2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly = 1;
79 int sysctl_tcp_window_scaling __read_mostly = 1;
80 int sysctl_tcp_sack __read_mostly = 1;
81 int sysctl_tcp_fack __read_mostly = 1;
82 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
83 EXPORT_SYMBOL(sysctl_tcp_reordering);
84 int sysctl_tcp_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit = 100;
92 int sysctl_tcp_stdurg __read_mostly;
93 int sysctl_tcp_rfc1337 __read_mostly;
94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
95 int sysctl_tcp_frto __read_mostly = 2;
97 int sysctl_tcp_thin_dupack __read_mostly;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
100 int sysctl_tcp_early_retrans __read_mostly = 3;
102 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
103 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
108 #define FLAG_ECE 0x40 /* ECE in this ACK */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 /* Adapt the MSS value used to make delayed ack decision to the
127 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
129 struct inet_connection_sock *icsk = inet_csk(sk);
130 const unsigned int lss = icsk->icsk_ack.last_seg_size;
133 icsk->icsk_ack.last_seg_size = 0;
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
138 len = skb_shinfo(skb)->gso_size ? : skb->len;
139 if (len >= icsk->icsk_ack.rcv_mss) {
140 icsk->icsk_ack.rcv_mss = len;
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
145 * "len" is invariant segment length, including TCP header.
147 len += skb->data - skb_transport_header(skb);
148 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
154 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
155 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
160 len -= tcp_sk(sk)->tcp_header_len;
161 icsk->icsk_ack.last_seg_size = len;
163 icsk->icsk_ack.rcv_mss = len;
167 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
173 static void tcp_incr_quickack(struct sock *sk)
175 struct inet_connection_sock *icsk = inet_csk(sk);
176 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
180 if (quickacks > icsk->icsk_ack.quick)
181 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
184 static void tcp_enter_quickack_mode(struct sock *sk)
186 struct inet_connection_sock *icsk = inet_csk(sk);
187 tcp_incr_quickack(sk);
188 icsk->icsk_ack.pingpong = 0;
189 icsk->icsk_ack.ato = TCP_ATO_MIN;
192 /* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
196 static inline bool tcp_in_quickack_mode(const struct sock *sk)
198 const struct inet_connection_sock *icsk = inet_csk(sk);
200 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
205 if (tp->ecn_flags & TCP_ECN_OK)
206 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
211 if (tcp_hdr(skb)->cwr)
212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
220 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
222 if (!(tp->ecn_flags & TCP_ECN_OK))
225 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
226 case INET_ECN_NOT_ECT:
227 /* Funny extension: if ECT is not set on a segment,
228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
231 if (tp->ecn_flags & TCP_ECN_SEEN)
232 tcp_enter_quickack_mode((struct sock *)tp);
235 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
236 /* Better not delay acks, sender can have a very low cwnd */
237 tcp_enter_quickack_mode((struct sock *)tp);
238 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
242 tp->ecn_flags |= TCP_ECN_SEEN;
246 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
248 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
249 tp->ecn_flags &= ~TCP_ECN_OK;
252 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
254 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
255 tp->ecn_flags &= ~TCP_ECN_OK;
258 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
260 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
265 /* Buffer size and advertised window tuning.
267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
270 static void tcp_fixup_sndbuf(struct sock *sk)
272 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
274 sndmem *= TCP_INIT_CWND;
275 if (sk->sk_sndbuf < sndmem)
276 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
279 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
281 * All tcp_full_space() is split to two parts: "network" buffer, allocated
282 * forward and advertised in receiver window (tp->rcv_wnd) and
283 * "application buffer", required to isolate scheduling/application
284 * latencies from network.
285 * window_clamp is maximal advertised window. It can be less than
286 * tcp_full_space(), in this case tcp_full_space() - window_clamp
287 * is reserved for "application" buffer. The less window_clamp is
288 * the smoother our behaviour from viewpoint of network, but the lower
289 * throughput and the higher sensitivity of the connection to losses. 8)
291 * rcv_ssthresh is more strict window_clamp used at "slow start"
292 * phase to predict further behaviour of this connection.
293 * It is used for two goals:
294 * - to enforce header prediction at sender, even when application
295 * requires some significant "application buffer". It is check #1.
296 * - to prevent pruning of receive queue because of misprediction
297 * of receiver window. Check #2.
299 * The scheme does not work when sender sends good segments opening
300 * window and then starts to feed us spaghetti. But it should work
301 * in common situations. Otherwise, we have to rely on queue collapsing.
304 /* Slow part of check#2. */
305 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
307 struct tcp_sock *tp = tcp_sk(sk);
309 int truesize = tcp_win_from_space(skb->truesize) >> 1;
310 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
312 while (tp->rcv_ssthresh <= window) {
313 if (truesize <= skb->len)
314 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
322 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
324 struct tcp_sock *tp = tcp_sk(sk);
327 if (tp->rcv_ssthresh < tp->window_clamp &&
328 (int)tp->rcv_ssthresh < tcp_space(sk) &&
329 !sk_under_memory_pressure(sk)) {
332 /* Check #2. Increase window, if skb with such overhead
333 * will fit to rcvbuf in future.
335 if (tcp_win_from_space(skb->truesize) <= skb->len)
336 incr = 2 * tp->advmss;
338 incr = __tcp_grow_window(sk, skb);
341 incr = max_t(int, incr, 2 * skb->len);
342 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
344 inet_csk(sk)->icsk_ack.quick |= 1;
349 /* 3. Tuning rcvbuf, when connection enters established state. */
350 static void tcp_fixup_rcvbuf(struct sock *sk)
352 u32 mss = tcp_sk(sk)->advmss;
355 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
356 tcp_default_init_rwnd(mss);
358 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
359 * Allow enough cushion so that sender is not limited by our window
361 if (sysctl_tcp_moderate_rcvbuf)
364 if (sk->sk_rcvbuf < rcvmem)
365 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
368 /* 4. Try to fixup all. It is made immediately after connection enters
371 void tcp_init_buffer_space(struct sock *sk)
373 struct tcp_sock *tp = tcp_sk(sk);
376 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
377 tcp_fixup_rcvbuf(sk);
378 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
379 tcp_fixup_sndbuf(sk);
381 tp->rcvq_space.space = tp->rcv_wnd;
382 tp->rcvq_space.time = tcp_time_stamp;
383 tp->rcvq_space.seq = tp->copied_seq;
385 maxwin = tcp_full_space(sk);
387 if (tp->window_clamp >= maxwin) {
388 tp->window_clamp = maxwin;
390 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
391 tp->window_clamp = max(maxwin -
392 (maxwin >> sysctl_tcp_app_win),
396 /* Force reservation of one segment. */
397 if (sysctl_tcp_app_win &&
398 tp->window_clamp > 2 * tp->advmss &&
399 tp->window_clamp + tp->advmss > maxwin)
400 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
402 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
403 tp->snd_cwnd_stamp = tcp_time_stamp;
406 /* 5. Recalculate window clamp after socket hit its memory bounds. */
407 static void tcp_clamp_window(struct sock *sk)
409 struct tcp_sock *tp = tcp_sk(sk);
410 struct inet_connection_sock *icsk = inet_csk(sk);
412 icsk->icsk_ack.quick = 0;
414 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
415 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
416 !sk_under_memory_pressure(sk) &&
417 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
418 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
421 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
422 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
425 /* Initialize RCV_MSS value.
426 * RCV_MSS is an our guess about MSS used by the peer.
427 * We haven't any direct information about the MSS.
428 * It's better to underestimate the RCV_MSS rather than overestimate.
429 * Overestimations make us ACKing less frequently than needed.
430 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
432 void tcp_initialize_rcv_mss(struct sock *sk)
434 const struct tcp_sock *tp = tcp_sk(sk);
435 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
437 hint = min(hint, tp->rcv_wnd / 2);
438 hint = min(hint, TCP_MSS_DEFAULT);
439 hint = max(hint, TCP_MIN_MSS);
441 inet_csk(sk)->icsk_ack.rcv_mss = hint;
443 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
445 /* Receiver "autotuning" code.
447 * The algorithm for RTT estimation w/o timestamps is based on
448 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
449 * <http://public.lanl.gov/radiant/pubs.html#DRS>
451 * More detail on this code can be found at
452 * <http://staff.psc.edu/jheffner/>,
453 * though this reference is out of date. A new paper
456 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
458 u32 new_sample = tp->rcv_rtt_est.rtt;
464 if (new_sample != 0) {
465 /* If we sample in larger samples in the non-timestamp
466 * case, we could grossly overestimate the RTT especially
467 * with chatty applications or bulk transfer apps which
468 * are stalled on filesystem I/O.
470 * Also, since we are only going for a minimum in the
471 * non-timestamp case, we do not smooth things out
472 * else with timestamps disabled convergence takes too
476 m -= (new_sample >> 3);
484 /* No previous measure. */
488 if (tp->rcv_rtt_est.rtt != new_sample)
489 tp->rcv_rtt_est.rtt = new_sample;
492 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
494 if (tp->rcv_rtt_est.time == 0)
496 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
498 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
501 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
502 tp->rcv_rtt_est.time = tcp_time_stamp;
505 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
506 const struct sk_buff *skb)
508 struct tcp_sock *tp = tcp_sk(sk);
509 if (tp->rx_opt.rcv_tsecr &&
510 (TCP_SKB_CB(skb)->end_seq -
511 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
512 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
516 * This function should be called every time data is copied to user space.
517 * It calculates the appropriate TCP receive buffer space.
519 void tcp_rcv_space_adjust(struct sock *sk)
521 struct tcp_sock *tp = tcp_sk(sk);
525 time = tcp_time_stamp - tp->rcvq_space.time;
526 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
529 /* Number of bytes copied to user in last RTT */
530 copied = tp->copied_seq - tp->rcvq_space.seq;
531 if (copied <= tp->rcvq_space.space)
535 * copied = bytes received in previous RTT, our base window
536 * To cope with packet losses, we need a 2x factor
537 * To cope with slow start, and sender growing its cwin by 100 %
538 * every RTT, we need a 4x factor, because the ACK we are sending
539 * now is for the next RTT, not the current one :
540 * <prev RTT . ><current RTT .. ><next RTT .... >
543 if (sysctl_tcp_moderate_rcvbuf &&
544 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
545 int rcvwin, rcvmem, rcvbuf;
547 /* minimal window to cope with packet losses, assuming
548 * steady state. Add some cushion because of small variations.
550 rcvwin = (copied << 1) + 16 * tp->advmss;
552 /* If rate increased by 25%,
553 * assume slow start, rcvwin = 3 * copied
554 * If rate increased by 50%,
555 * assume sender can use 2x growth, rcvwin = 4 * copied
558 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
560 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
563 rcvwin += (rcvwin >> 1);
566 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
567 while (tcp_win_from_space(rcvmem) < tp->advmss)
570 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
571 if (rcvbuf > sk->sk_rcvbuf) {
572 sk->sk_rcvbuf = rcvbuf;
574 /* Make the window clamp follow along. */
575 tp->window_clamp = rcvwin;
578 tp->rcvq_space.space = copied;
581 tp->rcvq_space.seq = tp->copied_seq;
582 tp->rcvq_space.time = tcp_time_stamp;
585 /* There is something which you must keep in mind when you analyze the
586 * behavior of the tp->ato delayed ack timeout interval. When a
587 * connection starts up, we want to ack as quickly as possible. The
588 * problem is that "good" TCP's do slow start at the beginning of data
589 * transmission. The means that until we send the first few ACK's the
590 * sender will sit on his end and only queue most of his data, because
591 * he can only send snd_cwnd unacked packets at any given time. For
592 * each ACK we send, he increments snd_cwnd and transmits more of his
595 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
597 struct tcp_sock *tp = tcp_sk(sk);
598 struct inet_connection_sock *icsk = inet_csk(sk);
601 inet_csk_schedule_ack(sk);
603 tcp_measure_rcv_mss(sk, skb);
605 tcp_rcv_rtt_measure(tp);
607 now = tcp_time_stamp;
609 if (!icsk->icsk_ack.ato) {
610 /* The _first_ data packet received, initialize
611 * delayed ACK engine.
613 tcp_incr_quickack(sk);
614 icsk->icsk_ack.ato = TCP_ATO_MIN;
616 int m = now - icsk->icsk_ack.lrcvtime;
618 if (m <= TCP_ATO_MIN / 2) {
619 /* The fastest case is the first. */
620 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
621 } else if (m < icsk->icsk_ack.ato) {
622 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
623 if (icsk->icsk_ack.ato > icsk->icsk_rto)
624 icsk->icsk_ack.ato = icsk->icsk_rto;
625 } else if (m > icsk->icsk_rto) {
626 /* Too long gap. Apparently sender failed to
627 * restart window, so that we send ACKs quickly.
629 tcp_incr_quickack(sk);
633 icsk->icsk_ack.lrcvtime = now;
635 TCP_ECN_check_ce(tp, skb);
638 tcp_grow_window(sk, skb);
641 /* Called to compute a smoothed rtt estimate. The data fed to this
642 * routine either comes from timestamps, or from segments that were
643 * known _not_ to have been retransmitted [see Karn/Partridge
644 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
645 * piece by Van Jacobson.
646 * NOTE: the next three routines used to be one big routine.
647 * To save cycles in the RFC 1323 implementation it was better to break
648 * it up into three procedures. -- erics
650 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
652 struct tcp_sock *tp = tcp_sk(sk);
653 long m = mrtt; /* RTT */
655 /* The following amusing code comes from Jacobson's
656 * article in SIGCOMM '88. Note that rtt and mdev
657 * are scaled versions of rtt and mean deviation.
658 * This is designed to be as fast as possible
659 * m stands for "measurement".
661 * On a 1990 paper the rto value is changed to:
662 * RTO = rtt + 4 * mdev
664 * Funny. This algorithm seems to be very broken.
665 * These formulae increase RTO, when it should be decreased, increase
666 * too slowly, when it should be increased quickly, decrease too quickly
667 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
668 * does not matter how to _calculate_ it. Seems, it was trap
669 * that VJ failed to avoid. 8)
674 m -= (tp->srtt >> 3); /* m is now error in rtt est */
675 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
677 m = -m; /* m is now abs(error) */
678 m -= (tp->mdev >> 2); /* similar update on mdev */
679 /* This is similar to one of Eifel findings.
680 * Eifel blocks mdev updates when rtt decreases.
681 * This solution is a bit different: we use finer gain
682 * for mdev in this case (alpha*beta).
683 * Like Eifel it also prevents growth of rto,
684 * but also it limits too fast rto decreases,
685 * happening in pure Eifel.
690 m -= (tp->mdev >> 2); /* similar update on mdev */
692 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
693 if (tp->mdev > tp->mdev_max) {
694 tp->mdev_max = tp->mdev;
695 if (tp->mdev_max > tp->rttvar)
696 tp->rttvar = tp->mdev_max;
698 if (after(tp->snd_una, tp->rtt_seq)) {
699 if (tp->mdev_max < tp->rttvar)
700 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
701 tp->rtt_seq = tp->snd_nxt;
702 tp->mdev_max = tcp_rto_min(sk);
705 /* no previous measure. */
706 tp->srtt = m << 3; /* take the measured time to be rtt */
707 tp->mdev = m << 1; /* make sure rto = 3*rtt */
708 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
709 tp->rtt_seq = tp->snd_nxt;
713 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
714 * Note: TCP stack does not yet implement pacing.
715 * FQ packet scheduler can be used to implement cheap but effective
716 * TCP pacing, to smooth the burst on large writes when packets
717 * in flight is significantly lower than cwnd (or rwin)
719 static void tcp_update_pacing_rate(struct sock *sk)
721 const struct tcp_sock *tp = tcp_sk(sk);
724 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
725 rate = (u64)tp->mss_cache * 2 * (HZ << 3);
727 rate *= max(tp->snd_cwnd, tp->packets_out);
729 /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
730 * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
731 * We probably need usec resolution in the future.
732 * Note: This also takes care of possible srtt=0 case,
733 * when tcp_rtt_estimator() was not yet called.
735 if (tp->srtt > 8 + 2)
736 do_div(rate, tp->srtt);
738 sk->sk_pacing_rate = min_t(u64, rate, sk->sk_max_pacing_rate);
741 /* Calculate rto without backoff. This is the second half of Van Jacobson's
742 * routine referred to above.
744 void tcp_set_rto(struct sock *sk)
746 const struct tcp_sock *tp = tcp_sk(sk);
747 /* Old crap is replaced with new one. 8)
750 * 1. If rtt variance happened to be less 50msec, it is hallucination.
751 * It cannot be less due to utterly erratic ACK generation made
752 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
753 * to do with delayed acks, because at cwnd>2 true delack timeout
754 * is invisible. Actually, Linux-2.4 also generates erratic
755 * ACKs in some circumstances.
757 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
759 /* 2. Fixups made earlier cannot be right.
760 * If we do not estimate RTO correctly without them,
761 * all the algo is pure shit and should be replaced
762 * with correct one. It is exactly, which we pretend to do.
765 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
766 * guarantees that rto is higher.
