1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2018 Mellanox Technologies LTD. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of Intel Corporation nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include "spdk/stdinc.h" 35 36 #include <infiniband/verbs.h> 37 #include <rdma/rdma_cma.h> 38 #include <rdma/rdma_verbs.h> 39 40 #include "nvmf_internal.h" 41 #include "transport.h" 42 43 #include "spdk/config.h" 44 #include "spdk/assert.h" 45 #include "spdk/thread.h" 46 #include "spdk/nvmf.h" 47 #include "spdk/nvmf_spec.h" 48 #include "spdk/string.h" 49 #include "spdk/trace.h" 50 #include "spdk/util.h" 51 52 #include "spdk_internal/log.h" 53 54 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 55 56 /* 57 RDMA Connection Resource Defaults 58 */ 59 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 60 #define NVMF_DEFAULT_RSP_SGE 1 61 #define NVMF_DEFAULT_RX_SGE 2 62 63 /* The RDMA completion queue size */ 64 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 65 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 66 67 /* Timeout for destroying defunct rqpairs */ 68 #define NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US 4000000 69 70 /* 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 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 455 }; 456 457 struct spdk_nvmf_rdma_port { 458 struct spdk_nvme_transport_id trid; 459 struct rdma_cm_id *id; 460 struct spdk_nvmf_rdma_device *device; 461 uint32_t ref; 462 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 463 }; 464 465 struct spdk_nvmf_rdma_transport { 466 struct spdk_nvmf_transport transport; 467 468 struct rdma_event_channel *event_channel; 469 470 struct spdk_mempool *data_wr_pool; 471 472 pthread_mutex_t lock; 473 474 /* fields used to poll RDMA/IB events */ 475 nfds_t npoll_fds; 476 struct pollfd *poll_fds; 477 478 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 479 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 480 }; 481 482 static inline int 483 spdk_nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 484 { 485 switch (state) { 486 case IBV_QPS_RESET: 487 case IBV_QPS_INIT: 488 case IBV_QPS_RTR: 489 case IBV_QPS_RTS: 490 case IBV_QPS_SQD: 491 case IBV_QPS_SQE: 492 case IBV_QPS_ERR: 493 return 0; 494 default: 495 return -1; 496 } 497 } 498 499 static enum ibv_qp_state 500 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 501 enum ibv_qp_state old_state, new_state; 502 struct ibv_qp_attr qp_attr; 503 struct ibv_qp_init_attr init_attr; 504 int rc; 505 506 old_state = rqpair->ibv_state; 507 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 508 g_spdk_nvmf_ibv_query_mask, &init_attr); 509 510 if (rc) 511 { 512 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 513 return IBV_QPS_ERR + 1; 514 } 515 516 new_state = qp_attr.qp_state; 517 rqpair->ibv_state = new_state; 518 qp_attr.ah_attr.port_num = qp_attr.port_num; 519 520 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 521 if (rc) 522 { 523 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 524 /* 525 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 526 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 527 */ 528 return IBV_QPS_ERR + 1; 529 } 530 531 if (old_state != new_state) 532 { 533 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 534 (uintptr_t)rqpair->cm_id, new_state); 535 } 536 return new_state; 537 } 538 539 static const char *str_ibv_qp_state[] = { 540 "IBV_QPS_RESET", 541 "IBV_QPS_INIT", 542 "IBV_QPS_RTR", 543 "IBV_QPS_RTS", 544 "IBV_QPS_SQD", 545 "IBV_QPS_SQE", 546 "IBV_QPS_ERR", 547 "IBV_QPS_UNKNOWN" 548 }; 549 550 static int 551 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 552 enum ibv_qp_state new_state) 553 { 554 struct ibv_qp_attr qp_attr; 555 struct ibv_qp_init_attr init_attr; 556 int rc; 557 enum ibv_qp_state state; 558 static int attr_mask_rc[] = { 559 [IBV_QPS_RESET] = IBV_QP_STATE, 560 [IBV_QPS_INIT] = (IBV_QP_STATE | 561 IBV_QP_PKEY_INDEX | 562 IBV_QP_PORT | 563 IBV_QP_ACCESS_FLAGS), 564 [IBV_QPS_RTR] = (IBV_QP_STATE | 565 IBV_QP_AV | 566 IBV_QP_PATH_MTU | 567 IBV_QP_DEST_QPN | 568 IBV_QP_RQ_PSN | 569 IBV_QP_MAX_DEST_RD_ATOMIC | 570 IBV_QP_MIN_RNR_TIMER), 571 [IBV_QPS_RTS] = (IBV_QP_STATE | 572 IBV_QP_SQ_PSN | 573 IBV_QP_TIMEOUT | 574 IBV_QP_RETRY_CNT | 575 IBV_QP_RNR_RETRY | 576 IBV_QP_MAX_QP_RD_ATOMIC), 577 [IBV_QPS_SQD] = IBV_QP_STATE, 578 [IBV_QPS_SQE] = IBV_QP_STATE, 579 [IBV_QPS_ERR] = IBV_QP_STATE, 580 }; 581 582 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 583 if (rc) { 584 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 585 rqpair->qpair.qid, new_state); 586 return rc; 587 } 588 589 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 590 g_spdk_nvmf_ibv_query_mask, &init_attr); 591 592 if (rc) { 593 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 594 assert(false); 595 } 596 597 qp_attr.cur_qp_state = rqpair->ibv_state; 598 qp_attr.qp_state = new_state; 599 600 rc = ibv_modify_qp(rqpair->cm_id->qp, &qp_attr, 601 attr_mask_rc[new_state]); 602 603 if (rc) { 604 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 605 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 606 return rc; 607 } 608 609 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 610 611 if (state != new_state) { 612 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 613 rqpair->qpair.qid, str_ibv_qp_state[new_state], 614 str_ibv_qp_state[state]); 615 return -1; 616 } 617 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 618 str_ibv_qp_state[state]); 619 return 0; 620 } 621 622 static void 623 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 624 struct spdk_nvmf_rdma_transport *rtransport) 625 { 626 struct spdk_nvmf_rdma_request_data *current_data_wr = NULL, *next_data_wr = NULL; 627 struct ibv_send_wr *send_wr; 628 int i; 629 630 rdma_req->num_outstanding_data_wr = 0; 631 current_data_wr = &rdma_req->data; 632 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 633 current_data_wr->wr.sg_list[i].addr = 0; 634 current_data_wr->wr.sg_list[i].length = 0; 635 current_data_wr->wr.sg_list[i].lkey = 0; 636 } 637 current_data_wr->wr.num_sge = 0; 638 639 send_wr = current_data_wr->wr.next; 640 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr) { 641 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 642 } 643 while (next_data_wr) { 644 current_data_wr = next_data_wr; 645 send_wr = current_data_wr->wr.next; 646 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr && 647 send_wr->wr_id == current_data_wr->wr.wr_id) { 648 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 649 } else { 650 next_data_wr = NULL; 651 } 652 653 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 654 current_data_wr->wr.sg_list[i].addr = 0; 655 current_data_wr->wr.sg_list[i].length = 0; 656 current_data_wr->wr.sg_list[i].lkey = 0; 657 } 658 current_data_wr->wr.num_sge = 0; 659 spdk_mempool_put(rtransport->data_wr_pool, current_data_wr); 660 } 661 } 662 663 static void 664 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 665 { 666 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->data_from_pool); 667 if (req->req.cmd) { 668 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 669 } 670 if (req->recv) { 671 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 672 } 673 } 674 675 static void 676 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 677 { 678 int i; 679 680 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 681 for (i = 0; i < rqpair->max_queue_depth; i++) { 682 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 683 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 684 } 685 } 686 } 687 688 static void 689 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 690 { 691 if (resources->cmds_mr) { 692 ibv_dereg_mr(resources->cmds_mr); 693 } 694 695 if (resources->cpls_mr) { 696 ibv_dereg_mr(resources->cpls_mr); 697 } 698 699 if (resources->bufs_mr) { 700 ibv_dereg_mr(resources->bufs_mr); 701 } 702 703 spdk_dma_free(resources->cmds); 704 spdk_dma_free(resources->cpls); 705 spdk_dma_free(resources->bufs); 706 free(resources->reqs); 707 free(resources->recvs); 708 free(resources); 709 } 710 711 712 static struct spdk_nvmf_rdma_resources * 713 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 714 { 715 struct spdk_nvmf_rdma_resources *resources; 716 struct spdk_nvmf_rdma_request *rdma_req; 717 struct spdk_nvmf_rdma_recv *rdma_recv; 718 struct ibv_qp *qp; 719 struct ibv_srq *srq; 720 uint32_t i; 721 int rc; 722 723 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 724 if (!resources) { 725 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 726 return NULL; 727 } 728 729 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 730 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 731 resources->cmds = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 732 0x1000, NULL); 733 resources->cpls = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 734 0x1000, NULL); 735 736 if (opts->in_capsule_data_size > 0) { 737 resources->bufs = spdk_dma_zmalloc(opts->max_queue_depth * 738 opts->in_capsule_data_size, 739 0x1000, NULL); 740 } 741 742 if (!resources->reqs || !resources->recvs || !resources->cmds || 743 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 744 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 745 goto cleanup; 746 } 747 748 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 749 opts->max_queue_depth * sizeof(*resources->cmds), 750 IBV_ACCESS_LOCAL_WRITE); 751 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 752 opts->max_queue_depth * sizeof(*resources->cpls), 753 0); 754 755 if (opts->in_capsule_data_size) { 756 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 757 opts->max_queue_depth * 758 opts->in_capsule_data_size, 759 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 760 } 761 762 if (!resources->cmds_mr || !resources->cpls_mr || 763 (opts->in_capsule_data_size && 764 !resources->bufs_mr)) { 765 goto cleanup; 766 } 767 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 768 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 769 resources->cmds_mr->lkey); 770 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 771 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 772 resources->cpls_mr->lkey); 773 if (resources->bufs && resources->bufs_mr) { 774 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 775 resources->bufs, opts->max_queue_depth * 776 opts->in_capsule_data_size, resources->bufs_mr->lkey); 777 } 778 779 /* Initialize queues */ 780 STAILQ_INIT(&resources->incoming_queue); 781 STAILQ_INIT(&resources->free_queue); 782 783 for (i = 0; i < opts->max_queue_depth; i++) { 784 struct ibv_recv_wr *bad_wr = NULL; 785 786 rdma_recv = &resources->recvs[i]; 787 rdma_recv->qpair = opts->qpair; 788 789 /* Set up memory to receive commands */ 790 if (resources->bufs) { 791 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 792 opts->in_capsule_data_size)); 793 } 794 795 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 796 797 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 798 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 799 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 800 rdma_recv->wr.num_sge = 1; 801 802 if (rdma_recv->buf && resources->bufs_mr) { 803 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 804 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 805 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 806 rdma_recv->wr.num_sge++; 807 } 808 809 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 810 rdma_recv->wr.sg_list = rdma_recv->sgl; 811 if (opts->shared) { 812 srq = (struct ibv_srq *)opts->qp; 813 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 814 } else { 815 qp = (struct ibv_qp *)opts->qp; 816 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 817 } 818 if (rc) { 819 goto cleanup; 820 } 821 } 822 823 for (i = 0; i < opts->max_queue_depth; i++) { 824 rdma_req = &resources->reqs[i]; 825 826 if (opts->qpair != NULL) { 827 rdma_req->req.qpair = &opts->qpair->qpair; 828 } else { 829 rdma_req->req.qpair = NULL; 830 } 831 rdma_req->req.cmd = NULL; 832 833 /* Set up memory to send responses */ 834 rdma_req->req.rsp = &resources->cpls[i]; 835 836 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 837 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 838 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 839 840 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 841 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 842 rdma_req->rsp.wr.next = NULL; 843 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 844 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 845 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 846 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 847 848 /* Set up memory for data buffers */ 849 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 850 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 851 rdma_req->data.wr.next = NULL; 852 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 