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/tcp_ao.h>
41#include <net/inet_ecn.h>
42#include <net/dst.h>
43#include <net/mptcp.h>
44#include <net/xfrm.h>
45
46#include <linux/seq_file.h>
47#include <linux/memcontrol.h>
48#include <linux/bpf-cgroup.h>
49#include <linux/siphash.h>
50
51extern struct inet_hashinfo tcp_hashinfo;
52
53DECLARE_PER_CPU(unsigned int, tcp_orphan_count);
54int tcp_orphan_count_sum(void);
55
56DECLARE_PER_CPU(u32, tcp_tw_isn);
57
58void tcp_time_wait(struct sock *sk, int state, int timeo);
59
60#define MAX_TCP_HEADER L1_CACHE_ALIGN(128 + MAX_HEADER)
61#define MAX_TCP_OPTION_SPACE 40
62#define TCP_MIN_SND_MSS 48
63#define TCP_MIN_GSO_SIZE (TCP_MIN_SND_MSS - MAX_TCP_OPTION_SPACE)
64
65/*
66 * Never offer a window over 32767 without using window scaling. Some
67 * poor stacks do signed 16bit maths!
68 */
69#define MAX_TCP_WINDOW 32767U
70
71/* Minimal accepted MSS. It is (60+60+8) - (20+20). */
72#define TCP_MIN_MSS 88U
73
74/* The initial MTU to use for probing */
75#define TCP_BASE_MSS 1024
76
77/* probing interval, default to 10 minutes as per RFC4821 */
78#define TCP_PROBE_INTERVAL 600
79
80/* Specify interval when tcp mtu probing will stop */
81#define TCP_PROBE_THRESHOLD 8
82
83/* After receiving this amount of duplicate ACKs fast retransmit starts. */
84#define TCP_FASTRETRANS_THRESH 3
85
86/* Maximal number of ACKs sent quickly to accelerate slow-start. */
87#define TCP_MAX_QUICKACKS 16U
88
89/* Maximal number of window scale according to RFC1323 */
90#define TCP_MAX_WSCALE 14U
91
92/* urg_data states */
93#define TCP_URG_VALID 0x0100
94#define TCP_URG_NOTYET 0x0200
95#define TCP_URG_READ 0x0400
96
97#define TCP_RETR1 3 /*
98 * This is how many retries it does before it
99 * tries to figure out if the gateway is
100 * down. Minimal RFC value is 3; it corresponds
101 * to ~3sec-8min depending on RTO.
102 */
103
104#define TCP_RETR2 15 /*
105 * This should take at least
106 * 90 minutes to time out.
107 * RFC1122 says that the limit is 100 sec.
108 * 15 is ~13-30min depending on RTO.
109 */
110
111#define TCP_SYN_RETRIES 6 /* This is how many retries are done
112 * when active opening a connection.
113 * RFC1122 says the minimum retry MUST
114 * be at least 180secs. Nevertheless
115 * this value is corresponding to
116 * 63secs of retransmission with the
117 * current initial RTO.
118 */
119
120#define TCP_SYNACK_RETRIES 5 /* This is how may retries are done
121 * when passive opening a connection.
122 * This is corresponding to 31secs of
123 * retransmission with the current
124 * initial RTO.
125 */
126
127#define TCP_TIMEWAIT_LEN (60*HZ) /* how long to wait to destroy TIME-WAIT
128 * state, about 60 seconds */
129#define TCP_FIN_TIMEOUT TCP_TIMEWAIT_LEN
130 /* BSD style FIN_WAIT2 deadlock breaker.
131 * It used to be 3min, new value is 60sec,
132 * to combine FIN-WAIT-2 timeout with
133 * TIME-WAIT timer.
134 */
135#define TCP_FIN_TIMEOUT_MAX (120 * HZ) /* max TCP_LINGER2 value (two minutes) */
136
137#define TCP_DELACK_MAX ((unsigned)(HZ/5)) /* maximal time to delay before sending an ACK */
138static_assert((1 << ATO_BITS) > TCP_DELACK_MAX);
139
140#if HZ >= 100
141#define TCP_DELACK_MIN ((unsigned)(HZ/25)) /* minimal time to delay before sending an ACK */
142#define TCP_ATO_MIN ((unsigned)(HZ/25))
143#else
144#define TCP_DELACK_MIN 4U
145#define TCP_ATO_MIN 4U
146#endif
147#define TCP_RTO_MAX ((unsigned)(120*HZ))
148#define TCP_RTO_MIN ((unsigned)(HZ/5))
149#define TCP_TIMEOUT_MIN (2U) /* Min timeout for TCP timers in jiffies */
150
151#define TCP_TIMEOUT_MIN_US (2*USEC_PER_MSEC) /* Min TCP timeout in microsecs */
152
153#define TCP_TIMEOUT_INIT ((unsigned)(1*HZ)) /* RFC6298 2.1 initial RTO value */
154#define TCP_TIMEOUT_FALLBACK ((unsigned)(3*HZ)) /* RFC 1122 initial RTO value, now
155 * used as a fallback RTO for the
156 * initial data transmission if no
157 * valid RTT sample has been acquired,
158 * most likely due to retrans in 3WHS.
159 */
160
161#define TCP_RESOURCE_PROBE_INTERVAL ((unsigned)(HZ/2U)) /* Maximal interval between probes
162 * for local resources.
163 */
164#define TCP_KEEPALIVE_TIME (120*60*HZ) /* two hours */
165#define TCP_KEEPALIVE_PROBES 9 /* Max of 9 keepalive probes */
166#define TCP_KEEPALIVE_INTVL (75*HZ)
167
168#define MAX_TCP_KEEPIDLE 32767
169#define MAX_TCP_KEEPINTVL 32767
170#define MAX_TCP_KEEPCNT 127
171#define MAX_TCP_SYNCNT 127
172
173/* Ensure that TCP PAWS checks are relaxed after ~2147 seconds
174 * to avoid overflows. This assumes a clock smaller than 1 Mhz.
175 * Default clock is 1 Khz, tcp_usec_ts uses 1 Mhz.
176 */
177#define TCP_PAWS_WRAP (INT_MAX / USEC_PER_SEC)
178
179#define TCP_PAWS_MSL 60 /* Per-host timestamps are invalidated
180 * after this time. It should be equal
181 * (or greater than) TCP_TIMEWAIT_LEN
182 * to provide reliability equal to one
183 * provided by timewait state.
184 */
185#define TCP_PAWS_WINDOW 1 /* Replay window for per-host
186 * timestamps. It must be less than
187 * minimal timewait lifetime.
188 */
189/*
190 * TCP option
191 */
192
193#define TCPOPT_NOP 1 /* Padding */
194#define TCPOPT_EOL 0 /* End of options */
195#define TCPOPT_MSS 2 /* Segment size negotiating */
196#define TCPOPT_WINDOW 3 /* Window scaling */
197#define TCPOPT_SACK_PERM 4 /* SACK Permitted */
198#define TCPOPT_SACK 5 /* SACK Block */
199#define TCPOPT_TIMESTAMP 8 /* Better RTT estimations/PAWS */
200#define TCPOPT_MD5SIG 19 /* MD5 Signature (RFC2385) */
201#define TCPOPT_AO 29 /* Authentication Option (RFC5925) */
202#define TCPOPT_MPTCP 30 /* Multipath TCP (RFC6824) */
203#define TCPOPT_FASTOPEN 34 /* Fast open (RFC7413) */
204#define TCPOPT_EXP 254 /* Experimental */
205/* Magic number to be after the option value for sharing TCP
206 * experimental options. See draft-ietf-tcpm-experimental-options-00.txt
207 */
208#define TCPOPT_FASTOPEN_MAGIC 0xF989
209#define TCPOPT_SMC_MAGIC 0xE2D4C3D9
210
211/*
212 * TCP option lengths
213 */
214
215#define TCPOLEN_MSS 4
216#define TCPOLEN_WINDOW 3
217#define TCPOLEN_SACK_PERM 2
218#define TCPOLEN_TIMESTAMP 10
219#define TCPOLEN_MD5SIG 18
220#define TCPOLEN_FASTOPEN_BASE 2
221#define TCPOLEN_EXP_FASTOPEN_BASE 4
222#define TCPOLEN_EXP_SMC_BASE 6
223
224/* But this is what stacks really send out. */
225#define TCPOLEN_TSTAMP_ALIGNED 12
226#define TCPOLEN_WSCALE_ALIGNED 4
227#define TCPOLEN_SACKPERM_ALIGNED 4
228#define TCPOLEN_SACK_BASE 2
229#define TCPOLEN_SACK_BASE_ALIGNED 4
230#define TCPOLEN_SACK_PERBLOCK 8
231#define TCPOLEN_MD5SIG_ALIGNED 20
232#define TCPOLEN_MSS_ALIGNED 4
233#define TCPOLEN_EXP_SMC_BASE_ALIGNED 8
234
235/* Flags in tp->nonagle */
236#define TCP_NAGLE_OFF 1 /* Nagle's algo is disabled */
237#define TCP_NAGLE_CORK 2 /* Socket is corked */
238#define TCP_NAGLE_PUSH 4 /* Cork is overridden for already queued data */
239
240/* TCP thin-stream limits */
241#define TCP_THIN_LINEAR_RETRIES 6 /* After 6 linear retries, do exp. backoff */
242
243/* TCP initial congestion window as per rfc6928 */
244#define TCP_INIT_CWND 10
245
246/* Bit Flags for sysctl_tcp_fastopen */
247#define TFO_CLIENT_ENABLE 1
248#define TFO_SERVER_ENABLE 2
249#define TFO_CLIENT_NO_COOKIE 4 /* Data in SYN w/o cookie option */
250
251/* Accept SYN data w/o any cookie option */
252#define TFO_SERVER_COOKIE_NOT_REQD 0x200
253
254/* Force enable TFO on all listeners, i.e., not requiring the
255 * TCP_FASTOPEN socket option.
256 */
257#define TFO_SERVER_WO_SOCKOPT1 0x400
258
259
260/* sysctl variables for tcp */
261extern int sysctl_tcp_max_orphans;
262extern long sysctl_tcp_mem[3];
263
264#define TCP_RACK_LOSS_DETECTION 0x1 /* Use RACK to detect losses */
265#define TCP_RACK_STATIC_REO_WND 0x2 /* Use static RACK reo wnd */
266#define TCP_RACK_NO_DUPTHRESH 0x4 /* Do not use DUPACK threshold in RACK */
267
268extern atomic_long_t tcp_memory_allocated;
269DECLARE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
270
271extern struct percpu_counter tcp_sockets_allocated;
272extern unsigned long tcp_memory_pressure;
273
274/* optimized version of sk_under_memory_pressure() for TCP sockets */
275static inline bool tcp_under_memory_pressure(const struct sock *sk)
276{
277 if (mem_cgroup_sockets_enabled && sk->sk_memcg &&
278 mem_cgroup_under_socket_pressure(sk->sk_memcg))
279 return true;
280
281 return READ_ONCE(tcp_memory_pressure);
282}
283/*
284 * The next routines deal with comparing 32 bit unsigned ints
285 * and worry about wraparound (automatic with unsigned arithmetic).
