1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2019, 2020 Mellanox Technologies LTD. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of Intel Corporation nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include "spdk/stdinc.h" 35 36 #include "spdk/config.h" 37 #include "spdk/thread.h" 38 #include "spdk/likely.h" 39 #include "spdk/nvmf_transport.h" 40 #include "spdk/string.h" 41 #include "spdk/trace.h" 42 #include "spdk/util.h" 43 44 #include "spdk_internal/assert.h" 45 #include "spdk/log.h" 46 #include "spdk_internal/rdma.h" 47 48 #include "nvmf_internal.h" 49 50 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 51 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 52 53 /* 54 RDMA Connection Resource Defaults 55 */ 56 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 57 #define NVMF_DEFAULT_RSP_SGE 1 58 #define NVMF_DEFAULT_RX_SGE 2 59 60 /* The RDMA completion queue size */ 61 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 62 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 63 64 /* rxe driver vendor_id has been changed from 0 to 0XFFFFFF in 0184afd15a141d7ce24c32c0d86a1e3ba6bc0eb3 */ 65 #define NVMF_RXE_VENDOR_ID_OLD 0 66 #define NVMF_RXE_VENDOR_ID_NEW 0XFFFFFF 67 68 static int g_spdk_nvmf_ibv_query_mask = 69 IBV_QP_STATE | 70 IBV_QP_PKEY_INDEX | 71 IBV_QP_PORT | 72 IBV_QP_ACCESS_FLAGS | 73 IBV_QP_AV | 74 IBV_QP_PATH_MTU | 75 IBV_QP_DEST_QPN | 76 IBV_QP_RQ_PSN | 77 IBV_QP_MAX_DEST_RD_ATOMIC | 78 IBV_QP_MIN_RNR_TIMER | 79 IBV_QP_SQ_PSN | 80 IBV_QP_TIMEOUT | 81 IBV_QP_RETRY_CNT | 82 IBV_QP_RNR_RETRY | 83 IBV_QP_MAX_QP_RD_ATOMIC; 84 85 enum spdk_nvmf_rdma_request_state { 86 /* The request is not currently in use */ 87 RDMA_REQUEST_STATE_FREE = 0, 88 89 /* Initial state when request first received */ 90 RDMA_REQUEST_STATE_NEW, 91 92 /* The request is queued until a data buffer is available. */ 93 RDMA_REQUEST_STATE_NEED_BUFFER, 94 95 /* The request is waiting on RDMA queue depth availability 96 * to transfer data from the host to the controller. 97 */ 98 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 99 100 /* The request is currently transferring data from the host to the controller. */ 101 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 102 103 /* The request is ready to execute at the block device */ 104 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 105 106 /* The request is currently executing at the block device */ 107 RDMA_REQUEST_STATE_EXECUTING, 108 109 /* The request finished executing at the block device */ 110 RDMA_REQUEST_STATE_EXECUTED, 111 112 /* The request is waiting on RDMA queue depth availability 113 * to transfer data from the controller to the host. 114 */ 115 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 116 117 /* The request is ready to send a completion */ 118 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 119 120 /* The request is currently transferring data from the controller to the host. */ 121 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 122 123 /* The request currently has an outstanding completion without an 124 * associated data transfer. 125 */ 126 RDMA_REQUEST_STATE_COMPLETING, 127 128 /* The request completed and can be marked free. */ 129 RDMA_REQUEST_STATE_COMPLETED, 130 131 /* Terminator */ 132 RDMA_REQUEST_NUM_STATES, 133 }; 134 135 #define OBJECT_NVMF_RDMA_IO 0x40 136 137 #define TRACE_GROUP_NVMF_RDMA 0x4 138 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 139 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 140 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 141 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 142 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 143 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 144 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 145 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 146 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 147 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 148 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 149 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 150 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 151 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 152 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 153 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 154 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 155 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 156 157 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 158 { 159 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 160 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 161 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 162 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 163 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 164 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 165 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 166 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 167 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 168 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 169 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 170 spdk_trace_register_description("RDMA_REQ_TX_H2C", 171 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 172 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 173 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 174 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 175 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 176 spdk_trace_register_description("RDMA_REQ_EXECUTING", 177 TRACE_RDMA_REQUEST_STATE_EXECUTING, 178 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 179 spdk_trace_register_description("RDMA_REQ_EXECUTED", 180 TRACE_RDMA_REQUEST_STATE_EXECUTED, 181 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 182 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 183 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 184 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 185 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 186 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 187 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 188 spdk_trace_register_description("RDMA_REQ_COMPLETING", 189 TRACE_RDMA_REQUEST_STATE_COMPLETING, 190 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 191 spdk_trace_register_description("RDMA_REQ_COMPLETED", 192 TRACE_RDMA_REQUEST_STATE_COMPLETED, 193 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 194 195 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 196 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 197 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 198 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 199 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 200 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 201 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 202 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 203 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 204 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 205 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 206 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 207 } 208 209 enum spdk_nvmf_rdma_wr_type { 210 RDMA_WR_TYPE_RECV, 211 RDMA_WR_TYPE_SEND, 212 RDMA_WR_TYPE_DATA, 213 }; 214 215 struct spdk_nvmf_rdma_wr { 216 enum spdk_nvmf_rdma_wr_type type; 217 }; 218 219 /* This structure holds commands as they are received off the wire. 220 * It must be dynamically paired with a full request object 221 * (spdk_nvmf_rdma_request) to service a request. It is separate 222 * from the request because RDMA does not appear to order 223 * completions, so occasionally we'll get a new incoming 224 * command when there aren't any free request objects. 225 */ 226 struct spdk_nvmf_rdma_recv { 227 struct ibv_recv_wr wr; 228 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 229 230 struct spdk_nvmf_rdma_qpair *qpair; 231 232 /* In-capsule data buffer */ 233 uint8_t *buf; 234 235 struct spdk_nvmf_rdma_wr rdma_wr; 236 uint64_t receive_tsc; 237 238 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 239 }; 240 241 struct spdk_nvmf_rdma_request_data { 242 struct spdk_nvmf_rdma_wr rdma_wr; 243 struct ibv_send_wr wr; 244 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 245 }; 246 247 struct spdk_nvmf_rdma_request { 248 struct spdk_nvmf_request req; 249 250 enum spdk_nvmf_rdma_request_state state; 251 252 struct spdk_nvmf_rdma_recv *recv; 253 254 struct { 255 struct spdk_nvmf_rdma_wr rdma_wr; 256 struct ibv_send_wr wr; 257 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 258 } rsp; 259 260 struct spdk_nvmf_rdma_request_data data; 261 262 uint32_t iovpos; 263 264 uint32_t num_outstanding_data_wr; 265 uint64_t receive_tsc; 266 267 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 268 }; 269 270 struct spdk_nvmf_rdma_resource_opts { 271 struct spdk_nvmf_rdma_qpair *qpair; 272 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 273 void *qp; 274 struct ibv_pd *pd; 275 uint32_t max_queue_depth; 276 uint32_t in_capsule_data_size; 277 bool shared; 278 }; 279 280 struct spdk_nvmf_recv_wr_list { 281 struct ibv_recv_wr *first; 282 struct ibv_recv_wr *last; 283 }; 284 285 struct spdk_nvmf_rdma_resources { 286 /* Array of size "max_queue_depth" containing RDMA requests. */ 287 struct spdk_nvmf_rdma_request *reqs; 288 289 /* Array of size "max_queue_depth" containing RDMA recvs. */ 290 struct spdk_nvmf_rdma_recv *recvs; 291 292 /* Array of size "max_queue_depth" containing 64 byte capsules 293 * used for receive. 294 */ 295 union nvmf_h2c_msg *cmds; 296 struct ibv_mr *cmds_mr; 297 298 /* Array of size "max_queue_depth" containing 16 byte completions 299 * to be sent back to the user. 300 */ 301 union nvmf_c2h_msg *cpls; 302 struct ibv_mr *cpls_mr; 303 304 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 305 * buffers to be used for in capsule data. 306 */ 307 void *bufs; 308 struct ibv_mr *bufs_mr; 309 310 /* The list of pending recvs to transfer */ 311 struct spdk_nvmf_recv_wr_list recvs_to_post; 312 313 /* Receives that are waiting for a request object */ 314 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 315 316 /* Queue to track free requests */ 317 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 318 }; 319 320 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 321 322 struct spdk_nvmf_rdma_ibv_event_ctx { 323 struct spdk_nvmf_rdma_qpair *rqpair; 324 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 325 /* Link to other ibv events associated with this qpair */ 326 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 327 }; 328 329 struct spdk_nvmf_rdma_qpair { 330 struct spdk_nvmf_qpair qpair; 331 332 struct spdk_nvmf_rdma_device *device; 333 struct spdk_nvmf_rdma_poller *poller; 334 335 struct spdk_rdma_qp *rdma_qp; 336 struct rdma_cm_id *cm_id; 337 struct ibv_srq *srq; 338 struct rdma_cm_id *listen_id; 339 340 /* The maximum number of I/O outstanding on this connection at one time */ 341 uint16_t max_queue_depth; 342 343 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 344 uint16_t max_read_depth; 345 346 /* The maximum number of RDMA SEND operations at one time */ 347 uint32_t max_send_depth; 348 349 /* The current number of outstanding WRs from this qpair's 350 * recv queue. Should not exceed device->attr.max_queue_depth. 351 */ 352 uint16_t current_recv_depth; 353 354 /* The current number of active RDMA READ operations */ 355 uint16_t current_read_depth; 356 357 /* The current number of posted WRs from this qpair's 358 * send queue. Should not exceed max_send_depth. 359 */ 360 uint32_t current_send_depth; 361 362 /* The maximum number of SGEs per WR on the send queue */ 363 uint32_t max_send_sge; 364 365 /* The maximum number of SGEs per WR on the recv queue */ 366 uint32_t max_recv_sge; 367 368 struct spdk_nvmf_rdma_resources *resources; 369 370 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 371 372 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 373 374 /* Number of requests not in the free state */ 375 uint32_t qd; 376 377 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 378 379 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 380 381 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 382 383 /* IBV queue pair attributes: they are used to manage 384 * qp state and recover from errors. 385 */ 386 enum ibv_qp_state ibv_state; 387 388 /* 389 * io_channel which is used to destroy qpair when it is removed from poll group 390 */ 391 struct spdk_io_channel *destruct_channel; 392 393 /* List of ibv async events */ 394 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 395 396 /* Lets us know that we have received the last_wqe event. */ 397 bool last_wqe_reached; 398 399 /* Indicate that nvmf_rdma_close_qpair is called */ 400 bool to_close; 401 }; 402 403 struct spdk_nvmf_rdma_poller_stat { 404 uint64_t completions; 405 uint64_t polls; 406 uint64_t requests; 407 uint64_t request_latency; 408 uint64_t pending_free_request; 409 uint64_t pending_rdma_read; 410 uint64_t pending_rdma_write; 411 }; 412 413 struct spdk_nvmf_rdma_poller { 414 struct spdk_nvmf_rdma_device *device; 415 struct spdk_nvmf_rdma_poll_group *group; 416 417 int num_cqe; 418 int required_num_wr; 419 struct ibv_cq *cq; 420 421 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 422 uint16_t max_srq_depth; 423 424 /* Shared receive queue */ 425 struct ibv_srq *srq; 426 427 struct spdk_nvmf_rdma_resources *resources; 428 struct spdk_nvmf_rdma_poller_stat stat; 429 430 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 431 432 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 433 434 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 435 436 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 437 }; 438 439 struct spdk_nvmf_rdma_poll_group_stat { 440 uint64_t pending_data_buffer; 441 }; 442 443 struct spdk_nvmf_rdma_poll_group { 444 struct spdk_nvmf_transport_poll_group group; 445 struct spdk_nvmf_rdma_poll_group_stat stat; 446 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 447 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 448 }; 449 450 struct spdk_nvmf_rdma_conn_sched { 451 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 452 struct spdk_nvmf_rdma_poll_group *next_io_pg; 453 }; 454 455 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 456 struct spdk_nvmf_rdma_device { 457 struct ibv_device_attr attr; 458 struct ibv_context *context; 459 460 struct spdk_rdma_mem_map *map; 461 struct ibv_pd *pd; 462 463 int num_srq; 464 465 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 466 }; 467 468 struct spdk_nvmf_rdma_port { 469 const struct spdk_nvme_transport_id *trid; 470 struct rdma_cm_id *id; 471 struct spdk_nvmf_rdma_device *device; 472 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 473 }; 474 475 struct rdma_transport_opts { 476 uint32_t max_srq_depth; 477 bool no_srq; 478 bool no_wr_batching; 479 int acceptor_backlog; 480 }; 481 482 struct spdk_nvmf_rdma_transport { 483 struct spdk_nvmf_transport transport; 484 struct rdma_transport_opts rdma_opts; 485 486 struct spdk_nvmf_rdma_conn_sched conn_sched; 487 488 struct rdma_event_channel *event_channel; 489 490 struct spdk_mempool *data_wr_pool; 491 492 pthread_mutex_t lock; 493 494 /* fields used to poll RDMA/IB events */ 495 nfds_t npoll_fds; 496 struct pollfd *poll_fds; 497 498 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 499 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 500 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 501 }; 502 503 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 504 { 505 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 506 spdk_json_decode_uint32, true 507 }, 508 { 509 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 510 spdk_json_decode_bool, true 511 }, 512 { 513 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 514 spdk_json_decode_bool, true 515 }, 516 { 517 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 518 spdk_json_decode_int32, true 519 }, 520 }; 521 522 static bool 523 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 524 struct spdk_nvmf_rdma_request *rdma_req); 525 526 static void 527 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 528 struct spdk_nvmf_rdma_poller *rpoller); 529 530 static void 531 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 532 struct spdk_nvmf_rdma_poller *rpoller); 533 534 static inline int 535 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 536 { 537 switch (state) { 538 case IBV_QPS_RESET: 539 case IBV_QPS_INIT: 540 case IBV_QPS_RTR: 541 case IBV_QPS_RTS: 542 case IBV_QPS_SQD: 543 case IBV_QPS_SQE: 544 case IBV_QPS_ERR: 545 return 0; 546 default: 547 return -1; 548 } 549 } 550 551 static inline enum