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 #include <net/rstreason.h>
16
17 #define MPTCP_SUPPORTED_VERSION 1
18
19 /* MPTCP option bits */
20 #define OPTION_MPTCP_MPC_SYN BIT(0)
21 #define OPTION_MPTCP_MPC_SYNACK BIT(1)
22 #define OPTION_MPTCP_MPC_ACK BIT(2)
23 #define OPTION_MPTCP_MPJ_SYN BIT(3)
24 #define OPTION_MPTCP_MPJ_SYNACK BIT(4)
25 #define OPTION_MPTCP_MPJ_ACK BIT(5)
26 #define OPTION_MPTCP_ADD_ADDR BIT(6)
27 #define OPTION_MPTCP_RM_ADDR BIT(7)
28 #define OPTION_MPTCP_FASTCLOSE BIT(8)
29 #define OPTION_MPTCP_PRIO BIT(9)
30 #define OPTION_MPTCP_RST BIT(10)
31 #define OPTION_MPTCP_DSS BIT(11)
32 #define OPTION_MPTCP_FAIL BIT(12)
33
34 #define OPTION_MPTCP_CSUMREQD BIT(13)
35
36 #define OPTIONS_MPTCP_MPC (OPTION_MPTCP_MPC_SYN | OPTION_MPTCP_MPC_SYNACK | \
37 OPTION_MPTCP_MPC_ACK)
38 #define OPTIONS_MPTCP_MPJ (OPTION_MPTCP_MPJ_SYN | OPTION_MPTCP_MPJ_SYNACK | \
39 OPTION_MPTCP_MPJ_ACK)
40 #define OPTIONS_MPTCP_DSS (OPTION_MPTCP_DSS | OPTION_MPTCP_CSUMREQD)
41
42 /* MPTCP option subtypes */
43 #define MPTCPOPT_MP_CAPABLE 0
44 #define MPTCPOPT_MP_JOIN 1
45 #define MPTCPOPT_DSS 2
46 #define MPTCPOPT_ADD_ADDR 3
47 #define MPTCPOPT_RM_ADDR 4
48 #define MPTCPOPT_MP_PRIO 5
49 #define MPTCPOPT_MP_FAIL 6
50 #define MPTCPOPT_MP_FASTCLOSE 7
51 #define MPTCPOPT_RST 8
52
53 /* MPTCP suboption lengths */
54 #define TCPOLEN_MPTCP_MPC_SYN 4
55 #define TCPOLEN_MPTCP_MPC_SYNACK 12
56 #define TCPOLEN_MPTCP_MPC_ACK 20
57 #define TCPOLEN_MPTCP_MPC_ACK_DATA 22
58 #define TCPOLEN_MPTCP_MPJ_SYN 12
59 #define TCPOLEN_MPTCP_MPJ_SYNACK 16
60 #define TCPOLEN_MPTCP_MPJ_ACK 24
61 #define TCPOLEN_MPTCP_DSS_BASE 4
62 #define TCPOLEN_MPTCP_DSS_ACK32 4
63 #define TCPOLEN_MPTCP_DSS_ACK64 8
64 #define TCPOLEN_MPTCP_DSS_MAP32 10
65 #define TCPOLEN_MPTCP_DSS_MAP64 14
66 #define TCPOLEN_MPTCP_DSS_CHECKSUM 2
67 #define TCPOLEN_MPTCP_ADD_ADDR 16
68 #define TCPOLEN_MPTCP_ADD_ADDR_PORT 18
69 #define TCPOLEN_MPTCP_ADD_ADDR_BASE 8
70 #define TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT 10
71 #define TCPOLEN_MPTCP_ADD_ADDR6 28
72 #define TCPOLEN_MPTCP_ADD_ADDR6_PORT 30
73 #define TCPOLEN_MPTCP_ADD_ADDR6_BASE 20
74 #define TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT 22
75 #define TCPOLEN_MPTCP_PORT_LEN 2
76 #define TCPOLEN_MPTCP_PORT_ALIGN 2
77 #define TCPOLEN_MPTCP_RM_ADDR_BASE 3
78 #define TCPOLEN_MPTCP_PRIO 3
79 #define TCPOLEN_MPTCP_PRIO_ALIGN 4
80 #define TCPOLEN_MPTCP_FASTCLOSE 12
81 #define TCPOLEN_MPTCP_RST 4
82 #define TCPOLEN_MPTCP_FAIL 12
83
84 #define TCPOLEN_MPTCP_MPC_ACK_DATA_CSUM (TCPOLEN_MPTCP_DSS_CHECKSUM + TCPOLEN_MPTCP_MPC_ACK_DATA)
85
86 /* MPTCP MP_JOIN flags */
87 #define MPTCPOPT_BACKUP BIT(0)
88 #define MPTCPOPT_THMAC_LEN 8
89
90 /* MPTCP MP_CAPABLE flags */
91 #define MPTCP_VERSION_MASK (0x0F)
92 #define MPTCP_CAP_CHECKSUM_REQD BIT(7)
93 #define MPTCP_CAP_EXTENSIBILITY BIT(6)
94 #define MPTCP_CAP_DENY_JOIN_ID0 BIT(5)
95 #define MPTCP_CAP_HMAC_SHA256 BIT(0)
96 #define MPTCP_CAP_FLAG_MASK (0x1F)
97
98 /* MPTCP DSS flags */
99 #define MPTCP_DSS_DATA_FIN BIT(4)
100 #define MPTCP_DSS_DSN64 BIT(3)
101 #define MPTCP_DSS_HAS_MAP BIT(2)
102 #define MPTCP_DSS_ACK64 BIT(1)
103 #define MPTCP_DSS_HAS_ACK BIT(0)
104 #define MPTCP_DSS_FLAG_MASK (0x1F)
105
106 /* MPTCP ADD_ADDR flags */
107 #define MPTCP_ADDR_ECHO BIT(0)
108
109 /* MPTCP MP_PRIO flags */
110 #define MPTCP_PRIO_BKUP BIT(0)
111
112 /* MPTCP TCPRST flags */
113 #define MPTCP_RST_TRANSIENT BIT(0)
114
115 /* MPTCP socket atomic flags */
116 #define MPTCP_WORK_RTX 1
117 #define MPTCP_FALLBACK_DONE 2
118 #define MPTCP_WORK_CLOSE_SUBFLOW 3
119
120 /* MPTCP socket release cb flags */
121 #define MPTCP_PUSH_PENDING 1
122 #define MPTCP_CLEAN_UNA 2
123 #define MPTCP_ERROR_REPORT 3
124 #define MPTCP_RETRANSMIT 4
125 #define MPTCP_FLUSH_JOIN_LIST 5
126 #define MPTCP_SYNC_STATE 6
127 #define MPTCP_SYNC_SNDBUF 7
128
129 struct mptcp_skb_cb {
130 u64 map_seq;
131 u64 end_seq;
132 u32 offset;
133 u8 has_rxtstamp;
134 u8 cant_coalesce;
135 };
136
137 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
138
before64(__u64 seq1,__u64 seq2)139 static inline bool before64(__u64 seq1, __u64 seq2)
140 {
141 return (__s64)(seq1 - seq2) < 0;
142 }
143
144 #define after64(seq2, seq1) before64(seq1, seq2)
145
146 struct mptcp_options_received {
147 u64 sndr_key;
148 u64 rcvr_key;
149 u64 data_ack;
150 u64 data_seq;
151 u32 subflow_seq;
152 u16 data_len;
153 __sum16 csum;
154 struct_group(status,
155 u16 suboptions;
156 u16 use_map:1,
157 dsn64:1,
158 data_fin:1,
159 use_ack:1,
160 ack64:1,
161 mpc_map:1,
162 reset_reason:4,
163 reset_transient:1,
164 echo:1,
165 backup:1,
166 deny_join_id0:1,
167 __unused:2;
168 );
169 u8 join_id;
170 u32 token;
171 u32 nonce;
172 u64 thmac;
173 u8 hmac[MPTCPOPT_HMAC_LEN];
174 struct mptcp_addr_info addr;
175 struct mptcp_rm_list rm_list;
