1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2019, 2020 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 "spdk/config.h" 37 #include "spdk/thread.h" 38 #include "spdk/likely.h" 39 #include "spdk/nvmf_transport.h" 40 #include "spdk/string.h" 41 #include "spdk/trace.h" 42 #include "spdk/util.h" 43 44 #include "spdk_internal/assert.h" 45 #include "spdk_internal/log.h" 46 #include "spdk_internal/rdma.h" 47 48 #include "nvmf_internal.h" 49 50 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 51 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 52 53 /* 54 RDMA Connection Resource Defaults 55 */ 56 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 57 #define NVMF_DEFAULT_RSP_SGE 1 58 #define NVMF_DEFAULT_RX_SGE 2 59 60 /* The RDMA completion queue size */ 61 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 62 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 63 64 static int g_spdk_nvmf_ibv_query_mask = 65 IBV_QP_STATE | 66 IBV_QP_PKEY_INDEX | 67 IBV_QP_PORT | 68 IBV_QP_ACCESS_FLAGS | 69 IBV_QP_AV | 70 IBV_QP_PATH_MTU | 71 IBV_QP_DEST_QPN | 72 IBV_QP_RQ_PSN | 73 IBV_QP_MAX_DEST_RD_ATOMIC | 74 IBV_QP_MIN_RNR_TIMER | 75 IBV_QP_SQ_PSN | 76 IBV_QP_TIMEOUT | 77 IBV_QP_RETRY_CNT | 78 IBV_QP_RNR_RETRY | 79 IBV_QP_MAX_QP_RD_ATOMIC; 80 81 enum spdk_nvmf_rdma_request_state { 82 /* The request is not currently in use */ 83 RDMA_REQUEST_STATE_FREE = 0, 84 85 /* Initial state when request first received */ 86 RDMA_REQUEST_STATE_NEW, 87 88 /* The request is queued until a data buffer is available. */ 89 RDMA_REQUEST_STATE_NEED_BUFFER, 90 91 /* The request is waiting on RDMA queue depth availability 92 * to transfer data from the host to the controller. 93 */ 94 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 95 96 /* The request is currently transferring data from the host to the controller. */ 97 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 98 99 /* The request is ready to execute at the block device */ 100 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 101 102 /* The request is currently executing at the block device */ 103 RDMA_REQUEST_STATE_EXECUTING, 104 105 /* The request finished executing at the block device */ 106 RDMA_REQUEST_STATE_EXECUTED, 107 108 /* The request is waiting on RDMA queue depth availability 109 * to transfer data from the controller to the host. 110 */ 111 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 112 113 /* The request is ready to send a completion */ 114 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 115 116 /* The request is currently transferring data from the controller to the host. */ 117 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 118 119 /* The request currently has an outstanding completion without an 120 * associated data transfer. 121 */ 122 RDMA_REQUEST_STATE_COMPLETING, 123 124 /* The request completed and can be marked free. */ 125 RDMA_REQUEST_STATE_COMPLETED, 126 127 /* Terminator */ 128 RDMA_REQUEST_NUM_STATES, 129 }; 130 131 #define OBJECT_NVMF_RDMA_IO 0x40 132 133 #define TRACE_GROUP_NVMF_RDMA 0x4 134 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 135 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 136 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 137 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 138 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 139 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 140 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 141 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 142 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 143 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 144 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 145 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 146 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 147 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 148 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 149 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 150 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 151 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 152 153 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 154 { 155 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 156 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 157 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 158 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 159 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 160 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 161 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 162 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 163 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 164 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 165 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 166 spdk_trace_register_description("RDMA_REQ_TX_H2C", 167 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 168 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 169 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 170 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 171 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 172 spdk_trace_register_description("RDMA_REQ_EXECUTING", 173 TRACE_RDMA_REQUEST_STATE_EXECUTING, 174 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 175 spdk_trace_register_description("RDMA_REQ_EXECUTED", 176 TRACE_RDMA_REQUEST_STATE_EXECUTED, 177 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 178 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 179 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 180 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 181 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 182 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 183 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 184 spdk_trace_register_description("RDMA_REQ_COMPLETING", 185 TRACE_RDMA_REQUEST_STATE_COMPLETING, 186 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 187 spdk_trace_register_description("RDMA_REQ_COMPLETED", 188 TRACE_RDMA_REQUEST_STATE_COMPLETED, 189 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 190 191 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 192 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 193 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 194 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 195 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 196 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 197 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 198 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 199 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 200 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 201 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 202 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 203 } 204 205 enum spdk_nvmf_rdma_wr_type { 206 RDMA_WR_TYPE_RECV, 207 RDMA_WR_TYPE_SEND, 208 RDMA_WR_TYPE_DATA, 209 }; 210 211 struct spdk_nvmf_rdma_wr { 212 enum spdk_nvmf_rdma_wr_type type; 213 }; 214 215 /* This structure holds commands as they are received off the wire. 216 * It must be dynamically paired with a full request object 217 * (spdk_nvmf_rdma_request) to service a request. It is separate 218 * from the request because RDMA does not appear to order 219 * completions, so occasionally we'll get a new incoming 220 * command when there aren't any free request objects. 221 */ 222 struct spdk_nvmf_rdma_recv { 223 struct ibv_recv_wr wr; 224 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 225 226 struct spdk_nvmf_rdma_qpair *qpair; 227 228 /* In-capsule data buffer */ 229 uint8_t *buf; 230 231 struct spdk_nvmf_rdma_wr rdma_wr; 232 uint64_t receive_tsc; 233 234 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 235 }; 236 237 struct spdk_nvmf_rdma_request_data { 238 struct spdk_nvmf_rdma_wr rdma_wr; 239 struct ibv_send_wr wr; 240 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 241 }; 242 243 struct spdk_nvmf_rdma_request { 244 struct spdk_nvmf_request req; 245 246 enum spdk_nvmf_rdma_request_state state; 247 248 struct spdk_nvmf_rdma_recv *recv; 249 250 struct { 251 struct spdk_nvmf_rdma_wr rdma_wr; 252 struct ibv_send_wr wr; 253 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 254 } rsp; 255 256 struct spdk_nvmf_rdma_request_data data; 257 258 uint32_t iovpos; 259 260 uint32_t num_outstanding_data_wr; 261 uint64_t receive_tsc; 262 263 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 264 }; 265 266 struct spdk_nvmf_rdma_resource_opts { 267 struct spdk_nvmf_rdma_qpair *qpair; 268 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 269 void *qp; 270 struct ibv_pd *pd; 271 uint32_t max_queue_depth; 272 uint32_t in_capsule_data_size; 273 bool shared; 274 }; 275 276 struct spdk_nvmf_recv_wr_list { 277 struct ibv_recv_wr *first; 278 struct ibv_recv_wr *last; 279 }; 280 281 struct spdk_nvmf_rdma_resources { 282 /* Array of size "max_queue_depth" containing RDMA requests. */ 283 struct spdk_nvmf_rdma_request *reqs; 284 285 /* Array of size "max_queue_depth" containing RDMA recvs. */ 286 struct spdk_nvmf_rdma_recv *recvs; 287 288 /* Array of size "max_queue_depth" containing 64 byte capsules 289 * used for receive. 290 */ 291 union nvmf_h2c_msg *cmds; 292 struct ibv_mr *cmds_mr; 293 294 /* Array of size "max_queue_depth" containing 16 byte completions 295 * to be sent back to the user. 296 */ 297 union nvmf_c2h_msg *cpls; 298 struct ibv_mr *cpls_mr; 299 300 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 301 * buffers to be used for in capsule data. 302 */ 303 void *bufs; 304 struct ibv_mr *bufs_mr; 305 306 /* The list of pending recvs to transfer */ 307 struct spdk_nvmf_recv_wr_list recvs_to_post; 308 309 /* Receives that are waiting for a request object */ 310 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 311 312 /* Queue to track free requests */ 313 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 314 }; 315 316 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 317 318 struct spdk_nvmf_rdma_ibv_event_ctx { 319 struct spdk_nvmf_rdma_qpair *rqpair; 320 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 321 /* Link to other ibv events associated with this qpair */ 322 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 323 }; 324 325 struct spdk_nvmf_rdma_qpair { 326 struct spdk_nvmf_qpair qpair; 327 328 struct spdk_nvmf_rdma_device *device; 329 struct spdk_nvmf_rdma_poller *poller; 330 331 struct spdk_rdma_qp *rdma_qp; 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 struct spdk_nvmf_rdma_resources *resources; 365 366 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 367 368 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 369 370 /* Number of requests not in the free state */ 371 uint32_t qd; 372 373 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 374 375 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 376 377 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 378 379 /* IBV queue pair attributes: they are used to manage 380 * qp state and recover from errors. 381 */ 382 enum ibv_qp_state ibv_state; 383 384 /* 385 * io_channel which is used to destroy qpair when it is removed from poll group 386 */ 387 struct spdk_io_channel *destruct_channel; 388 389 /* List of ibv async events */ 390 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 391 392 /* Lets us know that we have received the last_wqe event. */ 393 bool last_wqe_reached; 394 395 /* Indicate that nvmf_rdma_close_qpair is called */ 396 bool to_close; 397 }; 398 399 struct spdk_nvmf_rdma_poller_stat { 400 uint64_t completions; 401 uint64_t polls; 402 uint64_t requests; 403 uint64_t request_latency; 404 uint64_t pending_free_request; 405 uint64_t pending_rdma_read; 406 uint64_t pending_rdma_write; 407 }; 408 409 struct spdk_nvmf_rdma_poller { 410 struct spdk_nvmf_rdma_device *device; 411 struct spdk_nvmf_rdma_poll_group *group; 412 413 int num_cqe; 414 int required_num_wr; 415 struct ibv_cq *cq; 416 417 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 418 uint16_t max_srq_depth; 419 420 /* Shared receive queue */ 421 struct ibv_srq *srq; 422 423 struct spdk_nvmf_rdma_resources *resources; 424 struct spdk_nvmf_rdma_poller_stat stat; 425 426 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 427 428 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 429 430 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 431 432 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 433 }; 434 435 struct spdk_nvmf_rdma_poll_group_stat { 436 uint64_t pending_data_buffer; 437 }; 438 439 struct spdk_nvmf_rdma_poll_group { 440 struct spdk_nvmf_transport_poll_group group; 441 struct spdk_nvmf_rdma_poll_group_stat stat; 442 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 443 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 444 /* 445 * buffers which are split across multiple RDMA 446 * memory regions cannot be used by this transport. 447 */ 448 STAILQ_HEAD(, spdk_nvmf_transport_pg_cache_buf) retired_bufs; 449 }; 450 451 struct spdk_nvmf_rdma_conn_sched { 452 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 453 struct spdk_nvmf_rdma_poll_group *next_io_pg; 454 }; 455 456 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 457 struct spdk_nvmf_rdma_device { 458 struct ibv_device_attr attr; 459 struct ibv_context *context; 460 461 struct spdk_mem_map *map; 462 struct ibv_pd *pd; 463 464 int num_srq; 465 466 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 467 }; 468 469 struct spdk_nvmf_rdma_port { 470 const struct spdk_nvme_transport_id *trid; 471 struct rdma_cm_id *id; 472 struct spdk_nvmf_rdma_device *device; 473 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 474 }; 475 476 struct spdk_nvmf_rdma_transport { 477 struct spdk_nvmf_transport transport; 478 479 struct spdk_nvmf_rdma_conn_sched conn_sched; 480 481 struct rdma_event_channel *event_channel; 482 483 struct spdk_mempool *data_wr_pool; 484 485 pthread_mutex_t lock; 486 487 /* fields used to poll RDMA/IB events */ 488 nfds_t npoll_fds; 489 struct pollfd *poll_fds; 490 491 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 492 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 493 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 494 }; 495 496 static bool 497 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 498 struct spdk_nvmf_rdma_request *rdma_req); 499 500 static inline int 501 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 502 { 503 switch (state) { 504 case IBV_QPS_RESET: 505 case IBV_QPS_INIT: 506 case IBV_QPS_RTR: 507 case IBV_QPS_RTS: 508 case IBV_QPS_SQD: 509 case IBV_QPS_SQE: 510 case IBV_QPS_ERR: 511 return 0; 512 default: 513 return -1; 514 } 515 } 516 517 static inline enum spdk_nvme_media_error_status_code 518 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 519 enum spdk_nvme_media_error_status_code result; 520 switch (err_type) 521 { 522 case SPDK_DIF_REFTAG_ERROR: 523 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 524 break; 525 case SPDK_DIF_APPTAG_ERROR: 526 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 527 break; 528 case SPDK_DIF_GUARD_ERROR: 529 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 530 break; 531 default: 532 SPDK_UNREACHABLE(); 533 } 534 535 return result; 536 } 537 538 static enum ibv_qp_state 539 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 540 enum ibv_qp_state old_state, new_state; 541 struct ibv_qp_attr qp_attr; 542 struct ibv_qp_init_attr init_attr; 543 int rc; 544 545 old_state = rqpair->ibv_state; 546 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 547 g_spdk_nvmf_ibv_query_mask, &init_attr); 548 549 if (rc) 550 { 551 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 552 return IBV_QPS_ERR + 1; 553 } 554 555 new_state = qp_attr.qp_state; 556 rqpair->ibv_state = new_state; 557 qp_attr.ah_attr.port_num = qp_attr.port_num; 558 559 rc = nvmf_rdma_check_ibv_state(new_state); 560 if (rc) 561 { 562 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 563 /* 564 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 565 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 566 */ 567 return IBV_QPS_ERR + 1; 568 } 569 570 if (old_state != new_state) 571 { 572 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 573 (uintptr_t)rqpair->cm_id, new_state); 574 } 575 return new_state; 576 } 577 578 static void 579 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 580 struct spdk_nvmf_rdma_transport *rtransport) 581 { 582 struct spdk_nvmf_rdma_request_data *data_wr; 583 struct ibv_send_wr *next_send_wr; 584 uint64_t req_wrid; 585 586 rdma_req->num_outstanding_data_wr = 0; 587 data_wr = &rdma_req->data; 588 req_wrid = data_wr->wr.wr_id; 