1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (C) 2018 Intel Corporation. All rights reserved. 3 * Copyright (c) 2019, 2020 Mellanox Technologies LTD. All rights reserved. 4 * Copyright (c) 2022, 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. 5 */ 6 7 #include "spdk/accel.h" 8 #include "spdk/stdinc.h" 9 #include "spdk/crc32.h" 10 #include "spdk/endian.h" 11 #include "spdk/assert.h" 12 #include "spdk/thread.h" 13 #include "spdk/nvmf_transport.h" 14 #include "spdk/string.h" 15 #include "spdk/trace.h" 16 #include "spdk/util.h" 17 #include "spdk/log.h" 18 19 #include "spdk_internal/assert.h" 20 #include "spdk_internal/nvme_tcp.h" 21 #include "spdk_internal/sock.h" 22 23 #include "nvmf_internal.h" 24 25 #include "spdk_internal/trace_defs.h" 26 27 #define NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME 16 28 #define SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY 16 29 #define SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY 0 30 #define SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM 32 31 #define SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION true 32 33 #define SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH 2 34 #define SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH 65535 35 #define SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH 2 36 #define SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH 4096 37 38 #define SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH 128 39 #define SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH 128 40 #define SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 41 #define SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 42 #define SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE 131072 43 #define SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE 131072 44 #define SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS 511 45 #define SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX 46 #define SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP false 47 #define SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC 1 48 49 #define TCP_PSK_INVALID_PERMISSIONS 0177 50 51 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp; 52 static bool g_tls_log = false; 53 54 /* spdk nvmf related structure */ 55 enum spdk_nvmf_tcp_req_state { 56 57 /* The request is not currently in use */ 58 TCP_REQUEST_STATE_FREE = 0, 59 60 /* Initial state when request first received */ 61 TCP_REQUEST_STATE_NEW = 1, 62 63 /* The request is queued until a data buffer is available. */ 64 TCP_REQUEST_STATE_NEED_BUFFER = 2, 65 66 /* The request is waiting for zcopy_start to finish */ 67 TCP_REQUEST_STATE_AWAITING_ZCOPY_START = 3, 68 69 /* The request has received a zero-copy buffer */ 70 TCP_REQUEST_STATE_ZCOPY_START_COMPLETED = 4, 71 72 /* The request is currently transferring data from the host to the controller. */ 73 TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER = 5, 74 75 /* The request is waiting for the R2T send acknowledgement. */ 76 TCP_REQUEST_STATE_AWAITING_R2T_ACK = 6, 77 78 /* The request is ready to execute at the block device */ 79 TCP_REQUEST_STATE_READY_TO_EXECUTE = 7, 80 81 /* The request is currently executing at the block device */ 82 TCP_REQUEST_STATE_EXECUTING = 8, 83 84 /* The request is waiting for zcopy buffers to be committed */ 85 TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT = 9, 86 87 /* The request finished executing at the block device */ 88 TCP_REQUEST_STATE_EXECUTED = 10, 89 90 /* The request is ready to send a completion */ 91 TCP_REQUEST_STATE_READY_TO_COMPLETE = 11, 92 93 /* The request is currently transferring final pdus from the controller to the host. */ 94 TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST = 12, 95 96 /* The request is waiting for zcopy buffers to be released (without committing) */ 97 TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE = 13, 98 99 /* The request completed and can be marked free. */ 100 TCP_REQUEST_STATE_COMPLETED = 14, 101 102 /* Terminator */ 103 TCP_REQUEST_NUM_STATES, 104 }; 105 106 static const char *spdk_nvmf_tcp_term_req_fes_str[] = { 107 "Invalid PDU Header Field", 108 "PDU Sequence Error", 109 "Header Digiest Error", 110 "Data Transfer Out of Range", 111 "R2T Limit Exceeded", 112 "Unsupported parameter", 113 }; 114 115 SPDK_TRACE_REGISTER_FN(nvmf_tcp_trace, "nvmf_tcp", TRACE_GROUP_NVMF_TCP) 116 { 117 spdk_trace_register_owner(OWNER_NVMF_TCP, 't'); 118 spdk_trace_register_object(OBJECT_NVMF_TCP_IO, 'r'); 119 spdk_trace_register_description("TCP_REQ_NEW", 120 TRACE_TCP_REQUEST_STATE_NEW, 121 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 1, 122 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 123 spdk_trace_register_description("TCP_REQ_NEED_BUFFER", 124 TRACE_TCP_REQUEST_STATE_NEED_BUFFER, 125 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 126 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 127 spdk_trace_register_description("TCP_REQ_WAIT_ZCPY_START", 128 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, 129 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 130 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 131 spdk_trace_register_description("TCP_REQ_ZCPY_START_CPL", 132 TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, 133 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 134 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 135 spdk_trace_register_description("TCP_REQ_TX_H_TO_C", 136 TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 137 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 138 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 139 spdk_trace_register_description("TCP_REQ_RDY_TO_EXECUTE", 140 TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, 141 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 142 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 143 spdk_trace_register_description("TCP_REQ_EXECUTING", 144 TRACE_TCP_REQUEST_STATE_EXECUTING, 145 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 146 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 147 spdk_trace_register_description("TCP_REQ_WAIT_ZCPY_CMT", 148 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, 149 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 150 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 151 spdk_trace_register_description("TCP_REQ_EXECUTED", 152 TRACE_TCP_REQUEST_STATE_EXECUTED, 153 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 154 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 155 spdk_trace_register_description("TCP_REQ_RDY_TO_COMPLETE", 156 TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, 157 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 158 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 159 spdk_trace_register_description("TCP_REQ_TRANSFER_C2H", 160 TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 161 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 162 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 163 spdk_trace_register_description("TCP_REQ_AWAIT_ZCPY_RLS", 164 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, 165 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 166 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 167 spdk_trace_register_description("TCP_REQ_COMPLETED", 168 TRACE_TCP_REQUEST_STATE_COMPLETED, 169 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 170 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 171 spdk_trace_register_description("TCP_WRITE_START", 172 TRACE_TCP_FLUSH_WRITEBUF_START, 173 OWNER_NVMF_TCP, OBJECT_NONE, 0, 174 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 175 spdk_trace_register_description("TCP_WRITE_DONE", 176 TRACE_TCP_FLUSH_WRITEBUF_DONE, 177 OWNER_NVMF_TCP, OBJECT_NONE, 0, 178 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 179 spdk_trace_register_description("TCP_READ_DONE", 180 TRACE_TCP_READ_FROM_SOCKET_DONE, 181 OWNER_NVMF_TCP, OBJECT_NONE, 0, 182 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 183 spdk_trace_register_description("TCP_REQ_AWAIT_R2T_ACK", 184 TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, 185 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 186 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 187 188 spdk_trace_register_description("TCP_QP_CREATE", TRACE_TCP_QP_CREATE, 189 OWNER_NVMF_TCP, OBJECT_NONE, 0, 190 SPDK_TRACE_ARG_TYPE_INT, ""); 191 spdk_trace_register_description("TCP_QP_SOCK_INIT", TRACE_TCP_QP_SOCK_INIT, 192 OWNER_NVMF_TCP, OBJECT_NONE, 0, 193 SPDK_TRACE_ARG_TYPE_INT, ""); 194 spdk_trace_register_description("TCP_QP_STATE_CHANGE", TRACE_TCP_QP_STATE_CHANGE, 195 OWNER_NVMF_TCP, OBJECT_NONE, 0, 196 SPDK_TRACE_ARG_TYPE_INT, "state"); 197 spdk_trace_register_description("TCP_QP_DISCONNECT", TRACE_TCP_QP_DISCONNECT, 198 OWNER_NVMF_TCP, OBJECT_NONE, 0, 199 SPDK_TRACE_ARG_TYPE_INT, ""); 200 spdk_trace_register_description("TCP_QP_DESTROY", TRACE_TCP_QP_DESTROY, 201 OWNER_NVMF_TCP, OBJECT_NONE, 0, 202 SPDK_TRACE_ARG_TYPE_INT, ""); 203 spdk_trace_register_description("TCP_QP_ABORT_REQ", TRACE_TCP_QP_ABORT_REQ, 204 OWNER_NVMF_TCP, OBJECT_NONE, 0, 205 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 206 spdk_trace_register_description("TCP_QP_RCV_STATE_CHANGE", TRACE_TCP_QP_RCV_STATE_CHANGE, 207 OWNER_NVMF_TCP, OBJECT_NONE, 0, 208 SPDK_TRACE_ARG_TYPE_INT, "state"); 209 210 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_TCP_IO, 1); 211 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_TCP_IO, 0); 212 } 213 214 struct spdk_nvmf_tcp_req { 215 struct spdk_nvmf_request req; 216 struct spdk_nvme_cpl rsp; 217 struct spdk_nvme_cmd cmd; 218 219 /* A PDU that can be used for sending responses. This is 220 * not the incoming PDU! */ 221 struct nvme_tcp_pdu *pdu; 222 223 /* In-capsule data buffer */ 224 uint8_t *buf; 225 226 struct spdk_nvmf_tcp_req *fused_pair; 227 228 /* 229 * The PDU for a request may be used multiple times in serial over 230 * the request's lifetime. For example, first to send an R2T, then 231 * to send a completion. To catch mistakes where the PDU is used 232 * twice at the same time, add a debug flag here for init/fini. 233 */ 234 bool pdu_in_use; 235 bool has_in_capsule_data; 236 bool fused_failed; 237 238 /* transfer_tag */ 239 uint16_t ttag; 240 241 enum spdk_nvmf_tcp_req_state state; 242 243 /* 244 * h2c_offset is used when we receive the h2c_data PDU. 245 */ 246 uint32_t h2c_offset; 247 248 STAILQ_ENTRY(spdk_nvmf_tcp_req) link; 249 TAILQ_ENTRY(spdk_nvmf_tcp_req) state_link; 250 }; 251 252 struct spdk_nvmf_tcp_qpair { 253 struct spdk_nvmf_qpair qpair; 254 struct spdk_nvmf_tcp_poll_group *group; 255 struct spdk_sock *sock; 256 257 enum nvme_tcp_pdu_recv_state recv_state; 258 enum nvme_tcp_qpair_state state; 259 260 /* PDU being actively received */ 261 struct nvme_tcp_pdu *pdu_in_progress; 262 263 struct spdk_nvmf_tcp_req *fused_first; 264 265 /* Queues to track the requests in all states */ 266 TAILQ_HEAD(, spdk_nvmf_tcp_req) tcp_req_working_queue; 267 TAILQ_HEAD(, spdk_nvmf_tcp_req) tcp_req_free_queue; 268 SLIST_HEAD(, nvme_tcp_pdu) tcp_pdu_free_queue; 269 /* Number of working pdus */ 270 uint32_t tcp_pdu_working_count; 271 272 /* Number of requests in each state */ 273 uint32_t state_cntr[TCP_REQUEST_NUM_STATES]; 274 275 uint8_t cpda; 276 277 bool host_hdgst_enable; 278 bool host_ddgst_enable; 279 280 /* This is a spare PDU used for sending special management 281 * operations. Primarily, this is used for the initial 282 * connection response and c2h termination request. */ 283 struct nvme_tcp_pdu *mgmt_pdu; 284 285 /* Arrays of in-capsule buffers, requests, and pdus. 286 * Each array is 'resource_count' number of elements */ 287 void *bufs; 288 struct spdk_nvmf_tcp_req *reqs; 289 struct nvme_tcp_pdu *pdus; 290 uint32_t resource_count; 291 uint32_t recv_buf_size; 292 293 struct spdk_nvmf_tcp_port *port; 294 295 /* IP address */ 296 char initiator_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 297 char target_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 298 299 /* IP port */ 300 uint16_t initiator_port; 301 uint16_t target_port; 302 303 /* Wait until the host terminates the connection (e.g. after sending C2HTermReq) */ 304 bool wait_terminate; 305 306 /* Timer used to destroy qpair after detecting transport error issue if initiator does 307 * not close the connection. 308 */ 309 struct spdk_poller *timeout_poller; 310 311 spdk_nvmf_transport_qpair_fini_cb fini_cb_fn; 312 void *fini_cb_arg; 313 314 TAILQ_ENTRY(spdk_nvmf_tcp_qpair) link; 315 }; 316 317 struct spdk_nvmf_tcp_control_msg { 318 STAILQ_ENTRY(spdk_nvmf_tcp_control_msg) link; 319 }; 320 321 struct spdk_nvmf_tcp_control_msg_list { 322 void *msg_buf; 323 STAILQ_HEAD(, spdk_nvmf_tcp_control_msg) free_msgs; 324 }; 325 326 struct spdk_nvmf_tcp_poll_group { 327 struct spdk_nvmf_transport_poll_group group; 328 struct spdk_sock_group *sock_group; 329 330 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) qpairs; 331 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) await_req; 332 333 struct spdk_io_channel *accel_channel; 334 struct spdk_nvmf_tcp_control_msg_list *control_msg_list; 335 336 TAILQ_ENTRY(spdk_nvmf_tcp_poll_group) link; 337 }; 338 339 struct spdk_nvmf_tcp_port { 340 const struct spdk_nvme_transport_id *trid; 341 struct spdk_sock *listen_sock; 342 TAILQ_ENTRY(spdk_nvmf_tcp_port) link; 343 }; 344 345 struct tcp_transport_opts { 346 bool c2h_success; 347 uint16_t control_msg_num; 348 uint32_t sock_priority; 349 }; 350 351 struct tcp_psk_entry { 352 char hostnqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 353 char subnqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 354 char psk_identity[NVMF_PSK_IDENTITY_LEN]; 355 uint8_t psk[SPDK_TLS_PSK_MAX_LEN]; 356 357 /* Original path saved to emit SPDK configuration via "save_config". */ 358 char psk_path[PATH_MAX]; 359 uint32_t psk_size; 360 enum nvme_tcp_cipher_suite tls_cipher_suite; 361 TAILQ_ENTRY(tcp_psk_entry) link; 362 }; 363 364 struct spdk_nvmf_tcp_transport { 365 struct spdk_nvmf_transport transport; 366 struct tcp_transport_opts tcp_opts; 367 368 struct spdk_nvmf_tcp_poll_group *next_pg; 369 370 struct spdk_poller *accept_poller; 371 372 TAILQ_HEAD(, spdk_nvmf_tcp_port) ports; 373 TAILQ_HEAD(, spdk_nvmf_tcp_poll_group) poll_groups; 374 375 TAILQ_HEAD(, tcp_psk_entry) psks; 376 }; 377 378 static const struct spdk_json_object_decoder tcp_transport_opts_decoder[] = { 379 { 380 "c2h_success", offsetof(struct tcp_transport_opts, c2h_success), 381 spdk_json_decode_bool, true 382 }, 383 { 384 "control_msg_num", offsetof(struct tcp_transport_opts, control_msg_num), 385 spdk_json_decode_uint16, true 386 }, 387 { 388 "sock_priority", offsetof(struct tcp_transport_opts, sock_priority), 389 spdk_json_decode_uint32, true 390 }, 391 }; 392 393 static bool nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 394 struct spdk_nvmf_tcp_req *tcp_req); 395 static void nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 396 397 static void _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 398 struct spdk_nvmf_tcp_req *tcp_req); 399 400 static inline void 401 nvmf_tcp_req_set_state(struct spdk_nvmf_tcp_req *tcp_req, 402 enum spdk_nvmf_tcp_req_state state) 403 { 404 struct spdk_nvmf_qpair *qpair; 405 struct spdk_nvmf_tcp_qpair *tqpair; 406 407 qpair = tcp_req->req.qpair; 408 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 409 410 assert(tqpair->state_cntr[tcp_req->state] > 0); 411 tqpair->state_cntr[tcp_req->state]--; 412 tqpair->state_cntr[state]++; 413 414 tcp_req->state = state; 415 } 416 417 static inline struct nvme_tcp_pdu * 418 nvmf_tcp_req_pdu_init(struct spdk_nvmf_tcp_req *tcp_req) 419 { 420 assert(tcp_req->pdu_in_use == false); 421 422 memset(tcp_req->pdu, 0, sizeof(*tcp_req->pdu)); 423 tcp_req->pdu->qpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 424 425 return tcp_req->pdu; 426 } 427 428 static struct spdk_nvmf_tcp_req * 429 nvmf_tcp_req_get(struct spdk_nvmf_tcp_qpair *tqpair) 430 { 431 struct spdk_nvmf_tcp_req *tcp_req; 432 433 tcp_req = TAILQ_FIRST(&tqpair->tcp_req_free_queue); 434 if (spdk_unlikely(!tcp_req)) { 435 return NULL; 436 } 437 438 memset(&tcp_req->rsp, 0, sizeof(tcp_req->rsp)); 439 tcp_req->h2c_offset = 0; 440 tcp_req->has_in_capsule_data = false; 441 tcp_req->req.dif_enabled = false; 442 tcp_req->req.zcopy_phase = NVMF_ZCOPY_PHASE_NONE; 443 444 TAILQ_REMOVE(&tqpair->tcp_req_free_queue, tcp_req, state_link); 445 TAILQ_INSERT_TAIL(&tqpair->tcp_req_working_queue, tcp_req, state_link); 446 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEW); 447 return tcp_req; 448 } 449 450 static inline void 451 nvmf_tcp_req_put(struct spdk_nvmf_tcp_qpair *tqpair, struct spdk_nvmf_tcp_req *tcp_req) 452 { 453 assert(!tcp_req->pdu_in_use); 454 455 TAILQ_REMOVE(&tqpair->tcp_req_working_queue, tcp_req, state_link); 