1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (c) Intel Corporation. All rights reserved. 3 * Copyright (c) 2019, 2020 Mellanox Technologies LTD. All rights reserved. 4 * Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved. 5 */ 6 7 #include "spdk/accel.h" 8 #include "spdk/stdinc.h" 9 #include "spdk/crc32.h" 10 #include "spdk/endian.h" 11 #include "spdk/assert.h" 12 #include "spdk/thread.h" 13 #include "spdk/nvmf_transport.h" 14 #include "spdk/string.h" 15 #include "spdk/trace.h" 16 #include "spdk/util.h" 17 #include "spdk/log.h" 18 19 #include "spdk_internal/assert.h" 20 #include "spdk_internal/nvme_tcp.h" 21 #include "spdk_internal/sock.h" 22 23 #include "nvmf_internal.h" 24 25 #include "spdk_internal/trace_defs.h" 26 27 #define NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME 16 28 #define SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY 16 29 #define SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY 0 30 #define SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM 32 31 #define SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION true 32 33 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp; 34 35 /* spdk nvmf related structure */ 36 enum spdk_nvmf_tcp_req_state { 37 38 /* The request is not currently in use */ 39 TCP_REQUEST_STATE_FREE = 0, 40 41 /* Initial state when request first received */ 42 TCP_REQUEST_STATE_NEW = 1, 43 44 /* The request is queued until a data buffer is available. */ 45 TCP_REQUEST_STATE_NEED_BUFFER = 2, 46 47 /* The request is waiting for zcopy_start to finish */ 48 TCP_REQUEST_STATE_AWAITING_ZCOPY_START = 3, 49 50 /* The request has received a zero-copy buffer */ 51 TCP_REQUEST_STATE_ZCOPY_START_COMPLETED = 4, 52 53 /* The request is currently transferring data from the host to the controller. */ 54 TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER = 5, 55 56 /* The request is waiting for the R2T send acknowledgement. */ 57 TCP_REQUEST_STATE_AWAITING_R2T_ACK = 6, 58 59 /* The request is ready to execute at the block device */ 60 TCP_REQUEST_STATE_READY_TO_EXECUTE = 7, 61 62 /* The request is currently executing at the block device */ 63 TCP_REQUEST_STATE_EXECUTING = 8, 64 65 /* The request is waiting for zcopy buffers to be commited */ 66 TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT = 9, 67 68 /* The request finished executing at the block device */ 69 TCP_REQUEST_STATE_EXECUTED = 10, 70 71 /* The request is ready to send a completion */ 72 TCP_REQUEST_STATE_READY_TO_COMPLETE = 11, 73 74 /* The request is currently transferring final pdus from the controller to the host. */ 75 TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST = 12, 76 77 /* The request is waiting for zcopy buffers to be released (without committing) */ 78 TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE = 13, 79 80 /* The request completed and can be marked free. */ 81 TCP_REQUEST_STATE_COMPLETED = 14, 82 83 /* Terminator */ 84 TCP_REQUEST_NUM_STATES, 85 }; 86 87 static const char *spdk_nvmf_tcp_term_req_fes_str[] = { 88 "Invalid PDU Header Field", 89 "PDU Sequence Error", 90 "Header Digiest Error", 91 "Data Transfer Out of Range", 92 "R2T Limit Exceeded", 93 "Unsupported parameter", 94 }; 95 96 SPDK_TRACE_REGISTER_FN(nvmf_tcp_trace, "nvmf_tcp", TRACE_GROUP_NVMF_TCP) 97 { 98 spdk_trace_register_owner(OWNER_NVMF_TCP, 't'); 99 spdk_trace_register_object(OBJECT_NVMF_TCP_IO, 'r'); 100 spdk_trace_register_description("TCP_REQ_NEW", 101 TRACE_TCP_REQUEST_STATE_NEW, 102 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 1, 103 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 104 spdk_trace_register_description("TCP_REQ_NEED_BUFFER", 105 TRACE_TCP_REQUEST_STATE_NEED_BUFFER, 106 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 107 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 108 spdk_trace_register_description("TCP_REQ_WAIT_ZCPY_START", 109 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, 110 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 111 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 112 spdk_trace_register_description("TCP_REQ_ZCPY_START_CPL", 113 TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, 114 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 115 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 116 spdk_trace_register_description("TCP_REQ_TX_H_TO_C", 117 TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 118 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 119 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 120 spdk_trace_register_description("TCP_REQ_RDY_TO_EXECUTE", 121 TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, 122 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 123 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 124 spdk_trace_register_description("TCP_REQ_EXECUTING", 125 TRACE_TCP_REQUEST_STATE_EXECUTING, 126 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 127 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 128 spdk_trace_register_description("TCP_REQ_WAIT_ZCPY_CMT", 129 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, 130 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 131 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 132 spdk_trace_register_description("TCP_REQ_EXECUTED", 133 TRACE_TCP_REQUEST_STATE_EXECUTED, 134 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 135 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 136 spdk_trace_register_description("TCP_REQ_RDY_TO_COMPLETE", 137 TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, 138 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 139 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 140 spdk_trace_register_description("TCP_REQ_TRANSFER_C2H", 141 TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 142 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 143 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 144 spdk_trace_register_description("TCP_REQ_AWAIT_ZCPY_RLS", 145 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, 146 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 147 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 148 spdk_trace_register_description("TCP_REQ_COMPLETED", 149 TRACE_TCP_REQUEST_STATE_COMPLETED, 150 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 151 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 152 spdk_trace_register_description("TCP_WRITE_START", 153 TRACE_TCP_FLUSH_WRITEBUF_START, 154 OWNER_NVMF_TCP, OBJECT_NONE, 0, 155 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 156 spdk_trace_register_description("TCP_WRITE_DONE", 157 TRACE_TCP_FLUSH_WRITEBUF_DONE, 158 OWNER_NVMF_TCP, OBJECT_NONE, 0, 159 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 160 spdk_trace_register_description("TCP_READ_DONE", 161 TRACE_TCP_READ_FROM_SOCKET_DONE, 162 OWNER_NVMF_TCP, OBJECT_NONE, 0, 163 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 164 spdk_trace_register_description("TCP_REQ_AWAIT_R2T_ACK", 165 TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, 166 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 167 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 168 169 spdk_trace_register_description("TCP_QP_CREATE", TRACE_TCP_QP_CREATE, 170 OWNER_NVMF_TCP, OBJECT_NONE, 0, 171 SPDK_TRACE_ARG_TYPE_INT, ""); 172 spdk_trace_register_description("TCP_QP_SOCK_INIT", TRACE_TCP_QP_SOCK_INIT, 173 OWNER_NVMF_TCP, OBJECT_NONE, 0, 174 SPDK_TRACE_ARG_TYPE_INT, ""); 175 spdk_trace_register_description("TCP_QP_STATE_CHANGE", TRACE_TCP_QP_STATE_CHANGE, 176 OWNER_NVMF_TCP, OBJECT_NONE, 0, 177 SPDK_TRACE_ARG_TYPE_INT, "state"); 178 spdk_trace_register_description("TCP_QP_DISCONNECT", TRACE_TCP_QP_DISCONNECT, 179 OWNER_NVMF_TCP, OBJECT_NONE, 0, 180 SPDK_TRACE_ARG_TYPE_INT, ""); 181 spdk_trace_register_description("TCP_QP_DESTROY", TRACE_TCP_QP_DESTROY, 182 OWNER_NVMF_TCP, OBJECT_NONE, 0, 183 SPDK_TRACE_ARG_TYPE_INT, ""); 184 spdk_trace_register_description("TCP_QP_ABORT_REQ", TRACE_TCP_QP_ABORT_REQ, 185 OWNER_NVMF_TCP, OBJECT_NONE, 0, 186 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 187 spdk_trace_register_description("TCP_QP_RCV_STATE_CHANGE", TRACE_TCP_QP_RCV_STATE_CHANGE, 188 OWNER_NVMF_TCP, OBJECT_NONE, 0, 189 SPDK_TRACE_ARG_TYPE_INT, "state"); 190 191 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_TCP_IO, 1); 192 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_TCP_IO, 0); 193 } 194 195 struct spdk_nvmf_tcp_req { 196 struct spdk_nvmf_request req; 197 struct spdk_nvme_cpl rsp; 198 struct spdk_nvme_cmd cmd; 199 200 /* A PDU that can be used for sending responses. This is 201 * not the incoming PDU! */ 202 struct nvme_tcp_pdu *pdu; 203 204 /* In-capsule data buffer */ 205 uint8_t *buf; 206 207 struct spdk_nvmf_tcp_req *fused_pair; 208 209 /* 210 * The PDU for a request may be used multiple times in serial over 211 * the request's lifetime. For example, first to send an R2T, then 212 * to send a completion. To catch mistakes where the PDU is used 213 * twice at the same time, add a debug flag here for init/fini. 214 */ 215 bool pdu_in_use; 216 bool has_in_capsule_data; 217 bool fused_failed; 218 219 /* transfer_tag */ 220 uint16_t ttag; 221 222 enum spdk_nvmf_tcp_req_state state; 223 224 /* 225 * h2c_offset is used when we receive the h2c_data PDU. 226 */ 227 uint32_t h2c_offset; 228 229 STAILQ_ENTRY(spdk_nvmf_tcp_req) link; 230 TAILQ_ENTRY(spdk_nvmf_tcp_req) state_link; 231 }; 232 233 struct spdk_nvmf_tcp_qpair { 234 struct spdk_nvmf_qpair qpair; 235 struct spdk_nvmf_tcp_poll_group *group; 236 struct spdk_sock *sock; 237 238 enum nvme_tcp_pdu_recv_state recv_state; 239 enum nvme_tcp_qpair_state state; 240 241 /* PDU being actively received */ 242 struct nvme_tcp_pdu *pdu_in_progress; 243 244 struct spdk_nvmf_tcp_req *fused_first; 245 246 /* Queues to track the requests in all states */ 247 TAILQ_HEAD(, spdk_nvmf_tcp_req) tcp_req_working_queue; 248 TAILQ_HEAD(, spdk_nvmf_tcp_req) tcp_req_free_queue; 249 SLIST_HEAD(, nvme_tcp_pdu) tcp_pdu_free_queue; 250 251 /* Number of requests in each state */ 252 uint32_t state_cntr[TCP_REQUEST_NUM_STATES]; 253 254 uint8_t cpda; 255 256 bool host_hdgst_enable; 257 bool host_ddgst_enable; 258 259 /* This is a spare PDU used for sending special management 260 * operations. Primarily, this is used for the initial 261 * connection response and c2h termination request. */ 262 struct nvme_tcp_pdu *mgmt_pdu; 263 264 /* Arrays of in-capsule buffers, requests, and pdus. 265 * Each array is 'resource_count' number of elements */ 266 void *bufs; 267 struct spdk_nvmf_tcp_req *reqs; 268 struct nvme_tcp_pdu *pdus; 269 uint32_t resource_count; 270 uint32_t recv_buf_size; 271 272 struct spdk_nvmf_tcp_port *port; 273 274 /* IP address */ 275 char initiator_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 276 char target_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 277 278 /* IP port */ 279 uint16_t initiator_port; 280 uint16_t target_port; 281 282 /* Timer used to destroy qpair after detecting transport error issue if initiator does 283 * not close the connection. 284 */ 285 struct spdk_poller *timeout_poller; 286 287 spdk_nvmf_transport_qpair_fini_cb fini_cb_fn; 288 void *fini_cb_arg; 289 290 TAILQ_ENTRY(spdk_nvmf_tcp_qpair) link; 291 }; 292 293 struct spdk_nvmf_tcp_control_msg { 294 STAILQ_ENTRY(spdk_nvmf_tcp_control_msg) link; 295 }; 296 297 struct spdk_nvmf_tcp_control_msg_list { 298 void *msg_buf; 299 STAILQ_HEAD(, spdk_nvmf_tcp_control_msg) free_msgs; 300 }; 301 302 struct spdk_nvmf_tcp_poll_group { 303 struct spdk_nvmf_transport_poll_group group; 304 struct spdk_sock_group *sock_group; 305 306 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) qpairs; 307 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) await_req; 308 309 struct spdk_io_channel *accel_channel; 310 struct spdk_nvmf_tcp_control_msg_list *control_msg_list; 311 312 TAILQ_ENTRY(spdk_nvmf_tcp_poll_group) link; 313 }; 314 315 struct spdk_nvmf_tcp_port { 316 const struct spdk_nvme_transport_id *trid; 317 struct spdk_sock *listen_sock; 318 TAILQ_ENTRY(spdk_nvmf_tcp_port) link; 319 }; 320 321 struct tcp_transport_opts { 322 bool c2h_success; 323 uint16_t control_msg_num; 324 uint32_t sock_priority; 325 }; 326 327 struct spdk_nvmf_tcp_transport { 328 struct spdk_nvmf_transport transport; 329 struct tcp_transport_opts tcp_opts; 330 331 struct spdk_nvmf_tcp_poll_group *next_pg; 332 333 struct spdk_poller *accept_poller; 334 335 TAILQ_HEAD(, spdk_nvmf_tcp_port) ports; 336 TAILQ_HEAD(, spdk_nvmf_tcp_poll_group) poll_groups; 337 }; 338 339 static const struct spdk_json_object_decoder tcp_transport_opts_decoder[] = { 340 { 341 "c2h_success", offsetof(struct tcp_transport_opts, c2h_success), 342 spdk_json_decode_bool, true 343 }, 344 { 345 "control_msg_num", offsetof(struct tcp_transport_opts, control_msg_num), 346 spdk_json_decode_uint16, true 347 }, 348 { 349 "sock_priority", offsetof(struct tcp_transport_opts, sock_priority), 350 spdk_json_decode_uint32, true 351 }, 352 }; 353 354 static bool nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 355 struct spdk_nvmf_tcp_req *tcp_req); 356 static void nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 357 358 static void _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 359 struct spdk_nvmf_tcp_req *tcp_req); 360 361 static void 362 nvmf_tcp_req_set_state(struct spdk_nvmf_tcp_req *tcp_req, 363 enum spdk_nvmf_tcp_req_state state) 364 { 365 struct spdk_nvmf_qpair *qpair; 366 struct spdk_nvmf_tcp_qpair *tqpair; 367 368 qpair = tcp_req->req.qpair; 369 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 370 371 assert(tqpair->state_cntr[tcp_req->state] > 0); 372 tqpair->state_cntr[tcp_req->state]--; 373 tqpair->state_cntr[state]++; 374 375 tcp_req->state = state; 376 } 377 378 static inline struct nvme_tcp_pdu * 379 nvmf_tcp_req_pdu_init(struct spdk_nvmf_tcp_req *tcp_req) 380 { 381 assert(tcp_req->pdu_in_use == false); 382 383 memset(tcp_req->pdu, 0, sizeof(*tcp_req->pdu)); 384 tcp_req->pdu->qpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 385 386 return tcp_req->pdu; 387 } 388 389 static struct spdk_nvmf_tcp_req * 390 nvmf_tcp_req_get(struct spdk_nvmf_tcp_qpair *tqpair) 391 { 392 struct spdk_nvmf_tcp_req *tcp_req; 393 394 tcp_req = TAILQ_FIRST(&tqpair->tcp_req_free_queue); 395 if (spdk_unlikely(!tcp_req)) { 396 return NULL; 397 } 398 399 memset(&tcp_req->rsp, 0, sizeof(tcp_req->rsp)); 400 tcp_req->h2c_offset = 0; 401 tcp_req->has_in_capsule_data = false; 402 tcp_req->req.dif_enabled = false; 403 tcp_req->req.zcopy_phase = NVMF_ZCOPY_PHASE_NONE; 404 405 TAILQ_REMOVE(&tqpair->tcp_req_free_queue, tcp_req, state_link); 406 TAILQ_INSERT_TAIL(&tqpair->tcp_req_working_queue, tcp_req, state_link); 407 