Loading...
1/* SPDX-License-Identifier: GPL-2.0-or-later */
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Definitions for the TCP module.
8 *
9 * Version: @(#)tcp.h 1.0.5 05/23/93
10 *
11 * Authors: Ross Biro
12 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
13 */
14#ifndef _TCP_H
15#define _TCP_H
16
17#define FASTRETRANS_DEBUG 1
18
19#include <linux/list.h>
20#include <linux/tcp.h>
21#include <linux/bug.h>
22#include <linux/slab.h>
23#include <linux/cache.h>
24#include <linux/percpu.h>
25#include <linux/skbuff.h>
26#include <linux/kref.h>
27#include <linux/ktime.h>
28#include <linux/indirect_call_wrapper.h>
29
30#include <net/inet_connection_sock.h>
31#include <net/inet_timewait_sock.h>
32#include <net/inet_hashtables.h>
33#include <net/checksum.h>
34#include <net/request_sock.h>
35#include <net/sock_reuseport.h>
36#include <net/sock.h>
37#include <net/snmp.h>
38#include <net/ip.h>
39#include <net/tcp_states.h>
40#include <net/inet_ecn.h>
41#include <net/dst.h>
42#include <net/mptcp.h>
43
44#include <linux/seq_file.h>
45#include <linux/memcontrol.h>
46#include <linux/bpf-cgroup.h>
47#include <linux/siphash.h>
48
49extern struct inet_hashinfo tcp_hashinfo;
50
51extern struct percpu_counter tcp_orphan_count;
52void tcp_time_wait(struct sock *sk, int state, int timeo);
53
54#define MAX_TCP_HEADER L1_CACHE_ALIGN(128 + MAX_HEADER)
55#define MAX_TCP_OPTION_SPACE 40
56#define TCP_MIN_SND_MSS 48
57#define TCP_MIN_GSO_SIZE (TCP_MIN_SND_MSS - MAX_TCP_OPTION_SPACE)
58
59/*
60 * Never offer a window over 32767 without using window scaling. Some
61 * poor stacks do signed 16bit maths!
62 */
63#define MAX_TCP_WINDOW 32767U
64
65/* Minimal accepted MSS. It is (60+60+8) - (20+20). */
66#define TCP_MIN_MSS 88U
67
68/* The initial MTU to use for probing */
69#define TCP_BASE_MSS 1024
70
71/* probing interval, default to 10 minutes as per RFC4821 */
72#define TCP_PROBE_INTERVAL 600
73
74/* Specify interval when tcp mtu probing will stop */
75#define TCP_PROBE_THRESHOLD 8
76
77/* After receiving this amount of duplicate ACKs fast retransmit starts. */
78#define TCP_FASTRETRANS_THRESH 3
79
80/* Maximal number of ACKs sent quickly to accelerate slow-start. */
81#define TCP_MAX_QUICKACKS 16U
82
83/* Maximal number of window scale according to RFC1323 */
84#define TCP_MAX_WSCALE 14U
85
86/* urg_data states */
87#define TCP_URG_VALID 0x0100
88#define TCP_URG_NOTYET 0x0200
89#define TCP_URG_READ 0x0400
90
91#define TCP_RETR1 3 /*
92 * This is how many retries it does before it
93 * tries to figure out if the gateway is
94 * down. Minimal RFC value is 3; it corresponds
95 * to ~3sec-8min depending on RTO.
96 */
97
98#define TCP_RETR2 15 /*
99 * This should take at least
100 * 90 minutes to time out.
101 * RFC1122 says that the limit is 100 sec.
102 * 15 is ~13-30min depending on RTO.
103 */
104
105#define TCP_SYN_RETRIES 6 /* This is how many retries are done
106 * when active opening a connection.
107 * RFC1122 says the minimum retry MUST
108 * be at least 180secs. Nevertheless
109 * this value is corresponding to
110 * 63secs of retransmission with the
111 * current initial RTO.
112 */
113
114#define TCP_SYNACK_RETRIES 5 /* This is how may retries are done
115 * when passive opening a connection.
116 * This is corresponding to 31secs of
117 * retransmission with the current
118 * initial RTO.
119 */
120
121#define TCP_TIMEWAIT_LEN (60*HZ) /* how long to wait to destroy TIME-WAIT
122 * state, about 60 seconds */
123#define TCP_FIN_TIMEOUT TCP_TIMEWAIT_LEN
124 /* BSD style FIN_WAIT2 deadlock breaker.
125 * It used to be 3min, new value is 60sec,
126 * to combine FIN-WAIT-2 timeout with
127 * TIME-WAIT timer.
128 */
129#define TCP_FIN_TIMEOUT_MAX (120 * HZ) /* max TCP_LINGER2 value (two minutes) */
130
131#define TCP_DELACK_MAX ((unsigned)(HZ/5)) /* maximal time to delay before sending an ACK */
132#if HZ >= 100
133#define TCP_DELACK_MIN ((unsigned)(HZ/25)) /* minimal time to delay before sending an ACK */
134#define TCP_ATO_MIN ((unsigned)(HZ/25))
135#else
136#define TCP_DELACK_MIN 4U
137#define TCP_ATO_MIN 4U
138#endif
139#define TCP_RTO_MAX ((unsigned)(120*HZ))
140#define TCP_RTO_MIN ((unsigned)(HZ/5))
141#define TCP_TIMEOUT_MIN (2U) /* Min timeout for TCP timers in jiffies */
142#define TCP_TIMEOUT_INIT ((unsigned)(1*HZ)) /* RFC6298 2.1 initial RTO value */
143#define TCP_TIMEOUT_FALLBACK ((unsigned)(3*HZ)) /* RFC 1122 initial RTO value, now
144 * used as a fallback RTO for the
145 * initial data transmission if no
146 * valid RTT sample has been acquired,
147 * most likely due to retrans in 3WHS.
148 */
149
150#define TCP_RESOURCE_PROBE_INTERVAL ((unsigned)(HZ/2U)) /* Maximal interval between probes
151 * for local resources.
152 */
153#define TCP_KEEPALIVE_TIME (120*60*HZ) /* two hours */
154#define TCP_KEEPALIVE_PROBES 9 /* Max of 9 keepalive probes */
155#define TCP_KEEPALIVE_INTVL (75*HZ)
156
157#define MAX_TCP_KEEPIDLE 32767
158#define MAX_TCP_KEEPINTVL 32767
159#define MAX_TCP_KEEPCNT 127
160#define MAX_TCP_SYNCNT 127
161
162#define TCP_SYNQ_INTERVAL (HZ/5) /* Period of SYNACK timer */
163
164#define TCP_PAWS_24DAYS (60 * 60 * 24 * 24)
165#define TCP_PAWS_MSL 60 /* Per-host timestamps are invalidated
166 * after this time. It should be equal
167 * (or greater than) TCP_TIMEWAIT_LEN
168 * to provide reliability equal to one
169 * provided by timewait state.
170 */
171#define TCP_PAWS_WINDOW 1 /* Replay window for per-host
172 * timestamps. It must be less than
173 * minimal timewait lifetime.
174 */
175/*
176 * TCP option
177 */
178
179#define TCPOPT_NOP 1 /* Padding */
180#define TCPOPT_EOL 0 /* End of options */
181#define TCPOPT_MSS 2 /* Segment size negotiating */
182#define TCPOPT_WINDOW 3 /* Window scaling */
183#define TCPOPT_SACK_PERM 4 /* SACK Permitted */
184#define TCPOPT_SACK 5 /* SACK Block */
185#define TCPOPT_TIMESTAMP 8 /* Better RTT estimations/PAWS */
186#define TCPOPT_MD5SIG 19 /* MD5 Signature (RFC2385) */
187#define TCPOPT_MPTCP 30 /* Multipath TCP (RFC6824) */
188#define TCPOPT_FASTOPEN 34 /* Fast open (RFC7413) */
189#define TCPOPT_EXP 254 /* Experimental */
190/* Magic number to be after the option value for sharing TCP
191 * experimental options. See draft-ietf-tcpm-experimental-options-00.txt
192 */
193#define TCPOPT_FASTOPEN_MAGIC 0xF989
194#define TCPOPT_SMC_MAGIC 0xE2D4C3D9
195
196/*
197 * TCP option lengths
198 */
199
200#define TCPOLEN_MSS 4
201#define TCPOLEN_WINDOW 3
202#define TCPOLEN_SACK_PERM 2
203#define TCPOLEN_TIMESTAMP 10
204#define TCPOLEN_MD5SIG 18
205#define TCPOLEN_FASTOPEN_BASE 2
206#define TCPOLEN_EXP_FASTOPEN_BASE 4
207#define TCPOLEN_EXP_SMC_BASE 6
208
209/* But this is what stacks really send out. */
210#define TCPOLEN_TSTAMP_ALIGNED 12
211#define TCPOLEN_WSCALE_ALIGNED 4
212#define TCPOLEN_SACKPERM_ALIGNED 4
213#define TCPOLEN_SACK_BASE 2
214#define TCPOLEN_SACK_BASE_ALIGNED 4
215#define TCPOLEN_SACK_PERBLOCK 8
216#define TCPOLEN_MD5SIG_ALIGNED 20
217#define TCPOLEN_MSS_ALIGNED 4
218#define TCPOLEN_EXP_SMC_BASE_ALIGNED 8
219
220/* Flags in tp->nonagle */
221#define TCP_NAGLE_OFF 1 /* Nagle's algo is disabled */
222#define TCP_NAGLE_CORK 2 /* Socket is corked */
223#define TCP_NAGLE_PUSH 4 /* Cork is overridden for already queued data */
224
225/* TCP thin-stream limits */
226#define TCP_THIN_LINEAR_RETRIES 6 /* After 6 linear retries, do exp. backoff */
227
228/* TCP initial congestion window as per rfc6928 */
229#define TCP_INIT_CWND 10
230
231/* Bit Flags for sysctl_tcp_fastopen */
232#define TFO_CLIENT_ENABLE 1
233#define TFO_SERVER_ENABLE 2
234#define TFO_CLIENT_NO_COOKIE 4 /* Data in SYN w/o cookie option */
235
236/* Accept SYN data w/o any cookie option */
237#define TFO_SERVER_COOKIE_NOT_REQD 0x200
238
239/* Force enable TFO on all listeners, i.e., not requiring the
240 * TCP_FASTOPEN socket option.
241 */
242#define TFO_SERVER_WO_SOCKOPT1 0x400
243
244
245/* sysctl variables for tcp */
246extern int sysctl_tcp_max_orphans;
247extern long sysctl_tcp_mem[3];
248
249#define TCP_RACK_LOSS_DETECTION 0x1 /* Use RACK to detect losses */
250#define TCP_RACK_STATIC_REO_WND 0x2 /* Use static RACK reo wnd */
251#define TCP_RACK_NO_DUPTHRESH 0x4 /* Do not use DUPACK threshold in RACK */
252
253extern atomic_long_t tcp_memory_allocated;
254extern struct percpu_counter tcp_sockets_allocated;
255extern unsigned long tcp_memory_pressure;
256
257/* optimized version of sk_under_memory_pressure() for TCP sockets */
258static inline bool tcp_under_memory_pressure(const struct sock *sk)
259{
260 if (mem_cgroup_sockets_enabled && sk->sk_memcg &&
261 mem_cgroup_under_socket_pressure(sk->sk_memcg))
262 return true;
263
264 return READ_ONCE(tcp_memory_pressure);
265}
266/*
267 * The next routines deal with comparing 32 bit unsigned ints
268 * and worry about wraparound (automatic with unsigned arithmetic).
269 */
270
271static inline bool before(__u32 seq1, __u32 seq2)
272{
273 return (__s32)(seq1-seq2) < 0;
274}
275#define after(seq2, seq1) before(seq1, seq2)
276
277/* is s2<=s1<=s3 ? */
278static inline bool between(__u32 seq1, __u32 seq2, __u32 seq3)
279{
280 return seq3 - seq2 >= seq1 - seq2;
281}
282
283static inline bool tcp_out_of_memory(struct sock *sk)
284{
285 if (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
286 sk_memory_allocated(sk) > sk_prot_mem_limits(sk, 2))
287 return true;
288 return false;
289}
290
291void sk_forced_mem_schedule(struct sock *sk, int size);
292
293static inline bool tcp_too_many_orphans(struct sock *sk, int shift)
294{
295 struct percpu_counter *ocp = sk->sk_prot->orphan_count;
296 int orphans = percpu_counter_read_positive(ocp);
297
298 if (orphans << shift > sysctl_tcp_max_orphans) {
299 orphans = percpu_counter_sum_positive(ocp);
300 if (orphans << shift > sysctl_tcp_max_orphans)
301 return true;
302 }
303 return false;
304}
305
306bool tcp_check_oom(struct sock *sk, int shift);
307
308
309extern struct proto tcp_prot;
310
311#define TCP_INC_STATS(net, field) SNMP_INC_STATS((net)->mib.tcp_statistics, field)
312#define __TCP_INC_STATS(net, field) __SNMP_INC_STATS((net)->mib.tcp_statistics, field)
313#define TCP_DEC_STATS(net, field) SNMP_DEC_STATS((net)->mib.tcp_statistics, field)
314#define TCP_ADD_STATS(net, field, val) SNMP_ADD_STATS((net)->mib.tcp_statistics, field, val)
315
316void tcp_tasklet_init(void);
317
318int tcp_v4_err(struct sk_buff *skb, u32);
319
320void tcp_shutdown(struct sock *sk, int how);
321
322int tcp_v4_early_demux(struct sk_buff *skb);
323int tcp_v4_rcv(struct sk_buff *skb);
324
325int tcp_v4_tw_remember_stamp(struct inet_timewait_sock *tw);
326int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size);
327int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size);
328int tcp_sendpage(struct sock *sk, struct page *page, int offset, size_t size,
329 int flags);
330int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
331 size_t size, int flags);
332ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
333 size_t size, int flags);
334int tcp_send_mss(struct sock *sk, int *size_goal, int flags);
335void tcp_push(struct sock *sk, int flags, int mss_now, int nonagle,
336 int size_goal);
337void tcp_release_cb(struct sock *sk);
338void tcp_wfree(struct sk_buff *skb);
339void tcp_write_timer_handler(struct sock *sk);
340void tcp_delack_timer_handler(struct sock *sk);
341int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg);
342int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb);
343void tcp_rcv_established(struct sock *sk, struct sk_buff *skb);
344void tcp_rcv_space_adjust(struct sock *sk);
345int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp);
346void tcp_twsk_destructor(struct sock *sk);
347ssize_t tcp_splice_read(struct socket *sk, loff_t *ppos,
348 struct pipe_inode_info *pipe, size_t len,
349 unsigned int flags);
350
351void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks);
352static inline void tcp_dec_quickack_mode(struct sock *sk,
353 const unsigned int pkts)
354{
355 struct inet_connection_sock *icsk = inet_csk(sk);
356
357 if (icsk->icsk_ack.quick) {
358 if (pkts >= icsk->icsk_ack.quick) {
359 icsk->icsk_ack.quick = 0;
360 /* Leaving quickack mode we deflate ATO. */
361 icsk->icsk_ack.ato = TCP_ATO_MIN;
362 } else
363 icsk->icsk_ack.quick -= pkts;
364 }
365}
366
367#define TCP_ECN_OK 1
368#define TCP_ECN_QUEUE_CWR 2
369#define TCP_ECN_DEMAND_CWR 4
370#define TCP_ECN_SEEN 8
371
372enum tcp_tw_status {
373 TCP_TW_SUCCESS = 0,
374 TCP_TW_RST = 1,
375 TCP_TW_ACK = 2,
376 TCP_TW_SYN = 3
377};
378
379
380enum tcp_tw_status tcp_timewait_state_process(struct inet_timewait_sock *tw,
381 struct sk_buff *skb,
382 const struct tcphdr *th);
383struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
384 struct request_sock *req, bool fastopen,
385 bool *lost_race);
386int tcp_child_process(struct sock *parent, struct sock *child,
387 struct sk_buff *skb);
388void tcp_enter_loss(struct sock *sk);
389void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag);
390void tcp_clear_retrans(struct tcp_sock *tp);
391void tcp_update_metrics(struct sock *sk);
392void tcp_init_metrics(struct sock *sk);
393void tcp_metrics_init(void);
394bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst);
395void tcp_close(struct sock *sk, long timeout);
396void tcp_init_sock(struct sock *sk);
397void tcp_init_transfer(struct sock *sk, int bpf_op);
398__poll_t tcp_poll(struct file *file, struct socket *sock,
399 struct poll_table_struct *wait);
400int tcp_getsockopt(struct sock *sk, int level, int optname,
401 char __user *optval, int __user *optlen);
402int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
403 unsigned int optlen);
404void tcp_set_keepalive(struct sock *sk, int val);
405void tcp_syn_ack_timeout(const struct request_sock *req);
406int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
407 int flags, int *addr_len);
408int tcp_set_rcvlowat(struct sock *sk, int val);
409void tcp_data_ready(struct sock *sk);
410#ifdef CONFIG_MMU
411int tcp_mmap(struct file *file, struct socket *sock,
412 struct vm_area_struct *vma);
413#endif
414void tcp_parse_options(const struct net *net, const struct sk_buff *skb,
415 struct tcp_options_received *opt_rx,
416 int estab, struct tcp_fastopen_cookie *foc);
417const u8 *tcp_parse_md5sig_option(const struct tcphdr *th);
418
419/*
420 * BPF SKB-less helpers
421 */
422u16 tcp_v4_get_syncookie(struct sock *sk, struct iphdr *iph,
423 struct tcphdr *th, u32 *cookie);
424u16 tcp_v6_get_syncookie(struct sock *sk, struct ipv6hdr *iph,
425 struct tcphdr *th, u32 *cookie);
426u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
427 const struct tcp_request_sock_ops *af_ops,
428 struct sock *sk, struct tcphdr *th);
429/*
430 * TCP v4 functions exported for the inet6 API
431 */
432
433void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb);
434void tcp_v4_mtu_reduced(struct sock *sk);
435void tcp_req_err(struct sock *sk, u32 seq, bool abort);
436void tcp_ld_RTO_revert(struct sock *sk, u32 seq);
437int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb);
438struct sock *tcp_create_openreq_child(const struct sock *sk,
439 struct request_sock *req,
440 struct sk_buff *skb);
441void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst);
442struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
443 struct request_sock *req,
444 struct dst_entry *dst,
445 struct request_sock *req_unhash,
446 bool *own_req);
447int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb);
448int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len);
449int tcp_connect(struct sock *sk);
450enum tcp_synack_type {
451 TCP_SYNACK_NORMAL,
452 TCP_SYNACK_FASTOPEN,
453 TCP_SYNACK_COOKIE,
454};
455struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
456 struct request_sock *req,
457 struct tcp_fastopen_cookie *foc,
458 enum tcp_synack_type synack_type);
459int tcp_disconnect(struct sock *sk, int flags);
460
461void tcp_finish_connect(struct sock *sk, struct sk_buff *skb);
462int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size);
463void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb);
464
465/* From syncookies.c */
466struct sock *tcp_get_cookie_sock(struct sock *sk, struct sk_buff *skb,
467 struct request_sock *req,
468 struct dst_entry *dst, u32 tsoff);
469int __cookie_v4_check(const struct iphdr *iph, const struct tcphdr *th,
470 u32 cookie);
471struct sock *cookie_v4_check(struct sock *sk, struct sk_buff *skb);
472struct request_sock *cookie_tcp_reqsk_alloc(const struct request_sock_ops *ops,
473 struct sock *sk, struct sk_buff *skb);
474#ifdef CONFIG_SYN_COOKIES
475
476/* Syncookies use a monotonic timer which increments every 60 seconds.
477 * This counter is used both as a hash input and partially encoded into
478 * the cookie value. A cookie is only validated further if the delta
479 * between the current counter value and the encoded one is less than this,
480 * i.e. a sent cookie is valid only at most for 2*60 seconds (or less if
481 * the counter advances immediately after a cookie is generated).
482 */
483#define MAX_SYNCOOKIE_AGE 2
484#define TCP_SYNCOOKIE_PERIOD (60 * HZ)
485#define TCP_SYNCOOKIE_VALID (MAX_SYNCOOKIE_AGE * TCP_SYNCOOKIE_PERIOD)
486
487/* syncookies: remember time of last synqueue overflow
488 * But do not dirty this field too often (once per second is enough)
489 * It is racy as we do not hold a lock, but race is very minor.
490 */
491static inline void tcp_synq_overflow(const struct sock *sk)
492{
493 unsigned int last_overflow;
494 unsigned int now = jiffies;
495
496 if (sk->sk_reuseport) {
497 struct sock_reuseport *reuse;
498
499 reuse = rcu_dereference(sk->sk_reuseport_cb);
500 if (likely(reuse)) {
501 last_overflow = READ_ONCE(reuse->synq_overflow_ts);
502 if (!time_between32(now, last_overflow,
503 last_overflow + HZ))
504 WRITE_ONCE(reuse->synq_overflow_ts, now);
505 return;
506 }
507 }
508
509 last_overflow = READ_ONCE(tcp_sk(sk)->rx_opt.ts_recent_stamp);
510 if (!time_between32(now, last_overflow, last_overflow + HZ))
511 WRITE_ONCE(tcp_sk(sk)->rx_opt.ts_recent_stamp, now);
512}
513
514/* syncookies: no recent synqueue overflow on this listening socket? */
515static inline bool tcp_synq_no_recent_overflow(const struct sock *sk)
516{
517 unsigned int last_overflow;
518 unsigned int now = jiffies;
519
520 if (sk->sk_reuseport) {
521 struct sock_reuseport *reuse;
522
523 reuse = rcu_dereference(sk->sk_reuseport_cb);
524 if (likely(reuse)) {
525 last_overflow = READ_ONCE(reuse->synq_overflow_ts);
526 return !time_between32(now, last_overflow - HZ,
527 last_overflow +
528 TCP_SYNCOOKIE_VALID);
529 }
530 }
531
532 last_overflow = READ_ONCE(tcp_sk(sk)->rx_opt.ts_recent_stamp);
533
534 /* If last_overflow <= jiffies <= last_overflow + TCP_SYNCOOKIE_VALID,
535 * then we're under synflood. However, we have to use
536 * 'last_overflow - HZ' as lower bound. That's because a concurrent
537 * tcp_synq_overflow() could update .ts_recent_stamp after we read
538 * jiffies but before we store .ts_recent_stamp into last_overflow,
539 * which could lead to rejecting a valid syncookie.
540 */
541 return !time_between32(now, last_overflow - HZ,
542 last_overflow + TCP_SYNCOOKIE_VALID);
543}
544
545static inline u32 tcp_cookie_time(void)
546{
547 u64 val = get_jiffies_64();
548
549 do_div(val, TCP_SYNCOOKIE_PERIOD);
550 return val;
551}
552
553u32 __cookie_v4_init_sequence(const struct iphdr *iph, const struct tcphdr *th,
554 u16 *mssp);
555__u32 cookie_v4_init_sequence(const struct sk_buff *skb, __u16 *mss);
556u64 cookie_init_timestamp(struct request_sock *req, u64 now);
557bool cookie_timestamp_decode(const struct net *net,
558 struct tcp_options_received *opt);
559bool cookie_ecn_ok(const struct tcp_options_received *opt,
560 const struct net *net, const struct dst_entry *dst);
561
562/* From net/ipv6/syncookies.c */
563int __cookie_v6_check(const struct ipv6hdr *iph, const struct tcphdr *th,
564 u32 cookie);
565struct sock *cookie_v6_check(struct sock *sk, struct sk_buff *skb);
566
567u32 __cookie_v6_init_sequence(const struct ipv6hdr *iph,
568 const struct tcphdr *th, u16 *mssp);
569__u32 cookie_v6_init_sequence(const struct sk_buff *skb, __u16 *mss);
570#endif
571/* tcp_output.c */
572
573void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
574 int nonagle);
575int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs);
576int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs);
577void tcp_retransmit_timer(struct sock *sk);
578void tcp_xmit_retransmit_queue(struct sock *);
579void tcp_simple_retransmit(struct sock *);
580void tcp_enter_recovery(struct sock *sk, bool ece_ack);
581int tcp_trim_head(struct sock *, struct sk_buff *, u32);
582enum tcp_queue {
583 TCP_FRAG_IN_WRITE_QUEUE,
584 TCP_FRAG_IN_RTX_QUEUE,
585};
586int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
587 struct sk_buff *skb, u32 len,
588 unsigned int mss_now, gfp_t gfp);
589
590void tcp_send_probe0(struct sock *);
591void tcp_send_partial(struct sock *);
592int tcp_write_wakeup(struct sock *, int mib);
593void tcp_send_fin(struct sock *sk);
594void tcp_send_active_reset(struct sock *sk, gfp_t priority);
595int tcp_send_synack(struct sock *);
596void tcp_push_one(struct sock *, unsigned int mss_now);
597void __tcp_send_ack(struct sock *sk, u32 rcv_nxt);
598void tcp_send_ack(struct sock *sk);
599void tcp_send_delayed_ack(struct sock *sk);
600void tcp_send_loss_probe(struct sock *sk);
601bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto);
602void tcp_skb_collapse_tstamp(struct sk_buff *skb,
603 const struct sk_buff *next_skb);
604
605/* tcp_input.c */
606void tcp_rearm_rto(struct sock *sk);
607void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req);
608void tcp_reset(struct sock *sk);
609void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb);
610void tcp_fin(struct sock *sk);
611
612/* tcp_timer.c */
613void tcp_init_xmit_timers(struct sock *);
614static inline void tcp_clear_xmit_timers(struct sock *sk)
615{
616 if (hrtimer_try_to_cancel(&tcp_sk(sk)->pacing_timer) == 1)
617 __sock_put(sk);
618
619 if (hrtimer_try_to_cancel(&tcp_sk(sk)->compressed_ack_timer) == 1)
620 __sock_put(sk);
621
622 inet_csk_clear_xmit_timers(sk);
623}
624
625unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu);
626unsigned int tcp_current_mss(struct sock *sk);
627
628/* Bound MSS / TSO packet size with the half of the window */
629static inline int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
630{
631 int cutoff;
632
633 /* When peer uses tiny windows, there is no use in packetizing
634 * to sub-MSS pieces for the sake of SWS or making sure there
635 * are enough packets in the pipe for fast recovery.
