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/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 rdma_transport_opts { 477 uint32_t max_srq_depth; 478 bool no_srq; 479 int acceptor_backlog; 480 }; 481 482 struct spdk_nvmf_rdma_transport { 483 struct spdk_nvmf_transport transport; 484 struct rdma_transport_opts rdma_opts; 485 486 struct spdk_nvmf_rdma_conn_sched conn_sched; 487 488 struct rdma_event_channel *event_channel; 489 490 struct spdk_mempool *data_wr_pool; 491 492 pthread_mutex_t lock; 493 494 /* fields used to poll RDMA/IB events */ 495 nfds_t npoll_fds; 496 struct pollfd *poll_fds; 497 498 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 499 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 500 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 501 }; 502 503 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 504 { 505 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 506 spdk_json_decode_uint32, true 507 }, 508 { 509 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 510 spdk_json_decode_bool, true 511 }, 512 { 513 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 514 spdk_json_decode_int32, true 515 }, 516 }; 517 518 static bool 519 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 520 struct spdk_nvmf_rdma_request *rdma_req); 521 522 static inline int 523 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 524 { 525 switch (state) { 526 case IBV_QPS_RESET: 527 case IBV_QPS_INIT: 528 case IBV_QPS_RTR: 529 case IBV_QPS_RTS: 530 case IBV_QPS_SQD: 531 case IBV_QPS_SQE: 532 case IBV_QPS_ERR: 533 return 0; 534 default: 535 return -1; 536 } 537 } 538 539 static inline enum spdk_nvme_media_error_status_code 540 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 541 enum spdk_nvme_media_error_status_code result; 542 switch (err_type) 543 { 544 case SPDK_DIF_REFTAG_ERROR: 545 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 546 break; 547 case SPDK_DIF_APPTAG_ERROR: 548 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 549 break; 550 case SPDK_DIF_GUARD_ERROR: 551 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 552 break; 553 default: 554 SPDK_UNREACHABLE(); 555 } 556 557 return result; 558 } 559 560 static enum ibv_qp_state 561 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 562 enum ibv_qp_state old_state, new_state; 563 struct ibv_qp_attr qp_attr; 564 struct ibv_qp_init_attr init_attr; 565 int rc; 566 567 old_state = rqpair->ibv_state; 568 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 569 g_spdk_nvmf_ibv_query_mask, &init_attr); 570 571 if (rc) 572 { 573 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 574 return IBV_QPS_ERR + 1; 575 } 576 577 new_state = qp_attr.qp_state; 578 rqpair->ibv_state = new_state; 579 qp_attr.ah_attr.port_num = qp_attr.port_num; 580 581 rc = nvmf_rdma_check_ibv_state(new_state); 582 if (rc) 583 { 584 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 585 /* 586 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 587 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 588 */ 589 return IBV_QPS_ERR + 1; 590 } 591 592 if (old_state != new_state) 593 { 594 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 595 (uintptr_t)rqpair->cm_id, new_state); 596 } 597 return new_state; 598 } 599 600 static void 601 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 602 struct spdk_nvmf_rdma_transport *rtransport) 603 { 604 struct spdk_nvmf_rdma_request_data *data_wr; 605 struct ibv_send_wr *next_send_wr; 606 uint64_t req_wrid; 607 608 rdma_req->num_outstanding_data_wr = 0; 609 data_wr = &rdma_req->data; 610 req_wrid = data_wr->wr.wr_id; 611 while (data_wr && data_wr->wr.wr_id == req_wrid) { 612 memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge); 613 data_wr->wr.num_sge = 0; 614 next_send_wr = data_wr->wr.next; 615 if (data_wr != &rdma_req->data) { 616 spdk_mempool_put(rtransport->data_wr_pool, data_wr); 617 } 618 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : 619 SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr); 620 } 621 } 622 623 static void 624 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 625 { 626 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 627 if (req->req.cmd) { 628 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 629 } 630 if (req->recv) { 631 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 632 } 633 } 634 635 static void 636 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 637 { 638 int i; 639 640 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 641 for (i = 0; i < rqpair->max_queue_depth; i++) { 642 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 643 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 644 } 645 } 646 } 647 648 static void 649 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 650 { 651 if (resources->cmds_mr) { 652 ibv_dereg_mr(resources->cmds_mr); 653 } 654 655 if (resources->cpls_mr) { 656 ibv_dereg_mr(resources->cpls_mr); 657 } 658 659 if (resources->bufs_mr) { 660 ibv_dereg_mr(resources->bufs_mr); 661 } 662 663 spdk_free(resources->cmds); 664 spdk_free(resources->cpls); 665 spdk_free(resources->bufs); 666 free(resources->reqs); 667 free(resources->recvs); 668 free(resources); 669 } 670 671 672 static struct spdk_nvmf_rdma_resources * 673 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 674 { 675 struct spdk_nvmf_rdma_resources *resources; 676 struct spdk_nvmf_rdma_request *rdma_req; 677 struct spdk_nvmf_rdma_recv *rdma_recv; 678 struct ibv_qp *qp; 679 struct ibv_srq *srq; 680 uint32_t i; 681 int rc; 682 683 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 684 if (!resources) { 685 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 686 return NULL; 687 } 688 689 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 690 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 691 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 692 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 693 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 694 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 695 696 if (opts->in_capsule_data_size > 0) { 697 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 698 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 699 SPDK_MALLOC_DMA); 700 } 701 702 if (!resources->reqs || !resources->recvs || !resources->cmds || 703 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 704 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 705 goto cleanup; 706 } 707 708 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 709 opts->max_queue_depth * sizeof(*resources->cmds), 710 IBV_ACCESS_LOCAL_WRITE); 711 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 712 opts->max_queue_depth * sizeof(*resources->cpls), 713 0); 714 715 if (opts->in_capsule_data_size) { 716 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 717 opts->max_queue_depth * 718 opts->in_capsule_data_size, 719 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 720 } 721 722 if (!resources->cmds_mr || !resources->cpls_mr || 723 (opts->in_capsule_data_size && 724 !resources->bufs_mr)) { 725 goto cleanup; 726 } 727 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx LKey: %x\n", 728 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 729 resources->cmds_mr->lkey); 730 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx LKey: %x\n", 731 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 732 resources->cpls_mr->lkey); 733 if (resources->bufs && resources->bufs_mr) { 734 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x LKey: %x\n", 735 resources->bufs, opts->max_queue_depth * 736 opts->in_capsule_data_size, resources->bufs_mr->lkey); 737 } 738 739 /* Initialize queues */ 740 STAILQ_INIT(&resources->incoming_queue); 741 STAILQ_INIT(&resources->free_queue); 742 743 for (i = 0; i < opts->max_queue_depth; i++) { 744 struct ibv_recv_wr *bad_wr = NULL; 745 746 rdma_recv = &resources->recvs[i]; 747 rdma_recv->qpair = opts->qpair; 748 749 /* Set up memory to receive commands */ 750 if (resources->bufs) { 751 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 752 opts->in_capsule_data_size)); 753 } 754 755 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 756 757 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 758 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 759 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 760 rdma_recv->wr.num_sge = 1; 761 762 if (rdma_recv->buf && resources->bufs_mr) { 763 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 764 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 765 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 766 rdma_recv->wr.num_sge++; 767 } 768 769 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 770 rdma_recv->wr.sg_list = rdma_recv->sgl; 771 if (opts->shared) { 772 srq = (struct ibv_srq *)opts->qp; 773 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 774 } else { 775 qp = (struct ibv_qp *)opts->qp; 776 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 777 } 778 if (rc) { 779 goto cleanup; 780 } 781 } 782 783 for (i = 0; i < opts->max_queue_depth; i++) { 784 rdma_req = &resources->reqs[i]; 785 786 if (opts->qpair != NULL) { 787 rdma_req->req.qpair = &opts->qpair->qpair; 788 } else { 789 rdma_req->req.qpair = NULL; 790 } 791 rdma_req->req.cmd = NULL; 792 793 /* Set up memory to send responses */ 794 rdma_req->req.rsp = &resources->cpls[i]; 795 796 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 797 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 798 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 799 800 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 801 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 802 rdma_req->rsp.wr.next = NULL; 803 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 804 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 805 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 806 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 807 808 /* Set up memory for data buffers */ 809 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 810 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 811 rdma_req->data.wr.next = NULL; 812 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 813 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 814 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 815 816 /* Initialize request state to FREE */ 817 rdma_req->state = RDMA_REQUEST_STATE_FREE; 818 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 819 } 820 821 return resources; 822 823 cleanup: 824 nvmf_rdma_resources_destroy(resources); 825 return NULL; 826 } 827 828 static void 829 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 830 { 831 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 832 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 833 ctx->rqpair = NULL; 834 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 835 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 836 } 837 } 838 839 static void 840 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 841 { 842 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 843 struct ibv_recv_wr *bad_recv_wr = NULL; 844 int rc; 845 846 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 847 848 if (rqpair->qd != 0) { 849 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 850 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 851 struct spdk_nvmf_rdma_transport, transport); 852 struct spdk_nvmf_rdma_request *req; 853 uint32_t i, max_req_count = 0; 854 855 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 856 857 if (rqpair->srq == NULL) { 858 nvmf_rdma_dump_qpair_contents(rqpair); 859 max_req_count = rqpair->max_queue_depth; 860 } else if (rqpair->poller && rqpair->resources) { 861 max_req_count = rqpair->poller->max_srq_depth; 862 } 863 864 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 865 for (i = 0; i < max_req_count; i++) { 866 req = &rqpair->resources->reqs[i]; 867 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 868 /* nvmf_rdma_request_process checks qpair ibv and internal state 869 * and completes a request */ 870 nvmf_rdma_request_process(rtransport, req); 871 } 872 } 873 assert(rqpair->qd == 0); 874 } 875 876 if (rqpair->poller) { 877 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 878 879 if (rqpair->srq != NULL && rqpair->resources != NULL) { 880 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 881 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 882 if (rqpair == rdma_recv->qpair) { 883 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 884 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 885 if (rc) { 886 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 887 } 888 } 889 } 890 } 891 } 892 893 if (rqpair->cm_id) { 894 if (rqpair->rdma_qp != NULL) { 895 spdk_rdma_qp_destroy(rqpair->rdma_qp); 896 rqpair->rdma_qp = NULL; 897 } 898 rdma_destroy_id(rqpair->cm_id); 899 900 if (rqpair->poller != NULL && rqpair->srq == NULL) { 901 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 902 } 903 } 904 905 if (rqpair->srq == NULL && rqpair->resources != NULL) { 906 nvmf_rdma_resources_destroy(rqpair->resources); 907 } 908 909 nvmf_rdma_qpair_clean_ibv_events(rqpair); 910 911 if (rqpair->destruct_channel) { 912 spdk_put_io_channel(rqpair->destruct_channel); 913 rqpair->destruct_channel = NULL; 914 } 915 916 free(rqpair); 917 } 918 919 static int 920 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 921 { 922 struct spdk_nvmf_rdma_poller *rpoller; 923 int rc, num_cqe, required_num_wr; 924 925 /* Enlarge CQ size dynamically */ 926 rpoller = rqpair->poller; 927 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 928 num_cqe = rpoller->num_cqe; 929 if (num_cqe < required_num_wr) { 930 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 931 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 932 } 933 934 if (rpoller->num_cqe != num_cqe) { 935 if (required_num_wr > device->attr.max_cqe) { 936 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 937 required_num_wr, device->attr.max_cqe); 938 return -1; 939 } 940 941 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 942 rc = ibv_resize_cq(rpoller->cq, num_cqe); 943 if (rc) { 944 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 945 return -1; 946 } 947 948 rpoller->num_cqe = num_cqe; 949 } 950 951 rpoller->required_num_wr = required_num_wr; 952 return 0; 953 } 954 955 static int 956 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 957 { 958 struct spdk_nvmf_rdma_qpair *rqpair; 959 struct spdk_nvmf_rdma_transport *rtransport; 960 struct spdk_nvmf_transport *transport; 961 struct spdk_nvmf_rdma_resource_opts opts; 962 struct spdk_nvmf_rdma_device *device; 963 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 964 965 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 966 device = rqpair->device; 967 968 qp_init_attr.qp_context = rqpair; 969 qp_init_attr.pd = device->pd; 970 qp_init_attr.send_cq = rqpair->poller->cq; 971 qp_init_attr.recv_cq = rqpair->poller->cq; 972 973 if (rqpair->srq) { 974 qp_init_attr.srq = rqpair->srq; 975 } else { 976 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 977 } 978 979 /* SEND, READ, and WRITE operations */ 980 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 981 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 982 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 983 984 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 985 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 986 goto error; 987 } 988 989 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 990 if (!rqpair->rdma_qp) { 991 goto error; 992 } 993 994 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 995 qp_init_attr.cap.max_send_wr); 996 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 997 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 998 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 999 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1000 1001 if (rqpair->poller->srq == NULL) { 1002 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1003 transport = &rtransport->transport; 1004 1005 opts.qp = rqpair->rdma_qp->qp; 1006 opts.pd = rqpair->cm_id->pd; 1007 opts.qpair = rqpair; 1008 opts.shared = false; 1009 opts.max_queue_depth = rqpair->max_queue_depth; 1010 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1011 1012 rqpair->resources = nvmf_rdma_resources_create(&opts); 1013 1014 if (!rqpair->resources) { 1015 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1016 rdma_destroy_qp(rqpair->cm_id); 1017 goto error; 1018 } 1019 } else { 1020 rqpair->resources = rqpair->poller->resources; 1021 } 1022 1023 rqpair->current_recv_depth = 0; 1024 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1025 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1026 1027 return 0; 1028 1029 error: 1030 rdma_destroy_id(rqpair->cm_id); 1031 rqpair->cm_id = NULL; 1032 return -1; 1033 } 1034 1035 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1036 /* This function accepts either a single wr or the first wr in a linked list. */ 1037 static void 1038 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1039 { 1040 struct ibv_recv_wr *last; 1041 1042 last = first; 1043 while (last->next != NULL) { 1044 last = last->next; 1045 } 1046 1047 if (rqpair->resources->recvs_to_post.first == NULL) { 1048 rqpair->resources->recvs_to_post.first = first; 1049 rqpair->resources->recvs_to_post.last = last; 1050 if (rqpair->srq == NULL) { 1051 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1052 } 1053 } else { 1054 rqpair->resources->recvs_to_post.last->next = first; 1055 rqpair->resources->recvs_to_post.last = last; 1056 } 1057 } 1058 1059 static int 1060 request_transfer_in(struct spdk_nvmf_request *req) 1061 { 1062 struct spdk_nvmf_rdma_request *rdma_req; 1063 struct spdk_nvmf_qpair *qpair; 1064 struct spdk_nvmf_rdma_qpair *rqpair; 1065 1066 qpair = req->qpair; 1067 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1068 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1069 1070 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1071 assert(rdma_req != NULL); 1072 1073 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1074 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1075 } 1076 1077 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1078 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1079 return 0; 1080 } 1081 1082 static int 1083 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1084 { 1085 int num_outstanding_data_wr = 0; 1086 struct spdk_nvmf_rdma_request *rdma_req; 1087 struct spdk_nvmf_qpair *qpair; 1088 struct spdk_nvmf_rdma_qpair *rqpair; 1089 struct spdk_nvme_cpl *rsp; 1090 struct ibv_send_wr *first = NULL; 1091 1092 *data_posted = 0; 1093 qpair = req->qpair; 1094 rsp = &req->rsp->nvme_cpl; 1095 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1096 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1097 1098 /* Advance our sq_head pointer */ 1099 if (qpair->sq_head == qpair->sq_head_max) { 1100 qpair->sq_head = 0; 1101 } else { 1102 qpair->sq_head++; 1103 } 1104 rsp->sqhd = qpair->sq_head; 1105 1106 /* queue the capsule for the recv buffer */ 1107 assert(rdma_req->recv != NULL); 1108 1109 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1110 1111 rdma_req->recv = NULL; 1112 assert(rqpair->current_recv_depth > 0); 1113 rqpair->current_recv_depth--; 1114 1115 /* Build the response which consists of optional 1116 * RDMA WRITEs to transfer data, plus an RDMA SEND 1117 * containing the response. 1118 */ 1119 first = &rdma_req->rsp.wr; 1120 1121 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1122 /* On failure, data was not read from the controller. So clear the 1123 * number of outstanding data WRs to zero. 1124 */ 1125 rdma_req->num_outstanding_data_wr = 0; 1126 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1127 first = &rdma_req->data.wr; 1128 *data_posted = 1; 1129 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1130 } 1131 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1132 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1133 } 1134 1135 /* +1 for the rsp wr */ 1136 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1137 1138 return 0; 1139 } 1140 1141 static int 1142 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1143 { 1144 struct spdk_nvmf_rdma_accept_private_data accept_data; 1145 struct rdma_conn_param ctrlr_event_data = {}; 1146 int rc; 1147 1148 accept_data.recfmt = 0; 1149 accept_data.crqsize = rqpair->max_queue_depth; 1150 1151 ctrlr_event_data.private_data = &accept_data; 1152 ctrlr_event_data.private_data_len = sizeof(accept_data); 1153 if (id->ps == RDMA_PS_TCP) { 1154 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1155 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1156 } 1157 1158 /* Configure infinite retries for the initiator side qpair. 1159 * When using a shared receive queue on the target side, 1160 * we need to pass this value to the initiator to prevent the 1161 * initiator side NIC from completing SEND requests back to the 1162 * initiator with status rnr_retry_count_exceeded. */ 1163 if (rqpair->srq != NULL) { 1164 ctrlr_event_data.rnr_retry_count = 0x7; 1165 } 1166 1167 /* When qpair is created without use of rdma cm API, an additional 1168 * information must be provided to initiator in the connection response: 1169 * whether qpair is using SRQ and its qp_num 1170 * Fields below are ignored by rdma cm if qpair has been 1171 * created using rdma cm API. */ 1172 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1173 ctrlr_event_data.qp_num = rqpair->rdma_qp->qp->qp_num; 1174 1175 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1176 if (rc) { 1177 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1178 } else { 1179 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1180 } 1181 1182 return rc; 1183 } 1184 1185 static void 1186 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1187 { 1188 struct spdk_nvmf_rdma_reject_private_data rej_data; 1189 1190 rej_data.recfmt = 0; 1191 rej_data.sts = error; 1192 1193 rdma_reject(id, &rej_data, sizeof(rej_data)); 1194 } 1195 1196 static int 1197 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1198 { 1199 struct spdk_nvmf_rdma_transport *rtransport; 1200 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1201 struct spdk_nvmf_rdma_port *port; 1202 struct rdma_conn_param *rdma_param = NULL; 1203 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1204 uint16_t max_queue_depth; 1205 uint16_t max_read_depth; 1206 1207 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1208 1209 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1210 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1211 1212 rdma_param = &event->param.conn; 1213 if (rdma_param->private_data == NULL || 1214 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1215 SPDK_ERRLOG("connect request: no private data provided\n"); 1216 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1217 return -1; 1218 } 1219 1220 private_data = rdma_param->private_data; 1221 if (private_data->recfmt != 0) { 1222 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1223 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1224 return -1; 1225 } 1226 1227 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1228 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1229 1230 port = event->listen_id->context; 1231 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1232 event->listen_id, event->listen_id->verbs, port); 1233 1234 /* Figure out the supported queue depth. This is a multi-step process 1235 * that takes into account hardware maximums, host provided values, 1236 * and our target's internal memory limits */ 1237 1238 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1239 1240 /* Start with the maximum queue depth allowed by the target */ 1241 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1242 max_read_depth = rtransport->transport.opts.max_queue_depth; 1243 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1244 rtransport->transport.opts.max_queue_depth); 1245 1246 /* Next check the local NIC's hardware limitations */ 1247 SPDK_DEBUGLOG(rdma, 1248 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1249 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1250 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1251 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1252 1253 /* Next check the remote NIC's hardware limitations */ 1254 SPDK_DEBUGLOG(rdma, 1255 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1256 rdma_param->initiator_depth, rdma_param->responder_resources); 1257 if (rdma_param->initiator_depth > 0) { 1258 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1259 } 1260 1261 /* Finally check for the host software requested values, which are 1262 * optional. */ 1263 if (rdma_param->private_data != NULL && 1264 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1265 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1266 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1267 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1268 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1269 } 1270 1271 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1272 max_queue_depth, max_read_depth); 1273 1274 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1275 if (rqpair == NULL) { 1276 SPDK_ERRLOG("Could not allocate new connection.\n"); 1277 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1278 return -1; 1279 } 1280 1281 rqpair->device = port->device; 1282 rqpair->max_queue_depth = max_queue_depth; 1283 rqpair->max_read_depth = max_read_depth; 1284 rqpair->cm_id = event->id; 1285 rqpair->listen_id = event->listen_id; 1286 rqpair->qpair.transport = transport; 1287 rqpair->qpair.trid = port->trid; 1288 STAILQ_INIT(&rqpair->ibv_events); 1289 /* use qid from the private data to determine the qpair type 1290 qid will be set to the appropriate value when the controller is created */ 1291 rqpair->qpair.qid = private_data->qid; 1292 1293 event->id->context = &rqpair->qpair; 1294 1295 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1296 1297 return 0; 1298 } 1299 1300 static int 1301 nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1302 enum spdk_mem_map_notify_action action, 1303 void *vaddr, size_t size) 1304 { 1305 struct ibv_pd *pd = cb_ctx; 1306 struct ibv_mr *mr; 1307 int rc; 1308 1309 switch (action) { 1310 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1311 if (!g_nvmf_hooks.get_rkey) { 1312 mr = ibv_reg_mr(pd, vaddr, size, 1313 IBV_ACCESS_LOCAL_WRITE | 1314 IBV_ACCESS_REMOTE_READ | 1315 IBV_ACCESS_REMOTE_WRITE); 1316 if (mr == NULL) { 1317 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1318 return -1; 1319 } else { 1320 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1321 } 1322 } else { 1323 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1324 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1325 } 1326 break; 1327 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1328 if (!g_nvmf_hooks.get_rkey) { 1329 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1330 if (mr) { 1331 ibv_dereg_mr(mr); 1332 } 1333 } 1334 rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1335 break; 1336 default: 1337 SPDK_UNREACHABLE(); 1338 } 1339 1340 return rc; 1341 } 1342 1343 static int 1344 nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1345 { 1346 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1347 return addr_1 == addr_2; 1348 } 1349 1350 static inline void 1351 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1352 enum spdk_nvme_data_transfer xfer) 1353 { 1354 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1355 wr->opcode = IBV_WR_RDMA_WRITE; 1356 wr->send_flags = 0; 1357 wr->next = next; 1358 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1359 wr->opcode = IBV_WR_RDMA_READ; 1360 wr->send_flags = IBV_SEND_SIGNALED; 1361 wr->next = NULL; 1362 } else { 1363 assert(0); 1364 } 1365 } 1366 1367 static int 1368 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1369 struct spdk_nvmf_rdma_request *rdma_req, 1370 uint32_t num_sgl_descriptors) 1371 { 1372 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1373 struct spdk_nvmf_rdma_request_data *current_data_wr; 1374 uint32_t i; 1375 1376 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1377 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1378 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1379 return -EINVAL; 1380 } 1381 1382 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1383 return -ENOMEM; 1384 } 1385 1386 current_data_wr = &rdma_req->data; 1387 1388 for (i = 0; i < num_sgl_descriptors; i++) { 1389 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1390 current_data_wr->wr.next = &work_requests[i]->wr; 1391 current_data_wr = work_requests[i]; 1392 current_data_wr->wr.sg_list = current_data_wr->sgl; 1393 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1394 } 1395 1396 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1397 1398 return 0; 1399 } 1400 1401 static inline void 1402 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1403 { 1404 struct ibv_send_wr *wr = &rdma_req->data.wr; 1405 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1406 1407 wr->wr.rdma.rkey = sgl->keyed.key; 1408 wr->wr.rdma.remote_addr = sgl->address; 1409 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1410 } 1411 1412 static inline void 1413 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1414 { 1415 struct ibv_send_wr *wr = &rdma_req->data.wr; 1416 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1417 uint32_t i; 1418 int j; 1419 uint64_t remote_addr_offset = 0; 1420 1421 for (i = 0; i < num_wrs; ++i) { 1422 wr->wr.rdma.rkey = sgl->keyed.key; 1423 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1424 for (j = 0; j < wr->num_sge; ++j) { 1425 remote_addr_offset += wr->sg_list[j].length; 1426 } 1427 wr = wr->next; 1428 } 1429 } 1430 1431 /* This function is used in the rare case that we have a buffer split over multiple memory regions. */ 1432 static int 1433 nvmf_rdma_replace_buffer(struct spdk_nvmf_rdma_poll_group *rgroup, void **buf) 1434 { 1435 struct spdk_nvmf_transport_poll_group *group = &rgroup->group; 1436 struct spdk_nvmf_transport *transport = group->transport; 1437 struct spdk_nvmf_transport_pg_cache_buf *old_buf; 1438 void *new_buf; 1439 1440 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1441 group->buf_cache_count--; 1442 new_buf = STAILQ_FIRST(&group->buf_cache); 1443 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1444 assert(*buf != NULL); 1445 } else { 1446 new_buf = spdk_mempool_get(transport->data_buf_pool); 1447 } 1448 1449 if (*buf == NULL) { 1450 return -ENOMEM; 1451 } 1452 1453 old_buf = *buf; 1454 STAILQ_INSERT_HEAD(&rgroup->retired_bufs, old_buf, link); 1455 *buf = new_buf; 1456 return 0; 1457 } 1458 1459 static bool 1460 nvmf_rdma_get_lkey(struct spdk_nvmf_rdma_device *device, struct iovec *iov, 1461 uint32_t *_lkey) 1462 { 1463 uint64_t translation_len; 1464 uint32_t lkey; 1465 1466 translation_len = iov->iov_len; 1467 1468 if (!g_nvmf_hooks.get_rkey) { 1469 lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1470 (uint64_t)iov->iov_base, &translation_len))->lkey; 1471 } else { 1472 lkey = spdk_mem_map_translate(device->map, 1473 (uint64_t)iov->iov_base, &translation_len); 1474 } 1475 1476 if (spdk_unlikely(translation_len < iov->iov_len)) { 1477 return false; 1478 } 1479 1480 *_lkey = lkey; 1481 return true; 1482 } 1483 1484 static bool 1485 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device, 1486 struct iovec *iov, struct ibv_send_wr **_wr, 1487 uint32_t *_remaining_data_block, uint32_t *_offset, 1488 uint32_t *_num_extra_wrs, 1489 const struct spdk_dif_ctx *dif_ctx) 1490 { 1491 struct ibv_send_wr *wr = *_wr; 1492 struct ibv_sge *sg_ele = &wr->sg_list[wr->num_sge]; 1493 uint32_t lkey = 0; 1494 uint32_t remaining, data_block_size, md_size, sge_len; 1495 1496 if (spdk_unlikely(!nvmf_rdma_get_lkey(device, iov, &lkey))) { 1497 /* This is a very rare case that can occur when using DPDK version < 19.05 */ 1498 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions. Removing it from circulation.