1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2019-2021 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 (device->context->device->transport_type == IBV_TRANSPORT_IWARP) { 948 SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n" 949 "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe); 950 return -1; 951 } 952 if (required_num_wr > device->attr.max_cqe) { 953 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 954 required_num_wr, device->attr.max_cqe); 955 return -1; 956 } 957 958 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 959 rc = ibv_resize_cq(rpoller->cq, num_cqe); 960 if (rc) { 961 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 962 return -1; 963 } 964 965 rpoller->num_cqe = num_cqe; 966 } 967 968 rpoller->required_num_wr = required_num_wr; 969 return 0; 970 } 971 972 static int 973 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 974 { 975 struct spdk_nvmf_rdma_qpair *rqpair; 976 struct spdk_nvmf_rdma_transport *rtransport; 977 struct spdk_nvmf_transport *transport; 978 struct spdk_nvmf_rdma_resource_opts opts; 979 struct spdk_nvmf_rdma_device *device; 980 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 981 982 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 983 device = rqpair->device; 984 985 qp_init_attr.qp_context = rqpair; 986 qp_init_attr.pd = device->pd; 987 qp_init_attr.send_cq = rqpair->poller->cq; 988 qp_init_attr.recv_cq = rqpair->poller->cq; 989 990 if (rqpair->srq) { 991 qp_init_attr.srq = rqpair->srq; 992 } else { 993 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 994 } 995 996 /* SEND, READ, and WRITE operations */ 997 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 998 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 999 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1000 1001 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1002 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1003 goto error; 1004 } 1005 1006 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 1007 if (!rqpair->rdma_qp) { 1008 goto error; 1009 } 1010 1011 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1012 qp_init_attr.cap.max_send_wr); 1013 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1014 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1015 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 1016 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1017 1018 if (rqpair->poller->srq == NULL) { 1019 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1020 transport = &rtransport->transport; 1021 1022 opts.qp = rqpair->rdma_qp->qp; 1023 opts.pd = rqpair->cm_id->pd; 1024 opts.qpair = rqpair; 1025 opts.shared = false; 1026 opts.max_queue_depth = rqpair->max_queue_depth; 1027 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1028 1029 rqpair->resources = nvmf_rdma_resources_create(&opts); 1030 1031 if (!rqpair->resources) { 1032 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1033 rdma_destroy_qp(rqpair->cm_id); 1034 goto error; 1035 } 1036 } else { 1037 rqpair->resources = rqpair->poller->resources; 1038 } 1039 1040 rqpair->current_recv_depth = 0; 1041 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1042 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1043 1044 return 0; 1045 1046 error: 1047 rdma_destroy_id(rqpair->cm_id); 1048 rqpair->cm_id = NULL; 1049 return -1; 1050 } 1051 1052 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1053 /* This function accepts either a single wr or the first wr in a linked list. */ 1054 static void 1055 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1056 { 1057 struct ibv_recv_wr *last; 1058 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1059 struct spdk_nvmf_rdma_transport, transport); 1060 1061 last = first; 1062 while (last->next != NULL) { 1063 last = last->next; 1064 } 1065 1066 if (rqpair->resources->recvs_to_post.first == NULL) { 1067 rqpair->resources->recvs_to_post.first = first; 1068 rqpair->resources->recvs_to_post.last = last; 1069 if (rqpair->srq == NULL) { 1070 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1071 } 1072 } else { 1073 rqpair->resources->recvs_to_post.last->next = first; 1074 rqpair->resources->recvs_to_post.last = last; 1075 } 1076 1077 if (rtransport->rdma_opts.no_wr_batching) { 1078 _poller_submit_recvs(rtransport, rqpair->poller); 1079 } 1080 } 1081 1082 static int 1083 request_transfer_in(struct spdk_nvmf_request *req) 1084 { 1085 struct spdk_nvmf_rdma_request *rdma_req; 1086 struct spdk_nvmf_qpair *qpair; 1087 struct spdk_nvmf_rdma_qpair *rqpair; 1088 struct spdk_nvmf_rdma_transport *rtransport; 1089 1090 qpair = req->qpair; 1091 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1092 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1093 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1094 struct spdk_nvmf_rdma_transport, transport); 1095 1096 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1097 assert(rdma_req != NULL); 1098 1099 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1100 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1101 } 1102 if (rtransport->rdma_opts.no_wr_batching) { 1103 _poller_submit_sends(rtransport, rqpair->poller); 1104 } 1105 1106 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1107 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1108 return 0; 1109 } 1110 1111 static int 1112 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1113 { 1114 int num_outstanding_data_wr = 0; 1115 struct spdk_nvmf_rdma_request *rdma_req; 1116 struct spdk_nvmf_qpair *qpair; 1117 struct spdk_nvmf_rdma_qpair *rqpair; 1118 struct spdk_nvme_cpl *rsp; 1119 struct ibv_send_wr *first = NULL; 1120 struct spdk_nvmf_rdma_transport *rtransport; 1121 1122 *data_posted = 0; 1123 qpair = req->qpair; 1124 rsp = &req->rsp->nvme_cpl; 1125 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1126 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1127 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1128 struct spdk_nvmf_rdma_transport, transport); 1129 1130 /* Advance our sq_head pointer */ 1131 if (qpair->sq_head == qpair->sq_head_max) { 1132 qpair->sq_head = 0; 1133 } else { 1134 qpair->sq_head++; 1135 } 1136 rsp->sqhd = qpair->sq_head; 1137 1138 /* queue the capsule for the recv buffer */ 1139 assert(rdma_req->recv != NULL); 1140 1141 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1142 1143 rdma_req->recv = NULL; 1144 assert(rqpair->current_recv_depth > 0); 1145 rqpair->current_recv_depth--; 1146 1147 /* Build the response which consists of optional 1148 * RDMA WRITEs to transfer data, plus an RDMA SEND 1149 * containing the response. 1150 */ 1151 first = &rdma_req->rsp.wr; 1152 1153 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1154 /* On failure, data was not read from the controller. So clear the 1155 * number of outstanding data WRs to zero. 1156 */ 1157 rdma_req->num_outstanding_data_wr = 0; 1158 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1159 first = &rdma_req->data.wr; 1160 *data_posted = 1; 1161 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1162 } 1163 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1164 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1165 } 1166 if (rtransport->rdma_opts.no_wr_batching) { 1167 _poller_submit_sends(rtransport, rqpair->poller); 1168 } 1169 1170 /* +1 for the rsp wr */ 1171 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1172 1173 return 0; 1174 } 1175 1176 static int 1177 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1178 { 1179 struct spdk_nvmf_rdma_accept_private_data accept_data; 1180 struct rdma_conn_param ctrlr_event_data = {}; 1181 int rc; 1182 1183 accept_data.recfmt = 0; 1184 accept_data.crqsize = rqpair->max_queue_depth; 1185 1186 ctrlr_event_data.private_data = &accept_data; 1187 ctrlr_event_data.private_data_len = sizeof(accept_data); 1188 if (id->ps == RDMA_PS_TCP) { 1189 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1190 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1191 } 1192 1193 /* Configure infinite retries for the initiator side qpair. 1194 * When using a shared receive queue on the target side, 1195 * we need to pass this value to the initiator to prevent the 1196 * initiator side NIC from completing SEND requests back to the 1197 * initiator with status rnr_retry_count_exceeded. */ 1198 if (rqpair->srq != NULL) { 1199 ctrlr_event_data.rnr_retry_count = 0x7; 1200 } 1201 1202 /* When qpair is created without use of rdma cm API, an additional 1203 * information must be provided to initiator in the connection response: 1204 * whether qpair is using SRQ and its qp_num 1205 * Fields below are ignored by rdma cm if qpair has been 1206 * created using rdma cm API. */ 1207 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1208 ctrlr_event_data.qp_num = rqpair->rdma_qp->qp->qp_num; 1209 1210 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1211 if (rc) { 1212 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1213 } else { 1214 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1215 } 1216 1217 return rc; 1218 } 1219 1220 static void 1221 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1222 { 1223 struct spdk_nvmf_rdma_reject_private_data rej_data; 1224 1225 rej_data.recfmt = 0; 1226 rej_data.sts = error; 1227 1228 rdma_reject(id, &rej_data, sizeof(rej_data)); 1229 } 1230 1231 static int 1232 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1233 { 1234 struct spdk_nvmf_rdma_transport *rtransport; 1235 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1236 struct spdk_nvmf_rdma_port *port; 1237 struct rdma_conn_param *rdma_param = NULL; 1238 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1239 uint16_t max_queue_depth; 1240 uint16_t max_read_depth; 1241 1242 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1243 1244 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1245 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1246 1247 rdma_param = &event->param.conn; 1248 if (rdma_param->private_data == NULL || 1249 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1250 SPDK_ERRLOG("connect request: no private data provided\n"); 1251 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1252 return -1; 1253 } 1254 1255 private_data = rdma_param->private_data; 1256 if (private_data->recfmt != 0) { 1257 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1258 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1259 return -1; 1260 } 1261 1262 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1263 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1264 1265 port = event->listen_id->context; 1266 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1267 event->listen_id, event->listen_id->verbs, port); 1268 1269 /* Figure out the supported queue depth. This is a multi-step process 1270 * that takes into account hardware maximums, host provided values, 1271 * and our target's internal memory limits */ 1272 1273 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1274 1275 /* Start with the maximum queue depth allowed by the target */ 1276 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1277 max_read_depth = rtransport->transport.opts.max_queue_depth; 1278 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1279 rtransport->transport.opts.max_queue_depth); 1280 1281 /* Next check the local NIC's hardware limitations */ 1282 SPDK_DEBUGLOG(rdma, 1283 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1284 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1285 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1286 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1287 1288 /* Next check the remote NIC's hardware limitations */ 1289 SPDK_DEBUGLOG(rdma, 1290 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1291 rdma_param->initiator_depth, rdma_param->responder_resources); 1292 if (rdma_param->initiator_depth > 0) { 1293 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1294 } 1295 1296 /* Finally check for the host software requested values, which are 1297 * optional. */ 1298 if (rdma_param->private_data != NULL && 1299 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1300 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1301 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1302 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1303 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1304 } 1305 1306 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1307 max_queue_depth, max_read_depth); 1308 1309 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1310 if (rqpair == NULL) { 1311 SPDK_ERRLOG("Could not allocate new connection.