1 /* SPDX-License-Identifier: GPL-2.0 */ 2 /* Multipath TCP 3 * 4 * Copyright (c) 2017 - 2019, Intel Corporation. 5 */ 6 7 #ifndef __MPTCP_PROTOCOL_H 8 #define __MPTCP_PROTOCOL_H 9 10 #include <linux/random.h> 11 #include <net/tcp.h> 12 #include <net/inet_connection_sock.h> 13 #include <uapi/linux/mptcp.h> 14 #include <net/genetlink.h> 15 16 #define MPTCP_SUPPORTED_VERSION 1 17 18 /* MPTCP option bits */ 19 #define OPTION_MPTCP_MPC_SYN BIT(0) 20 #define OPTION_MPTCP_MPC_SYNACK BIT(1) 21 #define OPTION_MPTCP_MPC_ACK BIT(2) 22 #define OPTION_MPTCP_MPJ_SYN BIT(3) 23 #define OPTION_MPTCP_MPJ_SYNACK BIT(4) 24 #define OPTION_MPTCP_MPJ_ACK BIT(5) 25 #define OPTION_MPTCP_ADD_ADDR BIT(6) 26 #define OPTION_MPTCP_RM_ADDR BIT(7) 27 #define OPTION_MPTCP_FASTCLOSE BIT(8) 28 #define OPTION_MPTCP_PRIO BIT(9) 29 #define OPTION_MPTCP_RST BIT(10) 30 #define OPTION_MPTCP_DSS BIT(11) 31 #define OPTION_MPTCP_FAIL BIT(12) 32 33 #define OPTION_MPTCP_CSUMREQD BIT(13) 34 35 #define OPTIONS_MPTCP_MPC (OPTION_MPTCP_MPC_SYN | OPTION_MPTCP_MPC_SYNACK | \ 36 OPTION_MPTCP_MPC_ACK) 37 #define OPTIONS_MPTCP_MPJ (OPTION_MPTCP_MPJ_SYN | OPTION_MPTCP_MPJ_SYNACK | \ 38 OPTION_MPTCP_MPJ_ACK) 39 40 /* MPTCP option subtypes */ 41 #define MPTCPOPT_MP_CAPABLE 0 42 #define MPTCPOPT_MP_JOIN 1 43 #define MPTCPOPT_DSS 2 44 #define MPTCPOPT_ADD_ADDR 3 45 #define MPTCPOPT_RM_ADDR 4 46 #define MPTCPOPT_MP_PRIO 5 47 #define MPTCPOPT_MP_FAIL 6 48 #define MPTCPOPT_MP_FASTCLOSE 7 49 #define MPTCPOPT_RST 8 50 51 /* MPTCP suboption lengths */ 52 #define TCPOLEN_MPTCP_MPC_SYN 4 53 #define TCPOLEN_MPTCP_MPC_SYNACK 12 54 #define TCPOLEN_MPTCP_MPC_ACK 20 55 #define TCPOLEN_MPTCP_MPC_ACK_DATA 22 56 #define TCPOLEN_MPTCP_MPJ_SYN 12 57 #define TCPOLEN_MPTCP_MPJ_SYNACK 16 58 #define TCPOLEN_MPTCP_MPJ_ACK 24 59 #define TCPOLEN_MPTCP_DSS_BASE 4 60 #define TCPOLEN_MPTCP_DSS_ACK32 4 61 #define TCPOLEN_MPTCP_DSS_ACK64 8 62 #define TCPOLEN_MPTCP_DSS_MAP32 10 63 #define TCPOLEN_MPTCP_DSS_MAP64 14 64 #define TCPOLEN_MPTCP_DSS_CHECKSUM 2 65 #define TCPOLEN_MPTCP_ADD_ADDR 16 66 #define TCPOLEN_MPTCP_ADD_ADDR_PORT 18 67 #define TCPOLEN_MPTCP_ADD_ADDR_BASE 8 68 #define TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT 10 69 #define TCPOLEN_MPTCP_ADD_ADDR6 28 70 #define TCPOLEN_MPTCP_ADD_ADDR6_PORT 30 71 #define TCPOLEN_MPTCP_ADD_ADDR6_BASE 20 72 #define TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT 22 73 #define TCPOLEN_MPTCP_PORT_LEN 2 74 #define TCPOLEN_MPTCP_PORT_ALIGN 2 75 #define TCPOLEN_MPTCP_RM_ADDR_BASE 3 76 #define TCPOLEN_MPTCP_PRIO 3 77 #define TCPOLEN_MPTCP_PRIO_ALIGN 4 78 #define TCPOLEN_MPTCP_FASTCLOSE 12 79 #define TCPOLEN_MPTCP_RST 4 80 #define TCPOLEN_MPTCP_FAIL 12 81 82 #define TCPOLEN_MPTCP_MPC_ACK_DATA_CSUM (TCPOLEN_MPTCP_DSS_CHECKSUM + TCPOLEN_MPTCP_MPC_ACK_DATA) 83 84 /* MPTCP MP_JOIN flags */ 85 #define MPTCPOPT_BACKUP BIT(0) 86 #define MPTCPOPT_THMAC_LEN 8 87 88 /* MPTCP MP_CAPABLE flags */ 89 #define MPTCP_VERSION_MASK (0x0F) 90 #define MPTCP_CAP_CHECKSUM_REQD BIT(7) 91 #define MPTCP_CAP_EXTENSIBILITY BIT(6) 92 #define MPTCP_CAP_DENY_JOIN_ID0 BIT(5) 93 #define MPTCP_CAP_HMAC_SHA256 BIT(0) 94 #define MPTCP_CAP_FLAG_MASK (0x1F) 95 96 /* MPTCP DSS flags */ 97 #define MPTCP_DSS_DATA_FIN BIT(4) 98 #define MPTCP_DSS_DSN64 BIT(3) 99 #define MPTCP_DSS_HAS_MAP BIT(2) 100 #define MPTCP_DSS_ACK64 BIT(1) 101 #define MPTCP_DSS_HAS_ACK BIT(0) 102 #define MPTCP_DSS_FLAG_MASK (0x1F) 103 104 /* MPTCP ADD_ADDR flags */ 105 #define MPTCP_ADDR_ECHO BIT(0) 106 107 /* MPTCP MP_PRIO flags */ 108 #define MPTCP_PRIO_BKUP BIT(0) 109 110 /* MPTCP TCPRST flags */ 111 #define MPTCP_RST_TRANSIENT BIT(0) 112 113 /* MPTCP socket atomic flags */ 114 #define MPTCP_NOSPACE 1 115 #define MPTCP_WORK_RTX 2 116 #define MPTCP_FALLBACK_DONE 4 117 #define MPTCP_WORK_CLOSE_SUBFLOW 5 118 119 /* MPTCP socket release cb flags */ 120 #define MPTCP_PUSH_PENDING 1 121 #define MPTCP_CLEAN_UNA 2 122 #define MPTCP_ERROR_REPORT 3 123 #define MPTCP_RETRANSMIT 4 124 #define MPTCP_FLUSH_JOIN_LIST 5 125 #define MPTCP_CONNECTED 6 126 127 struct mptcp_skb_cb { 128 u64 map_seq; 129 u64 end_seq; 130 u32 offset; 131 u8 has_rxtstamp:1; 132 }; 133 134 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0])) 135 136 static inline bool before64(__u64 seq1, __u64 seq2) 137 { 138 return (__s64)(seq1 - seq2) < 0; 139 } 140 141 #define after64(seq2, seq1) before64(seq1, seq2) 142 143 struct mptcp_options_received { 144 u64 sndr_key; 145 u64 rcvr_key; 146 u64 data_ack; 147 u64 data_seq; 148 u32 subflow_seq; 149 u16 data_len; 150 __sum16 csum; 151 u16 suboptions; 152 u32 token; 