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 uint64_t translation_len; 1151 uint32_t i = 0; 1152 int rc = 0; 1153 1154 rdma_req->req.iovcnt = 0; 1155 while (length) { 1156 buf = spdk_mempool_get(rtransport->data_buf_pool); 1157 if (!buf) { 1158 rc = -ENOMEM; 1159 goto err_exit; 1160 } 1161 1162 rdma_req->req.iov[i].iov_base = (void *)((uintptr_t)(buf + NVMF_DATA_BUFFER_MASK) & 1163 ~NVMF_DATA_BUFFER_MASK); 1164 rdma_req->req.iov[i].iov_len = spdk_min(length, rtransport->transport.opts.io_unit_size); 1165 rdma_req->req.iovcnt++; 1166 rdma_req->data.buffers[i] = buf; 1167 rdma_req->data.wr.sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[i].iov_base); 1168 rdma_req->data.wr.sg_list[i].length = rdma_req->req.iov[i].iov_len; 1169 translation_len = rdma_req->req.iov[i].iov_len; 1170 rdma_req->data.wr.sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1171 (uint64_t)buf, &translation_len))->lkey; 1172 length -= rdma_req->req.iov[i].iov_len; 1173 1174 if (translation_len < rdma_req->req.iov[i].iov_len) { 1175 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions\n"); 1176 rc = -EINVAL; 1177 goto err_exit; 1178 } 1179 i++; 1180 } 1181 1182 rdma_req->data_from_pool = true; 1183 1184 return rc; 1185 1186 err_exit: 1187 while (i) { 1188 i--; 1189 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 1190 rdma_req->req.iov[i].iov_base = NULL; 1191 rdma_req->req.iov[i].iov_len = 0; 1192 1193 rdma_req->data.wr.sg_list[i].addr = 0; 1194 rdma_req->data.wr.sg_list[i].length = 0; 1195 rdma_req->data.wr.sg_list[i].lkey = 0; 1196 } 1197 rdma_req->req.iovcnt = 0; 1198 return rc; 1199 } 1200 1201 static int 1202 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1203 struct spdk_nvmf_rdma_device *device, 1204 struct spdk_nvmf_rdma_request *rdma_req) 1205 { 1206 struct spdk_nvme_cmd *cmd; 1207 struct spdk_nvme_cpl *rsp; 1208 struct spdk_nvme_sgl_descriptor *sgl; 1209 1210 cmd = &rdma_req->req.cmd->nvme_cmd; 1211 rsp = &rdma_req->req.rsp->nvme_cpl; 1212 sgl = &cmd->dptr.sgl1; 1213 1214 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1215 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1216 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1217 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1218 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1219 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1220 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1221 return -1; 1222 } 1223 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1224 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1225 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1226 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1227 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1228 } 1229 } 1230 #endif 1231 1232 /* fill request length and populate iovs */ 1233 rdma_req->req.length = sgl->keyed.length; 1234 1235 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1236 /* No available buffers. Queue this request up. */ 1237 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1238 return 0; 1239 } 1240 1241 /* backward compatible */ 1242 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1243 1244 /* rdma wr specifics */ 1245 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1246 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1247 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1248 1249 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1250 rdma_req->req.iovcnt); 1251 1252 return 0; 1253 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1254 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1255 uint64_t offset = sgl->address; 1256 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1257 1258 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1259 offset, sgl->unkeyed.length); 1260 1261 if (offset > max_len) { 1262 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1263 offset, max_len); 1264 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1265 return -1; 1266 } 1267 max_len -= (uint32_t)offset; 1268 1269 if (sgl->unkeyed.length > max_len) { 1270 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1271 sgl->unkeyed.length, max_len); 1272 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1273 return -1; 1274 } 1275 1276 rdma_req->req.data = rdma_req->recv->buf + offset; 1277 rdma_req->data_from_pool = false; 1278 rdma_req->req.length = sgl->unkeyed.length; 1279 1280 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1281 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1282 rdma_req->req.iovcnt = 1; 1283 1284 return 0; 1285 } 1286 1287 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1288 sgl->generic.type, sgl->generic.subtype); 1289 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1290 return -1; 1291 } 1292 1293 static bool 1294 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1295 struct spdk_nvmf_rdma_request *rdma_req) 1296 { 1297 struct spdk_nvmf_rdma_qpair *rqpair; 1298 struct spdk_nvmf_rdma_device *device; 1299 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1300 int rc; 1301 struct spdk_nvmf_rdma_recv *rdma_recv; 1302 enum spdk_nvmf_rdma_request_state prev_state; 1303 bool progress = false; 1304 int data_posted; 1305 int cur_rdma_rw_depth; 1306 1307 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1308 device = rqpair->port->device; 1309 1310 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1311 1312 /* If the queue pair is in an error state, force the request to the completed state 1313 * to release resources. */ 1314 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1315 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1316 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1317 } 1318 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1319 } 1320 1321 /* The loop here is to allow for several back-to-back state changes. */ 1322 do { 1323 prev_state = rdma_req->state; 1324 1325 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1326 1327 switch (rdma_req->state) { 1328 case RDMA_REQUEST_STATE_FREE: 1329 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1330 * to escape this state. */ 1331 break; 1332 case RDMA_REQUEST_STATE_NEW: 1333 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1334 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1335 rdma_recv = rdma_req->recv; 1336 1337 /* The first element of the SGL is the NVMe command */ 1338 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1339 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1340 1341 TAILQ_REMOVE(&rqpair->incoming_queue, rdma_recv, link); 1342 1343 if (rqpair->ibv_attr.qp_state == IBV_QPS_ERR) { 1344 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1345 break; 1346 } 1347 1348 /* The next state transition depends on the data transfer needs of this request. */ 1349 