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, 162 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 163 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 164 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 165 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 166 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 167 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 168 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 169 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 170 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 171 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 172 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 173 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 174 spdk_trace_register_description("RDMA_REQ_TX_H2C", 175 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 176 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 177 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 178 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 179 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 180 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 181 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 182 spdk_trace_register_description("RDMA_REQ_EXECUTING", 183 TRACE_RDMA_REQUEST_STATE_EXECUTING, 184 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 185 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 186 spdk_trace_register_description("RDMA_REQ_EXECUTED", 187 TRACE_RDMA_REQUEST_STATE_EXECUTED, 188 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 189 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 190 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 191 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 192 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 193 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 194 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 195 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 196 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 197 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 198 spdk_trace_register_description("RDMA_REQ_COMPLETING", 199 TRACE_RDMA_REQUEST_STATE_COMPLETING, 200 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 201 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 202 spdk_trace_register_description("RDMA_REQ_COMPLETED", 203 TRACE_RDMA_REQUEST_STATE_COMPLETED, 204 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 205 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 206 207 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 208 OWNER_NONE, OBJECT_NONE, 0, 209 SPDK_TRACE_ARG_TYPE_INT, ""); 210 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 211 OWNER_NONE, OBJECT_NONE, 0, 212 SPDK_TRACE_ARG_TYPE_INT, "type"); 213 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 214 OWNER_NONE, OBJECT_NONE, 0, 215 SPDK_TRACE_ARG_TYPE_INT, "type"); 216 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 217 OWNER_NONE, OBJECT_NONE, 0, 218 SPDK_TRACE_ARG_TYPE_PTR, "state"); 219 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 220 OWNER_NONE, OBJECT_NONE, 0, 221 SPDK_TRACE_ARG_TYPE_INT, ""); 222 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 223 OWNER_NONE, OBJECT_NONE, 0, 224 SPDK_TRACE_ARG_TYPE_INT, ""); 225 } 226 227 enum spdk_nvmf_rdma_wr_type { 228 RDMA_WR_TYPE_RECV, 229 RDMA_WR_TYPE_SEND, 230 RDMA_WR_TYPE_DATA, 231 }; 232 233 struct spdk_nvmf_rdma_wr { 234 enum spdk_nvmf_rdma_wr_type type; 235 }; 236 237 /* This structure holds commands as they are received off the wire. 238 * It must be dynamically paired with a full request object 239 * (spdk_nvmf_rdma_request) to service a request. It is separate 240 * from the request because RDMA does not appear to order 241 * completions, so occasionally we'll get a new incoming 242 * command when there aren't any free request objects. 243 */ 244 struct spdk_nvmf_rdma_recv { 245 struct ibv_recv_wr wr; 246 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 247 248 struct spdk_nvmf_rdma_qpair *qpair; 249 250 /* In-capsule data buffer */ 251 uint8_t *buf; 252 253 struct spdk_nvmf_rdma_wr rdma_wr; 254 uint64_t receive_tsc; 255 256 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 257 }; 258 259 struct spdk_nvmf_rdma_request_data { 260 struct spdk_nvmf_rdma_wr rdma_wr; 261 struct ibv_send_wr wr; 262 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 263 }; 264 265 struct spdk_nvmf_rdma_request { 266 struct spdk_nvmf_request req; 267 268 enum spdk_nvmf_rdma_request_state state; 269 270 /* Data offset in req.iov */ 271 uint32_t offset; 272 273 struct spdk_nvmf_rdma_recv *recv; 274 275 struct { 276 struct spdk_nvmf_rdma_wr rdma_wr; 277 struct ibv_send_wr wr; 278 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 279 } rsp; 280 281 struct spdk_nvmf_rdma_request_data data; 282 283 uint32_t iovpos; 284 285 uint32_t num_outstanding_data_wr; 286 uint64_t receive_tsc; 287 288 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 289 }; 290 291 struct spdk_nvmf_rdma_resource_opts { 292 struct spdk_nvmf_rdma_qpair *qpair; 293 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 294 void *qp; 295 struct ibv_pd *pd; 296 uint32_t max_queue_depth; 297 uint32_t in_capsule_data_size; 298 bool shared; 299 }; 300 301 struct spdk_nvmf_rdma_resources { 302 /* Array of size "max_queue_depth" containing RDMA requests. */ 303 struct spdk_nvmf_rdma_request *reqs; 304 305 /* Array of size "max_queue_depth" containing RDMA recvs. */ 306 struct spdk_nvmf_rdma_recv *recvs; 307 308 /* Array of size "max_queue_depth" containing 64 byte capsules 309 * used for receive. 310 */ 311 union nvmf_h2c_msg *cmds; 312 struct ibv_mr *cmds_mr; 313 314 /* Array of size "max_queue_depth" containing 16 byte completions 315 * to be sent back to the user. 316 */ 317 union nvmf_c2h_msg *cpls; 318 struct ibv_mr *cpls_mr; 319 320 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 321 * buffers to be used for in capsule data. 322 */ 323 void *bufs; 324 struct ibv_mr *bufs_mr; 325 326 /* Receives that are waiting for a request object */ 327 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 328 329 /* Queue to track free requests */ 330 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 331 }; 332 333 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 334 335 struct spdk_nvmf_rdma_ibv_event_ctx { 336 struct spdk_nvmf_rdma_qpair *rqpair; 337 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 338 /* Link to other ibv events associated with this qpair */ 339 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 340 }; 341 342 struct spdk_nvmf_rdma_qpair { 343 struct spdk_nvmf_qpair qpair; 344 345 struct spdk_nvmf_rdma_device *device; 346 struct spdk_nvmf_rdma_poller *poller; 347 348 struct spdk_rdma_qp *rdma_qp; 349 struct rdma_cm_id *cm_id; 350 struct spdk_rdma_srq *srq; 351 struct rdma_cm_id *listen_id; 352 353 /* The maximum number of I/O outstanding on this connection at one time */ 354 uint16_t max_queue_depth; 355 356 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 357 uint16_t max_read_depth; 358 359 /* The maximum number of RDMA SEND operations at one time */ 360 uint32_t max_send_depth; 361 362 /* The current number of outstanding WRs from this qpair's 363 * recv queue. Should not exceed device->attr.max_queue_depth. 364 */ 365 uint16_t current_recv_depth; 366 367 /* The current number of active RDMA READ operations */ 368 uint16_t current_read_depth; 369 370 /* The current number of posted WRs from this qpair's 371 * send queue. Should not exceed max_send_depth. 372 */ 373 uint32_t current_send_depth; 374 375 /* The maximum number of SGEs per WR on the send queue */ 376 uint32_t max_send_sge; 377 378 /* The maximum number of SGEs per WR on the recv queue */ 379 uint32_t max_recv_sge; 380 381 struct spdk_nvmf_rdma_resources *resources; 382 383 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 384 385 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 386 387 /* Number of requests not in the free state */ 388 uint32_t qd; 389 390 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 391 392 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 393 394 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 395 396 /* IBV queue pair attributes: they are used to manage 397 * qp state and recover from errors. 398 */ 399 enum ibv_qp_state ibv_state; 400 401 /* 402 * io_channel which is used to destroy qpair when it is removed from poll group 403 */ 404 struct spdk_io_channel *destruct_channel; 405 406 /* List of ibv async events */ 407 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 408 409 /* Lets us know that we have received the last_wqe event. */ 410 bool last_wqe_reached; 411 412 /* Indicate that nvmf_rdma_close_qpair is called */ 413 bool to_close; 414 }; 415 416 struct spdk_nvmf_rdma_poller_stat { 417 uint64_t completions; 418 uint64_t polls; 419 uint64_t idle_polls; 420 uint64_t requests; 421 uint64_t request_latency; 422 uint64_t pending_free_request; 423 uint64_t pending_rdma_read; 424 uint64_t pending_rdma_write; 425 struct spdk_rdma_qp_stats qp_stats; 426 }; 427 428 struct spdk_nvmf_rdma_poller { 429 struct spdk_nvmf_rdma_device *device; 430 struct spdk_nvmf_rdma_poll_group *group; 431 432 int num_cqe; 433 int required_num_wr; 434 struct ibv_cq *cq; 435 436 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 437 uint16_t max_srq_depth; 438 439 /* Shared receive queue */ 440 struct spdk_rdma_srq *srq; 441 442 struct spdk_nvmf_rdma_resources *resources; 443 struct spdk_nvmf_rdma_poller_stat stat; 444 445 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 446 447 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 448 449 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 450 451 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 452 }; 453 454 struct spdk_nvmf_rdma_poll_group_stat { 455 uint64_t pending_data_buffer; 456 }; 457 458 struct spdk_nvmf_rdma_poll_group { 459 struct spdk_nvmf_transport_poll_group group; 460 struct spdk_nvmf_rdma_poll_group_stat stat; 461 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 462 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 463 }; 464 465 struct spdk_nvmf_rdma_conn_sched { 466 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 467 struct spdk_nvmf_rdma_poll_group *next_io_pg; 468 }; 469 470 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 471 struct spdk_nvmf_rdma_device { 472 struct ibv_device_attr attr; 473 struct ibv_context *context; 474 475 struct spdk_rdma_mem_map *map; 476 struct ibv_pd *pd; 477 478 int num_srq; 479 480 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 481 }; 482 483 struct spdk_nvmf_rdma_port { 484 const struct spdk_nvme_transport_id *trid; 485 struct rdma_cm_id *id; 486 struct spdk_nvmf_rdma_device *device; 487 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 488 }; 489 490 struct rdma_transport_opts { 491 int num_cqe; 492 uint32_t max_srq_depth; 493 bool no_srq; 494 bool no_wr_batching; 495 int acceptor_backlog; 496 }; 497 498 struct spdk_nvmf_rdma_transport { 499 struct spdk_nvmf_transport transport; 500 struct rdma_transport_opts rdma_opts; 501 502 struct spdk_nvmf_rdma_conn_sched conn_sched; 503 504 struct rdma_event_channel *event_channel; 505 506 struct spdk_mempool *data_wr_pool; 507 508 pthread_mutex_t lock; 509 510 /* fields used to poll RDMA/IB events */ 511 nfds_t npoll_fds; 512 struct pollfd *poll_fds; 513 514 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 515 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 516 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 517 }; 518 519 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 520 { 521 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe), 522 spdk_json_decode_int32, true 523 }, 524 { 525 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 526 spdk_json_decode_uint32, true 527 }, 528 { 529 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 530 spdk_json_decode_bool, true 531 }, 532 { 533 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 534 spdk_json_decode_bool, true 535 }, 536 { 537 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 538 spdk_json_decode_int32, true 539 }, 540 }; 541 542 static bool 543 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 544 struct spdk_nvmf_rdma_request *rdma_req); 545 546 static void 547 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 548 struct spdk_nvmf_rdma_poller *rpoller); 549 550 static void 551 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 552 struct spdk_nvmf_rdma_poller *rpoller); 553 554 static inline int 555 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 556 { 557 switch (state) { 558 case IBV_QPS_RESET: 559 case IBV_QPS_INIT: 560 case IBV_QPS_RTR: 561 case IBV_QPS_RTS: 562 case IBV_QPS_SQD: 563 case IBV_QPS_SQE: 564 case IBV_QPS_ERR: 565 return 0; 566 default: 567 return -1; 568 } 569 } 570 571 static inline enum spdk_nvme_media_error_status_code 572 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 573 enum spdk_nvme_media_error_status_code result; 574 switch (err_type) 575 { 576 case SPDK_DIF_REFTAG_ERROR: 577 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 578 break; 579 case SPDK_DIF_APPTAG_ERROR: 580 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 581 break; 582 case SPDK_DIF_GUARD_ERROR: 583 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 584 break; 585 default: 586 SPDK_UNREACHABLE(); 587 } 588 589 return result; 590 } 591 592 static enum ibv_qp_state 593 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 594 enum ibv_qp_state old_state, new_state; 595 struct ibv_qp_attr qp_attr; 596 struct ibv_qp_init_attr init_attr; 597 int rc; 598 599 old_state = rqpair->ibv_state; 600 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 601 g_spdk_nvmf_ibv_query_mask, &init_attr); 602 603 if (rc) 604 { 605 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 606 return IBV_QPS_ERR + 1; 607 } 608 609 new_state = qp_attr.qp_state; 610 rqpair->ibv_state = new_state; 611 qp_attr.ah_attr.port_num = qp_attr.port_num; 612 613 rc = nvmf_rdma_check_ibv_state(new_state); 614 if (rc) 615 { 616 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 617 /* 618 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 619 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 620 */ 621 return IBV_QPS_ERR + 1; 622 } 623 624 if (old_state != new_state) 625 { 626 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, (uintptr_t)rqpair, new_state); 627 } 628 return new_state; 629 } 630 631 static void 632 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 633 struct spdk_nvmf_rdma_transport *rtransport) 634 { 635 struct spdk_nvmf_rdma_request_data *data_wr; 636 struct ibv_send_wr *next_send_wr; 637 uint64_t req_wrid; 638 639 rdma_req->num_outstanding_data_wr = 0; 640 data_wr = &rdma_req->data; 641 req_wrid = data_wr->wr.wr_id; 642 while (data_wr && data_wr->wr.wr_id == req_wrid) { 643 memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge); 644 data_wr->wr.num_sge = 0; 645 next_send_wr = data_wr->wr.next; 646 if (data_wr != &rdma_req->data) { 647 spdk_mempool_put(rtransport->data_wr_pool, data_wr); 648 } 649 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : 650 SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr); 651 } 652 } 653 654 static void 655 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 656 { 657 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 658 if (req->req.cmd) { 659 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 660 } 661 if (req->recv) { 662 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 663 } 664 } 665 666 static void 667 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 668 { 669 int i; 670 671 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 672 for (i = 0; i < rqpair->max_queue_depth; i++) { 673 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 674 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 675 } 676 } 677 } 678 679 static void 680 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 681 { 682 if (resources->cmds_mr) { 683 ibv_dereg_mr(resources->cmds_mr); 684 } 685 686 if (resources->cpls_mr) { 687 ibv_dereg_mr(resources->cpls_mr); 688 } 689 690 if (resources->bufs_mr) { 691 ibv_dereg_mr(resources->bufs_mr); 692 } 693 694 spdk_free(resources->cmds); 695 spdk_free(resources->cpls); 696 spdk_free(resources->bufs); 697 free(resources->reqs); 698 free(resources->recvs); 699 free(resources); 700 } 701 702 703 static struct spdk_nvmf_rdma_resources * 704 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 705 { 706 struct spdk_nvmf_rdma_resources *resources; 707 struct spdk_nvmf_rdma_request *rdma_req; 708 struct spdk_nvmf_rdma_recv *rdma_recv; 709 struct spdk_rdma_qp *qp = NULL; 710 struct spdk_rdma_srq *srq = NULL; 711 struct ibv_recv_wr *bad_wr = NULL; 712 uint32_t i; 713 int rc = 0; 714 715 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 716 if (!resources) { 717 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 718 return NULL; 719 } 720 721 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 722 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 723 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 724 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 725 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 726 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 727 728 if (opts->in_capsule_data_size > 0) { 729 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 730 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 731 SPDK_MALLOC_DMA); 732 } 733 734 if (!resources->reqs || !resources->recvs || !resources->cmds || 735 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 736 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 737 goto cleanup; 738 } 739 740 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 741 opts->max_queue_depth * sizeof(*resources->cmds), 742 IBV_ACCESS_LOCAL_WRITE); 743 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 744 opts->max_queue_depth * sizeof(*resources->cpls), 745 0); 746 747 if (opts->in_capsule_data_size) { 748 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 749 opts->max_queue_depth * 750 opts->in_capsule_data_size, 751 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 752 } 753 754 if (!resources->cmds_mr || !resources->cpls_mr || 755 (opts->in_capsule_data_size && 756 !resources->bufs_mr)) { 757 goto cleanup; 758 } 759 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx LKey: %x\n", 760 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 761 resources->cmds_mr->lkey); 762 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx LKey: %x\n", 763 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 764 resources->cpls_mr->lkey); 765 if (resources->bufs && resources->bufs_mr) { 766 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x LKey: %x\n", 767 resources->bufs, opts->max_queue_depth * 768 opts->in_capsule_data_size, resources->bufs_mr->lkey); 769 } 770 771 /* Initialize queues */ 772 STAILQ_INIT(&resources->incoming_queue); 773 STAILQ_INIT(&resources->free_queue); 774 775 if (opts->shared) { 776 srq = (struct spdk_rdma_srq *)opts->qp; 777 } else { 778 qp = (struct spdk_rdma_qp *)opts->qp; 779 } 780 781 for (i = 0; i < opts->max_queue_depth; i++) { 782 rdma_recv = &resources->recvs[i]; 783 rdma_recv->qpair = opts->qpair; 784 785 /* Set up memory to receive commands */ 786 if (resources->bufs) { 787 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 788 opts->in_capsule_data_size)); 789 } 790 791 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 792 793 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 794 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 795 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 796 rdma_recv->wr.num_sge = 1; 797 798 if (rdma_recv->buf && resources->bufs_mr) { 799 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 800 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 801 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 802 rdma_recv->wr.num_sge++; 803 } 804 805 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 806 rdma_recv->wr.sg_list = rdma_recv->sgl; 807 if (srq) { 808 spdk_rdma_srq_queue_recv_wrs(srq, &rdma_recv->wr); 809 } else { 810 spdk_rdma_qp_queue_recv_wrs(qp, &rdma_recv->wr); 811 } 812 } 813 814 for (i = 0; i < opts->max_queue_depth; i++) { 815 rdma_req = &resources->reqs[i]; 816 817 if (opts->qpair != NULL) { 818 rdma_req->req.qpair = &opts->qpair->qpair; 819 } else { 820 rdma_req->req.qpair = NULL; 821 } 822 rdma_req->req.cmd = NULL; 823 824 /* Set up memory to send responses */ 825 rdma_req->req.rsp = &resources->cpls[i]; 826 827 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 828 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 829 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 830 831 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 832 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 833 rdma_req->rsp.wr.next = NULL; 834 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 835 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 836 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 837 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 838 839 /* Set up memory for data buffers */ 840 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 841 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 842 rdma_req->data.wr.next = NULL; 843 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 844 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 845 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 846 847 /* Initialize request state to FREE */ 848 rdma_req->state = RDMA_REQUEST_STATE_FREE; 849 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 850 } 851 852 if (srq) { 853 rc = spdk_rdma_srq_flush_recv_wrs(srq, &bad_wr); 854 } else { 855 rc = spdk_rdma_qp_flush_recv_wrs(qp, &bad_wr); 856 } 857 858 if (rc) { 859 goto cleanup; 860 } 861 862 return resources; 863 864 cleanup: 865 nvmf_rdma_resources_destroy(resources); 866 return NULL; 867 } 868 869 static void 870 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 871 { 872 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 873 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 874 ctx->rqpair = NULL; 875 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 876 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 877 } 878 } 879 880 static void 881 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 882 { 883 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 884 struct ibv_recv_wr *bad_recv_wr = NULL; 885 int rc; 886 887 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair, 0); 888 889 if (rqpair->qd != 0) { 890 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 891 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 892 struct spdk_nvmf_rdma_transport, transport); 893 struct spdk_nvmf_rdma_request *req; 894 uint32_t i, max_req_count = 0; 895 896 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 897 898 if (rqpair->srq == NULL) { 899 nvmf_rdma_dump_qpair_contents(rqpair); 900 max_req_count = rqpair->max_queue_depth; 901 } else if (rqpair->poller && rqpair->resources) { 902 max_req_count = rqpair->poller->max_srq_depth; 903 } 904 905 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 906 for (i = 0; i < max_req_count; i++) { 907 req = &rqpair->resources->reqs[i]; 908 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 909 /* nvmf_rdma_request_process checks qpair ibv and internal state 910 * and completes a request */ 911 nvmf_rdma_request_process(rtransport, req); 912 } 913 } 914 assert(rqpair->qd == 0); 915 } 916 917 if (rqpair->poller) { 918 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 919 920 if (rqpair->srq != NULL && rqpair->resources != NULL) { 921 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 922 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 923 if (rqpair == rdma_recv->qpair) { 924 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 925 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr); 926 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 927 if (rc) { 928 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 929 } 930 } 931 } 932 } 933 } 934 935 if (rqpair->cm_id) { 936 if (rqpair->rdma_qp != NULL) { 937 spdk_rdma_qp_destroy(rqpair->rdma_qp); 938 rqpair->rdma_qp = NULL; 939 } 940 rdma_destroy_id(rqpair->cm_id); 941 942 if (rqpair->poller != NULL && rqpair->srq == NULL) { 943 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 944 } 945 } 946 947 if (rqpair->srq == NULL && rqpair->resources != NULL) { 948 nvmf_rdma_resources_destroy(rqpair->resources); 949 } 950 951 nvmf_rdma_qpair_clean_ibv_events(rqpair); 952 953 if (rqpair->destruct_channel) { 954 spdk_put_io_channel(rqpair->destruct_channel); 955 rqpair->destruct_channel = NULL; 956 } 957 958 free(rqpair); 959 } 960 961 static int 962 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 963 { 964 struct spdk_nvmf_rdma_poller *rpoller; 965 int rc, num_cqe, required_num_wr; 966 967 /* Enlarge CQ size dynamically */ 968 rpoller = rqpair->poller; 969 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 970 num_cqe = rpoller->num_cqe; 971 if (num_cqe < required_num_wr) { 972 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 973 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 974 } 975 976 if (rpoller->num_cqe != num_cqe) { 977 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) { 978 SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n" 979 "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe); 980 return -1; 981 } 982 if (required_num_wr > device->attr.max_cqe) { 983 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 984 required_num_wr, device->attr.max_cqe); 985 return -1; 986 } 987 988 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 989 rc = ibv_resize_cq(rpoller->cq, num_cqe); 990 if (rc) { 991 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 992 return -1; 993 } 994 995 rpoller->num_cqe = num_cqe; 996 } 997 998 rpoller->required_num_wr = required_num_wr; 999 return 0; 1000 } 1001 1002 static int 1003 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 1004 { 1005 struct spdk_nvmf_rdma_qpair *rqpair; 1006 struct spdk_nvmf_rdma_transport *rtransport; 1007 struct spdk_nvmf_transport *transport; 1008 struct spdk_nvmf_rdma_resource_opts opts; 1009 struct spdk_nvmf_rdma_device *device; 1010 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 1011 1012 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1013 device = rqpair->device; 1014 1015 qp_init_attr.qp_context = rqpair; 1016 qp_init_attr.pd = device->pd; 1017 qp_init_attr.send_cq = rqpair->poller->cq; 1018 qp_init_attr.recv_cq = rqpair->poller->cq; 1019 1020 if (rqpair->srq) { 1021 qp_init_attr.srq = rqpair->srq->srq; 1022 } else { 1023 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 1024 } 1025 1026 /* SEND, READ, and WRITE operations */ 1027 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 1028 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 1029 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1030 qp_init_attr.stats = &rqpair->poller->stat.qp_stats; 1031 1032 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1033 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1034 goto error; 1035 } 1036 1037 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 1038 if (!rqpair->rdma_qp) { 1039 goto error; 1040 } 1041 1042 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1043 qp_init_attr.cap.max_send_wr); 1044 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1045 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1046 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair, 0); 1047 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1048 1049 if (rqpair->poller->srq == NULL) { 1050 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1051 transport = &rtransport->transport; 1052 1053 opts.qp = rqpair->rdma_qp; 1054 opts.pd = rqpair->cm_id->pd; 1055 opts.qpair = rqpair; 1056 opts.shared = false; 1057 opts.max_queue_depth = rqpair->max_queue_depth; 1058 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1059 1060 rqpair->resources = nvmf_rdma_resources_create(&opts); 1061 1062 if (!rqpair->resources) { 1063 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1064 rdma_destroy_qp(rqpair->cm_id); 1065 goto error; 1066 } 1067 } else { 1068 rqpair->resources = rqpair->poller->resources; 1069 } 1070 1071 rqpair->current_recv_depth = 0; 1072 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1073 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1074 1075 return 0; 1076 1077 error: 1078 rdma_destroy_id(rqpair->cm_id); 1079 rqpair->cm_id = NULL; 1080 return -1; 1081 } 1082 1083 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1084 /* This function accepts either a single wr or the first wr in a linked list. */ 1085 static void 1086 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1087 { 1088 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1089 struct spdk_nvmf_rdma_transport, transport); 1090 1091 if (rqpair->srq != NULL) { 1092 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, first); 1093 } else { 1094 if (spdk_rdma_qp_queue_recv_wrs(rqpair->rdma_qp, first)) { 1095 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1096 } 1097 } 1098 1099 if (rtransport->rdma_opts.no_wr_batching) { 1100 _poller_submit_recvs(rtransport, rqpair->poller); 1101 } 1102 } 1103 1104 static int 1105 request_transfer_in(struct spdk_nvmf_request *req) 1106 { 1107 struct spdk_nvmf_rdma_request *rdma_req; 1108 struct spdk_nvmf_qpair *qpair; 1109 struct spdk_nvmf_rdma_qpair *rqpair; 1110 struct spdk_nvmf_rdma_transport *rtransport; 1111 1112 qpair = req->qpair; 1113 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1114 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1115 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1116 struct spdk_nvmf_rdma_transport, transport); 1117 1118 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1119 assert(rdma_req != NULL); 1120 1121 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1122 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1123 } 1124 if (rtransport->rdma_opts.no_wr_batching) { 1125 _poller_submit_sends(rtransport, rqpair->poller); 1126 } 1127 1128 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1129 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1130 return 0; 1131 } 1132 1133 static int 1134 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1135 { 1136 int num_outstanding_data_wr = 0; 1137 struct spdk_nvmf_rdma_request *rdma_req; 1138 struct spdk_nvmf_qpair *qpair; 1139 struct spdk_nvmf_rdma_qpair *rqpair; 1140 struct spdk_nvme_cpl *rsp; 1141 struct ibv_send_wr *first = NULL; 1142 struct spdk_nvmf_rdma_transport *rtransport; 1143 1144 *data_posted = 0; 1145 qpair = req->qpair; 1146 rsp = &req->rsp->nvme_cpl; 1147 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1148 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1149 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1150 struct spdk_nvmf_rdma_transport, transport); 1151 1152 /* Advance our sq_head pointer */ 1153 if (qpair->sq_head == qpair->sq_head_max) { 1154 qpair->sq_head = 0; 1155 } else { 1156 qpair->sq_head++; 1157 } 1158 rsp->sqhd = qpair->sq_head; 1159 1160 /* queue the capsule for the recv buffer */ 1161 assert(rdma_req->recv != NULL); 1162 1163 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1164 1165 rdma_req->recv = NULL; 1166 assert(rqpair->current_recv_depth > 0); 1167 rqpair->current_recv_depth--; 1168 1169 /* Build the response which consists of optional 1170 * RDMA WRITEs to transfer data, plus an RDMA SEND 1171 * containing the response. 1172 */ 1173 first = &rdma_req->rsp.wr; 1174 1175 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1176 /* On failure, data was not read from the controller. So clear the 1177 * number of outstanding data WRs to zero. 1178 */ 1179 rdma_req->num_outstanding_data_wr = 0; 1180 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1181 first = &rdma_req->data.wr; 1182 *data_posted = 1; 1183 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1184 } 1185 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1186 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1187 } 1188 if (rtransport->rdma_opts.no_wr_batching) { 1189 _poller_submit_sends(rtransport, rqpair->poller); 1190 } 1191 1192 /* +1 for the rsp wr */ 1193 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1194 1195 return 0; 1196 } 1197 1198 static int 1199 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1200 { 1201 struct spdk_nvmf_rdma_accept_private_data accept_data; 1202 struct rdma_conn_param ctrlr_event_data = {}; 1203 int rc; 1204 1205 accept_data.recfmt = 0; 1206 accept_data.crqsize = rqpair->max_queue_depth; 1207 1208 ctrlr_event_data.private_data = &accept_data; 1209 ctrlr_event_data.private_data_len = sizeof(accept_data); 1210 if (id->ps == RDMA_PS_TCP) { 1211 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1212 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1213 } 1214 1215 /* Configure infinite retries for the initiator side qpair. 