1 /* SPDX-License-Identifier: GPL-2.0 */
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #ifndef __MPTCP_PROTOCOL_H
8 #define __MPTCP_PROTOCOL_H
10 #include <linux/random.h>
12 #include <net/inet_connection_sock.h>
14 #define MPTCP_SUPPORTED_VERSION 1
16 /* MPTCP option bits */
17 #define OPTION_MPTCP_MPC_SYN BIT(0)
18 #define OPTION_MPTCP_MPC_SYNACK BIT(1)
19 #define OPTION_MPTCP_MPC_ACK BIT(2)
20 #define OPTION_MPTCP_MPJ_SYN BIT(3)
21 #define OPTION_MPTCP_MPJ_SYNACK BIT(4)
22 #define OPTION_MPTCP_MPJ_ACK BIT(5)
23 #define OPTION_MPTCP_ADD_ADDR BIT(6)
24 #define OPTION_MPTCP_ADD_ADDR6 BIT(7)
25 #define OPTION_MPTCP_RM_ADDR BIT(8)
27 /* MPTCP option subtypes */
28 #define MPTCPOPT_MP_CAPABLE 0
29 #define MPTCPOPT_MP_JOIN 1
30 #define MPTCPOPT_DSS 2
31 #define MPTCPOPT_ADD_ADDR 3
32 #define MPTCPOPT_RM_ADDR 4
33 #define MPTCPOPT_MP_PRIO 5
34 #define MPTCPOPT_MP_FAIL 6
35 #define MPTCPOPT_MP_FASTCLOSE 7
37 /* MPTCP suboption lengths */
38 #define TCPOLEN_MPTCP_MPC_SYN 4
39 #define TCPOLEN_MPTCP_MPC_SYNACK 12
40 #define TCPOLEN_MPTCP_MPC_ACK 20
41 #define TCPOLEN_MPTCP_MPC_ACK_DATA 22
42 #define TCPOLEN_MPTCP_MPJ_SYN 12
43 #define TCPOLEN_MPTCP_MPJ_SYNACK 16
44 #define TCPOLEN_MPTCP_MPJ_ACK 24
45 #define TCPOLEN_MPTCP_DSS_BASE 4
46 #define TCPOLEN_MPTCP_DSS_ACK32 4
47 #define TCPOLEN_MPTCP_DSS_ACK64 8
48 #define TCPOLEN_MPTCP_DSS_MAP32 10
49 #define TCPOLEN_MPTCP_DSS_MAP64 14
50 #define TCPOLEN_MPTCP_DSS_CHECKSUM 2
51 #define TCPOLEN_MPTCP_ADD_ADDR 16
52 #define TCPOLEN_MPTCP_ADD_ADDR_PORT 20
53 #define TCPOLEN_MPTCP_ADD_ADDR_BASE 8
54 #define TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT 12
55 #define TCPOLEN_MPTCP_ADD_ADDR6 28
56 #define TCPOLEN_MPTCP_ADD_ADDR6_PORT 32
57 #define TCPOLEN_MPTCP_ADD_ADDR6_BASE 20
58 #define TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT 24
59 #define TCPOLEN_MPTCP_PORT_LEN 4
60 #define TCPOLEN_MPTCP_RM_ADDR_BASE 4
62 /* MPTCP MP_JOIN flags */
63 #define MPTCPOPT_BACKUP BIT(0)
64 #define MPTCPOPT_HMAC_LEN 20
65 #define MPTCPOPT_THMAC_LEN 8
67 /* MPTCP MP_CAPABLE flags */
68 #define MPTCP_VERSION_MASK (0x0F)
69 #define MPTCP_CAP_CHECKSUM_REQD BIT(7)
70 #define MPTCP_CAP_EXTENSIBILITY BIT(6)
71 #define MPTCP_CAP_HMAC_SHA256 BIT(0)
72 #define MPTCP_CAP_FLAG_MASK (0x3F)
75 #define MPTCP_DSS_DATA_FIN BIT(4)
76 #define MPTCP_DSS_DSN64 BIT(3)
77 #define MPTCP_DSS_HAS_MAP BIT(2)
78 #define MPTCP_DSS_ACK64 BIT(1)
79 #define MPTCP_DSS_HAS_ACK BIT(0)
80 #define MPTCP_DSS_FLAG_MASK (0x1F)
82 /* MPTCP ADD_ADDR flags */
83 #define MPTCP_ADDR_ECHO BIT(0)
84 #define MPTCP_ADDR_IPVERSION_4 4
85 #define MPTCP_ADDR_IPVERSION_6 6
87 /* MPTCP socket flags */
88 #define MPTCP_DATA_READY 0
89 #define MPTCP_NOSPACE 1
90 #define MPTCP_WORK_RTX 2
91 #define MPTCP_WORK_EOF 3
92 #define MPTCP_FALLBACK_DONE 4
93 #define MPTCP_WORK_CLOSE_SUBFLOW 5
94 #define MPTCP_PUSH_PENDING 6
95 #define MPTCP_CLEAN_UNA 7
97 static inline bool before64(__u64 seq1, __u64 seq2)
