1446 case RDMA_REQUEST_STATE_FREE: 1447 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1448 * to escape this state. */ 1449 break; 1450 case RDMA_REQUEST_STATE_NEW: 1451 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1452 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1453 rdma_recv = rdma_req->recv; 1454 1455 /* The first element of the SGL is the NVMe command */ 1456 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1457 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1458 1459 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1460 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1461 break; 1462 } 1463 1464 /* The next state transition depends on the data transfer needs of this request. */ 1465 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1466 1467 /* If no data to transfer, ready to execute. */ 1468 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1469 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1470 break; 1471 } 1472 1473 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1474 TAILQ_INSERT_TAIL(&rgroup->pending_data_buf_queue, rdma_req, link); 1475 break; 1476 case RDMA_REQUEST_STATE_NEED_BUFFER: 1477 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1478 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1479 1480 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1481 1482 if (rdma_req != TAILQ_FIRST(&rgroup->pending_data_buf_queue)) { 1483 /* This request needs to wait in line to obtain a buffer */ 1484 break; 1485 } 1486 1487 /* Try to get a data buffer */ 1488 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1489 if (rc < 0) { 1490 TAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, link); 1491 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1492 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1493 break; 1494 } 1495 1496 if (!rdma_req->req.data) { 1497 /* No buffers available. */ 1498 break; 1499 } 1500 1501 TAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, link); 1502 1503 /* If data is transferring from host to controller and the data didn't 1504 * arrive using in capsule data, we need to do a transfer from the host. 1505 */ 1506 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1507 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1508 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1509 break; 1510 } 1511 1512 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1513 break; 1514 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1515 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1516 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1517 1518 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1519 /* This request needs to wait in line to perform RDMA */ 1520 break; 1521 } 1522 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1523 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1524 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1525 break; 1526 } 1527 1528 /* We have already verified that this request is the head of the queue. */ 1529 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 1530 1531 rc = request_transfer_in(&rdma_req->req); 1532 if (!rc) { 1533 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 1534 } else { 1535 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1536 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1537 } 1538 break; 1539 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1540 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1541 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1542 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1543 * to escape this state. */ 1544 break; 1545 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1546 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1547 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1548 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 1549 spdk_nvmf_request_exec(&rdma_req->req); 1550 break; 1551 case RDMA_REQUEST_STATE_EXECUTING: 1552 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1553 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1554 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1555 * to escape this state. */ 1556 break; 1557 case RDMA_REQUEST_STATE_EXECUTED: 1558 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1559 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1560 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1561 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 1562 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 1563 } else { 1564 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1565 } 1566 break; 1567 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 1568 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 1569 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1570 1571 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 1572 /* This request needs to wait in line to perform RDMA */ 1573 break; 1574 } 1575 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 1576 rqpair->max_send_depth) { 1577 /* We can only have so many WRs outstanding. we have to wait until some finish. 1578 * +1 since each request has an additional wr in the resp. */ 1579 break; 1580 } 1581 1582 /* We have already verified that this request is the head of the queue. */ 1583 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 1584 1585 /* The data transfer will be kicked off from 1586 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1587 */ 1588 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1589 break; 1590 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1591 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 1592 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1593 rc = request_transfer_out(&rdma_req->req, &data_posted); 1594 assert(rc == 0); /* No good way to handle this currently */ 1595 if (rc) { 1596 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1597 } else { 1598 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1599 RDMA_REQUEST_STATE_COMPLETING; 1600 } 1601 break; 1602 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1603 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 1604 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1605 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1606 * to escape this state. */ 1607 break; 1608 case RDMA_REQUEST_STATE_COMPLETING: 1609 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 1610 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1611 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1612 * to escape this state. */ 1613 break; 1614 case RDMA_REQUEST_STATE_COMPLETED: 1615 