286 */
287
288static inline bool before(__u32 seq1, __u32 seq2)
289{
290 return (__s32)(seq1-seq2) < 0;
291}
292#define after(seq2, seq1) before(seq1, seq2)
293
294/* is s2<=s1<=s3 ? */
295static inline bool between(__u32 seq1, __u32 seq2, __u32 seq3)
296{
297 return seq3 - seq2 >= seq1 - seq2;
298}
299
300static inline void tcp_wmem_free_skb(struct sock *sk, struct sk_buff *skb)
301{
302 sk_wmem_queued_add(sk, -skb->truesize);
303 if (!skb_zcopy_pure(skb))
304 sk_mem_uncharge(sk, skb->truesize);
305 else
306 sk_mem_uncharge(sk, SKB_TRUESIZE(skb_end_offset(skb)));
307 __kfree_skb(skb);
308}
309
310void sk_forced_mem_schedule(struct sock *sk, int size);
311
312bool tcp_check_oom(const struct sock *sk, int shift);
313
314
315extern struct proto tcp_prot;
316
317#define TCP_INC_STATS(net, field) SNMP_INC_STATS((net)->mib.tcp_statistics, field)
318#define __TCP_INC_STATS(net, field) __SNMP_INC_STATS((net)->mib.tcp_statistics, field)
319#define TCP_DEC_STATS(net, field) SNMP_DEC_STATS((net)->mib.tcp_statistics, field)
320#define TCP_ADD_STATS(net, field, val) SNMP_ADD_STATS((net)->mib.tcp_statistics, field, val)
321
322void tcp_tasklet_init(void);
323
324int tcp_v4_err(struct sk_buff *skb, u32);
325
326void tcp_shutdown(struct sock *sk, int how);
327
328int tcp_v4_early_demux(struct sk_buff *skb);
329int tcp_v4_rcv(struct sk_buff *skb);
330
331void tcp_remove_empty_skb(struct sock *sk);
332int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size);
333int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size);
334int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
335 size_t size, struct ubuf_info *uarg);
336void tcp_splice_eof(struct socket *sock);
337int tcp_send_mss(struct sock *sk, int *size_goal, int flags);
338int tcp_wmem_schedule(struct sock *sk, int copy);
339void tcp_push(struct sock *sk, int flags, int mss_now, int nonagle,
340 int size_goal);
341void tcp_release_cb(struct sock *sk);
342void tcp_wfree(struct sk_buff *skb);
343void tcp_write_timer_handler(struct sock *sk);
344void tcp_delack_timer_handler(struct sock *sk);
345int tcp_ioctl(struct sock *sk, int cmd, int *karg);
346enum skb_drop_reason tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb);
347void tcp_rcv_established(struct sock *sk, struct sk_buff *skb);
348void tcp_rcv_space_adjust(struct sock *sk);
349int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp);
350void tcp_twsk_destructor(struct sock *sk);
351void tcp_twsk_purge(struct list_head *net_exit_list);
352ssize_t tcp_splice_read(struct socket *sk, loff_t *ppos,
353 struct pipe_inode_info *pipe, size_t len,
354 unsigned int flags);
355struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
356 bool force_schedule);
357
358static inline void tcp_dec_quickack_mode(struct sock *sk)
359{
360 struct inet_connection_sock *icsk = inet_csk(sk);
361
362 if (icsk->icsk_ack.quick) {
363 /* How many ACKs S/ACKing new data have we sent? */
364 const unsigned int pkts = inet_csk_ack_scheduled(sk) ? 1 : 0;
365
366 if (pkts >= icsk->icsk_ack.quick) {
367 icsk->icsk_ack.quick = 0;
368 /* Leaving quickack mode we deflate ATO. */
369 icsk->icsk_ack.ato = TCP_ATO_MIN;
370 } else
371 icsk->icsk_ack.quick -= pkts;
372 }
373}
374
375#define TCP_ECN_OK 1
376#define TCP_ECN_QUEUE_CWR 2
377#define TCP_ECN_DEMAND_CWR 4
378#define TCP_ECN_SEEN 8
379
380enum tcp_tw_status {
381 TCP_TW_SUCCESS = 0,
382 TCP_TW_RST = 1,
383 TCP_TW_ACK = 2,
384 TCP_TW_SYN = 3
385};
386
387
388enum tcp_tw_status tcp_timewait_state_process(struct inet_timewait_sock *tw,
389 struct sk_buff *skb,
390 const struct tcphdr *th,
391 u32 *tw_isn);
392struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
393 struct request_sock *req, bool fastopen,
394 bool *lost_race);
395enum skb_drop_reason tcp_child_process(struct sock *parent, struct sock *child,
396 struct sk_buff *skb);
397void tcp_enter_loss(struct sock *sk);
398void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag);
399void tcp_clear_retrans(struct tcp_sock *tp);
400void tcp_update_metrics(struct sock *sk);
401void tcp_init_metrics(struct sock *sk);
402void tcp_metrics_init(void);
403bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst);
404void __tcp_close(struct sock *sk, long timeout);
405void tcp_close(struct sock *sk, long timeout);
406void tcp_init_sock(struct sock *sk);
407void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb);
408__poll_t tcp_poll(struct file *file, struct socket *sock,
409 struct poll_table_struct *wait);
410int do_tcp_getsockopt(struct sock *sk, int level,
411 int optname, sockptr_t optval, sockptr_t optlen);
412int tcp_getsockopt(struct sock *sk, int level, int optname,
413 char __user *optval, int __user *optlen);
414bool tcp_bpf_bypass_getsockopt(int level, int optname);
415int do_tcp_setsockopt(struct sock *sk, int level, int optname,
416 sockptr_t optval, unsigned int optlen);
417int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
418 unsigned int optlen);
419void tcp_set_keepalive(struct sock *sk, int val);
420void tcp_syn_ack_timeout(const struct request_sock *req);
421int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
422 int flags, int *addr_len);
423int tcp_set_rcvlowat(struct sock *sk, int val);
424int tcp_set_window_clamp(struct sock *sk, int val);
425void tcp_update_recv_tstamps(struct sk_buff *skb,
426 struct scm_timestamping_internal *tss);
427void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
428 struct scm_timestamping_internal *tss);
429void tcp_data_ready(struct sock *sk);
430#ifdef CONFIG_MMU
431int tcp_mmap(struct file *file, struct socket *sock,
432 struct vm_area_struct *vma);
433#endif
434void tcp_parse_options(const struct net *net, const struct sk_buff *skb,
435 struct tcp_options_received *opt_rx,
436 int estab, struct tcp_fastopen_cookie *foc);
437
438/*
439 * BPF SKB-less helpers
440 */
441u16 tcp_v4_get_syncookie(struct sock *sk, struct iphdr *iph,
442 struct tcphdr *th, u32 *cookie);
443u16 tcp_v6_get_syncookie(struct sock *sk, struct ipv6hdr *iph,
444 struct tcphdr *th, u32 *cookie);
445u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss);
446u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
447 const struct tcp_request_sock_ops *af_ops,
448 struct sock *sk, struct tcphdr *th);
449/*
450 * TCP v4 functions exported for the inet6 API
451 */
452
453void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb);
454void tcp_v4_mtu_reduced(struct sock *sk);
455void tcp_req_err(struct sock *sk, u32 seq, bool abort);
456void tcp_ld_RTO_revert(struct sock *sk, u32 seq);
457int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb);
458struct sock *tcp_create_openreq_child(const struct sock *sk,
459 struct request_sock *req,
460 struct sk_buff *skb);
461void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst);
462struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
463 struct request_sock *req,
464 struct dst_entry *dst,
465 struct request_sock *req_unhash,
466 bool *own_req);
467int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb);
468int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len);
469int tcp_connect(struct sock *sk);
470enum tcp_synack_type {
471 TCP_SYNACK_NORMAL,
472 TCP_SYNACK_FASTOPEN,
473 TCP_SYNACK_COOKIE,
474};
475struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
476 struct request_sock *req,
477 struct tcp_fastopen_cookie *foc,
478 enum tcp_synack_type synack_type,
479 struct sk_buff *syn_skb);
480int tcp_disconnect(struct sock *sk, int flags);
481
482void tcp_finish_connect(struct sock *sk, struct sk_buff *skb);
483int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size);
484void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb);
485
486/* From syncookies.c */
487struct sock *tcp_get_cookie_sock(struct sock *sk, struct sk_buff *skb,
488 struct request_sock *req,
489 struct dst_entry *dst);
490int __cookie_v4_check(const struct iphdr *iph, const struct tcphdr *th);
491struct sock *cookie_v4_check(struct sock *sk, struct sk_buff *skb);
492struct request_sock *cookie_tcp_reqsk_alloc(const struct request_sock_ops *ops,
493 struct sock *sk, struct sk_buff *skb,
494 struct tcp_options_received *tcp_opt,
495 int mss, u32 tsoff);
496
497#if IS_ENABLED(CONFIG_BPF)
498struct bpf_tcp_req_attrs {
499 u32 rcv_tsval;
500 u32 rcv_tsecr;
501 u16 mss;
502 u8 rcv_wscale;
503 u8 snd_wscale;
504 u8 ecn_ok;
505 u8 wscale_ok;
506 u8 sack_ok;
507 u8 tstamp_ok;
508 u8 usec_ts_ok;
509 u8 reserved[3];
510};
511#endif
512
513#ifdef CONFIG_SYN_COOKIES
514
515/* Syncookies use a monotonic timer which increments every 60 seconds.
516 * This counter is used both as a hash input and partially encoded into
517 * the cookie value. A cookie is only validated further if the delta
518 * between the current counter value and the encoded one is less than this,
519 * i.e. a sent cookie is valid only at most for 2*60 seconds (or less if
520 * the counter advances immediately after a cookie is generated).
521 */
522#define MAX_SYNCOOKIE_AGE 2
523#define TCP_SYNCOOKIE_PERIOD (60 * HZ)
524#define TCP_SYNCOOKIE_VALID (MAX_SYNCOOKIE_AGE * TCP_SYNCOOKIE_PERIOD)
525
526/* syncookies: remember time of last synqueue overflow
527 * But do not dirty this field too often (once per second is enough)
528 * It is racy as we do not hold a lock, but race is very minor.
529 */
530static inline void tcp_synq_overflow(const struct sock *sk)
531{
532 unsigned int last_overflow;
533 unsigned int now = jiffies;
534
535 if (sk->sk_reuseport) {
536 struct sock_reuseport *reuse;
537
538 reuse = rcu_dereference(sk->sk_reuseport_cb);
539 if (likely(reuse)) {
540 last_overflow = READ_ONCE(reuse->synq_overflow_ts);
541 if (!time_between32(now, last_overflow,
542 last_overflow + HZ))
543 WRITE_ONCE(reuse->synq_overflow_ts, now);
544 return;
545 }
546 }
547
548 last_overflow = READ_ONCE(tcp_sk(sk)->rx_opt.ts_recent_stamp);
549 if (!time_between32(now, last_overflow, last_overflow + HZ))
550 WRITE_ONCE(tcp_sk_rw(sk)->rx_opt.ts_recent_stamp, now);
551}
552
553/* syncookies: no recent synqueue overflow on this listening socket? */
554static inline bool tcp_synq_no_recent_overflow(const struct sock *sk)
555{
556 unsigned int last_overflow;
557 unsigned int now = jiffies;
558
559 if (sk->sk_reuseport) {
560 struct sock_reuseport *reuse;
561
562 reuse = rcu_dereference(sk->sk_reuseport_cb);
563 if (likely(reuse)) {
564 last_overflow = READ_ONCE(reuse->synq_overflow_ts);
565 return !time_between32(now, last_overflow - HZ,
566 last_overflow +
567 TCP_SYNCOOKIE_VALID);
568 }
569 }
570
571 last_overflow = READ_ONCE(tcp_sk(sk)->rx_opt.ts_recent_stamp);
572
573 /* If last_overflow <= jiffies <= last_overflow + TCP_SYNCOOKIE_VALID,
574 * then we're under synflood. However, we have to use
575 * 'last_overflow - HZ' as lower bound. That's because a concurrent
576 * tcp_synq_overflow() could update .ts_recent_stamp after we read
577 * jiffies but before we store .ts_recent_stamp into last_overflow,
578 * which could lead to rejecting a valid syncookie.
579 */
580 return !time_between32(now, last_overflow - HZ,
581 last_overflow + TCP_SYNCOOKIE_VALID);
582}
583
584static inline u32 tcp_cookie_time(void)
585{
586 u64 val = get_jiffies_64();
587
588 do_div(val, TCP_SYNCOOKIE_PERIOD);
589 return val;
590}
591
592/* Convert one nsec 64bit timestamp to ts (ms or usec resolution) */
593static inline u64 tcp_ns_to_ts(bool usec_ts, u64 val)
594{
595 if (usec_ts)
596 return div_u64(val, NSEC_PER_USEC);
597
598 return div_u64(val, NSEC_PER_MSEC);
599}
600
601u32 __cookie_v4_init_sequence(const struct iphdr *iph, const struct tcphdr *th,
602 u16 *mssp);
603__u32 cookie_v4_init_sequence(const struct sk_buff *skb, __u16 *mss);
604u64 cookie_init_timestamp(struct request_sock *req, u64 now);
605bool cookie_timestamp_decode(const struct net *net,
606 struct tcp_options_received *opt);
607
608static inline bool cookie_ecn_ok(const struct net *net, const struct dst_entry *dst)
609{
610 return READ_ONCE(net->ipv4.sysctl_tcp_ecn) ||
611 dst_feature(dst, RTAX_FEATURE_ECN);
612}
613
614#if IS_ENABLED(CONFIG_BPF)
615static inline bool cookie_bpf_ok(struct sk_buff *skb)
616{
617 return skb->sk;
618}
619
620struct request_sock *cookie_bpf_check(struct sock *sk, struct sk_buff *skb);
621#else
622static inline bool cookie_bpf_ok(struct sk_buff *skb)
623{
624 return false;
625}
626
627static inline struct request_sock *cookie_bpf_check(struct net *net, struct sock *sk,
628 struct sk_buff *skb)
629{
630 return NULL;
631}
632#endif
633
634/* From net/ipv6/syncookies.c */
635int __cookie_v6_check(const struct ipv6hdr *iph, const struct tcphdr *th);
636struct sock *cookie_v6_check(struct sock *sk, struct sk_buff *skb);
637
638u32 __cookie_v6_init_sequence(const struct ipv6hdr *iph,
639 const struct tcphdr *th, u16 *mssp);
640__u32 cookie_v6_init_sequence(const struct sk_buff *skb, __u16 *mss);
641#endif
642/* tcp_output.c */
643
644void tcp_skb_entail(struct sock *sk, struct sk_buff *skb);
645void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb);
646void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
647 int nonagle);
648int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs);
649int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs);
650void tcp_retransmit_timer(struct sock *sk);
651void tcp_xmit_retransmit_queue(struct sock *);
652void tcp_simple_retransmit(struct sock *);
653void tcp_enter_recovery(struct sock *sk, bool ece_ack);
654int tcp_trim_head(struct sock *, struct sk_buff *, u32);
655enum tcp_queue {
656 TCP_FRAG_IN_WRITE_QUEUE,
657 TCP_FRAG_IN_RTX_QUEUE,
658};
659int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
660 struct sk_buff *skb, u32 len,
661 unsigned int mss_now, gfp_t gfp);
662
663void tcp_send_probe0(struct sock *);
664int tcp_write_wakeup(struct sock *, int mib);
665void tcp_send_fin(struct sock *sk);
666void tcp_send_active_reset(struct sock *sk, gfp_t priority,
667 enum sk_rst_reason reason);
668int tcp_send_synack(struct sock *);
669void tcp_push_one(struct sock *, unsigned int mss_now);
670void __tcp_send_ack(struct sock *sk, u32 rcv_nxt);
671void tcp_send_ack(struct sock *sk);
672void tcp_send_delayed_ack(struct sock *sk);
673void tcp_send_loss_probe(struct sock *sk);
674bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto);
675void tcp_skb_collapse_tstamp(struct sk_buff *skb,
676 const struct sk_buff *next_skb);
677
678/* tcp_input.c */
679void tcp_rearm_rto(struct sock *sk);
680void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req);
681void tcp_done_with_error(struct sock *sk, int err);
682void tcp_reset(struct sock *sk, struct sk_buff *skb);
683void tcp_fin(struct sock *sk);
684void tcp_check_space(struct sock *sk);
685void tcp_sack_compress_send_ack(struct sock *sk);
686
687static inline void tcp_cleanup_skb(struct sk_buff *skb)
688{
689 skb_dst_drop(skb);
690 secpath_reset(skb);
691}
692
693static inline void tcp_add_receive_queue(struct sock *sk, struct sk_buff *skb)
694{
695 DEBUG_NET_WARN_ON_ONCE(skb_dst(skb));
696 DEBUG_NET_WARN_ON_ONCE(secpath_exists(skb));
697 __skb_queue_tail(&sk->sk_receive_queue, skb);
698}
699
700/* tcp_timer.c */
701void tcp_init_xmit_timers(struct sock *);
702static inline void tcp_clear_xmit_timers(struct sock *sk)
703{
704 if (hrtimer_try_to_cancel(&tcp_sk(sk)->pacing_timer) == 1)
705 __sock_put(sk);
706
707 if (hrtimer_try_to_cancel(&tcp_sk(sk)->compressed_ack_timer) == 1)
708 __sock_put(sk);
709
710 inet_csk_clear_xmit_timers(sk);
711}
712
713unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu);
714unsigned int tcp_current_mss(struct sock *sk);
715u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when);
716
717/* Bound MSS / TSO packet size with the half of the window */
718static inline int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
719{
720 int cutoff;
721
722 /* When peer uses tiny windows, there is no use in packetizing
723 * to sub-MSS pieces for the sake of SWS or making sure there
724 * are enough packets in the pipe for fast recovery.