spdk_nvme_media_error_status_code 552 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 553 enum spdk_nvme_media_error_status_code result; 554 switch (err_type) 555 { 556 case SPDK_DIF_REFTAG_ERROR: 557 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 558 break; 559 case SPDK_DIF_APPTAG_ERROR: 560 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 561 break; 562 case SPDK_DIF_GUARD_ERROR: 563 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 564 break; 565 default: 566 SPDK_UNREACHABLE(); 567 } 568 569 return result; 570 } 571 572 static enum ibv_qp_state 573 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 574 enum ibv_qp_state old_state, new_state; 575 struct ibv_qp_attr qp_attr; 576 struct ibv_qp_init_attr init_attr; 577 int rc; 578 579 old_state = rqpair->ibv_state; 580 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 581 g_spdk_nvmf_ibv_query_mask, &init_attr); 582 583 if (rc) 584 { 585 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 586 return IBV_QPS_ERR + 1; 587 } 588 589 new_state = qp_attr.qp_state; 590 rqpair->ibv_state = new_state; 591 qp_attr.ah_attr.port_num = qp_attr.port_num; 592 593 rc = nvmf_rdma_check_ibv_state(new_state); 594 if (rc) 595 { 596 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 597 /* 598 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 599 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 600 */ 601 return IBV_QPS_ERR + 1; 602 } 603 604 if (old_state != new_state) 605 { 606 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 607 (uintptr_t)rqpair->cm_id, new_state); 608 } 609 return new_state; 610 } 611 612 static void 613 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 614 struct spdk_nvmf_rdma_transport *rtransport) 615 { 616 struct spdk_nvmf_rdma_request_data *data_wr; 617 struct ibv_send_wr *next_send_wr; 618 uint64_t req_wrid; 619 620 rdma_req->num_outstanding_data_wr = 0; 621 data_wr = &rdma_req->data; 622 req_wrid = data_wr->wr.wr_id; 623 while (data_wr && data_wr->wr.wr_id == req_wrid) { 624 memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge); 625 data_wr->wr.num_sge = 0; 626 next_send_wr = data_wr->wr.next; 627 if (data_wr != &rdma_req->data) { 628 spdk_mempool_put(rtransport->data_wr_pool, data_wr); 629 } 630 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : 631 SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr); 632 } 633 } 634 635 static void 636 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 637 { 638 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 639 if (req->req.cmd) { 640 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 641 } 642 if (req->recv) { 643 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 644 } 645 } 646 647 static void 648 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 649 { 650 int i; 651 652 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 653 for (i = 0; i < rqpair->max_queue_depth; i++) { 654 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 655 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 656 } 657 } 658 } 659 660 static void 661 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 662 { 663 if (resources->cmds_mr) { 664 ibv_dereg_mr(resources->cmds_mr); 665 } 666 667 if (resources->cpls_mr) { 668 ibv_dereg_mr(resources->cpls_mr); 669 } 670 671 if (resources->bufs_mr) { 672 ibv_dereg_mr(resources->bufs_mr); 673 } 674 675 spdk_free(resources->cmds); 676 spdk_free(resources->cpls); 677 spdk_free(resources->bufs); 678 free(resources->reqs); 679 free(resources->recvs); 680 free(resources); 681 } 682 683 684 static struct spdk_nvmf_rdma_resources * 685 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 686 { 687 struct spdk_nvmf_rdma_resources *resources; 688 struct spdk_nvmf_rdma_request *rdma_req; 689 struct spdk_nvmf_rdma_recv *rdma_recv; 690 struct ibv_qp *qp; 691 struct ibv_srq *srq; 692 uint32_t i; 693 int rc; 694 695 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 696 if (!resources) { 697 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 698 return NULL; 699 } 700 701 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 702 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 703 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 704 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 705 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 706 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 707 708 if (opts->in_capsule_data_size > 0) { 709 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 710 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 711 SPDK_MALLOC_DMA); 712 } 713 714 if (!resources->reqs || !resources->recvs || !resources->cmds || 715 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 716 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 717 goto cleanup; 718 } 719 720 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 721 opts->max_queue_depth * sizeof(*resources->cmds), 722 IBV_ACCESS_LOCAL_WRITE); 723 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 724 opts->max_queue_depth * sizeof(*resources->cpls), 725 0); 726 727 if (opts->in_capsule_data_size) { 728 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 729 opts->max_queue_depth * 730 opts->in_capsule_data_size, 731 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 732 } 733 734 if (!resources->cmds_mr || !resources->cpls_mr || 735 (opts->in_capsule_data_size && 736 !resources->bufs_mr)) { 737 goto cleanup; 738 } 739 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx LKey: %x\n", 740 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 741 resources->cmds_mr->lkey); 742 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx LKey: %x\n", 743 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 744 resources->cpls_mr->lkey); 745 if (resources->bufs && resources->bufs_mr) { 746 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x LKey: %x\n", 747 resources->bufs, opts->max_queue_depth * 748 opts->in_capsule_data_size, resources->bufs_mr->lkey); 749 } 750 751 /* Initialize queues */ 752 STAILQ_INIT(&resources->incoming_queue); 753 STAILQ_INIT(&resources->free_queue); 754 755 for (i = 0; i < opts->max_queue_depth; i++) { 756 struct ibv_recv_wr *bad_wr = NULL; 757 758 rdma_recv = &resources->recvs[i]; 759 rdma_recv->qpair = opts->qpair; 760 761 /* Set up memory to receive commands */ 762 if (resources->bufs) { 763 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 764 opts->in_capsule_data_size)); 765 } 766 767 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 768 769 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 770 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 771 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 772 rdma_recv->wr.num_sge = 1; 773 774 if (rdma_recv->buf && resources->bufs_mr) { 775 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 776 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 777 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 778 rdma_recv->wr.num_sge++; 779 } 780 781 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 782 rdma_recv->wr.sg_list = rdma_recv->sgl; 783 if (opts->shared) { 784 srq = (struct ibv_srq *)opts->qp; 785 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 786 } else { 787 qp = (struct ibv_qp *)opts->qp; 788 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 789 } 790 if (rc) { 791 goto cleanup; 792 } 793 } 794 795 for (i = 0; i < opts->max_queue_depth; i++) { 796 rdma_req = &resources->reqs[i]; 797 798 if (opts->qpair != NULL) { 799 rdma_req->req.qpair = &opts->qpair->qpair; 800 } else { 801 rdma_req->req.qpair = NULL; 802 } 803 rdma_req->req.cmd = NULL; 804 805 /* Set up memory to send responses */ 806 rdma_req->req.rsp = &resources->cpls[i]; 807 808 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 809 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 810 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 811 812 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 813 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 814 rdma_req->rsp.wr.next = NULL; 815 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 816 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 817 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 818 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 819 820 /* Set up memory for data buffers */ 821 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 822 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 823 rdma_req->data.wr.next = NULL; 824 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 825 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 826 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 827 828 /* Initialize request state to FREE */ 829 rdma_req->state = RDMA_REQUEST_STATE_FREE; 830 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 831 } 832 833 return resources; 834 835 cleanup: 836 nvmf_rdma_resources_destroy(resources); 837 return NULL; 838 } 839 840 static void 841 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 842 { 843 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 844 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 845 ctx->rqpair = NULL; 846 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 847 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 848 } 849 } 850 851 static void 852 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 853 { 854 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 855 struct ibv_recv_wr *bad_recv_wr = NULL; 856 int rc; 857 858 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 859 860 if (rqpair->qd != 0) { 861 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 862 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 863 struct spdk_nvmf_rdma_transport, transport); 864 struct spdk_nvmf_rdma_request *req; 865 uint32_t i, max_req_count = 0; 866 867 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 868 869 if (rqpair->srq == NULL) { 870 nvmf_rdma_dump_qpair_contents(rqpair); 871 max_req_count = rqpair->max_queue_depth; 872 } else if (rqpair->poller && rqpair->resources) { 873 max_req_count = rqpair->poller->max_srq_depth; 874 } 875 876 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 877 for (i = 0; i < max_req_count; i++) { 878 req = &rqpair->resources->reqs[i]; 879 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 880 /* nvmf_rdma_request_process checks qpair ibv and internal state 881 * and completes a request */ 882 nvmf_rdma_request_process(rtransport, req); 883 } 884 } 885 assert(rqpair->qd == 0); 886 } 887 888 if (rqpair->poller) { 889 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 890 891 if (rqpair->srq != NULL && rqpair->resources != NULL) { 892 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 893 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 894 if (rqpair == rdma_recv->qpair) { 895 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 896 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 897 if (rc) { 898 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 899 } 900 } 901 } 902 } 903 } 904 905 if (rqpair->cm_id) { 906 if (rqpair->rdma_qp != NULL) { 907 spdk_rdma_qp_destroy(rqpair->rdma_qp); 908 rqpair->rdma_qp = NULL; 909 } 910 rdma_destroy_id(rqpair->cm_id); 911 912 if (rqpair->poller != NULL && rqpair->srq == NULL) { 913 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 914 } 915 } 916 917 if (rqpair->srq == NULL && rqpair->resources != NULL) { 918 nvmf_rdma_resources_destroy(rqpair->resources); 919 } 920 921 nvmf_rdma_qpair_clean_ibv_events(rqpair); 922 923 if (rqpair->destruct_channel) { 924 spdk_put_io_channel(rqpair->destruct_channel); 925 rqpair->destruct_channel = NULL; 926 } 927 928 free(rqpair); 929 } 930 931 static int 932 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 933 { 934 struct spdk_nvmf_rdma_poller *rpoller; 935 int rc, num_cqe, required_num_wr; 936 937 /* Enlarge CQ size dynamically */ 938 rpoller = rqpair->poller; 939 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 940 num_cqe = rpoller->num_cqe; 941 if (num_cqe < required_num_wr) { 942 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 943 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 944 } 945 946 if (rpoller->num_cqe != num_cqe) { 947 if (required_num_wr > device->attr.max_cqe) { 948 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 949 required_num_wr, device->attr.max_cqe); 950 return -1; 951 } 952 953 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 954 rc = ibv_resize_cq(rpoller->cq, num_cqe); 955 if (rc) { 956 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 957 return -1; 958 } 959 960 rpoller->num_cqe = num_cqe; 961 } 962 963 rpoller->required_num_wr = required_num_wr; 964 return 0; 965 } 966 967 static int 968 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 969 { 970 struct spdk_nvmf_rdma_qpair *rqpair; 971 struct spdk_nvmf_rdma_transport *rtransport; 972 struct spdk_nvmf_transport *transport; 973 struct spdk_nvmf_rdma_resource_opts opts; 974 struct spdk_nvmf_rdma_device *device; 975 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 976 977 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 978 device = rqpair->device; 979 980 qp_init_attr.qp_context = rqpair; 981 qp_init_attr.pd = device->pd; 982 qp_init_attr.send_cq = rqpair->poller->cq; 983 qp_init_attr.recv_cq = rqpair->poller->cq; 984 985 if (rqpair->srq) { 986 qp_init_attr.srq = rqpair->srq; 987 } else { 988 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 989 } 990 991 /* SEND, READ, and WRITE operations */ 992 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 993 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 994 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 995 996 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 997 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 998 goto error; 999 } 1000 1001 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 1002 if (!rqpair->rdma_qp) { 1003 goto error; 1004 } 1005 1006 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1007 qp_init_attr.cap.max_send_wr); 1008 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1009 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1010 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 1011 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1012 1013 if (rqpair->poller->srq == NULL) { 1014 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1015 transport = &rtransport->transport; 1016 1017 opts.qp = rqpair->rdma_qp->qp; 1018 opts.pd = rqpair->cm_id->pd; 1019 opts.qpair = rqpair; 1020 opts.shared = false; 1021 opts.max_queue_depth = rqpair->max_queue_depth; 1022 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1023 1024 rqpair->resources = nvmf_rdma_resources_create(&opts); 1025 1026 if (!rqpair->resources) { 1027 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1028 rdma_destroy_qp(rqpair->cm_id); 1029 goto error; 1030 } 1031 } else { 1032 rqpair->resources = rqpair->poller->resources; 1033 } 1034 1035 rqpair->current_recv_depth = 0; 1036 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1037 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1038 1039 return 0; 1040 1041 error: 1042 rdma_destroy_id(rqpair->cm_id); 1043 rqpair->cm_id = NULL; 1044 return -1; 1045 } 1046 1047 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1048 /* This function accepts either a single wr or the first wr in a linked list. */ 1049 static void 1050 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1051 { 1052 struct ibv_recv_wr *last; 1053 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1054 struct spdk_nvmf_rdma_transport, transport); 1055 1056 last = first; 1057 while (last->next != NULL) { 1058 last = last->next; 1059 } 1060 1061 if (rqpair->resources->recvs_to_post.first == NULL) { 1062 rqpair->resources->recvs_to_post.first = first; 1063 rqpair->resources->recvs_to_post.last = last; 1064 if (rqpair->srq == NULL) { 1065 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1066 } 1067 } else { 1068 rqpair->resources->recvs_to_post.last->next = first; 1069 rqpair->resources->recvs_to_post.last = last; 1070 } 1071 1072 if (rtransport->rdma_opts.no_wr_batching) { 1073 _poller_submit_recvs(rtransport, rqpair->poller); 1074 } 1075 } 1076 1077 static int 1078 request_transfer_in(struct spdk_nvmf_request *req) 1079 { 1080 struct spdk_nvmf_rdma_request *rdma_req; 1081 struct spdk_nvmf_qpair *qpair; 1082 struct spdk_nvmf_rdma_qpair *rqpair; 1083 struct spdk_nvmf_rdma_transport *rtransport; 1084 1085 qpair = req->qpair; 1086 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1087 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1088 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1089 struct spdk_nvmf_rdma_transport, transport); 1090 1091 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1092 assert(rdma_req != NULL); 1093 1094 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1095 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1096 } 1097 if (rtransport->rdma_opts.no_wr_batching) { 1098 _poller_submit_sends(rtransport, rqpair->poller); 1099 } 1100 1101 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1102 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1103 return 0; 1104 } 1105 1106 static int 1107 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1108 { 1109 int num_outstanding_data_wr = 0; 1110 struct spdk_nvmf_rdma_request *rdma_req; 1111 struct spdk_nvmf_qpair *qpair; 1112 struct spdk_nvmf_rdma_qpair *rqpair; 1113 struct spdk_nvme_cpl *rsp; 1114 struct ibv_send_wr *first = NULL; 1115 struct spdk_nvmf_rdma_transport *rtransport; 1116 1117 *data_posted = 0; 1118 qpair = req->qpair; 1119 rsp = &req->rsp->nvme_cpl; 1120 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1121 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1122 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1123 struct spdk_nvmf_rdma_transport, transport); 1124 1125 /* Advance our sq_head pointer */ 1126 if (qpair->sq_head == qpair->sq_head_max) { 1127 qpair->sq_head = 0; 1128 } else { 1129 qpair->sq_head++; 1130 } 1131 rsp->sqhd = qpair->sq_head; 1132 1133 /* queue the capsule for the recv buffer */ 1134 assert(rdma_req->recv != NULL); 1135 1136 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1137 1138 rdma_req->recv = NULL; 1139 assert(rqpair->current_recv_depth > 0); 1140 rqpair->current_recv_depth--; 1141 1142 /* Build the response which consists of optional 1143 * RDMA WRITEs to transfer data, plus an RDMA SEND 1144 * containing the response. 