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 /* 407 * io_channel which is used to destroy qpair when it is removed from poll group 408 */ 409 struct spdk_io_channel *destruct_channel; 410 411 /* List of ibv async events */ 412 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 413 414 /* There are several ways a disconnect can start on a qpair 415 * and they are not all mutually exclusive. It is important 416 * that we only initialize one of these paths. 417 */ 418 bool disconnect_started; 419 /* Lets us know that we have received the last_wqe event. */ 420 bool last_wqe_reached; 421 }; 422 423 struct spdk_nvmf_rdma_poller_stat { 424 uint64_t completions; 425 uint64_t polls; 426 uint64_t requests; 427 uint64_t request_latency; 428 uint64_t pending_free_request; 429 uint64_t pending_rdma_read; 430 uint64_t pending_rdma_write; 431 }; 432 433 struct spdk_nvmf_rdma_poller { 434 struct spdk_nvmf_rdma_device *device; 435 struct spdk_nvmf_rdma_poll_group *group; 436 437 int num_cqe; 438 int required_num_wr; 439 struct ibv_cq *cq; 440 441 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 442 uint16_t max_srq_depth; 443 444 /* Shared receive queue */ 445 struct ibv_srq *srq; 446 447 struct spdk_nvmf_rdma_resources *resources; 448 struct spdk_nvmf_rdma_poller_stat stat; 449 450 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 451 452 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 453 454 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 455 456 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 457 }; 458 459 struct spdk_nvmf_rdma_poll_group_stat { 460 uint64_t pending_data_buffer; 461 }; 462 463 struct spdk_nvmf_rdma_poll_group { 464 struct spdk_nvmf_transport_poll_group group; 465 struct spdk_nvmf_rdma_poll_group_stat stat; 466 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 467 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 468 /* 469 * buffers which are split across multiple RDMA 470 * memory regions cannot be used by this transport. 471 */ 472 STAILQ_HEAD(, spdk_nvmf_transport_pg_cache_buf) retired_bufs; 473 }; 474 475 struct spdk_nvmf_rdma_conn_sched { 476 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 477 struct spdk_nvmf_rdma_poll_group *next_io_pg; 478 }; 479 480 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 481 struct spdk_nvmf_rdma_device { 482 struct ibv_device_attr attr; 483 struct ibv_context *context; 484 485 struct spdk_mem_map *map; 486 struct ibv_pd *pd; 487 488 int num_srq; 489 490 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 491 }; 492 493 struct spdk_nvmf_rdma_port { 494 const struct spdk_nvme_transport_id *trid; 495 struct rdma_cm_id *id; 496 struct spdk_nvmf_rdma_device *device; 497 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 498 }; 499 500 struct spdk_nvmf_rdma_transport { 501 struct spdk_nvmf_transport transport; 502 503 struct spdk_nvmf_rdma_conn_sched conn_sched; 504 505 struct rdma_event_channel *event_channel; 506 507 struct spdk_mempool *data_wr_pool; 508 509 pthread_mutex_t lock; 510 511 /* fields used to poll RDMA/IB events */ 512 nfds_t npoll_fds; 513 struct pollfd *poll_fds; 514 515 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 516 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 517 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 518 }; 519 520 static inline void 521 nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair); 522 523 static bool 524 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 525 struct spdk_nvmf_rdma_request *rdma_req); 526 527 static inline int 528 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 529 { 530 switch (state) { 531 case IBV_QPS_RESET: 532 case IBV_QPS_INIT: 533 case IBV_QPS_RTR: 534 case IBV_QPS_RTS: 535 case IBV_QPS_SQD: 536 case IBV_QPS_SQE: 537 case IBV_QPS_ERR: 538 return 0; 539 default: 540 return -1; 541 } 542 } 543 544 static inline enum spdk_nvme_media_error_status_code 545 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 546 enum spdk_nvme_media_error_status_code result; 547 switch (err_type) 548 { 549 case SPDK_DIF_REFTAG_ERROR: 550 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 551 break; 552 case SPDK_DIF_APPTAG_ERROR: 553 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 554 break; 555 case SPDK_DIF_GUARD_ERROR: 556 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 557 break; 558 default: 559 SPDK_UNREACHABLE(); 560 } 561 562 return result; 563 } 564 565 static enum ibv_qp_state 566 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 567 enum ibv_qp_state old_state, new_state; 568 struct ibv_qp_attr qp_attr; 569 struct ibv_qp_init_attr init_attr; 570 int rc; 571 572 old_state = rqpair->ibv_state; 573 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 574 g_spdk_nvmf_ibv_query_mask, &init_attr); 575 576 if (rc) 577 { 578 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 579 return IBV_QPS_ERR + 1; 580 } 581 582 new_state = qp_attr.qp_state; 583 rqpair->ibv_state = new_state; 584 qp_attr.ah_attr.port_num = qp_attr.port_num; 585 586 rc = nvmf_rdma_check_ibv_state(new_state); 587 if (rc) 588 { 589 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 590 /* 591 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 592 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 593 */ 594 return IBV_QPS_ERR + 1; 595 } 596 597 if (old_state != new_state) 598 { 599 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 600 (uintptr_t)rqpair->cm_id, new_state); 601 } 602 return new_state; 603 } 604 605 static void 606 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 607 struct spdk_nvmf_rdma_transport *rtransport) 608 { 609 struct spdk_nvmf_rdma_request_data *data_wr; 610 struct ibv_send_wr *next_send_wr; 611 uint64_t req_wrid; 612 613 rdma_req->num_outstanding_data_wr = 0; 614 data_wr = &rdma_req->data; 615 req_wrid = data_wr->wr.wr_id; 616 while (data_wr && data_wr->wr.wr_id == req_wrid) { 617 memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge); 618 data_wr->wr.num_sge = 0; 619 next_send_wr = data_wr->wr.next; 620 if (data_wr != &rdma_req->data) { 621 spdk_mempool_put(rtransport->data_wr_pool, data_wr); 622 } 623 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : 624 SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr); 625 } 626 } 627 628 static void 629 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 630 { 631 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 632 if (req->req.cmd) { 633 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 634 } 635 if (req->recv) { 636 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 637 } 638 } 639 640 static void 641 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 642 { 643 int i; 644 645 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 646 for (i = 0; i < rqpair->max_queue_depth; i++) { 647 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 648 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 649 } 650 } 651 } 652 653 static void 654 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 655 { 656 if (resources->cmds_mr) { 657 ibv_dereg_mr(resources->cmds_mr); 658 } 659 660 if (resources->cpls_mr) { 661 ibv_dereg_mr(resources->cpls_mr); 662 } 663 664 if (resources->bufs_mr) { 665 ibv_dereg_mr(resources->bufs_mr); 666 } 667 668 spdk_free(resources->cmds); 669 spdk_free(resources->cpls); 670 spdk_free(resources->bufs); 671 free(resources->reqs); 672 free(resources->recvs); 673 free(resources); 674 } 675 676 677 static struct spdk_nvmf_rdma_resources * 678 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 679 { 680 struct spdk_nvmf_rdma_resources *resources; 681 struct spdk_nvmf_rdma_request *rdma_req; 682 struct spdk_nvmf_rdma_recv *rdma_recv; 683 struct ibv_qp *qp; 684 struct ibv_srq *srq; 685 uint32_t i; 686 int rc; 687 688 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 689 if (!resources) { 690 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 691 return NULL; 692 } 693 694 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 695 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 696 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 697 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 698 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 699 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 700 701 if (opts->in_capsule_data_size > 0) { 702 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 703 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 704 SPDK_MALLOC_DMA); 705 } 706 707 if (!resources->reqs || !resources->recvs || !resources->cmds || 708 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 709 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 710 goto cleanup; 711 } 712 713 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 714 opts->max_queue_depth * sizeof(*resources->cmds), 715 IBV_ACCESS_LOCAL_WRITE); 716 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 717 opts->max_queue_depth * sizeof(*resources->cpls), 718 0); 719 720 if (opts->in_capsule_data_size) { 721 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 722 opts->max_queue_depth * 723 opts->in_capsule_data_size, 724 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 725 } 726 727 if (!resources->cmds_mr || !resources->cpls_mr || 728 (opts->in_capsule_data_size && 729 !resources->bufs_mr)) { 730 goto cleanup; 731 } 732 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 733 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 734 resources->cmds_mr->lkey); 735 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 736 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 737 resources->cpls_mr->lkey); 738 if (resources->bufs && resources->bufs_mr) { 739 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 740 resources->bufs, opts->max_queue_depth * 741 opts->in_capsule_data_size, resources->bufs_mr->lkey); 742 } 743 744 /* Initialize queues */ 745 STAILQ_INIT(&resources->incoming_queue); 746 STAILQ_INIT(&resources->free_queue); 747 748 for (i = 0; i < opts->max_queue_depth; i++) { 749 struct ibv_recv_wr *bad_wr = NULL; 750 751 rdma_recv = &resources->recvs[i]; 752 rdma_recv->qpair = opts->qpair; 753 754 /* Set up memory to receive commands */ 755 if (resources->bufs) { 756 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 757 opts->in_capsule_data_size)); 758 } 759 760 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 761 762 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 763 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 764 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 765 rdma_recv->wr.num_sge = 1; 766 767 if (rdma_recv->buf && resources->bufs_mr) { 768 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 769 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 770 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 771 rdma_recv->wr.num_sge++; 772 } 773 774 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 775 rdma_recv->wr.sg_list = rdma_recv->sgl; 776 if (opts->shared) { 777 srq = (struct ibv_srq *)opts->qp; 778 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 779 } else { 780 qp = (struct ibv_qp *)opts->qp; 781 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 782 } 783 if (rc) { 784 goto cleanup; 785 } 786 } 787 788 for (i = 0; i < opts->max_queue_depth; i++) { 789 rdma_req = &resources->reqs[i]; 790 791 if (opts->qpair != NULL) { 792 rdma_req->req.qpair = &opts->qpair->qpair; 793 } else { 794 rdma_req->req.qpair = NULL; 795 } 796 rdma_req->req.cmd = NULL; 797 798 /* Set up memory to send responses */ 799 rdma_req->req.rsp = &resources->cpls[i]; 800 801 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 802 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 803 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 804 805 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 806 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 807 rdma_req->rsp.wr.next = NULL; 808 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 809 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 810 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 811 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 812 813 /* Set up memory for data buffers */ 814 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 815 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 816 rdma_req->data.wr.next = NULL; 817 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 818 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 819 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 820 821 /* Initialize request state to FREE */ 822 rdma_req->state = RDMA_REQUEST_STATE_FREE; 823 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 824 } 825 826 return resources; 827 828 cleanup: 829 nvmf_rdma_resources_destroy(resources); 830 return NULL; 831 } 832 833 static void 834 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 835 { 836 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 837 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 838 ctx->rqpair = NULL; 839 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 840 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 841 } 842 } 843 844 static void 845 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 846 { 847 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 848 struct ibv_recv_wr *bad_recv_wr = NULL; 849 int rc; 850 851 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 852 853 spdk_poller_unregister(&rqpair->destruct_poller); 854 855 if (rqpair->qd != 0) { 856 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 857 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 858 struct spdk_nvmf_rdma_transport, transport); 859 struct spdk_nvmf_rdma_request *req; 860 uint32_t i, max_req_count = 0; 861 862 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 863 864 if (rqpair->srq == NULL) { 865 nvmf_rdma_dump_qpair_contents(rqpair); 866 max_req_count = rqpair->max_queue_depth; 867 } else if (rqpair->poller && rqpair->resources) { 868 max_req_count = rqpair->poller->max_srq_depth; 869 } 870 871 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Release incomplete requests\n"); 872 for (i = 0; i < max_req_count; i++) { 873 req = &rqpair->resources->reqs[i]; 874 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 875 /* nvmf_rdma_request_process checks qpair ibv and internal state 876 * and completes a request */ 877 nvmf_rdma_request_process(rtransport, req); 878 } 879 } 880 assert(rqpair->qd == 0); 881 } 882 883 if (rqpair->poller) { 884 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 885 886 if (rqpair->srq != NULL && rqpair->resources != NULL) { 887 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 888 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 889 if (rqpair == rdma_recv->qpair) { 890 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 891 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 892 if (rc) { 893 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 894 } 895 } 896 } 897 } 898 } 899 900 if (rqpair->cm_id) { 901 if (rqpair->rdma_qp != NULL) { 902 spdk_rdma_qp_destroy(rqpair->rdma_qp); 903 rqpair->rdma_qp = NULL; 904 } 905 rdma_destroy_id(rqpair->cm_id); 906 907 if (rqpair->poller != NULL && rqpair->srq == NULL) { 908 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 909 } 910 } 911 912 if (rqpair->srq == NULL && rqpair->resources != NULL) { 913 nvmf_rdma_resources_destroy(rqpair->resources); 914 } 915 916 nvmf_rdma_qpair_clean_ibv_events(rqpair); 917 918 if (rqpair->destruct_channel) { 