153 u32 nonce; 154 u16 use_map:1, 155 dsn64:1, 156 data_fin:1, 157 use_ack:1, 158 ack64:1, 159 mpc_map:1, 160 reset_reason:4, 161 reset_transient:1, 162 echo:1, 163 backup:1, 164 deny_join_id0:1, 165 __unused:2; 166 u8 join_id; 167 u64 thmac; 168 u8 hmac[MPTCPOPT_HMAC_LEN]; 169 struct mptcp_addr_info addr; 170 struct mptcp_rm_list rm_list; 171 u64 ahmac; 172 u64 fail_seq; 173 }; 174 175 static inline __be32 mptcp_option(u8 subopt, u8 len, u8 nib, u8 field) 176 { 177 return htonl((TCPOPT_MPTCP << 24) | (len << 16) | (subopt << 12) | 178 ((nib & 0xF) << 8) | field); 179 } 180 181 enum mptcp_pm_status { 182 MPTCP_PM_ADD_ADDR_RECEIVED, 183 MPTCP_PM_ADD_ADDR_SEND_ACK, 184 MPTCP_PM_RM_ADDR_RECEIVED, 185 MPTCP_PM_ESTABLISHED, 186 MPTCP_PM_SUBFLOW_ESTABLISHED, 187 MPTCP_PM_ALREADY_ESTABLISHED, /* persistent status, set after ESTABLISHED event */ 188 MPTCP_PM_MPC_ENDPOINT_ACCOUNTED /* persistent status, set after MPC local address is 189 * accounted int id_avail_bitmap 190 */ 191 }; 192 193 enum mptcp_pm_type { 194 MPTCP_PM_TYPE_KERNEL = 0, 195 MPTCP_PM_TYPE_USERSPACE, 196 197 __MPTCP_PM_TYPE_NR, 198 __MPTCP_PM_TYPE_MAX = __MPTCP_PM_TYPE_NR - 1, 199 }; 200 201 /* Status bits below MPTCP_PM_ALREADY_ESTABLISHED need pm worker actions */ 202 #define MPTCP_PM_WORK_MASK ((1 << MPTCP_PM_ALREADY_ESTABLISHED) - 1) 203 204 enum mptcp_addr_signal_status { 205 MPTCP_ADD_ADDR_SIGNAL, 206 MPTCP_ADD_ADDR_ECHO, 207 MPTCP_RM_ADDR_SIGNAL, 208 }; 209 210 /* max value of mptcp_addr_info.id */ 211 #define MPTCP_PM_MAX_ADDR_ID U8_MAX 212 213 struct mptcp_pm_data { 214 struct mptcp_addr_info local; 215 struct mptcp_addr_info remote; 216 struct list_head anno_list; 217 struct list_head userspace_pm_local_addr_list; 218 219 spinlock_t lock; /*protects the whole PM data */ 220 221 u8 addr_signal; 222 bool server_side; 223 bool work_pending; 224 bool accept_addr; 225 bool accept_subflow; 226 bool remote_deny_join_id0; 227 u8 add_addr_signaled; 228 u8 add_addr_accepted; 229 u8 local_addr_used; 230 u8 pm_type; 231 u8 subflows; 232 u8 status; 233 DECLARE_BITMAP(id_avail_bitmap, MPTCP_PM_MAX_ADDR_ID + 1); 234 struct mptcp_rm_list rm_list_tx; 235 struct mptcp_rm_list rm_list_rx; 236 }; 237 238 struct mptcp_pm_addr_entry { 239 struct list_head list; 240 struct mptcp_addr_info addr; 241 u8 flags; 242 int ifindex; 243 struct socket *lsk; 244 }; 245 246 struct mptcp_data_frag { 247 struct list_head list; 248 u64 data_seq; 249 u16 data_len; 250 u16 offset; 251 u16 overhead; 252 u16 already_sent; 253 struct page *page; 254 }; 255 256 /* MPTCP connection sock */ 257 struct mptcp_sock { 258 /* inet_connection_sock must be the first member */ 259 struct inet_connection_sock sk; 260 u64 local_key; 261 u64 remote_key; 262 u64 write_seq; 263 u64 bytes_sent; 264 u64 snd_nxt; 265 u64 bytes_received; 266 u64 ack_seq; 267 atomic64_t rcv_wnd_sent; 268 u64 rcv_data_fin_seq; 269 u64 bytes_retrans; 270 int rmem_fwd_alloc; 271 int snd_burst; 272 int old_wspace; 273 u64 recovery_snd_nxt; /* in recovery mode accept up to this seq; 274 * recovery related fields are under data_lock 275 * protection 276 */ 277 u64 bytes_acked; 278 u64 snd_una; 279 u64 wnd_end; 280 unsigned long timer_ival; 281 u32 token; 282 int rmem_released; 283 unsigned long flags; 284 unsigned long cb_flags; 285 unsigned long push_pending; 286 bool recovery; /* closing subflow write queue reinjected */ 287 bool can_ack; 288 bool fully_established; 289 bool rcv_data_fin; 290 bool snd_data_fin_enable; 291 bool rcv_fastclose; 292 bool use_64bit_ack; /* Set when we received a 64-bit DSN */ 293 bool csum_enabled; 294 bool allow_infinite_fallback; 295 u8 mpc_endpoint_id; 296 u8 recvmsg_inq:1, 297 cork:1, 298 nodelay:1, 299 fastopening:1, 300 in_accept_queue:1, 301 free_first:1; 302 struct work_struct work; 303 struct sk_buff *ooo_last_skb; 304 struct rb_root out_of_order_queue; 305 struct sk_buff_head receive_queue; 306 struct list_head conn_list; 307 struct list_head rtx_queue; 308 struct mptcp_data_frag *first_pending; 309 struct list_head join_list; 310 struct sock *first; /* The mptcp ops can safely dereference, using suitable 311 * ONCE annotation, the subflow outside the socket 312 * lock as such sock is freed after close(). 