1 /* SPDX-License-Identifier: GPL-2.0-or-later */
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.
7 * Definitions for the TCP module.
9 * Version: @(#)tcp.h 1.0.5 05/23/93
12 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
17 #define FASTRETRANS_DEBUG 1
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>
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>
39 #include <net/tcp_states.h>
40 #include <net/inet_ecn.h>
42 #include <net/mptcp.h>
44 #include <linux/seq_file.h>
45 #include <linux/memcontrol.h>
46 #include <linux/bpf-cgroup.h>
47 #include <linux/siphash.h>
49 extern struct inet_hashinfo tcp_hashinfo;
51 DECLARE_PER_CPU(unsigned int, tcp_orphan_count);
52 int tcp_orphan_count_sum(void);
54 void tcp_time_wait(struct sock *sk, int state, int timeo);
56 #define MAX_TCP_HEADER L1_CACHE_ALIGN(128 + MAX_HEADER)
57 #define MAX_TCP_OPTION_SPACE 40
58 #define TCP_MIN_SND_MSS 48
59 #define TCP_MIN_GSO_SIZE (TCP_MIN_SND_MSS - MAX_TCP_OPTION_SPACE)
62 * Never offer a window over 32767 without using window scaling. Some
63 * poor stacks do signed 16bit maths!
65 #define MAX_TCP_WINDOW 32767U
67 /* Minimal accepted MSS. It is (60+60+8) - (20+20). */
68 #define TCP_MIN_MSS 88U
70 /* The initial MTU to use for probing */
71 #define TCP_BASE_MSS 1024
73 /* probing interval, default to 10 minutes as per RFC4821 */
74 #define TCP_PROBE_INTERVAL 600
76 /* Specify interval when tcp mtu probing will stop */
77 #define TCP_PROBE_THRESHOLD 8
79 /* After receiving this amount of duplicate ACKs fast retransmit starts. */
80 #define TCP_FASTRETRANS_THRESH 3
82 /* Maximal number of ACKs sent quickly to accelerate slow-start. */
83 #define TCP_MAX_QUICKACKS 16U
85 /* Maximal number of window scale according to RFC1323 */
86 #define TCP_MAX_WSCALE 14U
89 #define TCP_URG_VALID 0x0100
90 #define TCP_URG_NOTYET 0x0200
91 #define TCP_URG_READ 0x0400
93 #define TCP_RETR1 3 /*
94 * This is how many retries it does before it
95 * tries to figure out if the gateway is
96 * down. Minimal RFC value is 3; it corresponds
97 * to ~3sec-8min depending on RTO.
100 #define TCP_RETR2 15 /*
101 * This should take at least
102 * 90 minutes to time out.
103 * RFC1122 says that the limit is 100 sec.
104 * 15 is ~13-30min depending on RTO.
107 #define TCP_SYN_RETRIES 6 /* This is how many retries are done
108 * when active opening a connection.
109 * RFC1122 says the minimum retry MUST
110 * be at least 180secs. Nevertheless
111 * this value is corresponding to
112 * 63secs of retransmission with the
113 * current initial RTO.
116 #define TCP_SYNACK_RETRIES 5 /* This is how may retries are done
117 * when passive opening a connection.
118 * This is corresponding to 31secs of
119 * retransmission with the current
123 #define TCP_TIMEWAIT_LEN (60*HZ) /* how long to wait to destroy TIME-WAIT
124 * state, about 60 seconds */
125 #define TCP_FIN_TIMEOUT TCP_TIMEWAIT_LEN
126 /* BSD style FIN_WAIT2 deadlock breaker.
127 * It used to be 3min, new value is 60sec,
128 * to combine FIN-WAIT-2 timeout with
131 #define TCP_FIN_TIMEOUT_MAX (120 * HZ) /* max TCP_LINGER2 value (two minutes) */
133 #define TCP_DELACK_MAX ((unsigned)(HZ/5)) /* maximal time to delay before sending an ACK */
135 #define TCP_DELACK_MIN ((unsigned)(HZ/25)) /* minimal time to delay before sending an ACK */
136 #define TCP_ATO_MIN ((unsigned)(HZ/25))
138 #define TCP_DELACK_MIN 4U
139 #define TCP_ATO_MIN 4U
141 #define TCP_RTO_MAX ((unsigned)(120*HZ))
142 #define TCP_RTO_MIN ((unsigned)(HZ/5))
143 #define TCP_TIMEOUT_MIN (2U) /* Min timeout for TCP timers in jiffies */
145 #define TCP_TIMEOUT_MIN_US (2*USEC_PER_MSEC) /* Min TCP timeout in microsecs */
147 #define TCP_TIMEOUT_INIT ((unsigned)(1*HZ)) /* RFC6298 2.1 initial RTO value */
148 #define TCP_TIMEOUT_FALLBACK ((unsigned)(3*HZ)) /* RFC 1122 initial RTO value, now
149 * used as a fallback RTO for the
150 * initial data transmission if no
151 * valid RTT sample has been acquired,
152 * most likely due to retrans in 3WHS.
155 #define TCP_RESOURCE_PROBE_INTERVAL ((unsigned)(HZ/2U)) /* Maximal interval between probes
156 * for local resources.
158 #define TCP_KEEPALIVE_TIME (120*60*HZ) /* two hours */
159 #define TCP_KEEPALIVE_PROBES 9 /* Max of 9 keepalive probes */
160 #define TCP_KEEPALIVE_INTVL (75*HZ)
162 #define MAX_TCP_KEEPIDLE 32767
163 #define MAX_TCP_KEEPINTVL 32767
164 #define MAX_TCP_KEEPCNT 127
165 #define MAX_TCP_SYNCNT 127
167 #define TCP_PAWS_24DAYS (60 * 60 * 24 * 24)
168 #define TCP_PAWS_MSL 60 /* Per-host timestamps are invalidated
169 * after this time. It should be equal
170 * (or greater than) TCP_TIMEWAIT_LEN
171 * to provide reliability equal to one
172 * provided by timewait state.
174 #define TCP_PAWS_WINDOW 1 /* Replay window for per-host
175 * timestamps. It must be less than
176 * minimal timewait lifetime.
182 #define TCPOPT_NOP 1 /* Padding */
183 #define TCPOPT_EOL 0 /* End of options */
184 #define TCPOPT_MSS 2 /* Segment size negotiating */
185 #define TCPOPT_WINDOW 3 /* Window scaling */
186 #define TCPOPT_SACK_PERM 4 /* SACK Permitted */
187 #define TCPOPT_SACK 5 /* SACK Block */
188 #define TCPOPT_TIMESTAMP 8 /* Better RTT estimations/PAWS */
189 #define TCPOPT_MD5SIG 19 /* MD5 Signature (RFC2385) */
190 #define TCPOPT_MPTCP 30 /* Multipath TCP (RFC6824) */
191 #define TCPOPT_FASTOPEN 34 /* Fast open (RFC7413) */
192 #define TCPOPT_EXP 254 /* Experimental */
193 /* Magic number to be after the option value for sharing TCP
194 * experimental options. See draft-ietf-tcpm-experimental-options-00.txt
196 #define TCPOPT_FASTOPEN_MAGIC 0xF989
197 #define TCPOPT_SMC_MAGIC 0xE2D4C3D9
203 #define TCPOLEN_MSS 4
204 #define TCPOLEN_WINDOW 3
205 #define TCPOLEN_SACK_PERM 2
206 #define TCPOLEN_TIMESTAMP 10
207 #define TCPOLEN_MD5SIG 18
208 #define TCPOLEN_FASTOPEN_BASE 2
209 #define TCPOLEN_EXP_FASTOPEN_BASE 4
210 #define TCPOLEN_EXP_SMC_BASE 6
212 /* But this is what stacks really send out. */
213 #define TCPOLEN_TSTAMP_ALIGNED 12
214 #define TCPOLEN_WSCALE_ALIGNED 4
215 #define TCPOLEN_SACKPERM_ALIGNED 4
216 #define TCPOLEN_SACK_BASE 2
217 #define TCPOLEN_SACK_BASE_ALIGNED 4
218 #define TCPOLEN_SACK_PERBLOCK 8
219 #define TCPOLEN_MD5SIG_ALIGNED 20
220 #define TCPOLEN_MSS_ALIGNED 4
221 #define TCPOLEN_EXP_SMC_BASE_ALIGNED 8
223 /* Flags in tp->nonagle */
224 #define TCP_NAGLE_OFF 1 /* Nagle's algo is disabled */
225 #define TCP_NAGLE_CORK 2 /* Socket is corked */
226 #define TCP_NAGLE_PUSH 4 /* Cork is overridden for already queued data */
228 /* TCP thin-stream limits */
229 #define TCP_THIN_LINEAR_RETRIES 6 /* After 6 linear retries, do exp. backoff */
231 /* TCP initial congestion window as per rfc6928 */
232 #define TCP_INIT_CWND 10
234 /* Bit Flags for sysctl_tcp_fastopen */
235 #define TFO_CLIENT_ENABLE 1
236 #define TFO_SERVER_ENABLE 2
237 #define TFO_CLIENT_NO_COOKIE 4 /* Data in SYN w/o cookie option */
239 /* Accept SYN data w/o any cookie option */
240 #define TFO_SERVER_COOKIE_NOT_REQD 0x200
242 /* Force enable TFO on all listeners, i.e., not requiring the
243 * TCP_FASTOPEN socket option.
245 #define TFO_SERVER_WO_SOCKOPT1 0x400
248 /* sysctl variables for tcp */
249 extern int sysctl_tcp_max_orphans;
250 extern long sysctl_tcp_mem[3];
252 #define TCP_RACK_LOSS_DETECTION 0x1 /* Use RACK to detect losses */
253 #define TCP_RACK_STATIC_REO_WND 0x2 /* Use static RACK reo wnd */
254 #define TCP_RACK_NO_DUPTHRESH 0x4 /* Do not use DUPACK threshold in RACK */
256 extern atomic_long_t tcp_memory_allocated;
257 DECLARE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
259 extern struct percpu_counter tcp_sockets_allocated;
260 extern unsigned long tcp_memory_pressure;
262 /* optimized version of sk_under_memory_pressure() for TCP sockets */
263 static inline bool tcp_under_memory_pressure(const struct sock *sk)
265 if (mem_cgroup_sockets_enabled && sk->sk_memcg &&
266 mem_cgroup_under_socket_pressure(sk->sk_memcg))
269 return READ_ONCE(tcp_memory_pressure);
272 * The next routines deal with comparing 32 bit unsigned ints
273 * and worry about wraparound (automatic with unsigned arithmetic).
