1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2019-2021 Mellanox Technologies LTD. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of Intel Corporation nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include "spdk/stdinc.h" 35 36 #include "spdk/config.h" 37 #include "spdk/thread.h" 38 #include "spdk/likely.h" 39 #include "spdk/nvmf_transport.h" 40 #include "spdk/string.h" 41 #include "spdk/trace.h" 42 #include "spdk/util.h" 43 44 #include "spdk_internal/assert.h" 45 #include "spdk/log.h" 46 #include "spdk_internal/rdma.h" 47 48 #include "nvmf_internal.h" 49 50 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 51 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 52 53 /* 54 RDMA Connection Resource Defaults 55 */ 56 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 57 #define NVMF_DEFAULT_RSP_SGE 1 58 #define NVMF_DEFAULT_RX_SGE 2 59 60 /* The RDMA completion queue size */ 61 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 62 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 63 64 /* rxe driver vendor_id has been changed from 0 to 0XFFFFFF in 0184afd15a141d7ce24c32c0d86a1e3ba6bc0eb3 */ 65 #define NVMF_RXE_VENDOR_ID_OLD 0 66 #define NVMF_RXE_VENDOR_ID_NEW 0XFFFFFF 67 68 static int g_spdk_nvmf_ibv_query_mask = 69 IBV_QP_STATE | 70 IBV_QP_PKEY_INDEX | 71 IBV_QP_PORT | 72 IBV_QP_ACCESS_FLAGS | 73 IBV_QP_AV | 74 IBV_QP_PATH_MTU | 75 IBV_QP_DEST_QPN | 76 IBV_QP_RQ_PSN | 77 IBV_QP_MAX_DEST_RD_ATOMIC | 78 IBV_QP_MIN_RNR_TIMER | 79 IBV_QP_SQ_PSN | 80 IBV_QP_TIMEOUT | 81 IBV_QP_RETRY_CNT | 82 IBV_QP_RNR_RETRY | 83 IBV_QP_MAX_QP_RD_ATOMIC; 84 85 enum spdk_nvmf_rdma_request_state { 86 /* The request is not currently in use */ 87 RDMA_REQUEST_STATE_FREE = 0, 88 89 /* Initial state when request first received */ 90 RDMA_REQUEST_STATE_NEW, 91 92 /* The request is queued until a data buffer is available. */ 93 RDMA_REQUEST_STATE_NEED_BUFFER, 94 95 /* The request is waiting on RDMA queue depth availability 96 * to transfer data from the host to the controller. 97 */ 98 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 99 100 /* The request is currently transferring data from the host to the controller. */ 101 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 102 103 /* The request is ready to execute at the block device */ 104 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 105 106 /* The request is currently executing at the block device */ 107 RDMA_REQUEST_STATE_EXECUTING, 108 109 /* The request finished executing at the block device */ 110 RDMA_REQUEST_STATE_EXECUTED, 111 112 /* The request is waiting on RDMA queue depth availability 113 * to transfer data from the controller to the host. 114 */ 115 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 116 117 /* The request is ready to send a completion */ 118 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 119 120 /* The request is currently transferring data from the controller to the host. */ 121 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 122 123 /* The request currently has an outstanding completion without an 124 * associated data transfer. 125 */ 126 RDMA_REQUEST_STATE_COMPLETING, 127 128 /* The request completed and can be marked free. */ 129 RDMA_REQUEST_STATE_COMPLETED, 130 131 /* Terminator */ 132 RDMA_REQUEST_NUM_STATES, 133 }; 134 135 #define OBJECT_NVMF_RDMA_IO 0x40 136 137 #define TRACE_GROUP_NVMF_RDMA 0x4 138 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 139 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 140 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 141 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 142 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 143 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 144 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 145 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 146 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 147 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 148 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 149 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 150 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 151 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 152 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 153 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 154 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 155 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 156 157 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 158 { 159 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 160 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 161 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 162 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 163 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 164 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 165 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 166 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 167 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 168 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 169 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 170 spdk_trace_register_description("RDMA_REQ_TX_H2C", 171 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 172 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 173 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 174 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 175 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 176 spdk_trace_register_description("RDMA_REQ_EXECUTING", 177 TRACE_RDMA_REQUEST_STATE_EXECUTING, 178 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 179 spdk_trace_register_description("RDMA_REQ_EXECUTED", 180 TRACE_RDMA_REQUEST_STATE_EXECUTED, 181 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 182 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 183 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 184 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 185 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 186 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 187 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 188 spdk_trace_register_description("RDMA_REQ_COMPLETING", 189 TRACE_RDMA_REQUEST_STATE_COMPLETING, 190 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 191 spdk_trace_register_description("RDMA_REQ_COMPLETED", 192 TRACE_RDMA_REQUEST_STATE_COMPLETED, 193 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 194 195 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 196 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 197 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 198 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 199 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 200 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 201 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 202 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 203 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 204 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 205 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 206 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 207 } 208 209 enum spdk_nvmf_rdma_wr_type { 210 RDMA_WR_TYPE_RECV, 211 RDMA_WR_TYPE_SEND, 212 RDMA_WR_TYPE_DATA, 213 }; 214 215 struct spdk_nvmf_rdma_wr { 216 enum spdk_nvmf_rdma_wr_type type; 217 }; 218 219 /* This structure holds commands as they are received off the wire. 220 * It must be dynamically paired with a full request object 221 * (spdk_nvmf_rdma_request) to service a request. It is separate 222 * from the request because RDMA does not appear to order 223 * completions, so occasionally we'll get a new incoming 224 * command when there aren't any free request objects. 225 */ 226 struct spdk_nvmf_rdma_recv { 227 struct ibv_recv_wr wr; 228 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 229 230 struct spdk_nvmf_rdma_qpair *qpair; 231 232 /* In-capsule data buffer */ 233 uint8_t *buf; 234 235 struct spdk_nvmf_rdma_wr rdma_wr; 236 uint64_t receive_tsc; 237 238 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 239 }; 240 241 struct spdk_nvmf_rdma_request_data { 242 struct spdk_nvmf_rdma_wr rdma_wr; 243 struct ibv_send_wr wr; 244 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 245 }; 246 247 struct spdk_nvmf_rdma_request { 248 struct spdk_nvmf_request req; 249 250 enum spdk_nvmf_rdma_request_state state; 251 252 struct spdk_nvmf_rdma_recv *recv; 253 254 struct { 255 struct spdk_nvmf_rdma_wr rdma_wr; 256 struct ibv_send_wr wr; 257 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 258 } rsp; 259 260 struct spdk_nvmf_rdma_request_data data; 261 262 uint32_t iovpos; 263 264 uint32_t num_outstanding_data_wr; 265 uint64_t receive_tsc; 266 267 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 268 }; 269 270 struct spdk_nvmf_rdma_resource_opts { 271 struct spdk_nvmf_rdma_qpair *qpair; 272 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 273 void *qp; 274 struct ibv_pd *pd; 275 uint32_t max_queue_depth; 276 uint32_t in_capsule_data_size; 277 bool shared; 278 }; 279 280 struct spdk_nvmf_rdma_resources { 281 /* Array of size "max_queue_depth" containing RDMA requests. */ 282 struct spdk_nvmf_rdma_request *reqs; 283 284 /* Array of size "max_queue_depth" containing RDMA recvs. */ 285 struct spdk_nvmf_rdma_recv *recvs; 286 287 /* Array of size "max_queue_depth" containing 64 byte capsules 288 * used for receive. 289 */ 290 union nvmf_h2c_msg *cmds; 291 struct ibv_mr *cmds_mr; 292 293 /* Array of size "max_queue_depth" containing 16 byte completions 294 * to be sent back to the user. 295 */ 296 union nvmf_c2h_msg *cpls; 297 struct ibv_mr *cpls_mr; 298 299 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 300 * buffers to be used for in capsule data. 301 */ 302 void *bufs; 303 struct ibv_mr *bufs_mr; 304 305 /* Receives that are waiting for a request object */ 306 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 307 308 /* Queue to track free requests */ 309 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 310 }; 311 312 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 313 314 struct spdk_nvmf_rdma_ibv_event_ctx { 315 struct spdk_nvmf_rdma_qpair *rqpair; 316 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 317 /* Link to other ibv events associated with this qpair */ 318 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 319 }; 320 321 struct spdk_nvmf_rdma_qpair { 322 struct spdk_nvmf_qpair qpair; 323 324 struct spdk_nvmf_rdma_device *device; 325 struct spdk_nvmf_rdma_poller *poller; 326 327 struct spdk_rdma_qp *rdma_qp; 328 struct rdma_cm_id *cm_id; 329 struct spdk_rdma_srq *srq; 330 struct rdma_cm_id *listen_id; 331 332 /* The maximum number of I/O outstanding on this connection at one time */ 333 uint16_t max_queue_depth; 334 335 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 336 uint16_t max_read_depth; 337 338 /* The maximum number of RDMA SEND operations at one time */ 339 uint32_t max_send_depth; 340 341 /* The current number of outstanding WRs from this qpair's 342 * recv queue. Should not exceed device->attr.max_queue_depth. 343 */ 344 uint16_t current_recv_depth; 345 346 /* The current number of active RDMA READ operations */ 347 uint16_t current_read_depth; 348 349 /* The current number of posted WRs from this qpair's 350 * send queue. Should not exceed max_send_depth. 351 */ 352 uint32_t current_send_depth; 353 354 /* The maximum number of SGEs per WR on the send queue */ 355 uint32_t max_send_sge; 356 357 /* The maximum number of SGEs per WR on the recv queue */ 358 uint32_t max_recv_sge; 359 360 struct spdk_nvmf_rdma_resources *resources; 361 362 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 363 364 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 365 366 /* Number of requests not in the free state */ 367 uint32_t qd; 368 369 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 370 371 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 372 373 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 374 375 /* IBV queue pair attributes: they are used to manage 376 * qp state and recover from errors. 377 */ 378 enum ibv_qp_state ibv_state; 379 380 /* 381 * io_channel which is used to destroy qpair when it is removed from poll group 382 */ 383 struct spdk_io_channel *destruct_channel; 384 385 /* List of ibv async events */ 386 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 387 388 /* Lets us know that we have received the last_wqe event. */ 389 bool last_wqe_reached; 390 391 /* Indicate that nvmf_rdma_close_qpair is called */ 392 bool to_close; 393 }; 394 395 struct spdk_nvmf_rdma_poller_stat { 396 uint64_t completions; 397 uint64_t polls; 398 uint64_t idle_polls; 399 uint64_t requests; 400 uint64_t request_latency; 401 uint64_t pending_free_request; 402 uint64_t pending_rdma_read; 403 uint64_t pending_rdma_write; 404 struct spdk_rdma_qp_stats qp_stats; 405 }; 406 407 struct spdk_nvmf_rdma_poller { 408 struct spdk_nvmf_rdma_device *device; 409 struct spdk_nvmf_rdma_poll_group *group; 410 411 int num_cqe; 412 int required_num_wr; 413 struct ibv_cq *cq; 414 415 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 416 uint16_t max_srq_depth; 417 418 /* Shared receive queue */ 419 struct spdk_rdma_srq *srq; 420 421 struct spdk_nvmf_rdma_resources *resources; 422 struct spdk_nvmf_rdma_poller_stat stat; 423 424 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 425 426 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 427 428 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 429 430 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 431 }; 432 433 struct spdk_nvmf_rdma_poll_group_stat { 434 uint64_t pending_data_buffer; 435 }; 436 437 struct spdk_nvmf_rdma_poll_group { 438 struct spdk_nvmf_transport_poll_group group; 439 struct spdk_nvmf_rdma_poll_group_stat stat; 440 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 441 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 442 }; 443 444 struct spdk_nvmf_rdma_conn_sched { 445 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 446 struct spdk_nvmf_rdma_poll_group *next_io_pg; 447 }; 448 449 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 450 struct spdk_nvmf_rdma_device { 451 struct ibv_device_attr attr; 452 struct ibv_context *context; 453 454 struct spdk_rdma_mem_map *map; 455 struct ibv_pd *pd; 456 457 int num_srq; 458 459 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 460 }; 461 462 struct spdk_nvmf_rdma_port { 463 const struct spdk_nvme_transport_id *trid; 464 struct rdma_cm_id *id; 465 struct spdk_nvmf_rdma_device *device; 466 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 467 }; 468 469 struct rdma_transport_opts { 470 int num_cqe; 471 uint32_t max_srq_depth; 472 bool no_srq; 473 bool no_wr_batching; 474 int acceptor_backlog; 475 }; 476 477 struct spdk_nvmf_rdma_transport { 478 struct spdk_nvmf_transport transport; 479 struct rdma_transport_opts rdma_opts; 480 481 struct spdk_nvmf_rdma_conn_sched conn_sched; 482 483 struct rdma_event_channel *event_channel; 484 485 struct spdk_mempool *data_wr_pool; 486 487 pthread_mutex_t lock; 488 489 /* fields used to poll RDMA/IB events */ 490 nfds_t npoll_fds; 491 struct pollfd *poll_fds; 492 493 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 494 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 495 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 496 }; 497 498 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 499 { 500 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe), 501 spdk_json_decode_int32, true 502 }, 503 { 504 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 505 spdk_json_decode_uint32, true 506 }, 507 { 508 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 509 spdk_json_decode_bool, true 510 }, 511 { 512 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 513 spdk_json_decode_bool, true 514 }, 515 { 516 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 517 spdk_json_decode_int32, true 518 }, 519 }; 520 521 static bool 522 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 523 struct spdk_nvmf_rdma_request *rdma_req); 524 525 static void 526 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 527 struct spdk_nvmf_rdma_poller *rpoller); 528 529 static void 530 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 531 struct spdk_nvmf_rdma_poller *rpoller); 532 533 static inline int 534 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 535 { 536 switch (state) { 537 case IBV_QPS_RESET: 538 case IBV_QPS_INIT: 539 case IBV_QPS_RTR: 540 case IBV_QPS_RTS: 541 case IBV_QPS_SQD: 542 case IBV_QPS_SQE: 543 case IBV_QPS_ERR: 544 return 0; 545 default: 546 return -1; 547 } 548 } 549 550 static inline enum spdk_nvme_media_error_status_code 551 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 552 enum spdk_nvme_media_error_status_code result; 553 switch (err_type) 554 { 555 case SPDK_DIF_REFTAG_ERROR: 556 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 557 break; 558 case SPDK_DIF_APPTAG_ERROR: 559 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 560 break; 561 case SPDK_DIF_GUARD_ERROR: 562 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 