1216 * When using a shared receive queue on the target side, 1217 * we need to pass this value to the initiator to prevent the 1218 * initiator side NIC from completing SEND requests back to the 1219 * initiator with status rnr_retry_count_exceeded. */ 1220 if (rqpair->srq != NULL) { 1221 ctrlr_event_data.rnr_retry_count = 0x7; 1222 } 1223 1224 /* When qpair is created without use of rdma cm API, an additional 1225 * information must be provided to initiator in the connection response: 1226 * whether qpair is using SRQ and its qp_num 1227 * Fields below are ignored by rdma cm if qpair has been 1228 * created using rdma cm API. */ 1229 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1230 ctrlr_event_data.qp_num = rqpair->rdma_qp->qp->qp_num; 1231 1232 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1233 if (rc) { 1234 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1235 } else { 1236 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1237 } 1238 1239 return rc; 1240 } 1241 1242 static void 1243 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1244 { 1245 struct spdk_nvmf_rdma_reject_private_data rej_data; 1246 1247 rej_data.recfmt = 0; 1248 rej_data.sts = error; 1249 1250 rdma_reject(id, &rej_data, sizeof(rej_data)); 1251 } 1252 1253 static int 1254 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1255 { 1256 struct spdk_nvmf_rdma_transport *rtransport; 1257 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1258 struct spdk_nvmf_rdma_port *port; 1259 struct rdma_conn_param *rdma_param = NULL; 1260 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1261 uint16_t max_queue_depth; 1262 uint16_t max_read_depth; 1263 1264 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1265 1266 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1267 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1268 1269 rdma_param = &event->param.conn; 1270 if (rdma_param->private_data == NULL || 1271 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1272 SPDK_ERRLOG("connect request: no private data provided\n"); 1273 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1274 return -1; 1275 } 1276 1277 private_data = rdma_param->private_data; 1278 if (private_data->recfmt != 0) { 1279 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1280 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1281 return -1; 1282 } 1283 1284 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1285 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1286 1287 port = event->listen_id->context; 1288 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1289 event->listen_id, event->listen_id->verbs, port); 1290 1291 /* Figure out the supported queue depth. This is a multi-step process 1292 * that takes into account hardware maximums, host provided values, 1293 * and our target's internal memory limits */ 1294 1295 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1296 1297 /* Start with the maximum queue depth allowed by the target */ 1298 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1299 max_read_depth = rtransport->transport.opts.max_queue_depth; 1300 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1301 rtransport->transport.opts.max_queue_depth); 1302 1303 /* Next check the local NIC's hardware limitations */ 1304 SPDK_DEBUGLOG(rdma, 1305 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1306 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1307 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1308 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1309 1310 /* Next check the remote NIC's hardware limitations */ 1311 SPDK_DEBUGLOG(rdma, 1312 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1313 rdma_param->initiator_depth, rdma_param->responder_resources); 1314 if (rdma_param->initiator_depth > 0) { 1315 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1316 } 1317 1318 /* Finally check for the host software requested values, which are 1319 * optional. */ 1320 if (rdma_param->private_data != NULL && 1321 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1322 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1323 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1324 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1325 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1326 } 1327 1328 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1329 max_queue_depth, max_read_depth); 1330 1331 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1332 if (rqpair == NULL) { 1333 SPDK_ERRLOG("Could not allocate new connection.\n"); 1334 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1335 return -1; 1336 } 1337 1338 rqpair->device = port->device; 1339 rqpair->max_queue_depth = max_queue_depth; 1340 rqpair->max_read_depth = max_read_depth; 1341 rqpair->cm_id = event->id; 1342 rqpair->listen_id = event->listen_id; 1343 rqpair->qpair.transport = transport; 1344 STAILQ_INIT(&rqpair->ibv_events); 1345 /* use qid from the private data to determine the qpair type 1346 qid will be set to the appropriate value when the controller is created */ 1347 rqpair->qpair.qid = private_data->qid; 1348 1349 event->id->context = &rqpair->qpair; 1350 1351 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1352 1353 return 0; 1354 } 1355 1356 static inline void 1357 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1358 enum spdk_nvme_data_transfer xfer) 1359 { 1360 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1361 wr->opcode = IBV_WR_RDMA_WRITE; 1362 wr->send_flags = 0; 1363 wr->next = next; 1364 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1365 wr->opcode = IBV_WR_RDMA_READ; 1366 wr->send_flags = IBV_SEND_SIGNALED; 1367 wr->next = NULL; 1368 } else { 1369 assert(0); 1370 } 1371 } 1372 1373 static int 1374 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1375 struct spdk_nvmf_rdma_request *rdma_req, 1376 uint32_t num_sgl_descriptors) 1377 { 1378 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1379 struct spdk_nvmf_rdma_request_data *current_data_wr; 1380 uint32_t i; 1381 1382 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1383 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1384 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1385 return -EINVAL; 1386 } 1387 1388 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1389 return -ENOMEM; 1390 } 1391 1392 current_data_wr = &rdma_req->data; 1393 1394 for (i = 0; i < num_sgl_descriptors; i++) { 1395 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1396 current_data_wr->wr.next = &work_requests[i]->wr; 1397 current_data_wr = work_requests[i]; 1398 current_data_wr->wr.sg_list = current_data_wr->sgl; 1399 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1400 } 1401 1402 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1403 1404 return 0; 1405 } 1406 1407 static inline void 1408 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1409 { 1410 struct ibv_send_wr *wr = &rdma_req->data.wr; 1411 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1412 1413 wr->wr.rdma.rkey = sgl->keyed.key; 1414 wr->wr.rdma.remote_addr = sgl->address; 1415 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1416 } 1417 1418 static inline void 1419 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1420 { 1421 struct ibv_send_wr *wr = &rdma_req->data.wr; 1422 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1423 uint32_t i; 1424 int j; 1425 uint64_t remote_addr_offset = 0; 1426 1427 for (i = 0; i < num_wrs; ++i) { 1428 wr->wr.rdma.rkey = sgl->keyed.key; 1429 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1430 for (j = 0; j < wr->num_sge; ++j) { 1431 remote_addr_offset += wr->sg_list[j].length; 1432 } 1433 wr = wr->next; 1434 } 1435 } 1436 1437 static int 1438 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1439 struct spdk_nvmf_rdma_device *device, 1440 struct spdk_nvmf_rdma_request *rdma_req, 1441 struct ibv_send_wr *wr, 1442 uint32_t total_length, 1443 uint32_t num_extra_wrs) 1444 { 1445 struct spdk_rdma_memory_translation mem_translation; 1446 struct spdk_dif_ctx *dif_ctx = NULL; 1447 struct ibv_sge *sg_ele; 1448 struct iovec *iov; 1449 uint32_t remaining_data_block = 0; 1450 uint32_t lkey, remaining; 1451 int rc; 1452 1453 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 1454 dif_ctx = &rdma_req->req.dif.dif_ctx; 1455 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1456 } 1457 1458 wr->num_sge = 0; 1459 1460 while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1461 iov = &rdma_req->req.iov[rdma_req->iovpos]; 1462 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1463 if (spdk_unlikely(rc)) { 1464 return false; 1465 } 1466 1467 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1468 sg_ele = &wr->sg_list[wr->num_sge]; 1469 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length); 1470 1471 if (spdk_likely(!dif_ctx)) { 1472 sg_ele->lkey = lkey; 1473 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset; 1474 sg_ele->length = remaining; 1475 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1476 sg_ele->length); 1477 rdma_req->offset += sg_ele->length; 1478 total_length -= sg_ele->length; 1479 wr->num_sge++; 1480 1481 if (rdma_req->offset == iov->iov_len) { 1482 rdma_req->offset = 0; 1483 rdma_req->iovpos++; 1484 } 1485 } else { 1486 uint32_t data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1487 uint32_t md_size = dif_ctx->md_size; 1488 uint32_t sge_len; 1489 1490 while (remaining) { 1491 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1492 if (num_extra_wrs > 0 && wr->next) { 1493 wr = wr->next; 1494 wr->num_sge = 0; 1495 sg_ele = &wr->sg_list[wr->num_sge]; 1496 num_extra_wrs--; 1497 } else { 1498 break; 1499 } 1500 } 1501 sg_ele->lkey = lkey; 1502 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset); 1503 sge_len = spdk_min(remaining, remaining_data_block); 1504 sg_ele->length = sge_len; 1505 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr, 1506 sg_ele->length); 1507 remaining -= sge_len; 1508 remaining_data_block -= sge_len; 1509 rdma_req->offset += sge_len; 1510 total_length -= sge_len; 1511 1512 sg_ele++; 1513 wr->num_sge++; 1514 1515 if (remaining_data_block == 0) { 1516 /* skip metadata */ 1517 rdma_req->offset += md_size; 1518 total_length -= md_size; 1519 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1520 remaining -= spdk_min(remaining, md_size); 1521 remaining_data_block = data_block_size; 1522 } 1523 1524 if (remaining == 0) { 1525 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1526 the remaining metadata bits at the beginning of the next buffer */ 1527 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset); 1528 rdma_req->iovpos++; 1529 } 1530 } 1531 } 1532 } 1533 1534 if (total_length) { 1535 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1536 return -EINVAL; 1537 } 1538 1539 return 0; 1540 } 1541 1542 static inline uint32_t 1543 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1544 { 1545 /* estimate the number of SG entries and WRs needed to process the request */ 1546 uint32_t num_sge = 0; 1547 uint32_t i; 1548 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1549 1550 for (i = 0; i < num_buffers && length > 0; i++) { 1551 uint32_t buffer_len = spdk_min(length, io_unit_size); 1552 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1553 1554 if (num_sge_in_block * block_size > buffer_len) { 1555 ++num_sge_in_block; 1556 } 1557 num_sge += num_sge_in_block; 1558 length -= buffer_len; 1559 } 1560 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1561 } 1562 1563 static int 1564 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1565 struct spdk_nvmf_rdma_device *device, 1566 struct spdk_nvmf_rdma_request *rdma_req, 1567 uint32_t length) 1568 { 1569 struct spdk_nvmf_rdma_qpair *rqpair; 1570 struct spdk_nvmf_rdma_poll_group *rgroup; 1571 struct spdk_nvmf_request *req = &rdma_req->req; 1572 struct ibv_send_wr *wr = &rdma_req->data.wr; 1573 int rc; 1574 uint32_t num_wrs = 1; 1575 1576 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1577 rgroup = rqpair->poller->group; 1578 1579 /* rdma wr specifics */ 1580 nvmf_rdma_setup_request(rdma_req); 1581 1582 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1583 length); 1584 if (rc != 0) { 1585 return rc; 1586 } 1587 1588 assert(req->iovcnt <= rqpair->max_send_sge); 1589 1590 rdma_req->iovpos = 0; 1591 1592 if (spdk_unlikely(req->dif_enabled)) { 1593 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1594 req->dif.dif_ctx.block_size); 1595 if (num_wrs > 1) { 1596 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1597 if (rc != 0) { 1598 goto err_exit; 1599 } 1600 } 1601 } 1602 1603 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1604 if (spdk_unlikely(rc != 0)) { 1605 goto err_exit; 1606 } 1607 1608 if (spdk_unlikely(num_wrs > 1)) { 1609 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1610 } 1611 1612 /* set the number of outstanding data WRs for this request. */ 1613 rdma_req->num_outstanding_data_wr = num_wrs; 1614 1615 return rc; 1616 1617 err_exit: 1618 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1619 nvmf_rdma_request_free_data(rdma_req, rtransport); 1620 req->iovcnt = 0; 1621 return rc; 1622 } 1623 1624 static int 1625 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1626 struct spdk_nvmf_rdma_device *device, 1627 struct spdk_nvmf_rdma_request *rdma_req) 1628 { 1629 struct spdk_nvmf_rdma_qpair *rqpair; 1630 struct spdk_nvmf_rdma_poll_group *rgroup; 1631 struct ibv_send_wr *current_wr; 1632 struct spdk_nvmf_request *req = &rdma_req->req; 1633 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1634 uint32_t num_sgl_descriptors; 1635 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0; 1636 uint32_t i; 1637 int rc; 1638 1639 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1640 rgroup = rqpair->poller->group; 1641 1642 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1643 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1644 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1645 1646 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1647 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1648 1649 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1650 for (i = 0; i < num_sgl_descriptors; i++) { 1651 if (spdk_likely(!req->dif_enabled)) { 1652 lengths[i] = desc->keyed.length; 1653 } else { 1654 req->dif.orig_length += desc->keyed.length; 1655 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1656 req->dif.elba_length += lengths[i]; 1657 } 1658 total_length += lengths[i]; 1659 desc++; 1660 } 1661 1662 if (total_length > rtransport->transport.opts.max_io_size) { 1663 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n", 1664 total_length, rtransport->transport.opts.max_io_size); 1665 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1666 return -EINVAL; 1667 } 1668 1669 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1670 return -ENOMEM; 1671 } 1672 1673 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length); 1674 if (rc != 0) { 1675 nvmf_rdma_request_free_data(rdma_req, rtransport); 1676 return rc; 1677 } 1678 1679 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1680 current_wr = &rdma_req->data.wr; 1681 assert(current_wr != NULL); 1682 1683 req->length = 0; 1684 rdma_req->iovpos = 0; 1685 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1686 for (i = 0; i < num_sgl_descriptors; i++) { 1687 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1688 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1689 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1690 rc = -EINVAL; 1691 goto err_exit; 1692 } 1693 1694 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1695 if (rc != 0) { 1696 rc = -ENOMEM; 1697 goto err_exit; 1698 } 1699 1700 req->length += desc->keyed.length; 1701 current_wr->wr.rdma.rkey = desc->keyed.key; 1702 current_wr->wr.rdma.remote_addr = desc->address; 1703 current_wr = current_wr->next; 1704 desc++; 1705 } 1706 1707 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1708 /* Go back to the last descriptor in the list. */ 1709 desc--; 1710 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1711 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1712 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1713 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1714 } 1715 } 1716 #endif 1717 1718 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1719 1720 return 0; 1721 1722 err_exit: 1723 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1724 nvmf_rdma_request_free_data(rdma_req, rtransport); 1725 return rc; 1726 } 1727 1728 static int 1729 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1730 struct spdk_nvmf_rdma_device *device, 1731 struct spdk_nvmf_rdma_request *rdma_req) 1732 { 1733 struct spdk_nvmf_request *req = &rdma_req->req; 1734 struct spdk_nvme_cpl *rsp; 1735 struct spdk_nvme_sgl_descriptor *sgl; 1736 int rc; 1737 uint32_t length; 1738 1739 rsp = &req->rsp->nvme_cpl; 1740 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1741 