725 *
726 * On the other hand, for extremely large MSS devices, handling
727 * smaller than MSS windows in this way does make sense.
728 */
729 if (tp->max_window > TCP_MSS_DEFAULT)
730 cutoff = (tp->max_window >> 1);
731 else
732 cutoff = tp->max_window;
733
734 if (cutoff && pktsize > cutoff)
735 return max_t(int, cutoff, 68U - tp->tcp_header_len);
736 else
737 return pktsize;
738}
739
740/* tcp.c */
741void tcp_get_info(struct sock *, struct tcp_info *);
742
743/* Read 'sendfile()'-style from a TCP socket */
744int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
745 sk_read_actor_t recv_actor);
746int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor);
747struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off);
748void tcp_read_done(struct sock *sk, size_t len);
749
750void tcp_initialize_rcv_mss(struct sock *sk);
751
752int tcp_mtu_to_mss(struct sock *sk, int pmtu);
753int tcp_mss_to_mtu(struct sock *sk, int mss);
754void tcp_mtup_init(struct sock *sk);
755
756static inline void tcp_bound_rto(struct sock *sk)
757{
758 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
759 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
760}
761
762static inline u32 __tcp_set_rto(const struct tcp_sock *tp)
763{
764 return usecs_to_jiffies((tp->srtt_us >> 3) + tp->rttvar_us);
765}
766
767static inline void __tcp_fast_path_on(struct tcp_sock *tp, u32 snd_wnd)
768{
769 /* mptcp hooks are only on the slow path */
770 if (sk_is_mptcp((struct sock *)tp))
771 return;
772
773 tp->pred_flags = htonl((tp->tcp_header_len << 26) |
774 ntohl(TCP_FLAG_ACK) |
775 snd_wnd);
776}
777
778static inline void tcp_fast_path_on(struct tcp_sock *tp)
779{
780 __tcp_fast_path_on(tp, tp->snd_wnd >> tp->rx_opt.snd_wscale);
781}
782
783static inline void tcp_fast_path_check(struct sock *sk)
784{
785 struct tcp_sock *tp = tcp_sk(sk);
786
787 if (RB_EMPTY_ROOT(&tp->out_of_order_queue) &&
788 tp->rcv_wnd &&
789 atomic_read(&sk->sk_rmem_alloc) < sk->sk_rcvbuf &&
790 !tp->urg_data)
791 tcp_fast_path_on(tp);
792}
793
794u32 tcp_delack_max(const struct sock *sk);
795
796/* Compute the actual rto_min value */
797static inline u32 tcp_rto_min(const struct sock *sk)
798{
799 const struct dst_entry *dst = __sk_dst_get(sk);
800 u32 rto_min = inet_csk(sk)->icsk_rto_min;
801
802 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
803 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
804 return rto_min;
805}
806
807static inline u32 tcp_rto_min_us(const struct sock *sk)
808{
809 return jiffies_to_usecs(tcp_rto_min(sk));
810}
811
812static inline bool tcp_ca_dst_locked(const struct dst_entry *dst)
813{
814 return dst_metric_locked(dst, RTAX_CC_ALGO);
815}
816
817/* Minimum RTT in usec. ~0 means not available. */
818static inline u32 tcp_min_rtt(const struct tcp_sock *tp)
819{
820 return minmax_get(&tp->rtt_min);
821}
822
823/* Compute the actual receive window we are currently advertising.
824 * Rcv_nxt can be after the window if our peer push more data
825 * than the offered window.
826 */
827static inline u32 tcp_receive_window(const struct tcp_sock *tp)
828{
829 s32 win = tp->rcv_wup + tp->rcv_wnd - tp->rcv_nxt;
830
831 if (win < 0)
832 win = 0;
833 return (u32) win;
834}
835
836/* Choose a new window, without checks for shrinking, and without
837 * scaling applied to the result. The caller does these things
838 * if necessary. This is a "raw" window selection.
839 */
840u32 __tcp_select_window(struct sock *sk);
841
842void tcp_send_window_probe(struct sock *sk);
843
844/* TCP uses 32bit jiffies to save some space.
845 * Note that this is different from tcp_time_stamp, which
846 * historically has been the same until linux-4.13.
847 */
848#define tcp_jiffies32 ((u32)jiffies)
849
850/*
851 * Deliver a 32bit value for TCP timestamp option (RFC 7323)
852 * It is no longer tied to jiffies, but to 1 ms clock.
853 * Note: double check if you want to use tcp_jiffies32 instead of this.
854 */
855#define TCP_TS_HZ 1000
856
857static inline u64 tcp_clock_ns(void)
858{
859 return ktime_get_ns();
860}
861
862static inline u64 tcp_clock_us(void)
863{
864 return div_u64(tcp_clock_ns(), NSEC_PER_USEC);
865}
866
867static inline u64 tcp_clock_ms(void)
868{
869 return div_u64(tcp_clock_ns(), NSEC_PER_MSEC);
870}
871
872/* TCP Timestamp included in TS option (RFC 1323) can either use ms
873 * or usec resolution. Each socket carries a flag to select one or other
874 * resolution, as the route attribute could change anytime.
875 * Each flow must stick to initial resolution.
876 */
877static inline u32 tcp_clock_ts(bool usec_ts)
878{
879 return usec_ts ? tcp_clock_us() : tcp_clock_ms();
880}
881
882static inline u32 tcp_time_stamp_ms(const struct tcp_sock *tp)
883{
884 return div_u64(tp->tcp_mstamp, USEC_PER_MSEC);
885}
886
887static inline u32 tcp_time_stamp_ts(const struct tcp_sock *tp)
888{
889 if (tp->tcp_usec_ts)
890 return tp->tcp_mstamp;
891 return tcp_time_stamp_ms(tp);
892}
893
894void tcp_mstamp_refresh(struct tcp_sock *tp);
895
896static inline u32 tcp_stamp_us_delta(u64 t1, u64 t0)
897{
898 return max_t(s64, t1 - t0, 0);
899}
900
901/* provide the departure time in us unit */
902static inline u64 tcp_skb_timestamp_us(const struct sk_buff *skb)
903{
904 return div_u64(skb->skb_mstamp_ns, NSEC_PER_USEC);
905}
906
907/* Provide skb TSval in usec or ms unit */
908static inline u32 tcp_skb_timestamp_ts(bool usec_ts, const struct sk_buff *skb)
909{
910 if (usec_ts)
911 return tcp_skb_timestamp_us(skb);
912
913 return div_u64(skb->skb_mstamp_ns, NSEC_PER_MSEC);
914}
915
916static inline u32 tcp_tw_tsval(const struct tcp_timewait_sock *tcptw)
917{
918 return tcp_clock_ts(tcptw->tw_sk.tw_usec_ts) + tcptw->tw_ts_offset;
919}
920
921static inline u32 tcp_rsk_tsval(const struct tcp_request_sock *treq)
922{
923 return tcp_clock_ts(treq->req_usec_ts) + treq->ts_off;
924}
925
926#define tcp_flag_byte(th) (((u_int8_t *)th)[13])
927
928#define TCPHDR_FIN 0x01
929#define TCPHDR_SYN 0x02
930#define TCPHDR_RST 0x04
931#define TCPHDR_PSH 0x08
932#define TCPHDR_ACK 0x10
933#define TCPHDR_URG 0x20
934#define TCPHDR_ECE 0x40
935#define TCPHDR_CWR 0x80
936
937#define TCPHDR_SYN_ECN (TCPHDR_SYN | TCPHDR_ECE | TCPHDR_CWR)
938
939/* State flags for sacked in struct tcp_skb_cb */
940enum tcp_skb_cb_sacked_flags {
941 TCPCB_SACKED_ACKED = (1 << 0), /* SKB ACK'd by a SACK block */
942 TCPCB_SACKED_RETRANS = (1 << 1), /* SKB retransmitted */
943 TCPCB_LOST = (1 << 2), /* SKB is lost */
944 TCPCB_TAGBITS = (TCPCB_SACKED_ACKED | TCPCB_SACKED_RETRANS |
945 TCPCB_LOST), /* All tag bits */
946 TCPCB_REPAIRED = (1 << 4), /* SKB repaired (no skb_mstamp_ns) */
947 TCPCB_EVER_RETRANS = (1 << 7), /* Ever retransmitted frame */
948 TCPCB_RETRANS = (TCPCB_SACKED_RETRANS | TCPCB_EVER_RETRANS |
949 TCPCB_REPAIRED),
950};
951
952/* This is what the send packet queuing engine uses to pass
953 * TCP per-packet control information to the transmission code.
954 * We also store the host-order sequence numbers in here too.
955 * This is 44 bytes if IPV6 is enabled.
956 * If this grows please adjust skbuff.h:skbuff->cb[xxx] size appropriately.
957 */
958struct tcp_skb_cb {
959 __u32 seq; /* Starting sequence number */
960 __u32 end_seq; /* SEQ + FIN + SYN + datalen */
961 union {
962 /* Note :
963 * tcp_gso_segs/size are used in write queue only,
964 * cf tcp_skb_pcount()/tcp_skb_mss()
965 */
966 struct {
967 u16 tcp_gso_segs;
968 u16 tcp_gso_size;
969 };
970 };
971 __u8 tcp_flags; /* TCP header flags. (tcp[13]) */
972
973 __u8 sacked; /* State flags for SACK. */
974 __u8 ip_dsfield; /* IPv4 tos or IPv6 dsfield */
975 __u8 txstamp_ack:1, /* Record TX timestamp for ack? */
976 eor:1, /* Is skb MSG_EOR marked? */
977 has_rxtstamp:1, /* SKB has a RX timestamp */
978 unused:5;
979 __u32 ack_seq; /* Sequence number ACK'd */
980 union {
981 struct {
982#define TCPCB_DELIVERED_CE_MASK ((1U<<20) - 1)
983 /* There is space for up to 24 bytes */
984 __u32 is_app_limited:1, /* cwnd not fully used? */
985 delivered_ce:20,
986 unused:11;
987 /* pkts S/ACKed so far upon tx of skb, incl retrans: */
988 __u32 delivered;
989 /* start of send pipeline phase */
990 u64 first_tx_mstamp;
991 /* when we reached the "delivered" count */
992 u64 delivered_mstamp;
993 } tx; /* only used for outgoing skbs */
994 union {
995 struct inet_skb_parm h4;
996#if IS_ENABLED(CONFIG_IPV6)
997 struct inet6_skb_parm h6;
998#endif
999 } header; /* For incoming skbs */
1000 };
1001};
1002
1003#define TCP_SKB_CB(__skb) ((struct tcp_skb_cb *)&((__skb)->cb[0]))
1004
1005extern const struct inet_connection_sock_af_ops ipv4_specific;
1006
1007#if IS_ENABLED(CONFIG_IPV6)
1008/* This is the variant of inet6_iif() that must be used by TCP,
1009 * as TCP moves IP6CB into a different location in skb->cb[]
1010 */
1011static inline int tcp_v6_iif(const struct sk_buff *skb)
1012{
1013 return TCP_SKB_CB(skb)->header.h6.iif;
1014}
1015
1016static inline int tcp_v6_iif_l3_slave(const struct sk_buff *skb)
1017{
1018 bool l3_slave = ipv6_l3mdev_skb(TCP_SKB_CB(skb)->header.h6.flags);
1019
1020 return l3_slave ? skb->skb_iif : TCP_SKB_CB(skb)->header.h6.iif;
1021}
1022
1023/* TCP_SKB_CB reference means this can not be used from early demux */
1024static inline int tcp_v6_sdif(const struct sk_buff *skb)
1025{
1026#if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV)
1027 if (skb && ipv6_l3mdev_skb(TCP_SKB_CB(skb)->header.h6.flags))
1028 return TCP_SKB_CB(skb)->header.h6.iif;
1029#endif
1030 return 0;
1031}
1032
1033extern const struct inet_connection_sock_af_ops ipv6_specific;
1034
1035INDIRECT_CALLABLE_DECLARE(void tcp_v6_send_check(struct sock *sk, struct sk_buff *skb));
1036INDIRECT_CALLABLE_DECLARE(int tcp_v6_rcv(struct sk_buff *skb));
1037void tcp_v6_early_demux(struct sk_buff *skb);
1038
1039#endif
1040
1041/* TCP_SKB_CB reference means this can not be used from early demux */
1042static inline int tcp_v4_sdif(struct sk_buff *skb)
1043{
1044#if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV)
1045 if (skb && ipv4_l3mdev_skb(TCP_SKB_CB(skb)->header.h4.flags))
1046 return TCP_SKB_CB(skb)->header.h4.iif;
1047#endif
1048 return 0;
1049}
1050
1051/* Due to TSO, an SKB can be composed of multiple actual
1052 * packets. To keep these tracked properly, we use this.