\n"); 1499 return false; 1500 } 1501 1502 if (spdk_likely(!dif_ctx)) { 1503 sg_ele->lkey = lkey; 1504 sg_ele->addr = (uintptr_t)(iov->iov_base); 1505 sg_ele->length = iov->iov_len; 1506 wr->num_sge++; 1507 } else { 1508 remaining = iov->iov_len - *_offset; 1509 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1510 md_size = dif_ctx->md_size; 1511 1512 while (remaining) { 1513 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1514 if (*_num_extra_wrs > 0 && wr->next) { 1515 *_wr = wr->next; 1516 wr = *_wr; 1517 wr->num_sge = 0; 1518 sg_ele = &wr->sg_list[wr->num_sge]; 1519 (*_num_extra_wrs)--; 1520 } else { 1521 break; 1522 } 1523 } 1524 sg_ele->lkey = lkey; 1525 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset); 1526 sge_len = spdk_min(remaining, *_remaining_data_block); 1527 sg_ele->length = sge_len; 1528 remaining -= sge_len; 1529 *_remaining_data_block -= sge_len; 1530 *_offset += sge_len; 1531 1532 sg_ele++; 1533 wr->num_sge++; 1534 1535 if (*_remaining_data_block == 0) { 1536 /* skip metadata */ 1537 *_offset += md_size; 1538 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1539 remaining -= spdk_min(remaining, md_size); 1540 *_remaining_data_block = data_block_size; 1541 } 1542 1543 if (remaining == 0) { 1544 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1545 the remaining metadata bits at the beginning of the next buffer */ 1546 *_offset -= iov->iov_len; 1547 } 1548 } 1549 } 1550 1551 return true; 1552 } 1553 1554 static int 1555 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1556 struct spdk_nvmf_rdma_device *device, 1557 struct spdk_nvmf_rdma_request *rdma_req, 1558 struct ibv_send_wr *wr, 1559 uint32_t length, 1560 uint32_t num_extra_wrs) 1561 { 1562 struct spdk_nvmf_request *req = &rdma_req->req; 1563 struct spdk_dif_ctx *dif_ctx = NULL; 1564 uint32_t remaining_data_block = 0; 1565 uint32_t offset = 0; 1566 1567 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 1568 dif_ctx = &rdma_req->req.dif.dif_ctx; 1569 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1570 } 1571 1572 wr->num_sge = 0; 1573 1574 while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1575 while (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr, 1576 &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) { 1577 if (nvmf_rdma_replace_buffer(rgroup, &req->buffers[rdma_req->iovpos]) == -ENOMEM) { 1578 return -ENOMEM; 1579 } 1580 req->iov[rdma_req->iovpos].iov_base = (void *)((uintptr_t)(req->buffers[rdma_req->iovpos] + 1581 NVMF_DATA_BUFFER_MASK) & 1582 ~NVMF_DATA_BUFFER_MASK); 1583 } 1584 1585 length -= req->iov[rdma_req->iovpos].iov_len; 1586 rdma_req->iovpos++; 1587 } 1588 1589 if (length) { 1590 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1591 return -EINVAL; 1592 } 1593 1594 return 0; 1595 } 1596 1597 static inline uint32_t 1598 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1599 { 1600 /* estimate the number of SG entries and WRs needed to process the request */ 1601 uint32_t num_sge = 0; 1602 uint32_t i; 1603 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1604 1605 for (i = 0; i < num_buffers && length > 0; i++) { 1606 uint32_t buffer_len = spdk_min(length, io_unit_size); 1607 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1608 1609 if (num_sge_in_block * block_size > buffer_len) { 1610 ++num_sge_in_block; 1611 } 1612 num_sge += num_sge_in_block; 1613 length -= buffer_len; 1614 } 1615 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1616 } 1617 1618 static int 1619 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1620 struct spdk_nvmf_rdma_device *device, 1621 struct spdk_nvmf_rdma_request *rdma_req, 1622 uint32_t length) 1623 { 1624 struct spdk_nvmf_rdma_qpair *rqpair; 1625 struct spdk_nvmf_rdma_poll_group *rgroup; 1626 struct spdk_nvmf_request *req = &rdma_req->req; 1627 struct ibv_send_wr *wr = &rdma_req->data.wr; 1628 int rc; 1629 uint32_t num_wrs = 1; 1630 1631 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1632 rgroup = rqpair->poller->group; 1633 1634 /* rdma wr specifics */ 1635 nvmf_rdma_setup_request(rdma_req); 1636 1637 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1638 length); 1639 if (rc != 0) { 1640 return rc; 1641 } 1642 1643 assert(req->iovcnt <= rqpair->max_send_sge); 1644 1645 rdma_req->iovpos = 0; 1646 1647 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1648 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1649 req->dif.dif_ctx.block_size); 1650 if (num_wrs > 1) { 1651 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1652 if (rc != 0) { 1653 goto err_exit; 1654 } 1655 } 1656 } 1657 1658 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1659 if (spdk_unlikely(rc != 0)) { 1660 goto err_exit; 1661 } 1662 1663 if (spdk_unlikely(num_wrs > 1)) { 1664 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1665 } 1666 1667 /* set the number of outstanding data WRs for this request. */ 1668 rdma_req->num_outstanding_data_wr = num_wrs; 1669 1670 return rc; 1671 1672 err_exit: 1673 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1674 nvmf_rdma_request_free_data(rdma_req, rtransport); 1675 req->iovcnt = 0; 1676 return rc; 1677 } 1678 1679 static int 1680 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1681 struct spdk_nvmf_rdma_device *device, 1682 struct spdk_nvmf_rdma_request *rdma_req) 1683 { 1684 struct spdk_nvmf_rdma_qpair *rqpair; 1685 struct spdk_nvmf_rdma_poll_group *rgroup; 1686 struct ibv_send_wr *current_wr; 1687 struct spdk_nvmf_request *req = &rdma_req->req; 1688 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1689 uint32_t num_sgl_descriptors; 1690 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES]; 1691 uint32_t i; 1692 int rc; 1693 1694 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1695 rgroup = rqpair->poller->group; 1696 1697 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1698 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1699 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1700 1701 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1702 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1703 1704 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1705 return -ENOMEM; 1706 } 1707 1708 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1709 for (i = 0; i < num_sgl_descriptors; i++) { 1710 if (spdk_likely(!req->dif.dif_insert_or_strip)) { 1711 lengths[i] = desc->keyed.length; 1712 } else { 1713 req->dif.orig_length += desc->keyed.length; 1714 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1715 req->dif.elba_length += lengths[i]; 1716 } 1717 desc++; 1718 } 1719 1720 rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport, 1721 lengths, num_sgl_descriptors); 1722 if (rc != 0) { 1723 nvmf_rdma_request_free_data(rdma_req, rtransport); 1724 return rc; 1725 } 1726 1727 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1728 current_wr = &rdma_req->data.wr; 1729 assert(current_wr != NULL); 1730 1731 req->length = 0; 1732 rdma_req->iovpos = 0; 1733 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1734 for (i = 0; i < num_sgl_descriptors; i++) { 1735 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1736 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1737 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1738 rc = -EINVAL; 1739 goto err_exit; 1740 } 1741 1742 current_wr->num_sge = 0; 1743 1744 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1745 if (rc != 0) { 1746 rc = -ENOMEM; 1747 goto err_exit; 1748 } 1749 1750 req->length += desc->keyed.length; 1751 current_wr->wr.rdma.rkey = desc->keyed.key; 1752 current_wr->wr.rdma.remote_addr = desc->address; 1753 current_wr = current_wr->next; 1754 desc++; 1755 } 1756 1757 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1758 /* Go back to the last descriptor in the list. */ 1759 desc--; 1760 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1761 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1762 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1763 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1764 } 1765 } 1766 #endif 1767 1768 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1769 1770 return 0; 1771 1772 err_exit: 1773 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1774 nvmf_rdma_request_free_data(rdma_req, rtransport); 1775 return rc; 1776 } 1777 1778 static int 1779 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1780 struct spdk_nvmf_rdma_device *device, 1781 struct spdk_nvmf_rdma_request *rdma_req) 1782 { 1783 struct spdk_nvmf_request *req = &rdma_req->req; 1784 struct spdk_nvme_cpl *rsp; 1785 struct spdk_nvme_sgl_descriptor *sgl; 1786 int rc; 1787 uint32_t length; 1788 1789 rsp = &req->rsp->nvme_cpl; 1790 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1791 1792 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1793 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1794 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1795 1796 length = sgl->keyed.length; 1797 if (length > rtransport->transport.opts.max_io_size) { 1798 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1799 length, rtransport->transport.opts.max_io_size); 1800 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1801 return -1; 1802 } 1803 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1804 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1805 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1806 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1807 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1808 } 1809 } 1810 #endif 1811 1812 /* fill request length and populate iovs */ 1813 req->length = length; 1814 1815 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1816 req->dif.orig_length = length; 1817 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1818 req->dif.elba_length = length; 1819 } 1820 1821 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1822 if (spdk_unlikely(rc < 0)) { 1823 if (rc == -EINVAL) { 1824 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1825 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1826 return -1; 1827 } 1828 /* No available buffers. Queue this request up. */ 1829 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1830 return 0; 1831 } 1832 1833 /* backward compatible */ 1834 req->data = req->iov[0].iov_base; 1835 1836 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1837 req->iovcnt); 1838 1839 return 0; 1840 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1841 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1842 uint64_t offset = sgl->address; 1843 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1844 1845 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1846 offset, sgl->unkeyed.length); 1847 1848 if (offset > max_len) { 1849 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1850 offset, max_len); 1851 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1852 return -1; 1853 } 1854 max_len -= (uint32_t)offset; 1855 1856 if (sgl->unkeyed.length > max_len) { 1857 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1858 sgl->unkeyed.length, max_len); 1859 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1860 return -1; 1861 } 1862 1863 rdma_req->num_outstanding_data_wr = 0; 1864 req->data = rdma_req->recv->buf + offset; 1865 req->data_from_pool = false; 1866 req->length = sgl->unkeyed.length; 1867 1868 req->iov[0].iov_base = req->data; 1869 req->iov[0].iov_len = req->length; 1870 req->iovcnt = 1; 1871 1872 return 0; 1873 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1874 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1875 1876 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1877 if (rc == -ENOMEM) { 1878 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1879 return 0; 1880 } else if (rc == -EINVAL) { 1881 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1882 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1883 return -1; 1884 } 1885 1886 /* backward compatible */ 1887 req->data = req->iov[0].iov_base; 1888 1889 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1890 req->iovcnt); 1891 1892 return 0; 1893 } 1894 1895 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1896 sgl->generic.type, sgl->generic.subtype); 1897 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1898 return -1; 1899 } 1900 1901 static void 1902 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1903 struct spdk_nvmf_rdma_transport *rtransport) 1904 { 1905 struct spdk_nvmf_rdma_qpair *rqpair; 1906 struct spdk_nvmf_rdma_poll_group *rgroup; 1907 1908 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1909 if (rdma_req->req.data_from_pool) { 1910 rgroup = rqpair->poller->group; 1911 1912 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1913 } 1914 nvmf_rdma_request_free_data(rdma_req, rtransport); 1915 rdma_req->req.length = 0; 1916 rdma_req->req.iovcnt = 0; 1917 rdma_req->req.data = NULL; 1918 rdma_req->rsp.wr.next = NULL; 1919 rdma_req->data.wr.next = NULL; 1920 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1921 rqpair->qd--; 1922 1923 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1924 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1925 } 1926 1927 bool 1928 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1929 struct