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_engine.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; 249 250 /* Number of requests in each state */ 251 uint32_t state_cntr[TCP_REQUEST_NUM_STATES]; 252 253 uint8_t cpda; 254 255 bool host_hdgst_enable; 256 bool host_ddgst_enable; 257 258 /* This is a spare PDU used for sending special management 259 * operations. Primarily, this is used for the initial 260 * connection response and c2h termination request. */ 261 struct nvme_tcp_pdu *mgmt_pdu; 262 263 /* Arrays of in-capsule buffers, requests, and pdus. 264 * Each array is 'resource_count' number of elements */ 265 void *bufs; 266 struct spdk_nvmf_tcp_req *reqs; 267 struct nvme_tcp_pdu *pdus; 268 uint32_t resource_count; 269 uint32_t recv_buf_size; 270 271 struct spdk_nvmf_tcp_port *port; 272 273 /* IP address */ 274 char initiator_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 275 char target_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 276 277 /* IP port */ 278 uint16_t initiator_port; 279 uint16_t target_port; 280 281 /* Timer used to destroy qpair after detecting transport error issue if initiator does 282 * not close the connection. 283 */ 284 struct spdk_poller *timeout_poller; 285 286 spdk_nvmf_transport_qpair_fini_cb fini_cb_fn; 287 void *fini_cb_arg; 288 289 TAILQ_ENTRY(spdk_nvmf_tcp_qpair) link; 290 }; 291 292 struct spdk_nvmf_tcp_control_msg { 293 STAILQ_ENTRY(spdk_nvmf_tcp_control_msg) link; 294 }; 295 296 struct spdk_nvmf_tcp_control_msg_list { 297 void *msg_buf; 298 STAILQ_HEAD(, spdk_nvmf_tcp_control_msg) free_msgs; 299 }; 300 301 struct spdk_nvmf_tcp_poll_group { 302 struct spdk_nvmf_transport_poll_group group; 303 struct spdk_sock_group *sock_group; 304 305 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) qpairs; 306 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) await_req; 307 308 struct spdk_io_channel *accel_channel; 309 struct spdk_nvmf_tcp_control_msg_list *control_msg_list; 310 311 TAILQ_ENTRY(spdk_nvmf_tcp_poll_group) link; 312 }; 313 314 struct spdk_nvmf_tcp_port { 315 const struct spdk_nvme_transport_id *trid; 316 struct spdk_sock *listen_sock; 317 TAILQ_ENTRY(spdk_nvmf_tcp_port) link; 318 }; 319 320 struct tcp_transport_opts { 321 bool c2h_success; 322 uint16_t control_msg_num; 323 uint32_t sock_priority; 324 }; 325 326 struct spdk_nvmf_tcp_transport { 327 struct spdk_nvmf_transport transport; 328 struct tcp_transport_opts tcp_opts; 329 330 struct spdk_nvmf_tcp_poll_group *next_pg; 331 332 struct spdk_poller *accept_poller; 333 pthread_mutex_t lock; 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 (!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 pthread_mutex_destroy(&ttransport->lock); 575 free(ttransport); 576 577 if (cb_fn) { 578 cb_fn(cb_arg); 579 } 580 return 0; 581 } 582 583 static int nvmf_tcp_accept(void *ctx); 584 585 static struct spdk_nvmf_transport * 586 nvmf_tcp_create(struct spdk_nvmf_transport_opts *opts) 587 { 588 struct spdk_nvmf_tcp_transport *ttransport; 589 uint32_t sge_count; 590 uint32_t min_shared_buffers; 591 592 ttransport = calloc(1, sizeof(*ttransport)); 593 if (!ttransport) { 594 return NULL; 595 } 596 597 TAILQ_INIT(&ttransport->ports); 598 TAILQ_INIT(&ttransport->poll_groups); 599 600 ttransport->transport.ops = &spdk_nvmf_transport_tcp; 601 602 ttransport->tcp_opts.c2h_success = SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION; 603 ttransport->tcp_opts.sock_priority = SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY; 604 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 605 if (opts->transport_specific != NULL && 606 spdk_json_decode_object_relaxed(opts->transport_specific, tcp_transport_opts_decoder, 607 SPDK_COUNTOF(tcp_transport_opts_decoder), 608 &ttransport->tcp_opts)) { 609 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 610 free(ttransport); 611 return NULL; 612 } 613 614 SPDK_NOTICELOG("*** TCP Transport Init ***\n"); 615 616 SPDK_INFOLOG(nvmf_tcp, "*** TCP Transport Init ***\n" 617 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 618 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 619 " in_capsule_data_size=%d, max_aq_depth=%d\n" 620 " num_shared_buffers=%d, c2h_success=%d,\n" 621 " dif_insert_or_strip=%d, sock_priority=%d\n" 622 " abort_timeout_sec=%d, control_msg_num=%hu\n", 623 opts->max_queue_depth, 624 opts->max_io_size, 625 opts->max_qpairs_per_ctrlr - 1, 626 opts->io_unit_size, 627 opts->in_capsule_data_size, 628 opts->max_aq_depth, 629 opts->num_shared_buffers, 630 ttransport->tcp_opts.c2h_success, 631 opts->dif_insert_or_strip, 632 ttransport->tcp_opts.sock_priority, 633 opts->abort_timeout_sec, 634 ttransport->tcp_opts.control_msg_num); 635 636 if (ttransport->tcp_opts.sock_priority > SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY) { 637 SPDK_ERRLOG("Unsupported socket_priority=%d, the current range is: 0 to %d\n" 638 "you can use man 7 socket to view the range of priority under SO_PRIORITY item\n", 639 ttransport->tcp_opts.sock_priority, SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY); 640 free(ttransport); 641 return NULL; 642 } 643 644 if (ttransport->tcp_opts.control_msg_num == 0 && 645 opts->in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 646 SPDK_WARNLOG("TCP param control_msg_num can't be 0 if ICD is less than %u bytes. Using default value %u\n", 647 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM); 648 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 649 } 650 651 /* I/O unit size cannot be larger than max I/O size */ 652 if (opts->io_unit_size > opts->max_io_size) { 653 opts->io_unit_size = opts->max_io_size; 654 } 655 656 sge_count = opts->max_io_size / opts->io_unit_size; 657 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 658 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 659 free(ttransport); 660 return NULL; 661 } 662 663 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 664 if (min_shared_buffers > opts->num_shared_buffers) { 665 SPDK_ERRLOG("There are not enough buffers to satisfy " 666 "per-poll group caches for each thread. (%" PRIu32 ") " 667 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 668 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 669 free(ttransport); 670 return NULL; 671 } 672 673 pthread_mutex_init(&ttransport->lock, NULL); 674 675 ttransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_tcp_accept, &ttransport->transport, 676 opts->acceptor_poll_rate); 677 if (!ttransport->accept_poller) { 678 pthread_mutex_destroy(&ttransport->lock); 679 free(ttransport); 680 return NULL; 681 } 682 683 return &ttransport->transport; 684 } 685 686 static int 687 nvmf_tcp_trsvcid_to_int(const char *trsvcid) 688 { 689 unsigned long long ull; 690 char *end = NULL; 691 692 ull = strtoull(trsvcid, &end, 10); 693 if (end == NULL || end == trsvcid || *end != '\0') { 694 return -1; 695 } 696 697 /* Valid TCP/IP port numbers are in [0, 65535] */ 698 if (ull > 65535) { 699 return -1; 700 } 701 702 return (int)ull; 703 } 704 705 /** 706 * Canonicalize a listen address trid. 707 */ 708 static int 709 nvmf_tcp_canon_listen_trid(struct spdk_nvme_transport_id *canon_trid, 710 const struct spdk_nvme_transport_id *trid) 711 { 712 int trsvcid_int; 713 714 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 715 if (trsvcid_int < 0) { 716 return -EINVAL; 717 } 718 719 memset(canon_trid, 0, sizeof(*canon_trid)); 720 spdk_nvme_trid_populate_transport(canon_trid, SPDK_NVME_TRANSPORT_TCP); 721 canon_trid->adrfam = trid->adrfam; 722 snprintf(canon_trid->traddr, sizeof(canon_trid->traddr), "%s", trid->traddr); 723 snprintf(canon_trid->trsvcid, sizeof(canon_trid->trsvcid), "%d", trsvcid_int); 724 725 return 0; 726 } 727 728 /** 729 * Find an existing listening port. 730 * 731 * Caller must hold ttransport->lock. 732 */ 733 static struct spdk_nvmf_tcp_port * 734 nvmf_tcp_find_port(struct spdk_nvmf_tcp_transport *ttransport, 735 const struct spdk_nvme_transport_id *trid) 736 { 737 struct spdk_nvme_transport_id canon_trid; 738 struct spdk_nvmf_tcp_port *port; 739 740 if (nvmf_tcp_canon_listen_trid(&canon_trid, trid) != 0) { 741 return NULL; 742 } 743 744 TAILQ_FOREACH(port, &ttransport->ports, link) { 745 if (spdk_nvme_transport_id_compare(&canon_trid, port->trid) == 0) { 746 return port; 747 } 748 } 749 750 return NULL; 751 } 752 753 static int 754 nvmf_tcp_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 755 struct spdk_nvmf_listen_opts *listen_opts) 756 { 757 struct spdk_nvmf_tcp_transport *ttransport; 758 struct spdk_nvmf_tcp_port *port; 759 int trsvcid_int; 760 uint8_t adrfam; 761 struct spdk_sock_opts opts; 762 763 if (!strlen(trid->trsvcid)) { 764 SPDK_ERRLOG("Service id is required\n"); 765 return -EINVAL; 766 } 767 768 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 769 770 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 771 if (trsvcid_int < 0) { 772 SPDK_ERRLOG("Invalid trsvcid '%s'\n", trid->trsvcid); 773 return -EINVAL; 774 } 775 776 pthread_mutex_lock(&ttransport->lock); 777 port = calloc(1, sizeof(*port)); 778 if (!port) { 779 SPDK_ERRLOG("Port allocation failed\n"); 780 pthread_mutex_unlock(&ttransport->lock); 781 return -ENOMEM; 782 } 783 784 port->trid = trid; 785 opts.opts_size = sizeof(opts); 786 spdk_sock_get_default_opts(&opts); 787 opts.priority = ttransport->tcp_opts.sock_priority; 788 port->listen_sock = spdk_sock_listen_ext(trid->traddr, trsvcid_int, 789 NULL, &opts); 790 if (port->listen_sock == NULL) { 791 SPDK_ERRLOG("spdk_sock_listen(%s, %d) failed: %s (%d)\n", 792 trid->traddr, trsvcid_int, 793 spdk_strerror(errno), errno); 794 free(port); 795 pthread_mutex_unlock(&ttransport->lock); 796 return -errno; 797 } 798 799 if (spdk_sock_is_ipv4(port->listen_sock)) { 800 adrfam = SPDK_NVMF_ADRFAM_IPV4; 801 } else if (spdk_sock_is_ipv6(port->listen_sock)) { 802 adrfam = SPDK_NVMF_ADRFAM_IPV6; 803 } else { 804 SPDK_ERRLOG("Unhandled socket type\n"); 805 adrfam = 0; 806 } 807 808 if (adrfam != trid->adrfam) { 809 SPDK_ERRLOG("Socket address family mismatch\n"); 810 spdk_sock_close(&port->listen_sock); 811 free(port); 812 pthread_mutex_unlock(&ttransport->lock); 813 return -EINVAL; 814 } 815 816 SPDK_NOTICELOG("*** NVMe/TCP Target Listening on %s port %s ***\n", 817 trid->traddr, trid->trsvcid); 818 819 TAILQ_INSERT_TAIL(&ttransport->ports, port, link); 820 pthread_mutex_unlock(&ttransport->lock); 821 return 0; 822 } 823 824 static void 825 nvmf_tcp_stop_listen(struct spdk_nvmf_transport *transport, 826 const struct spdk_nvme_transport_id *trid) 827 { 828 struct spdk_nvmf_tcp_transport *ttransport; 829 struct spdk_nvmf_tcp_port *port; 830 831 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 832 833 SPDK_DEBUGLOG(nvmf_tcp, "Removing listen address %s port %s\n", 834 trid->traddr, trid->trsvcid); 835 836 pthread_mutex_lock(&ttransport->lock); 837 port = nvmf_tcp_find_port(ttransport, trid); 838 if (port) { 839 TAILQ_REMOVE(&ttransport->ports, port, link); 840 spdk_sock_close(&port->listen_sock); 841 free(port); 842 } 843 844 pthread_mutex_unlock(&ttransport->lock); 845 } 846 847 static void nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 848 enum nvme_tcp_pdu_recv_state state); 849 850 static void 851 nvmf_tcp_qpair_set_state(struct spdk_nvmf_tcp_qpair *tqpair, enum nvme_tcp_qpair_state state) 852 { 853 tqpair->state = state; 854 spdk_trace_record(TRACE_TCP_QP_STATE_CHANGE, tqpair->qpair.qid, 0, (uintptr_t)tqpair, 855 tqpair->state); 856 } 857 858 static void 859 nvmf_tcp_qpair_disconnect(struct spdk_nvmf_tcp_qpair *tqpair) 860 { 861 SPDK_DEBUGLOG(nvmf_tcp, "Disconnecting qpair %p\n", tqpair); 862 863 spdk_trace_record(TRACE_TCP_QP_DISCONNECT, 0, 0, (uintptr_t)tqpair); 864 865 if (tqpair->state <= NVME_TCP_QPAIR_STATE_RUNNING) { 866 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITING); 867 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 868 spdk_poller_unregister(&tqpair->timeout_poller); 869 870 /* This will end up calling nvmf_tcp_close_qpair */ 871 spdk_nvmf_qpair_disconnect(&tqpair->qpair, NULL, NULL); 872 } 873 } 874 875 static void 876 _mgmt_pdu_write_done(void *_tqpair, int err) 877 { 878 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 879 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 880 881 if (spdk_unlikely(err != 0)) { 882 nvmf_tcp_qpair_disconnect(tqpair); 883 return; 884 } 885 886 assert(pdu->cb_fn != NULL); 887 pdu->cb_fn(pdu->cb_arg); 888 } 889 890 static void 891 _req_pdu_write_done(void *req, int err) 892 { 893 struct spdk_nvmf_tcp_req *tcp_req = req; 894 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 895 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 896 897 assert(tcp_req->pdu_in_use); 898 tcp_req->pdu_in_use = false; 899 900 /* If the request is in a completed state, we're waiting for write completion to free it */ 901 if (spdk_unlikely(tcp_req->state == TCP_REQUEST_STATE_COMPLETED)) { 902 nvmf_tcp_request_free(tcp_req); 903 return; 904 } 905 906 if (spdk_unlikely(err != 0)) { 907 nvmf_tcp_qpair_disconnect(tqpair); 908 return; 909 } 910 911 assert(pdu->cb_fn != NULL); 912 pdu->cb_fn(pdu->cb_arg); 913 } 914 915 static void 916 _pdu_write_done(struct nvme_tcp_pdu *pdu, int err) 917 { 918 pdu->sock_req.cb_fn(pdu->sock_req.cb_arg, err); 919 } 920 921 static void 922 _tcp_write_pdu(struct nvme_tcp_pdu *pdu) 923 { 924 uint32_t mapped_length = 0; 925 ssize_t rc; 926 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 927 928 pdu->sock_req.iovcnt = nvme_tcp_build_iovs(pdu->iov, SPDK_COUNTOF(pdu->iov), pdu, 929 tqpair->host_hdgst_enable, tqpair->host_ddgst_enable, 930 &mapped_length); 931 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP || 932 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ) { 933 rc = spdk_sock_writev(tqpair->sock, pdu->iov, pdu->sock_req.iovcnt); 934 if (rc == mapped_length) { 935 _pdu_write_done(pdu, 0); 936 } else { 937 SPDK_ERRLOG("IC_RESP or TERM_REQ could not write to socket.