1 /*- 2 * BSD LICENSE 3 * 4 * Copyright (c) Intel Corporation. All rights reserved. 5 * Copyright (c) 2019 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 <infiniband/verbs.h> 37 #include <rdma/rdma_cma.h> 38 #include <rdma/rdma_verbs.h> 39 40 #include "nvmf_internal.h" 41 #include "transport.h" 42 43 #include "spdk/config.h" 44 #include "spdk/thread.h" 45 #include "spdk/nvmf.h" 46 #include "spdk/nvmf_spec.h" 47 #include "spdk/string.h" 48 #include "spdk/trace.h" 49 #include "spdk/util.h" 50 51 #include "spdk_internal/assert.h" 52 #include "spdk_internal/log.h" 53 54 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {}; 55 56 /* 57 RDMA Connection Resource Defaults 58 */ 59 #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES 60 #define NVMF_DEFAULT_RSP_SGE 1 61 #define NVMF_DEFAULT_RX_SGE 2 62 63 /* The RDMA completion queue size */ 64 #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096 65 #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2) 66 67 /* Timeout for destroying defunct rqpairs */ 68 #define NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US 4000000 69 70 static int g_spdk_nvmf_ibv_query_mask = 71 IBV_QP_STATE | 72 IBV_QP_PKEY_INDEX | 73 IBV_QP_PORT | 74 IBV_QP_ACCESS_FLAGS | 75 IBV_QP_AV | 76 IBV_QP_PATH_MTU | 77 IBV_QP_DEST_QPN | 78 IBV_QP_RQ_PSN | 79 IBV_QP_MAX_DEST_RD_ATOMIC | 80 IBV_QP_MIN_RNR_TIMER | 81 IBV_QP_SQ_PSN | 82 IBV_QP_TIMEOUT | 83 IBV_QP_RETRY_CNT | 84 IBV_QP_RNR_RETRY | 85 IBV_QP_MAX_QP_RD_ATOMIC; 86 87 enum spdk_nvmf_rdma_request_state { 88 /* The request is not currently in use */ 89 RDMA_REQUEST_STATE_FREE = 0, 90 91 /* Initial state when request first received */ 92 RDMA_REQUEST_STATE_NEW, 93 94 /* The request is queued until a data buffer is available. */ 95 RDMA_REQUEST_STATE_NEED_BUFFER, 96 97 /* The request is waiting on RDMA queue depth availability 98 * to transfer data from the host to the controller. 99 */ 100 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 101 102 /* The request is currently transferring data from the host to the controller. */ 103 RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 104 105 /* The request is ready to execute at the block device */ 106 RDMA_REQUEST_STATE_READY_TO_EXECUTE, 107 108 /* The request is currently executing at the block device */ 109 RDMA_REQUEST_STATE_EXECUTING, 110 111 /* The request finished executing at the block device */ 112 RDMA_REQUEST_STATE_EXECUTED, 113 114 /* The request is waiting on RDMA queue depth availability 115 * to transfer data from the controller to the host. 116 */ 117 RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 118 119 /* The request is ready to send a completion */ 120 RDMA_REQUEST_STATE_READY_TO_COMPLETE, 121 122 /* The request is currently transferring data from the controller to the host. */ 123 RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 124 125 /* The request currently has an outstanding completion without an 126 * associated data transfer. 127 */ 128 RDMA_REQUEST_STATE_COMPLETING, 129 130 /* The request completed and can be marked free. */ 131 RDMA_REQUEST_STATE_COMPLETED, 132 133 /* Terminator */ 134 RDMA_REQUEST_NUM_STATES, 135 }; 136 137 #define OBJECT_NVMF_RDMA_IO 0x40 138 139 #define TRACE_GROUP_NVMF_RDMA 0x4 140 #define TRACE_RDMA_REQUEST_STATE_NEW SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0) 141 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1) 142 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2) 143 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3) 144 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4) 145 #define TRACE_RDMA_REQUEST_STATE_EXECUTING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5) 146 #define TRACE_RDMA_REQUEST_STATE_EXECUTED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6) 147 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7) 148 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8) 149 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9) 150 #define TRACE_RDMA_REQUEST_STATE_COMPLETING SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA) 151 #define TRACE_RDMA_REQUEST_STATE_COMPLETED SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB) 152 #define TRACE_RDMA_QP_CREATE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC) 153 #define TRACE_RDMA_IBV_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD) 154 #define TRACE_RDMA_CM_ASYNC_EVENT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE) 155 #define TRACE_RDMA_QP_STATE_CHANGE SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF) 156 #define TRACE_RDMA_QP_DISCONNECT SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10) 157 #define TRACE_RDMA_QP_DESTROY SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11) 158 159 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA) 160 { 161 spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r'); 162 spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW, 163 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid: "); 164 spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 165 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 166 spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H", 167 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 168 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 169 spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C", 170 TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 171 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 172 spdk_trace_register_description("RDMA_REQ_TX_H2C", 173 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 174 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 175 spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE", 176 TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 177 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 178 spdk_trace_register_description("RDMA_REQ_EXECUTING", 179 TRACE_RDMA_REQUEST_STATE_EXECUTING, 180 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 181 spdk_trace_register_description("RDMA_REQ_EXECUTED", 182 TRACE_RDMA_REQUEST_STATE_EXECUTED, 183 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 184 spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL", 185 TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 186 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 187 spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H", 188 TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 189 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 190 spdk_trace_register_description("RDMA_REQ_COMPLETING", 191 TRACE_RDMA_REQUEST_STATE_COMPLETING, 192 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 193 spdk_trace_register_description("RDMA_REQ_COMPLETED", 194 TRACE_RDMA_REQUEST_STATE_COMPLETED, 195 OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid: "); 196 197 spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE, 198 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 199 spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT, 200 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 201 spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT, 202 OWNER_NONE, OBJECT_NONE, 0, 0, "type: "); 203 spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE, 204 OWNER_NONE, OBJECT_NONE, 0, 1, "state: "); 205 spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT, 206 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 207 spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY, 208 OWNER_NONE, OBJECT_NONE, 0, 0, ""); 209 } 210 211 enum spdk_nvmf_rdma_wr_type { 212 RDMA_WR_TYPE_RECV, 213 RDMA_WR_TYPE_SEND, 214 RDMA_WR_TYPE_DATA, 215 }; 216 217 struct spdk_nvmf_rdma_wr { 218 enum spdk_nvmf_rdma_wr_type type; 219 }; 220 221 /* This structure holds commands as they are received off the wire. 222 * It must be dynamically paired with a full request object 223 * (spdk_nvmf_rdma_request) to service a request. It is separate 224 * from the request because RDMA does not appear to order 225 * completions, so occasionally we'll get a new incoming 226 * command when there aren't any free request objects. 227 */ 228 struct spdk_nvmf_rdma_recv { 229 struct ibv_recv_wr wr; 230 struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE]; 231 232 struct spdk_nvmf_rdma_qpair *qpair; 233 234 /* In-capsule data buffer */ 235 uint8_t *buf; 236 237 struct spdk_nvmf_rdma_wr rdma_wr; 238 uint64_t receive_tsc; 239 240 STAILQ_ENTRY(spdk_nvmf_rdma_recv) link; 241 }; 242 243 struct spdk_nvmf_rdma_request_data { 244 struct spdk_nvmf_rdma_wr rdma_wr; 245 struct ibv_send_wr wr; 246 struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES]; 247 }; 248 249 struct spdk_nvmf_rdma_request { 250 struct spdk_nvmf_request req; 251 252 enum spdk_nvmf_rdma_request_state state; 253 254 struct spdk_nvmf_rdma_recv *recv; 255 256 struct { 257 struct spdk_nvmf_rdma_wr rdma_wr; 258 struct ibv_send_wr wr; 259 struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE]; 260 } rsp; 261 262 struct spdk_nvmf_rdma_request_data data; 263 264 uint32_t iovpos; 265 266 uint32_t num_outstanding_data_wr; 267 uint64_t receive_tsc; 268 269 STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link; 270 }; 271 272 enum spdk_nvmf_rdma_qpair_disconnect_flags { 273 RDMA_QP_DISCONNECTING = 1, 274 RDMA_QP_RECV_DRAINED = 1 << 1, 275 RDMA_QP_SEND_DRAINED = 1 << 2 276 }; 277 278 struct spdk_nvmf_rdma_resource_opts { 279 struct spdk_nvmf_rdma_qpair *qpair; 280 /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */ 281 void *qp; 282 struct ibv_pd *pd; 283 uint32_t max_queue_depth; 284 uint32_t in_capsule_data_size; 285 bool shared; 286 }; 287 288 struct spdk_nvmf_send_wr_list { 289 struct ibv_send_wr *first; 290 struct ibv_send_wr *last; 291 }; 292 293 struct spdk_nvmf_recv_wr_list { 294 struct ibv_recv_wr *first; 295 struct ibv_recv_wr *last; 296 }; 297 298 struct spdk_nvmf_rdma_resources { 299 /* Array of size "max_queue_depth" containing RDMA requests. */ 300 struct spdk_nvmf_rdma_request *reqs; 301 302 /* Array of size "max_queue_depth" containing RDMA recvs. */ 303 struct spdk_nvmf_rdma_recv *recvs; 304 305 /* Array of size "max_queue_depth" containing 64 byte capsules 306 * used for receive. 307 */ 308 union nvmf_h2c_msg *cmds; 309 struct ibv_mr *cmds_mr; 310 311 /* Array of size "max_queue_depth" containing 16 byte completions 312 * to be sent back to the user. 313 */ 314 union nvmf_c2h_msg *cpls; 315 struct ibv_mr *cpls_mr; 316 317 /* Array of size "max_queue_depth * InCapsuleDataSize" containing 318 * buffers to be used for in capsule data. 319 */ 320 void *bufs; 321 struct ibv_mr *bufs_mr; 322 323 /* The list of pending recvs to transfer */ 324 struct spdk_nvmf_recv_wr_list recvs_to_post; 325 326 /* Receives that are waiting for a request object */ 327 STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue; 328 329 /* Queue to track free requests */ 330 STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue; 331 }; 332 333 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair); 334 335 struct spdk_nvmf_rdma_ibv_event_ctx { 336 struct spdk_nvmf_rdma_qpair *rqpair; 337 spdk_nvmf_rdma_qpair_ibv_event cb_fn; 338 /* Link to other ibv events associated with this qpair */ 339 STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link; 340 }; 341 342 struct spdk_nvmf_rdma_qpair { 343 struct spdk_nvmf_qpair qpair; 344 345 struct spdk_nvmf_rdma_port *port; 346 struct spdk_nvmf_rdma_poller *poller; 347 348 struct rdma_cm_id *cm_id; 349 struct ibv_srq *srq; 350 struct rdma_cm_id *listen_id; 351 352 /* The maximum number of I/O outstanding on this connection at one time */ 353 uint16_t max_queue_depth; 354 355 /* The maximum number of active RDMA READ and ATOMIC operations at one time */ 356 uint16_t max_read_depth; 357 358 /* The maximum number of RDMA SEND operations at one time */ 359 uint32_t max_send_depth; 360 361 /* The current number of outstanding WRs from this qpair's 362 * recv queue. Should not exceed device->attr.max_queue_depth. 363 */ 364 uint16_t current_recv_depth; 365 366 /* The current number of active RDMA READ operations */ 367 uint16_t current_read_depth; 368 369 /* The current number of posted WRs from this qpair's 370 * send queue. Should not exceed max_send_depth. 371 */ 372 uint32_t current_send_depth; 373 374 /* The maximum number of SGEs per WR on the send queue */ 375 uint32_t max_send_sge; 376 377 /* The maximum number of SGEs per WR on the recv queue */ 378 uint32_t max_recv_sge; 379 380 /* The list of pending send requests for a transfer */ 381 struct spdk_nvmf_send_wr_list sends_to_post; 382 383 struct spdk_nvmf_rdma_resources *resources; 384 385 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue; 386 387 STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue; 388 389 /* Number of requests not in the free state */ 390 uint32_t qd; 391 392 TAILQ_ENTRY(spdk_nvmf_rdma_qpair) link; 393 394 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link; 395 396 STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link; 397 398 /* IBV queue pair attributes: they are used to manage 399 * qp state and recover from errors. 400 */ 401 enum ibv_qp_state ibv_state; 402 403 uint32_t disconnect_flags; 404 405 /* Poller registered in case the qpair doesn't properly 406 * complete the qpair destruct process and becomes defunct. 407 */ 408 409 struct spdk_poller *destruct_poller; 410 411 /* List of ibv async events */ 412 STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events; 413 414 /* There are several ways a disconnect can start on a qpair 415 * and they are not all mutually exclusive. It is important 416 * that we only initialize one of these paths. 417 */ 418 bool disconnect_started; 419 /* Lets us know that we have received the last_wqe event. */ 420 bool last_wqe_reached; 421 }; 422 423 struct spdk_nvmf_rdma_poller_stat { 424 uint64_t completions; 425 uint64_t polls; 426 uint64_t requests; 427 uint64_t request_latency; 428 uint64_t pending_free_request; 429 uint64_t pending_rdma_read; 430 uint64_t pending_rdma_write; 431 }; 432 433 struct spdk_nvmf_rdma_poller { 434 struct spdk_nvmf_rdma_device *device; 435 struct spdk_nvmf_rdma_poll_group *group; 436 437 int num_cqe; 438 int required_num_wr; 439 struct ibv_cq *cq; 440 441 /* The maximum number of I/O outstanding on the shared receive queue at one time */ 442 uint16_t max_srq_depth; 443 444 /* Shared receive queue */ 445 struct ibv_srq *srq; 446 447 struct spdk_nvmf_rdma_resources *resources; 448 struct spdk_nvmf_rdma_poller_stat stat; 449 450 TAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs; 451 452 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv; 453 454 STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send; 455 456 TAILQ_ENTRY(spdk_nvmf_rdma_poller) link; 457 }; 458 459 struct spdk_nvmf_rdma_poll_group_stat { 460 uint64_t pending_data_buffer; 461 }; 462 463 struct spdk_nvmf_rdma_poll_group { 464 struct spdk_nvmf_transport_poll_group group; 465 struct spdk_nvmf_rdma_poll_group_stat stat; 466 TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers; 467 TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link; 468 /* 469 * buffers which are split across multiple RDMA 470 * memory regions cannot be used by this transport. 