456 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 457 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_FREE); 458 } 459 460 static void 461 nvmf_tcp_request_free(void *cb_arg) 462 { 463 struct spdk_nvmf_tcp_transport *ttransport; 464 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 465 466 assert(tcp_req != NULL); 467 468 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req=%p will be freed\n", tcp_req); 469 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 470 struct spdk_nvmf_tcp_transport, transport); 471 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 472 nvmf_tcp_req_process(ttransport, tcp_req); 473 } 474 475 static int 476 nvmf_tcp_req_free(struct spdk_nvmf_request *req) 477 { 478 struct spdk_nvmf_tcp_req *tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 479 480 nvmf_tcp_request_free(tcp_req); 481 482 return 0; 483 } 484 485 static void 486 nvmf_tcp_drain_state_queue(struct spdk_nvmf_tcp_qpair *tqpair, 487 enum spdk_nvmf_tcp_req_state state) 488 { 489 struct spdk_nvmf_tcp_req *tcp_req, *req_tmp; 490 491 assert(state != TCP_REQUEST_STATE_FREE); 492 TAILQ_FOREACH_SAFE(tcp_req, &tqpair->tcp_req_working_queue, state_link, req_tmp) { 493 if (state == tcp_req->state) { 494 nvmf_tcp_request_free(tcp_req); 495 } 496 } 497 } 498 499 static void 500 nvmf_tcp_cleanup_all_states(struct spdk_nvmf_tcp_qpair *tqpair) 501 { 502 struct spdk_nvmf_tcp_req *tcp_req, *req_tmp; 503 504 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 505 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_NEW); 506 507 /* Wipe the requests waiting for buffer from the global list */ 508 TAILQ_FOREACH_SAFE(tcp_req, &tqpair->tcp_req_working_queue, state_link, req_tmp) { 509 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 510 STAILQ_REMOVE(&tqpair->group->group.pending_buf_queue, &tcp_req->req, 511 spdk_nvmf_request, buf_link); 512 } 513 } 514 515 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_NEED_BUFFER); 516 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_EXECUTING); 517 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 518 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 519 } 520 521 static void 522 nvmf_tcp_dump_qpair_req_contents(struct spdk_nvmf_tcp_qpair *tqpair) 523 { 524 int i; 525 struct spdk_nvmf_tcp_req *tcp_req; 526 527 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", tqpair->qpair.qid); 528 for (i = 1; i < TCP_REQUEST_NUM_STATES; i++) { 529 SPDK_ERRLOG("\tNum of requests in state[%d] = %u\n", i, tqpair->state_cntr[i]); 530 TAILQ_FOREACH(tcp_req, &tqpair->tcp_req_working_queue, state_link) { 531 if ((int)tcp_req->state == i) { 532 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", tcp_req->req.data_from_pool); 533 SPDK_ERRLOG("\t\tRequest opcode: %d\n", tcp_req->req.cmd->nvmf_cmd.opcode); 534 } 535 } 536 } 537 } 538 539 static void 540 _nvmf_tcp_qpair_destroy(void *_tqpair) 541 { 542 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 543 spdk_nvmf_transport_qpair_fini_cb cb_fn = tqpair->fini_cb_fn; 544 void *cb_arg = tqpair->fini_cb_arg; 545 int err = 0; 546 547 spdk_trace_record(TRACE_TCP_QP_DESTROY, 0, 0, (uintptr_t)tqpair); 548 549 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 550 551 err = spdk_sock_close(&tqpair->sock); 552 assert(err == 0); 553 nvmf_tcp_cleanup_all_states(tqpair); 554 555 if (tqpair->state_cntr[TCP_REQUEST_STATE_FREE] != tqpair->resource_count) { 556 SPDK_ERRLOG("tqpair(%p) free tcp request num is %u but should be %u\n", tqpair, 557 tqpair->state_cntr[TCP_REQUEST_STATE_FREE], 558 tqpair->resource_count); 559 err++; 560 } 561 562 if (err > 0) { 563 nvmf_tcp_dump_qpair_req_contents(tqpair); 564 } 565 566 /* The timeout poller might still be registered here if we close the qpair before host 567 * terminates the connection. 568 */ 569 spdk_poller_unregister(&tqpair->timeout_poller); 570 spdk_dma_free(tqpair->pdus); 571 free(tqpair->reqs); 572 spdk_free(tqpair->bufs); 573 free(tqpair); 574 575 if (cb_fn != NULL) { 576 cb_fn(cb_arg); 577 } 578 579 SPDK_DEBUGLOG(nvmf_tcp, "Leave\n"); 580 } 581 582 static void 583 nvmf_tcp_qpair_destroy(struct spdk_nvmf_tcp_qpair *tqpair) 584 { 585 /* Delay the destruction to make sure it isn't performed from the context of a sock 586 * callback. Otherwise, spdk_sock_close() might not abort pending requests, causing their 587 * completions to be executed after the qpair is freed. (Note: this fixed issue #2471.) 588 */ 589 spdk_thread_send_msg(spdk_get_thread(), _nvmf_tcp_qpair_destroy, tqpair); 590 } 591 592 static void 593 nvmf_tcp_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 594 { 595 struct spdk_nvmf_tcp_transport *ttransport; 596 assert(w != NULL); 597 598 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 599 spdk_json_write_named_bool(w, "c2h_success", ttransport->tcp_opts.c2h_success); 600 spdk_json_write_named_uint32(w, "sock_priority", ttransport->tcp_opts.sock_priority); 601 } 602 603 static int 604 nvmf_tcp_destroy(struct spdk_nvmf_transport *transport, 605 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 606 { 607 struct spdk_nvmf_tcp_transport *ttransport; 608 struct tcp_psk_entry *entry, *tmp; 609 610 assert(transport != NULL); 611 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 612 613 TAILQ_FOREACH_SAFE(entry, &ttransport->psks, link, tmp) { 614 TAILQ_REMOVE(&ttransport->psks, entry, link); 615 free(entry); 616 } 617 618 spdk_poller_unregister(&ttransport->accept_poller); 619 free(ttransport); 620 621 if (cb_fn) { 622 cb_fn(cb_arg); 623 } 624 return 0; 625 } 626 627 static int nvmf_tcp_accept(void *ctx); 628 629 static struct spdk_nvmf_transport * 630 nvmf_tcp_create(struct spdk_nvmf_transport_opts *opts) 631 { 632 struct spdk_nvmf_tcp_transport *ttransport; 633 uint32_t sge_count; 634 uint32_t min_shared_buffers; 635 636 ttransport = calloc(1, sizeof(*ttransport)); 637 if (!ttransport) { 638 return NULL; 639 } 640 641 TAILQ_INIT(&ttransport->ports); 642 TAILQ_INIT(&ttransport->poll_groups); 643 TAILQ_INIT(&ttransport->psks); 644 645 ttransport->transport.ops = &spdk_nvmf_transport_tcp; 646 647 ttransport->tcp_opts.c2h_success = SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION; 648 ttransport->tcp_opts.sock_priority = SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY; 649 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 650 if (opts->transport_specific != NULL && 651 spdk_json_decode_object_relaxed(opts->transport_specific, tcp_transport_opts_decoder, 652 SPDK_COUNTOF(tcp_transport_opts_decoder), 653 &ttransport->tcp_opts)) { 654 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 655 free(ttransport); 656 return NULL; 657 } 658 659 SPDK_NOTICELOG("*** TCP Transport Init ***\n"); 660 661 SPDK_INFOLOG(nvmf_tcp, "*** TCP Transport Init ***\n" 662 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 663 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 664 " in_capsule_data_size=%d, max_aq_depth=%d\n" 665 " num_shared_buffers=%d, c2h_success=%d,\n" 666 " dif_insert_or_strip=%d, sock_priority=%d\n" 667 " abort_timeout_sec=%d, control_msg_num=%hu\n", 668 opts->max_queue_depth, 669 opts->max_io_size, 670 opts->max_qpairs_per_ctrlr - 1, 671 opts->io_unit_size, 672 opts->in_capsule_data_size, 673 opts->max_aq_depth, 674 opts->num_shared_buffers, 675 ttransport->tcp_opts.c2h_success, 676 opts->dif_insert_or_strip, 677 ttransport->tcp_opts.sock_priority, 678 opts->abort_timeout_sec, 679 ttransport->tcp_opts.control_msg_num); 680 681 if (ttransport->tcp_opts.sock_priority > SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY) { 682 SPDK_ERRLOG("Unsupported socket_priority=%d, the current range is: 0 to %d\n" 683 "you can use man 7 socket to view the range of priority under SO_PRIORITY item\n", 684 ttransport->tcp_opts.sock_priority, SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY); 685 free(ttransport); 686 return NULL; 687 } 688 689 if (ttransport->tcp_opts.control_msg_num == 0 && 690 opts->in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 691 SPDK_WARNLOG("TCP param control_msg_num can't be 0 if ICD is less than %u bytes. Using default value %u\n", 692 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM); 693 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 694 } 695 696 /* I/O unit size cannot be larger than max I/O size */ 697 if (opts->io_unit_size > opts->max_io_size) { 698 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", 699 opts->io_unit_size, opts->max_io_size); 700 opts->io_unit_size = opts->max_io_size; 701 } 702 703 /* In capsule data size cannot be larger than max I/O size */ 704 if (opts->in_capsule_data_size > opts->max_io_size) { 705 SPDK_WARNLOG("TCP param ICD size %u can't be larger than max_io_size %u. Using max_io_size as ICD size\n", 706 opts->io_unit_size, opts->max_io_size); 707 opts->in_capsule_data_size = opts->max_io_size; 708 } 709 710 /* max IO queue depth cannot be smaller than 2 or larger than 65535. 711 * We will not check SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH, because max_queue_depth is 16bits and always not larger than 64k. */ 712 if (opts->max_queue_depth < SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH) { 713 SPDK_WARNLOG("TCP param max_queue_depth %u can't be smaller than %u or larger than %u. Using default value %u\n", 714 opts->max_queue_depth, SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH, 715 SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH, SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH); 716 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH; 717 } 718 719 /* max admin queue depth cannot be smaller than 2 or larger than 4096 */ 720 if (opts->max_aq_depth < SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH || 721 opts->max_aq_depth > SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH) { 722 SPDK_WARNLOG("TCP param max_aq_depth %u can't be smaller than %u or larger than %u. Using default value %u\n", 723 opts->max_aq_depth, SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH, 724 SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH, SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH); 725 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH; 726 } 727 728 sge_count = opts->max_io_size / opts->io_unit_size; 729 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 730 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 731 free(ttransport); 732 return NULL; 733 } 734 735 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 736 if (opts->buf_cache_size < UINT32_MAX) { 737 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 738 if (min_shared_buffers > opts->num_shared_buffers) { 739 SPDK_ERRLOG("There are not enough buffers to satisfy " 740 "per-poll group caches for each thread. (%" PRIu32 ") " 741 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 742 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 743 free(ttransport); 744 return NULL; 745 } 746 } 747 748 ttransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_tcp_accept, &ttransport->transport, 749 opts->acceptor_poll_rate); 750 if (!ttransport->accept_poller) { 751 free(ttransport); 752 return NULL; 753 } 754 755 return &ttransport->transport; 756 } 757 758 static int 759 nvmf_tcp_trsvcid_to_int(const char *trsvcid) 760 { 761 unsigned long long ull; 762 char *end = NULL; 763 764 ull = strtoull(trsvcid, &end, 10); 765 if (end == NULL || end == trsvcid || *end != '\0') { 766 return -1; 767 } 768 769 /* Valid TCP/IP port numbers are in [1, 65535] */ 770 if (ull == 0 || ull > 65535) { 771 return -1; 772 } 773 774 return (int)ull; 775 } 776 777 /** 778 * Canonicalize a listen address trid. 779 */ 780 static int 781 nvmf_tcp_canon_listen_trid(struct spdk_nvme_transport_id *canon_trid, 782 const struct spdk_nvme_transport_id *trid) 783 { 784 int trsvcid_int; 785 786 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 787 if (trsvcid_int < 0) { 788 return -EINVAL; 789 } 790 791 memset(canon_trid, 0, sizeof(*canon_trid)); 792 spdk_nvme_trid_populate_transport(canon_trid, SPDK_NVME_TRANSPORT_TCP); 793 canon_trid->adrfam = trid->adrfam; 794 snprintf(canon_trid->traddr, sizeof(canon_trid->traddr), "%s", trid->traddr); 795 snprintf(canon_trid->trsvcid, sizeof(canon_trid->trsvcid), "%d", trsvcid_int); 796 797 return 0; 798 } 799 800 /** 801 * Find an existing listening port. 802 */ 803 static struct spdk_nvmf_tcp_port * 804 nvmf_tcp_find_port(struct spdk_nvmf_tcp_transport *ttransport, 805 const struct spdk_nvme_transport_id *trid) 806 { 807 struct spdk_nvme_transport_id canon_trid; 808 struct spdk_nvmf_tcp_port *port; 809 810 if (nvmf_tcp_canon_listen_trid(&canon_trid, trid) != 0) { 811 return NULL; 812 } 813 814 TAILQ_FOREACH(port, &ttransport->ports, link) { 815 if (spdk_nvme_transport_id_compare(&canon_trid, port->trid) == 0) { 816 return port; 817 } 818 } 819 820 return NULL; 821 } 822 823 static int 824 tcp_sock_get_key(uint8_t *out, int out_len, const char **cipher, const char *psk_identity, 825 void *get_key_ctx) 826 { 827 struct tcp_psk_entry *entry; 828 struct spdk_nvmf_tcp_transport *ttransport = get_key_ctx; 829 size_t psk_len; 830 int rc; 831 832 TAILQ_FOREACH(entry, &ttransport->psks, link) { 833 if (strcmp(psk_identity, entry->psk_identity) != 0) { 834 continue; 835 } 836 837 psk_len = entry->psk_size; 838 if ((size_t)out_len < psk_len) { 839 SPDK_ERRLOG("Out buffer of size: %" PRIu32 " cannot fit PSK of len: %lu\n", 840 out_len, psk_len); 841 return -ENOBUFS; 842 } 843 844 /* Convert PSK to the TLS PSK format. */ 845 rc = nvme_tcp_derive_tls_psk(entry->psk, psk_len, psk_identity, out, out_len, 846 entry->tls_cipher_suite); 847 if (rc < 0) { 848 SPDK_ERRLOG("Could not generate TLS PSK\n"); 849 } 850 851 switch (entry->tls_cipher_suite) { 852 case NVME_TCP_CIPHER_AES_128_GCM_SHA256: 853 *cipher = "TLS_AES_128_GCM_SHA256"; 854 break; 855 case NVME_TCP_CIPHER_AES_256_GCM_SHA384: 856 *cipher = "TLS_AES_256_GCM_SHA384"; 857 break; 858 default: 859 *cipher = NULL; 860 return -ENOTSUP; 861 } 862 863 return rc; 864 } 865 866 SPDK_ERRLOG("Could not find PSK for identity: %s\n", psk_identity); 867 868 return -EINVAL; 869 } 870 871 static int 872 nvmf_tcp_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 873 struct spdk_nvmf_listen_opts *listen_opts) 874 { 875 struct spdk_nvmf_tcp_transport *ttransport; 876 struct spdk_nvmf_tcp_port *port; 877 int trsvcid_int; 878 uint8_t adrfam; 879 const char *sock_impl_name; 880 struct spdk_sock_impl_opts impl_opts; 881 size_t impl_opts_size = sizeof(impl_opts); 882 struct spdk_sock_opts opts; 883 884 if (!strlen(trid->trsvcid)) { 885 SPDK_ERRLOG("Service id is required\n"); 886 return -EINVAL; 887 } 888 889 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 890 891 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 892 if (trsvcid_int < 0) { 893 SPDK_ERRLOG("Invalid trsvcid '%s'\n", trid->trsvcid); 894 return -EINVAL; 895 } 896 897 port = calloc(1, sizeof(*port)); 898 if (!port) { 899 SPDK_ERRLOG("Port allocation failed\n"); 900 return -ENOMEM; 901 } 902 903 port->trid = trid; 904 905 sock_impl_name = NULL; 906 907 opts.opts_size = sizeof(opts); 908 spdk_sock_get_default_opts(&opts); 909 opts.priority = ttransport->tcp_opts.sock_priority; 910 if (listen_opts->secure_channel) { 911 if (!g_tls_log) { 912 SPDK_NOTICELOG("TLS support is considered experimental\n"); 913 g_tls_log = true; 914 } 915 sock_impl_name = "ssl"; 916 spdk_sock_impl_get_opts(sock_impl_name, &impl_opts, &impl_opts_size); 917 impl_opts.tls_version = SPDK_TLS_VERSION_1_3; 918 impl_opts.get_key = tcp_sock_get_key; 919 impl_opts.get_key_ctx = ttransport; 920 impl_opts.tls_cipher_suites = "TLS_AES_256_GCM_SHA384:TLS_AES_128_GCM_SHA256"; 921 opts.impl_opts = &impl_opts; 922 opts.impl_opts_size = sizeof(impl_opts); 923 } 924 925 port->listen_sock = spdk_sock_listen_ext(trid->traddr, trsvcid_int, 926 sock_impl_name, &opts); 927 if (port->listen_sock == NULL) { 928 SPDK_ERRLOG("spdk_sock_listen(%s, %d) failed: %s (%d)\n", 929 trid->traddr, trsvcid_int, 930 spdk_strerror(errno), errno); 931 free(port); 932 return -errno; 933 } 934 935 if (spdk_sock_is_ipv4(port->listen_sock)) { 936 adrfam = SPDK_NVMF_ADRFAM_IPV4; 937 } else if (spdk_sock_is_ipv6(port->listen_sock)) { 938 adrfam = SPDK_NVMF_ADRFAM_IPV6; 939 } else { 940 SPDK_ERRLOG("Unhandled socket type\n"); 941 adrfam = 0; 942 } 943 944 if (adrfam != trid->adrfam) { 945 SPDK_ERRLOG("Socket address family mismatch\n"); 946 spdk_sock_close(&port->listen_sock); 947 free(port); 948 return -EINVAL; 949 } 950 951 SPDK_NOTICELOG("*** NVMe/TCP Target Listening on %s port %s ***\n", 952 trid->traddr, trid->trsvcid); 953 954 TAILQ_INSERT_TAIL(&ttransport->ports, port, link); 955 return 0; 956 } 957 958 static void 959 nvmf_tcp_stop_listen(struct spdk_nvmf_transport *transport, 960 const struct spdk_nvme_transport_id *trid) 961 { 962 struct spdk_nvmf_tcp_transport *ttransport; 