276 static inline bool before(__u32 seq1, __u32 seq2)
278 return (__s32)(seq1-seq2) < 0;
280 #define after(seq2, seq1) before(seq1, seq2)
282 /* is s2<=s1<=s3 ? */
283 static inline bool between(__u32 seq1, __u32 seq2, __u32 seq3)
285 return seq3 - seq2 >= seq1 - seq2;
288 static inline bool tcp_out_of_memory(struct sock *sk)
290 if (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
291 sk_memory_allocated(sk) > sk_prot_mem_limits(sk, 2))
296 static inline void tcp_wmem_free_skb(struct sock *sk, struct sk_buff *skb)
298 sk_wmem_queued_add(sk, -skb->truesize);
299 if (!skb_zcopy_pure(skb))
300 sk_mem_uncharge(sk, skb->truesize);
302 sk_mem_uncharge(sk, SKB_TRUESIZE(skb_end_offset(skb)));
306 void sk_forced_mem_schedule(struct sock *sk, int size);
308 bool tcp_check_oom(struct sock *sk, int shift);
311 extern struct proto tcp_prot;
313 #define TCP_INC_STATS(net, field) SNMP_INC_STATS((net)->mib.tcp_statistics, field)
314 #define __TCP_INC_STATS(net, field) __SNMP_INC_STATS((net)->mib.tcp_statistics, field)
315 #define TCP_DEC_STATS(net, field) SNMP_DEC_STATS((net)->mib.tcp_statistics, field)
316 #define TCP_ADD_STATS(net, field, val) SNMP_ADD_STATS((net)->mib.tcp_statistics, field, val)
318 void tcp_tasklet_init(void);
320 int tcp_v4_err(struct sk_buff *skb, u32);
322 void tcp_shutdown(struct sock *sk, int how);
324 int tcp_v4_early_demux(struct sk_buff *skb);
325 int tcp_v4_rcv(struct sk_buff *skb);
327 void tcp_remove_empty_skb(struct sock *sk);
328 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size);
329 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size);
330 int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
331 size_t size, struct ubuf_info *uarg);
332 void tcp_splice_eof(struct socket *sock);
333 int tcp_send_mss(struct sock *sk, int *size_goal, int flags);
334 int tcp_wmem_schedule(struct sock *sk, int copy);
335 void tcp_push(struct sock *sk, int flags, int mss_now, int nonagle,
337 void tcp_release_cb(struct sock *sk);
338 void tcp_wfree(struct sk_buff *skb);
339 void tcp_write_timer_handler(struct sock *sk);
340 void tcp_delack_timer_handler(struct sock *sk);
341 int tcp_ioctl(struct sock *sk, int cmd, int *karg);
342 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb);
343 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb);
344 void tcp_rcv_space_adjust(struct sock *sk);
345 int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp);
346 void tcp_twsk_destructor(struct sock *sk);
347 void tcp_twsk_purge(struct list_head *net_exit_list, int family);
348 ssize_t tcp_splice_read(struct socket *sk, loff_t *ppos,
349 struct pipe_inode_info *pipe, size_t len,
351 struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
352 bool force_schedule);
354 static inline void tcp_dec_quickack_mode(struct sock *sk)
356 struct inet_connection_sock *icsk = inet_csk(sk);
358 if (icsk->icsk_ack.quick) {
359 /* How many ACKs S/ACKing new data have we sent? */
360 const unsigned int pkts = inet_csk_ack_scheduled(sk) ? 1 : 0;
362 if (pkts >= icsk->icsk_ack.quick) {
363 icsk->icsk_ack.quick = 0;
364 /* Leaving quickack mode we deflate ATO. */
365 icsk->icsk_ack.ato = TCP_ATO_MIN;
367 icsk->icsk_ack.quick -= pkts;
372 #define TCP_ECN_QUEUE_CWR 2
373 #define TCP_ECN_DEMAND_CWR 4
374 #define TCP_ECN_SEEN 8
384 enum tcp_tw_status tcp_timewait_state_process(struct inet_timewait_sock *tw,
386 const struct tcphdr *th);
387 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
388 struct request_sock *req, bool fastopen,
390 int tcp_child_process(struct sock *parent, struct sock *child,
391 struct sk_buff *skb);
392 void tcp_enter_loss(struct sock *sk);
393 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag);
394 void tcp_clear_retrans(struct tcp_sock *tp);
395 void tcp_update_metrics(struct sock *sk);
396 void tcp_init_metrics(struct sock *sk);
397 void tcp_metrics_init(void);
398 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst);
399 void __tcp_close(struct sock *sk, long timeout);
400 void tcp_close(struct sock *sk, long timeout);
401 void tcp_init_sock(struct sock *sk);
402 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb);
403 __poll_t tcp_poll(struct file *file, struct socket *sock,
404 struct poll_table_struct *wait);
405 int do_tcp_getsockopt(struct sock *sk, int level,
406 int optname, sockptr_t optval, sockptr_t optlen);
407 int tcp_getsockopt(struct sock *sk, int level, int optname,
408 char __user *optval, int __user *optlen);
409 bool tcp_bpf_bypass_getsockopt(int level, int optname);
410 int do_tcp_setsockopt(struct sock *sk, int level, int optname,
411 sockptr_t optval, unsigned int optlen);
412 int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
413 unsigned int optlen);
414 void tcp_set_keepalive(struct sock *sk, int val);
415 void tcp_syn_ack_timeout(const struct request_sock *req);
416 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
417 int flags, int *addr_len);
418 int tcp_set_rcvlowat(struct sock *sk, int val);
419 int tcp_set_window_clamp(struct sock *sk, int val);
420 void tcp_update_recv_tstamps(struct sk_buff *skb,
421 struct scm_timestamping_internal *tss);
422 void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
423 struct scm_timestamping_internal *tss);
424 void tcp_data_ready(struct sock *sk);
426 int tcp_mmap(struct file *file, struct socket *sock,
427 struct vm_area_struct *vma);
429 void tcp_parse_options(const struct net *net, const struct sk_buff *skb,
430 struct tcp_options_received *opt_rx,
431 int estab, struct tcp_fastopen_cookie *foc);
432 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th);
435 * BPF SKB-less helpers
437 u16 tcp_v4_get_syncookie(struct sock *sk, struct iphdr *iph,
438 struct tcphdr *th, u32 *cookie);
439 u16 tcp_v6_get_syncookie(struct sock *sk, struct ipv6hdr *iph,
440 struct tcphdr *th, u32 *cookie);
441 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss);
442 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
443 const struct tcp_request_sock_ops *af_ops,
444 struct sock *sk, struct tcphdr *th);
446 * TCP v4 functions exported for the inet6 API
449 void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb);
450 void tcp_v4_mtu_reduced(struct sock *sk);
451 void tcp_req_err(struct sock *sk, u32 seq, bool abort);
452 void tcp_ld_RTO_revert(struct sock *sk, u32 seq);
453 int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb);
454 struct sock *tcp_create_openreq_child(const struct sock *sk,
455 struct request_sock *req,
456 struct sk_buff *skb);
457 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst);
458 struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
459 struct request_sock *req,
460 struct dst_entry *dst,
461 struct request_sock *req_unhash,
463 int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb);
464 int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len);
465 int tcp_connect(struct sock *sk);
466 enum tcp_synack_type {
471 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
472 struct request_sock *req,
473 struct tcp_fastopen_cookie *foc,
474 enum tcp_synack_type synack_type,
475 struct sk_buff *syn_skb);
476 int tcp_disconnect(struct sock *sk, int flags);
478 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb);
479 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size);
480 void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb);
482 /* From syncookies.c */
483 struct sock *tcp_get_cookie_sock(struct sock *sk, struct sk_buff *skb,
484 struct request_sock *req,
485 struct dst_entry *dst, u32 tsoff);
486 int __cookie_v4_check(const struct iphdr *iph, const struct tcphdr *th,
488 struct sock *cookie_v4_check(struct sock *sk, struct sk_buff *skb);
489 struct request_sock *cookie_tcp_reqsk_alloc(const struct request_sock_ops *ops,
490 const struct tcp_request_sock_ops *af_ops,
491 struct sock *sk, struct sk_buff *skb);
492 #ifdef CONFIG_SYN_COOKIES
494 /* Syncookies use a monotonic timer which increments every 60 seconds.
495 * This counter is used both as a hash input and partially encoded into
496 * the cookie value. A cookie is only validated further if the delta
497 * between the current counter value and the encoded one is less than this,
498 * i.e. a sent cookie is valid only at most for 2*60 seconds (or less if
499 * the counter advances immediately after a cookie is generated).
501 #define MAX_SYNCOOKIE_AGE 2
502 #define TCP_SYNCOOKIE_PERIOD (60 * HZ)
503 #define TCP_SYNCOOKIE_VALID (MAX_SYNCOOKIE_AGE * TCP_SYNCOOKIE_PERIOD)
505 /* syncookies: remember time of last synqueue overflow
506 * But do not dirty this field too often (once per second is enough)
507 * It is racy as we do not hold a lock, but race is very minor.
509 static inline void tcp_synq_overflow(const struct sock *sk)
511 unsigned int last_overflow;
512 unsigned int now = jiffies;
514 if (sk->sk_reuseport) {
515 struct sock_reuseport *reuse;
517 reuse = rcu_dereference(sk->sk_reuseport_cb);
519 last_overflow = READ_ONCE(reuse->synq_overflow_ts);
520 if (!time_between32(now, last_overflow,
522 WRITE_ONCE(reuse->synq_overflow_ts, now);
527 last_overflow = READ_ONCE(tcp_sk(sk)->rx_opt.ts_recent_stamp);
528 if (!time_between32(now, last_overflow, last_overflow + HZ))
529 WRITE_ONCE(tcp_sk_rw(sk)->rx_opt.ts_recent_stamp, now);
532 /* syncookies: no recent synqueue overflow on this listening socket? */
533 static inline bool tcp_synq_no_recent_overflow(const struct sock *sk)
535 unsigned int last_overflow;
536 unsigned int now = jiffies;
538 if (sk->sk_reuseport) {
539 struct sock_reuseport *reuse;
541 reuse = rcu_dereference(sk->sk_reuseport_cb);
543 last_overflow = READ_ONCE(reuse->synq_overflow_ts);
544 return !time_between32(now, last_overflow - HZ,
546 TCP_SYNCOOKIE_VALID);
550 last_overflow = READ_ONCE(tcp_sk(sk)->rx_opt.ts_recent_stamp);
552 /* If last_overflow <= jiffies <= last_overflow + TCP_SYNCOOKIE_VALID,
553 * then we're under synflood. However, we have to use
554 * 'last_overflow - HZ' as lower bound. That's because a concurrent
555 * tcp_synq_overflow() could update .ts_recent_stamp after we read
556 * jiffies but before we store .ts_recent_stamp into last_overflow,
557 * which could lead to rejecting a valid syncookie.
