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 /* AIO backend requires block size aligned data buffers, 65 * extra 4KiB aligned data buffer should work for most devices. 66 */ 67 #define SHIFT_4KB 12 68 #define NVMF_DATA_BUFFER_ALIGNMENT (1 << SHIFT_4KB) 69 #define NVMF_DATA_BUFFER_MASK (NVMF_DATA_BUFFER_ALIGNMENT - 1) 70 71 enum spdk_nvmf_rdma_request_state { 72 /* The request is not currently in use */ 73 RDMA_REQUEST_STATE_FREE = 0, 74 75 /* Initial state when request first received */ 76 RDMA_REQUEST_STATE_NEW, 77 78 /* The request is queued until a data buffer is available. */ 79 RDMA_REQUEST_STATE_NEED_BUFFER, 80 81 /* The request is waiting on RDMA queue depth availability 82 * to transfer data between the host and the controller. 83 */ 84 RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 85 86 /* The request is currently transferring data from the host to the controller. */ 87 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 88 89 /* The request is ready to execute at the block device */ 90 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 91 92 /* The request is currently executing at the block device */ 93 RDMA_REQUEST_STATE_EXECUTING, 94 95 /* The request finished executing at the block device */ 96 RDMA_REQUEST_STATE_EXECUTED, 97 98 /* The request is ready to send a completion */ 99 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 100 101 /* The request is currently transferring data from the controller to the host. */ 102 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 103 104 /* The request currently has an outstanding completion without an 105 * associated data transfer. 106 */ 107 RDMA_REQUEST_STATE_COMPLETING, 108 109 /* The request completed and can be marked free. */ 110 RDMA_REQUEST_STATE_COMPLETED, 111 112 /* Terminator */ 113 RDMA_REQUEST_NUM_STATES, 114 }; 115 116 #define OBJECT_NVMF_RDMA_IO 0x40 117 118 #define TRACE_GROUP_NVMF_RDMA 0x4 119 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 120 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 121 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 122 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 123 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 124 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 125 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 126 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 127 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 128 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 129 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 130 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 131 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 132 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 133 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 134 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 135 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 136 137 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 138 { 139 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 140 spdk_trace_register_description("RDMA_REQ_NEW", "", 141 TRACE_RDMA_REQUEST_STATE_NEW, 142 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 143 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", "", 144 TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 145 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 146 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H_TO_C", "", 147 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 148 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 149 spdk_trace_register_description("RDMA_REQ_TX_H_TO_C", "", 150 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 151 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 152 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", "", 153 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 154 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 155 spdk_trace_register_description("RDMA_REQ_EXECUTING", "", 156 TRACE_RDMA_REQUEST_STATE_EXECUTING, 157 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 158 spdk_trace_register_description("RDMA_REQ_EXECUTED", "", 159 TRACE_RDMA_REQUEST_STATE_EXECUTED, 160 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 161 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPLETE", "", 162 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 163 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 164 spdk_trace_register_description("RDMA_REQ_COMPLETING_CONTROLLER_TO_HOST", "", 165 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 166 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 167 spdk_trace_register_description("RDMA_REQ_COMPLETING_INCAPSULE", "", 168 TRACE_RDMA_REQUEST_STATE_COMPLETING, 169 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 170 spdk_trace_register_description("RDMA_REQ_COMPLETED", "", 171 TRACE_RDMA_REQUEST_STATE_COMPLETED, 172 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 173 174 spdk_trace_register_description("RDMA_QP_CREATE", "", TRACE_RDMA_QP_CREATE, 175 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 176 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", "", TRACE_RDMA_IBV_ASYNC_EVENT, 177 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 178 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", "", TRACE_RDMA_CM_ASYNC_EVENT, 179 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 180 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", "", TRACE_RDMA_QP_STATE_CHANGE, 181 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 182 spdk_trace_register_description("RDMA_QP_DISCONNECT", "", TRACE_RDMA_QP_DISCONNECT, 183 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 184 spdk_trace_register_description("RDMA_QP_DESTROY", "", TRACE_RDMA_QP_DESTROY, 185 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 186 } 187 188 enum spdk_nvmf_rdma_wr_type { 189 RDMA_WR_TYPE_RECV, 190 RDMA_WR_TYPE_SEND, 191 RDMA_WR_TYPE_DATA, 192 RDMA_WR_TYPE_DRAIN_SEND, 193 RDMA_WR_TYPE_DRAIN_RECV 194 }; 195 196 struct spdk_nvmf_rdma_wr { 197 enum spdk_nvmf_rdma_wr_type type; 198 }; 199 200 /* This structure holds commands as they are received off the wire. 201 * It must be dynamically paired with a full request object 202 * (spdk_nvmf_rdma_request) to service a request. It is separate 203 * from the request because RDMA does not appear to order 204 * completions, so occasionally we'll get a new incoming 205 * command when there aren't any free request objects. 206 */ 207 struct spdk_nvmf_rdma_recv { 208 struct ibv_recv_wr wr; 209 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 210 211 struct spdk_nvmf_rdma_qpair *qpair; 212 213 /* In-capsule data buffer */ 214 uint8_t *buf; 215 216 struct spdk_nvmf_rdma_wr rdma_wr; 217 218 TAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 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 { 236 struct spdk_nvmf_rdma_wr rdma_wr; 237 struct ibv_send_wr wr; 238 struct ibv_sge sgl[NVMF_DEFAULT_TX_SGE]; 239 void *buffers[NVMF_DEFAULT_TX_SGE]; 240 } data; 241 242 struct spdk_nvmf_rdma_wr rdma_wr; 243 244 TAILQ_ENTRY(spdk_nvmf_rdma_request) link; 245 TAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 246 }; 247 248 enum spdk_nvmf_rdma_qpair_disconnect_flags { 249 RDMA_QP_DISCONNECTING = 1, 250 RDMA_QP_RECV_DRAINED = 1 << 1, 251 RDMA_QP_SEND_DRAINED = 1 << 2 252 }; 253 254 struct spdk_nvmf_rdma_qpair { 255 struct spdk_nvmf_qpair qpair; 256 257 struct spdk_nvmf_rdma_port *port; 258 struct spdk_nvmf_rdma_poller *poller; 259 260 struct rdma_cm_id *cm_id; 261 struct rdma_cm_id *listen_id; 262 263 /* The maximum number of I/O outstanding on this connection at one time */ 264 uint16_t max_queue_depth; 265 266 /* The maximum number of active RDMA READ and WRITE operations at one time */ 267 uint16_t max_rw_depth; 268 269 /* The maximum number of SGEs per WR on the send queue */ 270 uint32_t max_send_sge; 271 272 /* The maximum number of SGEs per WR on the recv queue */ 273 uint32_t max_recv_sge; 274 275 /* Receives that are waiting for a request object */ 276 TAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 277 278 /* Queues to track the requests in all states */ 279 TAILQ_HEAD(, spdk_nvmf_rdma_request) state_queue[RDMA_REQUEST_NUM_STATES]; 280 281 /* Number of requests in each state */ 282 uint32_t state_cntr[RDMA_REQUEST_NUM_STATES]; 283 284 /* Array of size "max_queue_depth" containing RDMA requests. */ 285 struct spdk_nvmf_rdma_request *reqs; 286 287 /* Array of size "max_queue_depth" containing RDMA recvs. */ 288 struct spdk_nvmf_rdma_recv *recvs; 289 290 /* Array of size "max_queue_depth" containing 64 byte capsules 291 * used for receive. 292 */ 293 union nvmf_h2c_msg *cmds; 294 struct ibv_mr *cmds_mr; 295 296 /* Array of size "max_queue_depth" containing 16 byte completions 297 * to be sent back to the user. 298 */ 299 union nvmf_c2h_msg *cpls; 300 struct ibv_mr *cpls_mr; 301 302 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 303 * buffers to be used for in capsule data. 304 */ 305 void *bufs; 306 struct ibv_mr *bufs_mr; 307 308 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 309 310 /* Mgmt channel */ 311 struct spdk_io_channel *mgmt_channel; 312 struct spdk_nvmf_rdma_mgmt_channel *ch; 313 314 /* IBV queue pair attributes: they are used to manage 315 * qp state and recover from errors. 316 */ 317 struct ibv_qp_attr ibv_attr; 318 319 uint32_t disconnect_flags; 320 struct spdk_nvmf_rdma_wr drain_send_wr; 321 struct spdk_nvmf_rdma_wr drain_recv_wr; 322 323 /* Reference counter for how many unprocessed messages 324 * from other threads are currently outstanding. The 325 * qpair cannot be destroyed until this is 0. This is 326 * atomically incremented from any thread, but only 327 * decremented and read from the thread that owns this 328 * qpair. 