1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (C) 2018 Intel Corporation. All rights reserved. 3 * Copyright (c) 2019, 2020 Mellanox Technologies LTD. All rights reserved. 4 * Copyright (c) 2022, 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. 5 */ 6 7 #include "spdk/accel.h" 8 #include "spdk/stdinc.h" 9 #include "spdk/crc32.h" 10 #include "spdk/endian.h" 11 #include "spdk/assert.h" 12 #include "spdk/thread.h" 13 #include "spdk/nvmf_transport.h" 14 #include "spdk/string.h" 15 #include "spdk/trace.h" 16 #include "spdk/util.h" 17 #include "spdk/log.h" 18 #include "spdk/keyring.h" 19 20 #include "spdk_internal/assert.h" 21 #include "spdk_internal/nvme_tcp.h" 22 #include "spdk_internal/sock.h" 23 24 #include "nvmf_internal.h" 25 26 #include "spdk_internal/trace_defs.h" 27 28 #define NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME 16 29 #define SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY 16 30 #define SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY 0 31 #define SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM 32 32 #define SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION true 33 34 #define SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH 2 35 #define SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH 65535 36 #define SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH 2 37 #define SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH 4096 38 39 #define SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH 128 40 #define SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH 128 41 #define SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 42 #define SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 43 #define SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE 131072 44 #define SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE 131072 45 #define SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS 511 46 #define SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX 47 #define SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP false 48 #define SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC 1 49 50 #define TCP_PSK_INVALID_PERMISSIONS 0177 51 52 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp; 53 static bool g_tls_log = false; 54 55 /* spdk nvmf related structure */ 56 enum spdk_nvmf_tcp_req_state { 57 58 /* The request is not currently in use */ 59 TCP_REQUEST_STATE_FREE = 0, 60 61 /* Initial state when request first received */ 62 TCP_REQUEST_STATE_NEW = 1, 63 64 /* The request is queued until a data buffer is available. */ 65 TCP_REQUEST_STATE_NEED_BUFFER = 2, 66 67 /* The request is waiting for zcopy_start to finish */ 68 TCP_REQUEST_STATE_AWAITING_ZCOPY_START = 3, 69 70 /* The request has received a zero-copy buffer */ 71 TCP_REQUEST_STATE_ZCOPY_START_COMPLETED = 4, 72 73 /* The request is currently transferring data from the host to the controller. */ 74 TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER = 5, 75 76 /* The request is waiting for the R2T send acknowledgement. */ 77 TCP_REQUEST_STATE_AWAITING_R2T_ACK = 6, 78 79 /* The request is ready to execute at the block device */ 80 TCP_REQUEST_STATE_READY_TO_EXECUTE = 7, 81 82 /* The request is currently executing at the block device */ 83 TCP_REQUEST_STATE_EXECUTING = 8, 84 85 /* The request is waiting for zcopy buffers to be committed */ 86 TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT = 9, 87 88 /* The request finished executing at the block device */ 89 TCP_REQUEST_STATE_EXECUTED = 10, 90 91 /* The request is ready to send a completion */ 92 TCP_REQUEST_STATE_READY_TO_COMPLETE = 11, 93 94 /* The request is currently transferring final pdus from the controller to the host. */ 95 TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST = 12, 96 97 /* The request is waiting for zcopy buffers to be released (without committing) */ 98 TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE = 13, 99 100 /* The request completed and can be marked free. */ 101 TCP_REQUEST_STATE_COMPLETED = 14, 102 103 /* Terminator */ 104 TCP_REQUEST_NUM_STATES, 105 }; 106 107 static const char *spdk_nvmf_tcp_term_req_fes_str[] = { 108 "Invalid PDU Header Field", 109 "PDU Sequence Error", 110 "Header Digiest Error", 111 "Data Transfer Out of Range", 112 "R2T Limit Exceeded", 113 "Unsupported parameter", 114 }; 115 116 SPDK_TRACE_REGISTER_FN(nvmf_tcp_trace, "nvmf_tcp", TRACE_GROUP_NVMF_TCP) 117 { 118 spdk_trace_register_owner(OWNER_NVMF_TCP, 't'); 119 spdk_trace_register_object(OBJECT_NVMF_TCP_IO, 'r'); 120 spdk_trace_register_description("TCP_REQ_NEW", 121 TRACE_TCP_REQUEST_STATE_NEW, 122 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 1, 123 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 124 spdk_trace_register_description("TCP_REQ_NEED_BUFFER", 125 TRACE_TCP_REQUEST_STATE_NEED_BUFFER, 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_START", 129 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, 130 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 131 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 132 spdk_trace_register_description("TCP_REQ_ZCPY_START_CPL", 133 TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, 134 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 135 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 136 spdk_trace_register_description("TCP_REQ_TX_H_TO_C", 137 TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 138 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 139 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 140 spdk_trace_register_description("TCP_REQ_RDY_TO_EXECUTE", 141 TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, 142 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 143 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 144 spdk_trace_register_description("TCP_REQ_EXECUTING", 145 TRACE_TCP_REQUEST_STATE_EXECUTING, 146 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 147 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 148 spdk_trace_register_description("TCP_REQ_WAIT_ZCPY_CMT", 149 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, 150 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 151 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 152 spdk_trace_register_description("TCP_REQ_EXECUTED", 153 TRACE_TCP_REQUEST_STATE_EXECUTED, 154 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 155 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 156 spdk_trace_register_description("TCP_REQ_RDY_TO_COMPLETE", 157 TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, 158 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 159 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 160 spdk_trace_register_description("TCP_REQ_TRANSFER_C2H", 161 TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 162 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 163 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 164 spdk_trace_register_description("TCP_REQ_AWAIT_ZCPY_RLS", 165 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, 166 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 167 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 168 spdk_trace_register_description("TCP_REQ_COMPLETED", 169 TRACE_TCP_REQUEST_STATE_COMPLETED, 170 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 171 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 172 spdk_trace_register_description("TCP_READ_DONE", 173 TRACE_TCP_READ_FROM_SOCKET_DONE, 174 OWNER_NVMF_TCP, OBJECT_NONE, 0, 175 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 176 spdk_trace_register_description("TCP_REQ_AWAIT_R2T_ACK", 177 TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, 178 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 179 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 180 181 spdk_trace_register_description("TCP_QP_CREATE", TRACE_TCP_QP_CREATE, 182 OWNER_NVMF_TCP, OBJECT_NONE, 0, 183 SPDK_TRACE_ARG_TYPE_INT, ""); 184 spdk_trace_register_description("TCP_QP_SOCK_INIT", TRACE_TCP_QP_SOCK_INIT, 185 OWNER_NVMF_TCP, OBJECT_NONE, 0, 186 SPDK_TRACE_ARG_TYPE_INT, ""); 187 spdk_trace_register_description("TCP_QP_STATE_CHANGE", TRACE_TCP_QP_STATE_CHANGE, 188 OWNER_NVMF_TCP, OBJECT_NONE, 0, 189 SPDK_TRACE_ARG_TYPE_INT, "state"); 190 spdk_trace_register_description("TCP_QP_DISCONNECT", TRACE_TCP_QP_DISCONNECT, 191 OWNER_NVMF_TCP, OBJECT_NONE, 0, 192 SPDK_TRACE_ARG_TYPE_INT, ""); 193 spdk_trace_register_description("TCP_QP_DESTROY", TRACE_TCP_QP_DESTROY, 194 OWNER_NVMF_TCP, OBJECT_NONE, 0, 195 SPDK_TRACE_ARG_TYPE_INT, ""); 196 spdk_trace_register_description("TCP_QP_ABORT_REQ", TRACE_TCP_QP_ABORT_REQ, 197 OWNER_NVMF_TCP, OBJECT_NONE, 0, 198 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 199 spdk_trace_register_description("TCP_QP_RCV_STATE_CHANGE", TRACE_TCP_QP_RCV_STATE_CHANGE, 200 OWNER_NVMF_TCP, OBJECT_NONE, 0, 201 SPDK_TRACE_ARG_TYPE_INT, "state"); 202 203 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_TCP_IO, 1); 204 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_TCP_IO, 0); 205 spdk_trace_tpoint_register_relation(TRACE_SOCK_REQ_QUEUE, OBJECT_NVMF_TCP_IO, 0); 206 spdk_trace_tpoint_register_relation(TRACE_SOCK_REQ_PEND, OBJECT_NVMF_TCP_IO, 0); 207 spdk_trace_tpoint_register_relation(TRACE_SOCK_REQ_COMPLETE, OBJECT_NVMF_TCP_IO, 0); 208 } 209 210 struct spdk_nvmf_tcp_req { 211 struct spdk_nvmf_request req; 212 struct spdk_nvme_cpl rsp; 213 struct spdk_nvme_cmd cmd; 214 215 /* A PDU that can be used for sending responses. This is 216 * not the incoming PDU! */ 217 struct nvme_tcp_pdu *pdu; 218 219 /* In-capsule data buffer */ 220 uint8_t *buf; 221 222 struct spdk_nvmf_tcp_req *fused_pair; 223 224 /* 225 * The PDU for a request may be used multiple times in serial over 226 * the request's lifetime. For example, first to send an R2T, then 227 * to send a completion. To catch mistakes where the PDU is used 228 * twice at the same time, add a debug flag here for init/fini. 229 */ 230 bool pdu_in_use; 231 bool has_in_capsule_data; 232 bool fused_failed; 233 234 /* transfer_tag */ 235 uint16_t ttag; 236 237 enum spdk_nvmf_tcp_req_state state; 238 239 /* 240 * h2c_offset is used when we receive the h2c_data PDU. 241 */ 242 uint32_t h2c_offset; 243 244 STAILQ_ENTRY(spdk_nvmf_tcp_req) link; 245 TAILQ_ENTRY(spdk_nvmf_tcp_req) state_link; 246 }; 247 248 struct spdk_nvmf_tcp_qpair { 249 struct spdk_nvmf_qpair qpair; 250 struct spdk_nvmf_tcp_poll_group *group; 251 struct spdk_sock *sock; 252 253 enum nvme_tcp_pdu_recv_state recv_state; 254 enum nvme_tcp_qpair_state state; 255 256 /* PDU being actively received */ 257 struct nvme_tcp_pdu *pdu_in_progress; 258 259 struct spdk_nvmf_tcp_req *fused_first; 260 261 /* Queues to track the requests in all states */ 262 TAILQ_HEAD(, spdk_nvmf_tcp_req) tcp_req_working_queue; 263 TAILQ_HEAD(, spdk_nvmf_tcp_req) tcp_req_free_queue; 264 SLIST_HEAD(, nvme_tcp_pdu) tcp_pdu_free_queue; 265 /* Number of working pdus */ 266 uint32_t tcp_pdu_working_count; 267 268 /* Number of requests in each state */ 269 uint32_t state_cntr[TCP_REQUEST_NUM_STATES]; 270 271 uint8_t cpda; 272 273 bool host_hdgst_enable; 274 bool host_ddgst_enable; 275 276 /* This is a spare PDU used for sending special management 277 * operations. Primarily, this is used for the initial 278 * connection response and c2h termination request. */ 279 struct nvme_tcp_pdu *mgmt_pdu; 280 281 /* Arrays of in-capsule buffers, requests, and pdus. 282 * Each array is 'resource_count' number of elements */ 283 void *bufs; 284 struct spdk_nvmf_tcp_req *reqs; 285 struct nvme_tcp_pdu *pdus; 286 uint32_t resource_count; 287 uint32_t recv_buf_size; 288 289 struct spdk_nvmf_tcp_port *port; 290 291 /* IP address */ 292 char initiator_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 293 char target_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 294 295 /* IP port */ 296 uint16_t initiator_port; 297 uint16_t target_port; 298 299 /* Wait until the host terminates the connection (e.g. after sending C2HTermReq) */ 300 bool wait_terminate; 301 302 /* Timer used to destroy qpair after detecting transport error issue if initiator does 303 * not close the connection. 304 */ 305 struct spdk_poller *timeout_poller; 306 307 spdk_nvmf_transport_qpair_fini_cb fini_cb_fn; 308 void *fini_cb_arg; 309 310 TAILQ_ENTRY(spdk_nvmf_tcp_qpair) link; 311 }; 312 313 struct spdk_nvmf_tcp_control_msg { 314 STAILQ_ENTRY(spdk_nvmf_tcp_control_msg) link; 315 }; 316 317 struct spdk_nvmf_tcp_control_msg_list { 318 void *msg_buf; 319 STAILQ_HEAD(, spdk_nvmf_tcp_control_msg) free_msgs; 320 }; 321 322 struct spdk_nvmf_tcp_poll_group { 323 struct spdk_nvmf_transport_poll_group group; 324 struct spdk_sock_group *sock_group; 325 326 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) qpairs; 327 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) await_req; 328 329 struct spdk_io_channel *accel_channel; 330 struct spdk_nvmf_tcp_control_msg_list *control_msg_list; 331 332 TAILQ_ENTRY(spdk_nvmf_tcp_poll_group) link; 333 }; 334 335 struct spdk_nvmf_tcp_port { 336 const struct spdk_nvme_transport_id *trid; 337 struct spdk_sock *listen_sock; 338 TAILQ_ENTRY(spdk_nvmf_tcp_port) link; 339 }; 340 341 struct tcp_transport_opts { 342 bool c2h_success; 343 uint16_t control_msg_num; 344 uint32_t sock_priority; 345 }; 346 347 struct tcp_psk_entry { 348 char hostnqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 349 char subnqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 350 char pskid[NVMF_PSK_IDENTITY_LEN]; 351 uint8_t psk[SPDK_TLS_PSK_MAX_LEN]; 352 struct spdk_key *key; 353 354 /* Original path saved to emit SPDK configuration via "save_config". */ 355 char psk_path[PATH_MAX]; 356 uint32_t psk_size; 357 enum nvme_tcp_cipher_suite tls_cipher_suite; 358 TAILQ_ENTRY(tcp_psk_entry) link; 359 }; 360 361 struct spdk_nvmf_tcp_transport { 362 struct spdk_nvmf_transport transport; 363 struct tcp_transport_opts tcp_opts; 364 uint32_t ack_timeout; 365 366 struct spdk_nvmf_tcp_poll_group *next_pg; 367 368 struct spdk_poller *accept_poller; 369 370 TAILQ_HEAD(, spdk_nvmf_tcp_port) ports; 371 TAILQ_HEAD(, spdk_nvmf_tcp_poll_group) poll_groups; 372 373 TAILQ_HEAD(, tcp_psk_entry) psks; 374 }; 375 376 static const struct spdk_json_object_decoder tcp_transport_opts_decoder[] = { 377 { 378 "c2h_success", offsetof(struct tcp_transport_opts, c2h_success), 379 spdk_json_decode_bool, true 380 }, 381 { 382 "control_msg_num", offsetof(struct tcp_transport_opts, control_msg_num), 383 spdk_json_decode_uint16, true 384 }, 385 { 386 "sock_priority", offsetof(struct tcp_transport_opts, sock_priority), 387 spdk_json_decode_uint32, true 388 }, 389 }; 390 391 static bool nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 392 struct spdk_nvmf_tcp_req *tcp_req); 393 static void nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 394 395 static void _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 396 struct spdk_nvmf_tcp_req *tcp_req); 397 398 static inline void 399 nvmf_tcp_req_set_state(struct spdk_nvmf_tcp_req *tcp_req, 400 enum spdk_nvmf_tcp_req_state state) 401 { 402 struct spdk_nvmf_qpair *qpair; 403 struct spdk_nvmf_tcp_qpair *tqpair; 404 405 qpair = tcp_req->req.qpair; 406 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 407 408 assert(tqpair->state_cntr[tcp_req->state] > 0); 409 tqpair->state_cntr[tcp_req->state]--; 410 tqpair->state_cntr[state]++; 411 412 tcp_req->state = state; 413 } 414 415 static inline struct nvme_tcp_pdu * 416 nvmf_tcp_req_pdu_init(struct spdk_nvmf_tcp_req *tcp_req) 417 { 418 assert(tcp_req->pdu_in_use == false); 419 420 memset(tcp_req->pdu, 0, sizeof(*tcp_req->pdu)); 421 tcp_req->pdu->qpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 422 423 return tcp_req->pdu; 424 } 425 426 static struct spdk_nvmf_tcp_req * 427 nvmf_tcp_req_get(struct spdk_nvmf_tcp_qpair *tqpair) 428 { 429 struct spdk_nvmf_tcp_req *tcp_req; 430 431 tcp_req = TAILQ_FIRST(&tqpair->tcp_req_free_queue); 432 if (spdk_unlikely(!tcp_req)) { 433 return NULL; 434 } 435 436 memset(&tcp_req->rsp, 0, sizeof(tcp_req->rsp)); 437 tcp_req->h2c_offset = 0; 438 tcp_req->has_in_capsule_data = false; 439 tcp_req->req.dif_enabled = false; 440 tcp_req->req.zcopy_phase = NVMF_ZCOPY_PHASE_NONE; 441 442 TAILQ_REMOVE(&tqpair->tcp_req_free_queue, tcp_req, state_link); 443 TAILQ_INSERT_TAIL(&tqpair->tcp_req_working_queue, tcp_req, state_link); 444 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEW); 445 return tcp_req; 446 } 447 448 static inline void 449 nvmf_tcp_req_put(struct spdk_nvmf_tcp_qpair *tqpair, struct spdk_nvmf_tcp_req *tcp_req) 450 { 451 assert(!tcp_req->pdu_in_use); 452 453 TAILQ_REMOVE(&tqpair->tcp_req_working_queue, tcp_req, state_link); 454 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 455 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_FREE); 456 } 457 458 static void 459 nvmf_tcp_request_free(void *cb_arg) 460 { 461 struct spdk_nvmf_tcp_transport *ttransport; 462 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 463 464 assert(tcp_req != NULL); 465 466 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req=%p will be freed\n", tcp_req); 467 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 468 struct spdk_nvmf_tcp_transport, transport); 469 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 470 nvmf_tcp_req_process(ttransport, tcp_req); 471 } 472 473 static int 474 nvmf_tcp_req_free(struct spdk_nvmf_request *req) 475 { 476 struct spdk_nvmf_tcp_req *tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 477 478 nvmf_tcp_request_free(tcp_req); 479 480 return 0; 481 } 482 483 static void 484 nvmf_tcp_drain_state_queue(struct spdk_nvmf_tcp_qpair *tqpair, 485 enum spdk_nvmf_tcp_req_state state) 486 { 487 struct spdk_nvmf_tcp_req *tcp_req, *req_tmp; 488 489 assert(state != TCP_REQUEST_STATE_FREE); 490 TAILQ_FOREACH_SAFE(tcp_req, &tqpair->tcp_req_working_queue, state_link, req_tmp) { 491 if (state == tcp_req->state) { 492 nvmf_tcp_request_free(tcp_req); 493 } 494 } 495 } 496 497 static void 498 nvmf_tcp_cleanup_all_states(struct spdk_nvmf_tcp_qpair *tqpair) 499 { 500 struct spdk_nvmf_tcp_req *tcp_req, *req_tmp; 501 502 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 503 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_NEW); 504 505 /* Wipe the requests waiting for buffer from the global list */ 506 TAILQ_FOREACH_SAFE(tcp_req, &tqpair->tcp_req_working_queue, state_link, req_tmp) { 507 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 508 STAILQ_REMOVE(&tqpair->group->group.pending_buf_queue, &tcp_req->req, 509 spdk_nvmf_request, buf_link); 510 } 511 } 512 513 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_NEED_BUFFER); 514 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_EXECUTING); 515 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 516 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 517 } 518 519 static void 520 nvmf_tcp_dump_qpair_req_contents(struct spdk_nvmf_tcp_qpair *tqpair) 521 { 522 int i; 523 struct spdk_nvmf_tcp_req *tcp_req; 524 525 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", tqpair->qpair.qid); 526 for (i = 1; i < TCP_REQUEST_NUM_STATES; i++) { 527 SPDK_ERRLOG("\tNum of requests in state[%d] = %u\n", i, tqpair->state_cntr[i]); 528 TAILQ_FOREACH(tcp_req, &tqpair->tcp_req_working_queue, state_link) { 529 if ((int)tcp_req->state == i) { 530 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", tcp_req->req.data_from_pool); 531 SPDK_ERRLOG("\t\tRequest opcode: %d\n", tcp_req->req.cmd->nvmf_cmd.opcode); 532 } 533 } 534 } 535 } 536 537 static void 538 _nvmf_tcp_qpair_destroy(void *_tqpair) 539 { 540 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 541 spdk_nvmf_transport_qpair_fini_cb cb_fn = tqpair->fini_cb_fn; 542 void *cb_arg = tqpair->fini_cb_arg; 543 int err = 0; 544 545 spdk_trace_record(TRACE_TCP_QP_DESTROY, 0, 0, (uintptr_t)tqpair); 546 547 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 548 549 err = spdk_sock_close(&tqpair->sock); 550 assert(err == 0); 551 nvmf_tcp_cleanup_all_states(tqpair); 552 553 if (tqpair->state_cntr[TCP_REQUEST_STATE_FREE] != tqpair->resource_count) { 554 SPDK_ERRLOG("tqpair(%p) free tcp request num is %u but should be %u\n", tqpair, 555 tqpair->state_cntr[TCP_REQUEST_STATE_FREE], 556 tqpair->resource_count); 557 err++; 558 } 559 560 if (err > 0) { 561 nvmf_tcp_dump_qpair_req_contents(tqpair); 562 } 563 564 /* The timeout poller might still be registered here if we close the qpair before host 565 * terminates the connection. 566 */ 567 spdk_poller_unregister(&tqpair->timeout_poller); 568 spdk_dma_free(tqpair->pdus); 569 free(tqpair->reqs); 570 spdk_free(tqpair->bufs); 571 free(tqpair); 572 573 if (cb_fn != NULL) { 574 cb_fn(cb_arg); 575 } 576 577 SPDK_DEBUGLOG(nvmf_tcp, "Leave\n"); 578 } 579 580 static void 581 nvmf_tcp_qpair_destroy(struct spdk_nvmf_tcp_qpair *tqpair) 582 { 583 /* Delay the destruction to make sure it isn't performed from the context of a sock 584 * callback. Otherwise, spdk_sock_close() might not abort pending requests, causing their 585 * completions to be executed after the qpair is freed. (Note: this fixed issue #2471.) 586 */ 587 spdk_thread_send_msg(spdk_get_thread(), _nvmf_tcp_qpair_destroy, tqpair); 588 } 589 590 static void 591 nvmf_tcp_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 592 { 593 struct spdk_nvmf_tcp_transport *ttransport; 594 assert(w != NULL); 595 596 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 597 spdk_json_write_named_bool(w, "c2h_success", ttransport->tcp_opts.c2h_success); 598 spdk_json_write_named_uint32(w, "sock_priority", ttransport->tcp_opts.sock_priority); 599 } 600 601 static void 602 nvmf_tcp_free_psk_entry(struct tcp_psk_entry *entry) 603 { 604 if (entry == NULL) { 605 return; 606 } 607 608 spdk_memset_s(entry->psk, sizeof(entry->psk), 0, sizeof(entry->psk)); 609 spdk_keyring_put_key(entry->key); 610 free(entry); 611 } 612 613 static int 614 nvmf_tcp_destroy(struct spdk_nvmf_transport *transport, 615 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 616 { 617 struct spdk_nvmf_tcp_transport *ttransport; 618 struct tcp_psk_entry *entry, *tmp; 619 620 assert(transport != NULL); 621 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 622 623 TAILQ_FOREACH_SAFE(entry, &ttransport->psks, link, tmp) { 624 TAILQ_REMOVE(&ttransport->psks, entry, link); 625 nvmf_tcp_free_psk_entry(entry); 626 } 627 628 spdk_poller_unregister(&ttransport->accept_poller); 629 free(ttransport); 630 631 if (cb_fn) { 632 cb_fn(cb_arg); 633 } 634 return 0; 635 } 636 637 static int nvmf_tcp_accept(void *ctx); 638 639 static struct spdk_nvmf_transport * 640 nvmf_tcp_create(struct spdk_nvmf_transport_opts *opts) 641 { 642 struct spdk_nvmf_tcp_transport *ttransport; 643 uint32_t sge_count; 644 uint32_t min_shared_buffers; 645 646 ttransport = calloc(1, sizeof(*ttransport)); 647 if (!ttransport) { 648 return NULL; 649 } 650 651 TAILQ_INIT(&ttransport->ports); 652 TAILQ_INIT(&ttransport->poll_groups); 653 TAILQ_INIT(&ttransport->psks); 654 655 ttransport->transport.ops = &spdk_nvmf_transport_tcp; 656 657 ttransport->tcp_opts.c2h_success = SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION; 658 ttransport->tcp_opts.sock_priority = SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY; 659 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 660 if (opts->transport_specific != NULL && 661 spdk_json_decode_object_relaxed(opts->transport_specific, tcp_transport_opts_decoder, 662 SPDK_COUNTOF(tcp_transport_opts_decoder), 663 &ttransport->tcp_opts)) { 664 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 665 free(ttransport); 666 return NULL; 667 } 668 669 SPDK_NOTICELOG("*** TCP Transport Init ***\n"); 670 671 SPDK_INFOLOG(nvmf_tcp, "*** TCP Transport Init ***\n" 672 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 673 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 674 " in_capsule_data_size=%d, max_aq_depth=%d\n" 675 " num_shared_buffers=%d, c2h_success=%d,\n" 676 " dif_insert_or_strip=%d, sock_priority=%d\n" 677 " abort_timeout_sec=%d, control_msg_num=%hu\n" 678 " ack_timeout=%d\n", 679 opts->max_queue_depth, 680 opts->max_io_size, 681 opts->max_qpairs_per_ctrlr - 1, 682 opts->io_unit_size, 683 opts->in_capsule_data_size, 684 opts->max_aq_depth, 685 opts->num_shared_buffers, 686 ttransport->tcp_opts.c2h_success, 687 opts->dif_insert_or_strip, 688 ttransport->tcp_opts.sock_priority, 689 opts->abort_timeout_sec, 690 ttransport->tcp_opts.control_msg_num, 691 opts->ack_timeout); 692 693 if (ttransport->tcp_opts.sock_priority > SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY) { 694 SPDK_ERRLOG("Unsupported socket_priority=%d, the current range is: 0 to %d\n" 695 "you can use man 7 socket to view the range of priority under SO_PRIORITY item\n", 696 ttransport->tcp_opts.sock_priority, SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY); 697 free(ttransport); 698 return NULL; 699 } 700 701 if (ttransport->tcp_opts.control_msg_num == 0 && 702 opts->in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 703 SPDK_WARNLOG("TCP param control_msg_num can't be 0 if ICD is less than %u bytes. Using default value %u\n", 704 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM); 705 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 706 } 707 708 /* I/O unit size cannot be larger than max I/O size */ 709 if (opts->io_unit_size > opts->max_io_size) { 710 SPDK_WARNLOG("TCP param io_unit_size %u can't be larger than max_io_size %u. Using max_io_size as io_unit_size\n", 711 opts->io_unit_size, opts->max_io_size); 712 opts->io_unit_size = opts->max_io_size; 713 } 714 715 /* In capsule data size cannot be larger than max I/O size */ 716 if (opts->in_capsule_data_size > opts->max_io_size) { 717 SPDK_WARNLOG("TCP param ICD size %u can't be larger than max_io_size %u. Using max_io_size as ICD size\n", 718 opts->io_unit_size, opts->max_io_size); 719 opts->in_capsule_data_size = opts->max_io_size; 720 } 721 722 /* max IO queue depth cannot be smaller than 2 or larger than 65535. 723 * We will not check SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH, because max_queue_depth is 16bits and always not larger than 64k. */ 724 if (opts->max_queue_depth < SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH) { 725 SPDK_WARNLOG("TCP param max_queue_depth %u can't be smaller than %u or larger than %u. Using default value %u\n", 726 opts->max_queue_depth, SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH, 727 SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH, SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH); 728 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH; 729 } 730 731 /* max admin queue depth cannot be smaller than 2 or larger than 4096 */ 732 if (opts->max_aq_depth < SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH || 733 opts->max_aq_depth > SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH) { 734 SPDK_WARNLOG("TCP param max_aq_depth %u can't be smaller than %u or larger than %u. Using default value %u\n", 735 opts->max_aq_depth, SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH, 736 SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH, SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH); 737 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH; 738 } 739 740 sge_count = opts->max_io_size / opts->io_unit_size; 741 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 742 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 743 free(ttransport); 744 return NULL; 745 } 746 747 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 748 if (opts->buf_cache_size < UINT32_MAX) { 749 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 750 if (min_shared_buffers > opts->num_shared_buffers) { 751 SPDK_ERRLOG("There are not enough buffers to satisfy " 752 "per-poll group caches for each thread. (%" PRIu32 ") " 753 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 754 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 755 free(ttransport); 756 return NULL; 757 } 758 } 759 760 ttransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_tcp_accept, &ttransport->transport, 761 opts->acceptor_poll_rate); 762 if (!ttransport->accept_poller) { 763 free(ttransport); 764 return NULL; 765 } 766 767 return &ttransport->transport; 768 } 769 770 static int 771 nvmf_tcp_trsvcid_to_int(const char *trsvcid) 772 { 773 unsigned long long ull; 774 char *end = NULL; 775 776 ull = strtoull(trsvcid, &end, 10); 777 if (end == NULL || end == trsvcid || *end != '\0') { 778 return -1; 779 } 780 781 /* Valid TCP/IP port numbers are in [1, 65535] */ 782 if (ull == 0 || ull > 65535) { 783 return -1; 784 } 785 786 return (int)ull; 787 } 788 789 /** 790 * Canonicalize a listen address trid. 791 */ 792 static int 793 nvmf_tcp_canon_listen_trid(struct spdk_nvme_transport_id *canon_trid, 794 const struct spdk_nvme_transport_id *trid) 795 { 796 int trsvcid_int; 797 798 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 799 if (trsvcid_int < 0) { 800 return -EINVAL; 801 } 802 803 memset(canon_trid, 0, sizeof(*canon_trid)); 804 spdk_nvme_trid_populate_transport(canon_trid, SPDK_NVME_TRANSPORT_TCP); 805 canon_trid->adrfam = trid->adrfam; 806 snprintf(canon_trid->traddr, sizeof(canon_trid->traddr), "%s", trid->traddr); 807 snprintf(canon_trid->trsvcid, sizeof(canon_trid->trsvcid), "%d", trsvcid_int); 808 809 return 0; 810 } 811 812 /** 813 * Find an existing listening port. 814 */ 815 static struct spdk_nvmf_tcp_port * 816 nvmf_tcp_find_port(struct spdk_nvmf_tcp_transport *ttransport, 817 const struct spdk_nvme_transport_id *trid) 818 { 819 struct spdk_nvme_transport_id canon_trid; 820 struct spdk_nvmf_tcp_port *port; 821 822 if (nvmf_tcp_canon_listen_trid(&canon_trid, trid) != 0) { 823 return NULL; 824 } 825 826 TAILQ_FOREACH(port, &ttransport->ports, link) { 827 if (spdk_nvme_transport_id_compare(&canon_trid, port->trid) == 0) { 828 return port; 829 } 830 } 831 832 return NULL; 833 } 834 835 static int 836 tcp_sock_get_key(uint8_t *out, int out_len, const char **cipher, const char *pskid, 837 void *get_key_ctx) 838 { 839 struct tcp_psk_entry *entry; 840 struct spdk_nvmf_tcp_transport *ttransport = get_key_ctx; 841 size_t psk_len; 842 int rc; 843 844 TAILQ_FOREACH(entry, &ttransport->psks, link) { 845 if (strcmp(pskid, entry->pskid) != 0) { 846 continue; 847 } 848 849 psk_len = entry->psk_size; 850 if ((size_t)out_len < psk_len) { 851 SPDK_ERRLOG("Out buffer of size: %" PRIu32 " cannot fit PSK of len: %lu\n", 852 out_len, psk_len); 853 return -ENOBUFS; 854 } 855 856 /* Convert PSK to the TLS PSK format. */ 857 rc = nvme_tcp_derive_tls_psk(entry->psk, psk_len, pskid, out, out_len, 858 entry->tls_cipher_suite); 859 if (rc < 0) { 860 SPDK_ERRLOG("Could not generate TLS PSK\n"); 861 } 862 863 switch (entry->tls_cipher_suite) { 864 case NVME_TCP_CIPHER_AES_128_GCM_SHA256: 865 *cipher = "TLS_AES_128_GCM_SHA256"; 866 break; 867 case NVME_TCP_CIPHER_AES_256_GCM_SHA384: 868 *cipher = "TLS_AES_256_GCM_SHA384"; 869 break; 870 default: 871 *cipher = NULL; 872 return -ENOTSUP; 873 } 874 875 return rc; 876 } 877 878 SPDK_ERRLOG("Could not find PSK for identity: %s\n", pskid); 879 880 return -EINVAL; 881 } 882 883 static int 884 nvmf_tcp_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 885 struct spdk_nvmf_listen_opts *listen_opts) 886 { 887 struct spdk_nvmf_tcp_transport *ttransport; 888 struct spdk_nvmf_tcp_port *port; 889 int trsvcid_int; 890 uint8_t adrfam; 891 const char *sock_impl_name; 892 struct spdk_sock_impl_opts impl_opts; 893 size_t impl_opts_size = sizeof(impl_opts); 894 struct spdk_sock_opts opts; 895 896 if (!strlen(trid->trsvcid)) { 897 SPDK_ERRLOG("Service id is required\n"); 898 return -EINVAL; 899 } 900 901 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 902 903 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 904 if (trsvcid_int < 0) { 905 SPDK_ERRLOG("Invalid trsvcid '%s'\n", trid->trsvcid); 906 return -EINVAL; 907 } 908 909 port = calloc(1, sizeof(*port)); 910 if (!port) { 911 SPDK_ERRLOG("Port allocation failed\n"); 912 return -ENOMEM; 913 } 914 915 port->trid = trid; 916 917 sock_impl_name = NULL; 918 919 opts.opts_size = sizeof(opts); 920 spdk_sock_get_default_opts(&opts); 921 opts.priority = ttransport->tcp_opts.sock_priority; 922 opts.ack_timeout = transport->opts.ack_timeout; 923 if (listen_opts->secure_channel) { 924 if (!g_tls_log) { 925 SPDK_NOTICELOG("TLS support is considered experimental\n"); 926 g_tls_log = true; 927 } 928 sock_impl_name = "ssl"; 929 spdk_sock_impl_get_opts(sock_impl_name, &impl_opts, &impl_opts_size); 930 impl_opts.tls_version = SPDK_TLS_VERSION_1_3; 931 impl_opts.get_key = tcp_sock_get_key; 932 impl_opts.get_key_ctx = ttransport; 933 impl_opts.tls_cipher_suites = "TLS_AES_256_GCM_SHA384:TLS_AES_128_GCM_SHA256"; 934 opts.impl_opts = &impl_opts; 935 opts.impl_opts_size = sizeof(impl_opts); 936 } 937 938 port->listen_sock = spdk_sock_listen_ext(trid->traddr, trsvcid_int, 939 sock_impl_name, &opts); 940 if (port->listen_sock == NULL) { 941 SPDK_ERRLOG("spdk_sock_listen(%s, %d) failed: %s (%d)\n", 942 trid->traddr, trsvcid_int, 943 spdk_strerror(errno), errno); 944 free(port); 945 return -errno; 946 } 947 948 if (spdk_sock_is_ipv4(port->listen_sock)) { 949 adrfam = SPDK_NVMF_ADRFAM_IPV4; 950 } else if (spdk_sock_is_ipv6(port->listen_sock)) { 951 adrfam = SPDK_NVMF_ADRFAM_IPV6; 952 } else { 953 SPDK_ERRLOG("Unhandled socket type\n"); 954 adrfam = 0; 955 } 956 957 if (adrfam != trid->adrfam) { 958 SPDK_ERRLOG("Socket address family mismatch\n"); 959 spdk_sock_close(&port->listen_sock); 960 free(port); 961 return -EINVAL; 962 } 963 964 SPDK_NOTICELOG("*** NVMe/TCP Target Listening on %s port %s ***\n", 965 trid->traddr, trid->trsvcid); 966 967 TAILQ_INSERT_TAIL(&ttransport->ports, port, link); 968 return 0; 969 } 970 971 static void 972 nvmf_tcp_stop_listen(struct spdk_nvmf_transport *transport, 