1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2019, 2020 Mellanox Technologies LTD. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * * Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * * Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * * Neither the name of Intel Corporation nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 */ 33 34 #include "spdk/stdinc.h" 35 36 #include "spdk/config.h" 37 #include "spdk/thread.h" 38 #include "spdk/likely.h" 39 #include "spdk/nvmf_transport.h" 40 #include "spdk/string.h" 41 #include "spdk/trace.h" 42 #include "spdk/util.h" 43 44 #include "spdk_internal/assert.h" 45 #include "spdk/log.h" 46 #include "spdk_internal/rdma.h" 47 48 #include "nvmf_internal.h" 49 50 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 51 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma; 52 53 /* 54 RDMA Connection Resource Defaults 55 */ 56 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 57 #define NVMF_DEFAULT_RSP_SGE 1 58 #define NVMF_DEFAULT_RX_SGE 2 59 60 /* The RDMA completion queue size */ 61 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 62 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 63 64 /* rxe driver vendor_id has been changed from 0 to 0XFFFFFF in 0184afd15a141d7ce24c32c0d86a1e3ba6bc0eb3 */ 65 #define NVMF_RXE_VENDOR_ID_OLD 0 66 #define NVMF_RXE_VENDOR_ID_NEW 0XFFFFFF 67 68 static int g_spdk_nvmf_ibv_query_mask = 69 IBV_QP_STATE | 70 IBV_QP_PKEY_INDEX | 71 IBV_QP_PORT | 72 IBV_QP_ACCESS_FLAGS | 73 IBV_QP_AV | 74 IBV_QP_PATH_MTU | 75 IBV_QP_DEST_QPN | 76 IBV_QP_RQ_PSN | 77 IBV_QP_MAX_DEST_RD_ATOMIC | 78 IBV_QP_MIN_RNR_TIMER | 79 IBV_QP_SQ_PSN | 80 IBV_QP_TIMEOUT | 81 IBV_QP_RETRY_CNT | 82 IBV_QP_RNR_RETRY | 83 IBV_QP_MAX_QP_RD_ATOMIC; 84 85 enum spdk_nvmf_rdma_request_state { 86 /* The request is not currently in use */ 87 RDMA_REQUEST_STATE_FREE = 0, 88 89 /* Initial state when request first received */ 90 RDMA_REQUEST_STATE_NEW, 91 92 /* The request is queued until a data buffer is available. */ 93 RDMA_REQUEST_STATE_NEED_BUFFER, 94 95 /* The request is waiting on RDMA queue depth availability 96 * to transfer data from the host to the controller. 97 */ 98 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 99 100 /* The request is currently transferring data from the host to the controller. */ 101 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 102 103 /* The request is ready to execute at the block device */ 104 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 105 106 /* The request is currently executing at the block device */ 107 RDMA_REQUEST_STATE_EXECUTING, 108 109 /* The request finished executing at the block device */ 110 RDMA_REQUEST_STATE_EXECUTED, 111 112 /* The request is waiting on RDMA queue depth availability 113 * to transfer data from the controller to the host. 114 */ 115 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 116 117 /* The request is ready to send a completion */ 118 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 119 120 /* The request is currently transferring data from the controller to the host. */ 121 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 122 123 /* The request currently has an outstanding completion without an 124 * associated data transfer. 125 */ 126 RDMA_REQUEST_STATE_COMPLETING, 127 128 /* The request completed and can be marked free. */ 129 RDMA_REQUEST_STATE_COMPLETED, 130 131 /* Terminator */ 132 RDMA_REQUEST_NUM_STATES, 133 }; 134 135 #define OBJECT_NVMF_RDMA_IO 0x40 136 137 #define TRACE_GROUP_NVMF_RDMA 0x4 138 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 139 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 140 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 141 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 142 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 143 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 144 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 145 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 146 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 147 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 148 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 149 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 150 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 151 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 152 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 153 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 154 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 155 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 156 157 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 158 { 159 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 160 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 161 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 162 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 163 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 164 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 165 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 166 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 167 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 168 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 169 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 170 spdk_trace_register_description("RDMA_REQ_TX_H2C", 171 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 172 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 173 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 174 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 175 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 176 spdk_trace_register_description("RDMA_REQ_EXECUTING", 177 TRACE_RDMA_REQUEST_STATE_EXECUTING, 178 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 179 spdk_trace_register_description("RDMA_REQ_EXECUTED", 180 TRACE_RDMA_REQUEST_STATE_EXECUTED, 181 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 182 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 183 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 184 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 185 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 186 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 187 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 188 spdk_trace_register_description("RDMA_REQ_COMPLETING", 189 TRACE_RDMA_REQUEST_STATE_COMPLETING, 190 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 191 spdk_trace_register_description("RDMA_REQ_COMPLETED", 192 TRACE_RDMA_REQUEST_STATE_COMPLETED, 193 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 194 195 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 196 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 197 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 198 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 199 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 200 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 201 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 202 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 203 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 204 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 205 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 206 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 207 } 208 209 enum spdk_nvmf_rdma_wr_type { 210 RDMA_WR_TYPE_RECV, 211 RDMA_WR_TYPE_SEND, 212 RDMA_WR_TYPE_DATA, 213 }; 214 215 struct spdk_nvmf_rdma_wr { 216 enum spdk_nvmf_rdma_wr_type type; 217 }; 218 219 /* This structure holds commands as they are received off the wire. 220 * It must be dynamically paired with a full request object 221 * (spdk_nvmf_rdma_request) to service a request. It is separate 222 * from the request because RDMA does not appear to order 223 * completions, so occasionally we'll get a new incoming 224 * command when there aren't any free request objects. 225 */ 226 struct spdk_nvmf_rdma_recv { 227 struct ibv_recv_wr wr; 228 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 229 230 struct spdk_nvmf_rdma_qpair *qpair; 231 232 /* In-capsule data buffer */ 233 uint8_t *buf; 234 235 struct spdk_nvmf_rdma_wr rdma_wr; 236 uint64_t receive_tsc; 237 238 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 239 }; 240 241 struct spdk_nvmf_rdma_request_data { 242 struct spdk_nvmf_rdma_wr rdma_wr; 243 struct ibv_send_wr wr; 244 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 245 }; 246 247 struct spdk_nvmf_rdma_request { 248 struct spdk_nvmf_request req; 249 250 enum spdk_nvmf_rdma_request_state state; 251 252 struct spdk_nvmf_rdma_recv *recv; 253 254 struct { 255 struct spdk_nvmf_rdma_wr rdma_wr; 256 struct ibv_send_wr wr; 257 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 258 } rsp; 259 260 struct spdk_nvmf_rdma_request_data data; 261 262 uint32_t iovpos; 263 264 uint32_t num_outstanding_data_wr; 265 uint64_t receive_tsc; 266 267 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 268 }; 269 270 struct spdk_nvmf_rdma_resource_opts { 271 struct spdk_nvmf_rdma_qpair *qpair; 272 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 273 void *qp; 274 struct ibv_pd *pd; 275 uint32_t max_queue_depth; 276 uint32_t in_capsule_data_size; 277 bool shared; 278 }; 279 280 struct spdk_nvmf_recv_wr_list { 281 struct ibv_recv_wr *first; 282 struct ibv_recv_wr *last; 283 }; 284 285 struct spdk_nvmf_rdma_resources { 286 /* Array of size "max_queue_depth" containing RDMA requests. */ 287 struct spdk_nvmf_rdma_request *reqs; 288 289 /* Array of size "max_queue_depth" containing RDMA recvs. */ 290 struct spdk_nvmf_rdma_recv *recvs; 291 292 /* Array of size "max_queue_depth" containing 64 byte capsules 293 * used for receive. 294 */ 295 union nvmf_h2c_msg *cmds; 296 struct ibv_mr *cmds_mr; 297 298 /* Array of size "max_queue_depth" containing 16 byte completions 299 * to be sent back to the user. 300 */ 301 union nvmf_c2h_msg *cpls; 302 struct ibv_mr *cpls_mr; 303 304 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 305 * buffers to be used for in capsule data. 306 */ 307 void *bufs; 308 struct ibv_mr *bufs_mr; 309 310 /* The list of pending recvs to transfer */ 311 struct spdk_nvmf_recv_wr_list recvs_to_post; 312 313 /* Receives that are waiting for a request object */ 314 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 315 316 /* Queue to track free requests */ 317 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 318 }; 319 320 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 321 322 struct spdk_nvmf_rdma_ibv_event_ctx { 323 struct spdk_nvmf_rdma_qpair *rqpair; 324 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 325 /* Link to other ibv events associated with this qpair */ 326 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 327 }; 328 329 struct spdk_nvmf_rdma_qpair { 330 struct spdk_nvmf_qpair qpair; 331 332 struct spdk_nvmf_rdma_device *device; 333 struct spdk_nvmf_rdma_poller *poller; 334 335 struct spdk_rdma_qp *rdma_qp; 336 struct rdma_cm_id *cm_id; 337 struct ibv_srq *srq; 338 struct rdma_cm_id *listen_id; 339 340 /* The maximum number of I/O outstanding on this connection at one time */ 341 uint16_t max_queue_depth; 342 343 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 344 uint16_t max_read_depth; 345 346 /* The maximum number of RDMA SEND operations at one time */ 347 uint32_t max_send_depth; 348 349 /* The current number of outstanding WRs from this qpair's 350 * recv queue. Should not exceed device->attr.max_queue_depth. 351 */ 352 uint16_t current_recv_depth; 353 354 /* The current number of active RDMA READ operations */ 355 uint16_t current_read_depth; 356 357 /* The current number of posted WRs from this qpair's 358 * send queue. Should not exceed max_send_depth. 359 */ 360 uint32_t current_send_depth; 361 362 /* The maximum number of SGEs per WR on the send queue */ 363 uint32_t max_send_sge; 364 365 /* The maximum number of SGEs per WR on the recv queue */ 366 uint32_t max_recv_sge; 367 368 struct spdk_nvmf_rdma_resources *resources; 369 370 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 371 372 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 373 374 /* Number of requests not in the free state */ 375 uint32_t qd; 376 377 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 378 379 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 380 381 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 382 383 /* IBV queue pair attributes: they are used to manage 384 * qp state and recover from errors. 385 */ 386 enum ibv_qp_state ibv_state; 387 388 /* 389 * io_channel which is used to destroy qpair when it is removed from poll group 390 */ 391 struct spdk_io_channel *destruct_channel; 392 393 /* List of ibv async events */ 394 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 395 396 /* Lets us know that we have received the last_wqe event. */ 397 bool last_wqe_reached; 398 399 /* Indicate that nvmf_rdma_close_qpair is called */ 400 bool to_close; 401 }; 402 403 struct spdk_nvmf_rdma_poller_stat { 404 uint64_t completions; 405 uint64_t polls; 406 uint64_t requests; 407 uint64_t request_latency; 408 uint64_t pending_free_request; 409 uint64_t pending_rdma_read; 410 uint64_t pending_rdma_write; 411 }; 412 413 struct spdk_nvmf_rdma_poller { 414 struct spdk_nvmf_rdma_device *device; 415 struct spdk_nvmf_rdma_poll_group *group; 416 417 int num_cqe; 418 int required_num_wr; 419 struct ibv_cq *cq; 420 421 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 422 uint16_t max_srq_depth; 423 424 /* Shared receive queue */ 425 struct ibv_srq *srq; 426 427 struct spdk_nvmf_rdma_resources *resources; 428 struct spdk_nvmf_rdma_poller_stat stat; 429 430 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 431 432 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 433 434 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 435 436 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 437 }; 438 439 struct spdk_nvmf_rdma_poll_group_stat { 440 uint64_t pending_data_buffer; 441 }; 442 443 struct spdk_nvmf_rdma_poll_group { 444 struct spdk_nvmf_transport_poll_group group; 445 struct spdk_nvmf_rdma_poll_group_stat stat; 446 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 447 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 448 /* 449 * buffers which are split across multiple RDMA 450 * memory regions cannot be used by this transport. 451 */ 452 STAILQ_HEAD(, spdk_nvmf_transport_pg_cache_buf) retired_bufs; 453 }; 454 455 struct spdk_nvmf_rdma_conn_sched { 456 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 457 struct spdk_nvmf_rdma_poll_group *next_io_pg; 458 }; 459 460 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 461 struct spdk_nvmf_rdma_device { 462 struct ibv_device_attr attr; 463 struct ibv_context *context; 464 465 struct spdk_mem_map *map; 466 struct ibv_pd *pd; 467 468 int num_srq; 469 470 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 471 }; 472 473 struct spdk_nvmf_rdma_port { 474 const struct spdk_nvme_transport_id *trid; 475 struct rdma_cm_id *id; 476 struct spdk_nvmf_rdma_device *device; 477 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 478 }; 479 480 struct rdma_transport_opts { 481 uint32_t max_srq_depth; 482 bool no_srq; 483 bool no_wr_batching; 484 int acceptor_backlog; 485 }; 486 487 struct spdk_nvmf_rdma_transport { 488 struct spdk_nvmf_transport transport; 489 struct rdma_transport_opts rdma_opts; 490 491 struct spdk_nvmf_rdma_conn_sched conn_sched; 492 493 struct rdma_event_channel *event_channel; 494 495 struct spdk_mempool *data_wr_pool; 496 497 pthread_mutex_t lock; 498 499 /* fields used to poll RDMA/IB events */ 500 nfds_t npoll_fds; 501 struct pollfd *poll_fds; 502 503 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 504 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 505 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 506 }; 507 508 static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = { 509 { 510 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth), 511 spdk_json_decode_uint32, true 512 }, 513 { 514 "no_srq", offsetof(struct rdma_transport_opts, no_srq), 515 spdk_json_decode_bool, true 516 }, 517 { 518 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching), 519 spdk_json_decode_bool, true 520 }, 521 { 522 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog), 523 spdk_json_decode_int32, true 524 }, 525 }; 526 527 static bool 528 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 529 struct spdk_nvmf_rdma_request *rdma_req); 530 531 static void 532 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 533 struct spdk_nvmf_rdma_poller *rpoller); 534 535 static void 536 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 537 struct spdk_nvmf_rdma_poller *rpoller); 538 539 static inline int 540 nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 541 { 542 switch (state) { 543 case IBV_QPS_RESET: 544 case IBV_QPS_INIT: 545 case IBV_QPS_RTR: 546 case IBV_QPS_RTS: 547 case IBV_QPS_SQD: 548 case IBV_QPS_SQE: 549 case IBV_QPS_ERR: 550 return 0; 551 default: 552 return -1; 553 } 554 } 555 556 static inline enum spdk_nvme_media_error_status_code 557 nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 558 enum spdk_nvme_media_error_status_code result; 559 switch (err_type) 560 { 561 case SPDK_DIF_REFTAG_ERROR: 562 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 