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 TAILQ_ENTRY(spdk_nvmf_rdma_request) link; 273 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 274 }; 275 276 enum spdk_nvmf_rdma_qpair_disconnect_flags { 277 RDMA_QP_DISCONNECTING = 1, 278 RDMA_QP_RECV_DRAINED = 1 << 1, 279 RDMA_QP_SEND_DRAINED = 1 << 2 280 }; 281 282 struct spdk_nvmf_rdma_resource_opts { 283 struct spdk_nvmf_rdma_qpair *qpair; 284 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 285 void *qp; 286 struct ibv_pd *pd; 287 uint32_t max_queue_depth; 288 uint32_t in_capsule_data_size; 289 bool shared; 290 }; 291 292 struct spdk_nvmf_rdma_resources { 293 /* Array of size "max_queue_depth" containing RDMA requests. */ 294 struct spdk_nvmf_rdma_request *reqs; 295 296 /* Array of size "max_queue_depth" containing RDMA recvs. */ 297 struct spdk_nvmf_rdma_recv *recvs; 298 299 /* Array of size "max_queue_depth" containing 64 byte capsules 300 * used for receive. 301 */ 302 union nvmf_h2c_msg *cmds; 303 struct ibv_mr *cmds_mr; 304 305 /* Array of size "max_queue_depth" containing 16 byte completions 306 * to be sent back to the user. 307 */ 308 union nvmf_c2h_msg *cpls; 309 struct ibv_mr *cpls_mr; 310 311 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 312 * buffers to be used for in capsule data. 313 */ 314 void *bufs; 315 struct ibv_mr *bufs_mr; 316 317 /* Receives that are waiting for a request object */ 318 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 319 320 /* Queue to track free requests */ 321 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 322 }; 323 324 struct spdk_nvmf_rdma_qpair { 325 struct spdk_nvmf_qpair qpair; 326 327 struct spdk_nvmf_rdma_port *port; 328 struct spdk_nvmf_rdma_poller *poller; 329 330 struct rdma_cm_id *cm_id; 331 struct ibv_srq *srq; 332 struct rdma_cm_id *listen_id; 333 334 /* The maximum number of I/O outstanding on this connection at one time */ 335 uint16_t max_queue_depth; 336 337 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 338 uint16_t max_read_depth; 339 340 /* The maximum number of RDMA SEND operations at one time */ 341 uint32_t max_send_depth; 342 343 /* The current number of outstanding WRs from this qpair's 344 * recv queue. Should not exceed device->attr.max_queue_depth. 345 */ 346 uint16_t current_recv_depth; 347 348 /* The current number of active RDMA READ operations */ 349 uint16_t current_read_depth; 350 351 /* The current number of posted WRs from this qpair's 352 * send queue. Should not exceed max_send_depth. 353 */ 354 uint32_t current_send_depth; 355 356 /* The maximum number of SGEs per WR on the send queue */ 357 uint32_t max_send_sge; 358 359 /* The maximum number of SGEs per WR on the recv queue */ 360 uint32_t max_recv_sge; 361 362 struct spdk_nvmf_rdma_resources *resources; 363 364 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 365 366 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 367 368 /* Number of requests not in the free state */ 369 uint32_t qd; 370 371 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 372 373 /* IBV queue pair attributes: they are used to manage 374 * qp state and recover from errors. 375 */ 376 enum ibv_qp_state ibv_state; 377 378 uint32_t disconnect_flags; 379 380 /* Poller registered in case the qpair doesn't properly 381 * complete the qpair destruct process and becomes defunct. 382 */ 383 384 struct spdk_poller *destruct_poller; 385 386 /* There are several ways a disconnect can start on a qpair 387 * and they are not all mutually exclusive. It is important 388 * that we only initialize one of these paths. 389 */ 390 bool disconnect_started; 391 /* Lets us know that we have received the last_wqe event. */ 392 bool last_wqe_reached; 393 }; 394 395 struct spdk_nvmf_rdma_poller { 396 struct spdk_nvmf_rdma_device *device; 397 struct spdk_nvmf_rdma_poll_group *group; 398 399 int num_cqe; 400 int required_num_wr; 401 struct ibv_cq *cq; 402 403 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 404 uint16_t max_srq_depth; 405 406 /* Shared receive queue */ 407 struct ibv_srq *srq; 408 409 struct spdk_nvmf_rdma_resources *resources; 410 411 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 412 413 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 414 }; 415 416 struct spdk_nvmf_rdma_poll_group { 417 struct spdk_nvmf_transport_poll_group group; 418 419 /* Requests that are waiting to obtain a data buffer */ 420 TAILQ_HEAD(, spdk_nvmf_rdma_request) pending_data_buf_queue; 421 422 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 423 }; 424 425 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 426 struct spdk_nvmf_rdma_device { 427 struct ibv_device_attr attr; 428 struct ibv_context *context; 429 430 struct spdk_mem_map *map; 431 struct ibv_pd *pd; 432 433 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 434 }; 435 436 struct spdk_nvmf_rdma_port { 437 struct spdk_nvme_transport_id trid; 438 struct rdma_cm_id *id; 439 struct spdk_nvmf_rdma_device *device; 440 uint32_t ref; 441 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 442 }; 443 444 struct spdk_nvmf_rdma_transport { 445 struct spdk_nvmf_transport transport; 446 447 struct rdma_event_channel *event_channel; 448 449 struct spdk_mempool *data_wr_pool; 450 451 pthread_mutex_t lock; 452 453 /* fields used to poll RDMA/IB events */ 454 nfds_t npoll_fds; 455 struct pollfd *poll_fds; 456 457 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 458 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 459 }; 460 461 static inline int 462 spdk_nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 463 { 464 switch (state) { 465 case IBV_QPS_RESET: 466 case IBV_QPS_INIT: 467 case IBV_QPS_RTR: 468 case IBV_QPS_RTS: 469 case IBV_QPS_SQD: 470 case IBV_QPS_SQE: 471 case IBV_QPS_ERR: 472 return 0; 473 default: 474 return -1; 475 } 476 } 477 478 static enum ibv_qp_state 479 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 480 enum ibv_qp_state old_state, new_state; 481 struct ibv_qp_attr qp_attr; 482 struct ibv_qp_init_attr init_attr; 483 int rc; 484 485 old_state = rqpair->ibv_state; 486 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 487 g_spdk_nvmf_ibv_query_mask, &init_attr); 488 489 if (rc) 490 { 491 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 492 return IBV_QPS_ERR + 1; 493 } 494 495 new_state = qp_attr.qp_state; 496 rqpair->ibv_state = new_state; 497 qp_attr.ah_attr.port_num = qp_attr.port_num; 498 499 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 500 if (rc) 501 { 502 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 503 /* 504 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 505 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 506 */ 507 return IBV_QPS_ERR + 1; 508 } 509 510 if (old_state != new_state) 511 { 512 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 513 (uintptr_t)rqpair->cm_id, new_state); 514 } 515 return new_state; 516 } 517 518 static const char *str_ibv_qp_state[] = { 519 "IBV_QPS_RESET", 520 "IBV_QPS_INIT", 521 "IBV_QPS_RTR", 522 "IBV_QPS_RTS", 523 "IBV_QPS_SQD", 524 "IBV_QPS_SQE", 525 "IBV_QPS_ERR", 526 "IBV_QPS_UNKNOWN" 527 }; 528 529 static int 530 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 531 enum ibv_qp_state new_state) 532 { 533 struct ibv_qp_attr qp_attr; 534 struct ibv_qp_init_attr init_attr; 535 int rc; 536 enum ibv_qp_state state; 537 static int attr_mask_rc[] = { 538 [IBV_QPS_RESET] = IBV_QP_STATE, 539 [IBV_QPS_INIT] = (IBV_QP_STATE | 540 IBV_QP_PKEY_INDEX | 541 IBV_QP_PORT | 542 IBV_QP_ACCESS_FLAGS), 543 [IBV_QPS_RTR] = (IBV_QP_STATE | 544 IBV_QP_AV | 545 IBV_QP_PATH_MTU | 546 IBV_QP_DEST_QPN | 547 IBV_QP_RQ_PSN | 548 IBV_QP_MAX_DEST_RD_ATOMIC | 549 IBV_QP_MIN_RNR_TIMER), 550 [IBV_QPS_RTS] = (IBV_QP_STATE | 551 IBV_QP_SQ_PSN | 552 IBV_QP_TIMEOUT | 553 IBV_QP_RETRY_CNT | 554 IBV_QP_RNR_RETRY | 555 IBV_QP_MAX_QP_RD_ATOMIC), 556 [IBV_QPS_SQD] = IBV_QP_STATE, 557 [IBV_QPS_SQE] = IBV_QP_STATE, 558 [IBV_QPS_ERR] = IBV_QP_STATE, 559 }; 560 561 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 562 if (rc) { 563 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 564 rqpair->qpair.qid, new_state); 565 return rc; 566 } 567 568 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 569 g_spdk_nvmf_ibv_query_mask, &init_attr); 570 571 if (rc) { 572 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 573 assert(false); 574 } 575 576 qp_attr.cur_qp_state = rqpair->ibv_state; 577 qp_attr.qp_state = new_state; 578 579 rc = ibv_modify_qp(rqpair->cm_id->qp, &qp_attr, 580 attr_mask_rc[new_state]); 581 582 if (rc) { 583 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 584 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 585 return rc; 586 } 587 588 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 589 590 if (state != new_state) { 591 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 592 rqpair->qpair.qid, str_ibv_qp_state[new_state], 593 str_ibv_qp_state[state]); 594 return -1; 595 } 596 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 597 str_ibv_qp_state[state]); 598 return 0; 599 } 600 601 static void 602 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 603 struct spdk_nvmf_rdma_transport *rtransport) 604 { 605 struct spdk_nvmf_rdma_request_data *current_data_wr = NULL, *next_data_wr = NULL; 606 struct ibv_send_wr *send_wr; 607 int i; 608 609 rdma_req->num_outstanding_data_wr = 0; 610 current_data_wr = &rdma_req->data; 611 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 612 current_data_wr->wr.sg_list[i].addr = 0; 613 current_data_wr->wr.sg_list[i].length = 0; 614 current_data_wr->wr.sg_list[i].lkey = 0; 615 } 616 current_data_wr->wr.num_sge = 0; 617 618 send_wr = current_data_wr->wr.next; 619 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr) { 620 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 621 } 622 while (next_data_wr) { 623 current_data_wr = next_data_wr; 624 send_wr = current_data_wr->wr.next; 625 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr) { 626 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 627 } else { 628 next_data_wr = NULL; 629 } 630 631 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 632 current_data_wr->wr.sg_list[i].addr = 0; 633 current_data_wr->wr.sg_list[i].length = 0; 634 current_data_wr->wr.sg_list[i].lkey = 0; 635 } 636 current_data_wr->wr.num_sge = 0; 637 spdk_mempool_put(rtransport->data_wr_pool, current_data_wr); 638 } 639 } 640 641 static void 642 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 643 { 644 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->data_from_pool); 645 if (req->req.cmd) { 646 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 647 } 648 if (req->recv) { 649 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 650 } 651 } 652 653 static void 654 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 655 { 656 int i; 657 658 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 659 for (i = 0; i < rqpair->max_queue_depth; i++) { 660 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 661 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 662 } 663 } 664 } 665 666 static void 667 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 668 { 669 if (resources->cmds_mr) { 670 ibv_dereg_mr(resources->cmds_mr); 671 } 672 673 if (resources->cpls_mr) { 674 ibv_dereg_mr(resources->cpls_mr); 675 } 676 677 if (resources->bufs_mr) { 678 ibv_dereg_mr(resources->bufs_mr); 679 } 680 681 spdk_dma_free(resources->cmds); 682 spdk_dma_free(resources->cpls); 683 spdk_dma_free(resources->bufs); 684 free(resources->reqs); 685 free(resources->recvs); 686 free(resources); 687 } 688 689 690 static struct spdk_nvmf_rdma_resources * 691 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 692 { 693 struct spdk_nvmf_rdma_resources *resources; 694 struct spdk_nvmf_rdma_request *rdma_req; 695 struct spdk_nvmf_rdma_recv *rdma_recv; 696 struct ibv_qp *qp; 697 struct ibv_srq *srq; 698 uint32_t i; 699 int rc; 700 701 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 702 if (!resources) { 703 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 704 return NULL; 705 } 706 707 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 708 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 709 resources->cmds = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 710 0x1000, NULL); 711 resources->cpls = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 712 0x1000, NULL); 713 714 if (opts->in_capsule_data_size > 0) { 715 resources->bufs = spdk_dma_zmalloc(opts->max_queue_depth * 716 opts->in_capsule_data_size, 717 0x1000, NULL); 