771 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
773 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
776 cwnd = TCP_INIT_CWND;
777 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
781 * Packet counting of FACK is based on in-order assumptions, therefore TCP
782 * disables it when reordering is detected
784 void tcp_disable_fack(struct tcp_sock *tp)
786 /* RFC3517 uses different metric in lost marker => reset on change */
788 tp->lost_skb_hint = NULL;
789 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
792 /* Take a notice that peer is sending D-SACKs */
793 static void tcp_dsack_seen(struct tcp_sock *tp)
795 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
798 static void tcp_update_reordering(struct sock *sk, const int metric,
801 struct tcp_sock *tp = tcp_sk(sk);
802 if (metric > tp->reordering) {
805 tp->reordering = min(TCP_MAX_REORDERING, metric);
807 /* This exciting event is worth to be remembered. 8) */
809 mib_idx = LINUX_MIB_TCPTSREORDER;
810 else if (tcp_is_reno(tp))
811 mib_idx = LINUX_MIB_TCPRENOREORDER;
812 else if (tcp_is_fack(tp))
813 mib_idx = LINUX_MIB_TCPFACKREORDER;
815 mib_idx = LINUX_MIB_TCPSACKREORDER;
817 NET_INC_STATS_BH(sock_net(sk), mib_idx);
818 #if FASTRETRANS_DEBUG > 1
819 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
820 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
824 tp->undo_marker ? tp->undo_retrans : 0);
826 tcp_disable_fack(tp);
830 tcp_disable_early_retrans(tp);
833 /* This must be called before lost_out is incremented */
834 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
836 if ((tp->retransmit_skb_hint == NULL) ||
837 before(TCP_SKB_CB(skb)->seq,
838 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
839 tp->retransmit_skb_hint = skb;
842 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
843 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
846 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
848 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
849 tcp_verify_retransmit_hint(tp, skb);
851 tp->lost_out += tcp_skb_pcount(skb);
852 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
856 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
859 tcp_verify_retransmit_hint(tp, skb);
861 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
862 tp->lost_out += tcp_skb_pcount(skb);
863 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
867 /* This procedure tags the retransmission queue when SACKs arrive.
869 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
870 * Packets in queue with these bits set are counted in variables
871 * sacked_out, retrans_out and lost_out, correspondingly.
873 * Valid combinations are:
874 * Tag InFlight Description
875 * 0 1 - orig segment is in flight.
876 * S 0 - nothing flies, orig reached receiver.
877 * L 0 - nothing flies, orig lost by net.
878 * R 2 - both orig and retransmit are in flight.
879 * L|R 1 - orig is lost, retransmit is in flight.
880 * S|R 1 - orig reached receiver, retrans is still in flight.
881 * (L|S|R is logically valid, it could occur when L|R is sacked,
882 * but it is equivalent to plain S and code short-curcuits it to S.
883 * L|S is logically invalid, it would mean -1 packet in flight 8))
885 * These 6 states form finite state machine, controlled by the following events:
886 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
887 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
888 * 3. Loss detection event of two flavors:
889 * A. Scoreboard estimator decided the packet is lost.
890 * A'. Reno "three dupacks" marks head of queue lost.
891 * A''. Its FACK modification, head until snd.fack is lost.
892 * B. SACK arrives sacking SND.NXT at the moment, when the
893 * segment was retransmitted.
894 * 4. D-SACK added new rule: D-SACK changes any tag to S.
896 * It is pleasant to note, that state diagram turns out to be commutative,
897 * so that we are allowed not to be bothered by order of our actions,
898 * when multiple events arrive simultaneously. (see the function below).
900 * Reordering detection.
901 * --------------------
902 * Reordering metric is maximal distance, which a packet can be displaced
903 * in packet stream. With SACKs we can estimate it:
905 * 1. SACK fills old hole and the corresponding segment was not
906 * ever retransmitted -> reordering. Alas, we cannot use it
907 * when segment was retransmitted.
908 * 2. The last flaw is solved with D-SACK. D-SACK arrives
909 * for retransmitted and already SACKed segment -> reordering..
910 * Both of these heuristics are not used in Loss state, when we cannot
911 * account for retransmits accurately.
913 * SACK block validation.
914 * ----------------------
916 * SACK block range validation checks that the received SACK block fits to
917 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
918 * Note that SND.UNA is not included to the range though being valid because
919 * it means that the receiver is rather inconsistent with itself reporting
920 * SACK reneging when it should advance SND.UNA. Such SACK block this is
921 * perfectly valid, however, in light of RFC2018 which explicitly states
922 * that "SACK block MUST reflect the newest segment. Even if the newest
923 * segment is going to be discarded ...", not that it looks very clever
924 * in case of head skb. Due to potentional receiver driven attacks, we
925 * choose to avoid immediate execution of a walk in write queue due to
926 * reneging and defer head skb's loss recovery to standard loss recovery
927 * procedure that will eventually trigger (nothing forbids us doing this).
929 * Implements also blockage to start_seq wrap-around. Problem lies in the
930 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
931 * there's no guarantee that it will be before snd_nxt (n). The problem
932 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
935 * <- outs wnd -> <- wrapzone ->
936 * u e n u_w e_w s n_w
938 * |<------------+------+----- TCP seqno space --------------+---------->|
939 * ...-- <2^31 ->| |<--------...
940 * ...---- >2^31 ------>| |<--------...
942 * Current code wouldn't be vulnerable but it's better still to discard such
943 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
944 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
945 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
946 * equal to the ideal case (infinite seqno space without wrap caused issues).
948 * With D-SACK the lower bound is extended to cover sequence space below
949 * SND.UNA down to undo_marker, which is the last point of interest. Yet
950 * again, D-SACK block must not to go across snd_una (for the same reason as
951 * for the normal SACK blocks, explained above). But there all simplicity
952 * ends, TCP might receive valid D-SACKs below that. As long as they reside
953 * fully below undo_marker they do not affect behavior in anyway and can
954 * therefore be safely ignored. In rare cases (which are more or less
955 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
956 * fragmentation and packet reordering past skb's retransmission. To consider
957 * them correctly, the acceptable range must be extended even more though
958 * the exact amount is rather hard to quantify. However, tp->max_window can
959 * be used as an exaggerated estimate.
961 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
962 u32 start_seq, u32 end_seq)
964 /* Too far in future, or reversed (interpretation is ambiguous) */
965 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
968 /* Nasty start_seq wrap-around check (see comments above) */
969 if (!before(start_seq, tp->snd_nxt))
972 /* In outstanding window? ...This is valid exit for D-SACKs too.
973 * start_seq == snd_una is non-sensical (see comments above)
975 if (after(start_seq, tp->snd_una))
978 if (!is_dsack || !tp->undo_marker)
981 /* ...Then it's D-SACK, and must reside below snd_una completely */
982 if (after(end_seq, tp->snd_una))
985 if (!before(start_seq, tp->undo_marker))
989 if (!after(end_seq, tp->undo_marker))
992 /* Undo_marker boundary crossing (overestimates a lot). Known already:
993 * start_seq < undo_marker and end_seq >= undo_marker.
995 return !before(start_seq, end_seq - tp->max_window);
998 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
999 * Event "B". Later note: FACK people cheated me again 8), we have to account
1000 * for reordering! Ugly, but should help.
1002 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1003 * less than what is now known to be received by the other end (derived from
1004 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1005 * retransmitted skbs to avoid some costly processing per ACKs.
1007 static void tcp_mark_lost_retrans(struct sock *sk)
1009 const struct inet_connection_sock *icsk = inet_csk(sk);
1010 struct tcp_sock *tp = tcp_sk(sk);
1011 struct sk_buff *skb;
1013 u32 new_low_seq = tp->snd_nxt;
1014 u32 received_upto = tcp_highest_sack_seq(tp);
1016 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1017 !after(received_upto, tp->lost_retrans_low) ||
1018 icsk->icsk_ca_state != TCP_CA_Recovery)
1021 tcp_for_write_queue(skb, sk) {
1022 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1024 if (skb == tcp_send_head(sk))
1026 if (cnt == tp->retrans_out)
1028 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1031 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1034 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1035 * constraint here (see above) but figuring out that at
1036 * least tp->reordering SACK blocks reside between ack_seq
1037 * and received_upto is not easy task to do cheaply with
1038 * the available datastructures.
1040 * Whether FACK should check here for tp->reordering segs
1041 * in-between one could argue for either way (it would be
1042 * rather simple to implement as we could count fack_count
1043 * during the walk and do tp->fackets_out - fack_count).
1045 if (after(received_upto, ack_seq)) {
1046 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1047 tp->retrans_out -= tcp_skb_pcount(skb);
1049 tcp_skb_mark_lost_uncond_verify(tp, skb);
1050 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1052 if (before(ack_seq, new_low_seq))
1053 new_low_seq = ack_seq;
1054 cnt += tcp_skb_pcount(skb);
1058 if (tp->retrans_out)
1059 tp->lost_retrans_low = new_low_seq;
1062 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1063 struct tcp_sack_block_wire *sp, int num_sacks,
1066 struct tcp_sock *tp = tcp_sk(sk);
1067 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1068 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1069 bool dup_sack = false;
1071 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1074 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1075 } else if (num_sacks > 1) {
1076 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1077 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1079 if (!after(end_seq_0, end_seq_1) &&
1080 !before(start_seq_0, start_seq_1)) {
1083 NET_INC_STATS_BH(sock_net(sk),
1084 LINUX_MIB_TCPDSACKOFORECV);
1088 /* D-SACK for already forgotten data... Do dumb counting. */
1089 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1090 !after(end_seq_0, prior_snd_una) &&
1091 after(end_seq_0, tp->undo_marker))
1097 struct tcp_sacktag_state {
1101 s32 rtt; /* RTT measured by SACKing never-retransmitted data */
1104 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1105 * the incoming SACK may not exactly match but we can find smaller MSS
1106 * aligned portion of it that matches. Therefore we might need to fragment
1107 * which may fail and creates some hassle (caller must handle error case
1110 * FIXME: this could be merged to shift decision code
1112 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1113 u32 start_seq, u32 end_seq)
1117 unsigned int pkt_len;
1120 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1121 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1123 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1124 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1125 mss = tcp_skb_mss(skb);
1126 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1129 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1133 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1138 /* Round if necessary so that SACKs cover only full MSSes
1139 * and/or the remaining small portion (if present)
1141 if (pkt_len > mss) {
1142 unsigned int new_len = (pkt_len / mss) * mss;
1143 if (!in_sack && new_len < pkt_len) {
1145 if (new_len > skb->len)
1150 err = tcp_fragment(sk, skb, pkt_len, mss);
1158 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1159 static u8 tcp_sacktag_one(struct sock *sk,
1160 struct tcp_sacktag_state *state, u8 sacked,
1161 u32 start_seq, u32 end_seq,
1162 int dup_sack, int pcount, u32 xmit_time)
1164 struct tcp_sock *tp = tcp_sk(sk);
1165 int fack_count = state->fack_count;
1167 /* Account D-SACK for retransmitted packet. */
1168 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1169 if (tp->undo_marker && tp->undo_retrans &&
1170 after(end_seq, tp->undo_marker))
1172 if (sacked & TCPCB_SACKED_ACKED)
1173 state->reord = min(fack_count, state->reord);
1176 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1177 if (!after(end_seq, tp->snd_una))
1180 if (!(sacked & TCPCB_SACKED_ACKED)) {
1181 if (sacked & TCPCB_SACKED_RETRANS) {
1182 /* If the segment is not tagged as lost,
1183 * we do not clear RETRANS, believing
1184 * that retransmission is still in flight.
1186 if (sacked & TCPCB_LOST) {
1187 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1188 tp->lost_out -= pcount;
1189 tp->retrans_out -= pcount;
1192 if (!(sacked & TCPCB_RETRANS)) {
1193 /* New sack for not retransmitted frame,
1194 * which was in hole. It is reordering.
1196 if (before(start_seq,
1197 tcp_highest_sack_seq(tp)))
1198 state->reord = min(fack_count,
1200 if (!after(end_seq, tp->high_seq))
1201 state->flag |= FLAG_ORIG_SACK_ACKED;
1202 /* Pick the earliest sequence sacked for RTT */
1204 state->rtt = tcp_time_stamp - xmit_time;
1207 if (sacked & TCPCB_LOST) {
1208 sacked &= ~TCPCB_LOST;
1209 tp->lost_out -= pcount;
1213 sacked |= TCPCB_SACKED_ACKED;
1214 state->flag |= FLAG_DATA_SACKED;
1215 tp->sacked_out += pcount;
1217 fack_count += pcount;
1219 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1220 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1221 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1222 tp->lost_cnt_hint += pcount;
1224 if (fack_count > tp->fackets_out)
1225 tp->fackets_out = fack_count;
1228 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1229 * frames and clear it. undo_retrans is decreased above, L|R frames
1230 * are accounted above as well.
1232 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1233 sacked &= ~TCPCB_SACKED_RETRANS;
1234 tp->retrans_out -= pcount;
1240 /* Shift newly-SACKed bytes from this skb to the immediately previous
1241 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1243 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1244 struct tcp_sacktag_state *state,
1245 unsigned int pcount, int shifted, int mss,
1248 struct tcp_sock *tp = tcp_sk(sk);
1249 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1250 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1251 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1255 /* Adjust counters and hints for the newly sacked sequence
1256 * range but discard the return value since prev is already
1257 * marked. We must tag the range first because the seq
1258 * advancement below implicitly advances
1259 * tcp_highest_sack_seq() when skb is highest_sack.
1261 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1262 start_seq, end_seq, dup_sack, pcount,
1263 TCP_SKB_CB(skb)->when);
1265 if (skb == tp->lost_skb_hint)
1266 tp->lost_cnt_hint += pcount;
1268 TCP_SKB_CB(prev)->end_seq += shifted;
1269 TCP_SKB_CB(skb)->seq += shifted;
1271 skb_shinfo(prev)->gso_segs += pcount;
1272 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1273 skb_shinfo(skb)->gso_segs -= pcount;
1275 /* When we're adding to gso_segs == 1, gso_size will be zero,
1276 * in theory this shouldn't be necessary but as long as DSACK
1277 * code can come after this skb later on it's better to keep
1278 * setting gso_size to something.
1280 if (!skb_shinfo(prev)->gso_size) {
1281 skb_shinfo(prev)->gso_size = mss;
1282 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1285 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1286 if (skb_shinfo(skb)->gso_segs <= 1) {
1287 skb_shinfo(skb)->gso_size = 0;
1288 skb_shinfo(skb)->gso_type = 0;
1291 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1292 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1295 BUG_ON(!tcp_skb_pcount(skb));
1296 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1300 /* Whole SKB was eaten :-) */
1302 if (skb == tp->retransmit_skb_hint)
1303 tp->retransmit_skb_hint = prev;
1304 if (skb == tp->lost_skb_hint) {
1305 tp->lost_skb_hint = prev;
1306 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1309 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1310 if (skb == tcp_highest_sack(sk))
1311 tcp_advance_highest_sack(sk, skb);
1313 tcp_unlink_write_queue(skb, sk);
1314 sk_wmem_free_skb(sk, skb);
1316 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1321 /* I wish gso_size would have a bit more sane initialization than
1322 * something-or-zero which complicates things
1324 static int tcp_skb_seglen(const struct sk_buff *skb)
1326 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1329 /* Shifting pages past head area doesn't work */
1330 static int skb_can_shift(const struct sk_buff *skb)
1332 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1335 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1338 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1339 struct tcp_sacktag_state *state,
1340 u32 start_seq, u32 end_seq,
1343 struct tcp_sock *tp = tcp_sk(sk);
1344 struct sk_buff *prev;
1350 if (!sk_can_gso(sk))
1353 /* Normally R but no L won't result in plain S */
1355 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1357 if (!skb_can_shift(skb))
1359 /* This frame is about to be dropped (was ACKed). */
1360 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1363 /* Can only happen with delayed DSACK + discard craziness */
1364 if (unlikely(skb == tcp_write_queue_head(sk)))
1366 prev = tcp_write_queue_prev(sk, skb);
1368 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1371 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1372 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1376 pcount = tcp_skb_pcount(skb);
1377 mss = tcp_skb_seglen(skb);
1379 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1380 * drop this restriction as unnecessary
1382 if (mss != tcp_skb_seglen(prev))
1385 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1387 /* CHECKME: This is non-MSS split case only?, this will
1388 * cause skipped skbs due to advancing loop btw, original
1389 * has that feature too
1391 if (tcp_skb_pcount(skb) <= 1)
1394 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1396 /* TODO: head merge to next could be attempted here
1397 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1398 * though it might not be worth of the additional hassle
1400 * ...we can probably just fallback to what was done
1401 * previously. We could try merging non-SACKed ones
1402 * as well but it probably isn't going to buy off
1403 * because later SACKs might again split them, and
1404 * it would make skb timestamp tracking considerably
1410 len = end_seq - TCP_SKB_CB(skb)->seq;
1412 BUG_ON(len > skb->len);
1414 /* MSS boundaries should be honoured or else pcount will
1415 * severely break even though it makes things bit trickier.