973 const struct spdk_nvme_transport_id *trid) 974 { 975 struct spdk_nvmf_tcp_transport *ttransport; 976 struct spdk_nvmf_tcp_port *port; 977 978 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 979 980 SPDK_DEBUGLOG(nvmf_tcp, "Removing listen address %s port %s\n", 981 trid->traddr, trid->trsvcid); 982 983 port = nvmf_tcp_find_port(ttransport, trid); 984 if (port) { 985 TAILQ_REMOVE(&ttransport->ports, port, link); 986 spdk_sock_close(&port->listen_sock); 987 free(port); 988 } 989 } 990 991 static void nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 992 enum nvme_tcp_pdu_recv_state state); 993 994 static void 995 nvmf_tcp_qpair_set_state(struct spdk_nvmf_tcp_qpair *tqpair, enum nvme_tcp_qpair_state state) 996 { 997 tqpair->state = state; 998 spdk_trace_record(TRACE_TCP_QP_STATE_CHANGE, tqpair->qpair.qid, 0, (uintptr_t)tqpair, 999 tqpair->state); 1000 } 1001 1002 static void 1003 nvmf_tcp_qpair_disconnect(struct spdk_nvmf_tcp_qpair *tqpair) 1004 { 1005 SPDK_DEBUGLOG(nvmf_tcp, "Disconnecting qpair %p\n", tqpair); 1006 1007 spdk_trace_record(TRACE_TCP_QP_DISCONNECT, 0, 0, (uintptr_t)tqpair); 1008 1009 if (tqpair->state <= NVME_TCP_QPAIR_STATE_RUNNING) { 1010 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITING); 1011 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1012 spdk_poller_unregister(&tqpair->timeout_poller); 1013 1014 /* This will end up calling nvmf_tcp_close_qpair */ 1015 spdk_nvmf_qpair_disconnect(&tqpair->qpair, NULL, NULL); 1016 } 1017 } 1018 1019 static void 1020 _mgmt_pdu_write_done(void *_tqpair, int err) 1021 { 1022 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 1023 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1024 1025 if (spdk_unlikely(err != 0)) { 1026 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1027 return; 1028 } 1029 1030 assert(pdu->cb_fn != NULL); 1031 pdu->cb_fn(pdu->cb_arg); 1032 } 1033 1034 static void 1035 _req_pdu_write_done(void *req, int err) 1036 { 1037 struct spdk_nvmf_tcp_req *tcp_req = req; 1038 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1039 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1040 1041 assert(tcp_req->pdu_in_use); 1042 tcp_req->pdu_in_use = false; 1043 1044 /* If the request is in a completed state, we're waiting for write completion to free it */ 1045 if (spdk_unlikely(tcp_req->state == TCP_REQUEST_STATE_COMPLETED)) { 1046 nvmf_tcp_request_free(tcp_req); 1047 return; 1048 } 1049 1050 if (spdk_unlikely(err != 0)) { 1051 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1052 return; 1053 } 1054 1055 assert(pdu->cb_fn != NULL); 1056 pdu->cb_fn(pdu->cb_arg); 1057 } 1058 1059 static void 1060 _pdu_write_done(struct nvme_tcp_pdu *pdu, int err) 1061 { 1062 pdu->sock_req.cb_fn(pdu->sock_req.cb_arg, err); 1063 } 1064 1065 static void 1066 _tcp_write_pdu(struct nvme_tcp_pdu *pdu) 1067 { 1068 int rc; 1069 uint32_t mapped_length; 1070 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1071 1072 pdu->sock_req.iovcnt = nvme_tcp_build_iovs(pdu->iov, SPDK_COUNTOF(pdu->iov), pdu, 1073 tqpair->host_hdgst_enable, tqpair->host_ddgst_enable, &mapped_length); 1074 spdk_sock_writev_async(tqpair->sock, &pdu->sock_req); 1075 1076 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP || 1077 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ) { 1078 /* Try to force the send immediately. */ 1079 rc = spdk_sock_flush(tqpair->sock); 1080 if (rc > 0 && (uint32_t)rc == mapped_length) { 1081 _pdu_write_done(pdu, 0); 1082 } else { 1083 SPDK_ERRLOG("Could not write %s to socket: rc=%d, errno=%d\n", 1084 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP ? 1085 "IC_RESP" : "TERM_REQ", rc, errno); 1086 _pdu_write_done(pdu, rc >= 0 ? -EAGAIN : -errno); 1087 } 1088 } 1089 } 1090 1091 static void 1092 data_crc32_accel_done(void *cb_arg, int status) 1093 { 1094 struct nvme_tcp_pdu *pdu = cb_arg; 1095 1096 if (spdk_unlikely(status)) { 1097 SPDK_ERRLOG("Failed to compute the data digest for pdu =%p\n", pdu); 1098 _pdu_write_done(pdu, status); 1099 return; 1100 } 1101 1102 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 1103 MAKE_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32); 1104 1105 _tcp_write_pdu(pdu); 1106 } 1107 1108 static void 1109 pdu_data_crc32_compute(struct nvme_tcp_pdu *pdu) 1110 { 1111 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1112 int rc = 0; 1113 1114 /* Data Digest */ 1115 if (pdu->data_len > 0 && g_nvme_tcp_ddgst[pdu->hdr.common.pdu_type] && tqpair->host_ddgst_enable) { 1116 /* Only support this limitated case for the first step */ 1117 if (spdk_likely(!pdu->dif_ctx && (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0) 1118 && tqpair->group)) { 1119 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 1120 pdu->data_iovcnt, 0, data_crc32_accel_done, pdu); 1121 if (spdk_likely(rc == 0)) { 1122 return; 1123 } 1124 } else { 1125 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 1126 } 1127 data_crc32_accel_done(pdu, rc); 1128 } else { 1129 _tcp_write_pdu(pdu); 1130 } 1131 } 1132 1133 static void 1134 nvmf_tcp_qpair_write_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1135 struct nvme_tcp_pdu *pdu, 1136 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1137 void *cb_arg) 1138 { 1139 int hlen; 1140 uint32_t crc32c; 1141 1142 assert(tqpair->pdu_in_progress != pdu); 1143 1144 hlen = pdu->hdr.common.hlen; 1145 pdu->cb_fn = cb_fn; 1146 pdu->cb_arg = cb_arg; 1147 1148 pdu->iov[0].iov_base = &pdu->hdr.raw; 1149 pdu->iov[0].iov_len = hlen; 1150 1151 /* Header Digest */ 1152 if (g_nvme_tcp_hdgst[pdu->hdr.common.pdu_type] && tqpair->host_hdgst_enable) { 1153 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 1154 MAKE_DIGEST_WORD((uint8_t *)pdu->hdr.raw + hlen, crc32c); 1155 } 1156 1157 /* Data Digest */ 1158 pdu_data_crc32_compute(pdu); 1159 } 1160 1161 static void 1162 nvmf_tcp_qpair_write_mgmt_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1163 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1164 void *cb_arg) 1165 { 1166 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1167 1168 pdu->sock_req.cb_fn = _mgmt_pdu_write_done; 1169 pdu->sock_req.cb_arg = tqpair; 1170 1171 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1172 } 1173 1174 static void 1175 nvmf_tcp_qpair_write_req_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1176 struct spdk_nvmf_tcp_req *tcp_req, 1177 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1178 void *cb_arg) 1179 { 1180 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1181 1182 pdu->sock_req.cb_fn = _req_pdu_write_done; 1183 pdu->sock_req.cb_arg = tcp_req; 1184 1185 assert(!tcp_req->pdu_in_use); 1186 tcp_req->pdu_in_use = true; 1187 1188 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1189 } 1190 1191 static int 1192 nvmf_tcp_qpair_init_mem_resource(struct spdk_nvmf_tcp_qpair *tqpair) 1193 { 1194 uint32_t i; 1195 struct spdk_nvmf_transport_opts *opts; 1196 uint32_t in_capsule_data_size; 1197 1198 opts = &tqpair->qpair.transport->opts; 1199 1200 in_capsule_data_size = opts->in_capsule_data_size; 1201 if (opts->dif_insert_or_strip) { 1202 in_capsule_data_size = SPDK_BDEV_BUF_SIZE_WITH_MD(in_capsule_data_size); 1203 } 1204 1205 tqpair->resource_count = opts->max_queue_depth; 1206 1207 tqpair->reqs = calloc(tqpair->resource_count, sizeof(*tqpair->reqs)); 1208 if (!tqpair->reqs) { 1209 SPDK_ERRLOG("Unable to allocate reqs on tqpair=%p\n", tqpair); 1210 return -1; 1211 } 1212 1213 if (in_capsule_data_size) { 1214 tqpair->bufs = spdk_zmalloc(tqpair->resource_count * in_capsule_data_size, 0x1000, 1215 NULL, SPDK_ENV_LCORE_ID_ANY, 1216 SPDK_MALLOC_DMA); 1217 if (!tqpair->bufs) { 1218 SPDK_ERRLOG("Unable to allocate bufs on tqpair=%p.\n", tqpair); 1219 return -1; 1220 } 1221 } 1222 /* prepare memory space for receiving pdus and tcp_req */ 1223 /* Add additional 1 member, which will be used for mgmt_pdu owned by the tqpair */ 1224 tqpair->pdus = spdk_dma_zmalloc((2 * tqpair->resource_count + 1) * sizeof(*tqpair->pdus), 0x1000, 1225 NULL); 1226 if (!tqpair->pdus) { 1227 SPDK_ERRLOG("Unable to allocate pdu pool on tqpair =%p.\n", tqpair); 1228 return -1; 1229 } 1230 1231 for (i = 0; i < tqpair->resource_count; i++) { 1232 struct spdk_nvmf_tcp_req *tcp_req = &tqpair->reqs[i]; 1233 1234 tcp_req->ttag = i + 1; 1235 tcp_req->req.qpair = &tqpair->qpair; 1236 1237 tcp_req->pdu = &tqpair->pdus[i]; 1238 tcp_req->pdu->qpair = tqpair; 1239 1240 /* Set up memory to receive commands */ 1241 if (tqpair->bufs) { 1242 tcp_req->buf = (void *)((uintptr_t)tqpair->bufs + (i * in_capsule_data_size)); 1243 } 1244 1245 /* Set the cmdn and rsp */ 1246 tcp_req->req.rsp = (union nvmf_c2h_msg *)&tcp_req->rsp; 1247 tcp_req->req.cmd = (union nvmf_h2c_msg *)&tcp_req->cmd; 1248 1249 tcp_req->req.stripped_data = NULL; 1250 1251 /* Initialize request state to FREE */ 1252 tcp_req->state = TCP_REQUEST_STATE_FREE; 1253 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 1254 tqpair->state_cntr[TCP_REQUEST_STATE_FREE]++; 1255 } 1256 1257 for (; i < 2 * tqpair->resource_count; i++) { 1258 struct nvme_tcp_pdu *pdu = &tqpair->pdus[i]; 1259 1260 pdu->qpair = tqpair; 1261 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 1262 } 1263 1264 tqpair->mgmt_pdu = &tqpair->pdus[i]; 1265 tqpair->mgmt_pdu->qpair = tqpair; 1266 tqpair->pdu_in_progress = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 1267 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 1268 tqpair->tcp_pdu_working_count = 1; 1269 1270 tqpair->recv_buf_size = (in_capsule_data_size + sizeof(struct spdk_nvme_tcp_cmd) + 2 * 1271 SPDK_NVME_TCP_DIGEST_LEN) * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1272 1273 return 0; 1274 } 1275 1276 static int 1277 nvmf_tcp_qpair_init(struct spdk_nvmf_qpair *qpair) 1278 { 1279 struct spdk_nvmf_tcp_qpair *tqpair; 1280 1281 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1282 1283 SPDK_DEBUGLOG(nvmf_tcp, "New TCP Connection: %p\n", qpair); 1284 1285 spdk_trace_record(TRACE_TCP_QP_CREATE, 0, 0, (uintptr_t)tqpair); 1286 1287 /* Initialise request state queues of the qpair */ 1288 TAILQ_INIT(&tqpair->tcp_req_free_queue); 1289 TAILQ_INIT(&tqpair->tcp_req_working_queue); 1290 SLIST_INIT(&tqpair->tcp_pdu_free_queue); 1291 1292 tqpair->host_hdgst_enable = true; 1293 tqpair->host_ddgst_enable = true; 1294 1295 return 0; 1296 } 1297 1298 static int 1299 nvmf_tcp_qpair_sock_init(struct spdk_nvmf_tcp_qpair *tqpair) 1300 { 1301 int rc; 1302 1303 spdk_trace_record(TRACE_TCP_QP_SOCK_INIT, 0, 0, (uintptr_t)tqpair); 1304 1305 /* set low water mark */ 1306 rc = spdk_sock_set_recvlowat(tqpair->sock, 1); 1307 if (rc != 0) { 1308 SPDK_ERRLOG("spdk_sock_set_recvlowat() failed\n"); 1309 return rc; 1310 } 1311 1312 return 0; 1313 } 1314 1315 static void 1316 nvmf_tcp_handle_connect(struct spdk_nvmf_transport *transport, 1317 struct spdk_nvmf_tcp_port *port, 1318 struct spdk_sock *sock) 1319 { 1320 struct spdk_nvmf_tcp_qpair *tqpair; 1321 int rc; 1322 1323 SPDK_DEBUGLOG(nvmf_tcp, "New connection accepted on %s port %s\n", 1324 port->trid->traddr, port->trid->trsvcid); 1325 1326 tqpair = calloc(1, sizeof(struct spdk_nvmf_tcp_qpair)); 1327 if (tqpair == NULL) { 1328 SPDK_ERRLOG("Could not allocate new connection.\n"); 1329 spdk_sock_close(&sock); 1330 return; 1331 } 1332 1333 tqpair->sock = sock; 1334 tqpair->state_cntr[TCP_REQUEST_STATE_FREE] = 0; 1335 tqpair->port = port; 1336 tqpair->qpair.transport = transport; 1337 1338 rc = spdk_sock_getaddr(tqpair->sock, tqpair->target_addr, 1339 sizeof(tqpair->target_addr), &tqpair->target_port, 1340 tqpair->initiator_addr, sizeof(tqpair->initiator_addr), 1341 &tqpair->initiator_port); 1342 if (rc < 0) { 1343 SPDK_ERRLOG("spdk_sock_getaddr() failed of tqpair=%p\n", tqpair); 1344 nvmf_tcp_qpair_destroy(tqpair); 1345 return; 1346 } 1347 1348 spdk_nvmf_tgt_new_qpair(transport->tgt, &tqpair->qpair); 1349 } 1350 1351 static uint32_t 1352 nvmf_tcp_port_accept(struct spdk_nvmf_transport *transport, struct spdk_nvmf_tcp_port *port) 1353 { 1354 struct spdk_sock *sock; 1355 uint32_t count = 0; 1356 int i; 1357 1358 for (i = 0; i < NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME; i++) { 1359 sock = spdk_sock_accept(port->listen_sock); 1360 if (sock == NULL) { 1361 break; 1362 } 1363 count++; 1364 nvmf_tcp_handle_connect(transport, port, sock); 1365 } 1366 1367 return count; 1368 } 1369 1370 static int 1371 nvmf_tcp_accept(void *ctx) 1372 { 1373 struct spdk_nvmf_transport *transport = ctx; 1374 struct spdk_nvmf_tcp_transport *ttransport; 1375 struct spdk_nvmf_tcp_port *port; 1376 uint32_t count = 0; 1377 1378 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1379 1380 TAILQ_FOREACH(port, &ttransport->ports, link) { 1381 count += nvmf_tcp_port_accept(transport, port); 1382 } 1383 1384 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 1385 } 1386 1387 static void 1388 nvmf_tcp_discover(struct spdk_nvmf_transport *transport, 1389 struct spdk_nvme_transport_id *trid, 1390 struct spdk_nvmf_discovery_log_page_entry *entry) 1391 { 1392 struct spdk_nvmf_tcp_port *port; 1393 struct spdk_nvmf_tcp_transport *ttransport; 1394 1395 entry->trtype = SPDK_NVMF_TRTYPE_TCP; 1396 entry->adrfam = trid->adrfam; 1397 1398 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 1399 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 1400 1401 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1402 port = nvmf_tcp_find_port(ttransport, trid); 1403 1404 assert(port != NULL); 1405 1406 if (strcmp(spdk_sock_get_impl_name(port->listen_sock), "ssl") == 0) { 1407 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_REQUIRED; 1408 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_TLS_1_3; 1409 } else { 1410 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 1411 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_NONE; 1412 } 1413 } 1414 1415 static struct spdk_nvmf_tcp_control_msg_list * 1416 nvmf_tcp_control_msg_list_create(uint16_t num_messages) 1417 { 1418 struct spdk_nvmf_tcp_control_msg_list *list; 1419 struct spdk_nvmf_tcp_control_msg *msg; 1420 uint16_t i; 1421 1422 list = calloc(1, sizeof(*list)); 1423 if (!list) { 1424 SPDK_ERRLOG("Failed to allocate memory for list structure\n"); 1425 return NULL; 1426 } 1427 1428 list->msg_buf = spdk_zmalloc(num_messages * SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, 1429 NVMF_DATA_BUFFER_ALIGNMENT, NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 1430 if (!list->msg_buf) { 1431 SPDK_ERRLOG("Failed to allocate memory for control message buffers\n"); 1432 free(list); 1433 return NULL; 1434 } 1435 1436 STAILQ_INIT(&list->free_msgs); 1437 1438 for (i = 0; i < num_messages; i++) { 1439 msg = (struct spdk_nvmf_tcp_control_msg *)((char *)list->msg_buf + i * 1440 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1441 STAILQ_INSERT_TAIL(&list->free_msgs, msg, link); 1442 } 1443 1444 return list; 1445 } 1446 1447 static void 1448 nvmf_tcp_control_msg_list_free(struct spdk_nvmf_tcp_control_msg_list *list) 1449 { 1450 if (!list) { 1451 return; 1452 } 1453 1454 spdk_free(list->msg_buf); 1455 free(list); 1456 } 1457 1458 static struct spdk_nvmf_transport_poll_group * 1459 nvmf_tcp_poll_group_create(struct spdk_nvmf_transport *transport, 1460 struct spdk_nvmf_poll_group *group) 1461 { 1462 struct spdk_nvmf_tcp_transport *ttransport; 1463 struct spdk_nvmf_tcp_poll_group *tgroup; 1464 1465 tgroup = calloc(1, sizeof(*tgroup)); 1466 if (!tgroup) { 1467 return NULL; 1468 } 1469 1470 tgroup->sock_group = spdk_sock_group_create(&tgroup->group); 1471 if (!tgroup->sock_group) { 1472 goto cleanup; 1473 } 1474 1475 TAILQ_INIT(&tgroup->qpairs); 1476 TAILQ_INIT(&tgroup->await_req); 1477 1478 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1479 1480 if (transport->opts.in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 1481 SPDK_DEBUGLOG(nvmf_tcp, "ICD %u is less than min required for admin/fabric commands (%u). " 1482 "Creating control messages list\n", transport->opts.in_capsule_data_size, 1483 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1484 tgroup->control_msg_list = nvmf_tcp_control_msg_list_create(ttransport->tcp_opts.control_msg_num); 1485 if (!tgroup->control_msg_list) { 1486 goto cleanup; 1487 } 1488 } 1489 1490 tgroup->accel_channel = spdk_accel_get_io_channel(); 1491 if (spdk_unlikely(!tgroup->accel_channel)) { 1492 SPDK_ERRLOG("Cannot create accel_channel for tgroup=%p\n", tgroup); 1493 goto cleanup; 1494 } 1495 1496 TAILQ_INSERT_TAIL(&ttransport->poll_groups, tgroup, link); 1497 if (ttransport->next_pg == NULL) { 1498 ttransport->next_pg = tgroup; 1499 } 1500 1501 return &tgroup->group; 1502 1503 cleanup: 1504 nvmf_tcp_poll_group_destroy(&tgroup->group); 1505 return NULL; 1506 } 1507 1508 static struct spdk_nvmf_transport_poll_group * 1509 nvmf_tcp_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 1510 { 1511 struct spdk_nvmf_tcp_transport *ttransport; 1512 struct spdk_nvmf_tcp_poll_group **pg; 1513 struct spdk_nvmf_tcp_qpair *tqpair; 1514 struct