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", TRACE_RDMA_REQUEST_STATE_NEW, 166 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 167 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 168 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 169 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 170 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 171 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 172 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 173 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 174 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 175 spdk_trace_register_description("RDMA_REQ_TX_H2C", 176 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 177 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 178 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 179 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 180 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 181 spdk_trace_register_description("RDMA_REQ_EXECUTING", 182 TRACE_RDMA_REQUEST_STATE_EXECUTING, 183 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 184 spdk_trace_register_description("RDMA_REQ_EXECUTED", 185 TRACE_RDMA_REQUEST_STATE_EXECUTED, 186 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 187 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 188 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 189 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 190 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 191 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 192 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 193 spdk_trace_register_description("RDMA_REQ_COMPLETING", 194 TRACE_RDMA_REQUEST_STATE_COMPLETING, 195 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 196 spdk_trace_register_description("RDMA_REQ_COMPLETED", 197 TRACE_RDMA_REQUEST_STATE_COMPLETED, 198 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 199 200 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 201 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 202 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 203 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 204 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 205 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 206 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 207 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 208 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 209 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 210 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 211 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 212 } 213 214 enum spdk_nvmf_rdma_wr_type { 215 RDMA_WR_TYPE_RECV, 216 RDMA_WR_TYPE_SEND, 217 RDMA_WR_TYPE_DATA, 218 }; 219 220 struct spdk_nvmf_rdma_wr { 221 enum spdk_nvmf_rdma_wr_type type; 222 }; 223 224 /* This structure holds commands as they are received off the wire. 225 * It must be dynamically paired with a full request object 226 * (spdk_nvmf_rdma_request) to service a request. It is separate 227 * from the request because RDMA does not appear to order 228 * completions, so occasionally we'll get a new incoming 229 * command when there aren't any free request objects. 230 */ 231 struct spdk_nvmf_rdma_recv { 232 struct ibv_recv_wr wr; 233 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 234 235 struct spdk_nvmf_rdma_qpair *qpair; 236 237 /* In-capsule data buffer */ 238 uint8_t *buf; 239 240 struct spdk_nvmf_rdma_wr rdma_wr; 241 242 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 243 }; 244 245 struct spdk_nvmf_rdma_request_data { 246 struct spdk_nvmf_rdma_wr rdma_wr; 247 struct ibv_send_wr wr; 248 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 249 }; 250 251 struct spdk_nvmf_rdma_request { 252 struct spdk_nvmf_request req; 253 bool data_from_pool; 254 255 enum spdk_nvmf_rdma_request_state state; 256 257 struct spdk_nvmf_rdma_recv *recv; 258 259 struct { 260 struct spdk_nvmf_rdma_wr rdma_wr; 261 struct ibv_send_wr wr; 262 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 263 } rsp; 264 265 struct spdk_nvmf_rdma_request_data data; 266 void *buffers[NVMF_REQ_MAX_BUFFERS]; 267 268 uint32_t num_outstanding_data_wr; 269 270 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 271 }; 272 273 enum spdk_nvmf_rdma_qpair_disconnect_flags { 274 RDMA_QP_DISCONNECTING = 1, 275 RDMA_QP_RECV_DRAINED = 1 << 1, 276 RDMA_QP_SEND_DRAINED = 1 << 2 277 }; 278 279 struct spdk_nvmf_rdma_resource_opts { 280 struct spdk_nvmf_rdma_qpair *qpair; 281 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 282 void *qp; 283 struct ibv_pd *pd; 284 uint32_t max_queue_depth; 285 uint32_t in_capsule_data_size; 286 bool shared; 287 }; 288 289 struct spdk_nvmf_send_wr_list { 290 struct ibv_send_wr *first; 291 struct ibv_send_wr *last; 292 }; 293 294 struct spdk_nvmf_rdma_resources { 295 /* Array of size "max_queue_depth" containing RDMA requests. */ 296 struct spdk_nvmf_rdma_request *reqs; 297 298 /* Array of size "max_queue_depth" containing RDMA recvs. */ 299 struct spdk_nvmf_rdma_recv *recvs; 300 301 /* Array of size "max_queue_depth" containing 64 byte capsules 302 * used for receive. 303 */ 304 union nvmf_h2c_msg *cmds; 305 struct ibv_mr *cmds_mr; 306 307 /* Array of size "max_queue_depth" containing 16 byte completions 308 * to be sent back to the user. 309 */ 310 union nvmf_c2h_msg *cpls; 311 struct ibv_mr *cpls_mr; 312 313 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 314 * buffers to be used for in capsule data. 315 */ 316 void *bufs; 317 struct ibv_mr *bufs_mr; 318 319 /* Receives that are waiting for a request object */ 320 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 321 322 /* Queue to track free requests */ 323 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 324 }; 325 326 struct spdk_nvmf_rdma_qpair { 327 struct spdk_nvmf_qpair qpair; 328 329 struct spdk_nvmf_rdma_port *port; 330 struct spdk_nvmf_rdma_poller *poller; 331 332 struct rdma_cm_id *cm_id; 333 struct ibv_srq *srq; 334 struct rdma_cm_id *listen_id; 335 336 /* The maximum number of I/O outstanding on this connection at one time */ 337 uint16_t max_queue_depth; 338 339 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 340 uint16_t max_read_depth; 341 342 /* The maximum number of RDMA SEND operations at one time */ 343 uint32_t max_send_depth; 344 345 /* The current number of outstanding WRs from this qpair's 346 * recv queue. Should not exceed device->attr.max_queue_depth. 347 */ 348 uint16_t current_recv_depth; 349 350 /* The current number of active RDMA READ operations */ 351 uint16_t current_read_depth; 352 353 /* The current number of posted WRs from this qpair's 354 * send queue. Should not exceed max_send_depth. 355 */ 356 uint32_t current_send_depth; 357 358 /* The maximum number of SGEs per WR on the send queue */ 359 uint32_t max_send_sge; 360 361 /* The maximum number of SGEs per WR on the recv queue */ 362 uint32_t max_recv_sge; 363 364 /* The list of pending send requests for a transfer */ 365 struct spdk_nvmf_send_wr_list sends_to_post; 366 367 struct spdk_nvmf_rdma_resources *resources; 368 369 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 370 371 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 372 373 /* Number of requests not in the free state */ 374 uint32_t qd; 375 376 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 377 378 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 379 380 /* IBV queue pair attributes: they are used to manage 381 * qp state and recover from errors. 382 */ 383 enum ibv_qp_state ibv_state; 384 385 uint32_t disconnect_flags; 386 387 /* Poller registered in case the qpair doesn't properly 388 * complete the qpair destruct process and becomes defunct. 389 */ 390 391 struct spdk_poller *destruct_poller; 392 393 /* There are several ways a disconnect can start on a qpair 394 * and they are not all mutually exclusive. It is important 395 * that we only initialize one of these paths. 396 */ 397 bool disconnect_started; 398 /* Lets us know that we have received the last_wqe event. */ 399 bool last_wqe_reached; 400 }; 401 402 struct spdk_nvmf_rdma_poller { 403 struct spdk_nvmf_rdma_device *device; 404 struct spdk_nvmf_rdma_poll_group *group; 405 406 int num_cqe; 407 int required_num_wr; 408 struct ibv_cq *cq; 409 410 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 411 uint16_t max_srq_depth; 412 413 /* Shared receive queue */ 414 struct ibv_srq *srq; 415 416 struct spdk_nvmf_rdma_resources *resources; 417 418 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 419 420 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 421 422 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 423 }; 424 425 struct spdk_nvmf_rdma_poll_group { 426 struct spdk_nvmf_transport_poll_group group; 427 428 /* Requests that are waiting to obtain a data buffer */ 429 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_data_buf_queue; 430 431 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 432 }; 433 434 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 435 struct spdk_nvmf_rdma_device { 436 struct ibv_device_attr attr; 437 struct ibv_context *context; 438 439 struct spdk_mem_map *map; 440 struct ibv_pd *pd; 441 442 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 443 }; 444 445 struct spdk_nvmf_rdma_port { 446 struct spdk_nvme_transport_id trid; 447 struct rdma_cm_id *id; 448 struct spdk_nvmf_rdma_device *device; 449 uint32_t ref; 450 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 451 }; 452 453 struct spdk_nvmf_rdma_transport { 454 struct spdk_nvmf_transport transport; 455 456 struct rdma_event_channel *event_channel; 457 458 struct spdk_mempool *data_wr_pool; 459 460 pthread_mutex_t lock; 461 462 /* fields used to poll RDMA/IB events */ 463 nfds_t npoll_fds; 464 struct pollfd *poll_fds; 465 466 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 467 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 468 }; 469 470 static inline int 471 spdk_nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 472 { 473 switch (state) { 474 case IBV_QPS_RESET: 475 case IBV_QPS_INIT: 476 case IBV_QPS_RTR: 477 case IBV_QPS_RTS: 478 case IBV_QPS_SQD: 479 case IBV_QPS_SQE: 480 case IBV_QPS_ERR: 481 return 0; 482 default: 483 return -1; 484 } 485 } 486 487 static enum ibv_qp_state 488 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 489 enum ibv_qp_state old_state, new_state; 490 struct ibv_qp_attr qp_attr; 491 struct ibv_qp_init_attr init_attr; 492 int rc; 493 494 old_state = rqpair->ibv_state; 495 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 496 g_spdk_nvmf_ibv_query_mask, &init_attr); 497 498 if (rc) 499 { 500 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 501 return IBV_QPS_ERR + 1; 502 } 503 504 new_state = qp_attr.qp_state; 505 rqpair->ibv_state = new_state; 506 qp_attr.ah_attr.port_num = qp_attr.port_num; 507 508 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 509 if (rc) 510 { 511 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 512 /* 513 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 514 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 515 */ 516 return IBV_QPS_ERR + 1; 517 } 518 519 if (old_state != new_state) 520 { 521 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 522 (uintptr_t)rqpair->cm_id, new_state); 523 } 524 return new_state; 525 } 526 527 static const char *str_ibv_qp_state[] = { 528 "IBV_QPS_RESET", 529 "IBV_QPS_INIT", 530 "IBV_QPS_RTR", 531 "IBV_QPS_RTS", 532 "IBV_QPS_SQD", 533 "IBV_QPS_SQE", 534 "IBV_QPS_ERR", 535 "IBV_QPS_UNKNOWN" 536 }; 537 538 static int 539 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 540 enum ibv_qp_state new_state) 541 { 542 struct ibv_qp_attr qp_attr; 543 struct ibv_qp_init_attr init_attr; 544 int rc; 545 enum ibv_qp_state state; 546 static int attr_mask_rc[] = { 547 [IBV_QPS_RESET] = IBV_QP_STATE, 548 [IBV_QPS_INIT] = (IBV_QP_STATE | 549 IBV_QP_PKEY_INDEX | 550 IBV_QP_PORT | 551 IBV_QP_ACCESS_FLAGS), 552 [IBV_QPS_RTR] = (IBV_QP_STATE | 553 IBV_QP_AV | 554 IBV_QP_PATH_MTU | 555 IBV_QP_DEST_QPN | 556 IBV_QP_RQ_PSN | 557 IBV_QP_MAX_DEST_RD_ATOMIC | 558 IBV_QP_MIN_RNR_TIMER), 559 [IBV_QPS_RTS] = (IBV_QP_STATE | 560 IBV_QP_SQ_PSN | 561 IBV_QP_TIMEOUT | 562 IBV_QP_RETRY_CNT | 563 IBV_QP_RNR_RETRY | 564 IBV_QP_MAX_QP_RD_ATOMIC), 565 [IBV_QPS_SQD] = IBV_QP_STATE, 566 [IBV_QPS_SQE] = IBV_QP_STATE, 567 [IBV_QPS_ERR] = IBV_QP_STATE, 568 }; 569 570 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 571 if (rc) { 572 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 573 rqpair->qpair.qid, new_state); 574 return rc; 575 } 576 577 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 578 g_spdk_nvmf_ibv_query_mask, &init_attr); 579 580 if (rc) { 581 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 582 assert(false); 583 } 584 585 qp_attr.cur_qp_state = rqpair->ibv_state; 586 qp_attr.qp_state = new_state; 587 588 rc = ibv_modify_qp(rqpair->cm_id->qp, &qp_attr, 589 attr_mask_rc[new_state]); 590 591 if (rc) { 592 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 593 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 594 return rc; 595 } 596 597 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 598 599 if (state != new_state) { 600 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 601 rqpair->qpair.qid, str_ibv_qp_state[new_state], 602 str_ibv_qp_state[state]); 603 return -1; 604 } 605 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 606 str_ibv_qp_state[state]); 607 return 0; 608 } 609 610 static void 611 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 612 struct spdk_nvmf_rdma_transport *rtransport) 613 { 614 struct spdk_nvmf_rdma_request_data *current_data_wr = NULL, *next_data_wr = NULL; 615 struct ibv_send_wr *send_wr; 616 int i; 617 618 rdma_req->num_outstanding_data_wr = 0; 619 current_data_wr = &rdma_req->data; 620 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 621 current_data_wr->wr.sg_list[i].addr = 0; 622 current_data_wr->wr.sg_list[i].length = 0; 623 current_data_wr->wr.sg_list[i].lkey = 0; 624 } 625 current_data_wr->wr.num_sge = 0; 626 627 send_wr = current_data_wr->wr.next; 628 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr) { 629 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 630 } 631 while (next_data_wr) { 632 current_data_wr = next_data_wr; 633 send_wr = current_data_wr->wr.next; 634 if (send_wr != NULL && send_wr != &rdma_req->rsp.wr && 635 send_wr->wr_id == current_data_wr->wr.wr_id) { 636 next_data_wr = SPDK_CONTAINEROF(send_wr, struct spdk_nvmf_rdma_request_data, wr); 637 } else { 638 next_data_wr = NULL; 639 } 640 641 for (i = 0; i < current_data_wr->wr.num_sge; i++) { 642 current_data_wr->wr.sg_list[i].addr = 0; 643 current_data_wr->wr.sg_list[i].length = 0; 644 current_data_wr->wr.sg_list[i].lkey = 0; 645 } 646 current_data_wr->wr.num_sge = 0; 647 spdk_mempool_put(rtransport->data_wr_pool, current_data_wr); 648 } 649 } 650 651 static void 652 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 653 { 654 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->data_from_pool); 655 if (req->req.cmd) { 656 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 657 } 658 if (req->recv) { 659 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 660 } 661 } 662 663 static void 664 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 665 { 666 int i; 667 668 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 669 for (i = 0; i < rqpair->max_queue_depth; i++) { 670 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 671 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 672 } 673 } 674 } 675 676 static void 677 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 678 { 679 if (resources->cmds_mr) { 680 ibv_dereg_mr(resources->cmds_mr); 681 } 682 683 if (resources->cpls_mr) { 684 ibv_dereg_mr(resources->cpls_mr); 685 } 686 687 if (resources->bufs_mr) { 688 ibv_dereg_mr(resources->bufs_mr); 689 } 690 691 spdk_dma_free(resources->cmds); 692 spdk_dma_free(resources->cpls); 693 spdk_dma_free(resources->bufs); 694 free(resources->reqs); 695 free(resources->recvs); 696 free(resources); 697 } 698 699 700 static struct spdk_nvmf_rdma_resources * 701 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 702 { 703 struct spdk_nvmf_rdma_resources *resources; 704 struct spdk_nvmf_rdma_request *rdma_req; 705 struct spdk_nvmf_rdma_recv *rdma_recv; 706 struct ibv_qp *qp; 707 struct ibv_srq *srq; 708 uint32_t i; 709 int rc; 710 711 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 712 if (!resources) { 713 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 714 return NULL; 715 } 716 717 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 718 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 719 resources->cmds = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 720 0x1000, NULL); 721 resources->cpls = spdk_dma_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 722 0x1000, NULL); 723 724 if (opts->in_capsule_data_size > 0) { 725 resources->bufs = spdk_dma_zmalloc(opts->max_queue_depth * 726 opts->in_capsule_data_size, 727 0x1000, NULL); 728 } 729 730 if (!resources->reqs || !resources->recvs || !resources->cmds || 731 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 732 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 733 goto cleanup; 734 } 735 736 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 737 opts->max_queue_depth * sizeof(*resources->cmds), 738 IBV_ACCESS_LOCAL_WRITE); 739 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 740 opts->max_queue_depth * sizeof(*resources->cpls), 741 0); 742 743 if (opts->in_capsule_data_size) { 744 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 745 opts->max_queue_depth * 746 opts->in_capsule_data_size, 747 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 748 } 749 750 if (!resources->cmds_mr || !resources->cpls_mr || 751 (opts->in_capsule_data_size && 752 !resources->bufs_mr)) { 753 goto cleanup; 754 } 755 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 756 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 757 resources->cmds_mr->lkey); 758 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 759 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 760 resources->cpls_mr->lkey); 761 if (resources->bufs && resources->bufs_mr) { 762 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 763 resources->bufs, opts->max_queue_depth * 764 opts->in_capsule_data_size, resources->bufs_mr->lkey); 765 } 766 767 /* Initialize queues */ 768 STAILQ_INIT(&resources->incoming_queue); 769 STAILQ_INIT(&resources->free_queue); 770 771 for (i = 0; i < opts->max_queue_depth; i++) { 772 struct ibv_recv_wr *bad_wr = NULL; 773 774 rdma_recv = &resources->recvs[i]; 775 rdma_recv->qpair = opts->qpair; 776 777 /* Set up memory to receive commands */ 778 if (resources->bufs) { 779 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 780 opts->in_capsule_data_size)); 781 } 782 783 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 784 785 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 786 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 787 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 788 rdma_recv->wr.num_sge = 1; 789 790 if (rdma_recv->buf && resources->bufs_mr) { 791 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 792 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 793 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 794 rdma_recv->wr.num_sge++; 795 } 796 797 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 798 rdma_recv->wr.sg_list = rdma_recv->sgl; 799 if (opts->shared) { 800 srq = (struct ibv_srq *)opts->qp; 801 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 802 } else { 803 qp = (struct ibv_qp *)opts->qp; 804 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 805 } 806 if (rc) { 807 goto cleanup; 808 } 809 } 810 811 for (i = 0; i < opts->max_queue_depth; i++) { 812 rdma_req = &resources->reqs[i]; 813 814 if (opts->qpair != NULL) { 815 rdma_req->req.qpair = &opts->qpair->qpair; 816 } else { 817 rdma_req->req.qpair = NULL; 818 } 819 rdma_req->req.cmd = NULL; 820 821 /* Set up memory to send responses */ 822 rdma_req->req.rsp = &resources->cpls[i]; 823 824 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 825 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 826 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 827 828 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 829 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 830 rdma_req->rsp.wr.next = NULL; 831 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 832 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 833 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 834 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 835 836 /* Set up memory for data buffers */ 837 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 838 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 839 rdma_req->data.wr.next = NULL; 840 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 841 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 842 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 843 844 /* Initialize request state to FREE */ 845 rdma_req->state = RDMA_REQUEST_STATE_FREE; 846 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 847 } 848 849 return resources; 850 851 cleanup: 852 nvmf_rdma_resources_destroy(resources); 853 return NULL; 854 } 855 856 static void 857 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 858 { 859 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 860 struct ibv_recv_wr *bad_recv_wr = NULL; 861 int rc; 862 863 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 864 865 spdk_poller_unregister(&rqpair->destruct_poller); 866 867 if (rqpair->qd != 0) { 868 if (rqpair->srq == NULL) { 869 nvmf_rdma_dump_qpair_contents(rqpair); 870 } 871 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 872 } 873 874 if (rqpair->poller) { 875 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 876 877 if (rqpair->srq != NULL && rqpair->resources != NULL) { 878 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 879 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 880 if (rqpair == rdma_recv->qpair) { 881 STAILQ_REMOVE_HEAD(&rqpair->resources->incoming_queue, link); 882 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 883 if (rc) { 884 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 885 } 886 } 887 } 888 } 889 } 890 891 if (rqpair->cm_id) { 892 if (rqpair->cm_id->qp != NULL) { 893 rdma_destroy_qp(rqpair->cm_id); 894 } 895 rdma_destroy_id(rqpair->cm_id); 896 897 if (rqpair->poller != NULL && rqpair->srq == NULL) { 898 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 899 } 900 } 901 902 if (rqpair->srq == NULL && rqpair->resources != NULL) { 903 nvmf_rdma_resources_destroy(rqpair->resources); 904 } 905 906 free(rqpair); 907 } 908 909 static int 910 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 911 { 912 struct spdk_nvmf_rdma_poller *rpoller; 913 int rc, num_cqe, required_num_wr; 914 915 /* Enlarge CQ size dynamically */ 916 rpoller = rqpair->poller; 917 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 918 num_cqe = rpoller->num_cqe; 919 if (num_cqe < required_num_wr) { 920 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 921 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 922 } 923 924 if (rpoller->num_cqe != num_cqe) { 925 if (required_num_wr > device->attr.max_cqe) { 926 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 927 required_num_wr, device->attr.max_cqe); 928 return -1; 929 } 930 931 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 932 rc = ibv_resize_cq(rpoller->cq, num_cqe); 933 if (rc) { 934 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 935 return -1; 936 } 937 938 rpoller->num_cqe = num_cqe; 939 } 940 941 rpoller->required_num_wr = required_num_wr; 942 return 0; 943 } 944 945 static int 946 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 947 { 948 struct spdk_nvmf_rdma_qpair *rqpair; 949 int rc; 950 struct spdk_nvmf_rdma_transport *rtransport; 951 struct spdk_nvmf_transport *transport; 952 struct spdk_nvmf_rdma_resource_opts opts; 953 struct spdk_nvmf_rdma_device *device; 954 struct ibv_qp_init_attr ibv_init_attr; 955 956 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 957 device = rqpair->port->device; 958 959 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 960 ibv_init_attr.qp_context = rqpair; 961 ibv_init_attr.qp_type = IBV_QPT_RC; 962 ibv_init_attr.send_cq = rqpair->poller->cq; 963 ibv_init_attr.recv_cq = rqpair->poller->cq; 964 965 if (rqpair->srq) { 966 ibv_init_attr.srq = rqpair->srq; 967 } else { 968 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 969 1; /* RECV operations + dummy drain WR */ 970 } 971 972 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 973 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 974 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 975 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 976 977 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 978 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 979 goto error; 980 } 981 982 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 983 if (rc) { 984 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 985 goto error; 986 } 987 988 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2 + 1), 989 ibv_init_attr.cap.max_send_wr); 990 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 991 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 992 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 993 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 994 995 rqpair->sends_to_post.first = NULL; 996 rqpair->sends_to_post.last = NULL; 997 998 if (rqpair->poller->srq == NULL) { 999 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1000 transport = &rtransport->transport; 1001 1002 opts.qp = rqpair->cm_id->qp; 1003 opts.pd = rqpair->cm_id->pd; 1004 opts.qpair = rqpair; 1005 opts.shared = false; 1006 opts.max_queue_depth = rqpair->max_queue_depth; 1007 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1008 1009 rqpair->resources = nvmf_rdma_resources_create(&opts); 1010 1011 if (!rqpair->resources) { 1012 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1013 goto error; 1014 } 1015 } else { 1016 rqpair->resources = rqpair->poller->resources; 1017 } 1018 1019 rqpair->current_recv_depth = 0; 1020 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1021 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1022 1023 return 0; 1024 1025 error: 1026 rdma_destroy_id(rqpair->cm_id); 1027 rqpair->cm_id = NULL; 1028 spdk_nvmf_rdma_qpair_destroy(rqpair); 1029 return -1; 1030 } 1031 1032 1033 /* Append the given send wr structure to the qpair's outstanding sends list. */ 1034 /* This function accepts either a single wr or the first wr in a linked list. */ 1035 static void 1036 nvmf_rdma_qpair_queue_send_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *first) 1037 { 1038 struct ibv_send_wr *last; 1039 1040 last = first; 1041 while (last->next != NULL) { 1042 last = last->next; 1043 } 1044 1045 if (rqpair->sends_to_post.first == NULL) { 1046 rqpair->sends_to_post.first = first; 1047 rqpair->sends_to_post.last = last; 1048 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1049 } else { 1050 rqpair->sends_to_post.last->next = first; 1051 rqpair->sends_to_post.last = last; 1052 } 1053 } 1054 1055 static int 1056 request_transfer_in(struct spdk_nvmf_request *req) 1057 { 1058 struct spdk_nvmf_rdma_request *rdma_req; 1059 struct spdk_nvmf_qpair *qpair; 1060 struct spdk_nvmf_rdma_qpair *rqpair; 1061 1062 qpair = req->qpair; 1063 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1064 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1065 1066 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1067 assert(rdma_req != NULL); 1068 1069 nvmf_rdma_qpair_queue_send_wrs(rqpair, &rdma_req->data.wr); 1070 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1071 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1072 return 0; 1073 } 1074 1075 static int 1076 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1077 { 1078 int rc; 1079 int num_outstanding_data_wr = 0; 1080 struct spdk_nvmf_rdma_request *rdma_req; 1081 struct spdk_nvmf_qpair *qpair; 1082 struct spdk_nvmf_rdma_qpair *rqpair; 1083 struct spdk_nvme_cpl *rsp; 1084 struct ibv_recv_wr *bad_recv_wr = NULL; 1085 struct ibv_send_wr *first = NULL; 1086 1087 *data_posted = 0; 1088 qpair = req->qpair; 1089 rsp = &req->rsp->nvme_cpl; 1090 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1091 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1092 1093 /* Advance our sq_head pointer */ 1094 if (qpair->sq_head == qpair->sq_head_max) { 1095 qpair->sq_head = 0; 1096 } else { 1097 qpair->sq_head++; 1098 } 1099 rsp->sqhd = qpair->sq_head; 1100 1101 /* Post the capsule to the recv buffer */ 1102 assert(rdma_req->recv != NULL); 1103 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "RDMA RECV POSTED. Recv: %p Connection: %p\n", rdma_req->recv, 1104 rqpair); 1105 if (rqpair->srq == NULL) { 1106 rc = ibv_post_recv(rqpair->cm_id->qp, &rdma_req->recv->wr, &bad_recv_wr); 1107 } else { 1108 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 1109 } 1110 1111 if (rc) { 1112 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 1113 return rc; 1114 } 1115 rdma_req->recv = NULL; 1116 assert(rqpair->current_recv_depth > 0); 1117 rqpair->current_recv_depth--; 1118 1119 /* Build the response which consists of optional 1120 * RDMA WRITEs to transfer data, plus an RDMA SEND 1121 * containing the response. 1122 */ 1123 first = &rdma_req->rsp.wr; 1124 1125 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 1126 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1127 first = &rdma_req->data.wr; 1128 *data_posted = 1; 1129 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1130 } 1131 nvmf_rdma_qpair_queue_send_wrs(rqpair, first); 1132 /* +1 for the rsp wr */ 1133 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1134 1135 return 0; 1136 } 1137 1138 static int 1139 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1140 { 1141 struct spdk_nvmf_rdma_accept_private_data accept_data; 1142 struct rdma_conn_param ctrlr_event_data = {}; 1143 int rc; 1144 1145 accept_data.recfmt = 0; 1146 accept_data.crqsize = rqpair->max_queue_depth; 1147 1148 ctrlr_event_data.private_data = &accept_data; 1149 ctrlr_event_data.private_data_len = sizeof(accept_data); 1150 if (id->ps == RDMA_PS_TCP) { 1151 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1152 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1153 } 1154 1155 /* Configure infinite retries for the initiator side qpair. 1156 * When using a shared receive queue on the target side, 1157 * we need to pass this value to the initiator to prevent the 1158 * initiator side NIC from completing SEND requests back to the 1159 * initiator with status rnr_retry_count_exceeded. */ 1160 if (rqpair->srq != NULL) { 1161 ctrlr_event_data.rnr_retry_count = 0x7; 1162 } 1163 1164 rc = rdma_accept(id, &ctrlr_event_data); 1165 if (rc) { 1166 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 1167 } else { 1168 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1169 } 1170 1171 return rc; 1172 } 1173 1174 static void 1175 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1176 { 1177 struct spdk_nvmf_rdma_reject_private_data rej_data; 1178 1179 rej_data.recfmt = 0; 1180 rej_data.sts = error; 1181 1182 rdma_reject(id, &rej_data, sizeof(rej_data)); 1183 } 1184 1185 static int 1186 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 1187 new_qpair_fn cb_fn) 1188 { 1189 struct spdk_nvmf_rdma_transport *rtransport; 1190 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1191 struct spdk_nvmf_rdma_port *port; 1192 struct rdma_conn_param *rdma_param = NULL; 1193 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1194 uint16_t max_queue_depth; 1195 uint16_t max_read_depth; 1196 1197 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1198 1199 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1200 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1201 1202 rdma_param = &event->param.conn; 1203 if (rdma_param->private_data == NULL || 1204 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1205 SPDK_ERRLOG("connect request: no private data provided\n"); 1206 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1207 return -1; 1208 } 1209 1210 private_data = rdma_param->private_data; 1211 if (private_data->recfmt != 0) { 1212 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1213 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1214 return -1; 1215 } 1216 1217 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1218 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1219 1220 port = event->listen_id->context; 1221 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1222 event->listen_id, event->listen_id->verbs, port); 1223 1224 /* Figure out the supported queue depth. This is a multi-step process 1225 * that takes into account hardware maximums, host provided values, 1226 * and our target's internal memory limits */ 1227 1228 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1229 1230 /* Start with the maximum queue depth allowed by the target */ 1231 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1232 max_read_depth = rtransport->transport.opts.max_queue_depth; 1233 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1234 rtransport->transport.opts.max_queue_depth); 1235 1236 /* Next check the local NIC's hardware limitations */ 1237 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1238 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1239 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1240 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1241 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1242 1243 /* Next check the remote NIC's hardware limitations */ 1244 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1245 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1246 rdma_param->initiator_depth, rdma_param->responder_resources); 1247 if (rdma_param->initiator_depth > 0) { 1248 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1249 } 1250 1251 /* Finally check for the host software requested values, which are 1252 * optional. */ 1253 if (rdma_param->private_data != NULL && 1254 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1255 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1256 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1257 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1258 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1259 } 1260 1261 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1262 max_queue_depth, max_read_depth); 1263 1264 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1265 if (rqpair == NULL) { 1266 SPDK_ERRLOG("Could not allocate new connection.\n"); 1267 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1268 return -1; 1269 } 1270 1271 rqpair->port = port; 1272 rqpair->max_queue_depth = max_queue_depth; 1273 rqpair->max_read_depth = max_read_depth; 1274 rqpair->cm_id = event->id; 1275 rqpair->listen_id = event->listen_id; 1276 rqpair->qpair.transport = transport; 1277 1278 event->id->context = &rqpair->qpair; 1279 1280 cb_fn(&rqpair->qpair); 1281 1282 return 0; 1283 } 1284 1285 static int 1286 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1287 enum spdk_mem_map_notify_action action, 1288 void *vaddr, size_t size) 1289 { 1290 struct ibv_pd *pd = cb_ctx; 1291 struct ibv_mr *mr; 1292 1293 switch (action) { 1294 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1295 if (!g_nvmf_hooks.get_rkey) { 1296 mr = ibv_reg_mr(pd, vaddr, size, 1297 IBV_ACCESS_LOCAL_WRITE | 1298 IBV_ACCESS_REMOTE_READ | 1299 IBV_ACCESS_REMOTE_WRITE); 1300 if (mr == NULL) { 1301 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1302 return -1; 1303 } else { 1304 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1305 } 1306 } else { 1307 spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1308 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1309 } 1310 break; 1311 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1312 if (!g_nvmf_hooks.get_rkey) { 1313 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1314 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1315 if (mr) { 1316 ibv_dereg_mr(mr); 1317 } 1318 } 1319 break; 1320 } 1321 1322 return 0; 1323 } 1324 1325 static int 1326 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1327 { 1328 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1329 return addr_1 == addr_2; 1330 } 1331 1332 static void 1333 spdk_nvmf_rdma_request_free_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1334 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1335 uint32_t num_buffers) 1336 { 1337 uint32_t i; 1338 1339 for (i = 0; i < num_buffers; i++) { 1340 if (group->buf_cache_count < group->buf_cache_size) { 1341 STAILQ_INSERT_HEAD(&group->buf_cache, 1342 (struct spdk_nvmf_transport_pg_cache_buf *)rdma_req->buffers[i], link); 1343 group->buf_cache_count++; 1344 } else { 1345 spdk_mempool_put(transport->data_buf_pool, rdma_req->buffers[i]); 1346 } 1347 rdma_req->req.iov[i].iov_base = NULL; 1348 rdma_req->buffers[i] = NULL; 1349 rdma_req->req.iov[i].iov_len = 0; 1350 1351 } 1352 rdma_req->data_from_pool = false; 1353 } 1354 1355 static int 1356 nvmf_rdma_request_get_buffers(struct spdk_nvmf_rdma_request *rdma_req, 1357 struct spdk_nvmf_transport_poll_group *group, struct spdk_nvmf_transport *transport, 1358 uint32_t num_buffers) 1359 { 1360 uint32_t i = 0; 1361 1362 while (i < num_buffers) { 1363 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1364 group->buf_cache_count--; 1365 rdma_req->buffers[i] = STAILQ_FIRST(&group->buf_cache); 1366 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1367 assert(rdma_req->buffers[i] != NULL); 1368 i++; 1369 } else { 1370 if (spdk_mempool_get_bulk(transport->data_buf_pool, &rdma_req->buffers[i], num_buffers - i)) { 1371 goto err_exit; 1372 } 1373 i += num_buffers - i; 1374 } 1375 } 1376 1377 return 0; 1378 1379 err_exit: 1380 spdk_nvmf_rdma_request_free_buffers(rdma_req, group, transport, i); 1381 return -ENOMEM; 1382 } 1383 1384 typedef enum spdk_nvme_data_transfer spdk_nvme_data_transfer_t; 1385 1386 static spdk_nvme_data_transfer_t 1387 spdk_nvmf_rdma_request_get_xfer(struct spdk_nvmf_rdma_request *rdma_req) 1388 { 1389 enum spdk_nvme_data_transfer xfer; 1390 struct spdk_nvme_cmd *cmd = &rdma_req->req.cmd->nvme_cmd; 1391 struct spdk_nvme_sgl_descriptor *sgl = &cmd->dptr.sgl1; 1392 1393 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1394 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1395 rdma_req->rsp.wr.imm_data = 0; 1396 #endif 1397 1398 /* Figure out data transfer direction */ 1399 if (cmd->opc == SPDK_NVME_OPC_FABRIC) { 1400 xfer = spdk_nvme_opc_get_data_transfer(rdma_req->req.cmd->nvmf_cmd.fctype); 1401 } else { 1402 xfer = spdk_nvme_opc_get_data_transfer(cmd->opc); 1403 1404 /* Some admin commands are special cases */ 1405 if ((rdma_req->req.qpair->qid == 0) && 1406 ((cmd->opc == SPDK_NVME_OPC_GET_FEATURES) || 1407 (cmd->opc == SPDK_NVME_OPC_SET_FEATURES))) { 1408 switch (cmd->cdw10 & 0xff) { 1409 case SPDK_NVME_FEAT_LBA_RANGE_TYPE: 1410 case SPDK_NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION: 1411 case SPDK_NVME_FEAT_HOST_IDENTIFIER: 1412 break; 1413 default: 1414 xfer = SPDK_NVME_DATA_NONE; 1415 } 1416 } 1417 } 1418 1419 if (xfer == SPDK_NVME_DATA_NONE) { 1420 return xfer; 1421 } 1422 1423 /* Even for commands that may transfer data, they could have specified 0 length. 