313 */ 314 struct mptcp_pm_data pm; 315 struct mptcp_sched_ops *sched; 316 struct { 317 u32 space; /* bytes copied in last measurement window */ 318 u32 copied; /* bytes copied in this measurement window */ 319 u64 time; /* start time of measurement window */ 320 u64 rtt_us; /* last maximum rtt of subflows */ 321 } rcvq_space; 322 u8 scaling_ratio; 323 324 u32 subflow_id; 325 u32 setsockopt_seq; 326 char ca_name[TCP_CA_NAME_MAX]; 327 }; 328 329 #define mptcp_data_lock(sk) spin_lock_bh(&(sk)->sk_lock.slock) 330 #define mptcp_data_unlock(sk) spin_unlock_bh(&(sk)->sk_lock.slock) 331 332 #define mptcp_for_each_subflow(__msk, __subflow) \ 333 list_for_each_entry(__subflow, &((__msk)->conn_list), node) 334 #define mptcp_for_each_subflow_safe(__msk, __subflow, __tmp) \ 335 list_for_each_entry_safe(__subflow, __tmp, &((__msk)->conn_list), node) 336 337 static inline void msk_owned_by_me(const struct mptcp_sock *msk) 338 { 339 sock_owned_by_me((const struct sock *)msk); 340 } 341 342 #define mptcp_sk(ptr) container_of_const(ptr, struct mptcp_sock, sk.icsk_inet.sk) 343 344 /* the msk socket don't use the backlog, also account for the bulk 345 * free memory 346 */ 347 static inline int __mptcp_rmem(const struct sock *sk) 348 { 349 return atomic_read(&sk->sk_rmem_alloc) - READ_ONCE(mptcp_sk(sk)->rmem_released); 350 } 351 352 static inline int mptcp_win_from_space(const struct sock *sk, int space) 353 { 354 return __tcp_win_from_space(mptcp_sk(sk)->scaling_ratio, space); 355 } 356 357 static inline int __mptcp_space(const struct sock *sk) 358 { 359 return mptcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf) - __mptcp_rmem(sk)); 360 } 361 362 static inline struct mptcp_data_frag *mptcp_send_head(const struct sock *sk) 363 { 364 const struct mptcp_sock *msk = mptcp_sk(sk); 365 366 return READ_ONCE(msk->first_pending); 367 } 368 369 static inline struct mptcp_data_frag *mptcp_send_next(struct sock *sk) 370 { 371 struct mptcp_sock *msk = mptcp_sk(sk); 372 struct mptcp_data_frag *cur; 373 374 cur = msk->first_pending; 375 return list_is_last(&cur->list, &msk->rtx_queue) ? NULL : 376 list_next_entry(cur, list); 377 } 378 379 static inline struct mptcp_data_frag *mptcp_pending_tail(const struct sock *sk) 380 { 381 const struct mptcp_sock *msk = mptcp_sk(sk); 382 383 if (!msk->first_pending) 384 return NULL; 385 386 if (WARN_ON_ONCE(list_empty(&msk->rtx_queue))) 387 return NULL; 388 389 return list_last_entry(&msk->rtx_queue, struct mptcp_data_frag, list); 390 } 391 392 static inline struct mptcp_data_frag *mptcp_rtx_head(struct sock *sk) 393 { 394 struct mptcp_sock *msk = mptcp_sk(sk); 395 396 if (msk->snd_una == READ_ONCE(msk->snd_nxt)) 397 return NULL; 398 399 return list_first_entry_or_null(&msk->rtx_queue, struct mptcp_data_frag, list); 400 } 401 402 struct csum_pseudo_header { 403 __be64 data_seq; 404 __be32 subflow_seq; 405 __be16 data_len; 406 __sum16 csum; 407 }; 408 409 struct mptcp_subflow_request_sock { 410 struct tcp_request_sock sk; 411 u16 mp_capable : 1, 412 mp_join : 1, 413 backup : 1, 414 csum_reqd : 1, 415 allow_join_id0 : 1; 416 u8 local_id; 417 u8 remote_id; 418 u64 local_key; 419 u64 idsn; 420 u32 token; 421 u32 ssn_offset; 422 u64 thmac; 423 u32 local_nonce; 424 u32 remote_nonce; 425 struct mptcp_sock *msk; 426 struct hlist_nulls_node token_node; 427 }; 428 429 static inline struct mptcp_subflow_request_sock * 430 mptcp_subflow_rsk(const struct request_sock *rsk) 431 { 432 return (struct mptcp_subflow_request_sock *)rsk; 433 } 434 435 enum mptcp_data_avail { 436 MPTCP_SUBFLOW_NODATA, 437 MPTCP_SUBFLOW_DATA_AVAIL, 438 }; 439 440 struct mptcp_delegated_action { 441 struct napi_struct napi; 442 struct list_head head; 443 }; 444 445 DECLARE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions); 446 447 #define MPTCP_DELEGATE_SEND 0 448 #define MPTCP_DELEGATE_ACK 1 449 450 /* MPTCP subflow context */ 451 struct mptcp_subflow_context { 452 struct list_head node;/* conn_list of subflows */ 453 454 struct_group(reset, 455 456 unsigned long avg_pacing_rate; /* protected by msk socket lock */ 457 u64 local_key; 458 u64 remote_key; 459 u64 idsn; 460 u64 map_seq; 461 u32 snd_isn; 462 u32 token; 463 u32 rel_write_seq; 464 u32 map_subflow_seq; 465 u32 ssn_offset; 466 u32 map_data_len; 467 __wsum map_data_csum; 468 u32 map_csum_len; 469 u32 request_mptcp : 1, /* send MP_CAPABLE */ 470 request_join : 1, /* send MP_JOIN */ 471 request_bkup : 1, 472 mp_capable : 1, /* remote is MPTCP capable */ 473 mp_join : 1, /* remote is JOINing */ 474 fully_established : 1, /* path validated */ 475 pm_notified : 1, /* PM hook called for established status */ 476 conn_finished : 1, 477 map_valid : 1, 478 map_csum_reqd : 1, 479 map_data_fin : 1, 480 mpc_map : 1, 481 backup : 1, 482 send_mp_prio : 1, 483 send_mp_fail : 1, 484 send_fastclose : 1, 485 send_infinite_map : 1, 486 remote_key_valid : 1, /* received the peer key from */ 487 