1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2018 Mellanox Technologies LTD. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of Intel Corporation nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include "spdk/stdinc.h" 35 36 #include <infiniband/verbs.h> 37 #include <rdma/rdma_cma.h> 38 #include <rdma/rdma_verbs.h> 39 40 #include "nvmf_internal.h" 41 #include "transport.h" 42 43 #include "spdk/config.h" 44 #include "spdk/assert.h" 45 #include "spdk/thread.h" 46 #include "spdk/nvmf.h" 47 #include "spdk/nvmf_spec.h" 48 #include "spdk/string.h" 49 #include "spdk/trace.h" 50 #include "spdk/util.h" 51 52 #include "spdk_internal/log.h" 53 54 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 55 56 /* 57 RDMA Connection Resource Defaults 58 */ 59 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 60 #define NVMF_DEFAULT_RSP_SGE 1 61 #define NVMF_DEFAULT_RX_SGE 2 62 63 /* The RDMA completion queue size */ 64 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 65 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 66 67 /* Timeout for destroying defunct rqpairs */ 68 #define NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US 4000000 69 70 /* The maximum number of buffers per request */ 71 #define NVMF_REQ_MAX_BUFFERS (SPDK_NVMF_MAX_SGL_ENTRIES * 2) 72 73 static int g_spdk_nvmf_ibv_query_mask = 74 IBV_QP_STATE | 75 IBV_QP_PKEY_INDEX | 76 IBV_QP_PORT | 77 IBV_QP_ACCESS_FLAGS | 78 IBV_QP_AV | 79 IBV_QP_PATH_MTU | 80 IBV_QP_DEST_QPN | 81 IBV_QP_RQ_PSN | 82 IBV_QP_MAX_DEST_RD_ATOMIC | 83 IBV_QP_MIN_RNR_TIMER | 84 IBV_QP_SQ_PSN | 85 IBV_QP_TIMEOUT | 86 IBV_QP_RETRY_CNT | 87 IBV_QP_RNR_RETRY | 88 IBV_QP_MAX_QP_RD_ATOMIC; 89 90 enum spdk_nvmf_rdma_request_state { 91 /* The request is not currently in use */ 92 RDMA_REQUEST_STATE_FREE = 0, 93 94 /* Initial state when request first received */ 95 RDMA_REQUEST_STATE_NEW, 96 97 /* The request is queued until a data buffer is available. */ 98 RDMA_REQUEST_STATE_NEED_BUFFER, 99 100 /* The request is waiting on RDMA queue depth availability 101 * to transfer data from the host to the controller. 102 */ 103 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 104 105 /* The request is currently transferring data from the host to the controller. */ 106 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 107 108 /* The request is ready to execute at the block device */ 109 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 110 111 /* The request is currently executing at the block device */ 112 RDMA_REQUEST_STATE_EXECUTING, 113 114 /* The request finished executing at the block device */ 115 RDMA_REQUEST_STATE_EXECUTED, 116 117 /* The request is waiting on RDMA queue depth availability 118 * to transfer data from the controller to the host. 119 */ 120 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 121 122 /* The request is ready to send a completion */ 123 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 124 125 /* The request is currently transferring data from the controller to the host. */ 126 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 127 128 /* The request currently has an outstanding completion without an 129 * associated data transfer. 130 */ 131 RDMA_REQUEST_STATE_COMPLETING, 132 133 /* The request completed and can be marked free. */ 134 RDMA_REQUEST_STATE_COMPLETED, 135 136 /* Terminator */ 137 RDMA_REQUEST_NUM_STATES, 138 }; 139 140 #define OBJECT_NVMF_RDMA_IO 0x40 141 142 #define TRACE_GROUP_NVMF_RDMA 0x4 143 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 144 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 145 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 146 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 147 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 148 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 149 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 150 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 151 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 152 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 153 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 154 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 155 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 156 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 157 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 158 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 159 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 160 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 161 162 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 163 { 164 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 165 spdk_trace_register_description("RDMA_REQ_NEW", "", 166 TRACE_RDMA_REQUEST_STATE_NEW, 167 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 168 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", "", 169 TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 170 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 171 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C_TO_H", "", 172 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 173 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 174 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H_TO_C", "", 175 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 176 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 177 spdk_trace_register_description("RDMA_REQ_TX_H_TO_C", "", 178 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 179 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 180 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", "", 181 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 182 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 183 spdk_trace_register_description("RDMA_REQ_EXECUTING", "", 184 TRACE_RDMA_REQUEST_STATE_EXECUTING, 185 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 186 spdk_trace_register_description("RDMA_REQ_EXECUTED", "", 187 TRACE_RDMA_REQUEST_STATE_EXECUTED, 188 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 189 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPLETE", "", 190 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 191 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 192 spdk_trace_register_description("RDMA_REQ_COMPLETING_CONTROLLER_TO_HOST", "", 193 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 194 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 195 spdk_trace_register_description("RDMA_REQ_COMPLETING_INCAPSULE", "", 196 TRACE_RDMA_REQUEST_STATE_COMPLETING, 197 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 198 spdk_trace_register_description("RDMA_REQ_COMPLETED", "", 199 TRACE_RDMA_REQUEST_STATE_COMPLETED, 200 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 201 202 spdk_trace_register_description("RDMA_QP_CREATE", "", TRACE_RDMA_QP_CREATE, 203 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 204 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", "", TRACE_RDMA_IBV_ASYNC_EVENT, 205 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 206 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", "", TRACE_RDMA_CM_ASYNC_EVENT, 207 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 208 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", "", TRACE_RDMA_QP_STATE_CHANGE, 209 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 210 spdk_trace_register_description("RDMA_QP_DISCONNECT", "", TRACE_RDMA_QP_DISCONNECT, 211 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 212 spdk_trace_register_description("RDMA_QP_DESTROY", "", TRACE_RDMA_QP_DESTROY, 213 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 214 } 215 216 enum spdk_nvmf_rdma_wr_type { 217 RDMA_WR_TYPE_RECV, 218 RDMA_WR_TYPE_SEND, 219 RDMA_WR_TYPE_DATA, 220 }; 221 222 struct spdk_nvmf_rdma_wr { 223 enum spdk_nvmf_rdma_wr_type type; 224 }; 225 226 /* This structure holds commands as they are received off the wire. 227 * It must be dynamically paired with a full request object 228 * (spdk_nvmf_rdma_request) to service a request. It is separate 229 * from the request because RDMA does not appear to order 230 * completions, so occasionally we'll get a new incoming 231 * command when there aren't any free request objects. 232 */ 233 struct spdk_nvmf_rdma_recv { 234 struct ibv_recv_wr wr; 235 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 236 237 struct spdk_nvmf_rdma_qpair *qpair; 238 239 /* In-capsule data buffer */ 240 uint8_t *buf; 241 242 struct spdk_nvmf_rdma_wr rdma_wr; 243 244 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 245 }; 246 247 struct spdk_nvmf_rdma_request_data { 248 struct spdk_nvmf_rdma_wr rdma_wr; 249 struct ibv_send_wr wr; 250 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 251 }; 252 253 struct spdk_nvmf_rdma_request { 254 struct spdk_nvmf_request req; 255 bool data_from_pool; 256 257 enum spdk_nvmf_rdma_request_state state; 258 259 struct spdk_nvmf_rdma_recv *recv; 260 261 struct { 262 struct spdk_nvmf_rdma_wr rdma_wr; 263 struct ibv_send_wr wr; 264 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 265 } rsp; 266 267 struct spdk_nvmf_rdma_request_data data; 268 void *buffers[NVMF_REQ_MAX_BUFFERS]; 269 270 uint32_t num_outstanding_data_wr; 271 272 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 273 }; 274 275 enum spdk_nvmf_rdma_qpair_disconnect_flags { 276 RDMA_QP_DISCONNECTING = 1, 277 RDMA_QP_RECV_DRAINED = 1 << 1, 278 RDMA_QP_SEND_DRAINED = 1 << 2 279 }; 280 281 struct spdk_nvmf_rdma_resource_opts { 282 struct spdk_nvmf_rdma_qpair *qpair; 283 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 284 void *qp; 285 struct ibv_pd *pd; 286 uint32_t max_queue_depth; 287 uint32_t in_capsule_data_size; 288 bool shared; 289 }; 290 291 struct spdk_nvmf_rdma_resources { 292 /* Array of size "max_queue_depth" containing RDMA requests. */ 293 struct spdk_nvmf_rdma_request *reqs; 294 295 /* Array of size "max_queue_depth" containing RDMA recvs. */ 296 struct spdk_nvmf_rdma_recv *recvs; 297 298 /* Array of size "max_queue_depth" containing 64 byte capsules 299 * used for receive. 300 */ 301 union nvmf_h2c_msg *cmds; 302 struct ibv_mr *cmds_mr; 303 304 /* Array of size "max_queue_depth" containing 16 byte completions 305 * to be sent back to the user. 306 */ 307 union nvmf_c2h_msg *cpls; 308 struct ibv_mr *cpls_mr; 309 310 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 311 * buffers to be used for in capsule data. 312 */ 313 void *bufs; 314 struct ibv_mr *bufs_mr; 315 316 /* Receives that are waiting for a request object */ 317 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 318 319 /* Queue to track free requests */ 320 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 321 }; 322 323 struct spdk_nvmf_rdma_qpair { 324 struct spdk_nvmf_qpair qpair; 325 326 struct spdk_nvmf_rdma_port *port; 327 struct spdk_nvmf_rdma_poller *poller; 328 329 struct rdma_cm_id *cm_id; 330 struct ibv_srq *srq; 331 struct rdma_cm_id *listen_id; 332 333 /* The maximum number of I/O outstanding on this connection at one time */ 334 uint16_t max_queue_depth; 335 336 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 337 uint16_t max_read_depth; 338 339 /* The maximum number of RDMA SEND operations at one time */ 340 uint32_t max_send_depth; 341 342 /* The current number of outstanding WRs from this qpair's 343 * recv queue. Should not exceed device->attr.max_queue_depth. 344 */ 345 uint16_t current_recv_depth; 346 347 /* The current number of active RDMA READ operations */ 348 uint16_t current_read_depth; 349 350 /* The current number of posted WRs from this qpair's 351 * send queue. Should not exceed max_send_depth. 352 */ 353 uint32_t current_send_depth; 354 355 /* The maximum number of SGEs per WR on the send queue */ 356 uint32_t max_send_sge; 357 358 /* The maximum number of SGEs per WR on the recv queue */ 359 uint32_t max_recv_sge; 360 361 struct spdk_nvmf_rdma_resources *resources; 362 363 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 364 365 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 366 367 /* Number of requests not in the free state */ 368 uint32_t qd; 369 370 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 371 372 /* IBV queue pair attributes: they are used to manage 373 * qp state and recover from errors. 374 */ 375 enum ibv_qp_state ibv_state; 376 377 uint32_t disconnect_flags; 378 379 /* Poller registered in case the qpair doesn't properly 380 * complete the qpair destruct process and becomes defunct. 381 */ 382 383 struct spdk_poller *destruct_poller; 384 385 /* There are several ways a disconnect can start on a qpair 386 * and they are not all mutually exclusive. It is important 387 * that we only initialize one of these paths. 