963 struct spdk_nvmf_tcp_port *port; 964 965 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 966 967 SPDK_DEBUGLOG(nvmf_tcp, "Removing listen address %s port %s\n", 968 trid->traddr, trid->trsvcid); 969 970 port = nvmf_tcp_find_port(ttransport, trid); 971 if (port) { 972 TAILQ_REMOVE(&ttransport->ports, port, link); 973 spdk_sock_close(&port->listen_sock); 974 free(port); 975 } 976 } 977 978 static void nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 979 enum nvme_tcp_pdu_recv_state state); 980 981 static void 982 nvmf_tcp_qpair_set_state(struct spdk_nvmf_tcp_qpair *tqpair, enum nvme_tcp_qpair_state state) 983 { 984 tqpair->state = state; 985 spdk_trace_record(TRACE_TCP_QP_STATE_CHANGE, tqpair->qpair.qid, 0, (uintptr_t)tqpair, 986 tqpair->state); 987 } 988 989 static void 990 nvmf_tcp_qpair_disconnect(struct spdk_nvmf_tcp_qpair *tqpair) 991 { 992 SPDK_DEBUGLOG(nvmf_tcp, "Disconnecting qpair %p\n", tqpair); 993 994 spdk_trace_record(TRACE_TCP_QP_DISCONNECT, 0, 0, (uintptr_t)tqpair); 995 996 if (tqpair->state <= NVME_TCP_QPAIR_STATE_RUNNING) { 997 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITING); 998 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 999 spdk_poller_unregister(&tqpair->timeout_poller); 1000 1001 /* This will end up calling nvmf_tcp_close_qpair */ 1002 spdk_nvmf_qpair_disconnect(&tqpair->qpair, NULL, NULL); 1003 } 1004 } 1005 1006 static void 1007 _mgmt_pdu_write_done(void *_tqpair, int err) 1008 { 1009 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 1010 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1011 1012 if (spdk_unlikely(err != 0)) { 1013 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1014 return; 1015 } 1016 1017 assert(pdu->cb_fn != NULL); 1018 pdu->cb_fn(pdu->cb_arg); 1019 } 1020 1021 static void 1022 _req_pdu_write_done(void *req, int err) 1023 { 1024 struct spdk_nvmf_tcp_req *tcp_req = req; 1025 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1026 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1027 1028 assert(tcp_req->pdu_in_use); 1029 tcp_req->pdu_in_use = false; 1030 1031 /* If the request is in a completed state, we're waiting for write completion to free it */ 1032 if (spdk_unlikely(tcp_req->state == TCP_REQUEST_STATE_COMPLETED)) { 1033 nvmf_tcp_request_free(tcp_req); 1034 return; 1035 } 1036 1037 if (spdk_unlikely(err != 0)) { 1038 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1039 return; 1040 } 1041 1042 assert(pdu->cb_fn != NULL); 1043 pdu->cb_fn(pdu->cb_arg); 1044 } 1045 1046 static void 1047 _pdu_write_done(struct nvme_tcp_pdu *pdu, int err) 1048 { 1049 pdu->sock_req.cb_fn(pdu->sock_req.cb_arg, err); 1050 } 1051 1052 static void 1053 _tcp_write_pdu(struct nvme_tcp_pdu *pdu) 1054 { 1055 int rc; 1056 uint32_t mapped_length; 1057 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1058 1059 pdu->sock_req.iovcnt = nvme_tcp_build_iovs(pdu->iov, SPDK_COUNTOF(pdu->iov), pdu, 1060 tqpair->host_hdgst_enable, tqpair->host_ddgst_enable, &mapped_length); 1061 spdk_sock_writev_async(tqpair->sock, &pdu->sock_req); 1062 1063 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP || 1064 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ) { 1065 /* Try to force the send immediately. */ 1066 rc = spdk_sock_flush(tqpair->sock); 1067 if (rc > 0 && (uint32_t)rc == mapped_length) { 1068 _pdu_write_done(pdu, 0); 1069 } else { 1070 SPDK_ERRLOG("Could not write %s to socket: rc=%d, errno=%d\n", 1071 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP ? 1072 "IC_RESP" : "TERM_REQ", rc, errno); 1073 _pdu_write_done(pdu, rc >= 0 ? -EAGAIN : -errno); 1074 } 1075 } 1076 } 1077 1078 static void 1079 data_crc32_accel_done(void *cb_arg, int status) 1080 { 1081 struct nvme_tcp_pdu *pdu = cb_arg; 1082 1083 if (spdk_unlikely(status)) { 1084 SPDK_ERRLOG("Failed to compute the data digest for pdu =%p\n", pdu); 1085 _pdu_write_done(pdu, status); 1086 return; 1087 } 1088 1089 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 1090 MAKE_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32); 1091 1092 _tcp_write_pdu(pdu); 1093 } 1094 1095 static void 1096 pdu_data_crc32_compute(struct nvme_tcp_pdu *pdu) 1097 { 1098 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1099 int rc = 0; 1100 1101 /* Data Digest */ 1102 if (pdu->data_len > 0 && g_nvme_tcp_ddgst[pdu->hdr.common.pdu_type] && tqpair->host_ddgst_enable) { 1103 /* Only support this limitated case for the first step */ 1104 if (spdk_likely(!pdu->dif_ctx && (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0) 1105 && tqpair->group)) { 1106 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 1107 pdu->data_iovcnt, 0, data_crc32_accel_done, pdu); 1108 if (spdk_likely(rc == 0)) { 1109 return; 1110 } 1111 } else { 1112 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 1113 } 1114 data_crc32_accel_done(pdu, rc); 1115 } else { 1116 _tcp_write_pdu(pdu); 1117 } 1118 } 1119 1120 static void 1121 nvmf_tcp_qpair_write_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1122 struct nvme_tcp_pdu *pdu, 1123 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1124 void *cb_arg) 1125 { 1126 int hlen; 1127 uint32_t crc32c; 1128 1129 assert(tqpair->pdu_in_progress != pdu); 1130 1131 hlen = pdu->hdr.common.hlen; 1132 pdu->cb_fn = cb_fn; 1133 pdu->cb_arg = cb_arg; 1134 1135 pdu->iov[0].iov_base = &pdu->hdr.raw; 1136 pdu->iov[0].iov_len = hlen; 1137 1138 /* Header Digest */ 1139 if (g_nvme_tcp_hdgst[pdu->hdr.common.pdu_type] && tqpair->host_hdgst_enable) { 1140 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 1141 MAKE_DIGEST_WORD((uint8_t *)pdu->hdr.raw + hlen, crc32c); 1142 } 1143 1144 /* Data Digest */ 1145 pdu_data_crc32_compute(pdu); 1146 } 1147 1148 static void 1149 nvmf_tcp_qpair_write_mgmt_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1150 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1151 void *cb_arg) 1152 { 1153 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1154 1155 pdu->sock_req.cb_fn = _mgmt_pdu_write_done; 1156 pdu->sock_req.cb_arg = tqpair; 1157 1158 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1159 } 1160 1161 static void 1162 nvmf_tcp_qpair_write_req_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1163 struct spdk_nvmf_tcp_req *tcp_req, 1164 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1165 void *cb_arg) 1166 { 1167 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1168 1169 pdu->sock_req.cb_fn = _req_pdu_write_done; 1170 pdu->sock_req.cb_arg = tcp_req; 1171 1172 assert(!tcp_req->pdu_in_use); 1173 tcp_req->pdu_in_use = true; 1174 1175 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1176 } 1177 1178 static int 1179 nvmf_tcp_qpair_init_mem_resource(struct spdk_nvmf_tcp_qpair *tqpair) 1180 { 1181 uint32_t i; 1182 struct spdk_nvmf_transport_opts *opts; 1183 uint32_t in_capsule_data_size; 1184 1185 opts = &tqpair->qpair.transport->opts; 1186 1187 in_capsule_data_size = opts->in_capsule_data_size; 1188 if (opts->dif_insert_or_strip) { 1189 in_capsule_data_size = SPDK_BDEV_BUF_SIZE_WITH_MD(in_capsule_data_size); 1190 } 1191 1192 tqpair->resource_count = opts->max_queue_depth; 1193 1194 tqpair->reqs = calloc(tqpair->resource_count, sizeof(*tqpair->reqs)); 1195 if (!tqpair->reqs) { 1196 SPDK_ERRLOG("Unable to allocate reqs on tqpair=%p\n", tqpair); 1197 return -1; 1198 } 1199 1200 if (in_capsule_data_size) { 1201 tqpair->bufs = spdk_zmalloc(tqpair->resource_count * in_capsule_data_size, 0x1000, 1202 NULL, SPDK_ENV_LCORE_ID_ANY, 1203 SPDK_MALLOC_DMA); 1204 if (!tqpair->bufs) { 1205 SPDK_ERRLOG("Unable to allocate bufs on tqpair=%p.\n", tqpair); 1206 return -1; 1207 } 1208 } 1209 /* prepare memory space for receiving pdus and tcp_req */ 1210 /* Add additional 1 member, which will be used for mgmt_pdu owned by the tqpair */ 1211 tqpair->pdus = spdk_dma_zmalloc((2 * tqpair->resource_count + 1) * sizeof(*tqpair->pdus), 0x1000, 1212 NULL); 1213 if (!tqpair->pdus) { 1214 SPDK_ERRLOG("Unable to allocate pdu pool on tqpair =%p.\n", tqpair); 1215 return -1; 1216 } 1217 1218 for (i = 0; i < tqpair->resource_count; i++) { 1219 struct spdk_nvmf_tcp_req *tcp_req = &tqpair->reqs[i]; 1220 1221 tcp_req->ttag = i + 1; 1222 tcp_req->req.qpair = &tqpair->qpair; 1223 1224 tcp_req->pdu = &tqpair->pdus[i]; 1225 tcp_req->pdu->qpair = tqpair; 1226 1227 /* Set up memory to receive commands */ 1228 if (tqpair->bufs) { 1229 tcp_req->buf = (void *)((uintptr_t)tqpair->bufs + (i * in_capsule_data_size)); 1230 } 1231 1232 /* Set the cmdn and rsp */ 1233 tcp_req->req.rsp = (union nvmf_c2h_msg *)&tcp_req->rsp; 1234 tcp_req->req.cmd = (union nvmf_h2c_msg *)&tcp_req->cmd; 1235 1236 tcp_req->req.stripped_data = NULL; 1237 1238 /* Initialize request state to FREE */ 1239 tcp_req->state = TCP_REQUEST_STATE_FREE; 1240 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 1241 tqpair->state_cntr[TCP_REQUEST_STATE_FREE]++; 1242 } 1243 1244 for (; i < 2 * tqpair->resource_count; i++) { 1245 struct nvme_tcp_pdu *pdu = &tqpair->pdus[i]; 1246 1247 pdu->qpair = tqpair; 1248 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 1249 } 1250 1251 tqpair->mgmt_pdu = &tqpair->pdus[i]; 1252 tqpair->mgmt_pdu->qpair = tqpair; 1253 tqpair->pdu_in_progress = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 1254 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 1255 tqpair->tcp_pdu_working_count = 1; 1256 1257 tqpair->recv_buf_size = (in_capsule_data_size + sizeof(struct spdk_nvme_tcp_cmd) + 2 * 1258 SPDK_NVME_TCP_DIGEST_LEN) * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1259 1260 return 0; 1261 } 1262 1263 static int 1264 nvmf_tcp_qpair_init(struct spdk_nvmf_qpair *qpair) 1265 { 1266 struct spdk_nvmf_tcp_qpair *tqpair; 1267 1268 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1269 1270 SPDK_DEBUGLOG(nvmf_tcp, "New TCP Connection: %p\n", qpair); 1271 1272 spdk_trace_record(TRACE_TCP_QP_CREATE, 0, 0, (uintptr_t)tqpair); 1273 1274 /* Initialise request state queues of the qpair */ 1275 TAILQ_INIT(&tqpair->tcp_req_free_queue); 1276 TAILQ_INIT(&tqpair->tcp_req_working_queue); 1277 SLIST_INIT(&tqpair->tcp_pdu_free_queue); 1278 1279 tqpair->host_hdgst_enable = true; 1280 tqpair->host_ddgst_enable = true; 1281 1282 return 0; 1283 } 1284 1285 static int 1286 nvmf_tcp_qpair_sock_init(struct spdk_nvmf_tcp_qpair *tqpair) 1287 { 1288 int rc; 1289 1290 spdk_trace_record(TRACE_TCP_QP_SOCK_INIT, 0, 0, (uintptr_t)tqpair); 1291 1292 /* set low water mark */ 1293 rc = spdk_sock_set_recvlowat(tqpair->sock, 1); 1294 if (rc != 0) { 1295 SPDK_ERRLOG("spdk_sock_set_recvlowat() failed\n"); 1296 return rc; 1297 } 1298 1299 return 0; 1300 } 1301 1302 static void 1303 nvmf_tcp_handle_connect(struct spdk_nvmf_transport *transport, 1304 struct spdk_nvmf_tcp_port *port, 1305 struct spdk_sock *sock) 1306 { 1307 struct spdk_nvmf_tcp_qpair *tqpair; 1308 int rc; 1309 1310 SPDK_DEBUGLOG(nvmf_tcp, "New connection accepted on %s port %s\n", 1311 port->trid->traddr, port->trid->trsvcid); 1312 1313 tqpair = calloc(1, sizeof(struct spdk_nvmf_tcp_qpair)); 1314 if (tqpair == NULL) { 1315 SPDK_ERRLOG("Could not allocate new connection.\n"); 1316 spdk_sock_close(&sock); 1317 return; 1318 } 1319 1320 tqpair->sock = sock; 1321 tqpair->state_cntr[TCP_REQUEST_STATE_FREE] = 0; 1322 tqpair->port = port; 1323 tqpair->qpair.transport = transport; 1324 1325 rc = spdk_sock_getaddr(tqpair->sock, tqpair->target_addr, 1326 sizeof(tqpair->target_addr), &tqpair->target_port, 1327 tqpair->initiator_addr, sizeof(tqpair->initiator_addr), 1328 &tqpair->initiator_port); 1329 if (rc < 0) { 1330 SPDK_ERRLOG("spdk_sock_getaddr() failed of tqpair=%p\n", tqpair); 1331 nvmf_tcp_qpair_destroy(tqpair); 1332 return; 1333 } 1334 1335 spdk_nvmf_tgt_new_qpair(transport->tgt, &tqpair->qpair); 1336 } 1337 1338 static uint32_t 1339 nvmf_tcp_port_accept(struct spdk_nvmf_transport *transport, struct spdk_nvmf_tcp_port *port) 1340 { 1341 struct spdk_sock *sock; 1342 uint32_t count = 0; 1343 int i; 1344 1345 for (i = 0; i < NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME; i++) { 1346 sock = spdk_sock_accept(port->listen_sock); 1347 if (sock == NULL) { 1348 break; 1349 } 1350 count++; 1351 nvmf_tcp_handle_connect(transport, port, sock); 1352 } 1353 1354 return count; 1355 } 1356 1357 static int 1358 nvmf_tcp_accept(void *ctx) 1359 { 1360 struct spdk_nvmf_transport *transport = ctx; 1361 struct spdk_nvmf_tcp_transport *ttransport; 1362 struct spdk_nvmf_tcp_port *port; 1363 uint32_t count = 0; 1364 1365 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1366 1367 TAILQ_FOREACH(port, &ttransport->ports, link) { 1368 count += nvmf_tcp_port_accept(transport, port); 1369 } 1370 1371 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 1372 } 1373 1374 static void 1375 nvmf_tcp_discover(struct spdk_nvmf_transport *transport, 1376 struct spdk_nvme_transport_id *trid, 1377 struct spdk_nvmf_discovery_log_page_entry *entry) 1378 { 1379 struct spdk_nvmf_tcp_port *port; 1380 struct spdk_nvmf_tcp_transport *ttransport; 1381 1382 entry->trtype = SPDK_NVMF_TRTYPE_TCP; 1383 entry->adrfam = trid->adrfam; 1384 1385 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 1386 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 1387 1388 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1389 port = nvmf_tcp_find_port(ttransport, trid); 1390 1391 assert(port != NULL); 1392 1393 if (strcmp(spdk_sock_get_impl_name(port->listen_sock), "ssl") == 0) { 1394 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_REQUIRED; 1395 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_TLS_1_3; 1396 } else { 1397 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 1398 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_NONE; 1399 } 1400 } 1401 1402 static struct spdk_nvmf_tcp_control_msg_list * 1403 nvmf_tcp_control_msg_list_create(uint16_t num_messages) 1404 { 1405 struct spdk_nvmf_tcp_control_msg_list *list; 1406 struct spdk_nvmf_tcp_control_msg *msg; 1407 uint16_t i; 1408 1409 list = calloc(1, sizeof(*list)); 1410 if (!list) { 1411 SPDK_ERRLOG("Failed to allocate memory for list structure\n"); 1412 return NULL; 1413 } 1414 1415 list->msg_buf = spdk_zmalloc(num_messages * SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, 1416 NVMF_DATA_BUFFER_ALIGNMENT, NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 1417 if (!list->msg_buf) { 1418 SPDK_ERRLOG("Failed to allocate memory for control message buffers\n"); 1419 free(list); 1420 return NULL; 1421 } 1422 1423 STAILQ_INIT(&list->free_msgs); 1424 1425 for (i = 0; i < num_messages; i++) { 1426 msg = (struct spdk_nvmf_tcp_control_msg *)((char *)list->msg_buf + i * 1427 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1428 STAILQ_INSERT_TAIL(&list->free_msgs, msg, link); 1429 } 1430 1431 return list; 1432 } 1433 1434 static void 1435 nvmf_tcp_control_msg_list_free(struct spdk_nvmf_tcp_control_msg_list *list) 1436 { 1437 if (!list) { 1438 return; 1439 } 1440 1441 spdk_free(list->msg_buf); 1442 free(list); 1443 } 1444 1445 static struct spdk_nvmf_transport_poll_group * 1446 nvmf_tcp_poll_group_create(struct spdk_nvmf_transport *transport, 1447 struct spdk_nvmf_poll_group *group) 1448 { 1449 struct spdk_nvmf_tcp_transport *ttransport; 1450 struct spdk_nvmf_tcp_poll_group *tgroup; 1451 1452 tgroup = calloc(1, sizeof(*tgroup)); 1453 if (!tgroup) { 1454 return NULL; 1455 } 1456 1457 tgroup->sock_group = spdk_sock_group_create(&tgroup->group); 1458 if (!tgroup->sock_group) { 1459 goto cleanup; 1460 } 1461 1462 TAILQ_INIT(&tgroup->qpairs); 1463 TAILQ_INIT(&tgroup->await_req); 1464 1465 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1466 1467 if (transport->opts.in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 1468 SPDK_DEBUGLOG(nvmf_tcp, "ICD %u is less than min required for admin/fabric commands (%u). " 1469 "Creating control messages list\n", transport->opts.in_capsule_data_size, 1470 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1471 tgroup->control_msg_list = nvmf_tcp_control_msg_list_create(ttransport->tcp_opts.control_msg_num); 1472 if (!tgroup->control_msg_list) { 1473 goto cleanup; 1474 } 1475 } 1476 1477 tgroup->accel_channel = spdk_accel_get_io_channel(); 1478 if (spdk_unlikely(!tgroup->accel_channel)) { 1479 SPDK_ERRLOG("Cannot create accel_channel for tgroup=%p\n", tgroup); 1480 goto cleanup; 1481 } 1482 1483 TAILQ_INSERT_TAIL(&ttransport->poll_groups, tgroup, link); 1484 if (ttransport->next_pg == NULL) { 1485 ttransport->next_pg = tgroup; 1486 } 1487 1488 return &tgroup->group; 1489 1490 cleanup: 1491 nvmf_tcp_poll_group_destroy(&tgroup->group); 1492 return NULL; 1493 } 1494 1495 static struct spdk_nvmf_transport_poll_group * 1496 nvmf_tcp_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 1497 { 1498 struct spdk_nvmf_tcp_transport *ttransport; 1499 struct spdk_nvmf_tcp_poll_group **pg; 1500 struct spdk_nvmf_tcp_qpair *tqpair; 1501 struct spdk_sock_group *group = NULL, *hint = NULL; 1502 int rc; 1503 1504 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 1505 1506 if (TAILQ_EMPTY(&ttransport->poll_groups)) { 1507 return NULL; 1508 } 1509 1510 pg = &ttransport->next_pg; 1511 assert(*pg != NULL); 1512 hint = (*pg)->sock_group; 1513 1514 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1515 rc = spdk_sock_get_optimal_sock_group(tqpair->sock, &group, hint); 1516 if (rc != 0) { 1517 return NULL; 1518 } else if (group != NULL) { 1519 /* Optimal poll group was found */ 1520 return spdk_sock_group_get_ctx(group); 1521 } 1522 1523 /* The hint was used for optimal poll group, advance next_pg. */ 1524 *pg = TAILQ_NEXT(*pg, link); 1525 if (*pg == NULL) { 1526 *pg = TAILQ_FIRST(&ttransport->poll_groups); 1527 } 1528 1529 return spdk_sock_group_get_ctx(hint); 1530 } 1531 1532 static void 1533 nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 1534 { 1535 struct spdk_nvmf_tcp_poll_group *tgroup, *next_tgroup; 1536 struct spdk_nvmf_tcp_transport *ttransport; 1537 1538 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 1539 spdk_sock_group_close(&tgroup->sock_group); 1540 if (tgroup->control_msg_list) { 1541 nvmf_tcp_control_msg_list_free(tgroup->control_msg_list); 1542 } 1543 1544 if (tgroup->accel_channel) { 1545 spdk_put_io_channel(tgroup->accel_channel); 1546 } 1547 1548 if (tgroup->group.transport == NULL) { 1549 /* Transport can be NULL when nvmf_tcp_poll_group_create() 1550 * calls this function directly in a failure path. */ 1551 free(tgroup); 1552 return; 1553 } 1554 1555 ttransport = SPDK_CONTAINEROF(tgroup->group.transport, struct spdk_nvmf_tcp_transport, transport); 1556 1557 next_tgroup = TAILQ_NEXT(tgroup, link); 1558 TAILQ_REMOVE(&ttransport->poll_groups, tgroup, link); 1559 if (next_tgroup == NULL) { 1560 next_tgroup = TAILQ_FIRST(&ttransport->poll_groups); 1561 } 1562 if (ttransport->next_pg == tgroup) { 1563 ttransport->next_pg = next_tgroup; 1564 } 1565 1566 free(tgroup); 1567 } 1568 1569 static void 1570 nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 1571 enum nvme_tcp_pdu_recv_state state) 1572 { 1573 if (tqpair->recv_state == state) { 1574 SPDK_ERRLOG("The recv state of tqpair=%p is same with the state(%d) to be set\n", 1575 tqpair, state); 1576 return; 1577 } 1578 1579 if (spdk_unlikely(state == NVME_TCP_PDU_RECV_STATE_QUIESCING)) { 1580 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH && tqpair->pdu_in_progress) { 1581 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, tqpair->pdu_in_progress, slist); 1582 tqpair->tcp_pdu_working_count--; 1583 } 1584 } 1585 1586 if (spdk_unlikely(state == NVME_TCP_PDU_RECV_STATE_ERROR)) { 1587 assert(tqpair->tcp_pdu_working_count == 0); 1588 } 1589 1590 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1591 /* When leaving the await req state, move the qpair to the main list */ 1592 TAILQ_REMOVE(&tqpair->group->await_req, tqpair, link); 1593 TAILQ_INSERT_TAIL(&tqpair->group->qpairs, tqpair, link); 1594 } else if (state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1595 TAILQ_REMOVE(&tqpair->group->qpairs, tqpair, link); 1596 TAILQ_INSERT_TAIL(&tqpair->group->await_req, tqpair, link); 1597 } 1598 1599 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv state=%d\n", tqpair, state); 1600 tqpair->recv_state = state; 1601 1602 spdk_trace_record(TRACE_TCP_QP_RCV_STATE_CHANGE, tqpair->qpair.qid, 0, (uintptr_t)tqpair, 1603 tqpair->recv_state); 1604 } 1605 1606 static int 1607 nvmf_tcp_qpair_handle_timeout(void *ctx) 1608 { 1609 struct spdk_nvmf_tcp_qpair *tqpair = ctx; 1610 1611 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1612 1613 SPDK_ERRLOG("No pdu coming for tqpair=%p within %d seconds\n", tqpair, 1614 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT); 1615 1616 nvmf_tcp_qpair_disconnect(tqpair); 1617 return SPDK_POLLER_BUSY; 1618 } 1619 1620 static void 1621 nvmf_tcp_send_c2h_term_req_complete(void *cb_arg) 1622 { 1623 struct spdk_nvmf_tcp_qpair *tqpair = (struct spdk_nvmf_tcp_qpair *)cb_arg; 1624 1625 if (!tqpair->timeout_poller) { 1626 tqpair->timeout_poller = SPDK_POLLER_REGISTER(nvmf_tcp_qpair_handle_timeout, tqpair, 1627 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT * 1000000); 1628 } 1629 } 1630 1631 static void 1632 nvmf_tcp_send_c2h_term_req(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu, 1633 enum spdk_nvme_tcp_term_req_fes fes, uint32_t error_offset) 1634 { 1635 struct nvme_tcp_pdu *rsp_pdu; 1636 struct spdk_nvme_tcp_term_req_hdr *c2h_term_req; 1637 uint32_t c2h_term_req_hdr_len = sizeof(*c2h_term_req); 1638 uint32_t copy_len; 1639 1640 rsp_pdu = tqpair->mgmt_pdu; 1641 1642 c2h_term_req = &rsp_pdu->hdr.term_req; 1643 c2h_term_req->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ; 1644 c2h_term_req->common.hlen = c2h_term_req_hdr_len; 1645 c2h_term_req->fes = fes; 1646 1647 if ((fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1648 (fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1649 DSET32(&c2h_term_req->fei, error_offset); 1650 } 1651 1652 copy_len = spdk_min(pdu->hdr.common.hlen, SPDK_NVME_TCP_TERM_REQ_ERROR_DATA_MAX_SIZE); 1653 1654 /* Copy the error info into the buffer */ 1655 memcpy((uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, pdu->hdr.raw, copy_len); 1656 nvme_tcp_pdu_set_data(rsp_pdu, (uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, copy_len); 1657 1658 /* Contain the header of the wrong received pdu */ 1659 c2h_term_req->common.plen = c2h_term_req->common.hlen + copy_len; 1660 tqpair->wait_terminate = true; 1661 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1662 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_c2h_term_req_complete, tqpair); 1663 } 1664 1665 static void 1666 nvmf_tcp_capsule_cmd_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1667 struct spdk_nvmf_tcp_qpair *tqpair, 1668 struct nvme_tcp_pdu *pdu) 1669 { 1670 struct spdk_nvmf_tcp_req *tcp_req; 1671 1672 assert(pdu->psh_valid_bytes == pdu->psh_len); 1673 assert(pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD); 1674 1675 tcp_req = nvmf_tcp_req_get(tqpair); 1676 if (!tcp_req) { 1677 /* Directly return and make the allocation retry again. This can happen if we're 1678 * using asynchronous writes to send the response to the host or when releasing 1679 * zero-copy buffers after a response has been sent. In both cases, the host might 1680 * receive the response before we've finished processing the request and is free to 1681 * send another one. 1682 */ 1683 if (tqpair->state_cntr[TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST] > 0 || 1684 tqpair->state_cntr[TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE] > 0) { 1685 return; 1686 } 1687 1688 /* The host sent more commands than the maximum queue depth. */ 1689 SPDK_ERRLOG("Cannot allocate tcp_req on tqpair=%p\n", tqpair); 1690 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1691 return; 1692 } 1693 1694 pdu->req = tcp_req; 1695 assert(tcp_req->state == TCP_REQUEST_STATE_NEW); 1696 nvmf_tcp_req_process(ttransport, tcp_req); 1697 } 1698 1699 static void 1700 nvmf_tcp_capsule_cmd_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1701 struct spdk_nvmf_tcp_qpair *tqpair, 1702 struct nvme_tcp_pdu *pdu) 1703 { 1704 struct spdk_nvmf_tcp_req *tcp_req; 1705 struct spdk_nvme_tcp_cmd *capsule_cmd; 1706 uint32_t error_offset = 0; 1707 enum spdk_nvme_tcp_term_req_fes fes; 1708 struct spdk_nvme_cpl *rsp; 1709 1710 capsule_cmd = &pdu->hdr.capsule_cmd; 1711 tcp_req = pdu->req; 1712 assert(tcp_req != NULL); 1713 1714 /* Zero-copy requests don't support ICD */ 1715 assert(!spdk_nvmf_request_using_zcopy(&tcp_req->req)); 1716 1717 if (capsule_cmd->common.pdo > SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET) { 1718 SPDK_ERRLOG("Expected ICReq capsule_cmd pdu offset <= %d, got %c\n", 1719 SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET, capsule_cmd->common.pdo); 1720 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1721 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 1722 goto err; 1723 } 1724 1725 rsp = &tcp_req->req.rsp->nvme_cpl; 1726 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1727 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1728 } else { 1729 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1730 } 1731 1732 nvmf_tcp_req_process(ttransport, tcp_req); 1733 1734 return; 1735 err: 1736 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1737 } 1738 1739 static void 1740 nvmf_tcp_h2c_data_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1741 struct spdk_nvmf_tcp_qpair *tqpair, 1742 struct nvme_tcp_pdu *pdu) 1743 { 1744 struct spdk_nvmf_tcp_req *tcp_req; 1745 uint32_t error_offset = 0; 1746 enum spdk_nvme_tcp_term_req_fes fes = 0; 1747 struct spdk_nvme_tcp_h2c_data_hdr *h2c_data; 1748 1749 h2c_data = &pdu->hdr.h2c_data; 1750 1751 SPDK_DEBUGLOG(nvmf_tcp, "tqpair=%p, r2t_info: datao=%u, datal=%u, cccid=%u, ttag=%u\n", 1752 tqpair, h2c_data->datao, h2c_data->datal, h2c_data->cccid, h2c_data->ttag); 1753 1754 if (h2c_data->ttag > tqpair->resource_count) { 1755 SPDK_DEBUGLOG(nvmf_tcp, "ttag %u is larger than allowed %u.\n", h2c_data->ttag, 1756 tqpair->resource_count); 1757 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1758 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1759 goto err; 1760 } 1761 1762 tcp_req = &tqpair->reqs[h2c_data->ttag - 1]; 1763 1764 if (spdk_unlikely(tcp_req->state != TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER && 1765 tcp_req->state != TCP_REQUEST_STATE_AWAITING_R2T_ACK)) { 1766 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, has error state in %d\n", tcp_req, tqpair, 1767 tcp_req->state); 1768 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1769 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1770 goto err; 1771 } 1772 1773 if (spdk_unlikely(tcp_req->req.cmd->nvme_cmd.cid != h2c_data->cccid)) { 1774 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, expected %u but %u for cccid.\n", tcp_req, tqpair, 1775 tcp_req->req.cmd->nvme_cmd.cid, h2c_data->cccid); 1776 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1777 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, cccid); 1778 goto err; 1779 } 1780 1781 if (tcp_req->h2c_offset != h2c_data->datao) { 1782 SPDK_DEBUGLOG(nvmf_tcp, 1783 "tcp_req(%p), tqpair=%p, expected data offset %u, but data offset is %u\n", 1784 tcp_req, tqpair, tcp_req->h2c_offset, h2c_data->datao); 1785 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1786 goto err; 1787 } 1788 1789 if ((h2c_data->datao + h2c_data->datal) > tcp_req->req.length) { 1790 SPDK_DEBUGLOG(nvmf_tcp, 1791 "tcp_req(%p), tqpair=%p, (datao=%u + datal=%u) exceeds requested length=%u\n", 1792 tcp_req, tqpair, h2c_data->datao, h2c_data->datal, tcp_req->req.length); 1793 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1794 goto err; 1795 } 1796 1797 pdu->req = tcp_req; 1798 1799 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 1800 pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 1801 } 1802 1803 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 1804 h2c_data->datao, h2c_data->datal); 1805 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1806 return; 1807 1808 err: 1809 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1810 } 1811 1812 static void 1813 nvmf_tcp_send_capsule_resp_pdu(struct spdk_nvmf_tcp_req *tcp_req, 1814 struct spdk_nvmf_tcp_qpair *tqpair) 1815 { 1816 struct nvme_tcp_pdu *rsp_pdu; 1817 struct spdk_nvme_tcp_rsp *capsule_resp; 1818 1819 SPDK_DEBUGLOG(nvmf_tcp, "enter, tqpair=%p\n", tqpair); 1820 1821 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1822 assert(rsp_pdu != NULL); 1823 1824 capsule_resp = &rsp_pdu->hdr.capsule_resp; 1825 capsule_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_CAPSULE_RESP; 1826 capsule_resp->common.plen = capsule_resp->common.hlen = sizeof(*capsule_resp); 1827 capsule_resp->rccqe = tcp_req->req.rsp->nvme_cpl; 1828 if (tqpair->host_hdgst_enable) { 1829 capsule_resp->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 1830 capsule_resp->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 1831 } 1832 1833 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_request_free, tcp_req); 1834 } 1835 1836 static void 1837 nvmf_tcp_pdu_c2h_data_complete(void *cb_arg) 1838 { 1839 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 1840 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, 1841 struct spdk_nvmf_tcp_qpair, qpair); 1842 1843 assert(tqpair != NULL); 1844 1845 if (spdk_unlikely(tcp_req->pdu->rw_offset < tcp_req->req.length)) { 1846 SPDK_DEBUGLOG(nvmf_tcp, "sending another C2H part, offset %u length %u\n", tcp_req->pdu->rw_offset, 1847 tcp_req->req.length); 1848 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 1849 return; 1850 } 1851 1852 if (tcp_req->pdu->hdr.c2h_data.common.flags & SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS) { 1853 nvmf_tcp_request_free(tcp_req); 1854 } else { 1855 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 1856 } 1857 } 1858 1859 static void 1860 nvmf_tcp_r2t_complete(void *cb_arg) 1861 { 1862 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 1863 struct spdk_nvmf_tcp_transport *ttransport; 1864 1865 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 1866 struct spdk_nvmf_tcp_transport, transport); 1867 1868 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1869 1870 if (tcp_req->h2c_offset == tcp_req->req.length) { 1871 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1872 nvmf_tcp_req_process(ttransport, tcp_req); 1873 } 1874 } 1875 1876 static void 1877 nvmf_tcp_send_r2t_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1878 struct spdk_nvmf_tcp_req *tcp_req) 1879 { 1880 struct nvme_tcp_pdu *rsp_pdu; 1881 struct spdk_nvme_tcp_r2t_hdr *r2t; 1882 1883 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1884 assert(rsp_pdu != NULL); 1885 1886 r2t = &rsp_pdu->hdr.r2t; 1887 r2t->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_R2T; 1888 r2t->common.plen = r2t->common.hlen = sizeof(*r2t); 1889 1890 if (tqpair->host_hdgst_enable) { 1891 r2t->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 1892 r2t->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 1893 } 1894 1895 r2t->cccid = tcp_req->req.cmd->nvme_cmd.cid; 1896 r2t->ttag = tcp_req->ttag; 1897 r2t->r2to = tcp_req->h2c_offset; 1898 r2t->r2tl = tcp_req->req.length; 1899 1900 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 1901 1902 SPDK_DEBUGLOG(nvmf_tcp, 1903 "tcp_req(%p) on tqpair(%p), r2t_info: cccid=%u, ttag=%u, r2to=%u, r2tl=%u\n", 1904 tcp_req, tqpair, r2t->cccid, r2t->ttag, r2t->r2to, r2t->r2tl); 1905 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_r2t_complete, tcp_req); 1906 } 1907 1908 static void 1909 nvmf_tcp_h2c_data_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1910 struct spdk_nvmf_tcp_qpair *tqpair, 1911 struct nvme_tcp_pdu *pdu) 1912 { 1913 struct spdk_nvmf_tcp_req *tcp_req; 1914 struct spdk_nvme_cpl *rsp; 1915 1916 tcp_req = pdu->req; 1917 assert(tcp_req != NULL); 1918 1919 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 1920 1921 tcp_req->h2c_offset += pdu->data_len; 1922 1923 /* Wait for all of the data to arrive AND for the initial R2T PDU send to be 1924 * acknowledged before moving on. */ 1925 if (tcp_req->h2c_offset == tcp_req->req.length && 1926 tcp_req->state == TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER) { 1927 /* After receiving all the h2c data, we need to check whether there is 1928 * transient transport error */ 1929 rsp = &tcp_req->req.rsp->nvme_cpl; 1930 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1931 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1932 } else { 1933 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1934 } 1935 nvmf_tcp_req_process(ttransport, tcp_req); 1936 } 1937 } 1938 1939 static void 1940 nvmf_tcp_h2c_term_req_dump(struct spdk_nvme_tcp_term_req_hdr *h2c_term_req) 1941 { 1942 SPDK_ERRLOG("Error info of pdu(%p): %s\n", h2c_term_req, 1943 spdk_nvmf_tcp_term_req_fes_str[h2c_term_req->fes]); 1944 if ((h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1945 (h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1946 SPDK_DEBUGLOG(nvmf_tcp, "The offset from the start of the PDU header is %u\n", 1947 DGET32(h2c_term_req->fei)); 1948 } 1949 } 1950 1951 static void 1952 nvmf_tcp_h2c_term_req_hdr_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1953 struct nvme_tcp_pdu *pdu) 1954 { 1955 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 1956 uint32_t error_offset = 0; 1957 enum spdk_nvme_tcp_term_req_fes fes; 1958 1959 if (h2c_term_req->fes > SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER) { 1960 SPDK_ERRLOG("Fatal Error Status(FES) is unknown for h2c_term_req pdu=%p\n", pdu); 1961 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1962 error_offset = offsetof(struct spdk_nvme_tcp_term_req_hdr, fes); 1963 goto end; 1964 } 1965 