\n"); 938 _pdu_write_done(pdu, -1); 939 } 940 } else { 941 spdk_sock_writev_async(tqpair->sock, &pdu->sock_req); 942 } 943 } 944 945 static void 946 data_crc32_accel_done(void *cb_arg, int status) 947 { 948 struct nvme_tcp_pdu *pdu = cb_arg; 949 950 if (spdk_unlikely(status)) { 951 SPDK_ERRLOG("Failed to compute the data digest for pdu =%p\n", pdu); 952 _pdu_write_done(pdu, status); 953 return; 954 } 955 956 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 957 MAKE_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32); 958 959 _tcp_write_pdu(pdu); 960 } 961 962 static void 963 pdu_data_crc32_compute(struct nvme_tcp_pdu *pdu) 964 { 965 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 966 uint32_t crc32c; 967 968 /* Data Digest */ 969 if (pdu->data_len > 0 && g_nvme_tcp_ddgst[pdu->hdr.common.pdu_type] && tqpair->host_ddgst_enable) { 970 /* Only suport this limitated case for the first step */ 971 if (spdk_likely(!pdu->dif_ctx && (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0) 972 && tqpair->group)) { 973 if (!spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 974 pdu->data_iovcnt, 0, data_crc32_accel_done, pdu)) { 975 return; 976 } 977 } 978 979 crc32c = nvme_tcp_pdu_calc_data_digest(pdu); 980 crc32c = crc32c ^ SPDK_CRC32C_XOR; 981 MAKE_DIGEST_WORD(pdu->data_digest, crc32c); 982 } 983 984 _tcp_write_pdu(pdu); 985 } 986 987 static void 988 nvmf_tcp_qpair_write_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 989 struct nvme_tcp_pdu *pdu, 990 nvme_tcp_qpair_xfer_complete_cb cb_fn, 991 void *cb_arg) 992 { 993 int hlen; 994 uint32_t crc32c; 995 996 assert(tqpair->pdu_in_progress != pdu); 997 998 hlen = pdu->hdr.common.hlen; 999 pdu->cb_fn = cb_fn; 1000 pdu->cb_arg = cb_arg; 1001 1002 pdu->iov[0].iov_base = &pdu->hdr.raw; 1003 pdu->iov[0].iov_len = hlen; 1004 1005 /* Header Digest */ 1006 if (g_nvme_tcp_hdgst[pdu->hdr.common.pdu_type] && tqpair->host_hdgst_enable) { 1007 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 1008 MAKE_DIGEST_WORD((uint8_t *)pdu->hdr.raw + hlen, crc32c); 1009 } 1010 1011 /* Data Digest */ 1012 pdu_data_crc32_compute(pdu); 1013 } 1014 1015 static void 1016 nvmf_tcp_qpair_write_mgmt_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1017 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1018 void *cb_arg) 1019 { 1020 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1021 1022 pdu->sock_req.cb_fn = _mgmt_pdu_write_done; 1023 pdu->sock_req.cb_arg = tqpair; 1024 1025 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1026 } 1027 1028 static void 1029 nvmf_tcp_qpair_write_req_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1030 struct spdk_nvmf_tcp_req *tcp_req, 1031 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1032 void *cb_arg) 1033 { 1034 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1035 1036 pdu->sock_req.cb_fn = _req_pdu_write_done; 1037 pdu->sock_req.cb_arg = tcp_req; 1038 1039 assert(!tcp_req->pdu_in_use); 1040 tcp_req->pdu_in_use = true; 1041 1042 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1043 } 1044 1045 static int 1046 nvmf_tcp_qpair_init_mem_resource(struct spdk_nvmf_tcp_qpair *tqpair) 1047 { 1048 uint32_t i; 1049 struct spdk_nvmf_transport_opts *opts; 1050 uint32_t in_capsule_data_size; 1051 1052 opts = &tqpair->qpair.transport->opts; 1053 1054 in_capsule_data_size = opts->in_capsule_data_size; 1055 if (opts->dif_insert_or_strip) { 1056 in_capsule_data_size = SPDK_BDEV_BUF_SIZE_WITH_MD(in_capsule_data_size); 1057 } 1058 1059 tqpair->resource_count = opts->max_queue_depth; 1060 1061 tqpair->reqs = calloc(tqpair->resource_count, sizeof(*tqpair->reqs)); 1062 if (!tqpair->reqs) { 1063 SPDK_ERRLOG("Unable to allocate reqs on tqpair=%p\n", tqpair); 1064 return -1; 1065 } 1066 1067 if (in_capsule_data_size) { 1068 tqpair->bufs = spdk_zmalloc(tqpair->resource_count * in_capsule_data_size, 0x1000, 1069 NULL, SPDK_ENV_LCORE_ID_ANY, 1070 SPDK_MALLOC_DMA); 1071 if (!tqpair->bufs) { 1072 SPDK_ERRLOG("Unable to allocate bufs on tqpair=%p.\n", tqpair); 1073 return -1; 1074 } 1075 } 1076 1077 /* Add additional 2 members, which will be used for mgmt_pdu and pdu_in_progress owned by the tqpair */ 1078 tqpair->pdus = spdk_dma_zmalloc((tqpair->resource_count + 2) * sizeof(*tqpair->pdus), 0x1000, NULL); 1079 if (!tqpair->pdus) { 1080 SPDK_ERRLOG("Unable to allocate pdu pool on tqpair =%p.\n", tqpair); 1081 return -1; 1082 } 1083 1084 for (i = 0; i < tqpair->resource_count; i++) { 1085 struct spdk_nvmf_tcp_req *tcp_req = &tqpair->reqs[i]; 1086 1087 tcp_req->ttag = i + 1; 1088 tcp_req->req.qpair = &tqpair->qpair; 1089 1090 tcp_req->pdu = &tqpair->pdus[i]; 1091 tcp_req->pdu->qpair = tqpair; 1092 1093 /* Set up memory to receive commands */ 1094 if (tqpair->bufs) { 1095 tcp_req->buf = (void *)((uintptr_t)tqpair->bufs + (i * in_capsule_data_size)); 1096 } 1097 1098 /* Set the cmdn and rsp */ 1099 tcp_req->req.rsp = (union nvmf_c2h_msg *)&tcp_req->rsp; 1100 tcp_req->req.cmd = (union nvmf_h2c_msg *)&tcp_req->cmd; 1101 1102 tcp_req->req.stripped_data = NULL; 1103 1104 /* Initialize request state to FREE */ 1105 tcp_req->state = TCP_REQUEST_STATE_FREE; 1106 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 1107 tqpair->state_cntr[TCP_REQUEST_STATE_FREE]++; 1108 } 1109 1110 tqpair->mgmt_pdu = &tqpair->pdus[i]; 1111 tqpair->mgmt_pdu->qpair = tqpair; 1112 tqpair->pdu_in_progress = &tqpair->pdus[i + 1]; 1113 1114 tqpair->recv_buf_size = (in_capsule_data_size + sizeof(struct spdk_nvme_tcp_cmd) + 2 * 1115 SPDK_NVME_TCP_DIGEST_LEN) * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1116 1117 return 0; 1118 } 1119 1120 static int 1121 nvmf_tcp_qpair_init(struct spdk_nvmf_qpair *qpair) 1122 { 1123 struct spdk_nvmf_tcp_qpair *tqpair; 1124 1125 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1126 1127 SPDK_DEBUGLOG(nvmf_tcp, "New TCP Connection: %p\n", qpair); 1128 1129 spdk_trace_record(TRACE_TCP_QP_CREATE, 0, 0, (uintptr_t)tqpair); 1130 1131 /* Initialise request state queues of the qpair */ 1132 TAILQ_INIT(&tqpair->tcp_req_free_queue); 1133 TAILQ_INIT(&tqpair->tcp_req_working_queue); 1134 1135 tqpair->host_hdgst_enable = true; 1136 tqpair->host_ddgst_enable = true; 1137 1138 return 0; 1139 } 1140 1141 static int 1142 nvmf_tcp_qpair_sock_init(struct spdk_nvmf_tcp_qpair *tqpair) 1143 { 1144 int rc; 1145 1146 spdk_trace_record(TRACE_TCP_QP_SOCK_INIT, 0, 0, (uintptr_t)tqpair); 1147 1148 /* set low water mark */ 1149 rc = spdk_sock_set_recvlowat(tqpair->sock, 1); 1150 if (rc != 0) { 1151 SPDK_ERRLOG("spdk_sock_set_recvlowat() failed\n"); 1152 return rc; 1153 } 1154 1155 return 0; 1156 } 1157 1158 static void 1159 nvmf_tcp_handle_connect(struct spdk_nvmf_transport *transport, 1160 struct spdk_nvmf_tcp_port *port, 1161 struct spdk_sock *sock) 1162 { 1163 struct spdk_nvmf_tcp_qpair *tqpair; 1164 int rc; 1165 1166 SPDK_DEBUGLOG(nvmf_tcp, "New connection accepted on %s port %s\n", 1167 port->trid->traddr, port->trid->trsvcid); 1168 1169 tqpair = calloc(1, sizeof(struct spdk_nvmf_tcp_qpair)); 1170 if (tqpair == NULL) { 1171 SPDK_ERRLOG("Could not allocate new connection.\n"); 1172 spdk_sock_close(&sock); 1173 return; 1174 } 1175 1176 tqpair->sock = sock; 1177 tqpair->state_cntr[TCP_REQUEST_STATE_FREE] = 0; 1178 tqpair->port = port; 1179 tqpair->qpair.transport = transport; 1180 1181 rc = spdk_sock_getaddr(tqpair->sock, tqpair->target_addr, 1182 sizeof(tqpair->target_addr), &tqpair->target_port, 1183 tqpair->initiator_addr, sizeof(tqpair->initiator_addr), 1184 &tqpair->initiator_port); 1185 if (rc < 0) { 1186 SPDK_ERRLOG("spdk_sock_getaddr() failed of tqpair=%p\n", tqpair); 1187 nvmf_tcp_qpair_destroy(tqpair); 1188 return; 1189 } 1190 1191 spdk_nvmf_tgt_new_qpair(transport->tgt, &tqpair->qpair); 1192 } 1193 1194 static uint32_t 1195 nvmf_tcp_port_accept(struct spdk_nvmf_transport *transport, struct spdk_nvmf_tcp_port *port) 1196 { 1197 struct spdk_sock *sock; 1198 uint32_t count = 0; 1199 int i; 1200 1201 for (i = 0; i < NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME; i++) { 1202 sock = spdk_sock_accept(port->listen_sock); 1203 if (sock == NULL) { 1204 break; 1205 } 1206 count++; 1207 nvmf_tcp_handle_connect(transport, port, sock); 1208 } 1209 1210 return count; 1211 } 1212 1213 static int 1214 nvmf_tcp_accept(void *ctx) 1215 { 1216 struct spdk_nvmf_transport *transport = ctx; 1217 struct spdk_nvmf_tcp_transport *ttransport; 1218 struct spdk_nvmf_tcp_port *port; 1219 uint32_t count = 0; 1220 1221 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1222 1223 TAILQ_FOREACH(port, &ttransport->ports, link) { 1224 count += nvmf_tcp_port_accept(transport, port); 1225 } 1226 1227 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 1228 } 1229 1230 static void 1231 nvmf_tcp_discover(struct spdk_nvmf_transport *transport, 1232 struct spdk_nvme_transport_id *trid, 1233 struct spdk_nvmf_discovery_log_page_entry *entry) 1234 { 1235 entry->trtype = SPDK_NVMF_TRTYPE_TCP; 1236 entry->adrfam = trid->adrfam; 1237 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 1238 1239 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 1240 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 1241 1242 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_NONE; 1243 } 1244 1245 static struct spdk_nvmf_tcp_control_msg_list * 1246 nvmf_tcp_control_msg_list_create(uint16_t num_messages) 1247 { 1248 struct spdk_nvmf_tcp_control_msg_list *list; 1249 struct spdk_nvmf_tcp_control_msg *msg; 1250 uint16_t i; 1251 1252 list = calloc(1, sizeof(*list)); 1253 if (!list) { 1254 SPDK_ERRLOG("Failed to allocate memory for list structure\n"); 1255 return NULL; 1256 } 1257 1258 list->msg_buf = spdk_zmalloc(num_messages * SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, 1259 NVMF_DATA_BUFFER_ALIGNMENT, NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 1260 if (!list->msg_buf) { 1261 SPDK_ERRLOG("Failed to allocate memory for control message buffers\n"); 1262 free(list); 1263 return NULL; 1264 } 1265 1266 STAILQ_INIT(&list->free_msgs); 1267 1268 for (i = 0; i < num_messages; i++) { 1269 msg = (struct spdk_nvmf_tcp_control_msg *)((char *)list->msg_buf + i * 1270 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1271 STAILQ_INSERT_TAIL(&list->free_msgs, msg, link); 1272 } 1273 1274 return list; 1275 } 1276 1277 static void 1278 nvmf_tcp_control_msg_list_free(struct spdk_nvmf_tcp_control_msg_list *list) 1279 { 1280 if (!list) { 1281 return; 1282 } 1283 1284 spdk_free(list->msg_buf); 1285 free(list); 1286 } 1287 1288 static struct spdk_nvmf_transport_poll_group * 1289 nvmf_tcp_poll_group_create(struct spdk_nvmf_transport *transport, 1290 struct spdk_nvmf_poll_group *group) 1291 { 1292 struct spdk_nvmf_tcp_transport *ttransport; 1293 struct spdk_nvmf_tcp_poll_group *tgroup; 1294 1295 tgroup = calloc(1, sizeof(*tgroup)); 1296 if (!tgroup) { 1297 return NULL; 1298 } 1299 1300 tgroup->sock_group = spdk_sock_group_create(&tgroup->group); 1301 if (!tgroup->sock_group) { 1302 goto cleanup; 1303 } 1304 1305 TAILQ_INIT(&tgroup->qpairs); 1306 TAILQ_INIT(&tgroup->await_req); 1307 1308 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1309 1310 if (transport->opts.in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 1311 SPDK_DEBUGLOG(nvmf_tcp, "ICD %u is less than min required for admin/fabric commands (%u). " 1312 "Creating control messages list\n", transport->opts.in_capsule_data_size, 1313 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1314 tgroup->control_msg_list = nvmf_tcp_control_msg_list_create(ttransport->tcp_opts.control_msg_num); 1315 if (!tgroup->control_msg_list) { 1316 goto cleanup; 1317 } 1318 } 1319 1320 tgroup->accel_channel = spdk_accel_engine_get_io_channel(); 1321 if (spdk_unlikely(!tgroup->accel_channel)) { 1322 SPDK_ERRLOG("Cannot create accel_channel for tgroup=%p\n", tgroup); 1323 goto cleanup; 1324 } 1325 1326 pthread_mutex_lock(&ttransport->lock); 1327 TAILQ_INSERT_TAIL(&ttransport->poll_groups, tgroup, link); 1328 if (ttransport->next_pg == NULL) { 1329 ttransport->next_pg = tgroup; 1330 } 1331 pthread_mutex_unlock(&ttransport->lock); 1332 1333 return &tgroup->group; 1334 1335 cleanup: 1336 nvmf_tcp_poll_group_destroy(&tgroup->group); 1337 return NULL; 1338 } 1339 1340 static struct spdk_nvmf_transport_poll_group * 1341 nvmf_tcp_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 1342 { 1343 struct spdk_nvmf_tcp_transport *ttransport; 1344 struct spdk_nvmf_tcp_poll_group **pg; 1345 struct spdk_nvmf_tcp_qpair *tqpair; 1346 struct spdk_sock_group *group = NULL, *hint = NULL; 1347 int rc; 1348 1349 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 1350 1351 pthread_mutex_lock(&ttransport->lock); 1352 1353 if (TAILQ_EMPTY(&ttransport->poll_groups)) { 1354 pthread_mutex_unlock(&ttransport->lock); 1355 return NULL; 1356 } 1357 1358 pg = &ttransport->next_pg; 1359 assert(*pg != NULL); 1360 hint = (*pg)->sock_group; 1361 1362 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1363 rc = spdk_sock_get_optimal_sock_group(tqpair->sock, &group, hint); 1364 if (rc != 0) { 1365 pthread_mutex_unlock(&ttransport->lock); 1366 return NULL; 1367 } else if (group != NULL) { 1368 /* Optimal poll group was found */ 1369 pthread_mutex_unlock(&ttransport->lock); 1370 return spdk_sock_group_get_ctx(group); 1371 } 1372 1373 /* The hint was used for optimal poll group, advance next_pg. */ 1374 *pg = TAILQ_NEXT(*pg, link); 1375 if (*pg == NULL) { 1376 *pg = TAILQ_FIRST(&ttransport->poll_groups); 1377 } 1378 1379 pthread_mutex_unlock(&ttransport->lock); 1380 return spdk_sock_group_get_ctx(hint); 1381 } 1382 1383 static void 1384 nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 1385 { 1386 struct spdk_nvmf_tcp_poll_group *tgroup, *next_tgroup; 1387 struct spdk_nvmf_tcp_transport *ttransport; 1388 1389 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 1390 spdk_sock_group_close(&tgroup->sock_group); 1391 if (tgroup->control_msg_list) { 1392 nvmf_tcp_control_msg_list_free(tgroup->control_msg_list); 1393 } 1394 1395 if (tgroup->accel_channel) { 1396 spdk_put_io_channel(tgroup->accel_channel); 1397 } 1398 1399 ttransport = SPDK_CONTAINEROF(tgroup->group.transport, struct spdk_nvmf_tcp_transport, transport); 1400 1401 pthread_mutex_lock(&ttransport->lock); 1402 next_tgroup = TAILQ_NEXT(tgroup, link); 1403 TAILQ_REMOVE(&ttransport->poll_groups, tgroup, link); 1404 if (next_tgroup == NULL) { 1405 next_tgroup = TAILQ_FIRST(&ttransport->poll_groups); 1406 } 1407 if (ttransport->next_pg == tgroup) { 1408 ttransport->next_pg = next_tgroup; 1409 } 1410 pthread_mutex_unlock(&ttransport->lock); 1411 1412 free(tgroup); 1413 } 1414 1415 static void 1416 nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 1417 enum nvme_tcp_pdu_recv_state state) 1418 { 1419 if (tqpair->recv_state == state) { 1420 SPDK_ERRLOG("The recv state of tqpair=%p is same with the state(%d) to be set\n", 1421 tqpair, state); 1422 return; 1423 } 1424 1425 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1426 /* When leaving the await req state, move the qpair to the main list */ 1427 TAILQ_REMOVE(&tqpair->group->await_req, tqpair, link); 1428 TAILQ_INSERT_TAIL(&tqpair->group->qpairs, tqpair, link); 1429 } 1430 1431 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv state=%d\n", tqpair, state); 1432 tqpair->recv_state = state; 1433 1434 spdk_trace_record(TRACE_TCP_QP_RCV_STATE_CHANGE, tqpair->qpair.qid, 0, (uintptr_t)tqpair, 1435 tqpair->recv_state); 1436 1437 switch (state) { 1438 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH: 1439 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH: 1440 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD: 1441 break; 1442 case NVME_TCP_PDU_RECV_STATE_AWAIT_REQ: 1443 TAILQ_REMOVE(&tqpair->group->qpairs, tqpair, link); 1444 TAILQ_INSERT_TAIL(&tqpair->group->await_req, tqpair, link); 1445 break; 1446 case NVME_TCP_PDU_RECV_STATE_ERROR: 1447 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY: 1448 memset(tqpair->pdu_in_progress, 0, sizeof(*(tqpair->pdu_in_progress))); 1449 break; 1450 default: 1451 SPDK_ERRLOG("The state(%d) is invalid\n", state); 1452 abort(); 1453 break; 1454 } 1455 } 1456 1457 static int 1458 nvmf_tcp_qpair_handle_timeout(void *ctx) 1459 { 1460 struct spdk_nvmf_tcp_qpair *tqpair = ctx; 1461 1462 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1463 1464 SPDK_ERRLOG("No pdu coming for tqpair=%p within %d seconds\n", tqpair, 1465 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT); 1466 1467 nvmf_tcp_qpair_disconnect(tqpair); 1468 return SPDK_POLLER_BUSY; 1469 } 1470 1471 static void 1472 nvmf_tcp_send_c2h_term_req_complete(void *cb_arg) 1473 { 1474 struct spdk_nvmf_tcp_qpair *tqpair = (struct spdk_nvmf_tcp_qpair *)cb_arg; 1475 1476 if (!tqpair->timeout_poller) { 1477 tqpair->timeout_poller = SPDK_POLLER_REGISTER(nvmf_tcp_qpair_handle_timeout, tqpair, 1478 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT * 1000000); 1479 } 1480 } 1481 1482 static void 1483 nvmf_tcp_send_c2h_term_req(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu, 1484 enum spdk_nvme_tcp_term_req_fes fes, uint32_t error_offset) 1485 { 1486 struct nvme_tcp_pdu *rsp_pdu; 1487 struct spdk_nvme_tcp_term_req_hdr *c2h_term_req; 1488 uint32_t c2h_term_req_hdr_len = sizeof(*c2h_term_req); 1489 uint32_t copy_len; 1490 1491 rsp_pdu = tqpair->mgmt_pdu; 1492 1493 c2h_term_req = &rsp_pdu->hdr.term_req; 1494 c2h_term_req->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ; 1495 c2h_term_req->common.hlen = c2h_term_req_hdr_len; 1496 1497 if ((fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1498 (fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1499 DSET32(&c2h_term_req->fei, error_offset); 1500 } 1501 1502 copy_len = spdk_min(pdu->hdr.common.hlen, SPDK_NVME_TCP_TERM_REQ_ERROR_DATA_MAX_SIZE); 1503 1504 /* Copy the error info into the buffer */ 1505 memcpy((uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, pdu->hdr.raw, copy_len); 1506 nvme_tcp_pdu_set_data(rsp_pdu, (uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, copy_len); 1507 1508 /* Contain the header of the wrong received pdu */ 1509 c2h_term_req->common.plen = c2h_term_req->common.hlen + copy_len; 1510 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 1511 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_c2h_term_req_complete, tqpair); 1512 } 1513 1514 static void 1515 nvmf_tcp_capsule_cmd_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1516 struct spdk_nvmf_tcp_qpair *tqpair, 1517 struct nvme_tcp_pdu *pdu) 1518 { 1519 struct spdk_nvmf_tcp_req *tcp_req; 1520 1521 assert(pdu->psh_valid_bytes == pdu->psh_len); 1522 assert(pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD); 1523 1524 tcp_req = nvmf_tcp_req_get(tqpair); 1525 if (!tcp_req) { 1526 /* Directly return and make the allocation retry again. This can happen if we're 1527 * using asynchronous writes to send the response to the host or when releasing 1528 * zero-copy buffers after a response has been sent. In both cases, the host might 1529 * receive the response before we've finished processing the request and is free to 1530 * send another one. 1531 */ 1532 if (tqpair->state_cntr[TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST] > 0 || 1533 tqpair->state_cntr[TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE] > 0) { 1534 return; 1535 } 1536 1537 /* The host sent more commands than the maximum queue depth. */ 1538 SPDK_ERRLOG("Cannot allocate tcp_req on tqpair=%p\n", tqpair); 1539 nvmf_tcp_qpair_disconnect(tqpair); 1540 return; 1541 } 1542 1543 pdu->req = tcp_req; 1544 assert(tcp_req->state == TCP_REQUEST_STATE_NEW); 1545 nvmf_tcp_req_process(ttransport, tcp_req); 1546 } 1547 1548 static void 1549 nvmf_tcp_capsule_cmd_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1550 struct spdk_nvmf_tcp_qpair *tqpair, 1551 struct nvme_tcp_pdu *pdu) 1552 { 1553 struct spdk_nvmf_tcp_req *tcp_req; 1554 struct spdk_nvme_tcp_cmd *capsule_cmd; 1555 uint32_t error_offset = 0; 1556 enum spdk_nvme_tcp_term_req_fes fes; 1557 struct spdk_nvme_cpl *rsp; 1558 1559 capsule_cmd = &pdu->hdr.capsule_cmd; 1560 tcp_req = pdu->req; 1561 assert(tcp_req != NULL); 1562 1563 /* Zero-copy requests don't support ICD */ 1564 assert(!spdk_nvmf_request_using_zcopy(&tcp_req->req)); 1565 1566 if (capsule_cmd->common.pdo > SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET) { 1567 SPDK_ERRLOG("Expected ICReq capsule_cmd pdu offset <= %d, got %c\n", 1568 SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET, capsule_cmd->common.pdo); 1569 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1570 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 1571 goto err; 1572 } 1573 1574 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 1575 1576 rsp = &tcp_req->req.rsp->nvme_cpl; 1577 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1578 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1579 } else { 1580 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1581 } 1582 1583 nvmf_tcp_req_process(ttransport, tcp_req); 1584 1585 return; 1586 err: 1587 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1588 } 1589 1590 static void 1591 nvmf_tcp_h2c_data_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1592 struct spdk_nvmf_tcp_qpair *tqpair, 1593 struct nvme_tcp_pdu *pdu) 1594 { 1595 struct spdk_nvmf_tcp_req *tcp_req; 1596 uint32_t error_offset = 0; 1597 enum spdk_nvme_tcp_term_req_fes fes = 0; 1598 struct spdk_nvme_tcp_h2c_data_hdr *h2c_data; 1599 1600 h2c_data = &pdu->hdr.h2c_data; 1601 1602 SPDK_DEBUGLOG(nvmf_tcp, "tqpair=%p, r2t_info: datao=%u, datal=%u, cccid=%u, ttag=%u\n", 1603 tqpair, h2c_data->datao, h2c_data->datal, h2c_data->cccid, h2c_data->ttag); 1604 1605 if (h2c_data->ttag > tqpair->resource_count) { 1606 SPDK_DEBUGLOG(nvmf_tcp, "ttag %u is larger than allowed %u.\n", h2c_data->ttag, 1607 tqpair->resource_count); 1608 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1609 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1610 goto err; 1611 } 1612 1613 tcp_req = &tqpair->reqs[h2c_data->ttag - 1]; 1614 1615 if (spdk_unlikely(tcp_req->state != TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER && 1616 tcp_req->state != TCP_REQUEST_STATE_AWAITING_R2T_ACK)) { 1617 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, has error state in %d\n", tcp_req, tqpair, 1618 tcp_req->state); 1619 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1620 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1621 goto err; 1622 } 1623 1624 if (spdk_unlikely(tcp_req->req.cmd->nvme_cmd.cid != h2c_data->cccid)) { 1625 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, expected %u but %u for cccid.