589 while (data_wr && data_wr->wr.wr_id == req_wrid) { 590 memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge); 591 data_wr->wr.num_sge = 0; 592 next_send_wr = data_wr->wr.next; 593 if (data_wr != &rdma_req->data) { 594 spdk_mempool_put(rtransport->data_wr_pool, data_wr); 595 } 596 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : 597 SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr); 598 } 599 } 600 601 static void 602 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 603 { 604 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 605 if (req->req.cmd) { 606 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 607 } 608 if (req->recv) { 609 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 610 } 611 } 612 613 static void 614 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 615 { 616 int i; 617 618 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 619 for (i = 0; i < rqpair->max_queue_depth; i++) { 620 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 621 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 622 } 623 } 624 } 625 626 static void 627 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 628 { 629 if (resources->cmds_mr) { 630 ibv_dereg_mr(resources->cmds_mr); 631 } 632 633 if (resources->cpls_mr) { 634 ibv_dereg_mr(resources->cpls_mr); 635 } 636 637 if (resources->bufs_mr) { 638 ibv_dereg_mr(resources->bufs_mr); 639 } 640 641 spdk_free(resources->cmds); 642 spdk_free(resources->cpls); 643 spdk_free(resources->bufs); 644 free(resources->reqs); 645 free(resources->recvs); 646 free(resources); 647 } 648 649 650 static struct spdk_nvmf_rdma_resources * 651 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 652 { 653 struct spdk_nvmf_rdma_resources *resources; 654 struct spdk_nvmf_rdma_request *rdma_req; 655 struct spdk_nvmf_rdma_recv *rdma_recv; 656 struct ibv_qp *qp; 657 struct ibv_srq *srq; 658 uint32_t i; 659 int rc; 660 661 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 662 if (!resources) { 663 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 664 return NULL; 665 } 666 667 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 668 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 669 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 670 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 671 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 672 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 673 674 if (opts->in_capsule_data_size > 0) { 675 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 676 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 677 SPDK_MALLOC_DMA); 678 } 679 680 if (!resources->reqs || !resources->recvs || !resources->cmds || 681 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 682 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 683 goto cleanup; 684 } 685 686 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 687 opts->max_queue_depth * sizeof(*resources->cmds), 688 IBV_ACCESS_LOCAL_WRITE); 689 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 690 opts->max_queue_depth * sizeof(*resources->cpls), 691 0); 692 693 if (opts->in_capsule_data_size) { 694 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 695 opts->max_queue_depth * 696 opts->in_capsule_data_size, 697 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 698 } 699 700 if (!resources->cmds_mr || !resources->cpls_mr || 701 (opts->in_capsule_data_size && 702 !resources->bufs_mr)) { 703 goto cleanup; 704 } 705 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 706 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 707 resources->cmds_mr->lkey); 708 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 709 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 710 resources->cpls_mr->lkey); 711 if (resources->bufs && resources->bufs_mr) { 712 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 713 resources->bufs, opts->max_queue_depth * 714 opts->in_capsule_data_size, resources->bufs_mr->lkey); 715 } 716 717 /* Initialize queues */ 718 STAILQ_INIT(&resources->incoming_queue); 719 STAILQ_INIT(&resources->free_queue); 720 721 for (i = 0; i < opts->max_queue_depth; i++) { 722 struct ibv_recv_wr *bad_wr = NULL; 723 724 rdma_recv = &resources->recvs[i]; 725 rdma_recv->qpair = opts->qpair; 726 727 /* Set up memory to receive commands */ 728 if (resources->bufs) { 729 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 730 opts->in_capsule_data_size)); 731 } 732 733 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 734 735 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 736 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 737 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 738 rdma_recv->wr.num_sge = 1; 739 740 if (rdma_recv->buf && resources->bufs_mr) { 741 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 742 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 743 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 744 rdma_recv->wr.num_sge++; 745 } 746 747 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 748 rdma_recv->wr.sg_list = rdma_recv->sgl; 749 if (opts->shared) { 750 srq = (struct ibv_srq *)opts->qp; 751 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 752 } else { 753 qp = (struct ibv_qp *)opts->qp; 754 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 755 } 756 if (rc) { 757 goto cleanup; 758 } 759 } 760 761 for (i = 0; i < opts->max_queue_depth; i++) { 762 rdma_req = &resources->reqs[i]; 763 764 if (opts->qpair != NULL) { 765 rdma_req->req.qpair = &opts->qpair->qpair; 766 } else { 767 rdma_req->req.qpair = NULL; 768 } 769 rdma_req->req.cmd = NULL; 770 771 /* Set up memory to send responses */ 772 rdma_req->req.rsp = &resources->cpls[i]; 773 774 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 775 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 776 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 777 778 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 779 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 780 rdma_req->rsp.wr.next = NULL; 781 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 782 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 783 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 784 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 785 786 /* Set up memory for data buffers */ 787 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 788 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 789 rdma_req->data.wr.next = NULL; 790 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 791 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 792 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 793 794 /* Initialize request state to FREE */ 795 rdma_req->state = RDMA_REQUEST_STATE_FREE; 796 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 797 } 798 799 return resources; 800 801 cleanup: 802 nvmf_rdma_resources_destroy(resources); 803 return NULL; 804 } 805 806 static void 807 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 808 { 809 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 810 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 811 ctx->rqpair = NULL; 812 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 813 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 814 } 815 } 816 817 static void 818 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 819 { 820 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 821 struct ibv_recv_wr *bad_recv_wr = NULL; 822 int rc; 823 824 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 825 826 if (rqpair->qd != 0) { 827 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 828 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 829 struct spdk_nvmf_rdma_transport, transport); 830 struct spdk_nvmf_rdma_request *req; 831 uint32_t i, max_req_count = 0; 832 833 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 834 835 if (rqpair->srq == NULL) { 836 nvmf_rdma_dump_qpair_contents(rqpair); 837 max_req_count = rqpair->max_queue_depth; 838 } else if (rqpair->poller && rqpair->resources) { 839 max_req_count = rqpair->poller->max_srq_depth; 840 } 841 842 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Release incomplete requests\n"); 843 for (i = 0; i < max_req_count; i++) { 844 req = &rqpair->resources->reqs[i]; 845 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 846 /* nvmf_rdma_request_process checks qpair ibv and internal state 847 * and completes a request */ 848 nvmf_rdma_request_process(rtransport, req); 849 } 850 } 851 assert(rqpair->qd == 0); 852 } 853 854 if (rqpair->poller) { 855 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 856 857 if (rqpair->srq != NULL && rqpair->resources != NULL) { 858 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 859 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 860 if (rqpair == rdma_recv->qpair) { 861 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 862 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 863 if (rc) { 864 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 865 } 866 } 867 } 868 } 869 } 870 871 if (rqpair->cm_id) { 872 if (rqpair->rdma_qp != NULL) { 873 spdk_rdma_qp_destroy(rqpair->rdma_qp); 874 rqpair->rdma_qp = NULL; 875 } 876 rdma_destroy_id(rqpair->cm_id); 877 878 if (rqpair->poller != NULL && rqpair->srq == NULL) { 879 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 880 } 881 } 882 883 if (rqpair->srq == NULL && rqpair->resources != NULL) { 884 nvmf_rdma_resources_destroy(rqpair->resources); 885 } 886 887 nvmf_rdma_qpair_clean_ibv_events(rqpair); 888 889 if (rqpair->destruct_channel) { 890 spdk_put_io_channel(rqpair->destruct_channel); 891 rqpair->destruct_channel = NULL; 892 } 893 894 free(rqpair); 895 } 896 897 static int 898 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 899 { 900 struct spdk_nvmf_rdma_poller *rpoller; 901 int rc, num_cqe, required_num_wr; 902 903 /* Enlarge CQ size dynamically */ 904 rpoller = rqpair->poller; 905 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 906 num_cqe = rpoller->num_cqe; 907 if (num_cqe < required_num_wr) { 908 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 909 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 910 } 911 912 if (rpoller->num_cqe != num_cqe) { 913 if (required_num_wr > device->attr.max_cqe) { 914 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 915 required_num_wr, device->attr.max_cqe); 916 return -1; 917 } 918 919 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 920 rc = ibv_resize_cq(rpoller->cq, num_cqe); 921 if (rc) { 922 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 923 return -1; 924 } 925 926 rpoller->num_cqe = num_cqe; 927 } 928 929 rpoller->required_num_wr = required_num_wr; 930 return 0; 931 } 932 933 static int 934 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 935 { 936 struct spdk_nvmf_rdma_qpair *rqpair; 937 struct spdk_nvmf_rdma_transport *rtransport; 938 struct spdk_nvmf_transport *transport; 939 struct spdk_nvmf_rdma_resource_opts opts; 940 struct spdk_nvmf_rdma_device *device; 941 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 942 943 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 944 device = rqpair->device; 945 946 qp_init_attr.qp_context = rqpair; 947 qp_init_attr.pd = device->pd; 948 qp_init_attr.send_cq = rqpair->poller->cq; 949 qp_init_attr.recv_cq = rqpair->poller->cq; 950 951 if (rqpair->srq) { 952 qp_init_attr.srq = rqpair->srq; 953 } else { 954 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 955 } 956 957 /* SEND, READ, and WRITE operations */ 958 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 959 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 960 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 961 962 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 963 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 964 goto error; 965 } 966 967 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 968 if (!rqpair->rdma_qp) { 969 goto error; 970 } 971 972 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 973 qp_init_attr.cap.max_send_wr); 974 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 975 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 976 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 977 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 978 979 if (rqpair->poller->srq == NULL) { 980 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 981 transport = &rtransport->transport; 982 983 opts.qp = rqpair->rdma_qp->qp; 984 opts.pd = rqpair->cm_id->pd; 985 opts.qpair = rqpair; 986 opts.shared = false; 987 opts.max_queue_depth = rqpair->max_queue_depth; 988 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 989 990 rqpair->resources = nvmf_rdma_resources_create(&opts); 991 992 if (!rqpair->resources) { 993 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 994 rdma_destroy_qp(rqpair->cm_id); 995 goto error; 996 } 997 } else { 998 rqpair->resources = rqpair->poller->resources; 999 } 1000 1001 rqpair->current_recv_depth = 0; 1002 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1003 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1004 1005 return 0; 1006 1007 error: 1008 rdma_destroy_id(rqpair->cm_id); 1009 rqpair->cm_id = NULL; 1010 return -1; 1011 } 1012 1013 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1014 /* This function accepts either a single wr or the first wr in a linked list. */ 1015 static void 1016 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1017 { 1018 struct ibv_recv_wr *last; 1019 1020 last = first; 1021 while (last->next != NULL) { 1022 last = last->next; 1023 } 1024 1025 if (rqpair->resources->recvs_to_post.first == NULL) { 1026 rqpair->resources->recvs_to_post.first = first; 1027 rqpair->resources->recvs_to_post.last = last; 1028 if (rqpair->srq == NULL) { 1029 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1030 } 1031 } else { 1032 rqpair->resources->recvs_to_post.last->next = first; 1033 rqpair->resources->recvs_to_post.last = last; 1034 } 1035 } 1036 1037 static int 1038 request_transfer_in(struct spdk_nvmf_request *req) 1039 { 1040 struct spdk_nvmf_rdma_request *rdma_req; 1041 struct spdk_nvmf_qpair *qpair; 1042 struct spdk_nvmf_rdma_qpair *rqpair; 1043 1044 qpair = req->qpair; 1045 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1046 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1047 1048 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1049 assert(rdma_req != NULL); 1050 1051 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1052 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1053 } 1054 1055 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1056 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1057 return 0; 1058 } 1059 1060 static int 1061 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1062 { 1063 int num_outstanding_data_wr = 0; 1064 struct spdk_nvmf_rdma_request *rdma_req; 1065 struct spdk_nvmf_qpair *qpair; 1066 struct spdk_nvmf_rdma_qpair *rqpair; 1067 struct spdk_nvme_cpl *rsp; 1068 struct ibv_send_wr *first = NULL; 1069 1070 *data_posted = 0; 1071 qpair = req->qpair; 1072 rsp = &req->rsp->nvme_cpl; 1073 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1074 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1075 1076 /* Advance our sq_head pointer */ 1077 if (qpair->sq_head == qpair->sq_head_max) { 1078 qpair->sq_head = 0; 1079 } else { 1080 qpair->sq_head++; 1081 } 1082 rsp->sqhd = qpair->sq_head; 1083 1084 /* queue the capsule for the recv buffer */ 1085 assert(rdma_req->recv != NULL); 1086 1087 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1088 1089 rdma_req->recv = NULL; 1090 assert(rqpair->current_recv_depth > 0); 1091 rqpair->current_recv_depth--; 1092 1093 /* Build the response which consists of optional 1094 * RDMA WRITEs to transfer data, plus an RDMA SEND 1095 * containing the response. 1096 */ 1097 first = &rdma_req->rsp.wr; 1098 1099 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1100 /* On failure, data was not read from the controller. So clear the 1101 * number of outstanding data WRs to zero. 1102 */ 1103 rdma_req->num_outstanding_data_wr = 0; 1104 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1105 first = &rdma_req->data.wr; 1106 *data_posted = 1; 1107 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1108 } 1109 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1110 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1111 } 1112 1113 /* +1 for the rsp wr */ 1114 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1115 1116 return 0; 1117 } 1118 1119 static int 1120 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1121 { 1122 struct spdk_nvmf_rdma_accept_private_data accept_data; 1123 struct rdma_conn_param ctrlr_event_data = {}; 1124 int rc; 1125 1126 accept_data.recfmt = 0; 1127 accept_data.crqsize = rqpair->max_queue_depth; 1128 1129 ctrlr_event_data.private_data = &accept_data; 1130 ctrlr_event_data.private_data_len = sizeof(accept_data); 1131 if (id->ps == RDMA_PS_TCP) { 1132 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1133 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1134 } 1135 1136 /* Configure infinite retries for the initiator side qpair. 