1966 /* set the data buffer */ 1967 nvme_tcp_pdu_set_data(pdu, (uint8_t *)pdu->hdr.raw + h2c_term_req->common.hlen, 1968 h2c_term_req->common.plen - h2c_term_req->common.hlen); 1969 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1970 return; 1971 end: 1972 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1973 } 1974 1975 static void 1976 nvmf_tcp_h2c_term_req_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1977 struct nvme_tcp_pdu *pdu) 1978 { 1979 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 1980 1981 nvmf_tcp_h2c_term_req_dump(h2c_term_req); 1982 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1983 } 1984 1985 static void 1986 _nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 1987 { 1988 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 1989 struct spdk_nvmf_tcp_transport, transport); 1990 1991 switch (pdu->hdr.common.pdu_type) { 1992 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 1993 nvmf_tcp_capsule_cmd_payload_handle(ttransport, tqpair, pdu); 1994 break; 1995 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 1996 nvmf_tcp_h2c_data_payload_handle(ttransport, tqpair, pdu); 1997 break; 1998 1999 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2000 nvmf_tcp_h2c_term_req_payload_handle(tqpair, pdu); 2001 break; 2002 2003 default: 2004 /* The code should not go to here */ 2005 SPDK_ERRLOG("ERROR pdu type %d\n", pdu->hdr.common.pdu_type); 2006 break; 2007 } 2008 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 2009 tqpair->tcp_pdu_working_count--; 2010 } 2011 2012 static inline void 2013 nvmf_tcp_req_set_cpl(struct spdk_nvmf_tcp_req *treq, int sct, int sc) 2014 { 2015 treq->req.rsp->nvme_cpl.status.sct = sct; 2016 treq->req.rsp->nvme_cpl.status.sc = sc; 2017 treq->req.rsp->nvme_cpl.cid = treq->req.cmd->nvme_cmd.cid; 2018 } 2019 2020 static void 2021 data_crc32_calc_done(void *cb_arg, int status) 2022 { 2023 struct nvme_tcp_pdu *pdu = cb_arg; 2024 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 2025 2026 /* async crc32 calculation is failed and use direct calculation to check */ 2027 if (spdk_unlikely(status)) { 2028 SPDK_ERRLOG("Data digest on tqpair=(%p) with pdu=%p failed to be calculated asynchronously\n", 2029 tqpair, pdu); 2030 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 2031 } 2032 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 2033 if (!MATCH_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32)) { 2034 SPDK_ERRLOG("Data digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 2035 assert(pdu->req != NULL); 2036 nvmf_tcp_req_set_cpl(pdu->req, SPDK_NVME_SCT_GENERIC, 2037 SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR); 2038 } 2039 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2040 } 2041 2042 static void 2043 nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 2044 { 2045 int rc = 0; 2046 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2047 tqpair->pdu_in_progress = NULL; 2048 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2049 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2050 /* check data digest if need */ 2051 if (pdu->ddgst_enable) { 2052 if (tqpair->qpair.qid != 0 && !pdu->dif_ctx && tqpair->group && 2053 (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0)) { 2054 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 2055 pdu->data_iovcnt, 0, data_crc32_calc_done, pdu); 2056 if (spdk_likely(rc == 0)) { 2057 return; 2058 } 2059 } else { 2060 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 2061 } 2062 data_crc32_calc_done(pdu, rc); 2063 } else { 2064 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2065 } 2066 } 2067 2068 static void 2069 nvmf_tcp_send_icresp_complete(void *cb_arg) 2070 { 2071 struct spdk_nvmf_tcp_qpair *tqpair = cb_arg; 2072 2073 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_RUNNING); 2074 } 2075 2076 static void 2077 nvmf_tcp_icreq_handle(struct spdk_nvmf_tcp_transport *ttransport, 2078 struct spdk_nvmf_tcp_qpair *tqpair, 2079 struct nvme_tcp_pdu *pdu) 2080 { 2081 struct spdk_nvme_tcp_ic_req *ic_req = &pdu->hdr.ic_req; 2082 struct nvme_tcp_pdu *rsp_pdu; 2083 struct spdk_nvme_tcp_ic_resp *ic_resp; 2084 uint32_t error_offset = 0; 2085 enum spdk_nvme_tcp_term_req_fes fes; 2086 2087 /* Only PFV 0 is defined currently */ 2088 if (ic_req->pfv != 0) { 2089 SPDK_ERRLOG("Expected ICReq PFV %u, got %u\n", 0u, ic_req->pfv); 2090 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2091 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, pfv); 2092 goto end; 2093 } 2094 2095 /* This value is 0’s based value in units of dwords should not be larger than SPDK_NVME_TCP_HPDA_MAX */ 2096 if (ic_req->hpda > SPDK_NVME_TCP_HPDA_MAX) { 2097 SPDK_ERRLOG("ICReq HPDA out of range 0 to 31, got %u\n", ic_req->hpda); 2098 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2099 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, hpda); 2100 goto end; 2101 } 2102 2103 /* MAXR2T is 0's based */ 2104 SPDK_DEBUGLOG(nvmf_tcp, "maxr2t =%u\n", (ic_req->maxr2t + 1u)); 2105 2106 tqpair->host_hdgst_enable = ic_req->dgst.bits.hdgst_enable ? true : false; 2107 if (!tqpair->host_hdgst_enable) { 2108 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 2109 } 2110 2111 tqpair->host_ddgst_enable = ic_req->dgst.bits.ddgst_enable ? true : false; 2112 if (!tqpair->host_ddgst_enable) { 2113 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 2114 } 2115 2116 tqpair->recv_buf_size = spdk_max(tqpair->recv_buf_size, MIN_SOCK_PIPE_SIZE); 2117 /* Now that we know whether digests are enabled, properly size the receive buffer */ 2118 if (spdk_sock_set_recvbuf(tqpair->sock, tqpair->recv_buf_size) < 0) { 2119 SPDK_WARNLOG("Unable to allocate enough memory for receive buffer on tqpair=%p with size=%d\n", 2120 tqpair, 2121 tqpair->recv_buf_size); 2122 /* Not fatal. */ 2123 } 2124 2125 tqpair->cpda = spdk_min(ic_req->hpda, SPDK_NVME_TCP_CPDA_MAX); 2126 SPDK_DEBUGLOG(nvmf_tcp, "cpda of tqpair=(%p) is : %u\n", tqpair, tqpair->cpda); 2127 2128 rsp_pdu = tqpair->mgmt_pdu; 2129 2130 ic_resp = &rsp_pdu->hdr.ic_resp; 2131 ic_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_IC_RESP; 2132 ic_resp->common.hlen = ic_resp->common.plen = sizeof(*ic_resp); 2133 ic_resp->pfv = 0; 2134 ic_resp->cpda = tqpair->cpda; 2135 ic_resp->maxh2cdata = ttransport->transport.opts.max_io_size; 2136 ic_resp->dgst.bits.hdgst_enable = tqpair->host_hdgst_enable ? 1 : 0; 2137 ic_resp->dgst.bits.ddgst_enable = tqpair->host_ddgst_enable ? 1 : 0; 2138 2139 SPDK_DEBUGLOG(nvmf_tcp, "host_hdgst_enable: %u\n", tqpair->host_hdgst_enable); 2140 SPDK_DEBUGLOG(nvmf_tcp, "host_ddgst_enable: %u\n", tqpair->host_ddgst_enable); 2141 2142 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INITIALIZING); 2143 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_icresp_complete, tqpair); 2144 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2145 return; 2146 end: 2147 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2148 } 2149 2150 static void 2151 nvmf_tcp_pdu_psh_handle(struct spdk_nvmf_tcp_qpair *tqpair, 2152 struct spdk_nvmf_tcp_transport *ttransport) 2153 { 2154 struct nvme_tcp_pdu *pdu; 2155 int rc; 2156 uint32_t crc32c, error_offset = 0; 2157 enum spdk_nvme_tcp_term_req_fes fes; 2158 2159 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2160 pdu = tqpair->pdu_in_progress; 2161 2162 SPDK_DEBUGLOG(nvmf_tcp, "pdu type of tqpair(%p) is %d\n", tqpair, 2163 pdu->hdr.common.pdu_type); 2164 /* check header digest if needed */ 2165 if (pdu->has_hdgst) { 2166 SPDK_DEBUGLOG(nvmf_tcp, "Compare the header of pdu=%p on tqpair=%p\n", pdu, tqpair); 2167 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 2168 rc = MATCH_DIGEST_WORD((uint8_t *)pdu->hdr.raw + pdu->hdr.common.hlen, crc32c); 2169 if (rc == 0) { 2170 SPDK_ERRLOG("Header digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 2171 fes = SPDK_NVME_TCP_TERM_REQ_FES_HDGST_ERROR; 2172 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2173 return; 2174 2175 } 2176 } 2177 2178 switch (pdu->hdr.common.pdu_type) { 2179 case SPDK_NVME_TCP_PDU_TYPE_IC_REQ: 2180 nvmf_tcp_icreq_handle(ttransport, tqpair, pdu); 2181 break; 2182 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2183 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_REQ); 2184 break; 2185 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2186 nvmf_tcp_h2c_data_hdr_handle(ttransport, tqpair, pdu); 2187 break; 2188 2189 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2190 nvmf_tcp_h2c_term_req_hdr_handle(tqpair, pdu); 2191 break; 2192 2193 default: 2194 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", tqpair->pdu_in_progress->hdr.common.pdu_type); 2195 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2196 error_offset = 1; 2197 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2198 break; 2199 } 2200 } 2201 2202 static void 2203 nvmf_tcp_pdu_ch_handle(struct spdk_nvmf_tcp_qpair *tqpair) 2204 { 2205 struct nvme_tcp_pdu *pdu; 2206 uint32_t error_offset = 0; 2207 enum spdk_nvme_tcp_term_req_fes fes; 2208 uint8_t expected_hlen, pdo; 2209 bool plen_error = false, pdo_error = false; 2210 2211 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2212 pdu = tqpair->pdu_in_progress; 2213 assert(pdu); 2214 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_REQ) { 2215 if (tqpair->state != NVME_TCP_QPAIR_STATE_INVALID) { 2216 SPDK_ERRLOG("Already received ICreq PDU, and reject this pdu=%p\n", pdu); 2217 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2218 goto err; 2219 } 2220 expected_hlen = sizeof(struct spdk_nvme_tcp_ic_req); 2221 if (pdu->hdr.common.plen != expected_hlen) { 2222 plen_error = true; 2223 } 2224 } else { 2225 if (tqpair->state != NVME_TCP_QPAIR_STATE_RUNNING) { 2226 SPDK_ERRLOG("The TCP/IP connection is not negotiated\n"); 2227 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2228 goto err; 2229 } 2230 2231 switch (pdu->hdr.common.pdu_type) { 2232 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2233 expected_hlen = sizeof(struct spdk_nvme_tcp_cmd); 2234 pdo = pdu->hdr.common.pdo; 2235 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2236 pdo_error = true; 2237 break; 2238 } 2239 2240 if (pdu->hdr.common.plen < expected_hlen) { 2241 plen_error = true; 2242 } 2243 break; 2244 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2245 expected_hlen = sizeof(struct spdk_nvme_tcp_h2c_data_hdr); 2246 pdo = pdu->hdr.common.pdo; 2247 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2248 pdo_error = true; 2249 break; 2250 } 2251 if (pdu->hdr.common.plen < expected_hlen) { 2252 plen_error = true; 2253 } 2254 break; 2255 2256 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2257 expected_hlen = sizeof(struct spdk_nvme_tcp_term_req_hdr); 2258 if ((pdu->hdr.common.plen <= expected_hlen) || 2259 (pdu->hdr.common.plen > SPDK_NVME_TCP_TERM_REQ_PDU_MAX_SIZE)) { 2260 plen_error = true; 2261 } 2262 break; 2263 2264 default: 2265 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", pdu->hdr.common.pdu_type); 2266 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2267 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdu_type); 2268 goto err; 2269 } 2270 } 2271 2272 if (pdu->hdr.common.hlen != expected_hlen) { 2273 SPDK_ERRLOG("PDU type=0x%02x, Expected ICReq header length %u, got %u on tqpair=%p\n", 2274 pdu->hdr.common.pdu_type, 2275 expected_hlen, pdu->hdr.common.hlen, tqpair); 2276 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2277 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, hlen); 2278 goto err; 2279 } else if (pdo_error) { 2280 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2281 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 2282 } else if (plen_error) { 2283 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2284 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2285 goto err; 2286 } else { 2287 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2288 nvme_tcp_pdu_calc_psh_len(tqpair->pdu_in_progress, tqpair->host_hdgst_enable); 2289 return; 2290 } 2291 err: 2292 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2293 } 2294 2295 static int 2296 nvmf_tcp_sock_process(struct spdk_nvmf_tcp_qpair *tqpair) 2297 { 2298 int rc = 0; 2299 struct nvme_tcp_pdu *pdu; 2300 enum nvme_tcp_pdu_recv_state prev_state; 2301 uint32_t data_len; 2302 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2303 struct spdk_nvmf_tcp_transport, transport); 2304 2305 /* The loop here is to allow for several back-to-back state changes. */ 2306 do { 2307 prev_state = tqpair->recv_state; 2308 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv pdu entering state %d\n", tqpair, prev_state); 2309 2310 pdu = tqpair->pdu_in_progress; 2311 assert(pdu != NULL || 2312 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY || 2313 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_QUIESCING || 2314 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 2315 2316 switch (tqpair->recv_state) { 2317 /* Wait for the common header */ 2318 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY: 2319 if (!pdu) { 2320 pdu = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 2321 if (spdk_unlikely(!pdu)) { 2322 return NVME_TCP_PDU_IN_PROGRESS; 2323 } 2324 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 2325 tqpair->pdu_in_progress = pdu; 2326 tqpair->tcp_pdu_working_count++; 2327 } 2328 memset(pdu, 0, offsetof(struct nvme_tcp_pdu, qpair)); 2329 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2330 /* FALLTHROUGH */ 2331 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH: 2332 if (spdk_unlikely(tqpair->state == NVME_TCP_QPAIR_STATE_INITIALIZING)) { 2333 return rc; 2334 } 2335 2336 rc = nvme_tcp_read_data(tqpair->sock, 2337 sizeof(struct spdk_nvme_tcp_common_pdu_hdr) - pdu->ch_valid_bytes, 2338 (void *)&pdu->hdr.common + pdu->ch_valid_bytes); 2339 if (rc < 0) { 2340 SPDK_DEBUGLOG(nvmf_tcp, "will disconnect tqpair=%p\n", tqpair); 2341 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2342 break; 2343 } else if (rc > 0) { 2344 pdu->ch_valid_bytes += rc; 2345 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.qid, rc, 0, tqpair); 2346 } 2347 2348 if (pdu->ch_valid_bytes < sizeof(struct spdk_nvme_tcp_common_pdu_hdr)) { 2349 return NVME_TCP_PDU_IN_PROGRESS; 2350 } 2351 2352 /* The command header of this PDU has now been read from the socket. */ 2353 nvmf_tcp_pdu_ch_handle(tqpair); 2354 break; 2355 /* Wait for the pdu specific header */ 2356 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH: 2357 rc = nvme_tcp_read_data(tqpair->sock, 2358 pdu->psh_len - pdu->psh_valid_bytes, 2359 (void *)&pdu->hdr.raw + sizeof(struct spdk_nvme_tcp_common_pdu_hdr) + pdu->psh_valid_bytes); 2360 if (rc < 0) { 2361 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2362 break; 2363 } else if (rc > 0) { 2364 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.qid, rc, 0, tqpair); 2365 pdu->psh_valid_bytes += rc; 2366 } 2367 2368 if (pdu->psh_valid_bytes < pdu->psh_len) { 2369 return NVME_TCP_PDU_IN_PROGRESS; 2370 } 2371 2372 /* All header(ch, psh, head digist) of this PDU has now been read from the socket. */ 2373 nvmf_tcp_pdu_psh_handle(tqpair, ttransport); 2374 break; 2375 /* Wait for the req slot */ 2376 case NVME_TCP_PDU_RECV_STATE_AWAIT_REQ: 2377 nvmf_tcp_capsule_cmd_hdr_handle(ttransport, tqpair, pdu); 2378 break; 2379 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD: 2380 /* check whether the data is valid, if not we just return */ 2381 if (!pdu->data_len) { 2382 return NVME_TCP_PDU_IN_PROGRESS; 2383 } 2384 2385 data_len = pdu->data_len; 2386 /* data digest */ 2387 if (spdk_unlikely((pdu->hdr.common.pdu_type != SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ) && 2388 tqpair->host_ddgst_enable)) { 2389 data_len += SPDK_NVME_TCP_DIGEST_LEN; 2390 pdu->ddgst_enable = true; 2391 } 2392 2393 rc = nvme_tcp_read_payload_data(tqpair->sock, pdu); 2394 if (rc < 0) { 2395 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2396 break; 2397 } 2398 pdu->rw_offset += rc; 2399 2400 if (pdu->rw_offset < data_len) { 2401 return NVME_TCP_PDU_IN_PROGRESS; 2402 } 2403 2404 /* Generate and insert DIF to whole data block received if DIF is enabled */ 2405 if (spdk_unlikely(pdu->dif_ctx != NULL) && 2406 spdk_dif_generate_stream(pdu->data_iov, pdu->data_iovcnt, 0, data_len, 2407 pdu->dif_ctx) != 0) { 2408 SPDK_ERRLOG("DIF generate failed\n"); 2409 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2410 break; 2411 } 2412 2413 /* All of this PDU has now been read from the socket. */ 2414 nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2415 break; 2416 case NVME_TCP_PDU_RECV_STATE_QUIESCING: 2417 if (tqpair->tcp_pdu_working_count != 0) { 2418 return NVME_TCP_PDU_IN_PROGRESS; 2419 } 2420 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 2421 break; 2422 case NVME_TCP_PDU_RECV_STATE_ERROR: 2423 if (spdk_sock_is_connected(tqpair->sock) && tqpair->wait_terminate) { 