\n", tcp_req, tqpair, 1626 tcp_req->req.cmd->nvme_cmd.cid, h2c_data->cccid); 1627 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1628 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, cccid); 1629 goto err; 1630 } 1631 1632 if (tcp_req->h2c_offset != h2c_data->datao) { 1633 SPDK_DEBUGLOG(nvmf_tcp, 1634 "tcp_req(%p), tqpair=%p, expected data offset %u, but data offset is %u\n", 1635 tcp_req, tqpair, tcp_req->h2c_offset, h2c_data->datao); 1636 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1637 goto err; 1638 } 1639 1640 if ((h2c_data->datao + h2c_data->datal) > tcp_req->req.length) { 1641 SPDK_DEBUGLOG(nvmf_tcp, 1642 "tcp_req(%p), tqpair=%p, (datao=%u + datal=%u) exceeds requested length=%u\n", 1643 tcp_req, tqpair, h2c_data->datao, h2c_data->datal, tcp_req->req.length); 1644 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1645 goto err; 1646 } 1647 1648 pdu->req = tcp_req; 1649 1650 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 1651 pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 1652 } 1653 1654 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 1655 h2c_data->datao, h2c_data->datal); 1656 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1657 return; 1658 1659 err: 1660 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1661 } 1662 1663 static void 1664 nvmf_tcp_send_capsule_resp_pdu(struct spdk_nvmf_tcp_req *tcp_req, 1665 struct spdk_nvmf_tcp_qpair *tqpair) 1666 { 1667 struct nvme_tcp_pdu *rsp_pdu; 1668 struct spdk_nvme_tcp_rsp *capsule_resp; 1669 1670 SPDK_DEBUGLOG(nvmf_tcp, "enter, tqpair=%p\n", tqpair); 1671 1672 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1673 assert(rsp_pdu != NULL); 1674 1675 capsule_resp = &rsp_pdu->hdr.capsule_resp; 1676 capsule_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_CAPSULE_RESP; 1677 capsule_resp->common.plen = capsule_resp->common.hlen = sizeof(*capsule_resp); 1678 capsule_resp->rccqe = tcp_req->req.rsp->nvme_cpl; 1679 if (tqpair->host_hdgst_enable) { 1680 capsule_resp->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 1681 capsule_resp->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 1682 } 1683 1684 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_request_free, tcp_req); 1685 } 1686 1687 static void 1688 nvmf_tcp_pdu_c2h_data_complete(void *cb_arg) 1689 { 1690 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 1691 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, 1692 struct spdk_nvmf_tcp_qpair, qpair); 1693 1694 assert(tqpair != NULL); 1695 1696 if (spdk_unlikely(tcp_req->pdu->rw_offset < tcp_req->req.length)) { 1697 SPDK_DEBUGLOG(nvmf_tcp, "sending another C2H part, offset %u length %u\n", tcp_req->pdu->rw_offset, 1698 tcp_req->req.length); 1699 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 1700 return; 1701 } 1702 1703 if (tcp_req->pdu->hdr.c2h_data.common.flags & SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS) { 1704 nvmf_tcp_request_free(tcp_req); 1705 } else { 1706 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 1707 } 1708 } 1709 1710 static void 1711 nvmf_tcp_r2t_complete(void *cb_arg) 1712 { 1713 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 1714 struct spdk_nvmf_tcp_transport *ttransport; 1715 1716 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 1717 struct spdk_nvmf_tcp_transport, transport); 1718 1719 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1720 1721 if (tcp_req->h2c_offset == tcp_req->req.length) { 1722 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1723 nvmf_tcp_req_process(ttransport, tcp_req); 1724 } 1725 } 1726 1727 static void 1728 nvmf_tcp_send_r2t_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1729 struct spdk_nvmf_tcp_req *tcp_req) 1730 { 1731 struct nvme_tcp_pdu *rsp_pdu; 1732 struct spdk_nvme_tcp_r2t_hdr *r2t; 1733 1734 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1735 assert(rsp_pdu != NULL); 1736 1737 r2t = &rsp_pdu->hdr.r2t; 1738 r2t->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_R2T; 1739 r2t->common.plen = r2t->common.hlen = sizeof(*r2t); 1740 1741 if (tqpair->host_hdgst_enable) { 1742 r2t->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 1743 r2t->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 1744 } 1745 1746 r2t->cccid = tcp_req->req.cmd->nvme_cmd.cid; 1747 r2t->ttag = tcp_req->ttag; 1748 r2t->r2to = tcp_req->h2c_offset; 1749 r2t->r2tl = tcp_req->req.length; 1750 1751 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 1752 1753 SPDK_DEBUGLOG(nvmf_tcp, 1754 "tcp_req(%p) on tqpair(%p), r2t_info: cccid=%u, ttag=%u, r2to=%u, r2tl=%u\n", 1755 tcp_req, tqpair, r2t->cccid, r2t->ttag, r2t->r2to, r2t->r2tl); 1756 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_r2t_complete, tcp_req); 1757 } 1758 1759 static void 1760 nvmf_tcp_h2c_data_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1761 struct spdk_nvmf_tcp_qpair *tqpair, 1762 struct nvme_tcp_pdu *pdu) 1763 { 1764 struct spdk_nvmf_tcp_req *tcp_req; 1765 struct spdk_nvme_cpl *rsp; 1766 1767 tcp_req = pdu->req; 1768 assert(tcp_req != NULL); 1769 1770 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 1771 1772 tcp_req->h2c_offset += pdu->data_len; 1773 1774 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 1775 1776 /* Wait for all of the data to arrive AND for the initial R2T PDU send to be 1777 * acknowledged before moving on. */ 1778 if (tcp_req->h2c_offset == tcp_req->req.length && 1779 tcp_req->state == TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER) { 1780 /* After receiving all the h2c data, we need to check whether there is 1781 * transient transport error */ 1782 rsp = &tcp_req->req.rsp->nvme_cpl; 1783 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1784 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1785 } else { 1786 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1787 } 1788 nvmf_tcp_req_process(ttransport, tcp_req); 1789 } 1790 } 1791 1792 static void 1793 nvmf_tcp_h2c_term_req_dump(struct spdk_nvme_tcp_term_req_hdr *h2c_term_req) 1794 { 1795 SPDK_ERRLOG("Error info of pdu(%p): %s\n", h2c_term_req, 1796 spdk_nvmf_tcp_term_req_fes_str[h2c_term_req->fes]); 1797 if ((h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1798 (h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1799 SPDK_DEBUGLOG(nvmf_tcp, "The offset from the start of the PDU header is %u\n", 1800 DGET32(h2c_term_req->fei)); 1801 } 1802 } 1803 1804 static void 1805 nvmf_tcp_h2c_term_req_hdr_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1806 struct nvme_tcp_pdu *pdu) 1807 { 1808 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 1809 uint32_t error_offset = 0; 1810 enum spdk_nvme_tcp_term_req_fes fes; 1811 1812 if (h2c_term_req->fes > SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER) { 1813 SPDK_ERRLOG("Fatal Error Status(FES) is unknown for h2c_term_req pdu=%p\n", pdu); 1814 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1815 error_offset = offsetof(struct spdk_nvme_tcp_term_req_hdr, fes); 1816 goto end; 1817 } 1818 1819 /* set the data buffer */ 1820 nvme_tcp_pdu_set_data(pdu, (uint8_t *)pdu->hdr.raw + h2c_term_req->common.hlen, 1821 h2c_term_req->common.plen - h2c_term_req->common.hlen); 1822 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1823 return; 1824 end: 1825 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1826 } 1827 1828 static void 1829 nvmf_tcp_h2c_term_req_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1830 struct nvme_tcp_pdu *pdu) 1831 { 1832 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 1833 1834 nvmf_tcp_h2c_term_req_dump(h2c_term_req); 1835 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 1836 } 1837 1838 static void 1839 _nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1840 struct spdk_nvmf_tcp_transport *ttransport) 1841 { 1842 struct nvme_tcp_pdu *pdu = tqpair->pdu_in_progress; 1843 1844 switch (pdu->hdr.common.pdu_type) { 1845 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 1846 nvmf_tcp_capsule_cmd_payload_handle(ttransport, tqpair, pdu); 1847 break; 1848 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 1849 nvmf_tcp_h2c_data_payload_handle(ttransport, tqpair, pdu); 1850 break; 1851 1852 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 1853 nvmf_tcp_h2c_term_req_payload_handle(tqpair, pdu); 1854 break; 1855 1856 default: 1857 /* The code should not go to here */ 1858 SPDK_ERRLOG("The code should not go to here\n"); 1859 break; 1860 } 1861 } 1862 1863 static void 1864 nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1865 struct spdk_nvmf_tcp_transport *ttransport) 1866 { 1867 int rc = 0; 1868 struct nvme_tcp_pdu *pdu; 1869 uint32_t crc32c; 1870 struct spdk_nvmf_tcp_req *tcp_req; 1871 struct spdk_nvme_cpl *rsp; 1872 1873 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1874 pdu = tqpair->pdu_in_progress; 1875 1876 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 1877 /* check data digest if need */ 1878 if (pdu->ddgst_enable) { 1879 crc32c = nvme_tcp_pdu_calc_data_digest(pdu); 1880 crc32c = crc32c ^ SPDK_CRC32C_XOR; 1881 rc = MATCH_DIGEST_WORD(pdu->data_digest, crc32c); 1882 if (rc == 0) { 1883 SPDK_ERRLOG("Data digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 1884 tcp_req = pdu->req; 1885 assert(tcp_req != NULL); 1886 rsp = &tcp_req->req.rsp->nvme_cpl; 1887 rsp->status.sc = SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR; 1888 } 1889 } 1890 1891 _nvmf_tcp_pdu_payload_handle(tqpair, ttransport); 1892 } 1893 1894 static void 1895 nvmf_tcp_send_icresp_complete(void *cb_arg) 1896 { 1897 struct spdk_nvmf_tcp_qpair *tqpair = cb_arg; 1898 1899 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_RUNNING); 1900 } 1901 1902 static void 1903 nvmf_tcp_icreq_handle(struct spdk_nvmf_tcp_transport *ttransport, 1904 struct spdk_nvmf_tcp_qpair *tqpair, 1905 struct nvme_tcp_pdu *pdu) 1906 { 1907 struct spdk_nvme_tcp_ic_req *ic_req = &pdu->hdr.ic_req; 1908 struct nvme_tcp_pdu *rsp_pdu; 1909 struct spdk_nvme_tcp_ic_resp *ic_resp; 1910 uint32_t error_offset = 0; 1911 enum spdk_nvme_tcp_term_req_fes fes; 1912 1913 /* Only PFV 0 is defined currently */ 1914 if (ic_req->pfv != 0) { 1915 SPDK_ERRLOG("Expected ICReq PFV %u, got %u\n", 0u, ic_req->pfv); 1916 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1917 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, pfv); 1918 goto end; 1919 } 1920 1921 /* MAXR2T is 0's based */ 1922 SPDK_DEBUGLOG(nvmf_tcp, "maxr2t =%u\n", (ic_req->maxr2t + 1u)); 1923 1924 tqpair->host_hdgst_enable = ic_req->dgst.bits.hdgst_enable ? true : false; 1925 if (!tqpair->host_hdgst_enable) { 1926 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1927 } 1928 1929 tqpair->host_ddgst_enable = ic_req->dgst.bits.ddgst_enable ? true : false; 1930 if (!tqpair->host_ddgst_enable) { 1931 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1932 } 1933 1934 tqpair->recv_buf_size = spdk_max(tqpair->recv_buf_size, MIN_SOCK_PIPE_SIZE); 1935 /* Now that we know whether digests are enabled, properly size the receive buffer */ 1936 if (spdk_sock_set_recvbuf(tqpair->sock, tqpair->recv_buf_size) < 0) { 1937 SPDK_WARNLOG("Unable to allocate enough memory for receive buffer on tqpair=%p with size=%d\n", 1938 tqpair, 1939 tqpair->recv_buf_size); 1940 /* Not fatal. */ 1941 } 1942 1943 tqpair->cpda = spdk_min(ic_req->hpda, SPDK_NVME_TCP_CPDA_MAX); 1944 SPDK_DEBUGLOG(nvmf_tcp, "cpda of tqpair=(%p) is : %u\n", tqpair, tqpair->cpda); 1945 1946 rsp_pdu = tqpair->mgmt_pdu; 1947 1948 ic_resp = &rsp_pdu->hdr.ic_resp; 1949 ic_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_IC_RESP; 1950 ic_resp->common.hlen = ic_resp->common.plen = sizeof(*ic_resp); 1951 ic_resp->pfv = 0; 1952 ic_resp->cpda = tqpair->cpda; 1953 ic_resp->maxh2cdata = ttransport->transport.opts.max_io_size; 1954 ic_resp->dgst.bits.hdgst_enable = tqpair->host_hdgst_enable ? 1 : 0; 1955 ic_resp->dgst.bits.ddgst_enable = tqpair->host_ddgst_enable ? 