1053 */
1054static inline int tcp_skb_pcount(const struct sk_buff *skb)
1055{
1056 return TCP_SKB_CB(skb)->tcp_gso_segs;
1057}
1058
1059static inline void tcp_skb_pcount_set(struct sk_buff *skb, int segs)
1060{
1061 TCP_SKB_CB(skb)->tcp_gso_segs = segs;
1062}
1063
1064static inline void tcp_skb_pcount_add(struct sk_buff *skb, int segs)
1065{
1066 TCP_SKB_CB(skb)->tcp_gso_segs += segs;
1067}
1068
1069/* This is valid iff skb is in write queue and tcp_skb_pcount() > 1. */
1070static inline int tcp_skb_mss(const struct sk_buff *skb)
1071{
1072 return TCP_SKB_CB(skb)->tcp_gso_size;
1073}
1074
1075static inline bool tcp_skb_can_collapse_to(const struct sk_buff *skb)
1076{
1077 return likely(!TCP_SKB_CB(skb)->eor);
1078}
1079
1080static inline bool tcp_skb_can_collapse(const struct sk_buff *to,
1081 const struct sk_buff *from)
1082{
1083 /* skb_cmp_decrypted() not needed, use tcp_write_collapse_fence() */
1084 return likely(tcp_skb_can_collapse_to(to) &&
1085 mptcp_skb_can_collapse(to, from) &&
1086 skb_pure_zcopy_same(to, from) &&
1087 skb_frags_readable(to) == skb_frags_readable(from));
1088}
1089
1090static inline bool tcp_skb_can_collapse_rx(const struct sk_buff *to,
1091 const struct sk_buff *from)
1092{
1093 return likely(mptcp_skb_can_collapse(to, from) &&
1094 !skb_cmp_decrypted(to, from));
1095}
1096
1097/* Events passed to congestion control interface */
1098enum tcp_ca_event {
1099 CA_EVENT_TX_START, /* first transmit when no packets in flight */
1100 CA_EVENT_CWND_RESTART, /* congestion window restart */
1101 CA_EVENT_COMPLETE_CWR, /* end of congestion recovery */
1102 CA_EVENT_LOSS, /* loss timeout */
1103 CA_EVENT_ECN_NO_CE, /* ECT set, but not CE marked */
1104 CA_EVENT_ECN_IS_CE, /* received CE marked IP packet */
1105};
1106
1107/* Information about inbound ACK, passed to cong_ops->in_ack_event() */
1108enum tcp_ca_ack_event_flags {
1109 CA_ACK_SLOWPATH = (1 << 0), /* In slow path processing */
1110 CA_ACK_WIN_UPDATE = (1 << 1), /* ACK updated window */
1111 CA_ACK_ECE = (1 << 2), /* ECE bit is set on ack */
1112};
1113
1114/*
1115 * Interface for adding new TCP congestion control handlers
1116 */
1117#define TCP_CA_NAME_MAX 16
1118#define TCP_CA_MAX 128
1119#define TCP_CA_BUF_MAX (TCP_CA_NAME_MAX*TCP_CA_MAX)
1120
1121#define TCP_CA_UNSPEC 0
1122
1123/* Algorithm can be set on socket without CAP_NET_ADMIN privileges */
1124#define TCP_CONG_NON_RESTRICTED 0x1
1125/* Requires ECN/ECT set on all packets */
1126#define TCP_CONG_NEEDS_ECN 0x2
1127#define TCP_CONG_MASK (TCP_CONG_NON_RESTRICTED | TCP_CONG_NEEDS_ECN)
1128
1129union tcp_cc_info;
1130
1131struct ack_sample {
1132 u32 pkts_acked;
1133 s32 rtt_us;
1134 u32 in_flight;
1135};
1136
1137/* A rate sample measures the number of (original/retransmitted) data
1138 * packets delivered "delivered" over an interval of time "interval_us".
1139 * The tcp_rate.c code fills in the rate sample, and congestion
1140 * control modules that define a cong_control function to run at the end
1141 * of ACK processing can optionally chose to consult this sample when
1142 * setting cwnd and pacing rate.
1143 * A sample is invalid if "delivered" or "interval_us" is negative.
1144 */
1145struct rate_sample {
1146 u64 prior_mstamp; /* starting timestamp for interval */
1147 u32 prior_delivered; /* tp->delivered at "prior_mstamp" */
1148 u32 prior_delivered_ce;/* tp->delivered_ce at "prior_mstamp" */
1149 s32 delivered; /* number of packets delivered over interval */
1150 s32 delivered_ce; /* number of packets delivered w/ CE marks*/
1151 long interval_us; /* time for tp->delivered to incr "delivered" */
1152 u32 snd_interval_us; /* snd interval for delivered packets */
1153 u32 rcv_interval_us; /* rcv interval for delivered packets */
1154 long rtt_us; /* RTT of last (S)ACKed packet (or -1) */
1155 int losses; /* number of packets marked lost upon ACK */
1156 u32 acked_sacked; /* number of packets newly (S)ACKed upon ACK */
1157 u32 prior_in_flight; /* in flight before this ACK */
1158 u32 last_end_seq; /* end_seq of most recently ACKed packet */
1159 bool is_app_limited; /* is sample from packet with bubble in pipe? */
1160 bool is_retrans; /* is sample from retransmission? */
1161 bool is_ack_delayed; /* is this (likely) a delayed ACK? */
1162};
1163
1164struct tcp_congestion_ops {
1165/* fast path fields are put first to fill one cache line */
1166
1167 /* return slow start threshold (required) */
1168 u32 (*ssthresh)(struct sock *sk);
1169
1170 /* do new cwnd calculation (required) */
1171 void (*cong_avoid)(struct sock *sk, u32 ack, u32 acked);
1172
1173 /* call before changing ca_state (optional) */
1174 void (*set_state)(struct sock *sk, u8 new_state);
1175
1176 /* call when cwnd event occurs (optional) */
1177 void (*cwnd_event)(struct sock *sk, enum tcp_ca_event ev);
1178
1179 /* call when ack arrives (optional) */
1180 void (*in_ack_event)(struct sock *sk, u32 flags);
1181
1182 /* hook for packet ack accounting (optional) */
1183 void (*pkts_acked)(struct sock *sk, const struct ack_sample *sample);
1184
1185 /* override sysctl_tcp_min_tso_segs */
1186 u32 (*min_tso_segs)(struct sock *sk);
1187
1188 /* call when packets are delivered to update cwnd and pacing rate,
1189 * after all the ca_state processing. (optional)
1190 */
1191 void (*cong_control)(struct sock *sk, u32 ack, int flag, const struct rate_sample *rs);
1192
1193
1194 /* new value of cwnd after loss (required) */
1195 u32 (*undo_cwnd)(struct sock *sk);
1196 /* returns the multiplier used in tcp_sndbuf_expand (optional) */
1197 u32 (*sndbuf_expand)(struct sock *sk);
1198
1199/* control/slow paths put last */
1200 /* get info for inet_diag (optional) */
1201 size_t (*get_info)(struct sock *sk, u32 ext, int *attr,
1202 union tcp_cc_info *info);
1203
1204 char name[TCP_CA_NAME_MAX];
1205 struct module *owner;
1206 struct list_head list;
1207 u32 key;
1208 u32 flags;
1209
1210 /* initialize private data (optional) */
1211 void (*init)(struct sock *sk);
1212 /* cleanup private data (optional) */
1213 void (*release)(struct sock *sk);
1214} ____cacheline_aligned_in_smp;
1215
1216int tcp_register_congestion_control(struct tcp_congestion_ops *type);
1217void tcp_unregister_congestion_control(struct tcp_congestion_ops *type);
1218int tcp_update_congestion_control(struct tcp_congestion_ops *type,
1219 struct tcp_congestion_ops *old_type);
1220int tcp_validate_congestion_control(struct tcp_congestion_ops *ca);
1221
1222void tcp_assign_congestion_control(struct sock *sk);
1223void tcp_init_congestion_control(struct sock *sk);
1224void tcp_cleanup_congestion_control(struct sock *sk);
1225int tcp_set_default_congestion_control(struct net *net, const char *name);
1226void tcp_get_default_congestion_control(struct net *net, char *name);
1227void tcp_get_available_congestion_control(char *buf, size_t len);
1228void tcp_get_allowed_congestion_control(char *buf, size_t len);
1229int tcp_set_allowed_congestion_control(char *allowed);
1230int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
1231 bool cap_net_admin);
1232u32 tcp_slow_start(struct tcp_sock *tp, u32 acked);
1233void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked);
1234
1235u32 tcp_reno_ssthresh(struct sock *sk);
1236u32 tcp_reno_undo_cwnd(struct sock *sk);
1237void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked);
1238extern struct tcp_congestion_ops tcp_reno;
1239
1240struct tcp_congestion_ops *tcp_ca_find(const char *name);
1241struct tcp_congestion_ops *tcp_ca_find_key(u32 key);
1242u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca);
1243#ifdef CONFIG_INET
1244char *tcp_ca_get_name_by_key(u32 key, char *buffer);
1245#else
1246static inline char *tcp_ca_get_name_by_key(u32 key, char *buffer)
1247{
1248 return NULL;
1249}
1250#endif
1251
1252static inline bool tcp_ca_needs_ecn(const struct sock *sk)
1253{
1254 const struct inet_connection_sock *icsk = inet_csk(sk);
1255
1256 return icsk->icsk_ca_ops->flags & TCP_CONG_NEEDS_ECN;
1257}
1258
1259static inline void tcp_ca_event(struct sock *sk, const enum tcp_ca_event event)
1260{
1261 const struct inet_connection_sock *icsk = inet_csk(sk);
1262
1263 if (icsk->icsk_ca_ops->cwnd_event)
1264 icsk->icsk_ca_ops->cwnd_event(sk, event);
1265}
1266
1267/* From tcp_cong.c */
1268void tcp_set_ca_state(struct sock *sk, const u8 ca_state);
1269
1270/* From tcp_rate.c */
1271void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb);
1272void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb,
1273 struct rate_sample *rs);
1274void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost,
1275 bool is_sack_reneg, struct rate_sample *rs);
1276void tcp_rate_check_app_limited(struct sock *sk);
1277
1278static inline bool tcp_skb_sent_after(u64 t1, u64 t2, u32 seq1, u32 seq2)
1279{
1280 return t1 > t2 || (t1 == t2 && after(seq1, seq2));
1281}
1282
1283/* These functions determine how the current flow behaves in respect of SACK
1284 * handling. SACK is negotiated with the peer, and therefore it can vary
1285 * between different flows.
1286 *
1287 * tcp_is_sack - SACK enabled
1288 * tcp_is_reno - No SACK
1289 */
1290static inline int tcp_is_sack(const struct tcp_sock *tp)
1291{
1292 return likely(tp->rx_opt.sack_ok);
1293}
1294
1295static inline bool tcp_is_reno(const struct tcp_sock *tp)
1296{
1297 return !tcp_is_sack(tp);
1298}
1299
1300static inline unsigned int tcp_left_out(const struct tcp_sock *tp)
1301{
1302 return tp->sacked_out + tp->lost_out;
1303}
1304
1305/* This determines how many packets are "in the network" to the best
1306 * of our knowledge. In many cases it is conservative, but where
1307 * detailed information is available from the receiver (via SACK
1308 * blocks etc.) we can make more aggressive calculations.
1309 *
1310 * Use this for decisions involving congestion control, use just
1311 * tp->packets_out to determine if the send queue is empty or not.
1312 *
1313 * Read this equation as:
1314 *
1315 * "Packets sent once on transmission queue" MINUS
1316 * "Packets left network, but not honestly ACKed yet" PLUS
1317 * "Packets fast retransmitted"
1318 */
1319static inline unsigned int tcp_packets_in_flight(const struct tcp_sock *tp)
1320{
1321 return tp->packets_out - tcp_left_out(tp) + tp->retrans_out;
1322}
1323
1324#define TCP_INFINITE_SSTHRESH 0x7fffffff
1325
1326static inline u32 tcp_snd_cwnd(const struct tcp_sock *tp)
1327{
1328 return tp->snd_cwnd;
1329}
1330
1331static inline void tcp_snd_cwnd_set(struct tcp_sock *tp, u32 val)
1332{
1333 WARN_ON_ONCE((int)val <= 0);
1334 tp->snd_cwnd = val;
1335}
1336
1337static inline bool tcp_in_slow_start(const struct tcp_sock *tp)
1338{
1339 return tcp_snd_cwnd(tp) < tp->snd_ssthresh;
1340}
1341
1342static inline bool tcp_in_initial_slowstart(const struct tcp_sock *tp)
1343{
1344 return tp->snd_ssthresh >= TCP_INFINITE_SSTHRESH;
1345}
1346
1347static inline bool tcp_in_cwnd_reduction(const struct sock *sk)
1348{
1349 return (TCPF_CA_CWR | TCPF_CA_Recovery) &
1350 (1 << inet_csk(sk)->icsk_ca_state);
1351}
1352
1353/* If cwnd > ssthresh, we may raise ssthresh to be half-way to cwnd.
1354 * The exception is cwnd reduction phase, when cwnd is decreasing towards
1355 * ssthresh.
1356 */
1357static inline __u32 tcp_current_ssthresh(const struct sock *sk)
1358{
1359 const struct tcp_sock *tp = tcp_sk(sk);
1360
1361 if (tcp_in_cwnd_reduction(sk))
1362 return tp->snd_ssthresh;
1363 else
1364 return max(tp->snd_ssthresh,
1365 ((tcp_snd_cwnd(tp) >> 1) +
1366 (tcp_snd_cwnd(tp) >> 2)));
1367}
1368
1369/* Use define here intentionally to get WARN_ON location shown at the caller */
1370#define tcp_verify_left_out(tp) WARN_ON(tcp_left_out(tp) > tp->packets_out)
1371
1372void tcp_enter_cwr(struct sock *sk);
1373__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst);
1374
1375/* The maximum number of MSS of available cwnd for which TSO defers
1376 * sending if not using sysctl_tcp_tso_win_divisor.
1377 */
1378static inline __u32 tcp_max_tso_deferred_mss(const struct tcp_sock *tp)
1379{
1380 return 3;
1381}
1382
1383/* Returns end sequence number of the receiver's advertised window */
1384static inline u32 tcp_wnd_end(const struct tcp_sock *tp)
1385{
1386 return tp->snd_una + tp->snd_wnd;
1387}
1388
1389/* We follow the spirit of RFC2861 to validate cwnd but implement a more
1390 * flexible approach. The RFC suggests cwnd should not be raised unless
1391 * it was fully used previously. And that's exactly what we do in
1392 * congestion avoidance mode. But in slow start we allow cwnd to grow
1393 * as long as the application has used half the cwnd.
1394 * Example :
1395 * cwnd is 10 (IW10), but application sends 9 frames.
1396 * We allow cwnd to reach 18 when all frames are ACKed.