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 1616 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1617 1618 nvmf_rdma_request_free(rdma_req, rtransport); 1619 break; 1620 case RDMA_REQUEST_NUM_STATES: 1621 default: 1622 assert(0); 1623 break; 1624 } 1625 1626 if (rdma_req->state != prev_state) { 1627 progress = true; 1628 } 1629 } while (rdma_req->state != prev_state); 1630 1631 return progress; 1632 } 1633 1634 /* Public API callbacks begin here */ 1635 1636 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 1637 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 1638 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 1639 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 1640 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 1641 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 1642 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4096 1643 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 1644 1645 static void 1646 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 1647 { 1648 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 1649 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 1650 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 1651 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 1652 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 1653 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 1654 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 1655 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 1656 } 1657 1658 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 1659 1660 static struct spdk_nvmf_transport * 1661 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 1662 { 1663 int rc; 1664 struct spdk_nvmf_rdma_transport *rtransport; 1665 struct spdk_nvmf_rdma_device *device, *tmp; 1666 struct ibv_context **contexts; 1667 uint32_t i; 1668 int flag; 1669 uint32_t sge_count; 1670 uint32_t min_shared_buffers; 1671 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 1672 1673 rtransport = calloc(1, sizeof(*rtransport)); 1674 if (!rtransport) { 1675 return NULL; 1676 } 1677 1678 if (pthread_mutex_init(&rtransport->lock, NULL)) { 1679 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 1680 free(rtransport); 1681 return NULL; 1682 } 1683 1684 TAILQ_INIT(&rtransport->devices); 1685 TAILQ_INIT(&rtransport->ports); 1686 1687 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 1688 1689 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 1690 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 1691 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 1692 " in_capsule_data_size=%d, max_aq_depth=%d\n" 1693 " num_shared_buffers=%d\n", 1694 opts->max_queue_depth, 1695 opts->max_io_size, 1696 opts->max_qpairs_per_ctrlr, 1697 opts->io_unit_size, 1698 opts->in_capsule_data_size, 1699 opts->max_aq_depth, 1700 opts->num_shared_buffers); 1701 1702 /* I/O unit size cannot be larger than max I/O size */ 1703 if (opts->io_unit_size > opts->max_io_size) { 1704 opts->io_unit_size = opts->max_io_size; 1705 } 1706 1707 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 1708 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 1709 "the minimum number required to guarantee that forward progress can be made (%d)\n", 1710 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 1711 spdk_nvmf_rdma_destroy(&rtransport->transport); 1712 return NULL; 1713 } 1714 1715 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 1716 if (min_shared_buffers > opts->num_shared_buffers) { 1717 SPDK_ERRLOG("There are not enough buffers to satisfy" 1718 "per-poll group caches for each thread. (%" PRIu32 ")" 1719 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 1720 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 1721 spdk_nvmf_rdma_destroy(&rtransport->transport); 1722 return NULL; 1723 } 1724 1725 sge_count = opts->max_io_size / opts->io_unit_size; 1726 if (sge_count > NVMF_DEFAULT_TX_SGE) { 1727 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 1728 spdk_nvmf_rdma_destroy(&rtransport->transport); 1729 return NULL; 1730 } 1731 1732 rtransport->event_channel = rdma_create_event_channel(); 1733 if (rtransport->event_channel == NULL) { 1734 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 1735 spdk_nvmf_rdma_destroy(&rtransport->transport); 1736 return NULL; 1737 } 1738 1739 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 1740 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 1741 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 1742 rtransport->event_channel->fd, spdk_strerror(errno)); 1743 spdk_nvmf_rdma_destroy(&rtransport->transport); 1744 return NULL; 1745 } 1746 1747 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 1748 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 1749 sizeof(struct spdk_nvmf_rdma_request_data), 1750 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 1751 SPDK_ENV_SOCKET_ID_ANY); 1752 if (!rtransport->data_wr_pool) { 1753 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 1754 spdk_nvmf_rdma_destroy(&rtransport->transport); 1755 return NULL; 1756 } 1757 1758 contexts = rdma_get_devices(NULL); 1759 if (contexts == NULL) { 1760 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 1761 spdk_nvmf_rdma_destroy(&rtransport->transport); 1762 return NULL; 1763 } 1764 1765 i = 0; 1766 rc = 0; 1767 while (contexts[i] != NULL) { 1768 device = calloc(1, sizeof(*device)); 1769 if (!device) { 1770 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 1771 rc = -ENOMEM; 1772 break; 1773 } 1774 device->context = contexts[i]; 1775 rc = ibv_query_device(device->context, &device->attr); 1776 if (rc < 0) { 1777 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1778 free(device); 1779 break; 1780 1781 } 1782 1783 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 1784 1785 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1786 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 1787 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 1788 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 1789 } 1790 1791 /** 1792 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 1793 * The Soft-RoCE RXE driver does not currently support send with invalidate, 1794 * but incorrectly reports that it does. There are changes making their way 1795 * through the kernel now that will enable this feature. When they are merged, 1796 * we can conditionally enable this feature. 