718 } 719 720 if (!resources->reqs || !resources->recvs || !resources->cmds || 721 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 722 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 723 goto cleanup; 724 } 725 726 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 727 opts->max_queue_depth * sizeof(*resources->cmds), 728 IBV_ACCESS_LOCAL_WRITE); 729 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 730 opts->max_queue_depth * sizeof(*resources->cpls), 731 0); 732 733 if (opts->in_capsule_data_size) { 734 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 735 opts->max_queue_depth * 736 opts->in_capsule_data_size, 737 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 738 } 739 740 if (!resources->cmds_mr || !resources->cpls_mr || 741 (opts->in_capsule_data_size && 742 !resources->bufs_mr)) { 743 goto cleanup; 744 } 745 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 746 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 747 resources->cmds_mr->lkey); 748 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 749 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 750 resources->cpls_mr->lkey); 751 if (resources->bufs && resources->bufs_mr) { 752 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 753 resources->bufs, opts->max_queue_depth * 754 opts->in_capsule_data_size, resources->bufs_mr->lkey); 755 } 756 757 /* Initialize queues */ 758 STAILQ_INIT(&resources->incoming_queue); 759 STAILQ_INIT(&resources->free_queue); 760 761 for (i = 0; i < opts->max_queue_depth; i++) { 762 struct ibv_recv_wr *bad_wr = NULL; 763 764 rdma_recv = &resources->recvs[i]; 765 rdma_recv->qpair = opts->qpair; 766 767 /* Set up memory to receive commands */ 768 if (resources->bufs) { 769 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 770 opts->in_capsule_data_size)); 771 } 772 773 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 774 775 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 776 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 777 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 778 rdma_recv->wr.num_sge = 1; 779 780 if (rdma_recv->buf && resources->bufs_mr) { 781 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 782 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 783 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 784 rdma_recv->wr.num_sge++; 785 } 786 787 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 788 rdma_recv->wr.sg_list = rdma_recv->sgl; 789 if (opts->shared) { 790 srq = (struct ibv_srq *)opts->qp; 791 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 792 } else { 793 qp = (struct ibv_qp *)opts->qp; 794 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 795 } 796 if (rc) { 797 goto cleanup; 798 } 799 } 800 801 for (i = 0; i < opts->max_queue_depth; i++) { 802 rdma_req = &resources->reqs[i]; 803 804 if (opts->qpair != NULL) { 805 rdma_req->req.qpair = &opts->qpair->qpair; 806 } else { 807 rdma_req->req.qpair = NULL; 808 } 809 rdma_req->req.cmd = NULL; 810 811 /* Set up memory to send responses */ 812 rdma_req->req.rsp = &resources->cpls[i]; 813 814 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 815 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 816 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 817 818 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 819 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 820 rdma_req->rsp.wr.next = NULL; 821 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 822 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 823 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 824 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 825 826 /* Set up memory for data buffers */ 827 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 828 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 829 rdma_req->data.wr.next = NULL; 830 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 831 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 832 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 833 834 /* Initialize request state to FREE */ 835 rdma_req->state = RDMA_REQUEST_STATE_FREE; 836 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 837 } 838 839 return resources; 840 841 cleanup: 842 nvmf_rdma_resources_destroy(resources); 843 return NULL; 844 } 845 846 static void 847 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 848 { 849 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 850 struct ibv_recv_wr *bad_recv_wr = NULL; 851 int rc; 852 853 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 854 855 spdk_poller_unregister(&rqpair->destruct_poller); 856 857 if (rqpair->qd != 0) { 858 if (rqpair->srq == NULL) { 859 nvmf_rdma_dump_qpair_contents(rqpair); 860 } 861 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 862 } 863 864 if (rqpair->poller) { 865 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 866 867 if (rqpair->srq != NULL) { 868 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 869 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 870 if (rqpair == rdma_recv->qpair) { 871 STAILQ_REMOVE_HEAD(&rqpair->resources->incoming_queue, link); 872 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 873 if (rc) { 874 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 875 } 876 } 877 } 878 } 879 } 880 881 if (rqpair->cm_id) { 882 rdma_destroy_qp(rqpair->cm_id); 883 rdma_destroy_id(rqpair->cm_id); 884 885 if (rqpair->poller != NULL && rqpair->srq == NULL) { 886 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 887 } 888 } 889 890 if (rqpair->srq == NULL) { 891 nvmf_rdma_resources_destroy(rqpair->resources); 892 } 893 894 free(rqpair); 895 } 896 897 static int 898 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 899 { 900 struct spdk_nvmf_rdma_poller *rpoller; 901 int rc, num_cqe, required_num_wr; 902 903 /* Enlarge CQ size dynamically */ 904 rpoller = rqpair->poller; 905 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 906 num_cqe = rpoller->num_cqe; 907 if (num_cqe < required_num_wr) { 908 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 909 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 910 } 911 912 if (rpoller->num_cqe != num_cqe) { 913 if (required_num_wr > device->attr.max_cqe) { 914 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 915 required_num_wr, device->attr.max_cqe); 916 return -1; 917 } 918 919 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 920 rc = ibv_resize_cq(rpoller->cq, num_cqe); 921 if (rc) { 922 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 923 return -1; 924 } 925 926 rpoller->num_cqe = num_cqe; 927 } 928 929 rpoller->required_num_wr = required_num_wr; 930 return 0; 931 } 932 933 static int 934 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 935 { 936 struct spdk_nvmf_rdma_qpair *rqpair; 937 int rc; 938 struct spdk_nvmf_rdma_transport *rtransport; 939 struct spdk_nvmf_transport *transport; 940 struct spdk_nvmf_rdma_resource_opts opts; 941 struct spdk_nvmf_rdma_device *device; 942 struct ibv_qp_init_attr ibv_init_attr; 943 944 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 945 device = rqpair->port->device; 946 947 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 948 ibv_init_attr.qp_context = rqpair; 949 ibv_init_attr.qp_type = IBV_QPT_RC; 950 ibv_init_attr.send_cq = rqpair->poller->cq; 951 ibv_init_attr.recv_cq = rqpair->poller->cq; 952 953 if (rqpair->srq) { 954 ibv_init_attr.srq = rqpair->srq; 955 } else { 956 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 957 1; /* RECV operations + dummy drain WR */ 958 } 959 960 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 961 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 962 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 963 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 964 965 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 966 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 967 goto error; 968 } 969 970 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 971 if (rc) { 972 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 973 goto error; 974 } 975 976 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2 + 1), 977 ibv_init_attr.cap.max_send_wr); 978 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 979 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 980 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 981 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 982 983 if (rqpair->poller->srq == NULL) { 984 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 985 transport = &rtransport->transport; 986 987 opts.qp = rqpair->cm_id->qp; 988 opts.pd = rqpair->cm_id->pd; 989 opts.qpair = rqpair; 990 opts.shared = false; 991 opts.max_queue_depth = rqpair->max_queue_depth; 992 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 993 994 rqpair->resources = nvmf_rdma_resources_create(&opts); 995 996 if (!rqpair->resources) { 997 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 998 goto error; 999 } 1000 } else { 1001 rqpair->resources = rqpair->poller->resources; 1002 } 1003 1004 rqpair->current_recv_depth = 0; 1005 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1006 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1007 1008 return 0; 1009 1010 error: 1011 rdma_destroy_id(rqpair->cm_id); 1012 rqpair->cm_id = NULL; 1013 spdk_nvmf_rdma_qpair_destroy(rqpair); 1014 return -1; 1015 } 1016 1017 static int 1018 request_transfer_in(struct spdk_nvmf_request *req) 1019 { 1020 int rc; 1021 struct spdk_nvmf_rdma_request *rdma_req; 1022 struct spdk_nvmf_qpair *qpair; 1023 struct spdk_nvmf_rdma_qpair *rqpair; 1024 struct ibv_send_wr *bad_wr = NULL; 1025 1026 qpair = req->qpair; 1027 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1028 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1029 1030 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1031 assert(rdma_req != NULL); 1032 1033 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA READ POSTED. Request: %p Connection: %p\n", req, qpair); 1034 1035 rc = ibv_post_send(rqpair->cm_id->qp, &rdma_req->data.wr, &bad_wr); 1036 if (rc) { 1037 SPDK_ERRLOG("Unable to transfer data from host to target\n"); 1038 return -1; 1039 } 1040 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1041 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1042 return 0; 1043 } 1044 1045 static int 1046 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1047 { 1048 int rc; 1049 struct spdk_nvmf_rdma_request *rdma_req; 1050 struct spdk_nvmf_qpair *qpair; 1051 struct spdk_nvmf_rdma_qpair *rqpair; 1052 struct spdk_nvme_cpl *rsp; 1053 struct ibv_recv_wr *bad_recv_wr = NULL; 1054 struct ibv_send_wr *send_wr, *bad_send_wr = NULL; 1055 1056 *data_posted = 0; 1057 qpair = req->qpair; 1058 rsp = &req->rsp->nvme_cpl; 1059 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1060 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1061 1062 /* Advance our sq_head pointer */ 1063 if (qpair->sq_head == qpair->sq_head_max) { 1064 qpair->sq_head = 0; 1065 } else { 1066 qpair->sq_head++; 1067 } 1068 rsp->sqhd = qpair->sq_head; 1069 1070 /* Post the capsule to the recv buffer */ 1071 assert(rdma_req->recv != NULL); 1072 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA RECV POSTED. Recv: %p Connection: %p\n", rdma_req->recv, 1073 rqpair); 1074 if (rqpair->srq == NULL) { 1075 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_req->recv->wr, &bad_recv_wr); 1076 } else { 1077 rdma_req->recv->qpair = NULL; 1078 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 1079 } 1080 1081 if (rc) { 1082 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 1083 return rc; 1084 } 1085 rdma_req->recv = NULL; 1086 assert(rqpair->current_recv_depth > 0); 1087 rqpair->current_recv_depth--; 1088 1089 /* Build the response which consists of optional 1090 * RDMA WRITEs to transfer data, plus an RDMA SEND 1091 * containing the response. 