1424 * We want those to show up with xfer SPDK_NVME_DATA_NONE. 1425 */ 1426 switch (sgl->generic.type) { 1427 case SPDK_NVME_SGL_TYPE_DATA_BLOCK: 1428 case SPDK_NVME_SGL_TYPE_BIT_BUCKET: 1429 case SPDK_NVME_SGL_TYPE_SEGMENT: 1430 case SPDK_NVME_SGL_TYPE_LAST_SEGMENT: 1431 case SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK: 1432 if (sgl->unkeyed.length == 0) { 1433 xfer = SPDK_NVME_DATA_NONE; 1434 } 1435 break; 1436 case SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK: 1437 if (sgl->keyed.length == 0) { 1438 xfer = SPDK_NVME_DATA_NONE; 1439 } 1440 break; 1441 } 1442 1443 return xfer; 1444 } 1445 1446 static int 1447 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1448 struct spdk_nvmf_rdma_request *rdma_req, 1449 uint32_t num_sgl_descriptors) 1450 { 1451 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1452 struct spdk_nvmf_rdma_request_data *current_data_wr; 1453 uint32_t i; 1454 1455 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1456 return -ENOMEM; 1457 } 1458 1459 current_data_wr = &rdma_req->data; 1460 1461 for (i = 0; i < num_sgl_descriptors; i++) { 1462 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1463 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1464 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1465 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1466 } else { 1467 assert(false); 1468 } 1469 work_requests[i]->wr.send_flags = IBV_SEND_SIGNALED; 1470 work_requests[i]->wr.sg_list = work_requests[i]->sgl; 1471 work_requests[i]->wr.wr_id = rdma_req->data.wr.wr_id; 1472 current_data_wr->wr.next = &work_requests[i]->wr; 1473 current_data_wr = work_requests[i]; 1474 } 1475 1476 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1477 current_data_wr->wr.opcode = IBV_WR_RDMA_WRITE; 1478 current_data_wr->wr.next = &rdma_req->rsp.wr; 1479 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1480 current_data_wr->wr.opcode = IBV_WR_RDMA_READ; 1481 current_data_wr->wr.next = NULL; 1482 } 1483 return 0; 1484 } 1485 1486 static int 1487 nvmf_rdma_fill_buffers(struct spdk_nvmf_rdma_transport *rtransport, 1488 struct spdk_nvmf_rdma_poll_group *rgroup, 1489 struct spdk_nvmf_rdma_device *device, 1490 struct spdk_nvmf_rdma_request *rdma_req, 1491 struct ibv_send_wr *wr, 1492 uint32_t length) 1493 { 1494 uint64_t translation_len; 1495 uint32_t remaining_length = length; 1496 uint32_t iovcnt; 1497 uint32_t i = 0; 1498 1499 1500 while (remaining_length) { 1501 iovcnt = rdma_req->req.iovcnt; 1502 rdma_req->req.iov[iovcnt].iov_base = (void *)((uintptr_t)(rdma_req->buffers[iovcnt] + 1503 NVMF_DATA_BUFFER_MASK) & 1504 ~NVMF_DATA_BUFFER_MASK); 1505 rdma_req->req.iov[iovcnt].iov_len = spdk_min(remaining_length, 1506 rtransport->transport.opts.io_unit_size); 1507 rdma_req->req.iovcnt++; 1508 wr->sg_list[i].addr = (uintptr_t)(rdma_req->req.iov[iovcnt].iov_base); 1509 wr->sg_list[i].length = rdma_req->req.iov[iovcnt].iov_len; 1510 translation_len = rdma_req->req.iov[iovcnt].iov_len; 1511 1512 if (!g_nvmf_hooks.get_rkey) { 1513 wr->sg_list[i].lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1514 (uint64_t)rdma_req->buffers[iovcnt], &translation_len))->lkey; 1515 } else { 1516 wr->sg_list[i].lkey = spdk_mem_map_translate(device->map, 1517 (uint64_t)rdma_req->buffers[iovcnt], &translation_len); 1518 } 1519 1520 remaining_length -= rdma_req->req.iov[iovcnt].iov_len; 1521 1522 if (translation_len < rdma_req->req.iov[iovcnt].iov_len) { 1523 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions\n"); 1524 return -EINVAL; 1525 } 1526 i++; 1527 } 1528 wr->num_sge = i; 1529 1530 return 0; 1531 } 1532 1533 static int 1534 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1535 struct spdk_nvmf_rdma_device *device, 1536 struct spdk_nvmf_rdma_request *rdma_req) 1537 { 1538 struct spdk_nvmf_rdma_qpair *rqpair; 1539 struct spdk_nvmf_rdma_poll_group *rgroup; 1540 uint32_t num_buffers; 1541 uint32_t i = 0; 1542 int rc = 0; 1543 1544 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1545 rgroup = rqpair->poller->group; 1546 rdma_req->req.iovcnt = 0; 1547 1548 num_buffers = rdma_req->req.length / rtransport->transport.opts.io_unit_size; 1549 if (rdma_req->req.length % rtransport->transport.opts.io_unit_size) { 1550 num_buffers++; 1551 } 1552 1553 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers)) { 1554 return -ENOMEM; 1555 } 1556 1557 rdma_req->req.iovcnt = 0; 1558 1559 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, &rdma_req->data.wr, 1560 rdma_req->req.length); 1561 if (rc != 0) { 1562 goto err_exit; 1563 } 1564 1565 assert(rdma_req->req.iovcnt <= rqpair->max_send_sge); 1566 1567 rdma_req->data_from_pool = true; 1568 1569 return rc; 1570 1571 err_exit: 1572 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1573 while (i) { 1574 i--; 1575 rdma_req->data.wr.sg_list[i].addr = 0; 1576 rdma_req->data.wr.sg_list[i].length = 0; 1577 rdma_req->data.wr.sg_list[i].lkey = 0; 1578 } 1579 rdma_req->req.iovcnt = 0; 1580 return rc; 1581 } 1582 1583 static int 1584 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1585 struct spdk_nvmf_rdma_device *device, 1586 struct spdk_nvmf_rdma_request *rdma_req) 1587 { 1588 struct spdk_nvmf_rdma_qpair *rqpair; 1589 struct spdk_nvmf_rdma_poll_group *rgroup; 1590 struct ibv_send_wr *current_wr; 1591 struct spdk_nvmf_request *req = &rdma_req->req; 1592 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1593 uint32_t num_sgl_descriptors; 1594 uint32_t num_buffers = 0; 1595 uint32_t i; 1596 int rc; 1597 1598 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1599 rgroup = rqpair->poller->group; 1600 1601 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1602 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1603 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1604 1605 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1606 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1607 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1608 1609 for (i = 0; i < num_sgl_descriptors; i++) { 1610 num_buffers += desc->keyed.length / rtransport->transport.opts.io_unit_size; 1611 if (desc->keyed.length % rtransport->transport.opts.io_unit_size) { 1612 num_buffers++; 1613 } 1614 desc++; 1615 } 1616 /* If the number of buffers is too large, then we know the I/O is larger than allowed. Fail it. */ 1617 if (num_buffers > NVMF_REQ_MAX_BUFFERS) { 1618 return -EINVAL; 1619 } 1620 if (nvmf_rdma_request_get_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1621 num_buffers) != 0) { 1622 return -ENOMEM; 1623 } 1624 1625 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1626 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1627 return -ENOMEM; 1628 } 1629 1630 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1631 current_wr = &rdma_req->data.wr; 1632 1633 req->iovcnt = 0; 1634 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1635 for (i = 0; i < num_sgl_descriptors; i++) { 1636 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1637 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1638 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1639 rc = -EINVAL; 1640 goto err_exit; 1641 } 1642 1643 current_wr->num_sge = 0; 1644 req->length += desc->keyed.length; 1645 1646 rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, rdma_req, current_wr, 1647 desc->keyed.length); 1648 if (rc != 0) { 1649 rc = -ENOMEM; 1650 goto err_exit; 1651 } 1652 1653 current_wr->wr.rdma.rkey = desc->keyed.key; 1654 current_wr->wr.rdma.remote_addr = desc->address; 1655 current_wr = current_wr->next; 1656 desc++; 1657 } 1658 1659 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1660 /* Go back to the last descriptor in the list. */ 1661 desc--; 1662 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1663 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1664 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1665 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1666 } 1667 } 1668 #endif 1669 1670 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1671 rdma_req->data_from_pool = true; 1672 1673 return 0; 1674 1675 err_exit: 1676 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, num_buffers); 1677 nvmf_rdma_request_free_data(rdma_req, rtransport); 1678 return rc; 1679 } 1680 1681 static int 1682 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1683 struct spdk_nvmf_rdma_device *device, 1684 struct spdk_nvmf_rdma_request *rdma_req) 1685 { 1686 struct spdk_nvme_cmd *cmd; 1687 struct spdk_nvme_cpl *rsp; 1688 struct spdk_nvme_sgl_descriptor *sgl; 1689 int rc; 1690 1691 cmd = &rdma_req->req.cmd->nvme_cmd; 1692 rsp = &rdma_req->req.rsp->nvme_cpl; 1693 sgl = &cmd->dptr.sgl1; 1694 1695 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1696 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1697 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1698 if (sgl->keyed.length > rtransport->transport.opts.max_io_size) { 1699 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1700 sgl->keyed.length, rtransport->transport.opts.max_io_size); 1701 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1702 return -1; 1703 } 1704 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1705 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1706 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1707 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1708 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1709 } 1710 } 1711 #endif 1712 1713 /* fill request length and populate iovs */ 1714 rdma_req->req.length = sgl->keyed.length; 1715 1716 if (spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req) < 0) { 1717 /* No available buffers. Queue this request up. */ 1718 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1719 return 0; 1720 } 1721 1722 /* backward compatible */ 1723 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1724 1725 /* rdma wr specifics */ 1726 rdma_req->data.wr.num_sge = rdma_req->req.iovcnt; 1727 rdma_req->data.wr.wr.rdma.rkey = sgl->keyed.key; 1728 rdma_req->data.wr.wr.rdma.remote_addr = sgl->address; 1729 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1730 rdma_req->data.wr.opcode = IBV_WR_RDMA_WRITE; 1731 rdma_req->data.wr.next = &rdma_req->rsp.wr; 1732 } else if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1733 rdma_req->data.wr.opcode = IBV_WR_RDMA_READ; 1734 rdma_req->data.wr.next = NULL; 1735 } 1736 1737 /* set the number of outstanding data WRs for this request. */ 1738 rdma_req->num_outstanding_data_wr = 1; 1739 1740 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1741 rdma_req->req.iovcnt); 1742 1743 return 0; 1744 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1745 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1746 uint64_t offset = sgl->address; 1747 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1748 1749 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1750 offset, sgl->unkeyed.length); 1751 1752 if (offset > max_len) { 1753 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1754 offset, max_len); 1755 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1756 return -1; 1757 } 1758 max_len -= (uint32_t)offset; 1759 1760 if (sgl->unkeyed.length > max_len) { 1761 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1762 sgl->unkeyed.length, max_len); 1763 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1764 return -1; 1765 } 1766 1767 rdma_req->num_outstanding_data_wr = 0; 1768 rdma_req->req.data = rdma_req->recv->buf + offset; 1769 rdma_req->data_from_pool = false; 1770 rdma_req->req.length = sgl->unkeyed.length; 1771 1772 rdma_req->req.iov[0].iov_base = rdma_req->req.data; 1773 rdma_req->req.iov[0].iov_len = rdma_req->req.length; 1774 rdma_req->req.iovcnt = 1; 1775 1776 return 0; 1777 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1778 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1779 1780 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1781 if (rc == -ENOMEM) { 1782 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1783 return 0; 1784 } else if (rc == -EINVAL) { 1785 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1786 return -1; 1787 } 1788 1789 /* backward compatible */ 1790 rdma_req->req.data = rdma_req->req.iov[0].iov_base; 1791 1792 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1793 rdma_req->req.iovcnt); 1794 1795 return 0; 1796 } 1797 1798 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1799 sgl->generic.type, sgl->generic.subtype); 1800 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1801 return -1; 1802 } 1803 1804 static void 1805 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1806 struct spdk_nvmf_rdma_transport *rtransport) 1807 { 1808 struct spdk_nvmf_rdma_qpair *rqpair; 1809 struct spdk_nvmf_rdma_poll_group *rgroup; 1810 1811 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1812 if (rdma_req->data_from_pool) { 1813 rgroup = rqpair->poller->group; 1814 1815 spdk_nvmf_rdma_request_free_buffers(rdma_req, &rgroup->group, &rtransport->transport, 1816 rdma_req->req.iovcnt); 1817 } 1818 nvmf_rdma_request_free_data(rdma_req, rtransport); 1819 rdma_req->req.length = 0; 1820 rdma_req->req.iovcnt = 0; 1821 rdma_req->req.data = NULL; 1822 rdma_req->rsp.wr.next = NULL; 1823 rdma_req->data.wr.next = NULL; 1824 rqpair->qd--; 1825 1826 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1827 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1828 } 1829 1830 static bool 1831 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1832 struct spdk_nvmf_rdma_request *rdma_req) 1833 { 1834 struct spdk_nvmf_rdma_qpair *rqpair; 1835 struct spdk_nvmf_rdma_device *device; 1836 struct spdk_nvmf_rdma_poll_group *rgroup; 1837 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1838 int rc; 1839 struct spdk_nvmf_rdma_recv *rdma_recv; 1840 enum spdk_nvmf_rdma_request_state prev_state; 1841 bool progress = false; 1842 int data_posted; 1843 1844 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1845 device = rqpair->port->device; 1846 rgroup = rqpair->poller->group; 1847 1848 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1849 1850 /* If the queue pair is in an error state, force the request to the completed state 1851 * to release resources. */ 1852 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1853 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1854 STAILQ_REMOVE(&rgroup->pending_data_buf_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1855 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1856 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1857 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1858 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1859 } 1860 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1861 } 1862 1863 /* The loop here is to allow for several back-to-back state changes. */ 1864 do { 1865 prev_state = rdma_req->state; 1866 1867 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1868 1869 switch (rdma_req->state) { 1870 case RDMA_REQUEST_STATE_FREE: 1871 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1872 * to escape this state. */ 1873 break; 1874 case RDMA_REQUEST_STATE_NEW: 