1742 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1743 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1744 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1745 1746 length = sgl->keyed.length; 1747 if (length > rtransport->transport.opts.max_io_size) { 1748 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1749 length, rtransport->transport.opts.max_io_size); 1750 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1751 return -1; 1752 } 1753 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1754 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1755 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1756 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1757 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1758 } 1759 } 1760 #endif 1761 1762 /* fill request length and populate iovs */ 1763 req->length = length; 1764 1765 if (spdk_unlikely(req->dif_enabled)) { 1766 req->dif.orig_length = length; 1767 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1768 req->dif.elba_length = length; 1769 } 1770 1771 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1772 if (spdk_unlikely(rc < 0)) { 1773 if (rc == -EINVAL) { 1774 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1775 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1776 return -1; 1777 } 1778 /* No available buffers. Queue this request up. */ 1779 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1780 return 0; 1781 } 1782 1783 /* backward compatible */ 1784 req->data = req->iov[0].iov_base; 1785 1786 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1787 req->iovcnt); 1788 1789 return 0; 1790 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1791 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1792 uint64_t offset = sgl->address; 1793 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1794 1795 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1796 offset, sgl->unkeyed.length); 1797 1798 if (offset > max_len) { 1799 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1800 offset, max_len); 1801 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1802 return -1; 1803 } 1804 max_len -= (uint32_t)offset; 1805 1806 if (sgl->unkeyed.length > max_len) { 1807 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1808 sgl->unkeyed.length, max_len); 1809 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1810 return -1; 1811 } 1812 1813 rdma_req->num_outstanding_data_wr = 0; 1814 req->data = rdma_req->recv->buf + offset; 1815 req->data_from_pool = false; 1816 req->length = sgl->unkeyed.length; 1817 1818 req->iov[0].iov_base = req->data; 1819 req->iov[0].iov_len = req->length; 1820 req->iovcnt = 1; 1821 1822 return 0; 1823 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1824 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1825 1826 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1827 if (rc == -ENOMEM) { 1828 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1829 return 0; 1830 } else if (rc == -EINVAL) { 1831 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1832 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1833 return -1; 1834 } 1835 1836 /* backward compatible */ 1837 req->data = req->iov[0].iov_base; 1838 1839 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1840 req->iovcnt); 1841 1842 return 0; 1843 } 1844 1845 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1846 sgl->generic.type, sgl->generic.subtype); 1847 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1848 return -1; 1849 } 1850 1851 static void 1852 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1853 struct spdk_nvmf_rdma_transport *rtransport) 1854 { 1855 struct spdk_nvmf_rdma_qpair *rqpair; 1856 struct spdk_nvmf_rdma_poll_group *rgroup; 1857 1858 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1859 if (rdma_req->req.data_from_pool) { 1860 rgroup = rqpair->poller->group; 1861 1862 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1863 } 1864 nvmf_rdma_request_free_data(rdma_req, rtransport); 1865 rdma_req->req.length = 0; 1866 rdma_req->req.iovcnt = 0; 1867 rdma_req->req.data = NULL; 1868 rdma_req->rsp.wr.next = NULL; 1869 rdma_req->data.wr.next = NULL; 1870 rdma_req->offset = 0; 1871 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1872 rqpair->qd--; 1873 1874 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1875 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1876 } 1877 1878 bool 1879 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1880 struct spdk_nvmf_rdma_request *rdma_req) 1881 { 1882 struct spdk_nvmf_rdma_qpair *rqpair; 1883 struct spdk_nvmf_rdma_device *device; 1884 struct spdk_nvmf_rdma_poll_group *rgroup; 1885 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1886 int rc; 1887 struct spdk_nvmf_rdma_recv *rdma_recv; 1888 enum spdk_nvmf_rdma_request_state prev_state; 1889 bool progress = false; 1890 int data_posted; 1891 uint32_t num_blocks; 1892 1893 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1894 device = rqpair->device; 1895 rgroup = rqpair->poller->group; 1896 1897 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1898 1899 /* If the queue pair is in an error state, force the request to the completed state 1900 * to release resources. */ 1901 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1902 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1903 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1904 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1905 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1906 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1907 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1908 } 1909 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1910 } 1911 1912 /* The loop here is to allow for several back-to-back state changes. */ 1913 do { 1914 prev_state = rdma_req->state; 1915 1916 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 1917 1918 switch (rdma_req->state) { 1919 case RDMA_REQUEST_STATE_FREE: 1920 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1921 * to escape this state. */ 1922 break; 1923 case RDMA_REQUEST_STATE_NEW: 1924 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1925 (uintptr_t)rdma_req, (uintptr_t)rqpair); 1926 rdma_recv = rdma_req->recv; 1927 1928 /* The first element of the SGL is the NVMe command */ 1929 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1930 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1931 1932 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1933 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1934 break; 1935 } 1936 1937 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 1938 rdma_req->req.dif_enabled = true; 1939 } 1940 1941 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1942 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1943 rdma_req->rsp.wr.imm_data = 0; 1944 #endif 1945 1946 /* The next state transition depends on the data transfer needs of this request. */ 1947 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 1948 1949 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 1950 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 1951 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 1952 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1953 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 1954 break; 1955 } 1956 1957 /* If no data to transfer, ready to execute. */ 1958 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1959 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1960 break; 1961 } 1962 1963 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1964 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 1965 break; 1966 case RDMA_REQUEST_STATE_NEED_BUFFER: 1967 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1968 (uintptr_t)rdma_req, (uintptr_t)rqpair); 1969 1970 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1971 1972 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 1973 /* This request needs to wait in line to obtain a buffer */ 1974 break; 1975 } 1976 1977 /* Try to get a data buffer */ 1978 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1979 if (rc < 0) { 1980 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 1981 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1982 break; 1983 } 1984 1985 if (!rdma_req->req.data) { 1986 /* No buffers available. */ 1987 rgroup->stat.pending_data_buffer++; 1988 break; 1989 } 1990 1991 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 1992 1993 /* If data is transferring from host to controller and the data didn't 1994 * arrive using in capsule data, we need to do a transfer from the host. 1995 */ 1996 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 1997 rdma_req->req.data_from_pool) { 1998 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1999 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2000 break; 2001 } 2002 2003 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2004 break; 2005 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2006 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2007 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2008 2009 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2010 /* This request needs to wait in line to perform RDMA */ 2011 break; 2012 } 2013 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2014 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2015 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2016 rqpair->poller->stat.pending_rdma_read++; 2017 break; 2018 } 2019 2020 /* We have already verified that this request is the head of the queue. */ 2021 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2022 2023 rc = request_transfer_in(&rdma_req->req); 2024 if (!rc) { 2025 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2026 } else { 2027 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2028 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2029 } 2030 break; 2031 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2032 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2033 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2034 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2035 * to escape this state. */ 2036 break; 2037 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2038 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2039 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2040 2041 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2042 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2043 /* generate DIF for write operation */ 2044 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2045 assert(num_blocks > 0); 2046 2047 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2048 num_blocks, &rdma_req->req.dif.dif_ctx); 2049 if (rc != 0) { 2050 SPDK_ERRLOG("DIF generation failed\n"); 2051 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2052 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2053 break; 2054 } 2055 } 2056 2057 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2058 /* set extended length before IO operation */ 2059 rdma_req->req.length = rdma_req->req.dif.elba_length; 2060 } 2061 2062 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2063 spdk_nvmf_request_exec(&rdma_req->req); 2064 break; 2065 case RDMA_REQUEST_STATE_EXECUTING: 2066 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2067 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2068 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2069 * to escape this state. */ 2070 break; 2071 case RDMA_REQUEST_STATE_EXECUTED: 2072 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2073 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2074 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2075 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2076 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2077 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2078 } else { 2079 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2080 } 2081 if (spdk_unlikely(rdma_req->req.dif_enabled)) { 2082 /* restore the original length */ 2083 rdma_req->req.length = rdma_req->req.dif.orig_length; 2084 2085 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2086 struct spdk_dif_error error_blk; 2087 2088 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2089 2090 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2091 &rdma_req->req.dif.dif_ctx, &error_blk); 2092 if (rc) { 2093 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2094 2095 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2096 error_blk.err_offset); 2097 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2098 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2099 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2100 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2101 } 2102 } 2103 } 2104 break; 2105 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2106 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2107 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2108 2109 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2110 /* This request needs to wait in line to perform RDMA */ 2111 break; 2112 } 2113 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2114 rqpair->max_send_depth) { 2115 /* We can only have so many WRs outstanding. we have to wait until some finish. 2116 * +1 since each request has an additional wr in the resp. */ 2117 rqpair->poller->stat.pending_rdma_write++; 2118 break; 2119 } 2120 2121 /* We have already verified that this request is the head of the queue. */ 2122 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2123 2124 /* The data transfer will be kicked off from 2125 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2126 */ 2127 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2128 break; 2129 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2130 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2131 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2132 rc = request_transfer_out(&rdma_req->req, &data_posted); 2133 assert(rc == 0); /* No good way to handle this currently */ 2134 if (rc) { 2135 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2136 } else { 2137 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2138 RDMA_REQUEST_STATE_COMPLETING; 2139 } 2140 break; 2141 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2142 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2143 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2144 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2145 * to escape this state. */ 2146 break; 2147 case RDMA_REQUEST_STATE_COMPLETING: 2148 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2149 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2150 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2151 * to escape this state. */ 2152 break; 2153 case RDMA_REQUEST_STATE_COMPLETED: 2154 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2155 (uintptr_t)rdma_req, (uintptr_t)rqpair); 2156 2157 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2158 _nvmf_rdma_request_free(rdma_req, rtransport); 2159 break; 2160 case RDMA_REQUEST_NUM_STATES: 2161 default: 2162 assert(0); 2163 break; 2164 } 2165 2166 if (rdma_req->state != prev_state) { 2167 progress = true; 2168 } 2169 } while (rdma_req->state != prev_state); 2170 2171 return progress; 2172 } 2173 2174 /* Public API callbacks begin here */ 2175 2176 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2177 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2178 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2179 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2180 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2181 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2182 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2183 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2184 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2185 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2186 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2187 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2188 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2189 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2190 2191 static void 2192 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2193 { 2194 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2195 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2196 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2197 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2198 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2199 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2200 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2201 