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 18
53 #define TCPOLEN_MPTCP_ADD_ADDR_BASE 8
54 #define TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT 10
55 #define TCPOLEN_MPTCP_ADD_ADDR6 28
56 #define TCPOLEN_MPTCP_ADD_ADDR6_PORT 30
57 #define TCPOLEN_MPTCP_ADD_ADDR6_BASE 20
58 #define TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT 22
59 #define TCPOLEN_MPTCP_PORT_LEN 2
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
95 static inline bool before64(__u64 seq1, __u64 seq2)
97 return (__s64)(seq1 - seq2) < 0;
100 #define after64(seq2, seq1) before64(seq1, seq2)
102 struct mptcp_options_received {
133 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
134 struct in6_addr addr6;
141 static inline __be32 mptcp_option(u8 subopt, u8 len, u8 nib, u8 field)
143 return htonl((TCPOPT_MPTCP << 24) | (len << 16) | (subopt << 12) |
144 ((nib & 0xF) << 8) | field);
147 struct mptcp_addr_info {
155 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
156 struct in6_addr addr6;
161 enum mptcp_pm_status {
162 MPTCP_PM_ADD_ADDR_RECEIVED,
163 MPTCP_PM_ADD_ADDR_SEND_ACK,
164 MPTCP_PM_RM_ADDR_RECEIVED,
165 MPTCP_PM_ESTABLISHED,
166 MPTCP_PM_SUBFLOW_ESTABLISHED,
169 enum mptcp_add_addr_status {
170 MPTCP_ADD_ADDR_SIGNAL,
175 struct mptcp_pm_data {
176 struct mptcp_addr_info local;
177 struct mptcp_addr_info remote;
178 struct list_head anno_list;
180 spinlock_t lock; /*protects the whole PM data */
188 u8 add_addr_signaled;
189 u8 add_addr_accepted;
192 u8 add_addr_signal_max;
193 u8 add_addr_accept_max;
200 struct mptcp_data_frag {
201 struct list_head list;
210 /* MPTCP connection sock */
212 /* inet_connection_sock must be the first member */
213 struct inet_connection_sock sk;
220 u64 rcv_data_fin_seq;
221 struct sock *last_snd;
226 unsigned long timer_ival;
231 bool fully_established;
233 bool snd_data_fin_enable;
234 bool use_64bit_ack; /* Set when we received a 64-bit DSN */
235 spinlock_t join_list_lock;
236 struct sock *ack_hint;
237 struct work_struct work;
238 struct sk_buff *ooo_last_skb;
239 struct rb_root out_of_order_queue;
240 struct list_head conn_list;
241 struct list_head rtx_queue;
242 struct mptcp_data_frag *first_pending;
243 struct list_head join_list;
244 struct skb_ext *cached_ext; /* for the next sendmsg */
245 struct socket *subflow; /* outgoing connect/listener/!mp_capable */
247 struct mptcp_pm_data pm;
249 u32 space; /* bytes copied in last measurement window */
250 u32 copied; /* bytes copied in this measurement window */
251 u64 time; /* start time of measurement window */
252 u64 rtt_us; /* last maximum rtt of subflows */
256 #define mptcp_lock_sock(___sk, cb) do { \
257 struct sock *__sk = (___sk); /* silence macro reuse warning */ \
259 spin_lock_bh(&__sk->sk_lock.slock); \
260 if (__sk->sk_lock.owned) \
263 __sk->sk_lock.owned = 1; \
264 spin_unlock(&__sk->sk_lock.slock); \
265 mutex_acquire(&__sk->sk_lock.dep_map, 0, 0, _RET_IP_); \
269 #define mptcp_for_each_subflow(__msk, __subflow) \
270 list_for_each_entry(__subflow, &((__msk)->conn_list), node)
272 static inline struct mptcp_sock *mptcp_sk(const struct sock *sk)
274 return (struct mptcp_sock *)sk;
277 static inline int __mptcp_space(const struct sock *sk)
279 return tcp_space(sk) + READ_ONCE(mptcp_sk(sk)->rmem_pending);
282 static inline struct mptcp_data_frag *mptcp_send_head(const struct sock *sk)
284 const struct mptcp_sock *msk = mptcp_sk(sk);
286 return READ_ONCE(msk->first_pending);
289 static inline struct mptcp_data_frag *mptcp_send_next(struct sock *sk)
291 struct mptcp_sock *msk = mptcp_sk(sk);
292 struct mptcp_data_frag *cur;
294 cur = msk->first_pending;
295 return list_is_last(&cur->list, &msk->rtx_queue) ? NULL :