spdk_nvmf_rdma_request *rdma_req) 1930 { 1931 struct spdk_nvmf_rdma_qpair *rqpair; 1932 struct spdk_nvmf_rdma_device *device; 1933 struct spdk_nvmf_rdma_poll_group *rgroup; 1934 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1935 int rc; 1936 struct spdk_nvmf_rdma_recv *rdma_recv; 1937 enum spdk_nvmf_rdma_request_state prev_state; 1938 bool progress = false; 1939 int data_posted; 1940 uint32_t num_blocks; 1941 1942 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1943 device = rqpair->device; 1944 rgroup = rqpair->poller->group; 1945 1946 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1947 1948 /* If the queue pair is in an error state, force the request to the completed state 1949 * to release resources. */ 1950 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1951 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1952 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1953 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1954 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1955 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1956 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1957 } 1958 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1959 } 1960 1961 /* The loop here is to allow for several back-to-back state changes. */ 1962 do { 1963 prev_state = rdma_req->state; 1964 1965 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 1966 1967 switch (rdma_req->state) { 1968 case RDMA_REQUEST_STATE_FREE: 1969 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1970 * to escape this state. */ 1971 break; 1972 case RDMA_REQUEST_STATE_NEW: 1973 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1974 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1975 rdma_recv = rdma_req->recv; 1976 1977 /* The first element of the SGL is the NVMe command */ 1978 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1979 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1980 1981 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1982 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1983 break; 1984 } 1985 1986 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 1987 rdma_req->req.dif.dif_insert_or_strip = true; 1988 } 1989 1990 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1991 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1992 rdma_req->rsp.wr.imm_data = 0; 1993 #endif 1994 1995 /* The next state transition depends on the data transfer needs of this request. */ 1996 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 1997 1998 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 1999 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2000 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2001 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2002 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 2003 break; 2004 } 2005 2006 /* If no data to transfer, ready to execute. */ 2007 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2008 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2009 break; 2010 } 2011 2012 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2013 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2014 break; 2015 case RDMA_REQUEST_STATE_NEED_BUFFER: 2016 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2017 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2018 2019 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2020 2021 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2022 /* This request needs to wait in line to obtain a buffer */ 2023 break; 2024 } 2025 2026 /* Try to get a data buffer */ 2027 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2028 if (rc < 0) { 2029 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2030 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2031 break; 2032 } 2033 2034 if (!rdma_req->req.data) { 2035 /* No buffers available. */ 2036 rgroup->stat.pending_data_buffer++; 2037 break; 2038 } 2039 2040 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2041 2042 /* If data is transferring from host to controller and the data didn't 2043 * arrive using in capsule data, we need to do a transfer from the host. 2044 */ 2045 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2046 rdma_req->req.data_from_pool) { 2047 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2048 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2049 break; 2050 } 2051 2052 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2053 break; 2054 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2055 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2056 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2057 2058 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2059 /* This request needs to wait in line to perform RDMA */ 2060 break; 2061 } 2062 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2063 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2064 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2065 rqpair->poller->stat.pending_rdma_read++; 2066 break; 2067 } 2068 2069 /* We have already verified that this request is the head of the queue. */ 2070 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2071 2072 rc = request_transfer_in(&rdma_req->req); 2073 if (!rc) { 2074 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2075 } else { 2076 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2077 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2078 } 2079 break; 2080 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2081 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2082 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2083 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2084 * to escape this state. */ 2085 break; 2086 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2087 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2088 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2089 2090 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2091 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2092 /* generate DIF for write operation */ 2093 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2094 assert(num_blocks > 0); 2095 2096 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2097 num_blocks, &rdma_req->req.dif.dif_ctx); 2098 if (rc != 0) { 2099 SPDK_ERRLOG("DIF generation failed\n"); 2100 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2101 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2102 break; 2103 } 2104 } 2105 2106 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2107 /* set extended length before IO operation */ 2108 rdma_req->req.length = rdma_req->req.dif.elba_length; 2109 } 2110 2111 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2112 spdk_nvmf_request_exec(&rdma_req->req); 2113 break; 2114 case RDMA_REQUEST_STATE_EXECUTING: 2115 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2116 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2117 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2118 * to escape this state. */ 2119 break; 2120 case RDMA_REQUEST_STATE_EXECUTED: 2121 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2122 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2123 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2124 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2125 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2126 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2127 } else { 2128 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2129 } 2130 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2131 /* restore the original length */ 2132 rdma_req->req.length = rdma_req->req.dif.orig_length; 2133 2134 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2135 struct spdk_dif_error error_blk; 2136 2137 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2138 2139 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2140 &rdma_req->req.dif.dif_ctx, &error_blk); 2141 if (rc) { 2142 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2143 2144 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2145 error_blk.err_offset); 2146 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2147 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2148 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2149 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2150 } 2151 } 2152 } 2153 break; 2154 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2155 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2156 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2157 2158 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2159 /* This request needs to wait in line to perform RDMA */ 2160 break; 2161 } 2162 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2163 rqpair->max_send_depth) { 2164 /* We can only have so many WRs outstanding. we have to wait until some finish. 2165 * +1 since each request has an additional wr in the resp. */ 2166 rqpair->poller->stat.pending_rdma_write++; 2167 break; 2168 } 2169 2170 /* We have already verified that this request is the head of the queue. */ 2171 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2172 2173 /* The data transfer will be kicked off from 2174 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2175 */ 2176 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2177 break; 2178 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2179 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2180 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2181 rc = request_transfer_out(&rdma_req->req, &data_posted); 2182 assert(rc == 0); /* No good way to handle this currently */ 2183 if (rc) { 2184 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2185 } else { 2186 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2187 RDMA_REQUEST_STATE_COMPLETING; 2188 } 2189 break; 2190 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2191 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2192 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2193 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2194 * to escape this state. */ 2195 break; 2196 case RDMA_REQUEST_STATE_COMPLETING: 2197 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2198 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2199 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2200 * to escape this state. */ 2201 break; 2202 case RDMA_REQUEST_STATE_COMPLETED: 2203 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2204 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2205 2206 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2207 _nvmf_rdma_request_free(rdma_req, rtransport); 2208 break; 2209 case RDMA_REQUEST_NUM_STATES: 2210 default: 2211 assert(0); 2212 break; 2213 } 2214 2215 if (rdma_req->state != prev_state) { 2216 progress = true; 2217 } 2218 } while (rdma_req->state != prev_state); 2219 2220 return progress; 2221 } 2222 2223 /* Public API callbacks begin here */ 2224 2225 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2226 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2227 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2228 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2229 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2230 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2231 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2232 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2233 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2234 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2235 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2236 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2237 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2238 2239 static void 2240 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2241 { 2242 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2243 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2244 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2245 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2246 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2247 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2248 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2249 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2250 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2251 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2252 opts->transport_specific = NULL; 2253 } 2254 2255 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2256 .notify_cb = nvmf_rdma_mem_notify, 2257 .are_contiguous = nvmf_rdma_check_contiguous_entries 2258 }; 2259 2260 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2261 2262 static struct spdk_nvmf_transport * 2263 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2264 { 2265 int rc; 2266 struct spdk_nvmf_rdma_transport *rtransport; 2267 struct spdk_nvmf_rdma_device *device, *tmp; 2268 struct ibv_context **contexts; 2269 uint32_t i; 2270 int flag; 2271 uint32_t sge_count; 2272 uint32_t min_shared_buffers; 2273 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2274 pthread_mutexattr_t attr; 2275 2276 rtransport = calloc(1, sizeof(*rtransport)); 2277 if (!rtransport) { 2278 return NULL; 2279 } 2280 2281 if (pthread_mutexattr_init(&attr)) { 2282 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2283 free(rtransport); 2284 return NULL; 2285 } 2286 2287 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2288 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2289 pthread_mutexattr_destroy(&attr); 2290 free(rtransport); 2291 return NULL; 2292 } 2293 2294 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2295 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2296 pthread_mutexattr_destroy(&attr); 2297 free(rtransport); 2298 return NULL; 2299 } 2300 2301 pthread_mutexattr_destroy(&attr); 2302 2303 TAILQ_INIT(&rtransport->devices); 2304 TAILQ_INIT(&rtransport->ports); 2305 TAILQ_INIT(&rtransport->poll_groups); 2306 2307 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2308 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2309 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2310 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2311 if (opts->transport_specific != NULL && 2312 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2313 SPDK_COUNTOF(rdma_transport_opts_decoder), 2314 &rtransport->rdma_opts)) { 2315 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2316 nvmf_rdma_destroy(&rtransport->transport); 2317 return NULL; 2318 } 2319 2320 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2321 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2322 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2323 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2324 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d," 2325 " acceptor_backlog=%d, abort_timeout_sec=%d\n", 2326 opts->max_queue_depth, 2327 opts->max_io_size, 2328 opts->max_qpairs_per_ctrlr - 1, 2329 opts->io_unit_size, 2330 opts->in_capsule_data_size, 2331 opts->max_aq_depth, 2332 opts->num_shared_buffers, 2333 rtransport->rdma_opts.max_srq_depth, 2334 rtransport->rdma_opts.no_srq, 2335 rtransport->rdma_opts.acceptor_backlog, 