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 void 701 nvmf_tcp_poller_set_interrupt_mode_nop(struct spdk_poller *poller, void *cb_arg, 702 bool interrupt_mode) 703 { 704 } 705 706 static struct spdk_nvmf_transport * 707 nvmf_tcp_create(struct spdk_nvmf_transport_opts *opts) 708 { 709 struct spdk_nvmf_tcp_transport *ttransport; 710 uint32_t sge_count; 711 uint32_t min_shared_buffers; 712 int rc; 713 uint64_t period; 714 715 ttransport = calloc(1, sizeof(*ttransport)); 716 if (!ttransport) { 717 return NULL; 718 } 719 720 TAILQ_INIT(&ttransport->ports); 721 TAILQ_INIT(&ttransport->poll_groups); 722 TAILQ_INIT(&ttransport->psks); 723 724 ttransport->transport.ops = &spdk_nvmf_transport_tcp; 725 726 ttransport->tcp_opts.c2h_success = SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION; 727 ttransport->tcp_opts.sock_priority = SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY; 728 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 729 if (opts->transport_specific != NULL && 730 spdk_json_decode_object_relaxed(opts->transport_specific, tcp_transport_opts_decoder, 731 SPDK_COUNTOF(tcp_transport_opts_decoder), 732 &ttransport->tcp_opts)) { 733 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 734 free(ttransport); 735 return NULL; 736 } 737 738 SPDK_NOTICELOG("*** TCP Transport Init ***\n"); 739 740 SPDK_INFOLOG(nvmf_tcp, "*** TCP Transport Init ***\n" 741 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 742 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 743 " in_capsule_data_size=%d, max_aq_depth=%d\n" 744 " num_shared_buffers=%d, c2h_success=%d,\n" 745 " dif_insert_or_strip=%d, sock_priority=%d\n" 746 " abort_timeout_sec=%d, control_msg_num=%hu\n" 747 " ack_timeout=%d\n", 748 opts->max_queue_depth, 749 opts->max_io_size, 750 opts->max_qpairs_per_ctrlr - 1, 751 opts->io_unit_size, 752 opts->in_capsule_data_size, 753 opts->max_aq_depth, 754 opts->num_shared_buffers, 755 ttransport->tcp_opts.c2h_success, 756 opts->dif_insert_or_strip, 757 ttransport->tcp_opts.sock_priority, 758 opts->abort_timeout_sec, 759 ttransport->tcp_opts.control_msg_num, 760 opts->ack_timeout); 761 762 if (ttransport->tcp_opts.sock_priority > SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY) { 763 SPDK_ERRLOG("Unsupported socket_priority=%d, the current range is: 0 to %d\n" 764 "you can use man 7 socket to view the range of priority under SO_PRIORITY item\n", 765 ttransport->tcp_opts.sock_priority, SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY); 766 free(ttransport); 767 return NULL; 768 } 769 770 if (ttransport->tcp_opts.control_msg_num == 0 && 771 opts->in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 772 SPDK_WARNLOG("TCP param control_msg_num can't be 0 if ICD is less than %u bytes. Using default value %u\n", 773 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM); 774 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 775 } 776 777 /* I/O unit size cannot be larger than max I/O size */ 778 if (opts->io_unit_size > opts->max_io_size) { 779 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", 780 opts->io_unit_size, opts->max_io_size); 781 opts->io_unit_size = opts->max_io_size; 782 } 783 784 /* In capsule data size cannot be larger than max I/O size */ 785 if (opts->in_capsule_data_size > opts->max_io_size) { 786 SPDK_WARNLOG("TCP param ICD size %u can't be larger than max_io_size %u. Using max_io_size as ICD size\n", 787 opts->io_unit_size, opts->max_io_size); 788 opts->in_capsule_data_size = opts->max_io_size; 789 } 790 791 /* max IO queue depth cannot be smaller than 2 or larger than 65535. 792 * We will not check SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH, because max_queue_depth is 16bits and always not larger than 64k. */ 793 if (opts->max_queue_depth < SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH) { 794 SPDK_WARNLOG("TCP param max_queue_depth %u can't be smaller than %u or larger than %u. Using default value %u\n", 795 opts->max_queue_depth, SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH, 796 SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH, SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH); 797 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH; 798 } 799 800 /* max admin queue depth cannot be smaller than 2 or larger than 4096 */ 801 if (opts->max_aq_depth < SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH || 802 opts->max_aq_depth > SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH) { 803 SPDK_WARNLOG("TCP param max_aq_depth %u can't be smaller than %u or larger than %u. Using default value %u\n", 804 opts->max_aq_depth, SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH, 805 SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH, SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH); 806 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH; 807 } 808 809 sge_count = opts->max_io_size / opts->io_unit_size; 810 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 811 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 812 free(ttransport); 813 return NULL; 814 } 815 816 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 817 if (opts->buf_cache_size < UINT32_MAX) { 818 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 819 if (min_shared_buffers > opts->num_shared_buffers) { 820 SPDK_ERRLOG("There are not enough buffers to satisfy " 821 "per-poll group caches for each thread. (%" PRIu32 ") " 822 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 823 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 824 free(ttransport); 825 return NULL; 826 } 827 } 828 829 period = spdk_interrupt_mode_is_enabled() ? 0 : opts->acceptor_poll_rate; 830 ttransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_tcp_accept, &ttransport->transport, period); 831 if (!ttransport->accept_poller) { 832 free(ttransport); 833 return NULL; 834 } 835 836 spdk_poller_register_interrupt(ttransport->accept_poller, nvmf_tcp_poller_set_interrupt_mode_nop, 837 NULL); 838 839 ttransport->listen_sock_group = spdk_sock_group_create(NULL); 840 if (ttransport->listen_sock_group == NULL) { 841 SPDK_ERRLOG("Failed to create socket group for listen sockets\n"); 842 spdk_poller_unregister(&ttransport->accept_poller); 843 free(ttransport); 844 return NULL; 845 } 846 847 if (spdk_interrupt_mode_is_enabled()) { 848 rc = SPDK_SOCK_GROUP_REGISTER_INTERRUPT(ttransport->listen_sock_group, 849 SPDK_INTERRUPT_EVENT_IN | SPDK_INTERRUPT_EVENT_OUT, nvmf_tcp_accept, &ttransport->transport); 850 if (rc != 0) { 851 SPDK_ERRLOG("Failed to register interrupt for listen socker sock group\n"); 852 spdk_sock_group_close(&ttransport->listen_sock_group); 853 spdk_poller_unregister(&ttransport->accept_poller); 854 free(ttransport); 855 return NULL; 856 } 857 } 858 859 return &ttransport->transport; 860 } 861 862 static int 863 nvmf_tcp_trsvcid_to_int(const char *trsvcid) 864 { 865 unsigned long long ull; 866 char *end = NULL; 867 868 ull = strtoull(trsvcid, &end, 10); 869 if (end == NULL || end == trsvcid || *end != '\0') { 870 return -1; 871 } 872 873 /* Valid TCP/IP port numbers are in [1, 65535] */ 874 if (ull == 0 || ull > 65535) { 875 return -1; 876 } 877 878 return (int)ull; 879 } 880 881 /** 882 * Canonicalize a listen address trid. 883 */ 884 static int 885 nvmf_tcp_canon_listen_trid(struct spdk_nvme_transport_id *canon_trid, 886 const struct spdk_nvme_transport_id *trid) 887 { 888 int trsvcid_int; 889 890 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 891 if (trsvcid_int < 0) { 892 return -EINVAL; 893 } 894 895 memset(canon_trid, 0, sizeof(*canon_trid)); 896 spdk_nvme_trid_populate_transport(canon_trid, SPDK_NVME_TRANSPORT_TCP); 897 canon_trid->adrfam = trid->adrfam; 898 snprintf(canon_trid->traddr, sizeof(canon_trid->traddr), "%s", trid->traddr); 899 snprintf(canon_trid->trsvcid, sizeof(canon_trid->trsvcid), "%d", trsvcid_int); 900 901 return 0; 902 } 903 904 /** 905 * Find an existing listening port. 906 */ 907 static struct spdk_nvmf_tcp_port * 908 nvmf_tcp_find_port(struct spdk_nvmf_tcp_transport *ttransport, 909 const struct spdk_nvme_transport_id *trid) 910 { 911 struct spdk_nvme_transport_id canon_trid; 912 struct spdk_nvmf_tcp_port *port; 913 914 if (nvmf_tcp_canon_listen_trid(&canon_trid, trid) != 0) { 915 return NULL; 916 } 917 918 TAILQ_FOREACH(port, &ttransport->ports, link) { 919 if (spdk_nvme_transport_id_compare(&canon_trid, port->trid) == 0) { 920 return port; 921 } 922 } 923 924 return NULL; 925 } 926 927 static int 928 tcp_sock_get_key(uint8_t *out, int out_len, const char **cipher, const char *pskid, 929 void *get_key_ctx) 930 { 931 struct tcp_psk_entry *entry; 932 struct spdk_nvmf_tcp_transport *ttransport = get_key_ctx; 933 size_t psk_len; 934 int rc; 935 936 TAILQ_FOREACH(entry, &ttransport->psks, link) { 937 if (strcmp(pskid, entry->pskid) != 0) { 938 continue; 939 } 940 941 psk_len = entry->psk_size; 942 if ((size_t)out_len < psk_len) { 943 SPDK_ERRLOG("Out buffer of size: %" PRIu32 " cannot fit PSK of len: %lu\n", 944 out_len, psk_len); 945 return -ENOBUFS; 946 } 947 948 /* Convert PSK to the TLS PSK format. */ 949 rc = nvme_tcp_derive_tls_psk(entry->psk, psk_len, pskid, out, out_len, 950 entry->tls_cipher_suite); 951 if (rc < 0) { 952 SPDK_ERRLOG("Could not generate TLS PSK\n"); 953 } 954 955 switch (entry->tls_cipher_suite) { 956 case NVME_TCP_CIPHER_AES_128_GCM_SHA256: 957 *cipher = "TLS_AES_128_GCM_SHA256"; 958 break; 959 case NVME_TCP_CIPHER_AES_256_GCM_SHA384: 960 *cipher = "TLS_AES_256_GCM_SHA384"; 961 break; 962 default: 963 *cipher = NULL; 964 return -ENOTSUP; 965 } 966 967 return rc; 968 } 969 970 SPDK_ERRLOG("Could not find PSK for identity: %s\n", pskid); 971 972 return -EINVAL; 973 } 974 975 static int 976 nvmf_tcp_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 977 struct spdk_nvmf_listen_opts *listen_opts) 978 { 979 struct spdk_nvmf_tcp_transport *ttransport; 980 struct spdk_nvmf_tcp_port *port; 981 int trsvcid_int; 982 uint8_t adrfam; 983 const char *sock_impl_name; 984 struct spdk_sock_impl_opts impl_opts; 985 size_t impl_opts_size = sizeof(impl_opts); 986 struct spdk_sock_opts opts; 987 int rc; 988 989 if (!strlen(trid->trsvcid)) { 990 SPDK_ERRLOG("Service id is required\n"); 991 return -EINVAL; 992 } 993 994 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 995 996 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 997 if (trsvcid_int < 0) { 998 SPDK_ERRLOG("Invalid trsvcid '%s'\n", trid->trsvcid); 999 return -EINVAL; 1000 } 1001 1002 port = calloc(1, sizeof(*port)); 1003 if (!port) { 1004 SPDK_ERRLOG("Port allocation failed\n"); 1005 return -ENOMEM; 1006 } 1007 1008 port->trid = trid; 1009 1010 sock_impl_name = NULL; 1011 1012 opts.opts_size = sizeof(opts); 1013 spdk_sock_get_default_opts(&opts); 1014 opts.priority = ttransport->tcp_opts.sock_priority; 1015 opts.ack_timeout = transport->opts.ack_timeout; 1016 if (listen_opts->secure_channel) { 1017 if (listen_opts->sock_impl && 1018 strncmp("ssl", listen_opts->sock_impl, strlen(listen_opts->sock_impl))) { 1019 SPDK_ERRLOG("Enabling secure_channel while specifying a sock_impl different from 'ssl' is unsupported"); 1020 free(port); 1021 return -EINVAL; 1022 } 1023 listen_opts->sock_impl = "ssl"; 1024 } 1025 1026 if (listen_opts->sock_impl) { 1027 sock_impl_name = listen_opts->sock_impl; 1028 spdk_sock_impl_get_opts(sock_impl_name, &impl_opts, &impl_opts_size); 1029 1030 if (!strncmp("ssl", sock_impl_name, strlen(sock_impl_name))) { 1031 if (!g_tls_log) { 1032 SPDK_NOTICELOG("TLS support is considered experimental\n"); 1033 g_tls_log = true; 1034 } 1035 impl_opts.tls_version = SPDK_TLS_VERSION_1_3; 1036 impl_opts.get_key = tcp_sock_get_key; 1037 impl_opts.get_key_ctx = ttransport; 1038 impl_opts.tls_cipher_suites = "TLS_AES_256_GCM_SHA384:TLS_AES_128_GCM_SHA256"; 1039 } 1040 1041 opts.impl_opts = &impl_opts; 1042 opts.impl_opts_size = sizeof(impl_opts); 1043 } 1044 1045 port->listen_sock = spdk_sock_listen_ext(trid->traddr, trsvcid_int, 1046 sock_impl_name, &opts); 1047 if (port->listen_sock == NULL) { 1048 SPDK_ERRLOG("spdk_sock_listen(%s, %d) failed: %s (%d)\n", 1049 trid->traddr, trsvcid_int, 1050 spdk_strerror(errno), errno); 1051 free(port); 1052 return -errno; 1053 } 1054 1055 if (spdk_sock_is_ipv4(port->listen_sock)) { 1056 adrfam = SPDK_NVMF_ADRFAM_IPV4; 1057 } else if (spdk_sock_is_ipv6(port->listen_sock)) { 1058 adrfam = SPDK_NVMF_ADRFAM_IPV6; 1059 } else { 1060 SPDK_ERRLOG("Unhandled socket type\n"); 1061 adrfam = 0; 1062 } 1063 1064 if (adrfam != trid->adrfam) { 1065 SPDK_ERRLOG("Socket address family mismatch\n"); 1066 spdk_sock_close(&port->listen_sock); 1067 free(port); 1068 return -EINVAL; 1069 } 1070 1071 rc = spdk_sock_group_add_sock(ttransport->listen_sock_group, port->listen_sock, nvmf_tcp_accept_cb, 1072 port); 1073 if (rc < 0) { 1074 SPDK_ERRLOG("Failed to add socket to the listen socket group\n"); 1075 spdk_sock_close(&port->listen_sock); 1076 free(port); 1077 return -errno; 1078 } 1079 1080 port->transport = transport; 1081 1082 SPDK_NOTICELOG("*** NVMe/TCP Target Listening on %s port %s ***\n", 1083 trid->traddr, trid->trsvcid); 1084 1085 TAILQ_INSERT_TAIL(&ttransport->ports, port, link); 1086 return 0; 1087 } 1088 1089 static void 1090 nvmf_tcp_stop_listen(struct spdk_nvmf_transport *transport, 1091 const struct spdk_nvme_transport_id *trid) 1092 { 1093 struct spdk_nvmf_tcp_transport *ttransport; 1094 struct spdk_nvmf_tcp_port *port; 1095 1096 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1097 1098 SPDK_DEBUGLOG(nvmf_tcp, "Removing listen address %s port %s\n", 1099 trid->traddr, trid->trsvcid); 1100 1101 port = nvmf_tcp_find_port(ttransport, trid); 1102 if (port) { 1103 spdk_sock_group_remove_sock(ttransport->listen_sock_group, port->listen_sock); 1104 TAILQ_REMOVE(&ttransport->ports, port, link); 1105 spdk_sock_close(&port->listen_sock); 1106 free(port); 1107 } 1108 } 1109 1110 static void nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 1111 enum nvme_tcp_pdu_recv_state state); 1112 1113 static void 1114 nvmf_tcp_qpair_set_state(struct spdk_nvmf_tcp_qpair *tqpair, enum nvme_tcp_qpair_state state) 1115 { 1116 tqpair->state = state; 1117 spdk_trace_record(TRACE_TCP_QP_STATE_CHANGE, tqpair->qpair.trace_id, 0, 0, 1118 (uint64_t)tqpair->state); 1119 } 1120 1121 static void 1122 nvmf_tcp_qpair_disconnect(struct spdk_nvmf_tcp_qpair *tqpair) 1123 { 1124 SPDK_DEBUGLOG(nvmf_tcp, "Disconnecting qpair %p\n", tqpair); 1125 1126 spdk_trace_record(TRACE_TCP_QP_DISCONNECT, tqpair->qpair.trace_id, 0, 0); 1127 1128 if (tqpair->state <= NVME_TCP_QPAIR_STATE_RUNNING) { 1129 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITING); 1130 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1131 