2424 return NVME_TCP_PDU_IN_PROGRESS; 2425 } 2426 return NVME_TCP_PDU_FATAL; 2427 default: 2428 SPDK_ERRLOG("The state(%d) is invalid\n", tqpair->recv_state); 2429 abort(); 2430 break; 2431 } 2432 } while (tqpair->recv_state != prev_state); 2433 2434 return rc; 2435 } 2436 2437 static inline void * 2438 nvmf_tcp_control_msg_get(struct spdk_nvmf_tcp_control_msg_list *list) 2439 { 2440 struct spdk_nvmf_tcp_control_msg *msg; 2441 2442 assert(list); 2443 2444 msg = STAILQ_FIRST(&list->free_msgs); 2445 if (!msg) { 2446 SPDK_DEBUGLOG(nvmf_tcp, "Out of control messages\n"); 2447 return NULL; 2448 } 2449 STAILQ_REMOVE_HEAD(&list->free_msgs, link); 2450 return msg; 2451 } 2452 2453 static inline void 2454 nvmf_tcp_control_msg_put(struct spdk_nvmf_tcp_control_msg_list *list, void *_msg) 2455 { 2456 struct spdk_nvmf_tcp_control_msg *msg = _msg; 2457 2458 assert(list); 2459 STAILQ_INSERT_HEAD(&list->free_msgs, msg, link); 2460 } 2461 2462 static int 2463 nvmf_tcp_req_parse_sgl(struct spdk_nvmf_tcp_req *tcp_req, 2464 struct spdk_nvmf_transport *transport, 2465 struct spdk_nvmf_transport_poll_group *group) 2466 { 2467 struct spdk_nvmf_request *req = &tcp_req->req; 2468 struct spdk_nvme_cmd *cmd; 2469 struct spdk_nvme_sgl_descriptor *sgl; 2470 struct spdk_nvmf_tcp_poll_group *tgroup; 2471 enum spdk_nvme_tcp_term_req_fes fes; 2472 struct nvme_tcp_pdu *pdu; 2473 struct spdk_nvmf_tcp_qpair *tqpair; 2474 uint32_t length, error_offset = 0; 2475 2476 cmd = &req->cmd->nvme_cmd; 2477 sgl = &cmd->dptr.sgl1; 2478 2479 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK && 2480 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_TRANSPORT) { 2481 /* get request length from sgl */ 2482 length = sgl->unkeyed.length; 2483 if (spdk_unlikely(length > transport->opts.max_io_size)) { 2484 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 2485 length, transport->opts.max_io_size); 2486 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_LIMIT_EXCEEDED; 2487 goto fatal_err; 2488 } 2489 2490 /* fill request length and populate iovs */ 2491 req->length = length; 2492 2493 SPDK_DEBUGLOG(nvmf_tcp, "Data requested length= 0x%x\n", length); 2494 2495 if (spdk_unlikely(req->dif_enabled)) { 2496 req->dif.orig_length = length; 2497 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2498 req->dif.elba_length = length; 2499 } 2500 2501 if (nvmf_ctrlr_use_zcopy(req)) { 2502 SPDK_DEBUGLOG(nvmf_tcp, "Using zero-copy to execute request %p\n", tcp_req); 2503 req->data_from_pool = false; 2504 return 0; 2505 } 2506 2507 if (spdk_nvmf_request_get_buffers(req, group, transport, length)) { 2508 /* No available buffers. Queue this request up. */ 2509 SPDK_DEBUGLOG(nvmf_tcp, "No available large data buffers. Queueing request %p\n", 2510 tcp_req); 2511 return 0; 2512 } 2513 2514 SPDK_DEBUGLOG(nvmf_tcp, "Request %p took %d buffer/s from central pool, and data=%p\n", 2515 tcp_req, req->iovcnt, req->iov[0].iov_base); 2516 2517 return 0; 2518 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 2519 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 2520 uint64_t offset = sgl->address; 2521 uint32_t max_len = transport->opts.in_capsule_data_size; 2522 2523 assert(tcp_req->has_in_capsule_data); 2524 /* Capsule Cmd with In-capsule Data should get data length from pdu header */ 2525 tqpair = tcp_req->pdu->qpair; 2526 /* receiving pdu is not same with the pdu in tcp_req */ 2527 pdu = tqpair->pdu_in_progress; 2528 length = pdu->hdr.common.plen - pdu->psh_len - sizeof(struct spdk_nvme_tcp_common_pdu_hdr); 2529 if (tqpair->host_ddgst_enable) { 2530 length -= SPDK_NVME_TCP_DIGEST_LEN; 2531 } 2532 /* This error is not defined in NVMe/TCP spec, take this error as fatal error */ 2533 if (spdk_unlikely(length != sgl->unkeyed.length)) { 2534 SPDK_ERRLOG("In-Capsule Data length 0x%x is not equal to SGL data length 0x%x\n", 2535 length, sgl->unkeyed.length); 2536 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2537 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2538 goto fatal_err; 2539 } 2540 2541 SPDK_DEBUGLOG(nvmf_tcp, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 2542 offset, length); 2543 2544 /* The NVMe/TCP transport does not use ICDOFF to control the in-capsule data offset. ICDOFF should be '0' */ 2545 if (spdk_unlikely(offset != 0)) { 2546 /* Not defined fatal error in NVMe/TCP spec, handle this error as a fatal error */ 2547 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " should be ZERO in NVMe/TCP\n", offset); 2548 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER; 2549 error_offset = offsetof(struct spdk_nvme_tcp_cmd, ccsqe.dptr.sgl1.address); 2550 goto fatal_err; 2551 } 2552 2553 if (spdk_unlikely(length > max_len)) { 2554 /* According to the SPEC we should support ICD up to 8192 bytes for admin and fabric commands */ 2555 if (length <= SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE && 2556 (cmd->opc == SPDK_NVME_OPC_FABRIC || req->qpair->qid == 0)) { 2557 2558 /* Get a buffer from dedicated list */ 2559 SPDK_DEBUGLOG(nvmf_tcp, "Getting a buffer from control msg list\n"); 2560 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 2561 assert(tgroup->control_msg_list); 2562 req->iov[0].iov_base = nvmf_tcp_control_msg_get(tgroup->control_msg_list); 2563 if (!req->iov[0].iov_base) { 2564 /* No available buffers. Queue this request up. */ 2565 SPDK_DEBUGLOG(nvmf_tcp, "No available ICD buffers. Queueing request %p\n", tcp_req); 2566 return 0; 2567 } 2568 } else { 2569 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 2570 length, max_len); 2571 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_LIMIT_EXCEEDED; 2572 goto fatal_err; 2573 } 2574 } else { 2575 req->iov[0].iov_base = tcp_req->buf; 2576 } 2577 2578 req->length = length; 2579 req->data_from_pool = false; 2580 2581 if (spdk_unlikely(req->dif_enabled)) { 2582 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2583 req->dif.elba_length = length; 2584 } 2585 2586 req->iov[0].iov_len = length; 2587 req->iovcnt = 1; 2588 2589 return 0; 2590 } 2591 /* If we want to handle the problem here, then we can't skip the following data segment. 2592 * Because this function runs before reading data part, now handle all errors as fatal errors. */ 2593 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 2594 sgl->generic.type, sgl->generic.subtype); 2595 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER; 2596 error_offset = offsetof(struct spdk_nvme_tcp_cmd, ccsqe.dptr.sgl1.generic); 2597 fatal_err: 2598 nvmf_tcp_send_c2h_term_req(tcp_req->pdu->qpair, tcp_req->pdu, fes, error_offset); 2599 return -1; 2600 } 2601 2602 static inline enum spdk_nvme_media_error_status_code 2603 nvmf_tcp_dif_error_to_compl_status(uint8_t err_type) { 2604 enum spdk_nvme_media_error_status_code result; 2605 2606 switch (err_type) 2607 { 2608 case SPDK_DIF_REFTAG_ERROR: 2609 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 2610 break; 2611 case SPDK_DIF_APPTAG_ERROR: 2612 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 2613 break; 2614 case SPDK_DIF_GUARD_ERROR: 2615 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 2616 break; 2617 default: 2618 SPDK_UNREACHABLE(); 2619 break; 2620 } 2621 2622 return result; 2623 } 2624 2625 static void 2626 _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2627 struct spdk_nvmf_tcp_req *tcp_req) 2628 { 2629 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF( 2630 tqpair->qpair.transport, struct spdk_nvmf_tcp_transport, transport); 2631 struct nvme_tcp_pdu *rsp_pdu; 2632 struct spdk_nvme_tcp_c2h_data_hdr *c2h_data; 2633 uint32_t plen, pdo, alignment; 2634 int rc; 2635 2636 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2637 2638 rsp_pdu = tcp_req->pdu; 2639 assert(rsp_pdu != NULL); 2640 2641 c2h_data = &rsp_pdu->hdr.c2h_data; 2642 c2h_data->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_DATA; 2643 plen = c2h_data->common.hlen = sizeof(*c2h_data); 2644 2645 if (tqpair->host_hdgst_enable) { 2646 plen += SPDK_NVME_TCP_DIGEST_LEN; 2647 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2648 } 2649 2650 /* set the psh */ 2651 c2h_data->cccid = tcp_req->req.cmd->nvme_cmd.cid; 2652 c2h_data->datal = tcp_req->req.length - tcp_req->pdu->rw_offset; 2653 c2h_data->datao = tcp_req->pdu->rw_offset; 2654 2655 /* set the padding */ 2656 rsp_pdu->padding_len = 0; 2657 pdo = plen; 2658 if (tqpair->cpda) { 2659 alignment = (tqpair->cpda + 1) << 2; 2660 if (plen % alignment != 0) { 2661 pdo = (plen + alignment) / alignment * alignment; 2662 rsp_pdu->padding_len = pdo - plen; 2663 plen = pdo; 2664 } 2665 } 2666 2667 c2h_data->common.pdo = pdo; 2668 plen += c2h_data->datal; 2669 if (tqpair->host_ddgst_enable) { 2670 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_DDGSTF; 2671 plen += SPDK_NVME_TCP_DIGEST_LEN; 2672 } 2673 2674 c2h_data->common.plen = plen; 2675 2676 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2677 rsp_pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 2678 } 2679 2680 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2681 c2h_data->datao, c2h_data->datal); 2682 2683 2684 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU; 2685 /* Need to send the capsule response if response is not all 0 */ 2686 if (ttransport->tcp_opts.c2h_success && 2687 tcp_req->rsp.cdw0 == 0 && tcp_req->rsp.cdw1 == 0) { 2688 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS; 2689 } 2690 2691 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2692 struct spdk_nvme_cpl *rsp = &tcp_req->req.rsp->nvme_cpl; 2693 struct spdk_dif_error err_blk = {}; 2694 uint32_t mapped_length = 0; 2695 uint32_t available_iovs = SPDK_COUNTOF(rsp_pdu->iov); 2696 uint32_t ddgst_len = 0; 2697 2698 if (tqpair->host_ddgst_enable) { 2699 /* Data digest consumes additional iov entry */ 2700 available_iovs--; 2701 /* plen needs to be updated since nvme_tcp_build_iovs compares expected and actual plen */ 2702 ddgst_len = SPDK_NVME_TCP_DIGEST_LEN; 2703 c2h_data->common.plen -= ddgst_len; 2704 } 2705 /* Temp call to estimate if data can be described by limited number of iovs. 2706 * iov vector will be rebuilt in nvmf_tcp_qpair_write_pdu */ 2707 nvme_tcp_build_iovs(rsp_pdu->iov, available_iovs, rsp_pdu, tqpair->host_hdgst_enable, 2708 false, &mapped_length); 2709 2710 if (mapped_length != c2h_data->common.plen) { 2711 c2h_data->datal = mapped_length - (c2h_data->common.plen - c2h_data->datal); 2712 SPDK_DEBUGLOG(nvmf_tcp, 2713 "Part C2H, data_len %u (of %u), PDU len %u, updated PDU len %u, offset %u\n", 2714 c2h_data->datal, tcp_req->req.length, c2h_data->common.plen, mapped_length, rsp_pdu->rw_offset); 2715 c2h_data->common.plen = mapped_length; 2716 2717 /* Rebuild pdu->data_iov since data length is changed */ 2718 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, c2h_data->datao, 2719 c2h_data->datal); 2720 2721 c2h_data->common.flags &= ~(SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU | 2722 SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS); 2723 } 2724 2725 c2h_data->common.plen += ddgst_len; 2726 2727 assert(rsp_pdu->rw_offset <= tcp_req->req.length); 2728 2729 rc = spdk_dif_verify_stream(rsp_pdu->data_iov, rsp_pdu->data_iovcnt, 2730 0, rsp_pdu->data_len, rsp_pdu->dif_ctx, &err_blk); 2731 if (rc != 0) { 2732 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", 2733 err_blk.err_type, err_blk.err_offset); 2734 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2735 rsp->status.sc = nvmf_tcp_dif_error_to_compl_status(err_blk.err_type); 2736 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2737 return; 2738 } 2739 } 2740 2741 rsp_pdu->rw_offset += c2h_data->datal; 2742 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_pdu_c2h_data_complete, tcp_req); 2743 } 2744 2745 static void 2746 nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2747 struct spdk_nvmf_tcp_req *tcp_req) 2748 { 2749 nvmf_tcp_req_pdu_init(tcp_req); 2750 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2751 } 2752 2753 static int 2754 request_transfer_out(struct spdk_nvmf_request *req) 2755 { 2756 struct spdk_nvmf_tcp_req *tcp_req; 2757 struct spdk_nvmf_qpair *qpair; 2758 struct spdk_nvmf_tcp_qpair *tqpair; 2759 struct spdk_nvme_cpl *rsp; 2760 2761 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2762 2763 qpair = req->qpair; 2764 rsp = &req->rsp->nvme_cpl; 2765 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 2766 2767 /* Advance our sq_head pointer */ 2768 if (qpair->sq_head == qpair->sq_head_max) { 2769 qpair->sq_head = 0; 2770 } else { 2771 qpair->sq_head++; 2772 } 2773 rsp->sqhd = qpair->sq_head; 2774 2775 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2776 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2777 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2778 nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2779 } else { 2780 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2781 } 2782 2783 return 0; 2784 } 2785 2786 static void 2787 nvmf_tcp_check_fused_ordering(struct spdk_nvmf_tcp_transport *ttransport, 2788 struct spdk_nvmf_tcp_qpair *tqpair, 2789 struct spdk_nvmf_tcp_req *tcp_req) 2790 { 2791 enum spdk_nvme_cmd_fuse last, next; 2792 2793 last = tqpair->fused_first ? tqpair->fused_first->cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2794 next = tcp_req->cmd.fuse; 2795 2796 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2797 2798 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2799 return; 2800 } 2801 2802 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2803 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2804 /* This is a valid pair of fused commands. Point them at each other 2805 * so they can be submitted consecutively once ready to be executed. 2806 */ 2807 tqpair->fused_first->fused_pair = tcp_req; 2808 tcp_req->fused_pair = tqpair->fused_first; 2809 tqpair->fused_first = NULL; 2810 return; 2811 } else { 2812 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2813 tqpair->fused_first->fused_failed = true; 2814 2815 /* 2816 * If the last req is in READY_TO_EXECUTE state, then call 2817 * nvmf_tcp_req_process(), otherwise nothing else will kick it. 2818 */ 2819 if (tqpair->fused_first->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 2820 nvmf_tcp_req_process(ttransport, tqpair->fused_first); 2821 } 2822 2823 tqpair->fused_first = NULL; 2824 } 2825 } 2826 2827 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2828 /* Set tqpair->fused_first here so that we know to check that the next request 2829 * is a SECOND (and to fail this one if it isn't). 