563 break; 564 default: 565 SPDK_UNREACHABLE(); 566 } 567 568 return result; 569 } 570 571 static enum ibv_qp_state 572 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 573 enum ibv_qp_state old_state, new_state; 574 struct ibv_qp_attr qp_attr; 575 struct ibv_qp_init_attr init_attr; 576 int rc; 577 578 old_state = rqpair->ibv_state; 579 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 580 g_spdk_nvmf_ibv_query_mask, &init_attr); 581 582 if (rc) 583 { 584 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 585 return IBV_QPS_ERR + 1; 586 } 587 588 new_state = qp_attr.qp_state; 589 rqpair->ibv_state = new_state; 590 qp_attr.ah_attr.port_num = qp_attr.port_num; 591 592 rc = nvmf_rdma_check_ibv_state(new_state); 593 if (rc) 594 { 595 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 596 /* 597 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 598 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 599 */ 600 return IBV_QPS_ERR + 1; 601 } 602 603 if (old_state != new_state) 604 { 605 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 606 (uintptr_t)rqpair->cm_id, new_state); 607 } 608 return new_state; 609 } 610 611 static void 612 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 613 struct spdk_nvmf_rdma_transport *rtransport) 614 { 615 struct spdk_nvmf_rdma_request_data *data_wr; 616 struct ibv_send_wr *next_send_wr; 617 uint64_t req_wrid; 618 619 rdma_req->num_outstanding_data_wr = 0; 620 data_wr = &rdma_req->data; 621 req_wrid = data_wr->wr.wr_id; 622 while (data_wr && data_wr->wr.wr_id == req_wrid) { 623 memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge); 624 data_wr->wr.num_sge = 0; 625 next_send_wr = data_wr->wr.next; 626 if (data_wr != &rdma_req->data) { 627 spdk_mempool_put(rtransport->data_wr_pool, data_wr); 628 } 629 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : 630 SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr); 631 } 632 } 633 634 static void 635 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 636 { 637 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 638 if (req->req.cmd) { 639 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 640 } 641 if (req->recv) { 642 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 643 } 644 } 645 646 static void 647 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 648 { 649 int i; 650 651 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 652 for (i = 0; i < rqpair->max_queue_depth; i++) { 653 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 654 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 655 } 656 } 657 } 658 659 static void 660 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 661 { 662 if (resources->cmds_mr) { 663 ibv_dereg_mr(resources->cmds_mr); 664 } 665 666 if (resources->cpls_mr) { 667 ibv_dereg_mr(resources->cpls_mr); 668 } 669 670 if (resources->bufs_mr) { 671 ibv_dereg_mr(resources->bufs_mr); 672 } 673 674 spdk_free(resources->cmds); 675 spdk_free(resources->cpls); 676 spdk_free(resources->bufs); 677 free(resources->reqs); 678 free(resources->recvs); 679 free(resources); 680 } 681 682 683 static struct spdk_nvmf_rdma_resources * 684 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 685 { 686 struct spdk_nvmf_rdma_resources *resources; 687 struct spdk_nvmf_rdma_request *rdma_req; 688 struct spdk_nvmf_rdma_recv *rdma_recv; 689 struct spdk_rdma_qp *qp = NULL; 690 struct spdk_rdma_srq *srq = NULL; 691 struct ibv_recv_wr *bad_wr = NULL; 692 uint32_t i; 693 int rc = 0; 694 695 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 696 if (!resources) { 697 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 698 return NULL; 699 } 700 701 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 702 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 703 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 704 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 705 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 706 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 707 708 if (opts->in_capsule_data_size > 0) { 709 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 710 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 711 SPDK_MALLOC_DMA); 712 } 713 714 if (!resources->reqs || !resources->recvs || !resources->cmds || 715 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 716 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 717 goto cleanup; 718 } 719 720 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 721 opts->max_queue_depth * sizeof(*resources->cmds), 722 IBV_ACCESS_LOCAL_WRITE); 723 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 724 opts->max_queue_depth * sizeof(*resources->cpls), 725 0); 726 727 if (opts->in_capsule_data_size) { 728 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 729 opts->max_queue_depth * 730 opts->in_capsule_data_size, 731 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 732 } 733 734 if (!resources->cmds_mr || !resources->cpls_mr || 735 (opts->in_capsule_data_size && 736 !resources->bufs_mr)) { 737 goto cleanup; 738 } 739 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx LKey: %x\n", 740 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 741 resources->cmds_mr->lkey); 742 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx LKey: %x\n", 743 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 744 resources->cpls_mr->lkey); 745 if (resources->bufs && resources->bufs_mr) { 746 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x LKey: %x\n", 747 resources->bufs, opts->max_queue_depth * 748 opts->in_capsule_data_size, resources->bufs_mr->lkey); 749 } 750 751 /* Initialize queues */ 752 STAILQ_INIT(&resources->incoming_queue); 753 STAILQ_INIT(&resources->free_queue); 754 755 if (opts->shared) { 756 srq = (struct spdk_rdma_srq *)opts->qp; 757 } else { 758 qp = (struct spdk_rdma_qp *)opts->qp; 759 } 760 761 for (i = 0; i < opts->max_queue_depth; i++) { 762 rdma_recv = &resources->recvs[i]; 763 rdma_recv->qpair = opts->qpair; 764 765 /* Set up memory to receive commands */ 766 if (resources->bufs) { 767 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 768 opts->in_capsule_data_size)); 769 } 770 771 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 772 773 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 774 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 775 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 776 rdma_recv->wr.num_sge = 1; 777 778 if (rdma_recv->buf && resources->bufs_mr) { 779 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 780 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 781 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 782 rdma_recv->wr.num_sge++; 783 } 784 785 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 786 rdma_recv->wr.sg_list = rdma_recv->sgl; 787 if (srq) { 788 spdk_rdma_srq_queue_recv_wrs(srq, &rdma_recv->wr); 789 } else { 790 spdk_rdma_qp_queue_recv_wrs(qp, &rdma_recv->wr); 791 } 792 } 793 794 for (i = 0; i < opts->max_queue_depth; i++) { 795 rdma_req = &resources->reqs[i]; 796 797 if (opts->qpair != NULL) { 798 rdma_req->req.qpair = &opts->qpair->qpair; 799 } else { 800 rdma_req->req.qpair = NULL; 801 } 802 rdma_req->req.cmd = NULL; 803 804 /* Set up memory to send responses */ 805 rdma_req->req.rsp = &resources->cpls[i]; 806 807 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 808 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 809 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 810 811 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 812 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 813 rdma_req->rsp.wr.next = NULL; 814 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 815 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 816 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 817 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 818 819 /* Set up memory for data buffers */ 820 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 821 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 822 rdma_req->data.wr.next = NULL; 823 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 824 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 825 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 826 827 /* Initialize request state to FREE */ 828 rdma_req->state = RDMA_REQUEST_STATE_FREE; 829 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 830 } 831 832 if (srq) { 833 rc = spdk_rdma_srq_flush_recv_wrs(srq, &bad_wr); 834 } else { 835 rc = spdk_rdma_qp_flush_recv_wrs(qp, &bad_wr); 836 } 837 838 if (rc) { 839 goto cleanup; 840 } 841 842 return resources; 843 844 cleanup: 845 nvmf_rdma_resources_destroy(resources); 846 return NULL; 847 } 848 849 static void 850 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 851 { 852 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 853 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 854 ctx->rqpair = NULL; 855 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 856 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 857 } 858 } 859 860 static void 861 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 862 { 863 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 864 struct ibv_recv_wr *bad_recv_wr = NULL; 865 int rc; 866 867 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 868 869 if (rqpair->qd != 0) { 870 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 871 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 872 struct spdk_nvmf_rdma_transport, transport); 873 struct spdk_nvmf_rdma_request *req; 874 uint32_t i, max_req_count = 0; 875 876 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 877 878 if (rqpair->srq == NULL) { 879 nvmf_rdma_dump_qpair_contents(rqpair); 880 max_req_count = rqpair->max_queue_depth; 881 } else if (rqpair->poller && rqpair->resources) { 882 max_req_count = rqpair->poller->max_srq_depth; 883 } 884 885 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 886 for (i = 0; i < max_req_count; i++) { 887 req = &rqpair->resources->reqs[i]; 888 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 889 /* nvmf_rdma_request_process checks qpair ibv and internal state 890 * and completes a request */ 891 nvmf_rdma_request_process(rtransport, req); 892 } 893 } 894 assert(rqpair->qd == 0); 895 } 896 897 if (rqpair->poller) { 898 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 899 900 if (rqpair->srq != NULL && rqpair->resources != NULL) { 901 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 902 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 903 if (rqpair == rdma_recv->qpair) { 904 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 905 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr); 906 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 907 if (rc) { 908 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 909 } 910 } 911 } 912 } 913 } 914 915 if (rqpair->cm_id) { 916 if (rqpair->rdma_qp != NULL) { 917 spdk_rdma_qp_destroy(rqpair->rdma_qp); 918 rqpair->rdma_qp = NULL; 919 } 920 rdma_destroy_id(rqpair->cm_id); 921 922 if (rqpair->poller != NULL && rqpair->srq == NULL) { 923 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 924 } 925 } 926 927 if (rqpair->srq == NULL && rqpair->resources != NULL) { 928 nvmf_rdma_resources_destroy(rqpair->resources); 929 } 930 931 nvmf_rdma_qpair_clean_ibv_events(rqpair); 932 933 if (rqpair->destruct_channel) { 934 spdk_put_io_channel(rqpair->destruct_channel); 935 rqpair->destruct_channel = NULL; 936 } 937 938 free(rqpair); 939 } 940 941 static int 942 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 943 { 944 struct spdk_nvmf_rdma_poller *rpoller; 945 int rc, num_cqe, required_num_wr; 946 947 /* Enlarge CQ size dynamically */ 948 rpoller = rqpair->poller; 949 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 950 num_cqe = rpoller->num_cqe; 951 if (num_cqe < required_num_wr) { 952 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 953 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 954 } 955 956 if (rpoller->num_cqe != num_cqe) { 957 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) { 958 SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n" 959 "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe); 960 return -1; 961 } 962 if (required_num_wr > device->attr.max_cqe) { 963 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 964 required_num_wr, device->attr.max_cqe); 965 return -1; 966 } 967 968 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 969 rc = ibv_resize_cq(rpoller->cq, num_cqe); 970 if (rc) { 971 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 972 return -1; 973 } 974 975 rpoller->num_cqe = num_cqe; 976 } 977 978 rpoller->required_num_wr = required_num_wr; 979 return 0; 980 } 981 982 static int 983 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 984 { 985 struct spdk_nvmf_rdma_qpair *rqpair; 986 struct spdk_nvmf_rdma_transport *rtransport; 987 struct spdk_nvmf_transport *transport; 988 struct spdk_nvmf_rdma_resource_opts opts; 989 struct spdk_nvmf_rdma_device *device; 990 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 991 992 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 993 device = rqpair->device; 994 995 qp_init_attr.qp_context = rqpair; 996 qp_init_attr.pd = device->pd; 997 qp_init_attr.send_cq = rqpair->poller->cq; 998 qp_init_attr.recv_cq = rqpair->poller->cq; 999 1000 if (rqpair->srq) { 1001 qp_init_attr.srq = rqpair->srq->srq; 1002 } else { 1003 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 1004 } 1005 1006 /* SEND, READ, and WRITE operations */ 1007 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 1008 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 1009 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1010 qp_init_attr.stats = &rqpair->poller->stat.qp_stats; 1011 1012 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1013 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1014 goto error; 1015 } 1016 1017 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 1018 if (!rqpair->rdma_qp) { 1019 goto error; 1020 } 1021 1022 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1023 qp_init_attr.cap.max_send_wr); 1024 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1025 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1026 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 1027 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1028 1029 if (rqpair->poller->srq == NULL) { 1030 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1031 transport = &rtransport->transport; 1032 1033 opts.qp = rqpair->rdma_qp; 1034 opts.pd = rqpair->cm_id->pd; 1035 opts.qpair = rqpair; 1036 opts.shared = false; 1037 opts.max_queue_depth = rqpair->max_queue_depth; 1038 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1039 1040 rqpair->resources = nvmf_rdma_resources_create(&opts); 1041 1042 if (!rqpair->resources) { 1043 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1044 rdma_destroy_qp(rqpair->cm_id); 1045 goto error; 1046 } 1047 } else { 1048 rqpair->resources = rqpair->poller->resources; 1049 } 1050 1051 rqpair->current_recv_depth = 0; 1052 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1053 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1054 1055 return 0; 1056 1057 error: 1058 rdma_destroy_id(rqpair->cm_id); 1059 rqpair->cm_id = NULL; 1060 return -1; 1061 } 1062 1063 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1064 /* This function accepts either a single wr or the first wr in a linked list. */ 1065 static void 1066 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1067 { 1068 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1069 struct spdk_nvmf_rdma_transport, transport); 1070 1071 if (rqpair->srq != NULL) { 1072 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, first); 1073 } else { 1074 if (spdk_rdma_qp_queue_recv_wrs(rqpair->rdma_qp, first)) { 1075 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1076 } 1077 } 1078 1079 if (rtransport->rdma_opts.no_wr_batching) { 1080 _poller_submit_recvs(rtransport, rqpair->poller); 1081 } 1082 } 1083 1084 static int 1085 request_transfer_in(struct spdk_nvmf_request *req) 1086 { 1087 struct spdk_nvmf_rdma_request *rdma_req; 1088 struct spdk_nvmf_qpair *qpair; 1089 struct spdk_nvmf_rdma_qpair *rqpair; 1090 struct spdk_nvmf_rdma_transport *rtransport; 1091 1092 qpair = req->qpair; 1093 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1094 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1095 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1096 struct spdk_nvmf_rdma_transport, transport); 1097 1098 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1099 assert(rdma_req != NULL); 1100 1101 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1102 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1103 } 1104 if (rtransport->rdma_opts.no_wr_batching) { 1105 _poller_submit_sends(rtransport, rqpair->poller); 1106 } 1107 1108 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1109 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1110 return 0; 1111 } 1112 1113 static int 1114 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1115 { 1116 int num_outstanding_data_wr = 0; 1117 struct spdk_nvmf_rdma_request *rdma_req; 1118 struct spdk_nvmf_qpair *qpair; 1119 struct spdk_nvmf_rdma_qpair *rqpair; 1120 struct spdk_nvme_cpl *rsp; 1121 struct ibv_send_wr *first = NULL; 1122 struct spdk_nvmf_rdma_transport *rtransport; 1123 1124 *data_posted = 0; 1125 qpair = req->qpair; 1126 rsp = &req->rsp->nvme_cpl; 1127 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1128 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1129 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1130 struct spdk_nvmf_rdma_transport, transport); 1131 1132 /* Advance our sq_head pointer */ 1133 if (qpair->sq_head == qpair->sq_head_max) { 1134 qpair->sq_head = 0; 1135 } else { 1136 qpair->sq_head++; 1137 } 1138 rsp->sqhd = qpair->sq_head; 1139 1140 /* queue the capsule for the recv buffer */ 1141 assert(rdma_req->recv != NULL); 1142 1143 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1144 1145 rdma_req->recv = NULL; 1146 assert(rqpair->current_recv_depth > 0); 1147 rqpair->current_recv_depth--; 1148 1149 /* Build the response which consists of optional 1150 * RDMA WRITEs to transfer data, plus an RDMA SEND 1151 * containing the response. 