853 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 854 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 855 856 /* Initialize request state to FREE */ 857 rdma_req->state = RDMA_REQUEST_STATE_FREE; 858 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 859 } 860 861 return resources; 862 863 cleanup: 864 nvmf_rdma_resources_destroy(resources); 865 return NULL; 866 } 867 868 static void 869 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 870 { 871 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 872 struct ibv_recv_wr *bad_recv_wr = NULL; 873 int rc; 874 875 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 876 877 spdk_poller_unregister(&rqpair->destruct_poller); 878 879 if (rqpair->qd != 0) { 880 if (rqpair->srq == NULL) { 881 nvmf_rdma_dump_qpair_contents(rqpair); 882 } 883 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 884 } 885 886 if (rqpair->poller) { 887 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 888 889 if (rqpair->srq != NULL && rqpair->resources != NULL) { 890 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 891 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 892 if (rqpair == rdma_recv->qpair) { 893 STAILQ_REMOVE_HEAD(&rqpair->resources->incoming_queue, link); 894 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 895 if (rc) { 896 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 897 } 898 } 899 } 900 } 901 } 902 903 if (rqpair->cm_id) { 904 if (rqpair->cm_id->qp != NULL) { 905 rdma_destroy_qp(rqpair->cm_id); 906 } 907 rdma_destroy_id(rqpair->cm_id); 908 909 if (rqpair->poller != NULL && rqpair->srq == NULL) { 910 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 911 } 912 } 913 914 if (rqpair->srq == NULL && rqpair->resources != NULL) { 915 nvmf_rdma_resources_destroy(rqpair->resources); 916 } 917 918 free(rqpair); 919 } 920 921 static int 922 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 923 { 924 struct spdk_nvmf_rdma_poller *rpoller; 925 int rc, num_cqe, required_num_wr; 926 927 /* Enlarge CQ size dynamically */ 928 rpoller = rqpair->poller; 929 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 930 num_cqe = rpoller->num_cqe; 931 if (num_cqe < required_num_wr) { 932 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 933 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 934 } 935 936 if (rpoller->num_cqe != num_cqe) { 937 if (required_num_wr > device->attr.max_cqe) { 938 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 939 required_num_wr, device->attr.max_cqe); 940 return -1; 941 } 942 943 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 944 rc = ibv_resize_cq(rpoller->cq, num_cqe); 945 if (rc) { 946 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 947 return -1; 948 } 949 950 rpoller->num_cqe = num_cqe; 951 } 952 953 rpoller->required_num_wr = required_num_wr; 954 return 0; 955 } 956 957 static int 958 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 959 { 960 struct spdk_nvmf_rdma_qpair *rqpair; 961 int rc; 962 struct spdk_nvmf_rdma_transport *rtransport; 963 struct spdk_nvmf_transport *transport; 964 struct spdk_nvmf_rdma_resource_opts opts; 965 struct spdk_nvmf_rdma_device *device; 966 struct ibv_qp_init_attr ibv_init_attr; 967 968 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 969 device = rqpair->port->device; 970 971 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 972 ibv_init_attr.qp_context = rqpair; 973 ibv_init_attr.qp_type = IBV_QPT_RC; 974 ibv_init_attr.send_cq = rqpair->poller->cq; 975 ibv_init_attr.recv_cq = rqpair->poller->cq; 976 977 if (rqpair->srq) { 978 ibv_init_attr.srq = rqpair->srq; 979 } else { 980 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 981 1; /* RECV operations + dummy drain WR */ 982 } 983 984 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 985 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 986 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 987 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 988 989 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 990 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 991 goto error; 992 } 993 994 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 995 if (rc) { 996 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 997 goto error; 998 } 999 1000 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2 + 1), 1001 ibv_init_attr.cap.max_send_wr); 1002 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 1003 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 1004 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 1005 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 1006 1007 rqpair->sends_to_post.first = NULL; 1008 rqpair->sends_to_post.last = NULL; 1009 1010 if (rqpair->poller->srq == NULL) { 1011 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1012 transport = &rtransport->transport; 1013 1014 opts.qp = rqpair->cm_id->qp; 1015 opts.pd = rqpair->cm_id->pd; 1016 opts.qpair = rqpair; 1017 opts.shared = false; 1018 opts.max_queue_depth = rqpair->max_queue_depth; 1019 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1020 1021 rqpair->resources = nvmf_rdma_resources_create(&opts); 1022 1023 if (!rqpair->resources) { 1024 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1025 goto error; 1026 } 1027 } else { 1028 rqpair->resources = rqpair->poller->resources; 1029 } 1030 1031 rqpair->current_recv_depth = 0; 1032 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1033 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1034 1035 return 0; 1036 1037 error: 1038 rdma_destroy_id(rqpair->cm_id); 1039 rqpair->cm_id = NULL; 1040 spdk_nvmf_rdma_qpair_destroy(rqpair); 1041 return -1; 1042 } 1043 1044 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1045 /* This function accepts either a single wr or the first wr in a linked list. */ 1046 static void 1047 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1048 { 1049 struct ibv_recv_wr *last; 1050 1051 last = first; 1052 while (last->next != NULL) { 1053 last = last->next; 1054 } 1055 1056 if (rqpair->resources->recvs_to_post.first == NULL) { 1057 rqpair->resources->recvs_to_post.first = first; 1058 rqpair->resources->recvs_to_post.last = last; 1059 if (rqpair->srq == NULL) { 1060 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1061 } 1062 } else { 1063 rqpair->resources->recvs_to_post.last->next = first; 1064 rqpair->resources->recvs_to_post.last = last; 1065 } 1066 } 1067 1068 /* Append the given send wr structure to the qpair's outstanding sends list. */ 1069 /* This function accepts either a single wr or the first wr in a linked list. */ 1070 static void 1071 nvmf_rdma_qpair_queue_send_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *first) 1072 { 1073 struct ibv_send_wr *last; 1074 1075 last = first; 1076 while (last->next != NULL) { 1077 last = last->next; 1078 } 1079 1080 if (rqpair->sends_to_post.first == NULL) { 1081 rqpair->sends_to_post.first = first; 1082 rqpair->sends_to_post.last = last; 1083 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1084 } else { 1085 rqpair->sends_to_post.last->next = first; 1086 rqpair->sends_to_post.last = last; 1087 } 1088 } 1089 1090 static int 1091 request_transfer_in(struct spdk_nvmf_request *req) 1092 { 1093 struct spdk_nvmf_rdma_request *rdma_req; 1094 struct spdk_nvmf_qpair *qpair; 1095 struct spdk_nvmf_rdma_qpair *rqpair; 1096 1097 qpair = req->qpair; 1098 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1099 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1100 1101 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1102 assert(rdma_req != NULL); 1103 1104 nvmf_rdma_qpair_queue_send_wrs(rqpair, &rdma_req->data.wr); 1105 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1106 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1107 return 0; 1108 } 1109 1110 static int 1111 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1112 { 1113 int num_outstanding_data_wr = 0; 1114 struct spdk_nvmf_rdma_request *rdma_req; 1115 struct spdk_nvmf_qpair *qpair; 1116 struct spdk_nvmf_rdma_qpair *rqpair; 1117 struct spdk_nvme_cpl *rsp; 1118 struct ibv_send_wr *first = NULL; 1119 1120 *data_posted = 0; 1121 qpair = req->qpair; 1122 rsp = &req->rsp->nvme_cpl; 1123 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1124 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1125 1126 /* Advance our sq_head pointer */ 1127 if (qpair->sq_head == qpair->sq_head_max) { 1128 qpair->sq_head = 0; 1129 } else { 1130 qpair->sq_head++; 1131 } 1132 rsp->sqhd = qpair->sq_head; 1133 1134 /* queue the capsule for the recv buffer */ 1135 assert(rdma_req->recv != NULL); 1136 1137 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1138 1139 rdma_req->recv = NULL; 1140 assert(rqpair->current_recv_depth > 0); 1141 rqpair->current_recv_depth--; 1142 1143 /* Build the response which consists of optional 1144 * RDMA WRITEs to transfer data, plus an RDMA SEND 1145 * containing the response. 1146 */ 1147 first = &rdma_req->rsp.wr; 1148 1149 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 1150 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1151 first = &rdma_req->data.wr; 1152 *data_posted = 1; 1153 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1154 } 1155 nvmf_rdma_qpair_queue_send_wrs(rqpair, first); 1156 /* +1 for the rsp wr */ 1157 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1158 1159 return 0; 1160 } 1161 1162 static int 1163 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1164 { 1165 struct spdk_nvmf_rdma_accept_private_data accept_data; 1166 struct rdma_conn_param ctrlr_event_data = {}; 1167 int rc; 1168 1169 accept_data.recfmt = 0; 1170 accept_data.crqsize = rqpair->max_queue_depth; 1171 1172 ctrlr_event_data.private_data = &accept_data; 1173 ctrlr_event_data.private_data_len = sizeof(accept_data); 1174 if (id->ps == RDMA_PS_TCP) { 1175 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1176 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1177 } 1178 1179 /* Configure infinite retries for the initiator side qpair. 1180 * When using a shared receive queue on the target side, 1181 * we need to pass this value to the initiator to prevent the 1182 * initiator side NIC from completing SEND requests back to the 1183 * initiator with status rnr_retry_count_exceeded. */ 1184 if (rqpair->srq != NULL) { 1185 ctrlr_event_data.rnr_retry_count = 0x7; 1186 } 1187 1188 rc = rdma_accept(id, &ctrlr_event_data); 1189 if (rc) { 1190 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 1191 } else { 1192 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1193 } 1194 1195 return rc; 1196 } 1197 1198 static void 1199 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1200 { 1201 struct spdk_nvmf_rdma_reject_private_data rej_data; 1202 1203 rej_data.recfmt = 0; 1204 rej_data.sts = error; 1205 1206 rdma_reject(id, &rej_data, sizeof(rej_data)); 1207 } 1208 1209 static int 1210 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 1211 new_qpair_fn cb_fn) 1212 { 1213 struct spdk_nvmf_rdma_transport *rtransport; 1214 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1215 struct spdk_nvmf_rdma_port *port; 1216 struct rdma_conn_param *rdma_param = NULL; 1217 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1218 uint16_t max_queue_depth; 1219 uint16_t max_read_depth; 1220 1221 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1222 1223 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1224 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1225 1226 rdma_param = &event->param.conn; 1227 if (rdma_param->private_data == NULL || 1228 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1229 SPDK_ERRLOG("connect request: no private data provided\n"); 1230 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1231 return -1; 1232 } 1233 1234 private_data = rdma_param->private_data; 1235 if (private_data->recfmt != 0) { 1236 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1237 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1238 return -1; 1239 } 1240 1241 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1242 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1243 1244 port = event->listen_id->context; 1245 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1246 event->listen_id, event->listen_id->verbs, port); 1247 1248 /* Figure out the supported queue depth. This is a multi-step process 1249 * that takes into account hardware maximums, host provided values, 1250 * and our target's internal memory limits */ 1251 1252 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1253 1254 /* Start with the maximum queue depth allowed by the target */ 1255 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1256 max_read_depth = rtransport->transport.opts.max_queue_depth; 1257 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1258 rtransport->transport.opts.max_queue_depth); 1259 1260 /* Next check the local NIC's hardware limitations */ 1261 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1262 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1263 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1264 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1265 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1266 1267 /* Next check the remote NIC's hardware limitations */ 1268 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1269 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1270 rdma_param->initiator_depth, rdma_param->responder_resources); 1271 if (rdma_param->initiator_depth > 0) { 1272 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1273 } 1274 1275 /* Finally check for the host software requested values, which are 1276 * optional. */ 1277 if (rdma_param->private_data != NULL && 1278 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1279 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1280 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1281 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1282 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1283 } 1284 1285 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1286 max_queue_depth, max_read_depth); 1287 1288 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1289 if (rqpair == NULL) { 1290 SPDK_ERRLOG("Could not allocate new connection.