919 spdk_put_io_channel(rqpair->destruct_channel); 920 rqpair->destruct_channel = NULL; 921 } 922 923 free(rqpair); 924 } 925 926 static int 927 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 928 { 929 struct spdk_nvmf_rdma_poller *rpoller; 930 int rc, num_cqe, required_num_wr; 931 932 /* Enlarge CQ size dynamically */ 933 rpoller = rqpair->poller; 934 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 935 num_cqe = rpoller->num_cqe; 936 if (num_cqe < required_num_wr) { 937 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 938 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 939 } 940 941 if (rpoller->num_cqe != num_cqe) { 942 if (required_num_wr > device->attr.max_cqe) { 943 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 944 required_num_wr, device->attr.max_cqe); 945 return -1; 946 } 947 948 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 949 rc = ibv_resize_cq(rpoller->cq, num_cqe); 950 if (rc) { 951 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 952 return -1; 953 } 954 955 rpoller->num_cqe = num_cqe; 956 } 957 958 rpoller->required_num_wr = required_num_wr; 959 return 0; 960 } 961 962 static int 963 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 964 { 965 struct spdk_nvmf_rdma_qpair *rqpair; 966 struct spdk_nvmf_rdma_transport *rtransport; 967 struct spdk_nvmf_transport *transport; 968 struct spdk_nvmf_rdma_resource_opts opts; 969 struct spdk_nvmf_rdma_device *device; 970 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 971 972 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 973 device = rqpair->device; 974 975 qp_init_attr.qp_context = rqpair; 976 qp_init_attr.pd = device->pd; 977 qp_init_attr.send_cq = rqpair->poller->cq; 978 qp_init_attr.recv_cq = rqpair->poller->cq; 979 980 if (rqpair->srq) { 981 qp_init_attr.srq = rqpair->srq; 982 } else { 983 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 984 } 985 986 /* SEND, READ, and WRITE operations */ 987 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 988 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 989 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 990 991 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 992 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 993 goto error; 994 } 995 996 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 997 if (!rqpair->rdma_qp) { 998 goto error; 999 } 1000 1001 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1002 qp_init_attr.cap.max_send_wr); 1003 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1004 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1005 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 1006 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 1007 1008 if (rqpair->poller->srq == NULL) { 1009 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1010 transport = &rtransport->transport; 1011 1012 opts.qp = rqpair->rdma_qp->qp; 1013 opts.pd = rqpair->cm_id->pd; 1014 opts.qpair = rqpair; 1015 opts.shared = false; 1016 opts.max_queue_depth = rqpair->max_queue_depth; 1017 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1018 1019 rqpair->resources = nvmf_rdma_resources_create(&opts); 1020 1021 if (!rqpair->resources) { 1022 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1023 rdma_destroy_qp(rqpair->cm_id); 1024 goto error; 1025 } 1026 } else { 1027 rqpair->resources = rqpair->poller->resources; 1028 } 1029 1030 rqpair->current_recv_depth = 0; 1031 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1032 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1033 1034 return 0; 1035 1036 error: 1037 rdma_destroy_id(rqpair->cm_id); 1038 rqpair->cm_id = NULL; 1039 return -1; 1040 } 1041 1042 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1043 /* This function accepts either a single wr or the first wr in a linked list. */ 1044 static void 1045 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1046 { 1047 struct ibv_recv_wr *last; 1048 1049 last = first; 1050 while (last->next != NULL) { 1051 last = last->next; 1052 } 1053 1054 if (rqpair->resources->recvs_to_post.first == NULL) { 1055 rqpair->resources->recvs_to_post.first = first; 1056 rqpair->resources->recvs_to_post.last = last; 1057 if (rqpair->srq == NULL) { 1058 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1059 } 1060 } else { 1061 rqpair->resources->recvs_to_post.last->next = first; 1062 rqpair->resources->recvs_to_post.last = last; 1063 } 1064 } 1065 1066 static int 1067 request_transfer_in(struct spdk_nvmf_request *req) 1068 { 1069 struct spdk_nvmf_rdma_request *rdma_req; 1070 struct spdk_nvmf_qpair *qpair; 1071 struct spdk_nvmf_rdma_qpair *rqpair; 1072 1073 qpair = req->qpair; 1074 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1075 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1076 1077 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1078 assert(rdma_req != NULL); 1079 1080 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1081 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1082 } 1083 1084 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1085 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1086 return 0; 1087 } 1088 1089 static int 1090 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1091 { 1092 int num_outstanding_data_wr = 0; 1093 struct spdk_nvmf_rdma_request *rdma_req; 1094 struct spdk_nvmf_qpair *qpair; 1095 struct spdk_nvmf_rdma_qpair *rqpair; 1096 struct spdk_nvme_cpl *rsp; 1097 struct ibv_send_wr *first = NULL; 1098 1099 *data_posted = 0; 1100 qpair = req->qpair; 1101 rsp = &req->rsp->nvme_cpl; 1102 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1103 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1104 1105 /* Advance our sq_head pointer */ 1106 if (qpair->sq_head == qpair->sq_head_max) { 1107 qpair->sq_head = 0; 1108 } else { 1109 qpair->sq_head++; 1110 } 1111 rsp->sqhd = qpair->sq_head; 1112 1113 /* queue the capsule for the recv buffer */ 1114 assert(rdma_req->recv != NULL); 1115 1116 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1117 1118 rdma_req->recv = NULL; 1119 assert(rqpair->current_recv_depth > 0); 1120 rqpair->current_recv_depth--; 1121 1122 /* Build the response which consists of optional 1123 * RDMA WRITEs to transfer data, plus an RDMA SEND 1124 * containing the response. 1125 */ 1126 first = &rdma_req->rsp.wr; 1127 1128 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1129 /* On failure, data was not read from the controller. So clear the 1130 * number of outstanding data WRs to zero. 1131 */ 1132 rdma_req->num_outstanding_data_wr = 0; 1133 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1134 first = &rdma_req->data.wr; 1135 *data_posted = 1; 1136 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1137 } 1138 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1139 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1140 } 1141 1142 /* +1 for the rsp wr */ 1143 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1144 1145 return 0; 1146 } 1147 1148 static int 1149 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1150 { 1151 struct spdk_nvmf_rdma_accept_private_data accept_data; 1152 struct rdma_conn_param ctrlr_event_data = {}; 1153 int rc; 1154 1155 accept_data.recfmt = 0; 1156 accept_data.crqsize = rqpair->max_queue_depth; 1157 1158 ctrlr_event_data.private_data = &accept_data; 1159 ctrlr_event_data.private_data_len = sizeof(accept_data); 1160 if (id->ps == RDMA_PS_TCP) { 1161 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1162 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1163 } 1164 1165 /* Configure infinite retries for the initiator side qpair. 1166 * When using a shared receive queue on the target side, 1167 * we need to pass this value to the initiator to prevent the 1168 * initiator side NIC from completing SEND requests back to the 1169 * initiator with status rnr_retry_count_exceeded. */ 1170 if (rqpair->srq != NULL) { 1171 ctrlr_event_data.rnr_retry_count = 0x7; 1172 } 1173 1174 /* When qpair is created without use of rdma cm API, an additional 1175 * information must be provided to initiator in the connection response: 1176 * whether qpair is using SRQ and its qp_num 1177 * Fields below are ignored by rdma cm if qpair has been 1178 * created using rdma cm API. */ 1179 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1180 ctrlr_event_data.qp_num = rqpair->rdma_qp->qp->qp_num; 1181 1182 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1183 if (rc) { 1184 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1185 } else { 1186 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1187 } 1188 1189 return rc; 1190 } 1191 1192 static void 1193 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1194 { 1195 struct spdk_nvmf_rdma_reject_private_data rej_data; 1196 1197 rej_data.recfmt = 0; 1198 rej_data.sts = error; 1199 1200 rdma_reject(id, &rej_data, sizeof(rej_data)); 1201 } 1202 1203 static int 1204 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1205 { 1206 struct spdk_nvmf_rdma_transport *rtransport; 1207 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1208 struct spdk_nvmf_rdma_port *port; 1209 struct rdma_conn_param *rdma_param = NULL; 1210 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1211 uint16_t max_queue_depth; 1212 uint16_t max_read_depth; 1213 1214 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1215 1216 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1217 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1218 1219 rdma_param = &event->param.conn; 1220 if (rdma_param->private_data == NULL || 1221 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1222 SPDK_ERRLOG("connect request: no private data provided\n"); 1223 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1224 return -1; 1225 } 1226 1227 private_data = rdma_param->private_data; 1228 if (private_data->recfmt != 0) { 1229 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1230 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1231 return -1; 1232 } 1233 1234 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1235 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1236 1237 port = event->listen_id->context; 1238 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1239 event->listen_id, event->listen_id->verbs, port); 1240 1241 /* Figure out the supported queue depth. This is a multi-step process 1242 * that takes into account hardware maximums, host provided values, 1243 * and our target's internal memory limits */ 1244 1245 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1246 1247 /* Start with the maximum queue depth allowed by the target */ 1248 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1249 max_read_depth = rtransport->transport.opts.max_queue_depth; 1250 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1251 rtransport->transport.opts.max_queue_depth); 1252 1253 /* Next check the local NIC's hardware limitations */ 1254 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1255 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1256 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1257 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1258 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1259 1260 /* Next check the remote NIC's hardware limitations */ 1261 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1262 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1263 rdma_param->initiator_depth, rdma_param->responder_resources); 1264 if (rdma_param->initiator_depth > 0) { 1265 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1266 } 1267 1268 /* Finally check for the host software requested values, which are 1269 * optional. */ 1270 if (rdma_param->private_data != NULL && 1271 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1272 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1273 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1274 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1275 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1276 } 1277 1278 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1279 max_queue_depth, max_read_depth); 1280 1281 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1282 if (rqpair == NULL) { 1283 SPDK_ERRLOG("Could not allocate new connection.\n"); 1284 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1285 return -1; 1286 } 1287 1288 rqpair->device = port->device; 1289 rqpair->max_queue_depth = max_queue_depth; 1290 rqpair->max_read_depth = max_read_depth; 1291 rqpair->cm_id = event->id; 1292 rqpair->listen_id = event->listen_id; 1293 rqpair->qpair.transport = transport; 1294 STAILQ_INIT(&rqpair->ibv_events); 1295 /* use qid from the private data to determine the qpair type 1296 qid will be set to the appropriate value when the controller is created */ 1297 rqpair->qpair.qid = private_data->qid; 1298 1299 event->id->context = &rqpair->qpair; 1300 1301 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1302 1303 return 0; 1304 } 1305 1306 static int 1307 nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1308 enum spdk_mem_map_notify_action action, 1309 void *vaddr, size_t size) 1310 { 1311 struct ibv_pd *pd = cb_ctx; 1312 struct ibv_mr *mr; 1313 int rc; 1314 1315 switch (action) { 1316 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1317 if (!g_nvmf_hooks.get_rkey) { 1318 mr = ibv_reg_mr(pd, vaddr, size, 1319 IBV_ACCESS_LOCAL_WRITE | 1320 IBV_ACCESS_REMOTE_READ | 1321 IBV_ACCESS_REMOTE_WRITE); 1322 if (mr == NULL) { 1323 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1324 return -1; 1325 } else { 1326 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1327 } 1328 } else { 1329 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1330 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1331 } 1332 break; 1333 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1334 if (!g_nvmf_hooks.get_rkey) { 1335 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1336 if (mr) { 1337 ibv_dereg_mr(mr); 1338 } 1339 } 1340 rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1341 break; 1342 default: 1343 SPDK_UNREACHABLE(); 1344 } 1345 1346 return rc; 1347 } 1348 1349 static int 1350 nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1351 { 1352 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1353 return addr_1 == addr_2; 1354 } 1355 1356 static inline void 1357 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1358 enum spdk_nvme_data_transfer xfer) 1359 { 1360 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1361 wr->opcode = IBV_WR_RDMA_WRITE; 1362 wr->send_flags = 0; 1363 wr->next = next; 1364 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1365 wr->opcode = IBV_WR_RDMA_READ; 1366 wr->send_flags = IBV_SEND_SIGNALED; 1367 wr->next = NULL; 1368 } else { 1369 assert(0); 1370 } 1371 } 1372 1373 static int 1374 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1375 struct spdk_nvmf_rdma_request *rdma_req, 1376 uint32_t num_sgl_descriptors) 1377 { 1378 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1379 struct spdk_nvmf_rdma_request_data *current_data_wr; 1380 uint32_t i; 1381 1382 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1383 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1384 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1385 return -EINVAL; 1386 } 1387 1388 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1389 return -ENOMEM; 1390 } 1391 1392 current_data_wr = &rdma_req->data; 1393 1394 for (i = 0; i < num_sgl_descriptors; i++) { 1395 