388 */ 389 bool disconnect_started; 390 /* Lets us know that we have received the last_wqe event. */ 391 bool last_wqe_reached; 392 }; 393 394 struct spdk_nvmf_rdma_poller { 395 struct spdk_nvmf_rdma_device *device; 396 struct spdk_nvmf_rdma_poll_group *group; 397 398 int num_cqe; 399 int required_num_wr; 400 struct ibv_cq *cq; 401 402 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 403 uint16_t max_srq_depth; 404 405 /* Shared receive queue */ 406 struct ibv_srq *srq; 407 408 struct spdk_nvmf_rdma_resources *resources; 409 410 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 411 412 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 413 }; 414 415 struct spdk_nvmf_rdma_poll_group { 416 struct spdk_nvmf_transport_poll_group group; 417 418 /* Requests that are waiting to obtain a data buffer */ 419 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_data_buf_queue; 420 421 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 422 }; 423 424 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 425 struct spdk_nvmf_rdma_device { 426 struct ibv_device_attr attr; 427 struct ibv_context *context; 428 429 struct spdk_mem_map *map; 430 struct ibv_pd *pd; 431 432 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 433 }; 434 435 struct spdk_nvmf_rdma_port { 436 struct spdk_nvme_transport_id trid; 437 struct rdma_cm_id *id; 438 struct spdk_nvmf_rdma_device *device; 439 uint32_t ref; 440 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 441 }; 442 443 struct spdk_nvmf_rdma_transport { 444 struct spdk_nvmf_transport transport; 445 446 struct rdma_event_channel *event_channel; 447 448 struct spdk_mempool *data_wr_pool; 449 450 pthread_mutex_t lock; 451 452 /* fields used to poll RDMA/IB events */ 453 nfds_t npoll_fds; 454 struct pollfd *poll_fds; 455 456 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 457 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 458 }; 459 460 static inline int 461 spdk_nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 462 { 463 switch (state) { 464 case IBV_QPS_RESET: 465 case IBV_QPS_INIT: 466 case IBV_QPS_RTR: 467 case IBV_QPS_RTS: 468 case IBV_QPS_SQD: 469 case IBV_QPS_SQE: 470 case IBV_QPS_ERR: 471 return 0; 472 default: 473 return -1; 474 } 475 } 476 477 static enum ibv_qp_state 478 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 479 enum ibv_qp_state old_state, new_state; 480 struct ibv_qp_attr qp_attr; 481 struct ibv_qp_init_attr init_attr; 482 int rc; 483 484 old_state = rqpair->ibv_state; 485 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 486 g_spdk_nvmf_ibv_query_mask, &init_attr); 487 488 if (rc) 489 { 490 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 491 return IBV_QPS_ERR + 1; 492 } 493 494 new_state = qp_attr.qp_state; 495 rqpair->ibv_state = new_state; 496 qp_attr.ah_attr.port_num = qp_attr.port_num; 497 498 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 499 if (rc) 500 { 501 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 502 /* 503 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 504 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 505 */ 506 return IBV_QPS_ERR + 1; 507 } 508 509 if (old_state != new_state) 510 { 511 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 512 (uintptr_t)rqpair->cm_id, new_state); 513 } 514 return new_state; 515 } 516 517 static const char *str_ibv_qp_state[] = { 518 "IBV_QPS_RESET", 519 "IBV_QPS_INIT", 520 "IBV_QPS_RTR", 521 "IBV_QPS_RTS", 522 "IBV_QPS_SQD", 523 "IBV_QPS_SQE", 524 "IBV_QPS_ERR", 525 "IBV_QPS_UNKNOWN" 526 }; 527 528 static int 529 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 530 enum ibv_qp_state new_state) 531 { 532 struct ibv_qp_attr qp_attr; 533 struct ibv_qp_init_attr init_attr; 534 int rc; 535 enum ibv_qp_state state; 536 static int attr_mask_rc[] = { 537 [IBV_QPS_RESET] = IBV_QP_STATE, 538 [IBV_QPS_INIT] = (IBV_QP_STATE | 539 IBV_QP_PKEY_INDEX | 540 IBV_QP_PORT | 541 IBV_QP_ACCESS_FLAGS), 542 [IBV_QPS_RTR] = (IBV_QP_STATE | 543 IBV_QP_AV | 544 IBV_QP_PATH_MTU | 545 IBV_QP_DEST_QPN | 546 IBV_QP_RQ_PSN | 547 IBV_QP_MAX_DEST_RD_ATOMIC | 548 IBV_QP_MIN_RNR_TIMER), 549 [IBV_QPS_RTS] = (IBV_QP_STATE | 550 IBV_QP_SQ_PSN | 551 IBV_QP_TIMEOUT | 552 IBV_QP_RETRY_CNT | 553 IBV_QP_RNR_RETRY | 554 IBV_QP_MAX_QP_RD_ATOMIC), 555 [IBV_QPS_SQD] = IBV_QP_STATE, 556 [IBV_QPS_SQE] = IBV_QP_STATE, 557 [IBV_QPS_ERR] = IBV_QP_STATE, 558 }; 559 560 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 561 if (rc) { 562 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 563 rqpair->qpair.qid, new_state); 564 return rc; 565 } 566 567 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 568 g_spdk_nvmf_ibv_query_mask, &init_attr); 569 570 if (rc) { 571 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 572 assert(false); 573 } 574 575 qp_attr.cur_qp_state = rqpair->ibv_state; 576 qp_attr.qp_state = new_state; 577 578 rc = ibv_modify_qp(rqpair->cm_id->qp, &qp_attr, 579 attr_mask_rc[new_state]); 580 581 if (rc) { 582 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 583 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 584 return rc; 585 } 586 587 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 588 589 if (state != new_state) { 590 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 591 rqpair->qpair.qid, str_ibv_qp_state[new_state], 592 str_ibv_qp_state[state]); 593 return -1; 594 } 595 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 596 str_ibv_qp_state[state]); 597 return 0; 598 } 599 600 static void 601 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 602 struct spdk_nvmf_rdma_transport *rtransport) 603 { 604 struct spdk_nvmf_rdma_request_data *current_data_wr = NULL, *next_data_wr = NULL; 605 struct ibv_send_wr *send_wr; 606 int i; 607 608 rdma_req->num_outstanding_data_wr = 0; 609 current_data_wr = &rdma_req->data; 610 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 611 current_data_wr->wr.sg_list[i].addr = 0; 612 current_data_wr->wr.sg_list[i].length = 0; 613 current_data_wr->wr.sg_list[i].lkey = 0; 614 } 615 current_data_wr->wr.num_sge = 0; 616 617 send_wr = current_data_wr->wr.next; 618 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr) { 619 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 620 } 621 while (next_data_wr) { 622 current_data_wr = next_data_wr; 623 send_wr = current_data_wr->wr.next; 624 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr) { 625 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 626 } else { 627 next_data_wr = NULL; 628 } 629 630 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 631 current_data_wr->wr.sg_list[i].addr = 0; 632 current_data_wr->wr.sg_list[i].length = 0; 633 current_data_wr->wr.sg_list[i].lkey = 0; 634 } 635 current_data_wr->wr.num_sge = 0; 636 spdk_mempool_put(rtransport->data_wr_pool, current_data_wr); 637 } 638 } 639 640 static void 641 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 642 { 643 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->data_from_pool); 644 if (req->req.cmd) { 645 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 646 } 647 if (req->recv) { 648 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 649 } 650 } 651 652 static void 653 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 654 { 655 int i; 656 657 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 658 for (i = 0; i < rqpair->max_queue_depth; i++) { 659 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 660 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 661 } 662 } 663 } 664 665 static void 666 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 667 { 668 if (resources->cmds_mr) { 669 ibv_dereg_mr(resources->cmds_mr); 670 } 671 672 if (resources->cpls_mr) { 673 ibv_dereg_mr(resources->cpls_mr); 674 } 675 676 if (resources->bufs_mr) { 677 ibv_dereg_mr(resources->bufs_mr); 678 } 679 680 spdk_dma_free(resources->cmds); 681 spdk_dma_free(resources->cpls); 682 spdk_dma_free(resources->bufs); 683 free(resources->reqs); 684 free(resources->recvs); 685 free(resources); 686 } 687 688 689 static struct spdk_nvmf_rdma_resources * 690 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 691 { 692 struct spdk_nvmf_rdma_resources *resources; 693 struct spdk_nvmf_rdma_request *rdma_req; 694 struct spdk_nvmf_rdma_recv *rdma_recv; 695 struct ibv_qp *qp; 696 struct ibv_srq *srq; 697 uint32_t i; 698 int rc; 699 700 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 701 if (!resources) { 702 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 703 return NULL; 704 } 705 706 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 707 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 708 resources->cmds = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 709 0x1000, NULL); 710 resources->cpls = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 711 0x1000, NULL); 712 713 if (opts->in_capsule_data_size > 0) { 714 resources->bufs = spdk_dma_zmalloc(opts->max_queue_depth * 715 opts->in_capsule_data_size, 716 0x1000, NULL); 717 } 718 719 if (!resources->reqs || !resources->recvs || !resources->cmds || 720 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 721 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 722 goto cleanup; 723 } 724 725 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 726 opts->max_queue_depth * sizeof(*resources->cmds), 727 IBV_ACCESS_LOCAL_WRITE); 728 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 729 opts->max_queue_depth * sizeof(*resources->cpls), 730 0); 731 732 if (opts->in_capsule_data_size) { 733 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 734 opts->max_queue_depth * 735 opts->in_capsule_data_size, 736 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 737 } 738 739 if (!resources->cmds_mr || !resources->cpls_mr || 740 (opts->in_capsule_data_size && 741 !resources->bufs_mr)) { 742 goto cleanup; 743 } 744 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 745 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 746 resources->cmds_mr->lkey); 747 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 748 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 749 resources->cpls_mr->lkey); 750 if (resources->bufs && resources->bufs_mr) { 751 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 752 resources->bufs, opts->max_queue_depth * 753 opts->in_capsule_data_size, resources->bufs_mr->lkey); 754 } 755 756 /* Initialize queues */ 757 STAILQ_INIT(&resources->incoming_queue); 758 STAILQ_INIT(&resources->free_queue); 759 760 for (i = 0; i < opts->max_queue_depth; i++) { 761 struct ibv_recv_wr *bad_wr = NULL; 762 763 rdma_recv = &resources->recvs[i]; 764 rdma_recv->qpair = opts->qpair; 765 766 /* Set up memory to receive commands */ 767 if (resources->bufs) { 768 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 769 opts->in_capsule_data_size)); 770 } 771 772 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 773 774 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 775 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 776 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 777 rdma_recv->wr.num_sge = 1; 778 779 if (rdma_recv->buf && resources->bufs_mr) { 780 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 781 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 782 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 783 rdma_recv->wr.num_sge++; 784 } 785 786 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 787 rdma_recv->wr.sg_list = rdma_recv->sgl; 788 if (opts->shared) { 789 srq = (struct ibv_srq *)opts->qp; 790 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 791 } else { 792 qp = (struct ibv_qp *)opts->qp; 793 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 794 } 795 if (rc) { 796 goto cleanup; 797 } 798 } 799 800 for (i = 0; i < opts->max_queue_depth; i++) { 801 rdma_req = &resources->reqs[i]; 802 803 if (opts->qpair != NULL) { 804 rdma_req->req.qpair = &opts->qpair->qpair; 805 } else { 806 rdma_req->req.qpair = NULL; 807 } 808 rdma_req->req.cmd = NULL; 809 810 /* Set up memory to send responses */ 811 rdma_req->req.rsp = &resources->cpls[i]; 812 813 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 814 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 815 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 816 817 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 818 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 819 rdma_req->rsp.wr.next = NULL; 820 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 821 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 822 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 823 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 824 825 /* Set up memory for data buffers */ 826 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 827 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 828 rdma_req->data.wr.next = NULL; 829 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 830 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 831 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 832 833 /* Initialize request state to FREE */ 834 rdma_req->state = RDMA_REQUEST_STATE_FREE; 835 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 836 } 837 838 return resources; 839 840 cleanup: 841 nvmf_rdma_resources_destroy(resources); 842 return NULL; 843 } 844 845 static void 846 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 847 { 848 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 849 struct ibv_recv_wr *bad_recv_wr = NULL; 850 int rc; 851 852 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 853 854 spdk_poller_unregister(&rqpair->destruct_poller); 855 856 if (rqpair->qd != 0) { 857 if (rqpair->srq == NULL) { 858 nvmf_rdma_dump_qpair_contents(rqpair); 859 } 860 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 861 } 862 863 if (rqpair->poller) { 864 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 865 866 if (rqpair->srq != NULL && rqpair->resources != NULL) { 867 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 868 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 869 if (rqpair == rdma_recv->qpair) { 870 STAILQ_REMOVE_HEAD(&rqpair->resources->incoming_queue, link); 871 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 872 if (rc) { 873 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 874 } 875 } 876 } 877 } 878 } 879 880 if (rqpair->cm_id) { 881 if (rqpair->cm_id->qp != NULL) { 882 rdma_destroy_qp(rqpair->cm_id); 883 } 884 rdma_destroy_id(rqpair->cm_id); 885 886 if (rqpair->poller != NULL && rqpair->srq == NULL) { 887 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 888 } 889 } 890 891 if (rqpair->srq == NULL && rqpair->resources != NULL) { 892 nvmf_rdma_resources_destroy(rqpair->resources); 893 } 894 895 free(rqpair); 