\n"); 1291 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1292 return -1; 1293 } 1294 1295 rqpair->port = port; 1296 rqpair->max_queue_depth = max_queue_depth; 1297 rqpair->max_read_depth = max_read_depth; 1298 rqpair->cm_id = event->id; 1299 rqpair->listen_id = event->listen_id; 1300 rqpair->qpair.transport = transport; 1301 1302 event->id->context = &rqpair->qpair; 1303 1304 cb_fn(&rqpair->qpair); 1305 1306 return 0; 1307 } 1308 1309 static int 1310 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1311 enum spdk_mem_map_notify_action action, 1312 void *vaddr, size_t size) 1313 { 1314 struct ibv_pd *pd = cb_ctx; 1315 struct ibv_mr *mr; 1316 1317 switch (action) { 1318 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1319 if (!g_nvmf_hooks.get_rkey) { 1320 mr = ibv_reg_mr(pd, vaddr, size, 1321 IBV_ACCESS_LOCAL_WRITE | 1322 IBV_ACCESS_REMOTE_READ | 1323 IBV_ACCESS_REMOTE_WRITE); 1324 if (mr == NULL) { 1325 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1326 return -1; 1327 } else { 1328 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1329 } 1330 } else { 1331 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1332 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1333 } 1334 break; 1335 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1336 if (!g_nvmf_hooks.get_rkey) { 1337 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1338 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1339 if (mr) { 1340 ibv_dereg_mr(mr); 1341 } 1342 } 1343 break; 1344 } 1345 1346 return 0; 1347 } 1348 1349 static int 1350 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1351 { 1352 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1353 return addr_1 == addr_2; 1354 } 1355 1356 static void 1357 spdk_nvmf_rdma_request_free_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1358 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1359 uint32_t num_buffers) 1360 { 1361 uint32_t i; 1362 1363 for (i = 0; i < num_buffers; i++) { 1364 if (group->buf_cache_count < group->buf_cache_size) { 1365 STAILQ_INSERT_HEAD(&group->buf_cache, 1366 (struct spdk_nvmf_transport_pg_cache_buf *)rdma_req->buffers[i], link); 1367 group->buf_cache_count++; 1368 } else { 1369 spdk_mempool_put(transport->data_buf_pool, rdma_req->buffers[i]); 1370 } 1371 rdma_req->req.iov[i].iov_base = NULL; 1372 rdma_req->buffers[i] = NULL; 1373 rdma_req->req.iov[i].iov_len = 0; 1374 1375 } 1376 rdma_req->data_from_pool = false; 1377 } 1378 1379 static int 1380 nvmf_rdma_request_get_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1381 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1382 uint32_t num_buffers) 1383 { 1384 uint32_t i = 0; 1385 1386 while (i < num_buffers) { 1387 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1388 group->buf_cache_count--; 1389 rdma_req->buffers[i] = STAILQ_FIRST(&group->buf_cache); 1390 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1391 assert(rdma_req->buffers[i] != NULL); 1392 i++; 1393 } else { 1394 if (spdk_mempool_get_bulk(transport->data_buf_pool, &rdma_req->buffers[i], num_buffers - i)) { 1395 goto err_exit; 1396 } 1397 i += num_buffers - i; 1398 } 1399 } 1400 1401 return 0; 1402 1403 err_exit: 1404 spdk_nvmf_rdma_request_free_buffers(rdma_req, group, transport, i); 1405 return -ENOMEM; 1406 } 1407 1408 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1409 1410 static spdk_nvme_data_transfer_t 1411 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1412 { 1413 enum spdk_nvme_data_transfer xfer; 1414 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1415 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1416 1417 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1418 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1419 rdma_req->rsp.wr.imm_data = 0; 1420 #endif 1421 1422 /* Figure out data transfer direction */ 1423 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1424 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1425 } else { 1426 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1427 1428 /* Some admin commands are special cases */ 1429 if ((rdma_req->req.qpair->qid == 0) && 1430 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1431 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1432 switch (cmd->cdw10 & 0xff) { 1433 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1434 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1435 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1436 break; 1437 default: 1438 xfer = SPDK_NVME_DATA_NONE; 1439 } 1440 } 1441 } 1442 1443 if (xfer == SPDK_NVME_DATA_NONE) { 1444 return xfer; 1445 } 1446 1447 /* Even for commands that may transfer data, they could have specified 0 length. 1448 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1449 */ 1450 switch (sgl->generic.type) { 1451 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1452 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1453 case SPDK_NVME_SGL_TYPE_SEGMENT: 1454 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1455 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1456 if (sgl->unkeyed.length == 0) { 1457 xfer = SPDK_NVME_DATA_NONE; 1458 } 1459 break; 1460 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1461 if (sgl->keyed.length == 0) { 1462 xfer = SPDK_NVME_DATA_NONE; 1463 } 1464 break; 1465 } 1466 1467 return xfer; 1468 } 1469 1470 static int 1471 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1472 struct spdk_nvmf_rdma_request *rdma_req, 1473 uint32_t num_sgl_descriptors) 1474 { 1475 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1476 struct spdk_nvmf_rdma_request_data *current_data_wr; 1477 uint32_t i; 1478 1479 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1480 return -ENOMEM; 1481 } 1482 1483 current_data_wr = &rdma_req->data; 1484 1485 for (i = 0; i < num_sgl_descriptors; i++) { 1486 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1487 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1488 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1489 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1490 } else { 1491 assert(false); 1492 } 1493 work_requests[i]->wr.send_flags = IBV_SEND_SIGNALED; 1494 work_requests[i]->wr.sg_list = work_requests[i]->sgl; 1495 work_requests[i]->wr.wr_id = rdma_req->data.wr.wr_id; 1496 current_data_wr->wr.next = &work_requests[i]->wr; 1497 current_data_wr = work_requests[i]; 1498 } 1499 1500 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1501 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1502 current_data_wr->wr.next = &rdma_req->rsp.wr; 1503 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1504 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1505 current_data_wr->wr.next = NULL; 1506 } 1507 return 0; 1508 } 1509 1510 static int 1511 nvmf_rdma_fill_buffers(struct spdk_nvmf_rdma_transport *rtransport, 1512 struct spdk_nvmf_rdma_poll_group *rgroup, 1513 struct spdk_nvmf_rdma_device *device, 1514 struct spdk_nvmf_rdma_request *rdma_req, 1515 struct ibv_send_wr *wr, 1516 uint32_t length) 1517 { 1518 uint64_t translation_len; 1519 uint32_t remaining_length = length; 1520 uint32_t iovcnt; 1521 uint32_t i = 0; 1522 1523 1524 while (remaining_length) { 1525 iovcnt = rdma_req->req.iovcnt; 1526 rdma_req->req.iov[iovcnt].iov_base = (void *)((uintptr_t)(rdma_req->buffers[iovcnt] + 1527 NVMF_DATA_BUFFER_MASK) & 1528 ~NVMF_DATA_BUFFER_MASK); 1529 rdma_req->req.iov[iovcnt].iov_len = spdk_min(remaining_length, 1530 rtransport->transport.opts.io_unit_size); 1531 rdma_req->req.iovcnt++; 1532 wr->sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[iovcnt].iov_base); 1533 wr->sg_list[i].length = rdma_req->req.iov[iovcnt].iov_len; 1534 translation_len = rdma_req->req.iov[iovcnt].iov_len; 1535 1536 if (!g_nvmf_hooks.get_rkey) { 1537 wr->sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1538 (uint64_t)rdma_req->buffers[iovcnt], &translation_len))->lkey; 1539 } else { 1540 wr->sg_list[i].lkey = spdk_mem_map_translate(device->map, 1541 (uint64_t)rdma_req->buffers[iovcnt], &translation_len); 1542 } 1543 1544 remaining_length -= rdma_req->req.iov[iovcnt].iov_len; 1545 1546 if (translation_len < rdma_req->req.iov[iovcnt].iov_len) { 1547 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions\n"); 1548 return -EINVAL; 1549 } 1550 i++; 1551 } 1552 wr->num_sge = i; 1553 1554 return 0; 1555 } 1556 1557 static int 1558 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1559 struct spdk_nvmf_rdma_device *device, 1560 struct spdk_nvmf_rdma_request *rdma_req) 1561 { 1562 struct spdk_nvmf_rdma_qpair *rqpair; 1563 struct spdk_nvmf_rdma_poll_group *rgroup; 1564 uint32_t num_buffers; 1565 uint32_t i = 0; 1566 int rc = 0; 1567 1568 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1569 rgroup = rqpair->poller->group; 1570 rdma_req->req.iovcnt = 0; 1571 1572 num_buffers = rdma_req->req.length / rtransport->transport.opts.io_unit_size; 1573 if (rdma_req->req.length % rtransport->transport.opts.io_unit_size) { 1574 num_buffers++; 1575 } 1576 1577 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers)) { 1578 return -ENOMEM; 1579 } 1580 1581 rdma_req->req.iovcnt = 0; 1582 1583 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, &rdma_req->data.wr, 1584 rdma_req->req.length); 1585 if (rc != 0) { 1586 goto err_exit; 1587 } 1588 1589 assert(rdma_req->req.iovcnt <= rqpair->max_send_sge); 1590 1591 rdma_req->data_from_pool = true; 1592 1593 return rc; 1594 1595 err_exit: 1596 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1597 while (i) { 1598 i--; 1599 rdma_req->data.wr.sg_list[i].addr = 0; 1600 rdma_req->data.wr.sg_list[i].length = 0; 1601 rdma_req->data.wr.sg_list[i].lkey = 0; 1602 } 1603 rdma_req->req.iovcnt = 0; 1604 return rc; 1605 } 1606 1607 static int 1608 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1609 struct spdk_nvmf_rdma_device *device, 1610 struct spdk_nvmf_rdma_request *rdma_req) 1611 { 1612 struct spdk_nvmf_rdma_qpair *rqpair; 1613 struct spdk_nvmf_rdma_poll_group *rgroup; 1614 struct ibv_send_wr *current_wr; 1615 struct spdk_nvmf_request *req = &rdma_req->req; 1616 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1617 uint32_t num_sgl_descriptors; 1618 uint32_t num_buffers = 0; 1619 uint32_t i; 1620 int rc; 1621 1622 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1623 rgroup = rqpair->poller->group; 1624 1625 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1626 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1627 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1628 1629 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1630 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1631 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1632 1633 for (i = 0; i < num_sgl_descriptors; i++) { 1634 num_buffers += desc->keyed.length / rtransport->transport.opts.io_unit_size; 1635 if (desc->keyed.length % rtransport->transport.opts.io_unit_size) { 1636 num_buffers++; 1637 } 1638 desc++; 1639 } 1640 /* If the number of buffers is too large, then we know the I/O is larger than allowed. Fail it. */ 1641 if (num_buffers > NVMF_REQ_MAX_BUFFERS) { 1642 return -EINVAL; 1643 } 1644 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1645 num_buffers) != 0) { 1646 return -ENOMEM; 1647 } 1648 1649 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1650 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1651 return -ENOMEM; 1652 } 1653 1654 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1655 current_wr = &rdma_req->data.wr; 1656 1657 req->iovcnt = 0; 1658 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1659 for (i = 0; i < num_sgl_descriptors; i++) { 1660 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1661 