471 */ 472 STAILQ_HEAD(, spdk_nvmf_transport_pg_cache_buf) retired_bufs; 473 }; 474 475 struct spdk_nvmf_rdma_conn_sched { 476 struct spdk_nvmf_rdma_poll_group *next_admin_pg; 477 struct spdk_nvmf_rdma_poll_group *next_io_pg; 478 }; 479 480 /* Assuming rdma_cm uses just one protection domain per ibv_context. */ 481 struct spdk_nvmf_rdma_device { 482 struct ibv_device_attr attr; 483 struct ibv_context *context; 484 485 struct spdk_mem_map *map; 486 struct ibv_pd *pd; 487 488 int num_srq; 489 490 TAILQ_ENTRY(spdk_nvmf_rdma_device) link; 491 }; 492 493 struct spdk_nvmf_rdma_port { 494 struct spdk_nvme_transport_id trid; 495 struct rdma_cm_id *id; 496 struct spdk_nvmf_rdma_device *device; 497 uint32_t ref; 498 TAILQ_ENTRY(spdk_nvmf_rdma_port) link; 499 }; 500 501 struct spdk_nvmf_rdma_transport { 502 struct spdk_nvmf_transport transport; 503 504 struct spdk_nvmf_rdma_conn_sched conn_sched; 505 506 struct rdma_event_channel *event_channel; 507 508 struct spdk_mempool *data_wr_pool; 509 510 pthread_mutex_t lock; 511 512 /* fields used to poll RDMA/IB events */ 513 nfds_t npoll_fds; 514 struct pollfd *poll_fds; 515 516 TAILQ_HEAD(, spdk_nvmf_rdma_device) devices; 517 TAILQ_HEAD(, spdk_nvmf_rdma_port) ports; 518 TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups; 519 }; 520 521 static inline void 522 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair); 523 524 static inline int 525 spdk_nvmf_rdma_check_ibv_state(enum ibv_qp_state state) 526 { 527 switch (state) { 528 case IBV_QPS_RESET: 529 case IBV_QPS_INIT: 530 case IBV_QPS_RTR: 531 case IBV_QPS_RTS: 532 case IBV_QPS_SQD: 533 case IBV_QPS_SQE: 534 case IBV_QPS_ERR: 535 return 0; 536 default: 537 return -1; 538 } 539 } 540 541 static inline enum spdk_nvme_media_error_status_code 542 spdk_nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) { 543 enum spdk_nvme_media_error_status_code result; 544 switch (err_type) 545 { 546 case SPDK_DIF_REFTAG_ERROR: 547 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 548 break; 549 case SPDK_DIF_APPTAG_ERROR: 550 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 551 break; 552 case SPDK_DIF_GUARD_ERROR: 553 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 554 break; 555 default: 556 SPDK_UNREACHABLE(); 557 } 558 559 return result; 560 } 561 562 static enum ibv_qp_state 563 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) { 564 enum ibv_qp_state old_state, new_state; 565 struct ibv_qp_attr qp_attr; 566 struct ibv_qp_init_attr init_attr; 567 int rc; 568 569 old_state = rqpair->ibv_state; 570 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 571 g_spdk_nvmf_ibv_query_mask, &init_attr); 572 573 if (rc) 574 { 575 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 576 return IBV_QPS_ERR + 1; 577 } 578 579 new_state = qp_attr.qp_state; 580 rqpair->ibv_state = new_state; 581 qp_attr.ah_attr.port_num = qp_attr.port_num; 582 583 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 584 if (rc) 585 { 586 SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state); 587 /* 588 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8 589 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR 590 */ 591 return IBV_QPS_ERR + 1; 592 } 593 594 if (old_state != new_state) 595 { 596 spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, 597 (uintptr_t)rqpair->cm_id, new_state); 598 } 599 return new_state; 600 } 601 602 static const char *str_ibv_qp_state[] = { 603 "IBV_QPS_RESET", 604 "IBV_QPS_INIT", 605 "IBV_QPS_RTR", 606 "IBV_QPS_RTS", 607 "IBV_QPS_SQD", 608 "IBV_QPS_SQE", 609 "IBV_QPS_ERR", 610 "IBV_QPS_UNKNOWN" 611 }; 612 613 static int 614 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair, 615 enum ibv_qp_state new_state) 616 { 617 struct ibv_qp_attr qp_attr; 618 struct ibv_qp_init_attr init_attr; 619 int rc; 620 enum ibv_qp_state state; 621 static int attr_mask_rc[] = { 622 [IBV_QPS_RESET] = IBV_QP_STATE, 623 [IBV_QPS_INIT] = (IBV_QP_STATE | 624 IBV_QP_PKEY_INDEX | 625 IBV_QP_PORT | 626 IBV_QP_ACCESS_FLAGS), 627 [IBV_QPS_RTR] = (IBV_QP_STATE | 628 IBV_QP_AV | 629 IBV_QP_PATH_MTU | 630 IBV_QP_DEST_QPN | 631 IBV_QP_RQ_PSN | 632 IBV_QP_MAX_DEST_RD_ATOMIC | 633 IBV_QP_MIN_RNR_TIMER), 634 [IBV_QPS_RTS] = (IBV_QP_STATE | 635 IBV_QP_SQ_PSN | 636 IBV_QP_TIMEOUT | 637 IBV_QP_RETRY_CNT | 638 IBV_QP_RNR_RETRY | 639 IBV_QP_MAX_QP_RD_ATOMIC), 640 [IBV_QPS_SQD] = IBV_QP_STATE, 641 [IBV_QPS_SQE] = IBV_QP_STATE, 642 [IBV_QPS_ERR] = IBV_QP_STATE, 643 }; 644 645 rc = spdk_nvmf_rdma_check_ibv_state(new_state); 646 if (rc) { 647 SPDK_ERRLOG("QP#%d: bad state requested: %u\n", 648 rqpair->qpair.qid, new_state); 649 return rc; 650 } 651 652 rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr, 653 g_spdk_nvmf_ibv_query_mask, &init_attr); 654 655 if (rc) { 656 SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n"); 657 assert(false); 658 } 659 660 qp_attr.cur_qp_state = rqpair->ibv_state; 661 qp_attr.qp_state = new_state; 662 663 rc = ibv_modify_qp(rqpair->cm_id->qp, &qp_attr, 664 attr_mask_rc[new_state]); 665 666 if (rc) { 667 SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n", 668 rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno)); 669 return rc; 670 } 671 672 state = spdk_nvmf_rdma_update_ibv_state(rqpair); 673 674 if (state != new_state) { 675 SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n", 676 rqpair->qpair.qid, str_ibv_qp_state[new_state], 677 str_ibv_qp_state[state]); 678 return -1; 679 } 680 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid, 681 str_ibv_qp_state[state]); 682 return 0; 683 } 684 685 static void 686 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req, 687 struct spdk_nvmf_rdma_transport *rtransport) 688 { 689 struct spdk_nvmf_rdma_request_data *data_wr; 690 struct ibv_send_wr *next_send_wr; 691 uint64_t req_wrid; 692 693 rdma_req->num_outstanding_data_wr = 0; 694 data_wr = &rdma_req->data; 695 req_wrid = data_wr->wr.wr_id; 696 while (data_wr && data_wr->wr.wr_id == req_wrid) { 697 memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge); 698 data_wr->wr.num_sge = 0; 699 next_send_wr = data_wr->wr.next; 700 if (data_wr != &rdma_req->data) { 701 spdk_mempool_put(rtransport->data_wr_pool, data_wr); 702 } 703 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : 704 SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr); 705 } 706 } 707 708 static void 709 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req) 710 { 711 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool); 712 if (req->req.cmd) { 713 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode); 714 } 715 if (req->recv) { 716 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id); 717 } 718 } 719 720 static void 721 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair) 722 { 723 int i; 724 725 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid); 726 for (i = 0; i < rqpair->max_queue_depth; i++) { 727 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) { 728 nvmf_rdma_dump_request(&rqpair->resources->reqs[i]); 729 } 730 } 731 } 732 733 static void 734 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources) 735 { 736 if (resources->cmds_mr) { 737 ibv_dereg_mr(resources->cmds_mr); 738 } 739 740 if (resources->cpls_mr) { 741 ibv_dereg_mr(resources->cpls_mr); 742 } 743 744 if (resources->bufs_mr) { 745 ibv_dereg_mr(resources->bufs_mr); 746 } 747 748 spdk_free(resources->cmds); 749 spdk_free(resources->cpls); 750 spdk_free(resources->bufs); 751 free(resources->reqs); 752 free(resources->recvs); 753 free(resources); 754 } 755 756 757 static struct spdk_nvmf_rdma_resources * 758 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts) 759 { 760 struct spdk_nvmf_rdma_resources *resources; 761 struct spdk_nvmf_rdma_request *rdma_req; 762 struct spdk_nvmf_rdma_recv *rdma_recv; 763 struct ibv_qp *qp; 764 struct ibv_srq *srq; 765 uint32_t i; 766 int rc; 767 768 resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources)); 769 if (!resources) { 770 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 771 return NULL; 772 } 773 774 resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs)); 775 resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs)); 776 resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds), 777 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 778 resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls), 779 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA); 780 781 if (opts->in_capsule_data_size > 0) { 782 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size, 783 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, 784 SPDK_MALLOC_DMA); 785 } 786 787 if (!resources->reqs || !resources->recvs || !resources->cmds || 788 !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) { 789 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n"); 790 goto cleanup; 791 } 792 793 resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds, 794 opts->max_queue_depth * sizeof(*resources->cmds), 795 IBV_ACCESS_LOCAL_WRITE); 796 resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls, 797 opts->max_queue_depth * sizeof(*resources->cpls), 798 0); 799 800 if (opts->in_capsule_data_size) { 801 resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs, 802 opts->max_queue_depth * 803 opts->in_capsule_data_size, 804 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); 805 } 806 807 if (!resources->cmds_mr || !resources->cpls_mr || 808 (opts->in_capsule_data_size && 809 !resources->bufs_mr)) { 810 goto cleanup; 811 } 812 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n", 813 resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds), 814 resources->cmds_mr->lkey); 815 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n", 816 resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls), 817 resources->cpls_mr->lkey); 818 if (resources->bufs && resources->bufs_mr) { 819 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n", 820 resources->bufs, opts->max_queue_depth * 821 opts->in_capsule_data_size, resources->bufs_mr->lkey); 822 } 823 824 /* Initialize queues */ 825 STAILQ_INIT(&resources->incoming_queue); 826 STAILQ_INIT(&resources->free_queue); 827 828 for (i = 0; i < opts->max_queue_depth; i++) { 829 struct ibv_recv_wr *bad_wr = NULL; 830 831 rdma_recv = &resources->recvs[i]; 832 rdma_recv->qpair = opts->qpair; 833 834 /* Set up memory to receive commands */ 835 if (resources->bufs) { 836 rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i * 837 opts->in_capsule_data_size)); 838 } 839 840 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV; 841 842 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i]; 843 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]); 844 rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey; 845 rdma_recv->wr.num_sge = 1; 846 847 if (rdma_recv->buf && resources->bufs_mr) { 848 rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf; 849 rdma_recv->sgl[1].length = opts->in_capsule_data_size; 850 rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey; 851 rdma_recv->wr.num_sge++; 852 } 853 854 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr; 855 rdma_recv->wr.sg_list = rdma_recv->sgl; 856 if (opts->shared) { 857 srq = (struct ibv_srq *)opts->qp; 858 rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr); 859 } else { 860 qp = (struct ibv_qp *)opts->qp; 861 rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr); 862 } 863 if (rc) { 864 goto cleanup; 865 } 866 } 867 868 for (i = 0; i < opts->max_queue_depth; i++) { 869 rdma_req = &resources->reqs[i]; 870 871 if (opts->qpair != NULL) { 872 rdma_req->req.qpair = &opts->qpair->qpair; 873 } else { 874 rdma_req->req.qpair = NULL; 875 } 876 rdma_req->req.cmd = NULL; 877 878 /* Set up memory to send responses */ 879 rdma_req->req.rsp = &resources->cpls[i]; 880 881 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i]; 882 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]); 883 rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey; 884 885 rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND; 886 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr; 887 rdma_req->rsp.wr.next = NULL; 888 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 889 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED; 890 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl; 891 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl); 892 893 /* Set up memory for data buffers */ 894 rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA; 895 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 896 rdma_req->data.wr.next = NULL; 897 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED; 898 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 899 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl); 900 901 /* Initialize request state to FREE */ 902 rdma_req->state = RDMA_REQUEST_STATE_FREE; 903 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link); 904 } 905 906 return resources; 907 908 cleanup: 909 nvmf_rdma_resources_destroy(resources); 910 return NULL; 911 } 912 913 static void 914 spdk_nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair) 915 { 916 struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx; 917 STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) { 918 ctx->rqpair = NULL; 919 /* Memory allocated for ctx is freed in spdk_nvmf_rdma_qpair_process_ibv_event */ 920 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 921 } 922 } 923 924 static void 925 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 926 { 927 struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp; 928 struct ibv_recv_wr *bad_recv_wr = NULL; 929 int rc; 930 931 spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0); 932 933 spdk_poller_unregister(&rqpair->destruct_poller); 934 935 if (rqpair->qd != 0) { 936 if (rqpair->srq == NULL) { 937 nvmf_rdma_dump_qpair_contents(rqpair); 938 } 939 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd); 940 } 941 942 if (rqpair->poller) { 943 TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link); 944 945 if (rqpair->srq != NULL && rqpair->resources != NULL) { 946 /* Drop all received but unprocessed commands for this queue and return them to SRQ */ 947 STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) { 948 if (rqpair == rdma_recv->qpair) { 949 STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link); 950 rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr); 951 if (rc) { 952 SPDK_ERRLOG("Unable to re-post rx descriptor\n"); 953 } 954 } 955 } 956 } 957 } 958 959 if (rqpair->cm_id) { 960 if (rqpair->cm_id->qp != NULL) { 961 rdma_destroy_qp(rqpair->cm_id); 962 } 963 rdma_destroy_id(rqpair->cm_id); 964 965 if (rqpair->poller != NULL && rqpair->srq == NULL) { 966 rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth); 967 } 968 } 969 970 if (rqpair->srq == NULL && rqpair->resources != NULL) { 971 nvmf_rdma_resources_destroy(rqpair->resources); 972 } 973 974 spdk_nvmf_rdma_qpair_clean_ibv_events(rqpair); 975 976 free(rqpair); 977 } 978 979 static int 980 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device) 981 { 982 struct spdk_nvmf_rdma_poller *rpoller; 983 int rc, num_cqe, required_num_wr; 984 985 /* Enlarge CQ size dynamically */ 986 rpoller = rqpair->poller; 987 required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth); 988 num_cqe = rpoller->num_cqe; 989 if (num_cqe < required_num_wr) { 990 num_cqe = spdk_max(num_cqe * 2, required_num_wr); 991 num_cqe = spdk_min(num_cqe, device->attr.max_cqe); 992 } 993 994 if (rpoller->num_cqe != num_cqe) { 995 if (required_num_wr > device->attr.max_cqe) { 996 SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n", 997 required_num_wr, device->attr.max_cqe); 998 return -1; 999 } 1000 1001 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe); 1002 rc = ibv_resize_cq(rpoller->cq, num_cqe); 1003 if (rc) { 1004 SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno)); 1005 return -1; 1006 } 1007 1008 rpoller->num_cqe = num_cqe; 1009 } 1010 1011 rpoller->required_num_wr = required_num_wr; 1012 return 0; 1013 } 1014 1015 static int 1016 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 1017 { 1018 struct spdk_nvmf_rdma_qpair *rqpair; 1019 int rc; 1020 struct spdk_nvmf_rdma_transport *rtransport; 1021 struct spdk_nvmf_transport *transport; 1022 struct spdk_nvmf_rdma_resource_opts opts; 1023 struct spdk_nvmf_rdma_device *device; 1024 struct ibv_qp_init_attr ibv_init_attr; 1025 1026 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1027 device = rqpair->port->device; 1028 1029 memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr)); 1030 ibv_init_attr.qp_context = rqpair; 1031 ibv_init_attr.qp_type = IBV_QPT_RC; 1032 ibv_init_attr.send_cq = rqpair->poller->cq; 1033 ibv_init_attr.recv_cq = rqpair->poller->cq; 1034 1035 if (rqpair->srq) { 1036 ibv_init_attr.srq = rqpair->srq; 1037 } else { 1038 ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth + 1039 1; /* RECV operations + dummy drain WR */ 1040 } 1041 1042 ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth * 1043 2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */ 1044 ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE); 1045 ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 1046 1047 if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) { 1048 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n"); 1049 goto error; 1050 } 1051 1052 rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr); 1053 if (rc) { 1054 SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno)); 1055 goto error; 1056 } 1057 1058 rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2 + 1), 1059 ibv_init_attr.cap.max_send_wr); 1060 rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge); 1061 rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge); 1062 spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0); 1063 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair); 1064 1065 rqpair->sends_to_post.first = NULL; 1066 rqpair->sends_to_post.last = NULL; 1067 1068 if (rqpair->poller->srq == NULL) { 1069 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 