1797 * 1798 * TODO: enable this for versions of the kernel rxe driver that support it. 1799 */ 1800 if (device->attr.vendor_id == 0) { 1801 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 1802 } 1803 #endif 1804 1805 /* set up device context async ev fd as NON_BLOCKING */ 1806 flag = fcntl(device->context->async_fd, F_GETFL); 1807 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 1808 if (rc < 0) { 1809 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 1810 free(device); 1811 break; 1812 } 1813 1814 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 1815 i++; 1816 } 1817 rdma_free_devices(contexts); 1818 1819 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 1820 /* divide and round up. */ 1821 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 1822 1823 /* round up to the nearest 4k. */ 1824 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 1825 1826 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 1827 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 1828 opts->io_unit_size); 1829 } 1830 1831 if (rc < 0) { 1832 spdk_nvmf_rdma_destroy(&rtransport->transport); 1833 return NULL; 1834 } 1835 1836 /* Set up poll descriptor array to monitor events from RDMA and IB 1837 * in a single poll syscall 1838 */ 1839 rtransport->npoll_fds = i + 1; 1840 i = 0; 1841 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 1842 if (rtransport->poll_fds == NULL) { 1843 SPDK_ERRLOG("poll_fds allocation failed\n"); 1844 spdk_nvmf_rdma_destroy(&rtransport->transport); 1845 return NULL; 1846 } 1847 1848 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 1849 rtransport->poll_fds[i++].events = POLLIN; 1850 1851 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 1852 rtransport->poll_fds[i].fd = device->context->async_fd; 1853 rtransport->poll_fds[i++].events = POLLIN; 1854 } 1855 1856 return &rtransport->transport; 1857 } 1858 1859 static int 1860 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 1861 { 1862 struct spdk_nvmf_rdma_transport *rtransport; 1863 struct spdk_nvmf_rdma_port *port, *port_tmp; 1864 struct spdk_nvmf_rdma_device *device, *device_tmp; 1865 1866 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1867 1868 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 1869 TAILQ_REMOVE(&rtransport->ports, port, link); 1870 rdma_destroy_id(port->id); 1871 free(port); 1872 } 1873 1874 if (rtransport->poll_fds != NULL) { 1875 free(rtransport->poll_fds); 1876 } 1877 1878 if (rtransport->event_channel != NULL) { 1879 rdma_destroy_event_channel(rtransport->event_channel); 1880 } 1881 1882 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 1883 TAILQ_REMOVE(&rtransport->devices, device, link); 1884 if (device->map) { 1885 spdk_mem_map_free(&device->map); 1886 } 1887 if (device->pd) { 1888 if (!g_nvmf_hooks.get_ibv_pd) { 1889 ibv_dealloc_pd(device->pd); 1890 } 1891 } 1892 free(device); 1893 } 1894 1895 if (rtransport->data_wr_pool != NULL) { 1896 if (spdk_mempool_count(rtransport->data_wr_pool) != 1897 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 1898 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 1899 spdk_mempool_count(rtransport->data_wr_pool), 1900 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 1901 } 1902 } 1903 1904 spdk_mempool_free(rtransport->data_wr_pool); 1905 pthread_mutex_destroy(&rtransport->lock); 1906 free(rtransport); 1907 1908 return 0; 1909 } 1910 1911 static int 1912 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 1913 struct spdk_nvme_transport_id *trid, 1914 bool peer); 1915 1916 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 1917 .notify_cb = spdk_nvmf_rdma_mem_notify, 1918 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 1919 }; 1920 1921 static int 1922 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 1923 const struct spdk_nvme_transport_id *trid) 1924 { 1925 struct spdk_nvmf_rdma_transport *rtransport; 1926 struct spdk_nvmf_rdma_device *device; 1927 struct spdk_nvmf_rdma_port *port_tmp, *port; 1928 struct ibv_pd *pd; 1929 struct addrinfo *res; 1930 struct addrinfo hints; 1931 int family; 1932 int rc; 1933 1934 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1935 1936 port = calloc(1, sizeof(*port)); 1937 if (!port) { 1938 return -ENOMEM; 1939 } 1940 1941 /* Selectively copy the trid. Things like NQN don't matter here - that 1942 * mapping is enforced elsewhere. 1943 */ 1944 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1945 port->trid.adrfam = trid->adrfam; 1946 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 1947 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 1948 1949 pthread_mutex_lock(&rtransport->lock); 1950 assert(rtransport->event_channel != NULL); 1951 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 1952 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 1953 port_tmp->ref++; 1954 free(port); 1955 /* Already listening at this address */ 1956 pthread_mutex_unlock(&rtransport->lock); 1957 return 0; 1958 } 1959 } 1960 1961 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 1962 if (rc < 0) { 1963 SPDK_ERRLOG("rdma_create_id() failed\n"); 1964 free(port); 1965 pthread_mutex_unlock(&rtransport->lock); 1966 return rc; 1967 } 1968 1969 switch (port->trid.adrfam) { 1970 case SPDK_NVMF_ADRFAM_IPV4: 1971 family = AF_INET; 1972 break; 1973 case SPDK_NVMF_ADRFAM_IPV6: 1974 family = AF_INET6; 1975 break; 1976 default: 1977 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 1978 free(port); 1979 pthread_mutex_unlock(&rtransport->lock); 1980 return -EINVAL; 1981 } 1982 1983 memset(&hints, 0, sizeof(hints)); 1984 hints.ai_family = family; 1985 hints.ai_flags = AI_NUMERICSERV; 1986 hints.ai_socktype = SOCK_STREAM; 1987 hints.ai_protocol = 0; 1988 1989 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 1990 if (rc) { 1991 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 1992 free(port); 1993 pthread_mutex_unlock(&rtransport->lock); 1994 return -EINVAL; 1995 } 1996 1997 rc = rdma_bind_addr(port->id, res->ai_addr); 1998 freeaddrinfo(res); 1999 2000 if (rc < 0) { 2001 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2002 rdma_destroy_id(port->id); 2003 free(port); 2004 pthread_mutex_unlock(&rtransport->lock); 2005 return rc; 2006 } 2007 2008 if (!port->id->verbs) { 2009 SPDK_ERRLOG("ibv_context is null\n"); 2010 rdma_destroy_id(port->id); 2011 free(port); 2012 pthread_mutex_unlock(&rtransport->lock); 2013 return -1; 2014 } 2015 2016 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 2017 if (rc < 0) { 2018 SPDK_ERRLOG("rdma_listen() failed\n"); 2019 rdma_destroy_id(port->id); 2020 free(port); 2021 pthread_mutex_unlock(&rtransport->lock); 2022 return rc; 2023 } 2024 2025 TAILQ_FOREACH(device, &rtransport->devices, link) { 2026 if (device->context == port->id->verbs) { 2027 port->device = device; 2028 break; 2029 } 2030 } 2031 if (!port->device) { 2032 