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2202 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2203 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2204 opts->transport_specific = NULL; 2205 } 2206 2207 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2208 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2209 2210 static inline bool 2211 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2212 { 2213 return device->attr.vendor_id == NVMF_RXE_VENDOR_ID_OLD || 2214 device->attr.vendor_id == NVMF_RXE_VENDOR_ID_NEW; 2215 } 2216 2217 static struct spdk_nvmf_transport * 2218 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2219 { 2220 int rc; 2221 struct spdk_nvmf_rdma_transport *rtransport; 2222 struct spdk_nvmf_rdma_device *device, *tmp; 2223 struct ibv_context **contexts; 2224 uint32_t i; 2225 int flag; 2226 uint32_t sge_count; 2227 uint32_t min_shared_buffers; 2228 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2229 pthread_mutexattr_t attr; 2230 2231 rtransport = calloc(1, sizeof(*rtransport)); 2232 if (!rtransport) { 2233 return NULL; 2234 } 2235 2236 if (pthread_mutexattr_init(&attr)) { 2237 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2238 free(rtransport); 2239 return NULL; 2240 } 2241 2242 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2243 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2244 pthread_mutexattr_destroy(&attr); 2245 free(rtransport); 2246 return NULL; 2247 } 2248 2249 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2250 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2251 pthread_mutexattr_destroy(&attr); 2252 free(rtransport); 2253 return NULL; 2254 } 2255 2256 pthread_mutexattr_destroy(&attr); 2257 2258 TAILQ_INIT(&rtransport->devices); 2259 TAILQ_INIT(&rtransport->ports); 2260 TAILQ_INIT(&rtransport->poll_groups); 2261 2262 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2263 rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2264 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2265 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2266 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2267 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2268 if (opts->transport_specific != NULL && 2269 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2270 SPDK_COUNTOF(rdma_transport_opts_decoder), 2271 &rtransport->rdma_opts)) { 2272 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2273 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2274 return NULL; 2275 } 2276 2277 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2278 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2279 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2280 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2281 " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d," 2282 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2283 opts->max_queue_depth, 2284 opts->max_io_size, 2285 opts->max_qpairs_per_ctrlr - 1, 2286 opts->io_unit_size, 2287 opts->in_capsule_data_size, 2288 opts->max_aq_depth, 2289 opts->num_shared_buffers, 2290 rtransport->rdma_opts.num_cqe, 2291 rtransport->rdma_opts.max_srq_depth, 2292 rtransport->rdma_opts.no_srq, 2293 rtransport->rdma_opts.acceptor_backlog, 2294 rtransport->rdma_opts.no_wr_batching, 2295 opts->abort_timeout_sec); 2296 2297 /* I/O unit size cannot be larger than max I/O size */ 2298 if (opts->io_unit_size > opts->max_io_size) { 2299 opts->io_unit_size = opts->max_io_size; 2300 } 2301 2302 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2303 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2304 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2305 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2306 } 2307 2308 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2309 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2310 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2311 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2312 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2313 return NULL; 2314 } 2315 2316 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2317 if (min_shared_buffers > opts->num_shared_buffers) { 2318 SPDK_ERRLOG("There are not enough buffers to satisfy" 2319 "per-poll group caches for each thread. (%" PRIu32 ")" 2320 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2321 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2322 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2323 return NULL; 2324 } 2325 2326 sge_count = opts->max_io_size / opts->io_unit_size; 2327 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2328 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2329 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2330 return NULL; 2331 } 2332 2333 rtransport->event_channel = rdma_create_event_channel(); 2334 if (rtransport->event_channel == NULL) { 2335 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2336 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2337 return NULL; 2338 } 2339 2340 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2341 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2342 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2343 rtransport->event_channel->fd, spdk_strerror(errno)); 2344 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2345 return NULL; 2346 } 2347 2348 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2349 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2350 sizeof(struct spdk_nvmf_rdma_request_data), 2351 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2352 SPDK_ENV_SOCKET_ID_ANY); 2353 if (!rtransport->data_wr_pool) { 2354 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2355 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2356 return NULL; 2357 } 2358 2359 contexts = rdma_get_devices(NULL); 2360 if (contexts == NULL) { 2361 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2362 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2363 return NULL; 2364 } 2365 2366 i = 0; 2367 rc = 0; 2368 while (contexts[i] != NULL) { 2369 device = calloc(1, sizeof(*device)); 2370 if (!device) { 2371 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2372 rc = -ENOMEM; 2373 break; 2374 } 2375 device->context = contexts[i]; 2376 rc = ibv_query_device(device->context, &device->attr); 2377 if (rc < 0) { 2378 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2379 free(device); 2380 break; 2381 2382 } 2383 2384 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2385 2386 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2387 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2388 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2389 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2390 } 2391 2392 /** 2393 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2394 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2395 * but incorrectly reports that it does. There are changes making their way 2396 * through the kernel now that will enable this feature. When they are merged, 2397 * we can conditionally enable this feature. 2398 * 2399 * TODO: enable this for versions of the kernel rxe driver that support it. 2400 */ 2401 if (nvmf_rdma_is_rxe_device(device)) { 2402 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2403 } 2404 #endif 2405 2406 /* set up device context async ev fd as NON_BLOCKING */ 2407 flag = fcntl(device->context->async_fd, F_GETFL); 2408 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2409 if (rc < 0) { 2410 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2411 free(device); 2412 break; 2413 } 2414 2415 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2416 i++; 2417 2418 if (g_nvmf_hooks.get_ibv_pd) { 2419 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2420 } else { 2421 device->pd = ibv_alloc_pd(device->context); 2422 } 2423 2424 if (!device->pd) { 2425 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2426 rc = -ENOMEM; 2427 break; 2428 } 2429 2430 assert(device->map == NULL); 2431 2432 device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks); 2433 if (!device->map) { 2434 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2435 rc = -ENOMEM; 2436 break; 2437 } 2438 2439 assert(device->map != NULL); 2440 assert(device->pd != NULL); 2441 } 2442 rdma_free_devices(contexts); 2443 2444 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2445 /* divide and round up. */ 2446 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2447 2448 /* round up to the nearest 4k. */ 2449 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2450 2451 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2452 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2453 opts->io_unit_size); 2454 } 2455 2456 if (rc < 0) { 2457 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2458 return NULL; 2459 } 2460 2461 /* Set up poll descriptor array to monitor events from RDMA and IB 2462 * in a single poll syscall 2463 */ 2464 rtransport->npoll_fds = i + 1; 2465 i = 0; 2466 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2467 if (rtransport->poll_fds == NULL) { 2468 SPDK_ERRLOG("poll_fds allocation failed\n"); 2469 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2470 return NULL; 2471 } 2472 2473 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2474 rtransport->poll_fds[i++].events = POLLIN; 2475 2476 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2477 rtransport->poll_fds[i].fd = device->context->async_fd; 2478 rtransport->poll_fds[i++].events = POLLIN; 2479 } 2480 2481 return &rtransport->transport; 2482 } 2483 2484 static void 2485 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2486 { 2487 struct spdk_nvmf_rdma_transport *rtransport; 2488 assert(w != NULL); 2489 2490 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2491 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2492 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2493 if (rtransport->rdma_opts.no_srq == true) { 2494 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2495 } 2496 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2497 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2498 } 2499 2500 static int 2501 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2502 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2503 { 2504 struct spdk_nvmf_rdma_transport *rtransport; 2505 struct spdk_nvmf_rdma_port *port, *port_tmp; 2506 struct spdk_nvmf_rdma_device *device, *device_tmp; 2507 2508 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2509 2510 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2511 TAILQ_REMOVE(&rtransport->ports, port, link); 2512 rdma_destroy_id(port->id); 2513 free(port); 2514 } 2515 2516 if (rtransport->poll_fds != NULL) { 2517 free(rtransport->poll_fds); 2518 } 2519 2520 if (rtransport->event_channel != NULL) { 2521 rdma_destroy_event_channel(rtransport->event_channel); 2522 } 2523 2524 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2525 TAILQ_REMOVE(&rtransport->devices, device, link); 2526 spdk_rdma_free_mem_map(&device->map); 2527 if (device->pd) { 2528 if (!g_nvmf_hooks.get_ibv_pd) { 2529 ibv_dealloc_pd(device->pd); 2530 } 2531 } 2532 free(device); 2533 } 2534 2535 if (rtransport->data_wr_pool != NULL) { 2536 if (spdk_mempool_count(rtransport->data_wr_pool) != 2537 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2538 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2539 spdk_mempool_count(rtransport->data_wr_pool), 2540 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2541 } 2542 } 2543 2544 spdk_mempool_free(rtransport->data_wr_pool); 2545 2546 pthread_mutex_destroy(&rtransport->lock); 2547 free(rtransport); 2548 2549 if (cb_fn) { 2550 cb_fn(cb_arg); 2551 } 2552 return 0; 2553 } 2554 2555 static int 2556 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2557 struct spdk_nvme_transport_id *trid, 2558 bool peer); 2559 2560 static int 2561 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2562 struct spdk_nvmf_listen_opts *listen_opts) 2563 { 2564 struct spdk_nvmf_rdma_transport *rtransport; 2565 struct spdk_nvmf_rdma_device *device; 2566 struct spdk_nvmf_rdma_port *port; 2567 struct addrinfo *res; 2568 struct addrinfo hints; 2569 int family; 2570 int rc; 2571 2572 if (!strlen(trid->trsvcid)) { 2573 SPDK_ERRLOG("Service id is required\n"); 2574 return -EINVAL; 2575 } 2576 2577 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2578 assert(rtransport->event_channel != NULL); 2579 2580 pthread_mutex_lock(&rtransport->lock); 2581 port = calloc(1, sizeof(*port)); 2582 if (!port) { 2583 SPDK_ERRLOG("Port allocation failed\n"); 2584 pthread_mutex_unlock(&rtransport->lock); 2585 return -ENOMEM; 2586 } 2587 2588 port->trid = trid; 2589 2590 switch (trid->adrfam) { 2591 case SPDK_NVMF_ADRFAM_IPV4: 2592 family = AF_INET; 2593 break; 2594 case SPDK_NVMF_ADRFAM_IPV6: 2595 family = AF_INET6; 2596 break; 2597 default: 2598 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2599 free(port); 2600 pthread_mutex_unlock(&rtransport->lock); 2601 return -EINVAL; 2602 } 2603 2604 memset(&hints, 0, sizeof(hints)); 2605 hints.ai_family = family; 2606 hints.ai_flags = AI_NUMERICSERV; 2607 hints.ai_socktype = SOCK_STREAM; 2608 hints.ai_protocol = 0; 2609 2610 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2611 if (rc) { 2612 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2613 free(port); 2614 pthread_mutex_unlock(&rtransport->lock); 2615 return -EINVAL; 2616 } 2617 2618 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2619 if (rc < 0) { 2620 SPDK_ERRLOG("rdma_create_id() failed\n"); 2621 freeaddrinfo(res); 2622 free(port); 2623 pthread_mutex_unlock(&rtransport->lock); 2624 return rc; 2625 } 2626 2627 rc = rdma_bind_addr(port->id, res->ai_addr); 2628 freeaddrinfo(res); 2629 2630 if (rc < 0) { 2631 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2632 rdma_destroy_id(port->id); 2633 free(port); 2634 pthread_mutex_unlock(&rtransport->lock); 2635 return rc; 2636 } 2637 2638 if (!port->id->verbs) { 2639 SPDK_ERRLOG("ibv_context is null\n"); 2640 rdma_destroy_id(port->id); 2641 free(port); 2642 pthread_mutex_unlock(&rtransport->lock); 2643 return -1; 2644 } 2645 2646 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2647 if (rc < 0) { 2648 SPDK_ERRLOG("rdma_listen() failed\n"); 2649 rdma_destroy_id(port->id); 2650 free(port); 2651 pthread_mutex_unlock(&rtransport->lock); 2652 return rc; 2653 } 2654 2655 TAILQ_FOREACH(device, &rtransport->devices, link) { 2656 if (device->context == port->id->verbs) { 2657 port->device = device; 2658 break; 2659 } 2660 } 2661 if (!port->device) { 2662 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2663 port->id->verbs); 2664 rdma_destroy_id(port->id); 2665 free(port); 2666 pthread_mutex_unlock(&rtransport->lock); 2667 return -EINVAL; 2668 } 2669 2670 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2671 trid->traddr, trid->trsvcid); 2672 2673 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2674 pthread_mutex_unlock(&rtransport->lock); 2675 return 0; 2676 } 2677 2678 static void 2679 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2680 const struct spdk_nvme_transport_id *trid) 2681 { 2682 struct spdk_nvmf_rdma_transport *rtransport; 2683 struct spdk_nvmf_rdma_port *port, *tmp; 2684 2685 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2686 2687 pthread_mutex_lock(&rtransport->lock); 2688 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2689 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2690 TAILQ_REMOVE(&rtransport->ports, port, link); 2691 rdma_destroy_id(port->id); 2692 free(port); 2693 break; 2694 } 2695 } 2696 2697 pthread_mutex_unlock(&rtransport->lock); 2698 } 2699 2700 static void 2701 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2702 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2703 { 2704 struct spdk_nvmf_request *req, *tmp; 2705 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2706 struct spdk_nvmf_rdma_resources *resources; 2707 2708 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2709 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2710 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2711 break; 2712 } 2713 } 2714 2715 /* Then RDMA writes since reads have stronger restrictions than writes */ 2716 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2717 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2718 break; 2719 } 2720 } 2721 2722 /* Then we handle request waiting on memory buffers. */ 2723 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2724 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2725 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2726 break; 2727 } 2728 } 2729 2730 resources = rqpair->resources; 2731 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2732 