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