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2033 port->id->verbs); 2034 rdma_destroy_id(port->id); 2035 free(port); 2036 pthread_mutex_unlock(&rtransport->lock); 2037 return -EINVAL; 2038 } 2039 2040 pd = NULL; 2041 if (g_nvmf_hooks.get_ibv_pd) { 2042 if (spdk_nvmf_rdma_trid_from_cm_id(port->id, &port->trid, 1) < 0) { 2043 rdma_destroy_id(port->id); 2044 free(port); 2045 pthread_mutex_unlock(&rtransport->lock); 2046 return -EINVAL; 2047 } 2048 2049 pd = g_nvmf_hooks.get_ibv_pd(&port->trid, port->id->verbs); 2050 } 2051 2052 if (device->pd == NULL) { 2053 /* Haven't created a protection domain yet. */ 2054 2055 if (!g_nvmf_hooks.get_ibv_pd) { 2056 device->pd = ibv_alloc_pd(device->context); 2057 if (!device->pd) { 2058 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2059 rdma_destroy_id(port->id); 2060 free(port); 2061 pthread_mutex_unlock(&rtransport->lock); 2062 return -ENOMEM; 2063 } 2064 } else { 2065 device->pd = pd; 2066 } 2067 2068 assert(device->map == NULL); 2069 2070 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2071 if (!device->map) { 2072 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2073 if (!g_nvmf_hooks.get_ibv_pd) { 2074 ibv_dealloc_pd(device->pd); 2075 } 2076 rdma_destroy_id(port->id); 2077 free(port); 2078 pthread_mutex_unlock(&rtransport->lock); 2079 return -ENOMEM; 2080 } 2081 } else if (g_nvmf_hooks.get_ibv_pd) { 2082 /* A protection domain exists for this device, but the user has 2083 * enabled hooks. Verify that they only supply one pd per device. */ 2084 if (device->pd != pd) { 2085 SPDK_ERRLOG("The NVMe-oF target only supports one protection domain per device.\n"); 2086 rdma_destroy_id(port->id); 2087 free(port); 2088 pthread_mutex_unlock(&rtransport->lock); 2089 return -EINVAL; 2090 } 2091 } 2092 2093 assert(device->map != NULL); 2094 assert(device->pd != NULL); 2095 2096 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 2097 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 2098 2099 port->ref = 1; 2100 2101 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2102 pthread_mutex_unlock(&rtransport->lock); 2103 2104 return 0; 2105 } 2106 2107 static int 2108 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2109 const struct spdk_nvme_transport_id *_trid) 2110 { 2111 struct spdk_nvmf_rdma_transport *rtransport; 2112 struct spdk_nvmf_rdma_port *port, *tmp; 2113 struct spdk_nvme_transport_id trid = {}; 2114 2115 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2116 2117 /* Selectively copy the trid. Things like NQN don't matter here - that 2118 * mapping is enforced elsewhere. 2119 */ 2120 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2121 trid.adrfam = _trid->adrfam; 2122 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2123 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2124 2125 pthread_mutex_lock(&rtransport->lock); 2126 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2127 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2128 assert(port->ref > 0); 2129 port->ref--; 2130 if (port->ref == 0) { 2131 TAILQ_REMOVE(&rtransport->ports, port, link); 2132 rdma_destroy_id(port->id); 2133 free(port); 2134 } 2135 break; 2136 } 2137 } 2138 2139 pthread_mutex_unlock(&rtransport->lock); 2140 return 0; 2141 } 2142 2143 static bool 2144 spdk_nvmf_rdma_qpair_is_idle(struct spdk_nvmf_qpair *qpair) 2145 { 2146 struct spdk_nvmf_rdma_qpair *rqpair; 2147 2148 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2149 2150 if (rqpair->qd == 0) { 2151 return true; 2152 } 2153 return false; 2154 } 2155 2156 static void 2157 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2158 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2159 { 2160 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2161 2162 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2163 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2164 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2165 break; 2166 } 2167 } 2168 2169 /* Then RDMA writes since reads have stronger restrictions than writes */ 2170 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2171 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2172 break; 2173 } 2174 } 2175 2176 /* The second highest priority is I/O waiting on memory buffers. */ 2177 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->poller->group->pending_data_buf_queue, link, 2178 req_tmp) { 2179 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2180 break; 2181 } 2182 } 2183 2184 while (!STAILQ_EMPTY(&rqpair->free_queue) && !STAILQ_EMPTY(&rqpair->incoming_queue)) { 2185 2186 rdma_req = STAILQ_FIRST(&rqpair->free_queue); 2187 STAILQ_REMOVE_HEAD(&rqpair->free_queue, state_link); 2188 rdma_req->recv = STAILQ_FIRST(&rqpair->incoming_queue); 2189 STAILQ_REMOVE_HEAD(&rqpair->incoming_queue, link); 2190 2191 rqpair->qd++; 2192 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2193 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2194 break; 2195 } 2196 } 2197 } 2198 2199 static void 2200 _nvmf_rdma_qpair_disconnect(void *ctx) 2201 { 2202 struct spdk_nvmf_qpair *qpair = ctx; 2203 2204 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2205 } 2206 2207 static void 2208 _nvmf_rdma_try_disconnect(void *ctx) 2209 { 2210 struct spdk_nvmf_qpair *qpair = ctx; 2211 struct spdk_nvmf_poll_group *group; 2212 2213 /* Read the group out of the qpair. This is normally set and accessed only from 2214 * the thread that created the group. Here, we're not on that thread necessarily. 2215 * The data member qpair->group begins it's life as NULL and then is assigned to 2216 * a pointer and never changes. So fortunately reading this and checking for 2217 * non-NULL is thread safe in the x86_64 memory model. */ 2218 group = qpair->group; 2219 2220 if (group == NULL) { 2221 /* The qpair hasn't been assigned to a group yet, so we can't 2222 * process a disconnect. Send a message to ourself and try again. */ 2223 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2224 return; 2225 } 2226 2227 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2228 } 2229 2230 static inline void 2231 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2232 { 2233 if (__sync_bool_compare_and_swap(&rqpair->disconnect_started, false, true)) { 2234 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2235 } 2236 } 2237 2238 2239 static int 2240 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2241 { 2242 struct spdk_nvmf_qpair *qpair; 2243 struct spdk_nvmf_rdma_qpair *rqpair; 2244 2245 if (evt->id == NULL) { 2246 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2247 return -1; 2248 } 2249 2250 qpair = evt->id->context; 2251 if (qpair == NULL) { 2252 