1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2019, 2020 Mellanox Technologies LTD. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of Intel Corporation nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include "spdk/stdinc.h" 35 36 #include "spdk/config.h" 37 #include "spdk/thread.h" 38 #include "spdk/likely.h" 39 #include "spdk/nvmf_transport.h" 40 #include "spdk/string.h" 41 #include "spdk/trace.h" 42 #include "spdk/util.h" 43 44 #include "spdk_internal/assert.h" 45 #include "spdk_internal/log.h" 46 #include "spdk_internal/rdma.h" 47 48 #include "nvmf_internal.h" 49 50 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 51 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 52 53 /* 54 RDMA Connection Resource Defaults 55 */ 56 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 57 #define NVMF_DEFAULT_RSP_SGE 1 58 #define NVMF_DEFAULT_RX_SGE 2 59 60 /* The RDMA completion queue size */ 61 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 62 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 63 64 /* Timeout for destroying defunct rqpairs */ 65 #define NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US 4000000 66 67 static int g_spdk_nvmf_ibv_query_mask = 68 IBV_QP_STATE | 69 IBV_QP_PKEY_INDEX | 70 IBV_QP_PORT | 71 IBV_QP_ACCESS_FLAGS | 72 IBV_QP_AV | 73 IBV_QP_PATH_MTU | 74 IBV_QP_DEST_QPN | 75 IBV_QP_RQ_PSN | 76 IBV_QP_MAX_DEST_RD_ATOMIC | 77 IBV_QP_MIN_RNR_TIMER | 78 IBV_QP_SQ_PSN | 79 IBV_QP_TIMEOUT | 80 IBV_QP_RETRY_CNT | 81 IBV_QP_RNR_RETRY | 82 IBV_QP_MAX_QP_RD_ATOMIC; 83 84 enum spdk_nvmf_rdma_request_state { 85 /* The request is not currently in use */ 86 RDMA_REQUEST_STATE_FREE = 0, 87 88 /* Initial state when request first received */ 89 RDMA_REQUEST_STATE_NEW, 90 91 /* The request is queued until a data buffer is available. */ 92 RDMA_REQUEST_STATE_NEED_BUFFER, 93 94 /* The request is waiting on RDMA queue depth availability 95 * to transfer data from the host to the controller. 96 */ 97 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 98 99 /* The request is currently transferring data from the host to the controller. */ 100 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 101 102 /* The request is ready to execute at the block device */ 103 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 104 105 /* The request is currently executing at the block device */ 106 RDMA_REQUEST_STATE_EXECUTING, 107 108 /* The request finished executing at the block device */ 109 RDMA_REQUEST_STATE_EXECUTED, 110 111 /* The request is waiting on RDMA queue depth availability 112 * to transfer data from the controller to the host. 113 */ 114 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 115 116 /* The request is ready to send a completion */ 117 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 118 119 /* The request is currently transferring data from the controller to the host. */ 120 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 121 122 /* The request currently has an outstanding completion without an 123 * associated data transfer. 124 */ 125 RDMA_REQUEST_STATE_COMPLETING, 126 127 /* The request completed and can be marked free. */ 128 RDMA_REQUEST_STATE_COMPLETED, 129 130 /* Terminator */ 131 RDMA_REQUEST_NUM_STATES, 132 }; 133 134 #define OBJECT_NVMF_RDMA_IO 0x40 135 136 #define TRACE_GROUP_NVMF_RDMA 0x4 137 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 138 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 139 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 140 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 141 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 142 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 143 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 144 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 145 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 146 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 147 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 148 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 149 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 150 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 151 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 152 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 153 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 154 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 155 156 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 157 { 158 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 159 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 160 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 161 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 162 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 163 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 164 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 165 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 166 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 167 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 168 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 169 spdk_trace_register_description("RDMA_REQ_TX_H2C", 170 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 171 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 172 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 173 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 174 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 175 spdk_trace_register_description("RDMA_REQ_EXECUTING", 176 TRACE_RDMA_REQUEST_STATE_EXECUTING, 177 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 178 spdk_trace_register_description("RDMA_REQ_EXECUTED", 179 TRACE_RDMA_REQUEST_STATE_EXECUTED, 180 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 181 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 182 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 183 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 184 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 185 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 186 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 187 spdk_trace_register_description("RDMA_REQ_COMPLETING", 188 TRACE_RDMA_REQUEST_STATE_COMPLETING, 189 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 190 spdk_trace_register_description("RDMA_REQ_COMPLETED", 191 TRACE_RDMA_REQUEST_STATE_COMPLETED, 192 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 193 194 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 195 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 196 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 197 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 198 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 199 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 200 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 201 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 202 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 203 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 204 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 205 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 206 } 207 208 enum spdk_nvmf_rdma_wr_type { 209 RDMA_WR_TYPE_RECV, 210 RDMA_WR_TYPE_SEND, 211 RDMA_WR_TYPE_DATA, 212 }; 213 214 struct spdk_nvmf_rdma_wr { 215 enum spdk_nvmf_rdma_wr_type type; 216 }; 217 218 /* This structure holds commands as they are received off the wire. 219 * It must be dynamically paired with a full request object 220 * (spdk_nvmf_rdma_request) to service a request. It is separate 221 * from the request because RDMA does not appear to order 222 * completions, so occasionally we'll get a new incoming 223 * command when there aren't any free request objects. 224 */ 225 struct spdk_nvmf_rdma_recv { 226 struct ibv_recv_wr wr; 227 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 228 229 struct spdk_nvmf_rdma_qpair *qpair; 230 231 /* In-capsule data buffer */ 232 uint8_t *buf; 233 234 struct spdk_nvmf_rdma_wr rdma_wr; 235 uint64_t receive_tsc; 236 237 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 238 }; 239 240 struct spdk_nvmf_rdma_request_data { 241 struct spdk_nvmf_rdma_wr rdma_wr; 242 struct ibv_send_wr wr; 243 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 244 }; 245 246 struct spdk_nvmf_rdma_request { 247 struct spdk_nvmf_request req; 248 249 enum spdk_nvmf_rdma_request_state state; 250 251 struct spdk_nvmf_rdma_recv *recv; 252 253 struct { 254 struct spdk_nvmf_rdma_wr rdma_wr; 255 struct ibv_send_wr wr; 256 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 257 } rsp; 258 259 struct spdk_nvmf_rdma_request_data data; 260 261 uint32_t iovpos; 262 263 uint32_t num_outstanding_data_wr; 264 uint64_t receive_tsc; 265 266 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 267 }; 268 269 enum spdk_nvmf_rdma_qpair_disconnect_flags { 270 RDMA_QP_DISCONNECTING = 1, 271 RDMA_QP_RECV_DRAINED = 1 << 1, 272 RDMA_QP_SEND_DRAINED = 1 << 2 273 }; 274 275 struct spdk_nvmf_rdma_resource_opts { 276 struct spdk_nvmf_rdma_qpair *qpair; 277 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 278 void *qp; 279 struct ibv_pd *pd; 280 uint32_t max_queue_depth; 281 uint32_t in_capsule_data_size; 282 bool shared; 283 }; 284 285 struct spdk_nvmf_send_wr_list { 286 struct ibv_send_wr *first; 287 struct ibv_send_wr *last; 288 }; 289 290 struct spdk_nvmf_recv_wr_list { 291 struct ibv_recv_wr *first; 292 struct ibv_recv_wr *last; 293 }; 294 295 struct spdk_nvmf_rdma_resources { 296 /* Array of size "max_queue_depth" containing RDMA requests. */ 297 struct spdk_nvmf_rdma_request *reqs; 298 299 /* Array of size "max_queue_depth" containing RDMA recvs. */ 300 struct spdk_nvmf_rdma_recv *recvs; 301 302 /* Array of size "max_queue_depth" containing 64 byte capsules 303 * used for receive. 304 */ 305 union nvmf_h2c_msg *cmds; 306 struct ibv_mr *cmds_mr; 307 308 /* Array of size "max_queue_depth" containing 16 byte completions 309 * to be sent back to the user. 310 */ 311 union nvmf_c2h_msg *cpls; 312 struct ibv_mr *cpls_mr; 313 314 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 315 * buffers to be used for in capsule data. 316 */ 317 void *bufs; 318 struct ibv_mr *bufs_mr; 319 320 /* The list of pending recvs to transfer */ 321 struct spdk_nvmf_recv_wr_list recvs_to_post; 322 323 /* Receives that are waiting for a request object */ 324 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 325 326 /* Queue to track free requests */ 327 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 328 }; 329 330 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 331 332 struct spdk_nvmf_rdma_ibv_event_ctx { 333 struct spdk_nvmf_rdma_qpair *rqpair; 334 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 335 /* Link to other ibv events associated with this qpair */ 336 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 337 }; 338 339 struct spdk_nvmf_rdma_qpair { 340 struct spdk_nvmf_qpair qpair; 341 342 struct spdk_nvmf_rdma_device *device; 343 struct spdk_nvmf_rdma_poller *poller; 344 345 struct spdk_rdma_qp *rdma_qp; 346 struct rdma_cm_id *cm_id; 347 struct ibv_srq *srq; 348 struct rdma_cm_id *listen_id; 349 350 /* The maximum number of I/O outstanding on this connection at one time */ 351 uint16_t max_queue_depth; 352 353 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 354 uint16_t max_read_depth; 355 356 /* The maximum number of RDMA SEND operations at one time */ 357 uint32_t max_send_depth; 358 359 /* The current number of outstanding WRs from this qpair's 360 * recv queue. Should not exceed device->attr.max_queue_depth. 361 */ 362 uint16_t current_recv_depth; 363 364 /* The current number of active RDMA READ operations */ 365 uint16_t current_read_depth; 366 367 /* The current number of posted WRs from this qpair's 368 * send queue. Should not exceed max_send_depth. 369 */ 370 uint32_t current_send_depth; 371 372 /* The maximum number of SGEs per WR on the send queue */ 373 uint32_t max_send_sge; 374 375 /* The maximum number of SGEs per WR on the recv queue */ 376 uint32_t max_recv_sge; 377 378 struct spdk_nvmf_rdma_resources *resources; 379 380 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 381 382 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 383 384 /* Number of requests not in the free state */ 385 uint32_t qd; 386 387 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 388 389 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 390 391 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 392 393 /* IBV queue pair attributes: they are used to manage 394 * qp state and recover from errors. 395 */ 396 enum ibv_qp_state ibv_state; 397 398 uint32_t disconnect_flags; 399 400 /* Poller registered in case the qpair doesn't properly 401 * complete the qpair destruct process and becomes defunct. 402 */ 403 404 struct spdk_poller *destruct_poller; 405 406 /* List of ibv async events */ 407 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 408 409 /* There are several ways a disconnect can start on a qpair 410 * and they are not all mutually exclusive. It is important 411 * that we only initialize one of these paths. 412 */ 413 bool disconnect_started; 414 /* Lets us know that we have received the last_wqe event. */ 415 bool last_wqe_reached; 416 }; 417 418 struct spdk_nvmf_rdma_poller_stat { 419 uint64_t completions; 420 uint64_t polls; 421 uint64_t requests; 422 uint64_t request_latency; 423 uint64_t pending_free_request; 424 uint64_t pending_rdma_read; 425 uint64_t pending_rdma_write; 426 }; 427 428 struct spdk_nvmf_rdma_poller { 429 struct spdk_nvmf_rdma_device *device; 430 struct spdk_nvmf_rdma_poll_group *group; 431 432 int num_cqe; 433 int required_num_wr; 434 struct ibv_cq *cq; 435 436 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 437 uint16_t max_srq_depth; 438 439 /* Shared receive queue */ 440 struct ibv_srq *srq; 441 442 struct spdk_nvmf_rdma_resources *resources; 443 struct spdk_nvmf_rdma_poller_stat stat; 444 445 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 446 447 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 448 449 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 450 451 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 452 }; 453 454 struct spdk_nvmf_rdma_poll_group_stat { 455 uint64_t pending_data_buffer; 456 }; 457 458 struct spdk_nvmf_rdma_poll_group { 459 struct spdk_nvmf_transport_poll_group group; 460 struct spdk_nvmf_rdma_poll_group_stat stat; 461 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 462 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 463 /* 464 * buffers which are split across multiple RDMA 465 * memory regions cannot be used by this transport. 