1 : 0; 1956 1957 SPDK_DEBUGLOG(nvmf_tcp, "host_hdgst_enable: %u\n", tqpair->host_hdgst_enable); 1958 SPDK_DEBUGLOG(nvmf_tcp, "host_ddgst_enable: %u\n", tqpair->host_ddgst_enable); 1959 1960 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INITIALIZING); 1961 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_icresp_complete, tqpair); 1962 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 1963 return; 1964 end: 1965 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1966 } 1967 1968 static void 1969 nvmf_tcp_pdu_psh_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1970 struct spdk_nvmf_tcp_transport *ttransport) 1971 { 1972 struct nvme_tcp_pdu *pdu; 1973 int rc; 1974 uint32_t crc32c, error_offset = 0; 1975 enum spdk_nvme_tcp_term_req_fes fes; 1976 1977 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 1978 pdu = tqpair->pdu_in_progress; 1979 1980 SPDK_DEBUGLOG(nvmf_tcp, "pdu type of tqpair(%p) is %d\n", tqpair, 1981 pdu->hdr.common.pdu_type); 1982 /* check header digest if needed */ 1983 if (pdu->has_hdgst) { 1984 SPDK_DEBUGLOG(nvmf_tcp, "Compare the header of pdu=%p on tqpair=%p\n", pdu, tqpair); 1985 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 1986 rc = MATCH_DIGEST_WORD((uint8_t *)pdu->hdr.raw + pdu->hdr.common.hlen, crc32c); 1987 if (rc == 0) { 1988 SPDK_ERRLOG("Header digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 1989 fes = SPDK_NVME_TCP_TERM_REQ_FES_HDGST_ERROR; 1990 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1991 return; 1992 1993 } 1994 } 1995 1996 switch (pdu->hdr.common.pdu_type) { 1997 case SPDK_NVME_TCP_PDU_TYPE_IC_REQ: 1998 nvmf_tcp_icreq_handle(ttransport, tqpair, pdu); 1999 break; 2000 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2001 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_REQ); 2002 break; 2003 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2004 nvmf_tcp_h2c_data_hdr_handle(ttransport, tqpair, pdu); 2005 break; 2006 2007 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2008 nvmf_tcp_h2c_term_req_hdr_handle(tqpair, pdu); 2009 break; 2010 2011 default: 2012 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", tqpair->pdu_in_progress->hdr.common.pdu_type); 2013 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2014 error_offset = 1; 2015 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2016 break; 2017 } 2018 } 2019 2020 static void 2021 nvmf_tcp_pdu_ch_handle(struct spdk_nvmf_tcp_qpair *tqpair) 2022 { 2023 struct nvme_tcp_pdu *pdu; 2024 uint32_t error_offset = 0; 2025 enum spdk_nvme_tcp_term_req_fes fes; 2026 uint8_t expected_hlen, pdo; 2027 bool plen_error = false, pdo_error = false; 2028 2029 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2030 pdu = tqpair->pdu_in_progress; 2031 2032 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_REQ) { 2033 if (tqpair->state != NVME_TCP_QPAIR_STATE_INVALID) { 2034 SPDK_ERRLOG("Already received ICreq PDU, and reject this pdu=%p\n", pdu); 2035 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2036 goto err; 2037 } 2038 expected_hlen = sizeof(struct spdk_nvme_tcp_ic_req); 2039 if (pdu->hdr.common.plen != expected_hlen) { 2040 plen_error = true; 2041 } 2042 } else { 2043 if (tqpair->state != NVME_TCP_QPAIR_STATE_RUNNING) { 2044 SPDK_ERRLOG("The TCP/IP connection is not negotiated\n"); 2045 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2046 goto err; 2047 } 2048 2049 switch (pdu->hdr.common.pdu_type) { 2050 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2051 expected_hlen = sizeof(struct spdk_nvme_tcp_cmd); 2052 pdo = pdu->hdr.common.pdo; 2053 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2054 pdo_error = true; 2055 break; 2056 } 2057 2058 if (pdu->hdr.common.plen < expected_hlen) { 2059 plen_error = true; 2060 } 2061 break; 2062 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2063 expected_hlen = sizeof(struct spdk_nvme_tcp_h2c_data_hdr); 2064 pdo = pdu->hdr.common.pdo; 2065 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2066 pdo_error = true; 2067 break; 2068 } 2069 if (pdu->hdr.common.plen < expected_hlen) { 2070 plen_error = true; 2071 } 2072 break; 2073 2074 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2075 expected_hlen = sizeof(struct spdk_nvme_tcp_term_req_hdr); 2076 if ((pdu->hdr.common.plen <= expected_hlen) || 2077 (pdu->hdr.common.plen > SPDK_NVME_TCP_TERM_REQ_PDU_MAX_SIZE)) { 2078 plen_error = true; 2079 } 2080 break; 2081 2082 default: 2083 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", pdu->hdr.common.pdu_type); 2084 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2085 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdu_type); 2086 goto err; 2087 } 2088 } 2089 2090 if (pdu->hdr.common.hlen != expected_hlen) { 2091 SPDK_ERRLOG("PDU type=0x%02x, Expected ICReq header length %u, got %u on tqpair=%p\n", 2092 pdu->hdr.common.pdu_type, 2093 expected_hlen, pdu->hdr.common.hlen, tqpair); 2094 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2095 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, hlen); 2096 goto err; 2097 } else if (pdo_error) { 2098 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2099 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 2100 } else if (plen_error) { 2101 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2102 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2103 goto err; 2104 } else { 2105 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2106 nvme_tcp_pdu_calc_psh_len(tqpair->pdu_in_progress, tqpair->host_hdgst_enable); 2107 return; 2108 } 2109 err: 2110 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2111 } 2112 2113 static int 2114 nvmf_tcp_pdu_payload_insert_dif(struct nvme_tcp_pdu *pdu, uint32_t read_offset, 2115 int read_len) 2116 { 2117 int rc; 2118 2119 rc = spdk_dif_generate_stream(pdu->data_iov, pdu->data_iovcnt, 2120 read_offset, read_len, pdu->dif_ctx); 2121 if (rc != 0) { 2122 SPDK_ERRLOG("DIF generate failed\n"); 2123 } 2124 2125 return rc; 2126 } 2127 2128 static int 2129 nvmf_tcp_sock_process(struct spdk_nvmf_tcp_qpair *tqpair) 2130 { 2131 int rc = 0; 2132 struct nvme_tcp_pdu *pdu; 2133 enum nvme_tcp_pdu_recv_state prev_state; 2134 uint32_t data_len; 2135 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2136 struct spdk_nvmf_tcp_transport, transport); 2137 2138 /* The loop here is to allow for several back-to-back state changes. */ 2139 do { 2140 prev_state = tqpair->recv_state; 2141 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv pdu entering state %d\n", tqpair, prev_state); 2142 2143 pdu = tqpair->pdu_in_progress; 2144 switch (tqpair->recv_state) { 2145 /* Wait for the common header */ 2146 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY: 2147 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH: 2148 if (spdk_unlikely(tqpair->state == NVME_TCP_QPAIR_STATE_INITIALIZING)) { 2149 return rc; 2150 } 2151 2152 rc = nvme_tcp_read_data(tqpair->sock, 2153 sizeof(struct spdk_nvme_tcp_common_pdu_hdr) - pdu->ch_valid_bytes, 2154 (void *)&pdu->hdr.common + pdu->ch_valid_bytes); 2155 if (rc < 0) { 2156 SPDK_DEBUGLOG(nvmf_tcp, "will disconnect tqpair=%p\n", tqpair); 2157 return NVME_TCP_PDU_FATAL; 2158 } else if (rc > 0) { 2159 pdu->ch_valid_bytes += rc; 2160 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.qid, rc, 0, tqpair); 2161 if (spdk_likely(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY)) { 2162 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2163 } 2164 } 2165 2166 if (pdu->ch_valid_bytes < sizeof(struct spdk_nvme_tcp_common_pdu_hdr)) { 2167 return NVME_TCP_PDU_IN_PROGRESS; 2168 } 2169 2170 /* The command header of this PDU has now been read from the socket. */ 2171 nvmf_tcp_pdu_ch_handle(tqpair); 2172 break; 2173 /* Wait for the pdu specific header */ 2174 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH: 2175 rc = nvme_tcp_read_data(tqpair->sock, 2176 pdu->psh_len - pdu->psh_valid_bytes, 2177 (void *)&pdu->hdr.raw + sizeof(struct spdk_nvme_tcp_common_pdu_hdr) + pdu->psh_valid_bytes); 2178 if (rc < 0) { 2179 return NVME_TCP_PDU_FATAL; 2180 } else if (rc > 0) { 2181 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.qid, rc, 0, tqpair); 2182 pdu->psh_valid_bytes += rc; 2183 } 2184 2185 if (pdu->psh_valid_bytes < pdu->psh_len) { 2186 return NVME_TCP_PDU_IN_PROGRESS; 2187 } 2188 2189 /* All header(ch, psh, head digist) of this PDU has now been read from the socket. */ 2190 nvmf_tcp_pdu_psh_handle(tqpair, ttransport); 2191 break; 2192 /* Wait for the req slot */ 2193 case NVME_TCP_PDU_RECV_STATE_AWAIT_REQ: 2194 nvmf_tcp_capsule_cmd_hdr_handle(ttransport, tqpair, pdu); 2195 break; 2196 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD: 2197 /* check whether the data is valid, if not we just return */ 2198 if (!pdu->data_len) { 2199 return NVME_TCP_PDU_IN_PROGRESS; 2200 } 2201 2202 data_len = pdu->data_len; 2203 /* data digest */ 2204 if (spdk_unlikely((pdu->hdr.common.pdu_type != SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ) && 2205 tqpair->host_ddgst_enable)) { 2206 data_len += SPDK_NVME_TCP_DIGEST_LEN; 2207 pdu->ddgst_enable = true; 2208 } 2209 2210 rc = nvme_tcp_read_payload_data(tqpair->sock, pdu); 2211 if (rc < 0) { 2212 return NVME_TCP_PDU_FATAL; 2213 } 2214 pdu->rw_offset += rc; 2215 2216 if (spdk_unlikely(pdu->dif_ctx != NULL)) { 2217 rc = nvmf_tcp_pdu_payload_insert_dif(pdu, pdu->rw_offset - rc, rc); 2218 if (rc != 0) { 2219 return NVME_TCP_PDU_FATAL; 2220 } 2221 } 2222 2223 if (pdu->rw_offset < data_len) { 2224 return NVME_TCP_PDU_IN_PROGRESS; 2225 } 2226 2227 /* All of this PDU has now been read from the socket. */ 2228 nvmf_tcp_pdu_payload_handle(tqpair, ttransport); 2229 break; 2230 case NVME_TCP_PDU_RECV_STATE_ERROR: 2231 if (!spdk_sock_is_connected(tqpair->sock)) { 2232 return NVME_TCP_PDU_FATAL; 2233 } 2234 break; 2235 default: 2236 assert(0); 2237 SPDK_ERRLOG("code should not come to here"); 2238 break; 2239 } 2240 } while (tqpair->recv_state != prev_state); 2241 2242 return rc; 2243 } 2244 2245 static inline void * 2246 nvmf_tcp_control_msg_get(struct spdk_nvmf_tcp_control_msg_list *list) 2247 { 2248 struct spdk_nvmf_tcp_control_msg *msg; 2249 2250 assert(list); 2251 2252 msg = STAILQ_FIRST(&list->free_msgs); 2253 if (!msg) { 2254 SPDK_DEBUGLOG(nvmf_tcp, "Out of control messages\n"); 2255 return NULL; 2256 } 2257 STAILQ_REMOVE_HEAD(&list->free_msgs, link); 2258 return msg; 2259 } 2260 2261 static inline void 2262 nvmf_tcp_control_msg_put(struct spdk_nvmf_tcp_control_msg_list *list, void *_msg) 2263 { 2264 struct spdk_nvmf_tcp_control_msg *msg = _msg; 2265 2266 assert(list); 2267 STAILQ_INSERT_HEAD(&list->free_msgs, msg, link); 2268 } 2269 2270 static int 2271 nvmf_tcp_req_parse_sgl(struct spdk_nvmf_tcp_req *tcp_req, 2272 struct spdk_nvmf_transport *transport, 2273 struct spdk_nvmf_transport_poll_group *group) 2274 { 2275 struct spdk_nvmf_request *req = &tcp_req->req; 2276 struct spdk_nvme_cmd *cmd; 2277 struct spdk_nvme_cpl *rsp; 2278 struct spdk_nvme_sgl_descriptor *sgl; 2279 struct spdk_nvmf_tcp_poll_group *tgroup; 2280 uint32_t length; 2281 2282 cmd = &req->cmd->nvme_cmd; 2283 rsp = &req->rsp->nvme_cpl; 2284 sgl = &cmd->dptr.sgl1; 2285 2286 length = sgl->unkeyed.length; 2287 2288 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK && 2289 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_TRANSPORT) { 2290 if (length > transport->opts.max_io_size) { 2291 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 2292 length, transport->opts.max_io_size); 2293 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 2294 return -1; 2295 } 2296 2297 /* fill request length and populate iovs */ 2298 req->length = length; 2299 2300 SPDK_DEBUGLOG(nvmf_tcp, "Data requested length= 0x%x\n", length); 2301 2302 if (spdk_unlikely(req->dif_enabled)) { 2303 req->dif.orig_length = length; 2304 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2305 req->dif.elba_length = length; 2306 } 2307 2308 if (nvmf_ctrlr_use_zcopy(req)) { 2309 SPDK_DEBUGLOG(nvmf_tcp, "Using zero-copy to execute request %p\n", tcp_req); 2310 req->data_from_pool = false; 2311 return 0; 2312 } 2313 2314 if (spdk_nvmf_request_get_buffers(req, group, transport, length)) { 2315 /* No available buffers. Queue this request up. */ 2316 SPDK_DEBUGLOG(nvmf_tcp, "No available large data buffers. Queueing request %p\n", 2317 tcp_req); 2318 return 0; 2319 } 2320 2321 /* backward compatible */ 2322 req->data = req->iov[0].iov_base; 2323 2324 SPDK_DEBUGLOG(nvmf_tcp, "Request %p took %d buffer/s from central pool, and data=%p\n", 2325 tcp_req, req->iovcnt, req->data); 2326 2327 return 0; 2328 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 2329 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 2330 uint64_t offset = sgl->address; 2331 uint32_t max_len = transport->opts.in_capsule_data_size; 2332 assert(tcp_req->has_in_capsule_data); 2333 2334 SPDK_DEBUGLOG(nvmf_tcp, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 2335 offset, length); 2336 2337 if (offset > max_len) { 2338 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n", 2339 offset, max_len); 2340 rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET; 2341 return -1; 2342 } 2343 max_len -= (uint32_t)offset; 2344 2345 if (spdk_unlikely(length > max_len)) { 2346 /* According to the SPEC we should support ICD up to 8192 bytes for admin and fabric commands */ 2347 if (length <= SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE && 2348 (cmd->opc == SPDK_NVME_OPC_FABRIC || req->qpair->qid == 0)) { 2349 2350 /* Get a buffer from dedicated list */ 2351 SPDK_DEBUGLOG(nvmf_tcp, "Getting a buffer from control msg list\n"); 2352 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 2353 assert(tgroup->control_msg_list); 2354 req->data = nvmf_tcp_control_msg_get(tgroup->control_msg_list); 2355 if (!req->data) { 2356 /* No available buffers. Queue this request up. */ 2357 SPDK_DEBUGLOG(nvmf_tcp, "No available ICD buffers. Queueing request %p\n", tcp_req); 2358 return 0; 2359 } 2360 } else { 2361 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 2362 length, max_len); 2363 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID; 2364 return -1; 2365 } 2366 } else { 2367 req->data = tcp_req->buf; 2368 } 2369 2370 req->length = length; 2371 