99 return (__s64)(seq1 - seq2) < 0;
102 #define after64(seq2, seq1) before64(seq1, seq2)
104 struct mptcp_options_received {
135 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
136 struct in6_addr addr6;
143 static inline __be32 mptcp_option(u8 subopt, u8 len, u8 nib, u8 field)
145 return htonl((TCPOPT_MPTCP << 24) | (len << 16) | (subopt << 12) |
146 ((nib & 0xF) << 8) | field);
149 struct mptcp_addr_info {
157 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
158 struct in6_addr addr6;
163 enum mptcp_pm_status {
164 MPTCP_PM_ADD_ADDR_RECEIVED,
165 MPTCP_PM_ADD_ADDR_SEND_ACK,
166 MPTCP_PM_RM_ADDR_RECEIVED,
167 MPTCP_PM_ESTABLISHED,
168 MPTCP_PM_SUBFLOW_ESTABLISHED,
171 enum mptcp_add_addr_status {
172 MPTCP_ADD_ADDR_SIGNAL,
177 struct mptcp_pm_data {
178 struct mptcp_addr_info local;
179 struct mptcp_addr_info remote;
180 struct list_head anno_list;
182 spinlock_t lock; /*protects the whole PM data */
190 u8 add_addr_signaled;
191 u8 add_addr_accepted;
194 u8 add_addr_signal_max;
195 u8 add_addr_accept_max;
202 struct mptcp_data_frag {
203 struct list_head list;
212 /* MPTCP connection sock */
214 /* inet_connection_sock must be the first member */
215 struct inet_connection_sock sk;
222 u64 rcv_data_fin_seq;
224 struct sock *last_snd;
229 unsigned long timer_ival;
235 bool fully_established;
237 bool snd_data_fin_enable;
238 bool use_64bit_ack; /* Set when we received a 64-bit DSN */
239 spinlock_t join_list_lock;
240 struct sock *ack_hint;
241 struct work_struct work;
242 struct sk_buff *ooo_last_skb;
243 struct rb_root out_of_order_queue;
244 struct sk_buff_head receive_queue;
245 struct sk_buff_head skb_tx_cache; /* this is wmem accounted */
248 struct list_head conn_list;
249 struct list_head rtx_queue;
250 struct mptcp_data_frag *first_pending;
251 struct list_head join_list;
252 struct socket *subflow; /* outgoing connect/listener/!mp_capable */
254 struct mptcp_pm_data pm;
256 u32 space; /* bytes copied in last measurement window */
257 u32 copied; /* bytes copied in this measurement window */
258 u64 time; /* start time of measurement window */
259 u64 rtt_us; /* last maximum rtt of subflows */
263 #define mptcp_lock_sock(___sk, cb) do { \
264 struct sock *__sk = (___sk); /* silence macro reuse warning */ \
266 spin_lock_bh(&__sk->sk_lock.slock); \
267 if (__sk->sk_lock.owned) \
270 __sk->sk_lock.owned = 1; \
271 spin_unlock(&__sk->sk_lock.slock); \
272 mutex_acquire(&__sk->sk_lock.dep_map, 0, 0, _RET_IP_); \
276 #define mptcp_data_lock(sk) spin_lock_bh(&(sk)->sk_lock.slock)
277 #define mptcp_data_unlock(sk) spin_unlock_bh(&(sk)->sk_lock.slock)
279 #define mptcp_for_each_subflow(__msk, __subflow) \
280 list_for_each_entry(__subflow, &((__msk)->conn_list), node)
282 static inline struct mptcp_sock *mptcp_sk(const struct sock *sk)
284 return (struct mptcp_sock *)sk;
287 static inline int __mptcp_space(const struct sock *sk)