1152 */ 1153 first = &rdma_req->rsp.wr; 1154 1155 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1156 /* On failure, data was not read from the controller. So clear the 1157 * number of outstanding data WRs to zero. 1158 */ 1159 rdma_req->num_outstanding_data_wr = 0; 1160 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1161 first = &rdma_req->data.wr; 1162 *data_posted = 1; 1163 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1164 } 1165 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1166 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1167 } 1168 if (rtransport->rdma_opts.no_wr_batching) { 1169 _poller_submit_sends(rtransport, rqpair->poller); 1170 } 1171 1172 /* +1 for the rsp wr */ 1173 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1174 1175 return 0; 1176 } 1177 1178 static int 1179 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1180 { 1181 struct spdk_nvmf_rdma_accept_private_data accept_data; 1182 struct rdma_conn_param ctrlr_event_data = {}; 1183 int rc; 1184 1185 accept_data.recfmt = 0; 1186 accept_data.crqsize = rqpair->max_queue_depth; 1187 1188 ctrlr_event_data.private_data = &accept_data; 1189 ctrlr_event_data.private_data_len = sizeof(accept_data); 1190 if (id->ps == RDMA_PS_TCP) { 1191 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1192 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1193 } 1194 1195 /* Configure infinite retries for the initiator side qpair. 1196 * When using a shared receive queue on the target side, 1197 * we need to pass this value to the initiator to prevent the 1198 * initiator side NIC from completing SEND requests back to the 1199 * initiator with status rnr_retry_count_exceeded. */ 1200 if (rqpair->srq != NULL) { 1201 ctrlr_event_data.rnr_retry_count = 0x7; 1202 } 1203 1204 /* When qpair is created without use of rdma cm API, an additional 1205 * information must be provided to initiator in the connection response: 1206 * whether qpair is using SRQ and its qp_num 1207 * Fields below are ignored by rdma cm if qpair has been 1208 * created using rdma cm API. */ 1209 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1210 ctrlr_event_data.qp_num = rqpair->rdma_qp->qp->qp_num; 1211 1212 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1213 if (rc) { 1214 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1215 } else { 1216 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1217 } 1218 1219 return rc; 1220 } 1221 1222 static void 1223 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1224 { 1225 struct spdk_nvmf_rdma_reject_private_data rej_data; 1226 1227 rej_data.recfmt = 0; 1228 rej_data.sts = error; 1229 1230 rdma_reject(id, &rej_data, sizeof(rej_data)); 1231 } 1232 1233 static int 1234 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1235 { 1236 struct spdk_nvmf_rdma_transport *rtransport; 1237 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1238 struct spdk_nvmf_rdma_port *port; 1239 struct rdma_conn_param *rdma_param = NULL; 1240 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1241 uint16_t max_queue_depth; 1242 uint16_t max_read_depth; 1243 1244 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1245 1246 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1247 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1248 1249 rdma_param = &event->param.conn; 1250 if (rdma_param->private_data == NULL || 1251 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1252 SPDK_ERRLOG("connect request: no private data provided\n"); 1253 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1254 return -1; 1255 } 1256 1257 private_data = rdma_param->private_data; 1258 if (private_data->recfmt != 0) { 1259 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1260 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1261 return -1; 1262 } 1263 1264 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1265 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1266 1267 port = event->listen_id->context; 1268 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1269 event->listen_id, event->listen_id->verbs, port); 1270 1271 /* Figure out the supported queue depth. This is a multi-step process 1272 * that takes into account hardware maximums, host provided values, 1273 * and our target's internal memory limits */ 1274 1275 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1276 1277 /* Start with the maximum queue depth allowed by the target */ 1278 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1279 max_read_depth = rtransport->transport.opts.max_queue_depth; 1280 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1281 rtransport->transport.opts.max_queue_depth); 1282 1283 /* Next check the local NIC's hardware limitations */ 1284 SPDK_DEBUGLOG(rdma, 1285 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1286 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1287 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1288 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1289 1290 /* Next check the remote NIC's hardware limitations */ 1291 SPDK_DEBUGLOG(rdma, 1292 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1293 rdma_param->initiator_depth, rdma_param->responder_resources); 1294 if (rdma_param->initiator_depth > 0) { 1295 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1296 } 1297 1298 /* Finally check for the host software requested values, which are 1299 * optional. */ 1300 if (rdma_param->private_data != NULL && 1301 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1302 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1303 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1304 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1305 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1306 } 1307 1308 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1309 max_queue_depth, max_read_depth); 1310 1311 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1312 if (rqpair == NULL) { 1313 SPDK_ERRLOG("Could not allocate new connection.\n"); 1314 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1315 return -1; 1316 } 1317 1318 rqpair->device = port->device; 1319 rqpair->max_queue_depth = max_queue_depth; 1320 rqpair->max_read_depth = max_read_depth; 1321 rqpair->cm_id = event->id; 1322 rqpair->listen_id = event->listen_id; 1323 rqpair->qpair.transport = transport; 1324 STAILQ_INIT(&rqpair->ibv_events); 1325 /* use qid from the private data to determine the qpair type 1326 qid will be set to the appropriate value when the controller is created */ 1327 rqpair->qpair.qid = private_data->qid; 1328 1329 event->id->context = &rqpair->qpair; 1330 1331 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1332 1333 return 0; 1334 } 1335 1336 static inline void 1337 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1338 enum spdk_nvme_data_transfer xfer) 1339 { 1340 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1341 wr->opcode = IBV_WR_RDMA_WRITE; 1342 wr->send_flags = 0; 1343 wr->next = next; 1344 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1345 wr->opcode = IBV_WR_RDMA_READ; 1346 wr->send_flags = IBV_SEND_SIGNALED; 1347 wr->next = NULL; 1348 } else { 1349 assert(0); 1350 } 1351 } 1352 1353 static int 1354 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1355 struct spdk_nvmf_rdma_request *rdma_req, 1356 uint32_t num_sgl_descriptors) 1357 { 1358 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1359 struct spdk_nvmf_rdma_request_data *current_data_wr; 1360 uint32_t i; 1361 1362 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1363 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1364 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1365 return -EINVAL; 1366 } 1367 1368 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1369 return -ENOMEM; 1370 } 1371 1372 current_data_wr = &rdma_req->data; 1373 1374 for (i = 0; i < num_sgl_descriptors; i++) { 1375 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1376 current_data_wr->wr.next = &work_requests[i]->wr; 1377 current_data_wr = work_requests[i]; 1378 current_data_wr->wr.sg_list = current_data_wr->sgl; 1379 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1380 } 1381 1382 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1383 1384 return 0; 1385 } 1386 1387 static inline void 1388 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1389 { 1390 struct ibv_send_wr *wr = &rdma_req->data.wr; 1391 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1392 1393 wr->wr.rdma.rkey = sgl->keyed.key; 1394 wr->wr.rdma.remote_addr = sgl->address; 1395 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1396 } 1397 1398 static inline void 1399 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1400 { 1401 struct ibv_send_wr *wr = &rdma_req->data.wr; 1402 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1403 uint32_t i; 1404 int j; 1405 uint64_t remote_addr_offset = 0; 1406 1407 for (i = 0; i < num_wrs; ++i) { 1408 wr->wr.rdma.rkey = sgl->keyed.key; 1409 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1410 for (j = 0; j < wr->num_sge; ++j) { 1411 remote_addr_offset += wr->sg_list[j].length; 1412 } 1413 wr = wr->next; 1414 } 1415 } 1416 1417 static bool 1418 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device, 1419 struct iovec *iov, struct ibv_send_wr **_wr, 1420 uint32_t *_remaining_data_block, uint32_t *_offset, 1421 uint32_t *_num_extra_wrs, 1422 const struct spdk_dif_ctx *dif_ctx) 1423 { 1424 struct ibv_send_wr *wr = *_wr; 1425 struct ibv_sge *sg_ele = &wr->sg_list[wr->num_sge]; 1426 struct spdk_rdma_memory_translation mem_translation; 1427 int rc; 1428 uint32_t lkey = 0; 1429 uint32_t remaining, data_block_size, md_size, sge_len; 1430 1431 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation); 1432 if (spdk_unlikely(rc)) { 1433 return false; 1434 } 1435 1436 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation); 1437 1438 if (spdk_likely(!dif_ctx)) { 1439 sg_ele->lkey = lkey; 1440 sg_ele->addr = (uintptr_t)(iov->iov_base); 1441 sg_ele->length = iov->iov_len; 1442 wr->num_sge++; 1443 } else { 1444 remaining = iov->iov_len - *_offset; 1445 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1446 md_size = dif_ctx->md_size; 1447 1448 while (remaining) { 1449 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1450 if (*_num_extra_wrs > 0 && wr->next) { 1451 *_wr = wr->next; 1452 wr = *_wr; 1453 wr->num_sge = 0; 1454 sg_ele = &wr->sg_list[wr->num_sge]; 1455 (*_num_extra_wrs)--; 1456 } else { 1457 break; 1458 } 1459 } 1460 sg_ele->lkey = lkey; 1461 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset); 1462 sge_len = spdk_min(remaining, *_remaining_data_block); 1463 sg_ele->length = sge_len; 1464 remaining -= sge_len; 1465 *_remaining_data_block -= sge_len; 1466 *_offset += sge_len; 1467 1468 sg_ele++; 1469 wr->num_sge++; 1470 1471 if (*_remaining_data_block == 0) { 1472 /* skip metadata */ 1473 *_offset += md_size; 1474 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1475 remaining -= spdk_min(remaining, md_size); 1476 *_remaining_data_block = data_block_size; 1477 } 1478 1479 if (remaining == 0) { 1480 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1481 the remaining metadata bits at the beginning of the next buffer */ 1482 *_offset -= iov->iov_len; 1483 } 1484 } 1485 } 1486 1487 return true; 1488 } 1489 1490 static int 1491 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1492 struct spdk_nvmf_rdma_device *device, 1493 struct spdk_nvmf_rdma_request *rdma_req, 1494 struct ibv_send_wr *wr, 1495 uint32_t length, 1496 uint32_t num_extra_wrs) 1497 { 1498 struct spdk_nvmf_request *req = &rdma_req->req; 1499 struct spdk_dif_ctx *dif_ctx = NULL; 1500 uint32_t remaining_data_block = 0; 1501 uint32_t offset = 0; 1502 1503 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 1504 dif_ctx = &rdma_req->req.dif.dif_ctx; 1505 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1506 } 1507 1508 wr->num_sge = 0; 1509 1510 while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1511 if (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr, 1512 &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) { 1513 return -EINVAL; 1514 } 1515 1516 length -= req->iov[rdma_req->iovpos].iov_len; 1517 rdma_req->iovpos++; 1518 } 1519 1520 if (length) { 1521 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1522 return -EINVAL; 1523 } 1524 1525 return 0; 1526 } 1527 1528 static inline uint32_t 1529 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1530 { 1531 /* estimate the number of SG entries and WRs needed to process the request */ 1532 uint32_t num_sge = 0; 1533 uint32_t i; 1534 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1535 1536 for (i = 0; i < num_buffers && length > 0; i++) { 1537 uint32_t buffer_len = spdk_min(length, io_unit_size); 1538 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1539 1540 if (num_sge_in_block * block_size > buffer_len) { 1541 ++num_sge_in_block; 1542 } 1543 num_sge += num_sge_in_block; 1544 length -= buffer_len; 1545 } 1546 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1547 } 1548 1549 static int 1550 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1551 struct spdk_nvmf_rdma_device *device, 1552 struct spdk_nvmf_rdma_request *rdma_req, 1553 uint32_t length) 1554 { 1555 struct spdk_nvmf_rdma_qpair *rqpair; 1556 struct spdk_nvmf_rdma_poll_group *rgroup; 1557 struct spdk_nvmf_request *req = &rdma_req->req; 1558 struct ibv_send_wr *wr = &rdma_req->data.wr; 1559 int rc; 1560 uint32_t num_wrs = 1; 1561 1562 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1563 rgroup = rqpair->poller->group; 1564 1565 /* rdma wr specifics */ 1566 nvmf_rdma_setup_request(rdma_req); 1567 1568 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1569 length); 1570 if (rc != 0) { 1571 return rc; 1572 } 1573 1574 assert(req->iovcnt <= rqpair->max_send_sge); 1575 1576 rdma_req->iovpos = 0; 1577 1578 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1579 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1580 req->dif.dif_ctx.block_size); 1581 if (num_wrs > 1) { 1582 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1583 if (rc != 0) { 1584 goto err_exit; 1585 } 1586 } 1587 } 1588 1589 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1590 if (spdk_unlikely(rc != 0)) { 1591 goto err_exit; 1592 } 1593 1594 if (spdk_unlikely(num_wrs > 1)) { 1595 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1596 } 1597 1598 /* set the number of outstanding data WRs for this request. */ 1599 rdma_req->num_outstanding_data_wr = num_wrs; 1600 1601 return rc; 1602 1603 err_exit: 1604 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1605 nvmf_rdma_request_free_data(rdma_req, rtransport); 1606 req->iovcnt = 0; 1607 return rc; 1608 } 1609 1610 static int 1611 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1612 struct spdk_nvmf_rdma_device *device, 1613 struct spdk_nvmf_rdma_request *rdma_req) 1614 { 1615 struct spdk_nvmf_rdma_qpair *rqpair; 1616 struct spdk_nvmf_rdma_poll_group *rgroup; 1617 struct ibv_send_wr *current_wr; 1618 struct spdk_nvmf_request *req = &rdma_req->req; 1619 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1620 uint32_t num_sgl_descriptors; 1621 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES]; 1622 uint32_t i; 1623 int rc; 1624 1625 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1626 rgroup = rqpair->poller->group; 1627 1628 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1629 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1630 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1631 1632 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1633 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1634 1635 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1636 return -ENOMEM; 1637 } 1638 1639 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1640 for (i = 0; i < num_sgl_descriptors; i++) { 1641 if (spdk_likely(!req->dif.dif_insert_or_strip)) { 1642 lengths[i] = desc->keyed.length; 1643 } else { 1644 req->dif.orig_length += desc->keyed.length; 1645 