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