289 return tcp_space(sk) + READ_ONCE(mptcp_sk(sk)->rmem_pending);
292 static inline struct mptcp_data_frag *mptcp_send_head(const struct sock *sk)
294 const struct mptcp_sock *msk = mptcp_sk(sk);
296 return READ_ONCE(msk->first_pending);
299 static inline struct mptcp_data_frag *mptcp_send_next(struct sock *sk)
301 struct mptcp_sock *msk = mptcp_sk(sk);
302 struct mptcp_data_frag *cur;
304 cur = msk->first_pending;
305 return list_is_last(&cur->list, &msk->rtx_queue) ? NULL :
306 list_next_entry(cur, list);
309 static inline struct mptcp_data_frag *mptcp_pending_tail(const struct sock *sk)
311 struct mptcp_sock *msk = mptcp_sk(sk);
313 if (!msk->first_pending)
316 if (WARN_ON_ONCE(list_empty(&msk->rtx_queue)))
319 return list_last_entry(&msk->rtx_queue, struct mptcp_data_frag, list);
322 static inline struct mptcp_data_frag *mptcp_rtx_tail(const struct sock *sk)
324 struct mptcp_sock *msk = mptcp_sk(sk);
326 if (!before64(msk->snd_nxt, READ_ONCE(msk->snd_una)))
329 return list_last_entry(&msk->rtx_queue, struct mptcp_data_frag, list);
332 static inline struct mptcp_data_frag *mptcp_rtx_head(const struct sock *sk)
334 struct mptcp_sock *msk = mptcp_sk(sk);
336 return list_first_entry_or_null(&msk->rtx_queue, struct mptcp_data_frag, list);
339 struct mptcp_subflow_request_sock {
340 struct tcp_request_sock sk;
353 struct mptcp_sock *msk;
354 struct hlist_nulls_node token_node;
357 static inline struct mptcp_subflow_request_sock *
358 mptcp_subflow_rsk(const struct request_sock *rsk)
360 return (struct mptcp_subflow_request_sock *)rsk;
363 enum mptcp_data_avail {
364 MPTCP_SUBFLOW_NODATA,
365 MPTCP_SUBFLOW_DATA_AVAIL,
366 MPTCP_SUBFLOW_OOO_DATA
369 /* MPTCP subflow context */
370 struct mptcp_subflow_context {
371 struct list_head node;/* conn_list of subflows */
382 u32 request_mptcp : 1, /* send MP_CAPABLE */
383 request_join : 1, /* send MP_JOIN */
385 mp_capable : 1, /* remote is MPTCP capable */
386 mp_join : 1, /* remote is JOINing */
387 fully_established : 1, /* path validated */
388 pm_notified : 1, /* PM hook called for established status */
394 can_ack : 1, /* only after processing the remote a key */
395 disposable : 1; /* ctx can be free at ulp release time */
396 enum mptcp_data_avail data_avail;
401 u8 hmac[MPTCPOPT_HMAC_LEN];
405 struct sock *tcp_sock; /* tcp sk backpointer */
406 struct sock *conn; /* parent mptcp_sock */
407 const struct inet_connection_sock_af_ops *icsk_af_ops;
408 void (*tcp_data_ready)(struct sock *sk);
409 void (*tcp_state_change)(struct sock *sk);
410 void (*tcp_write_space)(struct sock *sk);
415 static inline struct mptcp_subflow_context *
416 mptcp_subflow_ctx(const struct sock *sk)
418 struct inet_connection_sock *icsk = inet_csk(sk);
420 /* Use RCU on icsk_ulp_data only for sock diag code */
421 return (__force struct mptcp_subflow_context *)icsk->icsk_ulp_data;
424 static inline struct sock *