spdk_sock_group *group = NULL, *hint = NULL; 1515 int rc; 1516 1517 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 1518 1519 if (TAILQ_EMPTY(&ttransport->poll_groups)) { 1520 return NULL; 1521 } 1522 1523 pg = &ttransport->next_pg; 1524 assert(*pg != NULL); 1525 hint = (*pg)->sock_group; 1526 1527 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1528 rc = spdk_sock_get_optimal_sock_group(tqpair->sock, &group, hint); 1529 if (rc != 0) { 1530 return NULL; 1531 } else if (group != NULL) { 1532 /* Optimal poll group was found */ 1533 return spdk_sock_group_get_ctx(group); 1534 } 1535 1536 /* The hint was used for optimal poll group, advance next_pg. */ 1537 *pg = TAILQ_NEXT(*pg, link); 1538 if (*pg == NULL) { 1539 *pg = TAILQ_FIRST(&ttransport->poll_groups); 1540 } 1541 1542 return spdk_sock_group_get_ctx(hint); 1543 } 1544 1545 static void 1546 nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 1547 { 1548 struct spdk_nvmf_tcp_poll_group *tgroup, *next_tgroup; 1549 struct spdk_nvmf_tcp_transport *ttransport; 1550 1551 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 1552 spdk_sock_group_close(&tgroup->sock_group); 1553 if (tgroup->control_msg_list) { 1554 nvmf_tcp_control_msg_list_free(tgroup->control_msg_list); 1555 } 1556 1557 if (tgroup->accel_channel) { 1558 spdk_put_io_channel(tgroup->accel_channel); 1559 } 1560 1561 if (tgroup->group.transport == NULL) { 1562 /* Transport can be NULL when nvmf_tcp_poll_group_create() 1563 * calls this function directly in a failure path. */ 1564 free(tgroup); 1565 return; 1566 } 1567 1568 ttransport = SPDK_CONTAINEROF(tgroup->group.transport, struct spdk_nvmf_tcp_transport, transport); 1569 1570 next_tgroup = TAILQ_NEXT(tgroup, link); 1571 TAILQ_REMOVE(&ttransport->poll_groups, tgroup, link); 1572 if (next_tgroup == NULL) { 1573 next_tgroup = TAILQ_FIRST(&ttransport->poll_groups); 1574 } 1575 if (ttransport->next_pg == tgroup) { 1576 ttransport->next_pg = next_tgroup; 1577 } 1578 1579 free(tgroup); 1580 } 1581 1582 static void 1583 nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 1584 enum nvme_tcp_pdu_recv_state state) 1585 { 1586 if (tqpair->recv_state == state) { 1587 SPDK_ERRLOG("The recv state of tqpair=%p is same with the state(%d) to be set\n", 1588 tqpair, state); 1589 return; 1590 } 1591 1592 if (spdk_unlikely(state == NVME_TCP_PDU_RECV_STATE_QUIESCING)) { 1593 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH && tqpair->pdu_in_progress) { 1594 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, tqpair->pdu_in_progress, slist); 1595 tqpair->tcp_pdu_working_count--; 1596 } 1597 } 1598 1599 if (spdk_unlikely(state == NVME_TCP_PDU_RECV_STATE_ERROR)) { 1600 assert(tqpair->tcp_pdu_working_count == 0); 1601 } 1602 1603 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1604 /* When leaving the await req state, move the qpair to the main list */ 1605 TAILQ_REMOVE(&tqpair->group->await_req, tqpair, link); 1606 TAILQ_INSERT_TAIL(&tqpair->group->qpairs, tqpair, link); 1607 } else if (state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1608 TAILQ_REMOVE(&tqpair->group->qpairs, tqpair, link); 1609 TAILQ_INSERT_TAIL(&tqpair->group->await_req, tqpair, link); 1610 } 1611 1612 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv state=%d\n", tqpair, state); 1613 tqpair->recv_state = state; 1614 1615 spdk_trace_record(TRACE_TCP_QP_RCV_STATE_CHANGE, tqpair->qpair.qid, 0, (uintptr_t)tqpair, 1616 tqpair->recv_state); 1617 } 1618 1619 static int 1620 nvmf_tcp_qpair_handle_timeout(void *ctx) 1621 { 1622 struct spdk_nvmf_tcp_qpair *tqpair = ctx; 1623 1624 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1625 1626 SPDK_ERRLOG("No pdu coming for tqpair=%p within %d seconds\n", tqpair, 1627 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT); 1628 1629 nvmf_tcp_qpair_disconnect(tqpair); 1630 return SPDK_POLLER_BUSY; 1631 } 1632 1633 static void 1634 nvmf_tcp_send_c2h_term_req_complete(void *cb_arg) 1635 { 1636 struct spdk_nvmf_tcp_qpair *tqpair = (struct spdk_nvmf_tcp_qpair *)cb_arg; 1637 1638 if (!tqpair->timeout_poller) { 1639 tqpair->timeout_poller = SPDK_POLLER_REGISTER(nvmf_tcp_qpair_handle_timeout, tqpair, 1640 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT * 1000000); 1641 } 1642 } 1643 1644 static void 1645 nvmf_tcp_send_c2h_term_req(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu, 1646 enum spdk_nvme_tcp_term_req_fes fes, uint32_t error_offset) 1647 { 1648 struct nvme_tcp_pdu *rsp_pdu; 1649 struct spdk_nvme_tcp_term_req_hdr *c2h_term_req; 1650 uint32_t c2h_term_req_hdr_len = sizeof(*c2h_term_req); 1651 uint32_t copy_len; 1652 1653 rsp_pdu = tqpair->mgmt_pdu; 1654 1655 c2h_term_req = &rsp_pdu->hdr.term_req; 1656 c2h_term_req->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ; 1657 c2h_term_req->common.hlen = c2h_term_req_hdr_len; 1658 c2h_term_req->fes = fes; 1659 1660 if ((fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1661 (fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1662 DSET32(&c2h_term_req->fei, error_offset); 1663 } 1664 1665 copy_len = spdk_min(pdu->hdr.common.hlen, SPDK_NVME_TCP_TERM_REQ_ERROR_DATA_MAX_SIZE); 1666 1667 /* Copy the error info into the buffer */ 1668 memcpy((uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, pdu->hdr.raw, copy_len); 1669 nvme_tcp_pdu_set_data(rsp_pdu, (uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, copy_len); 1670 1671 /* Contain the header of the wrong received pdu */ 1672 c2h_term_req->common.plen = c2h_term_req->common.hlen + copy_len; 1673 tqpair->wait_terminate = true; 1674 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1675 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_c2h_term_req_complete, tqpair); 1676 } 1677 1678 static void 1679 nvmf_tcp_capsule_cmd_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1680 struct spdk_nvmf_tcp_qpair *tqpair, 1681 struct nvme_tcp_pdu *pdu) 1682 { 1683 struct spdk_nvmf_tcp_req *tcp_req; 1684 1685 assert(pdu->psh_valid_bytes == pdu->psh_len); 1686 assert(pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD); 1687 1688 tcp_req = nvmf_tcp_req_get(tqpair); 1689 if (!tcp_req) { 1690 /* Directly return and make the allocation retry again. This can happen if we're 1691 * using asynchronous writes to send the response to the host or when releasing 1692 * zero-copy buffers after a response has been sent. In both cases, the host might 1693 * receive the response before we've finished processing the request and is free to 1694 * send another one. 1695 */ 1696 if (tqpair->state_cntr[TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST] > 0 || 1697 tqpair->state_cntr[TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE] > 0) { 1698 return; 1699 } 1700 1701 /* The host sent more commands than the maximum queue depth. */ 1702 SPDK_ERRLOG("Cannot allocate tcp_req on tqpair=%p\n", tqpair); 1703 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1704 return; 1705 } 1706 1707 pdu->req = tcp_req; 1708 assert(tcp_req->state == TCP_REQUEST_STATE_NEW); 1709 nvmf_tcp_req_process(ttransport, tcp_req); 1710 } 1711 1712 static void 1713 nvmf_tcp_capsule_cmd_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1714 struct spdk_nvmf_tcp_qpair *tqpair, 1715 struct nvme_tcp_pdu *pdu) 1716 { 1717 struct spdk_nvmf_tcp_req *tcp_req; 1718 struct spdk_nvme_tcp_cmd *capsule_cmd; 1719 uint32_t error_offset = 0; 1720 enum spdk_nvme_tcp_term_req_fes fes; 1721 struct spdk_nvme_cpl *rsp; 1722 1723 capsule_cmd = &pdu->hdr.capsule_cmd; 1724 tcp_req = pdu->req; 1725 assert(tcp_req != NULL); 1726 1727 /* Zero-copy requests don't support ICD */ 1728 assert(!spdk_nvmf_request_using_zcopy(&tcp_req->req)); 1729 1730 if (capsule_cmd->common.pdo > SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET) { 1731 SPDK_ERRLOG("Expected ICReq capsule_cmd pdu offset <= %d, got %c\n", 1732 SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET, capsule_cmd->common.pdo); 1733 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1734 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 1735 goto err; 1736 } 1737 1738 rsp = &tcp_req->req.rsp->nvme_cpl; 1739 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1740 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1741 } else { 1742 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1743 } 1744 1745 nvmf_tcp_req_process(ttransport, tcp_req); 1746 1747 return; 1748 err: 1749 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1750 } 1751 1752 static void 1753 nvmf_tcp_h2c_data_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1754 struct spdk_nvmf_tcp_qpair *tqpair, 1755 struct nvme_tcp_pdu *pdu) 1756 { 1757 struct spdk_nvmf_tcp_req *tcp_req; 1758 uint32_t error_offset = 0; 1759 enum spdk_nvme_tcp_term_req_fes fes = 0; 1760 struct spdk_nvme_tcp_h2c_data_hdr *h2c_data; 1761 1762 h2c_data = &pdu->hdr.h2c_data; 1763 1764 SPDK_DEBUGLOG(nvmf_tcp, "tqpair=%p, r2t_info: datao=%u, datal=%u, cccid=%u, ttag=%u\n", 1765 tqpair, h2c_data->datao, h2c_data->datal, h2c_data->cccid, h2c_data->ttag); 1766 1767 if (h2c_data->ttag > tqpair->resource_count) { 1768 SPDK_DEBUGLOG(nvmf_tcp, "ttag %u is larger than allowed %u.\n", h2c_data->ttag, 1769 tqpair->resource_count); 1770 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1771 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1772 goto err; 1773 } 1774 1775 tcp_req = &tqpair->reqs[h2c_data->ttag - 1]; 1776 1777 if (spdk_unlikely(tcp_req->state != TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER && 1778 tcp_req->state != TCP_REQUEST_STATE_AWAITING_R2T_ACK)) { 1779 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, has error state in %d\n", tcp_req, tqpair, 1780 tcp_req->state); 1781 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1782 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1783 goto err; 1784 } 1785 1786 if (spdk_unlikely(tcp_req->req.cmd->nvme_cmd.cid != h2c_data->cccid)) { 1787 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, expected %u but %u for cccid.\n", tcp_req, tqpair, 1788 tcp_req->req.cmd->nvme_cmd.cid, h2c_data->cccid); 1789 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1790 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, cccid); 1791 goto err; 1792 } 1793 1794 if (tcp_req->h2c_offset != h2c_data->datao) { 1795 SPDK_DEBUGLOG(nvmf_tcp, 1796 "tcp_req(%p), tqpair=%p, expected data offset %u, but data offset is %u\n", 1797 tcp_req, tqpair, tcp_req->h2c_offset, h2c_data->datao); 1798 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1799 goto err; 1800 } 1801 1802 if ((h2c_data->datao + h2c_data->datal) > tcp_req->req.length) { 1803 SPDK_DEBUGLOG(nvmf_tcp, 1804 "tcp_req(%p), tqpair=%p, (datao=%u + datal=%u) exceeds requested length=%u\n", 1805 tcp_req, tqpair, h2c_data->datao, h2c_data->datal, tcp_req->req.length); 1806 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1807 goto err; 1808 } 1809 1810 pdu->req = tcp_req; 1811 1812 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 1813 pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 1814 } 1815 1816 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 1817 h2c_data->datao, h2c_data->datal); 1818 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1819 return; 1820 1821 err: 1822 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1823 } 1824 1825 static void 1826 nvmf_tcp_send_capsule_resp_pdu(struct spdk_nvmf_tcp_req *tcp_req, 1827 struct spdk_nvmf_tcp_qpair *tqpair) 1828 { 1829 struct nvme_tcp_pdu *rsp_pdu; 1830 struct spdk_nvme_tcp_rsp *capsule_resp; 1831 1832 SPDK_DEBUGLOG(nvmf_tcp, "enter, tqpair=%p\n", tqpair); 1833 1834 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1835 assert(rsp_pdu != NULL); 1836 1837 capsule_resp = &rsp_pdu->hdr.capsule_resp; 1838 capsule_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_CAPSULE_RESP; 1839 capsule_resp->common.plen = capsule_resp->common.hlen = sizeof(*capsule_resp); 1840 capsule_resp->rccqe = tcp_req->req.rsp->nvme_cpl; 1841 if (tqpair->host_hdgst_enable) { 1842 capsule_resp->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 1843 capsule_resp->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 1844 } 1845 1846 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_request_free, tcp_req); 1847 } 1848 1849 static void 1850 nvmf_tcp_pdu_c2h_data_complete(void *cb_arg) 1851 { 1852 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 1853 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, 1854 struct spdk_nvmf_tcp_qpair, qpair); 1855 1856 assert(tqpair != NULL); 1857 1858 if (spdk_unlikely(tcp_req->pdu->rw_offset < tcp_req->req.length)) { 1859 SPDK_DEBUGLOG(nvmf_tcp, "sending another C2H part, offset %u length %u\n", tcp_req->pdu->rw_offset, 1860 tcp_req->req.length); 1861 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 1862 return; 1863 } 1864 1865 if (tcp_req->pdu->hdr.c2h_data.common.flags & SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS) { 1866 nvmf_tcp_request_free(tcp_req); 1867 } else { 1868 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 1869 } 1870 } 1871 1872 static void 1873 nvmf_tcp_r2t_complete(void *cb_arg) 1874 { 1875 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 1876 struct spdk_nvmf_tcp_transport *ttransport; 1877 1878 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 1879 struct spdk_nvmf_tcp_transport, transport); 1880 1881 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1882 1883 if (tcp_req->h2c_offset == tcp_req->req.length) { 1884 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1885 nvmf_tcp_req_process(ttransport, tcp_req); 1886 } 1887 } 1888 1889 static void 1890 nvmf_tcp_send_r2t_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1891 struct spdk_nvmf_tcp_req *tcp_req) 1892 { 1893 struct nvme_tcp_pdu *rsp_pdu; 1894 struct spdk_nvme_tcp_r2t_hdr *r2t; 1895 1896 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1897 assert(rsp_pdu != NULL); 1898 1899 r2t = &rsp_pdu->hdr.r2t; 1900 r2t->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_R2T; 1901 r2t->common.plen = r2t->common.hlen = sizeof(*r2t); 1902 1903 if (tqpair->host_hdgst_enable) { 1904 r2t->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 1905 r2t->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 1906 } 1907 1908 r2t->cccid = tcp_req->req.cmd->nvme_cmd.cid; 1909 r2t->ttag = tcp_req->ttag; 1910 r2t->r2to = tcp_req->h2c_offset; 1911 r2t->r2tl = tcp_req->req.length; 1912 1913 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 1914 1915 SPDK_DEBUGLOG(nvmf_tcp, 1916 "tcp_req(%p) on tqpair(%p), r2t_info: cccid=%u, ttag=%u, r2to=%u, r2tl=%u\n", 1917 tcp_req, tqpair, r2t->cccid, r2t->ttag, r2t->r2to, r2t->r2tl); 1918 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_r2t_complete, tcp_req); 1919 } 1920 1921 static void 1922 nvmf_tcp_h2c_data_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1923 struct spdk_nvmf_tcp_qpair *tqpair, 1924 struct nvme_tcp_pdu *pdu) 1925 { 1926 struct spdk_nvmf_tcp_req *tcp_req; 1927 struct spdk_nvme_cpl *rsp; 1928 1929 tcp_req = pdu->req; 1930 assert(tcp_req != NULL); 1931 1932 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 1933 1934 tcp_req->h2c_offset += pdu->data_len; 1935 1936 /* Wait for all of the data to arrive AND for the initial R2T PDU send to be 1937 * acknowledged before moving on. */ 1938 if (tcp_req->h2c_offset == tcp_req->req.length && 1939 tcp_req->state == TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER) { 1940 /* After receiving all the h2c data, we need to check whether there is 1941 * transient transport error */ 1942 rsp = &tcp_req->req.rsp->nvme_cpl; 1943 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1944 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1945 } else { 1946 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1947 } 1948 nvmf_tcp_req_process(ttransport, tcp_req); 1949 } 1950 } 1951 1952 static void 1953 nvmf_tcp_h2c_term_req_dump(struct spdk_nvme_tcp_term_req_hdr *h2c_term_req) 1954 { 1955 SPDK_ERRLOG("Error info of pdu(%p): %s\n", h2c_term_req, 1956 spdk_nvmf_tcp_term_req_fes_str[h2c_term_req->fes]); 1957 if ((h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1958 (h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1959 SPDK_DEBUGLOG(nvmf_tcp, "The offset from the start of the PDU header is %u\n", 1960 DGET32(h2c_term_req->fei)); 1961 } 1962 } 1963 1964 static void 1965 nvmf_tcp_h2c_term_req_hdr_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1966 struct nvme_tcp_pdu *pdu) 1967 { 1968 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 1969 uint32_t error_offset = 0; 1970 enum spdk_nvme_tcp_term_req_fes fes; 1971 1972 if (h2c_term_req->fes > SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER) { 1973 SPDK_ERRLOG("Fatal Error Status(FES) is unknown for h2c_term_req pdu=%p\n", pdu); 1974 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1975 error_offset = offsetof(struct spdk_nvme_tcp_term_req_hdr, fes); 1976 goto end; 1977 } 1978 1979 /* set the data buffer */ 1980 