296 list_next_entry(cur, list);
299 static inline struct mptcp_data_frag *mptcp_pending_tail(const struct sock *sk)
301 struct mptcp_sock *msk = mptcp_sk(sk);
303 if (!msk->first_pending)
306 if (WARN_ON_ONCE(list_empty(&msk->rtx_queue)))
309 return list_last_entry(&msk->rtx_queue, struct mptcp_data_frag, list);
312 static inline struct mptcp_data_frag *mptcp_rtx_tail(const struct sock *sk)
314 struct mptcp_sock *msk = mptcp_sk(sk);
316 if (!before64(msk->snd_nxt, atomic64_read(&msk->snd_una)))
319 return list_last_entry(&msk->rtx_queue, struct mptcp_data_frag, list);
322 static inline struct mptcp_data_frag *mptcp_rtx_head(const struct sock *sk)
324 struct mptcp_sock *msk = mptcp_sk(sk);
326 return list_first_entry_or_null(&msk->rtx_queue, struct mptcp_data_frag, list);
329 struct mptcp_subflow_request_sock {
330 struct tcp_request_sock sk;
343 struct mptcp_sock *msk;
344 struct hlist_nulls_node token_node;
347 static inline struct mptcp_subflow_request_sock *
348 mptcp_subflow_rsk(const struct request_sock *rsk)
350 return (struct mptcp_subflow_request_sock *)rsk;
353 enum mptcp_data_avail {
354 MPTCP_SUBFLOW_NODATA,
355 MPTCP_SUBFLOW_DATA_AVAIL,
356 MPTCP_SUBFLOW_OOO_DATA
359 /* MPTCP subflow context */
360 struct mptcp_subflow_context {
361 struct list_head node;/* conn_list of subflows */
372 u32 request_mptcp : 1, /* send MP_CAPABLE */
373 request_join : 1, /* send MP_JOIN */
375 mp_capable : 1, /* remote is MPTCP capable */
376 mp_join : 1, /* remote is JOINing */
377 fully_established : 1, /* path validated */
378 pm_notified : 1, /* PM hook called for established status */
384 can_ack : 1, /* only after processing the remote a key */
385 disposable : 1; /* ctx can be free at ulp release time */
386 enum mptcp_data_avail data_avail;
391 u8 hmac[MPTCPOPT_HMAC_LEN];
395 struct sock *tcp_sock; /* tcp sk backpointer */
396 struct sock *conn; /* parent mptcp_sock */
397 const struct inet_connection_sock_af_ops *icsk_af_ops;
398 void (*tcp_data_ready)(struct sock *sk);
399 void (*tcp_state_change)(struct sock *sk);
400 void (*tcp_write_space)(struct sock *sk);
405 static inline struct mptcp_subflow_context *
406 mptcp_subflow_ctx(const struct sock *sk)
408 struct inet_connection_sock *icsk = inet_csk(sk);
410 /* Use RCU on icsk_ulp_data only for sock diag code */
411 return (__force struct mptcp_subflow_context *)icsk->icsk_ulp_data;
414 static inline struct sock *
415 mptcp_subflow_tcp_sock(const struct mptcp_subflow_context *subflow)
417 return subflow->tcp_sock;
421 mptcp_subflow_get_map_offset(const struct mptcp_subflow_context *subflow)
423 return tcp_sk(mptcp_subflow_tcp_sock(subflow))->copied_seq -
424 subflow->ssn_offset -
425 subflow->map_subflow_seq;
429 mptcp_subflow_get_mapped_dsn(const struct mptcp_subflow_context *subflow)
431 return subflow->map_seq + mptcp_subflow_get_map_offset(subflow);
434 static inline void mptcp_add_pending_subflow(struct mptcp_sock *msk,
435 struct mptcp_subflow_context *subflow)
437 sock_hold(mptcp_subflow_tcp_sock(subflow));
438 spin_lock_bh(&msk->join_list_lock);
439 list_add_tail(&subflow->node, &msk->join_list);
440 spin_unlock_bh(&msk->join_list_lock);
443 int mptcp_is_enabled(struct net *net);
444 unsigned int mptcp_get_add_addr_timeout(struct net *net);
445 void mptcp_subflow_fully_established(struct mptcp_subflow_context *subflow,
446 struct mptcp_options_received *mp_opt);
447 bool mptcp_subflow_data_available(struct sock *sk);
448 void __init mptcp_subflow_init(void);
449 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how);
450 void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
451 struct mptcp_subflow_context *subflow);
452 void mptcp_subflow_reset(struct sock *ssk);