1070 transport = &rtransport->transport; 1071 1072 opts.qp = rqpair->cm_id->qp; 1073 opts.pd = rqpair->cm_id->pd; 1074 opts.qpair = rqpair; 1075 opts.shared = false; 1076 opts.max_queue_depth = rqpair->max_queue_depth; 1077 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 1078 1079 rqpair->resources = nvmf_rdma_resources_create(&opts); 1080 1081 if (!rqpair->resources) { 1082 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n"); 1083 rdma_destroy_qp(rqpair->cm_id); 1084 goto error; 1085 } 1086 } else { 1087 rqpair->resources = rqpair->poller->resources; 1088 } 1089 1090 rqpair->current_recv_depth = 0; 1091 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 1092 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 1093 1094 return 0; 1095 1096 error: 1097 rdma_destroy_id(rqpair->cm_id); 1098 rqpair->cm_id = NULL; 1099 return -1; 1100 } 1101 1102 /* Append the given recv wr structure to the resource structs outstanding recvs list. */ 1103 /* This function accepts either a single wr or the first wr in a linked list. */ 1104 static void 1105 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first) 1106 { 1107 struct ibv_recv_wr *last; 1108 1109 last = first; 1110 while (last->next != NULL) { 1111 last = last->next; 1112 } 1113 1114 if (rqpair->resources->recvs_to_post.first == NULL) { 1115 rqpair->resources->recvs_to_post.first = first; 1116 rqpair->resources->recvs_to_post.last = last; 1117 if (rqpair->srq == NULL) { 1118 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link); 1119 } 1120 } else { 1121 rqpair->resources->recvs_to_post.last->next = first; 1122 rqpair->resources->recvs_to_post.last = last; 1123 } 1124 } 1125 1126 /* Append the given send wr structure to the qpair's outstanding sends list. */ 1127 /* This function accepts either a single wr or the first wr in a linked list. */ 1128 static void 1129 nvmf_rdma_qpair_queue_send_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *first) 1130 { 1131 struct ibv_send_wr *last; 1132 1133 last = first; 1134 while (last->next != NULL) { 1135 last = last->next; 1136 } 1137 1138 if (rqpair->sends_to_post.first == NULL) { 1139 rqpair->sends_to_post.first = first; 1140 rqpair->sends_to_post.last = last; 1141 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link); 1142 } else { 1143 rqpair->sends_to_post.last->next = first; 1144 rqpair->sends_to_post.last = last; 1145 } 1146 } 1147 1148 static int 1149 request_transfer_in(struct spdk_nvmf_request *req) 1150 { 1151 struct spdk_nvmf_rdma_request *rdma_req; 1152 struct spdk_nvmf_qpair *qpair; 1153 struct spdk_nvmf_rdma_qpair *rqpair; 1154 1155 qpair = req->qpair; 1156 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1157 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1158 1159 assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 1160 assert(rdma_req != NULL); 1161 1162 nvmf_rdma_qpair_queue_send_wrs(rqpair, &rdma_req->data.wr); 1163 rqpair->current_read_depth += rdma_req->num_outstanding_data_wr; 1164 rqpair->current_send_depth += rdma_req->num_outstanding_data_wr; 1165 return 0; 1166 } 1167 1168 static int 1169 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted) 1170 { 1171 int num_outstanding_data_wr = 0; 1172 struct spdk_nvmf_rdma_request *rdma_req; 1173 struct spdk_nvmf_qpair *qpair; 1174 struct spdk_nvmf_rdma_qpair *rqpair; 1175 struct spdk_nvme_cpl *rsp; 1176 struct ibv_send_wr *first = NULL; 1177 1178 *data_posted = 0; 1179 qpair = req->qpair; 1180 rsp = &req->rsp->nvme_cpl; 1181 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 1182 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 1183 1184 /* Advance our sq_head pointer */ 1185 if (qpair->sq_head == qpair->sq_head_max) { 1186 qpair->sq_head = 0; 1187 } else { 1188 qpair->sq_head++; 1189 } 1190 rsp->sqhd = qpair->sq_head; 1191 1192 /* queue the capsule for the recv buffer */ 1193 assert(rdma_req->recv != NULL); 1194 1195 nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr); 1196 1197 rdma_req->recv = NULL; 1198 assert(rqpair->current_recv_depth > 0); 1199 rqpair->current_recv_depth--; 1200 1201 /* Build the response which consists of optional 1202 * RDMA WRITEs to transfer data, plus an RDMA SEND 1203 * containing the response. 1204 */ 1205 first = &rdma_req->rsp.wr; 1206 1207 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && 1208 req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1209 first = &rdma_req->data.wr; 1210 *data_posted = 1; 1211 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr; 1212 } 1213 nvmf_rdma_qpair_queue_send_wrs(rqpair, first); 1214 /* +1 for the rsp wr */ 1215 rqpair->current_send_depth += num_outstanding_data_wr + 1; 1216 1217 return 0; 1218 } 1219 1220 static int 1221 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair) 1222 { 1223 struct spdk_nvmf_rdma_accept_private_data accept_data; 1224 struct rdma_conn_param ctrlr_event_data = {}; 1225 int rc; 1226 1227 accept_data.recfmt = 0; 1228 accept_data.crqsize = rqpair->max_queue_depth; 1229 1230 ctrlr_event_data.private_data = &accept_data; 1231 ctrlr_event_data.private_data_len = sizeof(accept_data); 1232 if (id->ps == RDMA_PS_TCP) { 1233 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */ 1234 ctrlr_event_data.initiator_depth = rqpair->max_read_depth; 1235 } 1236 1237 /* Configure infinite retries for the initiator side qpair. 1238 * When using a shared receive queue on the target side, 1239 * we need to pass this value to the initiator to prevent the 1240 * initiator side NIC from completing SEND requests back to the 1241 * initiator with status rnr_retry_count_exceeded. */ 1242 if (rqpair->srq != NULL) { 1243 ctrlr_event_data.rnr_retry_count = 0x7; 1244 } 1245 1246 rc = rdma_accept(id, &ctrlr_event_data); 1247 if (rc) { 1248 SPDK_ERRLOG("Error %d on rdma_accept\n", errno); 1249 } else { 1250 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n"); 1251 } 1252 1253 return rc; 1254 } 1255 1256 static void 1257 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error) 1258 { 1259 struct spdk_nvmf_rdma_reject_private_data rej_data; 1260 1261 rej_data.recfmt = 0; 1262 rej_data.sts = error; 1263 1264 rdma_reject(id, &rej_data, sizeof(rej_data)); 1265 } 1266 1267 static int 1268 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event, 1269 new_qpair_fn cb_fn, void *cb_arg) 1270 { 1271 struct spdk_nvmf_rdma_transport *rtransport; 1272 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 1273 struct spdk_nvmf_rdma_port *port; 1274 struct rdma_conn_param *rdma_param = NULL; 1275 const struct spdk_nvmf_rdma_request_private_data *private_data = NULL; 1276 uint16_t max_queue_depth; 1277 uint16_t max_read_depth; 1278 1279 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 1280 1281 assert(event->id != NULL); /* Impossible. Can't even reject the connection. */ 1282 assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */ 1283 1284 rdma_param = &event->param.conn; 1285 if (rdma_param->private_data == NULL || 1286 rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1287 SPDK_ERRLOG("connect request: no private data provided\n"); 1288 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH); 1289 return -1; 1290 } 1291 1292 private_data = rdma_param->private_data; 1293 if (private_data->recfmt != 0) { 1294 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n"); 1295 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT); 1296 return -1; 1297 } 1298 1299 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n", 1300 event->id->verbs->device->name, event->id->verbs->device->dev_name); 1301 1302 port = event->listen_id->context; 1303 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n", 1304 event->listen_id, event->listen_id->verbs, port); 1305 1306 /* Figure out the supported queue depth. This is a multi-step process 1307 * that takes into account hardware maximums, host provided values, 1308 * and our target's internal memory limits */ 1309 1310 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n"); 1311 1312 /* Start with the maximum queue depth allowed by the target */ 1313 max_queue_depth = rtransport->transport.opts.max_queue_depth; 1314 max_read_depth = rtransport->transport.opts.max_queue_depth; 1315 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n", 1316 rtransport->transport.opts.max_queue_depth); 1317 1318 /* Next check the local NIC's hardware limitations */ 1319 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1320 "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n", 1321 port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom); 1322 max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr); 1323 max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom); 1324 1325 /* Next check the remote NIC's hardware limitations */ 1326 SPDK_DEBUGLOG(SPDK_LOG_RDMA, 1327 "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n", 1328 rdma_param->initiator_depth, rdma_param->responder_resources); 1329 if (rdma_param->initiator_depth > 0) { 1330 max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth); 1331 } 1332 1333 /* Finally check for the host software requested values, which are 1334 * optional. */ 1335 if (rdma_param->private_data != NULL && 1336 rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) { 1337 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize); 1338 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize); 1339 max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize); 1340 max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1); 1341 } 1342 1343 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n", 1344 max_queue_depth, max_read_depth); 1345 1346 rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair)); 1347 if (rqpair == NULL) { 1348 SPDK_ERRLOG("Could not allocate new connection.\n"); 1349 spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 1350 return -1; 1351 } 1352 1353 rqpair->port = port; 1354 rqpair->max_queue_depth = max_queue_depth; 1355 rqpair->max_read_depth = max_read_depth; 1356 rqpair->cm_id = event->id; 1357 rqpair->listen_id = event->listen_id; 1358 rqpair->qpair.transport = transport; 1359 STAILQ_INIT(&rqpair->ibv_events); 1360 /* use qid from the private data to determine the qpair type 1361 qid will be set to the appropriate value when the controller is created */ 1362 rqpair->qpair.qid = private_data->qid; 1363 1364 event->id->context = &rqpair->qpair; 1365 1366 cb_fn(&rqpair->qpair, cb_arg); 1367 1368 return 0; 1369 } 1370 1371 static int 1372 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map, 1373 enum spdk_mem_map_notify_action action, 1374 void *vaddr, size_t size) 1375 { 1376 struct ibv_pd *pd = cb_ctx; 1377 struct ibv_mr *mr; 1378 int rc; 1379 1380 switch (action) { 1381 case SPDK_MEM_MAP_NOTIFY_REGISTER: 1382 if (!g_nvmf_hooks.get_rkey) { 1383 mr = ibv_reg_mr(pd, vaddr, size, 1384 IBV_ACCESS_LOCAL_WRITE | 1385 IBV_ACCESS_REMOTE_READ | 1386 IBV_ACCESS_REMOTE_WRITE); 1387 if (mr == NULL) { 1388 SPDK_ERRLOG("ibv_reg_mr() failed\n"); 1389 return -1; 1390 } else { 1391 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr); 1392 } 1393 } else { 1394 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, 1395 g_nvmf_hooks.get_rkey(pd, vaddr, size)); 1396 } 1397 break; 1398 case SPDK_MEM_MAP_NOTIFY_UNREGISTER: 1399 if (!g_nvmf_hooks.get_rkey) { 1400 mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL); 1401 if (mr) { 1402 ibv_dereg_mr(mr); 1403 } 1404 } 1405 rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size); 1406 break; 1407 default: 1408 SPDK_UNREACHABLE(); 1409 } 1410 1411 return rc; 1412 } 1413 1414 static int 1415 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2) 1416 { 1417 /* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */ 1418 return addr_1 == addr_2; 1419 } 1420 1421 static inline void 1422 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next, 1423 enum spdk_nvme_data_transfer xfer) 1424 { 1425 if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 1426 wr->opcode = IBV_WR_RDMA_WRITE; 1427 wr->send_flags = 0; 1428 wr->next = next; 1429 } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 1430 wr->opcode = IBV_WR_RDMA_READ; 1431 wr->send_flags = IBV_SEND_SIGNALED; 1432 wr->next = NULL; 1433 } else { 1434 assert(0); 1435 } 1436 } 1437 1438 static int 1439 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport, 1440 struct spdk_nvmf_rdma_request *rdma_req, 1441 uint32_t num_sgl_descriptors) 1442 { 1443 struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES]; 1444 struct spdk_nvmf_rdma_request_data *current_data_wr; 1445 uint32_t i; 1446 1447 if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) { 1448 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n", 1449 num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES); 1450 return -EINVAL; 1451 } 1452 1453 if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) { 1454 return -ENOMEM; 1455 } 1456 1457 current_data_wr = &rdma_req->data; 1458 1459 for (i = 0; i < num_sgl_descriptors; i++) { 1460 nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer); 1461 current_data_wr->wr.next = &work_requests[i]->wr; 1462 current_data_wr = work_requests[i]; 1463 current_data_wr->wr.sg_list = current_data_wr->sgl; 1464 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id; 1465 } 1466 1467 nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1468 1469 return 0; 1470 } 1471 1472 static inline void 1473 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req) 1474 { 1475 struct ibv_send_wr *wr = &rdma_req->data.wr; 1476 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1477 1478 wr->wr.rdma.rkey = sgl->keyed.key; 1479 wr->wr.rdma.remote_addr = sgl->address; 1480 nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer); 1481 } 1482 1483 static inline void 1484 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs) 1485 { 1486 struct ibv_send_wr *wr = &rdma_req->data.wr; 1487 struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1; 1488 uint32_t i; 1489 int j; 1490 uint64_t remote_addr_offset = 0; 1491 1492 for (i = 0; i < num_wrs; ++i) { 1493 wr->wr.rdma.rkey = sgl->keyed.key; 1494 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset; 1495 for (j = 0; j < wr->num_sge; ++j) { 1496 remote_addr_offset += wr->sg_list[j].length; 1497 } 1498 wr = wr->next; 1499 } 1500 } 1501 1502 /* This function is used in the rare case that we have a buffer split over multiple memory regions. */ 1503 static int 1504 nvmf_rdma_replace_buffer(struct spdk_nvmf_rdma_poll_group *rgroup, void **buf) 1505 { 1506 struct spdk_nvmf_transport_poll_group *group = &rgroup->group; 1507 struct spdk_nvmf_transport *transport = group->transport; 1508 struct spdk_nvmf_transport_pg_cache_buf *old_buf; 1509 void *new_buf; 1510 1511 if (!(STAILQ_EMPTY(&group->buf_cache))) { 1512 group->buf_cache_count--; 1513 new_buf = STAILQ_FIRST(&group->buf_cache); 1514 STAILQ_REMOVE_HEAD(&group->buf_cache, link); 1515 assert(*buf != NULL); 1516 } else { 1517 new_buf = spdk_mempool_get(transport->data_buf_pool); 1518 } 1519 1520 if (*buf == NULL) { 1521 return -ENOMEM; 1522 } 1523 1524 old_buf = *buf; 1525 STAILQ_INSERT_HEAD(&rgroup->retired_bufs, old_buf, link); 1526 *buf = new_buf; 1527 return 0; 1528 } 1529 1530 static bool 1531 nvmf_rdma_get_lkey(struct spdk_nvmf_rdma_device *device, struct iovec *iov, 1532 uint32_t *_lkey) 1533 { 1534 uint64_t translation_len; 1535 uint32_t lkey; 1536 1537 translation_len = iov->iov_len; 1538 1539 if (!g_nvmf_hooks.get_rkey) { 1540 lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map, 1541 (uint64_t)iov->iov_base, &translation_len))->lkey; 1542 } else { 1543 lkey = spdk_mem_map_translate(device->map, 1544 (uint64_t)iov->iov_base, &translation_len); 1545 } 1546 1547 if (spdk_unlikely(translation_len < iov->iov_len)) { 1548 return false; 1549 } 1550 1551 *_lkey = lkey; 1552 return true; 1553 } 1554 1555 static bool 1556 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device, 1557 struct iovec *iov, struct ibv_send_wr **_wr, 1558 uint32_t *_remaining_data_block, uint32_t *_offset, 1559 uint32_t *_num_extra_wrs, 1560 const struct spdk_dif_ctx *dif_ctx) 1561 { 1562 struct ibv_send_wr *wr = *_wr; 1563 struct ibv_sge *sg_ele = &wr->sg_list[wr->num_sge]; 1564 uint32_t lkey = 0; 1565 uint32_t remaining, data_block_size, md_size, sge_len; 1566 1567 if (spdk_unlikely(!nvmf_rdma_get_lkey(device, iov, &lkey))) { 1568 /* This is a very rare case that can occur when using DPDK version < 19.05 */ 1569 SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions. Removing it from circulation.