636 *
637 * On the other hand, for extremely large MSS devices, handling
638 * smaller than MSS windows in this way does make sense.
639 */
640 if (tp->max_window > TCP_MSS_DEFAULT)
641 cutoff = (tp->max_window >> 1);
642 else
643 cutoff = tp->max_window;
644
645 if (cutoff && pktsize > cutoff)
646 return max_t(int, cutoff, 68U - tp->tcp_header_len);
647 else
648 return pktsize;
649}
650
651/* tcp.c */
652void tcp_get_info(struct sock *, struct tcp_info *);
653
654/* Read 'sendfile()'-style from a TCP socket */
655int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
656 sk_read_actor_t recv_actor);
657
658void tcp_initialize_rcv_mss(struct sock *sk);
659
660int tcp_mtu_to_mss(struct sock *sk, int pmtu);
661int tcp_mss_to_mtu(struct sock *sk, int mss);
662void tcp_mtup_init(struct sock *sk);
663
664static inline void tcp_bound_rto(const struct sock *sk)
665{
666 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
667 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
668}
669
670static inline u32 __tcp_set_rto(const struct tcp_sock *tp)
671{
672 return usecs_to_jiffies((tp->srtt_us >> 3) + tp->rttvar_us);
673}
674
675static inline void __tcp_fast_path_on(struct tcp_sock *tp, u32 snd_wnd)
676{
677 tp->pred_flags = htonl((tp->tcp_header_len << 26) |
678 ntohl(TCP_FLAG_ACK) |
679 snd_wnd);
680}
681
682static inline void tcp_fast_path_on(struct tcp_sock *tp)
683{
684 __tcp_fast_path_on(tp, tp->snd_wnd >> tp->rx_opt.snd_wscale);
685}
686
687static inline void tcp_fast_path_check(struct sock *sk)
688{
689 struct tcp_sock *tp = tcp_sk(sk);
690
691 if (RB_EMPTY_ROOT(&tp->out_of_order_queue) &&
692 tp->rcv_wnd &&
693 atomic_read(&sk->sk_rmem_alloc) < sk->sk_rcvbuf &&
694 !tp->urg_data)
695 tcp_fast_path_on(tp);
696}
697
698/* Compute the actual rto_min value */
699static inline u32 tcp_rto_min(struct sock *sk)
700{
701 const struct dst_entry *dst = __sk_dst_get(sk);
702 u32 rto_min = TCP_RTO_MIN;
703
704 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
705 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
706 return rto_min;
707}
708
709static inline u32 tcp_rto_min_us(struct sock *sk)
710{
711 return jiffies_to_usecs(tcp_rto_min(sk));
712}
713
714static inline bool tcp_ca_dst_locked(const struct dst_entry *dst)
715{
716 return dst_metric_locked(dst, RTAX_CC_ALGO);
717}
718
719/* Minimum RTT in usec. ~0 means not available. */
720static inline u32 tcp_min_rtt(const struct tcp_sock *tp)
721{
722 return minmax_get(&tp->rtt_min);
723}
724
725/* Compute the actual receive window we are currently advertising.
726 * Rcv_nxt can be after the window if our peer push more data
727 * than the offered window.
728 */
729static inline u32 tcp_receive_window(const struct tcp_sock *tp)
730{
731 s32 win = tp->rcv_wup + tp->rcv_wnd - tp->rcv_nxt;
732
733 if (win < 0)
734 win = 0;
735 return (u32) win;
736}
737
738/* Choose a new window, without checks for shrinking, and without
739 * scaling applied to the result. The caller does these things
740 * if necessary. This is a "raw" window selection.
741 */
742u32 __tcp_select_window(struct sock *sk);
743
744void tcp_send_window_probe(struct sock *sk);
745
746/* TCP uses 32bit jiffies to save some space.
747 * Note that this is different from tcp_time_stamp, which
748 * historically has been the same until linux-4.13.
749 */
750#define tcp_jiffies32 ((u32)jiffies)
751
752/*
753 * Deliver a 32bit value for TCP timestamp option (RFC 7323)
754 * It is no longer tied to jiffies, but to 1 ms clock.
755 * Note: double check if you want to use tcp_jiffies32 instead of this.
756 */
757#define TCP_TS_HZ 1000
758
759static inline u64 tcp_clock_ns(void)
760{
761 return ktime_get_ns();
762}
763
764static inline u64 tcp_clock_us(void)
765{
766 return div_u64(tcp_clock_ns(), NSEC_PER_USEC);
767}
768
769/* This should only be used in contexts where tp->tcp_mstamp is up to date */
770static inline u32 tcp_time_stamp(const struct tcp_sock *tp)
771{
772 return div_u64(tp->tcp_mstamp, USEC_PER_SEC / TCP_TS_HZ);
773}
774
775/* Convert a nsec timestamp into TCP TSval timestamp (ms based currently) */
776static inline u32 tcp_ns_to_ts(u64 ns)
777{
778 return div_u64(ns, NSEC_PER_SEC / TCP_TS_HZ);
779}
780
781/* Could use tcp_clock_us() / 1000, but this version uses a single divide */
782static inline u32 tcp_time_stamp_raw(void)
783{
784 return tcp_ns_to_ts(tcp_clock_ns());
785}
786
787void tcp_mstamp_refresh(struct tcp_sock *tp);
788
789static inline u32 tcp_stamp_us_delta(u64 t1, u64 t0)
790{
791 return max_t(s64, t1 - t0, 0);
792}
793
794static inline u32 tcp_skb_timestamp(const struct sk_buff *skb)
795{
796 return tcp_ns_to_ts(skb->skb_mstamp_ns);
797}
798
799/* provide the departure time in us unit */
800static inline u64 tcp_skb_timestamp_us(const struct sk_buff *skb)
801{
802 return div_u64(skb->skb_mstamp_ns, NSEC_PER_USEC);
803}
804
805
806#define tcp_flag_byte(th) (((u_int8_t *)th)[13])
807
808#define TCPHDR_FIN 0x01
809#define TCPHDR_SYN 0x02
810#define TCPHDR_RST 0x04
811#define TCPHDR_PSH 0x08
812#define TCPHDR_ACK 0x10
813#define TCPHDR_URG 0x20
814#define TCPHDR_ECE 0x40
815#define TCPHDR_CWR 0x80
816
817#define TCPHDR_SYN_ECN (TCPHDR_SYN | TCPHDR_ECE | TCPHDR_CWR)
818
819/* This is what the send packet queuing engine uses to pass
820 * TCP per-packet control information to the transmission code.
821 * We also store the host-order sequence numbers in here too.
822 * This is 44 bytes if IPV6 is enabled.
823 * If this grows please adjust skbuff.h:skbuff->cb[xxx] size appropriately.
824 */
825struct tcp_skb_cb {
826 __u32 seq; /* Starting sequence number */
827 __u32 end_seq; /* SEQ + FIN + SYN + datalen */
828 union {
829 /* Note : tcp_tw_isn is used in input path only
830 * (isn chosen by tcp_timewait_state_process())
831 *
832 * tcp_gso_segs/size are used in write queue only,
833 * cf tcp_skb_pcount()/tcp_skb_mss()
834 */
835 __u32 tcp_tw_isn;
836 struct {
837 u16 tcp_gso_segs;
838 u16 tcp_gso_size;
839 };
840 };
841 __u8 tcp_flags; /* TCP header flags. (tcp[13]) */
842
843 __u8 sacked; /* State flags for SACK. */
844#define TCPCB_SACKED_ACKED 0x01 /* SKB ACK'd by a SACK block */
845#define TCPCB_SACKED_RETRANS 0x02 /* SKB retransmitted */
846#define TCPCB_LOST 0x04 /* SKB is lost */
847#define TCPCB_TAGBITS 0x07 /* All tag bits */
848#define TCPCB_REPAIRED 0x10 /* SKB repaired (no skb_mstamp_ns) */
849#define TCPCB_EVER_RETRANS 0x80 /* Ever retransmitted frame */
850#define TCPCB_RETRANS (TCPCB_SACKED_RETRANS|TCPCB_EVER_RETRANS| \
851 TCPCB_REPAIRED)
852
853 __u8 ip_dsfield; /* IPv4 tos or IPv6 dsfield */
854 __u8 txstamp_ack:1, /* Record TX timestamp for ack? */
855 eor:1, /* Is skb MSG_EOR marked? */
856 has_rxtstamp:1, /* SKB has a RX timestamp */
857 unused:5;
858 __u32 ack_seq; /* Sequence number ACK'd */
859 union {
860 struct {
861 /* There is space for up to 24 bytes */
862 __u32 in_flight:30,/* Bytes in flight at transmit */
863 is_app_limited:1, /* cwnd not fully used? */
864 unused:1;
865 /* pkts S/ACKed so far upon tx of skb, incl retrans: */
866 __u32 delivered;
867 /* start of send pipeline phase */
868 u64 first_tx_mstamp;
869 /* when we reached the "delivered" count */
870 u64 delivered_mstamp;
871 } tx; /* only used for outgoing skbs */
872 union {
873 struct inet_skb_parm h4;
874#if IS_ENABLED(CONFIG_IPV6)
875 struct inet6_skb_parm h6;
876#endif
877 } header; /* For incoming skbs */
878 struct {
879 __u32 flags;
880 struct sock *sk_redir;
881 void *data_end;
882 } bpf;
883 };
884};
885
886#define TCP_SKB_CB(__skb) ((struct tcp_skb_cb *)&((__skb)->cb[0]))
887
888static inline void bpf_compute_data_end_sk_skb(struct sk_buff *skb)
889{
890 TCP_SKB_CB(skb)->bpf.data_end = skb->data + skb_headlen(skb);
891}
892
893static inline bool tcp_skb_bpf_ingress(const struct sk_buff *skb)
894{
895 return TCP_SKB_CB(skb)->bpf.flags & BPF_F_INGRESS;
896}
897
898static inline struct sock *tcp_skb_bpf_redirect_fetch(struct sk_buff *skb)
899{
900 return TCP_SKB_CB(skb)->bpf.sk_redir;
901}
902
903static inline void tcp_skb_bpf_redirect_clear(struct sk_buff *skb)
904{
905 TCP_SKB_CB(skb)->bpf.sk_redir = NULL;
906}
907
908extern const struct inet_connection_sock_af_ops ipv4_specific;
909
910#if IS_ENABLED(CONFIG_IPV6)
911/* This is the variant of inet6_iif() that must be used by TCP,
912 * as TCP moves IP6CB into a different location in skb->cb[]
913 */
914static inline int tcp_v6_iif(const struct sk_buff *skb)
915{
916 return TCP_SKB_CB(skb)->header.h6.iif;
917}
918
919static inline int tcp_v6_iif_l3_slave(const struct sk_buff *skb)
920{
921 bool l3_slave = ipv6_l3mdev_skb(TCP_SKB_CB(skb)->header.h6.flags);
922
923 return l3_slave ? skb->skb_iif : TCP_SKB_CB(skb)->header.h6.iif;
924}
925
926/* TCP_SKB_CB reference means this can not be used from early demux */
927static inline int tcp_v6_sdif(const struct sk_buff *skb)
928{
929#if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV)
930 if (skb && ipv6_l3mdev_skb(TCP_SKB_CB(skb)->header.h6.flags))
931 return TCP_SKB_CB(skb)->header.h6.iif;
932#endif
933 return 0;
934}
935
936extern const struct inet_connection_sock_af_ops ipv6_specific;
937
938INDIRECT_CALLABLE_DECLARE(void tcp_v6_send_check(struct sock *sk, struct sk_buff *skb));
939INDIRECT_CALLABLE_DECLARE(int tcp_v6_rcv(struct sk_buff *skb));
940INDIRECT_CALLABLE_DECLARE(void tcp_v6_early_demux(struct sk_buff *skb));
941
942#endif
943
944static inline bool inet_exact_dif_match(struct net *net, struct sk_buff *skb)
945{
946#if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV)
947 if (!net->ipv4.sysctl_tcp_l3mdev_accept &&
948 skb && ipv4_l3mdev_skb(IPCB(skb)->flags))
949 return true;
950#endif
951 return false;
952}
953
954/* TCP_SKB_CB reference means this can not be used from early demux */
955static inline int tcp_v4_sdif(struct sk_buff *skb)
956{
957#if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV)
958 if (skb && ipv4_l3mdev_skb(TCP_SKB_CB(skb)->header.h4.flags))
959 return TCP_SKB_CB(skb)->header.h4.iif;
960#endif
961 return 0;
962}
963
964/* Due to TSO, an SKB can be composed of multiple actual
965 * packets. To keep these tracked properly, we use this.
966 */
967static inline int tcp_skb_pcount(const struct sk_buff *skb)
968{
969 return TCP_SKB_CB(skb)->tcp_gso_segs;
970}
971
972static inline void tcp_skb_pcount_set(struct sk_buff *skb, int segs)
973{
974 TCP_SKB_CB(skb)->tcp_gso_segs = segs;
975}
976
977static inline void tcp_skb_pcount_add(struct sk_buff *skb, int segs)
978{
979 TCP_SKB_CB(skb)->tcp_gso_segs += segs;
980}
981
982/* This is valid iff skb is in write queue and tcp_skb_pcount() > 1. */
983static inline int tcp_skb_mss(const struct sk_buff *skb)
984{
985 return TCP_SKB_CB(skb)->tcp_gso_size;
986}
987
988static inline bool tcp_skb_can_collapse_to(const struct sk_buff *skb)
989{
990 return likely(!TCP_SKB_CB(skb)->eor);
991}
992
993static inline bool tcp_skb_can_collapse(const struct sk_buff *to,
994 const struct sk_buff *from)
995{
996 return likely(tcp_skb_can_collapse_to(to) &&
997 mptcp_skb_can_collapse(to, from));
998}
999
1000/* Events passed to congestion control interface */
1001enum tcp_ca_event {
1002 CA_EVENT_TX_START, /* first transmit when no packets in flight */
1003 CA_EVENT_CWND_RESTART, /* congestion window restart */
1004 CA_EVENT_COMPLETE_CWR, /* end of congestion recovery */
1005 CA_EVENT_LOSS, /* loss timeout */
1006 CA_EVENT_ECN_NO_CE, /* ECT set, but not CE marked */
1007 CA_EVENT_ECN_IS_CE, /* received CE marked IP packet */
1008};
1009
1010/* Information about inbound ACK, passed to cong_ops->in_ack_event() */
1011enum tcp_ca_ack_event_flags {
1012 CA_ACK_SLOWPATH = (1 << 0), /* In slow path processing */
1013 CA_ACK_WIN_UPDATE = (1 << 1), /* ACK updated window */
1014 CA_ACK_ECE = (1 << 2), /* ECE bit is set on ack */
1015};
1016
1017/*
1018 * Interface for adding new TCP congestion control handlers
1019 */
1020#define TCP_CA_NAME_MAX 16
1021#define TCP_CA_MAX 128
1022#define TCP_CA_BUF_MAX (TCP_CA_NAME_MAX*TCP_CA_MAX)
1023
1024#define TCP_CA_UNSPEC 0
1025
1026/* Algorithm can be set on socket without CAP_NET_ADMIN privileges */
1027#define TCP_CONG_NON_RESTRICTED 0x1
1028/* Requires ECN/ECT set on all packets */
1029#define TCP_CONG_NEEDS_ECN 0x2
1030#define TCP_CONG_MASK (TCP_CONG_NON_RESTRICTED | TCP_CONG_NEEDS_ECN)
1031
1032union tcp_cc_info;
1033
1034struct ack_sample {
1035 u32 pkts_acked;
1036 s32 rtt_us;
1037 u32 in_flight;
1038};
1039
1040/* A rate sample measures the number of (original/retransmitted) data
1041 * packets delivered "delivered" over an interval of time "interval_us".
1042 * The tcp_rate.c code fills in the rate sample, and congestion
1043 * control modules that define a cong_control function to run at the end
1044 * of ACK processing can optionally chose to consult this sample when
1045 * setting cwnd and pacing rate.
1046 * A sample is invalid if "delivered" or "interval_us" is negative.
1047 */
1048struct rate_sample {
1049 u64 prior_mstamp; /* starting timestamp for interval */
1050 u32 prior_delivered; /* tp->delivered at "prior_mstamp" */
1051 s32 delivered; /* number of packets delivered over interval */
1052 long interval_us; /* time for tp->delivered to incr "delivered" */
1053 u32 snd_interval_us; /* snd interval for delivered packets */
1054 u32 rcv_interval_us; /* rcv interval for delivered packets */
1055 long rtt_us; /* RTT of last (S)ACKed packet (or -1) */
1056 int losses; /* number of packets marked lost upon ACK */
1057 u32 acked_sacked; /* number of packets newly (S)ACKed upon ACK */
1058 u32 prior_in_flight; /* in flight before this ACK */
1059 bool is_app_limited; /* is sample from packet with bubble in pipe? */
1060 bool is_retrans; /* is sample from retransmission? */
1061 bool is_ack_delayed; /* is this (likely) a delayed ACK? */
1062};
1063
1064struct tcp_congestion_ops {
1065 struct list_head list;
1066 u32 key;
1067 u32 flags;
1068
1069 /* initialize private data (optional) */
1070 void (*init)(struct sock *sk);
1071 /* cleanup private data (optional) */
1072 void (*release)(struct sock *sk);
1073
1074 /* return slow start threshold (required) */
1075 u32 (*ssthresh)(struct sock *sk);
1076 /* do new cwnd calculation (required) */
1077 void (*cong_avoid)(struct sock *sk, u32 ack, u32 acked);
1078 /* call before changing ca_state (optional) */
1079 void (*set_state)(struct sock *sk, u8 new_state);
1080 /* call when cwnd event occurs (optional) */
1081 void (*cwnd_event)(struct sock *sk, enum tcp_ca_event ev);
1082 /* call when ack arrives (optional) */
1083 void (*in_ack_event)(struct sock *sk, u32 flags);
1084 /* new value of cwnd after loss (required) */
1085 u32 (*undo_cwnd)(struct sock *sk);
1086 /* hook for packet ack accounting (optional) */
1087 void (*pkts_acked)(struct sock *sk, const struct ack_sample *sample);
1088 /* override sysctl_tcp_min_tso_segs */
1089 u32 (*min_tso_segs)(struct sock *sk);
1090 /* returns the multiplier used in tcp_sndbuf_expand (optional) */
1091 u32 (*sndbuf_expand)(struct sock *sk);
1092 /* call when packets are delivered to update cwnd and pacing rate,
1093 * after all the ca_state processing. (optional)
1094 */
1095 void (*cong_control)(struct sock *sk, const struct rate_sample *rs);
1096 /* get info for inet_diag (optional) */
1097 size_t (*get_info)(struct sock *sk, u32 ext, int *attr,
1098 union tcp_cc_info *info);
1099
1100 char name[TCP_CA_NAME_MAX];
1101 struct module *owner;
1102};
1103
1104int tcp_register_congestion_control(struct tcp_congestion_ops *type);
1105void tcp_unregister_congestion_control(struct tcp_congestion_ops *type);
1106
1107void tcp_assign_congestion_control(struct sock *sk);
1108void tcp_init_congestion_control(struct sock *sk);
1109void tcp_cleanup_congestion_control(struct sock *sk);
1110int tcp_set_default_congestion_control(struct net *net, const char *name);
1111void tcp_get_default_congestion_control(struct net *net, char *name);
1112void tcp_get_available_congestion_control(char *buf, size_t len);
1113void tcp_get_allowed_congestion_control(char *buf, size_t len);
1114int tcp_set_allowed_congestion_control(char *allowed);
1115int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
1116 bool reinit, bool cap_net_admin);
1117u32 tcp_slow_start(struct tcp_sock *tp, u32 acked);
1118void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked);
1119
1120u32 tcp_reno_ssthresh(struct sock *sk);
1121u32 tcp_reno_undo_cwnd(struct sock *sk);
1122void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked);
1123extern struct tcp_congestion_ops tcp_reno;
1124
1125struct tcp_congestion_ops *tcp_ca_find(const char *name);
1126struct tcp_congestion_ops *tcp_ca_find_key(u32 key);
1127u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca);
1128#ifdef CONFIG_INET
1129char *tcp_ca_get_name_by_key(u32 key, char *buffer);
1130#else
1131static inline char *tcp_ca_get_name_by_key(u32 key, char *buffer)
1132{
1133 return NULL;
1134}
1135#endif
1136
1137static inline bool tcp_ca_needs_ecn(const struct sock *sk)
1138{
1139 const struct inet_connection_sock *icsk = inet_csk(sk);
1140
1141 return icsk->icsk_ca_ops->flags & TCP_CONG_NEEDS_ECN;
1142}
1143
1144static inline void tcp_set_ca_state(struct sock *sk, const u8 ca_state)
1145{
1146 struct inet_connection_sock *icsk = inet_csk(sk);
1147
1148 if (icsk->icsk_ca_ops->set_state)
1149 icsk->icsk_ca_ops->set_state(sk, ca_state);
1150 icsk->icsk_ca_state = ca_state;
1151}
1152
1153static inline void tcp_ca_event(struct sock *sk, const enum tcp_ca_event event)
1154{
1155 const struct inet_connection_sock *icsk = inet_csk(sk);
1156
1157 if (icsk->icsk_ca_ops->cwnd_event)
1158 icsk->icsk_ca_ops->cwnd_event(sk, event);
1159}
1160
1161/* From tcp_rate.c */
1162void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb);
1163void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb,
1164 struct rate_sample *rs);
1165void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost,
1166 bool is_sack_reneg, struct rate_sample *rs);
1167void tcp_rate_check_app_limited(struct sock *sk);
1168
1169/* These functions determine how the current flow behaves in respect of SACK
1170 * handling. SACK is negotiated with the peer, and therefore it can vary
1171 * between different flows.
1172 *
1173 * tcp_is_sack - SACK enabled
1174 * tcp_is_reno - No SACK
1175 */
1176static inline int tcp_is_sack(const struct tcp_sock *tp)
1177{
1178 return likely(tp->rx_opt.sack_ok);
1179}
1180
1181static inline bool tcp_is_reno(const struct tcp_sock *tp)
1182{
1183 return !tcp_is_sack(tp);
1184}
1185
1186static inline unsigned int tcp_left_out(const struct tcp_sock *tp)
1187{
1188 return tp->sacked_out + tp->lost_out;
1189}
1190
1191/* This determines how many packets are "in the network" to the best
1192 * of our knowledge. In many cases it is conservative, but where
1193 * detailed information is available from the receiver (via SACK
1194 * blocks etc.) we can make more aggressive calculations.
1195 *
1196 * Use this for decisions involving congestion control, use just
1197 * tp->packets_out to determine if the send queue is empty or not.
1198 *
1199 * Read this equation as:
1200 *
1201 * "Packets sent once on transmission queue" MINUS
1202 * "Packets left network, but not honestly ACKed yet" PLUS
1203 * "Packets fast retransmitted"
1204 */
1205static inline unsigned int tcp_packets_in_flight(const struct tcp_sock *tp)
1206{
1207 return tp->packets_out - tcp_left_out(tp) + tp->retrans_out;
1208}
1209
1210#define TCP_INFINITE_SSTHRESH 0x7fffffff
1211
1212static inline bool tcp_in_slow_start(const struct tcp_sock *tp)
1213{
1214 return tp->snd_cwnd < tp->snd_ssthresh;
1215}
1216
1217static inline bool tcp_in_initial_slowstart(const struct tcp_sock *tp)
1218{
1219 return tp->snd_ssthresh >= TCP_INFINITE_SSTHRESH;
1220}
1221
1222static inline bool tcp_in_cwnd_reduction(const struct sock *sk)
1223{
1224 return (TCPF_CA_CWR | TCPF_CA_Recovery) &
1225 (1 << inet_csk(sk)->icsk_ca_state);
1226}
1227
1228/* If cwnd > ssthresh, we may raise ssthresh to be half-way to cwnd.
1229 * The exception is cwnd reduction phase, when cwnd is decreasing towards
1230 * ssthresh.
1231 */
1232static inline __u32 tcp_current_ssthresh(const struct sock *sk)
1233{
1234 const struct tcp_sock *tp = tcp_sk(sk);
1235
1236 if (tcp_in_cwnd_reduction(sk))
1237 return tp->snd_ssthresh;
1238 else
1239 return max(tp->snd_ssthresh,
1240 ((tp->snd_cwnd >> 1) +
1241 (tp->snd_cwnd >> 2)));
1242}
1243
1244/* Use define here intentionally to get WARN_ON location shown at the caller */
1245#define tcp_verify_left_out(tp) WARN_ON(tcp_left_out(tp) > tp->packets_out)
1246
1247void tcp_enter_cwr(struct sock *sk);
1248__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst);
1249
1250/* The maximum number of MSS of available cwnd for which TSO defers
1251 * sending if not using sysctl_tcp_tso_win_divisor.
1252 */
1253static inline __u32 tcp_max_tso_deferred_mss(const struct tcp_sock *tp)
1254{
1255 return 3;
1256}
1257
1258/* Returns end sequence number of the receiver's advertised window */
1259static inline u32 tcp_wnd_end(const struct tcp_sock *tp)
1260{
1261 return tp->snd_una + tp->snd_wnd;
1262}
1263
1264/* We follow the spirit of RFC2861 to validate cwnd but implement a more
1265 * flexible approach. The RFC suggests cwnd should not be raised unless
1266 * it was fully used previously. And that's exactly what we do in
1267 * congestion avoidance mode. But in slow start we allow cwnd to grow
1268 * as long as the application has used half the cwnd.
