1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (C) 2018 Intel Corporation. All rights reserved. 3 * Copyright (c) 2019, 2020 Mellanox Technologies LTD. All rights reserved. 4 * Copyright (c) 2022, 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. 5 */ 6 7 #include "spdk/accel.h" 8 #include "spdk/stdinc.h" 9 #include "spdk/crc32.h" 10 #include "spdk/endian.h" 11 #include "spdk/assert.h" 12 #include "spdk/thread.h" 13 #include "spdk/nvmf_transport.h" 14 #include "spdk/string.h" 15 #include "spdk/trace.h" 16 #include "spdk/util.h" 17 #include "spdk/log.h" 18 19 #include "spdk_internal/assert.h" 20 #include "spdk_internal/nvme_tcp.h" 21 #include "spdk_internal/sock.h" 22 23 #include "nvmf_internal.h" 24 25 #include "spdk_internal/trace_defs.h" 26 27 #define NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME 16 28 #define SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY 16 29 #define SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY 0 30 #define SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM 32 31 #define SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION true 32 33 #define SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH 2 34 #define SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH 65535 35 #define SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH 2 36 #define SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH 4096 37 38 #define SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH 128 39 #define SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH 128 40 #define SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR 128 41 #define SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE 4096 42 #define SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE 131072 43 #define SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE 131072 44 #define SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS 511 45 #define SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX 46 #define SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP false 47 #define SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC 1 48 49 #define TCP_PSK_INVALID_PERMISSIONS 0177 50 51 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp; 52 static bool g_tls_log = false; 53 54 /* spdk nvmf related structure */ 55 enum spdk_nvmf_tcp_req_state { 56 57 /* The request is not currently in use */ 58 TCP_REQUEST_STATE_FREE = 0, 59 60 /* Initial state when request first received */ 61 TCP_REQUEST_STATE_NEW = 1, 62 63 /* The request is queued until a data buffer is available. */ 64 TCP_REQUEST_STATE_NEED_BUFFER = 2, 65 66 /* The request is waiting for zcopy_start to finish */ 67 TCP_REQUEST_STATE_AWAITING_ZCOPY_START = 3, 68 69 /* The request has received a zero-copy buffer */ 70 TCP_REQUEST_STATE_ZCOPY_START_COMPLETED = 4, 71 72 /* The request is currently transferring data from the host to the controller. */ 73 TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER = 5, 74 75 /* The request is waiting for the R2T send acknowledgement. */ 76 TCP_REQUEST_STATE_AWAITING_R2T_ACK = 6, 77 78 /* The request is ready to execute at the block device */ 79 TCP_REQUEST_STATE_READY_TO_EXECUTE = 7, 80 81 /* The request is currently executing at the block device */ 82 TCP_REQUEST_STATE_EXECUTING = 8, 83 84 /* The request is waiting for zcopy buffers to be committed */ 85 TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT = 9, 86 87 /* The request finished executing at the block device */ 88 TCP_REQUEST_STATE_EXECUTED = 10, 89 90 /* The request is ready to send a completion */ 91 TCP_REQUEST_STATE_READY_TO_COMPLETE = 11, 92 93 /* The request is currently transferring final pdus from the controller to the host. */ 94 TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST = 12, 95 96 /* The request is waiting for zcopy buffers to be released (without committing) */ 97 TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE = 13, 98 99 /* The request completed and can be marked free. */ 100 TCP_REQUEST_STATE_COMPLETED = 14, 101 102 /* Terminator */ 103 TCP_REQUEST_NUM_STATES, 104 }; 105 106 static const char *spdk_nvmf_tcp_term_req_fes_str[] = { 107 "Invalid PDU Header Field", 108 "PDU Sequence Error", 109 "Header Digiest Error", 110 "Data Transfer Out of Range", 111 "R2T Limit Exceeded", 112 "Unsupported parameter", 113 }; 114 115 SPDK_TRACE_REGISTER_FN(nvmf_tcp_trace, "nvmf_tcp", TRACE_GROUP_NVMF_TCP) 116 { 117 spdk_trace_register_owner(OWNER_NVMF_TCP, 't'); 118 spdk_trace_register_object(OBJECT_NVMF_TCP_IO, 'r'); 119 spdk_trace_register_description("TCP_REQ_NEW", 120 TRACE_TCP_REQUEST_STATE_NEW, 121 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 1, 122 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 123 spdk_trace_register_description("TCP_REQ_NEED_BUFFER", 124 TRACE_TCP_REQUEST_STATE_NEED_BUFFER, 125 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 126 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 127 spdk_trace_register_description("TCP_REQ_WAIT_ZCPY_START", 128 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, 129 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 130 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 131 spdk_trace_register_description("TCP_REQ_ZCPY_START_CPL", 132 TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, 133 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 134 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 135 spdk_trace_register_description("TCP_REQ_TX_H_TO_C", 136 TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 137 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 138 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 139 spdk_trace_register_description("TCP_REQ_RDY_TO_EXECUTE", 140 TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, 141 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 142 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 143 spdk_trace_register_description("TCP_REQ_EXECUTING", 144 TRACE_TCP_REQUEST_STATE_EXECUTING, 145 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 146 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 147 spdk_trace_register_description("TCP_REQ_WAIT_ZCPY_CMT", 148 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, 149 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 150 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 151 spdk_trace_register_description("TCP_REQ_EXECUTED", 152 TRACE_TCP_REQUEST_STATE_EXECUTED, 153 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 154 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 155 spdk_trace_register_description("TCP_REQ_RDY_TO_COMPLETE", 156 TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, 157 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 158 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 159 spdk_trace_register_description("TCP_REQ_TRANSFER_C2H", 160 TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 161 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 162 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 163 spdk_trace_register_description("TCP_REQ_AWAIT_ZCPY_RLS", 164 TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, 165 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 166 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 167 spdk_trace_register_description("TCP_REQ_COMPLETED", 168 TRACE_TCP_REQUEST_STATE_COMPLETED, 169 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 170 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 171 spdk_trace_register_description("TCP_WRITE_START", 172 TRACE_TCP_FLUSH_WRITEBUF_START, 173 OWNER_NVMF_TCP, OBJECT_NONE, 0, 174 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 175 spdk_trace_register_description("TCP_WRITE_DONE", 176 TRACE_TCP_FLUSH_WRITEBUF_DONE, 177 OWNER_NVMF_TCP, OBJECT_NONE, 0, 178 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 179 spdk_trace_register_description("TCP_READ_DONE", 180 TRACE_TCP_READ_FROM_SOCKET_DONE, 181 OWNER_NVMF_TCP, OBJECT_NONE, 0, 182 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 183 spdk_trace_register_description("TCP_REQ_AWAIT_R2T_ACK", 184 TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, 185 OWNER_NVMF_TCP, OBJECT_NVMF_TCP_IO, 0, 186 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 187 188 spdk_trace_register_description("TCP_QP_CREATE", TRACE_TCP_QP_CREATE, 189 OWNER_NVMF_TCP, OBJECT_NONE, 0, 190 SPDK_TRACE_ARG_TYPE_INT, ""); 191 spdk_trace_register_description("TCP_QP_SOCK_INIT", TRACE_TCP_QP_SOCK_INIT, 192 OWNER_NVMF_TCP, OBJECT_NONE, 0, 193 SPDK_TRACE_ARG_TYPE_INT, ""); 194 spdk_trace_register_description("TCP_QP_STATE_CHANGE", TRACE_TCP_QP_STATE_CHANGE, 195 OWNER_NVMF_TCP, OBJECT_NONE, 0, 196 SPDK_TRACE_ARG_TYPE_INT, "state"); 197 spdk_trace_register_description("TCP_QP_DISCONNECT", TRACE_TCP_QP_DISCONNECT, 198 OWNER_NVMF_TCP, OBJECT_NONE, 0, 199 SPDK_TRACE_ARG_TYPE_INT, ""); 200 spdk_trace_register_description("TCP_QP_DESTROY", TRACE_TCP_QP_DESTROY, 201 OWNER_NVMF_TCP, OBJECT_NONE, 0, 202 SPDK_TRACE_ARG_TYPE_INT, ""); 203 spdk_trace_register_description("TCP_QP_ABORT_REQ", TRACE_TCP_QP_ABORT_REQ, 204 OWNER_NVMF_TCP, OBJECT_NONE, 0, 205 SPDK_TRACE_ARG_TYPE_PTR, "qpair"); 206 spdk_trace_register_description("TCP_QP_RCV_STATE_CHANGE", TRACE_TCP_QP_RCV_STATE_CHANGE, 207 OWNER_NVMF_TCP, OBJECT_NONE, 0, 208 SPDK_TRACE_ARG_TYPE_INT, "state"); 209 210 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_TCP_IO, 1); 211 spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_TCP_IO, 0); 212 } 213 214 struct spdk_nvmf_tcp_req { 215 struct spdk_nvmf_request req; 216 struct spdk_nvme_cpl rsp; 217 struct spdk_nvme_cmd cmd; 218 219 /* A PDU that can be used for sending responses. This is 220 * not the incoming PDU! */ 221 struct nvme_tcp_pdu *pdu; 222 223 /* In-capsule data buffer */ 224 uint8_t *buf; 225 226 struct spdk_nvmf_tcp_req *fused_pair; 227 228 /* 229 * The PDU for a request may be used multiple times in serial over 230 * the request's lifetime. For example, first to send an R2T, then 231 * to send a completion. To catch mistakes where the PDU is used 232 * twice at the same time, add a debug flag here for init/fini. 233 */ 234 bool pdu_in_use; 235 bool has_in_capsule_data; 236 bool fused_failed; 237 238 /* transfer_tag */ 239 uint16_t ttag; 240 241 enum spdk_nvmf_tcp_req_state state; 242 243 /* 244 * h2c_offset is used when we receive the h2c_data PDU. 245 */ 246 uint32_t h2c_offset; 247 248 STAILQ_ENTRY(spdk_nvmf_tcp_req) link; 249 TAILQ_ENTRY(spdk_nvmf_tcp_req) state_link; 250 }; 251 252 struct spdk_nvmf_tcp_qpair { 253 struct spdk_nvmf_qpair qpair; 254 struct spdk_nvmf_tcp_poll_group *group; 255 struct spdk_sock *sock; 256 257 enum nvme_tcp_pdu_recv_state recv_state; 258 enum nvme_tcp_qpair_state state; 259 260 /* PDU being actively received */ 261 struct nvme_tcp_pdu *pdu_in_progress; 262 263 struct spdk_nvmf_tcp_req *fused_first; 264 265 /* Queues to track the requests in all states */ 266 TAILQ_HEAD(, spdk_nvmf_tcp_req) tcp_req_working_queue; 267 TAILQ_HEAD(, spdk_nvmf_tcp_req) tcp_req_free_queue; 268 SLIST_HEAD(, nvme_tcp_pdu) tcp_pdu_free_queue; 269 /* Number of working pdus */ 270 uint32_t tcp_pdu_working_count; 271 272 /* Number of requests in each state */ 273 uint32_t state_cntr[TCP_REQUEST_NUM_STATES]; 274 275 uint8_t cpda; 276 277 bool host_hdgst_enable; 278 bool host_ddgst_enable; 279 280 /* This is a spare PDU used for sending special management 281 * operations. Primarily, this is used for the initial 282 * connection response and c2h termination request. */ 283 struct nvme_tcp_pdu *mgmt_pdu; 284 285 /* Arrays of in-capsule buffers, requests, and pdus. 286 * Each array is 'resource_count' number of elements */ 287 void *bufs; 288 struct spdk_nvmf_tcp_req *reqs; 289 struct nvme_tcp_pdu *pdus; 290 uint32_t resource_count; 291 uint32_t recv_buf_size; 292 293 struct spdk_nvmf_tcp_port *port; 294 295 /* IP address */ 296 char initiator_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 297 char target_addr[SPDK_NVMF_TRADDR_MAX_LEN]; 298 299 /* IP port */ 300 uint16_t initiator_port; 301 uint16_t target_port; 302 303 /* Wait until the host terminates the connection (e.g. after sending C2HTermReq) */ 304 bool wait_terminate; 305 306 /* Timer used to destroy qpair after detecting transport error issue if initiator does 307 * not close the connection. 308 */ 309 struct spdk_poller *timeout_poller; 310 311 spdk_nvmf_transport_qpair_fini_cb fini_cb_fn; 312 void *fini_cb_arg; 313 314 TAILQ_ENTRY(spdk_nvmf_tcp_qpair) link; 315 }; 316 317 struct spdk_nvmf_tcp_control_msg { 318 STAILQ_ENTRY(spdk_nvmf_tcp_control_msg) link; 319 }; 320 321 struct spdk_nvmf_tcp_control_msg_list { 322 void *msg_buf; 323 STAILQ_HEAD(, spdk_nvmf_tcp_control_msg) free_msgs; 324 }; 325 326 struct spdk_nvmf_tcp_poll_group { 327 struct spdk_nvmf_transport_poll_group group; 328 struct spdk_sock_group *sock_group; 329 330 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) qpairs; 331 TAILQ_HEAD(, spdk_nvmf_tcp_qpair) await_req; 332 333 struct spdk_io_channel *accel_channel; 334 struct spdk_nvmf_tcp_control_msg_list *control_msg_list; 335 336 TAILQ_ENTRY(spdk_nvmf_tcp_poll_group) link; 337 }; 338 339 struct spdk_nvmf_tcp_port { 340 const struct spdk_nvme_transport_id *trid; 341 struct spdk_sock *listen_sock; 342 TAILQ_ENTRY(spdk_nvmf_tcp_port) link; 343 }; 344 345 struct tcp_transport_opts { 346 bool c2h_success; 347 uint16_t control_msg_num; 348 uint32_t sock_priority; 349 }; 350 351 struct tcp_psk_entry { 352 char hostnqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 353 char subnqn[SPDK_NVMF_NQN_MAX_LEN + 1]; 354 char pskid[NVMF_PSK_IDENTITY_LEN]; 355 uint8_t psk[SPDK_TLS_PSK_MAX_LEN]; 356 357 /* Original path saved to emit SPDK configuration via "save_config". */ 358 char psk_path[PATH_MAX]; 359 uint32_t psk_size; 360 enum nvme_tcp_cipher_suite tls_cipher_suite; 361 TAILQ_ENTRY(tcp_psk_entry) link; 362 }; 363 364 struct spdk_nvmf_tcp_transport { 365 struct spdk_nvmf_transport transport; 366 struct tcp_transport_opts tcp_opts; 367 368 struct spdk_nvmf_tcp_poll_group *next_pg; 369 370 struct spdk_poller *accept_poller; 371 372 TAILQ_HEAD(, spdk_nvmf_tcp_port) ports; 373 TAILQ_HEAD(, spdk_nvmf_tcp_poll_group) poll_groups; 374 375 TAILQ_HEAD(, tcp_psk_entry) psks; 376 }; 377 378 static const struct spdk_json_object_decoder tcp_transport_opts_decoder[] = { 379 { 380 "c2h_success", offsetof(struct tcp_transport_opts, c2h_success), 381 spdk_json_decode_bool, true 382 }, 383 { 384 "control_msg_num", offsetof(struct tcp_transport_opts, control_msg_num), 385 spdk_json_decode_uint16, true 386 }, 387 { 388 "sock_priority", offsetof(struct tcp_transport_opts, sock_priority), 389 spdk_json_decode_uint32, true 390 }, 391 }; 392 393 static bool nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 394 struct spdk_nvmf_tcp_req *tcp_req); 395 static void nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group); 396 397 static void _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 398 struct spdk_nvmf_tcp_req *tcp_req); 399 400 static inline void 401 nvmf_tcp_req_set_state(struct spdk_nvmf_tcp_req *tcp_req, 402 enum spdk_nvmf_tcp_req_state state) 403 { 404 struct spdk_nvmf_qpair *qpair; 405 struct spdk_nvmf_tcp_qpair *tqpair; 406 407 qpair = tcp_req->req.qpair; 408 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 409 410 assert(tqpair->state_cntr[tcp_req->state] > 0); 411 tqpair->state_cntr[tcp_req->state]--; 412 tqpair->state_cntr[state]++; 413 414 tcp_req->state = state; 415 } 416 417 static inline struct nvme_tcp_pdu * 418 nvmf_tcp_req_pdu_init(struct spdk_nvmf_tcp_req *tcp_req) 419 { 420 assert(tcp_req->pdu_in_use == false); 421 422 memset(tcp_req->pdu, 0, sizeof(*tcp_req->pdu)); 423 tcp_req->pdu->qpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 424 425 return tcp_req->pdu; 426 } 427 428 static struct spdk_nvmf_tcp_req * 429 nvmf_tcp_req_get(struct spdk_nvmf_tcp_qpair *tqpair) 430 { 431 struct spdk_nvmf_tcp_req *tcp_req; 432 433 tcp_req = TAILQ_FIRST(&tqpair->tcp_req_free_queue); 434 if (spdk_unlikely(!tcp_req)) { 435 return NULL; 436 } 437 438 memset(&tcp_req->rsp, 0, sizeof(tcp_req->rsp)); 439 tcp_req->h2c_offset = 0; 440 tcp_req->has_in_capsule_data = false; 441 tcp_req->req.dif_enabled = false; 442 tcp_req->req.zcopy_phase = NVMF_ZCOPY_PHASE_NONE; 443 444 TAILQ_REMOVE(&tqpair->tcp_req_free_queue, tcp_req, state_link); 445 TAILQ_INSERT_TAIL(&tqpair->tcp_req_working_queue, tcp_req, state_link); 446 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEW); 447 return tcp_req; 448 } 449 450 static inline void 451 nvmf_tcp_req_put(struct spdk_nvmf_tcp_qpair *tqpair, struct spdk_nvmf_tcp_req *tcp_req) 452 { 453 assert(!tcp_req->pdu_in_use); 454 455 TAILQ_REMOVE(&tqpair->tcp_req_working_queue, tcp_req, state_link); 456 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 457 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_FREE); 458 } 459 460 static void 461 