1416 * Optimize common case to avoid most of the divides
1418 mss = tcp_skb_mss(skb);
1420 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1421 * drop this restriction as unnecessary
1423 if (mss != tcp_skb_seglen(prev))
1428 } else if (len < mss) {
1436 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1437 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1440 if (!skb_shift(prev, skb, len))
1442 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1445 /* Hole filled allows collapsing with the next as well, this is very
1446 * useful when hole on every nth skb pattern happens
1448 if (prev == tcp_write_queue_tail(sk))
1450 skb = tcp_write_queue_next(sk, prev);
1452 if (!skb_can_shift(skb) ||
1453 (skb == tcp_send_head(sk)) ||
1454 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1455 (mss != tcp_skb_seglen(skb)))
1459 if (skb_shift(prev, skb, len)) {
1460 pcount += tcp_skb_pcount(skb);
1461 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1465 state->fack_count += pcount;
1472 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1476 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1477 struct tcp_sack_block *next_dup,
1478 struct tcp_sacktag_state *state,
1479 u32 start_seq, u32 end_seq,
1482 struct tcp_sock *tp = tcp_sk(sk);
1483 struct sk_buff *tmp;
1485 tcp_for_write_queue_from(skb, sk) {
1487 bool dup_sack = dup_sack_in;
1489 if (skb == tcp_send_head(sk))
1492 /* queue is in-order => we can short-circuit the walk early */
1493 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1496 if ((next_dup != NULL) &&
1497 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1498 in_sack = tcp_match_skb_to_sack(sk, skb,
1499 next_dup->start_seq,
1505 /* skb reference here is a bit tricky to get right, since
1506 * shifting can eat and free both this skb and the next,
1507 * so not even _safe variant of the loop is enough.
1510 tmp = tcp_shift_skb_data(sk, skb, state,
1511 start_seq, end_seq, dup_sack);
1520 in_sack = tcp_match_skb_to_sack(sk, skb,
1526 if (unlikely(in_sack < 0))
1530 TCP_SKB_CB(skb)->sacked =
1533 TCP_SKB_CB(skb)->sacked,
1534 TCP_SKB_CB(skb)->seq,
1535 TCP_SKB_CB(skb)->end_seq,
1537 tcp_skb_pcount(skb),
1538 TCP_SKB_CB(skb)->when);
1540 if (!before(TCP_SKB_CB(skb)->seq,
1541 tcp_highest_sack_seq(tp)))
1542 tcp_advance_highest_sack(sk, skb);
1545 state->fack_count += tcp_skb_pcount(skb);
1550 /* Avoid all extra work that is being done by sacktag while walking in
1553 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1554 struct tcp_sacktag_state *state,
1557 tcp_for_write_queue_from(skb, sk) {
1558 if (skb == tcp_send_head(sk))
1561 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1564 state->fack_count += tcp_skb_pcount(skb);
1569 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1571 struct tcp_sack_block *next_dup,
1572 struct tcp_sacktag_state *state,
1575 if (next_dup == NULL)
1578 if (before(next_dup->start_seq, skip_to_seq)) {
1579 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1580 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1581 next_dup->start_seq, next_dup->end_seq,
1588 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1590 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1594 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1595 u32 prior_snd_una, s32 *sack_rtt)
1597 struct tcp_sock *tp = tcp_sk(sk);
1598 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1599 TCP_SKB_CB(ack_skb)->sacked);
1600 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1601 struct tcp_sack_block sp[TCP_NUM_SACKS];
1602 struct tcp_sack_block *cache;
1603 struct tcp_sacktag_state state;
1604 struct sk_buff *skb;
1605 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1607 bool found_dup_sack = false;
1609 int first_sack_index;
1612 state.reord = tp->packets_out;
1615 if (!tp->sacked_out) {
1616 if (WARN_ON(tp->fackets_out))
1617 tp->fackets_out = 0;
1618 tcp_highest_sack_reset(sk);
1621 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1622 num_sacks, prior_snd_una);
1624 state.flag |= FLAG_DSACKING_ACK;
1626 /* Eliminate too old ACKs, but take into
1627 * account more or less fresh ones, they can
1628 * contain valid SACK info.
1630 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1633 if (!tp->packets_out)
1637 first_sack_index = 0;
1638 for (i = 0; i < num_sacks; i++) {
1639 bool dup_sack = !i && found_dup_sack;
1641 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1642 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1644 if (!tcp_is_sackblock_valid(tp, dup_sack,
1645 sp[used_sacks].start_seq,
1646 sp[used_sacks].end_seq)) {
1650 if (!tp->undo_marker)
1651 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1653 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1655 /* Don't count olds caused by ACK reordering */
1656 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1657 !after(sp[used_sacks].end_seq, tp->snd_una))
1659 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1662 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1664 first_sack_index = -1;
1668 /* Ignore very old stuff early */
1669 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1675 /* order SACK blocks to allow in order walk of the retrans queue */
1676 for (i = used_sacks - 1; i > 0; i--) {
1677 for (j = 0; j < i; j++) {
1678 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1679 swap(sp[j], sp[j + 1]);
1681 /* Track where the first SACK block goes to */
1682 if (j == first_sack_index)
1683 first_sack_index = j + 1;
1688 skb = tcp_write_queue_head(sk);
1689 state.fack_count = 0;
1692 if (!tp->sacked_out) {
1693 /* It's already past, so skip checking against it */
1694 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1696 cache = tp->recv_sack_cache;
1697 /* Skip empty blocks in at head of the cache */
1698 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1703 while (i < used_sacks) {
1704 u32 start_seq = sp[i].start_seq;
1705 u32 end_seq = sp[i].end_seq;
1706 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1707 struct tcp_sack_block *next_dup = NULL;
1709 if (found_dup_sack && ((i + 1) == first_sack_index))
1710 next_dup = &sp[i + 1];
1712 /* Skip too early cached blocks */
1713 while (tcp_sack_cache_ok(tp, cache) &&
1714 !before(start_seq, cache->end_seq))
1717 /* Can skip some work by looking recv_sack_cache? */
1718 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1719 after(end_seq, cache->start_seq)) {
1722 if (before(start_seq, cache->start_seq)) {
1723 skb = tcp_sacktag_skip(skb, sk, &state,
1725 skb = tcp_sacktag_walk(skb, sk, next_dup,
1732 /* Rest of the block already fully processed? */
1733 if (!after(end_seq, cache->end_seq))
1736 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1740 /* ...tail remains todo... */
1741 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1742 /* ...but better entrypoint exists! */
1743 skb = tcp_highest_sack(sk);
1746 state.fack_count = tp->fackets_out;
1751 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1752 /* Check overlap against next cached too (past this one already) */
1757 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1758 skb = tcp_highest_sack(sk);
1761 state.fack_count = tp->fackets_out;
1763 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1766 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1767 start_seq, end_seq, dup_sack);
1773 /* Clear the head of the cache sack blocks so we can skip it next time */
1774 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1775 tp->recv_sack_cache[i].start_seq = 0;
1776 tp->recv_sack_cache[i].end_seq = 0;
1778 for (j = 0; j < used_sacks; j++)
1779 tp->recv_sack_cache[i++] = sp[j];
1781 tcp_mark_lost_retrans(sk);
1783 tcp_verify_left_out(tp);
1785 if ((state.reord < tp->fackets_out) &&
1786 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1787 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1791 #if FASTRETRANS_DEBUG > 0
1792 WARN_ON((int)tp->sacked_out < 0);
1793 WARN_ON((int)tp->lost_out < 0);
1794 WARN_ON((int)tp->retrans_out < 0);
1795 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1797 *sack_rtt = state.rtt;
1801 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1802 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1804 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1808 holes = max(tp->lost_out, 1U);
1809 holes = min(holes, tp->packets_out);
1811 if ((tp->sacked_out + holes) > tp->packets_out) {
1812 tp->sacked_out = tp->packets_out - holes;
1818 /* If we receive more dupacks than we expected counting segments
1819 * in assumption of absent reordering, interpret this as reordering.
1820 * The only another reason could be bug in receiver TCP.
1822 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1824 struct tcp_sock *tp = tcp_sk(sk);
1825 if (tcp_limit_reno_sacked(tp))
1826 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1829 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1831 static void tcp_add_reno_sack(struct sock *sk)
1833 struct tcp_sock *tp = tcp_sk(sk);
1835 tcp_check_reno_reordering(sk, 0);
1836 tcp_verify_left_out(tp);
1839 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1841 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1843 struct tcp_sock *tp = tcp_sk(sk);
1846 /* One ACK acked hole. The rest eat duplicate ACKs. */
1847 if (acked - 1 >= tp->sacked_out)
1850 tp->sacked_out -= acked - 1;
1852 tcp_check_reno_reordering(sk, acked);
1853 tcp_verify_left_out(tp);
1856 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1861 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1863 tp->retrans_out = 0;
1866 tp->undo_marker = 0;
1867 tp->undo_retrans = 0;
1870 void tcp_clear_retrans(struct tcp_sock *tp)
1872 tcp_clear_retrans_partial(tp);
1874 tp->fackets_out = 0;
1878 /* Enter Loss state. If "how" is not zero, forget all SACK information
1879 * and reset tags completely, otherwise preserve SACKs. If receiver
1880 * dropped its ofo queue, we will know this due to reneging detection.
1882 void tcp_enter_loss(struct sock *sk, int how)
1884 const struct inet_connection_sock *icsk = inet_csk(sk);
1885 struct tcp_sock *tp = tcp_sk(sk);
1886 struct sk_buff *skb;
1887 bool new_recovery = false;
1889 /* Reduce ssthresh if it has not yet been made inside this window. */
1890 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1891 !after(tp->high_seq, tp->snd_una) ||
1892 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1893 new_recovery = true;
1894 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1895 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1896 tcp_ca_event(sk, CA_EVENT_LOSS);
1899 tp->snd_cwnd_cnt = 0;
1900 tp->snd_cwnd_stamp = tcp_time_stamp;
1902 tcp_clear_retrans_partial(tp);
1904 if (tcp_is_reno(tp))
1905 tcp_reset_reno_sack(tp);
1907 tp->undo_marker = tp->snd_una;
1910 tp->fackets_out = 0;
1912 tcp_clear_all_retrans_hints(tp);
1914 tcp_for_write_queue(skb, sk) {
1915 if (skb == tcp_send_head(sk))
1918 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1919 tp->undo_marker = 0;
1920 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1921 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1922 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1923 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1924 tp->lost_out += tcp_skb_pcount(skb);
1925 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1928 tcp_verify_left_out(tp);
1930 /* Timeout in disordered state after receiving substantial DUPACKs
1931 * suggests that the degree of reordering is over-estimated.
1933 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1934 tp->sacked_out >= sysctl_tcp_reordering)
1935 tp->reordering = min_t(unsigned int, tp->reordering,
1936 sysctl_tcp_reordering);
1937 tcp_set_ca_state(sk, TCP_CA_Loss);
1938 tp->high_seq = tp->snd_nxt;
1939 TCP_ECN_queue_cwr(tp);
1941 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1942 * loss recovery is underway except recurring timeout(s) on
1943 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1945 tp->frto = sysctl_tcp_frto &&
1946 (new_recovery || icsk->icsk_retransmits) &&
1947 !inet_csk(sk)->icsk_mtup.probe_size;
1950 /* If ACK arrived pointing to a remembered SACK, it means that our
1951 * remembered SACKs do not reflect real state of receiver i.e.
1952 * receiver _host_ is heavily congested (or buggy).
1954 * Do processing similar to RTO timeout.
1956 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1958 if (flag & FLAG_SACK_RENEGING) {
1959 struct inet_connection_sock *icsk = inet_csk(sk);
1960 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1962 tcp_enter_loss(sk, 1);
1963 icsk->icsk_retransmits++;
1964 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1965 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1966 icsk->icsk_rto, TCP_RTO_MAX);
1972 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1974 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1977 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1978 * counter when SACK is enabled (without SACK, sacked_out is used for
1981 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1982 * segments up to the highest received SACK block so far and holes in
1985 * With reordering, holes may still be in flight, so RFC3517 recovery
1986 * uses pure sacked_out (total number of SACKed segments) even though
1987 * it violates the RFC that uses duplicate ACKs, often these are equal
1988 * but when e.g. out-of-window ACKs or packet duplication occurs,
1989 * they differ. Since neither occurs due to loss, TCP should really
1992 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
1994 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1997 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
1999 struct tcp_sock *tp = tcp_sk(sk);
2000 unsigned long delay;
2002 /* Delay early retransmit and entering fast recovery for
2003 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2004 * available, or RTO is scheduled to fire first.
2006 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2007 (flag & FLAG_ECE) || !tp->srtt)
2010 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
2011 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2014 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2019 /* Linux NewReno/SACK/FACK/ECN state machine.
2020 * --------------------------------------
2022 * "Open" Normal state, no dubious events, fast path.
2023 * "Disorder" In all the respects it is "Open",
2024 * but requires a bit more attention. It is entered when
2025 * we see some SACKs or dupacks. It is split of "Open"
2026 * mainly to move some processing from fast path to slow one.
2027 * "CWR" CWND was reduced due to some Congestion Notification event.
2028 * It can be ECN, ICMP source quench, local device congestion.
2029 * "Recovery" CWND was reduced, we are fast-retransmitting.
2030 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2032 * tcp_fastretrans_alert() is entered:
2033 * - each incoming ACK, if state is not "Open"
2034 * - when arrived ACK is unusual, namely:
2039 * Counting packets in flight is pretty simple.
2041 * in_flight = packets_out - left_out + retrans_out
2043 * packets_out is SND.NXT-SND.UNA counted in packets.
2045 * retrans_out is number of retransmitted segments.
2047 * left_out is number of segments left network, but not ACKed yet.
2049 * left_out = sacked_out + lost_out
2051 * sacked_out: Packets, which arrived to receiver out of order
2052 * and hence not ACKed. With SACKs this number is simply
2053 * amount of SACKed data. Even without SACKs
2054 * it is easy to give pretty reliable estimate of this number,
2055 * counting duplicate ACKs.
2057 * lost_out: Packets lost by network. TCP has no explicit
2058 * "loss notification" feedback from network (for now).
2059 * It means that this number can be only _guessed_.
2060 * Actually, it is the heuristics to predict lossage that
2061 * distinguishes different algorithms.
2063 * F.e. after RTO, when all the queue is considered as lost,
2064 * lost_out = packets_out and in_flight = retrans_out.
2066 * Essentially, we have now two algorithms counting
2069 * FACK: It is the simplest heuristics. As soon as we decided
2070 * that something is lost, we decide that _all_ not SACKed
2071 * packets until the most forward SACK are lost. I.e.
2072 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2073 * It is absolutely correct estimate, if network does not reorder
2074 * packets. And it loses any connection to reality when reordering
2075 * takes place. We use FACK by default until reordering
2076 * is suspected on the path to this destination.
2078 * NewReno: when Recovery is entered, we assume that one segment
2079 * is lost (classic Reno). While we are in Recovery and
2080 * a partial ACK arrives, we assume that one more packet
2081 * is lost (NewReno). This heuristics are the same in NewReno
2084 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2085 * deflation etc. CWND is real congestion window, never inflated, changes
2086 * only according to classic VJ rules.
2088 * Really tricky (and requiring careful tuning) part of algorithm
2089 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2090 * The first determines the moment _when_ we should reduce CWND and,
2091 * hence, slow down forward transmission. In fact, it determines the moment
2092 * when we decide that hole is caused by loss, rather than by a reorder.
2094 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2095 * holes, caused by lost packets.
2097 * And the most logically complicated part of algorithm is undo
2098 * heuristics. We detect false retransmits due to both too early
2099 * fast retransmit (reordering) and underestimated RTO, analyzing
2100 * timestamps and D-SACKs. When we detect that some segments were
2101 * retransmitted by mistake and CWND reduction was wrong, we undo
2102 * window reduction and abort recovery phase. This logic is hidden
2103 * inside several functions named tcp_try_undo_<something>.
2106 /* This function decides, when we should leave Disordered state
2107 * and enter Recovery phase, reducing congestion window.
2109 * Main question: may we further continue forward transmission
2110 * with the same cwnd?
2112 static bool tcp_time_to_recover(struct sock *sk, int flag)
2114 struct tcp_sock *tp = tcp_sk(sk);
2117 /* Trick#1: The loss is proven. */
2121 /* Not-A-Trick#2 : Classic rule... */
2122 if (tcp_dupack_heuristics(tp) > tp->reordering)
2125 /* Trick#4: It is still not OK... But will it be useful to delay
2128 packets_out = tp->packets_out;
2129 if (packets_out <= tp->reordering &&
2130 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2131 !tcp_may_send_now(sk)) {
2132 /* We have nothing to send. This connection is limited
2133 * either by receiver window or by application.
2138 /* If a thin stream is detected, retransmit after first
2139 * received dupack. Employ only if SACK is supported in order
2140 * to avoid possible corner-case series of spurious retransmissions
2141 * Use only if there are no unsent data.
2143 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2144 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2145 tcp_is_sack(tp) && !tcp_send_head(sk))
2148 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2149 * retransmissions due to small network reorderings, we implement
2150 * Mitigation A.3 in the RFC and delay the retransmission for a short
2151 * interval if appropriate.