1145 */ 1146 first = &rdma_req->rsp.wr; 1147 1148 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1149 /* On failure, data was not read from the controller. So clear the 1150 * number of outstanding data WRs to zero. 1151 */ 1152 rdma_req->num_outstanding_data_wr = 0; 1153 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1154 first = &rdma_req->data.wr; 1155 *data_posted = 1; 1156 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1157 } 1158 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1159 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1160 } 1161 if (rtransport->rdma_opts.no_wr_batching) { 1162 _poller_submit_sends(rtransport, rqpair->poller); 1163 } 1164 1165 /* +1 for the rsp wr */ 1166 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1167 1168 return 0; 1169 } 1170 1171 static int 1172 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1173 { 1174 struct spdk_nvmf_rdma_accept_private_data accept_data; 1175 struct rdma_conn_param ctrlr_event_data = {}; 1176 int rc; 1177 1178 accept_data.recfmt = 0; 1179 accept_data.crqsize = rqpair->max_queue_depth; 1180 1181 ctrlr_event_data.private_data = &accept_data; 1182 ctrlr_event_data.private_data_len = sizeof(accept_data); 1183 if (id->ps == RDMA_PS_TCP) { 1184 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1185 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1186 } 1187 1188 /* Configure infinite retries for the initiator side qpair. 1189 * When using a shared receive queue on the target side, 1190 * we need to pass this value to the initiator to prevent the 1191 * initiator side NIC from completing SEND requests back to the 1192 * initiator with status rnr_retry_count_exceeded. */ 1193 if (rqpair->srq != NULL) { 1194 ctrlr_event_data.rnr_retry_count = 0x7; 1195 } 1196 1197 /* When qpair is created without use of rdma cm API, an additional 1198 * information must be provided to initiator in the connection response: 1199 * whether qpair is using SRQ and its qp_num 1200 * Fields below are ignored by rdma cm if qpair has been 1201 * created using rdma cm API. */ 1202 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1203 ctrlr_event_data.qp_num = rqpair->rdma_qp->qp->qp_num; 1204 1205 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1206 if (rc) { 1207 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1208 } else { 1209 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1210 } 1211 1212 return rc; 1213 } 1214 1215 static void 1216 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1217 { 1218 struct spdk_nvmf_rdma_reject_private_data rej_data; 1219 1220 rej_data.recfmt = 0; 1221 rej_data.sts = error; 1222 1223 rdma_reject(id, &rej_data, sizeof(rej_data)); 1224 } 1225 1226 static int 1227 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1228 { 1229 struct spdk_nvmf_rdma_transport *rtransport; 1230 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1231 struct spdk_nvmf_rdma_port *port; 1232 struct rdma_conn_param *rdma_param = NULL; 1233 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1234 uint16_t max_queue_depth; 1235 uint16_t max_read_depth; 1236 1237 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1238 1239 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1240 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1241 1242 rdma_param = &event->param.conn; 1243 if (rdma_param->private_data == NULL || 1244 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1245 SPDK_ERRLOG("connect request: no private data provided\n"); 1246 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1247 return -1; 1248 } 1249 1250 private_data = rdma_param->private_data; 1251 if (private_data->recfmt != 0) { 1252 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1253 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1254 return -1; 1255 } 1256 1257 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1258 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1259 1260 port = event->listen_id->context; 1261 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1262 event->listen_id, event->listen_id->verbs, port); 1263 1264 /* Figure out the supported queue depth. This is a multi-step process 1265 * that takes into account hardware maximums, host provided values, 1266 * and our target's internal memory limits */ 1267 1268 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1269 1270 /* Start with the maximum queue depth allowed by the target */ 1271 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1272 max_read_depth = rtransport->transport.opts.max_queue_depth; 1273 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1274 rtransport->transport.opts.max_queue_depth); 1275 1276 /* Next check the local NIC's hardware limitations */ 1277 SPDK_DEBUGLOG(rdma, 1278 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1279 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1280 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1281 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1282 1283 /* Next check the remote NIC's hardware limitations */ 1284 SPDK_DEBUGLOG(rdma, 1285 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1286 rdma_param->initiator_depth, rdma_param->responder_resources); 1287 if (rdma_param->initiator_depth > 0) { 1288 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1289 } 1290 1291 /* Finally check for the host software requested values, which are 1292 * optional. */ 1293 if (rdma_param->private_data != NULL && 1294 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1295 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1296 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1297 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1298 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1299 } 1300 1301 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1302 max_queue_depth, max_read_depth); 1303 1304 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1305 if (rqpair == NULL) { 1306 SPDK_ERRLOG("Could not allocate new connection.\n"); 1307 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1308 return -1; 1309 } 1310 1311 rqpair->device = port->device; 1312 rqpair->max_queue_depth = max_queue_depth; 1313 rqpair->max_read_depth = max_read_depth; 1314 rqpair->cm_id = event->id; 1315 rqpair->listen_id = event->listen_id; 1316 rqpair->qpair.transport = transport; 1317 STAILQ_INIT(&rqpair->ibv_events); 1318 /* use qid from the private data to determine the qpair type 1319 qid will be set to the appropriate value when the controller is created */ 1320 rqpair->qpair.qid = private_data->qid; 1321 1322 event->id->context = &rqpair->qpair; 1323 1324 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1325 1326 return 0; 1327 } 1328 1329 static inline void 1330 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1331 enum spdk_nvme_data_transfer xfer) 1332 { 1333 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1334 wr->opcode = IBV_WR_RDMA_WRITE; 1335 wr->send_flags = 0; 1336 wr->next = next; 1337 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1338 wr->opcode = IBV_WR_RDMA_READ; 1339 wr->send_flags = IBV_SEND_SIGNALED; 1340 wr->next = NULL; 1341 } else { 1342 assert(0); 1343 } 1344 } 1345 1346 static int 1347 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1348 struct spdk_nvmf_rdma_request *rdma_req, 1349 uint32_t num_sgl_descriptors) 1350 { 1351 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1352 struct spdk_nvmf_rdma_request_data *current_data_wr; 1353 uint32_t i; 1354 1355 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1356 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1357 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1358 return -EINVAL; 1359 } 1360 1361 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1362 return -ENOMEM; 1363 } 1364 1365 current_data_wr = &rdma_req->data; 1366 1367 for (i = 0; i < num_sgl_descriptors; i++) { 1368 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1369 current_data_wr->wr.next = &work_requests[i]->wr; 1370 current_data_wr = work_requests[i]; 1371 current_data_wr->wr.sg_list = current_data_wr->sgl; 1372 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1373 } 1374 1375 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1376 1377 return 0; 1378 } 1379 1380 static inline void 1381 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1382 { 1383 struct ibv_send_wr *wr = &rdma_req->data.wr; 1384 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1385 1386 wr->wr.rdma.rkey = sgl->keyed.key; 1387 wr->wr.rdma.remote_addr = sgl->address; 1388 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1389 } 1390 1391 static inline void 1392 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1393 { 1394 struct ibv_send_wr *wr = &rdma_req->data.wr; 1395 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1396 uint32_t i; 1397 int j; 1398 uint64_t remote_addr_offset = 0; 1399 1400 for (i = 0; i < num_wrs; ++i) { 1401 wr->wr.rdma.rkey = sgl->keyed.key; 1402 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1403 for (j = 0; j < wr->num_sge; ++j) { 1404 remote_addr_offset += wr->sg_list[j].length; 1405 } 1406 wr = wr->next; 1407 } 1408 } 1409 1410 static bool 1411 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device, 1412 struct iovec *iov, struct ibv_send_wr **_wr, 1413 uint32_t *_remaining_data_block, uint32_t *_offset, 1414 uint32_t *_num_extra_wrs, 1415 const struct spdk_dif_ctx *dif_ctx) 1416 { 1417 struct ibv_send_wr *wr = *_wr; 1418 struct ibv_sge *sg_ele = &wr->sg_list[wr->num_sge]; 1419 struct spdk_rdma_memory_translation mem_translation; 1420 int rc; 1421 uint32_t lkey = 0; 1422 uint32_t remaining, data_block_size, md_size, sge_len; 1423 1424 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1425 if (spdk_unlikely(rc)) { 1426 return false; 1427 } 1428 1429 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1430 1431 if (spdk_likely(!dif_ctx)) { 1432 sg_ele->lkey = lkey; 1433 sg_ele->addr = (uintptr_t)(iov->iov_base); 1434 sg_ele->length = iov->iov_len; 1435 wr->num_sge++; 1436 } else { 1437 remaining = iov->iov_len - *_offset; 1438 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1439 md_size = dif_ctx->md_size; 1440 1441 while (remaining) { 1442 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1443 if (*_num_extra_wrs > 0 && wr->next) { 1444 *_wr = wr->next; 1445 wr = *_wr; 1446 wr->num_sge = 0; 1447 sg_ele = &wr->sg_list[wr->num_sge]; 1448 (*_num_extra_wrs)--; 1449 } else { 1450 break; 1451 } 1452 } 1453 sg_ele->lkey = lkey; 1454 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset); 1455 sge_len = spdk_min(remaining, *_remaining_data_block); 1456 sg_ele->length = sge_len; 1457 remaining -= sge_len; 1458 *_remaining_data_block -= sge_len; 1459 *_offset += sge_len; 1460 1461 sg_ele++; 1462 wr->num_sge++; 1463 1464 if (*_remaining_data_block == 0) { 1465 /* skip metadata */ 1466 *_offset += md_size; 1467 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1468 remaining -= spdk_min(remaining, md_size); 1469 *_remaining_data_block = data_block_size; 1470 } 1471 1472 if (remaining == 0) { 1473 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1474 the remaining metadata bits at the beginning of the next buffer */ 1475 *_offset -= iov->iov_len; 1476 } 1477 } 1478 } 1479 1480 return true; 1481 } 1482 1483 static int 1484 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1485 struct spdk_nvmf_rdma_device *device, 1486 struct spdk_nvmf_rdma_request *rdma_req, 1487 struct ibv_send_wr *wr, 1488 uint32_t length, 1489 uint32_t num_extra_wrs) 1490 { 1491 struct spdk_nvmf_request *req = &rdma_req->req; 1492 struct spdk_dif_ctx *dif_ctx = NULL; 1493 uint32_t remaining_data_block = 0; 1494 uint32_t offset = 0; 1495 1496 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 1497 dif_ctx = &rdma_req->req.dif.dif_ctx; 1498 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1499 } 1500 1501 wr->num_sge = 0; 1502 1503 while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1504 if (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr, 1505 &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) { 1506 return -EINVAL; 1507 } 1508 1509 length -= req->iov[rdma_req->iovpos].iov_len; 1510 rdma_req->iovpos++; 1511 } 1512 1513 if (length) { 1514 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1515 return -EINVAL; 1516 } 1517 1518 return 0; 1519 } 1520 1521 static inline uint32_t 1522 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1523 { 1524 /* estimate the number of SG entries and WRs needed to process the request */ 1525 uint32_t num_sge = 0; 1526 uint32_t i; 1527 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1528 1529 for (i = 0; i < num_buffers && length > 0; i++) { 1530 uint32_t buffer_len = spdk_min(length, io_unit_size); 1531 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1532 1533 if (num_sge_in_block * block_size > buffer_len) { 1534 ++num_sge_in_block; 1535 } 1536 num_sge += num_sge_in_block; 1537 length -= buffer_len; 1538 } 1539 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1540 } 1541 1542 static int 1543 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1544 struct spdk_nvmf_rdma_device *device, 1545 struct spdk_nvmf_rdma_request *rdma_req, 1546 uint32_t length) 1547 { 1548 struct spdk_nvmf_rdma_qpair *rqpair; 1549 struct spdk_nvmf_rdma_poll_group *rgroup; 1550 struct spdk_nvmf_request *req = &rdma_req->req; 1551 struct ibv_send_wr *wr = &rdma_req->data.wr; 1552 int rc; 1553 uint32_t num_wrs = 1; 1554 1555 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1556 rgroup = rqpair->poller->group; 1557 1558 /* rdma wr specifics */ 1559 nvmf_rdma_setup_request(rdma_req); 1560 1561 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1562 length); 1563 if (rc != 0) { 1564 return rc; 1565 } 1566 1567 assert(req->iovcnt <= rqpair->max_send_sge); 1568 1569 rdma_req->iovpos = 0; 1570 1571 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1572 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1573 req->dif.dif_ctx.block_size); 1574 if (num_wrs > 1) { 1575 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1576 if (rc != 0) { 1577 goto err_exit; 1578 } 1579 } 1580 } 1581 1582 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1583 if (spdk_unlikely(rc != 0)) { 1584 goto err_exit; 1585 } 1586 1587 if (spdk_unlikely(num_wrs > 1)) { 1588 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1589 } 1590 1591 /* set the number of outstanding data WRs for this request. */ 1592 rdma_req->num_outstanding_data_wr = num_wrs; 1593 1594 return rc; 1595 1596 err_exit: 1597 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1598 nvmf_rdma_request_free_data(rdma_req, rtransport); 1599 req->iovcnt = 0; 1600 return rc; 1601 } 1602 1603 static int 1604 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1605 struct spdk_nvmf_rdma_device *device, 1606 struct spdk_nvmf_rdma_request *rdma_req) 1607 { 1608 struct spdk_nvmf_rdma_qpair *rqpair; 1609 struct spdk_nvmf_rdma_poll_group *rgroup; 1610 struct ibv_send_wr *current_wr; 1611 struct spdk_nvmf_request *req = &rdma_req->req; 1612 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1613 uint32_t num_sgl_descriptors; 1614 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES]; 1615 uint32_t i; 1616 int rc; 1617 1618 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1619 rgroup = rqpair->poller->group; 1620 1621 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1622 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1623 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1624 1625 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1626 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1627 1628 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1629 return -ENOMEM; 1630 } 1631 1632 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1633 for (i = 0; i < num_sgl_descriptors; i++) { 1634 if (spdk_likely(!req->dif.dif_insert_or_strip)) { 1635 lengths[i] = desc->keyed.length; 1636 } else { 1637 req->dif.orig_length += desc->keyed.length; 1638 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1639 req->dif.elba_length += lengths[i]; 1640 } 1641 desc++; 1642 } 1643 1644 rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport, 1645 lengths, num_sgl_descriptors); 1646 if (rc != 0) { 1647 nvmf_rdma_request_free_data(rdma_req, rtransport); 1648 return rc; 1649 } 1650 1651 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1652 current_wr = &rdma_req->data.wr; 1653 assert(current_wr != NULL); 1654 1655 req->length = 0; 1656 rdma_req->iovpos = 0; 1657 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1658 for (i = 0; i < num_sgl_descriptors; i++) { 1659 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1660 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1661 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1662 rc = -EINVAL; 1663 goto err_exit; 1664 } 1665 1666 current_wr->num_sge = 0; 1667 1668 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1669 if (rc != 0) { 1670 rc = -ENOMEM; 1671 goto err_exit; 1672 } 1673 1674 req->length += desc->keyed.length; 1675 current_wr->wr.rdma.rkey = desc->keyed.key; 1676 current_wr->wr.rdma.remote_addr = desc->address; 1677 current_wr = current_wr->next; 1678 desc++; 1679 } 1680 1681 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1682 /* Go back to the last descriptor in the list. */ 1683 desc--; 1684 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1685 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1686 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1687 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1688 } 1689 } 1690 #endif 1691 1692 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1693 1694 return 0; 1695 1696 err_exit: 1697 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1698 nvmf_rdma_request_free_data(rdma_req, rtransport); 1699 return rc; 1700 } 1701 1702 static int 1703 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1704 struct spdk_nvmf_rdma_device *device, 1705 struct spdk_nvmf_rdma_request *rdma_req) 1706 { 1707 struct spdk_nvmf_request *req = &rdma_req->req; 1708 struct spdk_nvme_cpl *rsp; 1709 struct spdk_nvme_sgl_descriptor *sgl; 1710 int rc; 1711 uint32_t length; 1712 1713 rsp = &req->rsp->nvme_cpl; 1714 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1715 1716 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1717 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1718 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1719 1720 length = sgl->keyed.length; 1721 if (length > rtransport->transport.opts.max_io_size) { 1722 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1723 length, rtransport->transport.opts.max_io_size); 1724 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1725 return -1; 1726 } 1727 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1728 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1729 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1730 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1731 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1732 } 1733 } 1734 #endif 1735 1736 /* fill request length and populate iovs */ 1737 req->length = length; 1738 1739 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1740 req->dif.orig_length = length; 1741 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1742 req->dif.elba_length = length; 1743 } 1744 1745 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1746 if (spdk_unlikely(rc < 0)) { 1747 if (rc == -EINVAL) { 1748 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1749 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1750 return -1; 1751 } 1752 /* No available buffers. Queue this request up. */ 1753 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1754 return 0; 1755 } 1756 1757 /* backward compatible */ 1758 req->data = req->iov[0].iov_base; 1759 1760 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1761 req->iovcnt); 1762 1763 return 0; 1764 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1765 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1766 uint64_t offset = sgl->address; 1767 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1768 1769 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1770 offset, sgl->unkeyed.length); 1771 1772 if (offset > max_len) { 1773 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1774 offset, max_len); 1775 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1776 return -1; 1777 } 1778 max_len -= (uint32_t)offset; 1779 1780 if (sgl->unkeyed.length > max_len) { 1781 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1782 sgl->unkeyed.length, max_len); 1783 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1784 return -1; 1785 } 1786 1787 rdma_req->num_outstanding_data_wr = 0; 1788 req->data = rdma_req->recv->buf + offset; 1789 req->data_from_pool = false; 1790 req->length = sgl->unkeyed.length; 1791 1792 req->iov[0].iov_base = req->data; 1793 req->iov[0].iov_len = req->length; 1794 req->iovcnt = 1; 1795 1796 return 0; 1797 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1798 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1799 1800 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1801 if (rc == -ENOMEM) { 1802 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1803 return 0; 1804 } else if (rc == -EINVAL) { 1805 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1806 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1807 return -1; 1808 } 1809 1810 /* backward compatible */ 1811 req->data = req->iov[0].iov_base; 1812 1813 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1814 req->iovcnt); 1815 1816 return 0; 1817 } 1818 1819 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1820 sgl->generic.type, sgl->generic.subtype); 1821 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1822 return -1; 1823 } 1824 1825 static void 1826 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1827 struct spdk_nvmf_rdma_transport *rtransport) 1828 { 1829 struct spdk_nvmf_rdma_qpair *rqpair; 1830 struct spdk_nvmf_rdma_poll_group *rgroup; 1831 1832 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1833 if (rdma_req->req.data_from_pool) { 1834 rgroup = rqpair->poller->group; 1835 1836 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1837 } 1838 nvmf_rdma_request_free_data(rdma_req, rtransport); 1839 rdma_req->req.length = 0; 1840 rdma_req->req.iovcnt = 0; 1841 rdma_req->req.data = NULL; 1842 rdma_req->rsp.wr.next = NULL; 1843 rdma_req->data.wr.next = NULL; 1844 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1845 rqpair->qd--; 1846 1847 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1848 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1849 } 1850 1851 bool 1852 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1853 struct spdk_nvmf_rdma_request *rdma_req) 1854 { 1855 struct spdk_nvmf_rdma_qpair *rqpair; 1856 struct spdk_nvmf_rdma_device *device; 1857 struct spdk_nvmf_rdma_poll_group *rgroup; 1858 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1859 int rc; 1860 struct spdk_nvmf_rdma_recv *rdma_recv; 1861 enum spdk_nvmf_rdma_request_state prev_state; 1862 bool progress = false; 1863 int data_posted; 1864 uint32_t num_blocks; 1865 1866 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1867 device = rqpair->device; 1868 rgroup = rqpair->poller->group; 1869 1870 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1871 1872 /* If the queue pair is in an error state, force the request to the completed state 1873 * to release resources. */ 1874 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1875 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1876 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1877 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1878 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1879 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1880 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1881 } 1882 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1883 } 1884 1885 /* The loop here is to allow for several back-to-back state changes. */ 1886 do { 1887 prev_state = rdma_req->state; 1888 1889 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 1890 1891 switch (rdma_req->state) { 1892 case RDMA_REQUEST_STATE_FREE: 1893 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1894 * to escape this state. */ 1895 break; 1896 case RDMA_REQUEST_STATE_NEW: 1897 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1898 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1899 rdma_recv = rdma_req->recv; 1900 1901 /* The first element of the SGL is the NVMe command */ 1902 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1903 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1904 1905 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1906 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1907 break; 1908 } 1909 1910 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 1911 rdma_req->req.dif.dif_insert_or_strip = true; 1912 } 1913 1914 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1915 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1916 rdma_req->rsp.wr.imm_data = 0; 1917 #endif 1918 1919 /* The next state transition depends on the data transfer needs of this request. */ 1920 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 1921 1922 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 1923 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 1924 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 1925 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1926 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 1927 break; 1928 } 1929 1930 /* If no data to transfer, ready to execute. */ 1931 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1932 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1933 break; 1934 } 1935 1936 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1937 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 1938 break; 1939 case RDMA_REQUEST_STATE_NEED_BUFFER: 1940 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1941 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1942 1943 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1944 1945 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 1946 /* This request needs to wait in line to obtain a buffer */ 1947 break; 1948 } 1949 1950 /* Try to get a data buffer */ 1951 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1952 if (rc < 0) { 1953 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 1954 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1955 break; 1956 } 1957 1958 if (!rdma_req->req.data) { 1959 /* No buffers available. */ 1960 rgroup->stat.pending_data_buffer++; 1961 break; 1962 } 1963 1964 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 1965 1966 /* If data is transferring from host to controller and the data didn't 1967 * arrive using in capsule data, we need to do a transfer from the host. 1968 */ 1969 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 1970 rdma_req->req.data_from_pool) { 1971 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1972 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1973 break; 1974 } 1975 1976 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1977 break; 1978 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1979 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1980 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1981 1982 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1983 /* This request needs to wait in line to perform RDMA */ 1984 break; 1985 } 1986 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1987 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1988 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1989 rqpair->poller->stat.pending_rdma_read++; 1990 break; 1991 } 1992 1993 /* We have already verified that this request is the head of the queue. */ 1994 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 1995 1996 rc = request_transfer_in(&rdma_req->req); 1997 if (!rc) { 1998 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 1999 } else { 2000 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2001 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2002 } 2003 break; 2004 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2005 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2006 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2007 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2008 * to escape this state. */ 2009 break; 2010 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2011 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2012 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2013 2014 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2015 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2016 /* generate DIF for write operation */ 2017 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2018 assert(num_blocks > 0); 2019 2020 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2021 num_blocks, &rdma_req->req.dif.dif_ctx); 2022 if (rc != 0) { 2023 SPDK_ERRLOG("DIF generation failed\n"); 2024 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2025 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2026 break; 2027 } 2028 } 2029 2030 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2031 /* set extended length before IO operation */ 2032 rdma_req->req.length = rdma_req->req.dif.elba_length; 2033 } 2034 2035 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2036 spdk_nvmf_request_exec(&rdma_req->req); 2037 break; 2038 case RDMA_REQUEST_STATE_EXECUTING: 2039 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2040 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2041 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2042 * to escape this state. */ 2043 break; 2044 case RDMA_REQUEST_STATE_EXECUTED: 2045 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2046 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2047 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2048 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2049 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2050 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2051 } else { 2052 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2053 } 2054 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2055 /* restore the original length */ 2056 rdma_req->req.length = rdma_req->req.dif.orig_length; 2057 2058 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2059 struct spdk_dif_error error_blk; 2060 2061 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2062 2063 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2064 &rdma_req->req.dif.dif_ctx, &error_blk); 2065 if (rc) { 2066 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2067 2068 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2069 error_blk.err_offset); 2070 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2071 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2072 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2073 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2074 } 2075 } 2076 } 2077 break; 2078 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2079 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2080 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2081 2082 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2083 /* This request needs to wait in line to perform RDMA */ 2084 break; 2085 } 2086 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2087 rqpair->max_send_depth) { 2088 /* We can only have so many WRs outstanding. we have to wait until some finish. 