1137 * When using a shared receive queue on the target side, 1138 * we need to pass this value to the initiator to prevent the 1139 * initiator side NIC from completing SEND requests back to the 1140 * initiator with status rnr_retry_count_exceeded. */ 1141 if (rqpair->srq != NULL) { 1142 ctrlr_event_data.rnr_retry_count = 0x7; 1143 } 1144 1145 /* When qpair is created without use of rdma cm API, an additional 1146 * information must be provided to initiator in the connection response: 1147 * whether qpair is using SRQ and its qp_num 1148 * Fields below are ignored by rdma cm if qpair has been 1149 * created using rdma cm API. */ 1150 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1151 ctrlr_event_data.qp_num = rqpair->rdma_qp->qp->qp_num; 1152 1153 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1154 if (rc) { 1155 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1156 } else { 1157 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1158 } 1159 1160 return rc; 1161 } 1162 1163 static void 1164 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1165 { 1166 struct spdk_nvmf_rdma_reject_private_data rej_data; 1167 1168 rej_data.recfmt = 0; 1169 rej_data.sts = error; 1170 1171 rdma_reject(id, &rej_data, sizeof(rej_data)); 1172 } 1173 1174 static int 1175 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1176 { 1177 struct spdk_nvmf_rdma_transport *rtransport; 1178 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1179 struct spdk_nvmf_rdma_port *port; 1180 struct rdma_conn_param *rdma_param = NULL; 1181 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1182 uint16_t max_queue_depth; 1183 uint16_t max_read_depth; 1184 1185 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1186 1187 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1188 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1189 1190 rdma_param = &event->param.conn; 1191 if (rdma_param->private_data == NULL || 1192 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1193 SPDK_ERRLOG("connect request: no private data provided\n"); 1194 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1195 return -1; 1196 } 1197 1198 private_data = rdma_param->private_data; 1199 if (private_data->recfmt != 0) { 1200 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1201 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1202 return -1; 1203 } 1204 1205 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1206 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1207 1208 port = event->listen_id->context; 1209 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1210 event->listen_id, event->listen_id->verbs, port); 1211 1212 /* Figure out the supported queue depth. This is a multi-step process 1213 * that takes into account hardware maximums, host provided values, 1214 * and our target's internal memory limits */ 1215 1216 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1217 1218 /* Start with the maximum queue depth allowed by the target */ 1219 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1220 max_read_depth = rtransport->transport.opts.max_queue_depth; 1221 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1222 rtransport->transport.opts.max_queue_depth); 1223 1224 /* Next check the local NIC's hardware limitations */ 1225 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1226 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1227 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1228 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1229 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1230 1231 /* Next check the remote NIC's hardware limitations */ 1232 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1233 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1234 rdma_param->initiator_depth, rdma_param->responder_resources); 1235 if (rdma_param->initiator_depth > 0) { 1236 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1237 } 1238 1239 /* Finally check for the host software requested values, which are 1240 * optional. */ 1241 if (rdma_param->private_data != NULL && 1242 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1243 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1244 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1245 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1246 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1247 } 1248 1249 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1250 max_queue_depth, max_read_depth); 1251 1252 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1253 if (rqpair == NULL) { 1254 SPDK_ERRLOG("Could not allocate new connection.\n"); 1255 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1256 return -1; 1257 } 1258 1259 rqpair->device = port->device; 1260 rqpair->max_queue_depth = max_queue_depth; 1261 rqpair->max_read_depth = max_read_depth; 1262 rqpair->cm_id = event->id; 1263 rqpair->listen_id = event->listen_id; 1264 rqpair->qpair.transport = transport; 1265 STAILQ_INIT(&rqpair->ibv_events); 1266 /* use qid from the private data to determine the qpair type 1267 qid will be set to the appropriate value when the controller is created */ 1268 rqpair->qpair.qid = private_data->qid; 1269 1270 event->id->context = &rqpair->qpair; 1271 1272 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1273 1274 return 0; 1275 } 1276 1277 static int 1278 nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1279 enum spdk_mem_map_notify_action action, 1280 void *vaddr, size_t size) 1281 { 1282 struct ibv_pd *pd = cb_ctx; 1283 struct ibv_mr *mr; 1284 int rc; 1285 1286 switch (action) { 1287 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1288 if (!g_nvmf_hooks.get_rkey) { 1289 mr = ibv_reg_mr(pd, vaddr, size, 1290 IBV_ACCESS_LOCAL_WRITE | 1291 IBV_ACCESS_REMOTE_READ | 1292 IBV_ACCESS_REMOTE_WRITE); 1293 if (mr == NULL) { 1294 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1295 return -1; 1296 } else { 1297 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1298 } 1299 } else { 1300 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1301 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1302 } 1303 break; 1304 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1305 if (!g_nvmf_hooks.get_rkey) { 1306 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1307 if (mr) { 1308 ibv_dereg_mr(mr); 1309 } 1310 } 1311 rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1312 break; 1313 default: 1314 SPDK_UNREACHABLE(); 1315 } 1316 1317 return rc; 1318 } 1319 1320 static int 1321 nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1322 { 1323 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1324 return addr_1 == addr_2; 1325 } 1326 1327 static inline void 1328 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1329 enum spdk_nvme_data_transfer xfer) 1330 { 1331 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1332 wr->opcode = IBV_WR_RDMA_WRITE; 1333 wr->send_flags = 0; 1334 wr->next = next; 1335 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1336 wr->opcode = IBV_WR_RDMA_READ; 1337 wr->send_flags = IBV_SEND_SIGNALED; 1338 wr->next = NULL; 1339 } else { 1340 assert(0); 1341 } 1342 } 1343 1344 static int 1345 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1346 struct spdk_nvmf_rdma_request *rdma_req, 1347 uint32_t num_sgl_descriptors) 1348 { 1349 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1350 struct spdk_nvmf_rdma_request_data *current_data_wr; 1351 uint32_t i; 1352 1353 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1354 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1355 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1356 return -EINVAL; 1357 } 1358 1359 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1360 return -ENOMEM; 1361 } 1362 1363 current_data_wr = &rdma_req->data; 1364 1365 for (i = 0; i < num_sgl_descriptors; i++) { 1366 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1367 current_data_wr->wr.next = &work_requests[i]->wr; 1368 current_data_wr = work_requests[i]; 1369 current_data_wr->wr.sg_list = current_data_wr->sgl; 1370 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1371 } 1372 1373 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1374 1375 return 0; 1376 } 1377 1378 static inline void 1379 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1380 { 1381 struct ibv_send_wr *wr = &rdma_req->data.wr; 1382 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1383 1384 wr->wr.rdma.rkey = sgl->keyed.key; 1385 wr->wr.rdma.remote_addr = sgl->address; 1386 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1387 } 1388 1389 static inline void 1390 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1391 { 1392 struct ibv_send_wr *wr = &rdma_req->data.wr; 1393 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1394 uint32_t i; 1395 int j; 1396 uint64_t remote_addr_offset = 0; 1397 1398 for (i = 0; i < num_wrs; ++i) { 1399 wr->wr.rdma.rkey = sgl->keyed.key; 1400 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1401 for (j = 0; j < wr->num_sge; ++j) { 1402 remote_addr_offset += wr->sg_list[j].length; 1403 } 1404 wr = wr->next; 1405 } 1406 } 1407 1408 /* This function is used in the rare case that we have a buffer split over multiple memory regions. */ 1409 static int 1410 nvmf_rdma_replace_buffer(struct spdk_nvmf_rdma_poll_group *rgroup, void **buf) 1411 { 1412 struct spdk_nvmf_transport_poll_group *group = &rgroup->group; 1413 struct spdk_nvmf_transport *transport = group->transport; 1414 struct spdk_nvmf_transport_pg_cache_buf *old_buf; 1415 void *new_buf; 1416 1417 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1418 group->buf_cache_count--; 1419 new_buf = STAILQ_FIRST(&group->buf_cache); 1420 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1421 assert(*buf != NULL); 1422 } else { 1423 new_buf = spdk_mempool_get(transport->data_buf_pool); 1424 } 1425 1426 if (*buf == NULL) { 1427 return -ENOMEM; 1428 } 1429 1430 old_buf = *buf; 1431 STAILQ_INSERT_HEAD(&rgroup->retired_bufs, old_buf, link); 1432 *buf = new_buf; 1433 return 0; 1434 } 1435 1436 static bool 1437 nvmf_rdma_get_lkey(struct spdk_nvmf_rdma_device *device, struct iovec *iov, 1438 uint32_t *_lkey) 1439 { 1440 uint64_t translation_len; 1441 uint32_t lkey; 1442 1443 translation_len = iov->iov_len; 1444 1445 if (!g_nvmf_hooks.get_rkey) { 1446 lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1447 (uint64_t)iov->iov_base, &translation_len))->lkey; 1448 } else { 1449 lkey = spdk_mem_map_translate(device->map, 1450 (uint64_t)iov->iov_base, &translation_len); 1451 } 1452 1453 if (spdk_unlikely(translation_len < iov->iov_len)) { 1454 return false; 1455 } 1456 1457 *_lkey = lkey; 1458 return true; 1459 } 1460 1461 static bool 1462 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device, 1463 struct iovec *iov, struct ibv_send_wr **_wr, 1464 uint32_t *_remaining_data_block, uint32_t *_offset, 1465 uint32_t *_num_extra_wrs, 1466 const struct spdk_dif_ctx *dif_ctx) 1467 { 1468 struct ibv_send_wr *wr = *_wr; 1469 struct ibv_sge *sg_ele = &wr->sg_list[wr->num_sge]; 1470 uint32_t lkey = 0; 1471 uint32_t remaining, data_block_size, md_size, sge_len; 1472 1473 if (spdk_unlikely(!nvmf_rdma_get_lkey(device, iov, &lkey))) { 1474 /* This is a very rare case that can occur when using DPDK version < 19.05 */ 1475 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions. Removing it from circulation.\n"); 1476 return false; 1477 } 1478 1479 if (spdk_likely(!dif_ctx)) { 1480 sg_ele->lkey = lkey; 1481 sg_ele->addr = (uintptr_t)(iov->iov_base); 1482 sg_ele->length = iov->iov_len; 1483 wr->num_sge++; 1484 } else { 1485 remaining = iov->iov_len - *_offset; 1486 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1487 md_size = dif_ctx->md_size; 1488 1489 while (remaining) { 1490 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1491 if (*_num_extra_wrs > 0 && wr->next) { 1492 *_wr = wr->next; 1493 wr = *_wr; 1494 wr->num_sge = 0; 1495 sg_ele = &wr->sg_list[wr->num_sge]; 1496 (*_num_extra_wrs)--; 1497 } else { 1498 break; 1499 } 1500 } 1501 sg_ele->lkey = lkey; 1502 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset); 1503 sge_len = spdk_min(remaining, *_remaining_data_block); 1504 sg_ele->length = sge_len; 1505 remaining -= sge_len; 1506 *_remaining_data_block -= sge_len; 1507 *_offset += sge_len; 1508 1509 sg_ele++; 1510 wr->num_sge++; 1511 1512 if (*_remaining_data_block == 0) { 1513 /* skip metadata */ 1514 *_offset += md_size; 1515 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1516 remaining -= spdk_min(remaining, md_size); 1517 *_remaining_data_block = data_block_size; 1518 } 1519 1520 if (remaining == 0) { 1521 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1522 the remaining metadata bits at the beginning of the next buffer */ 1523 *_offset -= iov->iov_len; 1524 } 1525 } 1526 } 1527 1528 return true; 1529 } 1530 1531 static int 1532 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1533 struct spdk_nvmf_rdma_device *device, 1534 struct spdk_nvmf_rdma_request *rdma_req, 1535 struct ibv_send_wr *wr, 1536 uint32_t length, 1537 uint32_t num_extra_wrs) 1538 { 1539 struct spdk_nvmf_request *req = &rdma_req->req; 1540 struct spdk_dif_ctx *dif_ctx = NULL; 1541 uint32_t remaining_data_block = 0; 1542 uint32_t offset = 0; 1543 1544 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 1545 dif_ctx = &rdma_req->req.dif.dif_ctx; 1546 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1547 } 1548 1549 wr->num_sge = 0; 1550 1551 while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1552 while (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr, 1553 &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) { 1554 if (nvmf_rdma_replace_buffer(rgroup, &req->buffers[rdma_req->iovpos]) == -ENOMEM) { 1555 return -ENOMEM; 1556 } 1557 req->iov[rdma_req->iovpos].iov_base = (void *)((uintptr_t)(req->buffers[rdma_req->iovpos] + 1558 NVMF_DATA_BUFFER_MASK) & 1559 ~NVMF_DATA_BUFFER_MASK); 1560 } 1561 1562 length -= req->iov[rdma_req->iovpos].iov_len; 1563 rdma_req->iovpos++; 1564 } 1565 1566 if (length) { 1567 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1568 return -EINVAL; 1569 } 1570 1571 return 0; 1572 } 1573 1574 static inline uint32_t 1575 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1576 { 1577 /* estimate the number of SG entries and WRs needed to process the request */ 1578 uint32_t num_sge = 0; 1579 uint32_t i; 1580 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1581 1582 for (i = 0; i < num_buffers && length > 0; i++) { 1583 uint32_t buffer_len = spdk_min(length, io_unit_size); 1584 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1585 1586 if (num_sge_in_block * block_size > buffer_len) { 1587 ++num_sge_in_block; 1588 } 1589 num_sge += num_sge_in_block; 1590 length -= buffer_len; 1591 } 1592 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1593 } 1594 1595 static int 1596 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1597 struct spdk_nvmf_rdma_device *device, 1598 struct spdk_nvmf_rdma_request *rdma_req, 1599 uint32_t length) 1600 { 1601 struct spdk_nvmf_rdma_qpair *rqpair; 1602 struct spdk_nvmf_rdma_poll_group *rgroup; 1603 struct spdk_nvmf_request *req = &rdma_req->req; 1604 struct ibv_send_wr *wr = &rdma_req->data.wr; 1605 int rc; 1606 uint32_t num_wrs = 1; 1607 1608 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1609 rgroup = rqpair->poller->group; 1610 1611 /* rdma wr specifics */ 1612 nvmf_rdma_setup_request(rdma_req); 1613 1614 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1615 length); 1616 if (rc != 0) { 1617 return rc; 1618 } 1619 1620 assert(req->iovcnt <= rqpair->max_send_sge); 1621 1622 rdma_req->iovpos = 0; 1623 1624 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1625 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1626 req->dif.dif_ctx.block_size); 1627 if (num_wrs > 1) { 1628 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1629 if (rc != 0) { 1630 goto err_exit; 1631 } 1632 } 1633 } 1634 1635 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1636 if (spdk_unlikely(rc != 0)) { 1637 goto err_exit; 1638 } 1639 1640 if (spdk_unlikely(num_wrs > 1)) { 1641 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1642 } 1643 1644 /* set the