2153 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2154 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2155 !tcp_may_send_now(sk))
2156 return !tcp_pause_early_retransmit(sk, flag);
2161 /* Detect loss in event "A" above by marking head of queue up as lost.
2162 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2163 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2164 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2165 * the maximum SACKed segments to pass before reaching this limit.
2167 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2169 struct tcp_sock *tp = tcp_sk(sk);
2170 struct sk_buff *skb;
2174 /* Use SACK to deduce losses of new sequences sent during recovery */
2175 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2177 WARN_ON(packets > tp->packets_out);
2178 if (tp->lost_skb_hint) {
2179 skb = tp->lost_skb_hint;
2180 cnt = tp->lost_cnt_hint;
2181 /* Head already handled? */
2182 if (mark_head && skb != tcp_write_queue_head(sk))
2185 skb = tcp_write_queue_head(sk);
2189 tcp_for_write_queue_from(skb, sk) {
2190 if (skb == tcp_send_head(sk))
2192 /* TODO: do this better */
2193 /* this is not the most efficient way to do this... */
2194 tp->lost_skb_hint = skb;
2195 tp->lost_cnt_hint = cnt;
2197 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2201 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2202 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2203 cnt += tcp_skb_pcount(skb);
2205 if (cnt > packets) {
2206 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2207 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2208 (oldcnt >= packets))
2211 mss = skb_shinfo(skb)->gso_size;
2212 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2218 tcp_skb_mark_lost(tp, skb);
2223 tcp_verify_left_out(tp);
2226 /* Account newly detected lost packet(s) */
2228 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2230 struct tcp_sock *tp = tcp_sk(sk);
2232 if (tcp_is_reno(tp)) {
2233 tcp_mark_head_lost(sk, 1, 1);
2234 } else if (tcp_is_fack(tp)) {
2235 int lost = tp->fackets_out - tp->reordering;
2238 tcp_mark_head_lost(sk, lost, 0);
2240 int sacked_upto = tp->sacked_out - tp->reordering;
2241 if (sacked_upto >= 0)
2242 tcp_mark_head_lost(sk, sacked_upto, 0);
2243 else if (fast_rexmit)
2244 tcp_mark_head_lost(sk, 1, 1);
2248 /* CWND moderation, preventing bursts due to too big ACKs
2249 * in dubious situations.
2251 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2253 tp->snd_cwnd = min(tp->snd_cwnd,
2254 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2255 tp->snd_cwnd_stamp = tcp_time_stamp;
2258 /* Nothing was retransmitted or returned timestamp is less
2259 * than timestamp of the first retransmission.
2261 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2263 return !tp->retrans_stamp ||
2264 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2265 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2268 /* Undo procedures. */
2270 #if FASTRETRANS_DEBUG > 1
2271 static void DBGUNDO(struct sock *sk, const char *msg)
2273 struct tcp_sock *tp = tcp_sk(sk);
2274 struct inet_sock *inet = inet_sk(sk);
2276 if (sk->sk_family == AF_INET) {
2277 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2279 &inet->inet_daddr, ntohs(inet->inet_dport),
2280 tp->snd_cwnd, tcp_left_out(tp),
2281 tp->snd_ssthresh, tp->prior_ssthresh,
2284 #if IS_ENABLED(CONFIG_IPV6)
2285 else if (sk->sk_family == AF_INET6) {
2286 struct ipv6_pinfo *np = inet6_sk(sk);
2287 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2289 &np->daddr, ntohs(inet->inet_dport),
2290 tp->snd_cwnd, tcp_left_out(tp),
2291 tp->snd_ssthresh, tp->prior_ssthresh,
2297 #define DBGUNDO(x...) do { } while (0)
2300 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2302 struct tcp_sock *tp = tcp_sk(sk);
2305 struct sk_buff *skb;
2307 tcp_for_write_queue(skb, sk) {
2308 if (skb == tcp_send_head(sk))
2310 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2313 tcp_clear_all_retrans_hints(tp);
2316 if (tp->prior_ssthresh) {
2317 const struct inet_connection_sock *icsk = inet_csk(sk);
2319 if (icsk->icsk_ca_ops->undo_cwnd)
2320 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2322 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2324 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2325 tp->snd_ssthresh = tp->prior_ssthresh;
2326 TCP_ECN_withdraw_cwr(tp);
2329 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2331 tp->snd_cwnd_stamp = tcp_time_stamp;
2332 tp->undo_marker = 0;
2335 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2337 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2340 /* People celebrate: "We love our President!" */
2341 static bool tcp_try_undo_recovery(struct sock *sk)
2343 struct tcp_sock *tp = tcp_sk(sk);
2345 if (tcp_may_undo(tp)) {
2348 /* Happy end! We did not retransmit anything
2349 * or our original transmission succeeded.
2351 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2352 tcp_undo_cwnd_reduction(sk, false);
2353 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2354 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2356 mib_idx = LINUX_MIB_TCPFULLUNDO;
2358 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2360 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2361 /* Hold old state until something *above* high_seq
2362 * is ACKed. For Reno it is MUST to prevent false
2363 * fast retransmits (RFC2582). SACK TCP is safe. */
2364 tcp_moderate_cwnd(tp);
2367 tcp_set_ca_state(sk, TCP_CA_Open);
2371 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2372 static bool tcp_try_undo_dsack(struct sock *sk)
2374 struct tcp_sock *tp = tcp_sk(sk);
2376 if (tp->undo_marker && !tp->undo_retrans) {
2377 DBGUNDO(sk, "D-SACK");
2378 tcp_undo_cwnd_reduction(sk, false);
2379 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2385 /* We can clear retrans_stamp when there are no retransmissions in the
2386 * window. It would seem that it is trivially available for us in
2387 * tp->retrans_out, however, that kind of assumptions doesn't consider
2388 * what will happen if errors occur when sending retransmission for the
2389 * second time. ...It could the that such segment has only
2390 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2391 * the head skb is enough except for some reneging corner cases that
2392 * are not worth the effort.
2394 * Main reason for all this complexity is the fact that connection dying
2395 * time now depends on the validity of the retrans_stamp, in particular,
2396 * that successive retransmissions of a segment must not advance
2397 * retrans_stamp under any conditions.
2399 static bool tcp_any_retrans_done(const struct sock *sk)
2401 const struct tcp_sock *tp = tcp_sk(sk);
2402 struct sk_buff *skb;
2404 if (tp->retrans_out)
2407 skb = tcp_write_queue_head(sk);
2408 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2414 /* Undo during loss recovery after partial ACK or using F-RTO. */
2415 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2417 struct tcp_sock *tp = tcp_sk(sk);
2419 if (frto_undo || tcp_may_undo(tp)) {
2420 tcp_undo_cwnd_reduction(sk, true);
2422 DBGUNDO(sk, "partial loss");
2423 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2425 NET_INC_STATS_BH(sock_net(sk),
2426 LINUX_MIB_TCPSPURIOUSRTOS);
2427 inet_csk(sk)->icsk_retransmits = 0;
2428 if (frto_undo || tcp_is_sack(tp))
2429 tcp_set_ca_state(sk, TCP_CA_Open);
2435 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2436 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2437 * It computes the number of packets to send (sndcnt) based on packets newly
2439 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2440 * cwnd reductions across a full RTT.
2441 * 2) If packets in flight is lower than ssthresh (such as due to excess
2442 * losses and/or application stalls), do not perform any further cwnd
2443 * reductions, but instead slow start up to ssthresh.
2445 static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2447 struct tcp_sock *tp = tcp_sk(sk);
2449 tp->high_seq = tp->snd_nxt;
2450 tp->tlp_high_seq = 0;
2451 tp->snd_cwnd_cnt = 0;
2452 tp->prior_cwnd = tp->snd_cwnd;
2453 tp->prr_delivered = 0;
2456 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2457 TCP_ECN_queue_cwr(tp);
2460 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2463 struct tcp_sock *tp = tcp_sk(sk);
2465 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2466 int newly_acked_sacked = prior_unsacked -
2467 (tp->packets_out - tp->sacked_out);
2469 tp->prr_delivered += newly_acked_sacked;
2470 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2471 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2473 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2475 sndcnt = min_t(int, delta,
2476 max_t(int, tp->prr_delivered - tp->prr_out,
2477 newly_acked_sacked) + 1);
2480 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2481 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2484 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2486 struct tcp_sock *tp = tcp_sk(sk);
2488 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2489 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2490 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2491 tp->snd_cwnd = tp->snd_ssthresh;
2492 tp->snd_cwnd_stamp = tcp_time_stamp;
2494 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2497 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2498 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2500 struct tcp_sock *tp = tcp_sk(sk);
2502 tp->prior_ssthresh = 0;
2503 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2504 tp->undo_marker = 0;
2505 tcp_init_cwnd_reduction(sk, set_ssthresh);
2506 tcp_set_ca_state(sk, TCP_CA_CWR);
2510 static void tcp_try_keep_open(struct sock *sk)
2512 struct tcp_sock *tp = tcp_sk(sk);
2513 int state = TCP_CA_Open;
2515 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2516 state = TCP_CA_Disorder;
2518 if (inet_csk(sk)->icsk_ca_state != state) {
2519 tcp_set_ca_state(sk, state);
2520 tp->high_seq = tp->snd_nxt;
2524 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2526 struct tcp_sock *tp = tcp_sk(sk);
2528 tcp_verify_left_out(tp);
2530 if (!tcp_any_retrans_done(sk))
2531 tp->retrans_stamp = 0;
2533 if (flag & FLAG_ECE)
2534 tcp_enter_cwr(sk, 1);
2536 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2537 tcp_try_keep_open(sk);
2539 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2543 static void tcp_mtup_probe_failed(struct sock *sk)
2545 struct inet_connection_sock *icsk = inet_csk(sk);
2547 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2548 icsk->icsk_mtup.probe_size = 0;
2551 static void tcp_mtup_probe_success(struct sock *sk)
2553 struct tcp_sock *tp = tcp_sk(sk);
2554 struct inet_connection_sock *icsk = inet_csk(sk);
2556 /* FIXME: breaks with very large cwnd */
2557 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2558 tp->snd_cwnd = tp->snd_cwnd *
2559 tcp_mss_to_mtu(sk, tp->mss_cache) /
2560 icsk->icsk_mtup.probe_size;
2561 tp->snd_cwnd_cnt = 0;
2562 tp->snd_cwnd_stamp = tcp_time_stamp;
2563 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2565 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2566 icsk->icsk_mtup.probe_size = 0;
2567 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2570 /* Do a simple retransmit without using the backoff mechanisms in
2571 * tcp_timer. This is used for path mtu discovery.
2572 * The socket is already locked here.
2574 void tcp_simple_retransmit(struct sock *sk)
2576 const struct inet_connection_sock *icsk = inet_csk(sk);
2577 struct tcp_sock *tp = tcp_sk(sk);
2578 struct sk_buff *skb;
2579 unsigned int mss = tcp_current_mss(sk);
2580 u32 prior_lost = tp->lost_out;
2582 tcp_for_write_queue(skb, sk) {
2583 if (skb == tcp_send_head(sk))
2585 if (tcp_skb_seglen(skb) > mss &&
2586 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2587 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2588 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2589 tp->retrans_out -= tcp_skb_pcount(skb);
2591 tcp_skb_mark_lost_uncond_verify(tp, skb);
2595 tcp_clear_retrans_hints_partial(tp);
2597 if (prior_lost == tp->lost_out)
2600 if (tcp_is_reno(tp))
2601 tcp_limit_reno_sacked(tp);
2603 tcp_verify_left_out(tp);
2605 /* Don't muck with the congestion window here.
2606 * Reason is that we do not increase amount of _data_
2607 * in network, but units changed and effective
2608 * cwnd/ssthresh really reduced now.
2610 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2611 tp->high_seq = tp->snd_nxt;
2612 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2613 tp->prior_ssthresh = 0;
2614 tp->undo_marker = 0;
2615 tcp_set_ca_state(sk, TCP_CA_Loss);
2617 tcp_xmit_retransmit_queue(sk);
2619 EXPORT_SYMBOL(tcp_simple_retransmit);
2621 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2623 struct tcp_sock *tp = tcp_sk(sk);
2626 if (tcp_is_reno(tp))
2627 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2629 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2631 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2633 tp->prior_ssthresh = 0;
2634 tp->undo_marker = tp->snd_una;
2635 tp->undo_retrans = tp->retrans_out;
2637 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2639 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2640 tcp_init_cwnd_reduction(sk, true);
2642 tcp_set_ca_state(sk, TCP_CA_Recovery);
2645 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2646 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2648 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2650 struct inet_connection_sock *icsk = inet_csk(sk);
2651 struct tcp_sock *tp = tcp_sk(sk);
2652 bool recovered = !before(tp->snd_una, tp->high_seq);
2654 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2655 if (flag & FLAG_ORIG_SACK_ACKED) {
2656 /* Step 3.b. A timeout is spurious if not all data are
2657 * lost, i.e., never-retransmitted data are (s)acked.
2659 tcp_try_undo_loss(sk, true);
2662 if (after(tp->snd_nxt, tp->high_seq) &&
2663 (flag & FLAG_DATA_SACKED || is_dupack)) {
2664 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2665 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2666 tp->high_seq = tp->snd_nxt;
2667 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2669 if (after(tp->snd_nxt, tp->high_seq))
2670 return; /* Step 2.b */
2676 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2677 icsk->icsk_retransmits = 0;
2678 tcp_try_undo_recovery(sk);
2681 if (flag & FLAG_DATA_ACKED)
2682 icsk->icsk_retransmits = 0;
2683 if (tcp_is_reno(tp)) {
2684 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2685 * delivered. Lower inflight to clock out (re)tranmissions.
2687 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2688 tcp_add_reno_sack(sk);
2689 else if (flag & FLAG_SND_UNA_ADVANCED)
2690 tcp_reset_reno_sack(tp);
2692 if (tcp_try_undo_loss(sk, false))
2694 tcp_xmit_retransmit_queue(sk);
2697 /* Undo during fast recovery after partial ACK. */
2698 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2699 const int prior_unsacked)
2701 struct tcp_sock *tp = tcp_sk(sk);
2703 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2704 /* Plain luck! Hole if filled with delayed
2705 * packet, rather than with a retransmit.
2707 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2709 /* We are getting evidence that the reordering degree is higher
2710 * than we realized. If there are no retransmits out then we
2711 * can undo. Otherwise we clock out new packets but do not
2712 * mark more packets lost or retransmit more.
2714 if (tp->retrans_out) {
2715 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2719 if (!tcp_any_retrans_done(sk))
2720 tp->retrans_stamp = 0;
2722 DBGUNDO(sk, "partial recovery");
2723 tcp_undo_cwnd_reduction(sk, true);
2724 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2725 tcp_try_keep_open(sk);
2731 /* Process an event, which can update packets-in-flight not trivially.
2732 * Main goal of this function is to calculate new estimate for left_out,
2733 * taking into account both packets sitting in receiver's buffer and
2734 * packets lost by network.
2736 * Besides that it does CWND reduction, when packet loss is detected
2737 * and changes state of machine.
2739 * It does _not_ decide what to send, it is made in function
2740 * tcp_xmit_retransmit_queue().
2742 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2743 const int prior_unsacked,
2744 bool is_dupack, int flag)
2746 struct inet_connection_sock *icsk = inet_csk(sk);
2747 struct tcp_sock *tp = tcp_sk(sk);
2748 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2749 (tcp_fackets_out(tp) > tp->reordering));
2750 int fast_rexmit = 0;
2752 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2754 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2755 tp->fackets_out = 0;
2757 /* Now state machine starts.
2758 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2759 if (flag & FLAG_ECE)
2760 tp->prior_ssthresh = 0;
2762 /* B. In all the states check for reneging SACKs. */
2763 if (tcp_check_sack_reneging(sk, flag))
2766 /* C. Check consistency of the current state. */
2767 tcp_verify_left_out(tp);
2769 /* D. Check state exit conditions. State can be terminated
2770 * when high_seq is ACKed. */
2771 if (icsk->icsk_ca_state == TCP_CA_Open) {
2772 WARN_ON(tp->retrans_out != 0);
2773 tp->retrans_stamp = 0;
2774 } else if (!before(tp->snd_una, tp->high_seq)) {
2775 switch (icsk->icsk_ca_state) {
2777 /* CWR is to be held something *above* high_seq
2778 * is ACKed for CWR bit to reach receiver. */
2779 if (tp->snd_una != tp->high_seq) {
2780 tcp_end_cwnd_reduction(sk);
2781 tcp_set_ca_state(sk, TCP_CA_Open);
2785 case TCP_CA_Recovery:
2786 if (tcp_is_reno(tp))
2787 tcp_reset_reno_sack(tp);
2788 if (tcp_try_undo_recovery(sk))
2790 tcp_end_cwnd_reduction(sk);
2795 /* E. Process state. */
2796 switch (icsk->icsk_ca_state) {
2797 case TCP_CA_Recovery:
2798 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2799 if (tcp_is_reno(tp) && is_dupack)
2800 tcp_add_reno_sack(sk);
2802 if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2804 /* Partial ACK arrived. Force fast retransmit. */
2805 do_lost = tcp_is_reno(tp) ||
2806 tcp_fackets_out(tp) > tp->reordering;
2808 if (tcp_try_undo_dsack(sk)) {
2809 tcp_try_keep_open(sk);
2814 tcp_process_loss(sk, flag, is_dupack);
2815 if (icsk->icsk_ca_state != TCP_CA_Open)
2817 /* Fall through to processing in Open state. */
2819 if (tcp_is_reno(tp)) {
2820 if (flag & FLAG_SND_UNA_ADVANCED)
2821 tcp_reset_reno_sack(tp);
2823 tcp_add_reno_sack(sk);
2826 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2827 tcp_try_undo_dsack(sk);
2829 if (!tcp_time_to_recover(sk, flag)) {
2830 tcp_try_to_open(sk, flag, prior_unsacked);
2834 /* MTU probe failure: don't reduce cwnd */
2835 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2836 icsk->icsk_mtup.probe_size &&
2837 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2838 tcp_mtup_probe_failed(sk);
2839 /* Restores the reduction we did in tcp_mtup_probe() */
2841 tcp_simple_retransmit(sk);
2845 /* Otherwise enter Recovery state */
2846 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2851 tcp_update_scoreboard(sk, fast_rexmit);
2852 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2853 tcp_xmit_retransmit_queue(sk);
2856 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2857 s32 seq_rtt, s32 sack_rtt)
2859 const struct tcp_sock *tp = tcp_sk(sk);
2861 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2862 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2863 * Karn's algorithm forbids taking RTT if some retransmitted data
2864 * is acked (RFC6298).