rdma_req = STAILQ_FIRST(&resources->free_queue); 2733 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2734 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2735 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2736 2737 if (rqpair->srq != NULL) { 2738 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2739 rdma_req->recv->qpair->qd++; 2740 } else { 2741 rqpair->qd++; 2742 } 2743 2744 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2745 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2746 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2747 break; 2748 } 2749 } 2750 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2751 rqpair->poller->stat.pending_free_request++; 2752 } 2753 } 2754 2755 static inline bool 2756 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 2757 { 2758 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 2759 return nvmf_rdma_is_rxe_device(device) || 2760 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 2761 } 2762 2763 static void 2764 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2765 { 2766 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2767 struct spdk_nvmf_rdma_transport, transport); 2768 2769 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2770 2771 /* nvmr_rdma_close_qpair is not called */ 2772 if (!rqpair->to_close) { 2773 return; 2774 } 2775 2776 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2777 if (rqpair->current_send_depth != 0) { 2778 return; 2779 } 2780 2781 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2782 return; 2783 } 2784 2785 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 2786 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 2787 return; 2788 } 2789 2790 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2791 2792 nvmf_rdma_qpair_destroy(rqpair); 2793 } 2794 2795 static int 2796 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2797 { 2798 struct spdk_nvmf_qpair *qpair; 2799 struct spdk_nvmf_rdma_qpair *rqpair; 2800 2801 if (evt->id == NULL) { 2802 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2803 return -1; 2804 } 2805 2806 qpair = evt->id->context; 2807 if (qpair == NULL) { 2808 SPDK_ERRLOG("disconnect request: no active connection\n"); 2809 return -1; 2810 } 2811 2812 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2813 2814 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair, 0); 2815 2816 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2817 2818 return 0; 2819 } 2820 2821 #ifdef DEBUG 2822 static const char *CM_EVENT_STR[] = { 2823 "RDMA_CM_EVENT_ADDR_RESOLVED", 2824 "RDMA_CM_EVENT_ADDR_ERROR", 2825 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2826 "RDMA_CM_EVENT_ROUTE_ERROR", 2827 "RDMA_CM_EVENT_CONNECT_REQUEST", 2828 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2829 "RDMA_CM_EVENT_CONNECT_ERROR", 2830 "RDMA_CM_EVENT_UNREACHABLE", 2831 "RDMA_CM_EVENT_REJECTED", 2832 "RDMA_CM_EVENT_ESTABLISHED", 2833 "RDMA_CM_EVENT_DISCONNECTED", 2834 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2835 "RDMA_CM_EVENT_MULTICAST_JOIN", 2836 "RDMA_CM_EVENT_MULTICAST_ERROR", 2837 "RDMA_CM_EVENT_ADDR_CHANGE", 2838 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2839 }; 2840 #endif /* DEBUG */ 2841 2842 static void 2843 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2844 struct spdk_nvmf_rdma_port *port) 2845 { 2846 struct spdk_nvmf_rdma_poll_group *rgroup; 2847 struct spdk_nvmf_rdma_poller *rpoller; 2848 struct spdk_nvmf_rdma_qpair *rqpair; 2849 2850 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2851 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2852 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 2853 if (rqpair->listen_id == port->id) { 2854 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2855 } 2856 } 2857 } 2858 } 2859 } 2860 2861 static bool 2862 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2863 struct rdma_cm_event *event) 2864 { 2865 const struct spdk_nvme_transport_id *trid; 2866 struct spdk_nvmf_rdma_port *port; 2867 struct spdk_nvmf_rdma_transport *rtransport; 2868 bool event_acked = false; 2869 2870 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2871 TAILQ_FOREACH(port, &rtransport->ports, link) { 2872 if (port->id == event->id) { 2873 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2874 rdma_ack_cm_event(event); 2875 event_acked = true; 2876 trid = port->trid; 2877 break; 2878 } 2879 } 2880 2881 if (event_acked) { 2882 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2883 2884 nvmf_rdma_stop_listen(transport, trid); 2885 nvmf_rdma_listen(transport, trid, NULL); 2886 } 2887 2888 return event_acked; 2889 } 2890 2891 static void 2892 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2893 struct rdma_cm_event *event) 2894 { 2895 struct spdk_nvmf_rdma_port *port; 2896 struct spdk_nvmf_rdma_transport *rtransport; 2897 2898 port = event->id->context; 2899 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2900 2901 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2902 2903 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2904 2905 rdma_ack_cm_event(event); 2906 2907 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2908 ; 2909 } 2910 } 2911 2912 static void 2913 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2914 { 2915 struct spdk_nvmf_rdma_transport *rtransport; 2916 struct rdma_cm_event *event; 2917 int rc; 2918 bool event_acked; 2919 2920 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2921 2922 if (rtransport->event_channel == NULL) { 2923 return; 2924 } 2925 2926 while (1) { 2927 event_acked = false; 2928 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2929 if (rc) { 2930 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2931 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2932 } 2933 break; 2934 } 2935 2936 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2937 2938 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2939 2940 switch (event->event) { 2941 case RDMA_CM_EVENT_ADDR_RESOLVED: 2942 case RDMA_CM_EVENT_ADDR_ERROR: 2943 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2944 case RDMA_CM_EVENT_ROUTE_ERROR: 2945 /* No action required. The target never attempts to resolve routes. */ 2946 break; 2947 case RDMA_CM_EVENT_CONNECT_REQUEST: 2948 rc = nvmf_rdma_connect(transport, event); 2949 if (rc < 0) { 2950 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2951 break; 2952 } 2953 break; 2954 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2955 /* The target never initiates a new connection. So this will not occur. */ 2956 break; 2957 case RDMA_CM_EVENT_CONNECT_ERROR: 2958 /* Can this happen? The docs say it can, but not sure what causes it. */ 2959 break; 2960 case RDMA_CM_EVENT_UNREACHABLE: 2961 case RDMA_CM_EVENT_REJECTED: 2962 /* These only occur on the client side. */ 2963 break; 2964 case RDMA_CM_EVENT_ESTABLISHED: 2965 /* TODO: Should we be waiting for this event anywhere? */ 2966 break; 2967 case RDMA_CM_EVENT_DISCONNECTED: 2968 rc = nvmf_rdma_disconnect(event); 2969 if (rc < 0) { 2970 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2971 break; 2972 } 2973 break; 2974 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2975 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 2976 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 2977 * Once these events are sent to SPDK, we should release all IB resources and 2978 * don't make attempts to call any ibv_query/modify/create functions. We can only call 2979 * ibv_destory* functions to release user space memory allocated by IB. All kernel 2980 * resources are already cleaned. */ 2981 if (event->id->qp) { 2982 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 2983 * corresponding qpair. Otherwise the event refers to a listening device */ 2984 rc = nvmf_rdma_disconnect(event); 2985 if (rc < 0) { 2986 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2987 break; 2988 } 2989 } else { 2990 nvmf_rdma_handle_cm_event_port_removal(transport, event); 2991 event_acked = true; 2992 } 2993 break; 2994 case RDMA_CM_EVENT_MULTICAST_JOIN: 2995 case RDMA_CM_EVENT_MULTICAST_ERROR: 2996 /* Multicast is not used */ 2997 break; 2998 case RDMA_CM_EVENT_ADDR_CHANGE: 2999 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3000 break; 3001 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3002 /* For now, do nothing. The target never re-uses queue pairs. */ 3003 break; 3004 default: 3005 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3006 break; 3007 } 3008 if (!event_acked) { 3009 rdma_ack_cm_event(event); 3010 } 3011 } 3012 } 3013 3014 static void 3015 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3016 { 3017 rqpair->last_wqe_reached = true; 3018 nvmf_rdma_destroy_drained_qpair(rqpair); 3019 } 3020 3021 static void 3022 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3023 { 3024 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3025 3026 if (event_ctx->rqpair) { 3027 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3028 if (event_ctx->cb_fn) { 3029 event_ctx->cb_fn(event_ctx->rqpair); 3030 } 3031 } 3032 free(event_ctx); 3033 } 3034 3035 static int 3036 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3037 spdk_nvmf_rdma_qpair_ibv_event fn) 3038 { 3039 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3040 struct spdk_thread *thr = NULL; 3041 int rc; 3042 3043 if (rqpair->qpair.group) { 3044 thr = rqpair->qpair.group->thread; 3045 } else if (rqpair->destruct_channel) { 3046 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3047 } 3048 3049 if (!thr) { 3050 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3051 return -EINVAL; 3052 } 3053 3054 ctx = calloc(1, sizeof(*ctx)); 3055 if (!ctx) { 3056 return -ENOMEM; 3057 } 3058 3059 ctx->rqpair = rqpair; 3060 ctx->cb_fn = fn; 3061 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3062 3063 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3064 if (rc) { 3065 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3066 free(ctx); 3067 } 3068 3069 return rc; 3070 } 3071 3072 static int 3073 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3074 { 3075 int rc; 3076 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3077 struct ibv_async_event event; 3078 3079 rc = ibv_get_async_event(device->context, &event); 3080 3081 if (rc) { 3082 /* In non-blocking mode -1 means there are no events available */ 3083 return rc; 3084 } 3085 3086 switch (event.event_type) { 3087 case IBV_EVENT_QP_FATAL: 3088 rqpair = event.element.qp->qp_context; 3089 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3090 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3091 (uintptr_t)rqpair, event.event_type); 3092 nvmf_rdma_update_ibv_state(rqpair); 3093 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3094 break; 3095 case IBV_EVENT_QP_LAST_WQE_REACHED: 3096 /* This event only occurs for shared receive queues. */ 3097 rqpair = event.element.qp->qp_context; 3098 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3099 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3100 if (rc) { 3101 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3102 rqpair->last_wqe_reached = true; 3103 } 3104 break; 3105 case IBV_EVENT_SQ_DRAINED: 3106 /* This event occurs frequently in both error and non-error states. 3107 * Check if the qpair is in an error state before sending a message. */ 3108 rqpair = event.element.qp->qp_context; 3109 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3110 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3111 (uintptr_t)rqpair, event.event_type); 3112 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3113 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3114 } 3115 break; 3116 case IBV_EVENT_QP_REQ_ERR: 3117 case IBV_EVENT_QP_ACCESS_ERR: 3118 case IBV_EVENT_COMM_EST: 3119 case IBV_EVENT_PATH_MIG: 3120 case IBV_EVENT_PATH_MIG_ERR: 3121 SPDK_NOTICELOG("Async event: %s\n", 3122 ibv_event_type_str(event.event_type)); 3123 rqpair = event.element.qp->qp_context; 3124 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3125 (uintptr_t)rqpair, event.event_type); 3126 nvmf_rdma_update_ibv_state(rqpair); 3127 break; 3128 case IBV_EVENT_CQ_ERR: 3129 case IBV_EVENT_DEVICE_FATAL: 3130 case IBV_EVENT_PORT_ACTIVE: 3131 case IBV_EVENT_PORT_ERR: 3132 case IBV_EVENT_LID_CHANGE: 3133 case IBV_EVENT_PKEY_CHANGE: 3134 case IBV_EVENT_SM_CHANGE: 3135 case IBV_EVENT_SRQ_ERR: 3136 case IBV_EVENT_SRQ_LIMIT_REACHED: 3137 case IBV_EVENT_CLIENT_REREGISTER: 3138 case IBV_EVENT_GID_CHANGE: 3139 default: 3140 SPDK_NOTICELOG("Async event: %s\n", 3141 ibv_event_type_str(event.event_type)); 3142 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3143 break; 3144 } 3145 ibv_ack_async_event(&event); 3146 3147 return 0; 3148 } 3149 3150 static void 3151 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3152 { 3153 int rc = 0; 3154 uint32_t i = 0; 3155 3156 for (i = 0; i < max_events; i++) { 3157 rc = nvmf_process_ib_event(device); 3158 if (rc) { 3159 break; 3160 } 3161 } 3162 3163 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3164 } 3165 3166 static uint32_t 3167 nvmf_rdma_accept(struct spdk_nvmf_transport *transport) 3168 { 3169 int nfds, i = 0; 3170 struct spdk_nvmf_rdma_transport *rtransport; 3171 struct spdk_nvmf_rdma_device *device, *tmp; 3172 uint32_t count; 3173 3174 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3175 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3176 3177 if (nfds <= 0) { 3178 return 0; 3179 } 3180 3181 /* The first poll descriptor is RDMA CM event */ 3182 if (rtransport->poll_fds[i++].revents & POLLIN) { 3183 nvmf_process_cm_event(transport); 3184 nfds--; 3185 } 3186 3187 if (nfds == 0) { 3188 return count; 3189 } 3190 3191 /* Second and subsequent poll descriptors are IB async events */ 3192 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3193 if (rtransport->poll_fds[i++].revents & POLLIN) { 3194 nvmf_process_ib_events(device, 32); 3195 nfds--; 3196 } 3197 } 3198 /* check all flagged fd's have been served */ 3199 assert(nfds == 0); 3200 3201 return count; 3202 } 3203 3204 static void 3205 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3206 struct spdk_nvmf_ctrlr_data *cdata) 3207 { 3208 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3209 3210 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3211 since in-capsule data only works with NVME drives that support SGL memory layout */ 3212 if (transport->opts.dif_insert_or_strip) { 3213 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3214 } 3215 } 3216 3217 static void 3218 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3219 struct spdk_nvme_transport_id *trid, 3220 struct spdk_nvmf_discovery_log_page_entry *entry) 3221 { 3222 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3223 entry->adrfam = trid->adrfam; 3224 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3225 3226 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3227 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3228 3229 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3230 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3231 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3232 } 3233 3234 static void 3235 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3236 3237 static struct spdk_nvmf_transport_poll_group * 3238 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3239 { 3240 struct spdk_nvmf_rdma_transport *rtransport; 3241 struct spdk_nvmf_rdma_poll_group *rgroup; 3242 struct spdk_nvmf_rdma_poller *poller; 3243 struct spdk_nvmf_rdma_device *device; 3244 struct spdk_rdma_srq_init_attr srq_init_attr; 3245 struct spdk_nvmf_rdma_resource_opts opts; 3246 int num_cqe; 3247 3248 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3249 3250 rgroup = calloc(1, sizeof(*rgroup)); 3251 if (!rgroup) { 3252 return NULL; 3253 } 3254 3255 TAILQ_INIT(&rgroup->pollers); 3256 3257 pthread_mutex_lock(&rtransport->lock); 3258 TAILQ_FOREACH(device, &rtransport->devices, link) { 3259 poller = calloc(1, sizeof(*poller)); 3260 if (!poller) { 3261 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3262 nvmf_rdma_poll_group_destroy(&rgroup->group); 3263 pthread_mutex_unlock(&rtransport->lock); 3264 return NULL; 3265 } 3266 3267 poller->device = device; 3268 poller->group = rgroup; 3269 3270 TAILQ_INIT(&poller->qpairs); 3271 STAILQ_INIT(&poller->qpairs_pending_send); 3272 STAILQ_INIT(&poller->qpairs_pending_recv); 3273 3274 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3275 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3276 poller->max_srq_depth = rtransport->rdma_opts.max_srq_depth; 3277 3278 device->num_srq++; 3279 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 3280 srq_init_attr.pd = device->pd; 3281 srq_init_attr.stats = &poller->stat.qp_stats.recv; 3282 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 3283 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3284 poller->srq = spdk_rdma_srq_create(&srq_init_attr); 3285 if (!poller->srq) { 3286 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3287 nvmf_rdma_poll_group_destroy(&rgroup->group); 3288 pthread_mutex_unlock(&rtransport->lock); 3289 return NULL; 3290 } 3291 3292 opts.qp = poller->srq; 3293 opts.pd = device->pd; 3294 opts.qpair = NULL; 3295 opts.shared = true; 3296 opts.max_queue_depth = poller->max_srq_depth; 3297 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3298 3299 poller->resources = nvmf_rdma_resources_create(&opts); 3300 if (!poller->resources) { 3301 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3302 nvmf_rdma_poll_group_destroy(&rgroup->group); 3303 pthread_mutex_unlock(&rtransport->lock); 3304 return NULL; 3305 } 3306 } 3307 3308 /* 3309 * When using an srq, we can limit the completion queue at startup. 