1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2019 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 /* The maximum number of buffers per request */ 71 #define NVMF_REQ_MAX_BUFFERS (SPDK_NVMF_MAX_SGL_ENTRIES * 2) 72 73 static int g_spdk_nvmf_ibv_query_mask = 74 IBV_QP_STATE | 75 IBV_QP_PKEY_INDEX | 76 IBV_QP_PORT | 77 IBV_QP_ACCESS_FLAGS | 78 IBV_QP_AV | 79 IBV_QP_PATH_MTU | 80 IBV_QP_DEST_QPN | 81 IBV_QP_RQ_PSN | 82 IBV_QP_MAX_DEST_RD_ATOMIC | 83 IBV_QP_MIN_RNR_TIMER | 84 IBV_QP_SQ_PSN | 85 IBV_QP_TIMEOUT | 86 IBV_QP_RETRY_CNT | 87 IBV_QP_RNR_RETRY | 88 IBV_QP_MAX_QP_RD_ATOMIC; 89 90 enum spdk_nvmf_rdma_request_state { 91 /* The request is not currently in use */ 92 RDMA_REQUEST_STATE_FREE = 0, 93 94 /* Initial state when request first received */ 95 RDMA_REQUEST_STATE_NEW, 96 97 /* The request is queued until a data buffer is available. */ 98 RDMA_REQUEST_STATE_NEED_BUFFER, 99 100 /* The request is waiting on RDMA queue depth availability 101 * to transfer data from the host to the controller. 102 */ 103 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 104 105 /* The request is currently transferring data from the host to the controller. */ 106 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 107 108 /* The request is ready to execute at the block device */ 109 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 110 111 /* The request is currently executing at the block device */ 112 RDMA_REQUEST_STATE_EXECUTING, 113 114 /* The request finished executing at the block device */ 115 RDMA_REQUEST_STATE_EXECUTED, 116 117 /* The request is waiting on RDMA queue depth availability 118 * to transfer data from the controller to the host. 119 */ 120 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 121 122 /* The request is ready to send a completion */ 123 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 124 125 /* The request is currently transferring data from the controller to the host. */ 126 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 127 128 /* The request currently has an outstanding completion without an 129 * associated data transfer. 130 */ 131 RDMA_REQUEST_STATE_COMPLETING, 132 133 /* The request completed and can be marked free. */ 134 RDMA_REQUEST_STATE_COMPLETED, 135 136 /* Terminator */ 137 RDMA_REQUEST_NUM_STATES, 138 }; 139 140 #define OBJECT_NVMF_RDMA_IO 0x40 141 142 #define TRACE_GROUP_NVMF_RDMA 0x4 143 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 144 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 145 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 146 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 147 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 148 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 149 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 150 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 151 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 152 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 153 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 154 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 155 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 156 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 157 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 158 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 159 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 160 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 161 162 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 163 { 164 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 165 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 166 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 167 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 168 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 169 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 170 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 171 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 172 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 173 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 174 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 175 spdk_trace_register_description("RDMA_REQ_TX_H2C", 176 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 177 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 178 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 179 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 180 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 181 spdk_trace_register_description("RDMA_REQ_EXECUTING", 182 TRACE_RDMA_REQUEST_STATE_EXECUTING, 183 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 184 spdk_trace_register_description("RDMA_REQ_EXECUTED", 185 TRACE_RDMA_REQUEST_STATE_EXECUTED, 186 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 187 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 188 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 189 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 190 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 191 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 192 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 193 spdk_trace_register_description("RDMA_REQ_COMPLETING", 194 TRACE_RDMA_REQUEST_STATE_COMPLETING, 195 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 196 spdk_trace_register_description("RDMA_REQ_COMPLETED", 197 TRACE_RDMA_REQUEST_STATE_COMPLETED, 198 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 199 200 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 201 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 202 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 203 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 204 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 205 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 206 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 207 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 208 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 209 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 210 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 211 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 212 } 213 214 enum spdk_nvmf_rdma_wr_type { 215 RDMA_WR_TYPE_RECV, 216 RDMA_WR_TYPE_SEND, 217 RDMA_WR_TYPE_DATA, 218 }; 219 220 struct spdk_nvmf_rdma_wr { 221 enum spdk_nvmf_rdma_wr_type type; 222 }; 223 224 /* This structure holds commands as they are received off the wire. 225 * It must be dynamically paired with a full request object 226 * (spdk_nvmf_rdma_request) to service a request. It is separate 227 * from the request because RDMA does not appear to order 228 * completions, so occasionally we'll get a new incoming 229 * command when there aren't any free request objects. 230 */ 231 struct spdk_nvmf_rdma_recv { 232 struct ibv_recv_wr wr; 233 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 234 235 struct spdk_nvmf_rdma_qpair *qpair; 236 237 /* In-capsule data buffer */ 238 uint8_t *buf; 239 240 struct spdk_nvmf_rdma_wr rdma_wr; 241 242 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 243 }; 244 245 struct spdk_nvmf_rdma_request_data { 246 struct spdk_nvmf_rdma_wr rdma_wr; 247 struct ibv_send_wr wr; 248 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 249 }; 250 251 struct spdk_nvmf_rdma_request { 252 struct spdk_nvmf_request req; 253 bool data_from_pool; 254 255 enum spdk_nvmf_rdma_request_state state; 256 257 struct spdk_nvmf_rdma_recv *recv; 258 259 struct { 260 struct spdk_nvmf_rdma_wr rdma_wr; 261 struct ibv_send_wr wr; 262 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 263 } rsp; 264 265 struct spdk_nvmf_rdma_request_data data; 266 void *buffers[NVMF_REQ_MAX_BUFFERS]; 267 268 uint32_t num_outstanding_data_wr; 269 270 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 271 }; 272 273 enum spdk_nvmf_rdma_qpair_disconnect_flags { 274 RDMA_QP_DISCONNECTING = 1, 275 RDMA_QP_RECV_DRAINED = 1 << 1, 276 RDMA_QP_SEND_DRAINED = 1 << 2 277 }; 278 279 struct spdk_nvmf_rdma_resource_opts { 280 struct spdk_nvmf_rdma_qpair *qpair; 281 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 282 void *qp; 283 struct ibv_pd *pd; 284 uint32_t max_queue_depth; 285 uint32_t in_capsule_data_size; 286 bool shared; 287 }; 288 289 struct spdk_nvmf_send_wr_list { 290 struct ibv_send_wr *first; 291 struct ibv_send_wr *last; 292 }; 293 294 struct spdk_nvmf_recv_wr_list { 295 struct ibv_recv_wr *first; 296 struct ibv_recv_wr *last; 297 }; 298 299 struct spdk_nvmf_rdma_resources { 300 /* Array of size "max_queue_depth" containing RDMA requests. */ 301 struct spdk_nvmf_rdma_request *reqs; 302 303 /* Array of size "max_queue_depth" containing RDMA recvs. */ 304 struct spdk_nvmf_rdma_recv *recvs; 305 306 /* Array of size "max_queue_depth" containing 64 byte capsules 307 * used for receive. 308 */ 309 union nvmf_h2c_msg *cmds; 310 struct ibv_mr *cmds_mr; 311 312 /* Array of size "max_queue_depth" containing 16 byte completions 313 * to be sent back to the user. 314 */ 315 union nvmf_c2h_msg *cpls; 316 struct ibv_mr *cpls_mr; 317 318 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 319 * buffers to be used for in capsule data. 320 */ 321 void *bufs; 322 struct ibv_mr *bufs_mr; 323 324 /* The list of pending recvs to transfer */ 325 struct spdk_nvmf_recv_wr_list recvs_to_post; 326 327 /* Receives that are waiting for a request object */ 328 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 329 330 /* Queue to track free requests */ 331 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 332 }; 333 334 struct spdk_nvmf_rdma_qpair { 335 struct spdk_nvmf_qpair qpair; 336 337 struct spdk_nvmf_rdma_port *port; 338 struct spdk_nvmf_rdma_poller *poller; 339 340 struct rdma_cm_id *cm_id; 341 struct ibv_srq *srq; 342 struct rdma_cm_id *listen_id; 343 344 /* The maximum number of I/O outstanding on this connection at one time */ 345 uint16_t max_queue_depth; 346 347 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 348 uint16_t max_read_depth; 349 350 /* The maximum number of RDMA SEND operations at one time */ 351 uint32_t max_send_depth; 352 353 /* The current number of outstanding WRs from this qpair's 354 * recv queue. Should not exceed device->attr.max_queue_depth. 355 */ 356 uint16_t current_recv_depth; 357 358 /* The current number of active RDMA READ operations */ 359 uint16_t current_read_depth; 360 361 /* The current number of posted WRs from this qpair's 362 * send queue. Should not exceed max_send_depth. 363 */ 364 uint32_t current_send_depth; 365 366 /* The maximum number of SGEs per WR on the send queue */ 367 uint32_t max_send_sge; 368 369 /* The maximum number of SGEs per WR on the recv queue */ 370 uint32_t max_recv_sge; 371 372 /* The list of pending send requests for a transfer */ 373 struct spdk_nvmf_send_wr_list sends_to_post; 374 375 struct spdk_nvmf_rdma_resources *resources; 376 377 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 378 379 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 380 381 /* Number of requests not in the free state */ 382 uint32_t qd; 383 384 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 385 386 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 387 388 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 389 390 /* IBV queue pair attributes: they are used to manage 391 * qp state and recover from errors. 392 */ 393 enum ibv_qp_state ibv_state; 394 395 uint32_t disconnect_flags; 396 397 /* Poller registered in case the qpair doesn't properly 398 * complete the qpair destruct process and becomes defunct. 399 */ 400 401 struct spdk_poller *destruct_poller; 402 403 /* There are several ways a disconnect can start on a qpair 404 * and they are not all mutually exclusive. It is important 405 * that we only initialize one of these paths. 406 */ 407 bool disconnect_started; 408 /* Lets us know that we have received the last_wqe event. */ 409 bool last_wqe_reached; 410 }; 411 412 struct spdk_nvmf_rdma_poller { 413 struct spdk_nvmf_rdma_device *device; 414 struct spdk_nvmf_rdma_poll_group *group; 415 416 int num_cqe; 417 int required_num_wr; 418 struct ibv_cq *cq; 419 420 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 421 uint16_t max_srq_depth; 422 423 /* Shared receive queue */ 424 struct ibv_srq *srq; 425 426 struct spdk_nvmf_rdma_resources *resources; 427 428 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 429 430 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 431 432 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 433 434 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 435 }; 436 437 struct spdk_nvmf_rdma_poll_group { 438 struct spdk_nvmf_transport_poll_group group; 439 440 /* Requests that are waiting to obtain a data buffer */ 441 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_data_buf_queue; 442 443 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 444 }; 445 446 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 447 struct spdk_nvmf_rdma_device { 448 struct ibv_device_attr attr; 449 struct ibv_context *context; 450 451 struct spdk_mem_map *map; 452 struct ibv_pd *pd; 453 454 int num_srq; 455 456 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 457 }; 458 459 struct spdk_nvmf_rdma_port { 460 struct spdk_nvme_transport_id trid; 461 struct rdma_cm_id *id; 462 struct spdk_nvmf_rdma_device *device; 463 uint32_t ref; 464 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 465 }; 466 467 struct spdk_nvmf_rdma_transport { 468 struct spdk_nvmf_transport transport; 469 470 struct rdma_event_channel *event_channel; 471 472 struct spdk_mempool *data_wr_pool; 473 474 pthread_mutex_t lock; 475 476 /* fields used to poll RDMA/IB events */ 477 nfds_t npoll_fds; 478 struct pollfd *poll_fds; 479 480 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 481 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 482 }; 483 484 static inline int 485 spdk_nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 486 { 487 switch (state) { 488 case IBV_QPS_RESET: 489 case IBV_QPS_INIT: 490 case IBV_QPS_RTR: 491 case IBV_QPS_RTS: 492 case IBV_QPS_SQD: 493 case IBV_QPS_SQE: 494 case IBV_QPS_ERR: 495 return 0; 496 default: 497 return -1; 498 } 499 } 500 501 static enum ibv_qp_state 502 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 503 enum ibv_qp_state old_state, new_state; 504 struct ibv_qp_attr qp_attr; 505 struct ibv_qp_init_attr init_attr; 506 int rc; 507 508 old_state = rqpair->ibv_state; 509 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 