896 } 897 898 static int 899 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 900 { 901 struct spdk_nvmf_rdma_poller *rpoller; 902 int rc, num_cqe, required_num_wr; 903 904 /* Enlarge CQ size dynamically */ 905 rpoller = rqpair->poller; 906 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 907 num_cqe = rpoller->num_cqe; 908 if (num_cqe < required_num_wr) { 909 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 910 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 911 } 912 913 if (rpoller->num_cqe != num_cqe) { 914 if (required_num_wr > device->attr.max_cqe) { 915 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 916 required_num_wr, device->attr.max_cqe); 917 return -1; 918 } 919 920 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 921 rc = ibv_resize_cq(rpoller->cq, num_cqe); 922 if (rc) { 923 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 924 return -1; 925 } 926 927 rpoller->num_cqe = num_cqe; 928 } 929 930 rpoller->required_num_wr = required_num_wr; 931 return 0; 932 } 933 934 static int 935 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 936 { 937 struct spdk_nvmf_rdma_qpair *rqpair; 938 int rc; 939 struct spdk_nvmf_rdma_transport *rtransport; 940 struct spdk_nvmf_transport *transport; 941 struct spdk_nvmf_rdma_resource_opts opts; 942 struct spdk_nvmf_rdma_device *device; 943 struct ibv_qp_init_attr ibv_init_attr; 944 945 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 946 device = rqpair->port->device; 947 948 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 949 ibv_init_attr.qp_context = rqpair; 950 ibv_init_attr.qp_type = IBV_QPT_RC; 951 ibv_init_attr.send_cq = rqpair->poller->cq; 952 ibv_init_attr.recv_cq = rqpair->poller->cq; 953 954 if (rqpair->srq) { 955 ibv_init_attr.srq = rqpair->srq; 956 } else { 957 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 958 1; /* RECV operations + dummy drain WR */ 959 } 960 961 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 962 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 963 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 964 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 965 966 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 967 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 968 goto error; 969 } 970 971 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 972 if (rc) { 973 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 974 goto error; 975 } 976 977 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2 + 1), 978 ibv_init_attr.cap.max_send_wr); 979 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 980 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 981 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 982 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 983 984 if (rqpair->poller->srq == NULL) { 985 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 986 transport = &rtransport->transport; 987 988 opts.qp = rqpair->cm_id->qp; 989 opts.pd = rqpair->cm_id->pd; 990 opts.qpair = rqpair; 991 opts.shared = false; 992 opts.max_queue_depth = rqpair->max_queue_depth; 993 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 994 995 rqpair->resources = nvmf_rdma_resources_create(&opts); 996 997 if (!rqpair->resources) { 998 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 999 goto error; 1000 } 1001 } else { 1002 rqpair->resources = rqpair->poller->resources; 1003 } 1004 1005 rqpair->current_recv_depth = 0; 1006 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1007 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1008 1009 return 0; 1010 1011 error: 1012 rdma_destroy_id(rqpair->cm_id); 1013 rqpair->cm_id = NULL; 1014 spdk_nvmf_rdma_qpair_destroy(rqpair); 1015 return -1; 1016 } 1017 1018 static int 1019 request_transfer_in(struct spdk_nvmf_request *req) 1020 { 1021 int rc; 1022 struct spdk_nvmf_rdma_request *rdma_req; 1023 struct spdk_nvmf_qpair *qpair; 1024 struct spdk_nvmf_rdma_qpair *rqpair; 1025 struct ibv_send_wr *bad_wr = NULL; 1026 1027 qpair = req->qpair; 1028 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1029 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1030 1031 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1032 assert(rdma_req != NULL); 1033 1034 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA READ POSTED. Request: %p Connection: %p\n", req, qpair); 1035 1036 rc = ibv_post_send(rqpair->cm_id->qp, &rdma_req->data.wr, &bad_wr); 1037 if (rc) { 1038 SPDK_ERRLOG("Unable to transfer data from host to target\n"); 1039 return -1; 1040 } 1041 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1042 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1043 return 0; 1044 } 1045 1046 static int 1047 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1048 { 1049 int rc; 1050 int num_outstanding_data_wr = 0; 1051 struct spdk_nvmf_rdma_request *rdma_req; 1052 struct spdk_nvmf_qpair *qpair; 1053 struct spdk_nvmf_rdma_qpair *rqpair; 1054 struct spdk_nvme_cpl *rsp; 1055 struct ibv_recv_wr *bad_recv_wr = NULL; 1056 struct ibv_send_wr *send_wr, *bad_send_wr = NULL; 1057 1058 *data_posted = 0; 1059 qpair = req->qpair; 1060 rsp = &req->rsp->nvme_cpl; 1061 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1062 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1063 1064 /* Advance our sq_head pointer */ 1065 if (qpair->sq_head == qpair->sq_head_max) { 1066 qpair->sq_head = 0; 1067 } else { 1068 qpair->sq_head++; 1069 } 1070 rsp->sqhd = qpair->sq_head; 1071 1072 /* Post the capsule to the recv buffer */ 1073 assert(rdma_req->recv != NULL); 1074 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA RECV POSTED. Recv: %p Connection: %p\n", rdma_req->recv, 1075 rqpair); 1076 if (rqpair->srq == NULL) { 1077 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_req->recv->wr, &bad_recv_wr); 1078 } else { 1079 rdma_req->recv->qpair = NULL; 1080 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 1081 } 1082 1083 if (rc) { 1084 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 1085 return rc; 1086 } 1087 rdma_req->recv = NULL; 1088 assert(rqpair->current_recv_depth > 0); 1089 rqpair->current_recv_depth--; 1090 1091 /* Build the response which consists of optional 1092 * RDMA WRITEs to transfer data, plus an RDMA SEND 1093 * containing the response. 1094 */ 1095 send_wr = &rdma_req->rsp.wr; 1096 1097 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 1098 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1099 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA WRITE POSTED. Request: %p Connection: %p\n", req, qpair); 1100 send_wr = &rdma_req->data.wr; 1101 *data_posted = 1; 1102 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1103 } 1104 1105 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA SEND POSTED. Request: %p Connection: %p\n", req, qpair); 1106 1107 /* Send the completion */ 1108 rc = ibv_post_send(rqpair->cm_id->qp, send_wr, &bad_send_wr); 1109 if (rc) { 1110 SPDK_ERRLOG("Unable to send response capsule\n"); 1111 return rc; 1112 } 1113 /* +1 for the rsp wr */ 1114 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1115 1116 return 0; 1117 } 1118 1119 static int 1120 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1121 { 1122 struct spdk_nvmf_rdma_accept_private_data accept_data; 1123 struct rdma_conn_param ctrlr_event_data = {}; 1124 int rc; 1125 1126 accept_data.recfmt = 0; 1127 accept_data.crqsize = rqpair->max_queue_depth; 1128 1129 ctrlr_event_data.private_data = &accept_data; 1130 ctrlr_event_data.private_data_len = sizeof(accept_data); 1131 if (id->ps == RDMA_PS_TCP) { 1132 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1133 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1134 } 1135 1136 rc = rdma_accept(id, &ctrlr_event_data); 1137 if (rc) { 1138 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 1139 } else { 1140 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1141 } 1142 1143 return rc; 1144 } 1145 1146 static void 1147 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1148 { 1149 struct spdk_nvmf_rdma_reject_private_data rej_data; 1150 1151 rej_data.recfmt = 0; 1152 rej_data.sts = error; 1153 1154 rdma_reject(id, &rej_data, sizeof(rej_data)); 1155 } 1156 1157 static int 1158 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 1159 new_qpair_fn cb_fn) 1160 { 1161 struct spdk_nvmf_rdma_transport *rtransport; 1162 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1163 struct spdk_nvmf_rdma_port *port; 1164 struct rdma_conn_param *rdma_param = NULL; 1165 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1166 uint16_t max_queue_depth; 1167 uint16_t max_read_depth; 1168 1169 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1170 1171 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1172 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1173 1174 rdma_param = &event->param.conn; 1175 if (rdma_param->private_data == NULL || 1176 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1177 SPDK_ERRLOG("connect request: no private data provided\n"); 1178 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1179 return -1; 1180 } 1181 1182 private_data = rdma_param->private_data; 1183 if (private_data->recfmt != 0) { 1184 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1185 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1186 return -1; 1187 } 1188 1189 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1190 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1191 1192 port = event->listen_id->context; 1193 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1194 event->listen_id, event->listen_id->verbs, port); 1195 1196 /* Figure out the supported queue depth. This is a multi-step process 1197 * that takes into account hardware maximums, host provided values, 1198 * and our target's internal memory limits */ 1199 1200 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1201 1202 /* Start with the maximum queue depth allowed by the target */ 1203 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1204 max_read_depth = rtransport->transport.opts.max_queue_depth; 1205 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1206 rtransport->transport.opts.max_queue_depth); 1207 1208 /* Next check the local NIC's hardware limitations */ 1209 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1210 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1211 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1212 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1213 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1214 1215 /* Next check the remote NIC's hardware limitations */ 1216 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1217 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1218 rdma_param->initiator_depth, rdma_param->responder_resources); 1219 if (rdma_param->initiator_depth > 0) { 1220 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1221 } 1222 1223 /* Finally check for the host software requested values, which are 1224 * optional. */ 1225 if (rdma_param->private_data != NULL && 1226 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1227 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1228 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1229 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1230 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1231 } 1232 1233 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1234 max_queue_depth, max_read_depth); 1235 1236 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1237 if (rqpair == NULL) { 1238 SPDK_ERRLOG("Could not allocate new connection.\n"); 1239 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1240 return -1; 1241 } 1242 1243 rqpair->port = port; 1244 rqpair->max_queue_depth = max_queue_depth; 1245 rqpair->max_read_depth = max_read_depth; 1246 rqpair->cm_id = event->id; 1247 rqpair->listen_id = event->listen_id; 1248 rqpair->qpair.transport = transport; 1249 1250 event->id->context = &rqpair->qpair; 1251 1252 cb_fn(&rqpair->qpair); 1253 1254 return 0; 1255 } 1256 1257 static int 1258 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1259 enum spdk_mem_map_notify_action action, 1260 void *vaddr, size_t size) 1261 { 1262 struct ibv_pd *pd = cb_ctx; 1263 struct ibv_mr *mr; 1264 1265 switch (action) { 1266 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1267 if (!g_nvmf_hooks.get_rkey) { 1268 mr = ibv_reg_mr(pd, vaddr, size, 1269 IBV_ACCESS_LOCAL_WRITE | 1270 IBV_ACCESS_REMOTE_READ | 1271 IBV_ACCESS_REMOTE_WRITE); 1272 if (mr == NULL) { 1273 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1274 return -1; 1275 } else { 1276 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1277 } 1278 } else { 1279 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1280 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1281 } 1282 break; 1283 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1284 if (!g_nvmf_hooks.get_rkey) { 1285 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1286 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1287 if (mr) { 1288 ibv_dereg_mr(mr); 1289 } 1290 } 1291 break; 1292 } 1293 1294 return 0; 1295 } 1296 1297 static int 1298 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1299 { 1300 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1301 return addr_1 == addr_2; 1302 } 1303 1304 static void 1305 spdk_nvmf_rdma_request_free_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1306 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1307 uint32_t num_buffers) 1308 { 1309 uint32_t i; 1310 1311 for (i = 0; i < num_buffers; i++) { 1312 if (group->buf_cache_count < group->buf_cache_size) { 1313 STAILQ_INSERT_HEAD(&group->buf_cache, 1314 (struct spdk_nvmf_transport_pg_cache_buf *)rdma_req->buffers[i], link); 1315 group->buf_cache_count++; 1316 } else { 1317 spdk_mempool_put(transport->data_buf_pool, rdma_req->buffers[i]); 1318 } 1319 rdma_req->req.iov[i].iov_base = NULL; 1320 rdma_req->buffers[i] = NULL; 1321 rdma_req->req.iov[i].iov_len = 0; 1322 1323 } 1324 rdma_req->data_from_pool = false; 1325 } 1326 1327 static int 1328 nvmf_rdma_request_get_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1329 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1330 uint32_t num_buffers) 1331 { 1332 uint32_t i = 0; 1333 1334 while (i < num_buffers) { 1335 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1336 group->buf_cache_count--; 1337 rdma_req->buffers[i] = STAILQ_FIRST(&group->buf_cache); 1338 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1339 assert(rdma_req->buffers[i] != NULL); 1340 i++; 1341 } else { 1342 if (spdk_mempool_get_bulk(transport->data_buf_pool, &rdma_req->buffers[i], num_buffers - i)) { 1343 goto err_exit; 1344 } 1345 i += num_buffers - i; 1346 } 1347 } 1348 1349 return 0; 1350 1351 err_exit: 1352 spdk_nvmf_rdma_request_free_buffers(rdma_req, group, transport, i); 1353 return -ENOMEM; 1354 } 1355 1356 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1357 1358 static spdk_nvme_data_transfer_t 1359 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1360 { 1361 enum spdk_nvme_data_transfer xfer; 1362 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1363 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1364 1365 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1366 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1367 rdma_req->rsp.wr.imm_data = 0; 1368 #endif 1369 1370 /* Figure out data transfer direction */ 1371 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1372 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1373 } else { 1374 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1375 1376 /* Some admin commands are special cases */ 1377 if ((rdma_req->req.qpair->qid == 0) && 1378 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1379 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1380 switch (cmd->cdw10 & 0xff) { 1381 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1382 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1383 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1384 break; 1385 default: 1386 xfer = SPDK_NVME_DATA_NONE; 1387 } 1388 } 1389 } 1390 1391 if (xfer == SPDK_NVME_DATA_NONE) { 1392 return xfer; 1393 } 1394 1395 /* Even for commands that may transfer data, they could have specified 0 length. 1396 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1397 */ 1398 switch (sgl->generic.type) { 1399 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1400 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1401 case SPDK_NVME_SGL_TYPE_SEGMENT: 1402 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1403 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1404 if (sgl->unkeyed.length == 0) { 1405 xfer = SPDK_NVME_DATA_NONE; 1406 } 1407 break; 1408 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1409 if (sgl->keyed.length == 0) { 1410 xfer = SPDK_NVME_DATA_NONE; 1411 } 1412 break; 1413 } 1414 1415 return xfer; 1416 } 1417 1418 static int 1419 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1420 struct spdk_nvmf_rdma_request *rdma_req, 1421 uint32_t num_sgl_descriptors) 1422 { 1423 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1424 struct spdk_nvmf_rdma_request_data *current_data_wr; 1425 uint32_t i; 1426 1427 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1428 return -ENOMEM; 1429 } 1430 1431 current_data_wr = &rdma_req->data; 1432 1433 for (i = 0; i < num_sgl_descriptors; i++) { 1434 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1435 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1436 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1437 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1438 } else { 1439 assert(false); 1440 } 1441 work_requests[i]->wr.send_flags = IBV_SEND_SIGNALED; 1442 work_requests[i]->wr.sg_list = work_requests[i]->sgl; 1443 work_requests[i]->wr.wr_id = rdma_req->data.wr.wr_id; 1444 current_data_wr->wr.next = &work_requests[i]->wr; 1445 current_data_wr = work_requests[i]; 1446 } 1447 1448 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1449 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1450 current_data_wr->wr.next = &rdma_req->rsp.wr; 1451 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1452 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1453 current_data_wr->wr.next = NULL; 1454 } 1455 return 0; 1456 } 1457 1458 static int 1459 nvmf_rdma_fill_buffers(struct spdk_nvmf_rdma_transport *rtransport, 1460 struct spdk_nvmf_rdma_poll_group *rgroup, 1461 struct spdk_nvmf_rdma_device *device, 1462 struct spdk_nvmf_rdma_request *rdma_req, 1463 struct ibv_send_wr *wr, 1464 uint32_t length) 1465 { 1466 uint64_t translation_len; 1467 uint32_t remaining_length = length; 1468 uint32_t iovcnt; 1469 uint32_t i = 0; 1470 1471 1472 while (remaining_length) { 1473 iovcnt = rdma_req->req.iovcnt; 1474 rdma_req->req.iov[iovcnt].iov_base = (void *)((uintptr_t)(rdma_req->buffers[iovcnt] + 1475 NVMF_DATA_BUFFER_MASK) & 1476 ~NVMF_DATA_BUFFER_MASK); 1477 rdma_req->req.iov[iovcnt].iov_len = spdk_min(remaining_length, 1478 rtransport->transport.opts.io_unit_size); 1479 rdma_req->req.iovcnt++; 1480 wr->sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[iovcnt].iov_base); 1481 wr->sg_list[i].length = rdma_req->req.iov[iovcnt].iov_len; 1482 translation_len = rdma_req->req.iov[iovcnt].iov_len; 1483 1484 if (!g_nvmf_hooks.get_rkey) { 1485 wr->sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1486 (uint64_t)rdma_req->buffers[iovcnt], &translation_len))->lkey; 1487 } else { 1488 wr->sg_list[i].lkey = spdk_mem_map_translate(device->map, 1489 (uint64_t)rdma_req->buffers[iovcnt], &translation_len); 1490 } 1491 1492 remaining_length -= rdma_req->req.iov[iovcnt].iov_len; 1493 1494 if (translation_len < rdma_req->req.iov[iovcnt].iov_len) { 1495 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions\n"); 1496 return -EINVAL; 1497 } 1498 i++; 1499 } 1500 wr->num_sge = i; 1501 1502 return 0; 1503 } 1504 1505 static int 1506 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1507 struct spdk_nvmf_rdma_device *device, 1508 struct spdk_nvmf_rdma_request *rdma_req) 1509 { 1510 struct spdk_nvmf_rdma_qpair *rqpair; 1511 struct spdk_nvmf_rdma_poll_group *rgroup; 1512 uint32_t num_buffers; 1513 uint32_t i = 0; 1514 int rc = 0; 1515 1516 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1517 rgroup = rqpair->poller->group; 1518 rdma_req->req.iovcnt = 0; 1519 1520 num_buffers = rdma_req->req.length / rtransport->transport.opts.io_unit_size; 1521 if (rdma_req->req.length % rtransport->transport.opts.io_unit_size) { 1522 num_buffers++; 1523 } 1524 1525 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers)) { 1526 return -ENOMEM; 1527 } 1528 1529 rdma_req->req.iovcnt = 0; 1530 1531 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, &rdma_req->data.wr, 1532 rdma_req->req.length); 1533 if (rc != 0) { 1534 goto err_exit; 1535 } 1536 1537 assert(rdma_req->req.iovcnt <= rqpair->max_send_sge); 1538 1539 rdma_req->data_from_pool = true; 1540 1541 return rc; 1542 1543 err_exit: 1544 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1545 while (i) { 1546 i--; 1547 rdma_req->data.wr.sg_list[i].addr = 0; 1548 rdma_req->data.wr.sg_list[i].length = 0; 1549 rdma_req->data.wr.sg_list[i].lkey = 0; 1550 } 1551 rdma_req->req.iovcnt = 0; 1552 return rc; 1553 } 1554 1555 static int 1556 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1557 struct spdk_nvmf_rdma_device *device, 1558 struct spdk_nvmf_rdma_request *rdma_req) 1559 { 1560 struct spdk_nvmf_rdma_qpair *rqpair; 1561 struct spdk_nvmf_rdma_poll_group *rgroup; 1562 struct ibv_send_wr *current_wr; 1563 struct spdk_nvmf_request *req = &rdma_req->req; 1564 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1565 uint32_t num_sgl_descriptors; 1566 uint32_t num_buffers = 0; 1567 uint32_t i; 1568 int rc; 1569 1570 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1571 rgroup = rqpair->poller->group; 1572 1573 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1574 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1575 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1576 1577 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1578 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1579 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1580 1581 for (i = 0; i < num_sgl_descriptors; i++) { 1582 num_buffers += desc->keyed.length / rtransport->transport.opts.io_unit_size; 1583 if (desc->keyed.length % rtransport->transport.opts.io_unit_size) { 1584 num_buffers++; 1585 } 1586 desc++; 1587 } 1588 /* If the number of buffers is too large, then we know the I/O is larger than allowed. Fail it. */ 1589 if (num_buffers > NVMF_REQ_MAX_BUFFERS) { 1590 return -EINVAL; 1591 } 1592 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1593 num_buffers) != 0) { 1594 return -ENOMEM; 1595 } 1596 1597 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1598 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1599 return -ENOMEM; 1600 } 1601 1602 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1603 current_wr = &rdma_req->data.wr; 1604 1605 req->iovcnt = 0; 1606 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1607 for (i = 0; i < num_sgl_descriptors; i++) { 1608 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1609 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1610 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1611 rc = -EINVAL; 1612 goto err_exit; 1613 } 1614 1615 current_wr->num_sge = 0; 1616 req->length += desc->keyed.length; 1617 1618 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, current_wr, 1619 desc->keyed.length); 1620 if (rc != 0) { 1621 rc = -ENOMEM; 1622 goto err_exit; 1623 } 1624 1625 current_wr->wr.rdma.rkey = desc->keyed.key; 1626 current_wr->wr.rdma.remote_addr = desc->address; 1627 current_wr = current_wr->next; 1628 desc++; 1629 } 1630 1631 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1632 /* Go back to the last descriptor in the list. */ 1633 desc--; 1634 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1635 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1636 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1637 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1638 } 1639 } 1640 #endif 1641 1642 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1643 rdma_req->data_from_pool = true; 1644 1645 return 0; 1646 1647 err_exit: 1648 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1649 nvmf_rdma_request_free_data(rdma_req, rtransport); 1650 return rc; 1651 } 1652 1653 static int 1654 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1655 struct spdk_nvmf_rdma_device *device, 1656 struct spdk_nvmf_rdma_request *rdma_req) 1657 { 1658 struct spdk_nvme_cmd *cmd; 1659 struct spdk_nvme_cpl *rsp; 1660 struct spdk_nvme_sgl_descriptor *sgl; 1661 int rc; 1662 1663 cmd = &rdma_req->req.cmd->nvme_cmd; 1664 rsp = &rdma_req->req.rsp->nvme_cpl; 1665 sgl = &cmd->dptr.sgl1; 1666 1667 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1668 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1669 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1670 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1671 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1672 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1673 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1674 return -1; 1675 } 1676 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1677 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1678 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1679 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1680 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1681 } 1682 } 1683 #endif 1684 1685 /* fill request length and populate iovs */ 1686 rdma_req->req.length = sgl->keyed.length; 1687 1688 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1689 /* No available buffers. Queue this request up. */ 1690 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1691 return 0; 1692 } 1693 1694 /* backward compatible */ 1695 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1696 1697 /* rdma wr specifics */ 1698 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1699 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1700 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1701 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1702 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 1703 rdma_req->data.wr.next = &rdma_req->rsp.wr; 1704 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1705 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 1706 rdma_req->data.wr.next = NULL; 1707 } 1708 1709 /* set the number of outstanding data WRs for this request. */ 1710 rdma_req->num_outstanding_data_wr = 1; 1711 1712 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1713 rdma_req->req.iovcnt); 1714 1715 return 0; 1716 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1717 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1718 uint64_t offset = sgl->address; 1719 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1720 1721 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1722 offset, sgl->unkeyed.length); 1723 1724 if (offset > max_len) { 1725 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1726 offset, max_len); 1727 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1728 return -1; 1729 } 1730 max_len -= (uint32_t)offset; 1731 1732 if (sgl->unkeyed.length > max_len) { 1733 