req->data_from_pool = false; 2372 2373 if (spdk_unlikely(req->dif_enabled)) { 2374 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2375 req->dif.elba_length = length; 2376 } 2377 2378 req->iov[0].iov_base = req->data; 2379 req->iov[0].iov_len = length; 2380 req->iovcnt = 1; 2381 2382 return 0; 2383 } 2384 2385 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 2386 sgl->generic.type, sgl->generic.subtype); 2387 rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID; 2388 return -1; 2389 } 2390 2391 static inline enum spdk_nvme_media_error_status_code 2392 nvmf_tcp_dif_error_to_compl_status(uint8_t err_type) { 2393 enum spdk_nvme_media_error_status_code result; 2394 2395 switch (err_type) 2396 { 2397 case SPDK_DIF_REFTAG_ERROR: 2398 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 2399 break; 2400 case SPDK_DIF_APPTAG_ERROR: 2401 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 2402 break; 2403 case SPDK_DIF_GUARD_ERROR: 2404 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 2405 break; 2406 default: 2407 SPDK_UNREACHABLE(); 2408 break; 2409 } 2410 2411 return result; 2412 } 2413 2414 static void 2415 _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2416 struct spdk_nvmf_tcp_req *tcp_req) 2417 { 2418 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF( 2419 tqpair->qpair.transport, struct spdk_nvmf_tcp_transport, transport); 2420 struct nvme_tcp_pdu *rsp_pdu; 2421 struct spdk_nvme_tcp_c2h_data_hdr *c2h_data; 2422 uint32_t plen, pdo, alignment; 2423 int rc; 2424 2425 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2426 2427 rsp_pdu = tcp_req->pdu; 2428 assert(rsp_pdu != NULL); 2429 2430 c2h_data = &rsp_pdu->hdr.c2h_data; 2431 c2h_data->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_DATA; 2432 plen = c2h_data->common.hlen = sizeof(*c2h_data); 2433 2434 if (tqpair->host_hdgst_enable) { 2435 plen += SPDK_NVME_TCP_DIGEST_LEN; 2436 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2437 } 2438 2439 /* set the psh */ 2440 c2h_data->cccid = tcp_req->req.cmd->nvme_cmd.cid; 2441 c2h_data->datal = tcp_req->req.length - tcp_req->pdu->rw_offset; 2442 c2h_data->datao = tcp_req->pdu->rw_offset; 2443 2444 /* set the padding */ 2445 rsp_pdu->padding_len = 0; 2446 pdo = plen; 2447 if (tqpair->cpda) { 2448 alignment = (tqpair->cpda + 1) << 2; 2449 if (plen % alignment != 0) { 2450 pdo = (plen + alignment) / alignment * alignment; 2451 rsp_pdu->padding_len = pdo - plen; 2452 plen = pdo; 2453 } 2454 } 2455 2456 c2h_data->common.pdo = pdo; 2457 plen += c2h_data->datal; 2458 if (tqpair->host_ddgst_enable) { 2459 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_DDGSTF; 2460 plen += SPDK_NVME_TCP_DIGEST_LEN; 2461 } 2462 2463 c2h_data->common.plen = plen; 2464 2465 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2466 rsp_pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 2467 } 2468 2469 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2470 c2h_data->datao, c2h_data->datal); 2471 2472 2473 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU; 2474 /* Need to send the capsule response if response is not all 0 */ 2475 if (ttransport->tcp_opts.c2h_success && 2476 tcp_req->rsp.cdw0 == 0 && tcp_req->rsp.cdw1 == 0) { 2477 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS; 2478 } 2479 2480 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2481 struct spdk_nvme_cpl *rsp = &tcp_req->req.rsp->nvme_cpl; 2482 struct spdk_dif_error err_blk = {}; 2483 uint32_t mapped_length = 0; 2484 uint32_t available_iovs = SPDK_COUNTOF(rsp_pdu->iov); 2485 uint32_t ddgst_len = 0; 2486 2487 if (tqpair->host_ddgst_enable) { 2488 /* Data digest consumes additional iov entry */ 2489 available_iovs--; 2490 /* plen needs to be updated since nvme_tcp_build_iovs compares expected and actual plen */ 2491 ddgst_len = SPDK_NVME_TCP_DIGEST_LEN; 2492 c2h_data->common.plen -= ddgst_len; 2493 } 2494 /* Temp call to estimate if data can be described by limited number of iovs. 2495 * iov vector will be rebuilt in nvmf_tcp_qpair_write_pdu */ 2496 nvme_tcp_build_iovs(rsp_pdu->iov, available_iovs, rsp_pdu, tqpair->host_hdgst_enable, 2497 false, &mapped_length); 2498 2499 if (mapped_length != c2h_data->common.plen) { 2500 c2h_data->datal = mapped_length - (c2h_data->common.plen - c2h_data->datal); 2501 SPDK_DEBUGLOG(nvmf_tcp, 2502 "Part C2H, data_len %u (of %u), PDU len %u, updated PDU len %u, offset %u\n", 2503 c2h_data->datal, tcp_req->req.length, c2h_data->common.plen, mapped_length, rsp_pdu->rw_offset); 2504 c2h_data->common.plen = mapped_length; 2505 2506 /* Rebuild pdu->data_iov since data length is changed */ 2507 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, c2h_data->datao, 2508 c2h_data->datal); 2509 2510 c2h_data->common.flags &= ~(SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU | 2511 SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS); 2512 } 2513 2514 c2h_data->common.plen += ddgst_len; 2515 2516 assert(rsp_pdu->rw_offset <= tcp_req->req.length); 2517 2518 rc = spdk_dif_verify_stream(rsp_pdu->data_iov, rsp_pdu->data_iovcnt, 2519 0, rsp_pdu->data_len, rsp_pdu->dif_ctx, &err_blk); 2520 if (rc != 0) { 2521 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", 2522 err_blk.err_type, err_blk.err_offset); 2523 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2524 rsp->status.sc = nvmf_tcp_dif_error_to_compl_status(err_blk.err_type); 2525 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2526 return; 2527 } 2528 } 2529 2530 rsp_pdu->rw_offset += c2h_data->datal; 2531 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_pdu_c2h_data_complete, tcp_req); 2532 } 2533 2534 static void 2535 nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2536 struct spdk_nvmf_tcp_req *tcp_req) 2537 { 2538 nvmf_tcp_req_pdu_init(tcp_req); 2539 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2540 } 2541 2542 static int 2543 request_transfer_out(struct spdk_nvmf_request *req) 2544 { 2545 struct spdk_nvmf_tcp_req *tcp_req; 2546 struct spdk_nvmf_qpair *qpair; 2547 struct spdk_nvmf_tcp_qpair *tqpair; 2548 struct spdk_nvme_cpl *rsp; 2549 2550 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2551 2552 qpair = req->qpair; 2553 rsp = &req->rsp->nvme_cpl; 2554 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 2555 2556 /* Advance our sq_head pointer */ 2557 if (qpair->sq_head == qpair->sq_head_max) { 2558 qpair->sq_head = 0; 2559 } else { 2560 qpair->sq_head++; 2561 } 2562 rsp->sqhd = qpair->sq_head; 2563 2564 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2565 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2566 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2567 nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2568 } else { 2569 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2570 } 2571 2572 return 0; 2573 } 2574 2575 static void 2576 nvmf_tcp_set_in_capsule_data(struct spdk_nvmf_tcp_qpair *tqpair, 2577 struct spdk_nvmf_tcp_req *tcp_req) 2578 { 2579 struct nvme_tcp_pdu *pdu; 2580 uint32_t plen = 0; 2581 2582 pdu = tqpair->pdu_in_progress; 2583 plen = pdu->hdr.common.hlen; 2584 2585 if (tqpair->host_hdgst_enable) { 2586 plen += SPDK_NVME_TCP_DIGEST_LEN; 2587 } 2588 2589 if (pdu->hdr.common.plen != plen) { 2590 tcp_req->has_in_capsule_data = true; 2591 } 2592 } 2593 2594 static void 2595 nvmf_tcp_check_fused_ordering(struct spdk_nvmf_tcp_transport *ttransport, 2596 struct spdk_nvmf_tcp_qpair *tqpair, 2597 struct spdk_nvmf_tcp_req *tcp_req) 2598 { 2599 enum spdk_nvme_cmd_fuse last, next; 2600 2601 last = tqpair->fused_first ? tqpair->fused_first->cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2602 next = tcp_req->cmd.fuse; 2603 2604 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2605 2606 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2607 return; 2608 } 2609 2610 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2611 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2612 /* This is a valid pair of fused commands. Point them at each other 2613 * so they can be submitted consecutively once ready to be executed. 2614 */ 2615 tqpair->fused_first->fused_pair = tcp_req; 2616 tcp_req->fused_pair = tqpair->fused_first; 2617 tqpair->fused_first = NULL; 2618 return; 2619 } else { 2620 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2621 tqpair->fused_first->fused_failed = true; 2622 2623 /* 2624 * If the last req is in READY_TO_EXECUTE state, then call 2625 * nvmf_tcp_req_process(), otherwise nothing else will kick it. 2626 */ 2627 if (tqpair->fused_first->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 2628 nvmf_tcp_req_process(ttransport, tqpair->fused_first); 2629 } 2630 2631 tqpair->fused_first = NULL; 2632 } 2633 } 2634 2635 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2636 /* Set tqpair->fused_first here so that we know to check that the next request 2637 * is a SECOND (and to fail this one if it isn't). 2638 */ 2639 tqpair->fused_first = tcp_req; 2640 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2641 /* Mark this req failed since it is a SECOND and the last one was not a FIRST. */ 2642 tcp_req->fused_failed = true; 2643 } 2644 } 2645 2646 static bool 2647 nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 2648 struct spdk_nvmf_tcp_req *tcp_req) 2649 { 2650 struct spdk_nvmf_tcp_qpair *tqpair; 2651 int rc; 2652 enum spdk_nvmf_tcp_req_state prev_state; 2653 bool progress = false; 2654 struct spdk_nvmf_transport *transport = &ttransport->transport; 2655 struct spdk_nvmf_transport_poll_group *group; 2656 struct spdk_nvmf_tcp_poll_group *tgroup; 2657 2658 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2659 group = &tqpair->group->group; 2660 assert(tcp_req->state != TCP_REQUEST_STATE_FREE); 2661 2662 /* If the qpair is not active, we need to abort the outstanding requests. */ 2663 if (tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2664 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 2665 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2666 } 2667 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 2668 } 2669 2670 /* The loop here is to allow for several back-to-back state changes. */ 2671 do { 2672 prev_state = tcp_req->state; 2673 2674 SPDK_DEBUGLOG(nvmf_tcp, "Request %p entering state %d on tqpair=%p\n", tcp_req, prev_state, 2675 tqpair); 2676 2677 switch (tcp_req->state) { 2678 case TCP_REQUEST_STATE_FREE: 2679 /* Some external code must kick a request into TCP_REQUEST_STATE_NEW 2680 * to escape this state. */ 2681 break; 2682 case TCP_REQUEST_STATE_NEW: 2683 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEW, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, tqpair); 2684 2685 /* copy the cmd from the receive pdu */ 2686 tcp_req->cmd = tqpair->pdu_in_progress->hdr.capsule_cmd.ccsqe; 2687 2688 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&tcp_req->req, &tcp_req->req.dif.dif_ctx))) { 2689 tcp_req->req.dif_enabled = true; 2690 tqpair->pdu_in_progress->dif_ctx = &tcp_req->req.dif.dif_ctx; 2691 } 2692 2693 nvmf_tcp_check_fused_ordering(ttransport, tqpair, tcp_req); 2694 2695 /* The next state transition depends on the data transfer needs of this request. */ 2696 tcp_req->req.xfer = spdk_nvmf_req_get_xfer(&tcp_req->req); 2697 2698 if (spdk_unlikely(tcp_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2699 tcp_req->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 2700 tcp_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INVALID_OPCODE; 2701 tcp_req->req.rsp->nvme_cpl.cid = tcp_req->req.cmd->nvme_cmd.cid; 2702 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2703 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2704 SPDK_DEBUGLOG(nvmf_tcp, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", tcp_req); 2705 break; 2706 } 2707 2708 /* If no data to transfer, ready to execute. */ 2709 if (tcp_req->req.xfer == SPDK_NVME_DATA_NONE) { 2710 /* Reset the tqpair receiving pdu state */ 2711 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2712 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2713 break; 2714 } 2715 2716 nvmf_tcp_set_in_capsule_data(tqpair, tcp_req); 2717 2718 if (!tcp_req->has_in_capsule_data) { 2719 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2720 } 2721 2722 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEED_BUFFER); 2723 STAILQ_INSERT_TAIL(&group->pending_buf_queue, &tcp_req->req, buf_link); 2724 break; 2725 case TCP_REQUEST_STATE_NEED_BUFFER: 2726 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEED_BUFFER, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2727 tqpair); 2728 