disposable : 1, /* ctx can be free at ulp release time */ 488 stale : 1, /* unable to snd/rcv data, do not use for xmit */ 489 local_id_valid : 1, /* local_id is correctly initialized */ 490 valid_csum_seen : 1, /* at least one csum validated */ 491 is_mptfo : 1, /* subflow is doing TFO */ 492 __unused : 9; 493 enum mptcp_data_avail data_avail; 494 bool scheduled; 495 u32 remote_nonce; 496 u64 thmac; 497 u32 local_nonce; 498 u32 remote_token; 499 union { 500 u8 hmac[MPTCPOPT_HMAC_LEN]; /* MPJ subflow only */ 501 u64 iasn; /* initial ack sequence number, MPC subflows only */ 502 }; 503 u8 local_id; 504 u8 remote_id; 505 u8 reset_seen:1; 506 u8 reset_transient:1; 507 u8 reset_reason:4; 508 u8 stale_count; 509 510 u32 subflow_id; 511 512 long delegated_status; 513 unsigned long fail_tout; 514 515 ); 516 517 struct list_head delegated_node; /* link into delegated_action, protected by local BH */ 518 519 u32 setsockopt_seq; 520 u32 stale_rcv_tstamp; 521 522 struct sock *tcp_sock; /* tcp sk backpointer */ 523 struct sock *conn; /* parent mptcp_sock */ 524 const struct inet_connection_sock_af_ops *icsk_af_ops; 525 void (*tcp_state_change)(struct sock *sk); 526 void (*tcp_error_report)(struct sock *sk); 527 528 struct rcu_head rcu; 529 }; 530 531 static inline struct mptcp_subflow_context * 532 mptcp_subflow_ctx(const struct sock *sk) 533 { 534 struct inet_connection_sock *icsk = inet_csk(sk); 535 536 /* Use RCU on icsk_ulp_data only for sock diag code */ 537 return (__force struct mptcp_subflow_context *)icsk->icsk_ulp_data; 538 } 539 540 static inline struct sock * 541 mptcp_subflow_tcp_sock(const struct mptcp_subflow_context *subflow) 542 { 543 return subflow->tcp_sock; 544 } 545 546 static inline void 547 mptcp_subflow_ctx_reset(struct mptcp_subflow_context *subflow) 548 { 549 memset(&subflow->reset, 0, sizeof(subflow->reset)); 550 subflow->request_mptcp = 1; 551 } 552 553 static inline u64 554 mptcp_subflow_get_map_offset(const struct mptcp_subflow_context *subflow) 555 { 556 return tcp_sk(mptcp_subflow_tcp_sock(subflow))->copied_seq - 557 subflow->ssn_offset - 558 subflow->map_subflow_seq; 559 } 560 561 static inline u64 562 mptcp_subflow_get_mapped_dsn(const struct mptcp_subflow_context *subflow) 563 { 564 return subflow->map_seq + mptcp_subflow_get_map_offset(subflow); 565 } 566 567 void mptcp_subflow_process_delegated(struct sock *ssk); 568 569 static inline void mptcp_subflow_delegate(struct mptcp_subflow_context *subflow, int action) 570 { 571 struct mptcp_delegated_action *delegated; 572 bool schedule; 573 574 /* the caller held the subflow bh socket lock */ 575 lockdep_assert_in_softirq(); 576 577 /* The implied barrier pairs with mptcp_subflow_delegated_done(), and 578 * ensures the below list check sees list updates done prior to status 579 * bit changes 580 */ 581 if (!test_and_set_bit(action, &subflow->delegated_status)) { 582 /* still on delegated list from previous scheduling */ 583 if (!list_empty(&subflow->delegated_node)) 584 return; 585 586 delegated = this_cpu_ptr(&mptcp_delegated_actions); 587 schedule = list_empty(&delegated->head); 588 list_add_tail(&subflow->delegated_node, &delegated->head); 589 sock_hold(mptcp_subflow_tcp_sock(subflow)); 590 if (schedule) 591 napi_schedule(&delegated->napi); 592 } 593 } 594 595 static inline struct mptcp_subflow_context * 596 mptcp_subflow_delegated_next(struct mptcp_delegated_action *delegated) 597 { 598 struct mptcp_subflow_context *ret; 599 600 if (list_empty(&delegated->head)) 601 return NULL; 602 603 ret = list_first_entry(&delegated->head, struct mptcp_subflow_context, delegated_node); 604 list_del_init(&ret->delegated_node); 605 return ret; 606 } 607 608 static inline bool mptcp_subflow_has_delegated_action(const struct mptcp_subflow_context *subflow) 609 { 610 return !!READ_ONCE(subflow->delegated_status); 611 } 612 613 static inline void mptcp_subflow_delegated_done(struct mptcp_subflow_context *subflow, int action) 614 { 615 /* pairs with mptcp_subflow_delegate, ensures delegate_node is updated before 616 * touching the status bit 617 */ 618 smp_wmb(); 619 clear_bit(action, &subflow->delegated_status); 620 } 621 622 int mptcp_is_enabled(const struct net *net); 623 unsigned int mptcp_get_add_addr_timeout(const struct net *net); 624 int mptcp_is_checksum_enabled(const struct net *net); 625 int mptcp_allow_join_id0(const struct net *net); 626 unsigned int mptcp_stale_loss_cnt(const struct net *net); 627 int mptcp_get_pm_type(const struct net *net); 628 const char *mptcp_get_scheduler(const struct net *net); 629 