2830 */ 2831 tqpair->fused_first = tcp_req; 2832 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2833 /* Mark this req failed since it is a SECOND and the last one was not a FIRST. */ 2834 tcp_req->fused_failed = true; 2835 } 2836 } 2837 2838 static bool 2839 nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 2840 struct spdk_nvmf_tcp_req *tcp_req) 2841 { 2842 struct spdk_nvmf_tcp_qpair *tqpair; 2843 uint32_t plen; 2844 struct nvme_tcp_pdu *pdu; 2845 enum spdk_nvmf_tcp_req_state prev_state; 2846 bool progress = false; 2847 struct spdk_nvmf_transport *transport = &ttransport->transport; 2848 struct spdk_nvmf_transport_poll_group *group; 2849 struct spdk_nvmf_tcp_poll_group *tgroup; 2850 2851 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2852 group = &tqpair->group->group; 2853 assert(tcp_req->state != TCP_REQUEST_STATE_FREE); 2854 2855 /* If the qpair is not active, we need to abort the outstanding requests. */ 2856 if (tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2857 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 2858 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2859 } 2860 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 2861 } 2862 2863 /* The loop here is to allow for several back-to-back state changes. */ 2864 do { 2865 prev_state = tcp_req->state; 2866 2867 SPDK_DEBUGLOG(nvmf_tcp, "Request %p entering state %d on tqpair=%p\n", tcp_req, prev_state, 2868 tqpair); 2869 2870 switch (tcp_req->state) { 2871 case TCP_REQUEST_STATE_FREE: 2872 /* Some external code must kick a request into TCP_REQUEST_STATE_NEW 2873 * to escape this state. */ 2874 break; 2875 case TCP_REQUEST_STATE_NEW: 2876 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEW, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, tqpair); 2877 2878 /* copy the cmd from the receive pdu */ 2879 tcp_req->cmd = tqpair->pdu_in_progress->hdr.capsule_cmd.ccsqe; 2880 2881 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&tcp_req->req, &tcp_req->req.dif.dif_ctx))) { 2882 tcp_req->req.dif_enabled = true; 2883 tqpair->pdu_in_progress->dif_ctx = &tcp_req->req.dif.dif_ctx; 2884 } 2885 2886 nvmf_tcp_check_fused_ordering(ttransport, tqpair, tcp_req); 2887 2888 /* The next state transition depends on the data transfer needs of this request. */ 2889 tcp_req->req.xfer = spdk_nvmf_req_get_xfer(&tcp_req->req); 2890 2891 if (spdk_unlikely(tcp_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2892 nvmf_tcp_req_set_cpl(tcp_req, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_INVALID_OPCODE); 2893 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2894 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2895 SPDK_DEBUGLOG(nvmf_tcp, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", tcp_req); 2896 break; 2897 } 2898 2899 /* If no data to transfer, ready to execute. */ 2900 if (tcp_req->req.xfer == SPDK_NVME_DATA_NONE) { 2901 /* Reset the tqpair receiving pdu state */ 2902 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2903 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2904 break; 2905 } 2906 2907 pdu = tqpair->pdu_in_progress; 2908 plen = pdu->hdr.common.hlen; 2909 if (tqpair->host_hdgst_enable) { 2910 plen += SPDK_NVME_TCP_DIGEST_LEN; 2911 } 2912 if (pdu->hdr.common.plen != plen) { 2913 tcp_req->has_in_capsule_data = true; 2914 } else { 2915 /* Data is transmitted by C2H PDUs */ 2916 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2917 } 2918 2919 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEED_BUFFER); 2920 STAILQ_INSERT_TAIL(&group->pending_buf_queue, &tcp_req->req, buf_link); 2921 break; 2922 case TCP_REQUEST_STATE_NEED_BUFFER: 2923 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEED_BUFFER, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2924 tqpair); 2925 2926 assert(tcp_req->req.xfer != SPDK_NVME_DATA_NONE); 2927 2928 if (!tcp_req->has_in_capsule_data && (&tcp_req->req != STAILQ_FIRST(&group->pending_buf_queue))) { 2929 SPDK_DEBUGLOG(nvmf_tcp, 2930 "Not the first element to wait for the buf for tcp_req(%p) on tqpair=%p\n", 2931 tcp_req, tqpair); 2932 /* This request needs to wait in line to obtain a buffer */ 2933 break; 2934 } 2935 2936 /* Try to get a data buffer */ 2937 if (nvmf_tcp_req_parse_sgl(tcp_req, transport, group) < 0) { 2938 break; 2939 } 2940 2941 /* Get a zcopy buffer if the request can be serviced through zcopy */ 2942 if (spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 2943 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2944 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 2945 tcp_req->req.length = tcp_req->req.dif.elba_length; 2946 } 2947 2948 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2949 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_START); 2950 spdk_nvmf_request_zcopy_start(&tcp_req->req); 2951 break; 2952 } 2953 2954 if (tcp_req->req.iovcnt < 1) { 2955 SPDK_DEBUGLOG(nvmf_tcp, "No buffer allocated for tcp_req(%p) on tqpair(%p\n)", 2956 tcp_req, tqpair); 2957 /* No buffers available. */ 2958 break; 2959 } 2960 2961 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2962 2963 /* If data is transferring from host to controller, we need to do a transfer from the host. */ 2964 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2965 if (tcp_req->req.data_from_pool) { 2966 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 2967 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 2968 } else { 2969 struct nvme_tcp_pdu *pdu; 2970 2971 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2972 2973 pdu = tqpair->pdu_in_progress; 2974 SPDK_DEBUGLOG(nvmf_tcp, "Not need to send r2t for tcp_req(%p) on tqpair=%p\n", tcp_req, 2975 tqpair); 2976 /* No need to send r2t, contained in the capsuled data */ 2977 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2978 0, tcp_req->req.length); 2979 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2980 } 2981 break; 2982 } 2983 2984 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2985 break; 2986 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 2987 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, tqpair->qpair.qid, 0, 2988 (uintptr_t)tcp_req, tqpair); 2989 /* Some external code must kick a request into TCP_REQUEST_STATE_ZCOPY_START_COMPLETED 2990 * to escape this state. */ 2991 break; 2992 case TCP_REQUEST_STATE_ZCOPY_START_COMPLETED: 2993 spdk_trace_record(TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, tqpair->qpair.qid, 0, 2994 (uintptr_t)tcp_req, tqpair); 2995 if (spdk_unlikely(spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl))) { 2996 SPDK_DEBUGLOG(nvmf_tcp, "Zero-copy start failed for tcp_req(%p) on tqpair=%p\n", 2997 tcp_req, tqpair); 2998 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2999 break; 3000 } 3001 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 3002 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 3003 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 3004 } else { 3005 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3006 } 3007 break; 3008 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3009 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3010 tqpair); 3011 /* The R2T completion or the h2c data incoming will kick it out of this state. */ 3012 break; 3013 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3014 3015 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, tqpair->qpair.qid, 0, 3016 (uintptr_t)tcp_req, tqpair); 3017 /* Some external code must kick a request into TCP_REQUEST_STATE_READY_TO_EXECUTE 3018 * to escape this state. */ 3019 break; 3020 case TCP_REQUEST_STATE_READY_TO_EXECUTE: 3021 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, tqpair->qpair.qid, 0, 3022 (uintptr_t)tcp_req, tqpair); 3023 3024 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3025 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 3026 tcp_req->req.length = tcp_req->req.dif.elba_length; 3027 } 3028 3029 if (tcp_req->cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 3030 if (tcp_req->fused_failed) { 3031 /* This request failed FUSED semantics. Fail it immediately, without 3032 * even sending it to the target layer. 3033 */ 3034 nvmf_tcp_req_set_cpl(tcp_req, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_ABORTED_MISSING_FUSED); 3035 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3036 break; 3037 } 3038 3039 if (tcp_req->fused_pair == NULL || 3040 tcp_req->fused_pair->state != TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3041 /* This request is ready to execute, but either we don't know yet if it's 3042 * valid - i.e. this is a FIRST but we haven't received the next request yet), 3043 * or the other request of this fused pair isn't ready to execute. So 3044 * break here and this request will get processed later either when the 3045 * other request is ready or we find that this request isn't valid. 3046 */ 3047 break; 3048 } 3049 } 3050 3051 if (!spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 3052 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTING); 3053 /* If we get to this point, and this request is a fused command, we know that 3054 * it is part of a valid sequence (FIRST followed by a SECOND) and that both 3055 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 3056 * request, and the other request of the fused pair, in the correct order. 3057 * Also clear the ->fused_pair pointers on both requests, since after this point 3058 * we no longer need to maintain the relationship between these two requests. 3059 */ 3060 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 3061 assert(tcp_req->fused_pair != NULL); 3062 assert(tcp_req->fused_pair->fused_pair == tcp_req); 3063 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 3064 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 3065 tcp_req->fused_pair->fused_pair = NULL; 3066 tcp_req->fused_pair = NULL; 3067 } 3068 spdk_nvmf_request_exec(&tcp_req->req); 3069 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 3070 assert(tcp_req->fused_pair != NULL); 3071 assert(tcp_req->fused_pair->fused_pair == tcp_req); 3072 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 3073 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 3074 tcp_req->fused_pair->fused_pair = NULL; 3075 tcp_req->fused_pair = NULL; 3076 } 3077 } else { 3078 /* For zero-copy, only requests with data coming from host to the 3079 * controller can end up here. */ 3080 assert(tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 3081 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT); 3082 spdk_nvmf_request_zcopy_end(&tcp_req->req, true); 3083 } 3084 3085 break; 3086 case TCP_REQUEST_STATE_EXECUTING: 3087 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTING, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3088 tqpair); 3089 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 3090 * to escape this state. */ 3091 break; 3092 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3093 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, tqpair->qpair.qid, 0, 3094 (uintptr_t)tcp_req, tqpair); 3095 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 3096 * to escape this state. */ 3097 break; 3098 case TCP_REQUEST_STATE_EXECUTED: 3099 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTED, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3100 tqpair); 3101 3102 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3103 tcp_req->req.length = tcp_req->req.dif.orig_length; 3104 } 3105 3106 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3107 break; 3108 case TCP_REQUEST_STATE_READY_TO_COMPLETE: 3109 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, tqpair->qpair.qid, 0, 3110 (uintptr_t)tcp_req, tqpair); 3111 if (request_transfer_out(&tcp_req->req) != 0) { 3112 assert(0); /* No good way to handle this currently */ 3113 } 3114 break; 3115 case TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3116 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, tqpair->qpair.qid, 0, 3117 (uintptr_t)tcp_req, tqpair); 3118 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 3119 * to escape this state. */ 3120 break; 3121 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3122 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, tqpair->qpair.qid, 0, 3123 (uintptr_t)tcp_req, tqpair); 3124 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 3125 * to escape this state. */ 3126 break; 3127 case TCP_REQUEST_STATE_COMPLETED: 3128 spdk_trace_record(TRACE_TCP_REQUEST_STATE_COMPLETED, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3129 tqpair); 3130 /* If there's an outstanding PDU sent to the host, the request is completed 3131 * due to the qpair being disconnected. We must delay the completion until 3132 * that write is done to avoid freeing the request twice. */ 3133 if (spdk_unlikely(tcp_req->pdu_in_use)) { 3134 SPDK_DEBUGLOG(nvmf_tcp, "Delaying completion due to outstanding " 3135 "write on req=%p\n", tcp_req); 3136 /* This can only happen for zcopy requests */ 3137 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 3138 assert(tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE); 3139 break; 3140 } 3141 3142 if (tcp_req->req.data_from_pool) { 3143 spdk_nvmf_request_free_buffers(&tcp_req->req, group, transport); 3144 } else if (spdk_unlikely(tcp_req->has_in_capsule_data && 3145 (tcp_req->cmd.opc == SPDK_NVME_OPC_FABRIC || 3146 tqpair->qpair.qid == 0) && tcp_req->req.length > transport->opts.in_capsule_data_size)) { 3147 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3148 assert(tgroup->control_msg_list); 3149 SPDK_DEBUGLOG(nvmf_tcp, "Put buf to control msg list\n"); 3150 nvmf_tcp_control_msg_put(tgroup->control_msg_list, 3151 tcp_req->req.iov[0].iov_base); 3152 } else if (tcp_req->req.zcopy_bdev_io != NULL) { 3153 /* If the request has an unreleased zcopy bdev_io, it's either a 3154 * read, a failed write, or the qpair is being disconnected */ 3155 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 3156 assert(tcp_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST || 3157 spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl) || 3158 tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE); 3159 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE); 3160 spdk_nvmf_request_zcopy_end(&tcp_req->req, false); 3161 break; 3162 } 3163 tcp_req->req.length = 0; 3164 tcp_req->req.iovcnt = 0; 3165 tcp_req->fused_failed = false; 3166 if (tcp_req->fused_pair) { 3167 /* This req was part of a valid fused pair, but failed before it got to 3168 * READ_TO_EXECUTE state. This means we need to fail the other request 3169 * in the pair, because it is no longer part of a valid pair. If the pair 3170 * already reached READY_TO_EXECUTE state, we need to kick it. 3171 */ 3172 tcp_req->fused_pair->fused_failed = true; 3173 if (tcp_req->fused_pair->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3174 nvmf_tcp_req_process(ttransport, tcp_req->fused_pair); 3175 } 3176 tcp_req->fused_pair = NULL; 3177 } 3178 3179 nvmf_tcp_req_put(tqpair, tcp_req); 3180 break; 3181 case TCP_REQUEST_NUM_STATES: 3182 default: 3183 assert(0); 3184 break; 3185 } 3186 3187 if (tcp_req->state != prev_state) { 3188 progress = true; 3189 } 3190 } while (tcp_req->state != prev_state); 3191 3192 return progress; 3193 } 3194 3195 static void 3196 nvmf_tcp_sock_cb(void *arg, struct spdk_sock_group *group, struct spdk_sock *sock) 3197 { 3198 struct spdk_nvmf_tcp_qpair *tqpair = arg; 3199 int rc; 3200 3201 assert(tqpair != NULL); 3202 rc = nvmf_tcp_sock_process(tqpair); 3203 3204 /* If there was a new socket error, disconnect */ 3205 if (rc < 0) { 3206 nvmf_tcp_qpair_disconnect(tqpair); 3207 } 3208 } 3209 3210 static int 3211 nvmf_tcp_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3212 struct spdk_nvmf_qpair *qpair) 3213 { 3214 struct spdk_nvmf_tcp_poll_group *tgroup; 3215 struct spdk_nvmf_tcp_qpair *tqpair; 3216 int rc; 3217 3218 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3219 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3220 3221 rc = nvmf_tcp_qpair_sock_init(tqpair); 3222 if (rc != 0) { 3223 SPDK_ERRLOG("Cannot set sock opt for tqpair=%p\n", tqpair); 3224 return -1; 3225 } 3226 3227 rc = nvmf_tcp_qpair_init(&tqpair->qpair); 3228 if (rc < 0) { 3229 SPDK_ERRLOG("Cannot init tqpair=%p\n", tqpair); 3230 return -1; 3231 } 3232 3233 rc = nvmf_tcp_qpair_init_mem_resource(tqpair); 3234 if (rc < 0) { 3235 SPDK_ERRLOG("Cannot init memory resource info for tqpair=%p\n", tqpair); 3236 return -1; 3237 } 3238 3239 rc = spdk_sock_group_add_sock(tgroup->sock_group, tqpair->sock, 3240 nvmf_tcp_sock_cb, tqpair); 3241 if (rc != 0) { 3242 SPDK_ERRLOG("Could not add sock to sock_group: %s (%d)\n", 3243 spdk_strerror(errno), errno); 3244 return -1; 3245 } 3246 3247 tqpair->group = tgroup; 3248 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INVALID); 3249 TAILQ_INSERT_TAIL(&tgroup->qpairs, tqpair, link); 3250 3251 return 0; 3252 } 3253 3254 static int 3255 nvmf_tcp_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3256 struct spdk_nvmf_qpair *qpair) 3257 { 3258 struct spdk_nvmf_tcp_poll_group *tgroup; 3259 struct spdk_nvmf_tcp_qpair *tqpair; 3260 int rc; 3261 3262 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3263 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3264 3265 assert(tqpair->group == tgroup); 3266 3267 SPDK_DEBUGLOG(nvmf_tcp, "remove tqpair=%p from the tgroup=%p\n", tqpair, tgroup); 3268 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 3269 TAILQ_REMOVE(&tgroup->await_req, tqpair, link); 3270 } else { 3271 TAILQ_REMOVE(&tgroup->qpairs, tqpair, link); 3272 } 3273 3274 rc = spdk_sock_group_remove_sock(tgroup->sock_group, tqpair->sock); 3275 if (rc != 0) { 3276 SPDK_ERRLOG("Could not remove sock from sock_group: %s (%d)\n", 3277 spdk_strerror(errno), errno); 3278 } 3279 3280 return rc; 3281 } 3282 3283 static int 3284 nvmf_tcp_req_complete(struct spdk_nvmf_request *req) 3285 { 3286 struct spdk_nvmf_tcp_transport *ttransport; 3287 struct spdk_nvmf_tcp_req *tcp_req; 3288 3289 ttransport = SPDK_CONTAINEROF(req->qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3290 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3291 3292 switch (tcp_req->state) { 3293 case TCP_REQUEST_STATE_EXECUTING: 3294 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3295 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3296 break; 3297 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3298 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_ZCOPY_START_COMPLETED); 3299 break; 3300 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3301 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 3302 break; 3303 default: 3304 assert(0 && "Unexpected request state"); 3305 break; 3306 } 3307 3308 nvmf_tcp_req_process(ttransport, tcp_req); 3309 3310 return 0; 3311 } 3312 3313 static void 3314 nvmf_tcp_close_qpair(struct spdk_nvmf_qpair *qpair, 3315 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3316 { 3317 struct spdk_nvmf_tcp_qpair *tqpair; 3318 3319 SPDK_DEBUGLOG(nvmf_tcp, "Qpair: %p\n", qpair); 3320 3321 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3322 3323 assert(tqpair->fini_cb_fn == NULL); 3324 tqpair->fini_cb_fn = cb_fn; 3325 tqpair->fini_cb_arg = cb_arg; 3326 3327 