2089 * +1 since each request has an additional wr in the resp. */ 2090 rqpair->poller->stat.pending_rdma_write++; 2091 break; 2092 } 2093 2094 /* We have already verified that this request is the head of the queue. */ 2095 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2096 2097 /* The data transfer will be kicked off from 2098 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2099 */ 2100 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2101 break; 2102 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2103 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2104 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2105 rc = request_transfer_out(&rdma_req->req, &data_posted); 2106 assert(rc == 0); /* No good way to handle this currently */ 2107 if (rc) { 2108 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2109 } else { 2110 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2111 RDMA_REQUEST_STATE_COMPLETING; 2112 } 2113 break; 2114 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2115 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2116 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2117 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2118 * to escape this state. */ 2119 break; 2120 case RDMA_REQUEST_STATE_COMPLETING: 2121 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2122 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2123 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2124 * to escape this state. */ 2125 break; 2126 case RDMA_REQUEST_STATE_COMPLETED: 2127 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2128 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2129 2130 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2131 _nvmf_rdma_request_free(rdma_req, rtransport); 2132 break; 2133 case RDMA_REQUEST_NUM_STATES: 2134 default: 2135 assert(0); 2136 break; 2137 } 2138 2139 if (rdma_req->state != prev_state) { 2140 progress = true; 2141 } 2142 } while (rdma_req->state != prev_state); 2143 2144 return progress; 2145 } 2146 2147 /* Public API callbacks begin here */ 2148 2149 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2150 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2151 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2152 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2153 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2154 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2155 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2156 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2157 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2158 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2159 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2160 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2161 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2162 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2163 2164 static void 2165 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2166 { 2167 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2168 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2169 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2170 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2171 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2172 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2173 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2174 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2175 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2176 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2177 opts->transport_specific = NULL; 2178 } 2179 2180 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2181 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2182 2183 static inline bool 2184 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2185 { 2186 return device->attr.vendor_id == NVMF_RXE_VENDOR_ID_OLD || 2187 device->attr.vendor_id == NVMF_RXE_VENDOR_ID_NEW; 2188 } 2189 2190 static struct spdk_nvmf_transport * 2191 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2192 { 2193 int rc; 2194 struct spdk_nvmf_rdma_transport *rtransport; 2195 struct spdk_nvmf_rdma_device *device, *tmp; 2196 struct ibv_context **contexts; 2197 uint32_t i; 2198 int flag; 2199 uint32_t sge_count; 2200 uint32_t min_shared_buffers; 2201 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2202 pthread_mutexattr_t attr; 2203 2204 rtransport = calloc(1, sizeof(*rtransport)); 2205 if (!rtransport) { 2206 return NULL; 2207 } 2208 2209 if (pthread_mutexattr_init(&attr)) { 2210 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2211 free(rtransport); 2212 return NULL; 2213 } 2214 2215 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2216 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2217 pthread_mutexattr_destroy(&attr); 2218 free(rtransport); 2219 return NULL; 2220 } 2221 2222 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2223 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2224 pthread_mutexattr_destroy(&attr); 2225 free(rtransport); 2226 return NULL; 2227 } 2228 2229 pthread_mutexattr_destroy(&attr); 2230 2231 TAILQ_INIT(&rtransport->devices); 2232 TAILQ_INIT(&rtransport->ports); 2233 TAILQ_INIT(&rtransport->poll_groups); 2234 2235 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2236 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2237 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2238 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2239 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2240 if (opts->transport_specific != NULL && 2241 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2242 SPDK_COUNTOF(rdma_transport_opts_decoder), 2243 &rtransport->rdma_opts)) { 2244 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2245 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2246 return NULL; 2247 } 2248 2249 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2250 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2251 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2252 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2253 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d," 2254 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2255 opts->max_queue_depth, 2256 opts->max_io_size, 2257 opts->max_qpairs_per_ctrlr - 1, 2258 opts->io_unit_size, 2259 opts->in_capsule_data_size, 2260 opts->max_aq_depth, 2261 opts->num_shared_buffers, 2262 rtransport->rdma_opts.max_srq_depth, 2263 rtransport->rdma_opts.no_srq, 2264 rtransport->rdma_opts.acceptor_backlog, 2265 rtransport->rdma_opts.no_wr_batching, 2266 opts->abort_timeout_sec); 2267 2268 /* I/O unit size cannot be larger than max I/O size */ 2269 if (opts->io_unit_size > opts->max_io_size) { 2270 opts->io_unit_size = opts->max_io_size; 2271 } 2272 2273 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2274 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2275 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2276 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2277 } 2278 2279 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2280 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2281 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2282 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2283 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2284 return NULL; 2285 } 2286 2287 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2288 if (min_shared_buffers > opts->num_shared_buffers) { 2289 SPDK_ERRLOG("There are not enough buffers to satisfy" 2290 "per-poll group caches for each thread. (%" PRIu32 ")" 2291 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2292 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2293 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2294 return NULL; 2295 } 2296 2297 sge_count = opts->max_io_size / opts->io_unit_size; 2298 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2299 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2300 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2301 return NULL; 2302 } 2303 2304 rtransport->event_channel = rdma_create_event_channel(); 2305 if (rtransport->event_channel == NULL) { 2306 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2307 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2308 return NULL; 2309 } 2310 2311 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2312 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2313 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2314 rtransport->event_channel->fd, spdk_strerror(errno)); 2315 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2316 return NULL; 2317 } 2318 2319 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2320 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2321 sizeof(struct spdk_nvmf_rdma_request_data), 2322 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2323 SPDK_ENV_SOCKET_ID_ANY); 2324 if (!rtransport->data_wr_pool) { 2325 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2326 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2327 return NULL; 2328 } 2329 2330 contexts = rdma_get_devices(NULL); 2331 if (contexts == NULL) { 2332 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2333 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2334 return NULL; 2335 } 2336 2337 i = 0; 2338 rc = 0; 2339 while (contexts[i] != NULL) { 2340 device = calloc(1, sizeof(*device)); 2341 if (!device) { 2342 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2343 rc = -ENOMEM; 2344 break; 2345 } 2346 device->context = contexts[i]; 2347 rc = ibv_query_device(device->context, &device->attr); 2348 if (rc < 0) { 2349 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2350 free(device); 2351 break; 2352 2353 } 2354 2355 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2356 2357 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2358 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2359 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2360 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2361 } 2362 2363 /** 2364 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2365 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2366 * but incorrectly reports that it does. There are changes making their way 2367 * through the kernel now that will enable this feature. When they are merged, 2368 * we can conditionally enable this feature. 2369 * 2370 * TODO: enable this for versions of the kernel rxe driver that support it. 2371 */ 2372 if (nvmf_rdma_is_rxe_device(device)) { 2373 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2374 } 2375 #endif 2376 2377 /* set up device context async ev fd as NON_BLOCKING */ 2378 flag = fcntl(device->context->async_fd, F_GETFL); 2379 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2380 if (rc < 0) { 2381 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2382 free(device); 2383 break; 2384 } 2385 2386 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2387 i++; 2388 2389 if (g_nvmf_hooks.get_ibv_pd) { 2390 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2391 } else { 2392 device->pd = ibv_alloc_pd(device->context); 2393 } 2394 2395 if (!device->pd) { 2396 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2397 rc = -ENOMEM; 2398 break; 2399 } 2400 2401 assert(device->map == NULL); 2402 2403 device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks); 2404 if (!device->map) { 2405 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2406 rc = -ENOMEM; 2407 break; 2408 } 2409 2410 assert(device->map != NULL); 2411 assert(device->pd != NULL); 2412 } 2413 rdma_free_devices(contexts); 2414 2415 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2416 /* divide and round up. */ 2417 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2418 2419 /* round up to the nearest 4k. */ 2420 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2421 2422 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2423 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2424 opts->io_unit_size); 2425 } 2426 2427 if (rc < 0) { 2428 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2429 return NULL; 2430 } 2431 2432 /* Set up poll descriptor array to monitor events from RDMA and IB 2433 * in a single poll syscall 2434 */ 2435 rtransport->npoll_fds = i + 1; 2436 i = 0; 2437 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2438 if (rtransport->poll_fds == NULL) { 2439 SPDK_ERRLOG("poll_fds allocation failed\n"); 2440 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2441 return NULL; 2442 } 2443 2444 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2445 rtransport->poll_fds[i++].events = POLLIN; 2446 2447 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2448 rtransport->poll_fds[i].fd = device->context->async_fd; 2449 rtransport->poll_fds[i++].events = POLLIN; 2450 } 2451 2452 return &rtransport->transport; 2453 } 2454 2455 static void 2456 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2457 { 2458 struct spdk_nvmf_rdma_transport *rtransport; 2459 assert(w != NULL); 2460 2461 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2462 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2463 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2464 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2465 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2466 } 2467 2468 static int 2469 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2470 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2471 { 2472 struct spdk_nvmf_rdma_transport *rtransport; 2473 struct spdk_nvmf_rdma_port *port, *port_tmp; 2474 struct spdk_nvmf_rdma_device *device, *device_tmp; 2475 2476 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2477 2478 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2479 TAILQ_REMOVE(&rtransport->ports, port, link); 2480 rdma_destroy_id(port->id); 2481 free(port); 2482 } 2483 2484 if (rtransport->poll_fds != NULL) { 2485 free(rtransport->poll_fds); 2486 } 2487 2488 if (rtransport->event_channel != NULL) { 2489 rdma_destroy_event_channel(rtransport->event_channel); 2490 } 2491 2492 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2493 TAILQ_REMOVE(&rtransport->devices, device, link); 2494 spdk_rdma_free_mem_map(&device->map); 2495 if (device->pd) { 2496 if (!g_nvmf_hooks.get_ibv_pd) { 2497 ibv_dealloc_pd(device->pd); 2498 } 2499 } 2500 free(device); 2501 } 2502 2503 if (rtransport->data_wr_pool != NULL) { 2504 if (spdk_mempool_count(rtransport->data_wr_pool) != 2505 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2506 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2507 spdk_mempool_count(rtransport->data_wr_pool), 2508 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2509 } 2510 } 2511 2512 spdk_mempool_free(rtransport->data_wr_pool); 2513 2514 pthread_mutex_destroy(&rtransport->lock); 2515 free(rtransport); 2516 2517 if (cb_fn) { 2518 cb_fn(cb_arg); 2519 } 2520 return 0; 2521 } 2522 2523 static int 2524 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2525 struct spdk_nvme_transport_id *trid, 2526 bool peer); 2527 2528 static int 2529 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2530 struct spdk_nvmf_listen_opts *listen_opts) 2531 { 2532 struct spdk_nvmf_rdma_transport *rtransport; 2533 struct spdk_nvmf_rdma_device *device; 2534 struct spdk_nvmf_rdma_port *port; 2535 struct addrinfo *res; 2536 struct addrinfo hints; 2537 int family; 2538 int rc; 2539 2540 if (!strlen(trid->trsvcid)) { 2541 SPDK_ERRLOG("Service id is required\n"); 2542 return -EINVAL; 2543 } 2544 2545 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2546 assert(rtransport->event_channel != NULL); 2547 2548 pthread_mutex_lock(&rtransport->lock); 2549 port = calloc(1, sizeof(*port)); 2550 if (!port) { 2551 SPDK_ERRLOG("Port allocation failed\n"); 2552 pthread_mutex_unlock(&rtransport->lock); 2553 return -ENOMEM; 2554 } 2555 2556 port->trid = trid; 2557 2558 switch (trid->adrfam) { 2559 case SPDK_NVMF_ADRFAM_IPV4: 2560 family = AF_INET; 2561 break; 2562 case SPDK_NVMF_ADRFAM_IPV6: 2563 family = AF_INET6; 2564 break; 2565 default: 2566 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2567 free(port); 2568 pthread_mutex_unlock(&rtransport->lock); 2569 return -EINVAL; 2570 } 2571 2572 memset(&hints, 0, sizeof(hints)); 2573 hints.ai_family = family; 2574 hints.ai_flags = AI_NUMERICSERV; 2575 hints.ai_socktype = SOCK_STREAM; 2576 hints.ai_protocol = 0; 2577 2578 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2579 if (rc) { 2580 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2581 free(port); 2582 pthread_mutex_unlock(&rtransport->lock); 2583 return -EINVAL; 2584 } 2585 2586 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2587 if (rc < 0) { 2588 SPDK_ERRLOG("rdma_create_id() failed\n"); 2589 freeaddrinfo(res); 2590 free(port); 2591 pthread_mutex_unlock(&rtransport->lock); 2592 return rc; 2593 } 2594 2595 rc = rdma_bind_addr(port->id, res->ai_addr); 2596 freeaddrinfo(res); 2597 2598 if (rc < 0) { 2599 