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1662 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1663 rc = -EINVAL; 1664 goto err_exit; 1665 } 1666 1667 current_wr->num_sge = 0; 1668 req->length += desc->keyed.length; 1669 1670 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, current_wr, 1671 desc->keyed.length); 1672 if (rc != 0) { 1673 rc = -ENOMEM; 1674 goto err_exit; 1675 } 1676 1677 current_wr->wr.rdma.rkey = desc->keyed.key; 1678 current_wr->wr.rdma.remote_addr = desc->address; 1679 current_wr = current_wr->next; 1680 desc++; 1681 } 1682 1683 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1684 /* Go back to the last descriptor in the list. */ 1685 desc--; 1686 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1687 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1688 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1689 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1690 } 1691 } 1692 #endif 1693 1694 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1695 rdma_req->data_from_pool = true; 1696 1697 return 0; 1698 1699 err_exit: 1700 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1701 nvmf_rdma_request_free_data(rdma_req, rtransport); 1702 return rc; 1703 } 1704 1705 static int 1706 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1707 struct spdk_nvmf_rdma_device *device, 1708 struct spdk_nvmf_rdma_request *rdma_req) 1709 { 1710 struct spdk_nvme_cmd *cmd; 1711 struct spdk_nvme_cpl *rsp; 1712 struct spdk_nvme_sgl_descriptor *sgl; 1713 int rc; 1714 1715 cmd = &rdma_req->req.cmd->nvme_cmd; 1716 rsp = &rdma_req->req.rsp->nvme_cpl; 1717 sgl = &cmd->dptr.sgl1; 1718 1719 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1720 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1721 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1722 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1723 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1724 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1725 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1726 return -1; 1727 } 1728 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1729 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1730 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1731 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1732 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1733 } 1734 } 1735 #endif 1736 1737 /* fill request length and populate iovs */ 1738 rdma_req->req.length = sgl->keyed.length; 1739 1740 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1741 /* No available buffers. Queue this request up. */ 1742 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1743 return 0; 1744 } 1745 1746 /* backward compatible */ 1747 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1748 1749 /* rdma wr specifics */ 1750 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1751 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1752 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1753 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1754 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 1755 rdma_req->data.wr.next = &rdma_req->rsp.wr; 1756 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1757 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 1758 rdma_req->data.wr.next = NULL; 1759 } 1760 1761 /* set the number of outstanding data WRs for this request. */ 1762 rdma_req->num_outstanding_data_wr = 1; 1763 1764 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1765 rdma_req->req.iovcnt); 1766 1767 return 0; 1768 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1769 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1770 uint64_t offset = sgl->address; 1771 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1772 1773 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1774 offset, sgl->unkeyed.length); 1775 1776 if (offset > max_len) { 1777 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1778 offset, max_len); 1779 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1780 return -1; 1781 } 1782 max_len -= (uint32_t)offset; 1783 1784 if (sgl->unkeyed.length > max_len) { 1785 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1786 sgl->unkeyed.length, max_len); 1787 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1788 return -1; 1789 } 1790 1791 rdma_req->num_outstanding_data_wr = 0; 1792 rdma_req->req.data = rdma_req->recv->buf + offset; 1793 rdma_req->data_from_pool = false; 1794 rdma_req->req.length = sgl->unkeyed.length; 1795 1796 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1797 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1798 rdma_req->req.iovcnt = 1; 1799 1800 return 0; 1801 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1802 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1803 1804 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1805 if (rc == -ENOMEM) { 1806 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1807 return 0; 1808 } else if (rc == -EINVAL) { 1809 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1810 return -1; 1811 } 1812 1813 /* backward compatible */ 1814 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1815 1816 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1817 rdma_req->req.iovcnt); 1818 1819 return 0; 1820 } 1821 1822 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1823 sgl->generic.type, sgl->generic.subtype); 1824 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1825 return -1; 1826 } 1827 1828 static void 1829 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1830 struct spdk_nvmf_rdma_transport *rtransport) 1831 { 1832 struct spdk_nvmf_rdma_qpair *rqpair; 1833 struct spdk_nvmf_rdma_poll_group *rgroup; 1834 1835 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1836 if (rdma_req->data_from_pool) { 1837 rgroup = rqpair->poller->group; 1838 1839 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1840 rdma_req->req.iovcnt); 1841 } 1842 nvmf_rdma_request_free_data(rdma_req, rtransport); 1843 rdma_req->req.length = 0; 1844 rdma_req->req.iovcnt = 0; 1845 rdma_req->req.data = NULL; 1846 rdma_req->rsp.wr.next = NULL; 1847 rdma_req->data.wr.next = NULL; 1848 rqpair->qd--; 1849 1850 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1851 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1852 } 1853 1854 static bool 1855 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1856 struct spdk_nvmf_rdma_request *rdma_req) 1857 { 1858 struct spdk_nvmf_rdma_qpair *rqpair; 1859 struct spdk_nvmf_rdma_device *device; 1860 struct spdk_nvmf_rdma_poll_group *rgroup; 1861 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1862 int rc; 1863 struct spdk_nvmf_rdma_recv *rdma_recv; 1864 enum spdk_nvmf_rdma_request_state prev_state; 1865 bool progress = false; 1866 int data_posted; 1867 1868 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1869 device = rqpair->port->device; 1870 rgroup = rqpair->poller->group; 1871 1872 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1873 1874 /* If the queue pair is in an error state, force the request to the completed state 1875 * to release resources. */ 1876 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1877 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1878 STAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1879 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1880 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1881 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1882 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1883 } 1884 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1885 } 1886 1887 /* The loop here is to allow for several back-to-back state changes. */ 1888 do { 1889 prev_state = rdma_req->state; 1890 1891 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1892 1893 switch (rdma_req->state) { 1894 case RDMA_REQUEST_STATE_FREE: 1895 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1896 * to escape this state. */ 1897 break; 1898 case RDMA_REQUEST_STATE_NEW: 1899 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1900 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1901 rdma_recv = rdma_req->recv; 1902 1903 /* The first element of the SGL is the NVMe command */ 1904 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1905 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1906 1907 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1908 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1909 break; 1910 } 1911 1912 /* The next state transition depends on the data transfer needs of this request. */ 1913 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1914 1915 /* If no data to transfer, ready to execute. */ 1916 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1917 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1918 break; 1919 } 1920 1921 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1922 STAILQ_INSERT_TAIL(&rgroup->pending_data_buf_queue, rdma_req, state_link); 1923 break; 1924 case RDMA_REQUEST_STATE_NEED_BUFFER: 1925 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1926 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1927 1928 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1929 1930 if (rdma_req != STAILQ_FIRST(&rgroup->pending_data_buf_queue)) { 1931 /* This request needs to wait in line to obtain a buffer */ 1932 break; 1933 } 1934 1935 /* Try to get a data buffer */ 1936 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1937 if (rc < 0) { 1938 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1939 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1940 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1941 break; 1942 } 1943 1944 if (!rdma_req->req.data) { 1945 /* No buffers available. */ 1946 break; 1947 } 1948 1949 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1950 1951 /* If data is transferring from host to controller and the data didn't 1952 * arrive using in capsule data, we need to do a transfer from the host. 1953 */ 1954 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1955 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1956 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1957 break; 1958 } 1959 1960 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1961 break; 1962 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1963 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1964 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1965 1966 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1967 /* This request needs to wait in line to perform RDMA */ 1968 break; 1969 } 1970 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1971 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1972 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1973 break; 1974 } 1975 1976 /* We have already verified that this request is the head of the queue. */ 1977 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 1978 1979 rc = request_transfer_in(&rdma_req->req); 1980 if (!rc) { 1981 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 1982 } else { 1983 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1984 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1985 } 1986 break; 1987 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1988 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1989 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1990 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1991 * to escape this state. */ 1992 break; 1993 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1994 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1995 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1996 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 1997 spdk_nvmf_request_exec(&rdma_req->req); 1998 break; 1999 case RDMA_REQUEST_STATE_EXECUTING: 2000 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2001 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2002 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2003 * to escape this state. */ 2004 break; 2005 case RDMA_REQUEST_STATE_EXECUTED: 2006 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2007 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2008 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2009 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2010 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2011 } else { 2012 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2013 } 2014 break; 2015 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2016 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2017 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2018 2019 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2020 /* This request needs to wait in line to perform RDMA */ 2021 break; 2022 } 2023 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2024 rqpair->max_send_depth) { 2025 /* We can only have so many WRs outstanding. we have to wait until some finish. 