1269 * Example :
1270 * cwnd is 10 (IW10), but application sends 9 frames.
1271 * We allow cwnd to reach 18 when all frames are ACKed.
1272 * This check is safe because it's as aggressive as slow start which already
1273 * risks 100% overshoot. The advantage is that we discourage application to
1274 * either send more filler packets or data to artificially blow up the cwnd
1275 * usage, and allow application-limited process to probe bw more aggressively.
1276 */
1277static inline bool tcp_is_cwnd_limited(const struct sock *sk)
1278{
1279 const struct tcp_sock *tp = tcp_sk(sk);
1280
1281 /* If in slow start, ensure cwnd grows to twice what was ACKed. */
1282 if (tcp_in_slow_start(tp))
1283 return tp->snd_cwnd < 2 * tp->max_packets_out;
1284
1285 return tp->is_cwnd_limited;
1286}
1287
1288/* BBR congestion control needs pacing.
1289 * Same remark for SO_MAX_PACING_RATE.
1290 * sch_fq packet scheduler is efficiently handling pacing,
1291 * but is not always installed/used.
1292 * Return true if TCP stack should pace packets itself.
1293 */
1294static inline bool tcp_needs_internal_pacing(const struct sock *sk)
1295{
1296 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
1297}
1298
1299/* Estimates in how many jiffies next packet for this flow can be sent.
1300 * Scheduling a retransmit timer too early would be silly.
1301 */
1302static inline unsigned long tcp_pacing_delay(const struct sock *sk)
1303{
1304 s64 delay = tcp_sk(sk)->tcp_wstamp_ns - tcp_sk(sk)->tcp_clock_cache;
1305
1306 return delay > 0 ? nsecs_to_jiffies(delay) : 0;
1307}
1308
1309static inline void tcp_reset_xmit_timer(struct sock *sk,
1310 const int what,
1311 unsigned long when,
1312 const unsigned long max_when)
1313{
1314 inet_csk_reset_xmit_timer(sk, what, when + tcp_pacing_delay(sk),
1315 max_when);
1316}
1317
1318/* Something is really bad, we could not queue an additional packet,
1319 * because qdisc is full or receiver sent a 0 window, or we are paced.
1320 * We do not want to add fuel to the fire, or abort too early,
1321 * so make sure the timer we arm now is at least 200ms in the future,
1322 * regardless of current icsk_rto value (as it could be ~2ms)
1323 */
1324static inline unsigned long tcp_probe0_base(const struct sock *sk)
1325{
1326 return max_t(unsigned long, inet_csk(sk)->icsk_rto, TCP_RTO_MIN);
1327}
1328
1329/* Variant of inet_csk_rto_backoff() used for zero window probes */
1330static inline unsigned long tcp_probe0_when(const struct sock *sk,
1331 unsigned long max_when)
1332{
1333 u64 when = (u64)tcp_probe0_base(sk) << inet_csk(sk)->icsk_backoff;
1334
1335 return (unsigned long)min_t(u64, when, max_when);
1336}
1337
1338static inline void tcp_check_probe_timer(struct sock *sk)
1339{
1340 if (!tcp_sk(sk)->packets_out && !inet_csk(sk)->icsk_pending)
1341 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
1342 tcp_probe0_base(sk), TCP_RTO_MAX);
1343}
1344
1345static inline void tcp_init_wl(struct tcp_sock *tp, u32 seq)
1346{
1347 tp->snd_wl1 = seq;
1348}
1349
1350static inline void tcp_update_wl(struct tcp_sock *tp, u32 seq)
1351{
1352 tp->snd_wl1 = seq;
1353}
1354
1355/*
1356 * Calculate(/check) TCP checksum
1357 */
1358static inline __sum16 tcp_v4_check(int len, __be32 saddr,
1359 __be32 daddr, __wsum base)
1360{
1361 return csum_tcpudp_magic(saddr, daddr, len, IPPROTO_TCP, base);
1362}
1363
1364static inline bool tcp_checksum_complete(struct sk_buff *skb)
1365{
1366 return !skb_csum_unnecessary(skb) &&
1367 __skb_checksum_complete(skb);
1368}
1369
1370bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb);
1371int tcp_filter(struct sock *sk, struct sk_buff *skb);
1372void tcp_set_state(struct sock *sk, int state);
1373void tcp_done(struct sock *sk);
1374int tcp_abort(struct sock *sk, int err);
1375
1376static inline void tcp_sack_reset(struct tcp_options_received *rx_opt)
1377{
1378 rx_opt->dsack = 0;
1379 rx_opt->num_sacks = 0;
1380}
1381
1382void tcp_cwnd_restart(struct sock *sk, s32 delta);
1383
1384static inline void tcp_slow_start_after_idle_check(struct sock *sk)
1385{
1386 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1387 struct tcp_sock *tp = tcp_sk(sk);
1388 s32 delta;
1389
1390 if (!sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle || tp->packets_out ||
1391 ca_ops->cong_control)
1392 return;
1393 delta = tcp_jiffies32 - tp->lsndtime;
1394 if (delta > inet_csk(sk)->icsk_rto)
1395 tcp_cwnd_restart(sk, delta);
1396}
1397
1398/* Determine a window scaling and initial window to offer. */
1399void tcp_select_initial_window(const struct sock *sk, int __space,
1400 __u32 mss, __u32 *rcv_wnd,
1401 __u32 *window_clamp, int wscale_ok,
1402 __u8 *rcv_wscale, __u32 init_rcv_wnd);
1403
1404static inline int tcp_win_from_space(const struct sock *sk, int space)
1405{
1406 int tcp_adv_win_scale = sock_net(sk)->ipv4.sysctl_tcp_adv_win_scale;
1407
1408 return tcp_adv_win_scale <= 0 ?
1409 (space>>(-tcp_adv_win_scale)) :
1410 space - (space>>tcp_adv_win_scale);
1411}
1412
1413/* Note: caller must be prepared to deal with negative returns */
1414static inline int tcp_space(const struct sock *sk)
1415{
1416 return tcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf) -
1417 READ_ONCE(sk->sk_backlog.len) -
1418 atomic_read(&sk->sk_rmem_alloc));
1419}
1420
1421static inline int tcp_full_space(const struct sock *sk)
1422{
1423 return tcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf));
1424}
1425
1426/* We provision sk_rcvbuf around 200% of sk_rcvlowat.
1427 * If 87.5 % (7/8) of the space has been consumed, we want to override
1428 * SO_RCVLOWAT constraint, since we are receiving skbs with too small
1429 * len/truesize ratio.
1430 */
1431static inline bool tcp_rmem_pressure(const struct sock *sk)
1432{
1433 int rcvbuf = READ_ONCE(sk->sk_rcvbuf);
1434 int threshold = rcvbuf - (rcvbuf >> 3);
1435
1436 return atomic_read(&sk->sk_rmem_alloc) > threshold;
1437}
1438
1439extern void tcp_openreq_init_rwin(struct request_sock *req,
1440 const struct sock *sk_listener,
1441 const struct dst_entry *dst);
1442
1443void tcp_enter_memory_pressure(struct sock *sk);
1444void tcp_leave_memory_pressure(struct sock *sk);
1445
1446static inline int keepalive_intvl_when(const struct tcp_sock *tp)
1447{
1448 struct net *net = sock_net((struct sock *)tp);
1449
1450 return tp->keepalive_intvl ? : net->ipv4.sysctl_tcp_keepalive_intvl;
1451}
1452
1453static inline int keepalive_time_when(const struct tcp_sock *tp)
1454{
1455 struct net *net = sock_net((struct sock *)tp);
1456
1457 return tp->keepalive_time ? : net->ipv4.sysctl_tcp_keepalive_time;
1458}
1459
1460static inline int keepalive_probes(const struct tcp_sock *tp)
1461{
1462 struct net *net = sock_net((struct sock *)tp);
1463
1464 return tp->keepalive_probes ? : net->ipv4.sysctl_tcp_keepalive_probes;
1465}
1466
1467static inline u32 keepalive_time_elapsed(const struct tcp_sock *tp)
1468{
1469 const struct inet_connection_sock *icsk = &tp->inet_conn;
1470
1471 return min_t(u32, tcp_jiffies32 - icsk->icsk_ack.lrcvtime,
1472 tcp_jiffies32 - tp->rcv_tstamp);
1473}
1474
1475static inline int tcp_fin_time(const struct sock *sk)
1476{
1477 int fin_timeout = tcp_sk(sk)->linger2 ? : sock_net(sk)->ipv4.sysctl_tcp_fin_timeout;
1478 const int rto = inet_csk(sk)->icsk_rto;
1479
1480 if (fin_timeout < (rto << 2) - (rto >> 1))
1481 fin_timeout = (rto << 2) - (rto >> 1);
1482
1483 return fin_timeout;
1484}
1485
1486static inline bool tcp_paws_check(const struct tcp_options_received *rx_opt,
1487 int paws_win)
1488{
1489 if ((s32)(rx_opt->ts_recent - rx_opt->rcv_tsval) <= paws_win)
1490 return true;
1491 if (unlikely(!time_before32(ktime_get_seconds(),
1492 rx_opt->ts_recent_stamp + TCP_PAWS_24DAYS)))
1493 return true;
1494 /*
1495 * Some OSes send SYN and SYNACK messages with tsval=0 tsecr=0,
1496 * then following tcp messages have valid values. Ignore 0 value,
1497 * or else 'negative' tsval might forbid us to accept their packets.
1498 */
1499 if (!rx_opt->ts_recent)
1500 return true;
1501 return false;
1502}
1503
1504static inline bool tcp_paws_reject(const struct tcp_options_received *rx_opt,
1505 int rst)
1506{
1507 if (tcp_paws_check(rx_opt, 0))
1508 return false;
1509
1510 /* RST segments are not recommended to carry timestamp,
1511 and, if they do, it is recommended to ignore PAWS because
1512 "their cleanup function should take precedence over timestamps."
1513 Certainly, it is mistake. It is necessary to understand the reasons
1514 of this constraint to relax it: if peer reboots, clock may go
1515 out-of-sync and half-open connections will not be reset.
1516 Actually, the problem would be not existing if all
1517 the implementations followed draft about maintaining clock
1518 via reboots. Linux-2.2 DOES NOT!
1519
1520 However, we can relax time bounds for RST segments to MSL.
1521 */
1522 if (rst && !time_before32(ktime_get_seconds(),
1523 rx_opt->ts_recent_stamp + TCP_PAWS_MSL))
1524 return false;
1525 return true;
1526}
1527
1528bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
1529 int mib_idx, u32 *last_oow_ack_time);
1530
1531static inline void tcp_mib_init(struct net *net)
1532{
1533 /* See RFC 2012 */
1534 TCP_ADD_STATS(net, TCP_MIB_RTOALGORITHM, 1);
1535 TCP_ADD_STATS(net, TCP_MIB_RTOMIN, TCP_RTO_MIN*1000/HZ);
1536 TCP_ADD_STATS(net, TCP_MIB_RTOMAX, TCP_RTO_MAX*1000/HZ);
1537 TCP_ADD_STATS(net, TCP_MIB_MAXCONN, -1);
1538}
1539
1540/* from STCP */
1541static inline void tcp_clear_retrans_hints_partial(struct tcp_sock *tp)
1542{
1543 tp->lost_skb_hint = NULL;
1544}
1545
1546static inline void tcp_clear_all_retrans_hints(struct tcp_sock *tp)
1547{
1548 tcp_clear_retrans_hints_partial(tp);
1549 tp->retransmit_skb_hint = NULL;
1550}
1551
1552union tcp_md5_addr {
1553 struct in_addr a4;
1554#if IS_ENABLED(CONFIG_IPV6)
1555 struct in6_addr a6;
1556#endif
1557};
1558
1559/* - key database */
1560struct tcp_md5sig_key {
1561 struct hlist_node node;
1562 u8 keylen;
1563 u8 family; /* AF_INET or AF_INET6 */
1564 u8 prefixlen;
1565 union tcp_md5_addr addr;
1566 int l3index; /* set if key added with L3 scope */
1567 u8 key[TCP_MD5SIG_MAXKEYLEN];
1568 struct rcu_head rcu;
1569};
1570
1571/* - sock block */
1572struct tcp_md5sig_info {
1573 struct hlist_head head;
1574 struct rcu_head rcu;
1575};
1576
1577/* - pseudo header */
1578struct tcp4_pseudohdr {
1579 __be32 saddr;
1580 __be32 daddr;
1581 __u8 pad;
1582 __u8 protocol;
1583 __be16 len;
1584};
1585
1586struct tcp6_pseudohdr {
1587 struct in6_addr saddr;
1588 struct in6_addr daddr;
1589 __be32 len;
1590 __be32 protocol; /* including padding */
1591};
1592
1593union tcp_md5sum_block {
1594 struct tcp4_pseudohdr ip4;
1595#if IS_ENABLED(CONFIG_IPV6)
1596 struct tcp6_pseudohdr ip6;
1597#endif
1598};
1599
1600/* - pool: digest algorithm, hash description and scratch buffer */
1601struct tcp_md5sig_pool {
1602 struct ahash_request *md5_req;
1603 void *scratch;
1604};
1605
1606/* - functions */
1607int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1608 const struct sock *sk, const struct sk_buff *skb);
1609int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1610 int family, u8 prefixlen, int l3index,
1611 const u8 *newkey, u8 newkeylen, gfp_t gfp);
1612int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr,
1613 int family, u8 prefixlen, int l3index);
1614struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
1615 const struct sock *addr_sk);
1616
1617#ifdef CONFIG_TCP_MD5SIG
1618#include <linux/jump_label.h>
1619extern struct static_key_false tcp_md5_needed;
1620struct tcp_md5sig_key *__tcp_md5_do_lookup(const struct sock *sk, int l3index,
1621 const union tcp_md5_addr *addr,
1622 int family);
1623static inline struct tcp_md5sig_key *
1624tcp_md5_do_lookup(const struct sock *sk, int l3index,
1625 const union tcp_md5_addr *addr, int family)
1626{
1627 if (!static_branch_unlikely(&tcp_md5_needed))
1628 return NULL;
1629 return __tcp_md5_do_lookup(sk, l3index, addr, family);
1630}
1631
1632#define tcp_twsk_md5_key(twsk) ((twsk)->tw_md5_key)
1633#else
1634static inline struct tcp_md5sig_key *
1635tcp_md5_do_lookup(const struct sock *sk, int l3index,
1636 const union tcp_md5_addr *addr, int family)
1637{
1638 return NULL;
1639}
1640#define tcp_twsk_md5_key(twsk) NULL
1641#endif
1642
1643bool tcp_alloc_md5sig_pool(void);
1644
1645struct tcp_md5sig_pool *tcp_get_md5sig_pool(void);
1646static inline void tcp_put_md5sig_pool(void)
1647{
1648 local_bh_enable();
1649}
1650
1651int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *, const struct sk_buff *,
1652 unsigned int header_len);
1653int tcp_md5_hash_key(struct tcp_md5sig_pool *hp,
1654 const struct tcp_md5sig_key *key);
1655
1656/* From tcp_fastopen.c */
1657void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
1658 struct tcp_fastopen_cookie *cookie);
1659void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
1660 struct tcp_fastopen_cookie *cookie, bool syn_lost,
1661 u16 try_exp);
1662struct tcp_fastopen_request {
1663 /* Fast Open cookie. Size 0 means a cookie request */
1664 struct tcp_fastopen_cookie cookie;
1665 struct msghdr *data; /* data in MSG_FASTOPEN */
1666 size_t size;
1667 int copied; /* queued in tcp_connect() */
1668 struct ubuf_info *uarg;
1669};
1670void tcp_free_fastopen_req(struct tcp_sock *tp);
1671void tcp_fastopen_destroy_cipher(struct sock *sk);
1672void tcp_fastopen_ctx_destroy(struct net *net);
1673int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
1674 void *primary_key, void *backup_key);
1675int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
1676 u64 *key);
1677void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb);
1678struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
1679 struct request_sock *req,
1680 struct tcp_fastopen_cookie *foc,
1681 const struct dst_entry *dst);
1682void tcp_fastopen_init_key_once(struct net *net);
1683bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
1684 struct tcp_fastopen_cookie *cookie);
1685bool tcp_fastopen_defer_connect(struct sock *sk, int *err);
1686#define TCP_FASTOPEN_KEY_LENGTH sizeof(siphash_key_t)
1687#define TCP_FASTOPEN_KEY_MAX 2
1688#define TCP_FASTOPEN_KEY_BUF_LENGTH \
1689 (TCP_FASTOPEN_KEY_LENGTH * TCP_FASTOPEN_KEY_MAX)
1690
1691/* Fastopen key context */
1692struct tcp_fastopen_context {
1693 siphash_key_t key[TCP_FASTOPEN_KEY_MAX];
1694 int num;
1695 struct rcu_head rcu;
1696};
1697
1698extern unsigned int sysctl_tcp_fastopen_blackhole_timeout;
1699void tcp_fastopen_active_disable(struct sock *sk);
1700bool tcp_fastopen_active_should_disable(struct sock *sk);
1701void tcp_fastopen_active_disable_ofo_check(struct sock *sk);
1702void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired);
1703
1704/* Caller needs to wrap with rcu_read_(un)lock() */
1705static inline
1706struct tcp_fastopen_context *tcp_fastopen_get_ctx(const struct sock *sk)
1707{
1708 struct tcp_fastopen_context *ctx;
1709
1710 ctx = rcu_dereference(inet_csk(sk)->icsk_accept_queue.fastopenq.ctx);
1711 if (!ctx)
1712 ctx = rcu_dereference(sock_net(sk)->ipv4.tcp_fastopen_ctx);
1713 return ctx;
1714}
1715
1716static inline
1717bool tcp_fastopen_cookie_match(const struct tcp_fastopen_cookie *foc,
1718 const struct tcp_fastopen_cookie *orig)
1719{
1720 if (orig->len == TCP_FASTOPEN_COOKIE_SIZE &&
1721 orig->len == foc->len &&
1722 !memcmp(orig->val, foc->val, foc->len))
1723 return true;
1724 return false;
1725}
1726
1727static inline
1728int tcp_fastopen_context_len(const struct tcp_fastopen_context *ctx)
1729{
1730 return ctx->num;
1731}
1732
1733/* Latencies incurred by various limits for a sender. They are
1734 * chronograph-like stats that are mutually exclusive.
1735 */
1736enum tcp_chrono {
1737 TCP_CHRONO_UNSPEC,
1738 TCP_CHRONO_BUSY, /* Actively sending data (non-empty write queue) */
1739 TCP_CHRONO_RWND_LIMITED, /* Stalled by insufficient receive window */
1740 TCP_CHRONO_SNDBUF_LIMITED, /* Stalled by insufficient send buffer */
1741 __TCP_CHRONO_MAX,
1742};
1743
1744void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type);
1745void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type);
1746
1747/* This helper is needed, because skb->tcp_tsorted_anchor uses
1748 * the same memory storage than skb->destructor/_skb_refdst
1749 */
1750static inline void tcp_skb_tsorted_anchor_cleanup(struct sk_buff *skb)
1751{
1752 skb->destructor = NULL;
1753 skb->_skb_refdst = 0UL;
1754}
1755
1756#define tcp_skb_tsorted_save(skb) { \
1757 unsigned long _save = skb->_skb_refdst; \
1758 skb->_skb_refdst = 0UL;
1759
1760#define tcp_skb_tsorted_restore(skb) \
1761 skb->_skb_refdst = _save; \
1762}
1763
1764void tcp_write_queue_purge(struct sock *sk);
1765
1766static inline struct sk_buff *tcp_rtx_queue_head(const struct sock *sk)
1767{
1768 return skb_rb_first(&sk->tcp_rtx_queue);
1769}
1770
1771static inline struct sk_buff *tcp_rtx_queue_tail(const struct sock *sk)
1772{
1773 return skb_rb_last(&sk->tcp_rtx_queue);
1774}
1775
1776static inline struct sk_buff *tcp_write_queue_head(const struct sock *sk)
1777{
1778 return skb_peek(&sk->sk_write_queue);
1779}
1780
1781static inline struct sk_buff *tcp_write_queue_tail(const struct sock *sk)
1782{
1783 return skb_peek_tail(&sk->sk_write_queue);
1784}
1785
1786#define tcp_for_write_queue_from_safe(skb, tmp, sk) \
1787 skb_queue_walk_from_safe(&(sk)->sk_write_queue, skb, tmp)
1788
1789static inline struct sk_buff *tcp_send_head(const struct sock *sk)
1790{
1791 return skb_peek(&sk->sk_write_queue);
1792}
1793
1794static inline bool tcp_skb_is_last(const struct sock *sk,
1795 const struct sk_buff *skb)
1796{
1797 return skb_queue_is_last(&sk->sk_write_queue, skb);
1798}
1799
1800/**
1801 * tcp_write_queue_empty - test if any payload (or FIN) is available in write queue
1802 * @sk: socket
1803 *
1804 * Since the write queue can have a temporary empty skb in it,
1805 * we must not use "return skb_queue_empty(&sk->sk_write_queue)"
1806 */
1807static inline bool tcp_write_queue_empty(const struct sock *sk)
1808{
1809 const struct tcp_sock *tp = tcp_sk(sk);
1810
1811 return tp->write_seq == tp->snd_nxt;
1812}
1813
1814static inline bool tcp_rtx_queue_empty(const struct sock *sk)
1815{
1816 return RB_EMPTY_ROOT(&sk->tcp_rtx_queue);
1817}
1818
1819static inline bool tcp_rtx_and_write_queues_empty(const struct sock *sk)
1820{
1821 return tcp_rtx_queue_empty(sk) && tcp_write_queue_empty(sk);
1822}
1823
1824static inline void tcp_add_write_queue_tail(struct sock *sk, struct sk_buff *skb)
1825{
1826 __skb_queue_tail(&sk->sk_write_queue, skb);
1827
1828 /* Queue it, remembering where we must start sending. */
1829 if (sk->sk_write_queue.next == skb)
1830 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
1831}
1832
1833/* Insert new before skb on the write queue of sk. */
1834static inline void tcp_insert_write_queue_before(struct sk_buff *new,
1835 struct sk_buff *skb,
1836 struct sock *sk)
1837{
1838 __skb_queue_before(&sk->sk_write_queue, skb, new);
1839}
1840
1841static inline void tcp_unlink_write_queue(struct sk_buff *skb, struct sock *sk)
1842{
1843 tcp_skb_tsorted_anchor_cleanup(skb);
1844 __skb_unlink(skb, &sk->sk_write_queue);
1845}
1846
1847void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb);
1848
1849static inline void tcp_rtx_queue_unlink(struct sk_buff *skb, struct sock *sk)
1850{
1851 tcp_skb_tsorted_anchor_cleanup(skb);
1852 rb_erase(&skb->rbnode, &sk->tcp_rtx_queue);
1853}
1854
1855static inline void tcp_rtx_queue_unlink_and_free(struct sk_buff *skb, struct sock *sk)
1856{
1857 list_del(&skb->tcp_tsorted_anchor);
1858 tcp_rtx_queue_unlink(skb, sk);
1859 sk_wmem_free_skb(sk, skb);
1860}
1861
1862static inline void tcp_push_pending_frames(struct sock *sk)
1863{
1864 if (tcp_send_head(sk)) {
1865 struct tcp_sock *tp = tcp_sk(sk);
1866
1867 __tcp_push_pending_frames(sk, tcp_current_mss(sk), tp->nonagle);
1868 }
1869}
1870
1871/* Start sequence of the skb just after the highest skb with SACKed
1872 * bit, valid only if sacked_out > 0 or when the caller has ensured
1873 * validity by itself.
1874 */
1875static inline u32 tcp_highest_sack_seq(struct tcp_sock *tp)
1876{
1877 if (!tp->sacked_out)
1878 return tp->snd_una;
1879
1880 if (tp->highest_sack == NULL)
1881 return tp->snd_nxt;
1882
1883 return TCP_SKB_CB(tp->highest_sack)->seq;
1884}
1885
1886static inline void tcp_advance_highest_sack(struct sock *sk, struct sk_buff *skb)
1887{
1888 tcp_sk(sk)->highest_sack = skb_rb_next(skb);
1889}
1890
1891static inline struct sk_buff *tcp_highest_sack(struct sock *sk)
1892{
1893 return tcp_sk(sk)->highest_sack;
1894}
1895
1896static inline void tcp_highest_sack_reset(struct sock *sk)
1897{
1898 tcp_sk(sk)->highest_sack = tcp_rtx_queue_head(sk);
1899}
1900
1901/* Called when old skb is about to be deleted and replaced by new skb */
1902static inline void tcp_highest_sack_replace(struct sock *sk,
1903 struct sk_buff *old,
1904 struct sk_buff *new)
1905{
1906 if (old == tcp_highest_sack(sk))
1907 tcp_sk(sk)->highest_sack = new;
1908}
1909
1910/* This helper checks if socket has IP_TRANSPARENT set */
1911static inline bool inet_sk_transparent(const struct sock *sk)
1912{
1913 switch (sk->sk_state) {
1914 case TCP_TIME_WAIT:
1915 return inet_twsk(sk)->tw_transparent;
1916 case TCP_NEW_SYN_RECV:
1917 return inet_rsk(inet_reqsk(sk))->no_srccheck;
1918 }
1919 return inet_sk(sk)->transparent;
1920}
1921
1922/* Determines whether this is a thin stream (which may suffer from
1923 * increased latency). Used to trigger latency-reducing mechanisms.