559 return !time_between32(now, last_overflow - HZ,
560 last_overflow + TCP_SYNCOOKIE_VALID);
563 static inline u32 tcp_cookie_time(void)
565 u64 val = get_jiffies_64();
567 do_div(val, TCP_SYNCOOKIE_PERIOD);
571 u32 __cookie_v4_init_sequence(const struct iphdr *iph, const struct tcphdr *th,
573 __u32 cookie_v4_init_sequence(const struct sk_buff *skb, __u16 *mss);
574 u64 cookie_init_timestamp(struct request_sock *req, u64 now);
575 bool cookie_timestamp_decode(const struct net *net,
576 struct tcp_options_received *opt);
577 bool cookie_ecn_ok(const struct tcp_options_received *opt,
578 const struct net *net, const struct dst_entry *dst);
580 /* From net/ipv6/syncookies.c */
581 int __cookie_v6_check(const struct ipv6hdr *iph, const struct tcphdr *th,
583 struct sock *cookie_v6_check(struct sock *sk, struct sk_buff *skb);
585 u32 __cookie_v6_init_sequence(const struct ipv6hdr *iph,
586 const struct tcphdr *th, u16 *mssp);
587 __u32 cookie_v6_init_sequence(const struct sk_buff *skb, __u16 *mss);
591 void tcp_skb_entail(struct sock *sk, struct sk_buff *skb);
592 void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb);
593 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
595 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs);
596 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs);
597 void tcp_retransmit_timer(struct sock *sk);
598 void tcp_xmit_retransmit_queue(struct sock *);
599 void tcp_simple_retransmit(struct sock *);
600 void tcp_enter_recovery(struct sock *sk, bool ece_ack);
601 int tcp_trim_head(struct sock *, struct sk_buff *, u32);
603 TCP_FRAG_IN_WRITE_QUEUE,
604 TCP_FRAG_IN_RTX_QUEUE,
606 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
607 struct sk_buff *skb, u32 len,
608 unsigned int mss_now, gfp_t gfp);
610 void tcp_send_probe0(struct sock *);
611 int tcp_write_wakeup(struct sock *, int mib);
612 void tcp_send_fin(struct sock *sk);
613 void tcp_send_active_reset(struct sock *sk, gfp_t priority);
614 int tcp_send_synack(struct sock *);
615 void tcp_push_one(struct sock *, unsigned int mss_now);
616 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt);
617 void tcp_send_ack(struct sock *sk);
618 void tcp_send_delayed_ack(struct sock *sk);
619 void tcp_send_loss_probe(struct sock *sk);
620 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto);
621 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
622 const struct sk_buff *next_skb);
625 void tcp_rearm_rto(struct sock *sk);
626 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req);
627 void tcp_reset(struct sock *sk, struct sk_buff *skb);
628 void tcp_fin(struct sock *sk);
629 void tcp_check_space(struct sock *sk);
630 void tcp_sack_compress_send_ack(struct sock *sk);
633 void tcp_init_xmit_timers(struct sock *);
634 static inline void tcp_clear_xmit_timers(struct sock *sk)
636 if (hrtimer_try_to_cancel(&tcp_sk(sk)->pacing_timer) == 1)
639 if (hrtimer_try_to_cancel(&tcp_sk(sk)->compressed_ack_timer) == 1)
642 inet_csk_clear_xmit_timers(sk);
645 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu);
646 unsigned int tcp_current_mss(struct sock *sk);
647 u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when);
649 /* Bound MSS / TSO packet size with the half of the window */
650 static inline int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
654 /* When peer uses tiny windows, there is no use in packetizing
655 * to sub-MSS pieces for the sake of SWS or making sure there
656 * are enough packets in the pipe for fast recovery.
658 * On the other hand, for extremely large MSS devices, handling
659 * smaller than MSS windows in this way does make sense.
661 if (tp->max_window > TCP_MSS_DEFAULT)
662 cutoff = (tp->max_window >> 1);
664 cutoff = tp->max_window;
666 if (cutoff && pktsize > cutoff)
667 return max_t(int, cutoff, 68U - tp->tcp_header_len);
673 void tcp_get_info(struct sock *, struct tcp_info *);
675 /* Read 'sendfile()'-style from a TCP socket */
676 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
677 sk_read_actor_t recv_actor);
678 int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor);
679 struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off);
680 void tcp_read_done(struct sock *sk, size_t len);
682 void tcp_initialize_rcv_mss(struct sock *sk);
684 int tcp_mtu_to_mss(struct sock *sk, int pmtu);
685 int tcp_mss_to_mtu(struct sock *sk, int mss);
686 void tcp_mtup_init(struct sock *sk);
688 static inline void tcp_bound_rto(const struct sock *sk)
690 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
691 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
694 static inline u32 __tcp_set_rto(const struct tcp_sock *tp)
696 return usecs_to_jiffies((tp->srtt_us >> 3) + tp->rttvar_us);
699 static inline void __tcp_fast_path_on(struct tcp_sock *tp, u32 snd_wnd)
701 /* mptcp hooks are only on the slow path */
702 if (sk_is_mptcp((struct sock *)tp))
705 tp->pred_flags = htonl((tp->tcp_header_len << 26) |
706 ntohl(TCP_FLAG_ACK) |
710 static inline void tcp_fast_path_on(struct tcp_sock *tp)
712 __tcp_fast_path_on(tp, tp->snd_wnd >> tp->rx_opt.snd_wscale);
715 static inline void tcp_fast_path_check(struct sock *sk)
717 struct tcp_sock *tp = tcp_sk(sk);
719 if (RB_EMPTY_ROOT(&tp->out_of_order_queue) &&
721 atomic_read(&sk->sk_rmem_alloc) < sk->sk_rcvbuf &&
723 tcp_fast_path_on(tp);
726 /* Compute the actual rto_min value */
727 static inline u32 tcp_rto_min(struct sock *sk)
729 const struct dst_entry *dst = __sk_dst_get(sk);
730 u32 rto_min = inet_csk(sk)->icsk_rto_min;
732 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
733 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
737 static inline u32 tcp_rto_min_us(struct sock *sk)
739 return jiffies_to_usecs(tcp_rto_min(sk));
742 static inline bool tcp_ca_dst_locked(const struct dst_entry *dst)
744 return dst_metric_locked(dst, RTAX_CC_ALGO);
747 /* Minimum RTT in usec. ~0 means not available. */
748 static inline u32 tcp_min_rtt(const struct tcp_sock *tp)
750 return minmax_get(&tp->rtt_min);
753 /* Compute the actual receive window we are currently advertising.
754 * Rcv_nxt can be after the window if our peer push more data
755 * than the offered window.
757 static inline u32 tcp_receive_window(const struct tcp_sock *tp)
759 s32 win = tp->rcv_wup + tp->rcv_wnd - tp->rcv_nxt;
766 /* Choose a new window, without checks for shrinking, and without
767 * scaling applied to the result. The caller does these things
768 * if necessary. This is a "raw" window selection.
770 u32 __tcp_select_window(struct sock *sk);
772 void tcp_send_window_probe(struct sock *sk);
774 /* TCP uses 32bit jiffies to save some space.
775 * Note that this is different from tcp_time_stamp, which
776 * historically has been the same until linux-4.13.
778 #define tcp_jiffies32 ((u32)jiffies)
781 * Deliver a 32bit value for TCP timestamp option (RFC 7323)
782 * It is no longer tied to jiffies, but to 1 ms clock.
783 * Note: double check if you want to use tcp_jiffies32 instead of this.
785 #define TCP_TS_HZ 1000
787 static inline u64 tcp_clock_ns(void)
789 return ktime_get_ns();
792 static inline u64 tcp_clock_us(void)
794 return div_u64(tcp_clock_ns(), NSEC_PER_USEC);
797 /* This should only be used in contexts where tp->tcp_mstamp is up to date */
798 static inline u32 tcp_time_stamp(const struct tcp_sock *tp)
800 return div_u64(tp->tcp_mstamp, USEC_PER_SEC / TCP_TS_HZ);
803 /* Convert a nsec timestamp into TCP TSval timestamp (ms based currently) */
804 static inline u64 tcp_ns_to_ts(u64 ns)
806 return div_u64(ns, NSEC_PER_SEC / TCP_TS_HZ);
809 /* Could use tcp_clock_us() / 1000, but this version uses a single divide */
810 static inline u32 tcp_time_stamp_raw(void)
812 return tcp_ns_to_ts(tcp_clock_ns());
815 void tcp_mstamp_refresh(struct tcp_sock *tp);
817 static inline u32 tcp_stamp_us_delta(u64 t1, u64 t0)
819 return max_t(s64, t1 - t0, 0);
822 static inline u32 tcp_skb_timestamp(const struct sk_buff *skb)
824 return tcp_ns_to_ts(skb->skb_mstamp_ns);
827 /* provide the departure time in us unit */
828 static inline u64 tcp_skb_timestamp_us(const struct sk_buff *skb)
830 return div_u64(skb->skb_mstamp_ns, NSEC_PER_USEC);
834 #define tcp_flag_byte(th) (((u_int8_t *)th)[13])
836 #define TCPHDR_FIN 0x01
837 #define TCPHDR_SYN 0x02
838 #define TCPHDR_RST 0x04
839 #define TCPHDR_PSH 0x08
840 #define TCPHDR_ACK 0x10
841 #define TCPHDR_URG 0x20
842 #define TCPHDR_ECE 0x40
843 #define TCPHDR_CWR 0x80
845 #define TCPHDR_SYN_ECN (TCPHDR_SYN | TCPHDR_ECE | TCPHDR_CWR)
847 /* This is what the send packet queuing engine uses to pass
848 * TCP per-packet control information to the transmission code.
849 * We also store the host-order sequence numbers in here too.
850 * This is 44 bytes if IPV6 is enabled.
851 * If this grows please adjust skbuff.h:skbuff->cb[xxx] size appropriately.
854 __u32 seq; /* Starting sequence number */
855 __u32 end_seq; /* SEQ + FIN + SYN + datalen */
857 /* Note : tcp_tw_isn is used in input path only
858 * (isn chosen by tcp_timewait_state_process())
860 * tcp_gso_segs/size are used in write queue only,
861 * cf tcp_skb_pcount()/tcp_skb_mss()
869 __u8 tcp_flags; /* TCP header flags. (tcp[13]) */
871 __u8 sacked; /* State flags for SACK. */
872 #define TCPCB_SACKED_ACKED 0x01 /* SKB ACK'd by a SACK block */
873 #define TCPCB_SACKED_RETRANS 0x02 /* SKB retransmitted */
874 #define TCPCB_LOST 0x04 /* SKB is lost */
875 #define TCPCB_TAGBITS 0x07 /* All tag bits */
876 #define TCPCB_REPAIRED 0x10 /* SKB repaired (no skb_mstamp_ns) */
877 #define TCPCB_EVER_RETRANS 0x80 /* Ever retransmitted frame */
878 #define TCPCB_RETRANS (TCPCB_SACKED_RETRANS|TCPCB_EVER_RETRANS| \
881 __u8 ip_dsfield; /* IPv4 tos or IPv6 dsfield */
882 __u8 txstamp_ack:1, /* Record TX timestamp for ack? */
883 eor:1, /* Is skb MSG_EOR marked? */
884 has_rxtstamp:1, /* SKB has a RX timestamp */
886 __u32 ack_seq; /* Sequence number ACK'd */
889 #define TCPCB_DELIVERED_CE_MASK ((1U<<20) - 1)
890 /* There is space for up to 24 bytes */
891 __u32 is_app_limited:1, /* cwnd not fully used? */
894 /* pkts S/ACKed so far upon tx of skb, incl retrans: */
896 /* start of send pipeline phase */
898 /* when we reached the "delivered" count */
899 u64 delivered_mstamp;
900 } tx; /* only used for outgoing skbs */
902 struct inet_skb_parm h4;
903 #if IS_ENABLED(CONFIG_IPV6)
904 struct inet6_skb_parm h6;
906 } header; /* For incoming skbs */
910 #define TCP_SKB_CB(__skb) ((struct tcp_skb_cb *)&((__skb)->cb[0]))
912 extern const struct inet_connection_sock_af_ops ipv4_specific;
914 #if IS_ENABLED(CONFIG_IPV6)
915 /* This is the variant of inet6_iif() that must be used by TCP,
916 * as TCP moves IP6CB into a different location in skb->cb[]
918 static inline int tcp_v6_iif(const struct sk_buff *skb)
920 return TCP_SKB_CB(skb)->header.h6.iif;
923 static inline int tcp_v6_iif_l3_slave(const struct sk_buff *skb)
925 bool l3_slave = ipv6_l3mdev_skb(TCP_SKB_CB(skb)->header.h6.flags);
927 return l3_slave ? skb->skb_iif : TCP_SKB_CB(skb)->header.h6.iif;
930 /* TCP_SKB_CB reference means this can not be used from early demux */
931 static inline int tcp_v6_sdif(const struct sk_buff *skb)
933 #if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV)
934 if (skb && ipv6_l3mdev_skb(TCP_SKB_CB(skb)->header.h6.flags))
935 return TCP_SKB_CB(skb)->header.h6.iif;
940 extern const struct inet_connection_sock_af_ops ipv6_specific;
942 INDIRECT_CALLABLE_DECLARE(void tcp_v6_send_check(struct sock *sk, struct sk_buff *skb));
943 INDIRECT_CALLABLE_DECLARE(int tcp_v6_rcv(struct sk_buff *skb));
944 void tcp_v6_early_demux(struct sk_buff *skb);
948 /* TCP_SKB_CB reference means this can not be used from early demux */
949 static inline int tcp_v4_sdif(struct sk_buff *skb)
951 #if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV)
952 if (skb && ipv4_l3mdev_skb(TCP_SKB_CB(skb)->header.h4.flags))
953 return TCP_SKB_CB(skb)->header.h4.iif;
958 /* Due to TSO, an SKB can be composed of multiple actual
959 * packets. To keep these tracked properly, we use this.
