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