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1396 current_data_wr->wr.next = &work_requests[i]->wr; 1397 current_data_wr = work_requests[i]; 1398 current_data_wr->wr.sg_list = current_data_wr->sgl; 1399 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1400 } 1401 1402 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1403 1404 return 0; 1405 } 1406 1407 static inline void 1408 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1409 { 1410 struct ibv_send_wr *wr = &rdma_req->data.wr; 1411 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1412 1413 wr->wr.rdma.rkey = sgl->keyed.key; 1414 wr->wr.rdma.remote_addr = sgl->address; 1415 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1416 } 1417 1418 static inline void 1419 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1420 { 1421 struct ibv_send_wr *wr = &rdma_req->data.wr; 1422 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1423 uint32_t i; 1424 int j; 1425 uint64_t remote_addr_offset = 0; 1426 1427 for (i = 0; i < num_wrs; ++i) { 1428 wr->wr.rdma.rkey = sgl->keyed.key; 1429 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1430 for (j = 0; j < wr->num_sge; ++j) { 1431 remote_addr_offset += wr->sg_list[j].length; 1432 } 1433 wr = wr->next; 1434 } 1435 } 1436 1437 /* This function is used in the rare case that we have a buffer split over multiple memory regions. */ 1438 static int 1439 nvmf_rdma_replace_buffer(struct spdk_nvmf_rdma_poll_group *rgroup, void **buf) 1440 { 1441 struct spdk_nvmf_transport_poll_group *group = &rgroup->group; 1442 struct spdk_nvmf_transport *transport = group->transport; 1443 struct spdk_nvmf_transport_pg_cache_buf *old_buf; 1444 void *new_buf; 1445 1446 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1447 group->buf_cache_count--; 1448 new_buf = STAILQ_FIRST(&group->buf_cache); 1449 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1450 assert(*buf != NULL); 1451 } else { 1452 new_buf = spdk_mempool_get(transport->data_buf_pool); 1453 } 1454 1455 if (*buf == NULL) { 1456 return -ENOMEM; 1457 } 1458 1459 old_buf = *buf; 1460 STAILQ_INSERT_HEAD(&rgroup->retired_bufs, old_buf, link); 1461 *buf = new_buf; 1462 return 0; 1463 } 1464 1465 static bool 1466 nvmf_rdma_get_lkey(struct spdk_nvmf_rdma_device *device, struct iovec *iov, 1467 uint32_t *_lkey) 1468 { 1469 uint64_t translation_len; 1470 uint32_t lkey; 1471 1472 translation_len = iov->iov_len; 1473 1474 if (!g_nvmf_hooks.get_rkey) { 1475 lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1476 (uint64_t)iov->iov_base, &translation_len))->lkey; 1477 } else { 1478 lkey = spdk_mem_map_translate(device->map, 1479 (uint64_t)iov->iov_base, &translation_len); 1480 } 1481 1482 if (spdk_unlikely(translation_len < iov->iov_len)) { 1483 return false; 1484 } 1485 1486 *_lkey = lkey; 1487 return true; 1488 } 1489 1490 static bool 1491 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device, 1492 struct iovec *iov, struct ibv_send_wr **_wr, 1493 uint32_t *_remaining_data_block, uint32_t *_offset, 1494 uint32_t *_num_extra_wrs, 1495 const struct spdk_dif_ctx *dif_ctx) 1496 { 1497 struct ibv_send_wr *wr = *_wr; 1498 struct ibv_sge *sg_ele = &wr->sg_list[wr->num_sge]; 1499 uint32_t lkey = 0; 1500 uint32_t remaining, data_block_size, md_size, sge_len; 1501 1502 if (spdk_unlikely(!nvmf_rdma_get_lkey(device, iov, &lkey))) { 1503 /* This is a very rare case that can occur when using DPDK version < 19.05 */ 1504 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions. Removing it from circulation.\n"); 1505 return false; 1506 } 1507 1508 if (spdk_likely(!dif_ctx)) { 1509 sg_ele->lkey = lkey; 1510 sg_ele->addr = (uintptr_t)(iov->iov_base); 1511 sg_ele->length = iov->iov_len; 1512 wr->num_sge++; 1513 } else { 1514 remaining = iov->iov_len - *_offset; 1515 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1516 md_size = dif_ctx->md_size; 1517 1518 while (remaining) { 1519 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1520 if (*_num_extra_wrs > 0 && wr->next) { 1521 *_wr = wr->next; 1522 wr = *_wr; 1523 wr->num_sge = 0; 1524 sg_ele = &wr->sg_list[wr->num_sge]; 1525 (*_num_extra_wrs)--; 1526 } else { 1527 break; 1528 } 1529 } 1530 sg_ele->lkey = lkey; 1531 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset); 1532 sge_len = spdk_min(remaining, *_remaining_data_block); 1533 sg_ele->length = sge_len; 1534 remaining -= sge_len; 1535 *_remaining_data_block -= sge_len; 1536 *_offset += sge_len; 1537 1538 sg_ele++; 1539 wr->num_sge++; 1540 1541 if (*_remaining_data_block == 0) { 1542 /* skip metadata */ 1543 *_offset += md_size; 1544 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1545 remaining -= spdk_min(remaining, md_size); 1546 *_remaining_data_block = data_block_size; 1547 } 1548 1549 if (remaining == 0) { 1550 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1551 the remaining metadata bits at the beginning of the next buffer */ 1552 *_offset -= iov->iov_len; 1553 } 1554 } 1555 } 1556 1557 return true; 1558 } 1559 1560 static int 1561 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1562 struct spdk_nvmf_rdma_device *device, 1563 struct spdk_nvmf_rdma_request *rdma_req, 1564 struct ibv_send_wr *wr, 1565 uint32_t length, 1566 uint32_t num_extra_wrs) 1567 { 1568 struct spdk_nvmf_request *req = &rdma_req->req; 1569 struct spdk_dif_ctx *dif_ctx = NULL; 1570 uint32_t remaining_data_block = 0; 1571 uint32_t offset = 0; 1572 1573 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 1574 dif_ctx = &rdma_req->req.dif.dif_ctx; 1575 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1576 } 1577 1578 wr->num_sge = 0; 1579 1580 while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1581 while (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr, 1582 &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) { 1583 if (nvmf_rdma_replace_buffer(rgroup, &req->buffers[rdma_req->iovpos]) == -ENOMEM) { 1584 return -ENOMEM; 1585 } 1586 req->iov[rdma_req->iovpos].iov_base = (void *)((uintptr_t)(req->buffers[rdma_req->iovpos] + 1587 NVMF_DATA_BUFFER_MASK) & 1588 ~NVMF_DATA_BUFFER_MASK); 1589 } 1590 1591 length -= req->iov[rdma_req->iovpos].iov_len; 1592 rdma_req->iovpos++; 1593 } 1594 1595 if (length) { 1596 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1597 return -EINVAL; 1598 } 1599 1600 return 0; 1601 } 1602 1603 static inline uint32_t 1604 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1605 { 1606 /* estimate the number of SG entries and WRs needed to process the request */ 1607 uint32_t num_sge = 0; 1608 uint32_t i; 1609 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1610 1611 for (i = 0; i < num_buffers && length > 0; i++) { 1612 uint32_t buffer_len = spdk_min(length, io_unit_size); 1613 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1614 1615 if (num_sge_in_block * block_size > buffer_len) { 1616 ++num_sge_in_block; 1617 } 1618 num_sge += num_sge_in_block; 1619 length -= buffer_len; 1620 } 1621 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1622 } 1623 1624 static int 1625 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1626 struct spdk_nvmf_rdma_device *device, 1627 struct spdk_nvmf_rdma_request *rdma_req, 1628 uint32_t length) 1629 { 1630 struct spdk_nvmf_rdma_qpair *rqpair; 1631 struct spdk_nvmf_rdma_poll_group *rgroup; 1632 struct spdk_nvmf_request *req = &rdma_req->req; 1633 struct ibv_send_wr *wr = &rdma_req->data.wr; 1634 int rc; 1635 uint32_t num_wrs = 1; 1636 1637 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1638 rgroup = rqpair->poller->group; 1639 1640 /* rdma wr specifics */ 1641 nvmf_rdma_setup_request(rdma_req); 1642 1643 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1644 length); 1645 if (rc != 0) { 1646 return rc; 1647 } 1648 1649 assert(req->iovcnt <= rqpair->max_send_sge); 1650 1651 rdma_req->iovpos = 0; 1652 1653 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1654 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1655 req->dif.dif_ctx.block_size); 1656 if (num_wrs > 1) { 1657 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1658 if (rc != 0) { 1659 goto err_exit; 1660 } 1661 } 1662 } 1663 1664 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1665 if (spdk_unlikely(rc != 0)) { 1666 goto err_exit; 1667 } 1668 1669 if (spdk_unlikely(num_wrs > 1)) { 1670 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1671 } 1672 1673 /* set the number of outstanding data WRs for this request. */ 1674 rdma_req->num_outstanding_data_wr = num_wrs; 1675 1676 return rc; 1677 1678 err_exit: 1679 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1680 nvmf_rdma_request_free_data(rdma_req, rtransport); 1681 req->iovcnt = 0; 1682 return rc; 1683 } 1684 1685 static int 1686 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1687 struct spdk_nvmf_rdma_device *device, 1688 struct spdk_nvmf_rdma_request *rdma_req) 1689 { 1690 struct spdk_nvmf_rdma_qpair *rqpair; 1691 struct spdk_nvmf_rdma_poll_group *rgroup; 1692 struct ibv_send_wr *current_wr; 1693 struct spdk_nvmf_request *req = &rdma_req->req; 1694 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1695 uint32_t num_sgl_descriptors; 1696 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES]; 1697 uint32_t i; 1698 int rc; 1699 1700 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1701 rgroup = rqpair->poller->group; 1702 1703 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1704 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1705 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1706 1707 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1708 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1709 1710 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1711 return -ENOMEM; 1712 } 1713 1714 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1715 for (i = 0; i < num_sgl_descriptors; i++) { 1716 if (spdk_likely(!req->dif.dif_insert_or_strip)) { 1717 lengths[i] = desc->keyed.length; 1718 } else { 1719 req->dif.orig_length += desc->keyed.length; 1720 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1721 req->dif.elba_length += lengths[i]; 1722 } 1723 desc++; 1724 } 1725 1726 rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport, 1727 lengths, num_sgl_descriptors); 1728 if (rc != 0) { 1729 nvmf_rdma_request_free_data(rdma_req, rtransport); 1730 return rc; 1731 } 1732 1733 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1734 current_wr = &rdma_req->data.wr; 1735 assert(current_wr != NULL); 1736 1737 req->length = 0; 1738 rdma_req->iovpos = 0; 1739 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1740 for (i = 0; i < num_sgl_descriptors; i++) { 1741 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1742 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1743 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1744 rc = -EINVAL; 1745 goto err_exit; 1746 } 1747 1748 current_wr->num_sge = 0; 1749 1750 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1751 if (rc != 0) { 1752 rc = -ENOMEM; 1753 goto err_exit; 1754 } 1755 1756 req->length += desc->keyed.length; 1757 current_wr->wr.rdma.rkey = desc->keyed.key; 1758 current_wr->wr.rdma.remote_addr = desc->address; 1759 current_wr = current_wr->next; 1760 desc++; 1761 } 1762 1763 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1764 /* Go back to the last descriptor in the list. */ 1765 desc--; 1766 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1767 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1768 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1769 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1770 } 1771 } 1772 #endif 1773 1774 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1775 1776 return 0; 1777 1778 err_exit: 1779 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1780 nvmf_rdma_request_free_data(rdma_req, rtransport); 1781 return rc; 1782 } 1783 1784 static int 1785 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1786 struct spdk_nvmf_rdma_device *device, 1787 struct spdk_nvmf_rdma_request *rdma_req) 1788 { 1789 struct spdk_nvmf_request *req = &rdma_req->req; 1790 struct spdk_nvme_cpl *rsp; 1791 struct spdk_nvme_sgl_descriptor *sgl; 1792 int rc; 1793 uint32_t length; 1794 1795 rsp = &req->rsp->nvme_cpl; 1796 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1797 1798 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1799 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1800 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1801 1802 length = sgl->keyed.length; 1803 if (length > rtransport->transport.opts.max_io_size) { 1804 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1805 length, rtransport->transport.opts.max_io_size); 1806 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1807 return -1; 1808 } 1809 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1810 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1811 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1812 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1813 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1814 } 1815 } 1816 #endif 1817 1818 /* fill request length and populate iovs */ 1819 req->length = length; 1820 1821 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1822 req->dif.orig_length = length; 1823 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1824 req->dif.elba_length = length; 1825 } 1826 1827 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1828 if (spdk_unlikely(rc < 0)) { 1829 if (rc == -EINVAL) { 1830 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1831 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1832 return -1; 1833 } 1834 /* No available buffers. Queue this request up. */ 1835 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1836 return 0; 1837 } 1838 1839 /* backward compatible */ 1840 req->data = req->iov[0].iov_base; 1841 1842 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1843 req->iovcnt); 1844 1845 return 0; 1846 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1847 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1848 uint64_t offset = sgl->address; 1849 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1850 1851 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1852 offset, sgl->unkeyed.length); 1853 1854 if (offset > max_len) { 1855 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1856 offset, max_len); 1857 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1858 return -1; 1859 } 1860 max_len -= (uint32_t)offset; 1861 1862 if (sgl->unkeyed.length > max_len) { 1863 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1864 sgl->unkeyed.length, max_len); 1865 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1866 return -1; 1867 } 1868 1869 rdma_req->num_outstanding_data_wr = 0; 1870 req->data = rdma_req->recv->buf + offset; 1871 req->data_from_pool = false; 1872 req->length = sgl->unkeyed.length; 1873 1874 req->iov[0].iov_base = req->data; 1875 req->iov[0].iov_len = req->length; 1876 req->iovcnt = 1; 1877 1878 return 0; 1879 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1880 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1881 1882 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1883 if (rc == -ENOMEM) { 1884 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1885 return 0; 1886 } else if (rc == -EINVAL) { 1887 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1888 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1889 return -1; 1890 } 1891 1892 /* backward compatible */ 1893 req->data = req->iov[0].iov_base; 1894 1895 