\n"); 1312 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1313 return -1; 1314 } 1315 1316 rqpair->device = port->device; 1317 rqpair->max_queue_depth = max_queue_depth; 1318 rqpair->max_read_depth = max_read_depth; 1319 rqpair->cm_id = event->id; 1320 rqpair->listen_id = event->listen_id; 1321 rqpair->qpair.transport = transport; 1322 STAILQ_INIT(&rqpair->ibv_events); 1323 /* use qid from the private data to determine the qpair type 1324 qid will be set to the appropriate value when the controller is created */ 1325 rqpair->qpair.qid = private_data->qid; 1326 1327 event->id->context = &rqpair->qpair; 1328 1329 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1330 1331 return 0; 1332 } 1333 1334 static inline void 1335 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1336 enum spdk_nvme_data_transfer xfer) 1337 { 1338 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1339 wr->opcode = IBV_WR_RDMA_WRITE; 1340 wr->send_flags = 0; 1341 wr->next = next; 1342 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1343 wr->opcode = IBV_WR_RDMA_READ; 1344 wr->send_flags = IBV_SEND_SIGNALED; 1345 wr->next = NULL; 1346 } else { 1347 assert(0); 1348 } 1349 } 1350 1351 static int 1352 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1353 struct spdk_nvmf_rdma_request *rdma_req, 1354 uint32_t num_sgl_descriptors) 1355 { 1356 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1357 struct spdk_nvmf_rdma_request_data *current_data_wr; 1358 uint32_t i; 1359 1360 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1361 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1362 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1363 return -EINVAL; 1364 } 1365 1366 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1367 return -ENOMEM; 1368 } 1369 1370 current_data_wr = &rdma_req->data; 1371 1372 for (i = 0; i < num_sgl_descriptors; i++) { 1373 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1374 current_data_wr->wr.next = &work_requests[i]->wr; 1375 current_data_wr = work_requests[i]; 1376 current_data_wr->wr.sg_list = current_data_wr->sgl; 1377 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1378 } 1379 1380 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1381 1382 return 0; 1383 } 1384 1385 static inline void 1386 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1387 { 1388 struct ibv_send_wr *wr = &rdma_req->data.wr; 1389 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1390 1391 wr->wr.rdma.rkey = sgl->keyed.key; 1392 wr->wr.rdma.remote_addr = sgl->address; 1393 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1394 } 1395 1396 static inline void 1397 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1398 { 1399 struct ibv_send_wr *wr = &rdma_req->data.wr; 1400 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1401 uint32_t i; 1402 int j; 1403 uint64_t remote_addr_offset = 0; 1404 1405 for (i = 0; i < num_wrs; ++i) { 1406 wr->wr.rdma.rkey = sgl->keyed.key; 1407 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1408 for (j = 0; j < wr->num_sge; ++j) { 1409 remote_addr_offset += wr->sg_list[j].length; 1410 } 1411 wr = wr->next; 1412 } 1413 } 1414 1415 static bool 1416 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device, 1417 struct iovec *iov, struct ibv_send_wr **_wr, 1418 uint32_t *_remaining_data_block, uint32_t *_offset, 1419 uint32_t *_num_extra_wrs, 1420 const struct spdk_dif_ctx *dif_ctx) 1421 { 1422 struct ibv_send_wr *wr = *_wr; 1423 struct ibv_sge *sg_ele = &wr->sg_list[wr->num_sge]; 1424 struct spdk_rdma_memory_translation mem_translation; 1425 int rc; 1426 uint32_t lkey = 0; 1427 uint32_t remaining, data_block_size, md_size, sge_len; 1428 1429 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1430 if (spdk_unlikely(rc)) { 1431 return false; 1432 } 1433 1434 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1435 1436 if (spdk_likely(!dif_ctx)) { 1437 sg_ele->lkey = lkey; 1438 sg_ele->addr = (uintptr_t)(iov->iov_base); 1439 sg_ele->length = iov->iov_len; 1440 wr->num_sge++; 1441 } else { 1442 remaining = iov->iov_len - *_offset; 1443 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1444 md_size = dif_ctx->md_size; 1445 1446 while (remaining) { 1447 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1448 if (*_num_extra_wrs > 0 && wr->next) { 1449 *_wr = wr->next; 1450 wr = *_wr; 1451 wr->num_sge = 0; 1452 sg_ele = &wr->sg_list[wr->num_sge]; 1453 (*_num_extra_wrs)--; 1454 } else { 1455 break; 1456 } 1457 } 1458 sg_ele->lkey = lkey; 1459 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset); 1460 sge_len = spdk_min(remaining, *_remaining_data_block); 1461 sg_ele->length = sge_len; 1462 remaining -= sge_len; 1463 *_remaining_data_block -= sge_len; 1464 *_offset += sge_len; 1465 1466 sg_ele++; 1467 wr->num_sge++; 1468 1469 if (*_remaining_data_block == 0) { 1470 /* skip metadata */ 1471 *_offset += md_size; 1472 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1473 remaining -= spdk_min(remaining, md_size); 1474 *_remaining_data_block = data_block_size; 1475 } 1476 1477 if (remaining == 0) { 1478 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1479 the remaining metadata bits at the beginning of the next buffer */ 1480 *_offset -= iov->iov_len; 1481 } 1482 } 1483 } 1484 1485 return true; 1486 } 1487 1488 static int 1489 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1490 struct spdk_nvmf_rdma_device *device, 1491 struct spdk_nvmf_rdma_request *rdma_req, 1492 struct ibv_send_wr *wr, 1493 uint32_t length, 1494 uint32_t num_extra_wrs) 1495 { 1496 struct spdk_nvmf_request *req = &rdma_req->req; 1497 struct spdk_dif_ctx *dif_ctx = NULL; 1498 uint32_t remaining_data_block = 0; 1499 uint32_t offset = 0; 1500 1501 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 1502 dif_ctx = &rdma_req->req.dif.dif_ctx; 1503 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1504 } 1505 1506 wr->num_sge = 0; 1507 1508 while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1509 if (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr, 1510 &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) { 1511 return -EINVAL; 1512 } 1513 1514 length -= req->iov[rdma_req->iovpos].iov_len; 1515 rdma_req->iovpos++; 1516 } 1517 1518 if (length) { 1519 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1520 return -EINVAL; 1521 } 1522 1523 return 0; 1524 } 1525 1526 static inline uint32_t 1527 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1528 { 1529 /* estimate the number of SG entries and WRs needed to process the request */ 1530 uint32_t num_sge = 0; 1531 uint32_t i; 1532 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1533 1534 for (i = 0; i < num_buffers && length > 0; i++) { 1535 uint32_t buffer_len = spdk_min(length, io_unit_size); 1536 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1537 1538 if (num_sge_in_block * block_size > buffer_len) { 1539 ++num_sge_in_block; 1540 } 1541 num_sge += num_sge_in_block; 1542 length -= buffer_len; 1543 } 1544 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1545 } 1546 1547 static int 1548 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1549 struct spdk_nvmf_rdma_device *device, 1550 struct spdk_nvmf_rdma_request *rdma_req, 1551 uint32_t length) 1552 { 1553 struct spdk_nvmf_rdma_qpair *rqpair; 1554 struct spdk_nvmf_rdma_poll_group *rgroup; 1555 struct spdk_nvmf_request *req = &rdma_req->req; 1556 struct ibv_send_wr *wr = &rdma_req->data.wr; 1557 int rc; 1558 uint32_t num_wrs = 1; 1559 1560 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1561 rgroup = rqpair->poller->group; 1562 1563 /* rdma wr specifics */ 1564 nvmf_rdma_setup_request(rdma_req); 1565 1566 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1567 length); 1568 if (rc != 0) { 1569 return rc; 1570 } 1571 1572 assert(req->iovcnt <= rqpair->max_send_sge); 1573 1574 rdma_req->iovpos = 0; 1575 1576 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1577 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1578 req->dif.dif_ctx.block_size); 1579 if (num_wrs > 1) { 1580 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1581 if (rc != 0) { 1582 goto err_exit; 1583 } 1584 } 1585 } 1586 1587 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1588 if (spdk_unlikely(rc != 0)) { 1589 goto err_exit; 1590 } 1591 1592 if (spdk_unlikely(num_wrs > 1)) { 1593 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1594 } 1595 1596 /* set the number of outstanding data WRs for this request. */ 1597 rdma_req->num_outstanding_data_wr = num_wrs; 1598 1599 return rc; 1600 1601 err_exit: 1602 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1603 nvmf_rdma_request_free_data(rdma_req, rtransport); 1604 req->iovcnt = 0; 1605 return rc; 1606 } 1607 1608 static int 1609 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1610 struct spdk_nvmf_rdma_device *device, 1611 struct spdk_nvmf_rdma_request *rdma_req) 1612 { 1613 struct spdk_nvmf_rdma_qpair *rqpair; 1614 struct spdk_nvmf_rdma_poll_group *rgroup; 1615 struct ibv_send_wr *current_wr; 1616 struct spdk_nvmf_request *req = &rdma_req->req; 1617 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1618 uint32_t num_sgl_descriptors; 1619 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES]; 1620 uint32_t i; 1621 int rc; 1622 1623 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1624 rgroup = rqpair->poller->group; 1625 1626 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1627 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1628 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1629 1630 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1631 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1632 1633 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1634 return -ENOMEM; 1635 } 1636 1637 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1638 for (i = 0; i < num_sgl_descriptors; i++) { 1639 if (spdk_likely(!req->dif.dif_insert_or_strip)) { 1640 lengths[i] = desc->keyed.length; 1641 } else { 1642 req->dif.orig_length += desc->keyed.length; 1643 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1644 req->dif.elba_length += lengths[i]; 1645 } 1646 desc++; 1647 } 1648 1649 rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport, 1650 lengths, num_sgl_descriptors); 1651 if (rc != 0) { 1652 nvmf_rdma_request_free_data(rdma_req, rtransport); 1653 return rc; 1654 } 1655 1656 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1657 current_wr = &rdma_req->data.wr; 1658 assert(current_wr != NULL); 1659 1660 req->length = 0; 1661 rdma_req->iovpos = 0; 1662 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1663 for (i = 0; i < num_sgl_descriptors; i++) { 1664 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1665 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1666 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1667 rc = -EINVAL; 1668 goto err_exit; 1669 } 1670 1671 current_wr->num_sge = 0; 1672 1673 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1674 if (rc != 0) { 1675 rc = -ENOMEM; 1676 goto err_exit; 1677 } 1678 1679 req->length += desc->keyed.length; 1680 current_wr->wr.rdma.rkey = desc->keyed.key; 1681 current_wr->wr.rdma.remote_addr = desc->address; 1682 current_wr = current_wr->next; 1683 desc++; 1684 } 1685 1686 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1687 /* Go back to the last descriptor in the list. */ 1688 desc--; 1689 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1690 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1691 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1692 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1693 } 1694 } 1695 #endif 1696 1697 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1698 1699 return 0; 1700 1701 err_exit: 1702 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1703 nvmf_rdma_request_free_data(rdma_req, rtransport); 1704 return rc; 1705 } 1706 1707 static int 1708 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1709 struct spdk_nvmf_rdma_device *device, 1710 struct spdk_nvmf_rdma_request *rdma_req) 1711 { 1712 struct spdk_nvmf_request *req = &rdma_req->req; 1713 struct spdk_nvme_cpl *rsp; 1714 struct spdk_nvme_sgl_descriptor *sgl; 1715 int rc; 1716 uint32_t length; 1717 1718 rsp = &req->rsp->nvme_cpl; 1719 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1720 1721 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1722 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1723 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1724 1725 length = sgl->keyed.length; 1726 if (length > rtransport->transport.opts.max_io_size) { 1727 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1728 length, rtransport->transport.opts.max_io_size); 1729 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1730 return -1; 1731 } 1732 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1733 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1734 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1735 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1736 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1737 } 1738 } 1739 #endif 1740 1741 /* fill request length and populate iovs */ 1742 req->length = length; 1743 1744 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1745 req->dif.orig_length = length; 1746 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1747 req->dif.elba_length = length; 1748 } 1749 1750 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1751 if (spdk_unlikely(rc < 0)) { 1752 if (rc == -EINVAL) { 1753 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1754 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1755 return -1; 1756 } 1757 /* No available buffers. Queue this request up. */ 1758 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1759 return 0; 1760 } 1761 1762 /* backward compatible */ 1763 req->data = req->iov[0].iov_base; 1764 1765 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1766 req->iovcnt); 1767 1768 return 0; 1769 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1770 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1771 uint64_t offset = sgl->address; 1772 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1773 1774 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1775 offset, sgl->unkeyed.length); 1776 1777 if (offset > max_len) { 1778 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1779 offset, max_len); 1780 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1781 return -1; 1782 } 1783 max_len -= (uint32_t)offset; 1784 1785 if (sgl->unkeyed.length > max_len) { 1786 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1787 sgl->unkeyed.length, max_len); 1788 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1789 return -1; 1790 } 1791 1792 rdma_req->num_outstanding_data_wr = 0; 1793 req->data = rdma_req->recv->buf + offset; 1794 req->data_from_pool = false; 1795 req->length = sgl->unkeyed.length; 1796 1797 req->iov[0].iov_base = req->data; 1798 req->iov[0].iov_len = req->length; 1799 req->iovcnt = 1; 1800 1801 return 0; 1802 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1803 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1804 1805 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1806 if (rc == -ENOMEM) { 1807 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1808 return 0; 1809 } else if (rc == -EINVAL) { 1810 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1811 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1812 return -1; 1813 } 1814 1815 /* backward compatible */ 1816 req->data = req->iov[0].iov_base; 1817 1818 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1819 req->iovcnt); 1820 1821 return 0; 1822 } 1823 1824 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1825 sgl->generic.type, sgl->generic.subtype); 1826 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1827 return -1; 1828 } 1829 1830 static void 1831 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1832 struct spdk_nvmf_rdma_transport *rtransport) 1833 { 1834 struct spdk_nvmf_rdma_qpair *rqpair; 1835 struct spdk_nvmf_rdma_poll_group *rgroup; 1836 1837 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1838 if (rdma_req->req.data_from_pool) { 1839 rgroup = rqpair->poller->group; 1840 1841 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1842 } 1843 nvmf_rdma_request_free_data(rdma_req, rtransport); 1844 rdma_req->req.length = 0; 1845 rdma_req->req.iovcnt = 0; 1846 rdma_req->req.data = NULL; 1847 rdma_req->rsp.wr.next = NULL; 1848 rdma_req->data.wr.next = NULL; 1849 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1850 rqpair->qd--; 1851 1852 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1853 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1854 } 1855 1856 bool 1857 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1858 struct spdk_nvmf_rdma_request *rdma_req) 1859 { 1860 struct spdk_nvmf_rdma_qpair *rqpair; 1861 struct spdk_nvmf_rdma_device *device; 1862 struct spdk_nvmf_rdma_poll_group *rgroup; 1863 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1864 int rc; 1865 struct spdk_nvmf_rdma_recv *rdma_recv; 1866 enum spdk_nvmf_rdma_request_state prev_state; 1867 bool progress = false; 1868 int data_posted; 1869 uint32_t num_blocks; 1870 1871 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1872 device = rqpair->device; 1873 rgroup = rqpair->poller->group; 1874 1875 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1876 1877 /* If the queue pair is in an error state, force the request to the completed state 1878 * to release resources. */ 1879 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1880 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1881 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1882 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1883 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1884 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1885 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1886 } 1887 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1888 } 1889 1890 /* The loop here is to allow for several back-to-back state changes. */ 1891 do { 1892 prev_state = rdma_req->state; 1893 1894 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 1895 1896 switch (rdma_req->state) { 1897 case RDMA_REQUEST_STATE_FREE: 1898 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1899 * to escape this state. */ 1900 break; 1901 case RDMA_REQUEST_STATE_NEW: 1902 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1903 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1904 rdma_recv = rdma_req->recv; 1905 1906 /* The first element of the SGL is the NVMe command */ 1907 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1908 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1909 1910 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1911 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1912 break; 1913 } 1914 1915 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 1916 rdma_req->req.dif.dif_insert_or_strip = true; 1917 } 1918 1919 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1920 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1921 rdma_req->rsp.wr.imm_data = 0; 1922 #endif 1923 1924 /* The next state transition depends on the data transfer needs of this request. */ 1925 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 1926 1927 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 1928 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 1929 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 1930 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1931 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 1932 break; 1933 } 1934 1935 /* If no data to transfer, ready to execute. */ 1936 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1937 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1938 break; 1939 } 1940 1941 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1942 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 1943 break; 1944 case RDMA_REQUEST_STATE_NEED_BUFFER: 1945 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1946 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1947 1948 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1949 1950 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 1951 /* This request needs to wait in line to obtain a buffer */ 1952 break; 1953 } 1954 1955 /* Try to get a data buffer */ 1956 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1957 if (rc < 0) { 1958 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 1959 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1960 break; 1961 } 1962 1963 if (!rdma_req->req.data) { 1964 /* No buffers available. */ 1965 rgroup->stat.pending_data_buffer++; 1966 break; 1967 } 1968 1969 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 1970 1971 /* If data is transferring from host to controller and the data didn't 1972 * arrive using in capsule data, we need to do a transfer from the host. 1973 */ 1974 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 1975 rdma_req->req.data_from_pool) { 1976 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1977 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1978 break; 1979 } 1980 1981 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1982 break; 1983 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1984 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1985 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1986 1987 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1988 /* This request needs to wait in line to perform RDMA */ 1989 break; 1990 } 1991 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1992 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1993 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1994 rqpair->poller->stat.pending_rdma_read++; 1995 break; 1996 } 1997 1998 /* We have already verified that this request is the head of the queue. */ 1999 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2000 2001 rc = request_transfer_in(&rdma_req->req); 2002 if (!rc) { 2003 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2004 } else { 2005 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2006 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2007 } 2008 break; 2009 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2010 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2011 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2012 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2013 * to escape this state. */ 2014 break; 2015 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2016 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2017 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2018 2019 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2020 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2021 /* generate DIF for write operation */ 2022 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2023 assert(num_blocks > 0); 2024 2025 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2026 num_blocks, &rdma_req->req.dif.dif_ctx); 2027 if (rc != 0) { 2028 SPDK_ERRLOG("DIF generation failed\n"); 2029 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2030 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2031 break; 2032 } 2033 } 2034 2035 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2036 /* set extended length before IO operation */ 2037 rdma_req->req.length = rdma_req->req.dif.elba_length; 2038 } 2039 2040 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2041 spdk_nvmf_request_exec(&rdma_req->req); 2042 break; 2043 case RDMA_REQUEST_STATE_EXECUTING: 2044 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2045 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2046 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2047 * to escape this state. */ 2048 break; 2049 case RDMA_REQUEST_STATE_EXECUTED: 2050 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2051 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2052 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2053 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2054 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2055 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2056 } else { 2057 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2058 } 2059 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2060 /* restore the original length */ 2061 rdma_req->req.length = rdma_req->req.dif.orig_length; 2062 2063 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2064 struct spdk_dif_error error_blk; 2065 2066 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2067 2068 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2069 &rdma_req->req.dif.dif_ctx, &error_blk); 2070 if (rc) { 2071 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2072 2073 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2074 error_blk.err_offset); 2075 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2076 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2077 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2078 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2079 } 2080 } 2081 } 2082 break; 2083 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2084 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2085 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2086 2087 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2088 /* This request needs to wait in line to perform RDMA */ 2089 break; 2090 } 2091 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2092 rqpair->max_send_depth) { 2093 /* We can only have so many WRs outstanding. we have to wait until some finish. 