1092 */ 1093 send_wr = &rdma_req->rsp.wr; 1094 1095 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 1096 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1097 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA WRITE POSTED. Request: %p Connection: %p\n", req, qpair); 1098 send_wr = &rdma_req->data.wr; 1099 *data_posted = 1; 1100 } 1101 1102 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA SEND POSTED. Request: %p Connection: %p\n", req, qpair); 1103 1104 /* Send the completion */ 1105 rc = ibv_post_send(rqpair->cm_id->qp, send_wr, &bad_send_wr); 1106 if (rc) { 1107 SPDK_ERRLOG("Unable to send response capsule\n"); 1108 return rc; 1109 } 1110 /* +1 for the rsp wr */ 1111 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr + 1; 1112 1113 return 0; 1114 } 1115 1116 static int 1117 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1118 { 1119 struct spdk_nvmf_rdma_accept_private_data accept_data; 1120 struct rdma_conn_param ctrlr_event_data = {}; 1121 int rc; 1122 1123 accept_data.recfmt = 0; 1124 accept_data.crqsize = rqpair->max_queue_depth; 1125 1126 ctrlr_event_data.private_data = &accept_data; 1127 ctrlr_event_data.private_data_len = sizeof(accept_data); 1128 if (id->ps == RDMA_PS_TCP) { 1129 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1130 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1131 } 1132 1133 rc = rdma_accept(id, &ctrlr_event_data); 1134 if (rc) { 1135 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 1136 } else { 1137 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1138 } 1139 1140 return rc; 1141 } 1142 1143 static void 1144 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1145 { 1146 struct spdk_nvmf_rdma_reject_private_data rej_data; 1147 1148 rej_data.recfmt = 0; 1149 rej_data.sts = error; 1150 1151 rdma_reject(id, &rej_data, sizeof(rej_data)); 1152 } 1153 1154 static int 1155 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 1156 new_qpair_fn cb_fn) 1157 { 1158 struct spdk_nvmf_rdma_transport *rtransport; 1159 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1160 struct spdk_nvmf_rdma_port *port; 1161 struct rdma_conn_param *rdma_param = NULL; 1162 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1163 uint16_t max_queue_depth; 1164 uint16_t max_read_depth; 1165 1166 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1167 1168 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1169 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1170 1171 rdma_param = &event->param.conn; 1172 if (rdma_param->private_data == NULL || 1173 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1174 SPDK_ERRLOG("connect request: no private data provided\n"); 1175 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1176 return -1; 1177 } 1178 1179 private_data = rdma_param->private_data; 1180 if (private_data->recfmt != 0) { 1181 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1182 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1183 return -1; 1184 } 1185 1186 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1187 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1188 1189 port = event->listen_id->context; 1190 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1191 event->listen_id, event->listen_id->verbs, port); 1192 1193 /* Figure out the supported queue depth. This is a multi-step process 1194 * that takes into account hardware maximums, host provided values, 1195 * and our target's internal memory limits */ 1196 1197 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1198 1199 /* Start with the maximum queue depth allowed by the target */ 1200 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1201 max_read_depth = rtransport->transport.opts.max_queue_depth; 1202 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1203 rtransport->transport.opts.max_queue_depth); 1204 1205 /* Next check the local NIC's hardware limitations */ 1206 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1207 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1208 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1209 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1210 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1211 1212 /* Next check the remote NIC's hardware limitations */ 1213 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1214 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1215 rdma_param->initiator_depth, rdma_param->responder_resources); 1216 if (rdma_param->initiator_depth > 0) { 1217 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1218 } 1219 1220 /* Finally check for the host software requested values, which are 1221 * optional. */ 1222 if (rdma_param->private_data != NULL && 1223 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1224 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1225 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1226 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1227 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1228 } 1229 1230 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1231 max_queue_depth, max_read_depth); 1232 1233 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1234 if (rqpair == NULL) { 1235 SPDK_ERRLOG("Could not allocate new connection.\n"); 1236 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1237 return -1; 1238 } 1239 1240 rqpair->port = port; 1241 rqpair->max_queue_depth = max_queue_depth; 1242 rqpair->max_read_depth = max_read_depth; 1243 rqpair->cm_id = event->id; 1244 rqpair->listen_id = event->listen_id; 1245 rqpair->qpair.transport = transport; 1246 1247 event->id->context = &rqpair->qpair; 1248 1249 cb_fn(&rqpair->qpair); 1250 1251 return 0; 1252 } 1253 1254 static int 1255 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1256 enum spdk_mem_map_notify_action action, 1257 void *vaddr, size_t size) 1258 { 1259 struct ibv_pd *pd = cb_ctx; 1260 struct ibv_mr *mr; 1261 1262 switch (action) { 1263 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1264 if (!g_nvmf_hooks.get_rkey) { 1265 mr = ibv_reg_mr(pd, vaddr, size, 1266 IBV_ACCESS_LOCAL_WRITE | 1267 IBV_ACCESS_REMOTE_READ | 1268 IBV_ACCESS_REMOTE_WRITE); 1269 if (mr == NULL) { 1270 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1271 return -1; 1272 } else { 1273 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1274 } 1275 } else { 1276 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1277 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1278 } 1279 break; 1280 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1281 if (!g_nvmf_hooks.get_rkey) { 1282 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1283 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1284 if (mr) { 1285 ibv_dereg_mr(mr); 1286 } 1287 } 1288 break; 1289 } 1290 1291 return 0; 1292 } 1293 1294 static int 1295 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1296 { 1297 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1298 return addr_1 == addr_2; 1299 } 1300 1301 static void 1302 spdk_nvmf_rdma_request_free_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1303 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1304 uint32_t num_buffers) 1305 { 1306 uint32_t i; 1307 1308 for (i = 0; i < num_buffers; i++) { 1309 if (group->buf_cache_count < group->buf_cache_size) { 1310 STAILQ_INSERT_HEAD(&group->buf_cache, 1311 (struct spdk_nvmf_transport_pg_cache_buf *)rdma_req->buffers[i], link); 1312 group->buf_cache_count++; 1313 } else { 1314 spdk_mempool_put(transport->data_buf_pool, rdma_req->buffers[i]); 1315 } 1316 rdma_req->req.iov[i].iov_base = NULL; 1317 rdma_req->buffers[i] = NULL; 1318 rdma_req->req.iov[i].iov_len = 0; 1319 1320 } 1321 rdma_req->data_from_pool = false; 1322 } 1323 1324 static int 1325 nvmf_rdma_request_get_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1326 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1327 uint32_t num_buffers) 1328 { 1329 uint32_t i = 0; 1330 1331 while (i < num_buffers) { 1332 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1333 group->buf_cache_count--; 1334 rdma_req->buffers[i] = STAILQ_FIRST(&group->buf_cache); 1335 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1336 assert(rdma_req->buffers[i] != NULL); 1337 i++; 1338 } else { 1339 if (spdk_mempool_get_bulk(transport->data_buf_pool, &rdma_req->buffers[i], num_buffers - i)) { 1340 goto err_exit; 1341 } 1342 i += num_buffers - i; 1343 } 1344 } 1345 1346 return 0; 1347 1348 err_exit: 1349 spdk_nvmf_rdma_request_free_buffers(rdma_req, group, transport, i); 1350 return -ENOMEM; 1351 } 1352 1353 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1354 1355 static spdk_nvme_data_transfer_t 1356 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1357 { 1358 enum spdk_nvme_data_transfer xfer; 1359 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1360 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1361 1362 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1363 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1364 rdma_req->rsp.wr.imm_data = 0; 1365 #endif 1366 1367 /* Figure out data transfer direction */ 1368 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1369 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1370 } else { 1371 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1372 1373 /* Some admin commands are special cases */ 1374 if ((rdma_req->req.qpair->qid == 0) && 1375 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1376 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1377 switch (cmd->cdw10 & 0xff) { 1378 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1379 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1380 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1381 break; 1382 default: 1383 xfer = SPDK_NVME_DATA_NONE; 1384 } 1385 } 1386 } 1387 1388 if (xfer == SPDK_NVME_DATA_NONE) { 1389 return xfer; 1390 } 1391 1392 /* Even for commands that may transfer data, they could have specified 0 length. 1393 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1394 */ 1395 switch (sgl->generic.type) { 1396 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1397 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1398 case SPDK_NVME_SGL_TYPE_SEGMENT: 1399 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1400 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1401 if (sgl->unkeyed.length == 0) { 1402 xfer = SPDK_NVME_DATA_NONE; 1403 } 1404 break; 1405 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1406 if (sgl->keyed.length == 0) { 1407 xfer = SPDK_NVME_DATA_NONE; 1408 } 1409 break; 1410 } 1411 1412 return xfer; 1413 } 1414 1415 static int 1416 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1417 struct spdk_nvmf_rdma_request *rdma_req, 1418 uint32_t num_sgl_descriptors) 1419 { 1420 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1421 struct spdk_nvmf_rdma_request_data *current_data_wr; 1422 uint32_t i; 1423 1424 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1425 return -ENOMEM; 1426 } 1427 1428 current_data_wr = &rdma_req->data; 1429 1430 for (i = 0; i < num_sgl_descriptors; i++) { 1431 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1432 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1433 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1434 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1435 } else { 1436 assert(false); 1437 } 1438 work_requests[i]->wr.send_flags = IBV_SEND_SIGNALED; 1439 work_requests[i]->wr.sg_list = work_requests[i]->sgl; 1440 work_requests[i]->wr.wr_id = rdma_req->data.wr.wr_id; 1441 current_data_wr->wr.next = &work_requests[i]->wr; 1442 current_data_wr = work_requests[i]; 1443 } 1444 1445 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1446 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1447 current_data_wr->wr.next = &rdma_req->rsp.wr; 1448 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1449 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1450 current_data_wr->wr.next = NULL; 1451 } 1452 return 0; 1453 } 1454 1455 static int 1456 nvmf_rdma_fill_buffers(struct spdk_nvmf_rdma_transport *rtransport, 1457 struct spdk_nvmf_rdma_poll_group *rgroup, 1458 struct spdk_nvmf_rdma_device *device, 1459 struct spdk_nvmf_rdma_request *rdma_req, 1460 struct ibv_send_wr *wr, 1461 uint32_t length) 1462 { 1463 uint64_t translation_len; 1464 uint32_t remaining_length = length; 1465 uint32_t iovcnt; 1466 uint32_t i = 0; 1467 1468 1469 while (remaining_length) { 1470 iovcnt = rdma_req->req.iovcnt; 1471 rdma_req->req.iov[iovcnt].iov_base = (void *)((uintptr_t)(rdma_req->buffers[iovcnt] + 1472 NVMF_DATA_BUFFER_MASK) & 1473 ~NVMF_DATA_BUFFER_MASK); 1474 rdma_req->req.iov[iovcnt].iov_len = spdk_min(remaining_length, 1475 rtransport->transport.opts.io_unit_size); 1476 rdma_req->req.iovcnt++; 1477 wr->sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[iovcnt].iov_base); 1478 wr->sg_list[i].length = rdma_req->req.iov[iovcnt].iov_len; 1479 translation_len = rdma_req->req.iov[iovcnt].iov_len; 1480 1481 if (!g_nvmf_hooks.get_rkey) { 1482 wr->sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1483 (uint64_t)rdma_req->buffers[iovcnt], &translation_len))->lkey; 1484 } else { 1485 wr->sg_list[i].lkey = spdk_mem_map_translate(device->map, 