spdk_poller_unregister(&tqpair->timeout_poller); 1132 1133 /* This will end up calling nvmf_tcp_close_qpair */ 1134 spdk_nvmf_qpair_disconnect(&tqpair->qpair); 1135 } 1136 } 1137 1138 static void 1139 _mgmt_pdu_write_done(void *_tqpair, int err) 1140 { 1141 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 1142 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1143 1144 if (spdk_unlikely(err != 0)) { 1145 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1146 return; 1147 } 1148 1149 assert(pdu->cb_fn != NULL); 1150 pdu->cb_fn(pdu->cb_arg); 1151 } 1152 1153 static void 1154 _req_pdu_write_done(void *req, int err) 1155 { 1156 struct spdk_nvmf_tcp_req *tcp_req = req; 1157 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1158 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1159 1160 assert(tcp_req->pdu_in_use); 1161 tcp_req->pdu_in_use = false; 1162 1163 /* If the request is in a completed state, we're waiting for write completion to free it */ 1164 if (spdk_unlikely(tcp_req->state == TCP_REQUEST_STATE_COMPLETED)) { 1165 nvmf_tcp_request_free(tcp_req); 1166 return; 1167 } 1168 1169 if (spdk_unlikely(err != 0)) { 1170 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1171 return; 1172 } 1173 1174 assert(pdu->cb_fn != NULL); 1175 pdu->cb_fn(pdu->cb_arg); 1176 } 1177 1178 static void 1179 _pdu_write_done(struct nvme_tcp_pdu *pdu, int err) 1180 { 1181 pdu->sock_req.cb_fn(pdu->sock_req.cb_arg, err); 1182 } 1183 1184 static void 1185 _tcp_write_pdu(struct nvme_tcp_pdu *pdu) 1186 { 1187 int rc; 1188 uint32_t mapped_length; 1189 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1190 1191 pdu->sock_req.iovcnt = nvme_tcp_build_iovs(pdu->iov, SPDK_COUNTOF(pdu->iov), pdu, 1192 tqpair->host_hdgst_enable, tqpair->host_ddgst_enable, &mapped_length); 1193 spdk_sock_writev_async(tqpair->sock, &pdu->sock_req); 1194 1195 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP || 1196 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ) { 1197 /* Try to force the send immediately. */ 1198 rc = spdk_sock_flush(tqpair->sock); 1199 if (rc > 0 && (uint32_t)rc == mapped_length) { 1200 _pdu_write_done(pdu, 0); 1201 } else { 1202 SPDK_ERRLOG("Could not write %s to socket: rc=%d, errno=%d\n", 1203 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP ? 1204 "IC_RESP" : "TERM_REQ", rc, errno); 1205 _pdu_write_done(pdu, rc >= 0 ? -EAGAIN : -errno); 1206 } 1207 } 1208 } 1209 1210 static void 1211 data_crc32_accel_done(void *cb_arg, int status) 1212 { 1213 struct nvme_tcp_pdu *pdu = cb_arg; 1214 1215 if (spdk_unlikely(status)) { 1216 SPDK_ERRLOG("Failed to compute the data digest for pdu =%p\n", pdu); 1217 _pdu_write_done(pdu, status); 1218 return; 1219 } 1220 1221 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 1222 MAKE_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32); 1223 1224 _tcp_write_pdu(pdu); 1225 } 1226 1227 static void 1228 pdu_data_crc32_compute(struct nvme_tcp_pdu *pdu) 1229 { 1230 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1231 int rc = 0; 1232 1233 /* Data Digest */ 1234 if (pdu->data_len > 0 && g_nvme_tcp_ddgst[pdu->hdr.common.pdu_type] && tqpair->host_ddgst_enable) { 1235 /* Only support this limitated case for the first step */ 1236 if (spdk_likely(!pdu->dif_ctx && (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0) 1237 && tqpair->group)) { 1238 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 1239 pdu->data_iovcnt, 0, data_crc32_accel_done, pdu); 1240 if (spdk_likely(rc == 0)) { 1241 return; 1242 } 1243 } else { 1244 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 1245 } 1246 data_crc32_accel_done(pdu, rc); 1247 } else { 1248 _tcp_write_pdu(pdu); 1249 } 1250 } 1251 1252 static void 1253 nvmf_tcp_qpair_write_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1254 struct nvme_tcp_pdu *pdu, 1255 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1256 void *cb_arg) 1257 { 1258 int hlen; 1259 uint32_t crc32c; 1260 1261 assert(tqpair->pdu_in_progress != pdu); 1262 1263 hlen = pdu->hdr.common.hlen; 1264 pdu->cb_fn = cb_fn; 1265 pdu->cb_arg = cb_arg; 1266 1267 pdu->iov[0].iov_base = &pdu->hdr.raw; 1268 pdu->iov[0].iov_len = hlen; 1269 1270 /* Header Digest */ 1271 if (g_nvme_tcp_hdgst[pdu->hdr.common.pdu_type] && tqpair->host_hdgst_enable) { 1272 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 1273 MAKE_DIGEST_WORD((uint8_t *)pdu->hdr.raw + hlen, crc32c); 1274 } 1275 1276 /* Data Digest */ 1277 pdu_data_crc32_compute(pdu); 1278 } 1279 1280 static void 1281 nvmf_tcp_qpair_write_mgmt_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1282 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1283 void *cb_arg) 1284 { 1285 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1286 1287 pdu->sock_req.cb_fn = _mgmt_pdu_write_done; 1288 pdu->sock_req.cb_arg = tqpair; 1289 1290 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1291 } 1292 1293 static void 1294 nvmf_tcp_qpair_write_req_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1295 struct spdk_nvmf_tcp_req *tcp_req, 1296 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1297 void *cb_arg) 1298 { 1299 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1300 1301 pdu->sock_req.cb_fn = _req_pdu_write_done; 1302 pdu->sock_req.cb_arg = tcp_req; 1303 1304 assert(!tcp_req->pdu_in_use); 1305 tcp_req->pdu_in_use = true; 1306 1307 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1308 } 1309 1310 static int 1311 nvmf_tcp_qpair_init_mem_resource(struct spdk_nvmf_tcp_qpair *tqpair) 1312 { 1313 uint32_t i; 1314 struct spdk_nvmf_transport_opts *opts; 1315 uint32_t in_capsule_data_size; 1316 1317 opts = &tqpair->qpair.transport->opts; 1318 1319 in_capsule_data_size = opts->in_capsule_data_size; 1320 if (opts->dif_insert_or_strip) { 1321 in_capsule_data_size = SPDK_BDEV_BUF_SIZE_WITH_MD(in_capsule_data_size); 1322 } 1323 1324 tqpair->resource_count = opts->max_queue_depth; 1325 1326 tqpair->reqs = calloc(tqpair->resource_count, sizeof(*tqpair->reqs)); 1327 if (!tqpair->reqs) { 1328 SPDK_ERRLOG("Unable to allocate reqs on tqpair=%p\n", tqpair); 1329 return -1; 1330 } 1331 1332 if (in_capsule_data_size) { 1333 tqpair->bufs = spdk_zmalloc(tqpair->resource_count * in_capsule_data_size, 0x1000, 1334 NULL, SPDK_ENV_LCORE_ID_ANY, 1335 SPDK_MALLOC_DMA); 1336 if (!tqpair->bufs) { 1337 SPDK_ERRLOG("Unable to allocate bufs on tqpair=%p.\n", tqpair); 1338 return -1; 1339 } 1340 } 1341 /* prepare memory space for receiving pdus and tcp_req */ 1342 /* Add additional 1 member, which will be used for mgmt_pdu owned by the tqpair */ 1343 tqpair->pdus = spdk_dma_zmalloc((2 * tqpair->resource_count + 1) * sizeof(*tqpair->pdus), 0x1000, 1344 NULL); 1345 if (!tqpair->pdus) { 1346 SPDK_ERRLOG("Unable to allocate pdu pool on tqpair =%p.\n", tqpair); 1347 return -1; 1348 } 1349 1350 for (i = 0; i < tqpair->resource_count; i++) { 1351 struct spdk_nvmf_tcp_req *tcp_req = &tqpair->reqs[i]; 1352 1353 tcp_req->ttag = i + 1; 1354 tcp_req->req.qpair = &tqpair->qpair; 1355 1356 tcp_req->pdu = &tqpair->pdus[i]; 1357 tcp_req->pdu->qpair = tqpair; 1358 1359 /* Set up memory to receive commands */ 1360 if (tqpair->bufs) { 1361 tcp_req->buf = (void *)((uintptr_t)tqpair->bufs + (i * in_capsule_data_size)); 1362 } 1363 1364 /* Set the cmdn and rsp */ 1365 tcp_req->req.rsp = (union nvmf_c2h_msg *)&tcp_req->rsp; 1366 tcp_req->req.cmd = (union nvmf_h2c_msg *)&tcp_req->cmd; 1367 1368 tcp_req->req.stripped_data = NULL; 1369 1370 /* Initialize request state to FREE */ 1371 tcp_req->state = TCP_REQUEST_STATE_FREE; 1372 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 1373 tqpair->state_cntr[TCP_REQUEST_STATE_FREE]++; 1374 } 1375 1376 for (; i < 2 * tqpair->resource_count; i++) { 1377 struct nvme_tcp_pdu *pdu = &tqpair->pdus[i]; 1378 1379 pdu->qpair = tqpair; 1380 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 1381 } 1382 1383 tqpair->mgmt_pdu = &tqpair->pdus[i]; 1384 tqpair->mgmt_pdu->qpair = tqpair; 1385 tqpair->pdu_in_progress = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 1386 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 1387 tqpair->tcp_pdu_working_count = 1; 1388 1389 tqpair->recv_buf_size = (in_capsule_data_size + sizeof(struct spdk_nvme_tcp_cmd) + 2 * 1390 SPDK_NVME_TCP_DIGEST_LEN) * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1391 1392 return 0; 1393 } 1394 1395 static int 1396 nvmf_tcp_qpair_init(struct spdk_nvmf_qpair *qpair) 1397 { 1398 struct spdk_nvmf_tcp_qpair *tqpair; 1399 1400 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1401 1402 SPDK_DEBUGLOG(nvmf_tcp, "New TCP Connection: %p\n", qpair); 1403 1404 spdk_trace_record(TRACE_TCP_QP_CREATE, tqpair->qpair.trace_id, 0, 0); 1405 1406 /* Initialise request state queues of the qpair */ 1407 TAILQ_INIT(&tqpair->tcp_req_free_queue); 1408 TAILQ_INIT(&tqpair->tcp_req_working_queue); 1409 SLIST_INIT(&tqpair->tcp_pdu_free_queue); 1410 tqpair->qpair.queue_depth = 0; 1411 1412 tqpair->host_hdgst_enable = true; 1413 tqpair->host_ddgst_enable = true; 1414 1415 return 0; 1416 } 1417 1418 static int 1419 nvmf_tcp_qpair_sock_init(struct spdk_nvmf_tcp_qpair *tqpair) 1420 { 1421 char saddr[32], caddr[32]; 1422 uint16_t sport, cport; 1423 char owner[256]; 1424 int rc; 1425 1426 rc = spdk_sock_getaddr(tqpair->sock, saddr, sizeof(saddr), &sport, 1427 caddr, sizeof(caddr), &cport); 1428 if (rc != 0) { 1429 SPDK_ERRLOG("spdk_sock_getaddr() failed\n"); 1430 return rc; 1431 } 1432 snprintf(owner, sizeof(owner), "%s:%d", caddr, cport); 1433 tqpair->qpair.trace_id = spdk_trace_register_owner(OWNER_TYPE_NVMF_TCP, owner); 1434 spdk_trace_record(TRACE_TCP_QP_SOCK_INIT, tqpair->qpair.trace_id, 0, 0); 1435 1436 /* set low water mark */ 1437 rc = spdk_sock_set_recvlowat(tqpair->sock, 1); 1438 if (rc != 0) { 1439 SPDK_ERRLOG("spdk_sock_set_recvlowat() failed\n"); 1440 return rc; 1441 } 1442 1443 return 0; 1444 } 1445 1446 static void 1447 nvmf_tcp_handle_connect(struct spdk_nvmf_tcp_port *port, struct spdk_sock *sock) 1448 { 1449 struct spdk_nvmf_tcp_qpair *tqpair; 1450 int rc; 1451 1452 SPDK_DEBUGLOG(nvmf_tcp, "New connection accepted on %s port %s\n", 1453 port->trid->traddr, port->trid->trsvcid); 1454 1455 tqpair = calloc(1, sizeof(struct spdk_nvmf_tcp_qpair)); 1456 if (tqpair == NULL) { 1457 SPDK_ERRLOG("Could not allocate new connection.\n"); 1458 spdk_sock_close(&sock); 1459 return; 1460 } 1461 1462 tqpair->sock = sock; 1463 tqpair->state_cntr[TCP_REQUEST_STATE_FREE] = 0; 1464 tqpair->port = port; 1465 tqpair->qpair.transport = port->transport; 1466 1467 rc = spdk_sock_getaddr(tqpair->sock, tqpair->target_addr, 1468 sizeof(tqpair->target_addr), &tqpair->target_port, 1469 tqpair->initiator_addr, sizeof(tqpair->initiator_addr), 1470 &tqpair->initiator_port); 1471 if (rc < 0) { 1472 SPDK_ERRLOG("spdk_sock_getaddr() failed of tqpair=%p\n", tqpair); 1473 nvmf_tcp_qpair_destroy(tqpair); 1474 return; 1475 } 1476 1477 spdk_nvmf_tgt_new_qpair(port->transport->tgt, &tqpair->qpair); 1478 } 1479 1480 static uint32_t 1481 nvmf_tcp_port_accept(struct spdk_nvmf_tcp_port *port) 1482 { 1483 struct spdk_sock *sock; 1484 uint32_t count = 0; 1485 int i; 1486 1487 for (i = 0; i < NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME; i++) { 1488 sock = spdk_sock_accept(port->listen_sock); 1489 if (sock == NULL) { 1490 break; 1491 } 1492 count++; 1493 nvmf_tcp_handle_connect(port, sock); 1494 } 1495 1496 return count; 1497 } 1498 1499 static int 1500 nvmf_tcp_accept(void *ctx) 1501 { 1502 struct spdk_nvmf_transport *transport = ctx; 1503 struct spdk_nvmf_tcp_transport *ttransport; 1504 int count; 1505 1506 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1507 1508 count = spdk_sock_group_poll(ttransport->listen_sock_group); 1509 if (count < 0) { 1510 SPDK_ERRLOG("Fail in TCP listen socket group poll\n"); 1511 } 1512 1513 return count != 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 1514 } 1515 1516 static void 1517 nvmf_tcp_accept_cb(void *ctx, struct spdk_sock_group *group, struct spdk_sock *sock) 1518 { 1519 struct spdk_nvmf_tcp_port *port = ctx; 1520 1521 nvmf_tcp_port_accept(port); 1522 } 1523 1524 static void 1525 nvmf_tcp_discover(struct spdk_nvmf_transport *transport, 1526 struct spdk_nvme_transport_id *trid, 1527 struct spdk_nvmf_discovery_log_page_entry *entry) 1528 { 1529 struct spdk_nvmf_tcp_port *port; 1530 struct spdk_nvmf_tcp_transport *ttransport; 1531 1532 entry->trtype = SPDK_NVMF_TRTYPE_TCP; 1533 entry->adrfam = trid->adrfam; 1534 1535 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 1536 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 1537 1538 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1539 port = nvmf_tcp_find_port(ttransport, trid); 1540 1541 assert(port != NULL); 1542 1543 if (strcmp(spdk_sock_get_impl_name(port->listen_sock), "ssl") == 0) { 1544 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_REQUIRED; 1545 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_TLS_1_3; 1546 } else { 1547 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 1548 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_NONE; 1549 } 1550 } 1551 1552 static struct spdk_nvmf_tcp_control_msg_list * 1553 nvmf_tcp_control_msg_list_create(uint16_t num_messages) 1554 { 1555 struct spdk_nvmf_tcp_control_msg_list *list; 1556 struct spdk_nvmf_tcp_control_msg *msg; 1557 uint16_t i; 1558 1559 list = calloc(1, sizeof(*list)); 1560 if (!list) { 1561 SPDK_ERRLOG("Failed to allocate memory for list structure\n"); 1562 return NULL; 1563 } 1564 1565 list->msg_buf = spdk_zmalloc(num_messages * SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, 1566 NVMF_DATA_BUFFER_ALIGNMENT, NULL, SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA); 1567 if (!list->msg_buf) { 1568 SPDK_ERRLOG("Failed to allocate memory for control message buffers\n"); 1569 free(list); 1570 return NULL; 1571 } 1572 1573 STAILQ_INIT(&list->free_msgs); 1574 STAILQ_INIT(&list->waiting_for_msg_reqs); 1575 1576 for (i = 0; i < num_messages; i++) { 1577 msg = (struct spdk_nvmf_tcp_control_msg *)((char *)list->msg_buf + i * 1578 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1579 STAILQ_INSERT_TAIL(&list->free_msgs, msg, link); 1580 } 1581 1582 return list; 1583 } 1584 1585 static void 1586 nvmf_tcp_control_msg_list_free(struct spdk_nvmf_tcp_control_msg_list *list) 1587 { 1588 if (!list) { 1589 return; 1590 } 1591 1592 spdk_free(list->msg_buf); 1593 free(list); 1594 } 1595 1596 static int nvmf_tcp_poll_group_poll(struct spdk_nvmf_transport_poll_group *group); 1597 1598 static int 1599 nvmf_tcp_poll_group_intr(void *ctx) 1600 { 1601 struct spdk_nvmf_transport_poll_group *group = ctx; 1602 int ret = 0; 1603 1604 ret = nvmf_tcp_poll_group_poll(group); 1605 