2094 * +1 since each request has an additional wr in the resp. */ 2095 rqpair->poller->stat.pending_rdma_write++; 2096 break; 2097 } 2098 2099 /* We have already verified that this request is the head of the queue. */ 2100 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2101 2102 /* The data transfer will be kicked off from 2103 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2104 */ 2105 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2106 break; 2107 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2108 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2109 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2110 rc = request_transfer_out(&rdma_req->req, &data_posted); 2111 assert(rc == 0); /* No good way to handle this currently */ 2112 if (rc) { 2113 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2114 } else { 2115 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2116 RDMA_REQUEST_STATE_COMPLETING; 2117 } 2118 break; 2119 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2120 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2121 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2122 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2123 * to escape this state. */ 2124 break; 2125 case RDMA_REQUEST_STATE_COMPLETING: 2126 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2127 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2128 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2129 * to escape this state. */ 2130 break; 2131 case RDMA_REQUEST_STATE_COMPLETED: 2132 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2133 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2134 2135 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2136 _nvmf_rdma_request_free(rdma_req, rtransport); 2137 break; 2138 case RDMA_REQUEST_NUM_STATES: 2139 default: 2140 assert(0); 2141 break; 2142 } 2143 2144 if (rdma_req->state != prev_state) { 2145 progress = true; 2146 } 2147 } while (rdma_req->state != prev_state); 2148 2149 return progress; 2150 } 2151 2152 /* Public API callbacks begin here */ 2153 2154 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2155 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2156 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2157 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2158 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2159 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2160 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2161 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2162 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2163 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2164 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2165 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2166 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2167 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2168 2169 static void 2170 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2171 { 2172 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2173 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2174 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2175 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2176 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2177 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2178 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2179 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2180 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2181 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2182 opts->transport_specific = NULL; 2183 } 2184 2185 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2186 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2187 2188 static inline bool 2189 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2190 { 2191 return device->attr.vendor_id == NVMF_RXE_VENDOR_ID_OLD || 2192 device->attr.vendor_id == NVMF_RXE_VENDOR_ID_NEW; 2193 } 2194 2195 static struct spdk_nvmf_transport * 2196 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2197 { 2198 int rc; 2199 struct spdk_nvmf_rdma_transport *rtransport; 2200 struct spdk_nvmf_rdma_device *device, *tmp; 2201 struct ibv_context **contexts; 2202 uint32_t i; 2203 int flag; 2204 uint32_t sge_count; 2205 uint32_t min_shared_buffers; 2206 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2207 pthread_mutexattr_t attr; 2208 2209 rtransport = calloc(1, sizeof(*rtransport)); 2210 if (!rtransport) { 2211 return NULL; 2212 } 2213 2214 if (pthread_mutexattr_init(&attr)) { 2215 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2216 free(rtransport); 2217 return NULL; 2218 } 2219 2220 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2221 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2222 pthread_mutexattr_destroy(&attr); 2223 free(rtransport); 2224 return NULL; 2225 } 2226 2227 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2228 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2229 pthread_mutexattr_destroy(&attr); 2230 free(rtransport); 2231 return NULL; 2232 } 2233 2234 pthread_mutexattr_destroy(&attr); 2235 2236 TAILQ_INIT(&rtransport->devices); 2237 TAILQ_INIT(&rtransport->ports); 2238 TAILQ_INIT(&rtransport->poll_groups); 2239 2240 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2241 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2242 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2243 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2244 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2245 if (opts->transport_specific != NULL && 2246 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2247 SPDK_COUNTOF(rdma_transport_opts_decoder), 2248 &rtransport->rdma_opts)) { 2249 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2250 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2251 return NULL; 2252 } 2253 2254 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2255 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2256 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2257 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2258 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d," 2259 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2260 opts->max_queue_depth, 2261 opts->max_io_size, 2262 opts->max_qpairs_per_ctrlr - 1, 2263 opts->io_unit_size, 2264 opts->in_capsule_data_size, 2265 opts->max_aq_depth, 2266 opts->num_shared_buffers, 2267 rtransport->rdma_opts.max_srq_depth, 2268 rtransport->rdma_opts.no_srq, 2269 rtransport->rdma_opts.acceptor_backlog, 2270 rtransport->rdma_opts.no_wr_batching, 2271 opts->abort_timeout_sec); 2272 2273 /* I/O unit size cannot be larger than max I/O size */ 2274 if (opts->io_unit_size > opts->max_io_size) { 2275 opts->io_unit_size = opts->max_io_size; 2276 } 2277 2278 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2279 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2280 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2281 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2282 } 2283 2284 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2285 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2286 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2287 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2288 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2289 return NULL; 2290 } 2291 2292 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2293 if (min_shared_buffers > opts->num_shared_buffers) { 2294 SPDK_ERRLOG("There are not enough buffers to satisfy" 2295 "per-poll group caches for each thread. (%" PRIu32 ")" 2296 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2297 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2298 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2299 return NULL; 2300 } 2301 2302 sge_count = opts->max_io_size / opts->io_unit_size; 2303 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2304 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2305 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2306 return NULL; 2307 } 2308 2309 rtransport->event_channel = rdma_create_event_channel(); 2310 if (rtransport->event_channel == NULL) { 2311 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2312 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2313 return NULL; 2314 } 2315 2316 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2317 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2318 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2319 rtransport->event_channel->fd, spdk_strerror(errno)); 2320 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2321 return NULL; 2322 } 2323 2324 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2325 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2326 sizeof(struct spdk_nvmf_rdma_request_data), 2327 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2328 SPDK_ENV_SOCKET_ID_ANY); 2329 if (!rtransport->data_wr_pool) { 2330 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2331 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2332 return NULL; 2333 } 2334 2335 contexts = rdma_get_devices(NULL); 2336 if (contexts == NULL) { 2337 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2338 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2339 return NULL; 2340 } 2341 2342 i = 0; 2343 rc = 0; 2344 while (contexts[i] != NULL) { 2345 device = calloc(1, sizeof(*device)); 2346 if (!device) { 2347 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2348 rc = -ENOMEM; 2349 break; 2350 } 2351 device->context = contexts[i]; 2352 rc = ibv_query_device(device->context, &device->attr); 2353 if (rc < 0) { 2354 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2355 free(device); 2356 break; 2357 2358 } 2359 2360 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2361 2362 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2363 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2364 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2365 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2366 } 2367 2368 /** 2369 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2370 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2371 * but incorrectly reports that it does. There are changes making their way 2372 * through the kernel now that will enable this feature. When they are merged, 2373 * we can conditionally enable this feature. 2374 * 2375 * TODO: enable this for versions of the kernel rxe driver that support it. 2376 */ 2377 if (nvmf_rdma_is_rxe_device(device)) { 2378 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2379 } 2380 #endif 2381 2382 /* set up device context async ev fd as NON_BLOCKING */ 2383 flag = fcntl(device->context->async_fd, F_GETFL); 2384 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2385 if (rc < 0) { 2386 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2387 free(device); 2388 break; 2389 } 2390 2391 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2392 i++; 2393 2394 if (g_nvmf_hooks.get_ibv_pd) { 2395 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2396 } else { 2397 device->pd = ibv_alloc_pd(device->context); 2398 } 2399 2400 if (!device->pd) { 2401 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2402 rc = -ENOMEM; 2403 break; 2404 } 2405 2406 assert(device->map == NULL); 2407 2408 device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks); 2409 if (!device->map) { 2410 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2411 rc = -ENOMEM; 2412 break; 2413 } 2414 2415 assert(device->map != NULL); 2416 assert(device->pd != NULL); 2417 } 2418 rdma_free_devices(contexts); 2419 2420 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2421 /* divide and round up. */ 2422 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2423 2424 /* round up to the nearest 4k. */ 2425 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2426 2427 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2428 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2429 opts->io_unit_size); 2430 } 2431 2432 if (rc < 0) { 2433 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2434 return NULL; 2435 } 2436 2437 /* Set up poll descriptor array to monitor events from RDMA and IB 2438 * in a single poll syscall 2439 */ 2440 rtransport->npoll_fds = i + 1; 2441 i = 0; 2442 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2443 if (rtransport->poll_fds == NULL) { 2444 SPDK_ERRLOG("poll_fds allocation failed\n"); 2445 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2446 return NULL; 2447 } 2448 2449 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2450 rtransport->poll_fds[i++].events = POLLIN; 2451 2452 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2453 rtransport->poll_fds[i].fd = device->context->async_fd; 2454 rtransport->poll_fds[i++].events = POLLIN; 2455 } 2456 2457 return &rtransport->transport; 2458 } 2459 2460 static void 2461 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2462 { 2463 struct spdk_nvmf_rdma_transport *rtransport; 2464 assert(w != NULL); 2465 2466 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2467 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2468 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2469 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2470 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2471 } 2472 2473 static int 2474 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2475 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2476 { 2477 struct spdk_nvmf_rdma_transport *rtransport; 2478 struct spdk_nvmf_rdma_port *port, *port_tmp; 2479 struct spdk_nvmf_rdma_device *device, *device_tmp; 2480 2481 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2482 2483 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2484 TAILQ_REMOVE(&rtransport->ports, port, link); 2485 rdma_destroy_id(port->id); 2486 free(port); 2487 } 2488 2489 if (rtransport->poll_fds != NULL) { 2490 free(rtransport->poll_fds); 2491 } 2492 2493 if (rtransport->event_channel != NULL) { 2494 rdma_destroy_event_channel(rtransport->event_channel); 2495 } 2496 2497 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2498 TAILQ_REMOVE(&rtransport->devices, device, link); 2499 spdk_rdma_free_mem_map(&device->map); 2500 if (device->pd) { 2501 if (!g_nvmf_hooks.get_ibv_pd) { 2502 ibv_dealloc_pd(device->pd); 2503 } 2504 } 2505 free(device); 2506 } 2507 2508 if (rtransport->data_wr_pool != NULL) { 2509 if (spdk_mempool_count(rtransport->data_wr_pool) != 2510 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2511 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2512 spdk_mempool_count(rtransport->data_wr_pool), 