2026 * +1 since each request has an additional wr in the resp. */ 2027 break; 2028 } 2029 2030 /* We have already verified that this request is the head of the queue. */ 2031 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2032 2033 /* The data transfer will be kicked off from 2034 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2035 */ 2036 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2037 break; 2038 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2039 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2040 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2041 rc = request_transfer_out(&rdma_req->req, &data_posted); 2042 assert(rc == 0); /* No good way to handle this currently */ 2043 if (rc) { 2044 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2045 } else { 2046 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2047 RDMA_REQUEST_STATE_COMPLETING; 2048 } 2049 break; 2050 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2051 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2052 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2053 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2054 * to escape this state. */ 2055 break; 2056 case RDMA_REQUEST_STATE_COMPLETING: 2057 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2058 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2059 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2060 * to escape this state. */ 2061 break; 2062 case RDMA_REQUEST_STATE_COMPLETED: 2063 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2064 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2065 2066 nvmf_rdma_request_free(rdma_req, rtransport); 2067 break; 2068 case RDMA_REQUEST_NUM_STATES: 2069 default: 2070 assert(0); 2071 break; 2072 } 2073 2074 if (rdma_req->state != prev_state) { 2075 progress = true; 2076 } 2077 } while (rdma_req->state != prev_state); 2078 2079 return progress; 2080 } 2081 2082 /* Public API callbacks begin here */ 2083 2084 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2085 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2086 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2087 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 2088 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2089 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2090 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2091 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2092 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2093 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false; 2094 2095 static void 2096 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2097 { 2098 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2099 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2100 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2101 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2102 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2103 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2104 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2105 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2106 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2107 opts->no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ 2108 } 2109 2110 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2111 .notify_cb = spdk_nvmf_rdma_mem_notify, 2112 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 2113 }; 2114 2115 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2116 2117 static struct spdk_nvmf_transport * 2118 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2119 { 2120 int rc; 2121 struct spdk_nvmf_rdma_transport *rtransport; 2122 struct spdk_nvmf_rdma_device *device, *tmp; 2123 struct ibv_pd *pd; 2124 struct ibv_context **contexts; 2125 uint32_t i; 2126 int flag; 2127 uint32_t sge_count; 2128 uint32_t min_shared_buffers; 2129 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2130 2131 rtransport = calloc(1, sizeof(*rtransport)); 2132 if (!rtransport) { 2133 return NULL; 2134 } 2135 2136 if (pthread_mutex_init(&rtransport->lock, NULL)) { 2137 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2138 free(rtransport); 2139 return NULL; 2140 } 2141 2142 TAILQ_INIT(&rtransport->devices); 2143 TAILQ_INIT(&rtransport->ports); 2144 2145 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2146 2147 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2148 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2149 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2150 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2151 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d\n", 2152 opts->max_queue_depth, 2153 opts->max_io_size, 2154 opts->max_qpairs_per_ctrlr, 2155 opts->io_unit_size, 2156 opts->in_capsule_data_size, 2157 opts->max_aq_depth, 2158 opts->num_shared_buffers, 2159 opts->max_srq_depth, 2160 opts->no_srq); 2161 2162 /* I/O unit size cannot be larger than max I/O size */ 2163 if (opts->io_unit_size > opts->max_io_size) { 2164 opts->io_unit_size = opts->max_io_size; 2165 } 2166 2167 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2168 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2169 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2170 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2171 spdk_nvmf_rdma_destroy(&rtransport->transport); 2172 return NULL; 2173 } 2174 2175 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2176 if (min_shared_buffers > opts->num_shared_buffers) { 2177 SPDK_ERRLOG("There are not enough buffers to satisfy" 2178 "per-poll group caches for each thread. (%" PRIu32 ")" 2179 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2180 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2181 spdk_nvmf_rdma_destroy(&rtransport->transport); 2182 return NULL; 2183 } 2184 2185 sge_count = opts->max_io_size / opts->io_unit_size; 2186 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2187 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2188 spdk_nvmf_rdma_destroy(&rtransport->transport); 2189 return NULL; 2190 } 2191 2192 rtransport->event_channel = rdma_create_event_channel(); 2193 if (rtransport->event_channel == NULL) { 2194 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2195 spdk_nvmf_rdma_destroy(&rtransport->transport); 2196 return NULL; 2197 } 2198 2199 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2200 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2201 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2202 rtransport->event_channel->fd, spdk_strerror(errno)); 2203 spdk_nvmf_rdma_destroy(&rtransport->transport); 2204 return NULL; 2205 } 2206 2207 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2208 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2209 sizeof(struct spdk_nvmf_rdma_request_data), 2210 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2211 SPDK_ENV_SOCKET_ID_ANY); 2212 if (!rtransport->data_wr_pool) { 2213 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2214 spdk_nvmf_rdma_destroy(&rtransport->transport); 2215 return NULL; 2216 } 2217 2218 contexts = rdma_get_devices(NULL); 2219 if (contexts == NULL) { 2220 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2221 spdk_nvmf_rdma_destroy(&rtransport->transport); 2222 return NULL; 2223 } 2224 2225 i = 0; 2226 rc = 0; 2227 while (contexts[i] != NULL) { 2228 device = calloc(1, sizeof(*device)); 2229 if (!device) { 2230 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2231 rc = -ENOMEM; 2232 break; 2233 } 2234 device->context = contexts[i]; 2235 rc = ibv_query_device(device->context, &device->attr); 2236 if (rc < 0) { 2237 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2238 free(device); 2239 break; 2240 2241 } 2242 2243 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2244 2245 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2246 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2247 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2248 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2249 } 2250 2251 /** 2252 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2253 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2254 * but incorrectly reports that it does. There are changes making their way 2255 * through the kernel now that will enable this feature. When they are merged, 2256 * we can conditionally enable this feature. 2257 * 2258 * TODO: enable this for versions of the kernel rxe driver that support it. 2259 */ 2260 if (device->attr.vendor_id == 0) { 2261 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2262 } 2263 #endif 2264 2265 /* set up device context async ev fd as NON_BLOCKING */ 2266 flag = fcntl(device->context->async_fd, F_GETFL); 2267 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2268 if (rc < 0) { 2269 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2270 free(device); 2271 break; 2272 } 2273 2274 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2275 i++; 2276 2277 pd = NULL; 2278 if (g_nvmf_hooks.get_ibv_pd) { 2279 pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2280 } 2281 2282 if (!g_nvmf_hooks.get_ibv_pd) { 2283 device->pd = ibv_alloc_pd(device->context); 2284 if (!device->pd) { 2285 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2286 spdk_nvmf_rdma_destroy(&rtransport->transport); 2287 return NULL; 2288 } 2289 } else { 2290 device->pd = pd; 2291 } 2292 2293 assert(device->map == NULL); 2294 2295 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2296 if (!device->map) { 2297 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2298 spdk_nvmf_rdma_destroy(&rtransport->transport); 2299 return NULL; 2300 } 2301 2302 assert(device->map != NULL); 2303 assert(device->pd != NULL); 2304 } 2305 rdma_free_devices(contexts); 2306 2307 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2308 /* divide and round up. */ 2309 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2310 2311 /* round up to the nearest 4k. */ 2312 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2313 2314 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2315 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2316 opts->io_unit_size); 2317 } 2318 2319 if (rc < 0) { 2320 spdk_nvmf_rdma_destroy(&rtransport->transport); 2321 return NULL; 2322 } 2323 2324 /* Set up poll descriptor array to monitor events from RDMA and IB 2325 * in a single poll syscall 2326 */ 2327 rtransport->npoll_fds = i + 1; 2328 i = 0; 2329 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2330 if (rtransport->poll_fds == NULL) { 2331 SPDK_ERRLOG("poll_fds allocation failed\n"); 2332 spdk_nvmf_rdma_destroy(&rtransport->transport); 2333 return NULL; 2334 } 2335 2336 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2337 rtransport->poll_fds[i++].events = POLLIN; 2338 2339 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2340 rtransport->poll_fds[i].fd = device->context->async_fd; 2341 rtransport->poll_fds[i++].events = POLLIN; 2342 } 2343 2344 return &rtransport->transport; 2345 } 2346 2347 static int 2348 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2349 { 2350 struct spdk_nvmf_rdma_transport *rtransport; 2351 struct spdk_nvmf_rdma_port *port, *port_tmp; 2352 struct spdk_nvmf_rdma_device *device, *device_tmp; 2353 2354 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2355 2356 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2357 TAILQ_REMOVE(&rtransport->ports, port, link); 2358 rdma_destroy_id(port->id); 2359 free(port); 2360 } 2361 2362 if (rtransport->poll_fds != NULL) { 2363 free(rtransport->poll_fds); 2364 } 2365 2366 if (rtransport->event_channel != NULL) { 2367 rdma_destroy_event_channel(rtransport->event_channel); 2368 } 2369 2370 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2371 TAILQ_REMOVE(&rtransport->devices, device, link); 2372 if (device->map) { 2373 spdk_mem_map_free(&device->map); 2374 } 2375 if (device->pd) { 2376 if (!g_nvmf_hooks.get_ibv_pd) { 2377 ibv_dealloc_pd(device->pd); 2378 } 2379 } 2380 free(device); 2381 } 2382 2383 if (rtransport->data_wr_pool != NULL) { 2384 if (spdk_mempool_count(rtransport->data_wr_pool) != 2385 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2386 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2387 spdk_mempool_count(rtransport->data_wr_pool), 2388 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2389 } 2390 } 2391 2392 spdk_mempool_free(rtransport->data_wr_pool); 2393 pthread_mutex_destroy(&rtransport->lock); 2394 free(rtransport); 2395 2396 return 0; 2397 } 2398 2399 static int 2400 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2401 struct spdk_nvme_transport_id *trid, 2402 bool peer); 2403 2404 static int 2405 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2406 const struct spdk_nvme_transport_id *trid) 2407 { 2408 struct spdk_nvmf_rdma_transport *rtransport; 2409 struct spdk_nvmf_rdma_device *device; 2410 struct spdk_nvmf_rdma_port *port_tmp, *port; 2411 struct addrinfo *res; 2412 struct addrinfo hints; 2413 int family; 2414 int rc; 2415 2416 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2417 2418 port = calloc(1, sizeof(*port)); 2419 if (!port) { 2420 return -ENOMEM; 2421 } 2422 2423 /* Selectively copy the trid. Things like NQN don't matter here - that 2424 * mapping is enforced elsewhere. 