SPDK_ERRLOG("disconnect request: no active connection\n"); 2253 return -1; 2254 } 2255 2256 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2257 2258 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2259 2260 spdk_nvmf_rdma_update_ibv_state(rqpair); 2261 2262 spdk_nvmf_rdma_start_disconnect(rqpair); 2263 2264 return 0; 2265 } 2266 2267 #ifdef DEBUG 2268 static const char *CM_EVENT_STR[] = { 2269 "RDMA_CM_EVENT_ADDR_RESOLVED", 2270 "RDMA_CM_EVENT_ADDR_ERROR", 2271 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2272 "RDMA_CM_EVENT_ROUTE_ERROR", 2273 "RDMA_CM_EVENT_CONNECT_REQUEST", 2274 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2275 "RDMA_CM_EVENT_CONNECT_ERROR", 2276 "RDMA_CM_EVENT_UNREACHABLE", 2277 "RDMA_CM_EVENT_REJECTED", 2278 "RDMA_CM_EVENT_ESTABLISHED", 2279 "RDMA_CM_EVENT_DISCONNECTED", 2280 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2281 "RDMA_CM_EVENT_MULTICAST_JOIN", 2282 "RDMA_CM_EVENT_MULTICAST_ERROR", 2283 "RDMA_CM_EVENT_ADDR_CHANGE", 2284 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2285 }; 2286 #endif /* DEBUG */ 2287 2288 static void 2289 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2290 { 2291 struct spdk_nvmf_rdma_transport *rtransport; 2292 struct rdma_cm_event *event; 2293 int rc; 2294 2295 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2296 2297 if (rtransport->event_channel == NULL) { 2298 return; 2299 } 2300 2301 while (1) { 2302 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2303 if (rc == 0) { 2304 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2305 2306 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2307 2308 switch (event->event) { 2309 case RDMA_CM_EVENT_ADDR_RESOLVED: 2310 case RDMA_CM_EVENT_ADDR_ERROR: 2311 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2312 case RDMA_CM_EVENT_ROUTE_ERROR: 2313 /* No action required. The target never attempts to resolve routes. */ 2314 break; 2315 case RDMA_CM_EVENT_CONNECT_REQUEST: 2316 rc = nvmf_rdma_connect(transport, event, cb_fn); 2317 if (rc < 0) { 2318 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2319 break; 2320 } 2321 break; 2322 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2323 /* The target never initiates a new connection. So this will not occur. */ 2324 break; 2325 case RDMA_CM_EVENT_CONNECT_ERROR: 2326 /* Can this happen? The docs say it can, but not sure what causes it. */ 2327 break; 2328 case RDMA_CM_EVENT_UNREACHABLE: 2329 case RDMA_CM_EVENT_REJECTED: 2330 /* These only occur on the client side. */ 2331 break; 2332 case RDMA_CM_EVENT_ESTABLISHED: 2333 /* TODO: Should we be waiting for this event anywhere? */ 2334 break; 2335 case RDMA_CM_EVENT_DISCONNECTED: 2336 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2337 rc = nvmf_rdma_disconnect(event); 2338 if (rc < 0) { 2339 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2340 break; 2341 } 2342 break; 2343 case RDMA_CM_EVENT_MULTICAST_JOIN: 2344 case RDMA_CM_EVENT_MULTICAST_ERROR: 2345 /* Multicast is not used */ 2346 break; 2347 case RDMA_CM_EVENT_ADDR_CHANGE: 2348 /* Not utilizing this event */ 2349 break; 2350 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2351 /* For now, do nothing. The target never re-uses queue pairs. */ 2352 break; 2353 default: 2354 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2355 break; 2356 } 2357 2358 rdma_ack_cm_event(event); 2359 } else { 2360 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2361 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2362 } 2363 break; 2364 } 2365 } 2366 } 2367 2368 static void 2369 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2370 { 2371 int rc; 2372 struct spdk_nvmf_rdma_qpair *rqpair; 2373 struct ibv_async_event event; 2374 enum ibv_qp_state state; 2375 2376 rc = ibv_get_async_event(device->context, &event); 2377 2378 if (rc) { 2379 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2380 errno, spdk_strerror(errno)); 2381 return; 2382 } 2383 2384 SPDK_NOTICELOG("Async event: %s\n", 2385 ibv_event_type_str(event.event_type)); 2386 2387 switch (event.event_type) { 2388 case IBV_EVENT_QP_FATAL: 2389 rqpair = event.element.qp->qp_context; 2390 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2391 (uintptr_t)rqpair->cm_id, event.event_type); 2392 spdk_nvmf_rdma_update_ibv_state(rqpair); 2393 spdk_nvmf_rdma_start_disconnect(rqpair); 2394 break; 2395 case IBV_EVENT_QP_LAST_WQE_REACHED: 2396 /* This event only occurs for shared receive queues, which are not currently supported. */ 2397 break; 2398 case IBV_EVENT_SQ_DRAINED: 2399 /* This event occurs frequently in both error and non-error states. 2400 * Check if the qpair is in an error state before sending a message. 2401 * Note that we're not on the correct thread to access the qpair, but 2402 * the operations that the below calls make all happen to be thread 2403 * safe. */ 2404 rqpair = event.element.qp->qp_context; 2405 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2406 (uintptr_t)rqpair->cm_id, event.event_type); 2407 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2408 if (state == IBV_QPS_ERR) { 2409 spdk_nvmf_rdma_start_disconnect(rqpair); 2410 } 2411 break; 2412 case IBV_EVENT_QP_REQ_ERR: 2413 case IBV_EVENT_QP_ACCESS_ERR: 2414 case IBV_EVENT_COMM_EST: 2415 case IBV_EVENT_PATH_MIG: 2416 case IBV_EVENT_PATH_MIG_ERR: 2417 rqpair = event.element.qp->qp_context; 2418 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2419 (uintptr_t)rqpair->cm_id, event.event_type); 2420 spdk_nvmf_rdma_update_ibv_state(rqpair); 2421 break; 2422 case IBV_EVENT_CQ_ERR: 2423 case IBV_EVENT_DEVICE_FATAL: 2424 case IBV_EVENT_PORT_ACTIVE: 2425 case IBV_EVENT_PORT_ERR: 2426 case IBV_EVENT_LID_CHANGE: 2427 case IBV_EVENT_PKEY_CHANGE: 2428 case IBV_EVENT_SM_CHANGE: 2429 case IBV_EVENT_SRQ_ERR: 2430 case IBV_EVENT_SRQ_LIMIT_REACHED: 2431 case IBV_EVENT_CLIENT_REREGISTER: 2432 case IBV_EVENT_GID_CHANGE: 2433 default: 2434 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2435 break; 2436 } 2437 ibv_ack_async_event(&event); 2438 } 2439 2440 static void 2441 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2442 { 2443 int nfds, i = 0; 2444 struct spdk_nvmf_rdma_transport *rtransport; 2445 struct spdk_nvmf_rdma_device *device, *tmp; 2446 2447 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2448 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2449 2450 if (nfds <= 0) { 2451 return; 2452 } 2453 2454 /* The first poll descriptor is RDMA CM event */ 2455 if (rtransport->poll_fds[i++].revents & POLLIN) { 2456 spdk_nvmf_process_cm_event(transport, cb_fn); 2457 nfds--; 2458 } 2459 2460 if (nfds == 0) { 2461 return; 2462 } 2463 