nvmf_tcp_request_free(void *cb_arg) 462 { 463 struct spdk_nvmf_tcp_transport *ttransport; 464 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 465 466 assert(tcp_req != NULL); 467 468 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req=%p will be freed\n", tcp_req); 469 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 470 struct spdk_nvmf_tcp_transport, transport); 471 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 472 nvmf_tcp_req_process(ttransport, tcp_req); 473 } 474 475 static int 476 nvmf_tcp_req_free(struct spdk_nvmf_request *req) 477 { 478 struct spdk_nvmf_tcp_req *tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 479 480 nvmf_tcp_request_free(tcp_req); 481 482 return 0; 483 } 484 485 static void 486 nvmf_tcp_drain_state_queue(struct spdk_nvmf_tcp_qpair *tqpair, 487 enum spdk_nvmf_tcp_req_state state) 488 { 489 struct spdk_nvmf_tcp_req *tcp_req, *req_tmp; 490 491 assert(state != TCP_REQUEST_STATE_FREE); 492 TAILQ_FOREACH_SAFE(tcp_req, &tqpair->tcp_req_working_queue, state_link, req_tmp) { 493 if (state == tcp_req->state) { 494 nvmf_tcp_request_free(tcp_req); 495 } 496 } 497 } 498 499 static void 500 nvmf_tcp_cleanup_all_states(struct spdk_nvmf_tcp_qpair *tqpair) 501 { 502 struct spdk_nvmf_tcp_req *tcp_req, *req_tmp; 503 504 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 505 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_NEW); 506 507 /* Wipe the requests waiting for buffer from the global list */ 508 TAILQ_FOREACH_SAFE(tcp_req, &tqpair->tcp_req_working_queue, state_link, req_tmp) { 509 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 510 STAILQ_REMOVE(&tqpair->group->group.pending_buf_queue, &tcp_req->req, 511 spdk_nvmf_request, buf_link); 512 } 513 } 514 515 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_NEED_BUFFER); 516 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_EXECUTING); 517 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 518 nvmf_tcp_drain_state_queue(tqpair, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 519 } 520 521 static void 522 nvmf_tcp_dump_qpair_req_contents(struct spdk_nvmf_tcp_qpair *tqpair) 523 { 524 int i; 525 struct spdk_nvmf_tcp_req *tcp_req; 526 527 SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", tqpair->qpair.qid); 528 for (i = 1; i < TCP_REQUEST_NUM_STATES; i++) { 529 SPDK_ERRLOG("\tNum of requests in state[%d] = %u\n", i, tqpair->state_cntr[i]); 530 TAILQ_FOREACH(tcp_req, &tqpair->tcp_req_working_queue, state_link) { 531 if ((int)tcp_req->state == i) { 532 SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", tcp_req->req.data_from_pool); 533 SPDK_ERRLOG("\t\tRequest opcode: %d\n", tcp_req->req.cmd->nvmf_cmd.opcode); 534 } 535 } 536 } 537 } 538 539 static void 540 _nvmf_tcp_qpair_destroy(void *_tqpair) 541 { 542 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 543 spdk_nvmf_transport_qpair_fini_cb cb_fn = tqpair->fini_cb_fn; 544 void *cb_arg = tqpair->fini_cb_arg; 545 int err = 0; 546 547 spdk_trace_record(TRACE_TCP_QP_DESTROY, 0, 0, (uintptr_t)tqpair); 548 549 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 550 551 err = spdk_sock_close(&tqpair->sock); 552 assert(err == 0); 553 nvmf_tcp_cleanup_all_states(tqpair); 554 555 if (tqpair->state_cntr[TCP_REQUEST_STATE_FREE] != tqpair->resource_count) { 556 SPDK_ERRLOG("tqpair(%p) free tcp request num is %u but should be %u\n", tqpair, 557 tqpair->state_cntr[TCP_REQUEST_STATE_FREE], 558 tqpair->resource_count); 559 err++; 560 } 561 562 if (err > 0) { 563 nvmf_tcp_dump_qpair_req_contents(tqpair); 564 } 565 566 /* The timeout poller might still be registered here if we close the qpair before host 567 * terminates the connection. 568 */ 569 spdk_poller_unregister(&tqpair->timeout_poller); 570 spdk_dma_free(tqpair->pdus); 571 free(tqpair->reqs); 572 spdk_free(tqpair->bufs); 573 free(tqpair); 574 575 if (cb_fn != NULL) { 576 cb_fn(cb_arg); 577 } 578 579 SPDK_DEBUGLOG(nvmf_tcp, "Leave\n"); 580 } 581 582 static void 583 nvmf_tcp_qpair_destroy(struct spdk_nvmf_tcp_qpair *tqpair) 584 { 585 /* Delay the destruction to make sure it isn't performed from the context of a sock 586 * callback. Otherwise, spdk_sock_close() might not abort pending requests, causing their 587 * completions to be executed after the qpair is freed. (Note: this fixed issue #2471.) 588 */ 589 spdk_thread_send_msg(spdk_get_thread(), _nvmf_tcp_qpair_destroy, tqpair); 590 } 591 592 static void 593 nvmf_tcp_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w) 594 { 595 struct spdk_nvmf_tcp_transport *ttransport; 596 assert(w != NULL); 597 598 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 599 spdk_json_write_named_bool(w, "c2h_success", ttransport->tcp_opts.c2h_success); 600 spdk_json_write_named_uint32(w, "sock_priority", ttransport->tcp_opts.sock_priority); 601 } 602 603 static void 604 nvmf_tcp_free_psk_entry(struct tcp_psk_entry *entry) 605 { 606 if (entry == NULL) { 607 return; 608 } 609 610 spdk_memset_s(entry->psk, sizeof(entry->psk), 0, sizeof(entry->psk)); 611 free(entry); 612 } 613 614 static int 615 nvmf_tcp_destroy(struct spdk_nvmf_transport *transport, 616 spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg) 617 { 618 struct spdk_nvmf_tcp_transport *ttransport; 619 struct tcp_psk_entry *entry, *tmp; 620 621 assert(transport != NULL); 622 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 623 624 TAILQ_FOREACH_SAFE(entry, &ttransport->psks, link, tmp) { 625 TAILQ_REMOVE(&ttransport->psks, entry, link); 626 nvmf_tcp_free_psk_entry(entry); 627 } 628 629 spdk_poller_unregister(&ttransport->accept_poller); 630 free(ttransport); 631 632 if (cb_fn) { 633 cb_fn(cb_arg); 634 } 635 return 0; 636 } 637 638 static int nvmf_tcp_accept(void *ctx); 639 640 static struct spdk_nvmf_transport * 641 nvmf_tcp_create(struct spdk_nvmf_transport_opts *opts) 642 { 643 struct spdk_nvmf_tcp_transport *ttransport; 644 uint32_t sge_count; 645 uint32_t min_shared_buffers; 646 647 ttransport = calloc(1, sizeof(*ttransport)); 648 if (!ttransport) { 649 return NULL; 650 } 651 652 TAILQ_INIT(&ttransport->ports); 653 TAILQ_INIT(&ttransport->poll_groups); 654 TAILQ_INIT(&ttransport->psks); 655 656 ttransport->transport.ops = &spdk_nvmf_transport_tcp; 657 658 ttransport->tcp_opts.c2h_success = SPDK_NVMF_TCP_DEFAULT_SUCCESS_OPTIMIZATION; 659 ttransport->tcp_opts.sock_priority = SPDK_NVMF_TCP_DEFAULT_SOCK_PRIORITY; 660 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 661 if (opts->transport_specific != NULL && 662 spdk_json_decode_object_relaxed(opts->transport_specific, tcp_transport_opts_decoder, 663 SPDK_COUNTOF(tcp_transport_opts_decoder), 664 &ttransport->tcp_opts)) { 665 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n"); 666 free(ttransport); 667 return NULL; 668 } 669 670 SPDK_NOTICELOG("*** TCP Transport Init ***\n"); 671 672 SPDK_INFOLOG(nvmf_tcp, "*** TCP Transport Init ***\n" 673 " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n" 674 " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n" 675 " in_capsule_data_size=%d, max_aq_depth=%d\n" 676 " num_shared_buffers=%d, c2h_success=%d,\n" 677 " dif_insert_or_strip=%d, sock_priority=%d\n" 678 " abort_timeout_sec=%d, control_msg_num=%hu\n", 679 opts->max_queue_depth, 680 opts->max_io_size, 681 opts->max_qpairs_per_ctrlr - 1, 682 opts->io_unit_size, 683 opts->in_capsule_data_size, 684 opts->max_aq_depth, 685 opts->num_shared_buffers, 686 ttransport->tcp_opts.c2h_success, 687 opts->dif_insert_or_strip, 688 ttransport->tcp_opts.sock_priority, 689 opts->abort_timeout_sec, 690 ttransport->tcp_opts.control_msg_num); 691 692 if (ttransport->tcp_opts.sock_priority > SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY) { 693 SPDK_ERRLOG("Unsupported socket_priority=%d, the current range is: 0 to %d\n" 694 "you can use man 7 socket to view the range of priority under SO_PRIORITY item\n", 695 ttransport->tcp_opts.sock_priority, SPDK_NVMF_TCP_DEFAULT_MAX_SOCK_PRIORITY); 696 free(ttransport); 697 return NULL; 698 } 699 700 if (ttransport->tcp_opts.control_msg_num == 0 && 701 opts->in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 702 SPDK_WARNLOG("TCP param control_msg_num can't be 0 if ICD is less than %u bytes. Using default value %u\n", 703 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM); 704 ttransport->tcp_opts.control_msg_num = SPDK_NVMF_TCP_DEFAULT_CONTROL_MSG_NUM; 705 } 706 707 /* I/O unit size cannot be larger than max I/O size */ 708 if (opts->io_unit_size > opts->max_io_size) { 709 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", 710 opts->io_unit_size, opts->max_io_size); 711 opts->io_unit_size = opts->max_io_size; 712 } 713 714 /* In capsule data size cannot be larger than max I/O size */ 715 if (opts->in_capsule_data_size > opts->max_io_size) { 716 SPDK_WARNLOG("TCP param ICD size %u can't be larger than max_io_size %u. Using max_io_size as ICD size\n", 717 opts->io_unit_size, opts->max_io_size); 718 opts->in_capsule_data_size = opts->max_io_size; 719 } 720 721 /* max IO queue depth cannot be smaller than 2 or larger than 65535. 722 * We will not check SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH, because max_queue_depth is 16bits and always not larger than 64k. */ 723 if (opts->max_queue_depth < SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH) { 724 SPDK_WARNLOG("TCP param max_queue_depth %u can't be smaller than %u or larger than %u. Using default value %u\n", 725 opts->max_queue_depth, SPDK_NVMF_TCP_MIN_IO_QUEUE_DEPTH, 726 SPDK_NVMF_TCP_MAX_IO_QUEUE_DEPTH, SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH); 727 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH; 728 } 729 730 /* max admin queue depth cannot be smaller than 2 or larger than 4096 */ 731 if (opts->max_aq_depth < SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH || 732 opts->max_aq_depth > SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH) { 733 SPDK_WARNLOG("TCP param max_aq_depth %u can't be smaller than %u or larger than %u. Using default value %u\n", 734 opts->max_aq_depth, SPDK_NVMF_TCP_MIN_ADMIN_QUEUE_DEPTH, 735 SPDK_NVMF_TCP_MAX_ADMIN_QUEUE_DEPTH, SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH); 736 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH; 737 } 738 739 sge_count = opts->max_io_size / opts->io_unit_size; 740 if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) { 741 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size); 742 free(ttransport); 743 return NULL; 744 } 745 746 /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */ 747 if (opts->buf_cache_size < UINT32_MAX) { 748 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size; 749 if (min_shared_buffers > opts->num_shared_buffers) { 750 SPDK_ERRLOG("There are not enough buffers to satisfy " 751 "per-poll group caches for each thread. (%" PRIu32 ") " 752 "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers); 753 SPDK_ERRLOG("Please specify a larger number of shared buffers\n"); 754 free(ttransport); 755 return NULL; 756 } 757 } 758 759 ttransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_tcp_accept, &ttransport->transport, 760 opts->acceptor_poll_rate); 761 if (!ttransport->accept_poller) { 762 free(ttransport); 763 return NULL; 764 } 765 766 return &ttransport->transport; 767 } 768 769 static int 770 nvmf_tcp_trsvcid_to_int(const char *trsvcid) 771 { 772 unsigned long long ull; 773 char *end = NULL; 774 775 ull = strtoull(trsvcid, &end, 10); 776 if (end == NULL || end == trsvcid || *end != '\0') { 777 return -1; 778 } 779 780 /* Valid TCP/IP port numbers are in [1, 65535] */ 781 if (ull == 0 || ull > 65535) { 782 return -1; 783 } 784 785 return (int)ull; 786 } 787 788 /** 789 * Canonicalize a listen address trid. 790 */ 791 static int 792 nvmf_tcp_canon_listen_trid(struct spdk_nvme_transport_id *canon_trid, 793 const struct spdk_nvme_transport_id *trid) 794 { 795 int trsvcid_int; 796 797 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 798 if (trsvcid_int < 0) { 799 return -EINVAL; 800 } 801 802 memset(canon_trid, 0, sizeof(*canon_trid)); 803 spdk_nvme_trid_populate_transport(canon_trid, SPDK_NVME_TRANSPORT_TCP); 804 canon_trid->adrfam = trid->adrfam; 805 snprintf(canon_trid->traddr, sizeof(canon_trid->traddr), "%s", trid->traddr); 806 snprintf(canon_trid->trsvcid, sizeof(canon_trid->trsvcid), "%d", trsvcid_int); 807 808 return 0; 809 } 810 811 /** 812 * Find an existing listening port. 813 */ 814 static struct spdk_nvmf_tcp_port * 815 nvmf_tcp_find_port(struct spdk_nvmf_tcp_transport *ttransport, 816 const struct spdk_nvme_transport_id *trid) 817 { 818 struct spdk_nvme_transport_id canon_trid; 819 struct spdk_nvmf_tcp_port *port; 820 821 if (nvmf_tcp_canon_listen_trid(&canon_trid, trid) != 0) { 822 return NULL; 823 } 824 825 TAILQ_FOREACH(port, &ttransport->ports, link) { 826 if (spdk_nvme_transport_id_compare(&canon_trid, port->trid) == 0) { 827 return port; 828 } 829 } 830 831 return NULL; 832 } 833 834 static int 835 tcp_sock_get_key(uint8_t *out, int out_len, const char **cipher, const char *pskid, 836 void *get_key_ctx) 837 { 838 struct tcp_psk_entry *entry; 839 struct spdk_nvmf_tcp_transport *ttransport = get_key_ctx; 840 size_t psk_len; 841 int rc; 842 843 TAILQ_FOREACH(entry, &ttransport->psks, link) { 844 if (strcmp(pskid, entry->pskid) != 0) { 845 continue; 846 } 847 848 psk_len = entry->psk_size; 849 if ((size_t)out_len < psk_len) { 850 SPDK_ERRLOG("Out buffer of size: %" PRIu32 " cannot fit PSK of len: %lu\n", 851 out_len, psk_len); 852 return -ENOBUFS; 853 } 854 855 /* Convert PSK to the TLS PSK format. */ 856 rc = nvme_tcp_derive_tls_psk(entry->psk, psk_len, pskid, out, out_len, 857 entry->tls_cipher_suite); 858 if (rc < 0) { 859 SPDK_ERRLOG("Could not generate TLS PSK\n"); 860 } 861 862 switch (entry->tls_cipher_suite) { 863 case NVME_TCP_CIPHER_AES_128_GCM_SHA256: 864 *cipher = "TLS_AES_128_GCM_SHA256"; 865 break; 866 case NVME_TCP_CIPHER_AES_256_GCM_SHA384: 867 *cipher = "TLS_AES_256_GCM_SHA384"; 868 break; 869 default: 870 *cipher = NULL; 871 return -ENOTSUP; 872 } 873 874 return rc; 875 } 876 877 SPDK_ERRLOG("Could not find PSK for identity: %s\n", pskid); 878 879 return -EINVAL; 880 } 881 882 static int 883 nvmf_tcp_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid, 884 struct spdk_nvmf_listen_opts *listen_opts) 885 { 886 struct spdk_nvmf_tcp_transport *ttransport; 887 struct spdk_nvmf_tcp_port *port; 888 int trsvcid_int; 889 uint8_t adrfam; 890 const char *sock_impl_name; 891 struct spdk_sock_impl_opts impl_opts; 892 size_t impl_opts_size = sizeof(impl_opts); 893 struct spdk_sock_opts opts; 894 895 if (!strlen(trid->trsvcid)) { 896 SPDK_ERRLOG("Service id is required\n"); 897 return -EINVAL; 898 } 899 900 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 901 902 trsvcid_int = nvmf_tcp_trsvcid_to_int(trid->trsvcid); 903 if (trsvcid_int < 0) { 904 SPDK_ERRLOG("Invalid trsvcid '%s'\n", trid->trsvcid); 905 return -EINVAL; 906 } 907 908 port = calloc(1, sizeof(*port)); 909 if (!port) { 910 SPDK_ERRLOG("Port allocation failed\n"); 911 return -ENOMEM; 912 } 913 914 port->trid = trid; 915 916 sock_impl_name = NULL; 917 918 opts.opts_size = sizeof(opts); 919 spdk_sock_get_default_opts(&opts); 920 opts.priority = ttransport->tcp_opts.sock_priority; 921 if (listen_opts->secure_channel) { 922 if (!g_tls_log) { 923 SPDK_NOTICELOG("TLS support is considered experimental\n"); 924 g_tls_log = true; 925 } 926 sock_impl_name = "ssl"; 927 spdk_sock_impl_get_opts(sock_impl_name, &impl_opts, &impl_opts_size); 928 impl_opts.tls_version = SPDK_TLS_VERSION_1_3; 929 impl_opts.get_key = tcp_sock_get_key; 930 impl_opts.get_key_ctx = ttransport; 931 impl_opts.tls_cipher_suites = "TLS_AES_256_GCM_SHA384:TLS_AES_128_GCM_SHA256"; 932 opts.impl_opts = &impl_opts; 933 opts.impl_opts_size = sizeof(impl_opts); 934 } 935 936 port->listen_sock = spdk_sock_listen_ext(trid->traddr, trsvcid_int, 937 sock_impl_name, &opts); 938 if (port->listen_sock == NULL) { 939 SPDK_ERRLOG("spdk_sock_listen(%s, %d) failed: %s (%d)\n", 940 trid->traddr, trsvcid_int, 941 spdk_strerror(errno), errno); 942 free(port); 943 return -errno; 944 } 945 946 if (spdk_sock_is_ipv4(port->listen_sock)) { 947 adrfam = SPDK_NVMF_ADRFAM_IPV4; 948 } else if (spdk_sock_is_ipv6(port->listen_sock)) { 949 adrfam = SPDK_NVMF_ADRFAM_IPV6; 950 } else { 951 SPDK_ERRLOG("Unhandled socket type\n"); 952 adrfam = 0; 953 } 954 955 if (adrfam != trid->adrfam) { 956 SPDK_ERRLOG("Socket address family mismatch\n"); 957 spdk_sock_close(&port->listen_sock); 958 free(port); 959 return -EINVAL; 960 } 961 962 SPDK_NOTICELOG("*** NVMe/TCP Target Listening on %s port %s ***\n", 963 trid->traddr, trid->trsvcid); 964 965 TAILQ_INSERT_TAIL(&ttransport->ports, port, link); 966 return 0; 967 } 968 969 static void 970 nvmf_tcp_stop_listen(struct spdk_nvmf_transport *transport, 971 const struct spdk_nvme_transport_id *trid) 972 { 973 struct spdk_nvmf_tcp_transport *ttransport; 974 struct spdk_nvmf_tcp_port *port; 975 976 