number of outstanding data WRs for this request. */ 1645 rdma_req->num_outstanding_data_wr = num_wrs; 1646 1647 return rc; 1648 1649 err_exit: 1650 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1651 nvmf_rdma_request_free_data(rdma_req, rtransport); 1652 req->iovcnt = 0; 1653 return rc; 1654 } 1655 1656 static int 1657 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1658 struct spdk_nvmf_rdma_device *device, 1659 struct spdk_nvmf_rdma_request *rdma_req) 1660 { 1661 struct spdk_nvmf_rdma_qpair *rqpair; 1662 struct spdk_nvmf_rdma_poll_group *rgroup; 1663 struct ibv_send_wr *current_wr; 1664 struct spdk_nvmf_request *req = &rdma_req->req; 1665 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1666 uint32_t num_sgl_descriptors; 1667 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES]; 1668 uint32_t i; 1669 int rc; 1670 1671 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1672 rgroup = rqpair->poller->group; 1673 1674 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1675 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1676 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1677 1678 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1679 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1680 1681 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1682 return -ENOMEM; 1683 } 1684 1685 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1686 for (i = 0; i < num_sgl_descriptors; i++) { 1687 if (spdk_likely(!req->dif.dif_insert_or_strip)) { 1688 lengths[i] = desc->keyed.length; 1689 } else { 1690 req->dif.orig_length += desc->keyed.length; 1691 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1692 req->dif.elba_length += lengths[i]; 1693 } 1694 desc++; 1695 } 1696 1697 rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport, 1698 lengths, num_sgl_descriptors); 1699 if (rc != 0) { 1700 nvmf_rdma_request_free_data(rdma_req, rtransport); 1701 return rc; 1702 } 1703 1704 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1705 current_wr = &rdma_req->data.wr; 1706 assert(current_wr != NULL); 1707 1708 req->length = 0; 1709 rdma_req->iovpos = 0; 1710 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1711 for (i = 0; i < num_sgl_descriptors; i++) { 1712 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1713 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1714 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1715 rc = -EINVAL; 1716 goto err_exit; 1717 } 1718 1719 current_wr->num_sge = 0; 1720 1721 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1722 if (rc != 0) { 1723 rc = -ENOMEM; 1724 goto err_exit; 1725 } 1726 1727 req->length += desc->keyed.length; 1728 current_wr->wr.rdma.rkey = desc->keyed.key; 1729 current_wr->wr.rdma.remote_addr = desc->address; 1730 current_wr = current_wr->next; 1731 desc++; 1732 } 1733 1734 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1735 /* Go back to the last descriptor in the list. */ 1736 desc--; 1737 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1738 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1739 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1740 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1741 } 1742 } 1743 #endif 1744 1745 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1746 1747 return 0; 1748 1749 err_exit: 1750 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1751 nvmf_rdma_request_free_data(rdma_req, rtransport); 1752 return rc; 1753 } 1754 1755 static int 1756 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1757 struct spdk_nvmf_rdma_device *device, 1758 struct spdk_nvmf_rdma_request *rdma_req) 1759 { 1760 struct spdk_nvmf_request *req = &rdma_req->req; 1761 struct spdk_nvme_cpl *rsp; 1762 struct spdk_nvme_sgl_descriptor *sgl; 1763 int rc; 1764 uint32_t length; 1765 1766 rsp = &req->rsp->nvme_cpl; 1767 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1768 1769 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1770 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1771 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1772 1773 length = sgl->keyed.length; 1774 if (length > rtransport->transport.opts.max_io_size) { 1775 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1776 length, rtransport->transport.opts.max_io_size); 1777 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1778 return -1; 1779 } 1780 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1781 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1782 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1783 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1784 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1785 } 1786 } 1787 #endif 1788 1789 /* fill request length and populate iovs */ 1790 req->length = length; 1791 1792 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1793 req->dif.orig_length = length; 1794 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1795 req->dif.elba_length = length; 1796 } 1797 1798 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1799 if (spdk_unlikely(rc < 0)) { 1800 if (rc == -EINVAL) { 1801 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1802 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1803 return -1; 1804 } 1805 /* No available buffers. Queue this request up. */ 1806 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1807 return 0; 1808 } 1809 1810 /* backward compatible */ 1811 req->data = req->iov[0].iov_base; 1812 1813 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1814 req->iovcnt); 1815 1816 return 0; 1817 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1818 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1819 uint64_t offset = sgl->address; 1820 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1821 1822 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1823 offset, sgl->unkeyed.length); 1824 1825 if (offset > max_len) { 1826 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1827 offset, max_len); 1828 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1829 return -1; 1830 } 1831 max_len -= (uint32_t)offset; 1832 1833 if (sgl->unkeyed.length > max_len) { 1834 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1835 sgl->unkeyed.length, max_len); 1836 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1837 return -1; 1838 } 1839 1840 rdma_req->num_outstanding_data_wr = 0; 1841 req->data = rdma_req->recv->buf + offset; 1842 req->data_from_pool = false; 1843 req->length = sgl->unkeyed.length; 1844 1845 req->iov[0].iov_base = req->data; 1846 req->iov[0].iov_len = req->length; 1847 req->iovcnt = 1; 1848 1849 return 0; 1850 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1851 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1852 1853 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1854 if (rc == -ENOMEM) { 1855 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1856 return 0; 1857 } else if (rc == -EINVAL) { 1858 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1859 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1860 return -1; 1861 } 1862 1863 /* backward compatible */ 1864 req->data = req->iov[0].iov_base; 1865 1866 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1867 req->iovcnt); 1868 1869 return 0; 1870 } 1871 1872 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1873 sgl->generic.type, sgl->generic.subtype); 1874 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1875 return -1; 1876 } 1877 1878 static void 1879 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1880 struct spdk_nvmf_rdma_transport *rtransport) 1881 { 1882 struct spdk_nvmf_rdma_qpair *rqpair; 1883 struct spdk_nvmf_rdma_poll_group *rgroup; 1884 1885 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1886 if (rdma_req->req.data_from_pool) { 1887 rgroup = rqpair->poller->group; 1888 1889 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1890 } 1891 nvmf_rdma_request_free_data(rdma_req, rtransport); 1892 rdma_req->req.length = 0; 1893 rdma_req->req.iovcnt = 0; 1894 rdma_req->req.data = NULL; 1895 rdma_req->rsp.wr.next = NULL; 1896 rdma_req->data.wr.next = NULL; 1897 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1898 rqpair->qd--; 1899 1900 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1901 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1902 } 1903 1904 bool 1905 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1906 struct spdk_nvmf_rdma_request *rdma_req) 1907 { 1908 struct spdk_nvmf_rdma_qpair *rqpair; 1909 struct spdk_nvmf_rdma_device *device; 1910 struct spdk_nvmf_rdma_poll_group *rgroup; 1911 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1912 int rc; 1913 struct spdk_nvmf_rdma_recv *rdma_recv; 1914 enum spdk_nvmf_rdma_request_state prev_state; 1915 bool progress = false; 1916 int data_posted; 1917 uint32_t num_blocks; 1918 1919 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1920 device = rqpair->device; 1921 rgroup = rqpair->poller->group; 1922 1923 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1924 1925 /* If the queue pair is in an error state, force the request to the completed state 1926 * to release resources. */ 1927 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1928 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1929 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1930 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1931 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1932 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1933 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1934 } 1935 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1936 } 1937 1938 /* The loop here is to allow for several back-to-back state changes. */ 1939 do { 1940 prev_state = rdma_req->state; 1941 1942 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1943 1944 switch (rdma_req->state) { 1945 case RDMA_REQUEST_STATE_FREE: 1946 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1947 * to escape this state. */ 1948 break; 1949 case RDMA_REQUEST_STATE_NEW: 1950 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1951 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1952 rdma_recv = rdma_req->recv; 1953 1954 /* The first element of the SGL is the NVMe command */ 1955 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1956 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1957 1958 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1959 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1960 break; 1961 } 1962 1963 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 1964 rdma_req->req.dif.dif_insert_or_strip = true; 1965 } 1966 1967 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1968 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1969 rdma_req->rsp.wr.imm_data = 0; 1970 #endif 1971 1972 /* The next state transition depends on the data transfer needs of this request. */ 1973 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 1974 1975 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 1976 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 1977 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 1978 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1979 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 1980 break; 1981 } 1982 1983 /* If no data to transfer, ready to execute. */ 1984 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1985 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1986 break; 1987 } 1988 1989 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1990 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 1991 break; 1992 case RDMA_REQUEST_STATE_NEED_BUFFER: 1993 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1994 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1995 1996 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1997 1998 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 1999 /* This request needs to wait in line to obtain a buffer */ 2000 break; 2001 } 2002 2003 /* Try to get a data buffer */ 2004 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2005 if (rc < 0) { 2006 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2007 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2008 break; 2009 } 2010 2011 if (!rdma_req->req.data) { 2012 /* No buffers available. */ 2013 rgroup->stat.pending_data_buffer++; 2014 break; 2015 } 2016 2017 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2018 2019 /* If data is transferring from host to controller and the data didn't 2020 * arrive using in capsule data, we need to do a transfer from the host. 2021 */ 2022 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2023 rdma_req->req.data_from_pool) { 2024 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2025 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2026 break; 2027 } 2028 2029 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2030 break; 2031 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2032 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2033 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2034 2035 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2036 /* This request needs to wait in line to perform RDMA */ 2037 break; 2038 } 2039 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2040 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2041 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2042 rqpair->poller->stat.pending_rdma_read++; 2043 break; 2044 } 2045 2046 /* We have already verified that this request is the head of the queue. */ 2047 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2048 2049 rc = request_transfer_in(&rdma_req->req); 2050 if (!rc) { 2051 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2052 } else { 2053 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2054 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2055 } 2056 break; 2057 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2058 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2059 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2060 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2061 * to escape this state. */ 2062 break; 2063 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2064 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2065 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2066 2067 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2068 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2069 /* generate DIF for write operation */ 2070 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2071 assert(num_blocks > 0); 2072 2073 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2074 num_blocks, &rdma_req->req.dif.dif_ctx); 2075 if (rc != 0) { 2076 SPDK_ERRLOG("DIF generation failed\n"); 2077 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2078 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2079 break; 2080 } 2081 } 2082 2083 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2084 /* set extended length before IO operation */ 2085 rdma_req->req.length = rdma_req->req.dif.elba_length; 2086 } 2087 2088 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2089 spdk_nvmf_request_exec(&rdma_req->req); 2090 break; 2091 case RDMA_REQUEST_STATE_EXECUTING: 2092 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2093 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2094 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2095 * to escape this state. */ 2096 break; 2097 case RDMA_REQUEST_STATE_EXECUTED: 2098 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2099 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2100 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2101 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2102 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2103 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2104 } else { 2105 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2106 } 2107 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2108 /* restore the original length */ 2109 rdma_req->req.length = rdma_req->req.dif.orig_length; 2110 2111 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2112 struct spdk_dif_error error_blk; 2113 2114 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2115 2116 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2117 &rdma_req->req.dif.dif_ctx, &error_blk); 2118 if (rc) { 2119 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2120 2121 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2122 error_blk.err_offset); 2123 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2124 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2125 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2126 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2127 } 2128 } 2129 } 2130 break; 2131 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2132 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2133 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2134 2135 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2136 /* This request needs to wait in line to perform RDMA */ 2137 break; 2138 } 2139 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2140 rqpair->max_send_depth) { 2141 /* We can only have so many WRs outstanding. we have to wait until some finish. 