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 #define MPTCP_RTX_ENABLED 4
120
121 /* MPTCP socket release cb flags */
122 #define MPTCP_PUSH_PENDING 1
123 #define MPTCP_CLEAN_UNA 2
124 #define MPTCP_ERROR_REPORT 3
125 #define MPTCP_RETRANSMIT 4
126 #define MPTCP_FLUSH_JOIN_LIST 5
127 #define MPTCP_SYNC_STATE 6
128 #define MPTCP_SYNC_SNDBUF 7
129
130 struct mptcp_skb_cb {
131 u64 map_seq;
132 u64 end_seq;
133 u32 offset;
134 u8 has_rxtstamp;
135 u8 cant_coalesce;
136 };
137
138 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
139
before64(__u64 seq1,__u64 seq2)140 static inline bool before64(__u64 seq1, __u64 seq2)
141 {
142 return (__s64)(seq1 - seq2) < 0;
143 }
144
145 #define after64(seq2, seq1) before64(seq1, seq2)
146
147 struct mptcp_options_received {
148 u64 sndr_key;
149 u64 rcvr_key;
150 u64 data_ack;
151 u64 data_seq;
152 u32 subflow_seq;
153 u16 data_len;
154 __sum16 csum;
155 struct_group(status,
156 u16 suboptions;
157 u16 use_map:1,
158 dsn64:1,
159 data_fin:1,
160 use_ack:1,
161 ack64:1,
162 mpc_map:1,
163 reset_reason:4,
164 reset_transient:1,
165 echo:1,
166 backup:1,
167 deny_join_id0:1,
168 __unused:2;
169 );
170 u8 join_id;
171 u32 token;
172 u32 nonce;
173 u64 thmac;
174 u8 hmac[MPTCPOPT_HMAC_LEN];
175 struct mptcp_addr_info addr;
176 struct mptcp_rm_list rm_list;
177 u64 ahmac;
178 u64 fail_seq;
179 };
180
mptcp_option(u8 subopt,u8 len,u8 nib,u8 field)181 static inline __be32 mptcp_option(u8 subopt, u8 len, u8 nib, u8 field)
182 {
183 return htonl((TCPOPT_MPTCP << 24) | (len << 16) | (subopt << 12) |
184 ((nib & 0xF) << 8) | field);
185 }
186
187 enum mptcp_pm_status {
188 MPTCP_PM_ADD_ADDR_RECEIVED,
189 MPTCP_PM_ADD_ADDR_SEND_ACK,
190 MPTCP_PM_RM_ADDR_RECEIVED,
191 MPTCP_PM_ESTABLISHED,
192 MPTCP_PM_SUBFLOW_ESTABLISHED,
193 MPTCP_PM_ALREADY_ESTABLISHED, /* persistent status, set after ESTABLISHED event */
194 MPTCP_PM_MPC_ENDPOINT_ACCOUNTED, /* persistent status, set after MPC local address is
195 * accounted int id_avail_bitmap
196 */
197 MPTCP_PM_DESTROYING, /* To fence out PM list allocs */
198 };
199
200 enum mptcp_pm_type {
201 MPTCP_PM_TYPE_KERNEL = 0,
202 MPTCP_PM_TYPE_USERSPACE,
203
204 __MPTCP_PM_TYPE_NR,
205 __MPTCP_PM_TYPE_MAX = __MPTCP_PM_TYPE_NR - 1,
206 };
207
208 /* Status bits below MPTCP_PM_ALREADY_ESTABLISHED need pm worker actions */
209 #define MPTCP_PM_WORK_MASK ((1 << MPTCP_PM_ALREADY_ESTABLISHED) - 1)
210
211 enum mptcp_addr_signal_status {
212 MPTCP_ADD_ADDR_SIGNAL,
213 MPTCP_ADD_ADDR_ECHO,
214 MPTCP_RM_ADDR_SIGNAL,
215 };
216
217 /* max value of mptcp_addr_info.id */
218 #define MPTCP_PM_MAX_ADDR_ID U8_MAX
219
220 struct mptcp_pm_data {
221 struct mptcp_addr_info local;
222 struct mptcp_addr_info remote;
223 struct list_head anno_list;
224 struct list_head userspace_pm_local_addr_list;
225
226 spinlock_t lock; /*protects the whole PM data */
227
228 struct_group(reset,
229
230 u8 addr_signal;
231 bool server_side;
232 bool work_pending;
233 bool accept_addr;
234 bool accept_subflow;
235 bool remote_deny_join_id0;
236 u8 add_addr_signaled;
237 u8 add_addr_accepted;
238 u8 local_addr_used;
239 u8 pm_type;
240 u8 extra_subflows;
241 u8 status;
242
243 );
244
245 DECLARE_BITMAP(id_avail_bitmap, MPTCP_PM_MAX_ADDR_ID + 1);
246 struct mptcp_rm_list rm_list_tx;
247 struct mptcp_rm_list rm_list_rx;
248 };
249
250 struct mptcp_pm_local {
251 struct mptcp_addr_info addr;
252 u32 flags;
253 int ifindex;
254 };
255
256 struct mptcp_pm_addr_entry {
257 struct list_head list;
258 struct mptcp_addr_info addr;
259 u32 flags;
260 int ifindex;
261 struct socket *lsk;
262 };
263
264 struct mptcp_data_frag {
265 struct list_head list;
266 u64 data_seq;
267 u16 data_len;
268 u16 offset;
269 u8 overhead;
270 u8 eor; /* currently using 1 bit */
271 u16 already_sent;
272 struct page *page;
273 };
274
275 /* Arbitrary compromise between as low as possible to react timely to subflow
276 * close event and as big as possible to avoid being fouled by biased large
277 * samples due to peer sending data on a different subflow WRT to the incoming
278 * ack.
279 */
280 #define MPTCP_RTT_SAMPLES 5
281
282 /* MPTCP connection sock */
283 struct mptcp_sock {
284 /* inet_connection_sock must be the first member */
285 struct inet_connection_sock sk;
286 u64 local_key; /* protected by the first subflow socket lock
287 * lockless access read
288 */
289 u64 remote_key; /* same as above */
290 u64 write_seq;
291 u64 bytes_sent;
292 u64 snd_nxt;
293 u64 bytes_received;
294 u64 ack_seq;
295 atomic64_t rcv_wnd_sent;
296 u64 rcv_data_fin_seq;
297 u64 bytes_retrans;
298 u64 bytes_consumed;
299 int snd_burst;
300 int old_wspace;
301 u64 recovery_snd_nxt; /* in recovery mode accept up to this seq;
302 * recovery related fields are under data_lock
303 * protection
304 */
305 u64 bytes_acked;
306 u64 snd_una;
307 u64 wnd_end;
308 u32 last_data_sent;
309 u32 last_data_recv;
310 u32 last_ack_recv;
311 unsigned long timer_ival;
312 u32 token;
313 unsigned long flags;
314 unsigned long cb_flags;
315 bool recovery; /* closing subflow write queue reinjected */
316 bool can_ack;
317 bool fully_established;
318 bool rcv_data_fin;
319 bool snd_data_fin_enable;
320 bool rcv_fastclose;
321 bool use_64bit_ack; /* Set when we received a 64-bit DSN */
322 bool csum_enabled;
323 bool allow_infinite_fallback;
324 u8 pending_state; /* A subflow asked to set this sk_state,
325 * protected by the msk data lock
326 */
327 u8 mpc_endpoint_id;
328 u8 recvmsg_inq:1,
329 cork:1,
330 nodelay:1,
331 fastopening:1,
332 in_accept_queue:1,
333 free_first:1,
334 rcvspace_init:1,
335 fastclosing:1;
336 u32 notsent_lowat;
337 int keepalive_cnt;
338 int keepalive_idle;
339 int keepalive_intvl;
340 int maxseg;
341 struct work_struct work;
342 struct sk_buff *ooo_last_skb;
343 struct rb_root out_of_order_queue;
344 struct list_head conn_list;
345 struct list_head rtx_queue;
346 struct mptcp_data_frag *first_pending;
347 struct list_head join_list;
348 struct sock *first; /* The mptcp ops can safely dereference, using suitable
349 * ONCE annotation, the subflow outside the socket
350 * lock as such sock is freed after close().
