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 rdma_recv->qpair = NULL; 872 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 873 if (rc) { 874 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 875 } 876 } 877 } 878 } 879 } 880 881 if (rqpair->cm_id) { 882 if (rqpair->cm_id->qp != NULL) { 883 rdma_destroy_qp(rqpair->cm_id); 884 } 885 rdma_destroy_id(rqpair->cm_id); 886 887 if (rqpair->poller != NULL && rqpair->srq == NULL) { 888 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 889 } 890 } 891 892 if (rqpair->srq == NULL && rqpair->resources != NULL) { 893 nvmf_rdma_resources_destroy(rqpair->resources); 894 } 895 896 free(rqpair); 897 } 898 899 static int 900 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 901 { 902 struct spdk_nvmf_rdma_poller *rpoller; 903 int rc, num_cqe, required_num_wr; 904 905 /* Enlarge CQ size dynamically */ 906 rpoller = rqpair->poller; 907 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 908 num_cqe = rpoller->num_cqe; 909 if (num_cqe < required_num_wr) { 910 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 911 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 912 } 913 914 if (rpoller->num_cqe != num_cqe) { 915 if (required_num_wr > device->attr.max_cqe) { 916 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 917 required_num_wr, device->attr.max_cqe); 918 return -1; 919 } 920 921 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 922 rc = ibv_resize_cq(rpoller->cq, num_cqe); 923 if (rc) { 924 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 925 return -1; 926 } 927 928 rpoller->num_cqe = num_cqe; 929 } 930 931 rpoller->required_num_wr = required_num_wr; 932 return 0; 933 } 934 935 static int 936 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 937 { 938 struct spdk_nvmf_rdma_qpair *rqpair; 939 int rc; 940 struct spdk_nvmf_rdma_transport *rtransport; 941 struct spdk_nvmf_transport *transport; 942 struct spdk_nvmf_rdma_resource_opts opts; 943 struct spdk_nvmf_rdma_device *device; 944 struct ibv_qp_init_attr ibv_init_attr; 945 946 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 947 device = rqpair->port->device; 948 949 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 950 ibv_init_attr.qp_context = rqpair; 951 ibv_init_attr.qp_type = IBV_QPT_RC; 952 ibv_init_attr.send_cq = rqpair->poller->cq; 953 ibv_init_attr.recv_cq = rqpair->poller->cq; 954 955 if (rqpair->srq) { 956 ibv_init_attr.srq = rqpair->srq; 957 } else { 958 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 959 1; /* RECV operations + dummy drain WR */ 960 } 961 962 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 963 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 964 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 965 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 966 967 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 968 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 969 goto error; 970 } 971 972 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 973 if (rc) { 974 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 975 goto error; 976 } 977 978 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2 + 1), 979 ibv_init_attr.cap.max_send_wr); 980 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 981 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 982 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 983 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 984 985 if (rqpair->poller->srq == NULL) { 986 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 987 transport = &rtransport->transport; 988 989 opts.qp = rqpair->cm_id->qp; 990 opts.pd = rqpair->cm_id->pd; 991 opts.qpair = rqpair; 992 opts.shared = false; 993 opts.max_queue_depth = rqpair->max_queue_depth; 994 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 995 996 rqpair->resources = nvmf_rdma_resources_create(&opts); 997 998 if (!rqpair->resources) { 999 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1000 goto error; 1001 } 1002 } else { 1003 rqpair->resources = rqpair->poller->resources; 1004 } 1005 1006 rqpair->current_recv_depth = 0; 1007 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1008 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1009 1010 return 0; 1011 1012 error: 1013 rdma_destroy_id(rqpair->cm_id); 1014 rqpair->cm_id = NULL; 1015 spdk_nvmf_rdma_qpair_destroy(rqpair); 1016 return -1; 1017 } 1018 1019 static int 1020 request_transfer_in(struct spdk_nvmf_request *req) 1021 { 1022 int rc; 1023 struct spdk_nvmf_rdma_request *rdma_req; 1024 struct spdk_nvmf_qpair *qpair; 1025 struct spdk_nvmf_rdma_qpair *rqpair; 1026 struct ibv_send_wr *bad_wr = NULL; 1027 1028 qpair = req->qpair; 1029 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1030 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1031 1032 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1033 assert(rdma_req != NULL); 1034 1035 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA READ POSTED. Request: %p Connection: %p\n", req, qpair); 1036 1037 rc = ibv_post_send(rqpair->cm_id->qp, &rdma_req->data.wr, &bad_wr); 1038 if (rc) { 1039 SPDK_ERRLOG("Unable to transfer data from host to target\n"); 1040 return -1; 1041 } 1042 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1043 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1044 return 0; 1045 } 1046 1047 static int 1048 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1049 { 1050 int rc; 1051 int num_outstanding_data_wr = 0; 1052 struct spdk_nvmf_rdma_request *rdma_req; 1053 struct spdk_nvmf_qpair *qpair; 1054 struct spdk_nvmf_rdma_qpair *rqpair; 1055 struct spdk_nvme_cpl *rsp; 1056 struct ibv_recv_wr *bad_recv_wr = NULL; 1057 struct ibv_send_wr *send_wr, *bad_send_wr = NULL; 1058 1059 *data_posted = 0; 1060 qpair = req->qpair; 1061 rsp = &req->rsp->nvme_cpl; 1062 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1063 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1064 1065 /* Advance our sq_head pointer */ 1066 if (qpair->sq_head == qpair->sq_head_max) { 1067 qpair->sq_head = 0; 1068 } else { 1069 qpair->sq_head++; 1070 } 1071 rsp->sqhd = qpair->sq_head; 1072 1073 /* Post the capsule to the recv buffer */ 1074 assert(rdma_req->recv != NULL); 1075 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA RECV POSTED. Recv: %p Connection: %p\n", rdma_req->recv, 1076 rqpair); 1077 if (rqpair->srq == NULL) { 1078 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_req->recv->wr, &bad_recv_wr); 1079 } else { 1080 rdma_req->recv->qpair = NULL; 1081 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 1082 } 1083 1084 if (rc) { 1085 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 1086 return rc; 1087 } 1088 rdma_req->recv = NULL; 1089 assert(rqpair->current_recv_depth > 0); 1090 rqpair->current_recv_depth--; 1091 1092 /* Build the response which consists of optional 1093 * RDMA WRITEs to transfer data, plus an RDMA SEND 1094 * containing the response. 1095 */ 1096 send_wr = &rdma_req->rsp.wr; 1097 1098 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 1099 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1100 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA WRITE POSTED. Request: %p Connection: %p\n", req, qpair); 1101 send_wr = &rdma_req->data.wr; 1102 *data_posted = 1; 1103 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1104 } 1105 1106 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA SEND POSTED. Request: %p Connection: %p\n", req, qpair); 1107 1108 /* Send the completion */ 1109 rc = ibv_post_send(rqpair->cm_id->qp, send_wr, &bad_send_wr); 1110 if (rc) { 1111 SPDK_ERRLOG("Unable to send response capsule\n"); 1112 return rc; 1113 } 1114 /* +1 for the rsp wr */ 1115 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1116 1117 return 0; 1118 } 1119 1120 static int 1121 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1122 { 1123 struct spdk_nvmf_rdma_accept_private_data accept_data; 1124 struct rdma_conn_param ctrlr_event_data = {}; 1125 int rc; 1126 1127 accept_data.recfmt = 0; 1128 accept_data.crqsize = rqpair->max_queue_depth; 1129 1130 ctrlr_event_data.private_data = &accept_data; 1131 ctrlr_event_data.private_data_len = sizeof(accept_data); 1132 if (id->ps == RDMA_PS_TCP) { 1133 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1134 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1135 } 1136 1137 /* Configure infinite retries for the initiator side qpair. 1138 * When using a shared receive queue on the target side, 1139 * we need to pass this value to the initiator to prevent the 1140 * initiator side NIC from completing SEND requests back to the 1141 * initiator with status rnr_retry_count_exceeded. */ 1142 if (rqpair->srq != NULL) { 1143 ctrlr_event_data.rnr_retry_count = 0x7; 1144 } 1145 1146 rc = rdma_accept(id, &ctrlr_event_data); 1147 if (rc) { 1148 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 1149 } else { 1150 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1151 } 1152 1153 return rc; 1154 } 1155 1156 static void 1157 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1158 { 1159 struct spdk_nvmf_rdma_reject_private_data rej_data; 1160 1161 rej_data.recfmt = 0; 1162 rej_data.sts = error; 1163 1164 rdma_reject(id, &rej_data, sizeof(rej_data)); 1165 } 1166 1167 static int 1168 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 1169 new_qpair_fn cb_fn) 1170 { 1171 struct spdk_nvmf_rdma_transport *rtransport; 1172 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1173 struct spdk_nvmf_rdma_port *port; 1174 struct rdma_conn_param *rdma_param = NULL; 1175 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1176 uint16_t max_queue_depth; 1177 uint16_t max_read_depth; 1178 1179 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1180 1181 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1182 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1183 1184 rdma_param = &event->param.conn; 1185 if (rdma_param->private_data == NULL || 1186 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1187 SPDK_ERRLOG("connect request: no private data provided\n"); 1188 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1189 return -1; 1190 } 1191 1192 private_data = rdma_param->private_data; 1193 if (private_data->recfmt != 0) { 1194 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1195 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1196 return -1; 1197 } 1198 1199 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1200 