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