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