466 */ 467 STAILQ_HEAD(, spdk_nvmf_transport_pg_cache_buf) retired_bufs; 468 }; 469 470 struct spdk_nvmf_rdma_conn_sched { 471 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 472 struct spdk_nvmf_rdma_poll_group *next_io_pg; 473 }; 474 475 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 476 struct spdk_nvmf_rdma_device { 477 struct ibv_device_attr attr; 478 struct ibv_context *context; 479 480 struct spdk_mem_map *map; 481 struct ibv_pd *pd; 482 483 int num_srq; 484 485 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 486 }; 487 488 struct spdk_nvmf_rdma_port { 489 const struct spdk_nvme_transport_id *trid; 490 struct rdma_cm_id *id; 491 struct spdk_nvmf_rdma_device *device; 492 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 493 }; 494 495 struct spdk_nvmf_rdma_transport { 496 struct spdk_nvmf_transport transport; 497 498 struct spdk_nvmf_rdma_conn_sched conn_sched; 499 500 struct rdma_event_channel *event_channel; 501 502 struct spdk_mempool *data_wr_pool; 503 504 pthread_mutex_t lock; 505 506 /* fields used to poll RDMA/IB events */ 507 nfds_t npoll_fds; 508 struct pollfd *poll_fds; 509 510 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 511 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 512 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 513 }; 514 515 static inline void 516 nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair); 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(SPDK_LOG_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(SPDK_LOG_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(SPDK_LOG_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 spdk_poller_unregister(&rqpair->destruct_poller); 849 850 if (rqpair->qd != 0) { 851 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 852 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 853 struct spdk_nvmf_rdma_transport, transport); 854 struct spdk_nvmf_rdma_request *req; 855 uint32_t i, max_req_count = 0; 856 857 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 858 859 if (rqpair->srq == NULL) { 860 nvmf_rdma_dump_qpair_contents(rqpair); 861 max_req_count = rqpair->max_queue_depth; 862 } else if (rqpair->poller && rqpair->resources) { 863 max_req_count = rqpair->poller->max_srq_depth; 864 } 865 866 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Release incomplete requests\n"); 867 for (i = 0; i < max_req_count; i++) { 868 req = &rqpair->resources->reqs[i]; 869 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 870 /* nvmf_rdma_request_process checks qpair ibv and internal state 871 * and completes a request */ 872 nvmf_rdma_request_process(rtransport, req); 873 } 874 } 875 assert(rqpair->qd == 0); 876 } 877 878 if (rqpair->poller) { 879 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 880 881 if (rqpair->srq != NULL && rqpair->resources != NULL) { 882 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 883 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 884 if (rqpair == rdma_recv->qpair) { 885 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 886 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 887 if (rc) { 888 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 889 } 890 } 891 } 892 } 893 } 894 895 if (rqpair->cm_id) { 896 if (rqpair->rdma_qp != NULL) { 897 spdk_rdma_qp_destroy(rqpair->rdma_qp); 898 rqpair->rdma_qp = NULL; 899 } 900 rdma_destroy_id(rqpair->cm_id); 901 902 if (rqpair->poller != NULL && rqpair->srq == NULL) { 903 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 904 } 905 } 906 907 if (rqpair->srq == NULL && rqpair->resources != NULL) { 908 nvmf_rdma_resources_destroy(rqpair->resources); 909 } 910 911 nvmf_rdma_qpair_clean_ibv_events(rqpair); 912 913 free(rqpair); 914 } 915 916 static int 917 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 918 { 919 struct spdk_nvmf_rdma_poller *rpoller; 920 int rc, num_cqe, required_num_wr; 921 922 /* Enlarge CQ size dynamically */ 923 rpoller = rqpair->poller; 924 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 925 num_cqe = rpoller->num_cqe; 926 if (num_cqe < required_num_wr) { 927 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 928 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 929 } 930 931 if (rpoller->num_cqe != num_cqe) { 932 if (required_num_wr > device->attr.max_cqe) { 933 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 934 required_num_wr, device->attr.max_cqe); 935 return -1; 936 } 937 938 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 939 rc = ibv_resize_cq(rpoller->cq, num_cqe); 940 if (rc) { 941 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 942 return -1; 943 } 944 945 rpoller->num_cqe = num_cqe; 946 } 947 948 rpoller->required_num_wr = required_num_wr; 949 return 0; 950 } 951 952 static int 953 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 954 { 955 struct spdk_nvmf_rdma_qpair *rqpair; 956 struct spdk_nvmf_rdma_transport *rtransport; 957 struct spdk_nvmf_transport *transport; 958 struct spdk_nvmf_rdma_resource_opts opts; 959 struct spdk_nvmf_rdma_device *device; 960 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 961 962 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 963 device = rqpair->device; 964 965 qp_init_attr.qp_context = rqpair; 966 qp_init_attr.pd = device->pd; 967 qp_init_attr.send_cq = rqpair->poller->cq; 968 qp_init_attr.recv_cq = rqpair->poller->cq; 969 970 if (rqpair->srq) { 971 qp_init_attr.srq = rqpair->srq; 972 } else { 973 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 974 } 975 976 /* SEND, READ, and WRITE operations */ 977 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 978 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 979 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 980 981 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 982 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 983 goto error; 984 } 985 986 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 987 if (!rqpair->rdma_qp) { 988 goto error; 989 } 990 991 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 992 qp_init_attr.cap.max_send_wr); 993 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 994 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 995 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 996 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 997 998 if (rqpair->poller->srq == NULL) { 999 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1000 transport = &rtransport->transport; 1001 1002 opts.qp = rqpair->rdma_qp->qp; 1003 opts.pd = rqpair->cm_id->pd; 1004 opts.qpair = rqpair; 1005 opts.shared = false; 1006 opts.max_queue_depth = rqpair->max_queue_depth; 1007 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1008 1009 rqpair->resources = nvmf_rdma_resources_create(&opts); 1010 1011 if (!rqpair->resources) { 1012 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1013 rdma_destroy_qp(rqpair->cm_id); 1014 goto error; 1015 } 1016 } else { 1017 rqpair->resources = rqpair->poller->resources; 1018 } 1019 1020 rqpair->current_recv_depth = 0; 1021 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1022 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1023 1024 return 0; 1025 1026 error: 1027 rdma_destroy_id(rqpair->cm_id); 1028 rqpair->cm_id = NULL; 1029 return -1; 1030 } 1031 1032 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1033 /* This function accepts either a single wr or the first wr in a linked list. */ 1034 static void 1035 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1036 { 1037 struct ibv_recv_wr *last; 1038 1039 last = first; 1040 while (last->next != NULL) { 1041 last = last->next; 1042 } 1043 1044 if (rqpair->resources->recvs_to_post.first == NULL) { 1045 rqpair->resources->recvs_to_post.first = first; 1046 rqpair->resources->recvs_to_post.last = last; 1047 if (rqpair->srq == NULL) { 1048 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1049 } 1050 } else { 1051 rqpair->resources->recvs_to_post.last->next = first; 1052 rqpair->resources->recvs_to_post.last = last; 1053 } 1054 } 1055 1056 static int 1057 request_transfer_in(struct spdk_nvmf_request *req) 1058 { 1059 struct spdk_nvmf_rdma_request *rdma_req; 1060 struct spdk_nvmf_qpair *qpair; 1061 struct spdk_nvmf_rdma_qpair *rqpair; 1062 1063 qpair = req->qpair; 1064 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1065 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1066 1067 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1068 assert(rdma_req != NULL); 1069 1070 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1071 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1072 } 1073 1074 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1075 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1076 return 0; 1077 } 1078 1079 static int 1080 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1081 { 1082 int num_outstanding_data_wr = 0; 1083 struct spdk_nvmf_rdma_request *rdma_req; 1084 struct spdk_nvmf_qpair *qpair; 1085 struct spdk_nvmf_rdma_qpair *rqpair; 1086 struct spdk_nvme_cpl *rsp; 1087 struct ibv_send_wr *first = NULL; 1088 1089 *data_posted = 0; 1090 qpair = req->qpair; 1091 rsp = &req->rsp->nvme_cpl; 1092 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1093 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1094 1095 /* Advance our sq_head pointer */ 1096 if (qpair->sq_head == qpair->sq_head_max) { 1097 qpair->sq_head = 0; 1098 } else { 1099 qpair->sq_head++; 1100 } 1101 rsp->sqhd = qpair->sq_head; 1102 1103 /* queue the capsule for the recv buffer */ 1104 assert(rdma_req->recv != NULL); 1105 1106 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1107 1108 rdma_req->recv = NULL; 1109 assert(rqpair->current_recv_depth > 0); 1110 rqpair->current_recv_depth--; 1111 1112 /* Build the response which consists of optional 1113 * RDMA WRITEs to transfer data, plus an RDMA SEND 1114 * containing the response. 1115 */ 1116 first = &rdma_req->rsp.wr; 1117 1118 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1119 /* On failure, data was not read from the controller. So clear the 1120 * number of outstanding data WRs to zero. 1121 */ 1122 rdma_req->num_outstanding_data_wr = 0; 1123 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1124 first = &rdma_req->data.wr; 1125 *data_posted = 1; 1126 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1127 } 1128 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1129 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1130 } 1131 1132 /* +1 for the rsp wr */ 1133 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1134 1135 return 0; 1136 } 1137 1138 static int 1139 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1140 { 1141 struct spdk_nvmf_rdma_accept_private_data accept_data; 1142 struct rdma_conn_param ctrlr_event_data = {}; 1143 int rc; 1144 1145 accept_data.recfmt = 0; 1146 accept_data.crqsize = rqpair->max_queue_depth; 1147 1148 ctrlr_event_data.private_data = &accept_data; 1149 ctrlr_event_data.private_data_len = sizeof(accept_data); 1150 if (id->ps == RDMA_PS_TCP) { 1151 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1152 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1153 } 1154 1155 /* Configure infinite retries for the initiator side qpair. 1156 * When using a shared receive queue on the target side, 1157 * we need to pass this value to the initiator to prevent the 1158 * initiator side NIC from completing SEND requests back to the 1159 * initiator with status rnr_retry_count_exceeded. */ 1160 if (rqpair->srq != NULL) { 1161 ctrlr_event_data.rnr_retry_count = 0x7; 1162 } 1163 1164 /* When qpair is created without use of rdma cm API, an additional 1165 * information must be provided to initiator in the connection response: 1166 * whether qpair is using SRQ and its qp_num 1167 * Fields below are ignored by rdma cm if qpair has been 1168 * created using rdma cm API. */ 1169 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1170 ctrlr_event_data.qp_num = rqpair->rdma_qp->qp->qp_num; 1171 1172 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1173 if (rc) { 1174 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1175 } else { 1176 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1177 } 1178 1179 return rc; 1180 } 1181 1182 static void 1183 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1184 { 1185 struct spdk_nvmf_rdma_reject_private_data rej_data; 1186 1187 rej_data.recfmt = 0; 1188 rej_data.sts = error; 1189 1190 rdma_reject(id, &rej_data, sizeof(rej_data)); 1191 } 1192 1193 static int 1194 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1195 { 1196 struct spdk_nvmf_rdma_transport *rtransport; 1197 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1198 struct spdk_nvmf_rdma_port *port; 1199 struct rdma_conn_param *rdma_param = NULL; 1200 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1201 uint16_t max_queue_depth; 1202 uint16_t max_read_depth; 1203 1204 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1205 1206 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1207 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1208 1209 rdma_param = &event->param.conn; 1210 if (rdma_param->private_data == NULL || 1211 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1212 SPDK_ERRLOG("connect request: no private data provided\n"); 1213 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1214 return -1; 1215 } 1216 1217 private_data = rdma_param->private_data; 1218 if (private_data->recfmt != 0) { 1219 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1220 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1221 return -1; 1222 } 1223 1224 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1225 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1226 1227 port = event->listen_id->context; 1228 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1229 event->listen_id, event->listen_id->verbs, port); 1230 1231 /* Figure out the supported queue depth. This is a multi-step process 1232 * that takes into account hardware maximums, host provided values, 1233 * and our target's internal memory limits */ 1234 1235 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1236 1237 /* Start with the maximum queue depth allowed by the target */ 1238 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1239 max_read_depth = rtransport->transport.opts.max_queue_depth; 1240 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1241 rtransport->transport.opts.max_queue_depth); 1242 1243 /* Next check the local NIC's hardware limitations */ 1244 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1245 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1246 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1247 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1248 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1249 1250 /* Next check the remote NIC's hardware limitations */ 1251 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1252 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1253 rdma_param->initiator_depth, rdma_param->responder_resources); 1254 if (rdma_param->initiator_depth > 0) { 1255 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1256 } 1257 1258 /* Finally check for the host software requested values, which are 1259 * optional. */ 1260 if (rdma_param->private_data != NULL && 1261 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1262 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1263 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1264 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1265 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1266 } 1267 1268 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1269 max_queue_depth, max_read_depth); 1270 1271 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1272 if (rqpair == NULL) { 1273 SPDK_ERRLOG("Could not allocate new connection.