329 */ 330 uint32_t refcnt; 331 }; 332 333 struct spdk_nvmf_rdma_poller { 334 struct spdk_nvmf_rdma_device *device; 335 struct spdk_nvmf_rdma_poll_group *group; 336 337 struct ibv_cq *cq; 338 339 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 340 341 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 342 }; 343 344 struct spdk_nvmf_rdma_poll_group { 345 struct spdk_nvmf_transport_poll_group group; 346 347 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 348 }; 349 350 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 351 struct spdk_nvmf_rdma_device { 352 struct ibv_device_attr attr; 353 struct ibv_context *context; 354 355 struct spdk_mem_map *map; 356 struct ibv_pd *pd; 357 358 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 359 }; 360 361 struct spdk_nvmf_rdma_port { 362 struct spdk_nvme_transport_id trid; 363 struct rdma_cm_id *id; 364 struct spdk_nvmf_rdma_device *device; 365 uint32_t ref; 366 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 367 }; 368 369 struct spdk_nvmf_rdma_transport { 370 struct spdk_nvmf_transport transport; 371 372 struct rdma_event_channel *event_channel; 373 374 struct spdk_mempool *data_buf_pool; 375 376 pthread_mutex_t lock; 377 378 /* fields used to poll RDMA/IB events */ 379 nfds_t npoll_fds; 380 struct pollfd *poll_fds; 381 382 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 383 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 384 }; 385 386 struct spdk_nvmf_rdma_mgmt_channel { 387 /* Requests that are waiting to obtain a data buffer */ 388 TAILQ_HEAD(, spdk_nvmf_rdma_request) pending_data_buf_queue; 389 }; 390 391 static inline void 392 spdk_nvmf_rdma_qpair_inc_refcnt(struct spdk_nvmf_rdma_qpair *rqpair) 393 { 394 __sync_fetch_and_add(&rqpair->refcnt, 1); 395 } 396 397 static inline uint32_t 398 spdk_nvmf_rdma_qpair_dec_refcnt(struct spdk_nvmf_rdma_qpair *rqpair) 399 { 400 uint32_t old_refcnt, new_refcnt; 401 402 do { 403 old_refcnt = rqpair->refcnt; 404 assert(old_refcnt > 0); 405 new_refcnt = old_refcnt - 1; 406 } while (__sync_bool_compare_and_swap(&rqpair->refcnt, old_refcnt, new_refcnt) == false); 407 408 return new_refcnt; 409 } 410 411 static enum ibv_qp_state 412 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 413 enum ibv_qp_state old_state, new_state; 414 struct ibv_qp_init_attr init_attr; 415 int rc; 416 417 /* All the attributes needed for recovery */ 418 static int spdk_nvmf_ibv_attr_mask = 419 IBV_QP_STATE | 420 IBV_QP_PKEY_INDEX | 421 IBV_QP_PORT | 422 IBV_QP_ACCESS_FLAGS | 423 IBV_QP_AV | 424 IBV_QP_PATH_MTU | 425 IBV_QP_DEST_QPN | 426 IBV_QP_RQ_PSN | 427 IBV_QP_MAX_DEST_RD_ATOMIC | 428 IBV_QP_MIN_RNR_TIMER | 429 IBV_QP_SQ_PSN | 430 IBV_QP_TIMEOUT | 431 IBV_QP_RETRY_CNT | 432 IBV_QP_RNR_RETRY | 433 IBV_QP_MAX_QP_RD_ATOMIC; 434 435 old_state = rqpair->ibv_attr.qp_state; 436 rc = ibv_query_qp(rqpair->cm_id->qp, &rqpair->ibv_attr, 437 spdk_nvmf_ibv_attr_mask, &init_attr); 438 439 if (rc) 440 { 441 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 442 assert(false); 443 } 444 445 new_state = rqpair->ibv_attr.qp_state; 446 if (old_state != new_state) 447 { 448 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 449 (uintptr_t)rqpair->cm_id, new_state); 450 } 451 return new_state; 452 } 453 454 static const char *str_ibv_qp_state[] = { 455 "IBV_QPS_RESET", 456 "IBV_QPS_INIT", 457 "IBV_QPS_RTR", 458 "IBV_QPS_RTS", 459 "IBV_QPS_SQD", 460 "IBV_QPS_SQE", 461 "IBV_QPS_ERR" 462 }; 463 464 static int 465 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 466 enum ibv_qp_state new_state) 467 { 468 int rc; 469 enum ibv_qp_state state; 470 static int attr_mask_rc[] = { 471 [IBV_QPS_RESET] = IBV_QP_STATE, 472 [IBV_QPS_INIT] = (IBV_QP_STATE | 473 IBV_QP_PKEY_INDEX | 474 IBV_QP_PORT | 475 IBV_QP_ACCESS_FLAGS), 476 [IBV_QPS_RTR] = (IBV_QP_STATE | 477 IBV_QP_AV | 478 IBV_QP_PATH_MTU | 479 IBV_QP_DEST_QPN | 480 IBV_QP_RQ_PSN | 481 IBV_QP_MAX_DEST_RD_ATOMIC | 482 IBV_QP_MIN_RNR_TIMER), 483 [IBV_QPS_RTS] = (IBV_QP_STATE | 484 IBV_QP_SQ_PSN | 485 IBV_QP_TIMEOUT | 486 IBV_QP_RETRY_CNT | 487 IBV_QP_RNR_RETRY | 488 IBV_QP_MAX_QP_RD_ATOMIC), 489 [IBV_QPS_SQD] = IBV_QP_STATE, 490 [IBV_QPS_SQE] = IBV_QP_STATE, 491 [IBV_QPS_ERR] = IBV_QP_STATE, 492 }; 493 494 switch (new_state) { 495 case IBV_QPS_RESET: 496 case IBV_QPS_INIT: 497 case IBV_QPS_RTR: 498 case IBV_QPS_RTS: 499 case IBV_QPS_SQD: 500 case IBV_QPS_SQE: 501 case IBV_QPS_ERR: 502 break; 503 default: 504 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 505 rqpair->qpair.qid, new_state); 506 return -1; 507 } 508 rqpair->ibv_attr.cur_qp_state = rqpair->ibv_attr.qp_state; 509 rqpair->ibv_attr.qp_state = new_state; 510 rqpair->ibv_attr.ah_attr.port_num = rqpair->ibv_attr.port_num; 511 512 rc = ibv_modify_qp(rqpair->cm_id->qp, &rqpair->ibv_attr, 513 attr_mask_rc[new_state]); 514 515 if (rc) { 516 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 517 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 518 return rc; 519 } 520 521 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 522 523 if (state != new_state) { 524 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 525 rqpair->qpair.qid, str_ibv_qp_state[new_state], 526 str_ibv_qp_state[state]); 527 return -1; 528 } 529 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 530 str_ibv_qp_state[state]); 531 return 0; 532 } 533 534 static void 535 spdk_nvmf_rdma_request_set_state(struct spdk_nvmf_rdma_request *rdma_req, 536 enum spdk_nvmf_rdma_request_state state) 537 { 538 struct spdk_nvmf_qpair *qpair; 539 struct spdk_nvmf_rdma_qpair *rqpair; 540 541 qpair = rdma_req->req.qpair; 542 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 543 544 TAILQ_REMOVE(&rqpair->state_queue[rdma_req->state], rdma_req, state_link); 545 rqpair->state_cntr[rdma_req->state]--; 546 547 rdma_req->state = state; 548 549 TAILQ_INSERT_TAIL(&rqpair->state_queue[rdma_req->state], rdma_req, state_link); 550 rqpair->state_cntr[rdma_req->state]++; 551 } 552 553 static int 554 spdk_nvmf_rdma_mgmt_channel_create(void *io_device, void *ctx_buf) 555 { 556 struct spdk_nvmf_rdma_mgmt_channel *ch = ctx_buf; 557 558 TAILQ_INIT(&ch->pending_data_buf_queue); 559 return 0; 560 } 561 562 static void 563 spdk_nvmf_rdma_mgmt_channel_destroy(void *io_device, void *ctx_buf) 564 { 565 struct spdk_nvmf_rdma_mgmt_channel *ch = ctx_buf; 566 567 if (!TAILQ_EMPTY(&ch->pending_data_buf_queue)) { 568 SPDK_ERRLOG("Pending I/O list wasn't empty on channel destruction\n"); 569 } 570 } 571 572 static int 573 spdk_nvmf_rdma_cur_rw_depth(struct spdk_nvmf_rdma_qpair *rqpair) 574 { 575 return rqpair->state_cntr[RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER] + 576 rqpair->state_cntr[RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST]; 577 } 578 579 static int 580 spdk_nvmf_rdma_cur_queue_depth(struct spdk_nvmf_rdma_qpair *rqpair) 581 { 582 return rqpair->max_queue_depth - 583 rqpair->state_cntr[RDMA_REQUEST_STATE_FREE]; 584 } 585 586 static void 587 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 588 { 589 int qd; 590 591 if (rqpair->refcnt == 0) { 592 return; 593 } 594 595 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 596 597 qd = spdk_nvmf_rdma_cur_queue_depth(rqpair); 598 if (qd != 0) { 599 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", qd); 600 } 601 602 if (rqpair->poller) { 603 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 604 } 605 606 if (rqpair->cmds_mr) { 607 ibv_dereg_mr(rqpair->cmds_mr); 608 } 609 610 if (rqpair->cpls_mr) { 611 ibv_dereg_mr(rqpair->cpls_mr); 612 } 613 614 if (rqpair->bufs_mr) { 615 ibv_dereg_mr(rqpair->bufs_mr); 616 } 617 618 if (rqpair->cm_id) { 619 rdma_destroy_qp(rqpair->cm_id); 620 rdma_destroy_id(rqpair->cm_id); 621 } 622 623 if (rqpair->mgmt_channel) { 624 spdk_put_io_channel(rqpair->mgmt_channel); 625 } 626 627 /* Free all memory */ 628 spdk_dma_free(rqpair->cmds); 629 spdk_dma_free(rqpair->cpls); 630 spdk_dma_free(rqpair->bufs); 631 free(rqpair->reqs); 632 free(rqpair->recvs); 633 free(rqpair); 634 } 635 636 static int 637 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 638 { 639 struct spdk_nvmf_rdma_transport *rtransport; 640 struct spdk_nvmf_rdma_qpair *rqpair; 641 int rc, i; 642 struct spdk_nvmf_rdma_recv *rdma_recv; 643 struct spdk_nvmf_rdma_request *rdma_req; 644 struct spdk_nvmf_transport *transport; 645 struct spdk_nvmf_rdma_device *device; 646 struct ibv_qp_init_attr ibv_init_attr; 647 648 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 649 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 650 transport = &rtransport->transport; 651 device = rqpair->port->device; 652 653 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 654 ibv_init_attr.qp_context = rqpair; 655 ibv_init_attr.qp_type = IBV_QPT_RC; 656 ibv_init_attr.send_cq = rqpair->poller->cq; 657 ibv_init_attr.recv_cq = rqpair->poller->cq; 658 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 659 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 660 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 661 1; /* RECV operations + dummy drain WR */ 662 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 663 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 664 665 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 666 if (rc) { 667 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 668 rdma_destroy_id(rqpair->cm_id); 669 rqpair->cm_id = NULL; 670 spdk_nvmf_rdma_qpair_destroy(rqpair); 671 return -1; 672 } 673 674 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 675 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 676 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 677 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 678 679 rqpair->reqs = calloc(rqpair->max_queue_depth, sizeof(*rqpair->reqs)); 680 rqpair->recvs = calloc(rqpair->max_queue_depth, sizeof(*rqpair->recvs)); 681 rqpair->cmds = spdk_dma_zmalloc(rqpair->max_queue_depth * sizeof(*rqpair->cmds), 682 0x1000, NULL); 683 rqpair->cpls = spdk_dma_zmalloc(rqpair->max_queue_depth * sizeof(*rqpair->cpls), 684 0x1000, NULL); 685 686 687 if (transport->opts.in_capsule_data_size > 0) { 688 rqpair->bufs = spdk_dma_zmalloc(rqpair->max_queue_depth * 689 transport->opts.in_capsule_data_size, 690 0x1000, NULL); 691 } 692 693 if (!rqpair->reqs || !rqpair->recvs || !rqpair->cmds || 694 !rqpair->cpls || (transport->opts.in_capsule_data_size && !rqpair->bufs)) { 695 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 696 spdk_nvmf_rdma_qpair_destroy(rqpair); 697 return -1; 698 } 699 700 rqpair->cmds_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->cmds, 701 rqpair->max_queue_depth * sizeof(*rqpair->cmds), 702 IBV_ACCESS_LOCAL_WRITE); 703 rqpair->cpls_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->cpls, 704 rqpair->max_queue_depth * sizeof(*rqpair->cpls), 705 0); 706 707 if (transport->opts.in_capsule_data_size) { 708 rqpair->bufs_mr = ibv_reg_mr(rqpair->cm_id->pd, rqpair->bufs, 709 rqpair->max_queue_depth * 710 transport->opts.in_capsule_data_size, 711 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 712 } 713 714 if (!rqpair->cmds_mr || !rqpair->cpls_mr || (transport->opts.in_capsule_data_size && 715 !rqpair->bufs_mr)) { 716 SPDK_ERRLOG("Unable to register required memory for RDMA queue.