1924 */
1925static inline bool tcp_stream_is_thin(struct tcp_sock *tp)
1926{
1927 return tp->packets_out < 4 && !tcp_in_initial_slowstart(tp);
1928}
1929
1930/* /proc */
1931enum tcp_seq_states {
1932 TCP_SEQ_STATE_LISTENING,
1933 TCP_SEQ_STATE_ESTABLISHED,
1934};
1935
1936void *tcp_seq_start(struct seq_file *seq, loff_t *pos);
1937void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos);
1938void tcp_seq_stop(struct seq_file *seq, void *v);
1939
1940struct tcp_seq_afinfo {
1941 sa_family_t family;
1942};
1943
1944struct tcp_iter_state {
1945 struct seq_net_private p;
1946 enum tcp_seq_states state;
1947 struct sock *syn_wait_sk;
1948 struct tcp_seq_afinfo *bpf_seq_afinfo;
1949 int bucket, offset, sbucket, num;
1950 loff_t last_pos;
1951};
1952
1953extern struct request_sock_ops tcp_request_sock_ops;
1954extern struct request_sock_ops tcp6_request_sock_ops;
1955
1956void tcp_v4_destroy_sock(struct sock *sk);
1957
1958struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
1959 netdev_features_t features);
1960struct sk_buff *tcp_gro_receive(struct list_head *head, struct sk_buff *skb);
1961INDIRECT_CALLABLE_DECLARE(int tcp4_gro_complete(struct sk_buff *skb, int thoff));
1962INDIRECT_CALLABLE_DECLARE(struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb));
1963INDIRECT_CALLABLE_DECLARE(int tcp6_gro_complete(struct sk_buff *skb, int thoff));
1964INDIRECT_CALLABLE_DECLARE(struct sk_buff *tcp6_gro_receive(struct list_head *head, struct sk_buff *skb));
1965int tcp_gro_complete(struct sk_buff *skb);
1966
1967void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr);
1968
1969static inline u32 tcp_notsent_lowat(const struct tcp_sock *tp)
1970{
1971 struct net *net = sock_net((struct sock *)tp);
1972 return tp->notsent_lowat ?: net->ipv4.sysctl_tcp_notsent_lowat;
1973}
1974
1975/* @wake is one when sk_stream_write_space() calls us.
1976 * This sends EPOLLOUT only if notsent_bytes is half the limit.
1977 * This mimics the strategy used in sock_def_write_space().
1978 */
1979static inline bool tcp_stream_memory_free(const struct sock *sk, int wake)
1980{
1981 const struct tcp_sock *tp = tcp_sk(sk);
1982 u32 notsent_bytes = READ_ONCE(tp->write_seq) -
1983 READ_ONCE(tp->snd_nxt);
1984
1985 return (notsent_bytes << wake) < tcp_notsent_lowat(tp);
1986}
1987
1988#ifdef CONFIG_PROC_FS
1989int tcp4_proc_init(void);
1990void tcp4_proc_exit(void);
1991#endif
1992
1993int tcp_rtx_synack(const struct sock *sk, struct request_sock *req);
1994int tcp_conn_request(struct request_sock_ops *rsk_ops,
1995 const struct tcp_request_sock_ops *af_ops,
1996 struct sock *sk, struct sk_buff *skb);
1997
1998/* TCP af-specific functions */
1999struct tcp_sock_af_ops {
2000#ifdef CONFIG_TCP_MD5SIG
2001 struct tcp_md5sig_key *(*md5_lookup) (const struct sock *sk,
2002 const struct sock *addr_sk);
2003 int (*calc_md5_hash)(char *location,
2004 const struct tcp_md5sig_key *md5,
2005 const struct sock *sk,
2006 const struct sk_buff *skb);
2007 int (*md5_parse)(struct sock *sk,
2008 int optname,
2009 sockptr_t optval,
2010 int optlen);
2011#endif
2012};
2013
2014struct tcp_request_sock_ops {
2015 u16 mss_clamp;
2016#ifdef CONFIG_TCP_MD5SIG
2017 struct tcp_md5sig_key *(*req_md5_lookup)(const struct sock *sk,
2018 const struct sock *addr_sk);
2019 int (*calc_md5_hash) (char *location,
2020 const struct tcp_md5sig_key *md5,
2021 const struct sock *sk,
2022 const struct sk_buff *skb);
2023#endif
2024 void (*init_req)(struct request_sock *req,
2025 const struct sock *sk_listener,
2026 struct sk_buff *skb);
2027#ifdef CONFIG_SYN_COOKIES
2028 __u32 (*cookie_init_seq)(const struct sk_buff *skb,
2029 __u16 *mss);
2030#endif
2031 struct dst_entry *(*route_req)(const struct sock *sk, struct flowi *fl,
2032 const struct request_sock *req);
2033 u32 (*init_seq)(const struct sk_buff *skb);
2034 u32 (*init_ts_off)(const struct net *net, const struct sk_buff *skb);
2035 int (*send_synack)(const struct sock *sk, struct dst_entry *dst,
2036 struct flowi *fl, struct request_sock *req,
2037 struct tcp_fastopen_cookie *foc,
2038 enum tcp_synack_type synack_type);
2039};
2040
2041extern const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops;
2042#if IS_ENABLED(CONFIG_IPV6)
2043extern const struct tcp_request_sock_ops tcp_request_sock_ipv6_ops;
2044#endif
2045
2046#ifdef CONFIG_SYN_COOKIES
2047static inline __u32 cookie_init_sequence(const struct tcp_request_sock_ops *ops,
2048 const struct sock *sk, struct sk_buff *skb,
2049 __u16 *mss)
2050{
2051 tcp_synq_overflow(sk);
2052 __NET_INC_STATS(sock_net(sk), LINUX_MIB_SYNCOOKIESSENT);
2053 return ops->cookie_init_seq(skb, mss);
2054}
2055#else
2056static inline __u32 cookie_init_sequence(const struct tcp_request_sock_ops *ops,
2057 const struct sock *sk, struct sk_buff *skb,
2058 __u16 *mss)
2059{
2060 return 0;
2061}
2062#endif
2063
2064int tcpv4_offload_init(void);
2065
2066void tcp_v4_init(void);
2067void tcp_init(void);
2068
2069/* tcp_recovery.c */
2070void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb);
2071void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced);
2072extern s32 tcp_rack_skb_timeout(struct tcp_sock *tp, struct sk_buff *skb,
2073 u32 reo_wnd);
2074extern void tcp_rack_mark_lost(struct sock *sk);
2075extern void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
2076 u64 xmit_time);
2077extern void tcp_rack_reo_timeout(struct sock *sk);
2078extern void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs);
2079
2080/* At how many usecs into the future should the RTO fire? */
2081static inline s64 tcp_rto_delta_us(const struct sock *sk)
2082{
2083 const struct sk_buff *skb = tcp_rtx_queue_head(sk);
2084 u32 rto = inet_csk(sk)->icsk_rto;
2085 u64 rto_time_stamp_us = tcp_skb_timestamp_us(skb) + jiffies_to_usecs(rto);
2086
2087 return rto_time_stamp_us - tcp_sk(sk)->tcp_mstamp;
2088}
2089
2090/*
2091 * Save and compile IPv4 options, return a pointer to it
2092 */
2093static inline struct ip_options_rcu *tcp_v4_save_options(struct net *net,
2094 struct sk_buff *skb)
2095{
2096 const struct ip_options *opt = &TCP_SKB_CB(skb)->header.h4.opt;
2097 struct ip_options_rcu *dopt = NULL;
2098
2099 if (opt->optlen) {
2100 int opt_size = sizeof(*dopt) + opt->optlen;
2101
2102 dopt = kmalloc(opt_size, GFP_ATOMIC);
2103 if (dopt && __ip_options_echo(net, &dopt->opt, skb, opt)) {
2104 kfree(dopt);
2105 dopt = NULL;
2106 }
2107 }
2108 return dopt;
2109}
2110
2111/* locally generated TCP pure ACKs have skb->truesize == 2
2112 * (check tcp_send_ack() in net/ipv4/tcp_output.c )
2113 * This is much faster than dissecting the packet to find out.
2114 * (Think of GRE encapsulations, IPv4, IPv6, ...)
2115 */
2116static inline bool skb_is_tcp_pure_ack(const struct sk_buff *skb)
2117{
2118 return skb->truesize == 2;
2119}
2120
2121static inline void skb_set_tcp_pure_ack(struct sk_buff *skb)
2122{
2123 skb->truesize = 2;
2124}
2125
2126static inline int tcp_inq(struct sock *sk)
2127{
2128 struct tcp_sock *tp = tcp_sk(sk);
2129 int answ;
2130
2131 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
2132 answ = 0;
2133 } else if (sock_flag(sk, SOCK_URGINLINE) ||
2134 !tp->urg_data ||
2135 before(tp->urg_seq, tp->copied_seq) ||
2136 !before(tp->urg_seq, tp->rcv_nxt)) {
2137
2138 answ = tp->rcv_nxt - tp->copied_seq;
2139
2140 /* Subtract 1, if FIN was received */
2141 if (answ && sock_flag(sk, SOCK_DONE))
2142 answ--;
2143 } else {
2144 answ = tp->urg_seq - tp->copied_seq;
2145 }
2146
2147 return answ;
2148}
2149
2150int tcp_peek_len(struct socket *sock);
2151
2152static inline void tcp_segs_in(struct tcp_sock *tp, const struct sk_buff *skb)
2153{
2154 u16 segs_in;
2155
2156 segs_in = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
2157 tp->segs_in += segs_in;
2158 if (skb->len > tcp_hdrlen(skb))
2159 tp->data_segs_in += segs_in;
2160}
2161
2162/*
2163 * TCP listen path runs lockless.
2164 * We forced "struct sock" to be const qualified to make sure
2165 * we don't modify one of its field by mistake.
2166 * Here, we increment sk_drops which is an atomic_t, so we can safely
2167 * make sock writable again.
2168 */
2169static inline void tcp_listendrop(const struct sock *sk)
2170{
2171 atomic_inc(&((struct sock *)sk)->sk_drops);
2172 __NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENDROPS);
2173}
2174
2175enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer);
2176
2177/*
2178 * Interface for adding Upper Level Protocols over TCP
2179 */
2180
2181#define TCP_ULP_NAME_MAX 16
2182#define TCP_ULP_MAX 128
2183#define TCP_ULP_BUF_MAX (TCP_ULP_NAME_MAX*TCP_ULP_MAX)
2184
2185struct tcp_ulp_ops {
2186 struct list_head list;
2187
2188 /* initialize ulp */
2189 int (*init)(struct sock *sk);
2190 /* update ulp */
2191 void (*update)(struct sock *sk, struct proto *p,
2192 void (*write_space)(struct sock *sk));
2193 /* cleanup ulp */
2194 void (*release)(struct sock *sk);
2195 /* diagnostic */
2196 int (*get_info)(const struct sock *sk, struct sk_buff *skb);
2197 size_t (*get_info_size)(const struct sock *sk);
2198 /* clone ulp */
2199 void (*clone)(const struct request_sock *req, struct sock *newsk,
2200 const gfp_t priority);
2201
2202 char name[TCP_ULP_NAME_MAX];
2203 struct module *owner;
2204};
2205int tcp_register_ulp(struct tcp_ulp_ops *type);
2206void tcp_unregister_ulp(struct tcp_ulp_ops *type);
2207int tcp_set_ulp(struct sock *sk, const char *name);
2208void tcp_get_available_ulp(char *buf, size_t len);
2209void tcp_cleanup_ulp(struct sock *sk);
2210void tcp_update_ulp(struct sock *sk, struct proto *p,
2211 void (*write_space)(struct sock *sk));
2212
2213#define MODULE_ALIAS_TCP_ULP(name) \
2214 __MODULE_INFO(alias, alias_userspace, name); \
2215 __MODULE_INFO(alias, alias_tcp_ulp, "tcp-ulp-" name)
2216
2217struct sk_msg;
2218struct sk_psock;
2219
2220#ifdef CONFIG_BPF_STREAM_PARSER
2221struct proto *tcp_bpf_get_proto(struct sock *sk, struct sk_psock *psock);
2222void tcp_bpf_clone(const struct sock *sk, struct sock *newsk);
2223#else
2224static inline void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
2225{
2226}
2227#endif /* CONFIG_BPF_STREAM_PARSER */
2228
2229#ifdef CONFIG_NET_SOCK_MSG
2230int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg, u32 bytes,
2231 int flags);
2232int __tcp_bpf_recvmsg(struct sock *sk, struct sk_psock *psock,
2233 struct msghdr *msg, int len, int flags);
2234#endif /* CONFIG_NET_SOCK_MSG */
2235
2236/* Call BPF_SOCK_OPS program that returns an int. If the return value
2237 * is < 0, then the BPF op failed (for example if the loaded BPF
2238 * program does not support the chosen operation or there is no BPF
2239 * program loaded).
2240 */
2241#ifdef CONFIG_BPF
2242static inline int tcp_call_bpf(struct sock *sk, int op, u32 nargs, u32 *args)
2243{
2244 struct bpf_sock_ops_kern sock_ops;
2245 int ret;
2246
2247 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
2248 if (sk_fullsock(sk)) {
2249 sock_ops.is_fullsock = 1;
2250 sock_owned_by_me(sk);
2251 }
2252
2253 sock_ops.sk = sk;
2254 sock_ops.op = op;
2255 if (nargs > 0)
2256 memcpy(sock_ops.args, args, nargs * sizeof(*args));
2257
2258 ret = BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
2259 if (ret == 0)
2260 ret = sock_ops.reply;
2261 else
2262 ret = -1;
2263 return ret;
2264}
2265
2266static inline int tcp_call_bpf_2arg(struct sock *sk, int op, u32 arg1, u32 arg2)
2267{
2268 u32 args[2] = {arg1, arg2};
2269
2270 return tcp_call_bpf(sk, op, 2, args);
2271}
2272
2273static inline int tcp_call_bpf_3arg(struct sock *sk, int op, u32 arg1, u32 arg2,
2274 u32 arg3)
2275{
2276 u32 args[3] = {arg1, arg2, arg3};
2277
2278 return tcp_call_bpf(sk, op, 3, args);
2279}
2280
2281#else
2282static inline int tcp_call_bpf(struct sock *sk, int op, u32 nargs, u32 *args)
2283{
2284 return -EPERM;
2285}
2286
2287static inline int tcp_call_bpf_2arg(struct sock *sk, int op, u32 arg1, u32 arg2)
2288{
2289 return -EPERM;
2290}
2291
2292static inline int tcp_call_bpf_3arg(struct sock *sk, int op, u32 arg1, u32 arg2,
2293 u32 arg3)
2294{
2295 return -EPERM;
2296}
2297
2298#endif
2299
2300static inline u32 tcp_timeout_init(struct sock *sk)
2301{
2302 int timeout;
2303
2304 timeout = tcp_call_bpf(sk, BPF_SOCK_OPS_TIMEOUT_INIT, 0, NULL);
2305
2306 if (timeout <= 0)
2307 timeout = TCP_TIMEOUT_INIT;
2308 return timeout;
2309}
2310
2311static inline u32 tcp_rwnd_init_bpf(struct sock *sk)
2312{
2313 int rwnd;
2314
2315 rwnd = tcp_call_bpf(sk, BPF_SOCK_OPS_RWND_INIT, 0, NULL);
2316
2317 if (rwnd < 0)
2318 rwnd = 0;
2319 return rwnd;
2320}
2321
2322static inline bool tcp_bpf_ca_needs_ecn(struct sock *sk)
2323{
2324 return (tcp_call_bpf(sk, BPF_SOCK_OPS_NEEDS_ECN, 0, NULL) == 1);
2325}
2326
2327static inline void tcp_bpf_rtt(struct sock *sk)
2328{
2329 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_RTT_CB_FLAG))
2330 tcp_call_bpf(sk, BPF_SOCK_OPS_RTT_CB, 0, NULL);
2331}
2332
2333#if IS_ENABLED(CONFIG_SMC)
2334extern struct static_key_false tcp_have_smc;
2335#endif
2336
2337#if IS_ENABLED(CONFIG_TLS_DEVICE)
2338void clean_acked_data_enable(struct inet_connection_sock *icsk,
2339 void (*cad)(struct sock *sk, u32 ack_seq));
2340void clean_acked_data_disable(struct inet_connection_sock *icsk);
2341void clean_acked_data_flush(void);
2342#endif
2343
2344DECLARE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2345static inline void tcp_add_tx_delay(struct sk_buff *skb,
2346 const struct tcp_sock *tp)
2347{
2348 if (static_branch_unlikely(&tcp_tx_delay_enabled))
2349 skb->skb_mstamp_ns += (u64)tp->tcp_tx_delay * NSEC_PER_USEC;
2350}
2351
2352/* Compute Earliest Departure Time for some control packets
2353 * like ACK or RST for TIME_WAIT or non ESTABLISHED sockets.
2354 */
2355static inline u64 tcp_transmit_time(const struct sock *sk)
2356{
2357 if (static_branch_unlikely(&tcp_tx_delay_enabled)) {
2358 u32 delay = (sk->sk_state == TCP_TIME_WAIT) ?
2359 tcp_twsk(sk)->tw_tx_delay : tcp_sk(sk)->tcp_tx_delay;
2360
2361 return tcp_clock_ns() + (u64)delay * NSEC_PER_USEC;
2362 }
2363 return 0;
2364}
2365
2366#endif /* _TCP_H */
1/* SPDX-License-Identifier: GPL-2.0-or-later */
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Definitions for the TCP module.
8 *
9 * Version: @(#)tcp.h 1.0.5 05/23/93
10 *
11 * Authors: Ross Biro
12 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
13 */
14#ifndef _TCP_H
15#define _TCP_H
16
17#define FASTRETRANS_DEBUG 1
18
19#include <linux/list.h>
20#include <linux/tcp.h>
21#include <linux/bug.h>
22#include <linux/slab.h>
23#include <linux/cache.h>
24#include <linux/percpu.h>
25#include <linux/skbuff.h>
26#include <linux/kref.h>
27#include <linux/ktime.h>
28#include <linux/indirect_call_wrapper.h>
29
30#include <net/inet_connection_sock.h>
31#include <net/inet_timewait_sock.h>
32#include <net/inet_hashtables.h>
33#include <net/checksum.h>
34#include <net/request_sock.h>
35#include <net/sock_reuseport.h>
36#include <net/sock.h>
37#include <net/snmp.h>
38#include <net/ip.h>
39#include <net/tcp_states.h>
40#include <net/tcp_ao.h>
41#include <net/inet_ecn.h>
42#include <net/dst.h>
43#include <net/mptcp.h>
44
45#include <linux/seq_file.h>
46#include <linux/memcontrol.h>
47#include <linux/bpf-cgroup.h>
48#include <linux/siphash.h>
49
50extern struct inet_hashinfo tcp_hashinfo;
51
52DECLARE_PER_CPU(unsigned int, tcp_orphan_count);
53int tcp_orphan_count_sum(void);
54
55void tcp_time_wait(struct sock *sk, int state, int timeo);
56
57#define MAX_TCP_HEADER L1_CACHE_ALIGN(128 + MAX_HEADER)
58#define MAX_TCP_OPTION_SPACE 40
59#define TCP_MIN_SND_MSS 48
60#define TCP_MIN_GSO_SIZE (TCP_MIN_SND_MSS - MAX_TCP_OPTION_SPACE)
61
62/*
63 * Never offer a window over 32767 without using window scaling. Some
64 * poor stacks do signed 16bit maths!
65 */
66#define MAX_TCP_WINDOW 32767U
67
68/* Minimal accepted MSS. It is (60+60+8) - (20+20). */
69#define TCP_MIN_MSS 88U
70
71/* The initial MTU to use for probing */
72#define TCP_BASE_MSS 1024
73
74/* probing interval, default to 10 minutes as per RFC4821 */
75#define TCP_PROBE_INTERVAL 600
76
77/* Specify interval when tcp mtu probing will stop */
78#define TCP_PROBE_THRESHOLD 8
79
80/* After receiving this amount of duplicate ACKs fast retransmit starts. */
81#define TCP_FASTRETRANS_THRESH 3
82
83/* Maximal number of ACKs sent quickly to accelerate slow-start. */
84#define TCP_MAX_QUICKACKS 16U
85
86/* Maximal number of window scale according to RFC1323 */
87#define TCP_MAX_WSCALE 14U
88
89/* urg_data states */
90#define TCP_URG_VALID 0x0100
91#define TCP_URG_NOTYET 0x0200
92#define TCP_URG_READ 0x0400
93
94#define TCP_RETR1 3 /*
95 * This is how many retries it does before it
96 * tries to figure out if the gateway is
97 * down. Minimal RFC value is 3; it corresponds
98 * to ~3sec-8min depending on RTO.
99 */
100
101#define TCP_RETR2 15 /*
102 * This should take at least
103 * 90 minutes to time out.
104 * RFC1122 says that the limit is 100 sec.
105 * 15 is ~13-30min depending on RTO.
106 */
107
108#define TCP_SYN_RETRIES 6 /* This is how many retries are done
109 * when active opening a connection.
110 * RFC1122 says the minimum retry MUST
111 * be at least 180secs. Nevertheless
112 * this value is corresponding to
113 * 63secs of retransmission with the
114 * current initial RTO.
115 */
116
117#define TCP_SYNACK_RETRIES 5 /* This is how may retries are done
118 * when passive opening a connection.
119 * This is corresponding to 31secs of
120 * retransmission with the current
121 * initial RTO.
122 */
123
124#define TCP_TIMEWAIT_LEN (60*HZ) /* how long to wait to destroy TIME-WAIT
125 * state, about 60 seconds */
126#define TCP_FIN_TIMEOUT TCP_TIMEWAIT_LEN
127 /* BSD style FIN_WAIT2 deadlock breaker.
128 * It used to be 3min, new value is 60sec,
129 * to combine FIN-WAIT-2 timeout with
130 * TIME-WAIT timer.
131 */
132#define TCP_FIN_TIMEOUT_MAX (120 * HZ) /* max TCP_LINGER2 value (two minutes) */
133
134#define TCP_DELACK_MAX ((unsigned)(HZ/5)) /* maximal time to delay before sending an ACK */
135static_assert((1 << ATO_BITS) > TCP_DELACK_MAX);
136
137#if HZ >= 100
138#define TCP_DELACK_MIN ((unsigned)(HZ/25)) /* minimal time to delay before sending an ACK */
139#define TCP_ATO_MIN ((unsigned)(HZ/25))
140#else
141#define TCP_DELACK_MIN 4U
142#define TCP_ATO_MIN 4U
143#endif
144#define TCP_RTO_MAX ((unsigned)(120*HZ))
145#define TCP_RTO_MIN ((unsigned)(HZ/5))
146#define TCP_TIMEOUT_MIN (2U) /* Min timeout for TCP timers in jiffies */
147
148#define TCP_TIMEOUT_MIN_US (2*USEC_PER_MSEC) /* Min TCP timeout in microsecs */
149
150#define TCP_TIMEOUT_INIT ((unsigned)(1*HZ)) /* RFC6298 2.1 initial RTO value */
151#define TCP_TIMEOUT_FALLBACK ((unsigned)(3*HZ)) /* RFC 1122 initial RTO value, now
152 * used as a fallback RTO for the
153 * initial data transmission if no
154 * valid RTT sample has been acquired,
155 * most likely due to retrans in 3WHS.
156 */
157
158#define TCP_RESOURCE_PROBE_INTERVAL ((unsigned)(HZ/2U)) /* Maximal interval between probes
159 * for local resources.
160 */
161#define TCP_KEEPALIVE_TIME (120*60*HZ) /* two hours */
162#define TCP_KEEPALIVE_PROBES 9 /* Max of 9 keepalive probes */
163#define TCP_KEEPALIVE_INTVL (75*HZ)
164
165#define MAX_TCP_KEEPIDLE 32767
166#define MAX_TCP_KEEPINTVL 32767
167#define MAX_TCP_KEEPCNT 127
168#define MAX_TCP_SYNCNT 127
169
170/* Ensure that TCP PAWS checks are relaxed after ~2147 seconds
171 * to avoid overflows. This assumes a clock smaller than 1 Mhz.
172 * Default clock is 1 Khz, tcp_usec_ts uses 1 Mhz.
173 */
174#define TCP_PAWS_WRAP (INT_MAX / USEC_PER_SEC)
175
176#define TCP_PAWS_MSL 60 /* Per-host timestamps are invalidated
177 * after this time. It should be equal
178 * (or greater than) TCP_TIMEWAIT_LEN
179 * to provide reliability equal to one
180 * provided by timewait state.
181 */
182#define TCP_PAWS_WINDOW 1 /* Replay window for per-host
183 * timestamps. It must be less than
184 * minimal timewait lifetime.