nvme_tcp_pdu_set_data(pdu, (uint8_t *)pdu->hdr.raw + h2c_term_req->common.hlen, 1981 h2c_term_req->common.plen - h2c_term_req->common.hlen); 1982 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1983 return; 1984 end: 1985 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1986 } 1987 1988 static void 1989 nvmf_tcp_h2c_term_req_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1990 struct nvme_tcp_pdu *pdu) 1991 { 1992 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 1993 1994 nvmf_tcp_h2c_term_req_dump(h2c_term_req); 1995 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1996 } 1997 1998 static void 1999 _nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 2000 { 2001 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2002 struct spdk_nvmf_tcp_transport, transport); 2003 2004 switch (pdu->hdr.common.pdu_type) { 2005 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2006 nvmf_tcp_capsule_cmd_payload_handle(ttransport, tqpair, pdu); 2007 break; 2008 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2009 nvmf_tcp_h2c_data_payload_handle(ttransport, tqpair, pdu); 2010 break; 2011 2012 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2013 nvmf_tcp_h2c_term_req_payload_handle(tqpair, pdu); 2014 break; 2015 2016 default: 2017 /* The code should not go to here */ 2018 SPDK_ERRLOG("ERROR pdu type %d\n", pdu->hdr.common.pdu_type); 2019 break; 2020 } 2021 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 2022 tqpair->tcp_pdu_working_count--; 2023 } 2024 2025 static inline void 2026 nvmf_tcp_req_set_cpl(struct spdk_nvmf_tcp_req *treq, int sct, int sc) 2027 { 2028 treq->req.rsp->nvme_cpl.status.sct = sct; 2029 treq->req.rsp->nvme_cpl.status.sc = sc; 2030 treq->req.rsp->nvme_cpl.cid = treq->req.cmd->nvme_cmd.cid; 2031 } 2032 2033 static void 2034 data_crc32_calc_done(void *cb_arg, int status) 2035 { 2036 struct nvme_tcp_pdu *pdu = cb_arg; 2037 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 2038 2039 /* async crc32 calculation is failed and use direct calculation to check */ 2040 if (spdk_unlikely(status)) { 2041 SPDK_ERRLOG("Data digest on tqpair=(%p) with pdu=%p failed to be calculated asynchronously\n", 2042 tqpair, pdu); 2043 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 2044 } 2045 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 2046 if (!MATCH_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32)) { 2047 SPDK_ERRLOG("Data digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 2048 assert(pdu->req != NULL); 2049 nvmf_tcp_req_set_cpl(pdu->req, SPDK_NVME_SCT_GENERIC, 2050 SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR); 2051 } 2052 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2053 } 2054 2055 static void 2056 nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 2057 { 2058 int rc = 0; 2059 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2060 tqpair->pdu_in_progress = NULL; 2061 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2062 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2063 /* check data digest if need */ 2064 if (pdu->ddgst_enable) { 2065 if (tqpair->qpair.qid != 0 && !pdu->dif_ctx && tqpair->group && 2066 (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0)) { 2067 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 2068 pdu->data_iovcnt, 0, data_crc32_calc_done, pdu); 2069 if (spdk_likely(rc == 0)) { 2070 return; 2071 } 2072 } else { 2073 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 2074 } 2075 data_crc32_calc_done(pdu, rc); 2076 } else { 2077 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2078 } 2079 } 2080 2081 static void 2082 nvmf_tcp_send_icresp_complete(void *cb_arg) 2083 { 2084 struct spdk_nvmf_tcp_qpair *tqpair = cb_arg; 2085 2086 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_RUNNING); 2087 } 2088 2089 static void 2090 nvmf_tcp_icreq_handle(struct spdk_nvmf_tcp_transport *ttransport, 2091 struct spdk_nvmf_tcp_qpair *tqpair, 2092 struct nvme_tcp_pdu *pdu) 2093 { 2094 struct spdk_nvme_tcp_ic_req *ic_req = &pdu->hdr.ic_req; 2095 struct nvme_tcp_pdu *rsp_pdu; 2096 struct spdk_nvme_tcp_ic_resp *ic_resp; 2097 uint32_t error_offset = 0; 2098 enum spdk_nvme_tcp_term_req_fes fes; 2099 2100 /* Only PFV 0 is defined currently */ 2101 if (ic_req->pfv != 0) { 2102 SPDK_ERRLOG("Expected ICReq PFV %u, got %u\n", 0u, ic_req->pfv); 2103 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2104 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, pfv); 2105 goto end; 2106 } 2107 2108 /* This value is 0’s based value in units of dwords should not be larger than SPDK_NVME_TCP_HPDA_MAX */ 2109 if (ic_req->hpda > SPDK_NVME_TCP_HPDA_MAX) { 2110 SPDK_ERRLOG("ICReq HPDA out of range 0 to 31, got %u\n", ic_req->hpda); 2111 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2112 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, hpda); 2113 goto end; 2114 } 2115 2116 /* MAXR2T is 0's based */ 2117 SPDK_DEBUGLOG(nvmf_tcp, "maxr2t =%u\n", (ic_req->maxr2t + 1u)); 2118 2119 tqpair->host_hdgst_enable = ic_req->dgst.bits.hdgst_enable ? true : false; 2120 if (!tqpair->host_hdgst_enable) { 2121 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 2122 } 2123 2124 tqpair->host_ddgst_enable = ic_req->dgst.bits.ddgst_enable ? true : false; 2125 if (!tqpair->host_ddgst_enable) { 2126 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 2127 } 2128 2129 tqpair->recv_buf_size = spdk_max(tqpair->recv_buf_size, MIN_SOCK_PIPE_SIZE); 2130 /* Now that we know whether digests are enabled, properly size the receive buffer */ 2131 if (spdk_sock_set_recvbuf(tqpair->sock, tqpair->recv_buf_size) < 0) { 2132 SPDK_WARNLOG("Unable to allocate enough memory for receive buffer on tqpair=%p with size=%d\n", 2133 tqpair, 2134 tqpair->recv_buf_size); 2135 /* Not fatal. */ 2136 } 2137 2138 tqpair->cpda = spdk_min(ic_req->hpda, SPDK_NVME_TCP_CPDA_MAX); 2139 SPDK_DEBUGLOG(nvmf_tcp, "cpda of tqpair=(%p) is : %u\n", tqpair, tqpair->cpda); 2140 2141 rsp_pdu = tqpair->mgmt_pdu; 2142 2143 ic_resp = &rsp_pdu->hdr.ic_resp; 2144 ic_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_IC_RESP; 2145 ic_resp->common.hlen = ic_resp->common.plen = sizeof(*ic_resp); 2146 ic_resp->pfv = 0; 2147 ic_resp->cpda = tqpair->cpda; 2148 ic_resp->maxh2cdata = ttransport->transport.opts.max_io_size; 2149 ic_resp->dgst.bits.hdgst_enable = tqpair->host_hdgst_enable ? 1 : 0; 2150 ic_resp->dgst.bits.ddgst_enable = tqpair->host_ddgst_enable ? 1 : 0; 2151 2152 SPDK_DEBUGLOG(nvmf_tcp, "host_hdgst_enable: %u\n", tqpair->host_hdgst_enable); 2153 SPDK_DEBUGLOG(nvmf_tcp, "host_ddgst_enable: %u\n", tqpair->host_ddgst_enable); 2154 2155 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INITIALIZING); 2156 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_icresp_complete, tqpair); 2157 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2158 return; 2159 end: 2160 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2161 } 2162 2163 static void 2164 nvmf_tcp_pdu_psh_handle(struct spdk_nvmf_tcp_qpair *tqpair, 2165 struct spdk_nvmf_tcp_transport *ttransport) 2166 { 2167 struct nvme_tcp_pdu *pdu; 2168 int rc; 2169 uint32_t crc32c, error_offset = 0; 2170 enum spdk_nvme_tcp_term_req_fes fes; 2171 2172 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2173 pdu = tqpair->pdu_in_progress; 2174 2175 SPDK_DEBUGLOG(nvmf_tcp, "pdu type of tqpair(%p) is %d\n", tqpair, 2176 pdu->hdr.common.pdu_type); 2177 /* check header digest if needed */ 2178 if (pdu->has_hdgst) { 2179 SPDK_DEBUGLOG(nvmf_tcp, "Compare the header of pdu=%p on tqpair=%p\n", pdu, tqpair); 2180 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 2181 rc = MATCH_DIGEST_WORD((uint8_t *)pdu->hdr.raw + pdu->hdr.common.hlen, crc32c); 2182 if (rc == 0) { 2183 SPDK_ERRLOG("Header digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 2184 fes = SPDK_NVME_TCP_TERM_REQ_FES_HDGST_ERROR; 2185 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2186 return; 2187 2188 } 2189 } 2190 2191 switch (pdu->hdr.common.pdu_type) { 2192 case SPDK_NVME_TCP_PDU_TYPE_IC_REQ: 2193 nvmf_tcp_icreq_handle(ttransport, tqpair, pdu); 2194 break; 2195 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2196 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_REQ); 2197 break; 2198 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2199 nvmf_tcp_h2c_data_hdr_handle(ttransport, tqpair, pdu); 2200 break; 2201 2202 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2203 nvmf_tcp_h2c_term_req_hdr_handle(tqpair, pdu); 2204 break; 2205 2206 default: 2207 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", tqpair->pdu_in_progress->hdr.common.pdu_type); 2208 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2209 error_offset = 1; 2210 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2211 break; 2212 } 2213 } 2214 2215 static void 2216 nvmf_tcp_pdu_ch_handle(struct spdk_nvmf_tcp_qpair *tqpair) 2217 { 2218 struct nvme_tcp_pdu *pdu; 2219 uint32_t error_offset = 0; 2220 enum spdk_nvme_tcp_term_req_fes fes; 2221 uint8_t expected_hlen, pdo; 2222 bool plen_error = false, pdo_error = false; 2223 2224 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2225 pdu = tqpair->pdu_in_progress; 2226 assert(pdu); 2227 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_REQ) { 2228 if (tqpair->state != NVME_TCP_QPAIR_STATE_INVALID) { 2229 SPDK_ERRLOG("Already received ICreq PDU, and reject this pdu=%p\n", pdu); 2230 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2231 goto err; 2232 } 2233 expected_hlen = sizeof(struct spdk_nvme_tcp_ic_req); 2234 if (pdu->hdr.common.plen != expected_hlen) { 2235 plen_error = true; 2236 } 2237 } else { 2238 if (tqpair->state != NVME_TCP_QPAIR_STATE_RUNNING) { 2239 SPDK_ERRLOG("The TCP/IP connection is not negotiated\n"); 2240 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2241 goto err; 2242 } 2243 2244 switch (pdu->hdr.common.pdu_type) { 2245 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2246 expected_hlen = sizeof(struct spdk_nvme_tcp_cmd); 2247 pdo = pdu->hdr.common.pdo; 2248 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2249 pdo_error = true; 2250 break; 2251 } 2252 2253 if (pdu->hdr.common.plen < expected_hlen) { 2254 plen_error = true; 2255 } 2256 break; 2257 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2258 expected_hlen = sizeof(struct spdk_nvme_tcp_h2c_data_hdr); 2259 pdo = pdu->hdr.common.pdo; 2260 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2261 pdo_error = true; 2262 break; 2263 } 2264 if (pdu->hdr.common.plen < expected_hlen) { 2265 plen_error = true; 2266 } 2267 break; 2268 2269 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2270 expected_hlen = sizeof(struct spdk_nvme_tcp_term_req_hdr); 2271 if ((pdu->hdr.common.plen <= expected_hlen) || 2272 (pdu->hdr.common.plen > SPDK_NVME_TCP_TERM_REQ_PDU_MAX_SIZE)) { 2273 plen_error = true; 2274 } 2275 break; 2276 2277 default: 2278 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", pdu->hdr.common.pdu_type); 2279 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2280 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdu_type); 2281 goto err; 2282 } 2283 } 2284 2285 if (pdu->hdr.common.hlen != expected_hlen) { 2286 SPDK_ERRLOG("PDU type=0x%02x, Expected ICReq header length %u, got %u on tqpair=%p\n", 2287 pdu->hdr.common.pdu_type, 2288 expected_hlen, pdu->hdr.common.hlen, tqpair); 2289 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2290 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, hlen); 2291 goto err; 2292 } else if (pdo_error) { 2293 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2294 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 2295 } else if (plen_error) { 2296 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2297 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2298 goto err; 2299 } else { 2300 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2301 nvme_tcp_pdu_calc_psh_len(tqpair->pdu_in_progress, tqpair->host_hdgst_enable); 2302 return; 2303 } 2304 err: 2305 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2306 } 2307 2308 static int 2309 nvmf_tcp_sock_process(struct spdk_nvmf_tcp_qpair *tqpair) 2310 { 2311 int rc = 0; 2312 struct nvme_tcp_pdu *pdu; 2313 enum nvme_tcp_pdu_recv_state prev_state; 2314 uint32_t data_len; 2315 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2316 struct spdk_nvmf_tcp_transport, transport); 2317 2318 /* The loop here is to allow for several back-to-back state changes. */ 2319 do { 2320 prev_state = tqpair->recv_state; 2321 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv pdu entering state %d\n", tqpair, prev_state); 2322 2323 pdu = tqpair->pdu_in_progress; 2324 assert(pdu != NULL || 2325 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY || 2326 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_QUIESCING || 2327 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 2328 2329 switch (tqpair->recv_state) { 2330 /* Wait for the common header */ 2331 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY: 2332 if (!pdu) { 2333 pdu = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 2334 if (spdk_unlikely(!pdu)) { 2335 return NVME_TCP_PDU_IN_PROGRESS; 2336 } 2337 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 2338 tqpair->pdu_in_progress = pdu; 2339 tqpair->tcp_pdu_working_count++; 2340 } 2341 memset(pdu, 0, offsetof(struct nvme_tcp_pdu, qpair)); 2342 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2343 /* FALLTHROUGH */ 2344 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH: 2345 if (spdk_unlikely(tqpair->state == NVME_TCP_QPAIR_STATE_INITIALIZING)) { 2346 return rc; 2347 } 2348 2349 rc = nvme_tcp_read_data(tqpair->sock, 2350 sizeof(struct spdk_nvme_tcp_common_pdu_hdr) - pdu->ch_valid_bytes, 2351 (void *)&pdu->hdr.common + pdu->ch_valid_bytes); 2352 if (rc < 0) { 2353 SPDK_DEBUGLOG(nvmf_tcp, "will disconnect tqpair=%p\n", tqpair); 2354 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2355 break; 2356 } else if (rc > 0) { 2357 pdu->ch_valid_bytes += rc; 2358 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.qid, rc, 0, tqpair); 2359 } 2360 2361 if (pdu->ch_valid_bytes < sizeof(struct spdk_nvme_tcp_common_pdu_hdr)) { 2362 return NVME_TCP_PDU_IN_PROGRESS; 2363 } 2364 2365 /* The command header of this PDU has now been read from the socket. */ 2366 nvmf_tcp_pdu_ch_handle(tqpair); 2367 break; 2368 /* Wait for the pdu specific header */ 2369 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH: 2370 rc = nvme_tcp_read_data(tqpair->sock, 2371 pdu->psh_len - pdu->psh_valid_bytes, 2372 (void *)&pdu->hdr.raw + sizeof(struct spdk_nvme_tcp_common_pdu_hdr) + pdu->psh_valid_bytes); 2373 if (rc < 0) { 2374 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2375 break; 2376 } else if (rc > 0) { 2377 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.qid, rc, 0, tqpair); 2378 pdu->psh_valid_bytes += rc; 2379 } 2380 2381 if (pdu->psh_valid_bytes < pdu->psh_len) { 2382 return NVME_TCP_PDU_IN_PROGRESS; 2383 } 2384 2385 /* All header(ch, psh, head digist) of this PDU has now been read from the socket. */ 2386 nvmf_tcp_pdu_psh_handle(tqpair, ttransport); 2387 break; 2388 /* Wait for the req slot */ 2389 case NVME_TCP_PDU_RECV_STATE_AWAIT_REQ: 2390 nvmf_tcp_capsule_cmd_hdr_handle(ttransport, tqpair, pdu); 2391 break; 2392 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD: 2393 /* check whether the data is valid, if not we just return */ 2394 if (!pdu->data_len) { 2395 return NVME_TCP_PDU_IN_PROGRESS; 2396 } 2397 2398 data_len = pdu->data_len; 2399 /* data digest */ 2400 if (spdk_unlikely((pdu->hdr.common.pdu_type != SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ) && 2401 tqpair->host_ddgst_enable)) { 2402 data_len += SPDK_NVME_TCP_DIGEST_LEN; 2403 pdu->ddgst_enable = true; 2404 } 2405 2406 rc = nvme_tcp_read_payload_data(tqpair->sock, pdu); 2407 if (rc < 0) { 2408 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2409 break; 2410 } 2411 pdu->rw_offset += rc; 2412 2413 if (pdu->rw_offset < data_len) { 2414 return NVME_TCP_PDU_IN_PROGRESS; 2415 } 2416 2417 /* Generate and insert DIF to whole data block received if DIF is enabled */ 2418 if (spdk_unlikely(pdu->dif_ctx != NULL) && 2419 spdk_dif_generate_stream(pdu->data_iov, pdu->data_iovcnt, 0, data_len, 2420 pdu->dif_ctx) != 0) { 2421 SPDK_ERRLOG("DIF generate failed\n"); 2422 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2423 break; 2424 } 2425 2426 /* All of this PDU has now been read from the socket. */ 2427 nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2428 break; 2429 case NVME_TCP_PDU_RECV_STATE_QUIESCING: 2430 if (tqpair->tcp_pdu_working_count != 0) { 2431 return NVME_TCP_PDU_IN_PROGRESS; 2432 } 2433 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 2434 break; 2435 case NVME_TCP_PDU_RECV_STATE_ERROR: 2436 if (spdk_sock_is_connected(tqpair->sock) && tqpair->wait_terminate) { 2437 return NVME_TCP_PDU_IN_PROGRESS; 2438 } 2439 return NVME_TCP_PDU_FATAL; 2440 default: 2441 SPDK_ERRLOG("The state(%d) is