2425 */ 2426 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2427 port->trid.adrfam = trid->adrfam; 2428 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 2429 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 2430 2431 pthread_mutex_lock(&rtransport->lock); 2432 assert(rtransport->event_channel != NULL); 2433 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 2434 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 2435 port_tmp->ref++; 2436 free(port); 2437 /* Already listening at this address */ 2438 pthread_mutex_unlock(&rtransport->lock); 2439 return 0; 2440 } 2441 } 2442 2443 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2444 if (rc < 0) { 2445 SPDK_ERRLOG("rdma_create_id() failed\n"); 2446 free(port); 2447 pthread_mutex_unlock(&rtransport->lock); 2448 return rc; 2449 } 2450 2451 switch (port->trid.adrfam) { 2452 case SPDK_NVMF_ADRFAM_IPV4: 2453 family = AF_INET; 2454 break; 2455 case SPDK_NVMF_ADRFAM_IPV6: 2456 family = AF_INET6; 2457 break; 2458 default: 2459 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 2460 free(port); 2461 pthread_mutex_unlock(&rtransport->lock); 2462 return -EINVAL; 2463 } 2464 2465 memset(&hints, 0, sizeof(hints)); 2466 hints.ai_family = family; 2467 hints.ai_flags = AI_NUMERICSERV; 2468 hints.ai_socktype = SOCK_STREAM; 2469 hints.ai_protocol = 0; 2470 2471 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 2472 if (rc) { 2473 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2474 free(port); 2475 pthread_mutex_unlock(&rtransport->lock); 2476 return -EINVAL; 2477 } 2478 2479 rc = rdma_bind_addr(port->id, res->ai_addr); 2480 freeaddrinfo(res); 2481 2482 if (rc < 0) { 2483 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2484 rdma_destroy_id(port->id); 2485 free(port); 2486 pthread_mutex_unlock(&rtransport->lock); 2487 return rc; 2488 } 2489 2490 if (!port->id->verbs) { 2491 SPDK_ERRLOG("ibv_context is null\n"); 2492 rdma_destroy_id(port->id); 2493 free(port); 2494 pthread_mutex_unlock(&rtransport->lock); 2495 return -1; 2496 } 2497 2498 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 2499 if (rc < 0) { 2500 SPDK_ERRLOG("rdma_listen() failed\n"); 2501 rdma_destroy_id(port->id); 2502 free(port); 2503 pthread_mutex_unlock(&rtransport->lock); 2504 return rc; 2505 } 2506 2507 TAILQ_FOREACH(device, &rtransport->devices, link) { 2508 if (device->context == port->id->verbs) { 2509 port->device = device; 2510 break; 2511 } 2512 } 2513 if (!port->device) { 2514 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2515 port->id->verbs); 2516 rdma_destroy_id(port->id); 2517 free(port); 2518 pthread_mutex_unlock(&rtransport->lock); 2519 return -EINVAL; 2520 } 2521 2522 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 2523 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 2524 2525 port->ref = 1; 2526 2527 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2528 pthread_mutex_unlock(&rtransport->lock); 2529 2530 return 0; 2531 } 2532 2533 static int 2534 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2535 const struct spdk_nvme_transport_id *_trid) 2536 { 2537 struct spdk_nvmf_rdma_transport *rtransport; 2538 struct spdk_nvmf_rdma_port *port, *tmp; 2539 struct spdk_nvme_transport_id trid = {}; 2540 2541 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2542 2543 /* Selectively copy the trid. Things like NQN don't matter here - that 2544 * mapping is enforced elsewhere. 2545 */ 2546 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2547 trid.adrfam = _trid->adrfam; 2548 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2549 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2550 2551 pthread_mutex_lock(&rtransport->lock); 2552 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2553 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2554 assert(port->ref > 0); 2555 port->ref--; 2556 if (port->ref == 0) { 2557 TAILQ_REMOVE(&rtransport->ports, port, link); 2558 rdma_destroy_id(port->id); 2559 free(port); 2560 } 2561 break; 2562 } 2563 } 2564 2565 pthread_mutex_unlock(&rtransport->lock); 2566 return 0; 2567 } 2568 2569 static void 2570 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2571 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2572 { 2573 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2574 struct spdk_nvmf_rdma_resources *resources; 2575 2576 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2577 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2578 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2579 break; 2580 } 2581 } 2582 2583 /* Then RDMA writes since reads have stronger restrictions than writes */ 2584 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2585 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2586 break; 2587 } 2588 } 2589 2590 /* The second highest priority is I/O waiting on memory buffers. */ 2591 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->poller->group->pending_data_buf_queue, state_link, 2592 req_tmp) { 2593 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2594 break; 2595 } 2596 } 2597 2598 resources = rqpair->resources; 2599 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2600 rdma_req = STAILQ_FIRST(&resources->free_queue); 2601 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2602 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2603 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2604 2605 if (rqpair->srq != NULL) { 2606 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2607 rdma_req->recv->qpair->qd++; 2608 } else { 2609 rqpair->qd++; 2610 } 2611 2612 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2613 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2614 break; 2615 } 2616 } 2617 } 2618 2619 static void 2620 _nvmf_rdma_qpair_disconnect(void *ctx) 2621 { 2622 struct spdk_nvmf_qpair *qpair = ctx; 2623 2624 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2625 } 2626 2627 static void 2628 _nvmf_rdma_try_disconnect(void *ctx) 2629 { 2630 struct spdk_nvmf_qpair *qpair = ctx; 2631 struct spdk_nvmf_poll_group *group; 2632 2633 /* Read the group out of the qpair. This is normally set and accessed only from 2634 * the thread that created the group. Here, we're not on that thread necessarily. 2635 * The data member qpair->group begins it's life as NULL and then is assigned to 2636 * a pointer and never changes. So fortunately reading this and checking for 2637 * non-NULL is thread safe in the x86_64 memory model. */ 2638 group = qpair->group; 2639 2640 if (group == NULL) { 2641 /* The qpair hasn't been assigned to a group yet, so we can't 2642 * process a disconnect. Send a message to ourself and try again. */ 2643 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2644 return; 2645 } 2646 2647 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2648 } 2649 2650 static inline void 2651 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2652 { 2653 if (__sync_bool_compare_and_swap(&rqpair->disconnect_started, false, true)) { 2654 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2655 } 2656 } 2657 2658 static void nvmf_rdma_destroy_drained_qpair(void *ctx) 2659 { 2660 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2661 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2662 struct spdk_nvmf_rdma_transport, transport); 2663 2664 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2665 if (rqpair->current_send_depth != 0) { 2666 return; 2667 } 2668 2669 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2670 return; 2671 } 2672 2673 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) { 2674 return; 2675 } 2676 2677 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2678 spdk_nvmf_rdma_qpair_destroy(rqpair); 2679 } 2680 2681 2682 static int 2683 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2684 { 2685 struct spdk_nvmf_qpair *qpair; 2686 struct spdk_nvmf_rdma_qpair *rqpair; 2687 2688 if (evt->id == NULL) { 2689 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2690 return -1; 2691 } 2692 2693 qpair = evt->id->context; 2694 if (qpair == NULL) { 2695 SPDK_ERRLOG("disconnect request: no active connection\n"); 2696 return -1; 2697 } 2698 2699 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2700 2701 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2702 2703 spdk_nvmf_rdma_update_ibv_state(rqpair); 2704 2705 spdk_nvmf_rdma_start_disconnect(rqpair); 2706 2707 return 0; 2708 } 2709 2710 #ifdef DEBUG 2711 static const char *CM_EVENT_STR[] = { 2712 "RDMA_CM_EVENT_ADDR_RESOLVED", 2713 "RDMA_CM_EVENT_ADDR_ERROR", 2714 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2715 "RDMA_CM_EVENT_ROUTE_ERROR", 2716 "RDMA_CM_EVENT_CONNECT_REQUEST", 2717 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2718 "RDMA_CM_EVENT_CONNECT_ERROR", 2719 "RDMA_CM_EVENT_UNREACHABLE", 2720 "RDMA_CM_EVENT_REJECTED", 2721 "RDMA_CM_EVENT_ESTABLISHED", 2722 "RDMA_CM_EVENT_DISCONNECTED", 2723 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2724 "RDMA_CM_EVENT_MULTICAST_JOIN", 2725 "RDMA_CM_EVENT_MULTICAST_ERROR", 2726 "RDMA_CM_EVENT_ADDR_CHANGE", 2727 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2728 }; 2729 #endif /* DEBUG */ 2730 2731 static void 2732 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2733 { 2734 struct spdk_nvmf_rdma_transport *rtransport; 2735 struct rdma_cm_event *event; 2736 int rc; 2737 2738 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2739 2740 if (rtransport->event_channel == NULL) { 2741 return; 2742 } 2743 2744 while (1) { 2745 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2746 if (rc == 0) { 2747 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2748 2749 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2750 2751 switch (event->event) { 2752 case RDMA_CM_EVENT_ADDR_RESOLVED: 2753 case RDMA_CM_EVENT_ADDR_ERROR: 2754 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2755 case RDMA_CM_EVENT_ROUTE_ERROR: 2756 /* No action required. The target never attempts to resolve routes. */ 2757 break; 2758 case RDMA_CM_EVENT_CONNECT_REQUEST: 2759 rc = nvmf_rdma_connect(transport, event, cb_fn); 2760 if (rc < 0) { 2761 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2762 break; 2763 } 2764 break; 2765 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2766 /* The target never initiates a new connection. So this will not occur. */ 2767 break; 2768 case RDMA_CM_EVENT_CONNECT_ERROR: 2769 /* Can this happen? The docs say it can, but not sure what causes it. */ 2770 break; 2771 case RDMA_CM_EVENT_UNREACHABLE: 2772 case RDMA_CM_EVENT_REJECTED: 2773 /* These only occur on the client side. */ 2774 break; 2775 case RDMA_CM_EVENT_ESTABLISHED: 2776 /* TODO: Should we be waiting for this event anywhere? */ 2777 break; 2778 case RDMA_CM_EVENT_DISCONNECTED: 2779 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2780 rc = nvmf_rdma_disconnect(event); 2781 if (rc < 0) { 2782 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2783 break; 2784 } 2785 break; 2786 case RDMA_CM_EVENT_MULTICAST_JOIN: 2787 case RDMA_CM_EVENT_MULTICAST_ERROR: 2788 /* Multicast is not used */ 2789 break; 2790 case RDMA_CM_EVENT_ADDR_CHANGE: 2791 /* Not utilizing this event */ 2792 break; 2793 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2794 /* For now, do nothing. The target never re-uses queue pairs. */ 2795 break; 2796 default: 2797 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2798 break; 2799 } 2800 2801 rdma_ack_cm_event(event); 2802 } else { 2803 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2804 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2805 } 2806 break; 2807 } 2808 } 2809 } 2810 2811 static void 2812 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2813 { 2814 int rc; 2815 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 2816 struct ibv_async_event event; 2817 enum ibv_qp_state state; 2818 2819 rc = ibv_get_async_event(device->context, &event); 2820 2821 if (rc) { 2822 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2823 errno, spdk_strerror(errno)); 2824 return; 2825 } 2826 2827 switch (event.event_type) { 2828 case IBV_EVENT_QP_FATAL: 2829 rqpair = event.element.qp->qp_context; 2830 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 2831 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2832 (uintptr_t)rqpair->cm_id, event.event_type); 2833 spdk_nvmf_rdma_update_ibv_state(rqpair); 2834 spdk_nvmf_rdma_start_disconnect(rqpair); 2835 break; 2836 case IBV_EVENT_QP_LAST_WQE_REACHED: 2837 /* This event only occurs for shared receive queues. */ 2838 rqpair = event.element.qp->qp_context; 2839 rqpair->last_wqe_reached = true; 2840 2841 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last WQE reached event received for rqpair %p\n", rqpair); 2842 /* This must be handled on the polling thread if it exists. Otherwise the timeout will catch it. */ 2843 if (rqpair->qpair.group) { 2844 spdk_thread_send_msg(rqpair->qpair.group->thread, nvmf_rdma_destroy_drained_qpair, rqpair); 2845 } else { 2846 SPDK_ERRLOG("Unable to destroy the qpair %p since it does not have a poll group.