1875 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1876 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1877 rdma_recv = rdma_req->recv; 1878 1879 /* The first element of the SGL is the NVMe command */ 1880 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1881 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1882 1883 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1884 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1885 break; 1886 } 1887 1888 /* The next state transition depends on the data transfer needs of this request. */ 1889 rdma_req->req.xfer = spdk_nvmf_rdma_request_get_xfer(rdma_req); 1890 1891 /* If no data to transfer, ready to execute. */ 1892 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1893 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1894 break; 1895 } 1896 1897 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1898 STAILQ_INSERT_TAIL(&rgroup->pending_data_buf_queue, rdma_req, state_link); 1899 break; 1900 case RDMA_REQUEST_STATE_NEED_BUFFER: 1901 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1902 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1903 1904 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1905 1906 if (rdma_req != STAILQ_FIRST(&rgroup->pending_data_buf_queue)) { 1907 /* This request needs to wait in line to obtain a buffer */ 1908 break; 1909 } 1910 1911 /* Try to get a data buffer */ 1912 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1913 if (rc < 0) { 1914 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1915 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1916 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1917 break; 1918 } 1919 1920 if (!rdma_req->req.data) { 1921 /* No buffers available. */ 1922 break; 1923 } 1924 1925 STAILQ_REMOVE_HEAD(&rgroup->pending_data_buf_queue, state_link); 1926 1927 /* If data is transferring from host to controller and the data didn't 1928 * arrive using in capsule data, we need to do a transfer from the host. 1929 */ 1930 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && rdma_req->data_from_pool) { 1931 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1932 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1933 break; 1934 } 1935 1936 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1937 break; 1938 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1939 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1940 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1941 1942 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1943 /* This request needs to wait in line to perform RDMA */ 1944 break; 1945 } 1946 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1947 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1948 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1949 break; 1950 } 1951 1952 /* We have already verified that this request is the head of the queue. */ 1953 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 1954 1955 rc = request_transfer_in(&rdma_req->req); 1956 if (!rc) { 1957 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 1958 } else { 1959 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 1960 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1961 } 1962 break; 1963 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 1964 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 1965 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1966 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 1967 * to escape this state. */ 1968 break; 1969 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 1970 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 1971 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1972 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 1973 spdk_nvmf_request_exec(&rdma_req->req); 1974 break; 1975 case RDMA_REQUEST_STATE_EXECUTING: 1976 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 1977 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1978 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 1979 * to escape this state. */ 1980 break; 1981 case RDMA_REQUEST_STATE_EXECUTED: 1982 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 1983 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1984 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1985 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 1986 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 1987 } else { 1988 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1989 } 1990 break; 1991 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 1992 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 1993 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1994 1995 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 1996 /* This request needs to wait in line to perform RDMA */ 1997 break; 1998 } 1999 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2000 rqpair->max_send_depth) { 2001 /* We can only have so many WRs outstanding. we have to wait until some finish. 2002 * +1 since each request has an additional wr in the resp. */ 2003 break; 2004 } 2005 2006 /* We have already verified that this request is the head of the queue. */ 2007 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2008 2009 /* The data transfer will be kicked off from 2010 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2011 */ 2012 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2013 break; 2014 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2015 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2016 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2017 rc = request_transfer_out(&rdma_req->req, &data_posted); 2018 assert(rc == 0); /* No good way to handle this currently */ 2019 if (rc) { 2020 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2021 } else { 2022 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2023 RDMA_REQUEST_STATE_COMPLETING; 2024 } 2025 break; 2026 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2027 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2028 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2029 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2030 * to escape this state. */ 2031 break; 2032 case RDMA_REQUEST_STATE_COMPLETING: 2033 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2034 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2035 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2036 * to escape this state. */ 2037 break; 2038 case RDMA_REQUEST_STATE_COMPLETED: 2039 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2040 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2041 2042 nvmf_rdma_request_free(rdma_req, rtransport); 2043 break; 2044 case RDMA_REQUEST_NUM_STATES: 2045 default: 2046 assert(0); 2047 break; 2048 } 2049 2050 if (rdma_req->state != prev_state) { 2051 progress = true; 2052 } 2053 } while (rdma_req->state != prev_state); 2054 2055 return progress; 2056 } 2057 2058 /* Public API callbacks begin here */ 2059 2060 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2061 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2062 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2063 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 64 2064 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2065 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2066 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2067 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2068 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2069 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false; 2070 2071 static void 2072 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2073 { 2074 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2075 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2076 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2077 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2078 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2079 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2080 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2081 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2082 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2083 opts->no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ 2084 } 2085 2086 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2087 .notify_cb = spdk_nvmf_rdma_mem_notify, 2088 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 2089 }; 2090 2091 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2092 2093 static struct spdk_nvmf_transport * 2094 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2095 { 2096 int rc; 2097 struct spdk_nvmf_rdma_transport *rtransport; 2098 struct spdk_nvmf_rdma_device *device, *tmp; 2099 struct ibv_pd *pd; 2100 struct ibv_context **contexts; 2101 uint32_t i; 2102 int flag; 2103 uint32_t sge_count; 2104 uint32_t min_shared_buffers; 2105 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2106 2107 rtransport = calloc(1, sizeof(*rtransport)); 2108 if (!rtransport) { 2109 return NULL; 2110 } 2111 2112 if (pthread_mutex_init(&rtransport->lock, NULL)) { 2113 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2114 free(rtransport); 2115 return NULL; 2116 } 2117 2118 TAILQ_INIT(&rtransport->devices); 2119 TAILQ_INIT(&rtransport->ports); 2120 2121 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2122 2123 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2124 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2125 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2126 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2127 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d\n", 2128 opts->max_queue_depth, 2129 opts->max_io_size, 2130 opts->max_qpairs_per_ctrlr, 2131 opts->io_unit_size, 2132 opts->in_capsule_data_size, 2133 opts->max_aq_depth, 2134 opts->num_shared_buffers, 2135 opts->max_srq_depth, 2136 opts->no_srq); 2137 2138 /* I/O unit size cannot be larger than max I/O size */ 2139 if (opts->io_unit_size > opts->max_io_size) { 2140 opts->io_unit_size = opts->max_io_size; 2141 } 2142 2143 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2144 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2145 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2146 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2147 spdk_nvmf_rdma_destroy(&rtransport->transport); 2148 return NULL; 2149 } 2150 2151 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2152 if (min_shared_buffers > opts->num_shared_buffers) { 2153 SPDK_ERRLOG("There are not enough buffers to satisfy" 2154 "per-poll group caches for each thread. (%" PRIu32 ")" 2155 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2156 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2157 spdk_nvmf_rdma_destroy(&rtransport->transport); 2158 return NULL; 2159 } 2160 2161 sge_count = opts->max_io_size / opts->io_unit_size; 2162 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2163 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2164 spdk_nvmf_rdma_destroy(&rtransport->transport); 2165 return NULL; 2166 } 2167 2168 rtransport->event_channel = rdma_create_event_channel(); 2169 if (rtransport->event_channel == NULL) { 2170 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2171 spdk_nvmf_rdma_destroy(&rtransport->transport); 2172 return NULL; 2173 } 2174 2175 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2176 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2177 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2178 rtransport->event_channel->fd, spdk_strerror(errno)); 2179 spdk_nvmf_rdma_destroy(&rtransport->transport); 2180 return NULL; 2181 } 2182 2183 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2184 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2185 sizeof(struct spdk_nvmf_rdma_request_data), 2186 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2187 SPDK_ENV_SOCKET_ID_ANY); 2188 if (!rtransport->data_wr_pool) { 2189 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2190 spdk_nvmf_rdma_destroy(&rtransport->transport); 2191 return NULL; 2192 } 2193 2194 contexts = rdma_get_devices(NULL); 2195 if (contexts == NULL) { 2196 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2197 spdk_nvmf_rdma_destroy(&rtransport->transport); 2198 return NULL; 2199 } 2200 2201 i = 0; 2202 rc = 0; 2203 while (contexts[i] != NULL) { 2204 device = calloc(1, sizeof(*device)); 2205 if (!device) { 2206 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2207 rc = -ENOMEM; 2208 break; 2209 } 2210 device->context = contexts[i]; 2211 rc = ibv_query_device(device->context, &device->attr); 2212 if (rc < 0) { 2213 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2214 free(device); 2215 break; 2216 2217 } 2218 2219 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2220 2221 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2222 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2223 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2224 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2225 } 2226 2227 /** 2228 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2229 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2230 * but incorrectly reports that it does. There are changes making their way 2231 * through the kernel now that will enable this feature. When they are merged, 2232 * we can conditionally enable this feature. 