void mptcp_subflow_fully_established(struct mptcp_subflow_context *subflow, 630 const struct mptcp_options_received *mp_opt); 631 bool __mptcp_retransmit_pending_data(struct sock *sk); 632 void mptcp_check_and_set_pending(struct sock *sk); 633 void __mptcp_push_pending(struct sock *sk, unsigned int flags); 634 bool mptcp_subflow_data_available(struct sock *sk); 635 void __init mptcp_subflow_init(void); 636 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how); 637 void mptcp_close_ssk(struct sock *sk, struct sock *ssk, 638 struct mptcp_subflow_context *subflow); 639 void __mptcp_subflow_send_ack(struct sock *ssk); 640 void mptcp_subflow_reset(struct sock *ssk); 641 void mptcp_subflow_queue_clean(struct sock *sk, struct sock *ssk); 642 void mptcp_sock_graft(struct sock *sk, struct socket *parent); 643 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk); 644 bool __mptcp_close(struct sock *sk, long timeout); 645 void mptcp_cancel_work(struct sock *sk); 646 void __mptcp_unaccepted_force_close(struct sock *sk); 647 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk); 648 649 bool mptcp_addresses_equal(const struct mptcp_addr_info *a, 650 const struct mptcp_addr_info *b, bool use_port); 651 void mptcp_local_address(const struct sock_common *skc, struct mptcp_addr_info *addr); 652 653 /* called with sk socket lock held */ 654 int __mptcp_subflow_connect(struct sock *sk, const struct mptcp_addr_info *loc, 655 const struct mptcp_addr_info *remote); 656 int mptcp_subflow_create_socket(struct sock *sk, unsigned short family, 657 struct socket **new_sock); 658 void mptcp_info2sockaddr(const struct mptcp_addr_info *info, 659 struct sockaddr_storage *addr, 660 unsigned short family); 661 struct mptcp_sched_ops *mptcp_sched_find(const char *name); 662 int mptcp_register_scheduler(struct mptcp_sched_ops *sched); 663 void mptcp_unregister_scheduler(struct mptcp_sched_ops *sched); 664 void mptcp_sched_init(void); 665 int mptcp_init_sched(struct mptcp_sock *msk, 666 struct mptcp_sched_ops *sched); 667 void mptcp_release_sched(struct mptcp_sock *msk); 668 void mptcp_subflow_set_scheduled(struct mptcp_subflow_context *subflow, 669 bool scheduled); 670 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk); 671 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk); 672 int mptcp_sched_get_send(struct mptcp_sock *msk); 673 int mptcp_sched_get_retrans(struct mptcp_sock *msk); 674 675 static inline bool __tcp_can_send(const struct sock *ssk) 676 { 677 /* only send if our side has not closed yet */ 678 return ((1 << inet_sk_state_load(ssk)) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)); 679 } 680 681 static inline bool __mptcp_subflow_active(struct mptcp_subflow_context *subflow) 682 { 683 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */ 684 if (subflow->request_join && !subflow->fully_established) 685 return false; 686 687 return __tcp_can_send(mptcp_subflow_tcp_sock(subflow)); 688 } 689 690 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow); 691 692 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow); 693 694 void mptcp_subflow_drop_ctx(struct sock *ssk); 695 696 static inline void mptcp_subflow_tcp_fallback(struct sock *sk, 697 struct mptcp_subflow_context *ctx) 698 { 699 sk->sk_data_ready = sock_def_readable; 700 sk->sk_state_change = ctx->tcp_state_change; 701 sk->sk_write_space = sk_stream_write_space; 702 sk->sk_error_report = ctx->tcp_error_report; 703 704 inet_csk(sk)->icsk_af_ops = ctx->icsk_af_ops; 705 } 706 707 void __init mptcp_proto_init(void); 708 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 709 int __init mptcp_proto_v6_init(void); 710 #endif 711 712 struct sock *mptcp_sk_clone_init(const struct sock *sk, 713 const struct mptcp_options_received *mp_opt, 714 struct sock *ssk, 715 struct request_sock *req); 716 void mptcp_get_options(const struct sk_buff *skb, 717 struct mptcp_options_received *mp_opt); 718 719 void mptcp_finish_connect(struct sock *sk); 720 void __mptcp_set_connected(struct sock *sk); 721 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout); 722 static inline bool mptcp_is_fully_established(struct sock *sk) 723 { 724 return inet_sk_state_load(sk) == TCP_ESTABLISHED && 725 READ_ONCE(mptcp_sk(sk)->fully_established); 726 } 727 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk); 728 void mptcp_data_ready(struct sock *sk, struct sock *ssk); 729 bool mptcp_finish_join(struct sock *sk); 730 bool mptcp_schedule_work(struct sock *sk); 731 int mptcp_setsockopt(struct sock *sk, int level, int optname, 732 sockptr_t