2729 assert(tcp_req->req.xfer != SPDK_NVME_DATA_NONE); 2730 2731 if (!tcp_req->has_in_capsule_data && (&tcp_req->req != STAILQ_FIRST(&group->pending_buf_queue))) { 2732 SPDK_DEBUGLOG(nvmf_tcp, 2733 "Not the first element to wait for the buf for tcp_req(%p) on tqpair=%p\n", 2734 tcp_req, tqpair); 2735 /* This request needs to wait in line to obtain a buffer */ 2736 break; 2737 } 2738 2739 /* Try to get a data buffer */ 2740 rc = nvmf_tcp_req_parse_sgl(tcp_req, transport, group); 2741 if (rc < 0) { 2742 STAILQ_REMOVE_HEAD(&group->pending_buf_queue, buf_link); 2743 /* Reset the tqpair receiving pdu state */ 2744 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 2745 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2746 tcp_req->req.rsp->nvme_cpl.cid = tcp_req->req.cmd->nvme_cmd.cid; 2747 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2748 break; 2749 } 2750 2751 /* Get a zcopy buffer if the request can be serviced through zcopy */ 2752 if (spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 2753 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2754 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 2755 tcp_req->req.length = tcp_req->req.dif.elba_length; 2756 } 2757 2758 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2759 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_START); 2760 spdk_nvmf_request_zcopy_start(&tcp_req->req); 2761 break; 2762 } 2763 2764 if (!tcp_req->req.data) { 2765 SPDK_DEBUGLOG(nvmf_tcp, "No buffer allocated for tcp_req(%p) on tqpair(%p\n)", 2766 tcp_req, tqpair); 2767 /* No buffers available. */ 2768 break; 2769 } 2770 2771 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2772 2773 /* If data is transferring from host to controller, we need to do a transfer from the host. */ 2774 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2775 if (tcp_req->req.data_from_pool) { 2776 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 2777 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 2778 } else { 2779 struct nvme_tcp_pdu *pdu; 2780 2781 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2782 2783 pdu = tqpair->pdu_in_progress; 2784 SPDK_DEBUGLOG(nvmf_tcp, "Not need to send r2t for tcp_req(%p) on tqpair=%p\n", tcp_req, 2785 tqpair); 2786 /* No need to send r2t, contained in the capsuled data */ 2787 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2788 0, tcp_req->req.length); 2789 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2790 } 2791 break; 2792 } 2793 2794 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2795 break; 2796 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 2797 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, tqpair->qpair.qid, 0, 2798 (uintptr_t)tcp_req, tqpair); 2799 /* Some external code must kick a request into TCP_REQUEST_STATE_ZCOPY_START_COMPLETED 2800 * to escape this state. */ 2801 break; 2802 case TCP_REQUEST_STATE_ZCOPY_START_COMPLETED: 2803 spdk_trace_record(TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, tqpair->qpair.qid, 0, 2804 (uintptr_t)tcp_req, tqpair); 2805 if (spdk_unlikely(spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl))) { 2806 SPDK_DEBUGLOG(nvmf_tcp, "Zero-copy start failed for tcp_req(%p) on tqpair=%p\n", 2807 tcp_req, tqpair); 2808 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2809 break; 2810 } 2811 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2812 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 2813 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 2814 } else { 2815 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 2816 } 2817 break; 2818 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 2819 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2820 tqpair); 2821 /* The R2T completion or the h2c data incoming will kick it out of this state. */ 2822 break; 2823 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 2824 2825 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, tqpair->qpair.qid, 0, 2826 (uintptr_t)tcp_req, tqpair); 2827 /* Some external code must kick a request into TCP_REQUEST_STATE_READY_TO_EXECUTE 2828 * to escape this state. */ 2829 break; 2830 case TCP_REQUEST_STATE_READY_TO_EXECUTE: 2831 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, tqpair->qpair.qid, 0, 2832 (uintptr_t)tcp_req, tqpair); 2833 2834 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2835 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 2836 tcp_req->req.length = tcp_req->req.dif.elba_length; 2837 } 2838 2839 if (tcp_req->cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 2840 if (tcp_req->fused_failed) { 2841 /* This request failed FUSED semantics. Fail it immediately, without 2842 * even sending it to the target layer. 2843 */ 2844 tcp_req->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 2845 tcp_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED; 2846 tcp_req->req.rsp->nvme_cpl.cid = tcp_req->req.cmd->nvme_cmd.cid; 2847 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2848 break; 2849 } 2850 2851 if (tcp_req->fused_pair == NULL || 2852 tcp_req->fused_pair->state != TCP_REQUEST_STATE_READY_TO_EXECUTE) { 2853 /* This request is ready to execute, but either we don't know yet if it's 2854 * valid - i.e. this is a FIRST but we haven't received the next request yet), 2855 * or the other request of this fused pair isn't ready to execute. So 2856 * break here and this request will get processed later either when the 2857 * other request is ready or we find that this request isn't valid. 2858 */ 2859 break; 2860 } 2861 } 2862 2863 if (!spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 2864 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTING); 2865 /* If we get to this point, and this request is a fused command, we know that 2866 * it is part of a valid sequence (FIRST followed by a SECOND) and that both 2867 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 2868 * request, and the other request of the fused pair, in the correct order. 2869 * Also clear the ->fused_pair pointers on both requests, since after this point 2870 * we no longer need to maintain the relationship between these two requests. 2871 */ 2872 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 2873 assert(tcp_req->fused_pair != NULL); 2874 assert(tcp_req->fused_pair->fused_pair == tcp_req); 2875 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 2876 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 2877 tcp_req->fused_pair->fused_pair = NULL; 2878 tcp_req->fused_pair = NULL; 2879 } 2880 spdk_nvmf_request_exec(&tcp_req->req); 2881 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 2882 assert(tcp_req->fused_pair != NULL); 2883 assert(tcp_req->fused_pair->fused_pair == tcp_req); 2884 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 2885 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 2886 tcp_req->fused_pair->fused_pair = NULL; 2887 tcp_req->fused_pair = NULL; 2888 } 2889 } else { 2890 /* For zero-copy, only requests with data coming from host to the 2891 * controller can end up here. */ 2892 assert(tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 2893 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT); 2894 spdk_nvmf_request_zcopy_end(&tcp_req->req, true); 2895 } 2896 2897 break; 2898 case TCP_REQUEST_STATE_EXECUTING: 2899 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTING, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2900 tqpair); 2901 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 2902 * to escape this state. */ 2903 break; 2904 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 2905 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, tqpair->qpair.qid, 0, 2906 (uintptr_t)tcp_req, tqpair); 2907 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 2908 * to escape this state. */ 2909 break; 2910 case TCP_REQUEST_STATE_EXECUTED: 2911 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTED, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2912 tqpair); 2913 2914 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2915 tcp_req->req.length = tcp_req->req.dif.orig_length; 2916 } 2917 2918 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2919 break; 2920 case TCP_REQUEST_STATE_READY_TO_COMPLETE: 2921 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, tqpair->qpair.qid, 0, 2922 (uintptr_t)tcp_req, tqpair); 2923 rc = request_transfer_out(&tcp_req->req); 2924 assert(rc == 0); /* No good way to handle this currently */ 2925 break; 2926 case TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 2927 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, tqpair->qpair.qid, 0, 2928 (uintptr_t)tcp_req, tqpair); 2929 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 2930 * to escape this state. */ 2931 break; 2932 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 2933 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, tqpair->qpair.qid, 0, 2934 (uintptr_t)tcp_req, tqpair); 2935 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 2936 * to escape this state. */ 2937 break; 2938 case TCP_REQUEST_STATE_COMPLETED: 2939 spdk_trace_record(TRACE_TCP_REQUEST_STATE_COMPLETED, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2940 tqpair); 2941 /* If there's an outstanding PDU sent to the host, the request is completed 2942 * due to the qpair being disconnected. We must delay the completion until 2943 * that write is done to avoid freeing the request twice. */ 2944 if (spdk_unlikely(tcp_req->pdu_in_use)) { 2945 SPDK_DEBUGLOG(nvmf_tcp, "Delaying completion due to outstanding " 2946 "write on req=%p\n", tcp_req); 2947 /* This can only happen for zcopy requests */ 2948 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 2949 assert(tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE); 2950 break; 2951 } 2952 2953 if (tcp_req->req.data_from_pool) { 2954 spdk_nvmf_request_free_buffers(&tcp_req->req, group, transport); 2955 } else if (spdk_unlikely(tcp_req->has_in_capsule_data && 2956 (tcp_req->cmd.opc == SPDK_NVME_OPC_FABRIC || 2957 tqpair->qpair.qid == 0) && tcp_req->req.length > transport->opts.in_capsule_data_size)) { 2958 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 2959 assert(tgroup->control_msg_list); 2960 SPDK_DEBUGLOG(nvmf_tcp, "Put buf to control msg list\n"); 2961 nvmf_tcp_control_msg_put(tgroup->control_msg_list, tcp_req->req.data); 2962 } else if (tcp_req->req.zcopy_bdev_io != NULL) { 2963 /* If the request has an unreleased zcopy bdev_io, it's either a 2964 * read, a failed write, or the qpair is being disconnected */ 2965 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 2966 assert(tcp_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST || 2967 spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl) || 2968 tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE); 2969 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE); 2970 spdk_nvmf_request_zcopy_end(&tcp_req->req, false); 2971 break; 2972 } 2973 tcp_req->req.length = 0; 2974 tcp_req->req.iovcnt = 0; 2975 tcp_req->req.data = NULL; 2976 tcp_req->fused_failed = false; 2977 if (tcp_req->fused_pair) { 2978 /* This req was part of a valid fused pair, but failed before it got to 2979 * READ_TO_EXECUTE state. This means we need to fail the other request 2980 * in the pair, because it is no longer part of a valid pair. If the pair 2981 * already reached READY_TO_EXECUTE state, we need to kick it. 2982 */ 2983 tcp_req->fused_pair->fused_failed = true; 2984 if (tcp_req->fused_pair->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 2985 nvmf_tcp_req_process(ttransport, tcp_req->fused_pair); 2986 } 2987 tcp_req->fused_pair = NULL; 2988 } 2989 2990 nvmf_tcp_req_put(tqpair, tcp_req); 2991 break; 2992 case TCP_REQUEST_NUM_STATES: 2993 default: 2994 assert(0); 2995 break; 2996 } 2997 2998 if (tcp_req->state != prev_state) { 2999 progress = true; 3000 } 3001 } while (tcp_req->state != prev_state); 3002 3003 return progress; 3004 } 3005 3006 static void 3007 nvmf_tcp_sock_cb(void *arg, struct spdk_sock_group *group, struct spdk_sock *sock) 3008 { 3009 struct spdk_nvmf_tcp_qpair *tqpair = arg; 3010 int rc; 3011 3012 assert(tqpair != NULL); 3013 rc = nvmf_tcp_sock_process(tqpair); 3014 3015 /* If there was a new socket error, disconnect */ 3016 if (rc < 0) { 3017 nvmf_tcp_qpair_disconnect(tqpair); 3018 } 3019 } 3020 3021 static int 3022 nvmf_tcp_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3023 struct spdk_nvmf_qpair *qpair) 3024 { 3025 struct spdk_nvmf_tcp_poll_group *tgroup; 3026 struct spdk_nvmf_tcp_qpair *tqpair; 3027 int