961 static inline int tcp_skb_pcount(const struct sk_buff *skb)
963 return TCP_SKB_CB(skb)->tcp_gso_segs;
966 static inline void tcp_skb_pcount_set(struct sk_buff *skb, int segs)
968 TCP_SKB_CB(skb)->tcp_gso_segs = segs;
971 static inline void tcp_skb_pcount_add(struct sk_buff *skb, int segs)
973 TCP_SKB_CB(skb)->tcp_gso_segs += segs;
976 /* This is valid iff skb is in write queue and tcp_skb_pcount() > 1. */
977 static inline int tcp_skb_mss(const struct sk_buff *skb)
979 return TCP_SKB_CB(skb)->tcp_gso_size;
982 static inline bool tcp_skb_can_collapse_to(const struct sk_buff *skb)
984 return likely(!TCP_SKB_CB(skb)->eor);
987 static inline bool tcp_skb_can_collapse(const struct sk_buff *to,
988 const struct sk_buff *from)
990 return likely(tcp_skb_can_collapse_to(to) &&
991 mptcp_skb_can_collapse(to, from) &&
992 skb_pure_zcopy_same(to, from));
995 /* Events passed to congestion control interface */
997 CA_EVENT_TX_START, /* first transmit when no packets in flight */
998 CA_EVENT_CWND_RESTART, /* congestion window restart */
999 CA_EVENT_COMPLETE_CWR, /* end of congestion recovery */
1000 CA_EVENT_LOSS, /* loss timeout */
1001 CA_EVENT_ECN_NO_CE, /* ECT set, but not CE marked */
1002 CA_EVENT_ECN_IS_CE, /* received CE marked IP packet */
1005 /* Information about inbound ACK, passed to cong_ops->in_ack_event() */
1006 enum tcp_ca_ack_event_flags {
1007 CA_ACK_SLOWPATH = (1 << 0), /* In slow path processing */
1008 CA_ACK_WIN_UPDATE = (1 << 1), /* ACK updated window */
1009 CA_ACK_ECE = (1 << 2), /* ECE bit is set on ack */
1013 * Interface for adding new TCP congestion control handlers
1015 #define TCP_CA_NAME_MAX 16
1016 #define TCP_CA_MAX 128
1017 #define TCP_CA_BUF_MAX (TCP_CA_NAME_MAX*TCP_CA_MAX)
1019 #define TCP_CA_UNSPEC 0
1021 /* Algorithm can be set on socket without CAP_NET_ADMIN privileges */
1022 #define TCP_CONG_NON_RESTRICTED 0x1
1023 /* Requires ECN/ECT set on all packets */
1024 #define TCP_CONG_NEEDS_ECN 0x2
1025 #define TCP_CONG_MASK (TCP_CONG_NON_RESTRICTED | TCP_CONG_NEEDS_ECN)
1035 /* A rate sample measures the number of (original/retransmitted) data
1036 * packets delivered "delivered" over an interval of time "interval_us".
1037 * The tcp_rate.c code fills in the rate sample, and congestion
1038 * control modules that define a cong_control function to run at the end
1039 * of ACK processing can optionally chose to consult this sample when
1040 * setting cwnd and pacing rate.
1041 * A sample is invalid if "delivered" or "interval_us" is negative.
1043 struct rate_sample {
1044 u64 prior_mstamp; /* starting timestamp for interval */
1045 u32 prior_delivered; /* tp->delivered at "prior_mstamp" */
1046 u32 prior_delivered_ce;/* tp->delivered_ce at "prior_mstamp" */
1047 s32 delivered; /* number of packets delivered over interval */
1048 s32 delivered_ce; /* number of packets delivered w/ CE marks*/
1049 long interval_us; /* time for tp->delivered to incr "delivered" */
1050 u32 snd_interval_us; /* snd interval for delivered packets */
1051 u32 rcv_interval_us; /* rcv interval for delivered packets */
1052 long rtt_us; /* RTT of last (S)ACKed packet (or -1) */
1053 int losses; /* number of packets marked lost upon ACK */
1054 u32 acked_sacked; /* number of packets newly (S)ACKed upon ACK */
1055 u32 prior_in_flight; /* in flight before this ACK */
1056 u32 last_end_seq; /* end_seq of most recently ACKed packet */
1057 bool is_app_limited; /* is sample from packet with bubble in pipe? */
1058 bool is_retrans; /* is sample from retransmission? */
1059 bool is_ack_delayed; /* is this (likely) a delayed ACK? */
1062 struct tcp_congestion_ops {
1063 /* fast path fields are put first to fill one cache line */
1065 /* return slow start threshold (required) */
1066 u32 (*ssthresh)(struct sock *sk);
1068 /* do new cwnd calculation (required) */
1069 void (*cong_avoid)(struct sock *sk, u32 ack, u32 acked);
1071 /* call before changing ca_state (optional) */
1072 void (*set_state)(struct sock *sk, u8 new_state);
1074 /* call when cwnd event occurs (optional) */
1075 void (*cwnd_event)(struct sock *sk, enum tcp_ca_event ev);
1077 /* call when ack arrives (optional) */
1078 void (*in_ack_event)(struct sock *sk, u32 flags);
1080 /* hook for packet ack accounting (optional) */
1081 void (*pkts_acked)(struct sock *sk, const struct ack_sample *sample);
1083 /* override sysctl_tcp_min_tso_segs */
1084 u32 (*min_tso_segs)(struct sock *sk);
1086 /* call when packets are delivered to update cwnd and pacing rate,
1087 * after all the ca_state processing. (optional)
1089 void (*cong_control)(struct sock *sk, const struct rate_sample *rs);
1092 /* new value of cwnd after loss (required) */
1093 u32 (*undo_cwnd)(struct sock *sk);
1094 /* returns the multiplier used in tcp_sndbuf_expand (optional) */
1095 u32 (*sndbuf_expand)(struct sock *sk);
1097 /* control/slow paths put last */
1098 /* get info for inet_diag (optional) */
1099 size_t (*get_info)(struct sock *sk, u32 ext, int *attr,
1100 union tcp_cc_info *info);
1102 char name[TCP_CA_NAME_MAX];
1103 struct module *owner;
1104 struct list_head list;
1108 /* initialize private data (optional) */
1109 void (*init)(struct sock *sk);
1110 /* cleanup private data (optional) */
1111 void (*release)(struct sock *sk);
1112 } ____cacheline_aligned_in_smp;
1114 int tcp_register_congestion_control(struct tcp_congestion_ops *type);
1115 void tcp_unregister_congestion_control(struct tcp_congestion_ops *type);
1116 int tcp_update_congestion_control(struct tcp_congestion_ops *type,
1117 struct tcp_congestion_ops *old_type);
1118 int tcp_validate_congestion_control(struct tcp_congestion_ops *ca);
1120 void tcp_assign_congestion_control(struct sock *sk);
1121 void tcp_init_congestion_control(struct sock *sk);
1122 void tcp_cleanup_congestion_control(struct sock *sk);
1123 int tcp_set_default_congestion_control(struct net *net, const char *name);
1124 void tcp_get_default_congestion_control(struct net *net, char *name);
1125 void tcp_get_available_congestion_control(char *buf, size_t len);
1126 void tcp_get_allowed_congestion_control(char *buf, size_t len);
1127 int tcp_set_allowed_congestion_control(char *allowed);
1128 int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
1129 bool cap_net_admin);
1130 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked);
1131 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked);
1133 u32 tcp_reno_ssthresh(struct sock *sk);
1134 u32 tcp_reno_undo_cwnd(struct sock *sk);
1135 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked);
1136 extern struct tcp_congestion_ops tcp_reno;
1138 struct tcp_congestion_ops *tcp_ca_find(const char *name);
1139 struct tcp_congestion_ops *tcp_ca_find_key(u32 key);
1140 u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca);
1142 char *tcp_ca_get_name_by_key(u32 key, char *buffer);
1144 static inline char *tcp_ca_get_name_by_key(u32 key, char *buffer)
1150 static inline bool tcp_ca_needs_ecn(const struct sock *sk)
1152 const struct inet_connection_sock *icsk = inet_csk(sk);
1154 return icsk->icsk_ca_ops->flags & TCP_CONG_NEEDS_ECN;
1157 static inline void tcp_ca_event(struct sock *sk, const enum tcp_ca_event event)
1159 const struct inet_connection_sock *icsk = inet_csk(sk);
1161 if (icsk->icsk_ca_ops->cwnd_event)
1162 icsk->icsk_ca_ops->cwnd_event(sk, event);
1165 /* From tcp_cong.c */
1166 void tcp_set_ca_state(struct sock *sk, const u8 ca_state);
1168 /* From tcp_rate.c */
1169 void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb);
1170 void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb,
1171 struct rate_sample *rs);
1172 void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost,
1173 bool is_sack_reneg, struct rate_sample *rs);
1174 void tcp_rate_check_app_limited(struct sock *sk);
1176 static inline bool tcp_skb_sent_after(u64 t1, u64 t2, u32 seq1, u32 seq2)
1178 return t1 > t2 || (t1 == t2 && after(seq1, seq2));
1181 /* These functions determine how the current flow behaves in respect of SACK
1182 * handling. SACK is negotiated with the peer, and therefore it can vary
1183 * between different flows.
