1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2018 Mellanox Technologies LTD. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of Intel Corporation nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include "spdk/stdinc.h" 35 36 #include <infiniband/verbs.h> 37 #include <rdma/rdma_cma.h> 38 #include <rdma/rdma_verbs.h> 39 40 #include "nvmf_internal.h" 41 #include "transport.h" 42 43 #include "spdk/config.h" 44 #include "spdk/assert.h" 45 #include "spdk/thread.h" 46 #include "spdk/nvmf.h" 47 #include "spdk/nvmf_spec.h" 48 #include "spdk/string.h" 49 #include "spdk/trace.h" 50 #include "spdk/util.h" 51 52 #include "spdk_internal/log.h" 53 54 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 55 56 /* 57 RDMA Connection Resource Defaults 58 */ 59 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 60 #define NVMF_DEFAULT_RSP_SGE 1 61 #define NVMF_DEFAULT_RX_SGE 2 62 63 /* The RDMA completion queue size */ 64 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 65 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 66 67 /* Timeout for destroying defunct rqpairs */ 68 #define NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US 4000000 69 70 /* The maximum number of buffers per request */ 71 #define NVMF_REQ_MAX_BUFFERS (SPDK_NVMF_MAX_SGL_ENTRIES * 2) 72 73 static int g_spdk_nvmf_ibv_query_mask = 74 IBV_QP_STATE | 75 IBV_QP_PKEY_INDEX | 76 IBV_QP_PORT | 77 IBV_QP_ACCESS_FLAGS | 78 IBV_QP_AV | 79 IBV_QP_PATH_MTU | 80 IBV_QP_DEST_QPN | 81 IBV_QP_RQ_PSN | 82 IBV_QP_MAX_DEST_RD_ATOMIC | 83 IBV_QP_MIN_RNR_TIMER | 84 IBV_QP_SQ_PSN | 85 IBV_QP_TIMEOUT | 86 IBV_QP_RETRY_CNT | 87 IBV_QP_RNR_RETRY | 88 IBV_QP_MAX_QP_RD_ATOMIC; 89 90 enum spdk_nvmf_rdma_request_state { 91 /* The request is not currently in use */ 92 RDMA_REQUEST_STATE_FREE = 0, 93 94 /* Initial state when request first received */ 95 RDMA_REQUEST_STATE_NEW, 96 97 /* The request is queued until a data buffer is available. */ 98 RDMA_REQUEST_STATE_NEED_BUFFER, 99 100 /* The request is waiting on RDMA queue depth availability 101 * to transfer data from the host to the controller. 102 */ 103 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 104 105 /* The request is currently transferring data from the host to the controller. */ 106 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 107 108 /* The request is ready to execute at the block device */ 109 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 110 111 /* The request is currently executing at the block device */ 112 RDMA_REQUEST_STATE_EXECUTING, 113 114 /* The request finished executing at the block device */ 115 RDMA_REQUEST_STATE_EXECUTED, 116 117 /* The request is waiting on RDMA queue depth availability 118 * to transfer data from the controller to the host. 119 */ 120 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 121 122 /* The request is ready to send a completion */ 123 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 124 125 /* The request is currently transferring data from the controller to the host. */ 126 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 127 128 /* The request currently has an outstanding completion without an 129 * associated data transfer. 130 */ 131 RDMA_REQUEST_STATE_COMPLETING, 132 133 /* The request completed and can be marked free. */ 134 RDMA_REQUEST_STATE_COMPLETED, 135 136 /* Terminator */ 137 RDMA_REQUEST_NUM_STATES, 138 }; 139 140 #define OBJECT_NVMF_RDMA_IO 0x40 141 142 #define TRACE_GROUP_NVMF_RDMA 0x4 143 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 144 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 145 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 146 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 147 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 148 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 149 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 150 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 151 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 152 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 153 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 154 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 155 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 156 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 157 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 158 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 159 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 160 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 161 162 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 163 { 164 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 165 spdk_trace_register_description("RDMA_REQ_NEW", "", 166 TRACE_RDMA_REQUEST_STATE_NEW, 167 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 168 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", "", 169 TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 170 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 171 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C_TO_H", "", 172 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 173 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 174 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H_TO_C", "", 175 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 176 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 177 spdk_trace_register_description("RDMA_REQ_TX_H_TO_C", "", 178 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 179 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 180 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", "", 181 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 182 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 183 spdk_trace_register_description("RDMA_REQ_EXECUTING", "", 184 TRACE_RDMA_REQUEST_STATE_EXECUTING, 185 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 186 spdk_trace_register_description("RDMA_REQ_EXECUTED", "", 187 TRACE_RDMA_REQUEST_STATE_EXECUTED, 188 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 189 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPLETE", "", 190 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 191 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 192 spdk_trace_register_description("RDMA_REQ_COMPLETING_CONTROLLER_TO_HOST", "", 193 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 194 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 195 spdk_trace_register_description("RDMA_REQ_COMPLETING_INCAPSULE", "", 196 TRACE_RDMA_REQUEST_STATE_COMPLETING, 197 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 198 spdk_trace_register_description("RDMA_REQ_COMPLETED", "", 199 TRACE_RDMA_REQUEST_STATE_COMPLETED, 200 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 201 202 spdk_trace_register_description("RDMA_QP_CREATE", "", TRACE_RDMA_QP_CREATE, 203 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 204 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", "", TRACE_RDMA_IBV_ASYNC_EVENT, 205 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 206 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", "", TRACE_RDMA_CM_ASYNC_EVENT, 207 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 208 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", "", TRACE_RDMA_QP_STATE_CHANGE, 209 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 210 spdk_trace_register_description("RDMA_QP_DISCONNECT", "", TRACE_RDMA_QP_DISCONNECT, 211 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 212 spdk_trace_register_description("RDMA_QP_DESTROY", "", TRACE_RDMA_QP_DESTROY, 213 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 214 } 215 216 enum spdk_nvmf_rdma_wr_type { 217 RDMA_WR_TYPE_RECV, 218 RDMA_WR_TYPE_SEND, 219 RDMA_WR_TYPE_DATA, 220 }; 221 222 struct spdk_nvmf_rdma_wr { 223 enum spdk_nvmf_rdma_wr_type type; 224 }; 225 226 /* This structure holds commands as they are received off the wire. 227 * It must be dynamically paired with a full request object 228 * (spdk_nvmf_rdma_request) to service a request. It is separate 229 * from the request because RDMA does not appear to order 230 * completions, so occasionally we'll get a new incoming 231 * command when there aren't any free request objects. 232 */ 233 struct spdk_nvmf_rdma_recv { 234 struct ibv_recv_wr wr; 235 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 236 237 struct spdk_nvmf_rdma_qpair *qpair; 238 239 /* In-capsule data buffer */ 240 uint8_t *buf; 241 242 struct spdk_nvmf_rdma_wr rdma_wr; 243 244 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 245 }; 246 247 struct spdk_nvmf_rdma_request_data { 248 struct spdk_nvmf_rdma_wr rdma_wr; 249 struct ibv_send_wr wr; 250 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 251 }; 252 253 struct spdk_nvmf_rdma_request { 254 struct spdk_nvmf_request req; 255 bool data_from_pool; 256 257 enum spdk_nvmf_rdma_request_state state; 258 259 struct spdk_nvmf_rdma_recv *recv; 260 261 struct { 262 struct spdk_nvmf_rdma_wr rdma_wr; 263 struct ibv_send_wr wr; 264 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 265 } rsp; 266 267 struct spdk_nvmf_rdma_request_data data; 268 void *buffers[NVMF_REQ_MAX_BUFFERS]; 269 270 uint32_t num_outstanding_data_wr; 271 272 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 273 }; 274 275 enum spdk_nvmf_rdma_qpair_disconnect_flags { 276 RDMA_QP_DISCONNECTING = 1, 277 RDMA_QP_RECV_DRAINED = 1 << 1, 278 RDMA_QP_SEND_DRAINED = 1 << 2 279 }; 280 281 struct spdk_nvmf_rdma_resource_opts { 282 struct spdk_nvmf_rdma_qpair *qpair; 283 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 284 void *qp; 285 struct ibv_pd *pd; 286 uint32_t max_queue_depth; 287 uint32_t in_capsule_data_size; 288 bool shared; 289 }; 290 291 struct spdk_nvmf_rdma_resources { 292 /* Array of size "max_queue_depth" containing RDMA requests. */ 293 struct spdk_nvmf_rdma_request *reqs; 294 295 /* Array of size "max_queue_depth" containing RDMA recvs. */ 296 struct spdk_nvmf_rdma_recv *recvs; 297 298 /* Array of size "max_queue_depth" containing 64 byte capsules 299 * used for receive. 300 */ 301 union nvmf_h2c_msg *cmds; 302 struct ibv_mr *cmds_mr; 303 304 /* Array of size "max_queue_depth" containing 16 byte completions 305 * to be sent back to the user. 306 */ 307 union nvmf_c2h_msg *cpls; 308 struct ibv_mr *cpls_mr; 309 310 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 311 * buffers to be used for in capsule data. 312 */ 313 void *bufs; 314 struct ibv_mr *bufs_mr; 315 316 /* Receives that are waiting for a request object */ 317 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 318 319 /* Queue to track free requests */ 320 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 321 }; 322 323 struct spdk_nvmf_rdma_qpair { 324 struct spdk_nvmf_qpair qpair; 325 326 struct spdk_nvmf_rdma_port *port; 327 struct spdk_nvmf_rdma_poller *poller; 328 329 struct rdma_cm_id *cm_id; 330 struct ibv_srq *srq; 331 struct rdma_cm_id *listen_id; 332 333 /* The maximum number of I/O outstanding on this connection at one time */ 334 uint16_t max_queue_depth; 335 336 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 337 uint16_t max_read_depth; 338 339 /* The maximum number of RDMA SEND operations at one time */ 340 uint32_t max_send_depth; 341 342 /* The current number of outstanding WRs from this qpair's 343 * recv queue. Should not exceed device->attr.max_queue_depth. 344 */ 345 uint16_t current_recv_depth; 346 347 /* The current number of active RDMA READ operations */ 348 uint16_t current_read_depth; 349 350 /* The current number of posted WRs from this qpair's 351 * send queue. Should not exceed max_send_depth. 352 */ 353 uint32_t current_send_depth; 354 355 /* The maximum number of SGEs per WR on the send queue */ 356 uint32_t max_send_sge; 357 358 /* The maximum number of SGEs per WR on the recv queue */ 359 uint32_t max_recv_sge; 360 361 struct spdk_nvmf_rdma_resources *resources; 362 363 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 364 365 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 366 367 /* Number of requests not in the free state */ 368 uint32_t qd; 369 370 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 371 372 /* IBV queue pair attributes: they are used to manage 373 * qp state and recover from errors. 374 */ 375 enum ibv_qp_state ibv_state; 376 377 uint32_t disconnect_flags; 378 379 /* Poller registered in case the qpair doesn't properly 380 * complete the qpair destruct process and becomes defunct. 381 */ 382 383 struct spdk_poller *destruct_poller; 384 385 /* There are several ways a disconnect can start on a qpair 386 * and they are not all mutually exclusive. It is important 387 * that we only initialize one of these paths. 