351 */
352 struct mptcp_pm_data pm;
353 struct mptcp_sched_ops *sched;
354
355 /* Most recent rtt_us observed by in use incoming subflows. */
356 struct {
357 u32 samples[MPTCP_RTT_SAMPLES];
358 u32 next_sample;
359 } rcv_rtt_est;
360
361 struct {
362 int space; /* bytes copied in last measurement window */
363 int copied; /* bytes copied in this measurement window */
364 u64 time; /* start time of measurement window */
365 } rcvq_space;
366 u8 scaling_ratio;
367 bool allow_subflows;
368
369 u32 subflow_id;
370 u32 setsockopt_seq;
371 char ca_name[TCP_CA_NAME_MAX];
372
373 spinlock_t fallback_lock; /* protects fallback,
374 * allow_infinite_fallback and
375 * allow_join
376 */
377
378 struct list_head backlog_list; /* protected by the data lock */
379 u32 backlog_len;
380 u32 backlog_unaccounted;
381 };
382
383 #define mptcp_data_lock(sk) spin_lock_bh(&(sk)->sk_lock.slock)
384 #define mptcp_data_unlock(sk) spin_unlock_bh(&(sk)->sk_lock.slock)
385
386 #define mptcp_for_each_subflow(__msk, __subflow) \
387 list_for_each_entry(__subflow, &((__msk)->conn_list), node)
388 #define mptcp_for_each_subflow_safe(__msk, __subflow, __tmp) \
389 list_for_each_entry_safe(__subflow, __tmp, &((__msk)->conn_list), node)
390 #define mptcp_next_subflow(__msk, __subflow) \
391 list_next_entry_circular(__subflow, &((__msk)->conn_list), node)
392
393 extern struct genl_family mptcp_genl_family;
394
msk_owned_by_me(const struct mptcp_sock * msk)395 static inline void msk_owned_by_me(const struct mptcp_sock *msk)
396 {
397 sock_owned_by_me((const struct sock *)msk);
398 }
399
400 #ifdef CONFIG_DEBUG_NET
401 /* MPTCP-specific: we might (indirectly) call this helper with the wrong sk */
402 #undef tcp_sk
403 #define tcp_sk(ptr) ({ \
404 typeof(ptr) _ptr = (ptr); \
405 WARN_ON(_ptr->sk_protocol != IPPROTO_TCP); \
406 container_of_const(_ptr, struct tcp_sock, inet_conn.icsk_inet.sk); \
407 })
408 #define mptcp_sk(ptr) ({ \
409 typeof(ptr) _ptr = (ptr); \
410 WARN_ON(_ptr->sk_protocol != IPPROTO_MPTCP); \
411 container_of_const(_ptr, struct mptcp_sock, sk.icsk_inet.sk); \
412 })
413
414 #else /* !CONFIG_DEBUG_NET */
415 #define mptcp_sk(ptr) container_of_const(ptr, struct mptcp_sock, sk.icsk_inet.sk)
416 #endif
417
mptcp_win_from_space(const struct sock * sk,int space)418 static inline int mptcp_win_from_space(const struct sock *sk, int space)
419 {
420 return __tcp_win_from_space(mptcp_sk(sk)->scaling_ratio, space);
421 }
422
mptcp_space_from_win(const struct sock * sk,int win)423 static inline int mptcp_space_from_win(const struct sock *sk, int win)
424 {
425 return __tcp_space_from_win(mptcp_sk(sk)->scaling_ratio, win);
426 }
427
__mptcp_space(const struct sock * sk)428 static inline int __mptcp_space(const struct sock *sk)
429 {
430 return mptcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf) -
431 READ_ONCE(mptcp_sk(sk)->backlog_len) -
432 sk_rmem_alloc_get(sk));
433 }
434
mptcp_send_head(const struct sock * sk)435 static inline struct mptcp_data_frag *mptcp_send_head(const struct sock *sk)
436 {
437 const struct mptcp_sock *msk = mptcp_sk(sk);
438
439 return msk->first_pending;
440 }
441
mptcp_init_rtt_est(struct mptcp_sock * msk)442 static inline void mptcp_init_rtt_est(struct mptcp_sock *msk)
443 {
444 int i;
445
446 for (i = 0; i < MPTCP_RTT_SAMPLES; ++i)
447 msk->rcv_rtt_est.samples[i] = U32_MAX;
448 msk->rcv_rtt_est.next_sample = 0;
449 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
450 }
451
mptcp_rtt_us_est(const struct mptcp_sock * msk)452 static inline u32 mptcp_rtt_us_est(const struct mptcp_sock *msk)
453 {
454 u32 rtt_us = READ_ONCE(msk->rcv_rtt_est.samples[0]);
455 int i;
456
457 /* Lockless access of collected samples. */
458 for (i = 1; i < MPTCP_RTT_SAMPLES; ++i)
459 rtt_us = min(rtt_us, READ_ONCE(msk->rcv_rtt_est.samples[i]));
460 return rtt_us;
461 }
462
mptcp_send_next(struct sock * sk)463 static inline struct mptcp_data_frag *mptcp_send_next(struct sock *sk)
464 {
465 struct mptcp_sock *msk = mptcp_sk(sk);
466 struct mptcp_data_frag *cur;
467
468 cur = msk->first_pending;
469 return list_is_last(&cur->list, &msk->rtx_queue) ? NULL :
470 list_next_entry(cur, list);
471 }
472
mptcp_pending_tail(const struct sock * sk)473 static inline struct mptcp_data_frag *mptcp_pending_tail(const struct sock *sk)
474 {
475 const struct mptcp_sock *msk = mptcp_sk(sk);
476
477 if (!msk->first_pending)
478 return NULL;
479
480 if (WARN_ON_ONCE(list_empty(&msk->rtx_queue)))
481 return NULL;
482
483 return list_last_entry(&msk->rtx_queue, struct mptcp_data_frag, list);
484 }
485
mptcp_rtx_head(struct sock * sk)486 static inline struct mptcp_data_frag *mptcp_rtx_head(struct sock *sk)
487 {
488 struct mptcp_sock *msk = mptcp_sk(sk);
489
490 if (msk->snd_una == msk->snd_nxt)
491 return NULL;
492
493 return list_first_entry_or_null(&msk->rtx_queue, struct mptcp_data_frag, list);
494 }
495
496 struct csum_pseudo_header {
497 __be64 data_seq;
498 __be32 subflow_seq;
499 __be16 data_len;
500 __sum16 csum;
501 };
502
503 struct mptcp_subflow_request_sock {
504 struct tcp_request_sock sk;
505 u16 mp_capable : 1,
506 mp_join : 1,
507 backup : 1,
508 request_bkup : 1,
509 csum_reqd : 1,
510 allow_join_id0 : 1;
511 u8 local_id;
512 u8 remote_id;
513 u64 local_key;
514 u64 idsn;
515 u32 token;
516 u32 ssn_offset;
517 u64 thmac;
518 u32 local_nonce;
519 u32 remote_nonce;
520 struct mptcp_sock *msk;
521 struct hlist_nulls_node token_node;
522 };
523
524 static inline struct mptcp_subflow_request_sock *
mptcp_subflow_rsk(const struct request_sock * rsk)525 mptcp_subflow_rsk(const struct request_sock *rsk)
526 {