\n"); 1274 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1275 return -1; 1276 } 1277 1278 rqpair->device = port->device; 1279 rqpair->max_queue_depth = max_queue_depth; 1280 rqpair->max_read_depth = max_read_depth; 1281 rqpair->cm_id = event->id; 1282 rqpair->listen_id = event->listen_id; 1283 rqpair->qpair.transport = transport; 1284 STAILQ_INIT(&rqpair->ibv_events); 1285 /* use qid from the private data to determine the qpair type 1286 qid will be set to the appropriate value when the controller is created */ 1287 rqpair->qpair.qid = private_data->qid; 1288 1289 event->id->context = &rqpair->qpair; 1290 1291 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1292 1293 return 0; 1294 } 1295 1296 static int 1297 nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1298 enum spdk_mem_map_notify_action action, 1299 void *vaddr, size_t size) 1300 { 1301 struct ibv_pd *pd = cb_ctx; 1302 struct ibv_mr *mr; 1303 int rc; 1304 1305 switch (action) { 1306 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1307 if (!g_nvmf_hooks.get_rkey) { 1308 mr = ibv_reg_mr(pd, vaddr, size, 1309 IBV_ACCESS_LOCAL_WRITE | 1310 IBV_ACCESS_REMOTE_READ | 1311 IBV_ACCESS_REMOTE_WRITE); 1312 if (mr == NULL) { 1313 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1314 return -1; 1315 } else { 1316 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1317 } 1318 } else { 1319 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1320 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1321 } 1322 break; 1323 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1324 if (!g_nvmf_hooks.get_rkey) { 1325 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1326 if (mr) { 1327 ibv_dereg_mr(mr); 1328 } 1329 } 1330 rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1331 break; 1332 default: 1333 SPDK_UNREACHABLE(); 1334 } 1335 1336 return rc; 1337 } 1338 1339 static int 1340 nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1341 { 1342 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1343 return addr_1 == addr_2; 1344 } 1345 1346 static inline void 1347 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1348 enum spdk_nvme_data_transfer xfer) 1349 { 1350 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1351 wr->opcode = IBV_WR_RDMA_WRITE; 1352 wr->send_flags = 0; 1353 wr->next = next; 1354 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1355 wr->opcode = IBV_WR_RDMA_READ; 1356 wr->send_flags = IBV_SEND_SIGNALED; 1357 wr->next = NULL; 1358 } else { 1359 assert(0); 1360 } 1361 } 1362 1363 static int 1364 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1365 struct spdk_nvmf_rdma_request *rdma_req, 1366 uint32_t num_sgl_descriptors) 1367 { 1368 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1369 struct spdk_nvmf_rdma_request_data *current_data_wr; 1370 uint32_t i; 1371 1372 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1373 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1374 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1375 return -EINVAL; 1376 } 1377 1378 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1379 return -ENOMEM; 1380 } 1381 1382 current_data_wr = &rdma_req->data; 1383 1384 for (i = 0; i < num_sgl_descriptors; i++) { 1385 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1386 current_data_wr->wr.next = &work_requests[i]->wr; 1387 current_data_wr = work_requests[i]; 1388 current_data_wr->wr.sg_list = current_data_wr->sgl; 1389 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1390 } 1391 1392 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1393 1394 return 0; 1395 } 1396 1397 static inline void 1398 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1399 { 1400 struct ibv_send_wr *wr = &rdma_req->data.wr; 1401 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1402 1403 wr->wr.rdma.rkey = sgl->keyed.key; 1404 wr->wr.rdma.remote_addr = sgl->address; 1405 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1406 } 1407 1408 static inline void 1409 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1410 { 1411 struct ibv_send_wr *wr = &rdma_req->data.wr; 1412 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1413 uint32_t i; 1414 int j; 1415 uint64_t remote_addr_offset = 0; 1416 1417 for (i = 0; i < num_wrs; ++i) { 1418 wr->wr.rdma.rkey = sgl->keyed.key; 1419 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1420 for (j = 0; j < wr->num_sge; ++j) { 1421 remote_addr_offset += wr->sg_list[j].length; 1422 } 1423 wr = wr->next; 1424 } 1425 } 1426 1427 /* This function is used in the rare case that we have a buffer split over multiple memory regions. */ 1428 static int 1429 nvmf_rdma_replace_buffer(struct spdk_nvmf_rdma_poll_group *rgroup, void **buf) 1430 { 1431 struct spdk_nvmf_transport_poll_group *group = &rgroup->group; 1432 struct spdk_nvmf_transport *transport = group->transport; 1433 struct spdk_nvmf_transport_pg_cache_buf *old_buf; 1434 void *new_buf; 1435 1436 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1437 group->buf_cache_count--; 1438 new_buf = STAILQ_FIRST(&group->buf_cache); 1439 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1440 assert(*buf != NULL); 1441 } else { 1442 new_buf = spdk_mempool_get(transport->data_buf_pool); 1443 } 1444 1445 if (*buf == NULL) { 1446 return -ENOMEM; 1447 } 1448 1449 old_buf = *buf; 1450 STAILQ_INSERT_HEAD(&rgroup->retired_bufs, old_buf, link); 1451 *buf = new_buf; 1452 return 0; 1453 } 1454 1455 static bool 1456 nvmf_rdma_get_lkey(struct spdk_nvmf_rdma_device *device, struct iovec *iov, 1457 uint32_t *_lkey) 1458 { 1459 uint64_t translation_len; 1460 uint32_t lkey; 1461 1462 translation_len = iov->iov_len; 1463 1464 if (!g_nvmf_hooks.get_rkey) { 1465 lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1466 (uint64_t)iov->iov_base, &translation_len))->lkey; 1467 } else { 1468 lkey = spdk_mem_map_translate(device->map, 1469 (uint64_t)iov->iov_base, &translation_len); 1470 } 1471 1472 if (spdk_unlikely(translation_len < iov->iov_len)) { 1473 return false; 1474 } 1475 1476 *_lkey = lkey; 1477 return true; 1478 } 1479 1480 static bool 1481 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device, 1482 struct iovec *iov, struct ibv_send_wr **_wr, 1483 uint32_t *_remaining_data_block, uint32_t *_offset, 1484 uint32_t *_num_extra_wrs, 1485 const struct spdk_dif_ctx *dif_ctx) 1486 { 1487 struct ibv_send_wr *wr = *_wr; 1488 struct ibv_sge *sg_ele = &wr->sg_list[wr->num_sge]; 1489 uint32_t lkey = 0; 1490 uint32_t remaining, data_block_size, md_size, sge_len; 1491 1492 if (spdk_unlikely(!nvmf_rdma_get_lkey(device, iov, &lkey))) { 1493 /* This is a very rare case that can occur when using DPDK version < 19.05 */ 1494 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions. Removing it from circulation.\n"); 1495 return false; 1496 } 1497 1498 if (spdk_likely(!dif_ctx)) { 1499 sg_ele->lkey = lkey; 1500 sg_ele->addr = (uintptr_t)(iov->iov_base); 1501 sg_ele->length = iov->iov_len; 1502 wr->num_sge++; 1503 } else { 1504 remaining = iov->iov_len - *_offset; 1505 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1506 md_size = dif_ctx->md_size; 1507 1508 while (remaining) { 1509 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1510 if (*_num_extra_wrs > 0 && wr->next) { 1511 *_wr = wr->next; 1512 wr = *_wr; 1513 wr->num_sge = 0; 1514 sg_ele = &wr->sg_list[wr->num_sge]; 1515 (*_num_extra_wrs)--; 1516 } else { 1517 break; 1518 } 1519 } 1520 sg_ele->lkey = lkey; 1521 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset); 1522 sge_len = spdk_min(remaining, *_remaining_data_block); 1523 sg_ele->length = sge_len; 1524 remaining -= sge_len; 1525 *_remaining_data_block -= sge_len; 1526 *_offset += sge_len; 1527 1528 sg_ele++; 1529 wr->num_sge++; 1530 1531 if (*_remaining_data_block == 0) { 1532 /* skip metadata */ 1533 *_offset += md_size; 1534 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1535 remaining -= spdk_min(remaining, md_size); 1536 *_remaining_data_block = data_block_size; 1537 } 1538 1539 if (remaining == 0) { 1540 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1541 the remaining metadata bits at the beginning of the next buffer */ 1542 *_offset -= iov->iov_len; 1543 } 1544 } 1545 } 1546 1547 return true; 1548 } 1549 1550 static int 1551 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1552 struct spdk_nvmf_rdma_device *device, 1553 struct spdk_nvmf_rdma_request *rdma_req, 1554 struct ibv_send_wr *wr, 1555 uint32_t length, 1556 uint32_t num_extra_wrs) 1557 { 1558 struct spdk_nvmf_request *req = &rdma_req->req; 1559 struct spdk_dif_ctx *dif_ctx = NULL; 1560 uint32_t remaining_data_block = 0; 1561 uint32_t offset = 0; 1562 1563 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 1564 dif_ctx = &rdma_req->req.dif.dif_ctx; 1565 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1566 } 1567 1568 wr->num_sge = 0; 1569 1570 while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1571 while (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr, 1572 &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) { 1573 if (nvmf_rdma_replace_buffer(rgroup, &req->buffers[rdma_req->iovpos]) == -ENOMEM) { 1574 return -ENOMEM; 1575 } 1576 req->iov[rdma_req->iovpos].iov_base = (void *)((uintptr_t)(req->buffers[rdma_req->iovpos] + 1577 NVMF_DATA_BUFFER_MASK) & 1578 ~NVMF_DATA_BUFFER_MASK); 1579 } 1580 1581 length -= req->iov[rdma_req->iovpos].iov_len; 1582 rdma_req->iovpos++; 1583 } 1584 1585 if (length) { 1586 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1587 return -EINVAL; 1588 } 1589 1590 return 0; 1591 } 1592 1593 static inline uint32_t 1594 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1595 { 1596 /* estimate the number of SG entries and WRs needed to process the request */ 1597 uint32_t num_sge = 0; 1598 uint32_t i; 1599 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1600 1601 for (i = 0; i < num_buffers && length > 0; i++) { 1602 uint32_t buffer_len = spdk_min(length, io_unit_size); 1603 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1604 1605 if (num_sge_in_block * block_size > buffer_len) { 1606 ++num_sge_in_block; 1607 } 1608 num_sge += num_sge_in_block; 1609 length -= buffer_len; 1610 } 1611 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1612 } 1613 1614 static int 1615 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1616 struct spdk_nvmf_rdma_device *device, 1617 struct spdk_nvmf_rdma_request *rdma_req, 1618 uint32_t length) 1619 { 1620 struct spdk_nvmf_rdma_qpair *rqpair; 1621 struct spdk_nvmf_rdma_poll_group *rgroup; 1622 struct spdk_nvmf_request *req = &rdma_req->req; 1623 struct ibv_send_wr *wr = &rdma_req->data.wr; 1624 int rc; 1625 uint32_t num_wrs = 1; 1626 1627 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1628 rgroup = rqpair->poller->group; 1629 1630 /* rdma wr specifics */ 1631 nvmf_rdma_setup_request(rdma_req); 1632 1633 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1634 length); 1635 if (rc != 0) { 1636 return rc; 1637 } 1638 1639 assert(req->iovcnt <= rqpair->max_send_sge); 1640 1641 rdma_req->iovpos = 0; 1642 1643 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1644 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1645 req->dif.dif_ctx.block_size); 1646 if (num_wrs > 1) { 1647 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1648 if (rc != 0) { 1649 goto err_exit; 1650 } 1651 } 1652 } 1653 1654 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1655 if (spdk_unlikely(rc != 0)) { 1656 goto err_exit; 1657 } 1658 1659 if (spdk_unlikely(num_wrs > 1)) { 1660 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1661 } 1662 1663 /* set the number of outstanding data WRs for this request. */ 1664 rdma_req->num_outstanding_data_wr = num_wrs; 1665 1666 return rc; 1667 1668 err_exit: 1669 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1670 nvmf_rdma_request_free_data(rdma_req, rtransport); 1671 req->iovcnt = 0; 1672 return rc; 1673 } 1674 1675 static int 1676 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1677 struct spdk_nvmf_rdma_device *device, 1678 struct spdk_nvmf_rdma_request *rdma_req) 1679 { 1680 struct spdk_nvmf_rdma_qpair *rqpair; 1681 struct spdk_nvmf_rdma_poll_group *rgroup; 1682 struct ibv_send_wr *current_wr; 1683 struct spdk_nvmf_request *req = &rdma_req->req; 1684 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1685 uint32_t num_sgl_descriptors; 1686 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES]; 1687 uint32_t i; 1688 int rc; 1689 1690 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1691 rgroup = rqpair->poller->group; 1692 1693 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1694 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1695 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1696 1697 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1698 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1699 1700 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1701 return -ENOMEM; 1702 } 1703 1704 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1705 for (i = 0; i < num_sgl_descriptors; i++) { 1706 if (spdk_likely(!req->dif.dif_insert_or_strip)) { 1707 lengths[i] = desc->keyed.length; 1708 } else { 1709 req->dif.orig_length += desc->keyed.length; 1710 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1711 req->dif.elba_length += lengths[i]; 1712 } 1713 desc++; 1714 } 1715 1716 rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport, 1717 lengths, num_sgl_descriptors); 1718 if (rc != 0) { 1719 nvmf_rdma_request_free_data(rdma_req, rtransport); 1720 return rc; 1721 } 1722 1723 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1724 current_wr = &rdma_req->data.wr; 1725 assert(current_wr != NULL); 1726 1727 req->length = 0; 1728 rdma_req->iovpos = 0; 1729 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1730 for (i = 0; i < num_sgl_descriptors; i++) { 1731 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1732 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1733 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1734 rc = -EINVAL; 1735 goto err_exit; 1736 } 1737 1738 current_wr->num_sge = 0; 1739 1740 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1741 if (rc != 0) { 1742 rc = -ENOMEM; 1743 goto err_exit; 1744 } 1745 1746 req->length += desc->keyed.length; 1747 current_wr->wr.rdma.rkey = desc->keyed.key; 1748 current_wr->wr.rdma.remote_addr = desc->address; 1749 current_wr = current_wr->next; 1750 desc++; 1751 } 1752 1753 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1754 /* Go back to the last descriptor in the list. */ 1755 desc--; 1756 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1757 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1758 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1759 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1760 } 1761 } 1762 #endif 1763 1764 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1765 1766 return 0; 1767 1768 err_exit: 1769 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1770 nvmf_rdma_request_free_data(rdma_req, rtransport); 1771 return rc; 1772 } 1773 1774 static int 1775 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1776 struct spdk_nvmf_rdma_device *device, 1777 struct spdk_nvmf_rdma_request *rdma_req) 1778 { 1779 struct spdk_nvmf_request *req = &rdma_req->req; 1780 struct spdk_nvme_cpl *rsp; 1781 struct spdk_nvme_sgl_descriptor *sgl; 1782 int rc; 1783 uint32_t length; 1784 1785 rsp = &req->rsp->nvme_cpl; 1786 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1787 1788 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1789 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1790 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1791 1792 length = sgl->keyed.length; 1793 if (length > rtransport->transport.opts.max_io_size) { 1794 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1795 length, rtransport->transport.opts.max_io_size); 1796 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1797 return -1; 1798 } 1799 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1800 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1801 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1802 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1803 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1804 } 1805 } 1806 #endif 1807 1808 /* fill request length and populate iovs */ 1809 req->length = length; 1810 1811 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1812 req->dif.orig_length = length; 1813 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1814 req->dif.elba_length = length; 1815 } 1816 1817 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1818 if (spdk_unlikely(rc < 0)) { 1819 if (rc == -EINVAL) { 1820 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1821 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1822 return -1; 1823 } 1824 /* No available buffers. Queue this request up. */ 1825 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1826 return 0; 1827 } 1828 1829 /* backward compatible */ 1830 req->data = req->iov[0].iov_base; 1831 1832 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1833 req->iovcnt); 1834 1835 return 0; 1836 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1837 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1838 uint64_t offset = sgl->address; 1839 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1840 1841 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1842 offset, sgl->unkeyed.length); 1843 1844 if (offset > max_len) { 1845 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1846 offset, max_len); 1847 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1848 return -1; 1849 } 1850 max_len -= (uint32_t)offset; 1851 1852 if (sgl->unkeyed.length > max_len) { 1853 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1854 sgl->unkeyed.length, max_len); 1855 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1856 return -1; 1857 } 1858 1859 rdma_req->num_outstanding_data_wr = 0; 1860 req->data = rdma_req->recv->buf + offset; 1861 req->data_from_pool = false; 1862 req->length = sgl->unkeyed.length; 1863 1864 req->iov[0].iov_base = req->data; 1865 req->iov[0].iov_len = req->length; 1866 req->iovcnt = 1; 1867 1868 return 0; 1869 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1870 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1871 1872 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1873 if (rc == -ENOMEM) { 1874 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1875 return 0; 1876 } else if (rc == -EINVAL) { 1877 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1878 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1879 return -1; 1880 } 1881 1882 /* backward compatible */ 1883 req->data = req->iov[0].iov_base; 1884 1885 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1886 req->iovcnt); 1887 1888 return 0; 1889 } 1890 1891 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1892 sgl->generic.type, sgl->generic.subtype); 1893 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1894 return -1; 1895 } 1896 1897 static void 1898 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1899 struct spdk_nvmf_rdma_transport *rtransport) 1900 { 1901 struct spdk_nvmf_rdma_qpair *rqpair; 1902 struct spdk_nvmf_rdma_poll_group *rgroup; 1903 1904 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1905 if (rdma_req->req.data_from_pool) { 1906 rgroup = rqpair->poller->group; 1907 1908 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1909 } 1910 nvmf_rdma_request_free_data(rdma_req, rtransport); 1911 rdma_req->req.length = 0; 1912 rdma_req->req.iovcnt = 0; 1913 rdma_req->req.data = NULL; 1914 rdma_req->rsp.wr.next = NULL; 1915 rdma_req->data.wr.next = NULL; 1916 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1917 rqpair->qd--; 1918 1919 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1920 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1921 } 1922 1923 bool 1924 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1925 struct spdk_nvmf_rdma_request *rdma_req) 1926 { 1927 struct spdk_nvmf_rdma_qpair *rqpair; 1928 struct spdk_nvmf_rdma_device *device; 1929 struct spdk_nvmf_rdma_poll_group *rgroup; 1930 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1931 int rc; 1932 struct spdk_nvmf_rdma_recv *rdma_recv; 1933 enum spdk_nvmf_rdma_request_state prev_state; 1934 bool progress = false; 1935 int data_posted; 1936 uint32_t num_blocks; 1937 1938 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1939 device = rqpair->device; 1940 rgroup = rqpair->poller->group; 1941 1942 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1943 1944 /* If the queue pair is in an error state, force the request to the completed state 1945 * to release resources. */ 1946 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1947 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1948 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1949 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1950 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1951 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1952 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1953 } 1954 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1955 } 1956 1957 /* The loop here is to allow for several back-to-back state changes. */ 1958 do { 1959 prev_state = rdma_req->state; 1960 1961 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1962 1963 switch (rdma_req->state) { 1964 case RDMA_REQUEST_STATE_FREE: 1965 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1966 * to escape this state. */ 1967 break; 1968 case RDMA_REQUEST_STATE_NEW: 1969 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1970 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1971 rdma_recv = rdma_req->recv; 1972 1973 /* The first element of the SGL is the NVMe command */ 1974 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1975 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1976 1977 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1978 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1979 break; 1980 } 1981 1982 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 1983 rdma_req->req.dif.dif_insert_or_strip = true; 1984 } 1985 1986 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1987 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1988 rdma_req->rsp.wr.imm_data = 0; 1989 #endif 1990 1991 /* The next state transition depends on the data transfer needs of this request. */ 1992 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 1993 1994 /* If no data to transfer, ready to execute. */ 1995 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1996 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1997 break; 1998 } 1999 2000 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2001 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2002 break; 2003 case RDMA_REQUEST_STATE_NEED_BUFFER: 2004 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2005 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2006 2007 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2008 2009 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2010 /* This request needs to wait in line to obtain a buffer */ 2011 break; 2012 } 2013 2014 /* Try to get a data buffer */ 2015 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2016 if (rc < 0) { 2017 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2018 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2019 break; 2020 } 2021 2022 if (!rdma_req->req.data) { 2023 /* No buffers available. */ 2024 rgroup->stat.pending_data_buffer++; 2025 break; 2026 } 2027 2028 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2029 2030 /* If data is transferring from host to controller and the data didn't 2031 * arrive using in capsule data, we need to do a transfer from the host. 2032 */ 2033 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2034 rdma_req->req.data_from_pool) { 2035 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2036 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2037 break; 2038 } 2039 2040 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2041 break; 2042 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2043 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2044 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2045 2046 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2047 /* This request needs to wait in line to perform RDMA */ 2048 break; 2049 } 2050 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2051 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2052 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2053 rqpair->poller->stat.pending_rdma_read++; 2054 break; 2055 } 2056 2057 /* We have already verified that this request is the head of the queue. */ 2058 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2059 2060 rc = request_transfer_in(&rdma_req->req); 2061 if (!rc) { 2062 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2063 } else { 2064 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2065 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2066 } 2067 break; 2068 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2069 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2070 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2071 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2072 * to escape this state. */ 2073 break; 2074 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2075 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2076 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2077 2078 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2079 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2080 /* generate DIF for write operation */ 2081 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2082 assert(num_blocks > 0); 2083 2084 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2085 num_blocks, &rdma_req->req.dif.dif_ctx); 2086 if (rc != 0) { 2087 SPDK_ERRLOG("DIF generation failed\n"); 2088 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2089 nvmf_rdma_start_disconnect(rqpair); 2090 break; 2091 } 2092 } 2093 2094 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2095 /* set extended length before IO operation */ 2096 rdma_req->req.length = rdma_req->req.dif.elba_length; 2097 } 2098 2099 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2100 spdk_nvmf_request_exec(&rdma_req->req); 2101 break; 2102 case RDMA_REQUEST_STATE_EXECUTING: 2103 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2104 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2105 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2106 * to escape this state. */ 2107 break; 2108 case RDMA_REQUEST_STATE_EXECUTED: 2109 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2110 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2111 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2112 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2113 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2114 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2115 } else { 2116 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2117 } 2118 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2119 /* restore the original length */ 2120 rdma_req->req.length = rdma_req->req.dif.orig_length; 2121 2122 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2123 struct spdk_dif_error error_blk; 2124 2125 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2126 2127 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2128 &rdma_req->req.dif.dif_ctx, &error_blk); 2129 if (rc) { 2130 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2131 2132 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2133 error_blk.err_offset); 2134 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2135 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2136 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2137 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2138 } 2139 } 2140 } 2141 break; 2142 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2143 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2144 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2145 2146 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2147 /* This request needs to wait in line to perform RDMA */ 2148 break; 2149 } 2150 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2151 rqpair->max_send_depth) { 2152 /* We can only have so many WRs outstanding. we have to wait until some finish. 2153 * +1 since each request has an additional wr in the resp. */ 2154 rqpair->poller->stat.pending_rdma_write++; 2155 break; 2156 } 2157 2158 /* We have already verified that this request is the head of the queue. */ 2159 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2160 2161 /* The data transfer will be kicked off from 2162 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2163 */ 2164 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2165 break; 2166 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2167 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2168 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2169 rc = request_transfer_out(&rdma_req->req, &data_posted); 2170 assert(rc == 0); /* No good way to handle this currently */ 2171 if (rc) { 2172 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2173 } else { 2174 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2175 RDMA_REQUEST_STATE_COMPLETING; 2176 } 2177 break; 2178 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2179 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2180 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2181 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2182 * to escape this state. */ 2183 break; 2184 case RDMA_REQUEST_STATE_COMPLETING: 2185 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2186 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2187 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2188 * to escape this state. */ 2189 break; 2190 case RDMA_REQUEST_STATE_COMPLETED: 2191 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2192 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2193 2194 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2195 _nvmf_rdma_request_free(rdma_req, rtransport); 2196 break; 2197 case RDMA_REQUEST_NUM_STATES: 2198 default: 2199 assert(0); 2200 break; 2201 } 2202 2203 if (rdma_req->state != prev_state) { 2204 progress = true; 2205 } 2206 } while (rdma_req->state != prev_state); 2207 2208 return progress; 2209 } 2210 2211 /* Public API callbacks begin here */ 2212 2213 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2214 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2215 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2216 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2217 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2218 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2219 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2220 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2221 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2222 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2223 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2224 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2225 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2226 2227 static void 2228 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2229 { 2230 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2231 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2232 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2233 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2234 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2235 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2236 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2237 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2238 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2239 