invalid\n", tqpair->recv_state); 2442 abort(); 2443 break; 2444 } 2445 } while (tqpair->recv_state != prev_state); 2446 2447 return rc; 2448 } 2449 2450 static inline void * 2451 nvmf_tcp_control_msg_get(struct spdk_nvmf_tcp_control_msg_list *list) 2452 { 2453 struct spdk_nvmf_tcp_control_msg *msg; 2454 2455 assert(list); 2456 2457 msg = STAILQ_FIRST(&list->free_msgs); 2458 if (!msg) { 2459 SPDK_DEBUGLOG(nvmf_tcp, "Out of control messages\n"); 2460 return NULL; 2461 } 2462 STAILQ_REMOVE_HEAD(&list->free_msgs, link); 2463 return msg; 2464 } 2465 2466 static inline void 2467 nvmf_tcp_control_msg_put(struct spdk_nvmf_tcp_control_msg_list *list, void *_msg) 2468 { 2469 struct spdk_nvmf_tcp_control_msg *msg = _msg; 2470 2471 assert(list); 2472 STAILQ_INSERT_HEAD(&list->free_msgs, msg, link); 2473 } 2474 2475 static int 2476 nvmf_tcp_req_parse_sgl(struct spdk_nvmf_tcp_req *tcp_req, 2477 struct spdk_nvmf_transport *transport, 2478 struct spdk_nvmf_transport_poll_group *group) 2479 { 2480 struct spdk_nvmf_request *req = &tcp_req->req; 2481 struct spdk_nvme_cmd *cmd; 2482 struct spdk_nvme_sgl_descriptor *sgl; 2483 struct spdk_nvmf_tcp_poll_group *tgroup; 2484 enum spdk_nvme_tcp_term_req_fes fes; 2485 struct nvme_tcp_pdu *pdu; 2486 struct spdk_nvmf_tcp_qpair *tqpair; 2487 uint32_t length, error_offset = 0; 2488 2489 cmd = &req->cmd->nvme_cmd; 2490 sgl = &cmd->dptr.sgl1; 2491 2492 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK && 2493 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_TRANSPORT) { 2494 /* get request length from sgl */ 2495 length = sgl->unkeyed.length; 2496 if (spdk_unlikely(length > transport->opts.max_io_size)) { 2497 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 2498 length, transport->opts.max_io_size); 2499 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_LIMIT_EXCEEDED; 2500 goto fatal_err; 2501 } 2502 2503 /* fill request length and populate iovs */ 2504 req->length = length; 2505 2506 SPDK_DEBUGLOG(nvmf_tcp, "Data requested length= 0x%x\n", length); 2507 2508 if (spdk_unlikely(req->dif_enabled)) { 2509 req->dif.orig_length = length; 2510 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2511 req->dif.elba_length = length; 2512 } 2513 2514 if (nvmf_ctrlr_use_zcopy(req)) { 2515 SPDK_DEBUGLOG(nvmf_tcp, "Using zero-copy to execute request %p\n", tcp_req); 2516 req->data_from_pool = false; 2517 return 0; 2518 } 2519 2520 if (spdk_nvmf_request_get_buffers(req, group, transport, length)) { 2521 /* No available buffers. Queue this request up. */ 2522 SPDK_DEBUGLOG(nvmf_tcp, "No available large data buffers. Queueing request %p\n", 2523 tcp_req); 2524 return 0; 2525 } 2526 2527 SPDK_DEBUGLOG(nvmf_tcp, "Request %p took %d buffer/s from central pool, and data=%p\n", 2528 tcp_req, req->iovcnt, req->iov[0].iov_base); 2529 2530 return 0; 2531 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 2532 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 2533 uint64_t offset = sgl->address; 2534 uint32_t max_len = transport->opts.in_capsule_data_size; 2535 2536 assert(tcp_req->has_in_capsule_data); 2537 /* Capsule Cmd with In-capsule Data should get data length from pdu header */ 2538 tqpair = tcp_req->pdu->qpair; 2539 /* receiving pdu is not same with the pdu in tcp_req */ 2540 pdu = tqpair->pdu_in_progress; 2541 length = pdu->hdr.common.plen - pdu->psh_len - sizeof(struct spdk_nvme_tcp_common_pdu_hdr); 2542 if (tqpair->host_ddgst_enable) { 2543 length -= SPDK_NVME_TCP_DIGEST_LEN; 2544 } 2545 /* This error is not defined in NVMe/TCP spec, take this error as fatal error */ 2546 if (spdk_unlikely(length != sgl->unkeyed.length)) { 2547 SPDK_ERRLOG("In-Capsule Data length 0x%x is not equal to SGL data length 0x%x\n", 2548 length, sgl->unkeyed.length); 2549 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2550 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2551 goto fatal_err; 2552 } 2553 2554 SPDK_DEBUGLOG(nvmf_tcp, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 2555 offset, length); 2556 2557 /* The NVMe/TCP transport does not use ICDOFF to control the in-capsule data offset. ICDOFF should be '0' */ 2558 if (spdk_unlikely(offset != 0)) { 2559 /* Not defined fatal error in NVMe/TCP spec, handle this error as a fatal error */ 2560 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " should be ZERO in NVMe/TCP\n", offset); 2561 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER; 2562 error_offset = offsetof(struct spdk_nvme_tcp_cmd, ccsqe.dptr.sgl1.address); 2563 goto fatal_err; 2564 } 2565 2566 if (spdk_unlikely(length > max_len)) { 2567 /* According to the SPEC we should support ICD up to 8192 bytes for admin and fabric commands */ 2568 if (length <= SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE && 2569 (cmd->opc == SPDK_NVME_OPC_FABRIC || req->qpair->qid == 0)) { 2570 2571 /* Get a buffer from dedicated list */ 2572 SPDK_DEBUGLOG(nvmf_tcp, "Getting a buffer from control msg list\n"); 2573 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 2574 assert(tgroup->control_msg_list); 2575 req->iov[0].iov_base = nvmf_tcp_control_msg_get(tgroup->control_msg_list); 2576 if (!req->iov[0].iov_base) { 2577 /* No available buffers. Queue this request up. */ 2578 SPDK_DEBUGLOG(nvmf_tcp, "No available ICD buffers. Queueing request %p\n", tcp_req); 2579 return 0; 2580 } 2581 } else { 2582 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 2583 length, max_len); 2584 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_LIMIT_EXCEEDED; 2585 goto fatal_err; 2586 } 2587 } else { 2588 req->iov[0].iov_base = tcp_req->buf; 2589 } 2590 2591 req->length = length; 2592 req->data_from_pool = false; 2593 2594 if (spdk_unlikely(req->dif_enabled)) { 2595 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2596 req->dif.elba_length = length; 2597 } 2598 2599 req->iov[0].iov_len = length; 2600 req->iovcnt = 1; 2601 2602 return 0; 2603 } 2604 /* If we want to handle the problem here, then we can't skip the following data segment. 2605 * Because this function runs before reading data part, now handle all errors as fatal errors. */ 2606 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 2607 sgl->generic.type, sgl->generic.subtype); 2608 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER; 2609 error_offset = offsetof(struct spdk_nvme_tcp_cmd, ccsqe.dptr.sgl1.generic); 2610 fatal_err: 2611 nvmf_tcp_send_c2h_term_req(tcp_req->pdu->qpair, tcp_req->pdu, fes, error_offset); 2612 return -1; 2613 } 2614 2615 static inline enum spdk_nvme_media_error_status_code 2616 nvmf_tcp_dif_error_to_compl_status(uint8_t err_type) { 2617 enum spdk_nvme_media_error_status_code result; 2618 2619 switch (err_type) 2620 { 2621 case SPDK_DIF_REFTAG_ERROR: 2622 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 2623 break; 2624 case SPDK_DIF_APPTAG_ERROR: 2625 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 2626 break; 2627 case SPDK_DIF_GUARD_ERROR: 2628 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 2629 break; 2630 default: 2631 SPDK_UNREACHABLE(); 2632 break; 2633 } 2634 2635 return result; 2636 } 2637 2638 static void 2639 _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2640 struct spdk_nvmf_tcp_req *tcp_req) 2641 { 2642 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF( 2643 tqpair->qpair.transport, struct spdk_nvmf_tcp_transport, transport); 2644 struct nvme_tcp_pdu *rsp_pdu; 2645 struct spdk_nvme_tcp_c2h_data_hdr *c2h_data; 2646 uint32_t plen, pdo, alignment; 2647 int rc; 2648 2649 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2650 2651 rsp_pdu = tcp_req->pdu; 2652 assert(rsp_pdu != NULL); 2653 2654 c2h_data = &rsp_pdu->hdr.c2h_data; 2655 c2h_data->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_DATA; 2656 plen = c2h_data->common.hlen = sizeof(*c2h_data); 2657 2658 if (tqpair->host_hdgst_enable) { 2659 plen += SPDK_NVME_TCP_DIGEST_LEN; 2660 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2661 } 2662 2663 /* set the psh */ 2664 c2h_data->cccid = tcp_req->req.cmd->nvme_cmd.cid; 2665 c2h_data->datal = tcp_req->req.length - tcp_req->pdu->rw_offset; 2666 c2h_data->datao = tcp_req->pdu->rw_offset; 2667 2668 /* set the padding */ 2669 rsp_pdu->padding_len = 0; 2670 pdo = plen; 2671 if (tqpair->cpda) { 2672 alignment = (tqpair->cpda + 1) << 2; 2673 if (plen % alignment != 0) { 2674 pdo = (plen + alignment) / alignment * alignment; 2675 rsp_pdu->padding_len = pdo - plen; 2676 plen = pdo; 2677 } 2678 } 2679 2680 c2h_data->common.pdo = pdo; 2681 plen += c2h_data->datal; 2682 if (tqpair->host_ddgst_enable) { 2683 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_DDGSTF; 2684 plen += SPDK_NVME_TCP_DIGEST_LEN; 2685 } 2686 2687 c2h_data->common.plen = plen; 2688 2689 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2690 rsp_pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 2691 } 2692 2693 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2694 c2h_data->datao, c2h_data->datal); 2695 2696 2697 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU; 2698 /* Need to send the capsule response if response is not all 0 */ 2699 if (ttransport->tcp_opts.c2h_success && 2700 tcp_req->rsp.cdw0 == 0 && tcp_req->rsp.cdw1 == 0) { 2701 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS; 2702 } 2703 2704 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2705 struct spdk_nvme_cpl *rsp = &tcp_req->req.rsp->nvme_cpl; 2706 struct spdk_dif_error err_blk = {}; 2707 uint32_t mapped_length = 0; 2708 uint32_t available_iovs = SPDK_COUNTOF(rsp_pdu->iov); 2709 uint32_t ddgst_len = 0; 2710 2711 if (tqpair->host_ddgst_enable) { 2712 /* Data digest consumes additional iov entry */ 2713 available_iovs--; 2714 /* plen needs to be updated since nvme_tcp_build_iovs compares expected and actual plen */ 2715 ddgst_len = SPDK_NVME_TCP_DIGEST_LEN; 2716 c2h_data->common.plen -= ddgst_len; 2717 } 2718 /* Temp call to estimate if data can be described by limited number of iovs. 2719 * iov vector will be rebuilt in nvmf_tcp_qpair_write_pdu */ 2720 nvme_tcp_build_iovs(rsp_pdu->iov, available_iovs, rsp_pdu, tqpair->host_hdgst_enable, 2721 false, &mapped_length); 2722 2723 if (mapped_length != c2h_data->common.plen) { 2724 c2h_data->datal = mapped_length - (c2h_data->common.plen - c2h_data->datal); 2725 SPDK_DEBUGLOG(nvmf_tcp, 2726 "Part C2H, data_len %u (of %u), PDU len %u, updated PDU len %u, offset %u\n", 2727 c2h_data->datal, tcp_req->req.length, c2h_data->common.plen, mapped_length, rsp_pdu->rw_offset); 2728 c2h_data->common.plen = mapped_length; 2729 2730 /* Rebuild pdu->data_iov since data length is changed */ 2731 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, c2h_data->datao, 2732 c2h_data->datal); 2733 2734 c2h_data->common.flags &= ~(SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU | 2735 SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS); 2736 } 2737 2738 c2h_data->common.plen += ddgst_len; 2739 2740 assert(rsp_pdu->rw_offset <= tcp_req->req.length); 2741 2742 rc = spdk_dif_verify_stream(rsp_pdu->data_iov, rsp_pdu->data_iovcnt, 2743 0, rsp_pdu->data_len, rsp_pdu->dif_ctx, &err_blk); 2744 if (rc != 0) { 2745 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", 2746 err_blk.err_type, err_blk.err_offset); 2747 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2748 rsp->status.sc = nvmf_tcp_dif_error_to_compl_status(err_blk.err_type); 2749 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2750 return; 2751 } 2752 } 2753 2754 rsp_pdu->rw_offset += c2h_data->datal; 2755 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_pdu_c2h_data_complete, tcp_req); 2756 } 2757 2758 static void 2759 nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2760 struct spdk_nvmf_tcp_req *tcp_req) 2761 { 2762 nvmf_tcp_req_pdu_init(tcp_req); 2763 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2764 } 2765 2766 static int 2767 request_transfer_out(struct spdk_nvmf_request *req) 2768 { 2769 struct spdk_nvmf_tcp_req *tcp_req; 2770 struct spdk_nvmf_qpair *qpair; 2771 struct spdk_nvmf_tcp_qpair *tqpair; 2772 struct spdk_nvme_cpl *rsp; 2773 2774 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2775 2776 qpair = req->qpair; 2777 rsp = &req->rsp->nvme_cpl; 2778 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 2779 2780 /* Advance our sq_head pointer */ 2781 if (qpair->sq_head == qpair->sq_head_max) { 2782 qpair->sq_head = 0; 2783 } else { 2784 qpair->sq_head++; 2785 } 2786 rsp->sqhd = qpair->sq_head; 2787 2788 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2789 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2790 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2791 nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2792 } else { 2793 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2794 } 2795 2796 return 0; 2797 } 2798 2799 static void 2800 nvmf_tcp_check_fused_ordering(struct spdk_nvmf_tcp_transport *ttransport, 2801 struct spdk_nvmf_tcp_qpair *tqpair, 2802 struct spdk_nvmf_tcp_req *tcp_req) 2803 { 2804 enum spdk_nvme_cmd_fuse last, next; 2805 2806 last = tqpair->fused_first ? tqpair->fused_first->cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2807 next = tcp_req->cmd.fuse; 2808 2809 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2810 2811 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2812 return; 2813 } 2814 2815 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2816 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2817 /* This is a valid pair of fused commands. Point them at each other 2818 * so they can be submitted consecutively once ready to be executed. 2819 */ 2820 tqpair->fused_first->fused_pair = tcp_req; 2821 tcp_req->fused_pair = tqpair->fused_first; 2822 tqpair->fused_first = NULL; 2823 return; 2824 } else { 2825 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2826 tqpair->fused_first->fused_failed = true; 2827 2828 /* 2829 * If the last req is in READY_TO_EXECUTE state, then call 2830 * nvmf_tcp_req_process(), otherwise nothing else will kick it. 2831 */ 2832 if (tqpair->fused_first->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 2833 nvmf_tcp_req_process(ttransport, tqpair->fused_first); 2834 } 2835 2836 tqpair->fused_first = NULL; 2837 } 2838 } 2839 2840 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2841 /* Set tqpair->fused_first here so that we know to check that the next request 2842 * is a SECOND (and to fail this one if it isn't). 2843 */ 2844 tqpair->fused_first = tcp_req; 2845 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2846 /* Mark this req failed since it is a SECOND and the last one was not a FIRST. */ 2847 tcp_req->fused_failed = true; 2848 } 2849 } 2850 2851 static bool 2852 nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 2853 struct spdk_nvmf_tcp_req *tcp_req) 2854 { 2855 struct spdk_nvmf_tcp_qpair *tqpair; 2856 uint32_t plen; 2857 struct nvme_tcp_pdu *pdu; 2858 enum spdk_nvmf_tcp_req_state prev_state; 2859 bool progress = false; 2860 struct spdk_nvmf_transport *transport = &ttransport->transport; 2861 struct spdk_nvmf_transport_poll_group *group; 2862 struct spdk_nvmf_tcp_poll_group *tgroup; 2863 2864 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2865 group = &tqpair->group->group; 2866 assert(tcp_req->state != TCP_REQUEST_STATE_FREE); 2867 2868 /* If the qpair is not active, we need to abort the outstanding requests. */ 2869 if (tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2870 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 2871 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2872 } 2873 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 2874 } 2875 2876 /* The loop here is to allow for several back-to-back state changes. */ 2877 do { 2878 prev_state = tcp_req->state; 2879 2880 SPDK_DEBUGLOG(nvmf_tcp, "Request %p entering state %d on tqpair=%p\n", tcp_req, prev_state, 2881 tqpair); 2882 2883 switch (tcp_req->state) { 2884 case TCP_REQUEST_STATE_FREE: 2885 /* Some external code must kick a request into TCP_REQUEST_STATE_NEW 2886 * to escape this state. */ 2887 break; 2888 case TCP_REQUEST_STATE_NEW: 2889 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEW, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, tqpair); 2890 2891 /* copy the cmd from the receive pdu */ 2892 tcp_req->cmd = tqpair->pdu_in_progress->hdr.capsule_cmd.ccsqe; 2893 2894 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&tcp_req->req, &tcp_req->req.dif.dif_ctx))) { 2895 tcp_req->req.dif_enabled = true; 2896 tqpair->pdu_in_progress->dif_ctx = &tcp_req->req.dif.dif_ctx; 2897 } 2898 2899 nvmf_tcp_check_fused_ordering(ttransport, tqpair, tcp_req); 2900 2901 /* The next state transition depends on the data transfer needs of this request. */ 2902 tcp_req->req.xfer = spdk_nvmf_req_get_xfer(&tcp_req->req); 2903 2904 if (spdk_unlikely(tcp_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2905 nvmf_tcp_req_set_cpl(tcp_req, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_INVALID_OPCODE); 2906 