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEW); 408 return tcp_req; 409 } 410 411 static inline void 412 nvmf_tcp_req_put(struct spdk_nvmf_tcp_qpair *tqpair, struct spdk_nvmf_tcp_req *tcp_req) 413 { 414 assert(!tcp_req->pdu_in_use); 415 416 TAILQ_REMOVE(&tqpair->tcp_req_working_queue, tcp_req, state_link); 417 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 418 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_FREE); 419 } 420 421 static void 422 nvmf_tcp_request_free(void *cb_arg) 423 { 424 struct spdk_nvmf_tcp_transport *ttransport; 425 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 426 427 assert(tcp_req != NULL); 428 429 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req=%p will be freed\n", tcp_req); 430 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 431 struct spdk_nvmf_tcp_transport, transport); 432 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 433 nvmf_tcp_req_process(ttransport, tcp_req); 434 } 435 436 static int 437 nvmf_tcp_req_free(struct spdk_nvmf_request *req) 438 { 439 struct spdk_nvmf_tcp_req *tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 440 441 nvmf_tcp_request_free(tcp_req); 442 443 return 0; 444 } 445 446 static void 447 nvmf_tcp_drain_state_queue(struct spdk_nvmf_tcp_qpair *tqpair, 448 enum spdk_nvmf_tcp_req_state state) 449 { 450 struct spdk_nvmf_tcp_req *tcp_req, *req_tmp; 451 452 assert(state != TCP_REQUEST_STATE_FREE); 453 TAILQ_FOREACH_SAFE(tcp_req, &tqpair->tcp_req_working_queue, state_link, req_tmp) { 454 if (state == tcp_req->state) { 455 nvmf_tcp_request_free(tcp_req); 456 } 457 } 458 } 459 460 static void 461 nvmf_tcp_cleanup_all_states(struct spdk_nvmf_tcp_qpair *tqpair) 462 { 463 struct spdk_nvmf_tcp_req *tcp_req, *req_tmp; 464 465 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 466 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_NEW); 467 468 /* Wipe the requests waiting for buffer from the global list */ 469 TAILQ_FOREACH_SAFE(tcp_req, &tqpair->tcp_req_working_queue, state_link, req_tmp) { 470 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 471 STAILQ_REMOVE(&tqpair->group->group.pending_buf_queue, &tcp_req->req, 472 spdk_nvmf_request, buf_link); 473 } 474 } 475 476 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_NEED_BUFFER); 477 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_EXECUTING); 478 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 479 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 480 } 481 482 static void 483 nvmf_tcp_dump_qpair_req_contents(struct spdk_nvmf_tcp_qpair *tqpair) 484 { 485 int i; 486 struct spdk_nvmf_tcp_req *tcp_req; 487 488 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", tqpair->qpair.qid); 489 for (i = 1; i < TCP_REQUEST_NUM_STATES; i++) { 490 SPDK_ERRLOG("\tNum of requests in state[%d] = %u\n", i, tqpair->state_cntr[i]); 491 TAILQ_FOREACH(tcp_req, &tqpair->tcp_req_working_queue, state_link) { 492 if ((int)tcp_req->state == i) { 493 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", tcp_req->req.data_from_pool); 494 SPDK_ERRLOG("\t\tRequest opcode: %d\n", tcp_req->req.cmd->nvmf_cmd.opcode); 495 } 496 } 497 } 498 } 499 500 static void 501 _nvmf_tcp_qpair_destroy(void *_tqpair) 502 { 503 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 504 spdk_nvmf_transport_qpair_fini_cb cb_fn = tqpair->fini_cb_fn; 505 void *cb_arg = tqpair->fini_cb_arg; 506 int err = 0; 507 508 spdk_trace_record(TRACE_TCP_QP_DESTROY, 0, 0, (uintptr_t)tqpair); 509 510 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 511 512 err = spdk_sock_close(&tqpair->sock); 513 assert(err == 0); 514 nvmf_tcp_cleanup_all_states(tqpair); 515 516 if (tqpair->state_cntr[TCP_REQUEST_STATE_FREE] != tqpair->resource_count) { 517 SPDK_ERRLOG("tqpair(%p) free tcp request num is %u but should be %u\n", tqpair, 518 tqpair->state_cntr[TCP_REQUEST_STATE_FREE], 519 tqpair->resource_count); 520 err++; 521 } 522 523 if (err > 0) { 524 nvmf_tcp_dump_qpair_req_contents(tqpair); 525 } 526 527 /* The timeout poller might still be registered here if we close the qpair before host 528 * terminates the connection. 529 */ 530 spdk_poller_unregister(&tqpair->timeout_poller); 531 spdk_dma_free(tqpair->pdus); 532 free(tqpair->reqs); 533 spdk_free(tqpair->bufs); 534 free(tqpair); 535 536 if (cb_fn != NULL) { 537 cb_fn(cb_arg); 538 } 539 540 SPDK_DEBUGLOG(nvmf_tcp, "Leave\n"); 541 } 542 543 static void 544 nvmf_tcp_qpair_destroy(struct spdk_nvmf_tcp_qpair *tqpair) 545 { 546 /* Delay the destruction to make sure it isn't performed from the context of a sock 547 * callback. Otherwise, spdk_sock_close() might not abort pending requests, causing their 548 * completions to be executed after the qpair is freed. (Note: this fixed issue #2471.) 549 */ 550 spdk_thread_send_msg(spdk_get_thread(), _nvmf_tcp_qpair_destroy, tqpair); 551 } 552 553 static void 554 nvmf_tcp_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 555 { 556 struct spdk_nvmf_tcp_transport *ttransport; 557 assert(w != NULL); 558 559 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 560 spdk_json_write_named_bool(w, "c2h_success", ttransport->tcp_opts.c2h_success); 561 spdk_json_write_named_uint32(w, "sock_priority", ttransport->tcp_opts.sock_priority); 562 } 563 564 static int 565 nvmf_tcp_destroy(struct spdk_nvmf_transport *transport, 566 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 567 { 568 struct spdk_nvmf_tcp_transport *ttransport; 569 570 assert(transport != NULL); 571 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 572 573 spdk_poller_unregister(&ttransport->accept_poller); 574 free(ttransport); 575 576 if (cb_fn) { 577 cb_fn(cb_arg); 578 } 579 return 0; 580 } 581 582 static int nvmf_tcp_accept(void *ctx); 583 584 static struct spdk_nvmf_transport * 585 nvmf_tcp_create(struct spdk_nvmf_transport_opts *opts) 586 { 587 struct spdk_nvmf_tcp_transport *ttransport; 588 uint32_t sge_count; 589 uint32_t min_shared_buffers; 590 591 ttransport = calloc(1, sizeof(*ttransport)); 592 if (!ttransport) { 593 return NULL; 594 } 595 596 TAILQ_INIT(&ttransport->ports); 597 TAILQ_INIT(&ttransport->poll_groups); 598 599 ttransport->transport.ops = &spdk_nvmf_transport_tcp; 600 601 ttransport->tcp_opts.c2h_success = SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION; 602 ttransport->tcp_opts.sock_priority = SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY; 603 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 604 if (opts->transport_specific != NULL && 605 spdk_json_decode_object_relaxed(opts->transport_specific, tcp_transport_opts_decoder, 606 SPDK_COUNTOF(tcp_transport_opts_decoder), 607 &ttransport->tcp_opts)) { 608 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 609 free(ttransport); 610 return NULL; 611 } 612 613 SPDK_NOTICELOG("*** TCP Transport Init ***\n"); 614 615 SPDK_INFOLOG(nvmf_tcp, "*** TCP Transport Init ***\n" 616 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 617 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 618 " in_capsule_data_size=%d, max_aq_depth=%d\n" 619 " num_shared_buffers=%d, c2h_success=%d,\n" 620 " dif_insert_or_strip=%d, sock_priority=%d\n" 621 " abort_timeout_sec=%d, control_msg_num=%hu\n", 622 opts->max_queue_depth, 623 opts->max_io_size, 624 opts->max_qpairs_per_ctrlr - 1, 625 opts->io_unit_size, 626 opts->in_capsule_data_size, 627 opts->max_aq_depth, 628 opts->num_shared_buffers, 629 ttransport->tcp_opts.c2h_success, 630 opts->dif_insert_or_strip, 631 ttransport->tcp_opts.sock_priority, 632 opts->abort_timeout_sec, 633 ttransport->tcp_opts.control_msg_num); 634 635 if (ttransport->tcp_opts.sock_priority > SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY) { 636 SPDK_ERRLOG("Unsupported socket_priority=%d, the current range is: 0 to %d\n" 637 "you can use man 7 socket to view the range of priority under SO_PRIORITY item\n", 638 ttransport->tcp_opts.sock_priority, SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY); 639 free(ttransport); 640 return NULL; 641 } 642 643 if (ttransport->tcp_opts.control_msg_num == 0 && 644 opts->in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 645 SPDK_WARNLOG("TCP param control_msg_num can't be 0 if ICD is less than %u bytes. Using default value %u\n", 646 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM); 647 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 648 } 649 650 /* I/O unit size cannot be larger than max I/O size */ 651 if (opts->io_unit_size > opts->max_io_size) { 652 opts->io_unit_size = opts->max_io_size; 653 } 654 655 sge_count = opts->max_io_size / opts->io_unit_size; 656 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 657 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 658 free(ttransport); 659 return NULL; 660 } 661 662 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 663 if (min_shared_buffers > opts->num_shared_buffers) { 664 SPDK_ERRLOG("There are not enough buffers to satisfy " 665 "per-poll group caches for each thread. (%" PRIu32 ") " 666 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 667 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 668 free(ttransport); 669 return NULL; 670 } 671 672 ttransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_tcp_accept, &ttransport->transport, 673 opts->acceptor_poll_rate); 674 if (!ttransport->accept_poller) { 675 free(ttransport); 676 return NULL; 677 } 678 679 return &ttransport->transport; 680 } 681 682 static int 683 nvmf_tcp_trsvcid_to_int(const char *trsvcid) 684 { 685 unsigned long long ull; 686 char *end = NULL; 687 688 ull = strtoull(trsvcid, &end, 10); 689 if (end == NULL || end == trsvcid || *end != '\0') { 690 return -1; 691 } 692 693 /* Valid TCP/IP port numbers are in [0, 65535] */ 694 if (ull > 65535) { 695 return -1; 696 } 697 698 return (int)ull; 699 } 700 701 /** 702 * Canonicalize a listen address trid. 703 */ 704 static int 705 nvmf_tcp_canon_listen_trid(struct spdk_nvme_transport_id *canon_trid, 706 const struct spdk_nvme_transport_id *trid) 707 { 708 int trsvcid_int; 709 710 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 711 if (trsvcid_int < 0) { 712 return -EINVAL; 713 } 714 715 memset(canon_trid, 0, sizeof(*canon_trid)); 716 spdk_nvme_trid_populate_transport(canon_trid, SPDK_NVME_TRANSPORT_TCP); 717 canon_trid->adrfam = trid->adrfam; 718 snprintf(canon_trid->traddr, sizeof(canon_trid->traddr), "%s", trid->traddr); 719 snprintf(canon_trid->trsvcid, sizeof(canon_trid->trsvcid), "%d", trsvcid_int); 720 721 return 0; 722 } 723 724 /** 725 * Find an existing listening port. 726 */ 727 static struct spdk_nvmf_tcp_port * 728 nvmf_tcp_find_port(struct spdk_nvmf_tcp_transport *ttransport, 729 const struct spdk_nvme_transport_id *trid) 730 { 731 struct spdk_nvme_transport_id canon_trid; 732 struct spdk_nvmf_tcp_port *port; 733 734 if (nvmf_tcp_canon_listen_trid(&canon_trid, trid) != 0) { 735 return NULL; 736 } 737 738 TAILQ_FOREACH(port, &ttransport->ports, link) { 739 if (spdk_nvme_transport_id_compare(&canon_trid, port->trid) == 0) { 740 return port; 741 } 742 } 743 744 return NULL; 745 } 746 747 static int 748 nvmf_tcp_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 749 struct spdk_nvmf_listen_opts *listen_opts) 750 { 751 struct spdk_nvmf_tcp_transport *ttransport; 752 struct spdk_nvmf_tcp_port *port; 753 int trsvcid_int; 754 uint8_t adrfam; 755 struct spdk_sock_opts opts; 756 757 if (!strlen(trid->trsvcid)) { 758 SPDK_ERRLOG("Service id is required\n"); 759 return -EINVAL; 760 } 761 762 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 763 764 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 765 if (trsvcid_int < 0) { 766 SPDK_ERRLOG("Invalid trsvcid '%s'\n", trid->trsvcid); 767 return -EINVAL; 768 } 769 770 port = calloc(1, sizeof(*port)); 771 if (!port) { 772 SPDK_ERRLOG("Port allocation failed\n"); 773 return -ENOMEM; 774 } 775 776 port->trid = trid; 777 opts.opts_size = sizeof(opts); 778 spdk_sock_get_default_opts(&opts); 779 opts.priority = ttransport->tcp_opts.sock_priority; 780 /* TODO: also add impl_opts like on the initiator side */ 781 port->listen_sock = spdk_sock_listen_ext(trid->traddr, trsvcid_int, 782 NULL, &opts); 783 if (port->listen_sock == NULL) { 784 SPDK_ERRLOG("spdk_sock_listen(%s, %d) failed: %s (%d)\n", 785 trid->traddr, trsvcid_int, 786 spdk_strerror(errno), errno); 787 free(port); 788 return -errno; 789 } 790 791 if (spdk_sock_is_ipv4(port->listen_sock)) { 792 adrfam = SPDK_NVMF_ADRFAM_IPV4; 793 } else if (spdk_sock_is_ipv6(port->listen_sock)) { 794 adrfam = SPDK_NVMF_ADRFAM_IPV6; 795 } else { 796 SPDK_ERRLOG("Unhandled socket type\n"); 797 adrfam = 0; 798 } 799 800 if (adrfam != trid->adrfam) { 801 SPDK_ERRLOG("Socket address family mismatch\n"); 802 spdk_sock_close(&port->listen_sock); 803 free(port); 804 return -EINVAL; 805 } 806 807 SPDK_NOTICELOG("*** NVMe/TCP Target Listening on %s port %s ***\n", 808 trid->traddr, trid->trsvcid); 809 810 TAILQ_INSERT_TAIL(&ttransport->ports, port, link); 811 return 0; 812 } 813 814 static void 815 nvmf_tcp_stop_listen(struct spdk_nvmf_transport *transport, 816 const struct spdk_nvme_transport_id *trid) 817 { 818 struct spdk_nvmf_tcp_transport *ttransport; 819 struct spdk_nvmf_tcp_port *port; 820 821 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 822 823 SPDK_DEBUGLOG(nvmf_tcp, "Removing listen address %s port %s\n", 824 trid->traddr, trid->trsvcid); 825 826 port = nvmf_tcp_find_port(ttransport, trid); 827 if (port) { 828 TAILQ_REMOVE(&ttransport->ports, port, link); 829 spdk_sock_close(&port->listen_sock); 830 free(port); 831 } 832 } 833 834 static void nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 835 enum nvme_tcp_pdu_recv_state state); 836 837 static void 838 nvmf_tcp_qpair_set_state(struct spdk_nvmf_tcp_qpair *tqpair, enum nvme_tcp_qpair_state state) 839 { 840 tqpair->state = state; 841 spdk_trace_record(TRACE_TCP_QP_STATE_CHANGE, tqpair->qpair.qid, 0, (uintptr_t)tqpair, 842 tqpair->state); 843 } 844 845 static void 846 nvmf_tcp_qpair_disconnect(struct spdk_nvmf_tcp_qpair *tqpair) 847 { 848 SPDK_DEBUGLOG(nvmf_tcp, "Disconnecting qpair %p\n", tqpair); 849 850 spdk_trace_record(TRACE_TCP_QP_DISCONNECT, 0, 0, (uintptr_t)tqpair); 851 852 if (tqpair->state <= NVME_TCP_QPAIR_STATE_RUNNING) { 853 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITING); 854 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 855 spdk_poller_unregister(&tqpair->timeout_poller); 856 857 /* This will end up calling nvmf_tcp_close_qpair */ 858 spdk_nvmf_qpair_disconnect(&tqpair->qpair, NULL, NULL); 859 } 860 } 861 862 static void 863 _mgmt_pdu_write_done(void *_tqpair, int err) 864 { 865 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 866 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 867 868 if (spdk_unlikely(err != 0)) { 869 nvmf_tcp_qpair_disconnect(tqpair); 870 return; 871 } 872 873 assert(pdu->cb_fn != NULL); 874 pdu->cb_fn(pdu->cb_arg); 875 } 876 877 static void 878 _req_pdu_write_done(void *req, int err) 879 { 880 struct spdk_nvmf_tcp_req *tcp_req = req; 881 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 882 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 883 884 assert(tcp_req->pdu_in_use); 885 tcp_req->pdu_in_use = false; 886 887 /* If the request is in a completed state, we're waiting for write completion to free it */ 888 if (spdk_unlikely(tcp_req->state == TCP_REQUEST_STATE_COMPLETED)) { 889 nvmf_tcp_request_free(tcp_req); 890 return; 891 } 892 893 if (spdk_unlikely(err != 0)) { 894 nvmf_tcp_qpair_disconnect(tqpair); 895 return; 896 } 897 898 assert(pdu->cb_fn != NULL); 899 pdu->cb_fn(pdu->cb_arg); 900 } 901 902 static void 903 _pdu_write_done(struct nvme_tcp_pdu *pdu, int err) 904 { 905 pdu->sock_req.cb_fn(pdu->sock_req.cb_arg, err); 906 } 907 908 static void 909 _tcp_write_pdu(struct nvme_tcp_pdu *pdu) 910 { 911 uint32_t mapped_length = 0; 912 ssize_t rc; 913 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 914 915 pdu->sock_req.iovcnt = nvme_tcp_build_iovs(pdu->iov, SPDK_COUNTOF(pdu->iov), pdu, 916 tqpair->host_hdgst_enable, tqpair->host_ddgst_enable, 917 &mapped_length); 918 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP || 919 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ) { 920 rc = spdk_sock_writev(tqpair->sock, pdu->iov, pdu->sock_req.iovcnt); 921 if (rc == mapped_length) { 922 _pdu_write_done(pdu, 0); 923 } else { 924 SPDK_ERRLOG("IC_RESP or TERM_REQ could not write to socket.