185 */
186/*
187 * TCP option
188 */
189
190#define TCPOPT_NOP 1 /* Padding */
191#define TCPOPT_EOL 0 /* End of options */
192#define TCPOPT_MSS 2 /* Segment size negotiating */
193#define TCPOPT_WINDOW 3 /* Window scaling */
194#define TCPOPT_SACK_PERM 4 /* SACK Permitted */
195#define TCPOPT_SACK 5 /* SACK Block */
196#define TCPOPT_TIMESTAMP 8 /* Better RTT estimations/PAWS */
197#define TCPOPT_MD5SIG 19 /* MD5 Signature (RFC2385) */
198#define TCPOPT_AO 29 /* Authentication Option (RFC5925) */
199#define TCPOPT_MPTCP 30 /* Multipath TCP (RFC6824) */
200#define TCPOPT_FASTOPEN 34 /* Fast open (RFC7413) */
201#define TCPOPT_EXP 254 /* Experimental */
202/* Magic number to be after the option value for sharing TCP
203 * experimental options. See draft-ietf-tcpm-experimental-options-00.txt
204 */
205#define TCPOPT_FASTOPEN_MAGIC 0xF989
206#define TCPOPT_SMC_MAGIC 0xE2D4C3D9
207
208/*
209 * TCP option lengths
210 */
211
212#define TCPOLEN_MSS 4
213#define TCPOLEN_WINDOW 3
214#define TCPOLEN_SACK_PERM 2
215#define TCPOLEN_TIMESTAMP 10
216#define TCPOLEN_MD5SIG 18
217#define TCPOLEN_FASTOPEN_BASE 2
218#define TCPOLEN_EXP_FASTOPEN_BASE 4
219#define TCPOLEN_EXP_SMC_BASE 6
220
221/* But this is what stacks really send out. */
222#define TCPOLEN_TSTAMP_ALIGNED 12
223#define TCPOLEN_WSCALE_ALIGNED 4
224#define TCPOLEN_SACKPERM_ALIGNED 4
225#define TCPOLEN_SACK_BASE 2
226#define TCPOLEN_SACK_BASE_ALIGNED 4
227#define TCPOLEN_SACK_PERBLOCK 8
228#define TCPOLEN_MD5SIG_ALIGNED 20
229#define TCPOLEN_MSS_ALIGNED 4
230#define TCPOLEN_EXP_SMC_BASE_ALIGNED 8
231
232/* Flags in tp->nonagle */
233#define TCP_NAGLE_OFF 1 /* Nagle's algo is disabled */
234#define TCP_NAGLE_CORK 2 /* Socket is corked */
235#define TCP_NAGLE_PUSH 4 /* Cork is overridden for already queued data */
236
237/* TCP thin-stream limits */
238#define TCP_THIN_LINEAR_RETRIES 6 /* After 6 linear retries, do exp. backoff */
239
240/* TCP initial congestion window as per rfc6928 */
241#define TCP_INIT_CWND 10
242
243/* Bit Flags for sysctl_tcp_fastopen */
244#define TFO_CLIENT_ENABLE 1
245#define TFO_SERVER_ENABLE 2
246#define TFO_CLIENT_NO_COOKIE 4 /* Data in SYN w/o cookie option */
247
248/* Accept SYN data w/o any cookie option */
249#define TFO_SERVER_COOKIE_NOT_REQD 0x200
250
251/* Force enable TFO on all listeners, i.e., not requiring the
252 * TCP_FASTOPEN socket option.
253 */
254#define TFO_SERVER_WO_SOCKOPT1 0x400
255
256
257/* sysctl variables for tcp */
258extern int sysctl_tcp_max_orphans;
259extern long sysctl_tcp_mem[3];
260
261#define TCP_RACK_LOSS_DETECTION 0x1 /* Use RACK to detect losses */
262#define TCP_RACK_STATIC_REO_WND 0x2 /* Use static RACK reo wnd */
263#define TCP_RACK_NO_DUPTHRESH 0x4 /* Do not use DUPACK threshold in RACK */
264
265extern atomic_long_t tcp_memory_allocated;
266DECLARE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
267
268extern struct percpu_counter tcp_sockets_allocated;
269extern unsigned long tcp_memory_pressure;
270
271/* optimized version of sk_under_memory_pressure() for TCP sockets */
272static inline bool tcp_under_memory_pressure(const struct sock *sk)
273{
274 if (mem_cgroup_sockets_enabled && sk->sk_memcg &&
275 mem_cgroup_under_socket_pressure(sk->sk_memcg))
276 return true;
277
278 return READ_ONCE(tcp_memory_pressure);
279}
280/*
281 * The next routines deal with comparing 32 bit unsigned ints
282 * and worry about wraparound (automatic with unsigned arithmetic).
283 */
284
285static inline bool before(__u32 seq1, __u32 seq2)
286{
287 return (__s32)(seq1-seq2) < 0;
288}
289#define after(seq2, seq1) before(seq1, seq2)
290
291/* is s2<=s1<=s3 ? */
292static inline bool between(__u32 seq1, __u32 seq2, __u32 seq3)
293{
294 return seq3 - seq2 >= seq1 - seq2;
295}
296
297static inline bool tcp_out_of_memory(struct sock *sk)
298{
299 if (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
300 sk_memory_allocated(sk) > sk_prot_mem_limits(sk, 2))
301 return true;
302 return false;
303}
304
305static inline void tcp_wmem_free_skb(struct sock *sk, struct sk_buff *skb)
306{
307 sk_wmem_queued_add(sk, -skb->truesize);
308 if (!skb_zcopy_pure(skb))
309 sk_mem_uncharge(sk, skb->truesize);
310 else
311 sk_mem_uncharge(sk, SKB_TRUESIZE(skb_end_offset(skb)));
312 __kfree_skb(skb);
313}
314
315void sk_forced_mem_schedule(struct sock *sk, int size);
316
317bool tcp_check_oom(struct sock *sk, int shift);
318
319
320extern struct proto tcp_prot;
321
322#define TCP_INC_STATS(net, field) SNMP_INC_STATS((net)->mib.tcp_statistics, field)
323#define __TCP_INC_STATS(net, field) __SNMP_INC_STATS((net)->mib.tcp_statistics, field)
324#define TCP_DEC_STATS(net, field) SNMP_DEC_STATS((net)->mib.tcp_statistics, field)
325#define TCP_ADD_STATS(net, field, val) SNMP_ADD_STATS((net)->mib.tcp_statistics, field, val)
326
327void tcp_tasklet_init(void);
328
329int tcp_v4_err(struct sk_buff *skb, u32);
330
331void tcp_shutdown(struct sock *sk, int how);
332
333int tcp_v4_early_demux(struct sk_buff *skb);
334int tcp_v4_rcv(struct sk_buff *skb);
335
336void tcp_remove_empty_skb(struct sock *sk);
337int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size);
338int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size);
339int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
340 size_t size, struct ubuf_info *uarg);
341void tcp_splice_eof(struct socket *sock);
342int tcp_send_mss(struct sock *sk, int *size_goal, int flags);
343int tcp_wmem_schedule(struct sock *sk, int copy);
344void tcp_push(struct sock *sk, int flags, int mss_now, int nonagle,
345 int size_goal);
346void tcp_release_cb(struct sock *sk);
347void tcp_wfree(struct sk_buff *skb);
348void tcp_write_timer_handler(struct sock *sk);
349void tcp_delack_timer_handler(struct sock *sk);
350int tcp_ioctl(struct sock *sk, int cmd, int *karg);
351int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb);
352void tcp_rcv_established(struct sock *sk, struct sk_buff *skb);
353void tcp_rcv_space_adjust(struct sock *sk);
354int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp);
355void tcp_twsk_destructor(struct sock *sk);
356void tcp_twsk_purge(struct list_head *net_exit_list, int family);
357ssize_t tcp_splice_read(struct socket *sk, loff_t *ppos,
358 struct pipe_inode_info *pipe, size_t len,
359 unsigned int flags);
360struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
361 bool force_schedule);
362
363static inline void tcp_dec_quickack_mode(struct sock *sk)
364{
365 struct inet_connection_sock *icsk = inet_csk(sk);
366
367 if (icsk->icsk_ack.quick) {
368 /* How many ACKs S/ACKing new data have we sent? */
369 const unsigned int pkts = inet_csk_ack_scheduled(sk) ? 1 : 0;
370
371 if (pkts >= icsk->icsk_ack.quick) {
372 icsk->icsk_ack.quick = 0;
373 /* Leaving quickack mode we deflate ATO. */
374 icsk->icsk_ack.ato = TCP_ATO_MIN;
375 } else
376 icsk->icsk_ack.quick -= pkts;
377 }
378}
379
380#define TCP_ECN_OK 1
381#define TCP_ECN_QUEUE_CWR 2
382#define TCP_ECN_DEMAND_CWR 4
383#define TCP_ECN_SEEN 8
384
385enum tcp_tw_status {
386 TCP_TW_SUCCESS = 0,
387 TCP_TW_RST = 1,
388 TCP_TW_ACK = 2,
389 TCP_TW_SYN = 3
390};
391
392
393enum tcp_tw_status tcp_timewait_state_process(struct inet_timewait_sock *tw,
394 struct sk_buff *skb,
395 const struct tcphdr *th);
396struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
397 struct request_sock *req, bool fastopen,
398 bool *lost_race);
399int tcp_child_process(struct sock *parent, struct sock *child,
400 struct sk_buff *skb);
401void tcp_enter_loss(struct sock *sk);
402void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag);
403void tcp_clear_retrans(struct tcp_sock *tp);
404void tcp_update_metrics(struct sock *sk);
405void tcp_init_metrics(struct sock *sk);
406void tcp_metrics_init(void);
407bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst);
408void __tcp_close(struct sock *sk, long timeout);
409void tcp_close(struct sock *sk, long timeout);
410void tcp_init_sock(struct sock *sk);
411void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb);
412__poll_t tcp_poll(struct file *file, struct socket *sock,
413 struct poll_table_struct *wait);
414int do_tcp_getsockopt(struct sock *sk, int level,
415 int optname, sockptr_t optval, sockptr_t optlen);
416int tcp_getsockopt(struct sock *sk, int level, int optname,
417 char __user *optval, int __user *optlen);
418bool tcp_bpf_bypass_getsockopt(int level, int optname);
419int do_tcp_setsockopt(struct sock *sk, int level, int optname,
420 sockptr_t optval, unsigned int optlen);
421int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
422 unsigned int optlen);
423void tcp_set_keepalive(struct sock *sk, int val);
424void tcp_syn_ack_timeout(const struct request_sock *req);
425int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
426 int flags, int *addr_len);
427int tcp_set_rcvlowat(struct sock *sk, int val);
428int tcp_set_window_clamp(struct sock *sk, int val);
429void tcp_update_recv_tstamps(struct sk_buff *skb,
430 struct scm_timestamping_internal *tss);
431void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
432 struct scm_timestamping_internal *tss);
433void tcp_data_ready(struct sock *sk);
434#ifdef CONFIG_MMU
435int tcp_mmap(struct file *file, struct socket *sock,
436 struct vm_area_struct *vma);
437#endif
438void tcp_parse_options(const struct net *net, const struct sk_buff *skb,
439 struct tcp_options_received *opt_rx,
440 int estab, struct tcp_fastopen_cookie *foc);
441
442/*
443 * BPF SKB-less helpers
444 */
445u16 tcp_v4_get_syncookie(struct sock *sk, struct iphdr *iph,
446 struct tcphdr *th, u32 *cookie);
447u16 tcp_v6_get_syncookie(struct sock *sk, struct ipv6hdr *iph,
448 struct tcphdr *th, u32 *cookie);
449u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss);
450u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
451 const struct tcp_request_sock_ops *af_ops,
452 struct sock *sk, struct tcphdr *th);
453/*
454 * TCP v4 functions exported for the inet6 API
455 */
456
457void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb);
458void tcp_v4_mtu_reduced(struct sock *sk);
459void tcp_req_err(struct sock *sk, u32 seq, bool abort);
460void tcp_ld_RTO_revert(struct sock *sk, u32 seq);
461int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb);
462struct sock *tcp_create_openreq_child(const struct sock *sk,
463 struct request_sock *req,
464 struct sk_buff *skb);
465void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst);
466struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
467 struct request_sock *req,
468 struct dst_entry *dst,
469 struct request_sock *req_unhash,
470 bool *own_req);
471int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb);
472int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len);
473int tcp_connect(struct sock *sk);
474enum tcp_synack_type {
475 TCP_SYNACK_NORMAL,
476 TCP_SYNACK_FASTOPEN,
477 TCP_SYNACK_COOKIE,
478};
479struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
480 struct request_sock *req,
481 struct tcp_fastopen_cookie *foc,
482 enum tcp_synack_type synack_type,
483 struct sk_buff *syn_skb);
484int tcp_disconnect(struct sock *sk, int flags);
485
486void tcp_finish_connect(struct sock *sk, struct sk_buff *skb);
487int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size);
488void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb);
489
490/* From syncookies.c */
491struct sock *tcp_get_cookie_sock(struct sock *sk, struct sk_buff *skb,
492 struct request_sock *req,
493 struct dst_entry *dst);
494int __cookie_v4_check(const struct iphdr *iph, const struct tcphdr *th);
495struct sock *cookie_v4_check(struct sock *sk, struct sk_buff *skb);
496struct request_sock *cookie_tcp_reqsk_alloc(const struct request_sock_ops *ops,
497 struct sock *sk, struct sk_buff *skb,
498 struct tcp_options_received *tcp_opt,
499 int mss, u32 tsoff);
500
501#ifdef CONFIG_SYN_COOKIES
502
503/* Syncookies use a monotonic timer which increments every 60 seconds.
504 * This counter is used both as a hash input and partially encoded into
505 * the cookie value. A cookie is only validated further if the delta
506 * between the current counter value and the encoded one is less than this,
507 * i.e. a sent cookie is valid only at most for 2*60 seconds (or less if
508 * the counter advances immediately after a cookie is generated).
509 */
510#define MAX_SYNCOOKIE_AGE 2
511#define TCP_SYNCOOKIE_PERIOD (60 * HZ)
512#define TCP_SYNCOOKIE_VALID (MAX_SYNCOOKIE_AGE * TCP_SYNCOOKIE_PERIOD)
513
514/* syncookies: remember time of last synqueue overflow
515 * But do not dirty this field too often (once per second is enough)
516 * It is racy as we do not hold a lock, but race is very minor.
517 */
518static inline void tcp_synq_overflow(const struct sock *sk)
519{
520 unsigned int last_overflow;
521 unsigned int now = jiffies;
522
523 if (sk->sk_reuseport) {
524 struct sock_reuseport *reuse;
525
526 reuse = rcu_dereference(sk->sk_reuseport_cb);
527 if (likely(reuse)) {
528 last_overflow = READ_ONCE(reuse->synq_overflow_ts);
529 if (!time_between32(now, last_overflow,
530 last_overflow + HZ))
531 WRITE_ONCE(reuse->synq_overflow_ts, now);
532 return;
533 }
534 }
535
536 last_overflow = READ_ONCE(tcp_sk(sk)->rx_opt.ts_recent_stamp);
537 if (!time_between32(now, last_overflow, last_overflow + HZ))
538 WRITE_ONCE(tcp_sk_rw(sk)->rx_opt.ts_recent_stamp, now);
539}
540
541/* syncookies: no recent synqueue overflow on this listening socket? */
542static inline bool tcp_synq_no_recent_overflow(const struct sock *sk)
543{
544 unsigned int last_overflow;
545 unsigned int now = jiffies;
546
547 if (sk->sk_reuseport) {
548 struct sock_reuseport *reuse;
549
550 reuse = rcu_dereference(sk->sk_reuseport_cb);
551 if (likely(reuse)) {
552 last_overflow = READ_ONCE(reuse->synq_overflow_ts);
553 return !time_between32(now, last_overflow - HZ,
554 last_overflow +
555 TCP_SYNCOOKIE_VALID);
556 }
557 }
558
559 last_overflow = READ_ONCE(tcp_sk(sk)->rx_opt.ts_recent_stamp);
560
561 /* If last_overflow <= jiffies <= last_overflow + TCP_SYNCOOKIE_VALID,
562 * then we're under synflood. However, we have to use
563 * 'last_overflow - HZ' as lower bound. That's because a concurrent
564 * tcp_synq_overflow() could update .ts_recent_stamp after we read
565 * jiffies but before we store .ts_recent_stamp into last_overflow,
566 * which could lead to rejecting a valid syncookie.
567 */
568 return !time_between32(now, last_overflow - HZ,
569 last_overflow + TCP_SYNCOOKIE_VALID);
570}
571
572static inline u32 tcp_cookie_time(void)
573{
574 u64 val = get_jiffies_64();
575
576 do_div(val, TCP_SYNCOOKIE_PERIOD);
577 return val;
578}
579
580u32 __cookie_v4_init_sequence(const struct iphdr *iph, const struct tcphdr *th,
581 u16 *mssp);
582__u32 cookie_v4_init_sequence(const struct sk_buff *skb, __u16 *mss);
583u64 cookie_init_timestamp(struct request_sock *req, u64 now);
584bool cookie_timestamp_decode(const struct net *net,
585 struct tcp_options_received *opt);
586
587static inline bool cookie_ecn_ok(const struct net *net, const struct dst_entry *dst)
588{
589 return READ_ONCE(net->ipv4.sysctl_tcp_ecn) ||
590 dst_feature(dst, RTAX_FEATURE_ECN);
591}
592
593/* From net/ipv6/syncookies.c */
594int __cookie_v6_check(const struct ipv6hdr *iph, const struct tcphdr *th);
595struct sock *cookie_v6_check(struct sock *sk, struct sk_buff *skb);
596
597u32 __cookie_v6_init_sequence(const struct ipv6hdr *iph,
598 const struct tcphdr *th, u16 *mssp);
599__u32 cookie_v6_init_sequence(const struct sk_buff *skb, __u16 *mss);
600#endif
601/* tcp_output.c */
602
603void tcp_skb_entail(struct sock *sk, struct sk_buff *skb);
604void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb);
605void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
606 int nonagle);
607int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs);
608int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs);
609void tcp_retransmit_timer(struct sock *sk);
610void tcp_xmit_retransmit_queue(struct sock *);
611void tcp_simple_retransmit(struct sock *);
612void tcp_enter_recovery(struct sock *sk, bool ece_ack);
613int tcp_trim_head(struct sock *, struct sk_buff *, u32);
614enum tcp_queue {
615 TCP_FRAG_IN_WRITE_QUEUE,
616 TCP_FRAG_IN_RTX_QUEUE,
617};
618int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
619 struct sk_buff *skb, u32 len,
620 unsigned int mss_now, gfp_t gfp);
621
622void tcp_send_probe0(struct sock *);
623int tcp_write_wakeup(struct sock *, int mib);
624void tcp_send_fin(struct sock *sk);
625void tcp_send_active_reset(struct sock *sk, gfp_t priority);
626int tcp_send_synack(struct sock *);
627void tcp_push_one(struct sock *, unsigned int mss_now);
628void __tcp_send_ack(struct sock *sk, u32 rcv_nxt);
629void tcp_send_ack(struct sock *sk);
630void tcp_send_delayed_ack(struct sock *sk);
631void tcp_send_loss_probe(struct sock *sk);
632bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto);
633void tcp_skb_collapse_tstamp(struct sk_buff *skb,
634 const struct sk_buff *next_skb);
635
636/* tcp_input.c */
637void tcp_rearm_rto(struct sock *sk);
638void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req);
639void tcp_reset(struct sock *sk, struct sk_buff *skb);
640void tcp_fin(struct sock *sk);
641void tcp_check_space(struct sock *sk);
642void tcp_sack_compress_send_ack(struct sock *sk);
643
644/* tcp_timer.c */
645void tcp_init_xmit_timers(struct sock *);
646static inline void tcp_clear_xmit_timers(struct sock *sk)
647{
648 if (hrtimer_try_to_cancel(&tcp_sk(sk)->pacing_timer) == 1)
649 __sock_put(sk);
650
651 if (hrtimer_try_to_cancel(&tcp_sk(sk)->compressed_ack_timer) == 1)
652 __sock_put(sk);
653
654 inet_csk_clear_xmit_timers(sk);
655}
656
657unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu);
658unsigned int tcp_current_mss(struct sock *sk);
659u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when);
660
661/* Bound MSS / TSO packet size with the half of the window */
662static inline int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
663{
664 int cutoff;
665
666 /* When peer uses tiny windows, there is no use in packetizing
667 * to sub-MSS pieces for the sake of SWS or making sure there
668 * are enough packets in the pipe for fast recovery.
669 *
670 * On the other hand, for extremely large MSS devices, handling
671 * smaller than MSS windows in this way does make sense.
672 */
673 if (tp->max_window > TCP_MSS_DEFAULT)
674 cutoff = (tp->max_window >> 1);
675 else
676 cutoff = tp->max_window;
677
678 if (cutoff && pktsize > cutoff)
679 return max_t(int, cutoff, 68U - tp->tcp_header_len);
680 else
681 return pktsize;
682}
683
684/* tcp.c */
685void tcp_get_info(struct sock *, struct tcp_info *);
686
687/* Read 'sendfile()'-style from a TCP socket */
688int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
689 sk_read_actor_t recv_actor);
690int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor);
691struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off);
692void tcp_read_done(struct sock *sk, size_t len);
693
694void tcp_initialize_rcv_mss(struct sock *sk);
695
696int tcp_mtu_to_mss(struct sock *sk, int pmtu);
697int tcp_mss_to_mtu(struct sock *sk, int mss);
698void tcp_mtup_init(struct sock *sk);
699
700static inline void tcp_bound_rto(const struct sock *sk)
701{
702 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
703 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
704}
705
706static inline u32 __tcp_set_rto(const struct tcp_sock *tp)
707{
708 return usecs_to_jiffies((tp->srtt_us >> 3) + tp->rttvar_us);
709}
710
711static inline void __tcp_fast_path_on(struct tcp_sock *tp, u32 snd_wnd)
712{
713 /* mptcp hooks are only on the slow path */
714 if (sk_is_mptcp((struct sock *)tp))
715 return;
716
717 tp->pred_flags = htonl((tp->tcp_header_len << 26) |
718 ntohl(TCP_FLAG_ACK) |
719 snd_wnd);
720}
721
722static inline void tcp_fast_path_on(struct tcp_sock *tp)
723{
724 __tcp_fast_path_on(tp, tp->snd_wnd >> tp->rx_opt.snd_wscale);
725}
726
727static inline void tcp_fast_path_check(struct sock *sk)
728{
729 struct tcp_sock *tp = tcp_sk(sk);
730
731 if (RB_EMPTY_ROOT(&tp->out_of_order_queue) &&
732 tp->rcv_wnd &&
733 atomic_read(&sk->sk_rmem_alloc) < sk->sk_rcvbuf &&
734 !tp->urg_data)
735 tcp_fast_path_on(tp);
736}
737
738u32 tcp_delack_max(const struct sock *sk);
739
740/* Compute the actual rto_min value */
741static inline u32 tcp_rto_min(const struct sock *sk)
742{
743 const struct dst_entry *dst = __sk_dst_get(sk);
744 u32 rto_min = inet_csk(sk)->icsk_rto_min;
745
746 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
747 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
748 return rto_min;
749}
750
751static inline u32 tcp_rto_min_us(const struct sock *sk)
752{
753 return jiffies_to_usecs(tcp_rto_min(sk));
754}
755
756static inline bool tcp_ca_dst_locked(const struct dst_entry *dst)
757{
758 return dst_metric_locked(dst, RTAX_CC_ALGO);
759}
760
761/* Minimum RTT in usec. ~0 means not available. */
762static inline u32 tcp_min_rtt(const struct tcp_sock *tp)
763{
764 return minmax_get(&tp->rtt_min);
765}
766
767/* Compute the actual receive window we are currently advertising.
768 * Rcv_nxt can be after the window if our peer push more data
769 * than the offered window.
770 */
771static inline u32 tcp_receive_window(const struct tcp_sock *tp)
772{
773 s32 win = tp->rcv_wup + tp->rcv_wnd - tp->rcv_nxt;
774
775 if (win < 0)
776 win = 0;
777 return (u32) win;
778}
779
780/* Choose a new window, without checks for shrinking, and without
781 * scaling applied to the result. The caller does these things
782 * if necessary. This is a "raw" window selection.
783 */
784u32 __tcp_select_window(struct sock *sk);
785
786void tcp_send_window_probe(struct sock *sk);
787
788/* TCP uses 32bit jiffies to save some space.
789 * Note that this is different from tcp_time_stamp, which
790 * historically has been the same until linux-4.13.
791 */
792#define tcp_jiffies32 ((u32)jiffies)
793
794/*
795 * Deliver a 32bit value for TCP timestamp option (RFC 7323)
796 * It is no longer tied to jiffies, but to 1 ms clock.
797 * Note: double check if you want to use tcp_jiffies32 instead of this.
798 */
799#define TCP_TS_HZ 1000
800
801static inline u64 tcp_clock_ns(void)
802{
803 return ktime_get_ns();
804}
805
806static inline u64 tcp_clock_us(void)
807{
808 return div_u64(tcp_clock_ns(), NSEC_PER_USEC);
809}
810
811static inline u64 tcp_clock_ms(void)
812{
813 return div_u64(tcp_clock_ns(), NSEC_PER_MSEC);
814}
815
816/* TCP Timestamp included in TS option (RFC 1323) can either use ms
817 * or usec resolution. Each socket carries a flag to select one or other
818 * resolution, as the route attribute could change anytime.
819 * Each flow must stick to initial resolution.
820 */
821static inline u32 tcp_clock_ts(bool usec_ts)
822{
823 return usec_ts ? tcp_clock_us() : tcp_clock_ms();
824}
825
826static inline u32 tcp_time_stamp_ms(const struct tcp_sock *tp)
827{
828 return div_u64(tp->tcp_mstamp, USEC_PER_MSEC);
829}
830
831static inline u32 tcp_time_stamp_ts(const struct tcp_sock *tp)
832{
833 if (tp->tcp_usec_ts)
834 return tp->tcp_mstamp;
835 return tcp_time_stamp_ms(tp);
836}
837
838void tcp_mstamp_refresh(struct tcp_sock *tp);
839
840static inline u32 tcp_stamp_us_delta(u64 t1, u64 t0)
841{
842 return max_t(s64, t1 - t0, 0);
843}
844
845/* provide the departure time in us unit */
846static inline u64 tcp_skb_timestamp_us(const struct sk_buff *skb)
847{
848 return div_u64(skb->skb_mstamp_ns, NSEC_PER_USEC);
849}
850
851/* Provide skb TSval in usec or ms unit */
852static inline u32 tcp_skb_timestamp_ts(bool usec_ts, const struct sk_buff *skb)
853{
854 if (usec_ts)
855 return tcp_skb_timestamp_us(skb);
856
857 return div_u64(skb->skb_mstamp_ns, NSEC_PER_MSEC);
858}
859
860static inline u32 tcp_tw_tsval(const struct tcp_timewait_sock *tcptw)
861{
862 return tcp_clock_ts(tcptw->tw_sk.tw_usec_ts) + tcptw->tw_ts_offset;
863}
864
865static inline u32 tcp_rsk_tsval(const struct tcp_request_sock *treq)
866{
867 return tcp_clock_ts(treq->req_usec_ts) + treq->ts_off;
868}
869
870#define tcp_flag_byte(th) (((u_int8_t *)th)[13])
871
872#define TCPHDR_FIN 0x01
873#define TCPHDR_SYN 0x02
874#define TCPHDR_RST 0x04
875#define TCPHDR_PSH 0x08
876#define TCPHDR_ACK 0x10
877#define TCPHDR_URG 0x20
878#define TCPHDR_ECE 0x40
879#define TCPHDR_CWR 0x80
880
881#define TCPHDR_SYN_ECN (TCPHDR_SYN | TCPHDR_ECE | TCPHDR_CWR)
882
883/* This is what the send packet queuing engine uses to pass
884 * TCP per-packet control information to the transmission code.