510 g_spdk_nvmf_ibv_query_mask, &init_attr); 511 512 if (rc) 513 { 514 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 515 return IBV_QPS_ERR + 1; 516 } 517 518 new_state = qp_attr.qp_state; 519 rqpair->ibv_state = new_state; 520 qp_attr.ah_attr.port_num = qp_attr.port_num; 521 522 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 523 if (rc) 524 { 525 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 526 /* 527 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 528 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 529 */ 530 return IBV_QPS_ERR + 1; 531 } 532 533 if (old_state != new_state) 534 { 535 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 536 (uintptr_t)rqpair->cm_id, new_state); 537 } 538 return new_state; 539 } 540 541 static const char *str_ibv_qp_state[] = { 542 "IBV_QPS_RESET", 543 "IBV_QPS_INIT", 544 "IBV_QPS_RTR", 545 "IBV_QPS_RTS", 546 "IBV_QPS_SQD", 547 "IBV_QPS_SQE", 548 "IBV_QPS_ERR", 549 "IBV_QPS_UNKNOWN" 550 }; 551 552 static int 553 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 554 enum ibv_qp_state new_state) 555 { 556 struct ibv_qp_attr qp_attr; 557 struct ibv_qp_init_attr init_attr; 558 int rc; 559 enum ibv_qp_state state; 560 static int attr_mask_rc[] = { 561 [IBV_QPS_RESET] = IBV_QP_STATE, 562 [IBV_QPS_INIT] = (IBV_QP_STATE | 563 IBV_QP_PKEY_INDEX | 564 IBV_QP_PORT | 565 IBV_QP_ACCESS_FLAGS), 566 [IBV_QPS_RTR] = (IBV_QP_STATE | 567 IBV_QP_AV | 568 IBV_QP_PATH_MTU | 569 IBV_QP_DEST_QPN | 570 IBV_QP_RQ_PSN | 571 IBV_QP_MAX_DEST_RD_ATOMIC | 572 IBV_QP_MIN_RNR_TIMER), 573 [IBV_QPS_RTS] = (IBV_QP_STATE | 574 IBV_QP_SQ_PSN | 575 IBV_QP_TIMEOUT | 576 IBV_QP_RETRY_CNT | 577 IBV_QP_RNR_RETRY | 578 IBV_QP_MAX_QP_RD_ATOMIC), 579 [IBV_QPS_SQD] = IBV_QP_STATE, 580 [IBV_QPS_SQE] = IBV_QP_STATE, 581 [IBV_QPS_ERR] = IBV_QP_STATE, 582 }; 583 584 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 585 if (rc) { 586 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 587 rqpair->qpair.qid, new_state); 588 return rc; 589 } 590 591 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 592 g_spdk_nvmf_ibv_query_mask, &init_attr); 593 594 if (rc) { 595 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 596 assert(false); 597 } 598 599 qp_attr.cur_qp_state = rqpair->ibv_state; 600 qp_attr.qp_state = new_state; 601 602 rc = ibv_modify_qp(rqpair->cm_id->qp, &qp_attr, 603 attr_mask_rc[new_state]); 604 605 if (rc) { 606 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 607 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 608 return rc; 609 } 610 611 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 612 613 if (state != new_state) { 614 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 615 rqpair->qpair.qid, str_ibv_qp_state[new_state], 616 str_ibv_qp_state[state]); 617 return -1; 618 } 619 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 620 str_ibv_qp_state[state]); 621 return 0; 622 } 623 624 static void 625 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 626 struct spdk_nvmf_rdma_transport *rtransport) 627 { 628 struct spdk_nvmf_rdma_request_data *current_data_wr = NULL, *next_data_wr = NULL; 629 struct ibv_send_wr *send_wr; 630 int i; 631 632 rdma_req->num_outstanding_data_wr = 0; 633 current_data_wr = &rdma_req->data; 634 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 635 current_data_wr->wr.sg_list[i].addr = 0; 636 current_data_wr->wr.sg_list[i].length = 0; 637 current_data_wr->wr.sg_list[i].lkey = 0; 638 } 639 current_data_wr->wr.num_sge = 0; 640 641 send_wr = current_data_wr->wr.next; 642 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr) { 643 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 644 } 645 while (next_data_wr) { 646 current_data_wr = next_data_wr; 647 send_wr = current_data_wr->wr.next; 648 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr && 649 send_wr->wr_id == current_data_wr->wr.wr_id) { 650 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 651 } else { 652 next_data_wr = NULL; 653 } 654 655 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 656 current_data_wr->wr.sg_list[i].addr = 0; 657 current_data_wr->wr.sg_list[i].length = 0; 658 current_data_wr->wr.sg_list[i].lkey = 0; 659 } 660 current_data_wr->wr.num_sge = 0; 661 spdk_mempool_put(rtransport->data_wr_pool, current_data_wr); 662 } 663 } 664 665 static void 666 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 667 { 668 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->data_from_pool); 669 if (req->req.cmd) { 670 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 671 } 672 if (req->recv) { 673 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 674 } 675 } 676 677 static void 678 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 679 { 680 int i; 681 682 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 683 for (i = 0; i < rqpair->max_queue_depth; i++) { 684 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 685 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 686 } 687 } 688 } 689 690 static void 691 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 692 { 693 if (resources->cmds_mr) { 694 ibv_dereg_mr(resources->cmds_mr); 695 } 696 697 if (resources->cpls_mr) { 698 ibv_dereg_mr(resources->cpls_mr); 699 } 700 701 if (resources->bufs_mr) { 702 ibv_dereg_mr(resources->bufs_mr); 703 } 704 705 spdk_dma_free(resources->cmds); 706 spdk_dma_free(resources->cpls); 707 spdk_dma_free(resources->bufs); 708 free(resources->reqs); 709 free(resources->recvs); 710 free(resources); 711 } 712 713 714 static struct spdk_nvmf_rdma_resources * 715 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 716 { 717 struct spdk_nvmf_rdma_resources *resources; 718 struct spdk_nvmf_rdma_request *rdma_req; 719 struct spdk_nvmf_rdma_recv *rdma_recv; 720 struct ibv_qp *qp; 721 struct ibv_srq *srq; 722 uint32_t i; 723 int rc; 724 725 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 726 if (!resources) { 727 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 728 return NULL; 729 } 730 731 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 732 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 733 resources->cmds = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 734 0x1000, NULL); 735 resources->cpls = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 736 0x1000, NULL); 737 738 if (opts->in_capsule_data_size > 0) { 739 resources->bufs = spdk_dma_zmalloc(opts->max_queue_depth * 740 opts->in_capsule_data_size, 741 0x1000, NULL); 742 } 743 744 if (!resources->reqs || !resources->recvs || !resources->cmds || 745 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 746 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 747 goto cleanup; 748 } 749 750 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 751 opts->max_queue_depth * sizeof(*resources->cmds), 752 IBV_ACCESS_LOCAL_WRITE); 753 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 754 opts->max_queue_depth * sizeof(*resources->cpls), 755 0); 756 757 if (opts->in_capsule_data_size) { 758 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 759 opts->max_queue_depth * 760 opts->in_capsule_data_size, 761 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 762 } 763 764 if (!resources->cmds_mr || !resources->cpls_mr || 765 (opts->in_capsule_data_size && 766 !resources->bufs_mr)) { 767 goto cleanup; 768 } 769 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 770 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 771 resources->cmds_mr->lkey); 772 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 773 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 774 resources->cpls_mr->lkey); 775 if (resources->bufs && resources->bufs_mr) { 776 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 777 resources->bufs, opts->max_queue_depth * 778 opts->in_capsule_data_size, resources->bufs_mr->lkey); 779 } 780 781 /* Initialize queues */ 782 STAILQ_INIT(&resources->incoming_queue); 783 STAILQ_INIT(&resources->free_queue); 784 785 for (i = 0; i < opts->max_queue_depth; i++) { 786 struct ibv_recv_wr *bad_wr = NULL; 787 788 rdma_recv = &resources->recvs[i]; 789 rdma_recv->qpair = opts->qpair; 790 791 /* Set up memory to receive commands */ 792 if (resources->bufs) { 793 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 794 opts->in_capsule_data_size)); 795 } 796 797 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 798 799 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 800 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 801 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 802 rdma_recv->wr.num_sge = 1; 803 804 if (rdma_recv->buf && resources->bufs_mr) { 805 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 806 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 807 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 808 rdma_recv->wr.num_sge++; 809 } 810 811 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 812 rdma_recv->wr.sg_list = rdma_recv->sgl; 813 if (opts->shared) { 814 srq = (struct ibv_srq *)opts->qp; 815 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 816 } else { 817 qp = (struct ibv_qp *)opts->qp; 818 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 819 } 820 if (rc) { 821 goto cleanup; 822 } 823 } 824 825 for (i = 0; i < opts->max_queue_depth; i++) { 826 rdma_req = &resources->reqs[i]; 827 828 if (opts->qpair != NULL) { 829 rdma_req->req.qpair = &opts->qpair->qpair; 830 } else { 831 rdma_req->req.qpair = NULL; 832 } 833 rdma_req->req.cmd = NULL; 834 835 /* Set up memory to send responses */ 836 rdma_req->req.rsp = &resources->cpls[i]; 837 838 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 839 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 840 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 841 842 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 843 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 844 rdma_req->rsp.wr.next = NULL; 845 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 846 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 847 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 848 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 849 850 /* Set up memory for data buffers */ 851 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 852 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 853 rdma_req->data.wr.next = NULL; 854 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 855 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 856 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 857 858 /* Initialize request state to FREE */ 859 rdma_req->state = RDMA_REQUEST_STATE_FREE; 860 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 861 } 862 863 return resources; 864 865 cleanup: 866 nvmf_rdma_resources_destroy(resources); 867 return NULL; 868 } 869 870 static void 871 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 872 { 873 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 874 struct ibv_recv_wr *bad_recv_wr = NULL; 875 int rc; 876 877 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 878 879 spdk_poller_unregister(&rqpair->destruct_poller); 880 881 if (rqpair->qd != 0) { 882 if (rqpair->srq == NULL) { 883 nvmf_rdma_dump_qpair_contents(rqpair); 884 } 885 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 886 } 887 888 if (rqpair->poller) { 889 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 890 891 if (rqpair->srq != NULL && rqpair->resources != NULL) { 892 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 893 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 894 if (rqpair == rdma_recv->qpair) { 895 STAILQ_REMOVE_HEAD(&rqpair->resources->incoming_queue, link); 896 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 897 if (rc) { 898 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 899 } 900 } 901 } 902 } 903 } 904 905 if (rqpair->cm_id) { 906 if (rqpair->cm_id->qp != NULL) { 907 rdma_destroy_qp(rqpair->cm_id); 908 } 909 rdma_destroy_id(rqpair->cm_id); 910 911 if (rqpair->poller != NULL && rqpair->srq == NULL) { 912 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 913 } 914 } 915 916 if (rqpair->srq == NULL && rqpair->resources != NULL) { 917 nvmf_rdma_resources_destroy(rqpair->resources); 918 } 919 920 free(rqpair); 921 } 922 923 static int 924 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 925 { 926 struct spdk_nvmf_rdma_poller *rpoller; 927 int rc, num_cqe, required_num_wr; 928 929 /* Enlarge CQ size dynamically */ 930 rpoller = rqpair->poller; 931 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 932 num_cqe = rpoller->num_cqe; 933 if (num_cqe < required_num_wr) { 934 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 935 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 936 } 937 938 if (rpoller->num_cqe != num_cqe) { 939 if (required_num_wr > device->attr.max_cqe) { 940 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 941 required_num_wr, device->attr.max_cqe); 942 return -1; 943 } 944 945 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 946 rc = ibv_resize_cq(rpoller->cq, num_cqe); 947 if (rc) { 948 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 949 return -1; 950 } 951 952 rpoller->num_cqe = num_cqe; 953 } 954 955 rpoller->required_num_wr = required_num_wr; 956 return 0; 957 } 958 959 static int 960 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 961 { 962 struct spdk_nvmf_rdma_qpair *rqpair; 963 int rc; 964 struct spdk_nvmf_rdma_transport *rtransport; 965 struct spdk_nvmf_transport *transport; 966 struct spdk_nvmf_rdma_resource_opts opts; 967 struct spdk_nvmf_rdma_device *device; 968 struct ibv_qp_init_attr ibv_init_attr; 969 970 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 971 device = rqpair->port->device; 972 973 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 974 ibv_init_attr.qp_context = rqpair; 975 ibv_init_attr.qp_type = IBV_QPT_RC; 976 ibv_init_attr.send_cq = rqpair->poller->cq; 977 ibv_init_attr.recv_cq = rqpair->poller->cq; 978 979 if (rqpair->srq) { 980 ibv_init_attr.srq = rqpair->srq; 981 } else { 982 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 983 1; /* RECV operations + dummy drain WR */ 984 } 985 986 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 987 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 988 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 989 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 990 991 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 992 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 993 