425 mptcp_subflow_tcp_sock(const struct mptcp_subflow_context *subflow)
427 return subflow->tcp_sock;
431 mptcp_subflow_get_map_offset(const struct mptcp_subflow_context *subflow)
433 return tcp_sk(mptcp_subflow_tcp_sock(subflow))->copied_seq -
434 subflow->ssn_offset -
435 subflow->map_subflow_seq;
439 mptcp_subflow_get_mapped_dsn(const struct mptcp_subflow_context *subflow)
441 return subflow->map_seq + mptcp_subflow_get_map_offset(subflow);
444 static inline void mptcp_add_pending_subflow(struct mptcp_sock *msk,
445 struct mptcp_subflow_context *subflow)
447 sock_hold(mptcp_subflow_tcp_sock(subflow));
448 spin_lock_bh(&msk->join_list_lock);
449 list_add_tail(&subflow->node, &msk->join_list);
450 spin_unlock_bh(&msk->join_list_lock);
453 int mptcp_is_enabled(struct net *net);
454 unsigned int mptcp_get_add_addr_timeout(struct net *net);
455 void mptcp_subflow_fully_established(struct mptcp_subflow_context *subflow,
456 struct mptcp_options_received *mp_opt);
457 bool mptcp_subflow_data_available(struct sock *sk);
458 void __init mptcp_subflow_init(void);
459 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how);
460 void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
461 struct mptcp_subflow_context *subflow);
462 void mptcp_subflow_reset(struct sock *ssk);
464 /* called with sk socket lock held */
465 int __mptcp_subflow_connect(struct sock *sk, const struct mptcp_addr_info *loc,
466 const struct mptcp_addr_info *remote);
467 int mptcp_subflow_create_socket(struct sock *sk, struct socket **new_sock);
469 static inline void mptcp_subflow_tcp_fallback(struct sock *sk,
470 struct mptcp_subflow_context *ctx)
472 sk->sk_data_ready = ctx->tcp_data_ready;
473 sk->sk_state_change = ctx->tcp_state_change;
474 sk->sk_write_space = ctx->tcp_write_space;
476 inet_csk(sk)->icsk_af_ops = ctx->icsk_af_ops;
479 void __init mptcp_proto_init(void);
480 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
481 int __init mptcp_proto_v6_init(void);
484 struct sock *mptcp_sk_clone(const struct sock *sk,
485 const struct mptcp_options_received *mp_opt,
486 struct request_sock *req);
487 void mptcp_get_options(const struct sk_buff *skb,
488 struct mptcp_options_received *mp_opt);
490 void mptcp_finish_connect(struct sock *sk);
491 static inline bool mptcp_is_fully_established(struct sock *sk)
493 return inet_sk_state_load(sk) == TCP_ESTABLISHED &&
494 READ_ONCE(mptcp_sk(sk)->fully_established);
496 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk);
497 void mptcp_data_ready(struct sock *sk, struct sock *ssk);
498 bool mptcp_finish_join(struct sock *sk);
499 bool mptcp_schedule_work(struct sock *sk);
500 void __mptcp_wnd_updated(struct sock *sk, struct sock *ssk);
501 void __mptcp_data_acked(struct sock *sk);
502 void mptcp_subflow_eof(struct sock *sk);
503 bool mptcp_update_rcv_data_fin(struct mptcp_sock *msk, u64 data_fin_seq, bool use_64bit);
504 void __mptcp_flush_join_list(struct mptcp_sock *msk);