2866 if (flag & FLAG_RETRANS_DATA_ACKED)
2872 /* RTTM Rule: A TSecr value received in a segment is used to
2873 * update the averaged RTT measurement only if the segment
2874 * acknowledges some new data, i.e., only if it advances the
2875 * left edge of the send window.
2876 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2878 if (seq_rtt < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2879 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2884 tcp_rtt_estimator(sk, seq_rtt);
2887 /* RFC6298: only reset backoff on valid RTT measurement. */
2888 inet_csk(sk)->icsk_backoff = 0;
2892 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2893 static void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2895 struct tcp_sock *tp = tcp_sk(sk);
2898 if (tp->lsndtime && !tp->total_retrans)
2899 seq_rtt = tcp_time_stamp - tp->lsndtime;
2900 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt, -1);
2903 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2905 const struct inet_connection_sock *icsk = inet_csk(sk);
2906 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2907 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2910 /* Restart timer after forward progress on connection.
2911 * RFC2988 recommends to restart timer to now+rto.
2913 void tcp_rearm_rto(struct sock *sk)
2915 const struct inet_connection_sock *icsk = inet_csk(sk);
2916 struct tcp_sock *tp = tcp_sk(sk);
2918 /* If the retrans timer is currently being used by Fast Open
2919 * for SYN-ACK retrans purpose, stay put.
2921 if (tp->fastopen_rsk)
2924 if (!tp->packets_out) {
2925 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2927 u32 rto = inet_csk(sk)->icsk_rto;
2928 /* Offset the time elapsed after installing regular RTO */
2929 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2930 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2931 struct sk_buff *skb = tcp_write_queue_head(sk);
2932 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
2933 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2934 /* delta may not be positive if the socket is locked
2935 * when the retrans timer fires and is rescheduled.
2940 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2945 /* This function is called when the delayed ER timer fires. TCP enters
2946 * fast recovery and performs fast-retransmit.
2948 void tcp_resume_early_retransmit(struct sock *sk)
2950 struct tcp_sock *tp = tcp_sk(sk);
2954 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2955 if (!tp->do_early_retrans)
2958 tcp_enter_recovery(sk, false);
2959 tcp_update_scoreboard(sk, 1);
2960 tcp_xmit_retransmit_queue(sk);
2963 /* If we get here, the whole TSO packet has not been acked. */
2964 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2966 struct tcp_sock *tp = tcp_sk(sk);
2969 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2971 packets_acked = tcp_skb_pcount(skb);
2972 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2974 packets_acked -= tcp_skb_pcount(skb);
2976 if (packets_acked) {
2977 BUG_ON(tcp_skb_pcount(skb) == 0);
2978 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2981 return packets_acked;
2984 /* Remove acknowledged frames from the retransmission queue. If our packet
2985 * is before the ack sequence we can discard it as it's confirmed to have
2986 * arrived at the other end.
2988 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
2989 u32 prior_snd_una, s32 sack_rtt)
2991 struct tcp_sock *tp = tcp_sk(sk);
2992 const struct inet_connection_sock *icsk = inet_csk(sk);
2993 struct sk_buff *skb;
2994 u32 now = tcp_time_stamp;
2995 int fully_acked = true;
2998 u32 reord = tp->packets_out;
2999 u32 prior_sacked = tp->sacked_out;
3001 s32 ca_seq_rtt = -1;
3002 ktime_t last_ackt = net_invalid_timestamp();
3004 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3005 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3007 u8 sacked = scb->sacked;
3009 /* Determine how many packets and what bytes were acked, tso and else */
3010 if (after(scb->end_seq, tp->snd_una)) {
3011 if (tcp_skb_pcount(skb) == 1 ||
3012 !after(tp->snd_una, scb->seq))
3015 acked_pcount = tcp_tso_acked(sk, skb);
3019 fully_acked = false;
3021 acked_pcount = tcp_skb_pcount(skb);
3024 if (sacked & TCPCB_RETRANS) {
3025 if (sacked & TCPCB_SACKED_RETRANS)
3026 tp->retrans_out -= acked_pcount;
3027 flag |= FLAG_RETRANS_DATA_ACKED;
3029 ca_seq_rtt = now - scb->when;
3030 last_ackt = skb->tstamp;
3032 seq_rtt = ca_seq_rtt;
3034 if (!(sacked & TCPCB_SACKED_ACKED))
3035 reord = min(pkts_acked, reord);
3036 if (!after(scb->end_seq, tp->high_seq))
3037 flag |= FLAG_ORIG_SACK_ACKED;
3040 if (sacked & TCPCB_SACKED_ACKED)
3041 tp->sacked_out -= acked_pcount;
3042 if (sacked & TCPCB_LOST)
3043 tp->lost_out -= acked_pcount;
3045 tp->packets_out -= acked_pcount;
3046 pkts_acked += acked_pcount;
3048 /* Initial outgoing SYN's get put onto the write_queue
3049 * just like anything else we transmit. It is not
3050 * true data, and if we misinform our callers that
3051 * this ACK acks real data, we will erroneously exit
3052 * connection startup slow start one packet too
3053 * quickly. This is severely frowned upon behavior.
3055 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3056 flag |= FLAG_DATA_ACKED;
3058 flag |= FLAG_SYN_ACKED;
3059 tp->retrans_stamp = 0;
3065 tcp_unlink_write_queue(skb, sk);
3066 sk_wmem_free_skb(sk, skb);
3067 if (skb == tp->retransmit_skb_hint)
3068 tp->retransmit_skb_hint = NULL;
3069 if (skb == tp->lost_skb_hint)
3070 tp->lost_skb_hint = NULL;
3073 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3074 tp->snd_up = tp->snd_una;
3076 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3077 flag |= FLAG_SACK_RENEGING;
3079 if (tcp_ack_update_rtt(sk, flag, seq_rtt, sack_rtt) ||
3080 (flag & FLAG_ACKED))
3083 if (flag & FLAG_ACKED) {
3084 const struct tcp_congestion_ops *ca_ops
3085 = inet_csk(sk)->icsk_ca_ops;
3087 if (unlikely(icsk->icsk_mtup.probe_size &&
3088 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3089 tcp_mtup_probe_success(sk);
3092 if (tcp_is_reno(tp)) {
3093 tcp_remove_reno_sacks(sk, pkts_acked);
3097 /* Non-retransmitted hole got filled? That's reordering */
3098 if (reord < prior_fackets)
3099 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3101 delta = tcp_is_fack(tp) ? pkts_acked :
3102 prior_sacked - tp->sacked_out;
3103 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3106 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3108 if (ca_ops->pkts_acked) {
3111 /* Is the ACK triggering packet unambiguous? */
3112 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3113 /* High resolution needed and available? */
3114 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3115 !ktime_equal(last_ackt,
3116 net_invalid_timestamp()))
3117 rtt_us = ktime_us_delta(ktime_get_real(),
3119 else if (ca_seq_rtt >= 0)
3120 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3123 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3127 #if FASTRETRANS_DEBUG > 0
3128 WARN_ON((int)tp->sacked_out < 0);
3129 WARN_ON((int)tp->lost_out < 0);
3130 WARN_ON((int)tp->retrans_out < 0);
3131 if (!tp->packets_out && tcp_is_sack(tp)) {
3132 icsk = inet_csk(sk);
3134 pr_debug("Leak l=%u %d\n",
3135 tp->lost_out, icsk->icsk_ca_state);
3138 if (tp->sacked_out) {
3139 pr_debug("Leak s=%u %d\n",
3140 tp->sacked_out, icsk->icsk_ca_state);
3143 if (tp->retrans_out) {
3144 pr_debug("Leak r=%u %d\n",
3145 tp->retrans_out, icsk->icsk_ca_state);
3146 tp->retrans_out = 0;
3153 static void tcp_ack_probe(struct sock *sk)
3155 const struct tcp_sock *tp = tcp_sk(sk);
3156 struct inet_connection_sock *icsk = inet_csk(sk);
3158 /* Was it a usable window open? */
3160 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3161 icsk->icsk_backoff = 0;
3162 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3163 /* Socket must be waked up by subsequent tcp_data_snd_check().
3164 * This function is not for random using!
3167 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3168 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3173 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3175 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3176 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3179 /* Decide wheather to run the increase function of congestion control. */
3180 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3182 if (tcp_in_cwnd_reduction(sk))
3185 /* If reordering is high then always grow cwnd whenever data is
3186 * delivered regardless of its ordering. Otherwise stay conservative
3187 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3188 * new SACK or ECE mark may first advance cwnd here and later reduce
3189 * cwnd in tcp_fastretrans_alert() based on more states.
3191 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3192 return flag & FLAG_FORWARD_PROGRESS;
3194 return flag & FLAG_DATA_ACKED;
3197 /* Check that window update is acceptable.
3198 * The function assumes that snd_una<=ack<=snd_next.
3200 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3201 const u32 ack, const u32 ack_seq,
3204 return after(ack, tp->snd_una) ||
3205 after(ack_seq, tp->snd_wl1) ||
3206 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3209 /* Update our send window.
3211 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3212 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3214 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3217 struct tcp_sock *tp = tcp_sk(sk);
3219 u32 nwin = ntohs(tcp_hdr(skb)->window);
3221 if (likely(!tcp_hdr(skb)->syn))
3222 nwin <<= tp->rx_opt.snd_wscale;
3224 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3225 flag |= FLAG_WIN_UPDATE;
3226 tcp_update_wl(tp, ack_seq);
3228 if (tp->snd_wnd != nwin) {
3231 /* Note, it is the only place, where
3232 * fast path is recovered for sending TCP.
3235 tcp_fast_path_check(sk);
3237 if (nwin > tp->max_window) {
3238 tp->max_window = nwin;
3239 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3249 /* RFC 5961 7 [ACK Throttling] */
3250 static void tcp_send_challenge_ack(struct sock *sk)
3252 /* unprotected vars, we dont care of overwrites */
3253 static u32 challenge_timestamp;
3254 static unsigned int challenge_count;
3255 u32 now = jiffies / HZ;
3257 if (now != challenge_timestamp) {
3258 challenge_timestamp = now;
3259 challenge_count = 0;
3261 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3262 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3267 static void tcp_store_ts_recent(struct tcp_sock *tp)
3269 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3270 tp->rx_opt.ts_recent_stamp = get_seconds();
3273 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3275 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3276 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3277 * extra check below makes sure this can only happen
3278 * for pure ACK frames. -DaveM
3280 * Not only, also it occurs for expired timestamps.
3283 if (tcp_paws_check(&tp->rx_opt, 0))
3284 tcp_store_ts_recent(tp);
3288 /* This routine deals with acks during a TLP episode.
3289 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3291 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3293 struct tcp_sock *tp = tcp_sk(sk);
3294 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3295 !(flag & (FLAG_SND_UNA_ADVANCED |
3296 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3298 /* Mark the end of TLP episode on receiving TLP dupack or when
3299 * ack is after tlp_high_seq.
3301 if (is_tlp_dupack) {
3302 tp->tlp_high_seq = 0;
3306 if (after(ack, tp->tlp_high_seq)) {
3307 tp->tlp_high_seq = 0;
3308 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3309 if (!(flag & FLAG_DSACKING_ACK)) {
3310 tcp_init_cwnd_reduction(sk, true);
3311 tcp_set_ca_state(sk, TCP_CA_CWR);
3312 tcp_end_cwnd_reduction(sk);
3313 tcp_set_ca_state(sk, TCP_CA_Open);
3314 NET_INC_STATS_BH(sock_net(sk),
3315 LINUX_MIB_TCPLOSSPROBERECOVERY);
3320 /* This routine deals with incoming acks, but not outgoing ones. */
3321 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3323 struct inet_connection_sock *icsk = inet_csk(sk);
3324 struct tcp_sock *tp = tcp_sk(sk);
3325 u32 prior_snd_una = tp->snd_una;
3326 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3327 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3328 bool is_dupack = false;
3329 u32 prior_in_flight, prior_cwnd = tp->snd_cwnd, prior_rtt = tp->srtt;
3331 int prior_packets = tp->packets_out;
3332 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3333 int acked = 0; /* Number of packets newly acked */
3336 /* If the ack is older than previous acks
3337 * then we can probably ignore it.
3339 if (before(ack, prior_snd_una)) {
3340 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3341 if (before(ack, prior_snd_una - tp->max_window)) {
3342 tcp_send_challenge_ack(sk);
3348 /* If the ack includes data we haven't sent yet, discard
3349 * this segment (RFC793 Section 3.9).
3351 if (after(ack, tp->snd_nxt))
3354 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3355 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3358 if (after(ack, prior_snd_una))
3359 flag |= FLAG_SND_UNA_ADVANCED;
3361 prior_fackets = tp->fackets_out;
3362 prior_in_flight = tcp_packets_in_flight(tp);
3364 /* ts_recent update must be made after we are sure that the packet
3367 if (flag & FLAG_UPDATE_TS_RECENT)
3368 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3370 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3371 /* Window is constant, pure forward advance.
3372 * No more checks are required.
3373 * Note, we use the fact that SND.UNA>=SND.WL2.
3375 tcp_update_wl(tp, ack_seq);
3377 flag |= FLAG_WIN_UPDATE;
3379 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3381 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3383 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3386 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3388 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3390 if (TCP_SKB_CB(skb)->sacked)
3391 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3394 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3397 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3400 /* We passed data and got it acked, remove any soft error
3401 * log. Something worked...
3403 sk->sk_err_soft = 0;
3404 icsk->icsk_probes_out = 0;
3405 tp->rcv_tstamp = tcp_time_stamp;
3409 /* See if we can take anything off of the retransmit queue. */
3410 acked = tp->packets_out;
3411 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, sack_rtt);
3412 acked -= tp->packets_out;
3414 /* Advance cwnd if state allows */
3415 if (tcp_may_raise_cwnd(sk, flag))
3416 tcp_cong_avoid(sk, ack, prior_in_flight);
3418 if (tcp_ack_is_dubious(sk, flag)) {
3419 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3420 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3423 if (tp->tlp_high_seq)
3424 tcp_process_tlp_ack(sk, ack, flag);
3426 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3427 struct dst_entry *dst = __sk_dst_get(sk);
3432 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3433 tcp_schedule_loss_probe(sk);
3434 if (tp->srtt != prior_rtt || tp->snd_cwnd != prior_cwnd)
3435 tcp_update_pacing_rate(sk);
3439 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3440 if (flag & FLAG_DSACKING_ACK)
3441 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3443 /* If this ack opens up a zero window, clear backoff. It was
3444 * being used to time the probes, and is probably far higher than
3445 * it needs to be for normal retransmission.
3447 if (tcp_send_head(sk))
3450 if (tp->tlp_high_seq)
3451 tcp_process_tlp_ack(sk, ack, flag);
3455 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3459 /* If data was SACKed, tag it and see if we should send more data.
3460 * If data was DSACKed, see if we can undo a cwnd reduction.
3462 if (TCP_SKB_CB(skb)->sacked) {
3463 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3465 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3469 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3473 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3474 * But, this can also be called on packets in the established flow when
3475 * the fast version below fails.