454 /* called with sk socket lock held */
455 int __mptcp_subflow_connect(struct sock *sk, const struct mptcp_addr_info *loc,
456 const struct mptcp_addr_info *remote);
457 int mptcp_subflow_create_socket(struct sock *sk, struct socket **new_sock);
459 static inline void mptcp_subflow_tcp_fallback(struct sock *sk,
460 struct mptcp_subflow_context *ctx)
462 sk->sk_data_ready = ctx->tcp_data_ready;
463 sk->sk_state_change = ctx->tcp_state_change;
464 sk->sk_write_space = ctx->tcp_write_space;
466 inet_csk(sk)->icsk_af_ops = ctx->icsk_af_ops;
469 void __init mptcp_proto_init(void);
470 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
471 int __init mptcp_proto_v6_init(void);
474 struct sock *mptcp_sk_clone(const struct sock *sk,
475 const struct mptcp_options_received *mp_opt,
476 struct request_sock *req);
477 void mptcp_get_options(const struct sk_buff *skb,
478 struct mptcp_options_received *mp_opt);
480 void mptcp_finish_connect(struct sock *sk);
481 static inline bool mptcp_is_fully_established(struct sock *sk)
483 return inet_sk_state_load(sk) == TCP_ESTABLISHED &&
484 READ_ONCE(mptcp_sk(sk)->fully_established);
486 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk);
487 void mptcp_data_ready(struct sock *sk, struct sock *ssk);
488 bool mptcp_finish_join(struct sock *sk);
489 bool mptcp_schedule_work(struct sock *sk);
490 void mptcp_data_acked(struct sock *sk);
491 void mptcp_subflow_eof(struct sock *sk);
492 bool mptcp_update_rcv_data_fin(struct mptcp_sock *msk, u64 data_fin_seq, bool use_64bit);
493 void __mptcp_flush_join_list(struct mptcp_sock *msk);
494 static inline bool mptcp_data_fin_enabled(const struct mptcp_sock *msk)
496 return READ_ONCE(msk->snd_data_fin_enable) &&
497 READ_ONCE(msk->write_seq) == READ_ONCE(msk->snd_nxt);
500 void mptcp_destroy_common(struct mptcp_sock *msk);
502 void __init mptcp_token_init(void);
503 static inline void mptcp_token_init_request(struct request_sock *req)
505 mptcp_subflow_rsk(req)->token_node.pprev = NULL;
508 int mptcp_token_new_request(struct request_sock *req);
509 void mptcp_token_destroy_request(struct request_sock *req);
510 int mptcp_token_new_connect(struct sock *sk);
511 void mptcp_token_accept(struct mptcp_subflow_request_sock *r,
512 struct mptcp_sock *msk);
513 bool mptcp_token_exists(u32 token);
514 struct mptcp_sock *mptcp_token_get_sock(u32 token);
515 struct mptcp_sock *mptcp_token_iter_next(const struct net *net, long *s_slot,
517 void mptcp_token_destroy(struct mptcp_sock *msk);
519 void mptcp_crypto_key_sha(u64 key, u32 *token, u64 *idsn);
521 void mptcp_crypto_hmac_sha(u64 key1, u64 key2, u8 *msg, int len, void *hmac);
523 void __init mptcp_pm_init(void);
524 void mptcp_pm_data_init(struct mptcp_sock *msk);
525 void mptcp_pm_new_connection(struct mptcp_sock *msk, int server_side);
526 void mptcp_pm_fully_established(struct mptcp_sock *msk);
527 bool mptcp_pm_allow_new_subflow(struct mptcp_sock *msk);
528 void mptcp_pm_connection_closed(struct mptcp_sock *msk);
529 void mptcp_pm_subflow_established(struct mptcp_sock *msk,
530 struct mptcp_subflow_context *subflow);
531 void mptcp_pm_subflow_closed(struct mptcp_sock *msk, u8 id);
532 void mptcp_pm_add_addr_received(struct mptcp_sock *msk,
533 const struct mptcp_addr_info *addr);
534 void mptcp_pm_add_addr_send_ack(struct mptcp_sock *msk);
535 void mptcp_pm_rm_addr_received(struct mptcp_sock *msk, u8 rm_id);
536 void mptcp_pm_free_anno_list(struct mptcp_sock *msk);
537 struct mptcp_pm_add_entry *
538 mptcp_pm_del_add_timer(struct mptcp_sock *msk,
539 struct mptcp_addr_info *addr);
541 int mptcp_pm_announce_addr(struct mptcp_sock *msk,
542 const struct mptcp_addr_info *addr,
544 int mptcp_pm_remove_addr(struct mptcp_sock *msk, u8 local_id);