1606 return ret != 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 1607 } 1608 1609 static struct spdk_nvmf_transport_poll_group * 1610 nvmf_tcp_poll_group_create(struct spdk_nvmf_transport *transport, 1611 struct spdk_nvmf_poll_group *group) 1612 { 1613 struct spdk_nvmf_tcp_transport *ttransport; 1614 struct spdk_nvmf_tcp_poll_group *tgroup; 1615 int rc; 1616 1617 tgroup = calloc(1, sizeof(*tgroup)); 1618 if (!tgroup) { 1619 return NULL; 1620 } 1621 1622 tgroup->sock_group = spdk_sock_group_create(&tgroup->group); 1623 if (!tgroup->sock_group) { 1624 goto cleanup; 1625 } 1626 1627 TAILQ_INIT(&tgroup->qpairs); 1628 TAILQ_INIT(&tgroup->await_req); 1629 1630 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1631 1632 if (transport->opts.in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 1633 SPDK_DEBUGLOG(nvmf_tcp, "ICD %u is less than min required for admin/fabric commands (%u). " 1634 "Creating control messages list\n", transport->opts.in_capsule_data_size, 1635 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1636 tgroup->control_msg_list = nvmf_tcp_control_msg_list_create(ttransport->tcp_opts.control_msg_num); 1637 if (!tgroup->control_msg_list) { 1638 goto cleanup; 1639 } 1640 } 1641 1642 tgroup->accel_channel = spdk_accel_get_io_channel(); 1643 if (spdk_unlikely(!tgroup->accel_channel)) { 1644 SPDK_ERRLOG("Cannot create accel_channel for tgroup=%p\n", tgroup); 1645 goto cleanup; 1646 } 1647 1648 TAILQ_INSERT_TAIL(&ttransport->poll_groups, tgroup, link); 1649 if (ttransport->next_pg == NULL) { 1650 ttransport->next_pg = tgroup; 1651 } 1652 1653 if (spdk_interrupt_mode_is_enabled()) { 1654 rc = SPDK_SOCK_GROUP_REGISTER_INTERRUPT(tgroup->sock_group, 1655 SPDK_INTERRUPT_EVENT_IN | SPDK_INTERRUPT_EVENT_OUT, nvmf_tcp_poll_group_intr, &tgroup->group); 1656 if (rc != 0) { 1657 SPDK_ERRLOG("Failed to register interrupt for sock group\n"); 1658 goto cleanup; 1659 } 1660 } 1661 1662 return &tgroup->group; 1663 1664 cleanup: 1665 nvmf_tcp_poll_group_destroy(&tgroup->group); 1666 return NULL; 1667 } 1668 1669 static struct spdk_nvmf_transport_poll_group * 1670 nvmf_tcp_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 1671 { 1672 struct spdk_nvmf_tcp_transport *ttransport; 1673 struct spdk_nvmf_tcp_poll_group **pg; 1674 struct spdk_nvmf_tcp_qpair *tqpair; 1675 struct spdk_sock_group *group = NULL, *hint = NULL; 1676 int rc; 1677 1678 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 1679 1680 if (TAILQ_EMPTY(&ttransport->poll_groups)) { 1681 return NULL; 1682 } 1683 1684 pg = &ttransport->next_pg; 1685 assert(*pg != NULL); 1686 hint = (*pg)->sock_group; 1687 1688 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1689 rc = spdk_sock_get_optimal_sock_group(tqpair->sock, &group, hint); 1690 if (rc != 0) { 1691 return NULL; 1692 } else if (group != NULL) { 1693 /* Optimal poll group was found */ 1694 return spdk_sock_group_get_ctx(group); 1695 } 1696 1697 /* The hint was used for optimal poll group, advance next_pg. */ 1698 *pg = TAILQ_NEXT(*pg, link); 1699 if (*pg == NULL) { 1700 *pg = TAILQ_FIRST(&ttransport->poll_groups); 1701 } 1702 1703 return spdk_sock_group_get_ctx(hint); 1704 } 1705 1706 static void 1707 nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 1708 { 1709 struct spdk_nvmf_tcp_poll_group *tgroup, *next_tgroup; 1710 struct spdk_nvmf_tcp_transport *ttransport; 1711 1712 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 1713 spdk_sock_group_unregister_interrupt(tgroup->sock_group); 1714 spdk_sock_group_close(&tgroup->sock_group); 1715 if (tgroup->control_msg_list) { 1716 nvmf_tcp_control_msg_list_free(tgroup->control_msg_list); 1717 } 1718 1719 if (tgroup->accel_channel) { 1720 spdk_put_io_channel(tgroup->accel_channel); 1721 } 1722 1723 if (tgroup->group.transport == NULL) { 1724 /* Transport can be NULL when nvmf_tcp_poll_group_create() 1725 * calls this function directly in a failure path. */ 1726 free(tgroup); 1727 return; 1728 } 1729 1730 ttransport = SPDK_CONTAINEROF(tgroup->group.transport, struct spdk_nvmf_tcp_transport, transport); 1731 1732 next_tgroup = TAILQ_NEXT(tgroup, link); 1733 TAILQ_REMOVE(&ttransport->poll_groups, tgroup, link); 1734 if (next_tgroup == NULL) { 1735 next_tgroup = TAILQ_FIRST(&ttransport->poll_groups); 1736 } 1737 if (ttransport->next_pg == tgroup) { 1738 ttransport->next_pg = next_tgroup; 1739 } 1740 1741 free(tgroup); 1742 } 1743 1744 static void 1745 nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 1746 enum nvme_tcp_pdu_recv_state state) 1747 { 1748 if (tqpair->recv_state == state) { 1749 SPDK_ERRLOG("The recv state of tqpair=%p is same with the state(%d) to be set\n", 1750 tqpair, state); 1751 return; 1752 } 1753 1754 if (spdk_unlikely(state == NVME_TCP_PDU_RECV_STATE_QUIESCING)) { 1755 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH && tqpair->pdu_in_progress) { 1756 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, tqpair->pdu_in_progress, slist); 1757 tqpair->tcp_pdu_working_count--; 1758 } 1759 } 1760 1761 if (spdk_unlikely(state == NVME_TCP_PDU_RECV_STATE_ERROR)) { 1762 assert(tqpair->tcp_pdu_working_count == 0); 1763 } 1764 1765 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1766 /* When leaving the await req state, move the qpair to the main list */ 1767 TAILQ_REMOVE(&tqpair->group->await_req, tqpair, link); 1768 TAILQ_INSERT_TAIL(&tqpair->group->qpairs, tqpair, link); 1769 } else if (state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1770 TAILQ_REMOVE(&tqpair->group->qpairs, tqpair, link); 1771 TAILQ_INSERT_TAIL(&tqpair->group->await_req, tqpair, link); 1772 } 1773 1774 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv state=%d\n", tqpair, state); 1775 tqpair->recv_state = state; 1776 1777 spdk_trace_record(TRACE_TCP_QP_RCV_STATE_CHANGE, tqpair->qpair.trace_id, 0, 0, 1778 (uint64_t)tqpair->recv_state); 1779 } 1780 1781 static int 1782 nvmf_tcp_qpair_handle_timeout(void *ctx) 1783 { 1784 struct spdk_nvmf_tcp_qpair *tqpair = ctx; 1785 1786 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1787 1788 SPDK_ERRLOG("No pdu coming for tqpair=%p within %d seconds\n", tqpair, 1789 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT); 1790 1791 nvmf_tcp_qpair_disconnect(tqpair); 1792 return SPDK_POLLER_BUSY; 1793 } 1794 1795 static void 1796 nvmf_tcp_send_c2h_term_req_complete(void *cb_arg) 1797 { 1798 struct spdk_nvmf_tcp_qpair *tqpair = (struct spdk_nvmf_tcp_qpair *)cb_arg; 1799 1800 if (!tqpair->timeout_poller) { 1801 tqpair->timeout_poller = SPDK_POLLER_REGISTER(nvmf_tcp_qpair_handle_timeout, tqpair, 1802 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT * 1000000); 1803 } 1804 } 1805 1806 static void 1807 nvmf_tcp_send_c2h_term_req(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu, 1808 enum spdk_nvme_tcp_term_req_fes fes, uint32_t error_offset) 1809 { 1810 struct nvme_tcp_pdu *rsp_pdu; 1811 struct spdk_nvme_tcp_term_req_hdr *c2h_term_req; 1812 uint32_t c2h_term_req_hdr_len = sizeof(*c2h_term_req); 1813 uint32_t copy_len; 1814 1815 rsp_pdu = tqpair->mgmt_pdu; 1816 1817 c2h_term_req = &rsp_pdu->hdr.term_req; 1818 c2h_term_req->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ; 1819 c2h_term_req->common.hlen = c2h_term_req_hdr_len; 1820 c2h_term_req->fes = fes; 1821 1822 if ((fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1823 (fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1824 DSET32(&c2h_term_req->fei, error_offset); 1825 } 1826 1827 copy_len = spdk_min(pdu->hdr.common.hlen, SPDK_NVME_TCP_TERM_REQ_ERROR_DATA_MAX_SIZE); 1828 1829 /* Copy the error info into the buffer */ 1830 memcpy((uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, pdu->hdr.raw, copy_len); 1831 nvme_tcp_pdu_set_data(rsp_pdu, (uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, copy_len); 1832 1833 /* Contain the header of the wrong received pdu */ 1834 c2h_term_req->common.plen = c2h_term_req->common.hlen + copy_len; 1835 tqpair->wait_terminate = true; 1836 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1837 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_c2h_term_req_complete, tqpair); 1838 } 1839 1840 static void 1841 nvmf_tcp_capsule_cmd_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1842 struct spdk_nvmf_tcp_qpair *tqpair, 1843 struct nvme_tcp_pdu *pdu) 1844 { 1845 struct spdk_nvmf_tcp_req *tcp_req; 1846 1847 assert(pdu->psh_valid_bytes == pdu->psh_len); 1848 assert(pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD); 1849 1850 tcp_req = nvmf_tcp_req_get(tqpair); 1851 if (!tcp_req) { 1852 /* Directly return and make the allocation retry again. This can happen if we're 1853 * using asynchronous writes to send the response to the host or when releasing 1854 * zero-copy buffers after a response has been sent. In both cases, the host might 1855 * receive the response before we've finished processing the request and is free to 1856 * send another one. 1857 */ 1858 if (tqpair->state_cntr[TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST] > 0 || 1859 tqpair->state_cntr[TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE] > 0) { 1860 return; 1861 } 1862 1863 /* The host sent more commands than the maximum queue depth. */ 1864 SPDK_ERRLOG("Cannot allocate tcp_req on tqpair=%p\n", tqpair); 1865 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1866 return; 1867 } 1868 1869 pdu->req = tcp_req; 1870 assert(tcp_req->state == TCP_REQUEST_STATE_NEW); 1871 nvmf_tcp_req_process(ttransport, tcp_req); 1872 } 1873 1874 static void 1875 nvmf_tcp_capsule_cmd_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1876 struct spdk_nvmf_tcp_qpair *tqpair, 1877 struct nvme_tcp_pdu *pdu) 1878 { 1879 struct spdk_nvmf_tcp_req *tcp_req; 1880 struct spdk_nvme_tcp_cmd *capsule_cmd; 1881 uint32_t error_offset = 0; 1882 enum spdk_nvme_tcp_term_req_fes fes; 1883 struct spdk_nvme_cpl *rsp; 1884 1885 capsule_cmd = &pdu->hdr.capsule_cmd; 1886 tcp_req = pdu->req; 1887 assert(tcp_req != NULL); 1888 1889 /* Zero-copy requests don't support ICD */ 1890 assert(!spdk_nvmf_request_using_zcopy(&tcp_req->req)); 1891 1892 if (capsule_cmd->common.pdo > SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET) { 1893 SPDK_ERRLOG("Expected ICReq capsule_cmd pdu offset <= %d, got %c\n", 1894 SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET, capsule_cmd->common.pdo); 1895 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1896 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 1897 goto err; 1898 } 1899 1900 rsp = &tcp_req->req.rsp->nvme_cpl; 1901 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1902 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1903 } else { 1904 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1905 } 1906 1907 nvmf_tcp_req_process(ttransport, tcp_req); 1908 1909 return; 1910 err: 1911 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1912 } 1913 1914 static void 1915 nvmf_tcp_h2c_data_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1916 struct spdk_nvmf_tcp_qpair *tqpair, 1917 struct nvme_tcp_pdu *pdu) 1918 { 1919 struct spdk_nvmf_tcp_req *tcp_req; 1920 uint32_t error_offset = 0; 1921 enum spdk_nvme_tcp_term_req_fes fes = 0; 1922 struct spdk_nvme_tcp_h2c_data_hdr *h2c_data; 1923 1924 h2c_data = &pdu->hdr.h2c_data; 1925 1926 SPDK_DEBUGLOG(nvmf_tcp, "tqpair=%p, r2t_info: datao=%u, datal=%u, cccid=%u, ttag=%u\n", 1927 tqpair, h2c_data->datao, h2c_data->datal, h2c_data->cccid, h2c_data->ttag); 1928 1929 if (h2c_data->ttag > tqpair->resource_count) { 1930 SPDK_DEBUGLOG(nvmf_tcp, "ttag %u is larger than allowed %u.\n", h2c_data->ttag, 1931 tqpair->resource_count); 1932 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1933 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1934 goto err; 1935 } 1936 1937 tcp_req = &tqpair->reqs[h2c_data->ttag - 1]; 1938 1939 if (spdk_unlikely(tcp_req->state != TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER && 1940 tcp_req->state != TCP_REQUEST_STATE_AWAITING_R2T_ACK)) { 1941 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, has error state in %d\n", tcp_req, tqpair, 1942 tcp_req->state); 1943 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1944 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1945 goto err; 1946 } 1947 1948 if (spdk_unlikely(tcp_req->req.cmd->nvme_cmd.cid != h2c_data->cccid)) { 1949 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, expected %u but %u for cccid.\n", tcp_req, tqpair, 1950 tcp_req->req.cmd->nvme_cmd.cid, h2c_data->cccid); 1951 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1952 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, cccid); 1953 goto err; 1954 } 1955 1956 if (tcp_req->h2c_offset != h2c_data->datao) { 1957 SPDK_DEBUGLOG(nvmf_tcp, 1958 "tcp_req(%p), tqpair=%p, expected data offset %u, but data offset is %u\n", 1959 tcp_req, tqpair, tcp_req->h2c_offset, h2c_data->datao); 1960 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1961 goto err; 1962 } 1963 1964 if ((h2c_data->datao + h2c_data->datal) > tcp_req->req.length) { 1965 SPDK_DEBUGLOG(nvmf_tcp, 1966 "tcp_req(%p), tqpair=%p, (datao=%u + datal=%u) exceeds requested length=%u\n", 1967 tcp_req, tqpair, h2c_data->datao, h2c_data->datal, tcp_req->req.length); 1968 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1969 goto err; 1970 } 1971 1972 pdu->req = tcp_req; 1973 1974 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 1975 pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 1976 } 1977 1978 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 1979 h2c_data->datao, h2c_data->datal); 1980 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1981 return; 1982 1983 err: 1984 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1985 } 1986 1987 static void 1988 nvmf_tcp_send_capsule_resp_pdu(struct spdk_nvmf_tcp_req *tcp_req, 1989 struct spdk_nvmf_tcp_qpair *tqpair) 1990 { 1991 struct nvme_tcp_pdu *rsp_pdu; 1992 struct spdk_nvme_tcp_rsp *capsule_resp; 1993 1994 SPDK_DEBUGLOG(nvmf_tcp, "enter, tqpair=%p\n", tqpair); 1995 1996 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1997 assert(rsp_pdu != NULL); 1998 1999 capsule_resp = &rsp_pdu->hdr.capsule_resp; 2000 capsule_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_CAPSULE_RESP; 2001 capsule_resp->common.plen = capsule_resp->common.hlen = sizeof(*capsule_resp); 2002 capsule_resp->rccqe = tcp_req->req.rsp->nvme_cpl; 2003 if (tqpair->host_hdgst_enable) { 2004 capsule_resp->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2005 capsule_resp->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 2006 } 2007 2008 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_request_free, tcp_req); 2009 } 2010 2011 static void 2012 nvmf_tcp_pdu_c2h_data_complete(void *cb_arg) 2013 { 2014 