527 return (struct mptcp_subflow_request_sock *)rsk;
528 }
529
530 struct mptcp_delegated_action {
531 struct napi_struct napi;
532 local_lock_t bh_lock;
533 struct list_head head;
534 };
535
536 DECLARE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
537
538 #define MPTCP_DELEGATE_SCHEDULED 0
539 #define MPTCP_DELEGATE_SEND 1
540 #define MPTCP_DELEGATE_ACK 2
541 #define MPTCP_DELEGATE_SNDBUF 3
542
543 #define MPTCP_DELEGATE_ACTIONS_MASK (~BIT(MPTCP_DELEGATE_SCHEDULED))
544 /* MPTCP subflow context */
545 struct mptcp_subflow_context {
546 struct list_head node;/* conn_list of subflows */
547
548 struct_group(reset,
549
550 unsigned long avg_pacing_rate; /* protected by msk socket lock */
551 u64 local_key;
552 u64 remote_key;
553 u64 idsn;
554 u64 map_seq;
555 u64 rcv_wnd_sent;
556 u32 snd_isn;
557 u32 token;
558 u32 rel_write_seq;
559 u32 map_subflow_seq;
560 u32 ssn_offset;
561 u32 map_data_len;
562 __wsum map_data_csum;
563 u32 map_csum_len;
564 u32 prev_rtt_seq;
565 u32 request_mptcp : 1, /* send MP_CAPABLE */
566 request_join : 1, /* send MP_JOIN */
567 request_bkup : 1,
568 mp_capable : 1, /* remote is MPTCP capable */
569 mp_join : 1, /* remote is JOINing */
570 pm_notified : 1, /* PM hook called for established status */
571 conn_finished : 1,
572 map_valid : 1,
573 map_csum_reqd : 1,
574 map_data_fin : 1,
575 mpc_map : 1,
576 backup : 1,
577 send_mp_prio : 1,
578 send_mp_fail : 1,
579 send_fastclose : 1,
580 send_infinite_map : 1,
581 remote_key_valid : 1, /* received the peer key from */
582 disposable : 1, /* ctx can be free at ulp release time */
583 closing : 1, /* must not pass rx data to msk anymore */
584 valid_csum_seen : 1, /* at least one csum validated */
585 is_mptfo : 1, /* subflow is doing TFO */
586 close_event_done : 1, /* has done the post-closed part */
587 mpc_drop : 1, /* the MPC option has been dropped in a rtx */
588 resetting : 1, /* subflow is resetting */
589 __unused : 8;
590 bool data_avail;
591 bool scheduled;
592 bool pm_listener; /* a listener managed by the kernel PM? */
593 bool fully_established; /* path validated */
594 u32 lent_mem_frag;
595 u32 remote_nonce;
596 u64 thmac;
597 u32 local_nonce;
598 u32 remote_token;
599 union {
600 u8 hmac[MPTCPOPT_HMAC_LEN]; /* MPJ subflow only */
601 u64 iasn; /* initial ack sequence number, MPC subflows only */
602 };
603 s16 local_id; /* if negative not initialized yet */
604 u8 remote_id;
605 u8 reset_seen:1;
606 u8 reset_transient:1;
607 u8 reset_reason:4;
608 u8 stale_count; /* Protected by the msk socket lock */
609 u8 stale; /* Protected by the msk socket lock,
610 * if set the subflow is unable to snd/rcv
611 * data, the schedule should skip it
612 */
613
614 u32 subflow_id;
615
616 long delegated_status;
617 unsigned long fail_tout;
618
619 );
620
621 struct list_head delegated_node; /* link into delegated_action, protected by local BH */
622
623 u32 setsockopt_seq;
624 u32 stale_rcv_tstamp;
625 int cached_sndbuf; /* sndbuf size when last synced with the msk sndbuf,
626 * protected by the msk socket lock
627 */
628
629 struct sock *tcp_sock; /* tcp sk backpointer */
630 struct sock *conn; /* parent mptcp_sock */
631 const struct inet_connection_sock_af_ops *icsk_af_ops;
632 void (*tcp_state_change)(struct sock *sk);
633 void (*tcp_error_report)(struct sock *sk);
634
635 struct rcu_head rcu;
636 };
637
638 static inline struct mptcp_subflow_context *
mptcp_subflow_ctx(const struct sock * sk)639 mptcp_subflow_ctx(const struct sock *sk)
640 {
641 const struct inet_connection_sock *icsk = inet_csk(sk);
642
643 /* Use RCU on icsk_ulp_data only for sock diag code */
644 return (__force struct mptcp_subflow_context *)icsk->icsk_ulp_data;
645 }
646
647 static inline struct sock *
mptcp_subflow_tcp_sock(const struct mptcp_subflow_context * subflow)648 mptcp_subflow_tcp_sock(const struct mptcp_subflow_context *subflow)
649 {
650 return subflow->tcp_sock;
651 }
652
653 static inline void
mptcp_subflow_ctx_reset(struct mptcp_subflow_context * subflow)654 mptcp_subflow_ctx_reset(struct mptcp_subflow_context *subflow)
655 {
656 memset(&subflow->reset, 0, sizeof(subflow->reset));
657 subflow->request_mptcp = 1;
658 WRITE_ONCE(subflow->local_id, -1);
659 }
660
661 /* Convert reset reasons in MPTCP to enum sk_rst_reason type */
662 static inline enum sk_rst_reason
sk_rst_convert_mptcp_reason(u32 reason)663 sk_rst_convert_mptcp_reason(u32 reason)
664 {
665 switch (reason) {
666 case MPTCP_RST_EUNSPEC:
667 return SK_RST_REASON_MPTCP_RST_EUNSPEC;
668 case MPTCP_RST_EMPTCP:
669 return SK_RST_REASON_MPTCP_RST_EMPTCP;
670 case MPTCP_RST_ERESOURCE:
671 return SK_RST_REASON_MPTCP_RST_ERESOURCE;
672 case MPTCP_RST_EPROHIBIT:
673 return SK_RST_REASON_MPTCP_RST_EPROHIBIT;
674 case MPTCP_RST_EWQ2BIG:
675 return SK_RST_REASON_MPTCP_RST_EWQ2BIG;
676 case MPTCP_RST_EBADPERF:
677 return SK_RST_REASON_MPTCP_RST_EBADPERF;
678 case MPTCP_RST_EMIDDLEBOX:
679 return SK_RST_REASON_MPTCP_RST_EMIDDLEBOX;
680 default:
681 /* It should not happen, or else errors may occur
682 * in MPTCP layer
683 */
684 return SK_RST_REASON_ERROR;
685 }
686 }
687
688 static inline void
mptcp_send_active_reset_reason(struct sock * sk)689 mptcp_send_active_reset_reason(struct sock *sk)
690 {
691 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
692 enum sk_rst_reason reason;
693
694 reason = sk_rst_convert_mptcp_reason(subflow->reset_reason);
695 tcp_send_active_reset(sk, reason);
696 }
697
698 /* Made the fwd mem carried by the given skb available to the msk,
699 * To be paired with a previous mptcp_subflow_lend_fwdmem() before freeing
700 * the skb or setting the skb ownership.