\n"); 1570 return false; 1571 } 1572 1573 if (spdk_likely(!dif_ctx)) { 1574 sg_ele->lkey = lkey; 1575 sg_ele->addr = (uintptr_t)(iov->iov_base); 1576 sg_ele->length = iov->iov_len; 1577 wr->num_sge++; 1578 } else { 1579 remaining = iov->iov_len - *_offset; 1580 data_block_size = dif_ctx->block_size - dif_ctx->md_size; 1581 md_size = dif_ctx->md_size; 1582 1583 while (remaining) { 1584 if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) { 1585 if (*_num_extra_wrs > 0 && wr->next) { 1586 *_wr = wr->next; 1587 wr = *_wr; 1588 wr->num_sge = 0; 1589 sg_ele = &wr->sg_list[wr->num_sge]; 1590 (*_num_extra_wrs)--; 1591 } else { 1592 break; 1593 } 1594 } 1595 sg_ele->lkey = lkey; 1596 sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset); 1597 sge_len = spdk_min(remaining, *_remaining_data_block); 1598 sg_ele->length = sge_len; 1599 remaining -= sge_len; 1600 *_remaining_data_block -= sge_len; 1601 *_offset += sge_len; 1602 1603 sg_ele++; 1604 wr->num_sge++; 1605 1606 if (*_remaining_data_block == 0) { 1607 /* skip metadata */ 1608 *_offset += md_size; 1609 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */ 1610 remaining -= spdk_min(remaining, md_size); 1611 *_remaining_data_block = data_block_size; 1612 } 1613 1614 if (remaining == 0) { 1615 /* By subtracting the size of the last IOV from the offset, we ensure that we skip 1616 the remaining metadata bits at the beginning of the next buffer */ 1617 *_offset -= iov->iov_len; 1618 } 1619 } 1620 } 1621 1622 return true; 1623 } 1624 1625 static int 1626 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup, 1627 struct spdk_nvmf_rdma_device *device, 1628 struct spdk_nvmf_rdma_request *rdma_req, 1629 struct ibv_send_wr *wr, 1630 uint32_t length, 1631 uint32_t num_extra_wrs) 1632 { 1633 struct spdk_nvmf_request *req = &rdma_req->req; 1634 struct spdk_dif_ctx *dif_ctx = NULL; 1635 uint32_t remaining_data_block = 0; 1636 uint32_t offset = 0; 1637 1638 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 1639 dif_ctx = &rdma_req->req.dif.dif_ctx; 1640 remaining_data_block = dif_ctx->block_size - dif_ctx->md_size; 1641 } 1642 1643 wr->num_sge = 0; 1644 1645 while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) { 1646 while (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr, 1647 &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) { 1648 if (nvmf_rdma_replace_buffer(rgroup, &req->buffers[rdma_req->iovpos]) == -ENOMEM) { 1649 return -ENOMEM; 1650 } 1651 req->iov[rdma_req->iovpos].iov_base = (void *)((uintptr_t)(req->buffers[rdma_req->iovpos] + 1652 NVMF_DATA_BUFFER_MASK) & 1653 ~NVMF_DATA_BUFFER_MASK); 1654 } 1655 1656 length -= req->iov[rdma_req->iovpos].iov_len; 1657 rdma_req->iovpos++; 1658 } 1659 1660 if (length) { 1661 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n"); 1662 return -EINVAL; 1663 } 1664 1665 return 0; 1666 } 1667 1668 static inline uint32_t 1669 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size) 1670 { 1671 /* estimate the number of SG entries and WRs needed to process the request */ 1672 uint32_t num_sge = 0; 1673 uint32_t i; 1674 uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size); 1675 1676 for (i = 0; i < num_buffers && length > 0; i++) { 1677 uint32_t buffer_len = spdk_min(length, io_unit_size); 1678 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size); 1679 1680 if (num_sge_in_block * block_size > buffer_len) { 1681 ++num_sge_in_block; 1682 } 1683 num_sge += num_sge_in_block; 1684 length -= buffer_len; 1685 } 1686 return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES); 1687 } 1688 1689 static int 1690 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport, 1691 struct spdk_nvmf_rdma_device *device, 1692 struct spdk_nvmf_rdma_request *rdma_req, 1693 uint32_t length) 1694 { 1695 struct spdk_nvmf_rdma_qpair *rqpair; 1696 struct spdk_nvmf_rdma_poll_group *rgroup; 1697 struct spdk_nvmf_request *req = &rdma_req->req; 1698 struct ibv_send_wr *wr = &rdma_req->data.wr; 1699 int rc; 1700 uint32_t num_wrs = 1; 1701 1702 rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair); 1703 rgroup = rqpair->poller->group; 1704 1705 /* rdma wr specifics */ 1706 nvmf_rdma_setup_request(rdma_req); 1707 1708 rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, 1709 length); 1710 if (rc != 0) { 1711 return rc; 1712 } 1713 1714 assert(req->iovcnt <= rqpair->max_send_sge); 1715 1716 rdma_req->iovpos = 0; 1717 1718 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1719 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size, 1720 req->dif.dif_ctx.block_size); 1721 if (num_wrs > 1) { 1722 rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1); 1723 if (rc != 0) { 1724 goto err_exit; 1725 } 1726 } 1727 } 1728 1729 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1); 1730 if (spdk_unlikely(rc != 0)) { 1731 goto err_exit; 1732 } 1733 1734 if (spdk_unlikely(num_wrs > 1)) { 1735 nvmf_rdma_update_remote_addr(rdma_req, num_wrs); 1736 } 1737 1738 /* set the number of outstanding data WRs for this request. */ 1739 rdma_req->num_outstanding_data_wr = num_wrs; 1740 1741 return rc; 1742 1743 err_exit: 1744 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1745 nvmf_rdma_request_free_data(rdma_req, rtransport); 1746 req->iovcnt = 0; 1747 return rc; 1748 } 1749 1750 static int 1751 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1752 struct spdk_nvmf_rdma_device *device, 1753 struct spdk_nvmf_rdma_request *rdma_req) 1754 { 1755 struct spdk_nvmf_rdma_qpair *rqpair; 1756 struct spdk_nvmf_rdma_poll_group *rgroup; 1757 struct ibv_send_wr *current_wr; 1758 struct spdk_nvmf_request *req = &rdma_req->req; 1759 struct spdk_nvme_sgl_descriptor *inline_segment, *desc; 1760 uint32_t num_sgl_descriptors; 1761 uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES]; 1762 uint32_t i; 1763 int rc; 1764 1765 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1766 rgroup = rqpair->poller->group; 1767 1768 inline_segment = &req->cmd->nvme_cmd.dptr.sgl1; 1769 assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT); 1770 assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET); 1771 1772 num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor); 1773 assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES); 1774 1775 if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) { 1776 return -ENOMEM; 1777 } 1778 1779 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1780 for (i = 0; i < num_sgl_descriptors; i++) { 1781 if (spdk_likely(!req->dif.dif_insert_or_strip)) { 1782 lengths[i] = desc->keyed.length; 1783 } else { 1784 req->dif.orig_length += desc->keyed.length; 1785 lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx); 1786 req->dif.elba_length += lengths[i]; 1787 } 1788 desc++; 1789 } 1790 1791 rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport, 1792 lengths, num_sgl_descriptors); 1793 if (rc != 0) { 1794 nvmf_rdma_request_free_data(rdma_req, rtransport); 1795 return rc; 1796 } 1797 1798 /* The first WR must always be the embedded data WR. This is how we unwind them later. */ 1799 current_wr = &rdma_req->data.wr; 1800 assert(current_wr != NULL); 1801 1802 req->length = 0; 1803 rdma_req->iovpos = 0; 1804 desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address; 1805 for (i = 0; i < num_sgl_descriptors; i++) { 1806 /* The descriptors must be keyed data block descriptors with an address, not an offset. */ 1807 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK || 1808 desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) { 1809 rc = -EINVAL; 1810 goto err_exit; 1811 } 1812 1813 current_wr->num_sge = 0; 1814 1815 rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0); 1816 if (rc != 0) { 1817 rc = -ENOMEM; 1818 goto err_exit; 1819 } 1820 1821 req->length += desc->keyed.length; 1822 current_wr->wr.rdma.rkey = desc->keyed.key; 1823 current_wr->wr.rdma.remote_addr = desc->address; 1824 current_wr = current_wr->next; 1825 desc++; 1826 } 1827 1828 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1829 /* Go back to the last descriptor in the list. */ 1830 desc--; 1831 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1832 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1833 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1834 rdma_req->rsp.wr.imm_data = desc->keyed.key; 1835 } 1836 } 1837 #endif 1838 1839 rdma_req->num_outstanding_data_wr = num_sgl_descriptors; 1840 1841 return 0; 1842 1843 err_exit: 1844 spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport); 1845 nvmf_rdma_request_free_data(rdma_req, rtransport); 1846 return rc; 1847 } 1848 1849 static int 1850 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport, 1851 struct spdk_nvmf_rdma_device *device, 1852 struct spdk_nvmf_rdma_request *rdma_req) 1853 { 1854 struct spdk_nvmf_request *req = &rdma_req->req; 1855 struct spdk_nvme_cpl *rsp; 1856 struct spdk_nvme_sgl_descriptor *sgl; 1857 int rc; 1858 uint32_t length; 1859 1860 rsp = &req->rsp->nvme_cpl; 1861 sgl = &req->cmd->nvme_cmd.dptr.sgl1; 1862 1863 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK && 1864 (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS || 1865 sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) { 1866 1867 length = sgl->keyed.length; 1868 if (length > rtransport->transport.opts.max_io_size) { 1869 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 1870 length, rtransport->transport.opts.max_io_size); 1871 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1872 return -1; 1873 } 1874 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 1875 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) { 1876 if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) { 1877 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV; 1878 rdma_req->rsp.wr.imm_data = sgl->keyed.key; 1879 } 1880 } 1881 #endif 1882 1883 /* fill request length and populate iovs */ 1884 req->length = length; 1885 1886 if (spdk_unlikely(req->dif.dif_insert_or_strip)) { 1887 req->dif.orig_length = length; 1888 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 1889 req->dif.elba_length = length; 1890 } 1891 1892 rc = spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length); 1893 if (spdk_unlikely(rc < 0)) { 1894 if (rc == -EINVAL) { 1895 SPDK_ERRLOG("SGL length exceeds the max I/O size\n"); 1896 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1897 return -1; 1898 } 1899 /* No available buffers. Queue this request up. */ 1900 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1901 return 0; 1902 } 1903 1904 /* backward compatible */ 1905 req->data = req->iov[0].iov_base; 1906 1907 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1908 req->iovcnt); 1909 1910 return 0; 1911 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 1912 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1913 uint64_t offset = sgl->address; 1914 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size; 1915 1916 SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 1917 offset, sgl->unkeyed.length); 1918 1919 if (offset > max_len) { 1920 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 1921 offset, max_len); 1922 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 1923 return -1; 1924 } 1925 max_len -= (uint32_t)offset; 1926 1927 if (sgl->unkeyed.length > max_len) { 1928 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 1929 sgl->unkeyed.length, max_len); 1930 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1931 return -1; 1932 } 1933 1934 rdma_req->num_outstanding_data_wr = 0; 1935 req->data = rdma_req->recv->buf + offset; 1936 req->data_from_pool = false; 1937 req->length = sgl->unkeyed.length; 1938 1939 req->iov[0].iov_base = req->data; 1940 req->iov[0].iov_len = req->length; 1941 req->iovcnt = 1; 1942 1943 return 0; 1944 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT && 1945 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 1946 1947 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req); 1948 if (rc == -ENOMEM) { 1949 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req); 1950 return 0; 1951 } else if (rc == -EINVAL) { 1952 SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n"); 1953 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 1954 return -1; 1955 } 1956 1957 /* backward compatible */ 1958 req->data = req->iov[0].iov_base; 1959 1960 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req, 1961 req->iovcnt); 1962 1963 return 0; 1964 } 1965 1966 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 1967 sgl->generic.type, sgl->generic.subtype); 1968 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 1969 return -1; 1970 } 1971 1972 static void 1973 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req, 1974 struct spdk_nvmf_rdma_transport *rtransport) 1975 { 1976 struct spdk_nvmf_rdma_qpair *rqpair; 1977 struct spdk_nvmf_rdma_poll_group *rgroup; 1978 1979 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 1980 if (rdma_req->req.data_from_pool) { 1981 rgroup = rqpair->poller->group; 1982 1983 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport); 1984 } 1985 nvmf_rdma_request_free_data(rdma_req, rtransport); 1986 rdma_req->req.length = 0; 1987 rdma_req->req.iovcnt = 0; 1988 rdma_req->req.data = NULL; 1989 rdma_req->rsp.wr.next = NULL; 1990 rdma_req->data.wr.next = NULL; 1991 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 1992 rqpair->qd--; 1993 1994 STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link); 1995 rdma_req->state = RDMA_REQUEST_STATE_FREE; 1996 } 1997 1998 static bool 1999 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport, 2000 struct spdk_nvmf_rdma_request *rdma_req) 2001 { 2002 struct spdk_nvmf_rdma_qpair *rqpair; 2003 struct spdk_nvmf_rdma_device *device; 2004 struct spdk_nvmf_rdma_poll_group *rgroup; 2005 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2006 int rc; 2007 struct spdk_nvmf_rdma_recv *rdma_recv; 2008 enum spdk_nvmf_rdma_request_state prev_state; 2009 bool progress = false; 2010 int data_posted; 2011 uint32_t num_blocks; 2012 2013 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 2014 device = rqpair->port->device; 2015 rgroup = rqpair->poller->group; 2016 2017 assert(rdma_req->state != RDMA_REQUEST_STATE_FREE); 2018 2019 /* If the queue pair is in an error state, force the request to the completed state 2020 * to release resources. */ 2021 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2022 if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) { 2023 STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link); 2024 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) { 2025 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2026 } else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) { 2027 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2028 } 2029 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2030 } 2031 2032 /* The loop here is to allow for several back-to-back state changes. */ 2033 do { 2034 prev_state = rdma_req->state; 2035 2036 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state); 2037 2038 switch (rdma_req->state) { 2039 case RDMA_REQUEST_STATE_FREE: 2040 /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW 2041 * to escape this state. */ 2042 break; 2043 case RDMA_REQUEST_STATE_NEW: 2044 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0, 2045 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2046 rdma_recv = rdma_req->recv; 2047 2048 /* The first element of the SGL is the NVMe command */ 2049 rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr; 2050 memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp)); 2051 2052 if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2053 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2054 break; 2055 } 2056 2057 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) { 2058 rdma_req->req.dif.dif_insert_or_strip = true; 2059 } 2060 2061 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2062 rdma_req->rsp.wr.opcode = IBV_WR_SEND; 2063 rdma_req->rsp.wr.imm_data = 0; 2064 #endif 2065 2066 /* The next state transition depends on the data transfer needs of this request. */ 2067 rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req); 2068 2069 /* If no data to transfer, ready to execute. */ 2070 if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) { 2071 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2072 break; 2073 } 2074 2075 rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER; 2076 STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link); 2077 break; 2078 case RDMA_REQUEST_STATE_NEED_BUFFER: 2079 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0, 2080 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2081 2082 assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE); 2083 2084 if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) { 