1397 * This check is safe because it's as aggressive as slow start which already
1398 * risks 100% overshoot. The advantage is that we discourage application to
1399 * either send more filler packets or data to artificially blow up the cwnd
1400 * usage, and allow application-limited process to probe bw more aggressively.
1401 */
1402static inline bool tcp_is_cwnd_limited(const struct sock *sk)
1403{
1404 const struct tcp_sock *tp = tcp_sk(sk);
1405
1406 if (tp->is_cwnd_limited)
1407 return true;
1408
1409 /* If in slow start, ensure cwnd grows to twice what was ACKed. */
1410 if (tcp_in_slow_start(tp))
1411 return tcp_snd_cwnd(tp) < 2 * tp->max_packets_out;
1412
1413 return false;
1414}
1415
1416/* BBR congestion control needs pacing.
1417 * Same remark for SO_MAX_PACING_RATE.
1418 * sch_fq packet scheduler is efficiently handling pacing,
1419 * but is not always installed/used.
1420 * Return true if TCP stack should pace packets itself.
1421 */
1422static inline bool tcp_needs_internal_pacing(const struct sock *sk)
1423{
1424 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
1425}
1426
1427/* Estimates in how many jiffies next packet for this flow can be sent.
1428 * Scheduling a retransmit timer too early would be silly.
1429 */
1430static inline unsigned long tcp_pacing_delay(const struct sock *sk)
1431{
1432 s64 delay = tcp_sk(sk)->tcp_wstamp_ns - tcp_sk(sk)->tcp_clock_cache;
1433
1434 return delay > 0 ? nsecs_to_jiffies(delay) : 0;
1435}
1436
1437static inline void tcp_reset_xmit_timer(struct sock *sk,
1438 const int what,
1439 unsigned long when,
1440 const unsigned long max_when)
1441{
1442 inet_csk_reset_xmit_timer(sk, what, when + tcp_pacing_delay(sk),
1443 max_when);
1444}
1445
1446/* Something is really bad, we could not queue an additional packet,
1447 * because qdisc is full or receiver sent a 0 window, or we are paced.
1448 * We do not want to add fuel to the fire, or abort too early,
1449 * so make sure the timer we arm now is at least 200ms in the future,
1450 * regardless of current icsk_rto value (as it could be ~2ms)
1451 */
1452static inline unsigned long tcp_probe0_base(const struct sock *sk)
1453{
1454 return max_t(unsigned long, inet_csk(sk)->icsk_rto, TCP_RTO_MIN);
1455}
1456
1457/* Variant of inet_csk_rto_backoff() used for zero window probes */
1458static inline unsigned long tcp_probe0_when(const struct sock *sk,
1459 unsigned long max_when)
1460{
1461 u8 backoff = min_t(u8, ilog2(TCP_RTO_MAX / TCP_RTO_MIN) + 1,
1462 inet_csk(sk)->icsk_backoff);
1463 u64 when = (u64)tcp_probe0_base(sk) << backoff;
1464
1465 return (unsigned long)min_t(u64, when, max_when);
1466}
1467
1468static inline void tcp_check_probe_timer(struct sock *sk)
1469{
1470 if (!tcp_sk(sk)->packets_out && !inet_csk(sk)->icsk_pending)
1471 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
1472 tcp_probe0_base(sk), TCP_RTO_MAX);
1473}
1474
1475static inline void tcp_init_wl(struct tcp_sock *tp, u32 seq)
1476{
1477 tp->snd_wl1 = seq;
1478}
1479
1480static inline void tcp_update_wl(struct tcp_sock *tp, u32 seq)
1481{
1482 tp->snd_wl1 = seq;
1483}
1484
1485/*
1486 * Calculate(/check) TCP checksum
1487 */
1488static inline __sum16 tcp_v4_check(int len, __be32 saddr,
1489 __be32 daddr, __wsum base)
1490{
1491 return csum_tcpudp_magic(saddr, daddr, len, IPPROTO_TCP, base);
1492}
1493
1494static inline bool tcp_checksum_complete(struct sk_buff *skb)
1495{
1496 return !skb_csum_unnecessary(skb) &&
1497 __skb_checksum_complete(skb);
1498}
1499
1500bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
1501 enum skb_drop_reason *reason);
1502
1503
1504int tcp_filter(struct sock *sk, struct sk_buff *skb);
1505void tcp_set_state(struct sock *sk, int state);
1506void tcp_done(struct sock *sk);
1507int tcp_abort(struct sock *sk, int err);
1508
1509static inline void tcp_sack_reset(struct tcp_options_received *rx_opt)
1510{
1511 rx_opt->dsack = 0;
1512 rx_opt->num_sacks = 0;
1513}
1514
1515void tcp_cwnd_restart(struct sock *sk, s32 delta);
1516
1517static inline void tcp_slow_start_after_idle_check(struct sock *sk)
1518{
1519 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1520 struct tcp_sock *tp = tcp_sk(sk);
1521 s32 delta;
1522
1523 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) ||
1524 tp->packets_out || ca_ops->cong_control)
1525 return;
1526 delta = tcp_jiffies32 - tp->lsndtime;
1527 if (delta > inet_csk(sk)->icsk_rto)
1528 tcp_cwnd_restart(sk, delta);
1529}
1530
1531/* Determine a window scaling and initial window to offer. */
1532void tcp_select_initial_window(const struct sock *sk, int __space,
1533 __u32 mss, __u32 *rcv_wnd,
1534 __u32 *window_clamp, int wscale_ok,
1535 __u8 *rcv_wscale, __u32 init_rcv_wnd);
1536
1537static inline int __tcp_win_from_space(u8 scaling_ratio, int space)
1538{
1539 s64 scaled_space = (s64)space * scaling_ratio;
1540
1541 return scaled_space >> TCP_RMEM_TO_WIN_SCALE;
1542}
1543
1544static inline int tcp_win_from_space(const struct sock *sk, int space)
1545{
1546 return __tcp_win_from_space(tcp_sk(sk)->scaling_ratio, space);
1547}
1548
1549/* inverse of __tcp_win_from_space() */
1550static inline int __tcp_space_from_win(u8 scaling_ratio, int win)
1551{
1552 u64 val = (u64)win << TCP_RMEM_TO_WIN_SCALE;
1553
1554 do_div(val, scaling_ratio);
1555 return val;
1556}
1557
1558static inline int tcp_space_from_win(const struct sock *sk, int win)
1559{
1560 return __tcp_space_from_win(tcp_sk(sk)->scaling_ratio, win);
1561}
1562
1563/* Assume a 50% default for skb->len/skb->truesize ratio.
1564 * This may be adjusted later in tcp_measure_rcv_mss().
1565 */
1566#define TCP_DEFAULT_SCALING_RATIO (1 << (TCP_RMEM_TO_WIN_SCALE - 1))
1567
1568static inline void tcp_scaling_ratio_init(struct sock *sk)
1569{
1570 tcp_sk(sk)->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
1571}
1572
1573/* Note: caller must be prepared to deal with negative returns */
1574static inline int tcp_space(const struct sock *sk)
1575{
1576 return tcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf) -
1577 READ_ONCE(sk->sk_backlog.len) -
1578 atomic_read(&sk->sk_rmem_alloc));
1579}
1580
1581static inline int tcp_full_space(const struct sock *sk)
1582{
1583 return tcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf));
1584}
1585
1586static inline void __tcp_adjust_rcv_ssthresh(struct sock *sk, u32 new_ssthresh)
1587{
1588 int unused_mem = sk_unused_reserved_mem(sk);
1589 struct tcp_sock *tp = tcp_sk(sk);
1590
1591 tp->rcv_ssthresh = min(tp->rcv_ssthresh, new_ssthresh);
1592 if (unused_mem)
1593 tp->rcv_ssthresh = max_t(u32, tp->rcv_ssthresh,
1594 tcp_win_from_space(sk, unused_mem));
1595}
1596
1597static inline void tcp_adjust_rcv_ssthresh(struct sock *sk)
1598{
1599 __tcp_adjust_rcv_ssthresh(sk, 4U * tcp_sk(sk)->advmss);
1600}
1601
1602void tcp_cleanup_rbuf(struct sock *sk, int copied);
1603void __tcp_cleanup_rbuf(struct sock *sk, int copied);
1604
1605
1606/* We provision sk_rcvbuf around 200% of sk_rcvlowat.
1607 * If 87.5 % (7/8) of the space has been consumed, we want to override
1608 * SO_RCVLOWAT constraint, since we are receiving skbs with too small
1609 * len/truesize ratio.
1610 */
1611static inline bool tcp_rmem_pressure(const struct sock *sk)
1612{
1613 int rcvbuf, threshold;
1614
1615 if (tcp_under_memory_pressure(sk))
1616 return true;
1617
1618 rcvbuf = READ_ONCE(sk->sk_rcvbuf);
1619 threshold = rcvbuf - (rcvbuf >> 3);
1620
1621 return atomic_read(&sk->sk_rmem_alloc) > threshold;
1622}
1623
1624static inline bool tcp_epollin_ready(const struct sock *sk, int target)
1625{
1626 const struct tcp_sock *tp = tcp_sk(sk);
1627 int avail = READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq);
1628
1629 if (avail <= 0)
1630 return false;
1631
1632 return (avail >= target) || tcp_rmem_pressure(sk) ||
1633 (tcp_receive_window(tp) <= inet_csk(sk)->icsk_ack.rcv_mss);
1634}
1635
1636extern void tcp_openreq_init_rwin(struct request_sock *req,
1637 const struct sock *sk_listener,
1638 const struct dst_entry *dst);
1639
1640void tcp_enter_memory_pressure(struct sock *sk);
1641void tcp_leave_memory_pressure(struct sock *sk);
1642
1643static inline int keepalive_intvl_when(const struct tcp_sock *tp)
1644{
1645 struct net *net = sock_net((struct sock *)tp);
1646 int val;
1647
1648 /* Paired with WRITE_ONCE() in tcp_sock_set_keepintvl()
1649 * and do_tcp_setsockopt().
1650 */
1651 val = READ_ONCE(tp->keepalive_intvl);
1652
1653 return val ? : READ_ONCE(net->ipv4.sysctl_tcp_keepalive_intvl);
1654}
1655
1656static inline int keepalive_time_when(const struct tcp_sock *tp)
1657{
1658 struct net *net = sock_net((struct sock *)tp);
1659 int val;
1660
1661 /* Paired with WRITE_ONCE() in tcp_sock_set_keepidle_locked() */
1662 val = READ_ONCE(tp->keepalive_time);
1663
1664 return val ? : READ_ONCE(net->ipv4.sysctl_tcp_keepalive_time);
1665}
1666
1667static inline int keepalive_probes(const struct tcp_sock *tp)
1668{
1669 struct net *net = sock_net((struct sock *)tp);
1670 int val;
1671
1672 /* Paired with WRITE_ONCE() in tcp_sock_set_keepcnt()
1673 * and do_tcp_setsockopt().
1674 */
1675 val = READ_ONCE(tp->keepalive_probes);
1676
1677 return val ? : READ_ONCE(net->ipv4.sysctl_tcp_keepalive_probes);
1678}
1679
1680static inline u32 keepalive_time_elapsed(const struct tcp_sock *tp)
1681{
1682 const struct inet_connection_sock *icsk = &tp->inet_conn;
1683
1684 return min_t(u32, tcp_jiffies32 - icsk->icsk_ack.lrcvtime,
1685 tcp_jiffies32 - tp->rcv_tstamp);
1686}
1687
1688static inline int tcp_fin_time(const struct sock *sk)
1689{
1690 int fin_timeout = tcp_sk(sk)->linger2 ? :
1691 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fin_timeout);
1692 const int rto = inet_csk(sk)->icsk_rto;
1693
1694 if (fin_timeout < (rto << 2) - (rto >> 1))
1695 fin_timeout = (rto << 2) - (rto >> 1);
1696
1697 return fin_timeout;
1698}
1699
1700static inline bool tcp_paws_check(const struct tcp_options_received *rx_opt,
1701 int paws_win)
1702{
1703 if ((s32)(rx_opt->ts_recent - rx_opt->rcv_tsval) <= paws_win)
1704 return true;
1705 if (unlikely(!time_before32(ktime_get_seconds(),
1706 rx_opt->ts_recent_stamp + TCP_PAWS_WRAP)))
1707 return true;
1708 /*
1709 * Some OSes send SYN and SYNACK messages with tsval=0 tsecr=0,
1710 * then following tcp messages have valid values. Ignore 0 value,
1711 * or else 'negative' tsval might forbid us to accept their packets.
1712 */
1713 if (!rx_opt->ts_recent)
1714 return true;
1715 return false;
1716}
1717
1718static inline bool tcp_paws_reject(const struct tcp_options_received *rx_opt,
1719 int rst)
1720{
1721 if (tcp_paws_check(rx_opt, 0))
1722 return false;
1723
1724 /* RST segments are not recommended to carry timestamp,
1725 and, if they do, it is recommended to ignore PAWS because
1726 "their cleanup function should take precedence over timestamps."
1727 Certainly, it is mistake. It is necessary to understand the reasons
1728 of this constraint to relax it: if peer reboots, clock may go
1729 out-of-sync and half-open connections will not be reset.
1730 Actually, the problem would be not existing if all
1731 the implementations followed draft about maintaining clock
1732 via reboots. Linux-2.2 DOES NOT!