optval, unsigned int optlen); 733 int mptcp_getsockopt(struct sock *sk, int level, int optname, 734 char __user *optval, int __user *option); 735 736 u64 __mptcp_expand_seq(u64 old_seq, u64 cur_seq); 737 static inline u64 mptcp_expand_seq(u64 old_seq, u64 cur_seq, bool use_64bit) 738 { 739 if (use_64bit) 740 return cur_seq; 741 742 return __mptcp_expand_seq(old_seq, cur_seq); 743 } 744 void __mptcp_check_push(struct sock *sk, struct sock *ssk); 745 void __mptcp_data_acked(struct sock *sk); 746 void __mptcp_error_report(struct sock *sk); 747 bool mptcp_update_rcv_data_fin(struct mptcp_sock *msk, u64 data_fin_seq, bool use_64bit); 748 static inline bool mptcp_data_fin_enabled(const struct mptcp_sock *msk) 749 { 750 return READ_ONCE(msk->snd_data_fin_enable) && 751 READ_ONCE(msk->write_seq) == READ_ONCE(msk->snd_nxt); 752 } 753 754 static inline bool mptcp_propagate_sndbuf(struct sock *sk, struct sock *ssk) 755 { 756 if ((sk->sk_userlocks & SOCK_SNDBUF_LOCK) || ssk->sk_sndbuf <= READ_ONCE(sk->sk_sndbuf)) 757 return false; 758 759 WRITE_ONCE(sk->sk_sndbuf, ssk->sk_sndbuf); 760 return true; 761 } 762 763 static inline void mptcp_write_space(struct sock *sk) 764 { 765 if (sk_stream_is_writeable(sk)) { 766 /* pairs with memory barrier in mptcp_poll */ 767 smp_mb(); 768 if (test_and_clear_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags)) 769 sk_stream_write_space(sk); 770 } 771 } 772 773 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags); 774 775 #define MPTCP_TOKEN_MAX_RETRIES 4 776 777 void __init mptcp_token_init(void); 778 static inline void mptcp_token_init_request(struct request_sock *req) 779 { 780 mptcp_subflow_rsk(req)->token_node.pprev = NULL; 781 } 782 783 int mptcp_token_new_request(struct request_sock *req); 784 void mptcp_token_destroy_request(struct request_sock *req); 785 int mptcp_token_new_connect(struct sock *ssk); 786 void mptcp_token_accept(struct mptcp_subflow_request_sock *r, 787 struct mptcp_sock *msk); 788 bool mptcp_token_exists(u32 token); 789 struct mptcp_sock *mptcp_token_get_sock(struct net *net, u32 token); 790 struct mptcp_sock *mptcp_token_iter_next(const struct net *net, long *s_slot, 791 long *s_num); 792 void mptcp_token_destroy(struct mptcp_sock *msk); 793 794 void mptcp_crypto_key_sha(u64 key, u32 *token, u64 *idsn); 795 796 void mptcp_crypto_hmac_sha(u64 key1, u64 key2, u8 *msg, int len, void *hmac); 797 __sum16 __mptcp_make_csum(u64 data_seq, u32 subflow_seq, u16 data_len, __wsum sum); 798 799 void __init mptcp_pm_init(void); 800 void mptcp_pm_data_init(struct mptcp_sock *msk); 801 void mptcp_pm_data_reset(struct mptcp_sock *msk); 802 int mptcp_pm_parse_addr(struct nlattr *attr, struct genl_info *info, 803 struct mptcp_addr_info *addr); 804 int mptcp_pm_parse_entry(struct nlattr *attr, struct genl_info *info, 805 bool require_family, 806 struct mptcp_pm_addr_entry *entry); 807 bool mptcp_pm_addr_families_match(const struct sock *sk, 808 const struct mptcp_addr_info *loc, 809 const struct mptcp_addr_info *rem); 810 void mptcp_pm_subflow_chk_stale(const struct mptcp_sock *msk, struct sock *ssk); 811 void mptcp_pm_nl_subflow_chk_stale(const struct mptcp_sock *msk, struct sock *ssk); 812 void mptcp_pm_new_connection(struct mptcp_sock *msk, const struct sock *ssk, int server_side); 813 void mptcp_pm_fully_established(struct mptcp_sock *msk, const struct sock *ssk); 814 bool mptcp_pm_allow_new_subflow(struct mptcp_sock *msk); 815 void mptcp_pm_connection_closed(struct mptcp_sock *msk); 816 void mptcp_pm_subflow_established(struct mptcp_sock *msk); 817 bool mptcp_pm_nl_check_work_pending(struct mptcp_sock *msk); 818 void mptcp_pm_subflow_check_next(struct mptcp_sock *msk, const struct sock *ssk, 819 const struct mptcp_subflow_context *subflow); 820 void mptcp_pm_add_addr_received(const struct sock *ssk, 821 const struct mptcp_addr_info *addr); 822 void mptcp_pm_add_addr_echoed(struct mptcp_sock *msk, 823 const struct mptcp_addr_info *addr); 824 void mptcp_pm_add_addr_send_ack(struct mptcp_sock *msk); 825 void mptcp_pm_nl_addr_send_ack(struct mptcp_sock *msk); 826 void mptcp_pm_rm_addr_received(struct mptcp_sock *msk, 827 const struct mptcp_rm_list *rm_list); 828 void mptcp_pm_mp_prio_received(struct sock *sk, u8 bkup); 829 void mptcp_pm_mp_fail_received(struct sock *sk, u64 fail_seq); 830 int mptcp_pm_nl_mp_prio_send_ack(struct mptcp_sock *msk, 831 struct mptcp_addr_info *addr, 832 struct mptcp_addr_info *rem, 833 u8 bkup); 834 bool mptcp_pm_alloc_anno_list(struct mptcp_sock *msk, 835 const struct mptcp_addr_info *addr); 836 void