2085 /* This request needs to wait in line to obtain a buffer */ 2086 break; 2087 } 2088 2089 /* Try to get a data buffer */ 2090 rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req); 2091 if (rc < 0) { 2092 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2093 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2094 break; 2095 } 2096 2097 if (!rdma_req->req.data) { 2098 /* No buffers available. */ 2099 rgroup->stat.pending_data_buffer++; 2100 break; 2101 } 2102 2103 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link); 2104 2105 /* If data is transferring from host to controller and the data didn't 2106 * arrive using in capsule data, we need to do a transfer from the host. 2107 */ 2108 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER && 2109 rdma_req->req.data_from_pool) { 2110 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link); 2111 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING; 2112 break; 2113 } 2114 2115 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 2116 break; 2117 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING: 2118 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0, 2119 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2120 2121 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) { 2122 /* This request needs to wait in line to perform RDMA */ 2123 break; 2124 } 2125 if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth 2126 || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) { 2127 /* We can only have so many WRs outstanding. we have to wait until some finish. */ 2128 rqpair->poller->stat.pending_rdma_read++; 2129 break; 2130 } 2131 2132 /* We have already verified that this request is the head of the queue. */ 2133 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link); 2134 2135 rc = request_transfer_in(&rdma_req->req); 2136 if (!rc) { 2137 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER; 2138 } else { 2139 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 2140 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2141 } 2142 break; 2143 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2144 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0, 2145 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2146 /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE 2147 * to escape this state. */ 2148 break; 2149 case RDMA_REQUEST_STATE_READY_TO_EXECUTE: 2150 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0, 2151 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2152 2153 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2154 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2155 /* generate DIF for write operation */ 2156 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2157 assert(num_blocks > 0); 2158 2159 rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt, 2160 num_blocks, &rdma_req->req.dif.dif_ctx); 2161 if (rc != 0) { 2162 SPDK_ERRLOG("DIF generation failed\n"); 2163 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2164 spdk_nvmf_rdma_start_disconnect(rqpair); 2165 break; 2166 } 2167 } 2168 2169 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length); 2170 /* set extended length before IO operation */ 2171 rdma_req->req.length = rdma_req->req.dif.elba_length; 2172 } 2173 2174 rdma_req->state = RDMA_REQUEST_STATE_EXECUTING; 2175 spdk_nvmf_request_exec(&rdma_req->req); 2176 break; 2177 case RDMA_REQUEST_STATE_EXECUTING: 2178 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0, 2179 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2180 /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED 2181 * to escape this state. */ 2182 break; 2183 case RDMA_REQUEST_STATE_EXECUTED: 2184 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0, 2185 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2186 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2187 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link); 2188 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING; 2189 } else { 2190 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2191 } 2192 if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) { 2193 /* restore the original length */ 2194 rdma_req->req.length = rdma_req->req.dif.orig_length; 2195 2196 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2197 struct spdk_dif_error error_blk; 2198 2199 num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size); 2200 2201 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks, 2202 &rdma_req->req.dif.dif_ctx, &error_blk); 2203 if (rc) { 2204 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl; 2205 2206 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type, 2207 error_blk.err_offset); 2208 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2209 rsp->status.sc = spdk_nvmf_rdma_dif_error_to_compl_status(error_blk.err_type); 2210 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2211 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link); 2212 } 2213 } 2214 } 2215 break; 2216 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING: 2217 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0, 2218 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2219 2220 if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) { 2221 /* This request needs to wait in line to perform RDMA */ 2222 break; 2223 } 2224 if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) > 2225 rqpair->max_send_depth) { 2226 /* We can only have so many WRs outstanding. we have to wait until some finish. 2227 * +1 since each request has an additional wr in the resp. */ 2228 rqpair->poller->stat.pending_rdma_write++; 2229 break; 2230 } 2231 2232 /* We have already verified that this request is the head of the queue. */ 2233 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link); 2234 2235 /* The data transfer will be kicked off from 2236 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state. 2237 */ 2238 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 2239 break; 2240 case RDMA_REQUEST_STATE_READY_TO_COMPLETE: 2241 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0, 2242 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2243 rc = request_transfer_out(&rdma_req->req, &data_posted); 2244 assert(rc == 0); /* No good way to handle this currently */ 2245 if (rc) { 2246 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 2247 } else { 2248 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST : 2249 RDMA_REQUEST_STATE_COMPLETING; 2250 } 2251 break; 2252 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2253 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0, 2254 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2255 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2256 * to escape this state. */ 2257 break; 2258 case RDMA_REQUEST_STATE_COMPLETING: 2259 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0, 2260 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2261 /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED 2262 * to escape this state. */ 2263 break; 2264 case RDMA_REQUEST_STATE_COMPLETED: 2265 spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0, 2266 (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id); 2267 2268 rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc; 2269 nvmf_rdma_request_free(rdma_req, rtransport); 2270 break; 2271 case RDMA_REQUEST_NUM_STATES: 2272 default: 2273 assert(0); 2274 break; 2275 } 2276 2277 if (rdma_req->state != prev_state) { 2278 progress = true; 2279 } 2280 } while (rdma_req->state != prev_state); 2281 2282 return progress; 2283 } 2284 2285 /* Public API callbacks begin here */ 2286 2287 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128 2288 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128 2289 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096 2290 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 2291 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 2292 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072 2293 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES) 2294 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095 2295 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32 2296 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false 2297 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false 2298 2299 static void 2300 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts) 2301 { 2302 opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH; 2303 opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR; 2304 opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE; 2305 opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE; 2306 opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE; 2307 opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH; 2308 opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS; 2309 opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE; 2310 opts->max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH; 2311 opts->no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ; 2312 opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP; 2313 } 2314 2315 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = { 2316 .notify_cb = spdk_nvmf_rdma_mem_notify, 2317 .are_contiguous = spdk_nvmf_rdma_check_contiguous_entries 2318 }; 2319 2320 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport); 2321 2322 static struct spdk_nvmf_transport * 2323 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts) 2324 { 2325 int rc; 2326 struct spdk_nvmf_rdma_transport *rtransport; 2327 struct spdk_nvmf_rdma_device *device, *tmp; 2328 struct ibv_context **contexts; 2329 uint32_t i; 2330 int flag; 2331 uint32_t sge_count; 2332 uint32_t min_shared_buffers; 2333 int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 2334 pthread_mutexattr_t attr; 2335 2336 rtransport = calloc(1, sizeof(*rtransport)); 2337 if (!rtransport) { 2338 return NULL; 2339 } 2340 2341 if (pthread_mutexattr_init(&attr)) { 2342 SPDK_ERRLOG("pthread_mutexattr_init() failed\n"); 2343 free(rtransport); 2344 return NULL; 2345 } 2346 2347 if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) { 2348 SPDK_ERRLOG("pthread_mutexattr_settype() failed\n"); 2349 pthread_mutexattr_destroy(&attr); 2350 free(rtransport); 2351 return NULL; 2352 } 2353 2354 if (pthread_mutex_init(&rtransport->lock, &attr)) { 2355 SPDK_ERRLOG("pthread_mutex_init() failed\n"); 2356 pthread_mutexattr_destroy(&attr); 2357 free(rtransport); 2358 return NULL; 2359 } 2360 2361 pthread_mutexattr_destroy(&attr); 2362 2363 TAILQ_INIT(&rtransport->devices); 2364 TAILQ_INIT(&rtransport->ports); 2365 TAILQ_INIT(&rtransport->poll_groups); 2366 2367 rtransport->transport.ops = &spdk_nvmf_transport_rdma; 2368 2369 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n" 2370 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 2371 " max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 2372 " in_capsule_data_size=%d, max_aq_depth=%d,\n" 2373 " num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d\n", 2374 opts->max_queue_depth, 2375 opts->max_io_size, 2376 opts->max_qpairs_per_ctrlr, 2377 opts->io_unit_size, 2378 opts->in_capsule_data_size, 2379 opts->max_aq_depth, 2380 opts->num_shared_buffers, 2381 opts->max_srq_depth, 2382 opts->no_srq); 2383 2384 /* I/O unit size cannot be larger than max I/O size */ 2385 if (opts->io_unit_size > opts->max_io_size) { 2386 opts->io_unit_size = opts->max_io_size; 2387 } 2388 2389 if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) { 2390 SPDK_ERRLOG("The number of shared data buffers (%d) is less than" 2391 "the minimum number required to guarantee that forward progress can be made (%d)\n", 2392 opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2)); 2393 spdk_nvmf_rdma_destroy(&rtransport->transport); 2394 return NULL; 2395 } 2396 2397 min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size; 2398 if (min_shared_buffers > opts->num_shared_buffers) { 2399 SPDK_ERRLOG("There are not enough buffers to satisfy" 2400 "per-poll group caches for each thread. (%" PRIu32 ")" 2401 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 2402 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 2403 spdk_nvmf_rdma_destroy(&rtransport->transport); 2404 return NULL; 2405 } 2406 2407 sge_count = opts->max_io_size / opts->io_unit_size; 2408 if (sge_count > NVMF_DEFAULT_TX_SGE) { 2409 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 2410 spdk_nvmf_rdma_destroy(&rtransport->transport); 2411 return NULL; 2412 } 2413 2414 rtransport->event_channel = rdma_create_event_channel(); 2415 if (rtransport->event_channel == NULL) { 2416 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno)); 2417 spdk_nvmf_rdma_destroy(&rtransport->transport); 2418 return NULL; 2419 } 2420 2421 flag = fcntl(rtransport->event_channel->fd, F_GETFL); 2422 if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) { 2423 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n", 2424 rtransport->event_channel->fd, spdk_strerror(errno)); 2425 spdk_nvmf_rdma_destroy(&rtransport->transport); 2426 return NULL; 2427 } 2428 2429 rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 2430 opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES, 2431 sizeof(struct spdk_nvmf_rdma_request_data), 2432 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 2433 SPDK_ENV_SOCKET_ID_ANY); 2434 if (!rtransport->data_wr_pool) { 2435 SPDK_ERRLOG("Unable to allocate work request pool for poll group\n"); 2436 spdk_nvmf_rdma_destroy(&rtransport->transport); 2437 return NULL; 2438 } 2439 2440 contexts = rdma_get_devices(NULL); 2441 if (contexts == NULL) { 2442 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno); 2443 spdk_nvmf_rdma_destroy(&rtransport->transport); 2444 return NULL; 2445 } 2446 2447 i = 0; 2448 rc = 0; 2449 while (contexts[i] != NULL) { 2450 device = calloc(1, sizeof(*device)); 2451 if (!device) { 2452 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n"); 2453 rc = -ENOMEM; 2454 break; 2455 } 2456 device->context = contexts[i]; 2457 rc = ibv_query_device(device->context, &device->attr); 2458 if (rc < 0) { 2459 SPDK_ERRLOG("Failed to query RDMA device attributes.\n"); 2460 free(device); 2461 break; 2462 2463 } 2464 2465 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge); 2466 2467 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL 2468 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) { 2469 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,"); 2470 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id); 2471 } 2472 2473 /** 2474 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE. 2475 * The Soft-RoCE RXE driver does not currently support send with invalidate, 2476 * but incorrectly reports that it does. There are changes making their way 2477 * through the kernel now that will enable this feature. When they are merged, 2478 * we can conditionally enable this feature. 2479 * 2480 * TODO: enable this for versions of the kernel rxe driver that support it. 2481 */ 2482 if (device->attr.vendor_id == 0) { 2483 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS); 2484 } 2485 #endif 2486 2487 /* set up device context async ev fd as NON_BLOCKING */ 2488 flag = fcntl(device->context->async_fd, F_GETFL); 2489 rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK); 2490 if (rc < 0) { 2491 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n"); 2492 free(device); 2493 break; 2494 } 2495 2496 TAILQ_INSERT_TAIL(&rtransport->devices, device, link); 2497 i++; 2498 2499 if (g_nvmf_hooks.get_ibv_pd) { 2500 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context); 2501 } else { 2502 device->pd = ibv_alloc_pd(device->context); 2503 } 2504 2505 if (!device->pd) { 2506 SPDK_ERRLOG("Unable to allocate protection domain.