\n"); 717 spdk_nvmf_rdma_qpair_destroy(rqpair); 718 return -1; 719 } 720 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 721 rqpair->cmds, rqpair->max_queue_depth * sizeof(*rqpair->cmds), rqpair->cmds_mr->lkey); 722 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 723 rqpair->cpls, rqpair->max_queue_depth * sizeof(*rqpair->cpls), rqpair->cpls_mr->lkey); 724 if (rqpair->bufs && rqpair->bufs_mr) { 725 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 726 rqpair->bufs, rqpair->max_queue_depth * 727 transport->opts.in_capsule_data_size, rqpair->bufs_mr->lkey); 728 } 729 730 /* Initialise request state queues and counters of the queue pair */ 731 for (i = RDMA_REQUEST_STATE_FREE; i < RDMA_REQUEST_NUM_STATES; i++) { 732 TAILQ_INIT(&rqpair->state_queue[i]); 733 rqpair->state_cntr[i] = 0; 734 } 735 736 for (i = 0; i < rqpair->max_queue_depth; i++) { 737 struct ibv_recv_wr *bad_wr = NULL; 738 739 rdma_recv = &rqpair->recvs[i]; 740 rdma_recv->qpair = rqpair; 741 742 /* Set up memory to receive commands */ 743 if (rqpair->bufs) { 744 rdma_recv->buf = (void *)((uintptr_t)rqpair->bufs + (i * 745 transport->opts.in_capsule_data_size)); 746 } 747 748 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 749 750 rdma_recv->sgl[0].addr = (uintptr_t)&rqpair->cmds[i]; 751 rdma_recv->sgl[0].length = sizeof(rqpair->cmds[i]); 752 rdma_recv->sgl[0].lkey = rqpair->cmds_mr->lkey; 753 rdma_recv->wr.num_sge = 1; 754 755 if (rdma_recv->buf && rqpair->bufs_mr) { 756 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 757 rdma_recv->sgl[1].length = transport->opts.in_capsule_data_size; 758 rdma_recv->sgl[1].lkey = rqpair->bufs_mr->lkey; 759 rdma_recv->wr.num_sge++; 760 } 761 762 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 763 rdma_recv->wr.sg_list = rdma_recv->sgl; 764 765 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_recv->wr, &bad_wr); 766 if (rc) { 767 SPDK_ERRLOG("Unable to post capsule for RDMA RECV\n"); 768 spdk_nvmf_rdma_qpair_destroy(rqpair); 769 return -1; 770 } 771 } 772 773 for (i = 0; i < rqpair->max_queue_depth; i++) { 774 rdma_req = &rqpair->reqs[i]; 775 776 rdma_req->req.qpair = &rqpair->qpair; 777 rdma_req->req.cmd = NULL; 778 779 /* Set up memory to send responses */ 780 rdma_req->req.rsp = &rqpair->cpls[i]; 781 782 rdma_req->rsp.sgl[0].addr = (uintptr_t)&rqpair->cpls[i]; 783 rdma_req->rsp.sgl[0].length = sizeof(rqpair->cpls[i]); 784 rdma_req->rsp.sgl[0].lkey = rqpair->cpls_mr->lkey; 785 786 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 787 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 788 rdma_req->rsp.wr.next = NULL; 789 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 790 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 791 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 792 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 793 794 /* Set up memory for data buffers */ 795 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 796 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 797 rdma_req->data.wr.next = NULL; 798 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 799 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 800 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 801 802 /* Initialize request state to FREE */ 803 rdma_req->state = RDMA_REQUEST_STATE_FREE; 804 TAILQ_INSERT_TAIL(&rqpair->state_queue[rdma_req->state], rdma_req, state_link); 805 rqpair->state_cntr[rdma_req->state]++; 806 } 807 808 return 0; 809 } 810 811 static int 812 request_transfer_in(struct spdk_nvmf_request *req) 813 { 814 int rc; 815 struct spdk_nvmf_rdma_request *rdma_req; 816 struct spdk_nvmf_qpair *qpair; 817 struct spdk_nvmf_rdma_qpair *rqpair; 818 struct ibv_send_wr *bad_wr = NULL; 819 820 qpair = req->qpair; 821 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 822 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 823 824 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 825 826 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA READ POSTED. Request: %p Connection: %p\n", req, qpair); 827 828 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 829 rdma_req->data.wr.next = NULL; 830 rc = ibv_post_send(rqpair->cm_id->qp, &rdma_req->data.wr, &bad_wr); 831 if (rc) { 832 SPDK_ERRLOG("Unable to transfer data from host to target\n"); 833 return -1; 834 } 835 return 0; 836 } 837 838 static int 839 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 840 { 841 int rc; 842 struct spdk_nvmf_rdma_request *rdma_req; 843 struct spdk_nvmf_qpair *qpair; 844 struct spdk_nvmf_rdma_qpair *rqpair; 845 struct spdk_nvme_cpl *rsp; 846 struct ibv_recv_wr *bad_recv_wr = NULL; 847 struct ibv_send_wr *send_wr, *bad_send_wr = NULL; 848 849 *data_posted = 0; 850 qpair = req->qpair; 851 rsp = &req->rsp->nvme_cpl; 852 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 853 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 854 855 /* Advance our sq_head pointer */ 856 if (qpair->sq_head == qpair->sq_head_max) { 857 qpair->sq_head = 0; 858 } else { 859 qpair->sq_head++; 860 } 861 rsp->sqhd = qpair->sq_head; 862 863 /* Post the capsule to the recv buffer */ 864 assert(rdma_req->recv != NULL); 865 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA RECV POSTED. Recv: %p Connection: %p\n", rdma_req->recv, 866 rqpair); 867 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_req->recv->wr, &bad_recv_wr); 868 if (rc) { 869 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 870 return rc; 871 } 872 rdma_req->recv = NULL; 873 874 /* Build the response which consists of an optional 875 * RDMA WRITE to transfer data, plus an RDMA SEND 876 * containing the response. 877 */ 878 send_wr = &rdma_req->rsp.wr; 879 880 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 881 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 882 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA WRITE POSTED. Request: %p Connection: %p\n", req, qpair); 883 884 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 885 886 rdma_req->data.wr.next = send_wr; 887 *data_posted = 1; 888 send_wr = &rdma_req->data.wr; 889 } 890 891 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA SEND POSTED. Request: %p Connection: %p\n", req, qpair); 892 893 /* Send the completion */ 894 rc = ibv_post_send(rqpair->cm_id->qp, send_wr, &bad_send_wr); 895 if (rc) { 896 SPDK_ERRLOG("Unable to send response capsule\n"); 897 } 898 899 return rc; 900 } 901 902 static int 903 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 904 { 905 struct spdk_nvmf_rdma_accept_private_data accept_data; 906 struct rdma_conn_param ctrlr_event_data = {}; 907 int rc; 908 909 accept_data.recfmt = 0; 910 accept_data.crqsize = rqpair->max_queue_depth; 911 912 ctrlr_event_data.private_data = &accept_data; 913 ctrlr_event_data.private_data_len = sizeof(accept_data); 914 if (id->ps == RDMA_PS_TCP) { 915 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 916 ctrlr_event_data.initiator_depth = rqpair->max_rw_depth; 917 } 918 919 rc = rdma_accept(id, &ctrlr_event_data); 920 if (rc) { 921 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 922 } else { 923 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 924 } 925 926 return rc; 927 } 928 929 static void 930 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 931 { 932 struct spdk_nvmf_rdma_reject_private_data rej_data; 933 934 rej_data.recfmt = 0; 935 rej_data.sts = error; 936 937 rdma_reject(id, &rej_data, sizeof(rej_data)); 938 } 939 940 static int 941 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 942 new_qpair_fn cb_fn) 943 { 944 struct spdk_nvmf_rdma_transport *rtransport; 945 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 946 struct spdk_nvmf_rdma_port *port; 947 struct rdma_conn_param *rdma_param = NULL; 948 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 949 uint16_t max_queue_depth; 950 uint16_t max_rw_depth; 951 952 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 953 954 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 955 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 956 957 rdma_param = &event->param.conn; 958 if (rdma_param->private_data == NULL || 959 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 960 SPDK_ERRLOG("connect request: no private data provided\n"); 961 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 962 return -1; 963 } 964 965 private_data = rdma_param->private_data; 966 if (private_data->recfmt != 0) { 967 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 968 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 969 return -1; 970 } 971 972 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 973 event->id->verbs->device->name, event->id->verbs->device->dev_name); 974 975 port = event->listen_id->context; 976 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 977 event->listen_id, event->listen_id->verbs, port); 978 979 /* Figure out the supported queue depth. This is a multi-step process 980 * that takes into account hardware maximums, host provided values, 981 * and our target's internal memory limits */ 982 983 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 984 985 /* Start with the maximum queue depth allowed by the target */ 986 max_queue_depth = rtransport->transport.opts.max_queue_depth; 987 max_rw_depth = rtransport->transport.opts.max_queue_depth; 988 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 989 rtransport->transport.opts.max_queue_depth); 990 991 /* Next check the local NIC's hardware limitations */ 992 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 993 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 994 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 995 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 996 max_rw_depth = spdk_min(max_rw_depth, port->device->attr.max_qp_rd_atom); 997 998 /* Next check the remote NIC's hardware limitations */ 999 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1000 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1001 rdma_param->initiator_depth, rdma_param->responder_resources); 1002 if (rdma_param->initiator_depth > 0) { 1003 max_rw_depth = spdk_min(max_rw_depth, rdma_param->initiator_depth); 1004 } 1005 1006 /* Finally check for the host software requested values, which are 1007 * optional. */ 1008 if (rdma_param->private_data != NULL && 1009 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1010 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1011 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1012 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1013 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1014 } 1015 1016 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1017 max_queue_depth, max_rw_depth); 1018 1019 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1020 if (rqpair == NULL) { 1021 SPDK_ERRLOG("Could not allocate new connection.\n"); 1022 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1023 return -1; 1024 } 1025 1026 rqpair->port = port; 1027 rqpair->max_queue_depth = max_queue_depth; 1028 rqpair->max_rw_depth = max_rw_depth; 1029 rqpair->cm_id = event->id; 1030 rqpair->listen_id = event->listen_id; 1031 rqpair->qpair.transport = transport; 1032 TAILQ_INIT(&rqpair->incoming_queue); 1033 event->id->context = &rqpair->qpair; 1034 1035 cb_fn(&rqpair->qpair); 1036 1037 return 0; 1038 } 1039 1040 static int 1041 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1042 enum spdk_mem_map_notify_action action, 1043 void *vaddr, size_t size) 1044 { 1045 struct spdk_nvmf_rdma_device *device = cb_ctx; 1046 struct ibv_pd *pd = device->pd; 1047 struct ibv_mr *mr; 1048 1049 switch (action) { 1050 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1051 mr = ibv_reg_mr(pd, vaddr, size, 1052 IBV_ACCESS_LOCAL_WRITE | 1053 IBV_ACCESS_REMOTE_READ | 1054 IBV_ACCESS_REMOTE_WRITE); 1055 if (mr == NULL) { 1056 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1057 return -1; 1058 } else { 1059 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1060 } 1061 break; 1062 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1063 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1064 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1065 if (mr) { 1066 ibv_dereg_mr(mr); 1067 } 1068 break; 1069 } 1070 1071 return 0; 1072 } 1073 1074 static int 1075 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1076 { 1077 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1078 return addr_1 == addr_2; 1079 } 1080 1081 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1082 1083 static spdk_nvme_data_transfer_t 1084 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1085 { 1086 enum spdk_nvme_data_transfer xfer; 1087 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1088 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1089 1090 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1091 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1092 rdma_req->rsp.wr.imm_data = 0; 1093 #endif 1094 1095 /* Figure out data transfer direction */ 1096 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1097 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1098 } else { 1099 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1100 1101 /* Some admin commands are special cases */ 1102 if ((rdma_req->req.qpair->qid == 0) && 1103 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1104 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1105 switch (cmd->cdw10 & 0xff) { 1106 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1107 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1108 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1109 break; 1110 default: 1111 xfer = SPDK_NVME_DATA_NONE; 1112 } 1113 } 1114 } 1115 1116 if (xfer == SPDK_NVME_DATA_NONE) { 1117 return xfer; 1118 } 1119 1120 /* Even for commands that may transfer data, they could have specified 0 length. 