goto error; 994 } 995 996 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 997 if (rc) { 998 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 999 goto error; 1000 } 1001 1002 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2 + 1), 1003 ibv_init_attr.cap.max_send_wr); 1004 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 1005 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 1006 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 1007 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 1008 1009 rqpair->sends_to_post.first = NULL; 1010 rqpair->sends_to_post.last = NULL; 1011 1012 if (rqpair->poller->srq == NULL) { 1013 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1014 transport = &rtransport->transport; 1015 1016 opts.qp = rqpair->cm_id->qp; 1017 opts.pd = rqpair->cm_id->pd; 1018 opts.qpair = rqpair; 1019 opts.shared = false; 1020 opts.max_queue_depth = rqpair->max_queue_depth; 1021 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1022 1023 rqpair->resources = nvmf_rdma_resources_create(&opts); 1024 1025 if (!rqpair->resources) { 1026 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1027 goto error; 1028 } 1029 } else { 1030 rqpair->resources = rqpair->poller->resources; 1031 } 1032 1033 rqpair->current_recv_depth = 0; 1034 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1035 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1036 1037 return 0; 1038 1039 error: 1040 rdma_destroy_id(rqpair->cm_id); 1041 rqpair->cm_id = NULL; 1042 spdk_nvmf_rdma_qpair_destroy(rqpair); 1043 return -1; 1044 } 1045 1046 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1047 /* This function accepts either a single wr or the first wr in a linked list. */ 1048 static void 1049 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1050 { 1051 struct ibv_recv_wr *last; 1052 1053 last = first; 1054 while (last->next != NULL) { 1055 last = last->next; 1056 } 1057 1058 if (rqpair->resources->recvs_to_post.first == NULL) { 1059 rqpair->resources->recvs_to_post.first = first; 1060 rqpair->resources->recvs_to_post.last = last; 1061 if (rqpair->srq == NULL) { 1062 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1063 } 1064 } else { 1065 rqpair->resources->recvs_to_post.last->next = first; 1066 rqpair->resources->recvs_to_post.last = last; 1067 } 1068 } 1069 1070 /* Append the given send wr structure to the qpair's outstanding sends list. */ 1071 /* This function accepts either a single wr or the first wr in a linked list. */ 1072 static void 1073 nvmf_rdma_qpair_queue_send_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *first) 1074 { 1075 struct ibv_send_wr *last; 1076 1077 last = first; 1078 while (last->next != NULL) { 1079 last = last->next; 1080 } 1081 1082 if (rqpair->sends_to_post.first == NULL) { 1083 rqpair->sends_to_post.first = first; 1084 rqpair->sends_to_post.last = last; 1085 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1086 } else { 1087 rqpair->sends_to_post.last->next = first; 1088 rqpair->sends_to_post.last = last; 1089 } 1090 } 1091 1092 static int 1093 request_transfer_in(struct spdk_nvmf_request *req) 1094 { 1095 struct spdk_nvmf_rdma_request *rdma_req; 1096 struct spdk_nvmf_qpair *qpair; 1097 struct spdk_nvmf_rdma_qpair *rqpair; 1098 1099 qpair = req->qpair; 1100 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1101 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1102 1103 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1104 assert(rdma_req != NULL); 1105 1106 nvmf_rdma_qpair_queue_send_wrs(rqpair, &rdma_req->data.wr); 1107 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1108 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1109 return 0; 1110 } 1111 1112 static int 1113 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1114 { 1115 int num_outstanding_data_wr = 0; 1116 struct spdk_nvmf_rdma_request *rdma_req; 1117 struct spdk_nvmf_qpair *qpair; 1118 struct spdk_nvmf_rdma_qpair *rqpair; 1119 struct spdk_nvme_cpl *rsp; 1120 struct ibv_send_wr *first = NULL; 1121 1122 *data_posted = 0; 1123 qpair = req->qpair; 1124 rsp = &req->rsp->nvme_cpl; 1125 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1126 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1127 1128 /* Advance our sq_head pointer */ 1129 if (qpair->sq_head == qpair->sq_head_max) { 1130 qpair->sq_head = 0; 1131 } else { 1132 qpair->sq_head++; 1133 } 1134 rsp->sqhd = qpair->sq_head; 1135 1136 /* queue the capsule for the recv buffer */ 1137 assert(rdma_req->recv != NULL); 1138 1139 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1140 1141 rdma_req->recv = NULL; 1142 assert(rqpair->current_recv_depth > 0); 1143 rqpair->current_recv_depth--; 1144 1145 /* Build the response which consists of optional 1146 * RDMA WRITEs to transfer data, plus an RDMA SEND 1147 * containing the response. 1148 */ 1149 first = &rdma_req->rsp.wr; 1150 1151 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 1152 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1153 first = &rdma_req->data.wr; 1154 *data_posted = 1; 1155 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1156 } 1157 nvmf_rdma_qpair_queue_send_wrs(rqpair, first); 1158 /* +1 for the rsp wr */ 1159 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1160 1161 return 0; 1162 } 1163 1164 static int 1165 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1166 { 1167 struct spdk_nvmf_rdma_accept_private_data accept_data; 1168 struct rdma_conn_param ctrlr_event_data = {}; 1169 int rc; 1170 1171 accept_data.recfmt = 0; 1172 accept_data.crqsize = rqpair->max_queue_depth; 1173 1174 ctrlr_event_data.private_data = &accept_data; 1175 ctrlr_event_data.private_data_len = sizeof(accept_data); 1176 if (id->ps == RDMA_PS_TCP) { 1177 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1178 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1179 } 1180 1181 /* Configure infinite retries for the initiator side qpair. 1182 * When using a shared receive queue on the target side, 1183 * we need to pass this value to the initiator to prevent the 1184 * initiator side NIC from completing SEND requests back to the 1185 * initiator with status rnr_retry_count_exceeded. */ 1186 if (rqpair->srq != NULL) { 1187 ctrlr_event_data.rnr_retry_count = 0x7; 1188 } 1189 1190 rc = rdma_accept(id, &ctrlr_event_data); 1191 if (rc) { 1192 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 1193 } else { 1194 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1195 } 1196 1197 return rc; 1198 } 1199 1200 static void 1201 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1202 { 1203 struct spdk_nvmf_rdma_reject_private_data rej_data; 1204 1205 rej_data.recfmt = 0; 1206 rej_data.sts = error; 1207 1208 rdma_reject(id, &rej_data, sizeof(rej_data)); 1209 } 1210 1211 static int 1212 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 1213 new_qpair_fn cb_fn) 1214 { 1215 struct spdk_nvmf_rdma_transport *rtransport; 1216 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1217 struct spdk_nvmf_rdma_port *port; 1218 struct rdma_conn_param *rdma_param = NULL; 1219 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1220 uint16_t max_queue_depth; 1221 uint16_t max_read_depth; 1222 1223 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1224 1225 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1226 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1227 1228 rdma_param = &event->param.conn; 1229 if (rdma_param->private_data == NULL || 1230 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1231 SPDK_ERRLOG("connect request: no private data provided\n"); 1232 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1233 return -1; 1234 } 1235 1236 private_data = rdma_param->private_data; 1237 if (private_data->recfmt != 0) { 1238 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1239 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1240 return -1; 1241 } 1242 1243 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1244 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1245 1246 port = event->listen_id->context; 1247 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1248 event->listen_id, event->listen_id->verbs, port); 1249 1250 /* Figure out the supported queue depth. This is a multi-step process 1251 * that takes into account hardware maximums, host provided values, 1252 * and our target's internal memory limits */ 1253 1254 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1255 1256 /* Start with the maximum queue depth allowed by the target */ 1257 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1258 max_read_depth = rtransport->transport.opts.max_queue_depth; 1259 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1260 rtransport->transport.opts.max_queue_depth); 1261 1262 /* Next check the local NIC's hardware limitations */ 1263 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1264 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1265 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1266 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1267 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1268 1269 /* Next check the remote NIC's hardware limitations */ 1270 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1271 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1272 rdma_param->initiator_depth, rdma_param->responder_resources); 1273 if (rdma_param->initiator_depth > 0) { 1274 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1275 } 1276 1277 /* Finally check for the host software requested values, which are 1278 * optional. */ 1279 if (rdma_param->private_data != NULL && 1280 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1281 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1282 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1283 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1284 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1285 } 1286 1287 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1288 max_queue_depth, max_read_depth); 1289 1290 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1291 if (rqpair == NULL) { 1292 SPDK_ERRLOG("Could not allocate new connection.\n"); 1293 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1294 return -1; 1295 } 1296 1297 rqpair->port = port; 1298 rqpair->max_queue_depth = max_queue_depth; 1299 rqpair->max_read_depth = max_read_depth; 1300 rqpair->cm_id = event->id; 1301 rqpair->listen_id = event->listen_id; 1302 rqpair->qpair.transport = transport; 1303 1304 event->id->context = &rqpair->qpair; 1305 1306 cb_fn(&rqpair->qpair); 1307 1308 return 0; 1309 } 1310 1311 static int 1312 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1313 enum spdk_mem_map_notify_action action, 1314 void *vaddr, size_t size) 1315 { 1316 struct ibv_pd *pd = cb_ctx; 1317 struct ibv_mr *mr; 1318 1319 switch (action) { 1320 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1321 if (!g_nvmf_hooks.get_rkey) { 1322 mr = ibv_reg_mr(pd, vaddr, size, 1323 IBV_ACCESS_LOCAL_WRITE | 1324 IBV_ACCESS_REMOTE_READ | 1325 IBV_ACCESS_REMOTE_WRITE); 1326 if (mr == NULL) { 1327 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1328 return -1; 1329 } else { 1330 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1331 } 1332 } else { 1333 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1334 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1335 } 1336 break; 1337 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1338 if (!g_nvmf_hooks.get_rkey) { 1339 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1340 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1341 if (mr) { 1342 ibv_dereg_mr(mr); 1343 } 1344 } 1345 break; 1346 } 1347 1348 return 0; 1349 } 1350 1351 static int 1352 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1353 { 1354 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1355 return addr_1 == addr_2; 1356 } 1357 1358 static void 1359 spdk_nvmf_rdma_request_free_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1360 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1361 uint32_t num_buffers) 1362 { 1363 uint32_t i; 1364 1365 for (i = 0; i < num_buffers; i++) { 1366 if (group->buf_cache_count < group->buf_cache_size) { 1367 STAILQ_INSERT_HEAD(&group->buf_cache, 1368 (struct spdk_nvmf_transport_pg_cache_buf *)rdma_req->buffers[i], link); 1369 group->buf_cache_count++; 1370 } else { 1371 spdk_mempool_put(transport->data_buf_pool, rdma_req->buffers[i]); 1372 } 1373 rdma_req->req.iov[i].iov_base = NULL; 1374 rdma_req->buffers[i] = NULL; 1375 rdma_req->req.iov[i].iov_len = 0; 1376 1377 } 1378 rdma_req->data_from_pool = false; 1379 } 1380 1381 static int 1382 nvmf_rdma_request_get_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1383 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1384 uint32_t num_buffers) 1385 { 1386 uint32_t i = 0; 1387 1388 while (i < num_buffers) { 1389 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1390 group->buf_cache_count--; 1391 rdma_req->buffers[i] = STAILQ_FIRST(&group->buf_cache); 1392 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1393 assert(rdma_req->buffers[i] != NULL); 1394 i++; 1395 } else { 1396 if (spdk_mempool_get_bulk(transport->data_buf_pool, &rdma_req->buffers[i], num_buffers - i)) { 1397 goto err_exit; 1398 } 1399 i += num_buffers - i; 1400 } 1401 } 1402 1403 return 0; 1404 1405 err_exit: 1406 spdk_nvmf_rdma_request_free_buffers(rdma_req, group, transport, i); 1407 return -ENOMEM; 1408 } 1409 1410 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1411 1412 static spdk_nvme_data_transfer_t 1413 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1414 { 1415 enum spdk_nvme_data_transfer xfer; 1416 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1417 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1418 1419 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1420 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1421 rdma_req->rsp.wr.imm_data = 0; 1422 #endif 1423 1424 /* Figure out data transfer direction */ 1425 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1426 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1427 } else { 1428 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1429 1430 /* Some admin commands are special cases */ 1431 if ((rdma_req->req.qpair->qid == 0) && 1432 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1433 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1434 switch (cmd->cdw10 & 0xff) { 1435 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1436 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1437 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1438 break; 1439 default: 1440 xfer = SPDK_NVME_DATA_NONE; 1441 } 1442 } 1443 } 1444 1445 if (xfer == SPDK_NVME_DATA_NONE) { 1446 return xfer; 1447 } 1448 1449 /* Even for commands that may transfer data, they could have specified 0 length. 