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1896 req->iovcnt); 1897 1898 return 0; 1899 } 1900 1901 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1902 sgl->generic.type, sgl->generic.subtype); 1903 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1904 return -1; 1905 } 1906 1907 static void 1908 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1909 struct spdk_nvmf_rdma_transport *rtransport) 1910 { 1911 struct spdk_nvmf_rdma_qpair *rqpair; 1912 struct spdk_nvmf_rdma_poll_group *rgroup; 1913 1914 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1915 if (rdma_req->req.data_from_pool) { 1916 rgroup = rqpair->poller->group; 1917 1918 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1919 } 1920 nvmf_rdma_request_free_data(rdma_req, rtransport); 1921 rdma_req->req.length = 0; 1922 rdma_req->req.iovcnt = 0; 1923 rdma_req->req.data = NULL; 1924 rdma_req->rsp.wr.next = NULL; 1925 rdma_req->data.wr.next = NULL; 1926 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1927 rqpair->qd--; 1928 1929 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1930 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1931 } 1932 1933 bool 1934 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1935 struct spdk_nvmf_rdma_request *rdma_req) 1936 { 1937 struct spdk_nvmf_rdma_qpair *rqpair; 1938 struct spdk_nvmf_rdma_device *device; 1939 struct spdk_nvmf_rdma_poll_group *rgroup; 1940 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1941 int rc; 1942 struct spdk_nvmf_rdma_recv *rdma_recv; 1943 enum spdk_nvmf_rdma_request_state prev_state; 1944 bool progress = false; 1945 int data_posted; 1946 uint32_t num_blocks; 1947 1948 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1949 device = rqpair->device; 1950 rgroup = rqpair->poller->group; 1951 1952 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1953 1954 /* If the queue pair is in an error state, force the request to the completed state 1955 * to release resources. */ 1956 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1957 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1958 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1959 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1960 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1961 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1962 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1963 } 1964 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1965 } 1966 1967 /* The loop here is to allow for several back-to-back state changes. */ 1968 do { 1969 prev_state = rdma_req->state; 1970 1971 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 1972 1973 switch (rdma_req->state) { 1974 case RDMA_REQUEST_STATE_FREE: 1975 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1976 * to escape this state. */ 1977 break; 1978 case RDMA_REQUEST_STATE_NEW: 1979 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1980 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1981 rdma_recv = rdma_req->recv; 1982 1983 /* The first element of the SGL is the NVMe command */ 1984 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1985 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1986 1987 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1988 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1989 break; 1990 } 1991 1992 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 1993 rdma_req->req.dif.dif_insert_or_strip = true; 1994 } 1995 1996 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1997 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1998 rdma_req->rsp.wr.imm_data = 0; 1999 #endif 2000 2001 /* The next state transition depends on the data transfer needs of this request. */ 2002 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2003 2004 /* If no data to transfer, ready to execute. */ 2005 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2006 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2007 break; 2008 } 2009 2010 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2011 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2012 break; 2013 case RDMA_REQUEST_STATE_NEED_BUFFER: 2014 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2015 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2016 2017 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2018 2019 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2020 /* This request needs to wait in line to obtain a buffer */ 2021 break; 2022 } 2023 2024 /* Try to get a data buffer */ 2025 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2026 if (rc < 0) { 2027 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2028 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2029 break; 2030 } 2031 2032 if (!rdma_req->req.data) { 2033 /* No buffers available. */ 2034 rgroup->stat.pending_data_buffer++; 2035 break; 2036 } 2037 2038 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2039 2040 /* If data is transferring from host to controller and the data didn't 2041 * arrive using in capsule data, we need to do a transfer from the host. 2042 */ 2043 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2044 rdma_req->req.data_from_pool) { 2045 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2046 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2047 break; 2048 } 2049 2050 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2051 break; 2052 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2053 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2054 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2055 2056 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2057 /* This request needs to wait in line to perform RDMA */ 2058 break; 2059 } 2060 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2061 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2062 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2063 rqpair->poller->stat.pending_rdma_read++; 2064 break; 2065 } 2066 2067 /* We have already verified that this request is the head of the queue. */ 2068 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2069 2070 rc = request_transfer_in(&rdma_req->req); 2071 if (!rc) { 2072 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2073 } else { 2074 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2075 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2076 } 2077 break; 2078 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2079 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2080 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2081 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2082 * to escape this state. */ 2083 break; 2084 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2085 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2086 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2087 2088 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2089 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2090 /* generate DIF for write operation */ 2091 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2092 assert(num_blocks > 0); 2093 2094 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2095 num_blocks, &rdma_req->req.dif.dif_ctx); 2096 if (rc != 0) { 2097 SPDK_ERRLOG("DIF generation failed\n"); 2098 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2099 nvmf_rdma_start_disconnect(rqpair); 2100 break; 2101 } 2102 } 2103 2104 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2105 /* set extended length before IO operation */ 2106 rdma_req->req.length = rdma_req->req.dif.elba_length; 2107 } 2108 2109 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2110 spdk_nvmf_request_exec(&rdma_req->req); 2111 break; 2112 case RDMA_REQUEST_STATE_EXECUTING: 2113 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2114 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2115 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2116 * to escape this state. */ 2117 break; 2118 case RDMA_REQUEST_STATE_EXECUTED: 2119 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2120 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2121 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2122 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2123 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2124 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2125 } else { 2126 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2127 } 2128 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2129 /* restore the original length */ 2130 rdma_req->req.length = rdma_req->req.dif.orig_length; 2131 2132 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2133 struct spdk_dif_error error_blk; 2134 2135 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2136 2137 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2138 &rdma_req->req.dif.dif_ctx, &error_blk); 2139 if (rc) { 2140 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2141 2142 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2143 error_blk.err_offset); 2144 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2145 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2146 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2147 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2148 } 2149 } 2150 } 2151 break; 2152 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2153 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2154 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2155 2156 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2157 /* This request needs to wait in line to perform RDMA */ 2158 break; 2159 } 2160 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2161 rqpair->max_send_depth) { 2162 /* We can only have so many WRs outstanding. we have to wait until some finish. 2163 * +1 since each request has an additional wr in the resp. */ 2164 rqpair->poller->stat.pending_rdma_write++; 2165 break; 2166 } 2167 2168 /* We have already verified that this request is the head of the queue. */ 2169 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2170 2171 /* The data transfer will be kicked off from 2172 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2173 */ 2174 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2175 break; 2176 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2177 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2178 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2179 rc = request_transfer_out(&rdma_req->req, &data_posted); 2180 assert(rc == 0); /* No good way to handle this currently */ 2181 if (rc) { 2182 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2183 } else { 2184 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2185 RDMA_REQUEST_STATE_COMPLETING; 2186 } 2187 break; 2188 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2189 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2190 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2191 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2192 * to escape this state. */ 2193 break; 2194 case RDMA_REQUEST_STATE_COMPLETING: 2195 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2196 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2197 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2198 * to escape this state. */ 2199 break; 2200 case RDMA_REQUEST_STATE_COMPLETED: 2201 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2202 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2203 2204 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2205 _nvmf_rdma_request_free(rdma_req, rtransport); 2206 break; 2207 case RDMA_REQUEST_NUM_STATES: 2208 default: 2209 assert(0); 2210 break; 2211 } 2212 2213 if (rdma_req->state != prev_state) { 2214 progress = true; 2215 } 2216 } while (rdma_req->state != prev_state); 2217 2218 return progress; 2219 } 2220 2221 /* Public API callbacks begin here */ 2222 2223 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2224 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2225 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2226 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2227 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2228 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2229 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2230 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2231 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2232 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2233 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2234 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2235 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2236 2237 static void 2238 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2239 { 2240 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2241 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2242 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2243 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2244 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2245 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2246 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2247 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2248 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2249 opts->no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2250 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2251 opts->acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2252 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2253 } 2254 2255 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2256 .notify_cb = nvmf_rdma_mem_notify, 2257 .are_contiguous = nvmf_rdma_check_contiguous_entries 2258 }; 2259 2260 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2261 2262 static struct spdk_nvmf_transport * 2263 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2264 { 2265 int rc; 2266 struct spdk_nvmf_rdma_transport *rtransport; 2267 struct spdk_nvmf_rdma_device *device, *tmp; 2268 struct ibv_context **contexts; 2269 uint32_t i; 2270 int flag; 2271 uint32_t sge_count; 2272 uint32_t min_shared_buffers; 2273 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2274 pthread_mutexattr_t attr; 2275 2276 rtransport = calloc(1, sizeof(*rtransport)); 2277 if (!rtransport) { 2278 return NULL; 2279 } 2280 2281 if (pthread_mutexattr_init(&attr)) { 2282 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2283 free(rtransport); 2284 return NULL; 2285 } 2286 2287 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2288 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2289 pthread_mutexattr_destroy(&attr); 2290 free(rtransport); 2291 return NULL; 2292 } 2293 2294 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2295 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2296 pthread_mutexattr_destroy(&attr); 2297 free(rtransport); 2298 return NULL; 2299 } 2300 2301 pthread_mutexattr_destroy(&attr); 2302 2303 TAILQ_INIT(&rtransport->devices); 2304 TAILQ_INIT(&rtransport->ports); 2305 TAILQ_INIT(&rtransport->poll_groups); 2306 2307 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2308 2309 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2310 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2311 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2312 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2313 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d," 2314 " acceptor_backlog=%d, abort_timeout_sec=%d\n", 2315 opts->max_queue_depth, 2316 opts->max_io_size, 2317 opts->max_qpairs_per_ctrlr - 1, 2318 opts->io_unit_size, 