388 */ 389 bool disconnect_started; 390 /* Lets us know that we have received the last_wqe event. */ 391 bool last_wqe_reached; 392 }; 393 394 struct spdk_nvmf_rdma_poller { 395 struct spdk_nvmf_rdma_device *device; 396 struct spdk_nvmf_rdma_poll_group *group; 397 398 int num_cqe; 399 int required_num_wr; 400 struct ibv_cq *cq; 401 402 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 403 uint16_t max_srq_depth; 404 405 /* Shared receive queue */ 406 struct ibv_srq *srq; 407 408 struct spdk_nvmf_rdma_resources *resources; 409 410 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 411 412 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 413 }; 414 415 struct spdk_nvmf_rdma_poll_group { 416 struct spdk_nvmf_transport_poll_group group; 417 418 /* Requests that are waiting to obtain a data buffer */ 419 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_data_buf_queue; 420 421 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 422 }; 423 424 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 425 struct spdk_nvmf_rdma_device { 426 struct ibv_device_attr attr; 427 struct ibv_context *context; 428 429 struct spdk_mem_map *map; 430 struct ibv_pd *pd; 431 432 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 433 }; 434 435 struct spdk_nvmf_rdma_port { 436 struct spdk_nvme_transport_id trid; 437 struct rdma_cm_id *id; 438 struct spdk_nvmf_rdma_device *device; 439 uint32_t ref; 440 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 441 }; 442 443 struct spdk_nvmf_rdma_transport { 444 struct spdk_nvmf_transport transport; 445 446 struct rdma_event_channel *event_channel; 447 448 struct spdk_mempool *data_wr_pool; 449 450 pthread_mutex_t lock; 451 452 /* fields used to poll RDMA/IB events */ 453 nfds_t npoll_fds; 454 struct pollfd *poll_fds; 455 456 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 457 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 458 }; 459 460 static inline int 461 spdk_nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 462 { 463 switch (state) { 464 case IBV_QPS_RESET: 465 case IBV_QPS_INIT: 466 case IBV_QPS_RTR: 467 case IBV_QPS_RTS: 468 case IBV_QPS_SQD: 469 case IBV_QPS_SQE: 470 case IBV_QPS_ERR: 471 return 0; 472 default: 473 return -1; 474 } 475 } 476 477 static enum ibv_qp_state 478 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 479 enum ibv_qp_state old_state, new_state; 480 struct ibv_qp_attr qp_attr; 481 struct ibv_qp_init_attr init_attr; 482 int rc; 483 484 old_state = rqpair->ibv_state; 485 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 486 g_spdk_nvmf_ibv_query_mask, &init_attr); 487 488 if (rc) 489 { 490 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 491 return IBV_QPS_ERR + 1; 492 } 493 494 new_state = qp_attr.qp_state; 495 rqpair->ibv_state = new_state; 496 qp_attr.ah_attr.port_num = qp_attr.port_num; 497 498 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 499 if (rc) 500 { 501 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 502 /* 503 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 504 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 505 */ 506 return IBV_QPS_ERR + 1; 507 } 508 509 if (old_state != new_state) 510 { 511 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 512 (uintptr_t)rqpair->cm_id, new_state); 513 } 514 return new_state; 515 } 516 517 static const char *str_ibv_qp_state[] = { 518 "IBV_QPS_RESET", 519 "IBV_QPS_INIT", 520 "IBV_QPS_RTR", 521 "IBV_QPS_RTS", 522 "IBV_QPS_SQD", 523 "IBV_QPS_SQE", 524 "IBV_QPS_ERR", 525 "IBV_QPS_UNKNOWN" 526 }; 527 528 static int 529 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 530 enum ibv_qp_state new_state) 531 { 532 struct ibv_qp_attr qp_attr; 533 struct ibv_qp_init_attr init_attr; 534 int rc; 535 enum ibv_qp_state state; 536 static int attr_mask_rc[] = { 537 [IBV_QPS_RESET] = IBV_QP_STATE, 538 [IBV_QPS_INIT] = (IBV_QP_STATE | 539 IBV_QP_PKEY_INDEX | 540 IBV_QP_PORT | 541 IBV_QP_ACCESS_FLAGS), 542 [IBV_QPS_RTR] = (IBV_QP_STATE | 543 IBV_QP_AV | 544 IBV_QP_PATH_MTU | 545 IBV_QP_DEST_QPN | 546 IBV_QP_RQ_PSN | 547 IBV_QP_MAX_DEST_RD_ATOMIC | 548 IBV_QP_MIN_RNR_TIMER), 549 [IBV_QPS_RTS] = (IBV_QP_STATE | 550 IBV_QP_SQ_PSN | 551 IBV_QP_TIMEOUT | 552 IBV_QP_RETRY_CNT | 553 IBV_QP_RNR_RETRY | 554 IBV_QP_MAX_QP_RD_ATOMIC), 555 [IBV_QPS_SQD] = IBV_QP_STATE, 556 [IBV_QPS_SQE] = IBV_QP_STATE, 557 [IBV_QPS_ERR] = IBV_QP_STATE, 558 }; 559 560 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 561 if (rc) { 562 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 563 rqpair->qpair.qid, new_state); 564 return rc; 565 } 566 567 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 568 g_spdk_nvmf_ibv_query_mask, &init_attr); 569 570 if (rc) { 571 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 572 assert(false); 573 } 574 575 qp_attr.cur_qp_state = rqpair->ibv_state; 576 qp_attr.qp_state = new_state; 577 578 rc = ibv_modify_qp(rqpair->cm_id->qp, &qp_attr, 579 attr_mask_rc[new_state]); 580 581 if (rc) { 582 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 583 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 584 return rc; 585 } 586 587 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 588 589 if (state != new_state) { 590 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 591 rqpair->qpair.qid, str_ibv_qp_state[new_state], 592 str_ibv_qp_state[state]); 593 return -1; 594 } 595 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 596 str_ibv_qp_state[state]); 597 return 0; 598 } 599 600 static void 601 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 602 struct spdk_nvmf_rdma_transport *rtransport) 603 { 604 struct spdk_nvmf_rdma_request_data *current_data_wr = NULL, *next_data_wr = NULL; 605 struct ibv_send_wr *send_wr; 606 int i; 607 608 rdma_req->num_outstanding_data_wr = 0; 609 current_data_wr = &rdma_req->data; 610 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 611 current_data_wr->wr.sg_list[i].addr = 0; 612 current_data_wr->wr.sg_list[i].length = 0; 613 current_data_wr->wr.sg_list[i].lkey = 0; 614 } 615 current_data_wr->wr.num_sge = 0; 616 617 send_wr = current_data_wr->wr.next; 618 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr) { 619 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 620 } 621 while (next_data_wr) { 622 current_data_wr = next_data_wr; 623 send_wr = current_data_wr->wr.next; 624 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr) { 625 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 626 } else { 627 next_data_wr = NULL; 628 } 629 630 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 631 current_data_wr->wr.sg_list[i].addr = 0; 632 current_data_wr->wr.sg_list[i].length = 0; 633 current_data_wr->wr.sg_list[i].lkey = 0; 634 } 635 current_data_wr->wr.num_sge = 0; 636 spdk_mempool_put(rtransport->data_wr_pool, current_data_wr); 637 } 638 } 639 640 static void 641 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 642 { 643 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->data_from_pool); 644 if (req->req.cmd) { 645 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 646 } 647 if (req->recv) { 648 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 649 } 650 } 651 652 static void 653 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 654 { 655 int i; 656 657 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 658 for (i = 0; i < rqpair->max_queue_depth; i++) { 659 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 660 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 661 } 662 } 663 } 664 665 static void 666 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 667 { 668 if (resources->cmds_mr) { 669 ibv_dereg_mr(resources->cmds_mr); 670 } 671 672 if (resources->cpls_mr) { 673 ibv_dereg_mr(resources->cpls_mr); 674 } 675 676 if (resources->bufs_mr) { 677 ibv_dereg_mr(resources->bufs_mr); 678 } 679 680 spdk_dma_free(resources->cmds); 681 spdk_dma_free(resources->cpls); 682 spdk_dma_free(resources->bufs); 683 free(resources->reqs); 684 free(resources->recvs); 685 free(resources); 686 } 687 688 689 static struct spdk_nvmf_rdma_resources * 690 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 691 { 692 struct spdk_nvmf_rdma_resources *resources; 693 struct spdk_nvmf_rdma_request *rdma_req; 694 struct spdk_nvmf_rdma_recv *rdma_recv; 695 struct ibv_qp *qp; 696 struct ibv_srq *srq; 697 uint32_t i; 698 int rc; 699 700 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 701 if (!resources) { 702 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 703 return NULL; 704 } 705 706 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 707 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 708 resources->cmds = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 709 0x1000, NULL); 710 resources->cpls = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 711 0x1000, NULL); 712 713 if (opts->in_capsule_data_size > 0) { 714 resources->bufs = spdk_dma_zmalloc(opts->max_queue_depth * 715 opts->in_capsule_data_size, 716 0x1000, NULL); 717 } 718 719 if (!resources->reqs || !resources->recvs || !resources->cmds || 720 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 721 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 722 goto cleanup; 723 } 724 725 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 726 opts->max_queue_depth * sizeof(*resources->cmds), 727 IBV_ACCESS_LOCAL_WRITE); 728 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 729 opts->max_queue_depth * sizeof(*resources->cpls), 730 0); 731 732 if (opts->in_capsule_data_size) { 733 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 734 opts->max_queue_depth * 735 opts->in_capsule_data_size, 736 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 737 } 738 739 if (!resources->cmds_mr || !resources->cpls_mr || 740 (opts->in_capsule_data_size && 741 !resources->bufs_mr)) { 742 goto cleanup; 743 } 744 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 745 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 746 resources->cmds_mr->lkey); 747 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 748 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 749 resources->cpls_mr->lkey); 750 if (resources->bufs && resources->bufs_mr) { 751 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 752 resources->bufs, opts->max_queue_depth * 753 opts->in_capsule_data_size, resources->bufs_mr->lkey); 754 } 755 756 /* Initialize queues */ 757 STAILQ_INIT(&resources->incoming_queue); 758 STAILQ_INIT(&resources->free_queue); 759 760 for (i = 0; i < opts->max_queue_depth; i++) { 761 struct ibv_recv_wr *bad_wr = NULL; 762 763 rdma_recv = &resources->recvs[i]; 764 rdma_recv->qpair = opts->qpair; 765 766 /* Set up memory to receive commands */ 767 if (resources->bufs) { 768 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 769 opts->in_capsule_data_size)); 770 } 771 772 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 773 774 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 775 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 776 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 777 rdma_recv->wr.num_sge = 1; 778 779 if (rdma_recv->buf && resources->bufs_mr) { 780 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 781 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 782 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 783 rdma_recv->wr.num_sge++; 784 } 785 786 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 787 rdma_recv->wr.sg_list = rdma_recv->sgl; 788 if (opts->shared) { 789 srq = (struct ibv_srq *)opts->qp; 790 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 791 } else { 792 qp = (struct ibv_qp *)opts->qp; 793 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 794 } 795 if (rc) { 796 goto cleanup; 797 } 798 } 799 800 for (i = 0; i < opts->max_queue_depth; i++) { 801 rdma_req = &resources->reqs[i]; 802 803 if (opts->qpair != NULL) { 804 rdma_req->req.qpair = &opts->qpair->qpair; 805 } else { 806 rdma_req->req.qpair = NULL; 807 } 808 rdma_req->req.cmd = NULL; 809 810 /* Set up memory to send responses */ 811 rdma_req->req.rsp = &resources->cpls[i]; 812 813 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 814 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 815 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 816 817 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 818 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 819 rdma_req->rsp.wr.next = NULL; 820 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 821 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 822 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 823 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 824 825 /* Set up memory for data buffers */ 826 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 827 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 828 rdma_req->data.wr.next = NULL; 829 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 830 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 831 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 832 833 /* Initialize request state to FREE */ 834 rdma_req->state = RDMA_REQUEST_STATE_FREE; 835 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 836 } 837 838 return resources; 839 840 cleanup: 841 nvmf_rdma_resources_destroy(resources); 842 return NULL; 843 } 844 845 static void 846 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 847 { 848 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 849 struct ibv_recv_wr *bad_recv_wr = NULL; 850 int rc; 851 852 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 853 854 spdk_poller_unregister(&rqpair->destruct_poller); 855 856 if (rqpair->qd != 0) { 857 if (rqpair->srq == NULL) { 858 nvmf_rdma_dump_qpair_contents(rqpair); 859 } 860 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 861 } 862 863 if (rqpair->poller) { 864 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 865 866 if (rqpair->srq != NULL) { 867 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 868 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 869 if (rqpair == rdma_recv->qpair) { 870 STAILQ_REMOVE_HEAD(&rqpair->resources->incoming_queue, link); 871 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 872 if (rc) { 873 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 874 } 875 } 876 } 877 } 878 } 879 880 if (rqpair->cm_id) { 881 rdma_destroy_qp(rqpair->cm_id); 882 rdma_destroy_id(rqpair->cm_id); 883 884 if (rqpair->poller != NULL && rqpair->srq == NULL) { 885 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 886 } 887 } 888 889 if (rqpair->srq == NULL) { 890 nvmf_rdma_resources_destroy(rqpair->resources); 891 } 892 893 free(rqpair); 894 } 895 896 static int 897 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 