1450 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1451 */ 1452 switch (sgl->generic.type) { 1453 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1454 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1455 case SPDK_NVME_SGL_TYPE_SEGMENT: 1456 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1457 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1458 if (sgl->unkeyed.length == 0) { 1459 xfer = SPDK_NVME_DATA_NONE; 1460 } 1461 break; 1462 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1463 if (sgl->keyed.length == 0) { 1464 xfer = SPDK_NVME_DATA_NONE; 1465 } 1466 break; 1467 } 1468 1469 return xfer; 1470 } 1471 1472 static int 1473 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1474 struct spdk_nvmf_rdma_request *rdma_req, 1475 uint32_t num_sgl_descriptors) 1476 { 1477 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1478 struct spdk_nvmf_rdma_request_data *current_data_wr; 1479 uint32_t i; 1480 1481 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1482 return -ENOMEM; 1483 } 1484 1485 current_data_wr = &rdma_req->data; 1486 1487 for (i = 0; i < num_sgl_descriptors; i++) { 1488 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1489 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1490 current_data_wr->wr.send_flags = 0; 1491 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1492 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1493 current_data_wr->wr.send_flags = IBV_SEND_SIGNALED; 1494 } else { 1495 assert(false); 1496 } 1497 work_requests[i]->wr.sg_list = work_requests[i]->sgl; 1498 work_requests[i]->wr.wr_id = rdma_req->data.wr.wr_id; 1499 current_data_wr->wr.next = &work_requests[i]->wr; 1500 current_data_wr = work_requests[i]; 1501 } 1502 1503 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1504 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1505 current_data_wr->wr.next = &rdma_req->rsp.wr; 1506 current_data_wr->wr.send_flags = 0; 1507 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1508 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1509 current_data_wr->wr.next = NULL; 1510 current_data_wr->wr.send_flags = IBV_SEND_SIGNALED; 1511 } 1512 return 0; 1513 } 1514 1515 static int 1516 nvmf_rdma_fill_buffers(struct spdk_nvmf_rdma_transport *rtransport, 1517 struct spdk_nvmf_rdma_poll_group *rgroup, 1518 struct spdk_nvmf_rdma_device *device, 1519 struct spdk_nvmf_rdma_request *rdma_req, 1520 struct ibv_send_wr *wr, 1521 uint32_t length) 1522 { 1523 uint64_t translation_len; 1524 uint32_t remaining_length = length; 1525 uint32_t iovcnt; 1526 uint32_t i = 0; 1527 1528 1529 while (remaining_length) { 1530 iovcnt = rdma_req->req.iovcnt; 1531 rdma_req->req.iov[iovcnt].iov_base = (void *)((uintptr_t)(rdma_req->buffers[iovcnt] + 1532 NVMF_DATA_BUFFER_MASK) & 1533 ~NVMF_DATA_BUFFER_MASK); 1534 rdma_req->req.iov[iovcnt].iov_len = spdk_min(remaining_length, 1535 rtransport->transport.opts.io_unit_size); 1536 rdma_req->req.iovcnt++; 1537 wr->sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[iovcnt].iov_base); 1538 wr->sg_list[i].length = rdma_req->req.iov[iovcnt].iov_len; 1539 translation_len = rdma_req->req.iov[iovcnt].iov_len; 1540 1541 if (!g_nvmf_hooks.get_rkey) { 1542 wr->sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1543 (uint64_t)rdma_req->buffers[iovcnt], &translation_len))->lkey; 1544 } else { 1545 wr->sg_list[i].lkey = spdk_mem_map_translate(device->map, 1546 (uint64_t)rdma_req->buffers[iovcnt], &translation_len); 1547 } 1548 1549 remaining_length -= rdma_req->req.iov[iovcnt].iov_len; 1550 1551 if (translation_len < rdma_req->req.iov[iovcnt].iov_len) { 1552 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions\n"); 1553 return -EINVAL; 1554 } 1555 i++; 1556 } 1557 wr->num_sge = i; 1558 1559 return 0; 1560 } 1561 1562 static int 1563 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1564 struct spdk_nvmf_rdma_device *device, 1565 struct spdk_nvmf_rdma_request *rdma_req) 1566 { 1567 struct spdk_nvmf_rdma_qpair *rqpair; 1568 struct spdk_nvmf_rdma_poll_group *rgroup; 1569 uint32_t num_buffers; 1570 uint32_t i = 0; 1571 int rc = 0; 1572 1573 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1574 rgroup = rqpair->poller->group; 1575 rdma_req->req.iovcnt = 0; 1576 1577 num_buffers = rdma_req->req.length / rtransport->transport.opts.io_unit_size; 1578 if (rdma_req->req.length % rtransport->transport.opts.io_unit_size) { 1579 num_buffers++; 1580 } 1581 1582 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers)) { 1583 return -ENOMEM; 1584 } 1585 1586 rdma_req->req.iovcnt = 0; 1587 1588 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, &rdma_req->data.wr, 1589 rdma_req->req.length); 1590 if (rc != 0) { 1591 goto err_exit; 1592 } 1593 1594 assert(rdma_req->req.iovcnt <= rqpair->max_send_sge); 1595 1596 rdma_req->data_from_pool = true; 1597 1598 return rc; 1599 1600 err_exit: 1601 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1602 while (i) { 1603 i--; 1604 rdma_req->data.wr.sg_list[i].addr = 0; 1605 rdma_req->data.wr.sg_list[i].length = 0; 1606 rdma_req->data.wr.sg_list[i].lkey = 0; 1607 } 1608 rdma_req->req.iovcnt = 0; 1609 return rc; 1610 } 1611 1612 static int 1613 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1614 struct spdk_nvmf_rdma_device *device, 1615 struct spdk_nvmf_rdma_request *rdma_req) 1616 { 1617 struct spdk_nvmf_rdma_qpair *rqpair; 1618 struct spdk_nvmf_rdma_poll_group *rgroup; 1619 struct ibv_send_wr *current_wr; 1620 struct spdk_nvmf_request *req = &rdma_req->req; 1621 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1622 uint32_t num_sgl_descriptors; 1623 uint32_t num_buffers = 0; 1624 uint32_t i; 1625 int rc; 1626 1627 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1628 rgroup = rqpair->poller->group; 1629 1630 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1631 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1632 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1633 1634 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1635 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1636 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1637 1638 for (i = 0; i < num_sgl_descriptors; i++) { 1639 num_buffers += desc->keyed.length / rtransport->transport.opts.io_unit_size; 1640 if (desc->keyed.length % rtransport->transport.opts.io_unit_size) { 1641 num_buffers++; 1642 } 1643 desc++; 1644 } 1645 /* If the number of buffers is too large, then we know the I/O is larger than allowed. Fail it. */ 1646 if (num_buffers > NVMF_REQ_MAX_BUFFERS) { 1647 return -EINVAL; 1648 } 1649 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1650 num_buffers) != 0) { 1651 return -ENOMEM; 1652 } 1653 1654 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1655 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1656 return -ENOMEM; 1657 } 1658 1659 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1660 current_wr = &rdma_req->data.wr; 1661 assert(current_wr != NULL); 1662 1663 req->iovcnt = 0; 1664 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1665 for (i = 0; i < num_sgl_descriptors; i++) { 1666 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1667 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1668 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1669 rc = -EINVAL; 1670 goto err_exit; 1671 } 1672 1673 current_wr->num_sge = 0; 1674 req->length += desc->keyed.length; 1675 1676 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, current_wr, 1677 desc->keyed.length); 1678 if (rc != 0) { 1679 rc = -ENOMEM; 1680 goto err_exit; 1681 } 1682 1683 current_wr->wr.rdma.rkey = desc->keyed.key; 1684 current_wr->wr.rdma.remote_addr = desc->address; 1685 current_wr = current_wr->next; 1686 desc++; 1687 } 1688 1689 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1690 /* Go back to the last descriptor in the list. */ 1691 desc--; 1692 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1693 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1694 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1695 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1696 } 1697 } 1698 #endif 1699 1700 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1701 rdma_req->data_from_pool = true; 1702 1703 return 0; 1704 1705 err_exit: 1706 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1707 nvmf_rdma_request_free_data(rdma_req, rtransport); 1708 return rc; 1709 } 1710 1711 static int 1712 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1713 struct spdk_nvmf_rdma_device *device, 1714 struct spdk_nvmf_rdma_request *rdma_req) 1715 { 1716 struct spdk_nvme_cmd *cmd; 1717 struct spdk_nvme_cpl *rsp; 1718 struct spdk_nvme_sgl_descriptor *sgl; 1719 int rc; 1720 1721 cmd = &rdma_req->req.cmd->nvme_cmd; 1722 rsp = &rdma_req->req.rsp->nvme_cpl; 1723 sgl = &cmd->dptr.sgl1; 1724 1725 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1726 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1727 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1728 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1729 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1730 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1731 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1732 return -1; 1733 } 1734 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1735 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1736 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1737 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1738 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1739 } 1740 } 1741 #endif 1742 1743 /* fill request length and populate iovs */ 1744 rdma_req->req.length = sgl->keyed.length; 1745 1746 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1747 /* No available buffers. Queue this request up. */ 1748 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1749 return 0; 1750 } 1751 1752 /* backward compatible */ 1753 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1754 1755 /* rdma wr specifics */ 1756 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1757 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1758 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1759 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1760 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 1761 rdma_req->data.wr.next = &rdma_req->rsp.wr; 1762 rdma_req->data.wr.send_flags &= ~IBV_SEND_SIGNALED; 1763 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1764 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 1765 rdma_req->data.wr.next = NULL; 1766 rdma_req->data.wr.send_flags |= IBV_SEND_SIGNALED; 1767 } 1768 1769 /* set the number of outstanding data WRs for this request. */ 1770 rdma_req->num_outstanding_data_wr = 1; 1771 1772 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1773 rdma_req->req.iovcnt); 1774 1775 return 0; 1776 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1777 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1778 uint64_t offset = sgl->address; 1779 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1780 1781 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1782 offset, sgl->unkeyed.length); 1783 1784 if (offset > max_len) { 1785 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1786 offset, max_len); 1787 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1788 return -1; 1789 } 1790 max_len -= (uint32_t)offset; 1791 1792 if (sgl->unkeyed.length > max_len) { 1793 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1794 sgl->unkeyed.length, max_len); 1795 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1796 return -1; 1797 } 1798 1799 rdma_req->num_outstanding_data_wr = 0; 1800 rdma_req->req.data = rdma_req->recv->buf + offset; 1801 rdma_req->data_from_pool = false; 1802 rdma_req->req.length = sgl->unkeyed.length; 1803 1804 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1805 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1806 rdma_req->req.iovcnt = 1; 1807 1808 return 0; 1809 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1810 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1811 1812 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1813 if (rc == -ENOMEM) { 1814 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1815 return 0; 1816 } else if (rc == -EINVAL) { 1817 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1818 return -1; 1819 } 1820 1821 /* backward compatible */ 1822 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1823 1824 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1825 rdma_req->req.iovcnt); 1826 1827 return 0; 1828 } 1829 1830 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1831 sgl->generic.type, sgl->generic.subtype); 1832 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1833 return -1; 1834 } 1835 1836 static void 1837 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1838 struct spdk_nvmf_rdma_transport *rtransport) 1839 { 1840 struct spdk_nvmf_rdma_qpair *rqpair; 1841 struct spdk_nvmf_rdma_poll_group *rgroup; 1842 1843 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1844 if (rdma_req->data_from_pool) { 1845 rgroup = rqpair->poller->group; 1846 1847 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1848 rdma_req->req.iovcnt); 1849 } 1850 nvmf_rdma_request_free_data(rdma_req, rtransport); 1851 rdma_req->req.length = 0; 1852 rdma_req->req.iovcnt = 0; 1853 rdma_req->req.data = NULL; 1854 rdma_req->rsp.wr.next = NULL; 1855 rdma_req->data.wr.next = NULL; 1856 rqpair->qd--; 1857 1858 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1859 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1860 } 1861 1862 static bool 1863 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1864 struct spdk_nvmf_rdma_request *rdma_req) 1865 { 1866 struct spdk_nvmf_rdma_qpair *rqpair; 1867 struct spdk_nvmf_rdma_device *device; 1868 struct spdk_nvmf_rdma_poll_group *rgroup; 1869 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1870 int rc; 1871 struct spdk_nvmf_rdma_recv *rdma_recv; 1872 enum spdk_nvmf_rdma_request_state prev_state; 1873 bool progress = false; 1874 int data_posted; 1875 1876 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1877 device = rqpair->port->device; 1878 rgroup = rqpair->poller->group; 1879 1880 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1881 1882 /* If the queue pair is in an error state, force the request to the completed state 1883 * to release resources. */ 1884 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1885 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1886 STAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1887 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1888 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1889 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1890 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1891 } 1892 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1893 } 1894 1895 /* The loop here is to allow for several back-to-back state changes. */ 1896 do { 1897 prev_state = rdma_req->state; 1898 1899 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1900 1901 switch (rdma_req->state) { 1902 case RDMA_REQUEST_STATE_FREE: 1903 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1904 * to escape this state. */ 1905 break; 1906 case RDMA_REQUEST_STATE_NEW: 1907 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1908 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1909 rdma_recv = rdma_req->recv; 1910 1911 /* The first element of the SGL is the NVMe command */ 1912 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1913 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1914 1915 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1916 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1917 break; 1918 } 1919 1920 /* The next state transition depends on the data transfer needs of this request. */ 1921 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1922 1923 /* If no data to transfer, ready to execute. */ 1924 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1925 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1926 break; 1927 } 1928 1929 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1930 STAILQ_INSERT_TAIL(&rgroup->pending_data_buf_queue, rdma_req, state_link); 1931 break; 1932 case RDMA_REQUEST_STATE_NEED_BUFFER: 1933 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1934 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1935 1936 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1937 1938 if (rdma_req != STAILQ_FIRST(&rgroup->pending_data_buf_queue)) { 1939 /* This request needs to wait in line to obtain a buffer */ 1940 break; 1941 } 1942 1943 /* Try to get a data buffer */ 1944 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1945 if (rc < 0) { 1946 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1947 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1948 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1949 break; 1950 } 1951 1952 if (!rdma_req->req.data) { 1953 /* No buffers available. */ 1954 break; 1955 } 1956 1957 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1958 1959 /* If data is transferring from host to controller and the data didn't 1960 * arrive using in capsule data, we need to do a transfer from the host. 1961 */ 1962 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1963 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1964 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1965 break; 1966 } 1967 1968 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1969 break; 1970 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1971 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1972 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1973 1974 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1975 /* This request needs to wait in line to perform RDMA */ 1976 break; 1977 } 1978 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1979 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1980 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1981 break; 1982 } 1983 1984 /* We have already verified that this request is the head of the queue. */ 1985 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 1986 1987 rc = request_transfer_in(&rdma_req->req); 1988 if (!rc) { 1989 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 1990 } else { 1991 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1992 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1993 } 1994 break; 1995 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1996 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1997 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1998 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1999 * to escape this state. */ 2000 break; 2001 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2002 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2003 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2004 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2005 spdk_nvmf_request_exec(&rdma_req->req); 2006 break; 2007 case RDMA_REQUEST_STATE_EXECUTING: 2008 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2009 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2010 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2011 * to escape this state. */ 2012 break; 2013 case RDMA_REQUEST_STATE_EXECUTED: 2014 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2015 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2016 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2017 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2018 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2019 } else { 2020 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2021 } 2022 break; 2023 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2024 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2025 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2026 2027 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2028 /* This request needs to wait in line to perform RDMA */ 2029 break; 2030 } 2031 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2032 rqpair->max_send_depth) { 2033 /* We can only have so many WRs outstanding. we have to wait until some finish. 2034 * +1 since each request has an additional wr in the resp. */ 2035 break; 2036 } 2037 2038 /* We have already verified that this request is the head of the queue. */ 2039 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2040 2041 /* The data transfer will be kicked off from 2042 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2043 */ 2044 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2045 break; 2046 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2047 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2048 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2049 rc = request_transfer_out(&rdma_req->req, &data_posted); 2050 assert(rc == 0); /* No good way to handle this currently */ 2051 if (rc) { 2052 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2053 } else { 2054 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2055 RDMA_REQUEST_STATE_COMPLETING; 2056 } 2057 break; 2058 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2059 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2060 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2061 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2062 * to escape this state. */ 2063 break; 2064 case RDMA_REQUEST_STATE_COMPLETING: 2065 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2066 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2067 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2068 * to escape this state. */ 2069 break; 2070 case RDMA_REQUEST_STATE_COMPLETED: 2071 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2072 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2073 2074 nvmf_rdma_request_free(rdma_req, rtransport); 2075 break; 2076 case RDMA_REQUEST_NUM_STATES: 2077 default: 2078 assert(0); 2079 break; 2080 } 2081 2082 if (rdma_req->state != prev_state) { 2083 progress = true; 2084 } 2085 } while (rdma_req->state != prev_state); 2086 2087 return progress; 2088 } 2089 2090 /* Public API callbacks begin here */ 2091 2092 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2093 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2094 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2095 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2096 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2097 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2098 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2099 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2100 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2101 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false; 2102 2103 static void 2104 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2105 { 2106 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2107 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2108 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2109 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2110 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2111 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2112 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2113 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2114 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2115 opts->no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ 2116 } 2117 2118 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2119 .notify_cb = spdk_nvmf_rdma_mem_notify, 2120 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 2121 }; 2122 2123 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2124 2125 static struct spdk_nvmf_transport * 2126 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2127 { 2128 int rc; 2129 struct spdk_nvmf_rdma_transport *rtransport; 2130 struct spdk_nvmf_rdma_device *device, *tmp; 2131 struct ibv_pd *pd; 2132 struct ibv_context **contexts; 2133 uint32_t i; 2134 int flag; 2135 uint32_t sge_count; 2136 uint32_t min_shared_buffers; 2137 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2138 2139 rtransport = calloc(1, sizeof(*rtransport)); 2140 if (!rtransport) { 2141 return NULL; 2142 } 2143 2144 if (pthread_mutex_init(&rtransport->lock, NULL)) { 2145 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2146 free(rtransport); 2147 return NULL; 2148 } 2149 2150 TAILQ_INIT(&rtransport->devices); 2151 TAILQ_INIT(&rtransport->ports); 2152 2153 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2154 2155 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2156 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2157 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2158 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2159 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d\n", 2160 opts->max_queue_depth, 2161 opts->max_io_size, 2162 opts->max_qpairs_per_ctrlr, 2163 opts->io_unit_size, 2164 opts->in_capsule_data_size, 2165 opts->max_aq_depth, 2166 opts->num_shared_buffers, 2167 opts->max_srq_depth, 2168 opts->no_srq); 2169 2170 /* I/O unit size cannot be larger than max I/O size */ 2171 if (opts->io_unit_size > opts->max_io_size) { 2172 opts->io_unit_size = opts->max_io_size; 2173 } 2174 2175 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2176 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2177 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2178 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2179 spdk_nvmf_rdma_destroy(&rtransport->transport); 2180 return NULL; 2181 } 2182 2183 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2184 if (min_shared_buffers > opts->num_shared_buffers) { 2185 SPDK_ERRLOG("There are not enough buffers to satisfy" 2186 "per-poll group caches for each thread. (%" PRIu32 ")" 2187 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2188 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2189 spdk_nvmf_rdma_destroy(&rtransport->transport); 2190 return NULL; 2191 } 2192 2193 sge_count = opts->max_io_size / opts->io_unit_size; 2194 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2195 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2196 spdk_nvmf_rdma_destroy(&rtransport->transport); 2197 return NULL; 2198 } 2199 2200 rtransport->event_channel = rdma_create_event_channel(); 2201 if (rtransport->event_channel == NULL) { 2202 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2203 spdk_nvmf_rdma_destroy(&rtransport->transport); 2204 return NULL; 2205 } 2206 2207 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2208 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2209 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2210 rtransport->event_channel->fd, spdk_strerror(errno)); 2211 spdk_nvmf_rdma_destroy(&rtransport->transport); 2212 return NULL; 2213 } 2214 2215 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2216 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2217 sizeof(struct spdk_nvmf_rdma_request_data), 2218 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2219 SPDK_ENV_SOCKET_ID_ANY); 2220 if (!rtransport->data_wr_pool) { 2221 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2222 spdk_nvmf_rdma_destroy(&rtransport->transport); 2223 return NULL; 2224 } 2225 2226 contexts = rdma_get_devices(NULL); 2227 if (contexts == NULL) { 2228 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2229 spdk_nvmf_rdma_destroy(&rtransport->transport); 2230 return NULL; 2231 } 2232 2233 i = 0; 2234 rc = 0; 2235 while (contexts[i] != NULL) { 2236 device = calloc(1, sizeof(*device)); 2237 if (!device) { 2238 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2239 rc = -ENOMEM; 2240 break; 2241 } 2242 device->context = contexts[i]; 2243 rc = ibv_query_device(device->context, &device->attr); 2244 if (rc < 0) { 2245 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2246 free(device); 2247 break; 2248 2249 } 2250 2251 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2252 2253 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2254 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2255 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2256 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2257 } 2258 2259 /** 2260 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2261 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2262 * but incorrectly reports that it does. There are changes making their way 2263 * through the kernel now that will enable this feature. When they are merged, 2264 * we can conditionally enable this feature. 