701 */
mptcp_borrow_fwdmem(struct sock * sk,struct sk_buff * skb)702 static inline void mptcp_borrow_fwdmem(struct sock *sk, struct sk_buff *skb)
703 {
704 struct sock *ssk = skb->sk;
705
706 /* The subflow just lend the skb fwd memory; if the subflow meanwhile
707 * closed, mptcp_close_ssk() already released the ssk rcv memory.
708 */
709 DEBUG_NET_WARN_ON_ONCE(skb->destructor);
710 sk_forward_alloc_add(sk, skb->truesize);
711 if (!ssk)
712 return;
713
714 atomic_sub(skb->truesize, &ssk->sk_rmem_alloc);
715 skb->sk = NULL;
716 }
717
718 static inline void
__mptcp_subflow_lend_fwdmem(struct mptcp_subflow_context * subflow,int size)719 __mptcp_subflow_lend_fwdmem(struct mptcp_subflow_context *subflow, int size)
720 {
721 int frag = (subflow->lent_mem_frag + size) & (PAGE_SIZE - 1);
722
723 subflow->lent_mem_frag = frag;
724 }
725
726 static inline void
mptcp_subflow_lend_fwdmem(struct mptcp_subflow_context * subflow,struct sk_buff * skb)727 mptcp_subflow_lend_fwdmem(struct mptcp_subflow_context *subflow,
728 struct sk_buff *skb)
729 {
730 __mptcp_subflow_lend_fwdmem(subflow, skb->truesize);
731 skb->destructor = NULL;
732 }
733
734 static inline u64
mptcp_subflow_get_map_offset(const struct mptcp_subflow_context * subflow)735 mptcp_subflow_get_map_offset(const struct mptcp_subflow_context *subflow)
736 {
737 return tcp_sk(mptcp_subflow_tcp_sock(subflow))->copied_seq -
738 subflow->ssn_offset -
739 subflow->map_subflow_seq;
740 }
741
742 static inline u64
mptcp_subflow_get_mapped_dsn(const struct mptcp_subflow_context * subflow)743 mptcp_subflow_get_mapped_dsn(const struct mptcp_subflow_context *subflow)
744 {
745 return subflow->map_seq + mptcp_subflow_get_map_offset(subflow);
746 }
747
748 void mptcp_subflow_process_delegated(struct sock *ssk, long actions);
749
mptcp_subflow_delegate(struct mptcp_subflow_context * subflow,int action)750 static inline void mptcp_subflow_delegate(struct mptcp_subflow_context *subflow, int action)
751 {
752 long old, set_bits = BIT(MPTCP_DELEGATE_SCHEDULED) | BIT(action);
753 struct mptcp_delegated_action *delegated;
754 bool schedule;
755
756 /* the caller held the subflow bh socket lock */
757 lockdep_assert_in_softirq();
758
759 /* The implied barrier pairs with tcp_release_cb_override()
760 * mptcp_napi_poll(), and ensures the below list check sees list
761 * updates done prior to delegated status bits changes
762 */
763 old = set_mask_bits(&subflow->delegated_status, 0, set_bits);
764 if (!(old & BIT(MPTCP_DELEGATE_SCHEDULED))) {
765 if (WARN_ON_ONCE(!list_empty(&subflow->delegated_node)))
766 return;
767
768 local_lock_nested_bh(&mptcp_delegated_actions.bh_lock);
769 delegated = this_cpu_ptr(&mptcp_delegated_actions);
770 schedule = list_empty(&delegated->head);
771 list_add_tail(&subflow->delegated_node, &delegated->head);
772 local_unlock_nested_bh(&mptcp_delegated_actions.bh_lock);
773 sock_hold(mptcp_subflow_tcp_sock(subflow));
774 if (schedule)
775 napi_schedule(&delegated->napi);
776 }
777 }
778
779 static inline struct mptcp_subflow_context *
mptcp_subflow_delegated_next(struct mptcp_delegated_action * delegated)780 mptcp_subflow_delegated_next(struct mptcp_delegated_action *delegated)
781 {
782 struct mptcp_subflow_context *ret;
783
784 local_lock_nested_bh(&mptcp_delegated_actions.bh_lock);
785 if (list_empty(&delegated->head)) {
786 local_unlock_nested_bh(&mptcp_delegated_actions.bh_lock);
787 return NULL;
788 }
789
790 ret = list_first_entry(&delegated->head, struct mptcp_subflow_context, delegated_node);
791 list_del_init(&ret->delegated_node);
792 local_unlock_nested_bh(&mptcp_delegated_actions.bh_lock);
793 return ret;
794 }
795
796 void __mptcp_inherit_memcg(struct sock *sk, struct sock *ssk, gfp_t gfp);
797 void __mptcp_inherit_cgrp_data(struct sock *sk, struct sock *ssk);
798
799 int mptcp_is_enabled(const struct net *net);
800 unsigned int mptcp_get_add_addr_timeout(const struct net *net);
801 int mptcp_is_checksum_enabled(const struct net *net);
802 int mptcp_allow_join_id0(const struct net *net);
803 unsigned int mptcp_stale_loss_cnt(const struct net *net);
804 unsigned int mptcp_close_timeout(const struct sock *sk);
805 int mptcp_get_pm_type(const struct net *net);
806 const char *mptcp_get_path_manager(const struct net *net);
807 const char *mptcp_get_scheduler(const struct net *net);
808 unsigned int mptcp_add_addr_v6_port_drop_ts(const struct net *net);
809
810 void mptcp_active_disable(struct sock *sk);
811 bool mptcp_active_should_disable(struct sock *ssk);
812 void mptcp_active_enable(struct sock *sk);
813
814 void mptcp_get_available_schedulers(char *buf, size_t maxlen);
815 void __mptcp_subflow_fully_established(struct mptcp_sock *msk,
816 struct mptcp_subflow_context *subflow,
817 const struct mptcp_options_received *mp_opt);
818 bool __mptcp_retransmit_pending_data(struct sock *sk);
819 void mptcp_check_and_set_pending(struct sock *sk);
820 void __mptcp_push_pending(struct sock *sk, unsigned int flags);
821 bool mptcp_subflow_data_available(struct sock *sk);
822 void __init mptcp_subflow_init(void);
823 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how);
824 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
825 struct mptcp_subflow_context *subflow);
826 void __mptcp_subflow_send_ack(struct sock *ssk);
827 void mptcp_subflow_reset(struct sock *ssk);
828 void mptcp_subflow_queue_clean(struct sock *sk, struct sock *ssk);
829 void mptcp_sock_graft(struct sock *sk, struct socket *parent);
830 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk);
831 bool __mptcp_close(struct sock *sk, long timeout);
832 void mptcp_cancel_work(struct sock *sk);
833 void __mptcp_unaccepted_force_close(struct sock *sk);
834 void mptcp_set_state(struct sock *sk, int state);
835
836 bool mptcp_addresses_equal(const struct mptcp_addr_info *a,
