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) 2022-2024 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 #include "transport.h" 26 27 #include "spdk_internal/trace_defs.h" 28 29 #define NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME 16 30 #define SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY 16 31 #define SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY 0 32 #define SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM 32 33 #define SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION true 34 35 #define SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH 2 36 #define SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH 65535 37 #define SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH 2 38 #define SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH 4096 39 40 #define SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH 128 41 #define SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH 128 42 #define SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 43 #define SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 44 #define SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE 131072 45 #define SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE 131072 46 #define SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS 511 47 #define SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX 48 #define SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP false 49 #define SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC 1 50 51 #define TCP_PSK_INVALID_PERMISSIONS 0177 52 53 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp; 54 static bool g_tls_log = false; 55 56 /* spdk nvmf related structure */ 57 enum spdk_nvmf_tcp_req_state { 58 59 /* The request is not currently in use */ 60 TCP_REQUEST_STATE_FREE = 0, 61 62 /* Initial state when request first received */ 63 TCP_REQUEST_STATE_NEW = 1, 64 65 /* The request is queued until a data buffer is available. */ 66 TCP_REQUEST_STATE_NEED_BUFFER = 2, 67 68 /* The request has the data buffer available */ 69 TCP_REQUEST_STATE_HAVE_BUFFER = 3, 70 71 /* The request is waiting for zcopy_start to finish */ 72 TCP_REQUEST_STATE_AWAITING_ZCOPY_START = 4, 73 74 /* The request has received a zero-copy buffer */ 75 TCP_REQUEST_STATE_ZCOPY_START_COMPLETED = 5, 76 77 /* The request is currently transferring data from the host to the controller. */ 78 TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER = 6, 79 80 /* The request is waiting for the R2T send acknowledgement. */ 81 TCP_REQUEST_STATE_AWAITING_R2T_ACK = 7, 82 83 /* The request is ready to execute at the block device */ 84 TCP_REQUEST_STATE_READY_TO_EXECUTE = 8, 85 86 /* The request is currently executing at the block device */ 87 TCP_REQUEST_STATE_EXECUTING = 9, 88 89 /* The request is waiting for zcopy buffers to be committed */ 90 TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT = 10, 91 92 /* The request finished executing at the block device */ 93 TCP_REQUEST_STATE_EXECUTED = 11, 94 95 /* The request is ready to send a completion */ 96 TCP_REQUEST_STATE_READY_TO_COMPLETE = 12, 97 98 /* The request is currently transferring final pdus from the controller to the host. */ 99 TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST = 13, 100 101 /* The request is waiting for zcopy buffers to be released (without committing) */ 102 TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE = 14, 103 104 /* The request completed and can be marked free. */ 105 TCP_REQUEST_STATE_COMPLETED = 15, 106 107 /* Terminator */ 108 TCP_REQUEST_NUM_STATES, 109 }; 110 111 static const char *spdk_nvmf_tcp_term_req_fes_str[] = { 112 "Invalid PDU Header Field", 113 "PDU Sequence Error", 114 "Header Digiest Error", 115 "Data Transfer Out of Range", 116 "R2T Limit Exceeded", 117 "Unsupported parameter", 118 }; 119 120 SPDK_TRACE_REGISTER_FN(nvmf_tcp_trace, "nvmf_tcp", TRACE_GROUP_NVMF_TCP) 121 { 122 spdk_trace_register_owner_type(OWNER_TYPE_NVMF_TCP, 't'); 123 spdk_trace_register_object(OBJECT_NVMF_TCP_IO, 'r'); 124 spdk_trace_register_description("TCP_REQ_NEW", 125 TRACE_TCP_REQUEST_STATE_NEW, 126 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 1, 127 SPDK_TRACE_ARG_TYPE_INT, "qd"); 128 spdk_trace_register_description("TCP_REQ_NEED_BUFFER", 129 TRACE_TCP_REQUEST_STATE_NEED_BUFFER, 130 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 131 SPDK_TRACE_ARG_TYPE_INT, ""); 132 spdk_trace_register_description("TCP_REQ_HAVE_BUFFER", 133 TRACE_TCP_REQUEST_STATE_HAVE_BUFFER, 134 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 135 SPDK_TRACE_ARG_TYPE_INT, ""); 136 spdk_trace_register_description("TCP_REQ_WAIT_ZCPY_START", 137 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, 138 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 139 SPDK_TRACE_ARG_TYPE_INT, ""); 140 spdk_trace_register_description("TCP_REQ_ZCPY_START_CPL", 141 TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, 142 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 143 SPDK_TRACE_ARG_TYPE_INT, ""); 144 spdk_trace_register_description("TCP_REQ_TX_H_TO_C", 145 TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 146 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 147 SPDK_TRACE_ARG_TYPE_INT, ""); 148 spdk_trace_register_description("TCP_REQ_RDY_TO_EXECUTE", 149 TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, 150 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 151 SPDK_TRACE_ARG_TYPE_INT, ""); 152 spdk_trace_register_description("TCP_REQ_EXECUTING", 153 TRACE_TCP_REQUEST_STATE_EXECUTING, 154 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 155 SPDK_TRACE_ARG_TYPE_INT, ""); 156 spdk_trace_register_description("TCP_REQ_WAIT_ZCPY_CMT", 157 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, 158 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 159 SPDK_TRACE_ARG_TYPE_INT, ""); 160 spdk_trace_register_description("TCP_REQ_EXECUTED", 161 TRACE_TCP_REQUEST_STATE_EXECUTED, 162 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 163 SPDK_TRACE_ARG_TYPE_INT, ""); 164 spdk_trace_register_description("TCP_REQ_RDY_TO_COMPLETE", 165 TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, 166 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 167 SPDK_TRACE_ARG_TYPE_INT, ""); 168 spdk_trace_register_description("TCP_REQ_TRANSFER_C2H", 169 TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 170 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 171 SPDK_TRACE_ARG_TYPE_INT, ""); 172 spdk_trace_register_description("TCP_REQ_AWAIT_ZCPY_RLS", 173 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, 174 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 175 SPDK_TRACE_ARG_TYPE_INT, ""); 176 spdk_trace_register_description("TCP_REQ_COMPLETED", 177 TRACE_TCP_REQUEST_STATE_COMPLETED, 178 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 179 SPDK_TRACE_ARG_TYPE_INT, "qd"); 180 spdk_trace_register_description("TCP_READ_DONE", 181 TRACE_TCP_READ_FROM_SOCKET_DONE, 182 OWNER_TYPE_NVMF_TCP, OBJECT_NONE, 0, 183 SPDK_TRACE_ARG_TYPE_INT, ""); 184 spdk_trace_register_description("TCP_REQ_AWAIT_R2T_ACK", 185 TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, 186 OWNER_TYPE_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 187 SPDK_TRACE_ARG_TYPE_INT, ""); 188 189 spdk_trace_register_description("TCP_QP_CREATE", TRACE_TCP_QP_CREATE, 190 OWNER_TYPE_NVMF_TCP, OBJECT_NONE, 0, 191 SPDK_TRACE_ARG_TYPE_INT, ""); 192 spdk_trace_register_description("TCP_QP_SOCK_INIT", TRACE_TCP_QP_SOCK_INIT, 193 OWNER_TYPE_NVMF_TCP, OBJECT_NONE, 0, 194 SPDK_TRACE_ARG_TYPE_INT, ""); 195 spdk_trace_register_description("TCP_QP_STATE_CHANGE", TRACE_TCP_QP_STATE_CHANGE, 196 OWNER_TYPE_NVMF_TCP, OBJECT_NONE, 0, 197 SPDK_TRACE_ARG_TYPE_INT, "state"); 198 spdk_trace_register_description("TCP_QP_DISCONNECT", TRACE_TCP_QP_DISCONNECT, 199 OWNER_TYPE_NVMF_TCP, OBJECT_NONE, 0, 200 SPDK_TRACE_ARG_TYPE_INT, ""); 201 spdk_trace_register_description("TCP_QP_DESTROY", TRACE_TCP_QP_DESTROY, 202 OWNER_TYPE_NVMF_TCP, OBJECT_NONE, 0, 203 SPDK_TRACE_ARG_TYPE_INT, ""); 204 spdk_trace_register_description("TCP_QP_ABORT_REQ", TRACE_TCP_QP_ABORT_REQ, 205 OWNER_TYPE_NVMF_TCP, OBJECT_NONE, 0, 206 SPDK_TRACE_ARG_TYPE_INT, ""); 207 spdk_trace_register_description("TCP_QP_RCV_STATE_CHANGE", TRACE_TCP_QP_RCV_STATE_CHANGE, 208 OWNER_TYPE_NVMF_TCP, OBJECT_NONE, 0, 209 SPDK_TRACE_ARG_TYPE_INT, "state"); 210 211 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_TCP_IO, 1); 212 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_TCP_IO, 0); 213 spdk_trace_tpoint_register_relation(TRACE_SOCK_REQ_QUEUE, OBJECT_NVMF_TCP_IO, 0); 214 spdk_trace_tpoint_register_relation(TRACE_SOCK_REQ_PEND, OBJECT_NVMF_TCP_IO, 0); 215 spdk_trace_tpoint_register_relation(TRACE_SOCK_REQ_COMPLETE, OBJECT_NVMF_TCP_IO, 0); 216 } 217 218 struct spdk_nvmf_tcp_req { 219 struct spdk_nvmf_request req; 220 struct spdk_nvme_cpl rsp; 221 struct spdk_nvme_cmd cmd; 222 223 /* A PDU that can be used for sending responses. This is 224 * not the incoming PDU! */ 225 struct nvme_tcp_pdu *pdu; 226 227 /* In-capsule data buffer */ 228 uint8_t *buf; 229 230 struct spdk_nvmf_tcp_req *fused_pair; 231 232 /* 233 * The PDU for a request may be used multiple times in serial over 234 * the request's lifetime. For example, first to send an R2T, then 235 * to send a completion. To catch mistakes where the PDU is used 236 * twice at the same time, add a debug flag here for init/fini. 237 */ 238 bool pdu_in_use; 239 bool has_in_capsule_data; 240 bool fused_failed; 241 242 /* transfer_tag */ 243 uint16_t ttag; 244 245 enum spdk_nvmf_tcp_req_state state; 246 247 /* 248 * h2c_offset is used when we receive the h2c_data PDU. 249 */ 250 uint32_t h2c_offset; 251 252 STAILQ_ENTRY(spdk_nvmf_tcp_req) link; 253 TAILQ_ENTRY(spdk_nvmf_tcp_req) state_link; 254 STAILQ_ENTRY(spdk_nvmf_tcp_req) control_msg_link; 255 }; 256 257 struct spdk_nvmf_tcp_qpair { 258 struct spdk_nvmf_qpair qpair; 259 struct spdk_nvmf_tcp_poll_group *group; 260 struct spdk_sock *sock; 261 262 enum nvme_tcp_pdu_recv_state recv_state; 263 enum nvme_tcp_qpair_state state; 264 265 /* PDU being actively received */ 266 struct nvme_tcp_pdu *pdu_in_progress; 267 268 struct spdk_nvmf_tcp_req *fused_first; 269 270 /* Queues to track the requests in all states */ 271 TAILQ_HEAD(, spdk_nvmf_tcp_req) tcp_req_working_queue; 272 TAILQ_HEAD(, spdk_nvmf_tcp_req) tcp_req_free_queue; 273 SLIST_HEAD(, nvme_tcp_pdu) tcp_pdu_free_queue; 274 /* Number of working pdus */ 275 uint32_t tcp_pdu_working_count; 276 277 /* Number of requests in each state */ 278 uint32_t state_cntr[TCP_REQUEST_NUM_STATES]; 279 280 uint8_t cpda; 281 282 bool host_hdgst_enable; 283 bool host_ddgst_enable; 284 285 /* This is a spare PDU used for sending special management 286 * operations. Primarily, this is used for the initial 287 * connection response and c2h termination request. */ 288 struct nvme_tcp_pdu *mgmt_pdu; 289 290 /* Arrays of in-capsule buffers, requests, and pdus. 291 * Each array is 'resource_count' number of elements */ 292 void *bufs; 293 struct spdk_nvmf_tcp_req *reqs; 294 struct nvme_tcp_pdu *pdus; 295 uint32_t resource_count; 296 uint32_t recv_buf_size; 297 298 struct spdk_nvmf_tcp_port *port; 299 300 /* IP address */ 301 char initiator_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 302 char target_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 303 304 /* IP port */ 305 uint16_t initiator_port; 306 uint16_t target_port; 307 308 /* Wait until the host terminates the connection (e.g. after sending C2HTermReq) */ 309 bool wait_terminate; 310 311 /* Timer used to destroy qpair after detecting transport error issue if initiator does 312 * not close the connection. 313 */ 314 struct spdk_poller *timeout_poller; 315 316 spdk_nvmf_transport_qpair_fini_cb fini_cb_fn; 317 void *fini_cb_arg; 318 319 TAILQ_ENTRY(spdk_nvmf_tcp_qpair) link; 320 }; 321 322 struct spdk_nvmf_tcp_control_msg { 323 STAILQ_ENTRY(spdk_nvmf_tcp_control_msg) link; 324 }; 325 326 struct spdk_nvmf_tcp_control_msg_list { 327 void *msg_buf; 328 STAILQ_HEAD(, spdk_nvmf_tcp_control_msg) free_msgs; 329 STAILQ_HEAD(, spdk_nvmf_tcp_req) waiting_for_msg_reqs; 330 }; 331 332 struct spdk_nvmf_tcp_poll_group { 333 struct spdk_nvmf_transport_poll_group group; 334 struct spdk_sock_group *sock_group; 335 336 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) qpairs; 337 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) await_req; 338 339 struct spdk_io_channel *accel_channel; 340 struct spdk_nvmf_tcp_control_msg_list *control_msg_list; 341 342 TAILQ_ENTRY(spdk_nvmf_tcp_poll_group) link; 343 }; 344 345 struct spdk_nvmf_tcp_port { 346 const struct spdk_nvme_transport_id *trid; 347 struct spdk_sock *listen_sock; 348 struct spdk_nvmf_transport *transport; 349 TAILQ_ENTRY(spdk_nvmf_tcp_port) link; 350 }; 351 352 struct tcp_transport_opts { 353 bool c2h_success; 354 uint16_t control_msg_num; 355 uint32_t sock_priority; 356 }; 357 358 struct tcp_psk_entry { 359 char hostnqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 360 char subnqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 361 char pskid[NVMF_PSK_IDENTITY_LEN]; 362 uint8_t psk[SPDK_TLS_PSK_MAX_LEN]; 363 struct spdk_key *key; 364 365 /* Original path saved to emit SPDK configuration via "save_config". */ 366 char psk_path[PATH_MAX]; 367 uint32_t psk_size; 368 enum nvme_tcp_cipher_suite tls_cipher_suite; 369 TAILQ_ENTRY(tcp_psk_entry) link; 370 }; 371 372 struct spdk_nvmf_tcp_transport { 373 struct spdk_nvmf_transport transport; 374 struct tcp_transport_opts tcp_opts; 375 uint32_t ack_timeout; 376 377 struct spdk_nvmf_tcp_poll_group *next_pg; 378 379 struct spdk_poller *accept_poller; 380 struct spdk_sock_group *listen_sock_group; 381 382 TAILQ_HEAD(, spdk_nvmf_tcp_port) ports; 383 TAILQ_HEAD(, spdk_nvmf_tcp_poll_group) poll_groups; 384 385 TAILQ_HEAD(, tcp_psk_entry) psks; 386 }; 387 388 static const struct spdk_json_object_decoder tcp_transport_opts_decoder[] = { 389 { 390 "c2h_success", offsetof(struct tcp_transport_opts, c2h_success), 391 spdk_json_decode_bool, true 392 }, 393 { 394 "control_msg_num", offsetof(struct tcp_transport_opts, control_msg_num), 395 spdk_json_decode_uint16, true 396 }, 397 { 398 "sock_priority", offsetof(struct tcp_transport_opts, sock_priority), 399 spdk_json_decode_uint32, true 400 }, 401 }; 402 403 static bool nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 404 struct spdk_nvmf_tcp_req *tcp_req); 405 static void nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 406 407 static void _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 408 struct spdk_nvmf_tcp_req *tcp_req); 409 410 static inline void 411 nvmf_tcp_req_set_state(struct spdk_nvmf_tcp_req *tcp_req, 412 enum spdk_nvmf_tcp_req_state state) 413 { 414 struct spdk_nvmf_qpair *qpair; 415 struct spdk_nvmf_tcp_qpair *tqpair; 416 417 qpair = tcp_req->req.qpair; 418 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 419 420 assert(tqpair->state_cntr[tcp_req->state] > 0); 421 tqpair->state_cntr[tcp_req->state]--; 422 tqpair->state_cntr[state]++; 423 424 tcp_req->state = state; 425 } 426 427 static inline struct nvme_tcp_pdu * 428 nvmf_tcp_req_pdu_init(struct spdk_nvmf_tcp_req *tcp_req) 429 { 430 assert(tcp_req->pdu_in_use == false); 431 432 memset(tcp_req->pdu, 0, sizeof(*tcp_req->pdu)); 433 tcp_req->pdu->qpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 434 435 return tcp_req->pdu; 436 } 437 438 static struct spdk_nvmf_tcp_req * 439 nvmf_tcp_req_get(struct spdk_nvmf_tcp_qpair *tqpair) 440 { 441 struct spdk_nvmf_tcp_req *tcp_req; 442 443 tcp_req = TAILQ_FIRST(&tqpair->tcp_req_free_queue); 444 if (spdk_unlikely(!tcp_req)) { 445 return NULL; 446 } 447 448 memset(&tcp_req->rsp, 0, sizeof(tcp_req->rsp)); 449 tcp_req->h2c_offset = 0; 450 tcp_req->has_in_capsule_data = false; 451 tcp_req->req.dif_enabled = false; 452 tcp_req->req.zcopy_phase = NVMF_ZCOPY_PHASE_NONE; 453 454 TAILQ_REMOVE(&tqpair->tcp_req_free_queue, tcp_req, state_link); 455 TAILQ_INSERT_TAIL(&tqpair->tcp_req_working_queue, tcp_req, state_link); 456 tqpair->qpair.queue_depth++; 457 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEW); 458 return tcp_req; 459 } 460 461 static inline void 462 nvmf_tcp_req_put(struct spdk_nvmf_tcp_qpair *tqpair, struct spdk_nvmf_tcp_req *tcp_req) 463 { 464 assert(!tcp_req->pdu_in_use); 465 466 TAILQ_REMOVE(&tqpair->tcp_req_working_queue, tcp_req, state_link); 467 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 468 tqpair->qpair.queue_depth--; 469 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_FREE); 470 } 471 472 static void 473 nvmf_tcp_req_get_buffers_done(struct spdk_nvmf_request *req) 474 { 475 struct spdk_nvmf_tcp_req *tcp_req; 476 struct spdk_nvmf_transport *transport; 477 struct spdk_nvmf_tcp_transport *ttransport; 478 479 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 480 transport = req->qpair->transport; 481 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 482 483 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_HAVE_BUFFER); 484 nvmf_tcp_req_process(ttransport, tcp_req); 485 } 486 487 static void 488 nvmf_tcp_request_free(void *cb_arg) 489 { 490 struct spdk_nvmf_tcp_transport *ttransport; 491 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 492 493 assert(tcp_req != NULL); 494 495 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req=%p will be freed\n", tcp_req); 496 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 497 struct spdk_nvmf_tcp_transport, transport); 498 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 499 nvmf_tcp_req_process(ttransport, tcp_req); 500 } 501 502 static int 503 nvmf_tcp_req_free(struct spdk_nvmf_request *req) 504 { 505 struct spdk_nvmf_tcp_req *tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 506 507 nvmf_tcp_request_free(tcp_req); 508 509 return 0; 510 } 511 512 static void 513 nvmf_tcp_drain_state_queue(struct spdk_nvmf_tcp_qpair *tqpair, 514 enum spdk_nvmf_tcp_req_state state) 515 { 516 struct spdk_nvmf_tcp_req *tcp_req, *req_tmp; 517 518 assert(state != TCP_REQUEST_STATE_FREE); 519 TAILQ_FOREACH_SAFE(tcp_req, &tqpair->tcp_req_working_queue, state_link, req_tmp) { 520 if (state == tcp_req->state) { 521 nvmf_tcp_request_free(tcp_req); 522 } 523 } 524 } 525 526 static inline void 527 nvmf_tcp_request_get_buffers_abort(struct spdk_nvmf_tcp_req *tcp_req) 528 { 529 /* Request can wait either for the iobuf or control_msg */ 530 struct spdk_nvmf_poll_group *group = tcp_req->req.qpair->group; 531 struct spdk_nvmf_transport *transport = tcp_req->req.qpair->transport; 532 struct spdk_nvmf_transport_poll_group *tgroup = nvmf_get_transport_poll_group(group, transport); 533 struct spdk_nvmf_tcp_poll_group *tcp_group = SPDK_CONTAINEROF(tgroup, 534 struct spdk_nvmf_tcp_poll_group, group); 535 struct spdk_nvmf_tcp_req *tmp_req, *abort_req; 536 537 assert(tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER); 538 539 STAILQ_FOREACH_SAFE(abort_req, &tcp_group->control_msg_list->waiting_for_msg_reqs, control_msg_link, 540 tmp_req) { 541 if (abort_req == tcp_req) { 542 STAILQ_REMOVE(&tcp_group->control_msg_list->waiting_for_msg_reqs, abort_req, spdk_nvmf_tcp_req, 543 control_msg_link); 544 return; 545 } 546 } 547 548 if (!nvmf_request_get_buffers_abort(&tcp_req->req)) { 549 SPDK_ERRLOG("Failed to abort tcp_req=%p\n", tcp_req); 550 assert(0 && "Should never happen"); 551 } 552 } 553 554 static void 555 nvmf_tcp_cleanup_all_states(struct spdk_nvmf_tcp_qpair *tqpair) 556 { 557 struct spdk_nvmf_tcp_req *tcp_req, *req_tmp; 558 559 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 560 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_NEW); 561 562 /* Wipe the requests waiting for buffer from the waiting list */ 563 TAILQ_FOREACH_SAFE(tcp_req, &tqpair->tcp_req_working_queue, state_link, req_tmp) { 564 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 565 nvmf_tcp_request_get_buffers_abort(tcp_req); 566 } 567 } 568 569 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_NEED_BUFFER); 570 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_EXECUTING); 571 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 572 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 573 } 574 575 static void 576 nvmf_tcp_dump_qpair_req_contents(struct spdk_nvmf_tcp_qpair *tqpair) 577 { 578 int i; 579 struct spdk_nvmf_tcp_req *tcp_req; 580 581 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", tqpair->qpair.qid); 582 for (i = 1; i < TCP_REQUEST_NUM_STATES; i++) { 583 SPDK_ERRLOG("\tNum of requests in state[%d] = %u\n", i, tqpair->state_cntr[i]); 584 TAILQ_FOREACH(tcp_req, &tqpair->tcp_req_working_queue, state_link) { 585 if ((int)tcp_req->state == i) { 586 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", tcp_req->req.data_from_pool); 587 SPDK_ERRLOG("\t\tRequest opcode: %d\n", tcp_req->req.cmd->nvmf_cmd.opcode); 588 } 589 } 590 } 591 } 592 593 static void 594 _nvmf_tcp_qpair_destroy(void *_tqpair) 595 { 596 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 597 spdk_nvmf_transport_qpair_fini_cb cb_fn = tqpair->fini_cb_fn; 598 void *cb_arg = tqpair->fini_cb_arg; 599 int err = 0; 600 601 spdk_trace_record(TRACE_TCP_QP_DESTROY, tqpair->qpair.trace_id, 0, 0); 602 603 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 604 605 err = spdk_sock_close(&tqpair->sock); 606 assert(err == 0); 607 nvmf_tcp_cleanup_all_states(tqpair); 608 609 if (tqpair->state_cntr[TCP_REQUEST_STATE_FREE] != tqpair->resource_count) { 610 SPDK_ERRLOG("tqpair(%p) free tcp request num is %u but should be %u\n", tqpair, 611 tqpair->state_cntr[TCP_REQUEST_STATE_FREE], 612 tqpair->resource_count); 613 err++; 614 } 615 616 if (err > 0) { 617 nvmf_tcp_dump_qpair_req_contents(tqpair); 618 } 619 620 /* The timeout poller might still be registered here if we close the qpair before host 621 * terminates the connection. 622 */ 623 spdk_poller_unregister(&tqpair->timeout_poller); 624 spdk_dma_free(tqpair->pdus); 625 free(tqpair->reqs); 626 spdk_free(tqpair->bufs); 627 spdk_trace_unregister_owner(tqpair->qpair.trace_id); 628 free(tqpair); 629 630 if (cb_fn != NULL) { 631 cb_fn(cb_arg); 632 } 633 634 SPDK_DEBUGLOG(nvmf_tcp, "Leave\n"); 635 } 636 637 static void 638 nvmf_tcp_qpair_destroy(struct spdk_nvmf_tcp_qpair *tqpair) 639 { 640 /* Delay the destruction to make sure it isn't performed from the context of a sock 641 * callback. Otherwise, spdk_sock_close() might not abort pending requests, causing their 642 * completions to be executed after the qpair is freed. (Note: this fixed issue #2471.) 643 */ 644 spdk_thread_send_msg(spdk_get_thread(), _nvmf_tcp_qpair_destroy, tqpair); 645 } 646 647 static void 648 nvmf_tcp_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 649 { 650 struct spdk_nvmf_tcp_transport *ttransport; 651 assert(w != NULL); 652 653 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 654 spdk_json_write_named_bool(w, "c2h_success", ttransport->tcp_opts.c2h_success); 655 spdk_json_write_named_uint32(w, "sock_priority", ttransport->tcp_opts.sock_priority); 656 } 657 658 static void 659 nvmf_tcp_free_psk_entry(struct tcp_psk_entry *entry) 660 { 661 if (entry == NULL) { 662 return; 663 } 664 665 spdk_memset_s(entry->psk, sizeof(entry->psk), 0, sizeof(entry->psk)); 666 spdk_keyring_put_key(entry->key); 667 free(entry); 668 } 669 670 static int 671 nvmf_tcp_destroy(struct spdk_nvmf_transport *transport, 672 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 673 { 674 struct spdk_nvmf_tcp_transport *ttransport; 675 struct tcp_psk_entry *entry, *tmp; 676 677 assert(transport != NULL); 678 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 679 680 TAILQ_FOREACH_SAFE(entry, &ttransport->psks, link, tmp) { 681 TAILQ_REMOVE(&ttransport->psks, entry, link); 682 nvmf_tcp_free_psk_entry(entry); 683 } 684 685 spdk_poller_unregister(&ttransport->accept_poller); 686 spdk_sock_group_unregister_interrupt(ttransport->listen_sock_group); 687 spdk_sock_group_close(&ttransport->listen_sock_group); 688 free(ttransport); 689 690 if (cb_fn) { 691 cb_fn(cb_arg); 692 } 693 return 0; 694 } 695 696 static int nvmf_tcp_accept(void *ctx); 697 698 static void nvmf_tcp_accept_cb(void *ctx, struct spdk_sock_group *group, struct spdk_sock *sock); 699 700 static struct spdk_nvmf_transport * 701 nvmf_tcp_create(struct spdk_nvmf_transport_opts *opts) 702 { 703 struct spdk_nvmf_tcp_transport *ttransport; 704 uint32_t sge_count; 705 uint32_t min_shared_buffers; 706 int rc; 707 uint64_t period; 708 709 ttransport = calloc(1, sizeof(*ttransport)); 710 if (!ttransport) { 711 return NULL; 712 } 713 714 TAILQ_INIT(&ttransport->ports); 715 TAILQ_INIT(&ttransport->poll_groups); 716 TAILQ_INIT(&ttransport->psks); 717 718 ttransport->transport.ops = &spdk_nvmf_transport_tcp; 719 720 ttransport->tcp_opts.c2h_success = SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION; 721 ttransport->tcp_opts.sock_priority = SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY; 722 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 723 if (opts->transport_specific != NULL && 724 spdk_json_decode_object_relaxed(opts->transport_specific, tcp_transport_opts_decoder, 725 SPDK_COUNTOF(tcp_transport_opts_decoder), 726 &ttransport->tcp_opts)) { 727 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 728 free(ttransport); 729 return NULL; 730 } 731 732 SPDK_NOTICELOG("*** TCP Transport Init ***\n"); 733 734 SPDK_INFOLOG(nvmf_tcp, "*** TCP Transport Init ***\n" 735 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 736 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 737 " in_capsule_data_size=%d, max_aq_depth=%d\n" 738 " num_shared_buffers=%d, c2h_success=%d,\n" 739 " dif_insert_or_strip=%d, sock_priority=%d\n" 740 " abort_timeout_sec=%d, control_msg_num=%hu\n" 741 " ack_timeout=%d\n", 742 opts->max_queue_depth, 743 opts->max_io_size, 744 opts->max_qpairs_per_ctrlr - 1, 745 opts->io_unit_size, 746 opts->in_capsule_data_size, 747 opts->max_aq_depth, 748 opts->num_shared_buffers, 749 ttransport->tcp_opts.c2h_success, 750 opts->dif_insert_or_strip, 751 ttransport->tcp_opts.sock_priority, 752 opts->abort_timeout_sec, 753 ttransport->tcp_opts.control_msg_num, 754 opts->ack_timeout); 755 756 if (ttransport->tcp_opts.sock_priority > SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY) { 757 SPDK_ERRLOG("Unsupported socket_priority=%d, the current range is: 0 to %d\n" 758 "you can use man 7 socket to view the range of priority under SO_PRIORITY item\n", 759 ttransport->tcp_opts.sock_priority, SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY); 760 free(ttransport); 761 return NULL; 762 } 763 764 if (ttransport->tcp_opts.control_msg_num == 0 && 765 opts->in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 766 SPDK_WARNLOG("TCP param control_msg_num can't be 0 if ICD is less than %u bytes. Using default value %u\n", 767 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM); 768 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 769 } 770 771 /* I/O unit size cannot be larger than max I/O size */ 772 if (opts->io_unit_size > opts->max_io_size) { 773 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", 774 opts->io_unit_size, opts->max_io_size); 775 opts->io_unit_size = opts->max_io_size; 776 } 777 778 /* In capsule data size cannot be larger than max I/O size */ 779 if (opts->in_capsule_data_size > opts->max_io_size) { 780 SPDK_WARNLOG("TCP param ICD size %u can't be larger than max_io_size %u. Using max_io_size as ICD size\n", 781 opts->io_unit_size, opts->max_io_size); 782 opts->in_capsule_data_size = opts->max_io_size; 783 } 784 785 /* max IO queue depth cannot be smaller than 2 or larger than 65535. 786 * We will not check SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH, because max_queue_depth is 16bits and always not larger than 64k. */ 787 if (opts->max_queue_depth < SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH) { 788 SPDK_WARNLOG("TCP param max_queue_depth %u can't be smaller than %u or larger than %u. Using default value %u\n", 789 opts->max_queue_depth, SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH, 790 SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH, SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH); 791 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH; 792 } 793 794 /* max admin queue depth cannot be smaller than 2 or larger than 4096 */ 795 if (opts->max_aq_depth < SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH || 796 opts->max_aq_depth > SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH) { 797 SPDK_WARNLOG("TCP param max_aq_depth %u can't be smaller than %u or larger than %u. Using default value %u\n", 798 opts->max_aq_depth, SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH, 799 SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH, SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH); 800 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH; 801 } 802 803 sge_count = opts->max_io_size / opts->io_unit_size; 804 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 805 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 806 free(ttransport); 807 return NULL; 808 } 809 810 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 811 if (opts->buf_cache_size < UINT32_MAX) { 812 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 813 if (min_shared_buffers > opts->num_shared_buffers) { 814 SPDK_ERRLOG("There are not enough buffers to satisfy " 815 "per-poll group caches for each thread. (%" PRIu32 ") " 816 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 817 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 818 free(ttransport); 819 return NULL; 820 } 821 } 822 823 period = spdk_interrupt_mode_is_enabled() ? 0 : opts->acceptor_poll_rate; 824 ttransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_tcp_accept, &ttransport->transport, period); 825 if (!ttransport->accept_poller) { 826 free(ttransport); 827 return NULL; 828 } 829 830 spdk_poller_register_interrupt(ttransport->accept_poller, NULL, NULL); 831 832 ttransport->listen_sock_group = spdk_sock_group_create(NULL); 833 if (ttransport->listen_sock_group == NULL) { 834 SPDK_ERRLOG("Failed to create socket group for listen sockets\n"); 835 spdk_poller_unregister(&ttransport->accept_poller); 836 free(ttransport); 837 return NULL; 838 } 839 840 if (spdk_interrupt_mode_is_enabled()) { 841 rc = SPDK_SOCK_GROUP_REGISTER_INTERRUPT(ttransport->listen_sock_group, 842 SPDK_INTERRUPT_EVENT_IN | SPDK_INTERRUPT_EVENT_OUT, nvmf_tcp_accept, &ttransport->transport); 843 if (rc != 0) { 844 SPDK_ERRLOG("Failed to register interrupt for listen socker sock group\n"); 845 spdk_sock_group_close(&ttransport->listen_sock_group); 846 spdk_poller_unregister(&ttransport->accept_poller); 847 free(ttransport); 848 return NULL; 849 } 850 } 851 852 return &ttransport->transport; 853 } 854 855 static int 856 nvmf_tcp_trsvcid_to_int(const char *trsvcid) 857 { 858 unsigned long long ull; 859 char *end = NULL; 860 861 ull = strtoull(trsvcid, &end, 10); 862 if (end == NULL || end == trsvcid || *end != '\0') { 863 return -1; 864 } 865 866 /* Valid TCP/IP port numbers are in [1, 65535] */ 867 if (ull == 0 || ull > 65535) { 868 return -1; 869 } 870 871 return (int)ull; 872 } 873 874 /** 875 * Canonicalize a listen address trid. 876 */ 877 static int 878 nvmf_tcp_canon_listen_trid(struct spdk_nvme_transport_id *canon_trid, 879 const struct spdk_nvme_transport_id *trid) 880 { 881 int trsvcid_int; 882 883 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 884 if (trsvcid_int < 0) { 885 return -EINVAL; 886 } 887 888 memset(canon_trid, 0, sizeof(*canon_trid)); 889 spdk_nvme_trid_populate_transport(canon_trid, SPDK_NVME_TRANSPORT_TCP); 890 canon_trid->adrfam = trid->adrfam; 891 snprintf(canon_trid->traddr, sizeof(canon_trid->traddr), "%s", trid->traddr); 892 snprintf(canon_trid->trsvcid, sizeof(canon_trid->trsvcid), "%d", trsvcid_int); 893 894 return 0; 895 } 896 897 /** 898 * Find an existing listening port. 899 */ 900 static struct spdk_nvmf_tcp_port * 901 nvmf_tcp_find_port(struct spdk_nvmf_tcp_transport *ttransport, 902 const struct spdk_nvme_transport_id *trid) 903 { 904 struct spdk_nvme_transport_id canon_trid; 905 struct spdk_nvmf_tcp_port *port; 906 907 if (nvmf_tcp_canon_listen_trid(&canon_trid, trid) != 0) { 908 return NULL; 909 } 910 911 TAILQ_FOREACH(port, &ttransport->ports, link) { 912 if (spdk_nvme_transport_id_compare(&canon_trid, port->trid) == 0) { 913 return port; 914 } 915 } 916 917 return NULL; 918 } 919 920 static int 921 tcp_sock_get_key(uint8_t *out, int out_len, const char **cipher, const char *pskid, 922 void *get_key_ctx) 923 { 924 struct tcp_psk_entry *entry; 925 struct spdk_nvmf_tcp_transport *ttransport = get_key_ctx; 926 size_t psk_len; 927 int rc; 928 929 TAILQ_FOREACH(entry, &ttransport->psks, link) { 930 if (strcmp(pskid, entry->pskid) != 0) { 931 continue; 932 } 933 934 psk_len = entry->psk_size; 935 if ((size_t)out_len < psk_len) { 936 SPDK_ERRLOG("Out buffer of size: %" PRIu32 " cannot fit PSK of len: %lu\n", 937 out_len, psk_len); 938 return -ENOBUFS; 939 } 940 941 /* Convert PSK to the TLS PSK format. */ 942 rc = nvme_tcp_derive_tls_psk(entry->psk, psk_len, pskid, out, out_len, 943 entry->tls_cipher_suite); 944 if (rc < 0) { 945 SPDK_ERRLOG("Could not generate TLS PSK\n"); 946 } 947 948 switch (entry->tls_cipher_suite) { 949 case NVME_TCP_CIPHER_AES_128_GCM_SHA256: 950 *cipher = "TLS_AES_128_GCM_SHA256"; 951 break; 952 case NVME_TCP_CIPHER_AES_256_GCM_SHA384: 953 *cipher = "TLS_AES_256_GCM_SHA384"; 954 break; 955 default: 956 *cipher = NULL; 957 return -ENOTSUP; 958 } 959 960 return rc; 961 } 962 963 SPDK_ERRLOG("Could not find PSK for identity: %s\n", pskid); 964 965 return -EINVAL; 966 } 967 968 static int 969 nvmf_tcp_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 970 struct spdk_nvmf_listen_opts *listen_opts) 971 { 972 struct spdk_nvmf_tcp_transport *ttransport; 973 struct spdk_nvmf_tcp_port *port; 974 int trsvcid_int; 975 uint8_t adrfam; 976 const char *sock_impl_name; 977 struct spdk_sock_impl_opts impl_opts; 978 size_t impl_opts_size = sizeof(impl_opts); 979 struct spdk_sock_opts opts; 980 int rc; 981 982 if (!strlen(trid->trsvcid)) { 983 SPDK_ERRLOG("Service id is required\n"); 984 return -EINVAL; 985 } 986 987 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 988 989 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 990 if (trsvcid_int < 0) { 991 SPDK_ERRLOG("Invalid trsvcid '%s'\n", trid->trsvcid); 992 return -EINVAL; 993 } 994 995 port = calloc(1, sizeof(*port)); 996 if (!port) { 997 SPDK_ERRLOG("Port allocation failed\n"); 998 return -ENOMEM; 999 } 1000 1001 port->trid = trid; 1002 1003 sock_impl_name = NULL; 1004 1005 opts.opts_size = sizeof(opts); 1006 spdk_sock_get_default_opts(&opts); 1007 opts.priority = ttransport->tcp_opts.sock_priority; 1008 opts.ack_timeout = transport->opts.ack_timeout; 1009 if (listen_opts->secure_channel) { 1010 if (listen_opts->sock_impl && 1011 strncmp("ssl", listen_opts->sock_impl, strlen(listen_opts->sock_impl))) { 1012 SPDK_ERRLOG("Enabling secure_channel while specifying a sock_impl different from 'ssl' is unsupported"); 1013 free(port); 1014 return -EINVAL; 1015 } 1016 listen_opts->sock_impl = "ssl"; 1017 } 1018 1019 if (listen_opts->sock_impl) { 1020 sock_impl_name = listen_opts->sock_impl; 1021 spdk_sock_impl_get_opts(sock_impl_name, &impl_opts, &impl_opts_size); 1022 1023 if (!strncmp("ssl", sock_impl_name, strlen(sock_impl_name))) { 1024 if (!g_tls_log) { 1025 SPDK_NOTICELOG("TLS support is considered experimental\n"); 1026 g_tls_log = true; 1027 } 1028 impl_opts.tls_version = SPDK_TLS_VERSION_1_3; 1029 impl_opts.get_key = tcp_sock_get_key; 1030 impl_opts.get_key_ctx = ttransport; 1031 impl_opts.tls_cipher_suites = "TLS_AES_256_GCM_SHA384:TLS_AES_128_GCM_SHA256"; 1032 } 1033 1034 opts.impl_opts = &impl_opts; 1035 opts.impl_opts_size = sizeof(impl_opts); 1036 } 1037 1038 port->listen_sock = spdk_sock_listen_ext(trid->traddr, trsvcid_int, 1039 sock_impl_name, &opts); 1040 if (port->listen_sock == NULL) { 1041 SPDK_ERRLOG("spdk_sock_listen(%s, %d) failed: %s (%d)\n", 1042 trid->traddr, trsvcid_int, 1043 spdk_strerror(errno), errno); 1044 free(port); 1045 return -errno; 1046 } 1047 1048 if (spdk_sock_is_ipv4(port->listen_sock)) { 1049 adrfam = SPDK_NVMF_ADRFAM_IPV4; 1050 } else if (spdk_sock_is_ipv6(port->listen_sock)) { 1051 adrfam = SPDK_NVMF_ADRFAM_IPV6; 1052 } else { 1053 SPDK_ERRLOG("Unhandled socket type\n"); 1054 adrfam = 0; 1055 } 1056 1057 if (adrfam != trid->adrfam) { 1058 SPDK_ERRLOG("Socket address family mismatch\n"); 1059 spdk_sock_close(&port->listen_sock); 1060 free(port); 1061 return -EINVAL; 1062 } 1063 1064 rc = spdk_sock_group_add_sock(ttransport->listen_sock_group, port->listen_sock, nvmf_tcp_accept_cb, 1065 port); 1066 if (rc < 0) { 1067 SPDK_ERRLOG("Failed to add socket to the listen socket group\n"); 1068 spdk_sock_close(&port->listen_sock); 1069 free(port); 1070 return -errno; 1071 } 1072 1073 port->transport = transport; 1074 1075 SPDK_NOTICELOG("*** NVMe/TCP Target Listening on %s port %s ***\n", 1076 trid->traddr, trid->trsvcid); 1077 1078 TAILQ_INSERT_TAIL(&ttransport->ports, port, link); 1079 return 0; 1080 } 1081 1082 static void 1083 nvmf_tcp_stop_listen(struct spdk_nvmf_transport *transport, 1084 const struct spdk_nvme_transport_id *trid) 1085 { 1086 struct spdk_nvmf_tcp_transport *ttransport; 1087 struct spdk_nvmf_tcp_port *port; 1088 1089 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1090 1091 SPDK_DEBUGLOG(nvmf_tcp, "Removing listen address %s port %s\n", 1092 trid->traddr, trid->trsvcid); 1093 1094 port = nvmf_tcp_find_port(ttransport, trid); 1095 if (port) { 1096 spdk_sock_group_remove_sock(ttransport->listen_sock_group, port->listen_sock); 1097 TAILQ_REMOVE(&ttransport->ports, port, link); 1098 spdk_sock_close(&port->listen_sock); 1099 free(port); 1100 } 1101 } 1102 1103 static void nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 1104 enum nvme_tcp_pdu_recv_state state); 1105 1106 static void 1107 nvmf_tcp_qpair_set_state(struct spdk_nvmf_tcp_qpair *tqpair, enum nvme_tcp_qpair_state state) 1108 { 1109 tqpair->state = state; 1110 spdk_trace_record(TRACE_TCP_QP_STATE_CHANGE, tqpair->qpair.trace_id, 0, 0, 1111 (uint64_t)tqpair->state); 1112 } 1113 1114 static void 1115 nvmf_tcp_qpair_disconnect(struct spdk_nvmf_tcp_qpair *tqpair) 1116 { 1117 SPDK_DEBUGLOG(nvmf_tcp, "Disconnecting qpair %p\n", tqpair); 1118 1119 spdk_trace_record(TRACE_TCP_QP_DISCONNECT, tqpair->qpair.trace_id, 0, 0); 1120 1121 if (tqpair->state <= NVME_TCP_QPAIR_STATE_RUNNING) { 1122 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITING); 1123 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1124 spdk_poller_unregister(&tqpair->timeout_poller); 1125 1126 /* This will end up calling nvmf_tcp_close_qpair */ 1127 spdk_nvmf_qpair_disconnect(&tqpair->qpair); 1128 } 1129 } 1130 1131 static void 1132 _mgmt_pdu_write_done(void *_tqpair, int err) 1133 { 1134 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 1135 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1136 1137 if (spdk_unlikely(err != 0)) { 1138 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1139 return; 1140 } 1141 1142 assert(pdu->cb_fn != NULL); 1143 pdu->cb_fn(pdu->cb_arg); 1144 } 1145 1146 static void 1147 _req_pdu_write_done(void *req, int err) 1148 { 1149 struct spdk_nvmf_tcp_req *tcp_req = req; 1150 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1151 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1152 1153 assert(tcp_req->pdu_in_use); 1154 tcp_req->pdu_in_use = false; 1155 1156 /* If the request is in a completed state, we're waiting for write completion to free it */ 1157 if (spdk_unlikely(tcp_req->state == TCP_REQUEST_STATE_COMPLETED)) { 1158 nvmf_tcp_request_free(tcp_req); 1159 return; 1160 } 1161 1162 if (spdk_unlikely(err != 0)) { 1163 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1164 return; 1165 } 1166 1167 assert(pdu->cb_fn != NULL); 1168 pdu->cb_fn(pdu->cb_arg); 1169 } 1170 1171 static void 1172 _pdu_write_done(struct nvme_tcp_pdu *pdu, int err) 1173 { 1174 pdu->sock_req.cb_fn(pdu->sock_req.cb_arg, err); 1175 } 1176 1177 static void 1178 _tcp_write_pdu(struct nvme_tcp_pdu *pdu) 1179 { 1180 int rc; 1181 uint32_t mapped_length; 1182 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1183 1184 pdu->sock_req.iovcnt = nvme_tcp_build_iovs(pdu->iov, SPDK_COUNTOF(pdu->iov), pdu, 1185 tqpair->host_hdgst_enable, tqpair->host_ddgst_enable, &mapped_length); 1186 spdk_sock_writev_async(tqpair->sock, &pdu->sock_req); 1187 1188 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP || 1189 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ) { 1190 /* Try to force the send immediately. */ 1191 rc = spdk_sock_flush(tqpair->sock); 1192 if (rc > 0 && (uint32_t)rc == mapped_length) { 1193 _pdu_write_done(pdu, 0); 1194 } else { 1195 SPDK_ERRLOG("Could not write %s to socket: rc=%d, errno=%d\n", 1196 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP ? 1197 "IC_RESP" : "TERM_REQ", rc, errno); 1198 _pdu_write_done(pdu, rc >= 0 ? -EAGAIN : -errno); 1199 } 1200 } 1201 } 1202 1203 static void 1204 data_crc32_accel_done(void *cb_arg, int status) 1205 { 1206 struct nvme_tcp_pdu *pdu = cb_arg; 1207 1208 if (spdk_unlikely(status)) { 1209 SPDK_ERRLOG("Failed to compute the data digest for pdu =%p\n", pdu); 1210 _pdu_write_done(pdu, status); 1211 return; 1212 } 1213 1214 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 1215 MAKE_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32); 1216 1217 _tcp_write_pdu(pdu); 1218 } 1219 1220 static void 1221 pdu_data_crc32_compute(struct nvme_tcp_pdu *pdu) 1222 { 1223 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1224 int rc = 0; 1225 1226 /* Data Digest */ 1227 if (pdu->data_len > 0 && g_nvme_tcp_ddgst[pdu->hdr.common.pdu_type] && tqpair->host_ddgst_enable) { 1228 /* Only support this limitated case for the first step */ 1229 if (spdk_likely(!pdu->dif_ctx && (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0) 1230 && tqpair->group)) { 1231 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 1232 pdu->data_iovcnt, 0, data_crc32_accel_done, pdu); 1233 if (spdk_likely(rc == 0)) { 1234 return; 1235 } 1236 } else { 1237 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 1238 } 1239 data_crc32_accel_done(pdu, rc); 1240 } else { 1241 _tcp_write_pdu(pdu); 1242 } 1243 } 1244 1245 static void 1246 nvmf_tcp_qpair_write_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1247 struct nvme_tcp_pdu *pdu, 1248 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1249 void *cb_arg) 1250 { 1251 int hlen; 1252 uint32_t crc32c; 1253 1254 assert(tqpair->pdu_in_progress != pdu); 1255 1256 hlen = pdu->hdr.common.hlen; 1257 pdu->cb_fn = cb_fn; 1258 pdu->cb_arg = cb_arg; 1259 1260 pdu->iov[0].iov_base = &pdu->hdr.raw; 1261 pdu->iov[0].iov_len = hlen; 1262 1263 /* Header Digest */ 1264 if (g_nvme_tcp_hdgst[pdu->hdr.common.pdu_type] && tqpair->host_hdgst_enable) { 1265 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 1266 MAKE_DIGEST_WORD((uint8_t *)pdu->hdr.raw + hlen, crc32c); 1267 } 1268 1269 /* Data Digest */ 1270 pdu_data_crc32_compute(pdu); 1271 } 1272 1273 static void 1274 nvmf_tcp_qpair_write_mgmt_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1275 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1276 void *cb_arg) 1277 { 1278 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1279 1280 pdu->sock_req.cb_fn = _mgmt_pdu_write_done; 1281 pdu->sock_req.cb_arg = tqpair; 1282 1283 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1284 } 1285 1286 static void 1287 nvmf_tcp_qpair_write_req_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1288 struct spdk_nvmf_tcp_req *tcp_req, 1289 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1290 void *cb_arg) 1291 { 1292 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1293 1294 pdu->sock_req.cb_fn = _req_pdu_write_done; 1295 pdu->sock_req.cb_arg = tcp_req; 1296 1297 assert(!tcp_req->pdu_in_use); 1298 tcp_req->pdu_in_use = true; 1299 1300 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1301 } 1302 1303 static int 1304 nvmf_tcp_qpair_init_mem_resource(struct spdk_nvmf_tcp_qpair *tqpair) 1305 { 1306 uint32_t i; 1307 struct spdk_nvmf_transport_opts *opts; 1308 uint32_t in_capsule_data_size; 1309 1310 opts = &tqpair->qpair.transport->opts; 1311 1312 in_capsule_data_size = opts->in_capsule_data_size; 1313 if (opts->dif_insert_or_strip) { 1314 in_capsule_data_size = SPDK_BDEV_BUF_SIZE_WITH_MD(in_capsule_data_size); 1315 } 1316 1317 tqpair->resource_count = opts->max_queue_depth; 1318 1319 tqpair->reqs = calloc(tqpair->resource_count, sizeof(*tqpair->reqs)); 1320 if (!tqpair->reqs) { 1321 SPDK_ERRLOG("Unable to allocate reqs on tqpair=%p\n", tqpair); 1322 return -1; 1323 } 1324 1325 if (in_capsule_data_size) { 1326 tqpair->bufs = spdk_zmalloc(tqpair->resource_count * in_capsule_data_size, 0x1000, 1327 NULL, SPDK_ENV_LCORE_ID_ANY, 1328 SPDK_MALLOC_DMA); 1329 if (!tqpair->bufs) { 1330 SPDK_ERRLOG("Unable to allocate bufs on tqpair=%p.\n", tqpair); 1331 return -1; 1332 } 1333 } 1334 /* prepare memory space for receiving pdus and tcp_req */ 1335 /* Add additional 1 member, which will be used for mgmt_pdu owned by the tqpair */ 1336 tqpair->pdus = spdk_dma_zmalloc((2 * tqpair->resource_count + 1) * sizeof(*tqpair->pdus), 0x1000, 1337 NULL); 1338 if (!tqpair->pdus) { 1339 SPDK_ERRLOG("Unable to allocate pdu pool on tqpair =%p.\n", tqpair); 1340 return -1; 1341 } 1342 1343 for (i = 0; i < tqpair->resource_count; i++) { 1344 struct spdk_nvmf_tcp_req *tcp_req = &tqpair->reqs[i]; 1345 1346 tcp_req->ttag = i + 1; 1347 tcp_req->req.qpair = &tqpair->qpair; 1348 1349 tcp_req->pdu = &tqpair->pdus[i]; 1350 tcp_req->pdu->qpair = tqpair; 1351 1352 /* Set up memory to receive commands */ 1353 if (tqpair->bufs) { 1354 tcp_req->buf = (void *)((uintptr_t)tqpair->bufs + (i * in_capsule_data_size)); 1355 } 1356 1357 /* Set the cmdn and rsp */ 1358 tcp_req->req.rsp = (union nvmf_c2h_msg *)&tcp_req->rsp; 1359 tcp_req->req.cmd = (union nvmf_h2c_msg *)&tcp_req->cmd; 1360 1361 tcp_req->req.stripped_data = NULL; 1362 1363 /* Initialize request state to FREE */ 1364 tcp_req->state = TCP_REQUEST_STATE_FREE; 1365 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 1366 tqpair->state_cntr[TCP_REQUEST_STATE_FREE]++; 1367 } 1368 1369 for (; i < 2 * tqpair->resource_count; i++) { 1370 struct nvme_tcp_pdu *pdu = &tqpair->pdus[i]; 1371 1372 pdu->qpair = tqpair; 1373 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 1374 } 1375 1376 tqpair->mgmt_pdu = &tqpair->pdus[i]; 1377 tqpair->mgmt_pdu->qpair = tqpair; 1378 tqpair->pdu_in_progress = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 1379 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 1380 tqpair->tcp_pdu_working_count = 1; 1381 1382 tqpair->recv_buf_size = (in_capsule_data_size + sizeof(struct spdk_nvme_tcp_cmd) + 2 * 1383 SPDK_NVME_TCP_DIGEST_LEN) * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1384 1385 return 0; 1386 } 1387 1388 static int 1389 nvmf_tcp_qpair_init(struct spdk_nvmf_qpair *qpair) 1390 { 1391 struct spdk_nvmf_tcp_qpair *tqpair; 1392 1393 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1394 1395 SPDK_DEBUGLOG(nvmf_tcp, "New TCP Connection: %p\n", qpair); 1396 1397 spdk_trace_record(TRACE_TCP_QP_CREATE, tqpair->qpair.trace_id, 0, 0); 1398 1399 /* Initialise request state queues of the qpair */ 1400 TAILQ_INIT(&tqpair->tcp_req_free_queue); 1401 TAILQ_INIT(&tqpair->tcp_req_working_queue); 1402 SLIST_INIT(&tqpair->tcp_pdu_free_queue); 1403 tqpair->qpair.queue_depth = 0; 1404 1405 tqpair->host_hdgst_enable = true; 1406 tqpair->host_ddgst_enable = true; 1407 1408 return 0; 1409 } 1410 1411 static int 1412 nvmf_tcp_qpair_sock_init(struct spdk_nvmf_tcp_qpair *tqpair) 1413 { 1414 char saddr[32], caddr[32]; 1415 uint16_t sport, cport; 1416 char owner[256]; 1417 int rc; 1418 1419 rc = spdk_sock_getaddr(tqpair->sock, saddr, sizeof(saddr), &sport, 1420 caddr, sizeof(caddr), &cport); 1421 if (rc != 0) { 1422 SPDK_ERRLOG("spdk_sock_getaddr() failed\n"); 1423 return rc; 1424 } 1425 snprintf(owner, sizeof(owner), "%s:%d", caddr, cport); 1426 tqpair->qpair.trace_id = spdk_trace_register_owner(OWNER_TYPE_NVMF_TCP, owner); 1427 spdk_trace_record(TRACE_TCP_QP_SOCK_INIT, tqpair->qpair.trace_id, 0, 0); 1428 1429 /* set low water mark */ 1430 rc = spdk_sock_set_recvlowat(tqpair->sock, 1); 1431 if (rc != 0) { 1432 SPDK_ERRLOG("spdk_sock_set_recvlowat() failed\n"); 1433 return rc; 1434 } 1435 1436 return 0; 1437 } 1438 1439 static void 1440 nvmf_tcp_handle_connect(struct spdk_nvmf_tcp_port *port, struct spdk_sock *sock) 1441 { 1442 struct spdk_nvmf_tcp_qpair *tqpair; 1443 int rc; 1444 1445 SPDK_DEBUGLOG(nvmf_tcp, "New connection accepted on %s port %s\n", 1446 port->trid->traddr, port->trid->trsvcid); 1447 1448 tqpair = calloc(1, sizeof(struct spdk_nvmf_tcp_qpair)); 1449 if (tqpair == NULL) { 1450 SPDK_ERRLOG("Could not allocate new connection.\n"); 1451 spdk_sock_close(&sock); 1452 return; 1453 } 1454 1455 tqpair->sock = sock; 1456 tqpair->state_cntr[TCP_REQUEST_STATE_FREE] = 0; 1457 tqpair->port = port; 1458 tqpair->qpair.transport = port->transport; 1459 tqpair->qpair.numa.id_valid = 1; 1460 tqpair->qpair.numa.id = spdk_sock_get_numa_id(sock); 1461 1462 rc = spdk_sock_getaddr(tqpair->sock, tqpair->target_addr, 1463 sizeof(tqpair->target_addr), &tqpair->target_port, 1464 tqpair->initiator_addr, sizeof(tqpair->initiator_addr), 1465 &tqpair->initiator_port); 1466 if (rc < 0) { 1467 SPDK_ERRLOG("spdk_sock_getaddr() failed of tqpair=%p\n", tqpair); 1468 nvmf_tcp_qpair_destroy(tqpair); 1469 return; 1470 } 1471 1472 spdk_nvmf_tgt_new_qpair(port->transport->tgt, &tqpair->qpair); 1473 } 1474 1475 static uint32_t 1476 nvmf_tcp_port_accept(struct spdk_nvmf_tcp_port *port) 1477 { 1478 struct spdk_sock *sock; 1479 uint32_t count = 0; 1480 int i; 1481 1482 for (i = 0; i < NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME; i++) { 1483 sock = spdk_sock_accept(port->listen_sock); 1484 if (sock == NULL) { 1485 break; 1486 } 1487 count++; 1488 nvmf_tcp_handle_connect(port, sock); 1489 } 1490 1491 return count; 1492 } 1493 1494 static int 1495 nvmf_tcp_accept(void *ctx) 1496 { 1497 struct spdk_nvmf_transport *transport = ctx; 1498 struct spdk_nvmf_tcp_transport *ttransport; 1499 int count; 1500 1501 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1502 1503 count = spdk_sock_group_poll(ttransport->listen_sock_group); 1504 if (count < 0) { 1505 SPDK_ERRLOG("Fail in TCP listen socket group poll\n"); 1506 } 1507 1508 return count != 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 1509 } 1510 1511 static void 1512 nvmf_tcp_accept_cb(void *ctx, struct spdk_sock_group *group, struct spdk_sock *sock) 1513 { 1514 struct spdk_nvmf_tcp_port *port = ctx; 1515 1516 nvmf_tcp_port_accept(port); 1517 } 1518 1519 static void 1520 nvmf_tcp_discover(struct spdk_nvmf_transport *transport, 1521 struct spdk_nvme_transport_id *trid, 1522 struct spdk_nvmf_discovery_log_page_entry *entry) 1523 { 1524 struct spdk_nvmf_tcp_port *port; 1525 struct spdk_nvmf_tcp_transport *ttransport; 1526 1527 entry->trtype = SPDK_NVMF_TRTYPE_TCP; 1528 entry->adrfam = trid->adrfam; 1529 1530 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 1531 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 1532 1533 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1534 port = nvmf_tcp_find_port(ttransport, trid); 1535 1536 assert(port != NULL); 1537 1538 if (strcmp(spdk_sock_get_impl_name(port->listen_sock), "ssl") == 0) { 1539 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_REQUIRED; 1540 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_TLS_1_3; 1541 } else { 1542 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 1543 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_NONE; 1544 } 1545 } 1546 1547 static struct spdk_nvmf_tcp_control_msg_list * 1548 nvmf_tcp_control_msg_list_create(uint16_t num_messages) 1549 { 1550 struct spdk_nvmf_tcp_control_msg_list *list; 1551 struct spdk_nvmf_tcp_control_msg *msg; 1552 uint16_t i; 1553 1554 list = calloc(1, sizeof(*list)); 1555 if (!list) { 1556 SPDK_ERRLOG("Failed to allocate memory for list structure\n"); 1557 return NULL; 1558 } 1559 1560 list->msg_buf = spdk_zmalloc(num_messages * SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, 1561 NVMF_DATA_BUFFER_ALIGNMENT, NULL, SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 1562 if (!list->msg_buf) { 1563 SPDK_ERRLOG("Failed to allocate memory for control message buffers\n"); 1564 free(list); 1565 return NULL; 1566 } 1567 1568 STAILQ_INIT(&list->free_msgs); 1569 STAILQ_INIT(&list->waiting_for_msg_reqs); 1570 1571 for (i = 0; i < num_messages; i++) { 1572 msg = (struct spdk_nvmf_tcp_control_msg *)((char *)list->msg_buf + i * 1573 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1574 STAILQ_INSERT_TAIL(&list->free_msgs, msg, link); 1575 } 1576 1577 return list; 1578 } 1579 1580 static void 1581 nvmf_tcp_control_msg_list_free(struct spdk_nvmf_tcp_control_msg_list *list) 1582 { 1583 if (!list) { 1584 return; 