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 2015 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, 2016 struct spdk_nvmf_tcp_qpair, qpair); 2017 2018 assert(tqpair != NULL); 2019 2020 if (spdk_unlikely(tcp_req->pdu->rw_offset < tcp_req->req.length)) { 2021 SPDK_DEBUGLOG(nvmf_tcp, "sending another C2H part, offset %u length %u\n", tcp_req->pdu->rw_offset, 2022 tcp_req->req.length); 2023 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2024 return; 2025 } 2026 2027 if (tcp_req->pdu->hdr.c2h_data.common.flags & SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS) { 2028 nvmf_tcp_request_free(tcp_req); 2029 } else { 2030 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2031 } 2032 } 2033 2034 static void 2035 nvmf_tcp_r2t_complete(void *cb_arg) 2036 { 2037 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 2038 struct spdk_nvmf_tcp_transport *ttransport; 2039 2040 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 2041 struct spdk_nvmf_tcp_transport, transport); 2042 2043 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2044 2045 if (tcp_req->h2c_offset == tcp_req->req.length) { 2046 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2047 nvmf_tcp_req_process(ttransport, tcp_req); 2048 } 2049 } 2050 2051 static void 2052 nvmf_tcp_send_r2t_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 2053 struct spdk_nvmf_tcp_req *tcp_req) 2054 { 2055 struct nvme_tcp_pdu *rsp_pdu; 2056 struct spdk_nvme_tcp_r2t_hdr *r2t; 2057 2058 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 2059 assert(rsp_pdu != NULL); 2060 2061 r2t = &rsp_pdu->hdr.r2t; 2062 r2t->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_R2T; 2063 r2t->common.plen = r2t->common.hlen = sizeof(*r2t); 2064 2065 if (tqpair->host_hdgst_enable) { 2066 r2t->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2067 r2t->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 2068 } 2069 2070 r2t->cccid = tcp_req->req.cmd->nvme_cmd.cid; 2071 r2t->ttag = tcp_req->ttag; 2072 r2t->r2to = tcp_req->h2c_offset; 2073 r2t->r2tl = tcp_req->req.length; 2074 2075 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 2076 2077 SPDK_DEBUGLOG(nvmf_tcp, 2078 "tcp_req(%p) on tqpair(%p), r2t_info: cccid=%u, ttag=%u, r2to=%u, r2tl=%u\n", 2079 tcp_req, tqpair, r2t->cccid, r2t->ttag, r2t->r2to, r2t->r2tl); 2080 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_r2t_complete, tcp_req); 2081 } 2082 2083 static void 2084 nvmf_tcp_h2c_data_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 2085 struct spdk_nvmf_tcp_qpair *tqpair, 2086 struct nvme_tcp_pdu *pdu) 2087 { 2088 struct spdk_nvmf_tcp_req *tcp_req; 2089 struct spdk_nvme_cpl *rsp; 2090 2091 tcp_req = pdu->req; 2092 assert(tcp_req != NULL); 2093 2094 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2095 2096 tcp_req->h2c_offset += pdu->data_len; 2097 2098 /* Wait for all of the data to arrive AND for the initial R2T PDU send to be 2099 * acknowledged before moving on. */ 2100 if (tcp_req->h2c_offset == tcp_req->req.length && 2101 tcp_req->state == TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER) { 2102 /* After receiving all the h2c data, we need to check whether there is 2103 * transient transport error */ 2104 rsp = &tcp_req->req.rsp->nvme_cpl; 2105 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 2106 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2107 } else { 2108 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2109 } 2110 nvmf_tcp_req_process(ttransport, tcp_req); 2111 } 2112 } 2113 2114 static void 2115 nvmf_tcp_h2c_term_req_dump(struct spdk_nvme_tcp_term_req_hdr *h2c_term_req) 2116 { 2117 SPDK_ERRLOG("Error info of pdu(%p): %s\n", h2c_term_req, 2118 spdk_nvmf_tcp_term_req_fes_str[h2c_term_req->fes]); 2119 if ((h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 2120 (h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 2121 SPDK_DEBUGLOG(nvmf_tcp, "The offset from the start of the PDU header is %u\n", 2122 DGET32(h2c_term_req->fei)); 2123 } 2124 } 2125 2126 static void 2127 nvmf_tcp_h2c_term_req_hdr_handle(struct spdk_nvmf_tcp_qpair *tqpair, 2128 struct nvme_tcp_pdu *pdu) 2129 { 2130 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 2131 uint32_t error_offset = 0; 2132 enum spdk_nvme_tcp_term_req_fes fes; 2133 2134 if (h2c_term_req->fes > SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER) { 2135 SPDK_ERRLOG("Fatal Error Status(FES) is unknown for h2c_term_req pdu=%p\n", pdu); 2136 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2137 error_offset = offsetof(struct spdk_nvme_tcp_term_req_hdr, fes); 2138 goto end; 2139 } 2140 2141 /* set the data buffer */ 2142 nvme_tcp_pdu_set_data(pdu, (uint8_t *)pdu->hdr.raw + h2c_term_req->common.hlen, 2143 h2c_term_req->common.plen - h2c_term_req->common.hlen); 2144 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2145 return; 2146 end: 2147 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2148 } 2149 2150 static void 2151 nvmf_tcp_h2c_term_req_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, 2152 struct nvme_tcp_pdu *pdu) 2153 { 2154 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 2155 2156 nvmf_tcp_h2c_term_req_dump(h2c_term_req); 2157 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2158 } 2159 2160 static void 2161 _nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 2162 { 2163 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2164 struct spdk_nvmf_tcp_transport, transport); 2165 2166 switch (pdu->hdr.common.pdu_type) { 2167 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2168 nvmf_tcp_capsule_cmd_payload_handle(ttransport, tqpair, pdu); 2169 break; 2170 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2171 nvmf_tcp_h2c_data_payload_handle(ttransport, tqpair, pdu); 2172 break; 2173 2174 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2175 nvmf_tcp_h2c_term_req_payload_handle(tqpair, pdu); 2176 break; 2177 2178 default: 2179 /* The code should not go to here */ 2180 SPDK_ERRLOG("ERROR pdu type %d\n", pdu->hdr.common.pdu_type); 2181 break; 2182 } 2183 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 2184 tqpair->tcp_pdu_working_count--; 2185 } 2186 2187 static inline void 2188 nvmf_tcp_req_set_cpl(struct spdk_nvmf_tcp_req *treq, int sct, int sc) 2189 { 2190 treq->req.rsp->nvme_cpl.status.sct = sct; 2191 treq->req.rsp->nvme_cpl.status.sc = sc; 2192 treq->req.rsp->nvme_cpl.cid = treq->req.cmd->nvme_cmd.cid; 2193 } 2194 2195 static void 2196 data_crc32_calc_done(void *cb_arg, int status) 2197 { 2198 struct nvme_tcp_pdu *pdu = cb_arg; 2199 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 2200 2201 /* async crc32 calculation is failed and use direct calculation to check */ 2202 if (spdk_unlikely(status)) { 2203 SPDK_ERRLOG("Data digest on tqpair=(%p) with pdu=%p failed to be calculated asynchronously\n", 2204 tqpair, pdu); 2205 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 2206 } 2207 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 2208 if (!MATCH_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32)) { 2209 SPDK_ERRLOG("Data digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 2210 assert(pdu->req != NULL); 2211 nvmf_tcp_req_set_cpl(pdu->req, SPDK_NVME_SCT_GENERIC, 2212 SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR); 2213 } 2214 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2215 } 2216 2217 static void 2218 nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 2219 { 2220 int rc = 0; 2221 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2222 tqpair->pdu_in_progress = NULL; 2223 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2224 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2225 /* check data digest if need */ 2226 if (pdu->ddgst_enable) { 2227 if (tqpair->qpair.qid != 0 && !pdu->dif_ctx && tqpair->group && 2228 (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0)) { 2229 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 2230 pdu->data_iovcnt, 0, data_crc32_calc_done, pdu); 2231 if (spdk_likely(rc == 0)) { 2232 return; 2233 } 2234 } else { 2235 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 2236 } 2237 data_crc32_calc_done(pdu, rc); 2238 } else { 2239 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2240 } 2241 } 2242 2243 static void 2244 nvmf_tcp_send_icresp_complete(void *cb_arg) 2245 { 2246 struct spdk_nvmf_tcp_qpair *tqpair = cb_arg; 2247 2248 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_RUNNING); 2249 } 2250 2251 static void 2252 nvmf_tcp_icreq_handle(struct spdk_nvmf_tcp_transport *ttransport, 2253 struct spdk_nvmf_tcp_qpair *tqpair, 2254 struct nvme_tcp_pdu *pdu) 2255 { 2256 struct spdk_nvme_tcp_ic_req *ic_req = &pdu->hdr.ic_req; 2257 struct nvme_tcp_pdu *rsp_pdu; 2258 struct spdk_nvme_tcp_ic_resp *ic_resp; 2259 uint32_t error_offset = 0; 2260 enum spdk_nvme_tcp_term_req_fes fes; 2261 2262 /* Only PFV 0 is defined currently */ 2263 if (ic_req->pfv != 0) { 2264 SPDK_ERRLOG("Expected ICReq PFV %u, got %u\n", 0u, ic_req->pfv); 2265 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2266 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, pfv); 2267 goto end; 2268 } 2269 2270 /* This value is 0’s based value in units of dwords should not be larger than SPDK_NVME_TCP_HPDA_MAX */ 2271 if (ic_req->hpda > SPDK_NVME_TCP_HPDA_MAX) { 2272 SPDK_ERRLOG("ICReq HPDA out of range 0 to 31, got %u\n", ic_req->hpda); 2273 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2274 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, hpda); 2275 goto end; 2276 } 2277 2278 /* MAXR2T is 0's based */ 2279 SPDK_DEBUGLOG(nvmf_tcp, "maxr2t =%u\n", (ic_req->maxr2t + 1u)); 2280 2281 tqpair->host_hdgst_enable = ic_req->dgst.bits.hdgst_enable ? true : false; 2282 if (!tqpair->host_hdgst_enable) { 2283 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 2284 } 2285 2286 tqpair->host_ddgst_enable = ic_req->dgst.bits.ddgst_enable ? true : false; 2287 if (!tqpair->host_ddgst_enable) { 2288 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 2289 } 2290 2291 tqpair->recv_buf_size = spdk_max(tqpair->recv_buf_size, MIN_SOCK_PIPE_SIZE); 2292 /* Now that we know whether digests are enabled, properly size the receive buffer */ 2293 if (spdk_sock_set_recvbuf(tqpair->sock, tqpair->recv_buf_size) < 0) { 2294 SPDK_WARNLOG("Unable to allocate enough memory for receive buffer on tqpair=%p with size=%d\n", 2295 tqpair, 2296 tqpair->recv_buf_size); 2297 /* Not fatal. */ 2298 } 2299 2300 tqpair->cpda = spdk_min(ic_req->hpda, SPDK_NVME_TCP_CPDA_MAX); 2301 SPDK_DEBUGLOG(nvmf_tcp, "cpda of tqpair=(%p) is : %u\n", tqpair, tqpair->cpda); 2302 2303 rsp_pdu = tqpair->mgmt_pdu; 2304 2305 ic_resp = &rsp_pdu->hdr.ic_resp; 2306 ic_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_IC_RESP; 2307 ic_resp->common.hlen = ic_resp->common.plen = sizeof(*ic_resp); 2308 ic_resp->pfv = 0; 2309 ic_resp->cpda = tqpair->cpda; 2310 ic_resp->maxh2cdata = ttransport->transport.opts.max_io_size; 2311 ic_resp->dgst.bits.hdgst_enable = tqpair->host_hdgst_enable ? 1 : 0; 2312 ic_resp->dgst.bits.ddgst_enable = tqpair->host_ddgst_enable ? 1 : 0; 2313 2314 SPDK_DEBUGLOG(nvmf_tcp, "host_hdgst_enable: %u\n", tqpair->host_hdgst_enable); 2315 SPDK_DEBUGLOG(nvmf_tcp, "host_ddgst_enable: %u\n", tqpair->host_ddgst_enable); 2316 2317 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INITIALIZING); 2318 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_icresp_complete, tqpair); 2319 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2320 return; 2321 end: 2322 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2323 } 2324 2325 static void 2326 nvmf_tcp_pdu_psh_handle(struct spdk_nvmf_tcp_qpair *tqpair, 2327 struct spdk_nvmf_tcp_transport *ttransport) 2328 { 2329 struct nvme_tcp_pdu *pdu; 2330 int rc; 2331 uint32_t crc32c, error_offset = 0; 2332 enum spdk_nvme_tcp_term_req_fes fes; 2333 2334 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2335 pdu = tqpair->pdu_in_progress; 2336 2337 SPDK_DEBUGLOG(nvmf_tcp, "pdu type of tqpair(%p) is %d\n", tqpair, 2338 pdu->hdr.common.pdu_type); 2339 /* check header digest if needed */ 2340 if (pdu->has_hdgst) { 2341 SPDK_DEBUGLOG(nvmf_tcp, "Compare the header of pdu=%p on tqpair=%p\n", pdu, tqpair); 2342 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 2343 rc = MATCH_DIGEST_WORD((uint8_t *)pdu->hdr.raw + pdu->hdr.common.hlen, crc32c); 2344 if (rc == 0) { 2345 SPDK_ERRLOG("Header digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 2346 fes = SPDK_NVME_TCP_TERM_REQ_FES_HDGST_ERROR; 2347 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2348 return; 2349 2350 } 2351 } 2352 2353 switch (pdu->hdr.common.pdu_type) { 2354 case SPDK_NVME_TCP_PDU_TYPE_IC_REQ: 2355 nvmf_tcp_icreq_handle(ttransport, tqpair, pdu); 2356 break; 2357 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2358 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_REQ); 2359 break; 2360 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2361 nvmf_tcp_h2c_data_hdr_handle(ttransport, tqpair, pdu); 2362 break; 2363 2364 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2365 nvmf_tcp_h2c_term_req_hdr_handle(tqpair, pdu); 2366 break; 2367 2368 default: 2369 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", tqpair->pdu_in_progress->hdr.common.pdu_type); 2370 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2371 error_offset = 1; 2372 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2373 break; 2374 } 2375 } 2376 2377 static void 2378 nvmf_tcp_pdu_ch_handle(struct spdk_nvmf_tcp_qpair *tqpair) 2379 { 2380 struct nvme_tcp_pdu *pdu; 2381 uint32_t error_offset = 0; 2382 enum spdk_nvme_tcp_term_req_fes fes; 2383 uint8_t expected_hlen, pdo; 2384 bool plen_error = false, pdo_error = false; 2385 2386 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2387 pdu = tqpair->pdu_in_progress; 2388 assert(pdu); 2389 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_REQ) { 2390 if (tqpair->state != NVME_TCP_QPAIR_STATE_INVALID) { 2391 SPDK_ERRLOG("Already received ICreq PDU, and reject this pdu=%p\n", pdu); 2392 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2393 goto err; 2394 } 2395 expected_hlen = sizeof(struct spdk_nvme_tcp_ic_req); 2396 if (pdu->hdr.common.plen != expected_hlen) { 2397 plen_error = true; 2398 } 2399 } else { 2400 if (tqpair->state != NVME_TCP_QPAIR_STATE_RUNNING) { 2401 SPDK_ERRLOG("The TCP/IP connection is not negotiated\n"); 2402 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2403 goto err; 2404 } 2405 2406 switch (pdu->hdr.common.pdu_type) { 2407 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2408 expected_hlen = sizeof(struct spdk_nvme_tcp_cmd); 2409 pdo = pdu->hdr.common.pdo; 2410 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2411 pdo_error = true; 2412 break; 2413 } 2414 2415 if (pdu->hdr.common.plen < expected_hlen) { 2416 plen_error = true; 2417 } 2418 break; 2419 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2420 expected_hlen = sizeof(struct spdk_nvme_tcp_h2c_data_hdr); 2421 pdo = pdu->hdr.common.pdo; 2422 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2423 pdo_error = true; 2424 break; 2425 } 2426 if (pdu->hdr.common.plen < expected_hlen) { 2427 plen_error = true; 2428 } 2429 break; 2430 2431 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2432 expected_hlen = sizeof(struct spdk_nvme_tcp_term_req_hdr); 2433 if ((pdu->hdr.common.plen <= expected_hlen) || 2434 (pdu->hdr.common.plen > SPDK_NVME_TCP_TERM_REQ_PDU_MAX_SIZE)) { 2435 plen_error = true; 2436 } 2437 break; 2438 2439 default: 2440 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", pdu->hdr.common.pdu_type); 2441 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2442 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdu_type); 2443 goto err; 2444 } 2445 } 2446 2447 if (pdu->hdr.common.hlen != expected_hlen) { 2448 SPDK_ERRLOG("PDU type=0x%02x, Expected ICReq header length %u, got %u on tqpair=%p\n", 2449 pdu->hdr.common.pdu_type, 2450 expected_hlen, pdu->hdr.common.hlen, tqpair); 2451 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2452 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, hlen); 2453 goto err; 2454 } else if (pdo_error) { 2455 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2456 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 2457 } else if (plen_error) { 2458 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2459 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2460 goto err; 2461 } else { 2462 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2463 nvme_tcp_pdu_calc_psh_len(tqpair->pdu_in_progress, tqpair->host_hdgst_enable); 2464 return; 2465 } 2466 err: 2467 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2468 } 2469 2470 static int 2471 nvmf_tcp_sock_process(struct spdk_nvmf_tcp_qpair *tqpair) 2472 { 2473 int rc = 0; 2474 struct nvme_tcp_pdu *pdu; 2475 enum nvme_tcp_pdu_recv_state prev_state; 2476 uint32_t data_len; 2477 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2478 struct spdk_nvmf_tcp_transport, transport); 2479 2480 /* The loop here is to allow for several back-to-back state changes. */ 2481 do { 2482 prev_state = tqpair->recv_state; 2483 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv pdu entering state %d\n", tqpair, prev_state); 2484 2485 pdu = tqpair->pdu_in_progress; 2486 assert(pdu != NULL || 2487 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY || 2488 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_QUIESCING || 2489 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 2490 2491 switch (tqpair->recv_state) { 2492 /* Wait for the common header */ 2493 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY: 2494 if (!pdu) { 2495 pdu = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 2496 if (spdk_unlikely(!pdu)) { 2497 return NVME_TCP_PDU_IN_PROGRESS; 2498 } 2499 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 2500 tqpair->pdu_in_progress = pdu; 2501 tqpair->tcp_pdu_working_count++; 2502 } 2503 memset(pdu, 0, offsetof(struct nvme_tcp_pdu, qpair)); 2504 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2505 /* FALLTHROUGH */ 2506 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH: 2507 if (spdk_unlikely(tqpair->state == NVME_TCP_QPAIR_STATE_INITIALIZING)) { 2508 return rc; 2509 } 2510 2511 rc = nvme_tcp_read_data(tqpair->sock, 2512 sizeof(struct spdk_nvme_tcp_common_pdu_hdr) - pdu->ch_valid_bytes, 2513 (void *)&pdu->hdr.common + pdu->ch_valid_bytes); 2514 if (rc < 0) { 2515 SPDK_DEBUGLOG(nvmf_tcp, "will disconnect tqpair=%p\n", tqpair); 2516 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2517 break; 2518 } else if (rc > 0) { 2519 pdu->ch_valid_bytes += rc; 2520 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.trace_id, rc, 0); 2521 } 2522 2523 if (pdu->ch_valid_bytes < sizeof(struct spdk_nvme_tcp_common_pdu_hdr)) { 2524 return NVME_TCP_PDU_IN_PROGRESS; 2525 } 2526 2527 /* The command header of this PDU has now been read from the socket. */ 2528 nvmf_tcp_pdu_ch_handle(tqpair); 2529 break; 2530 /* Wait for the pdu specific header */ 2531 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH: 2532 rc = nvme_tcp_read_data(tqpair->sock, 2533 pdu->psh_len - pdu->psh_valid_bytes, 2534 (void *)&pdu->hdr.raw + sizeof(struct spdk_nvme_tcp_common_pdu_hdr) + pdu->psh_valid_bytes); 2535 if (rc < 0) { 2536 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2537 break; 2538 } else if (rc > 0) { 2539 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.trace_id, rc, 0); 2540 pdu->psh_valid_bytes += rc; 2541 } 2542 2543 if (pdu->psh_valid_bytes < pdu->psh_len) { 2544 return NVME_TCP_PDU_IN_PROGRESS; 2545 } 2546 2547 /* All header(ch, psh, head digits) of this PDU has now been read from the socket. */ 2548 nvmf_tcp_pdu_psh_handle(tqpair, ttransport); 2549 break; 2550 /* Wait for the req slot */ 2551 case NVME_TCP_PDU_RECV_STATE_AWAIT_REQ: 2552 nvmf_tcp_capsule_cmd_hdr_handle(ttransport, tqpair, pdu); 2553 break; 2554 /* Wait for the request processing loop to acquire a buffer for the PDU */ 2555 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_BUF: 2556 break; 2557 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD: 2558 /* check whether the data is valid, if not we just return */ 2559 if (!pdu->data_len) { 2560 return NVME_TCP_PDU_IN_PROGRESS; 2561 } 2562 2563 data_len = pdu->data_len; 2564 /* data digest */ 2565 if (spdk_unlikely((pdu->hdr.common.pdu_type != SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ) && 2566 tqpair->host_ddgst_enable)) { 2567 data_len += SPDK_NVME_TCP_DIGEST_LEN; 2568 pdu->ddgst_enable = true; 2569 } 2570 2571 rc = nvme_tcp_read_payload_data(tqpair->sock, pdu); 2572 if (rc < 0) { 2573 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2574 break; 2575 } 2576 pdu->rw_offset += rc; 2577 2578 if (pdu->rw_offset < data_len) { 2579 return NVME_TCP_PDU_IN_PROGRESS; 2580 } 2581 2582 /* Generate and insert DIF to whole data block received if DIF is enabled */ 2583 if (spdk_unlikely(pdu->dif_ctx != NULL) && 2584 spdk_dif_generate_stream(pdu->data_iov, pdu->data_iovcnt, 0, data_len, 2585 pdu->dif_ctx) != 0) { 2586 SPDK_ERRLOG("DIF generate failed\n"); 2587 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2588 break; 2589 } 2590 2591 /* All of this PDU has now been read from the socket. */ 2592 nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2593 break; 2594 case NVME_TCP_PDU_RECV_STATE_QUIESCING: 2595 if (tqpair->tcp_pdu_working_count != 0) { 2596 return NVME_TCP_PDU_IN_PROGRESS; 2597 } 2598 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 2599 break; 2600 case NVME_TCP_PDU_RECV_STATE_ERROR: 2601 if (spdk_sock_is_connected(tqpair->sock) && tqpair->wait_terminate) { 2602 return NVME_TCP_PDU_IN_PROGRESS; 2603 } 2604 return NVME_TCP_PDU_FATAL; 2605 default: 2606 SPDK_ERRLOG("The state(%d) is invalid\n", tqpair->recv_state); 2607 abort(); 2608 break; 2609 } 2610 } while (tqpair->recv_state != prev_state); 2611 2612 return rc; 2613 } 2614 2615 static inline void * 2616 nvmf_tcp_control_msg_get(struct spdk_nvmf_tcp_control_msg_list *list, 2617 struct spdk_nvmf_tcp_req *tcp_req) 2618 { 2619 struct spdk_nvmf_tcp_control_msg *msg; 2620 2621 assert(list); 2622 2623 msg = STAILQ_FIRST(&list->free_msgs); 2624 if (!msg) { 2625 SPDK_DEBUGLOG(nvmf_tcp, "Out of control messages\n"); 2626 STAILQ_INSERT_TAIL(&list->waiting_for_msg_reqs, tcp_req, control_msg_link); 2627 return NULL; 2628 } 2629 STAILQ_REMOVE_HEAD(&list->free_msgs, link); 2630 return msg; 2631 } 2632 2633 static inline void 2634 nvmf_tcp_control_msg_put(struct spdk_nvmf_tcp_control_msg_list *list, void *_msg) 2635 { 2636 struct spdk_nvmf_tcp_control_msg *msg = _msg; 2637 struct spdk_nvmf_tcp_req *tcp_req; 2638 struct spdk_nvmf_tcp_transport *ttransport; 2639 2640 assert(list); 2641 STAILQ_INSERT_HEAD(&list->free_msgs, msg, link); 2642 if (!STAILQ_EMPTY(&list->waiting_for_msg_reqs)) { 2643 tcp_req = STAILQ_FIRST(&list->waiting_for_msg_reqs); 2644 STAILQ_REMOVE_HEAD(&list->waiting_for_msg_reqs, control_msg_link); 2645 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 2646 struct spdk_nvmf_tcp_transport, transport); 2647 nvmf_tcp_req_process(ttransport, tcp_req); 2648 } 2649 } 2650 2651 static void 2652 nvmf_tcp_req_parse_sgl(struct spdk_nvmf_tcp_req *tcp_req, 2653 struct spdk_nvmf_transport *transport, 2654 struct spdk_nvmf_transport_poll_group *group) 2655 { 2656 struct spdk_nvmf_request *req = &tcp_req->req; 2657 struct spdk_nvme_cmd *cmd; 2658 struct spdk_nvme_sgl_descriptor *sgl; 2659 struct spdk_nvmf_tcp_poll_group *tgroup; 2660 enum spdk_nvme_tcp_term_req_fes fes; 2661 struct nvme_tcp_pdu *pdu; 2662 struct spdk_nvmf_tcp_qpair *tqpair; 2663 uint32_t length, error_offset = 0; 2664 2665 cmd = &req->cmd->nvme_cmd; 2666 sgl = &cmd->dptr.sgl1; 2667 2668 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK && 2669 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_TRANSPORT) { 2670 /* get request length from sgl */ 2671 length = sgl->unkeyed.length; 2672 if (spdk_unlikely(length > transport->opts.max_io_size)) { 2673 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 2674 length, transport->opts.max_io_size); 2675 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_LIMIT_EXCEEDED; 2676 goto fatal_err; 2677 } 2678 2679 /* fill request length and populate iovs */ 2680 req->length = length; 2681 2682 SPDK_DEBUGLOG(nvmf_tcp, "Data requested length= 0x%x\n", length); 2683 2684 if (spdk_unlikely(req->dif_enabled)) { 2685 req->dif.orig_length = length; 2686 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2687 req->dif.elba_length = length; 2688 } 2689 2690 if (nvmf_ctrlr_use_zcopy(req)) { 2691 SPDK_DEBUGLOG(nvmf_tcp, "Using zero-copy to execute request %p\n", tcp_req); 2692 req->data_from_pool = false; 2693 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_HAVE_BUFFER); 2694 return; 2695 } 2696 2697 if (spdk_nvmf_request_get_buffers(req, group, transport, length)) { 2698 /* No available buffers. Queue this request up. */ 2699 SPDK_DEBUGLOG(nvmf_tcp, "No available large data buffers. Queueing request %p\n", 2700 tcp_req); 2701 return; 2702 } 2703 2704 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_HAVE_BUFFER); 2705 SPDK_DEBUGLOG(nvmf_tcp, "Request %p took %d buffer/s from central pool, and data=%p\n", 2706 tcp_req, req->iovcnt, req->iov[0].iov_base); 2707 2708 return; 2709 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 2710 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 2711 uint64_t offset = sgl->address; 2712 uint32_t max_len = transport->opts.in_capsule_data_size; 2713 2714 assert(tcp_req->has_in_capsule_data); 2715 /* Capsule Cmd with In-capsule Data should get data length from pdu header */ 2716 tqpair = tcp_req->pdu->qpair; 2717 /* receiving pdu is not same with the pdu in tcp_req */ 2718 pdu = tqpair->pdu_in_progress; 2719 length = pdu->hdr.common.plen - pdu->psh_len - sizeof(struct spdk_nvme_tcp_common_pdu_hdr); 2720 if (tqpair->host_ddgst_enable) { 2721 length -= SPDK_NVME_TCP_DIGEST_LEN; 2722 } 2723 /* This error is not defined in NVMe/TCP spec, take this error as fatal error */ 2724 if (spdk_unlikely(length != sgl->unkeyed.length)) { 2725 SPDK_ERRLOG("In-Capsule Data length 0x%x is not equal to SGL data length 0x%x\n", 2726 length, sgl->unkeyed.length); 2727 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2728 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2729 goto fatal_err; 2730 } 2731 2732 SPDK_DEBUGLOG(nvmf_tcp, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 2733 offset, length); 2734 2735 /* The NVMe/TCP transport does not use ICDOFF to control the in-capsule data offset. ICDOFF should be '0' */ 2736 if (spdk_unlikely(offset != 0)) { 2737 /* Not defined fatal error in NVMe/TCP spec, handle this error as a fatal error */ 2738 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " should be ZERO in NVMe/TCP\n", offset); 2739 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER; 2740 error_offset = offsetof(struct spdk_nvme_tcp_cmd, ccsqe.dptr.sgl1.address); 2741 goto fatal_err; 2742 } 2743 2744 if (spdk_unlikely(length > max_len)) { 2745 /* According to the SPEC we should support ICD up to 8192 bytes for admin and fabric commands */ 2746 if (length <= SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE && 2747 (cmd->opc == SPDK_NVME_OPC_FABRIC || req->qpair->qid == 0)) { 2748 2749 /* Get a buffer from dedicated list */ 2750 SPDK_DEBUGLOG(nvmf_tcp, "Getting a buffer from control msg list\n"); 2751 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 2752 assert(tgroup->control_msg_list); 2753 req->iov[0].iov_base = nvmf_tcp_control_msg_get(tgroup->control_msg_list, tcp_req); 2754 if (!req->iov[0].iov_base) { 2755 /* No available buffers. Queue this request up. */ 2756 SPDK_DEBUGLOG(nvmf_tcp, "No available ICD buffers. Queueing request %p\n", tcp_req); 2757 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_BUF); 2758 return; 2759 } 2760 } else { 2761 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 2762 length, max_len); 2763 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_LIMIT_EXCEEDED; 2764 goto fatal_err; 2765 } 2766 } else { 2767 req->iov[0].iov_base = tcp_req->buf; 2768 } 2769 2770 req->length = length; 2771 req->data_from_pool = false; 2772 2773 if (spdk_unlikely(req->dif_enabled)) { 2774 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2775 req->dif.elba_length = length; 2776 } 2777 2778 req->iov[0].iov_len = length; 2779 req->iovcnt = 1; 2780 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_HAVE_BUFFER); 2781 2782 return; 2783 } 2784 /* If we want to handle the problem here, then we can't skip the following data segment. 