opts->no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2240 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2241 opts->acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2242 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2243 } 2244 2245 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2246 .notify_cb = nvmf_rdma_mem_notify, 2247 .are_contiguous = nvmf_rdma_check_contiguous_entries 2248 }; 2249 2250 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2251 2252 static struct spdk_nvmf_transport * 2253 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2254 { 2255 int rc; 2256 struct spdk_nvmf_rdma_transport *rtransport; 2257 struct spdk_nvmf_rdma_device *device, *tmp; 2258 struct ibv_context **contexts; 2259 uint32_t i; 2260 int flag; 2261 uint32_t sge_count; 2262 uint32_t min_shared_buffers; 2263 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2264 pthread_mutexattr_t attr; 2265 2266 rtransport = calloc(1, sizeof(*rtransport)); 2267 if (!rtransport) { 2268 return NULL; 2269 } 2270 2271 if (pthread_mutexattr_init(&attr)) { 2272 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2273 free(rtransport); 2274 return NULL; 2275 } 2276 2277 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2278 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2279 pthread_mutexattr_destroy(&attr); 2280 free(rtransport); 2281 return NULL; 2282 } 2283 2284 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2285 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2286 pthread_mutexattr_destroy(&attr); 2287 free(rtransport); 2288 return NULL; 2289 } 2290 2291 pthread_mutexattr_destroy(&attr); 2292 2293 TAILQ_INIT(&rtransport->devices); 2294 TAILQ_INIT(&rtransport->ports); 2295 TAILQ_INIT(&rtransport->poll_groups); 2296 2297 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2298 2299 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2300 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2301 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2302 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2303 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d," 2304 " acceptor_backlog=%d, abort_timeout_sec=%d\n", 2305 opts->max_queue_depth, 2306 opts->max_io_size, 2307 opts->max_qpairs_per_ctrlr - 1, 2308 opts->io_unit_size, 2309 opts->in_capsule_data_size, 2310 opts->max_aq_depth, 2311 opts->num_shared_buffers, 2312 opts->max_srq_depth, 2313 opts->no_srq, 2314 opts->acceptor_backlog, 2315 opts->abort_timeout_sec); 2316 2317 /* I/O unit size cannot be larger than max I/O size */ 2318 if (opts->io_unit_size > opts->max_io_size) { 2319 opts->io_unit_size = opts->max_io_size; 2320 } 2321 2322 if (opts->acceptor_backlog <= 0) { 2323 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2324 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2325 opts->acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2326 } 2327 2328 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2329 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2330 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2331 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2332 nvmf_rdma_destroy(&rtransport->transport); 2333 return NULL; 2334 } 2335 2336 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2337 if (min_shared_buffers > opts->num_shared_buffers) { 2338 SPDK_ERRLOG("There are not enough buffers to satisfy" 2339 "per-poll group caches for each thread. (%" PRIu32 ")" 2340 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2341 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2342 nvmf_rdma_destroy(&rtransport->transport); 2343 return NULL; 2344 } 2345 2346 sge_count = opts->max_io_size / opts->io_unit_size; 2347 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2348 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2349 nvmf_rdma_destroy(&rtransport->transport); 2350 return NULL; 2351 } 2352 2353 rtransport->event_channel = rdma_create_event_channel(); 2354 if (rtransport->event_channel == NULL) { 2355 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2356 nvmf_rdma_destroy(&rtransport->transport); 2357 return NULL; 2358 } 2359 2360 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2361 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2362 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2363 rtransport->event_channel->fd, spdk_strerror(errno)); 2364 nvmf_rdma_destroy(&rtransport->transport); 2365 return NULL; 2366 } 2367 2368 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2369 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2370 sizeof(struct spdk_nvmf_rdma_request_data), 2371 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2372 SPDK_ENV_SOCKET_ID_ANY); 2373 if (!rtransport->data_wr_pool) { 2374 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2375 nvmf_rdma_destroy(&rtransport->transport); 2376 return NULL; 2377 } 2378 2379 contexts = rdma_get_devices(NULL); 2380 if (contexts == NULL) { 2381 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2382 nvmf_rdma_destroy(&rtransport->transport); 2383 return NULL; 2384 } 2385 2386 i = 0; 2387 rc = 0; 2388 while (contexts[i] != NULL) { 2389 device = calloc(1, sizeof(*device)); 2390 if (!device) { 2391 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2392 rc = -ENOMEM; 2393 break; 2394 } 2395 device->context = contexts[i]; 2396 rc = ibv_query_device(device->context, &device->attr); 2397 if (rc < 0) { 2398 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2399 free(device); 2400 break; 2401 2402 } 2403 2404 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2405 2406 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2407 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2408 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2409 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2410 } 2411 2412 /** 2413 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2414 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2415 * but incorrectly reports that it does. There are changes making their way 2416 * through the kernel now that will enable this feature. When they are merged, 2417 * we can conditionally enable this feature. 2418 * 2419 * TODO: enable this for versions of the kernel rxe driver that support it. 2420 */ 2421 if (device->attr.vendor_id == 0) { 2422 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2423 } 2424 #endif 2425 2426 /* set up device context async ev fd as NON_BLOCKING */ 2427 flag = fcntl(device->context->async_fd, F_GETFL); 2428 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2429 if (rc < 0) { 2430 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2431 free(device); 2432 break; 2433 } 2434 2435 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2436 i++; 2437 2438 if (g_nvmf_hooks.get_ibv_pd) { 2439 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2440 } else { 2441 device->pd = ibv_alloc_pd(device->context); 2442 } 2443 2444 if (!device->pd) { 2445 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2446 rc = -ENOMEM; 2447 break; 2448 } 2449 2450 assert(device->map == NULL); 2451 2452 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2453 if (!device->map) { 2454 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2455 rc = -ENOMEM; 2456 break; 2457 } 2458 2459 assert(device->map != NULL); 2460 assert(device->pd != NULL); 2461 } 2462 rdma_free_devices(contexts); 2463 2464 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2465 /* divide and round up. */ 2466 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2467 2468 /* round up to the nearest 4k. */ 2469 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2470 2471 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2472 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2473 opts->io_unit_size); 2474 } 2475 2476 if (rc < 0) { 2477 nvmf_rdma_destroy(&rtransport->transport); 2478 return NULL; 2479 } 2480 2481 /* Set up poll descriptor array to monitor events from RDMA and IB 2482 * in a single poll syscall 2483 */ 2484 rtransport->npoll_fds = i + 1; 2485 i = 0; 2486 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2487 if (rtransport->poll_fds == NULL) { 2488 SPDK_ERRLOG("poll_fds allocation failed\n"); 2489 nvmf_rdma_destroy(&rtransport->transport); 2490 return NULL; 2491 } 2492 2493 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2494 rtransport->poll_fds[i++].events = POLLIN; 2495 2496 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2497 rtransport->poll_fds[i].fd = device->context->async_fd; 2498 rtransport->poll_fds[i++].events = POLLIN; 2499 } 2500 2501 return &rtransport->transport; 2502 } 2503 2504 static int 2505 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2506 { 2507 struct spdk_nvmf_rdma_transport *rtransport; 2508 struct spdk_nvmf_rdma_port *port, *port_tmp; 2509 struct spdk_nvmf_rdma_device *device, *device_tmp; 2510 2511 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2512 2513 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2514 TAILQ_REMOVE(&rtransport->ports, port, link); 2515 rdma_destroy_id(port->id); 2516 free(port); 2517 } 2518 2519 if (rtransport->poll_fds != NULL) { 2520 free(rtransport->poll_fds); 2521 } 2522 2523 if (rtransport->event_channel != NULL) { 2524 rdma_destroy_event_channel(rtransport->event_channel); 2525 } 2526 2527 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2528 TAILQ_REMOVE(&rtransport->devices, device, link); 2529 if (device->map) { 2530 spdk_mem_map_free(&device->map); 2531 } 2532 if (device->pd) { 2533 if (!g_nvmf_hooks.get_ibv_pd) { 2534 ibv_dealloc_pd(device->pd); 2535 } 2536 } 2537 free(device); 2538 } 2539 2540 if (rtransport->data_wr_pool != NULL) { 2541 if (spdk_mempool_count(rtransport->data_wr_pool) != 2542 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2543 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2544 spdk_mempool_count(rtransport->data_wr_pool), 2545 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2546 } 2547 } 2548 2549 spdk_mempool_free(rtransport->data_wr_pool); 2550 2551 pthread_mutex_destroy(&rtransport->lock); 2552 free(rtransport); 2553 2554 return 0; 2555 } 2556 2557 static int 2558 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2559 struct spdk_nvme_transport_id *trid, 2560 bool peer); 2561 2562 static int 2563 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2564 const struct spdk_nvme_transport_id *trid) 2565 { 2566 struct spdk_nvmf_rdma_transport *rtransport; 2567 struct spdk_nvmf_rdma_device *device; 2568 struct spdk_nvmf_rdma_port *port; 2569 struct addrinfo *res; 2570 struct addrinfo hints; 2571 int family; 2572 int rc; 2573 2574 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2575 assert(rtransport->event_channel != NULL); 2576 2577 pthread_mutex_lock(&rtransport->lock); 2578 port = calloc(1, sizeof(*port)); 2579 if (!port) { 2580 SPDK_ERRLOG("Port allocation failed\n"); 2581 pthread_mutex_unlock(&rtransport->lock); 2582 return -ENOMEM; 2583 } 2584 2585 port->trid = trid; 2586 2587 switch (trid->adrfam) { 2588 case SPDK_NVMF_ADRFAM_IPV4: 2589 family = AF_INET; 2590 break; 2591 case SPDK_NVMF_ADRFAM_IPV6: 2592 family = AF_INET6; 2593 break; 2594 default: 2595 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2596 free(port); 2597 pthread_mutex_unlock(&rtransport->lock); 2598 return -EINVAL; 2599 } 2600 2601 memset(&hints, 0, sizeof(hints)); 2602 hints.ai_family = family; 2603 hints.ai_flags = AI_NUMERICSERV; 2604 hints.ai_socktype = SOCK_STREAM; 2605 hints.ai_protocol = 0; 2606 2607 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2608 if (rc) { 2609 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2610 free(port); 2611 pthread_mutex_unlock(&rtransport->lock); 2612 return -EINVAL; 2613 } 2614 2615 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2616 if (rc < 0) { 2617 SPDK_ERRLOG("rdma_create_id() failed\n"); 2618 freeaddrinfo(res); 2619 free(port); 2620 pthread_mutex_unlock(&rtransport->lock); 2621 return rc; 2622 } 2623 2624 rc = rdma_bind_addr(port->id, res->ai_addr); 2625 freeaddrinfo(res); 2626 2627 if (rc < 0) { 2628 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2629 rdma_destroy_id(port->id); 2630 free(port); 2631 pthread_mutex_unlock(&rtransport->lock); 2632 return rc; 2633 } 2634 2635 if (!port->id->verbs) { 2636 SPDK_ERRLOG("ibv_context is null\n"); 2637 rdma_destroy_id(port->id); 2638 free(port); 2639 pthread_mutex_unlock(&rtransport->lock); 2640 return -1; 2641 } 2642 2643 rc = rdma_listen(port->id, transport->opts.acceptor_backlog); 2644 if (rc < 0) { 2645 SPDK_ERRLOG("rdma_listen() failed\n"); 2646 rdma_destroy_id(port->id); 2647 free(port); 2648 pthread_mutex_unlock(&rtransport->lock); 2649 return rc; 2650 } 2651 2652 TAILQ_FOREACH(device, &rtransport->devices, link) { 2653 if (device->context == port->id->verbs) { 2654 port->device = device; 2655 break; 2656 } 2657 } 2658 if (!port->device) { 2659 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2660 port->id->verbs); 2661 rdma_destroy_id(port->id); 2662 free(port); 2663 pthread_mutex_unlock(&rtransport->lock); 2664 return -EINVAL; 2665 } 2666 2667 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2668 trid->traddr, trid->trsvcid); 2669 2670 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2671 pthread_mutex_unlock(&rtransport->lock); 2672 return 0; 2673 } 2674 2675 static void 2676 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2677 const struct spdk_nvme_transport_id *trid) 2678 { 2679 struct spdk_nvmf_rdma_transport *rtransport; 2680 struct spdk_nvmf_rdma_port *port, *tmp; 2681 2682 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2683 2684 pthread_mutex_lock(&rtransport->lock); 2685 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2686 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2687 TAILQ_REMOVE(&rtransport->ports, port, link); 2688 rdma_destroy_id(port->id); 2689 free(port); 2690 break; 2691 } 2692 } 2693 2694 pthread_mutex_unlock(&rtransport->lock); 2695 } 2696 2697 static void 2698 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2699 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2700 { 2701 struct spdk_nvmf_request *req, *tmp; 2702 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2703 struct spdk_nvmf_rdma_resources *resources; 2704 2705 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2706 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2707 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2708 break; 2709 } 2710 } 2711 2712 /* Then RDMA writes since reads have stronger restrictions than writes */ 2713 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2714 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2715 break; 2716 } 2717 } 2718 2719 /* The second highest priority is I/O waiting on memory buffers. */ 2720 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2721 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2722 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2723 break; 2724 } 2725 } 2726 2727 resources = rqpair->resources; 2728 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2729 rdma_req = STAILQ_FIRST(&resources->free_queue); 2730 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2731 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2732 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2733 2734 if (rqpair->srq != NULL) { 2735 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2736 rdma_req->recv->qpair->qd++; 2737 } else { 2738 rqpair->qd++; 2739 } 2740 2741 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2742 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2743 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2744 break; 2745 } 2746 } 2747 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2748 rqpair->poller->stat.pending_free_request++; 2749 } 2750 } 2751 2752 static void 2753 _nvmf_rdma_qpair_disconnect(void *ctx) 2754 { 2755 struct spdk_nvmf_qpair *qpair = ctx; 2756 2757 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2758 } 2759 2760 static void 2761 _nvmf_rdma_try_disconnect(void *ctx) 2762 { 2763 struct spdk_nvmf_qpair *qpair = ctx; 2764 struct spdk_nvmf_poll_group *group; 2765 2766 /* Read the group out of the qpair. This is normally set and accessed only from 2767 * the thread that created the group. Here, we're not on that thread necessarily. 