176 u64 ahmac;
177 u64 fail_seq;
178 };
179
mptcp_option(u8 subopt,u8 len,u8 nib,u8 field)180 static inline __be32 mptcp_option(u8 subopt, u8 len, u8 nib, u8 field)
181 {
182 return htonl((TCPOPT_MPTCP << 24) | (len << 16) | (subopt << 12) |
183 ((nib & 0xF) << 8) | field);
184 }
185
186 enum mptcp_pm_status {
187 MPTCP_PM_ADD_ADDR_RECEIVED,
188 MPTCP_PM_ADD_ADDR_SEND_ACK,
189 MPTCP_PM_RM_ADDR_RECEIVED,
190 MPTCP_PM_ESTABLISHED,
191 MPTCP_PM_SUBFLOW_ESTABLISHED,
192 MPTCP_PM_ALREADY_ESTABLISHED, /* persistent status, set after ESTABLISHED event */
193 MPTCP_PM_MPC_ENDPOINT_ACCOUNTED, /* persistent status, set after MPC local address is
194 * accounted int id_avail_bitmap
195 */
196 MPTCP_PM_DESTROYING, /* To fence out PM list allocs */
197 };
198
199 enum mptcp_pm_type {
200 MPTCP_PM_TYPE_KERNEL = 0,
201 MPTCP_PM_TYPE_USERSPACE,
202
203 __MPTCP_PM_TYPE_NR,
204 __MPTCP_PM_TYPE_MAX = __MPTCP_PM_TYPE_NR - 1,
205 };
206
207 /* Status bits below MPTCP_PM_ALREADY_ESTABLISHED need pm worker actions */
208 #define MPTCP_PM_WORK_MASK ((1 << MPTCP_PM_ALREADY_ESTABLISHED) - 1)
209
210 enum mptcp_addr_signal_status {
211 MPTCP_ADD_ADDR_SIGNAL,
212 MPTCP_ADD_ADDR_ECHO,
213 MPTCP_RM_ADDR_SIGNAL,
214 };
215
216 /* max value of mptcp_addr_info.id */
217 #define MPTCP_PM_MAX_ADDR_ID U8_MAX
218
219 struct mptcp_pm_data {
220 struct mptcp_addr_info local;
221 struct mptcp_addr_info remote;
222 struct list_head anno_list;
223 struct list_head userspace_pm_local_addr_list;
224
225 spinlock_t lock; /*protects the whole PM data */
226
227 struct_group(reset,
228
229 u8 addr_signal;
230 bool server_side;
231 bool work_pending;
232 bool accept_addr;
233 bool accept_subflow;
234 bool remote_deny_join_id0;
235 u8 add_addr_signaled;
236 u8 add_addr_accepted;
237 u8 local_addr_used;
238 u8 pm_type;
239 u8 extra_subflows;
240 u8 status;
241
242 );
243
244 DECLARE_BITMAP(id_avail_bitmap, MPTCP_PM_MAX_ADDR_ID + 1);
245 struct mptcp_rm_list rm_list_tx;
246 struct mptcp_rm_list rm_list_rx;
247 };
248
249 struct mptcp_pm_local {
250 struct mptcp_addr_info addr;
251 u32 flags;
252 int ifindex;
253 };
254
255 struct mptcp_pm_addr_entry {
256 struct list_head list;
257 struct mptcp_addr_info addr;
258 u32 flags;
259 int ifindex;
260 struct socket *lsk;
261 };
262
263 struct mptcp_data_frag {
264 struct list_head list;
265 u64 data_seq;
266 u16 data_len;
267 u16 offset;
268 u8 overhead;
269 u8 eor; /* currently using 1 bit */
270 u16 already_sent;
271 struct page *page;
272 };
273
274 /* Arbitrary compromise between as low as possible to react timely to subflow
275 * close event and as big as possible to avoid being fouled by biased large
276 * samples due to peer sending data on a different subflow WRT to the incoming
277 * ack.
278 */
279 #define MPTCP_RTT_SAMPLES 5
280
281 /* MPTCP connection sock */
282 struct mptcp_sock {
283 /* inet_connection_sock must be the first member */
284 struct inet_connection_sock sk;
285 u64 local_key; /* protected by the first subflow socket lock
286 * lockless access read
287 */
288 u64 remote_key; /* same as above */
289 u64 write_seq;
290 u64 bytes_sent;
291 u64 snd_nxt;
292 u64 bytes_received;
293 u64 ack_seq;
294 atomic64_t rcv_wnd_sent;
295 u64 rcv_data_fin_seq;
296 u64 bytes_retrans;
297 u64 bytes_consumed;
298 int snd_burst;
299 int old_wspace;
300 u64 recovery_snd_nxt; /* in recovery mode accept up to this seq;
301 * recovery related fields are under data_lock
302 * protection
303 */
304 u64 bytes_acked;
305 u64 snd_una;
306 u64 wnd_end;
307 u32 last_data_sent;
308 u32 last_data_recv;
309 u32 last_ack_recv;
310 unsigned long timer_ival;
311 u32 token;
312 unsigned long flags;
313 unsigned long cb_flags;
314 bool recovery; /* closing subflow write queue reinjected */
315 bool can_ack;
316 bool fully_established;
317 bool rcv_data_fin;
318 bool snd_data_fin_enable;
319 bool rcv_fastclose;
320 bool use_64bit_ack; /* Set when we received a 64-bit DSN */
321 bool csum_enabled;
322 bool allow_infinite_fallback;
323 u8 pending_state; /* A subflow asked to set this sk_state,
324 * protected by the msk data lock
325 */
326 u8 mpc_endpoint_id;
327 u8 recvmsg_inq:1,
328 cork:1,
329 nodelay:1,
330 fastopening:1,
331 in_accept_queue:1,
332 free_first:1,
333 rcvspace_init:1,
334 fastclosing:1;
335 u32 notsent_lowat;
336 int keepalive_cnt;
337 int keepalive_idle;
338 int keepalive_intvl;
339 int maxseg;
340 struct work_struct work;
341 struct sk_buff *ooo_last_skb;
342 struct rb_root out_of_order_queue;
343 struct list_head conn_list;
344 struct list_head rtx_queue;
345 struct mptcp_data_frag *first_pending;
346 struct list_head join_list;
347 struct sock *first; /* The mptcp ops can safely dereference, using suitable
348 * ONCE annotation, the subflow outside the socket
349 * lock as such sock is freed after close().