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 977 978 SPDK_DEBUGLOG(nvmf_tcp, "Removing listen address %s port %s\n", 979 trid->traddr, trid->trsvcid); 980 981 port = nvmf_tcp_find_port(ttransport, trid); 982 if (port) { 983 TAILQ_REMOVE(&ttransport->ports, port, link); 984 spdk_sock_close(&port->listen_sock); 985 free(port); 986 } 987 } 988 989 static void nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 990 enum nvme_tcp_pdu_recv_state state); 991 992 static void 993 nvmf_tcp_qpair_set_state(struct spdk_nvmf_tcp_qpair *tqpair, enum nvme_tcp_qpair_state state) 994 { 995 tqpair->state = state; 996 spdk_trace_record(TRACE_TCP_QP_STATE_CHANGE, tqpair->qpair.qid, 0, (uintptr_t)tqpair, 997 tqpair->state); 998 } 999 1000 static void 1001 nvmf_tcp_qpair_disconnect(struct spdk_nvmf_tcp_qpair *tqpair) 1002 { 1003 SPDK_DEBUGLOG(nvmf_tcp, "Disconnecting qpair %p\n", tqpair); 1004 1005 spdk_trace_record(TRACE_TCP_QP_DISCONNECT, 0, 0, (uintptr_t)tqpair); 1006 1007 if (tqpair->state <= NVME_TCP_QPAIR_STATE_RUNNING) { 1008 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITING); 1009 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1010 spdk_poller_unregister(&tqpair->timeout_poller); 1011 1012 /* This will end up calling nvmf_tcp_close_qpair */ 1013 spdk_nvmf_qpair_disconnect(&tqpair->qpair, NULL, NULL); 1014 } 1015 } 1016 1017 static void 1018 _mgmt_pdu_write_done(void *_tqpair, int err) 1019 { 1020 struct spdk_nvmf_tcp_qpair *tqpair = _tqpair; 1021 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1022 1023 if (spdk_unlikely(err != 0)) { 1024 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1025 return; 1026 } 1027 1028 assert(pdu->cb_fn != NULL); 1029 pdu->cb_fn(pdu->cb_arg); 1030 } 1031 1032 static void 1033 _req_pdu_write_done(void *req, int err) 1034 { 1035 struct spdk_nvmf_tcp_req *tcp_req = req; 1036 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1037 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1038 1039 assert(tcp_req->pdu_in_use); 1040 tcp_req->pdu_in_use = false; 1041 1042 /* If the request is in a completed state, we're waiting for write completion to free it */ 1043 if (spdk_unlikely(tcp_req->state == TCP_REQUEST_STATE_COMPLETED)) { 1044 nvmf_tcp_request_free(tcp_req); 1045 return; 1046 } 1047 1048 if (spdk_unlikely(err != 0)) { 1049 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1050 return; 1051 } 1052 1053 assert(pdu->cb_fn != NULL); 1054 pdu->cb_fn(pdu->cb_arg); 1055 } 1056 1057 static void 1058 _pdu_write_done(struct nvme_tcp_pdu *pdu, int err) 1059 { 1060 pdu->sock_req.cb_fn(pdu->sock_req.cb_arg, err); 1061 } 1062 1063 static void 1064 _tcp_write_pdu(struct nvme_tcp_pdu *pdu) 1065 { 1066 int rc; 1067 uint32_t mapped_length; 1068 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1069 1070 pdu->sock_req.iovcnt = nvme_tcp_build_iovs(pdu->iov, SPDK_COUNTOF(pdu->iov), pdu, 1071 tqpair->host_hdgst_enable, tqpair->host_ddgst_enable, &mapped_length); 1072 spdk_sock_writev_async(tqpair->sock, &pdu->sock_req); 1073 1074 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP || 1075 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ) { 1076 /* Try to force the send immediately. */ 1077 rc = spdk_sock_flush(tqpair->sock); 1078 if (rc > 0 && (uint32_t)rc == mapped_length) { 1079 _pdu_write_done(pdu, 0); 1080 } else { 1081 SPDK_ERRLOG("Could not write %s to socket: rc=%d, errno=%d\n", 1082 pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_RESP ? 1083 "IC_RESP" : "TERM_REQ", rc, errno); 1084 _pdu_write_done(pdu, rc >= 0 ? -EAGAIN : -errno); 1085 } 1086 } 1087 } 1088 1089 static void 1090 data_crc32_accel_done(void *cb_arg, int status) 1091 { 1092 struct nvme_tcp_pdu *pdu = cb_arg; 1093 1094 if (spdk_unlikely(status)) { 1095 SPDK_ERRLOG("Failed to compute the data digest for pdu =%p\n", pdu); 1096 _pdu_write_done(pdu, status); 1097 return; 1098 } 1099 1100 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 1101 MAKE_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32); 1102 1103 _tcp_write_pdu(pdu); 1104 } 1105 1106 static void 1107 pdu_data_crc32_compute(struct nvme_tcp_pdu *pdu) 1108 { 1109 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 1110 int rc = 0; 1111 1112 /* Data Digest */ 1113 if (pdu->data_len > 0 && g_nvme_tcp_ddgst[pdu->hdr.common.pdu_type] && tqpair->host_ddgst_enable) { 1114 /* Only support this limitated case for the first step */ 1115 if (spdk_likely(!pdu->dif_ctx && (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0) 1116 && tqpair->group)) { 1117 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 1118 pdu->data_iovcnt, 0, data_crc32_accel_done, pdu); 1119 if (spdk_likely(rc == 0)) { 1120 return; 1121 } 1122 } else { 1123 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 1124 } 1125 data_crc32_accel_done(pdu, rc); 1126 } else { 1127 _tcp_write_pdu(pdu); 1128 } 1129 } 1130 1131 static void 1132 nvmf_tcp_qpair_write_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1133 struct nvme_tcp_pdu *pdu, 1134 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1135 void *cb_arg) 1136 { 1137 int hlen; 1138 uint32_t crc32c; 1139 1140 assert(tqpair->pdu_in_progress != pdu); 1141 1142 hlen = pdu->hdr.common.hlen; 1143 pdu->cb_fn = cb_fn; 1144 pdu->cb_arg = cb_arg; 1145 1146 pdu->iov[0].iov_base = &pdu->hdr.raw; 1147 pdu->iov[0].iov_len = hlen; 1148 1149 /* Header Digest */ 1150 if (g_nvme_tcp_hdgst[pdu->hdr.common.pdu_type] && tqpair->host_hdgst_enable) { 1151 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 1152 MAKE_DIGEST_WORD((uint8_t *)pdu->hdr.raw + hlen, crc32c); 1153 } 1154 1155 /* Data Digest */ 1156 pdu_data_crc32_compute(pdu); 1157 } 1158 1159 static void 1160 nvmf_tcp_qpair_write_mgmt_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1161 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1162 void *cb_arg) 1163 { 1164 struct nvme_tcp_pdu *pdu = tqpair->mgmt_pdu; 1165 1166 pdu->sock_req.cb_fn = _mgmt_pdu_write_done; 1167 pdu->sock_req.cb_arg = tqpair; 1168 1169 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1170 } 1171 1172 static void 1173 nvmf_tcp_qpair_write_req_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1174 struct spdk_nvmf_tcp_req *tcp_req, 1175 nvme_tcp_qpair_xfer_complete_cb cb_fn, 1176 void *cb_arg) 1177 { 1178 struct nvme_tcp_pdu *pdu = tcp_req->pdu; 1179 1180 pdu->sock_req.cb_fn = _req_pdu_write_done; 1181 pdu->sock_req.cb_arg = tcp_req; 1182 1183 assert(!tcp_req->pdu_in_use); 1184 tcp_req->pdu_in_use = true; 1185 1186 nvmf_tcp_qpair_write_pdu(tqpair, pdu, cb_fn, cb_arg); 1187 } 1188 1189 static int 1190 nvmf_tcp_qpair_init_mem_resource(struct spdk_nvmf_tcp_qpair *tqpair) 1191 { 1192 uint32_t i; 1193 struct spdk_nvmf_transport_opts *opts; 1194 uint32_t in_capsule_data_size; 1195 1196 opts = &tqpair->qpair.transport->opts; 1197 1198 in_capsule_data_size = opts->in_capsule_data_size; 1199 if (opts->dif_insert_or_strip) { 1200 in_capsule_data_size = SPDK_BDEV_BUF_SIZE_WITH_MD(in_capsule_data_size); 1201 } 1202 1203 tqpair->resource_count = opts->max_queue_depth; 1204 1205 tqpair->reqs = calloc(tqpair->resource_count, sizeof(*tqpair->reqs)); 1206 if (!tqpair->reqs) { 1207 SPDK_ERRLOG("Unable to allocate reqs on tqpair=%p\n", tqpair); 1208 return -1; 1209 } 1210 1211 if (in_capsule_data_size) { 1212 tqpair->bufs = spdk_zmalloc(tqpair->resource_count * in_capsule_data_size, 0x1000, 1213 NULL, SPDK_ENV_LCORE_ID_ANY, 1214 SPDK_MALLOC_DMA); 1215 if (!tqpair->bufs) { 1216 SPDK_ERRLOG("Unable to allocate bufs on tqpair=%p.\n", tqpair); 1217 return -1; 1218 } 1219 } 1220 /* prepare memory space for receiving pdus and tcp_req */ 1221 /* Add additional 1 member, which will be used for mgmt_pdu owned by the tqpair */ 1222 tqpair->pdus = spdk_dma_zmalloc((2 * tqpair->resource_count + 1) * sizeof(*tqpair->pdus), 0x1000, 1223 NULL); 1224 if (!tqpair->pdus) { 1225 SPDK_ERRLOG("Unable to allocate pdu pool on tqpair =%p.\n", tqpair); 1226 return -1; 1227 } 1228 1229 for (i = 0; i < tqpair->resource_count; i++) { 1230 struct spdk_nvmf_tcp_req *tcp_req = &tqpair->reqs[i]; 1231 1232 tcp_req->ttag = i + 1; 1233 tcp_req->req.qpair = &tqpair->qpair; 1234 1235 tcp_req->pdu = &tqpair->pdus[i]; 1236 tcp_req->pdu->qpair = tqpair; 1237 1238 /* Set up memory to receive commands */ 1239 if (tqpair->bufs) { 1240 tcp_req->buf = (void *)((uintptr_t)tqpair->bufs + (i * in_capsule_data_size)); 1241 } 1242 1243 /* Set the cmdn and rsp */ 1244 tcp_req->req.rsp = (union nvmf_c2h_msg *)&tcp_req->rsp; 1245 tcp_req->req.cmd = (union nvmf_h2c_msg *)&tcp_req->cmd; 1246 1247 tcp_req->req.stripped_data = NULL; 1248 1249 /* Initialize request state to FREE */ 1250 tcp_req->state = TCP_REQUEST_STATE_FREE; 1251 TAILQ_INSERT_TAIL(&tqpair->tcp_req_free_queue, tcp_req, state_link); 1252 tqpair->state_cntr[TCP_REQUEST_STATE_FREE]++; 1253 } 1254 1255 for (; i < 2 * tqpair->resource_count; i++) { 1256 struct nvme_tcp_pdu *pdu = &tqpair->pdus[i]; 1257 1258 pdu->qpair = tqpair; 1259 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 1260 } 1261 1262 tqpair->mgmt_pdu = &tqpair->pdus[i]; 1263 tqpair->mgmt_pdu->qpair = tqpair; 1264 tqpair->pdu_in_progress = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 1265 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 1266 tqpair->tcp_pdu_working_count = 1; 1267 1268 tqpair->recv_buf_size = (in_capsule_data_size + sizeof(struct spdk_nvme_tcp_cmd) + 2 * 1269 SPDK_NVME_TCP_DIGEST_LEN) * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 1270 1271 return 0; 1272 } 1273 1274 static int 1275 nvmf_tcp_qpair_init(struct spdk_nvmf_qpair *qpair) 1276 { 1277 struct spdk_nvmf_tcp_qpair *tqpair; 1278 1279 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1280 1281 SPDK_DEBUGLOG(nvmf_tcp, "New TCP Connection: %p\n", qpair); 1282 1283 spdk_trace_record(TRACE_TCP_QP_CREATE, 0, 0, (uintptr_t)tqpair); 1284 1285 /* Initialise request state queues of the qpair */ 1286 TAILQ_INIT(&tqpair->tcp_req_free_queue); 1287 TAILQ_INIT(&tqpair->tcp_req_working_queue); 1288 SLIST_INIT(&tqpair->tcp_pdu_free_queue); 1289 1290 tqpair->host_hdgst_enable = true; 1291 tqpair->host_ddgst_enable = true; 1292 1293 return 0; 1294 } 1295 1296 static int 1297 nvmf_tcp_qpair_sock_init(struct spdk_nvmf_tcp_qpair *tqpair) 1298 { 1299 int rc; 1300 1301 spdk_trace_record(TRACE_TCP_QP_SOCK_INIT, 0, 0, (uintptr_t)tqpair); 1302 1303 /* set low water mark */ 1304 rc = spdk_sock_set_recvlowat(tqpair->sock, 1); 1305 if (rc != 0) { 1306 SPDK_ERRLOG("spdk_sock_set_recvlowat() failed\n"); 1307 return rc; 1308 } 1309 1310 return 0; 1311 } 1312 1313 static void 1314 nvmf_tcp_handle_connect(struct spdk_nvmf_transport *transport, 1315 struct spdk_nvmf_tcp_port *port, 1316 struct spdk_sock *sock) 1317 { 1318 struct spdk_nvmf_tcp_qpair *tqpair; 1319 int rc; 1320 1321 SPDK_DEBUGLOG(nvmf_tcp, "New connection accepted on %s port %s\n", 1322 port->trid->traddr, port->trid->trsvcid); 1323 1324 tqpair = calloc(1, sizeof(struct spdk_nvmf_tcp_qpair)); 1325 if (tqpair == NULL) { 1326 SPDK_ERRLOG("Could not allocate new connection.\n"); 1327 spdk_sock_close(&sock); 1328 return; 1329 } 1330 1331 tqpair->sock = sock; 1332 tqpair->state_cntr[TCP_REQUEST_STATE_FREE] = 0; 1333 tqpair->port = port; 1334 tqpair->qpair.transport = transport; 1335 1336 rc = spdk_sock_getaddr(tqpair->sock, tqpair->target_addr, 1337 sizeof(tqpair->target_addr), &tqpair->target_port, 1338 tqpair->initiator_addr, sizeof(tqpair->initiator_addr), 1339 &tqpair->initiator_port); 1340 if (rc < 0) { 1341 SPDK_ERRLOG("spdk_sock_getaddr() failed of tqpair=%p\n", tqpair); 1342 nvmf_tcp_qpair_destroy(tqpair); 1343 return; 1344 } 1345 1346 spdk_nvmf_tgt_new_qpair(transport->tgt, &tqpair->qpair); 1347 } 1348 1349 static uint32_t 1350 nvmf_tcp_port_accept(struct spdk_nvmf_transport *transport, struct spdk_nvmf_tcp_port *port) 1351 { 1352 struct spdk_sock *sock; 1353 uint32_t count = 0; 1354 int i; 1355 1356 for (i = 0; i < NVMF_TCP_MAX_ACCEPT_SOCK_ONE_TIME; i++) { 1357 sock = spdk_sock_accept(port->listen_sock); 1358 if (sock == NULL) { 1359 break; 1360 } 1361 count++; 1362 nvmf_tcp_handle_connect(transport, port, sock); 1363 } 1364 1365 return count; 1366 } 1367 1368 static int 1369 nvmf_tcp_accept(void *ctx) 1370 { 1371 struct spdk_nvmf_transport *transport = ctx; 1372 struct spdk_nvmf_tcp_transport *ttransport; 1373 struct spdk_nvmf_tcp_port *port; 1374 uint32_t count = 0; 1375 1376 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1377 1378 TAILQ_FOREACH(port, &ttransport->ports, link) { 1379 count += nvmf_tcp_port_accept(transport, port); 1380 } 1381 1382 return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE; 1383 } 1384 1385 static void 1386 nvmf_tcp_discover(struct spdk_nvmf_transport *transport, 1387 struct spdk_nvme_transport_id *trid, 1388 struct spdk_nvmf_discovery_log_page_entry *entry) 1389 { 1390 struct spdk_nvmf_tcp_port *port; 1391 struct spdk_nvmf_tcp_transport *ttransport; 1392 1393 entry->trtype = SPDK_NVMF_TRTYPE_TCP; 1394 entry->adrfam = trid->adrfam; 1395 1396 spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' '); 1397 spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' '); 1398 1399 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1400 port = nvmf_tcp_find_port(ttransport, trid); 1401 1402 assert(port != NULL); 1403 1404 if (strcmp(spdk_sock_get_impl_name(port->listen_sock), "ssl") == 0) { 1405 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_REQUIRED; 1406 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_TLS_1_3; 1407 } else { 1408 entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED; 1409 entry->tsas.tcp.sectype = SPDK_NVME_TCP_SECURITY_NONE; 1410 } 1411 } 1412 1413 static struct spdk_nvmf_tcp_control_msg_list * 1414 nvmf_tcp_control_msg_list_create(uint16_t num_messages) 1415 { 1416 struct spdk_nvmf_tcp_control_msg_list *list; 1417 struct spdk_nvmf_tcp_control_msg *msg; 1418 uint16_t i; 1419 1420 list = calloc(1, sizeof(*list)); 1421 if (!list) { 1422 SPDK_ERRLOG("Failed to allocate memory for list structure\n"); 1423 return NULL; 1424 } 1425 1426 list->msg_buf = spdk_zmalloc(num_messages * SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE, 1427 NVMF_DATA_BUFFER_ALIGNMENT, NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA); 1428 if (!list->msg_buf) { 1429 SPDK_ERRLOG("Failed to allocate memory for control message buffers\n"); 1430 free(list); 1431 return NULL; 1432 } 1433 1434 STAILQ_INIT(&list->free_msgs); 1435 1436 for (i = 0; i < num_messages; i++) { 1437 msg = (struct spdk_nvmf_tcp_control_msg *)((char *)list->msg_buf + i * 1438 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1439 STAILQ_INSERT_TAIL(&list->free_msgs, msg, link); 1440 } 1441 1442 return list; 1443 } 1444 1445 static void 1446 nvmf_tcp_control_msg_list_free(struct spdk_nvmf_tcp_control_msg_list *list) 1447 { 1448 if (!list) { 1449 return; 1450 } 1451 1452 spdk_free(list->msg_buf); 1453 free(list); 1454 } 1455 1456 static struct spdk_nvmf_transport_poll_group * 1457 nvmf_tcp_poll_group_create(struct spdk_nvmf_transport *transport, 1458 struct spdk_nvmf_poll_group *group) 1459 { 1460 struct spdk_nvmf_tcp_transport *ttransport; 1461 struct spdk_nvmf_tcp_poll_group *tgroup; 1462 1463 tgroup = calloc(1, sizeof(*tgroup)); 1464 if (!tgroup) { 1465 return NULL; 1466 } 1467 1468 tgroup->sock_group = spdk_sock_group_create(&tgroup->group); 1469 if (!tgroup->sock_group) { 1470 goto cleanup; 1471 } 1472 1473 TAILQ_INIT(&tgroup->qpairs); 1474 TAILQ_INIT(&tgroup->await_req); 1475 1476 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 1477 1478 if (transport->opts.in_capsule_data_size < SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE) { 1479 SPDK_DEBUGLOG(nvmf_tcp, "ICD %u is less than min required for admin/fabric commands (%u). " 1480 "Creating control messages list\n", transport->opts.in_capsule_data_size, 1481 SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE); 1482 tgroup->control_msg_list = nvmf_tcp_control_msg_list_create(ttransport->tcp_opts.control_msg_num); 1483 if (!tgroup->control_msg_list) { 1484 goto cleanup; 1485 } 1486 } 1487 1488 tgroup->accel_channel = spdk_accel_get_io_channel(); 1489 if (spdk_unlikely(!tgroup->accel_channel)) { 1490 SPDK_ERRLOG("Cannot create accel_channel for tgroup=%p\n", tgroup); 1491 goto cleanup; 1492 } 1493 1494 TAILQ_INSERT_TAIL(&ttransport->poll_groups, tgroup, link); 1495 if (ttransport->next_pg == NULL) { 1496 ttransport->next_pg = tgroup; 1497 } 1498 1499 return &tgroup->group; 1500 1501 cleanup: 1502 nvmf_tcp_poll_group_destroy(&tgroup->group); 1503 return NULL; 1504 } 1505 1506 static struct spdk_nvmf_transport_poll_group * 1507 nvmf_tcp_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair) 1508 { 1509 struct spdk_nvmf_tcp_transport *ttransport; 1510 struct spdk_nvmf_tcp_poll_group **pg; 1511 struct spdk_nvmf_tcp_qpair *tqpair; 1512 struct spdk_sock_group *group = NULL, *hint = NULL; 1513 int rc; 1514 1515 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 1516 1517 if (TAILQ_EMPTY(&ttransport->poll_groups)) { 1518 return NULL; 1519 } 1520 1521 pg = &ttransport->next_pg; 1522 assert(*pg != NULL); 1523 hint = (*pg)->sock_group; 1524 1525 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 1526 rc = spdk_sock_get_optimal_sock_group(tqpair->sock, &group, hint); 1527 if (rc != 0) { 1528 return NULL; 1529 } else if (group != NULL) { 1530 /* Optimal poll group was found */ 1531 return spdk_sock_group_get_ctx(group); 1532 } 1533 1534 /* The hint was used for optimal poll group, advance next_pg. */ 1535 *pg = TAILQ_NEXT(*pg, link); 1536 if (*pg == NULL) { 1537 *pg = TAILQ_FIRST(&ttransport->poll_groups); 1538 } 1539 1540 return spdk_sock_group_get_ctx(hint); 1541 } 1542 1543 static void 1544 nvmf_tcp_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group) 1545 { 1546 struct spdk_nvmf_tcp_poll_group *tgroup, *next_tgroup; 1547 struct spdk_nvmf_tcp_transport *ttransport; 1548 1549 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 1550 spdk_sock_group_close(&tgroup->sock_group); 1551 if (tgroup->control_msg_list) { 1552 nvmf_tcp_control_msg_list_free(tgroup->control_msg_list); 1553 } 1554 1555 if (tgroup->accel_channel) { 1556 spdk_put_io_channel(tgroup->accel_channel); 1557 } 1558 1559 if (tgroup->group.transport == NULL) { 1560 /* Transport can be NULL when nvmf_tcp_poll_group_create() 1561 * calls this function directly in a failure path. */ 1562 free(tgroup); 1563 return; 1564 } 1565 1566 ttransport = SPDK_CONTAINEROF(tgroup->group.transport, struct spdk_nvmf_tcp_transport, transport); 1567 1568 next_tgroup = TAILQ_NEXT(tgroup, link); 1569 TAILQ_REMOVE(&ttransport->poll_groups, tgroup, link); 1570 if (next_tgroup == NULL) { 1571 next_tgroup = TAILQ_FIRST(&ttransport->poll_groups); 1572 } 1573 if (ttransport->next_pg == tgroup) { 1574 ttransport->next_pg = next_tgroup; 1575 } 1576 1577 free(tgroup); 1578 } 1579 1580 static void 1581 nvmf_tcp_qpair_set_recv_state(struct spdk_nvmf_tcp_qpair *tqpair, 1582 enum nvme_tcp_pdu_recv_state state) 1583 { 1584 if (tqpair->recv_state == state) { 1585 SPDK_ERRLOG("The recv state of tqpair=%p is same with the state(%d) to be set\n", 1586 tqpair, state); 1587 return; 1588 } 1589 1590 if (spdk_unlikely(state == NVME_TCP_PDU_RECV_STATE_QUIESCING)) { 1591 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH && tqpair->pdu_in_progress) { 1592 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, tqpair->pdu_in_progress, slist); 1593 tqpair->tcp_pdu_working_count--; 1594 } 1595 } 1596 1597 if (spdk_unlikely(state == NVME_TCP_PDU_RECV_STATE_ERROR)) { 1598 assert(tqpair->tcp_pdu_working_count == 0); 1599 } 1600 1601 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1602 /* When leaving the await req state, move the qpair to the main list */ 1603 TAILQ_REMOVE(&tqpair->group->await_req, tqpair, link); 1604 TAILQ_INSERT_TAIL(&tqpair->group->qpairs, tqpair, link); 1605 } else if (state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 1606 TAILQ_REMOVE(&tqpair->group->qpairs, tqpair, link); 1607 TAILQ_INSERT_TAIL(&tqpair->group->await_req, tqpair, link); 1608 } 1609 1610 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv state=%d\n", tqpair, state); 1611 tqpair->recv_state = state; 1612 1613 spdk_trace_record(TRACE_TCP_QP_RCV_STATE_CHANGE, tqpair->qpair.qid, 0, (uintptr_t)tqpair, 1614 tqpair->recv_state); 1615 } 1616 1617 static int 1618 nvmf_tcp_qpair_handle_timeout(void *ctx) 1619 { 1620 struct spdk_nvmf_tcp_qpair *tqpair = ctx; 1621 1622 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 1623 1624 SPDK_ERRLOG("No pdu coming for tqpair=%p within %d seconds\n", tqpair, 1625 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT); 1626 1627 nvmf_tcp_qpair_disconnect(tqpair); 1628 return SPDK_POLLER_BUSY; 1629 } 1630 1631 static void 1632 nvmf_tcp_send_c2h_term_req_complete(void *cb_arg) 1633 { 1634 struct spdk_nvmf_tcp_qpair *tqpair = (struct spdk_nvmf_tcp_qpair *)cb_arg; 1635 1636 if (!tqpair->timeout_poller) { 1637 tqpair->timeout_poller = SPDK_POLLER_REGISTER(nvmf_tcp_qpair_handle_timeout, tqpair, 1638 SPDK_NVME_TCP_QPAIR_EXIT_TIMEOUT * 1000000); 1639 } 1640 } 1641 1642 static void 1643 nvmf_tcp_send_c2h_term_req(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu, 1644 enum spdk_nvme_tcp_term_req_fes fes, uint32_t error_offset) 1645 { 1646 struct nvme_tcp_pdu *rsp_pdu; 1647 struct spdk_nvme_tcp_term_req_hdr *c2h_term_req; 1648 uint32_t c2h_term_req_hdr_len = sizeof(*c2h_term_req); 1649 uint32_t copy_len; 1650 1651 rsp_pdu = tqpair->mgmt_pdu; 1652 1653 c2h_term_req = &rsp_pdu->hdr.term_req; 1654 c2h_term_req->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_TERM_REQ; 1655 c2h_term_req->common.hlen = c2h_term_req_hdr_len; 1656 c2h_term_req->fes = fes; 1657 1658 if ((fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1659 (fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1660 DSET32(&c2h_term_req->fei, error_offset); 1661 } 1662 1663 copy_len = spdk_min(pdu->hdr.common.hlen, SPDK_NVME_TCP_TERM_REQ_ERROR_DATA_MAX_SIZE); 1664 1665 /* Copy the error info into the buffer */ 1666 memcpy((uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, pdu->hdr.raw, copy_len); 1667 nvme_tcp_pdu_set_data(rsp_pdu, (uint8_t *)rsp_pdu->hdr.raw + c2h_term_req_hdr_len, copy_len); 1668 1669 /* Contain the header of the wrong received pdu */ 1670 c2h_term_req->common.plen = c2h_term_req->common.hlen + copy_len; 1671 tqpair->wait_terminate = true; 1672 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1673 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_c2h_term_req_complete, tqpair); 1674 } 1675 1676 static void 1677 nvmf_tcp_capsule_cmd_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1678 struct spdk_nvmf_tcp_qpair *tqpair, 1679 struct nvme_tcp_pdu *pdu) 1680 { 1681 struct spdk_nvmf_tcp_req *tcp_req; 1682 1683 assert(pdu->psh_valid_bytes == pdu->psh_len); 1684 assert(pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD); 1685 1686 tcp_req = nvmf_tcp_req_get(tqpair); 1687 if (!tcp_req) { 1688 /* Directly return and make the allocation retry again. This can happen if we're 1689 * using asynchronous writes to send the response to the host or when releasing 1690 * zero-copy buffers after a response has been sent. In both cases, the host might 1691 * receive the response before we've finished processing the request and is free to 1692 * send another one. 1693 */ 1694 if (tqpair->state_cntr[TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST] > 0 || 1695 tqpair->state_cntr[TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE] > 0) { 1696 return; 1697 } 1698 1699 /* The host sent more commands than the maximum queue depth. */ 1700 SPDK_ERRLOG("Cannot allocate tcp_req on tqpair=%p\n", tqpair); 1701 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1702 return; 1703 } 1704 1705 pdu->req = tcp_req; 1706 assert(tcp_req->state == TCP_REQUEST_STATE_NEW); 1707 nvmf_tcp_req_process(ttransport, tcp_req); 1708 } 1709 1710 static void 1711 nvmf_tcp_capsule_cmd_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1712 struct spdk_nvmf_tcp_qpair *tqpair, 1713 struct nvme_tcp_pdu *pdu) 1714 { 1715 struct spdk_nvmf_tcp_req *tcp_req; 1716 struct spdk_nvme_tcp_cmd *capsule_cmd; 1717 uint32_t error_offset = 0; 1718 enum spdk_nvme_tcp_term_req_fes fes; 1719 struct spdk_nvme_cpl *rsp; 1720 1721 capsule_cmd = &pdu->hdr.capsule_cmd; 1722 tcp_req = pdu->req; 1723 assert(tcp_req != NULL); 1724 1725 /* Zero-copy requests don't support ICD */ 1726 assert(!spdk_nvmf_request_using_zcopy(&tcp_req->req)); 1727 1728 if (capsule_cmd->common.pdo > SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET) { 1729 SPDK_ERRLOG("Expected ICReq capsule_cmd pdu offset <= %d, got %c\n", 1730 SPDK_NVME_TCP_PDU_PDO_MAX_OFFSET, capsule_cmd->common.pdo); 1731 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1732 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 1733 goto err; 1734 } 1735 1736 rsp = &tcp_req->req.rsp->nvme_cpl; 1737 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1738 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1739 } else { 1740 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1741 } 1742 1743 nvmf_tcp_req_process(ttransport, tcp_req); 1744 1745 return; 1746 err: 1747 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1748 } 1749 1750 static void 1751 nvmf_tcp_h2c_data_hdr_handle(struct spdk_nvmf_tcp_transport *ttransport, 1752 struct spdk_nvmf_tcp_qpair *tqpair, 1753 struct nvme_tcp_pdu *pdu) 1754 { 1755 struct spdk_nvmf_tcp_req *tcp_req; 1756 uint32_t error_offset = 0; 1757 enum spdk_nvme_tcp_term_req_fes fes = 0; 1758 struct spdk_nvme_tcp_h2c_data_hdr *h2c_data; 1759 1760 h2c_data = &pdu->hdr.h2c_data; 1761 1762 SPDK_DEBUGLOG(nvmf_tcp, "tqpair=%p, r2t_info: datao=%u, datal=%u, cccid=%u, ttag=%u\n", 1763 tqpair, h2c_data->datao, h2c_data->datal, h2c_data->cccid, h2c_data->ttag); 1764 1765 if (h2c_data->ttag > tqpair->resource_count) { 1766 SPDK_DEBUGLOG(nvmf_tcp, "ttag %u is larger than allowed %u.\n", h2c_data->ttag, 1767 tqpair->resource_count); 1768 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1769 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1770 goto err; 1771 } 1772 1773 tcp_req = &tqpair->reqs[h2c_data->ttag - 1]; 1774 1775 if (spdk_unlikely(tcp_req->state != TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER && 1776 tcp_req->state != TCP_REQUEST_STATE_AWAITING_R2T_ACK)) { 1777 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, has error state in %d\n", tcp_req, tqpair, 1778 tcp_req->state); 1779 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1780 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, ttag); 1781 goto err; 1782 } 1783 1784 if (spdk_unlikely(tcp_req->req.cmd->nvme_cmd.cid != h2c_data->cccid)) { 1785 SPDK_DEBUGLOG(nvmf_tcp, "tcp_req(%p), tqpair=%p, expected %u but %u for cccid.\n", tcp_req, tqpair, 1786 tcp_req->req.cmd->nvme_cmd.cid, h2c_data->cccid); 1787 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 1788 error_offset = offsetof(struct spdk_nvme_tcp_h2c_data_hdr, cccid); 1789 goto err; 1790 } 1791 1792 if (tcp_req->h2c_offset != h2c_data->datao) { 1793 SPDK_DEBUGLOG(nvmf_tcp, 1794 "tcp_req(%p), tqpair=%p, expected data offset %u, but data offset is %u\n", 1795 tcp_req, tqpair, tcp_req->h2c_offset, h2c_data->datao); 1796 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1797 goto err; 1798 } 1799 1800 if ((h2c_data->datao + h2c_data->datal) > tcp_req->req.length) { 1801 SPDK_DEBUGLOG(nvmf_tcp, 1802 "tcp_req(%p), tqpair=%p, (datao=%u + datal=%u) exceeds requested length=%u\n", 1803 tcp_req, tqpair, h2c_data->datao, h2c_data->datal, tcp_req->req.length); 1804 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_OUT_OF_RANGE; 1805 goto err; 1806 } 1807 1808 pdu->req = tcp_req; 1809 1810 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 1811 pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 1812 } 1813 1814 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 1815 h2c_data->datao, h2c_data->datal); 1816 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1817 return; 1818 1819 err: 1820 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1821 } 1822 1823 static void 1824 nvmf_tcp_send_capsule_resp_pdu(struct spdk_nvmf_tcp_req *tcp_req, 1825 struct spdk_nvmf_tcp_qpair *tqpair) 1826 { 1827 struct nvme_tcp_pdu *rsp_pdu; 1828 struct spdk_nvme_tcp_rsp *capsule_resp; 1829 1830 SPDK_DEBUGLOG(nvmf_tcp, "enter, tqpair=%p\n", tqpair); 1831 1832 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1833 assert(rsp_pdu != NULL); 1834 1835 capsule_resp = &rsp_pdu->hdr.capsule_resp; 1836 capsule_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_CAPSULE_RESP; 1837 capsule_resp->common.plen = capsule_resp->common.hlen = sizeof(*capsule_resp); 1838 capsule_resp->rccqe = tcp_req->req.rsp->nvme_cpl; 1839 if (tqpair->host_hdgst_enable) { 1840 capsule_resp->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 1841 capsule_resp->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 1842 } 1843 1844 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_request_free, tcp_req); 1845 } 1846 1847 static void 1848 nvmf_tcp_pdu_c2h_data_complete(void *cb_arg) 1849 { 1850 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 1851 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, 1852 struct spdk_nvmf_tcp_qpair, qpair); 1853 1854 assert(tqpair != NULL); 1855 1856 if (spdk_unlikely(tcp_req->pdu->rw_offset < tcp_req->req.length)) { 1857 SPDK_DEBUGLOG(nvmf_tcp, "sending another C2H part, offset %u length %u\n", tcp_req->pdu->rw_offset, 1858 tcp_req->req.length); 1859 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 1860 return; 1861 } 1862 1863 if (tcp_req->pdu->hdr.c2h_data.common.flags & SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS) { 1864 nvmf_tcp_request_free(tcp_req); 1865 } else { 1866 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 1867 } 1868 } 1869 1870 static void 1871 nvmf_tcp_r2t_complete(void *cb_arg) 1872 { 1873 struct spdk_nvmf_tcp_req *tcp_req = cb_arg; 1874 struct spdk_nvmf_tcp_transport *ttransport; 1875 1876 ttransport = SPDK_CONTAINEROF(tcp_req->req.qpair->transport, 1877 struct spdk_nvmf_tcp_transport, transport); 1878 1879 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 1880 1881 if (tcp_req->h2c_offset == tcp_req->req.length) { 1882 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1883 nvmf_tcp_req_process(ttransport, tcp_req); 1884 } 1885 } 1886 1887 static void 1888 nvmf_tcp_send_r2t_pdu(struct spdk_nvmf_tcp_qpair *tqpair, 1889 struct spdk_nvmf_tcp_req *tcp_req) 1890 { 1891 struct nvme_tcp_pdu *rsp_pdu; 1892 struct spdk_nvme_tcp_r2t_hdr *r2t; 1893 1894 rsp_pdu = nvmf_tcp_req_pdu_init(tcp_req); 1895 assert(rsp_pdu != NULL); 1896 1897 r2t = &rsp_pdu->hdr.r2t; 1898 r2t->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_R2T; 1899 r2t->common.plen = r2t->common.hlen = sizeof(*r2t); 1900 1901 if (tqpair->host_hdgst_enable) { 1902 r2t->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 1903 r2t->common.plen += SPDK_NVME_TCP_DIGEST_LEN; 1904 } 1905 1906 r2t->cccid = tcp_req->req.cmd->nvme_cmd.cid; 1907 r2t->ttag = tcp_req->ttag; 1908 r2t->r2to = tcp_req->h2c_offset; 1909 r2t->r2tl = tcp_req->req.length; 1910 1911 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_R2T_ACK); 1912 1913 SPDK_DEBUGLOG(nvmf_tcp, 1914 "tcp_req(%p) on tqpair(%p), r2t_info: cccid=%u, ttag=%u, r2to=%u, r2tl=%u\n", 1915 tcp_req, tqpair, r2t->cccid, r2t->ttag, r2t->r2to, r2t->r2tl); 1916 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_r2t_complete, tcp_req); 1917 } 1918 1919 static void 1920 nvmf_tcp_h2c_data_payload_handle(struct spdk_nvmf_tcp_transport *ttransport, 1921 struct spdk_nvmf_tcp_qpair *tqpair, 1922 struct nvme_tcp_pdu *pdu) 1923 { 1924 struct spdk_nvmf_tcp_req *tcp_req; 1925 struct spdk_nvme_cpl *rsp; 1926 1927 tcp_req = pdu->req; 1928 assert(tcp_req != NULL); 1929 1930 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 1931 1932 tcp_req->h2c_offset += pdu->data_len; 1933 1934 /* Wait for all of the data to arrive AND for the initial R2T PDU send to be 1935 * acknowledged before moving on. */ 1936 if (tcp_req->h2c_offset == tcp_req->req.length && 1937 tcp_req->state == TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER) { 1938 /* After receiving all the h2c data, we need to check whether there is 1939 * transient transport error */ 1940 rsp = &tcp_req->req.rsp->nvme_cpl; 1941 if (spdk_unlikely(rsp->status.sc == SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR)) { 1942 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 1943 } else { 1944 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 1945 } 1946 nvmf_tcp_req_process(ttransport, tcp_req); 1947 } 1948 } 1949 1950 static void 1951 nvmf_tcp_h2c_term_req_dump(struct spdk_nvme_tcp_term_req_hdr *h2c_term_req) 1952 { 1953 SPDK_ERRLOG("Error info of pdu(%p): %s\n", h2c_term_req, 1954 spdk_nvmf_tcp_term_req_fes_str[h2c_term_req->fes]); 1955 if ((h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD) || 1956 (h2c_term_req->fes == SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER)) { 1957 SPDK_DEBUGLOG(nvmf_tcp, "The offset from the start of the PDU header is %u\n", 1958 DGET32(h2c_term_req->fei)); 1959 } 1960 } 1961 1962 static void 1963 nvmf_tcp_h2c_term_req_hdr_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1964 struct nvme_tcp_pdu *pdu) 1965 { 1966 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 1967 uint32_t error_offset = 0; 1968 enum spdk_nvme_tcp_term_req_fes fes; 1969 1970 if (h2c_term_req->fes > SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER) { 1971 SPDK_ERRLOG("Fatal Error Status(FES) is unknown for h2c_term_req pdu=%p\n", pdu); 1972 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 1973 error_offset = offsetof(struct spdk_nvme_tcp_term_req_hdr, fes); 1974 goto end; 1975 } 1976 1977 /* set the data buffer */ 1978 nvme_tcp_pdu_set_data(pdu, (uint8_t *)pdu->hdr.raw + h2c_term_req->common.hlen, 1979 