898 { 899 struct spdk_nvmf_rdma_poller *rpoller; 900 int rc, num_cqe, required_num_wr; 901 902 /* Enlarge CQ size dynamically */ 903 rpoller = rqpair->poller; 904 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 905 num_cqe = rpoller->num_cqe; 906 if (num_cqe < required_num_wr) { 907 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 908 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 909 } 910 911 if (rpoller->num_cqe != num_cqe) { 912 if (required_num_wr > device->attr.max_cqe) { 913 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 914 required_num_wr, device->attr.max_cqe); 915 return -1; 916 } 917 918 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 919 rc = ibv_resize_cq(rpoller->cq, num_cqe); 920 if (rc) { 921 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 922 return -1; 923 } 924 925 rpoller->num_cqe = num_cqe; 926 } 927 928 rpoller->required_num_wr = required_num_wr; 929 return 0; 930 } 931 932 static int 933 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 934 { 935 struct spdk_nvmf_rdma_qpair *rqpair; 936 int rc; 937 struct spdk_nvmf_rdma_transport *rtransport; 938 struct spdk_nvmf_transport *transport; 939 struct spdk_nvmf_rdma_resource_opts opts; 940 struct spdk_nvmf_rdma_device *device; 941 struct ibv_qp_init_attr ibv_init_attr; 942 943 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 944 device = rqpair->port->device; 945 946 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 947 ibv_init_attr.qp_context = rqpair; 948 ibv_init_attr.qp_type = IBV_QPT_RC; 949 ibv_init_attr.send_cq = rqpair->poller->cq; 950 ibv_init_attr.recv_cq = rqpair->poller->cq; 951 952 if (rqpair->srq) { 953 ibv_init_attr.srq = rqpair->srq; 954 } else { 955 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 956 1; /* RECV operations + dummy drain WR */ 957 } 958 959 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 960 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 961 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 962 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 963 964 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 965 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 966 goto error; 967 } 968 969 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 970 if (rc) { 971 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 972 goto error; 973 } 974 975 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2 + 1), 976 ibv_init_attr.cap.max_send_wr); 977 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 978 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 979 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 980 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 981 982 if (rqpair->poller->srq == NULL) { 983 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 984 transport = &rtransport->transport; 985 986 opts.qp = rqpair->cm_id->qp; 987 opts.pd = rqpair->cm_id->pd; 988 opts.qpair = rqpair; 989 opts.shared = false; 990 opts.max_queue_depth = rqpair->max_queue_depth; 991 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 992 993 rqpair->resources = nvmf_rdma_resources_create(&opts); 994 995 if (!rqpair->resources) { 996 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 997 goto error; 998 } 999 } else { 1000 rqpair->resources = rqpair->poller->resources; 1001 } 1002 1003 rqpair->current_recv_depth = 0; 1004 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1005 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1006 1007 return 0; 1008 1009 error: 1010 rdma_destroy_id(rqpair->cm_id); 1011 rqpair->cm_id = NULL; 1012 spdk_nvmf_rdma_qpair_destroy(rqpair); 1013 return -1; 1014 } 1015 1016 static int 1017 request_transfer_in(struct spdk_nvmf_request *req) 1018 { 1019 int rc; 1020 struct spdk_nvmf_rdma_request *rdma_req; 1021 struct spdk_nvmf_qpair *qpair; 1022 struct spdk_nvmf_rdma_qpair *rqpair; 1023 struct ibv_send_wr *bad_wr = NULL; 1024 1025 qpair = req->qpair; 1026 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1027 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1028 1029 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1030 assert(rdma_req != NULL); 1031 1032 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA READ POSTED. Request: %p Connection: %p\n", req, qpair); 1033 1034 rc = ibv_post_send(rqpair->cm_id->qp, &rdma_req->data.wr, &bad_wr); 1035 if (rc) { 1036 SPDK_ERRLOG("Unable to transfer data from host to target\n"); 1037 return -1; 1038 } 1039 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1040 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1041 return 0; 1042 } 1043 1044 static int 1045 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1046 { 1047 int rc; 1048 struct spdk_nvmf_rdma_request *rdma_req; 1049 struct spdk_nvmf_qpair *qpair; 1050 struct spdk_nvmf_rdma_qpair *rqpair; 1051 struct spdk_nvme_cpl *rsp; 1052 struct ibv_recv_wr *bad_recv_wr = NULL; 1053 struct ibv_send_wr *send_wr, *bad_send_wr = NULL; 1054 1055 *data_posted = 0; 1056 qpair = req->qpair; 1057 rsp = &req->rsp->nvme_cpl; 1058 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1059 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1060 1061 /* Advance our sq_head pointer */ 1062 if (qpair->sq_head == qpair->sq_head_max) { 1063 qpair->sq_head = 0; 1064 } else { 1065 qpair->sq_head++; 1066 } 1067 rsp->sqhd = qpair->sq_head; 1068 1069 /* Post the capsule to the recv buffer */ 1070 assert(rdma_req->recv != NULL); 1071 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA RECV POSTED. Recv: %p Connection: %p\n", rdma_req->recv, 1072 rqpair); 1073 if (rqpair->srq == NULL) { 1074 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_req->recv->wr, &bad_recv_wr); 1075 } else { 1076 rdma_req->recv->qpair = NULL; 1077 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 1078 } 1079 1080 if (rc) { 1081 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 1082 return rc; 1083 } 1084 rdma_req->recv = NULL; 1085 assert(rqpair->current_recv_depth > 0); 1086 rqpair->current_recv_depth--; 1087 1088 /* Build the response which consists of optional 1089 * RDMA WRITEs to transfer data, plus an RDMA SEND 1090 * containing the response. 1091 */ 1092 send_wr = &rdma_req->rsp.wr; 1093 1094 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 1095 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1096 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA WRITE POSTED. Request: %p Connection: %p\n", req, qpair); 1097 send_wr = &rdma_req->data.wr; 1098 *data_posted = 1; 1099 } 1100 1101 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA SEND POSTED. Request: %p Connection: %p\n", req, qpair); 1102 1103 /* Send the completion */ 1104 rc = ibv_post_send(rqpair->cm_id->qp, send_wr, &bad_send_wr); 1105 if (rc) { 1106 SPDK_ERRLOG("Unable to send response capsule\n"); 1107 return rc; 1108 } 1109 /* +1 for the rsp wr */ 1110 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr + 1; 1111 1112 return 0; 1113 } 1114 1115 static int 1116 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1117 { 1118 struct spdk_nvmf_rdma_accept_private_data accept_data; 1119 struct rdma_conn_param ctrlr_event_data = {}; 1120 int rc; 1121 1122 accept_data.recfmt = 0; 1123 accept_data.crqsize = rqpair->max_queue_depth; 1124 1125 ctrlr_event_data.private_data = &accept_data; 1126 ctrlr_event_data.private_data_len = sizeof(accept_data); 1127 if (id->ps == RDMA_PS_TCP) { 1128 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1129 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1130 } 1131 1132 rc = rdma_accept(id, &ctrlr_event_data); 1133 if (rc) { 1134 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 1135 } else { 1136 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1137 } 1138 1139 return rc; 1140 } 1141 1142 static void 1143 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1144 { 1145 struct spdk_nvmf_rdma_reject_private_data rej_data; 1146 1147 rej_data.recfmt = 0; 1148 rej_data.sts = error; 1149 1150 rdma_reject(id, &rej_data, sizeof(rej_data)); 1151 } 1152 1153 static int 1154 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 1155 new_qpair_fn cb_fn) 1156 { 1157 struct spdk_nvmf_rdma_transport *rtransport; 1158 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1159 struct spdk_nvmf_rdma_port *port; 1160 struct rdma_conn_param *rdma_param = NULL; 1161 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1162 uint16_t max_queue_depth; 1163 uint16_t max_read_depth; 1164 1165 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1166 1167 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1168 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1169 1170 rdma_param = &event->param.conn; 1171 if (rdma_param->private_data == NULL || 1172 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1173 SPDK_ERRLOG("connect request: no private data provided\n"); 1174 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1175 return -1; 1176 } 1177 1178 private_data = rdma_param->private_data; 1179 if (private_data->recfmt != 0) { 1180 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1181 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1182 return -1; 1183 } 1184 1185 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1186 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1187 1188 port = event->listen_id->context; 1189 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1190 event->listen_id, event->listen_id->verbs, port); 1191 1192 /* Figure out the supported queue depth. This is a multi-step process 1193 * that takes into account hardware maximums, host provided values, 1194 * and our target's internal memory limits */ 1195 1196 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1197 1198 /* Start with the maximum queue depth allowed by the target */ 1199 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1200 max_read_depth = rtransport->transport.opts.max_queue_depth; 1201 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1202 rtransport->transport.opts.max_queue_depth); 1203 1204 /* Next check the local NIC's hardware limitations */ 1205 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1206 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1207 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1208 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1209 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1210 1211 /* Next check the remote NIC's hardware limitations */ 1212 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1213 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1214 rdma_param->initiator_depth, rdma_param->responder_resources); 1215 if (rdma_param->initiator_depth > 0) { 1216 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1217 } 1218 1219 /* Finally check for the host software requested values, which are 1220 * optional. */ 1221 if (rdma_param->private_data != NULL && 1222 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1223 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1224 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1225 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1226 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1227 } 1228 1229 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1230 max_queue_depth, max_read_depth); 1231 1232 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1233 if (rqpair == NULL) { 1234 SPDK_ERRLOG("Could not allocate new connection.\n"); 1235 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1236 return -1; 1237 } 1238 1239 rqpair->port = port; 1240 rqpair->max_queue_depth = max_queue_depth; 1241 rqpair->max_read_depth = max_read_depth; 1242 rqpair->cm_id = event->id; 1243 rqpair->listen_id = event->listen_id; 1244 rqpair->qpair.transport = transport; 1245 1246 event->id->context = &rqpair->qpair; 1247 1248 cb_fn(&rqpair->qpair); 1249 1250 return 0; 1251 } 1252 1253 static int 1254 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1255 enum spdk_mem_map_notify_action action, 1256 void *vaddr, size_t size) 1257 { 1258 struct ibv_pd *pd = cb_ctx; 1259 struct ibv_mr *mr; 1260 1261 switch (action) { 1262 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1263 if (!g_nvmf_hooks.get_rkey) { 1264 mr = ibv_reg_mr(pd, vaddr, size, 1265 IBV_ACCESS_LOCAL_WRITE | 1266 IBV_ACCESS_REMOTE_READ | 1267 IBV_ACCESS_REMOTE_WRITE); 1268 if (mr == NULL) { 1269 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1270 return -1; 1271 } else { 1272 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1273 } 1274 } else { 1275 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1276 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1277 } 1278 break; 1279 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1280 if (!g_nvmf_hooks.get_rkey) { 1281 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1282 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1283 if (mr) { 1284 ibv_dereg_mr(mr); 1285 } 1286 } 1287 break; 1288 } 1289 1290 return 0; 1291 } 1292 1293 static int 1294 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1295 { 1296 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1297 return addr_1 == addr_2; 1298 } 1299 1300 static void 1301 spdk_nvmf_rdma_request_free_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1302 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1303 uint32_t num_buffers) 1304 { 1305 uint32_t i; 1306 1307 for (i = 0; i < num_buffers; i++) { 1308 if (group->buf_cache_count < group->buf_cache_size) { 1309 STAILQ_INSERT_HEAD(&group->buf_cache, 1310 (struct spdk_nvmf_transport_pg_cache_buf *)rdma_req->buffers[i], link); 1311 group->buf_cache_count++; 1312 } else { 1313 spdk_mempool_put(transport->data_buf_pool, rdma_req->buffers[i]); 