563 break; 564 case SPDK_DIF_APPTAG_ERROR: 565 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 566 break; 567 case SPDK_DIF_GUARD_ERROR: 568 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 569 break; 570 default: 571 SPDK_UNREACHABLE(); 572 } 573 574 return result; 575 } 576 577 static enum ibv_qp_state 578 nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 579 enum ibv_qp_state old_state, new_state; 580 struct ibv_qp_attr qp_attr; 581 struct ibv_qp_init_attr init_attr; 582 int rc; 583 584 old_state = rqpair->ibv_state; 585 rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr, 586 g_spdk_nvmf_ibv_query_mask, &init_attr); 587 588 if (rc) 589 { 590 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 591 return IBV_QPS_ERR + 1; 592 } 593 594 new_state = qp_attr.qp_state; 595 rqpair->ibv_state = new_state; 596 qp_attr.ah_attr.port_num = qp_attr.port_num; 597 598 rc = nvmf_rdma_check_ibv_state(new_state); 599 if (rc) 600 { 601 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 602 /* 603 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 604 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 605 */ 606 return IBV_QPS_ERR + 1; 607 } 608 609 if (old_state != new_state) 610 { 611 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 612 (uintptr_t)rqpair->cm_id, new_state); 613 } 614 return new_state; 615 } 616 617 static void 618 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 619 struct spdk_nvmf_rdma_transport *rtransport) 620 { 621 struct spdk_nvmf_rdma_request_data *data_wr; 622 struct ibv_send_wr *next_send_wr; 623 uint64_t req_wrid; 624 625 rdma_req->num_outstanding_data_wr = 0; 626 data_wr = &rdma_req->data; 627 req_wrid = data_wr->wr.wr_id; 628 while (data_wr && data_wr->wr.wr_id == req_wrid) { 629 memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge); 630 data_wr->wr.num_sge = 0; 631 next_send_wr = data_wr->wr.next; 632 if (data_wr != &rdma_req->data) { 633 spdk_mempool_put(rtransport->data_wr_pool, data_wr); 634 } 635 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : 636 SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr); 637 } 638 } 639 640 static void 641 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 642 { 643 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 644 if (req->req.cmd) { 645 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 646 } 647 if (req->recv) { 648 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 649 } 650 } 651 652 static void 653 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 654 { 655 int i; 656 657 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 658 for (i = 0; i < rqpair->max_queue_depth; i++) { 659 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 660 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 661 } 662 } 663 } 664 665 static void 666 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 667 { 668 if (resources->cmds_mr) { 669 ibv_dereg_mr(resources->cmds_mr); 670 } 671 672 if (resources->cpls_mr) { 673 ibv_dereg_mr(resources->cpls_mr); 674 } 675 676 if (resources->bufs_mr) { 677 ibv_dereg_mr(resources->bufs_mr); 678 } 679 680 spdk_free(resources->cmds); 681 spdk_free(resources->cpls); 682 spdk_free(resources->bufs); 683 free(resources->reqs); 684 free(resources->recvs); 685 free(resources); 686 } 687 688 689 static struct spdk_nvmf_rdma_resources * 690 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 691 { 692 struct spdk_nvmf_rdma_resources *resources; 693 struct spdk_nvmf_rdma_request *rdma_req; 694 struct spdk_nvmf_rdma_recv *rdma_recv; 695 struct ibv_qp *qp; 696 struct ibv_srq *srq; 697 uint32_t i; 698 int rc; 699 700 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 701 if (!resources) { 702 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 703 return NULL; 704 } 705 706 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 707 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 708 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 709 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 710 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 711 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 712 713 if (opts->in_capsule_data_size > 0) { 714 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 715 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 716 SPDK_MALLOC_DMA); 717 } 718 719 if (!resources->reqs || !resources->recvs || !resources->cmds || 720 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 721 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 722 goto cleanup; 723 } 724 725 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 726 opts->max_queue_depth * sizeof(*resources->cmds), 727 IBV_ACCESS_LOCAL_WRITE); 728 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 729 opts->max_queue_depth * sizeof(*resources->cpls), 730 0); 731 732 if (opts->in_capsule_data_size) { 733 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 734 opts->max_queue_depth * 735 opts->in_capsule_data_size, 736 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 737 } 738 739 if (!resources->cmds_mr || !resources->cpls_mr || 740 (opts->in_capsule_data_size && 741 !resources->bufs_mr)) { 742 goto cleanup; 743 } 744 SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx LKey: %x\n", 745 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 746 resources->cmds_mr->lkey); 747 SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx LKey: %x\n", 748 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 749 resources->cpls_mr->lkey); 750 if (resources->bufs && resources->bufs_mr) { 751 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x LKey: %x\n", 752 resources->bufs, opts->max_queue_depth * 753 opts->in_capsule_data_size, resources->bufs_mr->lkey); 754 } 755 756 /* Initialize queues */ 757 STAILQ_INIT(&resources->incoming_queue); 758 STAILQ_INIT(&resources->free_queue); 759 760 for (i = 0; i < opts->max_queue_depth; i++) { 761 struct ibv_recv_wr *bad_wr = NULL; 762 763 rdma_recv = &resources->recvs[i]; 764 rdma_recv->qpair = opts->qpair; 765 766 /* Set up memory to receive commands */ 767 if (resources->bufs) { 768 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 769 opts->in_capsule_data_size)); 770 } 771 772 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 773 774 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 775 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 776 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 777 rdma_recv->wr.num_sge = 1; 778 779 if (rdma_recv->buf && resources->bufs_mr) { 780 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 781 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 782 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 783 rdma_recv->wr.num_sge++; 784 } 785 786 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 787 rdma_recv->wr.sg_list = rdma_recv->sgl; 788 if (opts->shared) { 789 srq = (struct ibv_srq *)opts->qp; 790 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 791 } else { 792 qp = (struct ibv_qp *)opts->qp; 793 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 794 } 795 if (rc) { 796 goto cleanup; 797 } 798 } 799 800 for (i = 0; i < opts->max_queue_depth; i++) { 801 rdma_req = &resources->reqs[i]; 802 803 if (opts->qpair != NULL) { 804 rdma_req->req.qpair = &opts->qpair->qpair; 805 } else { 806 rdma_req->req.qpair = NULL; 807 } 808 rdma_req->req.cmd = NULL; 809 810 /* Set up memory to send responses */ 811 rdma_req->req.rsp = &resources->cpls[i]; 812 813 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 814 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 815 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 816 817 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 818 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 819 rdma_req->rsp.wr.next = NULL; 820 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 821 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 822 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 823 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 824 825 /* Set up memory for data buffers */ 826 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 827 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 828 rdma_req->data.wr.next = NULL; 829 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 830 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 831 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 832 833 /* Initialize request state to FREE */ 834 rdma_req->state = RDMA_REQUEST_STATE_FREE; 835 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 836 } 837 838 return resources; 839 840 cleanup: 841 nvmf_rdma_resources_destroy(resources); 842 return NULL; 843 } 844 845 static void 846 nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 847 { 848 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 849 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 850 ctx->rqpair = NULL; 851 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */ 852 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 853 } 854 } 855 856 static void 857 nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 858 { 859 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 860 struct ibv_recv_wr *bad_recv_wr = NULL; 861 int rc; 862 863 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 864 865 if (rqpair->qd != 0) { 866 struct spdk_nvmf_qpair *qpair = &rqpair->qpair; 867 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport, 868 struct spdk_nvmf_rdma_transport, transport); 869 struct spdk_nvmf_rdma_request *req; 870 uint32_t i, max_req_count = 0; 871 872 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 873 874 if (rqpair->srq == NULL) { 875 nvmf_rdma_dump_qpair_contents(rqpair); 876 max_req_count = rqpair->max_queue_depth; 877 } else if (rqpair->poller && rqpair->resources) { 878 max_req_count = rqpair->poller->max_srq_depth; 879 } 880 881 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n"); 882 for (i = 0; i < max_req_count; i++) { 883 req = &rqpair->resources->reqs[i]; 884 if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) { 885 /* nvmf_rdma_request_process checks qpair ibv and internal state 886 * and completes a request */ 887 nvmf_rdma_request_process(rtransport, req); 888 } 889 } 890 assert(rqpair->qd == 0); 891 } 892 893 if (rqpair->poller) { 894 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 895 896 if (rqpair->srq != NULL && rqpair->resources != NULL) { 897 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 898 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 899 if (rqpair == rdma_recv->qpair) { 900 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 901 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 902 if (rc) { 903 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 904 } 905 } 906 } 907 } 908 } 909 910 if (rqpair->cm_id) { 911 if (rqpair->rdma_qp != NULL) { 912 spdk_rdma_qp_destroy(rqpair->rdma_qp); 913 rqpair->rdma_qp = NULL; 914 } 915 rdma_destroy_id(rqpair->cm_id); 916 917 if (rqpair->poller != NULL && rqpair->srq == NULL) { 918 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 919 } 920 } 921 922 if (rqpair->srq == NULL && rqpair->resources != NULL) { 923 nvmf_rdma_resources_destroy(rqpair->resources); 924 } 925 926 nvmf_rdma_qpair_clean_ibv_events(rqpair); 927 928 if (rqpair->destruct_channel) { 929 spdk_put_io_channel(rqpair->destruct_channel); 930 rqpair->destruct_channel = NULL; 931 } 932 933 free(rqpair); 934 } 935 936 static int 937 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 938 { 939 struct spdk_nvmf_rdma_poller *rpoller; 940 int rc, num_cqe, required_num_wr; 941 942 /* Enlarge CQ size dynamically */ 943 rpoller = rqpair->poller; 944 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 945 num_cqe = rpoller->num_cqe; 946 if (num_cqe < required_num_wr) { 947 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 948 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 949 } 950 951 if (rpoller->num_cqe != num_cqe) { 952 if (required_num_wr > device->attr.max_cqe) { 953 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 954 required_num_wr, device->attr.max_cqe); 955 return -1; 956 } 957 958 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 959 rc = ibv_resize_cq(rpoller->cq, num_cqe); 960 if (rc) { 961 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 962 return -1; 963 } 964 965 rpoller->num_cqe = num_cqe; 966 } 967 968 rpoller->required_num_wr = required_num_wr; 969 return 0; 970 } 971 972 static int 973 nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 974 { 975 struct spdk_nvmf_rdma_qpair *rqpair; 976 struct spdk_nvmf_rdma_transport *rtransport; 977 struct spdk_nvmf_transport *transport; 978 struct spdk_nvmf_rdma_resource_opts opts; 979 struct spdk_nvmf_rdma_device *device; 980 struct spdk_rdma_qp_init_attr qp_init_attr = {}; 981 982 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 983 device = rqpair->device; 984 985 qp_init_attr.qp_context = rqpair; 986 qp_init_attr.pd = device->pd; 987 qp_init_attr.send_cq = rqpair->poller->cq; 988 qp_init_attr.recv_cq = rqpair->poller->cq; 989 990 if (rqpair->srq) { 991 qp_init_attr.srq = rqpair->srq; 992 } else { 993 qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth; 994 } 995 996 /* SEND, READ, and WRITE operations */ 997 qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2; 998 qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 999 qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1000 1001 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1002 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1003 goto error; 1004 } 1005 1006 rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr); 1007 if (!rqpair->rdma_qp) { 1008 goto error; 1009 } 1010 1011 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2), 1012 qp_init_attr.cap.max_send_wr); 1013 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge); 1014 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge); 1015 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 1016 SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair); 1017 1018 if (rqpair->poller->srq == NULL) { 1019 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1020 transport = &rtransport->transport; 1021 1022 opts.qp = rqpair->rdma_qp->qp; 1023 opts.pd = rqpair->cm_id->pd; 1024 opts.qpair = rqpair; 1025 opts.shared = false; 1026 opts.max_queue_depth = rqpair->max_queue_depth; 1027 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1028 1029 rqpair->resources = nvmf_rdma_resources_create(&opts); 1030 1031 if (!rqpair->resources) { 1032 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1033 rdma_destroy_qp(rqpair->cm_id); 1034 goto error; 1035 } 1036 } else { 1037 rqpair->resources = rqpair->poller->resources; 1038 } 1039 1040 rqpair->current_recv_depth = 0; 1041 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1042 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1043 1044 return 0; 1045 1046 error: 1047 rdma_destroy_id(rqpair->cm_id); 1048 rqpair->cm_id = NULL; 1049 return -1; 1050 } 1051 1052 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1053 /* This function accepts either a single wr or the first wr in a linked list. */ 1054 static void 1055 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1056 { 1057 struct ibv_recv_wr *last; 1058 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1059 struct spdk_nvmf_rdma_transport, transport); 1060 1061 last = first; 1062 while (last->next != NULL) { 1063 last = last->next; 1064 } 1065 1066 if (rqpair->resources->recvs_to_post.first == NULL) { 1067 rqpair->resources->recvs_to_post.first = first; 1068 rqpair->resources->recvs_to_post.last = last; 1069 if (rqpair->srq == NULL) { 1070 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1071 } 1072 } else { 1073 rqpair->resources->recvs_to_post.last->next = first; 1074 rqpair->resources->recvs_to_post.last = last; 1075 } 1076 1077 if (rtransport->rdma_opts.no_wr_batching) { 1078 _poller_submit_recvs(rtransport, rqpair->poller); 1079 } 1080 } 1081 1082 static int 1083 request_transfer_in(struct spdk_nvmf_request *req) 1084 { 1085 struct spdk_nvmf_rdma_request *rdma_req; 1086 struct spdk_nvmf_qpair *qpair; 1087 struct spdk_nvmf_rdma_qpair *rqpair; 1088 struct spdk_nvmf_rdma_transport *rtransport; 1089 1090 qpair = req->qpair; 1091 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1092 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1093 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1094 struct spdk_nvmf_rdma_transport, transport); 1095 1096 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1097 assert(rdma_req != NULL); 1098 1099 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, &rdma_req->data.wr)) { 1100 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1101 } 1102 if (rtransport->rdma_opts.no_wr_batching) { 1103 _poller_submit_sends(rtransport, rqpair->poller); 1104 } 1105 1106 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1107 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1108 return 0; 1109 } 1110 1111 static int 1112 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1113 { 1114 int num_outstanding_data_wr = 0; 1115 struct spdk_nvmf_rdma_request *rdma_req; 1116 struct spdk_nvmf_qpair *qpair; 1117 struct spdk_nvmf_rdma_qpair *rqpair; 1118 struct spdk_nvme_cpl *rsp; 1119 struct ibv_send_wr *first = NULL; 1120 struct spdk_nvmf_rdma_transport *rtransport; 1121 1122 *data_posted = 0; 1123 qpair = req->qpair; 1124 rsp = &req->rsp->nvme_cpl; 1125 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1126 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1127 rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 1128 struct spdk_nvmf_rdma_transport, transport); 1129 1130 /* Advance our sq_head pointer */ 1131 if (qpair->sq_head == qpair->sq_head_max) { 1132 qpair->sq_head = 0; 1133 } else { 1134 qpair->sq_head++; 1135 } 1136 rsp->sqhd = qpair->sq_head; 1137 1138 /* queue the capsule for the recv buffer */ 1139 assert(rdma_req->recv != NULL); 1140 1141 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1142 1143 rdma_req->recv = NULL; 1144 assert(rqpair->current_recv_depth > 0); 1145 rqpair->current_recv_depth--; 1146 1147 /* Build the response which consists of optional 1148 * RDMA WRITEs to transfer data, plus an RDMA SEND 1149 * containing the response. 