2336 opts->abort_timeout_sec); 2337 2338 /* I/O unit size cannot be larger than max I/O size */ 2339 if (opts->io_unit_size > opts->max_io_size) { 2340 opts->io_unit_size = opts->max_io_size; 2341 } 2342 2343 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2344 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2345 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2346 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2347 } 2348 2349 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2350 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2351 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2352 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2353 nvmf_rdma_destroy(&rtransport->transport); 2354 return NULL; 2355 } 2356 2357 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2358 if (min_shared_buffers > opts->num_shared_buffers) { 2359 SPDK_ERRLOG("There are not enough buffers to satisfy" 2360 "per-poll group caches for each thread. (%" PRIu32 ")" 2361 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2362 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2363 nvmf_rdma_destroy(&rtransport->transport); 2364 return NULL; 2365 } 2366 2367 sge_count = opts->max_io_size / opts->io_unit_size; 2368 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2369 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2370 nvmf_rdma_destroy(&rtransport->transport); 2371 return NULL; 2372 } 2373 2374 rtransport->event_channel = rdma_create_event_channel(); 2375 if (rtransport->event_channel == NULL) { 2376 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2377 nvmf_rdma_destroy(&rtransport->transport); 2378 return NULL; 2379 } 2380 2381 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2382 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2383 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2384 rtransport->event_channel->fd, spdk_strerror(errno)); 2385 nvmf_rdma_destroy(&rtransport->transport); 2386 return NULL; 2387 } 2388 2389 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2390 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2391 sizeof(struct spdk_nvmf_rdma_request_data), 2392 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2393 SPDK_ENV_SOCKET_ID_ANY); 2394 if (!rtransport->data_wr_pool) { 2395 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2396 nvmf_rdma_destroy(&rtransport->transport); 2397 return NULL; 2398 } 2399 2400 contexts = rdma_get_devices(NULL); 2401 if (contexts == NULL) { 2402 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2403 nvmf_rdma_destroy(&rtransport->transport); 2404 return NULL; 2405 } 2406 2407 i = 0; 2408 rc = 0; 2409 while (contexts[i] != NULL) { 2410 device = calloc(1, sizeof(*device)); 2411 if (!device) { 2412 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2413 rc = -ENOMEM; 2414 break; 2415 } 2416 device->context = contexts[i]; 2417 rc = ibv_query_device(device->context, &device->attr); 2418 if (rc < 0) { 2419 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2420 free(device); 2421 break; 2422 2423 } 2424 2425 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2426 2427 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2428 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2429 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2430 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2431 } 2432 2433 /** 2434 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2435 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2436 * but incorrectly reports that it does. There are changes making their way 2437 * through the kernel now that will enable this feature. When they are merged, 2438 * we can conditionally enable this feature. 2439 * 2440 * TODO: enable this for versions of the kernel rxe driver that support it. 2441 */ 2442 if (device->attr.vendor_id == 0) { 2443 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2444 } 2445 #endif 2446 2447 /* set up device context async ev fd as NON_BLOCKING */ 2448 flag = fcntl(device->context->async_fd, F_GETFL); 2449 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2450 if (rc < 0) { 2451 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2452 free(device); 2453 break; 2454 } 2455 2456 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2457 i++; 2458 2459 if (g_nvmf_hooks.get_ibv_pd) { 2460 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2461 } else { 2462 device->pd = ibv_alloc_pd(device->context); 2463 } 2464 2465 if (!device->pd) { 2466 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2467 rc = -ENOMEM; 2468 break; 2469 } 2470 2471 assert(device->map == NULL); 2472 2473 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2474 if (!device->map) { 2475 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2476 rc = -ENOMEM; 2477 break; 2478 } 2479 2480 assert(device->map != NULL); 2481 assert(device->pd != NULL); 2482 } 2483 rdma_free_devices(contexts); 2484 2485 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2486 /* divide and round up. */ 2487 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2488 2489 /* round up to the nearest 4k. */ 2490 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2491 2492 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2493 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2494 opts->io_unit_size); 2495 } 2496 2497 if (rc < 0) { 2498 nvmf_rdma_destroy(&rtransport->transport); 2499 return NULL; 2500 } 2501 2502 /* Set up poll descriptor array to monitor events from RDMA and IB 2503 * in a single poll syscall 2504 */ 2505 rtransport->npoll_fds = i + 1; 2506 i = 0; 2507 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2508 if (rtransport->poll_fds == NULL) { 2509 SPDK_ERRLOG("poll_fds allocation failed\n"); 2510 nvmf_rdma_destroy(&rtransport->transport); 2511 return NULL; 2512 } 2513 2514 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2515 rtransport->poll_fds[i++].events = POLLIN; 2516 2517 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2518 rtransport->poll_fds[i].fd = device->context->async_fd; 2519 rtransport->poll_fds[i++].events = POLLIN; 2520 } 2521 2522 return &rtransport->transport; 2523 } 2524 2525 static void 2526 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2527 { 2528 struct spdk_nvmf_rdma_transport *rtransport; 2529 assert(w != NULL); 2530 2531 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2532 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2533 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2534 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2535 } 2536 2537 static int 2538 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2539 { 2540 struct spdk_nvmf_rdma_transport *rtransport; 2541 struct spdk_nvmf_rdma_port *port, *port_tmp; 2542 struct spdk_nvmf_rdma_device *device, *device_tmp; 2543 2544 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2545 2546 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2547 TAILQ_REMOVE(&rtransport->ports, port, link); 2548 rdma_destroy_id(port->id); 2549 free(port); 2550 } 2551 2552 if (rtransport->poll_fds != NULL) { 2553 free(rtransport->poll_fds); 2554 } 2555 2556 if (rtransport->event_channel != NULL) { 2557 rdma_destroy_event_channel(rtransport->event_channel); 2558 } 2559 2560 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2561 TAILQ_REMOVE(&rtransport->devices, device, link); 2562 if (device->map) { 2563 spdk_mem_map_free(&device->map); 2564 } 2565 if (device->pd) { 2566 if (!g_nvmf_hooks.get_ibv_pd) { 2567 ibv_dealloc_pd(device->pd); 2568 } 2569 } 2570 free(device); 2571 } 2572 2573 if (rtransport->data_wr_pool != NULL) { 2574 if (spdk_mempool_count(rtransport->data_wr_pool) != 2575 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2576 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2577 spdk_mempool_count(rtransport->data_wr_pool), 2578 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2579 } 2580 } 2581 2582 spdk_mempool_free(rtransport->data_wr_pool); 2583 2584 pthread_mutex_destroy(&rtransport->lock); 2585 free(rtransport); 2586 2587 return 0; 2588 } 2589 2590 static int 2591 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2592 struct spdk_nvme_transport_id *trid, 2593 bool peer); 2594 2595 static int 2596 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2597 const struct spdk_nvme_transport_id *trid) 2598 { 2599 struct spdk_nvmf_rdma_transport *rtransport; 2600 struct spdk_nvmf_rdma_device *device; 2601 struct spdk_nvmf_rdma_port *port; 2602 struct addrinfo *res; 2603 struct addrinfo hints; 2604 int family; 2605 int rc; 2606 2607 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2608 assert(rtransport->event_channel != NULL); 2609 2610 pthread_mutex_lock(&rtransport->lock); 2611 port = calloc(1, sizeof(*port)); 2612 if (!port) { 2613 SPDK_ERRLOG("Port allocation failed\n"); 2614 pthread_mutex_unlock(&rtransport->lock); 2615 return -ENOMEM; 2616 } 2617 2618 port->trid = trid; 2619 2620 switch (trid->adrfam) { 2621 case SPDK_NVMF_ADRFAM_IPV4: 2622 family = AF_INET; 2623 break; 2624 case SPDK_NVMF_ADRFAM_IPV6: 2625 family = AF_INET6; 2626 break; 2627 default: 2628 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2629 free(port); 2630 pthread_mutex_unlock(&rtransport->lock); 2631 return -EINVAL; 2632 } 2633 2634 memset(&hints, 0, sizeof(hints)); 2635 hints.ai_family = family; 2636 hints.ai_flags = AI_NUMERICSERV; 2637 hints.ai_socktype = SOCK_STREAM; 2638 hints.ai_protocol = 0; 2639 2640 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2641 if (rc) { 2642 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2643 free(port); 2644 pthread_mutex_unlock(&rtransport->lock); 2645 return -EINVAL; 2646 } 2647 2648 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2649 if (rc < 0) { 2650 SPDK_ERRLOG("rdma_create_id() failed\n"); 2651 freeaddrinfo(res); 2652 free(port); 2653 pthread_mutex_unlock(&rtransport->lock); 2654 return rc; 2655 } 2656 2657 rc = rdma_bind_addr(port->id, res->ai_addr); 2658 freeaddrinfo(res); 2659 2660 if (rc < 0) { 2661 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2662 rdma_destroy_id(port->id); 2663 free(port); 2664 pthread_mutex_unlock(&rtransport->lock); 2665 return rc; 2666 } 2667 2668 if (!port->id->verbs) { 2669 SPDK_ERRLOG("ibv_context is null\n"); 2670 rdma_destroy_id(port->id); 2671 free(port); 2672 pthread_mutex_unlock(&rtransport->lock); 2673 return -1; 2674 } 2675 2676 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2677 if (rc < 0) { 2678 SPDK_ERRLOG("rdma_listen() failed\n"); 2679 rdma_destroy_id(port->id); 2680 free(port); 2681 pthread_mutex_unlock(&rtransport->lock); 2682 return rc; 2683 } 2684 2685 TAILQ_FOREACH(device, &rtransport->devices, link) { 2686 if (device->context == port->id->verbs) { 2687 port->device = device; 2688 break; 2689 } 2690 } 2691 if (!port->device) { 2692 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2693 port->id->verbs); 2694 rdma_destroy_id(port->id); 2695 free(port); 2696 pthread_mutex_unlock(&rtransport->lock); 2697 return -EINVAL; 2698 } 2699 2700 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2701 trid->traddr, trid->trsvcid); 2702 2703 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2704 pthread_mutex_unlock(&rtransport->lock); 2705 return 0; 2706 } 2707 2708 static void 2709 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2710 const struct spdk_nvme_transport_id *trid) 2711 { 2712 struct spdk_nvmf_rdma_transport *rtransport; 2713 struct spdk_nvmf_rdma_port *port, *tmp; 2714 2715 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2716 2717 pthread_mutex_lock(&rtransport->lock); 2718 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2719 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2720 TAILQ_REMOVE(&rtransport->ports, port, link); 2721 rdma_destroy_id(port->id); 2722 free(port); 2723 break; 2724 } 2725 } 2726 2727 pthread_mutex_unlock(&rtransport->lock); 2728 } 2729 2730 static void 2731 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2732 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2733 { 2734 struct spdk_nvmf_request *req, *tmp; 2735 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2736 struct spdk_nvmf_rdma_resources *resources; 2737 2738 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2739 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2740 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2741 break; 2742 } 2743 } 2744 2745 /* Then RDMA writes since reads have stronger restrictions than writes */ 2746 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2747 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2748 break; 2749 } 2750 } 2751 2752 /* The second highest priority is I/O waiting on memory buffers. */ 2753 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2754 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2755 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2756 break; 2757 } 2758 } 2759 2760 resources = rqpair->resources; 2761 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2762 rdma_req = STAILQ_FIRST(&resources->free_queue); 2763 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2764 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2765 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2766 2767 if (rqpair->srq != NULL) { 2768 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2769 rdma_req->recv->qpair->qd++; 2770 } else { 2771 rqpair->qd++; 2772 } 2773 2774 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2775 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2776 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2777 break; 2778 } 2779 } 2780 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2781 rqpair->poller->stat.pending_free_request++; 2782 } 2783 } 2784 2785 static void 2786 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2787 { 2788 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2789 struct spdk_nvmf_rdma_transport, transport); 2790 2791 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2792 2793 /* nvmr_rdma_close_qpair is not called */ 2794 if (!rqpair->to_close) { 2795 return; 2796 } 2797 2798 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2799 if (rqpair->current_send_depth != 0) { 2800 return; 2801 } 2802 2803 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2804 return; 2805 } 2806 2807 /* Judge whether the device is emulated by Software RoCE. 2808 * And it will not send last_wqe event 2809 */ 2810 if (rqpair->srq != NULL && rqpair->device->attr.vendor_id != 0 && 2811 rqpair->last_wqe_reached == false) { 2812 return; 2813 } 2814 2815 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2816 2817 nvmf_rdma_qpair_destroy(rqpair); 2818 } 2819 2820 static int 2821 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2822 { 2823 struct spdk_nvmf_qpair *qpair; 2824 struct spdk_nvmf_rdma_qpair *rqpair; 2825 2826 if (evt->id == NULL) { 2827 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2828 return -1; 2829 } 2830 2831 qpair = evt->id->context; 2832 if (qpair == NULL) { 2833 SPDK_ERRLOG("disconnect request: no active connection\n"); 2834 return -1; 2835 } 2836 2837 