\n"); 2507 rc = -ENOMEM; 2508 break; 2509 } 2510 2511 assert(device->map == NULL); 2512 2513 device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd); 2514 if (!device->map) { 2515 SPDK_ERRLOG("Unable to allocate memory map for listen address\n"); 2516 rc = -ENOMEM; 2517 break; 2518 } 2519 2520 assert(device->map != NULL); 2521 assert(device->pd != NULL); 2522 } 2523 rdma_free_devices(contexts); 2524 2525 if (opts->io_unit_size * max_device_sge < opts->max_io_size) { 2526 /* divide and round up. */ 2527 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge; 2528 2529 /* round up to the nearest 4k. */ 2530 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK; 2531 2532 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 2533 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n", 2534 opts->io_unit_size); 2535 } 2536 2537 if (rc < 0) { 2538 spdk_nvmf_rdma_destroy(&rtransport->transport); 2539 return NULL; 2540 } 2541 2542 /* Set up poll descriptor array to monitor events from RDMA and IB 2543 * in a single poll syscall 2544 */ 2545 rtransport->npoll_fds = i + 1; 2546 i = 0; 2547 rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd)); 2548 if (rtransport->poll_fds == NULL) { 2549 SPDK_ERRLOG("poll_fds allocation failed\n"); 2550 spdk_nvmf_rdma_destroy(&rtransport->transport); 2551 return NULL; 2552 } 2553 2554 rtransport->poll_fds[i].fd = rtransport->event_channel->fd; 2555 rtransport->poll_fds[i++].events = POLLIN; 2556 2557 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 2558 rtransport->poll_fds[i].fd = device->context->async_fd; 2559 rtransport->poll_fds[i++].events = POLLIN; 2560 } 2561 2562 return &rtransport->transport; 2563 } 2564 2565 static int 2566 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport) 2567 { 2568 struct spdk_nvmf_rdma_transport *rtransport; 2569 struct spdk_nvmf_rdma_port *port, *port_tmp; 2570 struct spdk_nvmf_rdma_device *device, *device_tmp; 2571 2572 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2573 2574 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) { 2575 TAILQ_REMOVE(&rtransport->ports, port, link); 2576 rdma_destroy_id(port->id); 2577 free(port); 2578 } 2579 2580 if (rtransport->poll_fds != NULL) { 2581 free(rtransport->poll_fds); 2582 } 2583 2584 if (rtransport->event_channel != NULL) { 2585 rdma_destroy_event_channel(rtransport->event_channel); 2586 } 2587 2588 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) { 2589 TAILQ_REMOVE(&rtransport->devices, device, link); 2590 if (device->map) { 2591 spdk_mem_map_free(&device->map); 2592 } 2593 if (device->pd) { 2594 if (!g_nvmf_hooks.get_ibv_pd) { 2595 ibv_dealloc_pd(device->pd); 2596 } 2597 } 2598 free(device); 2599 } 2600 2601 if (rtransport->data_wr_pool != NULL) { 2602 if (spdk_mempool_count(rtransport->data_wr_pool) != 2603 (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) { 2604 SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n", 2605 spdk_mempool_count(rtransport->data_wr_pool), 2606 transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES); 2607 } 2608 } 2609 2610 spdk_mempool_free(rtransport->data_wr_pool); 2611 2612 pthread_mutex_destroy(&rtransport->lock); 2613 free(rtransport); 2614 2615 return 0; 2616 } 2617 2618 static int 2619 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 2620 struct spdk_nvme_transport_id *trid, 2621 bool peer); 2622 2623 static int 2624 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport, 2625 const struct spdk_nvme_transport_id *trid) 2626 { 2627 struct spdk_nvmf_rdma_transport *rtransport; 2628 struct spdk_nvmf_rdma_device *device; 2629 struct spdk_nvmf_rdma_port *port_tmp, *port; 2630 struct addrinfo *res; 2631 struct addrinfo hints; 2632 int family; 2633 int rc; 2634 2635 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2636 2637 port = calloc(1, sizeof(*port)); 2638 if (!port) { 2639 return -ENOMEM; 2640 } 2641 2642 /* Selectively copy the trid. Things like NQN don't matter here - that 2643 * mapping is enforced elsewhere. 2644 */ 2645 port->trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2646 port->trid.adrfam = trid->adrfam; 2647 snprintf(port->trid.traddr, sizeof(port->trid.traddr), "%s", trid->traddr); 2648 snprintf(port->trid.trsvcid, sizeof(port->trid.trsvcid), "%s", trid->trsvcid); 2649 2650 switch (port->trid.adrfam) { 2651 case SPDK_NVMF_ADRFAM_IPV4: 2652 family = AF_INET; 2653 break; 2654 case SPDK_NVMF_ADRFAM_IPV6: 2655 family = AF_INET6; 2656 break; 2657 default: 2658 SPDK_ERRLOG("Unhandled ADRFAM %d\n", port->trid.adrfam); 2659 free(port); 2660 return -EINVAL; 2661 } 2662 2663 memset(&hints, 0, sizeof(hints)); 2664 hints.ai_family = family; 2665 hints.ai_flags = AI_NUMERICSERV; 2666 hints.ai_socktype = SOCK_STREAM; 2667 hints.ai_protocol = 0; 2668 2669 rc = getaddrinfo(port->trid.traddr, port->trid.trsvcid, &hints, &res); 2670 if (rc) { 2671 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc); 2672 free(port); 2673 return -EINVAL; 2674 } 2675 2676 pthread_mutex_lock(&rtransport->lock); 2677 assert(rtransport->event_channel != NULL); 2678 TAILQ_FOREACH(port_tmp, &rtransport->ports, link) { 2679 if (spdk_nvme_transport_id_compare(&port_tmp->trid, &port->trid) == 0) { 2680 port_tmp->ref++; 2681 freeaddrinfo(res); 2682 free(port); 2683 /* Already listening at this address */ 2684 pthread_mutex_unlock(&rtransport->lock); 2685 return 0; 2686 } 2687 } 2688 2689 rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP); 2690 if (rc < 0) { 2691 SPDK_ERRLOG("rdma_create_id() failed\n"); 2692 freeaddrinfo(res); 2693 free(port); 2694 pthread_mutex_unlock(&rtransport->lock); 2695 return rc; 2696 } 2697 2698 rc = rdma_bind_addr(port->id, res->ai_addr); 2699 freeaddrinfo(res); 2700 2701 if (rc < 0) { 2702 SPDK_ERRLOG("rdma_bind_addr() failed\n"); 2703 rdma_destroy_id(port->id); 2704 free(port); 2705 pthread_mutex_unlock(&rtransport->lock); 2706 return rc; 2707 } 2708 2709 if (!port->id->verbs) { 2710 SPDK_ERRLOG("ibv_context is null\n"); 2711 rdma_destroy_id(port->id); 2712 free(port); 2713 pthread_mutex_unlock(&rtransport->lock); 2714 return -1; 2715 } 2716 2717 rc = rdma_listen(port->id, 10); /* 10 = backlog */ 2718 if (rc < 0) { 2719 SPDK_ERRLOG("rdma_listen() failed\n"); 2720 rdma_destroy_id(port->id); 2721 free(port); 2722 pthread_mutex_unlock(&rtransport->lock); 2723 return rc; 2724 } 2725 2726 TAILQ_FOREACH(device, &rtransport->devices, link) { 2727 if (device->context == port->id->verbs) { 2728 port->device = device; 2729 break; 2730 } 2731 } 2732 if (!port->device) { 2733 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n", 2734 port->id->verbs); 2735 rdma_destroy_id(port->id); 2736 free(port); 2737 pthread_mutex_unlock(&rtransport->lock); 2738 return -EINVAL; 2739 } 2740 2741 SPDK_INFOLOG(SPDK_LOG_RDMA, "*** NVMf Target Listening on %s port %d ***\n", 2742 port->trid.traddr, ntohs(rdma_get_src_port(port->id))); 2743 2744 port->ref = 1; 2745 2746 TAILQ_INSERT_TAIL(&rtransport->ports, port, link); 2747 pthread_mutex_unlock(&rtransport->lock); 2748 2749 return 0; 2750 } 2751 2752 static int 2753 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport, 2754 const struct spdk_nvme_transport_id *_trid) 2755 { 2756 struct spdk_nvmf_rdma_transport *rtransport; 2757 struct spdk_nvmf_rdma_port *port, *tmp; 2758 struct spdk_nvme_transport_id trid = {}; 2759 2760 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2761 2762 /* Selectively copy the trid. Things like NQN don't matter here - that 2763 * mapping is enforced elsewhere. 2764 */ 2765 trid.trtype = SPDK_NVME_TRANSPORT_RDMA; 2766 trid.adrfam = _trid->adrfam; 2767 snprintf(trid.traddr, sizeof(port->trid.traddr), "%s", _trid->traddr); 2768 snprintf(trid.trsvcid, sizeof(port->trid.trsvcid), "%s", _trid->trsvcid); 2769 2770 pthread_mutex_lock(&rtransport->lock); 2771 TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) { 2772 if (spdk_nvme_transport_id_compare(&port->trid, &trid) == 0) { 2773 assert(port->ref > 0); 2774 port->ref--; 2775 if (port->ref == 0) { 2776 TAILQ_REMOVE(&rtransport->ports, port, link); 2777 rdma_destroy_id(port->id); 2778 free(port); 2779 } 2780 break; 2781 } 2782 } 2783 2784 pthread_mutex_unlock(&rtransport->lock); 2785 return 0; 2786 } 2787 2788 static void 2789 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport, 2790 struct spdk_nvmf_rdma_qpair *rqpair, bool drain) 2791 { 2792 struct spdk_nvmf_request *req, *tmp; 2793 struct spdk_nvmf_rdma_request *rdma_req, *req_tmp; 2794 struct spdk_nvmf_rdma_resources *resources; 2795 2796 /* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */ 2797 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) { 2798 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2799 break; 2800 } 2801 } 2802 2803 /* Then RDMA writes since reads have stronger restrictions than writes */ 2804 STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) { 2805 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2806 break; 2807 } 2808 } 2809 2810 /* The second highest priority is I/O waiting on memory buffers. */ 2811 STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) { 2812 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 2813 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) { 2814 break; 2815 } 2816 } 2817 2818 resources = rqpair->resources; 2819 while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) { 2820 rdma_req = STAILQ_FIRST(&resources->free_queue); 2821 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link); 2822 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue); 2823 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link); 2824 2825 if (rqpair->srq != NULL) { 2826 rdma_req->req.qpair = &rdma_req->recv->qpair->qpair; 2827 rdma_req->recv->qpair->qd++; 2828 } else { 2829 rqpair->qd++; 2830 } 2831 2832 rdma_req->receive_tsc = rdma_req->recv->receive_tsc; 2833 rdma_req->state = RDMA_REQUEST_STATE_NEW; 2834 if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) { 2835 break; 2836 } 2837 } 2838 if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) { 2839 rqpair->poller->stat.pending_free_request++; 2840 } 2841 } 2842 2843 static void 2844 _nvmf_rdma_qpair_disconnect(void *ctx) 2845 { 2846 struct spdk_nvmf_qpair *qpair = ctx; 2847 2848 spdk_nvmf_qpair_disconnect(qpair, NULL, NULL); 2849 } 2850 2851 static void 2852 _nvmf_rdma_try_disconnect(void *ctx) 2853 { 2854 struct spdk_nvmf_qpair *qpair = ctx; 2855 struct spdk_nvmf_poll_group *group; 2856 2857 /* Read the group out of the qpair. This is normally set and accessed only from 2858 * the thread that created the group. Here, we're not on that thread necessarily. 2859 * The data member qpair->group begins it's life as NULL and then is assigned to 2860 * a pointer and never changes. So fortunately reading this and checking for 2861 * non-NULL is thread safe in the x86_64 memory model. */ 2862 group = qpair->group; 2863 2864 if (group == NULL) { 2865 /* The qpair hasn't been assigned to a group yet, so we can't 2866 * process a disconnect. Send a message to ourself and try again. */ 2867 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair); 2868 return; 2869 } 2870 2871 spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair); 2872 } 2873 2874 static inline void 2875 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair) 2876 { 2877 if (!__atomic_test_and_set(&rqpair->disconnect_started, __ATOMIC_RELAXED)) { 2878 _nvmf_rdma_try_disconnect(&rqpair->qpair); 2879 } 2880 } 2881 2882 static void nvmf_rdma_destroy_drained_qpair(void *ctx) 2883 { 2884 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 2885 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 2886 struct spdk_nvmf_rdma_transport, transport); 2887 2888 /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */ 2889 if (rqpair->current_send_depth != 0) { 2890 return; 2891 } 2892 2893 if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) { 2894 return; 2895 } 2896 2897 if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) { 2898 return; 2899 } 2900 2901 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 2902 2903 /* Qpair will be destroyed after nvmf layer closes this qpair */ 2904 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ERROR) { 2905 return; 2906 } 2907 2908 spdk_nvmf_rdma_qpair_destroy(rqpair); 2909 } 2910 2911 2912 static int 2913 nvmf_rdma_disconnect(struct rdma_cm_event *evt) 2914 { 2915 struct spdk_nvmf_qpair *qpair; 2916 struct spdk_nvmf_rdma_qpair *rqpair; 2917 2918 if (evt->id == NULL) { 2919 SPDK_ERRLOG("disconnect request: missing cm_id\n"); 2920 return -1; 2921 } 2922 2923 qpair = evt->id->context; 2924 if (qpair == NULL) { 2925 SPDK_ERRLOG("disconnect request: no active connection\n"); 2926 return -1; 2927 } 2928 2929 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 2930 2931 spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0); 2932 2933 spdk_nvmf_rdma_start_disconnect(rqpair); 2934 2935 return 0; 2936 } 2937 2938 #ifdef DEBUG 2939 static const char *CM_EVENT_STR[] = { 2940 "RDMA_CM_EVENT_ADDR_RESOLVED", 2941 "RDMA_CM_EVENT_ADDR_ERROR", 2942 "RDMA_CM_EVENT_ROUTE_RESOLVED", 2943 "RDMA_CM_EVENT_ROUTE_ERROR", 2944 "RDMA_CM_EVENT_CONNECT_REQUEST", 2945 "RDMA_CM_EVENT_CONNECT_RESPONSE", 2946 "RDMA_CM_EVENT_CONNECT_ERROR", 2947 "RDMA_CM_EVENT_UNREACHABLE", 2948 "RDMA_CM_EVENT_REJECTED", 2949 "RDMA_CM_EVENT_ESTABLISHED", 2950 "RDMA_CM_EVENT_DISCONNECTED", 2951 "RDMA_CM_EVENT_DEVICE_REMOVAL", 2952 "RDMA_CM_EVENT_MULTICAST_JOIN", 2953 "RDMA_CM_EVENT_MULTICAST_ERROR", 2954 "RDMA_CM_EVENT_ADDR_CHANGE", 2955 "RDMA_CM_EVENT_TIMEWAIT_EXIT" 2956 }; 2957 #endif /* DEBUG */ 2958 2959 static void 2960 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn, void *cb_arg) 2961 { 2962 struct spdk_nvmf_rdma_transport *rtransport; 2963 struct rdma_cm_event *event; 2964 int rc; 2965 2966 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 2967 2968 if (rtransport->event_channel == NULL) { 2969 return; 2970 } 2971 2972 while (1) { 2973 rc = rdma_get_cm_event(rtransport->event_channel, &event); 2974 if (rc == 0) { 2975 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]); 2976 2977 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event); 2978 2979 switch (event->event) { 2980 case RDMA_CM_EVENT_ADDR_RESOLVED: 2981 case RDMA_CM_EVENT_ADDR_ERROR: 2982 case RDMA_CM_EVENT_ROUTE_RESOLVED: 2983 case RDMA_CM_EVENT_ROUTE_ERROR: 2984 /* No action required. The target never attempts to resolve routes. */ 2985 break; 2986 case RDMA_CM_EVENT_CONNECT_REQUEST: 2987 rc = nvmf_rdma_connect(transport, event, cb_fn, cb_arg); 2988 if (rc < 0) { 2989 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc); 2990 break; 2991 } 2992 break; 2993 case RDMA_CM_EVENT_CONNECT_RESPONSE: 2994 /* The target never initiates a new connection. So this will not occur. */ 2995 break; 2996 case RDMA_CM_EVENT_CONNECT_ERROR: 2997 /* Can this happen? The docs say it can, but not sure what causes it. */ 2998 break; 2999 case RDMA_CM_EVENT_UNREACHABLE: 3000 case RDMA_CM_EVENT_REJECTED: 3001 /* These only occur on the client side. */ 3002 break; 3003 case RDMA_CM_EVENT_ESTABLISHED: 3004 /* TODO: Should we be waiting for this event anywhere? */ 3005 break; 3006 case RDMA_CM_EVENT_DISCONNECTED: 3007 case RDMA_CM_EVENT_DEVICE_REMOVAL: 3008 rc = nvmf_rdma_disconnect(event); 3009 if (rc < 0) { 3010 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc); 3011 break; 3012 } 3013 break; 3014 case RDMA_CM_EVENT_MULTICAST_JOIN: 3015 case RDMA_CM_EVENT_MULTICAST_ERROR: 3016 /* Multicast is not used */ 3017 break; 3018 case RDMA_CM_EVENT_ADDR_CHANGE: 3019 /* Not utilizing this event */ 3020 break; 3021 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 3022 /* For now, do nothing. The target never re-uses queue pairs. */ 3023 break; 3024 default: 3025 SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event); 3026 break; 3027 } 3028 3029 rdma_ack_cm_event(event); 3030 } else { 3031 if (errno != EAGAIN && errno != EWOULDBLOCK) { 3032 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno)); 3033 } 3034 break; 3035 } 3036 } 3037 } 3038 3039 static void 3040 nvmf_rdma_handle_qp_fatal(struct spdk_nvmf_rdma_qpair *rqpair) 3041 { 3042 spdk_nvmf_rdma_update_ibv_state(rqpair); 3043 spdk_nvmf_rdma_start_disconnect(rqpair); 3044 } 3045 3046 static void 3047 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair) 3048 { 3049 rqpair->last_wqe_reached = true; 3050 nvmf_rdma_destroy_drained_qpair(rqpair); 3051 } 3052 3053 static void 3054 nvmf_rdma_handle_sq_drained(struct spdk_nvmf_rdma_qpair *rqpair) 3055 { 3056 spdk_nvmf_rdma_start_disconnect(rqpair); 3057 } 3058 3059 static void 3060 spdk_nvmf_rdma_qpair_process_ibv_event(void *ctx) 3061 { 3062 struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx; 3063 3064 if (event_ctx->rqpair) { 3065 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link); 3066 if (event_ctx->cb_fn) { 3067 event_ctx->cb_fn(event_ctx->rqpair); 3068 } 3069 } 3070 free(event_ctx); 3071 } 3072 3073 static int 3074 spdk_nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair, 3075 spdk_nvmf_rdma_qpair_ibv_event fn) 3076 { 3077 struct spdk_nvmf_rdma_ibv_event_ctx *ctx; 3078 3079 if (!rqpair->qpair.group) { 3080 return EINVAL; 3081 } 3082 3083 ctx = calloc(1, sizeof(*ctx)); 3084 if (!ctx) { 3085 return ENOMEM; 3086 } 3087 3088 ctx->rqpair = rqpair; 3089 ctx->cb_fn = fn; 3090 STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link); 3091 3092 return spdk_thread_send_msg(rqpair->qpair.group->thread, spdk_nvmf_rdma_qpair_process_ibv_event, 3093 ctx); 3094 } 3095 3096 static void 3097 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device) 3098 { 3099 int rc; 3100 struct spdk_nvmf_rdma_qpair *rqpair = NULL; 3101 struct ibv_async_event event; 3102 3103 rc = ibv_get_async_event(device->context, &event); 3104 3105 if (rc) { 3106 SPDK_ERRLOG("Failed to get async_event (%d): %s\n", 3107 errno, spdk_strerror(errno)); 3108 return; 3109 } 3110 3111 switch (event.event_type) { 3112 case IBV_EVENT_QP_FATAL: 3113 rqpair = event.element.qp->qp_context; 3114 SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair); 3115 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3116 (uintptr_t)rqpair->cm_id, event.event_type); 3117 if (spdk_nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_qp_fatal)) { 3118 SPDK_ERRLOG("Failed to send QP_FATAL event for rqpair %p\n", rqpair); 3119 nvmf_rdma_handle_qp_fatal(rqpair); 3120 } 3121 break; 3122 case IBV_EVENT_QP_LAST_WQE_REACHED: 3123 /* This event only occurs for shared receive queues. */ 3124 rqpair = event.element.qp->qp_context; 3125 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last WQE reached event received for rqpair %p\n", rqpair); 3126 if (spdk_nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached)) { 3127 SPDK_ERRLOG("Failed to send LAST_WQE_REACHED event for rqpair %p\n", rqpair); 3128 rqpair->last_wqe_reached = true; 3129 } 3130 break; 3131 case IBV_EVENT_SQ_DRAINED: 3132 /* This event occurs frequently in both error and non-error states. 