1150 */ 1151 first = &rdma_req->rsp.wr; 1152 1153 if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) { 1154 /* On failure, data was not read from the controller. So clear the 1155 * number of outstanding data WRs to zero. 1156 */ 1157 rdma_req->num_outstanding_data_wr = 0; 1158 } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1159 first = &rdma_req->data.wr; 1160 *data_posted = 1; 1161 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1162 } 1163 if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) { 1164 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1165 } 1166 if (rtransport->rdma_opts.no_wr_batching) { 1167 _poller_submit_sends(rtransport, rqpair->poller); 1168 } 1169 1170 /* +1 for the rsp wr */ 1171 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1172 1173 return 0; 1174 } 1175 1176 static int 1177 nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1178 { 1179 struct spdk_nvmf_rdma_accept_private_data accept_data; 1180 struct rdma_conn_param ctrlr_event_data = {}; 1181 int rc; 1182 1183 accept_data.recfmt = 0; 1184 accept_data.crqsize = rqpair->max_queue_depth; 1185 1186 ctrlr_event_data.private_data = &accept_data; 1187 ctrlr_event_data.private_data_len = sizeof(accept_data); 1188 if (id->ps == RDMA_PS_TCP) { 1189 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1190 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1191 } 1192 1193 /* Configure infinite retries for the initiator side qpair. 1194 * When using a shared receive queue on the target side, 1195 * we need to pass this value to the initiator to prevent the 1196 * initiator side NIC from completing SEND requests back to the 1197 * initiator with status rnr_retry_count_exceeded. */ 1198 if (rqpair->srq != NULL) { 1199 ctrlr_event_data.rnr_retry_count = 0x7; 1200 } 1201 1202 /* When qpair is created without use of rdma cm API, an additional 1203 * information must be provided to initiator in the connection response: 1204 * whether qpair is using SRQ and its qp_num 1205 * Fields below are ignored by rdma cm if qpair has been 1206 * created using rdma cm API. */ 1207 ctrlr_event_data.srq = rqpair->srq ? 1 : 0; 1208 ctrlr_event_data.qp_num = rqpair->rdma_qp->qp->qp_num; 1209 1210 rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data); 1211 if (rc) { 1212 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno); 1213 } else { 1214 SPDK_DEBUGLOG(rdma, "Sent back the accept\n"); 1215 } 1216 1217 return rc; 1218 } 1219 1220 static void 1221 nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1222 { 1223 struct spdk_nvmf_rdma_reject_private_data rej_data; 1224 1225 rej_data.recfmt = 0; 1226 rej_data.sts = error; 1227 1228 rdma_reject(id, &rej_data, sizeof(rej_data)); 1229 } 1230 1231 static int 1232 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event) 1233 { 1234 struct spdk_nvmf_rdma_transport *rtransport; 1235 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1236 struct spdk_nvmf_rdma_port *port; 1237 struct rdma_conn_param *rdma_param = NULL; 1238 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1239 uint16_t max_queue_depth; 1240 uint16_t max_read_depth; 1241 1242 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1243 1244 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1245 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1246 1247 rdma_param = &event->param.conn; 1248 if (rdma_param->private_data == NULL || 1249 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1250 SPDK_ERRLOG("connect request: no private data provided\n"); 1251 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1252 return -1; 1253 } 1254 1255 private_data = rdma_param->private_data; 1256 if (private_data->recfmt != 0) { 1257 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1258 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1259 return -1; 1260 } 1261 1262 SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n", 1263 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1264 1265 port = event->listen_id->context; 1266 SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1267 event->listen_id, event->listen_id->verbs, port); 1268 1269 /* Figure out the supported queue depth. This is a multi-step process 1270 * that takes into account hardware maximums, host provided values, 1271 * and our target's internal memory limits */ 1272 1273 SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n"); 1274 1275 /* Start with the maximum queue depth allowed by the target */ 1276 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1277 max_read_depth = rtransport->transport.opts.max_queue_depth; 1278 SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n", 1279 rtransport->transport.opts.max_queue_depth); 1280 1281 /* Next check the local NIC's hardware limitations */ 1282 SPDK_DEBUGLOG(rdma, 1283 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1284 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1285 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1286 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1287 1288 /* Next check the remote NIC's hardware limitations */ 1289 SPDK_DEBUGLOG(rdma, 1290 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1291 rdma_param->initiator_depth, rdma_param->responder_resources); 1292 if (rdma_param->initiator_depth > 0) { 1293 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1294 } 1295 1296 /* Finally check for the host software requested values, which are 1297 * optional. */ 1298 if (rdma_param->private_data != NULL && 1299 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1300 SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1301 SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize); 1302 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1303 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1304 } 1305 1306 SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1307 max_queue_depth, max_read_depth); 1308 1309 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1310 if (rqpair == NULL) { 1311 SPDK_ERRLOG("Could not allocate new connection.\n"); 1312 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1313 return -1; 1314 } 1315 1316 rqpair->device = port->device; 1317 rqpair->max_queue_depth = max_queue_depth; 1318 rqpair->max_read_depth = max_read_depth; 1319 rqpair->cm_id = event->id; 1320 rqpair->listen_id = event->listen_id; 1321 rqpair->qpair.transport = transport; 1322 rqpair->qpair.trid = port->trid; 1323 STAILQ_INIT(&rqpair->ibv_events); 1324 /* use qid from the private data to determine the qpair type 1325 qid will be set to the appropriate value when the controller is created */ 1326 rqpair->qpair.qid = private_data->qid; 1327 1328 event->id->context = &rqpair->qpair; 1329 1330 spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair); 1331 1332 return 0; 1333 } 1334 1335 static int 1336 nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1337 enum spdk_mem_map_notify_action action, 1338 void *vaddr, size_t size) 1339 { 1340 struct ibv_pd *pd = cb_ctx; 1341 struct ibv_mr *mr; 1342 int rc; 1343 1344 switch (action) { 1345 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1346 if (!g_nvmf_hooks.get_rkey) { 1347 mr = ibv_reg_mr(pd, vaddr, size, 1348 IBV_ACCESS_LOCAL_WRITE | 1349 IBV_ACCESS_REMOTE_READ | 1350 IBV_ACCESS_REMOTE_WRITE); 1351 if (mr == NULL) { 1352 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1353 return -1; 1354 } else { 1355 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1356 } 1357 } else { 1358 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1359 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1360 } 1361 break; 1362 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1363 if (!g_nvmf_hooks.get_rkey) { 1364 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1365 if (mr) { 1366 ibv_dereg_mr(mr); 1367 } 1368 } 1369 rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1370 break; 1371 default: 1372 SPDK_UNREACHABLE(); 1373 } 1374 1375 return rc; 1376 } 1377 1378 static int 1379 nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1380 { 1381 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1382 return addr_1 == addr_2; 1383 } 1384 1385 static inline void 1386 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1387 enum spdk_nvme_data_transfer xfer) 1388 { 1389 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1390 wr->opcode = IBV_WR_RDMA_WRITE; 1391 wr->send_flags = 0; 1392 wr->next = next; 1393 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1394 wr->opcode = IBV_WR_RDMA_READ; 1395 wr->send_flags = IBV_SEND_SIGNALED; 1396 wr->next = NULL; 1397 } else { 1398 assert(0); 1399 } 1400 } 1401 1402 static int 1403 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1404 struct spdk_nvmf_rdma_request *rdma_req, 1405 uint32_t num_sgl_descriptors) 1406 { 1407 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1408 struct spdk_nvmf_rdma_request_data *current_data_wr; 1409 uint32_t i; 1410 1411 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1412 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1413 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1414 return -EINVAL; 1415 } 1416 1417 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1418 return -ENOMEM; 1419 } 1420 1421 current_data_wr = &rdma_req->data; 1422 1423 for (i = 0; i < num_sgl_descriptors; i++) { 1424 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1425 current_data_wr->wr.next = &work_requests[i]->wr; 1426 current_data_wr = work_requests[i]; 1427 current_data_wr->wr.sg_list = current_data_wr->sgl; 1428 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1429 } 1430 1431 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1432 1433 return 0; 1434 } 1435 1436 static inline void 1437 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1438 { 1439 struct ibv_send_wr *wr = &rdma_req->data.wr; 1440 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1441 1442 wr->wr.rdma.rkey = sgl->keyed.key; 1443 wr->wr.rdma.remote_addr = sgl->address; 1444 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1445 } 1446 1447 static inline void 1448 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1449 { 1450 struct ibv_send_wr *wr = &rdma_req->data.wr; 1451 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1452 uint32_t i; 1453 int j; 1454 uint64_t remote_addr_offset = 0; 1455 1456 for (i = 0; i < num_wrs; ++i) { 1457 wr->wr.rdma.rkey = sgl->keyed.key; 1458 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1459 for (j = 0; j < wr->num_sge; ++j) { 1460 remote_addr_offset += wr->sg_list[j].length; 1461 } 1462 wr = wr->next; 1463 } 1464 } 1465 1466 /* This function is used in the rare case that we have a buffer split over multiple memory regions. */ 1467 static int 1468 nvmf_rdma_replace_buffer(struct spdk_nvmf_rdma_poll_group *rgroup, void **buf) 1469 { 1470 struct spdk_nvmf_transport_poll_group *group = &rgroup->group; 1471 struct spdk_nvmf_transport *transport = group->transport; 1472 struct spdk_nvmf_transport_pg_cache_buf *old_buf; 1473 void *new_buf; 1474 1475 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1476 group->buf_cache_count--; 1477 new_buf = STAILQ_FIRST(&group->buf_cache); 1478 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1479 assert(*buf != NULL); 1480 } else { 1481 new_buf = spdk_mempool_get(transport->data_buf_pool); 1482 } 1483 1484 if (*buf == NULL) { 1485 return -ENOMEM; 1486 } 1487 1488 old_buf = *buf; 1489 STAILQ_INSERT_HEAD(&rgroup->retired_bufs, old_buf, link); 1490 *buf = new_buf; 1491 return 0; 1492 } 1493 1494 static bool 1495 nvmf_rdma_get_lkey(struct spdk_nvmf_rdma_device *device, struct iovec *iov, 1496 uint32_t *_lkey) 1497 { 1498 uint64_t translation_len; 1499 uint32_t lkey; 1500 1501 translation_len = iov->iov_len; 1502 1503 if (!g_nvmf_hooks.get_rkey) { 1504 lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1505 (uint64_t)iov->iov_base, &translation_len))->lkey; 1506 } else { 1507 lkey = spdk_mem_map_translate(device->map, 1508 (uint64_t)iov->iov_base, &translation_len); 1509 } 1510 1511 if (spdk_unlikely(translation_len < iov->iov_len)) { 1512 return false; 1513 } 1514 1515 *_lkey = lkey; 1516 return true; 1517 } 1518 1519 static bool 1520 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device, 1521 struct iovec *iov, struct ibv_send_wr **_wr, 1522 uint32_t *_remaining_data_block, uint32_t *_offset, 1523 uint32_t *_num_extra_wrs, 1524 const struct spdk_dif_ctx *dif_ctx) 1525 { 1526 struct ibv_send_wr *wr = *_wr; 1527 struct ibv_sge *sg_ele = &wr->sg_list[wr->num_sge]; 1528 uint32_t lkey = 0; 1529 uint32_t remaining, data_block_size, md_size, sge_len; 1530 1531 if (spdk_unlikely(!nvmf_rdma_get_lkey(device, iov, &lkey))) { 1532 /* This is a very rare case that can occur when using DPDK version < 19.05 */ 1533 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions. Removing it from circulation.\n"); 1534 return false; 1535 } 1536 1537 if (spdk_likely(!dif_ctx)) { 1538 sg_ele->lkey = lkey; 1539 sg_ele->addr = (uintptr_t)(iov->iov_base); 1540 sg_ele->length = iov->iov_len; 1541 wr->num_sge++; 1542 } else { 1543 remaining = iov->iov_len - *_offset; 1544 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1545 md_size = dif_ctx->md_size; 1546 1547 while (remaining) { 1548 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1549 if (*_num_extra_wrs > 0 && wr->next) { 1550 *_wr = wr->next; 1551 wr = *_wr; 1552 wr->num_sge = 0; 1553 sg_ele = &wr->sg_list[wr->num_sge]; 1554 (*_num_extra_wrs)--; 1555 } else { 1556 break; 1557 } 1558 } 1559 sg_ele->lkey = lkey; 1560 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset); 1561 sge_len = spdk_min(remaining, *_remaining_data_block); 1562 sg_ele->length = sge_len; 1563 remaining -= sge_len; 1564 *_remaining_data_block -= sge_len; 1565 *_offset += sge_len; 1566 1567 sg_ele++; 1568 wr->num_sge++; 1569 1570 if (*_remaining_data_block == 0) { 1571 /* skip metadata */ 1572 *_offset += md_size; 1573 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1574 remaining -= spdk_min(remaining, md_size); 1575 *_remaining_data_block = data_block_size; 1576 } 1577 1578 if (remaining == 0) { 1579 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1580 the remaining metadata bits at the beginning of the next buffer */ 1581 *_offset -= iov->iov_len; 1582 } 1583 } 1584 } 1585 1586 return true; 1587 } 1588 1589 static int 1590 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1591 struct spdk_nvmf_rdma_device *device, 1592 struct spdk_nvmf_rdma_request *rdma_req, 1593 struct ibv_send_wr *wr, 1594 uint32_t length, 1595 uint32_t num_extra_wrs) 1596 { 1597 struct spdk_nvmf_request *req = &rdma_req->req; 1598 struct spdk_dif_ctx *dif_ctx = NULL; 1599 uint32_t remaining_data_block = 0; 1600 uint32_t offset = 0; 1601 1602 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 1603 dif_ctx = &rdma_req->req.dif.dif_ctx; 1604 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1605 } 1606 1607 wr->num_sge = 0; 1608 1609 while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1610 while (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr, 1611 &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) { 1612 if (nvmf_rdma_replace_buffer(rgroup, &req->buffers[rdma_req->iovpos]) == -ENOMEM) { 1613 return -ENOMEM; 1614 } 1615 req->iov[rdma_req->iovpos].iov_base = (void *)((uintptr_t)(req->buffers[rdma_req->iovpos] + 1616 NVMF_DATA_BUFFER_MASK) & 1617 ~NVMF_DATA_BUFFER_MASK); 1618 } 1619 1620 length -= req->iov[rdma_req->iovpos].iov_len; 1621 rdma_req->iovpos++; 1622 } 1623 1624 if (length) { 1625 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1626 return -EINVAL; 1627 } 1628 1629 return 0; 1630 } 1631 1632 static inline uint32_t 1633 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1634 { 1635 /* estimate the number of SG entries and WRs needed to process the request */ 1636 uint32_t num_sge = 0; 1637 uint32_t i; 1638 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1639 1640 for (i = 0; i < num_buffers && length > 0; i++) { 1641 uint32_t buffer_len = spdk_min(length, io_unit_size); 1642 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1643 1644 if (num_sge_in_block * block_size > buffer_len) { 1645 ++num_sge_in_block; 1646 } 1647 num_sge += num_sge_in_block; 1648 length -= buffer_len; 1649 } 1650 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1651 } 1652 1653 static int 1654 nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1655 struct spdk_nvmf_rdma_device *device, 1656 struct spdk_nvmf_rdma_request *rdma_req, 1657 uint32_t length) 1658 { 1659 struct spdk_nvmf_rdma_qpair *rqpair; 1660 struct spdk_nvmf_rdma_poll_group *rgroup; 1661 struct spdk_nvmf_request *req = &rdma_req->req; 1662 struct ibv_send_wr *wr = &rdma_req->data.wr; 1663 int rc; 1664 uint32_t num_wrs = 1; 1665 1666 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1667 rgroup = rqpair->poller->group; 1668 1669 /* rdma wr specifics */ 1670 nvmf_rdma_setup_request(rdma_req); 1671 1672 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1673 length); 1674 if (rc != 0) { 1675 return rc; 1676 } 1677 1678 assert(req->iovcnt <= rqpair->max_send_sge); 1679 1680 rdma_req->iovpos = 0; 1681 1682 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1683 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1684 req->dif.dif_ctx.block_size); 1685 if (num_wrs > 1) { 1686 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1687 if (rc != 0) { 1688 goto err_exit; 1689 } 1690 } 1691 } 1692 1693 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1694 if (spdk_unlikely(rc != 0)) { 1695 goto err_exit; 1696 } 1697 1698 if (spdk_unlikely(num_wrs > 1)) { 1699 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1700 } 1701 1702 /* set the number of outstanding data WRs for this request. */ 1703 rdma_req->num_outstanding_data_wr = num_wrs; 1704 1705 return rc; 1706 1707 err_exit: 1708 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1709 nvmf_rdma_request_free_data(rdma_req, rtransport); 1710 req->iovcnt = 0; 1711 return rc; 1712 } 1713 1714 static int 1715 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1716 struct spdk_nvmf_rdma_device *device, 1717 struct spdk_nvmf_rdma_request *rdma_req) 1718 { 1719 struct spdk_nvmf_rdma_qpair *rqpair; 1720 struct spdk_nvmf_rdma_poll_group *rgroup; 1721 struct ibv_send_wr *current_wr; 1722 struct spdk_nvmf_request *req = &rdma_req->req; 1723 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1724 uint32_t num_sgl_descriptors; 1725 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES]; 1726 uint32_t i; 1727 int rc; 1728 1729 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1730 rgroup = rqpair->poller->group; 1731 1732 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1733 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1734 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1735 1736 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1737 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1738 1739 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1740 return -ENOMEM; 1741 } 1742 1743 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1744 for (i = 0; i < num_sgl_descriptors; i++) { 1745 if (spdk_likely(!req->dif.dif_insert_or_strip)) { 1746 lengths[i] = desc->keyed.length; 1747 } else { 1748 req->dif.orig_length += desc->keyed.length; 1749 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1750 req->dif.elba_length += lengths[i]; 1751 } 1752 desc++; 1753 } 1754 1755 rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport, 1756 lengths, num_sgl_descriptors); 1757 if (rc != 0) { 1758 nvmf_rdma_request_free_data(rdma_req, rtransport); 1759 return rc; 1760 } 1761 1762 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1763 current_wr = &rdma_req->data.wr; 1764 assert(current_wr != NULL); 1765 1766 req->length = 0; 1767 rdma_req->iovpos = 0; 1768 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1769 for (i = 0; i < num_sgl_descriptors; i++) { 1770 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1771 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1772 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1773 rc = -EINVAL; 1774 goto err_exit; 1775 } 1776 1777 current_wr->num_sge = 0; 1778 1779 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1780 if (rc != 0) { 1781 rc = -ENOMEM; 1782 goto err_exit; 1783 } 1784 1785 req->length += desc->keyed.length; 1786 current_wr->wr.rdma.rkey = desc->keyed.key; 1787 current_wr->wr.rdma.remote_addr = desc->address; 1788 current_wr = current_wr->next; 1789 desc++; 1790 } 1791 1792 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1793 /* Go back to the last descriptor in the list. */ 1794 desc--; 1795 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1796 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1797 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1798 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1799 } 1800 } 1801 #endif 1802 1803 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1804 1805 return 0; 1806 1807 err_exit: 1808 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1809 nvmf_rdma_request_free_data(rdma_req, rtransport); 1810 return rc; 1811 } 1812 1813 static int 1814 nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1815 struct spdk_nvmf_rdma_device *device, 1816 struct spdk_nvmf_rdma_request *rdma_req) 1817 { 1818 struct spdk_nvmf_request *req = &rdma_req->req; 1819 struct spdk_nvme_cpl *rsp; 1820 struct spdk_nvme_sgl_descriptor *sgl; 1821 int rc; 1822 uint32_t length; 1823 1824 rsp = &req->rsp->nvme_cpl; 1825 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1826 1827 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1828 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1829 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1830 1831 length = sgl->keyed.length; 1832 if (length > rtransport->transport.opts.max_io_size) { 1833 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1834 length, rtransport->transport.opts.max_io_size); 1835 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1836 return -1; 1837 } 1838 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1839 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1840 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1841 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1842 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1843 } 1844 } 1845 #endif 1846 1847 /* fill request length and populate iovs */ 1848 req->length = length; 1849 1850 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1851 req->dif.orig_length = length; 1852 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1853 req->dif.elba_length = length; 1854 } 1855 1856 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1857 if (spdk_unlikely(rc < 0)) { 1858 if (rc == -EINVAL) { 1859 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1860 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1861 return -1; 1862 } 1863 /* No available buffers. Queue this request up. */ 1864 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1865 return 0; 1866 } 1867 1868 /* backward compatible */ 1869 req->data = req->iov[0].iov_base; 1870 1871 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1872 req->iovcnt); 1873 1874 return 0; 1875 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1876 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1877 uint64_t offset = sgl->address; 1878 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1879 1880 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1881 offset, sgl->unkeyed.length); 1882 1883 if (offset > max_len) { 1884 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1885 offset, max_len); 1886 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1887 return -1; 1888 } 1889 max_len -= (uint32_t)offset; 1890 1891 if (sgl->unkeyed.length > max_len) { 1892 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1893 sgl->unkeyed.length, max_len); 1894 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1895 return -1; 1896 } 1897 1898 rdma_req->num_outstanding_data_wr = 0; 1899 req->data = rdma_req->recv->buf + offset; 1900 req->data_from_pool = false; 1901 req->length = sgl->unkeyed.length; 1902 1903 req->iov[0].iov_base = req->data; 1904 req->iov[0].iov_len = req->length; 1905 req->iovcnt = 1; 1906 1907 return 0; 1908 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1909 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1910 1911 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1912 if (rc == -ENOMEM) { 1913 SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req); 1914 return 0; 1915 } else if (rc == -EINVAL) { 1916 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1917 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1918 return -1; 1919 } 1920 1921 /* backward compatible */ 1922 req->data = req->iov[0].iov_base; 1923 1924 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req, 1925 req->iovcnt); 1926 1927 return 0; 1928 } 1929 1930 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1931 sgl->generic.type, sgl->generic.subtype); 1932 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1933 return -1; 1934 } 1935 1936 static void 1937 _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1938 struct spdk_nvmf_rdma_transport *rtransport) 1939 { 1940 struct spdk_nvmf_rdma_qpair *rqpair; 1941 struct spdk_nvmf_rdma_poll_group *rgroup; 1942 1943 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1944 if (rdma_req->req.data_from_pool) { 1945 rgroup = rqpair->poller->group; 1946 1947 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1948 } 1949 nvmf_rdma_request_free_data(rdma_req, rtransport); 1950 rdma_req->req.length = 0; 1951 rdma_req->req.iovcnt = 0; 1952 rdma_req->req.data = NULL; 1953 rdma_req->rsp.wr.next = NULL; 1954 rdma_req->data.wr.next = NULL; 1955 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1956 rqpair->qd--; 1957 1958 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1959 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1960 } 1961 1962 bool 1963 nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 1964 struct spdk_nvmf_rdma_request *rdma_req) 1965 { 1966 struct spdk_nvmf_rdma_qpair *rqpair; 1967 struct spdk_nvmf_rdma_device *device; 1968 struct spdk_nvmf_rdma_poll_group *rgroup; 1969 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 1970 int rc; 1971 struct spdk_nvmf_rdma_recv *rdma_recv; 1972 enum spdk_nvmf_rdma_request_state prev_state; 1973 bool progress = false; 1974 int data_posted; 1975 uint32_t num_blocks; 1976 1977 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1978 device = rqpair->device; 1979 rgroup = rqpair->poller->group; 1980 1981 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 1982 1983 /* If the queue pair is in an error state, force the request to the completed state 1984 * to release resources. */ 1985 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 1986 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 1987 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 1988 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 1989 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1990 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 1991 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 1992 } 1993 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 1994 } 1995 1996 /* The loop here is to allow for several back-to-back state changes. */ 1997 do { 1998 prev_state = rdma_req->state; 1999 2000 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state); 2001 2002 switch (rdma_req->state) { 2003 case RDMA_REQUEST_STATE_FREE: 2004 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 2005 * to escape this state. */ 2006 break; 2007 case RDMA_REQUEST_STATE_NEW: 2008 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 2009 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2010 rdma_recv = rdma_req->recv; 2011 2012 /* The first element of the SGL is the NVMe command */ 2013 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 2014 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 2015 2016 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2017 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2018 break; 2019 } 2020 2021 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 2022 rdma_req->req.dif.dif_insert_or_strip = true; 2023 } 2024 2025 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2026 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 2027 rdma_req->rsp.wr.imm_data = 0; 2028 #endif 2029 2030 /* The next state transition depends on the data transfer needs of this request. */ 2031 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2032 2033 if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2034 rsp->status.sct = SPDK_NVME_SCT_GENERIC; 2035 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2036 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2037 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req); 2038 break; 2039 } 2040 2041 /* If no data to transfer, ready to execute. */ 2042 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2043 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2044 break; 2045 } 2046 2047 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2048 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2049 break; 2050 case RDMA_REQUEST_STATE_NEED_BUFFER: 2051 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2052 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2053 2054 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2055 2056 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2057 /* This request needs to wait in line to obtain a buffer */ 2058 break; 2059 } 2060 2061 /* Try to get a data buffer */ 2062 rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2063 if (rc < 0) { 2064 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2065 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2066 break; 2067 } 2068 2069 if (!rdma_req->req.data) { 2070 /* No buffers available. */ 2071 rgroup->stat.pending_data_buffer++; 2072 break; 2073 } 2074 2075 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2076 2077 /* If data is transferring from host to controller and the data didn't 2078 * arrive using in capsule data, we need to do a transfer from the host. 2079 */ 2080 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2081 rdma_req->req.data_from_pool) { 2082 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2083 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2084 break; 2085 } 2086 2087 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2088 break; 2089 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2090 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2091 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2092 2093 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2094 /* This request needs to wait in line to perform RDMA */ 2095 break; 2096 } 2097 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2098 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2099 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2100 rqpair->poller->stat.pending_rdma_read++; 2101 break; 2102 } 2103 2104 /* We have already verified that this request is the head of the queue. */ 2105 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2106 2107 rc = request_transfer_in(&rdma_req->req); 2108 if (!rc) { 2109 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2110 } else { 2111 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2112 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2113 } 2114 break; 2115 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2116 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2117 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2118 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2119 * to escape this state. */ 2120 break; 2121 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2122 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2123 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2124 2125 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2126 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2127 /* generate DIF for write operation */ 2128 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2129 assert(num_blocks > 0); 2130 2131 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2132 num_blocks, &rdma_req->req.dif.dif_ctx); 2133 if (rc != 0) { 2134 SPDK_ERRLOG("DIF generation failed\n"); 2135 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2136 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2137 break; 2138 } 2139 } 2140 2141 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2142 /* set extended length before IO operation */ 2143 rdma_req->req.length = rdma_req->req.dif.elba_length; 2144 } 2145 2146 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2147 spdk_nvmf_request_exec(&rdma_req->req); 2148 break; 2149 case RDMA_REQUEST_STATE_EXECUTING: 2150 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2151 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2152 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2153 * to escape this state. */ 2154 break; 2155 case RDMA_REQUEST_STATE_EXECUTED: 2156 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2157 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2158 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 2159 rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2160 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2161 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2162 } else { 2163 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2164 } 2165 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2166 /* restore the original length */ 2167 rdma_req->req.length = rdma_req->req.dif.orig_length; 2168 2169 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2170 struct spdk_dif_error error_blk; 2171 2172 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2173 2174 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2175 &rdma_req->req.dif.dif_ctx, &error_blk); 2176 if (rc) { 2177 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2178 2179 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2180 error_blk.err_offset); 2181 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2182 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2183 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2184 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2185 } 2186 } 2187 } 2188 break; 2189 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2190 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2191 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2192 2193 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2194 /* This request needs to wait in line to perform RDMA */ 2195 break; 2196 } 2197 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2198 rqpair->max_send_depth) { 2199 /* We can only have so many WRs outstanding. we have to wait until some finish. 