1585 } 1586 1587 spdk_free(list->msg_buf); 1588 free(list); 1589 } 1590 1591 static int nvmf_tcp_poll_group_poll(struct spdk_nvmf_transport_poll_group *group); 1592 1593 static int 1594 nvmf_tcp_poll_group_intr(void *ctx) 1595 { 1596 struct spdk_nvmf_transport_poll_group *group = ctx; 1597 int ret = 0; 1598 1599 ret = nvmf_tcp_poll_group_poll(group); 1600 1601 return ret != 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 1602 } 1603 1604 static struct spdk_nvmf_transport_poll_group * 1605 nvmf_tcp_poll_group_create(struct spdk_nvmf_transport *transport, 1606 struct spdk_nvmf_poll_group *group) 1607 { 1608 struct spdk_nvmf_tcp_transport *ttransport; 1609 struct spdk_nvmf_tcp_poll_group *tgroup; 1610 int rc; 1611 1612 tgroup = calloc(1, sizeof(*tgroup)); 1613 if (!tgroup) { 1614 return NULL; 1615 } 1616 1617 tgroup->sock_group = spdk_sock_group_create(&tgroup->group); 1618 if (!tgroup->sock_group) { 1619 goto cleanup; 1620 } 1621 1622 TAILQ_INIT(&tgroup->qpairs); 1623 TAILQ_INIT(&tgroup->await_req); 1624 1625 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1626 1627 if (transport->opts.in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 1628 SPDK_DEBUGLOG(nvmf_tcp, "ICD %u is less than min required for admin/fabric commands (%u). " 1629 "Creating control messages list\n", transport->opts.in_capsule_data_size, 1630 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1631 tgroup->control_msg_list = nvmf_tcp_control_msg_list_create(ttransport->tcp_opts.control_msg_num); 1632 if (!tgroup->control_msg_list) { 1633 goto cleanup; 1634 } 1635 } 1636 1637 tgroup->accel_channel = spdk_accel_get_io_channel(); 1638 if (spdk_unlikely(!tgroup->accel_channel)) { 1639 SPDK_ERRLOG("Cannot create accel_channel for tgroup=%p\n", tgroup); 1640 goto cleanup; 1641 } 1642 1643 TAILQ_INSERT_TAIL(&ttransport->poll_groups, tgroup, link); 1644 if (ttransport->next_pg == NULL) { 1645 ttransport->next_pg = tgroup; 1646 } 1647 1648 if (spdk_interrupt_mode_is_enabled()) { 1649 rc = SPDK_SOCK_GROUP_REGISTER_INTERRUPT(tgroup->sock_group, 1650 SPDK_INTERRUPT_EVENT_IN | SPDK_INTERRUPT_EVENT_OUT, nvmf_tcp_poll_group_intr, &tgroup->group); 1651 if (rc != 0) { 1652 SPDK_ERRLOG("Failed to register interrupt for sock group\n"); 1653 goto cleanup; 1654 } 1655 } 1656 1657 return &tgroup->group; 1658 1659 cleanup: 1660 nvmf_tcp_poll_group_destroy(&tgroup->group); 1661 return NULL; 1662 } 1663 1664 static struct spdk_nvmf_transport_poll_group * 1665 nvmf_tcp_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 1666 { 1667 struct spdk_nvmf_tcp_transport *ttransport; 1668 struct spdk_nvmf_tcp_poll_group **pg; 1669 struct spdk_nvmf_tcp_qpair *tqpair; 1670 struct spdk_sock_group *group = NULL, *hint = NULL; 1671 int rc; 1672 1673 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 1674 1675 if (TAILQ_EMPTY(&ttransport->poll_groups)) { 1676 return NULL; 1677 } 1678 1679 pg = &ttransport->next_pg; 1680 assert(*pg != NULL); 1681 hint = (*pg)->sock_group; 1682 1683 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1684 rc = spdk_sock_get_optimal_sock_group(tqpair->sock, &group, hint); 1685 if (rc != 0) { 1686 return NULL; 1687 } else if (group != NULL) { 1688 /* Optimal poll group was found */ 1689 return spdk_sock_group_get_ctx(group); 1690 } 1691 1692 /* The hint was used for optimal poll group, advance next_pg. */ 1693 *pg = TAILQ_NEXT(*pg, link); 1694 if (*pg == NULL) { 1695 *pg = TAILQ_FIRST(&ttransport->poll_groups); 1696 } 1697 1698 return spdk_sock_group_get_ctx(hint); 1699 } 1700 1701 static void 1702 nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 1703 { 1704 struct spdk_nvmf_tcp_poll_group *tgroup, *next_tgroup; 1705 struct spdk_nvmf_tcp_transport *ttransport; 1706 1707 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 1708 spdk_sock_group_unregister_interrupt(tgroup->sock_group); 1709 spdk_sock_group_close(&tgroup->sock_group); 1710 if (tgroup->control_msg_list) { 1711 nvmf_tcp_control_msg_list_free(tgroup->control_msg_list); 1712 } 1713 1714 if (tgroup->accel_channel) { 1715 spdk_put_io_channel(tgroup->accel_channel); 1716 } 1717 1718 if (tgroup->group.transport == NULL) { 1719 /* Transport can be NULL when nvmf_tcp_poll_group_create() 1720 * calls this function directly in a failure path. */ 1721 free(tgroup); 1722 return; 1723 } 1724 1725 ttransport = SPDK_CONTAINEROF(tgroup->group.transport, struct spdk_nvmf_tcp_transport, transport); 1726 1727 next_tgroup = TAILQ_NEXT(tgroup, link); 1728 TAILQ_REMOVE(&ttransport->poll_groups, tgroup, link); 1729 if (next_tgroup == NULL) { 1730 next_tgroup = TAILQ_FIRST(&ttransport->poll_groups); 1731 } 1732 if (ttransport->next_pg == tgroup) { 1733 ttransport->next_pg = next_tgroup; 1734 } 1735 1736 free(tgroup); 1737 } 1738 1739 static void 1740 nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 1741 enum nvme_tcp_pdu_recv_state state) 1742 { 1743 if (tqpair->recv_state == state) { 1744 SPDK_ERRLOG("The recv state of tqpair=%p is same with the state(%d) to be set\n", 1745 tqpair, state); 1746 return; 1747 } 1748 1749 if (spdk_unlikely(state == NVME_TCP_PDU_RECV_STATE_QUIESCING)) { 1750 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH && tqpair->pdu_in_progress) { 1751 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, tqpair->pdu_in_progress, slist); 1752 tqpair->tcp_pdu_working_count--; 1753 } 1754 } 1755 1756 if (spdk_unlikely(state == NVME_TCP_PDU_RECV_STATE_ERROR)) { 1757 assert(tqpair->tcp_pdu_working_count == 0); 1758 } 1759 1760 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1761 /* When leaving the await req state, move the qpair to the main list */ 1762 TAILQ_REMOVE(&tqpair->group->await_req, tqpair, link); 1763 TAILQ_INSERT_TAIL(&tqpair->group->qpairs, tqpair, link); 1764 } else if (state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1765 TAILQ_REMOVE(&tqpair->group->qpairs, tqpair, link); 1766 TAILQ_INSERT_TAIL(&tqpair->group->await_req, tqpair, link); 1767 } 1768 1769 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv state=%d\n", tqpair, state); 1770 tqpair->recv_state = state; 1771 1772 spdk_trace_record(TRACE_TCP_QP_RCV_STATE_CHANGE, tqpair->qpair.trace_id, 0, 0, 1773 (uint64_t)tqpair->recv_state); 1774 } 1775 1776 static int 1777 nvmf_tcp_qpair_handle_timeout(void *ctx) 1778 { 1779 struct spdk_nvmf_tcp_qpair *tqpair = ctx; 1780 1781 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1782 1783 SPDK_ERRLOG("No pdu coming for tqpair=%p within %d seconds\n", tqpair, 1784 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT); 1785 1786 nvmf_tcp_qpair_disconnect(tqpair); 1787 return SPDK_POLLER_BUSY; 1788 } 1789 1790 static void 1791 nvmf_tcp_send_c2h_term_req_complete(void *cb_arg) 1792 { 1793 struct spdk_nvmf_tcp_qpair *tqpair = (struct spdk_nvmf_tcp_qpair *)cb_arg; 1794 1795 if (!tqpair->timeout_poller) { 1796 tqpair->timeout_poller = SPDK_POLLER_REGISTER(nvmf_tcp_qpair_handle_timeout, tqpair, 1797 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT * 1000000); 1798 } 1799 } 1800 1801 static void 1802 nvmf_tcp_send_c2h_term_req(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu, 1803 enum spdk_nvme_tcp_term_req_fes fes, uint32_t error_offset) 1804 { 1805 struct nvme_tcp_pdu *rsp_pdu; 1806 struct spdk_nvme_tcp_term_req_hdr *c2h_term_req; 1807 uint32_t c2h_term_req_hdr_len = sizeof(*c2h_term_req); 1808 uint32_t copy_len; 1809 1810 rsp_pdu = tqpair->mgmt_pdu; 1811 1812 c2h_term_req = &rsp_pdu->hdr.term_req; 1813 c2h_term_req->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ; 1814 c2h_term_req->common.hlen = c2h_term_req_hdr_len; 1815 c2h_term_req->fes = fes; 1816 1817 if ((fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1818 (fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1819 DSET32(&c2h_term_req->fei, error_offset); 1820 } 1821 1822 copy_len = spdk_min(pdu->hdr.common.hlen, SPDK_NVME_TCP_TERM_REQ_ERROR_DATA_MAX_SIZE); 1823 1824 /* Copy the error info into the buffer */ 1825 memcpy((uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, pdu->hdr.raw, copy_len); 1826 nvme_tcp_pdu_set_data(rsp_pdu, (uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, copy_len); 1827 1828 /* Contain the header of the wrong received pdu */ 1829 c2h_term_req->common.plen = c2h_term_req->common.hlen + copy_len; 1830 tqpair->wait_terminate = true; 1831 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1832 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_c2h_term_req_complete, tqpair); 1833 } 1834 1835 static void 1836 nvmf_tcp_capsule_cmd_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1837 struct spdk_nvmf_tcp_qpair *tqpair, 1838 struct nvme_tcp_pdu *pdu) 1839 { 1840 struct spdk_nvmf_tcp_req *tcp_req; 1841 1842 assert(pdu->psh_valid_bytes == pdu->psh_len); 1843 assert(pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD); 1844 1845 tcp_req = nvmf_tcp_req_get(tqpair); 1846 if (!tcp_req) { 1847 /* Directly return and make the allocation retry again. This can happen if we're 1848 * using asynchronous writes to send the response to the host or when releasing 1849 * zero-copy buffers after a response has been sent. In both cases, the host might 1850 * receive the response before we've finished processing the request and is free to 1851 * send another one. 1852 */ 1853 if (tqpair->state_cntr[TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST] > 0 || 1854 tqpair->state_cntr[TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE] > 0) { 1855 return; 1856 } 1857 1858 /* The host sent more commands than the maximum queue depth. */ 1859 SPDK_ERRLOG("Cannot allocate tcp_req on tqpair=%p\n", tqpair); 1860 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1861 return; 1862 } 1863 1864 pdu->req = tcp_req; 1865 assert(tcp_req->state == TCP_REQUEST_STATE_NEW); 1866 nvmf_tcp_req_process(ttransport, tcp_req); 1867 } 1868 1869 static void 1870 nvmf_tcp_capsule_cmd_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1871 struct spdk_nvmf_tcp_qpair *tqpair, 1872 struct nvme_tcp_pdu *pdu) 1873 { 1874 struct spdk_nvmf_tcp_req *tcp_req; 1875 struct spdk_nvme_tcp_cmd *capsule_cmd; 1876 uint32_t error_offset = 0; 1877 enum spdk_nvme_tcp_term_req_fes fes; 1878 struct spdk_nvme_cpl *rsp; 1879 1880 capsule_cmd = &pdu->hdr.capsule_cmd; 1881 tcp_req = pdu->req; 1882 assert(tcp_req != NULL); 1883 1884 /* Zero-copy requests don't support ICD */ 1885 assert(!spdk_nvmf_request_using_zcopy(&tcp_req->req)); 1886 1887 if (capsule_cmd->common.pdo > SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET) { 1888 SPDK_ERRLOG("Expected ICReq capsule_cmd pdu offset <= %d, got %c\n", 1889 SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET, capsule_cmd->common.pdo); 1890 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1891 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 1892 goto err; 1893 } 1894 1895 rsp = &tcp_req->req.rsp->nvme_cpl; 1896 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1897 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1898 } else { 1899 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1900 } 1901 1902 nvmf_tcp_req_process(ttransport, tcp_req); 1903 1904 return; 1905 err: 1906 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1907 } 1908 1909 static void 1910 nvmf_tcp_h2c_data_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1911 struct spdk_nvmf_tcp_qpair *tqpair, 1912 struct nvme_tcp_pdu *pdu) 1913 { 1914 struct spdk_nvmf_tcp_req *tcp_req; 1915 uint32_t error_offset = 0; 1916 enum spdk_nvme_tcp_term_req_fes fes = 0; 1917 struct spdk_nvme_tcp_h2c_data_hdr *h2c_data; 1918 1919 h2c_data = &pdu->hdr.h2c_data; 1920 1921 SPDK_DEBUGLOG(nvmf_tcp, "tqpair=%p, r2t_info: datao=%u, datal=%u, cccid=%u, ttag=%u\n", 1922 tqpair, h2c_data->datao, h2c_data->datal, h2c_data->cccid, h2c_data->ttag); 1923 1924 if (h2c_data->ttag > tqpair->resource_count) { 1925 SPDK_DEBUGLOG(nvmf_tcp, "ttag %u is larger than allowed %u.\n", h2c_data->ttag, 1926 tqpair->resource_count); 1927 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1928 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1929 goto err; 1930 } 1931 1932 tcp_req = &tqpair->reqs[h2c_data->ttag - 1]; 1933 1934 if (spdk_unlikely(tcp_req->state != TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER && 1935 tcp_req->state != TCP_REQUEST_STATE_AWAITING_R2T_ACK)) { 1936 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, has error state in %d\n", tcp_req, tqpair, 1937 tcp_req->state); 1938 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1939 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1940 goto err; 1941 } 1942 1943 if (spdk_unlikely(tcp_req->req.cmd->nvme_cmd.cid != h2c_data->cccid)) { 1944 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, expected %u but %u for cccid.\n", tcp_req, tqpair, 1945 tcp_req->req.cmd->nvme_cmd.cid, h2c_data->cccid); 1946 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1947 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, cccid); 1948 goto err; 1949 } 1950 1951 if (tcp_req->h2c_offset != h2c_data->datao) { 1952 SPDK_DEBUGLOG(nvmf_tcp, 1953 "tcp_req(%p), tqpair=%p, expected data offset %u, but data offset is %u\n", 1954 tcp_req, tqpair, tcp_req->h2c_offset, h2c_data->datao); 1955 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1956 goto err; 1957 } 1958 1959 if ((h2c_data->datao + h2c_data->datal) > tcp_req->req.length) { 1960 SPDK_DEBUGLOG(nvmf_tcp, 1961 "tcp_req(%p), tqpair=%p, (datao=%u + datal=%u) exceeds requested length=%u\n", 1962 tcp_req, tqpair, h2c_data->datao, h2c_data->datal, tcp_req->req.length); 1963 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1964 goto err; 1965 } 1966 1967 pdu->req = tcp_req; 1968 1969 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 1970 pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 1971 } 1972 1973 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 1974 h2c_data->datao, h2c_data->datal); 1975 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1976 return; 1977 1978 err: 1979 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1980 } 1981 1982 static void 1983 nvmf_tcp_send_capsule_resp_pdu(struct spdk_nvmf_tcp_req *tcp_req, 1984 struct spdk_nvmf_tcp_qpair *tqpair) 1985 { 1986 struct nvme_tcp_pdu *rsp_pdu; 1987 struct spdk_nvme_tcp_rsp *capsule_resp; 1988 1989 SPDK_DEBUGLOG(nvmf_tcp, "enter, tqpair=%p\n", tqpair); 1990 1991 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1992 assert(rsp_pdu != NULL); 1993 1994 capsule_resp = &rsp_pdu->hdr.capsule_resp; 1995 capsule_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_CAPSULE_RESP; 1996 capsule_resp->common.plen = capsule_resp->common.hlen = sizeof(*capsule_resp); 1997 capsule_resp->rccqe = tcp_req->req.rsp->nvme_cpl; 1998 if (tqpair->host_hdgst_enable) { 1999 capsule_resp->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2000 capsule_resp->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 2001 } 2002 2003 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_request_free, tcp_req); 2004 } 2005 2006 static void 2007 nvmf_tcp_pdu_c2h_data_complete(void *cb_arg) 2008 { 2009 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 2010 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, 2011 struct spdk_nvmf_tcp_qpair, qpair); 2012 2013 assert(tqpair != NULL); 2014 2015 if (spdk_unlikely(tcp_req->pdu->rw_offset < tcp_req->req.length)) { 2016 SPDK_DEBUGLOG(nvmf_tcp, "sending another C2H part, offset %u length %u\n", tcp_req->pdu->rw_offset, 2017 tcp_req->req.length); 2018 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2019 return; 2020 } 2021 2022 if (tcp_req->pdu->hdr.c2h_data.common.flags & SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS) { 2023 nvmf_tcp_request_free(tcp_req); 2024 } else { 2025 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2026 } 2027 } 2028 2029 static void 2030 nvmf_tcp_r2t_complete(void *cb_arg) 2031 { 2032 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 2033 struct spdk_nvmf_tcp_transport *ttransport; 2034 2035 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 2036 struct spdk_nvmf_tcp_transport, transport); 2037 2038 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2039 2040 if (tcp_req->h2c_offset == tcp_req->req.length) { 2041 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2042 nvmf_tcp_req_process(ttransport, tcp_req); 2043 } 2044 } 2045 2046 static void 2047 nvmf_tcp_send_r2t_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 2048 struct spdk_nvmf_tcp_req *tcp_req) 2049 { 2050 struct nvme_tcp_pdu *rsp_pdu; 2051 struct spdk_nvme_tcp_r2t_hdr *r2t; 2052 2053 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 2054 assert(rsp_pdu != NULL); 2055 2056 r2t = &rsp_pdu->hdr.r2t; 2057 r2t->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_R2T; 2058 r2t->common.plen = r2t->common.hlen = sizeof(*r2t); 2059 2060 if (tqpair->host_hdgst_enable) { 2061 r2t->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2062 r2t->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 2063 } 2064 2065 r2t->cccid = tcp_req->req.cmd->nvme_cmd.cid; 2066 r2t->ttag = tcp_req->ttag; 2067 r2t->r2to = tcp_req->h2c_offset; 2068 r2t->r2tl = tcp_req->req.length; 2069 2070 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 2071 2072 SPDK_DEBUGLOG(nvmf_tcp, 2073 "tcp_req(%p) on tqpair(%p), r2t_info: cccid=%u, ttag=%u, r2to=%u, r2tl=%u\n", 2074 tcp_req, tqpair, r2t->cccid, r2t->ttag, r2t->r2to, r2t->r2tl); 2075 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_r2t_complete, tcp_req); 2076 } 2077 2078 static void 2079 nvmf_tcp_h2c_data_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 2080 struct spdk_nvmf_tcp_qpair *tqpair, 2081 struct nvme_tcp_pdu *pdu) 2082 { 2083 struct spdk_nvmf_tcp_req *tcp_req; 2084 struct spdk_nvme_cpl *rsp; 2085 