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