1733
1734 However, we can relax time bounds for RST segments to MSL.
1735 */
1736 if (rst && !time_before32(ktime_get_seconds(),
1737 rx_opt->ts_recent_stamp + TCP_PAWS_MSL))
1738 return false;
1739 return true;
1740}
1741
1742bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
1743 int mib_idx, u32 *last_oow_ack_time);
1744
1745static inline void tcp_mib_init(struct net *net)
1746{
1747 /* See RFC 2012 */
1748 TCP_ADD_STATS(net, TCP_MIB_RTOALGORITHM, 1);
1749 TCP_ADD_STATS(net, TCP_MIB_RTOMIN, TCP_RTO_MIN*1000/HZ);
1750 TCP_ADD_STATS(net, TCP_MIB_RTOMAX, TCP_RTO_MAX*1000/HZ);
1751 TCP_ADD_STATS(net, TCP_MIB_MAXCONN, -1);
1752}
1753
1754/* from STCP */
1755static inline void tcp_clear_retrans_hints_partial(struct tcp_sock *tp)
1756{
1757 tp->lost_skb_hint = NULL;
1758}
1759
1760static inline void tcp_clear_all_retrans_hints(struct tcp_sock *tp)
1761{
1762 tcp_clear_retrans_hints_partial(tp);
1763 tp->retransmit_skb_hint = NULL;
1764}
1765
1766#define tcp_md5_addr tcp_ao_addr
1767
1768/* - key database */
1769struct tcp_md5sig_key {
1770 struct hlist_node node;
1771 u8 keylen;
1772 u8 family; /* AF_INET or AF_INET6 */
1773 u8 prefixlen;
1774 u8 flags;
1775 union tcp_md5_addr addr;
1776 int l3index; /* set if key added with L3 scope */
1777 u8 key[TCP_MD5SIG_MAXKEYLEN];
1778 struct rcu_head rcu;
1779};
1780
1781/* - sock block */
1782struct tcp_md5sig_info {
1783 struct hlist_head head;
1784 struct rcu_head rcu;
1785};
1786
1787/* - pseudo header */
1788struct tcp4_pseudohdr {
1789 __be32 saddr;
1790 __be32 daddr;
1791 __u8 pad;
1792 __u8 protocol;
1793 __be16 len;
1794};
1795
1796struct tcp6_pseudohdr {
1797 struct in6_addr saddr;
1798 struct in6_addr daddr;
1799 __be32 len;
1800 __be32 protocol; /* including padding */
1801};
1802
1803union tcp_md5sum_block {
1804 struct tcp4_pseudohdr ip4;
1805#if IS_ENABLED(CONFIG_IPV6)
1806 struct tcp6_pseudohdr ip6;
1807#endif
1808};
1809
1810/*
1811 * struct tcp_sigpool - per-CPU pool of ahash_requests
1812 * @scratch: per-CPU temporary area, that can be used between
1813 * tcp_sigpool_start() and tcp_sigpool_end() to perform
1814 * crypto request
1815 * @req: pre-allocated ahash request
1816 */
1817struct tcp_sigpool {
1818 void *scratch;
1819 struct ahash_request *req;
1820};
1821
1822int tcp_sigpool_alloc_ahash(const char *alg, size_t scratch_size);
1823void tcp_sigpool_get(unsigned int id);
1824void tcp_sigpool_release(unsigned int id);
1825int tcp_sigpool_hash_skb_data(struct tcp_sigpool *hp,
1826 const struct sk_buff *skb,
1827 unsigned int header_len);
1828
1829/**
1830 * tcp_sigpool_start - disable bh and start using tcp_sigpool_ahash
1831 * @id: tcp_sigpool that was previously allocated by tcp_sigpool_alloc_ahash()
1832 * @c: returned tcp_sigpool for usage (uninitialized on failure)
1833 *
1834 * Returns 0 on success, error otherwise.
1835 */
1836int tcp_sigpool_start(unsigned int id, struct tcp_sigpool *c);
1837/**
1838 * tcp_sigpool_end - enable bh and stop using tcp_sigpool
1839 * @c: tcp_sigpool context that was returned by tcp_sigpool_start()
1840 */
1841void tcp_sigpool_end(struct tcp_sigpool *c);
1842size_t tcp_sigpool_algo(unsigned int id, char *buf, size_t buf_len);
1843/* - functions */
1844int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1845 const struct sock *sk, const struct sk_buff *skb);
1846int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1847 int family, u8 prefixlen, int l3index, u8 flags,
1848 const u8 *newkey, u8 newkeylen);
1849int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
1850 int family, u8 prefixlen, int l3index,
1851 struct tcp_md5sig_key *key);
1852
1853int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr,
1854 int family, u8 prefixlen, int l3index, u8 flags);
1855void tcp_clear_md5_list(struct sock *sk);
1856struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
1857 const struct sock *addr_sk);
1858
1859#ifdef CONFIG_TCP_MD5SIG
1860struct tcp_md5sig_key *__tcp_md5_do_lookup(const struct sock *sk, int l3index,
1861 const union tcp_md5_addr *addr,
1862 int family, bool any_l3index);
1863static inline struct tcp_md5sig_key *
1864tcp_md5_do_lookup(const struct sock *sk, int l3index,
1865 const union tcp_md5_addr *addr, int family)
1866{
1867 if (!static_branch_unlikely(&tcp_md5_needed.key))
1868 return NULL;
1869 return __tcp_md5_do_lookup(sk, l3index, addr, family, false);
1870}
1871
1872static inline struct tcp_md5sig_key *
1873tcp_md5_do_lookup_any_l3index(const struct sock *sk,
1874 const union tcp_md5_addr *addr, int family)
1875{
1876 if (!static_branch_unlikely(&tcp_md5_needed.key))
1877 return NULL;
1878 return __tcp_md5_do_lookup(sk, 0, addr, family, true);
1879}
1880
1881#define tcp_twsk_md5_key(twsk) ((twsk)->tw_md5_key)
1882#else
1883static inline struct tcp_md5sig_key *
1884tcp_md5_do_lookup(const struct sock *sk, int l3index,
1885 const union tcp_md5_addr *addr, int family)
1886{
1887 return NULL;
1888}
1889
1890static inline struct tcp_md5sig_key *
1891tcp_md5_do_lookup_any_l3index(const struct sock *sk,
1892 const union tcp_md5_addr *addr, int family)
1893{
1894 return NULL;
1895}
1896
1897#define tcp_twsk_md5_key(twsk) NULL
1898#endif
1899
1900int tcp_md5_alloc_sigpool(void);
1901void tcp_md5_release_sigpool(void);
1902void tcp_md5_add_sigpool(void);
1903extern int tcp_md5_sigpool_id;
1904
1905int tcp_md5_hash_key(struct tcp_sigpool *hp,
1906 const struct tcp_md5sig_key *key);
1907
1908/* From tcp_fastopen.c */
1909void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
1910 struct tcp_fastopen_cookie *cookie);
1911void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
1912 struct tcp_fastopen_cookie *cookie, bool syn_lost,
1913 u16 try_exp);
1914struct tcp_fastopen_request {
1915 /* Fast Open cookie. Size 0 means a cookie request */
1916 struct tcp_fastopen_cookie cookie;
1917 struct msghdr *data; /* data in MSG_FASTOPEN */
1918 size_t size;
1919 int copied; /* queued in tcp_connect() */
1920 struct ubuf_info *uarg;
1921};
1922void tcp_free_fastopen_req(struct tcp_sock *tp);
1923void tcp_fastopen_destroy_cipher(struct sock *sk);
1924void tcp_fastopen_ctx_destroy(struct net *net);
1925int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
1926 void *primary_key, void *backup_key);
1927int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
1928 u64 *key);
1929void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb);
1930struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
1931 struct request_sock *req,
1932 struct tcp_fastopen_cookie *foc,
1933 const struct dst_entry *dst);
1934void tcp_fastopen_init_key_once(struct net *net);
1935bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
1936 struct tcp_fastopen_cookie *cookie);
1937bool tcp_fastopen_defer_connect(struct sock *sk, int *err);
1938#define TCP_FASTOPEN_KEY_LENGTH sizeof(siphash_key_t)
1939#define TCP_FASTOPEN_KEY_MAX 2
1940#define TCP_FASTOPEN_KEY_BUF_LENGTH \
1941 (TCP_FASTOPEN_KEY_LENGTH * TCP_FASTOPEN_KEY_MAX)
1942
1943/* Fastopen key context */
1944struct tcp_fastopen_context {
1945 siphash_key_t key[TCP_FASTOPEN_KEY_MAX];
1946 int num;
1947 struct rcu_head rcu;
1948};
1949
1950void tcp_fastopen_active_disable(struct sock *sk);
1951bool tcp_fastopen_active_should_disable(struct sock *sk);
1952void tcp_fastopen_active_disable_ofo_check(struct sock *sk);
1953void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired);
1954
1955/* Caller needs to wrap with rcu_read_(un)lock() */
1956static inline
1957struct tcp_fastopen_context *tcp_fastopen_get_ctx(const struct sock *sk)
1958{
1959 struct tcp_fastopen_context *ctx;
1960
1961 ctx = rcu_dereference(inet_csk(sk)->icsk_accept_queue.fastopenq.ctx);
1962 if (!ctx)
1963 ctx = rcu_dereference(sock_net(sk)->ipv4.tcp_fastopen_ctx);
1964 return ctx;
1965}
1966
1967static inline
1968bool tcp_fastopen_cookie_match(const struct tcp_fastopen_cookie *foc,
1969 const struct tcp_fastopen_cookie *orig)
1970{
1971 if (orig->len == TCP_FASTOPEN_COOKIE_SIZE &&
1972 orig->len == foc->len &&
1973 !memcmp(orig->val, foc->val, foc->len))
1974 return true;
1975 return false;
1976}
1977
1978static inline
1979int tcp_fastopen_context_len(const struct tcp_fastopen_context *ctx)
1980{
1981 return ctx->num;
1982}
1983
1984/* Latencies incurred by various limits for a sender. They are
1985 * chronograph-like stats that are mutually exclusive.
1986 */
1987enum tcp_chrono {
1988 TCP_CHRONO_UNSPEC,
1989 TCP_CHRONO_BUSY, /* Actively sending data (non-empty write queue) */
1990 TCP_CHRONO_RWND_LIMITED, /* Stalled by insufficient receive window */
1991 TCP_CHRONO_SNDBUF_LIMITED, /* Stalled by insufficient send buffer */
1992 __TCP_CHRONO_MAX,
1993};
1994
1995void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type);
1996void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type);
1997
1998/* This helper is needed, because skb->tcp_tsorted_anchor uses
1999 * the same memory storage than skb->destructor/_skb_refdst
2000 */
2001static inline void tcp_skb_tsorted_anchor_cleanup(struct sk_buff *skb)
2002{
2003 skb->destructor = NULL;
2004 skb->_skb_refdst = 0UL;
2005}
2006
2007#define tcp_skb_tsorted_save(skb) { \
2008 unsigned long _save = skb->_skb_refdst; \
2009 skb->_skb_refdst = 0UL;
2010
2011#define tcp_skb_tsorted_restore(skb) \
2012 skb->_skb_refdst = _save; \
2013}
2014
2015void tcp_write_queue_purge(struct sock *sk);
2016
2017static inline struct sk_buff *tcp_rtx_queue_head(const struct sock *sk)
2018{
2019 return skb_rb_first(&sk->tcp_rtx_queue);
2020}
2021
2022static inline struct sk_buff *tcp_rtx_queue_tail(const struct sock *sk)
2023{
2024 return skb_rb_last(&sk->tcp_rtx_queue);
2025}
2026
2027static inline struct sk_buff *tcp_write_queue_tail(const struct sock *sk)
2028{
2029 return skb_peek_tail(&sk->sk_write_queue);
2030}
2031
2032#define tcp_for_write_queue_from_safe(skb, tmp, sk) \
2033 skb_queue_walk_from_safe(&(sk)->sk_write_queue, skb, tmp)
2034
2035static inline struct sk_buff *tcp_send_head(const struct sock *sk)
2036{
2037 return skb_peek(&sk->sk_write_queue);
2038}
2039
2040static inline bool tcp_skb_is_last(const struct sock *sk,
2041 const struct sk_buff *skb)
2042{
2043 return skb_queue_is_last(&sk->sk_write_queue, skb);
2044}
2045
2046/**
2047 * tcp_write_queue_empty - test if any payload (or FIN) is available in write queue
2048 * @sk: socket
2049 *
2050 * Since the write queue can have a temporary empty skb in it,
2051 * we must not use "return skb_queue_empty(&sk->sk_write_queue)"
2052 */
2053static inline bool tcp_write_queue_empty(const struct sock *sk)
2054{
2055 const struct tcp_sock *tp = tcp_sk(sk);
2056
2057 return tp->write_seq == tp->snd_nxt;
2058}
2059
2060static inline bool tcp_rtx_queue_empty(const struct sock *sk)
2061{
2062 return RB_EMPTY_ROOT(&sk->tcp_rtx_queue);
2063}
2064
2065static inline bool tcp_rtx_and_write_queues_empty(const struct sock *sk)
2066{
2067 return tcp_rtx_queue_empty(sk) && tcp_write_queue_empty(sk);
2068}
2069
2070static inline void tcp_add_write_queue_tail(struct sock *sk, struct sk_buff *skb)
2071{
2072 __skb_queue_tail(&sk->sk_write_queue, skb);
2073
2074 /* Queue it, remembering where we must start sending. */
2075 if (sk->sk_write_queue.next == skb)
2076 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
2077}
2078
2079/* Insert new before skb on the write queue of sk. */
2080static inline void tcp_insert_write_queue_before(struct sk_buff *new,
2081 struct sk_buff *skb,
2082 struct sock *sk)
2083{
2084 __skb_queue_before(&sk->sk_write_queue, skb, new);
2085}
2086
2087static inline void tcp_unlink_write_queue(struct sk_buff *skb, struct sock *sk)
2088{
2089 tcp_skb_tsorted_anchor_cleanup(skb);
2090 __skb_unlink(skb, &sk->sk_write_queue);
2091}
2092
2093void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb);
2094
2095static inline void tcp_rtx_queue_unlink(struct sk_buff *skb, struct sock *sk)
2096{
2097 tcp_skb_tsorted_anchor_cleanup(skb);
2098 rb_erase(&skb->rbnode, &sk->tcp_rtx_queue);
2099}
2100
2101static inline void tcp_rtx_queue_unlink_and_free(struct sk_buff *skb, struct sock *sk)
2102{
2103 list_del(&skb->tcp_tsorted_anchor);
2104 tcp_rtx_queue_unlink(skb, sk);
2105 tcp_wmem_free_skb(sk, skb);
2106}
2107
2108static inline void tcp_write_collapse_fence(struct sock *sk)
2109{
2110 struct sk_buff *skb = tcp_write_queue_tail(sk);
2111
2112 if (skb)
2113 TCP_SKB_CB(skb)->eor = 1;
2114}
2115
2116static inline void tcp_push_pending_frames(struct sock *sk)
2117{
2118 if (tcp_send_head(sk)) {
2119 struct tcp_sock *tp = tcp_sk(sk);
2120
2121 __tcp_push_pending_frames(sk, tcp_current_mss(sk), tp->nonagle);
2122 }
2123}
2124
2125/* Start sequence of the skb just after the highest skb with SACKed
2126 * bit, valid only if sacked_out > 0 or when the caller has ensured
2127 * validity by itself.