2086 tcp_req = pdu->req; 2087 assert(tcp_req != NULL); 2088 2089 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2090 2091 tcp_req->h2c_offset += pdu->data_len; 2092 2093 /* Wait for all of the data to arrive AND for the initial R2T PDU send to be 2094 * acknowledged before moving on. */ 2095 if (tcp_req->h2c_offset == tcp_req->req.length && 2096 tcp_req->state == TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER) { 2097 /* After receiving all the h2c data, we need to check whether there is 2098 * transient transport error */ 2099 rsp = &tcp_req->req.rsp->nvme_cpl; 2100 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 2101 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2102 } else { 2103 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2104 } 2105 nvmf_tcp_req_process(ttransport, tcp_req); 2106 } 2107 } 2108 2109 static void 2110 nvmf_tcp_h2c_term_req_dump(struct spdk_nvme_tcp_term_req_hdr *h2c_term_req) 2111 { 2112 SPDK_ERRLOG("Error info of pdu(%p): %s\n", h2c_term_req, 2113 spdk_nvmf_tcp_term_req_fes_str[h2c_term_req->fes]); 2114 if ((h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 2115 (h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 2116 SPDK_DEBUGLOG(nvmf_tcp, "The offset from the start of the PDU header is %u\n", 2117 DGET32(h2c_term_req->fei)); 2118 } 2119 } 2120 2121 static void 2122 nvmf_tcp_h2c_term_req_hdr_handle(struct spdk_nvmf_tcp_qpair *tqpair, 2123 struct nvme_tcp_pdu *pdu) 2124 { 2125 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 2126 uint32_t error_offset = 0; 2127 enum spdk_nvme_tcp_term_req_fes fes; 2128 2129 if (h2c_term_req->fes > SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER) { 2130 SPDK_ERRLOG("Fatal Error Status(FES) is unknown for h2c_term_req pdu=%p\n", pdu); 2131 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2132 error_offset = offsetof(struct spdk_nvme_tcp_term_req_hdr, fes); 2133 goto end; 2134 } 2135 2136 /* set the data buffer */ 2137 nvme_tcp_pdu_set_data(pdu, (uint8_t *)pdu->hdr.raw + h2c_term_req->common.hlen, 2138 h2c_term_req->common.plen - h2c_term_req->common.hlen); 2139 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2140 return; 2141 end: 2142 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2143 } 2144 2145 static void 2146 nvmf_tcp_h2c_term_req_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, 2147 struct nvme_tcp_pdu *pdu) 2148 { 2149 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 2150 2151 nvmf_tcp_h2c_term_req_dump(h2c_term_req); 2152 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2153 } 2154 2155 static void 2156 _nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 2157 { 2158 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2159 struct spdk_nvmf_tcp_transport, transport); 2160 2161 switch (pdu->hdr.common.pdu_type) { 2162 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2163 nvmf_tcp_capsule_cmd_payload_handle(ttransport, tqpair, pdu); 2164 break; 2165 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2166 nvmf_tcp_h2c_data_payload_handle(ttransport, tqpair, pdu); 2167 break; 2168 2169 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2170 nvmf_tcp_h2c_term_req_payload_handle(tqpair, pdu); 2171 break; 2172 2173 default: 2174 /* The code should not go to here */ 2175 SPDK_ERRLOG("ERROR pdu type %d\n", pdu->hdr.common.pdu_type); 2176 break; 2177 } 2178 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 2179 tqpair->tcp_pdu_working_count--; 2180 } 2181 2182 static inline void 2183 nvmf_tcp_req_set_cpl(struct spdk_nvmf_tcp_req *treq, int sct, int sc) 2184 { 2185 treq->req.rsp->nvme_cpl.status.sct = sct; 2186 treq->req.rsp->nvme_cpl.status.sc = sc; 2187 treq->req.rsp->nvme_cpl.cid = treq->req.cmd->nvme_cmd.cid; 2188 } 2189 2190 static void 2191 data_crc32_calc_done(void *cb_arg, int status) 2192 { 2193 struct nvme_tcp_pdu *pdu = cb_arg; 2194 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 2195 2196 /* async crc32 calculation is failed and use direct calculation to check */ 2197 if (spdk_unlikely(status)) { 2198 SPDK_ERRLOG("Data digest on tqpair=(%p) with pdu=%p failed to be calculated asynchronously\n", 2199 tqpair, pdu); 2200 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 2201 } 2202 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 2203 if (!MATCH_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32)) { 2204 SPDK_ERRLOG("Data digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 2205 assert(pdu->req != NULL); 2206 nvmf_tcp_req_set_cpl(pdu->req, SPDK_NVME_SCT_GENERIC, 2207 SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR); 2208 } 2209 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2210 } 2211 2212 static void 2213 nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 2214 { 2215 int rc = 0; 2216 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2217 tqpair->pdu_in_progress = NULL; 2218 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2219 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2220 /* check data digest if need */ 2221 if (pdu->ddgst_enable) { 2222 if (tqpair->qpair.qid != 0 && !pdu->dif_ctx && tqpair->group && 2223 (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0)) { 2224 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 2225 pdu->data_iovcnt, 0, data_crc32_calc_done, pdu); 2226 if (spdk_likely(rc == 0)) { 2227 return; 2228 } 2229 } else { 2230 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 2231 } 2232 data_crc32_calc_done(pdu, rc); 2233 } else { 2234 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2235 } 2236 } 2237 2238 static void 2239 nvmf_tcp_send_icresp_complete(void *cb_arg) 2240 { 2241 struct spdk_nvmf_tcp_qpair *tqpair = cb_arg; 2242 2243 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_RUNNING); 2244 } 2245 2246 static void 2247 nvmf_tcp_icreq_handle(struct spdk_nvmf_tcp_transport *ttransport, 2248 struct spdk_nvmf_tcp_qpair *tqpair, 2249 struct nvme_tcp_pdu *pdu) 2250 { 2251 struct spdk_nvme_tcp_ic_req *ic_req = &pdu->hdr.ic_req; 2252 struct nvme_tcp_pdu *rsp_pdu; 2253 struct spdk_nvme_tcp_ic_resp *ic_resp; 2254 uint32_t error_offset = 0; 2255 enum spdk_nvme_tcp_term_req_fes fes; 2256 2257 /* Only PFV 0 is defined currently */ 2258 if (ic_req->pfv != 0) { 2259 SPDK_ERRLOG("Expected ICReq PFV %u, got %u\n", 0u, ic_req->pfv); 2260 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2261 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, pfv); 2262 goto end; 2263 } 2264 2265 /* This value is 0’s based value in units of dwords should not be larger than SPDK_NVME_TCP_HPDA_MAX */ 2266 if (ic_req->hpda > SPDK_NVME_TCP_HPDA_MAX) { 2267 SPDK_ERRLOG("ICReq HPDA out of range 0 to 31, got %u\n", ic_req->hpda); 2268 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2269 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, hpda); 2270 goto end; 2271 } 2272 2273 /* MAXR2T is 0's based */ 2274 SPDK_DEBUGLOG(nvmf_tcp, "maxr2t =%u\n", (ic_req->maxr2t + 1u)); 2275 2276 tqpair->host_hdgst_enable = ic_req->dgst.bits.hdgst_enable ? true : false; 2277 if (!tqpair->host_hdgst_enable) { 2278 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 2279 } 2280 2281 tqpair->host_ddgst_enable = ic_req->dgst.bits.ddgst_enable ? true : false; 2282 if (!tqpair->host_ddgst_enable) { 2283 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 2284 } 2285 2286 tqpair->recv_buf_size = spdk_max(tqpair->recv_buf_size, MIN_SOCK_PIPE_SIZE); 2287 /* Now that we know whether digests are enabled, properly size the receive buffer */ 2288 if (spdk_sock_set_recvbuf(tqpair->sock, tqpair->recv_buf_size) < 0) { 2289 SPDK_WARNLOG("Unable to allocate enough memory for receive buffer on tqpair=%p with size=%d\n", 2290 tqpair, 2291 tqpair->recv_buf_size); 2292 /* Not fatal. */ 2293 } 2294 2295 tqpair->cpda = spdk_min(ic_req->hpda, SPDK_NVME_TCP_CPDA_MAX); 2296 SPDK_DEBUGLOG(nvmf_tcp, "cpda of tqpair=(%p) is : %u\n", tqpair, tqpair->cpda); 2297 2298 rsp_pdu = tqpair->mgmt_pdu; 2299 2300 ic_resp = &rsp_pdu->hdr.ic_resp; 2301 ic_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_IC_RESP; 2302 ic_resp->common.hlen = ic_resp->common.plen = sizeof(*ic_resp); 2303 ic_resp->pfv = 0; 2304 ic_resp->cpda = tqpair->cpda; 2305 ic_resp->maxh2cdata = ttransport->transport.opts.max_io_size; 2306 ic_resp->dgst.bits.hdgst_enable = tqpair->host_hdgst_enable ? 1 : 0; 2307 ic_resp->dgst.bits.ddgst_enable = tqpair->host_ddgst_enable ? 1 : 0; 2308 2309 SPDK_DEBUGLOG(nvmf_tcp, "host_hdgst_enable: %u\n", tqpair->host_hdgst_enable); 2310 SPDK_DEBUGLOG(nvmf_tcp, "host_ddgst_enable: %u\n", tqpair->host_ddgst_enable); 2311 2312 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INITIALIZING); 2313 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_icresp_complete, tqpair); 2314 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2315 return; 2316 end: 2317 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2318 } 2319 2320 static void 2321 nvmf_tcp_pdu_psh_handle(struct spdk_nvmf_tcp_qpair *tqpair, 2322 struct spdk_nvmf_tcp_transport *ttransport) 2323 { 2324 struct nvme_tcp_pdu *pdu; 2325 int rc; 2326 uint32_t crc32c, error_offset = 0; 2327 enum spdk_nvme_tcp_term_req_fes fes; 2328 2329 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2330 pdu = tqpair->pdu_in_progress; 2331 2332 SPDK_DEBUGLOG(nvmf_tcp, "pdu type of tqpair(%p) is %d\n", tqpair, 2333 pdu->hdr.common.pdu_type); 2334 /* check header digest if needed */ 2335 if (pdu->has_hdgst) { 2336 SPDK_DEBUGLOG(nvmf_tcp, "Compare the header of pdu=%p on tqpair=%p\n", pdu, tqpair); 2337 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 2338 rc = MATCH_DIGEST_WORD((uint8_t *)pdu->hdr.raw + pdu->hdr.common.hlen, crc32c); 2339 if (rc == 0) { 2340 SPDK_ERRLOG("Header digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 2341 fes = SPDK_NVME_TCP_TERM_REQ_FES_HDGST_ERROR; 2342 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2343 return; 2344 2345 } 2346 } 2347 2348 switch (pdu->hdr.common.pdu_type) { 2349 case SPDK_NVME_TCP_PDU_TYPE_IC_REQ: 2350 nvmf_tcp_icreq_handle(ttransport, tqpair, pdu); 2351 break; 2352 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2353 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_REQ); 2354 break; 2355 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2356 nvmf_tcp_h2c_data_hdr_handle(ttransport, tqpair, pdu); 2357 break; 2358 2359 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2360 nvmf_tcp_h2c_term_req_hdr_handle(tqpair, pdu); 2361 break; 2362 2363 default: 2364 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", tqpair->pdu_in_progress->hdr.common.pdu_type); 2365 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2366 error_offset = 1; 2367 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2368 break; 2369 } 2370 } 2371 2372 static void 2373 nvmf_tcp_pdu_ch_handle(struct spdk_nvmf_tcp_qpair *tqpair) 2374 { 2375 struct nvme_tcp_pdu *pdu; 2376 uint32_t error_offset = 0; 2377 enum spdk_nvme_tcp_term_req_fes fes; 2378 uint8_t expected_hlen, pdo; 2379 bool plen_error = false, pdo_error = false; 2380 2381 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2382 pdu = tqpair->pdu_in_progress; 2383 assert(pdu); 2384 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_REQ) { 2385 if (tqpair->state != NVME_TCP_QPAIR_STATE_INVALID) { 2386 SPDK_ERRLOG("Already received ICreq PDU, and reject this pdu=%p\n", pdu); 2387 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2388 goto err; 2389 } 2390 expected_hlen = sizeof(struct spdk_nvme_tcp_ic_req); 2391 if (pdu->hdr.common.plen != expected_hlen) { 2392 plen_error = true; 2393 } 2394 } else { 2395 if (tqpair->state != NVME_TCP_QPAIR_STATE_RUNNING) { 2396 SPDK_ERRLOG("The TCP/IP connection is not negotiated\n"); 2397 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2398 goto err; 2399 } 2400 2401 switch (pdu->hdr.common.pdu_type) { 2402 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2403 expected_hlen = sizeof(struct spdk_nvme_tcp_cmd); 2404 pdo = pdu->hdr.common.pdo; 2405 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2406 pdo_error = true; 2407 break; 2408 } 2409 2410 if (pdu->hdr.common.plen < expected_hlen) { 2411 plen_error = true; 2412 } 2413 break; 2414 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2415 expected_hlen = sizeof(struct spdk_nvme_tcp_h2c_data_hdr); 2416 pdo = pdu->hdr.common.pdo; 2417 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2418 pdo_error = true; 2419 break; 2420 } 2421 if (pdu->hdr.common.plen < expected_hlen) { 2422 plen_error = true; 2423 } 2424 break; 2425 2426 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2427 expected_hlen = sizeof(struct spdk_nvme_tcp_term_req_hdr); 2428 if ((pdu->hdr.common.plen <= expected_hlen) || 2429 (pdu->hdr.common.plen > SPDK_NVME_TCP_TERM_REQ_PDU_MAX_SIZE)) { 2430 plen_error = true; 2431 } 2432 break; 2433 2434 default: 2435 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", pdu->hdr.common.pdu_type); 2436 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2437 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdu_type); 2438 goto err; 2439 } 2440 } 2441 2442 if (pdu->hdr.common.hlen != expected_hlen) { 2443 SPDK_ERRLOG("PDU type=0x%02x, Expected ICReq header length %u, got %u on tqpair=%p\n", 2444 pdu->hdr.common.pdu_type, 2445 expected_hlen, pdu->hdr.common.hlen, tqpair); 2446 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2447 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, hlen); 2448 goto err; 2449 } else if (pdo_error) { 2450 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2451 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 2452 } else if (plen_error) { 2453 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2454 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2455 goto err; 2456 } else { 2457 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2458 nvme_tcp_pdu_calc_psh_len(tqpair->pdu_in_progress, tqpair->host_hdgst_enable); 2459 return; 2460 } 2461 err: 2462 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2463 } 2464 2465 static int 2466 nvmf_tcp_sock_process(struct spdk_nvmf_tcp_qpair *tqpair) 2467 { 2468 int rc = 0; 2469 struct nvme_tcp_pdu *pdu; 2470 enum nvme_tcp_pdu_recv_state prev_state; 2471 uint32_t data_len; 2472 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2473 struct spdk_nvmf_tcp_transport, transport); 2474 2475 /* The loop here is to allow for several back-to-back state changes. */ 2476 do { 2477 prev_state = tqpair->recv_state; 2478 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv pdu entering state %d\n", tqpair, prev_state); 2479 2480 pdu = tqpair->pdu_in_progress; 2481 assert(pdu != NULL || 2482 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY || 2483 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_QUIESCING || 2484 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 2485 2486 switch (tqpair->recv_state) { 2487 /* Wait for the common header */ 2488 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY: 2489 if (!pdu) { 2490 pdu = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 2491 if (spdk_unlikely(!pdu)) { 2492 return NVME_TCP_PDU_IN_PROGRESS; 2493 } 2494 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 2495 tqpair->pdu_in_progress = pdu; 2496 tqpair->tcp_pdu_working_count++; 2497 } 2498 memset(pdu, 0, offsetof(struct nvme_tcp_pdu, qpair)); 2499 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2500 /* FALLTHROUGH */ 2501 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH: 2502 if (spdk_unlikely(tqpair->state == NVME_TCP_QPAIR_STATE_INITIALIZING)) { 2503 return rc; 2504 } 2505 2506 rc = nvme_tcp_read_data(tqpair->sock, 2507 sizeof(struct spdk_nvme_tcp_common_pdu_hdr) - pdu->ch_valid_bytes, 2508 (void *)&pdu->hdr.common + pdu->ch_valid_bytes); 2509 if (rc < 0) { 2510 SPDK_DEBUGLOG(nvmf_tcp, "will disconnect tqpair=%p\n", tqpair); 2511 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2512 break; 2513 } else if (rc > 0) { 2514 pdu->ch_valid_bytes += rc; 2515 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.trace_id, rc, 0); 2516 } 2517 2518 if (pdu->ch_valid_bytes < sizeof(struct spdk_nvme_tcp_common_pdu_hdr)) { 2519 return NVME_TCP_PDU_IN_PROGRESS; 2520 } 2521 2522 /* The command header of this PDU has now been read from the socket. */ 2523 nvmf_tcp_pdu_ch_handle(tqpair); 2524 break; 2525 /* Wait for the pdu specific header */ 2526 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH: 2527 rc = nvme_tcp_read_data(tqpair->sock, 2528 pdu->psh_len - pdu->psh_valid_bytes, 2529 (void *)&pdu->hdr.raw + sizeof(struct spdk_nvme_tcp_common_pdu_hdr) + pdu->psh_valid_bytes); 2530 if (rc < 0) { 2531 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2532 break; 2533 } else if (rc > 0) { 2534 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.trace_id, rc, 0); 2535 pdu->psh_valid_bytes += rc; 2536 } 2537 2538 if (pdu->psh_valid_bytes < pdu->psh_len) { 2539 return NVME_TCP_PDU_IN_PROGRESS; 2540 } 2541 2542 /* All header(ch, psh, head digits) of this PDU has now been read from the socket. */ 2543 nvmf_tcp_pdu_psh_handle(tqpair, ttransport); 2544 break; 2545 /* Wait for the req slot */ 2546 case NVME_TCP_PDU_RECV_STATE_AWAIT_REQ: 2547 nvmf_tcp_capsule_cmd_hdr_handle(ttransport, tqpair, pdu); 2548 break; 2549 /* Wait for the request processing loop to acquire a buffer for the PDU */ 2550 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_BUF: 2551 break; 2552 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD: 2553 /* check whether the data is valid, if not we just return */ 2554 if (!pdu->data_len) { 2555 return NVME_TCP_PDU_IN_PROGRESS; 2556 } 2557 2558 data_len = pdu->data_len; 2559 /* data digest */ 2560 if (spdk_unlikely((pdu->hdr.common.pdu_type != SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ) && 2561 tqpair->host_ddgst_enable)) { 2562 data_len += SPDK_NVME_TCP_DIGEST_LEN; 2563 pdu->ddgst_enable = true; 2564 } 2565 2566 rc = nvme_tcp_read_payload_data(tqpair->sock, pdu); 2567 if (rc < 0) { 2568 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2569 break; 2570 } 2571 pdu->rw_offset += rc; 2572 2573 if (pdu->rw_offset < data_len) { 2574 return NVME_TCP_PDU_IN_PROGRESS; 2575 } 2576 2577 /* Generate and insert DIF to whole data block received if DIF is enabled */ 2578 if (spdk_unlikely(pdu->dif_ctx != NULL) && 2579 spdk_dif_generate_stream(pdu->data_iov, pdu->data_iovcnt, 0, data_len, 2580 pdu->dif_ctx) != 0) { 2581 SPDK_ERRLOG("DIF generate failed\n"); 2582 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2583 break; 2584 } 2585 2586 /* All of this PDU has now been read from the socket. */ 2587 nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2588 break; 2589 case NVME_TCP_PDU_RECV_STATE_QUIESCING: 2590 if (tqpair->tcp_pdu_working_count != 0) { 2591 return NVME_TCP_PDU_IN_PROGRESS; 2592 } 2593 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 2594 break; 2595 case NVME_TCP_PDU_RECV_STATE_ERROR: 2596 if (spdk_sock_is_connected(tqpair->sock) && tqpair->wait_terminate) { 2597 return NVME_TCP_PDU_IN_PROGRESS; 2598 } 2599 return NVME_TCP_PDU_FATAL; 2600 default: 2601 SPDK_ERRLOG("The state(%d) is invalid\n", tqpair->recv_state); 2602 abort(); 2603 break; 2604 } 2605 } while (tqpair->recv_state != prev_state); 2606 2607 return rc; 2608 } 2609 2610 static inline void * 2611 nvmf_tcp_control_msg_get(struct spdk_nvmf_tcp_control_msg_list *list, 2612 struct spdk_nvmf_tcp_req *tcp_req) 2613 { 2614 struct spdk_nvmf_tcp_control_msg *msg; 2615 2616 assert(list); 2617 2618 msg = STAILQ_FIRST(&list->free_msgs); 2619 if (!msg) { 2620 SPDK_DEBUGLOG(nvmf_tcp, "Out of control messages\n"); 2621 STAILQ_INSERT_TAIL(&list->waiting_for_msg_reqs, tcp_req, control_msg_link); 2622 return NULL; 2623 } 2624 STAILQ_REMOVE_HEAD(&list->free_msgs, link); 2625 return msg; 2626 } 2627 2628 static inline void 2629 nvmf_tcp_control_msg_put(struct spdk_nvmf_tcp_control_msg_list *list, void *_msg) 2630 { 2631 struct spdk_nvmf_tcp_control_msg *msg = _msg; 2632 struct spdk_nvmf_tcp_req *tcp_req; 2633 struct spdk_nvmf_tcp_transport *ttransport; 2634 2635 assert(list); 2636 STAILQ_INSERT_HEAD(&list->free_msgs, msg, link); 2637 if (!STAILQ_EMPTY(&list->waiting_for_msg_reqs)) { 2638 tcp_req = STAILQ_FIRST(&list->waiting_for_msg_reqs); 2639 STAILQ_REMOVE_HEAD(&list->waiting_for_msg_reqs, control_msg_link); 2640 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 2641 struct spdk_nvmf_tcp_transport, transport); 2642 nvmf_tcp_req_process(ttransport, tcp_req); 2643 } 2644 } 2645 2646 static void 2647 nvmf_tcp_req_parse_sgl(struct spdk_nvmf_tcp_req *tcp_req, 2648 struct spdk_nvmf_transport *transport, 2649 struct spdk_nvmf_transport_poll_group *group) 2650 { 2651 struct spdk_nvmf_request *req = &tcp_req->req; 2652 struct spdk_nvme_cmd *cmd; 2653 struct spdk_nvme_sgl_descriptor *sgl; 2654 struct spdk_nvmf_tcp_poll_group *tgroup; 2655 enum spdk_nvme_tcp_term_req_fes fes; 2656 struct nvme_tcp_pdu *pdu; 2657 struct spdk_nvmf_tcp_qpair *tqpair; 2658 uint32_t length, error_offset = 0; 2659 2660 cmd = &req->cmd->nvme_cmd; 2661 sgl = &cmd->dptr.sgl1; 2662 2663 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK && 2664 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_TRANSPORT) { 2665 /* get request length from sgl */ 2666 length = sgl->unkeyed.length; 2667 if (spdk_unlikely(length > transport->opts.max_io_size)) { 2668 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 2669 length, transport->opts.max_io_size); 2670 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_LIMIT_EXCEEDED; 2671 goto fatal_err; 2672 } 2673 2674 /* fill request length and populate iovs */ 2675 req->length = length; 2676 2677 SPDK_DEBUGLOG(nvmf_tcp, "Data requested length= 0x%x\n", length); 2678 2679 if (spdk_unlikely(req->dif_enabled)) { 2680 req->dif.orig_length = length; 2681 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2682 req->dif.elba_length = length; 2683 } 2684 2685 if (nvmf_ctrlr_use_zcopy(req)) { 2686 SPDK_DEBUGLOG(nvmf_tcp, "Using zero-copy to execute request %p\n", tcp_req); 2687 req->data_from_pool = false; 2688 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_HAVE_BUFFER); 2689 return; 2690 } 2691 2692 if (spdk_nvmf_request_get_buffers(req, group, transport, length)) { 2693 /* No available buffers. Queue this request up. */ 2694 SPDK_DEBUGLOG(nvmf_tcp, "No available large data buffers. Queueing request %p\n", 2695 tcp_req); 2696 return; 2697 } 2698 2699 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_HAVE_BUFFER); 2700 SPDK_DEBUGLOG(nvmf_tcp, "Request %p took %d buffer/s from central pool, and data=%p\n", 2701 tcp_req, req->iovcnt, req->iov[0].iov_base); 2702 2703 return; 2704 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 2705 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 2706 uint64_t offset = sgl->address; 2707 uint32_t max_len = transport->opts.in_capsule_data_size; 2708 2709 assert(tcp_req->has_in_capsule_data); 2710 /* Capsule Cmd with In-capsule Data should get data length from pdu header */ 2711 tqpair = tcp_req->pdu->qpair; 2712 /* receiving pdu is not same with the pdu in tcp_req */ 2713 pdu = tqpair->pdu_in_progress; 2714 length = pdu->hdr.common.plen - pdu->psh_len - sizeof(struct spdk_nvme_tcp_common_pdu_hdr); 2715 if (tqpair->host_ddgst_enable) { 2716 length -= SPDK_NVME_TCP_DIGEST_LEN; 2717 } 2718 /* This error is not defined in NVMe/TCP spec, take this error as fatal error */ 2719 if (spdk_unlikely(length != sgl->unkeyed.length)) { 2720 SPDK_ERRLOG("In-Capsule Data length 0x%x is not equal to SGL data length 0x%x\n", 2721 length, sgl->unkeyed.length); 2722 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2723 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2724 goto fatal_err; 2725 } 2726 2727 SPDK_DEBUGLOG(nvmf_tcp, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 2728 offset, length); 2729 2730 /* The NVMe/TCP transport does not use ICDOFF to control the in-capsule data offset. ICDOFF should be '0' */ 2731 if (spdk_unlikely(offset != 0)) { 2732 /* Not defined fatal error in NVMe/TCP spec, handle this error as a fatal error */ 2733 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " should be ZERO in NVMe/TCP\n", offset); 2734 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER; 2735 error_offset = offsetof(struct spdk_nvme_tcp_cmd, ccsqe.dptr.sgl1.address); 2736 goto fatal_err; 2737 } 2738 2739 if (spdk_unlikely(length > max_len)) { 2740 /* According to the SPEC we should support ICD up to 8192 bytes for admin and fabric commands */ 2741 if (length <= SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE && 2742 (cmd->opc == SPDK_NVME_OPC_FABRIC || req->qpair->qid == 0)) { 2743 2744 /* Get a buffer from dedicated list */ 2745 SPDK_DEBUGLOG(nvmf_tcp, "Getting a buffer from control msg list\n"); 2746 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 2747 assert(tgroup->control_msg_list); 2748 req->iov[0].iov_base = nvmf_tcp_control_msg_get(tgroup->control_msg_list, tcp_req); 2749 if (!req->iov[0].iov_base) { 2750 /* No available buffers. Queue this request up. */ 2751 SPDK_DEBUGLOG(nvmf_tcp, "No available ICD buffers. Queueing request %p\n", tcp_req); 2752 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_BUF); 2753 return; 2754 } 2755 } else { 2756 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 2757 length, max_len); 2758 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_LIMIT_EXCEEDED; 2759 goto fatal_err; 2760 } 2761 } else { 2762 req->iov[0].iov_base = tcp_req->buf; 2763 } 2764 2765 req->length = length; 2766 req->data_from_pool = false; 2767 2768 if (spdk_unlikely(req->dif_enabled)) { 2769 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2770 req->dif.elba_length = length; 2771 } 2772 2773 req->iov[0].iov_len = length; 2774 req->iovcnt = 1; 2775 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_HAVE_BUFFER); 2776 2777 return; 2778 } 2779 /* If we want to handle the problem here, then we can't skip the following data segment. 2780 * Because this function runs before reading data part, now handle all errors as fatal errors. */ 2781 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 2782 sgl->generic.type, sgl->generic.subtype); 2783 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER; 2784 error_offset = offsetof(struct spdk_nvme_tcp_cmd, ccsqe.dptr.sgl1.generic); 2785 fatal_err: 2786 nvmf_tcp_send_c2h_term_req(tcp_req->pdu->qpair, tcp_req->pdu, fes, error_offset); 2787 } 2788 2789 static inline enum spdk_nvme_media_error_status_code 2790 nvmf_tcp_dif_error_to_compl_status(uint8_t err_type) { 2791 enum spdk_nvme_media_error_status_code result; 2792 2793 switch (err_type) 2794 { 2795 case SPDK_DIF_REFTAG_ERROR: 2796 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 2797 break; 2798 case SPDK_DIF_APPTAG_ERROR: 2799 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 2800 break; 2801 case SPDK_DIF_GUARD_ERROR: 2802 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 2803 break; 2804 default: 2805 SPDK_UNREACHABLE(); 2806 break; 2807 } 2808 2809 return result; 2810 } 2811 2812 static void 2813 _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2814 struct spdk_nvmf_tcp_req *tcp_req) 2815 { 2816 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF( 2817 tqpair->qpair.transport, struct spdk_nvmf_tcp_transport, transport); 2818 struct nvme_tcp_pdu *rsp_pdu; 2819 struct spdk_nvme_tcp_c2h_data_hdr *c2h_data; 2820 uint32_t plen, pdo, alignment; 2821 int rc; 2822 2823 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2824 2825 rsp_pdu = tcp_req->pdu; 2826 assert(rsp_pdu != NULL); 2827 2828 c2h_data = &rsp_pdu->hdr.c2h_data; 2829 c2h_data->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_DATA; 2830 plen = c2h_data->common.hlen = sizeof(*c2h_data); 2831 2832 if (tqpair->host_hdgst_enable) { 2833 plen += SPDK_NVME_TCP_DIGEST_LEN; 2834 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2835 } 2836 2837 /* set the psh */ 2838 c2h_data->cccid = tcp_req->req.cmd->nvme_cmd.cid; 2839 c2h_data->datal = tcp_req->req.length - tcp_req->pdu->rw_offset; 2840 c2h_data->datao = tcp_req->pdu->rw_offset; 2841 2842 /* set the padding */ 2843 rsp_pdu->padding_len = 0; 2844 pdo = plen; 2845 if (tqpair->cpda) { 2846 alignment = (tqpair->cpda + 1) << 2; 2847 if (plen % alignment != 0) { 2848 pdo = (plen + alignment) / alignment * alignment; 2849 rsp_pdu->padding_len = pdo - plen; 2850 plen = pdo; 2851 } 2852 } 2853 2854 c2h_data->common.pdo = pdo; 2855 plen += c2h_data->datal; 2856 if (tqpair->host_ddgst_enable) { 2857 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_DDGSTF; 2858 plen += SPDK_NVME_TCP_DIGEST_LEN; 2859 } 2860 2861 c2h_data->common.plen = plen; 2862 2863 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2864 rsp_pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 2865 } 2866 2867 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2868 c2h_data->datao, c2h_data->datal); 2869 2870 2871 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU; 2872 /* Need to send the capsule response if response is not all 0 */ 2873 if (ttransport->tcp_opts.c2h_success && 2874 tcp_req->rsp.cdw0 == 0 && tcp_req->rsp.cdw1 == 0) { 2875 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS; 2876 } 2877 2878 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2879 struct spdk_nvme_cpl *rsp = &tcp_req->req.rsp->nvme_cpl; 2880 struct spdk_dif_error err_blk = {}; 2881 uint32_t mapped_length = 0; 2882 uint32_t available_iovs = SPDK_COUNTOF(rsp_pdu->iov); 2883 uint32_t ddgst_len = 0; 2884 2885 if (tqpair->host_ddgst_enable) { 2886 /* Data digest consumes additional iov entry */ 2887 available_iovs--; 2888 /* plen needs to be updated since nvme_tcp_build_iovs compares expected and actual plen */ 2889 ddgst_len = SPDK_NVME_TCP_DIGEST_LEN; 2890 c2h_data->common.plen -= ddgst_len; 2891 } 2892 /* Temp call to estimate if data can be described by limited number of iovs. 2893 * iov vector will be rebuilt in nvmf_tcp_qpair_write_pdu */ 2894 nvme_tcp_build_iovs(rsp_pdu->iov, available_iovs, rsp_pdu, tqpair->host_hdgst_enable, 2895 false, &mapped_length); 2896 2897 if (mapped_length != c2h_data->common.plen) { 2898 c2h_data->datal = mapped_length - (c2h_data->common.plen - c2h_data->datal); 2899 SPDK_DEBUGLOG(nvmf_tcp, 2900 "Part C2H, data_len %u (of %u), PDU len %u, updated PDU len %u, offset %u\n", 2901 c2h_data->datal, tcp_req->req.length, c2h_data->common.plen, mapped_length, rsp_pdu->rw_offset); 2902 c2h_data->common.plen = mapped_length; 2903 2904 /* Rebuild pdu->data_iov since data length is changed */ 2905 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, c2h_data->datao, 2906 c2h_data->datal); 2907 2908 c2h_data->common.flags &= ~(SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU | 2909 SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS); 2910 } 2911 2912 c2h_data->common.plen += ddgst_len; 2913 2914 assert(rsp_pdu->rw_offset <= tcp_req->req.length); 2915 2916 rc = spdk_dif_verify_stream(rsp_pdu->data_iov, rsp_pdu->data_iovcnt, 2917 0, rsp_pdu->data_len, rsp_pdu->dif_ctx, &err_blk); 2918 if (rc != 0) { 2919 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", 2920 err_blk.err_type, err_blk.err_offset); 2921 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2922 rsp->status.sc = nvmf_tcp_dif_error_to_compl_status(err_blk.err_type); 2923 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2924 return; 2925 } 2926 } 2927 2928 rsp_pdu->rw_offset += c2h_data->datal; 2929 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_pdu_c2h_data_complete, tcp_req); 2930 } 2931 2932 static void 2933 nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2934 struct spdk_nvmf_tcp_req *tcp_req) 2935 { 2936 nvmf_tcp_req_pdu_init(tcp_req); 2937 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2938 } 2939 2940 static int 2941 request_transfer_out(struct spdk_nvmf_request *req) 2942 { 2943 struct spdk_nvmf_tcp_req *tcp_req; 2944 struct spdk_nvmf_qpair *qpair; 2945 struct spdk_nvmf_tcp_qpair *tqpair; 2946 struct spdk_nvme_cpl *rsp; 2947 2948 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2949 2950 qpair = req->qpair; 2951 rsp = &req->rsp->nvme_cpl; 2952 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 2953 2954 /* Advance our sq_head pointer */ 2955 if (qpair->sq_head == qpair->sq_head_max) { 2956 qpair->sq_head = 0; 2957 } else { 2958 qpair->sq_head++; 2959 } 2960 rsp->sqhd = qpair->sq_head; 2961 2962 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2963 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2964 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2965 nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2966 } else { 2967 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2968 } 2969 2970 return 0; 2971 } 2972 2973 static void 2974 nvmf_tcp_check_fused_ordering(struct spdk_nvmf_tcp_transport *ttransport, 2975 struct spdk_nvmf_tcp_qpair *tqpair, 2976 struct spdk_nvmf_tcp_req *tcp_req) 2977 { 2978 enum spdk_nvme_cmd_fuse last, next; 2979 2980 last = tqpair->fused_first ? tqpair->fused_first->cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2981 next = tcp_req->cmd.fuse; 2982 2983 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2984 2985 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2986 return; 2987 } 2988 2989 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2990 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2991 /* This is a valid pair of fused commands. Point them at each other 2992 * so they can be submitted consecutively once ready to be executed. 2993 */ 2994 tqpair->fused_first->fused_pair = tcp_req; 2995 tcp_req->fused_pair = tqpair->fused_first; 2996 tqpair->fused_first = NULL; 2997 return; 2998 } else { 2999 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 3000 tqpair->fused_first->fused_failed = true; 3001 3002 /* 3003 * If the last req is in READY_TO_EXECUTE state, then call 3004 * nvmf_tcp_req_process(), otherwise nothing else will kick it. 3005 */ 3006 if (tqpair->fused_first->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3007 nvmf_tcp_req_process(ttransport, tqpair->fused_first); 3008 } 3009 3010 tqpair->fused_first = NULL; 3011 } 3012 } 3013 3014 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 3015 /* Set tqpair->fused_first here so that we know to check that the next request 3016 * is a SECOND (and to fail this one if it isn't). 