885 * We also store the host-order sequence numbers in here too.
886 * This is 44 bytes if IPV6 is enabled.
887 * If this grows please adjust skbuff.h:skbuff->cb[xxx] size appropriately.
888 */
889struct tcp_skb_cb {
890 __u32 seq; /* Starting sequence number */
891 __u32 end_seq; /* SEQ + FIN + SYN + datalen */
892 union {
893 /* Note : tcp_tw_isn is used in input path only
894 * (isn chosen by tcp_timewait_state_process())
895 *
896 * tcp_gso_segs/size are used in write queue only,
897 * cf tcp_skb_pcount()/tcp_skb_mss()
898 */
899 __u32 tcp_tw_isn;
900 struct {
901 u16 tcp_gso_segs;
902 u16 tcp_gso_size;
903 };
904 };
905 __u8 tcp_flags; /* TCP header flags. (tcp[13]) */
906
907 __u8 sacked; /* State flags for SACK. */
908#define TCPCB_SACKED_ACKED 0x01 /* SKB ACK'd by a SACK block */
909#define TCPCB_SACKED_RETRANS 0x02 /* SKB retransmitted */
910#define TCPCB_LOST 0x04 /* SKB is lost */
911#define TCPCB_TAGBITS 0x07 /* All tag bits */
912#define TCPCB_REPAIRED 0x10 /* SKB repaired (no skb_mstamp_ns) */
913#define TCPCB_EVER_RETRANS 0x80 /* Ever retransmitted frame */
914#define TCPCB_RETRANS (TCPCB_SACKED_RETRANS|TCPCB_EVER_RETRANS| \
915 TCPCB_REPAIRED)
916
917 __u8 ip_dsfield; /* IPv4 tos or IPv6 dsfield */
918 __u8 txstamp_ack:1, /* Record TX timestamp for ack? */
919 eor:1, /* Is skb MSG_EOR marked? */
920 has_rxtstamp:1, /* SKB has a RX timestamp */
921 unused:5;
922 __u32 ack_seq; /* Sequence number ACK'd */
923 union {
924 struct {
925#define TCPCB_DELIVERED_CE_MASK ((1U<<20) - 1)
926 /* There is space for up to 24 bytes */
927 __u32 is_app_limited:1, /* cwnd not fully used? */
928 delivered_ce:20,
929 unused:11;
930 /* pkts S/ACKed so far upon tx of skb, incl retrans: */
931 __u32 delivered;
932 /* start of send pipeline phase */
933 u64 first_tx_mstamp;
934 /* when we reached the "delivered" count */
935 u64 delivered_mstamp;
936 } tx; /* only used for outgoing skbs */
937 union {
938 struct inet_skb_parm h4;
939#if IS_ENABLED(CONFIG_IPV6)
940 struct inet6_skb_parm h6;
941#endif
942 } header; /* For incoming skbs */
943 };
944};
945
946#define TCP_SKB_CB(__skb) ((struct tcp_skb_cb *)&((__skb)->cb[0]))
947
948extern const struct inet_connection_sock_af_ops ipv4_specific;
949
950#if IS_ENABLED(CONFIG_IPV6)
951/* This is the variant of inet6_iif() that must be used by TCP,
952 * as TCP moves IP6CB into a different location in skb->cb[]
953 */
954static inline int tcp_v6_iif(const struct sk_buff *skb)
955{
956 return TCP_SKB_CB(skb)->header.h6.iif;
957}
958
959static inline int tcp_v6_iif_l3_slave(const struct sk_buff *skb)
960{
961 bool l3_slave = ipv6_l3mdev_skb(TCP_SKB_CB(skb)->header.h6.flags);
962
963 return l3_slave ? skb->skb_iif : TCP_SKB_CB(skb)->header.h6.iif;
964}
965
966/* TCP_SKB_CB reference means this can not be used from early demux */
967static inline int tcp_v6_sdif(const struct sk_buff *skb)
968{
969#if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV)
970 if (skb && ipv6_l3mdev_skb(TCP_SKB_CB(skb)->header.h6.flags))
971 return TCP_SKB_CB(skb)->header.h6.iif;
972#endif
973 return 0;
974}
975
976extern const struct inet_connection_sock_af_ops ipv6_specific;
977
978INDIRECT_CALLABLE_DECLARE(void tcp_v6_send_check(struct sock *sk, struct sk_buff *skb));
979INDIRECT_CALLABLE_DECLARE(int tcp_v6_rcv(struct sk_buff *skb));
980void tcp_v6_early_demux(struct sk_buff *skb);
981
982#endif
983
984/* TCP_SKB_CB reference means this can not be used from early demux */
985static inline int tcp_v4_sdif(struct sk_buff *skb)
986{
987#if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV)
988 if (skb && ipv4_l3mdev_skb(TCP_SKB_CB(skb)->header.h4.flags))
989 return TCP_SKB_CB(skb)->header.h4.iif;
990#endif
991 return 0;
992}
993
994/* Due to TSO, an SKB can be composed of multiple actual
995 * packets. To keep these tracked properly, we use this.
996 */
997static inline int tcp_skb_pcount(const struct sk_buff *skb)
998{
999 return TCP_SKB_CB(skb)->tcp_gso_segs;
1000}
1001
1002static inline void tcp_skb_pcount_set(struct sk_buff *skb, int segs)
1003{
1004 TCP_SKB_CB(skb)->tcp_gso_segs = segs;
1005}
1006
1007static inline void tcp_skb_pcount_add(struct sk_buff *skb, int segs)
1008{
1009 TCP_SKB_CB(skb)->tcp_gso_segs += segs;
1010}
1011
1012/* This is valid iff skb is in write queue and tcp_skb_pcount() > 1. */
1013static inline int tcp_skb_mss(const struct sk_buff *skb)
1014{
1015 return TCP_SKB_CB(skb)->tcp_gso_size;
1016}
1017
1018static inline bool tcp_skb_can_collapse_to(const struct sk_buff *skb)
1019{
1020 return likely(!TCP_SKB_CB(skb)->eor);
1021}
1022
1023static inline bool tcp_skb_can_collapse(const struct sk_buff *to,
1024 const struct sk_buff *from)
1025{
1026 return likely(tcp_skb_can_collapse_to(to) &&
1027 mptcp_skb_can_collapse(to, from) &&
1028 skb_pure_zcopy_same(to, from));
1029}
1030
1031/* Events passed to congestion control interface */
1032enum tcp_ca_event {
1033 CA_EVENT_TX_START, /* first transmit when no packets in flight */
1034 CA_EVENT_CWND_RESTART, /* congestion window restart */
1035 CA_EVENT_COMPLETE_CWR, /* end of congestion recovery */
1036 CA_EVENT_LOSS, /* loss timeout */
1037 CA_EVENT_ECN_NO_CE, /* ECT set, but not CE marked */
1038 CA_EVENT_ECN_IS_CE, /* received CE marked IP packet */
1039};
1040
1041/* Information about inbound ACK, passed to cong_ops->in_ack_event() */
1042enum tcp_ca_ack_event_flags {
1043 CA_ACK_SLOWPATH = (1 << 0), /* In slow path processing */
1044 CA_ACK_WIN_UPDATE = (1 << 1), /* ACK updated window */
1045 CA_ACK_ECE = (1 << 2), /* ECE bit is set on ack */
1046};
1047
1048/*
1049 * Interface for adding new TCP congestion control handlers
1050 */
1051#define TCP_CA_NAME_MAX 16
1052#define TCP_CA_MAX 128
1053#define TCP_CA_BUF_MAX (TCP_CA_NAME_MAX*TCP_CA_MAX)
1054
1055#define TCP_CA_UNSPEC 0
1056
1057/* Algorithm can be set on socket without CAP_NET_ADMIN privileges */
1058#define TCP_CONG_NON_RESTRICTED 0x1
1059/* Requires ECN/ECT set on all packets */
1060#define TCP_CONG_NEEDS_ECN 0x2
1061#define TCP_CONG_MASK (TCP_CONG_NON_RESTRICTED | TCP_CONG_NEEDS_ECN)
1062
1063union tcp_cc_info;
1064
1065struct ack_sample {
1066 u32 pkts_acked;
1067 s32 rtt_us;
1068 u32 in_flight;
1069};
1070
1071/* A rate sample measures the number of (original/retransmitted) data
1072 * packets delivered "delivered" over an interval of time "interval_us".
1073 * The tcp_rate.c code fills in the rate sample, and congestion
1074 * control modules that define a cong_control function to run at the end
1075 * of ACK processing can optionally chose to consult this sample when
1076 * setting cwnd and pacing rate.
1077 * A sample is invalid if "delivered" or "interval_us" is negative.
1078 */
1079struct rate_sample {
1080 u64 prior_mstamp; /* starting timestamp for interval */
1081 u32 prior_delivered; /* tp->delivered at "prior_mstamp" */
1082 u32 prior_delivered_ce;/* tp->delivered_ce at "prior_mstamp" */
1083 s32 delivered; /* number of packets delivered over interval */
1084 s32 delivered_ce; /* number of packets delivered w/ CE marks*/
1085 long interval_us; /* time for tp->delivered to incr "delivered" */
1086 u32 snd_interval_us; /* snd interval for delivered packets */
1087 u32 rcv_interval_us; /* rcv interval for delivered packets */
1088 long rtt_us; /* RTT of last (S)ACKed packet (or -1) */
1089 int losses; /* number of packets marked lost upon ACK */
1090 u32 acked_sacked; /* number of packets newly (S)ACKed upon ACK */
1091 u32 prior_in_flight; /* in flight before this ACK */
1092 u32 last_end_seq; /* end_seq of most recently ACKed packet */
1093 bool is_app_limited; /* is sample from packet with bubble in pipe? */
1094 bool is_retrans; /* is sample from retransmission? */
1095 bool is_ack_delayed; /* is this (likely) a delayed ACK? */
1096};
1097
1098struct tcp_congestion_ops {
1099/* fast path fields are put first to fill one cache line */
1100
1101 /* return slow start threshold (required) */
1102 u32 (*ssthresh)(struct sock *sk);
1103
1104 /* do new cwnd calculation (required) */
1105 void (*cong_avoid)(struct sock *sk, u32 ack, u32 acked);
1106
1107 /* call before changing ca_state (optional) */
1108 void (*set_state)(struct sock *sk, u8 new_state);
1109
1110 /* call when cwnd event occurs (optional) */
1111 void (*cwnd_event)(struct sock *sk, enum tcp_ca_event ev);
1112
1113 /* call when ack arrives (optional) */
1114 void (*in_ack_event)(struct sock *sk, u32 flags);
1115
1116 /* hook for packet ack accounting (optional) */
1117 void (*pkts_acked)(struct sock *sk, const struct ack_sample *sample);
1118
1119 /* override sysctl_tcp_min_tso_segs */
1120 u32 (*min_tso_segs)(struct sock *sk);
1121
1122 /* call when packets are delivered to update cwnd and pacing rate,
1123 * after all the ca_state processing. (optional)
1124 */
1125 void (*cong_control)(struct sock *sk, const struct rate_sample *rs);
1126
1127
1128 /* new value of cwnd after loss (required) */
1129 u32 (*undo_cwnd)(struct sock *sk);
1130 /* returns the multiplier used in tcp_sndbuf_expand (optional) */
1131 u32 (*sndbuf_expand)(struct sock *sk);
1132
1133/* control/slow paths put last */
1134 /* get info for inet_diag (optional) */
1135 size_t (*get_info)(struct sock *sk, u32 ext, int *attr,
1136 union tcp_cc_info *info);
1137
1138 char name[TCP_CA_NAME_MAX];
1139 struct module *owner;
1140 struct list_head list;
1141 u32 key;
1142 u32 flags;
1143
1144 /* initialize private data (optional) */
1145 void (*init)(struct sock *sk);
1146 /* cleanup private data (optional) */
1147 void (*release)(struct sock *sk);
1148} ____cacheline_aligned_in_smp;
1149
1150int tcp_register_congestion_control(struct tcp_congestion_ops *type);
1151void tcp_unregister_congestion_control(struct tcp_congestion_ops *type);
1152int tcp_update_congestion_control(struct tcp_congestion_ops *type,
1153 struct tcp_congestion_ops *old_type);
1154int tcp_validate_congestion_control(struct tcp_congestion_ops *ca);
1155
1156void tcp_assign_congestion_control(struct sock *sk);
1157void tcp_init_congestion_control(struct sock *sk);
1158void tcp_cleanup_congestion_control(struct sock *sk);
1159int tcp_set_default_congestion_control(struct net *net, const char *name);
1160void tcp_get_default_congestion_control(struct net *net, char *name);
1161void tcp_get_available_congestion_control(char *buf, size_t len);
1162void tcp_get_allowed_congestion_control(char *buf, size_t len);
1163int tcp_set_allowed_congestion_control(char *allowed);
1164int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
1165 bool cap_net_admin);
1166u32 tcp_slow_start(struct tcp_sock *tp, u32 acked);
1167void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked);
1168
1169u32 tcp_reno_ssthresh(struct sock *sk);
1170u32 tcp_reno_undo_cwnd(struct sock *sk);
1171void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked);
1172extern struct tcp_congestion_ops tcp_reno;
1173
1174struct tcp_congestion_ops *tcp_ca_find(const char *name);
1175struct tcp_congestion_ops *tcp_ca_find_key(u32 key);
1176u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca);
1177#ifdef CONFIG_INET
1178char *tcp_ca_get_name_by_key(u32 key, char *buffer);
1179#else
1180static inline char *tcp_ca_get_name_by_key(u32 key, char *buffer)
1181{
1182 return NULL;
1183}
1184#endif
1185
1186static inline bool tcp_ca_needs_ecn(const struct sock *sk)
1187{
1188 const struct inet_connection_sock *icsk = inet_csk(sk);
1189
1190 return icsk->icsk_ca_ops->flags & TCP_CONG_NEEDS_ECN;
1191}
1192
1193static inline void tcp_ca_event(struct sock *sk, const enum tcp_ca_event event)
1194{
1195 const struct inet_connection_sock *icsk = inet_csk(sk);
1196
1197 if (icsk->icsk_ca_ops->cwnd_event)
1198 icsk->icsk_ca_ops->cwnd_event(sk, event);
1199}
1200
1201/* From tcp_cong.c */
1202void tcp_set_ca_state(struct sock *sk, const u8 ca_state);
1203
1204/* From tcp_rate.c */
1205void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb);
1206void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb,
1207 struct rate_sample *rs);
1208void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost,
1209 bool is_sack_reneg, struct rate_sample *rs);
1210void tcp_rate_check_app_limited(struct sock *sk);
1211
1212static inline bool tcp_skb_sent_after(u64 t1, u64 t2, u32 seq1, u32 seq2)
1213{
1214 return t1 > t2 || (t1 == t2 && after(seq1, seq2));
1215}
1216
1217/* These functions determine how the current flow behaves in respect of SACK
1218 * handling. SACK is negotiated with the peer, and therefore it can vary
1219 * between different flows.
1220 *
1221 * tcp_is_sack - SACK enabled
1222 * tcp_is_reno - No SACK
1223 */
1224static inline int tcp_is_sack(const struct tcp_sock *tp)
1225{
1226 return likely(tp->rx_opt.sack_ok);
1227}
1228
1229static inline bool tcp_is_reno(const struct tcp_sock *tp)
1230{
1231 return !tcp_is_sack(tp);
1232}
1233
1234static inline unsigned int tcp_left_out(const struct tcp_sock *tp)
1235{
1236 return tp->sacked_out + tp->lost_out;
1237}
1238
1239/* This determines how many packets are "in the network" to the best
1240 * of our knowledge. In many cases it is conservative, but where
1241 * detailed information is available from the receiver (via SACK
1242 * blocks etc.) we can make more aggressive calculations.
1243 *
1244 * Use this for decisions involving congestion control, use just
1245 * tp->packets_out to determine if the send queue is empty or not.
1246 *
1247 * Read this equation as:
1248 *
1249 * "Packets sent once on transmission queue" MINUS
1250 * "Packets left network, but not honestly ACKed yet" PLUS
1251 * "Packets fast retransmitted"
1252 */
1253static inline unsigned int tcp_packets_in_flight(const struct tcp_sock *tp)
1254{
1255 return tp->packets_out - tcp_left_out(tp) + tp->retrans_out;
1256}
1257
1258#define TCP_INFINITE_SSTHRESH 0x7fffffff
1259
1260static inline u32 tcp_snd_cwnd(const struct tcp_sock *tp)
1261{
1262 return tp->snd_cwnd;
1263}
1264
1265static inline void tcp_snd_cwnd_set(struct tcp_sock *tp, u32 val)
1266{
1267 WARN_ON_ONCE((int)val <= 0);
1268 tp->snd_cwnd = val;
1269}
1270
1271static inline bool tcp_in_slow_start(const struct tcp_sock *tp)
1272{
1273 return tcp_snd_cwnd(tp) < tp->snd_ssthresh;
1274}
1275
1276static inline bool tcp_in_initial_slowstart(const struct tcp_sock *tp)
1277{
1278 return tp->snd_ssthresh >= TCP_INFINITE_SSTHRESH;
1279}
1280
1281static inline bool tcp_in_cwnd_reduction(const struct sock *sk)
1282{
1283 return (TCPF_CA_CWR | TCPF_CA_Recovery) &
1284 (1 << inet_csk(sk)->icsk_ca_state);
1285}
1286
1287/* If cwnd > ssthresh, we may raise ssthresh to be half-way to cwnd.
1288 * The exception is cwnd reduction phase, when cwnd is decreasing towards
1289 * ssthresh.
1290 */
1291static inline __u32 tcp_current_ssthresh(const struct sock *sk)
1292{
1293 const struct tcp_sock *tp = tcp_sk(sk);
1294
1295 if (tcp_in_cwnd_reduction(sk))
1296 return tp->snd_ssthresh;
1297 else
1298 return max(tp->snd_ssthresh,
1299 ((tcp_snd_cwnd(tp) >> 1) +
1300 (tcp_snd_cwnd(tp) >> 2)));
1301}
1302
1303/* Use define here intentionally to get WARN_ON location shown at the caller */
1304#define tcp_verify_left_out(tp) WARN_ON(tcp_left_out(tp) > tp->packets_out)
1305
1306void tcp_enter_cwr(struct sock *sk);
1307__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst);
1308
1309/* The maximum number of MSS of available cwnd for which TSO defers
1310 * sending if not using sysctl_tcp_tso_win_divisor.
1311 */
1312static inline __u32 tcp_max_tso_deferred_mss(const struct tcp_sock *tp)
1313{
1314 return 3;
1315}
1316
1317/* Returns end sequence number of the receiver's advertised window */
1318static inline u32 tcp_wnd_end(const struct tcp_sock *tp)
1319{
1320 return tp->snd_una + tp->snd_wnd;
1321}
1322
1323/* We follow the spirit of RFC2861 to validate cwnd but implement a more
1324 * flexible approach. The RFC suggests cwnd should not be raised unless
1325 * it was fully used previously. And that's exactly what we do in
1326 * congestion avoidance mode. But in slow start we allow cwnd to grow
1327 * as long as the application has used half the cwnd.
1328 * Example :
1329 * cwnd is 10 (IW10), but application sends 9 frames.
1330 * We allow cwnd to reach 18 when all frames are ACKed.
1331 * This check is safe because it's as aggressive as slow start which already
1332 * risks 100% overshoot. The advantage is that we discourage application to
1333 * either send more filler packets or data to artificially blow up the cwnd
1334 * usage, and allow application-limited process to probe bw more aggressively.
1335 */
1336static inline bool tcp_is_cwnd_limited(const struct sock *sk)
1337{
1338 const struct tcp_sock *tp = tcp_sk(sk);
1339
1340 if (tp->is_cwnd_limited)
1341 return true;
1342
1343 /* If in slow start, ensure cwnd grows to twice what was ACKed. */
1344 if (tcp_in_slow_start(tp))
1345 return tcp_snd_cwnd(tp) < 2 * tp->max_packets_out;
1346
1347 return false;
1348}
1349
1350/* BBR congestion control needs pacing.
1351 * Same remark for SO_MAX_PACING_RATE.
1352 * sch_fq packet scheduler is efficiently handling pacing,
1353 * but is not always installed/used.
1354 * Return true if TCP stack should pace packets itself.
1355 */
1356static inline bool tcp_needs_internal_pacing(const struct sock *sk)
1357{
1358 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
1359}
1360
1361/* Estimates in how many jiffies next packet for this flow can be sent.
1362 * Scheduling a retransmit timer too early would be silly.
1363 */
1364static inline unsigned long tcp_pacing_delay(const struct sock *sk)
1365{
1366 s64 delay = tcp_sk(sk)->tcp_wstamp_ns - tcp_sk(sk)->tcp_clock_cache;
1367
1368 return delay > 0 ? nsecs_to_jiffies(delay) : 0;
1369}
1370
1371static inline void tcp_reset_xmit_timer(struct sock *sk,
1372 const int what,
1373 unsigned long when,
1374 const unsigned long max_when)
1375{
1376 inet_csk_reset_xmit_timer(sk, what, when + tcp_pacing_delay(sk),
1377 max_when);
1378}
1379
1380/* Something is really bad, we could not queue an additional packet,
1381 * because qdisc is full or receiver sent a 0 window, or we are paced.
1382 * We do not want to add fuel to the fire, or abort too early,
1383 * so make sure the timer we arm now is at least 200ms in the future,
1384 * regardless of current icsk_rto value (as it could be ~2ms)
1385 */
1386static inline unsigned long tcp_probe0_base(const struct sock *sk)
1387{
1388 return max_t(unsigned long, inet_csk(sk)->icsk_rto, TCP_RTO_MIN);
1389}
1390
1391/* Variant of inet_csk_rto_backoff() used for zero window probes */
1392static inline unsigned long tcp_probe0_when(const struct sock *sk,
1393 unsigned long max_when)
1394{
1395 u8 backoff = min_t(u8, ilog2(TCP_RTO_MAX / TCP_RTO_MIN) + 1,
1396 inet_csk(sk)->icsk_backoff);
1397 u64 when = (u64)tcp_probe0_base(sk) << backoff;
1398
1399 return (unsigned long)min_t(u64, when, max_when);
1400}
1401
1402static inline void tcp_check_probe_timer(struct sock *sk)
1403{
1404 if (!tcp_sk(sk)->packets_out && !inet_csk(sk)->icsk_pending)
1405 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
1406 tcp_probe0_base(sk), TCP_RTO_MAX);
1407}
1408
1409static inline void tcp_init_wl(struct tcp_sock *tp, u32 seq)
1410{
1411 tp->snd_wl1 = seq;
1412}
1413
1414static inline void tcp_update_wl(struct tcp_sock *tp, u32 seq)
1415{
1416 tp->snd_wl1 = seq;
1417}
1418
1419/*
1420 * Calculate(/check) TCP checksum
1421 */
1422static inline __sum16 tcp_v4_check(int len, __be32 saddr,
1423 __be32 daddr, __wsum base)
1424{
1425 return csum_tcpudp_magic(saddr, daddr, len, IPPROTO_TCP, base);
1426}
1427
1428static inline bool tcp_checksum_complete(struct sk_buff *skb)
1429{
1430 return !skb_csum_unnecessary(skb) &&
1431 __skb_checksum_complete(skb);
1432}
1433
1434bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
1435 enum skb_drop_reason *reason);
1436
1437
1438int tcp_filter(struct sock *sk, struct sk_buff *skb);
1439void tcp_set_state(struct sock *sk, int state);
1440void tcp_done(struct sock *sk);
1441int tcp_abort(struct sock *sk, int err);
1442
1443static inline void tcp_sack_reset(struct tcp_options_received *rx_opt)
1444{
1445 rx_opt->dsack = 0;
1446 rx_opt->num_sacks = 0;
1447}
1448
1449void tcp_cwnd_restart(struct sock *sk, s32 delta);
1450
1451static inline void tcp_slow_start_after_idle_check(struct sock *sk)
1452{
1453 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1454 struct tcp_sock *tp = tcp_sk(sk);
1455 s32 delta;
1456
1457 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) ||
1458 tp->packets_out || ca_ops->cong_control)
1459 return;
1460 delta = tcp_jiffies32 - tp->lsndtime;
1461 if (delta > inet_csk(sk)->icsk_rto)
1462 tcp_cwnd_restart(sk, delta);
1463}
1464
1465/* Determine a window scaling and initial window to offer. */
1466void tcp_select_initial_window(const struct sock *sk, int __space,
1467 __u32 mss, __u32 *rcv_wnd,
1468 __u32 *window_clamp, int wscale_ok,
1469 __u8 *rcv_wscale, __u32 init_rcv_wnd);
1470
1471static inline int __tcp_win_from_space(u8 scaling_ratio, int space)
1472{
1473 s64 scaled_space = (s64)space * scaling_ratio;
1474
1475 return scaled_space >> TCP_RMEM_TO_WIN_SCALE;
1476}
1477
1478static inline int tcp_win_from_space(const struct sock *sk, int space)
1479{
1480 return __tcp_win_from_space(tcp_sk(sk)->scaling_ratio, space);
1481}
1482
1483/* inverse of __tcp_win_from_space() */
1484static inline int __tcp_space_from_win(u8 scaling_ratio, int win)
1485{
1486 u64 val = (u64)win << TCP_RMEM_TO_WIN_SCALE;
1487
1488 do_div(val, scaling_ratio);
1489 return val;
1490}
1491
1492static inline int tcp_space_from_win(const struct sock *sk, int win)
1493{
1494 return __tcp_space_from_win(tcp_sk(sk)->scaling_ratio, win);
1495}
1496
1497/* Assume a conservative default of 1200 bytes of payload per 4K page.