1185 * tcp_is_sack - SACK enabled
1186 * tcp_is_reno - No SACK
1188 static inline int tcp_is_sack(const struct tcp_sock *tp)
1190 return likely(tp->rx_opt.sack_ok);
1193 static inline bool tcp_is_reno(const struct tcp_sock *tp)
1195 return !tcp_is_sack(tp);
1198 static inline unsigned int tcp_left_out(const struct tcp_sock *tp)
1200 return tp->sacked_out + tp->lost_out;
1203 /* This determines how many packets are "in the network" to the best
1204 * of our knowledge. In many cases it is conservative, but where
1205 * detailed information is available from the receiver (via SACK
1206 * blocks etc.) we can make more aggressive calculations.
1208 * Use this for decisions involving congestion control, use just
1209 * tp->packets_out to determine if the send queue is empty or not.
1211 * Read this equation as:
1213 * "Packets sent once on transmission queue" MINUS
1214 * "Packets left network, but not honestly ACKed yet" PLUS
1215 * "Packets fast retransmitted"
1217 static inline unsigned int tcp_packets_in_flight(const struct tcp_sock *tp)
1219 return tp->packets_out - tcp_left_out(tp) + tp->retrans_out;
1222 #define TCP_INFINITE_SSTHRESH 0x7fffffff
1224 static inline u32 tcp_snd_cwnd(const struct tcp_sock *tp)
1226 return tp->snd_cwnd;
1229 static inline void tcp_snd_cwnd_set(struct tcp_sock *tp, u32 val)
1231 WARN_ON_ONCE((int)val <= 0);
1235 static inline bool tcp_in_slow_start(const struct tcp_sock *tp)
1237 return tcp_snd_cwnd(tp) < tp->snd_ssthresh;
1240 static inline bool tcp_in_initial_slowstart(const struct tcp_sock *tp)
1242 return tp->snd_ssthresh >= TCP_INFINITE_SSTHRESH;
1245 static inline bool tcp_in_cwnd_reduction(const struct sock *sk)
1247 return (TCPF_CA_CWR | TCPF_CA_Recovery) &
1248 (1 << inet_csk(sk)->icsk_ca_state);
1251 /* If cwnd > ssthresh, we may raise ssthresh to be half-way to cwnd.
1252 * The exception is cwnd reduction phase, when cwnd is decreasing towards
1255 static inline __u32 tcp_current_ssthresh(const struct sock *sk)
1257 const struct tcp_sock *tp = tcp_sk(sk);
1259 if (tcp_in_cwnd_reduction(sk))
1260 return tp->snd_ssthresh;
1262 return max(tp->snd_ssthresh,
1263 ((tcp_snd_cwnd(tp) >> 1) +
1264 (tcp_snd_cwnd(tp) >> 2)));
1267 /* Use define here intentionally to get WARN_ON location shown at the caller */
1268 #define tcp_verify_left_out(tp) WARN_ON(tcp_left_out(tp) > tp->packets_out)
1270 void tcp_enter_cwr(struct sock *sk);
1271 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst);
1273 /* The maximum number of MSS of available cwnd for which TSO defers
1274 * sending if not using sysctl_tcp_tso_win_divisor.
1276 static inline __u32 tcp_max_tso_deferred_mss(const struct tcp_sock *tp)
1281 /* Returns end sequence number of the receiver's advertised window */
1282 static inline u32 tcp_wnd_end(const struct tcp_sock *tp)
1284 return tp->snd_una + tp->snd_wnd;
1287 /* We follow the spirit of RFC2861 to validate cwnd but implement a more
1288 * flexible approach. The RFC suggests cwnd should not be raised unless
1289 * it was fully used previously. And that's exactly what we do in
1290 * congestion avoidance mode. But in slow start we allow cwnd to grow
1291 * as long as the application has used half the cwnd.
1293 * cwnd is 10 (IW10), but application sends 9 frames.
1294 * We allow cwnd to reach 18 when all frames are ACKed.
1295 * This check is safe because it's as aggressive as slow start which already
1296 * risks 100% overshoot. The advantage is that we discourage application to
1297 * either send more filler packets or data to artificially blow up the cwnd
1298 * usage, and allow application-limited process to probe bw more aggressively.
1300 static inline bool tcp_is_cwnd_limited(const struct sock *sk)
1302 const struct tcp_sock *tp = tcp_sk(sk);
1304 if (tp->is_cwnd_limited)
1307 /* If in slow start, ensure cwnd grows to twice what was ACKed. */
1308 if (tcp_in_slow_start(tp))
1309 return tcp_snd_cwnd(tp) < 2 * tp->max_packets_out;
1314 /* BBR congestion control needs pacing.
1315 * Same remark for SO_MAX_PACING_RATE.
1316 * sch_fq packet scheduler is efficiently handling pacing,
1317 * but is not always installed/used.
1318 * Return true if TCP stack should pace packets itself.
1320 static inline bool tcp_needs_internal_pacing(const struct sock *sk)
1322 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
1325 /* Estimates in how many jiffies next packet for this flow can be sent.
1326 * Scheduling a retransmit timer too early would be silly.
1328 static inline unsigned long tcp_pacing_delay(const struct sock *sk)
1330 s64 delay = tcp_sk(sk)->tcp_wstamp_ns - tcp_sk(sk)->tcp_clock_cache;
1332 return delay > 0 ? nsecs_to_jiffies(delay) : 0;
1335 static inline void tcp_reset_xmit_timer(struct sock *sk,
1338 const unsigned long max_when)
1340 inet_csk_reset_xmit_timer(sk, what, when + tcp_pacing_delay(sk),
1344 /* Something is really bad, we could not queue an additional packet,
1345 * because qdisc is full or receiver sent a 0 window, or we are paced.
1346 * We do not want to add fuel to the fire, or abort too early,
1347 * so make sure the timer we arm now is at least 200ms in the future,
1348 * regardless of current icsk_rto value (as it could be ~2ms)
1350 static inline unsigned long tcp_probe0_base(const struct sock *sk)
1352 return max_t(unsigned long, inet_csk(sk)->icsk_rto, TCP_RTO_MIN);
1355 /* Variant of inet_csk_rto_backoff() used for zero window probes */
1356 static inline unsigned long tcp_probe0_when(const struct sock *sk,
1357 unsigned long max_when)
1359 u8 backoff = min_t(u8, ilog2(TCP_RTO_MAX / TCP_RTO_MIN) + 1,
1360 inet_csk(sk)->icsk_backoff);
1361 u64 when = (u64)tcp_probe0_base(sk) << backoff;
1363 return (unsigned long)min_t(u64, when, max_when);
1366 static inline void tcp_check_probe_timer(struct sock *sk)
1368 if (!tcp_sk(sk)->packets_out && !inet_csk(sk)->icsk_pending)
1369 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
1370 tcp_probe0_base(sk), TCP_RTO_MAX);
1373 static inline void tcp_init_wl(struct tcp_sock *tp, u32 seq)
1378 static inline void tcp_update_wl(struct tcp_sock *tp, u32 seq)
1384 * Calculate(/check) TCP checksum
1386 static inline __sum16 tcp_v4_check(int len, __be32 saddr,
1387 __be32 daddr, __wsum base)
1389 return csum_tcpudp_magic(saddr, daddr, len, IPPROTO_TCP, base);
1392 static inline bool tcp_checksum_complete(struct sk_buff *skb)
1394 return !skb_csum_unnecessary(skb) &&
1395 __skb_checksum_complete(skb);
1398 bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
1399 enum skb_drop_reason *reason);
1402 int tcp_filter(struct sock *sk, struct sk_buff *skb);
1403 void tcp_set_state(struct sock *sk, int state);
1404 void tcp_done(struct sock *sk);
1405 int tcp_abort(struct sock *sk, int err);
1407 static inline void tcp_sack_reset(struct tcp_options_received *rx_opt)
1410 rx_opt->num_sacks = 0;
1413 void tcp_cwnd_restart(struct sock *sk, s32 delta);
1415 static inline void tcp_slow_start_after_idle_check(struct sock *sk)
1417 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1418 struct tcp_sock *tp = tcp_sk(sk);
1421 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) ||
1422 tp->packets_out || ca_ops->cong_control)
1424 delta = tcp_jiffies32 - tp->lsndtime;
1425 if (delta > inet_csk(sk)->icsk_rto)
1426 tcp_cwnd_restart(sk, delta);
1429 /* Determine a window scaling and initial window to offer. */
1430 void tcp_select_initial_window(const struct sock *sk, int __space,
1431 __u32 mss, __u32 *rcv_wnd,
1432 __u32 *window_clamp, int wscale_ok,
1433 __u8 *rcv_wscale, __u32 init_rcv_wnd);
1435 static inline int __tcp_win_from_space(u8 scaling_ratio, int space)
1437 s64 scaled_space = (s64)space * scaling_ratio;
1439 return scaled_space >> TCP_RMEM_TO_WIN_SCALE;
1442 static inline int tcp_win_from_space(const struct sock *sk, int space)
1444 return __tcp_win_from_space(tcp_sk(sk)->scaling_ratio, space);
1447 /* inverse of __tcp_win_from_space() */
1448 static inline int __tcp_space_from_win(u8 scaling_ratio, int win)
1450 u64 val = (u64)win << TCP_RMEM_TO_WIN_SCALE;
1452 do_div(val, scaling_ratio);
1456 static inline int tcp_space_from_win(const struct sock *sk, int win)
1458 return __tcp_space_from_win(tcp_sk(sk)->scaling_ratio, win);
1461 static inline void tcp_scaling_ratio_init(struct sock *sk)
1463 /* Assume a conservative default of 1200 bytes of payload per 4K page.
1464 * This may be adjusted later in tcp_measure_rcv_mss().
1466 tcp_sk(sk)->scaling_ratio = (1200 << TCP_RMEM_TO_WIN_SCALE) /
1470 /* Note: caller must be prepared to deal with negative returns */
1471 static inline int tcp_space(const struct sock *sk)
1473 return tcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf) -
1474 READ_ONCE(sk->sk_backlog.len) -
1475 atomic_read(&sk->sk_rmem_alloc));
1478 static inline int tcp_full_space(const struct sock *sk)
1480 return tcp_win_from_space(sk, READ_ONCE(sk->sk_rcvbuf));
1483 static inline void __tcp_adjust_rcv_ssthresh(struct sock *sk, u32 new_ssthresh)
1485 int unused_mem = sk_unused_reserved_mem(sk);
1486 struct tcp_sock *tp = tcp_sk(sk);
1488 tp->rcv_ssthresh = min(tp->rcv_ssthresh, new_ssthresh);
1490 tp->rcv_ssthresh = max_t(u32, tp->rcv_ssthresh,
1491 tcp_win_from_space(sk, unused_mem));
1494 static inline void tcp_adjust_rcv_ssthresh(struct sock *sk)
1496 __tcp_adjust_rcv_ssthresh(sk, 4U * tcp_sk(sk)->advmss);
1499 void tcp_cleanup_rbuf(struct sock *sk, int copied);
1500 void __tcp_cleanup_rbuf(struct sock *sk, int copied);
1503 /* We provision sk_rcvbuf around 200% of sk_rcvlowat.
1504 * If 87.5 % (7/8) of the space has been consumed, we want to override
1505 * SO_RCVLOWAT constraint, since we are receiving skbs with too small
1506 * len/truesize ratio.