505 static inline bool mptcp_data_fin_enabled(const struct mptcp_sock *msk)
507 return READ_ONCE(msk->snd_data_fin_enable) &&
508 READ_ONCE(msk->write_seq) == READ_ONCE(msk->snd_nxt);
511 void mptcp_destroy_common(struct mptcp_sock *msk);
513 void __init mptcp_token_init(void);
514 static inline void mptcp_token_init_request(struct request_sock *req)
516 mptcp_subflow_rsk(req)->token_node.pprev = NULL;
519 int mptcp_token_new_request(struct request_sock *req);
520 void mptcp_token_destroy_request(struct request_sock *req);
521 int mptcp_token_new_connect(struct sock *sk);
522 void mptcp_token_accept(struct mptcp_subflow_request_sock *r,
523 struct mptcp_sock *msk);
524 bool mptcp_token_exists(u32 token);
525 struct mptcp_sock *mptcp_token_get_sock(u32 token);
526 struct mptcp_sock *mptcp_token_iter_next(const struct net *net, long *s_slot,
528 void mptcp_token_destroy(struct mptcp_sock *msk);
530 void mptcp_crypto_key_sha(u64 key, u32 *token, u64 *idsn);
532 void mptcp_crypto_hmac_sha(u64 key1, u64 key2, u8 *msg, int len, void *hmac);
534 void __init mptcp_pm_init(void);
535 void mptcp_pm_data_init(struct mptcp_sock *msk);
536 void mptcp_pm_new_connection(struct mptcp_sock *msk, int server_side);
537 void mptcp_pm_fully_established(struct mptcp_sock *msk);
538 bool mptcp_pm_allow_new_subflow(struct mptcp_sock *msk);
539 void mptcp_pm_connection_closed(struct mptcp_sock *msk);
540 void mptcp_pm_subflow_established(struct mptcp_sock *msk,
541 struct mptcp_subflow_context *subflow);
542 void mptcp_pm_subflow_closed(struct mptcp_sock *msk, u8 id);
543 void mptcp_pm_add_addr_received(struct mptcp_sock *msk,
544 const struct mptcp_addr_info *addr);
545 void mptcp_pm_add_addr_send_ack(struct mptcp_sock *msk);
546 void mptcp_pm_rm_addr_received(struct mptcp_sock *msk, u8 rm_id);
547 void mptcp_pm_free_anno_list(struct mptcp_sock *msk);
548 struct mptcp_pm_add_entry *
549 mptcp_pm_del_add_timer(struct mptcp_sock *msk,
550 struct mptcp_addr_info *addr);
552 int mptcp_pm_announce_addr(struct mptcp_sock *msk,
553 const struct mptcp_addr_info *addr,
555 int mptcp_pm_remove_addr(struct mptcp_sock *msk, u8 local_id);
556 int mptcp_pm_remove_subflow(struct mptcp_sock *msk, u8 local_id);
558 static inline bool mptcp_pm_should_add_signal(struct mptcp_sock *msk)
560 return READ_ONCE(msk->pm.add_addr_signal) & BIT(MPTCP_ADD_ADDR_SIGNAL);
563 static inline bool mptcp_pm_should_add_signal_echo(struct mptcp_sock *msk)
565 return READ_ONCE(msk->pm.add_addr_signal) & BIT(MPTCP_ADD_ADDR_ECHO);
568 static inline bool mptcp_pm_should_add_signal_ipv6(struct mptcp_sock *msk)
570 return READ_ONCE(msk->pm.add_addr_signal) & BIT(MPTCP_ADD_ADDR_IPV6);
573 static inline bool mptcp_pm_should_rm_signal(struct mptcp_sock *msk)
575 return READ_ONCE(msk->pm.rm_addr_signal);
578 static inline unsigned int mptcp_add_addr_len(int family, bool echo)
580 if (family == AF_INET)