1486 (uint64_t)rdma_req->buffers[iovcnt], &translation_len); 1487 } 1488 1489 remaining_length -= rdma_req->req.iov[iovcnt].iov_len; 1490 1491 if (translation_len < rdma_req->req.iov[iovcnt].iov_len) { 1492 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions\n"); 1493 return -EINVAL; 1494 } 1495 i++; 1496 } 1497 1498 return 0; 1499 } 1500 1501 static int 1502 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1503 struct spdk_nvmf_rdma_device *device, 1504 struct spdk_nvmf_rdma_request *rdma_req) 1505 { 1506 struct spdk_nvmf_rdma_qpair *rqpair; 1507 struct spdk_nvmf_rdma_poll_group *rgroup; 1508 uint32_t num_buffers; 1509 uint32_t i = 0; 1510 int rc = 0; 1511 1512 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1513 rgroup = rqpair->poller->group; 1514 rdma_req->req.iovcnt = 0; 1515 1516 num_buffers = rdma_req->req.length / rtransport->transport.opts.io_unit_size; 1517 if (rdma_req->req.length % rtransport->transport.opts.io_unit_size) { 1518 num_buffers++; 1519 } 1520 1521 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers)) { 1522 return -ENOMEM; 1523 } 1524 1525 rdma_req->req.iovcnt = 0; 1526 1527 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, &rdma_req->data.wr, 1528 rdma_req->req.length); 1529 if (rc != 0) { 1530 goto err_exit; 1531 } 1532 1533 assert(rdma_req->req.iovcnt <= rqpair->max_send_sge); 1534 1535 rdma_req->data_from_pool = true; 1536 1537 return rc; 1538 1539 err_exit: 1540 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1541 while (i) { 1542 i--; 1543 rdma_req->data.wr.sg_list[i].addr = 0; 1544 rdma_req->data.wr.sg_list[i].length = 0; 1545 rdma_req->data.wr.sg_list[i].lkey = 0; 1546 } 1547 rdma_req->req.iovcnt = 0; 1548 return rc; 1549 } 1550 1551 static int 1552 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1553 struct spdk_nvmf_rdma_device *device, 1554 struct spdk_nvmf_rdma_request *rdma_req) 1555 { 1556 struct spdk_nvmf_rdma_qpair *rqpair; 1557 struct spdk_nvmf_rdma_poll_group *rgroup; 1558 struct ibv_send_wr *current_wr; 1559 struct spdk_nvmf_request *req = &rdma_req->req; 1560 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1561 uint32_t num_sgl_descriptors; 1562 uint32_t num_buffers = 0; 1563 uint32_t i; 1564 int rc; 1565 1566 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1567 rgroup = rqpair->poller->group; 1568 1569 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1570 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1571 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1572 1573 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1574 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1575 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1576 1577 for (i = 0; i < num_sgl_descriptors; i++) { 1578 num_buffers += desc->keyed.length / rtransport->transport.opts.io_unit_size; 1579 if (desc->keyed.length % rtransport->transport.opts.io_unit_size) { 1580 num_buffers++; 1581 } 1582 desc++; 1583 } 1584 /* If the number of buffers is too large, then we know the I/O is larger than allowed. Fail it. */ 1585 if (num_buffers > NVMF_REQ_MAX_BUFFERS) { 1586 return -EINVAL; 1587 } 1588 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1589 num_buffers) != 0) { 1590 return -ENOMEM; 1591 } 1592 1593 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1594 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1595 return -ENOMEM; 1596 } 1597 1598 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1599 current_wr = &rdma_req->data.wr; 1600 1601 req->iovcnt = 0; 1602 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1603 for (i = 0; i < num_sgl_descriptors; i++) { 1604 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1605 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1606 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1607 rc = -EINVAL; 1608 goto err_exit; 1609 } 1610 1611 current_wr->num_sge = 0; 1612 req->length += desc->keyed.length; 1613 1614 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, current_wr, 1615 desc->keyed.length); 1616 if (rc != 0) { 1617 rc = -ENOMEM; 1618 goto err_exit; 1619 } 1620 1621 current_wr->wr.rdma.rkey = desc->keyed.key; 1622 current_wr->wr.rdma.remote_addr = desc->address; 1623 current_wr = current_wr->next; 1624 desc++; 1625 } 1626 1627 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1628 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1629 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1630 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1631 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1632 } 1633 } 1634 #endif 1635 1636 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1637 rdma_req->data_from_pool = true; 1638 1639 return 0; 1640 1641 err_exit: 1642 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1643 nvmf_rdma_request_free_data(rdma_req, rtransport); 1644 return rc; 1645 } 1646 1647 static int 1648 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1649 struct spdk_nvmf_rdma_device *device, 1650 struct spdk_nvmf_rdma_request *rdma_req) 1651 { 1652 struct spdk_nvme_cmd *cmd; 1653 struct spdk_nvme_cpl *rsp; 1654 struct spdk_nvme_sgl_descriptor *sgl; 1655 int rc; 1656 1657 cmd = &rdma_req->req.cmd->nvme_cmd; 1658 rsp = &rdma_req->req.rsp->nvme_cpl; 1659 sgl = &cmd->dptr.sgl1; 1660 1661 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1662 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1663 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1664 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1665 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1666 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1667 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1668 return -1; 1669 } 1670 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1671 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1672 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1673 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1674 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1675 } 1676 } 1677 #endif 1678 1679 /* fill request length and populate iovs */ 1680 rdma_req->req.length = sgl->keyed.length; 1681 1682 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1683 /* No available buffers. Queue this request up. */ 1684 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1685 return 0; 1686 } 1687 1688 /* backward compatible */ 1689 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1690 1691 /* rdma wr specifics */ 1692 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1693 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1694 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1695 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1696 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 1697 rdma_req->data.wr.next = &rdma_req->rsp.wr; 1698 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1699 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 1700 rdma_req->data.wr.next = NULL; 1701 } 1702 1703 /* set the number of outstanding data WRs for this request. */ 1704 rdma_req->num_outstanding_data_wr = 1; 1705 1706 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1707 rdma_req->req.iovcnt); 1708 1709 return 0; 1710 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1711 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1712 uint64_t offset = sgl->address; 1713 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1714 1715 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1716 offset, sgl->unkeyed.length); 1717 1718 if (offset > max_len) { 1719 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1720 offset, max_len); 1721 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1722 return -1; 1723 } 1724 max_len -= (uint32_t)offset; 1725 1726 if (sgl->unkeyed.length > max_len) { 1727 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1728 sgl->unkeyed.length, max_len); 1729 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1730 return -1; 1731 } 1732 1733 rdma_req->num_outstanding_data_wr = 0; 1734 rdma_req->req.data = rdma_req->recv->buf + offset; 1735 rdma_req->data_from_pool = false; 1736 rdma_req->req.length = sgl->unkeyed.length; 1737 1738 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1739 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1740 rdma_req->req.iovcnt = 1; 1741 1742 return 0; 1743 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1744 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1745 1746 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1747 if (rc == -ENOMEM) { 1748 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1749 return 0; 1750 } else if (rc == -EINVAL) { 1751 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1752 return -1; 1753 } 1754 1755 /* backward compatible */ 1756 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1757 1758 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1759 rdma_req->req.iovcnt); 1760 1761 return 0; 1762 } 1763 1764 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1765 sgl->generic.type, sgl->generic.subtype); 1766 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1767 return -1; 1768 } 1769 1770 static void 1771 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1772 struct spdk_nvmf_rdma_transport *rtransport) 1773 { 1774 struct spdk_nvmf_rdma_qpair *rqpair; 1775 struct spdk_nvmf_rdma_poll_group *rgroup; 1776 1777 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1778 if (rdma_req->data_from_pool) { 1779 rgroup = rqpair->poller->group; 1780 1781 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1782 rdma_req->req.iovcnt); 1783 } 1784 nvmf_rdma_request_free_data(rdma_req, rtransport); 1785 rdma_req->req.length = 0; 1786 rdma_req->req.iovcnt = 0; 1787 rdma_req->req.data = NULL; 1788 rdma_req->data.wr.next = NULL; 1789 rqpair->qd--; 1790 1791 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1792 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1793 } 1794 1795 static bool 1796 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1797 struct spdk_nvmf_rdma_request *rdma_req) 1798 { 1799 struct spdk_nvmf_rdma_qpair *rqpair; 1800 struct spdk_nvmf_rdma_device *device; 1801 struct spdk_nvmf_rdma_poll_group *rgroup; 1802 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1803 int rc; 1804 struct spdk_nvmf_rdma_recv *rdma_recv; 1805 enum spdk_nvmf_rdma_request_state prev_state; 1806 bool progress = false; 1807 int data_posted; 1808 1809 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1810 device = rqpair->port->device; 1811 rgroup = rqpair->poller->group; 1812 1813 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1814 1815 /* If the queue pair is in an error state, force the request to the completed state 1816 * to release resources. */ 1817 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1818 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1819 TAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, link); 1820 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1821 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1822 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1823 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1824 } 1825 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1826 } 1827 1828 /* The loop here is to allow for several back-to-back state changes. */ 1829 do { 1830 prev_state = rdma_req->state; 1831 1832 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1833 1834 switch (rdma_req->state) { 1835 case RDMA_REQUEST_STATE_FREE: 1836 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1837 * to escape this state. */ 1838 break; 1839 case RDMA_REQUEST_STATE_NEW: 1840 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1841 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1842 rdma_recv = rdma_req->recv; 1843 1844 /* The first element of the SGL is the NVMe command */ 1845 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1846 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1847 1848 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1849 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1850 break; 1851 } 1852 1853 /* The next state transition depends on the data transfer needs of this request. */ 1854 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1855 1856 /* If no data to transfer, ready to execute. */ 1857 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1858 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1859 break; 1860 } 1861 1862 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1863 TAILQ_INSERT_TAIL(&rgroup->pending_data_buf_queue, rdma_req, link); 