1508 static inline bool tcp_rmem_pressure(const struct sock *sk)
1510 int rcvbuf, threshold;
1512 if (tcp_under_memory_pressure(sk))
1515 rcvbuf = READ_ONCE(sk->sk_rcvbuf);
1516 threshold = rcvbuf - (rcvbuf >> 3);
1518 return atomic_read(&sk->sk_rmem_alloc) > threshold;
1521 static inline bool tcp_epollin_ready(const struct sock *sk, int target)
1523 const struct tcp_sock *tp = tcp_sk(sk);
1524 int avail = READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq);
1529 return (avail >= target) || tcp_rmem_pressure(sk) ||
1530 (tcp_receive_window(tp) <= inet_csk(sk)->icsk_ack.rcv_mss);
1533 extern void tcp_openreq_init_rwin(struct request_sock *req,
1534 const struct sock *sk_listener,
1535 const struct dst_entry *dst);
1537 void tcp_enter_memory_pressure(struct sock *sk);
1538 void tcp_leave_memory_pressure(struct sock *sk);
1540 static inline int keepalive_intvl_when(const struct tcp_sock *tp)
1542 struct net *net = sock_net((struct sock *)tp);
1545 /* Paired with WRITE_ONCE() in tcp_sock_set_keepintvl()
1546 * and do_tcp_setsockopt().
1548 val = READ_ONCE(tp->keepalive_intvl);
1550 return val ? : READ_ONCE(net->ipv4.sysctl_tcp_keepalive_intvl);
1553 static inline int keepalive_time_when(const struct tcp_sock *tp)
1555 struct net *net = sock_net((struct sock *)tp);
1558 /* Paired with WRITE_ONCE() in tcp_sock_set_keepidle_locked() */
1559 val = READ_ONCE(tp->keepalive_time);
1561 return val ? : READ_ONCE(net->ipv4.sysctl_tcp_keepalive_time);
1564 static inline int keepalive_probes(const struct tcp_sock *tp)
1566 struct net *net = sock_net((struct sock *)tp);
1569 /* Paired with WRITE_ONCE() in tcp_sock_set_keepcnt()
1570 * and do_tcp_setsockopt().
1572 val = READ_ONCE(tp->keepalive_probes);
1574 return val ? : READ_ONCE(net->ipv4.sysctl_tcp_keepalive_probes);
1577 static inline u32 keepalive_time_elapsed(const struct tcp_sock *tp)
1579 const struct inet_connection_sock *icsk = &tp->inet_conn;
1581 return min_t(u32, tcp_jiffies32 - icsk->icsk_ack.lrcvtime,
1582 tcp_jiffies32 - tp->rcv_tstamp);
1585 static inline int tcp_fin_time(const struct sock *sk)
1587 int fin_timeout = tcp_sk(sk)->linger2 ? :
1588 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fin_timeout);
1589 const int rto = inet_csk(sk)->icsk_rto;
1591 if (fin_timeout < (rto << 2) - (rto >> 1))
1592 fin_timeout = (rto << 2) - (rto >> 1);
1597 static inline bool tcp_paws_check(const struct tcp_options_received *rx_opt,
1600 if ((s32)(rx_opt->ts_recent - rx_opt->rcv_tsval) <= paws_win)
1602 if (unlikely(!time_before32(ktime_get_seconds(),
1603 rx_opt->ts_recent_stamp + TCP_PAWS_24DAYS)))
1606 * Some OSes send SYN and SYNACK messages with tsval=0 tsecr=0,
1607 * then following tcp messages have valid values. Ignore 0 value,
1608 * or else 'negative' tsval might forbid us to accept their packets.
1610 if (!rx_opt->ts_recent)
1615 static inline bool tcp_paws_reject(const struct tcp_options_received *rx_opt,
1618 if (tcp_paws_check(rx_opt, 0))
1621 /* RST segments are not recommended to carry timestamp,
1622 and, if they do, it is recommended to ignore PAWS because
1623 "their cleanup function should take precedence over timestamps."
1624 Certainly, it is mistake. It is necessary to understand the reasons
1625 of this constraint to relax it: if peer reboots, clock may go
1626 out-of-sync and half-open connections will not be reset.
1627 Actually, the problem would be not existing if all
1628 the implementations followed draft about maintaining clock
1629 via reboots. Linux-2.2 DOES NOT!
1631 However, we can relax time bounds for RST segments to MSL.
1633 if (rst && !time_before32(ktime_get_seconds(),
1634 rx_opt->ts_recent_stamp + TCP_PAWS_MSL))
1639 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
1640 int mib_idx, u32 *last_oow_ack_time);
1642 static inline void tcp_mib_init(struct net *net)
1645 TCP_ADD_STATS(net, TCP_MIB_RTOALGORITHM, 1);
1646 TCP_ADD_STATS(net, TCP_MIB_RTOMIN, TCP_RTO_MIN*1000/HZ);
1647 TCP_ADD_STATS(net, TCP_MIB_RTOMAX, TCP_RTO_MAX*1000/HZ);
1648 TCP_ADD_STATS(net, TCP_MIB_MAXCONN, -1);
1652 static inline void tcp_clear_retrans_hints_partial(struct tcp_sock *tp)
1654 tp->lost_skb_hint = NULL;
1657 static inline void tcp_clear_all_retrans_hints(struct tcp_sock *tp)
1659 tcp_clear_retrans_hints_partial(tp);
1660 tp->retransmit_skb_hint = NULL;
1663 union tcp_md5_addr {
1665 #if IS_ENABLED(CONFIG_IPV6)
1670 /* - key database */
1671 struct tcp_md5sig_key {
1672 struct hlist_node node;
1674 u8 family; /* AF_INET or AF_INET6 */
1677 union tcp_md5_addr addr;
1678 int l3index; /* set if key added with L3 scope */
1679 u8 key[TCP_MD5SIG_MAXKEYLEN];
1680 struct rcu_head rcu;
1684 struct tcp_md5sig_info {
1685 struct hlist_head head;
1686 struct rcu_head rcu;
1689 /* - pseudo header */
1690 struct tcp4_pseudohdr {
1698 struct tcp6_pseudohdr {
1699 struct in6_addr saddr;
1700 struct in6_addr daddr;
1702 __be32 protocol; /* including padding */
1705 union tcp_md5sum_block {
1706 struct tcp4_pseudohdr ip4;
1707 #if IS_ENABLED(CONFIG_IPV6)
1708 struct tcp6_pseudohdr ip6;
1712 /* - pool: digest algorithm, hash description and scratch buffer */
1713 struct tcp_md5sig_pool {
1714 struct ahash_request *md5_req;
1719 int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1720 const struct sock *sk, const struct sk_buff *skb);
1721 int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1722 int family, u8 prefixlen, int l3index, u8 flags,
1723 const u8 *newkey, u8 newkeylen);
1724 int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
1725 int family, u8 prefixlen, int l3index,
1726 struct tcp_md5sig_key *key);
1728 int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr,
1729 int family, u8 prefixlen, int l3index, u8 flags);
1730 struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
1731 const struct sock *addr_sk);
1733 #ifdef CONFIG_TCP_MD5SIG
1734 #include <linux/jump_label.h>
1735 extern struct static_key_false_deferred tcp_md5_needed;
1736 struct tcp_md5sig_key *__tcp_md5_do_lookup(const struct sock *sk, int l3index,
1737 const union tcp_md5_addr *addr,
1739 static inline struct tcp_md5sig_key *
1740 tcp_md5_do_lookup(const struct sock *sk, int l3index,
1741 const union tcp_md5_addr *addr, int family)
1743 if (!static_branch_unlikely(&tcp_md5_needed.key))
1745 return __tcp_md5_do_lookup(sk, l3index, addr, family);
1748 enum skb_drop_reason
1749 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
1750 const void *saddr, const void *daddr,
1751 int family, int dif, int sdif);
1754 #define tcp_twsk_md5_key(twsk) ((twsk)->tw_md5_key)
1756 static inline struct tcp_md5sig_key *
1757 tcp_md5_do_lookup(const struct sock *sk, int l3index,
1758 const union tcp_md5_addr *addr, int family)
1763 static inline enum skb_drop_reason
1764 tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
1765 const void *saddr, const void *daddr,
1766 int family, int dif, int sdif)
1768 return SKB_NOT_DROPPED_YET;
1770 #define tcp_twsk_md5_key(twsk) NULL
1773 bool tcp_alloc_md5sig_pool(void);
1775 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void);
1776 static inline void tcp_put_md5sig_pool(void)
1781 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *, const struct sk_buff *,
1782 unsigned int header_len);
1783 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp,
1784 const struct tcp_md5sig_key *key);
1786 /* From tcp_fastopen.c */
1787 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
1788 struct tcp_fastopen_cookie *cookie);
1789 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
1790 struct tcp_fastopen_cookie *cookie, bool syn_lost,
1792 struct tcp_fastopen_request {
1793 /* Fast Open cookie. Size 0 means a cookie request */
1794 struct tcp_fastopen_cookie cookie;
1795 struct msghdr *data; /* data in MSG_FASTOPEN */
1797 int copied; /* queued in tcp_connect() */
1798 struct ubuf_info *uarg;
1800 void tcp_free_fastopen_req(struct tcp_sock *tp);
1801 void tcp_fastopen_destroy_cipher(struct sock *sk);
1802 void tcp_fastopen_ctx_destroy(struct net *net);
1803 int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
1804 void *primary_key, void *backup_key);
1805 int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
1807 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb);
1808 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
1809 struct request_sock *req,
1810 struct tcp_fastopen_cookie *foc,
1811 const struct dst_entry *dst);
1812 void tcp_fastopen_init_key_once(struct net *net);
1813 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
1814 struct tcp_fastopen_cookie *cookie);
1815 bool tcp_fastopen_defer_connect(struct sock *sk, int *err);
1816 #define TCP_FASTOPEN_KEY_LENGTH sizeof(siphash_key_t)
1817 #define TCP_FASTOPEN_KEY_MAX 2
1818 #define TCP_FASTOPEN_KEY_BUF_LENGTH \
1819 (TCP_FASTOPEN_KEY_LENGTH * TCP_FASTOPEN_KEY_MAX)
1821 /* Fastopen key context */
1822 struct tcp_fastopen_context {
1823 siphash_key_t key[TCP_FASTOPEN_KEY_MAX];
1825 struct rcu_head rcu;
1828 void tcp_fastopen_active_disable(struct sock *sk);
1829 bool tcp_fastopen_active_should_disable(struct sock *sk);
1830 void tcp_fastopen_active_disable_ofo_check(struct sock *sk);
1831 void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired);
1833 /* Caller needs to wrap with rcu_read_(un)lock() */
1835 struct tcp_fastopen_context *tcp_fastopen_get_ctx(const struct sock *sk)
1837 struct tcp_fastopen_context *ctx;
1839 ctx = rcu_dereference(inet_csk(sk)->icsk_accept_queue.fastopenq.ctx);
1841 ctx = rcu_dereference(sock_net(sk)->ipv4.tcp_fastopen_ctx);
1846 bool tcp_fastopen_cookie_match(const struct tcp_fastopen_cookie *foc,
1847 const struct tcp_fastopen_cookie *orig)
1849 if (orig->len == TCP_FASTOPEN_COOKIE_SIZE &&
1850 orig->len == foc->len &&
1851 !memcmp(orig->val, foc->val, foc->len))
1857 int tcp_fastopen_context_len(const struct tcp_fastopen_context *ctx)
1862 /* Latencies incurred by various limits for a sender. They are
1863 * chronograph-like stats that are mutually exclusive.