2319 opts->in_capsule_data_size, 2320 opts->max_aq_depth, 2321 opts->num_shared_buffers, 2322 opts->max_srq_depth, 2323 opts->no_srq, 2324 opts->acceptor_backlog, 2325 opts->abort_timeout_sec); 2326 2327 /* I/O unit size cannot be larger than max I/O size */ 2328 if (opts->io_unit_size > opts->max_io_size) { 2329 opts->io_unit_size = opts->max_io_size; 2330 } 2331 2332 if (opts->acceptor_backlog <= 0) { 2333 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2334 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2335 opts->acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2336 } 2337 2338 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2339 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2340 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2341 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2342 nvmf_rdma_destroy(&rtransport->transport); 2343 return NULL; 2344 } 2345 2346 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2347 if (min_shared_buffers > opts->num_shared_buffers) { 2348 SPDK_ERRLOG("There are not enough buffers to satisfy" 2349 "per-poll group caches for each thread. (%" PRIu32 ")" 2350 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2351 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2352 nvmf_rdma_destroy(&rtransport->transport); 2353 return NULL; 2354 } 2355 2356 sge_count = opts->max_io_size / opts->io_unit_size; 2357 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2358 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2359 nvmf_rdma_destroy(&rtransport->transport); 2360 return NULL; 2361 } 2362 2363 rtransport->event_channel = rdma_create_event_channel(); 2364 if (rtransport->event_channel == NULL) { 2365 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2366 nvmf_rdma_destroy(&rtransport->transport); 2367 return NULL; 2368 } 2369 2370 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2371 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2372 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2373 rtransport->event_channel->fd, spdk_strerror(errno)); 2374 nvmf_rdma_destroy(&rtransport->transport); 2375 return NULL; 2376 } 2377 2378 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2379 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2380 sizeof(struct spdk_nvmf_rdma_request_data), 2381 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2382 SPDK_ENV_SOCKET_ID_ANY); 2383 if (!rtransport->data_wr_pool) { 2384 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2385 nvmf_rdma_destroy(&rtransport->transport); 2386 return NULL; 2387 } 2388 2389 contexts = rdma_get_devices(NULL); 2390 if (contexts == NULL) { 2391 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2392 nvmf_rdma_destroy(&rtransport->transport); 2393 return NULL; 2394 } 2395 2396 i = 0; 2397 rc = 0; 2398 while (contexts[i] != NULL) { 2399 device = calloc(1, sizeof(*device)); 2400 if (!device) { 2401 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2402 rc = -ENOMEM; 2403 break; 2404 } 2405 device->context = contexts[i]; 2406 rc = ibv_query_device(device->context, &device->attr); 2407 if (rc < 0) { 2408 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2409 free(device); 2410 break; 2411 2412 } 2413 2414 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2415 2416 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2417 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2418 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2419 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2420 } 2421 2422 /** 2423 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2424 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2425 * but incorrectly reports that it does. There are changes making their way 2426 * through the kernel now that will enable this feature. When they are merged, 2427 * we can conditionally enable this feature. 2428 * 2429 * TODO: enable this for versions of the kernel rxe driver that support it. 2430 */ 2431 if (device->attr.vendor_id == 0) { 2432 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2433 } 2434 #endif 2435 2436 /* set up device context async ev fd as NON_BLOCKING */ 2437 flag = fcntl(device->context->async_fd, F_GETFL); 2438 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2439 if (rc < 0) { 2440 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2441 free(device); 2442 break; 2443 } 2444 2445 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2446 i++; 2447 2448 if (g_nvmf_hooks.get_ibv_pd) { 2449 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2450 } else { 2451 device->pd = ibv_alloc_pd(device->context); 2452 } 2453 2454 if (!device->pd) { 2455 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2456 rc = -ENOMEM; 2457 break; 2458 } 2459 2460 assert(device->map == NULL); 2461 2462 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2463 if (!device->map) { 2464 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2465 rc = -ENOMEM; 2466 break; 2467 } 2468 2469 assert(device->map != NULL); 2470 assert(device->pd != NULL); 2471 } 2472 rdma_free_devices(contexts); 2473 2474 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2475 /* divide and round up. */ 2476 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2477 2478 /* round up to the nearest 4k. */ 2479 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2480 2481 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2482 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2483 opts->io_unit_size); 2484 } 2485 2486 if (rc < 0) { 2487 nvmf_rdma_destroy(&rtransport->transport); 2488 return NULL; 2489 } 2490 2491 /* Set up poll descriptor array to monitor events from RDMA and IB 2492 * in a single poll syscall 2493 */ 2494 rtransport->npoll_fds = i + 1; 2495 i = 0; 2496 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2497 if (rtransport->poll_fds == NULL) { 2498 SPDK_ERRLOG("poll_fds allocation failed\n"); 2499 nvmf_rdma_destroy(&rtransport->transport); 2500 return NULL; 2501 } 2502 2503 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2504 rtransport->poll_fds[i++].events = POLLIN; 2505 2506 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2507 rtransport->poll_fds[i].fd = device->context->async_fd; 2508 rtransport->poll_fds[i++].events = POLLIN; 2509 } 2510 2511 return &rtransport->transport; 2512 } 2513 2514 static int 2515 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2516 { 2517 struct spdk_nvmf_rdma_transport *rtransport; 2518 struct spdk_nvmf_rdma_port *port, *port_tmp; 2519 struct spdk_nvmf_rdma_device *device, *device_tmp; 2520 2521 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2522 2523 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2524 TAILQ_REMOVE(&rtransport->ports, port, link); 2525 rdma_destroy_id(port->id); 2526 free(port); 2527 } 2528 2529 if (rtransport->poll_fds != NULL) { 2530 free(rtransport->poll_fds); 2531 } 2532 2533 if (rtransport->event_channel != NULL) { 2534 rdma_destroy_event_channel(rtransport->event_channel); 2535 } 2536 2537 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2538 TAILQ_REMOVE(&rtransport->devices, device, link); 2539 if (device->map) { 2540 spdk_mem_map_free(&device->map); 2541 } 2542 if (device->pd) { 2543 if (!g_nvmf_hooks.get_ibv_pd) { 2544 ibv_dealloc_pd(device->pd); 2545 } 2546 } 2547 free(device); 2548 } 2549 2550 if (rtransport->data_wr_pool != NULL) { 2551 if (spdk_mempool_count(rtransport->data_wr_pool) != 2552 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2553 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2554 spdk_mempool_count(rtransport->data_wr_pool), 2555 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2556 } 2557 } 2558 2559 spdk_mempool_free(rtransport->data_wr_pool); 2560 2561 pthread_mutex_destroy(&rtransport->lock); 2562 free(rtransport); 2563 2564 return 0; 2565 } 2566 2567 static int 2568 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2569 struct spdk_nvme_transport_id *trid, 2570 bool peer); 2571 2572 static int 2573 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2574 const struct spdk_nvme_transport_id *trid) 2575 { 2576 struct spdk_nvmf_rdma_transport *rtransport; 2577 struct spdk_nvmf_rdma_device *device; 2578 struct spdk_nvmf_rdma_port *port; 2579 struct addrinfo *res; 2580 struct addrinfo hints; 2581 int family; 2582 int rc; 2583 2584 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2585 assert(rtransport->event_channel != NULL); 2586 2587 pthread_mutex_lock(&rtransport->lock); 2588 port = calloc(1, sizeof(*port)); 2589 if (!port) { 2590 SPDK_ERRLOG("Port allocation failed\n"); 2591 pthread_mutex_unlock(&rtransport->lock); 2592 return -ENOMEM; 2593 } 2594 2595 port->trid = trid; 2596 2597 switch (trid->adrfam) { 2598 case SPDK_NVMF_ADRFAM_IPV4: 2599 family = AF_INET; 2600 break; 2601 case SPDK_NVMF_ADRFAM_IPV6: 2602 family = AF_INET6; 2603 break; 2604 default: 2605 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2606 free(port); 2607 pthread_mutex_unlock(&rtransport->lock); 2608 return -EINVAL; 2609 } 2610 2611 memset(&hints, 0, sizeof(hints)); 2612 hints.ai_family = family; 2613 hints.ai_flags = AI_NUMERICSERV; 2614 hints.ai_socktype = SOCK_STREAM; 2615 hints.ai_protocol = 0; 2616 2617 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2618 if (rc) { 2619 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2620 free(port); 2621 pthread_mutex_unlock(&rtransport->lock); 2622 return -EINVAL; 2623 } 2624 2625 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2626 if (rc < 0) { 2627 SPDK_ERRLOG("rdma_create_id() failed\n"); 2628 freeaddrinfo(res); 2629 free(port); 2630 pthread_mutex_unlock(&rtransport->lock); 2631 return rc; 2632 } 2633 2634 rc = rdma_bind_addr(port->id, res->ai_addr); 2635 freeaddrinfo(res); 2636 2637 if (rc < 0) { 2638 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2639 rdma_destroy_id(port->id); 2640 free(port); 2641 pthread_mutex_unlock(&rtransport->lock); 2642 return rc; 2643 } 2644 2645 if (!port->id->verbs) { 2646 SPDK_ERRLOG("ibv_context is null\n"); 2647 rdma_destroy_id(port->id); 2648 free(port); 2649 pthread_mutex_unlock(&rtransport->lock); 2650 return -1; 2651 } 2652 2653 rc = rdma_listen(port->id, transport->opts.acceptor_backlog); 2654 if (rc < 0) { 2655 SPDK_ERRLOG("rdma_listen() failed\n"); 2656 rdma_destroy_id(port->id); 2657 free(port); 2658 pthread_mutex_unlock(&rtransport->lock); 2659 return rc; 2660 } 2661 2662 TAILQ_FOREACH(device, &rtransport->devices, link) { 2663 if (device->context == port->id->verbs) { 2664 port->device = device; 2665 break; 2666 } 2667 } 2668 if (!port->device) { 2669 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2670 port->id->verbs); 2671 rdma_destroy_id(port->id); 2672 free(port); 2673 pthread_mutex_unlock(&rtransport->lock); 2674 return -EINVAL; 2675 } 2676 2677 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2678 trid->traddr, trid->trsvcid); 2679 2680 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2681 pthread_mutex_unlock(&rtransport->lock); 2682 return 0; 2683 } 2684 2685 static void 2686 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2687 const struct spdk_nvme_transport_id *trid) 2688 { 2689 struct spdk_nvmf_rdma_transport *rtransport; 2690 struct spdk_nvmf_rdma_port *port, *tmp; 2691 2692 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2693 2694 pthread_mutex_lock(&rtransport->lock); 2695 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2696 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2697 TAILQ_REMOVE(&rtransport->ports, port, link); 2698 rdma_destroy_id(port->id); 2699 free(port); 2700 break; 2701 } 2702 } 2703 2704 pthread_mutex_unlock(&rtransport->lock); 2705 } 2706 2707 static void 2708 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2709 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2710 { 2711 struct spdk_nvmf_request *req, *tmp; 2712 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2713 struct spdk_nvmf_rdma_resources *resources; 2714 2715 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2716 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2717 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2718 break; 2719 } 2720 } 2721 2722 /* Then RDMA writes since reads have stronger restrictions than writes */ 2723 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2724 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2725 break; 2726 } 2727 } 2728 2729 /* The second highest priority is I/O waiting on memory buffers. */ 2730 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2731 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2732 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2733 break; 2734 } 2735 } 2736 2737 resources = rqpair->resources; 2738 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2739 rdma_req = STAILQ_FIRST(&resources->free_queue); 2740 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2741 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2742 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2743 2744 if (rqpair->srq != NULL) { 2745 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2746 rdma_req->recv->qpair->qd++; 2747 } else { 2748 rqpair->qd++; 2749 } 2750 2751 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2752 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2753 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2754 break; 2755 } 2756 } 2757 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2758 rqpair->poller->stat.pending_free_request++; 2759 } 2760 } 2761 2762 static void 2763 _nvmf_rdma_qpair_disconnect(void *ctx) 2764 { 2765 struct spdk_nvmf_qpair *qpair = ctx; 2766 2767 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2768 } 2769 2770 static void 2771 _nvmf_rdma_try_disconnect(void *ctx) 2772 { 2773 struct spdk_nvmf_qpair *qpair = ctx; 2774 struct spdk_nvmf_poll_group *group; 2775 2776 /* Read the group out of the qpair. This is normally set and accessed only from 2777 * the thread that created the group. Here, we're not on that thread necessarily. 2778 * The data member qpair->group begins it's life as NULL and then is assigned to 2779 * a pointer and never changes. So fortunately reading this and checking for 2780 * non-NULL is thread safe in the x86_64 memory model. */ 2781 group = qpair->group; 2782 2783 if (group == NULL) { 2784 /* The qpair hasn't been assigned to a group yet, so we can't 2785 * process a disconnect. Send a message to ourself and try again. */ 2786 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2787 return; 2788 } 2789 2790 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2791 } 2792 2793 static inline void 2794 nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2795 { 2796 if (!