2265 * 2266 * TODO: enable this for versions of the kernel rxe driver that support it. 2267 */ 2268 if (device->attr.vendor_id == 0) { 2269 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2270 } 2271 #endif 2272 2273 /* set up device context async ev fd as NON_BLOCKING */ 2274 flag = fcntl(device->context->async_fd, F_GETFL); 2275 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2276 if (rc < 0) { 2277 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2278 free(device); 2279 break; 2280 } 2281 2282 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2283 i++; 2284 2285 pd = NULL; 2286 if (g_nvmf_hooks.get_ibv_pd) { 2287 pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2288 } 2289 2290 if (!g_nvmf_hooks.get_ibv_pd) { 2291 device->pd = ibv_alloc_pd(device->context); 2292 if (!device->pd) { 2293 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2294 spdk_nvmf_rdma_destroy(&rtransport->transport); 2295 return NULL; 2296 } 2297 } else { 2298 device->pd = pd; 2299 } 2300 2301 assert(device->map == NULL); 2302 2303 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2304 if (!device->map) { 2305 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2306 spdk_nvmf_rdma_destroy(&rtransport->transport); 2307 return NULL; 2308 } 2309 2310 assert(device->map != NULL); 2311 assert(device->pd != NULL); 2312 } 2313 rdma_free_devices(contexts); 2314 2315 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2316 /* divide and round up. */ 2317 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2318 2319 /* round up to the nearest 4k. */ 2320 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2321 2322 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2323 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2324 opts->io_unit_size); 2325 } 2326 2327 if (rc < 0) { 2328 spdk_nvmf_rdma_destroy(&rtransport->transport); 2329 return NULL; 2330 } 2331 2332 /* Set up poll descriptor array to monitor events from RDMA and IB 2333 * in a single poll syscall 2334 */ 2335 rtransport->npoll_fds = i + 1; 2336 i = 0; 2337 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2338 if (rtransport->poll_fds == NULL) { 2339 SPDK_ERRLOG("poll_fds allocation failed\n"); 2340 spdk_nvmf_rdma_destroy(&rtransport->transport); 2341 return NULL; 2342 } 2343 2344 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2345 rtransport->poll_fds[i++].events = POLLIN; 2346 2347 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2348 rtransport->poll_fds[i].fd = device->context->async_fd; 2349 rtransport->poll_fds[i++].events = POLLIN; 2350 } 2351 2352 return &rtransport->transport; 2353 } 2354 2355 static int 2356 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2357 { 2358 struct spdk_nvmf_rdma_transport *rtransport; 2359 struct spdk_nvmf_rdma_port *port, *port_tmp; 2360 struct spdk_nvmf_rdma_device *device, *device_tmp; 2361 2362 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2363 2364 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2365 TAILQ_REMOVE(&rtransport->ports, port, link); 2366 rdma_destroy_id(port->id); 2367 free(port); 2368 } 2369 2370 if (rtransport->poll_fds != NULL) { 2371 free(rtransport->poll_fds); 2372 } 2373 2374 if (rtransport->event_channel != NULL) { 2375 rdma_destroy_event_channel(rtransport->event_channel); 2376 } 2377 2378 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2379 TAILQ_REMOVE(&rtransport->devices, device, link); 2380 if (device->map) { 2381 spdk_mem_map_free(&device->map); 2382 } 2383 if (device->pd) { 2384 if (!g_nvmf_hooks.get_ibv_pd) { 2385 ibv_dealloc_pd(device->pd); 2386 } 2387 } 2388 free(device); 2389 } 2390 2391 if (rtransport->data_wr_pool != NULL) { 2392 if (spdk_mempool_count(rtransport->data_wr_pool) != 2393 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2394 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2395 spdk_mempool_count(rtransport->data_wr_pool), 2396 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2397 } 2398 } 2399 2400 spdk_mempool_free(rtransport->data_wr_pool); 2401 pthread_mutex_destroy(&rtransport->lock); 2402 free(rtransport); 2403 2404 return 0; 2405 } 2406 2407 static int 2408 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2409 struct spdk_nvme_transport_id *trid, 2410 bool peer); 2411 2412 static int 2413 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2414 const struct spdk_nvme_transport_id *trid) 2415 { 2416 struct spdk_nvmf_rdma_transport *rtransport; 2417 struct spdk_nvmf_rdma_device *device; 2418 struct spdk_nvmf_rdma_port *port_tmp, *port; 2419 struct addrinfo *res; 2420 struct addrinfo hints; 2421 int family; 2422 int rc; 2423 2424 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2425 2426 port = calloc(1, sizeof(*port)); 2427 if (!port) { 2428 return -ENOMEM; 2429 } 2430 2431 /* Selectively copy the trid. Things like NQN don't matter here - that 2432 * mapping is enforced elsewhere. 2433 */ 2434 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2435 port->trid.adrfam = trid->adrfam; 2436 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 2437 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 2438 2439 pthread_mutex_lock(&rtransport->lock); 2440 assert(rtransport->event_channel != NULL); 2441 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 2442 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 2443 port_tmp->ref++; 2444 free(port); 2445 /* Already listening at this address */ 2446 pthread_mutex_unlock(&rtransport->lock); 2447 return 0; 2448 } 2449 } 2450 2451 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2452 if (rc < 0) { 2453 SPDK_ERRLOG("rdma_create_id() failed\n"); 2454 free(port); 2455 pthread_mutex_unlock(&rtransport->lock); 2456 return rc; 2457 } 2458 2459 switch (port->trid.adrfam) { 2460 case SPDK_NVMF_ADRFAM_IPV4: 2461 family = AF_INET; 2462 break; 2463 case SPDK_NVMF_ADRFAM_IPV6: 2464 family = AF_INET6; 2465 break; 2466 default: 2467 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 2468 free(port); 2469 pthread_mutex_unlock(&rtransport->lock); 2470 return -EINVAL; 2471 } 2472 2473 memset(&hints, 0, sizeof(hints)); 2474 hints.ai_family = family; 2475 hints.ai_flags = AI_NUMERICSERV; 2476 hints.ai_socktype = SOCK_STREAM; 2477 hints.ai_protocol = 0; 2478 2479 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 2480 if (rc) { 2481 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2482 free(port); 2483 pthread_mutex_unlock(&rtransport->lock); 2484 return -EINVAL; 2485 } 2486 2487 rc = rdma_bind_addr(port->id, res->ai_addr); 2488 freeaddrinfo(res); 2489 2490 if (rc < 0) { 2491 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2492 rdma_destroy_id(port->id); 2493 free(port); 2494 pthread_mutex_unlock(&rtransport->lock); 2495 return rc; 2496 } 2497 2498 if (!port->id->verbs) { 2499 SPDK_ERRLOG("ibv_context is null\n"); 2500 rdma_destroy_id(port->id); 2501 free(port); 2502 pthread_mutex_unlock(&rtransport->lock); 2503 return -1; 2504 } 2505 2506 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 2507 if (rc < 0) { 2508 SPDK_ERRLOG("rdma_listen() failed\n"); 2509 rdma_destroy_id(port->id); 2510 free(port); 2511 pthread_mutex_unlock(&rtransport->lock); 2512 return rc; 2513 } 2514 2515 TAILQ_FOREACH(device, &rtransport->devices, link) { 2516 if (device->context == port->id->verbs) { 2517 port->device = device; 2518 break; 2519 } 2520 } 2521 if (!port->device) { 2522 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2523 port->id->verbs); 2524 rdma_destroy_id(port->id); 2525 free(port); 2526 pthread_mutex_unlock(&rtransport->lock); 2527 return -EINVAL; 2528 } 2529 2530 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 2531 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 2532 2533 port->ref = 1; 2534 2535 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2536 pthread_mutex_unlock(&rtransport->lock); 2537 2538 return 0; 2539 } 2540 2541 static int 2542 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2543 const struct spdk_nvme_transport_id *_trid) 2544 { 2545 struct spdk_nvmf_rdma_transport *rtransport; 2546 struct spdk_nvmf_rdma_port *port, *tmp; 2547 struct spdk_nvme_transport_id trid = {}; 2548 2549 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2550 2551 /* Selectively copy the trid. Things like NQN don't matter here - that 2552 * mapping is enforced elsewhere. 2553 */ 2554 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2555 trid.adrfam = _trid->adrfam; 2556 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2557 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2558 2559 pthread_mutex_lock(&rtransport->lock); 2560 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2561 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2562 assert(port->ref > 0); 2563 port->ref--; 2564 if (port->ref == 0) { 2565 TAILQ_REMOVE(&rtransport->ports, port, link); 2566 rdma_destroy_id(port->id); 2567 free(port); 2568 } 2569 break; 2570 } 2571 } 2572 2573 pthread_mutex_unlock(&rtransport->lock); 2574 return 0; 2575 } 2576 2577 static void 2578 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2579 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2580 { 2581 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2582 struct spdk_nvmf_rdma_resources *resources; 2583 2584 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2585 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2586 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2587 break; 2588 } 2589 } 2590 2591 /* Then RDMA writes since reads have stronger restrictions than writes */ 2592 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2593 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2594 break; 2595 } 2596 } 2597 2598 /* The second highest priority is I/O waiting on memory buffers. */ 2599 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->poller->group->pending_data_buf_queue, state_link, 2600 req_tmp) { 2601 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2602 break; 2603 } 2604 } 2605 2606 resources = rqpair->resources; 2607 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2608 rdma_req = STAILQ_FIRST(&resources->free_queue); 2609 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2610 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2611 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2612 2613 if (rqpair->srq != NULL) { 2614 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2615 rdma_req->recv->qpair->qd++; 2616 } else { 2617 rqpair->qd++; 2618 } 2619 2620 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2621 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2622 break; 2623 } 2624 } 2625 } 2626 2627 static void 2628 _nvmf_rdma_qpair_disconnect(void *ctx) 2629 { 2630 struct spdk_nvmf_qpair *qpair = ctx; 2631 2632 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2633 } 2634 2635 static void 2636 _nvmf_rdma_try_disconnect(void *ctx) 2637 { 2638 struct spdk_nvmf_qpair *qpair = ctx; 2639 struct spdk_nvmf_poll_group *group; 2640 2641 /* Read the group out of the qpair. This is normally set and accessed only from 2642 * the thread that created the group. Here, we're not on that thread necessarily. 2643 * The data member qpair->group begins it's life as NULL and then is assigned to 2644 * a pointer and never changes. So fortunately reading this and checking for 2645 * non-NULL is thread safe in the x86_64 memory model. */ 2646 group = qpair->group; 2647 2648 if (group == NULL) { 2649 /* The qpair hasn't been assigned to a group yet, so we can't 2650 * process a disconnect. Send a message to ourself and try again. */ 2651 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2652 return; 2653 } 2654 2655 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2656 } 2657 2658 static inline void 2659 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2660 { 2661 if (__sync_bool_compare_and_swap(&rqpair->disconnect_started, false, true)) { 2662 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2663 } 2664 } 2665 2666 static void nvmf_rdma_destroy_drained_qpair(void *ctx) 2667 { 2668 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2669 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2670 struct spdk_nvmf_rdma_transport, transport); 2671 2672 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2673 if (rqpair->current_send_depth != 0) { 2674 return; 2675 } 2676 2677 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2678 return; 2679 } 2680 2681 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) { 2682 return; 2683 } 2684 2685 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2686 spdk_nvmf_rdma_qpair_destroy(rqpair); 2687 } 2688 2689 2690 static int 2691 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2692 { 2693 struct spdk_nvmf_qpair *qpair; 2694 struct spdk_nvmf_rdma_qpair *rqpair; 2695 2696 if (evt->id == NULL) { 2697 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2698 return -1; 2699 } 2700 2701 qpair = evt->id->context; 2702 if (qpair == NULL) { 2703 SPDK_ERRLOG("disconnect request: no active connection\n"); 2704 return -1; 2705 } 2706 2707 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2708 2709 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2710 2711 spdk_nvmf_rdma_update_ibv_state(rqpair); 2712 2713 spdk_nvmf_rdma_start_disconnect(rqpair); 2714 2715 return 0; 2716 } 2717 2718 #ifdef DEBUG 2719 static const char *CM_EVENT_STR[] = { 2720 "RDMA_CM_EVENT_ADDR_RESOLVED", 2721 "RDMA_CM_EVENT_ADDR_ERROR", 2722 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2723 "RDMA_CM_EVENT_ROUTE_ERROR", 2724 "RDMA_CM_EVENT_CONNECT_REQUEST", 2725 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2726 "RDMA_CM_EVENT_CONNECT_ERROR", 2727 "RDMA_CM_EVENT_UNREACHABLE", 2728 "RDMA_CM_EVENT_REJECTED", 2729 "RDMA_CM_EVENT_ESTABLISHED", 2730 "RDMA_CM_EVENT_DISCONNECTED", 2731 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2732 "RDMA_CM_EVENT_MULTICAST_JOIN", 2733 "RDMA_CM_EVENT_MULTICAST_ERROR", 2734 "RDMA_CM_EVENT_ADDR_CHANGE", 2735 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2736 }; 2737 #endif /* DEBUG */ 2738 2739 static void 2740 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2741 { 2742 struct spdk_nvmf_rdma_transport *rtransport; 2743 struct rdma_cm_event *event; 2744 int rc; 2745 2746 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2747 2748 if (rtransport->event_channel == NULL) { 2749 return; 2750 } 2751 2752 while (1) { 2753 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2754 if (rc == 0) { 2755 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2756 2757 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2758 