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1201 1202 port = event->listen_id->context; 1203 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1204 event->listen_id, event->listen_id->verbs, port); 1205 1206 /* Figure out the supported queue depth. This is a multi-step process 1207 * that takes into account hardware maximums, host provided values, 1208 * and our target's internal memory limits */ 1209 1210 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1211 1212 /* Start with the maximum queue depth allowed by the target */ 1213 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1214 max_read_depth = rtransport->transport.opts.max_queue_depth; 1215 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1216 rtransport->transport.opts.max_queue_depth); 1217 1218 /* Next check the local NIC's hardware limitations */ 1219 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1220 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1221 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1222 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1223 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1224 1225 /* Next check the remote NIC's hardware limitations */ 1226 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1227 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1228 rdma_param->initiator_depth, rdma_param->responder_resources); 1229 if (rdma_param->initiator_depth > 0) { 1230 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1231 } 1232 1233 /* Finally check for the host software requested values, which are 1234 * optional. */ 1235 if (rdma_param->private_data != NULL && 1236 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1237 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1238 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1239 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1240 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1241 } 1242 1243 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1244 max_queue_depth, max_read_depth); 1245 1246 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1247 if (rqpair == NULL) { 1248 SPDK_ERRLOG("Could not allocate new connection.\n"); 1249 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1250 return -1; 1251 } 1252 1253 rqpair->port = port; 1254 rqpair->max_queue_depth = max_queue_depth; 1255 rqpair->max_read_depth = max_read_depth; 1256 rqpair->cm_id = event->id; 1257 rqpair->listen_id = event->listen_id; 1258 rqpair->qpair.transport = transport; 1259 1260 event->id->context = &rqpair->qpair; 1261 1262 cb_fn(&rqpair->qpair); 1263 1264 return 0; 1265 } 1266 1267 static int 1268 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1269 enum spdk_mem_map_notify_action action, 1270 void *vaddr, size_t size) 1271 { 1272 struct ibv_pd *pd = cb_ctx; 1273 struct ibv_mr *mr; 1274 1275 switch (action) { 1276 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1277 if (!g_nvmf_hooks.get_rkey) { 1278 mr = ibv_reg_mr(pd, vaddr, size, 1279 IBV_ACCESS_LOCAL_WRITE | 1280 IBV_ACCESS_REMOTE_READ | 1281 IBV_ACCESS_REMOTE_WRITE); 1282 if (mr == NULL) { 1283 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1284 return -1; 1285 } else { 1286 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1287 } 1288 } else { 1289 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1290 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1291 } 1292 break; 1293 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1294 if (!g_nvmf_hooks.get_rkey) { 1295 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1296 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1297 if (mr) { 1298 ibv_dereg_mr(mr); 1299 } 1300 } 1301 break; 1302 } 1303 1304 return 0; 1305 } 1306 1307 static int 1308 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1309 { 1310 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1311 return addr_1 == addr_2; 1312 } 1313 1314 static void 1315 spdk_nvmf_rdma_request_free_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1316 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1317 uint32_t num_buffers) 1318 { 1319 uint32_t i; 1320 1321 for (i = 0; i < num_buffers; i++) { 1322 if (group->buf_cache_count < group->buf_cache_size) { 1323 STAILQ_INSERT_HEAD(&group->buf_cache, 1324 (struct spdk_nvmf_transport_pg_cache_buf *)rdma_req->buffers[i], link); 1325 group->buf_cache_count++; 1326 } else { 1327 spdk_mempool_put(transport->data_buf_pool, rdma_req->buffers[i]); 1328 } 1329 rdma_req->req.iov[i].iov_base = NULL; 1330 rdma_req->buffers[i] = NULL; 1331 rdma_req->req.iov[i].iov_len = 0; 1332 1333 } 1334 rdma_req->data_from_pool = false; 1335 } 1336 1337 static int 1338 nvmf_rdma_request_get_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1339 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1340 uint32_t num_buffers) 1341 { 1342 uint32_t i = 0; 1343 1344 while (i < num_buffers) { 1345 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1346 group->buf_cache_count--; 1347 rdma_req->buffers[i] = STAILQ_FIRST(&group->buf_cache); 1348 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1349 assert(rdma_req->buffers[i] != NULL); 1350 i++; 1351 } else { 1352 if (spdk_mempool_get_bulk(transport->data_buf_pool, &rdma_req->buffers[i], num_buffers - i)) { 1353 goto err_exit; 1354 } 1355 i += num_buffers - i; 1356 } 1357 } 1358 1359 return 0; 1360 1361 err_exit: 1362 spdk_nvmf_rdma_request_free_buffers(rdma_req, group, transport, i); 1363 return -ENOMEM; 1364 } 1365 1366 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1367 1368 static spdk_nvme_data_transfer_t 1369 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1370 { 1371 enum spdk_nvme_data_transfer xfer; 1372 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1373 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1374 1375 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1376 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1377 rdma_req->rsp.wr.imm_data = 0; 1378 #endif 1379 1380 /* Figure out data transfer direction */ 1381 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1382 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1383 } else { 1384 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1385 1386 /* Some admin commands are special cases */ 1387 if ((rdma_req->req.qpair->qid == 0) && 1388 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1389 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1390 switch (cmd->cdw10 & 0xff) { 1391 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1392 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1393 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1394 break; 1395 default: 1396 xfer = SPDK_NVME_DATA_NONE; 1397 } 1398 } 1399 } 1400 1401 if (xfer == SPDK_NVME_DATA_NONE) { 1402 return xfer; 1403 } 1404 1405 /* Even for commands that may transfer data, they could have specified 0 length. 1406 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1407 */ 1408 switch (sgl->generic.type) { 1409 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1410 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1411 case SPDK_NVME_SGL_TYPE_SEGMENT: 1412 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1413 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1414 if (sgl->unkeyed.length == 0) { 1415 xfer = SPDK_NVME_DATA_NONE; 1416 } 1417 break; 1418 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1419 if (sgl->keyed.length == 0) { 1420 xfer = SPDK_NVME_DATA_NONE; 1421 } 1422 break; 1423 } 1424 1425 return xfer; 1426 } 1427 1428 static int 1429 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1430 struct spdk_nvmf_rdma_request *rdma_req, 1431 uint32_t num_sgl_descriptors) 1432 { 1433 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1434 struct spdk_nvmf_rdma_request_data *current_data_wr; 1435 uint32_t i; 1436 1437 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1438 return -ENOMEM; 1439 } 1440 1441 current_data_wr = &rdma_req->data; 1442 1443 for (i = 0; i < num_sgl_descriptors; i++) { 1444 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1445 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1446 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1447 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1448 } else { 1449 assert(false); 1450 } 1451 work_requests[i]->wr.send_flags = IBV_SEND_SIGNALED; 1452 work_requests[i]->wr.sg_list = work_requests[i]->sgl; 1453 work_requests[i]->wr.wr_id = rdma_req->data.wr.wr_id; 1454 current_data_wr->wr.next = &work_requests[i]->wr; 1455 current_data_wr = work_requests[i]; 1456 } 1457 1458 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1459 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1460 current_data_wr->wr.next = &rdma_req->rsp.wr; 1461 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1462 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1463 current_data_wr->wr.next = NULL; 1464 } 1465 return 0; 1466 } 1467 1468 static int 1469 nvmf_rdma_fill_buffers(struct spdk_nvmf_rdma_transport *rtransport, 1470 struct spdk_nvmf_rdma_poll_group *rgroup, 1471 struct spdk_nvmf_rdma_device *device, 1472 struct spdk_nvmf_rdma_request *rdma_req, 1473 struct ibv_send_wr *wr, 1474 uint32_t length) 1475 { 1476 uint64_t translation_len; 1477 uint32_t remaining_length = length; 1478 uint32_t iovcnt; 1479 uint32_t i = 0; 1480 1481 1482 while (remaining_length) { 1483 iovcnt = rdma_req->req.iovcnt; 1484 rdma_req->req.iov[iovcnt].iov_base = (void *)((uintptr_t)(rdma_req->buffers[iovcnt] + 1485 NVMF_DATA_BUFFER_MASK) & 1486 ~NVMF_DATA_BUFFER_MASK); 1487 rdma_req->req.iov[iovcnt].iov_len = spdk_min(remaining_length, 1488 rtransport->transport.opts.io_unit_size); 1489 rdma_req->req.iovcnt++; 1490 wr->sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[iovcnt].iov_base); 1491 wr->sg_list[i].length = rdma_req->req.iov[iovcnt].iov_len; 1492 translation_len = rdma_req->req.iov[iovcnt].iov_len; 1493 1494 if (!g_nvmf_hooks.get_rkey) { 1495 wr->sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1496 (uint64_t)rdma_req->buffers[iovcnt], &translation_len))->lkey; 1497 } else { 1498 wr->sg_list[i].lkey = spdk_mem_map_translate(device->map, 1499 (uint64_t)rdma_req->buffers[iovcnt], &translation_len); 1500 } 1501 1502 remaining_length -= rdma_req->req.iov[iovcnt].iov_len; 1503 1504 if (translation_len < rdma_req->req.iov[iovcnt].iov_len) { 1505 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions\n"); 1506 return -EINVAL; 1507 } 1508 i++; 1509 } 1510 wr->num_sge = i; 1511 1512 