837 const struct mptcp_addr_info *b, bool use_port);
838 void mptcp_local_address(const struct sock_common *skc,
839 struct mptcp_addr_info *addr);
840 void mptcp_remote_address(const struct sock_common *skc,
841 struct mptcp_addr_info *addr);
842
843 /* called with sk socket lock held */
844 int __mptcp_subflow_connect(struct sock *sk, const struct mptcp_pm_local *local,
845 const struct mptcp_addr_info *remote);
846 int mptcp_subflow_create_socket(struct sock *sk, unsigned short family,
847 struct socket **new_sock);
848 void mptcp_info2sockaddr(const struct mptcp_addr_info *info,
849 struct sockaddr_storage *addr,
850 unsigned short family);
851 struct mptcp_sched_ops *mptcp_sched_find(const char *name);
852 int mptcp_validate_scheduler(struct mptcp_sched_ops *sched);
853 int mptcp_register_scheduler(struct mptcp_sched_ops *sched);
854 void mptcp_unregister_scheduler(struct mptcp_sched_ops *sched);
855 void mptcp_sched_init(void);
856 int mptcp_init_sched(struct mptcp_sock *msk,
857 struct mptcp_sched_ops *sched);
858 void mptcp_release_sched(struct mptcp_sock *msk);
859 void mptcp_subflow_set_scheduled(struct mptcp_subflow_context *subflow,
860 bool scheduled);
861 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk);
862 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk);
863 int mptcp_sched_get_send(struct mptcp_sock *msk);
864 int mptcp_sched_get_retrans(struct mptcp_sock *msk);
865
mptcp_data_avail(const struct mptcp_sock * msk)866 static inline u64 mptcp_data_avail(const struct mptcp_sock *msk)
867 {
868 return READ_ONCE(msk->bytes_received) - READ_ONCE(msk->bytes_consumed);
869 }
870
mptcp_epollin_ready(const struct sock * sk)871 static inline bool mptcp_epollin_ready(const struct sock *sk)
872 {
873 u64 data_avail = mptcp_data_avail(mptcp_sk(sk));
874
875 if (!data_avail)
876 return false;
877
878 /* mptcp doesn't have to deal with small skbs in the receive queue,
879 * as it can always coalesce them
880 */
881 return (data_avail >= sk->sk_rcvlowat) ||
882 tcp_under_memory_pressure(sk);
883 }
884
885 int mptcp_set_rcvlowat(struct sock *sk, int val);
886
__tcp_can_send(const struct sock * ssk)887 static inline bool __tcp_can_send(const struct sock *ssk)
888 {
889 /* only send if our side has not closed yet */
890 return ((1 << inet_sk_state_load(ssk)) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
891 }
892
__mptcp_subflow_active(struct mptcp_subflow_context * subflow)893 static inline bool __mptcp_subflow_active(struct mptcp_subflow_context *subflow)
894 {
895 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
896 if (subflow->request_join && !READ_ONCE(subflow->fully_established))
897 return false;
898
899 return __tcp_can_send(mptcp_subflow_tcp_sock(subflow));
900 }
901
902 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow);
903
904 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow);
905
906 void mptcp_subflow_drop_ctx(struct sock *ssk);
907
mptcp_subflow_tcp_fallback(struct sock * sk,struct mptcp_subflow_context * ctx)908 static inline void mptcp_subflow_tcp_fallback(struct sock *sk,
909 struct mptcp_subflow_context *ctx)
910 {
911 sk->sk_data_ready = sock_def_readable;
912 sk->sk_state_change = ctx->tcp_state_change;
913 sk->sk_write_space = sk_stream_write_space;
914 sk->sk_error_report = ctx->tcp_error_report;
915
916 inet_csk(sk)->icsk_af_ops = ctx->icsk_af_ops;
917 }
918
919 void __init mptcp_proto_init(void);
920 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
921 int __init mptcp_proto_v6_init(void);
922 void __init mptcp_subflow_v6_init(void);
923 #endif
924
925 struct sock *mptcp_sk_clone_init(const struct sock *sk,
926 const struct mptcp_options_received *mp_opt,
927 struct sock *ssk,
928 struct request_sock *req);
929 void mptcp_get_options(const struct sk_buff *skb,
930 struct mptcp_options_received *mp_opt);
931
932 void mptcp_finish_connect(struct sock *sk);
933 void __mptcp_sync_state(struct sock *sk, int state);
934 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout);
935
mptcp_stop_tout_timer(struct sock * sk)936 static inline void mptcp_stop_tout_timer(struct sock *sk)
937 {
938 if (!inet_csk(sk)->icsk_mtup.probe_timestamp)
939 return;
940
941 sk_stop_timer(sk, &inet_csk(sk)->mptcp_tout_timer);
942 inet_csk(sk)->icsk_mtup.probe_timestamp = 0;
943 }
944
mptcp_set_close_tout(struct sock * sk,unsigned long tout)945 static inline void mptcp_set_close_tout(struct sock *sk, unsigned long tout)
946 {
947 /* avoid 0 timestamp, as that means no close timeout */
948 inet_csk(sk)->icsk_mtup.probe_timestamp = tout ? : 1;
949 }
950
mptcp_start_tout_timer(struct sock * sk)951 static inline void mptcp_start_tout_timer(struct sock *sk)
952 {
953 mptcp_set_close_tout(sk, tcp_jiffies32);
954 mptcp_reset_tout_timer(mptcp_sk(sk), 0);
955 }
956
mptcp_is_fully_established(struct sock * sk)957 static inline bool mptcp_is_fully_established(struct sock *sk)
958 {
959 return inet_sk_state_load(sk) == TCP_ESTABLISHED &&
960 READ_ONCE(mptcp_sk(sk)->fully_established);
961 }
962
mptcp_stamp(void)963 static inline u64 mptcp_stamp(void)
964 {
965 return div_u64(tcp_clock_ns(), NSEC_PER_USEC);
966 }
967
968 void mptcp_data_ready(struct sock *sk, struct sock *ssk);
969 bool mptcp_finish_join(struct sock *sk);
970 bool mptcp_schedule_work(struct sock *sk);
971 int mptcp_setsockopt(struct sock *sk, int level, int optname,
972 sockptr_t optval, unsigned int optlen);
973 int mptcp_getsockopt(struct sock *sk, int level, int optname,
974 char __user *optval, int __user *option);
975
976 u64 __mptcp_expand_seq(u64 old_seq, u64 cur_seq);
mptcp_expand_seq(u64 old_seq,u64 cur_seq,bool use_64bit)977 static inline u64 mptcp_expand_seq(u64 old_seq, u64 cur_seq, bool use_64bit)
978 {
979 if (use_64bit)
980 return cur_seq;
981
982 return __mptcp_expand_seq(old_seq, cur_seq);