350 */
351 struct mptcp_pm_data pm;
352 struct mptcp_sched_ops *sched;
353
354 /* Most recent rtt_us observed by in use incoming subflows. */
355 struct {
356 u32 samples[MPTCP_RTT_SAMPLES];
357 u32 next_sample;
358 } rcv_rtt_est;
359
360 struct {
361 int space; /* bytes copied in last measurement window */
362 int copied; /* bytes copied in this measurement window */
363 u64 time; /* start time of measurement window */
364 } rcvq_space;
365 u8 scaling_ratio;
366 bool allow_subflows;
367
368 u32 subflow_id;
369 u32 setsockopt_seq;
370 char ca_name[TCP_CA_NAME_MAX];
371
372 spinlock_t fallback_lock; /* protects fallback,
373 * allow_infinite_fallback and
374 * allow_join
375 */
376
377 struct list_head backlog_list; /* protected by the data lock */
378 u32 backlog_len;
379 u32 backlog_unaccounted;
380 };
381
382 #define mptcp_data_lock(sk) spin_lock_bh(&(sk)->sk_lock.slock)
383 #define mptcp_data_unlock(sk) spin_unlock_bh(&(sk)->sk_lock.slock)
384
385 #define mptcp_for_each_subflow(__msk, __subflow) \
386 list_for_each_entry(__subflow, &((__msk)->conn_list), node)
387 #define mptcp_for_each_subflow_safe(__msk, __subflow, __tmp) \
388 list_for_each_entry_safe(__subflow, __tmp, &((__msk)->conn_list), node)
389 #define mptcp_next_subflow(__msk, __subflow) \
390 list_next_entry_circular(__subflow, &((__msk)->conn_list), node)
391
392 extern struct genl_family mptcp_genl_family;
393
msk_owned_by_me(const struct mptcp_sock * msk)394 static inline void msk_owned_by_me(const struct mptcp_sock *msk)
395 {
396 sock_owned_by_me((const struct sock *)msk);
397 }
398
399 #ifdef CONFIG_DEBUG_NET
400 /* MPTCP-specific: we might (indirectly) call this helper with the wrong sk */
401 #undef tcp_sk
402 #define tcp_sk(ptr) ({ \
403 typeof(ptr) _ptr = (ptr); \
404 WARN_ON(_ptr->sk_protocol != IPPROTO_TCP); \
405 container_of_const(_ptr, struct tcp_sock, inet_conn.icsk_inet.sk); \
406 })
407 #define mptcp_sk(ptr) ({ \
408 typeof(ptr) _ptr = (ptr); \
409 WARN_ON(_ptr->sk_protocol != IPPROTO_MPTCP); \
410 container_of_const(_ptr, struct mptcp_sock, sk.icsk_inet.sk); \
411 })
412
413 #else /* !CONFIG_DEBUG_NET */
414 #define mptcp_sk(ptr) container_of_const(ptr, struct mptcp_sock, sk.icsk_inet.sk)
415 #endif
416
mptcp_win_from_space(const struct sock * sk,int space)417 static inline int mptcp_win_from_space(const struct sock *sk, int space)
418 {
419 return __tcp_win_from_space(mptcp_sk(sk)->scaling_ratio, space);
420 }
421
mptcp_space_from_win(const struct sock * sk,int win)422 static inline int mptcp_space_from_win(const struct sock *sk, int win)
423 {
424 return __tcp_space_from_win(mptcp_sk(sk)->scaling_ratio, win);
425 }
426
__mptcp_space(const struct sock * sk)427 static inline int __mptcp_space(const struct sock *sk)
428 {
429 return mptcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf) -
430 READ_ONCE(mptcp_sk(sk)->backlog_len) -
431 sk_rmem_alloc_get(sk));
432 }
433
mptcp_send_head(const struct sock * sk)434 static inline struct mptcp_data_frag *mptcp_send_head(const struct sock *sk)
435 {
436 const struct mptcp_sock *msk = mptcp_sk(sk);
437
438 return msk->first_pending;
439 }
440
mptcp_init_rtt_est(struct mptcp_sock * msk)441 static inline void mptcp_init_rtt_est(struct mptcp_sock *msk)
442 {
443 int i;
444
445 for (i = 0; i < MPTCP_RTT_SAMPLES; ++i)
446 msk->rcv_rtt_est.samples[i] = U32_MAX;
447 msk->rcv_rtt_est.next_sample = 0;
448 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
449 }
450
mptcp_rtt_us_est(const struct mptcp_sock * msk)451 static inline u32 mptcp_rtt_us_est(const struct mptcp_sock *msk)
452 {
453 u32 rtt_us = READ_ONCE(msk->rcv_rtt_est.samples[0]);
454 int i;
455
456 /* Lockless access of collected samples. */
457 for (i = 1; i < MPTCP_RTT_SAMPLES; ++i)
458 rtt_us = min(rtt_us, READ_ONCE(msk->rcv_rtt_est.samples[i]));
459 return rtt_us;
460 }
461
mptcp_send_next(struct sock * sk)462 static inline struct mptcp_data_frag *mptcp_send_next(struct sock *sk)
463 {
464 struct mptcp_sock *msk = mptcp_sk(sk);
465 struct mptcp_data_frag *cur;
466
467 cur = msk->first_pending;
468 return list_is_last(&cur->list, &msk->rtx_queue) ? NULL :
469 list_next_entry(cur, list);
470 }
471
mptcp_pending_tail(const struct sock * sk)472 static inline struct mptcp_data_frag *mptcp_pending_tail(const struct sock *sk)
473 {
474 const struct mptcp_sock *msk = mptcp_sk(sk);
475
476 if (!msk->first_pending)
477 return NULL;
478
479 if (WARN_ON_ONCE(list_empty(&msk->rtx_queue)))
480 return NULL;
481
482 return list_last_entry(&msk->rtx_queue, struct mptcp_data_frag, list);
483 }
484
mptcp_rtx_head(struct sock * sk)485 static inline struct mptcp_data_frag *mptcp_rtx_head(struct sock *sk)
486 {
487 struct mptcp_sock *msk = mptcp_sk(sk);
488
489 if (msk->snd_una == msk->snd_nxt)
490 return NULL;
491
492 return list_first_entry_or_null(&msk->rtx_queue, struct mptcp_data_frag, list);
493 }
494
495 struct csum_pseudo_header {
496 __be64 data_seq;
497 __be32 subflow_seq;
498 __be16 data_len;
499 __sum16 csum;
500 };
501
502 struct mptcp_subflow_request_sock {
503 struct tcp_request_sock sk;
504 u16 mp_capable : 1,
505 mp_join : 1,
506 backup : 1,
507 request_bkup : 1,
508 csum_reqd : 1,
509 allow_join_id0 : 1;
510 u8 local_id;
511 u8 remote_id;
512 u64 local_key;
513 u64 idsn;
514 u32 token;
515 u32 ssn_offset;
516 u64 thmac;
517 u32 local_nonce;
518 u32 remote_nonce;
519 struct mptcp_sock *msk;
520 struct hlist_nulls_node token_node;
521 };
522
523 static inline struct mptcp_subflow_request_sock *
mptcp_subflow_rsk(const struct request_sock * rsk)524 mptcp_subflow_rsk(const struct request_sock *rsk)
525 {
526 return (struct mptcp_subflow_request_sock *)rsk;
527 }
528
529 struct mptcp_delegated_action {
530 struct napi_struct napi;
531 local_lock_t bh_lock;
532 struct list_head head;
533 };
534
535 DECLARE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
536
537 #define MPTCP_DELEGATE_SCHEDULED 0
538 #define MPTCP_DELEGATE_SEND 1
539 #define MPTCP_DELEGATE_ACK 2
540 #define MPTCP_DELEGATE_SNDBUF 3
541
542 #define MPTCP_DELEGATE_ACTIONS_MASK (~BIT(MPTCP_DELEGATE_SCHEDULED))
543 /* MPTCP subflow context */
544 struct mptcp_subflow_context {
545 struct list_head node;/* conn_list of subflows */
546
547 struct_group(reset,
548
549 unsigned long avg_pacing_rate; /* protected by msk socket lock */
550 u64 local_key;
551 u64 remote_key;
552 u64 idsn;
553 u64 map_seq;
554 u64 rcv_wnd_sent;
555 u32 snd_isn;
556 u32 token;
557 u32 rel_write_seq;
558 u32 map_subflow_seq;
559 u32 ssn_offset;