2233 * 2234 * TODO: enable this for versions of the kernel rxe driver that support it. 2235 */ 2236 if (device->attr.vendor_id == 0) { 2237 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2238 } 2239 #endif 2240 2241 /* set up device context async ev fd as NON_BLOCKING */ 2242 flag = fcntl(device->context->async_fd, F_GETFL); 2243 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2244 if (rc < 0) { 2245 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2246 free(device); 2247 break; 2248 } 2249 2250 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2251 i++; 2252 2253 pd = NULL; 2254 if (g_nvmf_hooks.get_ibv_pd) { 2255 pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2256 } 2257 2258 if (!g_nvmf_hooks.get_ibv_pd) { 2259 device->pd = ibv_alloc_pd(device->context); 2260 if (!device->pd) { 2261 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2262 spdk_nvmf_rdma_destroy(&rtransport->transport); 2263 return NULL; 2264 } 2265 } else { 2266 device->pd = pd; 2267 } 2268 2269 assert(device->map == NULL); 2270 2271 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2272 if (!device->map) { 2273 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2274 spdk_nvmf_rdma_destroy(&rtransport->transport); 2275 return NULL; 2276 } 2277 2278 assert(device->map != NULL); 2279 assert(device->pd != NULL); 2280 } 2281 rdma_free_devices(contexts); 2282 2283 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2284 /* divide and round up. */ 2285 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2286 2287 /* round up to the nearest 4k. */ 2288 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2289 2290 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2291 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2292 opts->io_unit_size); 2293 } 2294 2295 if (rc < 0) { 2296 spdk_nvmf_rdma_destroy(&rtransport->transport); 2297 return NULL; 2298 } 2299 2300 /* Set up poll descriptor array to monitor events from RDMA and IB 2301 * in a single poll syscall 2302 */ 2303 rtransport->npoll_fds = i + 1; 2304 i = 0; 2305 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2306 if (rtransport->poll_fds == NULL) { 2307 SPDK_ERRLOG("poll_fds allocation failed\n"); 2308 spdk_nvmf_rdma_destroy(&rtransport->transport); 2309 return NULL; 2310 } 2311 2312 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2313 rtransport->poll_fds[i++].events = POLLIN; 2314 2315 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2316 rtransport->poll_fds[i].fd = device->context->async_fd; 2317 rtransport->poll_fds[i++].events = POLLIN; 2318 } 2319 2320 return &rtransport->transport; 2321 } 2322 2323 static int 2324 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2325 { 2326 struct spdk_nvmf_rdma_transport *rtransport; 2327 struct spdk_nvmf_rdma_port *port, *port_tmp; 2328 struct spdk_nvmf_rdma_device *device, *device_tmp; 2329 2330 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2331 2332 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2333 TAILQ_REMOVE(&rtransport->ports, port, link); 2334 rdma_destroy_id(port->id); 2335 free(port); 2336 } 2337 2338 if (rtransport->poll_fds != NULL) { 2339 free(rtransport->poll_fds); 2340 } 2341 2342 if (rtransport->event_channel != NULL) { 2343 rdma_destroy_event_channel(rtransport->event_channel); 2344 } 2345 2346 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2347 TAILQ_REMOVE(&rtransport->devices, device, link); 2348 if (device->map) { 2349 spdk_mem_map_free(&device->map); 2350 } 2351 if (device->pd) { 2352 if (!g_nvmf_hooks.get_ibv_pd) { 2353 ibv_dealloc_pd(device->pd); 2354 } 2355 } 2356 free(device); 2357 } 2358 2359 if (rtransport->data_wr_pool != NULL) { 2360 if (spdk_mempool_count(rtransport->data_wr_pool) != 2361 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2362 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2363 spdk_mempool_count(rtransport->data_wr_pool), 2364 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2365 } 2366 } 2367 2368 spdk_mempool_free(rtransport->data_wr_pool); 2369 pthread_mutex_destroy(&rtransport->lock); 2370 free(rtransport); 2371 2372 return 0; 2373 } 2374 2375 static int 2376 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2377 struct spdk_nvme_transport_id *trid, 2378 bool peer); 2379 2380 static int 2381 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2382 const struct spdk_nvme_transport_id *trid) 2383 { 2384 struct spdk_nvmf_rdma_transport *rtransport; 2385 struct spdk_nvmf_rdma_device *device; 2386 struct spdk_nvmf_rdma_port *port_tmp, *port; 2387 struct addrinfo *res; 2388 struct addrinfo hints; 2389 int family; 2390 int rc; 2391 2392 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2393 2394 port = calloc(1, sizeof(*port)); 2395 if (!port) { 2396 return -ENOMEM; 2397 } 2398 2399 /* Selectively copy the trid. Things like NQN don't matter here - that 2400 * mapping is enforced elsewhere. 2401 */ 2402 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2403 port->trid.adrfam = trid->adrfam; 2404 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 2405 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 2406 2407 pthread_mutex_lock(&rtransport->lock); 2408 assert(rtransport->event_channel != NULL); 2409 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 2410 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 2411 port_tmp->ref++; 2412 free(port); 2413 /* Already listening at this address */ 2414 pthread_mutex_unlock(&rtransport->lock); 2415 return 0; 2416 } 2417 } 2418 2419 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2420 if (rc < 0) { 2421 SPDK_ERRLOG("rdma_create_id() failed\n"); 2422 free(port); 2423 pthread_mutex_unlock(&rtransport->lock); 2424 return rc; 2425 } 2426 2427 switch (port->trid.adrfam) { 2428 case SPDK_NVMF_ADRFAM_IPV4: 2429 family = AF_INET; 2430 break; 2431 case SPDK_NVMF_ADRFAM_IPV6: 2432 family = AF_INET6; 2433 break; 2434 default: 2435 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 2436 free(port); 2437 pthread_mutex_unlock(&rtransport->lock); 2438 return -EINVAL; 2439 } 2440 2441 memset(&hints, 0, sizeof(hints)); 2442 hints.ai_family = family; 2443 hints.ai_flags = AI_NUMERICSERV; 2444 hints.ai_socktype = SOCK_STREAM; 2445 hints.ai_protocol = 0; 2446 2447 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 2448 if (rc) { 2449 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2450 free(port); 2451 pthread_mutex_unlock(&rtransport->lock); 2452 return -EINVAL; 2453 } 2454 2455 rc = rdma_bind_addr(port->id, res->ai_addr); 2456 freeaddrinfo(res); 2457 2458 if (rc < 0) { 2459 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2460 rdma_destroy_id(port->id); 2461 free(port); 2462 pthread_mutex_unlock(&rtransport->lock); 2463 return rc; 2464 } 2465 2466 if (!port->id->verbs) { 2467 SPDK_ERRLOG("ibv_context is null\n"); 2468 rdma_destroy_id(port->id); 2469 free(port); 2470 pthread_mutex_unlock(&rtransport->lock); 2471 return -1; 2472 } 2473 2474 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 2475 if (rc < 0) { 2476 SPDK_ERRLOG("rdma_listen() failed\n"); 2477 rdma_destroy_id(port->id); 2478 free(port); 2479 pthread_mutex_unlock(&rtransport->lock); 2480 return rc; 2481 } 2482 2483 TAILQ_FOREACH(device, &rtransport->devices, link) { 2484 if (device->context == port->id->verbs) { 2485 port->device = device; 2486 break; 2487 } 2488 } 2489 if (!port->device) { 2490 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2491 port->id->verbs); 2492 rdma_destroy_id(port->id); 2493 free(port); 2494 pthread_mutex_unlock(&rtransport->lock); 2495 return -EINVAL; 2496 } 2497 2498 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 2499 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 2500 2501 port->ref = 1; 2502 2503 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2504 pthread_mutex_unlock(&rtransport->lock); 2505 2506 return 0; 2507 } 2508 2509 static int 2510 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2511 const struct spdk_nvme_transport_id *_trid) 2512 { 2513 struct spdk_nvmf_rdma_transport *rtransport; 2514 struct spdk_nvmf_rdma_port *port, *tmp; 2515 struct spdk_nvme_transport_id trid = {}; 2516 2517 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2518 2519 /* Selectively copy the trid. Things like NQN don't matter here - that 2520 * mapping is enforced elsewhere. 2521 */ 2522 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2523 trid.adrfam = _trid->adrfam; 2524 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2525 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2526 2527 pthread_mutex_lock(&rtransport->lock); 2528 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2529 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2530 assert(port->ref > 0); 2531 port->ref--; 2532 if (port->ref == 0) { 2533 TAILQ_REMOVE(&rtransport->ports, port, link); 2534 rdma_destroy_id(port->id); 2535 free(port); 2536 } 2537 break; 2538 } 2539 } 2540 2541 pthread_mutex_unlock(&rtransport->lock); 2542 return 0; 2543 } 2544 2545 static void 2546 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2547 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2548 { 2549 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2550 struct spdk_nvmf_rdma_resources *resources; 2551 2552 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2553 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2554 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2555 break; 2556 } 2557 } 2558 2559 /* Then RDMA writes since reads have stronger restrictions than writes */ 2560 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2561 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2562 break; 2563 } 2564 } 2565 2566 /* The second highest priority is I/O waiting on memory buffers. */ 2567 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->poller->group->pending_data_buf_queue, state_link, 2568 req_tmp) { 2569 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2570 break; 2571 } 2572 } 2573 2574 resources = rqpair->resources; 2575 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2576 rdma_req = STAILQ_FIRST(&resources->free_queue); 2577 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2578 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2579 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2580 2581 if (rqpair->srq != NULL) { 2582 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2583 rdma_req->recv->qpair->qd++; 2584 } else { 2585 rqpair->qd++; 2586 } 2587 2588 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2589 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2590 break; 2591 } 2592 } 2593 } 2594 2595 static void 2596 _nvmf_rdma_qpair_disconnect(void *ctx) 2597 { 2598 struct spdk_nvmf_qpair *qpair = ctx; 2599 2600 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2601 } 2602 2603 static void 2604 _nvmf_rdma_try_disconnect(void *ctx) 2605 { 2606 struct spdk_nvmf_qpair *qpair = ctx; 2607 struct spdk_nvmf_poll_group *group; 2608 2609 /* Read the group out of the qpair. This is normally set and accessed only from 2610 * the thread that created the group. Here, we're not on that thread necessarily. 2611 * The data member qpair->group begins it's life as NULL and then is assigned to 2612 * a pointer and never changes. So fortunately reading this and checking for 2613 * non-NULL is thread safe in the x86_64 memory model. */ 2614 group = qpair->group; 2615 2616 if (group == NULL) { 2617 /* The qpair hasn't been assigned to a group yet, so we can't 2618 * process a disconnect. Send a message to ourself and try again. */ 2619 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2620 return; 2621 } 2622 2623 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2624 } 2625 2626 static inline void 2627 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2628 { 2629 if (__sync_bool_compare_and_swap(&rqpair->disconnect_started, false, true)) { 2630 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2631 } 2632 } 2633 2634 static void nvmf_rdma_destroy_drained_qpair(void *ctx) 2635 { 2636 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2637 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2638 struct spdk_nvmf_rdma_transport, transport); 2639 2640 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2641 if (rqpair->current_send_depth != 0) { 2642 return; 2643 } 2644 2645 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2646 return; 2647 } 2648 2649 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) { 2650 return; 2651 } 2652 2653 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2654 spdk_nvmf_rdma_qpair_destroy(rqpair); 2655 } 2656 2657 2658 static int 2659 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2660 { 2661 struct spdk_nvmf_qpair *qpair; 2662 struct spdk_nvmf_rdma_qpair *rqpair; 2663 2664 if (evt->id == NULL) { 2665 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2666 return -1; 2667 } 2668 2669 qpair = evt->id->context; 2670 if (qpair == NULL) { 2671 SPDK_ERRLOG("disconnect request: no active connection\n"); 2672 return -1; 2673 } 2674 2675 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2676 2677 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2678 2679 spdk_nvmf_rdma_update_ibv_state(rqpair); 2680 2681 spdk_nvmf_rdma_start_disconnect(rqpair); 2682 2683 return 0; 2684 } 2685 2686 #ifdef DEBUG 2687 static const char *CM_EVENT_STR[] = { 2688 "RDMA_CM_EVENT_ADDR_RESOLVED", 2689 "RDMA_CM_EVENT_ADDR_ERROR", 2690 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2691 "RDMA_CM_EVENT_ROUTE_ERROR", 2692 "RDMA_CM_EVENT_CONNECT_REQUEST", 2693 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2694 "RDMA_CM_EVENT_CONNECT_ERROR", 2695 "RDMA_CM_EVENT_UNREACHABLE", 2696 "RDMA_CM_EVENT_REJECTED", 2697 "RDMA_CM_EVENT_ESTABLISHED", 2698 "RDMA_CM_EVENT_DISCONNECTED", 2699 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2700 "RDMA_CM_EVENT_MULTICAST_JOIN", 2701 "RDMA_CM_EVENT_MULTICAST_ERROR", 2702 "RDMA_CM_EVENT_ADDR_CHANGE", 2703 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2704 }; 