__atomic_test_and_set(&rqpair->disconnect_started, __ATOMIC_RELAXED)) { 2797 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2798 } 2799 } 2800 2801 static void nvmf_rdma_destroy_drained_qpair(void *ctx) 2802 { 2803 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2804 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2805 struct spdk_nvmf_rdma_transport, transport); 2806 2807 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2808 if (rqpair->current_send_depth != 0) { 2809 return; 2810 } 2811 2812 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2813 return; 2814 } 2815 2816 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) { 2817 return; 2818 } 2819 2820 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2821 2822 /* Qpair will be destroyed after nvmf layer closes this qpair */ 2823 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ERROR) { 2824 return; 2825 } 2826 2827 nvmf_rdma_qpair_destroy(rqpair); 2828 } 2829 2830 2831 static int 2832 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2833 { 2834 struct spdk_nvmf_qpair *qpair; 2835 struct spdk_nvmf_rdma_qpair *rqpair; 2836 2837 if (evt->id == NULL) { 2838 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2839 return -1; 2840 } 2841 2842 qpair = evt->id->context; 2843 if (qpair == NULL) { 2844 SPDK_ERRLOG("disconnect request: no active connection\n"); 2845 return -1; 2846 } 2847 2848 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2849 2850 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2851 2852 nvmf_rdma_start_disconnect(rqpair); 2853 2854 return 0; 2855 } 2856 2857 #ifdef DEBUG 2858 static const char *CM_EVENT_STR[] = { 2859 "RDMA_CM_EVENT_ADDR_RESOLVED", 2860 "RDMA_CM_EVENT_ADDR_ERROR", 2861 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2862 "RDMA_CM_EVENT_ROUTE_ERROR", 2863 "RDMA_CM_EVENT_CONNECT_REQUEST", 2864 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2865 "RDMA_CM_EVENT_CONNECT_ERROR", 2866 "RDMA_CM_EVENT_UNREACHABLE", 2867 "RDMA_CM_EVENT_REJECTED", 2868 "RDMA_CM_EVENT_ESTABLISHED", 2869 "RDMA_CM_EVENT_DISCONNECTED", 2870 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2871 "RDMA_CM_EVENT_MULTICAST_JOIN", 2872 "RDMA_CM_EVENT_MULTICAST_ERROR", 2873 "RDMA_CM_EVENT_ADDR_CHANGE", 2874 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2875 }; 2876 #endif /* DEBUG */ 2877 2878 static void 2879 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2880 struct spdk_nvmf_rdma_port *port) 2881 { 2882 struct spdk_nvmf_rdma_poll_group *rgroup; 2883 struct spdk_nvmf_rdma_poller *rpoller; 2884 struct spdk_nvmf_rdma_qpair *rqpair; 2885 2886 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2887 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2888 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 2889 if (rqpair->listen_id == port->id) { 2890 nvmf_rdma_start_disconnect(rqpair); 2891 } 2892 } 2893 } 2894 } 2895 } 2896 2897 static bool 2898 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2899 struct rdma_cm_event *event) 2900 { 2901 const struct spdk_nvme_transport_id *trid; 2902 struct spdk_nvmf_rdma_port *port; 2903 struct spdk_nvmf_rdma_transport *rtransport; 2904 bool event_acked = false; 2905 2906 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2907 TAILQ_FOREACH(port, &rtransport->ports, link) { 2908 if (port->id == event->id) { 2909 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2910 rdma_ack_cm_event(event); 2911 event_acked = true; 2912 trid = port->trid; 2913 break; 2914 } 2915 } 2916 2917 if (event_acked) { 2918 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2919 2920 nvmf_rdma_stop_listen(transport, trid); 2921 nvmf_rdma_listen(transport, trid); 2922 } 2923 2924 return event_acked; 2925 } 2926 2927 static void 2928 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2929 struct rdma_cm_event *event) 2930 { 2931 struct spdk_nvmf_rdma_port *port; 2932 struct spdk_nvmf_rdma_transport *rtransport; 2933 2934 port = event->id->context; 2935 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2936 2937 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2938 2939 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2940 2941 rdma_ack_cm_event(event); 2942 2943 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2944 ; 2945 } 2946 } 2947 2948 static void 2949 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2950 { 2951 struct spdk_nvmf_rdma_transport *rtransport; 2952 struct rdma_cm_event *event; 2953 int rc; 2954 bool event_acked; 2955 2956 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2957 2958 if (rtransport->event_channel == NULL) { 2959 return; 2960 } 2961 2962 while (1) { 2963 event_acked = false; 2964 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2965 if (rc) { 2966 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2967 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2968 } 2969 break; 2970 } 2971 2972 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2973 2974 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2975 2976 switch (event->event) { 2977 case RDMA_CM_EVENT_ADDR_RESOLVED: 2978 case RDMA_CM_EVENT_ADDR_ERROR: 2979 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2980 case RDMA_CM_EVENT_ROUTE_ERROR: 2981 /* No action required. The target never attempts to resolve routes. */ 2982 break; 2983 case RDMA_CM_EVENT_CONNECT_REQUEST: 2984 rc = nvmf_rdma_connect(transport, event); 2985 if (rc < 0) { 2986 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2987 break; 2988 } 2989 break; 2990 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2991 /* The target never initiates a new connection. So this will not occur. */ 2992 break; 2993 case RDMA_CM_EVENT_CONNECT_ERROR: 2994 /* Can this happen? The docs say it can, but not sure what causes it. */ 2995 break; 2996 case RDMA_CM_EVENT_UNREACHABLE: 2997 case RDMA_CM_EVENT_REJECTED: 2998 /* These only occur on the client side. */ 2999 break; 3000 case RDMA_CM_EVENT_ESTABLISHED: 3001 /* TODO: Should we be waiting for this event anywhere? */ 3002 break; 3003 case RDMA_CM_EVENT_DISCONNECTED: 3004 rc = nvmf_rdma_disconnect(event); 3005 if (rc < 0) { 3006 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3007 break; 3008 } 3009 break; 3010 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3011 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3012 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3013 * Once these events are sent to SPDK, we should release all IB resources and 3014 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3015 * ibv_destory* functions to release user space memory allocated by IB. All kernel 3016 * resources are already cleaned. */ 3017 if (event->id->qp) { 3018 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3019 * corresponding qpair. Otherwise the event refers to a listening device */ 3020 rc = nvmf_rdma_disconnect(event); 3021 if (rc < 0) { 3022 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3023 break; 3024 } 3025 } else { 3026 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3027 event_acked = true; 3028 } 3029 break; 3030 case RDMA_CM_EVENT_MULTICAST_JOIN: 3031 case RDMA_CM_EVENT_MULTICAST_ERROR: 3032 /* Multicast is not used */ 3033 break; 3034 case RDMA_CM_EVENT_ADDR_CHANGE: 3035 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3036 break; 3037 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3038 /* For now, do nothing. The target never re-uses queue pairs. */ 3039 break; 3040 default: 3041 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3042 break; 3043 } 3044 if (!event_acked) { 3045 rdma_ack_cm_event(event); 3046 } 3047 } 3048 } 3049 3050 static void 3051 nvmf_rdma_handle_qp_fatal(struct spdk_nvmf_rdma_qpair *rqpair) 3052 { 3053 nvmf_rdma_update_ibv_state(rqpair); 3054 nvmf_rdma_start_disconnect(rqpair); 3055 } 3056 3057 static void 3058 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3059 { 3060 rqpair->last_wqe_reached = true; 3061 nvmf_rdma_destroy_drained_qpair(rqpair); 3062 } 3063 3064 static void 3065 nvmf_rdma_handle_sq_drained(struct spdk_nvmf_rdma_qpair *rqpair) 3066 { 3067 nvmf_rdma_start_disconnect(rqpair); 3068 } 3069 3070 static void 3071 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3072 { 3073 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3074 3075 if (event_ctx->rqpair) { 3076 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3077 if (event_ctx->cb_fn) { 3078 event_ctx->cb_fn(event_ctx->rqpair); 3079 } 3080 } 3081 free(event_ctx); 3082 } 3083 3084 static int 3085 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3086 spdk_nvmf_rdma_qpair_ibv_event fn) 3087 { 3088 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3089 struct spdk_thread *thr = NULL; 3090 int rc; 3091 3092 if (rqpair->qpair.group) { 3093 thr = rqpair->qpair.group->thread; 3094 } else if (rqpair->destruct_channel) { 3095 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3096 } 3097 3098 if (!thr) { 3099 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "rqpair %p has no thread\n", rqpair); 3100 return -EINVAL; 3101 } 3102 3103 ctx = calloc(1, sizeof(*ctx)); 3104 if (!ctx) { 3105 return -ENOMEM; 3106 } 3107 3108 ctx->rqpair = rqpair; 3109 ctx->cb_fn = fn; 3110 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3111 3112 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3113 if (rc) { 3114 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3115 free(ctx); 3116 } 3117 3118 return rc; 3119 } 3120 3121 static void 3122 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3123 { 3124 int rc; 3125 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3126 struct ibv_async_event event; 3127 3128 rc = ibv_get_async_event(device->context, &event); 3129 3130 if (rc) { 3131 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 3132 errno, spdk_strerror(errno)); 3133 return; 3134 } 3135 3136 switch (event.event_type) { 3137 case IBV_EVENT_QP_FATAL: 3138 rqpair = event.element.qp->qp_context; 3139 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3140 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3141 (uintptr_t)rqpair->cm_id, event.event_type); 3142 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_qp_fatal); 3143 if (rc) { 3144 SPDK_WARNLOG("Failed to send QP_FATAL event. rqpair %p, err %d\n", rqpair, rc); 3145 nvmf_rdma_handle_qp_fatal(rqpair); 3146 } 3147 break; 3148 case IBV_EVENT_QP_LAST_WQE_REACHED: 3149 /* This event only occurs for shared receive queues. */ 3150 rqpair = event.element.qp->qp_context; 3151 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last WQE reached event received for rqpair %p\n", rqpair); 3152 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3153 if (rc) { 3154 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3155 rqpair->last_wqe_reached = true; 3156 } 3157 break; 3158 case IBV_EVENT_SQ_DRAINED: 3159 /* This event occurs frequently in both error and non-error states. 3160 * Check if the qpair is in an error state before sending a message. */ 3161 rqpair = event.element.qp->qp_context; 3162 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last sq drained event received for rqpair %p\n", rqpair); 3163 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3164 (uintptr_t)rqpair->cm_id, event.event_type); 3165 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3166 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_sq_drained); 3167 if (rc) { 3168 SPDK_WARNLOG("Failed to send SQ_DRAINED event. rqpair %p, err %d\n", rqpair, rc); 3169 nvmf_rdma_handle_sq_drained(rqpair); 3170 } 3171 } 3172 break; 3173 case IBV_EVENT_QP_REQ_ERR: 3174 case IBV_EVENT_QP_ACCESS_ERR: 3175 case IBV_EVENT_COMM_EST: 3176 case IBV_EVENT_PATH_MIG: 3177 case IBV_EVENT_PATH_MIG_ERR: 3178 SPDK_NOTICELOG("Async event: %s\n", 3179 ibv_event_type_str(event.event_type)); 3180 rqpair = event.element.qp->qp_context; 3181 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3182 (uintptr_t)rqpair->cm_id, event.event_type); 3183 nvmf_rdma_update_ibv_state(rqpair); 3184 break; 3185 case IBV_EVENT_CQ_ERR: 3186 case IBV_EVENT_DEVICE_FATAL: 3187 case IBV_EVENT_PORT_ACTIVE: 3188 case IBV_EVENT_PORT_ERR: 3189 case IBV_EVENT_LID_CHANGE: 3190 case IBV_EVENT_PKEY_CHANGE: 3191 case IBV_EVENT_SM_CHANGE: 3192 case IBV_EVENT_SRQ_ERR: 3193 case IBV_EVENT_SRQ_LIMIT_REACHED: 3194 case IBV_EVENT_CLIENT_REREGISTER: 3195 case IBV_EVENT_GID_CHANGE: 3196 default: 3197 SPDK_NOTICELOG("Async event: %s\n", 3198 ibv_event_type_str(event.event_type)); 3199 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3200 break; 3201 } 3202 ibv_ack_async_event(&event); 3203 } 3204 3205 static uint32_t 3206 nvmf_rdma_accept(struct spdk_nvmf_transport *transport) 3207 { 3208 int nfds, i = 0; 3209 struct spdk_nvmf_rdma_transport *rtransport; 3210 struct spdk_nvmf_rdma_device *device, *tmp; 3211 uint32_t count; 3212 3213 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3214 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3215 3216 if (nfds <= 0) { 3217 return 0; 3218 } 3219 3220 /* The first poll descriptor is RDMA CM event */ 3221 if (rtransport->poll_fds[i++].revents & POLLIN) { 3222 nvmf_process_cm_event(transport); 3223 nfds--; 3224 } 3225 3226 if (nfds == 0) { 3227 return count; 3228 } 3229 3230 /* Second and subsequent poll descriptors are IB async events */ 3231 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3232 if (rtransport->poll_fds[i++].revents & POLLIN) { 3233 nvmf_process_ib_event(device); 3234 nfds--; 3235 } 3236 } 3237 /* check all flagged fd's have been served */ 3238 assert(nfds == 0); 3239 3240 return count; 3241 } 3242 3243 static void 3244 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3245 struct spdk_nvmf_ctrlr_data *cdata) 3246 { 3247 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3248 3249 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3250 since in-capsule data only works with NVME drives that support SGL memory layout */ 3251 if (transport->opts.dif_insert_or_strip) { 3252 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3253 } 3254 } 3255 3256 static void 3257 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3258 struct spdk_nvme_transport_id *trid, 3259 struct spdk_nvmf_discovery_log_page_entry *entry) 3260 { 3261 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3262 entry->adrfam = trid->adrfam; 3263 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3264 3265 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3266 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3267 3268 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3269 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3270 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3271 } 3272 3273 static void 3274 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3275 3276 static struct spdk_nvmf_transport_poll_group * 3277 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3278 { 3279 struct spdk_nvmf_rdma_transport *rtransport; 3280 struct spdk_nvmf_rdma_poll_group *rgroup; 3281 struct spdk_nvmf_rdma_poller *poller; 3282 struct spdk_nvmf_rdma_device *device; 3283 struct ibv_srq_init_attr srq_init_attr; 3284 struct spdk_nvmf_rdma_resource_opts opts; 3285 int num_cqe; 3286 3287 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3288 3289 rgroup = calloc(1, sizeof(*rgroup)); 3290 if (!rgroup) { 3291 return NULL; 3292 } 3293 3294 TAILQ_INIT(&rgroup->pollers); 3295 STAILQ_INIT(&rgroup->retired_bufs); 3296 3297 pthread_mutex_lock(&rtransport->lock); 3298 TAILQ_FOREACH(device, &rtransport->devices, link) { 3299 poller = calloc(1, sizeof(*poller)); 3300 if (!poller) { 3301 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3302 nvmf_rdma_poll_group_destroy(&rgroup->group); 3303 pthread_mutex_unlock(&rtransport->lock); 3304 return NULL; 3305 } 3306 3307 poller->device = device; 3308 poller->group = rgroup; 3309 3310 TAILQ_INIT(&poller->qpairs); 3311 STAILQ_INIT(&poller->qpairs_pending_send); 3312 STAILQ_INIT(&poller->qpairs_pending_recv); 3313 3314 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3315 if (transport->opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3316 poller->max_srq_depth = transport->opts.max_srq_depth; 3317 3318 device->num_srq++; 3319 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 3320 srq_init_attr.attr.max_wr = poller->max_srq_depth; 3321 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3322 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 3323 if (!poller->srq) { 3324 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3325 nvmf_rdma_poll_group_destroy(&rgroup->group); 3326 pthread_mutex_unlock(&rtransport->lock); 3327 return NULL; 3328 } 3329 3330 opts.qp = poller->srq; 3331 opts.pd = device->pd; 3332 opts.qpair = NULL; 3333 opts.shared = true; 3334 opts.max_queue_depth = poller->max_srq_depth; 3335 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3336 3337 poller->resources = nvmf_rdma_resources_create(&opts); 3338 if (!poller->resources) { 3339 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3340 nvmf_rdma_poll_group_destroy(&rgroup->group); 3341 pthread_mutex_unlock(&rtransport->lock); 3342 return NULL; 3343 } 3344 } 3345 3346 /* 3347 * When using an srq, we can limit the completion queue at startup. 