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1734 sgl->unkeyed.length, max_len); 1735 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1736 return -1; 1737 } 1738 1739 rdma_req->num_outstanding_data_wr = 0; 1740 rdma_req->req.data = rdma_req->recv->buf + offset; 1741 rdma_req->data_from_pool = false; 1742 rdma_req->req.length = sgl->unkeyed.length; 1743 1744 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1745 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1746 rdma_req->req.iovcnt = 1; 1747 1748 return 0; 1749 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1750 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1751 1752 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1753 if (rc == -ENOMEM) { 1754 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1755 return 0; 1756 } else if (rc == -EINVAL) { 1757 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1758 return -1; 1759 } 1760 1761 /* backward compatible */ 1762 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1763 1764 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1765 rdma_req->req.iovcnt); 1766 1767 return 0; 1768 } 1769 1770 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1771 sgl->generic.type, sgl->generic.subtype); 1772 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1773 return -1; 1774 } 1775 1776 static void 1777 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1778 struct spdk_nvmf_rdma_transport *rtransport) 1779 { 1780 struct spdk_nvmf_rdma_qpair *rqpair; 1781 struct spdk_nvmf_rdma_poll_group *rgroup; 1782 1783 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1784 if (rdma_req->data_from_pool) { 1785 rgroup = rqpair->poller->group; 1786 1787 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1788 rdma_req->req.iovcnt); 1789 } 1790 nvmf_rdma_request_free_data(rdma_req, rtransport); 1791 rdma_req->req.length = 0; 1792 rdma_req->req.iovcnt = 0; 1793 rdma_req->req.data = NULL; 1794 rdma_req->data.wr.next = NULL; 1795 rqpair->qd--; 1796 1797 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1798 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1799 } 1800 1801 static bool 1802 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1803 struct spdk_nvmf_rdma_request *rdma_req) 1804 { 1805 struct spdk_nvmf_rdma_qpair *rqpair; 1806 struct spdk_nvmf_rdma_device *device; 1807 struct spdk_nvmf_rdma_poll_group *rgroup; 1808 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1809 int rc; 1810 struct spdk_nvmf_rdma_recv *rdma_recv; 1811 enum spdk_nvmf_rdma_request_state prev_state; 1812 bool progress = false; 1813 int data_posted; 1814 1815 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1816 device = rqpair->port->device; 1817 rgroup = rqpair->poller->group; 1818 1819 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1820 1821 /* If the queue pair is in an error state, force the request to the completed state 1822 * to release resources. */ 1823 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1824 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1825 STAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1826 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1827 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1828 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1829 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1830 } 1831 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1832 } 1833 1834 /* The loop here is to allow for several back-to-back state changes. */ 1835 do { 1836 prev_state = rdma_req->state; 1837 1838 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1839 1840 switch (rdma_req->state) { 1841 case RDMA_REQUEST_STATE_FREE: 1842 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1843 * to escape this state. */ 1844 break; 1845 case RDMA_REQUEST_STATE_NEW: 1846 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1847 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1848 rdma_recv = rdma_req->recv; 1849 1850 /* The first element of the SGL is the NVMe command */ 1851 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1852 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1853 1854 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1855 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1856 break; 1857 } 1858 1859 /* The next state transition depends on the data transfer needs of this request. */ 1860 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1861 1862 /* If no data to transfer, ready to execute. */ 1863 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1864 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1865 break; 1866 } 1867 1868 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1869 STAILQ_INSERT_TAIL(&rgroup->pending_data_buf_queue, rdma_req, state_link); 1870 break; 1871 case RDMA_REQUEST_STATE_NEED_BUFFER: 1872 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1873 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1874 1875 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1876 1877 if (rdma_req != STAILQ_FIRST(&rgroup->pending_data_buf_queue)) { 1878 /* This request needs to wait in line to obtain a buffer */ 1879 break; 1880 } 1881 1882 /* Try to get a data buffer */ 1883 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1884 if (rc < 0) { 1885 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1886 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1887 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1888 break; 1889 } 1890 1891 if (!rdma_req->req.data) { 1892 /* No buffers available. */ 1893 break; 1894 } 1895 1896 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1897 1898 /* If data is transferring from host to controller and the data didn't 1899 * arrive using in capsule data, we need to do a transfer from the host. 1900 */ 1901 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1902 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1903 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1904 break; 1905 } 1906 1907 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1908 break; 1909 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1910 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1911 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1912 1913 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1914 /* This request needs to wait in line to perform RDMA */ 1915 break; 1916 } 1917 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1918 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1919 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1920 break; 1921 } 1922 1923 /* We have already verified that this request is the head of the queue. */ 1924 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 1925 1926 rc = request_transfer_in(&rdma_req->req); 1927 if (!rc) { 1928 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 1929 } else { 1930 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1931 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1932 } 1933 break; 1934 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1935 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1936 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1937 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1938 * to escape this state. */ 1939 break; 1940 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1941 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1942 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1943 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 1944 spdk_nvmf_request_exec(&rdma_req->req); 1945 break; 1946 case RDMA_REQUEST_STATE_EXECUTING: 1947 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1948 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1949 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1950 * to escape this state. */ 1951 break; 1952 case RDMA_REQUEST_STATE_EXECUTED: 1953 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1954 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1955 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1956 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 1957 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 1958 } else { 1959 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1960 } 1961 break; 1962 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 1963 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 1964 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1965 1966 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 1967 /* This request needs to wait in line to perform RDMA */ 1968 break; 1969 } 1970 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 1971 rqpair->max_send_depth) { 1972 /* We can only have so many WRs outstanding. we have to wait until some finish. 1973 * +1 since each request has an additional wr in the resp. */ 1974 break; 1975 } 1976 1977 /* We have already verified that this request is the head of the queue. */ 1978 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 1979 1980 /* The data transfer will be kicked off from 1981 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1982 */ 1983 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1984 break; 1985 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1986 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 1987 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1988 rc = request_transfer_out(&rdma_req->req, &data_posted); 1989 assert(rc == 0); /* No good way to handle this currently */ 1990 if (rc) { 1991 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1992 } else { 1993 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 1994 RDMA_REQUEST_STATE_COMPLETING; 1995 } 1996 break; 1997 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 1998 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 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_COMPLETING: 2004 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2005 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2006 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2007 * to escape this state. */ 2008 break; 2009 case RDMA_REQUEST_STATE_COMPLETED: 2010 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2011 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2012 2013 nvmf_rdma_request_free(rdma_req, rtransport); 2014 break; 2015 case RDMA_REQUEST_NUM_STATES: 2016 default: 2017 assert(0); 2018 break; 2019 } 2020 2021 if (rdma_req->state != prev_state) { 2022 progress = true; 2023 } 2024 } while (rdma_req->state != prev_state); 2025 2026 return progress; 2027 } 2028 2029 /* Public API callbacks begin here */ 2030 2031 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2032 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2033 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2034 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 2035 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2036 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2037 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2038 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4096 2039 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2040 2041 static void 2042 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2043 { 2044 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2045 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2046 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2047 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2048 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2049 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2050 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2051 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2052 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2053 } 2054 2055 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2056 .notify_cb = spdk_nvmf_rdma_mem_notify, 2057 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 2058 }; 2059 2060 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2061 2062 static struct spdk_nvmf_transport * 2063 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2064 { 2065 int rc; 2066 struct spdk_nvmf_rdma_transport *rtransport; 2067 struct spdk_nvmf_rdma_device *device, *tmp; 2068 struct ibv_pd *pd; 2069 struct ibv_context **contexts; 2070 uint32_t i; 2071 int flag; 2072 uint32_t sge_count; 2073 uint32_t min_shared_buffers; 2074 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2075 2076 rtransport = calloc(1, sizeof(*rtransport)); 2077 if (!rtransport) { 2078 return NULL; 2079 } 2080 2081 if (pthread_mutex_init(&rtransport->lock, NULL)) { 2082 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2083 free(rtransport); 2084 return NULL; 2085 } 2086 2087 TAILQ_INIT(&rtransport->devices); 2088 TAILQ_INIT(&rtransport->ports); 2089 2090 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2091 2092 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2093 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2094 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2095 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2096 " num_shared_buffers=%d, max_srq_depth=%d\n", 2097 opts->max_queue_depth, 2098 opts->max_io_size, 2099 opts->max_qpairs_per_ctrlr, 2100 opts->io_unit_size, 2101 opts->in_capsule_data_size, 2102 opts->max_aq_depth, 2103 opts->num_shared_buffers, 2104 opts->max_srq_depth); 2105 2106 /* I/O unit size cannot be larger than max I/O size */ 2107 if (opts->io_unit_size > opts->max_io_size) { 2108 opts->io_unit_size = opts->max_io_size; 2109 } 2110 2111 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2112 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2113 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2114 