2759 switch (event->event) { 2760 case RDMA_CM_EVENT_ADDR_RESOLVED: 2761 case RDMA_CM_EVENT_ADDR_ERROR: 2762 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2763 case RDMA_CM_EVENT_ROUTE_ERROR: 2764 /* No action required. The target never attempts to resolve routes. */ 2765 break; 2766 case RDMA_CM_EVENT_CONNECT_REQUEST: 2767 rc = nvmf_rdma_connect(transport, event, cb_fn); 2768 if (rc < 0) { 2769 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2770 break; 2771 } 2772 break; 2773 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2774 /* The target never initiates a new connection. So this will not occur. */ 2775 break; 2776 case RDMA_CM_EVENT_CONNECT_ERROR: 2777 /* Can this happen? The docs say it can, but not sure what causes it. */ 2778 break; 2779 case RDMA_CM_EVENT_UNREACHABLE: 2780 case RDMA_CM_EVENT_REJECTED: 2781 /* These only occur on the client side. */ 2782 break; 2783 case RDMA_CM_EVENT_ESTABLISHED: 2784 /* TODO: Should we be waiting for this event anywhere? */ 2785 break; 2786 case RDMA_CM_EVENT_DISCONNECTED: 2787 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2788 rc = nvmf_rdma_disconnect(event); 2789 if (rc < 0) { 2790 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2791 break; 2792 } 2793 break; 2794 case RDMA_CM_EVENT_MULTICAST_JOIN: 2795 case RDMA_CM_EVENT_MULTICAST_ERROR: 2796 /* Multicast is not used */ 2797 break; 2798 case RDMA_CM_EVENT_ADDR_CHANGE: 2799 /* Not utilizing this event */ 2800 break; 2801 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2802 /* For now, do nothing. The target never re-uses queue pairs. */ 2803 break; 2804 default: 2805 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2806 break; 2807 } 2808 2809 rdma_ack_cm_event(event); 2810 } else { 2811 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2812 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2813 } 2814 break; 2815 } 2816 } 2817 } 2818 2819 static void 2820 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2821 { 2822 int rc; 2823 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 2824 struct ibv_async_event event; 2825 enum ibv_qp_state state; 2826 2827 rc = ibv_get_async_event(device->context, &event); 2828 2829 if (rc) { 2830 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2831 errno, spdk_strerror(errno)); 2832 return; 2833 } 2834 2835 switch (event.event_type) { 2836 case IBV_EVENT_QP_FATAL: 2837 rqpair = event.element.qp->qp_context; 2838 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 2839 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2840 (uintptr_t)rqpair->cm_id, event.event_type); 2841 spdk_nvmf_rdma_update_ibv_state(rqpair); 2842 spdk_nvmf_rdma_start_disconnect(rqpair); 2843 break; 2844 case IBV_EVENT_QP_LAST_WQE_REACHED: 2845 /* This event only occurs for shared receive queues. */ 2846 rqpair = event.element.qp->qp_context; 2847 rqpair->last_wqe_reached = true; 2848 2849 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last WQE reached event received for rqpair %p\n", rqpair); 2850 /* This must be handled on the polling thread if it exists. Otherwise the timeout will catch it. */ 2851 if (rqpair->qpair.group) { 2852 spdk_thread_send_msg(rqpair->qpair.group->thread, nvmf_rdma_destroy_drained_qpair, rqpair); 2853 } else { 2854 SPDK_ERRLOG("Unable to destroy the qpair %p since it does not have a poll group.\n", rqpair); 2855 } 2856 2857 break; 2858 case IBV_EVENT_SQ_DRAINED: 2859 /* This event occurs frequently in both error and non-error states. 2860 * Check if the qpair is in an error state before sending a message. 2861 * Note that we're not on the correct thread to access the qpair, but 2862 * the operations that the below calls make all happen to be thread 2863 * safe. */ 2864 rqpair = event.element.qp->qp_context; 2865 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last sq drained event received for rqpair %p\n", rqpair); 2866 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2867 (uintptr_t)rqpair->cm_id, event.event_type); 2868 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2869 if (state == IBV_QPS_ERR) { 2870 spdk_nvmf_rdma_start_disconnect(rqpair); 2871 } 2872 break; 2873 case IBV_EVENT_QP_REQ_ERR: 2874 case IBV_EVENT_QP_ACCESS_ERR: 2875 case IBV_EVENT_COMM_EST: 2876 case IBV_EVENT_PATH_MIG: 2877 case IBV_EVENT_PATH_MIG_ERR: 2878 SPDK_NOTICELOG("Async event: %s\n", 2879 ibv_event_type_str(event.event_type)); 2880 rqpair = event.element.qp->qp_context; 2881 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2882 (uintptr_t)rqpair->cm_id, event.event_type); 2883 spdk_nvmf_rdma_update_ibv_state(rqpair); 2884 break; 2885 case IBV_EVENT_CQ_ERR: 2886 case IBV_EVENT_DEVICE_FATAL: 2887 case IBV_EVENT_PORT_ACTIVE: 2888 case IBV_EVENT_PORT_ERR: 2889 case IBV_EVENT_LID_CHANGE: 2890 case IBV_EVENT_PKEY_CHANGE: 2891 case IBV_EVENT_SM_CHANGE: 2892 case IBV_EVENT_SRQ_ERR: 2893 case IBV_EVENT_SRQ_LIMIT_REACHED: 2894 case IBV_EVENT_CLIENT_REREGISTER: 2895 case IBV_EVENT_GID_CHANGE: 2896 default: 2897 SPDK_NOTICELOG("Async event: %s\n", 2898 ibv_event_type_str(event.event_type)); 2899 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2900 break; 2901 } 2902 ibv_ack_async_event(&event); 2903 } 2904 2905 static void 2906 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2907 { 2908 int nfds, i = 0; 2909 struct spdk_nvmf_rdma_transport *rtransport; 2910 struct spdk_nvmf_rdma_device *device, *tmp; 2911 2912 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2913 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2914 2915 if (nfds <= 0) { 2916 return; 2917 } 2918 2919 /* The first poll descriptor is RDMA CM event */ 2920 if (rtransport->poll_fds[i++].revents & POLLIN) { 2921 spdk_nvmf_process_cm_event(transport, cb_fn); 2922 nfds--; 2923 } 2924 2925 if (nfds == 0) { 2926 return; 2927 } 2928 2929 /* Second and subsequent poll descriptors are IB async events */ 2930 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2931 if (rtransport->poll_fds[i++].revents & POLLIN) { 2932 spdk_nvmf_process_ib_event(device); 2933 nfds--; 2934 } 2935 } 2936 /* check all flagged fd's have been served */ 2937 assert(nfds == 0); 2938 } 2939 2940 static void 2941 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2942 struct spdk_nvme_transport_id *trid, 2943 struct spdk_nvmf_discovery_log_page_entry *entry) 2944 { 2945 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2946 entry->adrfam = trid->adrfam; 2947 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2948 2949 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2950 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2951 2952 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2953 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2954 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2955 } 2956 2957 static void 2958 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 2959 2960 static struct spdk_nvmf_transport_poll_group * 2961 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2962 { 2963 struct spdk_nvmf_rdma_transport *rtransport; 2964 struct spdk_nvmf_rdma_poll_group *rgroup; 2965 struct spdk_nvmf_rdma_poller *poller; 2966 struct spdk_nvmf_rdma_device *device; 2967 struct ibv_srq_init_attr srq_init_attr; 2968 struct spdk_nvmf_rdma_resource_opts opts; 2969 int num_cqe; 2970 2971 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2972 2973 rgroup = calloc(1, sizeof(*rgroup)); 2974 if (!rgroup) { 2975 return NULL; 2976 } 2977 2978 TAILQ_INIT(&rgroup->pollers); 2979 STAILQ_INIT(&rgroup->pending_data_buf_queue); 2980 2981 pthread_mutex_lock(&rtransport->lock); 2982 TAILQ_FOREACH(device, &rtransport->devices, link) { 2983 poller = calloc(1, sizeof(*poller)); 2984 if (!poller) { 2985 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2986 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2987 pthread_mutex_unlock(&rtransport->lock); 2988 return NULL; 2989 } 2990 2991 poller->device = device; 2992 poller->group = rgroup; 2993 2994 TAILQ_INIT(&poller->qpairs); 2995 STAILQ_INIT(&poller->qpairs_pending_send); 2996 STAILQ_INIT(&poller->qpairs_pending_recv); 2997 2998 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2999 if (transport->opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3000 poller->max_srq_depth = transport->opts.max_srq_depth; 3001 3002 device->num_srq++; 3003 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 3004 srq_init_attr.attr.max_wr = poller->max_srq_depth; 3005 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3006 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 3007 if (!poller->srq) { 3008 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3009 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3010 pthread_mutex_unlock(&rtransport->lock); 3011 return NULL; 3012 } 3013 3014 opts.qp = poller->srq; 3015 opts.pd = device->pd; 3016 opts.qpair = NULL; 3017 opts.shared = true; 3018 opts.max_queue_depth = poller->max_srq_depth; 3019 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3020 3021 poller->resources = nvmf_rdma_resources_create(&opts); 3022 if (!poller->resources) { 3023 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3024 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3025 pthread_mutex_unlock(&rtransport->lock); 3026 } 3027 } 3028 3029 /* 3030 * When using an srq, we can limit the completion queue at startup. 3031 * The following formula represents the calculation: 3032 * num_cqe = num_recv + num_data_wr + num_send_wr. 3033 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3034 */ 3035 if (poller->srq) { 3036 num_cqe = poller->max_srq_depth * 3; 3037 } else { 3038 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3039 } 3040 3041 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3042 if (!poller->cq) { 3043 SPDK_ERRLOG("Unable to create completion queue\n"); 3044 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3045 pthread_mutex_unlock(&rtransport->lock); 3046 return NULL; 3047 } 3048 poller->num_cqe = num_cqe; 3049 } 3050 3051 pthread_mutex_unlock(&rtransport->lock); 3052 return &rgroup->group; 3053 } 3054 3055 static void 3056 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3057 { 3058 struct spdk_nvmf_rdma_poll_group *rgroup; 3059 struct spdk_nvmf_rdma_poller *poller, *tmp; 3060 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3061 3062 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3063 3064 if (!rgroup) { 3065 return; 3066 } 3067 3068 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3069 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3070 3071 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3072 spdk_nvmf_rdma_qpair_destroy(qpair); 3073 } 3074 3075 if (poller->srq) { 3076 nvmf_rdma_resources_destroy(poller->resources); 3077 ibv_destroy_srq(poller->srq); 3078 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3079 } 3080 3081 if (poller->cq) { 3082 ibv_destroy_cq(poller->cq); 3083 } 3084 3085 free(poller); 3086 } 3087 3088 if (!STAILQ_EMPTY(&rgroup->pending_data_buf_queue)) { 3089 SPDK_ERRLOG("Pending I/O list wasn't empty on poll group destruction\n"); 3090 } 3091 3092 free(rgroup); 3093 } 3094 3095 static void 3096 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3097 { 3098 if (rqpair->cm_id != NULL) { 3099 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3100 } 3101 spdk_nvmf_rdma_qpair_destroy(rqpair); 3102 } 3103 3104 static int 3105 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3106 struct spdk_nvmf_qpair *qpair) 3107 { 3108 struct spdk_nvmf_rdma_poll_group *rgroup; 3109 struct spdk_nvmf_rdma_qpair *rqpair; 3110 struct spdk_nvmf_rdma_device *device; 3111 struct spdk_nvmf_rdma_poller *poller; 3112 int rc; 3113 3114 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3115 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3116 3117 device = rqpair->port->device; 3118 3119 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3120 if (poller->device == device) { 3121 break; 3122 } 3123 } 3124 3125 if (!poller) { 3126 SPDK_ERRLOG("No poller found for device.\n"); 3127 return -1; 3128 } 3129 3130 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3131 rqpair->poller = poller; 3132 rqpair->srq = rqpair->poller->srq; 3133 3134 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 3135 if (rc < 0) { 3136 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3137 return -1; 3138 } 3139 3140 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3141 if (rc) { 3142 /* Try to reject, but we probably can't */ 3143 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3144 return -1; 3145 } 3146 3147 spdk_nvmf_rdma_update_ibv_state(rqpair); 3148 3149 return 0; 3150 } 3151 3152 static int 3153 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3154 { 3155 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3156 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3157 struct spdk_nvmf_rdma_transport, transport); 3158 3159 nvmf_rdma_request_free(rdma_req, rtransport); 3160 return 0; 3161 } 3162 3163 static int 3164 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3165 { 3166 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3167 struct spdk_nvmf_rdma_transport, transport); 3168 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3169 struct spdk_nvmf_rdma_request, req); 3170 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3171 struct spdk_nvmf_rdma_qpair, qpair); 3172 3173 if (rqpair->ibv_state != IBV_QPS_ERR) { 3174 /* The connection is alive, so process the request as normal */ 3175 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3176 } else { 3177 /* The connection is dead. Move the request directly to the completed state. */ 3178 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3179 } 3180 3181 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3182 3183 return 0; 3184 } 3185 3186 static int 3187 spdk_nvmf_rdma_destroy_defunct_qpair(void *ctx) 3188 { 3189 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 3190 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3191 struct spdk_nvmf_rdma_transport, transport); 3192 3193 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3194 spdk_nvmf_rdma_qpair_destroy(rqpair); 3195 3196 return 0; 3197 } 3198 3199 static void 3200 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3201 { 3202 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3203 3204 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 3205 return; 3206 } 3207 3208 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 3209 3210 /* This happens only when