2200 * +1 since each request has an additional wr in the resp. */ 2201 rqpair->poller->stat.pending_rdma_write++; 2202 break; 2203 } 2204 2205 /* We have already verified that this request is the head of the queue. */ 2206 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2207 2208 /* The data transfer will be kicked off from 2209 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2210 */ 2211 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2212 break; 2213 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2214 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2215 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2216 rc = request_transfer_out(&rdma_req->req, &data_posted); 2217 assert(rc == 0); /* No good way to handle this currently */ 2218 if (rc) { 2219 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2220 } else { 2221 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2222 RDMA_REQUEST_STATE_COMPLETING; 2223 } 2224 break; 2225 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2226 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2227 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2228 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2229 * to escape this state. */ 2230 break; 2231 case RDMA_REQUEST_STATE_COMPLETING: 2232 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2233 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2234 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2235 * to escape this state. */ 2236 break; 2237 case RDMA_REQUEST_STATE_COMPLETED: 2238 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2239 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2240 2241 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2242 _nvmf_rdma_request_free(rdma_req, rtransport); 2243 break; 2244 case RDMA_REQUEST_NUM_STATES: 2245 default: 2246 assert(0); 2247 break; 2248 } 2249 2250 if (rdma_req->state != prev_state) { 2251 progress = true; 2252 } 2253 } while (rdma_req->state != prev_state); 2254 2255 return progress; 2256 } 2257 2258 /* Public API callbacks begin here */ 2259 2260 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2261 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2262 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2263 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2264 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2265 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2266 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2267 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2268 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2269 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2270 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2271 #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100 2272 #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1 2273 #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false 2274 2275 static void 2276 nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2277 { 2278 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2279 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2280 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2281 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2282 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2283 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2284 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2285 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2286 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2287 opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC; 2288 opts->transport_specific = NULL; 2289 } 2290 2291 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2292 .notify_cb = nvmf_rdma_mem_notify, 2293 .are_contiguous = nvmf_rdma_check_contiguous_entries 2294 }; 2295 2296 static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2297 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg); 2298 2299 static inline bool 2300 nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device) 2301 { 2302 return device->attr.vendor_id == NVMF_RXE_VENDOR_ID_OLD || 2303 device->attr.vendor_id == NVMF_RXE_VENDOR_ID_NEW; 2304 } 2305 2306 static struct spdk_nvmf_transport * 2307 nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2308 { 2309 int rc; 2310 struct spdk_nvmf_rdma_transport *rtransport; 2311 struct spdk_nvmf_rdma_device *device, *tmp; 2312 struct ibv_context **contexts; 2313 uint32_t i; 2314 int flag; 2315 uint32_t sge_count; 2316 uint32_t min_shared_buffers; 2317 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2318 pthread_mutexattr_t attr; 2319 2320 rtransport = calloc(1, sizeof(*rtransport)); 2321 if (!rtransport) { 2322 return NULL; 2323 } 2324 2325 if (pthread_mutexattr_init(&attr)) { 2326 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2327 free(rtransport); 2328 return NULL; 2329 } 2330 2331 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2332 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2333 pthread_mutexattr_destroy(&attr); 2334 free(rtransport); 2335 return NULL; 2336 } 2337 2338 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2339 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2340 pthread_mutexattr_destroy(&attr); 2341 free(rtransport); 2342 return NULL; 2343 } 2344 2345 pthread_mutexattr_destroy(&attr); 2346 2347 TAILQ_INIT(&rtransport->devices); 2348 TAILQ_INIT(&rtransport->ports); 2349 TAILQ_INIT(&rtransport->poll_groups); 2350 2351 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2352 rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2353 rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2354 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2355 rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING; 2356 if (opts->transport_specific != NULL && 2357 spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder, 2358 SPDK_COUNTOF(rdma_transport_opts_decoder), 2359 &rtransport->rdma_opts)) { 2360 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 2361 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2362 return NULL; 2363 } 2364 2365 SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n" 2366 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2367 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2368 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2369 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d," 2370 " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n", 2371 opts->max_queue_depth, 2372 opts->max_io_size, 2373 opts->max_qpairs_per_ctrlr - 1, 2374 opts->io_unit_size, 2375 opts->in_capsule_data_size, 2376 opts->max_aq_depth, 2377 opts->num_shared_buffers, 2378 rtransport->rdma_opts.max_srq_depth, 2379 rtransport->rdma_opts.no_srq, 2380 rtransport->rdma_opts.acceptor_backlog, 2381 rtransport->rdma_opts.no_wr_batching, 2382 opts->abort_timeout_sec); 2383 2384 /* I/O unit size cannot be larger than max I/O size */ 2385 if (opts->io_unit_size > opts->max_io_size) { 2386 opts->io_unit_size = opts->max_io_size; 2387 } 2388 2389 if (rtransport->rdma_opts.acceptor_backlog <= 0) { 2390 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n", 2391 SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG); 2392 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG; 2393 } 2394 2395 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2396 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2397 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2398 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2399 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2400 return NULL; 2401 } 2402 2403 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 2404 if (min_shared_buffers > opts->num_shared_buffers) { 2405 SPDK_ERRLOG("There are not enough buffers to satisfy" 2406 "per-poll group caches for each thread. (%" PRIu32 ")" 2407 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2408 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2409 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2410 return NULL; 2411 } 2412 2413 sge_count = opts->max_io_size / opts->io_unit_size; 2414 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2415 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2416 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2417 return NULL; 2418 } 2419 2420 rtransport->event_channel = rdma_create_event_channel(); 2421 if (rtransport->event_channel == NULL) { 2422 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2423 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2424 return NULL; 2425 } 2426 2427 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2428 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2429 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2430 rtransport->event_channel->fd, spdk_strerror(errno)); 2431 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2432 return NULL; 2433 } 2434 2435 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2436 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2437 sizeof(struct spdk_nvmf_rdma_request_data), 2438 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2439 SPDK_ENV_SOCKET_ID_ANY); 2440 if (!rtransport->data_wr_pool) { 2441 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2442 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2443 return NULL; 2444 } 2445 2446 contexts = rdma_get_devices(NULL); 2447 if (contexts == NULL) { 2448 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2449 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2450 return NULL; 2451 } 2452 2453 i = 0; 2454 rc = 0; 2455 while (contexts[i] != NULL) { 2456 device = calloc(1, sizeof(*device)); 2457 if (!device) { 2458 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2459 rc = -ENOMEM; 2460 break; 2461 } 2462 device->context = contexts[i]; 2463 rc = ibv_query_device(device->context, &device->attr); 2464 if (rc < 0) { 2465 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2466 free(device); 2467 break; 2468 2469 } 2470 2471 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2472 2473 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2474 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2475 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2476 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2477 } 2478 2479 /** 2480 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2481 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2482 * but incorrectly reports that it does. There are changes making their way 2483 * through the kernel now that will enable this feature. When they are merged, 2484 * we can conditionally enable this feature. 2485 * 2486 * TODO: enable this for versions of the kernel rxe driver that support it. 2487 */ 2488 if (nvmf_rdma_is_rxe_device(device)) { 2489 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2490 } 2491 #endif 2492 2493 /* set up device context async ev fd as NON_BLOCKING */ 2494 flag = fcntl(device->context->async_fd, F_GETFL); 2495 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2496 if (rc < 0) { 2497 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2498 free(device); 2499 break; 2500 } 2501 2502 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2503 i++; 2504 2505 if (g_nvmf_hooks.get_ibv_pd) { 2506 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2507 } else { 2508 device->pd = ibv_alloc_pd(device->context); 2509 } 2510 2511 if (!device->pd) { 2512 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2513 rc = -ENOMEM; 2514 break; 2515 } 2516 2517 assert(device->map == NULL); 2518 2519 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2520 if (!device->map) { 2521 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2522 rc = -ENOMEM; 2523 break; 2524 } 2525 2526 assert(device->map != NULL); 2527 assert(device->pd != NULL); 2528 } 2529 rdma_free_devices(contexts); 2530 2531 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2532 /* divide and round up. */ 2533 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2534 2535 /* round up to the nearest 4k. */ 2536 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2537 2538 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2539 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2540 opts->io_unit_size); 2541 } 2542 2543 if (rc < 0) { 2544 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2545 return NULL; 2546 } 2547 2548 /* Set up poll descriptor array to monitor events from RDMA and IB 2549 * in a single poll syscall 2550 */ 2551 rtransport->npoll_fds = i + 1; 2552 i = 0; 2553 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2554 if (rtransport->poll_fds == NULL) { 2555 SPDK_ERRLOG("poll_fds allocation failed\n"); 2556 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL); 2557 return NULL; 2558 } 2559 2560 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2561 rtransport->poll_fds[i++].events = POLLIN; 2562 2563 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2564 rtransport->poll_fds[i].fd = device->context->async_fd; 2565 rtransport->poll_fds[i++].events = POLLIN; 2566 } 2567 2568 return &rtransport->transport; 2569 } 2570 2571 static void 2572 nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 2573 { 2574 struct spdk_nvmf_rdma_transport *rtransport; 2575 assert(w != NULL); 2576 2577 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2578 spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth); 2579 spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq); 2580 spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog); 2581 spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching); 2582 } 2583 2584 static int 2585 nvmf_rdma_destroy(struct spdk_nvmf_transport *transport, 2586 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 2587 { 2588 struct spdk_nvmf_rdma_transport *rtransport; 2589 struct spdk_nvmf_rdma_port *port, *port_tmp; 2590 struct spdk_nvmf_rdma_device *device, *device_tmp; 2591 2592 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2593 2594 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2595 TAILQ_REMOVE(&rtransport->ports, port, link); 2596 rdma_destroy_id(port->id); 2597 free(port); 2598 } 2599 2600 if (rtransport->poll_fds != NULL) { 2601 free(rtransport->poll_fds); 2602 } 2603 2604 if (rtransport->event_channel != NULL) { 2605 rdma_destroy_event_channel(rtransport->event_channel); 2606 } 2607 2608 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2609 TAILQ_REMOVE(&rtransport->devices, device, link); 2610 if (device->map) { 2611 spdk_mem_map_free(&device->map); 2612 } 2613 if (device->pd) { 2614 if (!g_nvmf_hooks.get_ibv_pd) { 2615 ibv_dealloc_pd(device->pd); 2616 } 2617 } 2618 free(device); 2619 } 2620 2621 if (rtransport->data_wr_pool != NULL) { 2622 if (spdk_mempool_count(rtransport->data_wr_pool) != 2623 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2624 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2625 spdk_mempool_count(rtransport->data_wr_pool), 2626 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2627 } 2628 } 2629 2630 spdk_mempool_free(rtransport->data_wr_pool); 2631 2632 pthread_mutex_destroy(&rtransport->lock); 2633 free(rtransport); 2634 2635 if (cb_fn) { 2636 cb_fn(cb_arg); 2637 } 2638 return 0; 2639 } 2640 2641 static int 2642 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2643 struct spdk_nvme_transport_id *trid, 2644 bool peer); 2645 2646 static int 2647 nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2648 const struct spdk_nvme_transport_id *trid) 2649 { 2650 struct spdk_nvmf_rdma_transport *rtransport; 2651 struct spdk_nvmf_rdma_device *device; 2652 struct spdk_nvmf_rdma_port *port; 2653 struct addrinfo *res; 2654 struct addrinfo hints; 2655 int family; 2656 int rc; 2657 2658 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2659 assert(rtransport->event_channel != NULL); 2660 2661 pthread_mutex_lock(&rtransport->lock); 2662 port = calloc(1, sizeof(*port)); 2663 if (!port) { 2664 SPDK_ERRLOG("Port allocation failed\n"); 2665 pthread_mutex_unlock(&rtransport->lock); 2666 return -ENOMEM; 2667 } 2668 2669 port->trid = trid; 2670 2671 switch (trid->adrfam) { 2672 case SPDK_NVMF_ADRFAM_IPV4: 2673 family = AF_INET; 2674 break; 2675 case SPDK_NVMF_ADRFAM_IPV6: 2676 family = AF_INET6; 2677 break; 2678 default: 2679 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam); 2680 free(port); 2681 pthread_mutex_unlock(&rtransport->lock); 2682 return -EINVAL; 2683 } 2684 2685 memset(&hints, 0, sizeof(hints)); 2686 hints.ai_family = family; 2687 hints.ai_flags = AI_NUMERICSERV; 2688 hints.ai_socktype = SOCK_STREAM; 2689 hints.ai_protocol = 0; 2690 2691 rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res); 2692 if (rc) { 2693 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2694 free(port); 2695 pthread_mutex_unlock(&rtransport->lock); 2696 return -EINVAL; 2697 } 2698 2699 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2700 if (rc < 0) { 2701 SPDK_ERRLOG("rdma_create_id() failed\n"); 2702 freeaddrinfo(res); 2703 free(port); 2704 pthread_mutex_unlock(&rtransport->lock); 2705 return rc; 2706 } 2707 2708 rc = rdma_bind_addr(port->id, res->ai_addr); 2709 freeaddrinfo(res); 2710 2711 if (rc < 0) { 2712 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2713 rdma_destroy_id(port->id); 2714 free(port); 2715 pthread_mutex_unlock(&rtransport->lock); 2716 return rc; 2717 } 2718 2719 if (!port->id->verbs) { 2720 SPDK_ERRLOG("ibv_context is null\n"); 2721 rdma_destroy_id(port->id); 2722 free(port); 2723 pthread_mutex_unlock(&rtransport->lock); 2724 return -1; 2725 } 2726 2727 rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog); 2728 if (rc < 0) { 2729 SPDK_ERRLOG("rdma_listen() failed\n"); 2730 rdma_destroy_id(port->id); 2731 free(port); 2732 pthread_mutex_unlock(&rtransport->lock); 2733 return rc; 2734 } 2735 2736 TAILQ_FOREACH(device, &rtransport->devices, link) { 2737 if (device->context == port->id->verbs) { 2738 port->device = device; 2739 break; 2740 } 2741 } 2742 if (!port->device) { 2743 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2744 port->id->verbs); 2745 rdma_destroy_id(port->id); 2746 free(port); 2747 pthread_mutex_unlock(&rtransport->lock); 2748 return -EINVAL; 2749 } 2750 2751 SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n", 2752 trid->traddr, trid->trsvcid); 2753 2754 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2755 pthread_mutex_unlock(&rtransport->lock); 2756 return 0; 2757 } 2758 2759 static void 2760 nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2761 const struct spdk_nvme_transport_id *trid) 2762 { 2763 struct spdk_nvmf_rdma_transport *rtransport; 2764 struct spdk_nvmf_rdma_port *port, *tmp; 2765 2766 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2767 2768 pthread_mutex_lock(&rtransport->lock); 2769 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2770 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) { 2771 TAILQ_REMOVE(&rtransport->ports, port, link); 2772 rdma_destroy_id(port->id); 2773 free(port); 2774 break; 2775 } 2776 } 2777 2778 pthread_mutex_unlock(&rtransport->lock); 2779 } 2780 2781 static void 2782 nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2783 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2784 { 2785 struct spdk_nvmf_request *req, *tmp; 2786 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2787 struct spdk_nvmf_rdma_resources *resources; 2788 2789 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2790 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2791 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2792 break; 2793 } 2794 } 2795 2796 /* Then RDMA writes since reads have stronger restrictions than writes */ 2797 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2798 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2799 break; 2800 } 2801 } 2802 2803 /* The second highest priority is I/O waiting on memory buffers. */ 2804 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2805 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2806 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2807 break; 2808 } 2809 } 2810 2811 resources = rqpair->resources; 2812 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2813 rdma_req = STAILQ_FIRST(&resources->free_queue); 2814 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2815 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2816 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2817 2818 if (rqpair->srq != NULL) { 2819 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2820 rdma_req->recv->qpair->qd++; 2821 } else { 2822 rqpair->qd++; 2823 } 2824 2825 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2826 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2827 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2828 break; 2829 } 2830 } 2831 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2832 rqpair->poller->stat.pending_free_request++; 2833 } 2834 } 2835 2836 static void 2837 nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair) 2838 { 2839 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2840 struct spdk_nvmf_rdma_transport, transport); 2841 2842 nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2843 2844 /* nvmr_rdma_close_qpair is not called */ 2845 if (!rqpair->to_close) { 2846 return; 2847 } 2848 2849 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2850 if (rqpair->current_send_depth != 0) { 2851 return; 2852 } 2853 2854 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2855 return; 2856 } 2857 2858 /* Judge whether the device is emulated by Software RoCE. 2859 * And it will not send last_wqe event 2860 */ 2861 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false && 2862 !nvmf_rdma_is_rxe_device(rqpair->device)) { 2863 return; 2864 } 2865 2866 assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR); 2867 2868 nvmf_rdma_qpair_destroy(rqpair); 2869 } 2870 2871 static int 2872 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2873 { 2874 struct spdk_nvmf_qpair *qpair; 2875 struct spdk_nvmf_rdma_qpair *rqpair; 2876 2877 if (evt->id == NULL) { 2878 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2879 return -1; 2880 } 2881 2882 qpair = evt->id->context; 2883 if (qpair == NULL) { 2884 SPDK_ERRLOG("disconnect request: no active connection\n"); 2885 return -1; 2886 } 2887 2888 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2889 2890 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2891 2892 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2893 2894 return 0; 2895 } 2896 2897 #ifdef DEBUG 2898 static const char *CM_EVENT_STR[] = { 2899 "RDMA_CM_EVENT_ADDR_RESOLVED", 2900 "RDMA_CM_EVENT_ADDR_ERROR", 2901 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2902 "RDMA_CM_EVENT_ROUTE_ERROR", 2903 "RDMA_CM_EVENT_CONNECT_REQUEST", 2904 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2905 "RDMA_CM_EVENT_CONNECT_ERROR", 2906 "RDMA_CM_EVENT_UNREACHABLE", 2907 "RDMA_CM_EVENT_REJECTED", 2908 "RDMA_CM_EVENT_ESTABLISHED", 2909 "RDMA_CM_EVENT_DISCONNECTED", 2910 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2911 "RDMA_CM_EVENT_MULTICAST_JOIN", 2912 "RDMA_CM_EVENT_MULTICAST_ERROR", 2913 "RDMA_CM_EVENT_ADDR_CHANGE", 2914 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2915 }; 2916 #endif /* DEBUG */ 2917 2918 static void 2919 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport, 2920 struct spdk_nvmf_rdma_port *port) 2921 { 2922 struct spdk_nvmf_rdma_poll_group *rgroup; 2923 struct spdk_nvmf_rdma_poller *rpoller; 2924 struct spdk_nvmf_rdma_qpair *rqpair; 2925 2926 TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) { 2927 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 2928 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 2929 if (rqpair->listen_id == port->id) { 2930 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 2931 } 2932 } 2933 } 2934 } 2935 } 2936 2937 static bool 2938 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport, 2939 struct rdma_cm_event *event) 2940 { 2941 const struct spdk_nvme_transport_id *trid; 2942 struct spdk_nvmf_rdma_port *port; 2943 struct spdk_nvmf_rdma_transport *rtransport; 2944 bool event_acked = false; 2945 2946 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2947 TAILQ_FOREACH(port, &rtransport->ports, link) { 2948 if (port->id == event->id) { 2949 SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid); 2950 rdma_ack_cm_event(event); 2951 event_acked = true; 2952 trid = port->trid; 2953 break; 2954 } 2955 } 2956 2957 if (event_acked) { 2958 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2959 2960 nvmf_rdma_stop_listen(transport, trid); 2961 nvmf_rdma_listen(transport, trid); 2962 } 2963 2964 return event_acked; 2965 } 2966 2967 static void 2968 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport, 2969 struct rdma_cm_event *event) 2970 { 2971 struct spdk_nvmf_rdma_port *port; 2972 struct spdk_nvmf_rdma_transport *rtransport; 2973 2974 port = event->id->context; 2975 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2976 2977 SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid); 2978 2979 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port); 2980 2981 rdma_ack_cm_event(event); 2982 2983 while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) { 2984 ; 2985 } 2986 } 2987 2988 static void 2989 nvmf_process_cm_event(struct spdk_nvmf_transport *transport) 2990 { 2991 struct spdk_nvmf_rdma_transport *rtransport; 2992 struct rdma_cm_event *event; 2993 int rc; 2994 bool event_acked; 2995 2996 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2997 2998 if (rtransport->event_channel == NULL) { 2999 return; 3000 } 3001 3002 while (1) { 3003 event_acked = false; 3004 rc = rdma_get_cm_event(rtransport->event_channel, &event); 3005 if (rc) { 3006 if (errno != EAGAIN && errno != EWOULDBLOCK) { 3007 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 3008 } 3009 break; 3010 } 3011 3012 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 3013 3014 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 3015 3016 switch (event->event) { 3017 case RDMA_CM_EVENT_ADDR_RESOLVED: 3018 case RDMA_CM_EVENT_ADDR_ERROR: 3019 case RDMA_CM_EVENT_ROUTE_RESOLVED: 3020 case RDMA_CM_EVENT_ROUTE_ERROR: 3021 /* No action required. The target never attempts to resolve routes. */ 3022 break; 3023 case RDMA_CM_EVENT_CONNECT_REQUEST: 3024 rc = nvmf_rdma_connect(transport, event); 3025 if (rc < 0) { 3026 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 3027 break; 3028 } 3029 break; 3030 case RDMA_CM_EVENT_CONNECT_RESPONSE: 3031 /* The target never initiates a new connection. So this will not occur. */ 3032 break; 3033 case RDMA_CM_EVENT_CONNECT_ERROR: 3034 /* Can this happen? The docs say it can, but not sure what causes it. */ 3035 break; 3036 case RDMA_CM_EVENT_UNREACHABLE: 3037 case RDMA_CM_EVENT_REJECTED: 3038 /* These only occur on the client side. */ 3039 break; 3040 case RDMA_CM_EVENT_ESTABLISHED: 3041 /* TODO: Should we be waiting for this event anywhere? */ 3042 break; 3043 case RDMA_CM_EVENT_DISCONNECTED: 3044 rc = nvmf_rdma_disconnect(event); 3045 if (rc < 0) { 3046 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3047 break; 3048 } 3049 break; 3050 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3051 /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL 3052 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s. 3053 * Once these events are sent to SPDK, we should release all IB resources and 3054 * don't make attempts to call any ibv_query/modify/create functions. We can only call 3055 * ibv_destory* functions to release user space memory allocated by IB. All kernel 3056 * resources are already cleaned. */ 3057 if (event->id->qp) { 3058 /* If rdma_cm event has a valid `qp` pointer then the event refers to the 3059 * corresponding qpair. Otherwise the event refers to a listening device */ 3060 rc = nvmf_rdma_disconnect(event); 3061 if (rc < 0) { 3062 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3063 break; 3064 } 3065 } else { 3066 nvmf_rdma_handle_cm_event_port_removal(transport, event); 3067 event_acked = true; 3068 } 3069 break; 3070 case RDMA_CM_EVENT_MULTICAST_JOIN: 3071 case RDMA_CM_EVENT_MULTICAST_ERROR: 3072 /* Multicast is not used */ 3073 break; 3074 case RDMA_CM_EVENT_ADDR_CHANGE: 3075 event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event); 3076 break; 3077 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3078 /* For now, do nothing. The target never re-uses queue pairs. */ 3079 break; 3080 default: 3081 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3082 break; 3083 } 3084 if (!event_acked) { 3085 rdma_ack_cm_event(event); 3086 } 3087 } 3088 } 3089 3090 static void 3091 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3092 { 3093 rqpair->last_wqe_reached = true; 3094 nvmf_rdma_destroy_drained_qpair(rqpair); 3095 } 3096 3097 static void 3098 nvmf_rdma_qpair_process_ibv_event(void *ctx) 3099 { 3100 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3101 3102 if (event_ctx->rqpair) { 3103 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3104 if (event_ctx->cb_fn) { 3105 event_ctx->cb_fn(event_ctx->rqpair); 3106 } 3107 } 3108 free(event_ctx); 3109 } 3110 3111 static int 3112 nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3113 spdk_nvmf_rdma_qpair_ibv_event fn) 3114 { 3115 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3116 struct spdk_thread *thr = NULL; 3117 int rc; 3118 3119 if (rqpair->qpair.group) { 3120 thr = rqpair->qpair.group->thread; 3121 } else if (rqpair->destruct_channel) { 3122 thr = spdk_io_channel_get_thread(rqpair->destruct_channel); 3123 } 3124 3125 if (!thr) { 3126 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair); 3127 return -EINVAL; 3128 } 3129 3130 ctx = calloc(1, sizeof(*ctx)); 3131 if (!ctx) { 3132 return -ENOMEM; 3133 } 3134 3135 ctx->rqpair = rqpair; 3136 ctx->cb_fn = fn; 3137 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3138 3139 rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx); 3140 if (rc) { 3141 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3142 free(ctx); 3143 } 3144 3145 return rc; 3146 } 3147 3148 static int 3149 nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3150 { 3151 int rc; 3152 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3153 struct ibv_async_event event; 3154 3155 rc = ibv_get_async_event(device->context, &event); 3156 3157 if (rc) { 3158 /* In non-blocking mode -1 means there are no events available */ 3159 return rc; 3160 } 3161 3162 switch (event.event_type) { 3163 case IBV_EVENT_QP_FATAL: 3164 rqpair = event.element.qp->qp_context; 3165 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3166 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3167 (uintptr_t)rqpair->cm_id, event.event_type); 3168 nvmf_rdma_update_ibv_state(rqpair); 3169 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3170 break; 3171 case IBV_EVENT_QP_LAST_WQE_REACHED: 3172 /* This event only occurs for shared receive queues. */ 3173 rqpair = event.element.qp->qp_context; 3174 SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair); 3175 rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached); 3176 if (rc) { 3177 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc); 3178 rqpair->last_wqe_reached = true; 3179 } 3180 break; 3181 case IBV_EVENT_SQ_DRAINED: 3182 /* This event occurs frequently in both error and non-error states. 