\n"); 925 _pdu_write_done(pdu, -1); 926 } 927 } else { 928 spdk_sock_writev_async(tqpair->sock, &pdu->sock_req); 929 } 930 } 931 932 static void 933 data_crc32_accel_done(void *cb_arg, int status) 934 { 935 struct nvme_tcp_pdu *pdu = cb_arg; 936 937 if (spdk_unlikely(status)) { 938 SPDK_ERRLOG("Failed to compute the data digest for pdu =%p\n", pdu); 939 _pdu_write_done(pdu, status); 940 return; 941 } 942 943 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 944 MAKE_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32); 945 946 _tcp_write_pdu(pdu); 947 } 948 949 static void 950 pdu_data_crc32_compute(struct nvme_tcp_pdu *pdu) 951 { 952 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 953 int rc = 0; 954 955 /* Data Digest */ 956 if (pdu->data_len > 0 && g_nvme_tcp_ddgst[pdu->hdr.common.pdu_type] && tqpair->host_ddgst_enable) { 957 /* Only suport this limitated case for the first step */ 958 if (spdk_likely(!pdu->dif_ctx && (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0) 959 && tqpair->group)) { 960 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 961 pdu->data_iovcnt, 0, data_crc32_accel_done, pdu); 962 if (spdk_likely(rc == 0)) { 963 return; 964 } 965 } else { 966 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 967 } 968 data_crc32_accel_done(pdu, rc); 969 } else { 970 _tcp_write_pdu(pdu); 971 } 972 } 973 974 static void 975 nvmf_tcp_qpair_write_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 976 struct nvme_tcp_pdu *pdu, 977 nvme_tcp_qpair_xfer_complete_cb cb_fn, 978 void *cb_arg) 979 { 980 int hlen; 981 uint32_t crc32c; 982 983 assert(tqpair->pdu_in_progress != pdu); 984 985 hlen = pdu->hdr.common.hlen; 986 pdu->cb_fn = cb_fn; 987 pdu->cb_arg = cb_arg; 988 989 pdu->iov[0].iov_base = &pdu->hdr.raw; 990 pdu->iov[0].iov_len = hlen; 991 992 /* Header Digest */ 993 if (g_nvme_tcp_hdgst[pdu->hdr.common.pdu_type] && tqpair->host_hdgst_enable) { 994 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 995 MAKE_DIGEST_WORD((uint8_t *)pdu->hdr.raw + hlen, crc32c); 996 } 997 998 /* Data Digest */ 999 pdu_data_crc32_compute(pdu); 1000 } 1001 1002 static void 1003 nvmf_tcp_qpair_write_mgmt_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1004 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1005 void *cb_arg) 1006 { 1007 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1008 1009 pdu->sock_req.cb_fn = _mgmt_pdu_write_done; 1010 pdu->sock_req.cb_arg = tqpair; 1011 1012 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1013 } 1014 1015 static void 1016 nvmf_tcp_qpair_write_req_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1017 struct spdk_nvmf_tcp_req *tcp_req, 1018 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1019 void *cb_arg) 1020 { 1021 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1022 1023 pdu->sock_req.cb_fn = _req_pdu_write_done; 1024 pdu->sock_req.cb_arg = tcp_req; 1025 1026 assert(!tcp_req->pdu_in_use); 1027 tcp_req->pdu_in_use = true; 1028 1029 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1030 } 1031 1032 static int 1033 nvmf_tcp_qpair_init_mem_resource(struct spdk_nvmf_tcp_qpair *tqpair) 1034 { 1035 uint32_t i; 1036 struct spdk_nvmf_transport_opts *opts; 1037 uint32_t in_capsule_data_size; 1038 1039 opts = &tqpair->qpair.transport->opts; 1040 1041 in_capsule_data_size = opts->in_capsule_data_size; 1042 if (opts->dif_insert_or_strip) { 1043 in_capsule_data_size = SPDK_BDEV_BUF_SIZE_WITH_MD(in_capsule_data_size); 1044 } 1045 1046 tqpair->resource_count = opts->max_queue_depth; 1047 1048 tqpair->reqs = calloc(tqpair->resource_count, sizeof(*tqpair->reqs)); 1049 if (!tqpair->reqs) { 1050 SPDK_ERRLOG("Unable to allocate reqs on tqpair=%p\n", tqpair); 1051 return -1; 1052 } 1053 1054 if (in_capsule_data_size) { 1055 tqpair->bufs = spdk_zmalloc(tqpair->resource_count * in_capsule_data_size, 0x1000, 1056 NULL, SPDK_ENV_LCORE_ID_ANY, 1057 SPDK_MALLOC_DMA); 1058 if (!tqpair->bufs) { 1059 SPDK_ERRLOG("Unable to allocate bufs on tqpair=%p.\n", tqpair); 1060 return -1; 1061 } 1062 } 1063 /* prepare memory space for receiving pdus and tcp_req */ 1064 /* Add additional 1 member, which will be used for mgmt_pdu owned by the tqpair */ 1065 tqpair->pdus = spdk_dma_zmalloc((2 * tqpair->resource_count + 1) * sizeof(*tqpair->pdus), 0x1000, 1066 NULL); 1067 if (!tqpair->pdus) { 1068 SPDK_ERRLOG("Unable to allocate pdu pool on tqpair =%p.\n", tqpair); 1069 return -1; 1070 } 1071 1072 for (i = 0; i < tqpair->resource_count; i++) { 1073 struct spdk_nvmf_tcp_req *tcp_req = &tqpair->reqs[i]; 1074 1075 tcp_req->ttag = i + 1; 1076 tcp_req->req.qpair = &tqpair->qpair; 1077 1078 tcp_req->pdu = &tqpair->pdus[i]; 1079 tcp_req->pdu->qpair = tqpair; 1080 1081 /* Set up memory to receive commands */ 1082 if (tqpair->bufs) { 1083 tcp_req->buf = (void *)((uintptr_t)tqpair->bufs + (i * in_capsule_data_size)); 1084 } 1085 1086 /* Set the cmdn and rsp */ 1087 tcp_req->req.rsp = (union nvmf_c2h_msg *)&tcp_req->rsp; 1088 tcp_req->req.cmd = (union nvmf_h2c_msg *)&tcp_req->cmd; 1089 1090 tcp_req->req.stripped_data = NULL; 1091 1092 /* Initialize request state to FREE */ 1093 tcp_req->state = TCP_REQUEST_STATE_FREE; 1094 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 1095 tqpair->state_cntr[TCP_REQUEST_STATE_FREE]++; 1096 } 1097 1098 for (; i < 2 * tqpair->resource_count; i++) { 1099 struct nvme_tcp_pdu *pdu = &tqpair->pdus[i]; 1100 1101 pdu->qpair = tqpair; 1102 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 1103 } 1104 1105 tqpair->mgmt_pdu = &tqpair->pdus[i]; 1106 tqpair->mgmt_pdu->qpair = tqpair; 1107 tqpair->pdu_in_progress = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 1108 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 1109 1110 tqpair->recv_buf_size = (in_capsule_data_size + sizeof(struct spdk_nvme_tcp_cmd) + 2 * 1111 SPDK_NVME_TCP_DIGEST_LEN) * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1112 1113 return 0; 1114 } 1115 1116 static int 1117 nvmf_tcp_qpair_init(struct spdk_nvmf_qpair *qpair) 1118 { 1119 struct spdk_nvmf_tcp_qpair *tqpair; 1120 1121 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1122 1123 SPDK_DEBUGLOG(nvmf_tcp, "New TCP Connection: %p\n", qpair); 1124 1125 spdk_trace_record(TRACE_TCP_QP_CREATE, 0, 0, (uintptr_t)tqpair); 1126 1127 /* Initialise request state queues of the qpair */ 1128 TAILQ_INIT(&tqpair->tcp_req_free_queue); 1129 TAILQ_INIT(&tqpair->tcp_req_working_queue); 1130 SLIST_INIT(&tqpair->tcp_pdu_free_queue); 1131 1132 tqpair->host_hdgst_enable = true; 1133 tqpair->host_ddgst_enable = true; 1134 1135 return 0; 1136 } 1137 1138 static int 1139 nvmf_tcp_qpair_sock_init(struct spdk_nvmf_tcp_qpair *tqpair) 1140 { 1141 int rc; 1142 1143 spdk_trace_record(TRACE_TCP_QP_SOCK_INIT, 0, 0, (uintptr_t)tqpair); 1144 1145 /* set low water mark */ 1146 rc = spdk_sock_set_recvlowat(tqpair->sock, 1); 1147 if (rc != 0) { 1148 SPDK_ERRLOG("spdk_sock_set_recvlowat() failed\n"); 1149 return rc; 1150 } 1151 1152 return 0; 1153 } 1154 1155 static void 1156 nvmf_tcp_handle_connect(struct spdk_nvmf_transport *transport, 1157 struct spdk_nvmf_tcp_port *port, 1158 struct spdk_sock *sock) 1159 { 1160 struct spdk_nvmf_tcp_qpair *tqpair; 1161 int rc; 1162 1163 SPDK_DEBUGLOG(nvmf_tcp, "New connection accepted on %s port %s\n", 1164 port->trid->traddr, port->trid->trsvcid); 1165 1166 tqpair = calloc(1, sizeof(struct spdk_nvmf_tcp_qpair)); 1167 if (tqpair == NULL) { 1168 SPDK_ERRLOG("Could not allocate new connection.\n"); 1169 spdk_sock_close(&sock); 1170 return; 1171 } 1172 1173 tqpair->sock = sock; 1174 tqpair->state_cntr[TCP_REQUEST_STATE_FREE] = 0; 1175 tqpair->port = port; 1176 tqpair->qpair.transport = transport; 1177 1178 rc = spdk_sock_getaddr(tqpair->sock, tqpair->target_addr, 1179 sizeof(tqpair->target_addr), &tqpair->target_port, 1180 tqpair->initiator_addr, sizeof(tqpair->initiator_addr), 1181 &tqpair->initiator_port); 1182 if (rc < 0) { 1183 SPDK_ERRLOG("spdk_sock_getaddr() failed of tqpair=%p\n", tqpair); 1184 nvmf_tcp_qpair_destroy(tqpair); 1185 return; 1186 } 1187 1188 spdk_nvmf_tgt_new_qpair(transport->tgt, &tqpair->qpair); 1189 } 1190 1191 static uint32_t 1192 nvmf_tcp_port_accept(struct spdk_nvmf_transport *transport, struct spdk_nvmf_tcp_port *port) 1193 { 1194 struct spdk_sock *sock; 1195 uint32_t count = 0; 1196 int i; 1197 1198 for (i = 0; i < NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME; i++) { 1199 sock = spdk_sock_accept(port->listen_sock); 1200 if (sock == NULL) { 1201 break; 1202 } 1203 count++; 1204 nvmf_tcp_handle_connect(transport, port, sock); 1205 } 1206 1207 return count; 1208 } 1209 1210 static int 1211 nvmf_tcp_accept(void *ctx) 1212 { 1213 struct spdk_nvmf_transport *transport = ctx; 1214 struct spdk_nvmf_tcp_transport *ttransport; 1215 struct spdk_nvmf_tcp_port *port; 1216 uint32_t count = 0; 1217 1218 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1219 1220 TAILQ_FOREACH(port, &ttransport->ports, link) { 1221 count += nvmf_tcp_port_accept(transport, port); 1222 } 1223 1224 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 1225 } 1226 1227 static void 1228 nvmf_tcp_discover(struct spdk_nvmf_transport *transport, 1229 struct spdk_nvme_transport_id *trid, 1230 struct spdk_nvmf_discovery_log_page_entry *entry) 1231 { 1232 entry->trtype = SPDK_NVMF_TRTYPE_TCP; 1233 entry->adrfam = trid->adrfam; 1234 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 1235 1236 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 1237 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 1238 1239 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_NONE; 1240 } 1241 1242 static struct spdk_nvmf_tcp_control_msg_list * 1243 nvmf_tcp_control_msg_list_create(uint16_t num_messages) 1244 { 1245 struct spdk_nvmf_tcp_control_msg_list *list; 1246 struct spdk_nvmf_tcp_control_msg *msg; 1247 uint16_t i; 1248 1249 list = calloc(1, sizeof(*list)); 1250 if (!list) { 1251 SPDK_ERRLOG("Failed to allocate memory for list structure\n"); 1252 return NULL; 1253 } 1254 1255 list->msg_buf = spdk_zmalloc(num_messages * SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, 1256 NVMF_DATA_BUFFER_ALIGNMENT, NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 1257 if (!list->msg_buf) { 1258 SPDK_ERRLOG("Failed to allocate memory for control message buffers\n"); 1259 free(list); 1260 return NULL; 1261 } 1262 1263 STAILQ_INIT(&list->free_msgs); 1264 1265 for (i = 0; i < num_messages; i++) { 1266 msg = (struct spdk_nvmf_tcp_control_msg *)((char *)list->msg_buf + i * 1267 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1268 STAILQ_INSERT_TAIL(&list->free_msgs, msg, link); 1269 } 1270 1271 return list; 1272 } 1273 1274 static void 1275 nvmf_tcp_control_msg_list_free(struct spdk_nvmf_tcp_control_msg_list *list) 1276 { 1277 if (!list) { 1278 return; 1279 } 1280 1281 spdk_free(list->msg_buf); 1282 free(list); 1283 } 1284 1285 static struct spdk_nvmf_transport_poll_group * 1286 nvmf_tcp_poll_group_create(struct spdk_nvmf_transport *transport, 1287 struct spdk_nvmf_poll_group *group) 1288 { 1289 struct spdk_nvmf_tcp_transport *ttransport; 1290 struct spdk_nvmf_tcp_poll_group *tgroup; 1291 1292 tgroup = calloc(1, sizeof(*tgroup)); 1293 if (!tgroup) { 1294 return NULL; 1295 } 1296 1297 tgroup->sock_group = spdk_sock_group_create(&tgroup->group); 1298 if (!tgroup->sock_group) { 1299 goto cleanup; 1300 } 1301 1302 TAILQ_INIT(&tgroup->qpairs); 1303 TAILQ_INIT(&tgroup->await_req); 1304 1305 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1306 1307 if (transport->opts.in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 1308 SPDK_DEBUGLOG(nvmf_tcp, "ICD %u is less than min required for admin/fabric commands (%u). " 1309 "Creating control messages list\n", transport->opts.in_capsule_data_size, 1310 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1311 tgroup->control_msg_list = nvmf_tcp_control_msg_list_create(ttransport->tcp_opts.control_msg_num); 1312 if (!tgroup->control_msg_list) { 1313 goto cleanup; 1314 } 1315 } 1316 1317 tgroup->accel_channel = spdk_accel_get_io_channel(); 1318 if (spdk_unlikely(!tgroup->accel_channel)) { 1319 SPDK_ERRLOG("Cannot create accel_channel for tgroup=%p\n", tgroup); 1320 goto cleanup; 1321 } 1322 1323 TAILQ_INSERT_TAIL(&ttransport->poll_groups, tgroup, link); 1324 if (ttransport->next_pg == NULL) { 1325 ttransport->next_pg = tgroup; 1326 } 1327 1328 return &tgroup->group; 1329 1330 cleanup: 1331 nvmf_tcp_poll_group_destroy(&tgroup->group); 1332 return NULL; 1333 } 1334 1335 static struct spdk_nvmf_transport_poll_group * 1336 nvmf_tcp_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 1337 { 1338 struct spdk_nvmf_tcp_transport *ttransport; 1339 struct spdk_nvmf_tcp_poll_group **pg; 1340 struct spdk_nvmf_tcp_qpair *tqpair; 1341 struct spdk_sock_group *group = NULL, *hint = NULL; 1342 int rc; 1343 1344 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 1345 1346 if (TAILQ_EMPTY(&ttransport->poll_groups)) { 1347 return NULL; 1348 } 1349 1350 pg = &ttransport->next_pg; 1351 assert(*pg != NULL); 1352 hint = (*pg)->sock_group; 1353 1354 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1355 rc = spdk_sock_get_optimal_sock_group(tqpair->sock, &group, hint); 1356 if (rc != 0) { 1357 return NULL; 1358 } else if (group != NULL) { 1359 /* Optimal poll group was found */ 1360 return spdk_sock_group_get_ctx(group); 1361 } 1362 1363 /* The hint was used for optimal poll group, advance next_pg. */ 1364 *pg = TAILQ_NEXT(*pg, link); 1365 if (*pg == NULL) { 1366 *pg = TAILQ_FIRST(&ttransport->poll_groups); 1367 } 1368 1369 return spdk_sock_group_get_ctx(hint); 1370 } 1371 1372 static void 1373 nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 1374 { 1375 struct spdk_nvmf_tcp_poll_group *tgroup, *next_tgroup; 1376 struct spdk_nvmf_tcp_transport *ttransport; 1377 1378 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 1379 spdk_sock_group_close(&tgroup->sock_group); 1380 if (tgroup->control_msg_list) { 1381 nvmf_tcp_control_msg_list_free(tgroup->control_msg_list); 1382 } 1383 1384 if (tgroup->accel_channel) { 1385 spdk_put_io_channel(tgroup->accel_channel); 1386 } 1387 1388 ttransport = SPDK_CONTAINEROF(tgroup->group.transport, struct spdk_nvmf_tcp_transport, transport); 1389 1390 next_tgroup = TAILQ_NEXT(tgroup, link); 1391 TAILQ_REMOVE(&ttransport->poll_groups, tgroup, link); 1392 if (next_tgroup == NULL) { 1393 next_tgroup = TAILQ_FIRST(&ttransport->poll_groups); 1394 } 1395 if (ttransport->next_pg == tgroup) { 1396 ttransport->next_pg = next_tgroup; 1397 } 1398 1399 free(tgroup); 1400 } 1401 1402 static void 1403 nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 1404 enum nvme_tcp_pdu_recv_state state) 1405 { 1406 if (tqpair->recv_state == state) { 1407 SPDK_ERRLOG("The recv state of tqpair=%p is same with the state(%d) to be set\n", 1408 tqpair, state); 1409 return; 1410 } 1411 1412 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1413 /* When leaving the await req state, move the qpair to the main list */ 1414 TAILQ_REMOVE(&tqpair->group->await_req, tqpair, link); 1415 TAILQ_INSERT_TAIL(&tqpair->group->qpairs, tqpair, link); 1416 } else if (state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1417 TAILQ_REMOVE(&tqpair->group->qpairs, tqpair, link); 1418 TAILQ_INSERT_TAIL(&tqpair->group->await_req, tqpair, link); 1419 } 1420 1421 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv state=%d\n", tqpair, state); 1422 tqpair->recv_state = state; 1423 1424 spdk_trace_record(TRACE_TCP_QP_RCV_STATE_CHANGE, tqpair->qpair.qid, 0, (uintptr_t)tqpair, 1425 tqpair->recv_state); 1426 } 1427 1428 static int 1429 nvmf_tcp_qpair_handle_timeout(void *ctx) 1430 { 1431 struct spdk_nvmf_tcp_qpair *tqpair = ctx; 1432 1433 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1434 1435 SPDK_ERRLOG("No pdu coming for tqpair=%p within %d seconds\n", tqpair, 1436 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT); 1437 1438 nvmf_tcp_qpair_disconnect(tqpair); 1439 return SPDK_POLLER_BUSY; 1440 } 1441 1442 static void 1443 nvmf_tcp_send_c2h_term_req_complete(void *cb_arg) 1444 { 1445 struct spdk_nvmf_tcp_qpair *tqpair = (struct spdk_nvmf_tcp_qpair *)cb_arg; 1446 1447 if (!tqpair->timeout_poller) { 1448 tqpair->timeout_poller = SPDK_POLLER_REGISTER(nvmf_tcp_qpair_handle_timeout, tqpair, 1449 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT * 1000000); 1450 } 1451 } 1452 1453 static void 1454 nvmf_tcp_send_c2h_term_req(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu, 1455 enum spdk_nvme_tcp_term_req_fes fes, uint32_t error_offset) 1456 { 1457 struct nvme_tcp_pdu *rsp_pdu; 1458 struct spdk_nvme_tcp_term_req_hdr *c2h_term_req; 1459 uint32_t c2h_term_req_hdr_len = sizeof(*c2h_term_req); 1460 uint32_t copy_len; 1461 1462 rsp_pdu = tqpair->mgmt_pdu; 1463 1464 c2h_term_req = &rsp_pdu->hdr.term_req; 1465 c2h_term_req->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ; 1466 c2h_term_req->common.hlen = c2h_term_req_hdr_len; 1467 1468 if ((fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1469 (fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1470 DSET32(&c2h_term_req->fei, error_offset); 1471 } 1472 1473 copy_len = spdk_min(pdu->hdr.common.hlen, SPDK_NVME_TCP_TERM_REQ_ERROR_DATA_MAX_SIZE); 1474 1475 /* Copy the error info into the buffer */ 1476 memcpy((uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, pdu->hdr.raw, copy_len); 1477 nvme_tcp_pdu_set_data(rsp_pdu, (uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, copy_len); 1478 1479 /* Contain the header of the wrong received pdu */ 1480 c2h_term_req->common.plen = c2h_term_req->common.hlen + copy_len; 1481 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 1482 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_c2h_term_req_complete, tqpair); 1483 } 1484 1485 static void 1486 nvmf_tcp_capsule_cmd_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1487 struct spdk_nvmf_tcp_qpair *tqpair, 1488 struct nvme_tcp_pdu *pdu) 1489 { 1490 struct spdk_nvmf_tcp_req *tcp_req; 1491 1492 assert(pdu->psh_valid_bytes == pdu->psh_len); 1493 assert(pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD); 1494 1495 tcp_req = nvmf_tcp_req_get(tqpair); 1496 if (!tcp_req) { 1497 /* Directly return and make the allocation retry again. This can happen if we're 1498 * using asynchronous writes to send the response to the host or when releasing 1499 * zero-copy buffers after a response has been sent. In both cases, the host might 1500 * receive the response before we've finished processing the request and is free to 1501 * send another one. 1502 */ 1503 if (tqpair->state_cntr[TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST] > 0 || 1504 tqpair->state_cntr[TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE] > 0) { 1505 return; 1506 } 1507 1508 /* The host sent more commands than the maximum queue depth. */ 1509 SPDK_ERRLOG("Cannot allocate tcp_req on tqpair=%p\n", tqpair); 1510 nvmf_tcp_qpair_disconnect(tqpair); 1511 return; 1512 } 1513 1514 pdu->req = tcp_req; 1515 assert(tcp_req->state == TCP_REQUEST_STATE_NEW); 1516 nvmf_tcp_req_process(ttransport, tcp_req); 1517 } 1518 1519 static void 1520 nvmf_tcp_capsule_cmd_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1521 struct spdk_nvmf_tcp_qpair *tqpair, 1522 struct nvme_tcp_pdu *pdu) 1523 { 1524 struct spdk_nvmf_tcp_req *tcp_req; 1525 struct spdk_nvme_tcp_cmd *capsule_cmd; 1526 uint32_t error_offset = 0; 1527 enum spdk_nvme_tcp_term_req_fes fes; 1528 struct spdk_nvme_cpl *rsp; 1529 1530 capsule_cmd = &pdu->hdr.capsule_cmd; 1531 tcp_req = pdu->req; 1532 assert(tcp_req != NULL); 1533 1534 /* Zero-copy requests don't support ICD */ 1535 assert(!spdk_nvmf_request_using_zcopy(&tcp_req->req)); 1536 1537 if (capsule_cmd->common.pdo > SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET) { 1538 SPDK_ERRLOG("Expected ICReq capsule_cmd pdu offset <= %d, got %c\n", 1539 SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET, capsule_cmd->common.pdo); 1540 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1541 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 1542 goto err; 1543 } 1544 1545 rsp = &tcp_req->req.rsp->nvme_cpl; 1546 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1547 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1548 } else { 1549 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1550 } 1551 1552 nvmf_tcp_req_process(ttransport, tcp_req); 1553 1554 return; 1555 err: 1556 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1557 } 1558 1559 static void 1560 nvmf_tcp_h2c_data_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1561 struct spdk_nvmf_tcp_qpair *tqpair, 1562 struct nvme_tcp_pdu *pdu) 1563 { 1564 struct spdk_nvmf_tcp_req *tcp_req; 1565 uint32_t error_offset = 0; 1566 enum spdk_nvme_tcp_term_req_fes fes = 0; 1567 struct spdk_nvme_tcp_h2c_data_hdr *h2c_data; 1568 1569 h2c_data = &pdu->hdr.h2c_data; 1570 1571 SPDK_DEBUGLOG(nvmf_tcp, "tqpair=%p, r2t_info: datao=%u, datal=%u, cccid=%u, ttag=%u\n", 1572 tqpair, h2c_data->datao, h2c_data->datal, h2c_data->cccid, h2c_data->ttag); 1573 1574 if (h2c_data->ttag > tqpair->resource_count) { 1575 SPDK_DEBUGLOG(nvmf_tcp, "ttag %u is larger than allowed %u.\n", h2c_data->ttag, 1576 tqpair->resource_count); 1577 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1578 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1579 goto err; 1580 } 1581 1582 tcp_req = &tqpair->reqs[h2c_data->ttag - 1]; 1583 1584 if (spdk_unlikely(tcp_req->state != TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER && 1585 tcp_req->state != TCP_REQUEST_STATE_AWAITING_R2T_ACK)) { 1586 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, has error state in %d\n", tcp_req, tqpair, 1587 tcp_req->state); 1588 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1589 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1590 goto err; 1591 } 1592 1593 if (spdk_unlikely(tcp_req->req.cmd->nvme_cmd.cid != h2c_data->cccid)) { 1594 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, expected %u but %u for cccid.\n", tcp_req, tqpair, 1595 tcp_req->req.cmd->nvme_cmd.cid, h2c_data->cccid); 1596 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1597 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, cccid); 1598 goto err; 1599 } 1600 1601 if (tcp_req->h2c_offset != h2c_data->datao) { 1602 SPDK_DEBUGLOG(nvmf_tcp, 1603 "tcp_req(%p), tqpair=%p, expected data offset %u, but data offset is %u\n", 1604 tcp_req, tqpair, tcp_req->h2c_offset, h2c_data->datao); 1605 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1606 goto err; 1607 } 1608 1609 if ((h2c_data->datao + h2c_data->datal) > tcp_req->req.length) { 1610 SPDK_DEBUGLOG(nvmf_tcp, 1611 "tcp_req(%p), tqpair=%p, (datao=%u + datal=%u) exceeds requested length=%u\n", 1612 tcp_req, tqpair, h2c_data->datao, h2c_data->datal, tcp_req->req.length); 1613 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1614 goto err; 1615 } 1616 1617 pdu->req = tcp_req; 1618 1619 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 1620 pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 1621 } 1622 1623 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 1624 h2c_data->datao, h2c_data->datal); 1625 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1626 return; 1627 1628 err: 1629 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1630 } 1631 1632 static void 1633 nvmf_tcp_send_capsule_resp_pdu(struct spdk_nvmf_tcp_req *tcp_req, 1634 struct spdk_nvmf_tcp_qpair *tqpair) 1635 { 1636 struct nvme_tcp_pdu *rsp_pdu; 1637 struct spdk_nvme_tcp_rsp *capsule_resp; 1638 1639 SPDK_DEBUGLOG(nvmf_tcp, "enter, tqpair=%p\n", tqpair); 1640 1641 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1642 assert(rsp_pdu != NULL); 1643 1644 capsule_resp = &rsp_pdu->hdr.capsule_resp; 1645 capsule_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_CAPSULE_RESP; 1646 capsule_resp->common.plen = capsule_resp->common.hlen = sizeof(*capsule_resp); 1647 capsule_resp->rccqe = tcp_req->req.rsp->nvme_cpl; 1648 if (tqpair->host_hdgst_enable) { 1649 capsule_resp->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 1650 capsule_resp->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 1651 } 1652 1653 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_request_free, tcp_req); 1654 } 1655 1656 static void 1657 nvmf_tcp_pdu_c2h_data_complete(void *cb_arg) 1658 { 1659 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 1660 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, 1661 struct spdk_nvmf_tcp_qpair, qpair); 1662 1663 assert(tqpair != NULL); 1664 1665 if (spdk_unlikely(tcp_req->pdu->rw_offset < tcp_req->req.length)) { 1666 SPDK_DEBUGLOG(nvmf_tcp, "sending another C2H part, offset %u length %u\n", tcp_req->pdu->rw_offset, 1667 tcp_req->req.length); 1668 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 1669 return; 1670 } 1671 1672 if (tcp_req->pdu->hdr.c2h_data.common.flags & SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS) { 1673 nvmf_tcp_request_free(tcp_req); 1674 } else { 1675 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 1676 } 1677 } 1678 1679 static void 1680 nvmf_tcp_r2t_complete(void *cb_arg) 1681 { 1682 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 1683 struct spdk_nvmf_tcp_transport *ttransport; 1684 1685 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 1686 struct spdk_nvmf_tcp_transport, transport); 1687 1688 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1689 1690 if (tcp_req->h2c_offset == tcp_req->req.length) { 1691 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1692 nvmf_tcp_req_process(ttransport, tcp_req); 1693 } 1694 } 1695 1696 static void 1697 nvmf_tcp_send_r2t_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1698 struct spdk_nvmf_tcp_req *tcp_req) 1699 { 1700 struct nvme_tcp_pdu *rsp_pdu; 1701 struct spdk_nvme_tcp_r2t_hdr *r2t; 1702 1703 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1704 assert(rsp_pdu != NULL); 1705 1706 r2t = &rsp_pdu->hdr.r2t; 1707 r2t->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_R2T; 1708 r2t->common.plen = r2t->common.hlen = sizeof(*r2t); 1709 1710 if (tqpair->host_hdgst_enable) { 1711 r2t->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 1712 r2t->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 1713 } 1714 1715 r2t->cccid = tcp_req->req.cmd->nvme_cmd.cid; 1716 r2t->ttag = tcp_req->ttag; 1717 r2t->r2to = tcp_req->h2c_offset; 1718 r2t->r2tl = tcp_req->req.length; 1719 1720 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 1721 1722 SPDK_DEBUGLOG(nvmf_tcp, 1723 "tcp_req(%p) on tqpair(%p), r2t_info: cccid=%u, ttag=%u, r2to=%u, r2tl=%u\n", 1724 tcp_req, tqpair, r2t->cccid, r2t->ttag, r2t->r2to, r2t->r2tl); 1725 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_r2t_complete, tcp_req); 1726 } 1727 1728 static void 1729 nvmf_tcp_h2c_data_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1730 struct spdk_nvmf_tcp_qpair *tqpair, 1731 struct nvme_tcp_pdu *pdu) 1732 { 1733 struct spdk_nvmf_tcp_req *tcp_req; 1734 struct spdk_nvme_cpl *rsp; 1735 1736 tcp_req = pdu->req; 1737 assert(tcp_req != NULL); 1738 1739 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 1740 1741 tcp_req->h2c_offset += pdu->data_len; 1742 1743 /* Wait for all of the data to arrive AND for the initial R2T PDU send to be 1744 * acknowledged before moving on. */ 1745 if (tcp_req->h2c_offset == tcp_req->req.length && 1746 tcp_req->state == TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER) { 1747 /* After receiving all the h2c data, we need to check whether there is 1748 * transient transport error */ 1749 rsp = &tcp_req->req.rsp->nvme_cpl; 1750 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1751 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1752 } else { 1753 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1754 } 1755 nvmf_tcp_req_process(ttransport, tcp_req); 1756 } 1757 } 1758 1759 static void 1760 nvmf_tcp_h2c_term_req_dump(struct