return 0; 1513 } 1514 1515 static int 1516 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1517 struct spdk_nvmf_rdma_device *device, 1518 struct spdk_nvmf_rdma_request *rdma_req) 1519 { 1520 struct spdk_nvmf_rdma_qpair *rqpair; 1521 struct spdk_nvmf_rdma_poll_group *rgroup; 1522 uint32_t num_buffers; 1523 uint32_t i = 0; 1524 int rc = 0; 1525 1526 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1527 rgroup = rqpair->poller->group; 1528 rdma_req->req.iovcnt = 0; 1529 1530 num_buffers = rdma_req->req.length / rtransport->transport.opts.io_unit_size; 1531 if (rdma_req->req.length % rtransport->transport.opts.io_unit_size) { 1532 num_buffers++; 1533 } 1534 1535 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers)) { 1536 return -ENOMEM; 1537 } 1538 1539 rdma_req->req.iovcnt = 0; 1540 1541 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, &rdma_req->data.wr, 1542 rdma_req->req.length); 1543 if (rc != 0) { 1544 goto err_exit; 1545 } 1546 1547 assert(rdma_req->req.iovcnt <= rqpair->max_send_sge); 1548 1549 rdma_req->data_from_pool = true; 1550 1551 return rc; 1552 1553 err_exit: 1554 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1555 while (i) { 1556 i--; 1557 rdma_req->data.wr.sg_list[i].addr = 0; 1558 rdma_req->data.wr.sg_list[i].length = 0; 1559 rdma_req->data.wr.sg_list[i].lkey = 0; 1560 } 1561 rdma_req->req.iovcnt = 0; 1562 return rc; 1563 } 1564 1565 static int 1566 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1567 struct spdk_nvmf_rdma_device *device, 1568 struct spdk_nvmf_rdma_request *rdma_req) 1569 { 1570 struct spdk_nvmf_rdma_qpair *rqpair; 1571 struct spdk_nvmf_rdma_poll_group *rgroup; 1572 struct ibv_send_wr *current_wr; 1573 struct spdk_nvmf_request *req = &rdma_req->req; 1574 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1575 uint32_t num_sgl_descriptors; 1576 uint32_t num_buffers = 0; 1577 uint32_t i; 1578 int rc; 1579 1580 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1581 rgroup = rqpair->poller->group; 1582 1583 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1584 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1585 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1586 1587 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1588 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1589 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1590 1591 for (i = 0; i < num_sgl_descriptors; i++) { 1592 num_buffers += desc->keyed.length / rtransport->transport.opts.io_unit_size; 1593 if (desc->keyed.length % rtransport->transport.opts.io_unit_size) { 1594 num_buffers++; 1595 } 1596 desc++; 1597 } 1598 /* If the number of buffers is too large, then we know the I/O is larger than allowed. Fail it. */ 1599 if (num_buffers > NVMF_REQ_MAX_BUFFERS) { 1600 return -EINVAL; 1601 } 1602 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1603 num_buffers) != 0) { 1604 return -ENOMEM; 1605 } 1606 1607 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1608 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1609 return -ENOMEM; 1610 } 1611 1612 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1613 current_wr = &rdma_req->data.wr; 1614 1615 req->iovcnt = 0; 1616 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1617 for (i = 0; i < num_sgl_descriptors; i++) { 1618 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1619 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1620 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1621 rc = -EINVAL; 1622 goto err_exit; 1623 } 1624 1625 current_wr->num_sge = 0; 1626 req->length += desc->keyed.length; 1627 1628 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, current_wr, 1629 desc->keyed.length); 1630 if (rc != 0) { 1631 rc = -ENOMEM; 1632 goto err_exit; 1633 } 1634 1635 current_wr->wr.rdma.rkey = desc->keyed.key; 1636 current_wr->wr.rdma.remote_addr = desc->address; 1637 current_wr = current_wr->next; 1638 desc++; 1639 } 1640 1641 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1642 /* Go back to the last descriptor in the list. */ 1643 desc--; 1644 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1645 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1646 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1647 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1648 } 1649 } 1650 #endif 1651 1652 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1653 rdma_req->data_from_pool = true; 1654 1655 return 0; 1656 1657 err_exit: 1658 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1659 nvmf_rdma_request_free_data(rdma_req, rtransport); 1660 return rc; 1661 } 1662 1663 static int 1664 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1665 struct spdk_nvmf_rdma_device *device, 1666 struct spdk_nvmf_rdma_request *rdma_req) 1667 { 1668 struct spdk_nvme_cmd *cmd; 1669 struct spdk_nvme_cpl *rsp; 1670 struct spdk_nvme_sgl_descriptor *sgl; 1671 int rc; 1672 1673 cmd = &rdma_req->req.cmd->nvme_cmd; 1674 rsp = &rdma_req->req.rsp->nvme_cpl; 1675 sgl = &cmd->dptr.sgl1; 1676 1677 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1678 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1679 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1680 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1681 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1682 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1683 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1684 return -1; 1685 } 1686 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1687 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1688 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1689 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1690 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1691 } 1692 } 1693 #endif 1694 1695 /* fill request length and populate iovs */ 1696 rdma_req->req.length = sgl->keyed.length; 1697 1698 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1699 /* No available buffers. Queue this request up. */ 1700 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1701 return 0; 1702 } 1703 1704 /* backward compatible */ 1705 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1706 1707 /* rdma wr specifics */ 1708 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1709 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1710 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1711 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1712 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 1713 rdma_req->data.wr.next = &rdma_req->rsp.wr; 1714 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1715 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 1716 rdma_req->data.wr.next = NULL; 1717 } 1718 1719 /* set the number of outstanding data WRs for this request. */ 1720 rdma_req->num_outstanding_data_wr = 1; 1721 1722 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1723 rdma_req->req.iovcnt); 1724 1725 return 0; 1726 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1727 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1728 uint64_t offset = sgl->address; 1729 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1730 1731 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1732 offset, sgl->unkeyed.length); 1733 1734 if (offset > max_len) { 1735 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1736 offset, max_len); 1737 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1738 return -1; 1739 } 1740 max_len -= (uint32_t)offset; 1741 1742 if (sgl->unkeyed.length > max_len) { 1743 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1744 sgl->unkeyed.length, max_len); 1745 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1746 return -1; 1747 } 1748 1749 rdma_req->num_outstanding_data_wr = 0; 1750 rdma_req->req.data = rdma_req->recv->buf + offset; 1751 rdma_req->data_from_pool = false; 1752 rdma_req->req.length = sgl->unkeyed.length; 1753 1754 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1755 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1756 rdma_req->req.iovcnt = 1; 1757 1758 return 0; 1759 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1760 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1761 1762 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1763 if (rc == -ENOMEM) { 1764 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1765 return 0; 1766 } else if (rc == -EINVAL) { 1767 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1768 return -1; 1769 } 1770 1771 /* backward compatible */ 1772 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1773 1774 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1775 rdma_req->req.iovcnt); 1776 1777 return 0; 1778 } 1779 1780 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1781 sgl->generic.type, sgl->generic.subtype); 1782 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1783 return -1; 1784 } 1785 1786 static void 1787 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1788 struct spdk_nvmf_rdma_transport *rtransport) 1789 { 1790 struct spdk_nvmf_rdma_qpair *rqpair; 1791 struct spdk_nvmf_rdma_poll_group *rgroup; 1792 1793 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1794 if (rdma_req->data_from_pool) { 1795 rgroup = rqpair->poller->group; 1796 1797 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1798 rdma_req->req.iovcnt); 1799 } 1800 nvmf_rdma_request_free_data(rdma_req, rtransport); 1801 rdma_req->req.length = 0; 1802 rdma_req->req.iovcnt = 0; 1803 rdma_req->req.data = NULL; 1804 rdma_req->data.wr.next = NULL; 1805 rqpair->qd--; 1806 1807 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1808 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1809 } 1810 1811 static bool 1812 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1813 struct spdk_nvmf_rdma_request *rdma_req) 1814 { 1815 struct spdk_nvmf_rdma_qpair *rqpair; 1816 struct spdk_nvmf_rdma_device *device; 1817 struct spdk_nvmf_rdma_poll_group *rgroup; 1818 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1819 int rc; 1820 struct spdk_nvmf_rdma_recv *rdma_recv; 1821 enum spdk_nvmf_rdma_request_state prev_state; 1822 bool progress = false; 1823 int data_posted; 1824 1825 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1826 device = rqpair->port->device; 1827 rgroup = rqpair->poller->group; 1828 1829 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1830 1831 /* If the queue pair is in an error state, force the request to the completed state 1832 * to release resources. */ 1833 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1834 