545 int mptcp_pm_remove_subflow(struct mptcp_sock *msk, u8 local_id);
547 static inline bool mptcp_pm_should_add_signal(struct mptcp_sock *msk)
549 return READ_ONCE(msk->pm.add_addr_signal) & BIT(MPTCP_ADD_ADDR_SIGNAL);
552 static inline bool mptcp_pm_should_add_signal_echo(struct mptcp_sock *msk)
554 return READ_ONCE(msk->pm.add_addr_signal) & BIT(MPTCP_ADD_ADDR_ECHO);
557 static inline bool mptcp_pm_should_add_signal_ipv6(struct mptcp_sock *msk)
559 return READ_ONCE(msk->pm.add_addr_signal) & BIT(MPTCP_ADD_ADDR_IPV6);
562 static inline bool mptcp_pm_should_rm_signal(struct mptcp_sock *msk)
564 return READ_ONCE(msk->pm.rm_addr_signal);
567 static inline unsigned int mptcp_add_addr_len(int family, bool echo)
569 if (family == AF_INET)
570 return echo ? TCPOLEN_MPTCP_ADD_ADDR_BASE
571 : TCPOLEN_MPTCP_ADD_ADDR;
572 return echo ? TCPOLEN_MPTCP_ADD_ADDR6_BASE : TCPOLEN_MPTCP_ADD_ADDR6;
575 bool mptcp_pm_add_addr_signal(struct mptcp_sock *msk, unsigned int remaining,
576 struct mptcp_addr_info *saddr, bool *echo);
577 bool mptcp_pm_rm_addr_signal(struct mptcp_sock *msk, unsigned int remaining,
579 int mptcp_pm_get_local_id(struct mptcp_sock *msk, struct sock_common *skc);
581 void __init mptcp_pm_nl_init(void);
582 void mptcp_pm_nl_data_init(struct mptcp_sock *msk);
583 void mptcp_pm_nl_fully_established(struct mptcp_sock *msk);
584 void mptcp_pm_nl_subflow_established(struct mptcp_sock *msk);
585 void mptcp_pm_nl_add_addr_received(struct mptcp_sock *msk);
586 void mptcp_pm_nl_add_addr_send_ack(struct mptcp_sock *msk);
587 void mptcp_pm_nl_rm_addr_received(struct mptcp_sock *msk);
588 void mptcp_pm_nl_rm_subflow_received(struct mptcp_sock *msk, u8 rm_id);
589 int mptcp_pm_nl_get_local_id(struct mptcp_sock *msk, struct sock_common *skc);
591 static inline struct mptcp_ext *mptcp_get_ext(struct sk_buff *skb)
593 return (struct mptcp_ext *)skb_ext_find(skb, SKB_EXT_MPTCP);
596 void mptcp_diag_subflow_init(struct tcp_ulp_ops *ops);
598 static inline bool __mptcp_check_fallback(const struct mptcp_sock *msk)
600 return test_bit(MPTCP_FALLBACK_DONE, &msk->flags);
603 static inline bool mptcp_check_fallback(const struct sock *sk)
605 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
606 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
608 return __mptcp_check_fallback(msk);
611 static inline void __mptcp_do_fallback(struct mptcp_sock *msk)
613 if (test_bit(MPTCP_FALLBACK_DONE, &msk->flags)) {
614 pr_debug("TCP fallback already done (msk=%p)", msk);
617 set_bit(MPTCP_FALLBACK_DONE, &msk->flags);
620 static inline void mptcp_do_fallback(struct sock *sk)
622 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
623 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
625 __mptcp_do_fallback(msk);
628 #define pr_fallback(a) pr_debug("%s:fallback to TCP (msk=%p)", __func__, a)
630 static inline bool subflow_simultaneous_connect(struct sock *sk)
632 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
633 struct sock *parent = subflow->conn;
635 return sk->sk_state == TCP_ESTABLISHED &&
636 !mptcp_sk(parent)->pm.server_side &&
637 !subflow->conn_finished;
640 #ifdef CONFIG_SYN_COOKIES
641 void subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock *subflow_req,
642 struct sk_buff *skb);
643 bool mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock *subflow_req,
644 struct sk_buff *skb);
645 void __init mptcp_join_cookie_init(void);
648 subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock *subflow_req,
649 struct sk_buff *skb) {}
651 mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock *subflow_req,
657 static inline void mptcp_join_cookie_init(void) {}
660 #endif /* __MPTCP_PROTOCOL_H */