1867 TCP_CHRONO_BUSY, /* Actively sending data (non-empty write queue) */
1868 TCP_CHRONO_RWND_LIMITED, /* Stalled by insufficient receive window */
1869 TCP_CHRONO_SNDBUF_LIMITED, /* Stalled by insufficient send buffer */
1873 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type);
1874 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type);
1876 /* This helper is needed, because skb->tcp_tsorted_anchor uses
1877 * the same memory storage than skb->destructor/_skb_refdst
1879 static inline void tcp_skb_tsorted_anchor_cleanup(struct sk_buff *skb)
1881 skb->destructor = NULL;
1882 skb->_skb_refdst = 0UL;
1885 #define tcp_skb_tsorted_save(skb) { \
1886 unsigned long _save = skb->_skb_refdst; \
1887 skb->_skb_refdst = 0UL;
1889 #define tcp_skb_tsorted_restore(skb) \
1890 skb->_skb_refdst = _save; \
1893 void tcp_write_queue_purge(struct sock *sk);
1895 static inline struct sk_buff *tcp_rtx_queue_head(const struct sock *sk)
1897 return skb_rb_first(&sk->tcp_rtx_queue);
1900 static inline struct sk_buff *tcp_rtx_queue_tail(const struct sock *sk)
1902 return skb_rb_last(&sk->tcp_rtx_queue);
1905 static inline struct sk_buff *tcp_write_queue_tail(const struct sock *sk)
1907 return skb_peek_tail(&sk->sk_write_queue);
1910 #define tcp_for_write_queue_from_safe(skb, tmp, sk) \
1911 skb_queue_walk_from_safe(&(sk)->sk_write_queue, skb, tmp)
1913 static inline struct sk_buff *tcp_send_head(const struct sock *sk)
1915 return skb_peek(&sk->sk_write_queue);
1918 static inline bool tcp_skb_is_last(const struct sock *sk,
1919 const struct sk_buff *skb)
1921 return skb_queue_is_last(&sk->sk_write_queue, skb);
1925 * tcp_write_queue_empty - test if any payload (or FIN) is available in write queue
1928 * Since the write queue can have a temporary empty skb in it,
1929 * we must not use "return skb_queue_empty(&sk->sk_write_queue)"
1931 static inline bool tcp_write_queue_empty(const struct sock *sk)
1933 const struct tcp_sock *tp = tcp_sk(sk);
1935 return tp->write_seq == tp->snd_nxt;
1938 static inline bool tcp_rtx_queue_empty(const struct sock *sk)
1940 return RB_EMPTY_ROOT(&sk->tcp_rtx_queue);
1943 static inline bool tcp_rtx_and_write_queues_empty(const struct sock *sk)
1945 return tcp_rtx_queue_empty(sk) && tcp_write_queue_empty(sk);
1948 static inline void tcp_add_write_queue_tail(struct sock *sk, struct sk_buff *skb)
1950 __skb_queue_tail(&sk->sk_write_queue, skb);
1952 /* Queue it, remembering where we must start sending. */
1953 if (sk->sk_write_queue.next == skb)
1954 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
1957 /* Insert new before skb on the write queue of sk. */
1958 static inline void tcp_insert_write_queue_before(struct sk_buff *new,
1959 struct sk_buff *skb,
1962 __skb_queue_before(&sk->sk_write_queue, skb, new);
1965 static inline void tcp_unlink_write_queue(struct sk_buff *skb, struct sock *sk)
1967 tcp_skb_tsorted_anchor_cleanup(skb);
1968 __skb_unlink(skb, &sk->sk_write_queue);
1971 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb);
1973 static inline void tcp_rtx_queue_unlink(struct sk_buff *skb, struct sock *sk)
1975 tcp_skb_tsorted_anchor_cleanup(skb);
1976 rb_erase(&skb->rbnode, &sk->tcp_rtx_queue);
1979 static inline void tcp_rtx_queue_unlink_and_free(struct sk_buff *skb, struct sock *sk)
1981 list_del(&skb->tcp_tsorted_anchor);
1982 tcp_rtx_queue_unlink(skb, sk);
1983 tcp_wmem_free_skb(sk, skb);
1986 static inline void tcp_push_pending_frames(struct sock *sk)
1988 if (tcp_send_head(sk)) {
1989 struct tcp_sock *tp = tcp_sk(sk);
1991 __tcp_push_pending_frames(sk, tcp_current_mss(sk), tp->nonagle);
1995 /* Start sequence of the skb just after the highest skb with SACKed
1996 * bit, valid only if sacked_out > 0 or when the caller has ensured
1997 * validity by itself.
1999 static inline u32 tcp_highest_sack_seq(struct tcp_sock *tp)
2001 if (!tp->sacked_out)
2004 if (tp->highest_sack == NULL)
2007 return TCP_SKB_CB(tp->highest_sack)->seq;
2010 static inline void tcp_advance_highest_sack(struct sock *sk, struct sk_buff *skb)
2012 tcp_sk(sk)->highest_sack = skb_rb_next(skb);
2015 static inline struct sk_buff *tcp_highest_sack(struct sock *sk)
2017 return tcp_sk(sk)->highest_sack;
2020 static inline void tcp_highest_sack_reset(struct sock *sk)
2022 tcp_sk(sk)->highest_sack = tcp_rtx_queue_head(sk);
2025 /* Called when old skb is about to be deleted and replaced by new skb */
2026 static inline void tcp_highest_sack_replace(struct sock *sk,
2027 struct sk_buff *old,
2028 struct sk_buff *new)
2030 if (old == tcp_highest_sack(sk))
2031 tcp_sk(sk)->highest_sack = new;
2034 /* This helper checks if socket has IP_TRANSPARENT set */
2035 static inline bool inet_sk_transparent(const struct sock *sk)
2037 switch (sk->sk_state) {
2039 return inet_twsk(sk)->tw_transparent;
2040 case TCP_NEW_SYN_RECV:
2041 return inet_rsk(inet_reqsk(sk))->no_srccheck;
2043 return inet_test_bit(TRANSPARENT, sk);
2046 /* Determines whether this is a thin stream (which may suffer from
2047 * increased latency). Used to trigger latency-reducing mechanisms.
2049 static inline bool tcp_stream_is_thin(struct tcp_sock *tp)
2051 return tp->packets_out < 4 && !tcp_in_initial_slowstart(tp);
2055 enum tcp_seq_states {
2056 TCP_SEQ_STATE_LISTENING,
2057 TCP_SEQ_STATE_ESTABLISHED,
2060 void *tcp_seq_start(struct seq_file *seq, loff_t *pos);
2061 void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos);
2062 void tcp_seq_stop(struct seq_file *seq, void *v);
2064 struct tcp_seq_afinfo {
2068 struct tcp_iter_state {
2069 struct seq_net_private p;
2070 enum tcp_seq_states state;
2071 struct sock *syn_wait_sk;
2072 int bucket, offset, sbucket, num;
2076 extern struct request_sock_ops tcp_request_sock_ops;
2077 extern struct request_sock_ops tcp6_request_sock_ops;
2079 void tcp_v4_destroy_sock(struct sock *sk);
2081 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
2082 netdev_features_t features);
2083 struct sk_buff *tcp_gro_receive(struct list_head *head, struct sk_buff *skb);
2084 INDIRECT_CALLABLE_DECLARE(int tcp4_gro_complete(struct sk_buff *skb, int thoff));
2085 INDIRECT_CALLABLE_DECLARE(struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb));
2086 INDIRECT_CALLABLE_DECLARE(int tcp6_gro_complete(struct sk_buff *skb, int thoff));
2087 INDIRECT_CALLABLE_DECLARE(struct sk_buff *tcp6_gro_receive(struct list_head *head, struct sk_buff *skb));
2088 void tcp_gro_complete(struct sk_buff *skb);
2090 void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr);
2092 static inline u32 tcp_notsent_lowat(const struct tcp_sock *tp)
2094 struct net *net = sock_net((struct sock *)tp);
2097 val = READ_ONCE(tp->notsent_lowat);
2099 return val ?: READ_ONCE(net->ipv4.sysctl_tcp_notsent_lowat);
2102 bool tcp_stream_memory_free(const struct sock *sk, int wake);
2104 #ifdef CONFIG_PROC_FS
2105 int tcp4_proc_init(void);
2106 void tcp4_proc_exit(void);
2109 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req);
2110 int tcp_conn_request(struct request_sock_ops *rsk_ops,
2111 const struct tcp_request_sock_ops *af_ops,
2112 struct sock *sk, struct sk_buff *skb);
2114 /* TCP af-specific functions */
2115 struct tcp_sock_af_ops {
2116 #ifdef CONFIG_TCP_MD5SIG
2117 struct tcp_md5sig_key *(*md5_lookup) (const struct sock *sk,
2118 const struct sock *addr_sk);
2119 int (*calc_md5_hash)(char *location,
2120 const struct tcp_md5sig_key *md5,
2121 const struct sock *sk,
2122 const struct sk_buff *skb);
2123 int (*md5_parse)(struct sock *sk,
2130 struct tcp_request_sock_ops {
2132 #ifdef CONFIG_TCP_MD5SIG
2133 struct tcp_md5sig_key *(*req_md5_lookup)(const struct sock *sk,
2134 const struct sock *addr_sk);
2135 int (*calc_md5_hash) (char *location,
2136 const struct tcp_md5sig_key *md5,
2137 const struct sock *sk,
2138 const struct sk_buff *skb);
2140 #ifdef CONFIG_SYN_COOKIES
2141 __u32 (*cookie_init_seq)(const struct sk_buff *skb,
2144 struct dst_entry *(*route_req)(const struct sock *sk,
2145 struct sk_buff *skb,
2147 struct request_sock *req);
2148 u32 (*init_seq)(const struct sk_buff *skb);
2149 u32 (*init_ts_off)(const struct net *net, const struct sk_buff *skb);
2150 int (*send_synack)(const struct sock *sk, struct dst_entry *dst,
2151 struct flowi *fl, struct request_sock *req,
2152 struct tcp_fastopen_cookie *foc,
2153 enum tcp_synack_type synack_type,
2154 struct sk_buff *syn_skb);
2157 extern const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops;
2158 #if IS_ENABLED(CONFIG_IPV6)
2159 extern const struct tcp_request_sock_ops tcp_request_sock_ipv6_ops;
2162 #ifdef CONFIG_SYN_COOKIES
2163 static inline __u32 cookie_init_sequence(const struct tcp_request_sock_ops *ops,
2164 const struct sock *sk, struct sk_buff *skb,
2167 tcp_synq_overflow(sk);
2168 __NET_INC_STATS(sock_net(sk), LINUX_MIB_SYNCOOKIESSENT);
2169 return ops->cookie_init_seq(skb, mss);
2172 static inline __u32 cookie_init_sequence(const struct tcp_request_sock_ops *ops,
2173 const struct sock *sk, struct sk_buff *skb,
2180 int tcpv4_offload_init(void);
2182 void tcp_v4_init(void);
2183 void tcp_init(void);
2185 /* tcp_recovery.c */
2186 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb);
2187 void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced);
2188 extern s32 tcp_rack_skb_timeout(struct tcp_sock *tp, struct sk_buff *skb,
2190 extern bool tcp_rack_mark_lost(struct sock *sk);
2191 extern void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
2193 extern void tcp_rack_reo_timeout(struct sock *sk);
2194 extern void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs);
2199 * Scaling factor for fractions in PLB. For example, tcp_plb_update_state
2200 * expects cong_ratio which represents fraction of traffic that experienced
2201 * congestion over a single RTT. In order to avoid floating point operations,
2202 * this fraction should be mapped to (1 << TCP_PLB_SCALE) and passed in.