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1646 req->dif.elba_length += lengths[i]; 1647 } 1648 desc++; 1649 } 1650 1651 rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport, 1652 lengths, num_sgl_descriptors); 1653 if (rc != 0) { 1654 nvmf_rdma_request_free_data(rdma_req, rtransport); 1655 return rc; 1656 } 1657 1658 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1659 current_wr = &rdma_req->data.wr; 1660 assert(current_wr != NULL); 1661 1662 req->length = 0; 1663 rdma_req->iovpos = 0; 1664 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1665 for (i = 0; i < num_sgl_descriptors; i++) { 1666 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1667 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1668 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1669 rc = -EINVAL; 1670 goto err_exit; 1671 } 1672 1673 current_wr->num_sge = 0; 1674 1675 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1676 if (rc != 0) { 1677 rc = -ENOMEM; 1678 goto err_exit; 1679 } 1680 1681 req->length += desc->keyed.length; 1682 current_wr->wr.rdma.rkey = desc->keyed.key; 1683 current_wr->wr.rdma.remote_addr = desc->address; 1684 current_wr = current_wr->next; 1685 desc++; 1686 } 1687 1688 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1689 /* Go back to the last descriptor in the list. */ 1690 desc--; 1691 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1692 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1693 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1694 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1695 } 1696 } 1697 #endif 1698 1699 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1700 1701 return 0; 1702 1703 err_exit: 1704 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1705 nvmf_rdma_request_free_data(rdma_req, rtransport); 1706 return rc; 1707 } 1708 1709 static int 1710 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1711 struct spdk_nvmf_rdma_device *device, 1712 struct spdk_nvmf_rdma_request *rdma_req) 1713 { 1714 struct spdk_nvmf_request *req = &rdma_req->req; 1715 struct spdk_nvme_cpl *rsp; 1716 struct spdk_nvme_sgl_descriptor *sgl; 1717 int rc; 1718 uint32_t length; 1719 1720 rsp = &req->rsp->nvme_cpl; 1721 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1722 1723 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1724 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1725 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1726 1727 length = sgl->keyed.length; 1728 if (length > rtransport->transport.opts.max_io_size) { 1729 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1730 length, rtransport->transport.opts.max_io_size); 1731 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1732 return -1; 1733 } 1734 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1735 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1736 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1737 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1738 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1739 } 1740 } 1741 #endif 1742 1743 /* fill request length and populate iovs */ 1744 req->length = length; 1745 1746 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1747 req->dif.orig_length = length; 1748 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1749 req->dif.elba_length = length; 1750 } 1751 1752 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1753 if (spdk_unlikely(rc < 0)) { 1754 if (rc == -EINVAL) { 1755 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1756 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1757 return -1; 1758 } 1759 /* No available buffers. Queue this request up. */ 1760 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1761 return 0; 1762 } 1763 1764 /* backward compatible */ 1765 req->data = req->iov[0].iov_base; 1766 1767 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1768 req->iovcnt); 1769 1770 return 0; 1771 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1772 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1773 uint64_t offset = sgl->address; 1774 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1775 1776 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1777 offset, sgl->unkeyed.length); 1778 1779 if (offset > max_len) { 1780 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1781 offset, max_len); 1782 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1783 return -1; 1784 } 1785 max_len -= (uint32_t)offset; 1786 1787 if (sgl->unkeyed.length > max_len) { 1788 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1789 sgl->unkeyed.length, max_len); 1790 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1791 return -1; 1792 } 1793 1794 rdma_req->num_outstanding_data_wr = 0; 1795 req->data = rdma_req->recv->buf + offset; 1796 req->data_from_pool = false; 1797 req->length = sgl->unkeyed.length; 1798 1799 req->iov[0].iov_base = req->data; 1800 req->iov[0].iov_len = req->length; 1801 req->iovcnt = 1; 1802 1803 return 0; 1804 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1805 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1806 1807 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1808 if (rc == -ENOMEM) { 1809 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1810 return 0; 1811 } else if (rc == -EINVAL) { 1812 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1813 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1814 return -1; 1815 } 1816 1817 /* backward compatible */ 1818 req->data = req->iov[0].iov_base; 1819 1820 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1821 req->iovcnt); 1822 1823 return 0; 1824 } 1825 1826 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1827 sgl->generic.type, sgl->generic.subtype); 1828 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1829 return -1; 1830 } 1831 1832 static void 1833 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1834 struct spdk_nvmf_rdma_transport *rtransport) 1835 { 1836 struct spdk_nvmf_rdma_qpair *rqpair; 1837 struct spdk_nvmf_rdma_poll_group *rgroup; 1838 1839 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1840 if (rdma_req->req.data_from_pool) { 1841 rgroup = rqpair->poller->group; 1842 1843 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1844 } 1845 nvmf_rdma_request_free_data(rdma_req, rtransport); 1846 rdma_req->req.length = 0; 1847 rdma_req->req.iovcnt = 0; 1848 rdma_req->req.data = NULL; 1849 rdma_req->rsp.wr.next = NULL; 1850 rdma_req->data.wr.next = NULL; 1851 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1852 rqpair->qd--; 1853 1854 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1855 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1856 } 1857 1858 bool 1859 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1860 struct spdk_nvmf_rdma_request *rdma_req) 1861 { 1862 struct spdk_nvmf_rdma_qpair *rqpair; 1863 struct spdk_nvmf_rdma_device *device; 1864 struct spdk_nvmf_rdma_poll_group *rgroup; 1865 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1866 int rc; 1867 struct spdk_nvmf_rdma_recv *rdma_recv; 1868 enum spdk_nvmf_rdma_request_state prev_state; 1869 bool progress = false; 1870 int data_posted; 1871 uint32_t num_blocks; 1872 1873 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1874 device = rqpair->device; 1875 rgroup = rqpair->poller->group; 1876 1877 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1878 1879 /* If the queue pair is in an error state, force the request to the completed state 1880 * to release resources. */ 1881 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1882 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1883 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1884 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1885 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1886 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1887 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1888 } 1889 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1890 } 1891 1892 /* The loop here is to allow for several back-to-back state changes. */ 1893 do { 1894 prev_state = rdma_req->state; 1895 1896 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 1897 1898 switch (rdma_req->state) { 1899 case RDMA_REQUEST_STATE_FREE: 1900 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 1901 * to escape this state. */ 1902 break; 1903 case RDMA_REQUEST_STATE_NEW: 1904 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 1905 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1906 rdma_recv = rdma_req->recv; 1907 1908 /* The first element of the SGL is the NVMe command */ 1909 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 1910 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 1911 1912 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1913 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1914 break; 1915 } 1916 1917 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 1918 rdma_req->req.dif.dif_insert_or_strip = true; 1919 } 1920 1921 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1922 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 1923 rdma_req->rsp.wr.imm_data = 0; 1924 #endif 1925 1926 /* The next state transition depends on the data transfer needs of this request. */ 1927 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 1928 1929 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 1930 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 1931 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 1932 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1933 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 1934 break; 1935 } 1936 1937 /* If no data to transfer, ready to execute. */ 1938 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 1939 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1940 break; 1941 } 1942 1943 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 1944 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 1945 break; 1946 case RDMA_REQUEST_STATE_NEED_BUFFER: 1947 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 1948 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1949 1950 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 1951 1952 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 1953 /* This request needs to wait in line to obtain a buffer */ 1954 break; 1955 } 1956 1957 /* Try to get a data buffer */ 1958 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 1959 if (rc < 0) { 1960 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 1961 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 1962 break; 1963 } 1964 1965 if (!rdma_req->req.data) { 1966 /* No buffers available. */ 1967 rgroup->stat.pending_data_buffer++; 1968 break; 1969 } 1970 1971 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 1972 1973 /* If data is transferring from host to controller and the data didn't 1974 * arrive using in capsule data, we need to do a transfer from the host. 1975 */ 1976 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 1977 rdma_req->req.data_from_pool) { 1978 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 1979 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 1980 break; 1981 } 1982 1983 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 1984 break; 1985 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 1986 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 1987 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 1988 1989 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 1990 /* This request needs to wait in line to perform RDMA */ 1991 break; 1992 } 1993 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 1994 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 1995 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 1996 rqpair->poller->stat.pending_rdma_read++; 1997 break; 1998 } 1999 2000 /* We have already verified that this request is the head of the queue. */ 2001 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2002 2003 rc = request_transfer_in(&rdma_req->req); 2004 if (!rc) { 2005 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2006 } else { 2007 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2008 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2009 } 2010 break; 2011 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2012 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2013 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2014 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2015 * to escape this state. */ 2016 break; 2017 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2018 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2019 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2020 2021 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2022 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2023 /* generate DIF for write operation */ 2024 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2025 assert(num_blocks > 0); 2026 2027 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2028 num_blocks, &rdma_req->req.dif.dif_ctx); 2029 if (rc != 0) { 2030 SPDK_ERRLOG("DIF generation failed\n"); 2031 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2032 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2033 break; 2034 } 2035 } 2036 2037 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2038 /* set extended length before IO operation */ 2039 rdma_req->req.length = rdma_req->req.dif.elba_length; 2040 } 2041 2042 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2043 spdk_nvmf_request_exec(&rdma_req->req); 2044 break; 2045 case RDMA_REQUEST_STATE_EXECUTING: 2046 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2047 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2048 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2049 * to escape this state. */ 2050 break; 2051 case RDMA_REQUEST_STATE_EXECUTED: 2052 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2053 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2054 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2055 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2056 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2057 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2058 } else { 2059 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2060 } 2061 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2062 /* restore the original length */ 2063 rdma_req->req.length = rdma_req->req.dif.orig_length; 2064 2065 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2066 struct spdk_dif_error error_blk; 2067 2068 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2069 2070 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2071 &rdma_req->req.dif.dif_ctx, &error_blk); 2072 if (rc) { 2073 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2074 2075 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2076 error_blk.err_offset); 2077 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2078 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2079 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2080 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2081 } 2082 } 2083 } 2084 break; 2085 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2086 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2087 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2088 2089 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2090 /* This request needs to wait in line to perform RDMA */ 2091 break; 2092 } 2093 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2094 rqpair->max_send_depth) { 2095 /* We can only have so many WRs outstanding. we have to wait until some finish. 