2513 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2514 } 2515 } 2516 2517 spdk_mempool_free(rtransport->data_wr_pool); 2518 2519 pthread_mutex_destroy(&rtransport->lock); 2520 free(rtransport); 2521 2522 if (cb_fn) { 2523 cb_fn(cb_arg); 2524 } 2525 return 0; 2526 } 2527 2528 static int 2529 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2530 struct spdk_nvme_transport_id *trid, 2531 bool peer); 2532 2533 static int 2534 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2535 struct spdk_nvmf_listen_opts *listen_opts) 2536 { 2537 struct spdk_nvmf_rdma_transport *rtransport; 2538 struct spdk_nvmf_rdma_device *device; 2539 struct spdk_nvmf_rdma_port *port; 2540 struct addrinfo *res; 2541 struct addrinfo hints; 2542 int family; 2543 int rc; 2544 2545 if (!strlen(trid->trsvcid)) { 2546 SPDK_ERRLOG("Service id is required\n"); 2547 return -EINVAL; 2548 } 2549 2550 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2551 assert(rtransport->event_channel != NULL); 2552 2553 pthread_mutex_lock(&rtransport->lock); 2554 port = calloc(1, sizeof(*port)); 2555 if (!port) { 2556 SPDK_ERRLOG("Port allocation failed\n"); 2557 pthread_mutex_unlock(&rtransport->lock); 2558 return -ENOMEM; 2559 } 2560 2561 port->trid = trid; 2562 2563 switch (trid->adrfam) { 2564 case SPDK_NVMF_ADRFAM_IPV4: 2565 family = AF_INET; 2566 break; 2567 case SPDK_NVMF_ADRFAM_IPV6: 2568 family = AF_INET6; 2569 break; 2570 default: 2571 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2572 free(port); 2573 pthread_mutex_unlock(&rtransport->lock); 2574 return -EINVAL; 2575 } 2576 2577 memset(&hints, 0, sizeof(hints)); 2578 hints.ai_family = family; 2579 hints.ai_flags = AI_NUMERICSERV; 2580 hints.ai_socktype = SOCK_STREAM; 2581 hints.ai_protocol = 0; 2582 2583 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2584 if (rc) { 2585 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2586 free(port); 2587 pthread_mutex_unlock(&rtransport->lock); 2588 return -EINVAL; 2589 } 2590 2591 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2592 if (rc < 0) { 2593 SPDK_ERRLOG("rdma_create_id() failed\n"); 2594 freeaddrinfo(res); 2595 free(port); 2596 pthread_mutex_unlock(&rtransport->lock); 2597 return rc; 2598 } 2599 2600 rc = rdma_bind_addr(port->id, res->ai_addr); 2601 freeaddrinfo(res); 2602 2603 if (rc < 0) { 2604 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2605 rdma_destroy_id(port->id); 2606 free(port); 2607 pthread_mutex_unlock(&rtransport->lock); 2608 return rc; 2609 } 2610 2611 if (!port->id->verbs) { 2612 SPDK_ERRLOG("ibv_context is null\n"); 2613 rdma_destroy_id(port->id); 2614 free(port); 2615 pthread_mutex_unlock(&rtransport->lock); 2616 return -1; 2617 } 2618 2619 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2620 if (rc < 0) { 2621 SPDK_ERRLOG("rdma_listen() failed\n"); 2622 rdma_destroy_id(port->id); 2623 free(port); 2624 pthread_mutex_unlock(&rtransport->lock); 2625 return rc; 2626 } 2627 2628 TAILQ_FOREACH(device, &rtransport->devices, link) { 2629 if (device->context == port->id->verbs) { 2630 port->device = device; 2631 break; 2632 } 2633 } 2634 if (!port->device) { 2635 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2636 port->id->verbs); 2637 rdma_destroy_id(port->id); 2638 free(port); 2639 pthread_mutex_unlock(&rtransport->lock); 2640 return -EINVAL; 2641 } 2642 2643 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2644 trid->traddr, trid->trsvcid); 2645 2646 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2647 pthread_mutex_unlock(&rtransport->lock); 2648 return 0; 2649 } 2650 2651 static void 2652 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2653 const struct spdk_nvme_transport_id *trid) 2654 { 2655 struct spdk_nvmf_rdma_transport *rtransport; 2656 struct spdk_nvmf_rdma_port *port, *tmp; 2657 2658 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2659 2660 pthread_mutex_lock(&rtransport->lock); 2661 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2662 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2663 TAILQ_REMOVE(&rtransport->ports, port, link); 2664 rdma_destroy_id(port->id); 2665 free(port); 2666 break; 2667 } 2668 } 2669 2670 pthread_mutex_unlock(&rtransport->lock); 2671 } 2672 2673 static void 2674 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2675 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2676 { 2677 struct spdk_nvmf_request *req, *tmp; 2678 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2679 struct spdk_nvmf_rdma_resources *resources; 2680 2681 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2682 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2683 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2684 break; 2685 } 2686 } 2687 2688 /* Then RDMA writes since reads have stronger restrictions than writes */ 2689 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2690 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2691 break; 2692 } 2693 } 2694 2695 /* Then we handle request waiting on memory buffers. */ 2696 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2697 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2698 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2699 break; 2700 } 2701 } 2702 2703 resources = rqpair->resources; 2704 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2705 rdma_req = STAILQ_FIRST(&resources->free_queue); 2706 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2707 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2708 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2709 2710 if (rqpair->srq != NULL) { 2711 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2712 rdma_req->recv->qpair->qd++; 2713 } else { 2714 rqpair->qd++; 2715 } 2716 2717 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2718 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2719 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2720 break; 2721 } 2722 } 2723 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2724 rqpair->poller->stat.pending_free_request++; 2725 } 2726 } 2727 2728 static inline bool 2729 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 2730 { 2731 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 2732 return nvmf_rdma_is_rxe_device(device) || 2733 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 2734 } 2735 2736 static void 2737 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2738 { 2739 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2740 struct spdk_nvmf_rdma_transport, transport); 2741 2742 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2743 2744 /* nvmr_rdma_close_qpair is not called */ 2745 if (!rqpair->to_close) { 2746 return; 2747 } 2748 2749 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2750 if (rqpair->current_send_depth != 0) { 2751 return; 2752 } 2753 2754 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2755 return; 2756 } 2757 2758 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 2759 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 2760 return; 2761 } 2762 2763 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2764 2765 nvmf_rdma_qpair_destroy(rqpair); 2766 } 2767 2768 static int 2769 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2770 { 2771 struct spdk_nvmf_qpair *qpair; 2772 struct spdk_nvmf_rdma_qpair *rqpair; 2773 2774 if (evt->id == NULL) { 2775 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2776 return -1; 2777 } 2778 2779 qpair = evt->id->context; 2780 if (qpair == NULL) { 2781 SPDK_ERRLOG("disconnect request: no active connection\n"); 2782 return -1; 2783 } 2784 2785 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2786 2787 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2788 2789 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2790 2791 return 0; 2792 } 2793 2794 #ifdef DEBUG 2795 static const char *CM_EVENT_STR[] = { 2796 "RDMA_CM_EVENT_ADDR_RESOLVED", 2797 "RDMA_CM_EVENT_ADDR_ERROR", 2798 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2799 "RDMA_CM_EVENT_ROUTE_ERROR", 2800 "RDMA_CM_EVENT_CONNECT_REQUEST", 2801 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2802 "RDMA_CM_EVENT_CONNECT_ERROR", 2803 "RDMA_CM_EVENT_UNREACHABLE", 2804 "RDMA_CM_EVENT_REJECTED", 2805 "RDMA_CM_EVENT_ESTABLISHED", 2806 "RDMA_CM_EVENT_DISCONNECTED", 2807 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2808 "RDMA_CM_EVENT_MULTICAST_JOIN", 2809 "RDMA_CM_EVENT_MULTICAST_ERROR", 2810 "RDMA_CM_EVENT_ADDR_CHANGE", 2811 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2812 }; 2813 #endif /* DEBUG */ 2814 2815 static void 2816 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2817 struct spdk_nvmf_rdma_port *port) 2818 { 2819 struct spdk_nvmf_rdma_poll_group *rgroup; 2820 struct spdk_nvmf_rdma_poller *rpoller; 2821 struct spdk_nvmf_rdma_qpair *rqpair; 2822 2823 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2824 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2825 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 2826 if (rqpair->listen_id == port->id) { 2827 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2828 } 2829 } 2830 } 2831 } 2832 } 2833 2834 static bool 2835 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2836 struct rdma_cm_event *event) 2837 { 2838 const struct spdk_nvme_transport_id *trid; 2839 struct spdk_nvmf_rdma_port *port; 2840 struct spdk_nvmf_rdma_transport *rtransport; 2841 bool event_acked = false; 2842 2843 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2844 TAILQ_FOREACH(port, &rtransport->ports, link) { 2845 if (port->id == event->id) { 2846 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2847 rdma_ack_cm_event(event); 2848 event_acked = true; 2849 trid = port->trid; 2850 break; 2851 } 2852 } 2853 2854 if (event_acked) { 2855 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2856 2857 nvmf_rdma_stop_listen(transport, trid); 2858 nvmf_rdma_listen(transport, trid, NULL); 2859 } 2860 2861 return event_acked; 2862 } 2863 2864 static void 2865 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2866 struct rdma_cm_event *event) 2867 { 2868 struct spdk_nvmf_rdma_port *port; 2869 struct spdk_nvmf_rdma_transport *rtransport; 2870 2871 port = event->id->context; 2872 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2873 2874 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2875 2876 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2877 2878 rdma_ack_cm_event(event); 2879 2880 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2881 ; 2882 } 2883 } 2884 2885 static void 2886 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2887 { 2888 struct spdk_nvmf_rdma_transport *rtransport; 2889 struct rdma_cm_event *event; 2890 int rc; 2891 bool event_acked; 2892 2893 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2894 2895 if (rtransport->event_channel == NULL) { 2896 return; 2897 } 2898 2899 while (1) { 2900 event_acked = false; 2901 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2902 if (rc) { 2903 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2904 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2905 } 2906 break; 2907 } 2908 2909 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2910 2911 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2912 2913 switch (event->event) { 2914 case RDMA_CM_EVENT_ADDR_RESOLVED: 2915 case RDMA_CM_EVENT_ADDR_ERROR: 2916 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2917 case RDMA_CM_EVENT_ROUTE_ERROR: 2918 /* No action required. The target never attempts to resolve routes. */ 2919 break; 2920 case RDMA_CM_EVENT_CONNECT_REQUEST: 2921 rc = nvmf_rdma_connect(transport, event); 2922 if (rc < 0) { 2923 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2924 break; 2925 } 2926 break; 2927 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2928 /* The target never initiates a new connection. So this will not occur. */ 2929 break; 2930 case RDMA_CM_EVENT_CONNECT_ERROR: 2931 /* Can this happen? The docs say it can, but not sure what causes it. */ 2932 break; 2933 case RDMA_CM_EVENT_UNREACHABLE: 2934 case RDMA_CM_EVENT_REJECTED: 2935 /* These only occur on the client side. */ 2936 break; 2937 case RDMA_CM_EVENT_ESTABLISHED: 2938 /* TODO: Should we be waiting for this event anywhere? */ 2939 break; 2940 case RDMA_CM_EVENT_DISCONNECTED: 2941 rc = nvmf_rdma_disconnect(event); 2942 if (rc < 0) { 2943 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2944 break; 2945 } 2946 break; 2947 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2948 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 2949 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 2950 * Once these events are sent to SPDK, we should release all IB resources and 2951 * don't make attempts to call any ibv_query/modify/create functions. We can only call 2952 * ibv_destory* functions to release user space memory allocated by IB. All kernel 2953 * resources are already cleaned. */ 2954 if (event->id->qp) { 2955 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 2956 * corresponding qpair. Otherwise the event refers to a listening device */ 2957 rc = nvmf_rdma_disconnect(event); 2958 if (rc < 0) { 2959 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2960 break; 2961 } 2962 } else { 2963 