mptcp_pm_free_anno_list(struct mptcp_sock *msk); 837 bool mptcp_pm_sport_in_anno_list(struct mptcp_sock *msk, const struct sock *sk); 838 struct mptcp_pm_add_entry * 839 mptcp_pm_del_add_timer(struct mptcp_sock *msk, 840 const struct mptcp_addr_info *addr, bool check_id); 841 struct mptcp_pm_add_entry * 842 mptcp_lookup_anno_list_by_saddr(const struct mptcp_sock *msk, 843 const struct mptcp_addr_info *addr); 844 int mptcp_pm_get_flags_and_ifindex_by_id(struct mptcp_sock *msk, 845 unsigned int id, 846 u8 *flags, int *ifindex); 847 int mptcp_pm_nl_get_flags_and_ifindex_by_id(struct mptcp_sock *msk, unsigned int id, 848 u8 *flags, int *ifindex); 849 int mptcp_userspace_pm_get_flags_and_ifindex_by_id(struct mptcp_sock *msk, 850 unsigned int id, 851 u8 *flags, int *ifindex); 852 int mptcp_pm_set_flags(struct net *net, struct nlattr *token, 853 struct mptcp_pm_addr_entry *loc, 854 struct mptcp_pm_addr_entry *rem, u8 bkup); 855 int mptcp_pm_nl_set_flags(struct net *net, struct mptcp_pm_addr_entry *addr, u8 bkup); 856 int mptcp_userspace_pm_set_flags(struct net *net, struct nlattr *token, 857 struct mptcp_pm_addr_entry *loc, 858 struct mptcp_pm_addr_entry *rem, u8 bkup); 859 int mptcp_pm_announce_addr(struct mptcp_sock *msk, 860 const struct mptcp_addr_info *addr, 861 bool echo); 862 int mptcp_pm_remove_addr(struct mptcp_sock *msk, const struct mptcp_rm_list *rm_list); 863 int mptcp_pm_remove_subflow(struct mptcp_sock *msk, const struct mptcp_rm_list *rm_list); 864 void mptcp_pm_remove_addrs(struct mptcp_sock *msk, struct list_head *rm_list); 865 void mptcp_pm_remove_addrs_and_subflows(struct mptcp_sock *msk, 866 struct list_head *rm_list); 867 868 void mptcp_free_local_addr_list(struct mptcp_sock *msk); 869 int mptcp_nl_cmd_announce(struct sk_buff *skb, struct genl_info *info); 870 int mptcp_nl_cmd_remove(struct sk_buff *skb, struct genl_info *info); 871 int mptcp_nl_cmd_sf_create(struct sk_buff *skb, struct genl_info *info); 872 int mptcp_nl_cmd_sf_destroy(struct sk_buff *skb, struct genl_info *info); 873 874 void mptcp_event(enum mptcp_event_type type, const struct mptcp_sock *msk, 875 const struct sock *ssk, gfp_t gfp); 876 void mptcp_event_addr_announced(const struct sock *ssk, const struct mptcp_addr_info *info); 877 void mptcp_event_addr_removed(const struct mptcp_sock *msk, u8 id); 878 void mptcp_event_pm_listener(const struct sock *ssk, 879 enum mptcp_event_type event); 880 bool mptcp_userspace_pm_active(const struct mptcp_sock *msk); 881 882 void mptcp_fastopen_gen_msk_ackseq(struct mptcp_sock *msk, struct mptcp_subflow_context *subflow, 883 const struct mptcp_options_received *mp_opt); 884 void mptcp_fastopen_subflow_synack_set_params(struct mptcp_subflow_context *subflow, 885 struct request_sock *req); 886 887 static inline bool mptcp_pm_should_add_signal(struct mptcp_sock *msk) 888 { 889 return READ_ONCE(msk->pm.addr_signal) & 890 (BIT(MPTCP_ADD_ADDR_SIGNAL) | BIT(MPTCP_ADD_ADDR_ECHO)); 891 } 892 893 static inline bool mptcp_pm_should_add_signal_addr(struct mptcp_sock *msk) 894 { 895 return READ_ONCE(msk->pm.addr_signal) & BIT(MPTCP_ADD_ADDR_SIGNAL); 896 } 897 898 static inline bool mptcp_pm_should_add_signal_echo(struct mptcp_sock *msk) 899 { 900 return READ_ONCE(msk->pm.addr_signal) & BIT(MPTCP_ADD_ADDR_ECHO); 901 } 902 903 static inline bool mptcp_pm_should_rm_signal(struct mptcp_sock *msk) 904 { 905 return READ_ONCE(msk->pm.addr_signal) & BIT(MPTCP_RM_ADDR_SIGNAL); 906 } 907 908 static inline bool mptcp_pm_is_userspace(const struct mptcp_sock *msk) 909 { 910 return READ_ONCE(msk->pm.pm_type) == MPTCP_PM_TYPE_USERSPACE; 911 } 912 913 static inline bool mptcp_pm_is_kernel(const struct mptcp_sock *msk) 914 { 915 return READ_ONCE(msk->pm.pm_type) == MPTCP_PM_TYPE_KERNEL; 916 } 917 918 static inline unsigned int mptcp_add_addr_len(int family, bool echo, bool port) 919 { 920 u8 len = TCPOLEN_MPTCP_ADD_ADDR_BASE; 921 922 if (family == AF_INET6) 923 len = TCPOLEN_MPTCP_ADD_ADDR6_BASE; 924 if (!echo) 925 len += MPTCPOPT_THMAC_LEN; 926 /* account for 2 trailing 'nop' options */ 927 if (port) 928 len += TCPOLEN_MPTCP_PORT_LEN + TCPOLEN_MPTCP_PORT_ALIGN; 929 930 return len; 931 } 932 933 static inline int mptcp_rm_addr_len(const struct mptcp_rm_list *rm_list) 934 { 935 if (rm_list->nr == 0 || rm_list->nr > MPTCP_RM_IDS_MAX) 936 return -EINVAL; 937 938 return TCPOLEN_MPTCP_RM_ADDR_BASE + roundup(rm_list->nr - 1, 4) + 1; 939 } 940 941 bool mptcp_pm_add_addr_signal(struct mptcp_sock *msk, const struct sk_buff *skb, 942 unsigned int opt_size, unsigned int remaining, 943 struct