2096 * +1 since each request has an additional wr in the resp. */ 2097 rqpair->poller->stat.pending_rdma_write++; 2098 break; 2099 } 2100 2101 /* We have already verified that this request is the head of the queue. */ 2102 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2103 2104 /* The data transfer will be kicked off from 2105 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2106 */ 2107 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2108 break; 2109 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2110 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2111 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2112 rc = request_transfer_out(&rdma_req->req, &data_posted); 2113 assert(rc == 0); /* No good way to handle this currently */ 2114 if (rc) { 2115 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2116 } else { 2117 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2118 RDMA_REQUEST_STATE_COMPLETING; 2119 } 2120 break; 2121 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2122 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2123 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2124 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2125 * to escape this state. */ 2126 break; 2127 case RDMA_REQUEST_STATE_COMPLETING: 2128 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2129 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2130 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2131 * to escape this state. */ 2132 break; 2133 case RDMA_REQUEST_STATE_COMPLETED: 2134 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2135 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2136 2137 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2138 _nvmf_rdma_request_free(rdma_req, rtransport); 2139 break; 2140 case RDMA_REQUEST_NUM_STATES: 2141 default: 2142 assert(0); 2143 break; 2144 } 2145 2146 if (rdma_req->state != prev_state) { 2147 progress = true; 2148 } 2149 } while (rdma_req->state != prev_state); 2150 2151 return progress; 2152 } 2153 2154 /* Public API callbacks begin here */ 2155 2156 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2157 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2158 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2159 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2160 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2161 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2162 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2163 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2164 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2165 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2166 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2167 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2168 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2169 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2170 2171 static void 2172 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2173 { 2174 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2175 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2176 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2177 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2178 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2179 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2180 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2181 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2182 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2183 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2184 opts->transport_specific = NULL; 2185 } 2186 2187 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2188 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2189 2190 static inline bool 2191 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2192 { 2193 return device->attr.vendor_id == NVMF_RXE_VENDOR_ID_OLD || 2194 device->attr.vendor_id == NVMF_RXE_VENDOR_ID_NEW; 2195 } 2196 2197 static struct spdk_nvmf_transport * 2198 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2199 { 2200 int rc; 2201 struct spdk_nvmf_rdma_transport *rtransport; 2202 struct spdk_nvmf_rdma_device *device, *tmp; 2203 struct ibv_context **contexts; 2204 uint32_t i; 2205 int flag; 2206 uint32_t sge_count; 2207 uint32_t min_shared_buffers; 2208 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2209 pthread_mutexattr_t attr; 2210 2211 rtransport = calloc(1, sizeof(*rtransport)); 2212 if (!rtransport) { 2213 return NULL; 2214 } 2215 2216 if (pthread_mutexattr_init(&attr)) { 2217 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2218 free(rtransport); 2219 return NULL; 2220 } 2221 2222 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2223 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2224 pthread_mutexattr_destroy(&attr); 2225 free(rtransport); 2226 return NULL; 2227 } 2228 2229 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2230 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2231 pthread_mutexattr_destroy(&attr); 2232 free(rtransport); 2233 return NULL; 2234 } 2235 2236 pthread_mutexattr_destroy(&attr); 2237 2238 TAILQ_INIT(&rtransport->devices); 2239 TAILQ_INIT(&rtransport->ports); 2240 TAILQ_INIT(&rtransport->poll_groups); 2241 2242 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2243 rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 2244 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2245 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2246 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2247 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2248 if (opts->transport_specific != NULL && 2249 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2250 SPDK_COUNTOF(rdma_transport_opts_decoder), 2251 &rtransport->rdma_opts)) { 2252 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2253 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2254 return NULL; 2255 } 2256 2257 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2258 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2259 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2260 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2261 " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d," 2262 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2263 opts->max_queue_depth, 2264 opts->max_io_size, 2265 opts->max_qpairs_per_ctrlr - 1, 2266 opts->io_unit_size, 2267 opts->in_capsule_data_size, 2268 opts->max_aq_depth, 2269 opts->num_shared_buffers, 2270 rtransport->rdma_opts.num_cqe, 2271 rtransport->rdma_opts.max_srq_depth, 2272 rtransport->rdma_opts.no_srq, 2273 rtransport->rdma_opts.acceptor_backlog, 2274 rtransport->rdma_opts.no_wr_batching, 2275 opts->abort_timeout_sec); 2276 2277 /* I/O unit size cannot be larger than max I/O size */ 2278 if (opts->io_unit_size > opts->max_io_size) { 2279 opts->io_unit_size = opts->max_io_size; 2280 } 2281 2282 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2283 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2284 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2285 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2286 } 2287 2288 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2289 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2290 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2291 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2292 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2293 return NULL; 2294 } 2295 2296 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2297 if (min_shared_buffers > opts->num_shared_buffers) { 2298 SPDK_ERRLOG("There are not enough buffers to satisfy" 2299 "per-poll group caches for each thread. (%" PRIu32 ")" 2300 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2301 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2302 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2303 return NULL; 2304 } 2305 2306 sge_count = opts->max_io_size / opts->io_unit_size; 2307 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2308 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2309 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2310 return NULL; 2311 } 2312 2313 rtransport->event_channel = rdma_create_event_channel(); 2314 if (rtransport->event_channel == NULL) { 2315 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2316 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2317 return NULL; 2318 } 2319 2320 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2321 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2322 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2323 rtransport->event_channel->fd, spdk_strerror(errno)); 2324 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2325 return NULL; 2326 } 2327 2328 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2329 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2330 sizeof(struct spdk_nvmf_rdma_request_data), 2331 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2332 SPDK_ENV_SOCKET_ID_ANY); 2333 if (!rtransport->data_wr_pool) { 2334 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2335 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2336 return NULL; 2337 } 2338 2339 contexts = rdma_get_devices(NULL); 2340 if (contexts == NULL) { 2341 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2342 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2343 return NULL; 2344 } 2345 2346 i = 0; 2347 rc = 0; 2348 while (contexts[i] != NULL) { 2349 device = calloc(1, sizeof(*device)); 2350 if (!device) { 2351 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2352 rc = -ENOMEM; 2353 break; 2354 } 2355 device->context = contexts[i]; 2356 rc = ibv_query_device(device->context, &device->attr); 2357 if (rc < 0) { 2358 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2359 free(device); 2360 break; 2361 2362 } 2363 2364 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2365 2366 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2367 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2368 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2369 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2370 } 2371 2372 /** 2373 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2374 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2375 * but incorrectly reports that it does. There are changes making their way 2376 * through the kernel now that will enable this feature. When they are merged, 2377 * we can conditionally enable this feature. 2378 * 2379 * TODO: enable this for versions of the kernel rxe driver that support it. 2380 */ 2381 if (nvmf_rdma_is_rxe_device(device)) { 2382 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2383 } 2384 #endif 2385 2386 /* set up device context async ev fd as NON_BLOCKING */ 2387 flag = fcntl(device->context->async_fd, F_GETFL); 2388 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2389 if (rc < 0) { 2390 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2391 free(device); 2392 break; 2393 } 2394 2395 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2396 i++; 2397 2398 if (g_nvmf_hooks.get_ibv_pd) { 2399 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2400 } else { 2401 device->pd = ibv_alloc_pd(device->context); 2402 } 2403 2404 if (!device->pd) { 2405 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2406 rc = -ENOMEM; 2407 break; 2408 } 2409 2410 assert(device->map == NULL); 2411 2412 device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks); 2413 if (!device->map) { 2414 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2415 rc = -ENOMEM; 2416 break; 2417 } 2418 2419 assert(device->map != NULL); 2420 assert(device->pd != NULL); 2421 } 2422 rdma_free_devices(contexts); 2423 2424 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2425 /* divide and round up. */ 2426 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2427 2428 /* round up to the nearest 4k. */ 2429 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2430 2431 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2432 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2433 opts->io_unit_size); 2434 } 2435 2436 if (rc < 0) { 2437 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2438 return NULL; 2439 } 2440 2441 /* Set up poll descriptor array to monitor events from RDMA and IB 2442 * in a single poll syscall 2443 */ 2444 rtransport->npoll_fds = i + 1; 2445 i = 0; 2446 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2447 if (rtransport->poll_fds == NULL) { 2448 SPDK_ERRLOG("poll_fds allocation failed\n"); 2449 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2450 return NULL; 2451 } 2452 2453 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2454 rtransport->poll_fds[i++].events = POLLIN; 2455 2456 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2457 rtransport->poll_fds[i].fd = device->context->async_fd; 2458 rtransport->poll_fds[i++].events = POLLIN; 2459 } 2460 2461 return &rtransport->transport; 2462 } 2463 2464 static void 2465 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2466 { 2467 struct spdk_nvmf_rdma_transport *rtransport; 2468 assert(w != NULL); 2469 2470 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2471 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2472 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2473 if (rtransport->rdma_opts.no_srq == true) { 2474 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe); 2475 } 2476 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2477 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2478 } 2479 2480 static int 2481 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2482 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2483 { 2484 struct spdk_nvmf_rdma_transport *rtransport; 2485 struct spdk_nvmf_rdma_port *port, *port_tmp; 2486 struct spdk_nvmf_rdma_device *device, *device_tmp; 2487 2488 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2489 2490 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2491 TAILQ_REMOVE(&rtransport->ports, port, link); 2492 rdma_destroy_id(port->id); 2493 free(port); 2494 } 2495 2496 if (rtransport->poll_fds != NULL) { 2497 free(rtransport->poll_fds); 2498 } 2499 2500 if (rtransport->event_channel != NULL) { 2501 rdma_destroy_event_channel(rtransport->event_channel); 2502 } 2503 2504 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2505 TAILQ_REMOVE(&rtransport->devices, device, link); 2506 spdk_rdma_free_mem_map(&device->map); 2507 if (device->pd) { 2508 if (!g_nvmf_hooks.get_ibv_pd) { 2509 ibv_dealloc_pd(device->pd); 2510 } 2511 } 2512 free(device); 2513 } 2514 2515 if (rtransport->data_wr_pool != NULL) { 2516 if (spdk_mempool_count(rtransport->data_wr_pool) != 2517 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2518 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2519 spdk_mempool_count(rtransport->data_wr_pool), 2520 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2521 } 2522 } 2523 2524 spdk_mempool_free(rtransport->data_wr_pool); 2525 2526 pthread_mutex_destroy(&rtransport->lock); 2527 free(rtransport); 2528 2529 if (cb_fn) { 2530 cb_fn(cb_arg); 2531 } 2532 return 0; 2533 } 2534 2535 static int 2536 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2537 struct spdk_nvme_transport_id *trid, 2538 bool peer); 2539 2540 static int 2541 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 2542 struct spdk_nvmf_listen_opts *listen_opts) 2543 { 2544 struct spdk_nvmf_rdma_transport *rtransport; 2545 struct spdk_nvmf_rdma_device *device; 2546 struct spdk_nvmf_rdma_port *port; 2547 struct addrinfo *res; 2548 struct addrinfo hints; 2549 int family; 2550 int rc; 2551 2552 if (!strlen(trid->trsvcid)) { 2553 SPDK_ERRLOG("Service id is required\n"); 2554 return -EINVAL; 2555 } 2556 2557 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2558 assert(rtransport->event_channel != NULL); 2559 2560 pthread_mutex_lock(&rtransport->lock); 2561 port = calloc(1, sizeof(*port)); 2562 if (!port) { 2563 SPDK_ERRLOG("Port allocation failed\n"); 2564 pthread_mutex_unlock(&rtransport->lock); 2565 return -ENOMEM; 2566 } 2567 2568 port->trid = trid; 2569 2570 switch (trid->adrfam) { 2571 case SPDK_NVMF_ADRFAM_IPV4: 2572 family = AF_INET; 2573 break; 2574 case SPDK_NVMF_ADRFAM_IPV6: 2575 family = AF_INET6; 2576 break; 2577 default: 2578 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2579 free(port); 2580 pthread_mutex_unlock(&rtransport->lock); 2581 return -EINVAL; 2582 } 2583 2584 memset(&hints, 0, sizeof(hints)); 2585 hints.ai_family = family; 2586 hints.ai_flags = AI_NUMERICSERV; 2587 hints.ai_socktype = SOCK_STREAM; 2588 hints.ai_protocol = 0; 2589 2590 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2591 if (rc) { 2592 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2593 free(port); 2594 pthread_mutex_unlock(&rtransport->lock); 2595 return -EINVAL; 2596 } 2597 2598 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2599 if (rc < 0) { 2600 SPDK_ERRLOG("rdma_create_id() failed\n"); 2601 freeaddrinfo(res); 2602 free(port); 2603 pthread_mutex_unlock(&rtransport->lock); 2604 return rc; 2605 } 2606 2607 rc = rdma_bind_addr(port->id, res->ai_addr); 2608 freeaddrinfo(res); 2609 2610 if (rc < 0) { 2611 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2612 rdma_destroy_id(port->id); 2613 free(port); 2614 pthread_mutex_unlock(&rtransport->lock); 2615 return rc; 2616 } 2617 2618 if (!port->id->verbs) { 2619 SPDK_ERRLOG("ibv_context is null\n"); 2620 rdma_destroy_id(port->id); 2621 free(port); 2622 pthread_mutex_unlock(&rtransport->lock); 2623 return -1; 2624 } 2625 2626 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2627 if (rc < 0) { 2628 SPDK_ERRLOG("rdma_listen() failed\n"); 2629 rdma_destroy_id(port->id); 2630 free(port); 2631 pthread_mutex_unlock(&rtransport->lock); 2632 return rc; 2633 } 2634 2635 TAILQ_FOREACH(device, &rtransport->devices, link) { 2636 if (device->context == port->id->verbs) { 2637 port->device = device; 2638 break; 2639 } 2640 } 2641 if (!port->device) { 2642 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2643 port->id->verbs); 2644 rdma_destroy_id(port->id); 2645 free(port); 2646 pthread_mutex_unlock(&rtransport->lock); 2647 return -EINVAL; 2648 } 2649 2650 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2651 trid->traddr, trid->trsvcid); 2652 2653 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2654 pthread_mutex_unlock(&rtransport->lock); 2655 return 0; 2656 } 2657 2658 static void 2659 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2660 const struct spdk_nvme_transport_id *trid) 2661 { 2662 struct spdk_nvmf_rdma_transport *rtransport; 2663 struct spdk_nvmf_rdma_port *port, *tmp; 2664 2665 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2666 2667 pthread_mutex_lock(&rtransport->lock); 2668 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2669 