\n", rqpair); 2847 } 2848 2849 break; 2850 case IBV_EVENT_SQ_DRAINED: 2851 /* This event occurs frequently in both error and non-error states. 2852 * Check if the qpair is in an error state before sending a message. 2853 * Note that we're not on the correct thread to access the qpair, but 2854 * the operations that the below calls make all happen to be thread 2855 * safe. */ 2856 rqpair = event.element.qp->qp_context; 2857 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last sq drained event received for rqpair %p\n", rqpair); 2858 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2859 (uintptr_t)rqpair->cm_id, event.event_type); 2860 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2861 if (state == IBV_QPS_ERR) { 2862 spdk_nvmf_rdma_start_disconnect(rqpair); 2863 } 2864 break; 2865 case IBV_EVENT_QP_REQ_ERR: 2866 case IBV_EVENT_QP_ACCESS_ERR: 2867 case IBV_EVENT_COMM_EST: 2868 case IBV_EVENT_PATH_MIG: 2869 case IBV_EVENT_PATH_MIG_ERR: 2870 SPDK_NOTICELOG("Async event: %s\n", 2871 ibv_event_type_str(event.event_type)); 2872 rqpair = event.element.qp->qp_context; 2873 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2874 (uintptr_t)rqpair->cm_id, event.event_type); 2875 spdk_nvmf_rdma_update_ibv_state(rqpair); 2876 break; 2877 case IBV_EVENT_CQ_ERR: 2878 case IBV_EVENT_DEVICE_FATAL: 2879 case IBV_EVENT_PORT_ACTIVE: 2880 case IBV_EVENT_PORT_ERR: 2881 case IBV_EVENT_LID_CHANGE: 2882 case IBV_EVENT_PKEY_CHANGE: 2883 case IBV_EVENT_SM_CHANGE: 2884 case IBV_EVENT_SRQ_ERR: 2885 case IBV_EVENT_SRQ_LIMIT_REACHED: 2886 case IBV_EVENT_CLIENT_REREGISTER: 2887 case IBV_EVENT_GID_CHANGE: 2888 default: 2889 SPDK_NOTICELOG("Async event: %s\n", 2890 ibv_event_type_str(event.event_type)); 2891 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2892 break; 2893 } 2894 ibv_ack_async_event(&event); 2895 } 2896 2897 static void 2898 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2899 { 2900 int nfds, i = 0; 2901 struct spdk_nvmf_rdma_transport *rtransport; 2902 struct spdk_nvmf_rdma_device *device, *tmp; 2903 2904 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2905 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2906 2907 if (nfds <= 0) { 2908 return; 2909 } 2910 2911 /* The first poll descriptor is RDMA CM event */ 2912 if (rtransport->poll_fds[i++].revents & POLLIN) { 2913 spdk_nvmf_process_cm_event(transport, cb_fn); 2914 nfds--; 2915 } 2916 2917 if (nfds == 0) { 2918 return; 2919 } 2920 2921 /* Second and subsequent poll descriptors are IB async events */ 2922 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2923 if (rtransport->poll_fds[i++].revents & POLLIN) { 2924 spdk_nvmf_process_ib_event(device); 2925 nfds--; 2926 } 2927 } 2928 /* check all flagged fd's have been served */ 2929 assert(nfds == 0); 2930 } 2931 2932 static void 2933 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2934 struct spdk_nvme_transport_id *trid, 2935 struct spdk_nvmf_discovery_log_page_entry *entry) 2936 { 2937 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2938 entry->adrfam = trid->adrfam; 2939 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2940 2941 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2942 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2943 2944 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2945 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2946 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2947 } 2948 2949 static void 2950 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 2951 2952 static struct spdk_nvmf_transport_poll_group * 2953 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2954 { 2955 struct spdk_nvmf_rdma_transport *rtransport; 2956 struct spdk_nvmf_rdma_poll_group *rgroup; 2957 struct spdk_nvmf_rdma_poller *poller; 2958 struct spdk_nvmf_rdma_device *device; 2959 struct ibv_srq_init_attr srq_init_attr; 2960 struct spdk_nvmf_rdma_resource_opts opts; 2961 int num_cqe; 2962 2963 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2964 2965 rgroup = calloc(1, sizeof(*rgroup)); 2966 if (!rgroup) { 2967 return NULL; 2968 } 2969 2970 TAILQ_INIT(&rgroup->pollers); 2971 STAILQ_INIT(&rgroup->pending_data_buf_queue); 2972 2973 pthread_mutex_lock(&rtransport->lock); 2974 TAILQ_FOREACH(device, &rtransport->devices, link) { 2975 poller = calloc(1, sizeof(*poller)); 2976 if (!poller) { 2977 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2978 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2979 pthread_mutex_unlock(&rtransport->lock); 2980 return NULL; 2981 } 2982 2983 poller->device = device; 2984 poller->group = rgroup; 2985 2986 TAILQ_INIT(&poller->qpairs); 2987 STAILQ_INIT(&poller->qpairs_pending_send); 2988 STAILQ_INIT(&poller->qpairs_pending_recv); 2989 2990 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2991 if (transport->opts.no_srq == false && device->attr.max_srq != 0) { 2992 poller->max_srq_depth = transport->opts.max_srq_depth; 2993 2994 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 2995 srq_init_attr.attr.max_wr = poller->max_srq_depth; 2996 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 2997 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 2998 if (!poller->srq) { 2999 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3000 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3001 pthread_mutex_unlock(&rtransport->lock); 3002 return NULL; 3003 } 3004 3005 opts.qp = poller->srq; 3006 opts.pd = device->pd; 3007 opts.qpair = NULL; 3008 opts.shared = true; 3009 opts.max_queue_depth = poller->max_srq_depth; 3010 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3011 3012 poller->resources = nvmf_rdma_resources_create(&opts); 3013 if (!poller->resources) { 3014 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3015 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3016 pthread_mutex_unlock(&rtransport->lock); 3017 } 3018 } 3019 3020 /* 3021 * When using an srq, we can limit the completion queue at startup. 3022 * The following formula represents the calculation: 3023 * num_cqe = num_recv + num_data_wr + num_send_wr. 3024 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3025 */ 3026 if (poller->srq) { 3027 num_cqe = poller->max_srq_depth * 3; 3028 } else { 3029 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3030 } 3031 3032 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3033 if (!poller->cq) { 3034 SPDK_ERRLOG("Unable to create completion queue\n"); 3035 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3036 pthread_mutex_unlock(&rtransport->lock); 3037 return NULL; 3038 } 3039 poller->num_cqe = num_cqe; 3040 } 3041 3042 pthread_mutex_unlock(&rtransport->lock); 3043 return &rgroup->group; 3044 } 3045 3046 static void 3047 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3048 { 3049 struct spdk_nvmf_rdma_poll_group *rgroup; 3050 struct spdk_nvmf_rdma_poller *poller, *tmp; 3051 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3052 3053 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3054 3055 if (!rgroup) { 3056 return; 3057 } 3058 3059 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3060 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3061 3062 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3063 spdk_nvmf_rdma_qpair_destroy(qpair); 3064 } 3065 3066 if (poller->srq) { 3067 nvmf_rdma_resources_destroy(poller->resources); 3068 ibv_destroy_srq(poller->srq); 3069 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3070 } 3071 3072 if (poller->cq) { 3073 ibv_destroy_cq(poller->cq); 3074 } 3075 3076 free(poller); 3077 } 3078 3079 if (!STAILQ_EMPTY(&rgroup->pending_data_buf_queue)) { 3080 SPDK_ERRLOG("Pending I/O list wasn't empty on poll group destruction\n"); 3081 } 3082 3083 free(rgroup); 3084 } 3085 3086 static void 3087 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3088 { 3089 if (rqpair->cm_id != NULL) { 3090 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3091 } 3092 spdk_nvmf_rdma_qpair_destroy(rqpair); 3093 } 3094 3095 static int 3096 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3097 struct spdk_nvmf_qpair *qpair) 3098 { 3099 struct spdk_nvmf_rdma_poll_group *rgroup; 3100 struct spdk_nvmf_rdma_qpair *rqpair; 3101 struct spdk_nvmf_rdma_device *device; 3102 struct spdk_nvmf_rdma_poller *poller; 3103 int rc; 3104 3105 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3106 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3107 3108 device = rqpair->port->device; 3109 3110 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3111 if (poller->device == device) { 3112 break; 3113 } 3114 } 3115 3116 if (!poller) { 3117 SPDK_ERRLOG("No poller found for device.\n"); 3118 return -1; 3119 } 3120 3121 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3122 rqpair->poller = poller; 3123 rqpair->srq = rqpair->poller->srq; 3124 3125 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 3126 if (rc < 0) { 3127 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3128 return -1; 3129 } 3130 3131 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3132 if (rc) { 3133 /* Try to reject, but we probably can't */ 3134 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3135 return -1; 3136 } 3137 3138 spdk_nvmf_rdma_update_ibv_state(rqpair); 3139 3140 return 0; 3141 } 3142 3143 static int 3144 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3145 { 3146 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3147 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3148 struct spdk_nvmf_rdma_transport, transport); 3149 3150 nvmf_rdma_request_free(rdma_req, rtransport); 3151 return 0; 3152 } 3153 3154 static int 3155 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3156 { 3157 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3158 struct spdk_nvmf_rdma_transport, transport); 3159 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3160 struct spdk_nvmf_rdma_request, req); 3161 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3162 struct spdk_nvmf_rdma_qpair, qpair); 3163 3164 if (rqpair->ibv_state != IBV_QPS_ERR) { 3165 /* The connection is alive, so process the request as normal */ 3166 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3167 } else { 3168 /* The connection is dead. Move the request directly to the completed state. */ 3169 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3170 } 3171 3172 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3173 3174 return 0; 3175 } 3176 3177 static int 3178 spdk_nvmf_rdma_destroy_defunct_qpair(void *ctx) 3179 { 3180 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 3181 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3182 struct spdk_nvmf_rdma_transport, transport); 3183 3184 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3185 spdk_nvmf_rdma_qpair_destroy(rqpair); 3186 3187 return 0; 3188 } 3189 3190 static void 3191 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3192 { 3193 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3194 3195 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 3196 return; 3197 } 3198 3199 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 3200 3201 /* This happens only when the qpair is disconnected before 3202 * it is added to the poll group. Since there is no poll group, 3203 * the RDMA qp has not been initialized yet and the RDMA CM 3204 * event has not yet been acknowledged, so we need to reject it. 3205 */ 3206 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3207 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3208 return; 3209 } 3210 3211 if (rqpair->ibv_state != IBV_QPS_ERR) { 3212 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 3213 } 3214 3215 rqpair->destruct_poller = spdk_poller_register(spdk_nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 3216 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 3217 } 3218 3219 static struct spdk_nvmf_rdma_qpair * 3220 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3221 { 3222 struct spdk_nvmf_rdma_qpair *rqpair; 3223 /* @todo: improve QP search */ 3224 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3225 if (wc->qp_num == rqpair->cm_id->qp->qp_num) { 3226 return rqpair; 3227 } 3228 } 3229 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3230 return NULL; 3231 } 3232 3233 #ifdef DEBUG 3234 static int 3235 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3236 { 3237 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3238 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3239 } 3240 #endif 3241 3242 static void 3243 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3244 int rc) 3245 { 3246 struct spdk_nvmf_rdma_recv *rdma_recv; 3247 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3248 