3310 * The following formula represents the calculation: 3311 * num_cqe = num_recv + num_data_wr + num_send_wr. 3312 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3313 */ 3314 if (poller->srq) { 3315 num_cqe = poller->max_srq_depth * 3; 3316 } else { 3317 num_cqe = rtransport->rdma_opts.num_cqe; 3318 } 3319 3320 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3321 if (!poller->cq) { 3322 SPDK_ERRLOG("Unable to create completion queue\n"); 3323 nvmf_rdma_poll_group_destroy(&rgroup->group); 3324 pthread_mutex_unlock(&rtransport->lock); 3325 return NULL; 3326 } 3327 poller->num_cqe = num_cqe; 3328 } 3329 3330 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3331 if (rtransport->conn_sched.next_admin_pg == NULL) { 3332 rtransport->conn_sched.next_admin_pg = rgroup; 3333 rtransport->conn_sched.next_io_pg = rgroup; 3334 } 3335 3336 pthread_mutex_unlock(&rtransport->lock); 3337 return &rgroup->group; 3338 } 3339 3340 static struct spdk_nvmf_transport_poll_group * 3341 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3342 { 3343 struct spdk_nvmf_rdma_transport *rtransport; 3344 struct spdk_nvmf_rdma_poll_group **pg; 3345 struct spdk_nvmf_transport_poll_group *result; 3346 3347 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3348 3349 pthread_mutex_lock(&rtransport->lock); 3350 3351 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3352 pthread_mutex_unlock(&rtransport->lock); 3353 return NULL; 3354 } 3355 3356 if (qpair->qid == 0) { 3357 pg = &rtransport->conn_sched.next_admin_pg; 3358 } else { 3359 pg = &rtransport->conn_sched.next_io_pg; 3360 } 3361 3362 assert(*pg != NULL); 3363 3364 result = &(*pg)->group; 3365 3366 *pg = TAILQ_NEXT(*pg, link); 3367 if (*pg == NULL) { 3368 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3369 } 3370 3371 pthread_mutex_unlock(&rtransport->lock); 3372 3373 return result; 3374 } 3375 3376 static void 3377 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3378 { 3379 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3380 struct spdk_nvmf_rdma_poller *poller, *tmp; 3381 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3382 struct spdk_nvmf_rdma_transport *rtransport; 3383 3384 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3385 if (!rgroup) { 3386 return; 3387 } 3388 3389 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3390 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3391 3392 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3393 nvmf_rdma_qpair_destroy(qpair); 3394 } 3395 3396 if (poller->srq) { 3397 if (poller->resources) { 3398 nvmf_rdma_resources_destroy(poller->resources); 3399 } 3400 spdk_rdma_srq_destroy(poller->srq); 3401 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 3402 } 3403 3404 if (poller->cq) { 3405 ibv_destroy_cq(poller->cq); 3406 } 3407 3408 free(poller); 3409 } 3410 3411 if (rgroup->group.transport == NULL) { 3412 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3413 * calls this function directly in a failure path. */ 3414 free(rgroup); 3415 return; 3416 } 3417 3418 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3419 3420 pthread_mutex_lock(&rtransport->lock); 3421 next_rgroup = TAILQ_NEXT(rgroup, link); 3422 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3423 if (next_rgroup == NULL) { 3424 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3425 } 3426 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3427 rtransport->conn_sched.next_admin_pg = next_rgroup; 3428 } 3429 if (rtransport->conn_sched.next_io_pg == rgroup) { 3430 rtransport->conn_sched.next_io_pg = next_rgroup; 3431 } 3432 pthread_mutex_unlock(&rtransport->lock); 3433 3434 free(rgroup); 3435 } 3436 3437 static void 3438 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3439 { 3440 if (rqpair->cm_id != NULL) { 3441 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3442 } 3443 } 3444 3445 static int 3446 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3447 struct spdk_nvmf_qpair *qpair) 3448 { 3449 struct spdk_nvmf_rdma_poll_group *rgroup; 3450 struct spdk_nvmf_rdma_qpair *rqpair; 3451 struct spdk_nvmf_rdma_device *device; 3452 struct spdk_nvmf_rdma_poller *poller; 3453 int rc; 3454 3455 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3456 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3457 3458 device = rqpair->device; 3459 3460 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3461 if (poller->device == device) { 3462 break; 3463 } 3464 } 3465 3466 if (!poller) { 3467 SPDK_ERRLOG("No poller found for device.\n"); 3468 return -1; 3469 } 3470 3471 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3472 rqpair->poller = poller; 3473 rqpair->srq = rqpair->poller->srq; 3474 3475 rc = nvmf_rdma_qpair_initialize(qpair); 3476 if (rc < 0) { 3477 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3478 return -1; 3479 } 3480 3481 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3482 if (rc) { 3483 /* Try to reject, but we probably can't */ 3484 nvmf_rdma_qpair_reject_connection(rqpair); 3485 return -1; 3486 } 3487 3488 nvmf_rdma_update_ibv_state(rqpair); 3489 3490 return 0; 3491 } 3492 3493 static int 3494 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3495 struct spdk_nvmf_qpair *qpair) 3496 { 3497 struct spdk_nvmf_rdma_qpair *rqpair; 3498 3499 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3500 assert(group->transport->tgt != NULL); 3501 3502 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3503 3504 if (!rqpair->destruct_channel) { 3505 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3506 return 0; 3507 } 3508 3509 /* Sanity check that we get io_channel on the correct thread */ 3510 if (qpair->group) { 3511 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3512 } 3513 3514 return 0; 3515 } 3516 3517 static int 3518 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3519 { 3520 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3521 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3522 struct spdk_nvmf_rdma_transport, transport); 3523 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3524 struct spdk_nvmf_rdma_qpair, qpair); 3525 3526 /* 3527 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3528 * needs to be returned to the shared receive queue or the poll group will eventually be 3529 * starved of RECV structures. 3530 */ 3531 if (rqpair->srq && rdma_req->recv) { 3532 int rc; 3533 struct ibv_recv_wr *bad_recv_wr; 3534 3535 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 3536 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 3537 if (rc) { 3538 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3539 } 3540 } 3541 3542 _nvmf_rdma_request_free(rdma_req, rtransport); 3543 return 0; 3544 } 3545 3546 static int 3547 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3548 { 3549 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3550 struct spdk_nvmf_rdma_transport, transport); 3551 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3552 struct spdk_nvmf_rdma_request, req); 3553 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3554 struct spdk_nvmf_rdma_qpair, qpair); 3555 3556 if (rqpair->ibv_state != IBV_QPS_ERR) { 3557 /* The connection is alive, so process the request as normal */ 3558 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3559 } else { 3560 /* The connection is dead. Move the request directly to the completed state. */ 3561 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3562 } 3563 3564 nvmf_rdma_request_process(rtransport, rdma_req); 3565 3566 return 0; 3567 } 3568 3569 static void 3570 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 3571 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3572 { 3573 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3574 3575 rqpair->to_close = true; 3576 3577 /* This happens only when the qpair is disconnected before 3578 * it is added to the poll group. Since there is no poll group, 3579 * the RDMA qp has not been initialized yet and the RDMA CM 3580 * event has not yet been acknowledged, so we need to reject it. 3581 */ 3582 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3583 nvmf_rdma_qpair_reject_connection(rqpair); 3584 nvmf_rdma_qpair_destroy(rqpair); 3585 return; 3586 } 3587 3588 if (rqpair->rdma_qp) { 3589 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3590 } 3591 3592 nvmf_rdma_destroy_drained_qpair(rqpair); 3593 3594 if (cb_fn) { 3595 cb_fn(cb_arg); 3596 } 3597 } 3598 3599 static struct spdk_nvmf_rdma_qpair * 3600 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3601 { 3602 struct spdk_nvmf_rdma_qpair *rqpair; 3603 /* @todo: improve QP search */ 3604 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3605 if (wc->qp_num == rqpair->rdma_qp->qp->qp_num) { 3606 return rqpair; 3607 } 3608 } 3609 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3610 return NULL; 3611 } 3612 3613 #ifdef DEBUG 3614 static int 3615 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3616 { 3617 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3618 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3619 } 3620 #endif 3621 3622 static void 3623 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3624 int rc) 3625 { 3626 struct spdk_nvmf_rdma_recv *rdma_recv; 3627 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3628 3629 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3630 while (bad_recv_wr != NULL) { 3631 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3632 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3633 3634 rdma_recv->qpair->current_recv_depth++; 3635 bad_recv_wr = bad_recv_wr->next; 3636 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3637 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3638 } 3639 } 3640 3641 static void 3642 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3643 { 3644 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3645 while (bad_recv_wr != NULL) { 3646 bad_recv_wr = bad_recv_wr->next; 3647 rqpair->current_recv_depth++; 3648 } 3649 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3650 } 3651 3652 static void 3653 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3654 struct spdk_nvmf_rdma_poller *rpoller) 3655 { 3656 struct spdk_nvmf_rdma_qpair *rqpair; 3657 struct ibv_recv_wr *bad_recv_wr; 3658 int rc; 3659 3660 if (rpoller->srq) { 3661 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 3662 if (rc) { 3663 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3664 } 3665 } else { 3666 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3667 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3668 rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 3669 if (rc) { 3670 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3671 } 3672 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3673 } 3674 } 3675 } 3676 3677 static void 3678 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3679 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3680 { 3681 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3682 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3683 3684 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3685 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3686 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3687 assert(rqpair->current_send_depth > 0); 3688 rqpair->current_send_depth--; 3689 switch (bad_rdma_wr->type) { 3690 case RDMA_WR_TYPE_DATA: 3691 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3692 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3693 assert(rqpair->current_read_depth > 0); 3694 rqpair->current_read_depth--; 3695 } 3696 break; 3697 case RDMA_WR_TYPE_SEND: 3698 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3699 break; 3700 default: 3701 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3702 prev_rdma_req = cur_rdma_req; 3703 continue; 3704 } 3705 3706 if (prev_rdma_req == cur_rdma_req) { 3707 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3708 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3709 continue; 3710 } 3711 3712 switch (cur_rdma_req->state) { 3713 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3714 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3715 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3716 break; 3717 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3718 case RDMA_REQUEST_STATE_COMPLETING: 3719 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3720 break; 3721 default: 3722 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3723 cur_rdma_req->state, rqpair); 3724 continue; 3725 } 3726 3727 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3728 prev_rdma_req = cur_rdma_req; 3729 } 3730 3731 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3732 /* Disconnect the connection. */ 3733 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3734 } 3735 3736 } 3737 3738 static void 3739 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3740 struct spdk_nvmf_rdma_poller *rpoller) 3741 { 3742 struct spdk_nvmf_rdma_qpair *rqpair; 3743 struct ibv_send_wr *bad_wr = NULL; 3744 int rc; 3745 3746 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3747 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3748 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3749 3750 /* bad wr always points to the first wr that failed. */ 3751 if (rc) { 3752 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3753 } 3754 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3755 } 3756 } 3757 3758 static const char * 3759 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 3760 { 3761 switch (wr_type) { 3762 case RDMA_WR_TYPE_RECV: 3763 return "RECV"; 3764 case RDMA_WR_TYPE_SEND: 3765 return "SEND"; 3766 case RDMA_WR_TYPE_DATA: 3767 return "DATA"; 3768 default: 3769 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 3770 SPDK_UNREACHABLE(); 3771 } 3772 } 3773 3774 static inline void 3775 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 3776 { 3777 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 3778 3779 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 3780 /* If qpair is in ERR state, we will receive completions for all posted and not completed 3781 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 3782 SPDK_DEBUGLOG(rdma, 3783 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3784 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3785 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3786 } else { 3787 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3788 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3789 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3790 } 3791 } 3792 3793 static int 3794 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3795 struct spdk_nvmf_rdma_poller *rpoller) 3796 { 3797 struct ibv_wc wc[32]; 3798 struct spdk_nvmf_rdma_wr *rdma_wr; 3799 struct spdk_nvmf_rdma_request *rdma_req; 3800 struct spdk_nvmf_rdma_recv *rdma_recv; 3801 struct spdk_nvmf_rdma_qpair *rqpair; 3802 int reaped, i; 3803 int count = 0; 3804 bool error = false; 3805 uint64_t poll_tsc = spdk_get_ticks(); 3806 3807 /* Poll for completing operations. */ 3808 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3809 if (reaped < 0) { 3810 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3811 errno, spdk_strerror(errno)); 3812 return -1; 3813 } else if (reaped == 0) { 3814 rpoller->stat.idle_polls++; 3815 } 3816 3817 rpoller->stat.polls++; 3818 rpoller->stat.completions += reaped; 3819 3820 for (i = 0; i < reaped; i++) { 3821 3822 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3823 3824 switch (rdma_wr->type) { 3825 case RDMA_WR_TYPE_SEND: 3826 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3827 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3828 3829 if (!wc[i].status) { 3830 count++; 3831 assert(wc[i].opcode == IBV_WC_SEND); 3832 assert(nvmf_rdma_req_is_completing(rdma_req)); 3833 } 3834 3835 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3836 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3837 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3838 rdma_req->num_outstanding_data_wr = 0; 3839 3840 nvmf_rdma_request_process(rtransport, rdma_req); 3841 break; 3842 case RDMA_WR_TYPE_RECV: 3843 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3844 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3845 if (rpoller->srq != NULL) { 3846 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3847 /* It is possible that there are still some completions for destroyed QP 3848 * associated with SRQ. We just ignore these late completions and re-post 3849 * receive WRs back to SRQ. 3850 */ 3851 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3852 struct ibv_recv_wr *bad_wr; 3853 int rc; 3854 3855 rdma_recv->wr.next = NULL; 3856 spdk_rdma_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 3857 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 3858 if (rc) { 3859 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3860 } 3861 continue; 3862 } 3863 } 3864 rqpair = rdma_recv->qpair; 3865 3866 assert(rqpair != NULL); 3867 if (!wc[i].status) { 3868 assert(wc[i].opcode == IBV_WC_RECV); 3869 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3870 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3871 break; 3872 } 3873 } 3874 3875 rdma_recv->wr.next = NULL; 3876 rqpair->current_recv_depth++; 3877 rdma_recv->receive_tsc = poll_tsc; 3878 rpoller->stat.requests++; 3879 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3880 break; 3881 case RDMA_WR_TYPE_DATA: 3882 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3883 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3884 3885 assert(rdma_req->num_outstanding_data_wr > 0); 3886 3887 rqpair->current_send_depth--; 3888 rdma_req->num_outstanding_data_wr--; 3889 if (!wc[i].status) { 3890 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3891 rqpair->current_read_depth--; 3892 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3893 if (rdma_req->num_outstanding_data_wr == 0) { 3894 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3895 nvmf_rdma_request_process(rtransport, rdma_req); 3896 } 3897 } else { 3898 /* If the data transfer fails still force the queue into the error state, 3899 * if we were performing an RDMA_READ, we need to force the request into a 3900 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3901 * case, we should wait for the SEND to complete. */ 3902 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3903 rqpair->current_read_depth--; 3904 if (rdma_req->num_outstanding_data_wr == 0) { 3905 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3906 } 3907 } 3908 } 3909 break; 3910 default: 3911 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3912 continue; 3913 } 3914 3915 /* Handle error conditions */ 3916 if (wc[i].status) { 3917 nvmf_rdma_update_ibv_state(rqpair); 3918 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 3919 3920 error = true; 3921 3922 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3923 /* Disconnect the connection. */ 3924 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3925 } else { 3926 nvmf_rdma_destroy_drained_qpair(rqpair); 3927 } 3928 continue; 3929 } 3930 3931 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3932 3933 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3934 nvmf_rdma_destroy_drained_qpair(rqpair); 3935 } 3936 } 3937 3938 if (error == true) { 3939 return -1; 3940 } 3941 3942 /* submit outstanding work requests. */ 3943 _poller_submit_recvs(rtransport, rpoller); 3944 _poller_submit_sends(rtransport, rpoller); 3945 3946 return count; 3947 } 3948 3949 static int 3950 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3951 { 3952 struct spdk_nvmf_rdma_transport *rtransport; 3953 struct spdk_nvmf_rdma_poll_group *rgroup; 3954 struct spdk_nvmf_rdma_poller *rpoller; 3955 int count, rc; 3956 3957 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3958 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3959 3960 count = 0; 3961 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3962 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 3963 if (rc < 0) { 3964 return rc; 3965 } 3966 count += rc; 3967 } 3968 3969 return count; 3970 } 3971 3972 static int 3973 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3974 struct spdk_nvme_transport_id *trid, 3975 bool peer) 3976 { 3977 struct sockaddr *saddr; 3978 uint16_t port; 3979 3980 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 3981 3982 if (peer) { 3983 saddr = rdma_get_peer_addr(id); 3984 } else { 3985 saddr = rdma_get_local_addr(id); 3986 } 3987 switch (saddr->sa_family) { 3988 case AF_INET: { 3989 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3990 3991 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3992 inet_ntop(AF_INET, &saddr_in->sin_addr, 3993 trid->traddr, sizeof(trid->traddr)); 3994 if (peer) { 3995 port = ntohs(rdma_get_dst_port(id)); 3996 } else { 3997 port = ntohs(rdma_get_src_port(id)); 3998 } 3999 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4000 break; 4001 } 4002 case AF_INET6: { 4003 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4004 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4005 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4006 trid->traddr, sizeof(trid->traddr)); 4007 if (peer) { 4008 port = ntohs(rdma_get_dst_port(id)); 4009 } else { 4010 port = ntohs(rdma_get_src_port(id)); 4011 } 4012 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4013 break; 4014 } 4015 default: 4016 return -1; 4017 4018 } 4019 4020 return 0; 4021 } 4022 4023 static int 4024 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4025 struct spdk_nvme_transport_id *trid) 4026 { 4027 struct spdk_nvmf_rdma_qpair *rqpair; 4028 4029 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4030 4031 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4032 } 4033 4034 static int 4035 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4036 struct spdk_nvme_transport_id *trid) 4037 { 4038 struct spdk_nvmf_rdma_qpair *rqpair; 4039 4040 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4041 4042 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4043 } 4044 4045 static int 4046 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4047 struct spdk_nvme_transport_id *trid) 4048 { 4049 struct spdk_nvmf_rdma_qpair *rqpair; 4050 4051 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4052 4053 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4054 } 4055 4056 void 4057 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4058 { 4059 g_nvmf_hooks = *hooks; 4060 } 4061 4062 static void 4063 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4064 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4065 { 4066 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4067 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4068 4069 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4070 4071 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4072 } 4073 4074 static int 4075 _nvmf_rdma_qpair_abort_request(void *ctx) 4076 { 4077 struct spdk_nvmf_request *req = ctx; 4078 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4079 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4080 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4081 struct spdk_nvmf_rdma_qpair, qpair); 4082 int rc; 4083 4084 spdk_poller_unregister(&req->poller); 4085 4086 switch (rdma_req_to_abort->state) { 4087 case RDMA_REQUEST_STATE_EXECUTING: 4088 rc = nvmf_ctrlr_abort_request(req); 4089 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4090 return SPDK_POLLER_BUSY; 4091 } 4092 break; 4093 4094 case RDMA_REQUEST_STATE_NEED_BUFFER: 4095 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4096 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4097 4098 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4099 break; 4100 4101 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4102 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4103 spdk_nvmf_rdma_request, state_link); 4104 4105 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4106 break; 4107 4108 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4109 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4110 spdk_nvmf_rdma_request, state_link); 4111 4112 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4113 break; 4114 4115 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4116 if (spdk_get_ticks() < req->timeout_tsc) { 4117 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4118 return SPDK_POLLER_BUSY; 4119 } 4120 break; 4121 4122 default: 4123 break; 4124 } 4125 4126 spdk_nvmf_request_complete(req); 4127 return SPDK_POLLER_BUSY; 4128 } 4129 4130 static void 4131 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4132 struct spdk_nvmf_request *req) 4133 { 4134 struct spdk_nvmf_rdma_qpair *rqpair; 4135 struct spdk_nvmf_rdma_transport *rtransport; 4136 struct spdk_nvmf_transport *transport; 4137 uint16_t cid; 4138 uint32_t i, max_req_count; 4139 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 4140 4141 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4142 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4143 transport = &rtransport->transport; 4144 4145 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4146 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 4147 4148 for (i = 0; i < max_req_count; i++) { 4149 rdma_req = &rqpair->resources->reqs[i]; 4150 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 4151 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 4152 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 4153 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 4154 rdma_req->req.qpair == qpair) { 4155 rdma_req_to_abort = rdma_req; 4156 break; 4157 } 4158 } 4159 4160 if (rdma_req_to_abort == NULL) { 4161 spdk_nvmf_request_complete(req); 4162 return; 4163 } 4164 4165 req->req_to_abort = &rdma_req_to_abort->req; 4166 req->timeout_tsc = spdk_get_ticks() + 4167 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4168 req->poller = NULL; 4169 4170 _nvmf_rdma_qpair_abort_request(req); 4171 } 4172 4173 /* Deprecated, please use the flow with nvmf_rdma_poll_group_dump_stat. */ 4174 static int 4175 nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4176 struct spdk_nvmf_transport_poll_group_stat **stat) 4177 { 4178 struct spdk_io_channel *ch; 4179 struct spdk_nvmf_poll_group *group; 4180 struct spdk_nvmf_transport_poll_group *tgroup; 4181 struct spdk_nvmf_rdma_poll_group *rgroup; 4182 struct spdk_nvmf_rdma_poller *rpoller; 4183 struct spdk_nvmf_rdma_device_stat *device_stat; 4184 uint64_t num_devices = 0; 4185 4186 SPDK_ERRLOG("nvmf_rdma_poll_group_get_stat is deprecated and will be removed\n"); 4187 4188 if (tgt == NULL || stat == NULL) { 4189 return -EINVAL; 4190 } 4191 4192 ch = spdk_get_io_channel(tgt); 4193 group = spdk_io_channel_get_ctx(ch);; 4194 spdk_put_io_channel(ch); 4195 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4196 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4197 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4198 if (!*stat) { 4199 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4200 return -ENOMEM; 4201 } 4202 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4203 4204 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4205 /* Count devices to allocate enough memory */ 4206 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4207 ++num_devices; 4208 } 4209 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4210 if (!(*stat)->rdma.devices) { 4211 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4212 free(*stat); 4213 return -ENOMEM; 4214 } 4215 4216 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4217 (*stat)->rdma.num_devices = num_devices; 4218 num_devices = 0; 4219 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4220 device_stat = &(*stat)->rdma.devices[num_devices++]; 4221 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4222 device_stat->polls = rpoller->stat.polls; 4223 device_stat->idle_polls = rpoller->stat.idle_polls; 4224 device_stat->completions = rpoller->stat.completions; 4225 device_stat->requests = rpoller->stat.requests; 4226 device_stat->request_latency = rpoller->stat.request_latency; 4227 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4228 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4229 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4230 device_stat->total_send_wrs = rpoller->stat.qp_stats.send.num_submitted_wrs; 4231 device_stat->send_doorbell_updates = rpoller->stat.qp_stats.send.doorbell_updates; 4232 device_stat->total_recv_wrs = rpoller->stat.qp_stats.recv.num_submitted_wrs; 4233 device_stat->recv_doorbell_updates = rpoller->stat.qp_stats.recv.doorbell_updates; 4234 } 4235 return 0; 4236 } 4237 } 4238 return -ENOENT; 4239 } 4240 4241 /* Deprecated, please use the flow with nvmf_rdma_poll_group_dump_stat. */ 4242 static void 4243 nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4244 { 4245 SPDK_ERRLOG("nvmf_rdma_poll_group_free_stat is deprecated and will be removed\n"); 4246 4247 if (stat) { 4248 free(stat->rdma.devices); 4249 } 4250 free(stat); 4251 } 4252 4253 static void 4254 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 4255 struct spdk_json_write_ctx *w) 4256 { 4257 struct spdk_nvmf_rdma_poll_group *rgroup; 4258 struct spdk_nvmf_rdma_poller *rpoller; 4259 4260 assert(w != NULL); 4261 4262 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4263 4264 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 4265 4266 spdk_json_write_named_array_begin(w, "devices"); 4267 4268 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4269 spdk_json_write_object_begin(w); 4270 spdk_json_write_named_string(w, "name", 4271 ibv_get_device_name(rpoller->device->context->device)); 4272 spdk_json_write_named_uint64(w, "polls", 4273 rpoller->stat.polls); 4274 spdk_json_write_named_uint64(w, "idle_polls", 4275 rpoller->stat.idle_polls); 4276 spdk_json_write_named_uint64(w, "completions", 4277 rpoller->stat.completions); 4278 spdk_json_write_named_uint64(w, "requests", 4279 rpoller->stat.requests); 4280 spdk_json_write_named_uint64(w, "request_latency", 4281 rpoller->stat.request_latency); 4282 spdk_json_write_named_uint64(w, "pending_free_request", 4283 rpoller->stat.pending_free_request); 4284 spdk_json_write_named_uint64(w, "pending_rdma_read", 4285 rpoller->stat.pending_rdma_read); 4286 spdk_json_write_named_uint64(w, "pending_rdma_write", 4287 rpoller->stat.pending_rdma_write); 4288 spdk_json_write_named_uint64(w, "total_send_wrs", 4289 rpoller->stat.qp_stats.send.num_submitted_wrs); 4290 spdk_json_write_named_uint64(w, "send_doorbell_updates", 4291 rpoller->stat.qp_stats.send.doorbell_updates); 4292 spdk_json_write_named_uint64(w, "total_recv_wrs", 4293 rpoller->stat.qp_stats.recv.num_submitted_wrs); 4294 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 4295 rpoller->stat.qp_stats.recv.doorbell_updates); 4296 spdk_json_write_object_end(w); 4297 } 4298 4299 spdk_json_write_array_end(w); 4300 } 4301 4302 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4303 .name = "RDMA", 4304 .type = SPDK_NVME_TRANSPORT_RDMA, 4305 .opts_init = nvmf_rdma_opts_init, 4306 .create = nvmf_rdma_create, 4307 .dump_opts = nvmf_rdma_dump_opts, 4308 .destroy = nvmf_rdma_destroy, 4309 4310 .listen = nvmf_rdma_listen, 4311 .stop_listen = nvmf_rdma_stop_listen, 4312 .accept = nvmf_rdma_accept, 4313 .cdata_init = nvmf_rdma_cdata_init, 4314 4315 .listener_discover = nvmf_rdma_discover, 4316 4317 .poll_group_create = nvmf_rdma_poll_group_create, 4318 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4319 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4320 .poll_group_add = nvmf_rdma_poll_group_add, 4321 .poll_group_remove = nvmf_rdma_poll_group_remove, 4322 .poll_group_poll = nvmf_rdma_poll_group_poll, 4323 4324 .req_free = nvmf_rdma_request_free, 4325 .req_complete = nvmf_rdma_request_complete, 4326 4327 .qpair_fini = nvmf_rdma_close_qpair, 4328 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4329 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4330 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4331 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4332 4333 .poll_group_get_stat = nvmf_rdma_poll_group_get_stat, 4334 .poll_group_free_stat = nvmf_rdma_poll_group_free_stat, 4335 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 4336 }; 4337 4338 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4339 SPDK_LOG_REGISTER_COMPONENT(rdma) 4340