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2670 TAILQ_REMOVE(&rtransport->ports, port, link); 2671 rdma_destroy_id(port->id); 2672 free(port); 2673 break; 2674 } 2675 } 2676 2677 pthread_mutex_unlock(&rtransport->lock); 2678 } 2679 2680 static void 2681 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2682 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2683 { 2684 struct spdk_nvmf_request *req, *tmp; 2685 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2686 struct spdk_nvmf_rdma_resources *resources; 2687 2688 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2689 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2690 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2691 break; 2692 } 2693 } 2694 2695 /* Then RDMA writes since reads have stronger restrictions than writes */ 2696 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2697 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2698 break; 2699 } 2700 } 2701 2702 /* Then we handle request waiting on memory buffers. */ 2703 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2704 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2705 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2706 break; 2707 } 2708 } 2709 2710 resources = rqpair->resources; 2711 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2712 rdma_req = STAILQ_FIRST(&resources->free_queue); 2713 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2714 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2715 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2716 2717 if (rqpair->srq != NULL) { 2718 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2719 rdma_req->recv->qpair->qd++; 2720 } else { 2721 rqpair->qd++; 2722 } 2723 2724 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2725 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2726 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2727 break; 2728 } 2729 } 2730 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2731 rqpair->poller->stat.pending_free_request++; 2732 } 2733 } 2734 2735 static inline bool 2736 nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device) 2737 { 2738 /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */ 2739 return nvmf_rdma_is_rxe_device(device) || 2740 device->context->device->transport_type == IBV_TRANSPORT_IWARP; 2741 } 2742 2743 static void 2744 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2745 { 2746 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2747 struct spdk_nvmf_rdma_transport, transport); 2748 2749 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2750 2751 /* nvmr_rdma_close_qpair is not called */ 2752 if (!rqpair->to_close) { 2753 return; 2754 } 2755 2756 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2757 if (rqpair->current_send_depth != 0) { 2758 return; 2759 } 2760 2761 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2762 return; 2763 } 2764 2765 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 2766 !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) { 2767 return; 2768 } 2769 2770 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2771 2772 nvmf_rdma_qpair_destroy(rqpair); 2773 } 2774 2775 static int 2776 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2777 { 2778 struct spdk_nvmf_qpair *qpair; 2779 struct spdk_nvmf_rdma_qpair *rqpair; 2780 2781 if (evt->id == NULL) { 2782 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2783 return -1; 2784 } 2785 2786 qpair = evt->id->context; 2787 if (qpair == NULL) { 2788 SPDK_ERRLOG("disconnect request: no active connection\n"); 2789 return -1; 2790 } 2791 2792 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2793 2794 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2795 2796 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2797 2798 return 0; 2799 } 2800 2801 #ifdef DEBUG 2802 static const char *CM_EVENT_STR[] = { 2803 "RDMA_CM_EVENT_ADDR_RESOLVED", 2804 "RDMA_CM_EVENT_ADDR_ERROR", 2805 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2806 "RDMA_CM_EVENT_ROUTE_ERROR", 2807 "RDMA_CM_EVENT_CONNECT_REQUEST", 2808 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2809 "RDMA_CM_EVENT_CONNECT_ERROR", 2810 "RDMA_CM_EVENT_UNREACHABLE", 2811 "RDMA_CM_EVENT_REJECTED", 2812 "RDMA_CM_EVENT_ESTABLISHED", 2813 "RDMA_CM_EVENT_DISCONNECTED", 2814 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2815 "RDMA_CM_EVENT_MULTICAST_JOIN", 2816 "RDMA_CM_EVENT_MULTICAST_ERROR", 2817 "RDMA_CM_EVENT_ADDR_CHANGE", 2818 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2819 }; 2820 #endif /* DEBUG */ 2821 2822 static void 2823 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2824 struct spdk_nvmf_rdma_port *port) 2825 { 2826 struct spdk_nvmf_rdma_poll_group *rgroup; 2827 struct spdk_nvmf_rdma_poller *rpoller; 2828 struct spdk_nvmf_rdma_qpair *rqpair; 2829 2830 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2831 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2832 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 2833 if (rqpair->listen_id == port->id) { 2834 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2835 } 2836 } 2837 } 2838 } 2839 } 2840 2841 static bool 2842 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2843 struct rdma_cm_event *event) 2844 { 2845 const struct spdk_nvme_transport_id *trid; 2846 struct spdk_nvmf_rdma_port *port; 2847 struct spdk_nvmf_rdma_transport *rtransport; 2848 bool event_acked = false; 2849 2850 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2851 TAILQ_FOREACH(port, &rtransport->ports, link) { 2852 if (port->id == event->id) { 2853 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2854 rdma_ack_cm_event(event); 2855 event_acked = true; 2856 trid = port->trid; 2857 break; 2858 } 2859 } 2860 2861 if (event_acked) { 2862 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2863 2864 nvmf_rdma_stop_listen(transport, trid); 2865 nvmf_rdma_listen(transport, trid, NULL); 2866 } 2867 2868 return event_acked; 2869 } 2870 2871 static void 2872 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2873 struct rdma_cm_event *event) 2874 { 2875 struct spdk_nvmf_rdma_port *port; 2876 struct spdk_nvmf_rdma_transport *rtransport; 2877 2878 port = event->id->context; 2879 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2880 2881 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2882 2883 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2884 2885 rdma_ack_cm_event(event); 2886 2887 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2888 ; 2889 } 2890 } 2891 2892 static void 2893 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2894 { 2895 struct spdk_nvmf_rdma_transport *rtransport; 2896 struct rdma_cm_event *event; 2897 int rc; 2898 bool event_acked; 2899 2900 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2901 2902 if (rtransport->event_channel == NULL) { 2903 return; 2904 } 2905 2906 while (1) { 2907 event_acked = false; 2908 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2909 if (rc) { 2910 if (errno != EAGAIN && errno != EWOULDBLOCK) { 2911 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 2912 } 2913 break; 2914 } 2915 2916 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2917 2918 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2919 2920 switch (event->event) { 2921 case RDMA_CM_EVENT_ADDR_RESOLVED: 2922 case RDMA_CM_EVENT_ADDR_ERROR: 2923 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2924 case RDMA_CM_EVENT_ROUTE_ERROR: 2925 /* No action required. The target never attempts to resolve routes. */ 2926 break; 2927 case RDMA_CM_EVENT_CONNECT_REQUEST: 2928 rc = nvmf_rdma_connect(transport, event); 2929 if (rc < 0) { 2930 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2931 break; 2932 } 2933 break; 2934 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2935 /* The target never initiates a new connection. So this will not occur. */ 2936 break; 2937 case RDMA_CM_EVENT_CONNECT_ERROR: 2938 /* Can this happen? The docs say it can, but not sure what causes it. */ 2939 break; 2940 case RDMA_CM_EVENT_UNREACHABLE: 2941 case RDMA_CM_EVENT_REJECTED: 2942 /* These only occur on the client side. */ 2943 break; 2944 case RDMA_CM_EVENT_ESTABLISHED: 2945 /* TODO: Should we be waiting for this event anywhere? */ 2946 break; 2947 case RDMA_CM_EVENT_DISCONNECTED: 2948 rc = nvmf_rdma_disconnect(event); 2949 if (rc < 0) { 2950 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2951 break; 2952 } 2953 break; 2954 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2955 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 2956 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 2957 * Once these events are sent to SPDK, we should release all IB resources and 2958 * don't make attempts to call any ibv_query/modify/create functions. We can only call 2959 * ibv_destory* functions to release user space memory allocated by IB. All kernel 2960 * resources are already cleaned. */ 2961 if (event->id->qp) { 2962 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 2963 * corresponding qpair. Otherwise the event refers to a listening device */ 2964 rc = nvmf_rdma_disconnect(event); 2965 if (rc < 0) { 2966 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 2967 break; 2968 } 2969 } else { 2970 nvmf_rdma_handle_cm_event_port_removal(transport, event); 2971 event_acked = true; 2972 } 2973 break; 2974 case RDMA_CM_EVENT_MULTICAST_JOIN: 2975 case RDMA_CM_EVENT_MULTICAST_ERROR: 2976 /* Multicast is not used */ 2977 break; 2978 case RDMA_CM_EVENT_ADDR_CHANGE: 2979 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 2980 break; 2981 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2982 /* For now, do nothing. The target never re-uses queue pairs. */ 2983 break; 2984 default: 2985 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 2986 break; 2987 } 2988 if (!event_acked) { 2989 rdma_ack_cm_event(event); 2990 } 2991 } 2992 } 2993 2994 static void 2995 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 2996 { 2997 rqpair->last_wqe_reached = true; 2998 nvmf_rdma_destroy_drained_qpair(rqpair); 2999 } 3000 3001 static void 3002 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3003 { 3004 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3005 3006 if (event_ctx->rqpair) { 3007 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3008 if (event_ctx->cb_fn) { 3009 event_ctx->cb_fn(event_ctx->rqpair); 3010 } 3011 } 3012 free(event_ctx); 3013 } 3014 3015 static int 3016 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3017 spdk_nvmf_rdma_qpair_ibv_event fn) 3018 { 3019 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3020 struct spdk_thread *thr = NULL; 3021 int rc; 3022 3023 if (rqpair->qpair.group) { 3024 thr = rqpair->qpair.group->thread; 3025 } else if (rqpair->destruct_channel) { 3026 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3027 } 3028 3029 if (!thr) { 3030 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3031 return -EINVAL; 3032 } 3033 3034 ctx = calloc(1, sizeof(*ctx)); 3035 if (!ctx) { 3036 return -ENOMEM; 3037 } 3038 3039 ctx->rqpair = rqpair; 3040 ctx->cb_fn = fn; 3041 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3042 3043 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3044 if (rc) { 3045 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3046 free(ctx); 3047 } 3048 3049 return rc; 3050 } 3051 3052 static int 3053 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3054 { 3055 int rc; 3056 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3057 struct ibv_async_event event; 3058 3059 rc = ibv_get_async_event(device->context, &event); 3060 3061 if (rc) { 3062 /* In non-blocking mode -1 means there are no events available */ 3063 return rc; 3064 } 3065 3066 switch (event.event_type) { 3067 case IBV_EVENT_QP_FATAL: 3068 rqpair = event.element.qp->qp_context; 3069 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3070 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3071 (uintptr_t)rqpair->cm_id, event.event_type); 3072 nvmf_rdma_update_ibv_state(rqpair); 3073 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3074 break; 3075 case IBV_EVENT_QP_LAST_WQE_REACHED: 3076 /* This event only occurs for shared receive queues. */ 3077 rqpair = event.element.qp->qp_context; 3078 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3079 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3080 if (rc) { 3081 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3082 rqpair->last_wqe_reached = true; 3083 } 3084 break; 3085 case IBV_EVENT_SQ_DRAINED: 3086 /* This event occurs frequently in both error and non-error states. 3087 * Check if the qpair is in an error state before sending a message. */ 3088 rqpair = event.element.qp->qp_context; 3089 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3090 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3091 (uintptr_t)rqpair->cm_id, event.event_type); 3092 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3093 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3094 } 3095 break; 3096 case IBV_EVENT_QP_REQ_ERR: 3097 case IBV_EVENT_QP_ACCESS_ERR: 3098 case IBV_EVENT_COMM_EST: 3099 case IBV_EVENT_PATH_MIG: 3100 case IBV_EVENT_PATH_MIG_ERR: 3101 SPDK_NOTICELOG("Async event: %s\n", 3102 ibv_event_type_str(event.event_type)); 3103 rqpair = event.element.qp->qp_context; 3104 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3105 (uintptr_t)rqpair->cm_id, event.event_type); 3106 nvmf_rdma_update_ibv_state(rqpair); 3107 break; 3108 case IBV_EVENT_CQ_ERR: 3109 case IBV_EVENT_DEVICE_FATAL: 3110 case IBV_EVENT_PORT_ACTIVE: 3111 case IBV_EVENT_PORT_ERR: 3112 case IBV_EVENT_LID_CHANGE: 3113 case IBV_EVENT_PKEY_CHANGE: 3114 case IBV_EVENT_SM_CHANGE: 3115 case IBV_EVENT_SRQ_ERR: 3116 case IBV_EVENT_SRQ_LIMIT_REACHED: 3117 case IBV_EVENT_CLIENT_REREGISTER: 3118 case IBV_EVENT_GID_CHANGE: 3119 default: 3120 SPDK_NOTICELOG("Async event: %s\n", 3121 ibv_event_type_str(event.event_type)); 3122 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3123 break; 3124 } 3125 ibv_ack_async_event(&event); 3126 3127 return 0; 3128 } 3129 3130 static void 3131 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3132 { 3133 int rc = 0; 3134 uint32_t i = 0; 3135 3136 for (i = 0; i < max_events; i++) { 3137 rc = nvmf_process_ib_event(device); 3138 if (rc) { 3139 break; 3140 } 3141 } 3142 3143 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3144 } 3145 3146 static uint32_t 3147 nvmf_rdma_accept(struct spdk_nvmf_transport *transport) 3148 { 3149 int nfds, i = 0; 3150 struct spdk_nvmf_rdma_transport *rtransport; 3151 struct spdk_nvmf_rdma_device *device, *tmp; 3152 uint32_t count; 3153 3154 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3155 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3156 3157 if (nfds <= 0) { 3158 return 0; 3159 } 3160 3161 /* The first poll descriptor is RDMA CM event */ 3162 if (rtransport->poll_fds[i++].revents & POLLIN) { 3163 nvmf_process_cm_event(transport); 3164 nfds--; 3165 } 3166 3167 if (nfds == 0) { 3168 return count; 3169 } 3170 3171 /* Second and subsequent poll descriptors are IB async events */ 3172 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3173 if (rtransport->poll_fds[i++].revents & POLLIN) { 3174 nvmf_process_ib_events(device, 32); 3175 nfds--; 3176 } 3177 } 3178 /* check all flagged fd's have been served */ 3179 assert(nfds == 0); 3180 3181 return count; 3182 } 3183 3184 static void 3185 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3186 struct spdk_nvmf_ctrlr_data *cdata) 3187 { 3188 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3189 3190 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3191 since in-capsule data only works with NVME drives that support SGL memory layout */ 3192 if (transport->opts.dif_insert_or_strip) { 3193 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3194 } 3195 } 3196 3197 static void 3198 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3199 struct spdk_nvme_transport_id *trid, 3200 struct spdk_nvmf_discovery_log_page_entry *entry) 3201 { 3202 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3203 entry->adrfam = trid->adrfam; 3204 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3205 3206 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3207 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3208 3209 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3210 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3211 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3212 } 3213 3214 static void 3215 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3216 3217 static struct spdk_nvmf_transport_poll_group * 3218 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3219 { 3220 struct spdk_nvmf_rdma_transport *rtransport; 3221 struct spdk_nvmf_rdma_poll_group *rgroup; 3222 struct spdk_nvmf_rdma_poller *poller; 3223 struct spdk_nvmf_rdma_device *device; 3224 struct spdk_rdma_srq_init_attr srq_init_attr; 3225 struct spdk_nvmf_rdma_resource_opts opts; 3226 int num_cqe; 3227 3228 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3229 3230 rgroup = calloc(1, sizeof(*rgroup)); 3231 if (!rgroup) { 3232 return NULL; 3233 } 3234 3235 TAILQ_INIT(&rgroup->pollers); 3236 3237 pthread_mutex_lock(&rtransport->lock); 3238 TAILQ_FOREACH(device, &rtransport->devices, link) { 3239 poller = calloc(1, sizeof(*poller)); 3240 if (!poller) { 3241 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3242 nvmf_rdma_poll_group_destroy(&rgroup->group); 3243 pthread_mutex_unlock(&rtransport->lock); 3244 return NULL; 3245 } 3246 3247 poller->device = device; 3248 poller->group = rgroup; 3249 3250 TAILQ_INIT(&poller->qpairs); 3251 STAILQ_INIT(&poller->qpairs_pending_send); 3252 STAILQ_INIT(&poller->qpairs_pending_recv); 3253 3254 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3255 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3256 poller->max_srq_depth = rtransport->rdma_opts.max_srq_depth; 3257 3258 device->num_srq++; 3259 memset(&srq_init_attr, 0, sizeof(srq_init_attr)); 3260 srq_init_attr.pd = device->pd; 3261 srq_init_attr.stats = &poller->stat.qp_stats.recv; 3262 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth; 3263 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3264 poller->srq = spdk_rdma_srq_create(&srq_init_attr); 3265 if (!poller->srq) { 3266 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3267 nvmf_rdma_poll_group_destroy(&rgroup->group); 3268 pthread_mutex_unlock(&rtransport->lock); 3269 return NULL; 3270 } 3271 3272 opts.qp = poller->srq; 3273 opts.pd = device->pd; 3274 opts.qpair = NULL; 3275 opts.shared = true; 3276 opts.max_queue_depth = poller->max_srq_depth; 3277 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3278 3279 poller->resources = nvmf_rdma_resources_create(&opts); 3280 if (!poller->resources) { 3281 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3282 nvmf_rdma_poll_group_destroy(&rgroup->group); 3283 pthread_mutex_unlock(&rtransport->lock); 3284 return NULL; 3285 } 3286 } 3287 3288 /* 3289 * When using an srq, we can limit the completion queue at startup. 