1314 } 1315 rdma_req->req.iov[i].iov_base = NULL; 1316 rdma_req->buffers[i] = NULL; 1317 rdma_req->req.iov[i].iov_len = 0; 1318 1319 } 1320 rdma_req->data_from_pool = false; 1321 } 1322 1323 static int 1324 nvmf_rdma_request_get_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1325 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1326 uint32_t num_buffers) 1327 { 1328 uint32_t i = 0; 1329 1330 while (i < num_buffers) { 1331 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1332 group->buf_cache_count--; 1333 rdma_req->buffers[i] = STAILQ_FIRST(&group->buf_cache); 1334 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1335 assert(rdma_req->buffers[i] != NULL); 1336 i++; 1337 } else { 1338 if (spdk_mempool_get_bulk(transport->data_buf_pool, &rdma_req->buffers[i], num_buffers - i)) { 1339 goto err_exit; 1340 } 1341 i += num_buffers - i; 1342 } 1343 } 1344 1345 return 0; 1346 1347 err_exit: 1348 spdk_nvmf_rdma_request_free_buffers(rdma_req, group, transport, i); 1349 return -ENOMEM; 1350 } 1351 1352 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1353 1354 static spdk_nvme_data_transfer_t 1355 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1356 { 1357 enum spdk_nvme_data_transfer xfer; 1358 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1359 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1360 1361 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1362 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1363 rdma_req->rsp.wr.imm_data = 0; 1364 #endif 1365 1366 /* Figure out data transfer direction */ 1367 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1368 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1369 } else { 1370 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1371 1372 /* Some admin commands are special cases */ 1373 if ((rdma_req->req.qpair->qid == 0) && 1374 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1375 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1376 switch (cmd->cdw10 & 0xff) { 1377 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1378 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1379 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1380 break; 1381 default: 1382 xfer = SPDK_NVME_DATA_NONE; 1383 } 1384 } 1385 } 1386 1387 if (xfer == SPDK_NVME_DATA_NONE) { 1388 return xfer; 1389 } 1390 1391 /* Even for commands that may transfer data, they could have specified 0 length. 1392 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1393 */ 1394 switch (sgl->generic.type) { 1395 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1396 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1397 case SPDK_NVME_SGL_TYPE_SEGMENT: 1398 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1399 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1400 if (sgl->unkeyed.length == 0) { 1401 xfer = SPDK_NVME_DATA_NONE; 1402 } 1403 break; 1404 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1405 if (sgl->keyed.length == 0) { 1406 xfer = SPDK_NVME_DATA_NONE; 1407 } 1408 break; 1409 } 1410 1411 return xfer; 1412 } 1413 1414 static int 1415 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1416 struct spdk_nvmf_rdma_request *rdma_req, 1417 uint32_t num_sgl_descriptors) 1418 { 1419 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1420 struct spdk_nvmf_rdma_request_data *current_data_wr; 1421 uint32_t i; 1422 1423 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1424 return -ENOMEM; 1425 } 1426 1427 current_data_wr = &rdma_req->data; 1428 1429 for (i = 0; i < num_sgl_descriptors; i++) { 1430 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1431 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1432 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1433 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1434 } else { 1435 assert(false); 1436 } 1437 work_requests[i]->wr.send_flags = IBV_SEND_SIGNALED; 1438 work_requests[i]->wr.sg_list = work_requests[i]->sgl; 1439 work_requests[i]->wr.wr_id = rdma_req->data.wr.wr_id; 1440 current_data_wr->wr.next = &work_requests[i]->wr; 1441 current_data_wr = work_requests[i]; 1442 } 1443 1444 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1445 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1446 current_data_wr->wr.next = &rdma_req->rsp.wr; 1447 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1448 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1449 current_data_wr->wr.next = NULL; 1450 } 1451 return 0; 1452 } 1453 1454 static int 1455 nvmf_rdma_fill_buffers(struct spdk_nvmf_rdma_transport *rtransport, 1456 struct spdk_nvmf_rdma_poll_group *rgroup, 1457 struct spdk_nvmf_rdma_device *device, 1458 struct spdk_nvmf_rdma_request *rdma_req, 1459 struct ibv_send_wr *wr, 1460 uint32_t length) 1461 { 1462 uint64_t translation_len; 1463 uint32_t remaining_length = length; 1464 uint32_t iovcnt; 1465 uint32_t i = 0; 1466 1467 1468 while (remaining_length) { 1469 iovcnt = rdma_req->req.iovcnt; 1470 rdma_req->req.iov[iovcnt].iov_base = (void *)((uintptr_t)(rdma_req->buffers[iovcnt] + 1471 NVMF_DATA_BUFFER_MASK) & 1472 ~NVMF_DATA_BUFFER_MASK); 1473 rdma_req->req.iov[iovcnt].iov_len = spdk_min(remaining_length, 1474 rtransport->transport.opts.io_unit_size); 1475 rdma_req->req.iovcnt++; 1476 wr->sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[iovcnt].iov_base); 1477 wr->sg_list[i].length = rdma_req->req.iov[iovcnt].iov_len; 1478 translation_len = rdma_req->req.iov[iovcnt].iov_len; 1479 1480 if (!g_nvmf_hooks.get_rkey) { 1481 wr->sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1482 (uint64_t)rdma_req->buffers[iovcnt], &translation_len))->lkey; 1483 } else { 1484 wr->sg_list[i].lkey = spdk_mem_map_translate(device->map, 1485 (uint64_t)rdma_req->buffers[iovcnt], &translation_len); 1486 } 1487 1488 remaining_length -= rdma_req->req.iov[iovcnt].iov_len; 1489 1490 if (translation_len < rdma_req->req.iov[iovcnt].iov_len) { 1491 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions\n"); 1492 return -EINVAL; 1493 } 1494 i++; 1495 } 1496 1497 return 0; 1498 } 1499 1500 static int 1501 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1502 struct spdk_nvmf_rdma_device *device, 1503 struct spdk_nvmf_rdma_request *rdma_req) 1504 { 1505 struct spdk_nvmf_rdma_qpair *rqpair; 1506 struct spdk_nvmf_rdma_poll_group *rgroup; 1507 uint32_t num_buffers; 1508 uint32_t i = 0; 1509 int rc = 0; 1510 1511 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1512 rgroup = rqpair->poller->group; 1513 rdma_req->req.iovcnt = 0; 1514 1515 num_buffers = rdma_req->req.length / rtransport->transport.opts.io_unit_size; 1516 if (rdma_req->req.length % rtransport->transport.opts.io_unit_size) { 1517 num_buffers++; 1518 } 1519 1520 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers)) { 1521 return -ENOMEM; 1522 } 1523 1524 rdma_req->req.iovcnt = 0; 1525 1526 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, &rdma_req->data.wr, 1527 rdma_req->req.length); 1528 if (rc != 0) { 1529 goto err_exit; 1530 } 1531 1532 assert(rdma_req->req.iovcnt <= rqpair->max_send_sge); 1533 1534 rdma_req->data_from_pool = true; 1535 1536 return rc; 1537 1538 err_exit: 1539 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1540 while (i) { 1541 i--; 1542 rdma_req->data.wr.sg_list[i].addr = 0; 1543 rdma_req->data.wr.sg_list[i].length = 0; 1544 rdma_req->data.wr.sg_list[i].lkey = 0; 1545 } 1546 rdma_req->req.iovcnt = 0; 1547 return rc; 1548 } 1549 1550 static int 1551 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1552 struct spdk_nvmf_rdma_device *device, 1553 struct spdk_nvmf_rdma_request *rdma_req) 1554 { 1555 struct spdk_nvmf_rdma_qpair *rqpair; 1556 struct spdk_nvmf_rdma_poll_group *rgroup; 1557 struct ibv_send_wr *current_wr; 1558 struct spdk_nvmf_request *req = &rdma_req->req; 1559 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1560 uint32_t num_sgl_descriptors; 1561 uint32_t num_buffers = 0; 1562 uint32_t i; 1563 int rc; 1564 1565 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1566 rgroup = rqpair->poller->group; 1567 1568 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1569 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1570 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1571 1572 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1573 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1574 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1575 1576 for (i = 0; i < num_sgl_descriptors; i++) { 1577 num_buffers += desc->keyed.length / rtransport->transport.opts.io_unit_size; 1578 if (desc->keyed.length % rtransport->transport.opts.io_unit_size) { 1579 num_buffers++; 1580 } 1581 desc++; 1582 } 1583 /* If the number of buffers is too large, then we know the I/O is larger than allowed. Fail it. */ 1584 if (num_buffers > NVMF_REQ_MAX_BUFFERS) { 1585 return -EINVAL; 1586 } 1587 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1588 num_buffers) != 0) { 1589 return -ENOMEM; 1590 } 1591 1592 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1593 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1594 return -ENOMEM; 1595 } 1596 1597 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1598 current_wr = &rdma_req->data.wr; 1599 1600 req->iovcnt = 0; 1601 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1602 for (i = 0; i < num_sgl_descriptors; i++) { 1603 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1604 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1605 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1606 rc = -EINVAL; 1607 goto err_exit; 1608 } 1609 1610 current_wr->num_sge = 0; 1611 req->length += desc->keyed.length; 1612 1613 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, current_wr, 1614 desc->keyed.length); 1615 if (rc != 0) { 1616 rc = -ENOMEM; 1617 goto err_exit; 1618 } 1619 1620 current_wr->wr.rdma.rkey = desc->keyed.key; 1621 current_wr->wr.rdma.remote_addr = desc->address; 1622 current_wr = current_wr->next; 1623 desc++; 1624 } 1625 1626 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1627 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1628 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1629 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1630 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1631 } 1632 } 1633 #endif 1634 1635 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1636 rdma_req->data_from_pool = true; 1637 1638 return 0; 1639 1640 err_exit: 1641 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1642 nvmf_rdma_request_free_data(rdma_req, rtransport); 1643 return rc; 1644 } 1645 1646 static int 1647 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1648 struct spdk_nvmf_rdma_device *device, 1649 struct spdk_nvmf_rdma_request *rdma_req) 1650 { 1651 struct spdk_nvme_cmd *cmd; 1652 struct spdk_nvme_cpl *rsp; 1653 struct spdk_nvme_sgl_descriptor *sgl; 1654 int rc; 1655 1656 cmd = &rdma_req->req.cmd->nvme_cmd; 1657 rsp = &rdma_req->req.rsp->nvme_cpl; 1658 sgl = &cmd->dptr.sgl1; 1659 1660 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1661 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1662 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1663 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1664 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1665 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1666 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1667 return -1; 1668 } 1669 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1670 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1671 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1672 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1673 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1674 } 1675 } 1676 #endif 1677 1678 /* fill request length and populate iovs */ 1679 rdma_req->req.length = sgl->keyed.length; 1680 1681 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1682 /* No available buffers. Queue this request up. */ 1683 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1684 return 0; 1685 } 1686 1687 /* backward compatible */ 1688 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1689 1690 /* rdma wr specifics */ 1691 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1692 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1693 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1694 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1695 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 1696 rdma_req->data.wr.next = &rdma_req->rsp.wr; 1697 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1698 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 1699 rdma_req->data.wr.next = NULL; 1700 } 1701 1702 /* set the number of outstanding data WRs for this request. */ 1703 rdma_req->num_outstanding_data_wr = 1; 1704 1705 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1706 rdma_req->req.iovcnt); 1707 1708 return 0; 1709 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1710 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1711 uint64_t offset = sgl->address; 1712 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1713 1714 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1715 offset, sgl->unkeyed.length); 1716 1717 if (offset > max_len) { 1718 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1719 offset, max_len); 1720 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1721 return -1; 1722 } 1723 max_len -= (uint32_t)offset; 1724 1725 if (sgl->unkeyed.length > max_len) { 1726 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1727 sgl->unkeyed.length, max_len); 1728 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1729 return -1; 1730 } 1731 1732 rdma_req->num_outstanding_data_wr = 0; 1733 