3477 void tcp_parse_options(const struct sk_buff *skb,
3478 struct tcp_options_received *opt_rx, int estab,
3479 struct tcp_fastopen_cookie *foc)
3481 const unsigned char *ptr;
3482 const struct tcphdr *th = tcp_hdr(skb);
3483 int length = (th->doff * 4) - sizeof(struct tcphdr);
3485 ptr = (const unsigned char *)(th + 1);
3486 opt_rx->saw_tstamp = 0;
3488 while (length > 0) {
3489 int opcode = *ptr++;
3495 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3500 if (opsize < 2) /* "silly options" */
3502 if (opsize > length)
3503 return; /* don't parse partial options */
3506 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3507 u16 in_mss = get_unaligned_be16(ptr);
3509 if (opt_rx->user_mss &&
3510 opt_rx->user_mss < in_mss)
3511 in_mss = opt_rx->user_mss;
3512 opt_rx->mss_clamp = in_mss;
3517 if (opsize == TCPOLEN_WINDOW && th->syn &&
3518 !estab && sysctl_tcp_window_scaling) {
3519 __u8 snd_wscale = *(__u8 *)ptr;
3520 opt_rx->wscale_ok = 1;
3521 if (snd_wscale > 14) {
3522 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3527 opt_rx->snd_wscale = snd_wscale;
3530 case TCPOPT_TIMESTAMP:
3531 if ((opsize == TCPOLEN_TIMESTAMP) &&
3532 ((estab && opt_rx->tstamp_ok) ||
3533 (!estab && sysctl_tcp_timestamps))) {
3534 opt_rx->saw_tstamp = 1;
3535 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3536 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3539 case TCPOPT_SACK_PERM:
3540 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3541 !estab && sysctl_tcp_sack) {
3542 opt_rx->sack_ok = TCP_SACK_SEEN;
3543 tcp_sack_reset(opt_rx);
3548 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3549 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3551 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3554 #ifdef CONFIG_TCP_MD5SIG
3557 * The MD5 Hash has already been
3558 * checked (see tcp_v{4,6}_do_rcv()).
3563 /* Fast Open option shares code 254 using a
3564 * 16 bits magic number. It's valid only in
3565 * SYN or SYN-ACK with an even size.
3567 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3568 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3569 foc == NULL || !th->syn || (opsize & 1))
3571 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3572 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3573 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3574 memcpy(foc->val, ptr + 2, foc->len);
3575 else if (foc->len != 0)
3585 EXPORT_SYMBOL(tcp_parse_options);
3587 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3589 const __be32 *ptr = (const __be32 *)(th + 1);
3591 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3592 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3593 tp->rx_opt.saw_tstamp = 1;
3595 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3598 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3600 tp->rx_opt.rcv_tsecr = 0;
3606 /* Fast parse options. This hopes to only see timestamps.
3607 * If it is wrong it falls back on tcp_parse_options().
3609 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3610 const struct tcphdr *th, struct tcp_sock *tp)
3612 /* In the spirit of fast parsing, compare doff directly to constant
3613 * values. Because equality is used, short doff can be ignored here.
3615 if (th->doff == (sizeof(*th) / 4)) {
3616 tp->rx_opt.saw_tstamp = 0;
3618 } else if (tp->rx_opt.tstamp_ok &&
3619 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3620 if (tcp_parse_aligned_timestamp(tp, th))
3624 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3625 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3626 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3631 #ifdef CONFIG_TCP_MD5SIG
3633 * Parse MD5 Signature option
3635 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3637 int length = (th->doff << 2) - sizeof(*th);
3638 const u8 *ptr = (const u8 *)(th + 1);
3640 /* If the TCP option is too short, we can short cut */
3641 if (length < TCPOLEN_MD5SIG)
3644 while (length > 0) {
3645 int opcode = *ptr++;
3656 if (opsize < 2 || opsize > length)
3658 if (opcode == TCPOPT_MD5SIG)
3659 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3666 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3669 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3671 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3672 * it can pass through stack. So, the following predicate verifies that
3673 * this segment is not used for anything but congestion avoidance or
3674 * fast retransmit. Moreover, we even are able to eliminate most of such
3675 * second order effects, if we apply some small "replay" window (~RTO)
3676 * to timestamp space.
3678 * All these measures still do not guarantee that we reject wrapped ACKs
3679 * on networks with high bandwidth, when sequence space is recycled fastly,
3680 * but it guarantees that such events will be very rare and do not affect
3681 * connection seriously. This doesn't look nice, but alas, PAWS is really
3684 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3685 * states that events when retransmit arrives after original data are rare.
3686 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3687 * the biggest problem on large power networks even with minor reordering.
3688 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3689 * up to bandwidth of 18Gigabit/sec. 8) ]
3692 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3694 const struct tcp_sock *tp = tcp_sk(sk);
3695 const struct tcphdr *th = tcp_hdr(skb);
3696 u32 seq = TCP_SKB_CB(skb)->seq;
3697 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3699 return (/* 1. Pure ACK with correct sequence number. */
3700 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3702 /* 2. ... and duplicate ACK. */
3703 ack == tp->snd_una &&
3705 /* 3. ... and does not update window. */
3706 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3708 /* 4. ... and sits in replay window. */
3709 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3712 static inline bool tcp_paws_discard(const struct sock *sk,
3713 const struct sk_buff *skb)
3715 const struct tcp_sock *tp = tcp_sk(sk);
3717 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3718 !tcp_disordered_ack(sk, skb);
3721 /* Check segment sequence number for validity.
3723 * Segment controls are considered valid, if the segment
3724 * fits to the window after truncation to the window. Acceptability
3725 * of data (and SYN, FIN, of course) is checked separately.
3726 * See tcp_data_queue(), for example.
3728 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3729 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3730 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3731 * (borrowed from freebsd)
3734 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3736 return !before(end_seq, tp->rcv_wup) &&
3737 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3740 /* When we get a reset we do this. */
3741 void tcp_reset(struct sock *sk)
3743 /* We want the right error as BSD sees it (and indeed as we do). */
3744 switch (sk->sk_state) {
3746 sk->sk_err = ECONNREFUSED;
3748 case TCP_CLOSE_WAIT:
3754 sk->sk_err = ECONNRESET;
3756 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3759 if (!sock_flag(sk, SOCK_DEAD))
3760 sk->sk_error_report(sk);
3766 * Process the FIN bit. This now behaves as it is supposed to work
3767 * and the FIN takes effect when it is validly part of sequence
3768 * space. Not before when we get holes.
3770 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3771 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3774 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3775 * close and we go into CLOSING (and later onto TIME-WAIT)
3777 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3779 static void tcp_fin(struct sock *sk)
3781 struct tcp_sock *tp = tcp_sk(sk);
3782 const struct dst_entry *dst;
3784 inet_csk_schedule_ack(sk);
3786 sk->sk_shutdown |= RCV_SHUTDOWN;
3787 sock_set_flag(sk, SOCK_DONE);
3789 switch (sk->sk_state) {
3791 case TCP_ESTABLISHED:
3792 /* Move to CLOSE_WAIT */
3793 tcp_set_state(sk, TCP_CLOSE_WAIT);
3794 dst = __sk_dst_get(sk);
3795 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3796 inet_csk(sk)->icsk_ack.pingpong = 1;
3799 case TCP_CLOSE_WAIT:
3801 /* Received a retransmission of the FIN, do
3806 /* RFC793: Remain in the LAST-ACK state. */
3810 /* This case occurs when a simultaneous close
3811 * happens, we must ack the received FIN and
3812 * enter the CLOSING state.
3815 tcp_set_state(sk, TCP_CLOSING);
3818 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3820 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3823 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3824 * cases we should never reach this piece of code.
3826 pr_err("%s: Impossible, sk->sk_state=%d\n",
3827 __func__, sk->sk_state);
3831 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3832 * Probably, we should reset in this case. For now drop them.
3834 __skb_queue_purge(&tp->out_of_order_queue);
3835 if (tcp_is_sack(tp))
3836 tcp_sack_reset(&tp->rx_opt);
3839 if (!sock_flag(sk, SOCK_DEAD)) {
3840 sk->sk_state_change(sk);
3842 /* Do not send POLL_HUP for half duplex close. */
3843 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3844 sk->sk_state == TCP_CLOSE)
3845 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3847 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3851 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3854 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3855 if (before(seq, sp->start_seq))
3856 sp->start_seq = seq;
3857 if (after(end_seq, sp->end_seq))
3858 sp->end_seq = end_seq;
3864 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3866 struct tcp_sock *tp = tcp_sk(sk);
3868 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3871 if (before(seq, tp->rcv_nxt))
3872 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3874 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3876 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3878 tp->rx_opt.dsack = 1;
3879 tp->duplicate_sack[0].start_seq = seq;
3880 tp->duplicate_sack[0].end_seq = end_seq;
3884 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3886 struct tcp_sock *tp = tcp_sk(sk);
3888 if (!tp->rx_opt.dsack)
3889 tcp_dsack_set(sk, seq, end_seq);
3891 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3894 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3896 struct tcp_sock *tp = tcp_sk(sk);
3898 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3899 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3900 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3901 tcp_enter_quickack_mode(sk);
3903 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3904 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3906 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3907 end_seq = tp->rcv_nxt;
3908 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3915 /* These routines update the SACK block as out-of-order packets arrive or
3916 * in-order packets close up the sequence space.
3918 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3921 struct tcp_sack_block *sp = &tp->selective_acks[0];
3922 struct tcp_sack_block *swalk = sp + 1;
3924 /* See if the recent change to the first SACK eats into
3925 * or hits the sequence space of other SACK blocks, if so coalesce.
3927 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3928 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3931 /* Zap SWALK, by moving every further SACK up by one slot.
3932 * Decrease num_sacks.
3934 tp->rx_opt.num_sacks--;
3935 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3939 this_sack++, swalk++;
3943 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3945 struct tcp_sock *tp = tcp_sk(sk);
3946 struct tcp_sack_block *sp = &tp->selective_acks[0];
3947 int cur_sacks = tp->rx_opt.num_sacks;
3953 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3954 if (tcp_sack_extend(sp, seq, end_seq)) {
3955 /* Rotate this_sack to the first one. */
3956 for (; this_sack > 0; this_sack--, sp--)
3957 swap(*sp, *(sp - 1));
3959 tcp_sack_maybe_coalesce(tp);
3964 /* Could not find an adjacent existing SACK, build a new one,
3965 * put it at the front, and shift everyone else down. We
3966 * always know there is at least one SACK present already here.
3968 * If the sack array is full, forget about the last one.
3970 if (this_sack >= TCP_NUM_SACKS) {
3972 tp->rx_opt.num_sacks--;
3975 for (; this_sack > 0; this_sack--, sp--)
3979 /* Build the new head SACK, and we're done. */
3980 sp->start_seq = seq;
3981 sp->end_seq = end_seq;
3982 tp->rx_opt.num_sacks++;
3985 /* RCV.NXT advances, some SACKs should be eaten. */
3987 static void tcp_sack_remove(struct tcp_sock *tp)
3989 struct tcp_sack_block *sp = &tp->selective_acks[0];
3990 int num_sacks = tp->rx_opt.num_sacks;
3993 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3994 if (skb_queue_empty(&tp->out_of_order_queue)) {
3995 tp->rx_opt.num_sacks = 0;
3999 for (this_sack = 0; this_sack < num_sacks;) {
4000 /* Check if the start of the sack is covered by RCV.NXT. */
4001 if (!before(tp->rcv_nxt, sp->start_seq)) {
4004 /* RCV.NXT must cover all the block! */
4005 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4007 /* Zap this SACK, by moving forward any other SACKS. */
4008 for (i=this_sack+1; i < num_sacks; i++)
4009 tp->selective_acks[i-1] = tp->selective_acks[i];
4016 tp->rx_opt.num_sacks = num_sacks;
4019 /* This one checks to see if we can put data from the
4020 * out_of_order queue into the receive_queue.
4022 static void tcp_ofo_queue(struct sock *sk)
4024 struct tcp_sock *tp = tcp_sk(sk);
4025 __u32 dsack_high = tp->rcv_nxt;
4026 struct sk_buff *skb;
4028 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4029 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4032 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4033 __u32 dsack = dsack_high;
4034 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4035 dsack_high = TCP_SKB_CB(skb)->end_seq;
4036 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4039 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4040 SOCK_DEBUG(sk, "ofo packet was already received\n");
4041 __skb_unlink(skb, &tp->out_of_order_queue);
4045 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4046 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4047 TCP_SKB_CB(skb)->end_seq);
4049 __skb_unlink(skb, &tp->out_of_order_queue);
4050 __skb_queue_tail(&sk->sk_receive_queue, skb);
4051 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4052 if (tcp_hdr(skb)->fin)
4057 static bool tcp_prune_ofo_queue(struct sock *sk);
4058 static int tcp_prune_queue(struct sock *sk);
4060 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4063 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4064 !sk_rmem_schedule(sk, skb, size)) {
4066 if (tcp_prune_queue(sk) < 0)
4069 if (!sk_rmem_schedule(sk, skb, size)) {
4070 if (!tcp_prune_ofo_queue(sk))
4073 if (!sk_rmem_schedule(sk, skb, size))
4081 * tcp_try_coalesce - try to merge skb to prior one
4084 * @from: buffer to add in queue
4085 * @fragstolen: pointer to boolean
4087 * Before queueing skb @from after @to, try to merge them
4088 * to reduce overall memory use and queue lengths, if cost is small.
4089 * Packets in ofo or receive queues can stay a long time.
4090 * Better try to coalesce them right now to avoid future collapses.
4091 * Returns true if caller should free @from instead of queueing it
4093 static bool tcp_try_coalesce(struct sock *sk,
4095 struct sk_buff *from,
4100 *fragstolen = false;
4102 if (tcp_hdr(from)->fin)
4105 /* Its possible this segment overlaps with prior segment in queue */
4106 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4109 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4112 atomic_add(delta, &sk->sk_rmem_alloc);
4113 sk_mem_charge(sk, delta);
4114 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4115 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4116 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4120 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4122 struct tcp_sock *tp = tcp_sk(sk);
4123 struct sk_buff *skb1;
4126 TCP_ECN_check_ce(tp, skb);
4128 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4129 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4134 /* Disable header prediction. */
4136 inet_csk_schedule_ack(sk);
4138 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4139 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4140 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4142 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4144 /* Initial out of order segment, build 1 SACK. */
4145 if (tcp_is_sack(tp)) {
4146 tp->rx_opt.num_sacks = 1;
4147 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4148 tp->selective_acks[0].end_seq =
4149 TCP_SKB_CB(skb)->end_seq;
4151 __skb_queue_head(&tp->out_of_order_queue, skb);
4155 seq = TCP_SKB_CB(skb)->seq;
4156 end_seq = TCP_SKB_CB(skb)->end_seq;
4158 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4161 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4162 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4164 tcp_grow_window(sk, skb);
4165 kfree_skb_partial(skb, fragstolen);
4169 if (!tp->rx_opt.num_sacks ||
4170 tp->selective_acks[0].end_seq != seq)
4173 /* Common case: data arrive in order after hole. */
4174 tp->selective_acks[0].end_seq = end_seq;
4178 /* Find place to insert this segment. */
4180 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4182 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4186 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4189 /* Do skb overlap to previous one? */
4190 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4191 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4192 /* All the bits are present. Drop. */
4193 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4196 tcp_dsack_set(sk, seq, end_seq);
4199 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4200 /* Partial overlap. */
4201 tcp_dsack_set(sk, seq,
4202 TCP_SKB_CB(skb1)->end_seq);
4204 if (skb_queue_is_first(&tp->out_of_order_queue,
4208 skb1 = skb_queue_prev(
4209 &tp->out_of_order_queue,
4214 __skb_queue_head(&tp->out_of_order_queue, skb);
4216 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4218 /* And clean segments covered by new one as whole. */
4219 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4220 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4222 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4224 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4225 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4229 __skb_unlink(skb1, &tp->out_of_order_queue);
4230 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4231 TCP_SKB_CB(skb1)->end_seq);
4232 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4237 if (tcp_is_sack(tp))
4238 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4241 tcp_grow_window(sk, skb);
4242 skb_set_owner_r(skb, sk);
4246 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4250 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4252 __skb_pull(skb, hdrlen);
4254 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4255 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4257 __skb_queue_tail(&sk->sk_receive_queue, skb);
4258 skb_set_owner_r(skb, sk);
4263 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4265 struct sk_buff *skb = NULL;
4272 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4276 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4279 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4280 skb_reset_transport_header(skb);
4281 memset(th, 0, sizeof(*th));
4283 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4286 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4287 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4288 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4290 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4291 WARN_ON_ONCE(fragstolen); /* should not happen */
4302 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4304 const struct tcphdr *th = tcp_hdr(skb);
4305 struct tcp_sock *tp = tcp_sk(sk);
4307 bool fragstolen = false;
4309 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4313 __skb_pull(skb, th->doff * 4);
4315 TCP_ECN_accept_cwr(tp, skb);
4317 tp->rx_opt.dsack = 0;
4319 /* Queue data for delivery to the user.
4320 * Packets in sequence go to the receive queue.
4321 * Out of sequence packets to the out_of_order_queue.
4323 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4324 if (tcp_receive_window(tp) == 0)
4327 /* Ok. In sequence. In window. */
4328 if (tp->ucopy.task == current &&
4329 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4330 sock_owned_by_user(sk) && !tp->urg_data) {
4331 int chunk = min_t(unsigned int, skb->len,
4334 __set_current_state(TASK_RUNNING);
4337 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4338 tp->ucopy.len -= chunk;
4339 tp->copied_seq += chunk;
4340 eaten = (chunk == skb->len);
4341 tcp_rcv_space_adjust(sk);
4349 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4352 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4354 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4356 tcp_event_data_recv(sk, skb);
4360 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4363 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4364 * gap in queue is filled.
