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