3348 * The following formula represents the calculation: 3349 * num_cqe = num_recv + num_data_wr + num_send_wr. 3350 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3351 */ 3352 if (poller->srq) { 3353 num_cqe = poller->max_srq_depth * 3; 3354 } else { 3355 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3356 } 3357 3358 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3359 if (!poller->cq) { 3360 SPDK_ERRLOG("Unable to create completion queue\n"); 3361 nvmf_rdma_poll_group_destroy(&rgroup->group); 3362 pthread_mutex_unlock(&rtransport->lock); 3363 return NULL; 3364 } 3365 poller->num_cqe = num_cqe; 3366 } 3367 3368 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3369 if (rtransport->conn_sched.next_admin_pg == NULL) { 3370 rtransport->conn_sched.next_admin_pg = rgroup; 3371 rtransport->conn_sched.next_io_pg = rgroup; 3372 } 3373 3374 pthread_mutex_unlock(&rtransport->lock); 3375 return &rgroup->group; 3376 } 3377 3378 static struct spdk_nvmf_transport_poll_group * 3379 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3380 { 3381 struct spdk_nvmf_rdma_transport *rtransport; 3382 struct spdk_nvmf_rdma_poll_group **pg; 3383 struct spdk_nvmf_transport_poll_group *result; 3384 3385 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3386 3387 pthread_mutex_lock(&rtransport->lock); 3388 3389 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3390 pthread_mutex_unlock(&rtransport->lock); 3391 return NULL; 3392 } 3393 3394 if (qpair->qid == 0) { 3395 pg = &rtransport->conn_sched.next_admin_pg; 3396 } else { 3397 pg = &rtransport->conn_sched.next_io_pg; 3398 } 3399 3400 assert(*pg != NULL); 3401 3402 result = &(*pg)->group; 3403 3404 *pg = TAILQ_NEXT(*pg, link); 3405 if (*pg == NULL) { 3406 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3407 } 3408 3409 pthread_mutex_unlock(&rtransport->lock); 3410 3411 return result; 3412 } 3413 3414 static void 3415 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3416 { 3417 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3418 struct spdk_nvmf_rdma_poller *poller, *tmp; 3419 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3420 struct spdk_nvmf_transport_pg_cache_buf *buf, *tmp_buf; 3421 struct spdk_nvmf_rdma_transport *rtransport; 3422 3423 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3424 if (!rgroup) { 3425 return; 3426 } 3427 3428 /* free all retired buffers back to the transport so we don't short the mempool. */ 3429 STAILQ_FOREACH_SAFE(buf, &rgroup->retired_bufs, link, tmp_buf) { 3430 STAILQ_REMOVE(&rgroup->retired_bufs, buf, spdk_nvmf_transport_pg_cache_buf, link); 3431 assert(group->transport != NULL); 3432 spdk_mempool_put(group->transport->data_buf_pool, buf); 3433 } 3434 3435 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3436 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3437 3438 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3439 nvmf_rdma_qpair_destroy(qpair); 3440 } 3441 3442 if (poller->srq) { 3443 if (poller->resources) { 3444 nvmf_rdma_resources_destroy(poller->resources); 3445 } 3446 ibv_destroy_srq(poller->srq); 3447 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3448 } 3449 3450 if (poller->cq) { 3451 ibv_destroy_cq(poller->cq); 3452 } 3453 3454 free(poller); 3455 } 3456 3457 if (rgroup->group.transport == NULL) { 3458 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3459 * calls this function directly in a failure path. */ 3460 free(rgroup); 3461 return; 3462 } 3463 3464 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3465 3466 pthread_mutex_lock(&rtransport->lock); 3467 next_rgroup = TAILQ_NEXT(rgroup, link); 3468 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3469 if (next_rgroup == NULL) { 3470 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3471 } 3472 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3473 rtransport->conn_sched.next_admin_pg = next_rgroup; 3474 } 3475 if (rtransport->conn_sched.next_io_pg == rgroup) { 3476 rtransport->conn_sched.next_io_pg = next_rgroup; 3477 } 3478 pthread_mutex_unlock(&rtransport->lock); 3479 3480 free(rgroup); 3481 } 3482 3483 static void 3484 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3485 { 3486 if (rqpair->cm_id != NULL) { 3487 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3488 } 3489 nvmf_rdma_qpair_destroy(rqpair); 3490 } 3491 3492 static int 3493 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3494 struct spdk_nvmf_qpair *qpair) 3495 { 3496 struct spdk_nvmf_rdma_poll_group *rgroup; 3497 struct spdk_nvmf_rdma_qpair *rqpair; 3498 struct spdk_nvmf_rdma_device *device; 3499 struct spdk_nvmf_rdma_poller *poller; 3500 int rc; 3501 3502 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3503 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3504 3505 device = rqpair->device; 3506 3507 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3508 if (poller->device == device) { 3509 break; 3510 } 3511 } 3512 3513 if (!poller) { 3514 SPDK_ERRLOG("No poller found for device.\n"); 3515 return -1; 3516 } 3517 3518 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3519 rqpair->poller = poller; 3520 rqpair->srq = rqpair->poller->srq; 3521 3522 rc = nvmf_rdma_qpair_initialize(qpair); 3523 if (rc < 0) { 3524 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3525 return -1; 3526 } 3527 3528 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3529 if (rc) { 3530 /* Try to reject, but we probably can't */ 3531 nvmf_rdma_qpair_reject_connection(rqpair); 3532 return -1; 3533 } 3534 3535 nvmf_rdma_update_ibv_state(rqpair); 3536 3537 return 0; 3538 } 3539 3540 static int 3541 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3542 struct spdk_nvmf_qpair *qpair) 3543 { 3544 struct spdk_nvmf_rdma_qpair *rqpair; 3545 3546 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3547 assert(group->transport->tgt != NULL); 3548 3549 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3550 3551 if (!rqpair->destruct_channel) { 3552 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3553 return 0; 3554 } 3555 3556 /* Sanity check that we get io_channel on the correct thread */ 3557 if (qpair->group) { 3558 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3559 } 3560 3561 return 0; 3562 } 3563 3564 static int 3565 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3566 { 3567 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3568 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3569 struct spdk_nvmf_rdma_transport, transport); 3570 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3571 struct spdk_nvmf_rdma_qpair, qpair); 3572 3573 /* 3574 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3575 * needs to be returned to the shared receive queue or the poll group will eventually be 3576 * starved of RECV structures. 3577 */ 3578 if (rqpair->srq && rdma_req->recv) { 3579 int rc; 3580 struct ibv_recv_wr *bad_recv_wr; 3581 3582 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 3583 if (rc) { 3584 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3585 } 3586 } 3587 3588 _nvmf_rdma_request_free(rdma_req, rtransport); 3589 return 0; 3590 } 3591 3592 static int 3593 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3594 { 3595 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3596 struct spdk_nvmf_rdma_transport, transport); 3597 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3598 struct spdk_nvmf_rdma_request, req); 3599 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3600 struct spdk_nvmf_rdma_qpair, qpair); 3601 3602 if (rqpair->ibv_state != IBV_QPS_ERR) { 3603 /* The connection is alive, so process the request as normal */ 3604 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3605 } else { 3606 /* The connection is dead. Move the request directly to the completed state. */ 3607 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3608 } 3609 3610 nvmf_rdma_request_process(rtransport, rdma_req); 3611 3612 return 0; 3613 } 3614 3615 static int 3616 nvmf_rdma_destroy_defunct_qpair(void *ctx) 3617 { 3618 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 3619 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3620 struct spdk_nvmf_rdma_transport, transport); 3621 3622 SPDK_INFOLOG(SPDK_LOG_RDMA, "QP#%d hasn't been drained as expected, manually destroy it\n", 3623 rqpair->qpair.qid); 3624 3625 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3626 nvmf_rdma_qpair_destroy(rqpair); 3627 3628 return SPDK_POLLER_BUSY; 3629 } 3630 3631 static void 3632 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3633 { 3634 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3635 3636 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 3637 return; 3638 } 3639 3640 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 3641 3642 /* This happens only when the qpair is disconnected before 3643 * it is added to the poll group. Since there is no poll group, 3644 * the RDMA qp has not been initialized yet and the RDMA CM 3645 * event has not yet been acknowledged, so we need to reject it. 3646 */ 3647 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3648 nvmf_rdma_qpair_reject_connection(rqpair); 3649 return; 3650 } 3651 3652 if (rqpair->rdma_qp) { 3653 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3654 } 3655 3656 rqpair->destruct_poller = SPDK_POLLER_REGISTER(nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 3657 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 3658 } 3659 3660 static struct spdk_nvmf_rdma_qpair * 3661 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3662 { 3663 struct spdk_nvmf_rdma_qpair *rqpair; 3664 /* @todo: improve QP search */ 3665 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3666 if (wc->qp_num == rqpair->rdma_qp->qp->qp_num) { 3667 return rqpair; 3668 } 3669 } 3670 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3671 return NULL; 3672 } 3673 3674 #ifdef DEBUG 3675 static int 3676 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3677 { 3678 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3679 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3680 } 3681 #endif 3682 3683 static void 3684 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3685 int rc) 3686 { 3687 struct spdk_nvmf_rdma_recv *rdma_recv; 3688 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3689 3690 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3691 while (bad_recv_wr != NULL) { 3692 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3693 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3694 3695 rdma_recv->qpair->current_recv_depth++; 3696 bad_recv_wr = bad_recv_wr->next; 3697 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3698 nvmf_rdma_start_disconnect(rdma_recv->qpair); 3699 } 3700 } 3701 3702 static void 3703 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3704 { 3705 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3706 while (bad_recv_wr != NULL) { 3707 bad_recv_wr = bad_recv_wr->next; 3708 rqpair->current_recv_depth++; 3709 } 3710 nvmf_rdma_start_disconnect(rqpair); 3711 } 3712 3713 static void 3714 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3715 struct spdk_nvmf_rdma_poller *rpoller) 3716 { 3717 struct spdk_nvmf_rdma_qpair *rqpair; 3718 struct ibv_recv_wr *bad_recv_wr; 3719 int rc; 3720 3721 if (rpoller->srq) { 3722 if (rpoller->resources->recvs_to_post.first != NULL) { 3723 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3724 if (rc) { 3725 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3726 } 3727 rpoller->resources->recvs_to_post.first = NULL; 3728 rpoller->resources->recvs_to_post.last = NULL; 3729 } 3730 } else { 3731 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3732 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3733 assert(rqpair->resources->recvs_to_post.first != NULL); 3734 rc = ibv_post_recv(rqpair->rdma_qp->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3735 if (rc) { 3736 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3737 } 3738 rqpair->resources->recvs_to_post.first = NULL; 3739 rqpair->resources->recvs_to_post.last = NULL; 3740 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3741 } 3742 } 3743 } 3744 3745 static void 3746 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3747 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3748 { 3749 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3750 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3751 3752 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3753 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3754 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3755 assert(rqpair->current_send_depth > 0); 3756 rqpair->current_send_depth--; 3757 switch (bad_rdma_wr->type) { 3758 case RDMA_WR_TYPE_DATA: 3759 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3760 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3761 assert(rqpair->current_read_depth > 0); 3762 rqpair->current_read_depth--; 3763 } 3764 break; 3765 case RDMA_WR_TYPE_SEND: 3766 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3767 break; 3768 default: 3769 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3770 prev_rdma_req = cur_rdma_req; 3771 continue; 3772 } 3773 3774 if (prev_rdma_req == cur_rdma_req) { 3775 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3776 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3777 continue; 3778 } 3779 3780 switch (cur_rdma_req->state) { 3781 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3782 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3783 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3784 break; 3785 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3786 case RDMA_REQUEST_STATE_COMPLETING: 3787 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3788 break; 3789 default: 3790 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3791 cur_rdma_req->state, rqpair); 3792 continue; 3793 } 3794 3795 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3796 prev_rdma_req = cur_rdma_req; 3797 } 3798 3799 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3800 /* Disconnect the connection. */ 3801 nvmf_rdma_start_disconnect(rqpair); 3802 } 3803 3804 } 3805 3806 static void 3807 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3808 struct spdk_nvmf_rdma_poller *rpoller) 3809 { 3810 struct spdk_nvmf_rdma_qpair *rqpair; 3811 struct ibv_send_wr *bad_wr = NULL; 3812 int rc; 3813 3814 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3815 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3816 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3817 3818 /* bad wr always points to the first wr that failed. */ 