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2115 spdk_nvmf_rdma_destroy(&rtransport->transport); 2116 return NULL; 2117 } 2118 2119 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2120 if (min_shared_buffers > opts->num_shared_buffers) { 2121 SPDK_ERRLOG("There are not enough buffers to satisfy" 2122 "per-poll group caches for each thread. (%" PRIu32 ")" 2123 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2124 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2125 spdk_nvmf_rdma_destroy(&rtransport->transport); 2126 return NULL; 2127 } 2128 2129 sge_count = opts->max_io_size / opts->io_unit_size; 2130 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2131 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2132 spdk_nvmf_rdma_destroy(&rtransport->transport); 2133 return NULL; 2134 } 2135 2136 rtransport->event_channel = rdma_create_event_channel(); 2137 if (rtransport->event_channel == NULL) { 2138 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2139 spdk_nvmf_rdma_destroy(&rtransport->transport); 2140 return NULL; 2141 } 2142 2143 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2144 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2145 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2146 rtransport->event_channel->fd, spdk_strerror(errno)); 2147 spdk_nvmf_rdma_destroy(&rtransport->transport); 2148 return NULL; 2149 } 2150 2151 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2152 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2153 sizeof(struct spdk_nvmf_rdma_request_data), 2154 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2155 SPDK_ENV_SOCKET_ID_ANY); 2156 if (!rtransport->data_wr_pool) { 2157 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2158 spdk_nvmf_rdma_destroy(&rtransport->transport); 2159 return NULL; 2160 } 2161 2162 contexts = rdma_get_devices(NULL); 2163 if (contexts == NULL) { 2164 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2165 spdk_nvmf_rdma_destroy(&rtransport->transport); 2166 return NULL; 2167 } 2168 2169 i = 0; 2170 rc = 0; 2171 while (contexts[i] != NULL) { 2172 device = calloc(1, sizeof(*device)); 2173 if (!device) { 2174 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2175 rc = -ENOMEM; 2176 break; 2177 } 2178 device->context = contexts[i]; 2179 rc = ibv_query_device(device->context, &device->attr); 2180 if (rc < 0) { 2181 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2182 free(device); 2183 break; 2184 2185 } 2186 2187 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2188 2189 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2190 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2191 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2192 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2193 } 2194 2195 /** 2196 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2197 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2198 * but incorrectly reports that it does. There are changes making their way 2199 * through the kernel now that will enable this feature. When they are merged, 2200 * we can conditionally enable this feature. 2201 * 2202 * TODO: enable this for versions of the kernel rxe driver that support it. 2203 */ 2204 if (device->attr.vendor_id == 0) { 2205 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2206 } 2207 #endif 2208 2209 /* set up device context async ev fd as NON_BLOCKING */ 2210 flag = fcntl(device->context->async_fd, F_GETFL); 2211 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2212 if (rc < 0) { 2213 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2214 free(device); 2215 break; 2216 } 2217 2218 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2219 i++; 2220 2221 pd = NULL; 2222 if (g_nvmf_hooks.get_ibv_pd) { 2223 pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2224 } 2225 2226 if (!g_nvmf_hooks.get_ibv_pd) { 2227 device->pd = ibv_alloc_pd(device->context); 2228 if (!device->pd) { 2229 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2230 spdk_nvmf_rdma_destroy(&rtransport->transport); 2231 return NULL; 2232 } 2233 } else { 2234 device->pd = pd; 2235 } 2236 2237 assert(device->map == NULL); 2238 2239 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2240 if (!device->map) { 2241 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2242 spdk_nvmf_rdma_destroy(&rtransport->transport); 2243 return NULL; 2244 } 2245 2246 assert(device->map != NULL); 2247 assert(device->pd != NULL); 2248 } 2249 rdma_free_devices(contexts); 2250 2251 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2252 /* divide and round up. */ 2253 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2254 2255 /* round up to the nearest 4k. */ 2256 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2257 2258 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2259 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2260 opts->io_unit_size); 2261 } 2262 2263 if (rc < 0) { 2264 spdk_nvmf_rdma_destroy(&rtransport->transport); 2265 return NULL; 2266 } 2267 2268 /* Set up poll descriptor array to monitor events from RDMA and IB 2269 * in a single poll syscall 2270 */ 2271 rtransport->npoll_fds = i + 1; 2272 i = 0; 2273 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2274 if (rtransport->poll_fds == NULL) { 2275 SPDK_ERRLOG("poll_fds allocation failed\n"); 2276 spdk_nvmf_rdma_destroy(&rtransport->transport); 2277 return NULL; 2278 } 2279 2280 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2281 rtransport->poll_fds[i++].events = POLLIN; 2282 2283 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2284 rtransport->poll_fds[i].fd = device->context->async_fd; 2285 rtransport->poll_fds[i++].events = POLLIN; 2286 } 2287 2288 return &rtransport->transport; 2289 } 2290 2291 static int 2292 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2293 { 2294 struct spdk_nvmf_rdma_transport *rtransport; 2295 struct spdk_nvmf_rdma_port *port, *port_tmp; 2296 struct spdk_nvmf_rdma_device *device, *device_tmp; 2297 2298 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2299 2300 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2301 TAILQ_REMOVE(&rtransport->ports, port, link); 2302 rdma_destroy_id(port->id); 2303 free(port); 2304 } 2305 2306 if (rtransport->poll_fds != NULL) { 2307 free(rtransport->poll_fds); 2308 } 2309 2310 if (rtransport->event_channel != NULL) { 2311 rdma_destroy_event_channel(rtransport->event_channel); 2312 } 2313 2314 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2315 TAILQ_REMOVE(&rtransport->devices, device, link); 2316 if (device->map) { 2317 spdk_mem_map_free(&device->map); 2318 } 2319 if (device->pd) { 2320 if (!g_nvmf_hooks.get_ibv_pd) { 2321 ibv_dealloc_pd(device->pd); 2322 } 2323 } 2324 free(device); 2325 } 2326 2327 if (rtransport->data_wr_pool != NULL) { 2328 if (spdk_mempool_count(rtransport->data_wr_pool) != 2329 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2330 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2331 spdk_mempool_count(rtransport->data_wr_pool), 2332 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2333 } 2334 } 2335 2336 spdk_mempool_free(rtransport->data_wr_pool); 2337 pthread_mutex_destroy(&rtransport->lock); 2338 free(rtransport); 2339 2340 return 0; 2341 } 2342 2343 static int 2344 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2345 struct spdk_nvme_transport_id *trid, 2346 bool peer); 2347 2348 static int 2349 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2350 const struct spdk_nvme_transport_id *trid) 2351 { 2352 struct spdk_nvmf_rdma_transport *rtransport; 2353 struct spdk_nvmf_rdma_device *device; 2354 struct spdk_nvmf_rdma_port *port_tmp, *port; 2355 struct addrinfo *res; 2356 struct addrinfo hints; 2357 int family; 2358 int rc; 2359 2360 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2361 2362 port = calloc(1, sizeof(*port)); 2363 if (!port) { 2364 return -ENOMEM; 2365 } 2366 2367 /* Selectively copy the trid. Things like NQN don't matter here - that 2368 * mapping is enforced elsewhere. 2369 */ 2370 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2371 port->trid.adrfam = trid->adrfam; 2372 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 2373 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 2374 2375 pthread_mutex_lock(&rtransport->lock); 2376 assert(rtransport->event_channel != NULL); 2377 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 2378 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 2379 port_tmp->ref++; 2380 free(port); 2381 /* Already listening at this address */ 2382 pthread_mutex_unlock(&rtransport->lock); 2383 return 0; 2384 } 2385 } 2386 2387 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2388 if (rc < 0) { 2389 SPDK_ERRLOG("rdma_create_id() failed\n"); 2390 free(port); 2391 pthread_mutex_unlock(&rtransport->lock); 2392 return rc; 2393 } 2394 2395 switch (port->trid.adrfam) { 2396 case SPDK_NVMF_ADRFAM_IPV4: 2397 family = AF_INET; 2398 break; 2399 case SPDK_NVMF_ADRFAM_IPV6: 2400 family = AF_INET6; 2401 break; 2402 default: 2403 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 2404 free(port); 2405 pthread_mutex_unlock(&rtransport->lock); 2406 return -EINVAL; 2407 } 2408 2409 memset(&hints, 0, sizeof(hints)); 2410 hints.ai_family = family; 2411 hints.ai_flags = AI_NUMERICSERV; 2412 hints.ai_socktype = SOCK_STREAM; 2413 hints.ai_protocol = 0; 2414 2415 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 2416 if (rc) { 2417 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2418 free(port); 2419 pthread_mutex_unlock(&rtransport->lock); 2420 return -EINVAL; 2421 } 2422 2423 rc = rdma_bind_addr(port->id, res->ai_addr); 2424 freeaddrinfo(res); 2425 2426 if (rc < 0) { 2427 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2428 rdma_destroy_id(port->id); 2429 free(port); 2430 pthread_mutex_unlock(&rtransport->lock); 2431 return rc; 2432 } 2433 2434 if (!port->id->verbs) { 2435 SPDK_ERRLOG("ibv_context is null\n"); 2436 rdma_destroy_id(port->id); 2437 free(port); 2438 pthread_mutex_unlock(&rtransport->lock); 2439 return -1; 2440 } 2441 2442 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 2443 if (rc < 0) { 2444 SPDK_ERRLOG("rdma_listen() failed\n"); 2445 rdma_destroy_id(port->id); 2446 free(port); 2447 pthread_mutex_unlock(&rtransport->lock); 2448 return rc; 2449 } 2450 2451 TAILQ_FOREACH(device, &rtransport->devices, link) { 2452 if (device->context == port->id->verbs) { 2453 port->device = device; 2454 break; 2455 } 2456 } 2457 if (!port->device) { 2458 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2459 port->id->verbs); 2460 rdma_destroy_id(port->id); 2461 free(port); 2462 pthread_mutex_unlock(&rtransport->lock); 2463 return -EINVAL; 2464 } 2465 2466 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 2467 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 2468 2469 port->ref = 1; 2470 2471 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2472 pthread_mutex_unlock(&rtransport->lock); 2473 2474 return 0; 2475 } 2476 2477 static int 2478 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2479 const struct spdk_nvme_transport_id *_trid) 2480 { 2481 struct spdk_nvmf_rdma_transport *rtransport; 2482 struct spdk_nvmf_rdma_port *port, *tmp; 2483 struct spdk_nvme_transport_id trid = {}; 2484 2485 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2486 2487 /* Selectively copy the trid. Things like NQN don't matter here - that 2488 * mapping is enforced elsewhere. 2489 */ 2490 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2491 trid.adrfam = _trid->adrfam; 2492 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2493 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2494 2495 pthread_mutex_lock(&rtransport->lock); 2496 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2497 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2498 assert(port->ref > 0); 2499 port->ref--; 2500 if (port->ref == 0) { 2501 TAILQ_REMOVE(&rtransport->ports, port, link); 2502 rdma_destroy_id(port->id); 2503 free(port); 2504 } 2505 break; 2506 } 2507 } 2508 2509 pthread_mutex_unlock(&rtransport->lock); 2510 return 0; 2511 } 2512 2513 static void 2514 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2515 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2516 { 2517 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2518 struct spdk_nvmf_rdma_resources *resources; 2519 2520 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2521 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2522 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2523 break; 2524 } 2525 } 2526 2527 /* Then RDMA writes since reads have stronger restrictions than writes */ 2528 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2529 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2530 break; 2531 } 2532 } 2533 2534 /* The second highest priority is I/O waiting on memory buffers. */ 2535 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->poller->group->pending_data_buf_queue, state_link, 2536 req_tmp) { 2537 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2538 break; 2539 } 2540 } 2541 2542 resources = rqpair->resources; 2543 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2544 rdma_req = STAILQ_FIRST(&resources->free_queue); 2545 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2546 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2547 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2548 2549 if (rqpair->srq != NULL) { 2550 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2551 rdma_req->recv->qpair->qd++; 2552 } else { 2553 rqpair->qd++; 2554 } 2555 2556 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2557 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2558 break; 2559 } 2560 } 2561 } 2562 2563 static void 2564 _nvmf_rdma_qpair_disconnect(void *ctx) 2565 { 2566 struct spdk_nvmf_qpair *qpair = ctx; 2567 2568 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2569 } 2570 2571 static void 2572 _nvmf_rdma_try_disconnect(void *ctx) 2573 { 2574 struct spdk_nvmf_qpair *qpair = ctx; 2575 struct spdk_nvmf_poll_group *group; 2576 2577 /* Read the group out of the qpair. This is normally set and accessed only from 2578 * the thread that created the group. Here, we're not on that thread necessarily. 