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2907 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2908 SPDK_DEBUGLOG(nvmf_tcp, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", tcp_req); 2909 break; 2910 } 2911 2912 /* If no data to transfer, ready to execute. */ 2913 if (tcp_req->req.xfer == SPDK_NVME_DATA_NONE) { 2914 /* Reset the tqpair receiving pdu state */ 2915 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2916 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2917 break; 2918 } 2919 2920 pdu = tqpair->pdu_in_progress; 2921 plen = pdu->hdr.common.hlen; 2922 if (tqpair->host_hdgst_enable) { 2923 plen += SPDK_NVME_TCP_DIGEST_LEN; 2924 } 2925 if (pdu->hdr.common.plen != plen) { 2926 tcp_req->has_in_capsule_data = true; 2927 } else { 2928 /* Data is transmitted by C2H PDUs */ 2929 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2930 } 2931 2932 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEED_BUFFER); 2933 STAILQ_INSERT_TAIL(&group->pending_buf_queue, &tcp_req->req, buf_link); 2934 break; 2935 case TCP_REQUEST_STATE_NEED_BUFFER: 2936 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEED_BUFFER, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2937 tqpair); 2938 2939 assert(tcp_req->req.xfer != SPDK_NVME_DATA_NONE); 2940 2941 if (!tcp_req->has_in_capsule_data && (&tcp_req->req != STAILQ_FIRST(&group->pending_buf_queue))) { 2942 SPDK_DEBUGLOG(nvmf_tcp, 2943 "Not the first element to wait for the buf for tcp_req(%p) on tqpair=%p\n", 2944 tcp_req, tqpair); 2945 /* This request needs to wait in line to obtain a buffer */ 2946 break; 2947 } 2948 2949 /* Try to get a data buffer */ 2950 if (nvmf_tcp_req_parse_sgl(tcp_req, transport, group) < 0) { 2951 break; 2952 } 2953 2954 /* Get a zcopy buffer if the request can be serviced through zcopy */ 2955 if (spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 2956 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2957 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 2958 tcp_req->req.length = tcp_req->req.dif.elba_length; 2959 } 2960 2961 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2962 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_START); 2963 spdk_nvmf_request_zcopy_start(&tcp_req->req); 2964 break; 2965 } 2966 2967 if (tcp_req->req.iovcnt < 1) { 2968 SPDK_DEBUGLOG(nvmf_tcp, "No buffer allocated for tcp_req(%p) on tqpair(%p\n)", 2969 tcp_req, tqpair); 2970 /* No buffers available. */ 2971 break; 2972 } 2973 2974 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2975 2976 /* If data is transferring from host to controller, we need to do a transfer from the host. */ 2977 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2978 if (tcp_req->req.data_from_pool) { 2979 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 2980 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 2981 } else { 2982 struct nvme_tcp_pdu *pdu; 2983 2984 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2985 2986 pdu = tqpair->pdu_in_progress; 2987 SPDK_DEBUGLOG(nvmf_tcp, "Not need to send r2t for tcp_req(%p) on tqpair=%p\n", tcp_req, 2988 tqpair); 2989 /* No need to send r2t, contained in the capsuled data */ 2990 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2991 0, tcp_req->req.length); 2992 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2993 } 2994 break; 2995 } 2996 2997 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2998 break; 2999 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3000 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, tqpair->qpair.qid, 0, 3001 (uintptr_t)tcp_req, tqpair); 3002 /* Some external code must kick a request into TCP_REQUEST_STATE_ZCOPY_START_COMPLETED 3003 * to escape this state. */ 3004 break; 3005 case TCP_REQUEST_STATE_ZCOPY_START_COMPLETED: 3006 spdk_trace_record(TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, tqpair->qpair.qid, 0, 3007 (uintptr_t)tcp_req, tqpair); 3008 if (spdk_unlikely(spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl))) { 3009 SPDK_DEBUGLOG(nvmf_tcp, "Zero-copy start failed for tcp_req(%p) on tqpair=%p\n", 3010 tcp_req, tqpair); 3011 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3012 break; 3013 } 3014 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 3015 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 3016 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 3017 } else { 3018 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3019 } 3020 break; 3021 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3022 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3023 tqpair); 3024 /* The R2T completion or the h2c data incoming will kick it out of this state. */ 3025 break; 3026 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3027 3028 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, tqpair->qpair.qid, 0, 3029 (uintptr_t)tcp_req, tqpair); 3030 /* Some external code must kick a request into TCP_REQUEST_STATE_READY_TO_EXECUTE 3031 * to escape this state. */ 3032 break; 3033 case TCP_REQUEST_STATE_READY_TO_EXECUTE: 3034 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, tqpair->qpair.qid, 0, 3035 (uintptr_t)tcp_req, tqpair); 3036 3037 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3038 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 3039 tcp_req->req.length = tcp_req->req.dif.elba_length; 3040 } 3041 3042 if (tcp_req->cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 3043 if (tcp_req->fused_failed) { 3044 /* This request failed FUSED semantics. Fail it immediately, without 3045 * even sending it to the target layer. 3046 */ 3047 nvmf_tcp_req_set_cpl(tcp_req, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_ABORTED_MISSING_FUSED); 3048 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3049 break; 3050 } 3051 3052 if (tcp_req->fused_pair == NULL || 3053 tcp_req->fused_pair->state != TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3054 /* This request is ready to execute, but either we don't know yet if it's 3055 * valid - i.e. this is a FIRST but we haven't received the next request yet), 3056 * or the other request of this fused pair isn't ready to execute. So 3057 * break here and this request will get processed later either when the 3058 * other request is ready or we find that this request isn't valid. 3059 */ 3060 break; 3061 } 3062 } 3063 3064 if (!spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 3065 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTING); 3066 /* If we get to this point, and this request is a fused command, we know that 3067 * it is part of a valid sequence (FIRST followed by a SECOND) and that both 3068 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 3069 * request, and the other request of the fused pair, in the correct order. 3070 * Also clear the ->fused_pair pointers on both requests, since after this point 3071 * we no longer need to maintain the relationship between these two requests. 3072 */ 3073 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 3074 assert(tcp_req->fused_pair != NULL); 3075 assert(tcp_req->fused_pair->fused_pair == tcp_req); 3076 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 3077 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 3078 tcp_req->fused_pair->fused_pair = NULL; 3079 tcp_req->fused_pair = NULL; 3080 } 3081 spdk_nvmf_request_exec(&tcp_req->req); 3082 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 3083 assert(tcp_req->fused_pair != NULL); 3084 assert(tcp_req->fused_pair->fused_pair == tcp_req); 3085 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 3086 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 3087 tcp_req->fused_pair->fused_pair = NULL; 3088 tcp_req->fused_pair = NULL; 3089 } 3090 } else { 3091 /* For zero-copy, only requests with data coming from host to the 3092 * controller can end up here. */ 3093 assert(tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 3094 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT); 3095 spdk_nvmf_request_zcopy_end(&tcp_req->req, true); 3096 } 3097 3098 break; 3099 case TCP_REQUEST_STATE_EXECUTING: 3100 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTING, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3101 tqpair); 3102 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 3103 * to escape this state. */ 3104 break; 3105 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3106 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, tqpair->qpair.qid, 0, 3107 (uintptr_t)tcp_req, tqpair); 3108 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 3109 * to escape this state. */ 3110 break; 3111 case TCP_REQUEST_STATE_EXECUTED: 3112 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTED, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3113 tqpair); 3114 3115 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3116 tcp_req->req.length = tcp_req->req.dif.orig_length; 3117 } 3118 3119 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3120 break; 3121 case TCP_REQUEST_STATE_READY_TO_COMPLETE: 3122 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, tqpair->qpair.qid, 0, 3123 (uintptr_t)tcp_req, tqpair); 3124 if (request_transfer_out(&tcp_req->req) != 0) { 3125 assert(0); /* No good way to handle this currently */ 3126 } 3127 break; 3128 case TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3129 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, tqpair->qpair.qid, 0, 3130 (uintptr_t)tcp_req, tqpair); 3131 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 3132 * to escape this state. */ 3133 break; 3134 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3135 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, tqpair->qpair.qid, 0, 3136 (uintptr_t)tcp_req, tqpair); 3137 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 3138 * to escape this state. */ 3139 break; 3140 case TCP_REQUEST_STATE_COMPLETED: 3141 spdk_trace_record(TRACE_TCP_REQUEST_STATE_COMPLETED, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3142 tqpair); 3143 /* If there's an outstanding PDU sent to the host, the request is completed 3144 * due to the qpair being disconnected. We must delay the completion until 3145 * that write is done to avoid freeing the request twice. */ 3146 if (spdk_unlikely(tcp_req->pdu_in_use)) { 3147 SPDK_DEBUGLOG(nvmf_tcp, "Delaying completion due to outstanding " 3148 "write on req=%p\n", tcp_req); 3149 /* This can only happen for zcopy requests */ 3150 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 3151 assert(tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE); 3152 break; 3153 } 3154 3155 if (tcp_req->req.data_from_pool) { 3156 spdk_nvmf_request_free_buffers(&tcp_req->req, group, transport); 3157 } else if (spdk_unlikely(tcp_req->has_in_capsule_data && 3158 (tcp_req->cmd.opc == SPDK_NVME_OPC_FABRIC || 3159 tqpair->qpair.qid == 0) && tcp_req->req.length > transport->opts.in_capsule_data_size)) { 3160 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3161 assert(tgroup->control_msg_list); 3162 SPDK_DEBUGLOG(nvmf_tcp, "Put buf to control msg list\n"); 3163 nvmf_tcp_control_msg_put(tgroup->control_msg_list, 3164 tcp_req->req.iov[0].iov_base); 3165 } else if (tcp_req->req.zcopy_bdev_io != NULL) { 3166 /* If the request has an unreleased zcopy bdev_io, it's either a 3167 * read, a failed write, or the qpair is being disconnected */ 3168 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 3169 assert(tcp_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST || 3170 spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl) || 3171 tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE); 3172 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE); 3173 spdk_nvmf_request_zcopy_end(&tcp_req->req, false); 3174 break; 3175 } 3176 tcp_req->req.length = 0; 3177 tcp_req->req.iovcnt = 0; 3178 tcp_req->fused_failed = false; 3179 if (tcp_req->fused_pair) { 3180 /* This req was part of a valid fused pair, but failed before it got to 3181 * READ_TO_EXECUTE state. This means we need to fail the other request 3182 * in the pair, because it is no longer part of a valid pair. If the pair 3183 * already reached READY_TO_EXECUTE state, we need to kick it. 3184 */ 3185 tcp_req->fused_pair->fused_failed = true; 3186 if (tcp_req->fused_pair->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3187 nvmf_tcp_req_process(ttransport, tcp_req->fused_pair); 3188 } 3189 tcp_req->fused_pair = NULL; 3190 } 3191 3192 nvmf_tcp_req_put(tqpair, tcp_req); 3193 break; 3194 case TCP_REQUEST_NUM_STATES: 3195 default: 3196 assert(0); 3197 break; 3198 } 3199 3200 if (tcp_req->state != prev_state) { 3201 progress = true; 3202 } 3203 } while (tcp_req->state != prev_state); 3204 3205 return progress; 3206 } 3207 3208 static void 3209 nvmf_tcp_sock_cb(void *arg, struct spdk_sock_group *group, struct spdk_sock *sock) 3210 { 3211 struct spdk_nvmf_tcp_qpair *tqpair = arg; 3212 int rc; 3213 3214 assert(tqpair != NULL); 3215 rc = nvmf_tcp_sock_process(tqpair); 3216 3217 /* If there was a new socket error, disconnect */ 3218 if (rc < 0) { 3219 nvmf_tcp_qpair_disconnect(tqpair); 3220 } 3221 } 3222 3223 static int 3224 nvmf_tcp_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3225 struct spdk_nvmf_qpair *qpair) 3226 { 3227 struct spdk_nvmf_tcp_poll_group *tgroup; 3228 struct spdk_nvmf_tcp_qpair *tqpair; 3229 int rc; 3230 3231 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3232 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3233 3234 rc = nvmf_tcp_qpair_sock_init(tqpair); 3235 if (rc != 0) { 3236 SPDK_ERRLOG("Cannot set sock opt for tqpair=%p\n", tqpair); 3237 return -1; 3238 } 3239 3240 rc = nvmf_tcp_qpair_init(&tqpair->qpair); 3241 if (rc < 0) { 3242 SPDK_ERRLOG("Cannot init tqpair=%p\n", tqpair); 3243 return -1; 3244 } 3245 3246 rc = nvmf_tcp_qpair_init_mem_resource(tqpair); 3247 if (rc < 0) { 3248 SPDK_ERRLOG("Cannot init memory resource info for tqpair=%p\n", tqpair); 3249 return -1; 3250 } 3251 3252 rc = spdk_sock_group_add_sock(tgroup->sock_group, tqpair->sock, 3253 nvmf_tcp_sock_cb, tqpair); 3254 if (rc != 0) { 3255 SPDK_ERRLOG("Could not add sock to sock_group: %s (%d)\n", 3256 spdk_strerror(errno), errno); 3257 return -1; 3258 } 3259 3260 tqpair->group = tgroup; 3261 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INVALID); 3262 TAILQ_INSERT_TAIL(&tgroup->qpairs, tqpair, link); 3263 3264 return 0; 3265 } 3266 3267 static int 3268 nvmf_tcp_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3269 struct spdk_nvmf_qpair *qpair) 3270 { 3271 struct spdk_nvmf_tcp_poll_group *tgroup; 3272 struct spdk_nvmf_tcp_qpair *tqpair; 3273 int rc; 3274 3275 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3276 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3277 3278 assert(tqpair->group == tgroup); 3279 3280 SPDK_DEBUGLOG(nvmf_tcp, "remove tqpair=%p from the tgroup=%p\n", tqpair, tgroup); 3281 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 3282 /* Change the state to move the qpair from the await_req list to the main list 3283 * and prevent adding it again later by nvmf_tcp_qpair_set_recv_state() */ 3284 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 3285 } 3286 TAILQ_REMOVE(&tgroup->qpairs, tqpair, link); 3287 3288 rc = spdk_sock_group_remove_sock(tgroup->sock_group, tqpair->sock); 3289 if (rc != 0) { 3290 SPDK_ERRLOG("Could not remove sock from sock_group: %s (%d)\n", 3291 spdk_strerror(errno), errno); 3292 } 3293 3294 return rc; 3295 } 3296 3297 static int 3298 nvmf_tcp_req_complete(struct spdk_nvmf_request *req) 3299 { 3300 struct spdk_nvmf_tcp_transport *ttransport; 3301 struct spdk_nvmf_tcp_req *tcp_req; 3302 3303 ttransport = SPDK_CONTAINEROF(req->qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3304 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3305 3306 switch (tcp_req->state) { 3307 case TCP_REQUEST_STATE_EXECUTING: 3308 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3309 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3310 break; 3311 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3312 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_ZCOPY_START_COMPLETED); 3313 break; 3314 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3315 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 3316 break; 3317 default: 3318 assert(0 && "Unexpected request state"); 3319 break; 3320 } 3321 3322 nvmf_tcp_req_process(ttransport, tcp_req); 3323 3324 return 0; 3325 } 3326 3327 static void 3328 nvmf_tcp_close_qpair(struct spdk_nvmf_qpair *qpair, 3329 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3330 { 3331 struct spdk_nvmf_tcp_qpair *tqpair; 3332 3333 SPDK_DEBUGLOG(nvmf_tcp, "Qpair: %p\n", qpair); 3334 3335 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3336 3337 assert(tqpair->fini_cb_fn == NULL); 3338 tqpair->fini_cb_fn = cb_fn; 3339 tqpair->fini_cb_arg = cb_arg; 3340 3341 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITED); 3342 nvmf_tcp_qpair_destroy(tqpair); 3343 } 3344 3345 static int 3346 