4366 if (skb_queue_empty(&tp->out_of_order_queue))
4367 inet_csk(sk)->icsk_ack.pingpong = 0;
4370 if (tp->rx_opt.num_sacks)
4371 tcp_sack_remove(tp);
4373 tcp_fast_path_check(sk);
4376 kfree_skb_partial(skb, fragstolen);
4377 if (!sock_flag(sk, SOCK_DEAD))
4378 sk->sk_data_ready(sk, 0);
4382 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4383 /* A retransmit, 2nd most common case. Force an immediate ack. */
4384 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4385 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4388 tcp_enter_quickack_mode(sk);
4389 inet_csk_schedule_ack(sk);
4395 /* Out of window. F.e. zero window probe. */
4396 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4399 tcp_enter_quickack_mode(sk);
4401 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4402 /* Partial packet, seq < rcv_next < end_seq */
4403 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4404 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4405 TCP_SKB_CB(skb)->end_seq);
4407 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4409 /* If window is closed, drop tail of packet. But after
4410 * remembering D-SACK for its head made in previous line.
4412 if (!tcp_receive_window(tp))
4417 tcp_data_queue_ofo(sk, skb);
4420 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4421 struct sk_buff_head *list)
4423 struct sk_buff *next = NULL;
4425 if (!skb_queue_is_last(list, skb))
4426 next = skb_queue_next(list, skb);
4428 __skb_unlink(skb, list);
4430 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4435 /* Collapse contiguous sequence of skbs head..tail with
4436 * sequence numbers start..end.
4438 * If tail is NULL, this means until the end of the list.
4440 * Segments with FIN/SYN are not collapsed (only because this
4444 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4445 struct sk_buff *head, struct sk_buff *tail,
4448 struct sk_buff *skb, *n;
4451 /* First, check that queue is collapsible and find
4452 * the point where collapsing can be useful. */
4456 skb_queue_walk_from_safe(list, skb, n) {
4459 /* No new bits? It is possible on ofo queue. */
4460 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4461 skb = tcp_collapse_one(sk, skb, list);
4467 /* The first skb to collapse is:
4469 * - bloated or contains data before "start" or
4470 * overlaps to the next one.
4472 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4473 (tcp_win_from_space(skb->truesize) > skb->len ||
4474 before(TCP_SKB_CB(skb)->seq, start))) {
4475 end_of_skbs = false;
4479 if (!skb_queue_is_last(list, skb)) {
4480 struct sk_buff *next = skb_queue_next(list, skb);
4482 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4483 end_of_skbs = false;
4488 /* Decided to skip this, advance start seq. */
4489 start = TCP_SKB_CB(skb)->end_seq;
4491 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4494 while (before(start, end)) {
4495 struct sk_buff *nskb;
4496 unsigned int header = skb_headroom(skb);
4497 int copy = SKB_MAX_ORDER(header, 0);
4499 /* Too big header? This can happen with IPv6. */
4502 if (end - start < copy)
4504 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4508 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4509 skb_set_network_header(nskb, (skb_network_header(skb) -
4511 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4513 skb_reserve(nskb, header);
4514 memcpy(nskb->head, skb->head, header);
4515 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4516 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4517 __skb_queue_before(list, skb, nskb);
4518 skb_set_owner_r(nskb, sk);
4520 /* Copy data, releasing collapsed skbs. */
4522 int offset = start - TCP_SKB_CB(skb)->seq;
4523 int size = TCP_SKB_CB(skb)->end_seq - start;
4527 size = min(copy, size);
4528 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4530 TCP_SKB_CB(nskb)->end_seq += size;
4534 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4535 skb = tcp_collapse_one(sk, skb, list);
4538 tcp_hdr(skb)->syn ||
4546 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4547 * and tcp_collapse() them until all the queue is collapsed.
4549 static void tcp_collapse_ofo_queue(struct sock *sk)
4551 struct tcp_sock *tp = tcp_sk(sk);
4552 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4553 struct sk_buff *head;
4559 start = TCP_SKB_CB(skb)->seq;
4560 end = TCP_SKB_CB(skb)->end_seq;
4564 struct sk_buff *next = NULL;
4566 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4567 next = skb_queue_next(&tp->out_of_order_queue, skb);
4570 /* Segment is terminated when we see gap or when
4571 * we are at the end of all the queue. */
4573 after(TCP_SKB_CB(skb)->seq, end) ||
4574 before(TCP_SKB_CB(skb)->end_seq, start)) {
4575 tcp_collapse(sk, &tp->out_of_order_queue,
4576 head, skb, start, end);
4580 /* Start new segment */
4581 start = TCP_SKB_CB(skb)->seq;
4582 end = TCP_SKB_CB(skb)->end_seq;
4584 if (before(TCP_SKB_CB(skb)->seq, start))
4585 start = TCP_SKB_CB(skb)->seq;
4586 if (after(TCP_SKB_CB(skb)->end_seq, end))
4587 end = TCP_SKB_CB(skb)->end_seq;
4593 * Purge the out-of-order queue.
4594 * Return true if queue was pruned.
4596 static bool tcp_prune_ofo_queue(struct sock *sk)
4598 struct tcp_sock *tp = tcp_sk(sk);
4601 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4602 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4603 __skb_queue_purge(&tp->out_of_order_queue);
4605 /* Reset SACK state. A conforming SACK implementation will
4606 * do the same at a timeout based retransmit. When a connection
4607 * is in a sad state like this, we care only about integrity
4608 * of the connection not performance.
4610 if (tp->rx_opt.sack_ok)
4611 tcp_sack_reset(&tp->rx_opt);
4618 /* Reduce allocated memory if we can, trying to get
4619 * the socket within its memory limits again.
4621 * Return less than zero if we should start dropping frames
4622 * until the socket owning process reads some of the data
4623 * to stabilize the situation.
4625 static int tcp_prune_queue(struct sock *sk)
4627 struct tcp_sock *tp = tcp_sk(sk);
4629 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4631 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4633 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4634 tcp_clamp_window(sk);
4635 else if (sk_under_memory_pressure(sk))
4636 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4638 tcp_collapse_ofo_queue(sk);
4639 if (!skb_queue_empty(&sk->sk_receive_queue))
4640 tcp_collapse(sk, &sk->sk_receive_queue,
4641 skb_peek(&sk->sk_receive_queue),
4643 tp->copied_seq, tp->rcv_nxt);
4646 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4649 /* Collapsing did not help, destructive actions follow.
4650 * This must not ever occur. */
4652 tcp_prune_ofo_queue(sk);
4654 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4657 /* If we are really being abused, tell the caller to silently
4658 * drop receive data on the floor. It will get retransmitted
4659 * and hopefully then we'll have sufficient space.
4661 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4663 /* Massive buffer overcommit. */
4668 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4669 * As additional protections, we do not touch cwnd in retransmission phases,
4670 * and if application hit its sndbuf limit recently.
4672 void tcp_cwnd_application_limited(struct sock *sk)
4674 struct tcp_sock *tp = tcp_sk(sk);
4676 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4677 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4678 /* Limited by application or receiver window. */
4679 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4680 u32 win_used = max(tp->snd_cwnd_used, init_win);
4681 if (win_used < tp->snd_cwnd) {
4682 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4683 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4685 tp->snd_cwnd_used = 0;
4687 tp->snd_cwnd_stamp = tcp_time_stamp;
4690 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4692 const struct tcp_sock *tp = tcp_sk(sk);
4694 /* If the user specified a specific send buffer setting, do
4697 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4700 /* If we are under global TCP memory pressure, do not expand. */
4701 if (sk_under_memory_pressure(sk))
4704 /* If we are under soft global TCP memory pressure, do not expand. */
4705 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4708 /* If we filled the congestion window, do not expand. */
4709 if (tp->packets_out >= tp->snd_cwnd)
4715 /* When incoming ACK allowed to free some skb from write_queue,
4716 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4717 * on the exit from tcp input handler.
4719 * PROBLEM: sndbuf expansion does not work well with largesend.
4721 static void tcp_new_space(struct sock *sk)
4723 struct tcp_sock *tp = tcp_sk(sk);
4725 if (tcp_should_expand_sndbuf(sk)) {
4726 int sndmem = SKB_TRUESIZE(max_t(u32,
4727 tp->rx_opt.mss_clamp,
4730 int demanded = max_t(unsigned int, tp->snd_cwnd,
4731 tp->reordering + 1);
4732 sndmem *= 2 * demanded;
4733 if (sndmem > sk->sk_sndbuf)
4734 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4735 tp->snd_cwnd_stamp = tcp_time_stamp;
4738 sk->sk_write_space(sk);
4741 static void tcp_check_space(struct sock *sk)
4743 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4744 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4745 if (sk->sk_socket &&
4746 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4751 static inline void tcp_data_snd_check(struct sock *sk)
4753 tcp_push_pending_frames(sk);
4754 tcp_check_space(sk);
4758 * Check if sending an ack is needed.
4760 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4762 struct tcp_sock *tp = tcp_sk(sk);
4764 /* More than one full frame received... */
4765 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4766 /* ... and right edge of window advances far enough.
4767 * (tcp_recvmsg() will send ACK otherwise). Or...
4769 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4770 /* We ACK each frame or... */
4771 tcp_in_quickack_mode(sk) ||
4772 /* We have out of order data. */
4773 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4774 /* Then ack it now */
4777 /* Else, send delayed ack. */
4778 tcp_send_delayed_ack(sk);
4782 static inline void tcp_ack_snd_check(struct sock *sk)
4784 if (!inet_csk_ack_scheduled(sk)) {
4785 /* We sent a data segment already. */
4788 __tcp_ack_snd_check(sk, 1);
4792 * This routine is only called when we have urgent data
4793 * signaled. Its the 'slow' part of tcp_urg. It could be
4794 * moved inline now as tcp_urg is only called from one
4795 * place. We handle URGent data wrong. We have to - as
4796 * BSD still doesn't use the correction from RFC961.
4797 * For 1003.1g we should support a new option TCP_STDURG to permit
4798 * either form (or just set the sysctl tcp_stdurg).
4801 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4803 struct tcp_sock *tp = tcp_sk(sk);
4804 u32 ptr = ntohs(th->urg_ptr);
4806 if (ptr && !sysctl_tcp_stdurg)
4808 ptr += ntohl(th->seq);
4810 /* Ignore urgent data that we've already seen and read. */
4811 if (after(tp->copied_seq, ptr))
4814 /* Do not replay urg ptr.
4816 * NOTE: interesting situation not covered by specs.
4817 * Misbehaving sender may send urg ptr, pointing to segment,
4818 * which we already have in ofo queue. We are not able to fetch
4819 * such data and will stay in TCP_URG_NOTYET until will be eaten
4820 * by recvmsg(). Seems, we are not obliged to handle such wicked
4821 * situations. But it is worth to think about possibility of some
4822 * DoSes using some hypothetical application level deadlock.
4824 if (before(ptr, tp->rcv_nxt))
4827 /* Do we already have a newer (or duplicate) urgent pointer? */
4828 if (tp->urg_data && !after(ptr, tp->urg_seq))
4831 /* Tell the world about our new urgent pointer. */
4834 /* We may be adding urgent data when the last byte read was
4835 * urgent. To do this requires some care. We cannot just ignore
4836 * tp->copied_seq since we would read the last urgent byte again
4837 * as data, nor can we alter copied_seq until this data arrives
4838 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4840 * NOTE. Double Dutch. Rendering to plain English: author of comment
4841 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4842 * and expect that both A and B disappear from stream. This is _wrong_.
4843 * Though this happens in BSD with high probability, this is occasional.
4844 * Any application relying on this is buggy. Note also, that fix "works"
4845 * only in this artificial test. Insert some normal data between A and B and we will
4846 * decline of BSD again. Verdict: it is better to remove to trap
4849 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4850 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4851 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4853 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4854 __skb_unlink(skb, &sk->sk_receive_queue);
4859 tp->urg_data = TCP_URG_NOTYET;
4862 /* Disable header prediction. */
4866 /* This is the 'fast' part of urgent handling. */
4867 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4869 struct tcp_sock *tp = tcp_sk(sk);
4871 /* Check if we get a new urgent pointer - normally not. */
4873 tcp_check_urg(sk, th);
4875 /* Do we wait for any urgent data? - normally not... */
4876 if (tp->urg_data == TCP_URG_NOTYET) {
4877 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4880 /* Is the urgent pointer pointing into this packet? */
4881 if (ptr < skb->len) {
4883 if (skb_copy_bits(skb, ptr, &tmp, 1))
4885 tp->urg_data = TCP_URG_VALID | tmp;
4886 if (!sock_flag(sk, SOCK_DEAD))
4887 sk->sk_data_ready(sk, 0);
4892 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4894 struct tcp_sock *tp = tcp_sk(sk);
4895 int chunk = skb->len - hlen;
4899 if (skb_csum_unnecessary(skb))
4900 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4902 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4906 tp->ucopy.len -= chunk;
4907 tp->copied_seq += chunk;
4908 tcp_rcv_space_adjust(sk);
4915 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4916 struct sk_buff *skb)
4920 if (sock_owned_by_user(sk)) {
4922 result = __tcp_checksum_complete(skb);
4925 result = __tcp_checksum_complete(skb);
4930 static inline bool tcp_checksum_complete_user(struct sock *sk,
4931 struct sk_buff *skb)
4933 return !skb_csum_unnecessary(skb) &&
4934 __tcp_checksum_complete_user(sk, skb);
4937 #ifdef CONFIG_NET_DMA
4938 static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4941 struct tcp_sock *tp = tcp_sk(sk);
4942 int chunk = skb->len - hlen;
4944 bool copied_early = false;
4946 if (tp->ucopy.wakeup)
4949 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4950 tp->ucopy.dma_chan = net_dma_find_channel();
4952 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4954 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4956 tp->ucopy.iov, chunk,
4957 tp->ucopy.pinned_list);
4962 tp->ucopy.dma_cookie = dma_cookie;
4963 copied_early = true;
4965 tp->ucopy.len -= chunk;
4966 tp->copied_seq += chunk;
4967 tcp_rcv_space_adjust(sk);
4969 if ((tp->ucopy.len == 0) ||
4970 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4971 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4972 tp->ucopy.wakeup = 1;
4973 sk->sk_data_ready(sk, 0);
4975 } else if (chunk > 0) {
4976 tp->ucopy.wakeup = 1;
4977 sk->sk_data_ready(sk, 0);
4980 return copied_early;
4982 #endif /* CONFIG_NET_DMA */
4984 /* Does PAWS and seqno based validation of an incoming segment, flags will
4985 * play significant role here.
4987 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4988 const struct tcphdr *th, int syn_inerr)
4990 struct tcp_sock *tp = tcp_sk(sk);
4992 /* RFC1323: H1. Apply PAWS check first. */
4993 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4994 tcp_paws_discard(sk, skb)) {
4996 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4997 tcp_send_dupack(sk, skb);
5000 /* Reset is accepted even if it did not pass PAWS. */
5003 /* Step 1: check sequence number */
5004 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5005 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5006 * (RST) segments are validated by checking their SEQ-fields."
5007 * And page 69: "If an incoming segment is not acceptable,
5008 * an acknowledgment should be sent in reply (unless the RST
5009 * bit is set, if so drop the segment and return)".
5014 tcp_send_dupack(sk, skb);
5019 /* Step 2: check RST bit */
5022 * If sequence number exactly matches RCV.NXT, then
5023 * RESET the connection
5025 * Send a challenge ACK
5027 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5030 tcp_send_challenge_ack(sk);
5034 /* step 3: check security and precedence [ignored] */
5036 /* step 4: Check for a SYN
5037 * RFC 5691 4.2 : Send a challenge ack
5042 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5043 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5044 tcp_send_challenge_ack(sk);
5056 * TCP receive function for the ESTABLISHED state.
5058 * It is split into a fast path and a slow path. The fast path is
5060 * - A zero window was announced from us - zero window probing
5061 * is only handled properly in the slow path.
5062 * - Out of order segments arrived.
5063 * - Urgent data is expected.
5064 * - There is no buffer space left
5065 * - Unexpected TCP flags/window values/header lengths are received
5066 * (detected by checking the TCP header against pred_flags)
5067 * - Data is sent in both directions. Fast path only supports pure senders
5068 * or pure receivers (this means either the sequence number or the ack
5069 * value must stay constant)
5070 * - Unexpected TCP option.
5072 * When these conditions are not satisfied it drops into a standard
5073 * receive procedure patterned after RFC793 to handle all cases.
5074 * The first three cases are guaranteed by proper pred_flags setting,
5075 * the rest is checked inline. Fast processing is turned on in
5076 * tcp_data_queue when everything is OK.
5078 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5079 const struct tcphdr *th, unsigned int len)
5081 struct tcp_sock *tp = tcp_sk(sk);
5083 if (unlikely(sk->sk_rx_dst == NULL))
5084 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5086 * Header prediction.
5087 * The code loosely follows the one in the famous
5088 * "30 instruction TCP receive" Van Jacobson mail.