2785 * Because this function runs before reading data part, now handle all errors as fatal errors. */ 2786 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 2787 sgl->generic.type, sgl->generic.subtype); 2788 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER; 2789 error_offset = offsetof(struct spdk_nvme_tcp_cmd, ccsqe.dptr.sgl1.generic); 2790 fatal_err: 2791 nvmf_tcp_send_c2h_term_req(tcp_req->pdu->qpair, tcp_req->pdu, fes, error_offset); 2792 } 2793 2794 static inline enum spdk_nvme_media_error_status_code 2795 nvmf_tcp_dif_error_to_compl_status(uint8_t err_type) { 2796 enum spdk_nvme_media_error_status_code result; 2797 2798 switch (err_type) 2799 { 2800 case SPDK_DIF_REFTAG_ERROR: 2801 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 2802 break; 2803 case SPDK_DIF_APPTAG_ERROR: 2804 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 2805 break; 2806 case SPDK_DIF_GUARD_ERROR: 2807 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 2808 break; 2809 default: 2810 SPDK_UNREACHABLE(); 2811 break; 2812 } 2813 2814 return result; 2815 } 2816 2817 static void 2818 _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2819 struct spdk_nvmf_tcp_req *tcp_req) 2820 { 2821 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF( 2822 tqpair->qpair.transport, struct spdk_nvmf_tcp_transport, transport); 2823 struct nvme_tcp_pdu *rsp_pdu; 2824 struct spdk_nvme_tcp_c2h_data_hdr *c2h_data; 2825 uint32_t plen, pdo, alignment; 2826 int rc; 2827 2828 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2829 2830 rsp_pdu = tcp_req->pdu; 2831 assert(rsp_pdu != NULL); 2832 2833 c2h_data = &rsp_pdu->hdr.c2h_data; 2834 c2h_data->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_DATA; 2835 plen = c2h_data->common.hlen = sizeof(*c2h_data); 2836 2837 if (tqpair->host_hdgst_enable) { 2838 plen += SPDK_NVME_TCP_DIGEST_LEN; 2839 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2840 } 2841 2842 /* set the psh */ 2843 c2h_data->cccid = tcp_req->req.cmd->nvme_cmd.cid; 2844 c2h_data->datal = tcp_req->req.length - tcp_req->pdu->rw_offset; 2845 c2h_data->datao = tcp_req->pdu->rw_offset; 2846 2847 /* set the padding */ 2848 rsp_pdu->padding_len = 0; 2849 pdo = plen; 2850 if (tqpair->cpda) { 2851 alignment = (tqpair->cpda + 1) << 2; 2852 if (plen % alignment != 0) { 2853 pdo = (plen + alignment) / alignment * alignment; 2854 rsp_pdu->padding_len = pdo - plen; 2855 plen = pdo; 2856 } 2857 } 2858 2859 c2h_data->common.pdo = pdo; 2860 plen += c2h_data->datal; 2861 if (tqpair->host_ddgst_enable) { 2862 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_DDGSTF; 2863 plen += SPDK_NVME_TCP_DIGEST_LEN; 2864 } 2865 2866 c2h_data->common.plen = plen; 2867 2868 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2869 rsp_pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 2870 } 2871 2872 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2873 c2h_data->datao, c2h_data->datal); 2874 2875 2876 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU; 2877 /* Need to send the capsule response if response is not all 0 */ 2878 if (ttransport->tcp_opts.c2h_success && 2879 tcp_req->rsp.cdw0 == 0 && tcp_req->rsp.cdw1 == 0) { 2880 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS; 2881 } 2882 2883 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2884 struct spdk_nvme_cpl *rsp = &tcp_req->req.rsp->nvme_cpl; 2885 struct spdk_dif_error err_blk = {}; 2886 uint32_t mapped_length = 0; 2887 uint32_t available_iovs = SPDK_COUNTOF(rsp_pdu->iov); 2888 uint32_t ddgst_len = 0; 2889 2890 if (tqpair->host_ddgst_enable) { 2891 /* Data digest consumes additional iov entry */ 2892 available_iovs--; 2893 /* plen needs to be updated since nvme_tcp_build_iovs compares expected and actual plen */ 2894 ddgst_len = SPDK_NVME_TCP_DIGEST_LEN; 2895 c2h_data->common.plen -= ddgst_len; 2896 } 2897 /* Temp call to estimate if data can be described by limited number of iovs. 2898 * iov vector will be rebuilt in nvmf_tcp_qpair_write_pdu */ 2899 nvme_tcp_build_iovs(rsp_pdu->iov, available_iovs, rsp_pdu, tqpair->host_hdgst_enable, 2900 false, &mapped_length); 2901 2902 if (mapped_length != c2h_data->common.plen) { 2903 c2h_data->datal = mapped_length - (c2h_data->common.plen - c2h_data->datal); 2904 SPDK_DEBUGLOG(nvmf_tcp, 2905 "Part C2H, data_len %u (of %u), PDU len %u, updated PDU len %u, offset %u\n", 2906 c2h_data->datal, tcp_req->req.length, c2h_data->common.plen, mapped_length, rsp_pdu->rw_offset); 2907 c2h_data->common.plen = mapped_length; 2908 2909 /* Rebuild pdu->data_iov since data length is changed */ 2910 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, c2h_data->datao, 2911 c2h_data->datal); 2912 2913 c2h_data->common.flags &= ~(SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU | 2914 SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS); 2915 } 2916 2917 c2h_data->common.plen += ddgst_len; 2918 2919 assert(rsp_pdu->rw_offset <= tcp_req->req.length); 2920 2921 rc = spdk_dif_verify_stream(rsp_pdu->data_iov, rsp_pdu->data_iovcnt, 2922 0, rsp_pdu->data_len, rsp_pdu->dif_ctx, &err_blk); 2923 if (rc != 0) { 2924 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", 2925 err_blk.err_type, err_blk.err_offset); 2926 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2927 rsp->status.sc = nvmf_tcp_dif_error_to_compl_status(err_blk.err_type); 2928 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2929 return; 2930 } 2931 } 2932 2933 rsp_pdu->rw_offset += c2h_data->datal; 2934 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_pdu_c2h_data_complete, tcp_req); 2935 } 2936 2937 static void 2938 nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2939 struct spdk_nvmf_tcp_req *tcp_req) 2940 { 2941 nvmf_tcp_req_pdu_init(tcp_req); 2942 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2943 } 2944 2945 static int 2946 request_transfer_out(struct spdk_nvmf_request *req) 2947 { 2948 struct spdk_nvmf_tcp_req *tcp_req; 2949 struct spdk_nvmf_qpair *qpair; 2950 struct spdk_nvmf_tcp_qpair *tqpair; 2951 struct spdk_nvme_cpl *rsp; 2952 2953 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2954 2955 qpair = req->qpair; 2956 rsp = &req->rsp->nvme_cpl; 2957 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 2958 2959 /* Advance our sq_head pointer */ 2960 if (qpair->sq_head == qpair->sq_head_max) { 2961 qpair->sq_head = 0; 2962 } else { 2963 qpair->sq_head++; 2964 } 2965 rsp->sqhd = qpair->sq_head; 2966 2967 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2968 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2969 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2970 nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2971 } else { 2972 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2973 } 2974 2975 return 0; 2976 } 2977 2978 static void 2979 nvmf_tcp_check_fused_ordering(struct spdk_nvmf_tcp_transport *ttransport, 2980 struct spdk_nvmf_tcp_qpair *tqpair, 2981 struct spdk_nvmf_tcp_req *tcp_req) 2982 { 2983 enum spdk_nvme_cmd_fuse last, next; 2984 2985 last = tqpair->fused_first ? tqpair->fused_first->cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2986 next = tcp_req->cmd.fuse; 2987 2988 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2989 2990 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2991 return; 2992 } 2993 2994 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2995 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2996 /* This is a valid pair of fused commands. Point them at each other 2997 * so they can be submitted consecutively once ready to be executed. 2998 */ 2999 tqpair->fused_first->fused_pair = tcp_req; 3000 tcp_req->fused_pair = tqpair->fused_first; 3001 tqpair->fused_first = NULL; 3002 return; 3003 } else { 3004 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 3005 tqpair->fused_first->fused_failed = true; 3006 3007 /* 3008 * If the last req is in READY_TO_EXECUTE state, then call 3009 * nvmf_tcp_req_process(), otherwise nothing else will kick it. 3010 */ 3011 if (tqpair->fused_first->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3012 nvmf_tcp_req_process(ttransport, tqpair->fused_first); 3013 } 3014 3015 tqpair->fused_first = NULL; 3016 } 3017 } 3018 3019 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 3020 /* Set tqpair->fused_first here so that we know to check that the next request 3021 * is a SECOND (and to fail this one if it isn't). 3022 */ 3023 tqpair->fused_first = tcp_req; 3024 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 3025 /* Mark this req failed since it is a SECOND and the last one was not a FIRST. */ 3026 tcp_req->fused_failed = true; 3027 } 3028 } 3029 3030 static bool 3031 nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 3032 struct spdk_nvmf_tcp_req *tcp_req) 3033 { 3034 struct spdk_nvmf_tcp_qpair *tqpair; 3035 uint32_t plen; 3036 struct nvme_tcp_pdu *pdu; 3037 enum spdk_nvmf_tcp_req_state prev_state; 3038 bool progress = false; 3039 struct spdk_nvmf_transport *transport = &ttransport->transport; 3040 struct spdk_nvmf_transport_poll_group *group; 3041 struct spdk_nvmf_tcp_poll_group *tgroup; 3042 3043 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 3044 group = &tqpair->group->group; 3045 assert(tcp_req->state != TCP_REQUEST_STATE_FREE); 3046 3047 /* If the qpair is not active, we need to abort the outstanding requests. */ 3048 if (!spdk_nvmf_qpair_is_active(&tqpair->qpair)) { 3049 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 3050 nvmf_tcp_request_get_buffers_abort(tcp_req); 3051 } 3052 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 3053 } 3054 3055 /* The loop here is to allow for several back-to-back state changes. */ 3056 do { 3057 prev_state = tcp_req->state; 3058 3059 SPDK_DEBUGLOG(nvmf_tcp, "Request %p entering state %d on tqpair=%p\n", tcp_req, prev_state, 3060 tqpair); 3061 3062 switch (tcp_req->state) { 3063 case TCP_REQUEST_STATE_FREE: 3064 /* Some external code must kick a request into TCP_REQUEST_STATE_NEW 3065 * to escape this state. */ 3066 break; 3067 case TCP_REQUEST_STATE_NEW: 3068 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEW, tqpair->qpair.trace_id, 0, (uintptr_t)tcp_req, 3069 tqpair->qpair.queue_depth); 3070 3071 /* copy the cmd from the receive pdu */ 3072 tcp_req->cmd = tqpair->pdu_in_progress->hdr.capsule_cmd.ccsqe; 3073 3074 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&tcp_req->req, &tcp_req->req.dif.dif_ctx))) { 3075 tcp_req->req.dif_enabled = true; 3076 tqpair->pdu_in_progress->dif_ctx = &tcp_req->req.dif.dif_ctx; 3077 } 3078 3079 nvmf_tcp_check_fused_ordering(ttransport, tqpair, tcp_req); 3080 3081 /* The next state transition depends on the data transfer needs of this request. */ 3082 tcp_req->req.xfer = spdk_nvmf_req_get_xfer(&tcp_req->req); 3083 3084 if (spdk_unlikely(tcp_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 3085 nvmf_tcp_req_set_cpl(tcp_req, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_INVALID_OPCODE); 3086 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 3087 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3088 SPDK_DEBUGLOG(nvmf_tcp, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", tcp_req); 3089 break; 3090 } 3091 3092 /* If no data to transfer, ready to execute. */ 3093 if (tcp_req->req.xfer == SPDK_NVME_DATA_NONE) { 3094 /* Reset the tqpair receiving pdu state */ 3095 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 3096 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 3097 break; 3098 } 3099 3100 pdu = tqpair->pdu_in_progress; 3101 plen = pdu->hdr.common.hlen; 3102 if (tqpair->host_hdgst_enable) { 3103 plen += SPDK_NVME_TCP_DIGEST_LEN; 3104 } 3105 if (pdu->hdr.common.plen != plen) { 3106 tcp_req->has_in_capsule_data = true; 3107 } else { 3108 /* Data is transmitted by C2H PDUs */ 3109 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 3110 } 3111 3112 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEED_BUFFER); 3113 break; 3114 case TCP_REQUEST_STATE_NEED_BUFFER: 3115 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEED_BUFFER, tqpair->qpair.trace_id, 0, 3116 (uintptr_t)tcp_req); 3117 3118 assert(tcp_req->req.xfer != SPDK_NVME_DATA_NONE); 3119 3120 /* Try to get a data buffer */ 3121 nvmf_tcp_req_parse_sgl(tcp_req, transport, group); 3122 break; 3123 case TCP_REQUEST_STATE_HAVE_BUFFER: 3124 spdk_trace_record(TRACE_TCP_REQUEST_STATE_HAVE_BUFFER, tqpair->qpair.trace_id, 0, 3125 (uintptr_t)tcp_req); 3126 /* Get a zcopy buffer if the request can be serviced through zcopy */ 3127 if (spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 3128 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3129 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 3130 tcp_req->req.length = tcp_req->req.dif.elba_length; 3131 } 3132 3133 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_START); 3134 spdk_nvmf_request_zcopy_start(&tcp_req->req); 3135 break; 3136 } 3137 3138 assert(tcp_req->req.iovcnt > 0); 3139 3140 /* If data is transferring from host to controller, we need to do a transfer from the host. */ 3141 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 3142 if (tcp_req->req.data_from_pool) { 3143 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 3144 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 3145 } else { 3146 struct nvme_tcp_pdu *pdu; 3147 3148 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 3149 3150 pdu = tqpair->pdu_in_progress; 3151 SPDK_DEBUGLOG(nvmf_tcp, "Not need to send r2t for tcp_req(%p) on tqpair=%p\n", tcp_req, 3152 tqpair); 3153 /* No need to send r2t, contained in the capsuled data */ 3154 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 3155 0, tcp_req->req.length); 3156 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 3157 } 3158 break; 3159 } 3160 3161 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 3162 break; 3163 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3164 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, tqpair->qpair.trace_id, 0, 3165 (uintptr_t)tcp_req); 3166 /* Some external code must kick a request into TCP_REQUEST_STATE_ZCOPY_START_COMPLETED 3167 * to escape this state. */ 3168 break; 3169 case TCP_REQUEST_STATE_ZCOPY_START_COMPLETED: 3170 spdk_trace_record(TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, tqpair->qpair.trace_id, 0, 3171 (uintptr_t)tcp_req); 3172 if (spdk_unlikely(spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl))) { 3173 SPDK_DEBUGLOG(nvmf_tcp, "Zero-copy start failed for tcp_req(%p) on tqpair=%p\n", 3174 tcp_req, tqpair); 3175 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3176 break; 3177 } 3178 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 3179 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 3180 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 3181 } else { 3182 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3183 } 3184 break; 3185 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3186 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, tqpair->qpair.trace_id, 0, 3187 (uintptr_t)tcp_req); 3188 /* The R2T completion or the h2c data incoming will kick it out of this state. */ 3189 break; 3190 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3191 3192 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, tqpair->qpair.trace_id, 3193 0, (uintptr_t)tcp_req); 3194 /* Some external code must kick a request into TCP_REQUEST_STATE_READY_TO_EXECUTE 3195 * to escape this state. */ 3196 break; 3197 case TCP_REQUEST_STATE_READY_TO_EXECUTE: 3198 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, tqpair->qpair.trace_id, 0, 3199 (uintptr_t)tcp_req); 3200 3201 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3202 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 3203 tcp_req->req.length = tcp_req->req.dif.elba_length; 3204 } 3205 3206 if (tcp_req->cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 3207 if (tcp_req->fused_failed) { 3208 /* This request failed FUSED semantics. Fail it immediately, without 3209 * even sending it to the target layer. 