1864 break; 1865 case RDMA_REQUEST_STATE_NEED_BUFFER: 1866 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1867 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1868 1869 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1870 1871 if (rdma_req != TAILQ_FIRST(&rgroup->pending_data_buf_queue)) { 1872 /* This request needs to wait in line to obtain a buffer */ 1873 break; 1874 } 1875 1876 /* Try to get a data buffer */ 1877 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1878 if (rc < 0) { 1879 TAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, link); 1880 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1881 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1882 break; 1883 } 1884 1885 if (!rdma_req->req.data) { 1886 /* No buffers available. */ 1887 break; 1888 } 1889 1890 TAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, link); 1891 1892 /* If data is transferring from host to controller and the data didn't 1893 * arrive using in capsule data, we need to do a transfer from the host. 1894 */ 1895 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1896 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1897 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1898 break; 1899 } 1900 1901 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1902 break; 1903 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1904 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1905 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1906 1907 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1908 /* This request needs to wait in line to perform RDMA */ 1909 break; 1910 } 1911 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1912 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1913 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1914 break; 1915 } 1916 1917 /* We have already verified that this request is the head of the queue. */ 1918 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 1919 1920 rc = request_transfer_in(&rdma_req->req); 1921 if (!rc) { 1922 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 1923 } else { 1924 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1925 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1926 } 1927 break; 1928 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1929 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1930 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1931 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1932 * to escape this state. */ 1933 break; 1934 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1935 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1936 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1937 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 1938 spdk_nvmf_request_exec(&rdma_req->req); 1939 break; 1940 case RDMA_REQUEST_STATE_EXECUTING: 1941 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1942 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1943 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1944 * to escape this state. */ 1945 break; 1946 case RDMA_REQUEST_STATE_EXECUTED: 1947 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1948 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1949 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1950 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 1951 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 1952 } else { 1953 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1954 } 1955 break; 1956 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 1957 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 1958 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1959 1960 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 1961 /* This request needs to wait in line to perform RDMA */ 1962 break; 1963 } 1964 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 1965 rqpair->max_send_depth) { 1966 /* We can only have so many WRs outstanding. we have to wait until some finish. 1967 * +1 since each request has an additional wr in the resp. */ 1968 break; 1969 } 1970 1971 /* We have already verified that this request is the head of the queue. */ 1972 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 1973 1974 /* The data transfer will be kicked off from 1975 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1976 */ 1977 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1978 break; 1979 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1980 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 1981 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1982 rc = request_transfer_out(&rdma_req->req, &data_posted); 1983 assert(rc == 0); /* No good way to handle this currently */ 1984 if (rc) { 1985 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1986 } else { 1987 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1988 RDMA_REQUEST_STATE_COMPLETING; 1989 } 1990 break; 1991 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1992 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 1993 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1994 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 1995 * to escape this state. */ 1996 break; 1997 case RDMA_REQUEST_STATE_COMPLETING: 1998 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 1999 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2000 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2001 * to escape this state. */ 2002 break; 2003 case RDMA_REQUEST_STATE_COMPLETED: 2004 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2005 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2006 2007 nvmf_rdma_request_free(rdma_req, rtransport); 2008 break; 2009 case RDMA_REQUEST_NUM_STATES: 2010 default: 2011 assert(0); 2012 break; 2013 } 2014 2015 if (rdma_req->state != prev_state) { 2016 progress = true; 2017 } 2018 } while (rdma_req->state != prev_state); 2019 2020 return progress; 2021 } 2022 2023 /* Public API callbacks begin here */ 2024 2025 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2026 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2027 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2028 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 2029 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2030 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2031 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2032 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4096 2033 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2034 2035 static void 2036 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2037 { 2038 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2039 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2040 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2041 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2042 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2043 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2044 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2045 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2046 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2047 } 2048 2049 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2050 .notify_cb = spdk_nvmf_rdma_mem_notify, 2051 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 2052 }; 2053 2054 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2055 2056 static struct spdk_nvmf_transport * 2057 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2058 { 2059 int rc; 2060 struct spdk_nvmf_rdma_transport *rtransport; 2061 struct spdk_nvmf_rdma_device *device, *tmp; 2062 struct ibv_pd *pd; 2063 struct ibv_context **contexts; 2064 uint32_t i; 2065 int flag; 2066 uint32_t sge_count; 2067 uint32_t min_shared_buffers; 2068 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2069 2070 rtransport = calloc(1, sizeof(*rtransport)); 2071 if (!rtransport) { 2072 return NULL; 2073 } 2074 2075 if (pthread_mutex_init(&rtransport->lock, NULL)) { 2076 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2077 free(rtransport); 2078 return NULL; 2079 } 2080 2081 TAILQ_INIT(&rtransport->devices); 2082 TAILQ_INIT(&rtransport->ports); 2083 2084 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2085 2086 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2087 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2088 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2089 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2090 " num_shared_buffers=%d, max_srq_depth=%d\n", 2091 opts->max_queue_depth, 2092 opts->max_io_size, 2093 opts->max_qpairs_per_ctrlr, 2094 opts->io_unit_size, 2095 opts->in_capsule_data_size, 2096 opts->max_aq_depth, 2097 opts->num_shared_buffers, 2098 opts->max_srq_depth); 2099 2100 /* I/O unit size cannot be larger than max I/O size */ 2101 if (opts->io_unit_size > opts->max_io_size) { 2102 opts->io_unit_size = opts->max_io_size; 2103 } 2104 2105 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2106 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2107 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2108 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2109 spdk_nvmf_rdma_destroy(&rtransport->transport); 2110 return NULL; 2111 } 2112 2113 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2114 if (min_shared_buffers > opts->num_shared_buffers) { 2115 SPDK_ERRLOG("There are not enough buffers to satisfy" 2116 "per-poll group caches for each thread. (%" PRIu32 ")" 2117 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2118 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2119 spdk_nvmf_rdma_destroy(&rtransport->transport); 2120 return NULL; 2121 } 2122 2123 sge_count = opts->max_io_size / opts->io_unit_size; 2124 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2125 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2126 spdk_nvmf_rdma_destroy(&rtransport->transport); 2127 return NULL; 2128 } 2129 2130 rtransport->event_channel = rdma_create_event_channel(); 2131 if (rtransport->event_channel == NULL) { 2132 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2133 spdk_nvmf_rdma_destroy(&rtransport->transport); 2134 return NULL; 2135 } 2136 2137 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2138 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2139 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2140 rtransport->event_channel->fd, spdk_strerror(errno)); 2141 spdk_nvmf_rdma_destroy(&rtransport->transport); 2142 return NULL; 2143 } 2144 2145 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2146 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2147 sizeof(struct spdk_nvmf_rdma_request_data), 2148 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2149 SPDK_ENV_SOCKET_ID_ANY); 2150 if (!rtransport->data_wr_pool) { 2151 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2152 spdk_nvmf_rdma_destroy(&rtransport->transport); 2153 return NULL; 2154 } 2155 2156 contexts = rdma_get_devices(NULL); 2157 if (contexts == NULL) { 2158 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2159 spdk_nvmf_rdma_destroy(&rtransport->transport); 2160 return NULL; 2161 } 2162 2163 i = 0; 2164 rc = 0; 2165 while (contexts[i] != NULL) { 2166 device = calloc(1, sizeof(*device)); 2167 if (!device) { 2168 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2169 rc = -ENOMEM; 2170 break; 2171 } 2172 device->context = contexts[i]; 2173 rc = ibv_query_device(device->context, &device->attr); 2174 if (rc < 0) { 2175 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2176 free(device); 2177 break; 2178 2179 } 2180 2181 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2182 2183 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2184 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2185 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2186 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2187 } 2188 2189 /** 2190 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2191 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2192 * but incorrectly reports that it does. There are changes making their way 2193 * through the kernel now that will enable this feature. When they are merged, 2194 * we can conditionally enable this feature. 