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2600 rdma_destroy_id(port->id); 2601 free(port); 2602 pthread_mutex_unlock(&rtransport->lock); 2603 return rc; 2604 } 2605 2606 if (!port->id->verbs) { 2607 SPDK_ERRLOG("ibv_context is null\n"); 2608 rdma_destroy_id(port->id); 2609 free(port); 2610 pthread_mutex_unlock(&rtransport->lock); 2611 return -1; 2612 } 2613 2614 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2615 if (rc < 0) { 2616 SPDK_ERRLOG("rdma_listen() failed\n"); 2617 rdma_destroy_id(port->id); 2618 free(port); 2619 pthread_mutex_unlock(&rtransport->lock); 2620 return rc; 2621 } 2622 2623 TAILQ_FOREACH(device, &rtransport->devices, link) { 2624 if (device->context == port->id->verbs) { 2625 port->device = device; 2626 break; 2627 } 2628 } 2629 if (!port->device) { 2630 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2631 port->id->verbs); 2632 rdma_destroy_id(port->id); 2633 free(port); 2634 pthread_mutex_unlock(&rtransport->lock); 2635 return -EINVAL; 2636 } 2637 2638 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2639 trid->traddr, trid->trsvcid); 2640 2641 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2642 pthread_mutex_unlock(&rtransport->lock); 2643 return 0; 2644 } 2645 2646 static void 2647 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2648 const struct spdk_nvme_transport_id *trid) 2649 { 2650 struct spdk_nvmf_rdma_transport *rtransport; 2651 struct spdk_nvmf_rdma_port *port, *tmp; 2652 2653 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2654 2655 pthread_mutex_lock(&rtransport->lock); 2656 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2657 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2658 TAILQ_REMOVE(&rtransport->ports, port, link); 2659 rdma_destroy_id(port->id); 2660 free(port); 2661 break; 2662 } 2663 } 2664 2665 pthread_mutex_unlock(&rtransport->lock); 2666 } 2667 2668 static void 2669 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2670 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2671 { 2672 struct spdk_nvmf_request *req, *tmp; 2673 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2674 struct spdk_nvmf_rdma_resources *resources; 2675 2676 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2677 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2678 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2679 break; 2680 } 2681 } 2682 2683 /* Then RDMA writes since reads have stronger restrictions than writes */ 2684 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2685 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2686 break; 2687 } 2688 } 2689 2690 /* Then we handle request waiting on memory buffers. */ 2691 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2692 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2693 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2694 break; 2695 } 2696 } 2697 2698 resources = rqpair->resources; 2699 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2700 rdma_req = STAILQ_FIRST(&resources->free_queue); 2701 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2702 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2703 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2704 2705 if (rqpair->srq != NULL) { 2706 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2707 rdma_req->recv->qpair->qd++; 2708 } else { 2709 rqpair->qd++; 2710 } 2711 2712 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2713 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2714 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2715 break; 2716 } 2717 } 2718 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2719 rqpair->poller->stat.pending_free_request++; 2720 } 2721 } 2722 2723 static inline bool 2724 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 2725 { 2726 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 2727 return nvmf_rdma_is_rxe_device(device) || 2728 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 2729 } 2730 2731 static void 2732 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2733 { 2734 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2735 struct spdk_nvmf_rdma_transport, transport); 2736 2737 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2738 2739 /* nvmr_rdma_close_qpair is not called */ 2740 if (!rqpair->to_close) { 2741 return; 2742 } 2743 2744 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2745 if (rqpair->current_send_depth != 0) { 2746 return; 2747 } 2748 2749 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2750 return; 2751 } 2752 2753 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 2754 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 2755 return; 2756 } 2757 2758 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2759 2760 nvmf_rdma_qpair_destroy(rqpair); 2761 } 2762 2763 static int 2764 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2765 { 2766 struct spdk_nvmf_qpair *qpair; 2767 struct spdk_nvmf_rdma_qpair *rqpair; 2768 2769 if (evt->id == NULL) { 2770 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2771 return -1; 2772 } 2773 2774 qpair = evt->id->context; 2775 if (qpair == NULL) { 2776 SPDK_ERRLOG("disconnect request: no active connection\n"); 2777 return -1; 2778 } 2779 2780 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2781 2782 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2783 2784 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2785 2786 return 0; 2787 } 2788 2789 #ifdef DEBUG 2790 static const char *CM_EVENT_STR[] = { 2791 "RDMA_CM_EVENT_ADDR_RESOLVED", 2792 "RDMA_CM_EVENT_ADDR_ERROR", 2793 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2794 "RDMA_CM_EVENT_ROUTE_ERROR", 2795 "RDMA_CM_EVENT_CONNECT_REQUEST", 2796 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2797 "RDMA_CM_EVENT_CONNECT_ERROR", 2798 "RDMA_CM_EVENT_UNREACHABLE", 2799 "RDMA_CM_EVENT_REJECTED", 2800 "RDMA_CM_EVENT_ESTABLISHED", 2801 "RDMA_CM_EVENT_DISCONNECTED", 2802 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2803 "RDMA_CM_EVENT_MULTICAST_JOIN", 2804 "RDMA_CM_EVENT_MULTICAST_ERROR", 2805 "RDMA_CM_EVENT_ADDR_CHANGE", 2806 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2807 }; 2808 #endif /* DEBUG */ 2809 2810 static void 2811 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2812 struct spdk_nvmf_rdma_port *port) 2813 { 2814 struct spdk_nvmf_rdma_poll_group *rgroup; 2815 struct spdk_nvmf_rdma_poller *rpoller; 2816 struct spdk_nvmf_rdma_qpair *rqpair; 2817 2818 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2819 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2820 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 2821 if (rqpair->listen_id == port->id) { 2822 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2823 } 2824 } 2825 } 2826 } 2827 } 2828 2829 static bool 2830 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2831 struct rdma_cm_event *event) 2832 { 2833 const struct spdk_nvme_transport_id *trid; 2834 struct spdk_nvmf_rdma_port *port; 2835 struct spdk_nvmf_rdma_transport *rtransport; 2836 bool event_acked = false; 2837 2838 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2839 TAILQ_FOREACH(port, &rtransport->ports, link) { 2840 if (port->id == event->id) { 2841 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2842 rdma_ack_cm_event(event); 2843 event_acked = true; 2844 trid = port->trid; 2845 break; 2846 } 2847 } 2848 2849 if (event_acked) { 2850 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2851 2852 nvmf_rdma_stop_listen(transport, trid); 2853 nvmf_rdma_listen(transport, trid, NULL); 2854 } 2855 2856 return event_acked; 2857 } 2858 2859 static void 2860 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2861 struct rdma_cm_event *event) 2862 { 2863 struct spdk_nvmf_rdma_port *port; 2864 struct spdk_nvmf_rdma_transport *rtransport; 2865 2866 port = event->id->context; 2867 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2868 2869 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2870 2871 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2872 2873 rdma_ack_cm_event(event); 2874 2875 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2876 ; 2877 } 2878 } 2879 2880 static void 2881 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2882 { 2883 struct spdk_nvmf_rdma_transport *rtransport; 2884 struct rdma_cm_event *event; 2885 int rc; 2886 bool event_acked; 2887 2888 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2889 2890 if (rtransport->event_channel == NULL) { 2891 return; 2892 } 2893 2894 while (1) { 2895 event_acked = false; 2896 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2897 if (rc) { 2898 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2899 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2900 } 2901 break; 2902 } 2903 2904 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2905 2906 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2907 2908 switch (event->event) { 2909 case RDMA_CM_EVENT_ADDR_RESOLVED: 2910 case RDMA_CM_EVENT_ADDR_ERROR: 2911 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2912 case RDMA_CM_EVENT_ROUTE_ERROR: 2913 /* No action required. The target never attempts to resolve routes. */ 2914 break; 2915 case RDMA_CM_EVENT_CONNECT_REQUEST: 2916 rc = nvmf_rdma_connect(transport, event); 2917 if (rc < 0) { 2918 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2919 break; 2920 } 2921 break; 2922 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2923 /* The target never initiates a new connection. So this will not occur. */ 2924 break; 2925 case RDMA_CM_EVENT_CONNECT_ERROR: 2926 /* Can this happen? The docs say it can, but not sure what causes it. */ 2927 break; 2928 case RDMA_CM_EVENT_UNREACHABLE: 2929 case RDMA_CM_EVENT_REJECTED: 2930 /* These only occur on the client side. */ 2931 break; 2932 case RDMA_CM_EVENT_ESTABLISHED: 2933 /* TODO: Should we be waiting for this event anywhere? */ 2934 break; 2935 case RDMA_CM_EVENT_DISCONNECTED: 2936 rc = nvmf_rdma_disconnect(event); 2937 if (rc < 0) { 2938 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2939 break; 2940 } 2941 break; 2942 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2943 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 2944 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 2945 * Once these events are sent to SPDK, we should release all IB resources and 2946 * don't make attempts to call any ibv_query/modify/create functions. We can only call 2947 * ibv_destory* functions to release user space memory allocated by IB. All kernel 2948 * resources are already cleaned. */ 2949 if (event->id->qp) { 2950 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 2951 * corresponding qpair. Otherwise the event refers to a listening device */ 2952 rc = nvmf_rdma_disconnect(event); 2953 if (rc < 0) { 2954 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2955 break; 2956 } 2957 } else { 2958 nvmf_rdma_handle_cm_event_port_removal(transport, event); 2959 event_acked = true; 2960 } 2961 break; 2962 case RDMA_CM_EVENT_MULTICAST_JOIN: 2963 case RDMA_CM_EVENT_MULTICAST_ERROR: 2964 /* Multicast is not used */ 2965 break; 2966 case RDMA_CM_EVENT_ADDR_CHANGE: 2967 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 2968 break; 2969 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2970 /* For now, do nothing. The target never re-uses queue pairs. */ 2971 break; 2972 default: 2973 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2974 break; 2975 } 2976 if (!event_acked) { 2977 rdma_ack_cm_event(event); 2978 } 2979 } 2980 } 2981 2982 static void 2983 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 2984 { 2985 rqpair->last_wqe_reached = true; 2986 nvmf_rdma_destroy_drained_qpair(rqpair); 2987 } 2988 2989 static void 2990 nvmf_rdma_qpair_process_ibv_event(void *ctx) 2991 { 2992 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 2993 2994 if (event_ctx->rqpair) { 2995 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 2996 if (event_ctx->cb_fn) { 2997 event_ctx->cb_fn(event_ctx->rqpair); 2998 } 2999 } 3000 free(event_ctx); 3001 } 3002 3003 static int 3004 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3005 spdk_nvmf_rdma_qpair_ibv_event fn) 3006 { 3007 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3008 struct spdk_thread *thr = NULL; 3009 int rc; 3010 3011 if (rqpair->qpair.group) { 3012 thr = rqpair->qpair.group->thread; 3013 } else if (rqpair->destruct_channel) { 3014 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3015 } 3016 3017 if (!thr) { 3018 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3019 return -EINVAL; 3020 } 3021 3022 ctx = calloc(1, sizeof(*ctx)); 3023 if (!ctx) { 3024 return -ENOMEM; 3025 } 3026 3027 ctx->rqpair = rqpair; 3028 ctx->cb_fn = fn; 3029 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3030 3031 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3032 if (rc) { 3033 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3034 free(ctx); 3035 } 3036 3037 return rc; 3038 } 3039 3040 static int 3041 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3042 { 3043 int rc; 3044 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3045 struct ibv_async_event event; 3046 3047 rc = ibv_get_async_event(device->context, &event); 3048 3049 if (rc) { 3050 /* In non-blocking mode -1 means there are no events available */ 3051 return rc; 3052 } 3053 3054 switch (event.event_type) { 3055 case IBV_EVENT_QP_FATAL: 3056 rqpair = event.element.qp->qp_context; 3057 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3058 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3059 (uintptr_t)rqpair->cm_id, event.event_type); 3060 nvmf_rdma_update_ibv_state(rqpair); 3061 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3062 break; 3063 case IBV_EVENT_QP_LAST_WQE_REACHED: 3064 /* This event only occurs for shared receive queues. */ 3065 rqpair = event.element.qp->qp_context; 3066 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3067 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3068 if (rc) { 3069 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3070 rqpair->last_wqe_reached = true; 3071 } 3072 break; 3073 case IBV_EVENT_SQ_DRAINED: 3074 /* This event occurs frequently in both error and non-error states. 3075 * Check if the qpair is in an error state before sending a message. */ 3076 rqpair = event.element.qp->qp_context; 3077 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3078 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3079 (uintptr_t)rqpair->cm_id, event.event_type); 3080 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3081 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3082 } 3083 break; 3084 case IBV_EVENT_QP_REQ_ERR: 3085 case IBV_EVENT_QP_ACCESS_ERR: 3086 case IBV_EVENT_COMM_EST: 3087 case IBV_EVENT_PATH_MIG: 3088 case IBV_EVENT_PATH_MIG_ERR: 3089 SPDK_NOTICELOG("Async event: %s\n", 3090 ibv_event_type_str(event.event_type)); 3091 rqpair = event.element.qp->qp_context; 3092 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3093 (uintptr_t)rqpair->cm_id, event.event_type); 3094 nvmf_rdma_update_ibv_state(rqpair); 3095 break; 3096 case IBV_EVENT_CQ_ERR: 3097 case IBV_EVENT_DEVICE_FATAL: 3098 case IBV_EVENT_PORT_ACTIVE: 3099 case IBV_EVENT_PORT_ERR: 3100 case IBV_EVENT_LID_CHANGE: 3101 case IBV_EVENT_PKEY_CHANGE: 3102 case IBV_EVENT_SM_CHANGE: 3103 case IBV_EVENT_SRQ_ERR: 3104 case IBV_EVENT_SRQ_LIMIT_REACHED: 3105 case IBV_EVENT_CLIENT_REREGISTER: 3106 case IBV_EVENT_GID_CHANGE: 3107 default: 3108 SPDK_NOTICELOG("Async event: %s\n", 3109 ibv_event_type_str(event.event_type)); 3110 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3111 break; 3112 } 3113 ibv_ack_async_event(&event); 3114 3115 return 0; 3116 } 3117 3118 static void 3119 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3120 { 3121 int rc = 0; 3122 uint32_t i = 0; 3123 3124 for (i = 0; i < max_events; i++) { 3125 rc = nvmf_process_ib_event(device); 3126 if (rc) { 3127 break; 3128 } 3129 } 3130 3131 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3132 } 3133 3134 static uint32_t 3135 nvmf_rdma_accept(struct spdk_nvmf_transport *transport) 3136 { 3137 int nfds, i = 0; 3138 struct spdk_nvmf_rdma_transport *rtransport; 3139 struct spdk_nvmf_rdma_device *device, *tmp; 3140 uint32_t count; 3141 3142 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3143 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3144 3145 if (nfds <= 0) { 3146 return 0; 3147 } 3148 3149 /* The first poll descriptor is RDMA CM event */ 3150 if (rtransport->poll_fds[i++].revents & POLLIN) { 3151 nvmf_process_cm_event(transport); 3152 nfds--; 3153 } 3154 3155 if (nfds == 0) { 3156 return count; 3157 } 3158 3159 /* Second and subsequent poll descriptors are IB async events */ 3160 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3161 