1121 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1122 */ 1123 switch (sgl->generic.type) { 1124 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1125 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1126 case SPDK_NVME_SGL_TYPE_SEGMENT: 1127 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1128 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1129 if (sgl->unkeyed.length == 0) { 1130 xfer = SPDK_NVME_DATA_NONE; 1131 } 1132 break; 1133 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1134 if (sgl->keyed.length == 0) { 1135 xfer = SPDK_NVME_DATA_NONE; 1136 } 1137 break; 1138 } 1139 1140 return xfer; 1141 } 1142 1143 static int 1144 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1145 struct spdk_nvmf_rdma_device *device, 1146 struct spdk_nvmf_rdma_request *rdma_req) 1147 { 1148 void *buf = NULL; 1149 uint32_t length = rdma_req->req.length; 1150 uint32_t i = 0; 1151 1152 rdma_req->req.iovcnt = 0; 1153 while (length) { 1154 buf = spdk_mempool_get(rtransport->data_buf_pool); 1155 if (!buf) { 1156 goto nomem; 1157 } 1158 1159 rdma_req->req.iov[i].iov_base = (void *)((uintptr_t)(buf + NVMF_DATA_BUFFER_MASK) & 1160 ~NVMF_DATA_BUFFER_MASK); 1161 rdma_req->req.iov[i].iov_len = spdk_min(length, rtransport->transport.opts.io_unit_size); 1162 rdma_req->req.iovcnt++; 1163 rdma_req->data.buffers[i] = buf; 1164 rdma_req->data.wr.sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[i].iov_base); 1165 rdma_req->data.wr.sg_list[i].length = rdma_req->req.iov[i].iov_len; 1166 rdma_req->data.wr.sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1167 (uint64_t)buf, NULL))->lkey; 1168 1169 length -= rdma_req->req.iov[i].iov_len; 1170 i++; 1171 } 1172 1173 rdma_req->data_from_pool = true; 1174 1175 return 0; 1176 1177 nomem: 1178 while (i) { 1179 i--; 1180 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 1181 rdma_req->req.iov[i].iov_base = NULL; 1182 rdma_req->req.iov[i].iov_len = 0; 1183 1184 rdma_req->data.wr.sg_list[i].addr = 0; 1185 rdma_req->data.wr.sg_list[i].length = 0; 1186 rdma_req->data.wr.sg_list[i].lkey = 0; 1187 } 1188 rdma_req->req.iovcnt = 0; 1189 return -ENOMEM; 1190 } 1191 1192 static int 1193 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1194 struct spdk_nvmf_rdma_device *device, 1195 struct spdk_nvmf_rdma_request *rdma_req) 1196 { 1197 struct spdk_nvme_cmd *cmd; 1198 struct spdk_nvme_cpl *rsp; 1199 struct spdk_nvme_sgl_descriptor *sgl; 1200 1201 cmd = &rdma_req->req.cmd->nvme_cmd; 1202 rsp = &rdma_req->req.rsp->nvme_cpl; 1203 sgl = &cmd->dptr.sgl1; 1204 1205 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1206 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1207 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1208 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1209 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1210 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1211 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1212 return -1; 1213 } 1214 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1215 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1216 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1217 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1218 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1219 } 1220 } 1221 #endif 1222 1223 /* fill request length and populate iovs */ 1224 rdma_req->req.length = sgl->keyed.length; 1225 1226 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1227 /* No available buffers. Queue this request up. */ 1228 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1229 return 0; 1230 } 1231 1232 /* backward compatible */ 1233 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1234 1235 /* rdma wr specifics */ 1236 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1237 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1238 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1239 1240 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1241 rdma_req->req.iovcnt); 1242 1243 return 0; 1244 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1245 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1246 uint64_t offset = sgl->address; 1247 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1248 1249 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1250 offset, sgl->unkeyed.length); 1251 1252 if (offset > max_len) { 1253 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1254 offset, max_len); 1255 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1256 return -1; 1257 } 1258 max_len -= (uint32_t)offset; 1259 1260 if (sgl->unkeyed.length > max_len) { 1261 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1262 sgl->unkeyed.length, max_len); 1263 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1264 return -1; 1265 } 1266 1267 rdma_req->req.data = rdma_req->recv->buf + offset; 1268 rdma_req->data_from_pool = false; 1269 rdma_req->req.length = sgl->unkeyed.length; 1270 1271 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1272 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1273 rdma_req->req.iovcnt = 1; 1274 1275 return 0; 1276 } 1277 1278 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1279 sgl->generic.type, sgl->generic.subtype); 1280 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1281 return -1; 1282 } 1283 1284 static bool 1285 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1286 struct spdk_nvmf_rdma_request *rdma_req) 1287 { 1288 struct spdk_nvmf_rdma_qpair *rqpair; 1289 struct spdk_nvmf_rdma_device *device; 1290 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1291 int rc; 1292 struct spdk_nvmf_rdma_recv *rdma_recv; 1293 enum spdk_nvmf_rdma_request_state prev_state; 1294 bool progress = false; 1295 int data_posted; 1296 int cur_rdma_rw_depth; 1297 1298 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1299 device = rqpair->port->device; 1300 1301 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1302 1303 /* If the queue pair is in an error state, force the request to the completed state 1304 * to release resources. */ 1305 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1306 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1307 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1308 } 1309 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1310 } 1311 1312 /* The loop here is to allow for several back-to-back state changes. */ 1313 do { 1314 prev_state = rdma_req->state; 1315 1316 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1317 1318 switch (rdma_req->state) { 1319 case RDMA_REQUEST_STATE_FREE: 1320 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1321 * to escape this state. */ 1322 break; 1323 case RDMA_REQUEST_STATE_NEW: 1324 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1325 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1326 rdma_recv = rdma_req->recv; 1327 1328 /* The first element of the SGL is the NVMe command */ 1329 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1330 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1331 1332 TAILQ_REMOVE(&rqpair->incoming_queue, rdma_recv, link); 1333 1334 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 1335 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1336 break; 1337 } 1338 1339 /* The next state transition depends on the data transfer needs of this request. */ 1340 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1341 1342 /* If no data to transfer, ready to execute. */ 1343 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1344 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1345 break; 1346 } 1347 1348 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEED_BUFFER); 1349 TAILQ_INSERT_TAIL(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1350 break; 1351 case RDMA_REQUEST_STATE_NEED_BUFFER: 1352 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1353 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1354 1355 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1356 1357 if (rdma_req != TAILQ_FIRST(&rqpair->ch->pending_data_buf_queue)) { 1358 /* This request needs to wait in line to obtain a buffer */ 1359 break; 1360 } 1361 1362 /* Try to get a data buffer */ 1363 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1364 if (rc < 0) { 1365 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1366 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1367 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1368 break; 1369 } 1370 1371 if (!rdma_req->req.data) { 1372 /* No buffers available. */ 1373 break; 1374 } 1375 1376 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1377 1378 /* If data is transferring from host to controller and the data didn't 1379 * arrive using in capsule data, we need to do a transfer from the host. 1380 */ 1381 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1382 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1383 break; 1384 } 1385 1386 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1387 break; 1388 case RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING: 1389 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 0, 0, 1390 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1391 1392 if (rdma_req != TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING])) { 1393 /* This request needs to wait in line to perform RDMA */ 1394 break; 1395 } 1396 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1397 1398 if (cur_rdma_rw_depth >= rqpair->max_rw_depth) { 1399 /* R/W queue is full, need to wait */ 1400 break; 1401 } 1402 1403 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1404 rc = request_transfer_in(&rdma_req->req); 1405 if (!rc) { 1406 spdk_nvmf_rdma_request_set_state(rdma_req, 1407 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1408 } else { 1409 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1410 spdk_nvmf_rdma_request_set_state(rdma_req, 1411 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1412 } 1413 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1414 /* The data transfer will be kicked off