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2838 2839 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2840 2841 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2842 2843 return 0; 2844 } 2845 2846 #ifdef DEBUG 2847 static const char *CM_EVENT_STR[] = { 2848 "RDMA_CM_EVENT_ADDR_RESOLVED", 2849 "RDMA_CM_EVENT_ADDR_ERROR", 2850 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2851 "RDMA_CM_EVENT_ROUTE_ERROR", 2852 "RDMA_CM_EVENT_CONNECT_REQUEST", 2853 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2854 "RDMA_CM_EVENT_CONNECT_ERROR", 2855 "RDMA_CM_EVENT_UNREACHABLE", 2856 "RDMA_CM_EVENT_REJECTED", 2857 "RDMA_CM_EVENT_ESTABLISHED", 2858 "RDMA_CM_EVENT_DISCONNECTED", 2859 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2860 "RDMA_CM_EVENT_MULTICAST_JOIN", 2861 "RDMA_CM_EVENT_MULTICAST_ERROR", 2862 "RDMA_CM_EVENT_ADDR_CHANGE", 2863 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2864 }; 2865 #endif /* DEBUG */ 2866 2867 static void 2868 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2869 struct spdk_nvmf_rdma_port *port) 2870 { 2871 struct spdk_nvmf_rdma_poll_group *rgroup; 2872 struct spdk_nvmf_rdma_poller *rpoller; 2873 struct spdk_nvmf_rdma_qpair *rqpair; 2874 2875 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2876 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2877 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 2878 if (rqpair->listen_id == port->id) { 2879 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2880 } 2881 } 2882 } 2883 } 2884 } 2885 2886 static bool 2887 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2888 struct rdma_cm_event *event) 2889 { 2890 const struct spdk_nvme_transport_id *trid; 2891 struct spdk_nvmf_rdma_port *port; 2892 struct spdk_nvmf_rdma_transport *rtransport; 2893 bool event_acked = false; 2894 2895 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2896 TAILQ_FOREACH(port, &rtransport->ports, link) { 2897 if (port->id == event->id) { 2898 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2899 rdma_ack_cm_event(event); 2900 event_acked = true; 2901 trid = port->trid; 2902 break; 2903 } 2904 } 2905 2906 if (event_acked) { 2907 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2908 2909 nvmf_rdma_stop_listen(transport, trid); 2910 nvmf_rdma_listen(transport, trid); 2911 } 2912 2913 return event_acked; 2914 } 2915 2916 static void 2917 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2918 struct rdma_cm_event *event) 2919 { 2920 struct spdk_nvmf_rdma_port *port; 2921 struct spdk_nvmf_rdma_transport *rtransport; 2922 2923 port = event->id->context; 2924 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2925 2926 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2927 2928 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2929 2930 rdma_ack_cm_event(event); 2931 2932 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2933 ; 2934 } 2935 } 2936 2937 static void 2938 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2939 { 2940 struct spdk_nvmf_rdma_transport *rtransport; 2941 struct rdma_cm_event *event; 2942 int rc; 2943 bool event_acked; 2944 2945 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2946 2947 if (rtransport->event_channel == NULL) { 2948 return; 2949 } 2950 2951 while (1) { 2952 event_acked = false; 2953 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2954 if (rc) { 2955 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2956 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2957 } 2958 break; 2959 } 2960 2961 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2962 2963 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2964 2965 switch (event->event) { 2966 case RDMA_CM_EVENT_ADDR_RESOLVED: 2967 case RDMA_CM_EVENT_ADDR_ERROR: 2968 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2969 case RDMA_CM_EVENT_ROUTE_ERROR: 2970 /* No action required. The target never attempts to resolve routes. */ 2971 break; 2972 case RDMA_CM_EVENT_CONNECT_REQUEST: 2973 rc = nvmf_rdma_connect(transport, event); 2974 if (rc < 0) { 2975 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2976 break; 2977 } 2978 break; 2979 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2980 /* The target never initiates a new connection. So this will not occur. */ 2981 break; 2982 case RDMA_CM_EVENT_CONNECT_ERROR: 2983 /* Can this happen? The docs say it can, but not sure what causes it. */ 2984 break; 2985 case RDMA_CM_EVENT_UNREACHABLE: 2986 case RDMA_CM_EVENT_REJECTED: 2987 /* These only occur on the client side. */ 2988 break; 2989 case RDMA_CM_EVENT_ESTABLISHED: 2990 /* TODO: Should we be waiting for this event anywhere? */ 2991 break; 2992 case RDMA_CM_EVENT_DISCONNECTED: 2993 rc = nvmf_rdma_disconnect(event); 2994 if (rc < 0) { 2995 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2996 break; 2997 } 2998 break; 2999 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3000 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3001 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3002 * Once these events are sent to SPDK, we should release all IB resources and 3003 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3004 * ibv_destory* functions to release user space memory allocated by IB. All kernel 3005 * resources are already cleaned. */ 3006 if (event->id->qp) { 3007 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3008 * corresponding qpair. Otherwise the event refers to a listening device */ 3009 rc = nvmf_rdma_disconnect(event); 3010 if (rc < 0) { 3011 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3012 break; 3013 } 3014 } else { 3015 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3016 event_acked = true; 3017 } 3018 break; 3019 case RDMA_CM_EVENT_MULTICAST_JOIN: 3020 case RDMA_CM_EVENT_MULTICAST_ERROR: 3021 /* Multicast is not used */ 3022 break; 3023 case RDMA_CM_EVENT_ADDR_CHANGE: 3024 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3025 break; 3026 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3027 /* For now, do nothing. The target never re-uses queue pairs. */ 3028 break; 3029 default: 3030 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3031 break; 3032 } 3033 if (!event_acked) { 3034 rdma_ack_cm_event(event); 3035 } 3036 } 3037 } 3038 3039 static void 3040 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3041 { 3042 rqpair->last_wqe_reached = true; 3043 nvmf_rdma_destroy_drained_qpair(rqpair); 3044 } 3045 3046 static void 3047 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3048 { 3049 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3050 3051 if (event_ctx->rqpair) { 3052 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3053 if (event_ctx->cb_fn) { 3054 event_ctx->cb_fn(event_ctx->rqpair); 3055 } 3056 } 3057 free(event_ctx); 3058 } 3059 3060 static int 3061 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3062 spdk_nvmf_rdma_qpair_ibv_event fn) 3063 { 3064 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3065 struct spdk_thread *thr = NULL; 3066 int rc; 3067 3068 if (rqpair->qpair.group) { 3069 thr = rqpair->qpair.group->thread; 3070 } else if (rqpair->destruct_channel) { 3071 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3072 } 3073 3074 if (!thr) { 3075 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3076 return -EINVAL; 3077 } 3078 3079 ctx = calloc(1, sizeof(*ctx)); 3080 if (!ctx) { 3081 return -ENOMEM; 3082 } 3083 3084 ctx->rqpair = rqpair; 3085 ctx->cb_fn = fn; 3086 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3087 3088 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3089 if (rc) { 3090 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3091 free(ctx); 3092 } 3093 3094 return rc; 3095 } 3096 3097 static int 3098 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3099 { 3100 int rc; 3101 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3102 struct ibv_async_event event; 3103 3104 rc = ibv_get_async_event(device->context, &event); 3105 3106 if (rc) { 3107 /* In non-blocking mode -1 means there are no events available */ 3108 return rc; 3109 } 3110 3111 switch (event.event_type) { 3112 case IBV_EVENT_QP_FATAL: 3113 rqpair = event.element.qp->qp_context; 3114 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3115 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3116 (uintptr_t)rqpair->cm_id, event.event_type); 3117 nvmf_rdma_update_ibv_state(rqpair); 3118 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3119 break; 3120 case IBV_EVENT_QP_LAST_WQE_REACHED: 3121 /* This event only occurs for shared receive queues. */ 3122 rqpair = event.element.qp->qp_context; 3123 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3124 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3125 if (rc) { 3126 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3127 rqpair->last_wqe_reached = true; 3128 } 3129 break; 3130 case IBV_EVENT_SQ_DRAINED: 3131 /* This event occurs frequently in both error and non-error states. 3132 * Check if the qpair is in an error state before sending a message. */ 3133 rqpair = event.element.qp->qp_context; 3134 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3135 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3136 (uintptr_t)rqpair->cm_id, event.event_type); 3137 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3138 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3139 } 3140 break; 3141 case IBV_EVENT_QP_REQ_ERR: 3142 case IBV_EVENT_QP_ACCESS_ERR: 3143 case IBV_EVENT_COMM_EST: 3144 case IBV_EVENT_PATH_MIG: 3145 case IBV_EVENT_PATH_MIG_ERR: 3146 SPDK_NOTICELOG("Async event: %s\n", 3147 ibv_event_type_str(event.event_type)); 3148 rqpair = event.element.qp->qp_context; 3149 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3150 (uintptr_t)rqpair->cm_id, event.event_type); 3151 nvmf_rdma_update_ibv_state(rqpair); 3152 break; 3153 case IBV_EVENT_CQ_ERR: 3154 case IBV_EVENT_DEVICE_FATAL: 3155 case IBV_EVENT_PORT_ACTIVE: 3156 case IBV_EVENT_PORT_ERR: 3157 case IBV_EVENT_LID_CHANGE: 3158 case IBV_EVENT_PKEY_CHANGE: 3159 case IBV_EVENT_SM_CHANGE: 3160 case IBV_EVENT_SRQ_ERR: 3161 case IBV_EVENT_SRQ_LIMIT_REACHED: 3162 case IBV_EVENT_CLIENT_REREGISTER: 3163 case IBV_EVENT_GID_CHANGE: 3164 default: 3165 SPDK_NOTICELOG("Async event: %s\n", 3166 ibv_event_type_str(event.event_type)); 3167 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3168 break; 3169 } 3170 ibv_ack_async_event(&event); 3171 3172 return 0; 3173 } 3174 3175 static void 3176 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3177 { 3178 int rc = 0; 3179 uint32_t i = 0; 3180 3181 for (i = 0; i < max_events; i++) { 3182 rc = nvmf_process_ib_event(device); 3183 if (rc) { 3184 break; 3185 } 3186 } 3187 3188 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3189 } 3190 3191 static uint32_t 3192 nvmf_rdma_accept(struct spdk_nvmf_transport *transport) 3193 { 3194 int nfds, i = 0; 3195 struct spdk_nvmf_rdma_transport *rtransport; 3196 struct spdk_nvmf_rdma_device *device, *tmp; 3197 uint32_t count; 3198 3199 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3200 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3201 3202 if (nfds <= 0) { 3203 return 0; 3204 } 3205 3206 /* The first poll descriptor is RDMA CM event */ 3207 if (rtransport->poll_fds[i++].revents & POLLIN) { 3208 nvmf_process_cm_event(transport); 3209 nfds--; 3210 } 3211 3212 if (nfds == 0) { 3213 return count; 3214 } 3215 3216 /* Second and subsequent poll descriptors are IB async events */ 3217 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3218 if (rtransport->poll_fds[i++].revents & POLLIN) { 3219 nvmf_process_ib_events(device, 32); 3220 nfds--; 3221 } 3222 } 3223 /* check all flagged fd's have been served */ 3224 assert(nfds == 0); 3225 3226 return count; 3227 } 3228 3229 static void 3230 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3231 struct spdk_nvmf_ctrlr_data *cdata) 3232 { 3233 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3234 3235 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3236 since in-capsule data only works with NVME drives that support SGL memory layout */ 3237 if (transport->opts.dif_insert_or_strip) { 3238 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3239 } 3240 } 3241 3242 static void 3243 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3244 struct spdk_nvme_transport_id *trid, 3245 struct spdk_nvmf_discovery_log_page_entry *entry) 3246 { 3247 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3248 entry->adrfam = trid->adrfam; 3249 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3250 3251 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3252 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3253 3254 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3255 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3256 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3257 } 3258 3259 static void 3260 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3261 3262 static struct spdk_nvmf_transport_poll_group * 3263 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3264 { 3265 struct spdk_nvmf_rdma_transport *rtransport; 3266 struct spdk_nvmf_rdma_poll_group *rgroup; 3267 struct spdk_nvmf_rdma_poller *poller; 3268 struct spdk_nvmf_rdma_device *device; 3269 struct ibv_srq_init_attr srq_init_attr; 3270 struct spdk_nvmf_rdma_resource_opts opts; 3271 int num_cqe; 3272 3273 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3274 3275 rgroup = calloc(1, sizeof(*rgroup)); 3276 if (!rgroup) { 3277 return NULL; 3278 } 3279 3280 TAILQ_INIT(&rgroup->pollers); 3281 STAILQ_INIT(&rgroup->retired_bufs); 3282 3283 pthread_mutex_lock(&rtransport->lock); 3284 TAILQ_FOREACH(device, &rtransport->devices, link) { 3285 poller = calloc(1, sizeof(*poller)); 3286 if (!poller) { 3287 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3288 nvmf_rdma_poll_group_destroy(&rgroup->group); 3289 pthread_mutex_unlock(&rtransport->lock); 3290 return NULL; 3291 } 3292 3293 poller->device = device; 3294 poller->group = rgroup; 3295 3296 TAILQ_INIT(&poller->qpairs); 3297 STAILQ_INIT(&poller->qpairs_pending_send); 3298 STAILQ_INIT(&poller->qpairs_pending_recv); 3299 3300 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3301 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3302 poller->max_srq_depth = rtransport->rdma_opts.max_srq_depth; 3303 3304 device->num_srq++; 3305 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 3306 srq_init_attr.attr.max_wr = poller->max_srq_depth; 3307 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3308 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 3309 if (!poller->srq) { 3310 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3311 nvmf_rdma_poll_group_destroy(&rgroup->group); 3312 pthread_mutex_unlock(&rtransport->lock); 3313 return NULL; 3314 } 3315 3316 opts.qp = poller->srq; 3317 opts.pd = device->pd; 3318 opts.qpair = NULL; 3319 opts.shared = true; 3320 opts.max_queue_depth = poller->max_srq_depth; 3321 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3322 3323 poller->resources = nvmf_rdma_resources_create(&opts); 3324 if (!poller->resources) { 3325 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3326 nvmf_rdma_poll_group_destroy(&rgroup->group); 3327 pthread_mutex_unlock(&rtransport->lock); 3328 return NULL; 3329 } 3330 } 3331 3332 /* 3333 * When using an srq, we can limit the completion queue at startup. 