rdma_req->req.data = rdma_req->recv->buf + offset; 1734 rdma_req->data_from_pool = false; 1735 rdma_req->req.length = sgl->unkeyed.length; 1736 1737 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1738 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1739 rdma_req->req.iovcnt = 1; 1740 1741 return 0; 1742 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1743 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1744 1745 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1746 if (rc == -ENOMEM) { 1747 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1748 return 0; 1749 } else if (rc == -EINVAL) { 1750 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1751 return -1; 1752 } 1753 1754 /* backward compatible */ 1755 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1756 1757 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1758 rdma_req->req.iovcnt); 1759 1760 return 0; 1761 } 1762 1763 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1764 sgl->generic.type, sgl->generic.subtype); 1765 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1766 return -1; 1767 } 1768 1769 static void 1770 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1771 struct spdk_nvmf_rdma_transport *rtransport) 1772 { 1773 struct spdk_nvmf_rdma_qpair *rqpair; 1774 struct spdk_nvmf_rdma_poll_group *rgroup; 1775 1776 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1777 if (rdma_req->data_from_pool) { 1778 rgroup = rqpair->poller->group; 1779 1780 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1781 rdma_req->req.iovcnt); 1782 } 1783 nvmf_rdma_request_free_data(rdma_req, rtransport); 1784 rdma_req->req.length = 0; 1785 rdma_req->req.iovcnt = 0; 1786 rdma_req->req.data = NULL; 1787 rdma_req->data.wr.next = NULL; 1788 rqpair->qd--; 1789 1790 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1791 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1792 } 1793 1794 static bool 1795 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1796 struct spdk_nvmf_rdma_request *rdma_req) 1797 { 1798 struct spdk_nvmf_rdma_qpair *rqpair; 1799 struct spdk_nvmf_rdma_device *device; 1800 struct spdk_nvmf_rdma_poll_group *rgroup; 1801 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1802 int rc; 1803 struct spdk_nvmf_rdma_recv *rdma_recv; 1804 enum spdk_nvmf_rdma_request_state prev_state; 1805 bool progress = false; 1806 int data_posted; 1807 1808 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1809 device = rqpair->port->device; 1810 rgroup = rqpair->poller->group; 1811 1812 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1813 1814 /* If the queue pair is in an error state, force the request to the completed state 1815 * to release resources. */ 1816 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1817 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1818 STAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1819 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1820 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1821 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1822 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1823 } 1824 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1825 } 1826 1827 /* The loop here is to allow for several back-to-back state changes. */ 1828 do { 1829 prev_state = rdma_req->state; 1830 1831 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1832 1833 switch (rdma_req->state) { 1834 case RDMA_REQUEST_STATE_FREE: 1835 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1836 * to escape this state. */ 1837 break; 1838 case RDMA_REQUEST_STATE_NEW: 1839 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1840 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1841 rdma_recv = rdma_req->recv; 1842 1843 /* The first element of the SGL is the NVMe command */ 1844 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1845 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1846 1847 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1848 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1849 break; 1850 } 1851 1852 /* The next state transition depends on the data transfer needs of this request. */ 1853 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1854 1855 /* If no data to transfer, ready to execute. */ 1856 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1857 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1858 break; 1859 } 1860 1861 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1862 STAILQ_INSERT_TAIL(&rgroup->pending_data_buf_queue, rdma_req, state_link); 1863 break; 1864 case RDMA_REQUEST_STATE_NEED_BUFFER: 1865 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1866 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1867 1868 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1869 1870 if (rdma_req != STAILQ_FIRST(&rgroup->pending_data_buf_queue)) { 1871 /* This request needs to wait in line to obtain a buffer */ 1872 break; 1873 } 1874 1875 /* Try to get a data buffer */ 1876 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1877 if (rc < 0) { 1878 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1879 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1880 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1881 break; 1882 } 1883 1884 if (!rdma_req->req.data) { 1885 /* No buffers available. */ 1886 break; 1887 } 1888 1889 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1890 1891 /* If data is transferring from host to controller and the data didn't 1892 * arrive using in capsule data, we need to do a transfer from the host. 1893 */ 1894 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1895 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1896 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1897 break; 1898 } 1899 1900 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1901 break; 1902 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1903 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1904 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1905 1906 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1907 /* This request needs to wait in line to perform RDMA */ 1908 break; 1909 } 1910 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1911 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1912 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1913 break; 1914 } 1915 1916 /* We have already verified that this request is the head of the queue. */ 1917 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 1918 1919 rc = request_transfer_in(&rdma_req->req); 1920 if (!rc) { 1921 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 1922 } else { 1923 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1924 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1925 } 1926 break; 1927 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1928 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1929 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1930 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1931 * to escape this state. */ 1932 break; 1933 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1934 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1935 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1936 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 1937 spdk_nvmf_request_exec(&rdma_req->req); 1938 break; 1939 case RDMA_REQUEST_STATE_EXECUTING: 1940 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1941 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1942 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1943 * to escape this state. */ 1944 break; 1945 case RDMA_REQUEST_STATE_EXECUTED: 1946 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1947 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1948 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1949 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 1950 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 1951 } else { 1952 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1953 } 1954 break; 1955 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 1956 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 1957 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1958 1959 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 1960 /* This request needs to wait in line to perform RDMA */ 1961 break; 1962 } 1963 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 1964 rqpair->max_send_depth) { 1965 /* We can only have so many WRs outstanding. we have to wait until some finish. 1966 * +1 since each request has an additional wr in the resp. */ 1967 break; 1968 } 1969 1970 /* We have already verified that this request is the head of the queue. */ 1971 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 1972 1973 /* The data transfer will be kicked off from 1974 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1975 */ 1976 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1977 break; 1978 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1979 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 1980 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1981 rc = request_transfer_out(&rdma_req->req, &data_posted); 1982 assert(rc == 0); /* No good way to handle this currently */ 1983 if (rc) { 1984 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1985 } else { 1986 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1987 RDMA_REQUEST_STATE_COMPLETING; 1988 } 1989 break; 1990 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1991 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 1992 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1993 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1994 * to escape this state. */ 1995 break; 1996 case RDMA_REQUEST_STATE_COMPLETING: 1997 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 1998 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1999 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2000 * to escape this state. */ 2001 break; 2002 case RDMA_REQUEST_STATE_COMPLETED: 2003 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2004 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2005 2006 nvmf_rdma_request_free(rdma_req, rtransport); 2007 break; 2008 case RDMA_REQUEST_NUM_STATES: 2009 default: 2010 assert(0); 2011 break; 2012 } 2013 2014 if (rdma_req->state != prev_state) { 2015 progress = true; 2016 } 2017 } while (rdma_req->state != prev_state); 2018 2019 return progress; 2020 } 2021 2022 /* Public API callbacks begin here */ 2023 2024 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2025 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2026 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2027 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 2028 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2029 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2030 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2031 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4096 2032 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2033 2034 static void 2035 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2036 { 2037 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2038 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2039 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2040 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2041 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2042 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2043 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2044 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2045 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2046 } 2047 2048 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2049 .notify_cb = spdk_nvmf_rdma_mem_notify, 2050 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 2051 }; 2052 2053 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2054 2055 static struct spdk_nvmf_transport * 2056 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2057 { 2058 int rc; 2059 struct spdk_nvmf_rdma_transport *rtransport; 2060 struct spdk_nvmf_rdma_device *device, *tmp; 2061 struct ibv_pd *pd; 2062 struct ibv_context **contexts; 2063 uint32_t i; 2064 int flag; 2065 uint32_t sge_count; 2066 uint32_t min_shared_buffers; 2067 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2068 2069 rtransport = calloc(1, sizeof(*rtransport)); 2070 if (!rtransport) { 2071 return NULL; 2072 } 2073 2074 if (pthread_mutex_init(&rtransport->lock, NULL)) { 2075 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2076 free(rtransport); 2077 return NULL; 2078 } 2079 2080 TAILQ_INIT(&rtransport->devices); 2081 TAILQ_INIT(&rtransport->ports); 2082 2083 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2084 2085 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2086 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2087 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2088 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2089 " num_shared_buffers=%d, max_srq_depth=%d\n", 2090 opts->max_queue_depth, 2091 opts->max_io_size, 2092 opts->max_qpairs_per_ctrlr, 2093 opts->io_unit_size, 2094 opts->in_capsule_data_size, 2095 opts->max_aq_depth, 2096 opts->num_shared_buffers, 2097 opts->max_srq_depth); 2098 2099 /* I/O unit size cannot be larger than max I/O size */ 2100 if (opts->io_unit_size > opts->max_io_size) { 2101 opts->io_unit_size = opts->max_io_size; 2102 } 2103 2104 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2105 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2106 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2107 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2108 