h2c_term_req->common.plen - h2c_term_req->common.hlen); 1980 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 1981 return; 1982 end: 1983 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 1984 } 1985 1986 static void 1987 nvmf_tcp_h2c_term_req_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, 1988 struct nvme_tcp_pdu *pdu) 1989 { 1990 struct spdk_nvme_tcp_term_req_hdr *h2c_term_req = &pdu->hdr.term_req; 1991 1992 nvmf_tcp_h2c_term_req_dump(h2c_term_req); 1993 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 1994 } 1995 1996 static void 1997 _nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 1998 { 1999 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2000 struct spdk_nvmf_tcp_transport, transport); 2001 2002 switch (pdu->hdr.common.pdu_type) { 2003 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2004 nvmf_tcp_capsule_cmd_payload_handle(ttransport, tqpair, pdu); 2005 break; 2006 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2007 nvmf_tcp_h2c_data_payload_handle(ttransport, tqpair, pdu); 2008 break; 2009 2010 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2011 nvmf_tcp_h2c_term_req_payload_handle(tqpair, pdu); 2012 break; 2013 2014 default: 2015 /* The code should not go to here */ 2016 SPDK_ERRLOG("ERROR pdu type %d\n", pdu->hdr.common.pdu_type); 2017 break; 2018 } 2019 SLIST_INSERT_HEAD(&tqpair->tcp_pdu_free_queue, pdu, slist); 2020 tqpair->tcp_pdu_working_count--; 2021 } 2022 2023 static inline void 2024 nvmf_tcp_req_set_cpl(struct spdk_nvmf_tcp_req *treq, int sct, int sc) 2025 { 2026 treq->req.rsp->nvme_cpl.status.sct = sct; 2027 treq->req.rsp->nvme_cpl.status.sc = sc; 2028 treq->req.rsp->nvme_cpl.cid = treq->req.cmd->nvme_cmd.cid; 2029 } 2030 2031 static void 2032 data_crc32_calc_done(void *cb_arg, int status) 2033 { 2034 struct nvme_tcp_pdu *pdu = cb_arg; 2035 struct spdk_nvmf_tcp_qpair *tqpair = pdu->qpair; 2036 2037 /* async crc32 calculation is failed and use direct calculation to check */ 2038 if (spdk_unlikely(status)) { 2039 SPDK_ERRLOG("Data digest on tqpair=(%p) with pdu=%p failed to be calculated asynchronously\n", 2040 tqpair, pdu); 2041 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 2042 } 2043 pdu->data_digest_crc32 ^= SPDK_CRC32C_XOR; 2044 if (!MATCH_DIGEST_WORD(pdu->data_digest, pdu->data_digest_crc32)) { 2045 SPDK_ERRLOG("Data digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 2046 assert(pdu->req != NULL); 2047 nvmf_tcp_req_set_cpl(pdu->req, SPDK_NVME_SCT_GENERIC, 2048 SPDK_NVME_SC_COMMAND_TRANSIENT_TRANSPORT_ERROR); 2049 } 2050 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2051 } 2052 2053 static void 2054 nvmf_tcp_pdu_payload_handle(struct spdk_nvmf_tcp_qpair *tqpair, struct nvme_tcp_pdu *pdu) 2055 { 2056 int rc = 0; 2057 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2058 tqpair->pdu_in_progress = NULL; 2059 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2060 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2061 /* check data digest if need */ 2062 if (pdu->ddgst_enable) { 2063 if (tqpair->qpair.qid != 0 && !pdu->dif_ctx && tqpair->group && 2064 (pdu->data_len % SPDK_NVME_TCP_DIGEST_ALIGNMENT == 0)) { 2065 rc = spdk_accel_submit_crc32cv(tqpair->group->accel_channel, &pdu->data_digest_crc32, pdu->data_iov, 2066 pdu->data_iovcnt, 0, data_crc32_calc_done, pdu); 2067 if (spdk_likely(rc == 0)) { 2068 return; 2069 } 2070 } else { 2071 pdu->data_digest_crc32 = nvme_tcp_pdu_calc_data_digest(pdu); 2072 } 2073 data_crc32_calc_done(pdu, rc); 2074 } else { 2075 _nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2076 } 2077 } 2078 2079 static void 2080 nvmf_tcp_send_icresp_complete(void *cb_arg) 2081 { 2082 struct spdk_nvmf_tcp_qpair *tqpair = cb_arg; 2083 2084 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_RUNNING); 2085 } 2086 2087 static void 2088 nvmf_tcp_icreq_handle(struct spdk_nvmf_tcp_transport *ttransport, 2089 struct spdk_nvmf_tcp_qpair *tqpair, 2090 struct nvme_tcp_pdu *pdu) 2091 { 2092 struct spdk_nvme_tcp_ic_req *ic_req = &pdu->hdr.ic_req; 2093 struct nvme_tcp_pdu *rsp_pdu; 2094 struct spdk_nvme_tcp_ic_resp *ic_resp; 2095 uint32_t error_offset = 0; 2096 enum spdk_nvme_tcp_term_req_fes fes; 2097 2098 /* Only PFV 0 is defined currently */ 2099 if (ic_req->pfv != 0) { 2100 SPDK_ERRLOG("Expected ICReq PFV %u, got %u\n", 0u, ic_req->pfv); 2101 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2102 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, pfv); 2103 goto end; 2104 } 2105 2106 /* This value is 0’s based value in units of dwords should not be larger than SPDK_NVME_TCP_HPDA_MAX */ 2107 if (ic_req->hpda > SPDK_NVME_TCP_HPDA_MAX) { 2108 SPDK_ERRLOG("ICReq HPDA out of range 0 to 31, got %u\n", ic_req->hpda); 2109 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2110 error_offset = offsetof(struct spdk_nvme_tcp_ic_req, hpda); 2111 goto end; 2112 } 2113 2114 /* MAXR2T is 0's based */ 2115 SPDK_DEBUGLOG(nvmf_tcp, "maxr2t =%u\n", (ic_req->maxr2t + 1u)); 2116 2117 tqpair->host_hdgst_enable = ic_req->dgst.bits.hdgst_enable ? true : false; 2118 if (!tqpair->host_hdgst_enable) { 2119 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 2120 } 2121 2122 tqpair->host_ddgst_enable = ic_req->dgst.bits.ddgst_enable ? true : false; 2123 if (!tqpair->host_ddgst_enable) { 2124 tqpair->recv_buf_size -= SPDK_NVME_TCP_DIGEST_LEN * SPDK_NVMF_TCP_RECV_BUF_SIZE_FACTOR; 2125 } 2126 2127 tqpair->recv_buf_size = spdk_max(tqpair->recv_buf_size, MIN_SOCK_PIPE_SIZE); 2128 /* Now that we know whether digests are enabled, properly size the receive buffer */ 2129 if (spdk_sock_set_recvbuf(tqpair->sock, tqpair->recv_buf_size) < 0) { 2130 SPDK_WARNLOG("Unable to allocate enough memory for receive buffer on tqpair=%p with size=%d\n", 2131 tqpair, 2132 tqpair->recv_buf_size); 2133 /* Not fatal. */ 2134 } 2135 2136 tqpair->cpda = spdk_min(ic_req->hpda, SPDK_NVME_TCP_CPDA_MAX); 2137 SPDK_DEBUGLOG(nvmf_tcp, "cpda of tqpair=(%p) is : %u\n", tqpair, tqpair->cpda); 2138 2139 rsp_pdu = tqpair->mgmt_pdu; 2140 2141 ic_resp = &rsp_pdu->hdr.ic_resp; 2142 ic_resp->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_IC_RESP; 2143 ic_resp->common.hlen = ic_resp->common.plen = sizeof(*ic_resp); 2144 ic_resp->pfv = 0; 2145 ic_resp->cpda = tqpair->cpda; 2146 ic_resp->maxh2cdata = ttransport->transport.opts.max_io_size; 2147 ic_resp->dgst.bits.hdgst_enable = tqpair->host_hdgst_enable ? 1 : 0; 2148 ic_resp->dgst.bits.ddgst_enable = tqpair->host_ddgst_enable ? 1 : 0; 2149 2150 SPDK_DEBUGLOG(nvmf_tcp, "host_hdgst_enable: %u\n", tqpair->host_hdgst_enable); 2151 SPDK_DEBUGLOG(nvmf_tcp, "host_ddgst_enable: %u\n", tqpair->host_ddgst_enable); 2152 2153 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INITIALIZING); 2154 nvmf_tcp_qpair_write_mgmt_pdu(tqpair, nvmf_tcp_send_icresp_complete, tqpair); 2155 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2156 return; 2157 end: 2158 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2159 } 2160 2161 static void 2162 nvmf_tcp_pdu_psh_handle(struct spdk_nvmf_tcp_qpair *tqpair, 2163 struct spdk_nvmf_tcp_transport *ttransport) 2164 { 2165 struct nvme_tcp_pdu *pdu; 2166 int rc; 2167 uint32_t crc32c, error_offset = 0; 2168 enum spdk_nvme_tcp_term_req_fes fes; 2169 2170 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2171 pdu = tqpair->pdu_in_progress; 2172 2173 SPDK_DEBUGLOG(nvmf_tcp, "pdu type of tqpair(%p) is %d\n", tqpair, 2174 pdu->hdr.common.pdu_type); 2175 /* check header digest if needed */ 2176 if (pdu->has_hdgst) { 2177 SPDK_DEBUGLOG(nvmf_tcp, "Compare the header of pdu=%p on tqpair=%p\n", pdu, tqpair); 2178 crc32c = nvme_tcp_pdu_calc_header_digest(pdu); 2179 rc = MATCH_DIGEST_WORD((uint8_t *)pdu->hdr.raw + pdu->hdr.common.hlen, crc32c); 2180 if (rc == 0) { 2181 SPDK_ERRLOG("Header digest error on tqpair=(%p) with pdu=%p\n", tqpair, pdu); 2182 fes = SPDK_NVME_TCP_TERM_REQ_FES_HDGST_ERROR; 2183 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2184 return; 2185 2186 } 2187 } 2188 2189 switch (pdu->hdr.common.pdu_type) { 2190 case SPDK_NVME_TCP_PDU_TYPE_IC_REQ: 2191 nvmf_tcp_icreq_handle(ttransport, tqpair, pdu); 2192 break; 2193 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2194 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_REQ); 2195 break; 2196 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2197 nvmf_tcp_h2c_data_hdr_handle(ttransport, tqpair, pdu); 2198 break; 2199 2200 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2201 nvmf_tcp_h2c_term_req_hdr_handle(tqpair, pdu); 2202 break; 2203 2204 default: 2205 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", tqpair->pdu_in_progress->hdr.common.pdu_type); 2206 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2207 error_offset = 1; 2208 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2209 break; 2210 } 2211 } 2212 2213 static void 2214 nvmf_tcp_pdu_ch_handle(struct spdk_nvmf_tcp_qpair *tqpair) 2215 { 2216 struct nvme_tcp_pdu *pdu; 2217 uint32_t error_offset = 0; 2218 enum spdk_nvme_tcp_term_req_fes fes; 2219 uint8_t expected_hlen, pdo; 2220 bool plen_error = false, pdo_error = false; 2221 2222 assert(tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2223 pdu = tqpair->pdu_in_progress; 2224 assert(pdu); 2225 if (pdu->hdr.common.pdu_type == SPDK_NVME_TCP_PDU_TYPE_IC_REQ) { 2226 if (tqpair->state != NVME_TCP_QPAIR_STATE_INVALID) { 2227 SPDK_ERRLOG("Already received ICreq PDU, and reject this pdu=%p\n", pdu); 2228 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2229 goto err; 2230 } 2231 expected_hlen = sizeof(struct spdk_nvme_tcp_ic_req); 2232 if (pdu->hdr.common.plen != expected_hlen) { 2233 plen_error = true; 2234 } 2235 } else { 2236 if (tqpair->state != NVME_TCP_QPAIR_STATE_RUNNING) { 2237 SPDK_ERRLOG("The TCP/IP connection is not negotiated\n"); 2238 fes = SPDK_NVME_TCP_TERM_REQ_FES_PDU_SEQUENCE_ERROR; 2239 goto err; 2240 } 2241 2242 switch (pdu->hdr.common.pdu_type) { 2243 case SPDK_NVME_TCP_PDU_TYPE_CAPSULE_CMD: 2244 expected_hlen = sizeof(struct spdk_nvme_tcp_cmd); 2245 pdo = pdu->hdr.common.pdo; 2246 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2247 pdo_error = true; 2248 break; 2249 } 2250 2251 if (pdu->hdr.common.plen < expected_hlen) { 2252 plen_error = true; 2253 } 2254 break; 2255 case SPDK_NVME_TCP_PDU_TYPE_H2C_DATA: 2256 expected_hlen = sizeof(struct spdk_nvme_tcp_h2c_data_hdr); 2257 pdo = pdu->hdr.common.pdo; 2258 if ((tqpair->cpda != 0) && (pdo % ((tqpair->cpda + 1) << 2) != 0)) { 2259 pdo_error = true; 2260 break; 2261 } 2262 if (pdu->hdr.common.plen < expected_hlen) { 2263 plen_error = true; 2264 } 2265 break; 2266 2267 case SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ: 2268 expected_hlen = sizeof(struct spdk_nvme_tcp_term_req_hdr); 2269 if ((pdu->hdr.common.plen <= expected_hlen) || 2270 (pdu->hdr.common.plen > SPDK_NVME_TCP_TERM_REQ_PDU_MAX_SIZE)) { 2271 plen_error = true; 2272 } 2273 break; 2274 2275 default: 2276 SPDK_ERRLOG("Unexpected PDU type 0x%02x\n", pdu->hdr.common.pdu_type); 2277 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2278 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdu_type); 2279 goto err; 2280 } 2281 } 2282 2283 if (pdu->hdr.common.hlen != expected_hlen) { 2284 SPDK_ERRLOG("PDU type=0x%02x, Expected ICReq header length %u, got %u on tqpair=%p\n", 2285 pdu->hdr.common.pdu_type, 2286 expected_hlen, pdu->hdr.common.hlen, tqpair); 2287 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2288 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, hlen); 2289 goto err; 2290 } else if (pdo_error) { 2291 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2292 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, pdo); 2293 } else if (plen_error) { 2294 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2295 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2296 goto err; 2297 } else { 2298 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH); 2299 nvme_tcp_pdu_calc_psh_len(tqpair->pdu_in_progress, tqpair->host_hdgst_enable); 2300 return; 2301 } 2302 err: 2303 nvmf_tcp_send_c2h_term_req(tqpair, pdu, fes, error_offset); 2304 } 2305 2306 static int 2307 nvmf_tcp_sock_process(struct spdk_nvmf_tcp_qpair *tqpair) 2308 { 2309 int rc = 0; 2310 struct nvme_tcp_pdu *pdu; 2311 enum nvme_tcp_pdu_recv_state prev_state; 2312 uint32_t data_len; 2313 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 2314 struct spdk_nvmf_tcp_transport, transport); 2315 2316 /* The loop here is to allow for several back-to-back state changes. */ 2317 do { 2318 prev_state = tqpair->recv_state; 2319 SPDK_DEBUGLOG(nvmf_tcp, "tqpair(%p) recv pdu entering state %d\n", tqpair, prev_state); 2320 2321 pdu = tqpair->pdu_in_progress; 2322 assert(pdu != NULL || 2323 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY || 2324 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_QUIESCING || 2325 tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_ERROR); 2326 2327 switch (tqpair->recv_state) { 2328 /* Wait for the common header */ 2329 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY: 2330 if (!pdu) { 2331 pdu = SLIST_FIRST(&tqpair->tcp_pdu_free_queue); 2332 if (spdk_unlikely(!pdu)) { 2333 return NVME_TCP_PDU_IN_PROGRESS; 2334 } 2335 SLIST_REMOVE_HEAD(&tqpair->tcp_pdu_free_queue, slist); 2336 tqpair->pdu_in_progress = pdu; 2337 tqpair->tcp_pdu_working_count++; 2338 } 2339 memset(pdu, 0, offsetof(struct nvme_tcp_pdu, qpair)); 2340 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH); 2341 /* FALLTHROUGH */ 2342 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_CH: 2343 if (spdk_unlikely(tqpair->state == NVME_TCP_QPAIR_STATE_INITIALIZING)) { 2344 return rc; 2345 } 2346 2347 rc = nvme_tcp_read_data(tqpair->sock, 2348 sizeof(struct spdk_nvme_tcp_common_pdu_hdr) - pdu->ch_valid_bytes, 2349 (void *)&pdu->hdr.common + pdu->ch_valid_bytes); 2350 if (rc < 0) { 2351 SPDK_DEBUGLOG(nvmf_tcp, "will disconnect tqpair=%p\n", tqpair); 2352 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2353 break; 2354 } else if (rc > 0) { 2355 pdu->ch_valid_bytes += rc; 2356 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.qid, rc, 0, tqpair); 2357 } 2358 2359 if (pdu->ch_valid_bytes < sizeof(struct spdk_nvme_tcp_common_pdu_hdr)) { 2360 return NVME_TCP_PDU_IN_PROGRESS; 2361 } 2362 2363 /* The command header of this PDU has now been read from the socket. */ 2364 nvmf_tcp_pdu_ch_handle(tqpair); 2365 break; 2366 /* Wait for the pdu specific header */ 2367 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PSH: 2368 rc = nvme_tcp_read_data(tqpair->sock, 2369 pdu->psh_len - pdu->psh_valid_bytes, 2370 (void *)&pdu->hdr.raw + sizeof(struct spdk_nvme_tcp_common_pdu_hdr) + pdu->psh_valid_bytes); 2371 if (rc < 0) { 2372 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2373 break; 2374 } else if (rc > 0) { 2375 spdk_trace_record(TRACE_TCP_READ_FROM_SOCKET_DONE, tqpair->qpair.qid, rc, 0, tqpair); 2376 pdu->psh_valid_bytes += rc; 2377 } 2378 2379 if (pdu->psh_valid_bytes < pdu->psh_len) { 2380 return NVME_TCP_PDU_IN_PROGRESS; 2381 } 2382 2383 /* All header(ch, psh, head digist) of this PDU has now been read from the socket. */ 2384 nvmf_tcp_pdu_psh_handle(tqpair, ttransport); 2385 break; 2386 /* Wait for the req slot */ 2387 case NVME_TCP_PDU_RECV_STATE_AWAIT_REQ: 2388 nvmf_tcp_capsule_cmd_hdr_handle(ttransport, tqpair, pdu); 2389 break; 2390 case NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD: 2391 /* check whether the data is valid, if not we just return */ 2392 if (!pdu->data_len) { 2393 return NVME_TCP_PDU_IN_PROGRESS; 2394 } 2395 2396 data_len = pdu->data_len; 2397 /* data digest */ 2398 if (spdk_unlikely((pdu->hdr.common.pdu_type != SPDK_NVME_TCP_PDU_TYPE_H2C_TERM_REQ) && 2399 tqpair->host_ddgst_enable)) { 2400 data_len += SPDK_NVME_TCP_DIGEST_LEN; 2401 pdu->ddgst_enable = true; 2402 } 2403 2404 rc = nvme_tcp_read_payload_data(tqpair->sock, pdu); 2405 if (rc < 0) { 2406 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2407 break; 2408 } 2409 pdu->rw_offset += rc; 2410 2411 if (pdu->rw_offset < data_len) { 2412 return NVME_TCP_PDU_IN_PROGRESS; 2413 } 2414 2415 /* Generate and insert DIF to whole data block received if DIF is enabled */ 2416 if (spdk_unlikely(pdu->dif_ctx != NULL) && 2417 spdk_dif_generate_stream(pdu->data_iov, pdu->data_iovcnt, 0, data_len, 2418 pdu->dif_ctx) != 0) { 2419 SPDK_ERRLOG("DIF generate failed\n"); 2420 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 2421 break; 2422 } 2423 2424 /* All of this PDU has now been read from the socket. */ 2425 nvmf_tcp_pdu_payload_handle(tqpair, pdu); 2426 break; 2427 case NVME_TCP_PDU_RECV_STATE_QUIESCING: 2428 if (tqpair->tcp_pdu_working_count != 0) { 2429 return NVME_TCP_PDU_IN_PROGRESS; 2430 } 2431 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_ERROR); 2432 break; 2433 case NVME_TCP_PDU_RECV_STATE_ERROR: 2434 if (spdk_sock_is_connected(tqpair->sock) && tqpair->wait_terminate) { 2435 return NVME_TCP_PDU_IN_PROGRESS; 2436 } 2437 return NVME_TCP_PDU_FATAL; 2438 