3334 * The following formula represents the calculation: 3335 * num_cqe = num_recv + num_data_wr + num_send_wr. 3336 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3337 */ 3338 if (poller->srq) { 3339 num_cqe = poller->max_srq_depth * 3; 3340 } else { 3341 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3342 } 3343 3344 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3345 if (!poller->cq) { 3346 SPDK_ERRLOG("Unable to create completion queue\n"); 3347 nvmf_rdma_poll_group_destroy(&rgroup->group); 3348 pthread_mutex_unlock(&rtransport->lock); 3349 return NULL; 3350 } 3351 poller->num_cqe = num_cqe; 3352 } 3353 3354 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3355 if (rtransport->conn_sched.next_admin_pg == NULL) { 3356 rtransport->conn_sched.next_admin_pg = rgroup; 3357 rtransport->conn_sched.next_io_pg = rgroup; 3358 } 3359 3360 pthread_mutex_unlock(&rtransport->lock); 3361 return &rgroup->group; 3362 } 3363 3364 static struct spdk_nvmf_transport_poll_group * 3365 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3366 { 3367 struct spdk_nvmf_rdma_transport *rtransport; 3368 struct spdk_nvmf_rdma_poll_group **pg; 3369 struct spdk_nvmf_transport_poll_group *result; 3370 3371 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3372 3373 pthread_mutex_lock(&rtransport->lock); 3374 3375 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3376 pthread_mutex_unlock(&rtransport->lock); 3377 return NULL; 3378 } 3379 3380 if (qpair->qid == 0) { 3381 pg = &rtransport->conn_sched.next_admin_pg; 3382 } else { 3383 pg = &rtransport->conn_sched.next_io_pg; 3384 } 3385 3386 assert(*pg != NULL); 3387 3388 result = &(*pg)->group; 3389 3390 *pg = TAILQ_NEXT(*pg, link); 3391 if (*pg == NULL) { 3392 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3393 } 3394 3395 pthread_mutex_unlock(&rtransport->lock); 3396 3397 return result; 3398 } 3399 3400 static void 3401 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3402 { 3403 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3404 struct spdk_nvmf_rdma_poller *poller, *tmp; 3405 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3406 struct spdk_nvmf_transport_pg_cache_buf *buf, *tmp_buf; 3407 struct spdk_nvmf_rdma_transport *rtransport; 3408 3409 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3410 if (!rgroup) { 3411 return; 3412 } 3413 3414 /* free all retired buffers back to the transport so we don't short the mempool. */ 3415 STAILQ_FOREACH_SAFE(buf, &rgroup->retired_bufs, link, tmp_buf) { 3416 STAILQ_REMOVE(&rgroup->retired_bufs, buf, spdk_nvmf_transport_pg_cache_buf, link); 3417 assert(group->transport != NULL); 3418 spdk_mempool_put(group->transport->data_buf_pool, buf); 3419 } 3420 3421 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3422 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3423 3424 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3425 nvmf_rdma_qpair_destroy(qpair); 3426 } 3427 3428 if (poller->srq) { 3429 if (poller->resources) { 3430 nvmf_rdma_resources_destroy(poller->resources); 3431 } 3432 ibv_destroy_srq(poller->srq); 3433 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 3434 } 3435 3436 if (poller->cq) { 3437 ibv_destroy_cq(poller->cq); 3438 } 3439 3440 free(poller); 3441 } 3442 3443 if (rgroup->group.transport == NULL) { 3444 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3445 * calls this function directly in a failure path. */ 3446 free(rgroup); 3447 return; 3448 } 3449 3450 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3451 3452 pthread_mutex_lock(&rtransport->lock); 3453 next_rgroup = TAILQ_NEXT(rgroup, link); 3454 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3455 if (next_rgroup == NULL) { 3456 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3457 } 3458 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3459 rtransport->conn_sched.next_admin_pg = next_rgroup; 3460 } 3461 if (rtransport->conn_sched.next_io_pg == rgroup) { 3462 rtransport->conn_sched.next_io_pg = next_rgroup; 3463 } 3464 pthread_mutex_unlock(&rtransport->lock); 3465 3466 free(rgroup); 3467 } 3468 3469 static void 3470 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3471 { 3472 if (rqpair->cm_id != NULL) { 3473 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3474 } 3475 } 3476 3477 static int 3478 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3479 struct spdk_nvmf_qpair *qpair) 3480 { 3481 struct spdk_nvmf_rdma_poll_group *rgroup; 3482 struct spdk_nvmf_rdma_qpair *rqpair; 3483 struct spdk_nvmf_rdma_device *device; 3484 struct spdk_nvmf_rdma_poller *poller; 3485 int rc; 3486 3487 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3488 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3489 3490 device = rqpair->device; 3491 3492 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3493 if (poller->device == device) { 3494 break; 3495 } 3496 } 3497 3498 if (!poller) { 3499 SPDK_ERRLOG("No poller found for device.\n"); 3500 return -1; 3501 } 3502 3503 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3504 rqpair->poller = poller; 3505 rqpair->srq = rqpair->poller->srq; 3506 3507 rc = nvmf_rdma_qpair_initialize(qpair); 3508 if (rc < 0) { 3509 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3510 return -1; 3511 } 3512 3513 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3514 if (rc) { 3515 /* Try to reject, but we probably can't */ 3516 nvmf_rdma_qpair_reject_connection(rqpair); 3517 return -1; 3518 } 3519 3520 nvmf_rdma_update_ibv_state(rqpair); 3521 3522 return 0; 3523 } 3524 3525 static int 3526 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3527 struct spdk_nvmf_qpair *qpair) 3528 { 3529 struct spdk_nvmf_rdma_qpair *rqpair; 3530 3531 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3532 assert(group->transport->tgt != NULL); 3533 3534 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3535 3536 if (!rqpair->destruct_channel) { 3537 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3538 return 0; 3539 } 3540 3541 /* Sanity check that we get io_channel on the correct thread */ 3542 if (qpair->group) { 3543 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3544 } 3545 3546 return 0; 3547 } 3548 3549 static int 3550 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3551 { 3552 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3553 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3554 struct spdk_nvmf_rdma_transport, transport); 3555 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3556 struct spdk_nvmf_rdma_qpair, qpair); 3557 3558 /* 3559 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3560 * needs to be returned to the shared receive queue or the poll group will eventually be 3561 * starved of RECV structures. 3562 */ 3563 if (rqpair->srq && rdma_req->recv) { 3564 int rc; 3565 struct ibv_recv_wr *bad_recv_wr; 3566 3567 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 3568 if (rc) { 3569 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3570 } 3571 } 3572 3573 _nvmf_rdma_request_free(rdma_req, rtransport); 3574 return 0; 3575 } 3576 3577 static int 3578 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3579 { 3580 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3581 struct spdk_nvmf_rdma_transport, transport); 3582 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3583 struct spdk_nvmf_rdma_request, req); 3584 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3585 struct spdk_nvmf_rdma_qpair, qpair); 3586 3587 if (rqpair->ibv_state != IBV_QPS_ERR) { 3588 /* The connection is alive, so process the request as normal */ 3589 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3590 } else { 3591 /* The connection is dead. Move the request directly to the completed state. */ 3592 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3593 } 3594 3595 nvmf_rdma_request_process(rtransport, rdma_req); 3596 3597 return 0; 3598 } 3599 3600 static void 3601 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3602 { 3603 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3604 3605 rqpair->to_close = true; 3606 3607 /* This happens only when the qpair is disconnected before 3608 * it is added to the poll group. Since there is no poll group, 3609 * the RDMA qp has not been initialized yet and the RDMA CM 3610 * event has not yet been acknowledged, so we need to reject it. 3611 */ 3612 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3613 nvmf_rdma_qpair_reject_connection(rqpair); 3614 nvmf_rdma_qpair_destroy(rqpair); 3615 return; 3616 } 3617 3618 if (rqpair->rdma_qp) { 3619 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3620 } 3621 3622 nvmf_rdma_destroy_drained_qpair(rqpair); 3623 } 3624 3625 static struct spdk_nvmf_rdma_qpair * 3626 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3627 { 3628 struct spdk_nvmf_rdma_qpair *rqpair; 3629 /* @todo: improve QP search */ 3630 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3631 if (wc->qp_num == rqpair->rdma_qp->qp->qp_num) { 3632 return rqpair; 3633 } 3634 } 3635 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3636 return NULL; 3637 } 3638 3639 #ifdef DEBUG 3640 static int 3641 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3642 { 3643 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3644 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3645 } 3646 #endif 3647 3648 static void 3649 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3650 int rc) 3651 { 3652 struct spdk_nvmf_rdma_recv *rdma_recv; 3653 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3654 3655 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3656 while (bad_recv_wr != NULL) { 3657 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3658 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3659 3660 rdma_recv->qpair->current_recv_depth++; 3661 bad_recv_wr = bad_recv_wr->next; 3662 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3663 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3664 } 3665 } 3666 3667 static void 3668 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3669 { 3670 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3671 while (bad_recv_wr != NULL) { 3672 bad_recv_wr = bad_recv_wr->next; 3673 rqpair->current_recv_depth++; 3674 } 3675 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3676 } 3677 3678 static void 3679 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3680 struct spdk_nvmf_rdma_poller *rpoller) 3681 { 3682 struct spdk_nvmf_rdma_qpair *rqpair; 3683 struct ibv_recv_wr *bad_recv_wr; 3684 int rc; 3685 3686 if (rpoller->srq) { 3687 if (rpoller->resources->recvs_to_post.first != NULL) { 3688 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3689 if (rc) { 3690 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3691 } 3692 rpoller->resources->recvs_to_post.first = NULL; 3693 rpoller->resources->recvs_to_post.last = NULL; 3694 } 3695 } else { 3696 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3697 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3698 assert(rqpair->resources->recvs_to_post.first != NULL); 3699 rc = ibv_post_recv(rqpair->rdma_qp->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3700 if (rc) { 3701 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3702 } 3703 rqpair->resources->recvs_to_post.first = NULL; 3704 rqpair->resources->recvs_to_post.last = NULL; 3705 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3706 } 3707 } 3708 } 3709 3710 static void 3711 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3712 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3713 { 3714 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3715 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3716 3717 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3718 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3719 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3720 assert(rqpair->current_send_depth > 0); 3721 rqpair->current_send_depth--; 3722 switch (bad_rdma_wr->type) { 3723 case RDMA_WR_TYPE_DATA: 3724 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3725 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3726 assert(rqpair->current_read_depth > 0); 3727 rqpair->current_read_depth--; 3728 } 3729 break; 3730 case RDMA_WR_TYPE_SEND: 3731 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3732 break; 3733 default: 3734 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3735 prev_rdma_req = cur_rdma_req; 3736 continue; 3737 } 3738 3739 if (prev_rdma_req == cur_rdma_req) { 3740 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3741 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3742 continue; 3743 } 3744 3745 switch (cur_rdma_req->state) { 3746 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3747 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3748 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3749 break; 3750 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3751 case RDMA_REQUEST_STATE_COMPLETING: 3752 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3753 break; 3754 default: 3755 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3756 cur_rdma_req->state, rqpair); 3757 continue; 3758 } 3759 3760 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3761 prev_rdma_req = cur_rdma_req; 3762 } 3763 3764 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3765 /* Disconnect the connection. */ 3766 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3767 } 3768 3769 } 3770 3771 static void 3772 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3773 struct spdk_nvmf_rdma_poller *rpoller) 3774 { 3775 struct spdk_nvmf_rdma_qpair *rqpair; 3776 struct ibv_send_wr *bad_wr = NULL; 3777 int rc; 3778 3779 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3780 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3781 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3782 3783 /* bad wr always points to the first wr that failed. */ 3784 if (rc) { 3785 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3786 } 3787 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3788 } 3789 } 3790 3791 static const char * 3792 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 3793 { 3794 switch (wr_type) { 3795 case RDMA_WR_TYPE_RECV: 3796 return "RECV"; 3797 case RDMA_WR_TYPE_SEND: 3798 return "SEND"; 3799 case RDMA_WR_TYPE_DATA: 3800 return "DATA"; 3801 default: 3802 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 3803 SPDK_UNREACHABLE(); 3804 } 3805 } 3806 3807 static inline void 3808 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 3809 { 3810 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 3811 3812 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 3813 /* If qpair is in ERR state, we will receive completions for all posted and not completed 3814 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 3815 SPDK_DEBUGLOG(rdma, 3816 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3817 