1498 * This may be adjusted later in tcp_measure_rcv_mss().
1499 */
1500#define TCP_DEFAULT_SCALING_RATIO ((1200 << TCP_RMEM_TO_WIN_SCALE) / \
1501 SKB_TRUESIZE(4096))
1502
1503static inline void tcp_scaling_ratio_init(struct sock *sk)
1504{
1505 tcp_sk(sk)->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
1506}
1507
1508/* Note: caller must be prepared to deal with negative returns */
1509static inline int tcp_space(const struct sock *sk)
1510{
1511 return tcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf) -
1512 READ_ONCE(sk->sk_backlog.len) -
1513 atomic_read(&sk->sk_rmem_alloc));
1514}
1515
1516static inline int tcp_full_space(const struct sock *sk)
1517{
1518 return tcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf));
1519}
1520
1521static inline void __tcp_adjust_rcv_ssthresh(struct sock *sk, u32 new_ssthresh)
1522{
1523 int unused_mem = sk_unused_reserved_mem(sk);
1524 struct tcp_sock *tp = tcp_sk(sk);
1525
1526 tp->rcv_ssthresh = min(tp->rcv_ssthresh, new_ssthresh);
1527 if (unused_mem)
1528 tp->rcv_ssthresh = max_t(u32, tp->rcv_ssthresh,
1529 tcp_win_from_space(sk, unused_mem));
1530}
1531
1532static inline void tcp_adjust_rcv_ssthresh(struct sock *sk)
1533{
1534 __tcp_adjust_rcv_ssthresh(sk, 4U * tcp_sk(sk)->advmss);
1535}
1536
1537void tcp_cleanup_rbuf(struct sock *sk, int copied);
1538void __tcp_cleanup_rbuf(struct sock *sk, int copied);
1539
1540
1541/* We provision sk_rcvbuf around 200% of sk_rcvlowat.
1542 * If 87.5 % (7/8) of the space has been consumed, we want to override
1543 * SO_RCVLOWAT constraint, since we are receiving skbs with too small
1544 * len/truesize ratio.
1545 */
1546static inline bool tcp_rmem_pressure(const struct sock *sk)
1547{
1548 int rcvbuf, threshold;
1549
1550 if (tcp_under_memory_pressure(sk))
1551 return true;
1552
1553 rcvbuf = READ_ONCE(sk->sk_rcvbuf);
1554 threshold = rcvbuf - (rcvbuf >> 3);
1555
1556 return atomic_read(&sk->sk_rmem_alloc) > threshold;
1557}
1558
1559static inline bool tcp_epollin_ready(const struct sock *sk, int target)
1560{
1561 const struct tcp_sock *tp = tcp_sk(sk);
1562 int avail = READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq);
1563
1564 if (avail <= 0)
1565 return false;
1566
1567 return (avail >= target) || tcp_rmem_pressure(sk) ||
1568 (tcp_receive_window(tp) <= inet_csk(sk)->icsk_ack.rcv_mss);
1569}
1570
1571extern void tcp_openreq_init_rwin(struct request_sock *req,
1572 const struct sock *sk_listener,
1573 const struct dst_entry *dst);
1574
1575void tcp_enter_memory_pressure(struct sock *sk);
1576void tcp_leave_memory_pressure(struct sock *sk);
1577
1578static inline int keepalive_intvl_when(const struct tcp_sock *tp)
1579{
1580 struct net *net = sock_net((struct sock *)tp);
1581 int val;
1582
1583 /* Paired with WRITE_ONCE() in tcp_sock_set_keepintvl()
1584 * and do_tcp_setsockopt().
1585 */
1586 val = READ_ONCE(tp->keepalive_intvl);
1587
1588 return val ? : READ_ONCE(net->ipv4.sysctl_tcp_keepalive_intvl);
1589}
1590
1591static inline int keepalive_time_when(const struct tcp_sock *tp)
1592{
1593 struct net *net = sock_net((struct sock *)tp);
1594 int val;
1595
1596 /* Paired with WRITE_ONCE() in tcp_sock_set_keepidle_locked() */
1597 val = READ_ONCE(tp->keepalive_time);
1598
1599 return val ? : READ_ONCE(net->ipv4.sysctl_tcp_keepalive_time);
1600}
1601
1602static inline int keepalive_probes(const struct tcp_sock *tp)
1603{
1604 struct net *net = sock_net((struct sock *)tp);
1605 int val;
1606
1607 /* Paired with WRITE_ONCE() in tcp_sock_set_keepcnt()
1608 * and do_tcp_setsockopt().
1609 */
1610 val = READ_ONCE(tp->keepalive_probes);
1611
1612 return val ? : READ_ONCE(net->ipv4.sysctl_tcp_keepalive_probes);
1613}
1614
1615static inline u32 keepalive_time_elapsed(const struct tcp_sock *tp)
1616{
1617 const struct inet_connection_sock *icsk = &tp->inet_conn;
1618
1619 return min_t(u32, tcp_jiffies32 - icsk->icsk_ack.lrcvtime,
1620 tcp_jiffies32 - tp->rcv_tstamp);
1621}
1622
1623static inline int tcp_fin_time(const struct sock *sk)
1624{
1625 int fin_timeout = tcp_sk(sk)->linger2 ? :
1626 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fin_timeout);
1627 const int rto = inet_csk(sk)->icsk_rto;
1628
1629 if (fin_timeout < (rto << 2) - (rto >> 1))
1630 fin_timeout = (rto << 2) - (rto >> 1);
1631
1632 return fin_timeout;
1633}
1634
1635static inline bool tcp_paws_check(const struct tcp_options_received *rx_opt,
1636 int paws_win)
1637{
1638 if ((s32)(rx_opt->ts_recent - rx_opt->rcv_tsval) <= paws_win)
1639 return true;
1640 if (unlikely(!time_before32(ktime_get_seconds(),
1641 rx_opt->ts_recent_stamp + TCP_PAWS_WRAP)))
1642 return true;
1643 /*
1644 * Some OSes send SYN and SYNACK messages with tsval=0 tsecr=0,
1645 * then following tcp messages have valid values. Ignore 0 value,
1646 * or else 'negative' tsval might forbid us to accept their packets.
1647 */
1648 if (!rx_opt->ts_recent)
1649 return true;
1650 return false;
1651}
1652
1653static inline bool tcp_paws_reject(const struct tcp_options_received *rx_opt,
1654 int rst)
1655{
1656 if (tcp_paws_check(rx_opt, 0))
1657 return false;
1658
1659 /* RST segments are not recommended to carry timestamp,
1660 and, if they do, it is recommended to ignore PAWS because
1661 "their cleanup function should take precedence over timestamps."
1662 Certainly, it is mistake. It is necessary to understand the reasons
1663 of this constraint to relax it: if peer reboots, clock may go
1664 out-of-sync and half-open connections will not be reset.
1665 Actually, the problem would be not existing if all
1666 the implementations followed draft about maintaining clock
1667 via reboots. Linux-2.2 DOES NOT!
1668
1669 However, we can relax time bounds for RST segments to MSL.
1670 */
1671 if (rst && !time_before32(ktime_get_seconds(),
1672 rx_opt->ts_recent_stamp + TCP_PAWS_MSL))
1673 return false;
1674 return true;
1675}
1676
1677bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
1678 int mib_idx, u32 *last_oow_ack_time);
1679
1680static inline void tcp_mib_init(struct net *net)
1681{
1682 /* See RFC 2012 */
1683 TCP_ADD_STATS(net, TCP_MIB_RTOALGORITHM, 1);
1684 TCP_ADD_STATS(net, TCP_MIB_RTOMIN, TCP_RTO_MIN*1000/HZ);
1685 TCP_ADD_STATS(net, TCP_MIB_RTOMAX, TCP_RTO_MAX*1000/HZ);
1686 TCP_ADD_STATS(net, TCP_MIB_MAXCONN, -1);
1687}
1688
1689/* from STCP */
1690static inline void tcp_clear_retrans_hints_partial(struct tcp_sock *tp)
1691{
1692 tp->lost_skb_hint = NULL;
1693}
1694
1695static inline void tcp_clear_all_retrans_hints(struct tcp_sock *tp)
1696{
1697 tcp_clear_retrans_hints_partial(tp);
1698 tp->retransmit_skb_hint = NULL;
1699}
1700
1701#define tcp_md5_addr tcp_ao_addr
1702
1703/* - key database */
1704struct tcp_md5sig_key {
1705 struct hlist_node node;
1706 u8 keylen;
1707 u8 family; /* AF_INET or AF_INET6 */
1708 u8 prefixlen;
1709 u8 flags;
1710 union tcp_md5_addr addr;
1711 int l3index; /* set if key added with L3 scope */
1712 u8 key[TCP_MD5SIG_MAXKEYLEN];
1713 struct rcu_head rcu;
1714};
1715
1716/* - sock block */
1717struct tcp_md5sig_info {
1718 struct hlist_head head;
1719 struct rcu_head rcu;
1720};
1721
1722/* - pseudo header */
1723struct tcp4_pseudohdr {
1724 __be32 saddr;
1725 __be32 daddr;
1726 __u8 pad;
1727 __u8 protocol;
1728 __be16 len;
1729};
1730
1731struct tcp6_pseudohdr {
1732 struct in6_addr saddr;
1733 struct in6_addr daddr;
1734 __be32 len;
1735 __be32 protocol; /* including padding */
1736};
1737
1738union tcp_md5sum_block {
1739 struct tcp4_pseudohdr ip4;
1740#if IS_ENABLED(CONFIG_IPV6)
1741 struct tcp6_pseudohdr ip6;
1742#endif
1743};
1744
1745/*
1746 * struct tcp_sigpool - per-CPU pool of ahash_requests
1747 * @scratch: per-CPU temporary area, that can be used between
1748 * tcp_sigpool_start() and tcp_sigpool_end() to perform
1749 * crypto request
1750 * @req: pre-allocated ahash request
1751 */
1752struct tcp_sigpool {
1753 void *scratch;
1754 struct ahash_request *req;
1755};
1756
1757int tcp_sigpool_alloc_ahash(const char *alg, size_t scratch_size);
1758void tcp_sigpool_get(unsigned int id);
1759void tcp_sigpool_release(unsigned int id);
1760int tcp_sigpool_hash_skb_data(struct tcp_sigpool *hp,
1761 const struct sk_buff *skb,
1762 unsigned int header_len);
1763
1764/**
1765 * tcp_sigpool_start - disable bh and start using tcp_sigpool_ahash
1766 * @id: tcp_sigpool that was previously allocated by tcp_sigpool_alloc_ahash()
1767 * @c: returned tcp_sigpool for usage (uninitialized on failure)
1768 *
1769 * Returns 0 on success, error otherwise.
1770 */
1771int tcp_sigpool_start(unsigned int id, struct tcp_sigpool *c);
1772/**
1773 * tcp_sigpool_end - enable bh and stop using tcp_sigpool
1774 * @c: tcp_sigpool context that was returned by tcp_sigpool_start()
1775 */
1776void tcp_sigpool_end(struct tcp_sigpool *c);
1777size_t tcp_sigpool_algo(unsigned int id, char *buf, size_t buf_len);
1778/* - functions */
1779int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1780 const struct sock *sk, const struct sk_buff *skb);
1781int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1782 int family, u8 prefixlen, int l3index, u8 flags,
1783 const u8 *newkey, u8 newkeylen);
1784int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
1785 int family, u8 prefixlen, int l3index,
1786 struct tcp_md5sig_key *key);
1787
1788int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr,
1789 int family, u8 prefixlen, int l3index, u8 flags);
1790void tcp_clear_md5_list(struct sock *sk);
1791struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
1792 const struct sock *addr_sk);
1793
1794#ifdef CONFIG_TCP_MD5SIG
1795struct tcp_md5sig_key *__tcp_md5_do_lookup(const struct sock *sk, int l3index,
1796 const union tcp_md5_addr *addr,
1797 int family, bool any_l3index);
1798static inline struct tcp_md5sig_key *
1799tcp_md5_do_lookup(const struct sock *sk, int l3index,
1800 const union tcp_md5_addr *addr, int family)
1801{
1802 if (!static_branch_unlikely(&tcp_md5_needed.key))
1803 return NULL;
1804 return __tcp_md5_do_lookup(sk, l3index, addr, family, false);
1805}
1806
1807static inline struct tcp_md5sig_key *
1808tcp_md5_do_lookup_any_l3index(const struct sock *sk,
1809 const union tcp_md5_addr *addr, int family)
1810{
1811 if (!static_branch_unlikely(&tcp_md5_needed.key))
1812 return NULL;
1813 return __tcp_md5_do_lookup(sk, 0, addr, family, true);
1814}
1815
1816enum skb_drop_reason
1817tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
1818 const void *saddr, const void *daddr,
1819 int family, int l3index, const __u8 *hash_location);
1820
1821
1822#define tcp_twsk_md5_key(twsk) ((twsk)->tw_md5_key)
1823#else
1824static inline struct tcp_md5sig_key *
1825tcp_md5_do_lookup(const struct sock *sk, int l3index,
1826 const union tcp_md5_addr *addr, int family)
1827{
1828 return NULL;
1829}
1830
1831static inline struct tcp_md5sig_key *
1832tcp_md5_do_lookup_any_l3index(const struct sock *sk,
1833 const union tcp_md5_addr *addr, int family)
1834{
1835 return NULL;
1836}
1837
1838static inline enum skb_drop_reason
1839tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
1840 const void *saddr, const void *daddr,
1841 int family, int l3index, const __u8 *hash_location)
1842{
1843 return SKB_NOT_DROPPED_YET;
1844}
1845#define tcp_twsk_md5_key(twsk) NULL
1846#endif
1847
1848int tcp_md5_alloc_sigpool(void);
1849void tcp_md5_release_sigpool(void);
1850void tcp_md5_add_sigpool(void);
1851extern int tcp_md5_sigpool_id;
1852
1853int tcp_md5_hash_key(struct tcp_sigpool *hp,
1854 const struct tcp_md5sig_key *key);
1855
1856/* From tcp_fastopen.c */
1857void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
1858 struct tcp_fastopen_cookie *cookie);
1859void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
1860 struct tcp_fastopen_cookie *cookie, bool syn_lost,
1861 u16 try_exp);
1862struct tcp_fastopen_request {
1863 /* Fast Open cookie. Size 0 means a cookie request */
1864 struct tcp_fastopen_cookie cookie;
1865 struct msghdr *data; /* data in MSG_FASTOPEN */
1866 size_t size;
1867 int copied; /* queued in tcp_connect() */
1868 struct ubuf_info *uarg;
1869};
1870void tcp_free_fastopen_req(struct tcp_sock *tp);
1871void tcp_fastopen_destroy_cipher(struct sock *sk);
1872void tcp_fastopen_ctx_destroy(struct net *net);
1873int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
1874 void *primary_key, void *backup_key);
1875int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
1876 u64 *key);
1877void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb);
1878struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
1879 struct request_sock *req,
1880 struct tcp_fastopen_cookie *foc,
1881 const struct dst_entry *dst);
1882void tcp_fastopen_init_key_once(struct net *net);
1883bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
1884 struct tcp_fastopen_cookie *cookie);
1885bool tcp_fastopen_defer_connect(struct sock *sk, int *err);
1886#define TCP_FASTOPEN_KEY_LENGTH sizeof(siphash_key_t)
1887#define TCP_FASTOPEN_KEY_MAX 2
1888#define TCP_FASTOPEN_KEY_BUF_LENGTH \
1889 (TCP_FASTOPEN_KEY_LENGTH * TCP_FASTOPEN_KEY_MAX)
1890
1891/* Fastopen key context */
1892struct tcp_fastopen_context {
1893 siphash_key_t key[TCP_FASTOPEN_KEY_MAX];
1894 int num;
1895 struct rcu_head rcu;
1896};
1897
1898void tcp_fastopen_active_disable(struct sock *sk);
1899bool tcp_fastopen_active_should_disable(struct sock *sk);
1900void tcp_fastopen_active_disable_ofo_check(struct sock *sk);
1901void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired);
1902
1903/* Caller needs to wrap with rcu_read_(un)lock() */
1904static inline
1905struct tcp_fastopen_context *tcp_fastopen_get_ctx(const struct sock *sk)
1906{
1907 struct tcp_fastopen_context *ctx;
1908
1909 ctx = rcu_dereference(inet_csk(sk)->icsk_accept_queue.fastopenq.ctx);
1910 if (!ctx)
1911 ctx = rcu_dereference(sock_net(sk)->ipv4.tcp_fastopen_ctx);
1912 return ctx;
1913}
1914
1915static inline
1916bool tcp_fastopen_cookie_match(const struct tcp_fastopen_cookie *foc,
1917 const struct tcp_fastopen_cookie *orig)
1918{
1919 if (orig->len == TCP_FASTOPEN_COOKIE_SIZE &&
1920 orig->len == foc->len &&
1921 !memcmp(orig->val, foc->val, foc->len))
1922 return true;
1923 return false;
1924}
1925
1926static inline
1927int tcp_fastopen_context_len(const struct tcp_fastopen_context *ctx)
1928{
1929 return ctx->num;
1930}
1931
1932/* Latencies incurred by various limits for a sender. They are
1933 * chronograph-like stats that are mutually exclusive.
1934 */
1935enum tcp_chrono {
1936 TCP_CHRONO_UNSPEC,
1937 TCP_CHRONO_BUSY, /* Actively sending data (non-empty write queue) */
1938 TCP_CHRONO_RWND_LIMITED, /* Stalled by insufficient receive window */
1939 TCP_CHRONO_SNDBUF_LIMITED, /* Stalled by insufficient send buffer */
1940 __TCP_CHRONO_MAX,
1941};
1942
1943void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type);
1944void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type);
1945
1946/* This helper is needed, because skb->tcp_tsorted_anchor uses
1947 * the same memory storage than skb->destructor/_skb_refdst
1948 */
1949static inline void tcp_skb_tsorted_anchor_cleanup(struct sk_buff *skb)
1950{
1951 skb->destructor = NULL;
1952 skb->_skb_refdst = 0UL;
1953}
1954
1955#define tcp_skb_tsorted_save(skb) { \
1956 unsigned long _save = skb->_skb_refdst; \
1957 skb->_skb_refdst = 0UL;
1958
1959#define tcp_skb_tsorted_restore(skb) \
1960 skb->_skb_refdst = _save; \
1961}
1962
1963void tcp_write_queue_purge(struct sock *sk);
1964
1965static inline struct sk_buff *tcp_rtx_queue_head(const struct sock *sk)
1966{
1967 return skb_rb_first(&sk->tcp_rtx_queue);
1968}
1969
1970static inline struct sk_buff *tcp_rtx_queue_tail(const struct sock *sk)
1971{
1972 return skb_rb_last(&sk->tcp_rtx_queue);
1973}
1974
1975static inline struct sk_buff *tcp_write_queue_tail(const struct sock *sk)
1976{
1977 return skb_peek_tail(&sk->sk_write_queue);
1978}
1979
1980#define tcp_for_write_queue_from_safe(skb, tmp, sk) \
1981 skb_queue_walk_from_safe(&(sk)->sk_write_queue, skb, tmp)
1982
1983static inline struct sk_buff *tcp_send_head(const struct sock *sk)
1984{
1985 return skb_peek(&sk->sk_write_queue);
1986}
1987
1988static inline bool tcp_skb_is_last(const struct sock *sk,
1989 const struct sk_buff *skb)
1990{
1991 return skb_queue_is_last(&sk->sk_write_queue, skb);
1992}
1993
1994/**
1995 * tcp_write_queue_empty - test if any payload (or FIN) is available in write queue
1996 * @sk: socket
1997 *
1998 * Since the write queue can have a temporary empty skb in it,
1999 * we must not use "return skb_queue_empty(&sk->sk_write_queue)"
2000 */
2001static inline bool tcp_write_queue_empty(const struct sock *sk)
2002{
2003 const struct tcp_sock *tp = tcp_sk(sk);
2004
2005 return tp->write_seq == tp->snd_nxt;
2006}
2007
2008static inline bool tcp_rtx_queue_empty(const struct sock *sk)
2009{
2010 return RB_EMPTY_ROOT(&sk->tcp_rtx_queue);
2011}
2012
2013static inline bool tcp_rtx_and_write_queues_empty(const struct sock *sk)
2014{
2015 return tcp_rtx_queue_empty(sk) && tcp_write_queue_empty(sk);
2016}
2017
2018static inline void tcp_add_write_queue_tail(struct sock *sk, struct sk_buff *skb)
2019{
2020 __skb_queue_tail(&sk->sk_write_queue, skb);
2021
2022 /* Queue it, remembering where we must start sending. */
2023 if (sk->sk_write_queue.next == skb)
2024 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
2025}
2026
2027/* Insert new before skb on the write queue of sk. */
2028static inline void tcp_insert_write_queue_before(struct sk_buff *new,
2029 struct sk_buff *skb,
2030 struct sock *sk)
2031{
2032 __skb_queue_before(&sk->sk_write_queue, skb, new);
2033}
2034
2035static inline void tcp_unlink_write_queue(struct sk_buff *skb, struct sock *sk)
2036{
2037 tcp_skb_tsorted_anchor_cleanup(skb);
2038 __skb_unlink(skb, &sk->sk_write_queue);
2039}
2040
2041void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb);
2042
2043static inline void tcp_rtx_queue_unlink(struct sk_buff *skb, struct sock *sk)
2044{
2045 tcp_skb_tsorted_anchor_cleanup(skb);
2046 rb_erase(&skb->rbnode, &sk->tcp_rtx_queue);
2047}
2048
2049static inline void tcp_rtx_queue_unlink_and_free(struct sk_buff *skb, struct sock *sk)
2050{
2051 list_del(&skb->tcp_tsorted_anchor);
2052 tcp_rtx_queue_unlink(skb, sk);
2053 tcp_wmem_free_skb(sk, skb);
2054}
2055
2056static inline void tcp_push_pending_frames(struct sock *sk)
2057{
2058 if (tcp_send_head(sk)) {
2059 struct tcp_sock *tp = tcp_sk(sk);
2060
2061 __tcp_push_pending_frames(sk, tcp_current_mss(sk), tp->nonagle);
2062 }
2063}
2064
2065/* Start sequence of the skb just after the highest skb with SACKed
2066 * bit, valid only if sacked_out > 0 or when the caller has ensured
2067 * validity by itself.
2068 */
2069static inline u32 tcp_highest_sack_seq(struct tcp_sock *tp)
2070{
2071 if (!tp->sacked_out)
2072 return tp->snd_una;
2073
2074 if (tp->highest_sack == NULL)
2075 return tp->snd_nxt;
2076
2077 return TCP_SKB_CB(tp->highest_sack)->seq;
2078}
2079
2080static inline void tcp_advance_highest_sack(struct sock *sk, struct sk_buff *skb)
2081{
2082 tcp_sk(sk)->highest_sack = skb_rb_next(skb);
2083}
2084
2085static inline struct sk_buff *tcp_highest_sack(struct sock *sk)
2086{
2087 return tcp_sk(sk)->highest_sack;
2088}
2089
2090static inline void tcp_highest_sack_reset(struct sock *sk)
2091{
2092 tcp_sk(sk)->highest_sack = tcp_rtx_queue_head(sk);
2093}
2094
2095/* Called when old skb is about to be deleted and replaced by new skb */
2096static inline void tcp_highest_sack_replace(struct sock *sk,
2097 struct sk_buff *old,
2098 struct sk_buff *new)
2099{
2100 if (old == tcp_highest_sack(sk))
2101 tcp_sk(sk)->highest_sack = new;
2102}
2103
2104/* This helper checks if socket has IP_TRANSPARENT set */
2105static inline bool inet_sk_transparent(const struct sock *sk)
2106{
2107 switch (sk->sk_state) {
2108 case TCP_TIME_WAIT:
2109 return inet_twsk(sk)->tw_transparent;
2110 case TCP_NEW_SYN_RECV:
2111 return inet_rsk(inet_reqsk(sk))->no_srccheck;
2112 }
2113 return inet_test_bit(TRANSPARENT, sk);
2114}
2115
2116/* Determines whether this is a thin stream (which may suffer from
2117 * increased latency). Used to trigger latency-reducing mechanisms.