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1835 STAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1836 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1837 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1838 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1839 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1840 } 1841 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1842 } 1843 1844 /* The loop here is to allow for several back-to-back state changes. */ 1845 do { 1846 prev_state = rdma_req->state; 1847 1848 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1849 1850 switch (rdma_req->state) { 1851 case RDMA_REQUEST_STATE_FREE: 1852 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1853 * to escape this state. */ 1854 break; 1855 case RDMA_REQUEST_STATE_NEW: 1856 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1857 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1858 rdma_recv = rdma_req->recv; 1859 1860 /* The first element of the SGL is the NVMe command */ 1861 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1862 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1863 1864 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1865 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1866 break; 1867 } 1868 1869 /* The next state transition depends on the data transfer needs of this request. */ 1870 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1871 1872 /* If no data to transfer, ready to execute. */ 1873 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1874 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1875 break; 1876 } 1877 1878 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1879 STAILQ_INSERT_TAIL(&rgroup->pending_data_buf_queue, rdma_req, state_link); 1880 break; 1881 case RDMA_REQUEST_STATE_NEED_BUFFER: 1882 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1883 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1884 1885 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1886 1887 if (rdma_req != STAILQ_FIRST(&rgroup->pending_data_buf_queue)) { 1888 /* This request needs to wait in line to obtain a buffer */ 1889 break; 1890 } 1891 1892 /* Try to get a data buffer */ 1893 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1894 if (rc < 0) { 1895 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1896 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1897 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1898 break; 1899 } 1900 1901 if (!rdma_req->req.data) { 1902 /* No buffers available. */ 1903 break; 1904 } 1905 1906 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1907 1908 /* If data is transferring from host to controller and the data didn't 1909 * arrive using in capsule data, we need to do a transfer from the host. 1910 */ 1911 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1912 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1913 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1914 break; 1915 } 1916 1917 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1918 break; 1919 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1920 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1921 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1922 1923 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1924 /* This request needs to wait in line to perform RDMA */ 1925 break; 1926 } 1927 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1928 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1929 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1930 break; 1931 } 1932 1933 /* We have already verified that this request is the head of the queue. */ 1934 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 1935 1936 rc = request_transfer_in(&rdma_req->req); 1937 if (!rc) { 1938 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 1939 } else { 1940 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1941 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1942 } 1943 break; 1944 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1945 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1946 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1947 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1948 * to escape this state. */ 1949 break; 1950 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1951 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1952 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1953 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 1954 spdk_nvmf_request_exec(&rdma_req->req); 1955 break; 1956 case RDMA_REQUEST_STATE_EXECUTING: 1957 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1958 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1959 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1960 * to escape this state. */ 1961 break; 1962 case RDMA_REQUEST_STATE_EXECUTED: 1963 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1964 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1965 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1966 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 1967 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 1968 } else { 1969 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1970 } 1971 break; 1972 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 1973 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 1974 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1975 1976 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 1977 /* This request needs to wait in line to perform RDMA */ 1978 break; 1979 } 1980 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 1981 rqpair->max_send_depth) { 1982 /* We can only have so many WRs outstanding. we have to wait until some finish. 1983 * +1 since each request has an additional wr in the resp. */ 1984 break; 1985 } 1986 1987 /* We have already verified that this request is the head of the queue. */ 1988 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 1989 1990 /* The data transfer will be kicked off from 1991 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 1992 */ 1993 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1994 break; 1995 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 1996 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 1997 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1998 rc = request_transfer_out(&rdma_req->req, &data_posted); 1999 assert(rc == 0); /* No good way to handle this currently */ 2000 if (rc) { 2001 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2002 } else { 2003 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2004 RDMA_REQUEST_STATE_COMPLETING; 2005 } 2006 break; 2007 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2008 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2009 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2010 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2011 * to escape this state. */ 2012 break; 2013 case RDMA_REQUEST_STATE_COMPLETING: 2014 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2015 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2016 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2017 * to escape this state. */ 2018 break; 2019 case RDMA_REQUEST_STATE_COMPLETED: 2020 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2021 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2022 2023 nvmf_rdma_request_free(rdma_req, rtransport); 2024 break; 2025 case RDMA_REQUEST_NUM_STATES: 2026 default: 2027 assert(0); 2028 break; 2029 } 2030 2031 if (rdma_req->state != prev_state) { 2032 progress = true; 2033 } 2034 } while (rdma_req->state != prev_state); 2035 2036 return progress; 2037 } 2038 2039 /* Public API callbacks begin here */ 2040 2041 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2042 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2043 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2044 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 2045 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2046 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2047 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2048 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4096 2049 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2050 2051 static void 2052 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2053 { 2054 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2055 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2056 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2057 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2058 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2059 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2060 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2061 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2062 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2063 } 2064 2065 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2066 .notify_cb = spdk_nvmf_rdma_mem_notify, 2067 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 2068 }; 2069 2070 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2071 2072 static struct spdk_nvmf_transport * 2073 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2074 { 2075 int rc; 2076 struct spdk_nvmf_rdma_transport *rtransport; 2077 struct spdk_nvmf_rdma_device *device, *tmp; 2078 struct ibv_pd *pd; 2079 struct ibv_context **contexts; 2080 uint32_t i; 2081 int flag; 2082 uint32_t sge_count; 2083 uint32_t min_shared_buffers; 2084 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2085 2086 rtransport = calloc(1, sizeof(*rtransport)); 2087 if (!rtransport) { 2088 return NULL; 2089 } 2090 2091 if (pthread_mutex_init(&rtransport->lock, NULL)) { 2092 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2093 free(rtransport); 2094 return NULL; 2095 } 2096 2097 TAILQ_INIT(&rtransport->devices); 2098 TAILQ_INIT(&rtransport->ports); 2099 2100 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2101 2102 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2103 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2104 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2105 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2106 " num_shared_buffers=%d, max_srq_depth=%d\n", 2107 opts->max_queue_depth, 2108 opts->max_io_size, 2109 opts->max_qpairs_per_ctrlr, 2110 opts->io_unit_size, 2111 opts->in_capsule_data_size, 2112 opts->max_aq_depth, 2113 opts->num_shared_buffers, 2114 