2142 * +1 since each request has an additional wr in the resp. */ 2143 rqpair->poller->stat.pending_rdma_write++; 2144 break; 2145 } 2146 2147 /* We have already verified that this request is the head of the queue. */ 2148 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2149 2150 /* The data transfer will be kicked off from 2151 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2152 */ 2153 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2154 break; 2155 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2156 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2157 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2158 rc = request_transfer_out(&rdma_req->req, &data_posted); 2159 assert(rc == 0); /* No good way to handle this currently */ 2160 if (rc) { 2161 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2162 } else { 2163 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2164 RDMA_REQUEST_STATE_COMPLETING; 2165 } 2166 break; 2167 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2168 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2169 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2170 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2171 * to escape this state. */ 2172 break; 2173 case RDMA_REQUEST_STATE_COMPLETING: 2174 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2175 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2176 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2177 * to escape this state. */ 2178 break; 2179 case RDMA_REQUEST_STATE_COMPLETED: 2180 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2181 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2182 2183 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2184 _nvmf_rdma_request_free(rdma_req, rtransport); 2185 break; 2186 case RDMA_REQUEST_NUM_STATES: 2187 default: 2188 assert(0); 2189 break; 2190 } 2191 2192 if (rdma_req->state != prev_state) { 2193 progress = true; 2194 } 2195 } while (rdma_req->state != prev_state); 2196 2197 return progress; 2198 } 2199 2200 /* Public API callbacks begin here */ 2201 2202 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2203 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2204 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2205 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2206 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2207 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2208 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2209 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2210 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2211 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2212 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2213 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2214 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2215 2216 static void 2217 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2218 { 2219 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2220 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2221 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2222 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2223 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2224 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2225 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2226 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2227 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2228 opts->no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2229 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2230 opts->acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2231 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2232 } 2233 2234 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2235 .notify_cb = nvmf_rdma_mem_notify, 2236 .are_contiguous = nvmf_rdma_check_contiguous_entries 2237 }; 2238 2239 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2240 2241 static struct spdk_nvmf_transport * 2242 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2243 { 2244 int rc; 2245 struct spdk_nvmf_rdma_transport *rtransport; 2246 struct spdk_nvmf_rdma_device *device, *tmp; 2247 struct ibv_context **contexts; 2248 uint32_t i; 2249 int flag; 2250 uint32_t sge_count; 2251 uint32_t min_shared_buffers; 2252 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2253 pthread_mutexattr_t attr; 2254 2255 rtransport = calloc(1, sizeof(*rtransport)); 2256 if (!rtransport) { 2257 return NULL; 2258 } 2259 2260 if (pthread_mutexattr_init(&attr)) { 2261 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2262 free(rtransport); 2263 return NULL; 2264 } 2265 2266 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2267 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2268 pthread_mutexattr_destroy(&attr); 2269 free(rtransport); 2270 return NULL; 2271 } 2272 2273 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2274 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2275 pthread_mutexattr_destroy(&attr); 2276 free(rtransport); 2277 return NULL; 2278 } 2279 2280 pthread_mutexattr_destroy(&attr); 2281 2282 TAILQ_INIT(&rtransport->devices); 2283 TAILQ_INIT(&rtransport->ports); 2284 TAILQ_INIT(&rtransport->poll_groups); 2285 2286 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2287 2288 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2289 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2290 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2291 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2292 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d," 2293 " acceptor_backlog=%d, abort_timeout_sec=%d\n", 2294 opts->max_queue_depth, 2295 opts->max_io_size, 2296 opts->max_qpairs_per_ctrlr - 1, 2297 opts->io_unit_size, 2298 opts->in_capsule_data_size, 2299 opts->max_aq_depth, 2300 opts->num_shared_buffers, 2301 opts->max_srq_depth, 2302 opts->no_srq, 2303 opts->acceptor_backlog, 2304 opts->abort_timeout_sec); 2305 2306 /* I/O unit size cannot be larger than max I/O size */ 2307 if (opts->io_unit_size > opts->max_io_size) { 2308 opts->io_unit_size = opts->max_io_size; 2309 } 2310 2311 if (opts->acceptor_backlog <= 0) { 2312 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2313 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2314 opts->acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2315 } 2316 2317 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2318 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2319 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2320 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2321 nvmf_rdma_destroy(&rtransport->transport); 2322 return NULL; 2323 } 2324 2325 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2326 if (min_shared_buffers > opts->num_shared_buffers) { 2327 SPDK_ERRLOG("There are not enough buffers to satisfy" 2328 "per-poll group caches for each thread. (%" PRIu32 ")" 2329 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2330 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2331 nvmf_rdma_destroy(&rtransport->transport); 2332 return NULL; 2333 } 2334 2335 sge_count = opts->max_io_size / opts->io_unit_size; 2336 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2337 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2338 nvmf_rdma_destroy(&rtransport->transport); 2339 return NULL; 2340 } 2341 2342 rtransport->event_channel = rdma_create_event_channel(); 2343 if (rtransport->event_channel == NULL) { 2344 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2345 nvmf_rdma_destroy(&rtransport->transport); 2346 return NULL; 2347 } 2348 2349 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2350 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2351 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2352 rtransport->event_channel->fd, spdk_strerror(errno)); 2353 nvmf_rdma_destroy(&rtransport->transport); 2354 return NULL; 2355 } 2356 2357 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2358 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2359 sizeof(struct spdk_nvmf_rdma_request_data), 2360 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2361 SPDK_ENV_SOCKET_ID_ANY); 2362 if (!rtransport->data_wr_pool) { 2363 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2364 nvmf_rdma_destroy(&rtransport->transport); 2365 return NULL; 2366 } 2367 2368 contexts = rdma_get_devices(NULL); 2369 if (contexts == NULL) { 2370 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2371 nvmf_rdma_destroy(&rtransport->transport); 2372 return NULL; 2373 } 2374 2375 i = 0; 2376 rc = 0; 2377 while (contexts[i] != NULL) { 2378 device = calloc(1, sizeof(*device)); 2379 if (!device) { 2380 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2381 rc = -ENOMEM; 2382 break; 2383 } 2384 device->context = contexts[i]; 2385 rc = ibv_query_device(device->context, &device->attr); 2386 if (rc < 0) { 2387 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2388 free(device); 2389 break; 2390 2391 } 2392 2393 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2394 2395 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2396 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2397 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2398 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2399 } 2400 2401 /** 2402 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2403 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2404 * but incorrectly reports that it does. There are changes making their way 2405 * through the kernel now that will enable this feature. When they are merged, 2406 * we can conditionally enable this feature. 2407 * 2408 * TODO: enable this for versions of the kernel rxe driver that support it. 2409 */ 2410 if (device->attr.vendor_id == 0) { 2411 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2412 } 2413 #endif 2414 2415 /* set up device context async ev fd as NON_BLOCKING */ 2416 flag = fcntl(device->context->async_fd, F_GETFL); 2417 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2418 if (rc < 0) { 2419 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2420 free(device); 2421 break; 2422 } 2423 2424 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2425 i++; 2426 2427 if (g_nvmf_hooks.get_ibv_pd) { 2428 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2429 } else { 2430 device->pd = ibv_alloc_pd(device->context); 2431 } 2432 2433 if (!device->pd) { 2434 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2435 rc = -ENOMEM; 2436 break; 2437 } 2438 2439 assert(device->map == NULL); 2440 2441 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2442 if (!device->map) { 2443 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2444 rc = -ENOMEM; 2445 break; 2446 } 2447 2448 assert(device->map != NULL); 2449 assert(device->pd != NULL); 2450 } 2451 rdma_free_devices(contexts); 2452 2453 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2454 /* divide and round up. */ 2455 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2456 2457 /* round up to the nearest 4k. */ 2458 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2459 2460 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2461 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2462 opts->io_unit_size); 2463 } 2464 2465 if (rc < 0) { 2466 nvmf_rdma_destroy(&rtransport->transport); 2467 return NULL; 2468 } 2469 2470 /* Set up poll descriptor array to monitor events from RDMA and IB 2471 * in a single poll syscall 2472 */ 2473 rtransport->npoll_fds = i + 1; 2474 i = 0; 2475 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2476 if (rtransport->poll_fds == NULL) { 2477 SPDK_ERRLOG("poll_fds allocation failed\n"); 2478 nvmf_rdma_destroy(&rtransport->transport); 2479 return NULL; 2480 } 2481 2482 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2483 rtransport->poll_fds[i++].events = POLLIN; 2484 2485 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2486 rtransport->poll_fds[i].fd = device->context->async_fd; 2487 rtransport->poll_fds[i++].events = POLLIN; 2488 } 2489 2490 return &rtransport->transport; 2491 } 2492 2493 static int 2494 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2495 { 2496 struct spdk_nvmf_rdma_transport *rtransport; 2497 struct spdk_nvmf_rdma_port *port, *port_tmp; 2498 struct spdk_nvmf_rdma_device *device, *device_tmp; 2499 2500 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2501 2502 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2503 TAILQ_REMOVE(&rtransport->ports, port, link); 2504 rdma_destroy_id(port->id); 2505 free(port); 2506 } 2507 2508 if (rtransport->poll_fds != NULL) { 2509 free(rtransport->poll_fds); 2510 } 2511 2512 if (rtransport->event_channel != NULL) { 2513 rdma_destroy_event_channel(rtransport->event_channel); 2514 } 2515 2516 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2517 TAILQ_REMOVE(&rtransport->devices, device, link); 2518 if (device->map) { 2519 spdk_mem_map_free(&device->map); 2520 } 2521 if (device->pd) { 2522 if (!g_nvmf_hooks.get_ibv_pd) { 2523 ibv_dealloc_pd(device->pd); 2524 } 2525 } 2526 free(device); 2527 } 2528 2529 if (rtransport->data_wr_pool != NULL) { 2530 if (spdk_mempool_count(rtransport->data_wr_pool) != 2531 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2532 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2533 spdk_mempool_count(rtransport->data_wr_pool), 2534 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2535 } 2536 } 2537 2538 spdk_mempool_free(rtransport->data_wr_pool); 2539 2540 pthread_mutex_destroy(&rtransport->lock); 2541 free(rtransport); 2542 2543 return 0; 2544 } 2545 2546 static int 2547 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2548 struct spdk_nvme_transport_id *trid, 2549 bool peer); 2550 2551 static int 2552 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2553 const struct spdk_nvme_transport_id *trid) 2554 { 2555 struct spdk_nvmf_rdma_transport *rtransport; 2556 struct spdk_nvmf_rdma_device *device; 2557 struct spdk_nvmf_rdma_port *port; 2558 struct addrinfo *res; 2559 struct addrinfo hints; 2560 int family; 2561 int rc; 2562 2563 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2564 assert(rtransport->event_channel != NULL); 2565 2566 pthread_mutex_lock(&rtransport->lock); 2567 port = calloc(1, sizeof(*port)); 2568 if (!port) { 2569 SPDK_ERRLOG("Port allocation failed\n"); 2570 pthread_mutex_unlock(&rtransport->lock); 2571 return -ENOMEM; 2572 } 2573 2574 port->trid = trid; 2575 2576 switch (trid->adrfam) { 2577 case SPDK_NVMF_ADRFAM_IPV4: 2578 family = AF_INET; 2579 break; 2580 case SPDK_NVMF_ADRFAM_IPV6: 2581 family = AF_INET6; 2582 break; 2583 default: 2584 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2585 free(port); 2586 pthread_mutex_unlock(&rtransport->lock); 2587 return -EINVAL; 2588 } 2589 2590 memset(&hints, 0, sizeof(hints)); 2591 hints.ai_family = family; 2592 hints.ai_flags = AI_NUMERICSERV; 2593 hints.ai_socktype = SOCK_STREAM; 2594 hints.ai_protocol = 0; 2595 2596 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2597 if (rc) { 2598 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2599 free(port); 2600 pthread_mutex_unlock(&rtransport->lock); 2601 return -EINVAL; 2602 } 2603 2604 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2605 if (rc < 0) { 2606 SPDK_ERRLOG("rdma_create_id() failed\n"); 2607 freeaddrinfo(res); 2608 free(port); 2609 pthread_mutex_unlock(&rtransport->lock); 2610 return rc; 2611 } 2612 2613 rc = rdma_bind_addr(port->id, res->ai_addr); 2614 freeaddrinfo(res); 2615 2616 if (rc < 0) { 2617 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2618 rdma_destroy_id(port->id); 2619 free(port); 2620 pthread_mutex_unlock(&rtransport->lock); 2621 return rc; 2622 } 2623 2624 if (!port->id->verbs) { 2625 SPDK_ERRLOG("ibv_context is null\n"); 2626 rdma_destroy_id(port->id); 2627 free(port); 2628 pthread_mutex_unlock(&rtransport->lock); 2629 return -1; 2630 } 2631 2632 rc = rdma_listen(port->id, transport->opts.acceptor_backlog); 2633 if (rc < 0) { 2634 SPDK_ERRLOG("rdma_listen() failed\n"); 2635 rdma_destroy_id(port->id); 2636 free(port); 2637 pthread_mutex_unlock(&rtransport->lock); 2638 return rc; 2639 } 2640 2641 TAILQ_FOREACH(device, &rtransport->devices, link) { 