from 1415 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1416 */ 1417 spdk_nvmf_rdma_request_set_state(rdma_req, 1418 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1419 } else { 1420 SPDK_ERRLOG("Cannot perform data transfer, unknown state: %u\n", 1421 rdma_req->req.xfer); 1422 assert(0); 1423 } 1424 break; 1425 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1426 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1427 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1428 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1429 * to escape this state. */ 1430 break; 1431 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1432 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1433 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1434 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTING); 1435 spdk_nvmf_request_exec(&rdma_req->req); 1436 break; 1437 case RDMA_REQUEST_STATE_EXECUTING: 1438 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1439 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1440 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1441 * to escape this state. */ 1442 break; 1443 case RDMA_REQUEST_STATE_EXECUTED: 1444 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1445 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1446 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1447 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1448 } else { 1449 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1450 } 1451 break; 1452 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1453 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 1454 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1455 rc = request_transfer_out(&rdma_req->req, &data_posted); 1456 assert(rc == 0); /* No good way to handle this currently */ 1457 if (rc) { 1458 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1459 } else { 1460 spdk_nvmf_rdma_request_set_state(rdma_req, 1461 data_posted ? 1462 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1463 RDMA_REQUEST_STATE_COMPLETING); 1464 } 1465 break; 1466 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1467 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 1468 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1469 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1470 * to escape this state. */ 1471 break; 1472 case RDMA_REQUEST_STATE_COMPLETING: 1473 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 1474 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1475 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1476 * to escape this state. */ 1477 break; 1478 case RDMA_REQUEST_STATE_COMPLETED: 1479 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 1480 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1481 1482 if (rdma_req->data_from_pool) { 1483 /* Put the buffer/s back in the pool */ 1484 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 1485 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 1486 rdma_req->req.iov[i].iov_base = NULL; 1487 rdma_req->data.buffers[i] = NULL; 1488 } 1489 rdma_req->data_from_pool = false; 1490 } 1491 rdma_req->req.length = 0; 1492 rdma_req->req.iovcnt = 0; 1493 rdma_req->req.data = NULL; 1494 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 1495 break; 1496 case RDMA_REQUEST_NUM_STATES: 1497 default: 1498 assert(0); 1499 break; 1500 } 1501 1502 if (rdma_req->state != prev_state) { 1503 progress = true; 1504 } 1505 } while (rdma_req->state != prev_state); 1506 1507 return progress; 1508 } 1509 1510 /* Public API callbacks begin here */ 1511 1512 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 1513 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 1514 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 1515 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 1516 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 1517 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE 4096 1518 #define SPDK_NVMF_RDMA_DEFAULT_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 1519 1520 static void 1521 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 1522 { 1523 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 1524 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 1525 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 1526 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 1527 opts->io_unit_size = spdk_max(SPDK_NVMF_RDMA_DEFAULT_IO_BUFFER_SIZE, 1528 SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 1529 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 1530 } 1531 1532 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 1533 1534 static struct spdk_nvmf_transport * 1535 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 1536 { 1537 int rc; 1538 struct spdk_nvmf_rdma_transport *rtransport; 1539 struct spdk_nvmf_rdma_device *device, *tmp; 1540 struct ibv_context **contexts; 1541 uint32_t i; 1542 int flag; 1543 uint32_t sge_count; 1544 1545 const struct spdk_mem_map_ops nvmf_rdma_map_ops = { 1546 .notify_cb = spdk_nvmf_rdma_mem_notify, 1547 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 1548 }; 1549 1550 rtransport = calloc(1, sizeof(*rtransport)); 1551 if (!rtransport) { 1552 return NULL; 1553 } 1554 1555 if (pthread_mutex_init(&rtransport->lock, NULL)) { 1556 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 1557 free(rtransport); 1558 return NULL; 1559 } 1560 1561 spdk_io_device_register(rtransport, spdk_nvmf_rdma_mgmt_channel_create, 1562 spdk_nvmf_rdma_mgmt_channel_destroy, 1563 sizeof(struct spdk_nvmf_rdma_mgmt_channel), 1564 "rdma_transport"); 1565 1566 TAILQ_INIT(&rtransport->devices); 1567 TAILQ_INIT(&rtransport->ports); 1568 1569 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 1570 1571 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 1572 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 1573 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 1574 " in_capsule_data_size=%d, max_aq_depth=%d\n", 1575 opts->max_queue_depth, 1576 opts->max_io_size, 1577 opts->max_qpairs_per_ctrlr, 1578 opts->io_unit_size, 1579 opts->in_capsule_data_size, 1580 opts->max_aq_depth); 1581 1582 /* I/O unit size cannot be larger than max I/O size */ 1583 if (opts->io_unit_size > opts->max_io_size) { 1584 opts->io_unit_size = opts->max_io_size; 1585 } 1586 1587 sge_count = opts->max_io_size / opts->io_unit_size; 1588 if (sge_count > NVMF_DEFAULT_TX_SGE) { 1589 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 1590 spdk_nvmf_rdma_destroy(&rtransport->transport); 1591 return NULL; 1592 } 1593 1594 rtransport->event_channel = rdma_create_event_channel(); 1595 if (rtransport->event_channel == NULL) { 1596 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 1597 spdk_nvmf_rdma_destroy(&rtransport->transport); 1598 return NULL; 1599 } 1600 1601 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 1602 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 1603 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 1604 rtransport->event_channel->fd, spdk_strerror(errno)); 1605 spdk_nvmf_rdma_destroy(&rtransport->transport); 1606 return NULL; 1607 } 1608 1609 /* The maximum number of buffers we will need for a given request is equal to just less than double the number of SGL elements */ 1610 rtransport->data_buf_pool = spdk_mempool_create("spdk_nvmf_rdma", 1611 opts->max_queue_depth * (SPDK_NVMF_MAX_SGL_ENTRIES * 2) * 4, 1612 opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT, 1613 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 1614 SPDK_ENV_SOCKET_ID_ANY); 1615 if (!rtransport->data_buf_pool) { 1616 SPDK_ERRLOG("Unable to allocate buffer pool for poll group\n"); 1617 spdk_nvmf_rdma_destroy(&rtransport->transport); 1618 return NULL; 1619 } 1620 1621 contexts = rdma_get_devices(NULL); 1622 if (contexts == NULL) { 1623 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 1624 spdk_nvmf_rdma_destroy(&rtransport->transport); 1625 return NULL; 1626 } 1627 1628 i = 0; 1629 rc = 0; 1630 while (contexts[i] != NULL) { 1631 device = calloc(1, sizeof(*device)); 1632 if (!device) { 1633 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 1634 rc = -ENOMEM; 1635 break; 1636 } 1637 device->context = contexts[i]; 1638 rc = ibv_query_device(device->context, &device->attr); 1639 if (rc < 0) { 1640 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1641 free(device); 1642 break; 1643 1644 } 1645 1646 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1647 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 1648 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 1649 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 1650 } 1651 1652 /** 1653 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 1654 * The Soft-RoCE RXE driver does not currently support send with invalidate, 1655 * but incorrectly reports that it does. There are changes making their way 1656 * through the kernel now that will enable this feature. When they are merged, 1657 * we can conditionally enable this feature. 1658 * 1659 * TODO: enable this for versions of the kernel rxe driver that support it. 1660 */ 1661 if (device->attr.vendor_id == 0) { 1662 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 1663 } 1664 #endif 1665 1666 /* set up device context async ev fd as NON_BLOCKING */ 1667 flag = fcntl(device->context->async_fd, F_GETFL); 1668 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 1669 if (rc < 0) { 1670 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 1671 free(device); 1672 break; 1673 } 1674 1675 device->pd = ibv_alloc_pd(device->context); 1676 if (!device->pd) { 1677 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 1678 free(device); 1679 rc = -1; 1680 break; 1681 } 1682 1683 device->map = spdk_mem_map_alloc(0, &nvmf_rdma_map_ops, device); 1684 if (!device->map) { 1685 SPDK_ERRLOG("Unable to allocate memory map for new poll group\n"); 1686 ibv_dealloc_pd(device->pd); 1687 free(device); 1688 rc = -1; 1689 break; 1690 } 1691 1692 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 1693 i++; 1694 } 1695 rdma_free_devices(contexts); 1696 1697 if (rc < 0) { 1698 spdk_nvmf_rdma_destroy(&rtransport->transport); 1699 return NULL; 1700 } 1701 1702 /* Set up poll descriptor array to monitor events from RDMA and IB 1703 * in a single poll syscall 1704 */ 1705 rtransport->npoll_fds = i + 1; 1706 i = 0; 1707 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 1708 if (rtransport->poll_fds == NULL) { 1709 SPDK_ERRLOG("poll_fds allocation failed\n"); 1710 spdk_nvmf_rdma_destroy(&rtransport->transport); 1711 return NULL; 1712 } 1713 1714 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 1715 rtransport->poll_fds[i++].events = POLLIN; 1716 1717 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 1718 rtransport->poll_fds[i].fd = device->context->async_fd; 1719 rtransport->poll_fds[i++].events = POLLIN; 1720 } 1721 1722 return &rtransport->transport; 1723 } 1724 1725 static int 1726 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 1727 { 1728 struct spdk_nvmf_rdma_transport *rtransport; 1729 struct spdk_nvmf_rdma_port *port, *port_tmp; 1730 struct spdk_nvmf_rdma_device *device, *device_tmp; 1731 1732 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1733 