opts->max_srq_depth); 2115 2116 /* I/O unit size cannot be larger than max I/O size */ 2117 if (opts->io_unit_size > opts->max_io_size) { 2118 opts->io_unit_size = opts->max_io_size; 2119 } 2120 2121 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2122 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2123 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2124 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2125 spdk_nvmf_rdma_destroy(&rtransport->transport); 2126 return NULL; 2127 } 2128 2129 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2130 if (min_shared_buffers > opts->num_shared_buffers) { 2131 SPDK_ERRLOG("There are not enough buffers to satisfy" 2132 "per-poll group caches for each thread. (%" PRIu32 ")" 2133 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2134 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2135 spdk_nvmf_rdma_destroy(&rtransport->transport); 2136 return NULL; 2137 } 2138 2139 sge_count = opts->max_io_size / opts->io_unit_size; 2140 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2141 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2142 spdk_nvmf_rdma_destroy(&rtransport->transport); 2143 return NULL; 2144 } 2145 2146 rtransport->event_channel = rdma_create_event_channel(); 2147 if (rtransport->event_channel == NULL) { 2148 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2149 spdk_nvmf_rdma_destroy(&rtransport->transport); 2150 return NULL; 2151 } 2152 2153 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2154 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2155 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2156 rtransport->event_channel->fd, spdk_strerror(errno)); 2157 spdk_nvmf_rdma_destroy(&rtransport->transport); 2158 return NULL; 2159 } 2160 2161 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2162 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2163 sizeof(struct spdk_nvmf_rdma_request_data), 2164 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2165 SPDK_ENV_SOCKET_ID_ANY); 2166 if (!rtransport->data_wr_pool) { 2167 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2168 spdk_nvmf_rdma_destroy(&rtransport->transport); 2169 return NULL; 2170 } 2171 2172 contexts = rdma_get_devices(NULL); 2173 if (contexts == NULL) { 2174 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2175 spdk_nvmf_rdma_destroy(&rtransport->transport); 2176 return NULL; 2177 } 2178 2179 i = 0; 2180 rc = 0; 2181 while (contexts[i] != NULL) { 2182 device = calloc(1, sizeof(*device)); 2183 if (!device) { 2184 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2185 rc = -ENOMEM; 2186 break; 2187 } 2188 device->context = contexts[i]; 2189 rc = ibv_query_device(device->context, &device->attr); 2190 if (rc < 0) { 2191 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2192 free(device); 2193 break; 2194 2195 } 2196 2197 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2198 2199 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2200 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2201 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2202 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2203 } 2204 2205 /** 2206 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2207 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2208 * but incorrectly reports that it does. There are changes making their way 2209 * through the kernel now that will enable this feature. When they are merged, 2210 * we can conditionally enable this feature. 2211 * 2212 * TODO: enable this for versions of the kernel rxe driver that support it. 2213 */ 2214 if (device->attr.vendor_id == 0) { 2215 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2216 } 2217 #endif 2218 2219 /* set up device context async ev fd as NON_BLOCKING */ 2220 flag = fcntl(device->context->async_fd, F_GETFL); 2221 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2222 if (rc < 0) { 2223 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2224 free(device); 2225 break; 2226 } 2227 2228 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2229 i++; 2230 2231 pd = NULL; 2232 if (g_nvmf_hooks.get_ibv_pd) { 2233 pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2234 } 2235 2236 if (!g_nvmf_hooks.get_ibv_pd) { 2237 device->pd = ibv_alloc_pd(device->context); 2238 if (!device->pd) { 2239 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2240 spdk_nvmf_rdma_destroy(&rtransport->transport); 2241 return NULL; 2242 } 2243 } else { 2244 device->pd = pd; 2245 } 2246 2247 assert(device->map == NULL); 2248 2249 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2250 if (!device->map) { 2251 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2252 spdk_nvmf_rdma_destroy(&rtransport->transport); 2253 return NULL; 2254 } 2255 2256 assert(device->map != NULL); 2257 assert(device->pd != NULL); 2258 } 2259 rdma_free_devices(contexts); 2260 2261 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2262 /* divide and round up. */ 2263 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2264 2265 /* round up to the nearest 4k. */ 2266 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2267 2268 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2269 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2270 opts->io_unit_size); 2271 } 2272 2273 if (rc < 0) { 2274 spdk_nvmf_rdma_destroy(&rtransport->transport); 2275 return NULL; 2276 } 2277 2278 /* Set up poll descriptor array to monitor events from RDMA and IB 2279 * in a single poll syscall 2280 */ 2281 rtransport->npoll_fds = i + 1; 2282 i = 0; 2283 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2284 if (rtransport->poll_fds == NULL) { 2285 SPDK_ERRLOG("poll_fds allocation failed\n"); 2286 spdk_nvmf_rdma_destroy(&rtransport->transport); 2287 return NULL; 2288 } 2289 2290 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2291 rtransport->poll_fds[i++].events = POLLIN; 2292 2293 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2294 rtransport->poll_fds[i].fd = device->context->async_fd; 2295 rtransport->poll_fds[i++].events = POLLIN; 2296 } 2297 2298 return &rtransport->transport; 2299 } 2300 2301 static int 2302 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2303 { 2304 struct spdk_nvmf_rdma_transport *rtransport; 2305 struct spdk_nvmf_rdma_port *port, *port_tmp; 2306 struct spdk_nvmf_rdma_device *device, *device_tmp; 2307 2308 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2309 2310 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2311 TAILQ_REMOVE(&rtransport->ports, port, link); 2312 rdma_destroy_id(port->id); 2313 free(port); 2314 } 2315 2316 if (rtransport->poll_fds != NULL) { 2317 free(rtransport->poll_fds); 2318 } 2319 2320 if (rtransport->event_channel != NULL) { 2321 rdma_destroy_event_channel(rtransport->event_channel); 2322 } 2323 2324 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2325 TAILQ_REMOVE(&rtransport->devices, device, link); 2326 if (device->map) { 2327 spdk_mem_map_free(&device->map); 2328 } 2329 if (device->pd) { 2330 if (!g_nvmf_hooks.get_ibv_pd) { 2331 ibv_dealloc_pd(device->pd); 2332 } 2333 } 2334 free(device); 2335 } 2336 2337 if (rtransport->data_wr_pool != NULL) { 2338 if (spdk_mempool_count(rtransport->data_wr_pool) != 2339 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2340 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2341 spdk_mempool_count(rtransport->data_wr_pool), 2342 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2343 } 2344 } 2345 2346 spdk_mempool_free(rtransport->data_wr_pool); 2347 pthread_mutex_destroy(&rtransport->lock); 2348 free(rtransport); 2349 2350 return 0; 2351 } 2352 2353 static int 2354 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2355 struct spdk_nvme_transport_id *trid, 2356 bool peer); 2357 2358 static int 2359 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2360 const struct spdk_nvme_transport_id *trid) 2361 { 2362 struct spdk_nvmf_rdma_transport *rtransport; 2363 struct spdk_nvmf_rdma_device *device; 2364 struct spdk_nvmf_rdma_port *port_tmp, *port; 2365 struct addrinfo *res; 2366 struct addrinfo hints; 2367 int family; 2368 int rc; 2369 2370 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2371 2372 port = calloc(1, sizeof(*port)); 2373 if (!port) { 2374 return -ENOMEM; 2375 } 2376 2377 /* Selectively copy the trid. Things like NQN don't matter here - that 2378 * mapping is enforced elsewhere. 2379 */ 2380 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2381 port->trid.adrfam = trid->adrfam; 2382 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 2383 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 2384 2385 pthread_mutex_lock(&rtransport->lock); 2386 assert(rtransport->event_channel != NULL); 2387 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 2388 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 2389 port_tmp->ref++; 2390 free(port); 2391 /* Already listening at this address */ 2392 pthread_mutex_unlock(&rtransport->lock); 2393 return 0; 2394 } 2395 } 2396 2397 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2398 if (rc < 0) { 2399 SPDK_ERRLOG("rdma_create_id() failed\n"); 2400 free(port); 2401 pthread_mutex_unlock(&rtransport->lock); 2402 return rc; 2403 } 2404 2405 switch (port->trid.adrfam) { 2406 case SPDK_NVMF_ADRFAM_IPV4: 2407 family = AF_INET; 2408 break; 2409 case SPDK_NVMF_ADRFAM_IPV6: 2410 family = AF_INET6; 2411 break; 2412 default: 2413 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 2414 free(port); 2415 pthread_mutex_unlock(&rtransport->lock); 2416 return -EINVAL; 2417 } 2418 2419 memset(&hints, 0, sizeof(hints)); 2420 hints.ai_family = family; 2421 hints.ai_flags = AI_NUMERICSERV; 2422 hints.ai_socktype = SOCK_STREAM; 2423 hints.ai_protocol = 0; 2424 2425 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 2426 if (rc) { 2427 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2428 free(port); 2429 pthread_mutex_unlock(&rtransport->lock); 2430 return -EINVAL; 2431 } 2432 2433 rc = rdma_bind_addr(port->id, res->ai_addr); 2434 freeaddrinfo(res); 2435 2436 if (rc < 0) { 2437 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2438 rdma_destroy_id(port->id); 2439 free(port); 2440 pthread_mutex_unlock(&rtransport->lock); 2441 return rc; 2442 } 2443 2444 if (!port->id->verbs) { 2445 SPDK_ERRLOG("ibv_context is null\n"); 2446 rdma_destroy_id(port->id); 2447 free(port); 2448 pthread_mutex_unlock(&rtransport->lock); 2449 return -1; 2450 } 2451 2452 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 2453 if (rc < 0) { 2454 SPDK_ERRLOG("rdma_listen() failed\n"); 2455 rdma_destroy_id(port->id); 2456 free(port); 2457 pthread_mutex_unlock(&rtransport->lock); 2458 return rc; 2459 } 2460 2461 TAILQ_FOREACH(device, &rtransport->devices, link) { 2462 if (device->context == port->id->verbs) { 2463 port->device = device; 2464 break; 2465 } 2466 } 2467 if (!port->device) { 2468 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2469 port->id->verbs); 2470 rdma_destroy_id(port->id); 2471 free(port); 2472 pthread_mutex_unlock(&rtransport->lock); 2473 return -EINVAL; 2474 } 2475 2476 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 2477 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 2478 2479 port->ref = 1; 2480 2481 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2482 pthread_mutex_unlock(&rtransport->lock); 2483 2484 return 0; 2485 } 2486 2487 static int 2488 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2489 const struct spdk_nvme_transport_id *_trid) 2490 { 2491 struct spdk_nvmf_rdma_transport *rtransport; 2492 struct spdk_nvmf_rdma_port *port, *tmp; 2493 struct spdk_nvme_transport_id trid = {}; 2494 2495 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2496 2497 /* Selectively copy the trid. Things like NQN don't matter here - that 2498 * mapping is enforced elsewhere. 