3249 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3250 while (bad_recv_wr != NULL) { 3251 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3252 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3253 3254 rdma_recv->qpair->current_recv_depth++; 3255 bad_recv_wr = bad_recv_wr->next; 3256 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3257 spdk_nvmf_rdma_start_disconnect(rdma_recv->qpair); 3258 } 3259 } 3260 3261 static void 3262 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3263 { 3264 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3265 while (bad_recv_wr != NULL) { 3266 bad_recv_wr = bad_recv_wr->next; 3267 rqpair->current_recv_depth++; 3268 } 3269 spdk_nvmf_rdma_start_disconnect(rqpair); 3270 } 3271 3272 static void 3273 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3274 struct spdk_nvmf_rdma_poller *rpoller) 3275 { 3276 struct spdk_nvmf_rdma_qpair *rqpair; 3277 struct ibv_recv_wr *bad_recv_wr; 3278 int rc; 3279 3280 if (rpoller->srq) { 3281 if (rpoller->resources->recvs_to_post.first != NULL) { 3282 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3283 if (rc) { 3284 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3285 } 3286 rpoller->resources->recvs_to_post.first = NULL; 3287 rpoller->resources->recvs_to_post.last = NULL; 3288 } 3289 } else { 3290 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3291 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3292 assert(rqpair->resources->recvs_to_post.first != NULL); 3293 rc = ibv_post_recv(rqpair->cm_id->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3294 if (rc) { 3295 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3296 } 3297 rqpair->resources->recvs_to_post.first = NULL; 3298 rqpair->resources->recvs_to_post.last = NULL; 3299 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3300 } 3301 } 3302 } 3303 3304 static void 3305 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3306 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3307 { 3308 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3309 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3310 3311 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3312 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3313 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3314 assert(rqpair->current_send_depth > 0); 3315 rqpair->current_send_depth--; 3316 switch (bad_rdma_wr->type) { 3317 case RDMA_WR_TYPE_DATA: 3318 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3319 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3320 assert(rqpair->current_read_depth > 0); 3321 rqpair->current_read_depth--; 3322 } 3323 break; 3324 case RDMA_WR_TYPE_SEND: 3325 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3326 break; 3327 default: 3328 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3329 prev_rdma_req = cur_rdma_req; 3330 continue; 3331 } 3332 3333 if (prev_rdma_req == cur_rdma_req) { 3334 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3335 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3336 continue; 3337 } 3338 3339 switch (cur_rdma_req->state) { 3340 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3341 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3342 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3343 break; 3344 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3345 case RDMA_REQUEST_STATE_COMPLETING: 3346 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3347 break; 3348 default: 3349 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3350 cur_rdma_req->state, rqpair); 3351 continue; 3352 } 3353 3354 spdk_nvmf_rdma_request_process(rtransport, cur_rdma_req); 3355 prev_rdma_req = cur_rdma_req; 3356 } 3357 3358 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3359 /* Disconnect the connection. */ 3360 spdk_nvmf_rdma_start_disconnect(rqpair); 3361 } 3362 3363 } 3364 3365 static void 3366 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3367 struct spdk_nvmf_rdma_poller *rpoller) 3368 { 3369 struct spdk_nvmf_rdma_qpair *rqpair; 3370 struct ibv_send_wr *bad_wr = NULL; 3371 int rc; 3372 3373 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3374 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3375 assert(rqpair->sends_to_post.first != NULL); 3376 rc = ibv_post_send(rqpair->cm_id->qp, rqpair->sends_to_post.first, &bad_wr); 3377 3378 /* bad wr always points to the first wr that failed. */ 3379 if (rc) { 3380 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3381 } 3382 rqpair->sends_to_post.first = NULL; 3383 rqpair->sends_to_post.last = NULL; 3384 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3385 } 3386 } 3387 3388 static int 3389 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3390 struct spdk_nvmf_rdma_poller *rpoller) 3391 { 3392 struct ibv_wc wc[32]; 3393 struct spdk_nvmf_rdma_wr *rdma_wr; 3394 struct spdk_nvmf_rdma_request *rdma_req; 3395 struct spdk_nvmf_rdma_recv *rdma_recv; 3396 struct spdk_nvmf_rdma_qpair *rqpair; 3397 int reaped, i; 3398 int count = 0; 3399 bool error = false; 3400 3401 /* Poll for completing operations. */ 3402 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3403 if (reaped < 0) { 3404 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3405 errno, spdk_strerror(errno)); 3406 return -1; 3407 } 3408 3409 for (i = 0; i < reaped; i++) { 3410 3411 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3412 3413 switch (rdma_wr->type) { 3414 case RDMA_WR_TYPE_SEND: 3415 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3416 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3417 3418 if (!wc[i].status) { 3419 count++; 3420 assert(wc[i].opcode == IBV_WC_SEND); 3421 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 3422 } else { 3423 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3424 } 3425 3426 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3427 rqpair->current_send_depth--; 3428 3429 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3430 assert(rdma_req->num_outstanding_data_wr == 0); 3431 break; 3432 case RDMA_WR_TYPE_RECV: 3433 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3434 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3435 if (rpoller->srq != NULL) { 3436 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3437 } 3438 rqpair = rdma_recv->qpair; 3439 3440 assert(rqpair != NULL); 3441 if (!wc[i].status) { 3442 assert(wc[i].opcode == IBV_WC_RECV); 3443 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3444 spdk_nvmf_rdma_start_disconnect(rqpair); 3445 break; 3446 } 3447 } 3448 3449 rdma_recv->wr.next = NULL; 3450 rqpair->current_recv_depth++; 3451 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3452 break; 3453 case RDMA_WR_TYPE_DATA: 3454 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3455 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3456 3457 assert(rdma_req->num_outstanding_data_wr > 0); 3458 3459 rqpair->current_send_depth--; 3460 rdma_req->num_outstanding_data_wr--; 3461 if (!wc[i].status) { 3462 if (wc[i].opcode == IBV_WC_RDMA_READ) { 3463 rqpair->current_read_depth--; 3464 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3465 if (rdma_req->num_outstanding_data_wr == 0) { 3466 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3467 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3468 } 3469 } else { 3470 assert(wc[i].opcode == IBV_WC_RDMA_WRITE); 3471 } 3472 } else { 3473 /* If the data transfer fails still force the queue into the error state, 3474 * if we were performing an RDMA_READ, we need to force the request into a 3475 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3476 * case, we should wait for the SEND to complete. */ 3477 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3478 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3479 rqpair->current_read_depth--; 3480 if (rdma_req->num_outstanding_data_wr == 0) { 3481 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3482 } 3483 } 3484 } 3485 break; 3486 default: 3487 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3488 continue; 3489 } 3490 3491 /* Handle error conditions */ 3492 if (wc[i].status) { 3493 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 3494 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 3495 3496 error = true; 3497 3498 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3499 /* Disconnect the connection. */ 3500 spdk_nvmf_rdma_start_disconnect(rqpair); 3501 } else { 3502 nvmf_rdma_destroy_drained_qpair(rqpair); 3503 } 3504 continue; 3505 } 3506 3507 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3508 3509 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3510 nvmf_rdma_destroy_drained_qpair(rqpair); 3511 } 3512 } 3513 3514 if (error == true) { 3515 return -1; 3516 } 3517 3518 /* submit outstanding work requests. */ 3519 _poller_submit_recvs(rtransport, rpoller); 3520 _poller_submit_sends(rtransport, rpoller); 3521 3522 return count; 3523 } 3524 3525 static int 3526 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3527 { 3528 struct spdk_nvmf_rdma_transport *rtransport; 3529 struct spdk_nvmf_rdma_poll_group *rgroup; 3530 struct spdk_nvmf_rdma_poller *rpoller; 3531 int count, rc; 3532 3533 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3534 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3535 3536 count = 0; 3537 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3538 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 3539 if (rc < 0) { 3540 return rc; 3541 } 3542 count += rc; 3543 } 3544 3545 return count; 3546 } 3547 3548 static int 3549 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3550 struct spdk_nvme_transport_id *trid, 3551 bool peer) 3552 { 3553 struct sockaddr *saddr; 3554 uint16_t port; 3555 3556 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 3557 3558 if (peer) { 3559 saddr = rdma_get_peer_addr(id); 3560 } else { 3561 saddr = rdma_get_local_addr(id); 3562 } 3563 switch (saddr->sa_family) { 3564 case AF_INET: { 3565 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3566 3567 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3568 inet_ntop(AF_INET, &saddr_in->sin_addr, 3569 trid->traddr, sizeof(trid->traddr)); 3570 if (peer) { 3571 port = ntohs(rdma_get_dst_port(id)); 3572 } else { 3573 port = ntohs(rdma_get_src_port(id)); 3574 } 3575 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3576 break; 3577 } 3578 case AF_INET6: { 3579 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3580 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3581 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3582 trid->traddr, sizeof(trid->traddr)); 3583 if (peer) { 3584 port = ntohs(rdma_get_dst_port(id)); 3585 } else { 3586 port = ntohs(rdma_get_src_port(id)); 3587 } 3588 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3589 break; 3590 } 3591 default: 3592 return -1; 3593 3594 } 3595 3596 return 0; 3597 } 3598 3599 static int 3600 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3601 struct spdk_nvme_transport_id *trid) 3602 { 3603 struct spdk_nvmf_rdma_qpair *rqpair; 3604 3605 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3606 3607 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 3608 } 3609 3610 static int 3611 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3612 struct spdk_nvme_transport_id *trid) 3613 { 3614 struct spdk_nvmf_rdma_qpair *rqpair; 3615 3616 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3617 3618 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 3619 } 3620 3621 static int 3622 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3623 struct spdk_nvme_transport_id *trid) 3624 { 3625 struct spdk_nvmf_rdma_qpair *rqpair; 3626 3627 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3628 3629 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 3630 } 3631 3632 void 3633 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3634 { 3635 g_nvmf_hooks = *hooks; 3636 } 3637 3638 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 3639 .type = SPDK_NVME_TRANSPORT_RDMA, 3640 .opts_init = spdk_nvmf_rdma_opts_init, 3641 .create = spdk_nvmf_rdma_create, 3642 .destroy = spdk_nvmf_rdma_destroy, 3643 3644 .listen = spdk_nvmf_rdma_listen, 3645 .stop_listen = spdk_nvmf_rdma_stop_listen, 3646 .accept = spdk_nvmf_rdma_accept, 3647 3648 .listener_discover = spdk_nvmf_rdma_discover, 3649 3650 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 3651 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 3652 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 3653 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 3654 3655 .req_free = spdk_nvmf_rdma_request_free, 3656 .req_complete = spdk_nvmf_rdma_request_complete, 3657 3658 .qpair_fini = spdk_nvmf_rdma_close_qpair, 3659 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 3660 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 3661 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 3662 3663 }; 3664 3665 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 3666