2464 /* Second and subsequent poll descriptors are IB async events */ 2465 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2466 if (rtransport->poll_fds[i++].revents & POLLIN) { 2467 spdk_nvmf_process_ib_event(device); 2468 nfds--; 2469 } 2470 } 2471 /* check all flagged fd's have been served */ 2472 assert(nfds == 0); 2473 } 2474 2475 static void 2476 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2477 struct spdk_nvme_transport_id *trid, 2478 struct spdk_nvmf_discovery_log_page_entry *entry) 2479 { 2480 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2481 entry->adrfam = trid->adrfam; 2482 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2483 2484 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2485 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2486 2487 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2488 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2489 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2490 } 2491 2492 static struct spdk_nvmf_transport_poll_group * 2493 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2494 { 2495 struct spdk_nvmf_rdma_transport *rtransport; 2496 struct spdk_nvmf_rdma_poll_group *rgroup; 2497 struct spdk_nvmf_rdma_poller *poller, *tpoller; 2498 struct spdk_nvmf_rdma_device *device; 2499 2500 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2501 2502 rgroup = calloc(1, sizeof(*rgroup)); 2503 if (!rgroup) { 2504 return NULL; 2505 } 2506 2507 TAILQ_INIT(&rgroup->pollers); 2508 TAILQ_INIT(&rgroup->pending_data_buf_queue); 2509 2510 pthread_mutex_lock(&rtransport->lock); 2511 TAILQ_FOREACH(device, &rtransport->devices, link) { 2512 poller = calloc(1, sizeof(*poller)); 2513 if (!poller) { 2514 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2515 goto err_exit; 2516 } 2517 2518 poller->device = device; 2519 poller->group = rgroup; 2520 2521 TAILQ_INIT(&poller->qpairs); 2522 2523 poller->cq = ibv_create_cq(device->context, DEFAULT_NVMF_RDMA_CQ_SIZE, poller, NULL, 0); 2524 if (!poller->cq) { 2525 SPDK_ERRLOG("Unable to create completion queue\n"); 2526 free(poller); 2527 goto err_exit; 2528 } 2529 poller->num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2530 2531 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2532 } 2533 2534 pthread_mutex_unlock(&rtransport->lock); 2535 return &rgroup->group; 2536 2537 err_exit: 2538 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tpoller) { 2539 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2540 if (poller->cq) { 2541 ibv_destroy_cq(poller->cq); 2542 } 2543 free(poller); 2544 } 2545 2546 free(rgroup); 2547 pthread_mutex_unlock(&rtransport->lock); 2548 return NULL; 2549 } 2550 2551 static void 2552 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2553 { 2554 struct spdk_nvmf_rdma_poll_group *rgroup; 2555 struct spdk_nvmf_rdma_poller *poller, *tmp; 2556 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2557 2558 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2559 2560 if (!rgroup) { 2561 return; 2562 } 2563 2564 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2565 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2566 2567 if (poller->cq) { 2568 ibv_destroy_cq(poller->cq); 2569 } 2570 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 2571 spdk_nvmf_rdma_qpair_destroy(qpair); 2572 } 2573 2574 free(poller); 2575 } 2576 2577 if (!TAILQ_EMPTY(&rgroup->pending_data_buf_queue)) { 2578 SPDK_ERRLOG("Pending I/O list wasn't empty on poll group destruction\n"); 2579 } 2580 2581 free(rgroup); 2582 } 2583 2584 static void 2585 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 2586 { 2587 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2588 spdk_nvmf_rdma_qpair_destroy(rqpair); 2589 } 2590 2591 static int 2592 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 2593 struct spdk_nvmf_qpair *qpair) 2594 { 2595 struct spdk_nvmf_rdma_poll_group *rgroup; 2596 struct spdk_nvmf_rdma_qpair *rqpair; 2597 struct spdk_nvmf_rdma_device *device; 2598 struct spdk_nvmf_rdma_poller *poller; 2599 int rc; 2600 2601 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2602 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2603 2604 device = rqpair->port->device; 2605 2606 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 2607 if (poller->device == device) { 2608 break; 2609 } 2610 } 2611 2612 if (!poller) { 2613 SPDK_ERRLOG("No poller found for device.\n"); 2614 return -1; 2615 } 2616 2617 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 2618 rqpair->poller = poller; 2619 2620 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 2621 if (rc < 0) { 2622 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 2623 return -1; 2624 } 2625 2626 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 2627 if (rc) { 2628 /* Try to reject, but we probably can't */ 2629 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 2630 return -1; 2631 } 2632 2633 spdk_nvmf_rdma_update_ibv_state(rqpair); 2634 2635 return 0; 2636 } 2637 2638 static int 2639 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 2640 { 2641 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2642 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2643 struct spdk_nvmf_rdma_transport, transport); 2644 2645 nvmf_rdma_request_free(rdma_req, rtransport); 2646 return 0; 2647 } 2648 2649 static int 2650 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 2651 { 2652 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2653 struct spdk_nvmf_rdma_transport, transport); 2654 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 2655 struct spdk_nvmf_rdma_request, req); 2656 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 2657 struct spdk_nvmf_rdma_qpair, qpair); 2658 2659 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2660 /* The connection is alive, so process the request as normal */ 2661 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 2662 } else { 2663 /* The connection is dead. Move the request directly to the completed state. */ 2664 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2665 } 2666 2667 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2668 2669 return 0; 2670 } 2671 2672 static int 2673 spdk_nvmf_rdma_destroy_defunct_qpair(void *ctx) 2674 { 2675 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2676 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2677 struct spdk_nvmf_rdma_transport, transport); 2678 2679 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2680 spdk_nvmf_rdma_qpair_destroy(rqpair); 2681 2682 return 0; 2683 } 2684 2685 static void 2686 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 2687 { 2688 