1734 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 1735 TAILQ_REMOVE(&rtransport->ports, port, link); 1736 rdma_destroy_id(port->id); 1737 free(port); 1738 } 1739 1740 if (rtransport->poll_fds != NULL) { 1741 free(rtransport->poll_fds); 1742 } 1743 1744 if (rtransport->event_channel != NULL) { 1745 rdma_destroy_event_channel(rtransport->event_channel); 1746 } 1747 1748 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 1749 TAILQ_REMOVE(&rtransport->devices, device, link); 1750 if (device->map) { 1751 spdk_mem_map_free(&device->map); 1752 } 1753 if (device->pd) { 1754 ibv_dealloc_pd(device->pd); 1755 } 1756 free(device); 1757 } 1758 1759 if (rtransport->data_buf_pool != NULL) { 1760 if (spdk_mempool_count(rtransport->data_buf_pool) != 1761 (transport->opts.max_queue_depth * (SPDK_NVMF_MAX_SGL_ENTRIES * 2) * 4)) { 1762 SPDK_ERRLOG("transport buffer pool count is %zu but should be %u\n", 1763 spdk_mempool_count(rtransport->data_buf_pool), 1764 transport->opts.max_queue_depth * (SPDK_NVMF_MAX_SGL_ENTRIES * 2) * 4); 1765 } 1766 } 1767 1768 spdk_mempool_free(rtransport->data_buf_pool); 1769 spdk_io_device_unregister(rtransport, NULL); 1770 pthread_mutex_destroy(&rtransport->lock); 1771 free(rtransport); 1772 1773 return 0; 1774 } 1775 1776 static int 1777 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 1778 const struct spdk_nvme_transport_id *trid) 1779 { 1780 struct spdk_nvmf_rdma_transport *rtransport; 1781 struct spdk_nvmf_rdma_device *device; 1782 struct spdk_nvmf_rdma_port *port_tmp, *port; 1783 struct addrinfo *res; 1784 struct addrinfo hints; 1785 int family; 1786 int rc; 1787 1788 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1789 1790 port = calloc(1, sizeof(*port)); 1791 if (!port) { 1792 return -ENOMEM; 1793 } 1794 1795 /* Selectively copy the trid. Things like NQN don't matter here - that 1796 * mapping is enforced elsewhere. 1797 */ 1798 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1799 port->trid.adrfam = trid->adrfam; 1800 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 1801 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 1802 1803 pthread_mutex_lock(&rtransport->lock); 1804 assert(rtransport->event_channel != NULL); 1805 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 1806 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 1807 port_tmp->ref++; 1808 free(port); 1809 /* Already listening at this address */ 1810 pthread_mutex_unlock(&rtransport->lock); 1811 return 0; 1812 } 1813 } 1814 1815 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 1816 if (rc < 0) { 1817 SPDK_ERRLOG("rdma_create_id() failed\n"); 1818 free(port); 1819 pthread_mutex_unlock(&rtransport->lock); 1820 return rc; 1821 } 1822 1823 switch (port->trid.adrfam) { 1824 case SPDK_NVMF_ADRFAM_IPV4: 1825 family = AF_INET; 1826 break; 1827 case SPDK_NVMF_ADRFAM_IPV6: 1828 family = AF_INET6; 1829 break; 1830 default: 1831 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 1832 free(port); 1833 pthread_mutex_unlock(&rtransport->lock); 1834 return -EINVAL; 1835 } 1836 1837 memset(&hints, 0, sizeof(hints)); 1838 hints.ai_family = family; 1839 hints.ai_flags = AI_NUMERICSERV; 1840 hints.ai_socktype = SOCK_STREAM; 1841 hints.ai_protocol = 0; 1842 1843 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 1844 if (rc) { 1845 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 1846 free(port); 1847 pthread_mutex_unlock(&rtransport->lock); 1848 return -EINVAL; 1849 } 1850 1851 rc = rdma_bind_addr(port->id, res->ai_addr); 1852 freeaddrinfo(res); 1853 1854 if (rc < 0) { 1855 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 1856 rdma_destroy_id(port->id); 1857 free(port); 1858 pthread_mutex_unlock(&rtransport->lock); 1859 return rc; 1860 } 1861 1862 if (!port->id->verbs) { 1863 SPDK_ERRLOG("ibv_context is null\n"); 1864 rdma_destroy_id(port->id); 1865 free(port); 1866 pthread_mutex_unlock(&rtransport->lock); 1867 return -1; 1868 } 1869 1870 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 1871 if (rc < 0) { 1872 SPDK_ERRLOG("rdma_listen() failed\n"); 1873 rdma_destroy_id(port->id); 1874 free(port); 1875 pthread_mutex_unlock(&rtransport->lock); 1876 return rc; 1877 } 1878 1879 TAILQ_FOREACH(device, &rtransport->devices, link) { 1880 if (device->context == port->id->verbs) { 1881 port->device = device; 1882 break; 1883 } 1884 } 1885 if (!port->device) { 1886 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 1887 port->id->verbs); 1888 rdma_destroy_id(port->id); 1889 free(port); 1890 pthread_mutex_unlock(&rtransport->lock); 1891 return -EINVAL; 1892 } 1893 1894 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 1895 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 1896 1897 port->ref = 1; 1898 1899 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 1900 pthread_mutex_unlock(&rtransport->lock); 1901 1902 return 0; 1903 } 1904 1905 static int 1906 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 1907 const struct spdk_nvme_transport_id *_trid) 1908 { 1909 struct spdk_nvmf_rdma_transport *rtransport; 1910 struct spdk_nvmf_rdma_port *port, *tmp; 1911 struct spdk_nvme_transport_id trid = {}; 1912 1913 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1914 1915 /* Selectively copy the trid. Things like NQN don't matter here - that 1916 * mapping is enforced elsewhere. 1917 */ 1918 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1919 trid.adrfam = _trid->adrfam; 1920 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 1921 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 1922 1923 pthread_mutex_lock(&rtransport->lock); 1924 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 1925 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 1926 assert(port->ref > 0); 1927 port->ref--; 1928 if (port->ref == 0) { 1929 TAILQ_REMOVE(&rtransport->ports, port, link); 1930 rdma_destroy_id(port->id); 1931 free(port); 1932 } 1933 break; 1934 } 1935 } 1936 1937 pthread_mutex_unlock(&rtransport->lock); 1938 return 0; 1939 } 1940 1941 static bool 1942 spdk_nvmf_rdma_qpair_is_idle(struct spdk_nvmf_qpair *qpair) 1943 { 1944 int cur_queue_depth, cur_rdma_rw_depth; 1945 struct spdk_nvmf_rdma_qpair *rqpair; 1946 1947 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1948 cur_queue_depth = spdk_nvmf_rdma_cur_queue_depth(rqpair); 1949 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1950 1951 if (cur_queue_depth == 0 && cur_rdma_rw_depth == 0) { 1952 return true; 1953 } 1954 return false; 1955 } 1956 1957 static void 1958 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 1959 struct spdk_nvmf_rdma_qpair *rqpair) 1960 { 1961 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 1962 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 1963 1964 /* We process I/O in the data transfer pending queue at the highest priority. */ 1965 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING], 1966 state_link, req_tmp) { 1967 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1968 break; 1969 } 1970 } 1971 1972 /* The second highest priority is I/O waiting on memory buffers. */ 1973 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->ch->pending_data_buf_queue, link, 1974 req_tmp) { 1975 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1976 break; 1977 } 1978 } 1979 1980 /* The lowest priority is processing newly received commands */ 1981 TAILQ_FOREACH_SAFE(rdma_recv, &rqpair->incoming_queue, link, recv_tmp) { 1982 if (TAILQ_EMPTY(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE])) { 1983 break; 1984 } 1985 1986 rdma_req = TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE]); 1987 rdma_req->recv = rdma_recv; 1988 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEW); 1989 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1990 break; 1991 } 1992 } 1993 } 1994 1995 static void 1996 _nvmf_rdma_qpair_disconnect(void *ctx) 1997 { 1998 struct spdk_nvmf_qpair *qpair = ctx; 1999 struct spdk_nvmf_rdma_qpair *rqpair; 2000 2001 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2002 2003 spdk_nvmf_rdma_qpair_dec_refcnt(rqpair); 2004 2005 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2006 } 2007 2008 static void 2009 _nvmf_rdma_disconnect_retry(void *ctx) 2010 { 2011 struct spdk_nvmf_qpair *qpair = ctx; 2012 struct spdk_nvmf_poll_group *group; 2013 2014 /* Read the group out of the qpair. This is normally set and accessed only from 2015 * the thread that created the group. Here, we're not on that thread necessarily. 2016 * The data member qpair->group begins it's life as NULL and then is assigned to 2017 * a pointer and never changes. So fortunately reading this and checking for 2018 * non-NULL is thread safe in the x86_64 memory model. */ 2019 group = qpair->group; 2020 2021 if (group == NULL) { 2022 /* The qpair hasn't been assigned to a group yet, so we can't 2023 * process a disconnect. Send a message to ourself and try again. */ 2024 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_disconnect_retry, qpair); 2025 return; 2026 } 2027 2028 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2029 } 2030 2031 static int 2032 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2033 { 2034 struct spdk_nvmf_qpair *qpair; 2035 struct spdk_nvmf_rdma_qpair *rqpair; 2036 2037 if (evt->id == NULL) { 2038 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2039 return -1; 2040 } 2041 2042 qpair = evt->id->context; 2043 if (qpair == NULL) { 2044 SPDK_ERRLOG("disconnect request: no active connection\n"); 2045 return -1; 2046 } 2047 2048 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2049 2050 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2051 2052 spdk_nvmf_rdma_update_ibv_state(rqpair); 2053 spdk_nvmf_rdma_qpair_inc_refcnt(rqpair); 2054 2055 _nvmf_rdma_disconnect_retry(qpair); 2056 2057 return 0; 2058 } 2059 2060 #ifdef DEBUG 2061 static const char *CM_EVENT_STR[] = { 2062 "RDMA_CM_EVENT_ADDR_RESOLVED", 2063 "RDMA_CM_EVENT_ADDR_ERROR", 2064 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2065 "RDMA_CM_EVENT_ROUTE_ERROR", 2066 "RDMA_CM_EVENT_CONNECT_REQUEST", 2067 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2068 "RDMA_CM_EVENT_CONNECT_ERROR", 2069 "RDMA_CM_EVENT_UNREACHABLE", 2070 "RDMA_CM_EVENT_REJECTED", 2071 "RDMA_CM_EVENT_ESTABLISHED", 2072 "RDMA_CM_EVENT_DISCONNECTED", 2073 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2074 "RDMA_CM_EVENT_MULTICAST_JOIN", 2075 "RDMA_CM_EVENT_MULTICAST_ERROR", 2076 "RDMA_CM_EVENT_ADDR_CHANGE", 2077 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2078 }; 2079 #endif /* DEBUG */ 2080 2081 static void 2082 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2083 { 2084 struct spdk_nvmf_rdma_transport *rtransport; 2085 struct rdma_cm_event *event; 2086 int rc; 2087 2088 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2089 2090 if (rtransport->event_channel == NULL) { 2091 return; 2092 } 2093 2094 while (1) { 2095 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2096 if (rc == 0) { 2097 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2098 2099 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2100 2101 switch (event->event) { 2102 case RDMA_CM_EVENT_ADDR_RESOLVED: 2103 case RDMA_CM_EVENT_ADDR_ERROR: 2104 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2105 