560 u32 map_data_len;
561 __wsum map_data_csum;
562 u32 map_csum_len;
563 u32 prev_rtt_seq;
564 u32 request_mptcp : 1, /* send MP_CAPABLE */
565 request_join : 1, /* send MP_JOIN */
566 request_bkup : 1,
567 mp_capable : 1, /* remote is MPTCP capable */
568 mp_join : 1, /* remote is JOINing */
569 pm_notified : 1, /* PM hook called for established status */
570 conn_finished : 1,
571 map_valid : 1,
572 map_csum_reqd : 1,
573 map_data_fin : 1,
574 mpc_map : 1,
575 backup : 1,
576 send_mp_prio : 1,
577 send_mp_fail : 1,
578 send_fastclose : 1,
579 send_infinite_map : 1,
580 remote_key_valid : 1, /* received the peer key from */
581 disposable : 1, /* ctx can be free at ulp release time */
582 closing : 1, /* must not pass rx data to msk anymore */
583 valid_csum_seen : 1, /* at least one csum validated */
584 is_mptfo : 1, /* subflow is doing TFO */
585 close_event_done : 1, /* has done the post-closed part */
586 mpc_drop : 1, /* the MPC option has been dropped in a rtx */
587 __unused : 9;
588 bool data_avail;
589 bool scheduled;
590 bool pm_listener; /* a listener managed by the kernel PM? */
591 bool fully_established; /* path validated */
592 u32 lent_mem_frag;
593 u32 remote_nonce;
594 u64 thmac;
595 u32 local_nonce;
596 u32 remote_token;
597 union {
598 u8 hmac[MPTCPOPT_HMAC_LEN]; /* MPJ subflow only */
599 u64 iasn; /* initial ack sequence number, MPC subflows only */
600 };
601 s16 local_id; /* if negative not initialized yet */
602 u8 remote_id;
603 u8 reset_seen:1;
604 u8 reset_transient:1;
605 u8 reset_reason:4;
606 u8 stale_count; /* Protected by the msk socket lock */
607 u8 stale; /* Protected by the msk socket lock,
608 * if set the subflow is unable to snd/rcv
609 * data, the schedule should skip it
610 */
611
612 u32 subflow_id;
613
614 long delegated_status;
615 unsigned long fail_tout;
616
617 );
618
619 struct list_head delegated_node; /* link into delegated_action, protected by local BH */
620
621 u32 setsockopt_seq;
622 u32 stale_rcv_tstamp;
623 int cached_sndbuf; /* sndbuf size when last synced with the msk sndbuf,
624 * protected by the msk socket lock
625 */
626
627 struct sock *tcp_sock; /* tcp sk backpointer */
628 struct sock *conn; /* parent mptcp_sock */
629 const struct inet_connection_sock_af_ops *icsk_af_ops;
630 void (*tcp_state_change)(struct sock *sk);
631 void (*tcp_error_report)(struct sock *sk);
632
633 struct rcu_head rcu;
634 };
635
636 static inline struct mptcp_subflow_context *
mptcp_subflow_ctx(const struct sock * sk)637 mptcp_subflow_ctx(const struct sock *sk)
638 {
639 const struct inet_connection_sock *icsk = inet_csk(sk);
640
641 /* Use RCU on icsk_ulp_data only for sock diag code */
642 return (__force struct mptcp_subflow_context *)icsk->icsk_ulp_data;
643 }
644
645 static inline struct sock *
mptcp_subflow_tcp_sock(const struct mptcp_subflow_context * subflow)646 mptcp_subflow_tcp_sock(const struct mptcp_subflow_context *subflow)
647 {
648 return subflow->tcp_sock;
649 }
650
651 static inline void
mptcp_subflow_ctx_reset(struct mptcp_subflow_context * subflow)652 mptcp_subflow_ctx_reset(struct mptcp_subflow_context *subflow)
653 {
654 memset(&subflow->reset, 0, sizeof(subflow->reset));
655 subflow->request_mptcp = 1;
656 WRITE_ONCE(subflow->local_id, -1);
657 }
658
659 /* Convert reset reasons in MPTCP to enum sk_rst_reason type */
660 static inline enum sk_rst_reason
sk_rst_convert_mptcp_reason(u32 reason)661 sk_rst_convert_mptcp_reason(u32 reason)
662 {
663 switch (reason) {
664 case MPTCP_RST_EUNSPEC:
665 return SK_RST_REASON_MPTCP_RST_EUNSPEC;
666 case MPTCP_RST_EMPTCP:
667 return SK_RST_REASON_MPTCP_RST_EMPTCP;
668 case MPTCP_RST_ERESOURCE:
669 return SK_RST_REASON_MPTCP_RST_ERESOURCE;
670 case MPTCP_RST_EPROHIBIT:
671 return SK_RST_REASON_MPTCP_RST_EPROHIBIT;
672 case MPTCP_RST_EWQ2BIG:
673 return SK_RST_REASON_MPTCP_RST_EWQ2BIG;
674 case MPTCP_RST_EBADPERF:
675 return SK_RST_REASON_MPTCP_RST_EBADPERF;
676 case MPTCP_RST_EMIDDLEBOX:
677 return SK_RST_REASON_MPTCP_RST_EMIDDLEBOX;
678 default:
679 /* It should not happen, or else errors may occur
680 * in MPTCP layer
681 */
682 return SK_RST_REASON_ERROR;
683 }
684 }
685
686 static inline void
mptcp_send_active_reset_reason(struct sock * sk)687 mptcp_send_active_reset_reason(struct sock *sk)
688 {
689 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
690 enum sk_rst_reason reason;
691
692 reason = sk_rst_convert_mptcp_reason(subflow->reset_reason);
693 tcp_send_active_reset(sk, GFP_ATOMIC, reason);
694 }
695
696 /* Made the fwd mem carried by the given skb available to the msk,
697 * To be paired with a previous mptcp_subflow_lend_fwdmem() before freeing
698 * the skb or setting the skb ownership.
699 */
mptcp_borrow_fwdmem(struct sock * sk,struct sk_buff * skb)700 static inline void mptcp_borrow_fwdmem(struct sock *sk, struct sk_buff *skb)
701 {
702 struct sock *ssk = skb->sk;
703
704 /* The subflow just lend the skb fwd memory; if the subflow meanwhile
705 * closed, mptcp_close_ssk() already released the ssk rcv memory.
706 */
707 DEBUG_NET_WARN_ON_ONCE(skb->destructor);
708 sk_forward_alloc_add(sk, skb->truesize);
709 if (!ssk)
710 return;
711
712 atomic_sub(skb->truesize, &ssk->sk_rmem_alloc);
713 skb->sk = NULL;
714 }
715
716 static inline void
__mptcp_subflow_lend_fwdmem(struct mptcp_subflow_context * subflow,int size)717 __mptcp_subflow_lend_fwdmem(struct mptcp_subflow_context *subflow, int size)
718 {
719 int frag = (subflow->lent_mem_frag + size) & (PAGE_SIZE - 1);
720
721 subflow->lent_mem_frag = frag;
722 }
723
724 static inline void
mptcp_subflow_lend_fwdmem(struct mptcp_subflow_context * subflow,struct sk_buff * skb)725 mptcp_subflow_lend_fwdmem(struct mptcp_subflow_context *subflow,
726 struct sk_buff *skb)
727 {
728 __mptcp_subflow_lend_fwdmem(subflow, skb->truesize);
729 skb->destructor = NULL;
730 }
731
732 static inline u64
mptcp_subflow_get_map_offset(const struct mptcp_subflow_context * subflow)733 mptcp_subflow_get_map_offset(const struct mptcp_subflow_context *subflow)
734 {
735 return tcp_sk(mptcp_subflow_tcp_sock(subflow))->copied_seq -
736 subflow->ssn_offset -
737 subflow->map_subflow_seq;
738 }
739
740 static inline u64
mptcp_subflow_get_mapped_dsn(const struct mptcp_subflow_context * subflow)741 mptcp_subflow_get_mapped_dsn(const struct mptcp_subflow_context *subflow)
742 {
743 return subflow->map_seq + mptcp_subflow_get_map_offset(subflow);
744 }
745
746 void mptcp_subflow_process_delegated(struct sock *ssk, long actions);
