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