983 }
984 void __mptcp_check_push(struct sock *sk, struct sock *ssk);
985 void __mptcp_data_acked(struct sock *sk);
986 void __mptcp_error_report(struct sock *sk);
987 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)988 static inline bool mptcp_data_fin_enabled(const struct mptcp_sock *msk)
989 {
990 return READ_ONCE(msk->snd_data_fin_enable) &&
991 READ_ONCE(msk->write_seq) == READ_ONCE(msk->snd_nxt);
992 }
993
mptcp_notsent_lowat(const struct sock * sk)994 static inline u32 mptcp_notsent_lowat(const struct sock *sk)
995 {
996 struct net *net = sock_net(sk);
997 u32 val;
998
999 val = READ_ONCE(mptcp_sk(sk)->notsent_lowat);
1000 return val ?: READ_ONCE(net->ipv4.sysctl_tcp_notsent_lowat);
1001 }
1002
mptcp_stream_memory_free(const struct sock * sk,int wake)1003 static inline bool mptcp_stream_memory_free(const struct sock *sk, int wake)
1004 {
1005 const struct mptcp_sock *msk = mptcp_sk(sk);
1006 u32 notsent_bytes;
1007
1008 notsent_bytes = READ_ONCE(msk->write_seq) - READ_ONCE(msk->snd_nxt);
1009 return (notsent_bytes << wake) < mptcp_notsent_lowat(sk);
1010 }
1011
__mptcp_stream_is_writeable(const struct sock * sk,int wake)1012 static inline bool __mptcp_stream_is_writeable(const struct sock *sk, int wake)
1013 {
1014 return mptcp_stream_memory_free(sk, wake) &&
1015 __sk_stream_is_writeable(sk, wake);
1016 }
1017
mptcp_write_space(struct sock * sk)1018 static inline void mptcp_write_space(struct sock *sk)
1019 {
1020 /* pairs with memory barrier in mptcp_poll */
1021 smp_mb();
1022 if (mptcp_stream_memory_free(sk, 1))
1023 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
1024 }
1025
__mptcp_sync_sndbuf(struct sock * sk)1026 static inline void __mptcp_sync_sndbuf(struct sock *sk)
1027 {
1028 struct mptcp_subflow_context *subflow;
1029 int ssk_sndbuf, new_sndbuf;
1030
1031 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
1032 return;
1033
1034 new_sndbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[0]);
1035 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
1036 ssk_sndbuf = READ_ONCE(mptcp_subflow_tcp_sock(subflow)->sk_sndbuf);
1037
1038 subflow->cached_sndbuf = ssk_sndbuf;
1039 new_sndbuf += ssk_sndbuf;
1040 }
1041
1042 /* the msk max wmem limit is <nr_subflows> * tcp wmem[2] */
1043 WRITE_ONCE(sk->sk_sndbuf, new_sndbuf);
1044 mptcp_write_space(sk);
1045 }
1046
1047 /* The called held both the msk socket and the subflow socket locks,
1048 * possibly under BH
1049 */
__mptcp_propagate_sndbuf(struct sock * sk,struct sock * ssk)1050 static inline void __mptcp_propagate_sndbuf(struct sock *sk, struct sock *ssk)
1051 {
1052 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1053
1054 if (READ_ONCE(ssk->sk_sndbuf) != subflow->cached_sndbuf)
1055 __mptcp_sync_sndbuf(sk);
1056 }
1057
1058 /* the caller held only the subflow socket lock, either in process or
1059 * BH context. Additionally this can be called under the msk data lock,
1060 * so we can't acquire such lock here: let the delegate action acquires
1061 * the needed locks in suitable order.
1062 */
mptcp_propagate_sndbuf(struct sock * sk,struct sock * ssk)1063 static inline void mptcp_propagate_sndbuf(struct sock *sk, struct sock *ssk)
1064 {
1065 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1066
1067 if (likely(READ_ONCE(ssk->sk_sndbuf) == subflow->cached_sndbuf))
1068 return;
1069
1070 local_bh_disable();
1071 mptcp_subflow_delegate(subflow, MPTCP_DELEGATE_SNDBUF);
1072 local_bh_enable();
1073 }
1074
1075 #define MPTCP_TOKEN_MAX_RETRIES 4
1076
1077 void __init mptcp_token_init(void);
mptcp_token_init_request(struct request_sock * req)1078 static inline void mptcp_token_init_request(struct request_sock *req)
1079 {
1080 mptcp_subflow_rsk(req)->token_node.pprev = NULL;
1081 }
1082
1083 int mptcp_token_new_request(struct request_sock *req);
1084 void mptcp_token_destroy_request(struct request_sock *req);
1085 int mptcp_token_new_connect(struct sock *ssk);
1086 void mptcp_token_accept(struct mptcp_subflow_request_sock *r,
1087 struct mptcp_sock *msk);
1088 bool mptcp_token_exists(u32 token);
1089 struct mptcp_sock *mptcp_token_get_sock(struct net *net, u32 token);
1090 struct mptcp_sock *mptcp_token_iter_next(const struct net *net, long *s_slot,
1091 long *s_num);
1092 void mptcp_token_destroy(struct mptcp_sock *msk);
1093
1094 void mptcp_crypto_key_sha(u64 key, u32 *token, u64 *idsn);
1095
1096 void mptcp_crypto_hmac_sha(u64 key1, u64 key2, u8 *msg, int len, void *hmac);
1097 __sum16 __mptcp_make_csum(u64 data_seq, u32 subflow_seq, u16 data_len, __wsum sum);
1098
1099 void __init mptcp_pm_init(void);
1100 void mptcp_pm_data_init(struct mptcp_sock *msk);
1101 void mptcp_pm_data_reset(struct mptcp_sock *msk);
1102 void mptcp_pm_destroy(struct mptcp_sock *msk);
1103 int mptcp_pm_parse_addr(struct nlattr *attr, struct genl_info *info,
1104 struct mptcp_addr_info *addr);
1105 int mptcp_pm_parse_entry(struct nlattr *attr, struct genl_info *info,
1106 bool require_family,
1107 struct mptcp_pm_addr_entry *entry);
1108 bool mptcp_pm_addr_families_match(const struct sock *sk,
1109 const struct mptcp_addr_info *loc,
1110 const struct mptcp_addr_info *rem);
1111 void mptcp_pm_chk_stale(const struct mptcp_sock *msk);
1112 void mptcp_pm_new_connection(struct mptcp_sock *msk, const struct sock *ssk, int server_side);
1113 void mptcp_pm_fully_established(struct mptcp_sock *msk, const struct sock *ssk);
1114 bool mptcp_pm_allow_new_subflow(struct mptcp_sock *msk);
1115 void mptcp_pm_connection_closed(struct mptcp_sock *msk);
1116 void mptcp_pm_subflow_established(struct mptcp_sock *msk);
1117 bool mptcp_pm_nl_check_work_pending(struct mptcp_sock *msk);
1118 void mptcp_pm_subflow_check_next(struct mptcp_sock *msk,
1119 const struct mptcp_subflow_context *subflow);
1120 void mptcp_pm_add_addr_received(const struct sock *ssk,