2768 * The data member qpair->group begins it's life as NULL and then is assigned to 2769 * a pointer and never changes. So fortunately reading this and checking for 2770 * non-NULL is thread safe in the x86_64 memory model. */ 2771 group = qpair->group; 2772 2773 if (group == NULL) { 2774 /* The qpair hasn't been assigned to a group yet, so we can't 2775 * process a disconnect. Send a message to ourself and try again. */ 2776 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2777 return; 2778 } 2779 2780 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2781 } 2782 2783 static inline void 2784 nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2785 { 2786 if (!__atomic_test_and_set(&rqpair->disconnect_started, __ATOMIC_RELAXED)) { 2787 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2788 } 2789 } 2790 2791 static void nvmf_rdma_destroy_drained_qpair(void *ctx) 2792 { 2793 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2794 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2795 struct spdk_nvmf_rdma_transport, transport); 2796 2797 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2798 if (rqpair->current_send_depth != 0) { 2799 return; 2800 } 2801 2802 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2803 return; 2804 } 2805 2806 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) { 2807 return; 2808 } 2809 2810 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2811 2812 /* Qpair will be destroyed after nvmf layer closes this qpair */ 2813 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ERROR) { 2814 return; 2815 } 2816 2817 nvmf_rdma_qpair_destroy(rqpair); 2818 } 2819 2820 2821 static int 2822 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2823 { 2824 struct spdk_nvmf_qpair *qpair; 2825 struct spdk_nvmf_rdma_qpair *rqpair; 2826 2827 if (evt->id == NULL) { 2828 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2829 return -1; 2830 } 2831 2832 qpair = evt->id->context; 2833 if (qpair == NULL) { 2834 SPDK_ERRLOG("disconnect request: no active connection\n"); 2835 return -1; 2836 } 2837 2838 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2839 2840 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2841 2842 nvmf_rdma_start_disconnect(rqpair); 2843 2844 return 0; 2845 } 2846 2847 #ifdef DEBUG 2848 static const char *CM_EVENT_STR[] = { 2849 "RDMA_CM_EVENT_ADDR_RESOLVED", 2850 "RDMA_CM_EVENT_ADDR_ERROR", 2851 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2852 "RDMA_CM_EVENT_ROUTE_ERROR", 2853 "RDMA_CM_EVENT_CONNECT_REQUEST", 2854 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2855 "RDMA_CM_EVENT_CONNECT_ERROR", 2856 "RDMA_CM_EVENT_UNREACHABLE", 2857 "RDMA_CM_EVENT_REJECTED", 2858 "RDMA_CM_EVENT_ESTABLISHED", 2859 "RDMA_CM_EVENT_DISCONNECTED", 2860 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2861 "RDMA_CM_EVENT_MULTICAST_JOIN", 2862 "RDMA_CM_EVENT_MULTICAST_ERROR", 2863 "RDMA_CM_EVENT_ADDR_CHANGE", 2864 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2865 }; 2866 #endif /* DEBUG */ 2867 2868 static void 2869 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2870 struct spdk_nvmf_rdma_port *port) 2871 { 2872 struct spdk_nvmf_rdma_poll_group *rgroup; 2873 struct spdk_nvmf_rdma_poller *rpoller; 2874 struct spdk_nvmf_rdma_qpair *rqpair; 2875 2876 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2877 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2878 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 2879 if (rqpair->listen_id == port->id) { 2880 nvmf_rdma_start_disconnect(rqpair); 2881 } 2882 } 2883 } 2884 } 2885 } 2886 2887 static bool 2888 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2889 struct rdma_cm_event *event) 2890 { 2891 const struct spdk_nvme_transport_id *trid; 2892 struct spdk_nvmf_rdma_port *port; 2893 struct spdk_nvmf_rdma_transport *rtransport; 2894 bool event_acked = false; 2895 2896 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2897 TAILQ_FOREACH(port, &rtransport->ports, link) { 2898 if (port->id == event->id) { 2899 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2900 rdma_ack_cm_event(event); 2901 event_acked = true; 2902 trid = port->trid; 2903 break; 2904 } 2905 } 2906 2907 if (event_acked) { 2908 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2909 2910 nvmf_rdma_stop_listen(transport, trid); 2911 nvmf_rdma_listen(transport, trid); 2912 } 2913 2914 return event_acked; 2915 } 2916 2917 static void 2918 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2919 struct rdma_cm_event *event) 2920 { 2921 struct spdk_nvmf_rdma_port *port; 2922 struct spdk_nvmf_rdma_transport *rtransport; 2923 2924 port = event->id->context; 2925 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2926 2927 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2928 2929 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2930 2931 rdma_ack_cm_event(event); 2932 2933 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2934 ; 2935 } 2936 } 2937 2938 static void 2939 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2940 { 2941 struct spdk_nvmf_rdma_transport *rtransport; 2942 struct rdma_cm_event *event; 2943 int rc; 2944 bool event_acked; 2945 2946 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2947 2948 if (rtransport->event_channel == NULL) { 2949 return; 2950 } 2951 2952 while (1) { 2953 event_acked = false; 2954 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2955 if (rc) { 2956 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2957 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2958 } 2959 break; 2960 } 2961 2962 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2963 2964 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2965 2966 switch (event->event) { 2967 case RDMA_CM_EVENT_ADDR_RESOLVED: 2968 case RDMA_CM_EVENT_ADDR_ERROR: 2969 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2970 case RDMA_CM_EVENT_ROUTE_ERROR: 2971 /* No action required. The target never attempts to resolve routes. */ 2972 break; 2973 case RDMA_CM_EVENT_CONNECT_REQUEST: 2974 rc = nvmf_rdma_connect(transport, event); 2975 if (rc < 0) { 2976 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2977 break; 2978 } 2979 break; 2980 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2981 /* The target never initiates a new connection. So this will not occur. */ 2982 break; 2983 case RDMA_CM_EVENT_CONNECT_ERROR: 2984 /* Can this happen? The docs say it can, but not sure what causes it. */ 2985 break; 2986 case RDMA_CM_EVENT_UNREACHABLE: 2987 case RDMA_CM_EVENT_REJECTED: 2988 /* These only occur on the client side. */ 2989 break; 2990 case RDMA_CM_EVENT_ESTABLISHED: 2991 /* TODO: Should we be waiting for this event anywhere? */ 2992 break; 2993 case RDMA_CM_EVENT_DISCONNECTED: 2994 rc = nvmf_rdma_disconnect(event); 2995 if (rc < 0) { 2996 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2997 break; 2998 } 2999 break; 3000 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3001 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3002 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3003 * Once these events are sent to SPDK, we should release all IB resources and 3004 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3005 * ibv_destory* functions to release user space memory allocated by IB. All kernel 3006 * resources are already cleaned. */ 3007 if (event->id->qp) { 3008 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3009 * corresponding qpair. Otherwise the event refers to a listening device */ 3010 rc = nvmf_rdma_disconnect(event); 3011 if (rc < 0) { 3012 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3013 break; 3014 } 3015 } else { 3016 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3017 event_acked = true; 3018 } 3019 break; 3020 case RDMA_CM_EVENT_MULTICAST_JOIN: 3021 case RDMA_CM_EVENT_MULTICAST_ERROR: 3022 /* Multicast is not used */ 3023 break; 3024 case RDMA_CM_EVENT_ADDR_CHANGE: 3025 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3026 break; 3027 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3028 /* For now, do nothing. The target never re-uses queue pairs. */ 3029 break; 3030 default: 3031 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3032 break; 3033 } 3034 if (!event_acked) { 3035 rdma_ack_cm_event(event); 3036 } 3037 } 3038 } 3039 3040 static void 3041 nvmf_rdma_handle_qp_fatal(struct spdk_nvmf_rdma_qpair *rqpair) 3042 { 3043 nvmf_rdma_update_ibv_state(rqpair); 3044 nvmf_rdma_start_disconnect(rqpair); 3045 } 3046 3047 static void 3048 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3049 { 3050 rqpair->last_wqe_reached = true; 3051 nvmf_rdma_destroy_drained_qpair(rqpair); 3052 } 3053 3054 static void 3055 nvmf_rdma_handle_sq_drained(struct spdk_nvmf_rdma_qpair *rqpair) 3056 { 3057 nvmf_rdma_start_disconnect(rqpair); 3058 } 3059 3060 static void 3061 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3062 { 3063 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3064 3065 if (event_ctx->rqpair) { 3066 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3067 if (event_ctx->cb_fn) { 3068 event_ctx->cb_fn(event_ctx->rqpair); 3069 } 3070 } 3071 free(event_ctx); 3072 } 3073 3074 static int 3075 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3076 spdk_nvmf_rdma_qpair_ibv_event fn) 3077 { 3078 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3079 3080 if (!rqpair->qpair.group) { 3081 return EINVAL; 3082 } 3083 3084 ctx = calloc(1, sizeof(*ctx)); 3085 if (!ctx) { 3086 return ENOMEM; 3087 } 3088 3089 ctx->rqpair = rqpair; 3090 ctx->cb_fn = fn; 3091 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3092 3093 return spdk_thread_send_msg(rqpair->qpair.group->thread, nvmf_rdma_qpair_process_ibv_event, 3094 ctx); 3095 } 3096 3097 static void 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 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 3108 errno, spdk_strerror(errno)); 3109 return; 3110 } 3111 3112 switch (event.event_type) { 3113 case IBV_EVENT_QP_FATAL: 3114 rqpair = event.element.qp->qp_context; 3115 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3116 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3117 (uintptr_t)rqpair->cm_id, event.event_type); 3118 if (nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_qp_fatal)) { 3119 SPDK_ERRLOG("Failed to send QP_FATAL event for rqpair %p\n", rqpair); 3120 nvmf_rdma_handle_qp_fatal(rqpair); 3121 } 3122 break; 3123 case IBV_EVENT_QP_LAST_WQE_REACHED: 3124 /* This event only occurs for shared receive queues. */ 3125 rqpair = event.element.qp->qp_context; 3126 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last WQE reached event received for rqpair %p\n", rqpair); 3127 if (nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached)) { 3128 SPDK_ERRLOG("Failed to send LAST_WQE_REACHED event for rqpair %p\n", rqpair); 3129 rqpair->last_wqe_reached = true; 3130 } 3131 break; 3132 case IBV_EVENT_SQ_DRAINED: 3133 /* This event occurs frequently in both error and non-error states. 3134 * Check if the qpair is in an error state before sending a message. */ 3135 rqpair = event.element.qp->qp_context; 3136 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last sq drained event received for rqpair %p\n", rqpair); 3137 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3138 (uintptr_t)rqpair->cm_id, event.event_type); 3139 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3140 if (nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_sq_drained)) { 3141 SPDK_ERRLOG("Failed to send SQ_DRAINED event for rqpair %p\n", rqpair); 3142 nvmf_rdma_handle_sq_drained(rqpair); 3143 } 3144 } 3145 break; 3146 case IBV_EVENT_QP_REQ_ERR: 3147 case IBV_EVENT_QP_ACCESS_ERR: 3148 case IBV_EVENT_COMM_EST: 3149 case IBV_EVENT_PATH_MIG: 3150 case IBV_EVENT_PATH_MIG_ERR: 3151 SPDK_NOTICELOG("Async event: %s\n", 3152 ibv_event_type_str(event.event_type)); 3153 rqpair = event.element.qp->qp_context; 3154 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3155 (uintptr_t)rqpair->cm_id, event.event_type); 3156 nvmf_rdma_update_ibv_state(rqpair); 3157 break; 3158 case IBV_EVENT_CQ_ERR: 3159 case IBV_EVENT_DEVICE_FATAL: 3160 case IBV_EVENT_PORT_ACTIVE: 3161 case IBV_EVENT_PORT_ERR: 3162 case IBV_EVENT_LID_CHANGE: 3163 case IBV_EVENT_PKEY_CHANGE: 3164 case IBV_EVENT_SM_CHANGE: 3165 case IBV_EVENT_SRQ_ERR: 3166 case IBV_EVENT_SRQ_LIMIT_REACHED: 3167 case IBV_EVENT_CLIENT_REREGISTER: 3168 case IBV_EVENT_GID_CHANGE: 3169 default: 3170 SPDK_NOTICELOG("Async event: %s\n", 3171 ibv_event_type_str(event.event_type)); 3172 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3173 break; 3174 } 3175 ibv_ack_async_event(&event); 3176 } 3177 3178 static uint32_t 3179 nvmf_rdma_accept(struct spdk_nvmf_transport *transport) 3180 { 3181 int nfds, i = 0; 3182 struct spdk_nvmf_rdma_transport *rtransport; 3183 struct spdk_nvmf_rdma_device *device, *tmp; 3184 uint32_t count; 3185 3186 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3187 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3188 3189 if (nfds <= 0) { 3190 return 0; 3191 } 3192 3193 /* The first poll descriptor is RDMA CM event */ 3194 if (rtransport->poll_fds[i++].revents & POLLIN) { 3195 nvmf_process_cm_event(transport); 3196 nfds--; 3197 } 3198 3199 if (nfds == 0) { 3200 return count; 3201 } 3202 3203 /* Second and subsequent poll descriptors are IB async events */ 3204 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3205 if (rtransport->poll_fds[i++].revents & POLLIN) { 3206 nvmf_process_ib_event(device); 3207 nfds--; 3208 } 3209 } 3210 /* check all flagged fd's have been served */ 3211 assert(nfds == 0); 3212 3213 return count; 3214 } 3215 3216 static void 3217 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3218 struct spdk_nvmf_ctrlr_data *cdata) 3219 { 3220 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3221 3222 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3223 since in-capsule data only works with NVME drives that support SGL memory layout */ 3224 if (transport->opts.dif_insert_or_strip) { 3225 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3226 } 3227 } 3228 3229 static void 3230 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3231 struct spdk_nvme_transport_id *trid, 3232 struct spdk_nvmf_discovery_log_page_entry *entry) 3233 { 3234 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3235 entry->adrfam = trid->adrfam; 3236 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3237 3238 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3239 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3240 3241 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3242 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3243 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3244 } 3245 3246 static void 3247 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3248 3249 static struct spdk_nvmf_transport_poll_group * 3250 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3251 { 3252 struct spdk_nvmf_rdma_transport *rtransport; 3253 struct spdk_nvmf_rdma_poll_group *rgroup; 3254 struct spdk_nvmf_rdma_poller *poller; 3255 struct spdk_nvmf_rdma_device *device; 3256 struct ibv_srq_init_attr srq_init_attr; 3257 struct spdk_nvmf_rdma_resource_opts opts; 3258 int num_cqe; 3259 3260 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3261 3262 rgroup = calloc(1, sizeof(*rgroup)); 3263 if (!rgroup) { 3264 return NULL; 3265 } 3266 3267 TAILQ_INIT(&rgroup->pollers); 3268 STAILQ_INIT(&rgroup->retired_bufs); 3269 3270 pthread_mutex_lock(&rtransport->lock); 3271 TAILQ_FOREACH(device, &rtransport->devices, link) { 3272 poller = calloc(1, sizeof(*poller)); 3273 if (!poller) { 3274 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3275 nvmf_rdma_poll_group_destroy(&rgroup->group); 3276 pthread_mutex_unlock(&rtransport->lock); 3277 return NULL; 3278 } 3279 3280 poller->device = device; 3281 poller->group = rgroup; 3282 3283 TAILQ_INIT(&poller->qpairs); 3284 STAILQ_INIT(&poller->qpairs_pending_send); 3285 STAILQ_INIT(&poller->qpairs_pending_recv); 3286 3287 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3288 if (transport->opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3289 poller->max_srq_depth = transport->opts.max_srq_depth; 3290 3291 device->num_srq++; 3292 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 3293 srq_init_attr.attr.max_wr = poller->max_srq_depth; 3294 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3295 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 3296 if (!poller->srq) { 3297 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3298 nvmf_rdma_poll_group_destroy(&rgroup->group); 3299 pthread_mutex_unlock(&rtransport->lock); 3300 return NULL; 3301 } 3302 3303 opts.qp = poller->srq; 3304 opts.pd = device->pd; 3305 opts.qpair = NULL; 3306 opts.shared = true; 3307 opts.max_queue_depth = poller->max_srq_depth; 3308 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3309 3310 poller->resources = nvmf_rdma_resources_create(&opts); 3311 if (!poller->resources) { 3312 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3313 nvmf_rdma_poll_group_destroy(&rgroup->group); 3314 pthread_mutex_unlock(&rtransport->lock); 3315 return NULL; 3316 } 3317 } 3318 3319 /* 3320 * When using an srq, we can limit the completion queue at startup. 3321 * The following formula represents the calculation: 3322 * num_cqe = num_recv + num_data_wr + num_send_wr. 3323 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3324 */ 3325 if (poller->srq) { 3326 num_cqe = poller->max_srq_depth * 3; 3327 } else { 3328 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3329 } 3330 3331 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3332 if (!poller->cq) { 3333 SPDK_ERRLOG("Unable to create completion queue\n"); 3334 nvmf_rdma_poll_group_destroy(&rgroup->group); 3335 pthread_mutex_unlock(&rtransport->lock); 3336 return NULL; 3337 } 3338 poller->num_cqe = num_cqe; 3339 } 3340 3341 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3342 if (rtransport->conn_sched.next_admin_pg == NULL) { 3343 rtransport->conn_sched.next_admin_pg = rgroup; 3344 rtransport->conn_sched.next_io_pg = rgroup; 3345 } 3346 3347 pthread_mutex_unlock(&rtransport->lock); 3348 return &rgroup->group; 3349 } 3350 3351 static struct spdk_nvmf_transport_poll_group * 3352 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3353 { 3354 struct spdk_nvmf_rdma_transport *rtransport; 3355 struct spdk_nvmf_rdma_poll_group **pg; 3356 struct spdk_nvmf_transport_poll_group *result; 3357 3358 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3359 3360 pthread_mutex_lock(&rtransport->lock); 3361 3362 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3363 pthread_mutex_unlock(&rtransport->lock); 3364 return NULL; 3365 } 3366 3367 if (qpair->qid == 0) { 3368 pg = &rtransport->conn_sched.next_admin_pg; 3369 } else { 3370 pg = &rtransport->conn_sched.next_io_pg; 3371 } 3372 3373 assert(*pg != NULL); 3374 3375 result = &(*pg)->group; 3376 3377 *pg = TAILQ_NEXT(*pg, link); 3378 if (*pg == NULL) { 3379 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3380 } 3381 3382 pthread_mutex_unlock(&rtransport->lock); 3383 3384 return result; 3385 } 3386 3387 static void 3388 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3389 { 3390 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3391 struct spdk_nvmf_rdma_poller *poller, *tmp; 3392 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3393 struct spdk_nvmf_transport_pg_cache_buf *buf, *tmp_buf; 3394 struct spdk_nvmf_rdma_transport *rtransport; 3395 3396 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3397 if (!rgroup) { 3398 return; 3399 } 3400 3401 /* free all retired buffers back to the transport so we don't short the mempool. */ 3402 STAILQ_FOREACH_SAFE(buf, &rgroup->retired_bufs, link, tmp_buf) { 3403 STAILQ_REMOVE(&rgroup->retired_bufs, buf, spdk_nvmf_transport_pg_cache_buf, link); 3404 assert(group->transport != NULL); 3405 spdk_mempool_put(group->transport->data_buf_pool, buf); 3406 } 3407 3408 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3409 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3410 3411 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3412 nvmf_rdma_qpair_destroy(qpair); 3413 } 3414 3415 if (poller->srq) { 3416 if (poller->resources) { 3417 nvmf_rdma_resources_destroy(poller->resources); 3418 } 3419 ibv_destroy_srq(poller->srq); 3420 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3421 } 3422 3423 if (poller->cq) { 3424 ibv_destroy_cq(poller->cq); 3425 } 3426 3427 free(poller); 3428 } 3429 3430 if (rgroup->group.transport == NULL) { 3431 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3432 * calls this function directly in a failure path. */ 3433 free(rgroup); 3434 return; 3435 } 3436 3437 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3438 3439 pthread_mutex_lock(&rtransport->lock); 3440 next_rgroup = TAILQ_NEXT(rgroup, link); 3441 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3442 if (next_rgroup == NULL) { 3443 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3444 } 3445 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3446 rtransport->conn_sched.next_admin_pg = next_rgroup; 3447 } 3448 if (rtransport->conn_sched.next_io_pg == rgroup) { 3449 rtransport->conn_sched.next_io_pg = next_rgroup; 3450 } 3451 pthread_mutex_unlock(&rtransport->lock); 3452 3453 free(rgroup); 3454 } 3455 3456 static void 3457 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3458 { 3459 if (rqpair->cm_id != NULL) { 3460 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3461 } 3462 nvmf_rdma_qpair_destroy(rqpair); 3463 } 3464 3465 static int 3466 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3467 struct spdk_nvmf_qpair *qpair) 3468 { 3469 struct spdk_nvmf_rdma_poll_group *rgroup; 3470 struct spdk_nvmf_rdma_qpair *rqpair; 3471 struct spdk_nvmf_rdma_device *device; 3472 struct spdk_nvmf_rdma_poller *poller; 3473 int rc; 3474 3475 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3476 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3477 3478 device = rqpair->device; 3479 3480 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3481 if (poller->device == device) { 3482 break; 3483 } 3484 } 3485 3486 if (!poller) { 3487 SPDK_ERRLOG("No poller found for device.\n"); 3488 return -1; 3489 } 3490 3491 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3492 rqpair->poller = poller; 3493 rqpair->srq = rqpair->poller->srq; 3494 3495 rc = nvmf_rdma_qpair_initialize(qpair); 3496 if (rc < 0) { 3497 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3498 return -1; 3499 } 3500 3501 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3502 if (rc) { 3503 /* Try to reject, but we probably can't */ 3504 nvmf_rdma_qpair_reject_connection(rqpair); 3505 return -1; 3506 } 3507 3508 nvmf_rdma_update_ibv_state(rqpair); 3509 3510 return 0; 3511 } 3512 3513 static int 3514 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3515 { 3516 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3517 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3518 struct spdk_nvmf_rdma_transport, transport); 3519 3520 _nvmf_rdma_request_free(rdma_req, rtransport); 3521 return 0; 3522 } 3523 3524 static int 3525 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3526 { 3527 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3528 struct spdk_nvmf_rdma_transport, transport); 3529 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3530 struct spdk_nvmf_rdma_request, req); 3531 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3532 struct spdk_nvmf_rdma_qpair, qpair); 3533 3534 if (rqpair->ibv_state != IBV_QPS_ERR) { 3535 /* The connection is alive, so process the request as normal */ 3536 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3537 } else { 3538 /* The connection is dead. Move the request directly to the completed state. */ 3539 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3540 } 3541 3542 nvmf_rdma_request_process(rtransport, rdma_req); 3543 3544 return 0; 3545 } 3546 3547 static int 3548 nvmf_rdma_destroy_defunct_qpair(void *ctx) 3549 { 3550 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 3551 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3552 struct spdk_nvmf_rdma_transport, transport); 3553 3554 SPDK_INFOLOG(SPDK_LOG_RDMA, "QP#%d hasn't been drained as expected, manually destroy it\n", 3555 rqpair->qpair.qid); 3556 3557 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3558 nvmf_rdma_qpair_destroy(rqpair); 3559 3560 return SPDK_POLLER_BUSY; 3561 } 3562 3563 static void 3564 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3565 { 3566 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3567 3568 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 3569 return; 3570 } 3571 3572 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 3573 3574 /* This happens only when the qpair is disconnected before 3575 * it is added to the poll group. Since there is no poll group, 3576 * the RDMA qp has not been initialized yet and the RDMA CM 3577 * event has not yet been acknowledged, so we need to reject it. 3578 */ 3579 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3580 nvmf_rdma_qpair_reject_connection(rqpair); 3581 return; 3582 } 3583 3584 if (rqpair->rdma_qp) { 3585 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3586 } 3587 3588 rqpair->destruct_poller = SPDK_POLLER_REGISTER(nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 3589 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 3590 } 3591 3592 static struct spdk_nvmf_rdma_qpair * 3593 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3594 { 3595 struct spdk_nvmf_rdma_qpair *rqpair; 3596 /* @todo: improve QP search */ 3597 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3598 if (wc->qp_num == rqpair->rdma_qp->qp->qp_num) { 3599 return rqpair; 3600 } 3601 } 3602 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3603 return NULL; 3604 } 3605 3606 #ifdef DEBUG 3607 static int 3608 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3609 { 3610 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3611 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3612 } 3613 #endif 3614 3615 static void 3616 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3617 int rc) 3618 { 3619 struct spdk_nvmf_rdma_recv *rdma_recv; 3620 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3621 3622 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3623 while (bad_recv_wr != NULL) { 3624 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3625 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3626 3627 rdma_recv->qpair->current_recv_depth++; 3628 bad_recv_wr = bad_recv_wr->next; 3629 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3630 nvmf_rdma_start_disconnect(rdma_recv->qpair); 3631 } 3632 } 3633 3634 static void 3635 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3636 { 3637 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3638 while (bad_recv_wr != NULL) { 3639 bad_recv_wr = bad_recv_wr->next; 3640 rqpair->current_recv_depth++; 3641 } 3642 nvmf_rdma_start_disconnect(rqpair); 3643 } 3644 3645 static void 3646 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3647 struct spdk_nvmf_rdma_poller *rpoller) 3648 { 3649 struct spdk_nvmf_rdma_qpair *rqpair; 3650 struct ibv_recv_wr *bad_recv_wr; 3651 int rc; 3652 3653 if (rpoller->srq) { 3654 if (rpoller->resources->recvs_to_post.first != NULL) { 3655 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3656 if (rc) { 3657 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3658 } 3659 rpoller->resources->recvs_to_post.first = NULL; 3660 rpoller->resources->recvs_to_post.last = NULL; 3661 } 3662 } else { 3663 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3664 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3665 assert(rqpair->resources->recvs_to_post.first != NULL); 3666 rc = ibv_post_recv(rqpair->rdma_qp->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3667 if (rc) { 3668 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3669 } 3670 rqpair->resources->recvs_to_post.first = NULL; 3671 rqpair->resources->recvs_to_post.last = NULL; 3672 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3673 } 3674 } 3675 } 3676 3677 static void 3678 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3679 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3680 { 3681 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3682 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3683 3684 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3685 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3686 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3687 assert(rqpair->current_send_depth > 0); 3688 rqpair->current_send_depth--; 3689 switch (bad_rdma_wr->type) { 3690 case RDMA_WR_TYPE_DATA: 3691 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3692 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3693 assert(rqpair->current_read_depth > 0); 3694 rqpair->current_read_depth--; 3695 } 3696 break; 3697 case RDMA_WR_TYPE_SEND: 3698 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3699 break; 3700 default: 3701 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3702 prev_rdma_req = cur_rdma_req; 3703 continue; 3704 } 3705 3706 if (prev_rdma_req == cur_rdma_req) { 3707 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3708 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3709 continue; 3710 } 3711 3712 switch (cur_rdma_req->state) { 3713 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3714 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3715 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3716 break; 3717 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3718 case RDMA_REQUEST_STATE_COMPLETING: 3719 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3720 break; 3721 default: 3722 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3723 cur_rdma_req->state, rqpair); 3724 continue; 3725 } 3726 3727 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3728 prev_rdma_req = cur_rdma_req; 3729 } 3730 3731 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3732 /* Disconnect the connection. */ 3733 nvmf_rdma_start_disconnect(rqpair); 3734 } 3735 3736 } 3737 3738 static void 3739 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3740 struct spdk_nvmf_rdma_poller *rpoller) 3741 { 3742 struct spdk_nvmf_rdma_qpair *rqpair; 3743 struct ibv_send_wr *bad_wr = NULL; 3744 int rc; 3745 3746 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3747 