3183 * Check if the qpair is in an error state before sending a message. */ 3184 rqpair = event.element.qp->qp_context; 3185 SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair); 3186 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3187 (uintptr_t)rqpair->cm_id, event.event_type); 3188 if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3189 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3190 } 3191 break; 3192 case IBV_EVENT_QP_REQ_ERR: 3193 case IBV_EVENT_QP_ACCESS_ERR: 3194 case IBV_EVENT_COMM_EST: 3195 case IBV_EVENT_PATH_MIG: 3196 case IBV_EVENT_PATH_MIG_ERR: 3197 SPDK_NOTICELOG("Async event: %s\n", 3198 ibv_event_type_str(event.event_type)); 3199 rqpair = event.element.qp->qp_context; 3200 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3201 (uintptr_t)rqpair->cm_id, event.event_type); 3202 nvmf_rdma_update_ibv_state(rqpair); 3203 break; 3204 case IBV_EVENT_CQ_ERR: 3205 case IBV_EVENT_DEVICE_FATAL: 3206 case IBV_EVENT_PORT_ACTIVE: 3207 case IBV_EVENT_PORT_ERR: 3208 case IBV_EVENT_LID_CHANGE: 3209 case IBV_EVENT_PKEY_CHANGE: 3210 case IBV_EVENT_SM_CHANGE: 3211 case IBV_EVENT_SRQ_ERR: 3212 case IBV_EVENT_SRQ_LIMIT_REACHED: 3213 case IBV_EVENT_CLIENT_REREGISTER: 3214 case IBV_EVENT_GID_CHANGE: 3215 default: 3216 SPDK_NOTICELOG("Async event: %s\n", 3217 ibv_event_type_str(event.event_type)); 3218 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3219 break; 3220 } 3221 ibv_ack_async_event(&event); 3222 3223 return 0; 3224 } 3225 3226 static void 3227 nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events) 3228 { 3229 int rc = 0; 3230 uint32_t i = 0; 3231 3232 for (i = 0; i < max_events; i++) { 3233 rc = nvmf_process_ib_event(device); 3234 if (rc) { 3235 break; 3236 } 3237 } 3238 3239 SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i); 3240 } 3241 3242 static uint32_t 3243 nvmf_rdma_accept(struct spdk_nvmf_transport *transport) 3244 { 3245 int nfds, i = 0; 3246 struct spdk_nvmf_rdma_transport *rtransport; 3247 struct spdk_nvmf_rdma_device *device, *tmp; 3248 uint32_t count; 3249 3250 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3251 count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3252 3253 if (nfds <= 0) { 3254 return 0; 3255 } 3256 3257 /* The first poll descriptor is RDMA CM event */ 3258 if (rtransport->poll_fds[i++].revents & POLLIN) { 3259 nvmf_process_cm_event(transport); 3260 nfds--; 3261 } 3262 3263 if (nfds == 0) { 3264 return count; 3265 } 3266 3267 /* Second and subsequent poll descriptors are IB async events */ 3268 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3269 if (rtransport->poll_fds[i++].revents & POLLIN) { 3270 nvmf_process_ib_events(device, 32); 3271 nfds--; 3272 } 3273 } 3274 /* check all flagged fd's have been served */ 3275 assert(nfds == 0); 3276 3277 return count; 3278 } 3279 3280 static void 3281 nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem, 3282 struct spdk_nvmf_ctrlr_data *cdata) 3283 { 3284 cdata->nvmf_specific.msdbd = SPDK_NVMF_MAX_SGL_ENTRIES; 3285 3286 /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled 3287 since in-capsule data only works with NVME drives that support SGL memory layout */ 3288 if (transport->opts.dif_insert_or_strip) { 3289 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16; 3290 } 3291 } 3292 3293 static void 3294 nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3295 struct spdk_nvme_transport_id *trid, 3296 struct spdk_nvmf_discovery_log_page_entry *entry) 3297 { 3298 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3299 entry->adrfam = trid->adrfam; 3300 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3301 3302 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3303 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3304 3305 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3306 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3307 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3308 } 3309 3310 static void 3311 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3312 3313 static struct spdk_nvmf_transport_poll_group * 3314 nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3315 { 3316 struct spdk_nvmf_rdma_transport *rtransport; 3317 struct spdk_nvmf_rdma_poll_group *rgroup; 3318 struct spdk_nvmf_rdma_poller *poller; 3319 struct spdk_nvmf_rdma_device *device; 3320 struct ibv_srq_init_attr srq_init_attr; 3321 struct spdk_nvmf_rdma_resource_opts opts; 3322 int num_cqe; 3323 3324 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3325 3326 rgroup = calloc(1, sizeof(*rgroup)); 3327 if (!rgroup) { 3328 return NULL; 3329 } 3330 3331 TAILQ_INIT(&rgroup->pollers); 3332 STAILQ_INIT(&rgroup->retired_bufs); 3333 3334 pthread_mutex_lock(&rtransport->lock); 3335 TAILQ_FOREACH(device, &rtransport->devices, link) { 3336 poller = calloc(1, sizeof(*poller)); 3337 if (!poller) { 3338 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3339 nvmf_rdma_poll_group_destroy(&rgroup->group); 3340 pthread_mutex_unlock(&rtransport->lock); 3341 return NULL; 3342 } 3343 3344 poller->device = device; 3345 poller->group = rgroup; 3346 3347 TAILQ_INIT(&poller->qpairs); 3348 STAILQ_INIT(&poller->qpairs_pending_send); 3349 STAILQ_INIT(&poller->qpairs_pending_recv); 3350 3351 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3352 if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3353 poller->max_srq_depth = rtransport->rdma_opts.max_srq_depth; 3354 3355 device->num_srq++; 3356 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 3357 srq_init_attr.attr.max_wr = poller->max_srq_depth; 3358 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3359 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 3360 if (!poller->srq) { 3361 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3362 nvmf_rdma_poll_group_destroy(&rgroup->group); 3363 pthread_mutex_unlock(&rtransport->lock); 3364 return NULL; 3365 } 3366 3367 opts.qp = poller->srq; 3368 opts.pd = device->pd; 3369 opts.qpair = NULL; 3370 opts.shared = true; 3371 opts.max_queue_depth = poller->max_srq_depth; 3372 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3373 3374 poller->resources = nvmf_rdma_resources_create(&opts); 3375 if (!poller->resources) { 3376 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3377 nvmf_rdma_poll_group_destroy(&rgroup->group); 3378 pthread_mutex_unlock(&rtransport->lock); 3379 return NULL; 3380 } 3381 } 3382 3383 /* 3384 * When using an srq, we can limit the completion queue at startup. 3385 * The following formula represents the calculation: 3386 * num_cqe = num_recv + num_data_wr + num_send_wr. 3387 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3388 */ 3389 if (poller->srq) { 3390 num_cqe = poller->max_srq_depth * 3; 3391 } else { 3392 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3393 } 3394 3395 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3396 if (!poller->cq) { 3397 SPDK_ERRLOG("Unable to create completion queue\n"); 3398 nvmf_rdma_poll_group_destroy(&rgroup->group); 3399 pthread_mutex_unlock(&rtransport->lock); 3400 return NULL; 3401 } 3402 poller->num_cqe = num_cqe; 3403 } 3404 3405 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3406 if (rtransport->conn_sched.next_admin_pg == NULL) { 3407 rtransport->conn_sched.next_admin_pg = rgroup; 3408 rtransport->conn_sched.next_io_pg = rgroup; 3409 } 3410 3411 pthread_mutex_unlock(&rtransport->lock); 3412 return &rgroup->group; 3413 } 3414 3415 static struct spdk_nvmf_transport_poll_group * 3416 nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3417 { 3418 struct spdk_nvmf_rdma_transport *rtransport; 3419 struct spdk_nvmf_rdma_poll_group **pg; 3420 struct spdk_nvmf_transport_poll_group *result; 3421 3422 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3423 3424 pthread_mutex_lock(&rtransport->lock); 3425 3426 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3427 pthread_mutex_unlock(&rtransport->lock); 3428 return NULL; 3429 } 3430 3431 if (qpair->qid == 0) { 3432 pg = &rtransport->conn_sched.next_admin_pg; 3433 } else { 3434 pg = &rtransport->conn_sched.next_io_pg; 3435 } 3436 3437 assert(*pg != NULL); 3438 3439 result = &(*pg)->group; 3440 3441 *pg = TAILQ_NEXT(*pg, link); 3442 if (*pg == NULL) { 3443 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3444 } 3445 3446 pthread_mutex_unlock(&rtransport->lock); 3447 3448 return result; 3449 } 3450 3451 static void 3452 nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3453 { 3454 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3455 struct spdk_nvmf_rdma_poller *poller, *tmp; 3456 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3457 struct spdk_nvmf_transport_pg_cache_buf *buf, *tmp_buf; 3458 struct spdk_nvmf_rdma_transport *rtransport; 3459 3460 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3461 if (!rgroup) { 3462 return; 3463 } 3464 3465 /* free all retired buffers back to the transport so we don't short the mempool. */ 3466 STAILQ_FOREACH_SAFE(buf, &rgroup->retired_bufs, link, tmp_buf) { 3467 STAILQ_REMOVE(&rgroup->retired_bufs, buf, spdk_nvmf_transport_pg_cache_buf, link); 3468 assert(group->transport != NULL); 3469 spdk_mempool_put(group->transport->data_buf_pool, buf); 3470 } 3471 3472 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3473 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3474 3475 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3476 nvmf_rdma_qpair_destroy(qpair); 3477 } 3478 3479 if (poller->srq) { 3480 if (poller->resources) { 3481 nvmf_rdma_resources_destroy(poller->resources); 3482 } 3483 ibv_destroy_srq(poller->srq); 3484 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq); 3485 } 3486 3487 if (poller->cq) { 3488 ibv_destroy_cq(poller->cq); 3489 } 3490 3491 free(poller); 3492 } 3493 3494 if (rgroup->group.transport == NULL) { 3495 /* Transport can be NULL when nvmf_rdma_poll_group_create() 3496 * calls this function directly in a failure path. */ 3497 free(rgroup); 3498 return; 3499 } 3500 3501 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3502 3503 pthread_mutex_lock(&rtransport->lock); 3504 next_rgroup = TAILQ_NEXT(rgroup, link); 3505 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3506 if (next_rgroup == NULL) { 3507 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3508 } 3509 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3510 rtransport->conn_sched.next_admin_pg = next_rgroup; 3511 } 3512 if (rtransport->conn_sched.next_io_pg == rgroup) { 3513 rtransport->conn_sched.next_io_pg = next_rgroup; 3514 } 3515 pthread_mutex_unlock(&rtransport->lock); 3516 3517 free(rgroup); 3518 } 3519 3520 static void 3521 nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3522 { 3523 if (rqpair->cm_id != NULL) { 3524 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3525 } 3526 } 3527 3528 static int 3529 nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3530 struct spdk_nvmf_qpair *qpair) 3531 { 3532 struct spdk_nvmf_rdma_poll_group *rgroup; 3533 struct spdk_nvmf_rdma_qpair *rqpair; 3534 struct spdk_nvmf_rdma_device *device; 3535 struct spdk_nvmf_rdma_poller *poller; 3536 int rc; 3537 3538 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3539 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3540 3541 device = rqpair->device; 3542 3543 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3544 if (poller->device == device) { 3545 break; 3546 } 3547 } 3548 3549 if (!poller) { 3550 SPDK_ERRLOG("No poller found for device.\n"); 3551 return -1; 3552 } 3553 3554 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3555 rqpair->poller = poller; 3556 rqpair->srq = rqpair->poller->srq; 3557 3558 rc = nvmf_rdma_qpair_initialize(qpair); 3559 if (rc < 0) { 3560 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3561 return -1; 3562 } 3563 3564 rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3565 if (rc) { 3566 /* Try to reject, but we probably can't */ 3567 nvmf_rdma_qpair_reject_connection(rqpair); 3568 return -1; 3569 } 3570 3571 nvmf_rdma_update_ibv_state(rqpair); 3572 3573 return 0; 3574 } 3575 3576 static int 3577 nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3578 struct spdk_nvmf_qpair *qpair) 3579 { 3580 struct spdk_nvmf_rdma_qpair *rqpair; 3581 3582 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3583 assert(group->transport->tgt != NULL); 3584 3585 rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt); 3586 3587 if (!rqpair->destruct_channel) { 3588 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair); 3589 return 0; 3590 } 3591 3592 /* Sanity check that we get io_channel on the correct thread */ 3593 if (qpair->group) { 3594 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel)); 3595 } 3596 3597 return 0; 3598 } 3599 3600 static int 3601 nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3602 { 3603 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3604 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3605 struct spdk_nvmf_rdma_transport, transport); 3606 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3607 struct spdk_nvmf_rdma_qpair, qpair); 3608 3609 /* 3610 * AER requests are freed when a qpair is destroyed. The recv corresponding to that request 3611 * needs to be returned to the shared receive queue or the poll group will eventually be 3612 * starved of RECV structures. 3613 */ 3614 if (rqpair->srq && rdma_req->recv) { 3615 int rc; 3616 struct ibv_recv_wr *bad_recv_wr; 3617 3618 rc = ibv_post_srq_recv(rqpair->srq, &rdma_req->recv->wr, &bad_recv_wr); 3619 if (rc) { 3620 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 3621 } 3622 } 3623 3624 _nvmf_rdma_request_free(rdma_req, rtransport); 3625 return 0; 3626 } 3627 3628 static int 3629 nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3630 { 3631 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3632 struct spdk_nvmf_rdma_transport, transport); 3633 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3634 struct spdk_nvmf_rdma_request, req); 3635 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3636 struct spdk_nvmf_rdma_qpair, qpair); 3637 3638 if (rqpair->ibv_state != IBV_QPS_ERR) { 3639 /* The connection is alive, so process the request as normal */ 3640 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3641 } else { 3642 /* The connection is dead. Move the request directly to the completed state. */ 3643 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3644 } 3645 3646 nvmf_rdma_request_process(rtransport, rdma_req); 3647 3648 return 0; 3649 } 3650 3651 static void 3652 nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair, 3653 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3654 { 3655 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3656 3657 rqpair->to_close = true; 3658 3659 /* This happens only when the qpair is disconnected before 3660 * it is added to the poll group. Since there is no poll group, 3661 * the RDMA qp has not been initialized yet and the RDMA CM 3662 * event has not yet been acknowledged, so we need to reject it. 3663 */ 3664 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3665 nvmf_rdma_qpair_reject_connection(rqpair); 3666 nvmf_rdma_qpair_destroy(rqpair); 3667 return; 3668 } 3669 3670 if (rqpair->rdma_qp) { 3671 spdk_rdma_qp_disconnect(rqpair->rdma_qp); 3672 } 3673 3674 nvmf_rdma_destroy_drained_qpair(rqpair); 3675 3676 if (cb_fn) { 3677 cb_fn(cb_arg); 3678 } 3679 } 3680 3681 static struct spdk_nvmf_rdma_qpair * 3682 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3683 { 3684 struct spdk_nvmf_rdma_qpair *rqpair; 3685 /* @todo: improve QP search */ 3686 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3687 if (wc->qp_num == rqpair->rdma_qp->qp->qp_num) { 3688 return rqpair; 3689 } 3690 } 3691 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3692 return NULL; 3693 } 3694 3695 #ifdef DEBUG 3696 static int 3697 nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3698 { 3699 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3700 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3701 } 3702 #endif 3703 3704 static void 3705 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3706 int rc) 3707 { 3708 struct spdk_nvmf_rdma_recv *rdma_recv; 3709 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3710 3711 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3712 while (bad_recv_wr != NULL) { 3713 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3714 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3715 3716 rdma_recv->qpair->current_recv_depth++; 3717 bad_recv_wr = bad_recv_wr->next; 3718 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3719 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL); 3720 } 3721 } 3722 3723 static void 3724 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3725 { 3726 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3727 while (bad_recv_wr != NULL) { 3728 bad_recv_wr = bad_recv_wr->next; 3729 rqpair->current_recv_depth++; 3730 } 3731 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3732 } 3733 3734 static void 3735 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3736 struct spdk_nvmf_rdma_poller *rpoller) 3737 { 3738 struct spdk_nvmf_rdma_qpair *rqpair; 3739 struct ibv_recv_wr *bad_recv_wr; 3740 int rc; 3741 3742 if (rpoller->srq) { 3743 if (rpoller->resources->recvs_to_post.first != NULL) { 3744 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3745 if (rc) { 3746 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3747 } 3748 rpoller->resources->recvs_to_post.first = NULL; 3749 rpoller->resources->recvs_to_post.last = NULL; 3750 } 3751 } else { 3752 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3753 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3754 assert(rqpair->resources->recvs_to_post.first != NULL); 3755 rc = ibv_post_recv(rqpair->rdma_qp->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3756 if (rc) { 3757 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3758 } 3759 rqpair->resources->recvs_to_post.first = NULL; 3760 rqpair->resources->recvs_to_post.last = NULL; 3761 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3762 } 3763 } 3764 } 3765 3766 static void 3767 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3768 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3769 { 3770 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3771 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3772 3773 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3774 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3775 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3776 assert(rqpair->current_send_depth > 0); 3777 rqpair->current_send_depth--; 3778 switch (bad_rdma_wr->type) { 3779 case RDMA_WR_TYPE_DATA: 3780 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3781 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3782 assert(rqpair->current_read_depth > 0); 3783 rqpair->current_read_depth--; 3784 } 3785 break; 3786 case RDMA_WR_TYPE_SEND: 3787 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3788 break; 3789 default: 3790 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3791 prev_rdma_req = cur_rdma_req; 3792 continue; 3793 } 3794 3795 if (prev_rdma_req == cur_rdma_req) { 3796 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3797 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3798 continue; 3799 } 3800 3801 switch (cur_rdma_req->state) { 3802 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3803 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3804 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3805 