2128 */
2129static inline u32 tcp_highest_sack_seq(struct tcp_sock *tp)
2130{
2131 if (!tp->sacked_out)
2132 return tp->snd_una;
2133
2134 if (tp->highest_sack == NULL)
2135 return tp->snd_nxt;
2136
2137 return TCP_SKB_CB(tp->highest_sack)->seq;
2138}
2139
2140static inline void tcp_advance_highest_sack(struct sock *sk, struct sk_buff *skb)
2141{
2142 tcp_sk(sk)->highest_sack = skb_rb_next(skb);
2143}
2144
2145static inline struct sk_buff *tcp_highest_sack(struct sock *sk)
2146{
2147 return tcp_sk(sk)->highest_sack;
2148}
2149
2150static inline void tcp_highest_sack_reset(struct sock *sk)
2151{
2152 tcp_sk(sk)->highest_sack = tcp_rtx_queue_head(sk);
2153}
2154
2155/* Called when old skb is about to be deleted and replaced by new skb */
2156static inline void tcp_highest_sack_replace(struct sock *sk,
2157 struct sk_buff *old,
2158 struct sk_buff *new)
2159{
2160 if (old == tcp_highest_sack(sk))
2161 tcp_sk(sk)->highest_sack = new;
2162}
2163
2164/* This helper checks if socket has IP_TRANSPARENT set */
2165static inline bool inet_sk_transparent(const struct sock *sk)
2166{
2167 switch (sk->sk_state) {
2168 case TCP_TIME_WAIT:
2169 return inet_twsk(sk)->tw_transparent;
2170 case TCP_NEW_SYN_RECV:
2171 return inet_rsk(inet_reqsk(sk))->no_srccheck;
2172 }
2173 return inet_test_bit(TRANSPARENT, sk);
2174}
2175
2176/* Determines whether this is a thin stream (which may suffer from
2177 * increased latency). Used to trigger latency-reducing mechanisms.
2178 */
2179static inline bool tcp_stream_is_thin(struct tcp_sock *tp)
2180{
2181 return tp->packets_out < 4 && !tcp_in_initial_slowstart(tp);
2182}
2183
2184/* /proc */
2185enum tcp_seq_states {
2186 TCP_SEQ_STATE_LISTENING,
2187 TCP_SEQ_STATE_ESTABLISHED,
2188};
2189
2190void *tcp_seq_start(struct seq_file *seq, loff_t *pos);
2191void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos);
2192void tcp_seq_stop(struct seq_file *seq, void *v);
2193
2194struct tcp_seq_afinfo {
2195 sa_family_t family;
2196};
2197
2198struct tcp_iter_state {
2199 struct seq_net_private p;
2200 enum tcp_seq_states state;
2201 struct sock *syn_wait_sk;
2202 int bucket, offset, sbucket, num;
2203 loff_t last_pos;
2204};
2205
2206extern struct request_sock_ops tcp_request_sock_ops;
2207extern struct request_sock_ops tcp6_request_sock_ops;
2208
2209void tcp_v4_destroy_sock(struct sock *sk);
2210
2211struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
2212 netdev_features_t features);
2213struct tcphdr *tcp_gro_pull_header(struct sk_buff *skb);
2214struct sk_buff *tcp_gro_lookup(struct list_head *head, struct tcphdr *th);
2215struct sk_buff *tcp_gro_receive(struct list_head *head, struct sk_buff *skb,
2216 struct tcphdr *th);
2217INDIRECT_CALLABLE_DECLARE(int tcp4_gro_complete(struct sk_buff *skb, int thoff));
2218INDIRECT_CALLABLE_DECLARE(struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb));
2219INDIRECT_CALLABLE_DECLARE(int tcp6_gro_complete(struct sk_buff *skb, int thoff));
2220INDIRECT_CALLABLE_DECLARE(struct sk_buff *tcp6_gro_receive(struct list_head *head, struct sk_buff *skb));
2221#ifdef CONFIG_INET
2222void tcp_gro_complete(struct sk_buff *skb);
2223#else
2224static inline void tcp_gro_complete(struct sk_buff *skb) { }
2225#endif
2226
2227void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr);
2228
2229static inline u32 tcp_notsent_lowat(const struct tcp_sock *tp)
2230{
2231 struct net *net = sock_net((struct sock *)tp);
2232 u32 val;
2233
2234 val = READ_ONCE(tp->notsent_lowat);
2235
2236 return val ?: READ_ONCE(net->ipv4.sysctl_tcp_notsent_lowat);
2237}
2238
2239bool tcp_stream_memory_free(const struct sock *sk, int wake);
2240
2241#ifdef CONFIG_PROC_FS
2242int tcp4_proc_init(void);
2243void tcp4_proc_exit(void);
2244#endif
2245
2246int tcp_rtx_synack(const struct sock *sk, struct request_sock *req);
2247int tcp_conn_request(struct request_sock_ops *rsk_ops,
2248 const struct tcp_request_sock_ops *af_ops,
2249 struct sock *sk, struct sk_buff *skb);
2250
2251/* TCP af-specific functions */
2252struct tcp_sock_af_ops {
2253#ifdef CONFIG_TCP_MD5SIG
2254 struct tcp_md5sig_key *(*md5_lookup) (const struct sock *sk,
2255 const struct sock *addr_sk);
2256 int (*calc_md5_hash)(char *location,
2257 const struct tcp_md5sig_key *md5,
2258 const struct sock *sk,
2259 const struct sk_buff *skb);
2260 int (*md5_parse)(struct sock *sk,
2261 int optname,
2262 sockptr_t optval,
2263 int optlen);
2264#endif
2265#ifdef CONFIG_TCP_AO
2266 int (*ao_parse)(struct sock *sk, int optname, sockptr_t optval, int optlen);
2267 struct tcp_ao_key *(*ao_lookup)(const struct sock *sk,
2268 struct sock *addr_sk,
2269 int sndid, int rcvid);
2270 int (*ao_calc_key_sk)(struct tcp_ao_key *mkt, u8 *key,
2271 const struct sock *sk,
2272 __be32 sisn, __be32 disn, bool send);
2273 int (*calc_ao_hash)(char *location, struct tcp_ao_key *ao,
2274 const struct sock *sk, const struct sk_buff *skb,
2275 const u8 *tkey, int hash_offset, u32 sne);
2276#endif
2277};
2278
2279struct tcp_request_sock_ops {
2280 u16 mss_clamp;
2281#ifdef CONFIG_TCP_MD5SIG
2282 struct tcp_md5sig_key *(*req_md5_lookup)(const struct sock *sk,
2283 const struct sock *addr_sk);
2284 int (*calc_md5_hash) (char *location,
2285 const struct tcp_md5sig_key *md5,
2286 const struct sock *sk,
2287 const struct sk_buff *skb);
2288#endif
2289#ifdef CONFIG_TCP_AO
2290 struct tcp_ao_key *(*ao_lookup)(const struct sock *sk,
2291 struct request_sock *req,
2292 int sndid, int rcvid);
2293 int (*ao_calc_key)(struct tcp_ao_key *mkt, u8 *key, struct request_sock *sk);
2294 int (*ao_synack_hash)(char *ao_hash, struct tcp_ao_key *mkt,
2295 struct request_sock *req, const struct sk_buff *skb,
2296 int hash_offset, u32 sne);
2297#endif
2298#ifdef CONFIG_SYN_COOKIES
2299 __u32 (*cookie_init_seq)(const struct sk_buff *skb,
2300 __u16 *mss);
2301#endif
2302 struct dst_entry *(*route_req)(const struct sock *sk,
2303 struct sk_buff *skb,
2304 struct flowi *fl,
2305 struct request_sock *req,
2306 u32 tw_isn);
2307 u32 (*init_seq)(const struct sk_buff *skb);
2308 u32 (*init_ts_off)(const struct net *net, const struct sk_buff *skb);
2309 int (*send_synack)(const struct sock *sk, struct dst_entry *dst,
2310 struct flowi *fl, struct request_sock *req,
2311 struct tcp_fastopen_cookie *foc,
2312 enum tcp_synack_type synack_type,
2313 struct sk_buff *syn_skb);
2314};
2315
2316extern const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops;
2317#if IS_ENABLED(CONFIG_IPV6)
2318extern const struct tcp_request_sock_ops tcp_request_sock_ipv6_ops;
2319#endif
2320
2321#ifdef CONFIG_SYN_COOKIES
2322static inline __u32 cookie_init_sequence(const struct tcp_request_sock_ops *ops,
2323 const struct sock *sk, struct sk_buff *skb,
2324 __u16 *mss)
2325{
2326 tcp_synq_overflow(sk);
2327 __NET_INC_STATS(sock_net(sk), LINUX_MIB_SYNCOOKIESSENT);
2328 return ops->cookie_init_seq(skb, mss);
2329}
2330#else
2331static inline __u32 cookie_init_sequence(const struct tcp_request_sock_ops *ops,
2332 const struct sock *sk, struct sk_buff *skb,
2333 __u16 *mss)
2334{
2335 return 0;
2336}
2337#endif
2338
2339struct tcp_key {
2340 union {
2341 struct {
2342 struct tcp_ao_key *ao_key;
2343 char *traffic_key;
2344 u32 sne;
2345 u8 rcv_next;
2346 };
2347 struct tcp_md5sig_key *md5_key;
2348 };
2349 enum {
2350 TCP_KEY_NONE = 0,
2351 TCP_KEY_MD5,
2352 TCP_KEY_AO,
2353 } type;
2354};
2355
2356static inline void tcp_get_current_key(const struct sock *sk,
2357 struct tcp_key *out)
2358{
2359#if defined(CONFIG_TCP_AO) || defined(CONFIG_TCP_MD5SIG)
2360 const struct tcp_sock *tp = tcp_sk(sk);
2361#endif
2362
2363#ifdef CONFIG_TCP_AO
2364 if (static_branch_unlikely(&tcp_ao_needed.key)) {
2365 struct tcp_ao_info *ao;
2366
2367 ao = rcu_dereference_protected(tp->ao_info,
2368 lockdep_sock_is_held(sk));
2369 if (ao) {
2370 out->ao_key = READ_ONCE(ao->current_key);
2371 out->type = TCP_KEY_AO;
2372 return;
2373 }
2374 }
2375#endif
2376#ifdef CONFIG_TCP_MD5SIG
2377 if (static_branch_unlikely(&tcp_md5_needed.key) &&
2378 rcu_access_pointer(tp->md5sig_info)) {
2379 out->md5_key = tp->af_specific->md5_lookup(sk, sk);
2380 if (out->md5_key) {
2381 out->type = TCP_KEY_MD5;
2382 return;
2383 }
2384 }
2385#endif
2386 out->type = TCP_KEY_NONE;
2387}
2388
2389static inline bool tcp_key_is_md5(const struct tcp_key *key)
2390{
2391 if (static_branch_tcp_md5())
2392 return key->type == TCP_KEY_MD5;
2393 return false;
2394}
2395
2396static inline bool tcp_key_is_ao(const struct tcp_key *key)
2397{
2398 if (static_branch_tcp_ao())
2399 return key->type == TCP_KEY_AO;
2400 return false;
2401}
2402
2403int tcpv4_offload_init(void);
2404
2405void tcp_v4_init(void);
2406void tcp_init(void);
2407
2408/* tcp_recovery.c */
2409void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb);
2410void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced);
2411extern s32 tcp_rack_skb_timeout(struct tcp_sock *tp, struct sk_buff *skb,
2412 u32 reo_wnd);
2413extern bool tcp_rack_mark_lost(struct sock *sk);
2414extern void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
2415 u64 xmit_time);
2416extern void tcp_rack_reo_timeout(struct sock *sk);
2417extern void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs);
2418
2419/* tcp_plb.c */
2420
2421/*
2422 * Scaling factor for fractions in PLB. For example, tcp_plb_update_state
2423 * expects cong_ratio which represents fraction of traffic that experienced
2424 * congestion over a single RTT. In order to avoid floating point operations,
2425 * this fraction should be mapped to (1 << TCP_PLB_SCALE) and passed in.