2499 */ 2500 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2501 trid.adrfam = _trid->adrfam; 2502 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2503 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2504 2505 pthread_mutex_lock(&rtransport->lock); 2506 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2507 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2508 assert(port->ref > 0); 2509 port->ref--; 2510 if (port->ref == 0) { 2511 TAILQ_REMOVE(&rtransport->ports, port, link); 2512 rdma_destroy_id(port->id); 2513 free(port); 2514 } 2515 break; 2516 } 2517 } 2518 2519 pthread_mutex_unlock(&rtransport->lock); 2520 return 0; 2521 } 2522 2523 static void 2524 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2525 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2526 { 2527 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2528 struct spdk_nvmf_rdma_resources *resources; 2529 2530 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2531 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2532 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2533 break; 2534 } 2535 } 2536 2537 /* Then RDMA writes since reads have stronger restrictions than writes */ 2538 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2539 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2540 break; 2541 } 2542 } 2543 2544 /* The second highest priority is I/O waiting on memory buffers. */ 2545 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->poller->group->pending_data_buf_queue, state_link, 2546 req_tmp) { 2547 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2548 break; 2549 } 2550 } 2551 2552 resources = rqpair->resources; 2553 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2554 rdma_req = STAILQ_FIRST(&resources->free_queue); 2555 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2556 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2557 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2558 2559 if (rqpair->srq != NULL) { 2560 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2561 rdma_req->recv->qpair->qd++; 2562 } else { 2563 rqpair->qd++; 2564 } 2565 2566 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2567 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2568 break; 2569 } 2570 } 2571 } 2572 2573 static void 2574 _nvmf_rdma_qpair_disconnect(void *ctx) 2575 { 2576 struct spdk_nvmf_qpair *qpair = ctx; 2577 2578 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2579 } 2580 2581 static void 2582 _nvmf_rdma_try_disconnect(void *ctx) 2583 { 2584 struct spdk_nvmf_qpair *qpair = ctx; 2585 struct spdk_nvmf_poll_group *group; 2586 2587 /* Read the group out of the qpair. This is normally set and accessed only from 2588 * the thread that created the group. Here, we're not on that thread necessarily. 2589 * The data member qpair->group begins it's life as NULL and then is assigned to 2590 * a pointer and never changes. So fortunately reading this and checking for 2591 * non-NULL is thread safe in the x86_64 memory model. */ 2592 group = qpair->group; 2593 2594 if (group == NULL) { 2595 /* The qpair hasn't been assigned to a group yet, so we can't 2596 * process a disconnect. Send a message to ourself and try again. */ 2597 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2598 return; 2599 } 2600 2601 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2602 } 2603 2604 static inline void 2605 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2606 { 2607 if (__sync_bool_compare_and_swap(&rqpair->disconnect_started, false, true)) { 2608 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2609 } 2610 } 2611 2612 static void nvmf_rdma_destroy_drained_qpair(void *ctx) 2613 { 2614 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2615 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2616 struct spdk_nvmf_rdma_transport, transport); 2617 2618 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2619 if (rqpair->current_send_depth != 0) { 2620 return; 2621 } 2622 2623 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2624 return; 2625 } 2626 2627 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) { 2628 return; 2629 } 2630 2631 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2632 spdk_nvmf_rdma_qpair_destroy(rqpair); 2633 } 2634 2635 2636 static int 2637 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2638 { 2639 struct spdk_nvmf_qpair *qpair; 2640 struct spdk_nvmf_rdma_qpair *rqpair; 2641 2642 if (evt->id == NULL) { 2643 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2644 return -1; 2645 } 2646 2647 qpair = evt->id->context; 2648 if (qpair == NULL) { 2649 SPDK_ERRLOG("disconnect request: no active connection\n"); 2650 return -1; 2651 } 2652 2653 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2654 2655 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2656 2657 spdk_nvmf_rdma_update_ibv_state(rqpair); 2658 2659 spdk_nvmf_rdma_start_disconnect(rqpair); 2660 2661 return 0; 2662 } 2663 2664 #ifdef DEBUG 2665 static const char *CM_EVENT_STR[] = { 2666 "RDMA_CM_EVENT_ADDR_RESOLVED", 2667 "RDMA_CM_EVENT_ADDR_ERROR", 2668 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2669 "RDMA_CM_EVENT_ROUTE_ERROR", 2670 "RDMA_CM_EVENT_CONNECT_REQUEST", 2671 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2672 "RDMA_CM_EVENT_CONNECT_ERROR", 2673 "RDMA_CM_EVENT_UNREACHABLE", 2674 "RDMA_CM_EVENT_REJECTED", 2675 "RDMA_CM_EVENT_ESTABLISHED", 2676 "RDMA_CM_EVENT_DISCONNECTED", 2677 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2678 "RDMA_CM_EVENT_MULTICAST_JOIN", 2679 "RDMA_CM_EVENT_MULTICAST_ERROR", 2680 "RDMA_CM_EVENT_ADDR_CHANGE", 2681 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2682 }; 2683 #endif /* DEBUG */ 2684 2685 static void 2686 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2687 { 2688 struct spdk_nvmf_rdma_transport *rtransport; 2689 struct rdma_cm_event *event; 2690 int rc; 2691 2692 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2693 2694 if (rtransport->event_channel == NULL) { 2695 return; 2696 } 2697 2698 while (1) { 2699 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2700 if (rc == 0) { 2701 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2702 2703 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2704 2705 switch (event->event) { 2706 case RDMA_CM_EVENT_ADDR_RESOLVED: 2707 case RDMA_CM_EVENT_ADDR_ERROR: 2708 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2709 case RDMA_CM_EVENT_ROUTE_ERROR: 2710 /* No action required. The target never attempts to resolve routes. */ 2711 break; 2712 case RDMA_CM_EVENT_CONNECT_REQUEST: 2713 rc = nvmf_rdma_connect(transport, event, cb_fn); 2714 if (rc < 0) { 2715 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2716 break; 2717 } 2718 break; 2719 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2720 /* The target never initiates a new connection. So this will not occur. */ 2721 break; 2722 case RDMA_CM_EVENT_CONNECT_ERROR: 2723 /* Can this happen? The docs say it can, but not sure what causes it. */ 2724 break; 2725 case RDMA_CM_EVENT_UNREACHABLE: 2726 case RDMA_CM_EVENT_REJECTED: 2727 /* These only occur on the client side. */ 2728 break; 2729 case RDMA_CM_EVENT_ESTABLISHED: 2730 /* TODO: Should we be waiting for this event anywhere? */ 2731 break; 2732 case RDMA_CM_EVENT_DISCONNECTED: 2733 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2734 rc = nvmf_rdma_disconnect(event); 2735 if (rc < 0) { 2736 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2737 break; 2738 } 2739 break; 2740 case RDMA_CM_EVENT_MULTICAST_JOIN: 2741 case RDMA_CM_EVENT_MULTICAST_ERROR: 2742 /* Multicast is not used */ 2743 break; 2744 case RDMA_CM_EVENT_ADDR_CHANGE: 2745 /* Not utilizing this event */ 2746 break; 2747 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2748 /* For now, do nothing. The target never re-uses queue pairs. */ 2749 break; 2750 default: 2751 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2752 break; 2753 } 2754 2755 rdma_ack_cm_event(event); 2756 } else { 2757 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2758 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2759 } 2760 break; 2761 } 2762 } 2763 } 2764 2765 static void 2766 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2767 { 2768 int rc; 2769 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 2770 struct ibv_async_event event; 2771 enum ibv_qp_state state; 2772 2773 rc = ibv_get_async_event(device->context, &event); 2774 2775 if (rc) { 2776 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2777 errno, spdk_strerror(errno)); 2778 return; 2779 } 2780 2781 switch (event.event_type) { 2782 case IBV_EVENT_QP_FATAL: 2783 rqpair = event.element.qp->qp_context; 2784 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 2785 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2786 (uintptr_t)rqpair->cm_id, event.event_type); 2787 spdk_nvmf_rdma_update_ibv_state(rqpair); 2788 spdk_nvmf_rdma_start_disconnect(rqpair); 2789 break; 2790 case IBV_EVENT_QP_LAST_WQE_REACHED: 2791 /* This event only occurs for shared receive queues. */ 2792 rqpair = event.element.qp->qp_context; 2793 rqpair->last_wqe_reached = true; 2794 2795 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last WQE reached event received for rqpair %p\n", rqpair); 2796 /* This must be handled on the polling thread if it exists. Otherwise the timeout will catch it. */ 2797 if (rqpair->qpair.group) { 2798 spdk_thread_send_msg(rqpair->qpair.group->thread, nvmf_rdma_destroy_drained_qpair, rqpair); 2799 } else { 2800 SPDK_ERRLOG("Unable to destroy the qpair %p since it does not have a poll group.