2642 if (device->context == port->id->verbs) { 2643 port->device = device; 2644 break; 2645 } 2646 } 2647 if (!port->device) { 2648 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2649 port->id->verbs); 2650 rdma_destroy_id(port->id); 2651 free(port); 2652 pthread_mutex_unlock(&rtransport->lock); 2653 return -EINVAL; 2654 } 2655 2656 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2657 trid->traddr, trid->trsvcid); 2658 2659 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2660 pthread_mutex_unlock(&rtransport->lock); 2661 return 0; 2662 } 2663 2664 static void 2665 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2666 const struct spdk_nvme_transport_id *trid) 2667 { 2668 struct spdk_nvmf_rdma_transport *rtransport; 2669 struct spdk_nvmf_rdma_port *port, *tmp; 2670 2671 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2672 2673 pthread_mutex_lock(&rtransport->lock); 2674 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2675 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2676 TAILQ_REMOVE(&rtransport->ports, port, link); 2677 rdma_destroy_id(port->id); 2678 free(port); 2679 break; 2680 } 2681 } 2682 2683 pthread_mutex_unlock(&rtransport->lock); 2684 } 2685 2686 static void 2687 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2688 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2689 { 2690 struct spdk_nvmf_request *req, *tmp; 2691 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2692 struct spdk_nvmf_rdma_resources *resources; 2693 2694 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2695 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2696 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2697 break; 2698 } 2699 } 2700 2701 /* Then RDMA writes since reads have stronger restrictions than writes */ 2702 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2703 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2704 break; 2705 } 2706 } 2707 2708 /* The second highest priority is I/O waiting on memory buffers. */ 2709 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2710 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2711 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2712 break; 2713 } 2714 } 2715 2716 resources = rqpair->resources; 2717 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2718 rdma_req = STAILQ_FIRST(&resources->free_queue); 2719 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2720 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2721 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2722 2723 if (rqpair->srq != NULL) { 2724 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2725 rdma_req->recv->qpair->qd++; 2726 } else { 2727 rqpair->qd++; 2728 } 2729 2730 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2731 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2732 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2733 break; 2734 } 2735 } 2736 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2737 rqpair->poller->stat.pending_free_request++; 2738 } 2739 } 2740 2741 static void 2742 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2743 { 2744 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2745 struct spdk_nvmf_rdma_transport, transport); 2746 2747 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2748 2749 /* nvmr_rdma_close_qpair is not called */ 2750 if (!rqpair->to_close) { 2751 return; 2752 } 2753 2754 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2755 if (rqpair->current_send_depth != 0) { 2756 return; 2757 } 2758 2759 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2760 return; 2761 } 2762 2763 /* Judge whether the device is emulated by Software RoCE. 2764 * And it will not send last_wqe event 2765 */ 2766 if (rqpair->srq != NULL && rqpair->device->attr.vendor_id != 0 && 2767 rqpair->last_wqe_reached == false) { 2768 return; 2769 } 2770 2771 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2772 2773 nvmf_rdma_qpair_destroy(rqpair); 2774 } 2775 2776 static int 2777 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2778 { 2779 struct spdk_nvmf_qpair *qpair; 2780 struct spdk_nvmf_rdma_qpair *rqpair; 2781 2782 if (evt->id == NULL) { 2783 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2784 return -1; 2785 } 2786 2787 qpair = evt->id->context; 2788 if (qpair == NULL) { 2789 SPDK_ERRLOG("disconnect request: no active connection\n"); 2790 return -1; 2791 } 2792 2793 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2794 2795 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2796 2797 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2798 2799 return 0; 2800 } 2801 2802 #ifdef DEBUG 2803 static const char *CM_EVENT_STR[] = { 2804 "RDMA_CM_EVENT_ADDR_RESOLVED", 2805 "RDMA_CM_EVENT_ADDR_ERROR", 2806 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2807 "RDMA_CM_EVENT_ROUTE_ERROR", 2808 "RDMA_CM_EVENT_CONNECT_REQUEST", 2809 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2810 "RDMA_CM_EVENT_CONNECT_ERROR", 2811 "RDMA_CM_EVENT_UNREACHABLE", 2812 "RDMA_CM_EVENT_REJECTED", 2813 "RDMA_CM_EVENT_ESTABLISHED", 2814 "RDMA_CM_EVENT_DISCONNECTED", 2815 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2816 "RDMA_CM_EVENT_MULTICAST_JOIN", 2817 "RDMA_CM_EVENT_MULTICAST_ERROR", 2818 "RDMA_CM_EVENT_ADDR_CHANGE", 2819 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2820 }; 2821 #endif /* DEBUG */ 2822 2823 static void 2824 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2825 struct spdk_nvmf_rdma_port *port) 2826 { 2827 struct spdk_nvmf_rdma_poll_group *rgroup; 2828 struct spdk_nvmf_rdma_poller *rpoller; 2829 struct spdk_nvmf_rdma_qpair *rqpair; 2830 2831 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2832 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2833 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 2834 if (rqpair->listen_id == port->id) { 2835 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2836 } 2837 } 2838 } 2839 } 2840 } 2841 2842 static bool 2843 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2844 struct rdma_cm_event *event) 2845 { 2846 const struct spdk_nvme_transport_id *trid; 2847 struct spdk_nvmf_rdma_port *port; 2848 struct spdk_nvmf_rdma_transport *rtransport; 2849 bool event_acked = false; 2850 2851 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2852 TAILQ_FOREACH(port, &rtransport->ports, link) { 2853 if (port->id == event->id) { 2854 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2855 rdma_ack_cm_event(event); 2856 event_acked = true; 2857 trid = port->trid; 2858 break; 2859 } 2860 } 2861 2862 if (event_acked) { 2863 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2864 2865 nvmf_rdma_stop_listen(transport, trid); 2866 nvmf_rdma_listen(transport, trid); 2867 } 2868 2869 return event_acked; 2870 } 2871 2872 static void 2873 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2874 struct rdma_cm_event *event) 2875 { 2876 struct spdk_nvmf_rdma_port *port; 2877 struct spdk_nvmf_rdma_transport *rtransport; 2878 2879 port = event->id->context; 2880 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2881 2882 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2883 2884 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2885 2886 rdma_ack_cm_event(event); 2887 2888 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2889 ; 2890 } 2891 } 2892 2893 static void 2894 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2895 { 2896 struct spdk_nvmf_rdma_transport *rtransport; 2897 struct rdma_cm_event *event; 2898 int rc; 2899 bool event_acked; 2900 2901 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2902 2903 if (rtransport->event_channel == NULL) { 2904 return; 2905 } 2906 2907 while (1) { 2908 event_acked = false; 2909 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2910 if (rc) { 2911 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2912 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2913 } 2914 break; 2915 } 2916 2917 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2918 2919 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2920 2921 switch (event->event) { 2922 case RDMA_CM_EVENT_ADDR_RESOLVED: 2923 case RDMA_CM_EVENT_ADDR_ERROR: 2924 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2925 case RDMA_CM_EVENT_ROUTE_ERROR: 2926 /* No action required. The target never attempts to resolve routes. */ 2927 break; 2928 case RDMA_CM_EVENT_CONNECT_REQUEST: 2929 rc = nvmf_rdma_connect(transport, event); 2930 if (rc < 0) { 2931 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2932 break; 2933 } 2934 break; 2935 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2936 /* The target never initiates a new connection. So this will not occur. */ 2937 break; 2938 case RDMA_CM_EVENT_CONNECT_ERROR: 2939 /* Can this happen? The docs say it can, but not sure what causes it. */ 2940 break; 2941 case RDMA_CM_EVENT_UNREACHABLE: 2942 case RDMA_CM_EVENT_REJECTED: 2943 /* These only occur on the client side. */ 2944 break; 2945 case RDMA_CM_EVENT_ESTABLISHED: 2946 /* TODO: Should we be waiting for this event anywhere? */ 2947 break; 2948 case RDMA_CM_EVENT_DISCONNECTED: 2949 rc = nvmf_rdma_disconnect(event); 2950 if (rc < 0) { 2951 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2952 break; 2953 } 2954 break; 2955 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2956 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 2957 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 2958 * Once these events are sent to SPDK, we should release all IB resources and 2959 * don't make attempts to call any ibv_query/modify/create functions. We can only call 2960 * ibv_destory* functions to release user space memory allocated by IB. All kernel 2961 * resources are already cleaned. */ 2962 if (event->id->qp) { 2963 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 2964 * corresponding qpair. Otherwise the event refers to a listening device */ 2965 rc = nvmf_rdma_disconnect(event); 2966 if (rc < 0) { 2967 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2968 break; 2969 } 2970 } else { 2971 nvmf_rdma_handle_cm_event_port_removal(transport, event); 2972 event_acked = true; 2973 } 2974 break; 2975 case RDMA_CM_EVENT_MULTICAST_JOIN: 2976 case RDMA_CM_EVENT_MULTICAST_ERROR: 2977 /* Multicast is not used */ 2978 break; 2979 case RDMA_CM_EVENT_ADDR_CHANGE: 2980 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 2981 break; 2982 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2983 /* For now, do nothing. The target never re-uses queue pairs. */ 2984 break; 2985 default: 2986 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2987 break; 2988 } 2989 if (!event_acked) { 2990 rdma_ack_cm_event(event); 2991 } 2992 } 2993 } 2994 2995 static void 2996 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 2997 { 2998 rqpair->last_wqe_reached = true; 2999 nvmf_rdma_destroy_drained_qpair(rqpair); 3000 } 3001 3002 static void 3003 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3004 { 3005 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3006 3007 if (event_ctx->rqpair) { 3008 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3009 if (event_ctx->cb_fn) { 3010 event_ctx->cb_fn(event_ctx->rqpair); 3011 } 3012 } 3013 free(event_ctx); 3014 } 3015 3016 static int 3017 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3018 spdk_nvmf_rdma_qpair_ibv_event fn) 3019 { 3020 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3021 struct spdk_thread *thr = NULL; 3022 int rc; 3023 3024 if (rqpair->qpair.group) { 3025 thr = rqpair->qpair.group->thread; 3026 } else if (rqpair->destruct_channel) { 3027 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3028 } 3029 3030 if (!thr) { 3031 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "rqpair %p has no thread\n", rqpair); 3032 return -EINVAL; 3033 } 3034 3035 ctx = calloc(1, sizeof(*ctx)); 3036 if (!ctx) { 3037 return -ENOMEM; 3038 } 3039 3040 ctx->rqpair = rqpair; 3041 ctx->cb_fn = fn; 3042 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3043 3044 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3045 if (rc) { 3046 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3047 free(ctx); 3048 } 3049 3050 return rc; 3051 } 3052 3053 static int 3054 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3055 { 3056 int rc; 3057 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3058 struct ibv_async_event event; 3059 3060 rc = ibv_get_async_event(device->context, &event); 3061 3062 if (rc) { 3063 /* In non-blocking mode -1 means there are no events available */ 3064 return rc; 3065 } 3066 3067 switch (event.event_type) { 3068 case IBV_EVENT_QP_FATAL: 3069 rqpair = event.element.qp->qp_context; 3070 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3071 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3072 (uintptr_t)rqpair->cm_id, event.event_type); 3073 nvmf_rdma_update_ibv_state(rqpair); 3074 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3075 break; 3076 case IBV_EVENT_QP_LAST_WQE_REACHED: 3077 /* This event only occurs for shared receive queues. */ 3078 rqpair = event.element.qp->qp_context; 3079 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last WQE reached event received for rqpair %p\n", rqpair); 3080 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3081 if (rc) { 3082 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3083 rqpair->last_wqe_reached = true; 3084 } 3085 break; 3086 case IBV_EVENT_SQ_DRAINED: 3087 /* This event occurs frequently in both error and non-error states. 3088 * Check if the qpair is in an error state before sending a message. */ 3089 rqpair = event.element.qp->qp_context; 3090 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last sq drained event received for rqpair %p\n", rqpair); 3091 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3092 (uintptr_t)rqpair->cm_id, event.event_type); 3093 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3094 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3095 } 3096 break; 3097 case IBV_EVENT_QP_REQ_ERR: 3098 case IBV_EVENT_QP_ACCESS_ERR: 3099 case IBV_EVENT_COMM_EST: 3100 case IBV_EVENT_PATH_MIG: 3101 case IBV_EVENT_PATH_MIG_ERR: 3102 SPDK_NOTICELOG("Async event: %s\n", 3103 ibv_event_type_str(event.event_type)); 3104 rqpair = event.element.qp->qp_context; 3105 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3106 (uintptr_t)rqpair->cm_id, event.event_type); 3107 nvmf_rdma_update_ibv_state(rqpair); 3108 break; 3109 case IBV_EVENT_CQ_ERR: 3110 case IBV_EVENT_DEVICE_FATAL: 3111 case IBV_EVENT_PORT_ACTIVE: 3112 case IBV_EVENT_PORT_ERR: 3113 case IBV_EVENT_LID_CHANGE: 3114 case IBV_EVENT_PKEY_CHANGE: 3115 case IBV_EVENT_SM_CHANGE: 3116 case IBV_EVENT_SRQ_ERR: 3117 case IBV_EVENT_SRQ_LIMIT_REACHED: 3118 case IBV_EVENT_CLIENT_REREGISTER: 3119 case IBV_EVENT_GID_CHANGE: 3120 default: 3121 SPDK_NOTICELOG("Async event: %s\n", 3122 ibv_event_type_str(event.event_type)); 3123 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3124 break; 3125 } 3126 ibv_ack_async_event(&event); 3127 3128 return 0; 3129 } 3130 3131 static void 3132 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3133 { 3134 int rc = 0; 3135 uint32_t i = 0; 3136 3137 for (i = 0; i < max_events; i++) { 3138 rc = nvmf_process_ib_event(device); 3139 if (rc) { 3140 break; 3141 } 3142 } 3143 3144 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Device %s: %u events processed\n", device->context->device->name, i); 3145 } 3146 3147 static uint32_t 3148 nvmf_rdma_accept(struct spdk_nvmf_transport *transport) 3149 { 3150 int nfds, i = 0; 3151 struct spdk_nvmf_rdma_transport *rtransport; 3152 struct spdk_nvmf_rdma_device *device, *tmp; 3153 uint32_t count; 3154 3155 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3156 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3157 3158 if (nfds <= 0) { 3159 return 0; 3160 } 3161 3162 /* The first poll descriptor is RDMA CM event */ 3163 if (rtransport->poll_fds[i++].revents & POLLIN) { 3164 nvmf_process_cm_event(transport); 3165 nfds--; 3166 } 3167 3168 if (nfds == 0) { 3169 return count; 3170 } 3171 3172 /* Second and subsequent poll descriptors are IB async events */ 3173 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3174 if (rtransport->poll_fds[i++].revents & POLLIN) { 3175 nvmf_process_ib_events(device, 32); 3176 nfds--; 3177 } 3178 } 3179 /* check all flagged fd's have been served */ 3180 assert(nfds == 0); 3181 3182 return count; 3183 } 3184 3185 static void 3186 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3187 struct spdk_nvmf_ctrlr_data *cdata) 3188 { 3189 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3190 3191 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3192 since in-capsule