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1350 1351 /* If no data to transfer, ready to execute. */ 1352 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1353 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1354 break; 1355 } 1356 1357 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEED_BUFFER); 1358 TAILQ_INSERT_TAIL(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1359 break; 1360 case RDMA_REQUEST_STATE_NEED_BUFFER: 1361 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1362 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1363 1364 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1365 1366 if (rdma_req != TAILQ_FIRST(&rqpair->ch->pending_data_buf_queue)) { 1367 /* This request needs to wait in line to obtain a buffer */ 1368 break; 1369 } 1370 1371 /* Try to get a data buffer */ 1372 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1373 if (rc < 0) { 1374 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1375 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1376 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1377 break; 1378 } 1379 1380 if (!rdma_req->req.data) { 1381 /* No buffers available. */ 1382 break; 1383 } 1384 1385 TAILQ_REMOVE(&rqpair->ch->pending_data_buf_queue, rdma_req, link); 1386 1387 /* If data is transferring from host to controller and the data didn't 1388 * arrive using in capsule data, we need to do a transfer from the host. 1389 */ 1390 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1391 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1392 break; 1393 } 1394 1395 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 1396 break; 1397 case RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING: 1398 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING, 0, 0, 1399 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1400 1401 if (rdma_req != TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING])) { 1402 /* This request needs to wait in line to perform RDMA */ 1403 break; 1404 } 1405 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1406 1407 if (cur_rdma_rw_depth >= rqpair->max_rw_depth) { 1408 /* R/W queue is full, need to wait */ 1409 break; 1410 } 1411 1412 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1413 rc = request_transfer_in(&rdma_req->req); 1414 if (!rc) { 1415 spdk_nvmf_rdma_request_set_state(rdma_req, 1416 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1417 } else { 1418 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1419 spdk_nvmf_rdma_request_set_state(rdma_req, 1420 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1421 } 1422 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1423 /* The data transfer will be kicked off from 1424 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1425 */ 1426 spdk_nvmf_rdma_request_set_state(rdma_req, 1427 RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1428 } else { 1429 SPDK_ERRLOG("Cannot perform data transfer, unknown state: %u\n", 1430 rdma_req->req.xfer); 1431 assert(0); 1432 } 1433 break; 1434 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1435 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1436 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1437 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1438 * to escape this state. */ 1439 break; 1440 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1441 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1442 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1443 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTING); 1444 spdk_nvmf_request_exec(&rdma_req->req); 1445 break; 1446 case RDMA_REQUEST_STATE_EXECUTING: 1447 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1448 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1449 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1450 * to escape this state. */ 1451 break; 1452 case RDMA_REQUEST_STATE_EXECUTED: 1453 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1454 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1455 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1456 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING); 1457 } else { 1458 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_COMPLETE); 1459 } 1460 break; 1461 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1462 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 1463 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1464 rc = request_transfer_out(&rdma_req->req, &data_posted); 1465 assert(rc == 0); /* No good way to handle this currently */ 1466 if (rc) { 1467 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 1468 } else { 1469 spdk_nvmf_rdma_request_set_state(rdma_req, 1470 data_posted ? 1471 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1472 RDMA_REQUEST_STATE_COMPLETING); 1473 } 1474 break; 1475 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1476 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 1477 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1478 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1479 * to escape this state. */ 1480 break; 1481 case RDMA_REQUEST_STATE_COMPLETING: 1482 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 1483 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1484 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1485 * to escape this state. */ 1486 break; 1487 case RDMA_REQUEST_STATE_COMPLETED: 1488 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 1489 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1490 1491 if (rdma_req->data_from_pool) { 1492 /* Put the buffer/s back in the pool */ 1493 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 1494 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 1495 rdma_req->req.iov[i].iov_base = NULL; 1496 rdma_req->data.buffers[i] = NULL; 1497 } 1498 rdma_req->data_from_pool = false; 1499 } 1500 rdma_req->req.length = 0; 1501 rdma_req->req.iovcnt = 0; 1502 rdma_req->req.data = NULL; 1503 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 1504 break; 1505 case RDMA_REQUEST_NUM_STATES: 1506 default: 1507 assert(0); 1508 break; 1509 } 1510 1511 if (rdma_req->state != prev_state) { 1512 progress = true; 1513 } 1514 } while (rdma_req->state != prev_state); 1515 1516 return progress; 1517 } 1518 1519 /* Public API callbacks begin here */ 1520 1521 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 1522 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 1523 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 1524 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 1525 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 1526 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE 4096 1527 #define SPDK_NVMF_RDMA_DEFAULT_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 1528 1529 static