3210 */ 3211 nvmf_tcp_req_set_cpl(tcp_req, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_ABORTED_MISSING_FUSED); 3212 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3213 break; 3214 } 3215 3216 if (tcp_req->fused_pair == NULL || 3217 tcp_req->fused_pair->state != TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3218 /* This request is ready to execute, but either we don't know yet if it's 3219 * valid - i.e. this is a FIRST but we haven't received the next request yet), 3220 * or the other request of this fused pair isn't ready to execute. So 3221 * break here and this request will get processed later either when the 3222 * other request is ready or we find that this request isn't valid. 3223 */ 3224 break; 3225 } 3226 } 3227 3228 if (!spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 3229 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTING); 3230 /* If we get to this point, and this request is a fused command, we know that 3231 * it is part of a valid sequence (FIRST followed by a SECOND) and that both 3232 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 3233 * request, and the other request of the fused pair, in the correct order. 3234 * Also clear the ->fused_pair pointers on both requests, since after this point 3235 * we no longer need to maintain the relationship between these two requests. 3236 */ 3237 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 3238 assert(tcp_req->fused_pair != NULL); 3239 assert(tcp_req->fused_pair->fused_pair == tcp_req); 3240 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 3241 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 3242 tcp_req->fused_pair->fused_pair = NULL; 3243 tcp_req->fused_pair = NULL; 3244 } 3245 spdk_nvmf_request_exec(&tcp_req->req); 3246 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 3247 assert(tcp_req->fused_pair != NULL); 3248 assert(tcp_req->fused_pair->fused_pair == tcp_req); 3249 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 3250 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 3251 tcp_req->fused_pair->fused_pair = NULL; 3252 tcp_req->fused_pair = NULL; 3253 } 3254 } else { 3255 /* For zero-copy, only requests with data coming from host to the 3256 * controller can end up here. */ 3257 assert(tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 3258 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT); 3259 spdk_nvmf_request_zcopy_end(&tcp_req->req, true); 3260 } 3261 3262 break; 3263 case TCP_REQUEST_STATE_EXECUTING: 3264 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTING, tqpair->qpair.trace_id, 0, (uintptr_t)tcp_req); 3265 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 3266 * to escape this state. */ 3267 break; 3268 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3269 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, tqpair->qpair.trace_id, 0, 3270 (uintptr_t)tcp_req); 3271 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 3272 * to escape this state. */ 3273 break; 3274 case TCP_REQUEST_STATE_EXECUTED: 3275 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTED, tqpair->qpair.trace_id, 0, (uintptr_t)tcp_req); 3276 3277 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3278 tcp_req->req.length = tcp_req->req.dif.orig_length; 3279 } 3280 3281 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3282 break; 3283 case TCP_REQUEST_STATE_READY_TO_COMPLETE: 3284 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, tqpair->qpair.trace_id, 0, 3285 (uintptr_t)tcp_req); 3286 if (request_transfer_out(&tcp_req->req) != 0) { 3287 assert(0); /* No good way to handle this currently */ 3288 } 3289 break; 3290 case TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3291 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, tqpair->qpair.trace_id, 3292 0, (uintptr_t)tcp_req); 3293 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 3294 * to escape this state. */ 3295 break; 3296 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3297 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, tqpair->qpair.trace_id, 0, 3298 (uintptr_t)tcp_req); 3299 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 3300 * to escape this state. */ 3301 break; 3302 case TCP_REQUEST_STATE_COMPLETED: 3303 spdk_trace_record(TRACE_TCP_REQUEST_STATE_COMPLETED, tqpair->qpair.trace_id, 0, (uintptr_t)tcp_req, 3304 tqpair->qpair.queue_depth); 3305 /* If there's an outstanding PDU sent to the host, the request is completed 3306 * due to the qpair being disconnected. We must delay the completion until 3307 * that write is done to avoid freeing the request twice. */ 3308 if (spdk_unlikely(tcp_req->pdu_in_use)) { 3309 SPDK_DEBUGLOG(nvmf_tcp, "Delaying completion due to outstanding " 3310 "write on req=%p\n", tcp_req); 3311 /* This can only happen for zcopy requests */ 3312 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 3313 assert(!spdk_nvmf_qpair_is_active(&tqpair->qpair)); 3314 break; 3315 } 3316 3317 if (tcp_req->req.data_from_pool) { 3318 spdk_nvmf_request_free_buffers(&tcp_req->req, group, transport); 3319 } else if (spdk_unlikely(tcp_req->has_in_capsule_data && 3320 (tcp_req->cmd.opc == SPDK_NVME_OPC_FABRIC || 3321 tqpair->qpair.qid == 0) && tcp_req->req.length > transport->opts.in_capsule_data_size)) { 3322 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3323 assert(tgroup->control_msg_list); 3324 SPDK_DEBUGLOG(nvmf_tcp, "Put buf to control msg list\n"); 3325 nvmf_tcp_control_msg_put(tgroup->control_msg_list, 3326 tcp_req->req.iov[0].iov_base); 3327 } else if (tcp_req->req.zcopy_bdev_io != NULL) { 3328 /* If the request has an unreleased zcopy bdev_io, it's either a 3329 * read, a failed write, or the qpair is being disconnected */ 3330 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 3331 assert(tcp_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST || 3332 spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl) || 3333 !spdk_nvmf_qpair_is_active(&tqpair->qpair)); 3334 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE); 3335 spdk_nvmf_request_zcopy_end(&tcp_req->req, false); 3336 break; 3337 } 3338 tcp_req->req.length = 0; 3339 tcp_req->req.iovcnt = 0; 3340 tcp_req->fused_failed = false; 3341 if (tcp_req->fused_pair) { 3342 /* This req was part of a valid fused pair, but failed before it got to 3343 * READ_TO_EXECUTE state. This means we need to fail the other request 3344 * in the pair, because it is no longer part of a valid pair. If the pair 3345 * already reached READY_TO_EXECUTE state, we need to kick it. 3346 */ 3347 tcp_req->fused_pair->fused_failed = true; 3348 if (tcp_req->fused_pair->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3349 nvmf_tcp_req_process(ttransport, tcp_req->fused_pair); 3350 } 3351 tcp_req->fused_pair = NULL; 3352 } 3353 3354 nvmf_tcp_req_put(tqpair, tcp_req); 3355 break; 3356 case TCP_REQUEST_NUM_STATES: 3357 default: 3358 assert(0); 3359 break; 3360 } 3361 3362 if (tcp_req->state != prev_state) { 3363 progress = true; 3364 } 3365 } while (tcp_req->state != prev_state); 3366 3367 return progress; 3368 } 3369 3370 static void 3371 tcp_sock_cb(void *arg) 3372 { 3373 struct spdk_nvmf_tcp_qpair *tqpair = arg; 3374 int rc; 3375 3376 assert(tqpair != NULL); 3377 rc = nvmf_tcp_sock_process(tqpair); 3378 3379 /* If there was a new socket error, disconnect */ 3380 if (rc < 0) { 3381 nvmf_tcp_qpair_disconnect(tqpair); 3382 } 3383 } 3384 3385 static void 3386 nvmf_tcp_sock_cb(void *arg, struct spdk_sock_group *group, struct spdk_sock *sock) 3387 { 3388 tcp_sock_cb(arg); 3389 } 3390 3391 static int 3392 nvmf_tcp_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3393 struct spdk_nvmf_qpair *qpair) 3394 { 3395 struct spdk_nvmf_tcp_poll_group *tgroup; 3396 struct spdk_nvmf_tcp_qpair *tqpair; 3397 int rc; 3398 3399 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3400 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3401 3402 rc = nvmf_tcp_qpair_sock_init(tqpair); 3403 if (rc != 0) { 3404 SPDK_ERRLOG("Cannot set sock opt for tqpair=%p\n", tqpair); 3405 return -1; 3406 } 3407 3408 rc = nvmf_tcp_qpair_init(&tqpair->qpair); 3409 if (rc < 0) { 3410 SPDK_ERRLOG("Cannot init tqpair=%p\n", tqpair); 3411 return -1; 3412 } 3413 3414 rc = nvmf_tcp_qpair_init_mem_resource(tqpair); 3415 if (rc < 0) { 3416 SPDK_ERRLOG("Cannot init memory resource info for tqpair=%p\n", tqpair); 3417 return -1; 3418 } 3419 3420 rc = spdk_sock_group_add_sock(tgroup->sock_group, tqpair->sock, 3421 nvmf_tcp_sock_cb, tqpair); 3422 if (rc != 0) { 3423 SPDK_ERRLOG("Could not add sock to sock_group: %s (%d)\n", 3424 spdk_strerror(errno), errno); 3425 return -1; 3426 } 3427 3428 tqpair->group = tgroup; 3429 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INVALID); 3430 TAILQ_INSERT_TAIL(&tgroup->qpairs, tqpair, link); 3431 3432 return 0; 3433 } 3434 3435 static int 3436 nvmf_tcp_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3437 struct spdk_nvmf_qpair *qpair) 3438 { 3439 struct spdk_nvmf_tcp_poll_group *tgroup; 3440 struct spdk_nvmf_tcp_qpair *tqpair; 3441 int rc; 3442 3443 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3444 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3445 3446 assert(tqpair->group == tgroup); 3447 3448 SPDK_DEBUGLOG(nvmf_tcp, "remove tqpair=%p from the tgroup=%p\n", tqpair, tgroup); 3449 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 3450 /* Change the state to move the qpair from the await_req list to the main list 3451 * and prevent adding it again later by nvmf_tcp_qpair_set_recv_state() */ 3452 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 3453 } 3454 TAILQ_REMOVE(&tgroup->qpairs, tqpair, link); 3455 3456 /* Try to force out any pending writes */ 3457 spdk_sock_flush(tqpair->sock); 3458 3459 rc = spdk_sock_group_remove_sock(tgroup->sock_group, tqpair->sock); 3460 if (rc != 0) { 3461 SPDK_ERRLOG("Could not remove sock from sock_group: %s (%d)\n", 3462 spdk_strerror(errno), errno); 3463 } 3464 3465 return rc; 3466 } 3467 3468 static int 3469 nvmf_tcp_req_complete(struct spdk_nvmf_request *req) 3470 { 3471 struct spdk_nvmf_tcp_transport *ttransport; 3472 struct spdk_nvmf_tcp_req *tcp_req; 3473 struct spdk_nvmf_tcp_qpair *tqpair; 3474 3475 ttransport = SPDK_CONTAINEROF(req->qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3476 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3477 tqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_tcp_qpair, qpair); 3478 3479 switch (tcp_req->state) { 3480 case TCP_REQUEST_STATE_EXECUTING: 3481 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3482 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3483 break; 3484 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3485 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_ZCOPY_START_COMPLETED); 3486 break; 3487 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3488 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 3489 /* In the interrupt mode it's possible that all responses are already written out over 3490 * the socket but zero-copy buffers are still not released. In that case there won't be 3491 * any event to trigger further socket processing. Send msg to a thread to avoid deadlock. 3492 */ 3493 if (spdk_unlikely(spdk_interrupt_mode_is_enabled() && 3494 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ && 3495 spdk_nvmf_qpair_is_active(&tqpair->qpair))) { 3496 spdk_thread_send_msg(spdk_get_thread(), tcp_sock_cb, tqpair); 3497 } 3498 break; 3499 default: 3500 SPDK_ERRLOG("Unexpected request state %d (cntlid:%d, qid:%d)\n", 3501 tcp_req->state, req->qpair->ctrlr->cntlid, req->qpair->qid); 3502 assert(0 && "Unexpected request state"); 3503 break; 3504 } 3505 3506 nvmf_tcp_req_process(ttransport, tcp_req); 3507 3508 return 0; 3509 } 3510 3511 static void 3512 nvmf_tcp_close_qpair(struct spdk_nvmf_qpair *qpair, 3513 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3514 { 3515 struct spdk_nvmf_tcp_qpair *tqpair; 3516 3517 SPDK_DEBUGLOG(nvmf_tcp, "Qpair: %p\n", qpair); 3518 3519 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3520 3521 assert(tqpair->fini_cb_fn == NULL); 3522 tqpair->fini_cb_fn = cb_fn; 3523 tqpair->fini_cb_arg = cb_arg; 3524 3525 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITED); 3526 nvmf_tcp_qpair_destroy(tqpair); 3527 } 3528 3529 static int 3530 nvmf_tcp_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3531 { 3532 struct spdk_nvmf_tcp_poll_group *tgroup; 3533 int num_events, rc = 0, rc2; 3534 struct spdk_nvmf_tcp_qpair *tqpair, *tqpair_tmp; 3535 3536 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3537 3538 if (spdk_unlikely(TAILQ_EMPTY(&tgroup->qpairs) && TAILQ_EMPTY(&tgroup->await_req))) { 3539 return 0; 3540 } 3541 3542 num_events = spdk_sock_group_poll(tgroup->sock_group); 3543 if (spdk_unlikely(num_events < 0)) { 3544 SPDK_ERRLOG("Failed to poll sock_group=%p\n", tgroup->sock_group); 3545 } 3546 3547 TAILQ_FOREACH_SAFE(tqpair, &tgroup->await_req, link, tqpair_tmp) { 3548 rc2 = nvmf_tcp_sock_process(tqpair); 3549 3550 /* If there was a new socket error, disconnect */ 3551 if (spdk_unlikely(rc2 < 0)) { 3552 nvmf_tcp_qpair_disconnect(tqpair); 3553 if (rc == 0) { 3554 rc = rc2; 3555 } 3556 } 3557 } 3558 3559 return rc == 0 ? num_events : rc; 3560 } 3561 3562 static int 3563 nvmf_tcp_qpair_get_trid(struct spdk_nvmf_qpair *qpair, 3564 struct spdk_nvme_transport_id *trid, bool peer) 3565 { 3566 struct spdk_nvmf_tcp_qpair *tqpair; 3567 uint16_t port; 3568 3569 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3570 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_TCP); 3571 3572 if (peer) { 3573 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->initiator_addr); 3574 port = tqpair->initiator_port; 3575 } else { 3576 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->target_addr); 3577 port = tqpair->target_port; 3578 } 3579 3580 if (spdk_sock_is_ipv4(tqpair->sock)) { 3581 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3582 } else if (spdk_sock_is_ipv6(tqpair->sock)) { 3583 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3584 } else { 3585 return -1; 3586 } 3587 3588 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%d", port); 3589 return 0; 3590 } 3591 3592 static int 3593 nvmf_tcp_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3594 