data only works with NVME drives that support SGL memory layout */ 3193 if (transport->opts.dif_insert_or_strip) { 3194 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3195 } 3196 } 3197 3198 static void 3199 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3200 struct spdk_nvme_transport_id *trid, 3201 struct spdk_nvmf_discovery_log_page_entry *entry) 3202 { 3203 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3204 entry->adrfam = trid->adrfam; 3205 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3206 3207 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3208 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3209 3210 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3211 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3212 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3213 } 3214 3215 static void 3216 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3217 3218 static struct spdk_nvmf_transport_poll_group * 3219 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3220 { 3221 struct spdk_nvmf_rdma_transport *rtransport; 3222 struct spdk_nvmf_rdma_poll_group *rgroup; 3223 struct spdk_nvmf_rdma_poller *poller; 3224 struct spdk_nvmf_rdma_device *device; 3225 struct ibv_srq_init_attr srq_init_attr; 3226 struct spdk_nvmf_rdma_resource_opts opts; 3227 int num_cqe; 3228 3229 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3230 3231 rgroup = calloc(1, sizeof(*rgroup)); 3232 if (!rgroup) { 3233 return NULL; 3234 } 3235 3236 TAILQ_INIT(&rgroup->pollers); 3237 STAILQ_INIT(&rgroup->retired_bufs); 3238 3239 pthread_mutex_lock(&rtransport->lock); 3240 TAILQ_FOREACH(device, &rtransport->devices, link) { 3241 poller = calloc(1, sizeof(*poller)); 3242 if (!poller) { 3243 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3244 nvmf_rdma_poll_group_destroy(&rgroup->group); 3245 pthread_mutex_unlock(&rtransport->lock); 3246 return NULL; 3247 } 3248 3249 poller->device = device; 3250 poller->group = rgroup; 3251 3252 TAILQ_INIT(&poller->qpairs); 3253 STAILQ_INIT(&poller->qpairs_pending_send); 3254 STAILQ_INIT(&poller->qpairs_pending_recv); 3255 3256 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3257 if (transport->opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3258 poller->max_srq_depth = transport->opts.max_srq_depth; 3259 3260 device->num_srq++; 3261 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 3262 srq_init_attr.attr.max_wr = poller->max_srq_depth; 3263 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3264 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 3265 if (!poller->srq) { 3266 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3267 nvmf_rdma_poll_group_destroy(&rgroup->group); 3268 pthread_mutex_unlock(&rtransport->lock); 3269 return NULL; 3270 } 3271 3272 opts.qp = poller->srq; 3273 opts.pd = device->pd; 3274 opts.qpair = NULL; 3275 opts.shared = true; 3276 opts.max_queue_depth = poller->max_srq_depth; 3277 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3278 3279 poller->resources = nvmf_rdma_resources_create(&opts); 3280 if (!poller->resources) { 3281 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3282 nvmf_rdma_poll_group_destroy(&rgroup->group); 3283 pthread_mutex_unlock(&rtransport->lock); 3284 return NULL; 3285 } 3286 } 3287 3288 /* 3289 * When using an srq, we can limit the completion queue at startup. 3290 * The following formula represents the calculation: 3291 * num_cqe = num_recv + num_data_wr + num_send_wr. 3292 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3293 */ 3294 if (poller->srq) { 3295 num_cqe = poller->max_srq_depth * 3; 3296 } else { 3297 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3298 } 3299 3300 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3301 if (!poller->cq) { 3302 SPDK_ERRLOG("Unable to create completion queue\n"); 3303 nvmf_rdma_poll_group_destroy(&rgroup->group); 3304 pthread_mutex_unlock(&rtransport->lock); 3305 return NULL; 3306 } 3307 poller->num_cqe = num_cqe; 3308 } 3309 3310 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3311 if (rtransport->conn_sched.next_admin_pg == NULL) { 3312 rtransport->conn_sched.next_admin_pg = rgroup; 3313 rtransport->conn_sched.next_io_pg = rgroup; 3314 } 3315 3316 pthread_mutex_unlock(&rtransport->lock); 3317 return &rgroup->group; 3318 } 3319 3320 static struct spdk_nvmf_transport_poll_group * 3321 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3322 { 3323 struct spdk_nvmf_rdma_transport *rtransport; 3324 struct spdk_nvmf_rdma_poll_group **pg; 3325 struct spdk_nvmf_transport_poll_group *result; 3326 3327 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3328 3329 pthread_mutex_lock(&rtransport->lock); 3330 3331 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3332 pthread_mutex_unlock(&rtransport->lock); 3333 return NULL; 3334 } 3335 3336 if (qpair->qid == 0) { 3337 pg = &rtransport->conn_sched.next_admin_pg; 3338 } else { 3339 pg = &rtransport->conn_sched.next_io_pg; 3340 } 3341 3342 assert(*pg != NULL); 3343 3344 result = &(*pg)->group; 3345 3346 *pg = TAILQ_NEXT(*pg, link); 3347 if (*pg == NULL) { 3348 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3349 } 3350 3351 pthread_mutex_unlock(&rtransport->lock); 3352 3353 return result; 3354 } 3355 3356 static void 3357 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3358 { 3359 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3360 struct spdk_nvmf_rdma_poller *poller, *tmp; 3361 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3362 struct spdk_nvmf_transport_pg_cache_buf *buf, *tmp_buf; 3363 struct spdk_nvmf_rdma_transport *rtransport; 3364 3365 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3366 if (!rgroup) { 3367 return; 3368 } 3369 3370 /* free all retired buffers back to the transport so we don't short the mempool. */ 3371 STAILQ_FOREACH_SAFE(buf, &rgroup->retired_bufs, link, tmp_buf) { 3372 STAILQ_REMOVE(&rgroup->retired_bufs, buf, spdk_nvmf_transport_pg_cache_buf, link); 3373 assert(group->transport != NULL); 3374 spdk_mempool_put(group->transport->data_buf_pool, buf); 3375 } 3376 3377 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3378 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3379 3380 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3381 nvmf_rdma_qpair_destroy(qpair); 3382 } 3383 3384 if (poller->srq) { 3385 if (poller->resources) { 3386 nvmf_rdma_resources_destroy(poller->resources); 3387 } 3388 ibv_destroy_srq(poller->srq); 3389 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3390 } 3391 3392 if (poller->cq) { 3393 ibv_destroy_cq(poller->cq); 3394 } 3395 3396 free(poller); 3397 } 3398 3399 if (rgroup->group.transport == NULL) { 3400 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3401 * calls this function directly in a failure path. */ 3402 free(rgroup); 3403 return; 3404 } 3405 3406 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3407 3408 pthread_mutex_lock(&rtransport->lock); 3409 next_rgroup = TAILQ_NEXT(rgroup, link); 3410 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3411 if (next_rgroup == NULL) { 3412 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3413 } 3414 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3415 rtransport->conn_sched.next_admin_pg = next_rgroup; 3416 } 3417 if (rtransport->conn_sched.next_io_pg == rgroup) { 3418 rtransport->conn_sched.next_io_pg = next_rgroup; 3419 } 3420 pthread_mutex_unlock(&rtransport->lock); 3421 3422 free(rgroup); 3423 } 3424 3425 static void 3426 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3427 { 3428 if (rqpair->cm_id != NULL) { 3429 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3430 } 3431 } 3432 3433 static int 3434 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3435 struct spdk_nvmf_qpair *qpair) 3436 { 3437 struct spdk_nvmf_rdma_poll_group *rgroup; 3438 struct spdk_nvmf_rdma_qpair *rqpair; 3439 struct spdk_nvmf_rdma_device *device; 3440 struct spdk_nvmf_rdma_poller *poller; 3441 int rc; 3442 3443 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3444 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3445 3446 device = rqpair->device; 3447 3448 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3449 if (poller->device == device) { 3450 break; 3451 } 3452 } 3453 3454 if (!poller) { 3455 SPDK_ERRLOG("No poller found for device.\n"); 3456 return -1; 3457 } 3458 3459 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3460 rqpair->poller = poller; 3461 rqpair->srq = rqpair->poller->srq; 3462 3463 rc = nvmf_rdma_qpair_initialize(qpair); 3464 if (rc < 0) { 3465 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3466 return -1; 3467 } 3468 3469 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3470 if (rc) { 3471 /* Try to reject, but we probably can't */ 3472 nvmf_rdma_qpair_reject_connection(rqpair); 3473 return -1; 3474 } 3475 3476 nvmf_rdma_update_ibv_state(rqpair); 3477 3478 return 0; 3479 } 3480 3481 static int 3482 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3483 struct spdk_nvmf_qpair *qpair) 3484 { 3485 struct spdk_nvmf_rdma_qpair *rqpair; 3486 3487 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3488 assert(group->transport->tgt != NULL); 3489 3490 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3491 3492 if (!rqpair->destruct_channel) { 3493 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3494 return 0; 3495 } 3496 3497 /* Sanity check that we get io_channel on the correct thread */ 3498 if (qpair->group) { 3499 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3500 } 3501 3502 return 0; 3503 } 3504 3505 static int 3506 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3507 { 3508 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3509 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3510 struct spdk_nvmf_rdma_transport, transport); 3511 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3512 struct spdk_nvmf_rdma_qpair, qpair); 3513 3514 /* 3515 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3516 * needs to be returned to the shared receive queue or the poll group will eventually be 3517 * starved of RECV structures. 3518 */ 3519 if (rqpair->srq && rdma_req->recv) { 3520 int rc; 3521 struct ibv_recv_wr *bad_recv_wr; 3522 3523 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 3524 if (rc) { 3525 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3526 } 3527 } 3528 3529 _nvmf_rdma_request_free(rdma_req, rtransport); 3530 return 0; 3531 } 3532 3533 static int 3534 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3535 { 3536 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3537 struct spdk_nvmf_rdma_transport, transport); 3538 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3539 struct spdk_nvmf_rdma_request, req); 3540 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3541 struct spdk_nvmf_rdma_qpair, qpair); 3542 3543 if (rqpair->ibv_state != IBV_QPS_ERR) { 3544 /* The connection is alive, so process the request as normal */ 3545 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3546 } else { 3547 /* The connection is dead. Move the request directly to the completed state. */ 3548 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3549 } 3550 3551 nvmf_rdma_request_process(rtransport, rdma_req); 3552 3553 return 0; 3554 } 3555 3556 static void 3557 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3558 { 3559 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3560 3561 rqpair->to_close = true; 3562 3563 /* This happens only when the qpair is disconnected before 3564 * it is added to the poll group. Since there is no poll group, 3565 * the RDMA qp has not been initialized yet and the RDMA CM 3566 * event has not yet been acknowledged, so we need to reject it. 3567 */ 3568 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3569 nvmf_rdma_qpair_reject_connection(rqpair); 3570 nvmf_rdma_qpair_destroy(rqpair); 3571 return; 3572 } 3573 3574 if (rqpair->rdma_qp) { 3575 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3576 } 3577 3578 nvmf_rdma_destroy_drained_qpair(rqpair); 3579 } 3580 3581 static struct spdk_nvmf_rdma_qpair * 3582 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3583 { 3584 struct spdk_nvmf_rdma_qpair *rqpair; 3585 /* @todo: improve QP search */ 3586 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3587 if (wc->qp_num == rqpair->rdma_qp->qp->qp_num) { 3588 return rqpair; 3589 } 3590 } 3591 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3592 return NULL; 3593 } 3594 3595 #ifdef DEBUG 3596 static int 3597 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3598 { 3599 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3600 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3601 } 3602 #endif 3603 3604 static void 3605 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3606 int rc) 3607 { 3608 struct spdk_nvmf_rdma_recv *rdma_recv; 3609 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3610 3611 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3612 while (bad_recv_wr != NULL) { 3613 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3614 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3615 3616 rdma_recv->qpair->current_recv_depth++; 3617 bad_recv_wr = bad_recv_wr->next; 3618 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3619 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3620 } 3621 } 3622 3623 static void 3624 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3625 { 3626 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3627 while (bad_recv_wr != NULL) { 3628 bad_recv_wr = bad_recv_wr->next; 3629 rqpair->current_recv_depth++; 3630 } 3631 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3632 } 3633 3634 static void 3635 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3636 struct spdk_nvmf_rdma_poller *rpoller) 3637 { 3638 struct spdk_nvmf_rdma_qpair *rqpair; 3639 struct ibv_recv_wr *bad_recv_wr; 3640 int rc; 3641 3642 if (rpoller->srq) { 3643 if (rpoller->resources->recvs_to_post.first != NULL) { 3644 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3645 if (rc) { 3646 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3647 } 3648 rpoller->resources->recvs_to_post.first = NULL; 3649 rpoller->resources->recvs_to_post.last = NULL; 3650 } 3651 } else { 3652 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3653 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3654 assert(rqpair->resources->recvs_to_post.first != NULL); 3655 rc = ibv_post_recv(rqpair->rdma_qp->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3656 if (rc) { 3657 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3658 } 3659 rqpair->resources->recvs_to_post.first = NULL; 3660 rqpair->resources->recvs_to_post.last = NULL; 3661 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3662 } 3663 } 3664 } 3665 3666 static void 3667 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3668 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3669 { 3670 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3671 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3672 3673 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3674 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3675 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3676 assert(rqpair->current_send_depth > 0); 3677 rqpair->current_send_depth--; 3678 switch (bad_rdma_wr->type) { 3679 case RDMA_WR_TYPE_DATA: 3680 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3681 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3682 assert(rqpair->current_read_depth > 0); 3683 rqpair->current_read_depth--; 3684 } 3685 break; 3686 case RDMA_WR_TYPE_SEND: 3687 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3688 break; 3689 default: 3690 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3691 prev_rdma_req = cur_rdma_req; 3692 continue; 3693 } 3694 3695 if (prev_rdma_req == cur_rdma_req) { 3696 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3697 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3698 continue; 3699 } 3700 3701 switch (cur_rdma_req->state) { 3702 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3703 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3704 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3705 break; 3706 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3707 case RDMA_REQUEST_STATE_COMPLETING: 3708 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3709 break; 3710 default: 3711 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3712 cur_rdma_req->state, rqpair); 3713 continue; 3714 } 3715 3716 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3717 prev_rdma_req = cur_rdma_req; 3718 } 3719 3720 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3721 /* Disconnect