5090 * Van's trick is to deposit buffers into socket queue
5091 * on a device interrupt, to call tcp_recv function
5092 * on the receive process context and checksum and copy
5093 * the buffer to user space. smart...
5095 * Our current scheme is not silly either but we take the
5096 * extra cost of the net_bh soft interrupt processing...
5097 * We do checksum and copy also but from device to kernel.
5100 tp->rx_opt.saw_tstamp = 0;
5102 /* pred_flags is 0xS?10 << 16 + snd_wnd
5103 * if header_prediction is to be made
5104 * 'S' will always be tp->tcp_header_len >> 2
5105 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5106 * turn it off (when there are holes in the receive
5107 * space for instance)
5108 * PSH flag is ignored.
5111 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5112 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5113 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5114 int tcp_header_len = tp->tcp_header_len;
5116 /* Timestamp header prediction: tcp_header_len
5117 * is automatically equal to th->doff*4 due to pred_flags
5121 /* Check timestamp */
5122 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5123 /* No? Slow path! */
5124 if (!tcp_parse_aligned_timestamp(tp, th))
5127 /* If PAWS failed, check it more carefully in slow path */
5128 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5131 /* DO NOT update ts_recent here, if checksum fails
5132 * and timestamp was corrupted part, it will result
5133 * in a hung connection since we will drop all
5134 * future packets due to the PAWS test.
5138 if (len <= tcp_header_len) {
5139 /* Bulk data transfer: sender */
5140 if (len == tcp_header_len) {
5141 /* Predicted packet is in window by definition.
5142 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5143 * Hence, check seq<=rcv_wup reduces to:
5145 if (tcp_header_len ==
5146 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5147 tp->rcv_nxt == tp->rcv_wup)
5148 tcp_store_ts_recent(tp);
5150 /* We know that such packets are checksummed
5153 tcp_ack(sk, skb, 0);
5155 tcp_data_snd_check(sk);
5157 } else { /* Header too small */
5158 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5163 int copied_early = 0;
5164 bool fragstolen = false;
5166 if (tp->copied_seq == tp->rcv_nxt &&
5167 len - tcp_header_len <= tp->ucopy.len) {
5168 #ifdef CONFIG_NET_DMA
5169 if (tp->ucopy.task == current &&
5170 sock_owned_by_user(sk) &&
5171 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5176 if (tp->ucopy.task == current &&
5177 sock_owned_by_user(sk) && !copied_early) {
5178 __set_current_state(TASK_RUNNING);
5180 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5184 /* Predicted packet is in window by definition.
5185 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5186 * Hence, check seq<=rcv_wup reduces to:
5188 if (tcp_header_len ==
5189 (sizeof(struct tcphdr) +
5190 TCPOLEN_TSTAMP_ALIGNED) &&
5191 tp->rcv_nxt == tp->rcv_wup)
5192 tcp_store_ts_recent(tp);
5194 tcp_rcv_rtt_measure_ts(sk, skb);
5196 __skb_pull(skb, tcp_header_len);
5197 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5198 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5201 tcp_cleanup_rbuf(sk, skb->len);
5204 if (tcp_checksum_complete_user(sk, skb))
5207 if ((int)skb->truesize > sk->sk_forward_alloc)
5210 /* Predicted packet is in window by definition.
5211 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5212 * Hence, check seq<=rcv_wup reduces to:
5214 if (tcp_header_len ==
5215 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5216 tp->rcv_nxt == tp->rcv_wup)
5217 tcp_store_ts_recent(tp);
5219 tcp_rcv_rtt_measure_ts(sk, skb);
5221 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5223 /* Bulk data transfer: receiver */
5224 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5228 tcp_event_data_recv(sk, skb);
5230 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5231 /* Well, only one small jumplet in fast path... */
5232 tcp_ack(sk, skb, FLAG_DATA);
5233 tcp_data_snd_check(sk);
5234 if (!inet_csk_ack_scheduled(sk))
5238 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5239 __tcp_ack_snd_check(sk, 0);
5241 #ifdef CONFIG_NET_DMA
5243 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5247 kfree_skb_partial(skb, fragstolen);
5248 sk->sk_data_ready(sk, 0);
5254 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5257 if (!th->ack && !th->rst)
5261 * Standard slow path.
5264 if (!tcp_validate_incoming(sk, skb, th, 1))
5268 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5271 tcp_rcv_rtt_measure_ts(sk, skb);
5273 /* Process urgent data. */
5274 tcp_urg(sk, skb, th);
5276 /* step 7: process the segment text */
5277 tcp_data_queue(sk, skb);
5279 tcp_data_snd_check(sk);
5280 tcp_ack_snd_check(sk);
5284 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5285 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5290 EXPORT_SYMBOL(tcp_rcv_established);
5292 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5294 struct tcp_sock *tp = tcp_sk(sk);
5295 struct inet_connection_sock *icsk = inet_csk(sk);
5297 tcp_set_state(sk, TCP_ESTABLISHED);
5300 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5301 security_inet_conn_established(sk, skb);
5304 /* Make sure socket is routed, for correct metrics. */
5305 icsk->icsk_af_ops->rebuild_header(sk);
5307 tcp_init_metrics(sk);
5309 tcp_init_congestion_control(sk);
5311 /* Prevent spurious tcp_cwnd_restart() on first data
5314 tp->lsndtime = tcp_time_stamp;
5316 tcp_init_buffer_space(sk);
5318 if (sock_flag(sk, SOCK_KEEPOPEN))
5319 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5321 if (!tp->rx_opt.snd_wscale)
5322 __tcp_fast_path_on(tp, tp->snd_wnd);
5326 if (!sock_flag(sk, SOCK_DEAD)) {
5327 sk->sk_state_change(sk);
5328 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5332 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5333 struct tcp_fastopen_cookie *cookie)
5335 struct tcp_sock *tp = tcp_sk(sk);
5336 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5337 u16 mss = tp->rx_opt.mss_clamp;
5340 if (mss == tp->rx_opt.user_mss) {
5341 struct tcp_options_received opt;
5343 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5344 tcp_clear_options(&opt);
5345 opt.user_mss = opt.mss_clamp = 0;
5346 tcp_parse_options(synack, &opt, 0, NULL);
5347 mss = opt.mss_clamp;
5350 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5353 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5354 * the remote receives only the retransmitted (regular) SYNs: either
5355 * the original SYN-data or the corresponding SYN-ACK is lost.
5357 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5359 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5361 if (data) { /* Retransmit unacked data in SYN */
5362 tcp_for_write_queue_from(data, sk) {
5363 if (data == tcp_send_head(sk) ||
5364 __tcp_retransmit_skb(sk, data))
5370 tp->syn_data_acked = tp->syn_data;
5374 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5375 const struct tcphdr *th, unsigned int len)
5377 struct inet_connection_sock *icsk = inet_csk(sk);
5378 struct tcp_sock *tp = tcp_sk(sk);
5379 struct tcp_fastopen_cookie foc = { .len = -1 };
5380 int saved_clamp = tp->rx_opt.mss_clamp;
5382 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5383 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5384 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5388 * "If the state is SYN-SENT then
5389 * first check the ACK bit
5390 * If the ACK bit is set
5391 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5392 * a reset (unless the RST bit is set, if so drop
5393 * the segment and return)"
5395 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5396 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5397 goto reset_and_undo;
5399 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5400 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5402 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5403 goto reset_and_undo;
5406 /* Now ACK is acceptable.
5408 * "If the RST bit is set
5409 * If the ACK was acceptable then signal the user "error:
5410 * connection reset", drop the segment, enter CLOSED state,
5411 * delete TCB, and return."
5420 * "fifth, if neither of the SYN or RST bits is set then
5421 * drop the segment and return."
5427 goto discard_and_undo;
5430 * "If the SYN bit is on ...
5431 * are acceptable then ...
5432 * (our SYN has been ACKed), change the connection
5433 * state to ESTABLISHED..."
5436 TCP_ECN_rcv_synack(tp, th);
5438 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5439 tcp_ack(sk, skb, FLAG_SLOWPATH);
5441 /* Ok.. it's good. Set up sequence numbers and
5442 * move to established.
5444 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5445 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5447 /* RFC1323: The window in SYN & SYN/ACK segments is
5450 tp->snd_wnd = ntohs(th->window);
5452 if (!tp->rx_opt.wscale_ok) {
5453 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5454 tp->window_clamp = min(tp->window_clamp, 65535U);
5457 if (tp->rx_opt.saw_tstamp) {
5458 tp->rx_opt.tstamp_ok = 1;
5459 tp->tcp_header_len =
5460 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5461 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5462 tcp_store_ts_recent(tp);
5464 tp->tcp_header_len = sizeof(struct tcphdr);
5467 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5468 tcp_enable_fack(tp);
5471 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5472 tcp_initialize_rcv_mss(sk);
5474 /* Remember, tcp_poll() does not lock socket!
5475 * Change state from SYN-SENT only after copied_seq
5476 * is initialized. */
5477 tp->copied_seq = tp->rcv_nxt;
5481 tcp_finish_connect(sk, skb);
5483 if ((tp->syn_fastopen || tp->syn_data) &&
5484 tcp_rcv_fastopen_synack(sk, skb, &foc))
5487 if (sk->sk_write_pending ||
5488 icsk->icsk_accept_queue.rskq_defer_accept ||
5489 icsk->icsk_ack.pingpong) {
5490 /* Save one ACK. Data will be ready after
5491 * several ticks, if write_pending is set.
5493 * It may be deleted, but with this feature tcpdumps
5494 * look so _wonderfully_ clever, that I was not able
5495 * to stand against the temptation 8) --ANK
5497 inet_csk_schedule_ack(sk);
5498 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5499 tcp_enter_quickack_mode(sk);
5500 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5501 TCP_DELACK_MAX, TCP_RTO_MAX);
5512 /* No ACK in the segment */
5516 * "If the RST bit is set
5518 * Otherwise (no ACK) drop the segment and return."
5521 goto discard_and_undo;
5525 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5526 tcp_paws_reject(&tp->rx_opt, 0))
5527 goto discard_and_undo;
5530 /* We see SYN without ACK. It is attempt of
5531 * simultaneous connect with crossed SYNs.
5532 * Particularly, it can be connect to self.
5534 tcp_set_state(sk, TCP_SYN_RECV);
5536 if (tp->rx_opt.saw_tstamp) {
5537 tp->rx_opt.tstamp_ok = 1;
5538 tcp_store_ts_recent(tp);
5539 tp->tcp_header_len =
5540 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5542 tp->tcp_header_len = sizeof(struct tcphdr);
5545 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5546 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5548 /* RFC1323: The window in SYN & SYN/ACK segments is
5551 tp->snd_wnd = ntohs(th->window);
5552 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5553 tp->max_window = tp->snd_wnd;
5555 TCP_ECN_rcv_syn(tp, th);
5558 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5559 tcp_initialize_rcv_mss(sk);
5561 tcp_send_synack(sk);
5563 /* Note, we could accept data and URG from this segment.
5564 * There are no obstacles to make this (except that we must
5565 * either change tcp_recvmsg() to prevent it from returning data
5566 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5568 * However, if we ignore data in ACKless segments sometimes,
5569 * we have no reasons to accept it sometimes.
5570 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5571 * is not flawless. So, discard packet for sanity.
5572 * Uncomment this return to process the data.
5579 /* "fifth, if neither of the SYN or RST bits is set then
5580 * drop the segment and return."
5584 tcp_clear_options(&tp->rx_opt);
5585 tp->rx_opt.mss_clamp = saved_clamp;
5589 tcp_clear_options(&tp->rx_opt);
5590 tp->rx_opt.mss_clamp = saved_clamp;
5595 * This function implements the receiving procedure of RFC 793 for
5596 * all states except ESTABLISHED and TIME_WAIT.
5597 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5598 * address independent.
5601 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5602 const struct tcphdr *th, unsigned int len)
5604 struct tcp_sock *tp = tcp_sk(sk);
5605 struct inet_connection_sock *icsk = inet_csk(sk);
5606 struct request_sock *req;
5610 tp->rx_opt.saw_tstamp = 0;
5612 switch (sk->sk_state) {
5626 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5629 /* Now we have several options: In theory there is
5630 * nothing else in the frame. KA9Q has an option to
5631 * send data with the syn, BSD accepts data with the
5632 * syn up to the [to be] advertised window and
5633 * Solaris 2.1 gives you a protocol error. For now
5634 * we just ignore it, that fits the spec precisely
5635 * and avoids incompatibilities. It would be nice in
5636 * future to drop through and process the data.
5638 * Now that TTCP is starting to be used we ought to
5640 * But, this leaves one open to an easy denial of
5641 * service attack, and SYN cookies can't defend
5642 * against this problem. So, we drop the data
5643 * in the interest of security over speed unless
5644 * it's still in use.
5652 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5656 /* Do step6 onward by hand. */
5657 tcp_urg(sk, skb, th);
5659 tcp_data_snd_check(sk);
5663 req = tp->fastopen_rsk;
5665 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5666 sk->sk_state != TCP_FIN_WAIT1);
5668 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5672 if (!th->ack && !th->rst)
5675 if (!tcp_validate_incoming(sk, skb, th, 0))
5678 /* step 5: check the ACK field */
5679 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5680 FLAG_UPDATE_TS_RECENT) > 0;
5682 switch (sk->sk_state) {
5687 /* Once we leave TCP_SYN_RECV, we no longer need req
5691 tp->total_retrans = req->num_retrans;
5692 reqsk_fastopen_remove(sk, req, false);
5694 /* Make sure socket is routed, for correct metrics. */
5695 icsk->icsk_af_ops->rebuild_header(sk);
5696 tcp_init_congestion_control(sk);
5699 tp->copied_seq = tp->rcv_nxt;
5700 tcp_init_buffer_space(sk);
5703 tcp_set_state(sk, TCP_ESTABLISHED);
5704 sk->sk_state_change(sk);
5706 /* Note, that this wakeup is only for marginal crossed SYN case.
5707 * Passively open sockets are not waked up, because
5708 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5711 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5713 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5714 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5715 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5716 tcp_synack_rtt_meas(sk, req);
5718 if (tp->rx_opt.tstamp_ok)
5719 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5722 /* Re-arm the timer because data may have been sent out.
5723 * This is similar to the regular data transmission case
5724 * when new data has just been ack'ed.
5726 * (TFO) - we could try to be more aggressive and
5727 * retransmitting any data sooner based on when they
5732 tcp_init_metrics(sk);
5734 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5735 tp->lsndtime = tcp_time_stamp;
5737 tcp_initialize_rcv_mss(sk);
5738 tcp_fast_path_on(tp);
5741 case TCP_FIN_WAIT1: {
5742 struct dst_entry *dst;
5745 /* If we enter the TCP_FIN_WAIT1 state and we are a
5746 * Fast Open socket and this is the first acceptable
5747 * ACK we have received, this would have acknowledged
5748 * our SYNACK so stop the SYNACK timer.
5751 /* Return RST if ack_seq is invalid.
5752 * Note that RFC793 only says to generate a
5753 * DUPACK for it but for TCP Fast Open it seems
5754 * better to treat this case like TCP_SYN_RECV
5759 /* We no longer need the request sock. */
5760 reqsk_fastopen_remove(sk, req, false);
5763 if (tp->snd_una != tp->write_seq)
5766 tcp_set_state(sk, TCP_FIN_WAIT2);
5767 sk->sk_shutdown |= SEND_SHUTDOWN;
5769 dst = __sk_dst_get(sk);
5773 if (!sock_flag(sk, SOCK_DEAD)) {
5774 /* Wake up lingering close() */
5775 sk->sk_state_change(sk);
5779 if (tp->linger2 < 0 ||
5780 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5781 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5783 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5787 tmo = tcp_fin_time(sk);
5788 if (tmo > TCP_TIMEWAIT_LEN) {
5789 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5790 } else if (th->fin || sock_owned_by_user(sk)) {
5791 /* Bad case. We could lose such FIN otherwise.
5792 * It is not a big problem, but it looks confusing
5793 * and not so rare event. We still can lose it now,
5794 * if it spins in bh_lock_sock(), but it is really
5797 inet_csk_reset_keepalive_timer(sk, tmo);
5799 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5806 if (tp->snd_una == tp->write_seq) {
5807 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5813 if (tp->snd_una == tp->write_seq) {
5814 tcp_update_metrics(sk);
5821 /* step 6: check the URG bit */
5822 tcp_urg(sk, skb, th);
5824 /* step 7: process the segment text */
5825 switch (sk->sk_state) {
5826 case TCP_CLOSE_WAIT:
5829 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5833 /* RFC 793 says to queue data in these states,
5834 * RFC 1122 says we MUST send a reset.
5835 * BSD 4.4 also does reset.
5837 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5838 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5839 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5840 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5846 case TCP_ESTABLISHED:
5847 tcp_data_queue(sk, skb);
5852 /* tcp_data could move socket to TIME-WAIT */
5853 if (sk->sk_state != TCP_CLOSE) {
5854 tcp_data_snd_check(sk);
5855 tcp_ack_snd_check(sk);
5864 EXPORT_SYMBOL(tcp_rcv_state_process);