spdk_nvme_tcp_term_req_hdr *h2c_term_req) 1761 { 1762 SPDK_ERRLOG("Error info of pdu(%p): %s\n", h2c_term_req, 1763 spdk_nvmf_tcp_term_req_fes_str[h2c_term_req->fes]); 1764 if ((h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1765 (h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1766 SPDK_DEBUGLOG(nvmf_tcp, "The offset from the start of the PDU header is %u\n", 1767 DGET32(h2c_term_req->fei)); 1768 } 1769 } 1770 1771 static void 1772 nvmf_tcp_h2c_term_req_hdr_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1773 struct nvme_tcp_pdu *pdu) 1774 { 1775 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 1776 uint32_t error_offset = 0; 1777 enum spdk_nvme_tcp_term_req_fes fes; 1778 1779 if (h2c_term_req->fes > SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER) { 1780 SPDK_ERRLOG("Fatal Error Status(FES) is unknown for h2c_term_req pdu=%p\n", pdu); 1781 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1782 error_offset = offsetof(struct spdk_nvme_tcp_term_req_hdr, fes); 1783 goto end; 1784 } 1785 1786 /* set the data buffer */ 1787 nvme_tcp_pdu_set_data(pdu, (uint8_t *)pdu->hdr.raw + h2c_term_req->common.hlen, 1788 h2c_term_req->common.plen - h2c_term_req->common.hlen); 1789 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1790 return; 1791 end: 1792 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1793 } 1794 1795 static void 1796 nvmf_tcp_h2c_term_req_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1797 struct nvme_tcp_pdu *pdu) 1798 { 1799 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 1800 1801 nvmf_tcp_h2c_term_req_dump(h2c_term_req); 1802 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 1803 } 1804 1805 static void 1806 _nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 1807 { 1808 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 1809 struct spdk_nvmf_tcp_transport, transport); 1810 1811 switch (pdu->hdr.common.pdu_type) { 1812 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 1813 nvmf_tcp_capsule_cmd_payload_handle(ttransport, tqpair, pdu); 1814 break; 1815 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 1816 nvmf_tcp_h2c_data_payload_handle(ttransport, tqpair, pdu); 1817 break; 1818 1819 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 1820 nvmf_tcp_h2c_term_req_payload_handle(tqpair, pdu); 1821 break; 1822 1823 default: 1824 /* The code should not go to here */ 1825 SPDK_ERRLOG("ERROR pdu type %d\n", pdu->hdr.common.pdu_type); 1826 break; 1827 } 1828 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 1829 } 1830 1831 static void 1832 data_crc32_calc_done(void *cb_arg, int status) 1833 { 1834 struct nvme_tcp_pdu *pdu = cb_arg; 1835 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1836 struct spdk_nvmf_tcp_req *tcp_req; 1837 struct spdk_nvme_cpl *rsp; 1838 1839 /* async crc32 calculation is failed and use direct calculation to check */ 1840 if (spdk_unlikely(status)) { 1841 SPDK_ERRLOG("Data digest on tqpair=(%p) with pdu=%p failed to be calculated asynchronously\n", 1842 tqpair, pdu); 1843 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 1844 } 1845 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 1846 if (!MATCH_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32)) { 1847 SPDK_ERRLOG("Data digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 1848 tcp_req = pdu->req; 1849 assert(tcp_req != NULL); 1850 rsp = &tcp_req->req.rsp->nvme_cpl; 1851 rsp->status.sc = SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR; 1852 } 1853 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 1854 } 1855 1856 static void 1857 nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 1858 { 1859 int rc = 0; 1860 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1861 tqpair->pdu_in_progress = NULL; 1862 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 1863 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 1864 /* check data digest if need */ 1865 if (pdu->ddgst_enable) { 1866 if (!pdu->dif_ctx && tqpair->group && (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0)) { 1867 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 1868 pdu->data_iovcnt, 0, data_crc32_calc_done, pdu); 1869 if (spdk_likely(rc == 0)) { 1870 return; 1871 } 1872 } else { 1873 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 1874 } 1875 data_crc32_calc_done(pdu, rc); 1876 } else { 1877 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 1878 } 1879 } 1880 1881 static void 1882 nvmf_tcp_send_icresp_complete(void *cb_arg) 1883 { 1884 struct spdk_nvmf_tcp_qpair *tqpair = cb_arg; 1885 1886 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_RUNNING); 1887 } 1888 1889 static void 1890 nvmf_tcp_icreq_handle(struct spdk_nvmf_tcp_transport *ttransport, 1891 struct spdk_nvmf_tcp_qpair *tqpair, 1892 struct nvme_tcp_pdu *pdu) 1893 { 1894 struct spdk_nvme_tcp_ic_req *ic_req = &pdu->hdr.ic_req; 1895 struct nvme_tcp_pdu *rsp_pdu; 1896 struct spdk_nvme_tcp_ic_resp *ic_resp; 1897 uint32_t error_offset = 0; 1898 enum spdk_nvme_tcp_term_req_fes fes; 1899 1900 /* Only PFV 0 is defined currently */ 1901 if (ic_req->pfv != 0) { 1902 SPDK_ERRLOG("Expected ICReq PFV %u, got %u\n", 0u, ic_req->pfv); 1903 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1904 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, pfv); 1905 goto end; 1906 } 1907 1908 /* MAXR2T is 0's based */ 1909 SPDK_DEBUGLOG(nvmf_tcp, "maxr2t =%u\n", (ic_req->maxr2t + 1u)); 1910 1911 tqpair->host_hdgst_enable = ic_req->dgst.bits.hdgst_enable ? true : false; 1912 if (!tqpair->host_hdgst_enable) { 1913 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1914 } 1915 1916 tqpair->host_ddgst_enable = ic_req->dgst.bits.ddgst_enable ? true : false; 1917 if (!tqpair->host_ddgst_enable) { 1918 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1919 } 1920 1921 tqpair->recv_buf_size = spdk_max(tqpair->recv_buf_size, MIN_SOCK_PIPE_SIZE); 1922 /* Now that we know whether digests are enabled, properly size the receive buffer */ 1923 if (spdk_sock_set_recvbuf(tqpair->sock, tqpair->recv_buf_size) < 0) { 1924 SPDK_WARNLOG("Unable to allocate enough memory for receive buffer on tqpair=%p with size=%d\n", 1925 tqpair, 1926 tqpair->recv_buf_size); 1927 /* Not fatal. */ 1928 } 1929 1930 tqpair->cpda = spdk_min(ic_req->hpda, SPDK_NVME_TCP_CPDA_MAX); 1931 SPDK_DEBUGLOG(nvmf_tcp, "cpda of tqpair=(%p) is : %u\n", tqpair, tqpair->cpda); 1932 1933 rsp_pdu = tqpair->mgmt_pdu; 1934 1935 ic_resp = &rsp_pdu->hdr.ic_resp; 1936 ic_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_IC_RESP; 1937 ic_resp->common.hlen = ic_resp->common.plen = sizeof(*ic_resp); 1938 ic_resp->pfv = 0; 1939 ic_resp->cpda = tqpair->cpda; 1940 ic_resp->maxh2cdata = ttransport->transport.opts.max_io_size; 1941 ic_resp->dgst.bits.hdgst_enable = tqpair->host_hdgst_enable ? 1 : 0; 1942 ic_resp->dgst.bits.ddgst_enable = tqpair->host_ddgst_enable ? 1 : 0; 1943 1944 SPDK_DEBUGLOG(nvmf_tcp, "host_hdgst_enable: %u\n", tqpair->host_hdgst_enable); 1945 SPDK_DEBUGLOG(nvmf_tcp, "host_ddgst_enable: %u\n", tqpair->host_ddgst_enable); 1946 1947 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INITIALIZING); 1948 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_icresp_complete, tqpair); 1949 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 1950 return; 1951 end: 1952 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1953 } 1954 1955 static void 1956 nvmf_tcp_pdu_psh_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1957 struct spdk_nvmf_tcp_transport *ttransport) 1958 { 1959 struct nvme_tcp_pdu *pdu; 1960 int rc; 1961 uint32_t crc32c, error_offset = 0; 1962 enum spdk_nvme_tcp_term_req_fes fes; 1963 1964 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 1965 pdu = tqpair->pdu_in_progress; 1966 1967 SPDK_DEBUGLOG(nvmf_tcp, "pdu type of tqpair(%p) is %d\n", tqpair, 1968 pdu->hdr.common.pdu_type); 1969 /* check header digest if needed */ 1970 if (pdu->has_hdgst) { 1971 SPDK_DEBUGLOG(nvmf_tcp, "Compare the header of pdu=%p on tqpair=%p\n", pdu, tqpair); 1972 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 1973 rc = MATCH_DIGEST_WORD((uint8_t *)pdu->hdr.raw + pdu->hdr.common.hlen, crc32c); 1974 if (rc == 0) { 1975 SPDK_ERRLOG("Header digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 1976 fes = SPDK_NVME_TCP_TERM_REQ_FES_HDGST_ERROR; 1977 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1978 return; 1979 1980 } 1981 } 1982 1983 switch (pdu->hdr.common.pdu_type) { 1984 case SPDK_NVME_TCP_PDU_TYPE_IC_REQ: 1985 nvmf_tcp_icreq_handle(ttransport, tqpair, pdu); 1986 break; 1987 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 1988 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_REQ); 1989 break; 1990 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 1991 nvmf_tcp_h2c_data_hdr_handle(ttransport, tqpair, pdu); 1992 break; 1993 1994 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 1995 nvmf_tcp_h2c_term_req_hdr_handle(tqpair, pdu); 1996 break; 1997 1998 default: 1999 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", tqpair->pdu_in_progress->hdr.common.pdu_type); 2000 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2001 error_offset = 1; 2002 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2003 break; 2004 } 2005 } 2006 2007 static void 2008 nvmf_tcp_pdu_ch_handle(struct spdk_nvmf_tcp_qpair *tqpair) 2009 { 2010 struct nvme_tcp_pdu *pdu; 2011 uint32_t error_offset = 0; 2012 enum spdk_nvme_tcp_term_req_fes fes; 2013 uint8_t expected_hlen, pdo; 2014 bool plen_error = false, pdo_error = false; 2015 2016 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2017 pdu = tqpair->pdu_in_progress; 2018 assert(pdu); 2019 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_REQ) { 2020 if (tqpair->state != NVME_TCP_QPAIR_STATE_INVALID) { 2021 SPDK_ERRLOG("Already received ICreq PDU, and reject this pdu=%p\n", pdu); 2022 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2023 goto err; 2024 } 2025 expected_hlen = sizeof(struct spdk_nvme_tcp_ic_req); 2026 if (pdu->hdr.common.plen != expected_hlen) { 2027 plen_error = true; 2028 } 2029 } else { 2030 if (tqpair->state != NVME_TCP_QPAIR_STATE_RUNNING) { 2031 SPDK_ERRLOG("The TCP/IP connection is not negotiated\n"); 2032 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2033 goto err; 2034 } 2035 2036 switch (pdu->hdr.common.pdu_type) { 2037 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2038 expected_hlen = sizeof(struct spdk_nvme_tcp_cmd); 2039 pdo = pdu->hdr.common.pdo; 2040 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2041 pdo_error = true; 2042 break; 2043 } 2044 2045 if (pdu->hdr.common.plen < expected_hlen) { 2046 plen_error = true; 2047 } 2048 break; 2049 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2050 expected_hlen = sizeof(struct spdk_nvme_tcp_h2c_data_hdr); 2051 pdo = pdu->hdr.common.pdo; 2052 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2053 pdo_error = true; 2054 break; 2055 } 2056 if (pdu->hdr.common.plen < expected_hlen) { 2057 plen_error = true; 2058 } 2059 break; 2060 2061 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2062 expected_hlen = sizeof(struct spdk_nvme_tcp_term_req_hdr); 2063 if ((pdu->hdr.common.plen <= expected_hlen) || 2064 (pdu->hdr.common.plen > SPDK_NVME_TCP_TERM_REQ_PDU_MAX_SIZE)) { 2065 plen_error = true; 2066 } 2067 break; 2068 2069 default: 2070 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", pdu->hdr.common.pdu_type); 2071 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2072 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdu_type); 2073 goto err; 2074 } 2075 } 2076 2077 if (pdu->hdr.common.hlen != expected_hlen) { 2078 SPDK_ERRLOG("PDU type=0x%02x, Expected ICReq header length %u, got %u on tqpair=%p\n", 2079 pdu->hdr.common.pdu_type, 2080 expected_hlen, pdu->hdr.common.hlen, tqpair); 2081 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2082 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, hlen); 2083 goto err; 2084 } else if (pdo_error) { 2085 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2086 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 2087 } else if (plen_error) { 2088 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2089 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2090 goto err; 2091 } else { 2092 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2093 nvme_tcp_pdu_calc_psh_len(tqpair->pdu_in_progress, tqpair->host_hdgst_enable); 2094 return; 2095 } 2096 err: 2097 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2098 } 2099 2100 static int 2101 nvmf_tcp_pdu_payload_insert_dif(struct nvme_tcp_pdu *pdu, uint32_t read_offset, 2102 int read_len) 2103 { 2104 int rc; 2105 2106 rc = spdk_dif_generate_stream(pdu->data_iov, pdu->data_iovcnt, 2107 read_offset, read_len, pdu->dif_ctx); 2108 if (rc != 0) { 2109 SPDK_ERRLOG("DIF generate failed\n"); 2110 } 2111 2112 return rc; 2113 } 2114 2115 static int 2116 nvmf_tcp_sock_process(struct spdk_nvmf_tcp_qpair *tqpair) 2117 { 2118 int rc = 0; 2119 struct nvme_tcp_pdu *pdu; 2120 enum nvme_tcp_pdu_recv_state prev_state; 2121 uint32_t data_len; 2122 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2123 struct spdk_nvmf_tcp_transport, transport); 2124 2125 /* The loop here is to allow for several back-to-back state changes. */ 2126 do { 2127 prev_state = tqpair->recv_state; 2128 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv pdu entering state %d\n", tqpair, prev_state); 2129 2130 pdu = tqpair->pdu_in_progress; 2131 assert(pdu || tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2132 switch (tqpair->recv_state) { 2133 /* Wait for the common header */ 2134 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY: 2135 if (!pdu) { 2136 pdu = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 2137 if (spdk_unlikely(!pdu)) { 2138 return NVME_TCP_PDU_IN_PROGRESS; 2139 } 2140 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 2141 tqpair->pdu_in_progress = pdu; 2142 } 2143 memset(pdu, 0, offsetof(struct nvme_tcp_pdu, qpair)); 2144 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2145 /* FALLTHROUGH */ 2146 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH: 2147 if (spdk_unlikely(tqpair->state == NVME_TCP_QPAIR_STATE_INITIALIZING)) { 2148 return rc; 2149 } 2150 2151 rc = nvme_tcp_read_data(tqpair->sock, 2152 sizeof(struct spdk_nvme_tcp_common_pdu_hdr) - pdu->ch_valid_bytes, 2153 (void *)&pdu->hdr.common + pdu->ch_valid_bytes); 2154 if (rc < 0) { 2155 SPDK_DEBUGLOG(nvmf_tcp, "will disconnect tqpair=%p\n", tqpair); 2156 return NVME_TCP_PDU_FATAL; 2157 } else if (rc > 0) { 2158 pdu->ch_valid_bytes += rc; 2159 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.qid, rc, 0, tqpair); 2160 } 2161 2162 if (pdu->ch_valid_bytes < sizeof(struct spdk_nvme_tcp_common_pdu_hdr)) { 2163 return NVME_TCP_PDU_IN_PROGRESS; 