the connection. */ 3722 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3723 } 3724 3725 } 3726 3727 static void 3728 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3729 struct spdk_nvmf_rdma_poller *rpoller) 3730 { 3731 struct spdk_nvmf_rdma_qpair *rqpair; 3732 struct ibv_send_wr *bad_wr = NULL; 3733 int rc; 3734 3735 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3736 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3737 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3738 3739 /* bad wr always points to the first wr that failed. */ 3740 if (rc) { 3741 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3742 } 3743 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3744 } 3745 } 3746 3747 static const char * 3748 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 3749 { 3750 switch (wr_type) { 3751 case RDMA_WR_TYPE_RECV: 3752 return "RECV"; 3753 case RDMA_WR_TYPE_SEND: 3754 return "SEND"; 3755 case RDMA_WR_TYPE_DATA: 3756 return "DATA"; 3757 default: 3758 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 3759 SPDK_UNREACHABLE(); 3760 } 3761 } 3762 3763 static inline void 3764 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 3765 { 3766 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 3767 3768 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 3769 /* If qpair is in ERR state, we will receive completions for all posted and not completed 3770 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 3771 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 3772 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3773 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3774 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3775 } else { 3776 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3777 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3778 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3779 } 3780 } 3781 3782 static int 3783 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3784 struct spdk_nvmf_rdma_poller *rpoller) 3785 { 3786 struct ibv_wc wc[32]; 3787 struct spdk_nvmf_rdma_wr *rdma_wr; 3788 struct spdk_nvmf_rdma_request *rdma_req; 3789 struct spdk_nvmf_rdma_recv *rdma_recv; 3790 struct spdk_nvmf_rdma_qpair *rqpair; 3791 int reaped, i; 3792 int count = 0; 3793 bool error = false; 3794 uint64_t poll_tsc = spdk_get_ticks(); 3795 3796 /* Poll for completing operations. */ 3797 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3798 if (reaped < 0) { 3799 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3800 errno, spdk_strerror(errno)); 3801 return -1; 3802 } 3803 3804 rpoller->stat.polls++; 3805 rpoller->stat.completions += reaped; 3806 3807 for (i = 0; i < reaped; i++) { 3808 3809 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3810 3811 switch (rdma_wr->type) { 3812 case RDMA_WR_TYPE_SEND: 3813 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3814 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3815 3816 if (!wc[i].status) { 3817 count++; 3818 assert(wc[i].opcode == IBV_WC_SEND); 3819 assert(nvmf_rdma_req_is_completing(rdma_req)); 3820 } 3821 3822 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3823 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3824 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3825 rdma_req->num_outstanding_data_wr = 0; 3826 3827 nvmf_rdma_request_process(rtransport, rdma_req); 3828 break; 3829 case RDMA_WR_TYPE_RECV: 3830 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3831 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3832 if (rpoller->srq != NULL) { 3833 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3834 /* It is possible that there are still some completions for destroyed QP 3835 * associated with SRQ. We just ignore these late completions and re-post 3836 * receive WRs back to SRQ. 3837 */ 3838 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3839 struct ibv_recv_wr *bad_wr; 3840 int rc; 3841 3842 rdma_recv->wr.next = NULL; 3843 rc = ibv_post_srq_recv(rpoller->srq, 3844 &rdma_recv->wr, 3845 &bad_wr); 3846 if (rc) { 3847 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3848 } 3849 continue; 3850 } 3851 } 3852 rqpair = rdma_recv->qpair; 3853 3854 assert(rqpair != NULL); 3855 if (!wc[i].status) { 3856 assert(wc[i].opcode == IBV_WC_RECV); 3857 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3858 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3859 break; 3860 } 3861 } 3862 3863 rdma_recv->wr.next = NULL; 3864 rqpair->current_recv_depth++; 3865 rdma_recv->receive_tsc = poll_tsc; 3866 rpoller->stat.requests++; 3867 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3868 break; 3869 case RDMA_WR_TYPE_DATA: 3870 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3871 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3872 3873 assert(rdma_req->num_outstanding_data_wr > 0); 3874 3875 rqpair->current_send_depth--; 3876 rdma_req->num_outstanding_data_wr--; 3877 if (!wc[i].status) { 3878 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3879 rqpair->current_read_depth--; 3880 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3881 if (rdma_req->num_outstanding_data_wr == 0) { 3882 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3883 nvmf_rdma_request_process(rtransport, rdma_req); 3884 } 3885 } else { 3886 /* If the data transfer fails still force the queue into the error state, 3887 * if we were performing an RDMA_READ, we need to force the request into a 3888 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3889 * case, we should wait for the SEND to complete. */ 3890 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3891 rqpair->current_read_depth--; 3892 if (rdma_req->num_outstanding_data_wr == 0) { 3893 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3894 } 3895 } 3896 } 3897 break; 3898 default: 3899 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3900 continue; 3901 } 3902 3903 /* Handle error conditions */ 3904 if (wc[i].status) { 3905 nvmf_rdma_update_ibv_state(rqpair); 3906 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 3907 3908 error = true; 3909 3910 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3911 /* Disconnect the connection. */ 3912 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3913 } else { 3914 nvmf_rdma_destroy_drained_qpair(rqpair); 3915 } 3916 continue; 3917 } 3918 3919 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3920 3921 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3922 nvmf_rdma_destroy_drained_qpair(rqpair); 3923 } 3924 } 3925 3926 if (error == true) { 3927 return -1; 3928 } 3929 3930 /* submit outstanding work requests. */ 3931 _poller_submit_recvs(rtransport, rpoller); 3932 _poller_submit_sends(rtransport, rpoller); 3933 3934 return count; 3935 } 3936 3937 static int 3938 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3939 { 3940 struct spdk_nvmf_rdma_transport *rtransport; 3941 struct spdk_nvmf_rdma_poll_group *rgroup; 3942 struct spdk_nvmf_rdma_poller *rpoller; 3943 int count, rc; 3944 3945 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3946 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3947 3948 count = 0; 3949 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3950 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 3951 if (rc < 0) { 3952 return rc; 3953 } 3954 count += rc; 3955 } 3956 3957 return count; 3958 } 3959 3960 static int 3961 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3962 struct spdk_nvme_transport_id *trid, 3963 bool peer) 3964 { 3965 struct sockaddr *saddr; 3966 uint16_t port; 3967 3968 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 3969 3970 if (peer) { 3971 saddr = rdma_get_peer_addr(id); 3972 } else { 3973 saddr = rdma_get_local_addr(id); 3974 } 3975 switch (saddr->sa_family) { 3976 case AF_INET: { 3977 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3978 3979 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3980 inet_ntop(AF_INET, &saddr_in->sin_addr, 3981 trid->traddr, sizeof(trid->traddr)); 3982 if (peer) { 3983 port = ntohs(rdma_get_dst_port(id)); 3984 } else { 3985 port = ntohs(rdma_get_src_port(id)); 3986 } 3987 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3988 break; 3989 } 3990 case AF_INET6: { 3991 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3992 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3993 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3994 trid->traddr, sizeof(trid->traddr)); 3995 if (peer) { 3996 port = ntohs(rdma_get_dst_port(id)); 3997 } else { 3998 port = ntohs(rdma_get_src_port(id)); 3999 } 4000 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4001 break; 4002 } 4003 default: 4004 return -1; 4005 4006 } 4007 4008 return 0; 4009 } 4010 4011 static int 4012 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4013 struct spdk_nvme_transport_id *trid) 4014 { 4015 struct spdk_nvmf_rdma_qpair *rqpair; 4016 4017 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4018 4019 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4020 } 4021 4022 static int 4023 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4024 struct spdk_nvme_transport_id *trid) 4025 { 4026 struct spdk_nvmf_rdma_qpair *rqpair; 4027 4028 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4029 4030 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4031 } 4032 4033 static int 4034 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4035 struct spdk_nvme_transport_id *trid) 4036 { 4037 struct spdk_nvmf_rdma_qpair *rqpair; 4038 4039 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4040 4041 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4042 } 4043 4044 void 4045 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4046 { 4047 g_nvmf_hooks = *hooks; 4048 } 4049 4050 static void 4051 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4052 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4053 { 4054 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4055 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4056 4057 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4058 4059 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4060 } 4061 4062 static int 4063 _nvmf_rdma_qpair_abort_request(void *ctx) 4064 { 4065 struct spdk_nvmf_request *req = ctx; 4066 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4067 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4068 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4069 struct spdk_nvmf_rdma_qpair, qpair); 4070 int rc; 4071 4072 spdk_poller_unregister(&req->poller); 4073 4074 switch (rdma_req_to_abort->state) { 4075 case RDMA_REQUEST_STATE_EXECUTING: 4076 rc = nvmf_ctrlr_abort_request(req); 4077 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4078 return SPDK_POLLER_BUSY; 4079 } 4080 break; 4081 4082 case RDMA_REQUEST_STATE_NEED_BUFFER: 4083 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4084 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4085 4086 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4087 break; 4088 4089 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4090 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4091 spdk_nvmf_rdma_request, state_link); 4092 4093 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4094 break; 4095 4096 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4097 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4098 spdk_nvmf_rdma_request, state_link); 4099 4100 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4101 break; 4102 4103 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4104 if (spdk_get_ticks() < req->timeout_tsc) { 4105 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4106 return SPDK_POLLER_BUSY; 4107 } 4108 break; 4109 4110 default: 4111 break; 4112 } 4113 4114 spdk_nvmf_request_complete(req); 4115 return SPDK_POLLER_BUSY; 4116 } 4117 4118 static void 4119 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4120 struct spdk_nvmf_request *req) 4121 { 4122 struct spdk_nvmf_rdma_qpair *rqpair; 4123 struct spdk_nvmf_rdma_transport *rtransport; 4124 struct spdk_nvmf_transport *transport; 4125 uint16_t cid; 4126 uint32_t i; 4127 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL; 4128 4129 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4130 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4131 transport = &rtransport->transport; 4132 4133 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4134 4135 for (i = 0; i < rqpair->max_queue_depth; i++) { 4136 rdma_req_to_abort = &rqpair->resources->reqs[i]; 4137 4138 if (rdma_req_to_abort->state != RDMA_REQUEST_STATE_FREE && 4139 rdma_req_to_abort->req.cmd->nvme_cmd.cid == cid) { 4140 break; 4141 } 4142 } 4143 4144 if (rdma_req_to_abort == NULL) { 4145 spdk_nvmf_request_complete(req); 4146 return; 4147 } 4148 4149 req->req_to_abort = &rdma_req_to_abort->req; 4150 req->timeout_tsc = spdk_get_ticks() + 4151 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4152 req->poller = NULL; 4153 4154 _nvmf_rdma_qpair_abort_request(req); 4155 } 4156 4157 static int 4158 nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4159 struct spdk_nvmf_transport_poll_group_stat **stat) 4160 { 4161 struct spdk_io_channel *ch; 4162 struct spdk_nvmf_poll_group *group; 4163 struct spdk_nvmf_transport_poll_group *tgroup; 4164 struct spdk_nvmf_rdma_poll_group *rgroup; 4165 struct spdk_nvmf_rdma_poller *rpoller; 4166 struct spdk_nvmf_rdma_device_stat *device_stat; 4167 uint64_t num_devices = 0; 4168 4169 if (tgt == NULL || stat == NULL) { 4170 return -EINVAL; 4171 } 4172 4173 ch = spdk_get_io_channel(tgt); 4174 group = spdk_io_channel_get_ctx(ch);; 4175 spdk_put_io_channel(ch); 4176 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4177 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4178 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4179 if (!*stat) { 4180 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4181 return -ENOMEM; 4182 } 4183 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4184 4185 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4186 /* Count devices to allocate enough memory */ 4187 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4188 ++num_devices; 4189 } 4190 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4191 if (!(*stat)->rdma.devices) { 4192 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4193 free(*stat); 4194 return -ENOMEM; 4195 } 4196 4197 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4198 (*stat)->rdma.num_devices = num_devices; 4199 num_devices = 0; 4200 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4201 device_stat = &(*stat)->rdma.devices[num_devices++]; 4202 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4203 device_stat->polls = rpoller->stat.polls; 4204 device_stat->completions = rpoller->stat.completions; 4205 device_stat->requests = rpoller->stat.requests; 4206 device_stat->request_latency = rpoller->stat.request_latency; 4207 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4208 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4209 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4210 } 4211 return 0; 4212 } 4213 } 4214 return -ENOENT; 4215 } 4216 4217 static void 4218 nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4219 { 4220 if (stat) { 4221 free(stat->rdma.devices); 4222 } 4223 free(stat); 4224 } 4225 4226 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4227 .name = "RDMA", 4228 .type = SPDK_NVME_TRANSPORT_RDMA, 4229 .opts_init = nvmf_rdma_opts_init, 4230 .create = nvmf_rdma_create, 4231 .destroy = nvmf_rdma_destroy, 4232 4233 .listen = nvmf_rdma_listen, 4234 .stop_listen = nvmf_rdma_stop_listen, 4235 .accept = nvmf_rdma_accept, 4236 .cdata_init = nvmf_rdma_cdata_init, 4237 4238 .listener_discover = nvmf_rdma_discover, 4239 4240 .poll_group_create = nvmf_rdma_poll_group_create, 4241 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4242 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4243 .poll_group_add = nvmf_rdma_poll_group_add, 4244 .poll_group_remove = nvmf_rdma_poll_group_remove, 4245 .poll_group_poll = nvmf_rdma_poll_group_poll, 4246 4247 .req_free = nvmf_rdma_request_free, 4248 .req_complete = nvmf_rdma_request_complete, 4249 4250 .qpair_fini = nvmf_rdma_close_qpair, 4251 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4252 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4253 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4254 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4255 4256 .poll_group_get_stat = nvmf_rdma_poll_group_get_stat, 4257 .poll_group_free_stat = nvmf_rdma_poll_group_free_stat, 4258 }; 4259 4260 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4261 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 4262