\n", rqpair); 2801 } 2802 2803 break; 2804 case IBV_EVENT_SQ_DRAINED: 2805 /* This event occurs frequently in both error and non-error states. 2806 * Check if the qpair is in an error state before sending a message. 2807 * Note that we're not on the correct thread to access the qpair, but 2808 * the operations that the below calls make all happen to be thread 2809 * safe. */ 2810 rqpair = event.element.qp->qp_context; 2811 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last sq drained event received for rqpair %p\n", rqpair); 2812 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2813 (uintptr_t)rqpair->cm_id, event.event_type); 2814 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2815 if (state == IBV_QPS_ERR) { 2816 spdk_nvmf_rdma_start_disconnect(rqpair); 2817 } 2818 break; 2819 case IBV_EVENT_QP_REQ_ERR: 2820 case IBV_EVENT_QP_ACCESS_ERR: 2821 case IBV_EVENT_COMM_EST: 2822 case IBV_EVENT_PATH_MIG: 2823 case IBV_EVENT_PATH_MIG_ERR: 2824 SPDK_NOTICELOG("Async event: %s\n", 2825 ibv_event_type_str(event.event_type)); 2826 rqpair = event.element.qp->qp_context; 2827 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2828 (uintptr_t)rqpair->cm_id, event.event_type); 2829 spdk_nvmf_rdma_update_ibv_state(rqpair); 2830 break; 2831 case IBV_EVENT_CQ_ERR: 2832 case IBV_EVENT_DEVICE_FATAL: 2833 case IBV_EVENT_PORT_ACTIVE: 2834 case IBV_EVENT_PORT_ERR: 2835 case IBV_EVENT_LID_CHANGE: 2836 case IBV_EVENT_PKEY_CHANGE: 2837 case IBV_EVENT_SM_CHANGE: 2838 case IBV_EVENT_SRQ_ERR: 2839 case IBV_EVENT_SRQ_LIMIT_REACHED: 2840 case IBV_EVENT_CLIENT_REREGISTER: 2841 case IBV_EVENT_GID_CHANGE: 2842 default: 2843 SPDK_NOTICELOG("Async event: %s\n", 2844 ibv_event_type_str(event.event_type)); 2845 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2846 break; 2847 } 2848 ibv_ack_async_event(&event); 2849 } 2850 2851 static void 2852 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2853 { 2854 int nfds, i = 0; 2855 struct spdk_nvmf_rdma_transport *rtransport; 2856 struct spdk_nvmf_rdma_device *device, *tmp; 2857 2858 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2859 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2860 2861 if (nfds <= 0) { 2862 return; 2863 } 2864 2865 /* The first poll descriptor is RDMA CM event */ 2866 if (rtransport->poll_fds[i++].revents & POLLIN) { 2867 spdk_nvmf_process_cm_event(transport, cb_fn); 2868 nfds--; 2869 } 2870 2871 if (nfds == 0) { 2872 return; 2873 } 2874 2875 /* Second and subsequent poll descriptors are IB async events */ 2876 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2877 if (rtransport->poll_fds[i++].revents & POLLIN) { 2878 spdk_nvmf_process_ib_event(device); 2879 nfds--; 2880 } 2881 } 2882 /* check all flagged fd's have been served */ 2883 assert(nfds == 0); 2884 } 2885 2886 static void 2887 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2888 struct spdk_nvme_transport_id *trid, 2889 struct spdk_nvmf_discovery_log_page_entry *entry) 2890 { 2891 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2892 entry->adrfam = trid->adrfam; 2893 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2894 2895 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2896 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2897 2898 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2899 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2900 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2901 } 2902 2903 static void 2904 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 2905 2906 static struct spdk_nvmf_transport_poll_group * 2907 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2908 { 2909 struct spdk_nvmf_rdma_transport *rtransport; 2910 struct spdk_nvmf_rdma_poll_group *rgroup; 2911 struct spdk_nvmf_rdma_poller *poller; 2912 struct spdk_nvmf_rdma_device *device; 2913 struct ibv_srq_init_attr srq_init_attr; 2914 struct spdk_nvmf_rdma_resource_opts opts; 2915 int num_cqe; 2916 2917 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2918 2919 rgroup = calloc(1, sizeof(*rgroup)); 2920 if (!rgroup) { 2921 return NULL; 2922 } 2923 2924 TAILQ_INIT(&rgroup->pollers); 2925 STAILQ_INIT(&rgroup->pending_data_buf_queue); 2926 2927 pthread_mutex_lock(&rtransport->lock); 2928 TAILQ_FOREACH(device, &rtransport->devices, link) { 2929 poller = calloc(1, sizeof(*poller)); 2930 if (!poller) { 2931 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2932 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2933 pthread_mutex_unlock(&rtransport->lock); 2934 return NULL; 2935 } 2936 2937 poller->device = device; 2938 poller->group = rgroup; 2939 2940 TAILQ_INIT(&poller->qpairs); 2941 2942 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2943 if (device->attr.max_srq != 0) { 2944 poller->max_srq_depth = transport->opts.max_srq_depth; 2945 2946 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 2947 srq_init_attr.attr.max_wr = poller->max_srq_depth; 2948 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 2949 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 2950 if (!poller->srq) { 2951 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 2952 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2953 pthread_mutex_unlock(&rtransport->lock); 2954 return NULL; 2955 } 2956 2957 opts.qp = poller->srq; 2958 opts.pd = device->pd; 2959 opts.qpair = NULL; 2960 opts.shared = true; 2961 opts.max_queue_depth = poller->max_srq_depth; 2962 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 2963 2964 poller->resources = nvmf_rdma_resources_create(&opts); 2965 if (!poller->resources) { 2966 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 2967 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2968 pthread_mutex_unlock(&rtransport->lock); 2969 } 2970 } 2971 2972 /* 2973 * When using an srq, we can limit the completion queue at startup. 2974 * The following formula represents the calculation: 2975 * num_cqe = num_recv + num_data_wr + num_send_wr. 2976 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 2977 */ 2978 if (poller->srq) { 2979 num_cqe = poller->max_srq_depth * 3; 2980 } else { 2981 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2982 } 2983 2984 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 2985 if (!poller->cq) { 2986 SPDK_ERRLOG("Unable to create completion queue\n"); 2987 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2988 pthread_mutex_unlock(&rtransport->lock); 2989 return NULL; 2990 } 2991 poller->num_cqe = num_cqe; 2992 } 2993 2994 pthread_mutex_unlock(&rtransport->lock); 2995 return &rgroup->group; 2996 } 2997 2998 static void 2999 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3000 { 3001 struct spdk_nvmf_rdma_poll_group *rgroup; 3002 struct spdk_nvmf_rdma_poller *poller, *tmp; 3003 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3004 3005 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3006 3007 if (!rgroup) { 3008 return; 3009 } 3010 3011 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3012 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3013 3014 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3015 spdk_nvmf_rdma_qpair_destroy(qpair); 3016 } 3017 3018 if (poller->srq) { 3019 nvmf_rdma_resources_destroy(poller->resources); 3020 ibv_destroy_srq(poller->srq); 3021 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3022 } 3023 3024 if (poller->cq) { 3025 ibv_destroy_cq(poller->cq); 3026 } 3027 3028 free(poller); 3029 } 3030 3031 if (!STAILQ_EMPTY(&rgroup->pending_data_buf_queue)) { 3032 SPDK_ERRLOG("Pending I/O list wasn't empty on poll group destruction\n"); 3033 } 3034 3035 free(rgroup); 3036 } 3037 3038 static void 3039 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3040 { 3041 if (rqpair->cm_id != NULL) { 3042 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3043 } 3044 spdk_nvmf_rdma_qpair_destroy(rqpair); 3045 } 3046 3047 static int 3048 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3049 struct spdk_nvmf_qpair *qpair) 3050 { 3051 struct spdk_nvmf_rdma_poll_group *rgroup; 3052 struct spdk_nvmf_rdma_qpair *rqpair; 3053 struct spdk_nvmf_rdma_device *device; 3054 struct spdk_nvmf_rdma_poller *poller; 3055 int rc; 3056 3057 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3058 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3059 3060 device = rqpair->port->device; 3061 3062 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3063 if (poller->device == device) { 3064 break; 3065 } 3066 } 3067 3068 if (!poller) { 3069 SPDK_ERRLOG("No poller found for device.