2195 * 2196 * TODO: enable this for versions of the kernel rxe driver that support it. 2197 */ 2198 if (device->attr.vendor_id == 0) { 2199 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2200 } 2201 #endif 2202 2203 /* set up device context async ev fd as NON_BLOCKING */ 2204 flag = fcntl(device->context->async_fd, F_GETFL); 2205 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2206 if (rc < 0) { 2207 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2208 free(device); 2209 break; 2210 } 2211 2212 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2213 i++; 2214 2215 pd = NULL; 2216 if (g_nvmf_hooks.get_ibv_pd) { 2217 pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2218 } 2219 2220 if (!g_nvmf_hooks.get_ibv_pd) { 2221 device->pd = ibv_alloc_pd(device->context); 2222 if (!device->pd) { 2223 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2224 spdk_nvmf_rdma_destroy(&rtransport->transport); 2225 return NULL; 2226 } 2227 } else { 2228 device->pd = pd; 2229 } 2230 2231 assert(device->map == NULL); 2232 2233 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2234 if (!device->map) { 2235 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2236 spdk_nvmf_rdma_destroy(&rtransport->transport); 2237 return NULL; 2238 } 2239 2240 assert(device->map != NULL); 2241 assert(device->pd != NULL); 2242 } 2243 rdma_free_devices(contexts); 2244 2245 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2246 /* divide and round up. */ 2247 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2248 2249 /* round up to the nearest 4k. */ 2250 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2251 2252 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2253 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2254 opts->io_unit_size); 2255 } 2256 2257 if (rc < 0) { 2258 spdk_nvmf_rdma_destroy(&rtransport->transport); 2259 return NULL; 2260 } 2261 2262 /* Set up poll descriptor array to monitor events from RDMA and IB 2263 * in a single poll syscall 2264 */ 2265 rtransport->npoll_fds = i + 1; 2266 i = 0; 2267 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2268 if (rtransport->poll_fds == NULL) { 2269 SPDK_ERRLOG("poll_fds allocation failed\n"); 2270 spdk_nvmf_rdma_destroy(&rtransport->transport); 2271 return NULL; 2272 } 2273 2274 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2275 rtransport->poll_fds[i++].events = POLLIN; 2276 2277 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2278 rtransport->poll_fds[i].fd = device->context->async_fd; 2279 rtransport->poll_fds[i++].events = POLLIN; 2280 } 2281 2282 return &rtransport->transport; 2283 } 2284 2285 static int 2286 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2287 { 2288 struct spdk_nvmf_rdma_transport *rtransport; 2289 struct spdk_nvmf_rdma_port *port, *port_tmp; 2290 struct spdk_nvmf_rdma_device *device, *device_tmp; 2291 2292 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2293 2294 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2295 TAILQ_REMOVE(&rtransport->ports, port, link); 2296 rdma_destroy_id(port->id); 2297 free(port); 2298 } 2299 2300 if (rtransport->poll_fds != NULL) { 2301 free(rtransport->poll_fds); 2302 } 2303 2304 if (rtransport->event_channel != NULL) { 2305 rdma_destroy_event_channel(rtransport->event_channel); 2306 } 2307 2308 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2309 TAILQ_REMOVE(&rtransport->devices, device, link); 2310 if (device->map) { 2311 spdk_mem_map_free(&device->map); 2312 } 2313 if (device->pd) { 2314 if (!g_nvmf_hooks.get_ibv_pd) { 2315 ibv_dealloc_pd(device->pd); 2316 } 2317 } 2318 free(device); 2319 } 2320 2321 if (rtransport->data_wr_pool != NULL) { 2322 if (spdk_mempool_count(rtransport->data_wr_pool) != 2323 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2324 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2325 spdk_mempool_count(rtransport->data_wr_pool), 2326 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2327 } 2328 } 2329 2330 spdk_mempool_free(rtransport->data_wr_pool); 2331 pthread_mutex_destroy(&rtransport->lock); 2332 free(rtransport); 2333 2334 return 0; 2335 } 2336 2337 static int 2338 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2339 struct spdk_nvme_transport_id *trid, 2340 bool peer); 2341 2342 static int 2343 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2344 const struct spdk_nvme_transport_id *trid) 2345 { 2346 struct spdk_nvmf_rdma_transport *rtransport; 2347 struct spdk_nvmf_rdma_device *device; 2348 struct spdk_nvmf_rdma_port *port_tmp, *port; 2349 struct addrinfo *res; 2350 struct addrinfo hints; 2351 int family; 2352 int rc; 2353 2354 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2355 2356 port = calloc(1, sizeof(*port)); 2357 if (!port) { 2358 return -ENOMEM; 2359 } 2360 2361 /* Selectively copy the trid. Things like NQN don't matter here - that 2362 * mapping is enforced elsewhere. 2363 */ 2364 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2365 port->trid.adrfam = trid->adrfam; 2366 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 2367 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 2368 2369 pthread_mutex_lock(&rtransport->lock); 2370 assert(rtransport->event_channel != NULL); 2371 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 2372 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 2373 port_tmp->ref++; 2374 free(port); 2375 /* Already listening at this address */ 2376 pthread_mutex_unlock(&rtransport->lock); 2377 return 0; 2378 } 2379 } 2380 2381 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2382 if (rc < 0) { 2383 SPDK_ERRLOG("rdma_create_id() failed\n"); 2384 free(port); 2385 pthread_mutex_unlock(&rtransport->lock); 2386 return rc; 2387 } 2388 2389 switch (port->trid.adrfam) { 2390 case SPDK_NVMF_ADRFAM_IPV4: 2391 family = AF_INET; 2392 break; 2393 case SPDK_NVMF_ADRFAM_IPV6: 2394 family = AF_INET6; 2395 break; 2396 default: 2397 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 2398 free(port); 2399 pthread_mutex_unlock(&rtransport->lock); 2400 return -EINVAL; 2401 } 2402 2403 memset(&hints, 0, sizeof(hints)); 2404 hints.ai_family = family; 2405 hints.ai_flags = AI_NUMERICSERV; 2406 hints.ai_socktype = SOCK_STREAM; 2407 hints.ai_protocol = 0; 2408 2409 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 2410 if (rc) { 2411 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2412 free(port); 2413 pthread_mutex_unlock(&rtransport->lock); 2414 return -EINVAL; 2415 } 2416 2417 rc = rdma_bind_addr(port->id, res->ai_addr); 2418 freeaddrinfo(res); 2419 2420 if (rc < 0) { 2421 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2422 rdma_destroy_id(port->id); 2423 free(port); 2424 pthread_mutex_unlock(&rtransport->lock); 2425 return rc; 2426 } 2427 2428 if (!port->id->verbs) { 2429 SPDK_ERRLOG("ibv_context is null\n"); 2430 rdma_destroy_id(port->id); 2431 free(port); 2432 pthread_mutex_unlock(&rtransport->lock); 2433 return -1; 2434 } 2435 2436 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 2437 if (rc < 0) { 2438 SPDK_ERRLOG("rdma_listen() failed\n"); 2439 rdma_destroy_id(port->id); 2440 free(port); 2441 pthread_mutex_unlock(&rtransport->lock); 2442 return rc; 2443 } 2444 2445 TAILQ_FOREACH(device, &rtransport->devices, link) { 2446 if (device->context == port->id->verbs) { 2447 port->device = device; 2448 break; 2449 } 2450 } 2451 if (!port->device) { 2452 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2453 port->id->verbs); 2454 rdma_destroy_id(port->id); 2455 free(port); 2456 pthread_mutex_unlock(&rtransport->lock); 2457 return -EINVAL; 2458 } 2459 2460 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 2461 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 2462 2463 port->ref = 1; 2464 2465 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2466 pthread_mutex_unlock(&rtransport->lock); 2467 2468 return 0; 2469 } 2470 2471 static int 2472 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2473 const struct spdk_nvme_transport_id *_trid) 2474 { 2475 struct spdk_nvmf_rdma_transport *rtransport; 2476 struct spdk_nvmf_rdma_port *port, *tmp; 2477 struct spdk_nvme_transport_id trid = {}; 2478 2479 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2480 2481 /* Selectively copy the trid. Things like NQN don't matter here - that 2482 * mapping is enforced elsewhere. 2483 */ 2484 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2485 trid.adrfam = _trid->adrfam; 2486 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2487 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2488 2489 pthread_mutex_lock(&rtransport->lock); 2490 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2491 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2492 assert(port->ref > 0); 2493 port->ref--; 2494 if (port->ref == 0) { 2495 TAILQ_REMOVE(&rtransport->ports, port, link); 2496 rdma_destroy_id(port->id); 2497 free(port); 2498 } 2499 break; 2500 } 2501 } 2502 2503 pthread_mutex_unlock(&rtransport->lock); 2504 return 0; 2505 } 2506 2507 static void 2508 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2509 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2510 { 2511 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2512 struct spdk_nvmf_rdma_resources *resources; 2513 2514 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2515 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2516 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2517 break; 2518 } 2519 } 2520 2521 /* Then RDMA writes since reads have stronger restrictions than writes */ 2522 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2523 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2524 break; 2525 } 2526 } 2527 2528 /* The second highest priority is I/O waiting on memory buffers. */ 2529 TAILQ_FOREACH_SAFE(rdma_req, &rqpair->poller->group->pending_data_buf_queue, link, 2530 req_tmp) { 2531 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2532 break; 2533 } 2534 } 2535 2536 resources = rqpair->resources; 2537 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2538 rdma_req = STAILQ_FIRST(&resources->free_queue); 2539 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2540 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2541 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2542 2543 if (rqpair->srq != NULL) { 2544 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2545 rdma_req->recv->qpair->qd++; 2546 } else { 2547 rqpair->qd++; 2548 } 2549 2550 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2551 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2552 break; 2553 } 2554 } 2555 } 2556 2557 static void 2558 _nvmf_rdma_qpair_disconnect(void *ctx) 2559 { 2560 struct spdk_nvmf_qpair *qpair = ctx; 2561 2562 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2563 } 2564 2565 static void 2566 _nvmf_rdma_try_disconnect(void *ctx) 2567 { 2568 struct spdk_nvmf_qpair *qpair = ctx; 2569 struct spdk_nvmf_poll_group *group; 2570 2571 /* Read the group out of the qpair. This is normally set and accessed only from 2572 * the thread that created the group. Here, we're not on that thread necessarily. 