if (rtransport->poll_fds[i++].revents & POLLIN) { 3162 nvmf_process_ib_events(device, 32); 3163 nfds--; 3164 } 3165 } 3166 /* check all flagged fd's have been served */ 3167 assert(nfds == 0); 3168 3169 return count; 3170 } 3171 3172 static void 3173 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3174 struct spdk_nvmf_ctrlr_data *cdata) 3175 { 3176 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3177 3178 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3179 since in-capsule data only works with NVME drives that support SGL memory layout */ 3180 if (transport->opts.dif_insert_or_strip) { 3181 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3182 } 3183 } 3184 3185 static void 3186 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3187 struct spdk_nvme_transport_id *trid, 3188 struct spdk_nvmf_discovery_log_page_entry *entry) 3189 { 3190 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3191 entry->adrfam = trid->adrfam; 3192 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3193 3194 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3195 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3196 3197 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3198 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3199 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3200 } 3201 3202 static void 3203 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3204 3205 static struct spdk_nvmf_transport_poll_group * 3206 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3207 { 3208 struct spdk_nvmf_rdma_transport *rtransport; 3209 struct spdk_nvmf_rdma_poll_group *rgroup; 3210 struct spdk_nvmf_rdma_poller *poller; 3211 struct spdk_nvmf_rdma_device *device; 3212 struct ibv_srq_init_attr srq_init_attr; 3213 struct spdk_nvmf_rdma_resource_opts opts; 3214 int num_cqe; 3215 3216 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3217 3218 rgroup = calloc(1, sizeof(*rgroup)); 3219 if (!rgroup) { 3220 return NULL; 3221 } 3222 3223 TAILQ_INIT(&rgroup->pollers); 3224 3225 pthread_mutex_lock(&rtransport->lock); 3226 TAILQ_FOREACH(device, &rtransport->devices, link) { 3227 poller = calloc(1, sizeof(*poller)); 3228 if (!poller) { 3229 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3230 nvmf_rdma_poll_group_destroy(&rgroup->group); 3231 pthread_mutex_unlock(&rtransport->lock); 3232 return NULL; 3233 } 3234 3235 poller->device = device; 3236 poller->group = rgroup; 3237 3238 TAILQ_INIT(&poller->qpairs); 3239 STAILQ_INIT(&poller->qpairs_pending_send); 3240 STAILQ_INIT(&poller->qpairs_pending_recv); 3241 3242 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3243 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3244 poller->max_srq_depth = rtransport->rdma_opts.max_srq_depth; 3245 3246 device->num_srq++; 3247 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 3248 srq_init_attr.attr.max_wr = poller->max_srq_depth; 3249 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3250 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 3251 if (!poller->srq) { 3252 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3253 nvmf_rdma_poll_group_destroy(&rgroup->group); 3254 pthread_mutex_unlock(&rtransport->lock); 3255 return NULL; 3256 } 3257 3258 opts.qp = poller->srq; 3259 opts.pd = device->pd; 3260 opts.qpair = NULL; 3261 opts.shared = true; 3262 opts.max_queue_depth = poller->max_srq_depth; 3263 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3264 3265 poller->resources = nvmf_rdma_resources_create(&opts); 3266 if (!poller->resources) { 3267 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3268 nvmf_rdma_poll_group_destroy(&rgroup->group); 3269 pthread_mutex_unlock(&rtransport->lock); 3270 return NULL; 3271 } 3272 } 3273 3274 /* 3275 * When using an srq, we can limit the completion queue at startup. 3276 * The following formula represents the calculation: 3277 * num_cqe = num_recv + num_data_wr + num_send_wr. 3278 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3279 */ 3280 if (poller->srq) { 3281 num_cqe = poller->max_srq_depth * 3; 3282 } else { 3283 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3284 } 3285 3286 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3287 if (!poller->cq) { 3288 SPDK_ERRLOG("Unable to create completion queue\n"); 3289 nvmf_rdma_poll_group_destroy(&rgroup->group); 3290 pthread_mutex_unlock(&rtransport->lock); 3291 return NULL; 3292 } 3293 poller->num_cqe = num_cqe; 3294 } 3295 3296 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3297 if (rtransport->conn_sched.next_admin_pg == NULL) { 3298 rtransport->conn_sched.next_admin_pg = rgroup; 3299 rtransport->conn_sched.next_io_pg = rgroup; 3300 } 3301 3302 pthread_mutex_unlock(&rtransport->lock); 3303 return &rgroup->group; 3304 } 3305 3306 static struct spdk_nvmf_transport_poll_group * 3307 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3308 { 3309 struct spdk_nvmf_rdma_transport *rtransport; 3310 struct spdk_nvmf_rdma_poll_group **pg; 3311 struct spdk_nvmf_transport_poll_group *result; 3312 3313 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3314 3315 pthread_mutex_lock(&rtransport->lock); 3316 3317 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3318 pthread_mutex_unlock(&rtransport->lock); 3319 return NULL; 3320 } 3321 3322 if (qpair->qid == 0) { 3323 pg = &rtransport->conn_sched.next_admin_pg; 3324 } else { 3325 pg = &rtransport->conn_sched.next_io_pg; 3326 } 3327 3328 assert(*pg != NULL); 3329 3330 result = &(*pg)->group; 3331 3332 *pg = TAILQ_NEXT(*pg, link); 3333 if (*pg == NULL) { 3334 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3335 } 3336 3337 pthread_mutex_unlock(&rtransport->lock); 3338 3339 return result; 3340 } 3341 3342 static void 3343 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3344 { 3345 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3346 struct spdk_nvmf_rdma_poller *poller, *tmp; 3347 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3348 struct spdk_nvmf_rdma_transport *rtransport; 3349 3350 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3351 if (!rgroup) { 3352 return; 3353 } 3354 3355 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3356 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3357 3358 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3359 nvmf_rdma_qpair_destroy(qpair); 3360 } 3361 3362 if (poller->srq) { 3363 if (poller->resources) { 3364 nvmf_rdma_resources_destroy(poller->resources); 3365 } 3366 ibv_destroy_srq(poller->srq); 3367 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 3368 } 3369 3370 if (poller->cq) { 3371 ibv_destroy_cq(poller->cq); 3372 } 3373 3374 free(poller); 3375 } 3376 3377 if (rgroup->group.transport == NULL) { 3378 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3379 * calls this function directly in a failure path. */ 3380 free(rgroup); 3381 return; 3382 } 3383 3384 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3385 3386 pthread_mutex_lock(&rtransport->lock); 3387 next_rgroup = TAILQ_NEXT(rgroup, link); 3388 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3389 if (next_rgroup == NULL) { 3390 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3391 } 3392 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3393 rtransport->conn_sched.next_admin_pg = next_rgroup; 3394 } 3395 if (rtransport->conn_sched.next_io_pg == rgroup) { 3396 rtransport->conn_sched.next_io_pg = next_rgroup; 3397 } 3398 pthread_mutex_unlock(&rtransport->lock); 3399 3400 free(rgroup); 3401 } 3402 3403 static void 3404 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3405 { 3406 if (rqpair->cm_id != NULL) { 3407 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3408 } 3409 } 3410 3411 static int 3412 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3413 struct spdk_nvmf_qpair *qpair) 3414 { 3415 struct spdk_nvmf_rdma_poll_group *rgroup; 3416 struct spdk_nvmf_rdma_qpair *rqpair; 3417 struct spdk_nvmf_rdma_device *device; 3418 struct spdk_nvmf_rdma_poller *poller; 3419 int rc; 3420 3421 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3422 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3423 3424 device = rqpair->device; 3425 3426 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3427 if (poller->device == device) { 3428 break; 3429 } 3430 } 3431 3432 if (!poller) { 3433 SPDK_ERRLOG("No poller found for device.\n"); 3434 return -1; 3435 } 3436 3437 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3438 rqpair->poller = poller; 3439 rqpair->srq = rqpair->poller->srq; 3440 3441 rc = nvmf_rdma_qpair_initialize(qpair); 3442 if (rc < 0) { 3443 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3444 return -1; 3445 } 3446 3447 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3448 if (rc) { 3449 /* Try to reject, but we probably can't */ 3450 nvmf_rdma_qpair_reject_connection(rqpair); 3451 return -1; 3452 } 3453 3454 nvmf_rdma_update_ibv_state(rqpair); 3455 3456 return 0; 3457 } 3458 3459 static int 3460 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3461 struct spdk_nvmf_qpair *qpair) 3462 { 3463 struct spdk_nvmf_rdma_qpair *rqpair; 3464 3465 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3466 assert(group->transport->tgt != NULL); 3467 3468 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3469 3470 if (!rqpair->destruct_channel) { 3471 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3472 return 0; 3473 } 3474 3475 /* Sanity check that we get io_channel on the correct thread */ 3476 if (qpair->group) { 3477 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3478 } 3479 3480 return 0; 3481 } 3482 3483 static int 3484 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3485 { 3486 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3487 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3488 struct spdk_nvmf_rdma_transport, transport); 3489 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3490 struct spdk_nvmf_rdma_qpair, qpair); 3491 3492 /* 3493 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3494 * needs to be returned to the shared receive queue or the poll group will eventually be 3495 * starved of RECV structures. 3496 */ 3497 if (rqpair->srq && rdma_req->recv) { 3498 int rc; 3499 struct ibv_recv_wr *bad_recv_wr; 3500 3501 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 3502 if (rc) { 3503 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3504 } 3505 } 3506 3507 _nvmf_rdma_request_free(rdma_req, rtransport); 3508 return 0; 3509 } 3510 3511 static int 3512 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3513 { 3514 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3515 struct spdk_nvmf_rdma_transport, transport); 3516 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3517 struct spdk_nvmf_rdma_request, req); 3518 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3519 struct spdk_nvmf_rdma_qpair, qpair); 3520 3521 if (rqpair->ibv_state != IBV_QPS_ERR) { 3522 /* The connection is alive, so process the request as normal */ 3523 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3524 } else { 3525 /* The connection is dead. Move the request directly to the completed state. */ 3526 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3527 } 3528 3529 nvmf_rdma_request_process(rtransport, rdma_req); 3530 3531 return 0; 3532 } 3533 3534 static void 3535 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 3536 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3537 { 3538 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3539 3540 rqpair->to_close = true; 3541 3542 /* This happens only when the qpair is disconnected before 3543 * it is added to the poll group. Since there is no poll group, 3544 * the RDMA qp has not been initialized yet and the RDMA CM 3545 * event has not yet been acknowledged, so we need to reject it. 3546 */ 3547 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3548 nvmf_rdma_qpair_reject_connection(rqpair); 3549 nvmf_rdma_qpair_destroy(rqpair); 3550 return; 3551 } 3552 3553 if (rqpair->rdma_qp) { 3554 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3555 } 3556 3557 nvmf_rdma_destroy_drained_qpair(rqpair); 3558 3559 if (cb_fn) { 3560 cb_fn(cb_arg); 3561 } 3562 } 3563 3564 static struct spdk_nvmf_rdma_qpair * 3565 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3566 { 3567 struct spdk_nvmf_rdma_qpair *rqpair; 3568 /* @todo: improve QP search */ 3569 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3570 if (wc->qp_num == rqpair->rdma_qp->qp->qp_num) { 3571 return rqpair; 3572 } 3573 } 3574 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3575 return NULL; 3576 } 3577 3578 #ifdef DEBUG 3579 static int 3580 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3581 { 3582 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3583 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3584 } 3585 #endif 3586 3587 static void 3588 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3589 int rc) 3590 { 3591 struct spdk_nvmf_rdma_recv *rdma_recv; 3592 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3593 3594 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3595 while (bad_recv_wr != NULL) { 3596 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3597 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3598 3599 rdma_recv->qpair->current_recv_depth++; 3600 bad_recv_wr = bad_recv_wr->next; 3601 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3602 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3603 } 3604 } 3605 3606 static void 3607 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3608 { 3609 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3610 while (bad_recv_wr != NULL) { 3611 bad_recv_wr = bad_recv_wr->next; 3612 rqpair->current_recv_depth++; 3613 } 3614 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3615 } 3616 3617 static void 3618 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3619 struct spdk_nvmf_rdma_poller *rpoller) 3620 { 3621 struct spdk_nvmf_rdma_qpair *rqpair; 3622 struct ibv_recv_wr *bad_recv_wr; 3623 int rc; 3624 3625 if (rpoller->srq) { 3626 if (rpoller->resources->recvs_to_post.first != NULL) { 3627 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3628 if (rc) { 3629 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3630 } 3631 rpoller->resources->recvs_to_post.first = NULL; 3632 rpoller->resources->recvs_to_post.last = NULL; 3633 } 3634 } else { 3635 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3636 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3637 assert(rqpair->resources->recvs_to_post.first != NULL); 3638 rc = ibv_post_recv(rqpair->rdma_qp->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3639 if (rc) { 3640 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3641 } 3642 rqpair->resources->recvs_to_post.first = NULL; 3643 rqpair->resources->recvs_to_post.last = NULL; 3644 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3645 } 3646 } 3647 } 3648 3649 static void 3650 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3651 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3652 { 3653 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3654 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3655 3656 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3657 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3658 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3659 assert(rqpair->current_send_depth > 0); 3660 rqpair->current_send_depth--; 3661 switch (bad_rdma_wr->type) { 3662 case RDMA_WR_TYPE_DATA: 3663 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3664 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3665 assert(rqpair->current_read_depth > 0); 3666 rqpair->current_read_depth--; 3667 } 3668 break; 3669 case RDMA_WR_TYPE_SEND: 3670 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3671 break; 3672 default: 3673 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3674 prev_rdma_req = cur_rdma_req; 3675 continue; 3676 } 3677 3678 if (prev_rdma_req == cur_rdma_req) { 3679 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3680 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3681 continue; 3682 } 3683 3684 switch (cur_rdma_req->state) { 3685 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3686 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3687 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3688 break; 3689 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3690 case RDMA_REQUEST_STATE_COMPLETING: 3691 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3692 break; 3693 default: 3694 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3695 cur_rdma_req->state, rqpair); 3696 continue; 3697 } 3698 3699 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3700 prev_rdma_req = cur_rdma_req; 3701 } 