the qpair is disconnected before 3211 * it is added to the poll group. Since there is no poll group, 3212 * the RDMA qp has not been initialized yet and the RDMA CM 3213 * event has not yet been acknowledged, so we need to reject it. 3214 */ 3215 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3216 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3217 return; 3218 } 3219 3220 if (rqpair->ibv_state != IBV_QPS_ERR) { 3221 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 3222 } 3223 3224 rqpair->destruct_poller = spdk_poller_register(spdk_nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 3225 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 3226 } 3227 3228 static struct spdk_nvmf_rdma_qpair * 3229 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3230 { 3231 struct spdk_nvmf_rdma_qpair *rqpair; 3232 /* @todo: improve QP search */ 3233 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3234 if (wc->qp_num == rqpair->cm_id->qp->qp_num) { 3235 return rqpair; 3236 } 3237 } 3238 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3239 return NULL; 3240 } 3241 3242 #ifdef DEBUG 3243 static int 3244 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3245 { 3246 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3247 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3248 } 3249 #endif 3250 3251 static void 3252 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3253 int rc) 3254 { 3255 struct spdk_nvmf_rdma_recv *rdma_recv; 3256 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3257 3258 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3259 while (bad_recv_wr != NULL) { 3260 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3261 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3262 3263 rdma_recv->qpair->current_recv_depth++; 3264 bad_recv_wr = bad_recv_wr->next; 3265 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3266 spdk_nvmf_rdma_start_disconnect(rdma_recv->qpair); 3267 } 3268 } 3269 3270 static void 3271 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3272 { 3273 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3274 while (bad_recv_wr != NULL) { 3275 bad_recv_wr = bad_recv_wr->next; 3276 rqpair->current_recv_depth++; 3277 } 3278 spdk_nvmf_rdma_start_disconnect(rqpair); 3279 } 3280 3281 static void 3282 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3283 struct spdk_nvmf_rdma_poller *rpoller) 3284 { 3285 struct spdk_nvmf_rdma_qpair *rqpair; 3286 struct ibv_recv_wr *bad_recv_wr; 3287 int rc; 3288 3289 if (rpoller->srq) { 3290 if (rpoller->resources->recvs_to_post.first != NULL) { 3291 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3292 if (rc) { 3293 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3294 } 3295 rpoller->resources->recvs_to_post.first = NULL; 3296 rpoller->resources->recvs_to_post.last = NULL; 3297 } 3298 } else { 3299 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3300 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3301 assert(rqpair->resources->recvs_to_post.first != NULL); 3302 rc = ibv_post_recv(rqpair->cm_id->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3303 if (rc) { 3304 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3305 } 3306 rqpair->resources->recvs_to_post.first = NULL; 3307 rqpair->resources->recvs_to_post.last = NULL; 3308 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3309 } 3310 } 3311 } 3312 3313 static void 3314 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3315 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3316 { 3317 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3318 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3319 3320 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3321 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3322 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3323 assert(rqpair->current_send_depth > 0); 3324 rqpair->current_send_depth--; 3325 switch (bad_rdma_wr->type) { 3326 case RDMA_WR_TYPE_DATA: 3327 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3328 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3329 assert(rqpair->current_read_depth > 0); 3330 rqpair->current_read_depth--; 3331 } 3332 break; 3333 case RDMA_WR_TYPE_SEND: 3334 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3335 break; 3336 default: 3337 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3338 prev_rdma_req = cur_rdma_req; 3339 continue; 3340 } 3341 3342 if (prev_rdma_req == cur_rdma_req) { 3343 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3344 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3345 continue; 3346 } 3347 3348 switch (cur_rdma_req->state) { 3349 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3350 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3351 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3352 break; 3353 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3354 case RDMA_REQUEST_STATE_COMPLETING: 3355 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3356 break; 3357 default: 3358 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3359 cur_rdma_req->state, rqpair); 3360 continue; 3361 } 3362 3363 spdk_nvmf_rdma_request_process(rtransport, cur_rdma_req); 3364 prev_rdma_req = cur_rdma_req; 3365 } 3366 3367 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3368 /* Disconnect the connection. */ 3369 spdk_nvmf_rdma_start_disconnect(rqpair); 3370 } 3371 3372 } 3373 3374 static void 3375 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3376 struct spdk_nvmf_rdma_poller *rpoller) 3377 { 3378 struct spdk_nvmf_rdma_qpair *rqpair; 3379 struct ibv_send_wr *bad_wr = NULL; 3380 int rc; 3381 3382 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3383 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3384 assert(rqpair->sends_to_post.first != NULL); 3385 rc = ibv_post_send(rqpair->cm_id->qp, rqpair->sends_to_post.first, &bad_wr); 3386 3387 /* bad wr always points to the first wr that failed. */ 3388 if (rc) { 3389 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3390 } 3391 rqpair->sends_to_post.first = NULL; 3392 rqpair->sends_to_post.last = NULL; 3393 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3394 } 3395 } 3396 3397 static int 3398 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3399 struct spdk_nvmf_rdma_poller *rpoller) 3400 { 3401 struct ibv_wc wc[32]; 3402 struct spdk_nvmf_rdma_wr *rdma_wr; 3403 struct spdk_nvmf_rdma_request *rdma_req; 3404 struct spdk_nvmf_rdma_recv *rdma_recv; 3405 struct spdk_nvmf_rdma_qpair *rqpair; 3406 int reaped, i; 3407 int count = 0; 3408 bool error = false; 3409 3410 /* Poll for completing operations. */ 3411 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3412 if (reaped < 0) { 3413 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3414 errno, spdk_strerror(errno)); 3415 return -1; 3416 } 3417 3418 for (i = 0; i < reaped; i++) { 3419 3420 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3421 3422 switch (rdma_wr->type) { 3423 case RDMA_WR_TYPE_SEND: 3424 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3425 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3426 3427 if (!wc[i].status) { 3428 count++; 3429 assert(wc[i].opcode == IBV_WC_SEND); 3430 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 3431 } else { 3432 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3433 } 3434 3435 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3436 /* +1 for the response wr */ 3437 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3438 rdma_req->num_outstanding_data_wr = 0; 3439 3440 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3441 break; 3442 case RDMA_WR_TYPE_RECV: 3443 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3444 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3445 if (rpoller->srq != NULL) { 3446 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3447 } 3448 rqpair = rdma_recv->qpair; 3449 3450 assert(rqpair != NULL); 3451 if (!wc[i].status) { 3452 assert(wc[i].opcode == IBV_WC_RECV); 3453 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3454 spdk_nvmf_rdma_start_disconnect(rqpair); 3455 break; 3456 } 3457 } 3458 3459 rdma_recv->wr.next = NULL; 3460 rqpair->current_recv_depth++; 3461 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3462 break; 3463 case RDMA_WR_TYPE_DATA: 3464 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3465 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3466 3467 assert(rdma_req->num_outstanding_data_wr > 0); 3468 3469 rqpair->current_send_depth--; 3470 rdma_req->num_outstanding_data_wr--; 3471 if (!wc[i].status) { 3472 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3473 rqpair->current_read_depth--; 3474 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3475 if (rdma_req->num_outstanding_data_wr == 0) { 3476 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3477 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3478 } 3479 } else { 3480 /* If the data transfer fails still force the queue into the error state, 3481 * if we were performing an RDMA_READ, we need to force the request into a 3482 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3483 * case, we should wait for the SEND to complete. */ 3484 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3485 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3486 rqpair->current_read_depth--; 3487 if (rdma_req->num_outstanding_data_wr == 0) { 3488 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3489 } 3490 } 3491 } 3492 break; 3493 default: 3494 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3495 continue; 3496 } 3497 3498 /* Handle error conditions */ 3499 if (wc[i].status) { 3500 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 3501 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 3502 3503 error = true; 3504 3505 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3506 /* Disconnect the connection. */ 3507 spdk_nvmf_rdma_start_disconnect(rqpair); 3508 } else { 3509 nvmf_rdma_destroy_drained_qpair(rqpair); 3510 } 3511 continue; 3512 } 3513 3514 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3515 3516 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3517 nvmf_rdma_destroy_drained_qpair(rqpair); 3518 } 3519 } 3520 3521 if (error == true) { 3522 return -1; 3523 } 3524 3525 /* submit outstanding work requests. */ 3526 _poller_submit_recvs(rtransport, rpoller); 3527 _poller_submit_sends(rtransport, rpoller); 3528 3529 return count; 3530 } 3531 3532 static int 3533 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3534 { 3535 struct spdk_nvmf_rdma_transport *rtransport; 3536 struct spdk_nvmf_rdma_poll_group *rgroup; 3537 struct spdk_nvmf_rdma_poller *rpoller; 3538 int count, rc; 3539 3540 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3541 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3542 3543 count = 0; 3544 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3545 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 3546 if (rc < 0) { 3547 return rc; 3548 } 3549 count += rc; 3550 } 3551 3552 return count; 3553 } 3554 3555 static int 3556 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3557 struct spdk_nvme_transport_id *trid, 3558 bool peer) 3559 { 3560 struct sockaddr *saddr; 3561 uint16_t port; 3562 3563 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 3564 3565 if (peer) { 3566 saddr = rdma_get_peer_addr(id); 3567 } else { 3568 saddr = rdma_get_local_addr(id); 3569 } 3570 switch (saddr->sa_family) { 3571 case AF_INET: { 3572 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3573 3574 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3575 inet_ntop(AF_INET, &saddr_in->sin_addr, 3576 trid->traddr, sizeof(trid->traddr)); 3577 if (peer) { 3578 port = ntohs(rdma_get_dst_port(id)); 3579 } else { 3580 port = ntohs(rdma_get_src_port(id)); 3581 } 3582 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3583 break; 3584 } 3585 case AF_INET6: { 3586 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3587 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3588 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3589 trid->traddr, sizeof(trid->traddr)); 3590 if (peer) { 3591 port = ntohs(rdma_get_dst_port(id)); 3592 } else { 3593 port = ntohs(rdma_get_src_port(id)); 3594 } 3595 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3596 break; 3597 } 3598 default: 3599 return -1; 3600 3601 } 3602 3603 return 0; 3604 } 3605 3606 static int 3607 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3608 struct spdk_nvme_transport_id *trid) 3609 { 3610 struct spdk_nvmf_rdma_qpair *rqpair; 3611 3612 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3613 3614 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 3615 } 3616 3617 static int 3618 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3619 struct spdk_nvme_transport_id *trid) 3620 { 3621 struct spdk_nvmf_rdma_qpair *rqpair; 3622 3623 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3624 3625 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 3626 } 3627 3628 static int 3629 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3630 struct spdk_nvme_transport_id *trid) 3631 { 3632 struct spdk_nvmf_rdma_qpair *rqpair; 3633 3634 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3635 3636 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 3637 } 3638 3639 void 3640 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3641 { 3642 g_nvmf_hooks = *hooks; 3643 } 3644 3645 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 3646 .type = SPDK_NVME_TRANSPORT_RDMA, 3647 .opts_init = spdk_nvmf_rdma_opts_init, 3648 .create = spdk_nvmf_rdma_create, 3649 .destroy = spdk_nvmf_rdma_destroy, 3650 3651 .listen = spdk_nvmf_rdma_listen, 3652 .stop_listen = spdk_nvmf_rdma_stop_listen, 3653 .accept = spdk_nvmf_rdma_accept, 3654 3655 .listener_discover = spdk_nvmf_rdma_discover, 3656 3657 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 3658 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 3659 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 3660 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 3661 3662 .req_free = spdk_nvmf_rdma_request_free, 3663 .req_complete = spdk_nvmf_rdma_request_complete, 3664 3665 .qpair_fini = spdk_nvmf_rdma_close_qpair, 3666 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 3667 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 3668 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 3669 3670 }; 3671 3672 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 3673