3133 * Check if the qpair is in an error state before sending a message. */ 3134 rqpair = event.element.qp->qp_context; 3135 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last sq drained event received for rqpair %p\n", rqpair); 3136 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3137 (uintptr_t)rqpair->cm_id, event.event_type); 3138 if (spdk_nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) { 3139 if (spdk_nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_sq_drained)) { 3140 SPDK_ERRLOG("Failed to send SQ_DRAINED event for rqpair %p\n", rqpair); 3141 nvmf_rdma_handle_sq_drained(rqpair); 3142 } 3143 } 3144 break; 3145 case IBV_EVENT_QP_REQ_ERR: 3146 case IBV_EVENT_QP_ACCESS_ERR: 3147 case IBV_EVENT_COMM_EST: 3148 case IBV_EVENT_PATH_MIG: 3149 case IBV_EVENT_PATH_MIG_ERR: 3150 SPDK_NOTICELOG("Async event: %s\n", 3151 ibv_event_type_str(event.event_type)); 3152 rqpair = event.element.qp->qp_context; 3153 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 3154 (uintptr_t)rqpair->cm_id, event.event_type); 3155 spdk_nvmf_rdma_update_ibv_state(rqpair); 3156 break; 3157 case IBV_EVENT_CQ_ERR: 3158 case IBV_EVENT_DEVICE_FATAL: 3159 case IBV_EVENT_PORT_ACTIVE: 3160 case IBV_EVENT_PORT_ERR: 3161 case IBV_EVENT_LID_CHANGE: 3162 case IBV_EVENT_PKEY_CHANGE: 3163 case IBV_EVENT_SM_CHANGE: 3164 case IBV_EVENT_SRQ_ERR: 3165 case IBV_EVENT_SRQ_LIMIT_REACHED: 3166 case IBV_EVENT_CLIENT_REREGISTER: 3167 case IBV_EVENT_GID_CHANGE: 3168 default: 3169 SPDK_NOTICELOG("Async event: %s\n", 3170 ibv_event_type_str(event.event_type)); 3171 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type); 3172 break; 3173 } 3174 ibv_ack_async_event(&event); 3175 } 3176 3177 static void 3178 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn, void *cb_arg) 3179 { 3180 int nfds, i = 0; 3181 struct spdk_nvmf_rdma_transport *rtransport; 3182 struct spdk_nvmf_rdma_device *device, *tmp; 3183 3184 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3185 nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0); 3186 3187 if (nfds <= 0) { 3188 return; 3189 } 3190 3191 /* The first poll descriptor is RDMA CM event */ 3192 if (rtransport->poll_fds[i++].revents & POLLIN) { 3193 spdk_nvmf_process_cm_event(transport, cb_fn, cb_arg); 3194 nfds--; 3195 } 3196 3197 if (nfds == 0) { 3198 return; 3199 } 3200 3201 /* Second and subsequent poll descriptors are IB async events */ 3202 TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) { 3203 if (rtransport->poll_fds[i++].revents & POLLIN) { 3204 spdk_nvmf_process_ib_event(device); 3205 nfds--; 3206 } 3207 } 3208 /* check all flagged fd's have been served */ 3209 assert(nfds == 0); 3210 } 3211 3212 static void 3213 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport, 3214 struct spdk_nvme_transport_id *trid, 3215 struct spdk_nvmf_discovery_log_page_entry *entry) 3216 { 3217 entry->trtype = SPDK_NVMF_TRTYPE_RDMA; 3218 entry->adrfam = trid->adrfam; 3219 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 3220 3221 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 3222 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 3223 3224 entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED; 3225 entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE; 3226 entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM; 3227 } 3228 3229 static void 3230 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 3231 3232 static struct spdk_nvmf_transport_poll_group * 3233 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 3234 { 3235 struct spdk_nvmf_rdma_transport *rtransport; 3236 struct spdk_nvmf_rdma_poll_group *rgroup; 3237 struct spdk_nvmf_rdma_poller *poller; 3238 struct spdk_nvmf_rdma_device *device; 3239 struct ibv_srq_init_attr srq_init_attr; 3240 struct spdk_nvmf_rdma_resource_opts opts; 3241 int num_cqe; 3242 3243 rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport); 3244 3245 rgroup = calloc(1, sizeof(*rgroup)); 3246 if (!rgroup) { 3247 return NULL; 3248 } 3249 3250 TAILQ_INIT(&rgroup->pollers); 3251 STAILQ_INIT(&rgroup->retired_bufs); 3252 3253 pthread_mutex_lock(&rtransport->lock); 3254 TAILQ_FOREACH(device, &rtransport->devices, link) { 3255 poller = calloc(1, sizeof(*poller)); 3256 if (!poller) { 3257 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n"); 3258 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3259 pthread_mutex_unlock(&rtransport->lock); 3260 return NULL; 3261 } 3262 3263 poller->device = device; 3264 poller->group = rgroup; 3265 3266 TAILQ_INIT(&poller->qpairs); 3267 STAILQ_INIT(&poller->qpairs_pending_send); 3268 STAILQ_INIT(&poller->qpairs_pending_recv); 3269 3270 TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link); 3271 if (transport->opts.no_srq == false && device->num_srq < device->attr.max_srq) { 3272 poller->max_srq_depth = transport->opts.max_srq_depth; 3273 3274 device->num_srq++; 3275 memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr)); 3276 srq_init_attr.attr.max_wr = poller->max_srq_depth; 3277 srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE); 3278 poller->srq = ibv_create_srq(device->pd, &srq_init_attr); 3279 if (!poller->srq) { 3280 SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno); 3281 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3282 pthread_mutex_unlock(&rtransport->lock); 3283 return NULL; 3284 } 3285 3286 opts.qp = poller->srq; 3287 opts.pd = device->pd; 3288 opts.qpair = NULL; 3289 opts.shared = true; 3290 opts.max_queue_depth = poller->max_srq_depth; 3291 opts.in_capsule_data_size = transport->opts.in_capsule_data_size; 3292 3293 poller->resources = nvmf_rdma_resources_create(&opts); 3294 if (!poller->resources) { 3295 SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n"); 3296 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3297 pthread_mutex_unlock(&rtransport->lock); 3298 return NULL; 3299 } 3300 } 3301 3302 /* 3303 * When using an srq, we can limit the completion queue at startup. 3304 * The following formula represents the calculation: 3305 * num_cqe = num_recv + num_data_wr + num_send_wr. 3306 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth 3307 */ 3308 if (poller->srq) { 3309 num_cqe = poller->max_srq_depth * 3; 3310 } else { 3311 num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE; 3312 } 3313 3314 poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0); 3315 if (!poller->cq) { 3316 SPDK_ERRLOG("Unable to create completion queue\n"); 3317 spdk_nvmf_rdma_poll_group_destroy(&rgroup->group); 3318 pthread_mutex_unlock(&rtransport->lock); 3319 return NULL; 3320 } 3321 poller->num_cqe = num_cqe; 3322 } 3323 3324 TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link); 3325 if (rtransport->conn_sched.next_admin_pg == NULL) { 3326 rtransport->conn_sched.next_admin_pg = rgroup; 3327 rtransport->conn_sched.next_io_pg = rgroup; 3328 } 3329 3330 pthread_mutex_unlock(&rtransport->lock); 3331 return &rgroup->group; 3332 } 3333 3334 static struct spdk_nvmf_transport_poll_group * 3335 spdk_nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 3336 { 3337 struct spdk_nvmf_rdma_transport *rtransport; 3338 struct spdk_nvmf_rdma_poll_group **pg; 3339 struct spdk_nvmf_transport_poll_group *result; 3340 3341 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport); 3342 3343 pthread_mutex_lock(&rtransport->lock); 3344 3345 if (TAILQ_EMPTY(&rtransport->poll_groups)) { 3346 pthread_mutex_unlock(&rtransport->lock); 3347 return NULL; 3348 } 3349 3350 if (qpair->qid == 0) { 3351 pg = &rtransport->conn_sched.next_admin_pg; 3352 } else { 3353 pg = &rtransport->conn_sched.next_io_pg; 3354 } 3355 3356 assert(*pg != NULL); 3357 3358 result = &(*pg)->group; 3359 3360 *pg = TAILQ_NEXT(*pg, link); 3361 if (*pg == NULL) { 3362 *pg = TAILQ_FIRST(&rtransport->poll_groups); 3363 } 3364 3365 pthread_mutex_unlock(&rtransport->lock); 3366 3367 return result; 3368 } 3369 3370 static void 3371 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 3372 { 3373 struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup; 3374 struct spdk_nvmf_rdma_poller *poller, *tmp; 3375 struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair; 3376 struct spdk_nvmf_transport_pg_cache_buf *buf, *tmp_buf; 3377 struct spdk_nvmf_rdma_transport *rtransport; 3378 3379 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3380 if (!rgroup) { 3381 return; 3382 } 3383 3384 /* free all retired buffers back to the transport so we don't short the mempool. */ 3385 STAILQ_FOREACH_SAFE(buf, &rgroup->retired_bufs, link, tmp_buf) { 3386 STAILQ_REMOVE(&rgroup->retired_bufs, buf, spdk_nvmf_transport_pg_cache_buf, link); 3387 assert(group->transport != NULL); 3388 spdk_mempool_put(group->transport->data_buf_pool, buf); 3389 } 3390 3391 TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) { 3392 TAILQ_REMOVE(&rgroup->pollers, poller, link); 3393 3394 TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) { 3395 spdk_nvmf_rdma_qpair_destroy(qpair); 3396 } 3397 3398 if (poller->srq) { 3399 nvmf_rdma_resources_destroy(poller->resources); 3400 ibv_destroy_srq(poller->srq); 3401 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq); 3402 } 3403 3404 if (poller->cq) { 3405 ibv_destroy_cq(poller->cq); 3406 } 3407 3408 free(poller); 3409 } 3410 3411 if (rgroup->group.transport == NULL) { 3412 /* Transport can be NULL when spdk_nvmf_rdma_poll_group_create() 3413 * calls this function directly in a failure path. */ 3414 free(rgroup); 3415 return; 3416 } 3417 3418 rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport); 3419 3420 pthread_mutex_lock(&rtransport->lock); 3421 next_rgroup = TAILQ_NEXT(rgroup, link); 3422 TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link); 3423 if (next_rgroup == NULL) { 3424 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups); 3425 } 3426 if (rtransport->conn_sched.next_admin_pg == rgroup) { 3427 rtransport->conn_sched.next_admin_pg = next_rgroup; 3428 } 3429 if (rtransport->conn_sched.next_io_pg == rgroup) { 3430 rtransport->conn_sched.next_io_pg = next_rgroup; 3431 } 3432 pthread_mutex_unlock(&rtransport->lock); 3433 3434 free(rgroup); 3435 } 3436 3437 static void 3438 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair) 3439 { 3440 if (rqpair->cm_id != NULL) { 3441 spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES); 3442 } 3443 spdk_nvmf_rdma_qpair_destroy(rqpair); 3444 } 3445 3446 static int 3447 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3448 struct spdk_nvmf_qpair *qpair) 3449 { 3450 struct spdk_nvmf_rdma_poll_group *rgroup; 3451 struct spdk_nvmf_rdma_qpair *rqpair; 3452 struct spdk_nvmf_rdma_device *device; 3453 struct spdk_nvmf_rdma_poller *poller; 3454 int rc; 3455 3456 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3457 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3458 3459 device = rqpair->port->device; 3460 3461 TAILQ_FOREACH(poller, &rgroup->pollers, link) { 3462 if (poller->device == device) { 3463 break; 3464 } 3465 } 3466 3467 if (!poller) { 3468 SPDK_ERRLOG("No poller found for device.\n"); 3469 return -1; 3470 } 3471 3472 TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link); 3473 rqpair->poller = poller; 3474 rqpair->srq = rqpair->poller->srq; 3475 3476 rc = spdk_nvmf_rdma_qpair_initialize(qpair); 3477 if (rc < 0) { 3478 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair); 3479 return -1; 3480 } 3481 3482 rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair); 3483 if (rc) { 3484 /* Try to reject, but we probably can't */ 3485 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3486 return -1; 3487 } 3488 3489 spdk_nvmf_rdma_update_ibv_state(rqpair); 3490 3491 return 0; 3492 } 3493 3494 static int 3495 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req) 3496 { 3497 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req); 3498 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3499 struct spdk_nvmf_rdma_transport, transport); 3500 3501 nvmf_rdma_request_free(rdma_req, rtransport); 3502 return 0; 3503 } 3504 3505 static int 3506 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req) 3507 { 3508 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport, 3509 struct spdk_nvmf_rdma_transport, transport); 3510 struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, 3511 struct spdk_nvmf_rdma_request, req); 3512 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, 3513 struct spdk_nvmf_rdma_qpair, qpair); 3514 3515 if (rqpair->ibv_state != IBV_QPS_ERR) { 3516 /* The connection is alive, so process the request as normal */ 3517 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 3518 } else { 3519 /* The connection is dead. Move the request directly to the completed state. */ 3520 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3521 } 3522 3523 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3524 3525 return 0; 3526 } 3527 3528 static int 3529 spdk_nvmf_rdma_destroy_defunct_qpair(void *ctx) 3530 { 3531 struct spdk_nvmf_rdma_qpair *rqpair = ctx; 3532 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport, 3533 struct spdk_nvmf_rdma_transport, transport); 3534 3535 SPDK_INFOLOG(SPDK_LOG_RDMA, "QP#%d hasn't been drained as expected, manually destroy it\n", 3536 rqpair->qpair.qid); 3537 3538 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true); 3539 spdk_nvmf_rdma_qpair_destroy(rqpair); 3540 3541 return 0; 3542 } 3543 3544 static void 3545 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair) 3546 { 3547 struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3548 3549 if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) { 3550 return; 3551 } 3552 3553 rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING; 3554 3555 /* This happens only when the qpair is disconnected before 3556 * it is added to the poll group. Since there is no poll group, 3557 * the RDMA qp has not been initialized yet and the RDMA CM 3558 * event has not yet been acknowledged, so we need to reject it. 3559 */ 3560 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) { 3561 spdk_nvmf_rdma_qpair_reject_connection(rqpair); 3562 return; 3563 } 3564 3565 if (rqpair->ibv_state != IBV_QPS_ERR) { 3566 spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR); 3567 } 3568 3569 rqpair->destruct_poller = spdk_poller_register(spdk_nvmf_rdma_destroy_defunct_qpair, (void *)rqpair, 3570 NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US); 3571 } 3572 3573 static struct spdk_nvmf_rdma_qpair * 3574 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc) 3575 { 3576 struct spdk_nvmf_rdma_qpair *rqpair; 3577 /* @todo: improve QP search */ 3578 TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) { 3579 if (wc->qp_num == rqpair->cm_id->qp->qp_num) { 3580 return rqpair; 3581 } 3582 } 3583 SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num); 3584 return NULL; 3585 } 3586 3587 #ifdef DEBUG 3588 static int 3589 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req) 3590 { 3591 return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST || 3592 rdma_req->state == RDMA_REQUEST_STATE_COMPLETING; 3593 } 3594 #endif 3595 3596 static void 3597 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr, 3598 int rc) 3599 { 3600 struct spdk_nvmf_rdma_recv *rdma_recv; 3601 