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