1121 const struct mptcp_addr_info *addr);
1122 void mptcp_pm_add_addr_echoed(struct mptcp_sock *msk,
1123 const struct mptcp_addr_info *addr);
1124 void mptcp_pm_send_ack(struct mptcp_sock *msk,
1125 struct mptcp_subflow_context *subflow,
1126 bool prio, bool backup);
1127 void mptcp_pm_addr_send_ack(struct mptcp_sock *msk);
1128 void mptcp_pm_nl_rm_addr(struct mptcp_sock *msk, u8 rm_id);
1129 void mptcp_pm_rm_subflow(struct mptcp_sock *msk,
1130 const struct mptcp_rm_list *rm_list);
1131 void mptcp_pm_rm_addr_received(struct mptcp_sock *msk,
1132 const struct mptcp_rm_list *rm_list);
1133 void mptcp_pm_mp_prio_received(struct sock *sk, u8 bkup);
1134 void mptcp_pm_mp_fail_received(struct sock *sk, u64 fail_seq);
1135 int mptcp_pm_mp_prio_send_ack(struct mptcp_sock *msk,
1136 struct mptcp_addr_info *addr,
1137 struct mptcp_addr_info *rem,
1138 u8 bkup);
1139 bool mptcp_pm_announced_alloc(struct mptcp_sock *msk,
1140 const struct mptcp_addr_info *addr);
1141 bool mptcp_pm_announced_remove(struct mptcp_sock *msk,
1142 const struct mptcp_addr_info *addr);
1143 bool mptcp_pm_announced_has_ssk(struct mptcp_sock *msk, const struct sock *ssk);
1144 bool mptcp_pm_has_subflow_saddr(const struct mptcp_sock *msk,
1145 const struct mptcp_addr_info *saddr);
1146 int mptcp_pm_nl_set_flags(struct mptcp_pm_addr_entry *local,
1147 struct genl_info *info);
1148 int mptcp_userspace_pm_set_flags(struct mptcp_pm_addr_entry *local,
1149 struct genl_info *info);
1150 int mptcp_pm_announce_addr(struct mptcp_sock *msk,
1151 const struct mptcp_addr_info *addr,
1152 bool echo);
1153 int mptcp_pm_remove_addr(struct mptcp_sock *msk, const struct mptcp_rm_list *rm_list);
1154
1155 /* the default path manager, used in mptcp_pm_unregister */
1156 extern struct mptcp_pm_ops mptcp_pm_kernel;
1157
1158 struct mptcp_pm_ops *mptcp_pm_find(const char *name);
1159 int mptcp_pm_register(struct mptcp_pm_ops *pm_ops);
1160 void mptcp_pm_unregister(struct mptcp_pm_ops *pm_ops);
1161 int mptcp_pm_validate(struct mptcp_pm_ops *pm_ops);
1162 void mptcp_pm_get_available(char *buf, size_t maxlen);
1163
1164 void mptcp_userspace_pm_free_local_addr_list(struct mptcp_sock *msk);
1165
1166 void mptcp_event(enum mptcp_event_type type, const struct mptcp_sock *msk,
1167 const struct sock *ssk, gfp_t gfp);
1168 void mptcp_event_addr_announced(const struct sock *ssk, const struct mptcp_addr_info *info);
1169 void mptcp_event_addr_removed(const struct mptcp_sock *msk, u8 id);
1170 void mptcp_event_pm_listener(const struct sock *ssk,
1171 enum mptcp_event_type event);
1172 bool mptcp_userspace_pm_active(const struct mptcp_sock *msk);
1173
1174 void mptcp_fastopen_subflow_synack_set_params(struct mptcp_subflow_context *subflow,
1175 struct request_sock *req);
1176 int mptcp_pm_genl_fill_addr(struct sk_buff *msg,
1177 struct netlink_callback *cb,
1178 struct mptcp_pm_addr_entry *entry);
1179
mptcp_pm_should_add_signal(struct mptcp_sock * msk)1180 static inline bool mptcp_pm_should_add_signal(struct mptcp_sock *msk)
1181 {
1182 return READ_ONCE(msk->pm.addr_signal) &
1183 (BIT(MPTCP_ADD_ADDR_SIGNAL) | BIT(MPTCP_ADD_ADDR_ECHO));
1184 }
1185
mptcp_pm_should_add_signal_addr(struct mptcp_sock * msk)1186 static inline bool mptcp_pm_should_add_signal_addr(struct mptcp_sock *msk)
1187 {
1188 return READ_ONCE(msk->pm.addr_signal) & BIT(MPTCP_ADD_ADDR_SIGNAL);
1189 }
1190
mptcp_pm_should_add_signal_echo(struct mptcp_sock * msk)1191 static inline bool mptcp_pm_should_add_signal_echo(struct mptcp_sock *msk)
1192 {
1193 return READ_ONCE(msk->pm.addr_signal) & BIT(MPTCP_ADD_ADDR_ECHO);
1194 }
1195
mptcp_pm_should_rm_signal(struct mptcp_sock * msk)1196 static inline bool mptcp_pm_should_rm_signal(struct mptcp_sock *msk)
1197 {
1198 return READ_ONCE(msk->pm.addr_signal) & BIT(MPTCP_RM_ADDR_SIGNAL);
1199 }
1200
mptcp_pm_is_userspace(const struct mptcp_sock * msk)1201 static inline bool mptcp_pm_is_userspace(const struct mptcp_sock *msk)
1202 {
1203 return READ_ONCE(msk->pm.pm_type) == MPTCP_PM_TYPE_USERSPACE;
1204 }
1205
mptcp_pm_is_kernel(const struct mptcp_sock * msk)1206 static inline bool mptcp_pm_is_kernel(const struct mptcp_sock *msk)
1207 {
1208 return READ_ONCE(msk->pm.pm_type) == MPTCP_PM_TYPE_KERNEL;
1209 }
1210
1211 bool mptcp_pm_add_addr_signal(struct mptcp_sock *msk, int *size, int remaining,
1212 struct mptcp_addr_info *addr, bool *echo,
1213 bool *drop_ts);
1214 bool mptcp_pm_rm_addr_signal(struct mptcp_sock *msk, unsigned int remaining,
1215 struct mptcp_rm_list *rm_list, int *len);
1216 int mptcp_pm_get_local_id(struct mptcp_sock *msk, struct sock_common *skc);
1217 int mptcp_pm_nl_get_local_id(struct mptcp_sock *msk,
1218 struct mptcp_pm_addr_entry *skc);
1219 int mptcp_userspace_pm_get_local_id(struct mptcp_sock *msk,
1220 struct mptcp_pm_addr_entry *skc);
1221 bool mptcp_pm_is_backup(struct mptcp_sock *msk, struct sock_common *skc);
1222 bool mptcp_pm_nl_is_backup(struct mptcp_sock *msk, struct mptcp_addr_info *skc);
1223 bool mptcp_userspace_pm_is_backup(struct mptcp_sock *msk, struct mptcp_addr_info *skc);
1224 int mptcp_pm_nl_dump_addr(struct sk_buff *msg,
1225 struct netlink_callback *cb);
1226 int mptcp_userspace_pm_dump_addr(struct sk_buff *msg,
1227 struct netlink_callback *cb);
1228 int mptcp_pm_nl_get_addr(u8 id, struct mptcp_pm_addr_entry *addr,
1229 struct genl_info *info);
1230 int mptcp_userspace_pm_get_addr(u8 id, struct mptcp_pm_addr_entry *addr,
1231 struct genl_info *info);
1232
subflow_get_local_id(const struct mptcp_subflow_context * subflow)1233 static inline u8 subflow_get_local_id(const struct mptcp_subflow_context *subflow)
1234 {
1235 int local_id = READ_ONCE(subflow->local_id);
1236
1237 if (local_id < 0)
1238 return 0;
1239 return local_id;
1240 }
1241
1242 void __init mptcp_pm_kernel_register(void);