2426 */
2427#define TCP_PLB_SCALE 8
2428
2429/* State for PLB (Protective Load Balancing) for a single TCP connection. */
2430struct tcp_plb_state {
2431 u8 consec_cong_rounds:5, /* consecutive congested rounds */
2432 unused:3;
2433 u32 pause_until; /* jiffies32 when PLB can resume rerouting */
2434};
2435
2436static inline void tcp_plb_init(const struct sock *sk,
2437 struct tcp_plb_state *plb)
2438{
2439 plb->consec_cong_rounds = 0;
2440 plb->pause_until = 0;
2441}
2442void tcp_plb_update_state(const struct sock *sk, struct tcp_plb_state *plb,
2443 const int cong_ratio);
2444void tcp_plb_check_rehash(struct sock *sk, struct tcp_plb_state *plb);
2445void tcp_plb_update_state_upon_rto(struct sock *sk, struct tcp_plb_state *plb);
2446
2447static inline void tcp_warn_once(const struct sock *sk, bool cond, const char *str)
2448{
2449 WARN_ONCE(cond,
2450 "%scwn:%u out:%u sacked:%u lost:%u retrans:%u tlp_high_seq:%u sk_state:%u ca_state:%u advmss:%u mss_cache:%u pmtu:%u\n",
2451 str,
2452 tcp_snd_cwnd(tcp_sk(sk)),
2453 tcp_sk(sk)->packets_out, tcp_sk(sk)->sacked_out,
2454 tcp_sk(sk)->lost_out, tcp_sk(sk)->retrans_out,
2455 tcp_sk(sk)->tlp_high_seq, sk->sk_state,
2456 inet_csk(sk)->icsk_ca_state,
2457 tcp_sk(sk)->advmss, tcp_sk(sk)->mss_cache,
2458 inet_csk(sk)->icsk_pmtu_cookie);
2459}
2460
2461/* At how many usecs into the future should the RTO fire? */
2462static inline s64 tcp_rto_delta_us(const struct sock *sk)
2463{
2464 const struct sk_buff *skb = tcp_rtx_queue_head(sk);
2465 u32 rto = inet_csk(sk)->icsk_rto;
2466
2467 if (likely(skb)) {
2468 u64 rto_time_stamp_us = tcp_skb_timestamp_us(skb) + jiffies_to_usecs(rto);
2469
2470 return rto_time_stamp_us - tcp_sk(sk)->tcp_mstamp;
2471 } else {
2472 tcp_warn_once(sk, 1, "rtx queue empty: ");
2473 return jiffies_to_usecs(rto);
2474 }
2475
2476}
2477
2478/*
2479 * Save and compile IPv4 options, return a pointer to it
2480 */
2481static inline struct ip_options_rcu *tcp_v4_save_options(struct net *net,
2482 struct sk_buff *skb)
2483{
2484 const struct ip_options *opt = &TCP_SKB_CB(skb)->header.h4.opt;
2485 struct ip_options_rcu *dopt = NULL;
2486
2487 if (opt->optlen) {
2488 int opt_size = sizeof(*dopt) + opt->optlen;
2489
2490 dopt = kmalloc(opt_size, GFP_ATOMIC);
2491 if (dopt && __ip_options_echo(net, &dopt->opt, skb, opt)) {
2492 kfree(dopt);
2493 dopt = NULL;
2494 }
2495 }
2496 return dopt;
2497}
2498
2499/* locally generated TCP pure ACKs have skb->truesize == 2
2500 * (check tcp_send_ack() in net/ipv4/tcp_output.c )
2501 * This is much faster than dissecting the packet to find out.
2502 * (Think of GRE encapsulations, IPv4, IPv6, ...)
2503 */
2504static inline bool skb_is_tcp_pure_ack(const struct sk_buff *skb)
2505{
2506 return skb->truesize == 2;
2507}
2508
2509static inline void skb_set_tcp_pure_ack(struct sk_buff *skb)
2510{
2511 skb->truesize = 2;
2512}
2513
2514static inline int tcp_inq(struct sock *sk)
2515{
2516 struct tcp_sock *tp = tcp_sk(sk);
2517 int answ;
2518
2519 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
2520 answ = 0;
2521 } else if (sock_flag(sk, SOCK_URGINLINE) ||
2522 !tp->urg_data ||
2523 before(tp->urg_seq, tp->copied_seq) ||
2524 !before(tp->urg_seq, tp->rcv_nxt)) {
2525
2526 answ = tp->rcv_nxt - tp->copied_seq;
2527
2528 /* Subtract 1, if FIN was received */
2529 if (answ && sock_flag(sk, SOCK_DONE))
2530 answ--;
2531 } else {
2532 answ = tp->urg_seq - tp->copied_seq;
2533 }
2534
2535 return answ;
2536}
2537
2538int tcp_peek_len(struct socket *sock);
2539
2540static inline void tcp_segs_in(struct tcp_sock *tp, const struct sk_buff *skb)
2541{
2542 u16 segs_in;
2543
2544 segs_in = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
2545
2546 /* We update these fields while other threads might
2547 * read them from tcp_get_info()
2548 */
2549 WRITE_ONCE(tp->segs_in, tp->segs_in + segs_in);
2550 if (skb->len > tcp_hdrlen(skb))
2551 WRITE_ONCE(tp->data_segs_in, tp->data_segs_in + segs_in);
2552}
2553
2554/*
2555 * TCP listen path runs lockless.
2556 * We forced "struct sock" to be const qualified to make sure
2557 * we don't modify one of its field by mistake.
2558 * Here, we increment sk_drops which is an atomic_t, so we can safely
2559 * make sock writable again.
2560 */
2561static inline void tcp_listendrop(const struct sock *sk)
2562{
2563 atomic_inc(&((struct sock *)sk)->sk_drops);
2564 __NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENDROPS);
2565}
2566
2567enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer);
2568
2569/*
2570 * Interface for adding Upper Level Protocols over TCP
2571 */
2572
2573#define TCP_ULP_NAME_MAX 16
2574#define TCP_ULP_MAX 128
2575#define TCP_ULP_BUF_MAX (TCP_ULP_NAME_MAX*TCP_ULP_MAX)
2576
2577struct tcp_ulp_ops {
2578 struct list_head list;
2579
2580 /* initialize ulp */
2581 int (*init)(struct sock *sk);
2582 /* update ulp */
2583 void (*update)(struct sock *sk, struct proto *p,
2584 void (*write_space)(struct sock *sk));
2585 /* cleanup ulp */
2586 void (*release)(struct sock *sk);
2587 /* diagnostic */
2588 int (*get_info)(struct sock *sk, struct sk_buff *skb);
2589 size_t (*get_info_size)(const struct sock *sk);
2590 /* clone ulp */
2591 void (*clone)(const struct request_sock *req, struct sock *newsk,
2592 const gfp_t priority);
2593
2594 char name[TCP_ULP_NAME_MAX];
2595 struct module *owner;
2596};
2597int tcp_register_ulp(struct tcp_ulp_ops *type);
2598void tcp_unregister_ulp(struct tcp_ulp_ops *type);
2599int tcp_set_ulp(struct sock *sk, const char *name);
2600void tcp_get_available_ulp(char *buf, size_t len);
2601void tcp_cleanup_ulp(struct sock *sk);
2602void tcp_update_ulp(struct sock *sk, struct proto *p,
2603 void (*write_space)(struct sock *sk));
2604
2605#define MODULE_ALIAS_TCP_ULP(name) \
2606 __MODULE_INFO(alias, alias_userspace, name); \
2607 __MODULE_INFO(alias, alias_tcp_ulp, "tcp-ulp-" name)
2608
2609#ifdef CONFIG_NET_SOCK_MSG
2610struct sk_msg;
2611struct sk_psock;
2612
2613#ifdef CONFIG_BPF_SYSCALL
2614int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore);
2615void tcp_bpf_clone(const struct sock *sk, struct sock *newsk);
2616#endif /* CONFIG_BPF_SYSCALL */
2617
2618#ifdef CONFIG_INET
2619void tcp_eat_skb(struct sock *sk, struct sk_buff *skb);
2620#else
2621static inline void tcp_eat_skb(struct sock *sk, struct sk_buff *skb)
2622{
2623}
2624#endif
2625
2626int tcp_bpf_sendmsg_redir(struct sock *sk, bool ingress,
2627 struct sk_msg *msg, u32 bytes, int flags);
2628#endif /* CONFIG_NET_SOCK_MSG */
2629
2630#if !defined(CONFIG_BPF_SYSCALL) || !defined(CONFIG_NET_SOCK_MSG)
2631static inline void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
2632{
2633}
2634#endif
2635
2636#ifdef CONFIG_CGROUP_BPF
2637static inline void bpf_skops_init_skb(struct bpf_sock_ops_kern *skops,
2638 struct sk_buff *skb,
2639 unsigned int end_offset)
2640{
2641 skops->skb = skb;
2642 skops->skb_data_end = skb->data + end_offset;
2643}
2644#else
2645static inline void bpf_skops_init_skb(struct bpf_sock_ops_kern *skops,
2646 struct sk_buff *skb,
2647 unsigned int end_offset)
2648{
2649}
2650#endif
2651
2652/* Call BPF_SOCK_OPS program that returns an int. If the return value
2653 * is < 0, then the BPF op failed (for example if the loaded BPF
2654 * program does not support the chosen operation or there is no BPF
2655 * program loaded).
2656 */
2657#ifdef CONFIG_BPF
2658static inline int tcp_call_bpf(struct sock *sk, int op, u32 nargs, u32 *args)
2659{
2660 struct bpf_sock_ops_kern sock_ops;
2661 int ret;
2662
2663 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
2664 if (sk_fullsock(sk)) {
2665 sock_ops.is_fullsock = 1;
2666 sock_owned_by_me(sk);
2667 }
2668
2669 sock_ops.sk = sk;
2670 sock_ops.op = op;
2671 if (nargs > 0)
2672 memcpy(sock_ops.args, args, nargs * sizeof(*args));
2673
2674 ret = BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
2675 if (ret == 0)
2676 ret = sock_ops.reply;
2677 else
2678 ret = -1;
2679 return ret;
2680}
2681
2682static inline int tcp_call_bpf_2arg(struct sock *sk, int op, u32 arg1, u32 arg2)
2683{
2684 u32 args[2] = {arg1, arg2};
2685
2686 return tcp_call_bpf(sk, op, 2, args);
2687}
2688
2689static inline int tcp_call_bpf_3arg(struct sock *sk, int op, u32 arg1, u32 arg2,
2690 u32 arg3)
2691{
2692 u32 args[3] = {arg1, arg2, arg3};
2693
2694 return tcp_call_bpf(sk, op, 3, args);
2695}
2696
2697#else
2698static inline int tcp_call_bpf(struct sock *sk, int op, u32 nargs, u32 *args)
2699{
2700 return -EPERM;
2701}
2702
2703static inline int tcp_call_bpf_2arg(struct sock *sk, int op, u32 arg1, u32 arg2)
2704{
2705 return -EPERM;
2706}
2707
2708static inline int tcp_call_bpf_3arg(struct sock *sk, int op, u32 arg1, u32 arg2,
2709 u32 arg3)
2710{
2711 return -EPERM;
2712}
2713
2714#endif
2715
2716static inline u32 tcp_timeout_init(struct sock *sk)
2717{
2718 int timeout;
2719
2720 timeout = tcp_call_bpf(sk, BPF_SOCK_OPS_TIMEOUT_INIT, 0, NULL);
2721
2722 if (timeout <= 0)
2723 timeout = TCP_TIMEOUT_INIT;
2724 return min_t(int, timeout, TCP_RTO_MAX);
2725}
2726
2727static inline u32 tcp_rwnd_init_bpf(struct sock *sk)
2728{
2729 int rwnd;
2730
2731 rwnd = tcp_call_bpf(sk, BPF_SOCK_OPS_RWND_INIT, 0, NULL);
2732
2733 if (rwnd < 0)
2734 rwnd = 0;
2735 return rwnd;
2736}
2737
2738static inline bool tcp_bpf_ca_needs_ecn(struct sock *sk)
2739{
2740 return (tcp_call_bpf(sk, BPF_SOCK_OPS_NEEDS_ECN, 0, NULL) == 1);
2741}
2742
2743static inline void tcp_bpf_rtt(struct sock *sk, long mrtt, u32 srtt)
2744{
2745 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_RTT_CB_FLAG))
2746 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_RTT_CB, mrtt, srtt);
2747}
2748
2749#if IS_ENABLED(CONFIG_SMC)
2750extern struct static_key_false tcp_have_smc;
2751#endif
2752
2753#if IS_ENABLED(CONFIG_TLS_DEVICE)
2754void clean_acked_data_enable(struct inet_connection_sock *icsk,
2755 void (*cad)(struct sock *sk, u32 ack_seq));
2756void clean_acked_data_disable(struct inet_connection_sock *icsk);
2757void clean_acked_data_flush(void);
2758#endif
2759
2760DECLARE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2761static inline void tcp_add_tx_delay(struct sk_buff *skb,
2762 const struct tcp_sock *tp)
2763{
2764 if (static_branch_unlikely(&tcp_tx_delay_enabled))
2765 skb->skb_mstamp_ns += (u64)tp->tcp_tx_delay * NSEC_PER_USEC;
2766}
2767
2768/* Compute Earliest Departure Time for some control packets
2769 * like ACK or RST for TIME_WAIT or non ESTABLISHED sockets.
2770 */
2771static inline u64 tcp_transmit_time(const struct sock *sk)
2772{
2773 if (static_branch_unlikely(&tcp_tx_delay_enabled)) {
2774 u32 delay = (sk->sk_state == TCP_TIME_WAIT) ?
2775 tcp_twsk(sk)->tw_tx_delay : tcp_sk(sk)->tcp_tx_delay;
2776
2777 return tcp_clock_ns() + (u64)delay * NSEC_PER_USEC;
2778 }
2779 return 0;
2780}
2781
2782static inline int tcp_parse_auth_options(const struct tcphdr *th,
2783 const u8 **md5_hash, const struct tcp_ao_hdr **aoh)
2784{
2785 const u8 *md5_tmp, *ao_tmp;
2786 int ret;
2787
2788 ret = tcp_do_parse_auth_options(th, &md5_tmp, &ao_tmp);
2789 if (ret)
2790 return ret;
2791
2792 if (md5_hash)
2793 *md5_hash = md5_tmp;
2794
2795 if (aoh) {
2796 if (!ao_tmp)
2797 *aoh = NULL;
2798 else
2799 *aoh = (struct tcp_ao_hdr *)(ao_tmp - 2);
2800 }
2801
2802 return 0;
2803}
2804
2805static inline bool tcp_ao_required(struct sock *sk, const void *saddr,
2806 int family, int l3index, bool stat_inc)
2807{
2808#ifdef CONFIG_TCP_AO
2809 struct tcp_ao_info *ao_info;
2810 struct tcp_ao_key *ao_key;
2811
2812 if (!static_branch_unlikely(&tcp_ao_needed.key))
2813 return false;
2814
2815 ao_info = rcu_dereference_check(tcp_sk(sk)->ao_info,
2816 lockdep_sock_is_held(sk));
2817 if (!ao_info)
2818 return false;
2819
2820 ao_key = tcp_ao_do_lookup(sk, l3index, saddr, family, -1, -1);
2821 if (ao_info->ao_required || ao_key) {
2822 if (stat_inc) {
2823 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAOREQUIRED);
2824 atomic64_inc(&ao_info->counters.ao_required);
2825 }
2826 return true;
2827 }
2828#endif
2829 return false;
2830}
2831
2832enum skb_drop_reason tcp_inbound_hash(struct sock *sk,
2833 const struct request_sock *req, const struct sk_buff *skb,
2834 const void *saddr, const void *daddr,
2835 int family, int dif, int sdif);
2836
2837#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/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 */