void 1530 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 1531 { 1532 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 1533 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 1534 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 1535 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 1536 opts->io_unit_size = spdk_max(SPDK_NVMF_RDMA_DEFAULT_IO_BUFFER_SIZE, 1537 SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 1538 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 1539 } 1540 1541 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 1542 1543 static struct spdk_nvmf_transport * 1544 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 1545 { 1546 int rc; 1547 struct spdk_nvmf_rdma_transport *rtransport; 1548 struct spdk_nvmf_rdma_device *device, *tmp; 1549 struct ibv_context **contexts; 1550 uint32_t i; 1551 int flag; 1552 uint32_t sge_count; 1553 1554 const struct spdk_mem_map_ops nvmf_rdma_map_ops = { 1555 .notify_cb = spdk_nvmf_rdma_mem_notify, 1556 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 1557 }; 1558 1559 rtransport = calloc(1, sizeof(*rtransport)); 1560 if (!rtransport) { 1561 return NULL; 1562 } 1563 1564 if (pthread_mutex_init(&rtransport->lock, NULL)) { 1565 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 1566 free(rtransport); 1567 return NULL; 1568 } 1569 1570 spdk_io_device_register(rtransport, spdk_nvmf_rdma_mgmt_channel_create, 1571 spdk_nvmf_rdma_mgmt_channel_destroy, 1572 sizeof(struct spdk_nvmf_rdma_mgmt_channel), 1573 "rdma_transport"); 1574 1575 TAILQ_INIT(&rtransport->devices); 1576 TAILQ_INIT(&rtransport->ports); 1577 1578 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 1579 1580 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 1581 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 1582 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 1583 " in_capsule_data_size=%d, max_aq_depth=%d\n", 1584 opts->max_queue_depth, 1585 opts->max_io_size, 1586 opts->max_qpairs_per_ctrlr, 1587 opts->io_unit_size, 1588 opts->in_capsule_data_size, 1589 opts->max_aq_depth); 1590 1591 /* I/O unit size cannot be larger than max I/O size */ 1592 if (opts->io_unit_size > opts->max_io_size) { 1593 opts->io_unit_size = opts->max_io_size; 1594 } 1595 1596 sge_count = opts->max_io_size / opts->io_unit_size; 1597 if (sge_count > NVMF_DEFAULT_TX_SGE) { 1598 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 1599 spdk_nvmf_rdma_destroy(&rtransport->transport); 1600 return NULL; 1601 } 1602 1603 rtransport->event_channel = rdma_create_event_channel(); 1604 if (rtransport->event_channel == NULL) { 1605 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 1606 spdk_nvmf_rdma_destroy(&rtransport->transport); 1607 return NULL; 1608 } 1609 1610 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 1611 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 1612 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 1613 rtransport->event_channel->fd, spdk_strerror(errno)); 1614 spdk_nvmf_rdma_destroy(&rtransport->transport); 1615 return NULL; 1616 } 1617 1618 /* The maximum number of buffers we will need for a given request is equal to just less than double the number of SGL elements */ 1619 rtransport->data_buf_pool = spdk_mempool_create("spdk_nvmf_rdma", 1620 opts->max_queue_depth * (SPDK_NVMF_MAX_SGL_ENTRIES * 2) * 4, 1621 opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT, 1622 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 1623 SPDK_ENV_SOCKET_ID_ANY); 1624 if (!rtransport->data_buf_pool) { 1625 SPDK_ERRLOG("Unable to allocate buffer pool for poll group\n"); 1626 spdk_nvmf_rdma_destroy(&rtransport->transport); 1627 return NULL; 1628 } 1629 1630 contexts = rdma_get_devices(NULL); 1631 if (contexts == NULL) { 1632 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 1633 spdk_nvmf_rdma_destroy(&rtransport->transport); 1634 return NULL; 1635 } 1636 1637 i = 0; 1638 rc = 0; 1639 while (contexts[i] != NULL) { 1640 device = calloc(1, sizeof(*device)); 1641 if (!device) { 1642 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 1643 rc = -ENOMEM; 1644 break; 1645 } 1646 device->context = contexts[i]; 1647 rc = ibv_query_device(device->context, &device->attr); 1648 if (rc < 0) { 1649 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 1650 free(device); 1651 break; 1652 1653 } 1654 1655 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1656 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 1657 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 1658 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 1659 } 1660 1661 /** 1662 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 1663 * The Soft-RoCE RXE driver does not currently support send with invalidate, 1664 * but incorrectly reports that it does. There are changes making their way 1665 * through the kernel now that will enable this feature. When they are merged, 1666 * we can conditionally enable this feature. 1667 * 1668 * TODO: enable this for versions of the kernel rxe driver that support it. 1669 */ 1670 if (device->attr.vendor_id == 0) { 1671 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 1672 } 1673 #endif 1674 1675 /* set up device context async ev fd as NON_BLOCKING */ 1676 flag = fcntl(device->context->async_fd, F_GETFL); 1677 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 1678 if (rc < 0) { 1679 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 1680 free(device); 1681 break; 1682 } 1683 1684 device->pd = ibv_alloc_pd(device->context); 1685 if (!device->pd) { 1686 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 1687 free(device); 1688 rc = -1; 1689 break; 1690 } 1691 1692 device->map = spdk_mem_map_alloc(0, &nvmf_rdma_map_ops, device); 1693 if (!device->map) { 1694 SPDK_ERRLOG("Unable to allocate memory map for new poll group\n"); 1695 ibv_dealloc_pd(device->pd); 1696 free(device); 1697 rc = -1; 1698 break; 1699 } 1700 1701 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 1702 i++; 1703 } 1704 rdma_free_devices(contexts); 1705 1706 if (rc < 0) { 1707 spdk_nvmf_rdma_destroy(&rtransport->transport); 1708 return NULL; 1709 } 1710 1711 /* Set up poll descriptor array to monitor events from RDMA and IB 1712 * in a single poll syscall 1713 */ 1714 rtransport->npoll_fds = i + 1; 1715 i = 0; 1716 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 1717 if (rtransport->poll_fds == NULL) { 1718 SPDK_ERRLOG("poll_fds allocation failed\n"); 1719 spdk_nvmf_rdma_destroy(&rtransport->transport); 1720 return NULL; 1721 } 1722 1723 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 1724 rtransport->poll_fds[i++].events = POLLIN; 1725 1726 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 1727 rtransport->poll_fds[i].fd = device->context->async_fd; 1728 rtransport->poll_fds[i++].events = POLLIN; 1729 } 1730 1731 return &rtransport->transport; 1732 } 1733 1734 static int 1735 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 1736 { 1737 struct spdk_nvmf_rdma_transport *rtransport; 1738 struct spdk_nvmf_rdma_port *port, *port_tmp; 1739 struct spdk_nvmf_rdma_device *device, *device_tmp; 1740 1741 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1742 1743 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 1744 TAILQ_REMOVE(&rtransport->ports, port, link); 1745 rdma_destroy_id(port->id); 1746 free(port); 1747 } 1748 1749 if (rtransport->poll_fds != NULL) { 1750 free(rtransport->poll_fds); 1751 } 1752 1753 if (rtransport->event_channel != NULL) { 1754 rdma_destroy_event_channel(rtransport->event_channel); 1755 } 1756 1757 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 1758 TAILQ_REMOVE(&rtransport->devices, device, link); 1759 if (device->map) { 1760 spdk_mem_map_free(&device->map); 1761 } 1762 if (device->pd) { 1763 ibv_dealloc_pd(device->pd); 1764 } 1765 free(device); 1766 } 1767 1768 if (rtransport->data_buf_pool != NULL) { 1769 if (spdk_mempool_count(rtransport->data_buf_pool) != 1770 (transport->opts.max_queue_depth * (SPDK_NVMF_MAX_SGL_ENTRIES * 2) * 4)) { 1771 SPDK_ERRLOG("transport buffer pool count is %zu but should be %u\n", 1772 spdk_mempool_count(rtransport->data_buf_pool), 1773 transport->opts.max_queue_depth * (SPDK_NVMF_MAX_SGL_ENTRIES * 2) * 4); 1774 } 1775 } 1776 1777 spdk_mempool_free(rtransport->data_buf_pool); 1778 spdk_io_device_unregister(rtransport, NULL); 1779 pthread_mutex_destroy(&rtransport->lock); 1780 free(rtransport); 1781 1782 return 0; 1783 } 1784 1785 static int 1786 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 1787 const struct spdk_nvme_transport_id *trid) 1788 { 1789 struct spdk_nvmf_rdma_transport *rtransport; 1790 struct spdk_nvmf_rdma_device *device; 1791 struct spdk_nvmf_rdma_port *port_tmp, *port; 1792 struct addrinfo *res; 1793 struct addrinfo hints; 1794 int family; 1795 int rc; 1796 1797 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1798 1799 port = calloc(1, sizeof(*port)); 1800 if (!port) { 1801 return -ENOMEM; 1802 } 1803 1804 /* Selectively copy the trid. Things like NQN don't matter here - that 1805 * mapping is enforced elsewhere. 1806 */ 1807 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1808 port->trid.adrfam = trid->adrfam; 1809 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 1810 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 1811 1812 pthread_mutex_lock(&rtransport->lock); 1813 assert(rtransport->event_channel != NULL); 1814 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 1815 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 1816 port_tmp->ref++; 1817 free(port); 1818 /* Already listening at this address */ 1819 pthread_mutex_unlock(&rtransport->lock); 1820 return 0; 1821 } 1822 } 1823 1824 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 1825 if (rc < 0) { 1826 SPDK_ERRLOG("rdma_create_id() failed\n"); 1827 free(port); 1828 pthread_mutex_unlock(&rtransport->lock); 1829 return rc; 1830 } 1831 1832 switch (port->trid.adrfam) { 1833 case SPDK_NVMF_ADRFAM_IPV4: 1834 family = AF_INET; 1835 break; 1836 case SPDK_NVMF_ADRFAM_IPV6: 1837 family = AF_INET6; 1838 break; 1839 default: 1840 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 1841 free(port); 1842 pthread_mutex_unlock(&rtransport->lock); 1843 return -EINVAL; 1844 } 1845 1846 memset(&hints, 0, sizeof(hints)); 1847 hints.ai_family = family; 1848 hints.ai_flags = AI_NUMERICSERV; 1849 hints.ai_socktype = SOCK_STREAM; 1850 hints.ai_protocol = 0; 1851 1852 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 1853 if (rc) { 1854 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 1855 free(port); 1856 pthread_mutex_unlock(&rtransport->lock); 1857 return -EINVAL; 1858 } 1859 1860 rc = rdma_bind_addr(port->id, res->ai_addr); 1861 freeaddrinfo(res); 1862 1863 if (rc < 0) { 1864 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 1865 rdma_destroy_id(port->id); 1866 free(port); 1867 pthread_mutex_unlock(&rtransport->lock); 1868 return rc; 1869 } 1870 1871 if (!port->id->verbs) { 1872 SPDK_ERRLOG("ibv_context is null\n"); 1873 rdma_destroy_id(port->id); 1874 free(port); 1875 pthread_mutex_unlock(&rtransport->lock); 1876 return -1; 1877 } 1878 1879 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 1880 if (rc < 0) { 1881 SPDK_ERRLOG("rdma_listen() failed\n"); 1882 rdma_destroy_id(port->id); 1883 free(port); 1884 pthread_mutex_unlock(&rtransport->lock); 1885 return rc; 1886 } 1887 1888 TAILQ_FOREACH(device, &rtransport->devices, link) { 1889 if (device->context == port->id->verbs) { 1890 port->device = device; 1891 break; 1892 } 1893 } 1894 if (!port->device) { 1895 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 1896 port->id->verbs); 1897 rdma_destroy_id(port->id); 1898 free(port); 1899 pthread_mutex_unlock(&rtransport->lock); 1900 return -EINVAL; 1901 } 1902 1903 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 1904 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 1905 1906 port->ref = 1; 1907 1908 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 1909 pthread_mutex_unlock(&rtransport->lock); 1910 1911 return 0; 1912 } 1913 1914 static int 1915 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 1916 const struct spdk_nvme_transport_id *_trid) 1917 { 1918 struct spdk_nvmf_rdma_transport *rtransport; 1919 struct spdk_nvmf_rdma_port *port, *tmp; 1920 struct spdk_nvme_transport_id trid = {}; 1921 1922 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1923 1924 /* Selectively copy the trid. Things like NQN don't matter here - that 1925 * mapping is enforced elsewhere. 1926 */ 1927 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 1928 trid.adrfam = _trid->adrfam; 1929 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 1930 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 1931 1932 pthread_mutex_lock(&rtransport->lock); 1933 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 1934 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 1935 assert(port->ref > 0); 1936 port->ref--; 1937 if (port->ref == 0) { 1938 TAILQ_REMOVE(&rtransport->ports, port, link); 1939 rdma_destroy_id(port->id); 1940 free(port); 1941 } 1942 break; 1943 } 1944 } 1945 1946 pthread_mutex_unlock(&rtransport->lock); 1947 return 0; 1948 } 1949 1950 static bool 1951 spdk_nvmf_rdma_qpair_is_idle(struct spdk_nvmf_qpair *qpair) 1952 { 1953 int cur_queue_depth, cur_rdma_rw_depth; 1954 struct spdk_nvmf_rdma_qpair *rqpair; 1955 1956 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1957 cur_queue_depth = spdk_nvmf_rdma_cur_queue_depth(rqpair); 1958 cur_rdma_rw_depth = spdk_nvmf_rdma_cur_rw_depth(rqpair); 1959 1960 if (cur_queue_depth == 0 && cur_rdma_rw_depth == 0) { 1961 return true; 1962 } 1963 return false; 1964 } 1965 1966 static void 1967 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 1968 struct spdk_nvmf_rdma_qpair *rqpair) 1969 { 1970 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 1971 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 1972 1973 /* We process I/O in the data transfer pending queue at the highest priority. */ 1974 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->state_queue[RDMA_REQUEST_STATE_DATA_TRANSFER_PENDING], 1975 state_link, req_tmp) { 1976 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1977 break; 1978 } 1979 } 1980 1981 /* The second highest priority is I/O waiting on memory buffers. */ 1982 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->ch->pending_data_buf_queue, link, 1983 req_tmp) { 1984 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1985 break; 1986 } 1987 } 1988 1989 /* The lowest priority is processing newly received commands */ 1990 TAILQ_FOREACH_SAFE(rdma_recv, &rqpair->incoming_queue, link, recv_tmp) { 1991 if (TAILQ_EMPTY(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE])) { 1992 break; 1993 } 1994 1995 rdma_req = TAILQ_FIRST(&rqpair->state_queue[RDMA_REQUEST_STATE_FREE]); 1996 rdma_req->recv = rdma_recv; 1997 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_NEW); 1998 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 1999 break; 2000 } 2001 } 2002 } 2003 2004 static void 2005 _nvmf_rdma_qpair_disconnect(void *ctx) 2006 { 2007 struct spdk_nvmf_qpair *qpair = ctx; 2008 struct spdk_nvmf_rdma_qpair *rqpair; 2009 2010 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2011 2012 spdk_nvmf_rdma_qpair_dec_refcnt(rqpair); 2013 2014 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2015 } 2016 2017 static void 2018 _nvmf_rdma_disconnect_retry(void *ctx) 2019 { 2020 struct spdk_nvmf_qpair *qpair = ctx; 2021 struct spdk_nvmf_poll_group *group; 2022 2023 /* Read the group out of the qpair. This is normally set and accessed only from 2024 * the thread that created the group. Here, we're not on that thread necessarily. 2025 * The data member qpair->group begins it's life as NULL and then is assigned to 2026 * a pointer and never changes. So fortunately reading this and checking for 2027 * non-NULL is thread safe in the x86_64 memory model. */ 2028 group = qpair->group; 2029 2030 if (group == NULL) { 2031 /* The qpair hasn't been assigned to a group yet, so we can't 2032 * process a disconnect. Send a message to ourself and try again. */ 2033 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_disconnect_retry, qpair); 2034 return; 2035 } 2036 2037 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2038 } 2039 2040 static int 2041 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2042 { 2043 struct spdk_nvmf_qpair *qpair; 2044 struct spdk_nvmf_rdma_qpair *rqpair; 2045 2046 if (evt->id == NULL) { 2047 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2048 return -1; 2049 } 2050 2051 qpair = evt->id->context; 2052 if (qpair == NULL) { 2053 SPDK_ERRLOG("disconnect request: no active connection\n"); 2054 return -1; 2055 } 2056 2057 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2058 2059 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2060 2061 spdk_nvmf_rdma_update_ibv_state(rqpair); 2062 spdk_nvmf_rdma_qpair_inc_refcnt(rqpair); 2063 2064 _nvmf_rdma_disconnect_retry(qpair); 2065 2066 return 0; 2067 } 2068 2069 #ifdef DEBUG 2070 static const char *CM_EVENT_STR[] = { 2071 "RDMA_CM_EVENT_ADDR_RESOLVED", 2072 "RDMA_CM_EVENT_ADDR_ERROR", 2073 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2074 "RDMA_CM_EVENT_ROUTE_ERROR", 2075 "RDMA_CM_EVENT_CONNECT_REQUEST", 2076 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2077 "RDMA_CM_EVENT_CONNECT_ERROR", 2078 "RDMA_CM_EVENT_UNREACHABLE", 2079 "RDMA_CM_EVENT_REJECTED", 2080 "RDMA_CM_EVENT_ESTABLISHED", 2081 "RDMA_CM_EVENT_DISCONNECTED", 2082 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2083 "RDMA_CM_EVENT_MULTICAST_JOIN", 2084 "RDMA_CM_EVENT_MULTICAST_ERROR", 2085 "RDMA_CM_EVENT_ADDR_CHANGE", 2086 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2087 }; 2088 #endif /* DEBUG */ 2089 2090 static void 2091 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2092 { 2093 struct spdk_nvmf_rdma_transport *rtransport; 2094 struct rdma_cm_event *event; 2095 int rc; 2096 2097 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2098 2099 if (rtransport->event_channel == NULL) { 2100 return; 2101 } 2102 2103 while (1) { 2104 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2105 if (rc == 0) { 2106 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2107 2108 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2109 2110 switch (event->event) { 2111 case RDMA_CM_EVENT_ADDR_RESOLVED: 2112 case RDMA_CM_EVENT_ADDR_ERROR: 2113 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2114 case RDMA_CM_EVENT_ROUTE_ERROR: 2115 /* No action required. The target never attempts to resolve routes. */ 2116 break; 2117 case RDMA_CM_EVENT_CONNECT_REQUEST: 2118 rc = nvmf_rdma_connect(transport, event, cb_fn); 2119 if (rc < 0) { 2120 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2121 break; 2122 } 2123 break; 2124 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2125 /* The target never initiates a new connection. So this will not occur. */ 2126 break; 2127 case RDMA_CM_EVENT_CONNECT_ERROR: 2128 /* Can this happen? The docs say it can, but not sure what causes it. */ 2129 break; 2130 case RDMA_CM_EVENT_UNREACHABLE: 2131 case RDMA_CM_EVENT_REJECTED: 2132 /* These only occur on the client side. */ 2133 break; 2134 case RDMA_CM_EVENT_ESTABLISHED: 2135 /* TODO: Should we be waiting for this event anywhere? */ 2136 break; 2137 case RDMA_CM_EVENT_DISCONNECTED: 2138 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2139 rc = nvmf_rdma_disconnect(event); 2140 if (rc < 0) { 2141 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2142 break; 2143 } 2144 break; 2145 case RDMA_CM_EVENT_MULTICAST_JOIN: 2146 case RDMA_CM_EVENT_MULTICAST_ERROR: 2147 /* Multicast is not used */ 2148 break; 2149 case RDMA_CM_EVENT_ADDR_CHANGE: 2150 /* Not utilizing this event */ 2151 break; 2152 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2153 /* For now, do nothing. The target never re-uses queue pairs. */ 2154 break; 2155 default: 2156 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2157 break; 2158 } 2159 2160 rdma_ack_cm_event(event); 2161 } else { 2162 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2163 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2164 } 2165 break; 2166 } 2167 } 2168 } 2169 2170 static void 2171 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2172 { 2173 int rc; 2174 struct spdk_nvmf_rdma_qpair *rqpair; 2175 struct ibv_async_event event; 2176 enum ibv_qp_state state; 2177 2178 rc = ibv_get_async_event(device->context, &event); 2179 2180 if (rc) { 2181 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2182 errno, spdk_strerror(errno)); 2183 return; 2184 } 2185 2186 SPDK_NOTICELOG("Async event: %s\n", 2187 ibv_event_type_str(event.event_type)); 2188 2189 switch (event.event_type) { 2190 case IBV_EVENT_QP_FATAL: 2191 rqpair = event.element.qp->qp_context; 2192 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2193 (uintptr_t)rqpair->cm_id, event.event_type); 2194 spdk_nvmf_rdma_update_ibv_state(rqpair); 2195 spdk_nvmf_rdma_qpair_inc_refcnt(rqpair); 2196 spdk_thread_send_msg(rqpair->qpair.group->thread, _nvmf_rdma_qpair_disconnect, &rqpair->qpair); 2197 break; 2198 case IBV_EVENT_QP_LAST_WQE_REACHED: 2199 /* This event only occurs for shared receive queues, which are not currently supported. */ 2200 break; 2201 case IBV_EVENT_SQ_DRAINED: 2202 /* This event occurs frequently in both error and non-error states. 