nvmf_rdma_handle_cm_event_port_removal(transport, event); 2964 event_acked = true; 2965 } 2966 break; 2967 case RDMA_CM_EVENT_MULTICAST_JOIN: 2968 case RDMA_CM_EVENT_MULTICAST_ERROR: 2969 /* Multicast is not used */ 2970 break; 2971 case RDMA_CM_EVENT_ADDR_CHANGE: 2972 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 2973 break; 2974 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2975 /* For now, do nothing. The target never re-uses queue pairs. */ 2976 break; 2977 default: 2978 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2979 break; 2980 } 2981 if (!event_acked) { 2982 rdma_ack_cm_event(event); 2983 } 2984 } 2985 } 2986 2987 static void 2988 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 2989 { 2990 rqpair->last_wqe_reached = true; 2991 nvmf_rdma_destroy_drained_qpair(rqpair); 2992 } 2993 2994 static void 2995 nvmf_rdma_qpair_process_ibv_event(void *ctx) 2996 { 2997 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 2998 2999 if (event_ctx->rqpair) { 3000 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3001 if (event_ctx->cb_fn) { 3002 event_ctx->cb_fn(event_ctx->rqpair); 3003 } 3004 } 3005 free(event_ctx); 3006 } 3007 3008 static int 3009 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3010 spdk_nvmf_rdma_qpair_ibv_event fn) 3011 { 3012 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3013 struct spdk_thread *thr = NULL; 3014 int rc; 3015 3016 if (rqpair->qpair.group) { 3017 thr = rqpair->qpair.group->thread; 3018 } else if (rqpair->destruct_channel) { 3019 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3020 } 3021 3022 if (!thr) { 3023 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3024 return -EINVAL; 3025 } 3026 3027 ctx = calloc(1, sizeof(*ctx)); 3028 if (!ctx) { 3029 return -ENOMEM; 3030 } 3031 3032 ctx->rqpair = rqpair; 3033 ctx->cb_fn = fn; 3034 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3035 3036 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3037 if (rc) { 3038 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3039 free(ctx); 3040 } 3041 3042 return rc; 3043 } 3044 3045 static int 3046 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3047 { 3048 int rc; 3049 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3050 struct ibv_async_event event; 3051 3052 rc = ibv_get_async_event(device->context, &event); 3053 3054 if (rc) { 3055 /* In non-blocking mode -1 means there are no events available */ 3056 return rc; 3057 } 3058 3059 switch (event.event_type) { 3060 case IBV_EVENT_QP_FATAL: 3061 rqpair = event.element.qp->qp_context; 3062 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3063 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3064 (uintptr_t)rqpair->cm_id, event.event_type); 3065 nvmf_rdma_update_ibv_state(rqpair); 3066 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3067 break; 3068 case IBV_EVENT_QP_LAST_WQE_REACHED: 3069 /* This event only occurs for shared receive queues. */ 3070 rqpair = event.element.qp->qp_context; 3071 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3072 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3073 if (rc) { 3074 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3075 rqpair->last_wqe_reached = true; 3076 } 3077 break; 3078 case IBV_EVENT_SQ_DRAINED: 3079 /* This event occurs frequently in both error and non-error states. 3080 * Check if the qpair is in an error state before sending a message. */ 3081 rqpair = event.element.qp->qp_context; 3082 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3083 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3084 (uintptr_t)rqpair->cm_id, event.event_type); 3085 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3086 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3087 } 3088 break; 3089 case IBV_EVENT_QP_REQ_ERR: 3090 case IBV_EVENT_QP_ACCESS_ERR: 3091 case IBV_EVENT_COMM_EST: 3092 case IBV_EVENT_PATH_MIG: 3093 case IBV_EVENT_PATH_MIG_ERR: 3094 SPDK_NOTICELOG("Async event: %s\n", 3095 ibv_event_type_str(event.event_type)); 3096 rqpair = event.element.qp->qp_context; 3097 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3098 (uintptr_t)rqpair->cm_id, event.event_type); 3099 nvmf_rdma_update_ibv_state(rqpair); 3100 break; 3101 case IBV_EVENT_CQ_ERR: 3102 case IBV_EVENT_DEVICE_FATAL: 3103 case IBV_EVENT_PORT_ACTIVE: 3104 case IBV_EVENT_PORT_ERR: 3105 case IBV_EVENT_LID_CHANGE: 3106 case IBV_EVENT_PKEY_CHANGE: 3107 case IBV_EVENT_SM_CHANGE: 3108 case IBV_EVENT_SRQ_ERR: 3109 case IBV_EVENT_SRQ_LIMIT_REACHED: 3110 case IBV_EVENT_CLIENT_REREGISTER: 3111 case IBV_EVENT_GID_CHANGE: 3112 default: 3113 SPDK_NOTICELOG("Async event: %s\n", 3114 ibv_event_type_str(event.event_type)); 3115 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3116 break; 3117 } 3118 ibv_ack_async_event(&event); 3119 3120 return 0; 3121 } 3122 3123 static void 3124 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3125 { 3126 int rc = 0; 3127 uint32_t i = 0; 3128 3129 for (i = 0; i < max_events; i++) { 3130 rc = nvmf_process_ib_event(device); 3131 if (rc) { 3132 break; 3133 } 3134 } 3135 3136 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3137 } 3138 3139 static uint32_t 3140 nvmf_rdma_accept(struct spdk_nvmf_transport *transport) 3141 { 3142 int nfds, i = 0; 3143 struct spdk_nvmf_rdma_transport *rtransport; 3144 struct spdk_nvmf_rdma_device *device, *tmp; 3145 uint32_t count; 3146 3147 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3148 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3149 3150 if (nfds <= 0) { 3151 return 0; 3152 } 3153 3154 /* The first poll descriptor is RDMA CM event */ 3155 if (rtransport->poll_fds[i++].revents & POLLIN) { 3156 nvmf_process_cm_event(transport); 3157 nfds--; 3158 } 3159 3160 if (nfds == 0) { 3161 return count; 3162 } 3163 3164 /* Second and subsequent poll descriptors are IB async events */ 3165 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3166 if (rtransport->poll_fds[i++].revents & POLLIN) { 3167 nvmf_process_ib_events(device, 32); 3168 nfds--; 3169 } 3170 } 3171 /* check all flagged fd's have been served */ 3172 assert(nfds == 0); 3173 3174 return count; 3175 } 3176 3177 static void 3178 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3179 struct spdk_nvmf_ctrlr_data *cdata) 3180 { 3181 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3182 3183 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3184 since in-capsule data only works with NVME drives that support SGL memory layout */ 3185 if (transport->opts.dif_insert_or_strip) { 3186 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3187 } 3188 } 3189 3190 static void 3191 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3192 struct spdk_nvme_transport_id *trid, 3193 struct spdk_nvmf_discovery_log_page_entry *entry) 3194 { 3195 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3196 entry->adrfam = trid->adrfam; 3197 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3198 3199 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3200 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3201 3202 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3203 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3204 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3205 } 3206 3207 static void 3208 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3209 3210 static struct spdk_nvmf_transport_poll_group * 3211 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3212 { 3213 struct spdk_nvmf_rdma_transport *rtransport; 3214 struct spdk_nvmf_rdma_poll_group *rgroup; 3215 struct spdk_nvmf_rdma_poller *poller; 3216 struct spdk_nvmf_rdma_device *device; 3217 struct ibv_srq_init_attr srq_init_attr; 3218 struct spdk_nvmf_rdma_resource_opts opts; 3219 int num_cqe; 3220 3221 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3222 3223 rgroup = calloc(1, sizeof(*rgroup)); 3224 if (!rgroup) { 3225 return NULL; 3226 } 3227 3228 TAILQ_INIT(&rgroup->pollers); 3229 3230 pthread_mutex_lock(&rtransport->lock); 3231 TAILQ_FOREACH(device, &rtransport->devices, link) { 3232 poller = calloc(1, sizeof(*poller)); 3233 if (!poller) { 3234 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3235 nvmf_rdma_poll_group_destroy(&rgroup->group); 3236 pthread_mutex_unlock(&rtransport->lock); 3237 return NULL; 3238 } 3239 3240 poller->device = device; 3241 poller->group = rgroup; 3242 3243 TAILQ_INIT(&poller->qpairs); 3244 STAILQ_INIT(&poller->qpairs_pending_send); 3245 STAILQ_INIT(&poller->qpairs_pending_recv); 3246 3247 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3248 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3249 poller->max_srq_depth = rtransport->rdma_opts.max_srq_depth; 3250 3251 device->num_srq++; 3252 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 3253 srq_init_attr.attr.max_wr = poller->max_srq_depth; 3254 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3255 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 3256 if (!poller->srq) { 3257 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3258 nvmf_rdma_poll_group_destroy(&rgroup->group); 3259 pthread_mutex_unlock(&rtransport->lock); 3260 return NULL; 3261 } 3262 3263 opts.qp = poller->srq; 3264 opts.pd = device->pd; 3265 opts.qpair = NULL; 3266 opts.shared = true; 3267 opts.max_queue_depth = poller->max_srq_depth; 3268 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3269 3270 poller->resources = nvmf_rdma_resources_create(&opts); 3271 if (!poller->resources) { 3272 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3273 nvmf_rdma_poll_group_destroy(&rgroup->group); 3274 pthread_mutex_unlock(&rtransport->lock); 3275 return NULL; 3276 } 3277 } 3278 3279 /* 3280 * When using an srq, we can limit the completion queue at startup. 3281 * The following formula represents the calculation: 3282 * num_cqe = num_recv + num_data_wr + num_send_wr. 3283 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3284 */ 3285 if (poller->srq) { 3286 num_cqe = poller->max_srq_depth * 3; 3287 } else { 3288 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3289 } 3290 3291 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3292 if (!poller->cq) { 3293 SPDK_ERRLOG("Unable to create completion queue\n"); 3294 nvmf_rdma_poll_group_destroy(&rgroup->group); 3295 pthread_mutex_unlock(&rtransport->lock); 3296 return NULL; 3297 } 3298 poller->num_cqe = num_cqe; 3299 } 3300 3301 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3302 if (rtransport->conn_sched.next_admin_pg == NULL) { 3303 rtransport->conn_sched.next_admin_pg = rgroup; 3304 rtransport->conn_sched.next_io_pg = rgroup; 3305 } 3306 3307 pthread_mutex_unlock(&rtransport->lock); 3308 return &rgroup->group; 3309 } 3310 3311 static struct spdk_nvmf_transport_poll_group * 3312 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3313 { 3314 struct spdk_nvmf_rdma_transport *rtransport; 3315 struct spdk_nvmf_rdma_poll_group **pg; 3316 struct spdk_nvmf_transport_poll_group *result; 3317 3318 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3319 3320 pthread_mutex_lock(&rtransport->lock); 3321 3322 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3323 pthread_mutex_unlock(&rtransport->lock); 3324 return NULL; 3325 } 3326 3327 if (qpair->qid == 0) { 3328 pg = &rtransport->conn_sched.next_admin_pg; 3329 } else { 3330 pg = &rtransport->conn_sched.next_io_pg; 3331 } 3332 3333 assert(*pg != NULL); 3334 3335 result = &(*pg)->group; 3336 3337 *pg = TAILQ_NEXT(*pg, link); 3338 if (*pg == NULL) { 3339 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3340 } 3341 3342 pthread_mutex_unlock(&rtransport->lock); 3343 3344 return result; 3345 } 3346 3347 static void 3348 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3349 { 3350 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3351 struct spdk_nvmf_rdma_poller *poller, *tmp; 3352 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3353 struct spdk_nvmf_rdma_transport *rtransport; 3354 3355 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3356 if (!rgroup) { 3357 return; 3358 } 3359 3360 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3361 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3362 3363 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3364 nvmf_rdma_qpair_destroy(qpair); 3365 } 3366 3367 if (poller->srq) { 3368 if (poller->resources) { 3369 nvmf_rdma_resources_destroy(poller->resources); 3370 } 3371 ibv_destroy_srq(poller->srq); 3372 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 3373 } 3374 3375 if (poller->cq) { 3376 ibv_destroy_cq(poller->cq); 3377 } 3378 3379 free(poller); 3380 } 3381 3382 if (rgroup->group.transport == NULL) { 3383 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3384 * calls this function directly in a failure path. */ 3385 free(rgroup); 3386 return; 3387 } 3388 3389 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3390 3391 pthread_mutex_lock(&rtransport->lock); 3392 next_rgroup = TAILQ_NEXT(rgroup, link); 3393 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3394 if (next_rgroup == NULL) { 3395 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3396 } 3397 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3398 rtransport->conn_sched.next_admin_pg = next_rgroup; 3399 } 3400 if (rtransport->conn_sched.next_io_pg == rgroup) { 3401 rtransport->conn_sched.next_io_pg = next_rgroup; 3402 } 3403 pthread_mutex_unlock(&rtransport->lock); 3404 3405 free(rgroup); 3406 } 3407 3408 static void 3409 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3410 { 3411 if (rqpair->cm_id != NULL) { 3412 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3413 } 3414 } 3415 3416 static int 3417 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3418 struct spdk_nvmf_qpair *qpair) 3419 { 3420 struct spdk_nvmf_rdma_poll_group *rgroup; 3421 struct spdk_nvmf_rdma_qpair *rqpair; 3422 struct spdk_nvmf_rdma_device *device; 3423 struct spdk_nvmf_rdma_poller *poller; 3424 int rc; 3425 3426 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3427 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3428 3429 device = rqpair->device; 3430 3431 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3432 if (poller->device == device) { 3433 break; 3434 } 3435 } 3436 3437 if (!poller) { 3438 SPDK_ERRLOG("No poller found for device.