3290 * The following formula represents the calculation: 3291 * num_cqe = num_recv + num_data_wr + num_send_wr. 3292 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3293 */ 3294 if (poller->srq) { 3295 num_cqe = poller->max_srq_depth * 3; 3296 } else { 3297 num_cqe = rtransport->rdma_opts.num_cqe; 3298 } 3299 3300 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3301 if (!poller->cq) { 3302 SPDK_ERRLOG("Unable to create completion queue\n"); 3303 nvmf_rdma_poll_group_destroy(&rgroup->group); 3304 pthread_mutex_unlock(&rtransport->lock); 3305 return NULL; 3306 } 3307 poller->num_cqe = num_cqe; 3308 } 3309 3310 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3311 if (rtransport->conn_sched.next_admin_pg == NULL) { 3312 rtransport->conn_sched.next_admin_pg = rgroup; 3313 rtransport->conn_sched.next_io_pg = rgroup; 3314 } 3315 3316 pthread_mutex_unlock(&rtransport->lock); 3317 return &rgroup->group; 3318 } 3319 3320 static struct spdk_nvmf_transport_poll_group * 3321 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3322 { 3323 struct spdk_nvmf_rdma_transport *rtransport; 3324 struct spdk_nvmf_rdma_poll_group **pg; 3325 struct spdk_nvmf_transport_poll_group *result; 3326 3327 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3328 3329 pthread_mutex_lock(&rtransport->lock); 3330 3331 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3332 pthread_mutex_unlock(&rtransport->lock); 3333 return NULL; 3334 } 3335 3336 if (qpair->qid == 0) { 3337 pg = &rtransport->conn_sched.next_admin_pg; 3338 } else { 3339 pg = &rtransport->conn_sched.next_io_pg; 3340 } 3341 3342 assert(*pg != NULL); 3343 3344 result = &(*pg)->group; 3345 3346 *pg = TAILQ_NEXT(*pg, link); 3347 if (*pg == NULL) { 3348 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3349 } 3350 3351 pthread_mutex_unlock(&rtransport->lock); 3352 3353 return result; 3354 } 3355 3356 static void 3357 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3358 { 3359 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3360 struct spdk_nvmf_rdma_poller *poller, *tmp; 3361 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3362 struct spdk_nvmf_rdma_transport *rtransport; 3363 3364 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3365 if (!rgroup) { 3366 return; 3367 } 3368 3369 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3370 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3371 3372 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3373 nvmf_rdma_qpair_destroy(qpair); 3374 } 3375 3376 if (poller->srq) { 3377 if (poller->resources) { 3378 nvmf_rdma_resources_destroy(poller->resources); 3379 } 3380 spdk_rdma_srq_destroy(poller->srq); 3381 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 3382 } 3383 3384 if (poller->cq) { 3385 ibv_destroy_cq(poller->cq); 3386 } 3387 3388 free(poller); 3389 } 3390 3391 if (rgroup->group.transport == NULL) { 3392 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3393 * calls this function directly in a failure path. */ 3394 free(rgroup); 3395 return; 3396 } 3397 3398 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3399 3400 pthread_mutex_lock(&rtransport->lock); 3401 next_rgroup = TAILQ_NEXT(rgroup, link); 3402 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3403 if (next_rgroup == NULL) { 3404 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3405 } 3406 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3407 rtransport->conn_sched.next_admin_pg = next_rgroup; 3408 } 3409 if (rtransport->conn_sched.next_io_pg == rgroup) { 3410 rtransport->conn_sched.next_io_pg = next_rgroup; 3411 } 3412 pthread_mutex_unlock(&rtransport->lock); 3413 3414 free(rgroup); 3415 } 3416 3417 static void 3418 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3419 { 3420 if (rqpair->cm_id != NULL) { 3421 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3422 } 3423 } 3424 3425 static int 3426 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3427 struct spdk_nvmf_qpair *qpair) 3428 { 3429 struct spdk_nvmf_rdma_poll_group *rgroup; 3430 struct spdk_nvmf_rdma_qpair *rqpair; 3431 struct spdk_nvmf_rdma_device *device; 3432 struct spdk_nvmf_rdma_poller *poller; 3433 int rc; 3434 3435 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3436 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3437 3438 device = rqpair->device; 3439 3440 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3441 if (poller->device == device) { 3442 break; 3443 } 3444 } 3445 3446 if (!poller) { 3447 SPDK_ERRLOG("No poller found for device.\n"); 3448 return -1; 3449 } 3450 3451 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3452 rqpair->poller = poller; 3453 rqpair->srq = rqpair->poller->srq; 3454 3455 rc = nvmf_rdma_qpair_initialize(qpair); 3456 if (rc < 0) { 3457 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3458 return -1; 3459 } 3460 3461 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3462 if (rc) { 3463 /* Try to reject, but we probably can't */ 3464 nvmf_rdma_qpair_reject_connection(rqpair); 3465 return -1; 3466 } 3467 3468 nvmf_rdma_update_ibv_state(rqpair); 3469 3470 return 0; 3471 } 3472 3473 static int 3474 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3475 struct spdk_nvmf_qpair *qpair) 3476 { 3477 struct spdk_nvmf_rdma_qpair *rqpair; 3478 3479 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3480 assert(group->transport->tgt != NULL); 3481 3482 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3483 3484 if (!rqpair->destruct_channel) { 3485 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3486 return 0; 3487 } 3488 3489 /* Sanity check that we get io_channel on the correct thread */ 3490 if (qpair->group) { 3491 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3492 } 3493 3494 return 0; 3495 } 3496 3497 static int 3498 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3499 { 3500 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3501 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3502 struct spdk_nvmf_rdma_transport, transport); 3503 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3504 struct spdk_nvmf_rdma_qpair, qpair); 3505 3506 /* 3507 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3508 * needs to be returned to the shared receive queue or the poll group will eventually be 3509 * starved of RECV structures. 3510 */ 3511 if (rqpair->srq && rdma_req->recv) { 3512 int rc; 3513 struct ibv_recv_wr *bad_recv_wr; 3514 3515 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr); 3516 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr); 3517 if (rc) { 3518 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3519 } 3520 } 3521 3522 _nvmf_rdma_request_free(rdma_req, rtransport); 3523 return 0; 3524 } 3525 3526 static int 3527 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3528 { 3529 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3530 struct spdk_nvmf_rdma_transport, transport); 3531 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3532 struct spdk_nvmf_rdma_request, req); 3533 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3534 struct spdk_nvmf_rdma_qpair, qpair); 3535 3536 if (rqpair->ibv_state != IBV_QPS_ERR) { 3537 /* The connection is alive, so process the request as normal */ 3538 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3539 } else { 3540 /* The connection is dead. Move the request directly to the completed state. */ 3541 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3542 } 3543 3544 nvmf_rdma_request_process(rtransport, rdma_req); 3545 3546 return 0; 3547 } 3548 3549 static void 3550 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 3551 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3552 { 3553 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3554 3555 rqpair->to_close = true; 3556 3557 /* This happens only when the qpair is disconnected before 3558 * it is added to the poll group. Since there is no poll group, 3559 * the RDMA qp has not been initialized yet and the RDMA CM 3560 * event has not yet been acknowledged, so we need to reject it. 3561 */ 3562 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3563 nvmf_rdma_qpair_reject_connection(rqpair); 3564 nvmf_rdma_qpair_destroy(rqpair); 3565 return; 3566 } 3567 3568 if (rqpair->rdma_qp) { 3569 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3570 } 3571 3572 nvmf_rdma_destroy_drained_qpair(rqpair); 3573 3574 if (cb_fn) { 3575 cb_fn(cb_arg); 3576 } 3577 } 3578 3579 static struct spdk_nvmf_rdma_qpair * 3580 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3581 { 3582 struct spdk_nvmf_rdma_qpair *rqpair; 3583 /* @todo: improve QP search */ 3584 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3585 if (wc->qp_num == rqpair->rdma_qp->qp->qp_num) { 3586 return rqpair; 3587 } 3588 } 3589 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3590 return NULL; 3591 } 3592 3593 #ifdef DEBUG 3594 static int 3595 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3596 { 3597 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3598 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3599 } 3600 #endif 3601 3602 static void 3603 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3604 int rc) 3605 { 3606 struct spdk_nvmf_rdma_recv *rdma_recv; 3607 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3608 3609 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3610 while (bad_recv_wr != NULL) { 3611 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3612 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3613 3614 rdma_recv->qpair->current_recv_depth++; 3615 bad_recv_wr = bad_recv_wr->next; 3616 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3617 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3618 } 3619 } 3620 3621 static void 3622 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3623 { 3624 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3625 while (bad_recv_wr != NULL) { 3626 bad_recv_wr = bad_recv_wr->next; 3627 rqpair->current_recv_depth++; 3628 } 3629 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3630 } 3631 3632 static void 3633 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3634 struct spdk_nvmf_rdma_poller *rpoller) 3635 { 3636 struct spdk_nvmf_rdma_qpair *rqpair; 3637 struct ibv_recv_wr *bad_recv_wr; 3638 int rc; 3639 3640 if (rpoller->srq) { 3641 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr); 3642 if (rc) { 3643 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3644 } 3645 } else { 3646 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3647 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3648 rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr); 3649 if (rc) { 3650 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3651 } 3652 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3653 } 3654 } 3655 } 3656 3657 static void 3658 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3659 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3660 { 3661 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3662 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3663 3664 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3665 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3666 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3667 assert(rqpair->current_send_depth > 0); 3668 rqpair->current_send_depth--; 3669 switch (bad_rdma_wr->type) { 3670 case RDMA_WR_TYPE_DATA: 3671 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3672 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3673 assert(rqpair->current_read_depth > 0); 3674 rqpair->current_read_depth--; 3675 } 3676 break; 3677 case RDMA_WR_TYPE_SEND: 3678 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3679 break; 3680 default: 3681 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3682 prev_rdma_req = cur_rdma_req; 3683 continue; 3684 } 3685 3686 if (prev_rdma_req == cur_rdma_req) { 3687 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3688 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3689 continue; 3690 } 3691 3692 switch (cur_rdma_req->state) { 3693 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3694 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3695 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3696 break; 3697 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3698 case RDMA_REQUEST_STATE_COMPLETING: 3699 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3700 break; 3701 default: 3702 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3703 cur_rdma_req->state, rqpair); 3704 continue; 3705 } 3706 3707 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3708 prev_rdma_req = cur_rdma_req; 3709 } 3710 3711 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3712 /* Disconnect the connection. */ 3713 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3714 } 3715 3716 } 3717 3718 static void 3719 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3720 struct spdk_nvmf_rdma_poller *rpoller) 3721 { 3722 struct spdk_nvmf_rdma_qpair *rqpair; 3723 struct ibv_send_wr *bad_wr = NULL; 3724 int rc; 3725 3726 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3727 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3728 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3729 3730 /* bad wr always points to the first wr that failed. */ 3731 if (rc) { 3732 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3733 } 3734 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3735 } 3736 } 3737 3738 static const char * 3739 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 3740 { 3741 switch (wr_type) { 3742 case RDMA_WR_TYPE_RECV: 3743 return "RECV"; 3744 case RDMA_WR_TYPE_SEND: 3745 return "SEND"; 3746 case RDMA_WR_TYPE_DATA: 3747 return "DATA"; 3748 default: 3749 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 3750 SPDK_UNREACHABLE(); 3751 } 3752 } 3753 3754 static inline void 3755 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 3756 { 3757 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 3758 3759 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 3760 /* If qpair is in ERR state, we will receive completions for all posted and not completed 3761 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 3762 SPDK_DEBUGLOG(rdma, 3763 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3764 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3765 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3766 } else { 3767 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3768 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3769 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3770 } 3771 } 3772 3773 static int 3774 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3775 struct spdk_nvmf_rdma_poller *rpoller) 3776 { 3777 struct ibv_wc wc[32]; 3778 struct spdk_nvmf_rdma_wr *rdma_wr; 3779 struct spdk_nvmf_rdma_request *rdma_req; 3780 struct spdk_nvmf_rdma_recv *rdma_recv; 3781 struct spdk_nvmf_rdma_qpair *rqpair; 3782 int reaped, i; 3783 int count = 0; 3784 bool error = false; 3785 uint64_t poll_tsc = spdk_get_ticks(); 3786 3787 /* Poll for completing operations. */ 3788 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3789 if (reaped < 0) { 3790 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3791 errno, spdk_strerror(errno)); 3792 return -1; 3793 } else if (reaped == 0) { 3794 rpoller->stat.idle_polls++; 3795 } 3796 3797 rpoller->stat.polls++; 3798 rpoller->stat.completions += reaped; 3799 3800 for (i = 0; i < reaped; i++) { 3801 3802 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3803 3804 switch (rdma_wr->type) { 3805 case RDMA_WR_TYPE_SEND: 3806 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3807 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3808 3809 if (!wc[i].status) { 3810 count++; 3811 assert(wc[i].opcode == IBV_WC_SEND); 3812 assert(nvmf_rdma_req_is_completing(rdma_req)); 3813 } 3814 3815 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3816 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3817 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3818 rdma_req->num_outstanding_data_wr = 0; 3819 3820 nvmf_rdma_request_process(rtransport, rdma_req); 3821 break; 3822 case RDMA_WR_TYPE_RECV: 3823 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3824 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3825 if (rpoller->srq != NULL) { 3826 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3827 /* It is possible that there are still some completions for destroyed QP 3828 * associated with SRQ. We just ignore these late completions and re-post 3829 * receive WRs back to SRQ. 3830 */ 3831 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3832 struct ibv_recv_wr *bad_wr; 3833 int rc; 3834 3835 rdma_recv->wr.next = NULL; 3836 spdk_rdma_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr); 3837 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_wr); 3838 if (rc) { 3839 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3840 } 3841 continue; 3842 } 3843 } 3844 rqpair = rdma_recv->qpair; 3845 3846 assert(rqpair != NULL); 3847 if (!wc[i].status) { 3848 assert(wc[i].opcode == IBV_WC_RECV); 3849 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3850 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3851 break; 3852 } 3853 } 3854 3855 rdma_recv->wr.next = NULL; 3856 rqpair->current_recv_depth++; 3857 rdma_recv->receive_tsc = poll_tsc; 3858 rpoller->stat.requests++; 3859 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3860 break; 3861 case RDMA_WR_TYPE_DATA: 3862 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3863 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3864 3865 assert(rdma_req->num_outstanding_data_wr > 0); 3866 3867 rqpair->current_send_depth--; 3868 rdma_req->num_outstanding_data_wr--; 3869 if (!wc[i].status) { 3870 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3871 rqpair->current_read_depth--; 3872 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3873 if (rdma_req->num_outstanding_data_wr == 0) { 3874 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3875 nvmf_rdma_request_process(rtransport, rdma_req); 3876 } 3877 } else { 3878 /* If the data transfer fails still force the queue into the error state, 3879 * if we were performing an RDMA_READ, we need to force the request into a 3880 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3881 * case, we should wait for the SEND to complete. */ 3882 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3883 rqpair->current_read_depth--; 3884 if (rdma_req->num_outstanding_data_wr == 0) { 3885 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3886 } 3887 } 3888 } 3889 break; 3890 default: 3891 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3892 continue; 3893 } 3894 3895 /* Handle error conditions */ 3896 if (wc[i].status) { 3897 nvmf_rdma_update_ibv_state(rqpair); 3898 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 3899 3900 error = true; 3901 3902 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3903 /* Disconnect the connection. */ 3904 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3905 } else { 3906 nvmf_rdma_destroy_drained_qpair(rqpair); 3907 } 3908 continue; 3909 } 3910 3911 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3912 3913 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3914 nvmf_rdma_destroy_drained_qpair(rqpair); 3915 } 3916 } 3917 3918 if (error == true) { 3919 return -1; 3920 } 3921 3922 /* submit outstanding work requests. */ 3923 _poller_submit_recvs(rtransport, rpoller); 3924 _poller_submit_sends(rtransport, rpoller); 3925 3926 return count; 3927 } 3928 3929 static int 3930 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3931 { 3932 struct spdk_nvmf_rdma_transport *rtransport; 3933 struct spdk_nvmf_rdma_poll_group *rgroup; 3934 struct spdk_nvmf_rdma_poller *rpoller; 3935 int count, rc; 3936 3937 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3938 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3939 3940 count = 0; 3941 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3942 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 3943 if (rc < 0) { 3944 return rc; 3945 } 3946 count += rc; 3947 } 3948 3949 return count; 3950 } 3951 3952 static int 3953 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3954 struct spdk_nvme_transport_id *trid, 3955 bool peer) 3956 { 3957 struct sockaddr *saddr; 3958 uint16_t port; 3959 3960 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 3961 3962 if (peer) { 3963 saddr = rdma_get_peer_addr(id); 3964 } else { 3965 saddr = rdma_get_local_addr(id); 3966 } 3967 switch (saddr->sa_family) { 3968 case AF_INET: { 3969 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3970 3971 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3972 inet_ntop(AF_INET, &saddr_in->sin_addr, 3973 trid->traddr, sizeof(trid->traddr)); 3974 if (peer) { 3975 port = ntohs(rdma_get_dst_port(id)); 3976 } else { 3977 port = ntohs(rdma_get_src_port(id)); 3978 } 3979 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3980 break; 3981 } 3982 case AF_INET6: { 3983 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3984 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3985 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3986 trid->traddr, sizeof(trid->traddr)); 3987 if (peer) { 3988 port = ntohs(rdma_get_dst_port(id)); 3989 } else { 3990 port = ntohs(rdma_get_src_port(id)); 3991 } 3992 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3993 break; 3994 } 3995 default: 3996 return -1; 3997 3998 } 3999 4000 return 0; 4001 } 4002 4003 static int 4004 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4005 struct spdk_nvme_transport_id *trid) 4006 { 4007 struct spdk_nvmf_rdma_qpair *rqpair; 4008 4009 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4010 4011 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4012 } 4013 4014 static int 4015 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4016 struct spdk_nvme_transport_id *trid) 4017 { 4018 struct spdk_nvmf_rdma_qpair *rqpair; 4019 4020 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4021 4022 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4023 } 4024 4025 static int 4026 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4027 struct spdk_nvme_transport_id *trid) 4028 { 4029 struct spdk_nvmf_rdma_qpair *rqpair; 4030 4031 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4032 4033 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4034 } 4035 4036 void 4037 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4038 { 4039 g_nvmf_hooks = *hooks; 4040 } 4041 4042 static void 4043 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4044 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4045 { 4046 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4047 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4048 4049 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4050 4051 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4052 } 4053 4054 static int 4055 _nvmf_rdma_qpair_abort_request(void *ctx) 4056 { 4057 struct spdk_nvmf_request *req = ctx; 4058 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4059 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4060 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4061 struct spdk_nvmf_rdma_qpair, qpair); 4062 int rc; 4063 4064 spdk_poller_unregister(&req->poller); 4065 4066 switch (rdma_req_to_abort->state) { 4067 case RDMA_REQUEST_STATE_EXECUTING: 4068 rc = nvmf_ctrlr_abort_request(req); 4069 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4070 return SPDK_POLLER_BUSY; 4071 } 4072 break; 4073 4074 case RDMA_REQUEST_STATE_NEED_BUFFER: 4075 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4076 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4077 4078 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4079 break; 4080 4081 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4082 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4083 spdk_nvmf_rdma_request, state_link); 4084 4085 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4086 break; 4087 4088 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4089 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4090 spdk_nvmf_rdma_request, state_link); 4091 4092 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4093 break; 4094 4095 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4096 if (spdk_get_ticks() < req->timeout_tsc) { 4097 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4098 return SPDK_POLLER_BUSY; 4099 } 4100 break; 4101 4102 default: 4103 break; 4104 } 4105 4106 spdk_nvmf_request_complete(req); 4107 return SPDK_POLLER_BUSY; 4108 } 4109 4110 static void 4111 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4112 struct spdk_nvmf_request *req) 4113 { 4114 struct spdk_nvmf_rdma_qpair *rqpair; 4115 struct spdk_nvmf_rdma_transport *rtransport; 4116 struct spdk_nvmf_transport *transport; 4117 uint16_t cid; 4118 uint32_t i, max_req_count; 4119 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req; 4120 4121 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4122 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4123 transport = &rtransport->transport; 4124 4125 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4126 max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth; 4127 4128 for (i = 0; i < max_req_count; i++) { 4129 rdma_req = &rqpair->resources->reqs[i]; 4130 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair 4131 * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to 4132 * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */ 4133 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid && 4134 rdma_req->req.qpair == qpair) { 4135 rdma_req_to_abort = rdma_req; 4136 break; 4137 } 4138 } 4139 4140 if (rdma_req_to_abort == NULL) { 4141 spdk_nvmf_request_complete(req); 4142 return; 4143 } 4144 4145 req->req_to_abort = &rdma_req_to_abort->req; 4146 req->timeout_tsc = spdk_get_ticks() + 4147 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4148 req->poller = NULL; 4149 4150 _nvmf_rdma_qpair_abort_request(req); 4151 } 4152 4153 /* Deprecated, please use the flow with nvmf_rdma_poll_group_dump_stat. */ 4154 static int 4155 nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4156 struct spdk_nvmf_transport_poll_group_stat **stat) 4157 { 4158 struct spdk_io_channel *ch; 4159 struct spdk_nvmf_poll_group *group; 4160 struct spdk_nvmf_transport_poll_group *tgroup; 4161 struct spdk_nvmf_rdma_poll_group *rgroup; 4162 struct spdk_nvmf_rdma_poller *rpoller; 4163 struct spdk_nvmf_rdma_device_stat *device_stat; 4164 uint64_t num_devices = 0; 4165 4166 SPDK_ERRLOG("nvmf_rdma_poll_group_get_stat is deprecated and will be removed\n"); 4167 4168 if (tgt == NULL || stat == NULL) { 4169 return -EINVAL; 4170 } 4171 4172 ch = spdk_get_io_channel(tgt); 4173 group = spdk_io_channel_get_ctx(ch);; 4174 spdk_put_io_channel(ch); 4175 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4176 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4177 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4178 if (!*stat) { 4179 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4180 return -ENOMEM; 4181 } 4182 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4183 4184 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4185 /* Count devices to allocate enough memory */ 4186 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4187 ++num_devices; 4188 } 4189 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4190 if (!(*stat)->rdma.devices) { 4191 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4192 free(*stat); 4193 return -ENOMEM; 4194 } 4195 4196 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4197 (*stat)->rdma.num_devices = num_devices; 4198 num_devices = 0; 4199 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4200 device_stat = &(*stat)->rdma.devices[num_devices++]; 4201 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4202 device_stat->polls = rpoller->stat.polls; 4203 device_stat->idle_polls = rpoller->stat.idle_polls; 4204 device_stat->completions = rpoller->stat.completions; 4205 device_stat->requests = rpoller->stat.requests; 4206 device_stat->request_latency = rpoller->stat.request_latency; 4207 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4208 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4209 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4210 device_stat->total_send_wrs = rpoller->stat.qp_stats.send.num_submitted_wrs; 4211 device_stat->send_doorbell_updates = rpoller->stat.qp_stats.send.doorbell_updates; 4212 device_stat->total_recv_wrs = rpoller->stat.qp_stats.recv.num_submitted_wrs; 4213 device_stat->recv_doorbell_updates = rpoller->stat.qp_stats.recv.doorbell_updates; 4214 } 4215 return 0; 4216 } 4217 } 4218 return -ENOENT; 4219 } 4220 4221 /* Deprecated, please use the flow with nvmf_rdma_poll_group_dump_stat. */ 4222 static void 4223 nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4224 { 4225 SPDK_ERRLOG("nvmf_rdma_poll_group_free_stat is deprecated and will be removed\n"); 4226 4227 if (stat) { 4228 free(stat->rdma.devices); 4229 } 4230 free(stat); 4231 } 4232 4233 static void 4234 nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group, 4235 struct spdk_json_write_ctx *w) 4236 { 4237 struct spdk_nvmf_rdma_poll_group *rgroup; 4238 struct spdk_nvmf_rdma_poller *rpoller; 4239 4240 assert(w != NULL); 4241 4242 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4243 4244 spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer); 4245 4246 spdk_json_write_named_array_begin(w, "devices"); 4247 4248 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4249 spdk_json_write_object_begin(w); 4250 spdk_json_write_named_string(w, "name", 4251 ibv_get_device_name(rpoller->device->context->device)); 4252 spdk_json_write_named_uint64(w, "polls", 4253 rpoller->stat.polls); 4254 spdk_json_write_named_uint64(w, "idle_polls", 4255 rpoller->stat.idle_polls); 4256 spdk_json_write_named_uint64(w, "completions", 4257 rpoller->stat.completions); 4258 spdk_json_write_named_uint64(w, "requests", 4259 rpoller->stat.requests); 4260 spdk_json_write_named_uint64(w, "request_latency", 4261 rpoller->stat.request_latency); 4262 spdk_json_write_named_uint64(w, "pending_free_request", 4263 rpoller->stat.pending_free_request); 4264 spdk_json_write_named_uint64(w, "pending_rdma_read", 4265 rpoller->stat.pending_rdma_read); 4266 spdk_json_write_named_uint64(w, "pending_rdma_write", 4267 rpoller->stat.pending_rdma_write); 4268 spdk_json_write_named_uint64(w, "total_send_wrs", 4269 rpoller->stat.qp_stats.send.num_submitted_wrs); 4270 spdk_json_write_named_uint64(w, "send_doorbell_updates", 4271 rpoller->stat.qp_stats.send.doorbell_updates); 4272 spdk_json_write_named_uint64(w, "total_recv_wrs", 4273 rpoller->stat.qp_stats.recv.num_submitted_wrs); 4274 spdk_json_write_named_uint64(w, "recv_doorbell_updates", 4275 rpoller->stat.qp_stats.recv.doorbell_updates); 4276 spdk_json_write_object_end(w); 4277 } 4278 4279 spdk_json_write_array_end(w); 4280 } 4281 4282 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4283 .name = "RDMA", 4284 .type = SPDK_NVME_TRANSPORT_RDMA, 4285 .opts_init = nvmf_rdma_opts_init, 4286 .create = nvmf_rdma_create, 4287 .dump_opts = nvmf_rdma_dump_opts, 4288 .destroy = nvmf_rdma_destroy, 4289 4290 .listen = nvmf_rdma_listen, 4291 .stop_listen = nvmf_rdma_stop_listen, 4292 .accept = nvmf_rdma_accept, 4293 .cdata_init = nvmf_rdma_cdata_init, 4294 4295 .listener_discover = nvmf_rdma_discover, 4296 4297 .poll_group_create = nvmf_rdma_poll_group_create, 4298 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4299 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4300 .poll_group_add = nvmf_rdma_poll_group_add, 4301 .poll_group_remove = nvmf_rdma_poll_group_remove, 4302 .poll_group_poll = nvmf_rdma_poll_group_poll, 4303 4304 .req_free = nvmf_rdma_request_free, 4305 .req_complete = nvmf_rdma_request_complete, 4306 4307 .qpair_fini = nvmf_rdma_close_qpair, 4308 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4309 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4310 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4311 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4312 4313 .poll_group_get_stat = nvmf_rdma_poll_group_get_stat, 4314 .poll_group_free_stat = nvmf_rdma_poll_group_free_stat, 4315 .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat, 4316 }; 4317 4318 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4319 SPDK_LOG_REGISTER_COMPONENT(rdma) 4320