nvmf_tcp_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3347 { 3348 struct spdk_nvmf_tcp_poll_group *tgroup; 3349 int rc; 3350 struct spdk_nvmf_request *req, *req_tmp; 3351 struct spdk_nvmf_tcp_req *tcp_req; 3352 struct spdk_nvmf_tcp_qpair *tqpair, *tqpair_tmp; 3353 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(group->transport, 3354 struct spdk_nvmf_tcp_transport, transport); 3355 3356 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3357 3358 if (spdk_unlikely(TAILQ_EMPTY(&tgroup->qpairs) && TAILQ_EMPTY(&tgroup->await_req))) { 3359 return 0; 3360 } 3361 3362 STAILQ_FOREACH_SAFE(req, &group->pending_buf_queue, buf_link, req_tmp) { 3363 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3364 if (nvmf_tcp_req_process(ttransport, tcp_req) == false) { 3365 break; 3366 } 3367 } 3368 3369 rc = spdk_sock_group_poll(tgroup->sock_group); 3370 if (rc < 0) { 3371 SPDK_ERRLOG("Failed to poll sock_group=%p\n", tgroup->sock_group); 3372 } 3373 3374 TAILQ_FOREACH_SAFE(tqpair, &tgroup->await_req, link, tqpair_tmp) { 3375 rc = nvmf_tcp_sock_process(tqpair); 3376 3377 /* If there was a new socket error, disconnect */ 3378 if (rc < 0) { 3379 nvmf_tcp_qpair_disconnect(tqpair); 3380 } 3381 } 3382 3383 return rc; 3384 } 3385 3386 static int 3387 nvmf_tcp_qpair_get_trid(struct spdk_nvmf_qpair *qpair, 3388 struct spdk_nvme_transport_id *trid, bool peer) 3389 { 3390 struct spdk_nvmf_tcp_qpair *tqpair; 3391 uint16_t port; 3392 3393 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3394 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_TCP); 3395 3396 if (peer) { 3397 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->initiator_addr); 3398 port = tqpair->initiator_port; 3399 } else { 3400 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->target_addr); 3401 port = tqpair->target_port; 3402 } 3403 3404 if (spdk_sock_is_ipv4(tqpair->sock)) { 3405 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3406 } else if (spdk_sock_is_ipv6(tqpair->sock)) { 3407 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3408 } else { 3409 return -1; 3410 } 3411 3412 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%d", port); 3413 return 0; 3414 } 3415 3416 static int 3417 nvmf_tcp_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3418 struct spdk_nvme_transport_id *trid) 3419 { 3420 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3421 } 3422 3423 static int 3424 nvmf_tcp_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3425 struct spdk_nvme_transport_id *trid) 3426 { 3427 return nvmf_tcp_qpair_get_trid(qpair, trid, 1); 3428 } 3429 3430 static int 3431 nvmf_tcp_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3432 struct spdk_nvme_transport_id *trid) 3433 { 3434 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3435 } 3436 3437 static void 3438 nvmf_tcp_req_set_abort_status(struct spdk_nvmf_request *req, 3439 struct spdk_nvmf_tcp_req *tcp_req_to_abort) 3440 { 3441 nvmf_tcp_req_set_cpl(tcp_req_to_abort, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_ABORTED_BY_REQUEST); 3442 nvmf_tcp_req_set_state(tcp_req_to_abort, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3443 3444 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 3445 } 3446 3447 static int 3448 _nvmf_tcp_qpair_abort_request(void *ctx) 3449 { 3450 struct spdk_nvmf_request *req = ctx; 3451 struct spdk_nvmf_tcp_req *tcp_req_to_abort = SPDK_CONTAINEROF(req->req_to_abort, 3452 struct spdk_nvmf_tcp_req, req); 3453 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 3454 struct spdk_nvmf_tcp_qpair, qpair); 3455 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 3456 struct spdk_nvmf_tcp_transport, transport); 3457 int rc; 3458 3459 spdk_poller_unregister(&req->poller); 3460 3461 switch (tcp_req_to_abort->state) { 3462 case TCP_REQUEST_STATE_EXECUTING: 3463 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3464 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3465 rc = nvmf_ctrlr_abort_request(req); 3466 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 3467 return SPDK_POLLER_BUSY; 3468 } 3469 break; 3470 3471 case TCP_REQUEST_STATE_NEED_BUFFER: 3472 STAILQ_REMOVE(&tqpair->group->group.pending_buf_queue, 3473 &tcp_req_to_abort->req, spdk_nvmf_request, buf_link); 3474 3475 nvmf_tcp_req_set_abort_status(req, tcp_req_to_abort); 3476 nvmf_tcp_req_process(ttransport, tcp_req_to_abort); 3477 break; 3478 3479 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3480 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3481 if (spdk_get_ticks() < req->timeout_tsc) { 3482 req->poller = SPDK_POLLER_REGISTER(_nvmf_tcp_qpair_abort_request, req, 0); 3483 return SPDK_POLLER_BUSY; 3484 } 3485 break; 3486 3487 default: 3488 /* Requests in other states are either un-abortable (e.g. 3489 * TRANSFERRING_CONTROLLER_TO_HOST) or should never end up here, as they're 3490 * immediately transitioned to other states in nvmf_tcp_req_process() (e.g. 3491 * READY_TO_EXECUTE). But it is fine to end up here, as we'll simply complete the 3492 * abort request with the bit0 of dword0 set (command not aborted). 3493 */ 3494 break; 3495 } 3496 3497 spdk_nvmf_request_complete(req); 3498 return SPDK_POLLER_BUSY; 3499 } 3500 3501 static void 3502 nvmf_tcp_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 3503 struct spdk_nvmf_request *req) 3504 { 3505 struct spdk_nvmf_tcp_qpair *tqpair; 3506 struct spdk_nvmf_tcp_transport *ttransport; 3507 struct spdk_nvmf_transport *transport; 3508 uint16_t cid; 3509 uint32_t i; 3510 struct spdk_nvmf_tcp_req *tcp_req_to_abort = NULL; 3511 3512 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3513 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3514 transport = &ttransport->transport; 3515 3516 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 3517 3518 for (i = 0; i < tqpair->resource_count; i++) { 3519 if (tqpair->reqs[i].state != TCP_REQUEST_STATE_FREE && 3520 tqpair->reqs[i].req.cmd->nvme_cmd.cid == cid) { 3521 tcp_req_to_abort = &tqpair->reqs[i]; 3522 break; 3523 } 3524 } 3525 3526 spdk_trace_record(TRACE_TCP_QP_ABORT_REQ, qpair->qid, 0, (uintptr_t)req, tqpair); 3527 3528 if (tcp_req_to_abort == NULL) { 3529 spdk_nvmf_request_complete(req); 3530 return; 3531 } 3532 3533 req->req_to_abort = &tcp_req_to_abort->req; 3534 req->timeout_tsc = spdk_get_ticks() + 3535 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 3536 req->poller = NULL; 3537 3538 _nvmf_tcp_qpair_abort_request(req); 3539 } 3540 3541 struct tcp_subsystem_add_host_opts { 3542 char *psk; 3543 }; 3544 3545 static const struct spdk_json_object_decoder tcp_subsystem_add_host_opts_decoder[] = { 3546 {"psk", offsetof(struct tcp_subsystem_add_host_opts, psk), spdk_json_decode_string, true}, 3547 }; 3548 3549 static int 3550 tcp_load_psk(const char *fname, char *buf, size_t bufsz) 3551 { 3552 FILE *psk_file; 3553 struct stat statbuf; 3554 int rc; 3555 3556 if (stat(fname, &statbuf) != 0) { 3557 SPDK_ERRLOG("Could not read permissions for PSK file\n"); 3558 return -EACCES; 3559 } 3560 3561 if ((statbuf.st_mode & TCP_PSK_INVALID_PERMISSIONS) != 0) { 3562 SPDK_ERRLOG("Incorrect permissions for PSK file\n"); 3563 return -EPERM; 3564 } 3565 if ((size_t)statbuf.st_size > bufsz) { 3566 SPDK_ERRLOG("Invalid PSK: too long\n"); 3567 return -EINVAL; 3568 } 3569 psk_file = fopen(fname, "r"); 3570 if (psk_file == NULL) { 3571 SPDK_ERRLOG("Could not open PSK file\n"); 3572 return -EINVAL; 3573 } 3574 3575 rc = fread(buf, 1, statbuf.st_size, psk_file); 3576 if (rc != statbuf.st_size) { 3577 SPDK_ERRLOG("Failed to read PSK\n"); 3578 fclose(psk_file); 3579 return -EINVAL; 3580 } 3581 3582 fclose(psk_file); 3583 return 0; 3584 } 3585 3586 SPDK_LOG_DEPRECATION_REGISTER(nvmf_tcp_psk_path, "PSK path", "v24.09", 0); 3587 3588 static int 3589 nvmf_tcp_subsystem_add_host(struct spdk_nvmf_transport *transport, 3590 const struct spdk_nvmf_subsystem *subsystem, 3591 const char *hostnqn, 3592 const struct spdk_json_val *transport_specific) 3593 { 3594 struct tcp_subsystem_add_host_opts opts; 3595 struct spdk_nvmf_tcp_transport *ttransport; 3596 struct tcp_psk_entry *tmp, *entry = NULL; 3597 uint8_t psk_configured[SPDK_TLS_PSK_MAX_LEN] = {}; 3598 char psk_interchange[SPDK_TLS_PSK_MAX_LEN + 1] = {}; 3599 uint8_t tls_cipher_suite; 3600 int rc = 0; 3601 uint8_t psk_retained_hash; 3602 uint64_t psk_configured_size; 3603 3604 if (transport_specific == NULL) { 3605 return 0; 3606 } 3607 3608 assert(transport != NULL); 3609 assert(subsystem != NULL); 3610 3611 memset(&opts, 0, sizeof(opts)); 3612 3613 /* Decode PSK (either name of a key or file path) */ 3614 if (spdk_json_decode_object_relaxed(transport_specific, tcp_subsystem_add_host_opts_decoder, 3615 SPDK_COUNTOF(tcp_subsystem_add_host_opts_decoder), &opts)) { 3616 SPDK_ERRLOG("spdk_json_decode_object failed\n"); 3617 return -EINVAL; 3618 } 3619 3620 if (opts.psk == NULL) { 3621 return 0; 3622 } 3623 3624 entry = calloc(1, sizeof(struct tcp_psk_entry)); 3625 if (entry == NULL) { 3626 SPDK_ERRLOG("Unable to allocate memory for PSK entry!\n"); 3627 rc = -ENOMEM; 3628 goto end; 3629 } 3630 3631 entry->key = spdk_keyring_get_key(opts.psk); 3632 if (entry->key != NULL) { 3633 rc = spdk_key_get_key(entry->key, psk_interchange, SPDK_TLS_PSK_MAX_LEN); 3634 if (rc < 0) { 3635 SPDK_ERRLOG("Failed to retreive PSK '%s'\n", opts.psk); 3636 rc = -EINVAL; 3637 goto end; 3638 } 3639 } else { 3640 if (strlen(opts.psk) >= sizeof(entry->psk)) { 3641 SPDK_ERRLOG("PSK path too long\n"); 3642 rc = -EINVAL; 3643 goto end; 3644 } 3645 3646 rc = tcp_load_psk(opts.psk, psk_interchange, SPDK_TLS_PSK_MAX_LEN); 3647 if (rc) { 3648 SPDK_ERRLOG("Could not retrieve PSK from file\n"); 3649 goto end; 3650 } 3651 3652 SPDK_LOG_DEPRECATED(nvmf_tcp_psk_path); 3653 } 3654 3655 /* Parse PSK interchange to get length of base64 encoded data. 3656 * This is then used to decide which cipher suite should be used 3657 * to generate PSK identity and TLS PSK later on. */ 3658 rc = nvme_tcp_parse_interchange_psk(psk_interchange, psk_configured, sizeof(psk_configured), 3659 &psk_configured_size, &psk_retained_hash); 3660 if (rc < 0) { 3661 SPDK_ERRLOG("Failed to parse PSK interchange!\n"); 3662 goto end; 3663 } 3664 3665 /* The Base64 string encodes the configured PSK (32 or 48 bytes binary). 3666 * This check also ensures that psk_configured_size is smaller than 3667 * psk_retained buffer size. */ 3668 if (psk_configured_size == SHA256_DIGEST_LENGTH) { 3669 tls_cipher_suite = NVME_TCP_CIPHER_AES_128_GCM_SHA256; 3670 } else if (psk_configured_size == SHA384_DIGEST_LENGTH) { 3671 tls_cipher_suite = NVME_TCP_CIPHER_AES_256_GCM_SHA384; 3672 } else { 3673 SPDK_ERRLOG("Unrecognized cipher suite!\n"); 3674 rc = -EINVAL; 3675 goto end; 3676 } 3677 3678 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3679 /* Generate PSK identity. */ 3680 rc = nvme_tcp_generate_psk_identity(entry->pskid, sizeof(entry->pskid), hostnqn, 3681 subsystem->subnqn, tls_cipher_suite); 3682 if (rc) { 3683 rc = -EINVAL; 3684 goto end; 3685 } 3686 /* Check if PSK identity entry already exists. */ 3687 TAILQ_FOREACH(tmp, &ttransport->psks, link) { 3688 if (strncmp(tmp->pskid, entry->pskid, NVMF_PSK_IDENTITY_LEN) == 0) { 3689 SPDK_ERRLOG("Given PSK identity: %s entry already exists!\n", entry->pskid); 3690 rc = -EEXIST; 3691 goto end; 3692 } 3693 } 3694 3695 if (snprintf(entry->hostnqn, sizeof(entry->hostnqn), "%s", hostnqn) < 0) { 3696 SPDK_ERRLOG("Could not write hostnqn string!\n"); 3697 rc = -EINVAL; 3698 goto end; 3699 } 3700 if (snprintf(entry->subnqn, sizeof(entry->subnqn), "%s", subsystem->subnqn) < 0) { 3701 SPDK_ERRLOG("Could not write subnqn string!\n"); 3702 rc = -EINVAL; 3703 goto end; 3704 } 3705 3706 entry->tls_cipher_suite = tls_cipher_suite; 3707 3708 /* No hash indicates that Configured PSK must be used as Retained PSK. */ 3709 if (psk_retained_hash == NVME_TCP_HASH_ALGORITHM_NONE) { 3710 /* Psk configured is either 32 or 48 bytes long. */ 3711 memcpy(entry->psk, psk_configured, psk_configured_size); 3712 entry->psk_size = psk_configured_size; 3713 } else { 3714 /* Derive retained PSK. */ 3715 rc = nvme_tcp_derive_retained_psk(psk_configured, psk_configured_size, hostnqn, entry->psk, 3716 SPDK_TLS_PSK_MAX_LEN, psk_retained_hash); 3717 if (rc < 0) { 3718 SPDK_ERRLOG("Unable to derive retained PSK!\n"); 3719 goto end; 3720 } 3721 entry->psk_size = rc; 3722 } 3723 3724 if (entry->key == NULL) { 3725 rc = snprintf(entry->psk_path, sizeof(entry->psk_path), "%s", opts.psk); 3726 if (rc < 0 || (size_t)rc >= sizeof(entry->psk_path)) { 3727 SPDK_ERRLOG("Could not save PSK path!\n"); 3728 rc = -ENAMETOOLONG; 3729 goto end; 3730 } 3731 } 3732 3733 TAILQ_INSERT_TAIL(&ttransport->psks, entry, link); 3734 rc = 0; 3735 3736 end: 3737 spdk_memset_s(psk_configured, sizeof(psk_configured), 0, sizeof(psk_configured)); 3738 spdk_memset_s(psk_interchange, sizeof(psk_interchange), 0, sizeof(psk_interchange)); 3739 3740 free(opts.psk); 3741 if (rc != 0) { 3742 nvmf_tcp_free_psk_entry(entry); 3743 } 3744 3745 return rc; 3746 } 3747 3748 static void 3749 nvmf_tcp_subsystem_remove_host(struct spdk_nvmf_transport *transport, 3750 const struct spdk_nvmf_subsystem *subsystem, 3751 const char *hostnqn) 3752 { 3753 struct spdk_nvmf_tcp_transport *ttransport; 3754 struct tcp_psk_entry *entry, *tmp; 3755 3756 assert(transport != NULL); 3757 assert(subsystem != NULL); 3758 3759 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3760 TAILQ_FOREACH_SAFE(entry, &ttransport->psks, link, tmp) { 3761 if ((strncmp(entry->hostnqn, hostnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0 && 3762 (strncmp(entry->subnqn, subsystem->subnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0) { 3763 TAILQ_REMOVE(&ttransport->psks, entry, link); 3764 nvmf_tcp_free_psk_entry(entry); 3765 break; 3766 } 3767 } 3768 } 3769 3770 static void 3771 nvmf_tcp_subsystem_dump_host(struct spdk_nvmf_transport *transport, 3772 const struct spdk_nvmf_subsystem *subsystem, const char *hostnqn, 3773 struct spdk_json_write_ctx *w) 3774 { 3775 struct spdk_nvmf_tcp_transport *ttransport; 3776 struct tcp_psk_entry *entry; 3777 3778 assert(transport != NULL); 3779 assert(subsystem != NULL); 3780 3781 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3782 TAILQ_FOREACH(entry, &ttransport->psks, link) { 3783 if ((strncmp(entry->hostnqn, hostnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0 && 3784 (strncmp(entry->subnqn, subsystem->subnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0) { 3785 spdk_json_write_named_string(w, "psk", entry->key ? 3786 spdk_key_get_name(entry->key) : entry->psk_path); 3787 break; 3788 } 3789 } 3790 } 3791 3792 static void 3793 nvmf_tcp_opts_init(struct spdk_nvmf_transport_opts *opts) 3794 { 3795 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH; 3796 opts->max_qpairs_per_ctrlr = SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR; 3797 opts->in_capsule_data_size = SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE; 3798 opts->max_io_size = SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE; 3799 opts->io_unit_size = SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE; 3800 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH; 3801 opts->num_shared_buffers = SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS; 3802 opts->buf_cache_size = SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE; 3803 opts->dif_insert_or_strip = SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP; 3804 opts->abort_timeout_sec = SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC; 3805 opts->transport_specific = NULL; 3806 } 3807 3808 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp = { 3809 .name = "TCP", 3810 .type = SPDK_NVME_TRANSPORT_TCP, 3811 .opts_init = nvmf_tcp_opts_init, 3812 .create = nvmf_tcp_create, 3813 .dump_opts = nvmf_tcp_dump_opts, 3814 .destroy = nvmf_tcp_destroy, 3815 3816 .listen = nvmf_tcp_listen, 3817 .stop_listen = nvmf_tcp_stop_listen, 3818 3819 .listener_discover = nvmf_tcp_discover, 3820 3821 .poll_group_create = nvmf_tcp_poll_group_create, 3822 .get_optimal_poll_group = nvmf_tcp_get_optimal_poll_group, 3823 .poll_group_destroy = nvmf_tcp_poll_group_destroy, 3824 .poll_group_add = nvmf_tcp_poll_group_add, 3825 .poll_group_remove = nvmf_tcp_poll_group_remove, 3826 .poll_group_poll = nvmf_tcp_poll_group_poll, 3827 3828 .req_free = nvmf_tcp_req_free, 3829 .req_complete = nvmf_tcp_req_complete, 3830 3831 .qpair_fini = nvmf_tcp_close_qpair, 3832 .qpair_get_local_trid = nvmf_tcp_qpair_get_local_trid, 3833 .qpair_get_peer_trid = nvmf_tcp_qpair_get_peer_trid, 3834 .qpair_get_listen_trid = nvmf_tcp_qpair_get_listen_trid, 3835 .qpair_abort_request = nvmf_tcp_qpair_abort_request, 3836 .subsystem_add_host = nvmf_tcp_subsystem_add_host, 3837 .subsystem_remove_host = nvmf_tcp_subsystem_remove_host, 3838 .subsystem_dump_host = nvmf_tcp_subsystem_dump_host, 3839 }; 3840 3841 SPDK_NVMF_TRANSPORT_REGISTER(tcp, &spdk_nvmf_transport_tcp); 3842 SPDK_LOG_REGISTER_COMPONENT(nvmf_tcp) 3843