747
mptcp_subflow_delegate(struct mptcp_subflow_context * subflow,int action)748 static inline void mptcp_subflow_delegate(struct mptcp_subflow_context *subflow, int action)
749 {
750 long old, set_bits = BIT(MPTCP_DELEGATE_SCHEDULED) | BIT(action);
751 struct mptcp_delegated_action *delegated;
752 bool schedule;
753
754 /* the caller held the subflow bh socket lock */
755 lockdep_assert_in_softirq();
756
757 /* The implied barrier pairs with tcp_release_cb_override()
758 * mptcp_napi_poll(), and ensures the below list check sees list
759 * updates done prior to delegated status bits changes
760 */
761 old = set_mask_bits(&subflow->delegated_status, 0, set_bits);
762 if (!(old & BIT(MPTCP_DELEGATE_SCHEDULED))) {
763 if (WARN_ON_ONCE(!list_empty(&subflow->delegated_node)))
764 return;
765
766 local_lock_nested_bh(&mptcp_delegated_actions.bh_lock);
767 delegated = this_cpu_ptr(&mptcp_delegated_actions);
768 schedule = list_empty(&delegated->head);
769 list_add_tail(&subflow->delegated_node, &delegated->head);
770 local_unlock_nested_bh(&mptcp_delegated_actions.bh_lock);
771 sock_hold(mptcp_subflow_tcp_sock(subflow));
772 if (schedule)
773 napi_schedule(&delegated->napi);
774 }
775 }
776
777 static inline struct mptcp_subflow_context *
mptcp_subflow_delegated_next(struct mptcp_delegated_action * delegated)778 mptcp_subflow_delegated_next(struct mptcp_delegated_action *delegated)
779 {
780 struct mptcp_subflow_context *ret;
781
782 local_lock_nested_bh(&mptcp_delegated_actions.bh_lock);
783 if (list_empty(&delegated->head)) {
784 local_unlock_nested_bh(&mptcp_delegated_actions.bh_lock);
785 return NULL;
786 }
787
788 ret = list_first_entry(&delegated->head, struct mptcp_subflow_context, delegated_node);
789 list_del_init(&ret->delegated_node);
790 local_unlock_nested_bh(&mptcp_delegated_actions.bh_lock);
791 return ret;
792 }
793
794 void __mptcp_inherit_memcg(struct sock *sk, struct sock *ssk, gfp_t gfp);
795 void __mptcp_inherit_cgrp_data(struct sock *sk, struct sock *ssk);
796
797 int mptcp_is_enabled(const struct net *net);
798 unsigned int mptcp_get_add_addr_timeout(const struct net *net);
799 int mptcp_is_checksum_enabled(const struct net *net);
800 int mptcp_allow_join_id0(const struct net *net);
801 unsigned int mptcp_stale_loss_cnt(const struct net *net);
802 unsigned int mptcp_close_timeout(const struct sock *sk);
803 int mptcp_get_pm_type(const struct net *net);
804 const char *mptcp_get_path_manager(const struct net *net);
805 const char *mptcp_get_scheduler(const struct net *net);
806 unsigned int mptcp_add_addr_v6_port_drop_ts(const struct net *net);
807
808 void mptcp_active_disable(struct sock *sk);
809 bool mptcp_active_should_disable(struct sock *ssk);
810 void mptcp_active_enable(struct sock *sk);
811
812 void mptcp_get_available_schedulers(char *buf, size_t maxlen);
813 void __mptcp_subflow_fully_established(struct mptcp_sock *msk,
814 struct mptcp_subflow_context *subflow,
815 const struct mptcp_options_received *mp_opt);
816 bool __mptcp_retransmit_pending_data(struct sock *sk);
817 void mptcp_check_and_set_pending(struct sock *sk);
818 void __mptcp_push_pending(struct sock *sk, unsigned int flags);
819 bool mptcp_subflow_data_available(struct sock *sk);
820 void __init mptcp_subflow_init(void);
821 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how);
822 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
823 struct mptcp_subflow_context *subflow);
824 void __mptcp_subflow_send_ack(struct sock *ssk);
825 void mptcp_subflow_reset(struct sock *ssk);
826 void mptcp_subflow_queue_clean(struct sock *sk, struct sock *ssk);
827 void mptcp_sock_graft(struct sock *sk, struct socket *parent);
828 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk);
829 bool __mptcp_close(struct sock *sk, long timeout);
830 void mptcp_cancel_work(struct sock *sk);
831 void __mptcp_unaccepted_force_close(struct sock *sk);
832 void mptcp_set_state(struct sock *sk, int state);
833
834 bool mptcp_addresses_equal(const struct mptcp_addr_info *a,
835 const struct mptcp_addr_info *b, bool use_port);
836 void mptcp_local_address(const struct sock_common *skc,
837 struct mptcp_addr_info *addr);
838 void mptcp_remote_address(const struct sock_common *skc,
839 struct mptcp_addr_info *addr);
840
841 /* called with sk socket lock held */
842 int __mptcp_subflow_connect(struct sock *sk, const struct mptcp_pm_local *local,
843 const struct mptcp_addr_info *remote);
844 int mptcp_subflow_create_socket(struct sock *sk, unsigned short family,
845 struct socket **new_sock);
846 void mptcp_info2sockaddr(const struct mptcp_addr_info *info,
847 struct sockaddr_storage *addr,
848 unsigned short family);
849 struct mptcp_sched_ops *mptcp_sched_find(const char *name);
850 int mptcp_validate_scheduler(struct mptcp_sched_ops *sched);
851 int mptcp_register_scheduler(struct mptcp_sched_ops *sched);
852 void mptcp_unregister_scheduler(struct mptcp_sched_ops *sched);
853 void mptcp_sched_init(void);
854 int mptcp_init_sched(struct mptcp_sock *msk,
855 struct mptcp_sched_ops *sched);
856 void mptcp_release_sched(struct mptcp_sock *msk);
857 void mptcp_subflow_set_scheduled(struct mptcp_subflow_context *subflow,
858 bool scheduled);
859 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk);
860 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk);
861 int mptcp_sched_get_send(struct mptcp_sock *msk);
862 int mptcp_sched_get_retrans(struct mptcp_sock *msk);
863
mptcp_data_avail(const struct mptcp_sock * msk)864 static inline u64 mptcp_data_avail(const struct mptcp_sock *msk)
865 {
866 return READ_ONCE(msk->bytes_received) - READ_ONCE(msk->bytes_consumed);
867 }
868
mptcp_epollin_ready(const struct sock * sk)869 static inline bool mptcp_epollin_ready(const struct sock *sk)
870 {
871 u64 data_avail = mptcp_data_avail(mptcp_sk(sk));
872
873 if (!data_avail)
874 return false;
875
876 /* mptcp doesn't have to deal with small skbs in the receive queue,
877 * as it can always coalesce them
878 */
879 return (data_avail >= sk->sk_rcvlowat) ||
880 tcp_under_memory_pressure(sk);
881 }
882
883 int mptcp_set_rcvlowat(struct sock *sk, int val);
884
__tcp_can_send(const struct sock * ssk)885 static inline bool __tcp_can_send(const struct sock *ssk)
886 {
887 /* only send if our side has not closed yet */
888 return ((1 << inet_sk_state_load(ssk)) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
889 }
890