spdk_nvmf_rdma_destroy(&rtransport->transport); 2109 return NULL; 2110 } 2111 2112 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2113 if (min_shared_buffers > opts->num_shared_buffers) { 2114 SPDK_ERRLOG("There are not enough buffers to satisfy" 2115 "per-poll group caches for each thread. (%" PRIu32 ")" 2116 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2117 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2118 spdk_nvmf_rdma_destroy(&rtransport->transport); 2119 return NULL; 2120 } 2121 2122 sge_count = opts->max_io_size / opts->io_unit_size; 2123 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2124 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2125 spdk_nvmf_rdma_destroy(&rtransport->transport); 2126 return NULL; 2127 } 2128 2129 rtransport->event_channel = rdma_create_event_channel(); 2130 if (rtransport->event_channel == NULL) { 2131 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2132 spdk_nvmf_rdma_destroy(&rtransport->transport); 2133 return NULL; 2134 } 2135 2136 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2137 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2138 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2139 rtransport->event_channel->fd, spdk_strerror(errno)); 2140 spdk_nvmf_rdma_destroy(&rtransport->transport); 2141 return NULL; 2142 } 2143 2144 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2145 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2146 sizeof(struct spdk_nvmf_rdma_request_data), 2147 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2148 SPDK_ENV_SOCKET_ID_ANY); 2149 if (!rtransport->data_wr_pool) { 2150 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2151 spdk_nvmf_rdma_destroy(&rtransport->transport); 2152 return NULL; 2153 } 2154 2155 contexts = rdma_get_devices(NULL); 2156 if (contexts == NULL) { 2157 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2158 spdk_nvmf_rdma_destroy(&rtransport->transport); 2159 return NULL; 2160 } 2161 2162 i = 0; 2163 rc = 0; 2164 while (contexts[i] != NULL) { 2165 device = calloc(1, sizeof(*device)); 2166 if (!device) { 2167 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2168 rc = -ENOMEM; 2169 break; 2170 } 2171 device->context = contexts[i]; 2172 rc = ibv_query_device(device->context, &device->attr); 2173 if (rc < 0) { 2174 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2175 free(device); 2176 break; 2177 2178 } 2179 2180 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2181 2182 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2183 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2184 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2185 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2186 } 2187 2188 /** 2189 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2190 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2191 * but incorrectly reports that it does. There are changes making their way 2192 * through the kernel now that will enable this feature. When they are merged, 2193 * we can conditionally enable this feature. 2194 * 2195 * TODO: enable this for versions of the kernel rxe driver that support it. 2196 */ 2197 if (device->attr.vendor_id == 0) { 2198 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2199 } 2200 #endif 2201 2202 /* set up device context async ev fd as NON_BLOCKING */ 2203 flag = fcntl(device->context->async_fd, F_GETFL); 2204 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2205 if (rc < 0) { 2206 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2207 free(device); 2208 break; 2209 } 2210 2211 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2212 i++; 2213 2214 pd = NULL; 2215 if (g_nvmf_hooks.get_ibv_pd) { 2216 pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2217 } 2218 2219 if (!g_nvmf_hooks.get_ibv_pd) { 2220 device->pd = ibv_alloc_pd(device->context); 2221 if (!device->pd) { 2222 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2223 spdk_nvmf_rdma_destroy(&rtransport->transport); 2224 return NULL; 2225 } 2226 } else { 2227 device->pd = pd; 2228 } 2229 2230 assert(device->map == NULL); 2231 2232 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2233 if (!device->map) { 2234 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2235 spdk_nvmf_rdma_destroy(&rtransport->transport); 2236 return NULL; 2237 } 2238 2239 assert(device->map != NULL); 2240 assert(device->pd != NULL); 2241 } 2242 rdma_free_devices(contexts); 2243 2244 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2245 /* divide and round up. */ 2246 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2247 2248 /* round up to the nearest 4k. */ 2249 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2250 2251 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2252 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2253 opts->io_unit_size); 2254 } 2255 2256 if (rc < 0) { 2257 spdk_nvmf_rdma_destroy(&rtransport->transport); 2258 return NULL; 2259 } 2260 2261 /* Set up poll descriptor array to monitor events from RDMA and IB 2262 * in a single poll syscall 2263 */ 2264 rtransport->npoll_fds = i + 1; 2265 i = 0; 2266 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2267 if (rtransport->poll_fds == NULL) { 2268 SPDK_ERRLOG("poll_fds allocation failed\n"); 2269 spdk_nvmf_rdma_destroy(&rtransport->transport); 2270 return NULL; 2271 } 2272 2273 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2274 rtransport->poll_fds[i++].events = POLLIN; 2275 2276 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2277 rtransport->poll_fds[i].fd = device->context->async_fd; 2278 rtransport->poll_fds[i++].events = POLLIN; 2279 } 2280 2281 return &rtransport->transport; 2282 } 2283 2284 static int 2285 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2286 { 2287 struct spdk_nvmf_rdma_transport *rtransport; 2288 struct spdk_nvmf_rdma_port *port, *port_tmp; 2289 struct spdk_nvmf_rdma_device *device, *device_tmp; 2290 2291 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2292 2293 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2294 TAILQ_REMOVE(&rtransport->ports, port, link); 2295 rdma_destroy_id(port->id); 2296 free(port); 2297 } 2298 2299 if (rtransport->poll_fds != NULL) { 2300 free(rtransport->poll_fds); 2301 } 2302 2303 if (rtransport->event_channel != NULL) { 2304 rdma_destroy_event_channel(rtransport->event_channel); 2305 } 2306 2307 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2308 TAILQ_REMOVE(&rtransport->devices, device, link); 2309 if (device->map) { 2310 spdk_mem_map_free(&device->map); 2311 } 2312 if (device->pd) { 2313 if (!g_nvmf_hooks.get_ibv_pd) { 2314 ibv_dealloc_pd(device->pd); 2315 } 2316 } 2317 free(device); 2318 } 2319 2320 if (rtransport->data_wr_pool != NULL) { 2321 if (spdk_mempool_count(rtransport->data_wr_pool) != 2322 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2323 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2324 spdk_mempool_count(rtransport->data_wr_pool), 2325 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2326 } 2327 } 2328 2329 spdk_mempool_free(rtransport->data_wr_pool); 2330 pthread_mutex_destroy(&rtransport->lock); 2331 free(rtransport); 2332 2333 return 0; 2334 } 2335 2336 static int 2337 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2338 struct spdk_nvme_transport_id *trid, 2339 bool peer); 2340 2341 static int 2342 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2343 const struct spdk_nvme_transport_id *trid) 2344 { 2345 struct spdk_nvmf_rdma_transport *rtransport; 2346 struct spdk_nvmf_rdma_device *device; 2347 struct spdk_nvmf_rdma_port *port_tmp, *port; 2348 struct addrinfo *res; 2349 struct addrinfo hints; 2350 int family; 2351 int rc; 2352 2353 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2354 2355 port = calloc(1, sizeof(*port)); 2356 if (!port) { 2357 return -ENOMEM; 2358 } 2359 2360 /* Selectively copy the trid. Things like NQN don't matter here - that 2361 * mapping is enforced elsewhere. 2362 */ 2363 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2364 port->trid.adrfam = trid->adrfam; 2365 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 2366 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 2367 2368 pthread_mutex_lock(&rtransport->lock); 2369 assert(rtransport->event_channel != NULL); 2370 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 2371 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 2372 port_tmp->ref++; 2373 free(port); 2374 /* Already listening at this address */ 2375 pthread_mutex_unlock(&rtransport->lock); 2376 return 0; 2377 } 2378 } 2379 2380 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2381 if (rc < 0) { 2382 SPDK_ERRLOG("rdma_create_id() failed\n"); 2383 free(port); 2384 pthread_mutex_unlock(&rtransport->lock); 2385 return rc; 2386 } 2387 2388 switch (port->trid.adrfam) { 2389 case SPDK_NVMF_ADRFAM_IPV4: 2390 family = AF_INET; 2391 break; 2392 case SPDK_NVMF_ADRFAM_IPV6: 2393 family = AF_INET6; 2394 break; 2395 default: 2396 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 2397 free(port); 2398 pthread_mutex_unlock(&rtransport->lock); 2399 return -EINVAL; 2400 } 2401 2402 memset(&hints, 0, sizeof(hints)); 2403 hints.ai_family = family; 2404 hints.ai_flags = AI_NUMERICSERV; 2405 hints.ai_socktype = SOCK_STREAM; 2406 hints.ai_protocol = 0; 2407 2408 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 2409 if (rc) { 2410 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2411 free(port); 2412 pthread_mutex_unlock(&rtransport->lock); 2413 return -EINVAL; 2414 } 2415 2416 rc = rdma_bind_addr(port->id, res->ai_addr); 2417 freeaddrinfo(res); 2418 2419 if (rc < 0) { 2420 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2421 rdma_destroy_id(port->id); 2422 free(port); 2423 pthread_mutex_unlock(&rtransport->lock); 2424 return rc; 2425 } 2426 2427 if (!port->id->verbs) { 2428 SPDK_ERRLOG("ibv_context is null\n"); 2429 rdma_destroy_id(port->id); 2430 free(port); 2431 pthread_mutex_unlock(&rtransport->lock); 2432 return -1; 2433 } 2434 2435 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 2436 if (rc < 0) { 2437 SPDK_ERRLOG("rdma_listen() failed\n"); 2438 rdma_destroy_id(port->id); 2439 free(port); 2440 pthread_mutex_unlock(&rtransport->lock); 2441 return rc; 2442 } 2443 2444 TAILQ_FOREACH(device, &rtransport->devices, link) { 2445 if (device->context == port->id->verbs) { 2446 port->device = device; 2447 break; 2448 } 2449 } 2450 if (!port->device) { 2451 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2452 port->id->verbs); 2453 rdma_destroy_id(port->id); 2454 free(port); 2455 pthread_mutex_unlock(&rtransport->lock); 2456 return -EINVAL; 2457 } 2458 2459 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 2460 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 2461 2462 port->ref = 1; 2463 2464 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2465 pthread_mutex_unlock(&rtransport->lock); 2466 2467 return 0; 2468 } 2469 2470 static int 2471 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2472 const struct spdk_nvme_transport_id *_trid) 2473 { 2474 struct spdk_nvmf_rdma_transport *rtransport; 2475 struct spdk_nvmf_rdma_port *port, *tmp; 2476 struct spdk_nvme_transport_id trid = {}; 2477 2478 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2479 2480 /* Selectively copy the trid. Things like NQN don't matter here - that 2481 * mapping is enforced elsewhere. 2482 */ 2483 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2484 trid.adrfam = _trid->adrfam; 2485 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2486 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2487 2488 pthread_mutex_lock(&rtransport->lock); 2489 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2490 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2491 assert(port->ref > 0); 2492 port->ref--; 2493 if (port->ref == 0) { 2494 TAILQ_REMOVE(&rtransport->ports, port, link); 2495 rdma_destroy_id(port->id); 2496 free(port); 2497 } 2498 break; 2499 } 2500 } 2501 2502 pthread_mutex_unlock(&rtransport->lock); 2503 return 0; 2504 } 2505 2506 static void 2507 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2508 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2509 { 2510 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2511 struct spdk_nvmf_rdma_resources *resources; 2512 2513 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2514 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2515 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2516 break; 2517 } 2518 } 2519 2520 /* Then RDMA writes since reads have stronger restrictions than writes */ 2521 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2522 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2523 break; 2524 } 2525 } 2526 2527 /* The second highest priority is I/O waiting on memory buffers. */ 2528 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->poller->group->pending_data_buf_queue, state_link, 2529 req_tmp) { 2530 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2531 break; 2532 } 2533 } 2534 2535 resources = rqpair->resources; 2536 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2537 rdma_req = STAILQ_FIRST(&resources->free_queue); 2538 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2539 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2540 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2541 2542 if (rqpair->srq != NULL) { 2543 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2544 rdma_req->recv->qpair->qd++; 2545 } else { 2546 rqpair->qd++; 2547 } 2548 2549 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2550 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2551 break; 2552 } 2553 } 2554 } 2555 2556 static void 2557 _nvmf_rdma_qpair_disconnect(void *ctx) 2558 { 2559 struct spdk_nvmf_qpair *qpair = ctx; 2560 2561 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2562 } 2563 2564 static void 2565 _nvmf_rdma_try_disconnect(void *ctx) 2566 { 2567 struct spdk_nvmf_qpair *qpair = ctx; 2568 struct spdk_nvmf_poll_group *group; 2569 2570 /* Read the group out of the qpair. This is normally set and accessed only from 2571 * the thread that created the group. Here, we're not on that thread necessarily. 