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3748 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3749 3750 /* bad wr always points to the first wr that failed. */ 3751 if (rc) { 3752 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3753 } 3754 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3755 } 3756 } 3757 3758 static int 3759 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3760 struct spdk_nvmf_rdma_poller *rpoller) 3761 { 3762 struct ibv_wc wc[32]; 3763 struct spdk_nvmf_rdma_wr *rdma_wr; 3764 struct spdk_nvmf_rdma_request *rdma_req; 3765 struct spdk_nvmf_rdma_recv *rdma_recv; 3766 struct spdk_nvmf_rdma_qpair *rqpair; 3767 int reaped, i; 3768 int count = 0; 3769 bool error = false; 3770 uint64_t poll_tsc = spdk_get_ticks(); 3771 3772 /* Poll for completing operations. */ 3773 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3774 if (reaped < 0) { 3775 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3776 errno, spdk_strerror(errno)); 3777 return -1; 3778 } 3779 3780 rpoller->stat.polls++; 3781 rpoller->stat.completions += reaped; 3782 3783 for (i = 0; i < reaped; i++) { 3784 3785 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3786 3787 switch (rdma_wr->type) { 3788 case RDMA_WR_TYPE_SEND: 3789 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3790 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3791 3792 if (!wc[i].status) { 3793 count++; 3794 assert(wc[i].opcode == IBV_WC_SEND); 3795 assert(nvmf_rdma_req_is_completing(rdma_req)); 3796 } 3797 3798 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3799 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3800 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3801 rdma_req->num_outstanding_data_wr = 0; 3802 3803 nvmf_rdma_request_process(rtransport, rdma_req); 3804 break; 3805 case RDMA_WR_TYPE_RECV: 3806 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3807 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3808 if (rpoller->srq != NULL) { 3809 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3810 /* It is possible that there are still some completions for destroyed QP 3811 * associated with SRQ. We just ignore these late completions and re-post 3812 * receive WRs back to SRQ. 3813 */ 3814 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3815 struct ibv_recv_wr *bad_wr; 3816 int rc; 3817 3818 rdma_recv->wr.next = NULL; 3819 rc = ibv_post_srq_recv(rpoller->srq, 3820 &rdma_recv->wr, 3821 &bad_wr); 3822 if (rc) { 3823 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3824 } 3825 continue; 3826 } 3827 } 3828 rqpair = rdma_recv->qpair; 3829 3830 assert(rqpair != NULL); 3831 if (!wc[i].status) { 3832 assert(wc[i].opcode == IBV_WC_RECV); 3833 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3834 nvmf_rdma_start_disconnect(rqpair); 3835 break; 3836 } 3837 } 3838 3839 rdma_recv->wr.next = NULL; 3840 rqpair->current_recv_depth++; 3841 rdma_recv->receive_tsc = poll_tsc; 3842 rpoller->stat.requests++; 3843 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3844 break; 3845 case RDMA_WR_TYPE_DATA: 3846 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3847 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3848 3849 assert(rdma_req->num_outstanding_data_wr > 0); 3850 3851 rqpair->current_send_depth--; 3852 rdma_req->num_outstanding_data_wr--; 3853 if (!wc[i].status) { 3854 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3855 rqpair->current_read_depth--; 3856 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3857 if (rdma_req->num_outstanding_data_wr == 0) { 3858 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3859 nvmf_rdma_request_process(rtransport, rdma_req); 3860 } 3861 } else { 3862 /* If the data transfer fails still force the queue into the error state, 3863 * if we were performing an RDMA_READ, we need to force the request into a 3864 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3865 * case, we should wait for the SEND to complete. */ 3866 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3867 rqpair->current_read_depth--; 3868 if (rdma_req->num_outstanding_data_wr == 0) { 3869 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3870 } 3871 } 3872 } 3873 break; 3874 default: 3875 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3876 continue; 3877 } 3878 3879 /* Handle error conditions */ 3880 if (wc[i].status) { 3881 if ((rdma_wr->type == RDMA_WR_TYPE_RECV && !rpoller->srq)) { 3882 /* When we don't use SRQ and close a qpair, we will receive completions with error 3883 * status for all posted ibv_recv_wrs. This is expected and we don't want to log 3884 * an error in that case. */ 3885 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Error on CQ %p, request 0x%lu, type %d, status: (%d): %s\n", 3886 rpoller->cq, wc[i].wr_id, rdma_wr->type, wc[i].status, ibv_wc_status_str(wc[i].status)); 3887 } else { 3888 SPDK_ERRLOG("Error on CQ %p, request 0x%lu, type %d, status: (%d): %s\n", 3889 rpoller->cq, wc[i].wr_id, rdma_wr->type, wc[i].status, ibv_wc_status_str(wc[i].status)); 3890 } 3891 3892 error = true; 3893 3894 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3895 /* Disconnect the connection. */ 3896 nvmf_rdma_start_disconnect(rqpair); 3897 } else { 3898 nvmf_rdma_destroy_drained_qpair(rqpair); 3899 } 3900 continue; 3901 } 3902 3903 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3904 3905 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3906 nvmf_rdma_destroy_drained_qpair(rqpair); 3907 } 3908 } 3909 3910 if (error == true) { 3911 return -1; 3912 } 3913 3914 /* submit outstanding work requests. */ 3915 _poller_submit_recvs(rtransport, rpoller); 3916 _poller_submit_sends(rtransport, rpoller); 3917 3918 return count; 3919 } 3920 3921 static int 3922 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3923 { 3924 struct spdk_nvmf_rdma_transport *rtransport; 3925 struct spdk_nvmf_rdma_poll_group *rgroup; 3926 struct spdk_nvmf_rdma_poller *rpoller; 3927 int count, rc; 3928 3929 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3930 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3931 3932 count = 0; 3933 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3934 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 3935 if (rc < 0) { 3936 return rc; 3937 } 3938 count += rc; 3939 } 3940 3941 return count; 3942 } 3943 3944 static int 3945 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3946 struct spdk_nvme_transport_id *trid, 3947 bool peer) 3948 { 3949 struct sockaddr *saddr; 3950 uint16_t port; 3951 3952 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 3953 3954 if (peer) { 3955 saddr = rdma_get_peer_addr(id); 3956 } else { 3957 saddr = rdma_get_local_addr(id); 3958 } 3959 switch (saddr->sa_family) { 3960 case AF_INET: { 3961 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3962 3963 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3964 inet_ntop(AF_INET, &saddr_in->sin_addr, 3965 trid->traddr, sizeof(trid->traddr)); 3966 if (peer) { 3967 port = ntohs(rdma_get_dst_port(id)); 3968 } else { 3969 port = ntohs(rdma_get_src_port(id)); 3970 } 3971 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3972 break; 3973 } 3974 case AF_INET6: { 3975 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3976 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3977 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3978 trid->traddr, sizeof(trid->traddr)); 3979 if (peer) { 3980 port = ntohs(rdma_get_dst_port(id)); 3981 } else { 3982 port = ntohs(rdma_get_src_port(id)); 3983 } 3984 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3985 break; 3986 } 3987 default: 3988 return -1; 3989 3990 } 3991 3992 return 0; 3993 } 3994 3995 static int 3996 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3997 struct spdk_nvme_transport_id *trid) 3998 { 3999 struct spdk_nvmf_rdma_qpair *rqpair; 4000 4001 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4002 4003 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4004 } 4005 4006 static int 4007 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4008 struct spdk_nvme_transport_id *trid) 4009 { 4010 struct spdk_nvmf_rdma_qpair *rqpair; 4011 4012 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4013 4014 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4015 } 4016 4017 static int 4018 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4019 struct spdk_nvme_transport_id *trid) 4020 { 4021 struct spdk_nvmf_rdma_qpair *rqpair; 4022 4023 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4024 4025 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4026 } 4027 4028 void 4029 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4030 { 4031 g_nvmf_hooks = *hooks; 4032 } 4033 4034 static void 4035 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4036 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4037 { 4038 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4039 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4040 4041 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4042 4043 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4044 } 4045 4046 static int 4047 _nvmf_rdma_qpair_abort_request(void *ctx) 4048 { 4049 struct spdk_nvmf_request *req = ctx; 4050 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4051 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4052 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4053 struct spdk_nvmf_rdma_qpair, qpair); 4054 int rc; 4055 4056 spdk_poller_unregister(&req->poller); 4057 4058 switch (rdma_req_to_abort->state) { 4059 case RDMA_REQUEST_STATE_EXECUTING: 4060 rc = nvmf_ctrlr_abort_request(req, &rdma_req_to_abort->req); 4061 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4062 return SPDK_POLLER_BUSY; 4063 } 4064 break; 4065 4066 case RDMA_REQUEST_STATE_NEED_BUFFER: 4067 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4068 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4069 4070 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4071 break; 4072 4073 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4074 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4075 spdk_nvmf_rdma_request, state_link); 4076 4077 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4078 break; 4079 4080 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4081 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4082 spdk_nvmf_rdma_request, state_link); 4083 4084 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4085 break; 4086 4087 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4088 if (spdk_get_ticks() < req->timeout_tsc) { 4089 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4090 return SPDK_POLLER_BUSY; 4091 } 4092 break; 4093 4094 default: 4095 break; 4096 } 4097 4098 spdk_nvmf_request_complete(req); 4099 return SPDK_POLLER_BUSY; 4100 } 4101 4102 static void 4103 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4104 struct spdk_nvmf_request *req) 4105 { 4106 struct spdk_nvmf_rdma_qpair *rqpair; 4107 struct spdk_nvmf_rdma_transport *rtransport; 4108 struct spdk_nvmf_transport *transport; 4109 uint16_t cid; 4110 uint32_t i; 4111 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL; 4112 4113 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4114 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4115 transport = &rtransport->transport; 4116 4117 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4118 4119 for (i = 0; i < rqpair->max_queue_depth; i++) { 4120 rdma_req_to_abort = &rqpair->resources->reqs[i]; 4121 4122 if (rdma_req_to_abort->state != RDMA_REQUEST_STATE_FREE && 4123 rdma_req_to_abort->req.cmd->nvme_cmd.cid == cid) { 4124 break; 4125 } 4126 } 4127 4128 if (rdma_req_to_abort == NULL) { 4129 spdk_nvmf_request_complete(req); 4130 return; 4131 } 4132 4133 req->req_to_abort = &rdma_req_to_abort->req; 4134 req->timeout_tsc = spdk_get_ticks() + 4135 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4136 req->poller = NULL; 4137 4138 _nvmf_rdma_qpair_abort_request(req); 4139 } 4140 4141 static int 4142 nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4143 struct spdk_nvmf_transport_poll_group_stat **stat) 4144 { 4145 struct spdk_io_channel *ch; 4146 struct spdk_nvmf_poll_group *group; 4147 struct spdk_nvmf_transport_poll_group *tgroup; 4148 struct spdk_nvmf_rdma_poll_group *rgroup; 4149 struct spdk_nvmf_rdma_poller *rpoller; 4150 struct spdk_nvmf_rdma_device_stat *device_stat; 4151 uint64_t num_devices = 0; 4152 4153 if (tgt == NULL || stat == NULL) { 4154 return -EINVAL; 4155 } 4156 4157 ch = spdk_get_io_channel(tgt); 4158 group = spdk_io_channel_get_ctx(ch);; 4159 spdk_put_io_channel(ch); 4160 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4161 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4162 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4163 if (!*stat) { 4164 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4165 return -ENOMEM; 4166 } 4167 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4168 4169 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4170 /* Count devices to allocate enough memory */ 4171 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4172 ++num_devices; 4173 } 4174 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4175 if (!(*stat)->rdma.devices) { 4176 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4177 free(*stat); 4178 return -ENOMEM; 4179 } 4180 4181 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4182 (*stat)->rdma.num_devices = num_devices; 4183 num_devices = 0; 4184 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4185 device_stat = &(*stat)->rdma.devices[num_devices++]; 4186 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4187 device_stat->polls = rpoller->stat.polls; 4188 device_stat->completions = rpoller->stat.completions; 4189 device_stat->requests = rpoller->stat.requests; 4190 device_stat->request_latency = rpoller->stat.request_latency; 4191 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4192 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4193 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4194 } 4195 return 0; 4196 } 4197 } 4198 return -ENOENT; 4199 } 4200 4201 static void 4202 nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4203 { 4204 if (stat) { 4205 free(stat->rdma.devices); 4206 } 4207 free(stat); 4208 } 4209 4210 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4211 .name = "RDMA", 4212 .type = SPDK_NVME_TRANSPORT_RDMA, 4213 .opts_init = nvmf_rdma_opts_init, 4214 .create = nvmf_rdma_create, 4215 .destroy = nvmf_rdma_destroy, 4216 4217 .listen = nvmf_rdma_listen, 4218 .stop_listen = nvmf_rdma_stop_listen, 4219 .accept = nvmf_rdma_accept, 4220 .cdata_init = nvmf_rdma_cdata_init, 4221 4222 .listener_discover = nvmf_rdma_discover, 4223 4224 .poll_group_create = nvmf_rdma_poll_group_create, 4225 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4226 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4227 .poll_group_add = nvmf_rdma_poll_group_add, 4228 .poll_group_poll = nvmf_rdma_poll_group_poll, 4229 4230 .req_free = nvmf_rdma_request_free, 4231 .req_complete = nvmf_rdma_request_complete, 4232 4233 .qpair_fini = nvmf_rdma_close_qpair, 4234 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4235 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4236 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4237 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4238 4239 .poll_group_get_stat = nvmf_rdma_poll_group_get_stat, 4240 .poll_group_free_stat = nvmf_rdma_poll_group_free_stat, 4241 }; 4242 4243 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4244 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 4245