case RDMA_CM_EVENT_ROUTE_ERROR: 2106 /* No action required. The target never attempts to resolve routes. */ 2107 break; 2108 case RDMA_CM_EVENT_CONNECT_REQUEST: 2109 rc = nvmf_rdma_connect(transport, event, cb_fn); 2110 if (rc < 0) { 2111 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2112 break; 2113 } 2114 break; 2115 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2116 /* The target never initiates a new connection. So this will not occur. */ 2117 break; 2118 case RDMA_CM_EVENT_CONNECT_ERROR: 2119 /* Can this happen? The docs say it can, but not sure what causes it. */ 2120 break; 2121 case RDMA_CM_EVENT_UNREACHABLE: 2122 case RDMA_CM_EVENT_REJECTED: 2123 /* These only occur on the client side. */ 2124 break; 2125 case RDMA_CM_EVENT_ESTABLISHED: 2126 /* TODO: Should we be waiting for this event anywhere? */ 2127 break; 2128 case RDMA_CM_EVENT_DISCONNECTED: 2129 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2130 rc = nvmf_rdma_disconnect(event); 2131 if (rc < 0) { 2132 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2133 break; 2134 } 2135 break; 2136 case RDMA_CM_EVENT_MULTICAST_JOIN: 2137 case RDMA_CM_EVENT_MULTICAST_ERROR: 2138 /* Multicast is not used */ 2139 break; 2140 case RDMA_CM_EVENT_ADDR_CHANGE: 2141 /* Not utilizing this event */ 2142 break; 2143 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2144 /* For now, do nothing. The target never re-uses queue pairs. */ 2145 break; 2146 default: 2147 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2148 break; 2149 } 2150 2151 rdma_ack_cm_event(event); 2152 } else { 2153 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2154 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2155 } 2156 break; 2157 } 2158 } 2159 } 2160 2161 static void 2162 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2163 { 2164 int rc; 2165 struct spdk_nvmf_rdma_qpair *rqpair; 2166 struct ibv_async_event event; 2167 enum ibv_qp_state state; 2168 2169 rc = ibv_get_async_event(device->context, &event); 2170 2171 if (rc) { 2172 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2173 errno, spdk_strerror(errno)); 2174 return; 2175 } 2176 2177 SPDK_NOTICELOG("Async event: %s\n", 2178 ibv_event_type_str(event.event_type)); 2179 2180 switch (event.event_type) { 2181 case IBV_EVENT_QP_FATAL: 2182 rqpair = event.element.qp->qp_context; 2183 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2184 (uintptr_t)rqpair->cm_id, event.event_type); 2185 spdk_nvmf_rdma_update_ibv_state(rqpair); 2186 spdk_nvmf_rdma_qpair_inc_refcnt(rqpair); 2187 spdk_thread_send_msg(rqpair->qpair.group->thread, _nvmf_rdma_qpair_disconnect, &rqpair->qpair); 2188 break; 2189 case IBV_EVENT_QP_LAST_WQE_REACHED: 2190 /* This event only occurs for shared receive queues, which are not currently supported. */ 2191 break; 2192 case IBV_EVENT_SQ_DRAINED: 2193 /* This event occurs frequently in both error and non-error states. 2194 * Check if the qpair is in an error state before sending a message. 2195 * Note that we're not on the correct thread to access the qpair, but 2196 * the operations that the below calls make all happen to be thread 2197 * safe. */ 2198 rqpair = event.element.qp->qp_context; 2199 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2200 (uintptr_t)rqpair->cm_id, event.event_type); 2201 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2202 if (state == IBV_QPS_ERR) { 2203 spdk_nvmf_rdma_qpair_inc_refcnt(rqpair); 2204 spdk_thread_send_msg(rqpair->qpair.group->thread, _nvmf_rdma_qpair_disconnect, &rqpair->qpair); 2205 } 2206 break; 2207 case IBV_EVENT_QP_REQ_ERR: 2208 case IBV_EVENT_QP_ACCESS_ERR: 2209 case IBV_EVENT_COMM_EST: 2210 case IBV_EVENT_PATH_MIG: 2211 case IBV_EVENT_PATH_MIG_ERR: 2212 rqpair = event.element.qp->qp_context; 2213 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2214 (uintptr_t)rqpair->cm_id, event.event_type); 2215 spdk_nvmf_rdma_update_ibv_state(rqpair); 2216 break; 2217 case IBV_EVENT_CQ_ERR: 2218 case IBV_EVENT_DEVICE_FATAL: 2219 case IBV_EVENT_PORT_ACTIVE: 2220 case IBV_EVENT_PORT_ERR: 2221 case IBV_EVENT_LID_CHANGE: 2222 case IBV_EVENT_PKEY_CHANGE: 2223 case IBV_EVENT_SM_CHANGE: 2224 case IBV_EVENT_SRQ_ERR: 2225 case IBV_EVENT_SRQ_LIMIT_REACHED: 2226 case IBV_EVENT_CLIENT_REREGISTER: 2227 case IBV_EVENT_GID_CHANGE: 2228 default: 2229 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2230 break; 2231 } 2232 ibv_ack_async_event(&event); 2233 } 2234 2235 static void 2236 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2237 { 2238 int nfds, i = 0; 2239 struct spdk_nvmf_rdma_transport *rtransport; 2240 struct spdk_nvmf_rdma_device *device, *tmp; 2241 2242 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2243 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2244 2245 if (nfds <= 0) { 2246 return; 2247 } 2248 2249 /* The first poll descriptor is RDMA CM event */ 2250 if (rtransport->poll_fds[i++].revents & POLLIN) { 2251 spdk_nvmf_process_cm_event(transport, cb_fn); 2252 nfds--; 2253 } 2254 2255 if (nfds == 0) { 2256 return; 2257 } 2258 2259 /* Second and subsequent poll descriptors are IB async events */ 2260 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2261 if (rtransport->poll_fds[i++].revents & POLLIN) { 2262 spdk_nvmf_process_ib_event(device); 2263 nfds--; 2264 } 2265 } 2266 /* check all flagged fd's have been served */ 2267 assert(nfds == 0); 2268 } 2269 2270 static void 2271 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2272 struct spdk_nvme_transport_id *trid, 2273 struct spdk_nvmf_discovery_log_page_entry *entry) 2274 { 2275 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2276 entry->adrfam = trid->adrfam; 2277 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2278 2279 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2280 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2281 2282 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2283 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2284 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2285 } 2286 2287 static struct spdk_nvmf_transport_poll_group * 2288 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2289 { 2290 struct spdk_nvmf_rdma_transport *rtransport; 2291 struct spdk_nvmf_rdma_poll_group *rgroup; 2292 struct spdk_nvmf_rdma_poller *poller; 2293 struct spdk_nvmf_rdma_device *device; 2294 2295 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2296 2297 rgroup = calloc(1, sizeof(*rgroup)); 2298 if (!rgroup) { 2299 return NULL; 2300 } 2301 2302 TAILQ_INIT(&rgroup->pollers); 2303 2304 pthread_mutex_lock(&rtransport->lock); 2305 TAILQ_FOREACH(device, &rtransport->devices, link) { 2306 poller = calloc(1, sizeof(*poller)); 2307 if (!poller) { 2308 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2309 free(rgroup); 2310 pthread_mutex_unlock(&rtransport->lock); 2311 return NULL; 2312 } 2313 2314 poller->device = device; 2315 poller->group = rgroup; 2316 2317 TAILQ_INIT(&poller->qpairs); 2318 2319 poller->cq = ibv_create_cq(device->context, NVMF_RDMA_CQ_SIZE, poller, NULL, 0); 2320 if (!poller->cq) { 2321 SPDK_ERRLOG("Unable to create completion queue\n"); 2322 free(poller); 2323 free(rgroup); 2324 pthread_mutex_unlock(&rtransport->lock); 2325 return NULL; 2326 } 2327 2328 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2329 } 2330 2331 pthread_mutex_unlock(&rtransport->lock); 2332 return &rgroup->group; 2333 } 2334 2335 static void 2336 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2337 { 2338 struct spdk_nvmf_rdma_poll_group *rgroup; 2339 struct spdk_nvmf_rdma_poller *poller, *tmp; 2340 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2341 2342 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2343 2344 if (!rgroup) { 2345 return; 2346 } 2347 2348 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2349 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2350 2351 if (poller->cq) { 2352 ibv_destroy_cq(poller->cq); 2353 } 2354 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 2355 spdk_nvmf_rdma_qpair_destroy(qpair); 2356 } 2357 2358 free(poller); 2359 } 2360 2361 free(rgroup); 2362 } 2363 2364 static int 2365 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 2366 struct spdk_nvmf_qpair *qpair) 2367 { 2368 struct spdk_nvmf_rdma_transport *rtransport; 2369 struct spdk_nvmf_rdma_poll_group *rgroup; 2370 struct spdk_nvmf_rdma_qpair *rqpair; 2371 struct spdk_nvmf_rdma_device *device; 2372 struct spdk_nvmf_rdma_poller *poller; 2373 int rc; 2374 2375 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 2376 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2377 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2378 2379 device = rqpair->port->device; 2380 2381 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 2382 if (poller->device == device) { 2383 break; 2384 } 2385 } 2386 2387 if (!poller) { 2388 SPDK_ERRLOG("No poller found for device.\n"); 2389 return -1; 2390 } 2391 2392 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 2393 rqpair->poller = poller; 2394 2395 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 2396 if (rc < 0) { 2397 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 2398 return -1; 2399 } 2400 2401 rqpair->mgmt_channel = spdk_get_io_channel(rtransport); 2402 if (!rqpair->mgmt_channel) { 2403 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2404 spdk_nvmf_rdma_qpair_destroy(rqpair); 2405 return -1; 2406 } 2407 2408 rqpair->ch = spdk_io_channel_get_ctx(rqpair->mgmt_channel); 2409 assert(rqpair->ch != NULL); 2410 2411 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 2412 if (rc) { 2413 /* Try to reject, but we probably can't */ 2414 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2415 spdk_nvmf_rdma_qpair_destroy(rqpair); 2416 return -1; 2417 } 2418 2419 spdk_nvmf_rdma_update_ibv_state(rqpair); 2420 2421 return 0; 2422 } 2423 2424 static int 2425 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 2426 { 2427 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2428 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2429 struct spdk_nvmf_rdma_transport, transport); 2430 2431 if (rdma_req->data_from_pool) { 2432 /* Put the buffer/s back in the pool */ 2433 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 2434 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 2435 rdma_req->req.iov[i].iov_base = NULL; 2436 rdma_req->data.buffers[i] = NULL; 2437 } 2438 rdma_req->data_from_pool = false; 2439 } 2440 rdma_req->req.length = 0; 2441 rdma_req->req.iovcnt = 0; 2442 rdma_req->req.data = NULL; 2443 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 2444 return 0; 2445 } 2446 2447 static int 2448 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 2449 { 2450 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2451 struct spdk_nvmf_rdma_transport, transport); 2452 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 2453 struct spdk_nvmf_rdma_request, req); 2454 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 2455 struct spdk_nvmf_rdma_qpair, qpair); 2456 2457 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2458 /* The connection is alive, so process the request as normal */ 2459 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTED); 2460 } else { 2461 /* The connection is dead. Move the request directly to the completed state. */ 2462 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2463 } 2464 2465 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2466 2467 return 0; 2468 } 2469 2470 static void 2471 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 2472 { 2473 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2474 struct ibv_recv_wr recv_wr = {}; 2475 struct ibv_recv_wr *bad_recv_wr; 2476 struct ibv_send_wr send_wr = {}; 2477 struct ibv_send_wr *bad_send_wr; 2478 int rc; 2479 2480 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 2481 return; 2482 } 2483 2484 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 2485 2486 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2487 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 2488 } 2489 2490 rqpair->drain_recv_wr.type = RDMA_WR_TYPE_DRAIN_RECV; 2491 recv_wr.wr_id = (uintptr_t)&rqpair->drain_recv_wr; 2492 rc = ibv_post_recv(rqpair->cm_id->qp, &recv_wr, &bad_recv_wr); 2493 if (rc) { 2494 SPDK_ERRLOG("Failed to post dummy receive WR, errno %d\n", errno); 2495 assert(false); 2496 return; 2497 } 2498 2499 rqpair->drain_send_wr.type = RDMA_WR_TYPE_DRAIN_SEND; 2500 send_wr.wr_id = (uintptr_t)&rqpair->drain_send_wr; 2501 send_wr.opcode = IBV_WR_SEND; 2502 rc = ibv_post_send(rqpair->cm_id->qp, &send_wr, &bad_send_wr); 2503 if (rc) { 2504 SPDK_ERRLOG("Failed to post dummy send WR, errno %d\n", errno); 2505 assert(false); 2506 return; 2507 } 2508 } 2509 2510 #ifdef DEBUG 2511 static int 2512 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 2513 { 2514 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 2515 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 2516 } 2517 #endif 2518 2519 static int 2520 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 2521 struct spdk_nvmf_rdma_poller *rpoller) 2522 { 2523 struct ibv_wc wc[32]; 2524 struct spdk_nvmf_rdma_wr *rdma_wr; 2525 struct spdk_nvmf_rdma_request *rdma_req; 2526 struct spdk_nvmf_rdma_recv *rdma_recv; 2527 struct spdk_nvmf_rdma_qpair *rqpair; 2528 int reaped, i; 2529 int count = 0; 2530 bool error = false; 2531 2532 /* Poll for completing operations. */ 2533 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 2534 if (reaped < 0) { 2535 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2536 errno, spdk_strerror(errno)); 2537 return -1; 2538 } 2539 2540 for (i = 0; i < reaped; i++) { 2541 2542 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 2543 2544 /* Handle error conditions */ 2545 if (wc[i].status) { 2546 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 2547 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 2548 2549 error = true; 2550 2551 switch (rdma_wr->type) { 2552 case RDMA_WR_TYPE_SEND: 2553 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 2554 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2555 2556 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 2557 /* We're going to attempt an error recovery, so force the request into 2558 * the completed state. */ 2559 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2560 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2561 break; 2562 case RDMA_WR_TYPE_RECV: 2563 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 2564 rqpair = rdma_recv->qpair; 2565 2566 /* Dump this into the incoming queue. This gets cleaned up when 2567 * the queue pair disconnects or recovers. */ 2568 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2569 break; 2570 case RDMA_WR_TYPE_DATA: 2571 /* If the data transfer fails still force the queue into the error state, 2572 * but the rdma_req objects should only be manipulated in response to 2573 * SEND and RECV operations. */ 2574 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2575 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 2576 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2577 break; 2578 case RDMA_WR_TYPE_DRAIN_RECV: 2579 rqpair = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_qpair, drain_recv_wr); 2580 assert(rqpair->disconnect_flags & RDMA_QP_DISCONNECTING); 2581 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Drained QP RECV %u (%p)\n", rqpair->qpair.qid, rqpair); 2582 rqpair->disconnect_flags |= RDMA_QP_RECV_DRAINED; 2583 if (rqpair->disconnect_flags & RDMA_QP_SEND_DRAINED) { 2584 spdk_nvmf_rdma_qpair_destroy(rqpair); 2585 } 2586 /* Continue so that this does not trigger the disconnect path below. */ 2587 continue; 2588 case RDMA_WR_TYPE_DRAIN_SEND: 2589 rqpair = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_qpair, drain_send_wr); 2590 assert(rqpair->disconnect_flags & RDMA_QP_DISCONNECTING); 2591 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Drained QP SEND %u (%p)\n", rqpair->qpair.qid, rqpair); 2592 rqpair->disconnect_flags |= RDMA_QP_SEND_DRAINED; 2593 if (rqpair->disconnect_flags & RDMA_QP_RECV_DRAINED) { 2594 spdk_nvmf_rdma_qpair_destroy(rqpair); 2595 } 2596 /* Continue so that this does not trigger the disconnect path below. */ 2597 continue; 2598 default: 2599 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2600 continue; 2601 } 2602 2603 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 2604 /* Disconnect the connection. */ 2605 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2606 } 2607 continue; 2608 } 2609 2610 switch (wc[i].opcode) { 2611 case IBV_WC_SEND: 2612 assert(rdma_wr->type == RDMA_WR_TYPE_SEND); 2613 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 2614 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2615 2616 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 2617 2618 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2619 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2620 2621 count++; 2622 2623 /* Try to process other queued requests */ 2624 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2625 break; 2626 2627 case IBV_WC_RDMA_WRITE: 2628 assert(rdma_wr->type == RDMA_WR_TYPE_DATA); 2629 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2630 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2631 2632 /* Try to process other queued requests */ 2633 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2634 break; 2635 2636 case IBV_WC_RDMA_READ: 2637 assert(rdma_wr->type == RDMA_WR_TYPE_DATA); 2638 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2639 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2640 2641 assert(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2642 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 2643 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2644 2645 /* Try to process other queued requests */ 2646 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2647 break; 2648 2649 case IBV_WC_RECV: 2650 assert(rdma_wr->type == RDMA_WR_TYPE_RECV); 2651 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 2652 rqpair = rdma_recv->qpair; 2653 2654 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2655 /* Try to process other queued requests */ 2656 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2657 break; 2658 2659 default: 2660 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2661 continue; 2662 } 2663 } 2664 2665 if (error == true) { 2666 return -1; 2667 } 2668 2669 return count; 2670 } 2671 2672 static int 2673 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 2674 { 2675 struct spdk_nvmf_rdma_transport *rtransport; 2676 struct spdk_nvmf_rdma_poll_group *rgroup; 2677 struct spdk_nvmf_rdma_poller *rpoller; 2678 int count, rc; 2679 2680 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 2681 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2682 2683 count = 0; 2684 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2685 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 2686 if (rc < 0) { 2687 return rc; 2688 } 2689 count += rc; 2690 } 2691 2692 return count; 2693 } 2694 2695 static int 2696 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2697 struct spdk_nvme_transport_id *trid, 2698 bool peer) 2699 { 2700 struct sockaddr *saddr; 2701 uint16_t port; 2702 2703 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 2704 2705 if (peer) { 2706 saddr = rdma_get_peer_addr(id); 2707 } else { 2708 saddr = rdma_get_local_addr(id); 2709 } 2710 switch (saddr->sa_family) { 2711 case AF_INET: { 2712 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 2713 2714 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 2715 inet_ntop(AF_INET, &saddr_in->sin_addr, 2716 trid->traddr, sizeof(trid->traddr)); 2717 if (peer) { 2718 port = ntohs(rdma_get_dst_port(id)); 2719 } else { 2720 port = ntohs(rdma_get_src_port(id)); 2721 } 2722 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 2723 break; 2724 } 2725 case AF_INET6: { 2726 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 2727 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 2728 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 2729 trid->traddr, sizeof(trid->traddr)); 2730 if (peer) { 2731 port = ntohs(rdma_get_dst_port(id)); 2732 } else { 2733 port = ntohs(rdma_get_src_port(id)); 2734 } 2735 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 2736 break; 2737 } 2738 default: 2739 return -1; 2740 2741 } 2742 2743 return 0; 2744 } 2745 2746 static int 2747 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 2748 struct spdk_nvme_transport_id *trid) 2749 { 2750 struct spdk_nvmf_rdma_qpair *rqpair; 2751 2752 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2753 2754 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 2755 } 2756 2757 static int 2758 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 2759 struct spdk_nvme_transport_id *trid) 2760 { 2761 struct spdk_nvmf_rdma_qpair *rqpair; 2762 2763 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2764 2765 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 2766 } 2767 2768 static int 2769 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 2770 struct spdk_nvme_transport_id *trid) 2771 { 2772 struct spdk_nvmf_rdma_qpair *rqpair; 2773 2774 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2775 2776 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 2777 } 2778 2779 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 2780 .type = SPDK_NVME_TRANSPORT_RDMA, 2781 .opts_init = spdk_nvmf_rdma_opts_init, 2782 .create = spdk_nvmf_rdma_create, 2783 .destroy = spdk_nvmf_rdma_destroy, 2784 2785 .listen = spdk_nvmf_rdma_listen, 2786 .stop_listen = spdk_nvmf_rdma_stop_listen, 2787 .accept = spdk_nvmf_rdma_accept, 2788 2789 .listener_discover = spdk_nvmf_rdma_discover, 2790 2791 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 2792 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 2793 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 2794 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 2795 2796 .req_free = spdk_nvmf_rdma_request_free, 2797 .req_complete = spdk_nvmf_rdma_request_complete, 2798 2799 .qpair_fini = spdk_nvmf_rdma_close_qpair, 2800 .qpair_is_idle = spdk_nvmf_rdma_qpair_is_idle, 2801 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 2802 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 2803 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 2804 2805 }; 2806 2807 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 2808