\n"); 3070 return -1; 3071 } 3072 3073 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3074 rqpair->poller = poller; 3075 rqpair->srq = rqpair->poller->srq; 3076 3077 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 3078 if (rc < 0) { 3079 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3080 return -1; 3081 } 3082 3083 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3084 if (rc) { 3085 /* Try to reject, but we probably can't */ 3086 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3087 return -1; 3088 } 3089 3090 spdk_nvmf_rdma_update_ibv_state(rqpair); 3091 3092 return 0; 3093 } 3094 3095 static int 3096 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3097 { 3098 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3099 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3100 struct spdk_nvmf_rdma_transport, transport); 3101 3102 nvmf_rdma_request_free(rdma_req, rtransport); 3103 return 0; 3104 } 3105 3106 static int 3107 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3108 { 3109 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3110 struct spdk_nvmf_rdma_transport, transport); 3111 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3112 struct spdk_nvmf_rdma_request, req); 3113 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3114 struct spdk_nvmf_rdma_qpair, qpair); 3115 3116 if (rqpair->ibv_state != IBV_QPS_ERR) { 3117 /* The connection is alive, so process the request as normal */ 3118 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3119 } else { 3120 /* The connection is dead. Move the request directly to the completed state. */ 3121 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3122 } 3123 3124 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3125 3126 return 0; 3127 } 3128 3129 static int 3130 spdk_nvmf_rdma_destroy_defunct_qpair(void *ctx) 3131 { 3132 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 3133 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3134 struct spdk_nvmf_rdma_transport, transport); 3135 3136 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3137 spdk_nvmf_rdma_qpair_destroy(rqpair); 3138 3139 return 0; 3140 } 3141 3142 static void 3143 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3144 { 3145 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3146 3147 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 3148 return; 3149 } 3150 3151 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 3152 3153 /* This happens only when the qpair is disconnected before 3154 * it is added to the poll group. Since there is no poll group, 3155 * the RDMA qp has not been initialized yet and the RDMA CM 3156 * event has not yet been acknowledged, so we need to reject it. 3157 */ 3158 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3159 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3160 return; 3161 } 3162 3163 if (rqpair->ibv_state != IBV_QPS_ERR) { 3164 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 3165 } 3166 3167 rqpair->destruct_poller = spdk_poller_register(spdk_nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 3168 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 3169 } 3170 3171 static struct spdk_nvmf_rdma_qpair * 3172 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3173 { 3174 struct spdk_nvmf_rdma_qpair *rqpair; 3175 /* @todo: improve QP search */ 3176 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3177 if (wc->qp_num == rqpair->cm_id->qp->qp_num) { 3178 return rqpair; 3179 } 3180 } 3181 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3182 return NULL; 3183 } 3184 3185 #ifdef DEBUG 3186 static int 3187 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3188 { 3189 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3190 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3191 } 3192 #endif 3193 3194 static int 3195 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3196 struct spdk_nvmf_rdma_poller *rpoller) 3197 { 3198 struct ibv_wc wc[32]; 3199 struct spdk_nvmf_rdma_wr *rdma_wr; 3200 struct spdk_nvmf_rdma_request *rdma_req; 3201 struct spdk_nvmf_rdma_recv *rdma_recv; 3202 struct spdk_nvmf_rdma_qpair *rqpair; 3203 int reaped, i; 3204 int count = 0; 3205 bool error = false; 3206 3207 /* Poll for completing operations. */ 3208 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3209 if (reaped < 0) { 3210 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3211 errno, spdk_strerror(errno)); 3212 return -1; 3213 } 3214 3215 for (i = 0; i < reaped; i++) { 3216 3217 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3218 3219 switch (rdma_wr->type) { 3220 case RDMA_WR_TYPE_SEND: 3221 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3222 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3223 3224 if (!wc[i].status) { 3225 count++; 3226 assert(wc[i].opcode == IBV_WC_SEND); 3227 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 3228 } else { 3229 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3230 } 3231 3232 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3233 rqpair->current_send_depth--; 3234 3235 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3236 assert(rdma_req->num_outstanding_data_wr == 0); 3237 break; 3238 case RDMA_WR_TYPE_RECV: 3239 /* rdma_recv->qpair will be NULL if using an SRQ. In that case we have to get the qpair from the wc. */ 3240 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3241 if (rdma_recv->qpair == NULL) { 3242 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3243 } 3244 rqpair = rdma_recv->qpair; 3245 3246 assert(rqpair != NULL); 3247 if (!wc[i].status) { 3248 assert(wc[i].opcode == IBV_WC_RECV); 3249 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3250 spdk_nvmf_rdma_start_disconnect(rqpair); 3251 break; 3252 } 3253 } 3254 3255 rqpair->current_recv_depth++; 3256 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3257 break; 3258 case RDMA_WR_TYPE_DATA: 3259 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3260 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3261 3262 assert(rdma_req->num_outstanding_data_wr > 0); 3263 3264 rqpair->current_send_depth--; 3265 rdma_req->num_outstanding_data_wr--; 3266 if (!wc[i].status) { 3267 if (wc[i].opcode == IBV_WC_RDMA_READ) { 3268 rqpair->current_read_depth--; 3269 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3270 if (rdma_req->num_outstanding_data_wr == 0) { 3271 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3272 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3273 } 3274 } else { 3275 assert(wc[i].opcode == IBV_WC_RDMA_WRITE); 3276 } 3277 } else { 3278 /* If the data transfer fails still force the queue into the error state, 3279 * if we were performing an RDMA_READ, we need to force the request into a 3280 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3281 * case, we should wait for the SEND to complete. */ 3282 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3283 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3284 rqpair->current_read_depth--; 3285 if (rdma_req->num_outstanding_data_wr == 0) { 3286 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3287 } 3288 } 3289 } 3290 break; 3291 default: 3292 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3293 continue; 3294 } 3295 3296 /* Handle error conditions */ 3297 if (wc[i].status) { 3298 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 3299 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 3300 3301 error = true; 3302 3303 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3304 /* Disconnect the connection. */ 3305 spdk_nvmf_rdma_start_disconnect(rqpair); 3306 } else { 3307 nvmf_rdma_destroy_drained_qpair(rqpair); 3308 } 3309 continue; 3310 } 3311 3312 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3313 3314 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3315 nvmf_rdma_destroy_drained_qpair(rqpair); 3316 } 3317 } 3318 3319 if (error == true) { 3320 return -1; 3321 } 3322 3323 return count; 3324 } 3325 3326 static int 3327 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3328 { 3329 struct spdk_nvmf_rdma_transport *rtransport; 3330 struct spdk_nvmf_rdma_poll_group *rgroup; 3331 struct spdk_nvmf_rdma_poller *rpoller; 3332 int count, rc; 3333 3334 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3335 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3336 3337 count = 0; 3338 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3339 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 3340 if (rc < 0) { 3341 return rc; 3342 } 3343 count += rc; 3344 } 3345 3346 return count; 3347 } 3348 3349 static int 3350 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3351 struct spdk_nvme_transport_id *trid, 3352 bool peer) 3353 { 3354 struct sockaddr *saddr; 3355 uint16_t port; 3356 3357 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 3358 3359 if (peer) { 3360 saddr = rdma_get_peer_addr(id); 3361 } else { 3362 saddr = rdma_get_local_addr(id); 3363 } 3364 switch (saddr->sa_family) { 3365 case AF_INET: { 3366 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3367 3368 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3369 inet_ntop(AF_INET, &saddr_in->sin_addr, 3370 trid->traddr, sizeof(trid->traddr)); 3371 if (peer) { 3372 port = ntohs(rdma_get_dst_port(id)); 3373 } else { 3374 port = ntohs(rdma_get_src_port(id)); 3375 } 3376 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3377 break; 3378 } 3379 case AF_INET6: { 3380 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3381 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3382 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3383 trid->traddr, sizeof(trid->traddr)); 3384 if (peer) { 3385 port = ntohs(rdma_get_dst_port(id)); 3386 } else { 3387 port = ntohs(rdma_get_src_port(id)); 3388 } 3389 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3390 break; 3391 } 3392 default: 3393 return -1; 3394 3395 } 3396 3397 return 0; 3398 } 3399 3400 static int 3401 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3402 struct spdk_nvme_transport_id *trid) 3403 { 3404 struct spdk_nvmf_rdma_qpair *rqpair; 3405 3406 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3407 3408 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 3409 } 3410 3411 static int 3412 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3413 struct spdk_nvme_transport_id *trid) 3414 { 3415 struct spdk_nvmf_rdma_qpair *rqpair; 3416 3417 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3418 3419 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 3420 } 3421 3422 static int 3423 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3424 struct spdk_nvme_transport_id *trid) 3425 { 3426 struct spdk_nvmf_rdma_qpair *rqpair; 3427 3428 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3429 3430 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 3431 } 3432 3433 void 3434 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3435 { 3436 g_nvmf_hooks = *hooks; 3437 } 3438 3439 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 3440 .type = SPDK_NVME_TRANSPORT_RDMA, 3441 .opts_init = spdk_nvmf_rdma_opts_init, 3442 .create = spdk_nvmf_rdma_create, 3443 .destroy = spdk_nvmf_rdma_destroy, 3444 3445 .listen = spdk_nvmf_rdma_listen, 3446 .stop_listen = spdk_nvmf_rdma_stop_listen, 3447 .accept = spdk_nvmf_rdma_accept, 3448 3449 .listener_discover = spdk_nvmf_rdma_discover, 3450 3451 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 3452 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 3453 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 3454 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 3455 3456 .req_free = spdk_nvmf_rdma_request_free, 3457 .req_complete = spdk_nvmf_rdma_request_complete, 3458 3459 .qpair_fini = spdk_nvmf_rdma_close_qpair, 3460 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 3461 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 3462 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 3463 3464 }; 3465 3466 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 3467