\n"); 3439 return -1; 3440 } 3441 3442 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3443 rqpair->poller = poller; 3444 rqpair->srq = rqpair->poller->srq; 3445 3446 rc = nvmf_rdma_qpair_initialize(qpair); 3447 if (rc < 0) { 3448 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3449 return -1; 3450 } 3451 3452 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3453 if (rc) { 3454 /* Try to reject, but we probably can't */ 3455 nvmf_rdma_qpair_reject_connection(rqpair); 3456 return -1; 3457 } 3458 3459 nvmf_rdma_update_ibv_state(rqpair); 3460 3461 return 0; 3462 } 3463 3464 static int 3465 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3466 struct spdk_nvmf_qpair *qpair) 3467 { 3468 struct spdk_nvmf_rdma_qpair *rqpair; 3469 3470 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3471 assert(group->transport->tgt != NULL); 3472 3473 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3474 3475 if (!rqpair->destruct_channel) { 3476 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3477 return 0; 3478 } 3479 3480 /* Sanity check that we get io_channel on the correct thread */ 3481 if (qpair->group) { 3482 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3483 } 3484 3485 return 0; 3486 } 3487 3488 static int 3489 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3490 { 3491 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3492 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3493 struct spdk_nvmf_rdma_transport, transport); 3494 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3495 struct spdk_nvmf_rdma_qpair, qpair); 3496 3497 /* 3498 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3499 * needs to be returned to the shared receive queue or the poll group will eventually be 3500 * starved of RECV structures. 3501 */ 3502 if (rqpair->srq && rdma_req->recv) { 3503 int rc; 3504 struct ibv_recv_wr *bad_recv_wr; 3505 3506 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 3507 if (rc) { 3508 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3509 } 3510 } 3511 3512 _nvmf_rdma_request_free(rdma_req, rtransport); 3513 return 0; 3514 } 3515 3516 static int 3517 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3518 { 3519 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3520 struct spdk_nvmf_rdma_transport, transport); 3521 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3522 struct spdk_nvmf_rdma_request, req); 3523 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3524 struct spdk_nvmf_rdma_qpair, qpair); 3525 3526 if (rqpair->ibv_state != IBV_QPS_ERR) { 3527 /* The connection is alive, so process the request as normal */ 3528 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3529 } else { 3530 /* The connection is dead. Move the request directly to the completed state. */ 3531 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3532 } 3533 3534 nvmf_rdma_request_process(rtransport, rdma_req); 3535 3536 return 0; 3537 } 3538 3539 static void 3540 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 3541 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3542 { 3543 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3544 3545 rqpair->to_close = true; 3546 3547 /* This happens only when the qpair is disconnected before 3548 * it is added to the poll group. Since there is no poll group, 3549 * the RDMA qp has not been initialized yet and the RDMA CM 3550 * event has not yet been acknowledged, so we need to reject it. 3551 */ 3552 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3553 nvmf_rdma_qpair_reject_connection(rqpair); 3554 nvmf_rdma_qpair_destroy(rqpair); 3555 return; 3556 } 3557 3558 if (rqpair->rdma_qp) { 3559 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3560 } 3561 3562 nvmf_rdma_destroy_drained_qpair(rqpair); 3563 3564 if (cb_fn) { 3565 cb_fn(cb_arg); 3566 } 3567 } 3568 3569 static struct spdk_nvmf_rdma_qpair * 3570 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3571 { 3572 struct spdk_nvmf_rdma_qpair *rqpair; 3573 /* @todo: improve QP search */ 3574 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3575 if (wc->qp_num == rqpair->rdma_qp->qp->qp_num) { 3576 return rqpair; 3577 } 3578 } 3579 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3580 return NULL; 3581 } 3582 3583 #ifdef DEBUG 3584 static int 3585 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3586 { 3587 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3588 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3589 } 3590 #endif 3591 3592 static void 3593 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3594 int rc) 3595 { 3596 struct spdk_nvmf_rdma_recv *rdma_recv; 3597 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3598 3599 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3600 while (bad_recv_wr != NULL) { 3601 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3602 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3603 3604 rdma_recv->qpair->current_recv_depth++; 3605 bad_recv_wr = bad_recv_wr->next; 3606 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3607 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3608 } 3609 } 3610 3611 static void 3612 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3613 { 3614 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3615 while (bad_recv_wr != NULL) { 3616 bad_recv_wr = bad_recv_wr->next; 3617 rqpair->current_recv_depth++; 3618 } 3619 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3620 } 3621 3622 static void 3623 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3624 struct spdk_nvmf_rdma_poller *rpoller) 3625 { 3626 struct spdk_nvmf_rdma_qpair *rqpair; 3627 struct ibv_recv_wr *bad_recv_wr; 3628 int rc; 3629 3630 if (rpoller->srq) { 3631 if (rpoller->resources->recvs_to_post.first != NULL) { 3632 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3633 if (rc) { 3634 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3635 } 3636 rpoller->resources->recvs_to_post.first = NULL; 3637 rpoller->resources->recvs_to_post.last = NULL; 3638 } 3639 } else { 3640 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3641 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3642 assert(rqpair->resources->recvs_to_post.first != NULL); 3643 rc = ibv_post_recv(rqpair->rdma_qp->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3644 if (rc) { 3645 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3646 } 3647 rqpair->resources->recvs_to_post.first = NULL; 3648 rqpair->resources->recvs_to_post.last = NULL; 3649 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3650 } 3651 } 3652 } 3653 3654 static void 3655 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3656 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3657 { 3658 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3659 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3660 3661 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3662 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3663 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3664 assert(rqpair->current_send_depth > 0); 3665 rqpair->current_send_depth--; 3666 switch (bad_rdma_wr->type) { 3667 case RDMA_WR_TYPE_DATA: 3668 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3669 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3670 assert(rqpair->current_read_depth > 0); 3671 rqpair->current_read_depth--; 3672 } 3673 break; 3674 case RDMA_WR_TYPE_SEND: 3675 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3676 break; 3677 default: 3678 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3679 prev_rdma_req = cur_rdma_req; 3680 continue; 3681 } 3682 3683 if (prev_rdma_req == cur_rdma_req) { 3684 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3685 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3686 continue; 3687 } 3688 3689 switch (cur_rdma_req->state) { 3690 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3691 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3692 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3693 break; 3694 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3695 case RDMA_REQUEST_STATE_COMPLETING: 3696 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3697 break; 3698 default: 3699 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3700 cur_rdma_req->state, rqpair); 3701 continue; 3702 } 3703 3704 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3705 prev_rdma_req = cur_rdma_req; 3706 } 3707 3708 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3709 /* Disconnect the connection. */ 3710 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3711 } 3712 3713 } 3714 3715 static void 3716 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3717 struct spdk_nvmf_rdma_poller *rpoller) 3718 { 3719 struct spdk_nvmf_rdma_qpair *rqpair; 3720 struct ibv_send_wr *bad_wr = NULL; 3721 int rc; 3722 3723 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3724 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3725 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3726 3727 /* bad wr always points to the first wr that failed. */ 3728 if (rc) { 3729 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3730 } 3731 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3732 } 3733 } 3734 3735 static const char * 3736 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 3737 { 3738 switch (wr_type) { 3739 case RDMA_WR_TYPE_RECV: 3740 return "RECV"; 3741 case RDMA_WR_TYPE_SEND: 3742 return "SEND"; 3743 case RDMA_WR_TYPE_DATA: 3744 return "DATA"; 3745 default: 3746 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 3747 SPDK_UNREACHABLE(); 3748 } 3749 } 3750 3751 static inline void 3752 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 3753 { 3754 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 3755 3756 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 3757 /* If qpair is in ERR state, we will receive completions for all posted and not completed 3758 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 3759 SPDK_DEBUGLOG(rdma, 3760 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3761 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3762 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3763 } else { 3764 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3765 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3766 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3767 } 3768 } 3769 3770 static int 3771 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3772 struct spdk_nvmf_rdma_poller *rpoller) 3773 { 3774 struct ibv_wc wc[32]; 3775 struct spdk_nvmf_rdma_wr *rdma_wr; 3776 struct spdk_nvmf_rdma_request *rdma_req; 3777 struct spdk_nvmf_rdma_recv *rdma_recv; 3778 struct spdk_nvmf_rdma_qpair *rqpair; 3779 int reaped, i; 3780 int count = 0; 3781 bool error = false; 3782 uint64_t poll_tsc = spdk_get_ticks(); 3783 3784 /* Poll for completing operations. */ 3785 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3786 if (reaped < 0) { 3787 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3788 errno, spdk_strerror(errno)); 3789 return -1; 3790 } 3791 3792 rpoller->stat.polls++; 3793 rpoller->stat.completions += reaped; 3794 3795 for (i = 0; i < reaped; i++) { 3796 3797 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3798 3799 switch (rdma_wr->type) { 3800 case RDMA_WR_TYPE_SEND: 3801 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3802 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3803 3804 if (!wc[i].status) { 3805 count++; 3806 assert(wc[i].opcode == IBV_WC_SEND); 3807 assert(nvmf_rdma_req_is_completing(rdma_req)); 3808 } 3809 3810 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3811 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3812 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3813 rdma_req->num_outstanding_data_wr = 0; 3814 3815 nvmf_rdma_request_process(rtransport, rdma_req); 3816 break; 3817 case RDMA_WR_TYPE_RECV: 3818 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3819 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3820 if (rpoller->srq != NULL) { 3821 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3822 /* It is possible that there are still some completions for destroyed QP 3823 * associated with