581 return echo ? TCPOLEN_MPTCP_ADD_ADDR_BASE
582 : TCPOLEN_MPTCP_ADD_ADDR;
583 return echo ? TCPOLEN_MPTCP_ADD_ADDR6_BASE : TCPOLEN_MPTCP_ADD_ADDR6;
586 bool mptcp_pm_add_addr_signal(struct mptcp_sock *msk, unsigned int remaining,
587 struct mptcp_addr_info *saddr, bool *echo);
588 bool mptcp_pm_rm_addr_signal(struct mptcp_sock *msk, unsigned int remaining,
590 int mptcp_pm_get_local_id(struct mptcp_sock *msk, struct sock_common *skc);
592 void __init mptcp_pm_nl_init(void);
593 void mptcp_pm_nl_data_init(struct mptcp_sock *msk);
594 void mptcp_pm_nl_fully_established(struct mptcp_sock *msk);
595 void mptcp_pm_nl_subflow_established(struct mptcp_sock *msk);
596 void mptcp_pm_nl_add_addr_received(struct mptcp_sock *msk);
597 void mptcp_pm_nl_add_addr_send_ack(struct mptcp_sock *msk);
598 void mptcp_pm_nl_rm_addr_received(struct mptcp_sock *msk);
599 void mptcp_pm_nl_rm_subflow_received(struct mptcp_sock *msk, u8 rm_id);
600 int mptcp_pm_nl_get_local_id(struct mptcp_sock *msk, struct sock_common *skc);
602 static inline struct mptcp_ext *mptcp_get_ext(struct sk_buff *skb)
604 return (struct mptcp_ext *)skb_ext_find(skb, SKB_EXT_MPTCP);
607 void mptcp_diag_subflow_init(struct tcp_ulp_ops *ops);
609 static inline bool __mptcp_check_fallback(const struct mptcp_sock *msk)
611 return test_bit(MPTCP_FALLBACK_DONE, &msk->flags);
614 static inline bool mptcp_check_fallback(const struct sock *sk)
616 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
617 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
619 return __mptcp_check_fallback(msk);
622 static inline void __mptcp_do_fallback(struct mptcp_sock *msk)
624 if (test_bit(MPTCP_FALLBACK_DONE, &msk->flags)) {
625 pr_debug("TCP fallback already done (msk=%p)", msk);
628 set_bit(MPTCP_FALLBACK_DONE, &msk->flags);
631 static inline void mptcp_do_fallback(struct sock *sk)
633 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
634 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
636 __mptcp_do_fallback(msk);
639 #define pr_fallback(a) pr_debug("%s:fallback to TCP (msk=%p)", __func__, a)
641 static inline bool subflow_simultaneous_connect(struct sock *sk)
643 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
644 struct sock *parent = subflow->conn;
646 return sk->sk_state == TCP_ESTABLISHED &&
647 !mptcp_sk(parent)->pm.server_side &&
648 !subflow->conn_finished;
651 #ifdef CONFIG_SYN_COOKIES
652 void subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock *subflow_req,
653 struct sk_buff *skb);
654 bool mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock *subflow_req,
655 struct sk_buff *skb);
656 void __init mptcp_join_cookie_init(void);
659 subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock *subflow_req,
660 struct sk_buff *skb) {}
662 mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock *subflow_req,
668 static inline void mptcp_join_cookie_init(void) {}
671 #endif /* __MPTCP_PROTOCOL_H */