__mptcp_subflow_active(struct mptcp_subflow_context * subflow)891 static inline bool __mptcp_subflow_active(struct mptcp_subflow_context *subflow)
892 {
893 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
894 if (subflow->request_join && !READ_ONCE(subflow->fully_established))
895 return false;
896
897 return __tcp_can_send(mptcp_subflow_tcp_sock(subflow));
898 }
899
900 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow);
901
902 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow);
903
904 void mptcp_subflow_drop_ctx(struct sock *ssk);
905
mptcp_subflow_tcp_fallback(struct sock * sk,struct mptcp_subflow_context * ctx)906 static inline void mptcp_subflow_tcp_fallback(struct sock *sk,
907 struct mptcp_subflow_context *ctx)
908 {
909 sk->sk_data_ready = sock_def_readable;
910 sk->sk_state_change = ctx->tcp_state_change;
911 sk->sk_write_space = sk_stream_write_space;
912 sk->sk_error_report = ctx->tcp_error_report;
913
914 inet_csk(sk)->icsk_af_ops = ctx->icsk_af_ops;
915 }
916
917 void __init mptcp_proto_init(void);
918 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
919 int __init mptcp_proto_v6_init(void);
920 void __init mptcp_subflow_v6_init(void);
921 #endif
922
923 struct sock *mptcp_sk_clone_init(const struct sock *sk,
924 const struct mptcp_options_received *mp_opt,
925 struct sock *ssk,
926 struct request_sock *req);
927 void mptcp_get_options(const struct sk_buff *skb,
928 struct mptcp_options_received *mp_opt);
929
930 void mptcp_finish_connect(struct sock *sk);
931 void __mptcp_sync_state(struct sock *sk, int state);
932 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout);
933
mptcp_stop_tout_timer(struct sock * sk)934 static inline void mptcp_stop_tout_timer(struct sock *sk)
935 {
936 if (!inet_csk(sk)->icsk_mtup.probe_timestamp)
937 return;
938
939 sk_stop_timer(sk, &inet_csk(sk)->mptcp_tout_timer);
940 inet_csk(sk)->icsk_mtup.probe_timestamp = 0;
941 }
942
mptcp_set_close_tout(struct sock * sk,unsigned long tout)943 static inline void mptcp_set_close_tout(struct sock *sk, unsigned long tout)
944 {
945 /* avoid 0 timestamp, as that means no close timeout */
946 inet_csk(sk)->icsk_mtup.probe_timestamp = tout ? : 1;
947 }
948
mptcp_start_tout_timer(struct sock * sk)949 static inline void mptcp_start_tout_timer(struct sock *sk)
950 {
951 mptcp_set_close_tout(sk, tcp_jiffies32);
952 mptcp_reset_tout_timer(mptcp_sk(sk), 0);
953 }
954
mptcp_is_fully_established(struct sock * sk)955 static inline bool mptcp_is_fully_established(struct sock *sk)
956 {
957 return inet_sk_state_load(sk) == TCP_ESTABLISHED &&
958 READ_ONCE(mptcp_sk(sk)->fully_established);
959 }
960
mptcp_stamp(void)961 static inline u64 mptcp_stamp(void)
962 {
963 return div_u64(tcp_clock_ns(), NSEC_PER_USEC);
964 }
965
966 void mptcp_data_ready(struct sock *sk, struct sock *ssk);
967 bool mptcp_finish_join(struct sock *sk);
968 bool mptcp_schedule_work(struct sock *sk);
969 int mptcp_setsockopt(struct sock *sk, int level, int optname,
970 sockptr_t optval, unsigned int optlen);
971 int mptcp_getsockopt(struct sock *sk, int level, int optname,
972 char __user *optval, int __user *option);
973
974 u64 __mptcp_expand_seq(u64 old_seq, u64 cur_seq);
mptcp_expand_seq(u64 old_seq,u64 cur_seq,bool use_64bit)975 static inline u64 mptcp_expand_seq(u64 old_seq, u64 cur_seq, bool use_64bit)
976 {
977 if (use_64bit)
978 return cur_seq;
979
980 return __mptcp_expand_seq(old_seq, cur_seq);
981 }
982 void __mptcp_check_push(struct sock *sk, struct sock *ssk);
983 void __mptcp_data_acked(struct sock *sk);
984 void __mptcp_error_report(struct sock *sk);
985 bool mptcp_update_rcv_data_fin(struct mptcp_sock *msk, u64 data_fin_seq, bool use_64bit);
mptcp_data_fin_enabled(const struct mptcp_sock * msk)986 static inline bool mptcp_data_fin_enabled(const struct mptcp_sock *msk)
987 {
988 return READ_ONCE(msk->snd_data_fin_enable) &&
989 READ_ONCE(msk->write_seq) == READ_ONCE(msk->snd_nxt);
990 }
991
mptcp_notsent_lowat(const struct sock * sk)992 static inline u32 mptcp_notsent_lowat(const struct sock *sk)
993 {
994 struct net *net = sock_net(sk);
995 u32 val;
996
997 val = READ_ONCE(mptcp_sk(sk)->notsent_lowat);
998 return val ?: READ_ONCE(net->ipv4.sysctl_tcp_notsent_lowat);
999 }
1000
mptcp_stream_memory_free(const struct sock * sk,int wake)1001 static inline bool mptcp_stream_memory_free(const struct sock *sk, int wake)
1002 {
1003 const struct mptcp_sock *msk = mptcp_sk(sk);
1004 u32 notsent_bytes;
1005
1006 notsent_bytes = READ_ONCE(msk->write_seq) - READ_ONCE(msk->snd_nxt);
1007 return (notsent_bytes << wake) < mptcp_notsent_lowat(sk);
1008 }
1009
__mptcp_stream_is_writeable(const struct sock * sk,int wake)1010 static inline bool __mptcp_stream_is_writeable(const struct sock *sk, int wake)
1011 {
1012 return mptcp_stream_memory_free(sk, wake) &&
1013 __sk_stream_is_writeable(sk, wake);
1014 }
1015
mptcp_write_space(struct sock * sk)1016 static inline void mptcp_write_space(struct sock *sk)
1017 {
1018 /* pairs with memory barrier in mptcp_poll */
1019 smp_mb();
1020 if (mptcp_stream_memory_free(sk, 1))
1021 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
1022 }
1023
__mptcp_sync_sndbuf(struct sock * sk)1024 static inline void __mptcp_sync_sndbuf(struct sock *sk)
1025 {
1026 struct mptcp_subflow_context *subflow;
1027 int ssk_sndbuf, new_sndbuf;
1028
1029 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
1030 return;
1031
1032 new_sndbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[0]);
1033 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
1034 ssk_sndbuf = READ_ONCE(mptcp_subflow_tcp_sock(subflow)->sk_sndbuf);
1035
1036 subflow->cached_sndbuf = ssk_sndbuf;
1037 new_sndbuf += ssk_sndbuf;
1038 }
1039
1040 /* the msk max wmem limit is <nr_subflows> * tcp wmem[2] */
1041 WRITE_ONCE(sk->sk_sndbuf, new_sndbuf);
1042 mptcp_write_space(sk);
1043 }
1044
1045 /* The called held both the msk socket and the subflow socket locks,
1046 * possibly under BH
1047 */
__mptcp_propagate_sndbuf(struct sock * sk,struct sock * ssk)1048 static inline void __mptcp_propagate_sndbuf(struct sock *sk, struct sock *ssk)
1049 {
1050 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1051
1052 if (READ_ONCE(ssk->sk_sndbuf) != subflow->cached_sndbuf)
1053 __mptcp_sync_sndbuf(sk);
1054 }
1055
1056 /* the caller held only the subflow socket lock, either in process or
1057 * BH context. Additionally this can be called under the msk data lock,
1058 * so we can't acquire such lock here: let the delegate action acquires
1059 * the needed locks in suitable order.