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2689 struct ibv_recv_wr recv_wr = {}; 2690 struct ibv_recv_wr *bad_recv_wr; 2691 struct ibv_send_wr send_wr = {}; 2692 struct ibv_send_wr *bad_send_wr; 2693 int rc; 2694 2695 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 2696 return; 2697 } 2698 2699 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 2700 2701 /* This happens only when the qpair is disconnected before 2702 * it is added to the poll group. Since there is no poll group, 2703 * the RDMA qp has not been initialized yet and the RDMA CM 2704 * event has not yet been acknowledged, so we need to reject it. 2705 */ 2706 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 2707 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 2708 return; 2709 } 2710 2711 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2712 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 2713 } 2714 2715 rqpair->drain_recv_wr.type = RDMA_WR_TYPE_DRAIN_RECV; 2716 recv_wr.wr_id = (uintptr_t)&rqpair->drain_recv_wr; 2717 rc = ibv_post_recv(rqpair->cm_id->qp, &recv_wr, &bad_recv_wr); 2718 if (rc) { 2719 SPDK_ERRLOG("Failed to post dummy receive WR, errno %d\n", errno); 2720 assert(false); 2721 return; 2722 } 2723 2724 rqpair->drain_send_wr.type = RDMA_WR_TYPE_DRAIN_SEND; 2725 send_wr.wr_id = (uintptr_t)&rqpair->drain_send_wr; 2726 send_wr.opcode = IBV_WR_SEND; 2727 rc = ibv_post_send(rqpair->cm_id->qp, &send_wr, &bad_send_wr); 2728 if (rc) { 2729 SPDK_ERRLOG("Failed to post dummy send WR, errno %d\n", errno); 2730 assert(false); 2731 return; 2732 } 2733 rqpair->current_send_depth++; 2734 2735 rqpair->destruct_poller = spdk_poller_register(spdk_nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 2736 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 2737 } 2738 2739 #ifdef DEBUG 2740 static int 2741 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 2742 { 2743 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 2744 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 2745 } 2746 #endif 2747 2748 static int 2749 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 2750 struct spdk_nvmf_rdma_poller *rpoller) 2751 { 2752 struct ibv_wc wc[32]; 2753 struct spdk_nvmf_rdma_wr *rdma_wr; 2754 struct spdk_nvmf_rdma_request *rdma_req; 2755 struct spdk_nvmf_rdma_recv *rdma_recv; 2756 struct spdk_nvmf_rdma_qpair *rqpair; 2757 int reaped, i; 2758 int count = 0; 2759 bool error = false; 2760 2761 /* Poll for completing operations. */ 2762 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 2763 if (reaped < 0) { 2764 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2765 errno, spdk_strerror(errno)); 2766 return -1; 2767 } 2768 2769 for (i = 0; i < reaped; i++) { 2770 2771 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 2772 2773 /* Handle error conditions */ 2774 if (wc[i].status) { 2775 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 2776 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 2777 2778 error = true; 2779 2780 switch (rdma_wr->type) { 2781 case RDMA_WR_TYPE_SEND: 2782 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 2783 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2784 2785 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 2786 /* We're going to attempt an error recovery, so force the request into 2787 * the completed state. */ 2788 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2789 rqpair->current_send_depth--; 2790 2791 assert(rdma_req->num_outstanding_data_wr == 0); 2792 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2793 break; 2794 case RDMA_WR_TYPE_RECV: 2795 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 2796 rqpair = rdma_recv->qpair; 2797 2798 /* Dump this into the incoming queue. This gets cleaned up when 2799 * the queue pair disconnects or recovers. */ 2800 STAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2801 rqpair->current_recv_depth++; 2802 2803 /* Don't worry about responding to recv overflow, we are disconnecting anyways */ 2804 break; 2805 case RDMA_WR_TYPE_DATA: 2806 /* If the data transfer fails still force the queue into the error state, 2807 * if we were performing an RDMA_READ, we need to force the request into a 2808 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 2809 * case, we should wait for the SEND to complete. */ 2810 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2811 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2812 2813 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 2814 assert(rdma_req->num_outstanding_data_wr > 0); 2815 rdma_req->num_outstanding_data_wr--; 2816 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 2817 rqpair->current_read_depth--; 2818 if (rdma_req->num_outstanding_data_wr == 0) { 2819 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2820 } 2821 } 2822 rqpair->current_send_depth--; 2823 break; 2824 case RDMA_WR_TYPE_DRAIN_RECV: 2825 rqpair = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_qpair, drain_recv_wr); 2826 assert(rqpair->disconnect_flags & RDMA_QP_DISCONNECTING); 2827 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Drained QP RECV %u (%p)\n", rqpair->qpair.qid, rqpair); 2828 rqpair->disconnect_flags |= RDMA_QP_RECV_DRAINED; 2829 assert(rqpair->current_recv_depth == rqpair->max_queue_depth); 2830 /* Don't worry about responding to recv overflow, we are disconnecting anyways */ 2831 if (rqpair->disconnect_flags & RDMA_QP_SEND_DRAINED) { 2832 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2833 spdk_nvmf_rdma_qpair_destroy(rqpair); 2834 } 2835 /* Continue so that this does not trigger the disconnect path below. */ 2836 continue; 2837 case RDMA_WR_TYPE_DRAIN_SEND: 2838 rqpair = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_qpair, drain_send_wr); 2839 assert(rqpair->disconnect_flags & RDMA_QP_DISCONNECTING); 2840 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Drained QP SEND %u (%p)\n", rqpair->qpair.qid, rqpair); 2841 rqpair->disconnect_flags |= RDMA_QP_SEND_DRAINED; 2842 rqpair->current_send_depth--; 2843 if (rqpair->disconnect_flags & RDMA_QP_RECV_DRAINED) { 2844 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2845 spdk_nvmf_rdma_qpair_destroy(rqpair); 2846 } 2847 /* Continue so that this does not trigger the disconnect path below. */ 2848 continue; 2849 default: 2850 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2851 continue; 2852 } 2853 2854 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 2855 /* Disconnect the connection. */ 2856 spdk_nvmf_rdma_start_disconnect(rqpair); 2857 } 2858 continue; 2859 } 2860 2861 switch (wc[i].opcode) { 2862 case IBV_WC_SEND: 2863 assert(rdma_wr->type == RDMA_WR_TYPE_SEND); 2864 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 2865 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2866 2867 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 2868 2869 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2870 rqpair->current_send_depth--; 2871 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2872 2873 count++; 2874 2875 assert(rdma_req->num_outstanding_data_wr == 0); 2876 /* Try to process other queued requests */ 2877 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 2878 break; 2879 2880 case IBV_WC_RDMA_WRITE: 2881 assert(rdma_wr->type == RDMA_WR_TYPE_DATA); 2882 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2883 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2884 rqpair->current_send_depth--; 2885 rdma_req->num_outstanding_data_wr--; 2886 2887 /* Try to process other queued requests */ 2888 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 2889 break; 2890 2891 case IBV_WC_RDMA_READ: 2892 assert(rdma_wr->type == RDMA_WR_TYPE_DATA); 2893 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2894 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2895 rqpair->current_send_depth--; 2896 2897 assert(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2898 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 2899 assert(rdma_req->num_outstanding_data_wr > 0); 2900 rqpair->current_read_depth--; 2901 rdma_req->num_outstanding_data_wr--; 2902 if (rdma_req->num_outstanding_data_wr == 0) { 2903 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2904 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2905 } 2906 2907 /* Try to process other queued requests */ 2908 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 2909 break; 2910 2911 case IBV_WC_RECV: 2912 assert(rdma_wr->type == RDMA_WR_TYPE_RECV); 2913 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 2914 rqpair = rdma_recv->qpair; 2915 /* The qpair should not send more requests than are allowed per qpair. */ 2916 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 2917 spdk_nvmf_rdma_start_disconnect(rqpair); 2918 } else { 2919 rqpair->current_recv_depth++; 2920 } 2921 STAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2922 /* Try to process other queued requests */ 2923 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 2924 break; 2925 2926 default: 2927 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2928 continue; 2929 } 2930 } 2931 2932 if (error == true) { 2933 return -1; 2934 } 2935 2936 return count; 2937 } 2938 2939 static int 2940 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 2941 { 2942 struct spdk_nvmf_rdma_transport *rtransport; 2943 struct spdk_nvmf_rdma_poll_group *rgroup; 2944 struct spdk_nvmf_rdma_poller *rpoller; 2945 int count, rc; 2946 2947 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 2948 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2949 2950 count = 0; 2951 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2952 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 2953 if (rc < 0) { 2954 return rc; 2955 } 2956 count += rc; 2957 } 2958 2959 return count; 2960 } 2961 2962 static int 2963 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2964 struct spdk_nvme_transport_id *trid, 2965 bool peer) 2966 { 2967 struct sockaddr *saddr; 2968 uint16_t port; 2969 2970 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 2971 2972 if (peer) { 2973 saddr = rdma_get_peer_addr(id); 2974 } else { 2975 saddr = rdma_get_local_addr(id); 2976 } 2977 switch (saddr->sa_family) { 2978 case AF_INET: { 2979 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 2980 2981 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 2982 inet_ntop(AF_INET, &saddr_in->sin_addr, 2983 trid->traddr, sizeof(trid->traddr)); 2984 if (peer) { 2985 port = ntohs(rdma_get_dst_port(id)); 2986 } else { 2987 port = ntohs(rdma_get_src_port(id)); 2988 } 2989 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 2990 break; 2991 } 2992 case AF_INET6: { 2993 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 2994 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 2995 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 2996 trid->traddr, sizeof(trid->traddr)); 2997 if (peer) { 2998 port = ntohs(rdma_get_dst_port(id)); 2999 } else { 3000 port = ntohs(rdma_get_src_port(id)); 3001 } 3002 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3003 break; 3004 } 3005 default: 3006 return -1; 3007 3008 } 3009 3010 return 0; 3011 } 3012 3013 static int 3014 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3015 struct spdk_nvme_transport_id *trid) 3016 { 3017 struct spdk_nvmf_rdma_qpair *rqpair; 3018 3019 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3020 3021 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 3022 } 3023 3024 static int 3025 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3026 struct spdk_nvme_transport_id *trid) 3027 { 3028 struct spdk_nvmf_rdma_qpair *rqpair; 3029 3030 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3031 3032 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 3033 } 3034 3035 static int 3036 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3037 struct spdk_nvme_transport_id *trid) 3038 { 3039 struct spdk_nvmf_rdma_qpair *rqpair; 3040 3041 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3042 3043 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 3044 } 3045 3046 void 3047 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3048 { 3049 g_nvmf_hooks = *hooks; 3050 } 3051 3052 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 3053 .type = SPDK_NVME_TRANSPORT_RDMA, 3054 .opts_init = spdk_nvmf_rdma_opts_init, 3055 .create = spdk_nvmf_rdma_create, 3056 .destroy = spdk_nvmf_rdma_destroy, 3057 3058 .listen = spdk_nvmf_rdma_listen, 3059 .stop_listen = spdk_nvmf_rdma_stop_listen, 3060 .accept = spdk_nvmf_rdma_accept, 3061 3062 .listener_discover = spdk_nvmf_rdma_discover, 3063 3064 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 3065 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 3066 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 3067 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 3068 3069 .req_free = spdk_nvmf_rdma_request_free, 3070 .req_complete = spdk_nvmf_rdma_request_complete, 3071 3072 .qpair_fini = spdk_nvmf_rdma_close_qpair, 3073 .qpair_is_idle = spdk_nvmf_rdma_qpair_is_idle, 3074 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 3075 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 3076 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 3077 3078 }; 3079 3080 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 3081