3017 */ 3018 tqpair->fused_first = tcp_req; 3019 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 3020 /* Mark this req failed since it is a SECOND and the last one was not a FIRST. */ 3021 tcp_req->fused_failed = true; 3022 } 3023 } 3024 3025 static bool 3026 nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 3027 struct spdk_nvmf_tcp_req *tcp_req) 3028 { 3029 struct spdk_nvmf_tcp_qpair *tqpair; 3030 uint32_t plen; 3031 struct nvme_tcp_pdu *pdu; 3032 enum spdk_nvmf_tcp_req_state prev_state; 3033 bool progress = false; 3034 struct spdk_nvmf_transport *transport = &ttransport->transport; 3035 struct spdk_nvmf_transport_poll_group *group; 3036 struct spdk_nvmf_tcp_poll_group *tgroup; 3037 3038 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 3039 group = &tqpair->group->group; 3040 assert(tcp_req->state != TCP_REQUEST_STATE_FREE); 3041 3042 /* If the qpair is not active, we need to abort the outstanding requests. */ 3043 if (!spdk_nvmf_qpair_is_active(&tqpair->qpair)) { 3044 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 3045 nvmf_tcp_request_get_buffers_abort(tcp_req); 3046 } 3047 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 3048 } 3049 3050 /* The loop here is to allow for several back-to-back state changes. */ 3051 do { 3052 prev_state = tcp_req->state; 3053 3054 SPDK_DEBUGLOG(nvmf_tcp, "Request %p entering state %d on tqpair=%p\n", tcp_req, prev_state, 3055 tqpair); 3056 3057 switch (tcp_req->state) { 3058 case TCP_REQUEST_STATE_FREE: 3059 /* Some external code must kick a request into TCP_REQUEST_STATE_NEW 3060 * to escape this state. */ 3061 break; 3062 case TCP_REQUEST_STATE_NEW: 3063 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEW, tqpair->qpair.trace_id, 0, (uintptr_t)tcp_req, 3064 tqpair->qpair.queue_depth); 3065 3066 /* copy the cmd from the receive pdu */ 3067 tcp_req->cmd = tqpair->pdu_in_progress->hdr.capsule_cmd.ccsqe; 3068 3069 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&tcp_req->req, &tcp_req->req.dif.dif_ctx))) { 3070 tcp_req->req.dif_enabled = true; 3071 tqpair->pdu_in_progress->dif_ctx = &tcp_req->req.dif.dif_ctx; 3072 } 3073 3074 nvmf_tcp_check_fused_ordering(ttransport, tqpair, tcp_req); 3075 3076 /* The next state transition depends on the data transfer needs of this request. */ 3077 tcp_req->req.xfer = spdk_nvmf_req_get_xfer(&tcp_req->req); 3078 3079 if (spdk_unlikely(tcp_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 3080 nvmf_tcp_req_set_cpl(tcp_req, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_INVALID_OPCODE); 3081 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 3082 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3083 SPDK_DEBUGLOG(nvmf_tcp, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", tcp_req); 3084 break; 3085 } 3086 3087 /* If no data to transfer, ready to execute. */ 3088 if (tcp_req->req.xfer == SPDK_NVME_DATA_NONE) { 3089 /* Reset the tqpair receiving pdu state */ 3090 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 3091 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 3092 break; 3093 } 3094 3095 pdu = tqpair->pdu_in_progress; 3096 plen = pdu->hdr.common.hlen; 3097 if (tqpair->host_hdgst_enable) { 3098 plen += SPDK_NVME_TCP_DIGEST_LEN; 3099 } 3100 if (pdu->hdr.common.plen != plen) { 3101 tcp_req->has_in_capsule_data = true; 3102 } else { 3103 /* Data is transmitted by C2H PDUs */ 3104 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 3105 } 3106 3107 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEED_BUFFER); 3108 break; 3109 case TCP_REQUEST_STATE_NEED_BUFFER: 3110 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEED_BUFFER, tqpair->qpair.trace_id, 0, 3111 (uintptr_t)tcp_req); 3112 3113 assert(tcp_req->req.xfer != SPDK_NVME_DATA_NONE); 3114 3115 /* Try to get a data buffer */ 3116 nvmf_tcp_req_parse_sgl(tcp_req, transport, group); 3117 break; 3118 case TCP_REQUEST_STATE_HAVE_BUFFER: 3119 spdk_trace_record(TRACE_TCP_REQUEST_STATE_HAVE_BUFFER, tqpair->qpair.trace_id, 0, 3120 (uintptr_t)tcp_req); 3121 /* Get a zcopy buffer if the request can be serviced through zcopy */ 3122 if (spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 3123 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3124 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 3125 tcp_req->req.length = tcp_req->req.dif.elba_length; 3126 } 3127 3128 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_START); 3129 spdk_nvmf_request_zcopy_start(&tcp_req->req); 3130 break; 3131 } 3132 3133 assert(tcp_req->req.iovcnt > 0); 3134 3135 /* If data is transferring from host to controller, we need to do a transfer from the host. */ 3136 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 3137 if (tcp_req->req.data_from_pool) { 3138 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 3139 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 3140 } else { 3141 struct nvme_tcp_pdu *pdu; 3142 3143 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 3144 3145 pdu = tqpair->pdu_in_progress; 3146 SPDK_DEBUGLOG(nvmf_tcp, "Not need to send r2t for tcp_req(%p) on tqpair=%p\n", tcp_req, 3147 tqpair); 3148 /* No need to send r2t, contained in the capsuled data */ 3149 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 3150 0, tcp_req->req.length); 3151 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 3152 } 3153 break; 3154 } 3155 3156 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 3157 break; 3158 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3159 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, tqpair->qpair.trace_id, 0, 3160 (uintptr_t)tcp_req); 3161 /* Some external code must kick a request into TCP_REQUEST_STATE_ZCOPY_START_COMPLETED 3162 * to escape this state. */ 3163 break; 3164 case TCP_REQUEST_STATE_ZCOPY_START_COMPLETED: 3165 spdk_trace_record(TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, tqpair->qpair.trace_id, 0, 3166 (uintptr_t)tcp_req); 3167 if (spdk_unlikely(spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl))) { 3168 SPDK_DEBUGLOG(nvmf_tcp, "Zero-copy start failed for tcp_req(%p) on tqpair=%p\n", 3169 tcp_req, tqpair); 3170 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3171 break; 3172 } 3173 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 3174 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 3175 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 3176 } else { 3177 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3178 } 3179 break; 3180 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3181 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, tqpair->qpair.trace_id, 0, 3182 (uintptr_t)tcp_req); 3183 /* The R2T completion or the h2c data incoming will kick it out of this state. */ 3184 break; 3185 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3186 3187 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, tqpair->qpair.trace_id, 3188 0, (uintptr_t)tcp_req); 3189 /* Some external code must kick a request into TCP_REQUEST_STATE_READY_TO_EXECUTE 3190 * to escape this state. */ 3191 break; 3192 case TCP_REQUEST_STATE_READY_TO_EXECUTE: 3193 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, tqpair->qpair.trace_id, 0, 3194 (uintptr_t)tcp_req); 3195 3196 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3197 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 3198 tcp_req->req.length = tcp_req->req.dif.elba_length; 3199 } 3200 3201 if (tcp_req->cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 3202 if (tcp_req->fused_failed) { 3203 /* This request failed FUSED semantics. Fail it immediately, without 3204 * even sending it to the target layer. 3205 */ 3206 nvmf_tcp_req_set_cpl(tcp_req, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_ABORTED_MISSING_FUSED); 3207 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3208 break; 3209 } 3210 3211 if (tcp_req->fused_pair == NULL || 3212 tcp_req->fused_pair->state != TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3213 /* This request is ready to execute, but either we don't know yet if it's 3214 * valid - i.e. this is a FIRST but we haven't received the next request yet), 3215 * or the other request of this fused pair isn't ready to execute. So 3216 * break here and this request will get processed later either when the 3217 * other request is ready or we find that this request isn't valid. 3218 */ 3219 break; 3220 } 3221 } 3222 3223 if (!spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 3224 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTING); 3225 /* If we get to this point, and this request is a fused command, we know that 3226 * it is part of a valid sequence (FIRST followed by a SECOND) and that both 3227 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 3228 * request, and the other request of the fused pair, in the correct order. 3229 * Also clear the ->fused_pair pointers on both requests, since after this point 3230 * we no longer need to maintain the relationship between these two requests. 3231 */ 3232 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 3233 assert(tcp_req->fused_pair != NULL); 3234 assert(tcp_req->fused_pair->fused_pair == tcp_req); 3235 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 3236 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 3237 tcp_req->fused_pair->fused_pair = NULL; 3238 tcp_req->fused_pair = NULL; 3239 } 3240 spdk_nvmf_request_exec(&tcp_req->req); 3241 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 3242 assert(tcp_req->fused_pair != NULL); 3243 assert(tcp_req->fused_pair->fused_pair == tcp_req); 3244 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 3245 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 3246 tcp_req->fused_pair->fused_pair = NULL; 3247 tcp_req->fused_pair = NULL; 3248 } 3249 } else { 3250 /* For zero-copy, only requests with data coming from host to the 3251 * controller can end up here. */ 3252 assert(tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 3253 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT); 3254 spdk_nvmf_request_zcopy_end(&tcp_req->req, true); 3255 } 3256 3257 break; 3258 case TCP_REQUEST_STATE_EXECUTING: 3259 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTING, tqpair->qpair.trace_id, 0, (uintptr_t)tcp_req); 3260 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 3261 * to escape this state. */ 3262 break; 3263 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3264 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, tqpair->qpair.trace_id, 0, 3265 (uintptr_t)tcp_req); 3266 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 3267 * to escape this state. */ 3268 break; 3269 case TCP_REQUEST_STATE_EXECUTED: 3270 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTED, tqpair->qpair.trace_id, 0, (uintptr_t)tcp_req); 3271 3272 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3273 tcp_req->req.length = tcp_req->req.dif.orig_length; 3274 } 3275 3276 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3277 break; 3278 case TCP_REQUEST_STATE_READY_TO_COMPLETE: 3279 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, tqpair->qpair.trace_id, 0, 3280 (uintptr_t)tcp_req); 3281 if (request_transfer_out(&tcp_req->req) != 0) { 3282 assert(0); /* No good way to handle this currently */ 3283 } 3284 break; 3285 case TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3286 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, tqpair->qpair.trace_id, 3287 0, (uintptr_t)tcp_req); 3288 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 3289 * to escape this state. */ 3290 break; 3291 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3292 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, tqpair->qpair.trace_id, 0, 3293 (uintptr_t)tcp_req); 3294 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 3295 * to escape this state. */ 3296 break; 3297 case TCP_REQUEST_STATE_COMPLETED: 3298 spdk_trace_record(TRACE_TCP_REQUEST_STATE_COMPLETED, tqpair->qpair.trace_id, 0, (uintptr_t)tcp_req, 3299 tqpair->qpair.queue_depth); 3300 /* If there's an outstanding PDU sent to the host, the request is completed 3301 * due to the qpair being disconnected. We must delay the completion until 3302 * that write is done to avoid freeing the request twice. */ 3303 if (spdk_unlikely(tcp_req->pdu_in_use)) { 3304 SPDK_DEBUGLOG(nvmf_tcp, "Delaying completion due to outstanding " 3305 "write on req=%p\n", tcp_req); 3306 /* This can only happen for zcopy requests */ 3307 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 3308 assert(!spdk_nvmf_qpair_is_active(&tqpair->qpair)); 3309 break; 3310 } 3311 3312 if (tcp_req->req.data_from_pool) { 3313 spdk_nvmf_request_free_buffers(&tcp_req->req, group, transport); 3314 } else if (spdk_unlikely(tcp_req->has_in_capsule_data && 3315 (tcp_req->cmd.opc == SPDK_NVME_OPC_FABRIC || 3316 tqpair->qpair.qid == 0) && tcp_req->req.length > transport->opts.in_capsule_data_size)) { 3317 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3318 assert(tgroup->control_msg_list); 3319 SPDK_DEBUGLOG(nvmf_tcp, "Put buf to control msg list\n"); 3320 nvmf_tcp_control_msg_put(tgroup->control_msg_list, 3321 tcp_req->req.iov[0].iov_base); 3322 } else if (tcp_req->req.zcopy_bdev_io != NULL) { 3323 /* If the request has an unreleased zcopy bdev_io, it's either a 3324 * read, a failed write, or the qpair is being disconnected */ 3325 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 3326 assert(tcp_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST || 3327 spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl) || 3328 !spdk_nvmf_qpair_is_active(&tqpair->qpair)); 3329 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE); 3330 spdk_nvmf_request_zcopy_end(&tcp_req->req, false); 3331 break; 3332 } 3333 tcp_req->req.length = 0; 3334 tcp_req->req.iovcnt = 0; 3335 tcp_req->fused_failed = false; 3336 if (tcp_req->fused_pair) { 3337 /* This req was part of a valid fused pair, but failed before it got to 3338 * READ_TO_EXECUTE state. This means we need to fail the other request 3339 * in the pair, because it is no longer part of a valid pair. If the pair 3340 * already reached READY_TO_EXECUTE state, we need to kick it. 3341 */ 3342 tcp_req->fused_pair->fused_failed = true; 3343 if (tcp_req->fused_pair->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3344 nvmf_tcp_req_process(ttransport, tcp_req->fused_pair); 3345 } 3346 tcp_req->fused_pair = NULL; 3347 } 3348 3349 nvmf_tcp_req_put(tqpair, tcp_req); 3350 break; 3351 case TCP_REQUEST_NUM_STATES: 3352 default: 3353 assert(0); 3354 break; 3355 } 3356 3357 if (tcp_req->state != prev_state) { 3358 progress = true; 3359 } 3360 } while (tcp_req->state != prev_state); 3361 3362 return progress; 3363 } 3364 3365 static void 3366 tcp_sock_cb(void *arg) 3367 { 3368 struct spdk_nvmf_tcp_qpair *tqpair = arg; 3369 int rc; 3370 3371 assert(tqpair != NULL); 3372 rc = nvmf_tcp_sock_process(tqpair); 3373 3374 /* If there was a new socket error, disconnect */ 3375 if (rc < 0) { 3376 nvmf_tcp_qpair_disconnect(tqpair); 3377 } 3378 } 3379 3380 static void 3381 nvmf_tcp_sock_cb(void *arg, struct spdk_sock_group *group, struct spdk_sock *sock) 3382 { 3383 tcp_sock_cb(arg); 3384 } 3385 3386 static int 3387 nvmf_tcp_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3388 struct spdk_nvmf_qpair *qpair) 3389 { 3390 struct spdk_nvmf_tcp_poll_group *tgroup; 3391 struct spdk_nvmf_tcp_qpair *tqpair; 3392 int rc; 3393 3394 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3395 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3396 3397 rc = nvmf_tcp_qpair_sock_init(tqpair); 3398 if (rc != 0) { 3399 SPDK_ERRLOG("Cannot set sock opt for tqpair=%p\n", tqpair); 3400 return -1; 3401 } 3402 3403 rc = nvmf_tcp_qpair_init(&tqpair->qpair); 3404 if (rc < 0) { 3405 SPDK_ERRLOG("Cannot init tqpair=%p\n", tqpair); 3406 return -1; 3407 } 3408 3409 rc = nvmf_tcp_qpair_init_mem_resource(tqpair); 3410 if (rc < 0) { 3411 SPDK_ERRLOG("Cannot init memory resource info for tqpair=%p\n", tqpair); 3412 return -1; 3413 } 3414 3415 rc = spdk_sock_group_add_sock(tgroup->sock_group, tqpair->sock, 3416 nvmf_tcp_sock_cb, tqpair); 3417 if (rc != 0) { 3418 SPDK_ERRLOG("Could not add sock to sock_group: %s (%d)\n", 3419 spdk_strerror(errno), errno); 3420 return -1; 3421 } 3422 3423 tqpair->group = tgroup; 3424 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INVALID); 3425 TAILQ_INSERT_TAIL(&tgroup->qpairs, tqpair, link); 3426 3427 return 0; 3428 } 3429 3430 static int 3431 nvmf_tcp_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3432 struct spdk_nvmf_qpair *qpair) 3433 { 3434 struct spdk_nvmf_tcp_poll_group *tgroup; 3435 struct spdk_nvmf_tcp_qpair *tqpair; 3436 int rc; 3437 3438 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3439 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3440 3441 assert(tqpair->group == tgroup); 3442 3443 SPDK_DEBUGLOG(nvmf_tcp, "remove tqpair=%p from the tgroup=%p\n", tqpair, tgroup); 3444 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 3445 /* Change the state to move the qpair from the await_req list to the main list 3446 * and prevent adding it again later by nvmf_tcp_qpair_set_recv_state() */ 3447 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 3448 } 3449 TAILQ_REMOVE(&tgroup->qpairs, tqpair, link); 3450 3451 /* Try to force out any pending writes */ 3452 spdk_sock_flush(tqpair->sock); 3453 3454 rc = spdk_sock_group_remove_sock(tgroup->sock_group, tqpair->sock); 3455 if (rc != 0) { 3456 SPDK_ERRLOG("Could not remove sock from sock_group: %s (%d)\n", 3457 spdk_strerror(errno), errno); 3458 } 3459 3460 return rc; 3461 } 3462 3463 static int 3464 nvmf_tcp_req_complete(struct spdk_nvmf_request *req) 3465 { 3466 struct spdk_nvmf_tcp_transport *ttransport; 3467 struct spdk_nvmf_tcp_req *tcp_req; 3468 struct spdk_nvmf_tcp_qpair *tqpair; 3469 3470 ttransport = SPDK_CONTAINEROF(req->qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3471 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3472 tqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_tcp_qpair, qpair); 3473 3474 switch (tcp_req->state) { 3475 case TCP_REQUEST_STATE_EXECUTING: 3476 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3477 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3478 break; 3479 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3480 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_ZCOPY_START_COMPLETED); 3481 break; 3482 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3483 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 3484 /* In the interrupt mode it's possible that all responses are already written out over 3485 * the socket but zero-copy buffers are still not released. In that case there won't be 3486 * any event to trigger further socket processing. Send msg to a thread to avoid deadlock. 