2203 * Check if the qpair is in an error state before sending a message. 2204 * Note that we're not on the correct thread to access the qpair, but 2205 * the operations that the below calls make all happen to be thread 2206 * safe. */ 2207 rqpair = event.element.qp->qp_context; 2208 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2209 (uintptr_t)rqpair->cm_id, event.event_type); 2210 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2211 if (state == IBV_QPS_ERR) { 2212 spdk_nvmf_rdma_qpair_inc_refcnt(rqpair); 2213 spdk_thread_send_msg(rqpair->qpair.group->thread, _nvmf_rdma_qpair_disconnect, &rqpair->qpair); 2214 } 2215 break; 2216 case IBV_EVENT_QP_REQ_ERR: 2217 case IBV_EVENT_QP_ACCESS_ERR: 2218 case IBV_EVENT_COMM_EST: 2219 case IBV_EVENT_PATH_MIG: 2220 case IBV_EVENT_PATH_MIG_ERR: 2221 rqpair = event.element.qp->qp_context; 2222 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2223 (uintptr_t)rqpair->cm_id, event.event_type); 2224 spdk_nvmf_rdma_update_ibv_state(rqpair); 2225 break; 2226 case IBV_EVENT_CQ_ERR: 2227 case IBV_EVENT_DEVICE_FATAL: 2228 case IBV_EVENT_PORT_ACTIVE: 2229 case IBV_EVENT_PORT_ERR: 2230 case IBV_EVENT_LID_CHANGE: 2231 case IBV_EVENT_PKEY_CHANGE: 2232 case IBV_EVENT_SM_CHANGE: 2233 case IBV_EVENT_SRQ_ERR: 2234 case IBV_EVENT_SRQ_LIMIT_REACHED: 2235 case IBV_EVENT_CLIENT_REREGISTER: 2236 case IBV_EVENT_GID_CHANGE: 2237 default: 2238 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2239 break; 2240 } 2241 ibv_ack_async_event(&event); 2242 } 2243 2244 static void 2245 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2246 { 2247 int nfds, i = 0; 2248 struct spdk_nvmf_rdma_transport *rtransport; 2249 struct spdk_nvmf_rdma_device *device, *tmp; 2250 2251 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2252 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2253 2254 if (nfds <= 0) { 2255 return; 2256 } 2257 2258 /* The first poll descriptor is RDMA CM event */ 2259 if (rtransport->poll_fds[i++].revents & POLLIN) { 2260 spdk_nvmf_process_cm_event(transport, cb_fn); 2261 nfds--; 2262 } 2263 2264 if (nfds == 0) { 2265 return; 2266 } 2267 2268 /* Second and subsequent poll descriptors are IB async events */ 2269 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2270 if (rtransport->poll_fds[i++].revents & POLLIN) { 2271 spdk_nvmf_process_ib_event(device); 2272 nfds--; 2273 } 2274 } 2275 /* check all flagged fd's have been served */ 2276 assert(nfds == 0); 2277 } 2278 2279 static void 2280 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2281 struct spdk_nvme_transport_id *trid, 2282 struct spdk_nvmf_discovery_log_page_entry *entry) 2283 { 2284 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2285 entry->adrfam = trid->adrfam; 2286 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2287 2288 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2289 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2290 2291 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2292 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2293 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2294 } 2295 2296 static struct spdk_nvmf_transport_poll_group * 2297 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2298 { 2299 struct spdk_nvmf_rdma_transport *rtransport; 2300 struct spdk_nvmf_rdma_poll_group *rgroup; 2301 struct spdk_nvmf_rdma_poller *poller; 2302 struct spdk_nvmf_rdma_device *device; 2303 2304 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2305 2306 rgroup = calloc(1, sizeof(*rgroup)); 2307 if (!rgroup) { 2308 return NULL; 2309 } 2310 2311 TAILQ_INIT(&rgroup->pollers); 2312 2313 pthread_mutex_lock(&rtransport->lock); 2314 TAILQ_FOREACH(device, &rtransport->devices, link) { 2315 poller = calloc(1, sizeof(*poller)); 2316 if (!poller) { 2317 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2318 free(rgroup); 2319 pthread_mutex_unlock(&rtransport->lock); 2320 return NULL; 2321 } 2322 2323 poller->device = device; 2324 poller->group = rgroup; 2325 2326 TAILQ_INIT(&poller->qpairs); 2327 2328 poller->cq = ibv_create_cq(device->context, NVMF_RDMA_CQ_SIZE, poller, NULL, 0); 2329 if (!poller->cq) { 2330 SPDK_ERRLOG("Unable to create completion queue\n"); 2331 free(poller); 2332 free(rgroup); 2333 pthread_mutex_unlock(&rtransport->lock); 2334 return NULL; 2335 } 2336 2337 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2338 } 2339 2340 pthread_mutex_unlock(&rtransport->lock); 2341 return &rgroup->group; 2342 } 2343 2344 static void 2345 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2346 { 2347 struct spdk_nvmf_rdma_poll_group *rgroup; 2348 struct spdk_nvmf_rdma_poller *poller, *tmp; 2349 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2350 2351 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2352 2353 if (!rgroup) { 2354 return; 2355 } 2356 2357 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2358 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2359 2360 if (poller->cq) { 2361 ibv_destroy_cq(poller->cq); 2362 } 2363 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 2364 spdk_nvmf_rdma_qpair_destroy(qpair); 2365 } 2366 2367 free(poller); 2368 } 2369 2370 free(rgroup); 2371 } 2372 2373 static int 2374 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 2375 struct spdk_nvmf_qpair *qpair) 2376 { 2377 struct spdk_nvmf_rdma_transport *rtransport; 2378 struct spdk_nvmf_rdma_poll_group *rgroup; 2379 struct spdk_nvmf_rdma_qpair *rqpair; 2380 struct spdk_nvmf_rdma_device *device; 2381 struct spdk_nvmf_rdma_poller *poller; 2382 int rc; 2383 2384 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 2385 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2386 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2387 2388 device = rqpair->port->device; 2389 2390 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 2391 if (poller->device == device) { 2392 break; 2393 } 2394 } 2395 2396 if (!poller) { 2397 SPDK_ERRLOG("No poller found for device.