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3602 3603 SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc); 3604 while (bad_recv_wr != NULL) { 3605 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id; 3606 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3607 3608 rdma_recv->qpair->current_recv_depth++; 3609 bad_recv_wr = bad_recv_wr->next; 3610 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc); 3611 spdk_nvmf_rdma_start_disconnect(rdma_recv->qpair); 3612 } 3613 } 3614 3615 static void 3616 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc) 3617 { 3618 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc); 3619 while (bad_recv_wr != NULL) { 3620 bad_recv_wr = bad_recv_wr->next; 3621 rqpair->current_recv_depth++; 3622 } 3623 spdk_nvmf_rdma_start_disconnect(rqpair); 3624 } 3625 3626 static void 3627 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport, 3628 struct spdk_nvmf_rdma_poller *rpoller) 3629 { 3630 struct spdk_nvmf_rdma_qpair *rqpair; 3631 struct ibv_recv_wr *bad_recv_wr; 3632 int rc; 3633 3634 if (rpoller->srq) { 3635 if (rpoller->resources->recvs_to_post.first != NULL) { 3636 rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr); 3637 if (rc) { 3638 _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc); 3639 } 3640 rpoller->resources->recvs_to_post.first = NULL; 3641 rpoller->resources->recvs_to_post.last = NULL; 3642 } 3643 } else { 3644 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) { 3645 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv); 3646 assert(rqpair->resources->recvs_to_post.first != NULL); 3647 rc = ibv_post_recv(rqpair->cm_id->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr); 3648 if (rc) { 3649 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc); 3650 } 3651 rqpair->resources->recvs_to_post.first = NULL; 3652 rqpair->resources->recvs_to_post.last = NULL; 3653 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link); 3654 } 3655 } 3656 } 3657 3658 static void 3659 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport, 3660 struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc) 3661 { 3662 struct spdk_nvmf_rdma_wr *bad_rdma_wr; 3663 struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL; 3664 3665 SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc); 3666 for (; bad_wr != NULL; bad_wr = bad_wr->next) { 3667 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id; 3668 assert(rqpair->current_send_depth > 0); 3669 rqpair->current_send_depth--; 3670 switch (bad_rdma_wr->type) { 3671 case RDMA_WR_TYPE_DATA: 3672 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3673 if (bad_wr->opcode == IBV_WR_RDMA_READ) { 3674 assert(rqpair->current_read_depth > 0); 3675 rqpair->current_read_depth--; 3676 } 3677 break; 3678 case RDMA_WR_TYPE_SEND: 3679 cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3680 break; 3681 default: 3682 SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair); 3683 prev_rdma_req = cur_rdma_req; 3684 continue; 3685 } 3686 3687 if (prev_rdma_req == cur_rdma_req) { 3688 /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */ 3689 /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */ 3690 continue; 3691 } 3692 3693 switch (cur_rdma_req->state) { 3694 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3695 cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR; 3696 cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE; 3697 break; 3698 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3699 case RDMA_REQUEST_STATE_COMPLETING: 3700 cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3701 break; 3702 default: 3703 SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n", 3704 cur_rdma_req->state, rqpair); 3705 continue; 3706 } 3707 3708 spdk_nvmf_rdma_request_process(rtransport, cur_rdma_req); 3709 prev_rdma_req = cur_rdma_req; 3710 } 3711 3712 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3713 /* Disconnect the connection. */ 3714 spdk_nvmf_rdma_start_disconnect(rqpair); 3715 } 3716 3717 } 3718 3719 static void 3720 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport, 3721 struct spdk_nvmf_rdma_poller *rpoller) 3722 { 3723 struct spdk_nvmf_rdma_qpair *rqpair; 3724 struct ibv_send_wr *bad_wr = NULL; 3725 int rc; 3726 3727 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) { 3728 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send); 3729 assert(rqpair->sends_to_post.first != NULL); 3730 rc = ibv_post_send(rqpair->cm_id->qp, rqpair->sends_to_post.first, &bad_wr); 3731 3732 /* bad wr always points to the first wr that failed. */ 3733 if (rc) { 3734 _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc); 3735 } 3736 rqpair->sends_to_post.first = NULL; 3737 rqpair->sends_to_post.last = NULL; 3738 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link); 3739 } 3740 } 3741 3742 static int 3743 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport, 3744 struct spdk_nvmf_rdma_poller *rpoller) 3745 { 3746 struct ibv_wc wc[32]; 3747 struct spdk_nvmf_rdma_wr *rdma_wr; 3748 struct spdk_nvmf_rdma_request *rdma_req; 3749 struct spdk_nvmf_rdma_recv *rdma_recv; 3750 struct spdk_nvmf_rdma_qpair *rqpair; 3751 int reaped, i; 3752 int count = 0; 3753 bool error = false; 3754 uint64_t poll_tsc = spdk_get_ticks(); 3755 3756 /* Poll for completing operations. */ 3757 reaped = ibv_poll_cq(rpoller->cq, 32, wc); 3758 if (reaped < 0) { 3759 SPDK_ERRLOG("Error polling CQ! (%d): %s\n", 3760 errno, spdk_strerror(errno)); 3761 return -1; 3762 } 3763 3764 rpoller->stat.polls++; 3765 rpoller->stat.completions += reaped; 3766 3767 for (i = 0; i < reaped; i++) { 3768 3769 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id; 3770 3771 switch (rdma_wr->type) { 3772 case RDMA_WR_TYPE_SEND: 3773 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr); 3774 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3775 3776 if (!wc[i].status) { 3777 count++; 3778 assert(wc[i].opcode == IBV_WC_SEND); 3779 assert(spdk_nvmf_rdma_req_is_completing(rdma_req)); 3780 } else { 3781 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3782 } 3783 3784 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3785 /* +1 for the response wr */ 3786 rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1; 3787 rdma_req->num_outstanding_data_wr = 0; 3788 3789 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3790 break; 3791 case RDMA_WR_TYPE_RECV: 3792 /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */ 3793 rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr); 3794 if (rpoller->srq != NULL) { 3795 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]); 3796 /* It is possible that there are still some completions for destroyed QP 3797 * associated with SRQ. We just ignore these late completions and re-post 3798 * receive WRs back to SRQ. 3799 */ 3800 if (spdk_unlikely(NULL == rdma_recv->qpair)) { 3801 struct ibv_recv_wr *bad_wr; 3802 int rc; 3803 3804 rdma_recv->wr.next = NULL; 3805 rc = ibv_post_srq_recv(rpoller->srq, 3806 &rdma_recv->wr, 3807 &bad_wr); 3808 if (rc) { 3809 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc); 3810 } 3811 continue; 3812 } 3813 } 3814 rqpair = rdma_recv->qpair; 3815 3816 assert(rqpair != NULL); 3817 if (!wc[i].status) { 3818 assert(wc[i].opcode == IBV_WC_RECV); 3819 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) { 3820 spdk_nvmf_rdma_start_disconnect(rqpair); 3821 break; 3822 } 3823 } 3824 3825 rdma_recv->wr.next = NULL; 3826 rqpair->current_recv_depth++; 3827 rdma_recv->receive_tsc = poll_tsc; 3828 rpoller->stat.requests++; 3829 STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link); 3830 break; 3831 case RDMA_WR_TYPE_DATA: 3832 rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr); 3833 rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair); 3834 3835 assert(rdma_req->num_outstanding_data_wr > 0); 3836 3837 rqpair->current_send_depth--; 3838 rdma_req->num_outstanding_data_wr--; 3839 if (!wc[i].status) { 3840 assert(wc[i].opcode == IBV_WC_RDMA_READ); 3841 rqpair->current_read_depth--; 3842 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */ 3843 if (rdma_req->num_outstanding_data_wr == 0) { 3844 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 3845 spdk_nvmf_rdma_request_process(rtransport, rdma_req); 3846 } 3847 } else { 3848 /* If the data transfer fails still force the queue into the error state, 3849 * if we were performing an RDMA_READ, we need to force the request into a 3850 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE 3851 * case, we should wait for the SEND to complete. */ 3852 SPDK_ERRLOG("data=%p length=%u\n", rdma_req->req.data, rdma_req->req.length); 3853 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) { 3854 rqpair->current_read_depth--; 3855 if (rdma_req->num_outstanding_data_wr == 0) { 3856 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 3857 } 3858 } 3859 } 3860 break; 3861 default: 3862 SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode); 3863 continue; 3864 } 3865 3866 /* Handle error conditions */ 3867 if (wc[i].status) { 3868 SPDK_DEBUGLOG(SPDK_LOG_RDMA, "CQ error on CQ %p, Request 0x%lu (%d): %s\n", 3869 rpoller->cq, wc[i].wr_id, wc[i].status, ibv_wc_status_str(wc[i].status)); 3870 3871 error = true; 3872 3873 if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) { 3874 /* Disconnect the connection. */ 3875 spdk_nvmf_rdma_start_disconnect(rqpair); 3876 } else { 3877 nvmf_rdma_destroy_drained_qpair(rqpair); 3878 } 3879 continue; 3880 } 3881 3882 spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false); 3883 3884 if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 3885 nvmf_rdma_destroy_drained_qpair(rqpair); 3886 } 3887 } 3888 3889 if (error == true) { 3890 return -1; 3891 } 3892 3893 /* submit outstanding work requests. */ 3894 _poller_submit_recvs(rtransport, rpoller); 3895 _poller_submit_sends(rtransport, rpoller); 3896 3897 return count; 3898 } 3899 3900 static int 3901 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3902 { 3903 struct spdk_nvmf_rdma_transport *rtransport; 3904 struct spdk_nvmf_rdma_poll_group *rgroup; 3905 struct spdk_nvmf_rdma_poller *rpoller; 3906 int count, rc; 3907 3908 rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport); 3909 rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group); 3910 3911 count = 0; 3912 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 3913 rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller); 3914 if (rc < 0) { 3915 return rc; 3916 } 3917 count += rc; 3918 } 3919 3920 return count; 3921 } 3922 3923 static int 3924 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id, 3925 struct spdk_nvme_transport_id *trid, 3926 bool peer) 3927 { 3928 struct sockaddr *saddr; 3929 uint16_t port; 3930 3931 trid->trtype = SPDK_NVME_TRANSPORT_RDMA; 3932 3933 if (peer) { 3934 saddr = rdma_get_peer_addr(id); 3935 } else { 3936 saddr = rdma_get_local_addr(id); 3937 } 3938 switch (saddr->sa_family) { 3939 case AF_INET: { 3940 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr; 3941 3942 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3943 inet_ntop(AF_INET, &saddr_in->sin_addr, 3944 trid->traddr, sizeof(trid->traddr)); 3945 if (peer) { 3946 port = ntohs(rdma_get_dst_port(id)); 3947 } else { 3948 port = ntohs(rdma_get_src_port(id)); 3949 } 3950 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3951 break; 3952 } 3953 case AF_INET6: { 3954 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr; 3955 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3956 inet_ntop(AF_INET6, &saddr_in->sin6_addr, 3957 trid->traddr, sizeof(trid->traddr)); 3958 if (peer) { 3959 port = ntohs(rdma_get_dst_port(id)); 3960 } else { 3961 port = ntohs(rdma_get_src_port(id)); 3962 } 3963 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port); 3964 break; 3965 } 3966 default: 3967 return -1; 3968 3969 } 3970 3971 return 0; 3972 } 3973 3974 static int 3975 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3976 struct spdk_nvme_transport_id *trid) 3977 { 3978 struct spdk_nvmf_rdma_qpair *rqpair; 3979 3980 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3981 3982 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true); 3983 } 3984 3985 static int 3986 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3987 struct spdk_nvme_transport_id *trid) 3988 { 3989 struct spdk_nvmf_rdma_qpair *rqpair; 3990 3991 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 3992 3993 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false); 3994 } 3995 3996 static int 3997 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3998 struct spdk_nvme_transport_id *trid) 3999 { 4000 struct spdk_nvmf_rdma_qpair *rqpair; 4001 4002 rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair); 4003 4004 return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false); 4005 } 4006 4007 void 4008 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks) 4009 { 4010 g_nvmf_hooks = *hooks; 4011 } 4012 4013 static int 4014 spdk_nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt, 4015 struct spdk_nvmf_transport_poll_group_stat **stat) 4016 { 4017 struct spdk_io_channel *ch; 4018 struct spdk_nvmf_poll_group *group; 4019 struct spdk_nvmf_transport_poll_group *tgroup; 4020 struct spdk_nvmf_rdma_poll_group *rgroup; 4021 struct spdk_nvmf_rdma_poller *rpoller; 4022 struct spdk_nvmf_rdma_device_stat *device_stat; 4023 uint64_t num_devices = 0; 4024 4025 if (tgt == NULL || stat == NULL) { 4026 return -EINVAL; 4027 } 4028 4029 ch = spdk_get_io_channel(tgt); 4030 group = spdk_io_channel_get_ctx(ch);; 4031 spdk_put_io_channel(ch); 4032 TAILQ_FOREACH(tgroup, &group->tgroups, link) { 4033 if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) { 4034 *stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat)); 4035 if (!*stat) { 4036 SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n"); 4037 return -ENOMEM; 4038 } 4039 (*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA; 4040 4041 rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group); 4042 /* Count devices to allocate enough memory */ 4043 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4044 ++num_devices; 4045 } 4046 (*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat)); 4047 if (!(*stat)->rdma.devices) { 4048 SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n"); 4049 free(*stat); 4050 return -ENOMEM; 4051 } 4052 4053 (*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer; 4054 (*stat)->rdma.num_devices = num_devices; 4055 num_devices = 0; 4056 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) { 4057 device_stat = &(*stat)->rdma.devices[num_devices++]; 4058 device_stat->name = ibv_get_device_name(rpoller->device->context->device); 4059 device_stat->polls = rpoller->stat.polls; 4060 device_stat->completions = rpoller->stat.completions; 4061 device_stat->requests = rpoller->stat.requests; 4062 device_stat->request_latency = rpoller->stat.request_latency; 4063 device_stat->pending_free_request = rpoller->stat.pending_free_request; 4064 device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read; 4065 device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write; 4066 } 4067 return 0; 4068 } 4069 } 4070 return -ENOENT; 4071 } 4072 4073 static void 4074 spdk_nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat) 4075 { 4076 if (stat) { 4077 free(stat->rdma.devices); 4078 } 4079 free(stat); 4080 } 4081 4082 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = { 4083 .type = SPDK_NVME_TRANSPORT_RDMA, 4084 .opts_init = spdk_nvmf_rdma_opts_init, 4085 .create = spdk_nvmf_rdma_create, 4086 .destroy = spdk_nvmf_rdma_destroy, 4087 4088 .listen = spdk_nvmf_rdma_listen, 4089 .stop_listen = spdk_nvmf_rdma_stop_listen, 4090 .accept = spdk_nvmf_rdma_accept, 4091 4092 .listener_discover = spdk_nvmf_rdma_discover, 4093 4094 .poll_group_create = spdk_nvmf_rdma_poll_group_create, 4095 .get_optimal_poll_group = spdk_nvmf_rdma_get_optimal_poll_group, 4096 .poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy, 4097 .poll_group_add = spdk_nvmf_rdma_poll_group_add, 4098 .poll_group_poll = spdk_nvmf_rdma_poll_group_poll, 4099 4100 .req_free = spdk_nvmf_rdma_request_free, 4101 .req_complete = spdk_nvmf_rdma_request_complete, 4102 4103 .qpair_fini = spdk_nvmf_rdma_close_qpair, 4104 .qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid, 4105 .qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid, 4106 .qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid, 4107 4108 .poll_group_get_stat = spdk_nvmf_rdma_poll_group_get_stat, 4109 .poll_group_free_stat = spdk_nvmf_rdma_poll_group_free_stat, 4110 }; 4111 4112 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA) 4113