1060 */
mptcp_propagate_sndbuf(struct sock * sk,struct sock * ssk)1061 static inline void mptcp_propagate_sndbuf(struct sock *sk, struct sock *ssk)
1062 {
1063 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1064
1065 if (likely(READ_ONCE(ssk->sk_sndbuf) == subflow->cached_sndbuf))
1066 return;
1067
1068 local_bh_disable();
1069 mptcp_subflow_delegate(subflow, MPTCP_DELEGATE_SNDBUF);
1070 local_bh_enable();
1071 }
1072
1073 #define MPTCP_TOKEN_MAX_RETRIES 4
1074
1075 void __init mptcp_token_init(void);
mptcp_token_init_request(struct request_sock * req)1076 static inline void mptcp_token_init_request(struct request_sock *req)
1077 {
1078 mptcp_subflow_rsk(req)->token_node.pprev = NULL;
1079 }
1080
1081 int mptcp_token_new_request(struct request_sock *req);
1082 void mptcp_token_destroy_request(struct request_sock *req);
1083 int mptcp_token_new_connect(struct sock *ssk);
1084 void mptcp_token_accept(struct mptcp_subflow_request_sock *r,
1085 struct mptcp_sock *msk);
1086 bool mptcp_token_exists(u32 token);
1087 struct mptcp_sock *mptcp_token_get_sock(struct net *net, u32 token);
1088 struct mptcp_sock *mptcp_token_iter_next(const struct net *net, long *s_slot,
1089 long *s_num);
1090 void mptcp_token_destroy(struct mptcp_sock *msk);
1091
1092 void mptcp_crypto_key_sha(u64 key, u32 *token, u64 *idsn);
1093
1094 void mptcp_crypto_hmac_sha(u64 key1, u64 key2, u8 *msg, int len, void *hmac);
1095 __sum16 __mptcp_make_csum(u64 data_seq, u32 subflow_seq, u16 data_len, __wsum sum);
1096
1097 void __init mptcp_pm_init(void);
1098 void mptcp_pm_data_init(struct mptcp_sock *msk);
1099 void mptcp_pm_data_reset(struct mptcp_sock *msk);
1100 void mptcp_pm_destroy(struct mptcp_sock *msk);
1101 int mptcp_pm_parse_addr(struct nlattr *attr, struct genl_info *info,
1102 struct mptcp_addr_info *addr);
1103 int mptcp_pm_parse_entry(struct nlattr *attr, struct genl_info *info,
1104 bool require_family,
1105 struct mptcp_pm_addr_entry *entry);
1106 bool mptcp_pm_addr_families_match(const struct sock *sk,
1107 const struct mptcp_addr_info *loc,
1108 const struct mptcp_addr_info *rem);
1109 void mptcp_pm_chk_stale(const struct mptcp_sock *msk);
1110 void mptcp_pm_new_connection(struct mptcp_sock *msk, const struct sock *ssk, int server_side);
1111 void mptcp_pm_fully_established(struct mptcp_sock *msk, const struct sock *ssk);
1112 bool mptcp_pm_allow_new_subflow(struct mptcp_sock *msk);
1113 void mptcp_pm_connection_closed(struct mptcp_sock *msk);
1114 void mptcp_pm_subflow_established(struct mptcp_sock *msk);
1115 bool mptcp_pm_nl_check_work_pending(struct mptcp_sock *msk);
1116 void mptcp_pm_subflow_check_next(struct mptcp_sock *msk,
1117 const struct mptcp_subflow_context *subflow);
1118 void mptcp_pm_add_addr_received(const struct sock *ssk,
1119 const struct mptcp_addr_info *addr);
1120 void mptcp_pm_add_addr_echoed(struct mptcp_sock *msk,
1121 const struct mptcp_addr_info *addr);
1122 void mptcp_pm_send_ack(struct mptcp_sock *msk,
1123 struct mptcp_subflow_context *subflow,
1124 bool prio, bool backup);
1125 void mptcp_pm_addr_send_ack(struct mptcp_sock *msk);
1126 void mptcp_pm_nl_rm_addr(struct mptcp_sock *msk, u8 rm_id);
1127 void mptcp_pm_rm_subflow(struct mptcp_sock *msk,
1128 const struct mptcp_rm_list *rm_list);
1129 void mptcp_pm_rm_addr_received(struct mptcp_sock *msk,
1130 const struct mptcp_rm_list *rm_list);
1131 void mptcp_pm_mp_prio_received(struct sock *sk, u8 bkup);
1132 void mptcp_pm_mp_fail_received(struct sock *sk, u64 fail_seq);
1133 int mptcp_pm_mp_prio_send_ack(struct mptcp_sock *msk,
1134 struct mptcp_addr_info *addr,
1135 struct mptcp_addr_info *rem,
1136 u8 bkup);
1137 bool mptcp_pm_announced_alloc(struct mptcp_sock *msk,
1138 const struct mptcp_addr_info *addr);
1139 bool mptcp_pm_announced_remove(struct mptcp_sock *msk,
1140 const struct mptcp_addr_info *addr);
1141 bool mptcp_pm_announced_has_ssk(struct mptcp_sock *msk, const struct sock *ssk);
1142 bool mptcp_pm_has_subflow_saddr(const struct mptcp_sock *msk,
1143 const struct mptcp_addr_info *saddr);
1144 int mptcp_pm_nl_set_flags(struct mptcp_pm_addr_entry *local,
1145 struct genl_info *info);
1146 int mptcp_userspace_pm_set_flags(struct mptcp_pm_addr_entry *local,
1147 struct genl_info *info);
1148 int mptcp_pm_announce_addr(struct mptcp_sock *msk,
1149 const struct mptcp_addr_info *addr,
1150 bool echo);
1151 int mptcp_pm_remove_addr(struct mptcp_sock *msk, const struct mptcp_rm_list *rm_list);
1152
1153 /* the default path manager, used in mptcp_pm_unregister */
1154 extern struct mptcp_pm_ops mptcp_pm_kernel;
1155
1156 struct mptcp_pm_ops *mptcp_pm_find(const char *name);
1157 int mptcp_pm_register(struct mptcp_pm_ops *pm_ops);
1158 void mptcp_pm_unregister(struct mptcp_pm_ops *pm_ops);
1159 int mptcp_pm_validate(struct mptcp_pm_ops *pm_ops);
1160 void mptcp_pm_get_available(char *buf, size_t maxlen);
1161
1162 void mptcp_userspace_pm_free_local_addr_list(struct mptcp_sock *msk);
1163
1164 void mptcp_event(enum mptcp_event_type type, const struct mptcp_sock *msk,
1165 const struct sock *ssk, gfp_t gfp);
1166 void mptcp_event_addr_announced(const struct sock *ssk, const struct mptcp_addr_info *info);
1167 void mptcp_event_addr_removed(const struct mptcp_sock *msk, u8 id);
1168 void mptcp_event_pm_listener(const struct sock *ssk,
1169 enum mptcp_event_type event);
1170 bool mptcp_userspace_pm_active(const struct mptcp_sock *msk);
1171
1172 void mptcp_fastopen_subflow_synack_set_params(struct mptcp_subflow_context *subflow,
1173 struct request_sock *req);
1174 int mptcp_pm_genl_fill_addr(struct sk_buff *msg,
1175 struct netlink_callback *cb,
1176 struct mptcp_pm_addr_entry *entry);
1177
mptcp_pm_should_add_signal(struct mptcp_sock * msk)1178 static inline bool mptcp_pm_should_add_signal(struct mptcp_sock *msk)
1179 {
1180 return READ_ONCE(msk->pm.addr_signal) &
1181 (BIT(MPTCP_ADD_ADDR_SIGNAL) | BIT(MPTCP_ADD_ADDR_ECHO));
1182 }
1183
mptcp_pm_should_add_signal_addr(struct mptcp_sock * msk)1184 static inline bool mptcp_pm_should_add_signal_addr(struct mptcp_sock *msk)
1185 {
1186 return READ_ONCE(msk->pm.addr_signal) & BIT(MPTCP_ADD_ADDR_SIGNAL);
1187 }
1188
mptcp_pm_should_add_signal_echo(struct mptcp_sock * msk)1189 static inline bool mptcp_pm_should_add_signal_echo(struct mptcp_sock *msk)
1190 {
1191 return READ_ONCE(msk->pm.addr_signal) & BIT(MPTCP_ADD_ADDR_ECHO);
1192 }
1193
mptcp_pm_should_rm_signal(struct mptcp_sock * msk)1194 static inline bool mptcp_pm_should_rm_signal(struct mptcp_sock *msk)
1195 {
1196 return READ_ONCE(msk->pm.addr_signal) & BIT(MPTCP_RM_ADDR_SIGNAL);
1197 }
1198
mptcp_pm_is_userspace(const struct mptcp_sock * msk)1199 static inline bool mptcp_pm_is_userspace(const struct mptcp_sock *msk)
1200 {
1201 return READ_ONCE(msk->pm.pm_type) == MPTCP_PM_TYPE_USERSPACE;
1202 }
1203
mptcp_pm_is_kernel(const struct mptcp_sock * msk)1204 static inline bool mptcp_pm_is_kernel(const struct mptcp_sock *msk)
1205 {
1206 return READ_ONCE(msk->pm.pm_type) == MPTCP_PM_TYPE_KERNEL;
1207 }
1208
1209 bool mptcp_pm_add_addr_signal(struct mptcp_sock *msk, int *size, int remaining,
1210 struct mptcp_addr_info *addr, bool *echo,
1211 bool *drop_ts);