2705 #endif /* DEBUG */ 2706 2707 static void 2708 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2709 { 2710 struct spdk_nvmf_rdma_transport *rtransport; 2711 struct rdma_cm_event *event; 2712 int rc; 2713 2714 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2715 2716 if (rtransport->event_channel == NULL) { 2717 return; 2718 } 2719 2720 while (1) { 2721 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2722 if (rc == 0) { 2723 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2724 2725 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2726 2727 switch (event->event) { 2728 case RDMA_CM_EVENT_ADDR_RESOLVED: 2729 case RDMA_CM_EVENT_ADDR_ERROR: 2730 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2731 case RDMA_CM_EVENT_ROUTE_ERROR: 2732 /* No action required. The target never attempts to resolve routes. */ 2733 break; 2734 case RDMA_CM_EVENT_CONNECT_REQUEST: 2735 rc = nvmf_rdma_connect(transport, event, cb_fn); 2736 if (rc < 0) { 2737 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2738 break; 2739 } 2740 break; 2741 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2742 /* The target never initiates a new connection. So this will not occur. */ 2743 break; 2744 case RDMA_CM_EVENT_CONNECT_ERROR: 2745 /* Can this happen? The docs say it can, but not sure what causes it. */ 2746 break; 2747 case RDMA_CM_EVENT_UNREACHABLE: 2748 case RDMA_CM_EVENT_REJECTED: 2749 /* These only occur on the client side. */ 2750 break; 2751 case RDMA_CM_EVENT_ESTABLISHED: 2752 /* TODO: Should we be waiting for this event anywhere? */ 2753 break; 2754 case RDMA_CM_EVENT_DISCONNECTED: 2755 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2756 rc = nvmf_rdma_disconnect(event); 2757 if (rc < 0) { 2758 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2759 break; 2760 } 2761 break; 2762 case RDMA_CM_EVENT_MULTICAST_JOIN: 2763 case RDMA_CM_EVENT_MULTICAST_ERROR: 2764 /* Multicast is not used */ 2765 break; 2766 case RDMA_CM_EVENT_ADDR_CHANGE: 2767 /* Not utilizing this event */ 2768 break; 2769 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2770 /* For now, do nothing. The target never re-uses queue pairs. */ 2771 break; 2772 default: 2773 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2774 break; 2775 } 2776 2777 rdma_ack_cm_event(event); 2778 } else { 2779 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2780 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2781 } 2782 break; 2783 } 2784 } 2785 } 2786 2787 static void 2788 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 2789 { 2790 int rc; 2791 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 2792 struct ibv_async_event event; 2793 enum ibv_qp_state state; 2794 2795 rc = ibv_get_async_event(device->context, &event); 2796 2797 if (rc) { 2798 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 2799 errno, spdk_strerror(errno)); 2800 return; 2801 } 2802 2803 switch (event.event_type) { 2804 case IBV_EVENT_QP_FATAL: 2805 rqpair = event.element.qp->qp_context; 2806 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 2807 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2808 (uintptr_t)rqpair->cm_id, event.event_type); 2809 spdk_nvmf_rdma_update_ibv_state(rqpair); 2810 spdk_nvmf_rdma_start_disconnect(rqpair); 2811 break; 2812 case IBV_EVENT_QP_LAST_WQE_REACHED: 2813 /* This event only occurs for shared receive queues. */ 2814 rqpair = event.element.qp->qp_context; 2815 rqpair->last_wqe_reached = true; 2816 2817 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last WQE reached event received for rqpair %p\n", rqpair); 2818 /* This must be handled on the polling thread if it exists. Otherwise the timeout will catch it. */ 2819 if (rqpair->qpair.group) { 2820 spdk_thread_send_msg(rqpair->qpair.group->thread, nvmf_rdma_destroy_drained_qpair, rqpair); 2821 } else { 2822 SPDK_ERRLOG("Unable to destroy the qpair %p since it does not have a poll group.\n", rqpair); 2823 } 2824 2825 break; 2826 case IBV_EVENT_SQ_DRAINED: 2827 /* This event occurs frequently in both error and non-error states. 2828 * Check if the qpair is in an error state before sending a message. 2829 * Note that we're not on the correct thread to access the qpair, but 2830 * the operations that the below calls make all happen to be thread 2831 * safe. */ 2832 rqpair = event.element.qp->qp_context; 2833 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last sq drained event received for rqpair %p\n", rqpair); 2834 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2835 (uintptr_t)rqpair->cm_id, event.event_type); 2836 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 2837 if (state == IBV_QPS_ERR) { 2838 spdk_nvmf_rdma_start_disconnect(rqpair); 2839 } 2840 break; 2841 case IBV_EVENT_QP_REQ_ERR: 2842 case IBV_EVENT_QP_ACCESS_ERR: 2843 case IBV_EVENT_COMM_EST: 2844 case IBV_EVENT_PATH_MIG: 2845 case IBV_EVENT_PATH_MIG_ERR: 2846 SPDK_NOTICELOG("Async event: %s\n", 2847 ibv_event_type_str(event.event_type)); 2848 rqpair = event.element.qp->qp_context; 2849 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 2850 (uintptr_t)rqpair->cm_id, event.event_type); 2851 spdk_nvmf_rdma_update_ibv_state(rqpair); 2852 break; 2853 case IBV_EVENT_CQ_ERR: 2854 case IBV_EVENT_DEVICE_FATAL: 2855 case IBV_EVENT_PORT_ACTIVE: 2856 case IBV_EVENT_PORT_ERR: 2857 case IBV_EVENT_LID_CHANGE: 2858 case IBV_EVENT_PKEY_CHANGE: 2859 case IBV_EVENT_SM_CHANGE: 2860 case IBV_EVENT_SRQ_ERR: 2861 case IBV_EVENT_SRQ_LIMIT_REACHED: 2862 case IBV_EVENT_CLIENT_REREGISTER: 2863 case IBV_EVENT_GID_CHANGE: 2864 default: 2865 SPDK_NOTICELOG("Async event: %s\n", 2866 ibv_event_type_str(event.event_type)); 2867 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 2868 break; 2869 } 2870 ibv_ack_async_event(&event); 2871 } 2872 2873 static void 2874 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn) 2875 { 2876 int nfds, i = 0; 2877 struct spdk_nvmf_rdma_transport *rtransport; 2878 struct spdk_nvmf_rdma_device *device, *tmp; 2879 2880 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2881 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 2882 2883 if (nfds <= 0) { 2884 return; 2885 } 2886 2887 /* The first poll descriptor is RDMA CM event */ 2888 if (rtransport->poll_fds[i++].revents & POLLIN) { 2889 spdk_nvmf_process_cm_event(transport, cb_fn); 2890 nfds--; 2891 } 2892 2893 if (nfds == 0) { 2894 return; 2895 } 2896 2897 /* Second and subsequent poll descriptors are IB async events */ 2898 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2899 if (rtransport->poll_fds[i++].revents & POLLIN) { 2900 spdk_nvmf_process_ib_event(device); 2901 nfds--; 2902 } 2903 } 2904 /* check all flagged fd's have been served */ 2905 assert(nfds == 0); 2906 } 2907 2908 static void 2909 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 2910 struct spdk_nvme_transport_id *trid, 2911 struct spdk_nvmf_discovery_log_page_entry *entry) 2912 { 2913 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 2914 entry->adrfam = trid->adrfam; 2915 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED; 2916 2917 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 2918 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 2919 2920 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 2921 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 2922 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 2923 } 2924 2925 static void 2926 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 2927 2928 static struct spdk_nvmf_transport_poll_group * 2929 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 2930 { 2931 struct spdk_nvmf_rdma_transport *rtransport; 2932 struct spdk_nvmf_rdma_poll_group *rgroup; 2933 struct spdk_nvmf_rdma_poller *poller; 2934 struct spdk_nvmf_rdma_device *device; 2935 struct ibv_srq_init_attr srq_init_attr; 2936 struct spdk_nvmf_rdma_resource_opts opts; 2937 int num_cqe; 2938 2939 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2940 2941 rgroup = calloc(1, sizeof(*rgroup)); 2942 if (!rgroup) { 2943 return NULL; 2944 } 2945 2946 TAILQ_INIT(&rgroup->pollers); 2947 STAILQ_INIT(&rgroup->pending_data_buf_queue); 2948 2949 pthread_mutex_lock(&rtransport->lock); 2950 TAILQ_FOREACH(device, &rtransport->devices, link) { 2951 poller = calloc(1, sizeof(*poller)); 2952 if (!poller) { 2953 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 2954 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2955 pthread_mutex_unlock(&rtransport->lock); 2956 return NULL; 2957 } 2958 2959 poller->device = device; 2960 poller->group = rgroup; 2961 2962 TAILQ_INIT(&poller->qpairs); 2963 STAILQ_INIT(&poller->qpairs_pending_send); 2964 2965 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 2966 if (transport->opts.no_srq == false && device->attr.max_srq != 0) { 2967 poller->max_srq_depth = transport->opts.max_srq_depth; 2968 2969 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 2970 srq_init_attr.attr.max_wr = poller->max_srq_depth; 2971 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 2972 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 2973 if (!poller->srq) { 2974 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 2975 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2976 pthread_mutex_unlock(&rtransport->lock); 2977 return NULL; 2978 } 2979 2980 opts.qp = poller->srq; 2981 opts.pd = device->pd; 2982 opts.qpair = NULL; 2983 opts.shared = true; 2984 opts.max_queue_depth = poller->max_srq_depth; 2985 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 2986 2987 poller->resources = nvmf_rdma_resources_create(&opts); 2988 if (!poller->resources) { 2989 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 2990 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 2991 pthread_mutex_unlock(&rtransport->lock); 2992 } 2993 } 2994 2995 /* 2996 * When using an srq, we can limit the completion queue at startup. 2997 * The following formula represents the calculation: 2998 * num_cqe = num_recv + num_data_wr + num_send_wr. 2999 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3000 */ 3001 if (poller->srq) { 3002 num_cqe = poller->max_srq_depth * 3; 3003 } else { 3004 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3005 } 3006 3007 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3008 if (!poller->cq) { 3009 SPDK_ERRLOG("Unable to create completion queue\n"); 3010 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3011 pthread_mutex_unlock(&rtransport->lock); 3012 return NULL; 3013 } 3014 poller->num_cqe = num_cqe; 3015 } 3016 3017 pthread_mutex_unlock(&rtransport->lock); 3018 return &rgroup->group; 3019 } 3020 3021 static void 3022 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3023 { 3024 struct spdk_nvmf_rdma_poll_group *rgroup; 3025 struct spdk_nvmf_rdma_poller *poller, *tmp; 3026 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3027 3028 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3029 3030 if (!rgroup) { 3031 return; 3032 } 3033 3034 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3035 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3036 3037 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3038 spdk_nvmf_rdma_qpair_destroy(qpair); 3039 } 3040 3041 if (poller->srq) { 3042 nvmf_rdma_resources_destroy(poller->resources); 3043 ibv_destroy_srq(poller->srq); 3044 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3045 } 3046 3047 if (poller->cq) { 3048 ibv_destroy_cq(poller->cq); 3049 } 3050 3051 free(poller); 3052 } 3053 3054 if (!STAILQ_EMPTY(&rgroup->pending_data_buf_queue)) { 3055 SPDK_ERRLOG("Pending I/O list wasn't empty on poll group destruction\n"); 3056 } 3057 3058 free(rgroup); 3059 } 3060 3061 static void 3062 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3063 { 3064 if (rqpair->cm_id != NULL) { 3065 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3066 } 3067 spdk_nvmf_rdma_qpair_destroy(rqpair); 3068 } 3069 3070 static int 3071 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3072 struct spdk_nvmf_qpair *qpair) 3073 { 3074 struct spdk_nvmf_rdma_poll_group *rgroup; 3075 struct spdk_nvmf_rdma_qpair *rqpair; 3076 struct spdk_nvmf_rdma_device *device; 3077 struct spdk_nvmf_rdma_poller *poller; 3078 int rc; 3079 3080 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3081 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3082 3083 device = rqpair->port->device; 3084 3085 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3086 if (poller->device == device) { 3087 break; 3088 } 3089 } 3090 3091 if (!poller) { 3092 SPDK_ERRLOG("No poller found for device.