2164 } 2165 2166 /* The command header of this PDU has now been read from the socket. */ 2167 nvmf_tcp_pdu_ch_handle(tqpair); 2168 break; 2169 /* Wait for the pdu specific header */ 2170 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH: 2171 rc = nvme_tcp_read_data(tqpair->sock, 2172 pdu->psh_len - pdu->psh_valid_bytes, 2173 (void *)&pdu->hdr.raw + sizeof(struct spdk_nvme_tcp_common_pdu_hdr) + pdu->psh_valid_bytes); 2174 if (rc < 0) { 2175 return NVME_TCP_PDU_FATAL; 2176 } else if (rc > 0) { 2177 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.qid, rc, 0, tqpair); 2178 pdu->psh_valid_bytes += rc; 2179 } 2180 2181 if (pdu->psh_valid_bytes < pdu->psh_len) { 2182 return NVME_TCP_PDU_IN_PROGRESS; 2183 } 2184 2185 /* All header(ch, psh, head digist) of this PDU has now been read from the socket. */ 2186 nvmf_tcp_pdu_psh_handle(tqpair, ttransport); 2187 break; 2188 /* Wait for the req slot */ 2189 case NVME_TCP_PDU_RECV_STATE_AWAIT_REQ: 2190 nvmf_tcp_capsule_cmd_hdr_handle(ttransport, tqpair, pdu); 2191 break; 2192 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD: 2193 /* check whether the data is valid, if not we just return */ 2194 if (!pdu->data_len) { 2195 return NVME_TCP_PDU_IN_PROGRESS; 2196 } 2197 2198 data_len = pdu->data_len; 2199 /* data digest */ 2200 if (spdk_unlikely((pdu->hdr.common.pdu_type != SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ) && 2201 tqpair->host_ddgst_enable)) { 2202 data_len += SPDK_NVME_TCP_DIGEST_LEN; 2203 pdu->ddgst_enable = true; 2204 } 2205 2206 rc = nvme_tcp_read_payload_data(tqpair->sock, pdu); 2207 if (rc < 0) { 2208 return NVME_TCP_PDU_FATAL; 2209 } 2210 pdu->rw_offset += rc; 2211 2212 if (spdk_unlikely(pdu->dif_ctx != NULL)) { 2213 rc = nvmf_tcp_pdu_payload_insert_dif(pdu, pdu->rw_offset - rc, rc); 2214 if (rc != 0) { 2215 return NVME_TCP_PDU_FATAL; 2216 } 2217 } 2218 2219 if (pdu->rw_offset < data_len) { 2220 return NVME_TCP_PDU_IN_PROGRESS; 2221 } 2222 2223 /* All of this PDU has now been read from the socket. */ 2224 nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2225 break; 2226 case NVME_TCP_PDU_RECV_STATE_ERROR: 2227 if (!spdk_sock_is_connected(tqpair->sock)) { 2228 return NVME_TCP_PDU_FATAL; 2229 } 2230 break; 2231 default: 2232 SPDK_ERRLOG("The state(%d) is invalid\n", tqpair->recv_state); 2233 abort(); 2234 break; 2235 } 2236 } while (tqpair->recv_state != prev_state); 2237 2238 return rc; 2239 } 2240 2241 static inline void * 2242 nvmf_tcp_control_msg_get(struct spdk_nvmf_tcp_control_msg_list *list) 2243 { 2244 struct spdk_nvmf_tcp_control_msg *msg; 2245 2246 assert(list); 2247 2248 msg = STAILQ_FIRST(&list->free_msgs); 2249 if (!msg) { 2250 SPDK_DEBUGLOG(nvmf_tcp, "Out of control messages\n"); 2251 return NULL; 2252 } 2253 STAILQ_REMOVE_HEAD(&list->free_msgs, link); 2254 return msg; 2255 } 2256 2257 static inline void 2258 nvmf_tcp_control_msg_put(struct spdk_nvmf_tcp_control_msg_list *list, void *_msg) 2259 { 2260 struct spdk_nvmf_tcp_control_msg *msg = _msg; 2261 2262 assert(list); 2263 STAILQ_INSERT_HEAD(&list->free_msgs, msg, link); 2264 } 2265 2266 static int 2267 nvmf_tcp_req_parse_sgl(struct spdk_nvmf_tcp_req *tcp_req, 2268 struct spdk_nvmf_transport *transport, 2269 struct spdk_nvmf_transport_poll_group *group) 2270 { 2271 struct spdk_nvmf_request *req = &tcp_req->req; 2272 struct spdk_nvme_cmd *cmd; 2273 struct spdk_nvme_cpl *rsp; 2274 struct spdk_nvme_sgl_descriptor *sgl; 2275 struct spdk_nvmf_tcp_poll_group *tgroup; 2276 uint32_t length; 2277 2278 cmd = &req->cmd->nvme_cmd; 2279 rsp = &req->rsp->nvme_cpl; 2280 sgl = &cmd->dptr.sgl1; 2281 2282 length = sgl->unkeyed.length; 2283 2284 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK && 2285 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_TRANSPORT) { 2286 if (length > transport->opts.max_io_size) { 2287 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 2288 length, transport->opts.max_io_size); 2289 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 2290 return -1; 2291 } 2292 2293 /* fill request length and populate iovs */ 2294 req->length = length; 2295 2296 SPDK_DEBUGLOG(nvmf_tcp, "Data requested length= 0x%x\n", length); 2297 2298 if (spdk_unlikely(req->dif_enabled)) { 2299 req->dif.orig_length = length; 2300 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2301 req->dif.elba_length = length; 2302 } 2303 2304 if (nvmf_ctrlr_use_zcopy(req)) { 2305 SPDK_DEBUGLOG(nvmf_tcp, "Using zero-copy to execute request %p\n", tcp_req); 2306 req->data_from_pool = false; 2307 return 0; 2308 } 2309 2310 if (spdk_nvmf_request_get_buffers(req, group, transport, length)) { 2311 /* No available buffers. Queue this request up. */ 2312 SPDK_DEBUGLOG(nvmf_tcp, "No available large data buffers. Queueing request %p\n", 2313 tcp_req); 2314 return 0; 2315 } 2316 2317 /* backward compatible */ 2318 req->data = req->iov[0].iov_base; 2319 2320 SPDK_DEBUGLOG(nvmf_tcp, "Request %p took %d buffer/s from central pool, and data=%p\n", 2321 tcp_req, req->iovcnt, req->data); 2322 2323 return 0; 2324 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 2325 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 2326 uint64_t offset = sgl->address; 2327 uint32_t max_len = transport->opts.in_capsule_data_size; 2328 assert(tcp_req->has_in_capsule_data); 2329 2330 SPDK_DEBUGLOG(nvmf_tcp, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 2331 offset, length); 2332 2333 if (offset > max_len) { 2334 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 2335 offset, max_len); 2336 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 2337 return -1; 2338 } 2339 max_len -= (uint32_t)offset; 2340 2341 if (spdk_unlikely(length > max_len)) { 2342 /* According to the SPEC we should support ICD up to 8192 bytes for admin and fabric commands */ 2343 if (length <= SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE && 2344 (cmd->opc == SPDK_NVME_OPC_FABRIC || req->qpair->qid == 0)) { 2345 2346 /* Get a buffer from dedicated list */ 2347 SPDK_DEBUGLOG(nvmf_tcp, "Getting a buffer from control msg list\n"); 2348 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 2349 assert(tgroup->control_msg_list); 2350 req->data = nvmf_tcp_control_msg_get(tgroup->control_msg_list); 2351 if (!req->data) { 2352 /* No available buffers. Queue this request up. */ 2353 SPDK_DEBUGLOG(nvmf_tcp, "No available ICD buffers. Queueing request %p\n", tcp_req); 2354 return 0; 2355 } 2356 } else { 2357 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 2358 length, max_len); 2359 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 2360 return -1; 2361 } 2362 } else { 2363 req->data = tcp_req->buf; 2364 } 2365 2366 req->length = length; 2367 req->data_from_pool = false; 2368 2369 if (spdk_unlikely(req->dif_enabled)) { 2370 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2371 req->dif.elba_length = length; 2372 } 2373 2374 req->iov[0].iov_base = req->data; 2375 req->iov[0].iov_len = length; 2376 req->iovcnt = 1; 2377 2378 return 0; 2379 } 2380 2381 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 2382 sgl->generic.type, sgl->generic.subtype); 2383 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 2384 return -1; 2385 } 2386 2387 static inline enum spdk_nvme_media_error_status_code 2388 nvmf_tcp_dif_error_to_compl_status(uint8_t err_type) { 2389 enum spdk_nvme_media_error_status_code result; 2390 2391 switch (err_type) 2392 { 2393 case SPDK_DIF_REFTAG_ERROR: 2394 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 2395 break; 2396 case SPDK_DIF_APPTAG_ERROR: 2397 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 2398 break; 2399 case SPDK_DIF_GUARD_ERROR: 2400 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 2401 break; 2402 default: 2403 SPDK_UNREACHABLE(); 2404 break; 2405 } 2406 2407 return result; 2408 } 2409 2410 static void 2411 _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2412 struct spdk_nvmf_tcp_req *tcp_req) 2413 { 2414 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF( 2415 tqpair->qpair.transport, struct spdk_nvmf_tcp_transport, transport); 2416 struct nvme_tcp_pdu *rsp_pdu; 2417 struct spdk_nvme_tcp_c2h_data_hdr *c2h_data; 2418 uint32_t plen, pdo, alignment; 2419 int rc; 2420 2421 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2422 2423 rsp_pdu = tcp_req->pdu; 2424 assert(rsp_pdu != NULL); 2425 2426 c2h_data = &rsp_pdu->hdr.c2h_data; 2427 c2h_data->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_DATA; 2428 plen = c2h_data->common.hlen = sizeof(*c2h_data); 2429 2430 if (tqpair->host_hdgst_enable) { 2431 plen += SPDK_NVME_TCP_DIGEST_LEN; 2432 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2433 } 2434 2435 /* set the psh */ 2436 c2h_data->cccid = tcp_req->req.cmd->nvme_cmd.cid; 2437 c2h_data->datal = tcp_req->req.length - tcp_req->pdu->rw_offset; 2438 c2h_data->datao = tcp_req->pdu->rw_offset; 2439 2440 /* set the padding */ 2441 rsp_pdu->padding_len = 0; 2442 pdo = plen; 2443 if (tqpair->cpda) { 2444 alignment = (tqpair->cpda + 1) << 2; 2445 if (plen % alignment != 0) { 2446 pdo = (plen + alignment) / alignment * alignment; 2447 rsp_pdu->padding_len = pdo - plen; 2448 plen = pdo; 2449 } 2450 } 2451 2452 c2h_data->common.pdo = pdo; 2453 plen += c2h_data->datal; 2454 if (tqpair->host_ddgst_enable) { 2455 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_DDGSTF; 2456 plen += SPDK_NVME_TCP_DIGEST_LEN; 2457 } 2458 2459 c2h_data->common.plen = plen; 2460 2461 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2462 rsp_pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 2463 } 2464 2465 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2466 c2h_data->datao, c2h_data->datal); 2467 2468 2469 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU; 2470 /* Need to send the capsule response if response is not all 0 */ 2471 if (ttransport->tcp_opts.c2h_success && 2472 tcp_req->rsp.cdw0 == 0 && tcp_req->rsp.cdw1 == 0) { 2473 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS; 2474 } 2475 2476 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2477 struct spdk_nvme_cpl *rsp = &tcp_req->req.rsp->nvme_cpl; 2478 struct spdk_dif_error err_blk = {}; 2479 uint32_t mapped_length = 0; 2480 uint32_t available_iovs = SPDK_COUNTOF(rsp_pdu->iov); 2481 uint32_t ddgst_len = 0; 2482 2483 if (tqpair->host_ddgst_enable) { 2484 /* Data digest consumes additional iov entry */ 2485 available_iovs--; 2486 /* plen needs to be updated since nvme_tcp_build_iovs compares expected and actual plen */ 2487 ddgst_len = SPDK_NVME_TCP_DIGEST_LEN; 2488 c2h_data->common.plen -= ddgst_len; 2489 } 2490 /* Temp call to estimate if data can be described by limited number of iovs. 2491 * iov vector will be rebuilt in nvmf_tcp_qpair_write_pdu */ 2492 nvme_tcp_build_iovs(rsp_pdu->iov, available_iovs, rsp_pdu, tqpair->host_hdgst_enable, 2493 false, &mapped_length); 2494 2495 if (mapped_length != c2h_data->common.plen) { 2496 c2h_data->datal = mapped_length - (c2h_data->common.plen - c2h_data->datal); 2497 SPDK_DEBUGLOG(nvmf_tcp, 2498 "Part C2H, data_len %u (of %u), PDU len %u, updated PDU len %u, offset %u\n", 2499 c2h_data->datal, tcp_req->req.length, c2h_data->common.plen, mapped_length, rsp_pdu->rw_offset); 2500 c2h_data->common.plen = mapped_length; 2501 2502 /* Rebuild pdu->data_iov since data length is changed */ 2503 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, c2h_data->datao, 2504 c2h_data->datal); 2505 2506 c2h_data->common.flags &= ~(SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU | 2507 SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS); 2508 } 2509 2510 c2h_data->common.plen += ddgst_len; 2511 2512 assert(rsp_pdu->rw_offset <= tcp_req->req.length); 2513 2514 rc = spdk_dif_verify_stream(rsp_pdu->data_iov, rsp_pdu->data_iovcnt, 2515 0, rsp_pdu->data_len, rsp_pdu->dif_ctx, &err_blk); 2516 if (rc != 0) { 2517 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", 2518 err_blk.err_type, err_blk.err_offset); 2519 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2520 rsp->status.sc = nvmf_tcp_dif_error_to_compl_status(err_blk.err_type); 2521 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2522 return; 2523 } 2524 } 2525 2526 rsp_pdu->rw_offset += c2h_data->datal; 2527 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_pdu_c2h_data_complete, tcp_req); 2528 } 2529 2530 static void 2531 nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2532 struct spdk_nvmf_tcp_req *tcp_req) 2533 { 2534 nvmf_tcp_req_pdu_init(tcp_req); 2535 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2536 } 2537 2538 static int 2539 request_transfer_out(struct spdk_nvmf_request *req) 2540 { 2541 struct spdk_nvmf_tcp_req *tcp_req; 2542 struct spdk_nvmf_qpair *qpair; 2543 struct spdk_nvmf_tcp_qpair *tqpair; 2544 struct spdk_nvme_cpl *rsp; 2545 2546 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2547 2548 qpair = req->qpair; 2549 rsp = &req->rsp->nvme_cpl; 2550 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 2551 2552 /* Advance our sq_head pointer */ 2553 if (qpair->sq_head == qpair->sq_head_max) { 2554 qpair->sq_head = 0; 2555 } else { 2556 qpair->sq_head++; 2557 } 2558 rsp->sqhd = qpair->sq_head; 2559 2560 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2561 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2562 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2563 nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2564 } else { 2565 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2566 } 2567 2568 return 0; 2569 } 2570 2571 static void 2572 nvmf_tcp_set_in_capsule_data(struct spdk_nvmf_tcp_qpair *tqpair, 2573 struct spdk_nvmf_tcp_req *tcp_req) 2574 { 2575 struct nvme_tcp_pdu *pdu; 2576 uint32_t plen = 0; 2577 2578 pdu = tqpair->pdu_in_progress; 2579 plen = pdu->hdr.common.hlen; 2580 2581 if (tqpair->host_hdgst_enable) { 2582 plen += SPDK_NVME_TCP_DIGEST_LEN; 2583 } 2584 2585 if (pdu->hdr.common.plen != plen) { 2586 tcp_req->has_in_capsule_data = true; 2587 } 2588 } 2589 2590 static void 2591 nvmf_tcp_check_fused_ordering(struct spdk_nvmf_tcp_transport *ttransport, 2592 struct spdk_nvmf_tcp_qpair *tqpair, 2593 struct spdk_nvmf_tcp_req *tcp_req) 2594 { 2595 enum spdk_nvme_cmd_fuse last, next; 2596 2597 last = tqpair->fused_first ? tqpair->fused_first->cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2598 next = tcp_req->cmd.fuse; 2599 2600 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2601 2602 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2603 return; 2604 } 2605 2606 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2607 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2608 /* This is a valid pair of fused commands. Point them at each other 2609 * so they can be submitted consecutively once ready to be executed. 