1212 bool mptcp_pm_rm_addr_signal(struct mptcp_sock *msk, unsigned int remaining,
1213 struct mptcp_rm_list *rm_list, int *len);
1214 int mptcp_pm_get_local_id(struct mptcp_sock *msk, struct sock_common *skc);
1215 int mptcp_pm_nl_get_local_id(struct mptcp_sock *msk,
1216 struct mptcp_pm_addr_entry *skc);
1217 int mptcp_userspace_pm_get_local_id(struct mptcp_sock *msk,
1218 struct mptcp_pm_addr_entry *skc);
1219 bool mptcp_pm_is_backup(struct mptcp_sock *msk, struct sock_common *skc);
1220 bool mptcp_pm_nl_is_backup(struct mptcp_sock *msk, struct mptcp_addr_info *skc);
1221 bool mptcp_userspace_pm_is_backup(struct mptcp_sock *msk, struct mptcp_addr_info *skc);
1222 int mptcp_pm_nl_dump_addr(struct sk_buff *msg,
1223 struct netlink_callback *cb);
1224 int mptcp_userspace_pm_dump_addr(struct sk_buff *msg,
1225 struct netlink_callback *cb);
1226 int mptcp_pm_nl_get_addr(u8 id, struct mptcp_pm_addr_entry *addr,
1227 struct genl_info *info);
1228 int mptcp_userspace_pm_get_addr(u8 id, struct mptcp_pm_addr_entry *addr,
1229 struct genl_info *info);
1230
subflow_get_local_id(const struct mptcp_subflow_context * subflow)1231 static inline u8 subflow_get_local_id(const struct mptcp_subflow_context *subflow)
1232 {
1233 int local_id = READ_ONCE(subflow->local_id);
1234
1235 if (local_id < 0)
1236 return 0;
1237 return local_id;
1238 }
1239
1240 void __init mptcp_pm_kernel_register(void);
1241 void __init mptcp_pm_userspace_register(void);
1242 void __init mptcp_pm_nl_init(void);
1243 void mptcp_pm_worker(struct mptcp_sock *msk);
1244 void __mptcp_pm_kernel_worker(struct mptcp_sock *msk);
1245 u8 mptcp_pm_get_endp_signal_max(const struct mptcp_sock *msk);
1246 u8 mptcp_pm_get_endp_subflow_max(const struct mptcp_sock *msk);
1247 u8 mptcp_pm_get_endp_laminar_max(const struct mptcp_sock *msk);
1248 u8 mptcp_pm_get_endp_fullmesh_max(const struct mptcp_sock *msk);
1249 u8 mptcp_pm_get_limit_add_addr_accepted(const struct mptcp_sock *msk);
1250 u8 mptcp_pm_get_limit_extra_subflows(const struct mptcp_sock *msk);
1251
1252 /* called under PM lock */
__mptcp_pm_close_subflow(struct mptcp_sock * msk)1253 static inline void __mptcp_pm_close_subflow(struct mptcp_sock *msk)
1254 {
1255 if (!WARN_ON_ONCE(msk->pm.extra_subflows == 0) &&
1256 --msk->pm.extra_subflows < mptcp_pm_get_limit_extra_subflows(msk))
1257 WRITE_ONCE(msk->pm.accept_subflow, true);
1258 }
1259
mptcp_pm_close_subflow(struct mptcp_sock * msk)1260 static inline void mptcp_pm_close_subflow(struct mptcp_sock *msk)
1261 {
1262 spin_lock_bh(&msk->pm.lock);
1263 __mptcp_pm_close_subflow(msk);
1264 spin_unlock_bh(&msk->pm.lock);
1265 }
1266
mptcp_pm_add_addr_c_flag_case(struct mptcp_sock * msk)1267 static inline bool mptcp_pm_add_addr_c_flag_case(struct mptcp_sock *msk)
1268 {
1269 return READ_ONCE(msk->pm.remote_deny_join_id0) &&
1270 msk->pm.local_addr_used == 0 &&
1271 mptcp_pm_get_limit_add_addr_accepted(msk) == 0 &&
1272 msk->pm.extra_subflows < mptcp_pm_get_limit_extra_subflows(msk);
1273 }
1274
1275 void mptcp_sockopt_sync_locked(struct mptcp_sock *msk, struct sock *ssk);
1276
mptcp_get_ext(const struct sk_buff * skb)1277 static inline struct mptcp_ext *mptcp_get_ext(const struct sk_buff *skb)
1278 {
1279 return (struct mptcp_ext *)skb_ext_find(skb, SKB_EXT_MPTCP);
1280 }
1281
1282 void mptcp_diag_subflow_init(struct tcp_ulp_ops *ops);
1283
__mptcp_check_fallback(const struct mptcp_sock * msk)1284 static inline bool __mptcp_check_fallback(const struct mptcp_sock *msk)
1285 {
1286 return test_bit(MPTCP_FALLBACK_DONE, &msk->flags);
1287 }
1288
mptcp_check_fallback(const struct sock * sk)1289 static inline bool mptcp_check_fallback(const struct sock *sk)
1290 {
1291 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
1292 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
1293
1294 return __mptcp_check_fallback(msk);
1295 }
1296
__mptcp_has_initial_subflow(const struct mptcp_sock * msk)1297 static inline bool __mptcp_has_initial_subflow(const struct mptcp_sock *msk)
1298 {
1299 struct sock *ssk = READ_ONCE(msk->first);
1300
1301 return ssk && ((1 << inet_sk_state_load(ssk)) &
1302 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1303 TCPF_SYN_RECV | TCPF_LISTEN));
1304 }
1305
1306 bool __mptcp_try_fallback(struct mptcp_sock *msk, int fb_mib);
1307
mptcp_try_fallback(struct sock * ssk,int fb_mib)1308 static inline bool mptcp_try_fallback(struct sock *ssk, int fb_mib)
1309 {
1310 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1311 struct sock *sk = subflow->conn;
1312 struct mptcp_sock *msk;
1313
1314 msk = mptcp_sk(sk);
1315 if (!__mptcp_try_fallback(msk, fb_mib))
1316 return false;
1317 if (READ_ONCE(msk->snd_data_fin_enable) && !(ssk->sk_shutdown & SEND_SHUTDOWN)) {
1318 gfp_t saved_allocation = ssk->sk_allocation;
1319
1320 /* we are in a atomic (BH) scope, override ssk default for data
1321 * fin allocation
1322 */
1323 ssk->sk_allocation = GFP_ATOMIC;
1324 ssk->sk_shutdown |= SEND_SHUTDOWN;
1325 tcp_shutdown(ssk, SEND_SHUTDOWN);
1326 ssk->sk_allocation = saved_allocation;
1327 }
1328 return true;
1329 }
1330
mptcp_early_fallback(struct mptcp_sock * msk,struct mptcp_subflow_context * subflow,int fb_mib)1331 static inline void mptcp_early_fallback(struct mptcp_sock *msk,
1332 struct mptcp_subflow_context *subflow,
1333 int fb_mib)
1334 {
1335 subflow->request_mptcp = 0;
1336 WARN_ON_ONCE(!__mptcp_try_fallback(msk, fb_mib));
1337 }
1338
mptcp_check_infinite_map(struct sk_buff * skb)1339 static inline bool mptcp_check_infinite_map(struct sk_buff *skb)
1340 {
1341 struct mptcp_ext *mpext;
1342
1343 mpext = skb ? mptcp_get_ext(skb) : NULL;
1344 if (mpext && mpext->infinite_map)
1345 return true;
1346
1347 return false;
1348 }
1349
is_active_ssk(struct mptcp_subflow_context * subflow)1350 static inline bool is_active_ssk(struct mptcp_subflow_context *subflow)
1351 {
1352 return (subflow->request_mptcp || subflow->request_join);
1353 }
1354
subflow_simultaneous_connect(struct sock * sk)1355 static inline bool subflow_simultaneous_connect(struct sock *sk)
1356 {
1357 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
1358
1359 /* Note that the sk state implies !subflow->conn_finished. */
1360 return sk->sk_state == TCP_SYN_RECV && is_active_ssk(subflow);
1361 }
1362
1363 #ifdef CONFIG_SYN_COOKIES
1364 void subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock *subflow_req,
1365 struct sk_buff *skb);
1366 bool mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock *subflow_req,
1367 struct sk_buff *skb);
1368 void __init mptcp_join_cookie_init(void);
1369 #else
1370 static inline void
subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock * subflow_req,struct sk_buff * skb)1371 subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock *subflow_req,
1372 struct sk_buff *skb) {}
1373 static inline bool
mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock * subflow_req,struct sk_buff * skb)1374 mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock *subflow_req,
1375 struct sk_buff *skb)
1376 {
1377 return false;
1378 }
1379
mptcp_join_cookie_init(void)1380 static inline void mptcp_join_cookie_init(void) {}
1381 #endif
1382
1383 #endif /* __MPTCP_PROTOCOL_H */
1384