2572 * The data member qpair->group begins it's life as NULL and then is assigned to 2573 * a pointer and never changes. So fortunately reading this and checking for 2574 * non-NULL is thread safe in the x86_64 memory model. */ 2575 group = qpair->group; 2576 2577 if (group == NULL) { 2578 /* The qpair hasn't been assigned to a group yet, so we can't 2579 * process a disconnect. Send a message to ourself and try again. */ 2580 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2581 return; 2582 } 2583 2584 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2585 } 2586 2587 static inline void 2588 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2589 { 2590 if (__sync_bool_compare_and_swap(&rqpair->disconnect_started, false, true)) { 2591 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2592 } 2593 } 2594 2595 static void nvmf_rdma_destroy_drained_qpair(void *ctx) 2596 { 2597 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2598 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2599 struct spdk_nvmf_rdma_transport, transport); 2600 2601 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2602 if (rqpair->current_send_depth != 0) { 2603 return; 2604 } 2605 2606 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2607 return; 2608 } 2609 2610 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) { 2611 return; 2612 } 2613 2614 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2615 spdk_nvmf_rdma_qpair_destroy(rqpair); 2616 } 2617 2618 2619 static int 2620 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2621 { 2622 struct spdk_nvmf_qpair *qpair; 2623 struct spdk_nvmf_rdma_qpair *rqpair; 2624 2625 if (evt->id == NULL) { 2626 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2627 return -1; 2628 } 2629 2630 qpair = evt->id->context; 2631 if (qpair == NULL) { 2632 SPDK_ERRLOG("disconnect request: no active connection\n"); 2633 return -1; 2634 } 2635 2636 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2637 2638 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2639 2640 spdk_nvmf_rdma_update_ibv_state(rqpair); 2641 2642 spdk_nvmf_rdma_start_disconnect(rqpair); 2643 2644 return 0; 2645 } 2646 2647 #ifdef DEBUG 2648 static const char *CM_EVENT_STR[] = { 2649 "RDMA_CM_EVENT_ADDR_RESOLVED", 2650 "RDMA_CM_EVENT_ADDR_ERROR", 2651 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2652 "RDMA_CM_EVENT_ROUTE_ERROR", 2653 "RDMA_CM_EVENT_CONNECT_REQUEST", 2654 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2655 "RDMA_CM_EVENT_CONNECT_ERROR", 2656 "RDMA_CM_EVENT_UNREACHABLE", 2657 "RDMA_CM_EVENT_REJECTED", 2658 "RDMA_CM_EVENT_ESTABLISHED", 2659 "RDMA_CM_EVENT_DISCONNECTED", 2660 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2661 "RDMA_CM_EVENT_MULTICAST_JOIN", 2662 "RDMA_CM_EVENT_MULTICAST_ERROR", 2663 "RDMA_CM_EVENT_ADDR_CHANGE", 2664 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2665 }; 2666 #endif /* DEBUG */ 2667 2668 static void 2669 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2670 { 2671 struct spdk_nvmf_rdma_transport *rtransport; 2672 struct rdma_cm_event *event; 2673 int rc; 2674 2675 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2676 2677 if (rtransport->event_channel == NULL) { 2678 return; 2679 } 2680 2681 while (1) { 2682 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2683 if (rc == 0) { 2684 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2685 2686 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2687 2688 switch (event->event) { 2689 case RDMA_CM_EVENT_ADDR_RESOLVED: 2690 case RDMA_CM_EVENT_ADDR_ERROR: 2691 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2692 case RDMA_CM_EVENT_ROUTE_ERROR: 2693 /* No action required. The target never attempts to resolve routes. */ 2694 break; 2695 case RDMA_CM_EVENT_CONNECT_REQUEST: 2696 rc = nvmf_rdma_connect(transport, event, cb_fn); 2697 if (rc < 0) { 2698 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2699 break; 2700 } 2701 break; 2702 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2703 /* The target never initiates a new connection. So this will not occur. */ 2704 break; 2705 case RDMA_CM_EVENT_CONNECT_ERROR: 2706 /* Can this happen? The docs say it can, but not sure what causes it. */ 2707 break; 2708 case RDMA_CM_EVENT_UNREACHABLE: 2709 case RDMA_CM_EVENT_REJECTED: 2710 /* These only occur on the client side. */ 2711 break; 2712 case RDMA_CM_EVENT_ESTABLISHED: 2713 /* TODO: Should we be waiting for this event anywhere? */ 2714 break; 2715 case RDMA_CM_EVENT_DISCONNECTED: 2716 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2717 rc = nvmf_rdma_disconnect(event); 2718 if (rc < 0) { 2719 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2720 break; 2721 } 2722 break; 2723 case RDMA_CM_EVENT_MULTICAST_JOIN: 2724 case RDMA_CM_EVENT_MULTICAST_ERROR: 2725 /* Multicast is not used */ 2726 break; 2727 case RDMA_CM_EVENT_ADDR_CHANGE: 2728 /* Not utilizing this event */ 2729 break; 2730 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2731 /* For now, do nothing. The target never re-uses queue pairs. */ 2732 break; 2733 default: 2734 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2735 break; 2736 } 2737 2738 rdma_ack_cm_event(event); 2739 } else { 2740 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2741 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2742 } 2743 break; 2744 } 2745 } 2746 } 2747 2748 static void 2749 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2750 { 2751 int rc; 2752 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 2753 struct ibv_async_event event; 2754 enum ibv_qp_state state; 2755 2756 rc = ibv_get_async_event(device->context, &event); 2757 2758 if (rc) { 2759 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2760 errno, spdk_strerror(errno)); 2761 return; 2762 } 2763 2764 SPDK_NOTICELOG("Async event: %s\n", 2765 ibv_event_type_str(event.event_type)); 2766 2767 switch (event.event_type) { 2768 case IBV_EVENT_QP_FATAL: 2769 rqpair = event.element.qp->qp_context; 2770 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2771 (uintptr_t)rqpair->cm_id, event.event_type); 2772 spdk_nvmf_rdma_update_ibv_state(rqpair); 2773 spdk_nvmf_rdma_start_disconnect(rqpair); 2774 break; 2775 case IBV_EVENT_QP_LAST_WQE_REACHED: 2776 /* This event only occurs for shared receive queues. */ 2777 rqpair = event.element.qp->qp_context; 2778 rqpair->last_wqe_reached = true; 2779 2780 /* This must be handled on the polling thread if it exists. Otherwise the timeout will catch it. */ 2781 if (rqpair->qpair.group) { 2782 spdk_thread_send_msg(rqpair->qpair.group->thread, nvmf_rdma_destroy_drained_qpair, rqpair); 2783 } else { 2784 SPDK_ERRLOG("Unable to destroy the qpair %p since it does not have a poll group.\n", rqpair); 2785 } 2786 2787 break; 2788 case IBV_EVENT_SQ_DRAINED: 2789 /* This event occurs frequently in both error and non-error states. 2790 * Check if the qpair is in an error state before sending a message. 2791 * Note that we're not on the correct thread to access the qpair, but 2792 * the operations that the below calls make all happen to be thread 2793 * safe. */ 2794 rqpair = event.element.qp->qp_context; 2795 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2796 (uintptr_t)rqpair->cm_id, event.event_type); 2797 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2798 if (state == IBV_QPS_ERR) { 2799 spdk_nvmf_rdma_start_disconnect(rqpair); 2800 } 2801 break; 2802 case IBV_EVENT_QP_REQ_ERR: 2803 case IBV_EVENT_QP_ACCESS_ERR: 2804 case IBV_EVENT_COMM_EST: 2805 case IBV_EVENT_PATH_MIG: 2806 case IBV_EVENT_PATH_MIG_ERR: 2807 rqpair = event.element.qp->qp_context; 2808 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2809 (uintptr_t)rqpair->cm_id, event.event_type); 2810 spdk_nvmf_rdma_update_ibv_state(rqpair); 2811 break; 2812 case IBV_EVENT_CQ_ERR: 2813 case IBV_EVENT_DEVICE_FATAL: 2814 case IBV_EVENT_PORT_ACTIVE: 2815 case IBV_EVENT_PORT_ERR: 2816 case IBV_EVENT_LID_CHANGE: 2817 case IBV_EVENT_PKEY_CHANGE: 2818 case IBV_EVENT_SM_CHANGE: 2819 case IBV_EVENT_SRQ_ERR: 2820 case IBV_EVENT_SRQ_LIMIT_REACHED: 2821 case IBV_EVENT_CLIENT_REREGISTER: 2822 case IBV_EVENT_GID_CHANGE: 2823 default: 2824 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2825 break; 2826 } 2827 ibv_ack_async_event(&event); 2828 } 2829 2830 static void 2831 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2832 { 2833 int nfds, i = 0; 2834 struct spdk_nvmf_rdma_transport *rtransport; 2835 struct spdk_nvmf_rdma_device *device, *tmp; 2836 2837 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2838 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2839 2840 if (nfds <= 0) { 2841 return; 2842 } 2843 2844 /* The first poll descriptor is RDMA CM event */ 2845 if (rtransport->poll_fds[i++].revents & POLLIN) { 2846 spdk_nvmf_process_cm_event(transport, cb_fn); 2847 nfds--; 2848 } 2849 2850 if (nfds == 0) { 2851 return; 2852 } 2853 2854 /* Second and subsequent poll descriptors are IB async events */ 2855 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2856 if (rtransport->poll_fds[i++].revents & POLLIN) { 2857 spdk_nvmf_process_ib_event(device); 2858 nfds--; 2859 } 2860 } 2861 /* check all flagged fd's have been served */ 2862 assert(nfds == 0); 2863 } 2864 2865 static void 2866 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2867 struct spdk_nvme_transport_id *trid, 2868 struct spdk_nvmf_discovery_log_page_entry *entry) 2869 { 2870 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2871 entry->adrfam = trid->adrfam; 2872 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2873 2874 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2875 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2876 2877 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2878 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2879 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2880 } 2881 2882 static void 2883 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 2884 2885 static struct spdk_nvmf_transport_poll_group * 2886 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2887 { 2888 struct spdk_nvmf_rdma_transport *rtransport; 2889 struct spdk_nvmf_rdma_poll_group *rgroup; 2890 struct spdk_nvmf_rdma_poller *poller; 2891 struct spdk_nvmf_rdma_device *device; 2892 struct ibv_srq_init_attr srq_init_attr; 2893 struct spdk_nvmf_rdma_resource_opts opts; 2894 int num_cqe; 2895 2896 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2897 2898 rgroup = calloc(1, sizeof(*rgroup)); 2899 if (!rgroup) { 2900 return NULL; 2901 } 2902 2903 TAILQ_INIT(&rgroup->pollers); 2904 STAILQ_INIT(&rgroup->pending_data_buf_queue); 2905 2906 pthread_mutex_lock(&rtransport->lock); 2907 TAILQ_FOREACH(device, &rtransport->devices, link) { 2908 poller = calloc(1, sizeof(*poller)); 2909 if (!poller) { 2910 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2911 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2912 pthread_mutex_unlock(&rtransport->lock); 2913 return NULL; 2914 } 2915 2916 poller->device = device; 2917 poller->group = rgroup; 2918 2919 TAILQ_INIT(&poller->qpairs); 2920 2921 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2922 if (device->attr.max_srq != 0) { 2923 poller->max_srq_depth = transport->opts.max_srq_depth; 2924 2925 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 2926 srq_init_attr.attr.max_wr = poller->max_srq_depth; 2927 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 2928 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 2929 if (!poller->srq) { 2930 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 2931 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2932 pthread_mutex_unlock(&rtransport->lock); 2933 return NULL; 2934 } 2935 2936 opts.qp = poller->srq; 2937 opts.pd = device->pd; 2938 opts.qpair = NULL; 2939 opts.shared = true; 2940 opts.max_queue_depth = poller->max_srq_depth; 2941 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 2942 2943 poller->resources = nvmf_rdma_resources_create(&opts); 2944 if (!poller->resources) { 2945 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 2946 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2947 pthread_mutex_unlock(&rtransport->lock); 2948 } 2949 } 2950 2951 /* 2952 * When using an srq, we can limit the completion queue at startup. 