mptcp_addr_info *addr, bool *echo, 944 bool *drop_other_suboptions); 945 bool mptcp_pm_rm_addr_signal(struct mptcp_sock *msk, unsigned int remaining, 946 struct mptcp_rm_list *rm_list); 947 int mptcp_pm_get_local_id(struct mptcp_sock *msk, struct sock_common *skc); 948 int mptcp_pm_nl_get_local_id(struct mptcp_sock *msk, struct mptcp_addr_info *skc); 949 int mptcp_userspace_pm_get_local_id(struct mptcp_sock *msk, struct mptcp_addr_info *skc); 950 951 void __init mptcp_pm_nl_init(void); 952 void mptcp_pm_nl_work(struct mptcp_sock *msk); 953 void mptcp_pm_nl_rm_subflow_received(struct mptcp_sock *msk, 954 const struct mptcp_rm_list *rm_list); 955 unsigned int mptcp_pm_get_add_addr_signal_max(const struct mptcp_sock *msk); 956 unsigned int mptcp_pm_get_add_addr_accept_max(const struct mptcp_sock *msk); 957 unsigned int mptcp_pm_get_subflows_max(const struct mptcp_sock *msk); 958 unsigned int mptcp_pm_get_local_addr_max(const struct mptcp_sock *msk); 959 960 /* called under PM lock */ 961 static inline void __mptcp_pm_close_subflow(struct mptcp_sock *msk) 962 { 963 if (--msk->pm.subflows < mptcp_pm_get_subflows_max(msk)) 964 WRITE_ONCE(msk->pm.accept_subflow, true); 965 } 966 967 static inline void mptcp_pm_close_subflow(struct mptcp_sock *msk) 968 { 969 spin_lock_bh(&msk->pm.lock); 970 __mptcp_pm_close_subflow(msk); 971 spin_unlock_bh(&msk->pm.lock); 972 } 973 974 void mptcp_sockopt_sync(struct mptcp_sock *msk, struct sock *ssk); 975 void mptcp_sockopt_sync_locked(struct mptcp_sock *msk, struct sock *ssk); 976 977 static inline struct mptcp_ext *mptcp_get_ext(const struct sk_buff *skb) 978 { 979 return (struct mptcp_ext *)skb_ext_find(skb, SKB_EXT_MPTCP); 980 } 981 982 void mptcp_diag_subflow_init(struct tcp_ulp_ops *ops); 983 984 static inline bool __mptcp_check_fallback(const struct mptcp_sock *msk) 985 { 986 return test_bit(MPTCP_FALLBACK_DONE, &msk->flags); 987 } 988 989 static inline bool mptcp_check_fallback(const struct sock *sk) 990 { 991 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 992 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 993 994 return __mptcp_check_fallback(msk); 995 } 996 997 static inline void __mptcp_do_fallback(struct mptcp_sock *msk) 998 { 999 if (test_bit(MPTCP_FALLBACK_DONE, &msk->flags)) { 1000 pr_debug("TCP fallback already done (msk=%p)", msk); 1001 return; 1002 } 1003 set_bit(MPTCP_FALLBACK_DONE, &msk->flags); 1004 } 1005 1006 static inline void mptcp_do_fallback(struct sock *ssk) 1007 { 1008 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 1009 struct sock *sk = subflow->conn; 1010 struct mptcp_sock *msk; 1011 1012 msk = mptcp_sk(sk); 1013 __mptcp_do_fallback(msk); 1014 if (READ_ONCE(msk->snd_data_fin_enable) && !(ssk->sk_shutdown & SEND_SHUTDOWN)) { 1015 gfp_t saved_allocation = ssk->sk_allocation; 1016 1017 /* we are in a atomic (BH) scope, override ssk default for data 1018 * fin allocation 1019 */ 1020 ssk->sk_allocation = GFP_ATOMIC; 1021 ssk->sk_shutdown |= SEND_SHUTDOWN; 1022 tcp_shutdown(ssk, SEND_SHUTDOWN); 1023 ssk->sk_allocation = saved_allocation; 1024 } 1025 } 1026 1027 #define pr_fallback(a) pr_debug("%s:fallback to TCP (msk=%p)", __func__, a) 1028 1029 static inline bool mptcp_check_infinite_map(struct sk_buff *skb) 1030 { 1031 struct mptcp_ext *mpext; 1032 1033 mpext = skb ? mptcp_get_ext(skb) : NULL; 1034 if (mpext && mpext->infinite_map) 1035 return true; 1036 1037 return false; 1038 } 1039 1040 static inline bool is_active_ssk(struct mptcp_subflow_context *subflow) 1041 { 1042 return (subflow->request_mptcp || subflow->request_join); 1043 } 1044 1045 static inline bool subflow_simultaneous_connect(struct sock *sk) 1046 { 1047 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 1048 1049 return sk->sk_state == TCP_ESTABLISHED && 1050 is_active_ssk(subflow) && 1051 !subflow->conn_finished; 1052 } 1053 1054 #ifdef CONFIG_SYN_COOKIES 1055 void subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock *subflow_req, 1056 struct sk_buff *skb); 1057 bool mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock *subflow_req, 1058 struct sk_buff *skb); 1059 void __init mptcp_join_cookie_init(void); 1060 #else 1061 static inline void 1062 subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock *subflow_req, 1063 struct sk_buff *skb) {} 1064 static inline bool 1065 mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock *subflow_req, 1066 struct sk_buff *skb) 1067 { 1068 return false; 1069 } 1070 1071 static inline void mptcp_join_cookie_init(void) {} 1072 #endif 1073 1074 #endif /* __MPTCP_PROTOCOL_H */ 1075