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3818 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3819 } else { 3820 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3821 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3822 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3823 } 3824 } 3825 3826 static int 3827 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3828 struct spdk_nvmf_rdma_poller *rpoller) 3829 { 3830 struct ibv_wc wc[32]; 3831 struct spdk_nvmf_rdma_wr *rdma_wr; 3832 struct spdk_nvmf_rdma_request *rdma_req; 3833 struct spdk_nvmf_rdma_recv *rdma_recv; 3834 struct spdk_nvmf_rdma_qpair *rqpair; 3835 int reaped, i; 3836 int count = 0; 3837 bool error = false; 3838 uint64_t poll_tsc = spdk_get_ticks(); 3839 3840 /* Poll for completing operations. */ 3841 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3842 if (reaped < 0) { 3843 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3844 errno, spdk_strerror(errno)); 3845 return -1; 3846 } 3847 3848 rpoller->stat.polls++; 3849 rpoller->stat.completions += reaped; 3850 3851 for (i = 0; i < reaped; i++) { 3852 3853 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3854 3855 switch (rdma_wr->type) { 3856 case RDMA_WR_TYPE_SEND: 3857 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3858 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3859 3860 if (!wc[i].status) { 3861 count++; 3862 assert(wc[i].opcode == IBV_WC_SEND); 3863 assert(nvmf_rdma_req_is_completing(rdma_req)); 3864 } 3865 3866 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3867 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3868 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3869 rdma_req->num_outstanding_data_wr = 0; 3870 3871 nvmf_rdma_request_process(rtransport, rdma_req); 3872 break; 3873 case RDMA_WR_TYPE_RECV: 3874 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3875 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3876 if (rpoller->srq != NULL) { 3877 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3878 /* It is possible that there are still some completions for destroyed QP 3879 * associated with SRQ. We just ignore these late completions and re-post 3880 * receive WRs back to SRQ. 3881 */ 3882 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3883 struct ibv_recv_wr *bad_wr; 3884 int rc; 3885 3886 rdma_recv->wr.next = NULL; 3887 rc = ibv_post_srq_recv(rpoller->srq, 3888 &rdma_recv->wr, 3889 &bad_wr); 3890 if (rc) { 3891 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3892 } 3893 continue; 3894 } 3895 } 3896 rqpair = rdma_recv->qpair; 3897 3898 assert(rqpair != NULL); 3899 if (!wc[i].status) { 3900 assert(wc[i].opcode == IBV_WC_RECV); 3901 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3902 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3903 break; 3904 } 3905 } 3906 3907 rdma_recv->wr.next = NULL; 3908 rqpair->current_recv_depth++; 3909 rdma_recv->receive_tsc = poll_tsc; 3910 rpoller->stat.requests++; 3911 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3912 break; 3913 case RDMA_WR_TYPE_DATA: 3914 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3915 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3916 3917 assert(rdma_req->num_outstanding_data_wr > 0); 3918 3919 rqpair->current_send_depth--; 3920 rdma_req->num_outstanding_data_wr--; 3921 if (!wc[i].status) { 3922 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3923 rqpair->current_read_depth--; 3924 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3925 if (rdma_req->num_outstanding_data_wr == 0) { 3926 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3927 nvmf_rdma_request_process(rtransport, rdma_req); 3928 } 3929 } else { 3930 /* If the data transfer fails still force the queue into the error state, 3931 * if we were performing an RDMA_READ, we need to force the request into a 3932 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3933 * case, we should wait for the SEND to complete. */ 3934 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3935 rqpair->current_read_depth--; 3936 if (rdma_req->num_outstanding_data_wr == 0) { 3937 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3938 } 3939 } 3940 } 3941 break; 3942 default: 3943 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3944 continue; 3945 } 3946 3947 /* Handle error conditions */ 3948 if (wc[i].status) { 3949 nvmf_rdma_update_ibv_state(rqpair); 3950 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 3951 3952 error = true; 3953 3954 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3955 /* Disconnect the connection. */ 3956 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3957 } else { 3958 nvmf_rdma_destroy_drained_qpair(rqpair); 3959 } 3960 continue; 3961 } 3962 3963 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3964 3965 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3966 nvmf_rdma_destroy_drained_qpair(rqpair); 3967 } 3968 } 3969 3970 if (error == true) { 3971 return -1; 3972 } 3973 3974 /* submit outstanding work requests. */ 3975 _poller_submit_recvs(rtransport, rpoller); 3976 _poller_submit_sends(rtransport, rpoller); 3977 3978 return count; 3979 } 3980 3981 static int 3982 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3983 { 3984 struct spdk_nvmf_rdma_transport *rtransport; 3985 struct spdk_nvmf_rdma_poll_group *rgroup; 3986 struct spdk_nvmf_rdma_poller *rpoller; 3987 int count, rc; 3988 3989 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3990 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3991 3992 count = 0; 3993 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3994 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 3995 if (rc < 0) { 3996 return rc; 3997 } 3998 count += rc; 3999 } 4000 4001 return count; 4002 } 4003 4004 static int 4005 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4006 struct spdk_nvme_transport_id *trid, 4007 bool peer) 4008 { 4009 struct sockaddr *saddr; 4010 uint16_t port; 4011 4012 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4013 4014 if (peer) { 4015 saddr = rdma_get_peer_addr(id); 4016 } else { 4017 saddr = rdma_get_local_addr(id); 4018 } 4019 switch (saddr->sa_family) { 4020 case AF_INET: { 4021 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4022 4023 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4024 inet_ntop(AF_INET, &saddr_in->sin_addr, 4025 trid->traddr, sizeof(trid->traddr)); 4026 if (peer) { 4027 port = ntohs(rdma_get_dst_port(id)); 4028 } else { 4029 port = ntohs(rdma_get_src_port(id)); 4030 } 4031 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4032 break; 4033 } 4034 case AF_INET6: { 4035 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4036 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4037 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4038 trid->traddr, sizeof(trid->traddr)); 4039 if (peer) { 4040 port = ntohs(rdma_get_dst_port(id)); 4041 } else { 4042 port = ntohs(rdma_get_src_port(id)); 4043 } 4044 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4045 break; 4046 } 4047 default: 4048 return -1; 4049 4050 } 4051 4052 return 0; 4053 } 4054 4055 static int 4056 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4057 struct spdk_nvme_transport_id *trid) 4058 { 4059 struct spdk_nvmf_rdma_qpair *rqpair; 4060 4061 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4062 4063 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4064 } 4065 4066 static int 4067 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4068 struct spdk_nvme_transport_id *trid) 4069 { 4070 struct spdk_nvmf_rdma_qpair *rqpair; 4071 4072 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4073 4074 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4075 } 4076 4077 static int 4078 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4079 struct spdk_nvme_transport_id *trid) 4080 { 4081 struct spdk_nvmf_rdma_qpair *rqpair; 4082 4083 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4084 4085 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4086 } 4087 4088 void 4089 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4090 { 4091 g_nvmf_hooks = *hooks; 4092 } 4093 4094 static void 4095 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4096 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4097 { 4098 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4099 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4100 4101 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4102 4103 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4104 } 4105 4106 static int 4107 _nvmf_rdma_qpair_abort_request(void *ctx) 4108 { 4109 struct spdk_nvmf_request *req = ctx; 4110 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4111 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4112 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4113 struct spdk_nvmf_rdma_qpair, qpair); 4114 int rc; 4115 4116 spdk_poller_unregister(&req->poller); 4117 4118 switch (rdma_req_to_abort->state) { 4119 case RDMA_REQUEST_STATE_EXECUTING: 4120 rc = nvmf_ctrlr_abort_request(req); 4121 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4122 return SPDK_POLLER_BUSY; 4123 } 4124 break; 4125 4126 case RDMA_REQUEST_STATE_NEED_BUFFER: 4127 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4128 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4129 4130 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4131 break; 4132 4133 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4134 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4135 spdk_nvmf_rdma_request, state_link); 4136 4137 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4138 break; 4139 4140 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4141 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4142 spdk_nvmf_rdma_request, state_link); 4143 4144 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4145 break; 4146 4147 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4148 if (spdk_get_ticks() < req->timeout_tsc) { 4149 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4150 return SPDK_POLLER_BUSY; 4151 } 4152 break; 4153 4154 default: 4155 break; 4156 } 4157 4158 spdk_nvmf_request_complete(req); 4159 return SPDK_POLLER_BUSY; 4160 } 4161 4162 static void 4163 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4164 struct spdk_nvmf_request *req) 4165 { 4166 struct spdk_nvmf_rdma_qpair *rqpair; 4167 struct spdk_nvmf_rdma_transport *rtransport; 4168 struct spdk_nvmf_transport *transport; 4169 uint16_t cid; 4170 uint32_t i; 4171 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL; 4172 4173 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4174 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4175 transport = &rtransport->transport; 4176 4177 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4178 4179 for (i = 0; i < rqpair->max_queue_depth; i++) { 4180 rdma_req_to_abort = &rqpair->resources->reqs[i]; 4181 4182 if (rdma_req_to_abort->state != RDMA_REQUEST_STATE_FREE && 4183 rdma_req_to_abort->req.cmd->nvme_cmd.cid == cid) { 4184 break; 4185 } 4186 } 4187 4188 if (rdma_req_to_abort == NULL) { 4189 spdk_nvmf_request_complete(req); 4190 return; 4191 } 4192 4193 req->req_to_abort = &rdma_req_to_abort->req; 4194 req->timeout_tsc = spdk_get_ticks() + 4195 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4196 req->poller = NULL; 4197 4198 _nvmf_rdma_qpair_abort_request(req); 4199 } 4200 4201 static int 4202 nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4203 struct spdk_nvmf_transport_poll_group_stat **stat) 4204 { 4205 struct spdk_io_channel *ch; 4206 struct spdk_nvmf_poll_group *group; 4207 struct spdk_nvmf_transport_poll_group *tgroup; 4208 struct spdk_nvmf_rdma_poll_group *rgroup; 4209 struct spdk_nvmf_rdma_poller *rpoller; 4210 struct spdk_nvmf_rdma_device_stat *device_stat; 4211 uint64_t num_devices = 0; 4212 4213 if (tgt == NULL || stat == NULL) { 4214 return -EINVAL; 4215 } 4216 4217 ch = spdk_get_io_channel(tgt); 4218 group = spdk_io_channel_get_ctx(ch);; 4219 spdk_put_io_channel(ch); 4220 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4221 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4222 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4223 if (!*stat) { 4224 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4225 return -ENOMEM; 4226 } 4227 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4228 4229 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4230 /* Count devices to allocate enough memory */ 4231 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4232 ++num_devices; 4233 } 4234 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4235 if (!(*stat)->rdma.devices) { 4236 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4237 free(*stat); 4238 return -ENOMEM; 4239 } 4240 4241 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4242 (*stat)->rdma.num_devices = num_devices; 4243 num_devices = 0; 4244 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4245 device_stat = &(*stat)->rdma.devices[num_devices++]; 4246 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4247 device_stat->polls = rpoller->stat.polls; 4248 device_stat->completions = rpoller->stat.completions; 4249 device_stat->requests = rpoller->stat.requests; 4250 device_stat->request_latency = rpoller->stat.request_latency; 4251 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4252 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4253 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4254 } 4255 return 0; 4256 } 4257 } 4258 return -ENOENT; 4259 } 4260 4261 static void 4262 nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4263 { 4264 if (stat) { 4265 free(stat->rdma.devices); 4266 } 4267 free(stat); 4268 } 4269 4270 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4271 .name = "RDMA", 4272 .type = SPDK_NVME_TRANSPORT_RDMA, 4273 .opts_init = nvmf_rdma_opts_init, 4274 .create = nvmf_rdma_create, 4275 .dump_opts = nvmf_rdma_dump_opts, 4276 .destroy = nvmf_rdma_destroy, 4277 4278 .listen = nvmf_rdma_listen, 4279 .stop_listen = nvmf_rdma_stop_listen, 4280 .accept = nvmf_rdma_accept, 4281 .cdata_init = nvmf_rdma_cdata_init, 4282 4283 .listener_discover = nvmf_rdma_discover, 4284 4285 .poll_group_create = nvmf_rdma_poll_group_create, 4286 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4287 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4288 .poll_group_add = nvmf_rdma_poll_group_add, 4289 .poll_group_remove = nvmf_rdma_poll_group_remove, 4290 .poll_group_poll = nvmf_rdma_poll_group_poll, 4291 4292 .req_free = nvmf_rdma_request_free, 4293 .req_complete = nvmf_rdma_request_complete, 4294 4295 .qpair_fini = nvmf_rdma_close_qpair, 4296 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4297 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4298 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4299 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4300 4301 .poll_group_get_stat = nvmf_rdma_poll_group_get_stat, 4302 .poll_group_free_stat = nvmf_rdma_poll_group_free_stat, 4303 }; 4304 4305 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4306 SPDK_LOG_REGISTER_COMPONENT(rdma) 4307