2610 */ 2611 tqpair->fused_first->fused_pair = tcp_req; 2612 tcp_req->fused_pair = tqpair->fused_first; 2613 tqpair->fused_first = NULL; 2614 return; 2615 } else { 2616 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2617 tqpair->fused_first->fused_failed = true; 2618 2619 /* 2620 * If the last req is in READY_TO_EXECUTE state, then call 2621 * nvmf_tcp_req_process(), otherwise nothing else will kick it. 2622 */ 2623 if (tqpair->fused_first->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 2624 nvmf_tcp_req_process(ttransport, tqpair->fused_first); 2625 } 2626 2627 tqpair->fused_first = NULL; 2628 } 2629 } 2630 2631 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2632 /* Set tqpair->fused_first here so that we know to check that the next request 2633 * is a SECOND (and to fail this one if it isn't). 2634 */ 2635 tqpair->fused_first = tcp_req; 2636 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2637 /* Mark this req failed since it is a SECOND and the last one was not a FIRST. */ 2638 tcp_req->fused_failed = true; 2639 } 2640 } 2641 2642 static bool 2643 nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 2644 struct spdk_nvmf_tcp_req *tcp_req) 2645 { 2646 struct spdk_nvmf_tcp_qpair *tqpair; 2647 int rc; 2648 enum spdk_nvmf_tcp_req_state prev_state; 2649 bool progress = false; 2650 struct spdk_nvmf_transport *transport = &ttransport->transport; 2651 struct spdk_nvmf_transport_poll_group *group; 2652 struct spdk_nvmf_tcp_poll_group *tgroup; 2653 2654 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2655 group = &tqpair->group->group; 2656 assert(tcp_req->state != TCP_REQUEST_STATE_FREE); 2657 2658 /* If the qpair is not active, we need to abort the outstanding requests. */ 2659 if (tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2660 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 2661 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2662 } 2663 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 2664 } 2665 2666 /* The loop here is to allow for several back-to-back state changes. */ 2667 do { 2668 prev_state = tcp_req->state; 2669 2670 SPDK_DEBUGLOG(nvmf_tcp, "Request %p entering state %d on tqpair=%p\n", tcp_req, prev_state, 2671 tqpair); 2672 2673 switch (tcp_req->state) { 2674 case TCP_REQUEST_STATE_FREE: 2675 /* Some external code must kick a request into TCP_REQUEST_STATE_NEW 2676 * to escape this state. */ 2677 break; 2678 case TCP_REQUEST_STATE_NEW: 2679 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEW, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, tqpair); 2680 2681 /* copy the cmd from the receive pdu */ 2682 tcp_req->cmd = tqpair->pdu_in_progress->hdr.capsule_cmd.ccsqe; 2683 2684 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&tcp_req->req, &tcp_req->req.dif.dif_ctx))) { 2685 tcp_req->req.dif_enabled = true; 2686 tqpair->pdu_in_progress->dif_ctx = &tcp_req->req.dif.dif_ctx; 2687 } 2688 2689 nvmf_tcp_check_fused_ordering(ttransport, tqpair, tcp_req); 2690 2691 /* The next state transition depends on the data transfer needs of this request. */ 2692 tcp_req->req.xfer = spdk_nvmf_req_get_xfer(&tcp_req->req); 2693 2694 if (spdk_unlikely(tcp_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2695 tcp_req->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 2696 tcp_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2697 tcp_req->req.rsp->nvme_cpl.cid = tcp_req->req.cmd->nvme_cmd.cid; 2698 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2699 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2700 SPDK_DEBUGLOG(nvmf_tcp, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", tcp_req); 2701 break; 2702 } 2703 2704 /* If no data to transfer, ready to execute. */ 2705 if (tcp_req->req.xfer == SPDK_NVME_DATA_NONE) { 2706 /* Reset the tqpair receiving pdu state */ 2707 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2708 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2709 break; 2710 } 2711 2712 nvmf_tcp_set_in_capsule_data(tqpair, tcp_req); 2713 2714 if (!tcp_req->has_in_capsule_data) { 2715 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2716 } 2717 2718 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEED_BUFFER); 2719 STAILQ_INSERT_TAIL(&group->pending_buf_queue, &tcp_req->req, buf_link); 2720 break; 2721 case TCP_REQUEST_STATE_NEED_BUFFER: 2722 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEED_BUFFER, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2723 tqpair); 2724 2725 assert(tcp_req->req.xfer != SPDK_NVME_DATA_NONE); 2726 2727 if (!tcp_req->has_in_capsule_data && (&tcp_req->req != STAILQ_FIRST(&group->pending_buf_queue))) { 2728 SPDK_DEBUGLOG(nvmf_tcp, 2729 "Not the first element to wait for the buf for tcp_req(%p) on tqpair=%p\n", 2730 tcp_req, tqpair); 2731 /* This request needs to wait in line to obtain a buffer */ 2732 break; 2733 } 2734 2735 /* Try to get a data buffer */ 2736 rc = nvmf_tcp_req_parse_sgl(tcp_req, transport, group); 2737 if (rc < 0) { 2738 STAILQ_REMOVE_HEAD(&group->pending_buf_queue, buf_link); 2739 /* Reset the tqpair receiving pdu state */ 2740 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 2741 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2742 tcp_req->req.rsp->nvme_cpl.cid = tcp_req->req.cmd->nvme_cmd.cid; 2743 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2744 break; 2745 } 2746 2747 /* Get a zcopy buffer if the request can be serviced through zcopy */ 2748 if (spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 2749 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2750 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 2751 tcp_req->req.length = tcp_req->req.dif.elba_length; 2752 } 2753 2754 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2755 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_START); 2756 spdk_nvmf_request_zcopy_start(&tcp_req->req); 2757 break; 2758 } 2759 2760 if (!tcp_req->req.data) { 2761 SPDK_DEBUGLOG(nvmf_tcp, "No buffer allocated for tcp_req(%p) on tqpair(%p\n)", 2762 tcp_req, tqpair); 2763 /* No buffers available. */ 2764 break; 2765 } 2766 2767 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2768 2769 /* If data is transferring from host to controller, we need to do a transfer from the host. */ 2770 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2771 if (tcp_req->req.data_from_pool) { 2772 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 2773 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 2774 } else { 2775 struct nvme_tcp_pdu *pdu; 2776 2777 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2778 2779 pdu = tqpair->pdu_in_progress; 2780 SPDK_DEBUGLOG(nvmf_tcp, "Not need to send r2t for tcp_req(%p) on tqpair=%p\n", tcp_req, 2781 tqpair); 2782 /* No need to send r2t, contained in the capsuled data */ 2783 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2784 0, tcp_req->req.length); 2785 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2786 } 2787 break; 2788 } 2789 2790 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2791 break; 2792 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 2793 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, tqpair->qpair.qid, 0, 2794 (uintptr_t)tcp_req, tqpair); 2795 /* Some external code must kick a request into TCP_REQUEST_STATE_ZCOPY_START_COMPLETED 2796 * to escape this state. */ 2797 break; 2798 case TCP_REQUEST_STATE_ZCOPY_START_COMPLETED: 2799 spdk_trace_record(TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, tqpair->qpair.qid, 0, 2800 (uintptr_t)tcp_req, tqpair); 2801 if (spdk_unlikely(spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl))) { 2802 SPDK_DEBUGLOG(nvmf_tcp, "Zero-copy start failed for tcp_req(%p) on tqpair=%p\n", 2803 tcp_req, tqpair); 2804 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2805 break; 2806 } 2807 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2808 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 2809 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 2810 } else { 2811 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 2812 } 2813 break; 2814 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 2815 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2816 tqpair); 2817 /* The R2T completion or the h2c data incoming will kick it out of this state. */ 2818 break; 2819 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2820 2821 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, tqpair->qpair.qid, 0, 2822 (uintptr_t)tcp_req, tqpair); 2823 /* Some external code must kick a request into TCP_REQUEST_STATE_READY_TO_EXECUTE 2824 * to escape this state. */ 2825 break; 2826 case TCP_REQUEST_STATE_READY_TO_EXECUTE: 2827 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, tqpair->qpair.qid, 0, 2828 (uintptr_t)tcp_req, tqpair); 2829 2830 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2831 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 2832 tcp_req->req.length = tcp_req->req.dif.elba_length; 2833 } 2834 2835 if (tcp_req->cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 2836 if (tcp_req->fused_failed) { 2837 /* This request failed FUSED semantics. Fail it immediately, without 2838 * even sending it to the target layer. 2839 */ 2840 tcp_req->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 2841 tcp_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED; 2842 tcp_req->req.rsp->nvme_cpl.cid = tcp_req->req.cmd->nvme_cmd.cid; 2843 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2844 break; 2845 } 2846 2847 if (tcp_req->fused_pair == NULL || 2848 tcp_req->fused_pair->state != TCP_REQUEST_STATE_READY_TO_EXECUTE) { 2849 /* This request is ready to execute, but either we don't know yet if it's 2850 * valid - i.e. this is a FIRST but we haven't received the next request yet), 2851 * or the other request of this fused pair isn't ready to execute. So 2852 * break here and this request will get processed later either when the 2853 * other request is ready or we find that this request isn't valid. 2854 */ 2855 break; 2856 } 2857 } 2858 2859 if (!spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 2860 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTING); 2861 /* If we get to this point, and this request is a fused command, we know that 2862 * it is part of a valid sequence (FIRST followed by a SECOND) and that both 2863 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 2864 * request, and the other request of the fused pair, in the correct order. 2865 * Also clear the ->fused_pair pointers on both requests, since after this point 2866 * we no longer need to maintain the relationship between these two requests. 2867 */ 2868 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 2869 assert(tcp_req->fused_pair != NULL); 2870 assert(tcp_req->fused_pair->fused_pair == tcp_req); 2871 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 2872 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 2873 tcp_req->fused_pair->fused_pair = NULL; 2874 tcp_req->fused_pair = NULL; 2875 } 2876 spdk_nvmf_request_exec(&tcp_req->req); 2877 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 2878 assert(tcp_req->fused_pair != NULL); 2879 assert(tcp_req->fused_pair->fused_pair == tcp_req); 2880 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 2881 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 2882 tcp_req->fused_pair->fused_pair = NULL; 2883 tcp_req->fused_pair = NULL; 2884 } 2885 } else { 2886 /* For zero-copy, only requests with data coming from host to the 2887 * controller can end up here. */ 2888 assert(tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 2889 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT); 2890 spdk_nvmf_request_zcopy_end(&tcp_req->req, true); 2891 } 2892 2893 break; 2894 case TCP_REQUEST_STATE_EXECUTING: 2895 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTING, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2896 tqpair); 2897 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 2898 * to escape this state. */ 2899 break; 2900 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 2901 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, tqpair->qpair.qid, 0, 2902 (uintptr_t)tcp_req, tqpair); 2903 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 2904 * to escape this state. */ 2905 break; 2906 case TCP_REQUEST_STATE_EXECUTED: 2907 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTED, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2908 tqpair); 2909 2910 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2911 tcp_req->req.length = tcp_req->req.dif.orig_length; 2912 } 2913 2914 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2915 break; 2916 case TCP_REQUEST_STATE_READY_TO_COMPLETE: 2917 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, tqpair->qpair.qid, 0, 2918 (uintptr_t)tcp_req, tqpair); 2919 rc = request_transfer_out(&tcp_req->req); 2920 assert(rc == 0); /* No good way to handle this currently */ 2921 break; 2922 case TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2923 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, tqpair->qpair.qid, 0, 2924 (uintptr_t)tcp_req, tqpair); 2925 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 2926 * to escape this state. */ 2927 break; 2928 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 2929 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, tqpair->qpair.qid, 0, 2930 (uintptr_t)tcp_req, tqpair); 2931 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 2932 * to escape this state. */ 2933 break; 2934 case TCP_REQUEST_STATE_COMPLETED: 2935 spdk_trace_record(TRACE_TCP_REQUEST_STATE_COMPLETED, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2936 tqpair); 2937 /* If there's an outstanding PDU sent to the host, the request is completed 2938 * due to the qpair being disconnected. We must delay the completion until 2939 * that write is done to avoid freeing the request twice. */ 2940 if (spdk_unlikely(tcp_req->pdu_in_use)) { 2941 SPDK_DEBUGLOG(nvmf_tcp, "Delaying completion due to outstanding " 2942 "write on req=%p\n", tcp_req); 2943 /* This can only happen for zcopy requests */ 2944 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 2945 assert(tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE); 2946 break; 2947 } 2948 2949 if (tcp_req->req.data_from_pool) { 2950 spdk_nvmf_request_free_buffers(&tcp_req->req, group, transport); 2951 } else if (spdk_unlikely(tcp_req->has_in_capsule_data && 2952 (tcp_req->cmd.opc == SPDK_NVME_OPC_FABRIC || 2953 tqpair->qpair.qid == 0) && tcp_req->req.length > transport->opts.in_capsule_data_size)) { 2954 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 2955 assert(tgroup->control_msg_list); 2956 SPDK_DEBUGLOG(nvmf_tcp, "Put buf to control msg list\n"); 2957 nvmf_tcp_control_msg_put(tgroup->control_msg_list, tcp_req->req.data); 