2579 * The data member qpair->group begins it's life as NULL and then is assigned to 2580 * a pointer and never changes. So fortunately reading this and checking for 2581 * non-NULL is thread safe in the x86_64 memory model. */ 2582 group = qpair->group; 2583 2584 if (group == NULL) { 2585 /* The qpair hasn't been assigned to a group yet, so we can't 2586 * process a disconnect. Send a message to ourself and try again. */ 2587 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2588 return; 2589 } 2590 2591 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2592 } 2593 2594 static inline void 2595 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2596 { 2597 if (__sync_bool_compare_and_swap(&rqpair->disconnect_started, false, true)) { 2598 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2599 } 2600 } 2601 2602 static void nvmf_rdma_destroy_drained_qpair(void *ctx) 2603 { 2604 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2605 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2606 struct spdk_nvmf_rdma_transport, transport); 2607 2608 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2609 if (rqpair->current_send_depth != 0) { 2610 return; 2611 } 2612 2613 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2614 return; 2615 } 2616 2617 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) { 2618 return; 2619 } 2620 2621 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2622 spdk_nvmf_rdma_qpair_destroy(rqpair); 2623 } 2624 2625 2626 static int 2627 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2628 { 2629 struct spdk_nvmf_qpair *qpair; 2630 struct spdk_nvmf_rdma_qpair *rqpair; 2631 2632 if (evt->id == NULL) { 2633 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2634 return -1; 2635 } 2636 2637 qpair = evt->id->context; 2638 if (qpair == NULL) { 2639 SPDK_ERRLOG("disconnect request: no active connection\n"); 2640 return -1; 2641 } 2642 2643 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2644 2645 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2646 2647 spdk_nvmf_rdma_update_ibv_state(rqpair); 2648 2649 spdk_nvmf_rdma_start_disconnect(rqpair); 2650 2651 return 0; 2652 } 2653 2654 #ifdef DEBUG 2655 static const char *CM_EVENT_STR[] = { 2656 "RDMA_CM_EVENT_ADDR_RESOLVED", 2657 "RDMA_CM_EVENT_ADDR_ERROR", 2658 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2659 "RDMA_CM_EVENT_ROUTE_ERROR", 2660 "RDMA_CM_EVENT_CONNECT_REQUEST", 2661 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2662 "RDMA_CM_EVENT_CONNECT_ERROR", 2663 "RDMA_CM_EVENT_UNREACHABLE", 2664 "RDMA_CM_EVENT_REJECTED", 2665 "RDMA_CM_EVENT_ESTABLISHED", 2666 "RDMA_CM_EVENT_DISCONNECTED", 2667 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2668 "RDMA_CM_EVENT_MULTICAST_JOIN", 2669 "RDMA_CM_EVENT_MULTICAST_ERROR", 2670 "RDMA_CM_EVENT_ADDR_CHANGE", 2671 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2672 }; 2673 #endif /* DEBUG */ 2674 2675 static void 2676 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2677 { 2678 struct spdk_nvmf_rdma_transport *rtransport; 2679 struct rdma_cm_event *event; 2680 int rc; 2681 2682 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2683 2684 if (rtransport->event_channel == NULL) { 2685 return; 2686 } 2687 2688 while (1) { 2689 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2690 if (rc == 0) { 2691 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2692 2693 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2694 2695 switch (event->event) { 2696 case RDMA_CM_EVENT_ADDR_RESOLVED: 2697 case RDMA_CM_EVENT_ADDR_ERROR: 2698 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2699 case RDMA_CM_EVENT_ROUTE_ERROR: 2700 /* No action required. The target never attempts to resolve routes. */ 2701 break; 2702 case RDMA_CM_EVENT_CONNECT_REQUEST: 2703 rc = nvmf_rdma_connect(transport, event, cb_fn); 2704 if (rc < 0) { 2705 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2706 break; 2707 } 2708 break; 2709 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2710 /* The target never initiates a new connection. So this will not occur. */ 2711 break; 2712 case RDMA_CM_EVENT_CONNECT_ERROR: 2713 /* Can this happen? The docs say it can, but not sure what causes it. */ 2714 break; 2715 case RDMA_CM_EVENT_UNREACHABLE: 2716 case RDMA_CM_EVENT_REJECTED: 2717 /* These only occur on the client side. */ 2718 break; 2719 case RDMA_CM_EVENT_ESTABLISHED: 2720 /* TODO: Should we be waiting for this event anywhere? */ 2721 break; 2722 case RDMA_CM_EVENT_DISCONNECTED: 2723 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2724 rc = nvmf_rdma_disconnect(event); 2725 if (rc < 0) { 2726 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2727 break; 2728 } 2729 break; 2730 case RDMA_CM_EVENT_MULTICAST_JOIN: 2731 case RDMA_CM_EVENT_MULTICAST_ERROR: 2732 /* Multicast is not used */ 2733 break; 2734 case RDMA_CM_EVENT_ADDR_CHANGE: 2735 /* Not utilizing this event */ 2736 break; 2737 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2738 /* For now, do nothing. The target never re-uses queue pairs. */ 2739 break; 2740 default: 2741 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2742 break; 2743 } 2744 2745 rdma_ack_cm_event(event); 2746 } else { 2747 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2748 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2749 } 2750 break; 2751 } 2752 } 2753 } 2754 2755 static void 2756 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2757 { 2758 int rc; 2759 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 2760 struct ibv_async_event event; 2761 enum ibv_qp_state state; 2762 2763 rc = ibv_get_async_event(device->context, &event); 2764 2765 if (rc) { 2766 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2767 errno, spdk_strerror(errno)); 2768 return; 2769 } 2770 2771 SPDK_NOTICELOG("Async event: %s\n", 2772 ibv_event_type_str(event.event_type)); 2773 2774 switch (event.event_type) { 2775 case IBV_EVENT_QP_FATAL: 2776 rqpair = event.element.qp->qp_context; 2777 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2778 (uintptr_t)rqpair->cm_id, event.event_type); 2779 spdk_nvmf_rdma_update_ibv_state(rqpair); 2780 spdk_nvmf_rdma_start_disconnect(rqpair); 2781 break; 2782 case IBV_EVENT_QP_LAST_WQE_REACHED: 2783 /* This event only occurs for shared receive queues. */ 2784 rqpair = event.element.qp->qp_context; 2785 rqpair->last_wqe_reached = true; 2786 2787 /* This must be handled on the polling thread if it exists. Otherwise the timeout will catch it. */ 2788 if (rqpair->qpair.group) { 2789 spdk_thread_send_msg(rqpair->qpair.group->thread, nvmf_rdma_destroy_drained_qpair, rqpair); 2790 } else { 2791 SPDK_ERRLOG("Unable to destroy the qpair %p since it does not have a poll group.\n", rqpair); 2792 } 2793 2794 break; 2795 case IBV_EVENT_SQ_DRAINED: 2796 /* This event occurs frequently in both error and non-error states. 2797 * Check if the qpair is in an error state before sending a message. 2798 * Note that we're not on the correct thread to access the qpair, but 2799 * the operations that the below calls make all happen to be thread 2800 * safe. */ 2801 rqpair = event.element.qp->qp_context; 2802 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2803 (uintptr_t)rqpair->cm_id, event.event_type); 2804 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2805 if (state == IBV_QPS_ERR) { 2806 spdk_nvmf_rdma_start_disconnect(rqpair); 2807 } 2808 break; 2809 case IBV_EVENT_QP_REQ_ERR: 2810 case IBV_EVENT_QP_ACCESS_ERR: 2811 case IBV_EVENT_COMM_EST: 2812 case IBV_EVENT_PATH_MIG: 2813 case IBV_EVENT_PATH_MIG_ERR: 2814 rqpair = event.element.qp->qp_context; 2815 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2816 (uintptr_t)rqpair->cm_id, event.event_type); 2817 spdk_nvmf_rdma_update_ibv_state(rqpair); 2818 break; 2819 case IBV_EVENT_CQ_ERR: 2820 case IBV_EVENT_DEVICE_FATAL: 2821 case IBV_EVENT_PORT_ACTIVE: 2822 case IBV_EVENT_PORT_ERR: 2823 case IBV_EVENT_LID_CHANGE: 2824 case IBV_EVENT_PKEY_CHANGE: 2825 case IBV_EVENT_SM_CHANGE: 2826 case IBV_EVENT_SRQ_ERR: 2827 case IBV_EVENT_SRQ_LIMIT_REACHED: 2828 case IBV_EVENT_CLIENT_REREGISTER: 2829 case IBV_EVENT_GID_CHANGE: 2830 default: 2831 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2832 break; 2833 } 2834 ibv_ack_async_event(&event); 2835 } 2836 2837 static void 2838 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2839 { 2840 int nfds, i = 0; 2841 struct spdk_nvmf_rdma_transport *rtransport; 2842 struct spdk_nvmf_rdma_device *device, *tmp; 2843 2844 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2845 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2846 2847 if (nfds <= 0) { 2848 return; 2849 } 2850 2851 /* The first poll descriptor is RDMA CM event */ 2852 if (rtransport->poll_fds[i++].revents & POLLIN) { 2853 spdk_nvmf_process_cm_event(transport, cb_fn); 2854 nfds--; 2855 } 2856 2857 if (nfds == 0) { 2858 return; 2859 } 2860 2861 /* Second and subsequent poll descriptors are IB async events */ 2862 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2863 if (rtransport->poll_fds[i++].revents & POLLIN) { 2864 spdk_nvmf_process_ib_event(device); 2865 nfds--; 2866 } 2867 } 2868 /* check all flagged fd's have been served */ 2869 assert(nfds == 0); 2870 } 2871 2872 static void 2873 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2874 struct spdk_nvme_transport_id *trid, 2875 struct spdk_nvmf_discovery_log_page_entry *entry) 2876 { 2877 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2878 entry->adrfam = trid->adrfam; 2879 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2880 2881 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2882 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2883 2884 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2885 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2886 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2887 } 2888 2889 static void 2890 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 2891 2892 static struct spdk_nvmf_transport_poll_group * 2893 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2894 { 2895 struct spdk_nvmf_rdma_transport *rtransport; 2896 struct spdk_nvmf_rdma_poll_group *rgroup; 2897 struct spdk_nvmf_rdma_poller *poller; 2898 struct spdk_nvmf_rdma_device *device; 2899 struct ibv_srq_init_attr srq_init_attr; 2900 struct spdk_nvmf_rdma_resource_opts opts; 2901 int num_cqe; 2902 2903 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2904 2905 rgroup = calloc(1, sizeof(*rgroup)); 2906 if (!rgroup) { 2907 return NULL; 2908 } 2909 2910 TAILQ_INIT(&rgroup->pollers); 2911 STAILQ_INIT(&rgroup->pending_data_buf_queue); 2912 2913 pthread_mutex_lock(&rtransport->lock); 2914 TAILQ_FOREACH(device, &rtransport->devices, link) { 2915 poller = calloc(1, sizeof(*poller)); 2916 if (!poller) { 2917 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2918 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2919 pthread_mutex_unlock(&rtransport->lock); 2920 return NULL; 2921 } 2922 2923 poller->device = device; 2924 poller->group = rgroup; 2925 2926 TAILQ_INIT(&poller->qpairs); 2927 2928 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2929 if (device->attr.max_srq != 0) { 2930 poller->max_srq_depth = transport->opts.max_srq_depth; 2931 2932 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 2933 srq_init_attr.attr.max_wr = poller->max_srq_depth; 2934 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 2935 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 2936 if (!poller->srq) { 2937 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 2938 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2939 pthread_mutex_unlock(&rtransport->lock); 2940 return NULL; 2941 } 2942 2943 opts.qp = poller->srq; 2944 opts.pd = device->pd; 2945 opts.qpair = NULL; 2946 opts.shared = true; 2947 opts.max_queue_depth = poller->max_srq_depth; 2948 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 2949 2950 poller->resources = nvmf_rdma_resources_create(&opts); 2951 if (!poller->resources) { 2952 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 2953 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2954 pthread_mutex_unlock(&rtransport->lock); 2955 } 2956 } 2957 2958 /* 2959 * When using an srq, we can limit the completion queue at startup. 