default: 2439 SPDK_ERRLOG("The state(%d) is invalid\n", tqpair->recv_state); 2440 abort(); 2441 break; 2442 } 2443 } while (tqpair->recv_state != prev_state); 2444 2445 return rc; 2446 } 2447 2448 static inline void * 2449 nvmf_tcp_control_msg_get(struct spdk_nvmf_tcp_control_msg_list *list) 2450 { 2451 struct spdk_nvmf_tcp_control_msg *msg; 2452 2453 assert(list); 2454 2455 msg = STAILQ_FIRST(&list->free_msgs); 2456 if (!msg) { 2457 SPDK_DEBUGLOG(nvmf_tcp, "Out of control messages\n"); 2458 return NULL; 2459 } 2460 STAILQ_REMOVE_HEAD(&list->free_msgs, link); 2461 return msg; 2462 } 2463 2464 static inline void 2465 nvmf_tcp_control_msg_put(struct spdk_nvmf_tcp_control_msg_list *list, void *_msg) 2466 { 2467 struct spdk_nvmf_tcp_control_msg *msg = _msg; 2468 2469 assert(list); 2470 STAILQ_INSERT_HEAD(&list->free_msgs, msg, link); 2471 } 2472 2473 static int 2474 nvmf_tcp_req_parse_sgl(struct spdk_nvmf_tcp_req *tcp_req, 2475 struct spdk_nvmf_transport *transport, 2476 struct spdk_nvmf_transport_poll_group *group) 2477 { 2478 struct spdk_nvmf_request *req = &tcp_req->req; 2479 struct spdk_nvme_cmd *cmd; 2480 struct spdk_nvme_sgl_descriptor *sgl; 2481 struct spdk_nvmf_tcp_poll_group *tgroup; 2482 enum spdk_nvme_tcp_term_req_fes fes; 2483 struct nvme_tcp_pdu *pdu; 2484 struct spdk_nvmf_tcp_qpair *tqpair; 2485 uint32_t length, error_offset = 0; 2486 2487 cmd = &req->cmd->nvme_cmd; 2488 sgl = &cmd->dptr.sgl1; 2489 2490 if (sgl->generic.type == SPDK_NVME_SGL_TYPE_TRANSPORT_DATA_BLOCK && 2491 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_TRANSPORT) { 2492 /* get request length from sgl */ 2493 length = sgl->unkeyed.length; 2494 if (spdk_unlikely(length > transport->opts.max_io_size)) { 2495 SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n", 2496 length, transport->opts.max_io_size); 2497 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_LIMIT_EXCEEDED; 2498 goto fatal_err; 2499 } 2500 2501 /* fill request length and populate iovs */ 2502 req->length = length; 2503 2504 SPDK_DEBUGLOG(nvmf_tcp, "Data requested length= 0x%x\n", length); 2505 2506 if (spdk_unlikely(req->dif_enabled)) { 2507 req->dif.orig_length = length; 2508 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2509 req->dif.elba_length = length; 2510 } 2511 2512 if (nvmf_ctrlr_use_zcopy(req)) { 2513 SPDK_DEBUGLOG(nvmf_tcp, "Using zero-copy to execute request %p\n", tcp_req); 2514 req->data_from_pool = false; 2515 return 0; 2516 } 2517 2518 if (spdk_nvmf_request_get_buffers(req, group, transport, length)) { 2519 /* No available buffers. Queue this request up. */ 2520 SPDK_DEBUGLOG(nvmf_tcp, "No available large data buffers. Queueing request %p\n", 2521 tcp_req); 2522 return 0; 2523 } 2524 2525 SPDK_DEBUGLOG(nvmf_tcp, "Request %p took %d buffer/s from central pool, and data=%p\n", 2526 tcp_req, req->iovcnt, req->iov[0].iov_base); 2527 2528 return 0; 2529 } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK && 2530 sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) { 2531 uint64_t offset = sgl->address; 2532 uint32_t max_len = transport->opts.in_capsule_data_size; 2533 2534 assert(tcp_req->has_in_capsule_data); 2535 /* Capsule Cmd with In-capsule Data should get data length from pdu header */ 2536 tqpair = tcp_req->pdu->qpair; 2537 /* receiving pdu is not same with the pdu in tcp_req */ 2538 pdu = tqpair->pdu_in_progress; 2539 length = pdu->hdr.common.plen - pdu->psh_len - sizeof(struct spdk_nvme_tcp_common_pdu_hdr); 2540 if (tqpair->host_ddgst_enable) { 2541 length -= SPDK_NVME_TCP_DIGEST_LEN; 2542 } 2543 /* This error is not defined in NVMe/TCP spec, take this error as fatal error */ 2544 if (spdk_unlikely(length != sgl->unkeyed.length)) { 2545 SPDK_ERRLOG("In-Capsule Data length 0x%x is not equal to SGL data length 0x%x\n", 2546 length, sgl->unkeyed.length); 2547 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_HEADER_FIELD; 2548 error_offset = offsetof(struct spdk_nvme_tcp_common_pdu_hdr, plen); 2549 goto fatal_err; 2550 } 2551 2552 SPDK_DEBUGLOG(nvmf_tcp, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n", 2553 offset, length); 2554 2555 /* The NVMe/TCP transport does not use ICDOFF to control the in-capsule data offset. ICDOFF should be '0' */ 2556 if (spdk_unlikely(offset != 0)) { 2557 /* Not defined fatal error in NVMe/TCP spec, handle this error as a fatal error */ 2558 SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " should be ZERO in NVMe/TCP\n", offset); 2559 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER; 2560 error_offset = offsetof(struct spdk_nvme_tcp_cmd, ccsqe.dptr.sgl1.address); 2561 goto fatal_err; 2562 } 2563 2564 if (spdk_unlikely(length > max_len)) { 2565 /* According to the SPEC we should support ICD up to 8192 bytes for admin and fabric commands */ 2566 if (length <= SPDK_NVME_TCP_IN_CAPSULE_DATA_MAX_SIZE && 2567 (cmd->opc == SPDK_NVME_OPC_FABRIC || req->qpair->qid == 0)) { 2568 2569 /* Get a buffer from dedicated list */ 2570 SPDK_DEBUGLOG(nvmf_tcp, "Getting a buffer from control msg list\n"); 2571 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 2572 assert(tgroup->control_msg_list); 2573 req->iov[0].iov_base = nvmf_tcp_control_msg_get(tgroup->control_msg_list); 2574 if (!req->iov[0].iov_base) { 2575 /* No available buffers. Queue this request up. */ 2576 SPDK_DEBUGLOG(nvmf_tcp, "No available ICD buffers. Queueing request %p\n", tcp_req); 2577 return 0; 2578 } 2579 } else { 2580 SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n", 2581 length, max_len); 2582 fes = SPDK_NVME_TCP_TERM_REQ_FES_DATA_TRANSFER_LIMIT_EXCEEDED; 2583 goto fatal_err; 2584 } 2585 } else { 2586 req->iov[0].iov_base = tcp_req->buf; 2587 } 2588 2589 req->length = length; 2590 req->data_from_pool = false; 2591 2592 if (spdk_unlikely(req->dif_enabled)) { 2593 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx); 2594 req->dif.elba_length = length; 2595 } 2596 2597 req->iov[0].iov_len = length; 2598 req->iovcnt = 1; 2599 2600 return 0; 2601 } 2602 /* If we want to handle the problem here, then we can't skip the following data segment. 2603 * Because this function runs before reading data part, now handle all errors as fatal errors. */ 2604 SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n", 2605 sgl->generic.type, sgl->generic.subtype); 2606 fes = SPDK_NVME_TCP_TERM_REQ_FES_INVALID_DATA_UNSUPPORTED_PARAMETER; 2607 error_offset = offsetof(struct spdk_nvme_tcp_cmd, ccsqe.dptr.sgl1.generic); 2608 fatal_err: 2609 nvmf_tcp_send_c2h_term_req(tcp_req->pdu->qpair, tcp_req->pdu, fes, error_offset); 2610 return -1; 2611 } 2612 2613 static inline enum spdk_nvme_media_error_status_code 2614 nvmf_tcp_dif_error_to_compl_status(uint8_t err_type) { 2615 enum spdk_nvme_media_error_status_code result; 2616 2617 switch (err_type) 2618 { 2619 case SPDK_DIF_REFTAG_ERROR: 2620 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR; 2621 break; 2622 case SPDK_DIF_APPTAG_ERROR: 2623 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR; 2624 break; 2625 case SPDK_DIF_GUARD_ERROR: 2626 result = SPDK_NVME_SC_GUARD_CHECK_ERROR; 2627 break; 2628 default: 2629 SPDK_UNREACHABLE(); 2630 break; 2631 } 2632 2633 return result; 2634 } 2635 2636 static void 2637 _nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2638 struct spdk_nvmf_tcp_req *tcp_req) 2639 { 2640 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF( 2641 tqpair->qpair.transport, struct spdk_nvmf_tcp_transport, transport); 2642 struct nvme_tcp_pdu *rsp_pdu; 2643 struct spdk_nvme_tcp_c2h_data_hdr *c2h_data; 2644 uint32_t plen, pdo, alignment; 2645 int rc; 2646 2647 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2648 2649 rsp_pdu = tcp_req->pdu; 2650 assert(rsp_pdu != NULL); 2651 2652 c2h_data = &rsp_pdu->hdr.c2h_data; 2653 c2h_data->common.pdu_type = SPDK_NVME_TCP_PDU_TYPE_C2H_DATA; 2654 plen = c2h_data->common.hlen = sizeof(*c2h_data); 2655 2656 if (tqpair->host_hdgst_enable) { 2657 plen += SPDK_NVME_TCP_DIGEST_LEN; 2658 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_HDGSTF; 2659 } 2660 2661 /* set the psh */ 2662 c2h_data->cccid = tcp_req->req.cmd->nvme_cmd.cid; 2663 c2h_data->datal = tcp_req->req.length - tcp_req->pdu->rw_offset; 2664 c2h_data->datao = tcp_req->pdu->rw_offset; 2665 2666 /* set the padding */ 2667 rsp_pdu->padding_len = 0; 2668 pdo = plen; 2669 if (tqpair->cpda) { 2670 alignment = (tqpair->cpda + 1) << 2; 2671 if (plen % alignment != 0) { 2672 pdo = (plen + alignment) / alignment * alignment; 2673 rsp_pdu->padding_len = pdo - plen; 2674 plen = pdo; 2675 } 2676 } 2677 2678 c2h_data->common.pdo = pdo; 2679 plen += c2h_data->datal; 2680 if (tqpair->host_ddgst_enable) { 2681 c2h_data->common.flags |= SPDK_NVME_TCP_CH_FLAGS_DDGSTF; 2682 plen += SPDK_NVME_TCP_DIGEST_LEN; 2683 } 2684 2685 c2h_data->common.plen = plen; 2686 2687 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2688 rsp_pdu->dif_ctx = &tcp_req->req.dif.dif_ctx; 2689 } 2690 2691 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2692 c2h_data->datao, c2h_data->datal); 2693 2694 2695 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU; 2696 /* Need to send the capsule response if response is not all 0 */ 2697 if (ttransport->tcp_opts.c2h_success && 2698 tcp_req->rsp.cdw0 == 0 && tcp_req->rsp.cdw1 == 0) { 2699 c2h_data->common.flags |= SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS; 2700 } 2701 2702 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2703 struct spdk_nvme_cpl *rsp = &tcp_req->req.rsp->nvme_cpl; 2704 struct spdk_dif_error err_blk = {}; 2705 uint32_t mapped_length = 0; 2706 uint32_t available_iovs = SPDK_COUNTOF(rsp_pdu->iov); 2707 uint32_t ddgst_len = 0; 2708 2709 if (tqpair->host_ddgst_enable) { 2710 /* Data digest consumes additional iov entry */ 2711 available_iovs--; 2712 /* plen needs to be updated since nvme_tcp_build_iovs compares expected and actual plen */ 2713 ddgst_len = SPDK_NVME_TCP_DIGEST_LEN; 2714 c2h_data->common.plen -= ddgst_len; 2715 } 2716 /* Temp call to estimate if data can be described by limited number of iovs. 2717 * iov vector will be rebuilt in nvmf_tcp_qpair_write_pdu */ 2718 nvme_tcp_build_iovs(rsp_pdu->iov, available_iovs, rsp_pdu, tqpair->host_hdgst_enable, 2719 false, &mapped_length); 2720 2721 if (mapped_length != c2h_data->common.plen) { 2722 c2h_data->datal = mapped_length - (c2h_data->common.plen - c2h_data->datal); 2723 SPDK_DEBUGLOG(nvmf_tcp, 2724 "Part C2H, data_len %u (of %u), PDU len %u, updated PDU len %u, offset %u\n", 2725 c2h_data->datal, tcp_req->req.length, c2h_data->common.plen, mapped_length, rsp_pdu->rw_offset); 2726 c2h_data->common.plen = mapped_length; 2727 2728 /* Rebuild pdu->data_iov since data length is changed */ 2729 nvme_tcp_pdu_set_data_buf(rsp_pdu, tcp_req->req.iov, tcp_req->req.iovcnt, c2h_data->datao, 2730 c2h_data->datal); 2731 2732 c2h_data->common.flags &= ~(SPDK_NVME_TCP_C2H_DATA_FLAGS_LAST_PDU | 2733 SPDK_NVME_TCP_C2H_DATA_FLAGS_SUCCESS); 2734 } 2735 2736 c2h_data->common.plen += ddgst_len; 2737 2738 assert(rsp_pdu->rw_offset <= tcp_req->req.length); 2739 2740 rc = spdk_dif_verify_stream(rsp_pdu->data_iov, rsp_pdu->data_iovcnt, 2741 0, rsp_pdu->data_len, rsp_pdu->dif_ctx, &err_blk); 2742 if (rc != 0) { 2743 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", 2744 err_blk.err_type, err_blk.err_offset); 2745 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR; 2746 rsp->status.sc = nvmf_tcp_dif_error_to_compl_status(err_blk.err_type); 2747 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2748 return; 2749 } 2750 } 2751 2752 rsp_pdu->rw_offset += c2h_data->datal; 2753 nvmf_tcp_qpair_write_req_pdu(tqpair, tcp_req, nvmf_tcp_pdu_c2h_data_complete, tcp_req); 2754 } 2755 2756 static void 2757 nvmf_tcp_send_c2h_data(struct spdk_nvmf_tcp_qpair *tqpair, 2758 struct spdk_nvmf_tcp_req *tcp_req) 2759 { 2760 nvmf_tcp_req_pdu_init(tcp_req); 2761 _nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2762 } 2763 2764 static int 2765 request_transfer_out(struct spdk_nvmf_request *req) 2766 { 2767 struct spdk_nvmf_tcp_req *tcp_req; 2768 struct spdk_nvmf_qpair *qpair; 2769 struct spdk_nvmf_tcp_qpair *tqpair; 2770 struct spdk_nvme_cpl *rsp; 2771 2772 SPDK_DEBUGLOG(nvmf_tcp, "enter\n"); 2773 2774 qpair = req->qpair; 2775 rsp = &req->rsp->nvme_cpl; 2776 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 2777 2778 /* Advance our sq_head pointer */ 2779 if (qpair->sq_head == qpair->sq_head_max) { 2780 qpair->sq_head = 0; 2781 } else { 2782 qpair->sq_head++; 2783 } 2784 rsp->sqhd = qpair->sq_head; 2785 2786 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2787 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 2788 if (rsp->status.sc == SPDK_NVME_SC_SUCCESS && req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) { 2789 nvmf_tcp_send_c2h_data(tqpair, tcp_req); 2790 } else { 2791 nvmf_tcp_send_capsule_resp_pdu(tcp_req, tqpair); 2792 } 2793 2794 return 0; 2795 } 2796 2797 static void 2798 nvmf_tcp_check_fused_ordering(struct spdk_nvmf_tcp_transport *ttransport, 2799 struct spdk_nvmf_tcp_qpair *tqpair, 2800 struct spdk_nvmf_tcp_req *tcp_req) 2801 { 2802 enum spdk_nvme_cmd_fuse last, next; 2803 2804 last = tqpair->fused_first ? tqpair->fused_first->cmd.fuse : SPDK_NVME_CMD_FUSE_NONE; 2805 next = tcp_req->cmd.fuse; 2806 2807 assert(last != SPDK_NVME_CMD_FUSE_SECOND); 2808 2809 if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) { 2810 return; 2811 } 2812 2813 if (last == SPDK_NVME_CMD_FUSE_FIRST) { 2814 if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2815 /* This is a valid pair of fused commands. Point them at each other 2816 * so they can be submitted consecutively once ready to be executed. 2817 */ 2818 tqpair->fused_first->fused_pair = tcp_req; 2819 tcp_req->fused_pair = tqpair->fused_first; 2820 tqpair->fused_first = NULL; 2821 return; 2822 } else { 2823 /* Mark the last req as failed since it wasn't followed by a SECOND. */ 2824 tqpair->fused_first->fused_failed = true; 2825 2826 /* 2827 * If the last req is in READY_TO_EXECUTE state, then call 2828 * nvmf_tcp_req_process(), otherwise nothing else will kick it. 2829 */ 2830 if (tqpair->fused_first->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 2831 nvmf_tcp_req_process(ttransport, tqpair->fused_first); 2832 } 2833 2834 tqpair->fused_first = NULL; 2835 } 2836 } 2837 2838 if (next == SPDK_NVME_CMD_FUSE_FIRST) { 2839 /* Set tqpair->fused_first here so that we know to check that the next request 2840 * is a SECOND (and to fail this one if it isn't). 