rc; 3028 3029 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3030 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3031 3032 rc = nvmf_tcp_qpair_sock_init(tqpair); 3033 if (rc != 0) { 3034 SPDK_ERRLOG("Cannot set sock opt for tqpair=%p\n", tqpair); 3035 return -1; 3036 } 3037 3038 rc = nvmf_tcp_qpair_init(&tqpair->qpair); 3039 if (rc < 0) { 3040 SPDK_ERRLOG("Cannot init tqpair=%p\n", tqpair); 3041 return -1; 3042 } 3043 3044 rc = nvmf_tcp_qpair_init_mem_resource(tqpair); 3045 if (rc < 0) { 3046 SPDK_ERRLOG("Cannot init memory resource info for tqpair=%p\n", tqpair); 3047 return -1; 3048 } 3049 3050 rc = spdk_sock_group_add_sock(tgroup->sock_group, tqpair->sock, 3051 nvmf_tcp_sock_cb, tqpair); 3052 if (rc != 0) { 3053 SPDK_ERRLOG("Could not add sock to sock_group: %s (%d)\n", 3054 spdk_strerror(errno), errno); 3055 return -1; 3056 } 3057 3058 tqpair->group = tgroup; 3059 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INVALID); 3060 TAILQ_INSERT_TAIL(&tgroup->qpairs, tqpair, link); 3061 3062 return 0; 3063 } 3064 3065 static int 3066 nvmf_tcp_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3067 struct spdk_nvmf_qpair *qpair) 3068 { 3069 struct spdk_nvmf_tcp_poll_group *tgroup; 3070 struct spdk_nvmf_tcp_qpair *tqpair; 3071 int rc; 3072 3073 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3074 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3075 3076 assert(tqpair->group == tgroup); 3077 3078 SPDK_DEBUGLOG(nvmf_tcp, "remove tqpair=%p from the tgroup=%p\n", tqpair, tgroup); 3079 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 3080 TAILQ_REMOVE(&tgroup->await_req, tqpair, link); 3081 } else { 3082 TAILQ_REMOVE(&tgroup->qpairs, tqpair, link); 3083 } 3084 3085 rc = spdk_sock_group_remove_sock(tgroup->sock_group, tqpair->sock); 3086 if (rc != 0) { 3087 SPDK_ERRLOG("Could not remove sock from sock_group: %s (%d)\n", 3088 spdk_strerror(errno), errno); 3089 } 3090 3091 return rc; 3092 } 3093 3094 static int 3095 nvmf_tcp_req_complete(struct spdk_nvmf_request *req) 3096 { 3097 struct spdk_nvmf_tcp_transport *ttransport; 3098 struct spdk_nvmf_tcp_req *tcp_req; 3099 3100 ttransport = SPDK_CONTAINEROF(req->qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3101 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3102 3103 switch (tcp_req->state) { 3104 case TCP_REQUEST_STATE_EXECUTING: 3105 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3106 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3107 break; 3108 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3109 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_ZCOPY_START_COMPLETED); 3110 break; 3111 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3112 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 3113 break; 3114 default: 3115 assert(0 && "Unexpected request state"); 3116 break; 3117 } 3118 3119 nvmf_tcp_req_process(ttransport, tcp_req); 3120 3121 return 0; 3122 } 3123 3124 static void 3125 nvmf_tcp_close_qpair(struct spdk_nvmf_qpair *qpair, 3126 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3127 { 3128 struct spdk_nvmf_tcp_qpair *tqpair; 3129 3130 SPDK_DEBUGLOG(nvmf_tcp, "Qpair: %p\n", qpair); 3131 3132 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3133 3134 assert(tqpair->fini_cb_fn == NULL); 3135 tqpair->fini_cb_fn = cb_fn; 3136 tqpair->fini_cb_arg = cb_arg; 3137 3138 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITED); 3139 nvmf_tcp_qpair_destroy(tqpair); 3140 } 3141 3142 static int 3143 nvmf_tcp_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3144 { 3145 struct spdk_nvmf_tcp_poll_group *tgroup; 3146 int rc; 3147 struct spdk_nvmf_request *req, *req_tmp; 3148 struct spdk_nvmf_tcp_req *tcp_req; 3149 struct spdk_nvmf_tcp_qpair *tqpair, *tqpair_tmp; 3150 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(group->transport, 3151 struct spdk_nvmf_tcp_transport, transport); 3152 3153 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3154 3155 if (spdk_unlikely(TAILQ_EMPTY(&tgroup->qpairs) && TAILQ_EMPTY(&tgroup->await_req))) { 3156 return 0; 3157 } 3158 3159 STAILQ_FOREACH_SAFE(req, &group->pending_buf_queue, buf_link, req_tmp) { 3160 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3161 if (nvmf_tcp_req_process(ttransport, tcp_req) == false) { 3162 break; 3163 } 3164 } 3165 3166 rc = spdk_sock_group_poll(tgroup->sock_group); 3167 if (rc < 0) { 3168 SPDK_ERRLOG("Failed to poll sock_group=%p\n", tgroup->sock_group); 3169 } 3170 3171 TAILQ_FOREACH_SAFE(tqpair, &tgroup->await_req, link, tqpair_tmp) { 3172 nvmf_tcp_sock_process(tqpair); 3173 } 3174 3175 return rc; 3176 } 3177 3178 static int 3179 nvmf_tcp_qpair_get_trid(struct spdk_nvmf_qpair *qpair, 3180 struct spdk_nvme_transport_id *trid, bool peer) 3181 { 3182 struct spdk_nvmf_tcp_qpair *tqpair; 3183 uint16_t port; 3184 3185 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3186 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_TCP); 3187 3188 if (peer) { 3189 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->initiator_addr); 3190 port = tqpair->initiator_port; 3191 } else { 3192 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->target_addr); 3193 port = tqpair->target_port; 3194 } 3195 3196 if (spdk_sock_is_ipv4(tqpair->sock)) { 3197 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3198 } else if (spdk_sock_is_ipv6(tqpair->sock)) { 3199 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3200 } else { 3201 return -1; 3202 } 3203 3204 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%d", port); 3205 return 0; 3206 } 3207 3208 static int 3209 nvmf_tcp_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3210 struct spdk_nvme_transport_id *trid) 3211 { 3212 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3213 } 3214 3215 static int 3216 nvmf_tcp_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3217 struct spdk_nvme_transport_id *trid) 3218 { 3219 return nvmf_tcp_qpair_get_trid(qpair, trid, 1); 3220 } 3221 3222 static int 3223 nvmf_tcp_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3224 struct spdk_nvme_transport_id *trid) 3225 { 3226 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3227 } 3228 3229 static void 3230 nvmf_tcp_req_set_abort_status(struct spdk_nvmf_request *req, 3231 struct spdk_nvmf_tcp_req *tcp_req_to_abort) 3232 { 3233 tcp_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC; 3234 tcp_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST; 3235 tcp_req_to_abort->req.rsp->nvme_cpl.cid = tcp_req_to_abort->req.cmd->nvme_cmd.cid; 3236 3237 nvmf_tcp_req_set_state(tcp_req_to_abort, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3238 3239 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 3240 } 3241 3242 static int 3243 _nvmf_tcp_qpair_abort_request(void *ctx) 3244 { 3245 struct spdk_nvmf_request *req = ctx; 3246 struct spdk_nvmf_tcp_req *tcp_req_to_abort = SPDK_CONTAINEROF(req->req_to_abort, 3247 struct spdk_nvmf_tcp_req, req); 3248 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 3249 struct spdk_nvmf_tcp_qpair, qpair); 3250 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 3251 struct spdk_nvmf_tcp_transport, transport); 3252 int rc; 3253 3254 spdk_poller_unregister(&req->poller); 3255 3256 switch (tcp_req_to_abort->state) { 3257 case TCP_REQUEST_STATE_EXECUTING: 3258 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3259 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3260 rc = nvmf_ctrlr_abort_request(req); 3261 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 3262 return SPDK_POLLER_BUSY; 3263 } 3264 break; 3265 3266 case TCP_REQUEST_STATE_NEED_BUFFER: 3267 STAILQ_REMOVE(&tqpair->group->group.pending_buf_queue, 3268 &tcp_req_to_abort->req, spdk_nvmf_request, buf_link); 3269 3270 nvmf_tcp_req_set_abort_status(req, tcp_req_to_abort); 3271 nvmf_tcp_req_process(ttransport, tcp_req_to_abort); 3272 break; 3273 3274 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3275 case TCP_REQUEST_STATE_ZCOPY_START_COMPLETED: 3276 nvmf_tcp_req_set_abort_status(req, tcp_req_to_abort); 3277 break; 3278 3279 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3280 if (spdk_get_ticks() < req->timeout_tsc) { 3281 req->poller = SPDK_POLLER_REGISTER(_nvmf_tcp_qpair_abort_request, req, 0); 3282 return SPDK_POLLER_BUSY; 3283 } 3284 break; 3285 3286 default: 3287 break; 3288 } 3289 3290 spdk_nvmf_request_complete(req); 3291 return SPDK_POLLER_BUSY; 3292 } 3293 3294 static void 3295 nvmf_tcp_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 3296 struct spdk_nvmf_request *req) 3297 { 3298 struct spdk_nvmf_tcp_qpair *tqpair; 3299 struct spdk_nvmf_tcp_transport *ttransport; 3300 struct spdk_nvmf_transport *transport; 3301 uint16_t cid; 3302 uint32_t i; 3303 struct spdk_nvmf_tcp_req *tcp_req_to_abort = NULL; 3304 3305 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3306 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3307 transport = &ttransport->transport; 3308 3309 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 3310 3311 for (i = 0; i < tqpair->resource_count; i++) { 3312 if (tqpair->reqs[i].state != TCP_REQUEST_STATE_FREE && 3313 tqpair->reqs[i].req.cmd->nvme_cmd.cid == cid) { 3314 tcp_req_to_abort = &tqpair->reqs[i]; 3315 break; 3316 } 3317 } 3318 3319 spdk_trace_record(TRACE_TCP_QP_ABORT_REQ, qpair->qid, 0, (uintptr_t)req, tqpair); 3320 3321 if (tcp_req_to_abort == NULL) { 3322 spdk_nvmf_request_complete(req); 3323 return; 3324 } 3325 3326 req->req_to_abort = &tcp_req_to_abort->req; 3327 req->timeout_tsc = spdk_get_ticks() + 3328 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 3329 req->poller = NULL; 3330 3331 _nvmf_tcp_qpair_abort_request(req); 3332 } 3333 3334 #define SPDK_NVMF_TCP_DEFAULT_MAX_QUEUE_DEPTH 128 3335 #define SPDK_NVMF_TCP_DEFAULT_AQ_DEPTH 128 3336 #define SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 3337 #define SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 3338 #define SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE 131072 3339 #define SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE 131072 3340 #define SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS 511 3341 #define SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE 32 3342 #define SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP false 3343 #define SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC 1 3344 3345 static void 3346 nvmf_tcp_opts_init(struct spdk_nvmf_transport_opts *opts) 3347 { 3348 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_QUEUE_DEPTH; 3349 opts->max_qpairs_per_ctrlr = SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR; 3350 opts->in_capsule_data_size = SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE; 3351 opts->max_io_size = SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE; 3352 opts->io_unit_size = SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE; 3353 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_AQ_DEPTH; 3354 opts->num_shared_buffers = SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS; 3355 opts->buf_cache_size = SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE; 3356 opts->dif_insert_or_strip = SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP; 3357 opts->abort_timeout_sec = SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC; 3358 opts->transport_specific = NULL; 3359 } 3360 3361 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp = { 3362 .name = "TCP", 3363 .type = SPDK_NVME_TRANSPORT_TCP, 3364 .opts_init = nvmf_tcp_opts_init, 3365 .create = nvmf_tcp_create, 3366 .dump_opts = nvmf_tcp_dump_opts, 3367 .destroy = nvmf_tcp_destroy, 3368 3369 .listen = nvmf_tcp_listen, 3370 .stop_listen = nvmf_tcp_stop_listen, 3371 3372 .listener_discover = nvmf_tcp_discover, 3373 3374 .poll_group_create = nvmf_tcp_poll_group_create, 3375 .get_optimal_poll_group = nvmf_tcp_get_optimal_poll_group, 3376 .poll_group_destroy = nvmf_tcp_poll_group_destroy, 3377 .poll_group_add = nvmf_tcp_poll_group_add, 3378 .poll_group_remove = nvmf_tcp_poll_group_remove, 3379 .poll_group_poll = nvmf_tcp_poll_group_poll, 3380 3381 .req_free = nvmf_tcp_req_free, 3382 .req_complete = nvmf_tcp_req_complete, 3383 3384 .qpair_fini = nvmf_tcp_close_qpair, 3385 .qpair_get_local_trid = nvmf_tcp_qpair_get_local_trid, 3386 .qpair_get_peer_trid = nvmf_tcp_qpair_get_peer_trid, 3387 .qpair_get_listen_trid = nvmf_tcp_qpair_get_listen_trid, 3388 .qpair_abort_request = nvmf_tcp_qpair_abort_request, 3389 }; 3390 3391 SPDK_NVMF_TRANSPORT_REGISTER(tcp, &spdk_nvmf_transport_tcp); 3392 SPDK_LOG_REGISTER_COMPONENT(nvmf_tcp) 3393