\n"); 3093 return -1; 3094 } 3095 3096 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3097 rqpair->poller = poller; 3098 rqpair->srq = rqpair->poller->srq; 3099 3100 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 3101 if (rc < 0) { 3102 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3103 return -1; 3104 } 3105 3106 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3107 if (rc) { 3108 /* Try to reject, but we probably can't */ 3109 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3110 return -1; 3111 } 3112 3113 spdk_nvmf_rdma_update_ibv_state(rqpair); 3114 3115 return 0; 3116 } 3117 3118 static int 3119 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3120 { 3121 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3122 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3123 struct spdk_nvmf_rdma_transport, transport); 3124 3125 nvmf_rdma_request_free(rdma_req, rtransport); 3126 return 0; 3127 } 3128 3129 static int 3130 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3131 { 3132 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3133 struct spdk_nvmf_rdma_transport, transport); 3134 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3135 struct spdk_nvmf_rdma_request, req); 3136 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3137 struct spdk_nvmf_rdma_qpair, qpair); 3138 3139 if (rqpair->ibv_state != IBV_QPS_ERR) { 3140 /* The connection is alive, so process the request as normal */ 3141 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3142 } else { 3143 /* The connection is dead. Move the request directly to the completed state. */ 3144 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3145 } 3146 3147 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3148 3149 return 0; 3150 } 3151 3152 static int 3153 spdk_nvmf_rdma_destroy_defunct_qpair(void *ctx) 3154 { 3155 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 3156 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3157 struct spdk_nvmf_rdma_transport, transport); 3158 3159 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3160 spdk_nvmf_rdma_qpair_destroy(rqpair); 3161 3162 return 0; 3163 } 3164 3165 static void 3166 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3167 { 3168 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3169 3170 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 3171 return; 3172 } 3173 3174 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 3175 3176 /* This happens only when the qpair is disconnected before 3177 * it is added to the poll group. Since there is no poll group, 3178 * the RDMA qp has not been initialized yet and the RDMA CM 3179 * event has not yet been acknowledged, so we need to reject it. 3180 */ 3181 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3182 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3183 return; 3184 } 3185 3186 if (rqpair->ibv_state != IBV_QPS_ERR) { 3187 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 3188 } 3189 3190 rqpair->destruct_poller = spdk_poller_register(spdk_nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 3191 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 3192 } 3193 3194 static struct spdk_nvmf_rdma_qpair * 3195 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3196 { 3197 struct spdk_nvmf_rdma_qpair *rqpair; 3198 /* @todo: improve QP search */ 3199 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3200 if (wc->qp_num == rqpair->cm_id->qp->qp_num) { 3201 return rqpair; 3202 } 3203 } 3204 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3205 return NULL; 3206 } 3207 3208 #ifdef DEBUG 3209 static int 3210 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3211 { 3212 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3213 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3214 } 3215 #endif 3216 3217 static void 3218 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3219 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3220 { 3221 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3222 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3223 3224 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3225 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3226 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3227 assert(rqpair->current_send_depth > 0); 3228 rqpair->current_send_depth--; 3229 switch (bad_rdma_wr->type) { 3230 case RDMA_WR_TYPE_DATA: 3231 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3232 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3233 assert(rqpair->current_read_depth > 0); 3234 rqpair->current_read_depth--; 3235 } 3236 break; 3237 case RDMA_WR_TYPE_SEND: 3238 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3239 break; 3240 default: 3241 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3242 prev_rdma_req = cur_rdma_req; 3243 continue; 3244 } 3245 3246 if (prev_rdma_req == cur_rdma_req) { 3247 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3248 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3249 continue; 3250 } 3251 3252 switch (cur_rdma_req->state) { 3253 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3254 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3255 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3256 break; 3257 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3258 case RDMA_REQUEST_STATE_COMPLETING: 3259 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3260 break; 3261 default: 3262 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3263 cur_rdma_req->state, rqpair); 3264 continue; 3265 } 3266 3267 spdk_nvmf_rdma_request_process(rtransport, cur_rdma_req); 3268 prev_rdma_req = cur_rdma_req; 3269 } 3270 3271 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3272 /* Disconnect the connection. */ 3273 spdk_nvmf_rdma_start_disconnect(rqpair); 3274 } 3275 3276 } 3277 3278 static void 3279 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3280 struct spdk_nvmf_rdma_poller *rpoller) 3281 { 3282 struct spdk_nvmf_rdma_qpair *rqpair; 3283 struct ibv_send_wr *bad_wr = NULL; 3284 int rc; 3285 3286 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3287 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3288 assert(rqpair->sends_to_post.first != NULL); 3289 rc = ibv_post_send(rqpair->cm_id->qp, rqpair->sends_to_post.first, &bad_wr); 3290 3291 /* bad wr always points to the first wr that failed. */ 3292 if (rc) { 3293 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3294 } 3295 rqpair->sends_to_post.first = NULL; 3296 rqpair->sends_to_post.last = NULL; 3297 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3298 } 3299 } 3300 3301 static int 3302 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3303 struct spdk_nvmf_rdma_poller *rpoller) 3304 { 3305 struct ibv_wc wc[32]; 3306 struct spdk_nvmf_rdma_wr *rdma_wr; 3307 struct spdk_nvmf_rdma_request *rdma_req; 3308 struct spdk_nvmf_rdma_recv *rdma_recv; 3309 struct spdk_nvmf_rdma_qpair *rqpair; 3310 int reaped, i; 3311 int count = 0; 3312 bool error = false; 3313 3314 /* Poll for completing operations. */ 3315 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3316 if (reaped < 0) { 3317 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3318 errno, spdk_strerror(errno)); 3319 return -1; 3320 } 3321 3322 for (i = 0; i < reaped; i++) { 3323 3324 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3325 3326 switch (rdma_wr->type) { 3327 case RDMA_WR_TYPE_SEND: 3328 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3329 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3330 3331 if (!wc[i].status) { 3332 count++; 3333 assert(wc[i].opcode == IBV_WC_SEND); 3334 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 3335 } else { 3336 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3337 } 3338 3339 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3340 rqpair->current_send_depth--; 3341 3342 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3343 assert(rdma_req->num_outstanding_data_wr == 0); 3344 break; 3345 case RDMA_WR_TYPE_RECV: 3346 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3347 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3348 if (rpoller->srq != NULL) { 3349 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3350 } 3351 rqpair = rdma_recv->qpair; 3352 3353 assert(rqpair != NULL); 3354 if (!wc[i].status) { 3355 assert(wc[i].opcode == IBV_WC_RECV); 3356 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3357 spdk_nvmf_rdma_start_disconnect(rqpair); 3358 break; 3359 } 3360 } 3361 3362 rqpair->current_recv_depth++; 3363 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3364 break; 3365 case RDMA_WR_TYPE_DATA: 3366 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3367 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3368 3369 assert(rdma_req->num_outstanding_data_wr > 0); 3370 3371 rqpair->current_send_depth--; 3372 rdma_req->num_outstanding_data_wr--; 3373 if (!wc[i].status) { 3374 if (wc[i].opcode == IBV_WC_RDMA_READ) { 3375 rqpair->current_read_depth--; 3376 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3377 if (rdma_req->num_outstanding_data_wr == 0) { 3378 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3379 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3380 } 3381 } else { 3382 assert(wc[i].opcode == IBV_WC_RDMA_WRITE); 3383 } 3384 } else { 3385 /* If the data transfer fails still force the queue into the error state, 3386 * if we were performing an RDMA_READ, we need to force the request into a 3387 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3388 * case, we should wait for the SEND to complete. */ 3389 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3390 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3391 rqpair->current_read_depth--; 3392 if (rdma_req->num_outstanding_data_wr == 0) { 3393 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3394 } 3395 } 3396 } 3397 break; 3398 default: 3399 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3400 continue; 3401 } 3402 3403 /* Handle error conditions */ 3404 if (wc[i].status) { 3405 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 3406 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 3407 3408 error = true; 3409 3410 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3411 /* Disconnect the connection. */ 3412 spdk_nvmf_rdma_start_disconnect(rqpair); 3413 } else { 3414 nvmf_rdma_destroy_drained_qpair(rqpair); 3415 } 3416 continue; 3417 } 3418 3419 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3420 3421 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3422 nvmf_rdma_destroy_drained_qpair(rqpair); 3423 } 3424 } 3425 3426 if (error == true) { 3427 return -1; 3428 } 3429 3430 /* submit outstanding work requests. */ 3431 _poller_submit_sends(rtransport, rpoller); 3432 3433 return count; 3434 } 3435 3436 static int 3437 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3438 { 3439 struct spdk_nvmf_rdma_transport *rtransport; 3440 struct spdk_nvmf_rdma_poll_group *rgroup; 3441 struct spdk_nvmf_rdma_poller *rpoller; 3442 int count, rc; 3443 3444 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3445 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3446 3447 count = 0; 3448 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3449 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 3450 if (rc < 0) { 3451 return rc; 3452 } 3453 count += rc; 3454 } 3455 3456 return count; 3457 } 3458 3459 static int 3460 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3461 struct spdk_nvme_transport_id *trid, 3462 bool peer) 3463 { 3464 struct sockaddr *saddr; 3465 uint16_t port; 3466 3467 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 3468 3469 if (peer) { 3470 saddr = rdma_get_peer_addr(id); 3471 } else { 3472 saddr = rdma_get_local_addr(id); 3473 } 3474 switch (saddr->sa_family) { 3475 case AF_INET: { 3476 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3477 3478 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3479 inet_ntop(AF_INET, &saddr_in->sin_addr, 3480 trid->traddr, sizeof(trid->traddr)); 3481 if (peer) { 3482 port = ntohs(rdma_get_dst_port(id)); 3483 } else { 3484 port = ntohs(rdma_get_src_port(id)); 3485 } 3486 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3487 break; 3488 } 3489 case AF_INET6: { 3490 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3491 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3492 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3493 trid->traddr, sizeof(trid->traddr)); 3494 if (peer) { 3495 port = ntohs(rdma_get_dst_port(id)); 3496 } else { 3497 port = ntohs(rdma_get_src_port(id)); 3498 } 3499 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3500 break; 3501 } 3502 default: 3503 return -1; 3504 3505 } 3506 3507 return 0; 3508 } 3509 3510 static int 3511 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3512 struct spdk_nvme_transport_id *trid) 3513 { 3514 struct spdk_nvmf_rdma_qpair *rqpair; 3515 3516 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3517 3518 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 3519 } 3520 3521 static int 3522 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3523 struct spdk_nvme_transport_id *trid) 3524 { 3525 struct spdk_nvmf_rdma_qpair *rqpair; 3526 3527 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3528 3529 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 3530 } 3531 3532 static int 3533 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3534 struct spdk_nvme_transport_id *trid) 3535 { 3536 struct spdk_nvmf_rdma_qpair *rqpair; 3537 3538 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3539 3540 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 3541 } 3542 3543 void 3544 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 3545 { 3546 g_nvmf_hooks = *hooks; 3547 } 3548 3549 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 3550 .type = SPDK_NVME_TRANSPORT_RDMA, 3551 .opts_init = spdk_nvmf_rdma_opts_init, 3552 .create = spdk_nvmf_rdma_create, 3553 .destroy = spdk_nvmf_rdma_destroy, 3554 3555 .listen = spdk_nvmf_rdma_listen, 3556 .stop_listen = spdk_nvmf_rdma_stop_listen, 3557 .accept = spdk_nvmf_rdma_accept, 3558 3559 .listener_discover = spdk_nvmf_rdma_discover, 3560 3561 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 3562 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 3563 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 3564 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 3565 3566 .req_free = spdk_nvmf_rdma_request_free, 3567 .req_complete = spdk_nvmf_rdma_request_complete, 3568 3569 .qpair_fini = spdk_nvmf_rdma_close_qpair, 3570 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 3571 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 3572 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 3573 3574 }; 3575 3576 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 3577