2953 * The following formula represents the calculation: 2954 * num_cqe = num_recv + num_data_wr + num_send_wr. 2955 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 2956 */ 2957 if (poller->srq) { 2958 num_cqe = poller->max_srq_depth * 3; 2959 } else { 2960 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2961 } 2962 2963 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 2964 if (!poller->cq) { 2965 SPDK_ERRLOG("Unable to create completion queue\n"); 2966 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2967 pthread_mutex_unlock(&rtransport->lock); 2968 return NULL; 2969 } 2970 poller->num_cqe = num_cqe; 2971 } 2972 2973 pthread_mutex_unlock(&rtransport->lock); 2974 return &rgroup->group; 2975 } 2976 2977 static void 2978 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2979 { 2980 struct spdk_nvmf_rdma_poll_group *rgroup; 2981 struct spdk_nvmf_rdma_poller *poller, *tmp; 2982 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2983 2984 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2985 2986 if (!rgroup) { 2987 return; 2988 } 2989 2990 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2991 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2992 2993 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 2994 spdk_nvmf_rdma_qpair_destroy(qpair); 2995 } 2996 2997 if (poller->srq) { 2998 nvmf_rdma_resources_destroy(poller->resources); 2999 ibv_destroy_srq(poller->srq); 3000 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3001 } 3002 3003 if (poller->cq) { 3004 ibv_destroy_cq(poller->cq); 3005 } 3006 3007 free(poller); 3008 } 3009 3010 if (!STAILQ_EMPTY(&rgroup->pending_data_buf_queue)) { 3011 SPDK_ERRLOG("Pending I/O list wasn't empty on poll group destruction\n"); 3012 } 3013 3014 free(rgroup); 3015 } 3016 3017 static void 3018 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3019 { 3020 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3021 spdk_nvmf_rdma_qpair_destroy(rqpair); 3022 } 3023 3024 static int 3025 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3026 struct spdk_nvmf_qpair *qpair) 3027 { 3028 struct spdk_nvmf_rdma_poll_group *rgroup; 3029 struct spdk_nvmf_rdma_qpair *rqpair; 3030 struct spdk_nvmf_rdma_device *device; 3031 struct spdk_nvmf_rdma_poller *poller; 3032 int rc; 3033 3034 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3035 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3036 3037 device = rqpair->port->device; 3038 3039 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3040 if (poller->device == device) { 3041 break; 3042 } 3043 } 3044 3045 if (!poller) { 3046 SPDK_ERRLOG("No poller found for device.\n"); 3047 return -1; 3048 } 3049 3050 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3051 rqpair->poller = poller; 3052 rqpair->srq = rqpair->poller->srq; 3053 3054 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 3055 if (rc < 0) { 3056 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3057 return -1; 3058 } 3059 3060 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3061 if (rc) { 3062 /* Try to reject, but we probably can't */ 3063 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3064 return -1; 3065 } 3066 3067 spdk_nvmf_rdma_update_ibv_state(rqpair); 3068 3069 return 0; 3070 } 3071 3072 static int 3073 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3074 { 3075 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3076 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3077 struct spdk_nvmf_rdma_transport, transport); 3078 3079 nvmf_rdma_request_free(rdma_req, rtransport); 3080 return 0; 3081 } 3082 3083 static int 3084 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3085 { 3086 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3087 struct spdk_nvmf_rdma_transport, transport); 3088 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3089 struct spdk_nvmf_rdma_request, req); 3090 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3091 struct spdk_nvmf_rdma_qpair, qpair); 3092 3093 if (rqpair->ibv_state != IBV_QPS_ERR) { 3094 /* The connection is alive, so process the request as normal */ 3095 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3096 } else { 3097 /* The connection is dead. Move the request directly to the completed state. */ 3098 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3099 } 3100 3101 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3102 3103 return 0; 3104 } 3105 3106 static int 3107 spdk_nvmf_rdma_destroy_defunct_qpair(void *ctx) 3108 { 3109 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 3110 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3111 struct spdk_nvmf_rdma_transport, transport); 3112 3113 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3114 spdk_nvmf_rdma_qpair_destroy(rqpair); 3115 3116 return 0; 3117 } 3118 3119 static void 3120 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3121 { 3122 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3123 3124 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 3125 return; 3126 } 3127 3128 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 3129 3130 /* This happens only when the qpair is disconnected before 3131 * it is added to the poll group. Since there is no poll group, 3132 * the RDMA qp has not been initialized yet and the RDMA CM 3133 * event has not yet been acknowledged, so we need to reject it. 3134 */ 3135 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3136 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3137 return; 3138 } 3139 3140 if (rqpair->ibv_state != IBV_QPS_ERR) { 3141 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 3142 } 3143 3144 rqpair->destruct_poller = spdk_poller_register(spdk_nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 3145 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 3146 } 3147 3148 static struct spdk_nvmf_rdma_qpair * 3149 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3150 { 3151 struct spdk_nvmf_rdma_qpair *rqpair; 3152 /* @todo: improve QP search */ 3153 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3154 if (wc->qp_num == rqpair->cm_id->qp->qp_num) { 3155 return rqpair; 3156 } 3157 } 3158 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3159 return NULL; 3160 } 3161 3162 #ifdef DEBUG 3163 static int 3164 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3165 { 3166 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3167 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3168 } 3169 #endif 3170 3171 static int 3172 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3173 struct spdk_nvmf_rdma_poller *rpoller) 3174 { 3175 struct ibv_wc wc[32]; 3176 struct spdk_nvmf_rdma_wr *rdma_wr; 3177 struct spdk_nvmf_rdma_request *rdma_req; 3178 struct spdk_nvmf_rdma_recv *rdma_recv; 3179 struct spdk_nvmf_rdma_qpair *rqpair; 3180 int reaped, i; 3181 int count = 0; 3182 bool error = false; 3183 3184 /* Poll for completing operations. */ 3185 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3186 if (reaped < 0) { 3187 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3188 errno, spdk_strerror(errno)); 3189 return -1; 3190 } 3191 3192 for (i = 0; i < reaped; i++) { 3193 3194 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3195 3196 switch (rdma_wr->type) { 3197 case RDMA_WR_TYPE_SEND: 3198 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3199 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3200 3201 if (!wc[i].status) { 3202 count++; 3203 assert(wc[i].opcode == IBV_WC_SEND); 3204 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 3205 } else { 3206 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3207 } 3208 3209 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3210 rqpair->current_send_depth--; 3211 3212 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3213 assert(rdma_req->num_outstanding_data_wr == 0); 3214 break; 3215 case RDMA_WR_TYPE_RECV: 3216 /* rdma_recv->qpair will be NULL if using an SRQ. In that case we have to get the qpair from the wc. */ 3217 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3218 if (rdma_recv->qpair == NULL) { 3219 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3220 } 3221 rqpair = rdma_recv->qpair; 3222 3223 assert(rqpair != NULL); 3224 if (!wc[i].status) { 3225 assert(wc[i].opcode == IBV_WC_RECV); 3226 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3227 spdk_nvmf_rdma_start_disconnect(rqpair); 3228 break; 3229 } 3230 } 3231 3232 rqpair->current_recv_depth++; 3233 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3234 break; 3235 case RDMA_WR_TYPE_DATA: 3236 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3237 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3238 3239 assert(rdma_req->num_outstanding_data_wr > 0); 3240 3241 rqpair->current_send_depth--; 3242 rdma_req->num_outstanding_data_wr--; 3243 if (!wc[i].status) { 3244 if (wc[i].opcode == IBV_WC_RDMA_READ) { 3245 rqpair->current_read_depth--; 3246 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3247 if (rdma_req->num_outstanding_data_wr == 0) { 3248 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3249 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3250 } 3251 } else { 3252 assert(wc[i].opcode == IBV_WC_RDMA_WRITE); 3253 } 3254 } else { 3255 /* If the data transfer fails still force the queue into the error state, 3256 * if we were performing an RDMA_READ, we need to force the request into a 3257 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3258 * case, we should wait for the SEND to complete. */ 3259 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3260 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3261 rqpair->current_read_depth--; 3262 if (rdma_req->num_outstanding_data_wr == 0) { 3263 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3264 } 3265 } 3266 } 3267 break; 3268 default: 3269 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3270 continue; 3271 } 3272 3273 /* Handle error conditions */ 3274 if (wc[i].status) { 3275 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 3276 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 3277 3278 error = true; 3279 3280 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3281 /* Disconnect the connection. */ 3282 spdk_nvmf_rdma_start_disconnect(rqpair); 3283 } else { 3284 nvmf_rdma_destroy_drained_qpair(rqpair); 3285 } 3286 continue; 3287 } 3288 3289 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3290 3291 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3292 nvmf_rdma_destroy_drained_qpair(rqpair); 3293 } 3294 } 3295 3296 if (error == true) { 3297 return -1; 3298 } 3299 3300 return count; 3301 } 3302 3303 static int 3304 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3305 { 3306 struct spdk_nvmf_rdma_transport *rtransport; 3307 struct spdk_nvmf_rdma_poll_group *rgroup; 3308 struct spdk_nvmf_rdma_poller *rpoller; 3309 int count, rc; 3310 3311 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3312 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3313 3314 count = 0; 3315 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3316 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 3317 if (rc < 0) { 3318 return rc; 3319 } 3320 count += rc; 3321 } 3322 3323 return count; 3324 } 3325 3326 static int 3327 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3328 struct spdk_nvme_transport_id *trid, 3329 bool peer) 3330 { 3331 struct sockaddr *saddr; 3332 uint16_t port; 3333 3334 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 3335 3336 if (peer) { 3337 saddr = rdma_get_peer_addr(id); 3338 } else { 3339 saddr = rdma_get_local_addr(id); 3340 } 3341 switch (saddr->sa_family) { 3342 case AF_INET: { 3343 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3344 3345 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3346 inet_ntop(AF_INET, &saddr_in->sin_addr, 3347 trid->traddr, sizeof(trid->traddr)); 3348 if (peer) { 3349 port = ntohs(rdma_get_dst_port(id)); 3350 } else { 3351 port = ntohs(rdma_get_src_port(id)); 3352 } 3353 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3354 break; 3355 } 3356 case AF_INET6: { 3357 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3358 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3359 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3360 trid->traddr, sizeof(trid->traddr)); 3361 if (peer) { 3362 port = ntohs(rdma_get_dst_port(id)); 3363 } else { 3364 port = ntohs(rdma_get_src_port(id)); 3365 } 3366 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3367 break; 3368 } 3369 default: 3370 return -1; 3371 3372 } 3373 3374 return 0; 3375 } 3376 3377 static int 3378 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3379 struct spdk_nvme_transport_id *trid) 3380 { 3381 struct spdk_nvmf_rdma_qpair *rqpair; 3382 3383 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3384 3385 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 3386 } 3387 3388 static int 3389 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3390 struct spdk_nvme_transport_id *trid) 3391 { 3392 struct spdk_nvmf_rdma_qpair *rqpair; 3393 3394 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3395 3396 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 3397 } 3398 3399 static int 3400 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3401 struct spdk_nvme_transport_id *trid) 3402 { 3403 struct spdk_nvmf_rdma_qpair *rqpair; 3404 3405 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3406 3407 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 3408 } 3409 3410 void 3411 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3412 { 3413 g_nvmf_hooks = *hooks; 3414 } 3415 3416 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 3417 .type = SPDK_NVME_TRANSPORT_RDMA, 3418 .opts_init = spdk_nvmf_rdma_opts_init, 3419 .create = spdk_nvmf_rdma_create, 3420 .destroy = spdk_nvmf_rdma_destroy, 3421 3422 .listen = spdk_nvmf_rdma_listen, 3423 .stop_listen = spdk_nvmf_rdma_stop_listen, 3424 .accept = spdk_nvmf_rdma_accept, 3425 3426 .listener_discover = spdk_nvmf_rdma_discover, 3427 3428 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 3429 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 3430 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 3431 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 3432 3433 .req_free = spdk_nvmf_rdma_request_free, 3434 .req_complete = spdk_nvmf_rdma_request_complete, 3435 3436 .qpair_fini = spdk_nvmf_rdma_close_qpair, 3437 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 3438 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 3439 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 3440 3441 }; 3442 3443 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 3444