1243 void __init mptcp_pm_userspace_register(void);
1244 void __init mptcp_pm_nl_init(void);
1245 void mptcp_pm_worker(struct mptcp_sock *msk);
1246 void __mptcp_pm_kernel_worker(struct mptcp_sock *msk);
1247 u8 mptcp_pm_get_endp_signal_max(const struct mptcp_sock *msk);
1248 u8 mptcp_pm_get_endp_subflow_max(const struct mptcp_sock *msk);
1249 u8 mptcp_pm_get_endp_laminar_max(const struct mptcp_sock *msk);
1250 u8 mptcp_pm_get_endp_fullmesh_max(const struct mptcp_sock *msk);
1251 u8 mptcp_pm_get_limit_add_addr_accepted(const struct mptcp_sock *msk);
1252 u8 mptcp_pm_get_limit_extra_subflows(const struct mptcp_sock *msk);
1253
1254 /* called under PM lock */
__mptcp_pm_close_subflow(struct mptcp_sock * msk)1255 static inline void __mptcp_pm_close_subflow(struct mptcp_sock *msk)
1256 {
1257 if (!WARN_ON_ONCE(msk->pm.extra_subflows == 0) &&
1258 --msk->pm.extra_subflows < mptcp_pm_get_limit_extra_subflows(msk))
1259 WRITE_ONCE(msk->pm.accept_subflow, true);
1260 }
1261
mptcp_pm_close_subflow(struct mptcp_sock * msk)1262 static inline void mptcp_pm_close_subflow(struct mptcp_sock *msk)
1263 {
1264 spin_lock_bh(&msk->pm.lock);
1265 __mptcp_pm_close_subflow(msk);
1266 spin_unlock_bh(&msk->pm.lock);
1267 }
1268
mptcp_pm_add_addr_c_flag_case(struct mptcp_sock * msk)1269 static inline bool mptcp_pm_add_addr_c_flag_case(struct mptcp_sock *msk)
1270 {
1271 return READ_ONCE(msk->pm.remote_deny_join_id0) &&
1272 msk->pm.local_addr_used == 0 &&
1273 mptcp_pm_get_limit_add_addr_accepted(msk) == 0 &&
1274 msk->pm.extra_subflows < mptcp_pm_get_limit_extra_subflows(msk);
1275 }
1276
1277 void mptcp_sockopt_sync_locked(struct mptcp_sock *msk, struct sock *ssk);
1278
mptcp_get_ext(const struct sk_buff * skb)1279 static inline struct mptcp_ext *mptcp_get_ext(const struct sk_buff *skb)
1280 {
1281 return (struct mptcp_ext *)skb_ext_find(skb, SKB_EXT_MPTCP);
1282 }
1283
1284 void mptcp_diag_subflow_init(struct tcp_ulp_ops *ops);
1285
__mptcp_check_fallback(const struct mptcp_sock * msk)1286 static inline bool __mptcp_check_fallback(const struct mptcp_sock *msk)
1287 {
1288 return test_bit(MPTCP_FALLBACK_DONE, &msk->flags);
1289 }
1290
mptcp_check_fallback(const struct sock * sk)1291 static inline bool mptcp_check_fallback(const struct sock *sk)
1292 {
1293 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
1294 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
1295
1296 return __mptcp_check_fallback(msk);
1297 }
1298
__mptcp_has_initial_subflow(const struct mptcp_sock * msk)1299 static inline bool __mptcp_has_initial_subflow(const struct mptcp_sock *msk)
1300 {
1301 struct sock *ssk = READ_ONCE(msk->first);
1302
1303 return ssk && ((1 << inet_sk_state_load(ssk)) &
1304 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1305 TCPF_SYN_RECV | TCPF_LISTEN));
1306 }
1307
1308 bool __mptcp_try_fallback(struct mptcp_sock *msk, int fb_mib);
1309
mptcp_try_fallback(struct sock * ssk,int fb_mib)1310 static inline bool mptcp_try_fallback(struct sock *ssk, int fb_mib)
1311 {
1312 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1313 struct sock *sk = subflow->conn;
1314 struct mptcp_sock *msk;
1315
1316 msk = mptcp_sk(sk);
1317 if (!__mptcp_try_fallback(msk, fb_mib))
1318 return false;
1319 if (READ_ONCE(msk->snd_data_fin_enable) && !(ssk->sk_shutdown & SEND_SHUTDOWN)) {
1320 gfp_t saved_allocation = ssk->sk_allocation;
1321
1322 /* we are in a atomic (BH) scope, override ssk default for data
1323 * fin allocation
1324 */
1325 ssk->sk_allocation = GFP_ATOMIC;
1326 ssk->sk_shutdown |= SEND_SHUTDOWN;
1327 tcp_shutdown(ssk, SEND_SHUTDOWN);
1328 ssk->sk_allocation = saved_allocation;
1329 }
1330 return true;
1331 }
1332
mptcp_early_fallback(struct mptcp_sock * msk,struct mptcp_subflow_context * subflow,int fb_mib)1333 static inline void mptcp_early_fallback(struct mptcp_sock *msk,
1334 struct mptcp_subflow_context *subflow,
1335 int fb_mib)
1336 {
1337 subflow->request_mptcp = 0;
1338 WARN_ON_ONCE(!__mptcp_try_fallback(msk, fb_mib));
1339 }
1340
mptcp_check_infinite_map(struct sk_buff * skb)1341 static inline bool mptcp_check_infinite_map(struct sk_buff *skb)
1342 {
1343 struct mptcp_ext *mpext;
1344
1345 mpext = skb ? mptcp_get_ext(skb) : NULL;
1346 if (mpext && mpext->infinite_map)
1347 return true;
1348
1349 return false;
1350 }
1351
is_active_ssk(struct mptcp_subflow_context * subflow)1352 static inline bool is_active_ssk(struct mptcp_subflow_context *subflow)
1353 {
1354 return (subflow->request_mptcp || subflow->request_join);
1355 }
1356
subflow_simultaneous_connect(struct sock * sk)1357 static inline bool subflow_simultaneous_connect(struct sock *sk)
1358 {
1359 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
1360
1361 /* Note that the sk state implies !subflow->conn_finished. */
1362 return sk->sk_state == TCP_SYN_RECV && is_active_ssk(subflow);
1363 }
1364
1365 #ifdef CONFIG_SYN_COOKIES
1366 void subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock *subflow_req,
1367 struct sk_buff *skb);
1368 bool mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock *subflow_req,
1369 struct sk_buff *skb);
1370 void __init mptcp_join_cookie_init(void);
1371 #else
1372 static inline void
subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock * subflow_req,struct sk_buff * skb)1373 subflow_init_req_cookie_join_save(const struct mptcp_subflow_request_sock *subflow_req,
1374 struct sk_buff *skb) {}
1375 static inline bool
mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock * subflow_req,struct sk_buff * skb)1376 mptcp_token_join_cookie_init_state(struct mptcp_subflow_request_sock *subflow_req,
1377 struct sk_buff *skb)
1378 {
1379 return false;
1380 }
1381
mptcp_join_cookie_init(void)1382 static inline void mptcp_join_cookie_init(void) {}
1383 #endif
1384
1385 #endif /* __MPTCP_PROTOCOL_H */
1386