3487 */ 3488 if (spdk_unlikely(spdk_interrupt_mode_is_enabled() && 3489 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ && 3490 spdk_nvmf_qpair_is_active(&tqpair->qpair))) { 3491 spdk_thread_send_msg(spdk_get_thread(), tcp_sock_cb, tqpair); 3492 } 3493 break; 3494 default: 3495 SPDK_ERRLOG("Unexpected request state %d (cntlid:%d, qid:%d)\n", 3496 tcp_req->state, req->qpair->ctrlr->cntlid, req->qpair->qid); 3497 assert(0 && "Unexpected request state"); 3498 break; 3499 } 3500 3501 nvmf_tcp_req_process(ttransport, tcp_req); 3502 3503 return 0; 3504 } 3505 3506 static void 3507 nvmf_tcp_close_qpair(struct spdk_nvmf_qpair *qpair, 3508 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3509 { 3510 struct spdk_nvmf_tcp_qpair *tqpair; 3511 3512 SPDK_DEBUGLOG(nvmf_tcp, "Qpair: %p\n", qpair); 3513 3514 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3515 3516 assert(tqpair->fini_cb_fn == NULL); 3517 tqpair->fini_cb_fn = cb_fn; 3518 tqpair->fini_cb_arg = cb_arg; 3519 3520 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITED); 3521 nvmf_tcp_qpair_destroy(tqpair); 3522 } 3523 3524 static int 3525 nvmf_tcp_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3526 { 3527 struct spdk_nvmf_tcp_poll_group *tgroup; 3528 int num_events, rc = 0, rc2; 3529 struct spdk_nvmf_tcp_qpair *tqpair, *tqpair_tmp; 3530 3531 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3532 3533 if (spdk_unlikely(TAILQ_EMPTY(&tgroup->qpairs) && TAILQ_EMPTY(&tgroup->await_req))) { 3534 return 0; 3535 } 3536 3537 num_events = spdk_sock_group_poll(tgroup->sock_group); 3538 if (spdk_unlikely(num_events < 0)) { 3539 SPDK_ERRLOG("Failed to poll sock_group=%p\n", tgroup->sock_group); 3540 } 3541 3542 TAILQ_FOREACH_SAFE(tqpair, &tgroup->await_req, link, tqpair_tmp) { 3543 rc2 = nvmf_tcp_sock_process(tqpair); 3544 3545 /* If there was a new socket error, disconnect */ 3546 if (spdk_unlikely(rc2 < 0)) { 3547 nvmf_tcp_qpair_disconnect(tqpair); 3548 if (rc == 0) { 3549 rc = rc2; 3550 } 3551 } 3552 } 3553 3554 return rc == 0 ? num_events : rc; 3555 } 3556 3557 static int 3558 nvmf_tcp_qpair_get_trid(struct spdk_nvmf_qpair *qpair, 3559 struct spdk_nvme_transport_id *trid, bool peer) 3560 { 3561 struct spdk_nvmf_tcp_qpair *tqpair; 3562 uint16_t port; 3563 3564 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3565 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_TCP); 3566 3567 if (peer) { 3568 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->initiator_addr); 3569 port = tqpair->initiator_port; 3570 } else { 3571 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->target_addr); 3572 port = tqpair->target_port; 3573 } 3574 3575 if (spdk_sock_is_ipv4(tqpair->sock)) { 3576 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3577 } else if (spdk_sock_is_ipv6(tqpair->sock)) { 3578 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3579 } else { 3580 return -1; 3581 } 3582 3583 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%d", port); 3584 return 0; 3585 } 3586 3587 static int 3588 nvmf_tcp_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3589 struct spdk_nvme_transport_id *trid) 3590 { 3591 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3592 } 3593 3594 static int 3595 nvmf_tcp_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3596 struct spdk_nvme_transport_id *trid) 3597 { 3598 return nvmf_tcp_qpair_get_trid(qpair, trid, 1); 3599 } 3600 3601 static int 3602 nvmf_tcp_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3603 struct spdk_nvme_transport_id *trid) 3604 { 3605 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3606 } 3607 3608 static void 3609 nvmf_tcp_req_set_abort_status(struct spdk_nvmf_request *req, 3610 struct spdk_nvmf_tcp_req *tcp_req_to_abort) 3611 { 3612 nvmf_tcp_req_set_cpl(tcp_req_to_abort, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_ABORTED_BY_REQUEST); 3613 nvmf_tcp_req_set_state(tcp_req_to_abort, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3614 3615 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 3616 } 3617 3618 static int 3619 _nvmf_tcp_qpair_abort_request(void *ctx) 3620 { 3621 struct spdk_nvmf_request *req = ctx; 3622 struct spdk_nvmf_tcp_req *tcp_req_to_abort = SPDK_CONTAINEROF(req->req_to_abort, 3623 struct spdk_nvmf_tcp_req, req); 3624 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 3625 struct spdk_nvmf_tcp_qpair, qpair); 3626 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 3627 struct spdk_nvmf_tcp_transport, transport); 3628 int rc; 3629 3630 spdk_poller_unregister(&req->poller); 3631 3632 switch (tcp_req_to_abort->state) { 3633 case TCP_REQUEST_STATE_EXECUTING: 3634 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3635 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3636 rc = nvmf_ctrlr_abort_request(req); 3637 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 3638 return SPDK_POLLER_BUSY; 3639 } 3640 break; 3641 3642 case TCP_REQUEST_STATE_NEED_BUFFER: 3643 nvmf_tcp_request_get_buffers_abort(tcp_req_to_abort); 3644 nvmf_tcp_req_set_abort_status(req, tcp_req_to_abort); 3645 nvmf_tcp_req_process(ttransport, tcp_req_to_abort); 3646 break; 3647 3648 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3649 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3650 if (spdk_get_ticks() < req->timeout_tsc) { 3651 req->poller = SPDK_POLLER_REGISTER(_nvmf_tcp_qpair_abort_request, req, 0); 3652 return SPDK_POLLER_BUSY; 3653 } 3654 break; 3655 3656 default: 3657 /* Requests in other states are either un-abortable (e.g. 3658 * TRANSFERRING_CONTROLLER_TO_HOST) or should never end up here, as they're 3659 * immediately transitioned to other states in nvmf_tcp_req_process() (e.g. 3660 * READY_TO_EXECUTE). But it is fine to end up here, as we'll simply complete the 3661 * abort request with the bit0 of dword0 set (command not aborted). 3662 */ 3663 break; 3664 } 3665 3666 spdk_nvmf_request_complete(req); 3667 return SPDK_POLLER_BUSY; 3668 } 3669 3670 static void 3671 nvmf_tcp_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 3672 struct spdk_nvmf_request *req) 3673 { 3674 struct spdk_nvmf_tcp_qpair *tqpair; 3675 struct spdk_nvmf_tcp_transport *ttransport; 3676 struct spdk_nvmf_transport *transport; 3677 uint16_t cid; 3678 uint32_t i; 3679 struct spdk_nvmf_tcp_req *tcp_req_to_abort = NULL; 3680 3681 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3682 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3683 transport = &ttransport->transport; 3684 3685 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 3686 3687 for (i = 0; i < tqpair->resource_count; i++) { 3688 if (tqpair->reqs[i].state != TCP_REQUEST_STATE_FREE && 3689 tqpair->reqs[i].req.cmd->nvme_cmd.cid == cid) { 3690 tcp_req_to_abort = &tqpair->reqs[i]; 3691 break; 3692 } 3693 } 3694 3695 spdk_trace_record(TRACE_TCP_QP_ABORT_REQ, tqpair->qpair.trace_id, 0, (uintptr_t)req); 3696 3697 if (tcp_req_to_abort == NULL) { 3698 spdk_nvmf_request_complete(req); 3699 return; 3700 } 3701 3702 req->req_to_abort = &tcp_req_to_abort->req; 3703 req->timeout_tsc = spdk_get_ticks() + 3704 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 3705 req->poller = NULL; 3706 3707 _nvmf_tcp_qpair_abort_request(req); 3708 } 3709 3710 struct tcp_subsystem_add_host_opts { 3711 char *psk; 3712 }; 3713 3714 static const struct spdk_json_object_decoder tcp_subsystem_add_host_opts_decoder[] = { 3715 {"psk", offsetof(struct tcp_subsystem_add_host_opts, psk), spdk_json_decode_string, true}, 3716 }; 3717 3718 static int 3719 tcp_load_psk(const char *fname, char *buf, size_t bufsz) 3720 { 3721 FILE *psk_file; 3722 struct stat statbuf; 3723 int rc; 3724 3725 if (stat(fname, &statbuf) != 0) { 3726 SPDK_ERRLOG("Could not read permissions for PSK file\n"); 3727 return -EACCES; 3728 } 3729 3730 if ((statbuf.st_mode & TCP_PSK_INVALID_PERMISSIONS) != 0) { 3731 SPDK_ERRLOG("Incorrect permissions for PSK file\n"); 3732 return -EPERM; 3733 } 3734 if ((size_t)statbuf.st_size > bufsz) { 3735 SPDK_ERRLOG("Invalid PSK: too long\n"); 3736 return -EINVAL; 3737 } 3738 psk_file = fopen(fname, "r"); 3739 if (psk_file == NULL) { 3740 SPDK_ERRLOG("Could not open PSK file\n"); 3741 return -EINVAL; 3742 } 3743 3744 rc = fread(buf, 1, statbuf.st_size, psk_file); 3745 if (rc != statbuf.st_size) { 3746 SPDK_ERRLOG("Failed to read PSK\n"); 3747 fclose(psk_file); 3748 return -EINVAL; 3749 } 3750 3751 fclose(psk_file); 3752 return 0; 3753 } 3754 3755 SPDK_LOG_DEPRECATION_REGISTER(nvmf_tcp_psk_path, "PSK path", "v24.09", 0); 3756 3757 static int 3758 nvmf_tcp_subsystem_add_host(struct spdk_nvmf_transport *transport, 3759 const struct spdk_nvmf_subsystem *subsystem, 3760 const char *hostnqn, 3761 const struct spdk_json_val *transport_specific) 3762 { 3763 struct tcp_subsystem_add_host_opts opts; 3764 struct spdk_nvmf_tcp_transport *ttransport; 3765 struct tcp_psk_entry *tmp, *entry = NULL; 3766 uint8_t psk_configured[SPDK_TLS_PSK_MAX_LEN] = {}; 3767 char psk_interchange[SPDK_TLS_PSK_MAX_LEN + 1] = {}; 3768 uint8_t tls_cipher_suite; 3769 int rc = 0; 3770 uint8_t psk_retained_hash; 3771 uint64_t psk_configured_size; 3772 3773 if (transport_specific == NULL) { 3774 return 0; 3775 } 3776 3777 assert(transport != NULL); 3778 assert(subsystem != NULL); 3779 3780 memset(&opts, 0, sizeof(opts)); 3781 3782 /* Decode PSK (either name of a key or file path) */ 3783 if (spdk_json_decode_object_relaxed(transport_specific, tcp_subsystem_add_host_opts_decoder, 3784 SPDK_COUNTOF(tcp_subsystem_add_host_opts_decoder), &opts)) { 3785 SPDK_ERRLOG("spdk_json_decode_object failed\n"); 3786 return -EINVAL; 3787 } 3788 3789 if (opts.psk == NULL) { 3790 return 0; 3791 } 3792 3793 entry = calloc(1, sizeof(struct tcp_psk_entry)); 3794 if (entry == NULL) { 3795 SPDK_ERRLOG("Unable to allocate memory for PSK entry!\n"); 3796 rc = -ENOMEM; 3797 goto end; 3798 } 3799 3800 entry->key = spdk_keyring_get_key(opts.psk); 3801 if (entry->key != NULL) { 3802 rc = spdk_key_get_key(entry->key, psk_interchange, SPDK_TLS_PSK_MAX_LEN); 3803 if (rc < 0) { 3804 SPDK_ERRLOG("Failed to retrieve PSK '%s'\n", opts.psk); 3805 rc = -EINVAL; 3806 goto end; 3807 } 3808 } else { 3809 if (strlen(opts.psk) >= sizeof(entry->psk)) { 3810 SPDK_ERRLOG("PSK path too long\n"); 3811 rc = -EINVAL; 3812 goto end; 3813 } 3814 3815 rc = tcp_load_psk(opts.psk, psk_interchange, SPDK_TLS_PSK_MAX_LEN); 3816 if (rc) { 3817 SPDK_ERRLOG("Could not retrieve PSK from file\n"); 3818 goto end; 3819 } 3820 3821 SPDK_LOG_DEPRECATED(nvmf_tcp_psk_path); 3822 } 3823 3824 /* Parse PSK interchange to get length of base64 encoded data. 3825 * This is then used to decide which cipher suite should be used 3826 * to generate PSK identity and TLS PSK later on. */ 3827 rc = nvme_tcp_parse_interchange_psk(psk_interchange, psk_configured, sizeof(psk_configured), 3828 &psk_configured_size, &psk_retained_hash); 3829 if (rc < 0) { 3830 SPDK_ERRLOG("Failed to parse PSK interchange!\n"); 3831 goto end; 3832 } 3833 3834 /* The Base64 string encodes the configured PSK (32 or 48 bytes binary). 3835 * This check also ensures that psk_configured_size is smaller than 3836 * psk_retained buffer size. */ 3837 if (psk_configured_size == SHA256_DIGEST_LENGTH) { 3838 tls_cipher_suite = NVME_TCP_CIPHER_AES_128_GCM_SHA256; 3839 } else if (psk_configured_size == SHA384_DIGEST_LENGTH) { 3840 tls_cipher_suite = NVME_TCP_CIPHER_AES_256_GCM_SHA384; 3841 } else { 3842 SPDK_ERRLOG("Unrecognized cipher suite!\n"); 3843 rc = -EINVAL; 3844 goto end; 3845 } 3846 3847 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3848 /* Generate PSK identity. */ 3849 rc = nvme_tcp_generate_psk_identity(entry->pskid, sizeof(entry->pskid), hostnqn, 3850 subsystem->subnqn, tls_cipher_suite); 3851 if (rc) { 3852 rc = -EINVAL; 3853 goto end; 3854 } 3855 /* Check if PSK identity entry already exists. */ 3856 TAILQ_FOREACH(tmp, &ttransport->psks, link) { 3857 if (strncmp(tmp->pskid, entry->pskid, NVMF_PSK_IDENTITY_LEN) == 0) { 3858 SPDK_ERRLOG("Given PSK identity: %s entry already exists!\n", entry->pskid); 3859 rc = -EEXIST; 3860 goto end; 3861 } 3862 } 3863 3864 if (snprintf(entry->hostnqn, sizeof(entry->hostnqn), "%s", hostnqn) < 0) { 3865 SPDK_ERRLOG("Could not write hostnqn string!\n"); 3866 rc = -EINVAL; 3867 goto end; 3868 } 3869 if (snprintf(entry->subnqn, sizeof(entry->subnqn), "%s", subsystem->subnqn) < 0) { 3870 SPDK_ERRLOG("Could not write subnqn string!\n"); 3871 rc = -EINVAL; 3872 goto end; 3873 } 3874 3875 entry->tls_cipher_suite = tls_cipher_suite; 3876 3877 /* No hash indicates that Configured PSK must be used as Retained PSK. */ 3878 if (psk_retained_hash == NVME_TCP_HASH_ALGORITHM_NONE) { 3879 /* Psk configured is either 32 or 48 bytes long. */ 3880 memcpy(entry->psk, psk_configured, psk_configured_size); 3881 entry->psk_size = psk_configured_size; 3882 } else { 3883 /* Derive retained PSK. */ 3884 rc = nvme_tcp_derive_retained_psk(psk_configured, psk_configured_size, hostnqn, entry->psk, 3885 SPDK_TLS_PSK_MAX_LEN, psk_retained_hash); 3886 if (rc < 0) { 3887 SPDK_ERRLOG("Unable to derive retained PSK!\n"); 3888 goto end; 3889 } 3890 entry->psk_size = rc; 3891 } 3892 3893 if (entry->key == NULL) { 3894 rc = snprintf(entry->psk_path, sizeof(entry->psk_path), "%s", opts.psk); 3895 if (rc < 0 || (size_t)rc >= sizeof(entry->psk_path)) { 3896 SPDK_ERRLOG("Could not save PSK path!\n"); 3897 rc = -ENAMETOOLONG; 3898 goto end; 3899 } 3900 } 3901 3902 TAILQ_INSERT_TAIL(&ttransport->psks, entry, link); 3903 rc = 0; 3904 3905 end: 3906 spdk_memset_s(psk_configured, sizeof(psk_configured), 0, sizeof(psk_configured)); 3907 spdk_memset_s(psk_interchange, sizeof(psk_interchange), 0, sizeof(psk_interchange)); 3908 3909 free(opts.psk); 3910 if (rc != 0) { 3911 nvmf_tcp_free_psk_entry(entry); 3912 } 3913 3914 return rc; 3915 } 3916 3917 static void 3918 nvmf_tcp_subsystem_remove_host(struct spdk_nvmf_transport *transport, 3919 const struct spdk_nvmf_subsystem *subsystem, 3920 const char *hostnqn) 3921 { 3922 struct spdk_nvmf_tcp_transport *ttransport; 3923 struct tcp_psk_entry *entry, *tmp; 3924 3925 assert(transport != NULL); 3926 assert(subsystem != NULL); 3927 3928 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3929 TAILQ_FOREACH_SAFE(entry, &ttransport->psks, link, tmp) { 3930 if ((strncmp(entry->hostnqn, hostnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0 && 3931 (strncmp(entry->subnqn, subsystem->subnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0) { 3932 TAILQ_REMOVE(&ttransport->psks, entry, link); 3933 nvmf_tcp_free_psk_entry(entry); 3934 break; 3935 } 3936 } 3937 } 3938 3939 static void 3940 nvmf_tcp_subsystem_dump_host(struct spdk_nvmf_transport *transport, 3941 const struct spdk_nvmf_subsystem *subsystem, const char *hostnqn, 3942 struct spdk_json_write_ctx *w) 3943 { 3944 struct spdk_nvmf_tcp_transport *ttransport; 3945 struct tcp_psk_entry *entry; 3946 3947 assert(transport != NULL); 3948 assert(subsystem != NULL); 3949 3950 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3951 TAILQ_FOREACH(entry, &ttransport->psks, link) { 3952 if ((strncmp(entry->hostnqn, hostnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0 && 3953 (strncmp(entry->subnqn, subsystem->subnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0) { 3954 spdk_json_write_named_string(w, "psk", entry->key ? 3955 spdk_key_get_name(entry->key) : entry->psk_path); 3956 break; 3957 } 3958 } 3959 } 3960 3961 static void 3962 nvmf_tcp_opts_init(struct spdk_nvmf_transport_opts *opts) 3963 { 3964 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH; 3965 opts->max_qpairs_per_ctrlr = SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR; 3966 opts->in_capsule_data_size = SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE; 3967 opts->max_io_size = SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE; 3968 opts->io_unit_size = SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE; 3969 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH; 3970 opts->num_shared_buffers = SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS; 3971 opts->buf_cache_size = SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE; 3972 opts->dif_insert_or_strip = SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP; 3973 opts->abort_timeout_sec = SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC; 3974 opts->transport_specific = NULL; 3975 } 3976 3977 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp = { 3978 .name = "TCP", 3979 .type = SPDK_NVME_TRANSPORT_TCP, 3980 .opts_init = nvmf_tcp_opts_init, 3981 .create = nvmf_tcp_create, 3982 .dump_opts = nvmf_tcp_dump_opts, 3983 .destroy = nvmf_tcp_destroy, 3984 3985 .listen = nvmf_tcp_listen, 3986 .stop_listen = nvmf_tcp_stop_listen, 3987 3988 .listener_discover = nvmf_tcp_discover, 3989 3990 .poll_group_create = nvmf_tcp_poll_group_create, 3991 .get_optimal_poll_group = nvmf_tcp_get_optimal_poll_group, 3992 .poll_group_destroy = nvmf_tcp_poll_group_destroy, 3993 .poll_group_add = nvmf_tcp_poll_group_add, 3994 .poll_group_remove = nvmf_tcp_poll_group_remove, 3995 .poll_group_poll = nvmf_tcp_poll_group_poll, 3996 3997 .req_free = nvmf_tcp_req_free, 3998 .req_complete = nvmf_tcp_req_complete, 3999 .req_get_buffers_done = nvmf_tcp_req_get_buffers_done, 4000 4001 .qpair_fini = nvmf_tcp_close_qpair, 4002 .qpair_get_local_trid = nvmf_tcp_qpair_get_local_trid, 4003 .qpair_get_peer_trid = nvmf_tcp_qpair_get_peer_trid, 4004 .qpair_get_listen_trid = nvmf_tcp_qpair_get_listen_trid, 4005 .qpair_abort_request = nvmf_tcp_qpair_abort_request, 4006 .subsystem_add_host = nvmf_tcp_subsystem_add_host, 4007 .subsystem_remove_host = nvmf_tcp_subsystem_remove_host, 4008 .subsystem_dump_host = nvmf_tcp_subsystem_dump_host, 4009 }; 4010 4011 SPDK_NVMF_TRANSPORT_REGISTER(tcp, &spdk_nvmf_transport_tcp); 4012 SPDK_LOG_REGISTER_COMPONENT(nvmf_tcp) 4013