\n"); 2398 return -1; 2399 } 2400 2401 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 2402 rqpair->poller = poller; 2403 2404 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 2405 if (rc < 0) { 2406 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 2407 return -1; 2408 } 2409 2410 rqpair->mgmt_channel = spdk_get_io_channel(rtransport); 2411 if (!rqpair->mgmt_channel) { 2412 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2413 spdk_nvmf_rdma_qpair_destroy(rqpair); 2414 return -1; 2415 } 2416 2417 rqpair->ch = spdk_io_channel_get_ctx(rqpair->mgmt_channel); 2418 assert(rqpair->ch != NULL); 2419 2420 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 2421 if (rc) { 2422 /* Try to reject, but we probably can't */ 2423 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 2424 spdk_nvmf_rdma_qpair_destroy(rqpair); 2425 return -1; 2426 } 2427 2428 spdk_nvmf_rdma_update_ibv_state(rqpair); 2429 2430 return 0; 2431 } 2432 2433 static int 2434 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 2435 { 2436 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2437 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2438 struct spdk_nvmf_rdma_transport, transport); 2439 2440 if (rdma_req->data_from_pool) { 2441 /* Put the buffer/s back in the pool */ 2442 for (uint32_t i = 0; i < rdma_req->req.iovcnt; i++) { 2443 spdk_mempool_put(rtransport->data_buf_pool, rdma_req->data.buffers[i]); 2444 rdma_req->req.iov[i].iov_base = NULL; 2445 rdma_req->data.buffers[i] = NULL; 2446 } 2447 rdma_req->data_from_pool = false; 2448 } 2449 rdma_req->req.length = 0; 2450 rdma_req->req.iovcnt = 0; 2451 rdma_req->req.data = NULL; 2452 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_FREE); 2453 return 0; 2454 } 2455 2456 static int 2457 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 2458 { 2459 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 2460 struct spdk_nvmf_rdma_transport, transport); 2461 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 2462 struct spdk_nvmf_rdma_request, req); 2463 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 2464 struct spdk_nvmf_rdma_qpair, qpair); 2465 2466 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2467 /* The connection is alive, so process the request as normal */ 2468 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_EXECUTED); 2469 } else { 2470 /* The connection is dead. Move the request directly to the completed state. */ 2471 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2472 } 2473 2474 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2475 2476 return 0; 2477 } 2478 2479 static void 2480 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 2481 { 2482 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2483 struct ibv_recv_wr recv_wr = {}; 2484 struct ibv_recv_wr *bad_recv_wr; 2485 struct ibv_send_wr send_wr = {}; 2486 struct ibv_send_wr *bad_send_wr; 2487 int rc; 2488 2489 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 2490 return; 2491 } 2492 2493 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 2494 2495 if (rqpair->ibv_attr.qp_state != IBV_QPS_ERR) { 2496 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 2497 } 2498 2499 rqpair->drain_recv_wr.type = RDMA_WR_TYPE_DRAIN_RECV; 2500 recv_wr.wr_id = (uintptr_t)&rqpair->drain_recv_wr; 2501 rc = ibv_post_recv(rqpair->cm_id->qp, &recv_wr, &bad_recv_wr); 2502 if (rc) { 2503 SPDK_ERRLOG("Failed to post dummy receive WR, errno %d\n", errno); 2504 assert(false); 2505 return; 2506 } 2507 2508 rqpair->drain_send_wr.type = RDMA_WR_TYPE_DRAIN_SEND; 2509 send_wr.wr_id = (uintptr_t)&rqpair->drain_send_wr; 2510 send_wr.opcode = IBV_WR_SEND; 2511 rc = ibv_post_send(rqpair->cm_id->qp, &send_wr, &bad_send_wr); 2512 if (rc) { 2513 SPDK_ERRLOG("Failed to post dummy send WR, errno %d\n", errno); 2514 assert(false); 2515 return; 2516 } 2517 } 2518 2519 #ifdef DEBUG 2520 static int 2521 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 2522 { 2523 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 2524 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 2525 } 2526 #endif 2527 2528 static int 2529 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 2530 struct spdk_nvmf_rdma_poller *rpoller) 2531 { 2532 struct ibv_wc wc[32]; 2533 struct spdk_nvmf_rdma_wr *rdma_wr; 2534 struct spdk_nvmf_rdma_request *rdma_req; 2535 struct spdk_nvmf_rdma_recv *rdma_recv; 2536 struct spdk_nvmf_rdma_qpair *rqpair; 2537 int reaped, i; 2538 int count = 0; 2539 bool error = false; 2540 2541 /* Poll for completing operations. */ 2542 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 2543 if (reaped < 0) { 2544 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 2545 errno, spdk_strerror(errno)); 2546 return -1; 2547 } 2548 2549 for (i = 0; i < reaped; i++) { 2550 2551 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 2552 2553 /* Handle error conditions */ 2554 if (wc[i].status) { 2555 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 2556 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 2557 2558 error = true; 2559 2560 switch (rdma_wr->type) { 2561 case RDMA_WR_TYPE_SEND: 2562 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 2563 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2564 2565 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 2566 /* We're going to attempt an error recovery, so force the request into 2567 * the completed state. */ 2568 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2569 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2570 break; 2571 case RDMA_WR_TYPE_RECV: 2572 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 2573 rqpair = rdma_recv->qpair; 2574 2575 /* Dump this into the incoming queue. This gets cleaned up when 2576 * the queue pair disconnects or recovers. */ 2577 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2578 break; 2579 case RDMA_WR_TYPE_DATA: 2580 /* If the data transfer fails still force the queue into the error state, 2581 * but the rdma_req objects should only be manipulated in response to 2582 * SEND and RECV operations. */ 2583 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2584 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 2585 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2586 break; 2587 case RDMA_WR_TYPE_DRAIN_RECV: 2588 rqpair = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_qpair, drain_recv_wr); 2589 assert(rqpair->disconnect_flags & RDMA_QP_DISCONNECTING); 2590 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Drained QP RECV %u (%p)\n", rqpair->qpair.qid, rqpair); 2591 rqpair->disconnect_flags |= RDMA_QP_RECV_DRAINED; 2592 if (rqpair->disconnect_flags & RDMA_QP_SEND_DRAINED) { 2593 spdk_nvmf_rdma_qpair_destroy(rqpair); 2594 } 2595 /* Continue so that this does not trigger the disconnect path below. */ 2596 continue; 2597 case RDMA_WR_TYPE_DRAIN_SEND: 2598 rqpair = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_qpair, drain_send_wr); 2599 assert(rqpair->disconnect_flags & RDMA_QP_DISCONNECTING); 2600 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Drained QP SEND %u (%p)\n", rqpair->qpair.qid, rqpair); 2601 rqpair->disconnect_flags |= RDMA_QP_SEND_DRAINED; 2602 if (rqpair->disconnect_flags & RDMA_QP_RECV_DRAINED) { 2603 spdk_nvmf_rdma_qpair_destroy(rqpair); 2604 } 2605 /* Continue so that this does not trigger the disconnect path below. */ 2606 continue; 2607 default: 2608 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2609 continue; 2610 } 2611 2612 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 2613 /* Disconnect the connection. */ 2614 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2615 } 2616 continue; 2617 } 2618 2619 switch (wc[i].opcode) { 2620 case IBV_WC_SEND: 2621 assert(rdma_wr->type == RDMA_WR_TYPE_SEND); 2622 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 2623 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2624 2625 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 2626 2627 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_COMPLETED); 2628 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2629 2630 count++; 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_WRITE: 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 /* Try to process other queued requests */ 2642 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2643 break; 2644 2645 case IBV_WC_RDMA_READ: 2646 assert(rdma_wr->type == RDMA_WR_TYPE_DATA); 2647 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 2648 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2649 2650 assert(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2651 spdk_nvmf_rdma_request_set_state(rdma_req, RDMA_REQUEST_STATE_READY_TO_EXECUTE); 2652 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 2653 2654 /* Try to process other queued requests */ 2655 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2656 break; 2657 2658 case IBV_WC_RECV: 2659 assert(rdma_wr->type == RDMA_WR_TYPE_RECV); 2660 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 2661 rqpair = rdma_recv->qpair; 2662 2663 TAILQ_INSERT_TAIL(&rqpair->incoming_queue, rdma_recv, link); 2664 /* Try to process other queued requests */ 2665 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair); 2666 break; 2667 2668 default: 2669 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 2670 continue; 2671 } 2672 } 2673 2674 if (error == true) { 2675 return -1; 2676 } 2677 2678 return count; 2679 } 2680 2681 static int 2682 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 2683 { 2684 struct spdk_nvmf_rdma_transport *rtransport; 2685 struct spdk_nvmf_rdma_poll_group *rgroup; 2686 struct spdk_nvmf_rdma_poller *rpoller; 2687 int count, rc; 2688 2689 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 2690 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2691 2692 count = 0; 2693 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2694 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 2695 if (rc < 0) { 2696 return rc; 2697 } 2698 count += rc; 2699 } 2700 2701 return count; 2702 } 2703 2704 static int 2705 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2706 struct spdk_nvme_transport_id *trid, 2707 bool peer) 2708 { 2709 struct sockaddr *saddr; 2710 uint16_t port; 2711 2712 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 2713 2714 if (peer) { 2715 saddr = rdma_get_peer_addr(id); 2716 } else { 2717 saddr = rdma_get_local_addr(id); 2718 } 2719 switch (saddr->sa_family) { 2720 case AF_INET: { 2721 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 2722 2723 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 2724 inet_ntop(AF_INET, &saddr_in->sin_addr, 2725 trid->traddr, sizeof(trid->traddr)); 2726 if (peer) { 2727 port = ntohs(rdma_get_dst_port(id)); 2728 } else { 2729 port = ntohs(rdma_get_src_port(id)); 2730 } 2731 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 2732 break; 2733 } 2734 case AF_INET6: { 2735 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 2736 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 2737 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 2738 trid->traddr, sizeof(trid->traddr)); 2739 if (peer) { 2740 port = ntohs(rdma_get_dst_port(id)); 2741 } else { 2742 port = ntohs(rdma_get_src_port(id)); 2743 } 2744 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 2745 break; 2746 } 2747 default: 2748 return -1; 2749 2750 } 2751 2752 return 0; 2753 } 2754 2755 static int 2756 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 2757 struct spdk_nvme_transport_id *trid) 2758 { 2759 struct spdk_nvmf_rdma_qpair *rqpair; 2760 2761 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2762 2763 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 2764 } 2765 2766 static int 2767 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 2768 struct spdk_nvme_transport_id *trid) 2769 { 2770 struct spdk_nvmf_rdma_qpair *rqpair; 2771 2772 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2773 2774 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 2775 } 2776 2777 static int 2778 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 2779 struct spdk_nvme_transport_id *trid) 2780 { 2781 struct spdk_nvmf_rdma_qpair *rqpair; 2782 2783 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2784 2785 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 2786 } 2787 2788 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 2789 .type = SPDK_NVME_TRANSPORT_RDMA, 2790 .opts_init = spdk_nvmf_rdma_opts_init, 2791 .create = spdk_nvmf_rdma_create, 2792 .destroy = spdk_nvmf_rdma_destroy, 2793 2794 .listen = spdk_nvmf_rdma_listen, 2795 .stop_listen = spdk_nvmf_rdma_stop_listen, 2796 .accept = spdk_nvmf_rdma_accept, 2797 2798 .listener_discover = spdk_nvmf_rdma_discover, 2799 2800 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 2801 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 2802 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 2803 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 2804 2805 .req_free = spdk_nvmf_rdma_request_free, 2806 .req_complete = spdk_nvmf_rdma_request_complete, 2807 2808 .qpair_fini = spdk_nvmf_rdma_close_qpair, 2809 .qpair_is_idle = spdk_nvmf_rdma_qpair_is_idle, 2810 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 2811 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 2812 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 2813 2814 }; 2815 2816 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 2817