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITED); 3328 nvmf_tcp_qpair_destroy(tqpair); 3329 } 3330 3331 static int 3332 nvmf_tcp_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3333 { 3334 struct spdk_nvmf_tcp_poll_group *tgroup; 3335 int rc; 3336 struct spdk_nvmf_request *req, *req_tmp; 3337 struct spdk_nvmf_tcp_req *tcp_req; 3338 struct spdk_nvmf_tcp_qpair *tqpair, *tqpair_tmp; 3339 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(group->transport, 3340 struct spdk_nvmf_tcp_transport, transport); 3341 3342 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3343 3344 if (spdk_unlikely(TAILQ_EMPTY(&tgroup->qpairs) && TAILQ_EMPTY(&tgroup->await_req))) { 3345 return 0; 3346 } 3347 3348 STAILQ_FOREACH_SAFE(req, &group->pending_buf_queue, buf_link, req_tmp) { 3349 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3350 if (nvmf_tcp_req_process(ttransport, tcp_req) == false) { 3351 break; 3352 } 3353 } 3354 3355 rc = spdk_sock_group_poll(tgroup->sock_group); 3356 if (rc < 0) { 3357 SPDK_ERRLOG("Failed to poll sock_group=%p\n", tgroup->sock_group); 3358 } 3359 3360 TAILQ_FOREACH_SAFE(tqpair, &tgroup->await_req, link, tqpair_tmp) { 3361 rc = nvmf_tcp_sock_process(tqpair); 3362 3363 /* If there was a new socket error, disconnect */ 3364 if (rc < 0) { 3365 nvmf_tcp_qpair_disconnect(tqpair); 3366 } 3367 } 3368 3369 return rc; 3370 } 3371 3372 static int 3373 nvmf_tcp_qpair_get_trid(struct spdk_nvmf_qpair *qpair, 3374 struct spdk_nvme_transport_id *trid, bool peer) 3375 { 3376 struct spdk_nvmf_tcp_qpair *tqpair; 3377 uint16_t port; 3378 3379 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3380 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_TCP); 3381 3382 if (peer) { 3383 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->initiator_addr); 3384 port = tqpair->initiator_port; 3385 } else { 3386 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->target_addr); 3387 port = tqpair->target_port; 3388 } 3389 3390 if (spdk_sock_is_ipv4(tqpair->sock)) { 3391 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3392 } else if (spdk_sock_is_ipv6(tqpair->sock)) { 3393 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3394 } else { 3395 return -1; 3396 } 3397 3398 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%d", port); 3399 return 0; 3400 } 3401 3402 static int 3403 nvmf_tcp_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3404 struct spdk_nvme_transport_id *trid) 3405 { 3406 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3407 } 3408 3409 static int 3410 nvmf_tcp_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3411 struct spdk_nvme_transport_id *trid) 3412 { 3413 return nvmf_tcp_qpair_get_trid(qpair, trid, 1); 3414 } 3415 3416 static int 3417 nvmf_tcp_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3418 struct spdk_nvme_transport_id *trid) 3419 { 3420 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3421 } 3422 3423 static void 3424 nvmf_tcp_req_set_abort_status(struct spdk_nvmf_request *req, 3425 struct spdk_nvmf_tcp_req *tcp_req_to_abort) 3426 { 3427 nvmf_tcp_req_set_cpl(tcp_req_to_abort, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_ABORTED_BY_REQUEST); 3428 nvmf_tcp_req_set_state(tcp_req_to_abort, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3429 3430 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 3431 } 3432 3433 static int 3434 _nvmf_tcp_qpair_abort_request(void *ctx) 3435 { 3436 struct spdk_nvmf_request *req = ctx; 3437 struct spdk_nvmf_tcp_req *tcp_req_to_abort = SPDK_CONTAINEROF(req->req_to_abort, 3438 struct spdk_nvmf_tcp_req, req); 3439 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 3440 struct spdk_nvmf_tcp_qpair, qpair); 3441 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 3442 struct spdk_nvmf_tcp_transport, transport); 3443 int rc; 3444 3445 spdk_poller_unregister(&req->poller); 3446 3447 switch (tcp_req_to_abort->state) { 3448 case TCP_REQUEST_STATE_EXECUTING: 3449 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3450 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3451 rc = nvmf_ctrlr_abort_request(req); 3452 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 3453 return SPDK_POLLER_BUSY; 3454 } 3455 break; 3456 3457 case TCP_REQUEST_STATE_NEED_BUFFER: 3458 STAILQ_REMOVE(&tqpair->group->group.pending_buf_queue, 3459 &tcp_req_to_abort->req, spdk_nvmf_request, buf_link); 3460 3461 nvmf_tcp_req_set_abort_status(req, tcp_req_to_abort); 3462 nvmf_tcp_req_process(ttransport, tcp_req_to_abort); 3463 break; 3464 3465 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3466 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3467 if (spdk_get_ticks() < req->timeout_tsc) { 3468 req->poller = SPDK_POLLER_REGISTER(_nvmf_tcp_qpair_abort_request, req, 0); 3469 return SPDK_POLLER_BUSY; 3470 } 3471 break; 3472 3473 default: 3474 /* Requests in other states are either un-abortable (e.g. 3475 * TRANSFERRING_CONTROLLER_TO_HOST) or should never end up here, as they're 3476 * immediately transitioned to other states in nvmf_tcp_req_process() (e.g. 3477 * READY_TO_EXECUTE). But it is fine to end up here, as we'll simply complete the 3478 * abort request with the bit0 of dword0 set (command not aborted). 3479 */ 3480 break; 3481 } 3482 3483 spdk_nvmf_request_complete(req); 3484 return SPDK_POLLER_BUSY; 3485 } 3486 3487 static void 3488 nvmf_tcp_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 3489 struct spdk_nvmf_request *req) 3490 { 3491 struct spdk_nvmf_tcp_qpair *tqpair; 3492 struct spdk_nvmf_tcp_transport *ttransport; 3493 struct spdk_nvmf_transport *transport; 3494 uint16_t cid; 3495 uint32_t i; 3496 struct spdk_nvmf_tcp_req *tcp_req_to_abort = NULL; 3497 3498 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3499 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3500 transport = &ttransport->transport; 3501 3502 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 3503 3504 for (i = 0; i < tqpair->resource_count; i++) { 3505 if (tqpair->reqs[i].state != TCP_REQUEST_STATE_FREE && 3506 tqpair->reqs[i].req.cmd->nvme_cmd.cid == cid) { 3507 tcp_req_to_abort = &tqpair->reqs[i]; 3508 break; 3509 } 3510 } 3511 3512 spdk_trace_record(TRACE_TCP_QP_ABORT_REQ, qpair->qid, 0, (uintptr_t)req, tqpair); 3513 3514 if (tcp_req_to_abort == NULL) { 3515 spdk_nvmf_request_complete(req); 3516 return; 3517 } 3518 3519 req->req_to_abort = &tcp_req_to_abort->req; 3520 req->timeout_tsc = spdk_get_ticks() + 3521 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 3522 req->poller = NULL; 3523 3524 _nvmf_tcp_qpair_abort_request(req); 3525 } 3526 3527 struct tcp_subsystem_add_host_opts { 3528 char *psk; 3529 }; 3530 3531 static const struct spdk_json_object_decoder tcp_subsystem_add_host_opts_decoder[] = { 3532 {"psk", offsetof(struct tcp_subsystem_add_host_opts, psk), spdk_json_decode_string, true}, 3533 }; 3534 3535 static int 3536 tcp_load_psk(const char *fname, char *buf, size_t bufsz) 3537 { 3538 FILE *psk_file; 3539 struct stat statbuf; 3540 int rc; 3541 3542 if (stat(fname, &statbuf) != 0) { 3543 SPDK_ERRLOG("Could not read permissions for PSK file\n"); 3544 return -EACCES; 3545 } 3546 3547 if ((statbuf.st_mode & TCP_PSK_INVALID_PERMISSIONS) != 0) { 3548 SPDK_ERRLOG("Incorrect permissions for PSK file\n"); 3549 return -EPERM; 3550 } 3551 if ((size_t)statbuf.st_size >= bufsz) { 3552 SPDK_ERRLOG("Invalid PSK: too long\n"); 3553 return -EINVAL; 3554 } 3555 psk_file = fopen(fname, "r"); 3556 if (psk_file == NULL) { 3557 SPDK_ERRLOG("Could not open PSK file\n"); 3558 return -EINVAL; 3559 } 3560 3561 memset(buf, 0, bufsz); 3562 rc = fread(buf, 1, statbuf.st_size, psk_file); 3563 if (rc != statbuf.st_size) { 3564 SPDK_ERRLOG("Failed to read PSK\n"); 3565 fclose(psk_file); 3566 return -EINVAL; 3567 } 3568 3569 fclose(psk_file); 3570 return 0; 3571 } 3572 3573 static int 3574 nvmf_tcp_subsystem_add_host(struct spdk_nvmf_transport *transport, 3575 const struct spdk_nvmf_subsystem *subsystem, 3576 const char *hostnqn, 3577 const struct spdk_json_val *transport_specific) 3578 { 3579 struct tcp_subsystem_add_host_opts opts; 3580 struct spdk_nvmf_tcp_transport *ttransport; 3581 struct tcp_psk_entry *entry; 3582 char psk_identity[NVMF_PSK_IDENTITY_LEN]; 3583 uint8_t psk_configured[SPDK_TLS_PSK_MAX_LEN] = {}; 3584 char psk_interchange[SPDK_TLS_PSK_MAX_LEN] = {}; 3585 uint8_t tls_cipher_suite; 3586 int rc = 0; 3587 uint8_t psk_retained_hash; 3588 uint64_t psk_configured_size; 3589 3590 if (transport_specific == NULL) { 3591 return 0; 3592 } 3593 3594 assert(transport != NULL); 3595 assert(subsystem != NULL); 3596 3597 memset(&opts, 0, sizeof(opts)); 3598 3599 /* Decode PSK file path */ 3600 if (spdk_json_decode_object_relaxed(transport_specific, tcp_subsystem_add_host_opts_decoder, 3601 SPDK_COUNTOF(tcp_subsystem_add_host_opts_decoder), &opts)) { 3602 SPDK_ERRLOG("spdk_json_decode_object failed\n"); 3603 return -EINVAL; 3604 } 3605 3606 if (opts.psk == NULL) { 3607 return 0; 3608 } 3609 if (strlen(opts.psk) >= sizeof(entry->psk)) { 3610 SPDK_ERRLOG("PSK path too long\n"); 3611 rc = -EINVAL; 3612 goto end; 3613 } 3614 3615 rc = tcp_load_psk(opts.psk, psk_interchange, sizeof(psk_interchange)); 3616 if (rc) { 3617 SPDK_ERRLOG("Could not retrieve PSK from file\n"); 3618 goto end; 3619 } 3620 3621 /* Parse PSK interchange to get length of base64 encoded data. 3622 * This is then used to decide which cipher suite should be used 3623 * to generate PSK identity and TLS PSK later on. */ 3624 rc = nvme_tcp_parse_interchange_psk(psk_interchange, psk_configured, sizeof(psk_configured), 3625 &psk_configured_size, &psk_retained_hash); 3626 if (rc < 0) { 3627 SPDK_ERRLOG("Failed to parse PSK interchange!\n"); 3628 goto end; 3629 } 3630 3631 /* The Base64 string encodes the configured PSK (32 or 48 bytes binary). 3632 * This check also ensures that psk_configured_size is smaller than 3633 * psk_retained buffer size. */ 3634 if (psk_configured_size == SHA256_DIGEST_LENGTH) { 3635 tls_cipher_suite = NVME_TCP_CIPHER_AES_128_GCM_SHA256; 3636 } else if (psk_configured_size == SHA384_DIGEST_LENGTH) { 3637 tls_cipher_suite = NVME_TCP_CIPHER_AES_256_GCM_SHA384; 3638 } else { 3639 SPDK_ERRLOG("Unrecognized cipher suite!\n"); 3640 rc = -EINVAL; 3641 goto end; 3642 } 3643 3644 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3645 /* Generate PSK identity. */ 3646 rc = nvme_tcp_generate_psk_identity(psk_identity, NVMF_PSK_IDENTITY_LEN, hostnqn, 3647 subsystem->subnqn, tls_cipher_suite); 3648 if (rc) { 3649 rc = -EINVAL; 3650 goto end; 3651 } 3652 /* Check if PSK identity entry already exists. */ 3653 TAILQ_FOREACH(entry, &ttransport->psks, link) { 3654 if (strncmp(entry->psk_identity, psk_identity, NVMF_PSK_IDENTITY_LEN) == 0) { 3655 SPDK_ERRLOG("Given PSK identity: %s entry already exists!\n", psk_identity); 3656 rc = -EEXIST; 3657 goto end; 3658 } 3659 } 3660 entry = calloc(1, sizeof(struct tcp_psk_entry)); 3661 if (entry == NULL) { 3662 SPDK_ERRLOG("Unable to allocate memory for PSK entry!\n"); 3663 rc = -ENOMEM; 3664 goto end; 3665 } 3666 3667 if (snprintf(entry->hostnqn, sizeof(entry->hostnqn), "%s", hostnqn) < 0) { 3668 SPDK_ERRLOG("Could not write hostnqn string!\n"); 3669 rc = -EINVAL; 3670 free(entry); 3671 goto end; 3672 } 3673 if (snprintf(entry->subnqn, sizeof(entry->subnqn), "%s", subsystem->subnqn) < 0) { 3674 SPDK_ERRLOG("Could not write subnqn string!\n"); 3675 rc = -EINVAL; 3676 free(entry); 3677 goto end; 3678 } 3679 if (snprintf(entry->psk_identity, sizeof(entry->psk_identity), "%s", psk_identity) < 0) { 3680 SPDK_ERRLOG("Could not write PSK identity string!\n"); 3681 rc = -EINVAL; 3682 free(entry); 3683 goto end; 3684 } 3685 entry->tls_cipher_suite = tls_cipher_suite; 3686 3687 /* No hash indicates that Configured PSK must be used as Retained PSK. */ 3688 if (psk_retained_hash == NVME_TCP_HASH_ALGORITHM_NONE) { 3689 /* Psk configured is either 32 or 48 bytes long. */ 3690 memcpy(entry->psk, psk_configured, psk_configured_size); 3691 entry->psk_size = psk_configured_size; 3692 } else { 3693 /* Derive retained PSK. */ 3694 rc = nvme_tcp_derive_retained_psk(psk_configured, psk_configured_size, hostnqn, entry->psk, 3695 SPDK_TLS_PSK_MAX_LEN, psk_retained_hash); 3696 if (rc < 0) { 3697 SPDK_ERRLOG("Unable to derive retained PSK!\n"); 3698 free(entry); 3699 goto end; 3700 } 3701 entry->psk_size = rc; 3702 } 3703 3704 rc = snprintf(entry->psk_path, sizeof(entry->psk_path), "%s", opts.psk); 3705 if (rc < 0 || (size_t)rc >= sizeof(entry->psk_path)) { 3706 SPDK_ERRLOG("Could not save PSK path!\n"); 3707 rc = -ENAMETOOLONG; 3708 free(entry); 3709 goto end; 3710 } 3711 3712 TAILQ_INSERT_TAIL(&ttransport->psks, entry, link); 3713 rc = 0; 3714 3715 end: 3716 spdk_memset_s(psk_configured, sizeof(psk_configured), 0, sizeof(psk_configured)); 3717 spdk_memset_s(psk_interchange, sizeof(psk_interchange), 0, sizeof(psk_interchange)); 3718 3719 free(opts.psk); 3720 3721 return rc; 3722 } 3723 3724 static void 3725 nvmf_tcp_subsystem_remove_host(struct spdk_nvmf_transport *transport, 3726 const struct spdk_nvmf_subsystem *subsystem, 3727 const char *hostnqn) 3728 { 3729 struct spdk_nvmf_tcp_transport *ttransport; 3730 struct tcp_psk_entry *entry, *tmp; 3731 3732 assert(transport != NULL); 3733 assert(subsystem != NULL); 3734 3735 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3736 TAILQ_FOREACH_SAFE(entry, &ttransport->psks, link, tmp) { 3737 if ((strncmp(entry->hostnqn, hostnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0 && 3738 (strncmp(entry->subnqn, subsystem->subnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0) { 3739 TAILQ_REMOVE(&ttransport->psks, entry, link); 3740 spdk_memset_s(entry->psk, sizeof(entry->psk), 0, sizeof(entry->psk)); 3741 free(entry); 3742 break; 3743 } 3744 } 3745 } 3746 3747 static void 3748 nvmf_tcp_subsystem_dump_host(struct spdk_nvmf_transport *transport, 3749 const struct spdk_nvmf_subsystem *subsystem, const char *hostnqn, 3750 struct spdk_json_write_ctx *w) 3751 { 3752 struct spdk_nvmf_tcp_transport *ttransport; 3753 struct tcp_psk_entry *entry; 3754 3755 assert(transport != NULL); 3756 assert(subsystem != NULL); 3757 3758 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3759 TAILQ_FOREACH(entry, &ttransport->psks, link) { 3760 if ((strncmp(entry->hostnqn, hostnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0 && 3761 (strncmp(entry->subnqn, subsystem->subnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0) { 3762 spdk_json_write_named_string(w, "psk", entry->psk_path); 3763 break; 3764 } 3765 } 3766 } 3767 3768 static void 3769 nvmf_tcp_opts_init(struct spdk_nvmf_transport_opts *opts) 3770 { 3771 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH; 3772 opts->max_qpairs_per_ctrlr = SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR; 3773 opts->in_capsule_data_size = SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE; 3774 opts->max_io_size = SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE; 3775 opts->io_unit_size = SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE; 3776 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH; 3777 opts->num_shared_buffers = SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS; 3778 opts->buf_cache_size = SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE; 3779 opts->dif_insert_or_strip = SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP; 3780 opts->abort_timeout_sec = SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC; 3781 opts->transport_specific = NULL; 3782 } 3783 3784 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp = { 3785 .name = "TCP", 3786 .type = SPDK_NVME_TRANSPORT_TCP, 3787 .opts_init = nvmf_tcp_opts_init, 3788 .create = nvmf_tcp_create, 3789 .dump_opts = nvmf_tcp_dump_opts, 3790 .destroy = nvmf_tcp_destroy, 3791 3792 .listen = nvmf_tcp_listen, 3793 .stop_listen = nvmf_tcp_stop_listen, 3794 3795 .listener_discover = nvmf_tcp_discover, 3796 3797 .poll_group_create = nvmf_tcp_poll_group_create, 3798 .get_optimal_poll_group = nvmf_tcp_get_optimal_poll_group, 3799 .poll_group_destroy = nvmf_tcp_poll_group_destroy, 3800 .poll_group_add = nvmf_tcp_poll_group_add, 3801 .poll_group_remove = nvmf_tcp_poll_group_remove, 3802 .poll_group_poll = nvmf_tcp_poll_group_poll, 3803 3804 .req_free = nvmf_tcp_req_free, 3805 .req_complete = nvmf_tcp_req_complete, 3806 3807 .qpair_fini = nvmf_tcp_close_qpair, 3808 .qpair_get_local_trid = nvmf_tcp_qpair_get_local_trid, 3809 .qpair_get_peer_trid = nvmf_tcp_qpair_get_peer_trid, 3810 .qpair_get_listen_trid = nvmf_tcp_qpair_get_listen_trid, 3811 .qpair_abort_request = nvmf_tcp_qpair_abort_request, 3812 .subsystem_add_host = nvmf_tcp_subsystem_add_host, 3813 .subsystem_remove_host = nvmf_tcp_subsystem_remove_host, 3814 .subsystem_dump_host = nvmf_tcp_subsystem_dump_host, 3815 }; 3816 3817 SPDK_NVMF_TRANSPORT_REGISTER(tcp, &spdk_nvmf_transport_tcp); 3818 SPDK_LOG_REGISTER_COMPONENT(nvmf_tcp) 3819