break; 3806 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3807 case RDMA_REQUEST_STATE_COMPLETING: 3808 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3809 break; 3810 default: 3811 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3812 cur_rdma_req->state, rqpair); 3813 continue; 3814 } 3815 3816 nvmf_rdma_request_process(rtransport, cur_rdma_req); 3817 prev_rdma_req = cur_rdma_req; 3818 } 3819 3820 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3821 /* Disconnect the connection. */ 3822 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3823 } 3824 3825 } 3826 3827 static void 3828 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3829 struct spdk_nvmf_rdma_poller *rpoller) 3830 { 3831 struct spdk_nvmf_rdma_qpair *rqpair; 3832 struct ibv_send_wr *bad_wr = NULL; 3833 int rc; 3834 3835 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3836 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3837 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr); 3838 3839 /* bad wr always points to the first wr that failed. */ 3840 if (rc) { 3841 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3842 } 3843 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3844 } 3845 } 3846 3847 static const char * 3848 nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type) 3849 { 3850 switch (wr_type) { 3851 case RDMA_WR_TYPE_RECV: 3852 return "RECV"; 3853 case RDMA_WR_TYPE_SEND: 3854 return "SEND"; 3855 case RDMA_WR_TYPE_DATA: 3856 return "DATA"; 3857 default: 3858 SPDK_ERRLOG("Unknown WR type %d\n", wr_type); 3859 SPDK_UNREACHABLE(); 3860 } 3861 } 3862 3863 static inline void 3864 nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc) 3865 { 3866 enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type; 3867 3868 if (wc->status == IBV_WC_WR_FLUSH_ERR) { 3869 /* If qpair is in ERR state, we will receive completions for all posted and not completed 3870 * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */ 3871 SPDK_DEBUGLOG(rdma, 3872 "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3873 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3874 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3875 } else { 3876 SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n", 3877 rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id, 3878 nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status)); 3879 } 3880 } 3881 3882 static int 3883 nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3884 struct spdk_nvmf_rdma_poller *rpoller) 3885 { 3886 struct ibv_wc wc[32]; 3887 struct spdk_nvmf_rdma_wr *rdma_wr; 3888 struct spdk_nvmf_rdma_request *rdma_req; 3889 struct spdk_nvmf_rdma_recv *rdma_recv; 3890 struct spdk_nvmf_rdma_qpair *rqpair; 3891 int reaped, i; 3892 int count = 0; 3893 bool error = false; 3894 uint64_t poll_tsc = spdk_get_ticks(); 3895 3896 /* Poll for completing operations. */ 3897 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3898 if (reaped < 0) { 3899 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3900 errno, spdk_strerror(errno)); 3901 return -1; 3902 } 3903 3904 rpoller->stat.polls++; 3905 rpoller->stat.completions += reaped; 3906 3907 for (i = 0; i < reaped; i++) { 3908 3909 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3910 3911 switch (rdma_wr->type) { 3912 case RDMA_WR_TYPE_SEND: 3913 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3914 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3915 3916 if (!wc[i].status) { 3917 count++; 3918 assert(wc[i].opcode == IBV_WC_SEND); 3919 assert(nvmf_rdma_req_is_completing(rdma_req)); 3920 } 3921 3922 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3923 /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */ 3924 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3925 rdma_req->num_outstanding_data_wr = 0; 3926 3927 nvmf_rdma_request_process(rtransport, rdma_req); 3928 break; 3929 case RDMA_WR_TYPE_RECV: 3930 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3931 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3932 if (rpoller->srq != NULL) { 3933 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3934 /* It is possible that there are still some completions for destroyed QP 3935 * associated with SRQ. We just ignore these late completions and re-post 3936 * receive WRs back to SRQ. 3937 */ 3938 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3939 struct ibv_recv_wr *bad_wr; 3940 int rc; 3941 3942 rdma_recv->wr.next = NULL; 3943 rc = ibv_post_srq_recv(rpoller->srq, 3944 &rdma_recv->wr, 3945 &bad_wr); 3946 if (rc) { 3947 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3948 } 3949 continue; 3950 } 3951 } 3952 rqpair = rdma_recv->qpair; 3953 3954 assert(rqpair != NULL); 3955 if (!wc[i].status) { 3956 assert(wc[i].opcode == IBV_WC_RECV); 3957 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3958 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 3959 break; 3960 } 3961 } 3962 3963 rdma_recv->wr.next = NULL; 3964 rqpair->current_recv_depth++; 3965 rdma_recv->receive_tsc = poll_tsc; 3966 rpoller->stat.requests++; 3967 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3968 break; 3969 case RDMA_WR_TYPE_DATA: 3970 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3971 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3972 3973 assert(rdma_req->num_outstanding_data_wr > 0); 3974 3975 rqpair->current_send_depth--; 3976 rdma_req->num_outstanding_data_wr--; 3977 if (!wc[i].status) { 3978 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3979 rqpair->current_read_depth--; 3980 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3981 if (rdma_req->num_outstanding_data_wr == 0) { 3982 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3983 nvmf_rdma_request_process(rtransport, rdma_req); 3984 } 3985 } else { 3986 /* If the data transfer fails still force the queue into the error state, 3987 * if we were performing an RDMA_READ, we need to force the request into a 3988 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3989 * case, we should wait for the SEND to complete. */ 3990 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3991 rqpair->current_read_depth--; 3992 if (rdma_req->num_outstanding_data_wr == 0) { 3993 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3994 } 3995 } 3996 } 3997 break; 3998 default: 3999 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 4000 continue; 4001 } 4002 4003 /* Handle error conditions */ 4004 if (wc[i].status) { 4005 nvmf_rdma_update_ibv_state(rqpair); 4006 nvmf_rdma_log_wc_status(rqpair, &wc[i]); 4007 4008 error = true; 4009 4010 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 4011 /* Disconnect the connection. */ 4012 spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL); 4013 } else { 4014 nvmf_rdma_destroy_drained_qpair(rqpair); 4015 } 4016 continue; 4017 } 4018 4019 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 4020 4021 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 4022 nvmf_rdma_destroy_drained_qpair(rqpair); 4023 } 4024 } 4025 4026 if (error == true) { 4027 return -1; 4028 } 4029 4030 /* submit outstanding work requests. */ 4031 _poller_submit_recvs(rtransport, rpoller); 4032 _poller_submit_sends(rtransport, rpoller); 4033 4034 return count; 4035 } 4036 4037 static int 4038 nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 4039 { 4040 struct spdk_nvmf_rdma_transport *rtransport; 4041 struct spdk_nvmf_rdma_poll_group *rgroup; 4042 struct spdk_nvmf_rdma_poller *rpoller; 4043 int count, rc; 4044 4045 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 4046 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 4047 4048 count = 0; 4049 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4050 rc = nvmf_rdma_poller_poll(rtransport, rpoller); 4051 if (rc < 0) { 4052 return rc; 4053 } 4054 count += rc; 4055 } 4056 4057 return count; 4058 } 4059 4060 static int 4061 nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 4062 struct spdk_nvme_transport_id *trid, 4063 bool peer) 4064 { 4065 struct sockaddr *saddr; 4066 uint16_t port; 4067 4068 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA); 4069 4070 if (peer) { 4071 saddr = rdma_get_peer_addr(id); 4072 } else { 4073 saddr = rdma_get_local_addr(id); 4074 } 4075 switch (saddr->sa_family) { 4076 case AF_INET: { 4077 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 4078 4079 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 4080 inet_ntop(AF_INET, &saddr_in->sin_addr, 4081 trid->traddr, sizeof(trid->traddr)); 4082 if (peer) { 4083 port = ntohs(rdma_get_dst_port(id)); 4084 } else { 4085 port = ntohs(rdma_get_src_port(id)); 4086 } 4087 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4088 break; 4089 } 4090 case AF_INET6: { 4091 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 4092 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 4093 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 4094 trid->traddr, sizeof(trid->traddr)); 4095 if (peer) { 4096 port = ntohs(rdma_get_dst_port(id)); 4097 } else { 4098 port = ntohs(rdma_get_src_port(id)); 4099 } 4100 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 4101 break; 4102 } 4103 default: 4104 return -1; 4105 4106 } 4107 4108 return 0; 4109 } 4110 4111 static int 4112 nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 4113 struct spdk_nvme_transport_id *trid) 4114 { 4115 struct spdk_nvmf_rdma_qpair *rqpair; 4116 4117 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4118 4119 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 4120 } 4121 4122 static int 4123 nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 4124 struct spdk_nvme_transport_id *trid) 4125 { 4126 struct spdk_nvmf_rdma_qpair *rqpair; 4127 4128 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4129 4130 return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 4131 } 4132 4133 static int 4134 nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 4135 struct spdk_nvme_transport_id *trid) 4136 { 4137 struct spdk_nvmf_rdma_qpair *rqpair; 4138 4139 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4140 4141 return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4142 } 4143 4144 void 4145 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4146 { 4147 g_nvmf_hooks = *hooks; 4148 } 4149 4150 static void 4151 nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req, 4152 struct spdk_nvmf_rdma_request *rdma_req_to_abort) 4153 { 4154 rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 4155 rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 4156 4157 rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 4158 4159 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 4160 } 4161 4162 static int 4163 _nvmf_rdma_qpair_abort_request(void *ctx) 4164 { 4165 struct spdk_nvmf_request *req = ctx; 4166 struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF( 4167 req->req_to_abort, struct spdk_nvmf_rdma_request, req); 4168 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 4169 struct spdk_nvmf_rdma_qpair, qpair); 4170 int rc; 4171 4172 spdk_poller_unregister(&req->poller); 4173 4174 switch (rdma_req_to_abort->state) { 4175 case RDMA_REQUEST_STATE_EXECUTING: 4176 rc = nvmf_ctrlr_abort_request(req); 4177 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 4178 return SPDK_POLLER_BUSY; 4179 } 4180 break; 4181 4182 case RDMA_REQUEST_STATE_NEED_BUFFER: 4183 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue, 4184 &rdma_req_to_abort->req, spdk_nvmf_request, buf_link); 4185 4186 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4187 break; 4188 4189 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 4190 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort, 4191 spdk_nvmf_rdma_request, state_link); 4192 4193 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4194 break; 4195 4196 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 4197 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort, 4198 spdk_nvmf_rdma_request, state_link); 4199 4200 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort); 4201 break; 4202 4203 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 4204 if (spdk_get_ticks() < req->timeout_tsc) { 4205 req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0); 4206 return SPDK_POLLER_BUSY; 4207 } 4208 break; 4209 4210 default: 4211 break; 4212 } 4213 4214 spdk_nvmf_request_complete(req); 4215 return SPDK_POLLER_BUSY; 4216 } 4217 4218 static void 4219 nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 4220 struct spdk_nvmf_request *req) 4221 { 4222 struct spdk_nvmf_rdma_qpair *rqpair; 4223 struct spdk_nvmf_rdma_transport *rtransport; 4224 struct spdk_nvmf_transport *transport; 4225 uint16_t cid; 4226 uint32_t i; 4227 struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL; 4228 4229 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4230 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 4231 transport = &rtransport->transport; 4232 4233 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 4234 4235 for (i = 0; i < rqpair->max_queue_depth; i++) { 4236 rdma_req_to_abort = &rqpair->resources->reqs[i]; 4237 4238 if (rdma_req_to_abort->state != RDMA_REQUEST_STATE_FREE && 4239 rdma_req_to_abort->req.cmd->nvme_cmd.cid == cid) { 4240 break; 4241 } 4242 } 4243 4244 if (rdma_req_to_abort == NULL) { 4245 spdk_nvmf_request_complete(req); 4246 return; 4247 } 4248 4249 req->req_to_abort = &rdma_req_to_abort->req; 4250 req->timeout_tsc = spdk_get_ticks() + 4251 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 4252 req->poller = NULL; 4253 4254 _nvmf_rdma_qpair_abort_request(req); 4255 } 4256 4257 static int 4258 nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4259 struct spdk_nvmf_transport_poll_group_stat **stat) 4260 { 4261 struct spdk_io_channel *ch; 4262 struct spdk_nvmf_poll_group *group; 4263 struct spdk_nvmf_transport_poll_group *tgroup; 4264 struct spdk_nvmf_rdma_poll_group *rgroup; 4265 struct spdk_nvmf_rdma_poller *rpoller; 4266 struct spdk_nvmf_rdma_device_stat *device_stat; 4267 uint64_t num_devices = 0; 4268 4269 if (tgt == NULL || stat == NULL) { 4270 return -EINVAL; 4271 } 4272 4273 ch = spdk_get_io_channel(tgt); 4274 group = spdk_io_channel_get_ctx(ch);; 4275 spdk_put_io_channel(ch); 4276 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4277 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4278 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4279 if (!*stat) { 4280 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4281 return -ENOMEM; 4282 } 4283 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4284 4285 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4286 /* Count devices to allocate enough memory */ 4287 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4288 ++num_devices; 4289 } 4290 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4291 if (!(*stat)->rdma.devices) { 4292 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4293 free(*stat); 4294 return -ENOMEM; 4295 } 4296 4297 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4298 (*stat)->rdma.num_devices = num_devices; 4299 num_devices = 0; 4300 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4301 device_stat = &(*stat)->rdma.devices[num_devices++]; 4302 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4303 device_stat->polls = rpoller->stat.polls; 4304 device_stat->completions = rpoller->stat.completions; 4305 device_stat->requests = rpoller->stat.requests; 4306 device_stat->request_latency = rpoller->stat.request_latency; 4307 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4308 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4309 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4310 } 4311 return 0; 4312 } 4313 } 4314 return -ENOENT; 4315 } 4316 4317 static void 4318 nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4319 { 4320 if (stat) { 4321 free(stat->rdma.devices); 4322 } 4323 free(stat); 4324 } 4325 4326 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4327 .name = "RDMA", 4328 .type = SPDK_NVME_TRANSPORT_RDMA, 4329 .opts_init = nvmf_rdma_opts_init, 4330 .create = nvmf_rdma_create, 4331 .dump_opts = nvmf_rdma_dump_opts, 4332 .destroy = nvmf_rdma_destroy, 4333 4334 .listen = nvmf_rdma_listen, 4335 .stop_listen = nvmf_rdma_stop_listen, 4336 .accept = nvmf_rdma_accept, 4337 .cdata_init = nvmf_rdma_cdata_init, 4338 4339 .listener_discover = nvmf_rdma_discover, 4340 4341 .poll_group_create = nvmf_rdma_poll_group_create, 4342 .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group, 4343 .poll_group_destroy = nvmf_rdma_poll_group_destroy, 4344 .poll_group_add = nvmf_rdma_poll_group_add, 4345 .poll_group_remove = nvmf_rdma_poll_group_remove, 4346 .poll_group_poll = nvmf_rdma_poll_group_poll, 4347 4348 .req_free = nvmf_rdma_request_free, 4349 .req_complete = nvmf_rdma_request_complete, 4350 4351 .qpair_fini = nvmf_rdma_close_qpair, 4352 .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid, 4353 .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid, 4354 .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid, 4355 .qpair_abort_request = nvmf_rdma_qpair_abort_request, 4356 4357 .poll_group_get_stat = nvmf_rdma_poll_group_get_stat, 4358 .poll_group_free_stat = nvmf_rdma_poll_group_free_stat, 4359 }; 4360 4361 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma); 4362 SPDK_LOG_REGISTER_COMPONENT(rdma) 4363