2573 * The data member qpair->group begins it's life as NULL and then is assigned to 2574 * a pointer and never changes. So fortunately reading this and checking for 2575 * non-NULL is thread safe in the x86_64 memory model. */ 2576 group = qpair->group; 2577 2578 if (group == NULL) { 2579 /* The qpair hasn't been assigned to a group yet, so we can't 2580 * process a disconnect. Send a message to ourself and try again. */ 2581 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2582 return; 2583 } 2584 2585 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2586 } 2587 2588 static inline void 2589 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2590 { 2591 if (__sync_bool_compare_and_swap(&rqpair->disconnect_started, false, true)) { 2592 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2593 } 2594 } 2595 2596 static void nvmf_rdma_destroy_drained_qpair(void *ctx) 2597 { 2598 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2599 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2600 struct spdk_nvmf_rdma_transport, transport); 2601 2602 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2603 if (rqpair->current_send_depth != 0) { 2604 return; 2605 } 2606 2607 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2608 return; 2609 } 2610 2611 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) { 2612 return; 2613 } 2614 2615 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2616 spdk_nvmf_rdma_qpair_destroy(rqpair); 2617 } 2618 2619 2620 static int 2621 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2622 { 2623 struct spdk_nvmf_qpair *qpair; 2624 struct spdk_nvmf_rdma_qpair *rqpair; 2625 2626 if (evt->id == NULL) { 2627 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2628 return -1; 2629 } 2630 2631 qpair = evt->id->context; 2632 if (qpair == NULL) { 2633 SPDK_ERRLOG("disconnect request: no active connection\n"); 2634 return -1; 2635 } 2636 2637 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2638 2639 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2640 2641 spdk_nvmf_rdma_update_ibv_state(rqpair); 2642 2643 spdk_nvmf_rdma_start_disconnect(rqpair); 2644 2645 return 0; 2646 } 2647 2648 #ifdef DEBUG 2649 static const char *CM_EVENT_STR[] = { 2650 "RDMA_CM_EVENT_ADDR_RESOLVED", 2651 "RDMA_CM_EVENT_ADDR_ERROR", 2652 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2653 "RDMA_CM_EVENT_ROUTE_ERROR", 2654 "RDMA_CM_EVENT_CONNECT_REQUEST", 2655 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2656 "RDMA_CM_EVENT_CONNECT_ERROR", 2657 "RDMA_CM_EVENT_UNREACHABLE", 2658 "RDMA_CM_EVENT_REJECTED", 2659 "RDMA_CM_EVENT_ESTABLISHED", 2660 "RDMA_CM_EVENT_DISCONNECTED", 2661 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2662 "RDMA_CM_EVENT_MULTICAST_JOIN", 2663 "RDMA_CM_EVENT_MULTICAST_ERROR", 2664 "RDMA_CM_EVENT_ADDR_CHANGE", 2665 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2666 }; 2667 #endif /* DEBUG */ 2668 2669 static void 2670 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2671 { 2672 struct spdk_nvmf_rdma_transport *rtransport; 2673 struct rdma_cm_event *event; 2674 int rc; 2675 2676 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2677 2678 if (rtransport->event_channel == NULL) { 2679 return; 2680 } 2681 2682 while (1) { 2683 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2684 if (rc == 0) { 2685 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2686 2687 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2688 2689 switch (event->event) { 2690 case RDMA_CM_EVENT_ADDR_RESOLVED: 2691 case RDMA_CM_EVENT_ADDR_ERROR: 2692 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2693 case RDMA_CM_EVENT_ROUTE_ERROR: 2694 /* No action required. The target never attempts to resolve routes. */ 2695 break; 2696 case RDMA_CM_EVENT_CONNECT_REQUEST: 2697 rc = nvmf_rdma_connect(transport, event, cb_fn); 2698 if (rc < 0) { 2699 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2700 break; 2701 } 2702 break; 2703 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2704 /* The target never initiates a new connection. So this will not occur. */ 2705 break; 2706 case RDMA_CM_EVENT_CONNECT_ERROR: 2707 /* Can this happen? The docs say it can, but not sure what causes it. */ 2708 break; 2709 case RDMA_CM_EVENT_UNREACHABLE: 2710 case RDMA_CM_EVENT_REJECTED: 2711 /* These only occur on the client side. */ 2712 break; 2713 case RDMA_CM_EVENT_ESTABLISHED: 2714 /* TODO: Should we be waiting for this event anywhere? */ 2715 break; 2716 case RDMA_CM_EVENT_DISCONNECTED: 2717 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2718 rc = nvmf_rdma_disconnect(event); 2719 if (rc < 0) { 2720 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2721 break; 2722 } 2723 break; 2724 case RDMA_CM_EVENT_MULTICAST_JOIN: 2725 case RDMA_CM_EVENT_MULTICAST_ERROR: 2726 /* Multicast is not used */ 2727 break; 2728 case RDMA_CM_EVENT_ADDR_CHANGE: 2729 /* Not utilizing this event */ 2730 break; 2731 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2732 /* For now, do nothing. The target never re-uses queue pairs. */ 2733 break; 2734 default: 2735 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2736 break; 2737 } 2738 2739 rdma_ack_cm_event(event); 2740 } else { 2741 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2742 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2743 } 2744 break; 2745 } 2746 } 2747 } 2748 2749 static void 2750 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2751 { 2752 int rc; 2753 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 2754 struct ibv_async_event event; 2755 enum ibv_qp_state state; 2756 2757 rc = ibv_get_async_event(device->context, &event); 2758 2759 if (rc) { 2760 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2761 errno, spdk_strerror(errno)); 2762 return; 2763 } 2764 2765 SPDK_NOTICELOG("Async event: %s\n", 2766 ibv_event_type_str(event.event_type)); 2767 2768 switch (event.event_type) { 2769 case IBV_EVENT_QP_FATAL: 2770 rqpair = event.element.qp->qp_context; 2771 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2772 (uintptr_t)rqpair->cm_id, event.event_type); 2773 spdk_nvmf_rdma_update_ibv_state(rqpair); 2774 spdk_nvmf_rdma_start_disconnect(rqpair); 2775 break; 2776 case IBV_EVENT_QP_LAST_WQE_REACHED: 2777 /* This event only occurs for shared receive queues. */ 2778 rqpair = event.element.qp->qp_context; 2779 rqpair->last_wqe_reached = true; 2780 2781 /* This must be handled on the polling thread if it exists. Otherwise the timeout will catch it. */ 2782 if (rqpair->qpair.group) { 2783 spdk_thread_send_msg(rqpair->qpair.group->thread, nvmf_rdma_destroy_drained_qpair, rqpair); 2784 } else { 2785 SPDK_ERRLOG("Unable to destroy the qpair %p since it does not have a poll group.\n", rqpair); 2786 } 2787 2788 break; 2789 case IBV_EVENT_SQ_DRAINED: 2790 /* This event occurs frequently in both error and non-error states. 2791 * Check if the qpair is in an error state before sending a message. 2792 * Note that we're not on the correct thread to access the qpair, but 2793 * the operations that the below calls make all happen to be thread 2794 * safe. */ 2795 rqpair = event.element.qp->qp_context; 2796 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2797 (uintptr_t)rqpair->cm_id, event.event_type); 2798 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2799 if (state == IBV_QPS_ERR) { 2800 spdk_nvmf_rdma_start_disconnect(rqpair); 2801 } 2802 break; 2803 case IBV_EVENT_QP_REQ_ERR: 2804 case IBV_EVENT_QP_ACCESS_ERR: 2805 case IBV_EVENT_COMM_EST: 2806 case IBV_EVENT_PATH_MIG: 2807 case IBV_EVENT_PATH_MIG_ERR: 2808 rqpair = event.element.qp->qp_context; 2809 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2810 (uintptr_t)rqpair->cm_id, event.event_type); 2811 spdk_nvmf_rdma_update_ibv_state(rqpair); 2812 break; 2813 case IBV_EVENT_CQ_ERR: 2814 case IBV_EVENT_DEVICE_FATAL: 2815 case IBV_EVENT_PORT_ACTIVE: 2816 case IBV_EVENT_PORT_ERR: 2817 case IBV_EVENT_LID_CHANGE: 2818 case IBV_EVENT_PKEY_CHANGE: 2819 case IBV_EVENT_SM_CHANGE: 2820 case IBV_EVENT_SRQ_ERR: 2821 case IBV_EVENT_SRQ_LIMIT_REACHED: 2822 case IBV_EVENT_CLIENT_REREGISTER: 2823 case IBV_EVENT_GID_CHANGE: 2824 default: 2825 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2826 break; 2827 } 2828 ibv_ack_async_event(&event); 2829 } 2830 2831 static void 2832 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2833 { 2834 int nfds, i = 0; 2835 struct spdk_nvmf_rdma_transport *rtransport; 2836 struct spdk_nvmf_rdma_device *device, *tmp; 2837 2838 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2839 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2840 2841 if (nfds <= 0) { 2842 return; 2843 } 2844 2845 /* The first poll descriptor is RDMA CM event */ 2846 if (rtransport->poll_fds[i++].revents & POLLIN) { 2847 spdk_nvmf_process_cm_event(transport, cb_fn); 2848 nfds--; 2849 } 2850 2851 if (nfds == 0) { 2852 return; 2853 } 2854 2855 /* Second and subsequent poll descriptors are IB async events */ 2856 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2857 if (rtransport->poll_fds[i++].revents & POLLIN) { 2858 spdk_nvmf_process_ib_event(device); 2859 nfds--; 2860 } 2861 } 2862 /* check all flagged fd's have been served */ 2863 assert(nfds == 0); 2864 } 2865 2866 static void 2867 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2868 struct spdk_nvme_transport_id *trid, 2869 struct spdk_nvmf_discovery_log_page_entry *entry) 2870 { 2871 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2872 entry->adrfam = trid->adrfam; 2873 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2874 2875 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2876 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2877 2878 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2879 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2880 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2881 } 2882 2883 static void 2884 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 2885 2886 static struct spdk_nvmf_transport_poll_group * 2887 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2888 { 2889 struct spdk_nvmf_rdma_transport *rtransport; 2890 struct spdk_nvmf_rdma_poll_group *rgroup; 2891 struct spdk_nvmf_rdma_poller *poller; 2892 struct spdk_nvmf_rdma_device *device; 2893 struct ibv_srq_init_attr srq_init_attr; 2894 struct spdk_nvmf_rdma_resource_opts opts; 2895 int num_cqe; 2896 2897 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2898 2899 rgroup = calloc(1, sizeof(*rgroup)); 2900 if (!rgroup) { 2901 return NULL; 2902 } 2903 2904 TAILQ_INIT(&rgroup->pollers); 2905 TAILQ_INIT(&rgroup->pending_data_buf_queue); 2906 2907 pthread_mutex_lock(&rtransport->lock); 2908 TAILQ_FOREACH(device, &rtransport->devices, link) { 2909 poller = calloc(1, sizeof(*poller)); 2910 if (!poller) { 2911 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2912 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2913 pthread_mutex_unlock(&rtransport->lock); 2914 return NULL; 2915 } 2916 2917 poller->device = device; 2918 poller->group = rgroup; 2919 2920 TAILQ_INIT(&poller->qpairs); 2921 2922 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2923 if (device->attr.max_srq != 0) { 2924 poller->max_srq_depth = transport->opts.max_srq_depth; 2925 2926 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 2927 srq_init_attr.attr.max_wr = poller->max_srq_depth; 2928 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 2929 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 2930 if (!poller->srq) { 2931 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 2932 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2933 pthread_mutex_unlock(&rtransport->lock); 2934 return NULL; 2935 } 2936 2937 opts.qp = poller->srq; 2938 opts.pd = device->pd; 2939 opts.qpair = NULL; 2940 opts.shared = true; 2941 opts.max_queue_depth = poller->max_srq_depth; 2942 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 2943 2944 poller->resources = nvmf_rdma_resources_create(&opts); 2945 if (!poller->resources) { 2946 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 2947 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2948 pthread_mutex_unlock(&rtransport->lock); 2949 } 2950 } 2951 2952 /* 2953 * When using an srq, we can limit the completion queue at startup. 2954 * The following formula represents the calculation: 2955 * num_cqe = num_recv + num_data_wr + num_send_wr. 2956 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 2957 */ 2958 if (poller->srq) { 2959 num_cqe = poller->max_srq_depth * 3; 2960 } else { 2961 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2962 } 2963 2964 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 2965 if (!poller->cq) { 2966 SPDK_ERRLOG("Unable to create completion queue\n"); 2967 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2968 pthread_mutex_unlock(&rtransport->lock); 2969 return NULL; 2970 } 2971 poller->num_cqe = num_cqe; 2972 } 2973 2974 pthread_mutex_unlock(&rtransport->lock); 2975 return &rgroup->group; 2976 } 2977 2978 static void 2979 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2980 { 2981 struct spdk_nvmf_rdma_poll_group *rgroup; 2982 struct spdk_nvmf_rdma_poller *poller, *tmp; 2983 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2984 2985 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2986 2987 if (!rgroup) { 2988 return; 2989 } 2990 2991 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2992 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2993 2994 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 2995 spdk_nvmf_rdma_qpair_destroy(qpair); 2996 } 2997 2998 if (poller->srq) { 2999 nvmf_rdma_resources_destroy(poller->resources); 3000 ibv_destroy_srq(poller->srq); 3001 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3002 } 3003 3004 if (poller->cq) { 3005 ibv_destroy_cq(poller->cq); 3006 } 3007 3008 free(poller); 3009 } 3010 3011 if (!TAILQ_EMPTY(&rgroup->pending_data_buf_queue)) { 3012 SPDK_ERRLOG("Pending I/O list wasn't empty on poll group destruction\n"); 3013 } 3014 3015 free(rgroup); 3016 } 3017 3018 static void 3019 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3020 { 3021 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3022 spdk_nvmf_rdma_qpair_destroy(rqpair); 3023 } 3024 3025 static int 3026 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3027 struct spdk_nvmf_qpair *qpair) 3028 { 3029 struct spdk_nvmf_rdma_poll_group *rgroup; 3030 struct spdk_nvmf_rdma_qpair *rqpair; 3031 struct spdk_nvmf_rdma_device *device; 3032 struct spdk_nvmf_rdma_poller *poller; 3033 int rc; 3034 3035 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3036 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3037 3038 device = rqpair->port->device; 3039 3040 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3041 if (poller->device == device) { 3042 break; 3043 } 3044 } 3045 3046 if (!poller) { 3047 SPDK_ERRLOG("No poller found for device.