2958 } else if (tcp_req->req.zcopy_bdev_io != NULL) { 2959 /* If the request has an unreleased zcopy bdev_io, it's either a 2960 * read, a failed write, or the qpair is being disconnected */ 2961 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 2962 assert(tcp_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST || 2963 spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl) || 2964 tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE); 2965 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE); 2966 spdk_nvmf_request_zcopy_end(&tcp_req->req, false); 2967 break; 2968 } 2969 tcp_req->req.length = 0; 2970 tcp_req->req.iovcnt = 0; 2971 tcp_req->req.data = NULL; 2972 tcp_req->fused_failed = false; 2973 if (tcp_req->fused_pair) { 2974 /* This req was part of a valid fused pair, but failed before it got to 2975 * READ_TO_EXECUTE state. This means we need to fail the other request 2976 * in the pair, because it is no longer part of a valid pair. If the pair 2977 * already reached READY_TO_EXECUTE state, we need to kick it. 2978 */ 2979 tcp_req->fused_pair->fused_failed = true; 2980 if (tcp_req->fused_pair->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 2981 nvmf_tcp_req_process(ttransport, tcp_req->fused_pair); 2982 } 2983 tcp_req->fused_pair = NULL; 2984 } 2985 2986 nvmf_tcp_req_put(tqpair, tcp_req); 2987 break; 2988 case TCP_REQUEST_NUM_STATES: 2989 default: 2990 assert(0); 2991 break; 2992 } 2993 2994 if (tcp_req->state != prev_state) { 2995 progress = true; 2996 } 2997 } while (tcp_req->state != prev_state); 2998 2999 return progress; 3000 } 3001 3002 static void 3003 nvmf_tcp_sock_cb(void *arg, struct spdk_sock_group *group, struct spdk_sock *sock) 3004 { 3005 struct spdk_nvmf_tcp_qpair *tqpair = arg; 3006 int rc; 3007 3008 assert(tqpair != NULL); 3009 rc = nvmf_tcp_sock_process(tqpair); 3010 3011 /* If there was a new socket error, disconnect */ 3012 if (rc < 0) { 3013 nvmf_tcp_qpair_disconnect(tqpair); 3014 } 3015 } 3016 3017 static int 3018 nvmf_tcp_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3019 struct spdk_nvmf_qpair *qpair) 3020 { 3021 struct spdk_nvmf_tcp_poll_group *tgroup; 3022 struct spdk_nvmf_tcp_qpair *tqpair; 3023 int rc; 3024 3025 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3026 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3027 3028 rc = nvmf_tcp_qpair_sock_init(tqpair); 3029 if (rc != 0) { 3030 SPDK_ERRLOG("Cannot set sock opt for tqpair=%p\n", tqpair); 3031 return -1; 3032 } 3033 3034 rc = nvmf_tcp_qpair_init(&tqpair->qpair); 3035 if (rc < 0) { 3036 SPDK_ERRLOG("Cannot init tqpair=%p\n", tqpair); 3037 return -1; 3038 } 3039 3040 rc = nvmf_tcp_qpair_init_mem_resource(tqpair); 3041 if (rc < 0) { 3042 SPDK_ERRLOG("Cannot init memory resource info for tqpair=%p\n", tqpair); 3043 return -1; 3044 } 3045 3046 rc = spdk_sock_group_add_sock(tgroup->sock_group, tqpair->sock, 3047 nvmf_tcp_sock_cb, tqpair); 3048 if (rc != 0) { 3049 SPDK_ERRLOG("Could not add sock to sock_group: %s (%d)\n", 3050 spdk_strerror(errno), errno); 3051 return -1; 3052 } 3053 3054 tqpair->group = tgroup; 3055 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INVALID); 3056 TAILQ_INSERT_TAIL(&tgroup->qpairs, tqpair, link); 3057 3058 return 0; 3059 } 3060 3061 static int 3062 nvmf_tcp_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3063 struct spdk_nvmf_qpair *qpair) 3064 { 3065 struct spdk_nvmf_tcp_poll_group *tgroup; 3066 struct spdk_nvmf_tcp_qpair *tqpair; 3067 int rc; 3068 3069 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3070 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3071 3072 assert(tqpair->group == tgroup); 3073 3074 SPDK_DEBUGLOG(nvmf_tcp, "remove tqpair=%p from the tgroup=%p\n", tqpair, tgroup); 3075 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 3076 TAILQ_REMOVE(&tgroup->await_req, tqpair, link); 3077 } else { 3078 TAILQ_REMOVE(&tgroup->qpairs, tqpair, link); 3079 } 3080 3081 rc = spdk_sock_group_remove_sock(tgroup->sock_group, tqpair->sock); 3082 if (rc != 0) { 3083 SPDK_ERRLOG("Could not remove sock from sock_group: %s (%d)\n", 3084 spdk_strerror(errno), errno); 3085 } 3086 3087 return rc; 3088 } 3089 3090 static int 3091 nvmf_tcp_req_complete(struct spdk_nvmf_request *req) 3092 { 3093 struct spdk_nvmf_tcp_transport *ttransport; 3094 struct spdk_nvmf_tcp_req *tcp_req; 3095 3096 ttransport = SPDK_CONTAINEROF(req->qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3097 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3098 3099 switch (tcp_req->state) { 3100 case TCP_REQUEST_STATE_EXECUTING: 3101 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3102 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3103 break; 3104 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3105 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_ZCOPY_START_COMPLETED); 3106 break; 3107 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3108 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 3109 break; 3110 default: 3111 assert(0 && "Unexpected request state"); 3112 break; 3113 } 3114 3115 nvmf_tcp_req_process(ttransport, tcp_req); 3116 3117 return 0; 3118 } 3119 3120 static void 3121 nvmf_tcp_close_qpair(struct spdk_nvmf_qpair *qpair, 3122 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3123 { 3124 struct spdk_nvmf_tcp_qpair *tqpair; 3125 3126 SPDK_DEBUGLOG(nvmf_tcp, "Qpair: %p\n", qpair); 3127 3128 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3129 3130 assert(tqpair->fini_cb_fn == NULL); 3131 tqpair->fini_cb_fn = cb_fn; 3132 tqpair->fini_cb_arg = cb_arg; 3133 3134 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITED); 3135 nvmf_tcp_qpair_destroy(tqpair); 3136 } 3137 3138 static int 3139 nvmf_tcp_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3140 { 3141 struct spdk_nvmf_tcp_poll_group *tgroup; 3142 int rc; 3143 struct spdk_nvmf_request *req, *req_tmp; 3144 struct spdk_nvmf_tcp_req *tcp_req; 3145 struct spdk_nvmf_tcp_qpair *tqpair, *tqpair_tmp; 3146 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(group->transport, 3147 struct spdk_nvmf_tcp_transport, transport); 3148 3149 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3150 3151 if (spdk_unlikely(TAILQ_EMPTY(&tgroup->qpairs) && TAILQ_EMPTY(&tgroup->await_req))) { 3152 return 0; 3153 } 3154 3155 STAILQ_FOREACH_SAFE(req, &group->pending_buf_queue, buf_link, req_tmp) { 3156 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3157 if (nvmf_tcp_req_process(ttransport, tcp_req) == false) { 3158 break; 3159 } 3160 } 3161 3162 rc = spdk_sock_group_poll(tgroup->sock_group); 3163 if (rc < 0) { 3164 SPDK_ERRLOG("Failed to poll sock_group=%p\n", tgroup->sock_group); 3165 } 3166 3167 TAILQ_FOREACH_SAFE(tqpair, &tgroup->await_req, link, tqpair_tmp) { 3168 nvmf_tcp_sock_process(tqpair); 3169 } 3170 3171 return rc; 3172 } 3173 3174 static int 3175 nvmf_tcp_qpair_get_trid(struct spdk_nvmf_qpair *qpair, 3176 struct spdk_nvme_transport_id *trid, bool peer) 3177 { 3178 struct spdk_nvmf_tcp_qpair *tqpair; 3179 uint16_t port; 3180 3181 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3182 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_TCP); 3183 3184 if (peer) { 3185 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->initiator_addr); 3186 port = tqpair->initiator_port; 3187 } else { 3188 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->target_addr); 3189 port = tqpair->target_port; 3190 } 3191 3192 if (spdk_sock_is_ipv4(tqpair->sock)) { 3193 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3194 } else if (spdk_sock_is_ipv6(tqpair->sock)) { 3195 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3196 } else { 3197 return -1; 3198 } 3199 3200 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%d", port); 3201 return 0; 3202 } 3203 3204 static int 3205 nvmf_tcp_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3206 struct spdk_nvme_transport_id *trid) 3207 { 3208 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3209 } 3210 3211 static int 3212 nvmf_tcp_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3213 struct spdk_nvme_transport_id *trid) 3214 { 3215 return nvmf_tcp_qpair_get_trid(qpair, trid, 1); 3216 } 3217 3218 static int 3219 nvmf_tcp_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3220 struct spdk_nvme_transport_id *trid) 3221 { 3222 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3223 } 3224 3225 static void 3226 nvmf_tcp_req_set_abort_status(struct spdk_nvmf_request *req, 3227 struct spdk_nvmf_tcp_req *tcp_req_to_abort) 3228 { 3229 tcp_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 3230 tcp_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 3231 tcp_req_to_abort->req.rsp->nvme_cpl.cid = tcp_req_to_abort->req.cmd->nvme_cmd.cid; 3232 3233 nvmf_tcp_req_set_state(tcp_req_to_abort, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3234 3235 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 3236 } 3237 3238 static int 3239 _nvmf_tcp_qpair_abort_request(void *ctx) 3240 { 3241 struct spdk_nvmf_request *req = ctx; 3242 struct spdk_nvmf_tcp_req *tcp_req_to_abort = SPDK_CONTAINEROF(req->req_to_abort, 3243 struct spdk_nvmf_tcp_req, req); 3244 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 3245 struct spdk_nvmf_tcp_qpair, qpair); 3246 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 3247 struct spdk_nvmf_tcp_transport, transport); 3248 int rc; 3249 3250 spdk_poller_unregister(&req->poller); 3251 3252 switch (tcp_req_to_abort->state) { 3253 case TCP_REQUEST_STATE_EXECUTING: 3254 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3255 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3256 rc = nvmf_ctrlr_abort_request(req); 3257 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 3258 return SPDK_POLLER_BUSY; 3259 } 3260 break; 3261 3262 case TCP_REQUEST_STATE_NEED_BUFFER: 3263 STAILQ_REMOVE(&tqpair->group->group.pending_buf_queue, 3264 &tcp_req_to_abort->req, spdk_nvmf_request, buf_link); 3265 3266 nvmf_tcp_req_set_abort_status(req, tcp_req_to_abort); 3267 nvmf_tcp_req_process(ttransport, tcp_req_to_abort); 3268 break; 3269 3270 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3271 case TCP_REQUEST_STATE_ZCOPY_START_COMPLETED: 3272 nvmf_tcp_req_set_abort_status(req, tcp_req_to_abort); 3273 break; 3274 3275 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3276 if (spdk_get_ticks() < req->timeout_tsc) { 3277 req->poller = SPDK_POLLER_REGISTER(_nvmf_tcp_qpair_abort_request, req, 0); 3278 return SPDK_POLLER_BUSY; 3279 } 3280 break; 3281 3282 default: 3283 break; 3284 } 3285 3286 spdk_nvmf_request_complete(req); 3287 return SPDK_POLLER_BUSY; 3288 } 3289 3290 static void 3291 nvmf_tcp_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 3292 struct spdk_nvmf_request *req) 3293 { 3294 struct spdk_nvmf_tcp_qpair *tqpair; 3295 struct spdk_nvmf_tcp_transport *ttransport; 3296 struct spdk_nvmf_transport *transport; 3297 uint16_t cid; 3298 uint32_t i; 3299 struct spdk_nvmf_tcp_req *tcp_req_to_abort = NULL; 3300 3301 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3302 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3303 transport = &ttransport->transport; 3304 3305 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 3306 3307 for (i = 0; i < tqpair->resource_count; i++) { 3308 if (tqpair->reqs[i].state != TCP_REQUEST_STATE_FREE && 3309 tqpair->reqs[i].req.cmd->nvme_cmd.cid == cid) { 3310 tcp_req_to_abort = &tqpair->reqs[i]; 3311 break; 3312 } 3313 } 3314 3315 spdk_trace_record(TRACE_TCP_QP_ABORT_REQ, qpair->qid, 0, (uintptr_t)req, tqpair); 3316 3317 if (tcp_req_to_abort == NULL) { 3318 spdk_nvmf_request_complete(req); 3319 return; 3320 } 3321 3322 req->req_to_abort = &tcp_req_to_abort->req; 3323 req->timeout_tsc = spdk_get_ticks() + 3324 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 3325 req->poller = NULL; 3326 3327 _nvmf_tcp_qpair_abort_request(req); 3328 } 3329 3330 #define SPDK_NVMF_TCP_DEFAULT_MAX_QUEUE_DEPTH 128 3331 #define SPDK_NVMF_TCP_DEFAULT_AQ_DEPTH 128 3332 #define SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 3333 #define SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 3334 #define SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE 131072 3335 #define SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE 131072 3336 #define SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS 511 3337 #define SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE 32 3338 #define SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP false 3339 #define SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC 1 3340 3341 static void 3342 nvmf_tcp_opts_init(struct spdk_nvmf_transport_opts *opts) 3343 { 3344 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_QUEUE_DEPTH; 3345 opts->max_qpairs_per_ctrlr = SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR; 3346 opts->in_capsule_data_size = SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE; 3347 opts->max_io_size = SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE; 3348 opts->io_unit_size = SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE; 3349 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_AQ_DEPTH; 3350 opts->num_shared_buffers = SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS; 3351 opts->buf_cache_size = SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE; 3352 opts->dif_insert_or_strip = SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP; 3353 opts->abort_timeout_sec = SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC; 3354 opts->transport_specific = NULL; 3355 } 3356 3357 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp = { 3358 .name = "TCP", 3359 .type = SPDK_NVME_TRANSPORT_TCP, 3360 .opts_init = nvmf_tcp_opts_init, 3361 .create = nvmf_tcp_create, 3362 .dump_opts = nvmf_tcp_dump_opts, 3363 .destroy = nvmf_tcp_destroy, 3364 3365 .listen = nvmf_tcp_listen, 3366 .stop_listen = nvmf_tcp_stop_listen, 3367 3368 .listener_discover = nvmf_tcp_discover, 3369 3370 .poll_group_create = nvmf_tcp_poll_group_create, 3371 .get_optimal_poll_group = nvmf_tcp_get_optimal_poll_group, 3372 .poll_group_destroy = nvmf_tcp_poll_group_destroy, 3373 .poll_group_add = nvmf_tcp_poll_group_add, 3374 .poll_group_remove = nvmf_tcp_poll_group_remove, 3375 .poll_group_poll = nvmf_tcp_poll_group_poll, 3376 3377 .req_free = nvmf_tcp_req_free, 3378 .req_complete = nvmf_tcp_req_complete, 3379 3380 .qpair_fini = nvmf_tcp_close_qpair, 3381 .qpair_get_local_trid = nvmf_tcp_qpair_get_local_trid, 3382 .qpair_get_peer_trid = nvmf_tcp_qpair_get_peer_trid, 3383 .qpair_get_listen_trid = nvmf_tcp_qpair_get_listen_trid, 3384 .qpair_abort_request = nvmf_tcp_qpair_abort_request, 3385 }; 3386 3387 SPDK_NVMF_TRANSPORT_REGISTER(tcp, &spdk_nvmf_transport_tcp); 3388 SPDK_LOG_REGISTER_COMPONENT(nvmf_tcp) 3389