SRQ. We just ignore these late completions and re-post 3824 * receive WRs back to SRQ. 3825 */ 3826 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3827 struct ibv_recv_wr *bad_wr; 3828 int rc; 3829 3830 rdma_recv->wr.next = NULL; 3831 rc = ibv_post_srq_recv(rpoller->srq, 3832 &rdma_recv->wr, 3833 &bad_wr); 3834 if (rc) { 3835 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3836 } 3837 continue; 3838 } 3839 } 3840 rqpair = rdma_recv->qpair; 3841 3842 assert(rqpair != NULL); 3843 if (!wc[i].status) { 3844 assert(wc[i].opcode == IBV_WC_RECV); 3845 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3846 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3847 break; 3848 } 3849 } 3850 3851 rdma_recv->wr.next = NULL; 3852 rqpair->current_recv_depth++; 3853 rdma_recv->receive_tsc = poll_tsc; 3854 rpoller->stat.requests++; 3855 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3856 break; 3857 case RDMA_WR_TYPE_DATA: 3858 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3859 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3860 3861 assert(rdma_req->num_outstanding_data_wr > 0); 3862 3863 rqpair->current_send_depth--; 3864 rdma_req->num_outstanding_data_wr--; 3865 if (!wc[i].status) { 3866 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3867 rqpair->current_read_depth--; 3868 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3869 if (rdma_req->num_outstanding_data_wr == 0) { 3870 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3871 nvmf_rdma_request_process(rtransport, rdma_req); 3872 } 3873 } else { 3874 /* If the data transfer fails still force the queue into the error state, 3875 * if we were performing an RDMA_READ, we need to force the request into a 3876 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3877 * case, we should wait for the SEND to complete. */ 3878 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3879 rqpair->current_read_depth--; 3880 if (rdma_req->num_outstanding_data_wr == 0) { 3881 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3882 } 3883 } 3884 } 3885 break; 3886 default: 3887 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3888 continue; 3889 } 3890 3891 /* Handle error conditions */ 3892 if (wc[i].status) { 3893 nvmf_rdma_update_ibv_state(rqpair); 3894 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 3895 3896 error = true; 3897 3898 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3899 /* Disconnect the connection. */ 3900 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3901 } else { 3902 nvmf_rdma_destroy_drained_qpair(rqpair); 3903 } 3904 continue; 3905 } 3906 3907 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3908 3909 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3910 nvmf_rdma_destroy_drained_qpair(rqpair); 3911 } 3912 } 3913 3914 if (error == true) { 3915 return -1; 3916 } 3917 3918 /* submit outstanding work requests. */ 3919 _poller_submit_recvs(rtransport, rpoller); 3920 _poller_submit_sends(rtransport, rpoller); 3921 3922 return count; 3923 } 3924 3925 static int 3926 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3927 { 3928 struct spdk_nvmf_rdma_transport *rtransport; 3929 struct spdk_nvmf_rdma_poll_group *rgroup; 3930 struct spdk_nvmf_rdma_poller *rpoller; 3931 int count, rc; 3932 3933 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3934 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3935 3936 count = 0; 3937 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3938 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 3939 if (rc < 0) { 3940 return rc; 3941 } 3942 count += rc; 3943 } 3944 3945 return count; 3946 } 3947 3948 static int 3949 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3950 struct spdk_nvme_transport_id *trid, 3951 bool peer) 3952 { 3953 struct sockaddr *saddr; 3954 uint16_t port; 3955 3956 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 3957 3958 if (peer) { 3959 saddr = rdma_get_peer_addr(id); 3960 } else { 3961 saddr = rdma_get_local_addr(id); 3962 } 3963 switch (saddr->sa_family) { 3964 case AF_INET: { 3965 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3966 3967 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3968 inet_ntop(AF_INET, &saddr_in->sin_addr, 3969 trid->traddr, sizeof(trid->traddr)); 3970 if (peer) { 3971 port = ntohs(rdma_get_dst_port(id)); 3972 } else { 3973 port = ntohs(rdma_get_src_port(id)); 3974 } 3975 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3976 break; 3977 } 3978 case AF_INET6: { 3979 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3980 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3981 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3982 trid->traddr, sizeof(trid->traddr)); 3983 if (peer) { 3984 port = ntohs(rdma_get_dst_port(id)); 3985 } else { 3986 port = ntohs(rdma_get_src_port(id)); 3987 } 3988 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3989 break; 3990 } 3991 default: 3992 return -1; 3993 3994 } 3995 3996 return 0; 3997 } 3998 3999 static int 4000 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4001 struct spdk_nvme_transport_id *trid) 4002 { 4003 struct spdk_nvmf_rdma_qpair *rqpair; 4004 4005 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4006 4007 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4008 } 4009 4010 static int 4011 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4012 struct spdk_nvme_transport_id *trid) 4013 { 4014 struct spdk_nvmf_rdma_qpair *rqpair; 4015 4016 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4017 4018 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4019 } 4020 4021 static int 4022 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4023 struct spdk_nvme_transport_id *trid) 4024 { 4025 struct spdk_nvmf_rdma_qpair *rqpair; 4026 4027 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4028 4029 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4030 } 4031 4032 void 4033 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4034 { 4035 g_nvmf_hooks = *hooks; 4036 } 4037 4038 static void 4039 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4040 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4041 { 4042 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4043 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4044 4045 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4046 4047 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4048 } 4049 4050 static int 4051 _nvmf_rdma_qpair_abort_request(void *ctx) 4052 { 4053 struct spdk_nvmf_request *req = ctx; 4054 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4055 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4056 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4057 struct spdk_nvmf_rdma_qpair, qpair); 4058 int rc; 4059 4060 spdk_poller_unregister(&req->poller); 4061 4062 switch (rdma_req_to_abort->state) { 4063 case RDMA_REQUEST_STATE_EXECUTING: 4064 rc = nvmf_ctrlr_abort_request(req); 4065 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4066 return SPDK_POLLER_BUSY; 4067 } 4068 break; 4069 4070 case RDMA_REQUEST_STATE_NEED_BUFFER: 4071 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4072 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4073 4074 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4075 break; 4076 4077 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4078 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4079 spdk_nvmf_rdma_request, state_link); 4080 4081 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4082 break; 4083 4084 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4085 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4086 spdk_nvmf_rdma_request, state_link); 4087 4088 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4089 break; 4090 4091 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4092 if (spdk_get_ticks() < req->timeout_tsc) { 4093 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4094 return SPDK_POLLER_BUSY; 4095 } 4096 break; 4097 4098 default: 4099 break; 4100 } 4101 4102 spdk_nvmf_request_complete(req); 4103 return SPDK_POLLER_BUSY; 4104 } 4105 4106 static void 4107 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4108 struct spdk_nvmf_request *req) 4109 { 4110 struct spdk_nvmf_rdma_qpair *rqpair; 4111 struct spdk_nvmf_rdma_transport *rtransport; 4112 struct spdk_nvmf_transport *transport; 4113 uint16_t cid; 4114 uint32_t i, max_req_count; 4115 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 4116 4117 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4118 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4119 transport = &rtransport->transport; 4120 4121 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4122 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 4123 4124 for (i = 0; i < max_req_count; i++) { 4125 rdma_req = &rqpair->resources->reqs[i]; 4126 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 4127 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 4128 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 4129 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 4130 rdma_req->req.qpair == qpair) { 4131 rdma_req_to_abort = rdma_req; 4132 break; 4133 } 4134 } 4135 4136 if (rdma_req_to_abort == NULL) { 4137 spdk_nvmf_request_complete(req); 4138 return; 4139 } 4140 4141 req->req_to_abort = &rdma_req_to_abort->req; 4142 req->timeout_tsc = spdk_get_ticks() + 4143 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4144 req->poller = NULL; 4145 4146 _nvmf_rdma_qpair_abort_request(req); 4147 } 4148 4149 static int 4150 nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4151 struct spdk_nvmf_transport_poll_group_stat **stat) 4152 { 4153 struct spdk_io_channel *ch; 4154 struct spdk_nvmf_poll_group *group; 4155 struct spdk_nvmf_transport_poll_group *tgroup; 4156 struct spdk_nvmf_rdma_poll_group *rgroup; 4157 struct spdk_nvmf_rdma_poller *rpoller; 4158 struct spdk_nvmf_rdma_device_stat *device_stat; 4159 uint64_t num_devices = 0; 4160 4161 if (tgt == NULL || stat == NULL) { 4162 return -EINVAL; 4163 } 4164 4165 ch = spdk_get_io_channel(tgt); 4166 group = spdk_io_channel_get_ctx(ch);; 4167 spdk_put_io_channel(ch); 4168 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4169 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4170 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4171 if (!*stat) { 4172 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4173 return -ENOMEM; 4174 } 4175 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4176 4177 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4178 /* Count devices to allocate enough memory */ 4179 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4180 ++num_devices; 4181 } 4182 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4183 if (!(*stat)->rdma.devices) { 4184 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4185 free(*stat); 4186 return -ENOMEM; 4187 } 4188 4189 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4190 (*stat)->rdma.num_devices = num_devices; 4191 num_devices = 0; 4192 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4193 device_stat = &(*stat)->rdma.devices[num_devices++]; 4194 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4195 device_stat->polls = rpoller->stat.polls; 4196 device_stat->completions = rpoller->stat.completions; 4197 device_stat->requests = rpoller->stat.requests; 4198 device_stat->request_latency = rpoller->stat.request_latency; 4199 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4200 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4201 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4202 } 4203 return 0; 4204 } 4205 } 4206 return -ENOENT; 4207 } 4208 4209 static void 4210 nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4211 { 4212 if (stat) { 4213 free(stat->rdma.devices); 4214 } 4215 free(stat); 4216 } 4217 4218 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4219 .name = "RDMA", 4220 .type = SPDK_NVME_TRANSPORT_RDMA, 4221 .opts_init = nvmf_rdma_opts_init, 4222 .create = nvmf_rdma_create, 4223 .dump_opts = nvmf_rdma_dump_opts, 4224 .destroy = nvmf_rdma_destroy, 4225 4226 .listen = nvmf_rdma_listen, 4227 .stop_listen = nvmf_rdma_stop_listen, 4228 .accept = nvmf_rdma_accept, 4229 .cdata_init = nvmf_rdma_cdata_init, 4230 4231 .listener_discover = nvmf_rdma_discover, 4232 4233 .poll_group_create = nvmf_rdma_poll_group_create, 4234 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4235 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4236 .poll_group_add = nvmf_rdma_poll_group_add, 4237 .poll_group_remove = nvmf_rdma_poll_group_remove, 4238 .poll_group_poll = nvmf_rdma_poll_group_poll, 4239 4240 .req_free = nvmf_rdma_request_free, 4241 .req_complete = nvmf_rdma_request_complete, 4242 4243 .qpair_fini = nvmf_rdma_close_qpair, 4244 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4245 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4246 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4247 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4248 4249 .poll_group_get_stat = nvmf_rdma_poll_group_get_stat, 4250 .poll_group_free_stat = nvmf_rdma_poll_group_free_stat, 4251 }; 4252 4253 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4254 SPDK_LOG_REGISTER_COMPONENT(rdma) 4255