struct spdk_nvme_transport_id *trid) 3595 { 3596 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3597 } 3598 3599 static int 3600 nvmf_tcp_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3601 struct spdk_nvme_transport_id *trid) 3602 { 3603 return nvmf_tcp_qpair_get_trid(qpair, trid, 1); 3604 } 3605 3606 static int 3607 nvmf_tcp_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3608 struct spdk_nvme_transport_id *trid) 3609 { 3610 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3611 } 3612 3613 static void 3614 nvmf_tcp_req_set_abort_status(struct spdk_nvmf_request *req, 3615 struct spdk_nvmf_tcp_req *tcp_req_to_abort) 3616 { 3617 nvmf_tcp_req_set_cpl(tcp_req_to_abort, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_ABORTED_BY_REQUEST); 3618 nvmf_tcp_req_set_state(tcp_req_to_abort, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3619 3620 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 3621 } 3622 3623 static int 3624 _nvmf_tcp_qpair_abort_request(void *ctx) 3625 { 3626 struct spdk_nvmf_request *req = ctx; 3627 struct spdk_nvmf_tcp_req *tcp_req_to_abort = SPDK_CONTAINEROF(req->req_to_abort, 3628 struct spdk_nvmf_tcp_req, req); 3629 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 3630 struct spdk_nvmf_tcp_qpair, qpair); 3631 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 3632 struct spdk_nvmf_tcp_transport, transport); 3633 int rc; 3634 3635 spdk_poller_unregister(&req->poller); 3636 3637 switch (tcp_req_to_abort->state) { 3638 case TCP_REQUEST_STATE_EXECUTING: 3639 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3640 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3641 rc = nvmf_ctrlr_abort_request(req); 3642 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 3643 return SPDK_POLLER_BUSY; 3644 } 3645 break; 3646 3647 case TCP_REQUEST_STATE_NEED_BUFFER: 3648 nvmf_tcp_request_get_buffers_abort(tcp_req_to_abort); 3649 nvmf_tcp_req_set_abort_status(req, tcp_req_to_abort); 3650 nvmf_tcp_req_process(ttransport, tcp_req_to_abort); 3651 break; 3652 3653 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3654 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3655 if (spdk_get_ticks() < req->timeout_tsc) { 3656 req->poller = SPDK_POLLER_REGISTER(_nvmf_tcp_qpair_abort_request, req, 0); 3657 return SPDK_POLLER_BUSY; 3658 } 3659 break; 3660 3661 default: 3662 /* Requests in other states are either un-abortable (e.g. 3663 * TRANSFERRING_CONTROLLER_TO_HOST) or should never end up here, as they're 3664 * immediately transitioned to other states in nvmf_tcp_req_process() (e.g. 3665 * READY_TO_EXECUTE). But it is fine to end up here, as we'll simply complete the 3666 * abort request with the bit0 of dword0 set (command not aborted). 3667 */ 3668 break; 3669 } 3670 3671 spdk_nvmf_request_complete(req); 3672 return SPDK_POLLER_BUSY; 3673 } 3674 3675 static void 3676 nvmf_tcp_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 3677 struct spdk_nvmf_request *req) 3678 { 3679 struct spdk_nvmf_tcp_qpair *tqpair; 3680 struct spdk_nvmf_tcp_transport *ttransport; 3681 struct spdk_nvmf_transport *transport; 3682 uint16_t cid; 3683 uint32_t i; 3684 struct spdk_nvmf_tcp_req *tcp_req_to_abort = NULL; 3685 3686 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3687 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3688 transport = &ttransport->transport; 3689 3690 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 3691 3692 for (i = 0; i < tqpair->resource_count; i++) { 3693 if (tqpair->reqs[i].state != TCP_REQUEST_STATE_FREE && 3694 tqpair->reqs[i].req.cmd->nvme_cmd.cid == cid) { 3695 tcp_req_to_abort = &tqpair->reqs[i]; 3696 break; 3697 } 3698 } 3699 3700 spdk_trace_record(TRACE_TCP_QP_ABORT_REQ, tqpair->qpair.trace_id, 0, (uintptr_t)req); 3701 3702 if (tcp_req_to_abort == NULL) { 3703 spdk_nvmf_request_complete(req); 3704 return; 3705 } 3706 3707 req->req_to_abort = &tcp_req_to_abort->req; 3708 req->timeout_tsc = spdk_get_ticks() + 3709 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 3710 req->poller = NULL; 3711 3712 _nvmf_tcp_qpair_abort_request(req); 3713 } 3714 3715 struct tcp_subsystem_add_host_opts { 3716 char *psk; 3717 }; 3718 3719 static const struct spdk_json_object_decoder tcp_subsystem_add_host_opts_decoder[] = { 3720 {"psk", offsetof(struct tcp_subsystem_add_host_opts, psk), spdk_json_decode_string, true}, 3721 }; 3722 3723 static int 3724 tcp_load_psk(const char *fname, char *buf, size_t bufsz) 3725 { 3726 FILE *psk_file; 3727 struct stat statbuf; 3728 int rc; 3729 3730 if (stat(fname, &statbuf) != 0) { 3731 SPDK_ERRLOG("Could not read permissions for PSK file\n"); 3732 return -EACCES; 3733 } 3734 3735 if ((statbuf.st_mode & TCP_PSK_INVALID_PERMISSIONS) != 0) { 3736 SPDK_ERRLOG("Incorrect permissions for PSK file\n"); 3737 return -EPERM; 3738 } 3739 if ((size_t)statbuf.st_size > bufsz) { 3740 SPDK_ERRLOG("Invalid PSK: too long\n"); 3741 return -EINVAL; 3742 } 3743 psk_file = fopen(fname, "r"); 3744 if (psk_file == NULL) { 3745 SPDK_ERRLOG("Could not open PSK file\n"); 3746 return -EINVAL; 3747 } 3748 3749 rc = fread(buf, 1, statbuf.st_size, psk_file); 3750 if (rc != statbuf.st_size) { 3751 SPDK_ERRLOG("Failed to read PSK\n"); 3752 fclose(psk_file); 3753 return -EINVAL; 3754 } 3755 3756 fclose(psk_file); 3757 return 0; 3758 } 3759 3760 SPDK_LOG_DEPRECATION_REGISTER(nvmf_tcp_psk_path, "PSK path", "v24.09", 0); 3761 3762 static int 3763 nvmf_tcp_subsystem_add_host(struct spdk_nvmf_transport *transport, 3764 const struct spdk_nvmf_subsystem *subsystem, 3765 const char *hostnqn, 3766 const struct spdk_json_val *transport_specific) 3767 { 3768 struct tcp_subsystem_add_host_opts opts; 3769 struct spdk_nvmf_tcp_transport *ttransport; 3770 struct tcp_psk_entry *tmp, *entry = NULL; 3771 uint8_t psk_configured[SPDK_TLS_PSK_MAX_LEN] = {}; 3772 char psk_interchange[SPDK_TLS_PSK_MAX_LEN + 1] = {}; 3773 uint8_t tls_cipher_suite; 3774 int rc = 0; 3775 uint8_t psk_retained_hash; 3776 uint64_t psk_configured_size; 3777 3778 if (transport_specific == NULL) { 3779 return 0; 3780 } 3781 3782 assert(transport != NULL); 3783 assert(subsystem != NULL); 3784 3785 memset(&opts, 0, sizeof(opts)); 3786 3787 /* Decode PSK (either name of a key or file path) */ 3788 if (spdk_json_decode_object_relaxed(transport_specific, tcp_subsystem_add_host_opts_decoder, 3789 SPDK_COUNTOF(tcp_subsystem_add_host_opts_decoder), &opts)) { 3790 SPDK_ERRLOG("spdk_json_decode_object failed\n"); 3791 return -EINVAL; 3792 } 3793 3794 if (opts.psk == NULL) { 3795 return 0; 3796 } 3797 3798 entry = calloc(1, sizeof(struct tcp_psk_entry)); 3799 if (entry == NULL) { 3800 SPDK_ERRLOG("Unable to allocate memory for PSK entry!\n"); 3801 rc = -ENOMEM; 3802 goto end; 3803 } 3804 3805 entry->key = spdk_keyring_get_key(opts.psk); 3806 if (entry->key != NULL) { 3807 rc = spdk_key_get_key(entry->key, psk_interchange, SPDK_TLS_PSK_MAX_LEN); 3808 if (rc < 0) { 3809 SPDK_ERRLOG("Failed to retrieve PSK '%s'\n", opts.psk); 3810 rc = -EINVAL; 3811 goto end; 3812 } 3813 } else { 3814 if (strlen(opts.psk) >= sizeof(entry->psk)) { 3815 SPDK_ERRLOG("PSK path too long\n"); 3816 rc = -EINVAL; 3817 goto end; 3818 } 3819 3820 rc = tcp_load_psk(opts.psk, psk_interchange, SPDK_TLS_PSK_MAX_LEN); 3821 if (rc) { 3822 SPDK_ERRLOG("Could not retrieve PSK from file\n"); 3823 goto end; 3824 } 3825 3826 SPDK_LOG_DEPRECATED(nvmf_tcp_psk_path); 3827 } 3828 3829 /* Parse PSK interchange to get length of base64 encoded data. 3830 * This is then used to decide which cipher suite should be used 3831 * to generate PSK identity and TLS PSK later on. */ 3832 rc = nvme_tcp_parse_interchange_psk(psk_interchange, psk_configured, sizeof(psk_configured), 3833 &psk_configured_size, &psk_retained_hash); 3834 if (rc < 0) { 3835 SPDK_ERRLOG("Failed to parse PSK interchange!\n"); 3836 goto end; 3837 } 3838 3839 /* The Base64 string encodes the configured PSK (32 or 48 bytes binary). 3840 * This check also ensures that psk_configured_size is smaller than 3841 * psk_retained buffer size. */ 3842 if (psk_configured_size == SHA256_DIGEST_LENGTH) { 3843 tls_cipher_suite = NVME_TCP_CIPHER_AES_128_GCM_SHA256; 3844 } else if (psk_configured_size == SHA384_DIGEST_LENGTH) { 3845 tls_cipher_suite = NVME_TCP_CIPHER_AES_256_GCM_SHA384; 3846 } else { 3847 SPDK_ERRLOG("Unrecognized cipher suite!\n"); 3848 rc = -EINVAL; 3849 goto end; 3850 } 3851 3852 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3853 /* Generate PSK identity. */ 3854 rc = nvme_tcp_generate_psk_identity(entry->pskid, sizeof(entry->pskid), hostnqn, 3855 subsystem->subnqn, tls_cipher_suite); 3856 if (rc) { 3857 rc = -EINVAL; 3858 goto end; 3859 } 3860 /* Check if PSK identity entry already exists. */ 3861 TAILQ_FOREACH(tmp, &ttransport->psks, link) { 3862 if (strncmp(tmp->pskid, entry->pskid, NVMF_PSK_IDENTITY_LEN) == 0) { 3863 SPDK_ERRLOG("Given PSK identity: %s entry already exists!\n", entry->pskid); 3864 rc = -EEXIST; 3865 goto end; 3866 } 3867 } 3868 3869 if (snprintf(entry->hostnqn, sizeof(entry->hostnqn), "%s", hostnqn) < 0) { 3870 SPDK_ERRLOG("Could not write hostnqn string!\n"); 3871 rc = -EINVAL; 3872 goto end; 3873 } 3874 if (snprintf(entry->subnqn, sizeof(entry->subnqn), "%s", subsystem->subnqn) < 0) { 3875 SPDK_ERRLOG("Could not write subnqn string!\n"); 3876 rc = -EINVAL; 3877 goto end; 3878 } 3879 3880 entry->tls_cipher_suite = tls_cipher_suite; 3881 3882 /* No hash indicates that Configured PSK must be used as Retained PSK. */ 3883 if (psk_retained_hash == NVME_TCP_HASH_ALGORITHM_NONE) { 3884 /* Psk configured is either 32 or 48 bytes long. */ 3885 memcpy(entry->psk, psk_configured, psk_configured_size); 3886 entry->psk_size = psk_configured_size; 3887 } else { 3888 /* Derive retained PSK. */ 3889 rc = nvme_tcp_derive_retained_psk(psk_configured, psk_configured_size, hostnqn, entry->psk, 3890 SPDK_TLS_PSK_MAX_LEN, psk_retained_hash); 3891 if (rc < 0) { 3892 SPDK_ERRLOG("Unable to derive retained PSK!\n"); 3893 goto end; 3894 } 3895 entry->psk_size = rc; 3896 } 3897 3898 if (entry->key == NULL) { 3899 rc = snprintf(entry->psk_path, sizeof(entry->psk_path), "%s", opts.psk); 3900 if (rc < 0 || (size_t)rc >= sizeof(entry->psk_path)) { 3901 SPDK_ERRLOG("Could not save PSK path!\n"); 3902 rc = -ENAMETOOLONG; 3903 goto end; 3904 } 3905 } 3906 3907 TAILQ_INSERT_TAIL(&ttransport->psks, entry, link); 3908 rc = 0; 3909 3910 end: 3911 spdk_memset_s(psk_configured, sizeof(psk_configured), 0, sizeof(psk_configured)); 3912 spdk_memset_s(psk_interchange, sizeof(psk_interchange), 0, sizeof(psk_interchange)); 3913 3914 free(opts.psk); 3915 if (rc != 0) { 3916 nvmf_tcp_free_psk_entry(entry); 3917 } 3918 3919 return rc; 3920 } 3921 3922 static void 3923 nvmf_tcp_subsystem_remove_host(struct spdk_nvmf_transport *transport, 3924 const struct spdk_nvmf_subsystem *subsystem, 3925 const char *hostnqn) 3926 { 3927 struct spdk_nvmf_tcp_transport *ttransport; 3928 struct tcp_psk_entry *entry, *tmp; 3929 3930 assert(transport != NULL); 3931 assert(subsystem != NULL); 3932 3933 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3934 TAILQ_FOREACH_SAFE(entry, &ttransport->psks, link, tmp) { 3935 if ((strncmp(entry->hostnqn, hostnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0 && 3936 (strncmp(entry->subnqn, subsystem->subnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0) { 3937 TAILQ_REMOVE(&ttransport->psks, entry, link); 3938 nvmf_tcp_free_psk_entry(entry); 3939 break; 3940 } 3941 } 3942 } 3943 3944 static void 3945 nvmf_tcp_subsystem_dump_host(struct spdk_nvmf_transport *transport, 3946 const struct spdk_nvmf_subsystem *subsystem, const char *hostnqn, 3947 struct spdk_json_write_ctx *w) 3948 { 3949 struct spdk_nvmf_tcp_transport *ttransport; 3950 struct tcp_psk_entry *entry; 3951 3952 assert(transport != NULL); 3953 assert(subsystem != NULL); 3954 3955 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3956 TAILQ_FOREACH(entry, &ttransport->psks, link) { 3957 if ((strncmp(entry->hostnqn, hostnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0 && 3958 (strncmp(entry->subnqn, subsystem->subnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0) { 3959 spdk_json_write_named_string(w, "psk", entry->key ? 3960 spdk_key_get_name(entry->key) : entry->psk_path); 3961 break; 3962 } 3963 } 3964 } 3965 3966 static void 3967 nvmf_tcp_opts_init(struct spdk_nvmf_transport_opts *opts) 3968 { 3969 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH; 3970 opts->max_qpairs_per_ctrlr = SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR; 3971 opts->in_capsule_data_size = SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE; 3972 opts->max_io_size = SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE; 3973 opts->io_unit_size = SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE; 3974 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH; 3975 opts->num_shared_buffers = SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS; 3976 opts->buf_cache_size = SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE; 3977 opts->dif_insert_or_strip = SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP; 3978 opts->abort_timeout_sec = SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC; 3979 opts->transport_specific = NULL; 3980 } 3981 3982 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp = { 3983 .name = "TCP", 3984 .type = SPDK_NVME_TRANSPORT_TCP, 3985 .opts_init = nvmf_tcp_opts_init, 3986 .create = nvmf_tcp_create, 3987 .dump_opts = nvmf_tcp_dump_opts, 3988 .destroy = nvmf_tcp_destroy, 3989 3990 .listen = nvmf_tcp_listen, 3991 .stop_listen = nvmf_tcp_stop_listen, 3992 3993 .listener_discover = nvmf_tcp_discover, 3994 3995 .poll_group_create = nvmf_tcp_poll_group_create, 3996 .get_optimal_poll_group = nvmf_tcp_get_optimal_poll_group, 3997 .poll_group_destroy = nvmf_tcp_poll_group_destroy, 3998 .poll_group_add = nvmf_tcp_poll_group_add, 3999 .poll_group_remove = nvmf_tcp_poll_group_remove, 4000 .poll_group_poll = nvmf_tcp_poll_group_poll, 4001 4002 .req_free = nvmf_tcp_req_free, 4003 .req_complete = nvmf_tcp_req_complete, 4004 .req_get_buffers_done = nvmf_tcp_req_get_buffers_done, 4005 4006 .qpair_fini = nvmf_tcp_close_qpair, 4007 .qpair_get_local_trid = nvmf_tcp_qpair_get_local_trid, 4008 .qpair_get_peer_trid = nvmf_tcp_qpair_get_peer_trid, 4009 .qpair_get_listen_trid = nvmf_tcp_qpair_get_listen_trid, 4010 .qpair_abort_request = nvmf_tcp_qpair_abort_request, 4011 .subsystem_add_host = nvmf_tcp_subsystem_add_host, 4012 .subsystem_remove_host = nvmf_tcp_subsystem_remove_host, 4013 .subsystem_dump_host = nvmf_tcp_subsystem_dump_host, 4014 }; 4015 4016 SPDK_NVMF_TRANSPORT_REGISTER(tcp, &spdk_nvmf_transport_tcp); 4017 SPDK_LOG_REGISTER_COMPONENT(nvmf_tcp) 4018