3819 if (rc) { 3820 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3821 } 3822 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3823 } 3824 } 3825 3826 static int 3827 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3828 struct spdk_nvmf_rdma_poller *rpoller) 3829 { 3830 struct ibv_wc wc[32]; 3831 struct spdk_nvmf_rdma_wr *rdma_wr; 3832 struct spdk_nvmf_rdma_request *rdma_req; 3833 struct spdk_nvmf_rdma_recv *rdma_recv; 3834 struct spdk_nvmf_rdma_qpair *rqpair; 3835 int reaped, i; 3836 int count = 0; 3837 bool error = false; 3838 uint64_t poll_tsc = spdk_get_ticks(); 3839 3840 /* Poll for completing operations. */ 3841 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3842 if (reaped < 0) { 3843 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3844 errno, spdk_strerror(errno)); 3845 return -1; 3846 } 3847 3848 rpoller->stat.polls++; 3849 rpoller->stat.completions += reaped; 3850 3851 for (i = 0; i < reaped; i++) { 3852 3853 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3854 3855 switch (rdma_wr->type) { 3856 case RDMA_WR_TYPE_SEND: 3857 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3858 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3859 3860 if (!wc[i].status) { 3861 count++; 3862 assert(wc[i].opcode == IBV_WC_SEND); 3863 assert(nvmf_rdma_req_is_completing(rdma_req)); 3864 } 3865 3866 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3867 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3868 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3869 rdma_req->num_outstanding_data_wr = 0; 3870 3871 nvmf_rdma_request_process(rtransport, rdma_req); 3872 break; 3873 case RDMA_WR_TYPE_RECV: 3874 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3875 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3876 if (rpoller->srq != NULL) { 3877 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3878 /* It is possible that there are still some completions for destroyed QP 3879 * associated with SRQ. We just ignore these late completions and re-post 3880 * receive WRs back to SRQ. 3881 */ 3882 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3883 struct ibv_recv_wr *bad_wr; 3884 int rc; 3885 3886 rdma_recv->wr.next = NULL; 3887 rc = ibv_post_srq_recv(rpoller->srq, 3888 &rdma_recv->wr, 3889 &bad_wr); 3890 if (rc) { 3891 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3892 } 3893 continue; 3894 } 3895 } 3896 rqpair = rdma_recv->qpair; 3897 3898 assert(rqpair != NULL); 3899 if (!wc[i].status) { 3900 assert(wc[i].opcode == IBV_WC_RECV); 3901 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3902 nvmf_rdma_start_disconnect(rqpair); 3903 break; 3904 } 3905 } 3906 3907 rdma_recv->wr.next = NULL; 3908 rqpair->current_recv_depth++; 3909 rdma_recv->receive_tsc = poll_tsc; 3910 rpoller->stat.requests++; 3911 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3912 break; 3913 case RDMA_WR_TYPE_DATA: 3914 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3915 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3916 3917 assert(rdma_req->num_outstanding_data_wr > 0); 3918 3919 rqpair->current_send_depth--; 3920 rdma_req->num_outstanding_data_wr--; 3921 if (!wc[i].status) { 3922 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3923 rqpair->current_read_depth--; 3924 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3925 if (rdma_req->num_outstanding_data_wr == 0) { 3926 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3927 nvmf_rdma_request_process(rtransport, rdma_req); 3928 } 3929 } else { 3930 /* If the data transfer fails still force the queue into the error state, 3931 * if we were performing an RDMA_READ, we need to force the request into a 3932 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3933 * case, we should wait for the SEND to complete. */ 3934 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3935 rqpair->current_read_depth--; 3936 if (rdma_req->num_outstanding_data_wr == 0) { 3937 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3938 } 3939 } 3940 } 3941 break; 3942 default: 3943 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3944 continue; 3945 } 3946 3947 /* Handle error conditions */ 3948 if (wc[i].status) { 3949 if ((rdma_wr->type == RDMA_WR_TYPE_RECV && !rpoller->srq)) { 3950 /* When we don't use SRQ and close a qpair, we will receive completions with error 3951 * status for all posted ibv_recv_wrs. This is expected and we don't want to log 3952 * an error in that case. */ 3953 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Error on CQ %p, request 0x%lu, type %d, status: (%d): %s\n", 3954 rpoller->cq, wc[i].wr_id, rdma_wr->type, wc[i].status, ibv_wc_status_str(wc[i].status)); 3955 } else { 3956 SPDK_ERRLOG("Error on CQ %p, request 0x%lu, type %d, status: (%d): %s\n", 3957 rpoller->cq, wc[i].wr_id, rdma_wr->type, wc[i].status, ibv_wc_status_str(wc[i].status)); 3958 } 3959 3960 error = true; 3961 3962 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3963 /* Disconnect the connection. */ 3964 nvmf_rdma_start_disconnect(rqpair); 3965 } else { 3966 nvmf_rdma_destroy_drained_qpair(rqpair); 3967 } 3968 continue; 3969 } 3970 3971 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3972 3973 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3974 nvmf_rdma_destroy_drained_qpair(rqpair); 3975 } 3976 } 3977 3978 if (error == true) { 3979 return -1; 3980 } 3981 3982 /* submit outstanding work requests. */ 3983 _poller_submit_recvs(rtransport, rpoller); 3984 _poller_submit_sends(rtransport, rpoller); 3985 3986 return count; 3987 } 3988 3989 static int 3990 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3991 { 3992 struct spdk_nvmf_rdma_transport *rtransport; 3993 struct spdk_nvmf_rdma_poll_group *rgroup; 3994 struct spdk_nvmf_rdma_poller *rpoller; 3995 int count, rc; 3996 3997 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3998 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3999 4000 count = 0; 4001 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4002 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4003 if (rc < 0) { 4004 return rc; 4005 } 4006 count += rc; 4007 } 4008 4009 return count; 4010 } 4011 4012 static int 4013 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4014 struct spdk_nvme_transport_id *trid, 4015 bool peer) 4016 { 4017 struct sockaddr *saddr; 4018 uint16_t port; 4019 4020 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4021 4022 if (peer) { 4023 saddr = rdma_get_peer_addr(id); 4024 } else { 4025 saddr = rdma_get_local_addr(id); 4026 } 4027 switch (saddr->sa_family) { 4028 case AF_INET: { 4029 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4030 4031 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4032 inet_ntop(AF_INET, &saddr_in->sin_addr, 4033 trid->traddr, sizeof(trid->traddr)); 4034 if (peer) { 4035 port = ntohs(rdma_get_dst_port(id)); 4036 } else { 4037 port = ntohs(rdma_get_src_port(id)); 4038 } 4039 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4040 break; 4041 } 4042 case AF_INET6: { 4043 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4044 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4045 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4046 trid->traddr, sizeof(trid->traddr)); 4047 if (peer) { 4048 port = ntohs(rdma_get_dst_port(id)); 4049 } else { 4050 port = ntohs(rdma_get_src_port(id)); 4051 } 4052 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4053 break; 4054 } 4055 default: 4056 return -1; 4057 4058 } 4059 4060 return 0; 4061 } 4062 4063 static int 4064 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4065 struct spdk_nvme_transport_id *trid) 4066 { 4067 struct spdk_nvmf_rdma_qpair *rqpair; 4068 4069 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4070 4071 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4072 } 4073 4074 static int 4075 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4076 struct spdk_nvme_transport_id *trid) 4077 { 4078 struct spdk_nvmf_rdma_qpair *rqpair; 4079 4080 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4081 4082 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4083 } 4084 4085 static int 4086 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4087 struct spdk_nvme_transport_id *trid) 4088 { 4089 struct spdk_nvmf_rdma_qpair *rqpair; 4090 4091 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4092 4093 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4094 } 4095 4096 void 4097 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4098 { 4099 g_nvmf_hooks = *hooks; 4100 } 4101 4102 static void 4103 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4104 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4105 { 4106 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4107 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4108 4109 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4110 4111 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4112 } 4113 4114 static int 4115 _nvmf_rdma_qpair_abort_request(void *ctx) 4116 { 4117 struct spdk_nvmf_request *req = ctx; 4118 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4119 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4120 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4121 struct spdk_nvmf_rdma_qpair, qpair); 4122 int rc; 4123 4124 spdk_poller_unregister(&req->poller); 4125 4126 switch (rdma_req_to_abort->state) { 4127 case RDMA_REQUEST_STATE_EXECUTING: 4128 rc = nvmf_ctrlr_abort_request(req); 4129 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4130 return SPDK_POLLER_BUSY; 4131 } 4132 break; 4133 4134 case RDMA_REQUEST_STATE_NEED_BUFFER: 4135 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4136 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4137 4138 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4139 break; 4140 4141 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4142 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4143 spdk_nvmf_rdma_request, state_link); 4144 4145 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4146 break; 4147 4148 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4149 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4150 spdk_nvmf_rdma_request, state_link); 4151 4152 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4153 break; 4154 4155 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4156 if (spdk_get_ticks() < req->timeout_tsc) { 4157 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4158 return SPDK_POLLER_BUSY; 4159 } 4160 break; 4161 4162 default: 4163 break; 4164 } 4165 4166 spdk_nvmf_request_complete(req); 4167 return SPDK_POLLER_BUSY; 4168 } 4169 4170 static void 4171 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4172 struct spdk_nvmf_request *req) 4173 { 4174 struct spdk_nvmf_rdma_qpair *rqpair; 4175 struct spdk_nvmf_rdma_transport *rtransport; 4176 struct spdk_nvmf_transport *transport; 4177 uint16_t cid; 4178 uint32_t i; 4179 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL; 4180 4181 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4182 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4183 transport = &rtransport->transport; 4184 4185 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4186 4187 for (i = 0; i < rqpair->max_queue_depth; i++) { 4188 rdma_req_to_abort = &rqpair->resources->reqs[i]; 4189 4190 if (rdma_req_to_abort->state != RDMA_REQUEST_STATE_FREE && 4191 rdma_req_to_abort->req.cmd->nvme_cmd.cid == cid) { 4192 break; 4193 } 4194 } 4195 4196 if (rdma_req_to_abort == NULL) { 4197 spdk_nvmf_request_complete(req); 4198 return; 4199 } 4200 4201 req->req_to_abort = &rdma_req_to_abort->req; 4202 req->timeout_tsc = spdk_get_ticks() + 4203 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4204 req->poller = NULL; 4205 4206 _nvmf_rdma_qpair_abort_request(req); 4207 } 4208 4209 static int 4210 nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4211 struct spdk_nvmf_transport_poll_group_stat **stat) 4212 { 4213 struct spdk_io_channel *ch; 4214 struct spdk_nvmf_poll_group *group; 4215 struct spdk_nvmf_transport_poll_group *tgroup; 4216 struct spdk_nvmf_rdma_poll_group *rgroup; 4217 struct spdk_nvmf_rdma_poller *rpoller; 4218 struct spdk_nvmf_rdma_device_stat *device_stat; 4219 uint64_t num_devices = 0; 4220 4221 if (tgt == NULL || stat == NULL) { 4222 return -EINVAL; 4223 } 4224 4225 ch = spdk_get_io_channel(tgt); 4226 group = spdk_io_channel_get_ctx(ch);; 4227 spdk_put_io_channel(ch); 4228 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4229 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4230 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4231 if (!*stat) { 4232 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4233 return -ENOMEM; 4234 } 4235 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4236 4237 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4238 /* Count devices to allocate enough memory */ 4239 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4240 ++num_devices; 4241 } 4242 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4243 if (!(*stat)->rdma.devices) { 4244 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4245 free(*stat); 4246 return -ENOMEM; 4247 } 4248 4249 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4250 (*stat)->rdma.num_devices = num_devices; 4251 num_devices = 0; 4252 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4253 device_stat = &(*stat)->rdma.devices[num_devices++]; 4254 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4255 device_stat->polls = rpoller->stat.polls; 4256 device_stat->completions = rpoller->stat.completions; 4257 device_stat->requests = rpoller->stat.requests; 4258 device_stat->request_latency = rpoller->stat.request_latency; 4259 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4260 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4261 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4262 } 4263 return 0; 4264 } 4265 } 4266 return -ENOENT; 4267 } 4268 4269 static void 4270 nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4271 { 4272 if (stat) { 4273 free(stat->rdma.devices); 4274 } 4275 free(stat); 4276 } 4277 4278 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4279 .name = "RDMA", 4280 .type = SPDK_NVME_TRANSPORT_RDMA, 4281 .opts_init = nvmf_rdma_opts_init, 4282 .create = nvmf_rdma_create, 4283 .destroy = nvmf_rdma_destroy, 4284 4285 .listen = nvmf_rdma_listen, 4286 .stop_listen = nvmf_rdma_stop_listen, 4287 .accept = nvmf_rdma_accept, 4288 .cdata_init = nvmf_rdma_cdata_init, 4289 4290 .listener_discover = nvmf_rdma_discover, 4291 4292 .poll_group_create = nvmf_rdma_poll_group_create, 4293 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4294 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4295 .poll_group_add = nvmf_rdma_poll_group_add, 4296 .poll_group_remove = nvmf_rdma_poll_group_remove, 4297 .poll_group_poll = nvmf_rdma_poll_group_poll, 4298 4299 .req_free = nvmf_rdma_request_free, 4300 .req_complete = nvmf_rdma_request_complete, 4301 4302 .qpair_fini = nvmf_rdma_close_qpair, 4303 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4304 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4305 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4306 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4307 4308 .poll_group_get_stat = nvmf_rdma_poll_group_get_stat, 4309 .poll_group_free_stat = nvmf_rdma_poll_group_free_stat, 4310 }; 4311 4312 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4313 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 4314