2118 */
2119static inline bool tcp_stream_is_thin(struct tcp_sock *tp)
2120{
2121 return tp->packets_out < 4 && !tcp_in_initial_slowstart(tp);
2122}
2123
2124/* /proc */
2125enum tcp_seq_states {
2126 TCP_SEQ_STATE_LISTENING,
2127 TCP_SEQ_STATE_ESTABLISHED,
2128};
2129
2130void *tcp_seq_start(struct seq_file *seq, loff_t *pos);
2131void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos);
2132void tcp_seq_stop(struct seq_file *seq, void *v);
2133
2134struct tcp_seq_afinfo {
2135 sa_family_t family;
2136};
2137
2138struct tcp_iter_state {
2139 struct seq_net_private p;
2140 enum tcp_seq_states state;
2141 struct sock *syn_wait_sk;
2142 int bucket, offset, sbucket, num;
2143 loff_t last_pos;
2144};
2145
2146extern struct request_sock_ops tcp_request_sock_ops;
2147extern struct request_sock_ops tcp6_request_sock_ops;
2148
2149void tcp_v4_destroy_sock(struct sock *sk);
2150
2151struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
2152 netdev_features_t features);
2153struct sk_buff *tcp_gro_receive(struct list_head *head, struct sk_buff *skb);
2154INDIRECT_CALLABLE_DECLARE(int tcp4_gro_complete(struct sk_buff *skb, int thoff));
2155INDIRECT_CALLABLE_DECLARE(struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb));
2156INDIRECT_CALLABLE_DECLARE(int tcp6_gro_complete(struct sk_buff *skb, int thoff));
2157INDIRECT_CALLABLE_DECLARE(struct sk_buff *tcp6_gro_receive(struct list_head *head, struct sk_buff *skb));
2158#ifdef CONFIG_INET
2159void tcp_gro_complete(struct sk_buff *skb);
2160#else
2161static inline void tcp_gro_complete(struct sk_buff *skb) { }
2162#endif
2163
2164void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr);
2165
2166static inline u32 tcp_notsent_lowat(const struct tcp_sock *tp)
2167{
2168 struct net *net = sock_net((struct sock *)tp);
2169 u32 val;
2170
2171 val = READ_ONCE(tp->notsent_lowat);
2172
2173 return val ?: READ_ONCE(net->ipv4.sysctl_tcp_notsent_lowat);
2174}
2175
2176bool tcp_stream_memory_free(const struct sock *sk, int wake);
2177
2178#ifdef CONFIG_PROC_FS
2179int tcp4_proc_init(void);
2180void tcp4_proc_exit(void);
2181#endif
2182
2183int tcp_rtx_synack(const struct sock *sk, struct request_sock *req);
2184int tcp_conn_request(struct request_sock_ops *rsk_ops,
2185 const struct tcp_request_sock_ops *af_ops,
2186 struct sock *sk, struct sk_buff *skb);
2187
2188/* TCP af-specific functions */
2189struct tcp_sock_af_ops {
2190#ifdef CONFIG_TCP_MD5SIG
2191 struct tcp_md5sig_key *(*md5_lookup) (const struct sock *sk,
2192 const struct sock *addr_sk);
2193 int (*calc_md5_hash)(char *location,
2194 const struct tcp_md5sig_key *md5,
2195 const struct sock *sk,
2196 const struct sk_buff *skb);
2197 int (*md5_parse)(struct sock *sk,
2198 int optname,
2199 sockptr_t optval,
2200 int optlen);
2201#endif
2202#ifdef CONFIG_TCP_AO
2203 int (*ao_parse)(struct sock *sk, int optname, sockptr_t optval, int optlen);
2204 struct tcp_ao_key *(*ao_lookup)(const struct sock *sk,
2205 struct sock *addr_sk,
2206 int sndid, int rcvid);
2207 int (*ao_calc_key_sk)(struct tcp_ao_key *mkt, u8 *key,
2208 const struct sock *sk,
2209 __be32 sisn, __be32 disn, bool send);
2210 int (*calc_ao_hash)(char *location, struct tcp_ao_key *ao,
2211 const struct sock *sk, const struct sk_buff *skb,
2212 const u8 *tkey, int hash_offset, u32 sne);
2213#endif
2214};
2215
2216struct tcp_request_sock_ops {
2217 u16 mss_clamp;
2218#ifdef CONFIG_TCP_MD5SIG
2219 struct tcp_md5sig_key *(*req_md5_lookup)(const struct sock *sk,
2220 const struct sock *addr_sk);
2221 int (*calc_md5_hash) (char *location,
2222 const struct tcp_md5sig_key *md5,
2223 const struct sock *sk,
2224 const struct sk_buff *skb);
2225#endif
2226#ifdef CONFIG_TCP_AO
2227 struct tcp_ao_key *(*ao_lookup)(const struct sock *sk,
2228 struct request_sock *req,
2229 int sndid, int rcvid);
2230 int (*ao_calc_key)(struct tcp_ao_key *mkt, u8 *key, struct request_sock *sk);
2231 int (*ao_synack_hash)(char *ao_hash, struct tcp_ao_key *mkt,
2232 struct request_sock *req, const struct sk_buff *skb,
2233 int hash_offset, u32 sne);
2234#endif
2235#ifdef CONFIG_SYN_COOKIES
2236 __u32 (*cookie_init_seq)(const struct sk_buff *skb,
2237 __u16 *mss);
2238#endif
2239 struct dst_entry *(*route_req)(const struct sock *sk,
2240 struct sk_buff *skb,
2241 struct flowi *fl,
2242 struct request_sock *req);
2243 u32 (*init_seq)(const struct sk_buff *skb);
2244 u32 (*init_ts_off)(const struct net *net, const struct sk_buff *skb);
2245 int (*send_synack)(const struct sock *sk, struct dst_entry *dst,
2246 struct flowi *fl, struct request_sock *req,
2247 struct tcp_fastopen_cookie *foc,
2248 enum tcp_synack_type synack_type,
2249 struct sk_buff *syn_skb);
2250};
2251
2252extern const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops;
2253#if IS_ENABLED(CONFIG_IPV6)
2254extern const struct tcp_request_sock_ops tcp_request_sock_ipv6_ops;
2255#endif
2256
2257#ifdef CONFIG_SYN_COOKIES
2258static inline __u32 cookie_init_sequence(const struct tcp_request_sock_ops *ops,
2259 const struct sock *sk, struct sk_buff *skb,
2260 __u16 *mss)
2261{
2262 tcp_synq_overflow(sk);
2263 __NET_INC_STATS(sock_net(sk), LINUX_MIB_SYNCOOKIESSENT);
2264 return ops->cookie_init_seq(skb, mss);
2265}
2266#else
2267static inline __u32 cookie_init_sequence(const struct tcp_request_sock_ops *ops,
2268 const struct sock *sk, struct sk_buff *skb,
2269 __u16 *mss)
2270{
2271 return 0;
2272}
2273#endif
2274
2275struct tcp_key {
2276 union {
2277 struct {
2278 struct tcp_ao_key *ao_key;
2279 char *traffic_key;
2280 u32 sne;
2281 u8 rcv_next;
2282 };
2283 struct tcp_md5sig_key *md5_key;
2284 };
2285 enum {
2286 TCP_KEY_NONE = 0,
2287 TCP_KEY_MD5,
2288 TCP_KEY_AO,
2289 } type;
2290};
2291
2292static inline void tcp_get_current_key(const struct sock *sk,
2293 struct tcp_key *out)
2294{
2295#if defined(CONFIG_TCP_AO) || defined(CONFIG_TCP_MD5SIG)
2296 const struct tcp_sock *tp = tcp_sk(sk);
2297#endif
2298
2299#ifdef CONFIG_TCP_AO
2300 if (static_branch_unlikely(&tcp_ao_needed.key)) {
2301 struct tcp_ao_info *ao;
2302
2303 ao = rcu_dereference_protected(tp->ao_info,
2304 lockdep_sock_is_held(sk));
2305 if (ao) {
2306 out->ao_key = READ_ONCE(ao->current_key);
2307 out->type = TCP_KEY_AO;
2308 return;
2309 }
2310 }
2311#endif
2312#ifdef CONFIG_TCP_MD5SIG
2313 if (static_branch_unlikely(&tcp_md5_needed.key) &&
2314 rcu_access_pointer(tp->md5sig_info)) {
2315 out->md5_key = tp->af_specific->md5_lookup(sk, sk);
2316 if (out->md5_key) {
2317 out->type = TCP_KEY_MD5;
2318 return;
2319 }
2320 }
2321#endif
2322 out->type = TCP_KEY_NONE;
2323}
2324
2325static inline bool tcp_key_is_md5(const struct tcp_key *key)
2326{
2327#ifdef CONFIG_TCP_MD5SIG
2328 if (static_branch_unlikely(&tcp_md5_needed.key) &&
2329 key->type == TCP_KEY_MD5)
2330 return true;
2331#endif
2332 return false;
2333}
2334
2335static inline bool tcp_key_is_ao(const struct tcp_key *key)
2336{
2337#ifdef CONFIG_TCP_AO
2338 if (static_branch_unlikely(&tcp_ao_needed.key) &&
2339 key->type == TCP_KEY_AO)
2340 return true;
2341#endif
2342 return false;
2343}
2344
2345int tcpv4_offload_init(void);
2346
2347void tcp_v4_init(void);
2348void tcp_init(void);
2349
2350/* tcp_recovery.c */
2351void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb);
2352void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced);
2353extern s32 tcp_rack_skb_timeout(struct tcp_sock *tp, struct sk_buff *skb,
2354 u32 reo_wnd);
2355extern bool tcp_rack_mark_lost(struct sock *sk);
2356extern void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
2357 u64 xmit_time);
2358extern void tcp_rack_reo_timeout(struct sock *sk);
2359extern void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs);
2360
2361/* tcp_plb.c */
2362
2363/*
2364 * Scaling factor for fractions in PLB. For example, tcp_plb_update_state
2365 * expects cong_ratio which represents fraction of traffic that experienced
2366 * congestion over a single RTT. In order to avoid floating point operations,
2367 * this fraction should be mapped to (1 << TCP_PLB_SCALE) and passed in.
2368 */
2369#define TCP_PLB_SCALE 8
2370
2371/* State for PLB (Protective Load Balancing) for a single TCP connection. */
2372struct tcp_plb_state {
2373 u8 consec_cong_rounds:5, /* consecutive congested rounds */
2374 unused:3;
2375 u32 pause_until; /* jiffies32 when PLB can resume rerouting */
2376};
2377
2378static inline void tcp_plb_init(const struct sock *sk,
2379 struct tcp_plb_state *plb)
2380{
2381 plb->consec_cong_rounds = 0;
2382 plb->pause_until = 0;
2383}
2384void tcp_plb_update_state(const struct sock *sk, struct tcp_plb_state *plb,
2385 const int cong_ratio);
2386void tcp_plb_check_rehash(struct sock *sk, struct tcp_plb_state *plb);
2387void tcp_plb_update_state_upon_rto(struct sock *sk, struct tcp_plb_state *plb);
2388
2389/* At how many usecs into the future should the RTO fire? */
2390static inline s64 tcp_rto_delta_us(const struct sock *sk)
2391{
2392 const struct sk_buff *skb = tcp_rtx_queue_head(sk);
2393 u32 rto = inet_csk(sk)->icsk_rto;
2394 u64 rto_time_stamp_us = tcp_skb_timestamp_us(skb) + jiffies_to_usecs(rto);
2395
2396 return rto_time_stamp_us - tcp_sk(sk)->tcp_mstamp;
2397}
2398
2399/*
2400 * Save and compile IPv4 options, return a pointer to it
2401 */
2402static inline struct ip_options_rcu *tcp_v4_save_options(struct net *net,
2403 struct sk_buff *skb)
2404{
2405 const struct ip_options *opt = &TCP_SKB_CB(skb)->header.h4.opt;
2406 struct ip_options_rcu *dopt = NULL;
2407
2408 if (opt->optlen) {
2409 int opt_size = sizeof(*dopt) + opt->optlen;
2410
2411 dopt = kmalloc(opt_size, GFP_ATOMIC);
2412 if (dopt && __ip_options_echo(net, &dopt->opt, skb, opt)) {
2413 kfree(dopt);
2414 dopt = NULL;
2415 }
2416 }
2417 return dopt;
2418}
2419
2420/* locally generated TCP pure ACKs have skb->truesize == 2
2421 * (check tcp_send_ack() in net/ipv4/tcp_output.c )
2422 * This is much faster than dissecting the packet to find out.
2423 * (Think of GRE encapsulations, IPv4, IPv6, ...)
2424 */
2425static inline bool skb_is_tcp_pure_ack(const struct sk_buff *skb)
2426{
2427 return skb->truesize == 2;
2428}
2429
2430static inline void skb_set_tcp_pure_ack(struct sk_buff *skb)
2431{
2432 skb->truesize = 2;
2433}
2434
2435static inline int tcp_inq(struct sock *sk)
2436{
2437 struct tcp_sock *tp = tcp_sk(sk);
2438 int answ;
2439
2440 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
2441 answ = 0;
2442 } else if (sock_flag(sk, SOCK_URGINLINE) ||
2443 !tp->urg_data ||
2444 before(tp->urg_seq, tp->copied_seq) ||
2445 !before(tp->urg_seq, tp->rcv_nxt)) {
2446
2447 answ = tp->rcv_nxt - tp->copied_seq;
2448
2449 /* Subtract 1, if FIN was received */
2450 if (answ && sock_flag(sk, SOCK_DONE))
2451 answ--;
2452 } else {
2453 answ = tp->urg_seq - tp->copied_seq;
2454 }
2455
2456 return answ;
2457}
2458
2459int tcp_peek_len(struct socket *sock);
2460
2461static inline void tcp_segs_in(struct tcp_sock *tp, const struct sk_buff *skb)
2462{
2463 u16 segs_in;
2464
2465 segs_in = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
2466
2467 /* We update these fields while other threads might
2468 * read them from tcp_get_info()
2469 */
2470 WRITE_ONCE(tp->segs_in, tp->segs_in + segs_in);
2471 if (skb->len > tcp_hdrlen(skb))
2472 WRITE_ONCE(tp->data_segs_in, tp->data_segs_in + segs_in);
2473}
2474
2475/*
2476 * TCP listen path runs lockless.
2477 * We forced "struct sock" to be const qualified to make sure
2478 * we don't modify one of its field by mistake.
2479 * Here, we increment sk_drops which is an atomic_t, so we can safely
2480 * make sock writable again.
2481 */
2482static inline void tcp_listendrop(const struct sock *sk)
2483{
2484 atomic_inc(&((struct sock *)sk)->sk_drops);
2485 __NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENDROPS);
2486}
2487
2488enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer);
2489
2490/*
2491 * Interface for adding Upper Level Protocols over TCP
2492 */
2493
2494#define TCP_ULP_NAME_MAX 16
2495#define TCP_ULP_MAX 128
2496#define TCP_ULP_BUF_MAX (TCP_ULP_NAME_MAX*TCP_ULP_MAX)
2497
2498struct tcp_ulp_ops {
2499 struct list_head list;
2500
2501 /* initialize ulp */
2502 int (*init)(struct sock *sk);
2503 /* update ulp */
2504 void (*update)(struct sock *sk, struct proto *p,
2505 void (*write_space)(struct sock *sk));
2506 /* cleanup ulp */
2507 void (*release)(struct sock *sk);
2508 /* diagnostic */
2509 int (*get_info)(struct sock *sk, struct sk_buff *skb);
2510 size_t (*get_info_size)(const struct sock *sk);
2511 /* clone ulp */
2512 void (*clone)(const struct request_sock *req, struct sock *newsk,
2513 const gfp_t priority);
2514
2515 char name[TCP_ULP_NAME_MAX];
2516 struct module *owner;
2517};
2518int tcp_register_ulp(struct tcp_ulp_ops *type);
2519void tcp_unregister_ulp(struct tcp_ulp_ops *type);
2520int tcp_set_ulp(struct sock *sk, const char *name);
2521void tcp_get_available_ulp(char *buf, size_t len);
2522void tcp_cleanup_ulp(struct sock *sk);
2523void tcp_update_ulp(struct sock *sk, struct proto *p,
2524 void (*write_space)(struct sock *sk));
2525
2526#define MODULE_ALIAS_TCP_ULP(name) \
2527 __MODULE_INFO(alias, alias_userspace, name); \
2528 __MODULE_INFO(alias, alias_tcp_ulp, "tcp-ulp-" name)
2529
2530#ifdef CONFIG_NET_SOCK_MSG
2531struct sk_msg;
2532struct sk_psock;
2533
2534#ifdef CONFIG_BPF_SYSCALL
2535int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore);
2536void tcp_bpf_clone(const struct sock *sk, struct sock *newsk);
2537#endif /* CONFIG_BPF_SYSCALL */
2538
2539#ifdef CONFIG_INET
2540void tcp_eat_skb(struct sock *sk, struct sk_buff *skb);
2541#else
2542static inline void tcp_eat_skb(struct sock *sk, struct sk_buff *skb)
2543{
2544}
2545#endif
2546
2547int tcp_bpf_sendmsg_redir(struct sock *sk, bool ingress,
2548 struct sk_msg *msg, u32 bytes, int flags);
2549#endif /* CONFIG_NET_SOCK_MSG */
2550
2551#if !defined(CONFIG_BPF_SYSCALL) || !defined(CONFIG_NET_SOCK_MSG)
2552static inline void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
2553{
2554}
2555#endif
2556
2557#ifdef CONFIG_CGROUP_BPF
2558static inline void bpf_skops_init_skb(struct bpf_sock_ops_kern *skops,
2559 struct sk_buff *skb,
2560 unsigned int end_offset)
2561{
2562 skops->skb = skb;
2563 skops->skb_data_end = skb->data + end_offset;
2564}
2565#else
2566static inline void bpf_skops_init_skb(struct bpf_sock_ops_kern *skops,
2567 struct sk_buff *skb,
2568 unsigned int end_offset)
2569{
2570}
2571#endif
2572
2573/* Call BPF_SOCK_OPS program that returns an int. If the return value
2574 * is < 0, then the BPF op failed (for example if the loaded BPF
2575 * program does not support the chosen operation or there is no BPF
2576 * program loaded).
2577 */
2578#ifdef CONFIG_BPF
2579static inline int tcp_call_bpf(struct sock *sk, int op, u32 nargs, u32 *args)
2580{
2581 struct bpf_sock_ops_kern sock_ops;
2582 int ret;
2583
2584 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
2585 if (sk_fullsock(sk)) {
2586 sock_ops.is_fullsock = 1;
2587 sock_owned_by_me(sk);
2588 }
2589
2590 sock_ops.sk = sk;
2591 sock_ops.op = op;
2592 if (nargs > 0)
2593 memcpy(sock_ops.args, args, nargs * sizeof(*args));
2594
2595 ret = BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
2596 if (ret == 0)
2597 ret = sock_ops.reply;
2598 else
2599 ret = -1;
2600 return ret;
2601}
2602
2603static inline int tcp_call_bpf_2arg(struct sock *sk, int op, u32 arg1, u32 arg2)
2604{
2605 u32 args[2] = {arg1, arg2};
2606
2607 return tcp_call_bpf(sk, op, 2, args);
2608}
2609
2610static inline int tcp_call_bpf_3arg(struct sock *sk, int op, u32 arg1, u32 arg2,
2611 u32 arg3)
2612{
2613 u32 args[3] = {arg1, arg2, arg3};
2614
2615 return tcp_call_bpf(sk, op, 3, args);
2616}
2617
2618#else
2619static inline int tcp_call_bpf(struct sock *sk, int op, u32 nargs, u32 *args)
2620{
2621 return -EPERM;
2622}
2623
2624static inline int tcp_call_bpf_2arg(struct sock *sk, int op, u32 arg1, u32 arg2)
2625{
2626 return -EPERM;
2627}
2628
2629static inline int tcp_call_bpf_3arg(struct sock *sk, int op, u32 arg1, u32 arg2,
2630 u32 arg3)
2631{
2632 return -EPERM;
2633}
2634
2635#endif
2636
2637static inline u32 tcp_timeout_init(struct sock *sk)
2638{
2639 int timeout;
2640
2641 timeout = tcp_call_bpf(sk, BPF_SOCK_OPS_TIMEOUT_INIT, 0, NULL);
2642
2643 if (timeout <= 0)
2644 timeout = TCP_TIMEOUT_INIT;
2645 return min_t(int, timeout, TCP_RTO_MAX);
2646}
2647
2648static inline u32 tcp_rwnd_init_bpf(struct sock *sk)
2649{
2650 int rwnd;
2651
2652 rwnd = tcp_call_bpf(sk, BPF_SOCK_OPS_RWND_INIT, 0, NULL);
2653
2654 if (rwnd < 0)
2655 rwnd = 0;
2656 return rwnd;
2657}
2658
2659static inline bool tcp_bpf_ca_needs_ecn(struct sock *sk)
2660{
2661 return (tcp_call_bpf(sk, BPF_SOCK_OPS_NEEDS_ECN, 0, NULL) == 1);
2662}
2663
2664static inline void tcp_bpf_rtt(struct sock *sk)
2665{
2666 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_RTT_CB_FLAG))
2667 tcp_call_bpf(sk, BPF_SOCK_OPS_RTT_CB, 0, NULL);
2668}
2669
2670#if IS_ENABLED(CONFIG_SMC)
2671extern struct static_key_false tcp_have_smc;
2672#endif
2673
2674#if IS_ENABLED(CONFIG_TLS_DEVICE)
2675void clean_acked_data_enable(struct inet_connection_sock *icsk,
2676 void (*cad)(struct sock *sk, u32 ack_seq));
2677void clean_acked_data_disable(struct inet_connection_sock *icsk);
2678void clean_acked_data_flush(void);
2679#endif
2680
2681DECLARE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2682static inline void tcp_add_tx_delay(struct sk_buff *skb,
2683 const struct tcp_sock *tp)
2684{
2685 if (static_branch_unlikely(&tcp_tx_delay_enabled))
2686 skb->skb_mstamp_ns += (u64)tp->tcp_tx_delay * NSEC_PER_USEC;
2687}
2688
2689/* Compute Earliest Departure Time for some control packets
2690 * like ACK or RST for TIME_WAIT or non ESTABLISHED sockets.
2691 */
2692static inline u64 tcp_transmit_time(const struct sock *sk)
2693{
2694 if (static_branch_unlikely(&tcp_tx_delay_enabled)) {
2695 u32 delay = (sk->sk_state == TCP_TIME_WAIT) ?
2696 tcp_twsk(sk)->tw_tx_delay : tcp_sk(sk)->tcp_tx_delay;
2697
2698 return tcp_clock_ns() + (u64)delay * NSEC_PER_USEC;
2699 }
2700 return 0;
2701}
2702
2703static inline int tcp_parse_auth_options(const struct tcphdr *th,
2704 const u8 **md5_hash, const struct tcp_ao_hdr **aoh)
2705{
2706 const u8 *md5_tmp, *ao_tmp;
2707 int ret;
2708
2709 ret = tcp_do_parse_auth_options(th, &md5_tmp, &ao_tmp);
2710 if (ret)
2711 return ret;
2712
2713 if (md5_hash)
2714 *md5_hash = md5_tmp;
2715
2716 if (aoh) {
2717 if (!ao_tmp)
2718 *aoh = NULL;
2719 else
2720 *aoh = (struct tcp_ao_hdr *)(ao_tmp - 2);
2721 }
2722
2723 return 0;
2724}
2725
2726static inline bool tcp_ao_required(struct sock *sk, const void *saddr,
2727 int family, int l3index, bool stat_inc)
2728{
2729#ifdef CONFIG_TCP_AO
2730 struct tcp_ao_info *ao_info;
2731 struct tcp_ao_key *ao_key;
2732
2733 if (!static_branch_unlikely(&tcp_ao_needed.key))
2734 return false;
2735
2736 ao_info = rcu_dereference_check(tcp_sk(sk)->ao_info,
2737 lockdep_sock_is_held(sk));
2738 if (!ao_info)
2739 return false;
2740
2741 ao_key = tcp_ao_do_lookup(sk, l3index, saddr, family, -1, -1);
2742 if (ao_info->ao_required || ao_key) {
2743 if (stat_inc) {
2744 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAOREQUIRED);
2745 atomic64_inc(&ao_info->counters.ao_required);
2746 }
2747 return true;
2748 }
2749#endif
2750 return false;
2751}
2752
2753/* Called with rcu_read_lock() */
2754static inline enum skb_drop_reason
2755tcp_inbound_hash(struct sock *sk, const struct request_sock *req,
2756 const struct sk_buff *skb,
2757 const void *saddr, const void *daddr,
2758 int family, int dif, int sdif)
2759{
2760 const struct tcphdr *th = tcp_hdr(skb);
2761 const struct tcp_ao_hdr *aoh;
2762 const __u8 *md5_location;
2763 int l3index;
2764
2765 /* Invalid option or two times meet any of auth options */
2766 if (tcp_parse_auth_options(th, &md5_location, &aoh)) {
2767 tcp_hash_fail("TCP segment has incorrect auth options set",
2768 family, skb, "");
2769 return SKB_DROP_REASON_TCP_AUTH_HDR;
2770 }
2771
2772 if (req) {
2773 if (tcp_rsk_used_ao(req) != !!aoh) {
2774 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAOBAD);
2775 tcp_hash_fail("TCP connection can't start/end using TCP-AO",
2776 family, skb, "%s",
2777 !aoh ? "missing AO" : "AO signed");
2778 return SKB_DROP_REASON_TCP_AOFAILURE;
2779 }
2780 }
2781
2782 /* sdif set, means packet ingressed via a device
2783 * in an L3 domain and dif is set to the l3mdev
2784 */
2785 l3index = sdif ? dif : 0;
2786
2787 /* Fast path: unsigned segments */
2788 if (likely(!md5_location && !aoh)) {
2789 /* Drop if there's TCP-MD5 or TCP-AO key with any rcvid/sndid
2790 * for the remote peer. On TCP-AO established connection
2791 * the last key is impossible to remove, so there's
2792 * always at least one current_key.
2793 */
2794 if (tcp_ao_required(sk, saddr, family, l3index, true)) {
2795 tcp_hash_fail("AO hash is required, but not found",
2796 family, skb, "L3 index %d", l3index);
2797 return SKB_DROP_REASON_TCP_AONOTFOUND;
2798 }
2799 if (unlikely(tcp_md5_do_lookup(sk, l3index, saddr, family))) {
2800 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
2801 tcp_hash_fail("MD5 Hash not found",
2802 family, skb, "L3 index %d", l3index);
2803 return SKB_DROP_REASON_TCP_MD5NOTFOUND;
2804 }
2805 return SKB_NOT_DROPPED_YET;
2806 }
2807
2808 if (aoh)
2809 return tcp_inbound_ao_hash(sk, skb, family, req, l3index, aoh);
2810
2811 return tcp_inbound_md5_hash(sk, skb, saddr, daddr, family,
2812 l3index, md5_location);
2813}
2814
2815#endif /* _TCP_H */