2204 #define TCP_PLB_SCALE 8
2206 /* State for PLB (Protective Load Balancing) for a single TCP connection. */
2207 struct tcp_plb_state {
2208 u8 consec_cong_rounds:5, /* consecutive congested rounds */
2210 u32 pause_until; /* jiffies32 when PLB can resume rerouting */
2213 static inline void tcp_plb_init(const struct sock *sk,
2214 struct tcp_plb_state *plb)
2216 plb->consec_cong_rounds = 0;
2217 plb->pause_until = 0;
2219 void tcp_plb_update_state(const struct sock *sk, struct tcp_plb_state *plb,
2220 const int cong_ratio);
2221 void tcp_plb_check_rehash(struct sock *sk, struct tcp_plb_state *plb);
2222 void tcp_plb_update_state_upon_rto(struct sock *sk, struct tcp_plb_state *plb);
2224 /* At how many usecs into the future should the RTO fire? */
2225 static inline s64 tcp_rto_delta_us(const struct sock *sk)
2227 const struct sk_buff *skb = tcp_rtx_queue_head(sk);
2228 u32 rto = inet_csk(sk)->icsk_rto;
2229 u64 rto_time_stamp_us = tcp_skb_timestamp_us(skb) + jiffies_to_usecs(rto);
2231 return rto_time_stamp_us - tcp_sk(sk)->tcp_mstamp;
2235 * Save and compile IPv4 options, return a pointer to it
2237 static inline struct ip_options_rcu *tcp_v4_save_options(struct net *net,
2238 struct sk_buff *skb)
2240 const struct ip_options *opt = &TCP_SKB_CB(skb)->header.h4.opt;
2241 struct ip_options_rcu *dopt = NULL;
2244 int opt_size = sizeof(*dopt) + opt->optlen;
2246 dopt = kmalloc(opt_size, GFP_ATOMIC);
2247 if (dopt && __ip_options_echo(net, &dopt->opt, skb, opt)) {
2255 /* locally generated TCP pure ACKs have skb->truesize == 2
2256 * (check tcp_send_ack() in net/ipv4/tcp_output.c )
2257 * This is much faster than dissecting the packet to find out.
2258 * (Think of GRE encapsulations, IPv4, IPv6, ...)
2260 static inline bool skb_is_tcp_pure_ack(const struct sk_buff *skb)
2262 return skb->truesize == 2;
2265 static inline void skb_set_tcp_pure_ack(struct sk_buff *skb)
2270 static inline int tcp_inq(struct sock *sk)
2272 struct tcp_sock *tp = tcp_sk(sk);
2275 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
2277 } else if (sock_flag(sk, SOCK_URGINLINE) ||
2279 before(tp->urg_seq, tp->copied_seq) ||
2280 !before(tp->urg_seq, tp->rcv_nxt)) {
2282 answ = tp->rcv_nxt - tp->copied_seq;
2284 /* Subtract 1, if FIN was received */
2285 if (answ && sock_flag(sk, SOCK_DONE))
2288 answ = tp->urg_seq - tp->copied_seq;
2294 int tcp_peek_len(struct socket *sock);
2296 static inline void tcp_segs_in(struct tcp_sock *tp, const struct sk_buff *skb)
2300 segs_in = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
2302 /* We update these fields while other threads might
2303 * read them from tcp_get_info()
2305 WRITE_ONCE(tp->segs_in, tp->segs_in + segs_in);
2306 if (skb->len > tcp_hdrlen(skb))
2307 WRITE_ONCE(tp->data_segs_in, tp->data_segs_in + segs_in);
2311 * TCP listen path runs lockless.
2312 * We forced "struct sock" to be const qualified to make sure
2313 * we don't modify one of its field by mistake.
2314 * Here, we increment sk_drops which is an atomic_t, so we can safely
2315 * make sock writable again.
2317 static inline void tcp_listendrop(const struct sock *sk)
2319 atomic_inc(&((struct sock *)sk)->sk_drops);
2320 __NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENDROPS);
2323 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer);
2326 * Interface for adding Upper Level Protocols over TCP
2329 #define TCP_ULP_NAME_MAX 16
2330 #define TCP_ULP_MAX 128
2331 #define TCP_ULP_BUF_MAX (TCP_ULP_NAME_MAX*TCP_ULP_MAX)
2333 struct tcp_ulp_ops {
2334 struct list_head list;
2336 /* initialize ulp */
2337 int (*init)(struct sock *sk);
2339 void (*update)(struct sock *sk, struct proto *p,
2340 void (*write_space)(struct sock *sk));
2342 void (*release)(struct sock *sk);
2344 int (*get_info)(const struct sock *sk, struct sk_buff *skb);
2345 size_t (*get_info_size)(const struct sock *sk);
2347 void (*clone)(const struct request_sock *req, struct sock *newsk,
2348 const gfp_t priority);
2350 char name[TCP_ULP_NAME_MAX];
2351 struct module *owner;
2353 int tcp_register_ulp(struct tcp_ulp_ops *type);
2354 void tcp_unregister_ulp(struct tcp_ulp_ops *type);
2355 int tcp_set_ulp(struct sock *sk, const char *name);
2356 void tcp_get_available_ulp(char *buf, size_t len);
2357 void tcp_cleanup_ulp(struct sock *sk);
2358 void tcp_update_ulp(struct sock *sk, struct proto *p,
2359 void (*write_space)(struct sock *sk));
2361 #define MODULE_ALIAS_TCP_ULP(name) \
2362 __MODULE_INFO(alias, alias_userspace, name); \
2363 __MODULE_INFO(alias, alias_tcp_ulp, "tcp-ulp-" name)
2365 #ifdef CONFIG_NET_SOCK_MSG
2369 #ifdef CONFIG_BPF_SYSCALL
2370 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore);
2371 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk);
2372 #endif /* CONFIG_BPF_SYSCALL */
2375 void tcp_eat_skb(struct sock *sk, struct sk_buff *skb);
2377 static inline void tcp_eat_skb(struct sock *sk, struct sk_buff *skb)
2382 int tcp_bpf_sendmsg_redir(struct sock *sk, bool ingress,
2383 struct sk_msg *msg, u32 bytes, int flags);
2384 #endif /* CONFIG_NET_SOCK_MSG */
2386 #if !defined(CONFIG_BPF_SYSCALL) || !defined(CONFIG_NET_SOCK_MSG)
2387 static inline void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
2392 #ifdef CONFIG_CGROUP_BPF
2393 static inline void bpf_skops_init_skb(struct bpf_sock_ops_kern *skops,
2394 struct sk_buff *skb,
2395 unsigned int end_offset)
2398 skops->skb_data_end = skb->data + end_offset;
2401 static inline void bpf_skops_init_skb(struct bpf_sock_ops_kern *skops,
2402 struct sk_buff *skb,
2403 unsigned int end_offset)
2408 /* Call BPF_SOCK_OPS program that returns an int. If the return value
2409 * is < 0, then the BPF op failed (for example if the loaded BPF
2410 * program does not support the chosen operation or there is no BPF
2414 static inline int tcp_call_bpf(struct sock *sk, int op, u32 nargs, u32 *args)
2416 struct bpf_sock_ops_kern sock_ops;
2419 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
2420 if (sk_fullsock(sk)) {
2421 sock_ops.is_fullsock = 1;
2422 sock_owned_by_me(sk);
2428 memcpy(sock_ops.args, args, nargs * sizeof(*args));
2430 ret = BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
2432 ret = sock_ops.reply;
2438 static inline int tcp_call_bpf_2arg(struct sock *sk, int op, u32 arg1, u32 arg2)
2440 u32 args[2] = {arg1, arg2};
2442 return tcp_call_bpf(sk, op, 2, args);
2445 static inline int tcp_call_bpf_3arg(struct sock *sk, int op, u32 arg1, u32 arg2,
2448 u32 args[3] = {arg1, arg2, arg3};
2450 return tcp_call_bpf(sk, op, 3, args);
2454 static inline int tcp_call_bpf(struct sock *sk, int op, u32 nargs, u32 *args)
2459 static inline int tcp_call_bpf_2arg(struct sock *sk, int op, u32 arg1, u32 arg2)
2464 static inline int tcp_call_bpf_3arg(struct sock *sk, int op, u32 arg1, u32 arg2,
2472 static inline u32 tcp_timeout_init(struct sock *sk)
2476 timeout = tcp_call_bpf(sk, BPF_SOCK_OPS_TIMEOUT_INIT, 0, NULL);
2479 timeout = TCP_TIMEOUT_INIT;
2480 return min_t(int, timeout, TCP_RTO_MAX);
2483 static inline u32 tcp_rwnd_init_bpf(struct sock *sk)
2487 rwnd = tcp_call_bpf(sk, BPF_SOCK_OPS_RWND_INIT, 0, NULL);
2494 static inline bool tcp_bpf_ca_needs_ecn(struct sock *sk)
2496 return (tcp_call_bpf(sk, BPF_SOCK_OPS_NEEDS_ECN, 0, NULL) == 1);
2499 static inline void tcp_bpf_rtt(struct sock *sk)
2501 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_RTT_CB_FLAG))
2502 tcp_call_bpf(sk, BPF_SOCK_OPS_RTT_CB, 0, NULL);
2505 #if IS_ENABLED(CONFIG_SMC)
2506 extern struct static_key_false tcp_have_smc;
2509 #if IS_ENABLED(CONFIG_TLS_DEVICE)
2510 void clean_acked_data_enable(struct inet_connection_sock *icsk,
2511 void (*cad)(struct sock *sk, u32 ack_seq));
2512 void clean_acked_data_disable(struct inet_connection_sock *icsk);
2513 void clean_acked_data_flush(void);
2516 DECLARE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2517 static inline void tcp_add_tx_delay(struct sk_buff *skb,
2518 const struct tcp_sock *tp)
2520 if (static_branch_unlikely(&tcp_tx_delay_enabled))
2521 skb->skb_mstamp_ns += (u64)tp->tcp_tx_delay * NSEC_PER_USEC;
2524 /* Compute Earliest Departure Time for some control packets
2525 * like ACK or RST for TIME_WAIT or non ESTABLISHED sockets.
2527 static inline u64 tcp_transmit_time(const struct sock *sk)
2529 if (static_branch_unlikely(&tcp_tx_delay_enabled)) {
2530 u32 delay = (sk->sk_state == TCP_TIME_WAIT) ?
2531 tcp_twsk(sk)->tw_tx_delay : tcp_sk(sk)->tcp_tx_delay;
2533 return tcp_clock_ns() + (u64)delay * NSEC_PER_USEC;