3702 3703 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3704 /* Disconnect the connection. */ 3705 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3706 } 3707 3708 } 3709 3710 static void 3711 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3712 struct spdk_nvmf_rdma_poller *rpoller) 3713 { 3714 struct spdk_nvmf_rdma_qpair *rqpair; 3715 struct ibv_send_wr *bad_wr = NULL; 3716 int rc; 3717 3718 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3719 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3720 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3721 3722 /* bad wr always points to the first wr that failed. */ 3723 if (rc) { 3724 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3725 } 3726 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3727 } 3728 } 3729 3730 static const char * 3731 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 3732 { 3733 switch (wr_type) { 3734 case RDMA_WR_TYPE_RECV: 3735 return "RECV"; 3736 case RDMA_WR_TYPE_SEND: 3737 return "SEND"; 3738 case RDMA_WR_TYPE_DATA: 3739 return "DATA"; 3740 default: 3741 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 3742 SPDK_UNREACHABLE(); 3743 } 3744 } 3745 3746 static inline void 3747 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 3748 { 3749 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 3750 3751 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 3752 /* If qpair is in ERR state, we will receive completions for all posted and not completed 3753 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 3754 SPDK_DEBUGLOG(rdma, 3755 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3756 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3757 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3758 } else { 3759 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3760 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3761 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3762 } 3763 } 3764 3765 static int 3766 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3767 struct spdk_nvmf_rdma_poller *rpoller) 3768 { 3769 struct ibv_wc wc[32]; 3770 struct spdk_nvmf_rdma_wr *rdma_wr; 3771 struct spdk_nvmf_rdma_request *rdma_req; 3772 struct spdk_nvmf_rdma_recv *rdma_recv; 3773 struct spdk_nvmf_rdma_qpair *rqpair; 3774 int reaped, i; 3775 int count = 0; 3776 bool error = false; 3777 uint64_t poll_tsc = spdk_get_ticks(); 3778 3779 /* Poll for completing operations. */ 3780 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3781 if (reaped < 0) { 3782 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3783 errno, spdk_strerror(errno)); 3784 return -1; 3785 } 3786 3787 rpoller->stat.polls++; 3788 rpoller->stat.completions += reaped; 3789 3790 for (i = 0; i < reaped; i++) { 3791 3792 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3793 3794 switch (rdma_wr->type) { 3795 case RDMA_WR_TYPE_SEND: 3796 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3797 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3798 3799 if (!wc[i].status) { 3800 count++; 3801 assert(wc[i].opcode == IBV_WC_SEND); 3802 assert(nvmf_rdma_req_is_completing(rdma_req)); 3803 } 3804 3805 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3806 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3807 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3808 rdma_req->num_outstanding_data_wr = 0; 3809 3810 nvmf_rdma_request_process(rtransport, rdma_req); 3811 break; 3812 case RDMA_WR_TYPE_RECV: 3813 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3814 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3815 if (rpoller->srq != NULL) { 3816 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3817 /* It is possible that there are still some completions for destroyed QP 3818 * associated with SRQ. We just ignore these late completions and re-post 3819 * receive WRs back to SRQ. 3820 */ 3821 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3822 struct ibv_recv_wr *bad_wr; 3823 int rc; 3824 3825 rdma_recv->wr.next = NULL; 3826 rc = ibv_post_srq_recv(rpoller->srq, 3827 &rdma_recv->wr, 3828 &bad_wr); 3829 if (rc) { 3830 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3831 } 3832 continue; 3833 } 3834 } 3835 rqpair = rdma_recv->qpair; 3836 3837 assert(rqpair != NULL); 3838 if (!wc[i].status) { 3839 assert(wc[i].opcode == IBV_WC_RECV); 3840 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3841 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3842 break; 3843 } 3844 } 3845 3846 rdma_recv->wr.next = NULL; 3847 rqpair->current_recv_depth++; 3848 rdma_recv->receive_tsc = poll_tsc; 3849 rpoller->stat.requests++; 3850 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3851 break; 3852 case RDMA_WR_TYPE_DATA: 3853 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3854 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3855 3856 assert(rdma_req->num_outstanding_data_wr > 0); 3857 3858 rqpair->current_send_depth--; 3859 rdma_req->num_outstanding_data_wr--; 3860 if (!wc[i].status) { 3861 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3862 rqpair->current_read_depth--; 3863 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3864 if (rdma_req->num_outstanding_data_wr == 0) { 3865 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3866 nvmf_rdma_request_process(rtransport, rdma_req); 3867 } 3868 } else { 3869 /* If the data transfer fails still force the queue into the error state, 3870 * if we were performing an RDMA_READ, we need to force the request into a 3871 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3872 * case, we should wait for the SEND to complete. */ 3873 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3874 rqpair->current_read_depth--; 3875 if (rdma_req->num_outstanding_data_wr == 0) { 3876 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3877 } 3878 } 3879 } 3880 break; 3881 default: 3882 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3883 continue; 3884 } 3885 3886 /* Handle error conditions */ 3887 if (wc[i].status) { 3888 nvmf_rdma_update_ibv_state(rqpair); 3889 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 3890 3891 error = true; 3892 3893 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3894 /* Disconnect the connection. */ 3895 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3896 } else { 3897 nvmf_rdma_destroy_drained_qpair(rqpair); 3898 } 3899 continue; 3900 } 3901 3902 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3903 3904 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3905 nvmf_rdma_destroy_drained_qpair(rqpair); 3906 } 3907 } 3908 3909 if (error == true) { 3910 return -1; 3911 } 3912 3913 /* submit outstanding work requests. */ 3914 _poller_submit_recvs(rtransport, rpoller); 3915 _poller_submit_sends(rtransport, rpoller); 3916 3917 return count; 3918 } 3919 3920 static int 3921 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3922 { 3923 struct spdk_nvmf_rdma_transport *rtransport; 3924 struct spdk_nvmf_rdma_poll_group *rgroup; 3925 struct spdk_nvmf_rdma_poller *rpoller; 3926 int count, rc; 3927 3928 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3929 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3930 3931 count = 0; 3932 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3933 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 3934 if (rc < 0) { 3935 return rc; 3936 } 3937 count += rc; 3938 } 3939 3940 return count; 3941 } 3942 3943 static int 3944 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3945 struct spdk_nvme_transport_id *trid, 3946 bool peer) 3947 { 3948 struct sockaddr *saddr; 3949 uint16_t port; 3950 3951 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 3952 3953 if (peer) { 3954 saddr = rdma_get_peer_addr(id); 3955 } else { 3956 saddr = rdma_get_local_addr(id); 3957 } 3958 switch (saddr->sa_family) { 3959 case AF_INET: { 3960 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3961 3962 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3963 inet_ntop(AF_INET, &saddr_in->sin_addr, 3964 trid->traddr, sizeof(trid->traddr)); 3965 if (peer) { 3966 port = ntohs(rdma_get_dst_port(id)); 3967 } else { 3968 port = ntohs(rdma_get_src_port(id)); 3969 } 3970 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3971 break; 3972 } 3973 case AF_INET6: { 3974 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3975 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3976 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3977 trid->traddr, sizeof(trid->traddr)); 3978 if (peer) { 3979 port = ntohs(rdma_get_dst_port(id)); 3980 } else { 3981 port = ntohs(rdma_get_src_port(id)); 3982 } 3983 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3984 break; 3985 } 3986 default: 3987 return -1; 3988 3989 } 3990 3991 return 0; 3992 } 3993 3994 static int 3995 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3996 struct spdk_nvme_transport_id *trid) 3997 { 3998 struct spdk_nvmf_rdma_qpair *rqpair; 3999 4000 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4001 4002 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4003 } 4004 4005 static int 4006 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4007 struct spdk_nvme_transport_id *trid) 4008 { 4009 struct spdk_nvmf_rdma_qpair *rqpair; 4010 4011 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4012 4013 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4014 } 4015 4016 static int 4017 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4018 struct spdk_nvme_transport_id *trid) 4019 { 4020 struct spdk_nvmf_rdma_qpair *rqpair; 4021 4022 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4023 4024 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4025 } 4026 4027 void 4028 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4029 { 4030 g_nvmf_hooks = *hooks; 4031 } 4032 4033 static void 4034 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4035 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4036 { 4037 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4038 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4039 4040 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4041 4042 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4043 } 4044 4045 static int 4046 _nvmf_rdma_qpair_abort_request(void *ctx) 4047 { 4048 struct spdk_nvmf_request *req = ctx; 4049 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4050 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4051 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4052 struct spdk_nvmf_rdma_qpair, qpair); 4053 int rc; 4054 4055 spdk_poller_unregister(&req->poller); 4056 4057 switch (rdma_req_to_abort->state) { 4058 case RDMA_REQUEST_STATE_EXECUTING: 4059 rc = nvmf_ctrlr_abort_request(req); 4060 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4061 return SPDK_POLLER_BUSY; 4062 } 4063 break; 4064 4065 case RDMA_REQUEST_STATE_NEED_BUFFER: 4066 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4067 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4068 4069 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4070 break; 4071 4072 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4073 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4074 spdk_nvmf_rdma_request, state_link); 4075 4076 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4077 break; 4078 4079 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4080 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4081 spdk_nvmf_rdma_request, state_link); 4082 4083 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4084 break; 4085 4086 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4087 if (spdk_get_ticks() < req->timeout_tsc) { 4088 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4089 return SPDK_POLLER_BUSY; 4090 } 4091 break; 4092 4093 default: 4094 break; 4095 } 4096 4097 spdk_nvmf_request_complete(req); 4098 return SPDK_POLLER_BUSY; 4099 } 4100 4101 static void 4102 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4103 struct spdk_nvmf_request *req) 4104 { 4105 struct spdk_nvmf_rdma_qpair *rqpair; 4106 struct spdk_nvmf_rdma_transport *rtransport; 4107 struct spdk_nvmf_transport *transport; 4108 uint16_t cid; 4109 uint32_t i; 4110 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL; 4111 4112 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4113 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4114 transport = &rtransport->transport; 4115 4116 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4117 4118 for (i = 0; i < rqpair->max_queue_depth; i++) { 4119 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE && 4120 rqpair->resources->reqs[i].req.cmd->nvme_cmd.cid == cid) { 4121 rdma_req_to_abort = &rqpair->resources->reqs[i]; 4122 break; 4123 } 4124 } 4125 4126 if (rdma_req_to_abort == NULL) { 4127 spdk_nvmf_request_complete(req); 4128 return; 4129 } 4130 4131 req->req_to_abort = &rdma_req_to_abort->req; 4132 req->timeout_tsc = spdk_get_ticks() + 4133 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4134 req->poller = NULL; 4135 4136 _nvmf_rdma_qpair_abort_request(req); 4137 } 4138 4139 static int 4140 nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4141 struct spdk_nvmf_transport_poll_group_stat **stat) 4142 { 4143 struct spdk_io_channel *ch; 4144 struct spdk_nvmf_poll_group *group; 4145 struct spdk_nvmf_transport_poll_group *tgroup; 4146 struct spdk_nvmf_rdma_poll_group *rgroup; 4147 struct spdk_nvmf_rdma_poller *rpoller; 4148 struct spdk_nvmf_rdma_device_stat *device_stat; 4149 uint64_t num_devices = 0; 4150 4151 if (tgt == NULL || stat == NULL) { 4152 return -EINVAL; 4153 } 4154 4155 ch = spdk_get_io_channel(tgt); 4156 group = spdk_io_channel_get_ctx(ch);; 4157 spdk_put_io_channel(ch); 4158 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4159 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4160 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4161 if (!*stat) { 4162 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4163 return -ENOMEM; 4164 } 4165 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4166 4167 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4168 /* Count devices to allocate enough memory */ 4169 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4170 ++num_devices; 4171 } 4172 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4173 if (!(*stat)->rdma.devices) { 4174 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4175 free(*stat); 4176 return -ENOMEM; 4177 } 4178 4179 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4180 (*stat)->rdma.num_devices = num_devices; 4181 num_devices = 0; 4182 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4183 device_stat = &(*stat)->rdma.devices[num_devices++]; 4184 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4185 device_stat->polls = rpoller->stat.polls; 4186 device_stat->completions = rpoller->stat.completions; 4187 device_stat->requests = rpoller->stat.requests; 4188 device_stat->request_latency = rpoller->stat.request_latency; 4189 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4190 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4191 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4192 } 4193 return 0; 4194 } 4195 } 4196 return -ENOENT; 4197 } 4198 4199 static void 4200 nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4201 { 4202 if (stat) { 4203 free(stat->rdma.devices); 4204 } 4205 free(stat); 4206 } 4207 4208 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4209 .name = "RDMA", 4210 .type = SPDK_NVME_TRANSPORT_RDMA, 4211 .opts_init = nvmf_rdma_opts_init, 4212 .create = nvmf_rdma_create, 4213 .dump_opts = nvmf_rdma_dump_opts, 4214 .destroy = nvmf_rdma_destroy, 4215 4216 .listen = nvmf_rdma_listen, 4217 .stop_listen = nvmf_rdma_stop_listen, 4218 .accept = nvmf_rdma_accept, 4219 .cdata_init = nvmf_rdma_cdata_init, 4220 4221 .listener_discover = nvmf_rdma_discover, 4222 4223 .poll_group_create = nvmf_rdma_poll_group_create, 4224 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4225 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4226 .poll_group_add = nvmf_rdma_poll_group_add, 4227 .poll_group_remove = nvmf_rdma_poll_group_remove, 4228 .poll_group_poll = nvmf_rdma_poll_group_poll, 4229 4230 .req_free = nvmf_rdma_request_free, 4231 .req_complete = nvmf_rdma_request_complete, 4232 4233 .qpair_fini = nvmf_rdma_close_qpair, 4234 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4235 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4236 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4237 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4238 4239 .poll_group_get_stat = nvmf_rdma_poll_group_get_stat, 4240 .poll_group_free_stat = nvmf_rdma_poll_group_free_stat, 4241 }; 4242 4243 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4244 SPDK_LOG_REGISTER_COMPONENT(rdma) 4245