2960 * The following formula represents the calculation: 2961 * num_cqe = num_recv + num_data_wr + num_send_wr. 2962 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 2963 */ 2964 if (poller->srq) { 2965 num_cqe = poller->max_srq_depth * 3; 2966 } else { 2967 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2968 } 2969 2970 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 2971 if (!poller->cq) { 2972 SPDK_ERRLOG("Unable to create completion queue\n"); 2973 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2974 pthread_mutex_unlock(&rtransport->lock); 2975 return NULL; 2976 } 2977 poller->num_cqe = num_cqe; 2978 } 2979 2980 pthread_mutex_unlock(&rtransport->lock); 2981 return &rgroup->group; 2982 } 2983 2984 static void 2985 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 2986 { 2987 struct spdk_nvmf_rdma_poll_group *rgroup; 2988 struct spdk_nvmf_rdma_poller *poller, *tmp; 2989 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 2990 2991 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 2992 2993 if (!rgroup) { 2994 return; 2995 } 2996 2997 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 2998 TAILQ_REMOVE(&rgroup->pollers, poller, link); 2999 3000 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3001 spdk_nvmf_rdma_qpair_destroy(qpair); 3002 } 3003 3004 if (poller->srq) { 3005 nvmf_rdma_resources_destroy(poller->resources); 3006 ibv_destroy_srq(poller->srq); 3007 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3008 } 3009 3010 if (poller->cq) { 3011 ibv_destroy_cq(poller->cq); 3012 } 3013 3014 free(poller); 3015 } 3016 3017 if (!STAILQ_EMPTY(&rgroup->pending_data_buf_queue)) { 3018 SPDK_ERRLOG("Pending I/O list wasn't empty on poll group destruction\n"); 3019 } 3020 3021 free(rgroup); 3022 } 3023 3024 static void 3025 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3026 { 3027 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3028 spdk_nvmf_rdma_qpair_destroy(rqpair); 3029 } 3030 3031 static int 3032 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3033 struct spdk_nvmf_qpair *qpair) 3034 { 3035 struct spdk_nvmf_rdma_poll_group *rgroup; 3036 struct spdk_nvmf_rdma_qpair *rqpair; 3037 struct spdk_nvmf_rdma_device *device; 3038 struct spdk_nvmf_rdma_poller *poller; 3039 int rc; 3040 3041 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3042 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3043 3044 device = rqpair->port->device; 3045 3046 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3047 if (poller->device == device) { 3048 break; 3049 } 3050 } 3051 3052 if (!poller) { 3053 SPDK_ERRLOG("No poller found for device.\n"); 3054 return -1; 3055 } 3056 3057 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3058 rqpair->poller = poller; 3059 rqpair->srq = rqpair->poller->srq; 3060 3061 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 3062 if (rc < 0) { 3063 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3064 return -1; 3065 } 3066 3067 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3068 if (rc) { 3069 /* Try to reject, but we probably can't */ 3070 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3071 return -1; 3072 } 3073 3074 spdk_nvmf_rdma_update_ibv_state(rqpair); 3075 3076 return 0; 3077 } 3078 3079 static int 3080 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3081 { 3082 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3083 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3084 struct spdk_nvmf_rdma_transport, transport); 3085 3086 nvmf_rdma_request_free(rdma_req, rtransport); 3087 return 0; 3088 } 3089 3090 static int 3091 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3092 { 3093 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3094 struct spdk_nvmf_rdma_transport, transport); 3095 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3096 struct spdk_nvmf_rdma_request, req); 3097 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3098 struct spdk_nvmf_rdma_qpair, qpair); 3099 3100 if (rqpair->ibv_state != IBV_QPS_ERR) { 3101 /* The connection is alive, so process the request as normal */ 3102 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3103 } else { 3104 /* The connection is dead. Move the request directly to the completed state. */ 3105 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3106 } 3107 3108 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3109 3110 return 0; 3111 } 3112 3113 static int 3114 spdk_nvmf_rdma_destroy_defunct_qpair(void *ctx) 3115 { 3116 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 3117 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3118 struct spdk_nvmf_rdma_transport, transport); 3119 3120 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3121 spdk_nvmf_rdma_qpair_destroy(rqpair); 3122 3123 return 0; 3124 } 3125 3126 static void 3127 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3128 { 3129 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3130 3131 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 3132 return; 3133 } 3134 3135 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 3136 3137 /* This happens only when the qpair is disconnected before 3138 * it is added to the poll group. Since there is no poll group, 3139 * the RDMA qp has not been initialized yet and the RDMA CM 3140 * event has not yet been acknowledged, so we need to reject it. 3141 */ 3142 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3143 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3144 return; 3145 } 3146 3147 if (rqpair->ibv_state != IBV_QPS_ERR) { 3148 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 3149 } 3150 3151 rqpair->destruct_poller = spdk_poller_register(spdk_nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 3152 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 3153 } 3154 3155 static struct spdk_nvmf_rdma_qpair * 3156 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3157 { 3158 struct spdk_nvmf_rdma_qpair *rqpair; 3159 /* @todo: improve QP search */ 3160 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3161 if (wc->qp_num == rqpair->cm_id->qp->qp_num) { 3162 return rqpair; 3163 } 3164 } 3165 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3166 return NULL; 3167 } 3168 3169 #ifdef DEBUG 3170 static int 3171 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3172 { 3173 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3174 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3175 } 3176 #endif 3177 3178 static int 3179 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3180 struct spdk_nvmf_rdma_poller *rpoller) 3181 { 3182 struct ibv_wc wc[32]; 3183 struct spdk_nvmf_rdma_wr *rdma_wr; 3184 struct spdk_nvmf_rdma_request *rdma_req; 3185 struct spdk_nvmf_rdma_recv *rdma_recv; 3186 struct spdk_nvmf_rdma_qpair *rqpair; 3187 int reaped, i; 3188 int count = 0; 3189 bool error = false; 3190 3191 /* Poll for completing operations. */ 3192 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3193 if (reaped < 0) { 3194 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3195 errno, spdk_strerror(errno)); 3196 return -1; 3197 } 3198 3199 for (i = 0; i < reaped; i++) { 3200 3201 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3202 3203 switch (rdma_wr->type) { 3204 case RDMA_WR_TYPE_SEND: 3205 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3206 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3207 3208 if (!wc[i].status) { 3209 count++; 3210 assert(wc[i].opcode == IBV_WC_SEND); 3211 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 3212 } else { 3213 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3214 } 3215 3216 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3217 rqpair->current_send_depth--; 3218 3219 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3220 assert(rdma_req->num_outstanding_data_wr == 0); 3221 break; 3222 case RDMA_WR_TYPE_RECV: 3223 /* rdma_recv->qpair will be NULL if using an SRQ. In that case we have to get the qpair from the wc. */ 3224 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3225 if (rdma_recv->qpair == NULL) { 3226 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3227 } 3228 rqpair = rdma_recv->qpair; 3229 3230 assert(rqpair != NULL); 3231 if (!wc[i].status) { 3232 assert(wc[i].opcode == IBV_WC_RECV); 3233 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3234 spdk_nvmf_rdma_start_disconnect(rqpair); 3235 break; 3236 } 3237 } 3238 3239 rqpair->current_recv_depth++; 3240 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3241 break; 3242 case RDMA_WR_TYPE_DATA: 3243 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3244 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3245 3246 assert(rdma_req->num_outstanding_data_wr > 0); 3247 3248 rqpair->current_send_depth--; 3249 rdma_req->num_outstanding_data_wr--; 3250 if (!wc[i].status) { 3251 if (wc[i].opcode == IBV_WC_RDMA_READ) { 3252 rqpair->current_read_depth--; 3253 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3254 if (rdma_req->num_outstanding_data_wr == 0) { 3255 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3256 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3257 } 3258 } else { 3259 assert(wc[i].opcode == IBV_WC_RDMA_WRITE); 3260 } 3261 } else { 3262 /* If the data transfer fails still force the queue into the error state, 3263 * if we were performing an RDMA_READ, we need to force the request into a 3264 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3265 * case, we should wait for the SEND to complete. */ 3266 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3267 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3268 rqpair->current_read_depth--; 3269 if (rdma_req->num_outstanding_data_wr == 0) { 3270 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3271 } 3272 } 3273 } 3274 break; 3275 default: 3276 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3277 continue; 3278 } 3279 3280 /* Handle error conditions */ 3281 if (wc[i].status) { 3282 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 3283 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 3284 3285 error = true; 3286 3287 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3288 /* Disconnect the connection. */ 3289 spdk_nvmf_rdma_start_disconnect(rqpair); 3290 } else { 3291 nvmf_rdma_destroy_drained_qpair(rqpair); 3292 } 3293 continue; 3294 } 3295 3296 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3297 3298 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3299 nvmf_rdma_destroy_drained_qpair(rqpair); 3300 } 3301 } 3302 3303 if (error == true) { 3304 return -1; 3305 } 3306 3307 return count; 3308 } 3309 3310 static int 3311 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3312 { 3313 struct spdk_nvmf_rdma_transport *rtransport; 3314 struct spdk_nvmf_rdma_poll_group *rgroup; 3315 struct spdk_nvmf_rdma_poller *rpoller; 3316 int count, rc; 3317 3318 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3319 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3320 3321 count = 0; 3322 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3323 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 3324 if (rc < 0) { 3325 return rc; 3326 } 3327 count += rc; 3328 } 3329 3330 return count; 3331 } 3332 3333 static int 3334 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3335 struct spdk_nvme_transport_id *trid, 3336 bool peer) 3337 { 3338 struct sockaddr *saddr; 3339 uint16_t port; 3340 3341 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 3342 3343 if (peer) { 3344 saddr = rdma_get_peer_addr(id); 3345 } else { 3346 saddr = rdma_get_local_addr(id); 3347 } 3348 switch (saddr->sa_family) { 3349 case AF_INET: { 3350 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3351 3352 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3353 inet_ntop(AF_INET, &saddr_in->sin_addr, 3354 trid->traddr, sizeof(trid->traddr)); 3355 if (peer) { 3356 port = ntohs(rdma_get_dst_port(id)); 3357 } else { 3358 port = ntohs(rdma_get_src_port(id)); 3359 } 3360 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3361 break; 3362 } 3363 case AF_INET6: { 3364 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3365 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3366 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3367 trid->traddr, sizeof(trid->traddr)); 3368 if (peer) { 3369 port = ntohs(rdma_get_dst_port(id)); 3370 } else { 3371 port = ntohs(rdma_get_src_port(id)); 3372 } 3373 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3374 break; 3375 } 3376 default: 3377 return -1; 3378 3379 } 3380 3381 return 0; 3382 } 3383 3384 static int 3385 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3386 struct spdk_nvme_transport_id *trid) 3387 { 3388 struct spdk_nvmf_rdma_qpair *rqpair; 3389 3390 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3391 3392 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 3393 } 3394 3395 static int 3396 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3397 struct spdk_nvme_transport_id *trid) 3398 { 3399 struct spdk_nvmf_rdma_qpair *rqpair; 3400 3401 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3402 3403 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 3404 } 3405 3406 static int 3407 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3408 struct spdk_nvme_transport_id *trid) 3409 { 3410 struct spdk_nvmf_rdma_qpair *rqpair; 3411 3412 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3413 3414 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 3415 } 3416 3417 void 3418 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3419 { 3420 g_nvmf_hooks = *hooks; 3421 } 3422 3423 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 3424 .type = SPDK_NVME_TRANSPORT_RDMA, 3425 .opts_init = spdk_nvmf_rdma_opts_init, 3426 .create = spdk_nvmf_rdma_create, 3427 .destroy = spdk_nvmf_rdma_destroy, 3428 3429 .listen = spdk_nvmf_rdma_listen, 3430 .stop_listen = spdk_nvmf_rdma_stop_listen, 3431 .accept = spdk_nvmf_rdma_accept, 3432 3433 .listener_discover = spdk_nvmf_rdma_discover, 3434 3435 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 3436 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 3437 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 3438 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 3439 3440 .req_free = spdk_nvmf_rdma_request_free, 3441 .req_complete = spdk_nvmf_rdma_request_complete, 3442 3443 .qpair_fini = spdk_nvmf_rdma_close_qpair, 3444 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 3445 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 3446 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 3447 3448 }; 3449 3450 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 3451