\n"); 3048 return -1; 3049 } 3050 3051 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3052 rqpair->poller = poller; 3053 rqpair->srq = rqpair->poller->srq; 3054 3055 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 3056 if (rc < 0) { 3057 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3058 return -1; 3059 } 3060 3061 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3062 if (rc) { 3063 /* Try to reject, but we probably can't */ 3064 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3065 return -1; 3066 } 3067 3068 spdk_nvmf_rdma_update_ibv_state(rqpair); 3069 3070 return 0; 3071 } 3072 3073 static int 3074 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3075 { 3076 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3077 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3078 struct spdk_nvmf_rdma_transport, transport); 3079 3080 nvmf_rdma_request_free(rdma_req, rtransport); 3081 return 0; 3082 } 3083 3084 static int 3085 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3086 { 3087 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3088 struct spdk_nvmf_rdma_transport, transport); 3089 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3090 struct spdk_nvmf_rdma_request, req); 3091 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3092 struct spdk_nvmf_rdma_qpair, qpair); 3093 3094 if (rqpair->ibv_state != IBV_QPS_ERR) { 3095 /* The connection is alive, so process the request as normal */ 3096 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3097 } else { 3098 /* The connection is dead. Move the request directly to the completed state. */ 3099 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3100 } 3101 3102 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3103 3104 return 0; 3105 } 3106 3107 static int 3108 spdk_nvmf_rdma_destroy_defunct_qpair(void *ctx) 3109 { 3110 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 3111 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3112 struct spdk_nvmf_rdma_transport, transport); 3113 3114 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3115 spdk_nvmf_rdma_qpair_destroy(rqpair); 3116 3117 return 0; 3118 } 3119 3120 static void 3121 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3122 { 3123 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3124 3125 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 3126 return; 3127 } 3128 3129 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 3130 3131 /* This happens only when the qpair is disconnected before 3132 * it is added to the poll group. Since there is no poll group, 3133 * the RDMA qp has not been initialized yet and the RDMA CM 3134 * event has not yet been acknowledged, so we need to reject it. 3135 */ 3136 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3137 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3138 return; 3139 } 3140 3141 if (rqpair->ibv_state != IBV_QPS_ERR) { 3142 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 3143 } 3144 3145 rqpair->destruct_poller = spdk_poller_register(spdk_nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 3146 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 3147 } 3148 3149 static struct spdk_nvmf_rdma_qpair * 3150 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3151 { 3152 struct spdk_nvmf_rdma_qpair *rqpair; 3153 /* @todo: improve QP search */ 3154 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3155 if (wc->qp_num == rqpair->cm_id->qp->qp_num) { 3156 return rqpair; 3157 } 3158 } 3159 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3160 return NULL; 3161 } 3162 3163 #ifdef DEBUG 3164 static int 3165 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3166 { 3167 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3168 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3169 } 3170 #endif 3171 3172 static int 3173 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3174 struct spdk_nvmf_rdma_poller *rpoller) 3175 { 3176 struct ibv_wc wc[32]; 3177 struct spdk_nvmf_rdma_wr *rdma_wr; 3178 struct spdk_nvmf_rdma_request *rdma_req; 3179 struct spdk_nvmf_rdma_recv *rdma_recv; 3180 struct spdk_nvmf_rdma_qpair *rqpair; 3181 int reaped, i; 3182 int count = 0; 3183 bool error = false; 3184 3185 /* Poll for completing operations. */ 3186 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3187 if (reaped < 0) { 3188 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3189 errno, spdk_strerror(errno)); 3190 return -1; 3191 } 3192 3193 for (i = 0; i < reaped; i++) { 3194 3195 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3196 3197 switch (rdma_wr->type) { 3198 case RDMA_WR_TYPE_SEND: 3199 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3200 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3201 3202 if (!wc[i].status) { 3203 count++; 3204 assert(wc[i].opcode == IBV_WC_SEND); 3205 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 3206 } else { 3207 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3208 } 3209 3210 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3211 rqpair->current_send_depth--; 3212 3213 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3214 assert(rdma_req->num_outstanding_data_wr == 0); 3215 break; 3216 case RDMA_WR_TYPE_RECV: 3217 /* rdma_recv->qpair will be NULL if using an SRQ. In that case we have to get the qpair from the wc. */ 3218 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3219 if (rdma_recv->qpair == NULL) { 3220 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3221 } 3222 rqpair = rdma_recv->qpair; 3223 3224 assert(rqpair != NULL); 3225 if (!wc[i].status) { 3226 assert(wc[i].opcode == IBV_WC_RECV); 3227 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3228 spdk_nvmf_rdma_start_disconnect(rqpair); 3229 break; 3230 } 3231 } 3232 3233 rqpair->current_recv_depth++; 3234 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3235 break; 3236 case RDMA_WR_TYPE_DATA: 3237 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3238 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3239 3240 assert(rdma_req->num_outstanding_data_wr > 0); 3241 3242 rqpair->current_send_depth--; 3243 rdma_req->num_outstanding_data_wr--; 3244 if (!wc[i].status) { 3245 if (wc[i].opcode == IBV_WC_RDMA_READ) { 3246 rqpair->current_read_depth--; 3247 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3248 if (rdma_req->num_outstanding_data_wr == 0) { 3249 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3250 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3251 } 3252 } else { 3253 assert(wc[i].opcode == IBV_WC_RDMA_WRITE); 3254 } 3255 } else { 3256 /* If the data transfer fails still force the queue into the error state, 3257 * if we were performing an RDMA_READ, we need to force the request into a 3258 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3259 * case, we should wait for the SEND to complete. */ 3260 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3261 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3262 rqpair->current_read_depth--; 3263 if (rdma_req->num_outstanding_data_wr == 0) { 3264 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3265 } 3266 } 3267 } 3268 break; 3269 default: 3270 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3271 continue; 3272 } 3273 3274 /* Handle error conditions */ 3275 if (wc[i].status) { 3276 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 3277 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 3278 3279 error = true; 3280 3281 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3282 /* Disconnect the connection. */ 3283 spdk_nvmf_rdma_start_disconnect(rqpair); 3284 } else { 3285 nvmf_rdma_destroy_drained_qpair(rqpair); 3286 } 3287 continue; 3288 } 3289 3290 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3291 3292 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3293 nvmf_rdma_destroy_drained_qpair(rqpair); 3294 } 3295 } 3296 3297 if (error == true) { 3298 return -1; 3299 } 3300 3301 return count; 3302 } 3303 3304 static int 3305 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3306 { 3307 struct spdk_nvmf_rdma_transport *rtransport; 3308 struct spdk_nvmf_rdma_poll_group *rgroup; 3309 struct spdk_nvmf_rdma_poller *rpoller; 3310 int count, rc; 3311 3312 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3313 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3314 3315 count = 0; 3316 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3317 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 3318 if (rc < 0) { 3319 return rc; 3320 } 3321 count += rc; 3322 } 3323 3324 return count; 3325 } 3326 3327 static int 3328 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3329 struct spdk_nvme_transport_id *trid, 3330 bool peer) 3331 { 3332 struct sockaddr *saddr; 3333 uint16_t port; 3334 3335 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 3336 3337 if (peer) { 3338 saddr = rdma_get_peer_addr(id); 3339 } else { 3340 saddr = rdma_get_local_addr(id); 3341 } 3342 switch (saddr->sa_family) { 3343 case AF_INET: { 3344 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3345 3346 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3347 inet_ntop(AF_INET, &saddr_in->sin_addr, 3348 trid->traddr, sizeof(trid->traddr)); 3349 if (peer) { 3350 port = ntohs(rdma_get_dst_port(id)); 3351 } else { 3352 port = ntohs(rdma_get_src_port(id)); 3353 } 3354 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3355 break; 3356 } 3357 case AF_INET6: { 3358 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3359 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3360 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3361 trid->traddr, sizeof(trid->traddr)); 3362 if (peer) { 3363 port = ntohs(rdma_get_dst_port(id)); 3364 } else { 3365 port = ntohs(rdma_get_src_port(id)); 3366 } 3367 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3368 break; 3369 } 3370 default: 3371 return -1; 3372 3373 } 3374 3375 return 0; 3376 } 3377 3378 static int 3379 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3380 struct spdk_nvme_transport_id *trid) 3381 { 3382 struct spdk_nvmf_rdma_qpair *rqpair; 3383 3384 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3385 3386 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 3387 } 3388 3389 static int 3390 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3391 struct spdk_nvme_transport_id *trid) 3392 { 3393 struct spdk_nvmf_rdma_qpair *rqpair; 3394 3395 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3396 3397 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 3398 } 3399 3400 static int 3401 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3402 struct spdk_nvme_transport_id *trid) 3403 { 3404 struct spdk_nvmf_rdma_qpair *rqpair; 3405 3406 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3407 3408 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 3409 } 3410 3411 void 3412 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3413 { 3414 g_nvmf_hooks = *hooks; 3415 } 3416 3417 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 3418 .type = SPDK_NVME_TRANSPORT_RDMA, 3419 .opts_init = spdk_nvmf_rdma_opts_init, 3420 .create = spdk_nvmf_rdma_create, 3421 .destroy = spdk_nvmf_rdma_destroy, 3422 3423 .listen = spdk_nvmf_rdma_listen, 3424 .stop_listen = spdk_nvmf_rdma_stop_listen, 3425 .accept = spdk_nvmf_rdma_accept, 3426 3427 .listener_discover = spdk_nvmf_rdma_discover, 3428 3429 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 3430 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 3431 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 3432 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 3433 3434 .req_free = spdk_nvmf_rdma_request_free, 3435 .req_complete = spdk_nvmf_rdma_request_complete, 3436 3437 .qpair_fini = spdk_nvmf_rdma_close_qpair, 3438 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 3439 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 3440 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 3441 3442 }; 3443 3444 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 3445