2841 */ 2842 tqpair->fused_first = tcp_req; 2843 } else if (next == SPDK_NVME_CMD_FUSE_SECOND) { 2844 /* Mark this req failed since it is a SECOND and the last one was not a FIRST. */ 2845 tcp_req->fused_failed = true; 2846 } 2847 } 2848 2849 static bool 2850 nvmf_tcp_req_process(struct spdk_nvmf_tcp_transport *ttransport, 2851 struct spdk_nvmf_tcp_req *tcp_req) 2852 { 2853 struct spdk_nvmf_tcp_qpair *tqpair; 2854 uint32_t plen; 2855 struct nvme_tcp_pdu *pdu; 2856 enum spdk_nvmf_tcp_req_state prev_state; 2857 bool progress = false; 2858 struct spdk_nvmf_transport *transport = &ttransport->transport; 2859 struct spdk_nvmf_transport_poll_group *group; 2860 struct spdk_nvmf_tcp_poll_group *tgroup; 2861 2862 tqpair = SPDK_CONTAINEROF(tcp_req->req.qpair, struct spdk_nvmf_tcp_qpair, qpair); 2863 group = &tqpair->group->group; 2864 assert(tcp_req->state != TCP_REQUEST_STATE_FREE); 2865 2866 /* If the qpair is not active, we need to abort the outstanding requests. */ 2867 if (tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) { 2868 if (tcp_req->state == TCP_REQUEST_STATE_NEED_BUFFER) { 2869 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2870 } 2871 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 2872 } 2873 2874 /* The loop here is to allow for several back-to-back state changes. */ 2875 do { 2876 prev_state = tcp_req->state; 2877 2878 SPDK_DEBUGLOG(nvmf_tcp, "Request %p entering state %d on tqpair=%p\n", tcp_req, prev_state, 2879 tqpair); 2880 2881 switch (tcp_req->state) { 2882 case TCP_REQUEST_STATE_FREE: 2883 /* Some external code must kick a request into TCP_REQUEST_STATE_NEW 2884 * to escape this state. */ 2885 break; 2886 case TCP_REQUEST_STATE_NEW: 2887 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEW, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, tqpair); 2888 2889 /* copy the cmd from the receive pdu */ 2890 tcp_req->cmd = tqpair->pdu_in_progress->hdr.capsule_cmd.ccsqe; 2891 2892 if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&tcp_req->req, &tcp_req->req.dif.dif_ctx))) { 2893 tcp_req->req.dif_enabled = true; 2894 tqpair->pdu_in_progress->dif_ctx = &tcp_req->req.dif.dif_ctx; 2895 } 2896 2897 nvmf_tcp_check_fused_ordering(ttransport, tqpair, tcp_req); 2898 2899 /* The next state transition depends on the data transfer needs of this request. */ 2900 tcp_req->req.xfer = spdk_nvmf_req_get_xfer(&tcp_req->req); 2901 2902 if (spdk_unlikely(tcp_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) { 2903 nvmf_tcp_req_set_cpl(tcp_req, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_INVALID_OPCODE); 2904 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2905 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 2906 SPDK_DEBUGLOG(nvmf_tcp, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", tcp_req); 2907 break; 2908 } 2909 2910 /* If no data to transfer, ready to execute. */ 2911 if (tcp_req->req.xfer == SPDK_NVME_DATA_NONE) { 2912 /* Reset the tqpair receiving pdu state */ 2913 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2914 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2915 break; 2916 } 2917 2918 pdu = tqpair->pdu_in_progress; 2919 plen = pdu->hdr.common.hlen; 2920 if (tqpair->host_hdgst_enable) { 2921 plen += SPDK_NVME_TCP_DIGEST_LEN; 2922 } 2923 if (pdu->hdr.common.plen != plen) { 2924 tcp_req->has_in_capsule_data = true; 2925 } else { 2926 /* Data is transmitted by C2H PDUs */ 2927 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_READY); 2928 } 2929 2930 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_NEED_BUFFER); 2931 STAILQ_INSERT_TAIL(&group->pending_buf_queue, &tcp_req->req, buf_link); 2932 break; 2933 case TCP_REQUEST_STATE_NEED_BUFFER: 2934 spdk_trace_record(TRACE_TCP_REQUEST_STATE_NEED_BUFFER, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 2935 tqpair); 2936 2937 assert(tcp_req->req.xfer != SPDK_NVME_DATA_NONE); 2938 2939 if (!tcp_req->has_in_capsule_data && (&tcp_req->req != STAILQ_FIRST(&group->pending_buf_queue))) { 2940 SPDK_DEBUGLOG(nvmf_tcp, 2941 "Not the first element to wait for the buf for tcp_req(%p) on tqpair=%p\n", 2942 tcp_req, tqpair); 2943 /* This request needs to wait in line to obtain a buffer */ 2944 break; 2945 } 2946 2947 /* Try to get a data buffer */ 2948 if (nvmf_tcp_req_parse_sgl(tcp_req, transport, group) < 0) { 2949 break; 2950 } 2951 2952 /* Get a zcopy buffer if the request can be serviced through zcopy */ 2953 if (spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 2954 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 2955 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 2956 tcp_req->req.length = tcp_req->req.dif.elba_length; 2957 } 2958 2959 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2960 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_START); 2961 spdk_nvmf_request_zcopy_start(&tcp_req->req); 2962 break; 2963 } 2964 2965 if (tcp_req->req.iovcnt < 1) { 2966 SPDK_DEBUGLOG(nvmf_tcp, "No buffer allocated for tcp_req(%p) on tqpair(%p\n)", 2967 tcp_req, tqpair); 2968 /* No buffers available. */ 2969 break; 2970 } 2971 2972 STAILQ_REMOVE(&group->pending_buf_queue, &tcp_req->req, spdk_nvmf_request, buf_link); 2973 2974 /* If data is transferring from host to controller, we need to do a transfer from the host. */ 2975 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 2976 if (tcp_req->req.data_from_pool) { 2977 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 2978 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 2979 } else { 2980 struct nvme_tcp_pdu *pdu; 2981 2982 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 2983 2984 pdu = tqpair->pdu_in_progress; 2985 SPDK_DEBUGLOG(nvmf_tcp, "Not need to send r2t for tcp_req(%p) on tqpair=%p\n", tcp_req, 2986 tqpair); 2987 /* No need to send r2t, contained in the capsuled data */ 2988 nvme_tcp_pdu_set_data_buf(pdu, tcp_req->req.iov, tcp_req->req.iovcnt, 2989 0, tcp_req->req.length); 2990 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_AWAIT_PDU_PAYLOAD); 2991 } 2992 break; 2993 } 2994 2995 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_EXECUTE); 2996 break; 2997 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 2998 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_START, tqpair->qpair.qid, 0, 2999 (uintptr_t)tcp_req, tqpair); 3000 /* Some external code must kick a request into TCP_REQUEST_STATE_ZCOPY_START_COMPLETED 3001 * to escape this state. */ 3002 break; 3003 case TCP_REQUEST_STATE_ZCOPY_START_COMPLETED: 3004 spdk_trace_record(TRACE_TCP_REQUEST_STATE_ZCOPY_START_COMPLETED, tqpair->qpair.qid, 0, 3005 (uintptr_t)tcp_req, tqpair); 3006 if (spdk_unlikely(spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl))) { 3007 SPDK_DEBUGLOG(nvmf_tcp, "Zero-copy start failed for tcp_req(%p) on tqpair=%p\n", 3008 tcp_req, tqpair); 3009 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3010 break; 3011 } 3012 if (tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) { 3013 SPDK_DEBUGLOG(nvmf_tcp, "Sending R2T for tcp_req(%p) on tqpair=%p\n", tcp_req, tqpair); 3014 nvmf_tcp_send_r2t_pdu(tqpair, tcp_req); 3015 } else { 3016 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3017 } 3018 break; 3019 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3020 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_R2T_ACK, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3021 tqpair); 3022 /* The R2T completion or the h2c data incoming will kick it out of this state. */ 3023 break; 3024 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3025 3026 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, tqpair->qpair.qid, 0, 3027 (uintptr_t)tcp_req, tqpair); 3028 /* Some external code must kick a request into TCP_REQUEST_STATE_READY_TO_EXECUTE 3029 * to escape this state. */ 3030 break; 3031 case TCP_REQUEST_STATE_READY_TO_EXECUTE: 3032 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_EXECUTE, tqpair->qpair.qid, 0, 3033 (uintptr_t)tcp_req, tqpair); 3034 3035 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3036 assert(tcp_req->req.dif.elba_length >= tcp_req->req.length); 3037 tcp_req->req.length = tcp_req->req.dif.elba_length; 3038 } 3039 3040 if (tcp_req->cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) { 3041 if (tcp_req->fused_failed) { 3042 /* This request failed FUSED semantics. Fail it immediately, without 3043 * even sending it to the target layer. 3044 */ 3045 nvmf_tcp_req_set_cpl(tcp_req, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_ABORTED_MISSING_FUSED); 3046 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3047 break; 3048 } 3049 3050 if (tcp_req->fused_pair == NULL || 3051 tcp_req->fused_pair->state != TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3052 /* This request is ready to execute, but either we don't know yet if it's 3053 * valid - i.e. this is a FIRST but we haven't received the next request yet), 3054 * or the other request of this fused pair isn't ready to execute. So 3055 * break here and this request will get processed later either when the 3056 * other request is ready or we find that this request isn't valid. 3057 */ 3058 break; 3059 } 3060 } 3061 3062 if (!spdk_nvmf_request_using_zcopy(&tcp_req->req)) { 3063 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTING); 3064 /* If we get to this point, and this request is a fused command, we know that 3065 * it is part of a valid sequence (FIRST followed by a SECOND) and that both 3066 * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this 3067 * request, and the other request of the fused pair, in the correct order. 3068 * Also clear the ->fused_pair pointers on both requests, since after this point 3069 * we no longer need to maintain the relationship between these two requests. 3070 */ 3071 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) { 3072 assert(tcp_req->fused_pair != NULL); 3073 assert(tcp_req->fused_pair->fused_pair == tcp_req); 3074 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 3075 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 3076 tcp_req->fused_pair->fused_pair = NULL; 3077 tcp_req->fused_pair = NULL; 3078 } 3079 spdk_nvmf_request_exec(&tcp_req->req); 3080 if (tcp_req->cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) { 3081 assert(tcp_req->fused_pair != NULL); 3082 assert(tcp_req->fused_pair->fused_pair == tcp_req); 3083 nvmf_tcp_req_set_state(tcp_req->fused_pair, TCP_REQUEST_STATE_EXECUTING); 3084 spdk_nvmf_request_exec(&tcp_req->fused_pair->req); 3085 tcp_req->fused_pair->fused_pair = NULL; 3086 tcp_req->fused_pair = NULL; 3087 } 3088 } else { 3089 /* For zero-copy, only requests with data coming from host to the 3090 * controller can end up here. */ 3091 assert(tcp_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 3092 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT); 3093 spdk_nvmf_request_zcopy_end(&tcp_req->req, true); 3094 } 3095 3096 break; 3097 case TCP_REQUEST_STATE_EXECUTING: 3098 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTING, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3099 tqpair); 3100 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 3101 * to escape this state. */ 3102 break; 3103 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3104 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_COMMIT, tqpair->qpair.qid, 0, 3105 (uintptr_t)tcp_req, tqpair); 3106 /* Some external code must kick a request into TCP_REQUEST_STATE_EXECUTED 3107 * to escape this state. */ 3108 break; 3109 case TCP_REQUEST_STATE_EXECUTED: 3110 spdk_trace_record(TRACE_TCP_REQUEST_STATE_EXECUTED, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3111 tqpair); 3112 3113 if (spdk_unlikely(tcp_req->req.dif_enabled)) { 3114 tcp_req->req.length = tcp_req->req.dif.orig_length; 3115 } 3116 3117 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3118 break; 3119 case TCP_REQUEST_STATE_READY_TO_COMPLETE: 3120 spdk_trace_record(TRACE_TCP_REQUEST_STATE_READY_TO_COMPLETE, tqpair->qpair.qid, 0, 3121 (uintptr_t)tcp_req, tqpair); 3122 if (request_transfer_out(&tcp_req->req) != 0) { 3123 assert(0); /* No good way to handle this currently */ 3124 } 3125 break; 3126 case TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST: 3127 spdk_trace_record(TRACE_TCP_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, tqpair->qpair.qid, 0, 3128 (uintptr_t)tcp_req, tqpair); 3129 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 3130 * to escape this state. */ 3131 break; 3132 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3133 spdk_trace_record(TRACE_TCP_REQUEST_STATE_AWAIT_ZCOPY_RELEASE, tqpair->qpair.qid, 0, 3134 (uintptr_t)tcp_req, tqpair); 3135 /* Some external code must kick a request into TCP_REQUEST_STATE_COMPLETED 3136 * to escape this state. */ 3137 break; 3138 case TCP_REQUEST_STATE_COMPLETED: 3139 spdk_trace_record(TRACE_TCP_REQUEST_STATE_COMPLETED, tqpair->qpair.qid, 0, (uintptr_t)tcp_req, 3140 tqpair); 3141 /* If there's an outstanding PDU sent to the host, the request is completed 3142 * due to the qpair being disconnected. We must delay the completion until 3143 * that write is done to avoid freeing the request twice. */ 3144 if (spdk_unlikely(tcp_req->pdu_in_use)) { 3145 SPDK_DEBUGLOG(nvmf_tcp, "Delaying completion due to outstanding " 3146 "write on req=%p\n", tcp_req); 3147 /* This can only happen for zcopy requests */ 3148 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 3149 assert(tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE); 3150 break; 3151 } 3152 3153 if (tcp_req->req.data_from_pool) { 3154 spdk_nvmf_request_free_buffers(&tcp_req->req, group, transport); 3155 } else if (spdk_unlikely(tcp_req->has_in_capsule_data && 3156 (tcp_req->cmd.opc == SPDK_NVME_OPC_FABRIC || 3157 tqpair->qpair.qid == 0) && tcp_req->req.length > transport->opts.in_capsule_data_size)) { 3158 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3159 assert(tgroup->control_msg_list); 3160 SPDK_DEBUGLOG(nvmf_tcp, "Put buf to control msg list\n"); 3161 nvmf_tcp_control_msg_put(tgroup->control_msg_list, 3162 tcp_req->req.iov[0].iov_base); 3163 } else if (tcp_req->req.zcopy_bdev_io != NULL) { 3164 /* If the request has an unreleased zcopy bdev_io, it's either a 3165 * read, a failed write, or the qpair is being disconnected */ 3166 assert(spdk_nvmf_request_using_zcopy(&tcp_req->req)); 3167 assert(tcp_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST || 3168 spdk_nvme_cpl_is_error(&tcp_req->req.rsp->nvme_cpl) || 3169 tqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE); 3170 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE); 3171 spdk_nvmf_request_zcopy_end(&tcp_req->req, false); 3172 break; 3173 } 3174 tcp_req->req.length = 0; 3175 tcp_req->req.iovcnt = 0; 3176 tcp_req->fused_failed = false; 3177 if (tcp_req->fused_pair) { 3178 /* This req was part of a valid fused pair, but failed before it got to 3179 * READ_TO_EXECUTE state. This means we need to fail the other request 3180 * in the pair, because it is no longer part of a valid pair. If the pair 3181 * already reached READY_TO_EXECUTE state, we need to kick it. 3182 */ 3183 tcp_req->fused_pair->fused_failed = true; 3184 if (tcp_req->fused_pair->state == TCP_REQUEST_STATE_READY_TO_EXECUTE) { 3185 nvmf_tcp_req_process(ttransport, tcp_req->fused_pair); 3186 } 3187 tcp_req->fused_pair = NULL; 3188 } 3189 3190 nvmf_tcp_req_put(tqpair, tcp_req); 3191 break; 3192 case TCP_REQUEST_NUM_STATES: 3193 default: 3194 assert(0); 3195 break; 3196 } 3197 3198 if (tcp_req->state != prev_state) { 3199 progress = true; 3200 } 3201 } while (tcp_req->state != prev_state); 3202 3203 return progress; 3204 } 3205 3206 static void 3207 nvmf_tcp_sock_cb(void *arg, struct spdk_sock_group *group, struct spdk_sock *sock) 3208 { 3209 struct spdk_nvmf_tcp_qpair *tqpair = arg; 3210 int rc; 3211 3212 assert(tqpair != NULL); 3213 rc = nvmf_tcp_sock_process(tqpair); 3214 3215 /* If there was a new socket error, disconnect */ 3216 if (rc < 0) { 3217 nvmf_tcp_qpair_disconnect(tqpair); 3218 } 3219 } 3220 3221 static int 3222 nvmf_tcp_poll_group_add(struct spdk_nvmf_transport_poll_group *group, 3223 struct spdk_nvmf_qpair *qpair) 3224 { 3225 struct spdk_nvmf_tcp_poll_group *tgroup; 3226 struct spdk_nvmf_tcp_qpair *tqpair; 3227 int rc; 3228 3229 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3230 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3231 3232 rc = nvmf_tcp_qpair_sock_init(tqpair); 3233 if (rc != 0) { 3234 SPDK_ERRLOG("Cannot set sock opt for tqpair=%p\n", tqpair); 3235 return -1; 3236 } 3237 3238 rc = nvmf_tcp_qpair_init(&tqpair->qpair); 3239 if (rc < 0) { 3240 SPDK_ERRLOG("Cannot init tqpair=%p\n", tqpair); 3241 return -1; 3242 } 3243 3244 rc = nvmf_tcp_qpair_init_mem_resource(tqpair); 3245 if (rc < 0) { 3246 SPDK_ERRLOG("Cannot init memory resource info for tqpair=%p\n", tqpair); 3247 return -1; 3248 } 3249 3250 rc = spdk_sock_group_add_sock(tgroup->sock_group, tqpair->sock, 3251 nvmf_tcp_sock_cb, tqpair); 3252 if (rc != 0) { 3253 SPDK_ERRLOG("Could not add sock to sock_group: %s (%d)\n", 3254 spdk_strerror(errno), errno); 3255 return -1; 3256 } 3257 3258 tqpair->group = tgroup; 3259 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_INVALID); 3260 TAILQ_INSERT_TAIL(&tgroup->qpairs, tqpair, link); 3261 3262 return 0; 3263 } 3264 3265 static int 3266 nvmf_tcp_poll_group_remove(struct spdk_nvmf_transport_poll_group *group, 3267 struct spdk_nvmf_qpair *qpair) 3268 { 3269 struct spdk_nvmf_tcp_poll_group *tgroup; 3270 struct spdk_nvmf_tcp_qpair *tqpair; 3271 int rc; 3272 3273 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3274 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3275 3276 assert(tqpair->group == tgroup); 3277 3278 SPDK_DEBUGLOG(nvmf_tcp, "remove tqpair=%p from the tgroup=%p\n", tqpair, tgroup); 3279 if (tqpair->recv_state == NVME_TCP_PDU_RECV_STATE_AWAIT_REQ) { 3280 /* Change the state to move the qpair from the await_req list to the main list 3281 * and prevent adding it again later by nvmf_tcp_qpair_set_recv_state() */ 3282 nvmf_tcp_qpair_set_recv_state(tqpair, NVME_TCP_PDU_RECV_STATE_QUIESCING); 3283 } 3284 TAILQ_REMOVE(&tgroup->qpairs, tqpair, link); 3285 3286 rc = spdk_sock_group_remove_sock(tgroup->sock_group, tqpair->sock); 3287 if (rc != 0) { 3288 SPDK_ERRLOG("Could not remove sock from sock_group: %s (%d)\n", 3289 spdk_strerror(errno), errno); 3290 } 3291 3292 return rc; 3293 } 3294 3295 static int 3296 nvmf_tcp_req_complete(struct spdk_nvmf_request *req) 3297 { 3298 struct spdk_nvmf_tcp_transport *ttransport; 3299 struct spdk_nvmf_tcp_req *tcp_req; 3300 3301 ttransport = SPDK_CONTAINEROF(req->qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3302 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3303 3304 switch (tcp_req->state) { 3305 case TCP_REQUEST_STATE_EXECUTING: 3306 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3307 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_EXECUTED); 3308 break; 3309 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3310 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_ZCOPY_START_COMPLETED); 3311 break; 3312 case TCP_REQUEST_STATE_AWAITING_ZCOPY_RELEASE: 3313 nvmf_tcp_req_set_state(tcp_req, TCP_REQUEST_STATE_COMPLETED); 3314 break; 3315 default: 3316 assert(0 && "Unexpected request state"); 3317 break; 3318 } 3319 3320 nvmf_tcp_req_process(ttransport, tcp_req); 3321 3322 return 0; 3323 } 3324 3325 static void 3326 nvmf_tcp_close_qpair(struct spdk_nvmf_qpair *qpair, 3327 spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg) 3328 { 3329 struct spdk_nvmf_tcp_qpair *tqpair; 3330 3331 SPDK_DEBUGLOG(nvmf_tcp, "Qpair: %p\n", qpair); 3332 3333 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3334 3335 assert(tqpair->fini_cb_fn == NULL); 3336 tqpair->fini_cb_fn = cb_fn; 3337 tqpair->fini_cb_arg = cb_arg; 3338 3339 nvmf_tcp_qpair_set_state(tqpair, NVME_TCP_QPAIR_STATE_EXITED); 3340 nvmf_tcp_qpair_destroy(tqpair); 3341 } 3342 3343 static int 3344 nvmf_tcp_poll_group_poll(struct spdk_nvmf_transport_poll_group *group) 3345 { 3346 struct spdk_nvmf_tcp_poll_group *tgroup; 3347 int rc; 3348 struct spdk_nvmf_request *req, *req_tmp; 3349 struct spdk_nvmf_tcp_req *tcp_req; 3350 struct spdk_nvmf_tcp_qpair *tqpair, *tqpair_tmp; 3351 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(group->transport, 3352 struct spdk_nvmf_tcp_transport, transport); 3353 3354 tgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_tcp_poll_group, group); 3355 3356 if (spdk_unlikely(TAILQ_EMPTY(&tgroup->qpairs) && TAILQ_EMPTY(&tgroup->await_req))) { 3357 return 0; 3358 } 3359 3360 STAILQ_FOREACH_SAFE(req, &group->pending_buf_queue, buf_link, req_tmp) { 3361 tcp_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_tcp_req, req); 3362 if (nvmf_tcp_req_process(ttransport, tcp_req) == false) { 3363 break; 3364 } 3365 } 3366 3367 rc = spdk_sock_group_poll(tgroup->sock_group); 3368 if (rc < 0) { 3369 SPDK_ERRLOG("Failed to poll sock_group=%p\n", tgroup->sock_group); 3370 } 3371 3372 TAILQ_FOREACH_SAFE(tqpair, &tgroup->await_req, link, tqpair_tmp) { 3373 rc = nvmf_tcp_sock_process(tqpair); 3374 3375 /* If there was a new socket error, disconnect */ 3376 if (rc < 0) { 3377 nvmf_tcp_qpair_disconnect(tqpair); 3378 } 3379 } 3380 3381 return rc; 3382 } 3383 3384 static int 3385 nvmf_tcp_qpair_get_trid(struct spdk_nvmf_qpair *qpair, 3386 struct spdk_nvme_transport_id *trid, bool peer) 3387 { 3388 struct spdk_nvmf_tcp_qpair *tqpair; 3389 uint16_t port; 3390 3391 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3392 spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_TCP); 3393 3394 if (peer) { 3395 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->initiator_addr); 3396 port = tqpair->initiator_port; 3397 } else { 3398 snprintf(trid->traddr, sizeof(trid->traddr), "%s", tqpair->target_addr); 3399 port = tqpair->target_port; 3400 } 3401 3402 if (spdk_sock_is_ipv4(tqpair->sock)) { 3403 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4; 3404 } else if (spdk_sock_is_ipv6(tqpair->sock)) { 3405 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6; 3406 } else { 3407 return -1; 3408 } 3409 3410 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%d", port); 3411 return 0; 3412 } 3413 3414 static int 3415 nvmf_tcp_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair, 3416 struct spdk_nvme_transport_id *trid) 3417 { 3418 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3419 } 3420 3421 static int 3422 nvmf_tcp_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair, 3423 struct spdk_nvme_transport_id *trid) 3424 { 3425 return nvmf_tcp_qpair_get_trid(qpair, trid, 1); 3426 } 3427 3428 static int 3429 nvmf_tcp_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair, 3430 struct spdk_nvme_transport_id *trid) 3431 { 3432 return nvmf_tcp_qpair_get_trid(qpair, trid, 0); 3433 } 3434 3435 static void 3436 nvmf_tcp_req_set_abort_status(struct spdk_nvmf_request *req, 3437 struct spdk_nvmf_tcp_req *tcp_req_to_abort) 3438 { 3439 nvmf_tcp_req_set_cpl(tcp_req_to_abort, SPDK_NVME_SCT_GENERIC, SPDK_NVME_SC_ABORTED_BY_REQUEST); 3440 nvmf_tcp_req_set_state(tcp_req_to_abort, TCP_REQUEST_STATE_READY_TO_COMPLETE); 3441 3442 req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */ 3443 } 3444 3445 static int 3446 _nvmf_tcp_qpair_abort_request(void *ctx) 3447 { 3448 struct spdk_nvmf_request *req = ctx; 3449 struct spdk_nvmf_tcp_req *tcp_req_to_abort = SPDK_CONTAINEROF(req->req_to_abort, 3450 struct spdk_nvmf_tcp_req, req); 3451 struct spdk_nvmf_tcp_qpair *tqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair, 3452 struct spdk_nvmf_tcp_qpair, qpair); 3453 struct spdk_nvmf_tcp_transport *ttransport = SPDK_CONTAINEROF(tqpair->qpair.transport, 3454 struct spdk_nvmf_tcp_transport, transport); 3455 int rc; 3456 3457 spdk_poller_unregister(&req->poller); 3458 3459 switch (tcp_req_to_abort->state) { 3460 case TCP_REQUEST_STATE_EXECUTING: 3461 case TCP_REQUEST_STATE_AWAITING_ZCOPY_START: 3462 case TCP_REQUEST_STATE_AWAITING_ZCOPY_COMMIT: 3463 rc = nvmf_ctrlr_abort_request(req); 3464 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) { 3465 return SPDK_POLLER_BUSY; 3466 } 3467 break; 3468 3469 case TCP_REQUEST_STATE_NEED_BUFFER: 3470 STAILQ_REMOVE(&tqpair->group->group.pending_buf_queue, 3471 &tcp_req_to_abort->req, spdk_nvmf_request, buf_link); 3472 3473 nvmf_tcp_req_set_abort_status(req, tcp_req_to_abort); 3474 nvmf_tcp_req_process(ttransport, tcp_req_to_abort); 3475 break; 3476 3477 case TCP_REQUEST_STATE_AWAITING_R2T_ACK: 3478 case TCP_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER: 3479 if (spdk_get_ticks() < req->timeout_tsc) { 3480 req->poller = SPDK_POLLER_REGISTER(_nvmf_tcp_qpair_abort_request, req, 0); 3481 return SPDK_POLLER_BUSY; 3482 } 3483 break; 3484 3485 default: 3486 /* Requests in other states are either un-abortable (e.g. 3487 * TRANSFERRING_CONTROLLER_TO_HOST) or should never end up here, as they're 3488 * immediately transitioned to other states in nvmf_tcp_req_process() (e.g. 3489 * READY_TO_EXECUTE). But it is fine to end up here, as we'll simply complete the 3490 * abort request with the bit0 of dword0 set (command not aborted). 3491 */ 3492 break; 3493 } 3494 3495 spdk_nvmf_request_complete(req); 3496 return SPDK_POLLER_BUSY; 3497 } 3498 3499 static void 3500 nvmf_tcp_qpair_abort_request(struct spdk_nvmf_qpair *qpair, 3501 struct spdk_nvmf_request *req) 3502 { 3503 struct spdk_nvmf_tcp_qpair *tqpair; 3504 struct spdk_nvmf_tcp_transport *ttransport; 3505 struct spdk_nvmf_transport *transport; 3506 uint16_t cid; 3507 uint32_t i; 3508 struct spdk_nvmf_tcp_req *tcp_req_to_abort = NULL; 3509 3510 tqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_tcp_qpair, qpair); 3511 ttransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_tcp_transport, transport); 3512 transport = &ttransport->transport; 3513 3514 cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid; 3515 3516 for (i = 0; i < tqpair->resource_count; i++) { 3517 if (tqpair->reqs[i].state != TCP_REQUEST_STATE_FREE && 3518 tqpair->reqs[i].req.cmd->nvme_cmd.cid == cid) { 3519 tcp_req_to_abort = &tqpair->reqs[i]; 3520 break; 3521 } 3522 } 3523 3524 spdk_trace_record(TRACE_TCP_QP_ABORT_REQ, qpair->qid, 0, (uintptr_t)req, tqpair); 3525 3526 if (tcp_req_to_abort == NULL) { 3527 spdk_nvmf_request_complete(req); 3528 return; 3529 } 3530 3531 req->req_to_abort = &tcp_req_to_abort->req; 3532 req->timeout_tsc = spdk_get_ticks() + 3533 transport->opts.abort_timeout_sec * spdk_get_ticks_hz(); 3534 req->poller = NULL; 3535 3536 _nvmf_tcp_qpair_abort_request(req); 3537 } 3538 3539 struct tcp_subsystem_add_host_opts { 3540 char *psk; 3541 }; 3542 3543 static const struct spdk_json_object_decoder tcp_subsystem_add_host_opts_decoder[] = { 3544 {"psk", offsetof(struct tcp_subsystem_add_host_opts, psk), spdk_json_decode_string, true}, 3545 }; 3546 3547 static int 3548 tcp_load_psk(const char *fname, char *buf, size_t bufsz) 3549 { 3550 FILE *psk_file; 3551 struct stat statbuf; 3552 int rc; 3553 3554 if (stat(fname, &statbuf) != 0) { 3555 SPDK_ERRLOG("Could not read permissions for PSK file\n"); 3556 return -EACCES; 3557 } 3558 3559 if ((statbuf.st_mode & TCP_PSK_INVALID_PERMISSIONS) != 0) { 3560 SPDK_ERRLOG("Incorrect permissions for PSK file\n"); 3561 return -EPERM; 3562 } 3563 if ((size_t)statbuf.st_size > bufsz) { 3564 SPDK_ERRLOG("Invalid PSK: too long\n"); 3565 return -EINVAL; 3566 } 3567 psk_file = fopen(fname, "r"); 3568 if (psk_file == NULL) { 3569 SPDK_ERRLOG("Could not open PSK file\n"); 3570 return -EINVAL; 3571 } 3572 3573 rc = fread(buf, 1, statbuf.st_size, psk_file); 3574 if (rc != statbuf.st_size) { 3575 SPDK_ERRLOG("Failed to read PSK\n"); 3576 fclose(psk_file); 3577 return -EINVAL; 3578 } 3579 3580 fclose(psk_file); 3581 return 0; 3582 } 3583 3584 static int 3585 nvmf_tcp_subsystem_add_host(struct spdk_nvmf_transport *transport, 3586 const struct spdk_nvmf_subsystem *subsystem, 3587 const char *hostnqn, 3588 const struct spdk_json_val *transport_specific) 3589 { 3590 struct tcp_subsystem_add_host_opts opts; 3591 struct spdk_nvmf_tcp_transport *ttransport; 3592 struct tcp_psk_entry *tmp, *entry = NULL; 3593 uint8_t psk_configured[SPDK_TLS_PSK_MAX_LEN] = {}; 3594 char psk_interchange[SPDK_TLS_PSK_MAX_LEN + 1] = {}; 3595 uint8_t tls_cipher_suite; 3596 int rc = 0; 3597 uint8_t psk_retained_hash; 3598 uint64_t psk_configured_size; 3599 3600 if (transport_specific == NULL) { 3601 return 0; 3602 } 3603 3604 assert(transport != NULL); 3605 assert(subsystem != NULL); 3606 3607 memset(&opts, 0, sizeof(opts)); 3608 3609 /* Decode PSK file path */ 3610 if (spdk_json_decode_object_relaxed(transport_specific, tcp_subsystem_add_host_opts_decoder, 3611 SPDK_COUNTOF(tcp_subsystem_add_host_opts_decoder), &opts)) { 3612 SPDK_ERRLOG("spdk_json_decode_object failed\n"); 3613 return -EINVAL; 3614 } 3615 3616 if (opts.psk == NULL) { 3617 return 0; 3618 } 3619 3620 entry = calloc(1, sizeof(struct tcp_psk_entry)); 3621 if (entry == NULL) { 3622 SPDK_ERRLOG("Unable to allocate memory for PSK entry!\n"); 3623 rc = -ENOMEM; 3624 goto end; 3625 } 3626 3627 if (strlen(opts.psk) >= sizeof(entry->psk)) { 3628 SPDK_ERRLOG("PSK path too long\n"); 3629 rc = -EINVAL; 3630 goto end; 3631 } 3632 3633 rc = tcp_load_psk(opts.psk, psk_interchange, SPDK_TLS_PSK_MAX_LEN); 3634 if (rc) { 3635 SPDK_ERRLOG("Could not retrieve PSK from file\n"); 3636 goto end; 3637 } 3638 3639 /* Parse PSK interchange to get length of base64 encoded data. 3640 * This is then used to decide which cipher suite should be used 3641 * to generate PSK identity and TLS PSK later on. */ 3642 rc = nvme_tcp_parse_interchange_psk(psk_interchange, psk_configured, sizeof(psk_configured), 3643 &psk_configured_size, &psk_retained_hash); 3644 if (rc < 0) { 3645 SPDK_ERRLOG("Failed to parse PSK interchange!\n"); 3646 goto end; 3647 } 3648 3649 /* The Base64 string encodes the configured PSK (32 or 48 bytes binary). 3650 * This check also ensures that psk_configured_size is smaller than 3651 * psk_retained buffer size. */ 3652 if (psk_configured_size == SHA256_DIGEST_LENGTH) { 3653 tls_cipher_suite = NVME_TCP_CIPHER_AES_128_GCM_SHA256; 3654 } else if (psk_configured_size == SHA384_DIGEST_LENGTH) { 3655 tls_cipher_suite = NVME_TCP_CIPHER_AES_256_GCM_SHA384; 3656 } else { 3657 SPDK_ERRLOG("Unrecognized cipher suite!\n"); 3658 rc = -EINVAL; 3659 goto end; 3660 } 3661 3662 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3663 /* Generate PSK identity. */ 3664 rc = nvme_tcp_generate_psk_identity(entry->pskid, sizeof(entry->pskid), hostnqn, 3665 subsystem->subnqn, tls_cipher_suite); 3666 if (rc) { 3667 rc = -EINVAL; 3668 goto end; 3669 } 3670 /* Check if PSK identity entry already exists. */ 3671 TAILQ_FOREACH(tmp, &ttransport->psks, link) { 3672 if (strncmp(tmp->pskid, entry->pskid, NVMF_PSK_IDENTITY_LEN) == 0) { 3673 SPDK_ERRLOG("Given PSK identity: %s entry already exists!\n", entry->pskid); 3674 rc = -EEXIST; 3675 goto end; 3676 } 3677 } 3678 3679 if (snprintf(entry->hostnqn, sizeof(entry->hostnqn), "%s", hostnqn) < 0) { 3680 SPDK_ERRLOG("Could not write hostnqn string!\n"); 3681 rc = -EINVAL; 3682 goto end; 3683 } 3684 if (snprintf(entry->subnqn, sizeof(entry->subnqn), "%s", subsystem->subnqn) < 0) { 3685 SPDK_ERRLOG("Could not write subnqn string!\n"); 3686 rc = -EINVAL; 3687 goto end; 3688 } 3689 3690 entry->tls_cipher_suite = tls_cipher_suite; 3691 3692 /* No hash indicates that Configured PSK must be used as Retained PSK. */ 3693 if (psk_retained_hash == NVME_TCP_HASH_ALGORITHM_NONE) { 3694 /* Psk configured is either 32 or 48 bytes long. */ 3695 memcpy(entry->psk, psk_configured, psk_configured_size); 3696 entry->psk_size = psk_configured_size; 3697 } else { 3698 /* Derive retained PSK. */ 3699 rc = nvme_tcp_derive_retained_psk(psk_configured, psk_configured_size, hostnqn, entry->psk, 3700 SPDK_TLS_PSK_MAX_LEN, psk_retained_hash); 3701 if (rc < 0) { 3702 SPDK_ERRLOG("Unable to derive retained PSK!\n"); 3703 goto end; 3704 } 3705 entry->psk_size = rc; 3706 } 3707 3708 rc = snprintf(entry->psk_path, sizeof(entry->psk_path), "%s", opts.psk); 3709 if (rc < 0 || (size_t)rc >= sizeof(entry->psk_path)) { 3710 SPDK_ERRLOG("Could not save PSK path!\n"); 3711 rc = -ENAMETOOLONG; 3712 goto end; 3713 } 3714 3715 TAILQ_INSERT_TAIL(&ttransport->psks, entry, link); 3716 rc = 0; 3717 3718 end: 3719 spdk_memset_s(psk_configured, sizeof(psk_configured), 0, sizeof(psk_configured)); 3720 spdk_memset_s(psk_interchange, sizeof(psk_interchange), 0, sizeof(psk_interchange)); 3721 3722 free(opts.psk); 3723 if (rc != 0) { 3724 nvmf_tcp_free_psk_entry(entry); 3725 } 3726 3727 return rc; 3728 } 3729 3730 static void 3731 nvmf_tcp_subsystem_remove_host(struct spdk_nvmf_transport *transport, 3732 const struct spdk_nvmf_subsystem *subsystem, 3733 const char *hostnqn) 3734 { 3735 struct spdk_nvmf_tcp_transport *ttransport; 3736 struct tcp_psk_entry *entry, *tmp; 3737 3738 assert(transport != NULL); 3739 assert(subsystem != NULL); 3740 3741 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3742 TAILQ_FOREACH_SAFE(entry, &ttransport->psks, link, tmp) { 3743 if ((strncmp(entry->hostnqn, hostnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0 && 3744 (strncmp(entry->subnqn, subsystem->subnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0) { 3745 TAILQ_REMOVE(&ttransport->psks, entry, link); 3746 nvmf_tcp_free_psk_entry(entry); 3747 break; 3748 } 3749 } 3750 } 3751 3752 static void 3753 nvmf_tcp_subsystem_dump_host(struct spdk_nvmf_transport *transport, 3754 const struct spdk_nvmf_subsystem *subsystem, const char *hostnqn, 3755 struct spdk_json_write_ctx *w) 3756 { 3757 struct spdk_nvmf_tcp_transport *ttransport; 3758 struct tcp_psk_entry *entry; 3759 3760 assert(transport != NULL); 3761 assert(subsystem != NULL); 3762 3763 ttransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_tcp_transport, transport); 3764 TAILQ_FOREACH(entry, &ttransport->psks, link) { 3765 if ((strncmp(entry->hostnqn, hostnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0 && 3766 (strncmp(entry->subnqn, subsystem->subnqn, SPDK_NVMF_NQN_MAX_LEN)) == 0) { 3767 spdk_json_write_named_string(w, "psk", entry->psk_path); 3768 break; 3769 } 3770 } 3771 } 3772 3773 static void 3774 nvmf_tcp_opts_init(struct spdk_nvmf_transport_opts *opts) 3775 { 3776 opts->max_queue_depth = SPDK_NVMF_TCP_DEFAULT_MAX_IO_QUEUE_DEPTH; 3777 opts->max_qpairs_per_ctrlr = SPDK_NVMF_TCP_DEFAULT_MAX_QPAIRS_PER_CTRLR; 3778 opts->in_capsule_data_size = SPDK_NVMF_TCP_DEFAULT_IN_CAPSULE_DATA_SIZE; 3779 opts->max_io_size = SPDK_NVMF_TCP_DEFAULT_MAX_IO_SIZE; 3780 opts->io_unit_size = SPDK_NVMF_TCP_DEFAULT_IO_UNIT_SIZE; 3781 opts->max_aq_depth = SPDK_NVMF_TCP_DEFAULT_MAX_ADMIN_QUEUE_DEPTH; 3782 opts->num_shared_buffers = SPDK_NVMF_TCP_DEFAULT_NUM_SHARED_BUFFERS; 3783 opts->buf_cache_size = SPDK_NVMF_TCP_DEFAULT_BUFFER_CACHE_SIZE; 3784 opts->dif_insert_or_strip = SPDK_NVMF_TCP_DEFAULT_DIF_INSERT_OR_STRIP; 3785 opts->abort_timeout_sec = SPDK_NVMF_TCP_DEFAULT_ABORT_TIMEOUT_SEC; 3786 opts->transport_specific = NULL; 3787 } 3788 3789 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp = { 3790 .name = "TCP", 3791 .type = SPDK_NVME_TRANSPORT_TCP, 3792 .opts_init = nvmf_tcp_opts_init, 3793 .create = nvmf_tcp_create, 3794 .dump_opts = nvmf_tcp_dump_opts, 3795 .destroy = nvmf_tcp_destroy, 3796 3797 .listen = nvmf_tcp_listen, 3798 .stop_listen = nvmf_tcp_stop_listen, 3799 3800 .listener_discover = nvmf_tcp_discover, 3801 3802 .poll_group_create = nvmf_tcp_poll_group_create, 3803 .get_optimal_poll_group = nvmf_tcp_get_optimal_poll_group, 3804 .poll_group_destroy = nvmf_tcp_poll_group_destroy, 3805 .poll_group_add = nvmf_tcp_poll_group_add, 3806 .poll_group_remove = nvmf_tcp_poll_group_remove, 3807 .poll_group_poll = nvmf_tcp_poll_group_poll, 3808 3809 .req_free = nvmf_tcp_req_free, 3810 .req_complete = nvmf_tcp_req_complete, 3811 3812 .qpair_fini = nvmf_tcp_close_qpair, 3813 .qpair_get_local_trid = nvmf_tcp_qpair_get_local_trid, 3814 .qpair_get_peer_trid = nvmf_tcp_qpair_get_peer_trid, 3815 .qpair_get_listen_trid = nvmf_tcp_qpair_get_listen_trid, 3816 .qpair_abort_request = nvmf_tcp_qpair_abort_request, 3817 .subsystem_add_host = nvmf_tcp_subsystem_add_host, 3818 .subsystem_remove_host = nvmf_tcp_subsystem_remove_host, 3819 .subsystem_dump_host = nvmf_tcp_subsystem_dump_host, 3820 }; 3821 3822 SPDK_NVMF_TRANSPORT_REGISTER(tcp, &spdk_nvmf_transport_tcp); 3823 SPDK_LOG_REGISTER_COMPONENT(nvmf_tcp) 3824